From 13f2d39f55b17515b9637eaf4a24536d1c203f62 Mon Sep 17 00:00:00 2001 From: Stan Moore Date: Mon, 31 Jul 2017 10:34:21 -0600 Subject: [PATCH] Update Kokkos library to v2.03.13 --- lib/kokkos/CHANGELOG.md | 41 + lib/kokkos/Makefile.kokkos | 75 +- lib/kokkos/Makefile.targets | 8 +- .../algorithms/unit_tests/TestOpenMP.cpp | 13 +- .../algorithms/unit_tests/TestRandom.hpp | 38 +- .../algorithms/unit_tests/UnitTestMain.cpp | 8 +- .../benchmarks/bytes_and_flops/main.cpp | 9 +- lib/kokkos/benchmarks/gather/main.cpp | 2 +- .../benchmarks/policy_performance/Makefile | 44 + .../benchmarks/policy_performance/main.cpp | 170 + .../policy_performance/policy_perf_test.hpp | 354 + .../script_basic_testing.sh | 53 + .../policy_performance/script_sample_usage.sh | 126 + lib/kokkos/bin/hpcbind | 454 + lib/kokkos/bin/kokkos-bind | 221 + lib/kokkos/bin/runtest | 165 + lib/kokkos/cmake/kokkos.cmake | 4 + lib/kokkos/config/master_history.txt | 1 + lib/kokkos/config/query_cuda_arch.cpp | 24 + lib/kokkos/config/test_all_sandia | 11 +- .../shepard_jenkins_run_script_pthread_intel | 6 +- .../shepard_jenkins_run_script_serial_intel | 6 +- .../containers/performance_tests/Makefile | 1 - .../containers/performance_tests/TestMain.cpp | 11 +- .../performance_tests/TestOpenMP.cpp | 21 +- lib/kokkos/containers/src/Kokkos_DualView.hpp | 2 +- .../containers/src/Kokkos_DynRankView.hpp | 98 +- .../containers/src/Kokkos_DynamicView.hpp | 30 +- lib/kokkos/containers/unit_tests/TestCuda.cpp | 6 + .../containers/unit_tests/TestOpenMP.cpp | 15 +- .../containers/unit_tests/TestSerial.cpp | 6 + .../containers/unit_tests/TestThreads.cpp | 6 + .../TestViewCtorPropEmbeddedDim.hpp | 213 + .../containers/unit_tests/UnitTestMain.cpp | 10 +- lib/kokkos/core/perf_test/Makefile | 1 - lib/kokkos/core/perf_test/PerfTestMain.cpp | 10 +- .../KokkosExp_Cuda_IterateTile_Refactor.hpp | 2715 +++ lib/kokkos/core/src/Cuda/Kokkos_CudaExec.hpp | 108 +- lib/kokkos/core/src/Cuda/Kokkos_CudaSpace.cpp | 96 +- lib/kokkos/core/src/Cuda/Kokkos_Cuda_Impl.cpp | 85 +- .../core/src/Cuda/Kokkos_Cuda_Locks.cpp | 119 + .../core/src/Cuda/Kokkos_Cuda_Locks.hpp | 166 + .../core/src/Cuda/Kokkos_Cuda_Parallel.hpp | 404 +- .../core/src/Cuda/Kokkos_Cuda_ReduceScan.hpp | 19 +- lib/kokkos/core/src/Cuda/Kokkos_Cuda_Task.cpp | 128 +- lib/kokkos/core/src/Cuda/Kokkos_Cuda_Task.hpp | 23 +- lib/kokkos/core/src/Cuda/Kokkos_Cuda_Team.hpp | 26 +- .../core/src/Cuda/Kokkos_Cuda_UniqueToken.hpp | 133 + lib/kokkos/core/src/Cuda/Kokkos_Cuda_View.hpp | 6 +- .../src/Cuda/Kokkos_Cuda_WorkGraphPolicy.hpp | 119 + .../core/src/KokkosExp_MDRangePolicy.hpp | 175 +- lib/kokkos/core/src/Kokkos_Atomic.hpp | 33 +- lib/kokkos/core/src/Kokkos_Concepts.hpp | 16 + lib/kokkos/core/src/Kokkos_Core.hpp | 15 +- lib/kokkos/core/src/Kokkos_Core_fwd.hpp | 3 + lib/kokkos/core/src/Kokkos_Crs.hpp | 333 + lib/kokkos/core/src/Kokkos_Cuda.hpp | 3 +- lib/kokkos/core/src/Kokkos_CudaSpace.hpp | 6 +- lib/kokkos/core/src/Kokkos_ExecPolicy.hpp | 40 + lib/kokkos/core/src/Kokkos_HBWSpace.hpp | 8 - lib/kokkos/core/src/Kokkos_HostSpace.hpp | 10 +- lib/kokkos/core/src/Kokkos_Layout.hpp | 2 + lib/kokkos/core/src/Kokkos_Macros.hpp | 11 +- lib/kokkos/core/src/Kokkos_MasterLock.hpp | 73 + lib/kokkos/core/src/Kokkos_MemoryPool.hpp | 182 +- lib/kokkos/core/src/Kokkos_MemoryTraits.hpp | 16 +- lib/kokkos/core/src/Kokkos_OpenMP.hpp | 172 +- lib/kokkos/core/src/Kokkos_Parallel.hpp | 64 +- .../core/src/Kokkos_Parallel_Reduce.hpp | 19 +- lib/kokkos/core/src/Kokkos_Serial.hpp | 228 + lib/kokkos/core/src/Kokkos_TaskScheduler.hpp | 108 +- lib/kokkos/core/src/Kokkos_UniqueToken.hpp | 88 + lib/kokkos/core/src/Kokkos_View.hpp | 233 +- .../core/src/Kokkos_WorkGraphPolicy.hpp | 265 + .../core/src/OpenMP/Kokkos_OpenMP_Exec.cpp | 431 +- .../core/src/OpenMP/Kokkos_OpenMP_Exec.hpp | 490 +- .../src/OpenMP/Kokkos_OpenMP_Parallel.hpp | 419 +- .../core/src/OpenMP/Kokkos_OpenMP_Task.cpp | 94 +- .../core/src/OpenMP/Kokkos_OpenMP_Task.hpp | 4 +- .../core/src/OpenMP/Kokkos_OpenMP_Team.hpp | 245 + .../OpenMP/Kokkos_OpenMP_WorkGraphPolicy.hpp | 107 + .../OpenMPTarget/Kokkos_OpenMPTarget_Exec.hpp | 28 +- .../core/src/Qthreads/Kokkos_QthreadsExec.hpp | 2 +- .../core/src/Threads/Kokkos_ThreadsExec.cpp | 12 +- .../core/src/Threads/Kokkos_ThreadsExec.hpp | 70 +- .../core/src/Threads/Kokkos_ThreadsTeam.hpp | 6 +- .../src/Threads/Kokkos_Threads_Parallel.hpp | 267 + .../Kokkos_Threads_WorkGraphPolicy.hpp | 115 + .../src/impl/KokkosExp_Host_IterateTile.hpp | 566 +- .../core/src/impl/Kokkos_AnalyzePolicy.hpp | 35 +- .../Kokkos_Atomic_Compare_Exchange_Strong.hpp | 34 +- .../core/src/impl/Kokkos_Atomic_Decrement.hpp | 19 + .../core/src/impl/Kokkos_Atomic_Exchange.hpp | 32 + .../core/src/impl/Kokkos_Atomic_Fetch_Add.hpp | 62 +- .../core/src/impl/Kokkos_Atomic_Fetch_And.hpp | 32 +- .../core/src/impl/Kokkos_Atomic_Fetch_Or.hpp | 32 +- .../core/src/impl/Kokkos_Atomic_Fetch_Sub.hpp | 44 +- .../core/src/impl/Kokkos_Atomic_Increment.hpp | 16 + lib/kokkos/core/src/impl/Kokkos_Core.cpp | 20 +- .../core/src/impl/Kokkos_FunctorAdapter.hpp | 13 +- lib/kokkos/core/src/impl/Kokkos_HBWSpace.cpp | 56 - lib/kokkos/core/src/impl/Kokkos_HostSpace.cpp | 59 - .../core/src/impl/Kokkos_HostThreadTeam.cpp | 30 +- .../core/src/impl/Kokkos_HostThreadTeam.hpp | 33 +- .../src/impl/Kokkos_Profiling_Interface.cpp | 102 +- .../src/impl/Kokkos_Profiling_Interface.hpp | 53 +- .../core/src/impl/Kokkos_Rendezvous.cpp | 208 + .../core/src/impl/Kokkos_Rendezvous.hpp | 87 + lib/kokkos/core/src/impl/Kokkos_Serial.cpp | 4 +- .../core/src/impl/Kokkos_Serial_Task.cpp | 4 +- .../core/src/impl/Kokkos_Serial_Task.hpp | 2 +- .../impl/Kokkos_Serial_WorkGraphPolicy.hpp | 102 + .../core/src/impl/Kokkos_SharedAlloc.cpp | 24 +- .../core/src/impl/Kokkos_SharedAlloc.hpp | 16 +- lib/kokkos/core/src/impl/Kokkos_Spinwait.cpp | 210 + ...okkos_spinwait.hpp => Kokkos_Spinwait.hpp} | 17 + lib/kokkos/core/src/impl/Kokkos_TaskQueue.hpp | 606 +- .../core/src/impl/Kokkos_TaskQueue_impl.hpp | 86 +- lib/kokkos/core/src/impl/Kokkos_ViewArray.hpp | 14 +- lib/kokkos/core/src/impl/Kokkos_ViewCtor.hpp | 21 + .../core/src/impl/Kokkos_ViewMapping.hpp | 137 +- lib/kokkos/core/src/impl/Kokkos_ViewTile.hpp | 12 +- lib/kokkos/core/src/impl/Kokkos_spinwait.cpp | 183 - lib/kokkos/core/unit_test/CMakeLists.txt | 6 + lib/kokkos/core/unit_test/Makefile | 8 +- lib/kokkos/core/unit_test/TestAggregate.hpp | 2 +- .../unit_test/TestDefaultDeviceTypeInit.hpp | 21 +- lib/kokkos/core/unit_test/TestMDRange.hpp | 882 +- lib/kokkos/core/unit_test/TestMemoryPool.hpp | 228 +- lib/kokkos/core/unit_test/TestRange.hpp | 25 + lib/kokkos/core/unit_test/TestResize.hpp | 140 + .../core/unit_test/TestTaskScheduler.hpp | 32 +- lib/kokkos/core/unit_test/TestTeamVector.hpp | 12 + lib/kokkos/core/unit_test/TestTile.hpp | 2 +- lib/kokkos/core/unit_test/TestUniqueToken.hpp | 138 + .../unit_test/TestViewCtorPropEmbeddedDim.hpp | 160 + .../core/unit_test/TestViewMapping_a.hpp | 114 +- lib/kokkos/core/unit_test/TestViewSubview.hpp | 224 +- lib/kokkos/core/unit_test/TestWorkGraph.hpp | 172 + lib/kokkos/core/unit_test/UnitTestMain.cpp | 1 + .../core/unit_test/UnitTestMainInit.cpp | 2 + .../core/unit_test/cuda/TestCuda_Other.cpp | 2 + .../unit_test/cuda/TestCuda_UniqueToken.cpp | 46 + .../unit_test/cuda/TestCuda_WorkGraph.cpp | 45 + .../default/TestDefaultDeviceTypeResize.cpp | 57 + .../core/unit_test/openmp/TestOpenMP.hpp | 19 +- .../unit_test/openmp/TestOpenMP_Other.cpp | 90 + .../openmp/TestOpenMP_UniqueToken.cpp | 46 + .../unit_test/openmp/TestOpenMP_WorkGraph.cpp | 45 + .../unit_test/serial/TestSerial_Other.cpp | 2 + .../unit_test/serial/TestSerial_WorkGraph.cpp | 45 + .../unit_test/threads/TestThreads_Other.cpp | 2 + .../threads/TestThreads_WorkGraph.cpp | 45 + lib/kokkos/example/cmake_build/CMakeLists.txt | 2 + lib/kokkos/example/feint/main.cpp | 22 +- .../example/global_2_local_ids/G2L_Main.cpp | 11 +- lib/kokkos/example/grow_array/main.cpp | 2 +- .../example/tutorial/03_simple_view/Makefile | 1 + .../example/tutorial/Advanced_Views/Makefile | 86 +- .../example/tutorial/Algorithms/Makefile | 12 +- .../Hierarchical_Parallelism/Makefile | 60 +- lib/kokkos/example/tutorial/Makefile | 132 +- .../tutorial/launch_bounds/CMakeLists.txt | 10 + .../example/tutorial/launch_bounds/Makefile | 56 + .../launch_bounds/launch_bounds_reduce.cpp | 173 + lib/kokkos/generate_makefile.bash | 191 +- lib/kokkos/tpls/gtest/gtest/LICENSE | 28 + lib/kokkos/tpls/gtest/gtest/README | 13 + lib/kokkos/tpls/gtest/gtest/gtest-all.cc | 9594 ++++++++ lib/kokkos/tpls/gtest/gtest/gtest-test-part.h | 1 + lib/kokkos/tpls/gtest/gtest/gtest.h | 20065 ++++++++++++++++ 171 files changed, 44039 insertions(+), 2807 deletions(-) create mode 100644 lib/kokkos/benchmarks/policy_performance/Makefile create mode 100644 lib/kokkos/benchmarks/policy_performance/main.cpp create mode 100644 lib/kokkos/benchmarks/policy_performance/policy_perf_test.hpp create mode 100755 lib/kokkos/benchmarks/policy_performance/script_basic_testing.sh create mode 100755 lib/kokkos/benchmarks/policy_performance/script_sample_usage.sh create mode 100755 lib/kokkos/bin/hpcbind create mode 100755 lib/kokkos/bin/kokkos-bind create mode 100755 lib/kokkos/bin/runtest create mode 100644 lib/kokkos/config/query_cuda_arch.cpp create mode 100644 lib/kokkos/containers/unit_tests/TestViewCtorPropEmbeddedDim.hpp create mode 100644 lib/kokkos/core/src/Cuda/KokkosExp_Cuda_IterateTile_Refactor.hpp create mode 100644 lib/kokkos/core/src/Cuda/Kokkos_Cuda_Locks.cpp create mode 100644 lib/kokkos/core/src/Cuda/Kokkos_Cuda_Locks.hpp create mode 100644 lib/kokkos/core/src/Cuda/Kokkos_Cuda_UniqueToken.hpp create mode 100644 lib/kokkos/core/src/Cuda/Kokkos_Cuda_WorkGraphPolicy.hpp create mode 100644 lib/kokkos/core/src/Kokkos_Crs.hpp create mode 100644 lib/kokkos/core/src/Kokkos_MasterLock.hpp create mode 100644 lib/kokkos/core/src/Kokkos_UniqueToken.hpp create mode 100644 lib/kokkos/core/src/Kokkos_WorkGraphPolicy.hpp create mode 100644 lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Team.hpp create mode 100644 lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_WorkGraphPolicy.hpp create mode 100644 lib/kokkos/core/src/Threads/Kokkos_Threads_WorkGraphPolicy.hpp create mode 100644 lib/kokkos/core/src/impl/Kokkos_Rendezvous.cpp create mode 100644 lib/kokkos/core/src/impl/Kokkos_Rendezvous.hpp create mode 100644 lib/kokkos/core/src/impl/Kokkos_Serial_WorkGraphPolicy.hpp create mode 100644 lib/kokkos/core/src/impl/Kokkos_Spinwait.cpp rename lib/kokkos/core/src/impl/{Kokkos_spinwait.hpp => Kokkos_Spinwait.hpp} (82%) delete mode 100644 lib/kokkos/core/src/impl/Kokkos_spinwait.cpp create mode 100644 lib/kokkos/core/unit_test/TestResize.hpp create mode 100644 lib/kokkos/core/unit_test/TestUniqueToken.hpp create mode 100644 lib/kokkos/core/unit_test/TestViewCtorPropEmbeddedDim.hpp create mode 100644 lib/kokkos/core/unit_test/TestWorkGraph.hpp create mode 100644 lib/kokkos/core/unit_test/cuda/TestCuda_UniqueToken.cpp create mode 100644 lib/kokkos/core/unit_test/cuda/TestCuda_WorkGraph.cpp create mode 100644 lib/kokkos/core/unit_test/default/TestDefaultDeviceTypeResize.cpp create mode 100644 lib/kokkos/core/unit_test/openmp/TestOpenMP_UniqueToken.cpp create mode 100644 lib/kokkos/core/unit_test/openmp/TestOpenMP_WorkGraph.cpp create mode 100644 lib/kokkos/core/unit_test/serial/TestSerial_WorkGraph.cpp create mode 100644 lib/kokkos/core/unit_test/threads/TestThreads_WorkGraph.cpp create mode 100644 lib/kokkos/example/tutorial/launch_bounds/CMakeLists.txt create mode 100644 lib/kokkos/example/tutorial/launch_bounds/Makefile create mode 100644 lib/kokkos/example/tutorial/launch_bounds/launch_bounds_reduce.cpp create mode 100644 lib/kokkos/tpls/gtest/gtest/LICENSE create mode 100644 lib/kokkos/tpls/gtest/gtest/README create mode 100644 lib/kokkos/tpls/gtest/gtest/gtest-all.cc create mode 120000 lib/kokkos/tpls/gtest/gtest/gtest-test-part.h create mode 100644 lib/kokkos/tpls/gtest/gtest/gtest.h diff --git a/lib/kokkos/CHANGELOG.md b/lib/kokkos/CHANGELOG.md index acb54ff22f..3fe9e46111 100644 --- a/lib/kokkos/CHANGELOG.md +++ b/lib/kokkos/CHANGELOG.md @@ -1,5 +1,46 @@ # Change Log +## [2.03.13](https://github.com/kokkos/kokkos/tree/2.03.13) (2017-07-27) +[Full Changelog](https://github.com/kokkos/kokkos/compare/2.03.05...2.03.13) + +**Implemented enhancements:** + +- Disallow enabling both OpenMP and Threads in the same executable [\#406](https://github.com/kokkos/kokkos/issues/406) +- Make Kokkos::OpenMP respect OMP environment even if hwloc is available [\#630](https://github.com/kokkos/kokkos/issues/630) +- Improve Atomics Performance on KNL/Broadwell where PREFETCHW/RFO is Available [\#898](https://github.com/kokkos/kokkos/issues/898) +- Kokkos::resize should test whether dimensions have changed before resizing [\#904](https://github.com/kokkos/kokkos/issues/904) +- Develop performance-regression/acceptance tests [\#737](https://github.com/kokkos/kokkos/issues/737) +- Make the deep\_copy Profiling hook a start/end system [\#890](https://github.com/kokkos/kokkos/issues/890) +- Add deep\_copy Profiling hook [\#843](https://github.com/kokkos/kokkos/issues/843) +- Append tag name to parallel construct name for Profiling [\#842](https://github.com/kokkos/kokkos/issues/842) +- Add view label to `View bounds error` message for CUDA backend [\#870](https://github.com/kokkos/kokkos/issues/870) +- Disable printing the loaded profiling library [\#824](https://github.com/kokkos/kokkos/issues/824) +- "Declared but never referenced" warnings [\#853](https://github.com/kokkos/kokkos/issues/853) +- Warnings about lock\_address\_cuda\_space [\#852](https://github.com/kokkos/kokkos/issues/852) +- WorkGraph execution policy [\#771](https://github.com/kokkos/kokkos/issues/771) +- Simplify makefiles by guarding compilation with appropriate KOKKOS\_ENABLE\_\#\#\# macros [\#716](https://github.com/kokkos/kokkos/issues/716) +- Cmake build: wrong include install directory [\#668](https://github.com/kokkos/kokkos/issues/668) +- Derived View type and allocation [\#566](https://github.com/kokkos/kokkos/issues/566) +- Fix Compiler warnings when compiling core unit tests for Cuda [\#214](https://github.com/kokkos/kokkos/issues/214) + +**Fixed bugs:** + +- Out-of-bounds read in Kokkos\_Layout.hpp [\#975](https://github.com/kokkos/kokkos/issues/975) +- CudaClang: Fix failing test with Clang 4.0 [\#941](https://github.com/kokkos/kokkos/issues/941) +- Respawn when memory pool allocation fails \(not available memory\) [\#940](https://github.com/kokkos/kokkos/issues/940) +- Memory pool aborts on zero allocation request, returns NULL for \< minimum [\#939](https://github.com/kokkos/kokkos/issues/939) +- Error with TaskScheduler query of underlying memory pool [\#917](https://github.com/kokkos/kokkos/issues/917) +- Profiling::\*Callee static variables declared in header [\#863](https://github.com/kokkos/kokkos/issues/863) +- calling \*Space::name\(\) causes compile error [\#862](https://github.com/kokkos/kokkos/issues/862) +- bug in Profiling::deallocateData [\#860](https://github.com/kokkos/kokkos/issues/860) +- task\_depend test failing, CUDA 8.0 + Pascal + RDC [\#829](https://github.com/kokkos/kokkos/issues/829) +- \[develop branch\] Standalone cmake issues [\#826](https://github.com/kokkos/kokkos/issues/826) +- Kokkos CUDA failes to compile with OMPI\_CXX and MPICH\_CXX wrappers [\#776](https://github.com/kokkos/kokkos/issues/776) +- Task Team reduction on Pascal [\#767](https://github.com/kokkos/kokkos/issues/767) +- CUDA stack overflow with TaskDAG test [\#758](https://github.com/kokkos/kokkos/issues/758) +- TeamVector test on Cuda [\#670](https://github.com/kokkos/kokkos/issues/670) +- Clang 4.0 Cuda Build broken again [\#560](https://github.com/kokkos/kokkos/issues/560) + ## [2.03.05](https://github.com/kokkos/kokkos/tree/2.03.05) (2017-05-27) [Full Changelog](https://github.com/kokkos/kokkos/compare/2.03.00...2.03.05) diff --git a/lib/kokkos/Makefile.kokkos b/lib/kokkos/Makefile.kokkos index 24cd772e00..acb6c36134 100644 --- a/lib/kokkos/Makefile.kokkos +++ b/lib/kokkos/Makefile.kokkos @@ -33,6 +33,7 @@ KOKKOS_INTERNAL_USE_LIBRT := $(strip $(shell echo $(KOKKOS_USE_TPLS) | grep "lib KOKKOS_INTERNAL_USE_MEMKIND := $(strip $(shell echo $(KOKKOS_USE_TPLS) | grep "experimental_memkind" | wc -l)) # Check for advanced settings. +KOKKOS_INTERNAL_ENABLE_COMPILER_WARNINGS := $(strip $(shell echo $(KOKKOS_OPTIONS) | grep "compiler_warnings" | wc -l)) KOKKOS_INTERNAL_OPT_RANGE_AGGRESSIVE_VECTORIZATION := $(strip $(shell echo $(KOKKOS_OPTIONS) | grep "aggressive_vectorization" | wc -l)) KOKKOS_INTERNAL_DISABLE_PROFILING := $(strip $(shell echo $(KOKKOS_OPTIONS) | grep "disable_profiling" | wc -l)) KOKKOS_INTERNAL_DISABLE_DUALVIEW_MODIFY_CHECK := $(strip $(shell echo $(KOKKOS_OPTIONS) | grep "disable_dualview_modify_check" | wc -l)) @@ -78,6 +79,12 @@ KOKKOS_INTERNAL_COMPILER_PGI := $(strip $(shell $(CXX) --version 2 KOKKOS_INTERNAL_COMPILER_XL := $(strip $(shell $(CXX) -qversion 2>&1 | grep XL | wc -l)) KOKKOS_INTERNAL_COMPILER_CRAY := $(strip $(shell $(CXX) -craype-verbose 2>&1 | grep "CC-" | wc -l)) KOKKOS_INTERNAL_COMPILER_NVCC := $(strip $(shell $(CXX) --version 2>&1 | grep nvcc | wc -l)) +ifneq ($(OMPI_CXX),) + KOKKOS_INTERNAL_COMPILER_NVCC := $(strip $(shell $(OMPI_CXX) --version 2>&1 | grep nvcc | wc -l)) +endif +ifneq ($(MPICH_CXX),) + KOKKOS_INTERNAL_COMPILER_NVCC := $(strip $(shell $(MPICH_CXX) --version 2>&1 | grep nvcc | wc -l)) +endif KOKKOS_INTERNAL_COMPILER_CLANG := $(strip $(shell $(CXX) --version 2>&1 | grep clang | wc -l)) KOKKOS_INTERNAL_COMPILER_APPLE_CLANG := $(strip $(shell $(CXX) --version 2>&1 | grep "apple-darwin" | wc -l)) ifneq ($(OMPI_CXX),) @@ -111,6 +118,36 @@ ifeq ($(KOKKOS_INTERNAL_COMPILER_CLANG), 1) endif endif +# Set compiler warnings flags. +ifeq ($(KOKKOS_INTERNAL_ENABLE_COMPILER_WARNINGS), 1) + ifeq ($(KOKKOS_INTERNAL_COMPILER_PGI), 1) + # TODO check if PGI accepts GNU style warnings + KOKKOS_INTERNAL_COMPILER_WARNINGS = + else + ifeq ($(KOKKOS_INTERNAL_COMPILER_CLANG), 1) + KOKKOS_INTERNAL_COMPILER_WARNINGS = -Wall -Wshadow -pedantic -Wsign-compare -Wtype-limits -Wuninitialized + else + ifeq ($(KOKKOS_INTERNAL_COMPILER_APPLE_CLANG), 1) + KOKKOS_INTERNAL_COMPILER_WARNINGS = -Wall -Wshadow -pedantic -Wsign-compare -Wtype-limits -Wuninitialized + else + ifeq ($(KOKKOS_INTERNAL_COMPILER_XL), 1) + KOKKOS_INTERNAL_COMPILER_WARNINGS = -Wall -Wshadow -pedantic -Wsign-compare -Wtype-limits -Wuninitialized + else + ifeq ($(KOKKOS_INTERNAL_COMPILER_CRAY), 1) + # TODO check if cray accepts GNU style warnings + KOKKOS_INTERNAL_COMPILER_WARNINGS = + else + #gcc + KOKKOS_INTERNAL_COMPILER_WARNINGS = -Wall -Wshadow -pedantic -Wsign-compare -Wtype-limits -Wignored-qualifiers -Wempty-body -Wclobbered -Wuninitialized + endif + endif + endif + endif + endif +else + KOKKOS_INTERNAL_COMPILER_WARNINGS = +endif + # Set OpenMP flags. ifeq ($(KOKKOS_INTERNAL_COMPILER_PGI), 1) KOKKOS_INTERNAL_OPENMP_FLAG := -mp @@ -162,6 +199,7 @@ endif # Intel based. KOKKOS_INTERNAL_USE_ARCH_KNC := $(strip $(shell echo $(KOKKOS_ARCH) | grep KNC | wc -l)) +KOKKOS_INTERNAL_USE_ARCH_WSM := $(strip $(shell echo $(KOKKOS_ARCH) | grep WSM | wc -l)) KOKKOS_INTERNAL_USE_ARCH_SNB := $(strip $(shell echo $(KOKKOS_ARCH) | grep SNB | wc -l)) KOKKOS_INTERNAL_USE_ARCH_HSW := $(strip $(shell echo $(KOKKOS_ARCH) | grep HSW | wc -l)) KOKKOS_INTERNAL_USE_ARCH_BDW := $(strip $(shell echo $(KOKKOS_ARCH) | grep BDW | wc -l)) @@ -229,13 +267,14 @@ KOKKOS_INTERNAL_USE_ARCH_IBM := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_ KOKKOS_INTERNAL_USE_ARCH_AMDAVX := $(strip $(shell echo $(KOKKOS_ARCH) | grep AMDAVX | wc -l)) # Any AVX? +KOKKOS_INTERNAL_USE_ARCH_SSE42 := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_WSM) | bc )) KOKKOS_INTERNAL_USE_ARCH_AVX := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_SNB)+$(KOKKOS_INTERNAL_USE_ARCH_AMDAVX) | bc )) KOKKOS_INTERNAL_USE_ARCH_AVX2 := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_HSW)+$(KOKKOS_INTERNAL_USE_ARCH_BDW) | bc )) KOKKOS_INTERNAL_USE_ARCH_AVX512MIC := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_KNL) | bc )) KOKKOS_INTERNAL_USE_ARCH_AVX512XEON := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_SKX) | bc )) # Decide what ISA level we are able to support. -KOKKOS_INTERNAL_USE_ISA_X86_64 := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_SNB)+$(KOKKOS_INTERNAL_USE_ARCH_HSW)+$(KOKKOS_INTERNAL_USE_ARCH_BDW)+$(KOKKOS_INTERNAL_USE_ARCH_KNL)+$(KOKKOS_INTERNAL_USE_ARCH_SKX) | bc )) +KOKKOS_INTERNAL_USE_ISA_X86_64 := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_WSM)+$(KOKKOS_INTERNAL_USE_ARCH_SNB)+$(KOKKOS_INTERNAL_USE_ARCH_HSW)+$(KOKKOS_INTERNAL_USE_ARCH_BDW)+$(KOKKOS_INTERNAL_USE_ARCH_KNL)+$(KOKKOS_INTERNAL_USE_ARCH_SKX) | bc )) KOKKOS_INTERNAL_USE_ISA_KNC := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_KNC) | bc )) KOKKOS_INTERNAL_USE_ISA_POWERPCLE := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_POWER8)+$(KOKKOS_INTERNAL_USE_ARCH_POWER9) | bc )) @@ -243,7 +282,7 @@ KOKKOS_INTERNAL_USE_ISA_POWERPCLE := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ KOKKOS_INTERNAL_USE_TM := $(strip $(shell echo $(KOKKOS_INTERNAL_USE_ARCH_BDW)+$(KOKKOS_INTERNAL_USE_ARCH_SKX) | bc )) # Incompatible flags? -KOKKOS_INTERNAL_USE_ARCH_MULTIHOST := $(strip $(shell echo "$(KOKKOS_INTERNAL_USE_ARCH_AVX)+$(KOKKOS_INTERNAL_USE_ARCH_AVX2)+$(KOKKOS_INTERNAL_USE_ARCH_AVX512MIC)+$(KOKKOS_INTERNAL_USE_ARCH_AVX512XEON)+$(KOKKOS_INTERNAL_USE_ARCH_KNC)+$(KOKKOS_INTERNAL_USE_ARCH_IBM)+$(KOKKOS_INTERNAL_USE_ARCH_ARM)>1" | bc )) +KOKKOS_INTERNAL_USE_ARCH_MULTIHOST := $(strip $(shell echo "$(KOKKOS_INTERNAL_USE_ARCH_SSE42)+$(KOKKOS_INTERNAL_USE_ARCH_AVX)+$(KOKKOS_INTERNAL_USE_ARCH_AVX2)+$(KOKKOS_INTERNAL_USE_ARCH_AVX512MIC)+$(KOKKOS_INTERNAL_USE_ARCH_AVX512XEON)+$(KOKKOS_INTERNAL_USE_ARCH_KNC)+$(KOKKOS_INTERNAL_USE_ARCH_IBM)+$(KOKKOS_INTERNAL_USE_ARCH_ARM)>1" | bc )) KOKKOS_INTERNAL_USE_ARCH_MULTIGPU := $(strip $(shell echo "$(KOKKOS_INTERNAL_USE_ARCH_NVIDIA)>1" | bc)) ifeq ($(KOKKOS_INTERNAL_USE_ARCH_MULTIHOST), 1) @@ -257,12 +296,10 @@ endif KOKKOS_CPPFLAGS = -I./ -I$(KOKKOS_PATH)/core/src -I$(KOKKOS_PATH)/containers/src -I$(KOKKOS_PATH)/algorithms/src -# No warnings: KOKKOS_CXXFLAGS = -# INTEL and CLANG warnings: -#KOKKOS_CXXFLAGS = -Wall -Wshadow -pedantic -Wsign-compare -Wtype-limits -Wuninitialized -# GCC warnings: -#KOKKOS_CXXFLAGS = -Wall -Wshadow -pedantic -Wsign-compare -Wtype-limits -Wuninitialized -Wignored-qualifiers -Wempty-body -Wclobbered +ifeq ($(KOKKOS_INTERNAL_ENABLE_COMPILER_WARNINGS), 1) + KOKKOS_CXXFLAGS += $(KOKKOS_INTERNAL_COMPILER_WARNINGS) +endif KOKKOS_LIBS = -lkokkos -ldl KOKKOS_LDFLAGS = -L$(shell pwd) @@ -486,6 +523,28 @@ ifeq ($(KOKKOS_INTERNAL_USE_ARCH_ARMV8_THUNDERX), 1) endif endif +ifeq ($(KOKKOS_INTERNAL_USE_ARCH_SSE42), 1) + tmp := $(shell echo "\#define KOKKOS_ARCH_SSE42 1" >> KokkosCore_config.tmp ) + + ifeq ($(KOKKOS_INTERNAL_COMPILER_INTEL), 1) + KOKKOS_CXXFLAGS += -xSSE4.2 + KOKKOS_LDFLAGS += -xSSE4.2 + else + ifeq ($(KOKKOS_INTERNAL_COMPILER_CRAY), 1) + + else + ifeq ($(KOKKOS_INTERNAL_COMPILER_PGI), 1) + KOKKOS_CXXFLAGS += -tp=nehalem + KOKKOS_LDFLAGS += -tp=nehalem + else + # Assume that this is a really a GNU compiler. + KOKKOS_CXXFLAGS += -msse4.2 + KOKKOS_LDFLAGS += -msse4.2 + endif + endif + endif +endif + ifeq ($(KOKKOS_INTERNAL_USE_ARCH_AVX), 1) tmp := $(shell echo "\#define KOKKOS_ARCH_AVX 1" >> KokkosCore_config.tmp ) @@ -689,7 +748,7 @@ ifeq ($(KOKKOS_INTERNAL_USE_CUDA), 1) endif endif -KOKKOS_INTERNAL_LS_CONFIG := $(shell ls KokkosCore_config.h) +KOKKOS_INTERNAL_LS_CONFIG := $(shell ls KokkosCore_config.h 2>&1) ifeq ($(KOKKOS_INTERNAL_LS_CONFIG), KokkosCore_config.h) KOKKOS_INTERNAL_NEW_CONFIG := $(strip $(shell diff KokkosCore_config.h KokkosCore_config.tmp | grep define | wc -l)) else diff --git a/lib/kokkos/Makefile.targets b/lib/kokkos/Makefile.targets index 3cb52a04cd..a9341a907c 100644 --- a/lib/kokkos/Makefile.targets +++ b/lib/kokkos/Makefile.targets @@ -20,8 +20,10 @@ Kokkos_TaskQueue.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/impl/Kokkos_Ta $(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/impl/Kokkos_TaskQueue.cpp Kokkos_HostThreadTeam.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/impl/Kokkos_HostThreadTeam.cpp $(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/impl/Kokkos_HostThreadTeam.cpp -Kokkos_spinwait.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/impl/Kokkos_spinwait.cpp - $(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/impl/Kokkos_spinwait.cpp +Kokkos_Spinwait.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/impl/Kokkos_Spinwait.cpp + $(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/impl/Kokkos_Spinwait.cpp +Kokkos_Rendezvous.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/impl/Kokkos_Rendezvous.cpp + $(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/impl/Kokkos_Rendezvous.cpp Kokkos_Profiling_Interface.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/impl/Kokkos_Profiling_Interface.cpp $(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/impl/Kokkos_Profiling_Interface.cpp Kokkos_SharedAlloc.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/impl/Kokkos_SharedAlloc.cpp @@ -36,6 +38,8 @@ Kokkos_CudaSpace.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/Cuda/Kokkos_Cu $(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/Cuda/Kokkos_CudaSpace.cpp Kokkos_Cuda_Task.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/Cuda/Kokkos_Cuda_Task.cpp $(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/Cuda/Kokkos_Cuda_Task.cpp +Kokkos_Cuda_Locks.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/Cuda/Kokkos_Cuda_Locks.cpp + $(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/Cuda/Kokkos_Cuda_Locks.cpp endif ifeq ($(KOKKOS_INTERNAL_USE_PTHREADS), 1) diff --git a/lib/kokkos/algorithms/unit_tests/TestOpenMP.cpp b/lib/kokkos/algorithms/unit_tests/TestOpenMP.cpp index 1e7ee68549..c2c118ce1a 100644 --- a/lib/kokkos/algorithms/unit_tests/TestOpenMP.cpp +++ b/lib/kokkos/algorithms/unit_tests/TestOpenMP.cpp @@ -61,14 +61,19 @@ protected: { std::cout << std::setprecision(5) << std::scientific; - unsigned threads_count = omp_get_max_threads(); + int threads_count = 0; + #pragma omp parallel + { + #pragma omp atomic + ++threads_count; + } - if ( Kokkos::hwloc::available() ) { - threads_count = Kokkos::hwloc::get_available_numa_count() * - Kokkos::hwloc::get_available_cores_per_numa(); + if (threads_count > 3) { + threads_count /= 2; } Kokkos::OpenMP::initialize( threads_count ); + Kokkos::OpenMP::print_configuration( std::cout ); } static void TearDownTestCase() diff --git a/lib/kokkos/algorithms/unit_tests/TestRandom.hpp b/lib/kokkos/algorithms/unit_tests/TestRandom.hpp index 9cf02f74b4..2771f1793d 100644 --- a/lib/kokkos/algorithms/unit_tests/TestRandom.hpp +++ b/lib/kokkos/algorithms/unit_tests/TestRandom.hpp @@ -1,12 +1,12 @@ //@HEADER // ************************************************************************ -// +// // Kokkos v. 2.0 // Copyright (2014) Sandia Corporation -// +// // Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, // the U.S. Government retains certain rights in this software. -// +// // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are // met: @@ -35,7 +35,7 @@ // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. // // Questions? Contact H. Carter Edwards (hcedwar@sandia.gov) -// +// // ************************************************************************ //@HEADER @@ -283,12 +283,12 @@ struct test_random_scalar { RandomGenerator& pool, unsigned int num_draws) { - using std::cerr; + using std::cout; using std::endl; using Kokkos::parallel_reduce; { - cerr << " -- Testing randomness properties" << endl; + cout << " -- Testing randomness properties" << endl; RandomProperties result; typedef test_random_functor functor_type; @@ -307,7 +307,7 @@ struct test_random_scalar { ( 1.5*tolerance > variance_eps)) ? 1:0; pass_covar = ((-2.0*tolerance < covariance_eps) && ( 2.0*tolerance > covariance_eps)) ? 1:0; - cerr << "Pass: " << pass_mean + cout << "Pass: " << pass_mean << " " << pass_var << " " << mean_eps << " " << variance_eps @@ -315,7 +315,7 @@ struct test_random_scalar { << " || " << tolerance << endl; } { - cerr << " -- Testing 1-D histogram" << endl; + cout << " -- Testing 1-D histogram" << endl; RandomProperties result; typedef test_histogram1d_functor functor_type; @@ -335,7 +335,7 @@ struct test_random_scalar { pass_hist1d_covar = ((-0.06 < covariance_eps) && ( 0.06 > covariance_eps)) ? 1:0; - cerr << "Density 1D: " << mean_eps + cout << "Density 1D: " << mean_eps << " " << variance_eps << " " << (result.covariance/HIST_DIM1D/HIST_DIM1D) << " || " << tolerance @@ -348,7 +348,7 @@ struct test_random_scalar { << endl; } { - cerr << " -- Testing 3-D histogram" << endl; + cout << " -- Testing 3-D histogram" << endl; RandomProperties result; typedef test_histogram3d_functor functor_type; @@ -368,7 +368,7 @@ struct test_random_scalar { pass_hist3d_covar = ((-tolerance < covariance_eps) && ( tolerance > covariance_eps)) ? 1:0; - cerr << "Density 3D: " << mean_eps + cout << "Density 3D: " << mean_eps << " " << variance_eps << " " << result.covariance/HIST_DIM1D/HIST_DIM1D << " || " << tolerance @@ -381,18 +381,18 @@ struct test_random_scalar { template void test_random(unsigned int num_draws) { - using std::cerr; + using std::cout; using std::endl; typename test_random_functor::type_1d density_1d("D1d"); typename test_random_functor::type_3d density_3d("D3d"); uint64_t ticks = std::chrono::high_resolution_clock::now().time_since_epoch().count(); - cerr << "Test Seed:" << ticks << endl; + cout << "Test Seed:" << ticks << endl; RandomGenerator pool(ticks); - cerr << "Test Scalar=int" << endl; + cout << "Test Scalar=int" << endl; test_random_scalar test_int(density_1d,density_3d,pool,num_draws); ASSERT_EQ( test_int.pass_mean,1); ASSERT_EQ( test_int.pass_var,1); @@ -406,7 +406,7 @@ void test_random(unsigned int num_draws) deep_copy(density_1d,0); deep_copy(density_3d,0); - cerr << "Test Scalar=unsigned int" << endl; + cout << "Test Scalar=unsigned int" << endl; test_random_scalar test_uint(density_1d,density_3d,pool,num_draws); ASSERT_EQ( test_uint.pass_mean,1); ASSERT_EQ( test_uint.pass_var,1); @@ -420,7 +420,7 @@ void test_random(unsigned int num_draws) deep_copy(density_1d,0); deep_copy(density_3d,0); - cerr << "Test Scalar=int64_t" << endl; + cout << "Test Scalar=int64_t" << endl; test_random_scalar test_int64(density_1d,density_3d,pool,num_draws); ASSERT_EQ( test_int64.pass_mean,1); ASSERT_EQ( test_int64.pass_var,1); @@ -434,7 +434,7 @@ void test_random(unsigned int num_draws) deep_copy(density_1d,0); deep_copy(density_3d,0); - cerr << "Test Scalar=uint64_t" << endl; + cout << "Test Scalar=uint64_t" << endl; test_random_scalar test_uint64(density_1d,density_3d,pool,num_draws); ASSERT_EQ( test_uint64.pass_mean,1); ASSERT_EQ( test_uint64.pass_var,1); @@ -448,7 +448,7 @@ void test_random(unsigned int num_draws) deep_copy(density_1d,0); deep_copy(density_3d,0); - cerr << "Test Scalar=float" << endl; + cout << "Test Scalar=float" << endl; test_random_scalar test_float(density_1d,density_3d,pool,num_draws); ASSERT_EQ( test_float.pass_mean,1); ASSERT_EQ( test_float.pass_var,1); @@ -462,7 +462,7 @@ void test_random(unsigned int num_draws) deep_copy(density_1d,0); deep_copy(density_3d,0); - cerr << "Test Scalar=double" << endl; + cout << "Test Scalar=double" << endl; test_random_scalar test_double(density_1d,density_3d,pool,num_draws); ASSERT_EQ( test_double.pass_mean,1); ASSERT_EQ( test_double.pass_var,1); diff --git a/lib/kokkos/algorithms/unit_tests/UnitTestMain.cpp b/lib/kokkos/algorithms/unit_tests/UnitTestMain.cpp index f952ab3db5..9e75b580bc 100644 --- a/lib/kokkos/algorithms/unit_tests/UnitTestMain.cpp +++ b/lib/kokkos/algorithms/unit_tests/UnitTestMain.cpp @@ -1,13 +1,13 @@ /* //@HEADER // ************************************************************************ -// +// // Kokkos v. 2.0 // Copyright (2014) Sandia Corporation -// +// // Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, // the U.S. Government retains certain rights in this software. -// +// // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are // met: @@ -36,7 +36,7 @@ // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. // // Questions? Contact H. Carter Edwards (hcedwar@sandia.gov) -// +// // ************************************************************************ //@HEADER */ diff --git a/lib/kokkos/benchmarks/bytes_and_flops/main.cpp b/lib/kokkos/benchmarks/bytes_and_flops/main.cpp index f545247212..8db5ce0eb5 100644 --- a/lib/kokkos/benchmarks/bytes_and_flops/main.cpp +++ b/lib/kokkos/benchmarks/bytes_and_flops/main.cpp @@ -44,12 +44,13 @@ #include #include #include +#include int main(int argc, char* argv[]) { Kokkos::initialize(); - - if(argc<10) { + + if(argc<10) { printf("Arguments: N K R D U F T S\n"); printf(" P: Precision (1==float, 2==double)\n"); printf(" N,K: dimensions of the 2D array to allocate\n"); @@ -68,7 +69,7 @@ int main(int argc, char* argv[]) { Kokkos::finalize(); return 0; } - + int P = atoi(argv[1]); int N = atoi(argv[2]); @@ -80,7 +81,7 @@ int main(int argc, char* argv[]) { int T = atoi(argv[8]); int S = atoi(argv[9]); - if(U>8) {printf("U must be 1-8\n"); return 0;} + if(U>8) {printf("U must be 1-8\n"); return 0;} if( (D!=1) && (D!=2) && (D!=4) && (D!=8) && (D!=16) && (D!=32)) {printf("D must be one of 1,2,4,8,16,32\n"); return 0;} if( (P!=1) && (P!=2) ) {printf("P must be one of 1,2\n"); return 0;} diff --git a/lib/kokkos/benchmarks/gather/main.cpp b/lib/kokkos/benchmarks/gather/main.cpp index 161c6f2091..88eb0493c1 100644 --- a/lib/kokkos/benchmarks/gather/main.cpp +++ b/lib/kokkos/benchmarks/gather/main.cpp @@ -44,11 +44,11 @@ #include #include #include +#include int main(int argc, char* argv[]) { Kokkos::initialize(argc,argv); - if(argc<8) { printf("Arguments: S N K D\n"); printf(" S: Scalar Type Size (1==float, 2==double, 4=complex)\n"); diff --git a/lib/kokkos/benchmarks/policy_performance/Makefile b/lib/kokkos/benchmarks/policy_performance/Makefile new file mode 100644 index 0000000000..13aef3209c --- /dev/null +++ b/lib/kokkos/benchmarks/policy_performance/Makefile @@ -0,0 +1,44 @@ +KOKKOS_PATH = ../.. +SRC = $(wildcard *.cpp) + +default: build + echo "Start Build" + +ifneq (,$(findstring Cuda,$(KOKKOS_DEVICES))) +CXX = ${KOKKOS_PATH}/bin/nvcc_wrapper +CXXFLAGS = -O3 -g +LINK = ${CXX} +LINKFLAGS = +EXE = policy_performance.cuda +KOKKOS_DEVICES = "Cuda,OpenMP" +KOKKOS_ARCH = "SNB,Kepler35" +KOKKOS_CUDA_OPTIONS+=enable_lambda +else +CXX = g++ +CXXFLAGS = -O3 -g -Wall -Werror +LINK = ${CXX} +LINKFLAGS = +EXE = policy_performance.host +KOKKOS_DEVICES = "OpenMP" +KOKKOS_ARCH = "SNB" +endif + +DEPFLAGS = -M + +OBJ = $(SRC:.cpp=.o) +LIB = + +include $(KOKKOS_PATH)/Makefile.kokkos + +build: $(EXE) + +$(EXE): $(OBJ) $(KOKKOS_LINK_DEPENDS) + $(LINK) $(KOKKOS_LDFLAGS) $(LINKFLAGS) $(EXTRA_PATH) $(OBJ) $(KOKKOS_LIBS) $(LIB) -o $(EXE) + +clean: kokkos-clean + rm -f *.o *.cuda *.host + +# Compilation rules + +%.o:%.cpp $(KOKKOS_CPP_DEPENDS) main.cpp policy_perf_test.hpp + $(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) $(EXTRA_INC) -c $< diff --git a/lib/kokkos/benchmarks/policy_performance/main.cpp b/lib/kokkos/benchmarks/policy_performance/main.cpp new file mode 100644 index 0000000000..b0ed9bb512 --- /dev/null +++ b/lib/kokkos/benchmarks/policy_performance/main.cpp @@ -0,0 +1,170 @@ +/* +//@HEADER +// ************************************************************************ +// +// Kokkos v. 2.0 +// Copyright (2014) Sandia Corporation +// +// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, +// the U.S. Government retains certain rights in this software. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// 1. Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// 2. Redistributions in binary form must reproduce the above copyright +// notice, this list of conditions and the following disclaimer in the +// documentation and/or other materials provided with the distribution. +// +// 3. Neither the name of the Corporation nor the names of the +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY +// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR +// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE +// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, +// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, +// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR +// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF +// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING +// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS +// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +// +// Questions? Contact H. Carter Edwards (hcedwar@sandia.gov) +// +// ************************************************************************ +//@HEADER +*/ + +#include +#include "policy_perf_test.hpp" + +int main(int argc, char* argv[] ) { + Kokkos::initialize(argc,argv); + + if(argc<10) { + printf(" Ten arguments are needed to run this program:\n"); + printf(" (1)team_range, (2)thread_range, (3)vector_range, (4)outer_repeat, (5)thread_repeat, (6)vector_repeat, (7)team_size, (8)vector_size, (9)schedule, (10)test_type\n"); + printf(" team_range: number of teams (league_size)\n"); + printf(" thread_range: range for nested TeamThreadRange parallel_*\n"); + printf(" vector_range: range for nested ThreadVectorRange parallel_*\n"); + printf(" outer_repeat: number of repeats for outer parallel_* call\n"); + printf(" thread_repeat: number of repeats for TeamThreadRange parallel_* call\n"); + printf(" vector_repeat: number of repeats for ThreadVectorRange parallel_* call\n"); + printf(" team_size: number of team members (team_size)\n"); + printf(" vector_size: desired vectorization (if possible)\n"); + printf(" schedule: 1 == Static 2 == Dynamic\n"); + printf(" test_type: 3-digit code XYZ for testing (nested) parallel_*\n"); + printf(" code key: XYZ X in {1,2,3,4,5}, Y in {0,1,2}, Z in {0,1,2}\n"); + printf(" TeamPolicy:\n"); + printf(" X: 0 = none (never used, makes no sense); 1 = parallel_for; 2 = parallel_reduce\n"); + printf(" Y: 0 = none; 1 = parallel_for; 2 = parallel_reduce\n"); + printf(" Z: 0 = none; 1 = parallel_for; 2 = parallel_reduce\n"); + printf(" RangePolicy:\n"); + printf(" X: 3 = parallel_for; 4 = parallel_reduce; 5 = parallel_scan\n"); + printf(" Y: 0 = none\n"); + printf(" Z: 0 = none\n"); + printf(" Example Input:\n"); + printf(" 100000 32 32 100 100 100 8 1 1 100\n"); + Kokkos::finalize(); + return 0; + } + + int team_range = atoi(argv[1]); + int thread_range = atoi(argv[2]); + int vector_range = atoi(argv[3]); + + int outer_repeat = atoi(argv[4]); + int thread_repeat = atoi(argv[5]); + int vector_repeat = atoi(argv[6]); + + int team_size = atoi(argv[7]); + int vector_size = atoi(argv[8]); + int schedule = atoi(argv[9]); + int test_type = atoi(argv[10]); + + int disable_verbose_output = 0; + if ( argc > 11 ) { + disable_verbose_output = atoi(argv[11]); + } + + if ( schedule != 1 && schedule != 2 ) { + printf("schedule: %d\n", schedule); + printf("Options for schedule are: 1 == Static 2 == Dynamic\n"); + Kokkos::finalize(); + return -1; + } + + if ( test_type != 100 && test_type != 110 && test_type != 111 && test_type != 112 && test_type != 120 && test_type != 121 && test_type != 122 + && test_type != 200 && test_type != 210 && test_type != 211 && test_type != 212 && test_type != 220 && test_type != 221 && test_type != 222 + && test_type != 300 && test_type != 400 && test_type != 500 + ) + { + printf("Incorrect test_type option\n"); + Kokkos::finalize(); + return -2; + } + + double result = 0.0; + + Kokkos::parallel_reduce( "parallel_reduce warmup", Kokkos::TeamPolicy<>(10,1), + KOKKOS_LAMBDA(const Kokkos::TeamPolicy<>::member_type team, double& lval) { + lval += 1; + }, result); + + typedef Kokkos::View view_type_1d; + typedef Kokkos::View view_type_2d; + typedef Kokkos::View view_type_3d; + + // Allocate view without initializing + // Call a 'warmup' test with 1 repeat - this will initialize the corresponding view appropriately for test and should obey first-touch etc + // Second call to test is the one we actually care about and time + view_type_1d v_1( Kokkos::ViewAllocateWithoutInitializing("v_1"), team_range*team_size); + view_type_2d v_2( Kokkos::ViewAllocateWithoutInitializing("v_2"), team_range*team_size, thread_range); + view_type_3d v_3( Kokkos::ViewAllocateWithoutInitializing("v_3"), team_range*team_size, thread_range, vector_range); + + double result_computed = 0.0; + double result_expect = 0.0; + double time = 0.0; + + if(schedule==1) { + if ( test_type != 500 ) { + // warmup - no repeat of loops + test_policy,int>(team_range,thread_range,vector_range,1,1,1,team_size,vector_size,test_type,v_1,v_2,v_3,result_computed,result_expect,time); + test_policy,int>(team_range,thread_range,vector_range,outer_repeat,thread_repeat,vector_repeat,team_size,vector_size,test_type,v_1,v_2,v_3,result_computed,result_expect,time); + } + else { + // parallel_scan: initialize 1d view for parallel_scan + test_policy,int>(team_range,thread_range,vector_range,1,1,1,team_size,vector_size,100,v_1,v_2,v_3,result_computed,result_expect,time); + test_policy,int>(team_range,thread_range,vector_range,outer_repeat,thread_repeat,vector_repeat,team_size,vector_size,test_type,v_1,v_2,v_3,result_computed,result_expect,time); + } + } + if(schedule==2) { + if ( test_type != 500 ) { + // warmup - no repeat of loops + test_policy,int>(team_range,thread_range,vector_range,1,1,1,team_size,vector_size,test_type,v_1,v_2,v_3,result_computed,result_expect,time); + test_policy,int>(team_range,thread_range,vector_range,outer_repeat,thread_repeat,vector_repeat,team_size,vector_size,test_type,v_1,v_2,v_3,result_computed,result_expect,time); + } + else { + // parallel_scan: initialize 1d view for parallel_scan + test_policy,int>(team_range,thread_range,vector_range,1,1,1,team_size,vector_size,100,v_1,v_2,v_3,result_computed,result_expect,time); + test_policy,int>(team_range,thread_range,vector_range,outer_repeat,thread_repeat,vector_repeat,team_size,vector_size,test_type,v_1,v_2,v_3,result_computed,result_expect,time); + } + } + + if ( disable_verbose_output == 0 ) { + printf("%7i %4i %2i %9i %4i %4i %4i %2i %1i %3i %e %e %lf\n",team_range,thread_range,vector_range,outer_repeat,thread_repeat,vector_repeat,team_size,vector_size,schedule,test_type,result_computed,result_expect,time); + } + else { + printf("%lf\n",time); + } + + Kokkos::finalize(); + + return 0; +} diff --git a/lib/kokkos/benchmarks/policy_performance/policy_perf_test.hpp b/lib/kokkos/benchmarks/policy_performance/policy_perf_test.hpp new file mode 100644 index 0000000000..8c79f3b88d --- /dev/null +++ b/lib/kokkos/benchmarks/policy_performance/policy_perf_test.hpp @@ -0,0 +1,354 @@ +/* +//@HEADER +// ************************************************************************ +// +// Kokkos v. 2.0 +// Copyright (2014) Sandia Corporation +// +// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation, +// the U.S. Government retains certain rights in this software. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// 1. Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// +// 2. Redistributions in binary form must reproduce the above copyright +// notice, this list of conditions and the following disclaimer in the +// documentation and/or other materials provided with the distribution. +// +// 3. Neither the name of the Corporation nor the names of the +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY +// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR +// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE +// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, +// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, +// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR +// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF +// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING +// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS +// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +// +// Questions? Contact H. Carter Edwards (hcedwar@sandia.gov) +// +// ************************************************************************ +//@HEADER +*/ + +#include + +template < class ViewType > +struct ParallelScanFunctor { + using value_type = double; + ViewType v; + + ParallelScanFunctor( const ViewType & v_ ) + : v(v_) + {} + + KOKKOS_INLINE_FUNCTION + void operator()( const int idx, value_type& val, const bool& final ) const + { + // inclusive scan + val += v(idx); + if ( final ) { + v(idx) = val; + } + } +}; + +template +void test_policy(int team_range, int thread_range, int vector_range, + int outer_repeat, int thread_repeat, int inner_repeat, + int team_size, int vector_size, int test_type, + ViewType1 &v1, ViewType2 &v2, ViewType3 &v3, + double &result, double &result_expect, double &time) { + + typedef Kokkos::TeamPolicy t_policy; + typedef typename t_policy::member_type t_team; + Kokkos::Timer timer; + + for(int orep = 0; orep(v1) +#if 0 + // This does not compile with pre Cuda 8.0 - see Github Issue #913 for explanation + KOKKOS_LAMBDA (const int idx, double& val, const bool& final) { + // inclusive scan + val += v1(idx); + if ( final ) { + v1(idx) = val; + } + } +#endif + ); + // result = v1( team_size*team_range - 1 ); // won't work with Cuda - need to copy result back to host to print + // result_expect = 0.5*(team_size*team_range)*(team_size*team_range-1); + } + + } // end outer for loop + + time = timer.seconds(); +} //end test_policy diff --git a/lib/kokkos/benchmarks/policy_performance/script_basic_testing.sh b/lib/kokkos/benchmarks/policy_performance/script_basic_testing.sh new file mode 100755 index 0000000000..e621fffbd4 --- /dev/null +++ b/lib/kokkos/benchmarks/policy_performance/script_basic_testing.sh @@ -0,0 +1,53 @@ +#!/bin/bash + +# Script to check policy_perf_test code works with each possible combo of options + +echo "Performance test results for parallel_reduce code computing sum of sequence [0,N) with various (nested) policies" + +EXECUTABLE=policy_performance + +TEAMRANGE=1000 +THREADRANGE=4 +VECTORRANGE=32 +TEAMSIZE=4 +VECTORSIZE=1 +OREPEAT=1 +MREPEAT=1 +IREPEAT=1 +SCHEDULE=1 + +SUFFIX=host +if [ -e $EXECUTABLE.$SUFFIX ] +then +SCHEDULE=1 +echo "Host tests Static schedule" +for CODE in {100,110,111,112,120,121,122,200,210,211,212,220,221,222,300,400,500} +do + OMP_PROC_BIND=true ./$EXECUTABLE.$SUFFIX $TEAMRANGE $THREADRANGE $VECTORRANGE $OREPEAT $MREPEAT $IREPEAT $TEAMSIZE $VECTORSIZE $SCHEDULE $CODE +done + +SCHEDULE=2 +echo "Host tests Dynamic schedule" +for CODE in {100,110,111,112,120,121,122,200,210,211,212,220,221,222,300,400,500} +do + OMP_PROC_BIND=true ./$EXECUTABLE.$SUFFIX $TEAMRANGE $THREADRANGE $VECTORRANGE $OREPEAT $MREPEAT $IREPEAT $TEAMSIZE $VECTORSIZE $SCHEDULE $CODE +done +fi + +SUFFIX=cuda +if [ -e $EXECUTABLE.$SUFFIX ] +then +SCHEDULE=1 +echo "Cuda tests Static schedule" +for CODE in {100,110,111,112,120,121,122,200,210,211,212,220,221,222,300,400,500} +do + ./$EXECUTABLE.$SUFFIX $TEAMRANGE $THREADRANGE $VECTORRANGE $OREPEAT $MREPEAT $IREPEAT $TEAMSIZE $VECTORSIZE $SCHEDULE $CODE +done + +SCHEDULE=2 +echo "Cuda tests Dynamic schedule" +for CODE in {100,110,111,112,120,121,122,200,210,211,212,220,221,222,300,400,500} +do + ./$EXECUTABLE.$SUFFIX $TEAMRANGE $THREADRANGE $VECTORRANGE $OREPEAT $MREPEAT $IREPEAT $TEAMSIZE $VECTORSIZE $SCHEDULE $CODE +done +fi diff --git a/lib/kokkos/benchmarks/policy_performance/script_sample_usage.sh b/lib/kokkos/benchmarks/policy_performance/script_sample_usage.sh new file mode 100755 index 0000000000..f4bfb87f8f --- /dev/null +++ b/lib/kokkos/benchmarks/policy_performance/script_sample_usage.sh @@ -0,0 +1,126 @@ +#!/bin/bash + +# Sample script for benchmarking policy performance + +# Suggested enviroment variables to export prior to executing script: +# KNL: +# OMP_NUM_THREADS=256 KMP_AFFINITY=compact +# Power: +# OMP_NUM_THREADS=64 OMP_PROC_BIND=true + +# Constants and Variables: +# Vary: TEAMSIZE, and THREADRANGE +# for TEAMSIZE in {1,2,4,5,8}; do +# for THREADRANGE in {32,41,1000}; do +# Fixed: TEAMRANGE, VECTORRANGE, VECTORSIZE +# System specific: Adjust REPEAT values to architecture tests are run on + +# Tests +# Static SCHEDULE = 1 +# Tier 1: parallel_for + RangePolicy 300 +# Tier 2: parallel_reduce, parallel_scan + RangePolicy 400 500 +# Tier 3: 'outer' parallel_for with TeamPolicy (nested parallelism) 1XY +# Tier 4: 'outer' parallel_reduce with TeamPolicy (nested parallelism) 2XY +# Dynamic SCHEDULE = 2 +# Tier 5: parallel_for + RangePolicy 300 +# Tier 6: parallel_reduce, parallel_scan + RangePolicy 400 500 +# Tier 7: 'outer' parallel_for with TeamPolicy (nested parallelism) 1XY +# Tier 8: 'outer' parallel_reduce with TeamPolicy (nested parallelism) 2XY + +# Results grouped by: +# 0) SCHEDULE 1) CODE (test) 2) TEAMRANGE 3) TEAMSIZE 4) THREADRANGE + +EXECUTABLE=policy_performance + +# Default defined values +TEAMRANGE=1000 +THREADRANGE=1 +VECTORRANGE=32 +TEAMSIZE=1 +VECTORSIZE=1 +OREPEAT=1 +MREPEAT=1 +IREPEAT=1 +SCHEDULE=1 + +# Host tests +SUFFIX=host +if [ -e $EXECUTABLE.$SUFFIX ]; then +echo "Host" + +for SCHEDULE in {1,2}; do + +# Tier 1 and 2, 5 and 6 +for CODE in {300,400,500}; do + for TEAMSIZE in {1,2,4,5,8}; do + OMP_PROC_BIND=true ./$EXECUTABLE.$SUFFIX $TEAMRANGE $THREADRANGE $VECTORRANGE $OREPEAT $MREPEAT $IREPEAT $TEAMSIZE $VECTORSIZE $SCHEDULE $CODE + done +done + +# Tier 3, 7 +for CODE in {100,110,111,112,120,121,122}; do + for TEAMSIZE in {1,2,4,5,8}; do + for THREADRANGE in {32,41,1000}; do + OMP_PROC_BIND=true ./$EXECUTABLE.$SUFFIX $TEAMRANGE $THREADRANGE $VECTORRANGE $OREPEAT $MREPEAT $IREPEAT $TEAMSIZE $VECTORSIZE $SCHEDULE $CODE + done + done +done + +# Tier 4, 8 +for CODE in {200,210,211,212,220,221,222}; do + for TEAMSIZE in {1,2,4,5,8}; do + for THREADRANGE in {32,41,1000}; do + OMP_PROC_BIND=true ./$EXECUTABLE.$SUFFIX $TEAMRANGE $THREADRANGE $VECTORRANGE $OREPEAT $MREPEAT $IREPEAT $TEAMSIZE $VECTORSIZE $SCHEDULE $CODE + done + done +done + +done # end SCHEDULE + +fi # end host + + +# Cuda tests +SUFFIX=cuda +# TEAMRANGE=10000, TEAMSIZE=8 too large +# TEAMRANGE=10000, TEAMSIZE=8, THREADRANGE=1000 too large +if [ -e $EXECUTABLE.$SUFFIX ]; then +echo "Cuda" + +for SCHEDULE in {1,2}; do + +# Reset defaults +TEAMRANGE=1000 +THREADRANGE=1 +VECTORRANGE=32 +TEAMSIZE=1 +VECTORSIZE=1 + +# Tier 1 and 2, 5 and 6 +for CODE in {300,400,500}; do + for TEAMSIZE in {1,2,4,5,8}; do + ./$EXECUTABLE.$SUFFIX $TEAMRANGE $THREADRANGE $VECTORRANGE $OREPEAT $MREPEAT $IREPEAT $TEAMSIZE $VECTORSIZE $SCHEDULE $CODE + done +done + +# Tier 3, 7 +for CODE in {100,110,111,112,120,121,122}; do + for TEAMSIZE in {1,2,4,5,8}; do + for THREADRANGE in {32,41,1000}; do + ./$EXECUTABLE.$SUFFIX $TEAMRANGE $THREADRANGE $VECTORRANGE $OREPEAT $MREPEAT $IREPEAT $TEAMSIZE $VECTORSIZE $SCHEDULE $CODE + done + done +done + +# Tier 4, 8 +for CODE in {200,210,211,212,220,221,222}; do + for TEAMSIZE in {1,2,4,5,8}; do + for THREADRANGE in {32,41,1000}; do + ./$EXECUTABLE.$SUFFIX $TEAMRANGE $THREADRANGE $VECTORRANGE $OREPEAT $MREPEAT $IREPEAT $TEAMSIZE $VECTORSIZE $SCHEDULE $CODE + done + done +done + +done # end SCHEDULE + +fi #end cuda diff --git a/lib/kokkos/bin/hpcbind b/lib/kokkos/bin/hpcbind new file mode 100755 index 0000000000..ca34648780 --- /dev/null +++ b/lib/kokkos/bin/hpcbind @@ -0,0 +1,454 @@ +#!/usr/bin/env bash + +################################################################################ +# Check if hwloc commands exist +################################################################################ +declare -i HPCBIND_HAS_HWLOC=1 +type hwloc-bind >/dev/null 2>&1 +HPCBIND_HAS_HWLOC=$((HPCBIND_HAS_HWLOC & ! $?)) + +type hwloc-distrib >/dev/null 2>&1 +HPCBIND_HAS_HWLOC=$((HPCBIND_HAS_HWLOC & ! $?)) + +type hwloc-ls >/dev/null 2>&1 +HPCBIND_HAS_HWLOC=$((HPCBIND_HAS_HWLOC & ! $?)) + +type hwloc-calc >/dev/null 2>&1 +HPCBIND_HAS_HWLOC=$((HPCBIND_HAS_HWLOC & ! $?)) + +type hwloc-ps >/dev/null 2>&1 +HPCBIND_HAS_HWLOC=$((HPCBIND_HAS_HWLOC & ! $?)) + +if [[ ${HPCBIND_HAS_HWLOC} -eq 0 ]]; then + echo "hwloc not found, no process binding will occur" +fi + +# Get parent cpuset +HPCBIND_HWLOC_PARENT_CPUSET="" +if [[ ${HPCBIND_HAS_HWLOC} -eq 1 ]]; then + MY_PID="$BASHPID" + HPCBIND_HWLOC_PARENT_CPUSET=$(hwloc-ps --cpuset | grep "${MY_PID}" | cut -f 2) +fi + +################################################################################ +# Check if nvidia-smi exist +################################################################################ +declare -i HPCBIND_HAS_NVIDIA=0 +type nvidia-smi >/dev/null 2>&1 +HPCBIND_HAS_NVIDIA=$((!$?)) + + +################################################################################ +# Get visible gpu +################################################################################ +declare -i NUM_GPUS=0 +HPCBIND_VISIBLE_GPUS="" +if [[ ${HPCBIND_HAS_NVIDIA} -eq 1 ]]; then + NUM_GPUS=$(nvidia-smi -L | wc -l); + GPU_LIST="$( seq 0 $((NUM_GPUS-1)) )" + HPCBIND_VISIBLE_GPUS=${CUDA_VISIBLE_DEVICES:-${GPU_LIST}} +fi + +declare -i HPCBIND_ENABLE_GPU_MAPPING=$((NUM_GPUS > 0)) + + +################################################################################ +# Get queue id +# supports sbatch, bsub, aprun +################################################################################ +HPCBIND_QUEUE_NAME="" +declare -i HPCBIND_QUEUE_INDEX=0 +declare -i HPCBIND_QUEUE_GPU_MAPPING=0 + +if [[ ! -z "${SLURM_LOCAL_ID}" ]]; then + HPCBIND_QUEUE_GPU_MAPPING=1 + HPCBIND_QUEUE_NAME="sbatch" + HPCBIND_QUEUE_INDEX=${SLURM_LOCAL_ID} +elif [[ ! -z "${LBS_JOBINDEX}" ]]; then + HPCBIND_QUEUE_GPU_MAPPING=1 + HPCBIND_QUEUE_NAME="bsub" + HPCBIND_QUEUE_INDEX=${LBS_JOBINDEX} +elif [[ ! -z "${ALPS_APP_PE}" ]]; then + HPCBIND_QUEUE_GPU_MAPPING=1 + HPCBIND_QUEUE_NAME="aprun" + HPCBIND_QUEUE_INDEX=${ALPS_APP_PE} +fi + + +################################################################################ +# Show help +################################################################################ +function show_help { + local cmd=$(basename "$0") + echo "Usage: ${cmd} -- command ..." + echo " Set the process mask, OMP environment variables and CUDA environment" + echo " variables to sane values if possible. Uses hwloc and nvidia-smi if" + echo " available. Will preserve the current process binding, so it is safe" + echo " to use with a queuing system or mpiexec." + echo "" + echo "Options:" + echo " --no-hwloc-bind Disable binding" + echo " --proc-bind= Set the initial process mask for the script" + echo " LOC can be any valid location argument for" + echo " hwloc-calc Default: all" + echo " --distribute=N Distribute the current cpuset into N partitions" + echo " --distribute-partition=I" + echo " Use the i'th partition (zero based)" + echo " --visible-gpus= Comma separated list of gpu ids" + echo " Default: CUDA_VISIBLE_DEVICES or all gpus in" + echo " sequential order" + echo " --gpu-ignore-queue Ignore queue job id when choosing visible GPU" + echo " --no-gpu-mapping Do not set CUDA_VISIBLE_DEVICES" + echo " --openmp=M.m Set env variables for the given OpenMP version" + echo " Default: 4.0" + echo " --openmp-percent=N Integer percentage of cpuset to use for OpenMP" + echo " threads Default: 100" + echo " --openmp-places= Op=threads|cores|sockets. Default: threads" + echo " --no-openmp-proc-bind Set OMP_PROC_BIND to false and unset OMP_PLACES" + echo " --force-openmp-num-threads=N" + echo " Override logic for selecting OMP_NUM_THREADS" + echo " --force-openmp-proc-bind=" + echo " Override logic for selecting OMP_PROC_BIND" + echo " --no-openmp-nested Set OMP_NESTED to false" + echo " --show-bindings Show the bindings" + echo " --lstopo Show bindings in lstopo without executing a command" + echo " -v|--verbose Show options and relevant environment variables" + echo " -h|--help Show this message" + echo "" + echo "Sample Usage:" + echo " Split the current process cpuset into 4 and use the 3rd partition" + echo " ${cmd} --distribute=4 --distribute-partition=2 -v -- command ..." + echo " Bing the process to all even cores" + echo " ${cmd} --proc-bind=core:even -v -- command ..." + echo " Bind to the first 64 cores and split the current process cpuset into 4" + echo " ${cmd} --proc-bind=core:0-63 --distribute=4 --distribute-partition=0 -- command ..." + echo " skip GPU 0 when mapping visible devices" + echo " ${cmd} --distribute=4 --distribute-partition=0 --visible-gpus=1,2 -v -- command ..." + echo " Display the current bindings" + echo " ${cmd} --proc-bind=numa:0 --show-bindings -- command" + echo " Display the current bindings using lstopo" + echo " ${cmd} --proc-bind=numa:0.core:odd --lstopo" + echo "" +} + + +################################################################################ +# Parse command line arguments +################################################################################ +# Show help if no command line arguments given +if [[ "$#" -eq 0 ]]; then + show_help + exit 0 +fi + +declare -a UNKNOWN_ARGS=() +declare -i HPCBIND_ENABLE_HWLOC_BIND=${HPCBIND_HAS_HWLOC} +declare -i HPCBIND_DISTRIBUTE=1 +declare -i HPCBIND_PARTITION=0 +HPCBIND_PROC_BIND="all" +HPCBIND_OPENMP_VERSION=4.0 +declare -i HPCBIND_OPENMP_PERCENT=100 +HPCBIND_OPENMP_PLACES=${OMP_PLACES:-threads} +declare -i HPCBIND_OPENMP_PROC_BIND=1 +declare -i HPCBIND_OPENMP_FORCE_NUM_THREADS=-1 +HPCBIND_OPENMP_FORCE_PROC_BIND="" +HPCBIND_OPENMP_NESTED=${OMP_NESTED:-true} +declare -i HPCBIND_VERBOSE=0 + +declare -i HPCBIND_SHOW_BINDINGS=0 +declare -i HPCBIND_LSTOPO=0 + +for i in $@; do + case $i in + # number of partitions to create + --no-hwloc-bind) + HPCBIND_ENABLE_HWLOC_BIND=0 + shift + ;; + --proc-bind=*) + HPCBIND_PROC_BIND="${i#*=}" + shift + ;; + --distribute=*) + HPCBIND_DISTRIBUTE="${i#*=}" + shift + ;; + # which partition to use + --distribute-partition=*) + HPCBIND_PARTITION="${i#*=}" + shift + ;; + --visible-gpus=*) + HPCBIND_VISIBLE_GPUS=$(echo "${i#*=}" | tr ',' ' ') + shift + ;; + --gpu-ignore-queue) + HPCBIND_QUEUE_GPU_MAPPING=0 + shift + ;; + --no-gpu-mapping) + HPCBIND_ENABLE_GPU_MAPPING=0 + shift + ;; + --openmp=*) + HPCBIND_OPENMP_VERSION="${i#*=}" + shift + ;; + --openmp-percent=*) + HPCBIND_OPENMP_PERCENT="${i#*=}" + shift + ;; + --openmp-places=*) + HPCBIND_OPENMP_PLACES="${i#*=}" + shift + ;; + --no-openmp-proc-bind) + HPCBIND_OPENMP_PROC_BIND=0 + shift + ;; + --force-openmp-proc-bind=*) + HPCBIND_OPENMP_FORCE_PROC_BIND="${i#*=}" + shift + ;; + --force-openmp-num-threads=*) + HPCBIND_OPENMP_FORCE_NUM_THREADS="${i#*=}" + shift + ;; + --no-openmp-nested) + HPCBIND_OPENMP_NESTED="false" + shift + ;; + --show-bindings) + HPCBIND_VERBOSE=1 + HPCBIND_SHOW_BINDINGS=1 + shift + ;; + --lstopo) + HPCBIND_VERBOSE=1 + HPCBIND_SHOW_BINDINGS=0 + HPCBIND_LSTOPO=1 + shift + ;; + -v|--verbose) + HPCBIND_VERBOSE=1 + shift + ;; + -h|--help) + show_help + exit 0 + ;; + # ignore remaining arguments + --) + shift + break + ;; + # unknown option + *) + UNKNOWN_ARGS+=("$i") + shift + ;; + esac +done + + +################################################################################ +# Check unknown arguments +################################################################################ +if [[ ${#UNKNOWN_ARGS[*]} > 0 ]]; then + echo "Uknown options: ${UNKNOWN_ARGS[*]}" + exit 1 +fi + + +################################################################################ +# Check that visible gpus are valid +################################################################################ +HPCBIND_VISIBLE_GPUS=(${HPCBIND_VISIBLE_GPUS}) +if [[ ${HPCBIND_ENABLE_GPU_MAPPING} -eq 1 ]]; then + for ((i=0; i < ${#HPCBIND_VISIBLE_GPUS[*]}; i++)); do + if [[ ${HPCBIND_VISIBLE_GPUS[$i]} -ge ${NUM_GPUS} || + ${HPCBIND_VISIBLE_GPUS[$i]} -lt 0 ]]; then + echo "Invaild GPU ID ${HPCBIND_VISIBLE_GPUS[$i]}, setting to 0" + HPCBIND_VISIBLE_GPUS[$i]=0; + fi + done + NUM_GPUS=${#HPCBIND_VISIBLE_GPUS[@]} +fi + + +################################################################################ +# Check OpenMP percent +################################################################################ +if [[ ${HPCBIND_OPENMP_PERCENT} -lt 1 ]]; then + echo "OpenMP percent < 1, setting to 1" + HPCBIND_OPENMP_PERCENT=1 +elif [[ ${HPCBIND_OPENMP_PERCENT} -gt 100 ]]; then + echo "OpenMP percent > 100, setting to 100" + HPCBIND_OPENMP_PERCENT=100 +fi + +################################################################################ +# Check distribute +################################################################################ +if [[ ${HPCBIND_DISTRIBUTE} -le 0 ]]; then + echo "Invalid input for distribute, changing distribute to 1" + HPCBIND_DISTRIBUTE=1 +fi + +if [[ ${HPCBIND_PARTITION} -ge ${HPCBIND_DISTRIBUTE} ]]; then + echo "Invalid input for distribute-partition, changing to 0" + HPCBIND_PARTITION=0 +fi + + +################################################################################ +# Find cpuset and num threads +################################################################################ +HPCBIND_HWLOC_CPUSET="" +declare -i HPCBIND_NUM_PUS=0 + +if [[ ${HPCBIND_ENABLE_HWLOC_BIND} -eq 1 ]]; then + if [[ "${HPCBIND_HWLOC_PARENT_CPUSET}" == "" ]]; then + BINDING=$(hwloc-calc ${HPCBIND_PROC_BIND}) + else + BINDING=$(hwloc-calc --restrict ${HPCBIND_HWLOC_PARENT_CPUSET} ${HPCBIND_PROC_BIND}) + fi + + CPUSETS=($(hwloc-distrib --restrict ${BINDING} --at core ${HPCBIND_DISTRIBUTE})) + HPCBIND_HWLOC_CPUSET=${CPUSETS[${HPCBIND_PARTITION}]} + HPCBIND_NUM_PUS=$(hwloc-ls --restrict ${HPCBIND_HWLOC_CPUSET} --only pu | wc -l) +else + HPCBIND_NUM_PUS=$(cat /proc/cpuinfo | grep -c processor) +fi + +declare -i HPCBIND_OPENMP_NUM_THREADS=$((HPCBIND_NUM_PUS * HPCBIND_OPENMP_PERCENT)) +HPCBIND_OPENMP_NUM_THREADS=$((HPCBIND_OPENMP_NUM_THREADS / 100)) + + +if [[ ${HPCBIND_OPENMP_NUM_THREADS} -lt 1 ]]; then + HPCBIND_OPENMP_NUM_THREADS=1 +elif [[ ${HPCBIND_OPENMP_NUM_THREADS} -gt ${HPCBIND_NUM_PUS} ]]; then + HPCBIND_OPENMP_NUM_THREADS=${HPCBIND_NUM_PUS} +fi + +if [[ ${HPCBIND_OPENMP_FORCE_NUM_THREADS} -gt 0 ]]; then + HPCBIND_OPENMP_NUM_THREADS=${HPCBIND_OPENMP_FORCE_NUM_THREADS} +fi + +################################################################################ +# Set OpenMP environment variables +################################################################################ + +# set OMP_NUM_THREADS +export OMP_NUM_THREADS=${HPCBIND_OPENMP_NUM_THREADS} + +# set OMP_PROC_BIND and OMP_PLACES +if [[ ${HPCBIND_OPENMP_PROC_BIND} -eq 1 ]]; then + if [[ "${HPCBIND_OPENMP_FORCE_PROC_BIND}" == "" ]]; then + #default proc bind logic + if [[ "${HPCBIND_OPENMP_VERSION}" == "4.0" || "${HPCBIND_OPENMP_VERSION}" > "4.0" ]]; then + export OMP_PLACES="${HPCBIND_OPENMP_PLACES}" + export OMP_PROC_BIND="spread" + else + export OMP_PROC_BIND="true" + unset OMP_PLACES + fi + else + #force proc bind + export OMP_PLACES="${HPCBIND_OPENMP_PLACES}" + export OMP_PROC_BIND="${HPCBIND_OPENMP_FORCE_PROC_BIND}" + fi +else + # no openmp proc bind + unset OMP_PLACES + unset OMP_PROC_BIND +fi + +# set OMP_NESTED +export OMP_NESTED=${HPCBIND_OPENMP_NESTED} + + +################################################################################ +# Set CUDA environment variables +################################################################################ + +if [[ ${HPCBIND_ENABLE_GPU_MAPPING} -eq 1 ]]; then + if [[ ${HPCBIND_QUEUE_GPU_MAPPING} -eq 0 ]]; then + declare -i GPU_ID=$((HPCBIND_PARTITION % NUM_GPUS)) + export CUDA_VISIBLE_DEVICES=${HPCBIND_VISIBLE_GPUS[${GPU_ID}]} + else + declare -i MY_TASK_ID=$((HPCBIND_QUEUE_INDEX * HPCBIND_DISTRIBUTE + HPCBIND_PARTITION)) + declare -i GPU_ID=$((MY_TASK_ID % NUM_GPUS)) + export CUDA_VISIBLE_DEVICES=${HPCBIND_VISIBLE_GPUS[${GPU_ID}]} + fi +fi + +################################################################################ +# Set hpcbind environment variables +################################################################################ +export HPCBIND_HAS_HWLOC=${HPCBIND_HAS_HWLOC} +export HPCBIND_HAS_NVIDIA=${HPCBIND_HAS_NVIDIA} +export HPCBIND_NUM_PUS=${HPCBIND_NUM_PUS} +export HPCBIND_HWLOC_CPUSET=${HPCBIND_HWLOC_CPUSET} +export HPCBIND_HWLOC_DISTRIBUTE=${HPCBIND_DISTRIBUTE} +export HPCBIND_HWLOC_DISTRIBUTE_PARTITION=${HPCBIND_PARTITION} +if [[ "${HPCBIND_HWLOC_PARENT_CPUSET}" == "" ]]; then + export HPCBIND_HWLOC_PARENT_CPUSET="all" +else + export HPCBIND_HWLOC_PARENT_CPUSET=${HPCBIND_HWLOC_PARENT_CPUSET} +fi +export HPCBIND_HWLOC_PROC_BIND=${HPCBIND_PROC_BIND} +export HPCBIND_NVIDIA_ENABLE_GPU_MAPPING=${HPCBIND_ENABLE_GPU_MAPPING} +export HPCBIND_NVIDIA_VISIBLE_GPUS=$(echo "${HPCBIND_VISIBLE_GPUS[*]}" | tr ' ' ',') +export HPCBIND_OPENMP_VERSION=${HPCBIND_OPENMP_VERSION} +if [[ "${HPCBIND_QUEUE_NAME}" != "" ]]; then + export HPCBIND_QUEUE_INDEX=${HPCBIND_QUEUE_INDEX} + export HPCBIND_QUEUE_NAME=${HPCBIND_QUEUE_NAME} + export HPCBIND_QUEUE_GPU_MAPPING=${HPCBIND_QUEUE_GPU_MAPPING} +fi + + +################################################################################ +# Print verbose +################################################################################ + +if [[ ${HPCBIND_VERBOSE} -eq 1 ]]; then + MY_ENV=$(env | sort) + echo "[HPCBIND]" + echo "${MY_ENV}" | grep -E "^HPCBIND_" + echo "[CUDA]" + echo "${MY_ENV}" | grep -E "^CUDA_" + echo "[OPENMP]" + echo "${MY_ENV}" | grep -E "^OMP_" +fi + +if [[ ${HPCBIND_HAS_HWLOC} -eq 1 && ${HPCBIND_SHOW_BINDINGS} -eq 1 ]]; then + echo "[BINDINGS]" + hwloc-ls --restrict ${HPCBIND_HWLOC_CPUSET} --only pu +elif [[ ${HPCBIND_SHOW_BINDINGS} -eq 1 ]]; then + echo "Unable to show bindings, hwloc not available." +fi + +################################################################################ +# Run command +################################################################################ + +if [[ ${HPCBIND_LSTOPO} -eq 0 ]]; then + if [[ ${HPCBIND_ENABLE_HWLOC_BIND} -eq 1 ]]; then + hwloc-bind ${HPCBIND_HWLOC_CPUSET} -- $@ + else + eval $@ + fi +else + if [[ ${HPCBIND_HAS_HWLOC} -eq 1 ]]; then + if [[ ${HPCBIND_ENABLE_HWLOC_BIND} -eq 1 && ! -z ${DISPLAY} ]]; then + echo "[BINDINGS]" + hwloc-ls --restrict ${HPCBIND_HWLOC_CPUSET} --only pu + hwloc-bind ${HPCBIND_HWLOC_CPUSET} -- lstopo --pid 0 + else + hwloc-ls --restrict ${HPCBIND_HWLOC_CPUSET} + fi + else + echo "Unable to show bindings, hwloc not available." + fi +fi diff --git a/lib/kokkos/bin/kokkos-bind b/lib/kokkos/bin/kokkos-bind new file mode 100755 index 0000000000..b6fe07a1bd --- /dev/null +++ b/lib/kokkos/bin/kokkos-bind @@ -0,0 +1,221 @@ +#!/usr/bin/env bash + +# check if hwloc commands exist +declare -i HAS_HWLOC=0 +type hwloc-bind >/dev/null 2>&1 +HAS_HWLOC="${HAS_HWLOC} + $?" + +type hwloc-distrib >/dev/null 2>&1 +HAS_HWLOC="${HAS_HWLOC} + $?" + +type hwloc-ls >/dev/null 2>&1 +HAS_HWLOC="${HAS_HWLOC} + $?" + +type hwloc-calc >/dev/null 2>&1 +HAS_HWLOC="${HAS_HWLOC} + $?" + +type hwloc-ps >/dev/null 2>&1 +HAS_HWLOC="${HAS_HWLOC} + $?" + + +#parse args +declare -a UNKNOWN_ARGS=() +declare -i DISTRIBUTE=1 +declare -i INDEX=0 +PROC_BIND="all" +CURRENT_CPUSET="" +OPENMP_VERSION=4.0 +OPENMP_PROC_BIND=True +OPENMP_NESTED=True +VERBOSE=False + +#get the current process cpuset +if [[ ${HAS_HWLOC} -eq 0 ]]; then + MY_PID="$BASHPID" + CURRENT_CPUSET=$(hwloc-ps --cpuset | grep "${MY_PID}" | cut -f 2) + echo "$CURRENT_CPUSET" +fi + +function show_help { + local cmd=$(basename "$0") + echo "Usage: ${cmd} -- command ..." + echo " Uses hwloc to divide the node into the given number of groups," + echo " set the appropriate OMP_NUM_THREADS and execute the command on the" + echo " selected group." + echo "" + echo " NOTE: This command assumes it has exclusive use of the node" + echo "" + echo "Options:" + echo " --proc-bind= Set the initial process mask for the script. " + echo " LOC can be any valid location argumnet for" + echo " hwloc-calc. Defaults to the entire machine" + echo " --distribute=N Distribute the current proc-bind into N groups" + echo " --index=I Use the i'th group (zero based)" + echo " --openmp=M.m Set env variables for the given OpenMP version" + echo " (default 4.0)" + echo " --no-openmp-proc-bind Set OMP_PROC_BIND to false and unset OMP_PLACES" + echo " --no-openmp-nested Set OMP_NESTED to false" + echo " -v|--verbose" + echo " -h|--help" + echo "" + echo "Sample Usage:" + echo " ${cmd} --distribute=4 --index=2 -v -- command ..." + echo "" +} + +if [[ "$#" -eq 0 ]]; then + show_help + exit 0 +fi + + +for i in $@; do + case $i in + # number of partitions to create + --proc-bind=*) + PROC_BIND="${i#*=}" + shift + ;; + --distribute=*) + DISTRIBUTE="${i#*=}" + shift + ;; + # which group to use + --index=*) + INDEX="${i#*=}" + shift + ;; + --openmp=*) + OPENMP_VERSION="${i#*=}" + shift + ;; + --no-openmp-proc-bind) + OPENMP_PROC_BIND=False + shift + ;; + --no-openmp-nested) + OPENMP_NESTED=False + shift + ;; + -v|--verbose) + VERBOSE=True + shift + ;; + -h|--help) + show_help + exit 0 + ;; + # ignore remaining arguments + --) + shift + break + ;; + # unknown option + *) + UNKNOWN_ARGS+=("$i") + shift + ;; + esac +done + +if [[ ${#UNKNOWN_ARGS[*]} > 0 ]]; then + echo "Uknown options: ${UNKNOWN_ARGS[*]}" + exit 1 +fi + +if [[ ${DISTRIBUTE} -le 0 ]]; then + echo "Invalid input for distribute, changing distribute to 1" + DISTRIBUTE=1 +fi + +if [[ ${INDEX} -ge ${DISTRIBUTE} ]]; then + echo "Invalid input for index, changing index to 0" + INDEX=0 +fi + +if [[ ${HAS_HWLOC} -ne 0 ]]; then + echo "hwloc not found, no process binding will occur" + DISTRIBUTE=1 + INDEX=0 +fi + +if [[ ${HAS_HWLOC} -eq 0 ]]; then + + if [[ "${CURRENT_CPUSET}" == "" ]]; then + BINDING=$(hwloc-calc ${PROC_BIND}) + else + BINDING=$(hwloc-calc --restrict ${CURRENT_CPUSET} ${PROC_BIND}) + fi + + CPUSETS=($(hwloc-distrib --restrict ${BINDING} --at core ${DISTRIBUTE})) + CPUSET=${CPUSETS[${INDEX}]} + NUM_THREADS=$(hwloc-ls --restrict ${CPUSET} --only pu | wc -l) + + if [[ "${VERBOSE}" == "True" ]]; then + echo "hwloc: true" + echo " proc_bind: ${PROC_BIND}" + echo " distribute: ${DISTRIBUTE}" + echo " index: ${INDEX}" + echo " parent_cpuset: ${CURRENT_CPUSET}" + echo " cpuset: ${CPUSET}" + echo "omp_num_threads: ${NUM_THREADS}" + echo "omp_proc_bind: ${OPENMP_PROC_BIND}" + echo "omp_nested: ${OPENMP_NESTED}" + echo "OpenMP: ${OPENMP_VERSION}" + fi + + # set OMP env + if [[ "${OPENMP_PROC_BIND}" == "True" ]]; then + if [[ "${OPENMP_VERSION}" == "4.0" || "${OPENMP_VERSION}" > "4.0" ]]; then + export OMP_PLACES="threads" + export OMP_PROC_BIND="spread" + else + export OMP_PROC_BIND="true" + unset OMP_PLACES + fi + else + unset OMP_PLACES + unset OMP_PROC_BIND + fi + if [[ "${OPENMP_NESTED}" == "True" ]]; then + export OMP_NESTED="true" + else + export OMP_NESTED="false" + fi + export OMP_NUM_THREADS="${NUM_THREADS}" + + hwloc-bind ${CPUSET} -- $@ +else + NUM_THREADS=$(cat /proc/cpuinfo | grep -c processor) + + if [[ "${VERBOSE}" == "True" ]]; then + echo "hwloc: false" + echo "omp_num_threads: ${NUM_THREADS}" + echo "omp_proc_bind: ${OPENMP_PROC_BIND}" + echo "omp_nested: ${OPENMP_NESTED}" + echo "OpenMP: ${OPENMP_VERSION}" + fi + + # set OMP env + if [[ "${OPENMP_PROC_BIND}" == "True" ]]; then + if [[ "${OPENMP_VERSION}" == "4.0" || "${OPENMP_VERSION}" > "4.0" ]]; then + export OMP_PLACES="threads" + export OMP_PROC_BIND="spread" + else + export OMP_PROC_BIND="true" + unset OMP_PLACES + fi + else + unset OMP_PLACES + unset OMP_PROC_BIND + fi + if [[ "${OPENMP_NESTED}" == "True" ]]; then + export OMP_NESTED="true" + else + export OMP_NESTED="false" + fi + export OMP_NUM_THREADS="${NUM_THREADS}" + + eval $@ +fi + diff --git a/lib/kokkos/bin/runtest b/lib/kokkos/bin/runtest new file mode 100755 index 0000000000..92411fe5ba --- /dev/null +++ b/lib/kokkos/bin/runtest @@ -0,0 +1,165 @@ +#!/usr/bin/env bash + +function get_path() { + cd "$(dirname "$0")" + cd .. + echo "$(pwd -P)" +} + +KOKKOS_PATH="$(get_path "$0")" + +function show_help() { + local cmd=$(basename "$0") + echo "Usage: ${cmd} " + echo " Build and run the tests" + echo "" + echo "Options:" + echo " -j=N|--make-j=N Build the tests in parallel" + echo " -c|--clean Clean build and regenerate make files" + echo " --clean-on-pass Clean build when runtest passes" + echo " --output-prefix=
  Prefix of log files  Default: runtest"
+  echo "  --build-only           Only build the tests"
+  echo "  -v|--verbose           Tee STDOUT and STDERR to screen and files"
+  echo "  -h|--help              Show this message"
+  echo ""
+  ${KOKKOS_PATH}/generate_makefile.bash --help
+  return 0
+}
+
+
+declare -a GENERATE_ARGS=()
+declare -i VERBOSE=0
+declare -i CLEAN=0
+declare -i CLEAN_ON_PASS=0
+declare -i BUILD_ONLY=0
+OUTPUT="runtest"
+
+declare -i MAKE_J=${HPCBIND_NUM_PUS:-1}
+
+for i in $@; do
+  case $i in
+    -j=*|--make-j=*)
+      MAKE_J=${i#*=}
+      shift
+      ;;
+    -c|--clean)
+      CLEAN=1
+      shift
+      ;;
+    --clean-on-pass)
+      CLEAN_ON_PASS=1
+      shift
+      ;;
+    --output-prefix=*)
+      OUTPUT=${i#*=}
+      shift
+      ;;
+    --build-only)
+      BUILD_ONLY=1
+      shift
+      ;;
+    -v|--verbose)
+      VERBOSE=1
+      shift
+      ;;
+    -h|--help)
+      show_help
+      exit 0
+      ;;
+    *)
+      GENERATE_ARGS+=("$i")
+      shift
+      ;;
+  esac
+done
+
+if [[ "$(pwd -P)" == ${KOKKOS_PATH} ]]; then
+  echo "Cannot call $0 from root repository path ${KOKKOS_PATH}"
+  exit 1
+fi
+
+# Some makefile dependencies are incorrect, so clean needs to force
+# a new call to generate_makefiles.bash
+if [[ ${CLEAN} -eq 1 ]]; then
+  START=${SECONDS}
+  echo "Cleaning"
+  /bin/rm -rf algorithms containers core example install Makefile >/dev/null 2>&1
+  END=${SECONDS}
+  echo "    $((END-START)) seconds"
+  if [[ ${VERBOSE} -eq 1 ]]; then
+    echo ""
+    echo ""
+  fi
+fi
+
+declare -i START=${SECONDS}
+echo "Generating Makefile"
+echo "    ${KOKKOS_PATH}/generate_makefile.bash --kokkos-path=${KOKKOS_PATH} ${GENERATE_ARGS[@]}"
+
+if [[ ${VERBOSE} -eq 0 ]]; then
+  "${KOKKOS_PATH}"/generate_makefile.bash --kokkos-path="${KOKKOS_PATH}" "${GENERATE_ARGS[@]}" > ${OUTPUT}.out 2> >(tee ${OUTPUT}.err >&2)
+else
+  "${KOKKOS_PATH}"/generate_makefile.bash --kokkos-path="${KOKKOS_PATH}" "${GENERATE_ARGS[@]}" > >(tee ${OUTPUT}.out) 2> >(tee ${OUTPUT}.err >&2)
+fi
+declare -i RESULT=$?
+declare -i END=${SECONDS}
+if [[ ${RESULT} -eq 0 ]]; then
+  echo "    PASS:  $((END-START)) seconds"
+  if [[ ${VERBOSE} -eq 1 ]]; then
+    echo ""
+    echo ""
+  fi
+else
+  cat ${OUTPUT}.out | grep "FAIL"
+  cat ${OUTPUT}.err | grep "FAIL"
+  echo "    FAIL:  $((END-START)) seconds"
+  exit 1
+fi
+
+START=${SECONDS}
+echo "Building"
+if [[ ${VERBOSE} -eq 0 ]]; then
+  make --keep-going -j ${MAKE_J} build-test >> ${OUTPUT}.out 2> >(tee -a ${OUTPUT}.err >&2)
+else
+  make --keep-going -j ${MAKE_J} build-test > >(tee -a ${OUTPUT}.out) 2> >(tee -a ${OUTPUT}.err >&2)
+fi
+RESULT=$?
+END=${SECONDS}
+if [[ ${RESULT} -eq 0 ]]; then
+  echo "    PASS:  $((END-START)) seconds"
+  if [[ ${VERBOSE} -eq 1 ]]; then
+    echo ""
+    echo ""
+  fi
+else
+  cat ${OUTPUT}.out | grep -E "[[:space:]]error:[[:space:]]"
+  cat ${OUTPUT}.err | grep -E "[[:space:]]error:[[:space:]]"
+  echo "    FAIL:  $((END-START)) seconds"
+  exit 1
+fi
+
+if [[ ${BUILD_ONLY} -eq 0 ]]; then
+  START=${SECONDS}
+  echo "Testing"
+  if [[ ${VERBOSE} -eq 0 ]]; then
+    make --keep-going test >> ${OUTPUT}.out 2> >(tee -a ${OUTPUT}.err >&2)
+  else
+    make --keep-going test > >(tee -a ${OUTPUT}.out) 2> >(tee -a ${OUTPUT}.err >&2)
+  fi
+  RESULT=$?
+  END=${SECONDS}
+  if [[ ${RESULT} -eq 0 ]]; then
+    echo "    PASS:  $((END-START)) seconds"
+    if [[ ${CLEAN_ON_PASS} -eq 1 ]]; then
+      make clean
+    fi
+  else
+    cat ${OUTPUT}.out | grep "FAIL"
+    cat ${OUTPUT}.err | grep "FAIL"
+    echo "    FAIL:  $((END-START)) seconds"
+    exit 1
+  fi
+fi
+
+exit ${RESULT}
+
diff --git a/lib/kokkos/cmake/kokkos.cmake b/lib/kokkos/cmake/kokkos.cmake
index 235b7eaba4..396822c7fa 100644
--- a/lib/kokkos/cmake/kokkos.cmake
+++ b/lib/kokkos/cmake/kokkos.cmake
@@ -999,8 +999,12 @@ SET (Kokkos_INCLUDE_DIRS
     ${Kokkos_SOURCE_DIR}/containers/src
     ${Kokkos_SOURCE_DIR}/algorithms/src
     ${Kokkos_BINARY_DIR}  # to find KokkosCore_config.h
+    ${KOKKOS_INCLUDE_DIRS}
 )
 
+# pass include dirs back to parent scope
+SET(Kokkos_INCLUDE_DIRS_RET ${Kokkos_INCLUDE_DIRS} PARENT_SCOPE)
+
 INCLUDE_DIRECTORIES(${Kokkos_INCLUDE_DIRS})
 
 IF(KOKKOS_SEPARATE_LIBS)
diff --git a/lib/kokkos/config/master_history.txt b/lib/kokkos/config/master_history.txt
index cc6f4c97d7..0447db4b2b 100644
--- a/lib/kokkos/config/master_history.txt
+++ b/lib/kokkos/config/master_history.txt
@@ -7,3 +7,4 @@ tag:  2.02.07    date: 12:16:2016    master: 4b4cc4ba    develop: 382c0966
 tag:  2.02.15    date: 02:10:2017    master: 8c64cd93    develop: 28dea8b6
 tag:  2.03.00    date: 04:25:2017    master: 120d9ce7    develop: 015ba641
 tag:  2.03.05    date: 05:27:2017    master: 36b92f43    develop: 79073186
+tag:  2.03.13    date: 07:27:2017    master: da314444    develop: 29ccb58a
diff --git a/lib/kokkos/config/query_cuda_arch.cpp b/lib/kokkos/config/query_cuda_arch.cpp
new file mode 100644
index 0000000000..383f04e34e
--- /dev/null
+++ b/lib/kokkos/config/query_cuda_arch.cpp
@@ -0,0 +1,24 @@
+#include 
+#include 
+int main()
+{
+	cudaDeviceProp prop;
+  const cudaError_t err_code = cudaGetDeviceProperties(&prop, 0);
+  if (cudaSuccess != err_code) {
+		fprintf(stderr,"cudaGetDeviceProperties failed: %s\n", cudaGetErrorString(err_code));
+		return -1;
+	}
+  switch (prop.major) {
+    case 3:
+      printf("Kepler"); break;
+    case 5:
+      printf("Maxwell"); break;
+    case 6:
+      printf("Pascal"); break;
+    default:
+      fprintf(stderr, "Unspported Device %d%d\n", (int)prop.major, (int)prop.minor);
+      return -1;
+  }
+  printf("%d%d\n", (int)prop.major, (int)prop.minor);
+  return 0;
+}
diff --git a/lib/kokkos/config/test_all_sandia b/lib/kokkos/config/test_all_sandia
index 8e1246bf8b..005cd20721 100755
--- a/lib/kokkos/config/test_all_sandia
+++ b/lib/kokkos/config/test_all_sandia
@@ -160,9 +160,14 @@ if [ "$MACHINE" = "sems" ]; then
     # Format: (compiler module-list build-list exe-name warning-flag)
     COMPILERS=("gcc/4.7.2 $BASE_MODULE_LIST $GCC_BUILD_LIST g++ $GCC_WARNING_FLAGS"
                "gcc/4.8.4 $BASE_MODULE_LIST $GCC_BUILD_LIST g++ $GCC_WARNING_FLAGS"
+               "gcc/4.9.3 $BASE_MODULE_LIST $GCC_BUILD_LIST g++ $GCC_WARNING_FLAGS"
+               "gcc/5.3.0 $BASE_MODULE_LIST $GCC_BUILD_LIST g++ $GCC_WARNING_FLAGS"
+               "gcc/6.1.0 $BASE_MODULE_LIST $GCC_BUILD_LIST g++ $GCC_WARNING_FLAGS"
                "intel/14.0.4 $BASE_MODULE_LIST $INTEL_BUILD_LIST icpc $INTEL_WARNING_FLAGS"
                "intel/15.0.2 $BASE_MODULE_LIST $INTEL_BUILD_LIST icpc $INTEL_WARNING_FLAGS"
                "intel/16.0.1 $BASE_MODULE_LIST $INTEL_BUILD_LIST icpc $INTEL_WARNING_FLAGS"
+               "intel/16.0.3 $BASE_MODULE_LIST $INTEL_BUILD_LIST icpc $INTEL_WARNING_FLAGS"
+               "intel/17.0.1 $BASE_MODULE_LIST $INTEL_BUILD_LIST icpc $INTEL_WARNING_FLAGS"
                "clang/3.6.1 $BASE_MODULE_LIST $CLANG_BUILD_LIST clang++ $CLANG_WARNING_FLAGS"
                "clang/3.7.1 $BASE_MODULE_LIST $CLANG_BUILD_LIST clang++ $CLANG_WARNING_FLAGS"
                "clang/3.8.1 $BASE_MODULE_LIST $CLANG_BUILD_LIST clang++ $CLANG_WARNING_FLAGS"
@@ -280,13 +285,13 @@ elif [ "$MACHINE" = "apollo" ]; then
                "gcc/5.1.0 $BASE_MODULE_LIST "Serial" g++ $GCC_WARNING_FLAGS"
                "intel/16.0.1 $BASE_MODULE_LIST "OpenMP" icpc $INTEL_WARNING_FLAGS"
                "clang/3.9.0 $BASE_MODULE_LIST "Pthread_Serial" clang++ $CLANG_WARNING_FLAGS"
-               "clang/head $CLANG_MODULE_LIST "Cuda_Pthread" clang++ $CUDA_WARNING_FLAGS"
+               "clang/4.0.0 $CLANG_MODULE_LIST "Cuda_Pthread" clang++ $CUDA_WARNING_FLAGS"
                "cuda/8.0.44 $CUDA_MODULE_LIST "Cuda_OpenMP" $KOKKOS_PATH/bin/nvcc_wrapper $CUDA_WARNING_FLAGS"
     )
   else
     # Format: (compiler module-list build-list exe-name warning-flag)
     COMPILERS=("cuda/8.0.44 $CUDA8_MODULE_LIST $BUILD_LIST_CUDA_NVCC $KOKKOS_PATH/bin/nvcc_wrapper $CUDA_WARNING_FLAGS"
-               "clang/head $CLANG_MODULE_LIST $BUILD_LIST_CUDA_CLANG clang++ $CUDA_WARNING_FLAGS"
+               "clang/4.0.0 $CLANG_MODULE_LIST $BUILD_LIST_CUDA_CLANG clang++ $CUDA_WARNING_FLAGS"
                "clang/3.9.0 $CLANG_MODULE_LIST $BUILD_LIST_CLANG clang++ $CLANG_WARNING_FLAGS"
                "gcc/4.7.2 $BASE_MODULE_LIST $GCC_BUILD_LIST g++ $GCC_WARNING_FLAGS"
                "gcc/4.8.4 $BASE_MODULE_LIST $GCC_BUILD_LIST g++ $GCC_WARNING_FLAGS"
@@ -584,7 +589,7 @@ single_build_and_test() {
   else
     run_cmd ${KOKKOS_PATH}/generate_makefile.bash --with-devices=$build $ARCH_FLAG --compiler=$(which $compiler_exe) --cxxflags=\"$cxxflags\" $extra_args &>> ${desc}.configure.log || { report_and_log_test_result 1 ${desc} configure && return 0; }
     local -i build_start_time=$(date +%s)
-    run_cmd make build-test >& ${desc}.build.log || { report_and_log_test_result 1 ${desc} build && return 0; }
+    run_cmd make -j 32 build-test >& ${desc}.build.log || { report_and_log_test_result 1 ${desc} build && return 0; }
     local -i build_end_time=$(date +%s)
     comment="build_time=$(($build_end_time-$build_start_time))"
 
diff --git a/lib/kokkos/config/trilinos-integration/shepard_jenkins_run_script_pthread_intel b/lib/kokkos/config/trilinos-integration/shepard_jenkins_run_script_pthread_intel
index 23968e8c0f..6527df2eb9 100755
--- a/lib/kokkos/config/trilinos-integration/shepard_jenkins_run_script_pthread_intel
+++ b/lib/kokkos/config/trilinos-integration/shepard_jenkins_run_script_pthread_intel
@@ -28,14 +28,14 @@ export JENKINS_DO_PTHREAD=ON
 export JENKINS_DO_SERIAL=OFF
 export JENKINS_DO_COMPLEX=OFF
 
-export ARCH_CXX_FLAG="-xCORE-AVX2 -mkl"
-export ARCH_C_FLAG="-xCORE-AVX2 -mkl"
+export JENKINS_ARCH_CXX_FLAG="-xCORE-AVX2 -mkl"
+export JENKINS_ARCH_C_FLAG="-xCORE-AVX2 -mkl"
 export BLAS_LIBRARIES="-mkl;${MKLROOT}/lib/intel64/libmkl_intel_lp64.a;${MKLROOT}/lib/intel64/libmkl_intel_thread.a;${MKLROOT}/lib/intel64/libmkl_core.a"
 export LAPACK_LIBRARIES=${BLAS_LIBRARIES}
 
 export JENKINS_DO_TESTS=ON
 export JENKINS_DO_EXAMPLES=ON
-export JENKINS_DO_SHARED=OFF
+export JENKINS_DO_SHARED=ON
 
 export QUEUE=haswell
 
diff --git a/lib/kokkos/config/trilinos-integration/shepard_jenkins_run_script_serial_intel b/lib/kokkos/config/trilinos-integration/shepard_jenkins_run_script_serial_intel
index 964de3a002..1a306bc2b2 100755
--- a/lib/kokkos/config/trilinos-integration/shepard_jenkins_run_script_serial_intel
+++ b/lib/kokkos/config/trilinos-integration/shepard_jenkins_run_script_serial_intel
@@ -28,14 +28,14 @@ export JENKINS_DO_PTHREAD=OFF
 export JENKINS_DO_SERIAL=ON
 export JENKINS_DO_COMPLEX=ON
 
-export ARCH_CXX_FLAG="-xCORE-AVX2 -mkl"
-export ARCH_C_FLAG="-xCORE-AVX2 -mkl"
+export JENKINS_ARCH_CXX_FLAG="-xCORE-AVX2 -mkl"
+export JENKINS_ARCH_C_FLAG="-xCORE-AVX2 -mkl"
 export BLAS_LIBRARIES="-mkl;${MKLROOT}/lib/intel64/libmkl_intel_lp64.a;${MKLROOT}/lib/intel64/libmkl_intel_thread.a;${MKLROOT}/lib/intel64/libmkl_core.a"
 export LAPACK_LIBRARIES=${BLAS_LIBRARIES}
 
 export JENKINS_DO_TESTS=ON
 export JENKINS_DO_EXAMPLES=ON
-export JENKINS_DO_SHARED=OFF
+export JENKINS_DO_SHARED=ON
 
 export QUEUE=haswell
 
diff --git a/lib/kokkos/containers/performance_tests/Makefile b/lib/kokkos/containers/performance_tests/Makefile
index edaaf1ee51..ec69363a17 100644
--- a/lib/kokkos/containers/performance_tests/Makefile
+++ b/lib/kokkos/containers/performance_tests/Makefile
@@ -60,7 +60,6 @@ test-threads: KokkosContainers_PerformanceTest_Threads
 test-openmp: KokkosContainers_PerformanceTest_OpenMP
 	./KokkosContainers_PerformanceTest_OpenMP
 
-
 build_all: $(TARGETS)
 
 test: $(TEST_TARGETS)
diff --git a/lib/kokkos/containers/performance_tests/TestMain.cpp b/lib/kokkos/containers/performance_tests/TestMain.cpp
index f952ab3db5..1224af7cdb 100644
--- a/lib/kokkos/containers/performance_tests/TestMain.cpp
+++ b/lib/kokkos/containers/performance_tests/TestMain.cpp
@@ -1,13 +1,13 @@
 /*
 //@HEADER
 // ************************************************************************
-// 
+//
 //                        Kokkos v. 2.0
 //              Copyright (2014) Sandia Corporation
-// 
+//
 // Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
 // the U.S. Government retains certain rights in this software.
-// 
+//
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
@@ -36,12 +36,15 @@
 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 //
 // Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
-// 
+//
 // ************************************************************************
 //@HEADER
 */
 
 #include 
+#include 
+
+#include 
 
 int main(int argc, char *argv[]) {
   ::testing::InitGoogleTest(&argc,argv);
diff --git a/lib/kokkos/containers/performance_tests/TestOpenMP.cpp b/lib/kokkos/containers/performance_tests/TestOpenMP.cpp
index b674ec4a74..6631184624 100644
--- a/lib/kokkos/containers/performance_tests/TestOpenMP.cpp
+++ b/lib/kokkos/containers/performance_tests/TestOpenMP.cpp
@@ -69,30 +69,13 @@ protected:
   {
     std::cout << std::setprecision(5) << std::scientific;
 
-    unsigned num_threads = 4;
-
-    if (Kokkos::hwloc::available()) {
-      num_threads = Kokkos::hwloc::get_available_numa_count()
-                    * Kokkos::hwloc::get_available_cores_per_numa()
-                    * Kokkos::hwloc::get_available_threads_per_core()
-                    ;
-
-    }
-
-    std::cout << "OpenMP: " << num_threads << std::endl;
-
-    Kokkos::OpenMP::initialize( num_threads );
-
-    std::cout << "available threads: " << omp_get_max_threads() << std::endl;
+    Kokkos::OpenMP::initialize();
+    Kokkos::OpenMP::print_configuration( std::cout );
   }
 
   static void TearDownTestCase()
   {
     Kokkos::OpenMP::finalize();
-
-    omp_set_num_threads(1);
-
-    ASSERT_EQ( 1 , omp_get_max_threads() );
   }
 };
 
diff --git a/lib/kokkos/containers/src/Kokkos_DualView.hpp b/lib/kokkos/containers/src/Kokkos_DualView.hpp
index 937eab0d88..35cc8ec753 100644
--- a/lib/kokkos/containers/src/Kokkos_DualView.hpp
+++ b/lib/kokkos/containers/src/Kokkos_DualView.hpp
@@ -564,7 +564,7 @@ namespace Impl {
 template< class D, class A1, class A2, class A3, class ... Args >
 struct DualViewSubview {
 
-  typedef typename Kokkos::Experimental::Impl::ViewMapping
+  typedef typename Kokkos::Impl::ViewMapping
     < void
     , Kokkos::ViewTraits< D, A1, A2, A3 >
     , Args ...
diff --git a/lib/kokkos/containers/src/Kokkos_DynRankView.hpp b/lib/kokkos/containers/src/Kokkos_DynRankView.hpp
index 8e464506f9..d22d6b865d 100644
--- a/lib/kokkos/containers/src/Kokkos_DynRankView.hpp
+++ b/lib/kokkos/containers/src/Kokkos_DynRankView.hpp
@@ -46,19 +46,6 @@
 ///
 /// This header file declares and defines Kokkos::Experimental::DynRankView and its
 /// related nonmember functions.
-/*
- *   Changes from View
- *   1. The rank of the DynRankView is returned by the method rank()
- *   2. Max rank of a DynRankView is 7
- *   3. subview name is subdynrankview
- *   4. Every subdynrankview is returned with LayoutStride
- *
- *   NEW: Redesigned DynRankView
- *   5. subview function name now available
- *   6. Copy and Copy-Assign View to DynRankView
- *   7. deep_copy between Views and DynRankViews
- *   8. rank( view ); returns the rank of View or DynRankView
- */
 
 #ifndef KOKKOS_DYNRANKVIEW_HPP
 #define KOKKOS_DYNRANKVIEW_HPP
@@ -117,6 +104,14 @@ struct DynRankDimTraits {
                       , layout.dimension[7] );
   }
 
+  // Extra overload to match that for specialize types v2
+  template 
+  KOKKOS_INLINE_FUNCTION
+  static size_t computeRank( const Kokkos::Impl::ViewCtorProp& prop, const Layout& layout )
+  {
+    return computeRank(layout);
+  }
+
   // Create the layout for the rank-7 view.
   // Non-strided Layout
   template 
@@ -158,8 +153,17 @@ struct DynRankDimTraits {
                  );
   }
 
+  // Extra overload to match that for specialize types
+  template 
+  KOKKOS_INLINE_FUNCTION
+  static typename std::enable_if< (std::is_same::value || std::is_same::value || std::is_same::value) , typename Traits::array_layout >::type createLayout( const ViewCtorProp& prop, const typename Traits::array_layout& layout )
+  {
+    return createLayout( layout );
+  }
+
   // Create a view from the given dimension arguments.
   // This is only necessary because the shmem constructor doesn't take a layout.
+  //   NDE shmem View's are not compatible with the added view_alloc value_type / fad_dim deduction functionality
   template 
   static ViewType createView( const ViewArg& arg
                             , const size_t N0
@@ -186,7 +190,8 @@ struct DynRankDimTraits {
   // Non-strided Layout
   template 
   KOKKOS_INLINE_FUNCTION
-  static typename std::enable_if< (std::is_same::value || std::is_same::value) && std::is_integral::value , Layout >::type reconstructLayout( const Layout& layout , iType dynrank )
+  static typename std::enable_if< (std::is_same::value || std::is_same::value) && std::is_integral::value , Layout >::type
+  reconstructLayout( const Layout& layout , iType dynrank )
   {
     return Layout( dynrank > 0 ? layout.dimension[0] : ~size_t(0)
                  , dynrank > 1 ? layout.dimension[1] : ~size_t(0)
@@ -202,7 +207,8 @@ struct DynRankDimTraits {
   // LayoutStride
   template 
   KOKKOS_INLINE_FUNCTION
-  static typename std::enable_if< (std::is_same::value) && std::is_integral::value , Layout >::type reconstructLayout( const Layout& layout , iType dynrank )
+  static typename std::enable_if< (std::is_same::value) && std::is_integral::value , Layout >::type
+  reconstructLayout( const Layout& layout , iType dynrank )
   {
     return Layout( dynrank > 0 ? layout.dimension[0] : ~size_t(0)
                  , dynrank > 0 ? layout.stride[0] : (0)
@@ -311,6 +317,11 @@ void dyn_rank_view_verify_operator_bounds
 /** \brief  Assign compatible default mappings */
 struct ViewToDynRankViewTag {};
 
+} // namespace Impl
+} // namespace Experimental
+
+namespace Impl {
+
 template< class DstTraits , class SrcTraits >
 class ViewMapping< DstTraits , SrcTraits ,
   typename std::enable_if<(
@@ -337,7 +348,7 @@ class ViewMapping< DstTraits , SrcTraits ,
         )
       )
     )
-  ) , ViewToDynRankViewTag >::type >
+  ) , Kokkos::Experimental::Impl::ViewToDynRankViewTag >::type >
 {
 private:
 
@@ -376,7 +387,7 @@ public:
 
       typedef typename DstType::offset_type  dst_offset_type ;
       dst.m_map.m_offset = dst_offset_type(std::integral_constant() , src.layout() ); //Check this for integer input1 for padding, etc
-      dst.m_map.m_handle = Kokkos::Experimental::Impl::ViewDataHandle< DstTraits >::assign( src.m_map.m_handle , src.m_track );
+      dst.m_map.m_handle = Kokkos::Impl::ViewDataHandle< DstTraits >::assign( src.m_map.m_handle , src.m_track );
       dst.m_track.assign( src.m_track , DstTraits::is_managed );
       dst.m_rank = src.Rank ;
     }
@@ -384,22 +395,20 @@ public:
 
 } //end Impl
 
+namespace Experimental {
+
 /* \class DynRankView
  * \brief Container that creates a Kokkos view with rank determined at runtime.
- *   Essentially this is a rank 7 view that wraps the access operators
- *   to yield the functionality of a view
+ *   Essentially this is a rank 7 view
  *
  *   Changes from View
  *   1. The rank of the DynRankView is returned by the method rank()
  *   2. Max rank of a DynRankView is 7
- *   3. subview name is subdynrankview
- *   4. Every subdynrankview is returned with LayoutStride
- *
- *   NEW: Redesigned DynRankView
- *   5. subview function name now available
- *   6. Copy and Copy-Assign View to DynRankView
- *   7. deep_copy between Views and DynRankViews
- *   8. rank( view ); returns the rank of View or DynRankView
+ *   3. subview called with 'subview(...)' or 'subdynrankview(...)' (backward compatibility) 
+ *   4. Every subview is returned with LayoutStride
+ *   5. Copy and Copy-Assign View to DynRankView
+ *   6. deep_copy between Views and DynRankViews
+ *   7. rank( view ); returns the rank of View or DynRankView
  *
  */
 
@@ -427,7 +436,7 @@ public:
 
 
 private:
-  typedef Kokkos::Experimental::Impl::ViewMapping< traits , void > map_type ;
+  typedef Kokkos::Impl::ViewMapping< traits , void > map_type ;
   typedef Kokkos::Experimental::Impl::SharedAllocationTracker      track_type ;
 
   track_type  m_track ;
@@ -556,7 +565,7 @@ public:
   // Allow specializations to query their specialized map
 
   KOKKOS_INLINE_FUNCTION
-  const Kokkos::Experimental::Impl::ViewMapping< traits , void > &
+  const Kokkos::Impl::ViewMapping< traits , void > &
   implementation_map() const { return m_map ; }
 
   //----------------------------------------
@@ -803,7 +812,7 @@ public:
     , m_rank(rhs.m_rank)
     {
       typedef typename DynRankView ::traits SrcTraits ;
-      typedef Kokkos::Experimental::Impl::ViewMapping< traits , SrcTraits , void > Mapping ;
+      typedef Kokkos::Impl::ViewMapping< traits , SrcTraits , void > Mapping ;
       static_assert( Mapping::is_assignable , "Incompatible DynRankView copy construction" );
       Mapping::assign( m_map , rhs.m_map , rhs.m_track );
     }
@@ -813,7 +822,7 @@ public:
   DynRankView & operator = (const DynRankView & rhs )
     {
       typedef typename DynRankView ::traits SrcTraits ;
-      typedef Kokkos::Experimental::Impl::ViewMapping< traits , SrcTraits , void > Mapping ;
+      typedef Kokkos::Impl::ViewMapping< traits , SrcTraits , void > Mapping ;
       static_assert( Mapping::is_assignable , "Incompatible DynRankView copy construction" );
       Mapping::assign( m_map , rhs.m_map , rhs.m_track );
       m_track.assign( rhs.m_track , traits::is_managed );
@@ -831,7 +840,7 @@ public:
     , m_rank( rhs.Rank )
     {
       typedef typename View::traits  SrcTraits ;
-      typedef Kokkos::Experimental::Impl::ViewMapping< traits , SrcTraits , Kokkos::Experimental::Impl::ViewToDynRankViewTag >  Mapping ;
+      typedef Kokkos::Impl::ViewMapping< traits , SrcTraits , Kokkos::Experimental::Impl::ViewToDynRankViewTag >  Mapping ;
       static_assert( Mapping::is_assignable , "Incompatible DynRankView copy construction" );
       Mapping::assign( *this , rhs );
     }
@@ -841,7 +850,7 @@ public:
   DynRankView & operator = ( const View & rhs )
     {
       typedef typename View::traits  SrcTraits ;
-      typedef Kokkos::Experimental::Impl::ViewMapping< traits , SrcTraits , Kokkos::Experimental::Impl::ViewToDynRankViewTag >  Mapping ;
+      typedef Kokkos::Impl::ViewMapping< traits , SrcTraits , Kokkos::Experimental::Impl::ViewToDynRankViewTag >  Mapping ;
       static_assert( Mapping::is_assignable , "Incompatible View to DynRankView copy assignment" );
       Mapping::assign( *this , rhs );
       return *this ;
@@ -870,7 +879,7 @@ public:
       )
       : m_track()
       , m_map()
-      , m_rank( Impl::DynRankDimTraits::computeRank(arg_layout) )
+      , m_rank( Impl::DynRankDimTraits::template computeRank< typename traits::array_layout, P...>(arg_prop, arg_layout) )
     {
       // Append layout and spaces if not input
       typedef Impl::ViewCtorProp< P ... > alloc_prop_input ;
@@ -923,7 +932,7 @@ public:
 //------------------------------------------------------------
 
       Kokkos::Experimental::Impl::SharedAllocationRecord<> *
-        record = m_map.allocate_shared( prop , Impl::DynRankDimTraits::createLayout(arg_layout) );
+        record = m_map.allocate_shared( prop , Impl::DynRankDimTraits::template createLayout(arg_prop, arg_layout) );
 
 //------------------------------------------------------------
 #if defined( KOKKOS_ENABLE_CUDA )
@@ -947,8 +956,8 @@ public:
                                >::type const & arg_layout
       )
       : m_track() // No memory tracking
-      , m_map( arg_prop , Impl::DynRankDimTraits::createLayout(arg_layout) )
-      , m_rank( Impl::DynRankDimTraits::computeRank(arg_layout) )
+      , m_map( arg_prop , Impl::DynRankDimTraits::template createLayout(arg_prop, arg_layout) )
+      , m_rank( Impl::DynRankDimTraits::template computeRank< typename traits::array_layout, P...>(arg_prop, arg_layout) )
     {
       static_assert(
         std::is_same< pointer_type
@@ -1034,6 +1043,7 @@ public:
     {}
 
   // For backward compatibility
+  // NDE This ctor does not take ViewCtorProp argument - should not use alternative createLayout call
   explicit inline
   DynRankView( const ViewAllocateWithoutInitializing & arg_prop
       , const typename traits::array_layout & arg_layout
@@ -1179,6 +1189,11 @@ namespace Impl {
 
 struct DynRankSubviewTag {};
 
+} // namespace Impl
+} // namespace Experimental
+
+namespace Impl {
+
 template< class SrcTraits , class ... Args >
 struct ViewMapping
   < typename std::enable_if<(
@@ -1192,7 +1207,7 @@ struct ViewMapping
         std::is_same< typename SrcTraits::array_layout
                     , Kokkos::LayoutStride >::value
       )
-    ), DynRankSubviewTag >::type
+    ), Kokkos::Experimental::Impl::DynRankSubviewTag >::type
   , SrcTraits
   , Args ... >
 {
@@ -1264,7 +1279,7 @@ public:
   };
 
 
-  typedef DynRankView< value_type , array_layout , typename SrcTraits::device_type , typename SrcTraits::memory_traits >  ret_type;
+  typedef Kokkos::Experimental::DynRankView< value_type , array_layout , typename SrcTraits::device_type , typename SrcTraits::memory_traits >  ret_type;
 
   template < typename T , class ... P >
   KOKKOS_INLINE_FUNCTION
@@ -1336,9 +1351,10 @@ public:
 
 } // end Impl
 
+namespace Experimental {
 
 template< class V , class ... Args >
-using Subdynrankview = typename Kokkos::Experimental::Impl::ViewMapping< Kokkos::Experimental::Impl::DynRankSubviewTag , V , Args... >::ret_type ;
+using Subdynrankview = typename Kokkos::Impl::ViewMapping< Kokkos::Experimental::Impl::DynRankSubviewTag , V , Args... >::ret_type ;
 
 template< class D , class ... P , class ...Args >
 KOKKOS_INLINE_FUNCTION
@@ -1348,7 +1364,7 @@ subdynrankview( const Kokkos::Experimental::DynRankView< D , P... > &src , Args.
     if ( src.rank() > sizeof...(Args) ) //allow sizeof...(Args) >= src.rank(), ignore the remaining args
       { Kokkos::abort("subdynrankview: num of args must be >= rank of the source DynRankView"); }
 
-    typedef Kokkos::Experimental::Impl::ViewMapping< Kokkos::Experimental::Impl::DynRankSubviewTag , Kokkos::ViewTraits< D*******, P... > , Args... > metafcn ;
+    typedef Kokkos::Impl::ViewMapping< Kokkos::Experimental::Impl::DynRankSubviewTag , Kokkos::ViewTraits< D*******, P... > , Args... > metafcn ;
 
     return metafcn::subview( src.rank() , src , args... );
   }
diff --git a/lib/kokkos/containers/src/Kokkos_DynamicView.hpp b/lib/kokkos/containers/src/Kokkos_DynamicView.hpp
index da96db2d6b..e9059d64c4 100644
--- a/lib/kokkos/containers/src/Kokkos_DynamicView.hpp
+++ b/lib/kokkos/containers/src/Kokkos_DynamicView.hpp
@@ -1,13 +1,13 @@
 /*
 //@HEADER
 // ************************************************************************
-// 
+//
 //                        Kokkos v. 2.0
 //              Copyright (2014) Sandia Corporation
-// 
+//
 // Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
 // the U.S. Government retains certain rights in this software.
-// 
+//
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
@@ -36,7 +36,7 @@
 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 //
 // Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
-// 
+//
 // ************************************************************************
 //@HEADER
 */
@@ -57,7 +57,7 @@ namespace Experimental {
  */
 template< typename DataType , typename ... P >
 class DynamicView : public Kokkos::ViewTraits< DataType , P ... >
-{ 
+{
 public:
 
   typedef Kokkos::ViewTraits< DataType , P ... >  traits ;
@@ -68,7 +68,7 @@ private:
 
   typedef Kokkos::Experimental::Impl::SharedAllocationTracker   track_type ;
 
-  static_assert( traits::rank == 1 && traits::rank_dynamic == 1 
+  static_assert( traits::rank == 1 && traits::rank_dynamic == 1
                , "DynamicView must be rank-one" );
 
   static_assert( std::is_trivial< typename traits::value_type >::value &&
@@ -216,14 +216,14 @@ public:
         // Verify that allocation of the requested chunk in in progress.
 
         // The allocated chunk counter is m_chunks[ m_chunk_max ]
-        const uintptr_t n = 
+        const uintptr_t n =
           *reinterpret_cast( m_chunks + m_chunk_max );
 
         if ( n <= ic ) {
           Kokkos::abort("Kokkos::DynamicView array bounds error");
         }
 
-        // Allocation of this chunk is in progress 
+        // Allocation of this chunk is in progress
         // so wait for allocation to complete.
         while ( 0 == *ch );
       }
@@ -267,7 +267,7 @@ public:
         const uintptr_t jc_try = jc ;
 
         // Jump iteration to the chunk counter.
-        
+
         jc = atomic_compare_exchange( pc , jc_try , jc_try + 1 );
 
         if ( jc_try == jc ) {
@@ -316,7 +316,7 @@ public:
       }
       else {
         while ( NC + 1 <= *pc ) {
-          --*pc ;        
+          --*pc ;
           m_pool.deallocate( m_chunks[*pc]
                            , sizeof(value_type) << m_chunk_shift );
           m_chunks[*pc] = 0 ;
@@ -331,7 +331,7 @@ public:
     typename traits::value_type ** m_chunks ;
     uintptr_t                    * m_pc ;
     uintptr_t                      m_nc ;
-    unsigned                       m_chunk_shift ;  
+    unsigned                       m_chunk_shift ;
 
     KOKKOS_INLINE_FUNCTION
     void operator()( int ) const
@@ -348,7 +348,7 @@ public:
         }
         else {
           while ( m_nc + 1 <= *m_pc ) {
-            --*m_pc ;        
+            --*m_pc ;
             m_pool.deallocate( m_chunks[*m_pc]
                              , sizeof(value_type) << m_chunk_shift );
             m_chunks[*m_pc] = 0 ;
@@ -482,7 +482,7 @@ public:
   };
 
 
-  /**\brief  Allocation constructor 
+  /**\brief  Allocation constructor
    *
    *  Memory is allocated in chunks from the memory pool.
    *  The chunk size conforms to the memory pool's chunk size.
@@ -557,7 +557,7 @@ void deep_copy( const View & dst
 
   if ( DstExecCanAccessSrc ) {
     // Copying data between views in accessible memory spaces and either non-contiguous or incompatible shape.
-    Kokkos::Experimental::Impl::ViewRemap< dst_type , src_type >( dst , src );
+    Kokkos::Impl::ViewRemap< dst_type , src_type >( dst , src );
   }
   else {
     Kokkos::Impl::throw_runtime_exception("deep_copy given views that would require a temporary allocation");
@@ -581,7 +581,7 @@ void deep_copy( const DynamicView & dst
 
   if ( DstExecCanAccessSrc ) {
     // Copying data between views in accessible memory spaces and either non-contiguous or incompatible shape.
-    Kokkos::Experimental::Impl::ViewRemap< dst_type , src_type >( dst , src );
+    Kokkos::Impl::ViewRemap< dst_type , src_type >( dst , src );
   }
   else {
     Kokkos::Impl::throw_runtime_exception("deep_copy given views that would require a temporary allocation");
diff --git a/lib/kokkos/containers/unit_tests/TestCuda.cpp b/lib/kokkos/containers/unit_tests/TestCuda.cpp
index 5a78a5de9e..651a4e7eb8 100644
--- a/lib/kokkos/containers/unit_tests/TestCuda.cpp
+++ b/lib/kokkos/containers/unit_tests/TestCuda.cpp
@@ -69,6 +69,8 @@
 #include 
 #include 
 
+#include 
+
 //----------------------------------------------------------------------------
 
 
@@ -94,6 +96,10 @@ TEST_F( cuda , dyn_view_api) {
   TestDynViewAPI< double , Kokkos::Cuda >();
 }
 
+TEST_F( cuda, viewctorprop_embedded_dim ) {
+  TestViewCtorProp_EmbeddedDim< Kokkos::Cuda >::test_vcpt( 2, 3 );
+}
+
 TEST_F( cuda , staticcrsgraph )
 {
   TestStaticCrsGraph::run_test_graph< Kokkos::Cuda >();
diff --git a/lib/kokkos/containers/unit_tests/TestOpenMP.cpp b/lib/kokkos/containers/unit_tests/TestOpenMP.cpp
index 2448bd077b..5365d91361 100644
--- a/lib/kokkos/containers/unit_tests/TestOpenMP.cpp
+++ b/lib/kokkos/containers/unit_tests/TestOpenMP.cpp
@@ -66,6 +66,8 @@
 #include 
 #include 
 
+#include 
+
 #include 
 
 namespace Test {
@@ -76,14 +78,7 @@ protected:
   {
     std::cout << std::setprecision(5) << std::scientific;
 
-    unsigned threads_count = 4 ;
-
-    if ( Kokkos::hwloc::available() ) {
-      threads_count = Kokkos::hwloc::get_available_numa_count() *
-                      Kokkos::hwloc::get_available_cores_per_numa();
-    }
-
-    Kokkos::OpenMP::initialize( threads_count );
+    Kokkos::OpenMP::initialize();
   }
 
   static void TearDownTestCase()
@@ -96,6 +91,10 @@ TEST_F( openmp, dyn_view_api) {
   TestDynViewAPI< double , Kokkos::OpenMP >();
 }
 
+TEST_F( openmp, viewctorprop_embedded_dim ) {
+  TestViewCtorProp_EmbeddedDim< Kokkos::OpenMP >::test_vcpt( 2, 3 );
+}
+
 TEST_F( openmp, bitset )
 {
   test_bitset();
diff --git a/lib/kokkos/containers/unit_tests/TestSerial.cpp b/lib/kokkos/containers/unit_tests/TestSerial.cpp
index 06c4d9f6ed..1b9b5a2da3 100644
--- a/lib/kokkos/containers/unit_tests/TestSerial.cpp
+++ b/lib/kokkos/containers/unit_tests/TestSerial.cpp
@@ -67,6 +67,8 @@
 #include 
 #include 
 
+#include 
+
 namespace Test {
 
 class serial : public ::testing::Test {
@@ -85,6 +87,10 @@ TEST_F( serial, dyn_view_api) {
   TestDynViewAPI< double , Kokkos::Serial >();
 }
 
+TEST_F( serial, viewctorprop_embedded_dim ) {
+  TestViewCtorProp_EmbeddedDim< Kokkos::Serial >::test_vcpt( 2, 3 );
+}
+
 TEST_F( serial , staticcrsgraph )
 {
   TestStaticCrsGraph::run_test_graph< Kokkos::Serial >();
diff --git a/lib/kokkos/containers/unit_tests/TestThreads.cpp b/lib/kokkos/containers/unit_tests/TestThreads.cpp
index 938ec88e90..aca0b57d65 100644
--- a/lib/kokkos/containers/unit_tests/TestThreads.cpp
+++ b/lib/kokkos/containers/unit_tests/TestThreads.cpp
@@ -70,6 +70,8 @@
 #include 
 #include 
 
+#include 
+
 namespace Test {
 
 class threads : public ::testing::Test {
@@ -103,6 +105,10 @@ TEST_F( threads , dyn_view_api) {
   TestDynViewAPI< double , Kokkos::Threads >();
 }
 
+TEST_F( threads, viewctorprop_embedded_dim ) {
+  TestViewCtorProp_EmbeddedDim< Kokkos::Threads >::test_vcpt( 2, 3 );
+}
+
 TEST_F( threads , staticcrsgraph )
 {
   TestStaticCrsGraph::run_test_graph< Kokkos::Threads >();
diff --git a/lib/kokkos/containers/unit_tests/TestViewCtorPropEmbeddedDim.hpp b/lib/kokkos/containers/unit_tests/TestViewCtorPropEmbeddedDim.hpp
new file mode 100644
index 0000000000..1efd1ddc51
--- /dev/null
+++ b/lib/kokkos/containers/unit_tests/TestViewCtorPropEmbeddedDim.hpp
@@ -0,0 +1,213 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include 
+
+#include 
+
+#include 
+#include 
+
+#include 
+#include 
+
+namespace Test {
+
+namespace {
+
+template 
+struct TestViewCtorProp_EmbeddedDim {
+
+  using ViewIntType     = typename Kokkos::View< int**, ExecSpace >;
+  using ViewDoubleType     = typename Kokkos::View< double*, ExecSpace >;
+
+  using DynRankViewIntType     = typename Kokkos::DynRankView< int, ExecSpace >;
+  using DynRankViewDoubleType     = typename Kokkos::DynRankView< double, ExecSpace >;
+
+  // Cuda 7.0 has issues with using a lamda in parallel_for to initialize the view - replace with this functor
+  template < class ViewType >
+  struct Functor {
+
+    ViewType v;
+
+    Functor( const ViewType & v_ ) : v(v_) {}
+
+    KOKKOS_INLINE_FUNCTION
+    void operator()( const int i ) const {
+      v(i) = i;
+    }
+
+  };
+
+
+  static void test_vcpt( const int N0, const int N1 )
+  {
+
+    // Create two views to test
+    {
+      using VIT = typename TestViewCtorProp_EmbeddedDim::ViewIntType ;
+      using VDT = typename TestViewCtorProp_EmbeddedDim::ViewDoubleType ;
+
+      VIT vi1("vi1", N0, N1);
+      VDT vd1("vd1", N0);
+
+      // TEST: Test for common type between two views, one with type double, other with type int
+      // Deduce common value_type and construct a view with that type
+      {
+        // Two views
+        auto view_alloc_arg = Kokkos::common_view_alloc_prop(vi1, vd1);
+        typedef typename decltype( view_alloc_arg )::value_type                    CommonViewValueType;
+        typedef typename Kokkos::View< CommonViewValueType*, ExecSpace >  CVT;
+        typedef typename CVT::HostMirror                                           HostCVT;
+
+        // Construct View using the common type; for case of specialization, an 'embedded_dim' would be stored by view_alloc_arg
+        CVT cv1( Kokkos::view_alloc( "cv1", view_alloc_arg ), N0*N1 );
+
+        Kokkos::parallel_for( Kokkos::RangePolicy< ExecSpace >(0, N0*N1), 
+          Functor(cv1)
+        );
+
+        HostCVT hcv1 = Kokkos::create_mirror_view( cv1 );
+        Kokkos::deep_copy( hcv1, cv1 );
+
+        ASSERT_EQ( (std::is_same< CommonViewValueType, double >::value) , true ) ;
+      #if 0
+      // debug output
+      for ( int i = 0; i < N0*N1; ++i ) {
+        printf(" Output check: hcv1(%d) = %lf\n ", i, hcv1(i) );
+      }
+
+      printf( " Common value type view: %s \n", typeid( CVT() ).name() );
+      printf( " Common value type: %s \n", typeid( CommonViewValueType() ).name() );
+      if ( std::is_same< CommonViewValueType, double >::value == true ) {
+        printf("Proper common value_type\n");
+      }
+      else {
+        printf("WRONG common value_type\n");
+      }
+      // end debug output
+      #endif
+      }
+
+      {
+        // Single view
+        auto view_alloc_arg = Kokkos::common_view_alloc_prop(vi1);
+        typedef typename decltype( view_alloc_arg )::value_type                    CommonViewValueType;
+        typedef typename Kokkos::View< CommonViewValueType*, ExecSpace >  CVT;
+        typedef typename CVT::HostMirror                                           HostCVT;
+
+        // Construct View using the common type; for case of specialization, an 'embedded_dim' would be stored by view_alloc_arg
+        CVT cv1( Kokkos::view_alloc( "cv1", view_alloc_arg ), N0*N1 );
+
+        Kokkos::parallel_for( Kokkos::RangePolicy< ExecSpace >(0, N0*N1), 
+          Functor(cv1)
+        );
+
+        HostCVT hcv1 = Kokkos::create_mirror_view( cv1 );
+        Kokkos::deep_copy( hcv1, cv1 );
+
+        ASSERT_EQ( (std::is_same< CommonViewValueType, int>::value) , true ) ;
+      }
+
+    }
+
+    // Create two dynamic rank views to test
+    {
+      using VIT = typename TestViewCtorProp_EmbeddedDim::DynRankViewIntType ;
+      using VDT = typename TestViewCtorProp_EmbeddedDim::DynRankViewDoubleType ;
+
+      VIT vi1("vi1", N0, N1);
+      VDT vd1("vd1", N0);
+
+      // TEST: Test for common type between two views, one with type double, other with type int
+      // Deduce common value_type and construct a view with that type
+      {
+        // Two views
+        auto view_alloc_arg = Kokkos::common_view_alloc_prop( vi1, vd1 );
+        typedef typename decltype( view_alloc_arg )::value_type                    CommonViewValueType;
+        typedef typename Kokkos::View< CommonViewValueType*, ExecSpace >  CVT;
+        typedef typename CVT::HostMirror                                           HostCVT;
+
+        // Construct View using the common type; for case of specialization, an 'embedded_dim' would be stored by view_alloc_arg
+        CVT cv1( Kokkos::view_alloc( "cv1", view_alloc_arg ), N0*N1 );
+
+
+        Kokkos::parallel_for( Kokkos::RangePolicy< ExecSpace >(0, N0*N1), 
+          Functor(cv1)
+        );
+
+        HostCVT hcv1 = Kokkos::create_mirror_view( cv1 );
+        Kokkos::deep_copy( hcv1, cv1 );
+
+        ASSERT_EQ( (std::is_same< CommonViewValueType, double >::value) , true ) ;
+      }
+
+      {
+        // Single views
+        auto view_alloc_arg = Kokkos::common_view_alloc_prop( vi1 );
+        typedef typename decltype( view_alloc_arg )::value_type                    CommonViewValueType;
+        typedef typename Kokkos::View< CommonViewValueType*, ExecSpace >  CVT;
+        typedef typename CVT::HostMirror                                           HostCVT;
+
+        // Construct View using the common type; for case of specialization, an 'embedded_dim' would be stored by view_alloc_arg
+        CVT cv1( Kokkos::view_alloc( "cv1", view_alloc_arg ), N0*N1 );
+
+        Kokkos::parallel_for( Kokkos::RangePolicy< ExecSpace >(0, N0*N1), 
+          Functor(cv1)
+        );
+
+        HostCVT hcv1 = Kokkos::create_mirror_view( cv1 );
+        Kokkos::deep_copy( hcv1, cv1 );
+
+        ASSERT_EQ( (std::is_same< CommonViewValueType, int>::value) , true ) ;
+      }
+    }
+
+
+  } // end test_vcpt
+
+}; // end struct
+
+} // namespace
+
+} // namespace Test
diff --git a/lib/kokkos/containers/unit_tests/UnitTestMain.cpp b/lib/kokkos/containers/unit_tests/UnitTestMain.cpp
index f952ab3db5..2b73535c83 100644
--- a/lib/kokkos/containers/unit_tests/UnitTestMain.cpp
+++ b/lib/kokkos/containers/unit_tests/UnitTestMain.cpp
@@ -1,13 +1,13 @@
 /*
 //@HEADER
 // ************************************************************************
-// 
+//
 //                        Kokkos v. 2.0
 //              Copyright (2014) Sandia Corporation
-// 
+//
 // Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
 // the U.S. Government retains certain rights in this software.
-// 
+//
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
@@ -36,12 +36,14 @@
 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 //
 // Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
-// 
+//
 // ************************************************************************
 //@HEADER
 */
 
 #include 
+#include 
+#include 
 
 int main(int argc, char *argv[]) {
   ::testing::InitGoogleTest(&argc,argv);
diff --git a/lib/kokkos/core/perf_test/Makefile b/lib/kokkos/core/perf_test/Makefile
index f59e7bbe1c..bb9353f583 100644
--- a/lib/kokkos/core/perf_test/Makefile
+++ b/lib/kokkos/core/perf_test/Makefile
@@ -79,7 +79,6 @@ test-mempool: KokkosCore_PerformanceTest_Mempool
 test-taskdag: KokkosCore_PerformanceTest_TaskDAG
 	./KokkosCore_PerformanceTest_TaskDAG
 
-
 build_all: $(TARGETS)
 
 test: $(TEST_TARGETS)
diff --git a/lib/kokkos/core/perf_test/PerfTestMain.cpp b/lib/kokkos/core/perf_test/PerfTestMain.cpp
index d80cfab8b5..832f650b9a 100644
--- a/lib/kokkos/core/perf_test/PerfTestMain.cpp
+++ b/lib/kokkos/core/perf_test/PerfTestMain.cpp
@@ -1,13 +1,13 @@
 /*
 //@HEADER
 // ************************************************************************
-// 
+//
 //                        Kokkos v. 2.0
 //              Copyright (2014) Sandia Corporation
-// 
+//
 // Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
 // the U.S. Government retains certain rights in this software.
-// 
+//
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
@@ -36,12 +36,14 @@
 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 //
 // Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
-// 
+//
 // ************************************************************************
 //@HEADER
 */
 
 #include 
+#include 
+
 #include 
 
 namespace Test {
diff --git a/lib/kokkos/core/src/Cuda/KokkosExp_Cuda_IterateTile_Refactor.hpp b/lib/kokkos/core/src/Cuda/KokkosExp_Cuda_IterateTile_Refactor.hpp
new file mode 100644
index 0000000000..46321378d9
--- /dev/null
+++ b/lib/kokkos/core/src/Cuda/KokkosExp_Cuda_IterateTile_Refactor.hpp
@@ -0,0 +1,2715 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_CUDA_EXP_ITERATE_TILE_REFACTOR_HPP
+#define KOKKOS_CUDA_EXP_ITERATE_TILE_REFACTOR_HPP
+
+#include 
+#if defined( __CUDACC__ ) && defined( KOKKOS_ENABLE_CUDA )
+
+#include 
+#include 
+#include 
+
+#include 
+
+// #include
+// Including the file above leads to following type of errors:
+// /home/ndellin/kokkos/core/src/Cuda/Kokkos_CudaExec.hpp(84): error: incomplete type is not allowed
+// use existing Kokkos functionality, e.g. max blocks, once resolved
+
+#if defined(KOKKOS_ENABLE_PROFILING)
+#include 
+#include 
+#endif
+
+namespace Kokkos { namespace Experimental { namespace Impl {
+
+namespace Refactor {
+
+// ------------------------------------------------------------------ //
+// ParallelFor iteration pattern
+template< int N , typename RP , typename Functor , typename Tag >
+struct DeviceIterateTile;
+
+//Rank 2
+// Specializations for void tag type
+template< typename RP , typename Functor >
+struct DeviceIterateTile<2,RP,Functor,void >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ )
+  : m_rp(rp_)
+  , m_func(f_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    // LL
+    if (RP::inner_direction == RP::Left) {
+      for ( index_type tile_id1 = (index_type)blockIdx.y; tile_id1 < m_rp.m_tile_end[1]; tile_id1 += gridDim.y ) {
+        const index_type offset_1 = tile_id1*m_rp.m_tile[1] + (index_type)threadIdx.y;
+        if ( offset_1 < m_rp.m_upper[1] && (index_type)threadIdx.y < m_rp.m_tile[1] ) {
+
+          for ( index_type tile_id0 = (index_type)blockIdx.x; tile_id0 < m_rp.m_tile_end[0]; tile_id0 += gridDim.x ) {
+            const index_type offset_0 = tile_id0*m_rp.m_tile[0] + (index_type)threadIdx.x;
+            if ( offset_0 < m_rp.m_upper[0] && (index_type)threadIdx.x < m_rp.m_tile[0] ) {
+              m_func(offset_0 , offset_1);
+            }
+          }
+        }
+      }
+    }
+    // LR
+    else {
+      for ( index_type tile_id0 = (index_type)blockIdx.x; tile_id0 < m_rp.m_tile_end[0]; tile_id0 += gridDim.x ) {
+        const index_type offset_0 = tile_id0*m_rp.m_tile[0] + (index_type)threadIdx.x;
+        if ( offset_0 < m_rp.m_upper[0] && (index_type)threadIdx.x < m_rp.m_tile[0] ) {
+
+          for ( index_type tile_id1 = (index_type)blockIdx.y; tile_id1 < m_rp.m_tile_end[1]; tile_id1 += gridDim.y ) {
+            const index_type offset_1 = tile_id1*m_rp.m_tile[1] + (index_type)threadIdx.y;
+            if ( offset_1 < m_rp.m_upper[1] && (index_type)threadIdx.y < m_rp.m_tile[1] ) {
+              m_func(offset_0 , offset_1);
+            }
+          }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+};
+
+// Specializations for tag type
+template< typename RP , typename Functor , typename Tag >
+struct DeviceIterateTile<2,RP,Functor,Tag>
+{
+  using index_type = typename RP::index_type;
+
+  inline __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ )
+  : m_rp(rp_)
+  , m_func(f_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    if (RP::inner_direction == RP::Left) {
+      // Loop over size maxnumblocks until full range covered
+      for ( index_type tile_id1 = (index_type)blockIdx.y; tile_id1 < m_rp.m_tile_end[1]; tile_id1 += gridDim.y ) {
+        const index_type offset_1 = tile_id1*m_rp.m_tile[1] + (index_type)threadIdx.y;
+        if ( offset_1 < m_rp.m_upper[1] && (index_type)threadIdx.y < m_rp.m_tile[1] ) {
+
+          for ( index_type tile_id0 = (index_type)blockIdx.x; tile_id0 < m_rp.m_tile_end[0]; tile_id0 += gridDim.x ) {
+            const index_type offset_0 = tile_id0*m_rp.m_tile[0] + (index_type)threadIdx.x;
+            if ( offset_0 < m_rp.m_upper[0] && (index_type)threadIdx.x < m_rp.m_tile[0] ) {
+              m_func(Tag(), offset_0 , offset_1);
+            }
+          }
+        }
+      }
+    }
+    else {
+      for ( index_type tile_id0 = (index_type)blockIdx.x; tile_id0 < m_rp.m_tile_end[0]; tile_id0 += gridDim.x ) {
+        const index_type offset_0 = tile_id0*m_rp.m_tile[0] + (index_type)threadIdx.x;
+        if ( offset_0 < m_rp.m_upper[0] && (index_type)threadIdx.x < m_rp.m_tile[0] ) {
+
+          for ( index_type tile_id1 = (index_type)blockIdx.y; tile_id1 < m_rp.m_tile_end[1]; tile_id1 += gridDim.y ) {
+            const index_type offset_1 = tile_id1*m_rp.m_tile[1] + (index_type)threadIdx.y;
+            if ( offset_1 < m_rp.m_upper[1] && (index_type)threadIdx.y < m_rp.m_tile[1] ) {
+              m_func(Tag(), offset_0 , offset_1);
+            }
+          }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+};
+
+
+//Rank 3
+// Specializations for void tag type
+template< typename RP , typename Functor >
+struct DeviceIterateTile<3,RP,Functor,void >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ )
+  : m_rp(rp_)
+  , m_func(f_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    // LL
+    if (RP::inner_direction == RP::Left) {
+      for ( index_type tile_id2 = (index_type)blockIdx.z; tile_id2 < m_rp.m_tile_end[2]; tile_id2 += gridDim.z ) {
+        const index_type offset_2 = tile_id2*m_rp.m_tile[2] + (index_type)threadIdx.z;
+        if ( offset_2 < m_rp.m_upper[2] && (index_type)threadIdx.z < m_rp.m_tile[2] ) {
+
+          for ( index_type tile_id1 = (index_type)blockIdx.y; tile_id1 < m_rp.m_tile_end[1]; tile_id1 += gridDim.y ) {
+            const index_type offset_1 = tile_id1*m_rp.m_tile[1] + (index_type)threadIdx.y;
+            if ( offset_1 < m_rp.m_upper[1] && (index_type)threadIdx.y < m_rp.m_tile[1] ) {
+
+              for ( index_type tile_id0 = (index_type)blockIdx.x; tile_id0 < m_rp.m_tile_end[0]; tile_id0 += gridDim.x ) {
+                const index_type offset_0 = tile_id0*m_rp.m_tile[0] + (index_type)threadIdx.x;
+                if ( offset_0 < m_rp.m_upper[0] && (index_type)threadIdx.x < m_rp.m_tile[0] ) {
+                  m_func(offset_0 , offset_1 , offset_2);
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+    // LR
+    else {
+      for ( index_type tile_id0 = (index_type)blockIdx.x; tile_id0 < m_rp.m_tile_end[0]; tile_id0 += gridDim.x ) {
+        const index_type offset_0 = tile_id0*m_rp.m_tile[0] + (index_type)threadIdx.x;
+        if ( offset_0 < m_rp.m_upper[0] && (index_type)threadIdx.x < m_rp.m_tile[0] ) {
+
+          for ( index_type tile_id1 = (index_type)blockIdx.y; tile_id1 < m_rp.m_tile_end[1]; tile_id1 += gridDim.y ) {
+            const index_type offset_1 = tile_id1*m_rp.m_tile[1] + (index_type)threadIdx.y;
+            if ( offset_1 < m_rp.m_upper[1] && (index_type)threadIdx.y < m_rp.m_tile[1] ) {
+
+              for ( index_type tile_id2 = (index_type)blockIdx.z; tile_id2 < m_rp.m_tile_end[2]; tile_id2 += gridDim.z ) {
+                const index_type offset_2 = tile_id2*m_rp.m_tile[2] + (index_type)threadIdx.z;
+                if ( offset_2 < m_rp.m_upper[2] && (index_type)threadIdx.z < m_rp.m_tile[2] ) {
+                  m_func(offset_0 , offset_1 , offset_2);
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+};
+
+// Specializations for void tag type
+template< typename RP , typename Functor , typename Tag >
+struct DeviceIterateTile<3,RP,Functor,Tag>
+{
+  using index_type = typename RP::index_type;
+
+  inline __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ )
+  : m_rp(rp_)
+  , m_func(f_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    if (RP::inner_direction == RP::Left) {
+      for ( index_type tile_id2 = (index_type)blockIdx.z; tile_id2 < m_rp.m_tile_end[2]; tile_id2 += gridDim.z ) {
+        const index_type offset_2 = tile_id2*m_rp.m_tile[2] + (index_type)threadIdx.z;
+        if ( offset_2 < m_rp.m_upper[2] && (index_type)threadIdx.z < m_rp.m_tile[2] ) {
+
+          for ( index_type tile_id1 = (index_type)blockIdx.y; tile_id1 < m_rp.m_tile_end[1]; tile_id1 += gridDim.y ) {
+            const index_type offset_1 = tile_id1*m_rp.m_tile[1] + (index_type)threadIdx.y;
+            if ( offset_1 < m_rp.m_upper[1] && (index_type)threadIdx.y < m_rp.m_tile[1] ) {
+
+              for ( index_type tile_id0 = (index_type)blockIdx.x; tile_id0 < m_rp.m_tile_end[0]; tile_id0 += gridDim.x ) {
+                const index_type offset_0 = tile_id0*m_rp.m_tile[0] + (index_type)threadIdx.x;
+                if ( offset_0 < m_rp.m_upper[0] && (index_type)threadIdx.x < m_rp.m_tile[0] ) {
+                  m_func(Tag(), offset_0 , offset_1 , offset_2);
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+    else {
+      for ( index_type tile_id0 = (index_type)blockIdx.x; tile_id0 < m_rp.m_tile_end[0]; tile_id0 += gridDim.x ) {
+        const index_type offset_0 = tile_id0*m_rp.m_tile[0] + (index_type)threadIdx.x;
+        if ( offset_0 < m_rp.m_upper[0] && (index_type)threadIdx.x < m_rp.m_tile[0] ) {
+
+          for ( index_type tile_id1 = (index_type)blockIdx.y; tile_id1 < m_rp.m_tile_end[1]; tile_id1 += gridDim.y ) {
+            const index_type offset_1 = tile_id1*m_rp.m_tile[1] + (index_type)threadIdx.y;
+            if ( offset_1 < m_rp.m_upper[1] && (index_type)threadIdx.y < m_rp.m_tile[1] ) {
+
+              for ( index_type tile_id2 = (index_type)blockIdx.z; tile_id2 < m_rp.m_tile_end[2]; tile_id2 += gridDim.z ) {
+                const index_type offset_2 = tile_id2*m_rp.m_tile[2] + (index_type)threadIdx.z;
+                if ( offset_2 < m_rp.m_upper[2] && (index_type)threadIdx.z < m_rp.m_tile[2] ) {
+                  m_func(Tag(), offset_0 , offset_1 , offset_2);
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+};
+
+
+//Rank 4
+// Specializations for void tag type
+template< typename RP , typename Functor >
+struct DeviceIterateTile<4,RP,Functor,void >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ )
+  : m_rp(rp_)
+  , m_func(f_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    // LL
+    if (RP::inner_direction == RP::Left) {
+      const index_type temp0  =  m_rp.m_tile_end[0];
+      const index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl0 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl1 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl0 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x % numbl0;
+      const index_type tile_id1 = (index_type)blockIdx.x / numbl0;
+      const index_type thr_id0 = (index_type)threadIdx.x % m_rp.m_tile[0];
+      const index_type thr_id1 = (index_type)threadIdx.x / m_rp.m_tile[0];
+
+      for ( index_type tile_id3 = (index_type)blockIdx.z; tile_id3 < m_rp.m_tile_end[3]; tile_id3 += gridDim.z ) {
+        const index_type offset_3 = tile_id3*m_rp.m_tile[3] + (index_type)threadIdx.z;
+        if ( offset_3 < m_rp.m_upper[3] && (index_type)threadIdx.z < m_rp.m_tile[3] ) {
+
+          for ( index_type tile_id2 = (index_type)blockIdx.y; tile_id2 < m_rp.m_tile_end[2]; tile_id2 += gridDim.y ) {
+            const index_type offset_2 = tile_id2*m_rp.m_tile[2] + (index_type)threadIdx.y;
+            if ( offset_2 < m_rp.m_upper[2] && (index_type)threadIdx.y < m_rp.m_tile[2] ) {
+
+              for ( index_type j = tile_id1 ; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+                const index_type offset_1 = j*m_rp.m_tile[1] + thr_id1;
+                if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+                  for ( index_type i = tile_id0 ; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+                    const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+                    if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+                      m_func(offset_0 , offset_1 , offset_2 , offset_3);
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+    // LR
+    else {
+      const index_type temp0  =  m_rp.m_tile_end[0];
+      const index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl1 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl0 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl1 ) :
+          ( temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x / numbl1;
+      const index_type tile_id1 = (index_type)blockIdx.x % numbl1;
+      const index_type thr_id0 = (index_type)threadIdx.x / m_rp.m_tile[1];
+      const index_type thr_id1 = (index_type)threadIdx.x % m_rp.m_tile[1];
+
+      for ( index_type i = tile_id0; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+        const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+        if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+
+          for ( index_type j = tile_id1; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+            const index_type offset_1 = j*m_rp.m_tile[1] + thr_id1;
+            if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+              for ( index_type tile_id2 = (index_type)blockIdx.y; tile_id2 < m_rp.m_tile_end[2]; tile_id2 += gridDim.y ) {
+                const index_type offset_2 = tile_id2*m_rp.m_tile[2] + (index_type)threadIdx.y;
+                if ( offset_2 < m_rp.m_upper[2] && (index_type)threadIdx.y < m_rp.m_tile[2] ) {
+
+                  for ( index_type tile_id3 = (index_type)blockIdx.z; tile_id3 < m_rp.m_tile_end[3]; tile_id3 += gridDim.z ) {
+                    const index_type offset_3 = tile_id3*m_rp.m_tile[3] + (index_type)threadIdx.z;
+                    if ( offset_3 < m_rp.m_upper[3] && (index_type)threadIdx.z < m_rp.m_tile[3] ) {
+                      m_func(offset_0 , offset_1 , offset_2 , offset_3);
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+};
+
+// Specializations for void tag type
+template< typename RP , typename Functor , typename Tag >
+struct DeviceIterateTile<4,RP,Functor,Tag>
+{
+  using index_type = typename RP::index_type;
+
+  inline __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ )
+  : m_rp(rp_)
+  , m_func(f_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if (RP::inner_direction == RP::Left) {
+      const index_type temp0  =  m_rp.m_tile_end[0];
+      const index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl0 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl1 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl0 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x % numbl0;
+      const index_type tile_id1 = (index_type)blockIdx.x / numbl0;
+      const index_type thr_id0 = (index_type)threadIdx.x % m_rp.m_tile[0];
+      const index_type thr_id1 = (index_type)threadIdx.x / m_rp.m_tile[0];
+
+      for ( index_type tile_id3 = (index_type)blockIdx.z; tile_id3 < m_rp.m_tile_end[3]; tile_id3 += gridDim.z ) {
+        const index_type offset_3 = tile_id3*m_rp.m_tile[3] + (index_type)threadIdx.z;
+        if ( offset_3 < m_rp.m_upper[3] && (index_type)threadIdx.z < m_rp.m_tile[3] ) {
+
+          for ( index_type tile_id2 = (index_type)blockIdx.y; tile_id2 < m_rp.m_tile_end[2]; tile_id2 += gridDim.y ) {
+            const index_type offset_2 = tile_id2*m_rp.m_tile[2] + (index_type)threadIdx.y;
+            if ( offset_2 < m_rp.m_upper[2] && (index_type)threadIdx.y < m_rp.m_tile[2] ) {
+
+              for ( index_type j = tile_id1; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+                const index_type offset_1 = j*m_rp.m_tile[1] + thr_id1;
+                if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+                  for ( index_type i = tile_id0; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+                    const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+                    if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+                      m_func(Tag(), offset_0 , offset_1 , offset_2 , offset_3);
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+    else {
+      const index_type temp0  =  m_rp.m_tile_end[0];
+      const index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl1 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl0 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl1 ) :
+          ( temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x / numbl1;
+      const index_type tile_id1 = (index_type)blockIdx.x % numbl1;
+      const index_type thr_id0 = (index_type)threadIdx.x / m_rp.m_tile[1];
+      const index_type thr_id1 = (index_type)threadIdx.x % m_rp.m_tile[1];
+
+      for ( index_type i = tile_id0; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+        const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+        if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+
+          for ( index_type j = tile_id1; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+            const index_type offset_1 = tile_id1*m_rp.m_tile[1] + thr_id1;
+            if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+              for ( index_type tile_id2 = (index_type)blockIdx.y; tile_id2 < m_rp.m_tile_end[2]; tile_id2 += gridDim.y ) {
+                const index_type offset_2 = tile_id2*m_rp.m_tile[2] + (index_type)threadIdx.y;
+                if ( offset_2 < m_rp.m_upper[2] && (index_type)threadIdx.y < m_rp.m_tile[2] ) {
+
+                  for ( index_type tile_id3 = (index_type)blockIdx.z; tile_id3 < m_rp.m_tile_end[3]; tile_id3 += gridDim.z ) {
+                    const index_type offset_3 = tile_id3*m_rp.m_tile[3] + (index_type)threadIdx.z;
+                    if ( offset_3 < m_rp.m_upper[3] && (index_type)threadIdx.z < m_rp.m_tile[3] ) {
+                      m_func(Tag() , offset_0 , offset_1 , offset_2 , offset_3);
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+};
+
+
+//Rank 5
+// Specializations for void tag type
+template< typename RP , typename Functor >
+struct DeviceIterateTile<5,RP,Functor,void >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ )
+  : m_rp(rp_)
+  , m_func(f_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    // LL
+    if (RP::inner_direction == RP::Left) {
+
+      index_type temp0  =  m_rp.m_tile_end[0];
+      index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl0 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl1 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl0 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x % numbl0;
+      const index_type tile_id1 = (index_type)blockIdx.x / numbl0;
+      const index_type thr_id0 = (index_type)threadIdx.x % m_rp.m_tile[0];
+      const index_type thr_id1 = (index_type)threadIdx.x / m_rp.m_tile[0];
+
+      temp0  =  m_rp.m_tile_end[2];
+      temp1  =  m_rp.m_tile_end[3];
+      const index_type numbl2 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl3 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl2 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id2 = (index_type)blockIdx.y % numbl2;
+      const index_type tile_id3 = (index_type)blockIdx.y / numbl2;
+      const index_type thr_id2 = (index_type)threadIdx.y % m_rp.m_tile[2];
+      const index_type thr_id3 = (index_type)threadIdx.y / m_rp.m_tile[2];
+
+      for ( index_type tile_id4 = (index_type)blockIdx.z; tile_id4 < m_rp.m_tile_end[4]; tile_id4 += gridDim.z ) {
+        const index_type offset_4 = tile_id4*m_rp.m_tile[4] + (index_type)threadIdx.z;
+        if ( offset_4 < m_rp.m_upper[4] && (index_type)threadIdx.z < m_rp.m_tile[4] ) {
+
+          for ( index_type l = tile_id3; l < m_rp.m_tile_end[3]; l += numbl3 ) {
+            const index_type offset_3 = l*m_rp.m_tile[3] + thr_id3;
+            if ( offset_3 < m_rp.m_upper[3] && thr_id3 < m_rp.m_tile[3] ) {
+
+              for ( index_type k = tile_id2; k < m_rp.m_tile_end[2]; k += numbl2 ) {
+                const index_type offset_2 = k*m_rp.m_tile[2] + thr_id2;
+                if ( offset_2 < m_rp.m_upper[2] && thr_id2 < m_rp.m_tile[2] ) {
+
+                  for ( index_type j = tile_id1 ; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+                    const index_type offset_1 = j*m_rp.m_tile[1] + thr_id1;
+                    if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+                      for ( index_type i = tile_id0 ; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+                        const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+                        if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+                          m_func(offset_0 , offset_1 , offset_2 , offset_3, offset_4);
+                        }
+                      }
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+    // LR
+    else {
+      index_type temp0  =  m_rp.m_tile_end[0];
+      index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl1 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl0 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl1 ) :
+          ( temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x / numbl1;
+      const index_type tile_id1 = (index_type)blockIdx.x % numbl1;
+      const index_type thr_id0 = (index_type)threadIdx.x / m_rp.m_tile[1];
+      const index_type thr_id1 = (index_type)threadIdx.x % m_rp.m_tile[1];
+
+      temp0  =  m_rp.m_tile_end[2];
+      temp1  =  m_rp.m_tile_end[3];
+      const index_type numbl3 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl2 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl3 ) :
+          (  temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id2 = (index_type)blockIdx.y / numbl3;
+      const index_type tile_id3 = (index_type)blockIdx.y % numbl3;
+      const index_type thr_id2 = (index_type)threadIdx.y / m_rp.m_tile[3];
+      const index_type thr_id3 = (index_type)threadIdx.y % m_rp.m_tile[3];
+
+      for ( index_type i = tile_id0; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+        const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+        if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+
+          for ( index_type j = tile_id1; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+            const index_type offset_1 = j*m_rp.m_tile[1] + thr_id1;
+            if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+              for ( index_type k = tile_id2; k < m_rp.m_tile_end[2]; k += numbl2 ) {
+                const index_type offset_2 = k*m_rp.m_tile[2] + thr_id2;
+                if ( offset_2 < m_rp.m_upper[2] && thr_id2 < m_rp.m_tile[2] ) {
+
+                  for ( index_type l = tile_id3; l < m_rp.m_tile_end[3]; l += numbl3 ) {
+                    const index_type offset_3 = l*m_rp.m_tile[3] + thr_id3;
+                    if ( offset_3 < m_rp.m_upper[3] && thr_id3 < m_rp.m_tile[3] ) {
+
+                      for ( index_type tile_id4 = (index_type)blockIdx.z; tile_id4 < m_rp.m_tile_end[4]; tile_id4 += gridDim.z ) {
+                        const index_type offset_4 = tile_id4*m_rp.m_tile[4] + (index_type)threadIdx.z;
+                        if ( offset_4 < m_rp.m_upper[4] && (index_type)threadIdx.z < m_rp.m_tile[4] ) {
+                          m_func(offset_0 , offset_1 , offset_2 , offset_3 , offset_4);
+                        }
+                      }
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+};
+
+// Specializations for tag type
+template< typename RP , typename Functor , typename Tag >
+struct DeviceIterateTile<5,RP,Functor,Tag>
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ )
+  : m_rp(rp_)
+  , m_func(f_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    // LL
+    if (RP::inner_direction == RP::Left) {
+      index_type temp0  =  m_rp.m_tile_end[0];
+      index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl0 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl1 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl0 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x % numbl0;
+      const index_type tile_id1 = (index_type)blockIdx.x / numbl0;
+      const index_type thr_id0 = (index_type)threadIdx.x % m_rp.m_tile[0];
+      const index_type thr_id1 = (index_type)threadIdx.x / m_rp.m_tile[0];
+
+      temp0  =  m_rp.m_tile_end[2];
+      temp1  =  m_rp.m_tile_end[3];
+      const index_type numbl2 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl3 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl2 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id2 = (index_type)blockIdx.y % numbl2;
+      const index_type tile_id3 = (index_type)blockIdx.y / numbl2;
+      const index_type thr_id2 = (index_type)threadIdx.y % m_rp.m_tile[2];
+      const index_type thr_id3 = (index_type)threadIdx.y / m_rp.m_tile[2];
+
+      for ( index_type tile_id4 = (index_type)blockIdx.z; tile_id4 < m_rp.m_tile_end[4]; tile_id4 += gridDim.z ) {
+        const index_type offset_4 = tile_id4*m_rp.m_tile[4] + (index_type)threadIdx.z;
+        if ( offset_4 < m_rp.m_upper[4] && (index_type)threadIdx.z < m_rp.m_tile[4] ) {
+
+          for ( index_type l = tile_id3; l < m_rp.m_tile_end[3]; l += numbl3 ) {
+            const index_type offset_3 = l*m_rp.m_tile[3] + thr_id3;
+            if ( offset_3 < m_rp.m_upper[3] && thr_id3 < m_rp.m_tile[3] ) {
+
+              for ( index_type k = tile_id2; k < m_rp.m_tile_end[2]; k += numbl2 ) {
+                const index_type offset_2 = k*m_rp.m_tile[2] + thr_id2;
+                if ( offset_2 < m_rp.m_upper[2] && thr_id2 < m_rp.m_tile[2] ) {
+
+                  for ( index_type j = tile_id1 ; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+                    const index_type offset_1 = j*m_rp.m_tile[1] + thr_id1;
+                    if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+                      for ( index_type i = tile_id0 ; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+                        const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+                        if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+                          m_func(Tag() , offset_0 , offset_1 , offset_2 , offset_3, offset_4);
+                        }
+                      }
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+    // LR
+    else {
+      index_type temp0  =  m_rp.m_tile_end[0];
+      index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl1 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl0 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl1 ) :
+          ( temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x / numbl1;
+      const index_type tile_id1 = (index_type)blockIdx.x % numbl1;
+      const index_type thr_id0 = (index_type)threadIdx.x / m_rp.m_tile[1];
+      const index_type thr_id1 = (index_type)threadIdx.x % m_rp.m_tile[1];
+
+      temp0  =  m_rp.m_tile_end[2];
+      temp1  =  m_rp.m_tile_end[3];
+      const index_type numbl3 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl2 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl3 ) :
+          (  temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id2 = (index_type)blockIdx.y / numbl3;
+      const index_type tile_id3 = (index_type)blockIdx.y % numbl3;
+      const index_type thr_id2 = (index_type)threadIdx.y / m_rp.m_tile[3];
+      const index_type thr_id3 = (index_type)threadIdx.y % m_rp.m_tile[3];
+
+      for ( index_type i = tile_id0; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+        const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+        if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+
+          for ( index_type j = tile_id1; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+            const index_type offset_1 = j*m_rp.m_tile[1] + thr_id1;
+            if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+              for ( index_type k = tile_id2; k < m_rp.m_tile_end[2]; k += numbl2 ) {
+                const index_type offset_2 = k*m_rp.m_tile[2] + thr_id2;
+                if ( offset_2 < m_rp.m_upper[2] && thr_id2 < m_rp.m_tile[2] ) {
+
+                  for ( index_type l = tile_id3; l < m_rp.m_tile_end[3]; l += numbl3 ) {
+                    const index_type offset_3 = l*m_rp.m_tile[3] + thr_id3;
+                    if ( offset_3 < m_rp.m_upper[3] && thr_id3 < m_rp.m_tile[3] ) {
+
+                      for ( index_type tile_id4 = (index_type)blockIdx.z; tile_id4 < m_rp.m_tile_end[4]; tile_id4 += gridDim.z ) {
+                        const index_type offset_4 = tile_id4*m_rp.m_tile[4] + (index_type)threadIdx.z;
+                        if ( offset_4 < m_rp.m_upper[4] && (index_type)threadIdx.z < m_rp.m_tile[4] ) {
+                          m_func(Tag() , offset_0 , offset_1 , offset_2 , offset_3 , offset_4);
+                        }
+                      }
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+};
+
+
+//Rank 6
+// Specializations for void tag type
+template< typename RP , typename Functor >
+struct DeviceIterateTile<6,RP,Functor,void >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ )
+  : m_rp(rp_)
+  , m_func(f_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    // LL
+    if (RP::inner_direction == RP::Left) {
+      index_type temp0  =  m_rp.m_tile_end[0];
+      index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl0 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl1 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl0 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x % numbl0;
+      const index_type tile_id1 = (index_type)blockIdx.x / numbl0;
+      const index_type thr_id0 = (index_type)threadIdx.x % m_rp.m_tile[0];
+      const index_type thr_id1 = (index_type)threadIdx.x / m_rp.m_tile[0];
+
+      temp0  =  m_rp.m_tile_end[2];
+      temp1  =  m_rp.m_tile_end[3];
+      const index_type numbl2 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl3 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl2 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id2 = (index_type)blockIdx.y % numbl2;
+      const index_type tile_id3 = (index_type)blockIdx.y / numbl2;
+      const index_type thr_id2 = (index_type)threadIdx.y % m_rp.m_tile[2];
+      const index_type thr_id3 = (index_type)threadIdx.y / m_rp.m_tile[2];
+
+      temp0  =  m_rp.m_tile_end[4];
+      temp1  =  m_rp.m_tile_end[5];
+      const index_type numbl4 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl5 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl4 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id4 = (index_type)blockIdx.z % numbl4;
+      const index_type tile_id5 = (index_type)blockIdx.z / numbl4;
+      const index_type thr_id4 = (index_type)threadIdx.z % m_rp.m_tile[4];
+      const index_type thr_id5 = (index_type)threadIdx.z / m_rp.m_tile[4];
+
+      for ( index_type n = tile_id5; n < m_rp.m_tile_end[5]; n += numbl5 ) {
+        const index_type offset_5 = n*m_rp.m_tile[5] + thr_id5;
+        if ( offset_5 < m_rp.m_upper[5] && thr_id5 < m_rp.m_tile[5] ) {
+
+          for ( index_type m = tile_id4; m < m_rp.m_tile_end[4]; m += numbl4 ) {
+            const index_type offset_4 = m*m_rp.m_tile[4] + thr_id4;
+            if ( offset_4 < m_rp.m_upper[4] && thr_id4 < m_rp.m_tile[4] ) {
+
+              for ( index_type l = tile_id3; l < m_rp.m_tile_end[3]; l += numbl3 ) {
+                const index_type offset_3 = l*m_rp.m_tile[3] + thr_id3;
+                if ( offset_3 < m_rp.m_upper[3] && thr_id3 < m_rp.m_tile[3] ) {
+
+                  for ( index_type k = tile_id2; k < m_rp.m_tile_end[2]; k += numbl2 ) {
+                    const index_type offset_2 = k*m_rp.m_tile[2] + thr_id2;
+                    if ( offset_2 < m_rp.m_upper[2] && thr_id2 < m_rp.m_tile[2] ) {
+
+                      for ( index_type j = tile_id1 ; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+                        const index_type offset_1 = j*m_rp.m_tile[1] + thr_id1;
+                        if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+                          for ( index_type i = tile_id0 ; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+                            const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+                            if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+                              m_func(offset_0 , offset_1 , offset_2 , offset_3, offset_4, offset_5);
+                            }
+                          }
+                        }
+                      }
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+    // LR
+    else {
+      index_type temp0  =  m_rp.m_tile_end[0];
+      index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl1 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl0 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl1 ) :
+          ( temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x / numbl1;
+      const index_type tile_id1 = (index_type)blockIdx.x % numbl1;
+      const index_type thr_id0 = (index_type)threadIdx.x / m_rp.m_tile[1];
+      const index_type thr_id1 = (index_type)threadIdx.x % m_rp.m_tile[1];
+
+      temp0  =  m_rp.m_tile_end[2];
+      temp1  =  m_rp.m_tile_end[3];
+      const index_type numbl3 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl2 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl3 ) :
+          (  temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id2 = (index_type)blockIdx.y / numbl3;
+      const index_type tile_id3 = (index_type)blockIdx.y % numbl3;
+      const index_type thr_id2 = (index_type)threadIdx.y / m_rp.m_tile[3];
+      const index_type thr_id3 = (index_type)threadIdx.y % m_rp.m_tile[3];
+
+      temp0  =  m_rp.m_tile_end[4];
+      temp1  =  m_rp.m_tile_end[5];
+      const index_type numbl5 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl4 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl5 ) :
+          (  temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id4 = (index_type)blockIdx.z / numbl5;
+      const index_type tile_id5 = (index_type)blockIdx.z % numbl5;
+      const index_type thr_id4 = (index_type)threadIdx.z / m_rp.m_tile[5];
+      const index_type thr_id5 = (index_type)threadIdx.z % m_rp.m_tile[5];
+
+      for ( index_type i = tile_id0; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+        const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+        if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+
+          for ( index_type j = tile_id1; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+            const index_type offset_1 = j*m_rp.m_tile[1] + thr_id1;
+            if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+              for ( index_type k = tile_id2; k < m_rp.m_tile_end[2]; k += numbl2 ) {
+                const index_type offset_2 = k*m_rp.m_tile[2] + thr_id2;
+                if ( offset_2 < m_rp.m_upper[2] && thr_id2 < m_rp.m_tile[2] ) {
+
+                  for ( index_type l = tile_id3; l < m_rp.m_tile_end[3]; l += numbl3 ) {
+                    const index_type offset_3 = l*m_rp.m_tile[3] + thr_id3;
+                    if ( offset_3 < m_rp.m_upper[3] && thr_id3 < m_rp.m_tile[3] ) {
+
+                      for ( index_type m = tile_id4; m < m_rp.m_tile_end[4]; m += numbl4 ) {
+                        const index_type offset_4 = m*m_rp.m_tile[4] + thr_id4;
+                        if ( offset_4 < m_rp.m_upper[4] && thr_id4 < m_rp.m_tile[4] ) {
+
+                          for ( index_type n = tile_id5; n < m_rp.m_tile_end[5]; n += numbl5 ) {
+                            const index_type offset_5 = n*m_rp.m_tile[5] + thr_id5;
+                            if ( offset_5 < m_rp.m_upper[5] && thr_id5 < m_rp.m_tile[5] ) {
+                              m_func(offset_0 , offset_1 , offset_2 , offset_3 , offset_4 , offset_5);
+                            }
+                          }
+                        }
+                      }
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+};
+
+// Specializations for tag type
+template< typename RP , typename Functor , typename Tag >
+struct DeviceIterateTile<6,RP,Functor,Tag>
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ )
+  : m_rp(rp_)
+  , m_func(f_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    // LL
+    if (RP::inner_direction == RP::Left) {
+      index_type temp0  =  m_rp.m_tile_end[0];
+      index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl0 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl1 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl0 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x % numbl0;
+      const index_type tile_id1 = (index_type)blockIdx.x / numbl0;
+      const index_type thr_id0 = (index_type)threadIdx.x % m_rp.m_tile[0];
+      const index_type thr_id1 = (index_type)threadIdx.x / m_rp.m_tile[0];
+
+      temp0  =  m_rp.m_tile_end[2];
+      temp1  =  m_rp.m_tile_end[3];
+      const index_type numbl2 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl3 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl2 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id2 = (index_type)blockIdx.y % numbl2;
+      const index_type tile_id3 = (index_type)blockIdx.y / numbl2;
+      const index_type thr_id2 = (index_type)threadIdx.y % m_rp.m_tile[2];
+      const index_type thr_id3 = (index_type)threadIdx.y / m_rp.m_tile[2];
+
+      temp0  =  m_rp.m_tile_end[4];
+      temp1  =  m_rp.m_tile_end[5];
+      const index_type numbl4 = ( temp0 <= max_blocks ? temp0 : max_blocks ) ;
+      const index_type numbl5 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl4 ) :
+          (  temp1 <= max_blocks ? temp1 : max_blocks ) );
+
+      const index_type tile_id4 = (index_type)blockIdx.z % numbl4;
+      const index_type tile_id5 = (index_type)blockIdx.z / numbl4;
+      const index_type thr_id4 = (index_type)threadIdx.z % m_rp.m_tile[4];
+      const index_type thr_id5 = (index_type)threadIdx.z / m_rp.m_tile[4];
+
+      for ( index_type n = tile_id5; n < m_rp.m_tile_end[5]; n += numbl5 ) {
+        const index_type offset_5 = n*m_rp.m_tile[5] + thr_id5;
+        if ( offset_5 < m_rp.m_upper[5] && thr_id5 < m_rp.m_tile[5] ) {
+
+          for ( index_type m = tile_id4; m < m_rp.m_tile_end[4]; m += numbl4 ) {
+            const index_type offset_4 = m*m_rp.m_tile[4] + thr_id4;
+            if ( offset_4 < m_rp.m_upper[4] && thr_id4 < m_rp.m_tile[4] ) {
+
+              for ( index_type l = tile_id3; l < m_rp.m_tile_end[3]; l += numbl3 ) {
+                const index_type offset_3 = l*m_rp.m_tile[3] + thr_id3;
+                if ( offset_3 < m_rp.m_upper[3] && thr_id3 < m_rp.m_tile[3] ) {
+
+                  for ( index_type k = tile_id2; k < m_rp.m_tile_end[2]; k += numbl2 ) {
+                    const index_type offset_2 = k*m_rp.m_tile[2] + thr_id2;
+                    if ( offset_2 < m_rp.m_upper[2] && thr_id2 < m_rp.m_tile[2] ) {
+
+                      for ( index_type j = tile_id1 ; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+                        const index_type offset_1 = j*m_rp.m_tile[1] + thr_id1;
+                        if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+                          for ( index_type i = tile_id0 ; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+                            const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+                            if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+                              m_func(Tag() , offset_0 , offset_1 , offset_2 , offset_3, offset_4, offset_5);
+                            }
+                          }
+                        }
+                      }
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+    // LR
+    else {
+      index_type temp0  =  m_rp.m_tile_end[0];
+      index_type temp1  =  m_rp.m_tile_end[1];
+      const index_type numbl1 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl0 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl1 ) :
+          ( temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id0 = (index_type)blockIdx.x / numbl1;
+      const index_type tile_id1 = (index_type)blockIdx.x % numbl1;
+      const index_type thr_id0 = (index_type)threadIdx.x / m_rp.m_tile[1];
+      const index_type thr_id1 = (index_type)threadIdx.x % m_rp.m_tile[1];
+
+      temp0  =  m_rp.m_tile_end[2];
+      temp1  =  m_rp.m_tile_end[3];
+      const index_type numbl3 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl2 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl3 ) :
+          (  temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id2 = (index_type)blockIdx.y / numbl3;
+      const index_type tile_id3 = (index_type)blockIdx.y % numbl3;
+      const index_type thr_id2 = (index_type)threadIdx.y / m_rp.m_tile[3];
+      const index_type thr_id3 = (index_type)threadIdx.y % m_rp.m_tile[3];
+
+      temp0  =  m_rp.m_tile_end[4];
+      temp1  =  m_rp.m_tile_end[5];
+      const index_type numbl5 = ( temp1 <= max_blocks ? temp1 : max_blocks ) ;
+      const index_type numbl4 = ( temp0*temp1 > max_blocks ? index_type( max_blocks / numbl5 ) :
+          (  temp0 <= max_blocks ? temp0 : max_blocks ) );
+
+      const index_type tile_id4 = (index_type)blockIdx.z / numbl5;
+      const index_type tile_id5 = (index_type)blockIdx.z % numbl5;
+      const index_type thr_id4 = (index_type)threadIdx.z / m_rp.m_tile[5];
+      const index_type thr_id5 = (index_type)threadIdx.z % m_rp.m_tile[5];
+
+      for ( index_type i = tile_id0; i < m_rp.m_tile_end[0]; i += numbl0 ) {
+        const index_type offset_0 = i*m_rp.m_tile[0] + thr_id0;
+        if ( offset_0 < m_rp.m_upper[0] && thr_id0 < m_rp.m_tile[0] ) {
+
+          for ( index_type j = tile_id1; j < m_rp.m_tile_end[1]; j += numbl1 ) {
+            const index_type offset_1 = j*m_rp.m_tile[1] + thr_id1;
+            if ( offset_1 < m_rp.m_upper[1] && thr_id1 < m_rp.m_tile[1] ) {
+
+              for ( index_type k = tile_id2; k < m_rp.m_tile_end[2]; k += numbl2 ) {
+                const index_type offset_2 = k*m_rp.m_tile[2] + thr_id2;
+                if ( offset_2 < m_rp.m_upper[2] && thr_id2 < m_rp.m_tile[2] ) {
+
+                  for ( index_type l = tile_id3; l < m_rp.m_tile_end[3]; l += numbl3 ) {
+                    const index_type offset_3 = l*m_rp.m_tile[3] + thr_id3;
+                    if ( offset_3 < m_rp.m_upper[3] && thr_id3 < m_rp.m_tile[3] ) {
+
+                      for ( index_type m = tile_id4; m < m_rp.m_tile_end[4]; m += numbl4 ) {
+                        const index_type offset_4 = m*m_rp.m_tile[4] + thr_id4;
+                        if ( offset_4 < m_rp.m_upper[4] && thr_id4 < m_rp.m_tile[4] ) {
+
+                          for ( index_type n = tile_id5; n < m_rp.m_tile_end[5]; n += numbl5 ) {
+                            const index_type offset_5 = n*m_rp.m_tile[5] + thr_id5;
+                            if ( offset_5 < m_rp.m_upper[5] && thr_id5 < m_rp.m_tile[5] ) {
+                              m_func(Tag() , offset_0 , offset_1 , offset_2 , offset_3 , offset_4 , offset_5);
+                            }
+                          }
+                        }
+                      }
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+};
+
+} // Refactor
+
+// ----------------------------------------------------------------------------------
+
+namespace Reduce {
+
+template < typename T >
+using is_void = std::is_same< T, void >;
+
+template < typename T >
+struct is_array_type : std::false_type
+{
+  using value_type = T;
+};
+
+template < typename T >
+struct is_array_type< T* > : std::true_type
+{
+  using value_type = T;
+};
+
+template < typename T >
+struct is_array_type< T[] > : std::true_type
+{
+  using value_type = T;
+};
+
+// ------------------------------------------------------------------ //
+template< int N , typename RP , typename Functor , typename Tag , typename ValueType , typename Enable = void >
+struct DeviceIterateTile;
+
+// ParallelReduce iteration pattern
+// Scalar reductions
+
+// num_blocks = min( num_tiles, max_num_blocks ); //i.e. determined by number of tiles and reduction algorithm constraints
+// extract n-dim tile offsets (i.e. tile's global starting mulit-index) from the tileid = blockid using tile dimensions
+// local indices within a tile extracted from (index_type)threadIdx.x using tile dims, constrained by blocksize
+// combine tile and local id info for multi-dim global ids
+
+// Pattern:
+// Each block+thread is responsible for a tile+local_id combo (additional when striding by num_blocks)
+// 1. create offset arrays
+// 2. loop over number of tiles, striding by griddim (equal to num tiles, or max num blocks)
+// 3. temps set for tile_idx and thrd_idx, which will be modified
+// 4. if LL vs LR:
+//      determine tile starting point offsets (multidim)
+//      determine local index offsets (multidim)
+//      concatentate tile offset + local offset for global multi-dim index
+//    if offset withinin range bounds AND local offset within tile bounds, call functor
+
+// ValueType = T
+//Rank 2
+// Specializations for void tag type
+template< typename RP , typename Functor , typename ValueType >
+struct DeviceIterateTile<2,RP,Functor,void,ValueType, typename std::enable_if< !is_array_type::value >::type >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , ValueType & v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( m_offset[0], m_offset[1], m_v ); }
+        }
+      }
+    }
+
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  ValueType & m_v;
+};
+
+
+// Specializations for tag type
+template< typename RP , typename Functor , typename Tag, typename ValueType >
+struct DeviceIterateTile<2,RP,Functor,Tag, ValueType, typename std::enable_if< !is_array_type::value && !is_void< Tag >::value >::type >
+{
+  using index_type = typename RP::index_type;
+
+  inline __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , ValueType & v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]); // Move this to first computation, add to m_offset right away
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( Tag(), m_offset[0], m_offset[1], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  ValueType & m_v;
+};
+
+
+//Rank 3
+// Specializations for void tag type
+template< typename RP , typename Functor , typename ValueType >
+struct DeviceIterateTile<3,RP,Functor,void,ValueType , typename std::enable_if< !is_array_type::value >::type >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , ValueType & v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]); // Move this to first computation, add to m_offset right away
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( m_offset[0], m_offset[1], m_offset[2], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  ValueType & m_v;
+};
+
+
+// Specializations for void tag type
+template< typename RP , typename Functor , typename Tag, typename ValueType >
+struct DeviceIterateTile<3,RP,Functor,Tag, ValueType, typename std::enable_if< !is_array_type::value && !is_void< Tag >::value >::type >
+{
+  using index_type = typename RP::index_type;
+
+  inline __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , ValueType & v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]); // Move this to first computation, add to m_offset right away
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( Tag(), m_offset[0], m_offset[1], m_offset[2], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  ValueType & m_v;
+};
+
+
+//Rank 4
+// Specializations for void tag type
+template< typename RP , typename Functor , typename ValueType >
+struct DeviceIterateTile<4,RP,Functor,void,ValueType , typename std::enable_if< !is_array_type::value >::type >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , ValueType & v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  ValueType & m_v;
+};
+
+
+// Specializations for void tag type
+template< typename RP , typename Functor , typename Tag, typename ValueType >
+struct DeviceIterateTile<4,RP,Functor,Tag,ValueType, typename std::enable_if< !is_array_type::value && !is_void< Tag >::value >::type >
+{
+  using index_type = typename RP::index_type;
+
+  inline __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , ValueType & v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( Tag(), m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  ValueType & m_v;
+};
+
+
+//Rank 5
+// Specializations for void tag type
+template< typename RP , typename Functor , typename ValueType >
+struct DeviceIterateTile<5,RP,Functor,void,ValueType , typename std::enable_if< !is_array_type::value >::type >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , ValueType & v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_offset[4], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  ValueType & m_v;
+};
+
+
+// Specializations for tag type
+template< typename RP , typename Functor , typename Tag, typename ValueType >
+struct DeviceIterateTile<5,RP,Functor,Tag,ValueType, typename std::enable_if< !is_array_type::value && !is_void< Tag >::value >::type >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , ValueType & v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( Tag(), m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_offset[4], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  ValueType & m_v;
+};
+
+
+//Rank 6
+// Specializations for void tag type
+template< typename RP , typename Functor , typename ValueType >
+struct DeviceIterateTile<6,RP,Functor,void,ValueType , typename std::enable_if< !is_array_type::value >::type >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , ValueType & v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_offset[4], m_offset[5], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  ValueType & m_v;
+};
+
+
+// Specializations for tag type
+template< typename RP , typename Functor , typename Tag, typename ValueType >
+struct DeviceIterateTile<6,RP,Functor,Tag,ValueType, typename std::enable_if< !is_array_type::value && !is_void< Tag >::value >::type >
+{
+  using index_type = typename RP::index_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , ValueType & v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( Tag(), m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_offset[4], m_offset[5], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  ValueType & m_v;
+};
+
+
+// ValueType = T[], T*
+//Rank 2
+// Specializations for void tag type
+template< typename RP , typename Functor , typename ValueType >
+struct DeviceIterateTile<2,RP,Functor,void,ValueType, typename std::enable_if< is_array_type::value >::type >
+{
+  using index_type = typename RP::index_type;
+  using value_type = typename is_array_type< ValueType >::value_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , value_type* v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]); // Move this to first computation, add to m_offset right away
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( m_offset[0], m_offset[1], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  value_type* m_v;
+};
+
+
+// Specializations for tag type
+template< typename RP , typename Functor , typename Tag, typename ValueType >
+struct DeviceIterateTile<2,RP,Functor,Tag, ValueType, typename std::enable_if< is_array_type::value && !is_void< Tag >::value >::type >
+{
+  using index_type = typename RP::index_type;
+  using value_type = typename is_array_type< ValueType >::value_type;
+
+  inline __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , value_type* v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( Tag(), m_offset[0], m_offset[1], m_v ); }
+        }
+      } //end for loop over num_tiles - product of tiles in each direction
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  value_type* m_v;
+};
+
+
+//Rank 3
+// Specializations for void tag type
+template< typename RP , typename Functor , typename ValueType >
+struct DeviceIterateTile<3,RP,Functor,void,ValueType , typename std::enable_if< is_array_type::value >::type >
+{
+  using index_type = typename RP::index_type;
+  using value_type = typename is_array_type< ValueType >::value_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , value_type* v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]); // Move this to first computation, add to m_offset right away
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( m_offset[0], m_offset[1], m_offset[2], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  value_type* m_v;
+};
+
+
+// Specializations for void tag type
+template< typename RP , typename Functor , typename Tag, typename ValueType >
+struct DeviceIterateTile<3,RP,Functor,Tag, ValueType, typename std::enable_if< is_array_type::value && !is_void< Tag >::value >::type >
+{
+  using index_type = typename RP::index_type;
+  using value_type = typename is_array_type< ValueType >::value_type;
+
+  inline __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , value_type* v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  inline __device__
+  void exec_range() const
+  {
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( Tag(), m_offset[0], m_offset[1], m_offset[2], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  value_type* m_v;
+};
+
+
+//Rank 4
+// Specializations for void tag type
+template< typename RP , typename Functor , typename ValueType >
+struct DeviceIterateTile<4,RP,Functor,void,ValueType , typename std::enable_if< is_array_type::value >::type >
+{
+  using index_type = typename RP::index_type;
+  using value_type = typename is_array_type< ValueType >::value_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , value_type* v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  value_type* m_v;
+};
+
+
+// Specializations for void tag type
+template< typename RP , typename Functor , typename Tag, typename ValueType >
+struct DeviceIterateTile<4,RP,Functor,Tag,ValueType, typename std::enable_if< is_array_type::value && !is_void< Tag >::value >::type >
+{
+  using index_type = typename RP::index_type;
+  using value_type = typename is_array_type< ValueType >::value_type;
+
+  inline __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , value_type* v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( Tag(), m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  value_type* m_v;
+};
+
+
+//Rank 5
+// Specializations for void tag type
+template< typename RP , typename Functor , typename ValueType >
+struct DeviceIterateTile<5,RP,Functor,void,ValueType , typename std::enable_if< is_array_type::value >::type >
+{
+  using index_type = typename RP::index_type;
+  using value_type = typename is_array_type< ValueType >::value_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , value_type* v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_offset[4], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  value_type* m_v;
+};
+
+
+// Specializations for tag type
+template< typename RP , typename Functor , typename Tag, typename ValueType >
+struct DeviceIterateTile<5,RP,Functor,Tag,ValueType, typename std::enable_if< is_array_type::value && !is_void< Tag >::value >::type >
+{
+  using index_type = typename RP::index_type;
+  using value_type = typename is_array_type< ValueType >::value_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , value_type* v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( Tag(), m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_offset[4], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  value_type* m_v;
+};
+
+
+//Rank 6
+// Specializations for void tag type
+template< typename RP , typename Functor , typename ValueType >
+struct DeviceIterateTile<6,RP,Functor,void,ValueType , typename std::enable_if< is_array_type::value >::type >
+{
+  using index_type = typename RP::index_type;
+  using value_type = typename is_array_type< ValueType >::value_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , value_type* v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_offset[4], m_offset[5], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  value_type* m_v;
+};
+
+
+// Specializations for tag type
+template< typename RP , typename Functor , typename Tag, typename ValueType >
+struct DeviceIterateTile<6,RP,Functor,Tag,ValueType, typename std::enable_if< is_array_type::value && !is_void< Tag >::value >::type >
+{
+  using index_type = typename RP::index_type;
+  using value_type = typename is_array_type< ValueType >::value_type;
+
+  __device__
+  DeviceIterateTile( const RP & rp_ , const Functor & f_ , value_type* v_)
+  : m_rp(rp_)
+  , m_func(f_)
+  , m_v(v_)
+  {}
+
+  static constexpr index_type max_blocks = 65535;
+  //static constexpr index_type max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+
+  inline __device__
+  void exec_range() const
+  {
+    //enum { max_blocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount) };
+    //const index_type max_blocks = static_cast( Kokkos::Impl::cuda_internal_maximum_grid_count() );
+    if ( (index_type)blockIdx.x < m_rp.m_num_tiles && (index_type)threadIdx.y < m_rp.m_prod_tile_dims ) {
+      index_type m_offset[RP::rank]; // tile starting global id offset
+      index_type m_local_offset[RP::rank]; // tile starting global id offset
+
+      for ( index_type tileidx = (index_type)blockIdx.x; tileidx < m_rp.m_num_tiles; tileidx += gridDim.x ) {
+        index_type tile_idx = tileidx; // temp because tile_idx will be modified while determining tile starting point offsets
+        index_type thrd_idx = (index_type)threadIdx.y;
+        bool in_bounds = true;
+
+        // LL
+        if (RP::inner_direction == RP::Left) {
+          for (int i=0; i=0; --i) {
+            m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+            tile_idx /= m_rp.m_tile_end[i];
+
+            // tile-local indices identified with (index_type)threadIdx.y
+            m_local_offset[i] = (thrd_idx % m_rp.m_tile[i]);
+            thrd_idx /= m_rp.m_tile[i];
+
+            m_offset[i] += m_local_offset[i];
+            if ( !(m_offset[i] < m_rp.m_upper[i] && m_local_offset[i] < m_rp.m_tile[i]) ) {
+              in_bounds &= false;
+            }
+          }
+          if ( in_bounds )
+          { m_func( Tag(), m_offset[0], m_offset[1], m_offset[2], m_offset[3], m_offset[4], m_offset[5], m_v ); }
+        }
+      }
+    }
+  } //end exec_range
+
+private:
+  const RP & m_rp;
+  const Functor & m_func;
+  value_type* m_v;
+};
+
+} // Reduce
+
+// ----------------------------------------------------------------------------------
+
+} } } //end namespace Kokkos::Experimental::Impl
+
+#endif
+#endif
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_CudaExec.hpp b/lib/kokkos/core/src/Cuda/Kokkos_CudaExec.hpp
index 13abcfd93c..cae8ecd489 100644
--- a/lib/kokkos/core/src/Cuda/Kokkos_CudaExec.hpp
+++ b/lib/kokkos/core/src/Cuda/Kokkos_CudaExec.hpp
@@ -53,6 +53,7 @@
 #include 
 #include 
 #include 
+#include 
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
@@ -125,53 +126,12 @@ unsigned long kokkos_impl_cuda_constant_memory_buffer[ Kokkos::Impl::CudaTraits:
 
 #endif
 
-
-namespace Kokkos {
-namespace Impl {
-  struct CudaLockArraysStruct {
-    int* atomic;
-    int* scratch;
-    int* threadid;
-    int n;
-  };
-}
-}
-__device__ __constant__
-#ifdef KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE
-extern
-#endif
-Kokkos::Impl::CudaLockArraysStruct kokkos_impl_cuda_lock_arrays ;
-
-#define CUDA_SPACE_ATOMIC_MASK 0x1FFFF
-#define CUDA_SPACE_ATOMIC_XOR_MASK 0x15A39
-
 namespace Kokkos {
 namespace Impl {
   void* cuda_resize_scratch_space(std::int64_t bytes, bool force_shrink = false);
 }
 }
 
-namespace Kokkos {
-namespace Impl {
-__device__ inline
-bool lock_address_cuda_space(void* ptr) {
-  size_t offset = size_t(ptr);
-  offset = offset >> 2;
-  offset = offset & CUDA_SPACE_ATOMIC_MASK;
-  return (0 == atomicCAS(&kokkos_impl_cuda_lock_arrays.atomic[offset],0,1));
-}
-
-__device__ inline
-void unlock_address_cuda_space(void* ptr) {
-  size_t offset = size_t(ptr);
-  offset = offset >> 2;
-  offset = offset & CUDA_SPACE_ATOMIC_MASK;
-  atomicExch( &kokkos_impl_cuda_lock_arrays.atomic[ offset ], 0);
-}
-
-}
-}
-
 template< typename T >
 inline
 __device__
@@ -192,7 +152,7 @@ namespace Impl {
 // For 2.0 capability: 48 KB L1 and 16 KB shared
 //----------------------------------------------------------------------------
 
-template< class DriverType >
+template< class DriverType>
 __global__
 static void cuda_parallel_launch_constant_memory()
 {
@@ -202,19 +162,39 @@ static void cuda_parallel_launch_constant_memory()
   driver();
 }
 
-template< class DriverType >
+template< class DriverType, unsigned int maxTperB, unsigned int minBperSM >
+__global__
+__launch_bounds__(maxTperB, minBperSM)
+static void cuda_parallel_launch_constant_memory()
+{
+  const DriverType & driver =
+    *((const DriverType *) kokkos_impl_cuda_constant_memory_buffer );
+
+  driver();
+}
+
+template< class DriverType>
 __global__
 static void cuda_parallel_launch_local_memory( const DriverType driver )
 {
   driver();
 }
 
-template < class DriverType ,
-           bool Large = ( CudaTraits::ConstantMemoryUseThreshold < sizeof(DriverType) ) >
+template< class DriverType, unsigned int maxTperB, unsigned int minBperSM >
+__global__
+__launch_bounds__(maxTperB, minBperSM)
+static void cuda_parallel_launch_local_memory( const DriverType driver )
+{
+  driver();
+}
+
+template < class DriverType
+         , class LaunchBounds = Kokkos::LaunchBounds<>
+         , bool Large = ( CudaTraits::ConstantMemoryUseThreshold < sizeof(DriverType) ) >
 struct CudaParallelLaunch ;
 
-template < class DriverType >
-struct CudaParallelLaunch< DriverType , true > {
+template < class DriverType, class LaunchBounds >
+struct CudaParallelLaunch< DriverType, LaunchBounds, true > {
 
   inline
   CudaParallelLaunch( const DriverType & driver
@@ -238,26 +218,19 @@ struct CudaParallelLaunch< DriverType , true > {
       }
       #ifndef KOKKOS_ARCH_KEPLER //On Kepler the L1 has no benefit since it doesn't cache reads
       else if ( shmem ) {
-        CUDA_SAFE_CALL( cudaFuncSetCacheConfig( cuda_parallel_launch_constant_memory< DriverType > , cudaFuncCachePreferShared ) );
+        CUDA_SAFE_CALL( cudaFuncSetCacheConfig( cuda_parallel_launch_constant_memory< DriverType, LaunchBounds::maxTperB, LaunchBounds::minBperSM > , cudaFuncCachePreferShared ) );
       } else {
-        CUDA_SAFE_CALL( cudaFuncSetCacheConfig( cuda_parallel_launch_constant_memory< DriverType > , cudaFuncCachePreferL1 ) );
+        CUDA_SAFE_CALL( cudaFuncSetCacheConfig( cuda_parallel_launch_constant_memory< DriverType, LaunchBounds::maxTperB, LaunchBounds::minBperSM > , cudaFuncCachePreferL1 ) );
       }
       #endif
 
       // Copy functor to constant memory on the device
       cudaMemcpyToSymbol( kokkos_impl_cuda_constant_memory_buffer , & driver , sizeof(DriverType) );
 
-      #ifndef KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE
-      Kokkos::Impl::CudaLockArraysStruct locks;
-      locks.atomic = atomic_lock_array_cuda_space_ptr(false);
-      locks.scratch = scratch_lock_array_cuda_space_ptr(false);
-      locks.threadid = threadid_lock_array_cuda_space_ptr(false);
-      locks.n = Kokkos::Cuda::concurrency();
-      cudaMemcpyToSymbol( kokkos_impl_cuda_lock_arrays , & locks , sizeof(CudaLockArraysStruct) );
-      #endif
+      KOKKOS_ENSURE_CUDA_LOCK_ARRAYS_ON_DEVICE();
 
       // Invoke the driver function on the device
-      cuda_parallel_launch_constant_memory< DriverType ><<< grid , block , shmem , stream >>>();
+      cuda_parallel_launch_constant_memory< DriverType, LaunchBounds::maxTperB, LaunchBounds::minBperSM ><<< grid , block , shmem , stream >>>();
 
 #if defined( KOKKOS_ENABLE_DEBUG_BOUNDS_CHECK )
       CUDA_SAFE_CALL( cudaGetLastError() );
@@ -267,8 +240,8 @@ struct CudaParallelLaunch< DriverType , true > {
   }
 };
 
-template < class DriverType >
-struct CudaParallelLaunch< DriverType , false > {
+template < class DriverType, class LaunchBounds >
+struct CudaParallelLaunch< DriverType, LaunchBounds, false > {
 
   inline
   CudaParallelLaunch( const DriverType & driver
@@ -284,22 +257,15 @@ struct CudaParallelLaunch< DriverType , false > {
       }
       #ifndef KOKKOS_ARCH_KEPLER //On Kepler the L1 has no benefit since it doesn't cache reads
       else if ( shmem ) {
-        CUDA_SAFE_CALL( cudaFuncSetCacheConfig( cuda_parallel_launch_local_memory< DriverType > , cudaFuncCachePreferShared ) );
+        CUDA_SAFE_CALL( cudaFuncSetCacheConfig( cuda_parallel_launch_local_memory< DriverType, LaunchBounds::maxTperB, LaunchBounds::minBperSM > , cudaFuncCachePreferShared ) );
       } else {
-        CUDA_SAFE_CALL( cudaFuncSetCacheConfig( cuda_parallel_launch_local_memory< DriverType > , cudaFuncCachePreferL1 ) );
+        CUDA_SAFE_CALL( cudaFuncSetCacheConfig( cuda_parallel_launch_local_memory< DriverType, LaunchBounds::maxTperB, LaunchBounds::minBperSM > , cudaFuncCachePreferL1 ) );
       }
       #endif
 
-      #ifndef KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE
-      Kokkos::Impl::CudaLockArraysStruct locks;
-      locks.atomic = atomic_lock_array_cuda_space_ptr(false);
-      locks.scratch = scratch_lock_array_cuda_space_ptr(false);
-      locks.threadid = threadid_lock_array_cuda_space_ptr(false);
-      locks.n = Kokkos::Cuda::concurrency();
-      cudaMemcpyToSymbol( kokkos_impl_cuda_lock_arrays , & locks , sizeof(CudaLockArraysStruct) );
-      #endif
+      KOKKOS_ENSURE_CUDA_LOCK_ARRAYS_ON_DEVICE();
 
-      cuda_parallel_launch_local_memory< DriverType ><<< grid , block , shmem , stream >>>( driver );
+      cuda_parallel_launch_local_memory< DriverType, LaunchBounds::maxTperB, LaunchBounds::minBperSM ><<< grid , block , shmem , stream >>>( driver );
 
 #if defined( KOKKOS_ENABLE_DEBUG_BOUNDS_CHECK )
       CUDA_SAFE_CALL( cudaGetLastError() );
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_CudaSpace.cpp b/lib/kokkos/core/src/Cuda/Kokkos_CudaSpace.cpp
index 406b4f1e22..b699f0d6ba 100644
--- a/lib/kokkos/core/src/Cuda/Kokkos_CudaSpace.cpp
+++ b/lib/kokkos/core/src/Cuda/Kokkos_CudaSpace.cpp
@@ -230,18 +230,6 @@ void CudaHostPinnedSpace::deallocate( void * const arg_alloc_ptr , const size_t
   } catch(...) {}
 }
 
-constexpr const char* CudaSpace::name() {
-  return m_name;
-}
-
-constexpr const char* CudaUVMSpace::name() {
-  return m_name;
-}
-
-constexpr const char* CudaHostPinnedSpace::name() {
-  return m_name;
-}
-
 } // namespace Kokkos
 
 //----------------------------------------------------------------------------
@@ -655,11 +643,12 @@ reallocate_tracked( void * const arg_alloc_ptr
 SharedAllocationRecord< Kokkos::CudaSpace , void > *
 SharedAllocationRecord< Kokkos::CudaSpace , void >::get_record( void * alloc_ptr )
 {
-  using Header     = SharedAllocationHeader ;
   using RecordBase = SharedAllocationRecord< void , void > ;
   using RecordCuda = SharedAllocationRecord< Kokkos::CudaSpace , void > ;
 
 #if 0
+  using Header     = SharedAllocationHeader ;
+
   // Copy the header from the allocation
   Header head ;
 
@@ -812,83 +801,6 @@ print_records( std::ostream & s , const Kokkos::CudaHostPinnedSpace & space , bo
   SharedAllocationRecord< void , void >::print_host_accessible_records( s , "CudaHostPinned" , & s_root_record , detail );
 }
 
-} // namespace Impl
-} // namespace Kokkos
-
-/*--------------------------------------------------------------------------*/
-/*--------------------------------------------------------------------------*/
-
-namespace Kokkos {
-namespace {
-  __global__ void init_lock_array_kernel_atomic() {
-    unsigned i = blockIdx.x*blockDim.x + threadIdx.x;
-
-    if(i>>();
-    init_lock_array_kernel_scratch_threadid<<<(Kokkos::Cuda::concurrency()+255)/256,256>>>(Kokkos::Cuda::concurrency());
-  }
-}
-
 void* cuda_resize_scratch_space(std::int64_t bytes, bool force_shrink) {
   static void* ptr = NULL;
   static std::int64_t current_size = 0;
@@ -908,8 +820,8 @@ void* cuda_resize_scratch_space(std::int64_t bytes, bool force_shrink) {
   return ptr;
 }
 
-}
-}
+} // namespace Impl
+} // namespace Kokkos
 #else
 void KOKKOS_CORE_SRC_CUDA_CUDASPACE_PREVENT_LINK_ERROR() {}
 #endif // KOKKOS_ENABLE_CUDA
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Impl.cpp b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Impl.cpp
index daf55cbd97..80e8f9bd8a 100644
--- a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Impl.cpp
+++ b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Impl.cpp
@@ -51,6 +51,7 @@
 
 #include 
 #include 
+#include 
 #include 
 #include 
 
@@ -69,9 +70,6 @@
 __device__ __constant__
 unsigned long kokkos_impl_cuda_constant_memory_buffer[ Kokkos::Impl::CudaTraits::ConstantMemoryUsage / sizeof(unsigned long) ] ;
 
-__device__ __constant__
-Kokkos::Impl::CudaLockArraysStruct kokkos_impl_cuda_lock_arrays ;
-
 #endif
 
 /*--------------------------------------------------------------------------*/
@@ -103,6 +101,7 @@ int cuda_kernel_arch()
   return arch ;
 }
 
+#ifdef KOKKOS_ENABLE_CUDA_UVM
 bool cuda_launch_blocking()
 {
   const char * env = getenv("CUDA_LAUNCH_BLOCKING");
@@ -111,16 +110,13 @@ bool cuda_launch_blocking()
 
   return atoi(env);
 }
+#endif
 
 }
 
 void cuda_device_synchronize()
 {
-//  static const bool launch_blocking = cuda_launch_blocking();
-
-//  if (!launch_blocking) {
-    CUDA_SAFE_CALL( cudaDeviceSynchronize() );
-//  }
+  CUDA_SAFE_CALL( cudaDeviceSynchronize() );
 }
 
 void cuda_internal_error_throw( cudaError e , const char * name, const char * file, const int line )
@@ -240,6 +236,7 @@ public:
   unsigned    m_maxWarpCount ;
   unsigned    m_maxBlock ;
   unsigned    m_maxSharedWords ;
+  uint32_t    m_maxConcurrency ;
   size_type   m_scratchSpaceCount ;
   size_type   m_scratchFlagsCount ;
   size_type   m_scratchUnifiedCount ;
@@ -248,6 +245,7 @@ public:
   size_type * m_scratchSpace ;
   size_type * m_scratchFlags ;
   size_type * m_scratchUnified ;
+  uint32_t  * m_scratchConcurrentBitset ;
   cudaStream_t * m_stream ;
 
   static int was_initialized;
@@ -274,6 +272,7 @@ public:
     , m_maxWarpCount( 0 )
     , m_maxBlock( 0 )
     , m_maxSharedWords( 0 )
+    , m_maxConcurrency( 0 )
     , m_scratchSpaceCount( 0 )
     , m_scratchFlagsCount( 0 )
     , m_scratchUnifiedCount( 0 )
@@ -282,6 +281,7 @@ public:
     , m_scratchSpace( 0 )
     , m_scratchFlags( 0 )
     , m_scratchUnified( 0 )
+    , m_scratchConcurrentBitset( 0 )
     , m_stream( 0 )
     {}
 
@@ -327,7 +327,8 @@ CudaInternal::~CudaInternal()
   if ( m_stream ||
        m_scratchSpace ||
        m_scratchFlags ||
-       m_scratchUnified ) {
+       m_scratchUnified ||
+       m_scratchConcurrentBitset ) {
     std::cerr << "Kokkos::Cuda ERROR: Failed to call Kokkos::Cuda::finalize()"
               << std::endl ;
     std::cerr.flush();
@@ -339,6 +340,7 @@ CudaInternal::~CudaInternal()
   m_maxWarpCount            = 0 ;
   m_maxBlock                = 0 ;
   m_maxSharedWords          = 0 ;
+  m_maxConcurrency          = 0 ;
   m_scratchSpaceCount       = 0 ;
   m_scratchFlagsCount       = 0 ;
   m_scratchUnifiedCount     = 0 ;
@@ -347,6 +349,7 @@ CudaInternal::~CudaInternal()
   m_scratchSpace            = 0 ;
   m_scratchFlags            = 0 ;
   m_scratchUnified          = 0 ;
+  m_scratchConcurrentBitset = 0 ;
   m_stream                  = 0 ;
 }
 
@@ -485,6 +488,33 @@ void CudaInternal::initialize( int cuda_device_id , int stream_count )
       (void) scratch_space( reduce_block_count * 16 * sizeof(size_type) );
     }
     //----------------------------------
+    // Concurrent bitset for obtaining unique tokens from within
+    // an executing kernel.
+    {
+      const unsigned max_threads_per_sm = 2048 ; // up to capability 7.0
+
+      m_maxConcurrency =
+        max_threads_per_sm * cudaProp.multiProcessorCount ;
+
+      const int32_t buffer_bound =
+         Kokkos::Impl::concurrent_bitset::buffer_bound( m_maxConcurrency );
+
+      // Allocate and initialize uint32_t[ buffer_bound ]
+
+      typedef Kokkos::Experimental::Impl::SharedAllocationRecord< Kokkos::CudaSpace , void > Record ;
+
+      Record * const r = Record::allocate( Kokkos::CudaSpace()
+                                         , "InternalScratchBitset"
+                                         , sizeof(uint32_t) * buffer_bound );
+
+      Record::increment( r );
+
+      m_scratchConcurrentBitset = reinterpret_cast( r->data() );
+
+      CUDA_SAFE_CALL( cudaMemset( m_scratchConcurrentBitset , 0 , sizeof(uint32_t) * buffer_bound ) );
+
+    }
+    //----------------------------------
 
     if ( stream_count ) {
       m_stream = (cudaStream_t*) ::malloc( stream_count * sizeof(cudaStream_t) );
@@ -543,16 +573,7 @@ void CudaInternal::initialize( int cuda_device_id , int stream_count )
   cudaThreadSetCacheConfig(cudaFuncCachePreferShared);
 
   // Init the array for used for arbitrarily sized atomics
-  Impl::init_lock_arrays_cuda_space();
-
-  #ifdef KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE
-  Kokkos::Impl::CudaLockArraysStruct locks;
-  locks.atomic = atomic_lock_array_cuda_space_ptr(false);
-  locks.scratch = scratch_lock_array_cuda_space_ptr(false);
-  locks.threadid = threadid_lock_array_cuda_space_ptr(false);
-  locks.n = Kokkos::Cuda::concurrency();
-  cudaMemcpyToSymbol( kokkos_impl_cuda_lock_arrays , & locks , sizeof(CudaLockArraysStruct) );
-  #endif
+  Impl::initialize_host_cuda_lock_arrays();
 }
 
 //----------------------------------------------------------------------------
@@ -635,9 +656,7 @@ void CudaInternal::finalize()
   was_finalized = 1;
   if ( 0 != m_scratchSpace || 0 != m_scratchFlags ) {
 
-    atomic_lock_array_cuda_space_ptr(true);
-    scratch_lock_array_cuda_space_ptr(true);
-    threadid_lock_array_cuda_space_ptr(true);
+    Impl::finalize_host_cuda_lock_arrays();
 
     if ( m_stream ) {
       for ( size_type i = 1 ; i < m_streamCount ; ++i ) {
@@ -653,6 +672,7 @@ void CudaInternal::finalize()
     RecordCuda::decrement( RecordCuda::get_record( m_scratchFlags ) );
     RecordCuda::decrement( RecordCuda::get_record( m_scratchSpace ) );
     RecordHost::decrement( RecordHost::get_record( m_scratchUnified ) );
+    RecordCuda::decrement( RecordCuda::get_record( m_scratchConcurrentBitset ) );
 
     m_cudaDev             = -1 ;
     m_multiProcCount      = 0 ;
@@ -666,6 +686,7 @@ void CudaInternal::finalize()
     m_scratchSpace        = 0 ;
     m_scratchFlags        = 0 ;
     m_scratchUnified      = 0 ;
+    m_scratchConcurrentBitset = 0 ;
     m_stream              = 0 ;
   }
 }
@@ -713,9 +734,8 @@ namespace Kokkos {
 Cuda::size_type Cuda::detect_device_count()
 { return Impl::CudaInternalDevices::singleton().m_cudaDevCount ; }
 
-int Cuda::concurrency() {
-  return 131072;
-}
+int Cuda::concurrency()
+{ return Impl::CudaInternal::singleton().m_maxConcurrency ; }
 
 int Cuda::is_initialized()
 { return Impl::CudaInternal::singleton().is_initialized(); }
@@ -798,7 +818,22 @@ void Cuda::fence()
 const char* Cuda::name() { return "Cuda"; }
 
 } // namespace Kokkos
+
+namespace Kokkos {
+namespace Experimental {
+
+UniqueToken< Kokkos::Cuda , Kokkos::Experimental::UniqueTokenScope::Global >::
+UniqueToken( Kokkos::Cuda const & )
+  : m_buffer( Kokkos::Impl::CudaInternal::singleton().m_scratchConcurrentBitset )
+  , m_count(  Kokkos::Impl::CudaInternal::singleton().m_maxConcurrency )
+  {}
+
+} // namespace Experimental
+} // namespace Kokkos
+
 #else
+
 void KOKKOS_CORE_SRC_CUDA_IMPL_PREVENT_LINK_ERROR() {}
+
 #endif // KOKKOS_ENABLE_CUDA
 
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Locks.cpp b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Locks.cpp
new file mode 100644
index 0000000000..237022ad23
--- /dev/null
+++ b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Locks.cpp
@@ -0,0 +1,119 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include 
+
+#ifdef KOKKOS_ENABLE_CUDA
+
+#include 
+#include 
+#include 
+
+#ifdef KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE
+namespace Kokkos {
+namespace Impl {
+__device__ __constant__
+CudaLockArrays g_device_cuda_lock_arrays = { nullptr, nullptr, 0 };
+}
+}
+#endif
+
+namespace Kokkos {
+
+namespace {
+
+__global__ void init_lock_array_kernel_atomic() {
+  unsigned i = blockIdx.x*blockDim.x + threadIdx.x;
+  if(i>>();
+  init_lock_array_kernel_threadid<<<(Kokkos::Cuda::concurrency()+255)/256,256>>>(Kokkos::Cuda::concurrency());
+  CUDA_SAFE_CALL(cudaDeviceSynchronize());
+}
+
+void finalize_host_cuda_lock_arrays() {
+  if (g_host_cuda_lock_arrays.atomic == nullptr) return;
+  cudaFree(g_host_cuda_lock_arrays.atomic);
+  g_host_cuda_lock_arrays.atomic = nullptr;
+  cudaFree(g_host_cuda_lock_arrays.scratch);
+  g_host_cuda_lock_arrays.scratch = nullptr;
+  g_host_cuda_lock_arrays.n = 0;
+#ifdef KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE
+  KOKKOS_COPY_CUDA_LOCK_ARRAYS_TO_DEVICE();
+#endif
+}
+
+} // namespace Impl
+
+} // namespace Kokkos
+
+#else
+
+void KOKKOS_CORE_SRC_CUDA_CUDA_LOCKS_PREVENT_LINK_ERROR() {}
+
+#endif
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Locks.hpp b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Locks.hpp
new file mode 100644
index 0000000000..d01f06fb4f
--- /dev/null
+++ b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Locks.hpp
@@ -0,0 +1,166 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_CUDA_LOCKS_HPP
+#define KOKKOS_CUDA_LOCKS_HPP
+
+#include 
+
+#ifdef KOKKOS_ENABLE_CUDA
+
+#include 
+
+#include 
+
+namespace Kokkos {
+namespace Impl {
+
+struct CudaLockArrays {
+  std::int32_t* atomic;
+  std::int32_t* scratch;
+  std::int32_t n;
+};
+
+/// \brief This global variable in Host space is the central definition
+///        of these arrays.
+extern Kokkos::Impl::CudaLockArrays g_host_cuda_lock_arrays ;
+
+/// \brief After this call, the g_host_cuda_lock_arrays variable has
+///        valid, initialized arrays.
+///
+/// This call is idempotent.
+void initialize_host_cuda_lock_arrays();
+
+/// \brief After this call, the g_host_cuda_lock_arrays variable has
+///        all null pointers, and all array memory has been freed.
+///
+/// This call is idempotent.
+void finalize_host_cuda_lock_arrays();
+
+} // namespace Impl
+} // namespace Kokkos
+
+#if defined( __CUDACC__ )
+
+namespace Kokkos {
+namespace Impl {
+
+/// \brief This global variable in CUDA space is what kernels use
+///        to get access to the lock arrays.
+///
+/// When relocatable device code is enabled, there can be one single
+/// instance of this global variable for the entire executable,
+/// whose definition will be in Kokkos_Cuda_Locks.cpp (and whose declaration
+/// here must then be extern.
+/// This one instance will be initialized by initialize_host_cuda_lock_arrays
+/// and need not be modified afterwards.
+///
+/// When relocatable device code is disabled, an instance of this variable
+/// will be created in every translation unit that sees this header file
+/// (we make this clear by marking it static, meaning no other translation
+///  unit can link to it).
+/// Since the Kokkos_Cuda_Locks.cpp translation unit cannot initialize the
+/// instances in other translation units, we must update this CUDA global
+/// variable based on the Host global variable prior to running any kernels
+/// that will use it.
+/// That is the purpose of the KOKKOS_ENSURE_CUDA_LOCK_ARRAYS_ON_DEVICE macro.
+__device__ __constant__
+#ifdef KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE
+extern
+#endif
+Kokkos::Impl::CudaLockArrays g_device_cuda_lock_arrays ;
+
+#define CUDA_SPACE_ATOMIC_MASK 0x1FFFF
+
+/// \brief Aquire a lock for the address
+///
+/// This function tries to aquire the lock for the hash value derived
+/// from the provided ptr. If the lock is successfully aquired the
+/// function returns true. Otherwise it returns false.
+__device__ inline
+bool lock_address_cuda_space(void* ptr) {
+  size_t offset = size_t(ptr);
+  offset = offset >> 2;
+  offset = offset & CUDA_SPACE_ATOMIC_MASK;
+  return (0 == atomicCAS(&Kokkos::Impl::g_device_cuda_lock_arrays.atomic[offset],0,1));
+}
+
+/// \brief Release lock for the address
+///
+/// This function releases the lock for the hash value derived
+/// from the provided ptr. This function should only be called
+/// after previously successfully aquiring a lock with
+/// lock_address.
+__device__ inline
+void unlock_address_cuda_space(void* ptr) {
+  size_t offset = size_t(ptr);
+  offset = offset >> 2;
+  offset = offset & CUDA_SPACE_ATOMIC_MASK;
+  atomicExch( &Kokkos::Impl::g_device_cuda_lock_arrays.atomic[ offset ], 0);
+}
+
+} // namespace Impl
+} // namespace Kokkos
+
+/* Dan Ibanez: it is critical that this code be a macro, so that it will
+   capture the right address for Kokkos::Impl::g_device_cuda_lock_arrays!
+   putting this in an inline function will NOT do the right thing! */
+#define KOKKOS_COPY_CUDA_LOCK_ARRAYS_TO_DEVICE() \
+{ \
+  CUDA_SAFE_CALL(cudaMemcpyToSymbol( \
+        Kokkos::Impl::g_device_cuda_lock_arrays , \
+        & Kokkos::Impl::g_host_cuda_lock_arrays , \
+        sizeof(Kokkos::Impl::CudaLockArrays) ) ); \
+}
+
+#ifdef KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE
+#define KOKKOS_ENSURE_CUDA_LOCK_ARRAYS_ON_DEVICE()
+#else
+#define KOKKOS_ENSURE_CUDA_LOCK_ARRAYS_ON_DEVICE() KOKKOS_COPY_CUDA_LOCK_ARRAYS_TO_DEVICE()
+#endif
+
+#endif /* defined( __CUDACC__ ) */
+
+#endif /* defined( KOKKOS_ENABLE_CUDA ) */
+
+#endif /* #ifndef KOKKOS_CUDA_LOCKS_HPP */
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Parallel.hpp b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Parallel.hpp
index 0c8c700e8f..e2eab19e45 100644
--- a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Parallel.hpp
+++ b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Parallel.hpp
@@ -58,6 +58,7 @@
 #include 
 #include 
 #include 
+#include 
 #include 
 
 #if defined(KOKKOS_ENABLE_PROFILING)
@@ -65,6 +66,8 @@
 #include 
 #endif
 
+#include 
+
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
@@ -318,6 +321,7 @@ private:
   typedef Kokkos::RangePolicy< Traits ... > Policy;
   typedef typename Policy::member_type  Member ;
   typedef typename Policy::work_tag     WorkTag ;
+  typedef typename Policy::launch_bounds LaunchBounds ;
 
   const FunctorType  m_functor ;
   const Policy       m_policy ;
@@ -363,7 +367,7 @@ public:
       const dim3 block(  1 , CudaTraits::WarpSize * cuda_internal_maximum_warp_count(), 1);
       const dim3 grid( std::min( ( nwork + block.y - 1 ) / block.y , cuda_internal_maximum_grid_count() ) , 1 , 1);
 
-      CudaParallelLaunch< ParallelFor >( *this , grid , block , 0 );
+      CudaParallelLaunch< ParallelFor, LaunchBounds >( *this , grid , block , 0 );
     }
 
   ParallelFor( const FunctorType  & arg_functor ,
@@ -373,6 +377,115 @@ public:
     { }
 };
 
+
+// MDRangePolicy impl
+template< class FunctorType , class ... Traits >
+class ParallelFor< FunctorType
+                 , Kokkos::Experimental::MDRangePolicy< Traits ... >
+                 , Kokkos::Cuda
+                 >
+{
+private:
+  typedef Kokkos::Experimental::MDRangePolicy< Traits ...  > Policy ;
+  using RP = Policy;
+  typedef typename Policy::array_index_type array_index_type;
+  typedef typename Policy::index_type index_type;
+  typedef typename Policy::launch_bounds LaunchBounds;
+
+
+  const FunctorType m_functor ;
+  const Policy      m_rp ;
+
+public:
+
+  inline
+  __device__
+  void operator()(void) const
+    {
+      Kokkos::Experimental::Impl::Refactor::DeviceIterateTile(m_rp,m_functor).exec_range();
+    }
+
+
+  inline
+  void execute() const
+  {
+    const array_index_type maxblocks = static_cast(Kokkos::Impl::CudaTraits::UpperBoundGridCount);
+    if ( RP::rank == 2 )
+    {
+      const dim3 block( m_rp.m_tile[0] , m_rp.m_tile[1] , 1);
+      const dim3 grid(
+            std::min( ( m_rp.m_upper[0] - m_rp.m_lower[0] + block.x - 1 ) / block.x , maxblocks )
+          , std::min( ( m_rp.m_upper[1] - m_rp.m_lower[1] + block.y - 1 ) / block.y , maxblocks )
+          , 1
+          );
+      CudaParallelLaunch< ParallelFor, LaunchBounds >( *this , grid , block , 0 );
+    }
+    else if ( RP::rank == 3 )
+    {
+      const dim3 block( m_rp.m_tile[0] , m_rp.m_tile[1] , m_rp.m_tile[2] );
+      const dim3 grid(
+          std::min( ( m_rp.m_upper[0] - m_rp.m_lower[0] + block.x - 1 ) / block.x , maxblocks )
+        , std::min( ( m_rp.m_upper[1] - m_rp.m_lower[1] + block.y - 1 ) / block.y , maxblocks )
+        , std::min( ( m_rp.m_upper[2] - m_rp.m_lower[2] + block.z - 1 ) / block.z , maxblocks )
+        );
+      CudaParallelLaunch< ParallelFor, LaunchBounds >( *this , grid , block , 0 );
+    }
+    else if ( RP::rank == 4 )
+    {
+      // id0,id1 encoded within threadIdx.x; id2 to threadIdx.y; id3 to threadIdx.z
+      const dim3 block( m_rp.m_tile[0]*m_rp.m_tile[1] , m_rp.m_tile[2] , m_rp.m_tile[3] );
+      const dim3 grid(
+          std::min( static_cast( m_rp.m_tile_end[0] * m_rp.m_tile_end[1] )
+                  , static_cast(maxblocks) )
+        , std::min( ( m_rp.m_upper[2] - m_rp.m_lower[2] + block.y - 1 ) / block.y , maxblocks )
+        , std::min( ( m_rp.m_upper[3] - m_rp.m_lower[3] + block.z - 1 ) / block.z , maxblocks )
+        );
+      CudaParallelLaunch< ParallelFor, LaunchBounds >( *this , grid , block , 0 );
+    }
+    else if ( RP::rank == 5 )
+    {
+      // id0,id1 encoded within threadIdx.x; id2,id3 to threadIdx.y; id4 to threadIdx.z
+      const dim3 block( m_rp.m_tile[0]*m_rp.m_tile[1] , m_rp.m_tile[2]*m_rp.m_tile[3] , m_rp.m_tile[4] );
+      const dim3 grid(
+          std::min( static_cast( m_rp.m_tile_end[0] * m_rp.m_tile_end[1] )
+                  , static_cast(maxblocks) )
+        , std::min( static_cast( m_rp.m_tile_end[2] * m_rp.m_tile_end[3] )
+                  , static_cast(maxblocks) )
+        , std::min( ( m_rp.m_upper[4] - m_rp.m_lower[4] + block.z - 1 ) / block.z , maxblocks )
+        );
+      CudaParallelLaunch< ParallelFor, LaunchBounds >( *this , grid , block , 0 );
+    }
+    else if ( RP::rank == 6 )
+    {
+      // id0,id1 encoded within threadIdx.x; id2,id3 to threadIdx.y; id4,id5 to threadIdx.z
+      const dim3 block( m_rp.m_tile[0]*m_rp.m_tile[1] , m_rp.m_tile[2]*m_rp.m_tile[3] , m_rp.m_tile[4]*m_rp.m_tile[5] );
+      const dim3 grid(
+          std::min( static_cast( m_rp.m_tile_end[0] * m_rp.m_tile_end[1] )
+                  , static_cast(maxblocks) )
+        ,  std::min( static_cast( m_rp.m_tile_end[2] * m_rp.m_tile_end[3] )
+                  , static_cast(maxblocks) )
+        , std::min( static_cast( m_rp.m_tile_end[4] * m_rp.m_tile_end[5] )
+                  , static_cast(maxblocks) )
+        );
+      CudaParallelLaunch< ParallelFor, LaunchBounds >( *this , grid , block , 0 );
+    }
+    else
+    {
+      printf("Kokkos::MDRange Error: Exceeded rank bounds with Cuda\n");
+      Kokkos::abort("Aborting");
+    }
+
+  } //end execute
+
+//  inline
+  ParallelFor( const FunctorType & arg_functor
+             , Policy arg_policy )
+    : m_functor( arg_functor )
+    , m_rp(  arg_policy )
+    {}
+};
+
+
 template< class FunctorType , class ... Properties >
 class ParallelFor< FunctorType
                  , Kokkos::TeamPolicy< Properties ... >
@@ -384,6 +497,7 @@ private:
   typedef TeamPolicyInternal< Kokkos::Cuda , Properties ... >   Policy ;
   typedef typename Policy::member_type  Member ;
   typedef typename Policy::work_tag     WorkTag ;
+  typedef typename Policy::launch_bounds  LaunchBounds ;
 
 public:
 
@@ -430,15 +544,15 @@ public:
     if ( m_scratch_size[1]>0 ) {
       __shared__ int base_thread_id;
       if (threadIdx.x==0 && threadIdx.y==0 ) {
-        threadid = ((blockIdx.x*blockDim.z + threadIdx.z) * blockDim.x * blockDim.y) % kokkos_impl_cuda_lock_arrays.n;
+        threadid = ((blockIdx.x*blockDim.z + threadIdx.z) * blockDim.x * blockDim.y) % Kokkos::Impl::g_device_cuda_lock_arrays.n;
         threadid = ((threadid + blockDim.x * blockDim.y-1)/(blockDim.x * blockDim.y)) * blockDim.x * blockDim.y;
-        if(threadid > kokkos_impl_cuda_lock_arrays.n) threadid-=blockDim.x * blockDim.y;
+        if(threadid > Kokkos::Impl::g_device_cuda_lock_arrays.n) threadid-=blockDim.x * blockDim.y;
         int done = 0;
         while (!done) {
-          done = (0 == atomicCAS(&kokkos_impl_cuda_lock_arrays.atomic[threadid],0,1));
+          done = (0 == atomicCAS(&Kokkos::Impl::g_device_cuda_lock_arrays.scratch[threadid],0,1));
           if(!done) {
             threadid += blockDim.x * blockDim.y;
-            if(threadid > kokkos_impl_cuda_lock_arrays.n) threadid = 0;
+            if(threadid > Kokkos::Impl::g_device_cuda_lock_arrays.n) threadid = 0;
           }
         }
         base_thread_id = threadid;
@@ -448,7 +562,8 @@ public:
     }
 
 
-    for ( int league_rank = blockIdx.x ; league_rank < m_league_size ; league_rank += gridDim.x ) {
+    const int int_league_size = (int)m_league_size;
+    for ( int league_rank = blockIdx.x ; league_rank < int_league_size ; league_rank += gridDim.x ) {
 
       this-> template exec_team< WorkTag >(
         typename Policy::member_type( kokkos_impl_cuda_shared_memory()
@@ -462,7 +577,7 @@ public:
     if ( m_scratch_size[1]>0 ) {
       __syncthreads();
       if (threadIdx.x==0 && threadIdx.y==0 )
-        kokkos_impl_cuda_lock_arrays.atomic[threadid]=0;
+        Kokkos::Impl::g_device_cuda_lock_arrays.scratch[threadid]=0;
     }
   }
 
@@ -473,7 +588,7 @@ public:
       const dim3 grid( int(m_league_size) , 1 , 1 );
       const dim3 block( int(m_vector_size) , int(m_team_size) , 1 );
 
-      CudaParallelLaunch< ParallelFor >( *this, grid, block, shmem_size_total ); // copy to device and execute
+      CudaParallelLaunch< ParallelFor, LaunchBounds >( *this, grid, block, shmem_size_total ); // copy to device and execute
 
     }
 
@@ -529,6 +644,7 @@ private:
   typedef typename Policy::WorkRange    WorkRange ;
   typedef typename Policy::work_tag     WorkTag ;
   typedef typename Policy::member_type  Member ;
+  typedef typename Policy::launch_bounds LaunchBounds ;
 
   typedef Kokkos::Impl::if_c< std::is_same::value, FunctorType, ReducerType> ReducerConditional;
   typedef typename ReducerConditional::type ReducerTypeFwd;
@@ -563,6 +679,7 @@ private:
   typedef int DummySHMEMReductionType;
 
 public:
+  // Make the exec_range calls call to Reduce::DeviceIterateTile
   template< class TagType >
   __device__ inline
   typename std::enable_if< std::is_same< TagType , void >::value >::type
@@ -686,7 +803,7 @@ public:
 
       const int shmem = UseShflReduction?0:cuda_single_inter_block_reduce_scan_shmem( m_functor , block.y );
 
-      CudaParallelLaunch< ParallelReduce >( *this, grid, block, shmem ); // copy to device and execute
+      CudaParallelLaunch< ParallelReduce, LaunchBounds >( *this, grid, block, shmem ); // copy to device and execute
 
       Cuda::fence();
 
@@ -737,6 +854,232 @@ public:
   { }
 };
 
+
+// MDRangePolicy impl
+template< class FunctorType , class ReducerType, class ... Traits >
+class ParallelReduce< FunctorType
+                    , Kokkos::Experimental::MDRangePolicy< Traits ... >
+                    , ReducerType
+                    , Kokkos::Cuda
+                    >
+{
+private:
+
+  typedef Kokkos::Experimental::MDRangePolicy< Traits ... > Policy ;
+  typedef typename Policy::array_index_type                 array_index_type;
+  typedef typename Policy::index_type                       index_type;
+
+  typedef typename Policy::work_tag     WorkTag ;
+  typedef typename Policy::member_type  Member ;
+  typedef typename Policy::launch_bounds LaunchBounds;
+
+  typedef Kokkos::Impl::if_c< std::is_same::value, FunctorType, ReducerType> ReducerConditional;
+  typedef typename ReducerConditional::type ReducerTypeFwd;
+
+  typedef Kokkos::Impl::FunctorValueTraits< ReducerTypeFwd, WorkTag > ValueTraits ;
+  typedef Kokkos::Impl::FunctorValueInit<   ReducerTypeFwd, WorkTag > ValueInit ;
+  typedef Kokkos::Impl::FunctorValueJoin<   ReducerTypeFwd, WorkTag > ValueJoin ;
+
+public:
+
+  typedef typename ValueTraits::pointer_type    pointer_type ;
+  typedef typename ValueTraits::value_type      value_type ;
+  typedef typename ValueTraits::reference_type  reference_type ;
+  typedef FunctorType                           functor_type ;
+  typedef Cuda::size_type                       size_type ;
+
+  // Algorithmic constraints: blockSize is a power of two AND blockDim.y == blockDim.z == 1
+
+  const FunctorType   m_functor ;
+  const Policy        m_policy ; // used for workrange and nwork
+  const ReducerType   m_reducer ;
+  const pointer_type  m_result_ptr ;
+  size_type *         m_scratch_space ;
+  size_type *         m_scratch_flags ;
+  size_type *         m_unified_space ;
+
+  typedef typename Kokkos::Experimental::Impl::Reduce::DeviceIterateTile DeviceIteratePattern;
+
+  // Shall we use the shfl based reduction or not (only use it for static sized types of more than 128bit
+  enum { UseShflReduction = ((sizeof(value_type)>2*sizeof(double)) && ValueTraits::StaticValueSize) };
+  // Some crutch to do function overloading
+private:
+  typedef double DummyShflReductionType;
+  typedef int DummySHMEMReductionType;
+
+public:
+  inline
+  __device__
+  void
+  exec_range( reference_type update ) const
+  {
+    Kokkos::Experimental::Impl::Reduce::DeviceIterateTile(m_policy, m_functor, update).exec_range();
+  }
+
+  inline
+  __device__
+  void operator() (void) const {
+    run(Kokkos::Impl::if_c::select(1,1.0) );
+  }
+
+  __device__ inline
+  void run(const DummySHMEMReductionType& ) const
+  {
+    const integral_nonzero_constant< size_type , ValueTraits::StaticValueSize / sizeof(size_type) >
+      word_count( ValueTraits::value_size( ReducerConditional::select(m_functor , m_reducer) ) / sizeof(size_type) );
+
+    {
+      reference_type value =
+        ValueInit::init( ReducerConditional::select(m_functor , m_reducer) , kokkos_impl_cuda_shared_memory() + threadIdx.y * word_count.value );
+
+      // Number of blocks is bounded so that the reduction can be limited to two passes.
+      // Each thread block is given an approximately equal amount of work to perform.
+      // Accumulate the values for this block.
+      // The accumulation ordering does not match the final pass, but is arithmatically equivalent.
+
+      this-> exec_range( value );
+    }
+
+    // Reduce with final value at blockDim.y - 1 location.
+    // Problem: non power-of-two blockDim
+    if ( cuda_single_inter_block_reduce_scan(
+           ReducerConditional::select(m_functor , m_reducer) , blockIdx.x , gridDim.x ,
+           kokkos_impl_cuda_shared_memory() , m_scratch_space , m_scratch_flags ) ) {
+
+      // This is the final block with the final result at the final threads' location
+      size_type * const shared = kokkos_impl_cuda_shared_memory() + ( blockDim.y - 1 ) * word_count.value ;
+      size_type * const global = m_unified_space ? m_unified_space : m_scratch_space ;
+
+      if ( threadIdx.y == 0 ) {
+        Kokkos::Impl::FunctorFinal< ReducerTypeFwd , WorkTag >::final( ReducerConditional::select(m_functor , m_reducer) , shared );
+      }
+
+      if ( CudaTraits::WarpSize < word_count.value ) { __syncthreads(); }
+
+      for ( unsigned i = threadIdx.y ; i < word_count.value ; i += blockDim.y ) { global[i] = shared[i]; }
+    }
+  }
+
+  __device__ inline
+   void run(const DummyShflReductionType&) const
+   {
+
+     value_type value;
+     ValueInit::init( ReducerConditional::select(m_functor , m_reducer) , &value);
+     // Number of blocks is bounded so that the reduction can be limited to two passes.
+     // Each thread block is given an approximately equal amount of work to perform.
+     // Accumulate the values for this block.
+     // The accumulation ordering does not match the final pass, but is arithmatically equivalent.
+
+     const Member work_part =
+       ( ( m_policy.m_num_tiles + ( gridDim.x - 1 ) ) / gridDim.x ); //portion of tiles handled by each block
+
+     this-> exec_range( value );
+
+     pointer_type const result = (pointer_type) (m_unified_space ? m_unified_space : m_scratch_space) ;
+
+     int max_active_thread = work_part < blockDim.y ? work_part:blockDim.y;
+     max_active_thread = (max_active_thread == 0)?blockDim.y:max_active_thread;
+
+     value_type init;
+     ValueInit::init( ReducerConditional::select(m_functor , m_reducer) , &init);
+     if(Impl::cuda_inter_block_reduction
+         (value,init,ValueJoin(ReducerConditional::select(m_functor , m_reducer)),m_scratch_space,result,m_scratch_flags,max_active_thread)) {
+       const unsigned id = threadIdx.y*blockDim.x + threadIdx.x;
+       if(id==0) {
+         Kokkos::Impl::FunctorFinal< ReducerTypeFwd , WorkTag >::final( ReducerConditional::select(m_functor , m_reducer) , (void*) &value );
+         *result = value;
+       }
+     }
+   }
+
+  // Determine block size constrained by shared memory:
+  static inline
+  unsigned local_block_size( const FunctorType & f )
+    {
+      unsigned n = CudaTraits::WarpSize * 8 ;
+      while ( n && CudaTraits::SharedMemoryCapacity < cuda_single_inter_block_reduce_scan_shmem( f , n ) ) { n >>= 1 ; }
+      return n ;
+    }
+
+  inline
+  void execute()
+    {
+      const int nwork = m_policy.m_num_tiles;
+      if ( nwork ) {
+        int block_size = m_policy.m_prod_tile_dims;
+        // CONSTRAINT: Algorithm requires block_size >= product of tile dimensions
+        // Nearest power of two
+        int exponent_pow_two = std::ceil( std::log2(block_size) );
+        block_size = std::pow(2, exponent_pow_two);
+        int suggested_blocksize = local_block_size( m_functor );
+
+        block_size = (block_size > suggested_blocksize) ? block_size : suggested_blocksize ; //Note: block_size must be less than or equal to 512
+
+
+        m_scratch_space = cuda_internal_scratch_space( ValueTraits::value_size( ReducerConditional::select(m_functor , m_reducer) ) * block_size /* block_size == max block_count */ );
+        m_scratch_flags = cuda_internal_scratch_flags( sizeof(size_type) );
+        m_unified_space = cuda_internal_scratch_unified( ValueTraits::value_size( ReducerConditional::select(m_functor , m_reducer) ) );
+
+        // REQUIRED ( 1 , N , 1 )
+        const dim3 block( 1 , block_size , 1 );
+        // Required grid.x <= block.y
+        const dim3 grid( std::min( int(block.y) , int( nwork ) ) , 1 , 1 );
+
+      const int shmem = UseShflReduction?0:cuda_single_inter_block_reduce_scan_shmem( m_functor , block.y );
+
+      CudaParallelLaunch< ParallelReduce, LaunchBounds >( *this, grid, block, shmem ); // copy to device and execute
+
+      Cuda::fence();
+
+      if ( m_result_ptr ) {
+        if ( m_unified_space ) {
+          const int count = ValueTraits::value_count( ReducerConditional::select(m_functor , m_reducer)  );
+          for ( int i = 0 ; i < count ; ++i ) { m_result_ptr[i] = pointer_type(m_unified_space)[i] ; }
+        }
+        else {
+          const int size = ValueTraits::value_size( ReducerConditional::select(m_functor , m_reducer)  );
+          DeepCopy( m_result_ptr , m_scratch_space , size );
+        }
+      }
+    }
+    else {
+      if (m_result_ptr) {
+        ValueInit::init( ReducerConditional::select(m_functor , m_reducer) , m_result_ptr );
+      }
+    }
+  }
+
+  template< class HostViewType >
+  ParallelReduce( const FunctorType  & arg_functor
+                , const Policy       & arg_policy
+                , const HostViewType & arg_result
+                , typename std::enable_if<
+                   Kokkos::is_view< HostViewType >::value
+                ,void*>::type = NULL)
+  : m_functor( arg_functor )
+  , m_policy(  arg_policy )
+  , m_reducer( InvalidType() )
+  , m_result_ptr( arg_result.ptr_on_device() )
+  , m_scratch_space( 0 )
+  , m_scratch_flags( 0 )
+  , m_unified_space( 0 )
+  {}
+
+  ParallelReduce( const FunctorType  & arg_functor
+                , const Policy       & arg_policy
+                , const ReducerType & reducer)
+  : m_functor( arg_functor )
+  , m_policy(  arg_policy )
+  , m_reducer( reducer )
+  , m_result_ptr( reducer.view().ptr_on_device() )
+  , m_scratch_space( 0 )
+  , m_scratch_flags( 0 )
+  , m_unified_space( 0 )
+  {}
+};
+
+
 //----------------------------------------------------------------------------
 
 #if 1
@@ -753,6 +1096,7 @@ private:
   typedef TeamPolicyInternal< Kokkos::Cuda, Properties ... >  Policy ;
   typedef typename Policy::member_type  Member ;
   typedef typename Policy::work_tag     WorkTag ;
+  typedef typename Policy::launch_bounds     LaunchBounds ;
 
   typedef Kokkos::Impl::if_c< std::is_same::value, FunctorType, ReducerType> ReducerConditional;
   typedef typename ReducerConditional::type ReducerTypeFwd;
@@ -819,15 +1163,15 @@ public:
     if ( m_scratch_size[1]>0 ) {
       __shared__ int base_thread_id;
       if (threadIdx.x==0 && threadIdx.y==0 ) {
-        threadid = ((blockIdx.x*blockDim.z + threadIdx.z) * blockDim.x * blockDim.y) % kokkos_impl_cuda_lock_arrays.n;
+        threadid = ((blockIdx.x*blockDim.z + threadIdx.z) * blockDim.x * blockDim.y) % Kokkos::Impl::g_device_cuda_lock_arrays.n;
         threadid = ((threadid + blockDim.x * blockDim.y-1)/(blockDim.x * blockDim.y)) * blockDim.x * blockDim.y;
-        if(threadid > kokkos_impl_cuda_lock_arrays.n) threadid-=blockDim.x * blockDim.y;
+        if(threadid > Kokkos::Impl::g_device_cuda_lock_arrays.n) threadid-=blockDim.x * blockDim.y;
         int done = 0;
         while (!done) {
-          done = (0 == atomicCAS(&kokkos_impl_cuda_lock_arrays.atomic[threadid],0,1));
+          done = (0 == atomicCAS(&Kokkos::Impl::g_device_cuda_lock_arrays.scratch[threadid],0,1));
           if(!done) {
             threadid += blockDim.x * blockDim.y;
-            if(threadid > kokkos_impl_cuda_lock_arrays.n) threadid = 0;
+            if(threadid > Kokkos::Impl::g_device_cuda_lock_arrays.n) threadid = 0;
           }
         }
         base_thread_id = threadid;
@@ -840,7 +1184,7 @@ public:
     if ( m_scratch_size[1]>0 ) {
       __syncthreads();
       if (threadIdx.x==0 && threadIdx.y==0 )
-        kokkos_impl_cuda_lock_arrays.atomic[threadid]=0;
+        Kokkos::Impl::g_device_cuda_lock_arrays.scratch[threadid]=0;
     }
   }
 
@@ -854,7 +1198,8 @@ public:
       ValueInit::init( ReducerConditional::select(m_functor , m_reducer) , kokkos_impl_cuda_shared_memory() + threadIdx.y * word_count.value );
 
     // Iterate this block through the league
-    for ( int league_rank = blockIdx.x ; league_rank < m_league_size ; league_rank += gridDim.x ) {
+    const int int_league_size = (int)m_league_size;
+    for ( int league_rank = blockIdx.x ; league_rank < int_league_size ; league_rank += gridDim.x ) {
       this-> template exec_team< WorkTag >
         ( Member( kokkos_impl_cuda_shared_memory() + m_team_begin
                                         , m_shmem_begin
@@ -894,7 +1239,8 @@ public:
     ValueInit::init( ReducerConditional::select(m_functor , m_reducer) , &value);
 
     // Iterate this block through the league
-    for ( int league_rank = blockIdx.x ; league_rank < m_league_size ; league_rank += gridDim.x ) {
+    const int int_league_size = (int)m_league_size;
+    for ( int league_rank = blockIdx.x ; league_rank < int_league_size ; league_rank += gridDim.x ) {
       this-> template exec_team< WorkTag >
         ( Member( kokkos_impl_cuda_shared_memory() + m_team_begin
                                         , m_shmem_begin
@@ -936,7 +1282,7 @@ public:
         const dim3 grid( block_count , 1 , 1 );
         const int shmem_size_total = m_team_begin + m_shmem_begin + m_shmem_size ;
 
-        CudaParallelLaunch< ParallelReduce >( *this, grid, block, shmem_size_total ); // copy to device and execute
+        CudaParallelLaunch< ParallelReduce, LaunchBounds >( *this, grid, block, shmem_size_total ); // copy to device and execute
 
         Cuda::fence();
 
@@ -975,12 +1321,6 @@ public:
   , m_shmem_begin( 0 )
   , m_shmem_size( 0 )
   , m_scratch_ptr{NULL,NULL}
-  , m_league_size( arg_policy.league_size() )
-  , m_team_size( 0 <= arg_policy.team_size() ? arg_policy.team_size() :
-      Kokkos::Impl::cuda_get_opt_block_size< ParallelReduce >( arg_functor , arg_policy.vector_length(),
-                                                               arg_policy.team_scratch_size(0),arg_policy.thread_scratch_size(0) ) /
-                                                               arg_policy.vector_length() )
-  , m_vector_size( arg_policy.vector_length() )
   , m_scratch_size{
     arg_policy.scratch_size(0,( 0 <= arg_policy.team_size() ? arg_policy.team_size() :
         Kokkos::Impl::cuda_get_opt_block_size< ParallelReduce >( arg_functor , arg_policy.vector_length(),
@@ -991,6 +1331,12 @@ public:
                                                                  arg_policy.team_scratch_size(0),arg_policy.thread_scratch_size(0) ) /
                                                                  arg_policy.vector_length() )
         )}
+  , m_league_size( arg_policy.league_size() )
+  , m_team_size( 0 <= arg_policy.team_size() ? arg_policy.team_size() :
+      Kokkos::Impl::cuda_get_opt_block_size< ParallelReduce >( arg_functor , arg_policy.vector_length(),
+                                                               arg_policy.team_scratch_size(0),arg_policy.thread_scratch_size(0) ) /
+                                                               arg_policy.vector_length() )
+  , m_vector_size( arg_policy.vector_length() )
   {
     // Return Init value if the number of worksets is zero
     if( arg_policy.league_size() == 0) {
@@ -1150,6 +1496,7 @@ private:
   typedef typename reducer_type<>::pointer_type    pointer_type ;
   typedef typename reducer_type<>::reference_type  reference_type ;
   typedef typename reducer_type<>::value_type      value_type ;
+  typedef typename Policy::launch_bounds           LaunchBounds ;
 
   typedef Kokkos::Impl::FunctorAnalysis
     < Kokkos::Impl::FunctorPatternInterface::REDUCE
@@ -1273,7 +1620,7 @@ public:
         const int  shmem = m_shmem_team_begin + m_shmem_team_size ;
 
         // copy to device and execute
-        CudaParallelLaunch( *this, grid, block, shmem );
+        CudaParallelLaunch( *this, grid, block, shmem );
 
         Cuda::fence();
 
@@ -1373,7 +1720,7 @@ public:
 
     if ( CudaTraits::WarpSize < team_threads ) {
       // Need inter-warp team reduction (collectives) shared memory
-      // Speculate an upper bound for the value size 
+      // Speculate an upper bound for the value size
 
       m_shmem_team_begin =
         align_scratch( CudaTraits::warp_count(team_threads) * sizeof(double) );
@@ -1426,7 +1773,7 @@ public:
 
     // Reduce space has claim flag followed by vaue buffer
     const int global_reduce_value_size =
-      max_concurrent_block * 
+      max_concurrent_block *
       ( aligned_flag_size + align_scratch( value_size ) );
 
     // Scratch space has claim flag followed by scratch buffer
@@ -1469,6 +1816,7 @@ private:
   typedef typename Policy::member_type  Member ;
   typedef typename Policy::work_tag     WorkTag ;
   typedef typename Policy::WorkRange    WorkRange ;
+  typedef typename Policy::launch_bounds  LaunchBounds ;
 
   typedef Kokkos::Impl::FunctorValueTraits< FunctorType, WorkTag > ValueTraits ;
   typedef Kokkos::Impl::FunctorValueInit<   FunctorType, WorkTag > ValueInit ;
@@ -1655,10 +2003,10 @@ public:
         const int shmem = ValueTraits::value_size( m_functor ) * ( block_size + 2 );
 
         m_final = false ;
-        CudaParallelLaunch< ParallelScan >( *this, grid, block, shmem ); // copy to device and execute
+        CudaParallelLaunch< ParallelScan, LaunchBounds >( *this, grid, block, shmem ); // copy to device and execute
 
         m_final = true ;
-        CudaParallelLaunch< ParallelScan >( *this, grid, block, shmem ); // copy to device and execute
+        CudaParallelLaunch< ParallelScan, LaunchBounds >( *this, grid, block, shmem ); // copy to device and execute
       }
     }
 
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_ReduceScan.hpp b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_ReduceScan.hpp
index 432c7895cc..709cbbd534 100644
--- a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_ReduceScan.hpp
+++ b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_ReduceScan.hpp
@@ -151,7 +151,7 @@ template< class ValueType , class JoinOp>
 __device__
 inline void cuda_intra_warp_reduction( ValueType& result,
                                        const JoinOp& join,
-                                       const int max_active_thread = blockDim.y) {
+                                       const uint32_t max_active_thread = blockDim.y) {
 
   unsigned int shift = 1;
 
@@ -268,29 +268,33 @@ bool cuda_inter_block_reduction( typename FunctorValueTraits< FunctorType , ArgT
         if( id + 1 < int(gridDim.x) )
           join(value, tmp);
       }
+      int active = __ballot(1);
       if (int(blockDim.x*blockDim.y) > 2) {
         value_type tmp = Kokkos::shfl_down(value, 2,32);
         if( id + 2 < int(gridDim.x) )
           join(value, tmp);
       }
+      active += __ballot(1);
       if (int(blockDim.x*blockDim.y) > 4) {
         value_type tmp = Kokkos::shfl_down(value, 4,32);
         if( id + 4 < int(gridDim.x) )
           join(value, tmp);
       }
+      active += __ballot(1);
       if (int(blockDim.x*blockDim.y) > 8) {
         value_type tmp = Kokkos::shfl_down(value, 8,32);
         if( id + 8 < int(gridDim.x) )
           join(value, tmp);
       }
+      active += __ballot(1);
       if (int(blockDim.x*blockDim.y) > 16) {
         value_type tmp = Kokkos::shfl_down(value, 16,32);
         if( id + 16 < int(gridDim.x) )
           join(value, tmp);
       }
+      active += __ballot(1);
     }
   }
-
   //The last block has in its thread=0 the global reduction value through "value"
   return last_block;
 #else
@@ -302,7 +306,7 @@ template< class ReducerType >
 __device__ inline
 typename std::enable_if< Kokkos::is_reducer::value >::type
 cuda_intra_warp_reduction( const ReducerType& reducer,
-                           const int max_active_thread = blockDim.y) {
+                           const uint32_t max_active_thread = blockDim.y) {
 
   typedef typename ReducerType::value_type ValueType;
 
@@ -428,26 +432,31 @@ cuda_inter_block_reduction( const ReducerType& reducer,
         if( id + 1 < int(gridDim.x) )
           reducer.join(value, tmp);
       }
+      int active = __ballot(1);
       if (int(blockDim.x*blockDim.y) > 2) {
         value_type tmp = Kokkos::shfl_down(value, 2,32);
         if( id + 2 < int(gridDim.x) )
           reducer.join(value, tmp);
       }
+      active += __ballot(1);
       if (int(blockDim.x*blockDim.y) > 4) {
         value_type tmp = Kokkos::shfl_down(value, 4,32);
         if( id + 4 < int(gridDim.x) )
           reducer.join(value, tmp);
       }
+      active += __ballot(1);
       if (int(blockDim.x*blockDim.y) > 8) {
         value_type tmp = Kokkos::shfl_down(value, 8,32);
         if( id + 8 < int(gridDim.x) )
           reducer.join(value, tmp);
       }
+      active += __ballot(1);
       if (int(blockDim.x*blockDim.y) > 16) {
         value_type tmp = Kokkos::shfl_down(value, 16,32);
         if( id + 16 < int(gridDim.x) )
           reducer.join(value, tmp);
       }
+      active += __ballot(1);
     }
   }
 
@@ -594,7 +603,7 @@ bool cuda_single_inter_block_reduce_scan( const FunctorType     & functor ,
   typedef FunctorValueOps<    FunctorType , ArgTag >  ValueOps ;
 
   typedef typename ValueTraits::pointer_type    pointer_type ;
-  typedef typename ValueTraits::reference_type  reference_type ;
+  //typedef typename ValueTraits::reference_type  reference_type ;
 
   // '__ffs' = position of the least significant bit set to 1.
   // 'blockDim.y' is guaranteed to be a power of two so this
@@ -637,7 +646,7 @@ bool cuda_single_inter_block_reduce_scan( const FunctorType     & functor ,
 
     {
       void * const shared_ptr = shared_data + word_count.value * threadIdx.y ;
-      reference_type shared_value = ValueInit::init( functor , shared_ptr );
+      /* reference_type shared_value = */ ValueInit::init( functor , shared_ptr );
 
       for ( size_type i = b ; i < e ; ++i ) {
         ValueJoin::join( functor , shared_ptr , global_data + word_count.value * i );
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Task.cpp b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Task.cpp
index 3c6f0a5dda..5f08800c40 100644
--- a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Task.cpp
+++ b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Task.cpp
@@ -58,25 +58,56 @@ template class TaskQueue< Kokkos::Cuda > ;
 
 //----------------------------------------------------------------------------
 
+#if defined( KOKKOS_DEBUG )
+
+__device__
+void verify_warp_convergence( const char * const where )
+{
+  const unsigned b = __ballot(1);
+
+  if ( b != ~0u ) {
+
+printf(" verify_warp_convergence( %s ) (%d,%d,%d) (%d,%d,%d) failed %x\n"
+      , where
+      , blockIdx.x
+      , blockIdx.y
+      , blockIdx.z
+      , threadIdx.x
+      , threadIdx.y
+      , threadIdx.z
+      , b );
+
+  }
+}
+
+#endif // #if defined( KOKKOS_DEBUG )
+
+//----------------------------------------------------------------------------
+
 __device__
 void TaskQueueSpecialization< Kokkos::Cuda >::driver
-  ( TaskQueueSpecialization< Kokkos::Cuda >::queue_type * const queue )
+  ( TaskQueueSpecialization< Kokkos::Cuda >::queue_type * const queue 
+  , int32_t shmem_per_warp )
 {
   using Member = TaskExec< Kokkos::Cuda > ;
   using Queue  = TaskQueue< Kokkos::Cuda > ;
-  using task_root_type = TaskBase< Kokkos::Cuda , void , void > ;
+  using task_root_type = TaskBase< void , void , void > ;
+
+  extern __shared__ int32_t shmem_all[];
 
   task_root_type * const end = (task_root_type *) task_root_type::EndTag ;
 
-  Member single_exec( 1 );
-  Member team_exec( blockDim.y );
+  int32_t * const warp_shmem =
+    shmem_all + ( threadIdx.z * shmem_per_warp ) / sizeof(int32_t);
+
+  task_root_type * const task_shmem = (task_root_type *) warp_shmem ;
 
   const int warp_lane = threadIdx.x + threadIdx.y * blockDim.x ;
 
-  union {
-    task_root_type * ptr ;
-    int              raw[2] ;
-  } task ;
+  Member single_exec( warp_shmem , 1 );
+  Member team_exec( warp_shmem , blockDim.y );
+
+  task_root_type * task_ptr ;
 
   // Loop until all queues are empty and no tasks in flight
 
@@ -87,41 +118,86 @@ void TaskQueueSpecialization< Kokkos::Cuda >::driver
 
     if ( 0 == warp_lane ) {
 
-      task.ptr = 0 < *((volatile int *) & queue->m_ready_count) ? end : 0 ;
+      task_ptr = 0 < *((volatile int *) & queue->m_ready_count) ? end : 0 ;
 
       // Loop by priority and then type
-      for ( int i = 0 ; i < Queue::NumQueue && end == task.ptr ; ++i ) {
-        for ( int j = 0 ; j < 2 && end == task.ptr ; ++j ) {
-          task.ptr = Queue::pop_ready_task( & queue->m_ready[i][j] );
+      for ( int i = 0 ; i < Queue::NumQueue && end == task_ptr ; ++i ) {
+        for ( int j = 0 ; j < 2 && end == task_ptr ; ++j ) {
+          task_ptr = Queue::pop_ready_task( & queue->m_ready[i][j] );
         }
       }
 
 #if 0
 printf("TaskQueue::driver(%d,%d) task(%lx)\n",threadIdx.z,blockIdx.x
-      , uintptr_t(task.ptr));
+      , uintptr_t(task_ptr));
 #endif
 
     }
 
     // shuffle broadcast
 
-    task.raw[0] = __shfl( task.raw[0] , 0 );
-    task.raw[1] = __shfl( task.raw[1] , 0 );
+    ((int*) & task_ptr )[0] = __shfl( ((int*) & task_ptr )[0] , 0 );
+    ((int*) & task_ptr )[1] = __shfl( ((int*) & task_ptr )[1] , 0 );
 
-    if ( 0 == task.ptr ) break ; // 0 == queue->m_ready_count
+#if defined( KOKKOS_DEBUG )
+    verify_warp_convergence("task_ptr");
+#endif
 
-    if ( end != task.ptr ) {
-      if ( task_root_type::TaskTeam == task.ptr->m_task_type ) {
+    if ( 0 == task_ptr ) break ; // 0 == queue->m_ready_count
+
+    if ( end != task_ptr ) {
+
+      // Whole warp copy task's closure to/from shared memory.
+      // Use all threads of warp for coalesced read/write.
+
+      int32_t const b = sizeof(task_root_type) / sizeof(int32_t);
+      int32_t const e = *((int32_t volatile *)( & task_ptr->m_alloc_size )) / sizeof(int32_t);
+
+      int32_t volatile * const task_mem = (int32_t volatile *) task_ptr ;
+
+      // copy global to shared memory:
+
+      for ( int32_t i = warp_lane ; i < e ; i += CudaTraits::WarpSize ) {
+        warp_shmem[i] = task_mem[i] ;
+      }
+
+      Kokkos::memory_fence();
+
+      // Copy done - use memory fence so that memory writes are visible.
+      // For reliable warp convergence on Pascal and Volta an explicit
+      // warp level synchronization will also be required.
+
+      if ( task_root_type::TaskTeam == task_shmem->m_task_type ) {
         // Thread Team Task
-        (*task.ptr->m_apply)( task.ptr , & team_exec );
+        (*task_shmem->m_apply)( task_shmem , & team_exec );
       }
       else if ( 0 == threadIdx.y ) {
         // Single Thread Task
-        (*task.ptr->m_apply)( task.ptr , & single_exec );
+        (*task_shmem->m_apply)( task_shmem , & single_exec );
       }
 
+      // copy shared to global memory:
+
+      for ( int32_t i = b + warp_lane ; i < e ; i += CudaTraits::WarpSize ) {
+        task_mem[i] = warp_shmem[i] ;
+      }
+
+      Kokkos::memory_fence();
+
+#if defined( KOKKOS_DEBUG )
+    verify_warp_convergence("apply");
+#endif
+
+      // If respawn requested copy respawn data back to main memory
+
       if ( 0 == warp_lane ) {
-        queue->complete( task.ptr );
+
+        if ( ((task_root_type *) task_root_type::LockTag) != task_shmem->m_next ) {
+          ( (volatile task_root_type *) task_ptr )->m_next = task_shmem->m_next ;
+          ( (volatile task_root_type *) task_ptr )->m_priority = task_shmem->m_priority ;
+        }
+
+        queue->complete( task_ptr );
       }
     }
   } while(1);
@@ -130,18 +206,20 @@ printf("TaskQueue::driver(%d,%d) task(%lx)\n",threadIdx.z,blockIdx.x
 namespace {
 
 __global__
-void cuda_task_queue_execute( TaskQueue< Kokkos::Cuda > * queue )
-{ TaskQueueSpecialization< Kokkos::Cuda >::driver( queue ); }
+void cuda_task_queue_execute( TaskQueue< Kokkos::Cuda > * queue 
+                            , int32_t shmem_size )
+{ TaskQueueSpecialization< Kokkos::Cuda >::driver( queue , shmem_size ); }
 
 }
 
 void TaskQueueSpecialization< Kokkos::Cuda >::execute
   ( TaskQueue< Kokkos::Cuda > * const queue )
 {
+  const int shared_per_warp = 2048 ;
   const int warps_per_block = 4 ;
   const dim3 grid( Kokkos::Impl::cuda_internal_multiprocessor_count() , 1 , 1 );
   const dim3 block( 1 , Kokkos::Impl::CudaTraits::WarpSize , warps_per_block );
-  const int shared = 0 ;
+  const int shared_total = shared_per_warp * warps_per_block ;
   const cudaStream_t stream = 0 ;
 
   CUDA_SAFE_CALL( cudaDeviceSynchronize() );
@@ -159,7 +237,7 @@ printf("cuda_task_queue_execute before\n");
   //
   // CUDA_SAFE_CALL( cudaDeviceSetLimit( cudaLimitStackSize , stack_size ) );
 
-  cuda_task_queue_execute<<< grid , block , shared , stream >>>( queue );
+  cuda_task_queue_execute<<< grid , block , shared_total , stream >>>( queue , shared_per_warp );
 
   CUDA_SAFE_CALL( cudaGetLastError() );
 
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Task.hpp b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Task.hpp
index 5d08219ea5..4a52985d29 100644
--- a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Task.hpp
+++ b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Task.hpp
@@ -57,7 +57,7 @@ namespace {
 template< typename TaskType >
 __global__
 void set_cuda_task_base_apply_function_pointer
-  ( TaskBase::function_type * ptr )
+  ( TaskBase::function_type * ptr )
 { *ptr = TaskType::apply ; }
 
 }
@@ -78,7 +78,7 @@ public:
   void iff_single_thread_recursive_execute( queue_type * const ) {}
 
   __device__
-  static void driver( queue_type * const );
+  static void driver( queue_type * const , int32_t );
 
   static
   void execute( queue_type * const );
@@ -106,7 +106,14 @@ public:
 
 extern template class TaskQueue< Kokkos::Cuda > ;
 
+}} /* namespace Kokkos::Impl */
+
 //----------------------------------------------------------------------------
+//----------------------------------------------------------------------------
+
+namespace Kokkos {
+namespace Impl {
+
 /**\brief  Impl::TaskExec is the TaskScheduler::member_type
  *         passed to tasks running in a Cuda space.
  *
@@ -134,11 +141,13 @@ private:
   friend class Kokkos::Impl::TaskQueue< Kokkos::Cuda > ;
   friend class Kokkos::Impl::TaskQueueSpecialization< Kokkos::Cuda > ;
 
+  int32_t * m_team_shmem ;
   const int m_team_size ;
 
   __device__
-  TaskExec( int arg_team_size = blockDim.y )
-    : m_team_size( arg_team_size ) {}
+  TaskExec( int32_t * arg_team_shmem , int arg_team_size = blockDim.y )
+    : m_team_shmem( arg_team_shmem )
+    , m_team_size( arg_team_size ) {}
 
 public:
 
@@ -154,7 +163,13 @@ public:
 
 };
 
+}} /* namespace Kokkos::Impl */
+
 //----------------------------------------------------------------------------
+//----------------------------------------------------------------------------
+
+namespace Kokkos {
+namespace Impl {
 
 template
 struct TeamThreadRangeBoundariesStruct >
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Team.hpp b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Team.hpp
index 084daa098b..3f3d85ecd1 100644
--- a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Team.hpp
+++ b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_Team.hpp
@@ -106,7 +106,7 @@ private:
   typedef Kokkos::Cuda                           execution_space ;
   typedef execution_space::scratch_memory_space  scratch_memory_space ;
 
-  void                * m_team_reduce ;
+  mutable void        * m_team_reduce ;
   scratch_memory_space  m_team_shared ;
   int                   m_team_reduce_size ;
   int                   m_league_rank ;
@@ -166,7 +166,7 @@ public:
       if ( 1 == blockDim.z ) { // team == block
         __syncthreads();
         // Wait for shared data write until all threads arrive here
-        if ( threadIdx.x == 0 && threadIdx.y == thread_id ) {
+        if ( threadIdx.x == 0u && threadIdx.y == (uint32_t)thread_id ) {
           *((ValueType*) m_team_reduce) = val ;
         }
         __syncthreads(); // Wait for shared data read until root thread writes
@@ -210,7 +210,7 @@ public:
       const int wx =
         ( threadIdx.x + blockDim.x * threadIdx.y ) & CudaTraits::WarpIndexMask ;
 
-      for ( int i = CudaTraits::WarpSize ; blockDim.x <= ( i >>= 1 ) ; ) {
+      for ( int i = CudaTraits::WarpSize ; (int)blockDim.x <= ( i >>= 1 ) ; ) {
 
         cuda_shfl_down( reducer.reference() , tmp , i , CudaTraits::WarpSize );
 
@@ -354,7 +354,7 @@ public:
 
       for ( int i = blockDim.x ; ( i >>= 1 ) ; ) {
         cuda_shfl_down( reducer.reference() , tmp , i , blockDim.x );
-        if ( threadIdx.x < i ) { reducer.join( tmp , reducer.reference() ); }
+        if ( (int)threadIdx.x < i ) { reducer.join( tmp , reducer.reference() ); }
       }
 
       // Broadcast from root lane to all other lanes.
@@ -410,7 +410,7 @@ public:
 
         value_type tmp( reducer.reference() );
 
-        for ( int i = CudaTraits::WarpSize ; blockDim.x <= ( i >>= 1 ) ; ) {
+        for ( int i = CudaTraits::WarpSize ; (int)blockDim.x <= ( i >>= 1 ) ; ) {
 
           cuda_shfl_down( reducer.reference(), tmp, i, CudaTraits::WarpSize );
 
@@ -479,7 +479,7 @@ public:
 
           __threadfence(); // Wait until global write is visible.
 
-          last_block = gridDim.x ==
+          last_block = (int)gridDim.x ==
                        1 + Kokkos::atomic_fetch_add(global_scratch_flags,1);
 
           // If last block then reset count
@@ -509,7 +509,7 @@ public:
         reducer.copy( ((pointer_type)shmem) + offset
                     , ((pointer_type)global_scratch_space) + offset );
 
-        for ( int i = nentry + tid ; i < gridDim.x ; i += nentry ) {
+        for ( int i = nentry + tid ; i < (int)gridDim.x ; i += nentry ) {
           reducer.join( ((pointer_type)shmem) + offset
                       , ((pointer_type)global_scratch_space)
                         + i * reducer.length() );
@@ -576,6 +576,14 @@ public:
     , m_league_size( arg_league_size )
     {}
 
+public:
+  // Declare to avoid unused private member warnings which are trigger
+  // when SFINAE excludes the member function which uses these variables
+  // Making another class a friend also surpresses these warnings
+  bool impl_avoid_sfinae_warning() const noexcept
+  {
+    return m_team_reduce_size > 0 && m_team_reduce != nullptr;
+  }
 };
 
 } // namspace Impl
@@ -913,10 +921,10 @@ void parallel_scan
     //  [t] += [t-4] if t >= 4
     //  ...
 
-    for ( int j = 1 ; j < blockDim.x ; j <<= 1 ) {
+    for ( int j = 1 ; j < (int)blockDim.x ; j <<= 1 ) {
       value_type tmp = 0 ;
       Impl::cuda_shfl_up( tmp , sval , j , blockDim.x );
-      if ( j <= threadIdx.x ) { sval += tmp ; }
+      if ( j <= (int)threadIdx.x ) { sval += tmp ; }
     }
 
     // Include accumulation and remove value for exclusive scan:
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_UniqueToken.hpp b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_UniqueToken.hpp
new file mode 100644
index 0000000000..e11ae4798f
--- /dev/null
+++ b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_UniqueToken.hpp
@@ -0,0 +1,133 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_CUDA_UNIQUE_TOKEN_HPP
+#define KOKKOS_CUDA_UNIQUE_TOKEN_HPP
+
+#include 
+#ifdef KOKKOS_ENABLE_CUDA
+
+#include 
+#include 
+#include 
+#include 
+
+namespace Kokkos { namespace Experimental {
+
+// both global and instance Unique Tokens are implemented in the same way
+template<>
+class UniqueToken< Cuda, UniqueTokenScope::Global >
+{
+private:
+
+  uint32_t volatile * m_buffer ;
+  uint32_t            m_count ;
+
+public:
+
+  using execution_space = Cuda;
+
+  explicit
+  UniqueToken( execution_space const& );
+
+  KOKKOS_INLINE_FUNCTION
+  UniqueToken() : m_buffer(0), m_count(0) {}
+
+  KOKKOS_INLINE_FUNCTION
+  UniqueToken( const UniqueToken & ) = default;
+
+  KOKKOS_INLINE_FUNCTION
+  UniqueToken( UniqueToken && )      = default;
+
+  KOKKOS_INLINE_FUNCTION
+  UniqueToken & operator=( const UniqueToken & ) = default ;
+
+  KOKKOS_INLINE_FUNCTION
+  UniqueToken & operator=( UniqueToken && ) = default ;
+
+  /// \brief upper bound for acquired values, i.e. 0 <= value < size()
+  KOKKOS_INLINE_FUNCTION
+  int32_t size() const noexcept { return m_count ; }
+
+  /// \brief acquire value such that 0 <= value < size()
+  KOKKOS_INLINE_FUNCTION
+  int32_t acquire() const
+  {
+    const Kokkos::pair result =
+      Kokkos::Impl::concurrent_bitset::
+        acquire_bounded( m_buffer
+                       , m_count
+                       , Kokkos::Impl::clock_tic() % m_count
+                       );
+
+   if ( result.first < 0 ) {
+     Kokkos::abort("UniqueToken failure to release tokens, no tokens available" );
+   }
+
+    return result.first;
+  }
+
+  /// \brief release an acquired value
+  KOKKOS_INLINE_FUNCTION
+  void release( int32_t i ) const noexcept
+  {
+    Kokkos::Impl::concurrent_bitset::release( m_buffer, i );
+  }
+};
+
+template<>
+class UniqueToken< Cuda, UniqueTokenScope::Instance >
+  : public UniqueToken< Cuda, UniqueTokenScope::Global >
+{
+public:
+
+  explicit
+  UniqueToken( execution_space const& arg )
+    : UniqueToken< Cuda, UniqueTokenScope::Global >( arg ) {}
+};
+
+}} // namespace Kokkos::Experimental
+
+#endif // KOKKOS_ENABLE_CUDA
+#endif // KOKKOS_CUDA_UNIQUE_TOKEN_HPP
+
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_View.hpp b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_View.hpp
index f5e2d87fb6..d641622bb6 100644
--- a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_View.hpp
+++ b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_View.hpp
@@ -221,7 +221,6 @@ struct CudaLDGFetch {
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
-namespace Experimental {
 namespace Impl {
 
 /** \brief  Replace Default ViewDataHandle with Cuda texture fetch specialization
@@ -294,9 +293,8 @@ public:
     }
 };
 
-}
-}
-}
+} // namespace Impl
+} // namespace Kokkos
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
diff --git a/lib/kokkos/core/src/Cuda/Kokkos_Cuda_WorkGraphPolicy.hpp b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_WorkGraphPolicy.hpp
new file mode 100644
index 0000000000..99778c64b1
--- /dev/null
+++ b/lib/kokkos/core/src/Cuda/Kokkos_Cuda_WorkGraphPolicy.hpp
@@ -0,0 +1,119 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_CUDA_WORKGRAPHPOLICY_HPP
+#define KOKKOS_CUDA_WORKGRAPHPOLICY_HPP
+
+namespace Kokkos {
+namespace Impl {
+
+template< class FunctorType , class ... Traits >
+class ParallelFor< FunctorType ,
+                   Kokkos::Experimental::WorkGraphPolicy< Traits ... > ,
+                   Kokkos::Cuda
+                 >
+  : public Kokkos::Impl::Experimental::
+           WorkGraphExec< FunctorType,
+                          Kokkos::Cuda,
+                          Traits ...
+                        >
+{
+public:
+
+  typedef Kokkos::Experimental::WorkGraphPolicy< Traits ... >   Policy ;
+  typedef Kokkos::Impl::Experimental::
+          WorkGraphExec Base ;
+  typedef ParallelFor        Self ;
+
+private:
+
+  template< class TagType >
+  __device__
+  typename std::enable_if< std::is_same< TagType , void >::value >::type
+  exec_one(const typename Policy::member_type& i) const {
+    Base::m_functor( i );
+  }
+
+  template< class TagType >
+  __device__
+  typename std::enable_if< ! std::is_same< TagType , void >::value >::type
+  exec_one(const typename Policy::member_type& i) const {
+    const TagType t{} ;
+    Base::m_functor( t , i );
+  }
+
+public:
+
+  __device__
+  inline
+  void operator()() const {
+    for (std::int32_t i; (-1 != (i = Base::before_work())); ) {
+      exec_one< typename Policy::work_tag >( i );
+      Base::after_work(i);
+    }
+  }
+
+  inline
+  void execute()
+  {
+    const int warps_per_block = 4 ;
+    const dim3 grid( Kokkos::Impl::cuda_internal_multiprocessor_count() , 1 , 1 );
+    const dim3 block( 1 , Kokkos::Impl::CudaTraits::WarpSize , warps_per_block );
+    const int shared = 0 ;
+    const cudaStream_t stream = 0 ;
+
+    Kokkos::Impl::CudaParallelLaunch(*this, grid, block, shared, stream);
+  }
+
+  inline
+  ParallelFor( const FunctorType & arg_functor
+             , const Policy      & arg_policy )
+    : Base( arg_functor, arg_policy )
+  {
+  }
+};
+
+} // namespace Impl
+} // namespace Kokkos
+
+#endif /* #define KOKKOS_CUDA_WORKGRAPHPOLICY_HPP */
diff --git a/lib/kokkos/core/src/KokkosExp_MDRangePolicy.hpp b/lib/kokkos/core/src/KokkosExp_MDRangePolicy.hpp
index 4f68d9c2c0..6ef7443a14 100644
--- a/lib/kokkos/core/src/KokkosExp_MDRangePolicy.hpp
+++ b/lib/kokkos/core/src/KokkosExp_MDRangePolicy.hpp
@@ -52,6 +52,7 @@
 
 #if defined( __CUDACC__ ) && defined( KOKKOS_ENABLE_CUDA )
 #include
+#include 
 #endif
 
 namespace Kokkos { namespace Experimental {
@@ -120,28 +121,17 @@ struct MDRangePolicy
                                        , typename traits::index_type
                                        > ;
 
+  typedef MDRangePolicy execution_policy; // needed for is_execution_space interrogation
+
   static_assert( !std::is_same::value
                , "Kokkos Error: MD iteration pattern not defined" );
 
   using iteration_pattern   = typename traits::iteration_pattern;
   using work_tag            = typename traits::work_tag;
+  using launch_bounds       = typename traits::launch_bounds;
+  using member_type = typename range_policy::member_type;
 
-  static constexpr int rank = iteration_pattern::rank;
-
-  static constexpr int outer_direction = static_cast (
-      (iteration_pattern::outer_direction != Iterate::Default)
-    ? iteration_pattern::outer_direction
-    : default_outer_direction< typename traits::execution_space>::value );
-
-  static constexpr int inner_direction = static_cast (
-      iteration_pattern::inner_direction != Iterate::Default
-    ? iteration_pattern::inner_direction
-    : default_inner_direction< typename traits::execution_space>::value ) ;
-
-
-  // Ugly ugly workaround intel 14 not handling scoped enum correctly
-  static constexpr int Right = static_cast( Iterate::Right );
-  static constexpr int Left  = static_cast( Iterate::Left );
+  enum { rank = static_cast(iteration_pattern::rank) };
 
   using index_type  = typename traits::index_type;
   using array_index_type = long;
@@ -155,11 +145,50 @@ struct MDRangePolicy
   // This would require the user to either pass a matching index_type parameter
   // as template parameter to the MDRangePolicy or static_cast the individual values
 
+  point_type m_lower;
+  point_type m_upper;
+  tile_type  m_tile;
+  point_type m_tile_end;
+  index_type m_num_tiles;
+  index_type m_prod_tile_dims;
+
+/*
+  // NDE enum impl definition alternative - replace static constexpr int ? 
+  enum { outer_direction = static_cast (
+      (iteration_pattern::outer_direction != Iterate::Default)
+    ? iteration_pattern::outer_direction
+    : default_outer_direction< typename traits::execution_space>::value ) };
+
+  enum { inner_direction = static_cast (
+      iteration_pattern::inner_direction != Iterate::Default
+    ? iteration_pattern::inner_direction
+    : default_inner_direction< typename traits::execution_space>::value ) };
+
+  enum { Right = static_cast( Iterate::Right ) };
+  enum { Left  = static_cast( Iterate::Left ) };
+*/
+  //static constexpr int rank = iteration_pattern::rank;
+
+  static constexpr int outer_direction = static_cast (
+      (iteration_pattern::outer_direction != Iterate::Default)
+    ? iteration_pattern::outer_direction
+    : default_outer_direction< typename traits::execution_space>::value );
+
+  static constexpr int inner_direction = static_cast (
+      iteration_pattern::inner_direction != Iterate::Default
+    ? iteration_pattern::inner_direction
+    : default_inner_direction< typename traits::execution_space>::value ) ;
+
+  // Ugly ugly workaround intel 14 not handling scoped enum correctly
+  static constexpr int Right = static_cast( Iterate::Right );
+  static constexpr int Left  = static_cast( Iterate::Left );
+
   MDRangePolicy( point_type const& lower, point_type const& upper, tile_type const& tile = tile_type{} )
     : m_lower(lower)
     , m_upper(upper)
     , m_tile(tile)
     , m_num_tiles(1)
+    , m_prod_tile_dims(1)
   {
     // Host
     if ( true
@@ -172,8 +201,8 @@ struct MDRangePolicy
       for (int i=0; i 0)) )
+          if (  ((int)inner_direction == (int)Right && (i < rank-1))
+              || ((int)inner_direction == (int)Left && (i > 0)) )
           {
             m_tile[i] = 2;
           }
@@ -183,6 +212,7 @@ struct MDRangePolicy
         }
         m_tile_end[i] = static_cast((span + m_tile[i] - 1) / m_tile[i]);
         m_num_tiles *= m_tile_end[i];
+        m_prod_tile_dims *= m_tile[i];
       }
     }
     #if defined(KOKKOS_ENABLE_CUDA)
@@ -190,14 +220,18 @@ struct MDRangePolicy
     {
       index_type span;
       for (int i=0; i 0)) )
+          if (  ((int)inner_direction == (int)Right && (i < rank-1))
+              || ((int)inner_direction == (int)Left && (i > 0)) )
           {
-            m_tile[i] = 2;
+            if ( m_prod_tile_dims < 512 ) {
+              m_tile[i] = 2;
+            } else {
+              m_tile[i] = 1;
+            }
           }
           else {
             m_tile[i] = 16;
@@ -205,12 +239,9 @@ struct MDRangePolicy
         }
         m_tile_end[i] = static_cast((span + m_tile[i] - 1) / m_tile[i]);
         m_num_tiles *= m_tile_end[i];
+        m_prod_tile_dims *= m_tile[i];
       }
-      index_type total_tile_size_check = 1;
-      for (int i=0; i= 1024 ) { // improve this check - 1024,1024,64 max per dim (Kepler), but product num_threads < 1024; more restrictions pending register limit
+      if ( m_prod_tile_dims > 512 ) { // Match Cuda restriction for ParallelReduce; 1024,1024,64 max per dim (Kepler), but product num_threads < 1024
         printf(" Tile dimensions exceed Cuda limits\n");
         Kokkos::abort(" Cuda ExecSpace Error: MDRange tile dims exceed maximum number of threads per block - choose smaller tile dims");
         //Kokkos::Impl::throw_runtime_exception( " Cuda ExecSpace Error: MDRange tile dims exceed maximum number of threads per block - choose smaller tile dims");
@@ -223,19 +254,7 @@ struct MDRangePolicy
   template < typename LT , typename UT , typename TT = array_index_type >
   MDRangePolicy( std::initializer_list const& lower, std::initializer_list const& upper, std::initializer_list const& tile = {} )
   {
-#if 0
-    // This should work, less duplicated code but not yet extensively tested
-    point_type lower_tmp, upper_tmp;
-    tile_type tile_tmp;
-    for ( auto i = 0; i < rank; ++i ) {
-      lower_tmp[i] = static_cast(lower.begin()[i]);
-      upper_tmp[i] = static_cast(upper.begin()[i]);
-      tile_tmp[i]  = static_cast(tile.begin()[i]);
-    }
 
-    MDRangePolicy( lower_tmp, upper_tmp, tile_tmp );
-
-#else
     if(static_cast(m_lower.size()) != rank || static_cast(m_upper.size()) != rank)
       Kokkos::abort("MDRangePolicy: Constructor initializer lists have wrong size");
 
@@ -249,7 +268,7 @@ struct MDRangePolicy
     }
 
     m_num_tiles = 1;
-
+    m_prod_tile_dims = 1;
 
     // Host
     if ( true
@@ -262,8 +281,8 @@ struct MDRangePolicy
       for (int i=0; i 0)) )
+          if (  ((int)inner_direction == (int)Right && (i < rank-1))
+              || ((int)inner_direction == (int)Left && (i > 0)) )
           {
             m_tile[i] = 2;
           }
@@ -273,6 +292,7 @@ struct MDRangePolicy
         }
         m_tile_end[i] = static_cast((span + m_tile[i] - 1) / m_tile[i]);
         m_num_tiles *= m_tile_end[i];
+        m_prod_tile_dims *= m_tile[i];
       }
     }
     #if defined(KOKKOS_ENABLE_CUDA)
@@ -284,10 +304,14 @@ struct MDRangePolicy
         if ( m_tile[i] <= 0 ) {
           // TODO: determine what is a good default tile size for cuda
           // may be rank dependent
-          if (  (inner_direction == Right && (i < rank-1))
-              || (inner_direction == Left && (i > 0)) )
+          if (  ((int)inner_direction == (int)Right && (i < rank-1))
+              || ((int)inner_direction == (int)Left && (i > 0)) )
           {
-            m_tile[i] = 2;
+            if ( m_prod_tile_dims < 512 ) {
+              m_tile[i] = 2;
+            } else {
+              m_tile[i] = 1;
+            }
           }
           else {
             m_tile[i] = 16;
@@ -295,32 +319,22 @@ struct MDRangePolicy
         }
         m_tile_end[i] = static_cast((span + m_tile[i] - 1) / m_tile[i]);
         m_num_tiles *= m_tile_end[i];
+        m_prod_tile_dims *= m_tile[i];
       }
-      index_type total_tile_size_check = 1;
-      for (int i=0; i= 1024 ) { // improve this check - 1024,1024,64 max per dim (Kepler), but product num_threads < 1024; more restrictions pending register limit
+      if ( m_prod_tile_dims > 512 ) { // Match Cuda restriction for ParallelReduce; 1024,1024,64 max per dim (Kepler), but product num_threads < 1024
         printf(" Tile dimensions exceed Cuda limits\n");
         Kokkos::abort(" Cuda ExecSpace Error: MDRange tile dims exceed maximum number of threads per block - choose smaller tile dims");
         //Kokkos::Impl::throw_runtime_exception( " Cuda ExecSpace Error: MDRange tile dims exceed maximum number of threads per block - choose smaller tile dims");
       }
     }
     #endif
-#endif
   }
 
-
-  point_type m_lower;
-  point_type m_upper;
-  tile_type  m_tile;
-  point_type m_tile_end;
-  index_type m_num_tiles;
 };
 // ------------------------------------------------------------------ //
 
 // ------------------------------------------------------------------ //
-//md_parallel_for
+//md_parallel_for - deprecated use parallel_for
 // ------------------------------------------------------------------ //
 template 
 void md_parallel_for( MDRange const& range
@@ -335,7 +349,6 @@ void md_parallel_for( MDRange const& range
 {
   Impl::MDFunctor g(range, f);
 
-  //using range_policy = typename MDRange::range_policy;
   using range_policy = typename MDRange::impl_range_policy;
 
   Kokkos::parallel_for( range_policy(0, range.m_num_tiles).set_chunk_size(1), g, str );
@@ -354,7 +367,6 @@ void md_parallel_for( const std::string& str
 {
   Impl::MDFunctor g(range, f);
 
-  //using range_policy = typename MDRange::range_policy;
   using range_policy = typename MDRange::impl_range_policy;
 
   Kokkos::parallel_for( range_policy(0, range.m_num_tiles).set_chunk_size(1), g, str );
@@ -395,7 +407,7 @@ void md_parallel_for( MDRange const& range
 // ------------------------------------------------------------------ //
 
 // ------------------------------------------------------------------ //
-//md_parallel_reduce
+//md_parallel_reduce - deprecated use parallel_reduce
 // ------------------------------------------------------------------ //
 template 
 void md_parallel_reduce( MDRange const& range
@@ -409,9 +421,8 @@ void md_parallel_reduce( MDRange const& range
                       ) >::type* = 0
                     )
 {
-  Impl::MDFunctor g(range, f, v);
+  Impl::MDFunctor g(range, f);
 
-  //using range_policy = typename MDRange::range_policy;
   using range_policy = typename MDRange::impl_range_policy;
   Kokkos::parallel_reduce( str, range_policy(0, range.m_num_tiles).set_chunk_size(1), g, v );
 }
@@ -428,48 +439,14 @@ void md_parallel_reduce( const std::string& str
                       ) >::type* = 0
                     )
 {
-  Impl::MDFunctor g(range, f, v);
+  Impl::MDFunctor g(range, f);
 
-  //using range_policy = typename MDRange::range_policy;
   using range_policy = typename MDRange::impl_range_policy;
 
   Kokkos::parallel_reduce( str, range_policy(0, range.m_num_tiles).set_chunk_size(1), g, v );
 }
 
-// Cuda - parallel_reduce not implemented yet
-/*
-template 
-void md_parallel_reduce( MDRange const& range
-                    , Functor const& f
-                    , ValueType & v
-                    , const std::string& str = ""
-                    , typename std::enable_if<( true
-                      #if defined( KOKKOS_ENABLE_CUDA)
-                      && std::is_same< typename MDRange::range_policy::execution_space, Kokkos::Cuda>::value
-                      #endif
-                      ) >::type* = 0
-                    )
-{
-  Impl::DeviceIterateTile closure(range, f, v);
-  closure.execute();
-}
-
-template 
-void md_parallel_reduce( const std::string& str
-                    , MDRange const& range
-                    , Functor const& f
-                    , ValueType & v
-                    , typename std::enable_if<( true
-                      #if defined( KOKKOS_ENABLE_CUDA)
-                      && std::is_same< typename MDRange::range_policy::execution_space, Kokkos::Cuda>::value
-                      #endif
-                      ) >::type* = 0
-                    )
-{
-  Impl::DeviceIterateTile closure(range, f, v);
-  closure.execute();
-}
-*/
+// Cuda - md_parallel_reduce not implemented - use parallel_reduce
 
 }} // namespace Kokkos::Experimental
 
diff --git a/lib/kokkos/core/src/Kokkos_Atomic.hpp b/lib/kokkos/core/src/Kokkos_Atomic.hpp
index 3ecae24da4..3c8673c66a 100644
--- a/lib/kokkos/core/src/Kokkos_Atomic.hpp
+++ b/lib/kokkos/core/src/Kokkos_Atomic.hpp
@@ -114,40 +114,9 @@
 #endif /* Not pre-selected atomic implementation */
 #endif
 
-//----------------------------------------------------------------------------
-
-// Forward decalaration of functions supporting arbitrary sized atomics
-// This is necessary since Kokkos_Atomic.hpp is internally included very early
-// through Kokkos_HostSpace.hpp as well as the allocation tracker.
 #ifdef KOKKOS_ENABLE_CUDA
-namespace Kokkos {
-namespace Impl {
-/// \brief Aquire a lock for the address
-///
-/// This function tries to aquire the lock for the hash value derived
-/// from the provided ptr. If the lock is successfully aquired the
-/// function returns true. Otherwise it returns false.
-#ifdef KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE
-extern
+#include 
 #endif
-__device__ inline
-bool lock_address_cuda_space(void* ptr);
-
-/// \brief Release lock for the address
-///
-/// This function releases the lock for the hash value derived
-/// from the provided ptr. This function should only be called
-/// after previously successfully aquiring a lock with
-/// lock_address.
-#ifdef KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE
-extern
-#endif
-__device__ inline
-void unlock_address_cuda_space(void* ptr);
-}
-}
-#endif
-
 
 namespace Kokkos {
 template 
diff --git a/lib/kokkos/core/src/Kokkos_Concepts.hpp b/lib/kokkos/core/src/Kokkos_Concepts.hpp
index 9a2b53e157..5480dbf40c 100644
--- a/lib/kokkos/core/src/Kokkos_Concepts.hpp
+++ b/lib/kokkos/core/src/Kokkos_Concepts.hpp
@@ -79,6 +79,21 @@ struct IndexType
   using type = T;
 };
 
+/**\brief Specify Launch Bounds for CUDA execution.
+ *
+ *  The "best" defaults may be architecture specific.
+ */
+template< unsigned int maxT = 1024 /* Max threads per block */
+        , unsigned int minB = 1    /* Min blocks per SM */
+        >
+struct LaunchBounds
+{
+  using launch_bounds = LaunchBounds;
+  using type = LaunchBounds;
+  static unsigned int constexpr maxTperB {maxT};
+  static unsigned int constexpr minBperSM {minB};
+};
+
 } // namespace Kokkos
 
 //----------------------------------------------------------------------------
@@ -119,6 +134,7 @@ using Kokkos::is_array_layout ;
 KOKKOS_IMPL_IS_CONCEPT( iteration_pattern )
 KOKKOS_IMPL_IS_CONCEPT( schedule_type )
 KOKKOS_IMPL_IS_CONCEPT( index_type )
+KOKKOS_IMPL_IS_CONCEPT( launch_bounds )
 
 }
 
diff --git a/lib/kokkos/core/src/Kokkos_Core.hpp b/lib/kokkos/core/src/Kokkos_Core.hpp
index 19de791c0f..ddb11d2894 100644
--- a/lib/kokkos/core/src/Kokkos_Core.hpp
+++ b/lib/kokkos/core/src/Kokkos_Core.hpp
@@ -96,11 +96,13 @@ struct InitArguments {
   int num_numa;
   int device_id;
 
-  InitArguments() {
-    num_threads = -1;
-    num_numa = -1;
-    device_id = -1;
-  }
+  InitArguments( int nt = -1
+               , int nn = -1
+               , int dv = -1)
+    : num_threads( nt )
+    , num_numa( nn )
+    , device_id( dv )
+  {}
 };
 
 void initialize(int& narg, char* arg[]);
@@ -168,6 +170,9 @@ void * kokkos_realloc( void * arg_alloc , const size_t arg_alloc_size )
 
 } // namespace Kokkos
 
+#include 
+#include 
+
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
diff --git a/lib/kokkos/core/src/Kokkos_Core_fwd.hpp b/lib/kokkos/core/src/Kokkos_Core_fwd.hpp
index 09081d2387..8c080f7a8f 100644
--- a/lib/kokkos/core/src/Kokkos_Core_fwd.hpp
+++ b/lib/kokkos/core/src/Kokkos_Core_fwd.hpp
@@ -51,6 +51,9 @@
 #include 
 #include 
 
+#include 
+#include 
+
 //----------------------------------------------------------------------------
 // Have assumed a 64bit build (8byte pointers) throughout the code base.
 
diff --git a/lib/kokkos/core/src/Kokkos_Crs.hpp b/lib/kokkos/core/src/Kokkos_Crs.hpp
new file mode 100644
index 0000000000..93b3fa5ca9
--- /dev/null
+++ b/lib/kokkos/core/src/Kokkos_Crs.hpp
@@ -0,0 +1,333 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_CRS_HPP
+#define KOKKOS_CRS_HPP
+
+namespace Kokkos {
+namespace Experimental {
+
+/// \class Crs
+/// \brief Compressed row storage array.
+///
+/// \tparam DataType The type of stored entries.  If a Crs is
+///   used as the graph of a sparse matrix, then this is usually an
+///   integer type, the type of the column indices in the sparse
+///   matrix.
+///
+/// \tparam Arg1Type The second template parameter, corresponding
+///   either to the Device type (if there are no more template
+///   parameters) or to the Layout type (if there is at least one more
+///   template parameter).
+///
+/// \tparam Arg2Type The third template parameter, which if provided
+///   corresponds to the Device type.
+///
+/// \tparam SizeType The type of row offsets.  Usually the default
+///   parameter suffices.  However, setting a nondefault value is
+///   necessary in some cases, for example, if you want to have a
+///   sparse matrices with dimensions (and therefore column indices)
+///   that fit in \c int, but want to store more than INT_MAX
+///   entries in the sparse matrix.
+///
+/// A row has a range of entries:
+/// 
    
+/// 
  •   row_map[i0] <= entry < row_map[i0+1]  
    • 
+/// 
  •   0 <= i1 < row_map[i0+1] - row_map[i0]  
    • 
+/// 
  •   entries( entry ,            i2 , i3 , ... );  
    • 
+/// 
  •   entries( row_map[i0] + i1 , i2 , i3 , ... );  
    • 
+/// 

+template< class DataType,
+          class Arg1Type,
+          class Arg2Type = void,
+          typename SizeType = typename ViewTraits::size_type>
+class Crs {
+protected:
+  typedef ViewTraits traits;
+
+public:
+  typedef DataType                                            data_type;
+  typedef typename traits::array_layout                       array_layout;
+  typedef typename traits::execution_space                    execution_space;
+  typedef typename traits::memory_space                       memory_space;
+  typedef typename traits::device_type                        device_type;
+  typedef SizeType                                            size_type;
+
+  typedef Crs< DataType , Arg1Type , Arg2Type , SizeType > staticcrsgraph_type;
+  typedef Crs< DataType , array_layout , typename traits::host_mirror_space , SizeType > HostMirror;
+  typedef View row_map_type;
+  typedef View entries_type;
+
+  entries_type entries;
+  row_map_type row_map;
+
+  //! Construct an empty view.
+  Crs () : entries(), row_map() {}
+
+  //! Copy constructor (shallow copy).
+  Crs (const Crs& rhs) : entries (rhs.entries), row_map (rhs.row_map)
+  {}
+
+  template
+  Crs (const EntriesType& entries_,const RowMapType& row_map_) : entries (entries_), row_map (row_map_)
+  {}
+
+  /** \brief  Assign to a view of the rhs array.
+   *          If the old view is the last view
+   *          then allocated memory is deallocated.
+   */
+  Crs& operator= (const Crs& rhs) {
+    entries = rhs.entries;
+    row_map = rhs.row_map;
+    return *this;
+  }
+
+  /**  \brief  Destroy this view of the array.
+   *           If the last view then allocated memory is deallocated.
+   */
+  ~Crs() {}
+
+  /**  \brief  Return number of rows in the graph
+   */
+  KOKKOS_INLINE_FUNCTION
+  size_type numRows() const {
+    return (row_map.dimension_0 () != 0) ?
+      row_map.dimension_0 () - static_cast (1) :
+      static_cast (0);
+  }
+};
+
+/*--------------------------------------------------------------------------*/
+
+template< class OutCounts,
+          class DataType,
+          class Arg1Type,
+          class Arg2Type,
+          class SizeType>
+void get_crs_transpose_counts(
+    OutCounts& out,
+    Crs const& in,
+    std::string const& name = "transpose_counts");
+
+template< class OutCounts,
+          class InCrs>
+void get_crs_row_map_from_counts(
+    OutCounts& out,
+    InCrs const& in,
+    std::string const& name = "row_map");
+
+template< class DataType,
+          class Arg1Type,
+          class Arg2Type,
+          class SizeType>
+void transpose_crs(
+    Crs& out,
+    Crs const& in);
+
+}} // namespace Kokkos::Experimental
+
+/*--------------------------------------------------------------------------*/
+
+/*--------------------------------------------------------------------------*/
+
+namespace Kokkos {
+namespace Impl {
+namespace Experimental {
+
+template 
+class GetCrsTransposeCounts {
+ public:
+  using execution_space = typename InCrs::execution_space;
+  using self_type = GetCrsTransposeCounts;
+  using index_type = typename InCrs::size_type;
+ private:
+  InCrs in;
+  OutCounts out;
+ public:
+  KOKKOS_INLINE_FUNCTION
+  void operator()(index_type i) const {
+    atomic_increment( &out[in.entries(i)] );
+  }
+  GetCrsTransposeCounts(InCrs const& arg_in, OutCounts const& arg_out):
+    in(arg_in),out(arg_out) {
+    using policy_type = RangePolicy;
+    using closure_type = Kokkos::Impl::ParallelFor;
+    const closure_type closure(*this, policy_type(0, index_type(in.entries.size())));
+    closure.execute();
+    execution_space::fence();
+  }
+};
+
+template 
+class CrsRowMapFromCounts {
+ public:
+  using execution_space = typename InCounts::execution_space;
+  using value_type = typename OutRowMap::value_type;
+  using index_type = typename InCounts::size_type;
+ private:
+  InCounts in;
+  OutRowMap out;
+ public:
+  KOKKOS_INLINE_FUNCTION
+  void operator()(index_type i, value_type& update, bool final_pass) const {
+    update += in(i);
+    if (final_pass) {
+      out(i + 1) = update;
+      if (i == 0) {
+        out(0) = 0;
+      }
+    }
+  }
+  KOKKOS_INLINE_FUNCTION
+  void init(value_type& update) const { update = 0; }
+  KOKKOS_INLINE_FUNCTION
+  void join(volatile value_type& update, const volatile value_type& input) const {
+    update += input;
+  }
+  using self_type = CrsRowMapFromCounts;
+  CrsRowMapFromCounts(InCounts const& arg_in, OutRowMap const& arg_out):
+    in(arg_in),out(arg_out) {
+    using policy_type = RangePolicy;
+    using closure_type = Kokkos::Impl::ParallelScan;
+    closure_type closure(*this, policy_type(0, in.size()));
+    closure.execute();
+    execution_space::fence();
+  }
+};
+
+template 
+class FillCrsTransposeEntries {
+ public:
+  using execution_space = typename InCrs::execution_space;
+  using memory_space = typename InCrs::memory_space;
+  using value_type = typename OutCrs::entries_type::value_type;
+  using index_type = typename InCrs::size_type;
+ private:
+  using counters_type = View;
+  InCrs in;
+  OutCrs out;
+  counters_type counters;
+ public:
+  KOKKOS_INLINE_FUNCTION
+  void operator()(index_type i) const {
+    auto begin = in.row_map(i);
+    auto end = in.row_map(i + 1);
+    for (auto j = begin; j < end; ++j) {
+      auto ti = in.entries(j);
+      auto tbegin = out.row_map(ti);
+      auto tj = atomic_fetch_add( &counters(ti), 1 );
+      out.entries( tbegin + tj ) = i;
+    }
+  }
+  using self_type = FillCrsTransposeEntries;
+  FillCrsTransposeEntries(InCrs const& arg_in, OutCrs const& arg_out):
+    in(arg_in),out(arg_out),
+    counters("counters", arg_out.numRows()) {
+    using policy_type = RangePolicy;
+    using closure_type = Kokkos::Impl::ParallelFor;
+    const closure_type closure(*this, policy_type(0, index_type(in.numRows())));
+    closure.execute();
+    execution_space::fence();
+  }
+};
+
+}}} // namespace Kokkos::Impl::Experimental
+
+/*--------------------------------------------------------------------------*/
+
+/*--------------------------------------------------------------------------*/
+
+namespace Kokkos {
+namespace Experimental {
+
+template< class OutCounts,
+          class DataType,
+          class Arg1Type,
+          class Arg2Type,
+          class SizeType>
+void get_crs_transpose_counts(
+    OutCounts& out,
+    Crs const& in,
+    std::string const& name) {
+  using InCrs = Crs;
+  out = OutCounts(name, in.numRows());
+  Kokkos::Impl::Experimental::
+    GetCrsTransposeCounts functor(in, out);
+}
+
+template< class OutRowMap,
+          class InCounts>
+void get_crs_row_map_from_counts(
+    OutRowMap& out,
+    InCounts const& in,
+    std::string const& name) {
+  out = OutRowMap(ViewAllocateWithoutInitializing(name), in.size() + 1);
+  Kokkos::Impl::Experimental::
+    CrsRowMapFromCounts functor(in, out);
+}
+
+template< class DataType,
+          class Arg1Type,
+          class Arg2Type,
+          class SizeType>
+void transpose_crs(
+    Crs& out,
+    Crs const& in)
+{
+  typedef Crs crs_type ;
+  typedef typename crs_type::memory_space             memory_space ;
+  typedef View               counts_type ;
+  {
+  counts_type counts;
+  Kokkos::Experimental::get_crs_transpose_counts(counts, in);
+  Kokkos::Experimental::get_crs_row_map_from_counts(out.row_map, counts,
+      "tranpose_row_map");
+  }
+  out.entries = decltype(out.entries)("transpose_entries", in.entries.size());
+  Kokkos::Impl::Experimental::
+    FillCrsTransposeEntries entries_functor(in, out);
+}
+
+}} // namespace Kokkos::Experimental
+
+#endif /* #define KOKKOS_CRS_HPP */
diff --git a/lib/kokkos/core/src/Kokkos_Cuda.hpp b/lib/kokkos/core/src/Kokkos_Cuda.hpp
index f0f0f87458..197831dee5 100644
--- a/lib/kokkos/core/src/Kokkos_Cuda.hpp
+++ b/lib/kokkos/core/src/Kokkos_Cuda.hpp
@@ -217,8 +217,8 @@ public:
 
 private:
 
-  cudaStream_t m_stream ;
   int          m_device ;
+  cudaStream_t m_stream ;
 };
 
 } // namespace Kokkos
@@ -295,6 +295,7 @@ struct VerifyExecutionCanAccessMemorySpace
 #include 
 #include 
 #include 
+#include 
 
 #include 
 //----------------------------------------------------------------------------
diff --git a/lib/kokkos/core/src/Kokkos_CudaSpace.hpp b/lib/kokkos/core/src/Kokkos_CudaSpace.hpp
index 307ab193b1..fb5985e164 100644
--- a/lib/kokkos/core/src/Kokkos_CudaSpace.hpp
+++ b/lib/kokkos/core/src/Kokkos_CudaSpace.hpp
@@ -90,7 +90,7 @@ public:
                  , const size_t arg_alloc_size ) const ;
 
   /**\brief Return Name of the MemorySpace */
-  static constexpr const char* name();
+  static constexpr const char* name() { return m_name; }
 
   /*--------------------------------*/
   /** \brief  Error reporting for HostSpace attempt to access CudaSpace */
@@ -186,7 +186,7 @@ public:
                  , const size_t arg_alloc_size ) const ;
 
   /**\brief Return Name of the MemorySpace */
-  static constexpr const char* name();
+  static constexpr const char* name() { return m_name; }
 
   /*--------------------------------*/
 
@@ -234,7 +234,7 @@ public:
                  , const size_t arg_alloc_size ) const ;
 
   /**\brief Return Name of the MemorySpace */
-  static constexpr const char* name();
+  static constexpr const char* name() { return m_name; }
 
 private:
 
diff --git a/lib/kokkos/core/src/Kokkos_ExecPolicy.hpp b/lib/kokkos/core/src/Kokkos_ExecPolicy.hpp
index 375a2d3744..a8c4d77c62 100644
--- a/lib/kokkos/core/src/Kokkos_ExecPolicy.hpp
+++ b/lib/kokkos/core/src/Kokkos_ExecPolicy.hpp
@@ -384,6 +384,7 @@ Impl::PerThreadValue PerThread(const int& arg);
  *    WorkTag (none): Tag which is used as the first argument for the functor operator.
  *    Schedule (Schedule): Scheduling Policy (Dynamic, or Static).
  *    IndexType (IndexType: Integer Index type used to iterate over the Index space.
+ *    LaunchBounds (LaunchBounds<1024,1>: Launch Bounds for CUDA compilation.
  */
 template< class ... Properties>
 class TeamPolicy: public
@@ -561,6 +562,45 @@ KOKKOS_INLINE_FUNCTION
 Impl::ThreadVectorRangeBoundariesStruct
 ThreadVectorRange( const TeamMemberType&, const iType& count );
 
+#if defined(KOKKOS_ENABLE_PROFILING)
+namespace Impl {
+
+template::value >
+struct ParallelConstructName;
+
+template
+struct ParallelConstructName {
+  ParallelConstructName(std::string const& label):label_ref(label) {
+    if (label.empty()) {
+      default_name = std::string(typeid(FunctorType).name()) + "/" +
+        typeid(TagType).name();
+    }
+  }
+  std::string const& get() {
+    return (label_ref.empty()) ? default_name : label_ref;
+  }
+  std::string const& label_ref;
+  std::string default_name;
+};
+
+template
+struct ParallelConstructName {
+  ParallelConstructName(std::string const& label):label_ref(label) {
+    if (label.empty()) {
+      default_name = std::string(typeid(FunctorType).name());
+    }
+  }
+  std::string const& get() {
+    return (label_ref.empty()) ? default_name : label_ref;
+  }
+  std::string const& label_ref;
+  std::string default_name;
+};
+
+} // namespace Impl
+#endif /* defined KOKKOS_ENABLE_PROFILING */
+
 } // namespace Kokkos
 
 #endif /* #define KOKKOS_EXECPOLICY_HPP */
diff --git a/lib/kokkos/core/src/Kokkos_HBWSpace.hpp b/lib/kokkos/core/src/Kokkos_HBWSpace.hpp
index e224cd4e84..9c9af0dd8b 100644
--- a/lib/kokkos/core/src/Kokkos_HBWSpace.hpp
+++ b/lib/kokkos/core/src/Kokkos_HBWSpace.hpp
@@ -126,14 +126,6 @@ public:
   //! This memory space preferred device_type
   typedef Kokkos::Device< execution_space, memory_space > device_type;
 
-  /*--------------------------------*/
-  /* Functions unique to the HBWSpace */
-  static int in_parallel();
-
-  static void register_in_parallel( int (*)() );
-
-  /*--------------------------------*/
-
   /**\brief  Default memory space instance */
   HBWSpace();
   HBWSpace( const HBWSpace & rhs ) = default;
diff --git a/lib/kokkos/core/src/Kokkos_HostSpace.hpp b/lib/kokkos/core/src/Kokkos_HostSpace.hpp
index d00cce8f60..431635047a 100644
--- a/lib/kokkos/core/src/Kokkos_HostSpace.hpp
+++ b/lib/kokkos/core/src/Kokkos_HostSpace.hpp
@@ -130,14 +130,6 @@ public:
   //! This memory space preferred device_type
   typedef Kokkos::Device< execution_space, memory_space > device_type;
 
-  /*--------------------------------*/
-  /* Functions unique to the HostSpace */
-  static int in_parallel();
-
-  static void register_in_parallel( int (*)() );
-
-  /*--------------------------------*/
-
   /**\brief  Default memory space instance */
   HostSpace();
   HostSpace( HostSpace && rhs ) = default;
@@ -161,7 +153,7 @@ public:
                  , const size_t arg_alloc_size ) const;
 
   /**\brief Return Name of the MemorySpace */
-  static constexpr const char* name();
+  static constexpr const char* name() { return m_name; }
 
 private:
   AllocationMechanism  m_alloc_mech;
diff --git a/lib/kokkos/core/src/Kokkos_Layout.hpp b/lib/kokkos/core/src/Kokkos_Layout.hpp
index f300a6d9f6..87c705153e 100644
--- a/lib/kokkos/core/src/Kokkos_Layout.hpp
+++ b/lib/kokkos/core/src/Kokkos_Layout.hpp
@@ -156,6 +156,8 @@ struct LayoutStride {
       for ( int r = 0 ; r < ARRAY_LAYOUT_MAX_RANK ; ++r ) {
         tmp.dimension[r] = 0 ;
         tmp.stride[r]    = 0 ;
+      }
+      for ( int r = 0 ; r < rank ; ++r ) {
         check_input &= ~int( 1 << order[r] );
       }
       if ( 0 == check_input ) {
diff --git a/lib/kokkos/core/src/Kokkos_Macros.hpp b/lib/kokkos/core/src/Kokkos_Macros.hpp
index 1439dbd3f8..250ef6630a 100644
--- a/lib/kokkos/core/src/Kokkos_Macros.hpp
+++ b/lib/kokkos/core/src/Kokkos_Macros.hpp
@@ -297,6 +297,10 @@
     #endif
   #endif
 
+  #if defined( KOKKOS_ARCH_AVX512MIC )
+      #define KOKKOS_ENABLE_RFO_PREFETCH 1
+  #endif 
+
   #if defined( __MIC__ )
     // Compiling for Xeon Phi
   #endif
@@ -344,13 +348,18 @@
   //#define KOKKOS_ENABLE_PRAGMA_VECTOR 1
   //#define KOKKOS_ENABLE_PRAGMA_SIMD 1
 
+  #if defined( KOKKOS_ARCH_AVX512MIC )
+      #define KOKKOS_ENABLE_RFO_PREFETCH 1
+  #endif
+
   #if !defined( KOKKOS_FORCEINLINE_FUNCTION )
     #define KOKKOS_FORCEINLINE_FUNCTION inline __attribute__((always_inline))
   #endif
 
   #if !defined( KOKKOS_ENABLE_ASM ) && !defined( __PGIC__ ) && \
       ( defined( __amd64 ) || defined( __amd64__ ) || \
-        defined( __x86_64 ) || defined( __x86_64__ ) )
+        defined( __x86_64 ) || defined( __x86_64__ ) || \
+	defined(__PPC64__) )
     #define KOKKOS_ENABLE_ASM 1
   #endif
 #endif
diff --git a/lib/kokkos/core/src/Kokkos_MasterLock.hpp b/lib/kokkos/core/src/Kokkos_MasterLock.hpp
new file mode 100644
index 0000000000..81564b8eac
--- /dev/null
+++ b/lib/kokkos/core/src/Kokkos_MasterLock.hpp
@@ -0,0 +1,73 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_MASTER_LOCK_HPP
+#define KOKKOS_MASTER_LOCK_HPP
+
+#include 
+
+namespace Kokkos { namespace Experimental {
+
+// my be used to coordinate work between master instances
+// SHOULD NOT be used within a parallel algorithm
+//
+// This lock should be used with with a scoped lock guard
+// i.e. std::unique_lock, std::lock_guard
+//
+// cannot be copied or moved
+// has the following functions available
+//
+// Lock()
+// ~Lock()
+//
+// void lock()
+// void unlock()
+// bool try_lock()
+//
+template 
+class MasterLock;
+
+}} // namespace Kokkos::Experimental
+
+#endif //KOKKOS_MASTER_LOCK_HPP
+
diff --git a/lib/kokkos/core/src/Kokkos_MemoryPool.hpp b/lib/kokkos/core/src/Kokkos_MemoryPool.hpp
index dbf1ad8057..1da936067d 100644
--- a/lib/kokkos/core/src/Kokkos_MemoryPool.hpp
+++ b/lib/kokkos/core/src/Kokkos_MemoryPool.hpp
@@ -66,11 +66,6 @@ private:
   enum : uint32_t { max_bit_count_lg2 = CB::max_bit_count_lg2 };
   enum : uint32_t { max_bit_count     = CB::max_bit_count };
 
-  /*  Defaults for min block, max block, and superblock sizes */
-  enum : uint32_t { MIN_BLOCK_SIZE_LG2  =  6  /*   64 bytes */ };
-  enum : uint32_t { MAX_BLOCK_SIZE_LG2  = 12  /*   4k bytes */ };
-  enum : uint32_t { SUPERBLOCK_SIZE_LG2 = 16  /*  64k bytes */ };
-
   enum : uint32_t { HINT_PER_BLOCK_SIZE = 2 };
 
   /*  Each superblock has a concurrent bitset state
@@ -85,6 +80,14 @@ private:
    *  is concurrently updated.
    */
 
+  /*  Mapping between block_size <-> block_state
+   *
+   *  block_state = ( m_sb_size_lg2 - block_size_lg2 ) << state_shift
+   *  block_size  = m_sb_size_lg2 - ( block_state >> state_shift )
+   *
+   *  Thus A_block_size < B_block_size  <=>  A_block_state > B_block_state
+   */
+
   typedef typename DeviceType::memory_space base_memory_space ;
 
   enum { accessible =
@@ -251,10 +254,10 @@ public:
    *  significant runtime performance improvements.
    */
   MemoryPool( const base_memory_space & memspace
-            , const size_t   min_total_alloc_size
-            , const uint32_t min_block_alloc_size // = 1 << MIN_BLOCK_SIZE_LG2
-            , const uint32_t max_block_alloc_size // = 1 << MAX_BLOCK_SIZE_LG2
-            , const uint32_t min_superblock_size  // = 1 << SUPERBLOCK_SIZE_LG2
+            , const size_t min_total_alloc_size
+            , size_t min_block_alloc_size = 0
+            , size_t max_block_alloc_size = 0
+            , size_t min_superblock_size  = 0
             )
     : m_tracker()
     , m_sb_state_array(0)
@@ -267,8 +270,43 @@ public:
     , m_data_offset(0)
     , m_unused_padding(0)
     {
-      const uint32_t int_align_lg2  = 3 ; /* align as int[8] */
-      const uint32_t int_align_mask = ( 1u << int_align_lg2 ) - 1 ;
+      const uint32_t int_align_lg2   = 3 ; /* align as int[8] */
+      const uint32_t int_align_mask  = ( 1u << int_align_lg2 ) - 1 ;
+
+      // Constraints and defaults:
+      //   min_block_alloc_size <= max_block_alloc_size
+      //   max_block_alloc_size <= min_superblock_size 
+      //   min_superblock_size  <= min_total_alloc_size
+
+      const uint32_t MIN_BLOCK_SIZE  = 1u << 6   /*   64 bytes */ ;
+      const uint32_t MAX_BLOCK_SIZE  = 1u << 12  /*   4k bytes */ ;
+
+      if ( 0 == min_block_alloc_size ) min_block_alloc_size = MIN_BLOCK_SIZE ;
+
+      if ( 0 == max_block_alloc_size ) {
+
+        max_block_alloc_size = MAX_BLOCK_SIZE ;
+
+        // Upper bound of total allocation size
+        max_block_alloc_size = std::min( size_t(max_block_alloc_size)
+                                       , min_total_alloc_size );
+
+        // Lower bound of minimum block size
+        max_block_alloc_size = std::max( max_block_alloc_size
+                                       , min_block_alloc_size );
+      }
+
+      if ( 0 == min_superblock_size ) {
+        min_superblock_size = max_block_alloc_size ;
+
+        // Upper bound of total allocation size
+        min_superblock_size = std::min( size_t(min_superblock_size)
+                                      , min_total_alloc_size );
+
+        // Lower bound of maximum block size
+        min_superblock_size = std::max( min_superblock_size
+                                      , max_block_alloc_size );
+      }
 
       // Block and superblock size is power of two:
 
@@ -435,6 +473,8 @@ public:
   void * allocate( size_t alloc_size
                  , int32_t attempt_limit = 1 ) const noexcept
     {
+      if ( 0 == alloc_size ) return (void*) 0 ;
+
       void * p = 0 ;
 
       const uint32_t block_size_lg2 = get_block_size_lg2( alloc_size );
@@ -444,10 +484,9 @@ public:
         // Allocation will fit within a superblock
         // that has block sizes ( 1 << block_size_lg2 )
 
-        const uint32_t block_count_lg2  = m_sb_size_lg2 - block_size_lg2 ;
-        const uint32_t block_state      = block_count_lg2 << state_shift ;
-        const uint32_t block_count      = 1u << block_count_lg2 ;
-        const uint32_t block_count_mask = block_count - 1 ;
+        const uint32_t block_count_lg2 = m_sb_size_lg2 - block_size_lg2 ;
+        const uint32_t block_state     = block_count_lg2 << state_shift ;
+        const uint32_t block_count     = 1u << block_count_lg2 ;
 
         // Superblock hints for this block size:
         //   hint_sb_id_ptr[0] is the dynamically changing hint
@@ -465,7 +504,7 @@ public:
         // the guess for which block within a superblock should
         // be claimed.  If not available then a search occurs.
 
-        const uint32_t block_id_hint = block_count_mask &
+        const uint32_t block_id_hint =
           (uint32_t)( Kokkos::Impl::clock_tic()
 #if defined( KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_CUDA )
           // Spread out potentially concurrent access
@@ -474,6 +513,9 @@ public:
 #endif
           );
 
+        // expected state of superblock for allocation
+        uint32_t sb_state = block_state ;
+
         int32_t sb_id = -1 ;
 
         volatile uint32_t * sb_state_array = 0 ;
@@ -484,6 +526,8 @@ public:
 
           if ( sb_id < 0 ) {
 
+            // No superblock specified, try the hint for this block size
+
             sb_id = hint_sb_id = int32_t( *hint_sb_id_ptr );
 
             sb_state_array = m_sb_state_array + ( sb_id * m_sb_state_size );
@@ -493,16 +537,20 @@ public:
           //   0 <= sb_id
           //   sb_state_array == m_sb_state_array + m_sb_state_size * sb_id
 
-          if ( block_state == ( state_header_mask & *sb_state_array ) ) {
+          if ( sb_state == ( state_header_mask & *sb_state_array ) ) {
 
-            // This superblock state is assigned to this block size.
-            // Try to claim a bit.
+            // This superblock state is as expected, for the moment.
+            // Attempt to claim a bit.  The attempt updates the state
+            // so have already made sure the state header is as expected.
+
+            const uint32_t count_lg2 = sb_state >> state_shift ;
+            const uint32_t mask      = ( 1u << count_lg2 ) - 1 ;
 
             const Kokkos::pair result =
               CB::acquire_bounded_lg2( sb_state_array
-                                     , block_count_lg2
-                                     , block_id_hint
-                                     , block_state
+                                     , count_lg2
+                                     , block_id_hint & mask
+                                     , sb_state
                                      );
 
             // If result.first < 0 then failed to acquire
@@ -512,16 +560,18 @@ public:
 
             if ( 0 <= result.first ) { // acquired a bit
 
+              const uint32_t size_lg2 = m_sb_size_lg2 - count_lg2 ;
+
               // Set the allocated block pointer
 
               p = ((char*)( m_sb_state_array + m_data_offset ))
                 + ( uint32_t(sb_id) << m_sb_size_lg2 ) // superblock memory
-                + ( result.first    << block_size_lg2 ); // block memory
+                + ( result.first    << size_lg2 );     // block memory
 
               break ; // Success
             }
 
-// printf("  acquire block_count_lg2(%d) block_state(0x%x) sb_id(%d) result(%d,%d)\n" , block_count_lg2 , block_state , sb_id , result.first , result.second );
+// printf("  acquire count_lg2(%d) sb_state(0x%x) sb_id(%d) result(%d,%d)\n" , count_lg2 , sb_state , sb_id , result.first , result.second );
 
           }
           //------------------------------------------------------------------
@@ -529,12 +579,18 @@ public:
           //  Must find a new superblock.
 
           //  Start searching at designated index for this block size.
-          //  Look for a partially full superblock of this block size.
-          //  Look for an empty superblock just in case cannot find partfull.
+          //  Look for superblock that, in preferential order,
+          //  1) part-full superblock of this block size
+          //  2) empty superblock to claim for this block size
+          //  3) part-full superblock of the next larger block size
 
+          sb_state = block_state ; // Expect to find the desired state
           sb_id = -1 ;
 
+          bool update_hint = false ;
           int32_t sb_id_empty = -1 ;
+          int32_t sb_id_large = -1 ;
+          uint32_t sb_state_large = 0 ;
 
           sb_state_array = m_sb_state_array + sb_id_begin * m_sb_state_size ;
 
@@ -544,38 +600,54 @@ public:
             //  Note that the state may change at any moment
             //  as concurrent allocations and deallocations occur.
             
-            const uint32_t state = *sb_state_array ;
-            const uint32_t used  = state & state_used_mask ;
+            const uint32_t full_state = *sb_state_array ;
+            const uint32_t used       = full_state & state_used_mask ;
+            const uint32_t state      = full_state & state_header_mask ;
 
-            if ( block_state == ( state & state_header_mask ) ) {
+            if ( state == block_state ) {
 
               //  Superblock is assigned to this block size
 
-              if ( used < block_count ) { 
+              if ( used < block_count ) {
 
                 // There is room to allocate one block
 
                 sb_id = id ;
 
-                if ( used + 1 < block_count ) {
+                // Is there room to allocate more than one block?
 
-                  // There is room to allocate more than one block
-
-                  Kokkos::atomic_compare_exchange
-                    ( hint_sb_id_ptr , uint32_t(hint_sb_id) , uint32_t(sb_id) );
-                }
+                update_hint = used + 1 < block_count ;
 
                 break ;
               }
             }
-            else if ( ( used == 0 ) && ( sb_id_empty == -1 ) ) {
+            else if ( 0 == used ) {
 
-              // Superblock is not assigned to this block size
-              // and is the first empty superblock encountered.
-              // Save this id to use if a partfull superblock is not found.
+              // Superblock is empty
 
-              sb_id_empty = id ;
+              if ( -1 == sb_id_empty ) {
+
+                // Superblock is not assigned to this block size
+                // and is the first empty superblock encountered.
+                // Save this id to use if a partfull superblock is not found.
+
+                sb_id_empty = id ;
+              }
             }
+            else if ( ( -1 == sb_id_empty /* have not found an empty */ ) &&
+                      ( -1 == sb_id_large /* have not found a larger */ ) &&
+                      ( state < block_state /* a larger block */ ) &&
+                      // is not full:
+                      ( used < ( 1u << ( state >> state_shift ) ) ) ) {
+              //  First superblock encountered that is
+              //  larger than this block size and
+              //  has room for an allocation.
+              //  Save this id to use of partfull or empty superblock not found
+              sb_id_large    = id ;
+              sb_state_large = state ;
+            }
+
+            // Iterate around the superblock array:
 
             if ( ++id < m_sb_count ) {
               sb_state_array += m_sb_state_size ;
@@ -586,7 +658,7 @@ public:
             }
           }
 
-// printf("  search m_sb_count(%d) sb_id(%d) sb_id_empty(%d)\n" , m_sb_count , sb_id , sb_id_empty );
+// printf("  search m_sb_count(%d) sb_id(%d) sb_id_empty(%d) sb_id_large(%d)\n" , m_sb_count , sb_id , sb_id_empty , sb_id_large);
 
           if ( sb_id < 0 ) {
 
@@ -609,21 +681,31 @@ public:
 
               const uint32_t state_empty = state_header_mask & *sb_state_array ;
 
-              if ( state_empty ==
-                     Kokkos::atomic_compare_exchange
-                       (sb_state_array,state_empty,block_state) ) {
+              // If this thread claims the empty block then update the hint
+              update_hint =
+                state_empty ==
+                  Kokkos::atomic_compare_exchange
+                    (sb_state_array,state_empty,block_state);
+            }
+            else if ( 0 <= sb_id_large ) {
 
-                // If this thread claimed the block then update the hint
+              // Found a larger superblock with space available
 
-                Kokkos::atomic_compare_exchange
-                  ( hint_sb_id_ptr , uint32_t(hint_sb_id) , uint32_t(sb_id) );
-              }
+              sb_id    = sb_id_large ;
+              sb_state = sb_state_large ;
+
+              sb_state_array = m_sb_state_array + ( sb_id * m_sb_state_size );
             }
             else {
               // Did not find a potentially usable superblock
               --attempt_limit ;
             }
           }
+
+          if ( update_hint ) {
+            Kokkos::atomic_compare_exchange
+              ( hint_sb_id_ptr , uint32_t(hint_sb_id) , uint32_t(sb_id) );
+          }
         } // end allocation attempt loop
 
         //--------------------------------------------------------------------
@@ -646,6 +728,8 @@ public:
   KOKKOS_INLINE_FUNCTION
   void deallocate( void * p , size_t /* alloc_size */ ) const noexcept
     {
+      if ( 0 == p ) return ;
+
       // Determine which superblock and block
       const ptrdiff_t d =
         ((char*)p) - ((char*)( m_sb_state_array + m_data_offset ));
diff --git a/lib/kokkos/core/src/Kokkos_MemoryTraits.hpp b/lib/kokkos/core/src/Kokkos_MemoryTraits.hpp
index 94b58b8aff..af9c8ea782 100644
--- a/lib/kokkos/core/src/Kokkos_MemoryTraits.hpp
+++ b/lib/kokkos/core/src/Kokkos_MemoryTraits.hpp
@@ -72,11 +72,11 @@ struct MemoryTraits {
   //! Tag this class as a kokkos memory traits:
   typedef MemoryTraits memory_traits ;
 
-  enum { Unmanaged    = T & unsigned(Kokkos::Unmanaged) };
-  enum { RandomAccess = T & unsigned(Kokkos::RandomAccess) };
-  enum { Atomic       = T & unsigned(Kokkos::Atomic) };
-  enum { Restrict     = T & unsigned(Kokkos::Restrict) };
-  enum { Aligned      = T & unsigned(Kokkos::Aligned) };
+  enum : bool { Unmanaged    = (unsigned(0) != (T & unsigned(Kokkos::Unmanaged))) };
+  enum : bool { RandomAccess = (unsigned(0) != (T & unsigned(Kokkos::RandomAccess))) };
+  enum : bool { Atomic       = (unsigned(0) != (T & unsigned(Kokkos::Atomic))) };
+  enum : bool { Restrict     = (unsigned(0) != (T & unsigned(Kokkos::Restrict))) };
+  enum : bool { Aligned      = (unsigned(0) != (T & unsigned(Kokkos::Aligned))) };
 
 };
 
@@ -109,7 +109,11 @@ enum { MEMORY_ALIGNMENT =
 #else
     ( 1 << Kokkos::Impl::integral_power_of_two( 128 ) )
 #endif
-  , MEMORY_ALIGNMENT_THRESHOLD = 4 
+#if defined( KOKKOS_MEMORY_ALIGNMENT_THRESHOLD )
+  , MEMORY_ALIGNMENT_THRESHOLD = KOKKOS_MEMORY_ALIGNMENT_THRESHOLD
+#else
+  , MEMORY_ALIGNMENT_THRESHOLD = 4
+#endif
   };
 
 
diff --git a/lib/kokkos/core/src/Kokkos_OpenMP.hpp b/lib/kokkos/core/src/Kokkos_OpenMP.hpp
index 3e11621ce6..d5de01cf2f 100644
--- a/lib/kokkos/core/src/Kokkos_OpenMP.hpp
+++ b/lib/kokkos/core/src/Kokkos_OpenMP.hpp
@@ -47,10 +47,6 @@
 #include 
 #if defined( KOKKOS_ENABLE_OPENMP)
 
-#if !defined(_OPENMP)
-#error "You enabled Kokkos OpenMP support without enabling OpenMP in the compiler!"
-#endif
-
 #include 
 
 #include 
@@ -67,95 +63,144 @@
 #include 
 #include 
 
+#include 
+
 /*--------------------------------------------------------------------------*/
 
 namespace Kokkos {
 
+namespace Impl {
+class OpenMPExec;
+}
+
 /// \class OpenMP
 /// \brief Kokkos device for multicore processors in the host memory space.
 class OpenMP {
 public:
-  //------------------------------------
-  //! \name Type declarations that all Kokkos devices must provide.
-  //@{
-
   //! Tag this class as a kokkos execution space
   using execution_space = OpenMP;
+
+  using memory_space =
   #ifdef KOKKOS_ENABLE_HBWSPACE
-  using memory_space = Experimental::HBWSpace;
+    Experimental::HBWSpace;
   #else
-  using memory_space = HostSpace;
+    HostSpace;
   #endif
+
   //! This execution space preferred device_type
-  using device_type = Kokkos::Device;
-
-  using array_layout = LayoutRight;
-  using size_type = memory_space::size_type;
-
+  using device_type          = Kokkos::Device< execution_space, memory_space >;
+  using array_layout         = LayoutRight;
+  using size_type            = memory_space::size_type;
   using scratch_memory_space = ScratchMemorySpace< OpenMP >;
 
-  //@}
-  //------------------------------------
-  //! \name Functions that all Kokkos execution spaces must implement.
-  //@{
+  /// \brief Get a handle to the default execution space instance
+  inline
+  OpenMP() noexcept;
 
-  inline static bool in_parallel();
+  // Using omp_get_max_threads(); is problematic
+  // On Intel (essentially an initial call to the OpenMP runtime
+  // without a parallel region before will set a process mask for a single core
+  // The runtime will than bind threads for a parallel region to other cores on the
+  // entering the first parallel region and make the process mask the aggregate of
+  // the thread masks. The intend seems to be to make serial code run fast, if you
+  // compile with OpenMP enabled but don't actually use parallel regions or so
+  // static int omp_max_threads = omp_get_max_threads();
+  static int get_current_max_threads() noexcept;
 
-  /** \brief  Set the device in a "sleep" state. A noop for OpenMP.  */
-  static bool sleep();
+  /// \brief Initialize the default execution space
+  ///
+  /// if ( thread_count == -1 )
+  ///   then use the number of threads that openmp defaults to
+  /// if ( thread_count == 0 && Kokkos::hwlow_available() )
+  ///   then use hwloc to choose the number of threads and change
+  ///   the default number of threads
+  /// if ( thread_count > 0 )
+  ///   then force openmp to use the given number of threads and change
+  ///   the default number of threads
+  static void initialize( int thread_count = -1 );
 
-  /** \brief Wake the device from the 'sleep' state. A noop for OpenMP. */
-  static bool wake();
-
-  /** \brief Wait until all dispatched functors complete. A noop for OpenMP. */
-  static void fence() {}
-
-  /// \brief Print configuration information to the given output stream.
-  static void print_configuration( std::ostream & , const bool detail = false );
-
-  /// \brief Free any resources being consumed by the device.
+  /// \brief Free any resources being consumed by the default execution space
   static void finalize();
 
-  /** \brief  Initialize the device.
-   *
-   *  1) If the hardware locality library is enabled and OpenMP has not
-   *     already bound threads then bind OpenMP threads to maximize
-   *     core utilization and group for memory hierarchy locality.
-   *
-   *  2) Allocate a HostThread for each OpenMP thread to hold its
-   *     topology and fan in/out data.
-   */
-  static void initialize( unsigned thread_count = 0 ,
-                          unsigned use_numa_count = 0 ,
-                          unsigned use_cores_per_numa = 0 );
+  /// \brief is the default execution space initialized for current 'master' thread
+  static bool is_initialized() noexcept;
 
-  static int is_initialized();
+  /// \brief Print configuration information to the given output stream.
+  static void print_configuration( std::ostream & , const bool verbose = false );
 
-  /** \brief  Return the maximum amount of concurrency.  */
-  static int concurrency();
+  /// \brief is the instance running a parallel algorithm
+  inline
+  static bool in_parallel( OpenMP const& = OpenMP() ) noexcept;
 
-  //@}
-  //------------------------------------
-  /** \brief  This execution space has a topological thread pool which can be queried.
-   *
-   *  All threads within a pool have a common memory space for which they are cache coherent.
-   *    depth = 0  gives the number of threads in the whole pool.
-   *    depth = 1  gives the number of threads in a NUMA region, typically sharing L3 cache.
-   *    depth = 2  gives the number of threads at the finest granularity, typically sharing L1 cache.
-   */
-  inline static int thread_pool_size( int depth = 0 );
+  /// \brief Wait until all dispatched functors complete on the given instance
+  ///
+  ///  This is a no-op on OpenMP
+  inline
+  static void fence( OpenMP const& = OpenMP() ) noexcept;
+
+  /// \brief Does the given instance return immediately after launching
+  /// a parallel algorithm
+  ///
+  /// This always returns false on OpenMP
+  inline
+  static bool is_asynchronous( OpenMP const& = OpenMP() ) noexcept;
+
+
+  /// \brief Partition the default instance into new instances without creating
+  ///  new masters
+  ///
+  /// This is a no-op on OpenMP since the default instance cannot be partitioned
+  /// without promoting other threads to 'master'
+  static std::vector partition(...);
+
+  /// Non-default instances should be ref-counted so that when the last
+  /// is destroyed the instance resources are released
+  ///
+  /// This is a no-op on OpenMP since a non default instance cannot be created
+  static OpenMP create_instance(...);
+
+  /// \brief Partition the default instance and call 'f' on each new 'master' thread
+  ///
+  /// Func is a functor with the following signiture
+  ///   void( int partition_id, int num_partitions )
+  template 
+  static void partition_master( F const& f
+                              , int requested_num_partitions = 0
+                              , int requested_partition_size = 0
+                              );
+
+  inline
+  static int thread_pool_size() noexcept;
 
   /** \brief  The rank of the executing thread in this thread pool */
-  KOKKOS_INLINE_FUNCTION static int thread_pool_rank();
+  KOKKOS_INLINE_FUNCTION
+  static int thread_pool_rank() noexcept;
 
-  //------------------------------------
+#if !defined( KOKKOS_DISABLE_DEPRECATED )
+  /// \brief Initialize the default execution space
+  static void initialize( int thread_count,
+                          int use_numa_count,
+                          int use_cores_per_numa = 0);
 
-  inline static unsigned max_hardware_threads() { return thread_pool_size(0); }
+  inline
+  static int thread_pool_size( int depth );
 
-  KOKKOS_INLINE_FUNCTION static
-  unsigned hardware_thread_id() { return thread_pool_rank(); }
+  static void sleep() {};
+  static void wake() {};
 
-  static const char* name();
+  // use UniqueToken
+  static int concurrency();
+
+  // use UniqueToken
+  inline
+  static int max_hardware_threads() noexcept;
+
+  // use UniqueToken
+  KOKKOS_INLINE_FUNCTION
+  static int hardware_thread_id() noexcept;
+#endif
+
+  static constexpr const char* name() noexcept { return "OpenMP"; }
 };
 
 } // namespace Kokkos
@@ -195,6 +240,7 @@ struct VerifyExecutionCanAccessMemorySpace
 /*--------------------------------------------------------------------------*/
 
 #include 
+#include 
 #include 
 #include 
 
diff --git a/lib/kokkos/core/src/Kokkos_Parallel.hpp b/lib/kokkos/core/src/Kokkos_Parallel.hpp
index e412e608b2..fc8d6bec81 100644
--- a/lib/kokkos/core/src/Kokkos_Parallel.hpp
+++ b/lib/kokkos/core/src/Kokkos_Parallel.hpp
@@ -177,22 +177,23 @@ void parallel_for( const ExecPolicy  & policy
                  )
 {
 #if defined(KOKKOS_ENABLE_PROFILING)
-    uint64_t kpID = 0;
-     if(Kokkos::Profiling::profileLibraryLoaded()) {
-     	Kokkos::Profiling::beginParallelFor("" == str ? typeid(FunctorType).name() : str, 0, &kpID);
-     }
+  uint64_t kpID = 0;
+  if(Kokkos::Profiling::profileLibraryLoaded()) {
+    Kokkos::Impl::ParallelConstructName name(str);
+    Kokkos::Profiling::beginParallelFor(name.get(), 0, &kpID);
+  }
 #endif
 
-    Kokkos::Impl::shared_allocation_tracking_claim_and_disable();
+    Kokkos::Impl::shared_allocation_tracking_disable();
     Impl::ParallelFor< FunctorType , ExecPolicy > closure( functor , policy );
-    Kokkos::Impl::shared_allocation_tracking_release_and_enable();
+    Kokkos::Impl::shared_allocation_tracking_enable();
 
    closure.execute();
 
 #if defined(KOKKOS_ENABLE_PROFILING)
-     if(Kokkos::Profiling::profileLibraryLoaded()) {
-        Kokkos::Profiling::endParallelFor(kpID);
-     }
+  if(Kokkos::Profiling::profileLibraryLoaded()) {
+    Kokkos::Profiling::endParallelFor(kpID);
+  }
 #endif
 }
 
@@ -210,14 +211,15 @@ void parallel_for( const size_t        work_count
 
 #if defined(KOKKOS_ENABLE_PROFILING)
   uint64_t kpID = 0;
-     if(Kokkos::Profiling::profileLibraryLoaded()) {
-  	Kokkos::Profiling::beginParallelFor("" == str ? typeid(FunctorType).name() : str, 0, &kpID);
-     }
+  if(Kokkos::Profiling::profileLibraryLoaded()) {
+    Kokkos::Impl::ParallelConstructName name(str);
+    Kokkos::Profiling::beginParallelFor(name.get(), 0, &kpID);
+  }
 #endif
 
-  Kokkos::Impl::shared_allocation_tracking_claim_and_disable();
+  Kokkos::Impl::shared_allocation_tracking_disable();
   Impl::ParallelFor< FunctorType , policy > closure( functor , policy(0,work_count) );
-  Kokkos::Impl::shared_allocation_tracking_release_and_enable();
+  Kokkos::Impl::shared_allocation_tracking_enable();
 
   closure.execute();
 
@@ -420,21 +422,22 @@ void parallel_scan( const ExecutionPolicy & policy
 {
 #if defined(KOKKOS_ENABLE_PROFILING)
   uint64_t kpID = 0;
-     if(Kokkos::Profiling::profileLibraryLoaded()) {
-	Kokkos::Profiling::beginParallelScan("" == str ? typeid(FunctorType).name() : str, 0, &kpID);
-     }
+  if(Kokkos::Profiling::profileLibraryLoaded()) {
+    Kokkos::Impl::ParallelConstructName name(str);
+    Kokkos::Profiling::beginParallelScan(name.get(), 0, &kpID);
+  }
 #endif
 
-  Kokkos::Impl::shared_allocation_tracking_claim_and_disable();
+  Kokkos::Impl::shared_allocation_tracking_disable();
   Impl::ParallelScan< FunctorType , ExecutionPolicy > closure( functor , policy );
-  Kokkos::Impl::shared_allocation_tracking_release_and_enable();
+  Kokkos::Impl::shared_allocation_tracking_enable();
 
   closure.execute();
 
 #if defined(KOKKOS_ENABLE_PROFILING)
-     if(Kokkos::Profiling::profileLibraryLoaded()) {
-	Kokkos::Profiling::endParallelScan(kpID);
-     }
+  if(Kokkos::Profiling::profileLibraryLoaded()) {
+    Kokkos::Profiling::endParallelScan(kpID);
+  }
 #endif
 
 }
@@ -453,21 +456,22 @@ void parallel_scan( const size_t        work_count
 
 #if defined(KOKKOS_ENABLE_PROFILING)
   uint64_t kpID = 0;
-     if(Kokkos::Profiling::profileLibraryLoaded()) {
-	Kokkos::Profiling::beginParallelScan("" == str ? typeid(FunctorType).name() : str, 0, &kpID);
-     }
+  if(Kokkos::Profiling::profileLibraryLoaded()) {
+    Kokkos::Impl::ParallelConstructName name(str);
+    Kokkos::Profiling::beginParallelScan(name.get(), 0, &kpID);
+  }
 #endif
 
-  Kokkos::Impl::shared_allocation_tracking_claim_and_disable();
+  Kokkos::Impl::shared_allocation_tracking_disable();
   Impl::ParallelScan< FunctorType , policy > closure( functor , policy(0,work_count) );
-  Kokkos::Impl::shared_allocation_tracking_release_and_enable();
+  Kokkos::Impl::shared_allocation_tracking_enable();
 
   closure.execute();
 
 #if defined(KOKKOS_ENABLE_PROFILING)
-     if(Kokkos::Profiling::profileLibraryLoaded()) {
-	Kokkos::Profiling::endParallelScan(kpID);
-     }
+  if(Kokkos::Profiling::profileLibraryLoaded()) {
+    Kokkos::Profiling::endParallelScan(kpID);
+  }
 #endif
 
 }
diff --git a/lib/kokkos/core/src/Kokkos_Parallel_Reduce.hpp b/lib/kokkos/core/src/Kokkos_Parallel_Reduce.hpp
index 8ea5183e35..9df6d4ba09 100644
--- a/lib/kokkos/core/src/Kokkos_Parallel_Reduce.hpp
+++ b/lib/kokkos/core/src/Kokkos_Parallel_Reduce.hpp
@@ -872,13 +872,14 @@ namespace Impl {
         const FunctorType& functor,
         ReturnType& return_value) {
           #if defined(KOKKOS_ENABLE_PROFILING)
-            uint64_t kpID = 0;
-            if(Kokkos::Profiling::profileLibraryLoaded()) {
-              Kokkos::Profiling::beginParallelReduce("" == label ? typeid(FunctorType).name() : label, 0, &kpID);
-            }
+          uint64_t kpID = 0;
+          if(Kokkos::Profiling::profileLibraryLoaded()) {
+            Kokkos::Impl::ParallelConstructName name(label);
+            Kokkos::Profiling::beginParallelReduce(name.get(), 0, &kpID);
+          }
           #endif
 
-          Kokkos::Impl::shared_allocation_tracking_claim_and_disable();
+          Kokkos::Impl::shared_allocation_tracking_disable();
           #ifdef KOKKOS_IMPL_NEED_FUNCTOR_WRAPPER
           Impl::ParallelReduce
              closure(functor_adaptor::functor(functor),
@@ -890,13 +891,13 @@ namespace Impl {
                      policy,
                      return_value_adapter::return_value(return_value,functor));
           #endif
-          Kokkos::Impl::shared_allocation_tracking_release_and_enable();
+          Kokkos::Impl::shared_allocation_tracking_enable();
           closure.execute();
 
           #if defined(KOKKOS_ENABLE_PROFILING)
-            if(Kokkos::Profiling::profileLibraryLoaded()) {
-              Kokkos::Profiling::endParallelReduce(kpID);
-            }
+          if(Kokkos::Profiling::profileLibraryLoaded()) {
+            Kokkos::Profiling::endParallelReduce(kpID);
+          }
           #endif
         }
 
diff --git a/lib/kokkos/core/src/Kokkos_Serial.hpp b/lib/kokkos/core/src/Kokkos_Serial.hpp
index 73e8ae3030..539761a1f9 100644
--- a/lib/kokkos/core/src/Kokkos_Serial.hpp
+++ b/lib/kokkos/core/src/Kokkos_Serial.hpp
@@ -66,6 +66,7 @@
 
 #include 
 
+#include 
 
 namespace Kokkos {
 
@@ -526,6 +527,7 @@ public:
     }
 };
 
+
 /*--------------------------------------------------------------------------*/
 
 template< class FunctorType , class ... Traits >
@@ -604,6 +606,178 @@ public:
     {}
 };
 
+} // namespace Impl
+} // namespace Kokkos
+
+
+/*--------------------------------------------------------------------------*/
+/*--------------------------------------------------------------------------*/
+/* Parallel patterns for Kokkos::Serial with MDRangePolicy */
+
+namespace Kokkos {
+namespace Impl {
+
+template< class FunctorType , class ... Traits >
+class ParallelFor< FunctorType ,
+                   Kokkos::Experimental::MDRangePolicy< Traits ... > ,
+                   Kokkos::Serial
+                 >
+{
+private:
+
+  typedef Kokkos::Experimental::MDRangePolicy< Traits ... > MDRangePolicy ;
+  typedef typename MDRangePolicy::impl_range_policy Policy ;
+
+  typedef typename Kokkos::Experimental::Impl::HostIterateTile< MDRangePolicy, FunctorType, typename MDRangePolicy::work_tag, void > iterate_type;
+
+  const FunctorType   m_functor ;
+  const MDRangePolicy m_mdr_policy ;
+  const Policy        m_policy ;
+
+  void
+  exec() const
+    {
+      const typename Policy::member_type e = m_policy.end();
+      for ( typename Policy::member_type i = m_policy.begin() ; i < e ; ++i ) {
+        iterate_type( m_mdr_policy, m_functor )( i );
+      }
+    }
+
+public:
+
+  inline
+  void execute() const
+    { this->exec(); }
+
+  inline
+  ParallelFor( const FunctorType   & arg_functor
+             , const MDRangePolicy & arg_policy )
+    : m_functor( arg_functor )
+    , m_mdr_policy(  arg_policy )
+    , m_policy( Policy(0, m_mdr_policy.m_num_tiles).set_chunk_size(1) )
+    {}
+};
+
+
+template< class FunctorType , class ReducerType , class ... Traits >
+class ParallelReduce< FunctorType
+                    , Kokkos::Experimental::MDRangePolicy< Traits ... >
+                    , ReducerType
+                    , Kokkos::Serial
+                    >
+{
+private:
+
+  typedef Kokkos::Experimental::MDRangePolicy< Traits ... > MDRangePolicy ;
+  typedef typename MDRangePolicy::impl_range_policy Policy ;
+
+  typedef typename MDRangePolicy::work_tag                                  WorkTag ;
+
+  typedef Kokkos::Impl::if_c< std::is_same::value, FunctorType, ReducerType> ReducerConditional;
+  typedef typename ReducerConditional::type ReducerTypeFwd;
+
+  typedef typename ReducerTypeFwd::value_type ValueType; 
+
+  typedef FunctorAnalysis< FunctorPatternInterface::REDUCE , Policy , FunctorType > Analysis ;
+
+  typedef Kokkos::Impl::FunctorValueInit<   ReducerTypeFwd , WorkTag >  ValueInit ;
+
+  typedef typename Analysis::pointer_type    pointer_type ;
+  typedef typename Analysis::reference_type  reference_type ;
+
+
+  using iterate_type = typename Kokkos::Experimental::Impl::HostIterateTile< MDRangePolicy
+                                                                           , FunctorType
+                                                                           , WorkTag
+                                                                           , ValueType
+                                                                           >;
+
+
+  const FunctorType   m_functor ;
+  const MDRangePolicy m_mdr_policy ;
+  const Policy        m_policy ;
+  const ReducerType   m_reducer ;
+  const pointer_type  m_result_ptr ;
+
+  inline
+  void
+  exec( reference_type update ) const
+    {
+      const typename Policy::member_type e = m_policy.end();
+      for ( typename Policy::member_type i = m_policy.begin() ; i < e ; ++i ) {
+        iterate_type( m_mdr_policy, m_functor, update )( i );
+      }
+    }
+
+public:
+
+  inline
+  void execute() const
+    {
+      const size_t pool_reduce_size =
+        Analysis::value_size( ReducerConditional::select(m_functor , m_reducer) );
+      const size_t team_reduce_size  = 0 ; // Never shrinks
+      const size_t team_shared_size  = 0 ; // Never shrinks
+      const size_t thread_local_size = 0 ; // Never shrinks
+
+      serial_resize_thread_team_data( pool_reduce_size
+                                    , team_reduce_size
+                                    , team_shared_size
+                                    , thread_local_size );
+
+      HostThreadTeamData & data = *serial_get_thread_team_data();
+
+      pointer_type ptr =
+        m_result_ptr ? m_result_ptr : pointer_type(data.pool_reduce_local());
+
+      reference_type update =
+        ValueInit::init(  ReducerConditional::select(m_functor , m_reducer) , ptr );
+
+      this-> exec( update );
+
+      Kokkos::Impl::FunctorFinal< ReducerTypeFwd , WorkTag >::
+        final(  ReducerConditional::select(m_functor , m_reducer) , ptr );
+    }
+
+  template< class HostViewType >
+  ParallelReduce( const FunctorType  & arg_functor ,
+                  const MDRangePolicy       & arg_policy ,
+                  const HostViewType & arg_result_view ,
+                  typename std::enable_if<
+                               Kokkos::is_view< HostViewType >::value &&
+                              !Kokkos::is_reducer_type::value
+                  ,void*>::type = NULL)
+    : m_functor( arg_functor )
+    , m_mdr_policy( arg_policy )
+    , m_policy( Policy(0, m_mdr_policy.m_num_tiles).set_chunk_size(1) )
+    , m_reducer( InvalidType() )
+    , m_result_ptr( arg_result_view.data() )
+    {
+      static_assert( Kokkos::is_view< HostViewType >::value
+        , "Kokkos::Serial reduce result must be a View" );
+
+      static_assert( std::is_same< typename HostViewType::memory_space , HostSpace >::value
+        , "Kokkos::Serial reduce result must be a View in HostSpace" );
+    }
+
+  inline
+  ParallelReduce( const FunctorType & arg_functor
+                , MDRangePolicy       arg_policy
+                , const ReducerType& reducer )
+    : m_functor( arg_functor )
+    , m_mdr_policy(  arg_policy )
+    , m_policy( Policy(0, m_mdr_policy.m_num_tiles).set_chunk_size(1) )
+    , m_reducer( reducer )
+    , m_result_ptr(  reducer.view().data() )
+    {
+      /*static_assert( std::is_same< typename ViewType::memory_space
+                                      , Kokkos::HostSpace >::value
+        , "Reduction result on Kokkos::OpenMP must be a Kokkos::View in HostSpace" );*/
+    }
+};
+
+
+
 } // namespace Impl
 } // namespace Kokkos
 
@@ -819,6 +993,60 @@ public:
 /*--------------------------------------------------------------------------*/
 /*--------------------------------------------------------------------------*/
 
+namespace Kokkos { namespace Experimental {
+
+template<>
+class UniqueToken< Serial, UniqueTokenScope::Instance>
+{
+public:
+  using execution_space = Serial;
+  using size_type       = int;
+
+  /// \brief create object size for concurrency on the given instance
+  ///
+  /// This object should not be shared between instances
+  UniqueToken( execution_space const& = execution_space() ) noexcept {}
+
+  /// \brief upper bound for acquired values, i.e. 0 <= value < size()
+  inline
+  int size() const noexcept { return 1; }
+
+  /// \brief acquire value such that 0 <= value < size()
+  inline
+  int acquire() const  noexcept { return 0; }
+
+  /// \brief release a value acquired by generate
+  inline
+  void release( int ) const noexcept {}
+};
+
+template<>
+class UniqueToken< Serial, UniqueTokenScope::Global>
+{
+public:
+  using execution_space = Serial;
+  using size_type       = int;
+
+  /// \brief create object size for concurrency on the given instance
+  ///
+  /// This object should not be shared between instances
+  UniqueToken( execution_space const& = execution_space() ) noexcept {}
+
+  /// \brief upper bound for acquired values, i.e. 0 <= value < size()
+  inline
+  int size() const noexcept { return 1; }
+
+  /// \brief acquire value such that 0 <= value < size()
+  inline
+  int acquire() const  noexcept { return 0; }
+
+  /// \brief release a value acquired by generate
+  inline
+  void release( int ) const noexcept {}
+};
+
+}} // namespace Kokkos::Experimental
+
 #include 
 
 #endif // defined( KOKKOS_ENABLE_SERIAL )
diff --git a/lib/kokkos/core/src/Kokkos_TaskScheduler.hpp b/lib/kokkos/core/src/Kokkos_TaskScheduler.hpp
index 7edda7aa75..fcfc91a4ee 100644
--- a/lib/kokkos/core/src/Kokkos_TaskScheduler.hpp
+++ b/lib/kokkos/core/src/Kokkos_TaskScheduler.hpp
@@ -148,7 +148,7 @@ private:
     typename std::conditional< Arg2_is_space , Arg2 , void
     >::type >::type ;
 
-  using task_base  = Impl::TaskBase< Space , ValueType , void > ;
+  using task_base  = Impl::TaskBase< void , void , void > ;
   using queue_type = Impl::TaskQueue< Space > ;
 
   task_base * m_task ;
@@ -293,13 +293,17 @@ public:
   //----------------------------------------
 
   KOKKOS_INLINE_FUNCTION
-  typename task_base::get_return_type
+  int is_ready() const noexcept
+    { return ( 0 == m_task ) || ( ((task_base*) task_base::LockTag) == m_task->m_wait ); }
+
+  KOKKOS_INLINE_FUNCTION
+  const typename Impl::TaskResult< ValueType >::reference_type
   get() const
     {
       if ( 0 == m_task ) {
         Kokkos::abort( "Kokkos:::Future::get ERROR: is_null()");
       }
-      return m_task->get();
+      return Impl::TaskResult< ValueType >::get( m_task );
     }
 };
 
@@ -396,7 +400,7 @@ private:
 
   using track_type = Kokkos::Impl::SharedAllocationTracker ;
   using queue_type = Kokkos::Impl::TaskQueue< ExecSpace > ;
-  using task_base  = Impl::TaskBase< ExecSpace , void , void > ;
+  using task_base  = Impl::TaskBase< void , void , void > ;
 
   track_type   m_track ;
   queue_type * m_queue ;
@@ -464,29 +468,19 @@ public:
 
   KOKKOS_INLINE_FUNCTION
   memory_pool * memory() const noexcept
-    { return m_queue ? m_queue->m_memory : (memory_pool*) 0 ; }
+    { return m_queue ? &( m_queue->m_memory ) : (memory_pool*) 0 ; }
 
   //----------------------------------------
   /**\brief  Allocation size for a spawned task */
   template< typename FunctorType >
   KOKKOS_FUNCTION
   size_t spawn_allocation_size() const
-    {
-      using task_type  = Impl::TaskBase< execution_space
-                                       , typename FunctorType::value_type
-                                       , FunctorType > ;
-
-      return m_queue->allocate_block_size( sizeof(task_type) );
-    }
+    { return m_queue->template spawn_allocation_size< FunctorType >(); }
 
   /**\brief  Allocation size for a when_all aggregate */
   KOKKOS_FUNCTION
   size_t when_all_allocation_size( int narg ) const
-    {
-      using task_base  = Kokkos::Impl::TaskBase< ExecSpace , void , void > ;
-
-      return m_queue->allocate_block_size( sizeof(task_base) + narg * sizeof(task_base*) );
-    }
+    { return m_queue->when_all_allocation_size( narg ); }
 
   //----------------------------------------
 
@@ -507,7 +501,7 @@ public:
       queue_type * const queue =
         arg_policy.m_scheduler ? arg_policy.m_scheduler->m_queue : (
         arg_policy.m_dependence.m_task
-          ? arg_policy.m_dependence.m_task->m_queue
+          ? static_cast(arg_policy.m_dependence.m_task->m_queue)
           : (queue_type*) 0 );
 
       if ( 0 == queue ) {
@@ -530,8 +524,12 @@ public:
       future_type f ;
 
       // Allocate task from memory pool
+
+      const size_t alloc_size =
+        queue->template spawn_allocation_size< FunctorType >();
+
       f.m_task =
-        reinterpret_cast< task_type * >(queue->allocate(sizeof(task_type)));
+        reinterpret_cast< task_type * >(queue->allocate(alloc_size) );
 
       if ( f.m_task ) {
 
@@ -539,15 +537,17 @@ public:
         // Reference count starts at two:
         //   +1 for the matching decrement when task is complete
         //   +1 for the future
-        new ( f.m_task )
-          task_type( arg_function
-                   , queue
-                   , arg_policy.m_dependence.m_task /* dependence */
-                   , 2                              /* reference count */
-                   , int(sizeof(task_type))         /* allocation size */
-                   , int(arg_policy.m_task_type)
-                   , int(arg_policy.m_priority)
-                   , std::move(arg_functor) );
+        new ( f.m_task ) task_type( std::move(arg_functor) );
+
+        f.m_task->m_apply      = arg_function ;
+        f.m_task->m_queue      = queue ;
+        f.m_task->m_next       = arg_policy.m_dependence.m_task ;
+        f.m_task->m_ref_count  = 2 ;
+        f.m_task->m_alloc_size = alloc_size ;
+        f.m_task->m_task_type  = arg_policy.m_task_type ;
+        f.m_task->m_priority   = arg_policy.m_priority ;
+
+        Kokkos::memory_fence();
 
         // The dependence (if any) is processed immediately
         // within the schedule function, as such the dependence's
@@ -586,6 +586,30 @@ public:
       // Postcondition: task is in Executing-Respawn state
     }
 
+  template< typename FunctorType >
+  KOKKOS_FUNCTION static
+  void
+  respawn( FunctorType         * arg_self
+         , TaskScheduler const &
+         , TaskPriority  const & arg_priority
+         )
+    {
+      // Precondition: task is in Executing state
+
+      using value_type  = typename FunctorType::value_type ;
+      using task_type   = Impl::TaskBase< execution_space
+                                        , value_type
+                                        , FunctorType > ;
+
+      task_type * const task = static_cast< task_type * >( arg_self );
+
+      task->m_priority = static_cast(arg_priority);
+
+      task->add_dependence( (task_base*) 0 );
+
+      // Postcondition: task is in Executing-Respawn state
+    }
+
   //----------------------------------------
   /**\brief  Return a future that is complete
    *         when all input futures are complete.
@@ -596,7 +620,7 @@ public:
   when_all( Future< A1 , A2 > const arg[] , int narg )
     {
       using future_type = Future< execution_space > ;
-      using task_base   = Kokkos::Impl::TaskBase< execution_space , void , void > ;
+      using task_base   = Kokkos::Impl::TaskBase< void , void , void > ;
 
       future_type f ;
 
@@ -610,9 +634,9 @@ public:
             // Increment reference count to track subsequent assignment.
             Kokkos::atomic_increment( &(t->m_ref_count) );
             if ( queue == 0 ) {
-              queue = t->m_queue ;
+              queue = static_cast< queue_type * >( t->m_queue );
             }
-            else if ( queue != t->m_queue ) {
+            else if ( queue != static_cast< queue_type * >( t->m_queue ) ) {
               Kokkos::abort("Kokkos when_all Futures must be in the same scheduler" );
             }
           }
@@ -620,28 +644,34 @@ public:
 
         if ( queue != 0 ) {
 
-          size_t const size  = sizeof(task_base) + narg * sizeof(task_base*);
+          size_t const alloc_size = queue->when_all_allocation_size( narg );
 
           f.m_task =
-            reinterpret_cast< task_base * >( queue->allocate( size ) );
+            reinterpret_cast< task_base * >( queue->allocate( alloc_size ) );
 
           if ( f.m_task ) {
 
             // Reference count starts at two:
             // +1 to match decrement when task completes
             // +1 for the future
-            new( f.m_task ) task_base( queue
-                                     , 2     /* reference count */
-                                     , size  /* allocation size */
-                                     , narg  /* dependence count */
-                                     );
+
+            new( f.m_task ) task_base();
+
+            f.m_task->m_queue      = queue ;
+            f.m_task->m_ref_count  = 2 ;
+            f.m_task->m_alloc_size = alloc_size ;
+            f.m_task->m_dep_count  = narg ;
+            f.m_task->m_task_type  = task_base::Aggregate ;
 
             // Assign dependences, reference counts were already incremented
 
-            task_base ** const dep = f.m_task->aggregate_dependences();
+            task_base * volatile * const dep =
+              f.m_task->aggregate_dependences();
 
             for ( int i = 0 ; i < narg ; ++i ) { dep[i] = arg[i].m_task ; }
 
+            Kokkos::memory_fence();
+
             queue->schedule_aggregate( f.m_task );
             // this when_all may be processed at any moment
           }
diff --git a/lib/kokkos/core/src/Kokkos_UniqueToken.hpp b/lib/kokkos/core/src/Kokkos_UniqueToken.hpp
new file mode 100644
index 0000000000..1ffb07a6db
--- /dev/null
+++ b/lib/kokkos/core/src/Kokkos_UniqueToken.hpp
@@ -0,0 +1,88 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_UNIQUE_TOKEN_HPP
+#define KOKKOS_UNIQUE_TOKEN_HPP
+
+#include 
+
+namespace Kokkos { namespace Experimental {
+
+enum class UniqueTokenScope : int
+{
+  Instance,
+  Global
+};
+
+/// \brief class to generate unique ids base on the required amount of concurrency
+///
+/// This object should behave like a ref-counted object, so that when the last
+/// instance is destroy resources are free if needed
+template 
+class UniqueToken
+{
+public:
+  using execution_space = ExecutionSpace;
+  using size_type       = typename execution_space::size_type;
+
+  /// \brief create object size for concurrency on the given instance
+  ///
+  /// This object should not be shared between instances
+  UniqueToken( execution_space const& = execution_space() );
+
+  /// \brief upper bound for acquired values, i.e. 0 <= value < size()
+  KOKKOS_INLINE_FUNCTION
+  size_type size() const ;
+
+  /// \brief acquire value such that 0 <= value < size()
+  KOKKOS_INLINE_FUNCTION
+  size_type acquire() const ;
+
+  /// \brief release a value acquired by generate
+  KOKKOS_INLINE_FUNCTION
+  void release( size_type ) const ;
+};
+
+}} // namespace Kokkos::Experimental
+
+#endif //KOKKOS_UNIQUE_TOKEN_HPP
diff --git a/lib/kokkos/core/src/Kokkos_View.hpp b/lib/kokkos/core/src/Kokkos_View.hpp
index 3312aa6a96..1754e4a8fb 100644
--- a/lib/kokkos/core/src/Kokkos_View.hpp
+++ b/lib/kokkos/core/src/Kokkos_View.hpp
@@ -1,13 +1,13 @@
 /*
 //@HEADER
 // ************************************************************************
-// 
+//
 //                        Kokkos v. 2.0
 //              Copyright (2014) Sandia Corporation
-// 
+//
 // Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
 // the U.S. Government retains certain rights in this software.
-// 
+//
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
@@ -36,7 +36,7 @@
 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 //
 // Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
-// 
+//
 // ************************************************************************
 //@HEADER
 */
@@ -54,11 +54,14 @@
 #include 
 #include 
 
+#if defined(KOKKOS_ENABLE_PROFILING)
+#include 
+#endif
+
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
-namespace Experimental {
 namespace Impl {
 
 template< class DataType >
@@ -73,16 +76,6 @@ struct ViewDataAnalysis ;
 template< class , class ... >
 class ViewMapping { public: enum { is_assignable = false }; };
 
-} /* namespace Impl */
-} /* namespace Experimental */
-} /* namespace Kokkos */
-
-namespace Kokkos {
-namespace Impl {
-
-using Kokkos::Experimental::Impl::ViewMapping ;
-using Kokkos::Experimental::Impl::ViewDataAnalysis ;
-
 } /* namespace Impl */
 } /* namespace Kokkos */
 
@@ -1563,12 +1556,12 @@ namespace Kokkos {
 namespace Impl {
 
 inline
-void shared_allocation_tracking_claim_and_disable()
-{ Kokkos::Impl::SharedAllocationRecord::tracking_claim_and_disable(); }
+void shared_allocation_tracking_disable()
+{ Kokkos::Impl::SharedAllocationRecord::tracking_disable(); }
 
 inline
-void shared_allocation_tracking_release_and_enable()
-{ Kokkos::Impl::SharedAllocationRecord::tracking_release_and_enable(); }
+void shared_allocation_tracking_enable()
+{ Kokkos::Impl::SharedAllocationRecord::tracking_enable(); }
 
 } /* namespace Impl */
 } /* namespace Kokkos */
@@ -1795,6 +1788,20 @@ void deep_copy
 
   if ( (void *) dst.data() != (void*) src.data() ) {
 
+#if defined(KOKKOS_ENABLE_PROFILING)
+    if (Kokkos::Profiling::profileLibraryLoaded()) {
+      const size_t nbytes = sizeof(typename dst_type::value_type) * dst.span();
+      Kokkos::Profiling::beginDeepCopy(
+          Kokkos::Profiling::SpaceHandle(dst_memory_space::name()),
+          dst.label(),
+          dst.data(),
+          Kokkos::Profiling::SpaceHandle(src_memory_space::name()),
+          src.label(),
+          src.data(),
+          nbytes);
+    }
+#endif
+
     // Concern: If overlapping views then a parallel copy will be erroneous.
     // ...
 
@@ -1882,7 +1889,14 @@ void deep_copy
     else {
       Kokkos::Impl::throw_runtime_exception("deep_copy given views that would require a temporary allocation");
     }
-  }
+
+#if defined(KOKKOS_ENABLE_PROFILING)
+    if (Kokkos::Profiling::profileLibraryLoaded()) {
+      Kokkos::Profiling::endDeepCopy();
+    }
+#endif
+
+  } // ( (void *) dst.data() != (void*) src.data() )
 }
 
 } /* namespace Kokkos */
@@ -2249,6 +2263,82 @@ resize( Kokkos::View & v ,
 
   static_assert( Kokkos::ViewTraits::is_managed , "Can only resize managed views" );
 
+  // Fix #904 by checking dimensions before actually resizing.
+  //
+  // Rank is known at compile time, so hopefully the compiler will
+  // remove branches that are compile-time false.  The upcoming "if
+  // constexpr" language feature would make this certain.
+  if (view_type::Rank == 1 &&
+      n0 == static_cast (v.extent(0))) {
+    return;
+  }
+  if (view_type::Rank == 2 &&
+      n0 == static_cast (v.extent(0)) &&
+      n1 == static_cast (v.extent(1))) {
+    return;
+  }
+  if (view_type::Rank == 3 &&
+      n0 == static_cast (v.extent(0)) &&
+      n1 == static_cast (v.extent(1)) &&
+      n2 == static_cast (v.extent(2))) {
+    return;
+  }
+  if (view_type::Rank == 4 &&
+      n0 == static_cast (v.extent(0)) &&
+      n1 == static_cast (v.extent(1)) &&
+      n2 == static_cast (v.extent(2)) &&
+      n3 == static_cast (v.extent(3))) {
+    return;
+  }
+  if (view_type::Rank == 5 &&
+      n0 == static_cast (v.extent(0)) &&
+      n1 == static_cast (v.extent(1)) &&
+      n2 == static_cast (v.extent(2)) &&
+      n3 == static_cast (v.extent(3)) &&
+      n4 == static_cast (v.extent(4))) {
+    return;
+  }
+  if (view_type::Rank == 6 &&
+      n0 == static_cast (v.extent(0)) &&
+      n1 == static_cast (v.extent(1)) &&
+      n2 == static_cast (v.extent(2)) &&
+      n3 == static_cast (v.extent(3)) &&
+      n4 == static_cast (v.extent(4)) &&
+      n5 == static_cast (v.extent(5))) {
+    return;
+  }
+  if (view_type::Rank == 7 &&
+      n0 == static_cast (v.extent(0)) &&
+      n1 == static_cast (v.extent(1)) &&
+      n2 == static_cast (v.extent(2)) &&
+      n3 == static_cast (v.extent(3)) &&
+      n4 == static_cast (v.extent(4)) &&
+      n5 == static_cast (v.extent(5)) &&
+      n6 == static_cast (v.extent(6))) {
+    return;
+  }
+  if (view_type::Rank == 8 &&
+      n0 == static_cast (v.extent(0)) &&
+      n1 == static_cast (v.extent(1)) &&
+      n2 == static_cast (v.extent(2)) &&
+      n3 == static_cast (v.extent(3)) &&
+      n4 == static_cast (v.extent(4)) &&
+      n5 == static_cast (v.extent(5)) &&
+      n6 == static_cast (v.extent(6)) &&
+      n7 == static_cast (v.extent(7))) {
+    return;
+  }
+  // If Kokkos ever supports Views of rank > 8, the above code won't
+  // be incorrect, because avoiding reallocation in resize() is just
+  // an optimization.
+
+  // TODO (mfh 27 Jun 2017) If the old View has enough space but just
+  // different dimensions (e.g., if the product of the dimensions,
+  // including extra space for alignment, will not change), then
+  // consider just reusing storage.  For now, Kokkos always
+  // reallocates if any of the dimensions change, even if the old View
+  // has enough space.
+
   view_type v_resized( v.label(), n0, n1, n2, n3, n4, n5, n6, n7 );
 
   Kokkos::Impl::ViewRemap< view_type , view_type >( v_resized , v );
@@ -2317,6 +2407,106 @@ void realloc(      Kokkos::View & v ,
 }
 } /* namespace Kokkos */
 
+//----------------------------------------------------------------------------
+//----------------------------------------------------------------------------
+
+namespace Kokkos { namespace Impl {
+
+template < class Specialize, typename A, typename B >
+struct CommonViewValueType;
+
+template < typename A, typename B >
+struct CommonViewValueType< void, A, B >
+{
+  using value_type = typename std::common_type< A , B >::type;
+};
+
+
+template < class Specialize, class ValueType >
+struct CommonViewAllocProp;
+
+template < class ValueType >
+struct CommonViewAllocProp< void, ValueType >
+{
+  using value_type = ValueType;
+
+  template < class ... Views >
+  CommonViewAllocProp( const Views & ... ) {}
+};
+
+
+template < class ... Views >
+struct DeduceCommonViewAllocProp;
+
+// Base case must provide types for:
+// 1. specialize  2. value_type  3. is_view  4. prop_type
+template < class FirstView >
+struct DeduceCommonViewAllocProp< FirstView >
+{
+  using specialize = typename FirstView::traits::specialize;
+
+  using value_type = typename FirstView::traits::value_type;
+
+  enum : bool { is_view = is_view< FirstView >::value };
+
+  using prop_type = CommonViewAllocProp< specialize, value_type >;
+};
+
+
+template < class FirstView, class ... NextViews >
+struct DeduceCommonViewAllocProp< FirstView, NextViews... >
+{
+  using NextTraits = DeduceCommonViewAllocProp< NextViews... >;
+
+  using first_specialize = typename FirstView::traits::specialize;
+  using first_value_type = typename FirstView::traits::value_type;
+
+  enum : bool { first_is_view = is_view< FirstView >::value };
+
+  using next_specialize = typename NextTraits::specialize;
+  using next_value_type = typename NextTraits::value_type;
+
+  enum : bool { next_is_view = NextTraits::is_view };
+
+  // common types
+
+  // determine specialize type
+  // if first and next specialize differ, but are not the same specialize, error out
+  static_assert( !(!std::is_same< first_specialize, next_specialize >::value && !std::is_same< first_specialize, void>::value && !std::is_same< void, next_specialize >::value)  , "Kokkos DeduceCommonViewAllocProp ERROR: Only one non-void specialize trait allowed" );
+
+  // otherwise choose non-void specialize if either/both are non-void
+  using specialize = typename std::conditional< std::is_same< first_specialize, next_specialize >::value
+                                              , first_specialize
+                                              , typename std::conditional< ( std::is_same< first_specialize, void >::value
+                                                                             && !std::is_same< next_specialize, void >::value)
+                                                                           , next_specialize
+                                                                           , first_specialize
+                                                                         >::type
+                                               >::type;
+
+  using value_type = typename CommonViewValueType< specialize, first_value_type, next_value_type >::value_type;
+
+  enum : bool { is_view = (first_is_view && next_is_view) };
+
+  using prop_type = CommonViewAllocProp< specialize, value_type >;
+};
+
+} // end namespace Impl
+
+template < class ... Views >
+using DeducedCommonPropsType = typename Impl::DeduceCommonViewAllocProp::prop_type ;
+
+// User function
+template < class ... Views >
+DeducedCommonPropsType 
+common_view_alloc_prop( Views const & ... views )
+{
+  return DeducedCommonPropsType( views... );
+}
+
+} // namespace Kokkos
+
+
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 // For backward compatibility:
@@ -2350,6 +2540,9 @@ using Kokkos::Impl::WithoutInitializing_t ;
 using Kokkos::Impl::AllowPadding_t ;
 using Kokkos::Impl::SharedAllocationRecord ;
 using Kokkos::Impl::SharedAllocationTracker ;
+using Kokkos::Impl::ViewMapping ;
+using Kokkos::Impl::ViewDataAnalysis ;
+
 
 } /* namespace Impl */
 } /* namespace Experimental */
diff --git a/lib/kokkos/core/src/Kokkos_WorkGraphPolicy.hpp b/lib/kokkos/core/src/Kokkos_WorkGraphPolicy.hpp
new file mode 100644
index 0000000000..58b0f72f51
--- /dev/null
+++ b/lib/kokkos/core/src/Kokkos_WorkGraphPolicy.hpp
@@ -0,0 +1,265 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_WORKGRAPHPOLICY_HPP
+#define KOKKOS_WORKGRAPHPOLICY_HPP
+
+namespace Kokkos {
+namespace Impl {
+namespace Experimental {
+
+template< class functor_type , class execution_space, class ... policy_args >
+class WorkGraphExec;
+
+}}} // namespace Kokkos::Impl::Experimental
+
+namespace Kokkos {
+namespace Experimental {
+
+template< class ... Properties >
+class WorkGraphPolicy
+{
+public:
+
+  using self_type = WorkGraphPolicy;
+  using traits = Kokkos::Impl::PolicyTraits;
+  using index_type = typename traits::index_type;
+  using execution_space = typename traits::execution_space;
+  using work_tag = typename traits::work_tag;
+  using memory_space = typename execution_space::memory_space;
+  using graph_type = Kokkos::Experimental::Crs;
+  using member_type = index_type;
+
+private:
+   
+  graph_type m_graph;
+
+  using ints_type = Kokkos::View;
+  using range_type = Kokkos::pair;
+  using ranges_type = Kokkos::View;
+  const std::int32_t m_total_work;
+  ints_type m_counts;
+  ints_type m_queue;
+  ranges_type m_ranges;
+
+public:
+
+  struct TagZeroRanges {};
+  KOKKOS_INLINE_FUNCTION
+  void operator()(TagZeroRanges, std::int32_t i) const {
+    m_ranges[i] = range_type(0, 0);
+  }
+  void zero_ranges() {
+    using policy_type = RangePolicy;
+    using closure_type = Kokkos::Impl::ParallelFor;
+    const closure_type closure(*this, policy_type(0, 1));
+    closure.execute();
+    execution_space::fence();
+  }
+
+  struct TagFillQueue {};
+  KOKKOS_INLINE_FUNCTION
+  void operator()(TagFillQueue, std::int32_t i) const {
+    if (*((volatile std::int32_t*)(&m_counts(i))) == 0) push_work(i);
+  }
+  void fill_queue() {
+    using policy_type = RangePolicy;
+    using closure_type = Kokkos::Impl::ParallelFor;
+    const closure_type closure(*this, policy_type(0, m_total_work));
+    closure.execute();
+    execution_space::fence();
+  }
+
+private:
+
+  inline
+  void setup() {
+    if (m_graph.numRows() > std::numeric_limits::max()) {
+      Kokkos::abort("WorkGraphPolicy work must be indexable using int32_t");
+    }
+    get_crs_transpose_counts(m_counts, m_graph);
+    m_queue = ints_type(ViewAllocateWithoutInitializing("queue"), m_total_work);
+    deep_copy(m_queue, std::int32_t(-1));
+    m_ranges = ranges_type("ranges", 1);
+    fill_queue();
+  }
+
+  KOKKOS_INLINE_FUNCTION
+  std::int32_t pop_work() const {
+    range_type w(-1,-1);
+    while (true) {
+      const range_type w_new( w.first + 1 , w.second );
+      w = atomic_compare_exchange( &m_ranges(0) , w , w_new );
+      if ( w.first < w.second ) { // there was work in the queue
+        if ( w_new.first == w.first + 1 && w_new.second == w.second ) {
+          // we got a work item
+          std::int32_t i;
+          // the push_work function may have incremented the end counter
+          // but not yet written the work index into the queue.
+          // wait until the entry is valid.
+          while ( -1 == ( i = *((volatile std::int32_t*)(&m_queue( w.first ))) ) );
+          return i;
+        } // we got a work item
+      } else { // there was no work in the queue
+#ifdef KOKKOS_DEBUG
+        if ( w_new.first == w.first + 1 && w_new.second == w.second ) {
+          Kokkos::abort("bug in pop_work");
+        }
+#endif
+        if (w.first == m_total_work) { // all work is done
+          return -1;
+        } else { // need to wait for more work to be pushed
+          // take a guess that one work item will be pushed
+          // the key thing is we can't leave (w) alone, because
+          // otherwise the next compare_exchange may succeed in
+          // popping work from an empty queue
+          w.second++;
+        }
+      } // there was no work in the queue
+    } // while (true)
+  }
+
+  KOKKOS_INLINE_FUNCTION
+  void push_work(std::int32_t i) const {
+    range_type w(-1,-1);
+    while (true) {
+      const range_type w_new( w.first , w.second + 1 );
+      // try to increment the end counter
+      w = atomic_compare_exchange( &m_ranges(0) , w , w_new );
+      // stop trying if the increment was successful
+      if ( w.first == w_new.first && w.second + 1 == w_new.second ) break;
+    }
+    // write the work index into the claimed spot in the queue
+    *((volatile std::int32_t*)(&m_queue( w.second ))) = i;
+    // push this write out into the memory system
+    memory_fence();
+  }
+
+  template< class functor_type , class execution_space, class ... policy_args >
+  friend class Kokkos::Impl::Experimental::WorkGraphExec;
+
+public:
+
+  WorkGraphPolicy(graph_type arg_graph)
+    : m_graph(arg_graph)
+    , m_total_work( arg_graph.numRows() )
+  {
+    setup();
+  }
+
+};
+
+}} // namespace Kokkos::Experimental
+
+/*--------------------------------------------------------------------------*/
+
+/*--------------------------------------------------------------------------*/
+
+namespace Kokkos {
+namespace Impl {
+namespace Experimental {
+
+template< class functor_type , class execution_space, class ... policy_args >
+class WorkGraphExec
+{
+ public:
+
+  using self_type = WorkGraphExec< functor_type, execution_space, policy_args ... >;
+  using policy_type = Kokkos::Experimental::WorkGraphPolicy< policy_args ... >;
+  using member_type = typename policy_type::member_type;
+  using memory_space = typename execution_space::memory_space;
+
+ protected:
+
+  const functor_type m_functor;
+  const policy_type  m_policy;
+
+ protected:
+
+  KOKKOS_INLINE_FUNCTION
+  std::int32_t before_work() const {
+    return m_policy.pop_work();
+  }
+
+  KOKKOS_INLINE_FUNCTION
+  void after_work(std::int32_t i) const {
+    /* fence any writes that were done by the work item itself
+       (usually writing its result to global memory) */
+    memory_fence();
+    const std::int32_t begin = m_policy.m_graph.row_map( i );
+    const std::int32_t end = m_policy.m_graph.row_map( i + 1 );
+    for (std::int32_t j = begin; j < end; ++j) {
+      const std::int32_t next = m_policy.m_graph.entries( j );
+      const std::int32_t old_count = atomic_fetch_add( &(m_policy.m_counts(next)), -1 );
+      if ( old_count == 1 )  m_policy.push_work( next );
+    }
+  }
+
+  inline
+  WorkGraphExec( const functor_type & arg_functor
+               , const policy_type  & arg_policy )
+    : m_functor( arg_functor )
+    , m_policy(  arg_policy )
+  {
+  }
+};
+
+}}} // namespace Kokkos::Impl::Experimental
+
+#ifdef KOKKOS_ENABLE_SERIAL
+#include "impl/Kokkos_Serial_WorkGraphPolicy.hpp"
+#endif
+
+#ifdef KOKKOS_ENABLE_OPENMP
+#include "OpenMP/Kokkos_OpenMP_WorkGraphPolicy.hpp"
+#endif
+
+#ifdef KOKKOS_ENABLE_CUDA
+#include "Cuda/Kokkos_Cuda_WorkGraphPolicy.hpp"
+#endif
+
+#ifdef KOKKOS_ENABLE_THREADS
+#include "Threads/Kokkos_Threads_WorkGraphPolicy.hpp"
+#endif
+
+#endif /* #define KOKKOS_WORKGRAPHPOLICY_HPP */
diff --git a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Exec.cpp b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Exec.cpp
index 4e0ea93920..915fbe52c1 100644
--- a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Exec.cpp
+++ b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Exec.cpp
@@ -45,75 +45,100 @@
 #if defined( KOKKOS_ENABLE_OPENMP )
 
 #include 
+#include 
+
 #include 
 #include 
 #include 
+
 #include 
+
 #include 
-#include 
 #include 
 #include 
 
 
 namespace Kokkos {
 namespace Impl {
-namespace {
 
-KOKKOS_INLINE_FUNCTION
-int kokkos_omp_in_parallel();
+int g_openmp_hardware_max_threads = 1;
 
-int kokkos_omp_in_critical_region = ( Kokkos::HostSpace::register_in_parallel( kokkos_omp_in_parallel ) , 0 );
+__thread int t_openmp_hardware_id = 0;
+__thread Impl::OpenMPExec * t_openmp_instance = nullptr;
 
-KOKKOS_INLINE_FUNCTION
-int kokkos_omp_in_parallel()
+void OpenMPExec::validate_partition( const int nthreads
+                                   , int & num_partitions
+                                   , int & partition_size
+                                  )
 {
-#ifndef __CUDA_ARCH__
-  return omp_in_parallel() && ! kokkos_omp_in_critical_region ;
-#else
-  return 0;
-#endif
+  if (nthreads == 1) {
+    num_partitions = 1;
+    partition_size = 1;
+  }
+  else if( num_partitions < 1 && partition_size < 1) {
+    int idle = nthreads;
+    for (int np = 2; np <= nthreads ; ++np) {
+      for (int ps = 1; ps <= nthreads/np; ++ps) {
+        if (nthreads - np*ps < idle) {
+          idle = nthreads - np*ps;
+          num_partitions = np;
+          partition_size = ps;
+        }
+        if (idle == 0) {
+          break;
+        }
+      }
+    }
+  }
+  else if( num_partitions < 1 && partition_size > 0 ) {
+    if ( partition_size <= nthreads ) {
+      num_partitions = nthreads / partition_size;
+    }
+    else {
+      num_partitions = 1;
+      partition_size = nthreads;
+    }
+  }
+  else if( num_partitions > 0 && partition_size < 1 ) {
+    if ( num_partitions <= nthreads ) {
+      partition_size = nthreads / num_partitions;
+    }
+    else {
+      num_partitions = nthreads;
+      partition_size = 1;
+    }
+  }
+  else if ( num_partitions * partition_size > nthreads ) {
+    int idle = nthreads;
+    const int NP = num_partitions;
+    const int PS = partition_size;
+    for (int np = NP; np > 0; --np) {
+      for (int ps = PS; ps > 0; --ps) {
+        if (  (np*ps <= nthreads)
+           && (nthreads - np*ps < idle) ) {
+          idle = nthreads - np*ps;
+          num_partitions = np;
+          partition_size = ps;
+        }
+        if (idle == 0) {
+          break;
+        }
+      }
+    }
+  }
+
 }
 
-bool s_using_hwloc = false;
-
-} // namespace
-} // namespace Impl
-} // namespace Kokkos
-
-
-namespace Kokkos {
-namespace Impl {
-
-int OpenMPExec::m_map_rank[ OpenMPExec::MAX_THREAD_COUNT ] = { 0 };
-
-int OpenMPExec::m_pool_topo[ 4 ] = { 0 };
-
-HostThreadTeamData * OpenMPExec::m_pool[ OpenMPExec::MAX_THREAD_COUNT ] = { 0 };
-
-void OpenMPExec::verify_is_process( const char * const label )
+void OpenMPExec::verify_is_master( const char * const label )
 {
-  if ( omp_in_parallel() ) {
+  if ( !t_openmp_instance )
+  {
     std::string msg( label );
-    msg.append( " ERROR: in parallel" );
+    msg.append( " ERROR: in parallel or not initialized" );
     Kokkos::Impl::throw_runtime_exception( msg );
   }
 }
 
-void OpenMPExec::verify_initialized( const char * const label )
-{
-  if ( 0 == m_pool[0] ) {
-    std::string msg( label );
-    msg.append( " ERROR: not initialized" );
-    Kokkos::Impl::throw_runtime_exception( msg );
-  }
-
-  if ( omp_get_max_threads() != Kokkos::OpenMP::thread_pool_size(0) ) {
-    std::string msg( label );
-    msg.append( " ERROR: Initialized but threads modified inappropriately" );
-    Kokkos::Impl::throw_runtime_exception( msg );
-  }
-
-}
 
 } // namespace Impl
 } // namespace Kokkos
@@ -133,11 +158,11 @@ void OpenMPExec::clear_thread_data()
   const int old_alloc_bytes =
     m_pool[0] ? ( member_bytes + m_pool[0]->scratch_bytes() ) : 0 ;
 
-  Kokkos::HostSpace space ;
+  OpenMP::memory_space space ;
 
-#pragma omp parallel
+  #pragma omp parallel num_threads( m_pool_size )
   {
-    const int rank = m_map_rank[ omp_get_thread_num() ];
+    const int rank = omp_get_thread_num();
 
     if ( 0 != m_pool[rank] ) {
 
@@ -189,13 +214,13 @@ void OpenMPExec::resize_thread_data( size_t pool_reduce_bytes
                                       , team_shared_bytes
                                       , thread_local_bytes );
 
-    const int pool_size = omp_get_max_threads();
+    OpenMP::memory_space space ;
 
-    Kokkos::HostSpace space ;
+    memory_fence();
 
-#pragma omp parallel
+    #pragma omp parallel num_threads(m_pool_size)
     {
-      const int rank = m_map_rank[ omp_get_thread_num() ];
+      const int rank = omp_get_thread_num();
 
       if ( 0 != m_pool[rank] ) {
 
@@ -214,11 +239,14 @@ void OpenMPExec::resize_thread_data( size_t pool_reduce_bytes
                       , pool_reduce_bytes
                       , team_reduce_bytes
                       , team_shared_bytes
-                      , thread_local_bytes );
+                      , thread_local_bytes
+                      );
+
+      memory_fence();
     }
 /* END #pragma omp parallel */
 
-    HostThreadTeamData::organize_pool( m_pool , pool_size );
+    HostThreadTeamData::organize_pool( m_pool , m_pool_size );
   }
 }
 
@@ -232,16 +260,8 @@ namespace Kokkos {
 
 //----------------------------------------------------------------------------
 
-int OpenMP::is_initialized()
-{ return 0 != Impl::OpenMPExec::m_pool[0]; }
-
-void OpenMP::initialize( unsigned thread_count ,
-                         unsigned use_numa_count ,
-                         unsigned use_cores_per_numa )
+int OpenMP::get_current_max_threads() noexcept
 {
-  // Before any other call to OMP query the maximum number of threads
-  // and save the value for re-initialization unit testing.
-
   // Using omp_get_max_threads(); is problematic in conjunction with
   // Hwloc on Intel (essentially an initial call to the OpenMP runtime
   // without a parallel region before will set a process mask for a single core
@@ -250,110 +270,99 @@ void OpenMP::initialize( unsigned thread_count ,
   // the thread masks. The intend seems to be to make serial code run fast, if you
   // compile with OpenMP enabled but don't actually use parallel regions or so
   // static int omp_max_threads = omp_get_max_threads();
-  int nthreads = 0;
+
+  int count = 0;
   #pragma omp parallel
   {
     #pragma omp atomic
-    nthreads++;
+     ++count;
   }
+  return count;
+}
 
-  static int omp_max_threads = nthreads;
-
-  const bool is_initialized = 0 != Impl::OpenMPExec::m_pool[0] ;
-
-  bool thread_spawn_failed = false ;
-
-  if ( ! is_initialized ) {
-
-    // Use hwloc thread pinning if concerned with locality.
-    // If spreading threads across multiple NUMA regions.
-    // If hyperthreading is enabled.
-    Impl::s_using_hwloc = hwloc::available() && (
-                            ( 1 < Kokkos::hwloc::get_available_numa_count() ) ||
-                            ( 1 < Kokkos::hwloc::get_available_threads_per_core() ) );
-
-    std::pair threads_coord[ Impl::OpenMPExec::MAX_THREAD_COUNT ];
-
-    // If hwloc available then use it's maximum value.
-
-    if ( thread_count == 0 ) {
-      thread_count = Impl::s_using_hwloc
-      ? Kokkos::hwloc::get_available_numa_count() *
-        Kokkos::hwloc::get_available_cores_per_numa() *
-        Kokkos::hwloc::get_available_threads_per_core()
-      : omp_max_threads ;
-    }
-
-    if(Impl::s_using_hwloc)
-      hwloc::thread_mapping( "Kokkos::OpenMP::initialize" ,
-                           false /* do not allow asynchronous */ ,
-                           thread_count ,
-                           use_numa_count ,
-                           use_cores_per_numa ,
-                           threads_coord );
-
-    // Spawn threads:
-
-    omp_set_num_threads( thread_count );
-
-    // Verify OMP interaction:
-    if ( int(thread_count) != omp_get_max_threads() ) {
-      thread_spawn_failed = true ;
-    }
-
-    // Verify spawning and bind threads:
-#pragma omp parallel
-    {
-#pragma omp critical
-      {
-        if ( int(thread_count) != omp_get_num_threads() ) {
-          thread_spawn_failed = true ;
-        }
-
-        // Call to 'bind_this_thread' is not thread safe so place this whole block in a critical region.
-        // Call to 'new' may not be thread safe as well.
-
-        const unsigned omp_rank    = omp_get_thread_num();
-        const unsigned thread_r    = Impl::s_using_hwloc && Kokkos::hwloc::can_bind_threads()
-                                   ? Kokkos::hwloc::bind_this_thread( thread_count , threads_coord )
-                                   : omp_rank ;
-
-        Impl::OpenMPExec::m_map_rank[ omp_rank ] = thread_r ;
-      }
-/* END #pragma omp critical */
-    }
-/* END #pragma omp parallel */
-
-    if ( ! thread_spawn_failed ) {
-      Impl::OpenMPExec::m_pool_topo[0] = thread_count ;
-      Impl::OpenMPExec::m_pool_topo[1] = Impl::s_using_hwloc ? thread_count / use_numa_count : thread_count;
-      Impl::OpenMPExec::m_pool_topo[2] = Impl::s_using_hwloc ? thread_count / ( use_numa_count * use_cores_per_numa ) : 1;
-
-      // New, unified host thread team data:
-      {
-        size_t pool_reduce_bytes  =   32 * thread_count ;
-        size_t team_reduce_bytes  =   32 * thread_count ;
-        size_t team_shared_bytes  = 1024 * thread_count ;
-        size_t thread_local_bytes = 1024 ;
-
-        Impl::OpenMPExec::resize_thread_data( pool_reduce_bytes
-                                            , team_reduce_bytes
-                                            , team_shared_bytes
-                                            , thread_local_bytes
-                                            );
-      }
-    }
-  }
-
-  if ( is_initialized || thread_spawn_failed ) {
-    std::string msg("Kokkos::OpenMP::initialize ERROR");
-
-    if ( is_initialized ) { msg.append(" : already initialized"); }
-    if ( thread_spawn_failed ) { msg.append(" : failed spawning threads"); }
 
+void OpenMP::initialize( int thread_count )
+{
+  if ( omp_in_parallel() ) {
+    std::string msg("Kokkos::OpenMP::initialize ERROR : in parallel");
     Kokkos::Impl::throw_runtime_exception(msg);
   }
 
+  if ( Impl::t_openmp_instance )
+  {
+    finalize();
+  }
+
+  {
+    if (nullptr == std::getenv("OMP_PROC_BIND") ) {
+      printf("Kokkos::OpenMP::initialize WARNING: OMP_PROC_BIND environment variable not set\n");
+      printf("  In general, for best performance with OpenMP 4.0 or better set OMP_PROC_BIND=spread and OMP_PLACES=threads\n");
+      printf("  For best performance with OpenMP 3.1 set OMP_PROC_BIND=true\n");
+      printf("  For unit testing set OMP_PROC_BIND=false\n");
+    }
+
+    OpenMP::memory_space space ;
+
+    // Before any other call to OMP query the maximum number of threads
+    // and save the value for re-initialization unit testing.
+
+    Impl::g_openmp_hardware_max_threads = get_current_max_threads();
+
+    int process_num_threads = Impl::g_openmp_hardware_max_threads;
+
+    if ( Kokkos::hwloc::available() ) {
+      process_num_threads = Kokkos::hwloc::get_available_numa_count()
+                          * Kokkos::hwloc::get_available_cores_per_numa()
+                          * Kokkos::hwloc::get_available_threads_per_core();
+    }
+
+    // if thread_count  < 0, use g_openmp_hardware_max_threads;
+    // if thread_count == 0, set g_openmp_hardware_max_threads to process_num_threads
+    // if thread_count  > 0, set g_openmp_hardware_max_threads to thread_count
+    if (thread_count < 0 ) {
+      thread_count = Impl::g_openmp_hardware_max_threads;
+    }
+    else if( thread_count == 0 && Impl::g_openmp_hardware_max_threads != process_num_threads ) {
+      Impl::g_openmp_hardware_max_threads = process_num_threads;
+      omp_set_num_threads(Impl::g_openmp_hardware_max_threads);
+    }
+    else {
+      if( thread_count > process_num_threads ) {
+        printf( "Kokkos::OpenMP::initialize WARNING: You are likely oversubscribing your CPU cores.\n");
+        printf( "  process threads available : %3d,  requested thread : %3d\n", process_num_threads, thread_count );
+      }
+      Impl::g_openmp_hardware_max_threads = thread_count;
+      omp_set_num_threads(Impl::g_openmp_hardware_max_threads);
+    }
+
+    // setup thread local
+    #pragma omp parallel num_threads(Impl::g_openmp_hardware_max_threads)
+    {
+      Impl::t_openmp_instance = nullptr;
+      Impl::t_openmp_hardware_id = omp_get_thread_num();
+      Impl::SharedAllocationRecord< void, void >::tracking_enable();
+    }
+
+    void * const ptr = space.allocate( sizeof(Impl::OpenMPExec) );
+
+    Impl::t_openmp_instance = new (ptr) Impl::OpenMPExec( Impl::g_openmp_hardware_max_threads );
+
+    // New, unified host thread team data:
+    {
+      size_t pool_reduce_bytes  =   32 * thread_count ;
+      size_t team_reduce_bytes  =   32 * thread_count ;
+      size_t team_shared_bytes  = 1024 * thread_count ;
+      size_t thread_local_bytes = 1024 ;
+
+      Impl::t_openmp_instance->resize_thread_data( pool_reduce_bytes
+                                                 , team_reduce_bytes
+                                                 , team_shared_bytes
+                                                 , thread_local_bytes
+                                                 );
+    }
+  }
+
+
   // Check for over-subscription
   //if( Impl::mpi_ranks_per_node() * long(thread_count) > Impl::processors_per_node() ) {
   //  std::cout << "Kokkos::OpenMP::initialize WARNING: You are likely oversubscribing your CPU cores." << std::endl;
@@ -373,20 +382,38 @@ void OpenMP::initialize( unsigned thread_count ,
 
 void OpenMP::finalize()
 {
-  Impl::OpenMPExec::verify_initialized( "OpenMP::finalize" );
-  Impl::OpenMPExec::verify_is_process( "OpenMP::finalize" );
+  if ( omp_in_parallel() )
+  {
+    std::string msg("Kokkos::OpenMP::finalize ERROR ");
+    if( !Impl::t_openmp_instance ) msg.append(": not initialized");
+    if( omp_in_parallel() ) msg.append(": in parallel");
+    Kokkos::Impl::throw_runtime_exception(msg);
+  }
 
-  // New, unified host thread team data:
-  Impl::OpenMPExec::clear_thread_data();
+  if ( Impl::t_openmp_instance ) {
 
-  Impl::OpenMPExec::m_pool_topo[0] = 0 ;
-  Impl::OpenMPExec::m_pool_topo[1] = 0 ;
-  Impl::OpenMPExec::m_pool_topo[2] = 0 ;
+    const int nthreads = Impl::t_openmp_instance->m_pool_size <= Impl::g_openmp_hardware_max_threads
+                       ? Impl::g_openmp_hardware_max_threads
+                       : Impl::t_openmp_instance->m_pool_size;
 
-  omp_set_num_threads(1);
+    using Exec = Impl::OpenMPExec;
+    Exec * instance = Impl::t_openmp_instance;
+    instance->~Exec();
 
-  if ( Impl::s_using_hwloc && Kokkos::hwloc::can_bind_threads() ) {
-    hwloc::unbind_this_thread();
+    OpenMP::memory_space space;
+    space.deallocate( instance, sizeof(Exec) );
+
+    #pragma omp parallel num_threads(nthreads)
+    {
+      Impl::t_openmp_hardware_id = 0;
+      Impl::t_openmp_instance    = nullptr;
+      Impl::SharedAllocationRecord< void, void >::tracking_disable();
+    }
+
+    // allow main thread to track
+    Impl::SharedAllocationRecord< void, void >::tracking_enable();
+
+    Impl::g_openmp_hardware_max_threads = 1;
   }
 
   #if defined(KOKKOS_ENABLE_PROFILING)
@@ -396,70 +423,48 @@ void OpenMP::finalize()
 
 //----------------------------------------------------------------------------
 
-void OpenMP::print_configuration( std::ostream & s , const bool detail )
+void OpenMP::print_configuration( std::ostream & s , const bool verbose )
 {
-  Impl::OpenMPExec::verify_is_process( "OpenMP::print_configuration" );
-
   s << "Kokkos::OpenMP" ;
 
-#if defined( KOKKOS_ENABLE_OPENMP )
-  s << " KOKKOS_ENABLE_OPENMP" ;
-#endif
-#if defined( KOKKOS_ENABLE_HWLOC )
-
-  const unsigned numa_count_       = Kokkos::hwloc::get_available_numa_count();
-  const unsigned cores_per_numa   = Kokkos::hwloc::get_available_cores_per_numa();
-  const unsigned threads_per_core = Kokkos::hwloc::get_available_threads_per_core();
-
-  s << " hwloc[" << numa_count_ << "x" << cores_per_numa << "x" << threads_per_core << "]"
-    << " hwloc_binding_" << ( Impl::s_using_hwloc ? "enabled" : "disabled" )
-    ;
-#endif
-
-  const bool is_initialized = 0 != Impl::OpenMPExec::m_pool[0] ;
+  const bool is_initialized =  Impl::t_openmp_instance != nullptr;
 
   if ( is_initialized ) {
-    const int numa_count      = Kokkos::Impl::OpenMPExec::m_pool_topo[0] / Kokkos::Impl::OpenMPExec::m_pool_topo[1] ;
-    const int core_per_numa   = Kokkos::Impl::OpenMPExec::m_pool_topo[1] / Kokkos::Impl::OpenMPExec::m_pool_topo[2] ;
-    const int thread_per_core = Kokkos::Impl::OpenMPExec::m_pool_topo[2] ;
+    Impl::OpenMPExec::verify_is_master( "OpenMP::print_configuration" );
+
+    const int numa_count      = 1;
+    const int core_per_numa   = Impl::g_openmp_hardware_max_threads;
+    const int thread_per_core = 1;
 
     s << " thread_pool_topology[ " << numa_count
       << " x " << core_per_numa
       << " x " << thread_per_core
       << " ]"
       << std::endl ;
-
-    if ( detail ) {
-      std::vector< std::pair > coord( Kokkos::Impl::OpenMPExec::m_pool_topo[0] );
-
-#pragma omp parallel
-      {
-#pragma omp critical
-        {
-          coord[ omp_get_thread_num() ] = hwloc::get_this_thread_coordinate();
-        }
-/* END #pragma omp critical */
-      }
-/* END #pragma omp parallel */
-
-      for ( unsigned i = 0 ; i < coord.size() ; ++i ) {
-        s << "  thread omp_rank[" << i << "]"
-          << " kokkos_rank[" << Impl::OpenMPExec::m_map_rank[ i ] << "]"
-          << " hwloc_coord[" << coord[i].first << "." << coord[i].second << "]"
-          << std::endl ;
-      }
-    }
   }
   else {
     s << " not initialized" << std::endl ;
   }
 }
 
+std::vector OpenMP::partition(...)
+{ return std::vector(1); }
+
+OpenMP OpenMP::create_instance(...) { return OpenMP(); }
+
+
+#if !defined( KOKKOS_DISABLE_DEPRECATED )
+
 int OpenMP::concurrency() {
-  return thread_pool_size(0);
+  return Impl::g_openmp_hardware_max_threads;
 }
 
-const char* OpenMP::name() { return "OpenMP"; }
+void OpenMP::initialize( int thread_count , int, int )
+{
+  initialize(thread_count);
+}
+
+#endif
 
 } // namespace Kokkos
 
diff --git a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Exec.hpp b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Exec.hpp
index 75b7f5da4a..37d2ac8318 100644
--- a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Exec.hpp
+++ b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Exec.hpp
@@ -47,6 +47,10 @@
 #include 
 #if defined( KOKKOS_ENABLE_OPENMP )
 
+#if !defined(_OPENMP)
+#error "You enabled Kokkos OpenMP support without enabling OpenMP in the compiler!"
+#endif
+
 #include 
 
 #include 
@@ -54,6 +58,8 @@
 
 #include 
 
+#include 
+
 #include 
 #include 
 #include 
@@ -63,8 +69,14 @@
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
-namespace Kokkos {
-namespace Impl {
+namespace Kokkos { namespace Impl {
+
+class OpenMPExec;
+
+extern int g_openmp_hardware_max_threads;
+
+extern __thread int t_openmp_hardware_id;
+extern __thread OpenMPExec * t_openmp_instance;
 
 //----------------------------------------------------------------------------
 /** \brief  Data for OpenMP thread execution */
@@ -74,279 +86,279 @@ public:
 
   friend class Kokkos::OpenMP ;
 
-  enum { MAX_THREAD_COUNT = 4096 };
+  enum { MAX_THREAD_COUNT = 512 };
+
+  void clear_thread_data();
+
+  static void validate_partition( const int nthreads
+                                , int & num_partitions
+                                , int & partition_size
+                                );
 
 private:
+  OpenMPExec( int arg_pool_size )
+    : m_pool_size{ arg_pool_size }
+    , m_level{ omp_get_level() }
+    , m_pool()
+  {}
 
-  static int          m_pool_topo[ 4 ];
-  static int          m_map_rank[ MAX_THREAD_COUNT ];
+  ~OpenMPExec()
+  {
+    clear_thread_data();
+  }
 
-  static HostThreadTeamData * m_pool[ MAX_THREAD_COUNT ];
+  int m_pool_size;
+  int m_level;
 
-  static
-  void clear_thread_data();
+  HostThreadTeamData * m_pool[ MAX_THREAD_COUNT ];
 
 public:
 
-  // Topology of a cache coherent thread pool:
-  //   TOTAL = NUMA x GRAIN
-  //   pool_size( depth = 0 )
-  //   pool_size(0) = total number of threads
-  //   pool_size(1) = number of threads per NUMA
-  //   pool_size(2) = number of threads sharing finest grain memory hierarchy
+  static void verify_is_master( const char * const );
 
-  inline static
-  int pool_size( int depth = 0 ) { return m_pool_topo[ depth ]; }
-
-  static void finalize();
-
-  static void initialize( const unsigned team_count ,
-                          const unsigned threads_per_team ,
-                          const unsigned numa_count ,
-                          const unsigned cores_per_numa );
-
-  static void verify_is_process( const char * const );
-  static void verify_initialized( const char * const );
-
-
-  static
   void resize_thread_data( size_t pool_reduce_bytes
                          , size_t team_reduce_bytes
                          , size_t team_shared_bytes
                          , size_t thread_local_bytes );
 
-  inline static
-  HostThreadTeamData * get_thread_data() noexcept
-    { return m_pool[ m_map_rank[ omp_get_thread_num() ] ]; }
+  inline
+  HostThreadTeamData * get_thread_data() const noexcept
+  { return m_pool[ m_level == omp_get_level() ? 0 : omp_get_thread_num() ]; }
 
-  inline static
-  HostThreadTeamData * get_thread_data( int i ) noexcept
-    { return m_pool[i]; }
+  inline
+  HostThreadTeamData * get_thread_data( int i ) const noexcept
+  { return m_pool[i]; }
 };
 
-} // namespace Impl
-} // namespace Kokkos
-
-//----------------------------------------------------------------------------
-//----------------------------------------------------------------------------
-
-namespace Kokkos {
-namespace Impl {
-
-template< class ... Properties >
-class TeamPolicyInternal< Kokkos::OpenMP, Properties ... >: public PolicyTraits
-{
-public:
-
-  //! Tag this class as a kokkos execution policy
-  typedef TeamPolicyInternal      execution_policy ;
-
-  typedef PolicyTraits traits;
-
-  TeamPolicyInternal& operator = (const TeamPolicyInternal& p) {
-    m_league_size = p.m_league_size;
-    m_team_size = p.m_team_size;
-    m_team_alloc = p.m_team_alloc;
-    m_team_iter = p.m_team_iter;
-    m_team_scratch_size[0] = p.m_team_scratch_size[0];
-    m_thread_scratch_size[0] = p.m_thread_scratch_size[0];
-    m_team_scratch_size[1] = p.m_team_scratch_size[1];
-    m_thread_scratch_size[1] = p.m_thread_scratch_size[1];
-    m_chunk_size = p.m_chunk_size;
-    return *this;
-  }
-
-  //----------------------------------------
-
-  template< class FunctorType >
-  inline static
-  int team_size_max( const FunctorType & ) {
-      int pool_size = traits::execution_space::thread_pool_size(1);
-      int max_host_team_size =  Impl::HostThreadTeamData::max_team_members;
-      return pool_size
-  inline static
-  int team_size_recommended( const FunctorType & )
-    { return traits::execution_space::thread_pool_size(2); }
-
-  template< class FunctorType >
-  inline static
-  int team_size_recommended( const FunctorType &, const int& )
-    { return traits::execution_space::thread_pool_size(2); }
-
-  //----------------------------------------
-
-private:
-
-  int m_league_size ;
-  int m_team_size ;
-  int m_team_alloc ;
-  int m_team_iter ;
-
-  size_t m_team_scratch_size[2];
-  size_t m_thread_scratch_size[2];
-
-  int m_chunk_size;
-
-  inline void init( const int league_size_request
-                  , const int team_size_request )
-    {
-      const int pool_size  = traits::execution_space::thread_pool_size(0);
-      const int max_host_team_size =  Impl::HostThreadTeamData::max_team_members;
-      const int team_max   = pool_size 0) {
-      if(!Impl::is_integral_power_of_two( m_chunk_size ))
-        Kokkos::abort("TeamPolicy blocking granularity must be power of two" );
-    }
-
-    int new_chunk_size = 1;
-    while(new_chunk_size*100*concurrency < m_league_size)
-      new_chunk_size *= 2;
-    if(new_chunk_size < 128) {
-      new_chunk_size = 1;
-      while( (new_chunk_size*40*concurrency < m_league_size ) && (new_chunk_size<128) )
-        new_chunk_size*=2;
-    }
-    m_chunk_size = new_chunk_size;
-  }
-
-public:
-  typedef Impl::HostThreadTeamMember< Kokkos::OpenMP > member_type ;
-};
-} // namespace Impl
-
-} // namespace Kokkos
+}} // namespace Kokkos::Impl
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
 
-inline
-bool OpenMP::in_parallel()
-{ return omp_in_parallel(); }
+inline OpenMP::OpenMP() noexcept
+{}
 
 inline
-int OpenMP::thread_pool_size( int depth )
+bool OpenMP::is_initialized() noexcept
+{ return Impl::t_openmp_instance != nullptr; }
+
+inline
+bool OpenMP::in_parallel( OpenMP const& ) noexcept
 {
-  return Impl::OpenMPExec::pool_size(depth);
+  //t_openmp_instance is only non-null on a master thread
+  return   !Impl::t_openmp_instance
+         || Impl::t_openmp_instance->m_level < omp_get_level()
+         ;
+}
+
+inline
+int OpenMP::thread_pool_size() noexcept
+{
+  return   OpenMP::in_parallel()
+         ? omp_get_num_threads()
+         : Impl::t_openmp_instance->m_pool_size
+         ;
 }
 
 KOKKOS_INLINE_FUNCTION
-int OpenMP::thread_pool_rank()
+int OpenMP::thread_pool_rank() noexcept
 {
 #if defined( KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_HOST )
-  return Impl::OpenMPExec::m_map_rank[ omp_get_thread_num() ];
+  return Impl::t_openmp_instance ? 0 : omp_get_thread_num();
 #else
   return -1 ;
 #endif
 }
 
+inline
+void OpenMP::fence( OpenMP const& instance ) noexcept {}
+
+inline
+bool OpenMP::is_asynchronous( OpenMP const& instance ) noexcept
+{ return false; }
+
+template 
+void OpenMP::partition_master( F const& f
+                             , int num_partitions
+                             , int partition_size
+                             )
+{
+  if (omp_get_nested()) {
+    using Exec = Impl::OpenMPExec;
+
+    Exec * prev_instance = Impl::t_openmp_instance;
+
+    Exec::validate_partition( prev_instance->m_pool_size, num_partitions, partition_size );
+
+    OpenMP::memory_space space;
+
+    #pragma omp parallel num_threads(num_partitions)
+    {
+      void * const ptr = space.allocate( sizeof(Exec) );
+
+      Impl::t_openmp_instance = new (ptr) Exec( partition_size );
+
+      size_t pool_reduce_bytes  =   32 * partition_size ;
+      size_t team_reduce_bytes  =   32 * partition_size ;
+      size_t team_shared_bytes  = 1024 * partition_size ;
+      size_t thread_local_bytes = 1024 ;
+
+      Impl::t_openmp_instance->resize_thread_data( pool_reduce_bytes
+                                                 , team_reduce_bytes
+                                                 , team_shared_bytes
+                                                 , thread_local_bytes
+                                                 );
+
+      f( omp_get_thread_num(), omp_get_num_threads() );
+
+      Impl::t_openmp_instance->~Exec();
+      space.deallocate( Impl::t_openmp_instance, sizeof(Exec) );
+      Impl::t_openmp_instance = nullptr;
+    }
+
+    Impl::t_openmp_instance  = prev_instance;
+  }
+  else {
+    // nested openmp not enabled
+    f(0,1);
+  }
+}
+
+
+namespace Experimental {
+
+template<>
+class MasterLock
+{
+public:
+  void lock()     { omp_set_lock( &m_lock );   }
+  void unlock()   { omp_unset_lock( &m_lock ); }
+  bool try_lock() { return static_cast(omp_test_lock( &m_lock )); }
+
+  MasterLock()  { omp_init_lock( &m_lock ); }
+  ~MasterLock() { omp_destroy_lock( &m_lock ); }
+
+  MasterLock( MasterLock const& ) = delete;
+  MasterLock( MasterLock && )     = delete;
+  MasterLock & operator=( MasterLock const& ) = delete;
+  MasterLock & operator=( MasterLock && )     = delete;
+
+private:
+  omp_lock_t m_lock;
+
+};
+
+template<>
+class UniqueToken< OpenMP, UniqueTokenScope::Instance>
+{
+public:
+  using execution_space = OpenMP;
+  using size_type       = int;
+
+  /// \brief create object size for concurrency on the given instance
+  ///
+  /// This object should not be shared between instances
+  UniqueToken( execution_space const& = execution_space() ) noexcept {}
+
+  /// \brief upper bound for acquired values, i.e. 0 <= value < size()
+  KOKKOS_INLINE_FUNCTION
+  int size() const noexcept
+    {
+      #if defined( KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_HOST )
+      return Kokkos::OpenMP::thread_pool_size();
+      #else
+      return 0 ;
+      #endif
+    }
+
+  /// \brief acquire value such that 0 <= value < size()
+  KOKKOS_INLINE_FUNCTION
+  int acquire() const  noexcept
+    {
+      #if defined( KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_HOST )
+      return Kokkos::OpenMP::thread_pool_rank();
+      #else
+      return 0 ;
+      #endif
+    }
+
+  /// \brief release a value acquired by generate
+  KOKKOS_INLINE_FUNCTION
+  void release( int ) const noexcept {}
+};
+
+template<>
+class UniqueToken< OpenMP, UniqueTokenScope::Global>
+{
+public:
+  using execution_space = OpenMP;
+  using size_type       = int;
+
+  /// \brief create object size for concurrency on the given instance
+  ///
+  /// This object should not be shared between instances
+  UniqueToken( execution_space const& = execution_space() ) noexcept {}
+
+  /// \brief upper bound for acquired values, i.e. 0 <= value < size()
+  KOKKOS_INLINE_FUNCTION
+  int size() const noexcept
+    {
+      #if defined( KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_HOST )
+      return Kokkos::Impl::g_openmp_hardware_max_threads ;
+      #else
+      return 0 ;
+      #endif
+    }
+
+  /// \brief acquire value such that 0 <= value < size()
+  KOKKOS_INLINE_FUNCTION
+  int acquire() const noexcept
+    {
+      #if defined( KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_HOST )
+      return Kokkos::Impl::t_openmp_hardware_id ;
+      #else
+      return 0 ;
+      #endif
+    }
+
+  /// \brief release a value acquired by generate
+  KOKKOS_INLINE_FUNCTION
+  void release( int ) const noexcept {}
+};
+
+} // namespace Experimental
+
+
+#if !defined( KOKKOS_DISABLE_DEPRECATED )
+
+inline
+int OpenMP::thread_pool_size( int depth )
+{
+  return depth < 2
+         ? thread_pool_size()
+         : 1;
+}
+
+KOKKOS_INLINE_FUNCTION
+int OpenMP::hardware_thread_id() noexcept
+{
+#if defined( KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_HOST )
+  return Impl::t_openmp_hardware_id;
+#else
+  return -1 ;
+#endif
+}
+
+inline
+int OpenMP::max_hardware_threads() noexcept
+{
+  return Impl::g_openmp_hardware_max_threads;
+}
+
+#endif // KOKKOS_DISABLE_DEPRECATED
+
 } // namespace Kokkos
 
 #endif
diff --git a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Parallel.hpp b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Parallel.hpp
index c47e0fc654..b54abb0068 100644
--- a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Parallel.hpp
+++ b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Parallel.hpp
@@ -52,6 +52,8 @@
 #include 
 #include 
 
+#include 
+
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
@@ -71,8 +73,9 @@ private:
   typedef typename Policy::WorkRange    WorkRange ;
   typedef typename Policy::member_type  Member ;
 
-  const FunctorType m_functor ;
-  const Policy      m_policy ;
+        OpenMPExec   * m_instance ;
+  const FunctorType    m_functor ;
+  const Policy         m_policy ;
 
   template< class TagType >
   inline static
@@ -110,16 +113,120 @@ private:
 public:
 
   inline void execute() const
+  {
+    enum { is_dynamic = std::is_same< typename Policy::schedule_type::type
+         , Kokkos::Dynamic >::value
+         };
+
+    if ( OpenMP::in_parallel() ) {
+      exec_range< WorkTag >( m_functor
+                           , m_policy.begin()
+                           , m_policy.end() );
+    }
+    else {
+
+      OpenMPExec::verify_is_master("Kokkos::OpenMP parallel_for");
+
+      const int pool_size = OpenMP::thread_pool_size();
+      #pragma omp parallel num_threads(pool_size)
+      {
+        HostThreadTeamData & data = *(m_instance->get_thread_data());
+
+        data.set_work_partition( m_policy.end() - m_policy.begin()
+            , m_policy.chunk_size() );
+
+        if ( is_dynamic ) {
+          // Make sure work partition is set before stealing
+          if ( data.pool_rendezvous() ) data.pool_rendezvous_release();
+        }
+
+        std::pair range(0,0);
+
+        do {
+
+          range = is_dynamic ? data.get_work_stealing_chunk()
+            : data.get_work_partition();
+
+          ParallelFor::template
+            exec_range< WorkTag >( m_functor
+                , range.first  + m_policy.begin()
+                , range.second + m_policy.begin() );
+
+        } while ( is_dynamic && 0 <= range.first );
+      }
+    }
+  }
+
+  inline
+  ParallelFor( const FunctorType & arg_functor
+             , Policy arg_policy )
+    : m_instance( t_openmp_instance )
+    , m_functor( arg_functor )
+    , m_policy(  arg_policy )
+    {}
+};
+
+
+// MDRangePolicy impl
+template< class FunctorType , class ... Traits >
+class ParallelFor< FunctorType
+                 , Kokkos::Experimental::MDRangePolicy< Traits ... >
+                 , Kokkos::OpenMP
+                 >
+{
+private:
+
+  typedef Kokkos::Experimental::MDRangePolicy< Traits ... > MDRangePolicy ;
+  typedef typename MDRangePolicy::impl_range_policy         Policy ;
+  typedef typename MDRangePolicy::work_tag                  WorkTag ;
+
+  typedef typename Policy::WorkRange    WorkRange ;
+  typedef typename Policy::member_type  Member ;
+
+  typedef typename Kokkos::Experimental::Impl::HostIterateTile< MDRangePolicy, FunctorType, typename MDRangePolicy::work_tag, void > iterate_type;
+
+        OpenMPExec   * m_instance ;
+  const FunctorType   m_functor ;
+  const MDRangePolicy m_mdr_policy ;
+  const Policy        m_policy ;  // construct as RangePolicy( 0, num_tiles ).set_chunk_size(1) in ctor
+
+  inline static
+  void
+  exec_range( const MDRangePolicy & mdr_policy 
+            , const FunctorType & functor
+            , const Member ibeg , const Member iend )
     {
+      #ifdef KOKKOS_ENABLE_AGGRESSIVE_VECTORIZATION
+      #ifdef KOKKOS_ENABLE_PRAGMA_IVDEP
+      #pragma ivdep
+      #endif
+      #endif
+      for ( Member iwork = ibeg ; iwork < iend ; ++iwork ) {
+        iterate_type( mdr_policy, functor )( iwork );
+      }
+    }
+
+public:
+
+  inline void execute() const
+  {
       enum { is_dynamic = std::is_same< typename Policy::schedule_type::type
                                       , Kokkos::Dynamic >::value };
 
-      OpenMPExec::verify_is_process("Kokkos::OpenMP parallel_for");
-      OpenMPExec::verify_initialized("Kokkos::OpenMP parallel_for");
+    if ( OpenMP::in_parallel() ) {
+      ParallelFor::exec_range ( m_mdr_policy
+                              , m_functor
+                              , m_policy.begin()
+                              , m_policy.end() );
+    }
+    else {
 
-#pragma omp parallel
+      OpenMPExec::verify_is_master("Kokkos::OpenMP parallel_for");
+
+      const int pool_size = OpenMP::thread_pool_size();
+      #pragma omp parallel num_threads(pool_size)
       {
-        HostThreadTeamData & data = *OpenMPExec::get_thread_data();
+        HostThreadTeamData & data = *(m_instance->get_thread_data());
 
         data.set_work_partition( m_policy.end() - m_policy.begin()
                                , m_policy.chunk_size() );
@@ -136,8 +243,8 @@ public:
           range = is_dynamic ? data.get_work_stealing_chunk()
                              : data.get_work_partition();
 
-          ParallelFor::template
-            exec_range< WorkTag >( m_functor
+          ParallelFor::exec_range( m_mdr_policy 
+                                 , m_functor
                                  , range.first  + m_policy.begin()
                                  , range.second + m_policy.begin() );
 
@@ -145,12 +252,15 @@ public:
       }
       // END #pragma omp parallel
     }
+  }
 
   inline
   ParallelFor( const FunctorType & arg_functor
-             , Policy arg_policy )
-    : m_functor( arg_functor )
-    , m_policy(  arg_policy )
+             , MDRangePolicy arg_policy )
+    : m_instance( t_openmp_instance )
+    , m_functor( arg_functor )
+    , m_mdr_policy( arg_policy )
+    , m_policy( Policy(0, m_mdr_policy.m_num_tiles).set_chunk_size(1) )
     {}
 };
 
@@ -191,10 +301,11 @@ private:
   typedef typename Analysis::pointer_type    pointer_type ;
   typedef typename Analysis::reference_type  reference_type ;
 
-  const FunctorType   m_functor ;
-  const Policy        m_policy ;
-  const ReducerType   m_reducer ;
-  const pointer_type  m_result_ptr ;
+        OpenMPExec   * m_instance;
+  const FunctorType    m_functor;
+  const Policy         m_policy;
+  const ReducerType    m_reducer;
+  const pointer_type   m_result_ptr;
 
   template< class TagType >
   inline static
@@ -228,21 +339,21 @@ public:
       enum { is_dynamic = std::is_same< typename Policy::schedule_type::type
                                       , Kokkos::Dynamic >::value };
 
-      OpenMPExec::verify_is_process("Kokkos::OpenMP parallel_reduce");
-      OpenMPExec::verify_initialized("Kokkos::OpenMP parallel_reduce");
+      OpenMPExec::verify_is_master("Kokkos::OpenMP parallel_reduce");
 
       const size_t pool_reduce_bytes =
         Analysis::value_size( ReducerConditional::select(m_functor, m_reducer));
 
-      OpenMPExec::resize_thread_data( pool_reduce_bytes
+      m_instance->resize_thread_data( pool_reduce_bytes
                                     , 0 // team_reduce_bytes
                                     , 0 // team_shared_bytes
                                     , 0 // thread_local_bytes
                                     );
 
-#pragma omp parallel
+      const int pool_size = OpenMP::thread_pool_size();
+      #pragma omp parallel num_threads(pool_size)
       {
-        HostThreadTeamData & data = *OpenMPExec::get_thread_data();
+        HostThreadTeamData & data = *(m_instance->get_thread_data());
 
         data.set_work_partition( m_policy.end() - m_policy.begin()
                                , m_policy.chunk_size() );
@@ -271,16 +382,15 @@ public:
 
         } while ( is_dynamic && 0 <= range.first );
       }
-// END #pragma omp parallel
 
       // Reduction:
 
-      const pointer_type ptr = pointer_type( OpenMPExec::get_thread_data(0)->pool_reduce_local() );
+      const pointer_type ptr = pointer_type( m_instance->get_thread_data(0)->pool_reduce_local() );
 
-      for ( int i = 1 ; i < OpenMPExec::pool_size() ; ++i ) {
+      for ( int i = 1 ; i < pool_size ; ++i ) {
         ValueJoin::join( ReducerConditional::select(m_functor , m_reducer)
                        , ptr
-                       , OpenMPExec::get_thread_data(i)->pool_reduce_local() );
+                       , m_instance->get_thread_data(i)->pool_reduce_local() );
       }
 
       Kokkos::Impl::FunctorFinal<  ReducerTypeFwd , WorkTag >::final( ReducerConditional::select(m_functor , m_reducer) , ptr );
@@ -303,7 +413,8 @@ public:
                            Kokkos::is_view< ViewType >::value &&
                            !Kokkos::is_reducer_type::value
                   ,void*>::type = NULL)
-    : m_functor( arg_functor )
+    : m_instance( t_openmp_instance )
+    , m_functor( arg_functor )
     , m_policy(  arg_policy )
     , m_reducer( InvalidType() )
     , m_result_ptr(  arg_view.data() )
@@ -317,7 +428,8 @@ public:
   ParallelReduce( const FunctorType & arg_functor
                 , Policy       arg_policy
                 , const ReducerType& reducer )
-    : m_functor( arg_functor )
+    : m_instance( t_openmp_instance )
+    , m_functor( arg_functor )
     , m_policy(  arg_policy )
     , m_reducer( reducer )
     , m_result_ptr(  reducer.view().data() )
@@ -329,6 +441,173 @@ public:
 
 };
 
+
+// MDRangePolicy impl
+template< class FunctorType , class ReducerType, class ... Traits >
+class ParallelReduce< FunctorType
+                    , Kokkos::Experimental::MDRangePolicy< Traits ...>
+                    , ReducerType
+                    , Kokkos::OpenMP
+                    >
+{
+private:
+
+  typedef Kokkos::Experimental::MDRangePolicy< Traits ... > MDRangePolicy ;
+  typedef typename MDRangePolicy::impl_range_policy         Policy ;
+
+  typedef typename MDRangePolicy::work_tag                  WorkTag ;
+  typedef typename Policy::WorkRange                        WorkRange ;
+  typedef typename Policy::member_type                      Member ;
+
+  typedef FunctorAnalysis< FunctorPatternInterface::REDUCE , Policy , FunctorType > Analysis ;
+
+  typedef Kokkos::Impl::if_c< std::is_same::value, FunctorType, ReducerType> ReducerConditional;
+  typedef typename ReducerConditional::type ReducerTypeFwd;
+
+  typedef typename ReducerTypeFwd::value_type ValueType; 
+
+  typedef Kokkos::Impl::FunctorValueInit<   ReducerTypeFwd, WorkTag > ValueInit ;
+  typedef Kokkos::Impl::FunctorValueJoin<   ReducerTypeFwd, WorkTag > ValueJoin ;
+
+  typedef typename Analysis::pointer_type    pointer_type ;
+  typedef typename Analysis::reference_type  reference_type ;
+
+  using iterate_type = typename Kokkos::Experimental::Impl::HostIterateTile< MDRangePolicy
+                                                                           , FunctorType
+                                                                           , WorkTag
+                                                                           , ValueType
+                                                                           >;
+
+        OpenMPExec   * m_instance ;
+  const FunctorType   m_functor ;
+  const MDRangePolicy m_mdr_policy ;
+  const Policy        m_policy ;     // construct as RangePolicy( 0, num_tiles ).set_chunk_size(1) in ctor
+  const ReducerType   m_reducer ;
+  const pointer_type  m_result_ptr ;
+
+  inline static
+  void
+  exec_range( const MDRangePolicy & mdr_policy
+            , const FunctorType & functor
+            , const Member ibeg , const Member iend
+            , reference_type update )
+    {
+      for ( Member iwork = ibeg ; iwork < iend ; ++iwork ) {
+        iterate_type( mdr_policy, functor, update )( iwork );
+      }
+    }
+
+public:
+
+  inline void execute() const
+    {
+      enum { is_dynamic = std::is_same< typename Policy::schedule_type::type
+                                      , Kokkos::Dynamic >::value };
+
+      OpenMPExec::verify_is_master("Kokkos::OpenMP parallel_reduce");
+
+      const size_t pool_reduce_bytes =
+        Analysis::value_size( ReducerConditional::select(m_functor, m_reducer));
+
+      m_instance->resize_thread_data( pool_reduce_bytes
+                                    , 0 // team_reduce_bytes
+                                    , 0 // team_shared_bytes
+                                    , 0 // thread_local_bytes
+                                    );
+
+      const int pool_size = OpenMP::thread_pool_size();
+      #pragma omp parallel num_threads(pool_size)
+      {
+        HostThreadTeamData & data = *(m_instance->get_thread_data());
+
+        data.set_work_partition( m_policy.end() - m_policy.begin()
+                               , m_policy.chunk_size() );
+
+        if ( is_dynamic ) {
+          // Make sure work partition is set before stealing
+          if ( data.pool_rendezvous() ) data.pool_rendezvous_release();
+        }
+
+        reference_type update =
+          ValueInit::init( ReducerConditional::select(m_functor , m_reducer)
+                         , data.pool_reduce_local() );
+
+        std::pair range(0,0);
+
+        do {
+
+          range = is_dynamic ? data.get_work_stealing_chunk()
+                             : data.get_work_partition();
+
+          ParallelReduce::exec_range ( m_mdr_policy, m_functor
+                                     , range.first  + m_policy.begin()
+                                     , range.second + m_policy.begin()
+                                     , update );
+
+        } while ( is_dynamic && 0 <= range.first );
+      }
+// END #pragma omp parallel
+
+      // Reduction:
+
+      const pointer_type ptr = pointer_type( m_instance->get_thread_data(0)->pool_reduce_local() );
+
+      for ( int i = 1 ; i < pool_size ; ++i ) {
+        ValueJoin::join( ReducerConditional::select(m_functor , m_reducer)
+                       , ptr
+                       , m_instance->get_thread_data(i)->pool_reduce_local() );
+      }
+
+      Kokkos::Impl::FunctorFinal<  ReducerTypeFwd , WorkTag >::final( ReducerConditional::select(m_functor , m_reducer) , ptr );
+
+      if ( m_result_ptr ) {
+        const int n = Analysis::value_count( ReducerConditional::select(m_functor , m_reducer) );
+
+        for ( int j = 0 ; j < n ; ++j ) { m_result_ptr[j] = ptr[j] ; }
+      }
+    }
+
+  //----------------------------------------
+
+  template< class ViewType >
+  inline
+  ParallelReduce( const FunctorType & arg_functor
+                , MDRangePolicy       arg_policy
+                , const ViewType    & arg_view
+                , typename std::enable_if<
+                           Kokkos::is_view< ViewType >::value &&
+                           !Kokkos::is_reducer_type::value
+                  ,void*>::type = NULL)
+    : m_instance( t_openmp_instance )
+    , m_functor( arg_functor )
+    , m_mdr_policy(  arg_policy )
+    , m_policy( Policy(0, m_mdr_policy.m_num_tiles).set_chunk_size(1) )
+    , m_reducer( InvalidType() )
+    , m_result_ptr(  arg_view.data() )
+    {
+      /*static_assert( std::is_same< typename ViewType::memory_space
+                                      , Kokkos::HostSpace >::value
+        , "Reduction result on Kokkos::OpenMP must be a Kokkos::View in HostSpace" );*/
+    }
+
+  inline
+  ParallelReduce( const FunctorType & arg_functor
+                , MDRangePolicy       arg_policy
+                , const ReducerType& reducer )
+    : m_instance( t_openmp_instance )
+    , m_functor( arg_functor )
+    , m_mdr_policy(  arg_policy )
+    , m_policy( Policy(0, m_mdr_policy.m_num_tiles).set_chunk_size(1) )
+    , m_reducer( reducer )
+    , m_result_ptr(  reducer.view().data() )
+    {
+      /*static_assert( std::is_same< typename ViewType::memory_space
+                                      , Kokkos::HostSpace >::value
+        , "Reduction result on Kokkos::OpenMP must be a Kokkos::View in HostSpace" );*/
+    }
+
+};
+
 } // namespace Impl
 } // namespace Kokkos
 
@@ -361,8 +640,9 @@ private:
   typedef typename Analysis::pointer_type    pointer_type ;
   typedef typename Analysis::reference_type  reference_type ;
 
-  const FunctorType   m_functor ;
-  const Policy        m_policy ;
+        OpenMPExec   * m_instance;
+  const FunctorType    m_functor;
+  const Policy         m_policy;
 
   template< class TagType >
   inline static
@@ -394,23 +674,23 @@ public:
   inline
   void execute() const
     {
-      OpenMPExec::verify_is_process("Kokkos::OpenMP parallel_scan");
-      OpenMPExec::verify_initialized("Kokkos::OpenMP parallel_scan");
+      OpenMPExec::verify_is_master("Kokkos::OpenMP parallel_scan");
 
       const int    value_count       = Analysis::value_count( m_functor );
       const size_t pool_reduce_bytes = 2 * Analysis::value_size( m_functor );
 
-      OpenMPExec::resize_thread_data( pool_reduce_bytes
+      m_instance->resize_thread_data( pool_reduce_bytes
                                     , 0 // team_reduce_bytes
                                     , 0 // team_shared_bytes
                                     , 0 // thread_local_bytes
                                     );
 
-#pragma omp parallel
+      const int pool_size = OpenMP::thread_pool_size();
+      #pragma omp parallel num_threads(pool_size)
       {
-        HostThreadTeamData & data = *OpenMPExec::get_thread_data();
+        HostThreadTeamData & data = *(m_instance->get_thread_data());
 
-        const WorkRange range( m_policy, data.pool_rank(), data.pool_size() );
+        const WorkRange range( m_policy, omp_get_thread_num(), omp_get_num_threads() );
 
         reference_type update_sum =
           ValueInit::init( m_functor , data.pool_reduce_local() );
@@ -422,7 +702,7 @@ public:
 
           pointer_type ptr_prev = 0 ;
 
-          const int n = data.pool_size();
+          const int n = omp_get_num_threads();
 
           for ( int i = 0 ; i < n ; ++i ) {
 
@@ -452,7 +732,6 @@ public:
         ParallelScan::template exec_range< WorkTag >
           ( m_functor , range.begin() , range.end() , update_base , true );
       }
-/* END #pragma omp parallel */
 
     }
 
@@ -461,7 +740,8 @@ public:
   inline
   ParallelScan( const FunctorType & arg_functor
               , const Policy      & arg_policy )
-    : m_functor( arg_functor )
+    : m_instance( t_openmp_instance )
+    , m_functor( arg_functor )
     , m_policy(  arg_policy )
   {}
 
@@ -492,9 +772,10 @@ private:
   typedef typename Policy::schedule_type::type  SchedTag ;
   typedef typename Policy::member_type          Member ;
 
-  const FunctorType  m_functor ;
-  const Policy       m_policy ;
-  const int          m_shmem_size ;
+        OpenMPExec   * m_instance;
+  const FunctorType    m_functor;
+  const Policy         m_policy;
+  const int            m_shmem_size;
 
   template< class TagType >
   inline static
@@ -548,22 +829,22 @@ public:
     {
       enum { is_dynamic = std::is_same< SchedTag , Kokkos::Dynamic >::value };
 
-      OpenMPExec::verify_is_process("Kokkos::OpenMP parallel_for");
-      OpenMPExec::verify_initialized("Kokkos::OpenMP parallel_for");
+      OpenMPExec::verify_is_master("Kokkos::OpenMP parallel_for");
 
       const size_t pool_reduce_size = 0 ; // Never shrinks
       const size_t team_reduce_size = TEAM_REDUCE_SIZE * m_policy.team_size();
       const size_t team_shared_size = m_shmem_size + m_policy.scratch_size(1);
       const size_t thread_local_size = 0 ; // Never shrinks
 
-      OpenMPExec::resize_thread_data( pool_reduce_size
+      m_instance->resize_thread_data( pool_reduce_size
                                     , team_reduce_size
                                     , team_shared_size
                                     , thread_local_size );
 
-#pragma omp parallel
+      const int pool_size = OpenMP::thread_pool_size();
+      #pragma omp parallel num_threads(pool_size)
       {
-        HostThreadTeamData & data = *OpenMPExec::get_thread_data();
+        HostThreadTeamData & data = *(m_instance->get_thread_data());
 
         const int active = data.organize_team( m_policy.team_size() );
 
@@ -598,14 +879,14 @@ public:
 
         data.disband_team();
       }
-// END #pragma omp parallel
     }
 
 
   inline
   ParallelFor( const FunctorType & arg_functor ,
                const Policy      & arg_policy )
-    : m_functor( arg_functor )
+    : m_instance( t_openmp_instance )
+    , m_functor( arg_functor )
     , m_policy(  arg_policy )
     , m_shmem_size( arg_policy.scratch_size(0) +
                     arg_policy.scratch_size(1) +
@@ -646,11 +927,12 @@ private:
   typedef typename Analysis::pointer_type    pointer_type ;
   typedef typename Analysis::reference_type  reference_type ;
 
-  const FunctorType  m_functor ;
-  const Policy       m_policy ;
-  const ReducerType  m_reducer ;
-  const pointer_type m_result_ptr ;
-  const int          m_shmem_size ;
+        OpenMPExec   * m_instance;
+  const FunctorType    m_functor;
+  const Policy         m_policy;
+  const ReducerType    m_reducer;
+  const pointer_type   m_result_ptr;
+  const int            m_shmem_size;
 
   template< class TagType >
   inline static
@@ -706,8 +988,7 @@ public:
     {
       enum { is_dynamic = std::is_same< SchedTag , Kokkos::Dynamic >::value };
 
-      OpenMPExec::verify_is_process("Kokkos::OpenMP parallel_reduce");
-      OpenMPExec::verify_initialized("Kokkos::OpenMP parallel_reduce");
+      OpenMPExec::verify_is_master("Kokkos::OpenMP parallel_reduce");
 
       const size_t pool_reduce_size =
         Analysis::value_size( ReducerConditional::select(m_functor, m_reducer));
@@ -716,14 +997,15 @@ public:
       const size_t team_shared_size = m_shmem_size + m_policy.scratch_size(1);
       const size_t thread_local_size = 0 ; // Never shrinks
 
-      OpenMPExec::resize_thread_data( pool_reduce_size
+      m_instance->resize_thread_data( pool_reduce_size
                                     , team_reduce_size
                                     , team_shared_size
                                     , thread_local_size );
 
-#pragma omp parallel
+      const int pool_size = OpenMP::thread_pool_size();
+      #pragma omp parallel num_threads(pool_size)
       {
-        HostThreadTeamData & data = *OpenMPExec::get_thread_data();
+        HostThreadTeamData & data = *(m_instance->get_thread_data());
 
         const int active = data.organize_team( m_policy.team_size() );
 
@@ -763,17 +1045,26 @@ public:
         }
 
         data.disband_team();
+
+        //  This thread has updated 'pool_reduce_local()' with its
+        //  contributions to the reduction.  The parallel region is
+        //  about to terminate and the master thread will load and
+        //  reduce each 'pool_reduce_local()' contribution.
+        //  Must 'memory_fence()' to guarantee that storing the update to
+        //  'pool_reduce_local()' will complete before this thread
+        //  exits the parallel region.
+
+        memory_fence();
       }
-// END #pragma omp parallel
 
       // Reduction:
 
-      const pointer_type ptr = pointer_type( OpenMPExec::get_thread_data(0)->pool_reduce_local() );
+      const pointer_type ptr = pointer_type( m_instance->get_thread_data(0)->pool_reduce_local() );
 
-      for ( int i = 1 ; i < OpenMPExec::pool_size() ; ++i ) {
+      for ( int i = 1 ; i < pool_size ; ++i ) {
         ValueJoin::join( ReducerConditional::select(m_functor , m_reducer)
                        , ptr
-                       , OpenMPExec::get_thread_data(i)->pool_reduce_local() );
+                       , m_instance->get_thread_data(i)->pool_reduce_local() );
       }
 
       Kokkos::Impl::FunctorFinal<  ReducerTypeFwd , WorkTag >::final( ReducerConditional::select(m_functor , m_reducer) , ptr );
@@ -796,7 +1087,8 @@ public:
                     Kokkos::is_view< ViewType >::value &&
                     !Kokkos::is_reducer_type::value
                     ,void*>::type = NULL)
-    : m_functor( arg_functor )
+    : m_instance( t_openmp_instance )
+    , m_functor( arg_functor )
     , m_policy(  arg_policy )
     , m_reducer( InvalidType() )
     , m_result_ptr( arg_result.ptr_on_device() )
@@ -810,7 +1102,8 @@ public:
   ParallelReduce( const FunctorType & arg_functor
     , Policy       arg_policy
     , const ReducerType& reducer )
-  : m_functor( arg_functor )
+  : m_instance( t_openmp_instance )
+  , m_functor( arg_functor )
   , m_policy(  arg_policy )
   , m_reducer( reducer )
   , m_result_ptr(  reducer.view().data() )
diff --git a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Task.cpp b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Task.cpp
index d4ade211f8..77363876b0 100644
--- a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Task.cpp
+++ b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Task.cpp
@@ -105,7 +105,7 @@ void TaskQueueSpecialization< Kokkos::OpenMP >::execute
 {
   using execution_space = Kokkos::OpenMP ;
   using queue_type      = TaskQueue< execution_space > ;
-  using task_root_type  = TaskBase< execution_space , void , void > ;
+  using task_root_type  = TaskBase< void , void , void > ;
   using Member          = Impl::HostThreadTeamMember< execution_space > ;
 
   static task_root_type * const end =
@@ -115,23 +115,19 @@ void TaskQueueSpecialization< Kokkos::OpenMP >::execute
   HostThreadTeamData & team_data_single =
     HostThreadTeamDataSingleton::singleton();
 
-  const int team_size = Impl::OpenMPExec::pool_size(2); // Threads per core
-  // const int team_size = Impl::OpenMPExec::pool_size(1); // Threads per NUMA
+  Impl::OpenMPExec * instance = t_openmp_instance;
+  const int pool_size = OpenMP::thread_pool_size();
 
-#if 0
-fprintf(stdout,"TaskQueue execute %d\n", team_size );
-fflush(stdout);
-#endif
+  const int team_size = 1;  // Threads per core
+  instance->resize_thread_data( 0 /* global reduce buffer */
+                              , 512 * team_size /* team reduce buffer */
+                              , 0 /* team shared buffer */
+                              , 0 /* thread local buffer */
+                              );
 
-  OpenMPExec::resize_thread_data( 0 /* global reduce buffer */
-                                , 512 * team_size /* team reduce buffer */
-                                , 0 /* team shared buffer */
-                                , 0 /* thread local buffer */
-                                );
-
-#pragma omp parallel
+  #pragma omp parallel num_threads(pool_size)
   {
-    Impl::HostThreadTeamData & self = *Impl::OpenMPExec::get_thread_data();
+    Impl::HostThreadTeamData & self = *(instance->get_thread_data());
 
     // Organizing threads into a team performs a barrier across the
     // entire pool to insure proper initialization of the team
@@ -142,18 +138,6 @@ fflush(stdout);
       Member single_exec( team_data_single );
       Member team_exec( self );
 
-#if 0
-fprintf(stdout,"TaskQueue pool(%d of %d) team(%d of %d) league(%d of %d) running\n"
-       , self.pool_rank()
-       , self.pool_size()
-       , team_exec.team_rank()
-       , team_exec.team_size()
-       , team_exec.league_rank()
-       , team_exec.league_size()
-       );
-fflush(stdout);
-#endif
-
       // Loop until all queues are empty and no tasks in flight
 
       task_root_type * task = 0 ;
@@ -197,15 +181,6 @@ fflush(stdout);
 
               // if a single thread task then execute now
 
-#if 0
-fprintf(stdout,"TaskQueue pool(%d of %d) executing single task 0x%lx\n"
-       , self.pool_rank()
-       , self.pool_size()
-       , int64_t(task)
-       );
-fflush(stdout);
-#endif
-
               (*task->m_apply)( task , & single_exec );
 
               leader_loop = true ;
@@ -220,57 +195,14 @@ fflush(stdout);
 
         if ( 0 != task ) { // Thread Team Task
 
-#if 0
-fprintf(stdout,"TaskQueue pool(%d of %d) team((%d of %d) league(%d of %d) executing team task 0x%lx\n"
-       , self.pool_rank()
-       , self.pool_size()
-       , team_exec.team_rank()
-       , team_exec.team_size()
-       , team_exec.league_rank()
-       , team_exec.league_size()
-       , int64_t(task)
-       );
-fflush(stdout);
-#endif
-
           (*task->m_apply)( task , & team_exec );
 
           // The m_apply function performs a barrier
         }
       } while( 0 != task );
-
-#if 0
-fprintf(stdout,"TaskQueue pool(%d of %d) team(%d of %d) league(%d of %d) ending\n"
-       , self.pool_rank()
-       , self.pool_size()
-       , team_exec.team_rank()
-       , team_exec.team_size()
-       , team_exec.league_rank()
-       , team_exec.league_size()
-       );
-fflush(stdout);
-#endif
-
     }
-
     self.disband_team();
-
-#if 0
-fprintf(stdout,"TaskQueue pool(%d of %d) disbanded\n"
-       , self.pool_rank()
-       , self.pool_size()
-       );
-fflush(stdout);
-#endif
-
   }
-// END #pragma omp parallel
-
-#if 0
-fprintf(stdout,"TaskQueue execute %d end\n", team_size );
-fflush(stdout);
-#endif
-
 }
 
 void TaskQueueSpecialization< Kokkos::OpenMP >::
@@ -279,10 +211,10 @@ void TaskQueueSpecialization< Kokkos::OpenMP >::
 {
   using execution_space = Kokkos::OpenMP ;
   using queue_type      = TaskQueue< execution_space > ;
-  using task_root_type  = TaskBase< execution_space , void , void > ;
+  using task_root_type  = TaskBase< void , void , void > ;
   using Member          = Impl::HostThreadTeamMember< execution_space > ;
 
-  if ( 1 == omp_get_num_threads() ) {
+  if ( 1 == OpenMP::thread_pool_size() ) {
 
     task_root_type * const end = (task_root_type *) task_root_type::EndTag ;
 
diff --git a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Task.hpp b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Task.hpp
index 82fbef255b..dfa1635e08 100644
--- a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Task.hpp
+++ b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Task.hpp
@@ -45,7 +45,7 @@
 #define KOKKOS_IMPL_OPENMP_TASK_HPP
 
 #include 
-#if defined( KOKKOS_ENABLE_TASKDAG )
+#if defined( KOKKOS_ENABLE_OPENMP ) && defined( KOKKOS_ENABLE_TASKDAG )
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
@@ -60,7 +60,7 @@ public:
 
   using execution_space = Kokkos::OpenMP ;
   using queue_type      = Kokkos::Impl::TaskQueue< execution_space > ;
-  using task_base_type  = Kokkos::Impl::TaskBase< execution_space , void , void > ;
+  using task_base_type  = Kokkos::Impl::TaskBase< void , void , void > ;
   using member_type     = Kokkos::Impl::HostThreadTeamMember< execution_space > ;
 
   // Must specify memory space
diff --git a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Team.hpp b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Team.hpp
new file mode 100644
index 0000000000..743e6b6e62
--- /dev/null
+++ b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Team.hpp
@@ -0,0 +1,245 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_OPENMP_TEAM_HPP
+#define KOKKOS_OPENMP_TEAM_HPP
+
+#include 
+#if defined( KOKKOS_ENABLE_OPENMP )
+
+#include 
+
+namespace Kokkos { namespace Impl {
+
+template< class ... Properties >
+class TeamPolicyInternal< Kokkos::OpenMP, Properties ... >: public PolicyTraits
+{
+public:
+
+  //! Tag this class as a kokkos execution policy
+  typedef TeamPolicyInternal      execution_policy ;
+
+  typedef PolicyTraits traits;
+
+  TeamPolicyInternal& operator = (const TeamPolicyInternal& p) {
+    m_league_size = p.m_league_size;
+    m_team_size = p.m_team_size;
+    m_team_alloc = p.m_team_alloc;
+    m_team_iter = p.m_team_iter;
+    m_team_scratch_size[0] = p.m_team_scratch_size[0];
+    m_thread_scratch_size[0] = p.m_thread_scratch_size[0];
+    m_team_scratch_size[1] = p.m_team_scratch_size[1];
+    m_thread_scratch_size[1] = p.m_thread_scratch_size[1];
+    m_chunk_size = p.m_chunk_size;
+    return *this;
+  }
+
+  //----------------------------------------
+
+  template< class FunctorType >
+  inline static
+  int team_size_max( const FunctorType & ) {
+      int pool_size = traits::execution_space::thread_pool_size(1);
+      int max_host_team_size =  Impl::HostThreadTeamData::max_team_members;
+      return pool_size
+  inline static
+  int team_size_recommended( const FunctorType & )
+    { return traits::execution_space::thread_pool_size(2); }
+
+  template< class FunctorType >
+  inline static
+  int team_size_recommended( const FunctorType &, const int& )
+    { return traits::execution_space::thread_pool_size(2); }
+
+  //----------------------------------------
+
+private:
+
+  int m_league_size ;
+  int m_team_size ;
+  int m_team_alloc ;
+  int m_team_iter ;
+
+  size_t m_team_scratch_size[2];
+  size_t m_thread_scratch_size[2];
+
+  int m_chunk_size;
+
+  inline void init( const int league_size_request
+                  , const int team_size_request )
+    {
+      const int pool_size  = traits::execution_space::thread_pool_size(0);
+      const int max_host_team_size =  Impl::HostThreadTeamData::max_team_members;
+      const int team_max   = pool_size 0) {
+      if(!Impl::is_integral_power_of_two( m_chunk_size ))
+        Kokkos::abort("TeamPolicy blocking granularity must be power of two" );
+    }
+
+    int new_chunk_size = 1;
+    while(new_chunk_size*100*concurrency < m_league_size)
+      new_chunk_size *= 2;
+    if(new_chunk_size < 128) {
+      new_chunk_size = 1;
+      while( (new_chunk_size*40*concurrency < m_league_size ) && (new_chunk_size<128) )
+        new_chunk_size*=2;
+    }
+    m_chunk_size = new_chunk_size;
+  }
+
+public:
+  typedef Impl::HostThreadTeamMember< Kokkos::OpenMP > member_type ;
+};
+
+}} // namespace Kokkos::Impl
+
+#endif
+#endif /* KOKKOS_OPENMP_TEAM_HPP */
+
+
diff --git a/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_WorkGraphPolicy.hpp b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_WorkGraphPolicy.hpp
new file mode 100644
index 0000000000..289ad15451
--- /dev/null
+++ b/lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_WorkGraphPolicy.hpp
@@ -0,0 +1,107 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_OPENMP_WORKGRAPHPOLICY_HPP
+#define KOKKOS_OPENMP_WORKGRAPHPOLICY_HPP
+
+namespace Kokkos {
+namespace Impl {
+
+template< class FunctorType , class ... Traits >
+class ParallelFor< FunctorType ,
+                   Kokkos::Experimental::WorkGraphPolicy< Traits ... > ,
+                   Kokkos::OpenMP
+                 >
+  : public Kokkos::Impl::Experimental::
+           WorkGraphExec< FunctorType,
+                          Kokkos::OpenMP,
+                          Traits ...
+                        >
+{
+private:
+
+  typedef Kokkos::Experimental::WorkGraphPolicy< Traits ... > Policy ;
+  typedef Kokkos::Impl::Experimental::
+          WorkGraphExec Base ;
+
+  template< class TagType >
+  typename std::enable_if< std::is_same< TagType , void >::value >::type
+  exec_one(const typename Policy::member_type& i) const {
+    Base::m_functor( i );
+  }
+
+  template< class TagType >
+  typename std::enable_if< ! std::is_same< TagType , void >::value >::type
+  exec_one(const typename Policy::member_type& i) const {
+    const TagType t{} ;
+    Base::m_functor( t , i );
+  }
+
+public:
+
+  inline
+  void execute()
+  {
+    const int pool_size = OpenMP::thread_pool_size();
+
+    #pragma omp parallel num_threads(pool_size)
+    {
+      for (std::int32_t i; (-1 != (i = Base::before_work())); ) {
+        exec_one< typename Policy::work_tag >( i );
+        Base::after_work(i);
+      }
+    }
+  }
+
+  inline
+  ParallelFor( const FunctorType & arg_functor
+             , const Policy      & arg_policy )
+    : Base( arg_functor, arg_policy )
+  {
+  }
+};
+
+} // namespace Impl
+} // namespace Kokkos
+
+#endif /* #define KOKKOS_OPENMP_WORKGRAPHPOLICY_HPP */
diff --git a/lib/kokkos/core/src/OpenMPTarget/Kokkos_OpenMPTarget_Exec.hpp b/lib/kokkos/core/src/OpenMPTarget/Kokkos_OpenMPTarget_Exec.hpp
index bec7844ed6..258a9d2ff7 100644
--- a/lib/kokkos/core/src/OpenMPTarget/Kokkos_OpenMPTarget_Exec.hpp
+++ b/lib/kokkos/core/src/OpenMPTarget/Kokkos_OpenMPTarget_Exec.hpp
@@ -1,13 +1,13 @@
 /*
 //@HEADER
 // ************************************************************************
-// 
+//
 //                        Kokkos v. 2.0
 //              Copyright (2014) Sandia Corporation
-// 
+//
 // Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
 // the U.S. Government retains certain rights in this software.
-// 
+//
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
@@ -36,7 +36,7 @@
 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 //
 // Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
-// 
+//
 // ************************************************************************
 //@HEADER
 */
@@ -45,7 +45,7 @@
 #define KOKKOS_OPENMPTARGETEXEC_HPP
 
 #include 
-#include 
+#include 
 
 #include 
 #include 
@@ -59,10 +59,10 @@ namespace Impl {
 
 
 class OpenMPTargetExec {
-public: 
+public:
   enum { MAX_ACTIVE_THREADS = 256*8*56*4 };
   enum { MAX_ACTIVE_TEAMS = MAX_ACTIVE_THREADS/32 };
-  
+
 private:
   static void* scratch_ptr;
 
@@ -70,7 +70,7 @@ public:
   static void verify_is_process( const char * const );
   static void verify_initialized( const char * const );
 
-  static void* get_scratch_ptr();  
+  static void* get_scratch_ptr();
   static void clear_scratch();
   static void resize_scratch( int64_t reduce_bytes , int64_t team_reduce_bytes, int64_t team_shared_bytes, int64_t thread_local_bytes );
 
@@ -159,7 +159,7 @@ public:
 
   KOKKOS_INLINE_FUNCTION void team_barrier() const
     {
-      #pragma omp barrier  
+      #pragma omp barrier
     }
 
   template
@@ -191,13 +191,13 @@ public:
 
       typedef ValueType value_type;
       const JoinLambdaAdapter op(op_in);
-      
+
       // Make sure there is enough scratch space:
       typedef typename if_c< sizeof(value_type) < TEAM_REDUCE_SIZE
                            , value_type , void >::type type ;
 
       const int n_values = TEAM_REDUCE_SIZE/sizeof(value_type);
-      type * team_scratch = (type*) ((char*)m_glb_scratch + TEAM_REDUCE_SIZE*omp_get_team_num()); 
+      type * team_scratch = (type*) ((char*)m_glb_scratch + TEAM_REDUCE_SIZE*omp_get_team_num());
       for(int i = m_team_rank; i < n_values; i+= m_team_size) {
         team_scratch[i] = value_type();
       }
@@ -209,7 +209,7 @@ public:
           team_scratch[m_team_rank%n_values]+=value;
         #pragma omp barrier
       }
-      
+
       for(int d = 1; d
 #if defined( KOKKOS_ENABLE_QTHREADS )
 
-#include 
+#include 
 
 //----------------------------------------------------------------------------
 
diff --git a/lib/kokkos/core/src/Threads/Kokkos_ThreadsExec.cpp b/lib/kokkos/core/src/Threads/Kokkos_ThreadsExec.cpp
index 4c805310cc..35b2163ae5 100644
--- a/lib/kokkos/core/src/Threads/Kokkos_ThreadsExec.cpp
+++ b/lib/kokkos/core/src/Threads/Kokkos_ThreadsExec.cpp
@@ -45,14 +45,14 @@
 #include 
 #if defined( KOKKOS_ENABLE_THREADS )
 
-#include 
-
 #include 
 #include 
 #include 
 #include 
 #include 
+
 #include 
+
 #include 
 #include 
 #include 
@@ -80,9 +80,7 @@ const void * volatile s_current_function_arg = 0 ;
 
 struct Sentinel {
   Sentinel()
-  {
-    HostSpace::register_in_parallel( ThreadsExec::in_parallel );
-  }
+  {}
 
   ~Sentinel()
   {
@@ -122,6 +120,8 @@ void execute_function_noop( ThreadsExec & , const void * ) {}
 
 void ThreadsExec::driver(void)
 {
+  SharedAllocationRecord< void, void >::tracking_enable();
+
   ThreadsExec this_thread ;
 
   while ( ThreadsExec::Active == this_thread.m_pool_state ) {
@@ -726,6 +726,8 @@ void ThreadsExec::initialize( unsigned thread_count ,
   // Init the array for used for arbitrarily sized atomics
   Impl::init_lock_array_host_space();
 
+  Impl::SharedAllocationRecord< void, void >::tracking_enable();
+
   #if defined(KOKKOS_ENABLE_PROFILING)
     Kokkos::Profiling::initialize();
   #endif
diff --git a/lib/kokkos/core/src/Threads/Kokkos_ThreadsExec.hpp b/lib/kokkos/core/src/Threads/Kokkos_ThreadsExec.hpp
index 74de3a2596..7557bad7d9 100644
--- a/lib/kokkos/core/src/Threads/Kokkos_ThreadsExec.hpp
+++ b/lib/kokkos/core/src/Threads/Kokkos_ThreadsExec.hpp
@@ -50,11 +50,12 @@
 #include 
 
 #include 
-#include 
+#include 
 #include 
 
 #include 
 
+#include 
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
@@ -275,6 +276,17 @@ public:
       if ( ! rev_rank ) {
         Final::final( f , reduce_memory() );
       }
+
+      //  This thread has updated 'reduce_memory()' and upon returning
+      //  from this function will set 'm_pool_state' to inactive.
+      //  If this is a non-root thread then setting 'm_pool_state'
+      //  to inactive triggers another thread to exit a spinwait
+      //  and read the 'reduce_memory'.
+      //  Must 'memory_fence()' to guarantee that storing the update to
+      //  'reduce_memory()' will complete before storing the the update to
+      //  'm_pool_state'.
+
+      memory_fence();
     }
 
   inline
@@ -627,6 +639,62 @@ inline void Threads::fence()
 
 } /* namespace Kokkos */
 
+//----------------------------------------------------------------------------
+//----------------------------------------------------------------------------
+
+namespace Kokkos { namespace Experimental {
+
+template<>
+class UniqueToken< Threads, UniqueTokenScope::Instance>
+{
+public:
+  using execution_space = Threads;
+  using size_type       = int;
+
+  /// \brief create object size for concurrency on the given instance
+  ///
+  /// This object should not be shared between instances
+  UniqueToken( execution_space const& = execution_space() ) noexcept {}
+
+  /// \brief upper bound for acquired values, i.e. 0 <= value < size()
+  inline
+  int size() const noexcept { return Threads::thread_pool_size(); }
+
+  /// \brief acquire value such that 0 <= value < size()
+  inline
+  int acquire() const  noexcept { return Threads::thread_pool_rank(); }
+
+  /// \brief release a value acquired by generate
+  inline
+  void release( int ) const noexcept {}
+};
+
+template<>
+class UniqueToken< Threads, UniqueTokenScope::Global>
+{
+public:
+  using execution_space = Threads;
+  using size_type       = int;
+
+  /// \brief create object size for concurrency on the given instance
+  ///
+  /// This object should not be shared between instances
+  UniqueToken( execution_space const& = execution_space() ) noexcept {}
+
+  /// \brief upper bound for acquired values, i.e. 0 <= value < size()
+  inline
+  int size() const noexcept { return Threads::thread_pool_size(); }
+
+  /// \brief acquire value such that 0 <= value < size()
+  inline
+  int acquire() const  noexcept { return Threads::thread_pool_rank(); }
+
+  /// \brief release a value acquired by generate
+  inline
+  void release( int ) const noexcept {}
+};
+
+}} // namespace Kokkos::Experimental
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 #endif
diff --git a/lib/kokkos/core/src/Threads/Kokkos_ThreadsTeam.hpp b/lib/kokkos/core/src/Threads/Kokkos_ThreadsTeam.hpp
index c12019413b..6060bf191f 100644
--- a/lib/kokkos/core/src/Threads/Kokkos_ThreadsTeam.hpp
+++ b/lib/kokkos/core/src/Threads/Kokkos_ThreadsTeam.hpp
@@ -50,7 +50,7 @@
 #include 
 
 #include 
-#include 
+#include 
 #include 
 #include 
 
@@ -482,6 +482,8 @@ public:
   void next_static()
     {
       if ( m_league_rank < m_league_end ) {
+        // Make sure all stores are complete before entering the barrier
+        memory_fence();
         team_barrier();
         set_team_shared();
       }
@@ -518,6 +520,8 @@ public:
       return;
 
     if ( m_league_rank < m_league_chunk_end ) {
+      // Make sure all stores are complete before entering the barrier
+      memory_fence();
       team_barrier();
       set_team_shared();
     }
diff --git a/lib/kokkos/core/src/Threads/Kokkos_Threads_Parallel.hpp b/lib/kokkos/core/src/Threads/Kokkos_Threads_Parallel.hpp
index 0ee0cd3280..18ac7d26ad 100644
--- a/lib/kokkos/core/src/Threads/Kokkos_Threads_Parallel.hpp
+++ b/lib/kokkos/core/src/Threads/Kokkos_Threads_Parallel.hpp
@@ -55,6 +55,8 @@
 #include 
 #include 
 
+#include 
+
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
@@ -174,6 +176,108 @@ public:
     {}
 };
 
+
+// MDRangePolicy impl
+template< class FunctorType , class ... Traits >
+class ParallelFor< FunctorType
+                 , Kokkos::Experimental::MDRangePolicy< Traits ... >
+                 , Kokkos::Threads
+                 >
+{
+private:
+  typedef Kokkos::Experimental::MDRangePolicy< Traits ... > MDRangePolicy ;
+  typedef typename MDRangePolicy::impl_range_policy         Policy ;
+
+  typedef typename MDRangePolicy::work_tag                  WorkTag ;
+
+  typedef typename Policy::WorkRange   WorkRange ;
+  typedef typename Policy::member_type Member ;
+
+  typedef typename Kokkos::Experimental::Impl::HostIterateTile< MDRangePolicy, FunctorType, typename MDRangePolicy::work_tag, void > iterate_type;
+
+  const FunctorType   m_functor ;
+  const MDRangePolicy m_mdr_policy ;
+  const Policy        m_policy ;  // construct as RangePolicy( 0, num_tiles ).set_chunk_size(1) in ctor
+
+  inline static
+  void
+  exec_range( const MDRangePolicy & mdr_policy 
+            , const FunctorType & functor
+            , const Member ibeg , const Member iend )
+    {
+      #if defined( KOKKOS_ENABLE_AGGRESSIVE_VECTORIZATION ) && \
+          defined( KOKKOS_ENABLE_PRAGMA_IVDEP )
+      #pragma ivdep
+      #endif
+      for ( Member i = ibeg ; i < iend ; ++i ) {
+        iterate_type( mdr_policy, functor )( i );
+      }
+    }
+
+  static void exec( ThreadsExec & exec , const void * arg )
+  {
+    exec_schedule(exec,arg);
+  }
+
+  template
+  static
+  typename std::enable_if< std::is_same::value >::type
+  exec_schedule( ThreadsExec & exec , const void * arg )
+  {
+    const ParallelFor & self = * ((const ParallelFor *) arg );
+
+    WorkRange range( self.m_policy , exec.pool_rank() , exec.pool_size() );
+
+    ParallelFor::exec_range
+      ( self.m_mdr_policy, self.m_functor , range.begin() , range.end() );
+
+    exec.fan_in();
+  }
+
+  template
+  static
+  typename std::enable_if< std::is_same::value >::type
+  exec_schedule( ThreadsExec & exec , const void * arg )
+  {
+    const ParallelFor & self = * ((const ParallelFor *) arg );
+
+    WorkRange range( self.m_policy , exec.pool_rank() , exec.pool_size() );
+
+    exec.set_work_range(range.begin(),range.end(),self.m_policy.chunk_size());
+    exec.reset_steal_target();
+    exec.barrier();
+
+    long work_index = exec.get_work_index();
+
+    while(work_index != -1) {
+      const Member begin = static_cast(work_index) * self.m_policy.chunk_size();
+      const Member end = begin + self.m_policy.chunk_size() < self.m_policy.end()?begin+self.m_policy.chunk_size():self.m_policy.end();
+
+      ParallelFor::exec_range
+        ( self.m_mdr_policy, self.m_functor , begin , end );
+      work_index = exec.get_work_index();
+    }
+
+    exec.fan_in();
+  }
+
+public:
+
+  inline
+  void execute() const
+    {
+      ThreadsExec::start( & ParallelFor::exec , this );
+      ThreadsExec::fence();
+    }
+
+  ParallelFor( const FunctorType & arg_functor
+             , const MDRangePolicy      & arg_policy )
+    : m_functor( arg_functor )
+    , m_mdr_policy( arg_policy )
+    , m_policy( Policy(0, m_mdr_policy.m_num_tiles).set_chunk_size(1) )
+    {}
+};
+
 //----------------------------------------------------------------------------
 /* ParallelFor Kokkos::Threads with TeamPolicy */
 
@@ -440,6 +544,169 @@ public:
 
 };
 
+
+// MDRangePolicy impl
+template< class FunctorType , class ReducerType, class ... Traits >
+class ParallelReduce< FunctorType
+                    , Kokkos::Experimental::MDRangePolicy< Traits ... >
+                    , ReducerType
+                    , Kokkos::Threads
+                    >
+{
+private:
+
+  typedef Kokkos::Experimental::MDRangePolicy< Traits ... > MDRangePolicy ;
+  typedef typename MDRangePolicy::impl_range_policy Policy ;
+
+  typedef typename MDRangePolicy::work_tag    WorkTag ;
+  typedef typename Policy::WorkRange   WorkRange ;
+  typedef typename Policy::member_type Member ;
+
+  typedef Kokkos::Impl::if_c< std::is_same::value, FunctorType, ReducerType> ReducerConditional;
+  typedef typename ReducerConditional::type ReducerTypeFwd;
+
+  typedef typename ReducerTypeFwd::value_type ValueType; 
+
+  typedef Kokkos::Impl::FunctorValueTraits< ReducerTypeFwd, WorkTag > ValueTraits ;
+  typedef Kokkos::Impl::FunctorValueInit<   ReducerTypeFwd, WorkTag > ValueInit ;
+
+  typedef typename ValueTraits::pointer_type    pointer_type ;
+  typedef typename ValueTraits::reference_type  reference_type ;
+
+  using iterate_type = typename Kokkos::Experimental::Impl::HostIterateTile< MDRangePolicy
+                                                                           , FunctorType
+                                                                           , WorkTag
+                                                                           , ValueType
+                                                                           >;
+
+  const FunctorType   m_functor ;
+  const MDRangePolicy m_mdr_policy ;
+  const Policy        m_policy ;  // construct as RangePolicy( 0, num_tiles ).set_chunk_size(1) in ctor
+  const ReducerType   m_reducer ;
+  const pointer_type  m_result_ptr ;
+
+  inline static
+  void
+  exec_range( const MDRangePolicy & mdr_policy
+            , const FunctorType & functor
+            , const Member & ibeg , const Member & iend
+            , reference_type update )
+    {
+      #if defined( KOKKOS_ENABLE_AGGRESSIVE_VECTORIZATION ) && \
+          defined( KOKKOS_ENABLE_PRAGMA_IVDEP )
+      #pragma ivdep
+      #endif
+      for ( Member i = ibeg ; i < iend ; ++i ) {
+        iterate_type( mdr_policy, functor, update )( i );
+      }
+    }
+
+  static void
+  exec( ThreadsExec & exec , const void * arg ) {
+    exec_schedule(exec, arg);
+  }
+
+  template
+  static
+  typename std::enable_if< std::is_same::value >::type
+  exec_schedule( ThreadsExec & exec , const void * arg )
+  {
+    const ParallelReduce & self = * ((const ParallelReduce *) arg );
+    const WorkRange range( self.m_policy, exec.pool_rank(), exec.pool_size() );
+
+    ParallelReduce::exec_range
+      ( self.m_mdr_policy, self.m_functor , range.begin() , range.end()
+      , ValueInit::init( ReducerConditional::select(self.m_functor , self.m_reducer) , exec.reduce_memory() ) );
+
+    exec.template fan_in_reduce< ReducerTypeFwd , WorkTag >( ReducerConditional::select(self.m_functor , self.m_reducer) );
+  }
+
+  template
+  static
+  typename std::enable_if< std::is_same::value >::type
+    exec_schedule( ThreadsExec & exec , const void * arg )
+  {
+    const ParallelReduce & self = * ((const ParallelReduce *) arg );
+    const WorkRange range( self.m_policy, exec.pool_rank(), exec.pool_size() );
+
+    exec.set_work_range(range.begin(),range.end(),self.m_policy.chunk_size());
+    exec.reset_steal_target();
+    exec.barrier();
+
+    long work_index = exec.get_work_index();
+    reference_type update = ValueInit::init( ReducerConditional::select(self.m_functor , self.m_reducer) , exec.reduce_memory() );
+    while(work_index != -1) {
+      const Member begin = static_cast(work_index) * self.m_policy.chunk_size();
+      const Member end = begin + self.m_policy.chunk_size() < self.m_policy.end()?begin+self.m_policy.chunk_size():self.m_policy.end();
+      ParallelReduce::exec_range
+        ( self.m_mdr_policy, self.m_functor , begin , end
+        , update );
+      work_index = exec.get_work_index();
+    }
+
+    exec.template fan_in_reduce< ReducerTypeFwd , WorkTag >( ReducerConditional::select(self.m_functor , self.m_reducer) );
+  }
+
+public:
+
+  inline
+  void execute() const
+    {
+      ThreadsExec::resize_scratch( ValueTraits::value_size( ReducerConditional::select(m_functor , m_reducer) ) , 0 );
+
+      ThreadsExec::start( & ParallelReduce::exec , this );
+
+      ThreadsExec::fence();
+
+      if ( m_result_ptr ) {
+
+        const pointer_type data =
+          (pointer_type) ThreadsExec::root_reduce_scratch();
+
+        const unsigned n = ValueTraits::value_count( ReducerConditional::select(m_functor , m_reducer) );
+        for ( unsigned i = 0 ; i < n ; ++i ) { m_result_ptr[i] = data[i]; }
+      }
+    }
+
+  template< class HostViewType >
+  ParallelReduce( const FunctorType  & arg_functor ,
+                  const MDRangePolicy       & arg_policy ,
+                  const HostViewType & arg_result_view ,
+                  typename std::enable_if<
+                               Kokkos::is_view< HostViewType >::value &&
+                              !Kokkos::is_reducer_type::value
+                  ,void*>::type = NULL)
+    : m_functor( arg_functor )
+    , m_mdr_policy( arg_policy )
+    , m_policy( Policy(0, m_mdr_policy.m_num_tiles).set_chunk_size(1) )
+    , m_reducer( InvalidType() )
+    , m_result_ptr( arg_result_view.ptr_on_device() )
+    {
+      static_assert( Kokkos::is_view< HostViewType >::value
+        , "Kokkos::Threads reduce result must be a View" );
+
+      static_assert( std::is_same< typename HostViewType::memory_space , HostSpace >::value
+        , "Kokkos::Threads reduce result must be a View in HostSpace" );
+    }
+
+  inline
+  ParallelReduce( const FunctorType & arg_functor
+                , MDRangePolicy       arg_policy
+                , const ReducerType& reducer )
+    : m_functor( arg_functor )
+    , m_mdr_policy(  arg_policy )
+    , m_policy( Policy(0, m_mdr_policy.m_num_tiles).set_chunk_size(1) )
+    , m_reducer( reducer )
+    , m_result_ptr(  reducer.view().data() )
+    {
+      /*static_assert( std::is_same< typename ViewType::memory_space
+                                      , Kokkos::HostSpace >::value
+        , "Reduction result on Kokkos::OpenMP must be a Kokkos::View in HostSpace" );*/
+    }
+
+};
+
+
 //----------------------------------------------------------------------------
 /* ParallelReduce with Kokkos::Threads and TeamPolicy */
 
diff --git a/lib/kokkos/core/src/Threads/Kokkos_Threads_WorkGraphPolicy.hpp b/lib/kokkos/core/src/Threads/Kokkos_Threads_WorkGraphPolicy.hpp
new file mode 100644
index 0000000000..be904a1670
--- /dev/null
+++ b/lib/kokkos/core/src/Threads/Kokkos_Threads_WorkGraphPolicy.hpp
@@ -0,0 +1,115 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_THREADS_WORKGRAPHPOLICY_HPP
+#define KOKKOS_THREADS_WORKGRAPHPOLICY_HPP
+
+namespace Kokkos {
+namespace Impl {
+
+template< class FunctorType , class ... Traits >
+class ParallelFor< FunctorType ,
+                   Kokkos::Experimental::WorkGraphPolicy< Traits ... > ,
+                   Kokkos::Threads
+                 >
+  : public Kokkos::Impl::Experimental::
+           WorkGraphExec< FunctorType,
+                          Kokkos::Threads,
+                          Traits ...
+                        >
+{
+private:
+
+  typedef Kokkos::Experimental::WorkGraphPolicy< Traits ... > Policy ;
+  typedef Kokkos::Impl::Experimental::
+          WorkGraphExec Base ;
+  typedef ParallelFor,
+                      Kokkos::Threads> Self ;
+
+  template< class TagType >
+  typename std::enable_if< std::is_same< TagType , void >::value >::type
+  exec_one(const typename Policy::member_type& i) const {
+    Base::m_functor( i );
+  }
+
+  template< class TagType >
+  typename std::enable_if< ! std::is_same< TagType , void >::value >::type
+  exec_one(const typename Policy::member_type& i) const {
+    const TagType t{} ;
+    Base::m_functor( t , i );
+  }
+
+  inline void exec_one_thread() const {
+    for (std::int32_t i; (-1 != (i = Base::before_work())); ) {
+      exec_one< typename Policy::work_tag >( i );
+      Base::after_work(i);
+    }
+  }
+
+  static inline void thread_main( ThreadsExec&, const void* arg ) {
+    const Self& self = *(static_cast(arg));
+    self.exec_one_thread();
+  }
+
+public:
+
+  inline
+  void execute()
+  {
+    ThreadsExec::start( & Self::thread_main, this );
+    ThreadsExec::fence();
+  }
+
+  inline
+  ParallelFor( const FunctorType & arg_functor
+             , const Policy      & arg_policy )
+    : Base( arg_functor, arg_policy )
+  {
+  }
+};
+
+} // namespace Impl
+} // namespace Kokkos
+
+#endif /* #define KOKKOS_THREADS_WORKGRAPHPOLICY_HPP */
diff --git a/lib/kokkos/core/src/impl/KokkosExp_Host_IterateTile.hpp b/lib/kokkos/core/src/impl/KokkosExp_Host_IterateTile.hpp
index 77a1e8754d..0171b209e5 100644
--- a/lib/kokkos/core/src/impl/KokkosExp_Host_IterateTile.hpp
+++ b/lib/kokkos/core/src/impl/KokkosExp_Host_IterateTile.hpp
@@ -141,7 +141,6 @@ namespace Kokkos { namespace Experimental { namespace Impl {
 #define LOOP_ARGS_8 LOOP_ARGS_7, i7 + m_offset[7]
 
 
-
 // New Loop Macros...
 // parallel_for, non-tagged
 #define APPLY( func, ... ) \
@@ -1010,8 +1009,6 @@ namespace Kokkos { namespace Experimental { namespace Impl {
 // end tagged macros
 
 
-
-
 // Structs for calling loops
 template < int Rank, bool IsLeft, typename IType, typename Tagged, typename Enable = void >
 struct Tile_Loop_Type;
@@ -1279,6 +1276,19 @@ struct Tile_Loop_Type<8, IsLeft, IType, Tagged, typename std::enable_if< !std::i
 template 
 using is_void = std::is_same< T , void >;
 
+template 
+struct is_type_array : std::false_type 
+{
+  using value_type = T;
+};
+
+template 
+struct is_type_array< T[] > : std::true_type
+{
+  using value_type = T;
+};
+
+
 template < typename RP
          , typename Functor
          , typename Tag = void
@@ -1761,18 +1771,17 @@ struct HostIterateTile < RP , Functor , Tag , ValueType , typename std::enable_i
   RP         const& m_rp;
   Functor    const& m_func;
   typename std::conditional< std::is_same::value,int,Tag>::type m_tag;
-//  value_type  & m_v;
-
 };
 
 
-// ValueType: For reductions
+// For ParallelReduce
+// ValueType - scalar: For reductions
 template < typename RP
          , typename Functor
          , typename Tag
          , typename ValueType
          >
-struct HostIterateTile < RP , Functor , Tag , ValueType , typename std::enable_if< !is_void::value >::type >
+struct HostIterateTile < RP , Functor , Tag , ValueType , typename std::enable_if< !is_void::value && !is_type_array::value >::type >
 {
   using index_type = typename RP::index_type;
   using point_type = typename RP::point_type;
@@ -2251,12 +2260,497 @@ struct HostIterateTile < RP , Functor , Tag , ValueType , typename std::enable_i
 };
 
 
+// For ParallelReduce
+// Extra specialization for array reductions
+// ValueType[]: For array reductions
+template < typename RP
+         , typename Functor
+         , typename Tag
+         , typename ValueType
+         >
+struct HostIterateTile < RP , Functor , Tag , ValueType , typename std::enable_if< !is_void::value && is_type_array::value >::type >
+{
+  using index_type = typename RP::index_type;
+  using point_type = typename RP::point_type;
+
+  using value_type = typename is_type_array::value_type; // strip away the 'array-ness' [], only underlying type remains
+
+  inline
+  HostIterateTile( RP const& rp, Functor const& func, value_type *v ) // v should be an array; treat as pointer for compatibility since size is not known nor needed here
+    : m_rp(rp) //Cuda 7.0 does not like braces...
+    , m_func(func)
+    , m_v(v) // use with non-void ValueType struct
+  {}
+
+  inline
+  bool check_iteration_bounds( point_type& partial_tile , point_type& offset ) const {
+    bool is_full_tile = true;
+
+      for ( int i = 0; i < RP::rank; ++i ) {
+        if ((offset[i] + m_rp.m_tile[i]) <= m_rp.m_upper[i]) {
+            partial_tile[i] = m_rp.m_tile[i] ;
+        }
+        else {
+          is_full_tile = false ;
+            partial_tile[i] = (m_rp.m_upper[i] - 1 - offset[i]) == 0 ? 1
+                            : (m_rp.m_upper[i] - m_rp.m_tile[i]) > 0 ? (m_rp.m_upper[i] - offset[i])
+                            : (m_rp.m_upper[i] - m_rp.m_lower[i]) ; // when single tile encloses range
+        }
+      }
+
+    return is_full_tile ;
+  } // end check bounds
+
+
+  template 
+  struct RankTag
+  {
+    typedef RankTag type;
+    enum { value = (int)Rank };
+  };
+
+
+#if KOKKOS_ENABLE_NEW_LOOP_MACROS
+  template 
+  inline
+  void
+  operator()(IType tile_idx) const
+  {
+    point_type m_offset;
+    point_type m_tiledims;
+
+    if (RP::outer_direction == RP::Left) {
+      for (int i=0; i=0; --i) {
+        m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+        tile_idx /= m_rp.m_tile_end[i];
+      }
+    }
+
+    //Check if offset+tiledim in bounds - if not, replace tile dims with the partial tile dims
+    const bool full_tile = check_iteration_bounds(m_tiledims , m_offset) ;
+
+    Tile_Loop_Type< RP::rank, (RP::inner_direction == RP::Left), index_type, Tag >::apply( m_v, m_func, full_tile, m_offset, m_rp.m_tile, m_tiledims );
+
+  }
+
+#else
+  template 
+  inline
+  void
+  operator()(IType tile_idx) const
+  { operator_impl( tile_idx , RankTag() ); }
+  // added due to compiler error when using sfinae to choose operator based on rank
+
+
+  template 
+  inline
+  void operator_impl( IType tile_idx , const RankTag<2> ) const
+  {
+    point_type m_offset;
+    point_type m_tiledims;
+
+    if (RP::outer_direction == RP::Left) {
+      for (int i=0; i=0; --i) {
+        m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+        tile_idx /= m_rp.m_tile_end[i];
+      }
+    }
+
+    //Check if offset+tiledim in bounds - if not, replace tile dims with the partial tile dims
+    const bool full_tile = check_iteration_bounds(m_tiledims , m_offset) ;
+
+    if (RP::inner_direction == RP::Left) {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_2L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_2 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_2L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_2 );
+        }
+      }
+    } // end RP::Left
+    else {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_2R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_2 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_2R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_2 );
+        }
+      }
+    } // end RP::Right
+
+  } //end op() rank == 2
+
+
+  template 
+  inline
+  void operator_impl( IType tile_idx , const RankTag<3> ) const
+  {
+    point_type m_offset;
+    point_type m_tiledims;
+
+    if (RP::outer_direction == RP::Left) {
+      for (int i=0; i=0; --i) {
+        m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+        tile_idx /= m_rp.m_tile_end[i];
+      }
+    }
+
+    //Check if offset+tiledim in bounds - if not, replace tile dims with the partial tile dims
+    const bool full_tile = check_iteration_bounds(m_tiledims , m_offset) ;
+
+    if (RP::inner_direction == RP::Left) {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_3L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_3 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_3L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_3 );
+        }
+      }
+    } // end RP::Left
+    else {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_3R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_3 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_3R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_3 );
+        }
+      }
+    } // end RP::Right
+
+  } //end op() rank == 3
+
+
+  template 
+  inline
+  void operator_impl( IType tile_idx , const RankTag<4> ) const
+  {
+    point_type m_offset;
+    point_type m_tiledims;
+
+    if (RP::outer_direction == RP::Left) {
+      for (int i=0; i=0; --i) {
+        m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+        tile_idx /= m_rp.m_tile_end[i];
+      }
+    }
+
+    //Check if offset+tiledim in bounds - if not, replace tile dims with the partial tile dims
+    const bool full_tile = check_iteration_bounds(m_tiledims , m_offset) ;
+
+    if (RP::inner_direction == RP::Left) {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_4L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_4 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_4L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_4 );
+        }
+      }
+    } // end RP::Left
+    else {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_4R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_4 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_4R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_4 );
+        }
+      }
+    } // end RP::Right
+
+  } //end op() rank == 4
+
+
+  template 
+  inline
+  void operator_impl( IType tile_idx , const RankTag<5> ) const
+  {
+    point_type m_offset;
+    point_type m_tiledims;
+
+    if (RP::outer_direction == RP::Left) {
+      for (int i=0; i=0; --i) {
+        m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+        tile_idx /= m_rp.m_tile_end[i];
+      }
+    }
+
+    //Check if offset+tiledim in bounds - if not, replace tile dims with the partial tile dims
+    const bool full_tile = check_iteration_bounds(m_tiledims , m_offset) ;
+
+    if (RP::inner_direction == RP::Left) {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_5L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_5 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_5L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_5 );
+        }
+      }
+    } // end RP::Left
+    else {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_5R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_5 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_5R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_5 );
+        }
+      }
+    } // end RP::Right
+
+  } //end op() rank == 5
+
+
+  template 
+  inline
+  void operator_impl( IType tile_idx , const RankTag<6> ) const
+  {
+    point_type m_offset;
+    point_type m_tiledims;
+
+    if (RP::outer_direction == RP::Left) {
+      for (int i=0; i=0; --i) {
+        m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+        tile_idx /= m_rp.m_tile_end[i];
+      }
+    }
+
+    //Check if offset+tiledim in bounds - if not, replace tile dims with the partial tile dims
+    const bool full_tile = check_iteration_bounds(m_tiledims , m_offset) ;
+
+    if (RP::inner_direction == RP::Left) {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_6L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_6 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_6L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_6 );
+        }
+      }
+    } // end RP::Left
+    else {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_6R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_6 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_6R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_6 );
+        }
+      }
+    } // end RP::Right
+
+  } //end op() rank == 6
+
+
+  template 
+  inline
+  void operator_impl( IType tile_idx , const RankTag<7> ) const
+  {
+    point_type m_offset;
+    point_type m_tiledims;
+
+    if (RP::outer_direction == RP::Left) {
+      for (int i=0; i=0; --i) {
+        m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+        tile_idx /= m_rp.m_tile_end[i];
+      }
+    }
+
+    //Check if offset+tiledim in bounds - if not, replace tile dims with the partial tile dims
+    const bool full_tile = check_iteration_bounds(m_tiledims , m_offset) ;
+
+    if (RP::inner_direction == RP::Left) {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_7L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_7 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_7L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_7 );
+        }
+      }
+    } // end RP::Left
+    else {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_7R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_7 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_7R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_7 );
+        }
+      }
+    } // end RP::Right
+
+  } //end op() rank == 7
+
+
+  template 
+  inline
+  void operator_impl( IType tile_idx , const RankTag<8> ) const
+  {
+    point_type m_offset;
+    point_type m_tiledims;
+
+    if (RP::outer_direction == RP::Left) {
+      for (int i=0; i=0; --i) {
+        m_offset[i] = (tile_idx % m_rp.m_tile_end[i]) * m_rp.m_tile[i] + m_rp.m_lower[i] ;
+        tile_idx /= m_rp.m_tile_end[i];
+      }
+    }
+
+    //Check if offset+tiledim in bounds - if not, replace tile dims with the partial tile dims
+    const bool full_tile = check_iteration_bounds(m_tiledims , m_offset) ;
+
+    if (RP::inner_direction == RP::Left) {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_8L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_8 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_8L(index_type, m_tiledims) {
+          apply( LOOP_ARGS_8 );
+        }
+      }
+    } // end RP::Left
+    else {
+     if ( full_tile ) {
+//      #pragma simd
+        LOOP_8R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_8 );
+        }
+      } else {
+//      #pragma simd
+        LOOP_8R(index_type, m_tiledims) {
+          apply( LOOP_ARGS_8 );
+        }
+      }
+    } // end RP::Right
+
+  } //end op() rank == 8
+#endif
+
+
+    template 
+    typename std::enable_if<( sizeof...(Args) == RP::rank && std::is_same::value), void>::type
+    apply(Args &&... args) const
+    {
+      m_func(args... , m_v);
+    }
+
+    template 
+    typename std::enable_if<( sizeof...(Args) == RP::rank && !std::is_same::value), void>::type
+    apply(Args &&... args) const
+    {
+      m_func( m_tag, args... , m_v);
+    }
+
+
+  RP         const& m_rp;
+  Functor    const& m_func;
+  value_type * m_v;
+  typename std::conditional< std::is_same::value,int,Tag>::type m_tag;
+
+};
+
+
 // ------------------------------------------------------------------ //
 
 // MDFunctor - wraps the range_policy and functor to pass to IterateTile
-// Serial, Threads, OpenMP
+// Used for md_parallel_{for,reduce} with Serial, Threads, OpenMP
 // Cuda uses DeviceIterateTile directly within md_parallel_for
-// ParallelReduce
+// TODO Once md_parallel_{for,reduce} removed, this can be removed
+
+// ParallelReduce - scalar reductions
 template < typename MDRange, typename Functor, typename ValueType = void >
 struct MDFunctor
 {
@@ -2273,7 +2767,7 @@ struct MDFunctor
 
 
   inline
-  MDFunctor( MDRange const& range, Functor const& f, ValueType & v )
+  MDFunctor( MDRange const& range, Functor const& f )
     : m_range( range )
     , m_func( f )
   {}
@@ -2290,7 +2784,6 @@ struct MDFunctor
   inline
   MDFunctor& operator=( MDFunctor && ) = default;
 
-//  KOKKOS_FORCEINLINE_FUNCTION //Caused cuda warning - __host__ warning
   inline
   void operator()(index_type t, value_type & v) const
   {
@@ -2301,6 +2794,56 @@ struct MDFunctor
   Functor   m_func;
 };
 
+
+// ParallelReduce - array reductions 
+template < typename MDRange, typename Functor, typename ValueType >
+struct MDFunctor< MDRange, Functor, ValueType[] >
+{
+  using range_policy = MDRange;
+  using functor_type = Functor;
+  using value_type   = ValueType[];
+  using work_tag     = typename range_policy::work_tag;
+  using index_type   = typename range_policy::index_type;
+  using iterate_type = typename Kokkos::Experimental::Impl::HostIterateTile< MDRange
+                                                                           , Functor
+                                                                           , work_tag
+                                                                           , value_type
+                                                                           >;
+
+
+  inline
+  MDFunctor( MDRange const& range, Functor const& f )
+    : m_range( range )
+    , m_func( f )
+    , value_count( f.value_count )
+  {}
+
+  inline
+  MDFunctor( MDFunctor const& ) = default;
+
+  inline
+  MDFunctor& operator=( MDFunctor const& ) = default;
+
+  inline
+  MDFunctor( MDFunctor && ) = default;
+
+  inline
+  MDFunctor& operator=( MDFunctor && ) = default;
+
+  // FIXME Init and Join, as defined in m_func, are not working through the MDFunctor
+  // Best path forward is to eliminate need for MDFunctor, directly use MDRangePolicy within Parallel{For,Reduce} ??
+  inline
+  void operator()(index_type t, value_type v) const
+  {
+    iterate_type(m_range, m_func, v)(t);
+  }
+
+  MDRange   m_range;
+  Functor   m_func;
+  size_t    value_count;
+};
+
+
 // ParallelFor
 template < typename MDRange, typename Functor >
 struct MDFunctor< MDRange, Functor, void >
@@ -2349,4 +2892,3 @@ struct MDFunctor< MDRange, Functor, void >
 } } } //end namespace Kokkos::Experimental::Impl
 
 #endif
-
diff --git a/lib/kokkos/core/src/impl/Kokkos_AnalyzePolicy.hpp b/lib/kokkos/core/src/impl/Kokkos_AnalyzePolicy.hpp
index c5685c5b62..3fb15c8d1e 100644
--- a/lib/kokkos/core/src/impl/Kokkos_AnalyzePolicy.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_AnalyzePolicy.hpp
@@ -55,16 +55,19 @@ template < typename ExecutionSpace   = void
          , typename WorkTag          = void
          , typename IndexType        = void
          , typename IterationPattern = void
+         , typename LaunchBounds     = void
          >
 struct PolicyTraitsBase
 {
-  using type = PolicyTraitsBase< ExecutionSpace, Schedule, WorkTag, IndexType, IterationPattern>;
+  using type = PolicyTraitsBase< ExecutionSpace, Schedule, WorkTag, IndexType, 
+               IterationPattern, LaunchBounds>;
 
   using execution_space   = ExecutionSpace;
   using schedule_type     = Schedule;
   using work_tag          = WorkTag;
   using index_type        = IndexType;
   using iteration_pattern = IterationPattern;
+  using launch_bounds     = LaunchBounds;
 };
 
 
@@ -78,6 +81,7 @@ struct SetExecutionSpace
                                , typename PolicyBase::work_tag
                                , typename PolicyBase::index_type
                                , typename PolicyBase::iteration_pattern
+                               , typename PolicyBase::launch_bounds
                                >;
 };
 
@@ -91,6 +95,7 @@ struct SetSchedule
                                , typename PolicyBase::work_tag
                                , typename PolicyBase::index_type
                                , typename PolicyBase::iteration_pattern
+                               , typename PolicyBase::launch_bounds
                                >;
 };
 
@@ -104,6 +109,7 @@ struct SetWorkTag
                                , WorkTag
                                , typename PolicyBase::index_type
                                , typename PolicyBase::iteration_pattern
+                               , typename PolicyBase::launch_bounds
                                >;
 };
 
@@ -117,6 +123,7 @@ struct SetIndexType
                                , typename PolicyBase::work_tag
                                , IndexType
                                , typename PolicyBase::iteration_pattern
+                               , typename PolicyBase::launch_bounds
                                >;
 };
 
@@ -131,6 +138,22 @@ struct SetIterationPattern
                                , typename PolicyBase::work_tag
                                , typename PolicyBase::index_type
                                , IterationPattern
+                               , typename PolicyBase::launch_bounds
+                               >;
+};
+
+
+template 
+struct SetLaunchBounds
+{
+  static_assert( is_void::value
+               , "Kokkos Error: More than one launch_bounds given" );
+  using type = PolicyTraitsBase< typename PolicyBase::execution_space
+                               , typename PolicyBase::schedule_type
+                               , typename PolicyBase::work_tag
+                               , typename PolicyBase::index_type
+                               , typename PolicyBase::iteration_pattern
+                               , LaunchBounds
                                >;
 };
 
@@ -146,8 +169,9 @@ struct AnalyzePolicy : public
     , typename std::conditional< is_index_type::value       , SetIndexType
     , typename std::conditional< std::is_integral::value    , SetIndexType >
     , typename std::conditional< is_iteration_pattern::value, SetIterationPattern
+    , typename std::conditional< is_launch_bounds::value    , SetLaunchBounds
     , SetWorkTag
-    >::type >::type >::type >::type>::type::type
+    >::type >::type >::type >::type >::type>::type::type
   , Traits...
   >
 {};
@@ -178,11 +202,18 @@ struct AnalyzePolicy
                                                      , void // TODO set default iteration pattern
                                                      , typename Base::iteration_pattern
                                                      >::type;
+
+  using launch_bounds = typename std::conditional< is_void< typename Base::launch_bounds >::value
+                                                     , LaunchBounds<>
+                                                     , typename Base::launch_bounds
+                                                     >::type;
+
   using type = PolicyTraitsBase< execution_space
                                , schedule_type
                                , work_tag
                                , index_type
                                , iteration_pattern
+                               , launch_bounds
                                >;
 };
 
diff --git a/lib/kokkos/core/src/impl/Kokkos_Atomic_Compare_Exchange_Strong.hpp b/lib/kokkos/core/src/impl/Kokkos_Atomic_Compare_Exchange_Strong.hpp
index 010b15064e..5b894b037b 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Atomic_Compare_Exchange_Strong.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Atomic_Compare_Exchange_Strong.hpp
@@ -41,6 +41,10 @@
 //@HEADER
 */
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+#include 
+#endif
+
 #include 
 #if defined( KOKKOS_ATOMIC_HPP ) && ! defined( KOKKOS_ATOMIC_COMPARE_EXCHANGE_STRONG_HPP )
 #define KOKKOS_ATOMIC_COMPARE_EXCHANGE_STRONG_HPP
@@ -126,11 +130,21 @@ T atomic_compare_exchange( volatile T * const dest , const T & compare ,
 
 inline
 int atomic_compare_exchange( volatile int * const dest, const int compare, const int val)
-{ return __sync_val_compare_and_swap(dest,compare,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_val_compare_and_swap(dest,compare,val);
+}
 
 inline
 long atomic_compare_exchange( volatile long * const dest, const long compare, const long val )
-{ return __sync_val_compare_and_swap(dest,compare,val); }
+{ 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_val_compare_and_swap(dest,compare,val);
+}
 
 #if defined( KOKKOS_ENABLE_GNU_ATOMICS )
 
@@ -159,6 +173,10 @@ T atomic_compare_exchange( volatile T * const dest, const T & compare,
     KOKKOS_INLINE_FUNCTION U() {};
   } tmp ;
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   tmp.i = __sync_val_compare_and_swap( (int*) dest , *((int*)&compare) , *((int*)&val) );
   return tmp.t ;
 }
@@ -175,6 +193,10 @@ T atomic_compare_exchange( volatile T * const dest, const T & compare,
     KOKKOS_INLINE_FUNCTION U() {};
   } tmp ;
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   tmp.i = __sync_val_compare_and_swap( (long*) dest , *((long*)&compare) , *((long*)&val) );
   return tmp.t ;
 }
@@ -193,6 +215,10 @@ T atomic_compare_exchange( volatile T * const dest, const T & compare,
     KOKKOS_INLINE_FUNCTION U() {};
   } tmp ;
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   tmp.i = Impl::cas128( (Impl::cas128_t*) dest , *((Impl::cas128_t*)&compare) , *((Impl::cas128_t*)&val) );
   return tmp.t ;
 }
@@ -209,6 +235,10 @@ T atomic_compare_exchange( volatile T * const dest , const T compare ,
             #endif
              , const T >::type& val )
 {
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   while( !Impl::lock_address_host_space( (void*) dest ) );
   T return_val = *dest;
   if( return_val == compare ) {
diff --git a/lib/kokkos/core/src/impl/Kokkos_Atomic_Decrement.hpp b/lib/kokkos/core/src/impl/Kokkos_Atomic_Decrement.hpp
index 127de528f5..2a13a4865c 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Atomic_Decrement.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Atomic_Decrement.hpp
@@ -41,6 +41,10 @@
 //@HEADER
 */
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+#include 
+#endif
+
 #include 
 #if defined( KOKKOS_ATOMIC_HPP) && ! defined( KOKKOS_ATOMIC_DECREMENT_HPP )
 #define KOKKOS_ATOMIC_DECREMENT_HPP
@@ -54,6 +58,10 @@ template<>
 KOKKOS_INLINE_FUNCTION
 void atomic_decrement(volatile char* a) {
 #if defined( KOKKOS_ENABLE_ASM ) && defined( KOKKOS_ENABLE_ISA_X86_64 ) && ! defined(_WIN32) && ! defined(__CUDA_ARCH__)
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) a, _MM_HINT_ET0 );
+#endif
+
   __asm__ __volatile__(
       "lock decb %0"
       : /* no output registers */
@@ -69,6 +77,10 @@ template<>
 KOKKOS_INLINE_FUNCTION
 void atomic_decrement(volatile short* a) {
 #if defined( KOKKOS_ENABLE_ASM ) && defined( KOKKOS_ENABLE_ISA_X86_64 ) && ! defined(_WIN32) && ! defined(__CUDA_ARCH__)
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) a, _MM_HINT_ET0 );
+#endif
+
   __asm__ __volatile__(
       "lock decw %0"
       : /* no output registers */
@@ -84,6 +96,10 @@ template<>
 KOKKOS_INLINE_FUNCTION
 void atomic_decrement(volatile int* a) {
 #if defined( KOKKOS_ENABLE_ASM ) && defined( KOKKOS_ENABLE_ISA_X86_64 ) && ! defined(_WIN32) && ! defined(__CUDA_ARCH__)
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) a, _MM_HINT_ET0 );
+#endif
+
   __asm__ __volatile__(
       "lock decl %0"
       : /* no output registers */
@@ -99,6 +115,9 @@ template<>
 KOKKOS_INLINE_FUNCTION
 void atomic_decrement(volatile long long int* a) {
 #if defined( KOKKOS_ENABLE_ASM ) && defined( KOKKOS_ENABLE_ISA_X86_64 ) && ! defined(_WIN32) && ! defined(__CUDA_ARCH__)
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) a, _MM_HINT_ET0 );
+#endif
   __asm__ __volatile__(
       "lock decq %0"
       : /* no output registers */
diff --git a/lib/kokkos/core/src/impl/Kokkos_Atomic_Exchange.hpp b/lib/kokkos/core/src/impl/Kokkos_Atomic_Exchange.hpp
index a1ff47abce..9ba3cae9fc 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Atomic_Exchange.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Atomic_Exchange.hpp
@@ -41,6 +41,10 @@
 //@HEADER
 */
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+#include 
+#endif
+
 #include 
 #if defined( KOKKOS_ATOMIC_HPP ) && ! defined( KOKKOS_ATOMIC_EXCHANGE_HPP )
 #define KOKKOS_ATOMIC_EXCHANGE_HPP
@@ -81,6 +85,10 @@ T atomic_exchange(
   typename Kokkos::Impl::enable_if< sizeof(T) == sizeof(int) , const T & >::type val )
 {
   // int tmp = __ullAtomicExch( (int*) dest , *((int*)&val) );
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   int tmp = atomicExch( ((int*)dest) , *((int*)&val) );
   return *((T*)&tmp);
 }
@@ -93,6 +101,11 @@ T atomic_exchange(
                                     sizeof(T) == sizeof(unsigned long long int) , const T & >::type val )
 {
   typedef unsigned long long int type ;
+
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   // type tmp = __ullAtomicExch( (type*) dest , *((type*)&val) );
   type tmp = atomicExch( ((type*)dest) , *((type*)&val) );
   return *((T*)&tmp);
@@ -108,6 +121,10 @@ T atomic_exchange( volatile T * const dest ,
 {
   T return_val;
   // This is a way to (hopefully) avoid dead lock in a warp
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   int done = 0;
   unsigned int active = __ballot(1);
   unsigned int done_active = 0;
@@ -173,6 +190,9 @@ T atomic_exchange( volatile T * const dest ,
                                   , const T & >::type val )
 {
   typedef typename Kokkos::Impl::if_c< sizeof(T) == sizeof(int) , int , long >::type type ;
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
 
   const type v = *((type*)&val); // Extract to be sure the value doesn't change
 
@@ -201,6 +221,10 @@ T atomic_exchange( volatile T * const dest ,
   typename Kokkos::Impl::enable_if< sizeof(T) == sizeof(Impl::cas128_t)
                                   , const T & >::type val )
 {
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   union U {
     Impl::cas128_t i ;
     T t ;
@@ -260,6 +284,10 @@ void atomic_assign( volatile T * const dest ,
 {
   typedef typename Kokkos::Impl::if_c< sizeof(T) == sizeof(int) , int , long >::type type ;
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   const type v = *((type*)&val); // Extract to be sure the value doesn't change
 
   type assumed ;
@@ -285,6 +313,10 @@ void atomic_assign( volatile T * const dest ,
   typename Kokkos::Impl::enable_if< sizeof(T) == sizeof(Impl::cas128_t)
                                   , const T & >::type val )
 {
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   union U {
     Impl::cas128_t i ;
     T t ;
diff --git a/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Add.hpp b/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Add.hpp
index 860c8e0e43..084c55efed 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Add.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Add.hpp
@@ -41,6 +41,10 @@
 //@HEADER
 */
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+#include 
+#endif
+
 #include 
 #if defined( KOKKOS_ATOMIC_HPP ) && ! defined( KOKKOS_ATOMIC_FETCH_ADD_HPP )
 #define KOKKOS_ATOMIC_FETCH_ADD_HPP
@@ -161,36 +165,60 @@ T atomic_fetch_add( volatile T * const dest ,
 inline
 int atomic_fetch_add( volatile int * dest , const int val )
 {
-        int original = val;
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH ) 
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
 
-        __asm__ __volatile__(
-                "lock xadd %1, %0"
-                : "+m" (*dest), "+r" (original)
-                : "m" (*dest), "r" (original)
-                : "memory"
+  int original = val;
+
+  __asm__ __volatile__(
+  	"lock xadd %1, %0"
+        : "+m" (*dest), "+r" (original)
+        : "m" (*dest), "r" (original)
+        : "memory"
         );
 
-        return original;
+  return original;
 }
 #else
 inline
 int atomic_fetch_add( volatile int * const dest , const int val )
-{ return __sync_fetch_and_add(dest, val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_add(dest, val);
+}
 #endif
 
 inline
 long int atomic_fetch_add( volatile long int * const dest , const long int val )
-{ return __sync_fetch_and_add(dest,val); }
+{ 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_add(dest,val);
+}
 
 #if defined( KOKKOS_ENABLE_GNU_ATOMICS )
 
 inline
 unsigned int atomic_fetch_add( volatile unsigned int * const dest , const unsigned int val )
-{ return __sync_fetch_and_add(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_add(dest,val);
+}
 
 inline
 unsigned long int atomic_fetch_add( volatile unsigned long int * const dest , const unsigned long int val )
-{ return __sync_fetch_and_add(dest,val); }
+{ 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_add(dest,val);
+}
 
 #endif
 
@@ -205,6 +233,10 @@ T atomic_fetch_add( volatile T * const dest ,
     inline U() {};
   } assume , oldval , newval ;
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   oldval.t = *dest ;
 
   do {
@@ -228,6 +260,10 @@ T atomic_fetch_add( volatile T * const dest ,
     inline U() {};
   } assume , oldval , newval ;
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   oldval.t = *dest ;
 
   do {
@@ -253,6 +289,10 @@ T atomic_fetch_add( volatile T * const dest ,
     inline U() {};
   } assume , oldval , newval ;
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   oldval.t = *dest ;
 
   do {
diff --git a/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_And.hpp b/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_And.hpp
index 83f5b2a5aa..6ecb65336c 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_And.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_And.hpp
@@ -41,6 +41,10 @@
 //@HEADER
 */
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+#include 
+#endif
+
 #include 
 #if defined( KOKKOS_ATOMIC_HPP ) && ! defined( KOKKOS_ATOMIC_FETCH_AND_HPP )
 #define KOKKOS_ATOMIC_FETCH_AND_HPP
@@ -76,21 +80,41 @@ unsigned long long int atomic_fetch_and( volatile unsigned long long int * const
 
 inline
 int atomic_fetch_and( volatile int * const dest , const int val )
-{ return __sync_fetch_and_and(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_and(dest,val);
+}
 
 inline
 long int atomic_fetch_and( volatile long int * const dest , const long int val )
-{ return __sync_fetch_and_and(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_and(dest,val);
+}
 
 #if defined( KOKKOS_ENABLE_GNU_ATOMICS )
 
 inline
 unsigned int atomic_fetch_and( volatile unsigned int * const dest , const unsigned int val )
-{ return __sync_fetch_and_and(dest,val); }
+{ 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_and(dest,val);
+}
 
 inline
 unsigned long int atomic_fetch_and( volatile unsigned long int * const dest , const unsigned long int val )
-{ return __sync_fetch_and_and(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_and(dest,val);
+}
 
 #endif
 
diff --git a/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Or.hpp b/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Or.hpp
index 8c73b4c3ef..ed3b438f89 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Or.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Or.hpp
@@ -41,6 +41,10 @@
 //@HEADER
 */
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+#include 
+#endif
+
 #include 
 #if defined( KOKKOS_ATOMIC_HPP ) && ! defined( KOKKOS_ATOMIC_FETCH_OR_HPP )
 #define KOKKOS_ATOMIC_FETCH_OR_HPP
@@ -76,21 +80,41 @@ unsigned long long int atomic_fetch_or( volatile unsigned long long int * const
 
 inline
 int atomic_fetch_or( volatile int * const dest , const int val )
-{ return __sync_fetch_and_or(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_or(dest,val);
+}
 
 inline
 long int atomic_fetch_or( volatile long int * const dest , const long int val )
-{ return __sync_fetch_and_or(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_or(dest,val);
+}
 
 #if defined( KOKKOS_ENABLE_GNU_ATOMICS )
 
 inline
 unsigned int atomic_fetch_or( volatile unsigned int * const dest , const unsigned int val )
-{ return __sync_fetch_and_or(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_or(dest,val);
+}
 
 inline
 unsigned long int atomic_fetch_or( volatile unsigned long int * const dest , const unsigned long int val )
-{ return __sync_fetch_and_or(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_or(dest,val);
+}
 
 #endif
 
diff --git a/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Sub.hpp b/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Sub.hpp
index 504731d3a2..038cc13e9a 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Sub.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Atomic_Fetch_Sub.hpp
@@ -41,6 +41,10 @@
 //@HEADER
 */
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+#include 
+#endif
+
 #include 
 #if defined( KOKKOS_ATOMIC_HPP ) && ! defined( KOKKOS_ATOMIC_FETCH_SUB_HPP )
 #define KOKKOS_ATOMIC_FETCH_SUB_HPP
@@ -136,21 +140,41 @@ T atomic_fetch_sub( volatile T * const dest ,
 
 inline
 int atomic_fetch_sub( volatile int * const dest , const int val )
-{ return __sync_fetch_and_sub(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_sub(dest,val);
+}
 
 inline
 long int atomic_fetch_sub( volatile long int * const dest , const long int val )
-{ return __sync_fetch_and_sub(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_sub(dest,val);
+}
 
 #if defined( KOKKOS_ENABLE_GNU_ATOMICS )
 
 inline
 unsigned int atomic_fetch_sub( volatile unsigned int * const dest , const unsigned int val )
-{ return __sync_fetch_and_sub(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_sub(dest,val);
+}
 
 inline
 unsigned long int atomic_fetch_sub( volatile unsigned long int * const dest , const unsigned long int val )
-{ return __sync_fetch_and_sub(dest,val); }
+{
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+  return __sync_fetch_and_sub(dest,val);
+}
 
 #endif
 
@@ -161,6 +185,10 @@ T atomic_fetch_sub( volatile T * const dest ,
 {
   union { int i ; T t ; } assume , oldval , newval ;
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   oldval.t = *dest ;
 
   do {
@@ -178,6 +206,10 @@ T atomic_fetch_sub( volatile T * const dest ,
   typename Kokkos::Impl::enable_if< sizeof(T) != sizeof(int) &&
                                     sizeof(T) == sizeof(long) , const T >::type val )
 {
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   union { long i ; T t ; } assume , oldval , newval ;
 
   oldval.t = *dest ;
@@ -202,6 +234,10 @@ T atomic_fetch_sub( volatile T * const dest ,
                && ( sizeof(T) != 8 )
              , const T >::type& val )
 {
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) dest, _MM_HINT_ET0 );
+#endif
+
   while( !Impl::lock_address_host_space( (void*) dest ) );
   T return_val = *dest;
   *dest = return_val - val;
diff --git a/lib/kokkos/core/src/impl/Kokkos_Atomic_Increment.hpp b/lib/kokkos/core/src/impl/Kokkos_Atomic_Increment.hpp
index 2985fad95e..e7626603fc 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Atomic_Increment.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Atomic_Increment.hpp
@@ -41,6 +41,10 @@
 //@HEADER
 */
 
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+#include 
+#endif
+
 #include 
 #if defined( KOKKOS_ATOMIC_HPP) && ! defined( KOKKOS_ATOMIC_INCREMENT_HPP )
 #define KOKKOS_ATOMIC_INCREMENT_HPP
@@ -52,6 +56,9 @@ template<>
 KOKKOS_INLINE_FUNCTION
 void atomic_increment(volatile char* a) {
 #if defined( KOKKOS_ENABLE_ASM ) && defined( KOKKOS_ENABLE_ISA_X86_64 ) && ! defined(_WIN32) && ! defined(__CUDA_ARCH__)
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) a, _MM_HINT_ET0 );
+#endif
   __asm__ __volatile__(
       "lock incb %0"
       : /* no output registers */
@@ -67,6 +74,9 @@ template<>
 KOKKOS_INLINE_FUNCTION
 void atomic_increment(volatile short* a) {
 #if defined( KOKKOS_ENABLE_ASM ) && defined( KOKKOS_ENABLE_ISA_X86_64 ) && ! defined(_WIN32) && ! defined(__CUDA_ARCH__)
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) a, _MM_HINT_ET0 );
+#endif
   __asm__ __volatile__(
       "lock incw %0"
       : /* no output registers */
@@ -82,6 +92,9 @@ template<>
 KOKKOS_INLINE_FUNCTION
 void atomic_increment(volatile int* a) {
 #if defined( KOKKOS_ENABLE_ASM ) && defined( KOKKOS_ENABLE_ISA_X86_64 ) && ! defined(_WIN32) && ! defined(__CUDA_ARCH__)
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) a, _MM_HINT_ET0 );
+#endif
   __asm__ __volatile__(
       "lock incl %0"
       : /* no output registers */
@@ -97,6 +110,9 @@ template<>
 KOKKOS_INLINE_FUNCTION
 void atomic_increment(volatile long long int* a) {
 #if defined( KOKKOS_ENABLE_ASM ) && defined( KOKKOS_ENABLE_ISA_X86_64 ) && ! defined(_WIN32) && ! defined(__CUDA_ARCH__)
+#if defined( KOKKOS_ENABLE_RFO_PREFETCH )
+  _mm_prefetch( (const char*) a, _MM_HINT_ET0 );
+#endif
   __asm__ __volatile__(
       "lock incq %0"
       : /* no output registers */
diff --git a/lib/kokkos/core/src/impl/Kokkos_Core.cpp b/lib/kokkos/core/src/impl/Kokkos_Core.cpp
index f0ff6d78ec..f52cc469ac 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Core.cpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Core.cpp
@@ -87,17 +87,12 @@ setenv("MEMKIND_HBW_NODES", "1", 0);
 #if defined( KOKKOS_ENABLE_OPENMP )
   if( std::is_same< Kokkos::OpenMP , Kokkos::DefaultExecutionSpace >::value ||
       std::is_same< Kokkos::OpenMP , Kokkos::HostSpace::execution_space >::value ) {
-    if(num_threads>0) {
-      if(use_numa>0) {
-        Kokkos::OpenMP::initialize(num_threads,use_numa);
-      }
-      else {
-        Kokkos::OpenMP::initialize(num_threads);
-      }
-    } else {
-      Kokkos::OpenMP::initialize();
+    if(use_numa>0) {
+      Kokkos::OpenMP::initialize(num_threads,use_numa);
+    }
+    else {
+      Kokkos::OpenMP::initialize(num_threads);
     }
-    //std::cout << "Kokkos::initialize() fyi: OpenMP enabled and initialized" << std::endl ;
   }
   else {
     //std::cout << "Kokkos::initialize() fyi: OpenMP enabled but not initialized" << std::endl ;
@@ -437,10 +432,7 @@ void initialize(int& narg, char* arg[])
       iarg++;
     }
 
-    InitArguments arguments;
-    arguments.num_threads = num_threads;
-    arguments.num_numa = numa;
-    arguments.device_id = device;
+    InitArguments arguments{num_threads, numa, device};
     Impl::initialize_internal(arguments);
 }
 
diff --git a/lib/kokkos/core/src/impl/Kokkos_FunctorAdapter.hpp b/lib/kokkos/core/src/impl/Kokkos_FunctorAdapter.hpp
index dc75fb072f..fccd8e090f 100644
--- a/lib/kokkos/core/src/impl/Kokkos_FunctorAdapter.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_FunctorAdapter.hpp
@@ -170,28 +170,31 @@ struct FunctorValueTraits< FunctorType , ArgTag , true /* == exists FunctorType:
   static_assert( 0 == ( sizeof(value_type) % sizeof(int) ) ,
     "Reduction functor's declared value_type requires: 0 == sizeof(value_type) % sizeof(int)" );
 
+  /* this cast to bool is needed for correctness by NVCC */
+  enum : bool { IsArray = static_cast(Impl::is_array< typename FunctorType::value_type >::value) };
+
   // If not an array then what is the sizeof(value_type)
-  enum { StaticValueSize = Impl::is_array< typename FunctorType::value_type >::value ? 0 : sizeof(value_type) };
+  enum { StaticValueSize = IsArray ? 0 : sizeof(value_type) };
 
   typedef value_type                 * pointer_type ;
 
   // The reference_type for an array is 'value_type *'
   // The reference_type for a single value is 'value_type &'
 
-  typedef typename Impl::if_c< ! StaticValueSize , value_type *
-                                                 , value_type & >::type  reference_type ;
+  typedef typename Impl::if_c< IsArray , value_type *
+                                       , value_type & >::type  reference_type ;
 
   // Number of values if single value
   template< class F >
   KOKKOS_FORCEINLINE_FUNCTION static
-  typename Impl::enable_if< std::is_same::value && StaticValueSize , unsigned >::type
+  typename Impl::enable_if< std::is_same::value && ! IsArray , unsigned >::type
     value_count( const F & ) { return 1 ; }
 
   // Number of values if an array, protect via templating because 'f.value_count'
   // will only exist when the functor declares the value_type to be an array.
   template< class F >
   KOKKOS_FORCEINLINE_FUNCTION static
-  typename Impl::enable_if< std::is_same::value && ! StaticValueSize , unsigned >::type
+  typename Impl::enable_if< std::is_same::value && IsArray , unsigned >::type
     value_count( const F & f ) { return f.value_count ; }
 
   // Total size of the value
diff --git a/lib/kokkos/core/src/impl/Kokkos_HBWSpace.cpp b/lib/kokkos/core/src/impl/Kokkos_HBWSpace.cpp
index 8cb7430035..e11f8b6d34 100644
--- a/lib/kokkos/core/src/impl/Kokkos_HBWSpace.cpp
+++ b/lib/kokkos/core/src/impl/Kokkos_HBWSpace.cpp
@@ -70,62 +70,6 @@
 #ifdef KOKKOS_ENABLE_HBWSPACE
 #define MEMKIND_TYPE MEMKIND_HBW //hbw_get_kind(HBW_PAGESIZE_4KB)
 
-namespace Kokkos {
-namespace Experimental {
-namespace {
-
-static const int QUERY_SPACE_IN_PARALLEL_MAX = 16 ;
-
-typedef int (* QuerySpaceInParallelPtr )();
-
-QuerySpaceInParallelPtr s_in_parallel_query[ QUERY_SPACE_IN_PARALLEL_MAX ] ;
-int s_in_parallel_query_count = 0 ;
-
-} // namespace 
-
-void HBWSpace::register_in_parallel( int (*device_in_parallel)() )
-{
-  if ( 0 == device_in_parallel ) {
-    Kokkos::Impl::throw_runtime_exception( std::string("Kokkos::Experimental::HBWSpace::register_in_parallel ERROR : given NULL" ) );
-  }
-
-  int i = -1 ;
-
-  if ( ! (device_in_parallel)() ) {
-    for ( i = 0 ; i < s_in_parallel_query_count && ! (*(s_in_parallel_query[i]))() ; ++i );
-  }
-
-  if ( i < s_in_parallel_query_count ) {
-    Kokkos::Impl::throw_runtime_exception( std::string("Kokkos::Experimental::HBWSpace::register_in_parallel_query ERROR : called in_parallel" ) );
-
-  }
-
-  if ( QUERY_SPACE_IN_PARALLEL_MAX <= i ) {
-    Kokkos::Impl::throw_runtime_exception( std::string("Kokkos::Experimental::HBWSpace::register_in_parallel_query ERROR : exceeded maximum" ) );
-
-  }
-
-  for ( i = 0 ; i < s_in_parallel_query_count && s_in_parallel_query[i] != device_in_parallel ; ++i );
-
-  if ( i == s_in_parallel_query_count ) {
-    s_in_parallel_query[s_in_parallel_query_count++] = device_in_parallel ;
-  }
-}
-
-int HBWSpace::in_parallel()
-{
-  const int n = s_in_parallel_query_count ;
-
-  int i = 0 ;
-
-  while ( i < n && ! (*(s_in_parallel_query[i]))() ) { ++i ; }
-
-  return i < n ;
-}
-
-} // namespace Experiemtal
-} // namespace Kokkos
-
 /*--------------------------------------------------------------------------*/
 
 namespace Kokkos {
diff --git a/lib/kokkos/core/src/impl/Kokkos_HostSpace.cpp b/lib/kokkos/core/src/impl/Kokkos_HostSpace.cpp
index 2a5c34c375..a5a73ddebb 100644
--- a/lib/kokkos/core/src/impl/Kokkos_HostSpace.cpp
+++ b/lib/kokkos/core/src/impl/Kokkos_HostSpace.cpp
@@ -106,62 +106,6 @@
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
-namespace Kokkos {
-namespace {
-
-static const int QUERY_SPACE_IN_PARALLEL_MAX = 16 ;
-
-typedef int (* QuerySpaceInParallelPtr )();
-
-QuerySpaceInParallelPtr s_in_parallel_query[ QUERY_SPACE_IN_PARALLEL_MAX ] ;
-int s_in_parallel_query_count = 0 ;
-
-} // namespace 
-
-void HostSpace::register_in_parallel( int (*device_in_parallel)() )
-{
-  if ( 0 == device_in_parallel ) {
-    Kokkos::Impl::throw_runtime_exception( std::string("Kokkos::HostSpace::register_in_parallel ERROR : given NULL" ) );
-  }
-
-  int i = -1 ;
-
-  if ( ! (device_in_parallel)() ) {
-    for ( i = 0 ; i < s_in_parallel_query_count && ! (*(s_in_parallel_query[i]))() ; ++i );
-  }
-
-  if ( i < s_in_parallel_query_count ) {
-    Kokkos::Impl::throw_runtime_exception( std::string("Kokkos::HostSpace::register_in_parallel_query ERROR : called in_parallel" ) );
-
-  }
-
-  if ( QUERY_SPACE_IN_PARALLEL_MAX <= i ) {
-    Kokkos::Impl::throw_runtime_exception( std::string("Kokkos::HostSpace::register_in_parallel_query ERROR : exceeded maximum" ) );
-
-  }
-
-  for ( i = 0 ; i < s_in_parallel_query_count && s_in_parallel_query[i] != device_in_parallel ; ++i );
-
-  if ( i == s_in_parallel_query_count ) {
-    s_in_parallel_query[s_in_parallel_query_count++] = device_in_parallel ;
-  }
-}
-
-int HostSpace::in_parallel()
-{
-  const int n = s_in_parallel_query_count ;
-
-  int i = 0 ;
-
-  while ( i < n && ! (*(s_in_parallel_query[i]))() ) { ++i ; }
-
-  return i < n ;
-}
-
-} // namespace Kokkos
-
-/*--------------------------------------------------------------------------*/
-
 namespace Kokkos {
 
 /* Default allocation mechanism */
@@ -340,9 +284,6 @@ void HostSpace::deallocate( void * const arg_alloc_ptr , const size_t arg_alloc_
   }
 }
 
-constexpr const char* HostSpace::name() {
-  return m_name;
-}
 } // namespace Kokkos
 
 //----------------------------------------------------------------------------
diff --git a/lib/kokkos/core/src/impl/Kokkos_HostThreadTeam.cpp b/lib/kokkos/core/src/impl/Kokkos_HostThreadTeam.cpp
index ac200209c7..d2446bde09 100644
--- a/lib/kokkos/core/src/impl/Kokkos_HostThreadTeam.cpp
+++ b/lib/kokkos/core/src/impl/Kokkos_HostThreadTeam.cpp
@@ -45,7 +45,7 @@
 #include 
 #include 
 #include 
-#include 
+#include 
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
@@ -58,9 +58,11 @@ void HostThreadTeamData::organize_pool
 {
   bool ok = true ;
 
+  memory_fence();
+
   // Verify not already a member of a pool:
   for ( int rank = 0 ; rank < size && ok ; ++rank ) {
-    ok = ( 0 != members[rank] ) && ( 0 == members[rank]->m_pool_scratch );
+    ok = ( nullptr != members[rank] ) && ( 0 == members[rank]->m_pool_scratch );
   }
 
   if ( ok ) {
@@ -89,7 +91,6 @@ void HostThreadTeamData::organize_pool
         mem->m_team_alloc   = 1 ;
         mem->m_league_rank  = rank ;
         mem->m_league_size  = size ;
-        mem->m_pool_rendezvous_step = 0 ;
         mem->m_team_rendezvous_step = 0 ;
         pool[ rank ] = mem ;
       }
@@ -116,7 +117,6 @@ void HostThreadTeamData::disband_pool()
    m_team_alloc   = 1 ;
    m_league_rank  = 0 ;
    m_league_size  = 1 ;
-   m_pool_rendezvous_step = 0 ;
    m_team_rendezvous_step = 0 ;
 }
 
@@ -256,11 +256,6 @@ int HostThreadTeamData::rendezvous( int64_t * const buffer
 
   const int sync_offset = ( step & mask_mem_cycle ) + size_mem_cycle ;
 
-  union {
-    int64_t full ;
-    int8_t  byte[8] ;
-  } value ;
-
   if ( rank ) {
 
     const int group_begin = rank << shift_byte ; // == rank * size_byte
@@ -275,13 +270,14 @@ int HostThreadTeamData::rendezvous( int64_t * const buffer
       const int end = group_begin + size_byte < size
                     ? size_byte : size - group_begin ;
 
-      value.full = 0 ;
-      for ( int i = 0 ; i < end ; ++i ) value.byte[i] = int8_t( step );
+      int64_t value = 0 ;
 
-      store_fence(); // This should not be needed but fixes #742
+      for ( int i = 0 ; i < end ; ++i ) {
+        ((int8_t*) & value )[i] = int8_t( step );
+      }
 
       spinwait_until_equal( buffer[ (rank << shift_mem_cycle) + sync_offset ]
-                          , value.full );
+                          , value );
     }
 
     {
@@ -316,10 +312,12 @@ int HostThreadTeamData::rendezvous( int64_t * const buffer
 
     const int end = size_byte < size ? 8 : size ;
 
-    value.full = 0 ;
-    for ( int i = 1 ; i < end ; ++i ) value.byte[i] = int8_t( step );
+    int64_t value = 0 ;
+    for ( int i = 1 ; i < end ; ++i ) {
+      ((int8_t *) & value)[i] = int8_t( step );
+    }
 
-    spinwait_until_equal( buffer[ sync_offset ], value.full );
+    spinwait_until_equal( buffer[ sync_offset ], value );
   }
 
   return rank ? 0 : 1 ;
diff --git a/lib/kokkos/core/src/impl/Kokkos_HostThreadTeam.hpp b/lib/kokkos/core/src/impl/Kokkos_HostThreadTeam.hpp
index c050a16eae..7facc0a410 100644
--- a/lib/kokkos/core/src/impl/Kokkos_HostThreadTeam.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_HostThreadTeam.hpp
@@ -50,6 +50,7 @@
 #include 
 #include 
 #include 
+#include 
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
@@ -67,14 +68,12 @@ public:
 
   // Assume upper bounds on number of threads:
   //   pool size       <= 1024 threads
-  //   pool rendezvous <= ( 1024 / 8 ) * 4 + 4 = 2052
   //   team size       <= 64 threads
-  //   team rendezvous <= ( 64 / 8 ) * 4 + 4 = 36
 
   enum : int { max_pool_members  = 1024 };
   enum : int { max_team_members  = 64 };
-  enum : int { max_pool_rendezvous  = ( max_pool_members / 8 ) * 4 + 4 };
-  enum : int { max_team_rendezvous  = ( max_team_members / 8 ) * 4 + 4 };
+  enum : int { max_pool_rendezvous = rendezvous_buffer_size( max_pool_members ) };
+  enum : int { max_team_rendezvous = rendezvous_buffer_size( max_team_members ) };
 
 private:
 
@@ -114,7 +113,6 @@ private:
   int         m_league_size ;
   int         m_work_chunk ;
   int         m_steal_rank ; // work stealing rank
-  int mutable m_pool_rendezvous_step ;
   int mutable m_team_rendezvous_step ;
 
   HostThreadTeamData * team_member( int r ) const noexcept
@@ -147,6 +145,7 @@ public:
   int team_rendezvous( int const root ) const noexcept
     {
       return 1 == m_team_size ? 1 :
+             HostThreadTeamData::
              rendezvous( m_team_scratch + m_team_rendezvous
                        , m_team_rendezvous_step
                        , m_team_size
@@ -157,6 +156,7 @@ public:
   int team_rendezvous() const noexcept
     {
       return 1 == m_team_size ? 1 :
+             HostThreadTeamData::
              rendezvous( m_team_scratch + m_team_rendezvous
                        , m_team_rendezvous_step
                        , m_team_size
@@ -167,6 +167,7 @@ public:
   void team_rendezvous_release() const noexcept
     {
       if ( 1 < m_team_size ) {
+        HostThreadTeamData::
         rendezvous_release( m_team_scratch + m_team_rendezvous
                           , m_team_rendezvous_step );
       }
@@ -175,19 +176,30 @@ public:
   inline
   int pool_rendezvous() const noexcept
     {
+      static constexpr int yield_wait =
+        #if defined( KOKKOS_COMPILER_IBM )
+            // If running on IBM POWER architecture the global
+            // level rendzvous should immediately yield when
+            // waiting for other threads in the pool to arrive.
+          1
+        #else
+          0
+        #endif
+          ;
       return 1 == m_pool_size ? 1 :
+             Kokkos::Impl::
              rendezvous( m_pool_scratch + m_pool_rendezvous
-                       , m_pool_rendezvous_step
                        , m_pool_size
-                       , m_pool_rank );
+                       , m_pool_rank
+                       , yield_wait );
     }
 
   inline
   void pool_rendezvous_release() const noexcept
     {
       if ( 1 < m_pool_size ) {
-        rendezvous_release( m_pool_scratch + m_pool_rendezvous
-                          , m_pool_rendezvous_step );
+        Kokkos::Impl::
+        rendezvous_release( m_pool_scratch + m_pool_rendezvous );
       }
     }
 
@@ -213,7 +225,6 @@ public:
     , m_league_size(1)
     , m_work_chunk(0)
     , m_steal_rank(0)
-    , m_pool_rendezvous_step(0)
     , m_team_rendezvous_step(0)
     {}
 
@@ -406,7 +417,7 @@ fflush(stdout);
       // Steal from next team, round robin
       // The next team is offset by m_team_alloc if it fits in the pool.
 
-      m_steal_rank = m_team_base + m_team_alloc + m_team_size <= m_pool_size ? 
+      m_steal_rank = m_team_base + m_team_alloc + m_team_size <= m_pool_size ?
                      m_team_base + m_team_alloc : 0 ;
     }
 
diff --git a/lib/kokkos/core/src/impl/Kokkos_Profiling_Interface.cpp b/lib/kokkos/core/src/impl/Kokkos_Profiling_Interface.cpp
index 98482cfab6..608d514c79 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Profiling_Interface.cpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Profiling_Interface.cpp
@@ -50,51 +50,70 @@
 namespace Kokkos {
 namespace Profiling {
 
+static initFunction initProfileLibrary = nullptr;
+static finalizeFunction finalizeProfileLibrary = nullptr;
+
+static beginFunction beginForCallee = nullptr;
+static beginFunction beginScanCallee = nullptr;
+static beginFunction beginReduceCallee = nullptr;
+static endFunction endForCallee = nullptr;
+static endFunction endScanCallee = nullptr;
+static endFunction endReduceCallee = nullptr;
+
+static pushFunction pushRegionCallee = nullptr;
+static popFunction popRegionCallee = nullptr;
+
+static allocateDataFunction allocateDataCallee = nullptr;
+static deallocateDataFunction deallocateDataCallee = nullptr;
+
+static beginDeepCopyFunction beginDeepCopyCallee = nullptr;
+static endDeepCopyFunction endDeepCopyCallee = nullptr;
+
 SpaceHandle::SpaceHandle(const char* space_name) {
   strncpy(name,space_name,64);
 }
 
 bool profileLibraryLoaded() {
-  return (NULL != initProfileLibrary);
+  return (nullptr != initProfileLibrary);
 }
 
 void beginParallelFor(const std::string& kernelPrefix, const uint32_t devID, uint64_t* kernelID) {
-  if(NULL != beginForCallee) {
+  if(nullptr != beginForCallee) {
     Kokkos::fence();
     (*beginForCallee)(kernelPrefix.c_str(), devID, kernelID);
   }
 }
 
 void endParallelFor(const uint64_t kernelID) {
-  if(NULL != endForCallee) {
+  if(nullptr != endForCallee) {
     Kokkos::fence();
     (*endForCallee)(kernelID);
   }
 }
 
 void beginParallelScan(const std::string& kernelPrefix, const uint32_t devID, uint64_t* kernelID) {
-  if(NULL != beginScanCallee) {
+  if(nullptr != beginScanCallee) {
     Kokkos::fence();
     (*beginScanCallee)(kernelPrefix.c_str(), devID, kernelID);
   }
 }
 
 void endParallelScan(const uint64_t kernelID) {
-  if(NULL != endScanCallee) {
+  if(nullptr != endScanCallee) {
     Kokkos::fence();
     (*endScanCallee)(kernelID);
   }
 }
 
 void beginParallelReduce(const std::string& kernelPrefix, const uint32_t devID, uint64_t* kernelID) {
-  if(NULL != beginReduceCallee) {
+  if(nullptr != beginReduceCallee) {
     Kokkos::fence();
     (*beginReduceCallee)(kernelPrefix.c_str(), devID, kernelID);
   }
 }
 
 void endParallelReduce(const uint64_t kernelID) {
-  if(NULL != endReduceCallee) {
+  if(nullptr != endReduceCallee) {
     Kokkos::fence();
     (*endReduceCallee)(kernelID);
   }
@@ -102,31 +121,47 @@ void endParallelReduce(const uint64_t kernelID) {
 
 
 void pushRegion(const std::string& kName) {
-  if( NULL != pushRegionCallee ) {
+  if( nullptr != pushRegionCallee ) {
     Kokkos::fence();
     (*pushRegionCallee)(kName.c_str());
   }
 }
 
 void popRegion() {
-  if( NULL != popRegionCallee ) {
+  if( nullptr != popRegionCallee ) {
     Kokkos::fence();
     (*popRegionCallee)();
   }
 }
 
 void allocateData(const SpaceHandle space, const std::string label, const void* ptr, const uint64_t size) {
-  if(NULL != allocateDataCallee) {
+  if(nullptr != allocateDataCallee) {
     (*allocateDataCallee)(space,label.c_str(),ptr,size);
   }
 }
 
 void deallocateData(const SpaceHandle space, const std::string label, const void* ptr, const uint64_t size) {
-  if(NULL != allocateDataCallee) {
+  if(nullptr != deallocateDataCallee) {
     (*deallocateDataCallee)(space,label.c_str(),ptr,size);
   }
 }
 
+void beginDeepCopy(const SpaceHandle dst_space, const std::string dst_label, const void* dst_ptr,
+    const SpaceHandle src_space, const std::string src_label, const void* src_ptr,
+    const uint64_t size) {
+  if(nullptr != beginDeepCopyCallee) {
+    (*beginDeepCopyCallee)(dst_space, dst_label.c_str(), dst_ptr,
+                      src_space, src_label.c_str(), src_ptr,
+                      size);
+  }
+}
+
+void endDeepCopy() {
+  if(nullptr != endDeepCopyCallee) {
+    (*endDeepCopyCallee)();
+  }
+}
+
 void initialize() {
 
   // Make sure initialize calls happens only once
@@ -140,7 +175,7 @@ void initialize() {
 
   // If we do not find a profiling library in the environment then exit
   // early.
-  if( NULL == envProfileLibrary ) {
+  if( nullptr == envProfileLibrary ) {
     return ;
   }
 
@@ -149,10 +184,10 @@ void initialize() {
 
   char* profileLibraryName = strtok(envProfileCopy, ";");
 
-  if( (NULL != profileLibraryName) && (strcmp(profileLibraryName, "") != 0) ) {
+  if( (nullptr != profileLibraryName) && (strcmp(profileLibraryName, "") != 0) ) {
     firstProfileLibrary = dlopen(profileLibraryName, RTLD_NOW | RTLD_GLOBAL);
 
-    if(NULL == firstProfileLibrary) {
+    if(nullptr == firstProfileLibrary) {
       std::cerr << "Error: Unable to load KokkosP library: " <<
         profileLibraryName << std::endl;
     } else {
@@ -191,14 +226,19 @@ void initialize() {
       auto p12 = dlsym(firstProfileLibrary, "kokkosp_deallocate_data");
       deallocateDataCallee = *((deallocateDataFunction*) &p12);
 
+      auto p13 = dlsym(firstProfileLibrary, "kokkosp_begin_deep_copy");
+      beginDeepCopyCallee = *((beginDeepCopyFunction*) &p13);
+      auto p14 = dlsym(firstProfileLibrary, "kokkosp_end_deep_copy");
+      endDeepCopyCallee = *((endDeepCopyFunction*) &p14);
+
     }
   }
 
-  if(NULL != initProfileLibrary) {
+  if(nullptr != initProfileLibrary) {
     (*initProfileLibrary)(0,
         (uint64_t) KOKKOSP_INTERFACE_VERSION,
         (uint32_t) 0,
-        NULL);
+        nullptr);
   }
 
   free(envProfileCopy);
@@ -210,28 +250,30 @@ void finalize() {
   if(is_finalized) return;
   is_finalized = 1;
 
-  if(NULL != finalizeProfileLibrary) {
+  if(nullptr != finalizeProfileLibrary) {
     (*finalizeProfileLibrary)();
 
-    // Set all profile hooks to NULL to prevent
+    // Set all profile hooks to nullptr to prevent
     // any additional calls. Once we are told to
     // finalize, we mean it
-    initProfileLibrary = NULL;
-    finalizeProfileLibrary = NULL;
+    initProfileLibrary = nullptr;
+    finalizeProfileLibrary = nullptr;
 
-    beginForCallee = NULL;
-    beginScanCallee = NULL;
-    beginReduceCallee = NULL;
-    endScanCallee = NULL;
-    endForCallee = NULL;
-    endReduceCallee = NULL;
+    beginForCallee = nullptr;
+    beginScanCallee = nullptr;
+    beginReduceCallee = nullptr;
+    endScanCallee = nullptr;
+    endForCallee = nullptr;
+    endReduceCallee = nullptr;
 
-    pushRegionCallee = NULL;
-    popRegionCallee = NULL;
+    pushRegionCallee = nullptr;
+    popRegionCallee = nullptr;
 
-    allocateDataCallee = NULL;
-    deallocateDataCallee = NULL;
+    allocateDataCallee = nullptr;
+    deallocateDataCallee = nullptr;
 
+    beginDeepCopyCallee = nullptr;
+    endDeepCopyCallee = nullptr;
   }
 }
 }
diff --git a/lib/kokkos/core/src/impl/Kokkos_Profiling_Interface.hpp b/lib/kokkos/core/src/impl/Kokkos_Profiling_Interface.hpp
index f76e5dfa04..2c2e524d9d 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Profiling_Interface.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Profiling_Interface.hpp
@@ -81,23 +81,11 @@ typedef void (*popFunction)();
 typedef void (*allocateDataFunction)(const SpaceHandle, const char*, const void*, const uint64_t);
 typedef void (*deallocateDataFunction)(const SpaceHandle, const char*, const void*, const uint64_t);
 
-
-static initFunction initProfileLibrary = NULL;
-static finalizeFunction finalizeProfileLibrary = NULL;
-
-static beginFunction beginForCallee = NULL;
-static beginFunction beginScanCallee = NULL;
-static beginFunction beginReduceCallee = NULL;
-static endFunction endForCallee = NULL;
-static endFunction endScanCallee = NULL;
-static endFunction endReduceCallee = NULL;
-
-static pushFunction pushRegionCallee = NULL;
-static popFunction popRegionCallee = NULL;
-
-static allocateDataFunction allocateDataCallee = NULL;
-static deallocateDataFunction deallocateDataCallee = NULL;
-
+typedef void (*beginDeepCopyFunction)(
+    SpaceHandle, const char*, const void*,
+    SpaceHandle, const char*, const void*,
+    uint64_t);
+typedef void (*endDeepCopyFunction)();
 
 bool profileLibraryLoaded();
 
@@ -114,35 +102,14 @@ void popRegion();
 void allocateData(const SpaceHandle space, const std::string label, const void* ptr, const uint64_t size);
 void deallocateData(const SpaceHandle space, const std::string label, const void* ptr, const uint64_t size);
 
+void beginDeepCopy(const SpaceHandle dst_space, const std::string dst_label, const void* dst_ptr,
+    const SpaceHandle src_space, const std::string src_label, const void* src_ptr,
+    const uint64_t size);
+void endDeepCopy();
+
 void initialize();
 void finalize();
 
-//Define finalize_fake inline to get rid of warnings for unused static variables
-inline void finalize_fake() {
-  if(NULL != finalizeProfileLibrary) {
-    (*finalizeProfileLibrary)();
-
-    // Set all profile hooks to NULL to prevent
-    // any additional calls. Once we are told to
-    // finalize, we mean it
-    beginForCallee = NULL;
-    beginScanCallee = NULL;
-    beginReduceCallee = NULL;
-    endScanCallee = NULL;
-    endForCallee = NULL;
-    endReduceCallee = NULL;
-
-    allocateDataCallee = NULL;
-    deallocateDataCallee = NULL;
-
-    initProfileLibrary = NULL;
-    finalizeProfileLibrary = NULL;
-    pushRegionCallee = NULL;
-    popRegionCallee = NULL;
-  }
-}
-
-
 }
 }
 
diff --git a/lib/kokkos/core/src/impl/Kokkos_Rendezvous.cpp b/lib/kokkos/core/src/impl/Kokkos_Rendezvous.cpp
new file mode 100644
index 0000000000..ac697fce4b
--- /dev/null
+++ b/lib/kokkos/core/src/impl/Kokkos_Rendezvous.cpp
@@ -0,0 +1,208 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include 
+#include 
+#include 
+#include 
+
+namespace Kokkos { namespace Impl {
+
+//----------------------------------------------------------------------------
+/* pattern for rendezvous
+ *
+ *  if ( rendezvous() ) {
+ *     ... all other threads are still in team_rendezvous() ...
+ *     rendezvous_release();
+ *     ... all other threads are released from team_rendezvous() ...
+ *  }
+ */
+
+int rendezvous( volatile int64_t * const buffer
+              , int const size
+              , int const rank
+              , int const slow
+              ) noexcept
+{
+  enum : int { shift_byte = 3 };
+  enum : int { size_byte  = ( 01 << shift_byte ) }; // == 8
+  enum : int { mask_byte  = size_byte - 1 };
+
+  enum : int { shift_mem_cycle = 2 };
+  enum : int { size_mem_cycle  = ( 01 << shift_mem_cycle ) }; // == 4
+  enum : int { mask_mem_cycle  = size_mem_cycle - 1 };
+
+  // Cycle step values: 1 <= step <= size_val_cycle
+  // An odd multiple of memory cycle so that when a memory location
+  // is reused it has a different value.
+  // Must be representable within a single byte: size_val_cycle < 16
+
+  enum : int { size_val_cycle = 3 * size_mem_cycle };
+
+  // Requires:
+  //   Called by rank = [ 0 .. size )
+  //   buffer aligned to int64_t[4]
+
+  // A sequence of rendezvous uses four cycled locations in memory
+  // and non-equal cycled synchronization values to
+  // 1) prevent rendezvous from overtaking one another and
+  // 2) give each spin wait location an int64_t[4] span
+  //    so that it has its own cache line.
+
+  const int64_t step = (buffer[0] % size_val_cycle ) + 1 ;
+
+  // The leading int64_t[4] span is for thread 0 to write
+  // and all other threads to read spin-wait.
+  // sync_offset is the index into this array for this step.
+
+  const int sync_offset = ( step & mask_mem_cycle ) + size_mem_cycle + size_mem_cycle ;
+
+  if ( rank ) {
+
+    const int group_begin = rank << shift_byte ; // == rank * size_byte
+
+    if ( group_begin < size ) {
+
+      //  This thread waits for threads
+      //   [ group_begin .. group_begin + 8 )
+      //   [ rank*8      .. rank*8 + 8      )
+      // to write to their designated bytes.
+
+      const int end = group_begin + size_byte < size
+                    ? size_byte : size - group_begin ;
+
+      int64_t value = 0;
+      for ( int i = 0 ; i < end ; ++i ) {
+        value |= step << (i * size_byte );
+      }
+
+      store_fence(); // This should not be needed but fixes #742
+
+      if ( slow ) {
+        yield_until_equal( buffer[ (rank << shift_mem_cycle) + sync_offset ]
+                          , value );
+      }
+      else {
+        spinwait_until_equal( buffer[ (rank << shift_mem_cycle) + sync_offset ]
+                            , value );
+      }
+    }
+
+    {
+      // This thread sets its designated byte.
+      //   ( rank % size_byte ) +
+      //   ( ( rank / size_byte ) * size_byte * size_mem_cycle ) +
+      //   ( sync_offset * size_byte )
+      const int offset = ( rank & mask_byte )
+                       + ( ( rank & ~mask_byte ) << shift_mem_cycle )
+                       + ( sync_offset << shift_byte );
+
+      // All of this thread's previous memory stores must be complete before
+      // this thread stores the step value at this thread's designated byte
+      // in the shared synchronization array.
+
+      Kokkos::memory_fence();
+
+      ((volatile int8_t*) buffer)[ offset ] = int8_t( step );
+
+      // Memory fence to push the previous store out
+      Kokkos::memory_fence();
+    }
+
+    // Wait for thread 0 to release all other threads
+
+    if ( slow ) {
+      yield_until_equal( buffer[ (step & mask_mem_cycle) + size_mem_cycle ] , int64_t(step) );
+    }
+    else {
+      spinwait_until_equal( buffer[ (step & mask_mem_cycle) + size_mem_cycle ] , int64_t(step) );
+    }
+  }
+  else {
+    // Thread 0 waits for threads [1..7]
+    // to write to their designated bytes.
+
+    const int end = size_byte < size ? 8 : size ;
+
+    int64_t value = 0;
+    for ( int i = 1 ; i < end ; ++i ) {
+      value |= step << (i * size_byte );
+    }
+
+    if ( slow ) {
+      yield_until_equal( buffer[ sync_offset ], value );
+    }
+    else {
+      spinwait_until_equal( buffer[ sync_offset ], value );
+    }
+  }
+
+  return rank ? 0 : 1 ;
+}
+
+void rendezvous_release( volatile int64_t * const buffer ) noexcept
+{
+  enum : int { shift_mem_cycle = 2 };
+  enum : int { size_mem_cycle  = ( 01 << shift_mem_cycle ) }; // == 4
+  enum : int { mask_mem_cycle  = size_mem_cycle - 1 };
+  enum : int { size_val_cycle = 3 * size_mem_cycle };
+
+  // Requires:
+  //   Called after team_rendezvous
+  //   Called only by true == team_rendezvous(root)
+
+  // update step
+  const int64_t step = (buffer[0] % size_val_cycle ) + 1;
+  buffer[0] = step;
+
+  // Memory fence to be sure all previous writes are complete:
+  Kokkos::memory_fence();
+
+  buffer[ (step & mask_mem_cycle) + size_mem_cycle ] = step;
+
+  // Memory fence to push the store out
+  Kokkos::memory_fence();
+}
+
+}} // namespace Kokkos::Impl
+
diff --git a/lib/kokkos/core/src/impl/Kokkos_Rendezvous.hpp b/lib/kokkos/core/src/impl/Kokkos_Rendezvous.hpp
new file mode 100644
index 0000000000..57f8633bca
--- /dev/null
+++ b/lib/kokkos/core/src/impl/Kokkos_Rendezvous.hpp
@@ -0,0 +1,87 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_IMPL_RENDEZVOUS_HPP
+#define KOKKOS_IMPL_RENDEZVOUS_HPP
+
+#include 
+
+namespace Kokkos { namespace Impl {
+
+inline
+constexpr int rendezvous_buffer_size( int max_members ) noexcept
+{
+  return (((max_members + 7) / 8) * 4) + 4 + 4;
+}
+
+/** \brief  Thread pool rendezvous
+ *
+ *  Rendezvous pattern:
+ *   if ( rendezvous(root) ) {
+ *     ... only root thread here while all others wait ...
+ *     rendezvous_release();
+ *   }
+ *   else {
+ *     ... all other threads release here ...
+ *   }
+ *
+ *  Requires: buffer[ rendezvous_buffer_size( max_threads ) ];
+ *
+ *  When slow != 0 the expectation is thread arrival will be 
+ *  slow so the threads that arrive early should quickly yield
+ *  their core to the runtime thus possibly allowing the late
+ *  arriving threads to have more resources
+ *  (e.g., power and clock frequency).
+ */
+int rendezvous( volatile int64_t * const buffer
+              , int const size
+              , int const rank
+              , int const slow = 0 ) noexcept ;
+
+void rendezvous_release( volatile int64_t * const buffer ) noexcept ;
+
+
+}} // namespace Kokkos::Impl
+
+#endif // KOKKOS_IMPL_RENDEZVOUS_HPP
+
diff --git a/lib/kokkos/core/src/impl/Kokkos_Serial.cpp b/lib/kokkos/core/src/impl/Kokkos_Serial.cpp
index 755271c07e..dfbeba461e 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Serial.cpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Serial.cpp
@@ -50,6 +50,7 @@
 #include 
 #include 
 
+#include 
 
 /*--------------------------------------------------------------------------*/
 
@@ -123,7 +124,6 @@ void serial_resize_thread_team_data( size_t pool_reduce_bytes
   }
 }
 
-// Get thread team data structure for omp_get_thread_num()
 HostThreadTeamData * serial_get_thread_team_data()
 {
   return & g_serial_thread_team_data ;
@@ -151,6 +151,8 @@ void Serial::initialize( unsigned threads_count
   (void) use_cores_per_numa;
   (void) allow_asynchronous_threadpool;
 
+  Impl::SharedAllocationRecord< void, void >::tracking_enable();
+
   // Init the array of locks used for arbitrarily sized atomics
   Impl::init_lock_array_host_space();
   #if defined(KOKKOS_ENABLE_PROFILING)
diff --git a/lib/kokkos/core/src/impl/Kokkos_Serial_Task.cpp b/lib/kokkos/core/src/impl/Kokkos_Serial_Task.cpp
index 76297161b1..0b6fbd9af0 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Serial_Task.cpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Serial_Task.cpp
@@ -62,7 +62,7 @@ void TaskQueueSpecialization< Kokkos::Serial >::execute
 {
   using execution_space = Kokkos::Serial ;
   using queue_type      = TaskQueue< execution_space > ;
-  using task_root_type  = TaskBase< execution_space , void , void > ;
+  using task_root_type  = TaskBase< void , void , void > ;
   using Member          = Impl::HostThreadTeamMember< execution_space > ;
 
   task_root_type * const end = (task_root_type *) task_root_type::EndTag ;
@@ -122,7 +122,7 @@ void TaskQueueSpecialization< Kokkos::Serial > ::
 {
   using execution_space = Kokkos::Serial ;
   using queue_type      = TaskQueue< execution_space > ;
-  using task_root_type  = TaskBase< execution_space , void , void > ;
+  using task_root_type  = TaskBase< void , void , void > ;
   using Member          = Impl::HostThreadTeamMember< execution_space > ;
 
   task_root_type * const end = (task_root_type *) task_root_type::EndTag ;
diff --git a/lib/kokkos/core/src/impl/Kokkos_Serial_Task.hpp b/lib/kokkos/core/src/impl/Kokkos_Serial_Task.hpp
index 2eb2b5cf52..39deebbbf1 100644
--- a/lib/kokkos/core/src/impl/Kokkos_Serial_Task.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Serial_Task.hpp
@@ -65,7 +65,7 @@ public:
   using execution_space = Kokkos::Serial ;
   using memory_space    = Kokkos::HostSpace ;
   using queue_type      = Kokkos::Impl::TaskQueue< execution_space > ;
-  using task_base_type  = Kokkos::Impl::TaskBase< execution_space , void , void > ;
+  using task_base_type  = Kokkos::Impl::TaskBase< void , void , void > ;
   using member_type     = Kokkos::Impl::HostThreadTeamMember< execution_space > ;
 
   static
diff --git a/lib/kokkos/core/src/impl/Kokkos_Serial_WorkGraphPolicy.hpp b/lib/kokkos/core/src/impl/Kokkos_Serial_WorkGraphPolicy.hpp
new file mode 100644
index 0000000000..dc30ffe9e0
--- /dev/null
+++ b/lib/kokkos/core/src/impl/Kokkos_Serial_WorkGraphPolicy.hpp
@@ -0,0 +1,102 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#ifndef KOKKOS_SERIAL_WORKGRAPHPOLICY_HPP
+#define KOKKOS_SERIAL_WORKGRAPHPOLICY_HPP
+
+namespace Kokkos {
+namespace Impl {
+
+template< class FunctorType , class ... Traits >
+class ParallelFor< FunctorType ,
+                   Kokkos::Experimental::WorkGraphPolicy< Traits ... > ,
+                   Kokkos::Serial
+                 >
+  : public Kokkos::Impl::Experimental::
+           WorkGraphExec< FunctorType,
+                          Kokkos::Serial,
+                          Traits ...
+                        >
+{
+private:
+
+  typedef Kokkos::Experimental::WorkGraphPolicy< Traits ... > Policy ;
+  typedef Kokkos::Impl::Experimental::
+          WorkGraphExec Base ;
+
+  template< class TagType >
+  typename std::enable_if< std::is_same< TagType , void >::value >::type
+  exec_one(const typename Policy::member_type& i) const {
+    Base::m_functor( i );
+  }
+
+  template< class TagType >
+  typename std::enable_if< ! std::is_same< TagType , void >::value >::type
+  exec_one(const typename Policy::member_type& i) const {
+    const TagType t{} ;
+    Base::m_functor( t , i );
+  }
+
+public:
+
+  inline
+  void execute()
+  {
+    for (std::int32_t i; (-1 != (i = Base::before_work())); ) {
+      exec_one< typename Policy::work_tag >( i );
+      Base::after_work(i);
+    }
+  }
+
+  inline
+  ParallelFor( const FunctorType & arg_functor
+             , const Policy      & arg_policy )
+    : Base( arg_functor, arg_policy )
+  {
+  }
+};
+
+} // namespace Impl
+} // namespace Kokkos
+
+#endif /* #define KOKKOS_SERIAL_WORKGRAPHPOLICY_HPP */
diff --git a/lib/kokkos/core/src/impl/Kokkos_SharedAlloc.cpp b/lib/kokkos/core/src/impl/Kokkos_SharedAlloc.cpp
index e28c1194a7..af79523e0c 100644
--- a/lib/kokkos/core/src/impl/Kokkos_SharedAlloc.cpp
+++ b/lib/kokkos/core/src/impl/Kokkos_SharedAlloc.cpp
@@ -46,23 +46,23 @@
 namespace Kokkos {
 namespace Impl {
 
-int SharedAllocationRecord< void , void >::s_tracking_enabled = 1 ;
+namespace {
 
-void SharedAllocationRecord< void , void >::tracking_claim_and_disable()
-{
-  // A host thread claim and disable tracking flag
+__thread int t_tracking_enabled = 1;
 
-  while ( ! Kokkos::atomic_compare_exchange_strong( & s_tracking_enabled, 1, 0 ) );
 }
 
-void SharedAllocationRecord< void , void >::tracking_release_and_enable()
-{
-  // The host thread that claimed and disabled the tracking flag
-  // now release and enable tracking.
+int SharedAllocationRecord< void , void >::tracking_enabled()
+{ return t_tracking_enabled; }
 
-  if ( ! Kokkos::atomic_compare_exchange_strong( & s_tracking_enabled, 0, 1 ) ){
-    Kokkos::Impl::throw_runtime_exception("Kokkos::Impl::SharedAllocationRecord<>::tracking_release_and_enable FAILED, this host process thread did not hold the lock" );
-  }
+void SharedAllocationRecord< void , void >::tracking_disable()
+{
+  t_tracking_enabled = 0;
+}
+
+void SharedAllocationRecord< void , void >::tracking_enable()
+{
+  t_tracking_enabled = 1;
 }
 
 //----------------------------------------------------------------------------
diff --git a/lib/kokkos/core/src/impl/Kokkos_SharedAlloc.hpp b/lib/kokkos/core/src/impl/Kokkos_SharedAlloc.hpp
index 4dc61bb02e..2e3cc1a163 100644
--- a/lib/kokkos/core/src/impl/Kokkos_SharedAlloc.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_SharedAlloc.hpp
@@ -71,6 +71,9 @@ public:
   KOKKOS_INLINE_FUNCTION static
   const SharedAllocationHeader * get_header( void * alloc_ptr )
     { return reinterpret_cast( reinterpret_cast(alloc_ptr) - sizeof(SharedAllocationHeader) ); }
+
+  KOKKOS_INLINE_FUNCTION
+  const char* label() const { return m_label; }
 };
 
 template<>
@@ -83,8 +86,6 @@ protected:
 
   typedef void (* function_type )( SharedAllocationRecord * );
 
-  static int s_tracking_enabled ;
-
   SharedAllocationHeader * const m_alloc_ptr ;
   size_t                   const m_alloc_size ;
   function_type            const m_dealloc ;
@@ -110,17 +111,17 @@ protected:
 public:
   inline std::string get_label() const { return std::string("Unmanaged"); }
 
-  static int tracking_enabled() { return s_tracking_enabled ; }
+  static int tracking_enabled();
 
   /**\brief A host process thread claims and disables the
    *        shared allocation tracking flag.
    */
-  static void tracking_claim_and_disable();
+  static void tracking_disable();
 
   /**\brief A host process thread releases and enables the
    *        shared allocation tracking flag.
    */
-  static void tracking_release_and_enable();
+  static void tracking_enable();
 
   ~SharedAllocationRecord() = default ;
 
@@ -317,6 +318,11 @@ public:
 #endif
     }
 
+  KOKKOS_INLINE_FUNCTION
+  bool has_record() const {
+    return (m_record_bits & (~DO_NOT_DEREF_FLAG)) != 0;
+  }
+
   KOKKOS_FORCEINLINE_FUNCTION
   ~SharedAllocationTracker()
     { KOKKOS_IMPL_SHARED_ALLOCATION_TRACKER_DECREMENT }
diff --git a/lib/kokkos/core/src/impl/Kokkos_Spinwait.cpp b/lib/kokkos/core/src/impl/Kokkos_Spinwait.cpp
new file mode 100644
index 0000000000..3d3f83ed85
--- /dev/null
+++ b/lib/kokkos/core/src/impl/Kokkos_Spinwait.cpp
@@ -0,0 +1,210 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include 
+#if defined( KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_HOST )
+
+#include 
+#include 
+#include 
+
+#if defined( KOKKOS_ENABLE_STDTHREAD )
+  #include 
+#elif !defined( _WIN32 )
+  #include 
+  #include 
+#else
+  #include 
+  #include 
+  #include 
+#endif
+
+/*--------------------------------------------------------------------------*/
+
+namespace Kokkos {
+namespace Impl {
+namespace {
+
+void host_thread_yield( const uint32_t i , const int force_yield )
+{
+  static constexpr uint32_t sleep_limit = 1 << 13 ;
+  static constexpr uint32_t yield_limit = 1 << 12 ;
+
+  const int c = Kokkos::Impl::bit_scan_reverse(i);
+
+  if ( sleep_limit < i ) {
+
+    // Attempt to put the thread to sleep for 'c' milliseconds
+
+    #if defined( KOKKOS_ENABLE_STDTHREAD )
+      std::this_thread::sleep_for( std::chrono::nanoseconds( c * 1000 ) )
+    #elif !defined( _WIN32 )
+      timespec req ;
+      req.tv_sec  = 0 ;
+      req.tv_nsec = 1000 * c ;
+      nanosleep( &req, nullptr );
+    #else /* defined( _WIN32 ) IS Microsoft Windows */
+      Sleep(c);
+    #endif
+  }
+
+  else if ( force_yield || yield_limit < i ) {
+
+    // Attempt to yield thread resources to runtime
+
+    #if defined( KOKKOS_ENABLE_STDTHREAD )
+      std::this_thread::yield();
+    #elif !defined( _WIN32 )
+      sched_yield();
+    #else /* defined( _WIN32 ) IS Microsoft Windows */
+      YieldProcessor();
+    #endif
+  }
+
+  #if defined( KOKKOS_ENABLE_ASM )
+
+  else if ( (1u<<4) < i ) {
+
+    // Insert a few no-ops to quiet the thread:
+
+    for ( int k = 0 ; k < c ; ++k ) {
+      #if defined( __amd64 ) || defined( __amd64__ ) || \
+       	    defined( __x86_64 ) || defined( __x86_64__ )
+    	#if !defined( _WIN32 ) /* IS NOT Microsoft Windows */
+          asm volatile( "nop\n" );
+	#else
+          __asm__ __volatile__( "nop\n" );
+        #endif
+      #elif defined(__PPC64__)
+          asm volatile( "nop\n" );
+      #endif
+    }
+  }
+
+  {
+    // Insert memory pause
+      #if defined( __amd64 ) || defined( __amd64__ ) || \
+       	    defined( __x86_64 ) || defined( __x86_64__ )
+    	#if !defined( _WIN32 ) /* IS NOT Microsoft Windows */
+          asm volatile( "pause\n":::"memory" );
+	#else
+          __asm__ __volatile__( "pause\n":::"memory" );
+        #endif
+      #elif defined(__PPC64__)
+	asm volatile( "or 27, 27, 27" ::: "memory" );
+      #endif
+  }
+
+  #endif /* defined( KOKKOS_ENABLE_ASM ) */
+}
+
+}}} // namespace Kokkos::Impl::{anonymous}
+
+/*--------------------------------------------------------------------------*/
+
+namespace Kokkos {
+namespace Impl {
+
+void spinwait_while_equal( volatile int32_t & flag , const int32_t value )
+{
+  Kokkos::store_fence();
+  uint32_t i = 0 ; while( value == flag ) host_thread_yield(++i,0);
+  Kokkos::load_fence();
+}
+
+void spinwait_until_equal( volatile int32_t & flag , const int32_t value )
+{
+  Kokkos::store_fence();
+  uint32_t i = 0 ; while( value != flag ) host_thread_yield(++i,0);
+  Kokkos::load_fence();
+}
+
+void spinwait_while_equal( volatile int64_t & flag , const int64_t value )
+{
+  Kokkos::store_fence();
+  uint32_t i = 0 ; while( value == flag ) host_thread_yield(++i,0);
+  Kokkos::load_fence();
+}
+
+void spinwait_until_equal( volatile int64_t & flag , const int64_t value )
+{
+  Kokkos::store_fence();
+  uint32_t i = 0 ; while( value != flag ) host_thread_yield(++i,0);
+  Kokkos::load_fence();
+}
+
+void yield_while_equal( volatile int32_t & flag , const int32_t value )
+{
+  Kokkos::store_fence();
+  uint32_t i = 0 ; while( value == flag ) host_thread_yield(++i,1);
+  Kokkos::load_fence();
+}
+
+void yield_until_equal( volatile int32_t & flag , const int32_t value )
+{
+  Kokkos::store_fence();
+  uint32_t i = 0 ; while( value != flag ) host_thread_yield(++i,1);
+  Kokkos::load_fence();
+}
+
+void yield_while_equal( volatile int64_t & flag , const int64_t value )
+{
+  Kokkos::store_fence();
+  uint32_t i = 0 ; while( value == flag ) host_thread_yield(++i,1);
+  Kokkos::load_fence();
+}
+
+void yield_until_equal( volatile int64_t & flag , const int64_t value )
+{
+  Kokkos::store_fence();
+  uint32_t i = 0 ; while( value != flag ) host_thread_yield(++i,1);
+  Kokkos::load_fence();
+}
+
+} /* namespace Impl */
+} /* namespace Kokkos */
+
+#else
+void KOKKOS_CORE_SRC_IMPL_SPINWAIT_PREVENT_LINK_ERROR() {}
+#endif
+
diff --git a/lib/kokkos/core/src/impl/Kokkos_spinwait.hpp b/lib/kokkos/core/src/impl/Kokkos_Spinwait.hpp
similarity index 82%
rename from lib/kokkos/core/src/impl/Kokkos_spinwait.hpp
rename to lib/kokkos/core/src/impl/Kokkos_Spinwait.hpp
index 6e34b8a943..b49e308566 100644
--- a/lib/kokkos/core/src/impl/Kokkos_spinwait.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_Spinwait.hpp
@@ -59,6 +59,13 @@ void spinwait_until_equal( volatile int32_t & flag , const int32_t value );
 
 void spinwait_while_equal( volatile int64_t & flag , const int64_t value );
 void spinwait_until_equal( volatile int64_t & flag , const int64_t value );
+
+void yield_while_equal( volatile int32_t & flag , const int32_t value );
+void yield_until_equal( volatile int32_t & flag , const int32_t value );
+
+void yield_while_equal( volatile int64_t & flag , const int64_t value );
+void yield_until_equal( volatile int64_t & flag , const int64_t value );
+
 #else
 
 KOKKOS_INLINE_FUNCTION
@@ -71,6 +78,16 @@ void spinwait_while_equal( volatile int64_t & , const int64_t ) {}
 KOKKOS_INLINE_FUNCTION
 void spinwait_until_equal( volatile int64_t & , const int64_t ) {}
 
+KOKKOS_INLINE_FUNCTION
+void yield_while_equal( volatile int32_t & , const int32_t ) {}
+KOKKOS_INLINE_FUNCTION
+void yield_until_equal( volatile int32_t & , const int32_t ) {}
+
+KOKKOS_INLINE_FUNCTION
+void yield_while_equal( volatile int64_t & , const int64_t ) {}
+KOKKOS_INLINE_FUNCTION
+void yield_until_equal( volatile int64_t & , const int64_t ) {}
+
 #endif
 
 } /* namespace Impl */
diff --git a/lib/kokkos/core/src/impl/Kokkos_TaskQueue.hpp b/lib/kokkos/core/src/impl/Kokkos_TaskQueue.hpp
index bee98e6745..5f8699302d 100644
--- a/lib/kokkos/core/src/impl/Kokkos_TaskQueue.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_TaskQueue.hpp
@@ -59,24 +59,15 @@
 namespace Kokkos {
 namespace Impl {
 
-/*\brief  Implementation data for task data management, access, and execution.
- *
- *  Curiously recurring template pattern (CRTP)
- *  to allow static_cast from the
- *  task root type and a task's FunctorType.
- *
- *    TaskBase< Space , ResultType , FunctorType >
- *      : TaskBase< Space , ResultType , void >
- *      , FunctorType
- *      { ... };
- *
- *    TaskBase< Space , ResultType , void >
- *      : TaskBase< Space , void , void >
- *      { ... };
- */
-template< typename Space , typename ResultType , typename FunctorType >
+template< class Space , typename ResultType , class FunctorType >
 class TaskBase ;
 
+template< typename Space >
+class TaskQueue ;
+
+template< typename Space >
+class TaskQueueSpecialization ;
+
 } /* namespace Impl */
 } /* namespace Kokkos */
 
@@ -86,8 +77,217 @@ class TaskBase ;
 namespace Kokkos {
 namespace Impl {
 
-template< typename Space >
-class TaskQueueSpecialization ;
+/** \brief  Base class for task management, access, and execution.
+ *
+ *  Inheritance structure to allow static_cast from the task root type
+ *  and a task's FunctorType.
+ *
+ *    // Enable a functor to access the base class
+ *    // and provide memory for result value.
+ *    TaskBase< Space , ResultType , FunctorType >
+ *      : TaskBase< void , void , void >
+ *      , FunctorType
+ *      { ... };
+ *    Followed by memory allocated for result value.
+ *
+ *
+ *  States of a task:
+ *
+ *    Constructing State, NOT IN a linked list
+ *      m_wait == 0
+ *      m_next == 0
+ *
+ *    Scheduling transition : Constructing -> Waiting
+ *      before:
+ *        m_wait == 0
+ *        m_next == this task's initial dependence, 0 if none
+ *      after:
+ *        m_wait == EndTag
+ *        m_next == EndTag
+ *
+ *    Waiting State, IN a linked list
+ *      m_apply != 0
+ *      m_queue != 0
+ *      m_ref_count > 0
+ *      m_wait == head of linked list of tasks waiting on this task
+ *      m_next == next of linked list of tasks
+ *
+ *    transition : Waiting -> Executing
+ *      before:
+ *        m_next == EndTag
+ *      after::
+ *        m_next == LockTag
+ *
+ *    Executing State, NOT IN a linked list
+ *      m_apply != 0
+ *      m_queue != 0
+ *      m_ref_count > 0
+ *      m_wait == head of linked list of tasks waiting on this task
+ *      m_next == LockTag
+ *
+ *    Respawn transition : Executing -> Executing-Respawn
+ *      before:
+ *        m_next == LockTag
+ *      after:
+ *        m_next == this task's updated dependence, 0 if none
+ *
+ *    Executing-Respawn State, NOT IN a linked list
+ *      m_apply != 0
+ *      m_queue != 0
+ *      m_ref_count > 0
+ *      m_wait == head of linked list of tasks waiting on this task
+ *      m_next == this task's updated dependence, 0 if none
+ *
+ *    transition : Executing -> Complete
+ *      before:
+ *        m_wait == head of linked list
+ *      after:
+ *        m_wait == LockTag
+ *
+ *    Complete State, NOT IN a linked list
+ *      m_wait == LockTag: cannot add dependence (<=> complete)
+ *      m_next == LockTag: not a member of a wait queue
+ *
+ */
+template<>
+class TaskBase< void , void , void >
+{
+public:
+
+  enum : int16_t   { TaskTeam = 0 , TaskSingle = 1 , Aggregate = 2 };
+  enum : uintptr_t { LockTag = ~uintptr_t(0) , EndTag = ~uintptr_t(1) };
+
+  template< typename > friend class Kokkos::TaskScheduler ;
+
+  typedef TaskQueue< void > queue_type ;
+
+  typedef void (* function_type) ( TaskBase * , void * );
+
+  // sizeof(TaskBase) == 48
+
+  function_type  m_apply ;       ///< Apply function pointer
+  queue_type   * m_queue ;       ///< Pointer to queue
+  TaskBase     * m_wait ;        ///< Linked list of tasks waiting on this
+  TaskBase     * m_next ;        ///< Waiting linked-list next
+  int32_t        m_ref_count ;   ///< Reference count
+  int32_t        m_alloc_size ;  ///< Allocation size
+  int32_t        m_dep_count ;   ///< Aggregate's number of dependences
+  int16_t        m_task_type ;   ///< Type of task
+  int16_t        m_priority ;    ///< Priority of runnable task
+
+  TaskBase( TaskBase && ) = delete ;
+  TaskBase( const TaskBase & ) = delete ;
+  TaskBase & operator = ( TaskBase && ) = delete ;
+  TaskBase & operator = ( const TaskBase & ) = delete ;
+
+  KOKKOS_INLINE_FUNCTION ~TaskBase() = default ;
+
+  KOKKOS_INLINE_FUNCTION constexpr
+  TaskBase()
+    : m_apply(      0 )
+    , m_queue(      0 )
+    , m_wait(       0 )
+    , m_next(       0 )
+    , m_ref_count(  0 )
+    , m_alloc_size( 0 )
+    , m_dep_count(  0 )
+    , m_task_type(  0 )
+    , m_priority(   0 )
+    {}
+
+  //----------------------------------------
+
+  KOKKOS_INLINE_FUNCTION
+  TaskBase * volatile * aggregate_dependences() volatile
+    { return reinterpret_cast( this + 1 ); }
+
+  KOKKOS_INLINE_FUNCTION
+  bool requested_respawn()
+    {
+      // This should only be called when a task has finished executing and is
+      // in the transition to either the complete or executing-respawn state.
+      TaskBase * const lock = reinterpret_cast< TaskBase * >( LockTag );
+      return lock != m_next;
+    }
+
+  KOKKOS_INLINE_FUNCTION
+  void add_dependence( TaskBase* dep )
+    {
+      // Precondition: lock == m_next
+
+      TaskBase * const lock = (TaskBase *) LockTag ;
+
+      // Assign dependence to m_next.  It will be processed in the subsequent
+      // call to schedule.  Error if the dependence is reset.
+      if ( lock != Kokkos::atomic_exchange( & m_next, dep ) ) {
+        Kokkos::abort("TaskScheduler ERROR: resetting task dependence");
+      }
+
+      if ( 0 != dep ) {
+        // The future may be destroyed upon returning from this call
+        // so increment reference count to track this assignment.
+        Kokkos::atomic_increment( &(dep->m_ref_count) );
+      }
+    }
+
+  //----------------------------------------
+
+  KOKKOS_INLINE_FUNCTION
+  int32_t reference_count() const
+    { return *((int32_t volatile *)( & m_ref_count )); }
+
+};
+
+static_assert( sizeof(TaskBase) == 48
+             , "Verifying expected sizeof(TaskBase)" );
+
+//----------------------------------------------------------------------------
+//----------------------------------------------------------------------------
+
+template< typename ResultType >
+struct TaskResult {
+
+  enum : int32_t { size = sizeof(ResultType) };
+
+  using reference_type = ResultType & ;
+
+  KOKKOS_INLINE_FUNCTION static
+  ResultType * ptr( TaskBase * task )
+    {
+      return reinterpret_cast< ResultType * >
+        ( reinterpret_cast< char * >(task) + task->m_alloc_size - sizeof(ResultType) );
+    }
+
+  KOKKOS_INLINE_FUNCTION static
+  reference_type get( TaskBase * task )
+    { return *ptr( task ); }
+};
+
+template<>
+struct TaskResult< void > {
+
+  enum : int32_t { size = 0 };
+
+  using reference_type = void ;
+
+  KOKKOS_INLINE_FUNCTION static
+  void * ptr( TaskBase * ) { return (void*) 0 ; }
+
+  KOKKOS_INLINE_FUNCTION static
+  reference_type get( TaskBase * ) {}
+};
+
+} /* namespace Impl */
+} /* namespace Kokkos */
+
+//----------------------------------------------------------------------------
+//----------------------------------------------------------------------------
+
+namespace Kokkos {
+namespace Impl {
+
+template<>
+class TaskQueue< void > {};
 
 /** \brief  Manage task allocation, deallocation, and scheduling.
  *
@@ -95,7 +295,7 @@ class TaskQueueSpecialization ;
  *  All other aspects of task management have shared implementation.
  */
 template< typename ExecSpace >
-class TaskQueue {
+class TaskQueue : public TaskQueue {
 private:
 
   friend class TaskQueueSpecialization< ExecSpace > ;
@@ -106,7 +306,7 @@ private:
   using memory_space    = typename specialization::memory_space ;
   using device_type     = Kokkos::Device< execution_space , memory_space > ;
   using memory_pool     = Kokkos::MemoryPool< device_type > ;
-  using task_root_type  = Kokkos::Impl::TaskBase ;
+  using task_root_type  = Kokkos::Impl::TaskBase ;
 
   struct Destroy {
     TaskQueue * m_queue ;
@@ -198,12 +398,10 @@ public:
     }
 
   // Assign task pointer with reference counting of assigned tasks
-  template< typename LV , typename RV >
   KOKKOS_FUNCTION static
-  void assign( TaskBase< execution_space,LV,void> ** const lhs
-             , TaskBase< execution_space,RV,void> *  const rhs )
+  void assign( task_root_type ** const lhs
+             , task_root_type *  const rhs )
     {
-      using task_lhs = TaskBase< execution_space,LV,void> ;
 #if 0
   {
     printf( "assign( 0x%lx { 0x%lx %d %d } , 0x%lx { 0x%lx %d %d } )\n"
@@ -225,7 +423,7 @@ public:
 
       // Force write of *lhs
 
-      *static_cast< task_lhs * volatile * >(lhs) = rhs ;
+      *static_cast< task_root_type * volatile * >(lhs) = rhs ;
 
       Kokkos::memory_fence();
     }
@@ -238,6 +436,38 @@ public:
 
   KOKKOS_FUNCTION
   void deallocate( void * p , size_t n ); ///< Deallocate to the memory pool
+
+
+  //----------------------------------------
+  /**\brief  Allocation size for a spawned task */
+
+  template< typename FunctorType >
+  KOKKOS_FUNCTION
+  size_t spawn_allocation_size() const
+    {
+      using value_type = typename FunctorType::value_type ;
+
+      using task_type = Impl::TaskBase< execution_space
+                                      , value_type
+                                      , FunctorType > ;
+
+      enum : size_t { align = ( 1 << 4 ) , align_mask = align - 1 };
+      enum : size_t { task_size   = sizeof(task_type) };
+      enum : size_t { result_size = Impl::TaskResult< value_type >::size };
+      enum : size_t { alloc_size =
+        ( ( task_size   + align_mask ) & ~align_mask ) +
+        ( ( result_size + align_mask ) & ~align_mask ) };
+
+      return m_memory.allocate_block_size( task_size );
+    }
+
+  /**\brief  Allocation size for a when_all aggregate */
+
+  KOKKOS_FUNCTION
+  size_t when_all_allocation_size( int narg ) const
+    {
+      return m_memory.allocate_block_size( sizeof(task_root_type) + narg * sizeof(task_root_type*) );
+    }
 };
 
 } /* namespace Impl */
@@ -249,261 +479,9 @@ public:
 namespace Kokkos {
 namespace Impl {
 
-template<>
-class TaskBase< void , void , void > {
-public:
-  enum : int16_t   { TaskTeam = 0 , TaskSingle = 1 , Aggregate = 2 };
-  enum : uintptr_t { LockTag = ~uintptr_t(0) , EndTag = ~uintptr_t(1) };
-};
-
-/** \brief  Base class for task management, access, and execution.
- *
- *  Inheritance structure to allow static_cast from the task root type
- *  and a task's FunctorType.
- *
- *    // Enable a Future to access result data
- *    TaskBase< Space , ResultType , void >
- *      : TaskBase< void , void , void >
- *      { ... };
- *
- *    // Enable a functor to access the base class
- *    TaskBase< Space , ResultType , FunctorType >
- *      : TaskBase< Space , ResultType , void >
- *      , FunctorType
- *      { ... };
- *
- *
- *  States of a task:
- *
- *    Constructing State, NOT IN a linked list
- *      m_wait == 0
- *      m_next == 0
- *
- *    Scheduling transition : Constructing -> Waiting
- *      before:
- *        m_wait == 0
- *        m_next == this task's initial dependence, 0 if none
- *      after:
- *        m_wait == EndTag
- *        m_next == EndTag
- *
- *    Waiting State, IN a linked list
- *      m_apply != 0
- *      m_queue != 0
- *      m_ref_count > 0
- *      m_wait == head of linked list of tasks waiting on this task
- *      m_next == next of linked list of tasks
- *
- *    transition : Waiting -> Executing
- *      before:
- *        m_next == EndTag
- *      after::
- *        m_next == LockTag
- *
- *    Executing State, NOT IN a linked list
- *      m_apply != 0
- *      m_queue != 0
- *      m_ref_count > 0
- *      m_wait == head of linked list of tasks waiting on this task
- *      m_next == LockTag
- *
- *    Respawn transition : Executing -> Executing-Respawn
- *      before:
- *        m_next == LockTag
- *      after:
- *        m_next == this task's updated dependence, 0 if none
- *
- *    Executing-Respawn State, NOT IN a linked list
- *      m_apply != 0
- *      m_queue != 0
- *      m_ref_count > 0
- *      m_wait == head of linked list of tasks waiting on this task
- *      m_next == this task's updated dependence, 0 if none
- *
- *    transition : Executing -> Complete
- *      before:
- *        m_wait == head of linked list
- *      after:
- *        m_wait == LockTag
- *
- *    Complete State, NOT IN a linked list
- *      m_wait == LockTag: cannot add dependence
- *      m_next == LockTag: not a member of a wait queue
- *
- */
-template< typename ExecSpace >
-class TaskBase< ExecSpace , void , void >
-{
-public:
-
-  enum : int16_t   { TaskTeam   = TaskBase::TaskTeam
-                   , TaskSingle = TaskBase::TaskSingle
-                   , Aggregate  = TaskBase::Aggregate };
-
-  enum : uintptr_t { LockTag = TaskBase::LockTag
-                   , EndTag  = TaskBase::EndTag };
-
-  using execution_space = ExecSpace ;
-  using queue_type      = TaskQueue< execution_space > ;
-
-  template< typename > friend class Kokkos::TaskScheduler ;
-
-  typedef void (* function_type) ( TaskBase * , void * );
-
-  // sizeof(TaskBase) == 48
-
-  function_type  m_apply ;       ///< Apply function pointer
-  queue_type   * m_queue ;       ///< Queue in which this task resides
-  TaskBase     * m_wait ;        ///< Linked list of tasks waiting on this
-  TaskBase     * m_next ;        ///< Waiting linked-list next
-  int32_t        m_ref_count ;   ///< Reference count
-  int32_t        m_alloc_size ;  ///< Allocation size
-  int32_t        m_dep_count ;   ///< Aggregate's number of dependences
-  int16_t        m_task_type ;   ///< Type of task
-  int16_t        m_priority ;    ///< Priority of runnable task
-
-  TaskBase() = delete ;
-  TaskBase( TaskBase && ) = delete ;
-  TaskBase( const TaskBase & ) = delete ;
-  TaskBase & operator = ( TaskBase && ) = delete ;
-  TaskBase & operator = ( const TaskBase & ) = delete ;
-
-  KOKKOS_INLINE_FUNCTION ~TaskBase() = default ;
-
-  // Constructor for a runnable task
-  KOKKOS_INLINE_FUNCTION
-  constexpr TaskBase( function_type arg_apply
-                    , queue_type  * arg_queue
-                    , TaskBase    * arg_dependence
-                    , int           arg_ref_count
-                    , int           arg_alloc_size
-                    , int           arg_task_type
-                    , int           arg_priority
-                    ) noexcept
-    : m_apply(      arg_apply )
-    , m_queue(      arg_queue )
-    , m_wait( 0 )
-    , m_next(       arg_dependence )
-    , m_ref_count(  arg_ref_count )
-    , m_alloc_size( arg_alloc_size )
-    , m_dep_count( 0 )
-    , m_task_type(  arg_task_type )
-    , m_priority(   arg_priority )
-    {}
-
-  // Constructor for an aggregate task
-  KOKKOS_INLINE_FUNCTION
-  constexpr TaskBase( queue_type  * arg_queue
-                    , int           arg_ref_count
-                    , int           arg_alloc_size
-                    , int           arg_dep_count
-                    ) noexcept
-    : m_apply( 0 )
-    , m_queue( arg_queue )
-    , m_wait( 0 )
-    , m_next( 0 )
-    , m_ref_count(  arg_ref_count )
-    , m_alloc_size( arg_alloc_size )
-    , m_dep_count(  arg_dep_count )
-    , m_task_type(  Aggregate )
-    , m_priority( 0 )
-    {}
-
-  //----------------------------------------
-
-  KOKKOS_INLINE_FUNCTION
-  TaskBase ** aggregate_dependences()
-    { return reinterpret_cast( this + 1 ); }
-
-  KOKKOS_INLINE_FUNCTION
-  bool requested_respawn()
-    {
-      // This should only be called when a task has finished executing and is
-      // in the transition to either the complete or executing-respawn state.
-      TaskBase * const lock = reinterpret_cast< TaskBase * >( LockTag );
-      return lock != m_next;
-    }
-
-  KOKKOS_INLINE_FUNCTION
-  void add_dependence( TaskBase* dep )
-    {
-      // Precondition: lock == m_next
-
-      TaskBase * const lock = (TaskBase *) LockTag ;
-
-      // Assign dependence to m_next.  It will be processed in the subsequent
-      // call to schedule.  Error if the dependence is reset.
-      if ( lock != Kokkos::atomic_exchange( & m_next, dep ) ) {
-        Kokkos::abort("TaskScheduler ERROR: resetting task dependence");
-      }
-
-      if ( 0 != dep ) {
-        // The future may be destroyed upon returning from this call
-        // so increment reference count to track this assignment.
-        Kokkos::atomic_increment( &(dep->m_ref_count) );
-      }
-    }
-
-  using get_return_type = void ;
-
-  KOKKOS_INLINE_FUNCTION
-  get_return_type get() const {}
-};
-
-template < typename ExecSpace , typename ResultType >
-class TaskBase< ExecSpace , ResultType , void >
-  : public TaskBase< ExecSpace , void , void >
-{
-private:
-
-  using root_type     = TaskBase ;
-  using function_type = typename root_type::function_type ;
-  using queue_type    = typename root_type::queue_type ;
-
-  static_assert( sizeof(root_type) == 48 , "" );
-
-  TaskBase() = delete ;
-  TaskBase( TaskBase && ) = delete ;
-  TaskBase( const TaskBase & ) = delete ;
-  TaskBase & operator = ( TaskBase && ) = delete ;
-  TaskBase & operator = ( const TaskBase & ) = delete ;
-
-public:
-
-  ResultType   m_result ;
-
-  KOKKOS_INLINE_FUNCTION ~TaskBase() = default ;
-
-  // Constructor for runnable task
-  KOKKOS_INLINE_FUNCTION
-  constexpr TaskBase( function_type arg_apply
-                    , queue_type  * arg_queue
-                    , root_type   * arg_dependence
-                    , int           arg_ref_count
-                    , int           arg_alloc_size
-                    , int           arg_task_type
-                    , int           arg_priority
-                    )
-    : root_type( arg_apply
-               , arg_queue
-               , arg_dependence
-               , arg_ref_count
-               , arg_alloc_size
-               , arg_task_type
-               , arg_priority
-               )
-    , m_result()
-    {}
-
-  using get_return_type = ResultType const & ;
-
-  KOKKOS_INLINE_FUNCTION
-  get_return_type get() const { return m_result ; }
-};
-
-template< typename ExecSpace , typename ResultType , typename FunctorType >
+template< class ExecSpace , typename ResultType , class FunctorType >
 class TaskBase
-  : public TaskBase< ExecSpace , ResultType , void >
+  : public TaskBase< void , void , void >
   , public FunctorType
 {
 private:
@@ -516,50 +494,31 @@ private:
 
 public:
 
-  using root_type       = TaskBase< ExecSpace , void , void > ;
-  using base_type       = TaskBase< ExecSpace , ResultType , void > ;
-  using specialization  = TaskQueueSpecialization< ExecSpace > ;
-  using function_type   = typename root_type::function_type ;
-  using queue_type      = typename root_type::queue_type ;
-  using member_type     = typename specialization::member_type ;
+  using root_type       = TaskBase< void , void , void > ;
   using functor_type    = FunctorType ;
   using result_type     = ResultType ;
 
-  template< typename Type >
-  KOKKOS_INLINE_FUNCTION static
-  void apply_functor
-    ( Type * const task
-    , typename std::enable_if
-        < std::is_same< typename Type::result_type , void >::value
-        , member_type * const
-        >::type member
-    )
-    {
-      using fType = typename Type::functor_type ;
-      static_cast(task)->operator()( *member );
-    }
+  using specialization  = TaskQueueSpecialization< ExecSpace > ;
+  using member_type     = typename specialization::member_type ;
 
-  template< typename Type >
-  KOKKOS_INLINE_FUNCTION static
-  void apply_functor
-    ( Type * const task
-    , typename std::enable_if
-        < ! std::is_same< typename Type::result_type , void >::value
-        , member_type * const
-        >::type member
-    )
-    {
-      using fType = typename Type::functor_type ;
-      static_cast(task)->operator()( *member , task->m_result );
-    }
+  KOKKOS_INLINE_FUNCTION
+  void apply_functor( member_type * const member , void * )
+    { functor_type::operator()( *member ); }
+
+  template< typename T >
+  KOKKOS_INLINE_FUNCTION
+  void apply_functor( member_type * const member
+                    , T           * const result )
+    { functor_type::operator()( *member , *result ); }
 
   KOKKOS_FUNCTION static
   void apply( root_type * root , void * exec )
     {
       TaskBase    * const task   = static_cast< TaskBase * >( root );
       member_type * const member = reinterpret_cast< member_type * >( exec );
+      result_type * const result = TaskResult< result_type >::ptr( task );
 
-      TaskBase::template apply_functor( task , member );
+      task->apply_functor( member , result );
 
       // Task may be serial or team.
       // If team then must synchronize before querying if respawn was requested.
@@ -576,26 +535,9 @@ public:
     }
 
   // Constructor for runnable task
-  KOKKOS_INLINE_FUNCTION
-  constexpr TaskBase( function_type arg_apply
-                    , queue_type  * arg_queue
-                    , root_type   * arg_dependence
-                    , int           arg_ref_count
-                    , int           arg_alloc_size
-                    , int           arg_task_type
-                    , int           arg_priority
-                    , FunctorType && arg_functor
-                    )
-    : base_type( arg_apply
-               , arg_queue
-               , arg_dependence
-               , arg_ref_count
-               , arg_alloc_size
-               , arg_task_type
-               , arg_priority
-               )
-    , functor_type( arg_functor )
-    {}
+  KOKKOS_INLINE_FUNCTION constexpr
+  TaskBase( FunctorType && arg_functor )
+    : root_type() , functor_type( std::move(arg_functor) ) {}
 
   KOKKOS_INLINE_FUNCTION
   ~TaskBase() {}
diff --git a/lib/kokkos/core/src/impl/Kokkos_TaskQueue_impl.hpp b/lib/kokkos/core/src/impl/Kokkos_TaskQueue_impl.hpp
index aee381afad..1974f7e1ca 100644
--- a/lib/kokkos/core/src/impl/Kokkos_TaskQueue_impl.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_TaskQueue_impl.hpp
@@ -44,6 +44,8 @@
 #include 
 #if defined( KOKKOS_ENABLE_TASKDAG )
 
+#define KOKKOS_IMPL_DEBUG_TASKDAG_SCHEDULING 0
+
 namespace Kokkos {
 namespace Impl {
 
@@ -100,9 +102,11 @@ KOKKOS_FUNCTION
 void TaskQueue< ExecSpace >::decrement
   ( TaskQueue< ExecSpace >::task_root_type * task )
 {
-  const int count = Kokkos::atomic_fetch_add(&(task->m_ref_count),-1);
+  task_root_type volatile & t = *task ;
 
-#if 0
+  const int count = Kokkos::atomic_fetch_add(&(t.m_ref_count),-1);
+
+#if KOKKOS_IMPL_DEBUG_TASKDAG_SCHEDULING
   if ( 1 == count ) {
     printf( "decrement-destroy( 0x%lx { 0x%lx %d %d } )\n"
           , uintptr_t( task )
@@ -114,9 +118,13 @@ void TaskQueue< ExecSpace >::decrement
 #endif
 
   if ( ( 1 == count ) &&
-       ( task->m_next == (task_root_type *) task_root_type::LockTag ) ) {
+       ( t.m_next == (task_root_type *) task_root_type::LockTag ) ) {
     // Reference count is zero and task is complete, deallocate.
-    task->m_queue->deallocate( task , task->m_alloc_size );
+
+    TaskQueue< ExecSpace > * const queue =
+      static_cast< TaskQueue< ExecSpace > * >( t.m_queue );
+
+    queue->deallocate( task , t.m_alloc_size );
   }
   else if ( count <= 1 ) {
     Kokkos::abort("TaskScheduler task has negative reference count or is incomplete" );
@@ -171,7 +179,7 @@ bool TaskQueue< ExecSpace >::push_task
   // Fail the push attempt if the queue is locked;
   // otherwise retry until the push succeeds.
 
-#if 0
+#if KOKKOS_IMPL_DEBUG_TASKDAG_SCHEDULING
   printf( "push_task( 0x%lx { 0x%lx } 0x%lx { 0x%lx 0x%lx %d %d %d } )\n"
         , uintptr_t(queue)
         , uintptr_t(*queue)
@@ -186,9 +194,9 @@ bool TaskQueue< ExecSpace >::push_task
   task_root_type * const zero = (task_root_type *) 0 ;
   task_root_type * const lock = (task_root_type *) task_root_type::LockTag ;
 
-  task_root_type * volatile * const next = & task->m_next ;
+  task_root_type * volatile & next = task->m_next ;
 
-  if ( zero != *next ) {
+  if ( zero != next ) {
     Kokkos::abort("TaskQueue::push_task ERROR: already a member of another queue" );
   }
 
@@ -196,9 +204,9 @@ bool TaskQueue< ExecSpace >::push_task
 
   while ( lock != y ) {
 
-    *next = y ;
+    next = y ;
 
-    // Do not proceed until '*next' has been stored.
+    // Do not proceed until 'next' has been stored.
     Kokkos::memory_fence();
 
     task_root_type * const x = y ;
@@ -211,9 +219,9 @@ bool TaskQueue< ExecSpace >::push_task
   // Failed, replace 'task->m_next' value since 'task' remains
   // not a member of a queue.
 
-  *next = zero ;
+  next = zero ;
 
-  // Do not proceed until '*next' has been stored.
+  // Do not proceed until 'next' has been stored.
   Kokkos::memory_fence();
 
   return false ;
@@ -270,11 +278,13 @@ TaskQueue< ExecSpace >::pop_ready_task
       // This thread has exclusive access to
       // the queue and the popped task's m_next.
 
-      *queue = task->m_next ; task->m_next = lock ;
+      task_root_type * volatile & next = task->m_next ;
+
+      *queue = next ; next = lock ;
 
       Kokkos::memory_fence();
 
-#if 0
+#if KOKKOS_IMPL_DEBUG_TASKDAG_SCHEDULING
       printf( "pop_ready_task( 0x%lx 0x%lx { 0x%lx 0x%lx %d %d %d } )\n"
             , uintptr_t(queue)
             , uintptr_t(task)
@@ -323,7 +333,7 @@ void TaskQueue< ExecSpace >::schedule_runnable
   //     task->m_wait == head of linked list (queue)
   //     task->m_next == member of linked list (queue)
 
-#if 0
+#if KOKKOS_IMPL_DEBUG_TASKDAG_SCHEDULING
   printf( "schedule_runnable( 0x%lx { 0x%lx 0x%lx %d %d %d }\n"
         , uintptr_t(task)
         , uintptr_t(task->m_wait)
@@ -337,20 +347,22 @@ void TaskQueue< ExecSpace >::schedule_runnable
   task_root_type * const lock = (task_root_type *) task_root_type::LockTag ;
   task_root_type * const end  = (task_root_type *) task_root_type::EndTag ;
 
+  task_root_type volatile & t = *task ;
+
   bool respawn = false ;
 
   //----------------------------------------
 
-  if ( zero == task->m_wait ) {
+  if ( zero == t.m_wait ) {
     // Task in Constructing state
     // - Transition to Waiting state
     // Preconditions:
     // - call occurs exclusively within a single thread
 
-    task->m_wait = end ;
+    t.m_wait = end ;
     // Task in Waiting state
   }
-  else if ( lock != task->m_wait ) {
+  else if ( lock != t.m_wait ) {
     // Task in Executing state with Respawn request
     // - Update dependence
     // - Transition to Waiting state
@@ -373,7 +385,9 @@ void TaskQueue< ExecSpace >::schedule_runnable
 
   // Exclusive access so don't need an atomic exchange
   // task_root_type * dep = Kokkos::atomic_exchange( & task->m_next , zero );
-  task_root_type * dep = task->m_next ; task->m_next = zero ;
+  task_root_type * dep = t.m_next ; t.m_next = zero ;
+
+  Kokkos::memory_fence();
 
   const bool is_ready =
     ( 0 == dep ) || ( ! push_task( & dep->m_wait , task ) );
@@ -398,7 +412,7 @@ void TaskQueue< ExecSpace >::schedule_runnable
     Kokkos::atomic_increment( & m_ready_count );
 
     task_root_type * volatile * const ready_queue =
-      & m_ready[ task->m_priority ][ task->m_task_type ];
+      & m_ready[ t.m_priority ][ t.m_task_type ];
 
     // A push_task fails if the ready queue is locked.
     // A ready queue is only locked during a push or pop;
@@ -441,7 +455,7 @@ void TaskQueue< ExecSpace >::schedule_aggregate
   //     task->m_wait == head of linked list (queue)
   //     task->m_next == member of linked list (queue)
 
-#if 0
+#if KOKKOS_IMPL_DEBUG_TASKDAG_SCHEDULING
   printf( "schedule_aggregate( 0x%lx { 0x%lx 0x%lx %d %d %d }\n"
         , uintptr_t(task)
         , uintptr_t(task->m_wait)
@@ -455,18 +469,20 @@ void TaskQueue< ExecSpace >::schedule_aggregate
   task_root_type * const lock = (task_root_type *) task_root_type::LockTag ;
   task_root_type * const end  = (task_root_type *) task_root_type::EndTag ;
 
+  task_root_type volatile & t = *task ;
+
   //----------------------------------------
 
-  if ( zero == task->m_wait ) {
+  if ( zero == t.m_wait ) {
     // Task in Constructing state
     // - Transition to Waiting state
     // Preconditions:
     // - call occurs exclusively within a single thread
 
-    task->m_wait = end ;
+    t.m_wait = end ;
     // Task in Waiting state
   }
-  else if ( lock == task->m_wait ) {
+  else if ( lock == t.m_wait ) {
     // Task in Complete state
     Kokkos::abort("TaskQueue::schedule_aggregate ERROR: task is complete");
   }
@@ -477,14 +493,14 @@ void TaskQueue< ExecSpace >::schedule_aggregate
   // (1) created or
   // (2) being removed from a completed task's wait list.
 
-  task_root_type ** const aggr = task->aggregate_dependences();
+  task_root_type * volatile * const aggr = t.aggregate_dependences();
 
   // Assume the 'when_all' is complete until a dependence is
   // found that is not complete.
 
   bool is_complete = true ;
 
-  for ( int i = task->m_dep_count ; 0 < i && is_complete ; ) {
+  for ( int i = t.m_dep_count ; 0 < i && is_complete ; ) {
 
     --i ;
 
@@ -523,7 +539,7 @@ void TaskQueue< ExecSpace >::schedule_aggregate
     // Complete the when_all 'task' to schedule other tasks
     // that are waiting for the when_all 'task' to complete.
 
-    task->m_next = lock ;
+    t.m_next = lock ;
 
     complete( task );
 
@@ -573,7 +589,7 @@ void TaskQueue< ExecSpace >::complete
   task_root_type * const lock = (task_root_type *) task_root_type::LockTag ;
   task_root_type * const end  = (task_root_type *) task_root_type::EndTag ;
 
-#if 0
+#if KOKKOS_IMPL_DEBUG_TASKDAG_SCHEDULING
   printf( "complete( 0x%lx { 0x%lx 0x%lx %d %d %d }\n"
         , uintptr_t(task)
         , uintptr_t(task->m_wait)
@@ -584,11 +600,13 @@ void TaskQueue< ExecSpace >::complete
   fflush( stdout );
 #endif
 
-  const bool runnable = task_root_type::Aggregate != task->m_task_type ;
+  task_root_type volatile & t = *task ;
+
+  const bool runnable = task_root_type::Aggregate != t.m_task_type ;
 
   //----------------------------------------
 
-  if ( runnable && lock != task->m_next ) {
+  if ( runnable && lock != t.m_next ) {
     // Is a runnable task has finished executing and requested respawn.
     // Schedule the task for subsequent execution.
 
@@ -607,7 +625,7 @@ void TaskQueue< ExecSpace >::complete
     // Stop other tasks from adding themselves to this task's wait queue
     // by locking the head of this task's wait queue.
 
-    task_root_type * x = Kokkos::atomic_exchange( & task->m_wait , lock );
+    task_root_type * x = Kokkos::atomic_exchange( & t.m_wait , lock );
 
     if ( x != (task_root_type *) lock ) {
 
@@ -627,9 +645,13 @@ void TaskQueue< ExecSpace >::complete
         // Have exclusive access to 'x' until it is scheduled
         // Set x->m_next = zero  <=  no dependence, not a respawn
 
-        task_root_type * const next = x->m_next ; x->m_next = 0 ;
+        task_root_type volatile & vx = *x ;
 
-        if ( task_root_type::Aggregate != x->m_task_type ) {
+        task_root_type * const next = vx.m_next ; vx.m_next = 0 ;
+
+        Kokkos::memory_fence();
+
+        if ( task_root_type::Aggregate != vx.m_task_type ) {
           schedule_runnable( x );
         }
         else {
diff --git a/lib/kokkos/core/src/impl/Kokkos_ViewArray.hpp b/lib/kokkos/core/src/impl/Kokkos_ViewArray.hpp
index c55636b64e..ed1a71bea7 100644
--- a/lib/kokkos/core/src/impl/Kokkos_ViewArray.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_ViewArray.hpp
@@ -1,13 +1,13 @@
 /*
 //@HEADER
 // ************************************************************************
-// 
+//
 //                        Kokkos v. 2.0
 //              Copyright (2014) Sandia Corporation
-// 
+//
 // Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
 // the U.S. Government retains certain rights in this software.
-// 
+//
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
@@ -36,7 +36,7 @@
 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 //
 // Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
-// 
+//
 // ************************************************************************
 //@HEADER
 */
@@ -47,7 +47,6 @@
 #include 
 
 namespace Kokkos {
-namespace Experimental {
 namespace Impl {
 
 template< class DataType , class ArrayLayout , class V , size_t N , class P >
@@ -94,13 +93,12 @@ public:
   typedef typename ViewDataType< non_const_scalar_type , array_scalar_dimension >::type  non_const_scalar_array_type ;
 };
 
-}}} // namespace Kokkos::Experimental::Impl
+}} // namespace Kokkos::Impl
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
-namespace Experimental {
 namespace Impl {
 
 /** \brief  View mapping for non-specialized data type and standard layout */
@@ -597,7 +595,7 @@ public:
     }
 };
 
-}}} // namespace Kokkos::Experimental::Impl
+}} // namespace Kokkos::Impl
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
diff --git a/lib/kokkos/core/src/impl/Kokkos_ViewCtor.hpp b/lib/kokkos/core/src/impl/Kokkos_ViewCtor.hpp
index 6381aee468..f32c6bb2ee 100644
--- a/lib/kokkos/core/src/impl/Kokkos_ViewCtor.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_ViewCtor.hpp
@@ -96,6 +96,27 @@ struct is_view_label< const char[N] > : public std::true_type {};
 template< typename ... P >
 struct ViewCtorProp ;
 
+// Forward declare
+template< typename Specialize , typename T >
+struct CommonViewAllocProp ;
+
+/* Common value_type stored as ViewCtorProp
+ */
+template< typename Specialize , typename T >
+struct ViewCtorProp< void , CommonViewAllocProp >
+{
+  ViewCtorProp() = default ;
+  ViewCtorProp( const ViewCtorProp & ) = default ;
+  ViewCtorProp & operator = ( const ViewCtorProp & ) = default ;
+
+  using type = CommonViewAllocProp ;
+
+  ViewCtorProp( const type & arg ) : value( arg ) {}
+  ViewCtorProp( type && arg ) : value( arg ) {}
+
+  type value ;
+};
+
 /*  std::integral_constant are dummy arguments
  *  that avoid duplicate base class errors
  */
diff --git a/lib/kokkos/core/src/impl/Kokkos_ViewMapping.hpp b/lib/kokkos/core/src/impl/Kokkos_ViewMapping.hpp
index 900bd88f1c..d346f9e639 100644
--- a/lib/kokkos/core/src/impl/Kokkos_ViewMapping.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_ViewMapping.hpp
@@ -62,7 +62,6 @@
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
-namespace Experimental {
 namespace Impl {
 
 template< unsigned I , size_t ... Args >
@@ -250,7 +249,7 @@ struct ViewDimensionAssignable< ViewDimension< DstArgs ... >
 
 };
 
-}}} // namespace Kokkos::Experimental::Impl
+}} // namespace Kokkos::Impl
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
@@ -266,14 +265,11 @@ struct ALL_t {
 }} // namespace Kokkos::Impl
 
 namespace Kokkos {
-namespace Experimental {
 namespace Impl {
 
-using Kokkos::Impl::ALL_t ;
-
 template< class T >
 struct is_integral_extent_type
-{ enum { value = std::is_same::value ? 1 : 0 }; };
+{ enum { value = std::is_same::value ? 1 : 0 }; };
 
 template< class iType >
 struct is_integral_extent_type< std::pair >
@@ -314,10 +310,10 @@ struct SubviewLegalArgsCompileTime;
 
 template
 struct SubviewLegalArgsCompileTime {
-  enum { value      =(((CurrentArg==RankDest-1) && (Kokkos::Experimental::Impl::is_integral_extent_type::value)) ||
+  enum { value      =(((CurrentArg==RankDest-1) && (Kokkos::Impl::is_integral_extent_type::value)) ||
                       ((CurrentArg>=RankDest) && (std::is_integral::value)) ||
                       ((CurrentArg::value)) ||
-                      ((CurrentArg==0) && (Kokkos::Experimental::Impl::is_integral_extent_type::value))
+                      ((CurrentArg==0) && (Kokkos::Impl::is_integral_extent_type::value))
                      ) && (SubviewLegalArgsCompileTime::value)};
 };
 
@@ -331,7 +327,7 @@ struct SubviewLegalArgsCompileTime
 struct SubviewLegalArgsCompileTime {
-  enum { value      =(((CurrentArg==RankSrc-RankDest) && (Kokkos::Experimental::Impl::is_integral_extent_type::value)) ||
+  enum { value      =(((CurrentArg==RankSrc-RankDest) && (Kokkos::Impl::is_integral_extent_type::value)) ||
                       ((CurrentArg::value)) ||
                       ((CurrentArg>=RankSrc-RankDest) && (std::is_same::value))
                      ) && (SubviewLegalArgsCompileTime::value)};
@@ -403,7 +399,7 @@ private:
   bool set( unsigned domain_rank
           , unsigned range_rank
           , const ViewDimension< DimArgs ... > & dim
-          , const Kokkos::Experimental::Impl::ALL_t
+          , const Kokkos::Impl::ALL_t
           , Args ... args )
     {
       m_begin[  domain_rank ] = 0 ;
@@ -519,7 +515,7 @@ private:
             , unsigned domain_rank
             , unsigned range_rank
             , const ViewDimension< DimArgs ... > & dim
-            , const Kokkos::Experimental::Impl::ALL_t
+            , const Kokkos::Impl::ALL_t
             , Args ... args ) const
     {
       const int n = std::min( buf_len ,
@@ -670,13 +666,12 @@ public:
     { return unsigned(i) < InternalRangeRank ? m_index[i] : ~0u ; }
 };
 
-}}} // namespace Kokkos::Experimental::Impl
+}} // namespace Kokkos::Impl
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
-namespace Experimental {
 namespace Impl {
 
 /** \brief  Given a value type and dimension generate the View data type */
@@ -814,13 +809,12 @@ public:
   typedef non_const_type  non_const_scalar_array_type ;
 };
 
-}}} // namespace Kokkos::Experimental::Impl
+}} // namespace Kokkos::Impl
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
-namespace Experimental {
 namespace Impl {
 
 template < class Dimension , class Layout , typename Enable = void >
@@ -1228,14 +1222,14 @@ private:
 
     // If memory alignment is a multiple of the trivial scalar size then attempt to align.
     enum { align = 0 != TrivialScalarSize && 0 == mod ? div : 0 };
-    enum { div_ok = div ? div : 1 }; // To valid modulo zero in constexpr
+    enum { div_ok = (div != 0) ? div : 1 }; // To valid modulo zero in constexpr
 
     KOKKOS_INLINE_FUNCTION
     static constexpr size_t stride( size_t const N )
-      {
-        return ( align && ( Kokkos::Impl::MEMORY_ALIGNMENT_THRESHOLD * align < N ) && ( N % div_ok ) )
-               ? N + align - ( N % div_ok ) : N ;
-      }
+    {
+      return ( (align != 0) && ((Kokkos::Impl::MEMORY_ALIGNMENT_THRESHOLD * align) < N) && ((N % div_ok) != 0) )
+             ? N + align - ( N % div_ok ) : N ;
+    }
   };
 
 public:
@@ -1707,12 +1701,12 @@ private:
 
     // If memory alignment is a multiple of the trivial scalar size then attempt to align.
     enum { align = 0 != TrivialScalarSize && 0 == mod ? div : 0 };
-    enum { div_ok = div ? div : 1 }; // To valid modulo zero in constexpr
+    enum { div_ok = (div != 0) ? div : 1 }; // To valid modulo zero in constexpr
 
     KOKKOS_INLINE_FUNCTION
     static constexpr size_t stride( size_t const N )
     {
-      return ( align && ( Kokkos::Impl::MEMORY_ALIGNMENT_THRESHOLD * align < N ) && ( N % div_ok ) )
+      return ( (align != 0) && ((Kokkos::Impl::MEMORY_ALIGNMENT_THRESHOLD * align) < N) && ((N % div_ok) != 0) )
              ? N + align - ( N % div_ok ) : N ;
     }
   };
@@ -2225,13 +2219,12 @@ public:
     {}
 };
 
-}}} // namespace Kokkos::Experimental::Impl
+}} // namespace Kokkos::Impl
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
-namespace Experimental {
 namespace Impl {
 
 /** \brief  ViewDataHandle provides the type of the 'data handle' which the view
@@ -2422,13 +2415,12 @@ struct ViewDataHandle< Traits ,
     return handle_type( arg_data_ptr + offset );
   }
 };
-}}} // namespace Kokkos::Experimental::Impl
+}} // namespace Kokkos::Impl
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
-namespace Experimental {
 namespace Impl {
 
 //----------------------------------------------------------------------------
@@ -2451,8 +2443,9 @@ template< class ExecSpace , class ValueType >
 struct ViewValueFunctor< ExecSpace , ValueType , false /* is_scalar */ >
 {
   typedef Kokkos::RangePolicy< ExecSpace > PolicyType ;
+  typedef typename ExecSpace::execution_space Exec;
 
-  ExecSpace   space ;
+  Exec        space ;
   ValueType * ptr ;
   size_t      n ;
   bool        destroy ;
@@ -2597,6 +2590,9 @@ private:
 
 public:
 
+  typedef void printable_label_typedef;
+  enum { is_managed = Traits::is_managed };
+
   //----------------------------------------
   // Domain dimensions
 
@@ -2944,7 +2940,7 @@ public:
           Kokkos::abort("View Assignment: trying to assign runtime dimension to non matching compile time dimension.");
       }
       dst.m_offset = dst_offset_type( src.m_offset );
-      dst.m_handle = Kokkos::Experimental::Impl::ViewDataHandle< DstTraits >::assign( src.m_handle , src_track );
+      dst.m_handle = Kokkos::Impl::ViewDataHandle< DstTraits >::assign( src.m_handle , src_track );
     }
 };
 
@@ -3102,7 +3098,7 @@ public:
 
 //----------------------------------------------------------------------------
 
-}}} // namespace Kokkos::Experimental::Impl
+}} // namespace Kokkos::Impl
 
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
@@ -3151,6 +3147,77 @@ void view_error_operator_bounds
   view_error_operator_bounds(buf+n,len-n,map,args...);
 }
 
+#if ! defined( KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_HOST )
+
+/* Check #3: is the View managed as determined by the MemoryTraits? */
+template< class MapType,
+  bool is_managed = (MapType::is_managed != 0) >
+struct OperatorBoundsErrorOnDevice;
+
+template< class MapType >
+struct OperatorBoundsErrorOnDevice< MapType, false > {
+KOKKOS_INLINE_FUNCTION
+static void run(MapType const&) {
+  Kokkos::abort("View bounds error");
+}
+};
+
+template< class MapType >
+struct OperatorBoundsErrorOnDevice< MapType, true > {
+KOKKOS_INLINE_FUNCTION
+static void run(MapType const& map) {
+  char const* const user_alloc_start = reinterpret_cast(map.data());
+  char const* const header_start = user_alloc_start - sizeof(SharedAllocationHeader);
+  SharedAllocationHeader const* const header =
+    reinterpret_cast(header_start);
+  char const* const label = header->label();
+  enum { LEN = 128 };
+  char msg[LEN];
+  char const* const first_part = "View bounds error of view ";
+  char* p = msg;
+  char* const end = msg + LEN - 1;
+  for (char const* p2 = first_part; (*p2 != '\0') && (p < end); ++p, ++p2) {
+    *p = *p2;
+  }
+  for (char const* p2 = label; (*p2 != '\0') && (p < end); ++p, ++p2) {
+    *p = *p2;
+  }
+  *p = '\0';
+  Kokkos::abort(msg);
+}
+};
+
+/* Check #2: does the ViewMapping have the printable_label_typedef defined?
+   See above that only the non-specialized standard-layout ViewMapping has
+   this defined by default.
+   The existence of this typedef indicates the existence of MapType::is_managed */
+template< class T, class Enable = void >
+struct has_printable_label_typedef : public std::false_type {};
+
+template
+struct has_printable_label_typedef<
+  T, typename enable_if_type::type>
+  : public std::true_type
+{};
+
+template< class MapType >
+KOKKOS_INLINE_FUNCTION
+void operator_bounds_error_on_device(
+    MapType const&,
+    std::false_type) {
+  Kokkos::abort("View bounds error");
+}
+
+template< class MapType >
+KOKKOS_INLINE_FUNCTION
+void operator_bounds_error_on_device(
+    MapType const& map,
+    std::true_type) {
+  OperatorBoundsErrorOnDevice< MapType >::run(map);
+}
+
+#endif // ! defined( KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_HOST )
+
 template< class MemorySpace , class MapType , class ... Args >
 KOKKOS_INLINE_FUNCTION
 void view_verify_operator_bounds
@@ -3166,7 +3233,17 @@ void view_verify_operator_bounds
     view_error_operator_bounds<0>( buffer + n , LEN - n , map , args ... );
     Kokkos::Impl::throw_runtime_exception(std::string(buffer));
 #else
-    Kokkos::abort("View bounds error");
+    /* Check #1: is there a SharedAllocationRecord?
+       (we won't use it, but if its not there then there isn't
+        a corresponding SharedAllocationHeader containing a label).
+       This check should cover the case of Views that don't
+       have the Unmanaged trait but were initialized by pointer. */
+    if (tracker.has_record()) {
+      operator_bounds_error_on_device(
+          map, has_printable_label_typedef());
+    } else {
+      Kokkos::abort("View bounds error");
+    }
 #endif
   }
 }
diff --git a/lib/kokkos/core/src/impl/Kokkos_ViewTile.hpp b/lib/kokkos/core/src/impl/Kokkos_ViewTile.hpp
index ecbcf72fe0..5a8600e0ae 100644
--- a/lib/kokkos/core/src/impl/Kokkos_ViewTile.hpp
+++ b/lib/kokkos/core/src/impl/Kokkos_ViewTile.hpp
@@ -1,13 +1,13 @@
 /*
 //@HEADER
 // ************************************************************************
-// 
+//
 //                        Kokkos v. 2.0
 //              Copyright (2014) Sandia Corporation
-// 
+//
 // Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
 // the U.S. Government retains certain rights in this software.
-// 
+//
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
@@ -36,7 +36,7 @@
 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 //
 // Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
-// 
+//
 // ************************************************************************
 //@HEADER
 */
@@ -48,7 +48,6 @@
 //----------------------------------------------------------------------------
 
 namespace Kokkos {
-namespace Experimental {
 namespace Impl {
 
 // View mapping for rank two tiled array
@@ -195,11 +194,9 @@ struct ViewMapping
 };
 
 } /* namespace Impl */
-} /* namespace Experimental */
 } /* namespace Kokkos */
 
 namespace Kokkos {
-namespace Experimental {
 
 template< typename T , unsigned N0 , unsigned N1 , class ... P >
 KOKKOS_INLINE_FUNCTION
@@ -217,7 +214,6 @@ tile_subview( const Kokkos::View,P...> &
     ( src , SrcLayout() , i_tile0 , i_tile1 );
 }
 
-} /* namespace Experimental */
 } /* namespace Kokkos */
 
 //----------------------------------------------------------------------------
diff --git a/lib/kokkos/core/src/impl/Kokkos_spinwait.cpp b/lib/kokkos/core/src/impl/Kokkos_spinwait.cpp
deleted file mode 100644
index 101b714fcd..0000000000
--- a/lib/kokkos/core/src/impl/Kokkos_spinwait.cpp
+++ /dev/null
@@ -1,183 +0,0 @@
-/*
-//@HEADER
-// ************************************************************************
-//
-//                        Kokkos v. 2.0
-//              Copyright (2014) Sandia Corporation
-//
-// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
-// the U.S. Government retains certain rights in this software.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-// 1. Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//
-// 2. Redistributions in binary form must reproduce the above copyright
-// notice, this list of conditions and the following disclaimer in the
-// documentation and/or other materials provided with the distribution.
-//
-// 3. Neither the name of the Corporation nor the names of the
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
-// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
-// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
-// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
-// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
-// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
-// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
-// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
-// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
-// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
-// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-//
-// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
-//
-// ************************************************************************
-//@HEADER
-*/
-
-#include 
-#if defined( KOKKOS_ACTIVE_EXECUTION_MEMORY_SPACE_HOST )
-
-#include 
-
-#include 
-#include 
-
-/*--------------------------------------------------------------------------*/
-
-#if !defined( _WIN32 )
-  #if defined( KOKKOS_ENABLE_ASM )
-    #if defined( __arm__ ) || defined( __aarch64__ )
-      /* No-operation instruction to idle the thread. */
-      #define KOKKOS_INTERNAL_PAUSE
-    #else
-      /* Pause instruction to prevent excess processor bus usage */
-      #define KOKKOS_INTERNAL_PAUSE   asm volatile("pause\n":::"memory")
-    #endif
-    #define KOKKOS_INTERNAL_NOP2    asm volatile("nop\n" "nop\n")
-    #define KOKKOS_INTERNAL_NOP4    KOKKOS_INTERNAL_NOP2;  KOKKOS_INTERNAL_NOP2
-    #define KOKKOS_INTERNAL_NOP8    KOKKOS_INTERNAL_NOP4;  KOKKOS_INTERNAL_NOP4;
-    #define KOKKOS_INTERNAL_NOP16   KOKKOS_INTERNAL_NOP8;  KOKKOS_INTERNAL_NOP8;
-    #define KOKKOS_INTERNAL_NOP32   KOKKOS_INTERNAL_NOP16; KOKKOS_INTERNAL_NOP16;
-    namespace {
-    inline void kokkos_internal_yield( const unsigned i ) noexcept {
-      switch (Kokkos::Impl::bit_scan_reverse((i >> 2)+1u)) {
-      case 0u:  KOKKOS_INTERNAL_NOP2;  break;
-      case 1u:  KOKKOS_INTERNAL_NOP4;  break;
-      case 2u:  KOKKOS_INTERNAL_NOP8;  break;
-      case 3u:  KOKKOS_INTERNAL_NOP16; break;
-      default: KOKKOS_INTERNAL_NOP32;
-      }
-      KOKKOS_INTERNAL_PAUSE;
-    }
-    }
-  #else
-    #include 
-    namespace {
-    inline void kokkos_internal_yield( const unsigned ) noexcept {
-      sched_yield();
-    }
-    }
-  #endif
-#else // defined( _WIN32 )
-  #if defined ( KOKKOS_ENABLE_WINTHREAD )
-    #include 
-    namespace {
-    inline void kokkos_internal_yield( const unsigned ) noexcept {
-      Sleep(0);
-    }
-    }
-  #elif defined( _MSC_VER )
-    #define NOMINMAX
-    #include 
-    #include 
-    namespace {
-    inline void kokkos_internal_yield( const unsigned ) noexcept {
-      YieldProcessor();
-    }
-    }
-  #else
-    #define KOKKOS_INTERNAL_PAUSE   __asm__ __volatile__("pause\n":::"memory")
-    #define KOKKOS_INTERNAL_NOP2    __asm__ __volatile__("nop\n" "nop")
-    #define KOKKOS_INTERNAL_NOP4    KOKKOS_INTERNAL_NOP2;  KOKKOS_INTERNAL_NOP2
-    #define KOKKOS_INTERNAL_NOP8    KOKKOS_INTERNAL_NOP4;  KOKKOS_INTERNAL_NOP4;
-    #define KOKKOS_INTERNAL_NOP16   KOKKOS_INTERNAL_NOP8;  KOKKOS_INTERNAL_NOP8;
-    #define KOKKOS_INTERNAL_NOP32   KOKKOS_INTERNAL_NOP16; KOKKOS_INTERNAL_NOP16;
-    namespace {
-    inline void kokkos_internal_yield( const unsigned i ) noexcept {
-      switch (Kokkos::Impl::bit_scan_reverse((i >> 2)+1u)) {
-      case 0:  KOKKOS_INTERNAL_NOP2;  break;
-      case 1:  KOKKOS_INTERNAL_NOP4;  break;
-      case 2:  KOKKOS_INTERNAL_NOP8;  break;
-      case 3:  KOKKOS_INTERNAL_NOP16; break;
-      default: KOKKOS_INTERNAL_NOP32;
-      }
-      KOKKOS_INTERNAL_PAUSE;
-    }
-    }
-  #endif
-#endif
-
-
-/*--------------------------------------------------------------------------*/
-
-namespace Kokkos {
-namespace Impl {
-
-void spinwait_while_equal( volatile int32_t & flag , const int32_t value )
-{
-  Kokkos::store_fence();
-  unsigned i = 0;
-  while ( value == flag ) {
-    kokkos_internal_yield(i);
-    ++i;
-  }
-  Kokkos::load_fence();
-}
-
-void spinwait_until_equal( volatile int32_t & flag , const int32_t value )
-{
-  Kokkos::store_fence();
-  unsigned i = 0;
-  while ( value != flag ) {
-    kokkos_internal_yield(i);
-    ++i;
-  }
-  Kokkos::load_fence();
-}
-
-void spinwait_while_equal( volatile int64_t & flag , const int64_t value )
-{
-  Kokkos::store_fence();
-  unsigned i = 0;
-  while ( value == flag ) {
-    kokkos_internal_yield(i);
-    ++i;
-  }
-  Kokkos::load_fence();
-}
-
-void spinwait_until_equal( volatile int64_t & flag , const int64_t value )
-{
-  Kokkos::store_fence();
-  unsigned i = 0;
-  while ( value != flag ) {
-    kokkos_internal_yield(i);
-    ++i;
-  }
-  Kokkos::load_fence();
-}
-
-} /* namespace Impl */
-} /* namespace Kokkos */
-
-#else
-void KOKKOS_CORE_SRC_IMPL_SPINWAIT_PREVENT_LINK_ERROR() {}
-#endif
-
diff --git a/lib/kokkos/core/unit_test/CMakeLists.txt b/lib/kokkos/core/unit_test/CMakeLists.txt
index 5d6f25ac95..475b6bb48a 100644
--- a/lib/kokkos/core/unit_test/CMakeLists.txt
+++ b/lib/kokkos/core/unit_test/CMakeLists.txt
@@ -57,6 +57,7 @@ IF(Kokkos_ENABLE_Serial)
       serial/TestSerial_ViewMapping_b.cpp
       serial/TestSerial_ViewMapping_subview.cpp
       serial/TestSerial_ViewOfClass.cpp
+      serial/TestSerial_WorkGraph.cpp
     COMM serial mpi
     NUM_MPI_PROCS 1
     FAIL_REGULAR_EXPRESSION "  FAILED  "
@@ -102,6 +103,7 @@ IF(Kokkos_ENABLE_Pthread)
       threads/TestThreads_ViewMapping_b.cpp
       threads/TestThreads_ViewMapping_subview.cpp
       threads/TestThreads_ViewOfClass.cpp
+      threads/TestThreads_WorkGraph.cpp
     COMM serial mpi
     NUM_MPI_PROCS 1
     FAIL_REGULAR_EXPRESSION "  FAILED  "
@@ -147,6 +149,8 @@ IF(Kokkos_ENABLE_OpenMP)
       openmp/TestOpenMP_ViewMapping_b.cpp
       openmp/TestOpenMP_ViewMapping_subview.cpp
       openmp/TestOpenMP_ViewOfClass.cpp
+      openmp/TestOpenMP_WorkGraph.cpp
+      openmp/TestOpenMP_UniqueToken.cpp
     COMM serial mpi
     NUM_MPI_PROCS 1
     FAIL_REGULAR_EXPRESSION "  FAILED  "
@@ -237,6 +241,7 @@ IF(Kokkos_ENABLE_Cuda)
       cuda/TestCuda_ViewMapping_b.cpp
       cuda/TestCuda_ViewMapping_subview.cpp
       cuda/TestCuda_ViewOfClass.cpp
+      cuda/TestCuda_WorkGraph.cpp
     COMM serial mpi
     NUM_MPI_PROCS 1
     FAIL_REGULAR_EXPRESSION "  FAILED  "
@@ -253,6 +258,7 @@ TRIBITS_ADD_EXECUTABLE_AND_TEST(
     default/TestDefaultDeviceType_b.cpp
     default/TestDefaultDeviceType_c.cpp
     default/TestDefaultDeviceType_d.cpp
+    default/TestDefaultDeviceTypeResize.cpp
   COMM serial mpi
   NUM_MPI_PROCS 1
   FAIL_REGULAR_EXPRESSION "  FAILED  "
diff --git a/lib/kokkos/core/unit_test/Makefile b/lib/kokkos/core/unit_test/Makefile
index 41f192a486..c877aa7dd2 100644
--- a/lib/kokkos/core/unit_test/Makefile
+++ b/lib/kokkos/core/unit_test/Makefile
@@ -62,8 +62,9 @@ endif
 	OBJ_CUDA += TestCuda_TeamReductionScan.o
 	OBJ_CUDA += TestCuda_Other.o
 	OBJ_CUDA += TestCuda_MDRange.o
-	OBJ_CUDA += TestCuda_Task.o
+	OBJ_CUDA += TestCuda_Task.o TestCuda_WorkGraph.o
 	OBJ_CUDA += TestCuda_Spaces.o
+	OBJ_CUDA += TestCuda_UniqueToken.o
 	
 	TARGETS += KokkosCore_UnitTest_Cuda
 
@@ -121,7 +122,8 @@ endif
 	OBJ_OPENMP += TestOpenMP_TeamReductionScan.o
 	OBJ_OPENMP += TestOpenMP_Other.o
 	OBJ_OPENMP += TestOpenMP_MDRange.o
-	OBJ_OPENMP += TestOpenMP_Task.o
+	OBJ_OPENMP += TestOpenMP_Task.o TestOpenMP_WorkGraph.o
+	OBJ_OPENMP += TestOpenMP_UniqueToken.o
 	
 	TARGETS += KokkosCore_UnitTest_OpenMP
 
@@ -208,7 +210,7 @@ endif
         OBJ_SERIAL += TestSerial_TeamReductionScan.o
         OBJ_SERIAL += TestSerial_Other.o
         OBJ_SERIAL += TestSerial_MDRange.o
-        OBJ_SERIAL += TestSerial_Task.o
+        OBJ_SERIAL += TestSerial_Task.o TestSerial_WorkGraph.o
 	
 	TARGETS += KokkosCore_UnitTest_Serial
 
diff --git a/lib/kokkos/core/unit_test/TestAggregate.hpp b/lib/kokkos/core/unit_test/TestAggregate.hpp
index 6896a27bfb..87440c36be 100644
--- a/lib/kokkos/core/unit_test/TestAggregate.hpp
+++ b/lib/kokkos/core/unit_test/TestAggregate.hpp
@@ -58,7 +58,7 @@ template< class DeviceType >
 void TestViewAggregate()
 {
   typedef Kokkos::Array< double, 32 >  value_type;
-  typedef Kokkos::Experimental::Impl::ViewDataAnalysis< value_type *, Kokkos::LayoutLeft, value_type > analysis_1d;
+  typedef Kokkos::Impl::ViewDataAnalysis< value_type *, Kokkos::LayoutLeft, value_type > analysis_1d;
 
   static_assert( std::is_same< typename analysis_1d::specialize, Kokkos::Array<> >::value, "" );
 
diff --git a/lib/kokkos/core/unit_test/TestDefaultDeviceTypeInit.hpp b/lib/kokkos/core/unit_test/TestDefaultDeviceTypeInit.hpp
index 401da58a58..68864c8d66 100644
--- a/lib/kokkos/core/unit_test/TestDefaultDeviceTypeInit.hpp
+++ b/lib/kokkos/core/unit_test/TestDefaultDeviceTypeInit.hpp
@@ -186,6 +186,21 @@ void check_correct_initialization( const Kokkos::InitArguments & argstruct ) {
   // Figure out the number of threads the HostSpace ExecutionSpace should have initialized to.
   int expected_nthreads = argstruct.num_threads;
 
+#ifdef KOKKOS_ENABLE_OPENMP
+  if ( std::is_same< Kokkos::HostSpace::execution_space, Kokkos::OpenMP >::value ) {
+    // use openmp default num threads
+    if ( expected_nthreads < 0 || ( expected_nthreads == 0 && !Kokkos::hwloc::available() ) ) {
+      expected_nthreads = omp_get_max_threads();
+    }
+    // use hwloc if available
+    else if ( expected_nthreads == 0 && Kokkos::hwloc::available() ) {
+      expected_nthreads = Kokkos::hwloc::get_available_numa_count()
+                        * Kokkos::hwloc::get_available_cores_per_numa()
+                        * Kokkos::hwloc::get_available_threads_per_core();
+    }
+  }
+#endif
+
   if ( expected_nthreads < 1 ) {
     if ( Kokkos::hwloc::available() ) {
       expected_nthreads = Kokkos::hwloc::get_available_numa_count()
@@ -193,12 +208,6 @@ void check_correct_initialization( const Kokkos::InitArguments & argstruct ) {
                         * Kokkos::hwloc::get_available_threads_per_core();
     }
     else {
-#ifdef KOKKOS_ENABLE_OPENMP
-      if ( std::is_same< Kokkos::HostSpace::execution_space, Kokkos::OpenMP >::value ) {
-        expected_nthreads = omp_get_max_threads();
-      }
-      else
-#endif
         expected_nthreads = 1;
     }
 
diff --git a/lib/kokkos/core/unit_test/TestMDRange.hpp b/lib/kokkos/core/unit_test/TestMDRange.hpp
index 091591bcbf..f579ddf02c 100644
--- a/lib/kokkos/core/unit_test/TestMDRange.hpp
+++ b/lib/kokkos/core/unit_test/TestMDRange.hpp
@@ -51,6 +51,180 @@ namespace Test {
 
 namespace {
 
+template 
+struct TestMDRange_ReduceArray_2D {
+
+  using DataType       = int;
+  using ViewType_2     = typename Kokkos::View< DataType**, ExecSpace >;
+  using HostViewType_2 = typename ViewType_2::HostMirror;
+
+  ViewType_2 input_view;
+
+  using scalar_type = double;
+  using value_type = scalar_type[];
+  const unsigned value_count;
+
+  TestMDRange_ReduceArray_2D( const int N0, const int N1, const unsigned array_size ) 
+    : input_view( "input_view", N0, N1 ) 
+    , value_count( array_size )
+  {}
+
+  KOKKOS_INLINE_FUNCTION
+  void init( scalar_type dst[] ) const
+  {
+    for ( unsigned i = 0; i < value_count; ++i ) {
+      dst[i] = 0.0;
+    }
+  }
+
+  KOKKOS_INLINE_FUNCTION
+  void join( volatile scalar_type dst[],
+             const volatile scalar_type src[] ) const
+  {
+    for ( unsigned i = 0; i < value_count; ++i ) {
+        dst[i] += src[i];
+    }
+  }
+
+
+  KOKKOS_INLINE_FUNCTION
+  void operator()( const int i, const int j ) const
+  {
+    input_view( i, j ) = 1;
+  }
+
+  KOKKOS_INLINE_FUNCTION
+  void operator()( const int i, const int j, value_type lsum ) const
+  {
+    lsum[0] += input_view( i, j ) * 2; //+=6 each time if InitTag => N0*N1*6
+    lsum[1] += input_view( i, j ) ;    //+=3 each time if InitTag => N0*N1*3
+  }
+
+  // tagged operators
+  struct InitTag {};
+  KOKKOS_INLINE_FUNCTION
+  void operator()( const InitTag &, const int i, const int j ) const
+  {
+    input_view( i, j ) = 3;
+  }
+
+  static void test_arrayreduce2( const int N0, const int N1 )
+  {
+    using namespace Kokkos::Experimental;
+
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<2>, Kokkos::IndexType, InitTag > range_type_init;
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<2>, Kokkos::IndexType > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type_init range_init( point_type{ { 0, 0 } }, point_type{ { N0, N1 } }, tile_type{ { 3, 3 } } );
+      range_type range( point_type{ { 0, 0 } }, point_type{ { N0, N1 } }, tile_type{ { 3, 3 } } );
+
+      const unsigned array_size = 2;
+
+      TestMDRange_ReduceArray_2D functor( N0, N1, array_size );
+
+      parallel_for( range_init, functor ); // Init the view to 3's
+
+      double sums[ array_size ];
+      parallel_reduce( range, functor, sums );
+
+      // Check output
+      //printf("Array Reduce result. N0 = %d  N1 = %d  N0*N1 = %d  sums[0] = %lf  sums[1] = %lf \n", N0, N1, N0*N1, sums[0], sums[1]);
+
+      ASSERT_EQ( sums[0], 6 * N0 * N1 );
+      ASSERT_EQ( sums[1], 3 * N0 * N1 );
+    }
+  }
+};
+
+template 
+struct TestMDRange_ReduceArray_3D {
+
+  using DataType       = int;
+  using ViewType_3     = typename Kokkos::View< DataType***, ExecSpace >;
+  using HostViewType_3 = typename ViewType_3::HostMirror;
+
+  ViewType_3 input_view;
+
+  using scalar_type = double;
+  using value_type = scalar_type[];
+  const unsigned value_count;
+
+  TestMDRange_ReduceArray_3D( const int N0, const int N1, const int N2, const unsigned array_size ) 
+    : input_view( "input_view", N0, N1, N2 ) 
+    , value_count( array_size )
+  {}
+
+  KOKKOS_INLINE_FUNCTION
+  void init( scalar_type dst[] ) const
+  {
+    for ( unsigned i = 0; i < value_count; ++i ) {
+      dst[i] = 0.0;
+    }
+  }
+
+  KOKKOS_INLINE_FUNCTION
+  void join( volatile scalar_type dst[],
+             const volatile scalar_type src[] ) const
+  {
+    for ( unsigned i = 0; i < value_count; ++i ) {
+        dst[i] += src[i];
+    }
+  }
+
+
+  KOKKOS_INLINE_FUNCTION
+  void operator()( const int i, const int j, const int k ) const
+  {
+    input_view( i, j, k ) = 1;
+  }
+
+  KOKKOS_INLINE_FUNCTION
+  void operator()( const int i, const int j, const int k, value_type lsum ) const
+  {
+    lsum[0] += input_view( i, j, k ) * 2; //+=6 each time if InitTag => N0*N1*N2*6
+    lsum[1] += input_view( i, j, k ) ;    //+=3 each time if InitTag => N0*N1*N2*3
+  }
+
+  // tagged operators
+  struct InitTag {};
+  KOKKOS_INLINE_FUNCTION
+  void operator()( const InitTag &, const int i, const int j, const int k ) const
+  {
+    input_view( i, j, k ) = 3;
+  }
+
+  static void test_arrayreduce3( const int N0, const int N1, const int N2 )
+  {
+    using namespace Kokkos::Experimental;
+
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<3>, Kokkos::IndexType, InitTag > range_type_init;
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<3>, Kokkos::IndexType > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type_init range_init( point_type{ { 0, 0, 0 } }, point_type{ { N0, N1, N2 } }, tile_type{ { 3, 3, 3 } } );
+      range_type range( point_type{ { 0, 0, 0 } }, point_type{ { N0, N1, N2 } }, tile_type{ { 3, 3, 3 } } );
+
+      const unsigned array_size = 2;
+
+      TestMDRange_ReduceArray_3D functor( N0, N1, N2, array_size );
+
+      parallel_for( range_init, functor ); // Init the view to 3's
+
+      double sums[ array_size ];
+      parallel_reduce( range, functor, sums );
+
+      ASSERT_EQ( sums[0], 6 * N0 * N1 * N2 );
+      ASSERT_EQ( sums[1], 3 * N0 * N1 * N2 );
+    }
+  }
+};
+
+
 template 
 struct TestMDRange_2D {
   using DataType     = int;
@@ -58,6 +232,7 @@ struct TestMDRange_2D {
   using HostViewType = typename ViewType::HostMirror;
 
   ViewType input_view;
+  using value_type = double;
 
   TestMDRange_2D( const DataType N0, const DataType N1 ) : input_view( "input_view", N0, N1 ) {}
 
@@ -68,7 +243,7 @@ struct TestMDRange_2D {
   }
 
   KOKKOS_INLINE_FUNCTION
-  void operator()( const int i, const int j, double &lsum ) const
+  void operator()( const int i, const int j, value_type &lsum ) const
   {
     lsum += input_view( i, j ) * 2;
   }
@@ -81,6 +256,13 @@ struct TestMDRange_2D {
     input_view( i, j ) = 3;
   }
 
+  // reduction tagged operators
+  KOKKOS_INLINE_FUNCTION
+  void operator()( const InitTag &, const int i, const int j, value_type &lsum ) const
+  {
+    lsum += input_view( i, j ) * 3;
+  }
+
   static void test_reduce2( const int N0, const int N1 )
   {
     using namespace Kokkos::Experimental;
@@ -94,13 +276,85 @@ struct TestMDRange_2D {
 
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 );
     }
 
+    // Test with reducers - scalar
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<2>, Kokkos::IndexType > range_type;
+      range_type range( {{ 0, 0 }}, {{ N0, N1 }}, {{ 3, 3 }} );
+
+      TestMDRange_2D functor( N0, N1 );
+
+      parallel_for( range, functor );
+
+      value_type sum = 0.0;
+      Kokkos::Experimental::Sum< value_type > reducer_scalar( sum );
+
+      parallel_reduce( range, functor, reducer_scalar );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 );
+    }
+    // Test with reducers - scalar view
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<2>, Kokkos::IndexType > range_type;
+      range_type range( {{ 0, 0 }}, {{ N0, N1 }}, {{ 3, 3 }} );
+
+      TestMDRange_2D functor( N0, N1 );
+
+      parallel_for( range, functor );
+
+      value_type sum = 0.0;
+      Kokkos::View< value_type, Kokkos::HostSpace > sum_view("sum_view");
+      sum_view() = sum;
+      Kokkos::Experimental::Sum< value_type > reducer_view( sum_view );
+
+      parallel_reduce( range, functor, reducer_view);
+      sum = sum_view();
+
+      ASSERT_EQ( sum, 2 * N0 * N1 );
+    }
+
+    // Tagged operator test
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<2, Iterate::Default, Iterate::Default >, Kokkos::IndexType, InitTag > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0 } }, point_type{ { N0, N1 } }, tile_type{ { 2, 4 } } );
+
+      TestMDRange_2D functor( N0, N1 );
+
+      parallel_for( range, functor );
+
+      // check parallel_for results correct with InitTag
+      HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
+      Kokkos::deep_copy( h_view, functor.input_view );
+      int counter = 0;
+      for ( int i = 0; i < N0; ++i )
+      for ( int j = 0; j < N1; ++j )
+      {
+        if ( h_view( i, j ) != 3 ) {
+          ++counter;
+        }
+      }
+
+      if ( counter != 0 ) {
+        printf( "Defaults + InitTag op(): Errors in test_for3; mismatches = %d\n\n", counter );
+      }
+      ASSERT_EQ( counter, 0 );
+
+
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 9 * N0 * N1 );
+    }
+
     {
       typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<2, Iterate::Default, Iterate::Default>, Kokkos::IndexType > range_type;
       typedef typename range_type::tile_type tile_type;
@@ -110,9 +364,9 @@ struct TestMDRange_2D {
 
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 );
     }
@@ -126,9 +380,9 @@ struct TestMDRange_2D {
 
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 );
     }
@@ -142,9 +396,9 @@ struct TestMDRange_2D {
 
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 );
     }
@@ -158,9 +412,9 @@ struct TestMDRange_2D {
 
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 );
     }
@@ -174,9 +428,9 @@ struct TestMDRange_2D {
 
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 );
     }
@@ -194,7 +448,7 @@ struct TestMDRange_2D {
       range_type range( point_type{ { 0, 0 } }, point_type{ { N0, N1 } }, tile_type{ { 3, 3 } } );
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -223,7 +477,7 @@ struct TestMDRange_2D {
       range_type range( point_type{ { 0, 0 } }, point_type{ { N0, N1 } }, tile_type{ { 3, 3 } } );
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -251,7 +505,7 @@ struct TestMDRange_2D {
       range_type range( point_type{ { 0, 0 } }, point_type{ { N0, N1 } } );
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -280,7 +534,7 @@ struct TestMDRange_2D {
       range_type range( point_type{ { 0, 0 } }, point_type{ { N0, N1 } }, tile_type{ { 3, 3 } } );
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -309,7 +563,7 @@ struct TestMDRange_2D {
       range_type range( point_type{ { 0, 0 } }, point_type{ { N0, N1 } }, tile_type{ { 4, 4 } } );
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -338,7 +592,7 @@ struct TestMDRange_2D {
       range_type range( point_type{ { 0, 0 } }, point_type{ { N0, N1 } }, tile_type{ { 3, 3 } } );
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -367,7 +621,7 @@ struct TestMDRange_2D {
       range_type range( point_type{ { 0, 0 } }, point_type{ { N0, N1 } }, tile_type{ { 7, 7 } } );
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -396,7 +650,7 @@ struct TestMDRange_2D {
       range_type range( point_type{ { 0, 0 } }, point_type{ { N0, N1 } }, tile_type{ { 16, 16 } } );
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -425,7 +679,7 @@ struct TestMDRange_2D {
       range_type range( point_type{ { 0, 0 } }, point_type{ { N0, N1 } }, tile_type{ { 5, 16 } } );
       TestMDRange_2D functor( N0, N1 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -455,6 +709,7 @@ struct TestMDRange_3D {
   using HostViewType = typename ViewType::HostMirror;
 
   ViewType input_view;
+  using value_type = double;
 
   TestMDRange_3D( const DataType N0, const DataType N1, const DataType N2 ) : input_view( "input_view", N0, N1, N2 ) {}
 
@@ -478,6 +733,13 @@ struct TestMDRange_3D {
     input_view( i, j, k ) = 3;
   }
 
+  // reduction tagged operators
+  KOKKOS_INLINE_FUNCTION
+  void operator()( const InitTag &, const int i, const int j, const int k, value_type &lsum ) const
+  {
+    lsum += input_view( i, j, k ) * 3;
+  }
+
   static void test_reduce3( const int N0, const int N1, const int N2 )
   {
     using namespace Kokkos::Experimental;
@@ -491,13 +753,86 @@ struct TestMDRange_3D {
 
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 * N2 );
     }
 
+    // Test with reducers - scalar
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<3>, Kokkos::IndexType > range_type;
+      range_type range( {{ 0, 0, 0 }}, {{ N0, N1, N2 }}, {{ 3, 3, 3 }} );
+
+      TestMDRange_3D functor( N0, N1, N2 );
+
+      parallel_for( range, functor );
+
+      value_type sum = 0.0;
+      Kokkos::Experimental::Sum< value_type > reducer_scalar( sum );
+
+      parallel_reduce( range, functor, reducer_scalar );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 );
+    }
+    // Test with reducers - scalar view
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<3>, Kokkos::IndexType > range_type;
+      range_type range( {{ 0, 0, 0 }}, {{ N0, N1, N2 }}, {{ 3, 3, 3 }} );
+
+      TestMDRange_3D functor( N0, N1, N2 );
+
+      parallel_for( range, functor );
+
+      value_type sum = 0.0;
+      Kokkos::View< value_type, Kokkos::HostSpace > sum_view("sum_view");
+      sum_view() = sum;
+      Kokkos::Experimental::Sum< value_type > reducer_view( sum_view );
+
+      parallel_reduce( range, functor, reducer_view);
+      sum = sum_view();
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 );
+    }
+
+    // Tagged operator test
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<3, Iterate::Default, Iterate::Default >, Kokkos::IndexType, InitTag > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0 } }, point_type{ { N0, N1, N2 } }, tile_type{ { 2, 4, 6 } } );
+
+      TestMDRange_3D functor( N0, N1, N2 );
+
+      parallel_for( range, functor );
+
+      // check parallel_for results correct with InitTag
+      HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
+      Kokkos::deep_copy( h_view, functor.input_view );
+      int counter = 0;
+      for ( int i = 0; i < N0; ++i )
+      for ( int j = 0; j < N1; ++j )
+      for ( int k = 0; k < N2; ++k )
+      {
+        if ( h_view( i, j, k ) != 3 ) {
+          ++counter;
+        }
+      }
+
+      if ( counter != 0 ) {
+        printf( "Defaults + InitTag op(): Errors in test_for3; mismatches = %d\n\n", counter );
+      }
+      ASSERT_EQ( counter, 0 );
+
+
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 9 * N0 * N1 * N2 );
+    }
+
     {
       typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<3, Iterate::Default, Iterate::Default >, Kokkos::IndexType > range_type;
       typedef typename range_type::tile_type tile_type;
@@ -507,9 +842,9 @@ struct TestMDRange_3D {
 
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 * N2 );
     }
@@ -523,9 +858,9 @@ struct TestMDRange_3D {
 
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 * N2 );
     }
@@ -539,9 +874,9 @@ struct TestMDRange_3D {
 
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 * N2 );
     }
@@ -555,9 +890,9 @@ struct TestMDRange_3D {
 
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 * N2 );
     }
@@ -571,9 +906,9 @@ struct TestMDRange_3D {
 
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
       double sum = 0.0;
-      md_parallel_reduce( range, functor, sum );
+      parallel_reduce( range, functor, sum );
 
       ASSERT_EQ( sum, 2 * N0 * N1 * N2 );
     }
@@ -590,7 +925,7 @@ struct TestMDRange_3D {
       range_type range( point_type{ { 0, 0, 0 } }, point_type{ { N0, N1, N2 } } );
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -620,7 +955,7 @@ struct TestMDRange_3D {
       range_type range( point_type{ { 0, 0, 0 } }, point_type{ { N0, N1, N2 } }, tile_type{ { 3, 3, 3 } } );
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -651,7 +986,7 @@ struct TestMDRange_3D {
 
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -681,7 +1016,7 @@ struct TestMDRange_3D {
       range_type range( point_type{ { 0, 0, 0 } }, point_type{ { N0, N1, N2 } }, tile_type{ { 3, 3, 3 } } );
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -711,7 +1046,7 @@ struct TestMDRange_3D {
       range_type range( point_type{ { 0, 0, 0 } }, point_type{ { N0, N1, N2 } }, tile_type{ { 2, 4, 2 } } );
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -741,7 +1076,7 @@ struct TestMDRange_3D {
       range_type range( point_type{ { 0, 0, 0 } }, point_type{ { N0, N1, N2 } }, tile_type{ { 3, 5, 7 } } );
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -771,7 +1106,7 @@ struct TestMDRange_3D {
       range_type range( point_type{ { 0, 0, 0 } }, point_type{ { N0, N1, N2 } }, tile_type{ { 8, 8, 8 } } );
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -801,7 +1136,7 @@ struct TestMDRange_3D {
       range_type range( point_type{ { 0, 0, 0 } }, point_type{ { N0, N1, N2 } }, tile_type{ { 2, 4, 2 } } );
       TestMDRange_3D functor( N0, N1, N2 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -832,6 +1167,7 @@ struct TestMDRange_4D {
   using HostViewType = typename ViewType::HostMirror;
 
   ViewType input_view;
+  using value_type = double;
 
   TestMDRange_4D( const DataType N0, const DataType N1, const DataType N2, const DataType N3 ) : input_view( "input_view", N0, N1, N2, N3 ) {}
 
@@ -855,6 +1191,191 @@ struct TestMDRange_4D {
     input_view( i, j, k, l ) = 3;
   }
 
+  // reduction tagged operators
+  KOKKOS_INLINE_FUNCTION
+  void operator()( const InitTag &, const int i, const int j, const int k, const int l, value_type &lsum ) const
+  {
+    lsum += input_view( i, j, k, l ) * 3;
+  }
+
+  static void test_reduce4( const int N0, const int N1, const int N2, const int N3 )
+  {
+    using namespace Kokkos::Experimental;
+
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<4>, Kokkos::IndexType > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3 } }, tile_type{ { 3, 3, 3, 3 } } );
+
+      TestMDRange_4D functor( N0, N1, N2, N3 );
+
+      parallel_for( range, functor );
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 );
+    }
+
+    // Test with reducers - scalar
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<4>, Kokkos::IndexType > range_type;
+      range_type range( {{ 0, 0, 0, 0 }}, {{ N0, N1, N2, N3 }}, {{ 3, 3, 3, 3 }} );
+
+      TestMDRange_4D functor( N0, N1, N2, N3 );
+
+      parallel_for( range, functor );
+
+      value_type sum = 0.0;
+      Kokkos::Experimental::Sum< value_type > reducer_scalar( sum );
+
+      parallel_reduce( range, functor, reducer_scalar );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 );
+    }
+
+    // Test with reducers - scalar view
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<4>, Kokkos::IndexType > range_type;
+      range_type range( {{ 0, 0, 0, 0 }}, {{ N0, N1, N2, N3 }}, {{ 3, 3, 3, 3 }} );
+
+      TestMDRange_4D functor( N0, N1, N2, N3 );
+
+      parallel_for( range, functor );
+
+      value_type sum = 0.0;
+      Kokkos::View< value_type, Kokkos::HostSpace > sum_view("sum_view");
+      sum_view() = sum;
+      Kokkos::Experimental::Sum< value_type > reducer_view( sum_view );
+
+      parallel_reduce( range, functor, reducer_view);
+      sum = sum_view();
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 );
+    }
+
+    // Tagged operator test
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<4, Iterate::Default, Iterate::Default >, Kokkos::IndexType, InitTag > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3 } }, tile_type{ { 2, 4, 6, 2 } } );
+
+      TestMDRange_4D functor( N0, N1, N2, N3 );
+
+      parallel_for( range, functor );
+
+      // check parallel_for results correct with InitTag
+      HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
+      Kokkos::deep_copy( h_view, functor.input_view );
+      int counter = 0;
+      for ( int i = 0; i < N0; ++i )
+      for ( int j = 0; j < N1; ++j )
+      for ( int k = 0; k < N2; ++k )
+      for ( int l = 0; l < N3; ++l )
+      {
+        if ( h_view( i, j, k, l ) != 3 ) {
+          ++counter;
+        }
+      }
+
+      if ( counter != 0 ) {
+        printf( "Defaults + InitTag op(): Errors in test_reduce4 parallel_for init; mismatches = %d\n\n", counter );
+      }
+      ASSERT_EQ( counter, 0 );
+
+
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 9 * N0 * N1 * N2 * N3 );
+    }
+
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<4, Iterate::Default, Iterate::Default >, Kokkos::IndexType > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3 } }, tile_type{ { 2, 4, 6, 2 } } );
+
+      TestMDRange_4D functor( N0, N1, N2, N3 );
+
+      parallel_for( range, functor );
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 );
+    }
+
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<4, Iterate::Left, Iterate::Left>, Kokkos::IndexType > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3 } }, tile_type{ { 2, 4, 6, 2 } } );
+
+      TestMDRange_4D functor( N0, N1, N2, N3 );
+
+      parallel_for( range, functor );
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 );
+    }
+
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<4, Iterate::Left, Iterate::Right>, Kokkos::IndexType > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3 } }, tile_type{ { 2, 4, 6, 2 } } );
+
+      TestMDRange_4D functor( N0, N1, N2, N3 );
+
+      parallel_for( range, functor );
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 );
+    }
+
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<4, Iterate::Right, Iterate::Left>, Kokkos::IndexType > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3 } }, tile_type{ { 2, 4, 6, 2 } } );
+
+      TestMDRange_4D functor( N0, N1, N2, N3 );
+
+      parallel_for( range, functor );
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 );
+    }
+
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<4, Iterate::Right, Iterate::Right>, Kokkos::IndexType > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3 } }, tile_type{ { 2, 4, 6, 2 } } );
+
+      TestMDRange_4D functor( N0, N1, N2, N3 );
+
+      parallel_for( range, functor );
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 );
+    }
+  } // end test_reduce
+
+
+
   static void test_for4( const int N0, const int N1, const int N2, const int N3 )
   {
     using namespace Kokkos::Experimental;
@@ -866,7 +1387,7 @@ struct TestMDRange_4D {
       range_type range( point_type{ { 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3 } } );
       TestMDRange_4D functor( N0, N1, N2, N3 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -897,7 +1418,7 @@ struct TestMDRange_4D {
       range_type range( point_type{ { 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3 } }, tile_type{ { 3, 11, 3, 3 } } );
       TestMDRange_4D functor( N0, N1, N2, N3 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -929,7 +1450,7 @@ struct TestMDRange_4D {
 
       TestMDRange_4D functor( N0, N1, N2, N3 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -961,7 +1482,7 @@ struct TestMDRange_4D {
 
       TestMDRange_4D functor( N0, N1, N2, N3 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -993,7 +1514,7 @@ struct TestMDRange_4D {
 
       TestMDRange_4D functor( N0, N1, N2, N3 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1025,7 +1546,7 @@ struct TestMDRange_4D {
 
       TestMDRange_4D functor( N0, N1, N2, N3 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1057,7 +1578,7 @@ struct TestMDRange_4D {
 
       TestMDRange_4D functor( N0, N1, N2, N3 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1089,7 +1610,7 @@ struct TestMDRange_4D {
 
       TestMDRange_4D functor( N0, N1, N2, N3 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1121,6 +1642,7 @@ struct TestMDRange_5D {
   using HostViewType = typename ViewType::HostMirror;
 
   ViewType input_view;
+  using value_type = double;
 
   TestMDRange_5D( const DataType N0, const DataType N1, const DataType N2, const DataType N3, const DataType N4 ) : input_view( "input_view", N0, N1, N2, N3, N4 ) {}
 
@@ -1131,7 +1653,7 @@ struct TestMDRange_5D {
   }
 
   KOKKOS_INLINE_FUNCTION
-  void operator()( const int i, const int j, const int k, const int l, const int m, double &lsum ) const
+  void operator()( const int i, const int j, const int k, const int l, const int m, value_type &lsum ) const
   {
     lsum += input_view( i, j, k, l, m ) * 2;
   }
@@ -1144,6 +1666,110 @@ struct TestMDRange_5D {
     input_view( i, j, k, l, m ) = 3;
   }
 
+  // reduction tagged operators
+  KOKKOS_INLINE_FUNCTION
+  void operator()( const InitTag &, const int i, const int j, const int k, const int l, const int m, value_type &lsum ) const
+  {
+    lsum += input_view( i, j, k, l, m ) * 3;
+  }
+
+  static void test_reduce5( const int N0, const int N1, const int N2, const int N3, const int N4 )
+  {
+    using namespace Kokkos::Experimental;
+
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<5>, Kokkos::IndexType > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3, N4 } }, tile_type{ { 3, 3, 3, 3, 3 } } );
+
+      TestMDRange_5D functor( N0, N1, N2, N3, N4 );
+
+      parallel_for( range, functor );
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 * N4 );
+    }
+
+    // Test with reducers - scalar
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<5>, Kokkos::IndexType > range_type;
+      range_type range( {{ 0, 0, 0, 0, 0 }}, {{ N0, N1, N2, N3, N4 }}, {{ 3, 3, 3, 3, 3 }} );
+
+      TestMDRange_5D functor( N0, N1, N2, N3, N4 );
+
+      parallel_for( range, functor );
+
+      value_type sum = 0.0;
+      Kokkos::Experimental::Sum< value_type > reducer_scalar( sum );
+
+      parallel_reduce( range, functor, reducer_scalar );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 * N4 );
+    }
+
+    // Test with reducers - scalar view
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<5>, Kokkos::IndexType > range_type;
+      range_type range( {{ 0, 0, 0, 0, 0 }}, {{ N0, N1, N2, N3, N4 }}, {{ 3, 3, 3, 3, 3 }} );
+
+      TestMDRange_5D functor( N0, N1, N2, N3, N4 );
+
+      parallel_for( range, functor );
+
+      value_type sum = 0.0;
+      Kokkos::View< value_type, Kokkos::HostSpace > sum_view("sum_view");
+      sum_view() = sum;
+      Kokkos::Experimental::Sum< value_type > reducer_view( sum_view );
+
+      parallel_reduce( range, functor, reducer_view);
+      sum = sum_view();
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 * N4 );
+    }
+
+    // Tagged operator test
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<5, Iterate::Default, Iterate::Default >, Kokkos::IndexType, InitTag > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3, N4 } }, tile_type{ { 2, 4, 6, 2, 2 } } );
+
+      TestMDRange_5D functor( N0, N1, N2, N3, N4 );
+
+      parallel_for( range, functor );
+
+      // check parallel_for results correct with InitTag
+      HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
+      Kokkos::deep_copy( h_view, functor.input_view );
+      int counter = 0;
+      for ( int i = 0; i < N0; ++i )
+      for ( int j = 0; j < N1; ++j )
+      for ( int k = 0; k < N2; ++k )
+      for ( int l = 0; l < N3; ++l )
+      for ( int m = 0; m < N4; ++m )
+      {
+        if ( h_view( i, j, k, l, m ) != 3 ) {
+          ++counter;
+        }
+      }
+
+      if ( counter != 0 ) {
+        printf( "Defaults + InitTag op(): Errors in test_reduce5 parallel_for init; mismatches = %d\n\n", counter );
+      }
+      ASSERT_EQ( counter, 0 );
+
+
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 9 * N0 * N1 * N2 * N3 * N4 );
+    }
+  }
+
   static void test_for5( const int N0, const int N1, const int N2, const int N3, const int N4 )
   {
     using namespace Kokkos::Experimental;
@@ -1155,7 +1781,7 @@ struct TestMDRange_5D {
       range_type range( point_type{ { 0, 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3, N4 } } );
       TestMDRange_5D functor( N0, N1, N2, N3, N4 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1184,10 +1810,10 @@ struct TestMDRange_5D {
       typedef typename range_type::tile_type tile_type;
       typedef typename range_type::point_type point_type;
 
-      range_type range( point_type{ { 0, 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3, N4 } }, tile_type{ { 3, 3, 3, 3, 7 } } );
+      range_type range( point_type{ { 0, 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3, N4 } }, tile_type{ { 3, 3, 3, 3, 5 } } );
       TestMDRange_5D functor( N0, N1, N2, N3, N4 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1220,7 +1846,7 @@ struct TestMDRange_5D {
 
       TestMDRange_5D functor( N0, N1, N2, N3, N4 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1253,7 +1879,7 @@ struct TestMDRange_5D {
 
       TestMDRange_5D functor( N0, N1, N2, N3, N4 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1286,7 +1912,7 @@ struct TestMDRange_5D {
 
       TestMDRange_5D functor( N0, N1, N2, N3, N4 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1319,7 +1945,7 @@ struct TestMDRange_5D {
 
       TestMDRange_5D functor( N0, N1, N2, N3, N4 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1352,7 +1978,7 @@ struct TestMDRange_5D {
 
       TestMDRange_5D functor( N0, N1, N2, N3, N4 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1385,7 +2011,7 @@ struct TestMDRange_5D {
 
       TestMDRange_5D functor( N0, N1, N2, N3, N4 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1418,6 +2044,7 @@ struct TestMDRange_6D {
   using HostViewType = typename ViewType::HostMirror;
 
   ViewType input_view;
+  using value_type = double;
 
   TestMDRange_6D( const DataType N0, const DataType N1, const DataType N2, const DataType N3, const DataType N4, const DataType N5 ) : input_view( "input_view", N0, N1, N2, N3, N4, N5 ) {}
 
@@ -1428,7 +2055,7 @@ struct TestMDRange_6D {
   }
 
   KOKKOS_INLINE_FUNCTION
-  void operator()( const int i, const int j, const int k, const int l, const int m, const int n, double &lsum ) const
+  void operator()( const int i, const int j, const int k, const int l, const int m, const int n, value_type &lsum ) const
   {
     lsum += input_view( i, j, k, l, m, n ) * 2;
   }
@@ -1441,6 +2068,111 @@ struct TestMDRange_6D {
     input_view( i, j, k, l, m, n ) = 3;
   }
 
+  // reduction tagged operators
+  KOKKOS_INLINE_FUNCTION
+  void operator()( const InitTag &, const int i, const int j, const int k, const int l, const int m, const int n, value_type &lsum ) const
+  {
+    lsum += input_view( i, j, k, l, m, n ) * 3;
+  }
+
+  static void test_reduce6( const int N0, const int N1, const int N2, const int N3, const int N4, const int N5 )
+  {
+    using namespace Kokkos::Experimental;
+
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<6>, Kokkos::IndexType > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3, N4, N5 } }, tile_type{ { 3, 3, 3, 3, 3, 2 } } );
+
+      TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
+
+      parallel_for( range, functor );
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 * N4 * N5 );
+    }
+
+    // Test with reducers - scalar
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<6>, Kokkos::IndexType > range_type;
+      range_type range( {{ 0, 0, 0, 0, 0, 0 }}, {{ N0, N1, N2, N3, N4, N5 }}, {{ 3, 3, 3, 3, 3, 2 }} );
+
+      TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
+
+      parallel_for( range, functor );
+
+      value_type sum = 0.0;
+      Kokkos::Experimental::Sum< value_type > reducer_scalar( sum );
+
+      parallel_reduce( range, functor, reducer_scalar );
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 * N4 * N5 );
+    }
+
+    // Test with reducers - scalar view
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<6>, Kokkos::IndexType > range_type;
+      range_type range( {{ 0, 0, 0, 0, 0, 0 }}, {{ N0, N1, N2, N3, N4, N5 }}, {{ 3, 3, 3, 3, 3, 2 }} );
+
+      TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
+
+      parallel_for( range, functor );
+
+      value_type sum = 0.0;
+      Kokkos::View< value_type, Kokkos::HostSpace > sum_view("sum_view");
+      sum_view() = sum;
+      Kokkos::Experimental::Sum< value_type > reducer_view( sum_view );
+
+      parallel_reduce( range, functor, reducer_view);
+      sum = sum_view();
+
+      ASSERT_EQ( sum, 2 * N0 * N1 * N2 * N3 * N4 * N5 );
+    }
+
+    // Tagged operator test
+    {
+      typedef typename Kokkos::Experimental::MDRangePolicy< ExecSpace, Rank<6, Iterate::Default, Iterate::Default >, Kokkos::IndexType, InitTag > range_type;
+      typedef typename range_type::tile_type tile_type;
+      typedef typename range_type::point_type point_type;
+
+      range_type range( point_type{ { 0, 0, 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3, N4, N5 } }, tile_type{ { 2, 4, 6, 2, 2, 2 } } );
+
+      TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
+
+      parallel_for( range, functor );
+
+      // check parallel_for results correct with InitTag
+      HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
+      Kokkos::deep_copy( h_view, functor.input_view );
+      int counter = 0;
+      for ( int i = 0; i < N0; ++i )
+      for ( int j = 0; j < N1; ++j )
+      for ( int k = 0; k < N2; ++k )
+      for ( int l = 0; l < N3; ++l )
+      for ( int m = 0; m < N4; ++m )
+      for ( int n = 0; n < N5; ++n )
+      {
+        if ( h_view( i, j, k, l, m, n ) != 3 ) {
+          ++counter;
+        }
+      }
+
+      if ( counter != 0 ) {
+        printf( "Defaults + InitTag op(): Errors in test_reduce6 parallel_for init; mismatches = %d\n\n", counter );
+      }
+      ASSERT_EQ( counter, 0 );
+
+
+      double sum = 0.0;
+      parallel_reduce( range, functor, sum );
+
+      ASSERT_EQ( sum, 9 * N0 * N1 * N2 * N3 * N4 * N5 );
+    }
+  }
+
   static void test_for6( const int N0, const int N1, const int N2, const int N3, const int N4, const int N5 )
   {
     using namespace Kokkos::Experimental;
@@ -1452,7 +2184,7 @@ struct TestMDRange_6D {
       range_type range( point_type{ { 0, 0, 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3, N4, N5 } } );
       TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1485,7 +2217,7 @@ struct TestMDRange_6D {
       range_type range( point_type{ { 0, 0, 0, 0, 0, 0 } }, point_type{ { N0, N1, N2, N3, N4, N5 } }, tile_type{ { 3, 3, 3, 3, 2, 3 } } ); //tile dims 3,3,3,3,3,3 more than cuda can handle with debugging
       TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1519,7 +2251,7 @@ struct TestMDRange_6D {
 
       TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1553,7 +2285,7 @@ struct TestMDRange_6D {
 
       TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1587,7 +2319,7 @@ struct TestMDRange_6D {
 
       TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1621,7 +2353,7 @@ struct TestMDRange_6D {
 
       TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1655,7 +2387,7 @@ struct TestMDRange_6D {
 
       TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1689,7 +2421,7 @@ struct TestMDRange_6D {
 
       TestMDRange_6D functor( N0, N1, N2, N3, N4, N5 );
 
-      md_parallel_for( range, functor );
+      parallel_for( range, functor );
 
       HostViewType h_view = Kokkos::create_mirror_view( functor.input_view );
       Kokkos::deep_copy( h_view, functor.input_view );
@@ -1726,11 +2458,19 @@ TEST_F( TEST_CATEGORY , mdrange_for ) {
   TestMDRange_6D< TEST_EXECSPACE >::test_for6( 10, 10, 10, 10, 5, 5 );
 }
 
-#ifndef KOKKOS_ENABLE_CUDA
 TEST_F( TEST_CATEGORY , mdrange_reduce ) {
   TestMDRange_2D< TEST_EXECSPACE >::test_reduce2( 100, 100 );
   TestMDRange_3D< TEST_EXECSPACE >::test_reduce3( 100, 10, 100 );
+  TestMDRange_4D< TEST_EXECSPACE >::test_reduce4( 100, 10, 10, 10 );
+  TestMDRange_5D< TEST_EXECSPACE >::test_reduce5( 100, 10, 10, 10, 5 );
+  TestMDRange_6D< TEST_EXECSPACE >::test_reduce6( 100, 10, 10, 10, 5, 5 );
 }
-#endif
+
+//#ifndef KOKKOS_ENABLE_CUDA
+TEST_F( TEST_CATEGORY , mdrange_array_reduce ) {
+  TestMDRange_ReduceArray_2D< TEST_EXECSPACE >::test_arrayreduce2( 4, 5 );
+  TestMDRange_ReduceArray_3D< TEST_EXECSPACE >::test_arrayreduce3( 4, 5, 10 );
+}
+//#endif
 
 } // namespace Test
diff --git a/lib/kokkos/core/unit_test/TestMemoryPool.hpp b/lib/kokkos/core/unit_test/TestMemoryPool.hpp
index 941cd6c26d..9f708390c2 100644
--- a/lib/kokkos/core/unit_test/TestMemoryPool.hpp
+++ b/lib/kokkos/core/unit_test/TestMemoryPool.hpp
@@ -54,6 +54,96 @@
 
 namespace TestMemoryPool {
 
+template< typename MemSpace = Kokkos::HostSpace >
+void test_host_memory_pool_defaults()
+{
+  typedef typename MemSpace::execution_space   Space ;
+  typedef typename Kokkos::MemoryPool< Space > MemPool ;
+
+  {
+    const size_t MemoryCapacity = 32000 ;
+    const size_t MinBlockSize   =    64 ;
+    const size_t MaxBlockSize   =  1024 ;
+    const size_t SuperBlockSize =  4096 ;
+
+    MemPool pool( MemSpace()
+                , MemoryCapacity
+                , MinBlockSize
+                , MaxBlockSize
+                , SuperBlockSize
+                );
+
+    typename MemPool::usage_statistics stats ;
+
+    pool.get_usage_statistics( stats );
+
+    ASSERT_LE( MemoryCapacity , stats.capacity_bytes );
+    ASSERT_LE( MinBlockSize , stats.min_block_bytes );
+    ASSERT_LE( MaxBlockSize , stats.max_block_bytes );
+    ASSERT_LE( SuperBlockSize , stats.superblock_bytes );
+  }
+
+  {
+    const size_t MemoryCapacity = 10000 ;
+
+    MemPool pool( MemSpace()
+                , MemoryCapacity
+                );
+
+    typename MemPool::usage_statistics stats ;
+
+    pool.get_usage_statistics( stats );
+
+    ASSERT_LE( MemoryCapacity , stats.capacity_bytes );
+    ASSERT_LE( 64u /* default */ , stats.min_block_bytes );
+    ASSERT_LE( stats.min_block_bytes , stats.max_block_bytes );
+    ASSERT_LE( stats.max_block_bytes , stats.superblock_bytes );
+    ASSERT_LE( stats.superblock_bytes , stats.capacity_bytes );
+  }
+
+  {
+    const size_t MemoryCapacity = 10000 ;
+    const size_t MinBlockSize   =    32 ; // power of two is exact
+
+    MemPool pool( MemSpace()
+                , MemoryCapacity
+                , MinBlockSize
+                );
+
+    typename MemPool::usage_statistics stats ;
+
+    pool.get_usage_statistics( stats );
+
+    ASSERT_LE( MemoryCapacity , stats.capacity_bytes );
+    ASSERT_EQ( MinBlockSize , stats.min_block_bytes );
+    ASSERT_LE( stats.min_block_bytes , stats.max_block_bytes );
+    ASSERT_LE( stats.max_block_bytes , stats.superblock_bytes );
+    ASSERT_LE( stats.superblock_bytes , stats.capacity_bytes );
+  }
+
+  {
+    const size_t MemoryCapacity = 32000 ;
+    const size_t MinBlockSize   =    32 ; // power of two is exact
+    const size_t MaxBlockSize   =  1024 ; // power of two is exact
+
+    MemPool pool( MemSpace()
+                , MemoryCapacity
+                , MinBlockSize
+                , MaxBlockSize
+                );
+
+    typename MemPool::usage_statistics stats ;
+
+    pool.get_usage_statistics( stats );
+
+    ASSERT_LE( MemoryCapacity , stats.capacity_bytes );
+    ASSERT_EQ( MinBlockSize , stats.min_block_bytes );
+    ASSERT_EQ( MaxBlockSize , stats.max_block_bytes );
+    ASSERT_LE( stats.max_block_bytes , stats.superblock_bytes );
+    ASSERT_LE( stats.superblock_bytes , stats.capacity_bytes );
+  }
+}
+
 template< typename MemSpace = Kokkos::HostSpace >
 void test_host_memory_pool_stats()
 {
@@ -188,8 +278,8 @@ void print_memory_pool_stats
             << "  bytes reserved = " << stats.reserved_bytes << std::endl
             << "  bytes free     = " << ( stats.capacity_bytes -
                ( stats.consumed_bytes + stats.reserved_bytes ) ) << std::endl
-            << "  alloc used     = " << stats.consumed_blocks << std::endl
-            << "  alloc reserved = " << stats.reserved_blocks << std::endl
+            << "  block used     = " << stats.consumed_blocks << std::endl
+            << "  block reserved = " << stats.reserved_blocks << std::endl
             << "  super used     = " << stats.consumed_superblocks << std::endl
             << "  super reserved = " << ( stats.capacity_superblocks -
                                     stats.consumed_superblocks ) << std::endl
@@ -302,15 +392,147 @@ void test_memory_pool_v2( const bool print_statistics
 //----------------------------------------------------------------------------
 //----------------------------------------------------------------------------
 
-} // namespace TestMemoryPool {
+template< class DeviceType >
+struct TestMemoryPoolCorners {
+
+  typedef Kokkos::View< uintptr_t * , DeviceType >  ptrs_type ;
+  typedef Kokkos::MemoryPool< DeviceType >          pool_type ;
+
+  pool_type pool ;
+  ptrs_type ptrs ;
+  uint32_t  size ;
+  uint32_t  stride ;
+
+  TestMemoryPoolCorners( const pool_type & arg_pool
+                       , const ptrs_type & arg_ptrs
+                       , const uint32_t arg_base
+                       , const uint32_t arg_stride
+                       )
+    : pool( arg_pool )
+    , ptrs( arg_ptrs )
+    , size( arg_base )
+    , stride( arg_stride )
+    {}
+
+  // Specify reduction argument value_type to
+  // avoid confusion with tag-dispatch.
+
+  using value_type = long ;
+
+  KOKKOS_INLINE_FUNCTION
+  void operator()( int i , long & err ) const noexcept
+    {
+      unsigned alloc_size = size << ( i % stride );
+      if ( 0 == ptrs(i) ) {
+        ptrs(i) = (uintptr_t) pool.allocate( alloc_size );
+        if ( ptrs(i) && ! alloc_size ) { ++err ; }
+      }
+    }
+
+  struct TagDealloc {};
+
+  KOKKOS_INLINE_FUNCTION
+  void operator()( int i ) const noexcept
+    {
+      unsigned alloc_size = size << ( i % stride );
+      if ( ptrs(i) ) { pool.deallocate( (void*) ptrs(i) , alloc_size ); }
+      ptrs(i) = 0 ;
+    }
+};
+
+template< class DeviceType >
+void test_memory_pool_corners( const bool print_statistics
+                             , const bool print_superblocks )
+{
+  typedef typename DeviceType::memory_space     memory_space ;
+  typedef typename DeviceType::execution_space  execution_space ;
+  typedef Kokkos::MemoryPool< DeviceType >      pool_type ;
+  typedef TestMemoryPoolCorners< DeviceType >   functor_type ;
+  typedef typename functor_type::ptrs_type      ptrs_type ;
+
+  {
+    // superblock size 1 << 14 
+    const size_t  min_superblock_size = 1u << 14 ;
+
+    // four superblocks
+    const size_t total_alloc_size = min_superblock_size * 4 ;
+
+    // block sizes  {  64 , 128 , 256 , 512 }
+    // block counts { 256 , 128 ,  64 ,  32 }
+    const unsigned  min_block_size  = 64 ;
+    const unsigned  max_block_size  = 512 ;
+    const unsigned  num_blocks      = 480 ;
+
+    pool_type pool( memory_space()
+                  , total_alloc_size
+                  , min_block_size
+                  , max_block_size
+                  , min_superblock_size );
+
+    // Allocate one block from each superblock to lock that
+    // superblock into the block size.
+
+    ptrs_type ptrs("ptrs",num_blocks);
+
+    long err = 0 ;
+
+    Kokkos::parallel_reduce
+      ( Kokkos::RangePolicy< execution_space >(0,4)
+      , functor_type( pool , ptrs , 64 , 4 )
+      , err
+      );
+
+    if ( print_statistics || err ) {
+
+      typename pool_type::usage_statistics stats ;
+
+      pool.get_usage_statistics( stats );
+
+      print_memory_pool_stats< pool_type >( stats );
+    }
+
+    if ( print_superblocks || err ) {
+      pool.print_state( std::cout );
+    }
+
+    // Now fill remaining allocations with small size
+
+    Kokkos::parallel_reduce
+      ( Kokkos::RangePolicy< execution_space >(0,num_blocks)
+      , functor_type( pool , ptrs , 64 , 1 )
+      , err
+      );
+
+    if ( print_statistics || err ) {
+
+      typename pool_type::usage_statistics stats ;
+
+      pool.get_usage_statistics( stats );
+
+      print_memory_pool_stats< pool_type >( stats );
+    }
+
+    if ( print_superblocks || err ) {
+      pool.print_state( std::cout );
+    }
+  }
+}
+
+//----------------------------------------------------------------------------
+//----------------------------------------------------------------------------
+
+} // namespace TestMemoryPool
 
 namespace Test {
 
 TEST_F( TEST_CATEGORY, memory_pool )
 {
+  TestMemoryPool::test_host_memory_pool_defaults<>();
   TestMemoryPool::test_host_memory_pool_stats<>();
   TestMemoryPool::test_memory_pool_v2< TEST_EXECSPACE >(false,false);
+  TestMemoryPool::test_memory_pool_corners< TEST_EXECSPACE >(false,false);
 }
+
 }
 
 #endif
diff --git a/lib/kokkos/core/unit_test/TestRange.hpp b/lib/kokkos/core/unit_test/TestRange.hpp
index f55574761b..3cea1ad4a0 100644
--- a/lib/kokkos/core/unit_test/TestRange.hpp
+++ b/lib/kokkos/core/unit_test/TestRange.hpp
@@ -72,8 +72,33 @@ struct TestRange {
     typename view_type::HostMirror host_flags = Kokkos::create_mirror_view( m_flags );
 
     Kokkos::parallel_for( Kokkos::RangePolicy< ExecSpace, ScheduleType >( 0, N ), *this );
+
+#if defined(KOKKOS_ENABLE_PROFILING)
+    {
+      typedef TestRange< ExecSpace, ScheduleType > ThisType;
+      std::string label("parallel_for");
+      Kokkos::Impl::ParallelConstructName< ThisType, void> pcn(label);
+      ASSERT_EQ( pcn.get(), label );
+      std::string empty_label("");
+      Kokkos::Impl::ParallelConstructName< ThisType, void> empty_pcn(empty_label);
+      ASSERT_EQ( empty_pcn.get(), typeid(ThisType).name() );
+    }
+#endif
+
     Kokkos::parallel_for( Kokkos::RangePolicy< ExecSpace, ScheduleType, VerifyInitTag >( 0, N ), *this );
 
+#if defined(KOKKOS_ENABLE_PROFILING)
+    {
+      typedef TestRange< ExecSpace, ScheduleType > ThisType;
+      std::string label("parallel_for");
+      Kokkos::Impl::ParallelConstructName< ThisType, VerifyInitTag> pcn(label);
+      ASSERT_EQ( pcn.get(), label );
+      std::string empty_label("");
+      Kokkos::Impl::ParallelConstructName< ThisType, VerifyInitTag> empty_pcn(empty_label);
+      ASSERT_EQ( empty_pcn.get(), std::string(typeid(ThisType).name()) + "/" + typeid(VerifyInitTag).name() );
+    }
+#endif
+
     Kokkos::deep_copy( host_flags, m_flags );
 
     int error_count = 0;
diff --git a/lib/kokkos/core/unit_test/TestResize.hpp b/lib/kokkos/core/unit_test/TestResize.hpp
new file mode 100644
index 0000000000..aaf0422b19
--- /dev/null
+++ b/lib/kokkos/core/unit_test/TestResize.hpp
@@ -0,0 +1,140 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+#ifndef TESTVIEWSUBVIEW_HPP_
+#define TESTVIEWSUBVIEW_HPP_
+
+#include 
+#include 
+
+namespace TestViewResize {
+
+template
+void testResize ()
+{
+  const int sizes[8] = {2, 3, 4, 5, 6, 7, 8, 9};
+
+  // Check #904 fix (no reallocation if dimensions didn't change).
+  {
+    typedef Kokkos::View view_type;
+    view_type view_1d ("view_1d", sizes[0]);
+    const int* oldPointer = view_1d.data ();
+    EXPECT_TRUE( oldPointer != NULL );
+    Kokkos::resize (view_1d, sizes[0]);
+    const int* newPointer = view_1d.data ();
+    EXPECT_TRUE( oldPointer == newPointer );
+  }
+  {
+    typedef Kokkos::View view_type;
+    view_type view_2d ("view_2d", sizes[0], sizes[1]);
+    const int* oldPointer = view_2d.data ();
+    EXPECT_TRUE( oldPointer != NULL );
+    Kokkos::resize (view_2d, sizes[0], sizes[1]);
+    const int* newPointer = view_2d.data ();
+    EXPECT_TRUE( oldPointer == newPointer );
+  }
+  {
+    typedef Kokkos::View view_type;
+    view_type view_3d ("view_3d", sizes[0], sizes[1], sizes[2]);
+    const int* oldPointer = view_3d.data ();
+    EXPECT_TRUE( oldPointer != NULL );
+    Kokkos::resize (view_3d, sizes[0], sizes[1], sizes[2]);
+    const int* newPointer = view_3d.data ();
+    EXPECT_TRUE( oldPointer == newPointer );
+  }
+  {
+    typedef Kokkos::View view_type;
+    view_type view_4d ("view_4d", sizes[0], sizes[1], sizes[2], sizes[3]);
+    const int* oldPointer = view_4d.data ();
+    EXPECT_TRUE( oldPointer != NULL );
+    Kokkos::resize (view_4d, sizes[0], sizes[1], sizes[2], sizes[3]);
+    const int* newPointer = view_4d.data ();
+    EXPECT_TRUE( oldPointer == newPointer );
+  }
+  {
+    typedef Kokkos::View view_type;
+    view_type view_5d ("view_5d", sizes[0], sizes[1], sizes[2], sizes[3],
+                       sizes[4]);
+    const int* oldPointer = view_5d.data ();
+    EXPECT_TRUE( oldPointer != NULL );
+    Kokkos::resize (view_5d, sizes[0], sizes[1], sizes[2], sizes[3], sizes[4]);
+    const int* newPointer = view_5d.data ();
+    EXPECT_TRUE( oldPointer == newPointer );
+  }
+  {
+    typedef Kokkos::View view_type;
+    view_type view_6d ("view_6d", sizes[0], sizes[1], sizes[2], sizes[3],
+                       sizes[4], sizes[5]);
+    const int* oldPointer = view_6d.data ();
+    EXPECT_TRUE( oldPointer != NULL );
+    Kokkos::resize (view_6d, sizes[0], sizes[1], sizes[2], sizes[3], sizes[4],
+                    sizes[5]);
+    const int* newPointer = view_6d.data ();
+    EXPECT_TRUE( oldPointer == newPointer );
+  }
+  {
+    typedef Kokkos::View view_type;
+    view_type view_7d ("view_7d", sizes[0], sizes[1], sizes[2], sizes[3],
+                       sizes[4], sizes[5], sizes[6]);
+    const int* oldPointer = view_7d.data ();
+    EXPECT_TRUE( oldPointer != NULL );
+    Kokkos::resize (view_7d, sizes[0], sizes[1], sizes[2], sizes[3], sizes[4],
+                    sizes[5], sizes[6]);
+    const int* newPointer = view_7d.data ();
+    EXPECT_TRUE( oldPointer == newPointer );
+  }
+  {
+    typedef Kokkos::View view_type;
+    view_type view_8d ("view_8d", sizes[0], sizes[1], sizes[2], sizes[3],
+                       sizes[4], sizes[5], sizes[6], sizes[7]);
+    const int* oldPointer = view_8d.data ();
+    EXPECT_TRUE( oldPointer != NULL );
+    Kokkos::resize (view_8d, sizes[0], sizes[1], sizes[2], sizes[3], sizes[4],
+                    sizes[5], sizes[6], sizes[7]);
+    const int* newPointer = view_8d.data ();
+    EXPECT_TRUE( oldPointer == newPointer );
+  }
+}
+
+} // namespace TestViewSubview
+
+#endif // TESTVIEWSUBVIEW_HPP_
diff --git a/lib/kokkos/core/unit_test/TestTaskScheduler.hpp b/lib/kokkos/core/unit_test/TestTaskScheduler.hpp
index 3a88475620..4e66543857 100644
--- a/lib/kokkos/core/unit_test/TestTaskScheduler.hpp
+++ b/lib/kokkos/core/unit_test/TestTaskScheduler.hpp
@@ -250,13 +250,21 @@ struct TestTaskDependence {
     const int n = CHUNK < m_count ? CHUNK : m_count;
 
     if ( 1 < m_count ) {
-      future_type f[ CHUNK ];
+      // Test use of memory pool for temporary allocation:
+
+      // Raw allocation:
+      future_type * const f =
+        (future_type *) m_sched.memory()->allocate( sizeof(future_type) * n );
+
+      // In-place construction:
+      for ( int i = 0; i < n; ++i ) new(f+i) future_type();
 
       const int inc = ( m_count + n - 1 ) / n;
 
       for ( int i = 0; i < n; ++i ) {
         long begin = i * inc;
         long count = begin + inc < m_count ? inc : m_count - begin;
+
         f[i] = Kokkos::task_spawn( Kokkos::TaskSingle( m_sched )
                                  , TestTaskDependence( count, m_sched, m_accum ) );
       }
@@ -264,6 +272,12 @@ struct TestTaskDependence {
       m_count = 0;
 
       Kokkos::respawn( this, Kokkos::when_all( f, n ) );
+
+      // In-place destruction to release future:
+      for ( int i = 0; i < n; ++i ) (f+i)->~future_type();
+
+      // Raw deallocation:
+      m_sched.memory()->deallocate( f , sizeof(future_type) * n );
     }
     else if ( 1 == m_count ) {
       Kokkos::atomic_increment( & m_accum() );
@@ -641,19 +655,12 @@ namespace Test {
 
 TEST_F( TEST_CATEGORY, task_fib )
 {
-  const int N = 24 ; // 25 triggers tbd bug on Cuda/Pascal
+  const int N = 27 ;
   for ( int i = 0; i < N; ++i ) {
-    TestTaskScheduler::TestFib< TEST_EXECSPACE >::run( i , ( i + 1 ) * ( i + 1 ) * 10000 );
+    TestTaskScheduler::TestFib< TEST_EXECSPACE >::run( i , ( i + 1 ) * ( i + 1 ) * 2000 );
   }
 }
 
-#if defined(KOKKOS_ARCH_MAXWELL) || defined(KOKKOS_ARCH_PASCAL)
-  // TODO: Resolve bug in task DAG for Pascal
-  #define KOKKOS_IMPL_DISABLE_UNIT_TEST_TASK_DAG_PASCAL
-#endif
-
-#ifndef KOKKOS_IMPL_DISABLE_UNIT_TEST_TASK_DAG_PASCAL
-
 TEST_F( TEST_CATEGORY, task_depend )
 {
   for ( int i = 0; i < 25; ++i ) {
@@ -667,11 +674,8 @@ TEST_F( TEST_CATEGORY, task_team )
   //TestTaskScheduler::TestTaskTeamValue< TEST_EXECSPACE >::run( 1000 ); // Put back after testing.
 }
 
-#else //ndef KOKKOS_IMPL_DISABLE_UNIT_TEST_TASK_DAG_PASCAL
-#undef KOKKOS_IMPL_DISABLE_UNIT_TEST_TASK_DAG_PASCAL
-#endif //ndef KOKKOS_IMPL_DISABLE_UNIT_TEST_TASK_DAG_PASCAL
-
 }
+
 #endif // #if defined( KOKKOS_ENABLE_TASKDAG )
 #endif // #ifndef KOKKOS_UNITTEST_TASKSCHEDULER_HPP
 
diff --git a/lib/kokkos/core/unit_test/TestTeamVector.hpp b/lib/kokkos/core/unit_test/TestTeamVector.hpp
index e9e2f7548a..7f4663d0f9 100644
--- a/lib/kokkos/core/unit_test/TestTeamVector.hpp
+++ b/lib/kokkos/core/unit_test/TestTeamVector.hpp
@@ -838,6 +838,18 @@ public:
     }, result );
 
     const ScalarType solution = (ScalarType) nrows * (ScalarType) ncols;
+
+    if ( int64_t(solution) != int64_t(result) ) {
+      printf( "  TestTripleNestedReduce failed solution(%ld) != result(%ld), nrows(%d) ncols(%d) league_size(%d) team_size(%d)\n"
+            , int64_t(solution)
+            , int64_t(result)
+            , int32_t(nrows)
+            , int32_t(ncols)
+            , int32_t(nrows/chunk_size)
+            , int32_t(team_size)
+            );
+    }
+
     ASSERT_EQ( solution, result );
   }
 };
diff --git a/lib/kokkos/core/unit_test/TestTile.hpp b/lib/kokkos/core/unit_test/TestTile.hpp
index 8f57dfea75..f15667322f 100644
--- a/lib/kokkos/core/unit_test/TestTile.hpp
+++ b/lib/kokkos/core/unit_test/TestTile.hpp
@@ -94,7 +94,7 @@ struct ReduceTileErrors
     const size_t jtile = iwork / tile_dim0;
 
     if ( jtile < tile_dim1 ) {
-      tile_type tile = Kokkos::Experimental::tile_subview( m_array, itile, jtile );
+      tile_type tile = Kokkos::tile_subview( m_array, itile, jtile );
 
       if ( tile( 0, 0 ) != ptrdiff_t( ( itile + jtile * tile_dim0 ) * TileLayout::N0 * TileLayout::N1 ) ) {
         ++errors;
diff --git a/lib/kokkos/core/unit_test/TestUniqueToken.hpp b/lib/kokkos/core/unit_test/TestUniqueToken.hpp
new file mode 100644
index 0000000000..28add61a8a
--- /dev/null
+++ b/lib/kokkos/core/unit_test/TestUniqueToken.hpp
@@ -0,0 +1,138 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include 
+
+#include 
+
+namespace Test {
+
+template< class Space >
+class TestUniqueToken
+{
+public:
+  typedef typename Space::execution_space  execution_space;
+  typedef Kokkos::View< int * , execution_space > view_type ;
+
+  Kokkos::Experimental::UniqueToken< execution_space , Kokkos::Experimental::UniqueTokenScope::Global > tokens ;
+
+  view_type verify ;
+  view_type counts ;
+  view_type errors ;
+
+  KOKKOS_INLINE_FUNCTION
+  void operator()( long ) const
+  {
+    const int32_t t = tokens.acquire();
+
+    bool ok = true ;
+
+    ok = ok && 0 <= t ;
+    ok = ok && t < tokens.size();
+    ok = ok && 0 == Kokkos::atomic_fetch_add( & verify(t) , 1 );
+
+    Kokkos::atomic_fetch_add( & counts(t) , 1 );
+
+    ok = ok && 1 == Kokkos::atomic_fetch_add( & verify(t) , -1 );
+
+    if ( ! ok ) { Kokkos::atomic_fetch_add( & errors(0) , 1 ) ; }
+
+    tokens.release(t);
+  }
+
+  TestUniqueToken()
+    : tokens( execution_space() )
+    , verify( "TestUniqueTokenVerify" , tokens.size() )
+    , counts( "TestUniqueTokenCounts" , tokens.size() )
+    , errors( "TestUniqueTokenErrors" , 1 )
+    {}
+
+  static void run()
+    {
+      using policy = Kokkos::RangePolicy ;
+
+      TestUniqueToken self ;
+
+      {
+        const int duplicate = 100 ;
+        const long n = duplicate * self.tokens.size();
+
+        Kokkos::parallel_for( policy(0,n) , self );
+        Kokkos::parallel_for( policy(0,n) , self );
+        Kokkos::parallel_for( policy(0,n) , self );
+        Kokkos::fence();
+      }
+
+      typename view_type::HostMirror host_counts =
+        Kokkos::create_mirror_view( self.counts );
+
+      Kokkos::deep_copy( host_counts , self.counts );
+
+      int32_t max = 0 ;
+
+      {
+        const long n = host_counts.extent(0);
+        for ( long i = 0 ; i < n ; ++i ) {
+          if ( max < host_counts[i] ) max = host_counts[i] ;
+        }
+      }
+
+      std::cout << "TestUniqueToken max reuse = " << max << std::endl ;
+
+      typename view_type::HostMirror host_errors =
+        Kokkos::create_mirror_view( self.errors );
+
+      Kokkos::deep_copy( host_errors , self.errors );
+
+      ASSERT_EQ( host_errors(0) , 0 );
+    }
+};
+
+
+TEST_F( TEST_CATEGORY, unique_token )
+{
+  TestUniqueToken< TEST_EXECSPACE >::run();
+}
+
+} // namespace Test
+
diff --git a/lib/kokkos/core/unit_test/TestViewCtorPropEmbeddedDim.hpp b/lib/kokkos/core/unit_test/TestViewCtorPropEmbeddedDim.hpp
new file mode 100644
index 0000000000..305ddb2a1d
--- /dev/null
+++ b/lib/kokkos/core/unit_test/TestViewCtorPropEmbeddedDim.hpp
@@ -0,0 +1,160 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include 
+
+#include 
+
+#include 
+
+#include 
+#include 
+
+namespace Test {
+
+namespace {
+
+template 
+struct TestViewCtorProp_EmbeddedDim {
+
+  using ViewIntType     = typename Kokkos::View< int**, ExecSpace >;
+  using ViewDoubleType     = typename Kokkos::View< double*, ExecSpace >;
+
+  // Cuda 7.0 has issues with using a lamda in parallel_for to initialize the view - replace with this functor
+  template < class ViewType >
+  struct Functor {
+
+    ViewType v;
+
+    Functor( const ViewType & v_ ) : v(v_) {}
+
+    KOKKOS_INLINE_FUNCTION
+    void operator()( const int i ) const {
+      v(i) = i;
+    }
+
+  };
+
+
+  static void test_vcpt( const int N0, const int N1 )
+  {
+
+    // Create views to test
+    {
+      using VIT = typename TestViewCtorProp_EmbeddedDim::ViewIntType ;
+      using VDT = typename TestViewCtorProp_EmbeddedDim::ViewDoubleType ;
+
+      VIT vi1("vi1", N0, N1);
+      VDT vd1("vd1", N0);
+
+      // TEST: Test for common type between two views, one with type double, other with type int
+      // Deduce common value_type and construct a view with that type
+      {
+        // Two views
+        auto view_alloc_arg = Kokkos::common_view_alloc_prop(vi1, vd1);
+        typedef typename decltype( view_alloc_arg )::value_type                    CommonViewValueType;
+        typedef typename Kokkos::View< CommonViewValueType*, ExecSpace >  CVT;
+        typedef typename CVT::HostMirror                                           HostCVT;
+
+        // Construct View using the common type; for case of specialization, an 'embedded_dim' would be stored by view_alloc_arg
+        CVT cv1( Kokkos::view_alloc( "cv1", view_alloc_arg ), N0*N1 );
+
+        Kokkos::parallel_for( Kokkos::RangePolicy< ExecSpace >(0, N0*N1), 
+          Functor(cv1)
+        );
+
+        HostCVT hcv1 = Kokkos::create_mirror_view( cv1 );
+        Kokkos::deep_copy( hcv1, cv1 );
+
+        ASSERT_EQ( (std::is_same< CommonViewValueType, double >::value) , true ) ;
+      #if 0
+      // debug output
+      for ( int i = 0; i < N0*N1; ++i ) {
+        printf(" Output check: hcv1(%d) = %lf\n ", i, hcv1(i) );
+      }
+
+      printf( " Common value type view: %s \n", typeid( CVT() ).name() );
+      printf( " Common value type: %s \n", typeid( CommonViewValueType() ).name() );
+      if ( std::is_same< CommonViewValueType, double >::value == true ) {
+        printf("Proper common value_type\n");
+      }
+      else {
+        printf("WRONG common value_type\n");
+      }
+      // end debug output
+      #endif
+      }
+
+      {
+        // Single view
+        auto view_alloc_arg = Kokkos::common_view_alloc_prop(vi1);
+        typedef typename decltype( view_alloc_arg )::value_type                    CommonViewValueType;
+        typedef typename Kokkos::View< CommonViewValueType*, ExecSpace >  CVT;
+        typedef typename CVT::HostMirror                                           HostCVT;
+
+        // Construct View using the common type; for case of specialization, an 'embedded_dim' would be stored by view_alloc_arg
+        CVT cv1( Kokkos::view_alloc( "cv1", view_alloc_arg ), N0*N1 );
+
+        Kokkos::parallel_for( Kokkos::RangePolicy< ExecSpace >(0, N0*N1), 
+          Functor(cv1)
+        );
+
+        HostCVT hcv1 = Kokkos::create_mirror_view( cv1 );
+        Kokkos::deep_copy( hcv1, cv1 );
+
+        ASSERT_EQ( (std::is_same< CommonViewValueType, int>::value) , true ) ;
+      }
+
+    }
+
+  } // end test_vcpt
+
+}; // end struct
+
+} // namespace
+
+TEST_F( TEST_CATEGORY , viewctorprop_embedded_dim ) {
+  TestViewCtorProp_EmbeddedDim< TEST_EXECSPACE >::test_vcpt( 2, 3 );
+}
+
+} // namespace Test
diff --git a/lib/kokkos/core/unit_test/TestViewMapping_a.hpp b/lib/kokkos/core/unit_test/TestViewMapping_a.hpp
index 6830c2e049..810ae72e73 100644
--- a/lib/kokkos/core/unit_test/TestViewMapping_a.hpp
+++ b/lib/kokkos/core/unit_test/TestViewMapping_a.hpp
@@ -56,24 +56,24 @@ void test_view_mapping()
 {
   typedef typename Space::execution_space ExecSpace;
 
-  typedef Kokkos::Experimental::Impl::ViewDimension<>  dim_0;
-  typedef Kokkos::Experimental::Impl::ViewDimension< 2 > dim_s2;
-  typedef Kokkos::Experimental::Impl::ViewDimension< 2, 3 > dim_s2_s3;
-  typedef Kokkos::Experimental::Impl::ViewDimension< 2, 3, 4 > dim_s2_s3_s4;
+  typedef Kokkos::Impl::ViewDimension<>  dim_0;
+  typedef Kokkos::Impl::ViewDimension< 2 > dim_s2;
+  typedef Kokkos::Impl::ViewDimension< 2, 3 > dim_s2_s3;
+  typedef Kokkos::Impl::ViewDimension< 2, 3, 4 > dim_s2_s3_s4;
 
-  typedef Kokkos::Experimental::Impl::ViewDimension< 0 > dim_s0;
-  typedef Kokkos::Experimental::Impl::ViewDimension< 0, 3 > dim_s0_s3;
-  typedef Kokkos::Experimental::Impl::ViewDimension< 0, 3, 4 > dim_s0_s3_s4;
+  typedef Kokkos::Impl::ViewDimension< 0 > dim_s0;
+  typedef Kokkos::Impl::ViewDimension< 0, 3 > dim_s0_s3;
+  typedef Kokkos::Impl::ViewDimension< 0, 3, 4 > dim_s0_s3_s4;
 
-  typedef Kokkos::Experimental::Impl::ViewDimension< 0, 0 > dim_s0_s0;
-  typedef Kokkos::Experimental::Impl::ViewDimension< 0, 0, 4 > dim_s0_s0_s4;
+  typedef Kokkos::Impl::ViewDimension< 0, 0 > dim_s0_s0;
+  typedef Kokkos::Impl::ViewDimension< 0, 0, 4 > dim_s0_s0_s4;
 
-  typedef Kokkos::Experimental::Impl::ViewDimension< 0, 0, 0 > dim_s0_s0_s0;
-  typedef Kokkos::Experimental::Impl::ViewDimension< 0, 0, 0, 0 > dim_s0_s0_s0_s0;
-  typedef Kokkos::Experimental::Impl::ViewDimension< 0, 0, 0, 0, 0 > dim_s0_s0_s0_s0_s0;
-  typedef Kokkos::Experimental::Impl::ViewDimension< 0, 0, 0, 0, 0, 0 > dim_s0_s0_s0_s0_s0_s0;
-  typedef Kokkos::Experimental::Impl::ViewDimension< 0, 0, 0, 0, 0, 0, 0 > dim_s0_s0_s0_s0_s0_s0_s0;
-  typedef Kokkos::Experimental::Impl::ViewDimension< 0, 0, 0, 0, 0, 0, 0, 0 > dim_s0_s0_s0_s0_s0_s0_s0_s0;
+  typedef Kokkos::Impl::ViewDimension< 0, 0, 0 > dim_s0_s0_s0;
+  typedef Kokkos::Impl::ViewDimension< 0, 0, 0, 0 > dim_s0_s0_s0_s0;
+  typedef Kokkos::Impl::ViewDimension< 0, 0, 0, 0, 0 > dim_s0_s0_s0_s0_s0;
+  typedef Kokkos::Impl::ViewDimension< 0, 0, 0, 0, 0, 0 > dim_s0_s0_s0_s0_s0_s0;
+  typedef Kokkos::Impl::ViewDimension< 0, 0, 0, 0, 0, 0, 0 > dim_s0_s0_s0_s0_s0_s0_s0;
+  typedef Kokkos::Impl::ViewDimension< 0, 0, 0, 0, 0, 0, 0, 0 > dim_s0_s0_s0_s0_s0_s0_s0_s0;
 
   // Fully static dimensions should not be larger than an int.
   ASSERT_LE( sizeof( dim_0 ), sizeof( int ) );
@@ -186,12 +186,12 @@ void test_view_mapping()
 
   //----------------------------------------
 
-  typedef Kokkos::Experimental::Impl::ViewOffset< dim_s0_s0_s0, Kokkos::LayoutStride > stride_s0_s0_s0;
+  typedef Kokkos::Impl::ViewOffset< dim_s0_s0_s0, Kokkos::LayoutStride > stride_s0_s0_s0;
 
   //----------------------------------------
   // Static dimension.
   {
-    typedef Kokkos::Experimental::Impl::ViewOffset< dim_s2_s3_s4, Kokkos::LayoutLeft > left_s2_s3_s4;
+    typedef Kokkos::Impl::ViewOffset< dim_s2_s3_s4, Kokkos::LayoutLeft > left_s2_s3_s4;
 
     ASSERT_EQ( sizeof( left_s2_s3_s4 ), sizeof( dim_s2_s3_s4 ) );
 
@@ -223,7 +223,7 @@ void test_view_mapping()
   //----------------------------------------
   // Small dimension is unpadded.
   {
-    typedef Kokkos::Experimental::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutLeft > left_s0_s0_s4;
+    typedef Kokkos::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutLeft > left_s0_s0_s4;
 
     left_s0_s0_s4 dyn_off3( std::integral_constant< unsigned, sizeof( int ) >()
                           , Kokkos::LayoutLeft( 2, 3, 0, 0, 0, 0, 0, 0 ) );
@@ -275,7 +275,7 @@ void test_view_mapping()
     constexpr int N0 = 2000;
     constexpr int N1 = 300;
 
-    typedef Kokkos::Experimental::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutLeft > left_s0_s0_s4;
+    typedef Kokkos::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutLeft > left_s0_s0_s4;
 
     left_s0_s0_s4 dyn_off3( std::integral_constant< unsigned, sizeof( int ) >()
                           , Kokkos::LayoutLeft( N0, N1, 0, 0, 0, 0, 0, 0 ) );
@@ -314,7 +314,7 @@ void test_view_mapping()
   //----------------------------------------
   // Static dimension.
   {
-    typedef Kokkos::Experimental::Impl::ViewOffset< dim_s2_s3_s4, Kokkos::LayoutRight > right_s2_s3_s4;
+    typedef Kokkos::Impl::ViewOffset< dim_s2_s3_s4, Kokkos::LayoutRight > right_s2_s3_s4;
 
     ASSERT_EQ( sizeof( right_s2_s3_s4 ), sizeof( dim_s2_s3_s4 ) );
 
@@ -350,7 +350,7 @@ void test_view_mapping()
   //----------------------------------------
   // Small dimension is unpadded.
   {
-    typedef Kokkos::Experimental::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutRight > right_s0_s0_s4;
+    typedef Kokkos::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutRight > right_s0_s0_s4;
 
     right_s0_s0_s4 dyn_off3( std::integral_constant< unsigned, sizeof( int ) >()
                            , Kokkos::LayoutRight( 2, 3, 0, 0, 0, 0, 0, 0 ) );
@@ -391,7 +391,7 @@ void test_view_mapping()
     constexpr int N0 = 2000;
     constexpr int N1 = 300;
 
-    typedef Kokkos::Experimental::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutRight > right_s0_s0_s4;
+    typedef Kokkos::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutRight > right_s0_s0_s4;
 
     right_s0_s0_s4 dyn_off3( std::integral_constant< unsigned, sizeof( int ) >()
                            , Kokkos::LayoutRight( N0, N1, 0, 0, 0, 0, 0, 0 ) );
@@ -431,18 +431,18 @@ void test_view_mapping()
   // Subview.
   {
     // Mapping rank 4 to rank 3
-    typedef Kokkos::Experimental::Impl::SubviewExtents< 4, 3 > SubviewExtents;
+    typedef Kokkos::Impl::SubviewExtents< 4, 3 > SubviewExtents;
 
     constexpr int N0 = 1000;
     constexpr int N1 = 2000;
     constexpr int N2 = 3000;
     constexpr int N3 = 4000;
 
-    Kokkos::Experimental::Impl::ViewDimension< N0, N1, N2, N3 > dim;
+    Kokkos::Impl::ViewDimension< N0, N1, N2, N3 > dim;
 
     SubviewExtents tmp( dim
                       , N0 / 2
-                      , Kokkos::Experimental::ALL
+                      , Kokkos::ALL
                       , std::pair< int, int >( N2 / 4, 10 + N2 / 4 )
                       , Kokkos::pair< int, int >( N3 / 4, 20 + N3 / 4 )
                       );
@@ -469,12 +469,12 @@ void test_view_mapping()
     constexpr int sub_N1 = 200;
     constexpr int sub_N2 = 4;
 
-    typedef Kokkos::Experimental::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutLeft > left_s0_s0_s4;
+    typedef Kokkos::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutLeft > left_s0_s0_s4;
 
     left_s0_s0_s4 dyn_off3( std::integral_constant< unsigned, sizeof( int ) >()
                           , Kokkos::LayoutLeft( N0, N1, 0, 0, 0, 0, 0, 0 ) );
 
-    Kokkos::Experimental::Impl::SubviewExtents< 3, 3 >
+    Kokkos::Impl::SubviewExtents< 3, 3 >
       sub( dyn_off3.m_dim
          , Kokkos::pair< int, int >( 0, sub_N0 )
          , Kokkos::pair< int, int >( 0, sub_N1 )
@@ -509,12 +509,12 @@ void test_view_mapping()
     constexpr int sub_N1 = 200;
     constexpr int sub_N2 = 4;
 
-    typedef Kokkos::Experimental::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutRight > right_s0_s0_s4;
+    typedef Kokkos::Impl::ViewOffset< dim_s0_s0_s4, Kokkos::LayoutRight > right_s0_s0_s4;
 
     right_s0_s0_s4 dyn_off3( std::integral_constant< unsigned, sizeof( int ) >()
                            , Kokkos::LayoutRight( N0, N1, 0, 0, 0, 0, 0, 0 ) );
 
-    Kokkos::Experimental::Impl::SubviewExtents< 3, 3 >
+    Kokkos::Impl::SubviewExtents< 3, 3 >
       sub( dyn_off3.m_dim
          , Kokkos::pair< int, int >( 0, sub_N0 )
          , Kokkos::pair< int, int >( 0, sub_N1 )
@@ -544,7 +544,7 @@ void test_view_mapping()
   //----------------------------------------
   // View data analysis.
   {
-    using namespace Kokkos::Experimental::Impl;
+    using namespace Kokkos::Impl;
 
     static_assert( rank_dynamic<>::value == 0, "" );
     static_assert( rank_dynamic< 1 >::value == 0, "" );
@@ -554,7 +554,7 @@ void test_view_mapping()
   }
 
   {
-    using namespace Kokkos::Experimental::Impl;
+    using namespace Kokkos::Impl;
 
     typedef ViewArrayAnalysis< int[] >                 a_int_r1;
     typedef ViewArrayAnalysis< int**[4][5][6] >        a_int_r5;
@@ -598,7 +598,7 @@ void test_view_mapping()
   }
 
   {
-    using namespace Kokkos::Experimental::Impl;
+    using namespace Kokkos::Impl;
 
     typedef int t_i4[4];
 
@@ -616,12 +616,12 @@ void test_view_mapping()
   }
 
   {
-    using namespace Kokkos::Experimental::Impl;
+    using namespace Kokkos::Impl;
 
     typedef ViewDataAnalysis< const int[], void >  a_const_int_r1;
 
     static_assert( std::is_same< typename a_const_int_r1::specialize, void >::value, "" );
-    static_assert( std::is_same< typename a_const_int_r1::dimension, Kokkos::Experimental::Impl::ViewDimension<0> >::value, "" );
+    static_assert( std::is_same< typename a_const_int_r1::dimension, Kokkos::Impl::ViewDimension<0> >::value, "" );
 
     static_assert( std::is_same< typename a_const_int_r1::type, const int * >::value, "" );
     static_assert( std::is_same< typename a_const_int_r1::value_type, const int >::value, "" );
@@ -637,7 +637,7 @@ void test_view_mapping()
 
     static_assert( std::is_same< typename a_const_int_r3::specialize, void >::value, "" );
 
-    static_assert( std::is_same< typename a_const_int_r3::dimension, Kokkos::Experimental::Impl::ViewDimension<0, 0, 4> >::value, "" );
+    static_assert( std::is_same< typename a_const_int_r3::dimension, Kokkos::Impl::ViewDimension<0, 0, 4> >::value, "" );
 
     static_assert( std::is_same< typename a_const_int_r3::type, const int**[4] >::value, "" );
     static_assert( std::is_same< typename a_const_int_r3::value_type, const int >::value, "" );
@@ -786,7 +786,7 @@ void test_view_mapping()
   // The execution space of the memory space must be available for view data initialization.
   if ( std::is_same< ExecSpace, typename ExecSpace::memory_space::execution_space >::value ) {
 
-    using namespace Kokkos::Experimental;
+    using namespace Kokkos;
 
     typedef typename ExecSpace::memory_space  memory_space;
     typedef View< int*, memory_space >        V;
@@ -811,8 +811,8 @@ void test_view_mapping()
 
   {
     typedef Kokkos::ViewTraits< int***, Kokkos::LayoutStride, ExecSpace >           traits_t;
-    typedef Kokkos::Experimental::Impl::ViewDimension< 0, 0, 0 >                    dims_t;
-    typedef Kokkos::Experimental::Impl::ViewOffset< dims_t, Kokkos::LayoutStride >  offset_t;
+    typedef Kokkos::Impl::ViewDimension< 0, 0, 0 >                    dims_t;
+    typedef Kokkos::Impl::ViewOffset< dims_t, Kokkos::LayoutStride >  offset_t;
 
     Kokkos::LayoutStride stride;
 
@@ -836,8 +836,8 @@ void test_view_mapping()
     ASSERT_EQ( offset.span(), 60 );
     ASSERT_TRUE( offset.span_is_contiguous() );
 
-    Kokkos::Experimental::Impl::ViewMapping< traits_t, void >
-      v( Kokkos::Experimental::Impl::ViewCtorProp< int* >( (int*) 0 ), stride );
+    Kokkos::Impl::ViewMapping< traits_t, void >
+      v( Kokkos::Impl::ViewCtorProp< int* >( (int*) 0 ), stride );
   }
 
   {
@@ -849,8 +849,8 @@ void test_view_mapping()
     constexpr int N1 = 11;
 
     V a( "a", N0, N1 );
-    M b = Kokkos::Experimental::create_mirror( a );
-    M c = Kokkos::Experimental::create_mirror_view( a );
+    M b = Kokkos::create_mirror( a );
+    M c = Kokkos::create_mirror_view( a );
     M d;
 
     for ( int i0 = 0; i0 < N0; ++i0 )
@@ -859,8 +859,8 @@ void test_view_mapping()
       b( i0, i1 ) = 1 + i0 + i1 * N0;
     }
 
-    Kokkos::Experimental::deep_copy( a, b );
-    Kokkos::Experimental::deep_copy( c, a );
+    Kokkos::deep_copy( a, b );
+    Kokkos::deep_copy( c, a );
 
     for ( int i0 = 0; i0 < N0; ++i0 )
     for ( int i1 = 0; i1 < N1; ++i1 )
@@ -868,7 +868,7 @@ void test_view_mapping()
       ASSERT_EQ( b( i0, i1 ), c( i0, i1 ) );
     }
 
-    Kokkos::Experimental::resize( b, 5, 6 );
+    Kokkos::resize( b, 5, 6 );
 
     for ( int i0 = 0; i0 < 5; ++i0 )
     for ( int i1 = 0; i1 < 6; ++i1 )
@@ -878,8 +878,8 @@ void test_view_mapping()
       ASSERT_EQ( b( i0, i1 ), val );
     }
 
-    Kokkos::Experimental::realloc( c, 5, 6 );
-    Kokkos::Experimental::realloc( d, 5, 6 );
+    Kokkos::realloc( c, 5, 6 );
+    Kokkos::realloc( d, 5, 6 );
 
     ASSERT_EQ( b.dimension_0(), 5 );
     ASSERT_EQ( b.dimension_1(), 6 );
@@ -889,7 +889,7 @@ void test_view_mapping()
     ASSERT_EQ( d.dimension_1(), 6 );
 
     layout_type layout( 7, 8 );
-    Kokkos::Experimental::resize( b, layout );
+    Kokkos::resize( b, layout );
     for ( int i0 = 0; i0 < 7; ++i0 )
     for ( int i1 = 6; i1 < 8; ++i1 )
     {
@@ -909,8 +909,8 @@ void test_view_mapping()
        ASSERT_EQ( b( i0, i1 ), val );
     }
 
-    Kokkos::Experimental::realloc( c, layout );
-    Kokkos::Experimental::realloc( d, layout );
+    Kokkos::realloc( c, layout );
+    Kokkos::realloc( d, layout );
 
     ASSERT_EQ( b.dimension_0(), 7 );
     ASSERT_EQ( b.dimension_1(), 8 );
@@ -932,8 +932,8 @@ void test_view_mapping()
     const int order[] = { 1, 0 };
 
     V a( "a", Kokkos::LayoutStride::order_dimensions( 2, order, dimensions ) );
-    M b = Kokkos::Experimental::create_mirror( a );
-    M c = Kokkos::Experimental::create_mirror_view( a );
+    M b = Kokkos::create_mirror( a );
+    M c = Kokkos::create_mirror_view( a );
     M d;
 
     for ( int i0 = 0; i0 < N0; ++i0 )
@@ -942,8 +942,8 @@ void test_view_mapping()
       b( i0, i1 ) = 1 + i0 + i1 * N0;
     }
 
-    Kokkos::Experimental::deep_copy( a, b );
-    Kokkos::Experimental::deep_copy( c, a );
+    Kokkos::deep_copy( a, b );
+    Kokkos::deep_copy( c, a );
 
     for ( int i0 = 0; i0 < N0; ++i0 )
     for ( int i1 = 0; i1 < N1; ++i1 )
@@ -954,7 +954,7 @@ void test_view_mapping()
     const int dimensions2[] = { 7, 8 };
     const int order2[] = { 1, 0 };
     layout_type layout = layout_type::order_dimensions( 2, order2, dimensions2 );
-    Kokkos::Experimental::resize( b, layout );
+    Kokkos::resize( b, layout );
 
     for ( int i0 = 0; i0 < 7; ++i0 )
     for ( int i1 = 0; i1 < 8; ++i1 )
@@ -964,8 +964,8 @@ void test_view_mapping()
        ASSERT_EQ( b( i0, i1 ), val );
     }
 
-    Kokkos::Experimental::realloc( c, layout );
-    Kokkos::Experimental::realloc( d, layout );
+    Kokkos::realloc( c, layout );
+    Kokkos::realloc( d, layout );
 
     ASSERT_EQ( b.dimension_0(), 7 );
     ASSERT_EQ( b.dimension_1(), 8 );
diff --git a/lib/kokkos/core/unit_test/TestViewSubview.hpp b/lib/kokkos/core/unit_test/TestViewSubview.hpp
index e3a12e684e..106323492a 100644
--- a/lib/kokkos/core/unit_test/TestViewSubview.hpp
+++ b/lib/kokkos/core/unit_test/TestViewSubview.hpp
@@ -915,134 +915,134 @@ void test_3d_subview_5d_impl_layout() {
 
 inline
 void test_subview_legal_args_right() {
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair, int, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair, int, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t, int, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair, int, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::pair, Kokkos::pair, int, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::pair, Kokkos::pair, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t, int, int >::value ) );
 
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::pair, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, int, Kokkos::pair, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::pair, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, int, Kokkos::pair, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
 
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::pair, int, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::pair, int, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
 
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
 
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, int, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, int, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, int, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, int, Kokkos::Impl::ALL_t >::value ) );
 
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::pair, Kokkos::pair, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::pair, Kokkos::pair, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 5, 0, int, int, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
 
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::pair, Kokkos::pair, Kokkos::pair >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutRight, Kokkos::LayoutRight, 3, 3, 0, Kokkos::pair, Kokkos::pair, Kokkos::pair >::value ) );
 }
 
 inline
 void test_subview_legal_args_left() {
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair, int, int >::value ) );
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair, int, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t, int, int >::value ) );
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair, int, int >::value ) );
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::pair, Kokkos::pair, int, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t, int, int >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair, int, int >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t, int, int >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair, int, int >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::pair, Kokkos::pair, int, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t, int, int >::value ) );
 
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::pair, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, int, Kokkos::pair, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::pair, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, int, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, int, Kokkos::pair, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, int, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
 
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::pair, int, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::pair, int, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair, int >::value ) );
 
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, Kokkos::pair, int >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, Kokkos::pair, int >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t, int >::value ) );
 
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, int, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, int, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::Impl::ALL_t, Kokkos::pair, int, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::Impl::ALL_t, int, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, int, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, Kokkos::pair, Kokkos::pair, int, Kokkos::Impl::ALL_t >::value ) );
 
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::pair, Kokkos::pair, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::pair, Kokkos::pair, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 5, 0, int, int, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
 
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
-  ASSERT_EQ( 1, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair >::value ) );
-  ASSERT_EQ( 0, (  Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
-  ASSERT_EQ( 0, ( Kokkos::Experimental::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::pair, Kokkos::pair, Kokkos::pair >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::pair >::value ) );
+  ASSERT_EQ( 1, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::pair, Kokkos::Impl::ALL_t, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::Impl::ALL_t, Kokkos::pair, Kokkos::pair >::value ) );
+  ASSERT_EQ( 0, (  Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::pair, Kokkos::pair, Kokkos::Impl::ALL_t >::value ) );
+  ASSERT_EQ( 0, ( Kokkos::Impl::SubviewLegalArgsCompileTime< Kokkos::LayoutLeft, Kokkos::LayoutLeft, 3, 3, 0, Kokkos::pair, Kokkos::pair, Kokkos::pair >::value ) );
 }
 
 } // namespace Impl
diff --git a/lib/kokkos/core/unit_test/TestWorkGraph.hpp b/lib/kokkos/core/unit_test/TestWorkGraph.hpp
new file mode 100644
index 0000000000..70cf6b47c0
--- /dev/null
+++ b/lib/kokkos/core/unit_test/TestWorkGraph.hpp
@@ -0,0 +1,172 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include 
+#include 
+
+#include 
+
+namespace Test {
+
+namespace {
+
+/* This test is meant to be the WorkGraph equivalent of the Task DAG Scheduler test,
+   please see TestTaskScheduler.hpp for that test.
+   The algorithm computes the N-th fibonacci number as follows:
+    - Each "task" or "work item" computes the i-th fibonacci number
+    - If a task as (i < 2), it will record the known answer ahead of time.
+    - If a taks has (i >= 2), it will "spawn" two more tasks to compute
+      the (i - 1) and (i - 2) fibonacci numbers.
+      We do NOT do any de-duplication of these tasks.
+      De-duplication would result in only (N - 2) tasks which must be run in serial.
+      We allow duplicates both to increase the number of tasks and to increase the
+      amount of available parallelism.
+ */
+
+template< class ExecSpace >
+struct TestWorkGraph {
+
+  using MemorySpace = typename ExecSpace::memory_space;
+  using Policy = Kokkos::Experimental::WorkGraphPolicy;
+  using Graph = typename Policy::graph_type;
+  using RowMap = typename Graph::row_map_type;
+  using Entries = typename Graph::entries_type;
+  using Values = Kokkos::View;
+
+  long m_input;
+  Graph m_graph;
+  Graph m_transpose;
+  Values m_values;
+
+  TestWorkGraph(long arg_input):m_input(arg_input) {
+    form_graph();
+    transpose_crs(m_transpose, m_graph);
+  }
+
+  inline
+  long full_fibonacci( long n ) {
+    constexpr long mask = 0x03;
+    long fib[4] = { 0, 1, 1, 2 };
+    for ( long i = 2; i <= n; ++i ) {
+      fib[ i & mask ] = fib[ ( i - 1 ) & mask ] + fib[ ( i - 2 ) & mask ];
+    }
+    return fib[ n & mask ];
+  }
+
+  struct HostEntry {
+    long input;
+    std::int32_t parent;
+  };
+  std::vector form_host_graph() {
+    std::vector g;
+    g.push_back({ m_input , -1 });
+    for (std::int32_t i = 0; i < std::int32_t(g.size()); ++i) {
+      auto e = g.at(std::size_t(i));
+      if (e.input < 2) continue;
+      /* This part of the host graph formation is the equivalent of task spawning
+         in the Task DAG system. Notice how each task which is not a base case
+         spawns two more tasks, without any de-duplication */
+      g.push_back({ e.input - 1, i });
+      g.push_back({ e.input - 2, i });
+    }
+    return g;
+  }
+
+  void form_graph() {
+    auto hg = form_host_graph();
+    m_graph.row_map = RowMap("row_map", hg.size() + 1); // row map always has one more
+    m_graph.entries = Entries("entries", hg.size() - 1); // all but the first have a parent
+    m_values = Values("values", hg.size());
+    auto h_row_map = Kokkos::create_mirror_view(m_graph.row_map);
+    auto h_entries = Kokkos::create_mirror_view(m_graph.entries);
+    auto h_values = Kokkos::create_mirror_view(m_values);
+    h_row_map(0) = 0;
+    for (std::int32_t i = 0; i < std::int32_t(hg.size()); ++i) {
+      auto& e = hg.at(std::size_t(i));
+      h_row_map(i + 1) = i;
+      if (e.input < 2) {
+        h_values(i) = e.input;
+      }
+      if (e.parent == -1) continue;
+      h_entries(i - 1) = e.parent;
+    }
+    Kokkos::deep_copy(m_graph.row_map, h_row_map);
+    Kokkos::deep_copy(m_graph.entries, h_entries);
+    Kokkos::deep_copy(m_values, h_values);
+  }
+
+  KOKKOS_INLINE_FUNCTION
+  void operator()(std::int32_t i) const {
+    auto begin = m_transpose.row_map(i);
+    auto end = m_transpose.row_map(i + 1);
+    for (auto j = begin; j < end; ++j) {
+      auto k = m_transpose.entries(j);
+      m_values(i) += m_values( k );
+    }
+  }
+
+  void test_for() {
+    Kokkos::parallel_for(Policy(m_graph), *this);
+    auto h_values = Kokkos::create_mirror_view(m_values);
+    Kokkos::deep_copy(h_values, m_values);
+    ASSERT_EQ( h_values(0), full_fibonacci(m_input) );
+  }
+
+};
+
+} // anonymous namespace
+
+TEST_F( TEST_CATEGORY, DISABLED_workgraph_fib )
+{
+  #ifdef KOKKOS_IMPL_CUDA_CLANG_WORKAROUND
+  int limit = 15;
+  #else
+  int limit = 27;
+  #endif
+  for ( int i = 0; i < limit; ++i) {
+    TestWorkGraph< TEST_EXECSPACE > f(i);
+    f.test_for();
+  }
+}
+
+} // namespace Test
diff --git a/lib/kokkos/core/unit_test/UnitTestMain.cpp b/lib/kokkos/core/unit_test/UnitTestMain.cpp
index 4f52fc9567..a7dc7c4973 100644
--- a/lib/kokkos/core/unit_test/UnitTestMain.cpp
+++ b/lib/kokkos/core/unit_test/UnitTestMain.cpp
@@ -42,6 +42,7 @@
 */
 
 #include 
+#include 
 
 int main( int argc, char *argv[] ) {
   ::testing::InitGoogleTest( &argc, argv );
diff --git a/lib/kokkos/core/unit_test/UnitTestMainInit.cpp b/lib/kokkos/core/unit_test/UnitTestMainInit.cpp
index 21f851274b..62a01e9033 100644
--- a/lib/kokkos/core/unit_test/UnitTestMainInit.cpp
+++ b/lib/kokkos/core/unit_test/UnitTestMainInit.cpp
@@ -42,6 +42,8 @@
 */
 
 #include 
+#include 
+
 #include 
 
 int main( int argc, char *argv[] ) {
diff --git a/lib/kokkos/core/unit_test/cuda/TestCuda_Other.cpp b/lib/kokkos/core/unit_test/cuda/TestCuda_Other.cpp
index ba06b71192..fa6722615c 100644
--- a/lib/kokkos/core/unit_test/cuda/TestCuda_Other.cpp
+++ b/lib/kokkos/core/unit_test/cuda/TestCuda_Other.cpp
@@ -48,3 +48,5 @@
 #include
 #include
 #include
+
+#include
diff --git a/lib/kokkos/core/unit_test/cuda/TestCuda_UniqueToken.cpp b/lib/kokkos/core/unit_test/cuda/TestCuda_UniqueToken.cpp
new file mode 100644
index 0000000000..8424ae10d6
--- /dev/null
+++ b/lib/kokkos/core/unit_test/cuda/TestCuda_UniqueToken.cpp
@@ -0,0 +1,46 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include
+#include
+
diff --git a/lib/kokkos/core/unit_test/cuda/TestCuda_WorkGraph.cpp b/lib/kokkos/core/unit_test/cuda/TestCuda_WorkGraph.cpp
new file mode 100644
index 0000000000..663ca1d560
--- /dev/null
+++ b/lib/kokkos/core/unit_test/cuda/TestCuda_WorkGraph.cpp
@@ -0,0 +1,45 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include
+#include
diff --git a/lib/kokkos/core/unit_test/default/TestDefaultDeviceTypeResize.cpp b/lib/kokkos/core/unit_test/default/TestDefaultDeviceTypeResize.cpp
new file mode 100644
index 0000000000..c02905535b
--- /dev/null
+++ b/lib/kokkos/core/unit_test/default/TestDefaultDeviceTypeResize.cpp
@@ -0,0 +1,57 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include 
+#include "TestResize.hpp"
+
+namespace Test {
+
+TEST( kokkosresize, host_space_access )
+{
+  // Test with the default device type.
+  using TestViewResize::testResize;
+  typedef Kokkos::View::device_type device_type;
+  testResize ();
+}
+
+} // namespace Test
diff --git a/lib/kokkos/core/unit_test/openmp/TestOpenMP.hpp b/lib/kokkos/core/unit_test/openmp/TestOpenMP.hpp
index 2f8daf7ad7..c12574a65a 100644
--- a/lib/kokkos/core/unit_test/openmp/TestOpenMP.hpp
+++ b/lib/kokkos/core/unit_test/openmp/TestOpenMP.hpp
@@ -86,25 +86,26 @@ class openmp : public ::testing::Test {
 protected:
   static void SetUpTestCase()
   {
-    const unsigned numa_count       = Kokkos::hwloc::get_available_numa_count();
-    const unsigned cores_per_numa   = Kokkos::hwloc::get_available_cores_per_numa();
-    const unsigned threads_per_core = Kokkos::hwloc::get_available_threads_per_core();
+    int threads_count = 0;
+    #pragma omp parallel
+    {
+      #pragma omp atomic
+      ++threads_count;
+    }
 
-    const unsigned threads_count = std::max( 1u, numa_count ) *
-                                   std::max( 2u, ( cores_per_numa * threads_per_core ) / 2 );
+    if (threads_count > 3) {
+      threads_count /= 2;
+    }
 
     Kokkos::OpenMP::initialize( threads_count );
     Kokkos::print_configuration( std::cout, true );
+
     srand( 10231 );
   }
 
   static void TearDownTestCase()
   {
     Kokkos::OpenMP::finalize();
-
-    omp_set_num_threads( 1 );
-
-    ASSERT_EQ( 1, omp_get_max_threads() );
   }
 };
 
diff --git a/lib/kokkos/core/unit_test/openmp/TestOpenMP_Other.cpp b/lib/kokkos/core/unit_test/openmp/TestOpenMP_Other.cpp
index 5e9535638d..33e7402ce6 100644
--- a/lib/kokkos/core/unit_test/openmp/TestOpenMP_Other.cpp
+++ b/lib/kokkos/core/unit_test/openmp/TestOpenMP_Other.cpp
@@ -48,3 +48,93 @@
 #include
 #include
 #include
+
+#include
+
+#include 
+
+namespace Test {
+
+TEST_F( openmp, partition_master )
+{
+  using Mutex = Kokkos::Experimental::MasterLock;
+
+  Mutex mtx;
+  int errors = 0;
+
+  auto master = [&errors, &mtx](int partition_id, int num_partitions) {
+
+    const int pool_size = Kokkos::OpenMP::thread_pool_size();
+
+    {
+      std::unique_lock lock(mtx);
+      if ( Kokkos::OpenMP::in_parallel() ) {
+        ++errors;
+      }
+      if ( Kokkos::OpenMP::thread_pool_rank() != 0 ) {
+        ++errors;
+      }
+    }
+
+    {
+      int local_errors = 0;
+      Kokkos::parallel_reduce( Kokkos::RangePolicy(0,1000)
+                           , [pool_size]( const int , int & errs ) {
+          if ( Kokkos::OpenMP::thread_pool_size() != pool_size ) {
+            ++errs;
+          }
+        }
+        , local_errors
+      );
+      Kokkos::atomic_add( &errors, local_errors );
+    }
+
+    Kokkos::Experimental::UniqueToken< Kokkos::OpenMP > token;
+
+    Kokkos::View count( "",  token.size() );
+
+    Kokkos::parallel_for( Kokkos::RangePolicy(0,1000),
+        [=] ( const int ) {
+      int i = token.acquire();
+      ++count[i];
+      token.release(i);
+    });
+
+    Kokkos::View sum ("");
+    Kokkos::parallel_for( Kokkos::RangePolicy(0,token.size()),
+        [=] ( const int i ) {
+      Kokkos::atomic_add( sum.data(), count[i] );
+    });
+
+    if (sum() != 1000) {
+      Kokkos::atomic_add( &errors, 1 );
+    }
+  };
+
+  master(0,1);
+
+  ASSERT_EQ( errors, 0 );
+
+  Kokkos::OpenMP::partition_master( master );
+  ASSERT_EQ( errors, 0 );
+
+  Kokkos::OpenMP::partition_master( master, 4, 0 );
+  ASSERT_EQ( errors, 0 );
+
+  Kokkos::OpenMP::partition_master( master, 0, 4 );
+  ASSERT_EQ( errors, 0 );
+
+  Kokkos::OpenMP::partition_master( master, 2, 2 );
+  ASSERT_EQ( errors, 0 );
+
+  Kokkos::OpenMP::partition_master( master, 8, 0 );
+  ASSERT_EQ( errors, 0 );
+
+  Kokkos::OpenMP::partition_master( master, 0, 8 );
+  ASSERT_EQ( errors, 0 );
+
+  Kokkos::OpenMP::partition_master( master, 8, 8 );
+  ASSERT_EQ( errors, 0 );
+}
+
+} // namespace Test
diff --git a/lib/kokkos/core/unit_test/openmp/TestOpenMP_UniqueToken.cpp b/lib/kokkos/core/unit_test/openmp/TestOpenMP_UniqueToken.cpp
new file mode 100644
index 0000000000..143a6d9910
--- /dev/null
+++ b/lib/kokkos/core/unit_test/openmp/TestOpenMP_UniqueToken.cpp
@@ -0,0 +1,46 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include
+#include
+
diff --git a/lib/kokkos/core/unit_test/openmp/TestOpenMP_WorkGraph.cpp b/lib/kokkos/core/unit_test/openmp/TestOpenMP_WorkGraph.cpp
new file mode 100644
index 0000000000..ec6fa1653c
--- /dev/null
+++ b/lib/kokkos/core/unit_test/openmp/TestOpenMP_WorkGraph.cpp
@@ -0,0 +1,45 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include
+#include
diff --git a/lib/kokkos/core/unit_test/serial/TestSerial_Other.cpp b/lib/kokkos/core/unit_test/serial/TestSerial_Other.cpp
index a6a76a03bd..bc39b1e160 100644
--- a/lib/kokkos/core/unit_test/serial/TestSerial_Other.cpp
+++ b/lib/kokkos/core/unit_test/serial/TestSerial_Other.cpp
@@ -48,3 +48,5 @@
 #include
 #include
 #include
+
+#include
diff --git a/lib/kokkos/core/unit_test/serial/TestSerial_WorkGraph.cpp b/lib/kokkos/core/unit_test/serial/TestSerial_WorkGraph.cpp
new file mode 100644
index 0000000000..de1638de5e
--- /dev/null
+++ b/lib/kokkos/core/unit_test/serial/TestSerial_WorkGraph.cpp
@@ -0,0 +1,45 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include
+#include
diff --git a/lib/kokkos/core/unit_test/threads/TestThreads_Other.cpp b/lib/kokkos/core/unit_test/threads/TestThreads_Other.cpp
index c11155c5c0..160b37a2c8 100644
--- a/lib/kokkos/core/unit_test/threads/TestThreads_Other.cpp
+++ b/lib/kokkos/core/unit_test/threads/TestThreads_Other.cpp
@@ -48,3 +48,5 @@
 #include
 #include
 #include
+
+#include
diff --git a/lib/kokkos/core/unit_test/threads/TestThreads_WorkGraph.cpp b/lib/kokkos/core/unit_test/threads/TestThreads_WorkGraph.cpp
new file mode 100644
index 0000000000..6b7dbb26db
--- /dev/null
+++ b/lib/kokkos/core/unit_test/threads/TestThreads_WorkGraph.cpp
@@ -0,0 +1,45 @@
+/*
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+//
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+//
+// ************************************************************************
+//@HEADER
+*/
+
+#include
+#include
diff --git a/lib/kokkos/example/cmake_build/CMakeLists.txt b/lib/kokkos/example/cmake_build/CMakeLists.txt
index 4e149726ee..f92c5c6513 100644
--- a/lib/kokkos/example/cmake_build/CMakeLists.txt
+++ b/lib/kokkos/example/cmake_build/CMakeLists.txt
@@ -40,5 +40,7 @@ list(APPEND CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS} -O3)
 
 add_subdirectory(${Example_SOURCE_DIR}/../.. ${Example_BINARY_DIR}/kokkos)
 
+include_directories(${Kokkos_INCLUDE_DIRS_RET})
+
 add_executable(example cmake_example.cpp)
 target_link_libraries(example kokkos)
diff --git a/lib/kokkos/example/feint/main.cpp b/lib/kokkos/example/feint/main.cpp
index 616e584bf6..57a8f8fafb 100644
--- a/lib/kokkos/example/feint/main.cpp
+++ b/lib/kokkos/example/feint/main.cpp
@@ -69,12 +69,26 @@ int main()
 
 #if defined( KOKKOS_ENABLE_OPENMP )
   {
-    // Use 4 cores per NUMA region, unless fewer available
 
-    const unsigned use_numa_count     = Kokkos::hwloc::get_available_numa_count();
-    const unsigned use_cores_per_numa = std::min( 4u , Kokkos::hwloc::get_available_cores_per_numa() );
+    int num_threads  = 0;
+    if ( Kokkos::hwloc::available() ) {
+      // Use 4 cores per NUMA region, unless fewer available
+      const unsigned use_numa_count     = Kokkos::hwloc::get_available_numa_count();
+      const unsigned use_cores_per_numa = std::min( 4u , Kokkos::hwloc::get_available_cores_per_numa() );
+      num_threads = use_numa_count * use_cores_per_numa;
 
-    Kokkos::OpenMP::initialize( use_numa_count * use_cores_per_numa );
+    }
+    else {
+      #pragma omp parallel
+      {
+        #pragma omp atomic
+        ++num_threads;
+      }
+      num_threads = std::max(4, num_threads/4);
+    }
+
+
+    Kokkos::OpenMP::initialize( num_threads );
 
     std::cout << "feint< OpenMP , NotUsingAtomic >" << std::endl ;
     Kokkos::Example::feint< Kokkos::OpenMP , false >();
diff --git a/lib/kokkos/example/global_2_local_ids/G2L_Main.cpp b/lib/kokkos/example/global_2_local_ids/G2L_Main.cpp
index fb33aef56e..b6b8b2f5e0 100644
--- a/lib/kokkos/example/global_2_local_ids/G2L_Main.cpp
+++ b/lib/kokkos/example/global_2_local_ids/G2L_Main.cpp
@@ -138,7 +138,16 @@ int main(int argc, char *argv[])
 #endif
 
 #ifdef KOKKOS_ENABLE_OPENMP
-  Kokkos::OpenMP::initialize( threads_count );
+  int num_threads = 0;
+  #pragma omp parallel
+  {
+    #pragma omp atomic
+    ++num_threads;
+  }
+  if( num_threads > 3 ) {
+    num_threads = std::max(4, num_threads/4);
+  }
+  Kokkos::OpenMP::initialize( num_threads );
   num_errors += G2L::run_openmp(num_ids,num_find_iterations);
   Kokkos::OpenMP::finalize();
 #endif
diff --git a/lib/kokkos/example/grow_array/main.cpp b/lib/kokkos/example/grow_array/main.cpp
index e7438a9bf4..3f1d534d93 100644
--- a/lib/kokkos/example/grow_array/main.cpp
+++ b/lib/kokkos/example/grow_array/main.cpp
@@ -88,7 +88,7 @@ int main( int argc , char ** argv )
 #if defined( KOKKOS_ENABLE_OPENMP )
   {
     std::cout << "Kokkos::OpenMP" << std::endl ;
-    Kokkos::OpenMP::initialize( num_threads , use_numa , use_core );
+    Kokkos::OpenMP::initialize();
     Example::grow_array< Kokkos::OpenMP >( length_array , span_values );
     Kokkos::OpenMP::finalize();
   }
diff --git a/lib/kokkos/example/tutorial/03_simple_view/Makefile b/lib/kokkos/example/tutorial/03_simple_view/Makefile
index e716b765e7..32483a2555 100644
--- a/lib/kokkos/example/tutorial/03_simple_view/Makefile
+++ b/lib/kokkos/example/tutorial/03_simple_view/Makefile
@@ -33,6 +33,7 @@ include $(KOKKOS_PATH)/Makefile.kokkos
 
 build: $(EXE)
 
+#for unit testing only, for best preformance with OpenMP 4.0 or better
 test: $(EXE)
 	./$(EXE)
 
diff --git a/lib/kokkos/example/tutorial/Advanced_Views/Makefile b/lib/kokkos/example/tutorial/Advanced_Views/Makefile
index bc4012f68c..12ac5652e5 100644
--- a/lib/kokkos/example/tutorial/Advanced_Views/Makefile
+++ b/lib/kokkos/example/tutorial/Advanced_Views/Makefile
@@ -22,100 +22,102 @@ endif
 build:
 	mkdir -p 01_data_layouts
 	cd ./01_data_layouts; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/01_data_layouts/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/01_data_layouts/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 02_memory_traits
 	cd ./02_memory_traits; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/02_memory_traits/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/02_memory_traits/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 03_subviews
 	cd ./03_subviews; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/03_subviews/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/03_subviews/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 04_dualviews
 	cd ./04_dualviews; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/04_dualviews/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/04_dualviews/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 05_NVIDIA_UVM
 	cd ./05_NVIDIA_UVM; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/05_NVIDIA_UVM/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/05_NVIDIA_UVM/Makefile ${KOKKOS_SETTINGS}
 	#mkdir -p 06_AtomicViews
 	#cd ./06_AtomicViews; \
-	#make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/06_AtomicViews/Makefile ${KOKKOS_SETTINGS}
+	#$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/06_AtomicViews/Makefile ${KOKKOS_SETTINGS}
 	#mkdir -p 07_Overlapping_DeepCopy
 	#cd ./07_Overlapping_DeepCopy; \
-	#make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/07_Overlapping_DeepCopy/Makefile ${KOKKOS_SETTINGS}
+	#$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/07_Overlapping_DeepCopy/Makefile ${KOKKOS_SETTINGS}
 
 build-insource:
 	cd ./01_data_layouts; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./02_memory_traits; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./03_subviews; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./04_dualviews; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./05_NVIDIA_UVM; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	#cd ./06_AtomicViews; \
-	#make build -j 4 ${KOKKOS_SETTINGS}
+	#$(MAKE) build ${KOKKOS_SETTINGS}
 	#cd ./07_Overlapping_DeepCopy; \
-	#make build -j 4 ${KOKKOS_SETTINGS}
+	#$(MAKE) build ${KOKKOS_SETTINGS}
+
 test:
 	cd ./01_data_layouts; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/01_data_layouts/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/01_data_layouts/Makefile ${KOKKOS_SETTINGS}
 	cd ./02_memory_traits; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/02_memory_traits/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/02_memory_traits/Makefile ${KOKKOS_SETTINGS}
 	cd ./03_subviews; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/03_subviews/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/03_subviews/Makefile ${KOKKOS_SETTINGS}
 	cd ./04_dualviews; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/04_dualviews/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/04_dualviews/Makefile ${KOKKOS_SETTINGS}
 	cd ./05_NVIDIA_UVM; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/05_NVIDIA_UVM/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/05_NVIDIA_UVM/Makefile ${KOKKOS_SETTINGS}
 	#cd ./06_AtomicViews; \
-	#make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/06_AtomicViews/Makefile ${KOKKOS_SETTINGS}
+	#$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/06_AtomicViews/Makefile ${KOKKOS_SETTINGS}
 	#cd ./07_Overlapping_DeepCopy; \
-	#make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/07_Overlapping_DeepCopy/Makefile ${KOKKOS_SETTINGS}
+	#$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/07_Overlapping_DeepCopy/Makefile ${KOKKOS_SETTINGS}
 
 test-insource:
 	cd ./01_data_layouts; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./02_memory_traits; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./03_subviews; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./04_dualviews; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./05_NVIDIA_UVM; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	#cd ./06_AtomicViews; \
-	#make test -j 4 ${KOKKOS_SETTINGS}
+	#$(MAKE) test ${KOKKOS_SETTINGS}
 	#cd ./07_Overlapping_DeepCopy; \
-	#make test -j 4 ${KOKKOS_SETTINGS}
+	#$(MAKE) test ${KOKKOS_SETTINGS}
+
 clean:
 	cd ./01_data_layouts; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/01_data_layouts/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/01_data_layouts/Makefile ${KOKKOS_SETTINGS}
 	cd ./02_memory_traits; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/02_memory_traits/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/02_memory_traits/Makefile ${KOKKOS_SETTINGS}
 	cd ./03_subviews; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/03_subviews/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/03_subviews/Makefile ${KOKKOS_SETTINGS}
 	cd ./04_dualviews; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/04_dualviews/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/04_dualviews/Makefile ${KOKKOS_SETTINGS}
 	cd ./05_NVIDIA_UVM; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/05_NVIDIA_UVM/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/05_NVIDIA_UVM/Makefile ${KOKKOS_SETTINGS}
 	#cd ./06_AtomicViews; \
-	#make clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/06_AtomicViews/Makefile ${KOKKOS_SETTINGS}
+	#$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/06_AtomicViews/Makefile ${KOKKOS_SETTINGS}
 	#cd ./07_Overlapping_DeepCopy; \
-	#make clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/07_Overlapping_DeepCopy/Makefile ${KOKKOS_SETTINGS}
+	#$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/07_Overlapping_DeepCopy/Makefile ${KOKKOS_SETTINGS}
 
 clean-insource:
 	cd ./01_data_layouts; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./02_memory_traits; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./03_subviews; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./04_dualviews; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./05_NVIDIA_UVM; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	#cd ./06_AtomicViews; \
-	#make clean ${KOKKOS_SETTINGS}
+	#$(MAKE) clean ${KOKKOS_SETTINGS}
 	#cd ./07_Overlapping_DeepCopy; \
-	#make clean ${KOKKOS_SETTINGS}
+	#$(MAKE) clean ${KOKKOS_SETTINGS}
diff --git a/lib/kokkos/example/tutorial/Algorithms/Makefile b/lib/kokkos/example/tutorial/Algorithms/Makefile
index ad0b76f9d6..4e70ba7d97 100644
--- a/lib/kokkos/example/tutorial/Algorithms/Makefile
+++ b/lib/kokkos/example/tutorial/Algorithms/Makefile
@@ -22,22 +22,22 @@ endif
 build:
 	mkdir -p 01_random_numbers
 	cd ./01_random_numbers; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Algorithms/01_random_numbers/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Algorithms/01_random_numbers/Makefile ${KOKKOS_SETTINGS}
 
 build-insource:
 	cd ./01_random_numbers; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 test:
 	cd ./01_random_numbers; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Algorithms/01_random_numbers/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Algorithms/01_random_numbers/Makefile ${KOKKOS_SETTINGS}
 
 test-insource:
 	cd ./01_random_numbers; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 clean:
 	cd ./01_random_numbers; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Algorithms/01_random_numbers/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Algorithms/01_random_numbers/Makefile ${KOKKOS_SETTINGS}
 
 clean-insource:
 	cd ./01_random_numbers; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
diff --git a/lib/kokkos/example/tutorial/Hierarchical_Parallelism/Makefile b/lib/kokkos/example/tutorial/Hierarchical_Parallelism/Makefile
index 44fdf90f8a..4bf6d487ae 100644
--- a/lib/kokkos/example/tutorial/Hierarchical_Parallelism/Makefile
+++ b/lib/kokkos/example/tutorial/Hierarchical_Parallelism/Makefile
@@ -22,74 +22,74 @@ endif
 build:
 	mkdir -p 01_thread_teams
 	cd ./01_thread_teams; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 01_thread_teams_lambda
 	cd ./01_thread_teams_lambda; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams_lambda/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 02_nested_parallel_for
 	cd ./02_nested_parallel_for; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/02_nested_parallel_for/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/02_nested_parallel_for/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 03_vectorization
 	cd ./03_vectorization; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/03_vectorization/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/03_vectorization/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 04_team_scan
 	cd ./04_team_scan; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/04_team_scan/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/04_team_scan/Makefile ${KOKKOS_SETTINGS}
 
 build-insource:
 	cd ./01_thread_teams; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./01_thread_teams_lambda; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./02_nested_parallel_for; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./03_vectorization; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./04_team_scan; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 test:
 	cd ./01_thread_teams; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams/Makefile ${KOKKOS_SETTINGS}
 	cd ./01_thread_teams_lambda; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams_lambda/Makefile ${KOKKOS_SETTINGS}
 	cd ./02_nested_parallel_for; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/02_nested_parallel_for/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/02_nested_parallel_for/Makefile ${KOKKOS_SETTINGS}
 	cd ./03_vectorization; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/03_vectorization/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/03_vectorization/Makefile ${KOKKOS_SETTINGS}
 	cd ./04_team_scan; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/04_team_scan/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/04_team_scan/Makefile ${KOKKOS_SETTINGS}
 
 test-insource:
 	cd ./01_thread_teams; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./01_thread_teams_lambda; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./02_nested_parallel_for; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./03_vectorization; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./04_team_scan; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 clean:
 	cd ./01_thread_teams; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams/Makefile ${KOKKOS_SETTINGS}
 	cd ./01_thread_teams_lambda; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/01_thread_teams_lambda/Makefile ${KOKKOS_SETTINGS}
 	cd ./02_nested_parallel_for; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/02_nested_parallel_for/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/02_nested_parallel_for/Makefile ${KOKKOS_SETTINGS}
 	cd ./03_vectorization; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/03_vectorization/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/03_vectorization/Makefile ${KOKKOS_SETTINGS}
 	cd ./04_team_scan; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/04_team_scan/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/04_team_scan/Makefile ${KOKKOS_SETTINGS}
 
 clean-insource:
 	cd ./01_thread_teams; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./01_thread_teams_lambda; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./02_nested_parallel_for; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./03_vectorization; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./04_team_scan; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
diff --git a/lib/kokkos/example/tutorial/Makefile b/lib/kokkos/example/tutorial/Makefile
index 063ace8aab..7b2732eeed 100644
--- a/lib/kokkos/example/tutorial/Makefile
+++ b/lib/kokkos/example/tutorial/Makefile
@@ -23,152 +23,152 @@ endif
 build:
 	mkdir -p 01_hello_world
 	cd ./01_hello_world; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/01_hello_world/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/01_hello_world/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 01_hello_world_lambda
 	cd ./01_hello_world_lambda; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/01_hello_world_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/01_hello_world_lambda/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 02_simple_reduce
 	cd ./02_simple_reduce; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 02_simple_reduce_lambda
 	cd ./02_simple_reduce_lambda; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce_lambda/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 03_simple_view
 	cd ./03_simple_view; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/03_simple_view/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/03_simple_view/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 03_simple_view_lambda
 	cd ./03_simple_view_lambda; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/03_simple_view_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/03_simple_view_lambda/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 04_simple_memoryspaces
 	cd ./04_simple_memoryspaces; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/04_simple_memoryspaces/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/04_simple_memoryspaces/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p 05_simple_atomics
 	cd ./05_simple_atomics; \
-	make build -j 4 -f ${KOKKOS_PATH}/example/tutorial/05_simple_atomics/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/05_simple_atomics/Makefile ${KOKKOS_SETTINGS}
 	mkdir -p Advanced_Views
 	cd ./Advanced_Views; \
-	make build -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	mkdir -p Algorithms
 	cd ./Algorithms; \
-	make build -f ${KOKKOS_PATH}/example/tutorial/Algorithms/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Algorithms/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	mkdir -p Hierarchical_Parallelism
 	cd ./Hierarchical_Parallelism; \
-	make build -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) build -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 
 build-insource:
 	cd ./01_hello_world; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./01_hello_world_lambda; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./02_simple_reduce; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./02_simple_reduce_lambda; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./03_simple_view; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./03_simple_view_lambda; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./04_simple_memoryspaces; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./05_simple_atomics; \
-	make build -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) build ${KOKKOS_SETTINGS}
 	cd ./Advanced_Views; \
-	make build KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) build KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	cd ./Algorithms; \
-	make build KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) build KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	cd ./Hierarchical_Parallelism; \
-	make build KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) build KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 test:
 	cd ./01_hello_world; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/01_hello_world/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/01_hello_world/Makefile ${KOKKOS_SETTINGS}
 	cd ./01_hello_world_lambda; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/01_hello_world_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/01_hello_world_lambda/Makefile ${KOKKOS_SETTINGS}
 	cd ./02_simple_reduce; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce/Makefile ${KOKKOS_SETTINGS}
 	cd ./02_simple_reduce_lambda; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce_lambda/Makefile ${KOKKOS_SETTINGS}
 	cd ./03_simple_view; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/03_simple_view/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/03_simple_view/Makefile ${KOKKOS_SETTINGS}
 	cd ./03_simple_view_lambda; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/03_simple_view_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/03_simple_view_lambda/Makefile ${KOKKOS_SETTINGS}
 	cd ./04_simple_memoryspaces; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/04_simple_memoryspaces/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/04_simple_memoryspaces/Makefile ${KOKKOS_SETTINGS}
 	cd ./05_simple_atomics; \
-	make test -j 4 -f ${KOKKOS_PATH}/example/tutorial/05_simple_atomics/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/05_simple_atomics/Makefile ${KOKKOS_SETTINGS}
 	cd ./Advanced_Views; \
-	make test -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	cd ./Algorithms; \
-	make test -f ${KOKKOS_PATH}/example/tutorial/Algorithms/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Algorithms/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	cd ./Hierarchical_Parallelism; \
-	make test -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) test -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 
 test-insource:
 	cd ./01_hello_world; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./01_hello_world_lambda; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./02_simple_reduce; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./02_simple_reduce_lambda; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./03_simple_view; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./03_simple_view_lambda; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./04_simple_memoryspaces; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./05_simple_atomics; \
-	make test -j 4 ${KOKKOS_SETTINGS}
+	$(MAKE) test ${KOKKOS_SETTINGS}
 	cd ./Advanced_Views; \
-	make test KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) test KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	cd ./Algorithms; \
-	make test KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) test KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	cd ./Hierarchical_Parallelism; \
-	make test KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) test KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 clean:
 	cd ./01_hello_world; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/01_hello_world/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/01_hello_world/Makefile ${KOKKOS_SETTINGS}
 	cd ./01_hello_world_lambda; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/01_hello_world_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/01_hello_world_lambda/Makefile ${KOKKOS_SETTINGS}
 	cd ./02_simple_reduce; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce/Makefile ${KOKKOS_SETTINGS}
 	cd ./02_simple_reduce_lambda; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/02_simple_reduce_lambda/Makefile ${KOKKOS_SETTINGS}
 	cd ./03_simple_view; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/03_simple_view/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/03_simple_view/Makefile ${KOKKOS_SETTINGS}
 	cd ./03_simple_view_lambda; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/03_simple_view_lambda/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/03_simple_view_lambda/Makefile ${KOKKOS_SETTINGS}
 	cd ./04_simple_memoryspaces; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/04_simple_memoryspaces/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/04_simple_memoryspaces/Makefile ${KOKKOS_SETTINGS}
 	cd ./05_simple_atomics; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/05_simple_atomics/Makefile ${KOKKOS_SETTINGS}
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/05_simple_atomics/Makefile ${KOKKOS_SETTINGS}
 	cd ./Advanced_Views; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Advanced_Views/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	cd ./Algorithms; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Algorithms/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Algorithms/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	cd ./Hierarchical_Parallelism; \
-	make clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) clean -f ${KOKKOS_PATH}/example/tutorial/Hierarchical_Parallelism/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 
 clean-insource:
 	cd ./01_hello_world; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./01_hello_world_lambda; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./02_simple_reduce; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./02_simple_reduce_lambda; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./03_simple_view; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./03_simple_view_lambda; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./04_simple_memoryspaces; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./05_simple_atomics; \
-	make clean ${KOKKOS_SETTINGS}
+	$(MAKE) clean ${KOKKOS_SETTINGS}
 	cd ./Advanced_Views; \
-	make clean KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) clean KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	cd ./Algorithms; \
-	make clean KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) clean KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
 	cd ./Hierarchical_Parallelism; \
-	make clean KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
+	$(MAKE) clean KOKKOS_SETTINGS='${KOKKOS_SETTINGS}'
diff --git a/lib/kokkos/example/tutorial/launch_bounds/CMakeLists.txt b/lib/kokkos/example/tutorial/launch_bounds/CMakeLists.txt
new file mode 100644
index 0000000000..7c78db840f
--- /dev/null
+++ b/lib/kokkos/example/tutorial/launch_bounds/CMakeLists.txt
@@ -0,0 +1,10 @@
+
+INCLUDE_DIRECTORIES(${CMAKE_CURRENT_BINARY_DIR})
+INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
+
+# This is a tutorial, not a test, so we don't ask CTest to run it.
+TRIBITS_ADD_EXECUTABLE(
+  tutorial_02_simple_reduce
+  SOURCES simple_reduce.cpp
+  COMM serial mpi
+  )
diff --git a/lib/kokkos/example/tutorial/launch_bounds/Makefile b/lib/kokkos/example/tutorial/launch_bounds/Makefile
new file mode 100644
index 0000000000..5b605a4119
--- /dev/null
+++ b/lib/kokkos/example/tutorial/launch_bounds/Makefile
@@ -0,0 +1,56 @@
+KOKKOS_PATH = ../../..
+KOKKOS_SRC_PATH = ${KOKKOS_PATH}
+SRC = $(wildcard ${KOKKOS_SRC_PATH}/example/tutorial/launch_bounds/*.cpp)
+vpath %.cpp $(sort $(dir $(SRC)))
+
+default: build
+	echo "Start Build"
+
+ifneq (,$(findstring Cuda,$(KOKKOS_DEVICES)))
+CXX = ${KOKKOS_PATH}/bin/nvcc_wrapper
+CXXFLAGS = -O3
+LINK = ${CXX}
+LINKFLAGS = 
+EXE = launch_bounds.cuda
+KOKKOS_DEVICES = "Cuda,OpenMP"
+KOKKOS_ARCH = "SNB,Kepler35"
+else
+CXX = g++
+CXXFLAGS = -O3
+LINK = ${CXX}
+LINKFLAGS =  
+EXE = launch_bounds.host
+KOKKOS_DEVICES = "OpenMP"
+KOKKOS_ARCH = "SNB"
+endif
+
+# WAR for "undefined memcpy" w/ Ubuntu + CUDA 7.5
+CXXFLAGS += -D_FORCE_INLINES
+# Additional compile-time information
+CXXFLAGS += -Xptxas=-v
+
+DEPFLAGS = -M
+
+OBJ = $(notdir $(SRC:.cpp=.o))
+LIB =
+
+include $(KOKKOS_PATH)/Makefile.kokkos
+
+temp:
+	echo $(KOKKOS_INTERNAL_USE_CUDA) $(CUDA_PATH)
+
+build: $(EXE)
+
+test: $(EXE)
+	./$(EXE)
+
+$(EXE): $(OBJ) $(KOKKOS_LINK_DEPENDS)
+	$(LINK) $(KOKKOS_LDFLAGS) $(LINKFLAGS) $(EXTRA_PATH) $(OBJ) $(KOKKOS_LIBS) $(LIB) -o $(EXE)
+
+clean: kokkos-clean 
+	rm -f *.o *.cuda *.host
+
+# Compilation rules
+
+%.o:%.cpp $(KOKKOS_CPP_DEPENDS)
+	$(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) $(EXTRA_INC) -c $< -o $(notdir $@)
diff --git a/lib/kokkos/example/tutorial/launch_bounds/launch_bounds_reduce.cpp b/lib/kokkos/example/tutorial/launch_bounds/launch_bounds_reduce.cpp
new file mode 100644
index 0000000000..9a26eda507
--- /dev/null
+++ b/lib/kokkos/example/tutorial/launch_bounds/launch_bounds_reduce.cpp
@@ -0,0 +1,173 @@
+/*
+//@HEADER
+// ************************************************************************
+// 
+//                        Kokkos v. 2.0
+//              Copyright (2014) Sandia Corporation
+// 
+// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
+// the U.S. Government retains certain rights in this software.
+// 
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// 1. Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//
+// 2. Redistributions in binary form must reproduce the above copyright
+// notice, this list of conditions and the following disclaimer in the
+// documentation and/or other materials provided with the distribution.
+//
+// 3. Neither the name of the Corporation nor the names of the
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Questions? Contact  H. Carter Edwards (hcedwar@sandia.gov)
+// 
+// ************************************************************************
+//@HEADER
+*/
+
+#include 
+#include 
+
+//
+// First reduction (parallel_reduce) example:
+//   1. Start up Kokkos
+//   2. Execute a parallel_reduce loop in the default execution space,
+//      using a functor to define the loop body
+//   3. Shut down Kokkos
+//
+struct collision {
+// Reduction functor
+// For each i, we generate 10 hashes, look for and count collisions
+// We use parallel_reduce to count the total collisions
+// Note that we're just counting collisions within the 10 generated
+// one i.
+// This function was chosen as one that very simply can increase the
+// register count.
+  typedef int value_type;
+
+  KOKKOS_INLINE_FUNCTION
+  int hash(int q) const {
+	  // A simple hash by Justin Sobel
+	  // Thanks to Arash Partow (partow.net)
+	  char* fourchars = (char*)&q;
+	  int hash = 1315423911;
+	  for (int i=0; i<4; fourchars++, i++) {
+		  hash ^= ((hash<<5) + *fourchars + (hash >> 2));
+	  }
+	  return hash;
+  }
+
+  KOKKOS_INLINE_FUNCTION
+  void operator () (const int i, int& lsum) const {
+	  //This is a silly function which generates 10 hashes
+	  // then checks for collisions
+	  int a = hash(i)%64;
+	  int b = hash(i*3)%64;
+	  int c = hash(i*5)%64;
+	  int d = hash(i*7)%64;
+	  int e = hash(i*11)%64;
+	  int f = hash(i*17)%64;
+	  int g = hash(i*23)%64;
+	  int h = hash(i*29)%64;
+	  int j = hash(i*31)%64;
+	  int k = hash(i*37)%64;
+
+
+	  if (a==b) lsum++;
+	  if (a==c) lsum++;
+	  if (a==d) lsum++;
+	  if (a==e) lsum++;
+	  if (a==f) lsum++;
+	  if (a==g) lsum++;
+	  if (a==h) lsum++;
+	  if (a==j) lsum++;
+	  if (a==k) lsum++;
+	  if (b==c) lsum++;
+	  if (b==d) lsum++;
+	  if (b==e) lsum++;
+	  if (b==f) lsum++;
+	  if (b==g) lsum++;
+	  if (b==h) lsum++;
+	  if (b==j) lsum++;
+	  if (b==k) lsum++;
+	  if (c==d) lsum++;
+	  if (c==e) lsum++;
+	  if (c==f) lsum++;
+	  if (c==g) lsum++;
+	  if (c==h) lsum++;
+	  if (c==j) lsum++;
+	  if (c==k) lsum++;
+	  if (d==e) lsum++;
+	  if (d==f) lsum++;
+	  if (d==g) lsum++;
+	  if (d==h) lsum++;
+	  if (d==j) lsum++;
+	  if (d==k) lsum++;
+	  if (e==f) lsum++;
+	  if (e==g) lsum++;
+	  if (e==h) lsum++;
+	  if (e==j) lsum++;
+	  if (e==k) lsum++;
+	  if (f==g) lsum++;
+	  if (f==h) lsum++;
+	  if (f==j) lsum++;
+	  if (f==k) lsum++;
+	  if (g==h) lsum++;
+	  if (g==j) lsum++;
+	  if (g==k) lsum++;
+	  if (h==j) lsum++;
+	  if (h==k) lsum++;
+	  if (j==k) lsum++;
+  }
+
+
+
+};
+
+int main (int argc, char* argv[]) {
+  Kokkos::initialize (argc, argv);
+  const int n = 10000;
+
+  // Compute and count hash collisions in
+  // parallel, using Kokkos.
+  // This is not really a useful algorithm, but it demonstrates the
+  // LaunchBounds functionality
+  int sum1 = 0;
+  int sum2 = 0;
+  
+  //Without LaunchBounds, the kernel uses 56 registers
+  Kokkos::parallel_reduce (n, collision (), sum1);
+
+  //With LaunchBounds, we can reduce the register usage to 32
+  Kokkos::parallel_reduce (Kokkos::RangePolicy>(0,n), collision (), sum2);
+
+  printf ("Number of collisions, "
+          "computed in parallel, is %i\n", sum1);
+
+  if (sum1 != sum2) {
+	  printf( "Uh-oh! Results do not match\n");
+	  return -1;
+  }
+
+  Kokkos::finalize();
+  
+
+  return 0;
+}
+
diff --git a/lib/kokkos/generate_makefile.bash b/lib/kokkos/generate_makefile.bash
index 5f2442102d..6d636dc7e4 100755
--- a/lib/kokkos/generate_makefile.bash
+++ b/lib/kokkos/generate_makefile.bash
@@ -1,7 +1,6 @@
 #!/bin/bash
 
 KOKKOS_DEVICES=""
-MAKE_J_OPTION="32"
 
 KOKKOS_DO_EXAMPLES="1"
 
@@ -70,7 +69,8 @@ do
       KOKKOS_DEBUG=yes
       ;;
     --make-j*)
-      MAKE_J_OPTION="${key#*=}"
+      echo "Warning: ${key} is deprecated"
+      echo "Call make with appropriate -j flag"
       ;;
     --no-examples)
       KOKKOS_DO_EXAMPLES="0"
@@ -110,23 +110,34 @@ do
       echo "--with-devices:                       Explicitly add a set of backends."
       echo ""
       echo "--arch=[OPT]:  Set target architectures. Options are:"
+      echo "               [AMD]"
+      echo "                 AMDAVX         = AMD CPU"
+      echo "               [ARM]"
       echo "                 ARMv80         = ARMv8.0 Compatible CPU"
       echo "                 ARMv81         = ARMv8.1 Compatible CPU"
       echo "                 ARMv8-ThunderX = ARMv8 Cavium ThunderX CPU"
+      echo "               [IBM]"
+      echo "                 Power8         = IBM POWER8 CPUs"
+      echo "                 Power9         = IBM POWER9 CPUs"
+      echo "               [Intel]"
+      echo "                 WSM            = Intel Westmere CPUs"
       echo "                 SNB            = Intel Sandy/Ivy Bridge CPUs"
       echo "                 HSW            = Intel Haswell CPUs"
       echo "                 BDW            = Intel Broadwell Xeon E-class CPUs"
       echo "                 SKX            = Intel Sky Lake Xeon E-class HPC CPUs (AVX512)"
+      echo "               [Intel Xeon Phi]"
       echo "                 KNC            = Intel Knights Corner Xeon Phi"
       echo "                 KNL            = Intel Knights Landing Xeon Phi"
+      echo "               [NVIDIA]"
       echo "                 Kepler30       = NVIDIA Kepler generation CC 3.0"
+      echo "                 Kepler32       = NVIDIA Kepler generation CC 3.2"
       echo "                 Kepler35       = NVIDIA Kepler generation CC 3.5"
       echo "                 Kepler37       = NVIDIA Kepler generation CC 3.7"
+      echo "                 Maxwell50      = NVIDIA Maxwell generation CC 5.0"
+      echo "                 Maxwell52      = NVIDIA Maxwell generation CC 5.2"
+      echo "                 Maxwell53      = NVIDIA Maxwell generation CC 5.3"
       echo "                 Pascal60       = NVIDIA Pascal generation CC 6.0"
       echo "                 Pascal61       = NVIDIA Pascal generation CC 6.1"
-      echo "                 Maxwell50      = NVIDIA Maxwell generation CC 5.0"
-      echo "                 Power8         = IBM POWER8 CPUs"
-      echo "                 Power9         = IBM POWER9 CPUs"
       echo ""
       echo "--compiler=/Path/To/Compiler  Set the compiler."
       echo "--debug,-dbg:                 Enable Debugging."
@@ -142,10 +153,14 @@ do
       echo "                                tests.)"
       echo "--with-hwloc=/Path/To/Hwloc:  Set path to hwloc."
       echo "--with-options=[OPT]:         Additional options to Kokkos:"
+      echo "                                compiler_warnings"
       echo "                                aggressive_vectorization = add ivdep on loops"
+      echo "                                disable_profiling = do not compile with profiling hooks"
+      echo "                                "
       echo "--with-cuda-options=[OPT]:    Additional options to CUDA:"
       echo "                                force_uvm, use_ldg, enable_lambda, rdc"
-      echo "--make-j=[NUM]:               Set -j flag used during build."
+      echo "--make-j=[NUM]:               DEPRECATED: call make with appropriate"
+      echo "                                -j flag"
       exit 0
       ;;
     *)
@@ -237,27 +252,27 @@ else
   KOKKOS_INSTALL_PATH=${KOKKOS_TEST_INSTALL_PATH}
 fi
 
-mkdir install
+mkdir -p install
 echo "#Makefile to satisfy existens of target kokkos-clean before installing the library" > install/Makefile.kokkos
 echo "kokkos-clean:" >> install/Makefile.kokkos
 echo "" >> install/Makefile.kokkos
-mkdir core
-mkdir core/unit_test
-mkdir core/perf_test
-mkdir containers
-mkdir containers/unit_tests
-mkdir containers/performance_tests
-mkdir algorithms
-mkdir algorithms/unit_tests
-mkdir algorithms/performance_tests
-mkdir example
-mkdir example/fixture
-mkdir example/feint
-mkdir example/fenl
-mkdir example/tutorial
+mkdir -p core
+mkdir -p core/unit_test
+mkdir -p core/perf_test
+mkdir -p containers
+mkdir -p containers/unit_tests
+mkdir -p containers/performance_tests
+mkdir -p algorithms
+mkdir -p algorithms/unit_tests
+mkdir -p algorithms/performance_tests
+mkdir -p example
+mkdir -p example/fixture
+mkdir -p example/feint
+mkdir -p example/fenl
+mkdir -p example/tutorial
 
 if [ ${#KOKKOS_ENABLE_EXAMPLE_ICHOL} -gt 0 ]; then
-  mkdir example/ichol
+  mkdir -p example/ichol
 fi
 
 KOKKOS_SETTINGS="${KOKKOS_SETTINGS_NO_KOKKOS_PATH} KOKKOS_PATH=${KOKKOS_PATH}"
@@ -266,115 +281,115 @@ KOKKOS_SETTINGS="${KOKKOS_SETTINGS_NO_KOKKOS_PATH} KOKKOS_PATH=${KOKKOS_PATH}"
 echo "KOKKOS_SETTINGS=${KOKKOS_SETTINGS}" > core/unit_test/Makefile
 echo "" >> core/unit_test/Makefile
 echo "all:" >> core/unit_test/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/core/unit_test/Makefile ${KOKKOS_SETTINGS}" >> core/unit_test/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/core/unit_test/Makefile ${KOKKOS_SETTINGS}" >> core/unit_test/Makefile
 echo "" >> core/unit_test/Makefile
 echo "test: all" >> core/unit_test/Makefile
-echo -e "\tmake -f ${KOKKOS_PATH}/core/unit_test/Makefile ${KOKKOS_SETTINGS} test" >> core/unit_test/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/core/unit_test/Makefile ${KOKKOS_SETTINGS} test" >> core/unit_test/Makefile
 echo "" >> core/unit_test/Makefile
 echo "clean:" >> core/unit_test/Makefile
-echo -e "\tmake -f ${KOKKOS_PATH}/core/unit_test/Makefile ${KOKKOS_SETTINGS} clean" >> core/unit_test/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/core/unit_test/Makefile ${KOKKOS_SETTINGS} clean" >> core/unit_test/Makefile
 
 echo "KOKKOS_SETTINGS=${KOKKOS_SETTINGS}" > core/perf_test/Makefile
 echo "" >> core/perf_test/Makefile
 echo "all:" >> core/perf_test/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/core/perf_test/Makefile ${KOKKOS_SETTINGS}" >> core/perf_test/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/core/perf_test/Makefile ${KOKKOS_SETTINGS}" >> core/perf_test/Makefile
 echo "" >> core/perf_test/Makefile
 echo "test: all" >> core/perf_test/Makefile
-echo -e "\tmake -f ${KOKKOS_PATH}/core/perf_test/Makefile ${KOKKOS_SETTINGS} test" >> core/perf_test/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/core/perf_test/Makefile ${KOKKOS_SETTINGS} test" >> core/perf_test/Makefile
 echo "" >> core/perf_test/Makefile
 echo "clean:" >> core/perf_test/Makefile
-echo -e "\tmake -f ${KOKKOS_PATH}/core/perf_test/Makefile ${KOKKOS_SETTINGS} clean" >> core/perf_test/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/core/perf_test/Makefile ${KOKKOS_SETTINGS} clean" >> core/perf_test/Makefile
 
 echo "KOKKOS_SETTINGS=${KOKKOS_SETTINGS}" > containers/unit_tests/Makefile
 echo "" >> containers/unit_tests/Makefile
 echo "all:" >> containers/unit_tests/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/containers/unit_tests/Makefile ${KOKKOS_SETTINGS}" >> containers/unit_tests/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/containers/unit_tests/Makefile ${KOKKOS_SETTINGS}" >> containers/unit_tests/Makefile
 echo "" >> containers/unit_tests/Makefile
 echo "test: all" >> containers/unit_tests/Makefile
-echo -e "\tmake -f ${KOKKOS_PATH}/containers/unit_tests/Makefile ${KOKKOS_SETTINGS} test" >> containers/unit_tests/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/containers/unit_tests/Makefile ${KOKKOS_SETTINGS} test" >> containers/unit_tests/Makefile
 echo "" >> containers/unit_tests/Makefile
 echo "clean:" >> containers/unit_tests/Makefile
-echo -e "\tmake -f ${KOKKOS_PATH}/containers/unit_tests/Makefile ${KOKKOS_SETTINGS} clean" >> containers/unit_tests/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/containers/unit_tests/Makefile ${KOKKOS_SETTINGS} clean" >> containers/unit_tests/Makefile
 
 echo "KOKKOS_SETTINGS=${KOKKOS_SETTINGS}" > containers/performance_tests/Makefile
 echo "" >> containers/performance_tests/Makefile
 echo "all:" >> containers/performance_tests/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/containers/performance_tests/Makefile ${KOKKOS_SETTINGS}" >> containers/performance_tests/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/containers/performance_tests/Makefile ${KOKKOS_SETTINGS}" >> containers/performance_tests/Makefile
 echo "" >> containers/performance_tests/Makefile
 echo "test: all" >> containers/performance_tests/Makefile
-echo -e "\tmake -f ${KOKKOS_PATH}/containers/performance_tests/Makefile ${KOKKOS_SETTINGS} test" >> containers/performance_tests/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/containers/performance_tests/Makefile ${KOKKOS_SETTINGS} test" >> containers/performance_tests/Makefile
 echo "" >> containers/performance_tests/Makefile
 echo "clean:" >> containers/performance_tests/Makefile
-echo -e "\tmake -f ${KOKKOS_PATH}/containers/performance_tests/Makefile ${KOKKOS_SETTINGS} clean" >> containers/performance_tests/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/containers/performance_tests/Makefile ${KOKKOS_SETTINGS} clean" >> containers/performance_tests/Makefile
 
 echo "KOKKOS_SETTINGS=${KOKKOS_SETTINGS}" > algorithms/unit_tests/Makefile
 echo "" >> algorithms/unit_tests/Makefile
 echo "all:" >> algorithms/unit_tests/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/algorithms/unit_tests/Makefile ${KOKKOS_SETTINGS}" >> algorithms/unit_tests/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/algorithms/unit_tests/Makefile ${KOKKOS_SETTINGS}" >> algorithms/unit_tests/Makefile
 echo "" >> algorithms/unit_tests/Makefile
 echo "test: all" >> algorithms/unit_tests/Makefile
-echo -e "\tmake -f ${KOKKOS_PATH}/algorithms/unit_tests/Makefile ${KOKKOS_SETTINGS} test" >> algorithms/unit_tests/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/algorithms/unit_tests/Makefile ${KOKKOS_SETTINGS} test" >> algorithms/unit_tests/Makefile
 echo "" >> algorithms/unit_tests/Makefile
 echo "clean:" >> algorithms/unit_tests/Makefile
-echo -e "\tmake -f ${KOKKOS_PATH}/algorithms/unit_tests/Makefile ${KOKKOS_SETTINGS} clean" >> algorithms/unit_tests/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/algorithms/unit_tests/Makefile ${KOKKOS_SETTINGS} clean" >> algorithms/unit_tests/Makefile
 
 KOKKOS_SETTINGS="${KOKKOS_SETTINGS_NO_KOKKOS_PATH} KOKKOS_PATH=${KOKKOS_TEST_INSTALL_PATH}"
 
 echo "KOKKOS_SETTINGS=${KOKKOS_SETTINGS}" > example/fixture/Makefile
 echo "" >> example/fixture/Makefile
 echo "all:" >> example/fixture/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/fixture/Makefile ${KOKKOS_SETTINGS}" >> example/fixture/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/fixture/Makefile ${KOKKOS_SETTINGS}" >> example/fixture/Makefile
 echo "" >> example/fixture/Makefile
 echo "test: all" >> example/fixture/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/fixture/Makefile ${KOKKOS_SETTINGS} test" >> example/fixture/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/fixture/Makefile ${KOKKOS_SETTINGS} test" >> example/fixture/Makefile
 echo "" >> example/fixture/Makefile
 echo "clean:" >> example/fixture/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/fixture/Makefile ${KOKKOS_SETTINGS} clean" >> example/fixture/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/fixture/Makefile ${KOKKOS_SETTINGS} clean" >> example/fixture/Makefile
 
 echo "KOKKOS_SETTINGS=${KOKKOS_SETTINGS}" > example/feint/Makefile
 echo "" >> example/feint/Makefile
 echo "all:" >> example/feint/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/feint/Makefile ${KOKKOS_SETTINGS}" >> example/feint/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/feint/Makefile ${KOKKOS_SETTINGS}" >> example/feint/Makefile
 echo "" >> example/feint/Makefile
 echo "test: all" >> example/feint/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/feint/Makefile ${KOKKOS_SETTINGS} test" >> example/feint/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/feint/Makefile ${KOKKOS_SETTINGS} test" >> example/feint/Makefile
 echo "" >> example/feint/Makefile
 echo "clean:" >> example/feint/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/feint/Makefile ${KOKKOS_SETTINGS} clean" >> example/feint/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/feint/Makefile ${KOKKOS_SETTINGS} clean" >> example/feint/Makefile
 
 echo "KOKKOS_SETTINGS=${KOKKOS_SETTINGS}" > example/fenl/Makefile
 echo "" >> example/fenl/Makefile
 echo "all:" >> example/fenl/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/fenl/Makefile ${KOKKOS_SETTINGS}" >> example/fenl/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/fenl/Makefile ${KOKKOS_SETTINGS}" >> example/fenl/Makefile
 echo "" >> example/fenl/Makefile
 echo "test: all" >> example/fenl/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/fenl/Makefile ${KOKKOS_SETTINGS} test" >> example/fenl/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/fenl/Makefile ${KOKKOS_SETTINGS} test" >> example/fenl/Makefile
 echo "" >> example/fenl/Makefile
 echo "clean:" >> example/fenl/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/fenl/Makefile ${KOKKOS_SETTINGS} clean" >> example/fenl/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/fenl/Makefile ${KOKKOS_SETTINGS} clean" >> example/fenl/Makefile
 
 echo "KOKKOS_SETTINGS=${KOKKOS_SETTINGS}" > example/tutorial/Makefile
 echo "" >> example/tutorial/Makefile
 echo "build:" >> example/tutorial/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/tutorial/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}' KOKKOS_PATH=${KOKKOS_PATH} build">> example/tutorial/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/tutorial/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}' KOKKOS_PATH=${KOKKOS_PATH} build">> example/tutorial/Makefile
 echo "" >> example/tutorial/Makefile
 echo "test: build" >> example/tutorial/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/tutorial/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}' KOKKOS_PATH=${KOKKOS_PATH} test" >> example/tutorial/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/tutorial/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}' KOKKOS_PATH=${KOKKOS_PATH} test" >> example/tutorial/Makefile
 echo "" >> example/tutorial/Makefile
 echo "clean:" >> example/tutorial/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/tutorial/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}' KOKKOS_PATH=${KOKKOS_PATH} clean" >> example/tutorial/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/tutorial/Makefile KOKKOS_SETTINGS='${KOKKOS_SETTINGS}' KOKKOS_PATH=${KOKKOS_PATH} clean" >> example/tutorial/Makefile
 
 if [ ${#KOKKOS_ENABLE_EXAMPLE_ICHOL} -gt 0 ]; then
 echo "KOKKOS_SETTINGS=${KOKKOS_SETTINGS}" > example/ichol/Makefile
 echo "" >> example/ichol/Makefile
 echo "all:" >> example/ichol/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/ichol/Makefile ${KOKKOS_SETTINGS}" >> example/ichol/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/ichol/Makefile ${KOKKOS_SETTINGS}" >> example/ichol/Makefile
 echo "" >> example/ichol/Makefile
 echo "test: all" >> example/ichol/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/ichol/Makefile ${KOKKOS_SETTINGS} test" >> example/ichol/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/ichol/Makefile ${KOKKOS_SETTINGS} test" >> example/ichol/Makefile
 echo "" >> example/ichol/Makefile
 echo "clean:" >> example/ichol/Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/example/ichol/Makefile ${KOKKOS_SETTINGS} clean" >> example/ichol/Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/example/ichol/Makefile ${KOKKOS_SETTINGS} clean" >> example/ichol/Makefile
 fi
 
 KOKKOS_SETTINGS="${KOKKOS_SETTINGS_NO_KOKKOS_PATH} KOKKOS_PATH=${KOKKOS_PATH}"
@@ -385,62 +400,64 @@ echo "KOKKOS_SETTINGS=${KOKKOS_SETTINGS}" > Makefile
 echo "" >> Makefile
 echo "kokkoslib:" >> Makefile
 echo -e "\tcd core; \\" >> Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/core/src/Makefile ${KOKKOS_SETTINGS} PREFIX=${KOKKOS_INSTALL_PATH} build-lib" >> Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/core/src/Makefile ${KOKKOS_SETTINGS} PREFIX=${KOKKOS_INSTALL_PATH} build-lib" >> Makefile
 echo "" >> Makefile
 echo "install: kokkoslib" >> Makefile
 echo -e "\tcd core; \\" >> Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/core/src/Makefile ${KOKKOS_SETTINGS} PREFIX=${KOKKOS_INSTALL_PATH} install" >> Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/core/src/Makefile ${KOKKOS_SETTINGS} PREFIX=${KOKKOS_INSTALL_PATH} install" >> Makefile
 echo "" >> Makefile
 echo "kokkoslib-test:" >> Makefile
 echo -e "\tcd core; \\" >> Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/core/src/Makefile ${KOKKOS_SETTINGS} PREFIX=${KOKKOS_TEST_INSTALL_PATH} build-lib" >> Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/core/src/Makefile ${KOKKOS_SETTINGS} PREFIX=${KOKKOS_TEST_INSTALL_PATH} build-lib" >> Makefile
 echo "" >> Makefile
 echo "install-test: kokkoslib-test" >> Makefile
 echo -e "\tcd core; \\" >> Makefile
-echo -e "\tmake -j ${MAKE_J_OPTION} -f ${KOKKOS_PATH}/core/src/Makefile ${KOKKOS_SETTINGS} PREFIX=${KOKKOS_TEST_INSTALL_PATH} install" >> Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/core/src/Makefile ${KOKKOS_SETTINGS} PREFIX=${KOKKOS_TEST_INSTALL_PATH} install" >> Makefile
 echo "" >> Makefile
 echo "build-test: install-test" >> Makefile
-echo -e "\tmake -C core/unit_test" >> Makefile
-echo -e "\tmake -C core/perf_test" >> Makefile
-echo -e "\tmake -C containers/unit_tests" >> Makefile
-echo -e "\tmake -C containers/performance_tests" >> Makefile
-echo -e "\tmake -C algorithms/unit_tests" >> Makefile
+echo -e "\t\$(MAKE) -C core/unit_test" >> Makefile
+echo -e "\t\$(MAKE) -C core/perf_test" >> Makefile
+echo -e "\t\$(MAKE) -C containers/unit_tests" >> Makefile
+echo -e "\t\$(MAKE) -C containers/performance_tests" >> Makefile
+echo -e "\t\$(MAKE) -C algorithms/unit_tests" >> Makefile
 if [ ${KOKKOS_DO_EXAMPLES} -gt 0 ]; then
-echo -e "\tmake -C example/fixture" >> Makefile
-echo -e "\tmake -C example/feint" >> Makefile
-echo -e "\tmake -C example/fenl" >> Makefile
-echo -e "\tmake -C example/tutorial build" >> Makefile
+echo -e "\t\$(MAKE) -C example/fixture" >> Makefile
+echo -e "\t\$(MAKE) -C example/feint" >> Makefile
+echo -e "\t\$(MAKE) -C example/fenl" >> Makefile
+echo -e "\t\$(MAKE) -C example/tutorial build" >> Makefile
 fi
 echo "" >> Makefile
 echo "test: build-test" >> Makefile
-echo -e "\tmake -C core/unit_test test" >> Makefile
-echo -e "\tmake -C core/perf_test test" >> Makefile
-echo -e "\tmake -C containers/unit_tests test" >> Makefile
-echo -e "\tmake -C containers/performance_tests test" >> Makefile
-echo -e "\tmake -C algorithms/unit_tests test" >> Makefile
+echo -e "\t\$(MAKE) -C core/unit_test test" >> Makefile
+echo -e "\t\$(MAKE) -C core/perf_test test" >> Makefile
+echo -e "\t\$(MAKE) -C containers/unit_tests test" >> Makefile
+echo -e "\t\$(MAKE) -C containers/performance_tests test" >> Makefile
+echo -e "\t\$(MAKE) -C algorithms/unit_tests test" >> Makefile
 if [ ${KOKKOS_DO_EXAMPLES} -gt 0 ]; then
-echo -e "\tmake -C example/fixture test" >> Makefile
-echo -e "\tmake -C example/feint test" >> Makefile
-echo -e "\tmake -C example/fenl test" >> Makefile
-echo -e "\tmake -C example/tutorial test" >> Makefile
+echo -e "\t\$(MAKE) -C example/fixture test" >> Makefile
+echo -e "\t\$(MAKE) -C example/feint test" >> Makefile
+echo -e "\t\$(MAKE) -C example/fenl test" >> Makefile
+echo -e "\t\$(MAKE) -C example/tutorial test" >> Makefile
 fi
 echo "" >> Makefile
 echo "unit-tests-only:" >> Makefile
-echo -e "\tmake -C core/unit_test test" >> Makefile
-echo -e "\tmake -C containers/unit_tests test" >> Makefile
-echo -e "\tmake -C algorithms/unit_tests test" >> Makefile
+echo -e "\t\$(MAKE) -C core/unit_test test" >> Makefile
+echo -e "\t\$(MAKE) -C containers/unit_tests test" >> Makefile
+echo -e "\t\$(MAKE) -C algorithms/unit_tests test" >> Makefile
 echo "" >> Makefile
+
 echo "clean:" >> Makefile
-echo -e "\tmake -C core/unit_test clean" >> Makefile
-echo -e "\tmake -C core/perf_test clean" >> Makefile
-echo -e "\tmake -C containers/unit_tests clean" >> Makefile
-echo -e "\tmake -C containers/performance_tests clean" >> Makefile
-echo -e "\tmake -C algorithms/unit_tests clean" >> Makefile
+echo -e "\t\$(MAKE) -C core/unit_test clean" >> Makefile
+echo -e "\t\$(MAKE) -C core/perf_test clean" >> Makefile
+echo -e "\t\$(MAKE) -C containers/unit_tests clean" >> Makefile
+echo -e "\t\$(MAKE) -C containers/performance_tests clean" >> Makefile
+echo -e "\t\$(MAKE) -C algorithms/unit_tests clean" >> Makefile
 if [ ${KOKKOS_DO_EXAMPLES} -gt 0 ]; then
-echo -e "\tmake -C example/fixture clean" >> Makefile
-echo -e "\tmake -C example/feint clean" >> Makefile
-echo -e "\tmake -C example/fenl clean" >> Makefile
-echo -e "\tmake -C example/tutorial clean" >> Makefile
+echo -e "\t\$(MAKE) -C example/fixture clean" >> Makefile
+echo -e "\t\$(MAKE) -C example/feint clean" >> Makefile
+echo -e "\t\$(MAKE) -C example/fenl clean" >> Makefile
+echo -e "\t\$(MAKE) -C example/tutorial clean" >> Makefile
 fi
 echo -e "\tcd core; \\" >> Makefile
-echo -e "\tmake -f ${KOKKOS_PATH}/core/src/Makefile ${KOKKOS_SETTINGS} clean" >> Makefile
+echo -e "\t\$(MAKE) -f ${KOKKOS_PATH}/core/src/Makefile ${KOKKOS_SETTINGS} clean" >> Makefile
+
diff --git a/lib/kokkos/tpls/gtest/gtest/LICENSE b/lib/kokkos/tpls/gtest/gtest/LICENSE
new file mode 100644
index 0000000000..1941a11f8c
--- /dev/null
+++ b/lib/kokkos/tpls/gtest/gtest/LICENSE
@@ -0,0 +1,28 @@
+Copyright 2008, Google Inc.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+    * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+    * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/lib/kokkos/tpls/gtest/gtest/README b/lib/kokkos/tpls/gtest/gtest/README
new file mode 100644
index 0000000000..82964ecc32
--- /dev/null
+++ b/lib/kokkos/tpls/gtest/gtest/README
@@ -0,0 +1,13 @@
+This is a fused source version of gtest 1.7.0. All that should be necessary to
+start using gtest in your package is to declare the dependency and include
+gtest/gtest.h.
+
+However, because some of the packages that are developed in Sierra do not use a
+fused source version of gtest we need to make it possible for them to build with
+this version as well as with their native build. To facilitate this we have
+created symlinks for the other gtest headers that they use to the fused source
+gtest.h. This will make it possible for them find the headers while still using
+the fuse source version. This should not have any ill effects since the header is
+protected and allows for only using the non-gtest.h headers in their files.
+
+
diff --git a/lib/kokkos/tpls/gtest/gtest/gtest-all.cc b/lib/kokkos/tpls/gtest/gtest/gtest-all.cc
new file mode 100644
index 0000000000..538c78db93
--- /dev/null
+++ b/lib/kokkos/tpls/gtest/gtest/gtest-all.cc
@@ -0,0 +1,9594 @@
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: mheule@google.com (Markus Heule)
+//
+// Google C++ Testing Framework (Google Test)
+//
+// Sometimes it's desirable to build Google Test by compiling a single file.
+// This file serves this purpose.
+
+// This line ensures that gtest.h can be compiled on its own, even
+// when it's fused.
+#include "gtest/gtest.h"
+
+// The following lines pull in the real gtest *.cc files.
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+
+// Copyright 2007, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// Utilities for testing Google Test itself and code that uses Google Test
+// (e.g. frameworks built on top of Google Test).
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_SPI_H_
+#define GTEST_INCLUDE_GTEST_GTEST_SPI_H_
+
+
+namespace testing {
+
+// This helper class can be used to mock out Google Test failure reporting
+// so that we can test Google Test or code that builds on Google Test.
+//
+// An object of this class appends a TestPartResult object to the
+// TestPartResultArray object given in the constructor whenever a Google Test
+// failure is reported. It can either intercept only failures that are
+// generated in the same thread that created this object or it can intercept
+// all generated failures. The scope of this mock object can be controlled with
+// the second argument to the two arguments constructor.
+class GTEST_API_ ScopedFakeTestPartResultReporter
+    : public TestPartResultReporterInterface {
+ public:
+  // The two possible mocking modes of this object.
+  enum InterceptMode {
+    INTERCEPT_ONLY_CURRENT_THREAD,  // Intercepts only thread local failures.
+    INTERCEPT_ALL_THREADS           // Intercepts all failures.
+  };
+
+  // The c'tor sets this object as the test part result reporter used
+  // by Google Test.  The 'result' parameter specifies where to report the
+  // results. This reporter will only catch failures generated in the current
+  // thread. DEPRECATED
+  explicit ScopedFakeTestPartResultReporter(TestPartResultArray* result);
+
+  // Same as above, but you can choose the interception scope of this object.
+  ScopedFakeTestPartResultReporter(InterceptMode intercept_mode,
+                                   TestPartResultArray* result);
+
+  // The d'tor restores the previous test part result reporter.
+  virtual ~ScopedFakeTestPartResultReporter();
+
+  // Appends the TestPartResult object to the TestPartResultArray
+  // received in the constructor.
+  //
+  // This method is from the TestPartResultReporterInterface
+  // interface.
+  virtual void ReportTestPartResult(const TestPartResult& result);
+ private:
+  void Init();
+
+  const InterceptMode intercept_mode_;
+  TestPartResultReporterInterface* old_reporter_;
+  TestPartResultArray* const result_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ScopedFakeTestPartResultReporter);
+};
+
+namespace internal {
+
+// A helper class for implementing EXPECT_FATAL_FAILURE() and
+// EXPECT_NONFATAL_FAILURE().  Its destructor verifies that the given
+// TestPartResultArray contains exactly one failure that has the given
+// type and contains the given substring.  If that's not the case, a
+// non-fatal failure will be generated.
+class GTEST_API_ SingleFailureChecker {
+ public:
+  // The constructor remembers the arguments.
+  SingleFailureChecker(const TestPartResultArray* results,
+                       TestPartResult::Type type,
+                       const string& substr);
+  ~SingleFailureChecker();
+ private:
+  const TestPartResultArray* const results_;
+  const TestPartResult::Type type_;
+  const string substr_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(SingleFailureChecker);
+};
+
+}  // namespace internal
+
+}  // namespace testing
+
+// A set of macros for testing Google Test assertions or code that's expected
+// to generate Google Test fatal failures.  It verifies that the given
+// statement will cause exactly one fatal Google Test failure with 'substr'
+// being part of the failure message.
+//
+// There are two different versions of this macro. EXPECT_FATAL_FAILURE only
+// affects and considers failures generated in the current thread and
+// EXPECT_FATAL_FAILURE_ON_ALL_THREADS does the same but for all threads.
+//
+// The verification of the assertion is done correctly even when the statement
+// throws an exception or aborts the current function.
+//
+// Known restrictions:
+//   - 'statement' cannot reference local non-static variables or
+//     non-static members of the current object.
+//   - 'statement' cannot return a value.
+//   - You cannot stream a failure message to this macro.
+//
+// Note that even though the implementations of the following two
+// macros are much alike, we cannot refactor them to use a common
+// helper macro, due to some peculiarity in how the preprocessor
+// works.  The AcceptsMacroThatExpandsToUnprotectedComma test in
+// gtest_unittest.cc will fail to compile if we do that.
+#define EXPECT_FATAL_FAILURE(statement, substr) \
+  do { \
+    class GTestExpectFatalFailureHelper {\
+     public:\
+      static void Execute() { statement; }\
+    };\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        >est_failures, ::testing::TestPartResult::kFatalFailure, (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter:: \
+          INTERCEPT_ONLY_CURRENT_THREAD, >est_failures);\
+      GTestExpectFatalFailureHelper::Execute();\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+#define EXPECT_FATAL_FAILURE_ON_ALL_THREADS(statement, substr) \
+  do { \
+    class GTestExpectFatalFailureHelper {\
+     public:\
+      static void Execute() { statement; }\
+    };\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        >est_failures, ::testing::TestPartResult::kFatalFailure, (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter:: \
+          INTERCEPT_ALL_THREADS, >est_failures);\
+      GTestExpectFatalFailureHelper::Execute();\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+// A macro for testing Google Test assertions or code that's expected to
+// generate Google Test non-fatal failures.  It asserts that the given
+// statement will cause exactly one non-fatal Google Test failure with 'substr'
+// being part of the failure message.
+//
+// There are two different versions of this macro. EXPECT_NONFATAL_FAILURE only
+// affects and considers failures generated in the current thread and
+// EXPECT_NONFATAL_FAILURE_ON_ALL_THREADS does the same but for all threads.
+//
+// 'statement' is allowed to reference local variables and members of
+// the current object.
+//
+// The verification of the assertion is done correctly even when the statement
+// throws an exception or aborts the current function.
+//
+// Known restrictions:
+//   - You cannot stream a failure message to this macro.
+//
+// Note that even though the implementations of the following two
+// macros are much alike, we cannot refactor them to use a common
+// helper macro, due to some peculiarity in how the preprocessor
+// works.  If we do that, the code won't compile when the user gives
+// EXPECT_NONFATAL_FAILURE() a statement that contains a macro that
+// expands to code containing an unprotected comma.  The
+// AcceptsMacroThatExpandsToUnprotectedComma test in gtest_unittest.cc
+// catches that.
+//
+// For the same reason, we have to write
+//   if (::testing::internal::AlwaysTrue()) { statement; }
+// instead of
+//   GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement)
+// to avoid an MSVC warning on unreachable code.
+#define EXPECT_NONFATAL_FAILURE(statement, substr) \
+  do {\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        >est_failures, ::testing::TestPartResult::kNonFatalFailure, \
+        (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter:: \
+          INTERCEPT_ONLY_CURRENT_THREAD, >est_failures);\
+      if (::testing::internal::AlwaysTrue()) { statement; }\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+#define EXPECT_NONFATAL_FAILURE_ON_ALL_THREADS(statement, substr) \
+  do {\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        >est_failures, ::testing::TestPartResult::kNonFatalFailure, \
+        (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter::INTERCEPT_ALL_THREADS, \
+          >est_failures);\
+      if (::testing::internal::AlwaysTrue()) { statement; }\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_SPI_H_
+
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+
+#include 
+#include 
+#include 
+#include   // NOLINT
+#include 
+#include 
+
+#if GTEST_OS_LINUX
+
+// TODO(kenton@google.com): Use autoconf to detect availability of
+// gettimeofday().
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+
+# include   // NOLINT
+# include   // NOLINT
+# include   // NOLINT
+// Declares vsnprintf().  This header is not available on Windows.
+# include   // NOLINT
+# include   // NOLINT
+# include   // NOLINT
+# include   // NOLINT
+# include 
+
+#elif GTEST_OS_SYMBIAN
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+# include   // NOLINT
+
+#elif GTEST_OS_ZOS
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+# include   // NOLINT
+
+// On z/OS we additionally need strings.h for strcasecmp.
+# include   // NOLINT
+
+#elif GTEST_OS_WINDOWS_MOBILE  // We are on Windows CE.
+
+# include   // NOLINT
+
+#elif GTEST_OS_WINDOWS  // We are on Windows proper.
+
+# include   // NOLINT
+# include   // NOLINT
+# include   // NOLINT
+# include   // NOLINT
+
+# if GTEST_OS_WINDOWS_MINGW
+// MinGW has gettimeofday() but not _ftime64().
+// TODO(kenton@google.com): Use autoconf to detect availability of
+//   gettimeofday().
+// TODO(kenton@google.com): There are other ways to get the time on
+//   Windows, like GetTickCount() or GetSystemTimeAsFileTime().  MinGW
+//   supports these.  consider using them instead.
+#  define GTEST_HAS_GETTIMEOFDAY_ 1
+#  include   // NOLINT
+# endif  // GTEST_OS_WINDOWS_MINGW
+
+// cpplint thinks that the header is already included, so we want to
+// silence it.
+# include   // NOLINT
+
+#else
+
+// Assume other platforms have gettimeofday().
+// TODO(kenton@google.com): Use autoconf to detect availability of
+//   gettimeofday().
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+
+// cpplint thinks that the header is already included, so we want to
+// silence it.
+# include   // NOLINT
+# include   // NOLINT
+
+#endif  // GTEST_OS_LINUX
+
+#if GTEST_HAS_EXCEPTIONS
+# include 
+#endif
+
+#if GTEST_CAN_STREAM_RESULTS_
+# include   // NOLINT
+# include   // NOLINT
+#endif
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick is to
+// prevent a user from accidentally including gtest-internal-inl.h in
+// his code.
+#define GTEST_IMPLEMENTATION_ 1
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Utility functions and classes used by the Google C++ testing framework.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// This file contains purely Google Test's internal implementation.  Please
+// DO NOT #INCLUDE IT IN A USER PROGRAM.
+
+#ifndef GTEST_SRC_GTEST_INTERNAL_INL_H_
+#define GTEST_SRC_GTEST_INTERNAL_INL_H_
+
+// GTEST_IMPLEMENTATION_ is defined to 1 iff the current translation unit is
+// part of Google Test's implementation; otherwise it's undefined.
+#if !GTEST_IMPLEMENTATION_
+// A user is trying to include this from his code - just say no.
+# error "gtest-internal-inl.h is part of Google Test's internal implementation."
+# error "It must not be included except by Google Test itself."
+#endif  // GTEST_IMPLEMENTATION_
+
+#ifndef _WIN32_WCE
+# include 
+#endif  // !_WIN32_WCE
+#include 
+#include   // For strtoll/_strtoul64/malloc/free.
+#include   // For memmove.
+
+#include 
+#include 
+#include 
+
+
+#if GTEST_CAN_STREAM_RESULTS_
+# include   // NOLINT
+# include   // NOLINT
+#endif
+
+#if GTEST_OS_WINDOWS
+# include   // NOLINT
+#endif  // GTEST_OS_WINDOWS
+
+
+namespace testing {
+
+// Declares the flags.
+//
+// We don't want the users to modify this flag in the code, but want
+// Google Test's own unit tests to be able to access it. Therefore we
+// declare it here as opposed to in gtest.h.
+GTEST_DECLARE_bool_(death_test_use_fork);
+
+namespace internal {
+
+// The value of GetTestTypeId() as seen from within the Google Test
+// library.  This is solely for testing GetTestTypeId().
+GTEST_API_ extern const TypeId kTestTypeIdInGoogleTest;
+
+// Names of the flags (needed for parsing Google Test flags).
+const char kAlsoRunDisabledTestsFlag[] = "also_run_disabled_tests";
+const char kBreakOnFailureFlag[] = "break_on_failure";
+const char kCatchExceptionsFlag[] = "catch_exceptions";
+const char kColorFlag[] = "color";
+const char kFilterFlag[] = "filter";
+const char kListTestsFlag[] = "list_tests";
+const char kOutputFlag[] = "output";
+const char kPrintTimeFlag[] = "print_time";
+const char kRandomSeedFlag[] = "random_seed";
+const char kRepeatFlag[] = "repeat";
+const char kShuffleFlag[] = "shuffle";
+const char kStackTraceDepthFlag[] = "stack_trace_depth";
+const char kStreamResultToFlag[] = "stream_result_to";
+const char kThrowOnFailureFlag[] = "throw_on_failure";
+
+// A valid random seed must be in [1, kMaxRandomSeed].
+const int kMaxRandomSeed = 99999;
+
+// g_help_flag is true iff the --help flag or an equivalent form is
+// specified on the command line.
+GTEST_API_ extern bool g_help_flag;
+
+// Returns the current time in milliseconds.
+GTEST_API_ TimeInMillis GetTimeInMillis();
+
+// Returns true iff Google Test should use colors in the output.
+GTEST_API_ bool ShouldUseColor(bool stdout_is_tty);
+
+// Formats the given time in milliseconds as seconds.
+GTEST_API_ std::string FormatTimeInMillisAsSeconds(TimeInMillis ms);
+
+// Converts the given time in milliseconds to a date string in the ISO 8601
+// format, without the timezone information.  N.B.: due to the use the
+// non-reentrant localtime() function, this function is not thread safe.  Do
+// not use it in any code that can be called from multiple threads.
+GTEST_API_ std::string FormatEpochTimeInMillisAsIso8601(TimeInMillis ms);
+
+// Parses a string for an Int32 flag, in the form of "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+GTEST_API_ bool ParseInt32Flag(
+    const char* str, const char* flag, Int32* value);
+
+// Returns a random seed in range [1, kMaxRandomSeed] based on the
+// given --gtest_random_seed flag value.
+inline int GetRandomSeedFromFlag(Int32 random_seed_flag) {
+  const unsigned int raw_seed = (random_seed_flag == 0) ?
+      static_cast(GetTimeInMillis()) :
+      static_cast(random_seed_flag);
+
+  // Normalizes the actual seed to range [1, kMaxRandomSeed] such that
+  // it's easy to type.
+  const int normalized_seed =
+      static_cast((raw_seed - 1U) %
+                       static_cast(kMaxRandomSeed)) + 1;
+  return normalized_seed;
+}
+
+// Returns the first valid random seed after 'seed'.  The behavior is
+// undefined if 'seed' is invalid.  The seed after kMaxRandomSeed is
+// considered to be 1.
+inline int GetNextRandomSeed(int seed) {
+  GTEST_CHECK_(1 <= seed && seed <= kMaxRandomSeed)
+      << "Invalid random seed " << seed << " - must be in [1, "
+      << kMaxRandomSeed << "].";
+  const int next_seed = seed + 1;
+  return (next_seed > kMaxRandomSeed) ? 1 : next_seed;
+}
+
+// This class saves the values of all Google Test flags in its c'tor, and
+// restores them in its d'tor.
+class GTestFlagSaver {
+ public:
+  // The c'tor.
+  GTestFlagSaver() {
+    also_run_disabled_tests_ = GTEST_FLAG(also_run_disabled_tests);
+    break_on_failure_ = GTEST_FLAG(break_on_failure);
+    catch_exceptions_ = GTEST_FLAG(catch_exceptions);
+    color_ = GTEST_FLAG(color);
+    death_test_style_ = GTEST_FLAG(death_test_style);
+    death_test_use_fork_ = GTEST_FLAG(death_test_use_fork);
+    filter_ = GTEST_FLAG(filter);
+    internal_run_death_test_ = GTEST_FLAG(internal_run_death_test);
+    list_tests_ = GTEST_FLAG(list_tests);
+    output_ = GTEST_FLAG(output);
+    print_time_ = GTEST_FLAG(print_time);
+    random_seed_ = GTEST_FLAG(random_seed);
+    repeat_ = GTEST_FLAG(repeat);
+    shuffle_ = GTEST_FLAG(shuffle);
+    stack_trace_depth_ = GTEST_FLAG(stack_trace_depth);
+    stream_result_to_ = GTEST_FLAG(stream_result_to);
+    throw_on_failure_ = GTEST_FLAG(throw_on_failure);
+  }
+
+  // The d'tor is not virtual.  DO NOT INHERIT FROM THIS CLASS.
+  ~GTestFlagSaver() {
+    GTEST_FLAG(also_run_disabled_tests) = also_run_disabled_tests_;
+    GTEST_FLAG(break_on_failure) = break_on_failure_;
+    GTEST_FLAG(catch_exceptions) = catch_exceptions_;
+    GTEST_FLAG(color) = color_;
+    GTEST_FLAG(death_test_style) = death_test_style_;
+    GTEST_FLAG(death_test_use_fork) = death_test_use_fork_;
+    GTEST_FLAG(filter) = filter_;
+    GTEST_FLAG(internal_run_death_test) = internal_run_death_test_;
+    GTEST_FLAG(list_tests) = list_tests_;
+    GTEST_FLAG(output) = output_;
+    GTEST_FLAG(print_time) = print_time_;
+    GTEST_FLAG(random_seed) = random_seed_;
+    GTEST_FLAG(repeat) = repeat_;
+    GTEST_FLAG(shuffle) = shuffle_;
+    GTEST_FLAG(stack_trace_depth) = stack_trace_depth_;
+    GTEST_FLAG(stream_result_to) = stream_result_to_;
+    GTEST_FLAG(throw_on_failure) = throw_on_failure_;
+  }
+
+ private:
+  // Fields for saving the original values of flags.
+  bool also_run_disabled_tests_;
+  bool break_on_failure_;
+  bool catch_exceptions_;
+  std::string color_;
+  std::string death_test_style_;
+  bool death_test_use_fork_;
+  std::string filter_;
+  std::string internal_run_death_test_;
+  bool list_tests_;
+  std::string output_;
+  bool print_time_;
+  internal::Int32 random_seed_;
+  internal::Int32 repeat_;
+  bool shuffle_;
+  internal::Int32 stack_trace_depth_;
+  std::string stream_result_to_;
+  bool throw_on_failure_;
+} GTEST_ATTRIBUTE_UNUSED_;
+
+// Converts a Unicode code point to a narrow string in UTF-8 encoding.
+// code_point parameter is of type UInt32 because wchar_t may not be
+// wide enough to contain a code point.
+// If the code_point is not a valid Unicode code point
+// (i.e. outside of Unicode range U+0 to U+10FFFF) it will be converted
+// to "(Invalid Unicode 0xXXXXXXXX)".
+GTEST_API_ std::string CodePointToUtf8(UInt32 code_point);
+
+// Converts a wide string to a narrow string in UTF-8 encoding.
+// The wide string is assumed to have the following encoding:
+//   UTF-16 if sizeof(wchar_t) == 2 (on Windows, Cygwin, Symbian OS)
+//   UTF-32 if sizeof(wchar_t) == 4 (on Linux)
+// Parameter str points to a null-terminated wide string.
+// Parameter num_chars may additionally limit the number
+// of wchar_t characters processed. -1 is used when the entire string
+// should be processed.
+// If the string contains code points that are not valid Unicode code points
+// (i.e. outside of Unicode range U+0 to U+10FFFF) they will be output
+// as '(Invalid Unicode 0xXXXXXXXX)'. If the string is in UTF16 encoding
+// and contains invalid UTF-16 surrogate pairs, values in those pairs
+// will be encoded as individual Unicode characters from Basic Normal Plane.
+GTEST_API_ std::string WideStringToUtf8(const wchar_t* str, int num_chars);
+
+// Reads the GTEST_SHARD_STATUS_FILE environment variable, and creates the file
+// if the variable is present. If a file already exists at this location, this
+// function will write over it. If the variable is present, but the file cannot
+// be created, prints an error and exits.
+void WriteToShardStatusFileIfNeeded();
+
+// Checks whether sharding is enabled by examining the relevant
+// environment variable values. If the variables are present,
+// but inconsistent (e.g., shard_index >= total_shards), prints
+// an error and exits. If in_subprocess_for_death_test, sharding is
+// disabled because it must only be applied to the original test
+// process. Otherwise, we could filter out death tests we intended to execute.
+GTEST_API_ bool ShouldShard(const char* total_shards_str,
+                            const char* shard_index_str,
+                            bool in_subprocess_for_death_test);
+
+// Parses the environment variable var as an Int32. If it is unset,
+// returns default_val. If it is not an Int32, prints an error and
+// and aborts.
+GTEST_API_ Int32 Int32FromEnvOrDie(const char* env_var, Int32 default_val);
+
+// Given the total number of shards, the shard index, and the test id,
+// returns true iff the test should be run on this shard. The test id is
+// some arbitrary but unique non-negative integer assigned to each test
+// method. Assumes that 0 <= shard_index < total_shards.
+GTEST_API_ bool ShouldRunTestOnShard(
+    int total_shards, int shard_index, int test_id);
+
+// STL container utilities.
+
+// Returns the number of elements in the given container that satisfy
+// the given predicate.
+template 
+inline int CountIf(const Container& c, Predicate predicate) {
+  // Implemented as an explicit loop since std::count_if() in libCstd on
+  // Solaris has a non-standard signature.
+  int count = 0;
+  for (typename Container::const_iterator it = c.begin(); it != c.end(); ++it) {
+    if (predicate(*it))
+      ++count;
+  }
+  return count;
+}
+
+// Applies a function/functor to each element in the container.
+template 
+void ForEach(const Container& c, Functor functor) {
+  std::for_each(c.begin(), c.end(), functor);
+}
+
+// Returns the i-th element of the vector, or default_value if i is not
+// in range [0, v.size()).
+template 
+inline E GetElementOr(const std::vector& v, int i, E default_value) {
+  return (i < 0 || i >= static_cast(v.size())) ? default_value : v[i];
+}
+
+// Performs an in-place shuffle of a range of the vector's elements.
+// 'begin' and 'end' are element indices as an STL-style range;
+// i.e. [begin, end) are shuffled, where 'end' == size() means to
+// shuffle to the end of the vector.
+template 
+void ShuffleRange(internal::Random* random, int begin, int end,
+                  std::vector* v) {
+  const int size = static_cast(v->size());
+  GTEST_CHECK_(0 <= begin && begin <= size)
+      << "Invalid shuffle range start " << begin << ": must be in range [0, "
+      << size << "].";
+  GTEST_CHECK_(begin <= end && end <= size)
+      << "Invalid shuffle range finish " << end << ": must be in range ["
+      << begin << ", " << size << "].";
+
+  // Fisher-Yates shuffle, from
+  // http://en.wikipedia.org/wiki/Fisher-Yates_shuffle
+  for (int range_width = end - begin; range_width >= 2; range_width--) {
+    const int last_in_range = begin + range_width - 1;
+    const int selected = begin + random->Generate(range_width);
+    std::swap((*v)[selected], (*v)[last_in_range]);
+  }
+}
+
+// Performs an in-place shuffle of the vector's elements.
+template 
+inline void Shuffle(internal::Random* random, std::vector* v) {
+  ShuffleRange(random, 0, static_cast(v->size()), v);
+}
+
+// A function for deleting an object.  Handy for being used as a
+// functor.
+template 
+static void Delete(T* x) {
+  delete x;
+}
+
+// A predicate that checks the key of a TestProperty against a known key.
+//
+// TestPropertyKeyIs is copyable.
+class TestPropertyKeyIs {
+ public:
+  // Constructor.
+  //
+  // TestPropertyKeyIs has NO default constructor.
+  explicit TestPropertyKeyIs(const std::string& key) : key_(key) {}
+
+  // Returns true iff the test name of test property matches on key_.
+  bool operator()(const TestProperty& test_property) const {
+    return test_property.key() == key_;
+  }
+
+ private:
+  std::string key_;
+};
+
+// Class UnitTestOptions.
+//
+// This class contains functions for processing options the user
+// specifies when running the tests.  It has only static members.
+//
+// In most cases, the user can specify an option using either an
+// environment variable or a command line flag.  E.g. you can set the
+// test filter using either GTEST_FILTER or --gtest_filter.  If both
+// the variable and the flag are present, the latter overrides the
+// former.
+class GTEST_API_ UnitTestOptions {
+ public:
+  // Functions for processing the gtest_output flag.
+
+  // Returns the output format, or "" for normal printed output.
+  static std::string GetOutputFormat();
+
+  // Returns the absolute path of the requested output file, or the
+  // default (test_detail.xml in the original working directory) if
+  // none was explicitly specified.
+  static std::string GetAbsolutePathToOutputFile();
+
+  // Functions for processing the gtest_filter flag.
+
+  // Returns true iff the wildcard pattern matches the string.  The
+  // first ':' or '\0' character in pattern marks the end of it.
+  //
+  // This recursive algorithm isn't very efficient, but is clear and
+  // works well enough for matching test names, which are short.
+  static bool PatternMatchesString(const char *pattern, const char *str);
+
+  // Returns true iff the user-specified filter matches the test case
+  // name and the test name.
+  static bool FilterMatchesTest(const std::string &test_case_name,
+                                const std::string &test_name);
+
+#if GTEST_OS_WINDOWS
+  // Function for supporting the gtest_catch_exception flag.
+
+  // Returns EXCEPTION_EXECUTE_HANDLER if Google Test should handle the
+  // given SEH exception, or EXCEPTION_CONTINUE_SEARCH otherwise.
+  // This function is useful as an __except condition.
+  static int GTestShouldProcessSEH(DWORD exception_code);
+#endif  // GTEST_OS_WINDOWS
+
+  // Returns true if "name" matches the ':' separated list of glob-style
+  // filters in "filter".
+  static bool MatchesFilter(const std::string& name, const char* filter);
+};
+
+// Returns the current application's name, removing directory path if that
+// is present.  Used by UnitTestOptions::GetOutputFile.
+GTEST_API_ FilePath GetCurrentExecutableName();
+
+// The role interface for getting the OS stack trace as a string.
+class OsStackTraceGetterInterface {
+ public:
+  OsStackTraceGetterInterface() {}
+  virtual ~OsStackTraceGetterInterface() {}
+
+  // Returns the current OS stack trace as an std::string.  Parameters:
+  //
+  //   max_depth  - the maximum number of stack frames to be included
+  //                in the trace.
+  //   skip_count - the number of top frames to be skipped; doesn't count
+  //                against max_depth.
+  virtual string CurrentStackTrace(int max_depth, int skip_count) = 0;
+
+  // UponLeavingGTest() should be called immediately before Google Test calls
+  // user code. It saves some information about the current stack that
+  // CurrentStackTrace() will use to find and hide Google Test stack frames.
+  virtual void UponLeavingGTest() = 0;
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(OsStackTraceGetterInterface);
+};
+
+// A working implementation of the OsStackTraceGetterInterface interface.
+class OsStackTraceGetter : public OsStackTraceGetterInterface {
+ public:
+  OsStackTraceGetter() : caller_frame_(NULL) {}
+
+  virtual string CurrentStackTrace(int max_depth, int skip_count)
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  virtual void UponLeavingGTest() GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // This string is inserted in place of stack frames that are part of
+  // Google Test's implementation.
+  static const char* const kElidedFramesMarker;
+
+ private:
+  Mutex mutex_;  // protects all internal state
+
+  // We save the stack frame below the frame that calls user code.
+  // We do this because the address of the frame immediately below
+  // the user code changes between the call to UponLeavingGTest()
+  // and any calls to CurrentStackTrace() from within the user code.
+  void* caller_frame_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(OsStackTraceGetter);
+};
+
+// Information about a Google Test trace point.
+struct TraceInfo {
+  const char* file;
+  int line;
+  std::string message;
+};
+
+// This is the default global test part result reporter used in UnitTestImpl.
+// This class should only be used by UnitTestImpl.
+class DefaultGlobalTestPartResultReporter
+  : public TestPartResultReporterInterface {
+ public:
+  explicit DefaultGlobalTestPartResultReporter(UnitTestImpl* unit_test);
+  // Implements the TestPartResultReporterInterface. Reports the test part
+  // result in the current test.
+  virtual void ReportTestPartResult(const TestPartResult& result);
+
+ private:
+  UnitTestImpl* const unit_test_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(DefaultGlobalTestPartResultReporter);
+};
+
+// This is the default per thread test part result reporter used in
+// UnitTestImpl. This class should only be used by UnitTestImpl.
+class DefaultPerThreadTestPartResultReporter
+    : public TestPartResultReporterInterface {
+ public:
+  explicit DefaultPerThreadTestPartResultReporter(UnitTestImpl* unit_test);
+  // Implements the TestPartResultReporterInterface. The implementation just
+  // delegates to the current global test part result reporter of *unit_test_.
+  virtual void ReportTestPartResult(const TestPartResult& result);
+
+ private:
+  UnitTestImpl* const unit_test_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(DefaultPerThreadTestPartResultReporter);
+};
+
+// The private implementation of the UnitTest class.  We don't protect
+// the methods under a mutex, as this class is not accessible by a
+// user and the UnitTest class that delegates work to this class does
+// proper locking.
+class GTEST_API_ UnitTestImpl {
+ public:
+  explicit UnitTestImpl(UnitTest* parent);
+  virtual ~UnitTestImpl();
+
+  // There are two different ways to register your own TestPartResultReporter.
+  // You can register your own repoter to listen either only for test results
+  // from the current thread or for results from all threads.
+  // By default, each per-thread test result repoter just passes a new
+  // TestPartResult to the global test result reporter, which registers the
+  // test part result for the currently running test.
+
+  // Returns the global test part result reporter.
+  TestPartResultReporterInterface* GetGlobalTestPartResultReporter();
+
+  // Sets the global test part result reporter.
+  void SetGlobalTestPartResultReporter(
+      TestPartResultReporterInterface* reporter);
+
+  // Returns the test part result reporter for the current thread.
+  TestPartResultReporterInterface* GetTestPartResultReporterForCurrentThread();
+
+  // Sets the test part result reporter for the current thread.
+  void SetTestPartResultReporterForCurrentThread(
+      TestPartResultReporterInterface* reporter);
+
+  // Gets the number of successful test cases.
+  int successful_test_case_count() const;
+
+  // Gets the number of failed test cases.
+  int failed_test_case_count() const;
+
+  // Gets the number of all test cases.
+  int total_test_case_count() const;
+
+  // Gets the number of all test cases that contain at least one test
+  // that should run.
+  int test_case_to_run_count() const;
+
+  // Gets the number of successful tests.
+  int successful_test_count() const;
+
+  // Gets the number of failed tests.
+  int failed_test_count() const;
+
+  // Gets the number of disabled tests that will be reported in the XML report.
+  int reportable_disabled_test_count() const;
+
+  // Gets the number of disabled tests.
+  int disabled_test_count() const;
+
+  // Gets the number of tests to be printed in the XML report.
+  int reportable_test_count() const;
+
+  // Gets the number of all tests.
+  int total_test_count() const;
+
+  // Gets the number of tests that should run.
+  int test_to_run_count() const;
+
+  // Gets the time of the test program start, in ms from the start of the
+  // UNIX epoch.
+  TimeInMillis start_timestamp() const { return start_timestamp_; }
+
+  // Gets the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const { return elapsed_time_; }
+
+  // Returns true iff the unit test passed (i.e. all test cases passed).
+  bool Passed() const { return !Failed(); }
+
+  // Returns true iff the unit test failed (i.e. some test case failed
+  // or something outside of all tests failed).
+  bool Failed() const {
+    return failed_test_case_count() > 0 || ad_hoc_test_result()->Failed();
+  }
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  const TestCase* GetTestCase(int i) const {
+    const int index = GetElementOr(test_case_indices_, i, -1);
+    return index < 0 ? NULL : test_cases_[i];
+  }
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  TestCase* GetMutableTestCase(int i) {
+    const int index = GetElementOr(test_case_indices_, i, -1);
+    return index < 0 ? NULL : test_cases_[index];
+  }
+
+  // Provides access to the event listener list.
+  TestEventListeners* listeners() { return &listeners_; }
+
+  // Returns the TestResult for the test that's currently running, or
+  // the TestResult for the ad hoc test if no test is running.
+  TestResult* current_test_result();
+
+  // Returns the TestResult for the ad hoc test.
+  const TestResult* ad_hoc_test_result() const { return &ad_hoc_test_result_; }
+
+  // Sets the OS stack trace getter.
+  //
+  // Does nothing if the input and the current OS stack trace getter
+  // are the same; otherwise, deletes the old getter and makes the
+  // input the current getter.
+  void set_os_stack_trace_getter(OsStackTraceGetterInterface* getter);
+
+  // Returns the current OS stack trace getter if it is not NULL;
+  // otherwise, creates an OsStackTraceGetter, makes it the current
+  // getter, and returns it.
+  OsStackTraceGetterInterface* os_stack_trace_getter();
+
+  // Returns the current OS stack trace as an std::string.
+  //
+  // The maximum number of stack frames to be included is specified by
+  // the gtest_stack_trace_depth flag.  The skip_count parameter
+  // specifies the number of top frames to be skipped, which doesn't
+  // count against the number of frames to be included.
+  //
+  // For example, if Foo() calls Bar(), which in turn calls
+  // CurrentOsStackTraceExceptTop(1), Foo() will be included in the
+  // trace but Bar() and CurrentOsStackTraceExceptTop() won't.
+  std::string CurrentOsStackTraceExceptTop(int skip_count) GTEST_NO_INLINE_;
+
+  // Finds and returns a TestCase with the given name.  If one doesn't
+  // exist, creates one and returns it.
+  //
+  // Arguments:
+  //
+  //   test_case_name: name of the test case
+  //   type_param:     the name of the test's type parameter, or NULL if
+  //                   this is not a typed or a type-parameterized test.
+  //   set_up_tc:      pointer to the function that sets up the test case
+  //   tear_down_tc:   pointer to the function that tears down the test case
+  TestCase* GetTestCase(const char* test_case_name,
+                        const char* type_param,
+                        Test::SetUpTestCaseFunc set_up_tc,
+                        Test::TearDownTestCaseFunc tear_down_tc);
+
+  // Adds a TestInfo to the unit test.
+  //
+  // Arguments:
+  //
+  //   set_up_tc:    pointer to the function that sets up the test case
+  //   tear_down_tc: pointer to the function that tears down the test case
+  //   test_info:    the TestInfo object
+  void AddTestInfo(Test::SetUpTestCaseFunc set_up_tc,
+                   Test::TearDownTestCaseFunc tear_down_tc,
+                   TestInfo* test_info) {
+    // In order to support thread-safe death tests, we need to
+    // remember the original working directory when the test program
+    // was first invoked.  We cannot do this in RUN_ALL_TESTS(), as
+    // the user may have changed the current directory before calling
+    // RUN_ALL_TESTS().  Therefore we capture the current directory in
+    // AddTestInfo(), which is called to register a TEST or TEST_F
+    // before main() is reached.
+    if (original_working_dir_.IsEmpty()) {
+      original_working_dir_.Set(FilePath::GetCurrentDir());
+      GTEST_CHECK_(!original_working_dir_.IsEmpty())
+          << "Failed to get the current working directory.";
+    }
+
+    GetTestCase(test_info->test_case_name(),
+                test_info->type_param(),
+                set_up_tc,
+                tear_down_tc)->AddTestInfo(test_info);
+  }
+
+#if GTEST_HAS_PARAM_TEST
+  // Returns ParameterizedTestCaseRegistry object used to keep track of
+  // value-parameterized tests and instantiate and register them.
+  internal::ParameterizedTestCaseRegistry& parameterized_test_registry() {
+    return parameterized_test_registry_;
+  }
+#endif  // GTEST_HAS_PARAM_TEST
+
+  // Sets the TestCase object for the test that's currently running.
+  void set_current_test_case(TestCase* a_current_test_case) {
+    current_test_case_ = a_current_test_case;
+  }
+
+  // Sets the TestInfo object for the test that's currently running.  If
+  // current_test_info is NULL, the assertion results will be stored in
+  // ad_hoc_test_result_.
+  void set_current_test_info(TestInfo* a_current_test_info) {
+    current_test_info_ = a_current_test_info;
+  }
+
+  // Registers all parameterized tests defined using TEST_P and
+  // INSTANTIATE_TEST_CASE_P, creating regular tests for each test/parameter
+  // combination. This method can be called more then once; it has guards
+  // protecting from registering the tests more then once.  If
+  // value-parameterized tests are disabled, RegisterParameterizedTests is
+  // present but does nothing.
+  void RegisterParameterizedTests();
+
+  // Runs all tests in this UnitTest object, prints the result, and
+  // returns true if all tests are successful.  If any exception is
+  // thrown during a test, this test is considered to be failed, but
+  // the rest of the tests will still be run.
+  bool RunAllTests();
+
+  // Clears the results of all tests, except the ad hoc tests.
+  void ClearNonAdHocTestResult() {
+    ForEach(test_cases_, TestCase::ClearTestCaseResult);
+  }
+
+  // Clears the results of ad-hoc test assertions.
+  void ClearAdHocTestResult() {
+    ad_hoc_test_result_.Clear();
+  }
+
+  // Adds a TestProperty to the current TestResult object when invoked in a
+  // context of a test or a test case, or to the global property set. If the
+  // result already contains a property with the same key, the value will be
+  // updated.
+  void RecordProperty(const TestProperty& test_property);
+
+  enum ReactionToSharding {
+    HONOR_SHARDING_PROTOCOL,
+    IGNORE_SHARDING_PROTOCOL
+  };
+
+  // Matches the full name of each test against the user-specified
+  // filter to decide whether the test should run, then records the
+  // result in each TestCase and TestInfo object.
+  // If shard_tests == HONOR_SHARDING_PROTOCOL, further filters tests
+  // based on sharding variables in the environment.
+  // Returns the number of tests that should run.
+  int FilterTests(ReactionToSharding shard_tests);
+
+  // Prints the names of the tests matching the user-specified filter flag.
+  void ListTestsMatchingFilter();
+
+  const TestCase* current_test_case() const { return current_test_case_; }
+  TestInfo* current_test_info() { return current_test_info_; }
+  const TestInfo* current_test_info() const { return current_test_info_; }
+
+  // Returns the vector of environments that need to be set-up/torn-down
+  // before/after the tests are run.
+  std::vector& environments() { return environments_; }
+
+  // Getters for the per-thread Google Test trace stack.
+  std::vector& gtest_trace_stack() {
+    return *(gtest_trace_stack_.pointer());
+  }
+  const std::vector& gtest_trace_stack() const {
+    return gtest_trace_stack_.get();
+  }
+
+#if GTEST_HAS_DEATH_TEST
+  void InitDeathTestSubprocessControlInfo() {
+    internal_run_death_test_flag_.reset(ParseInternalRunDeathTestFlag());
+  }
+  // Returns a pointer to the parsed --gtest_internal_run_death_test
+  // flag, or NULL if that flag was not specified.
+  // This information is useful only in a death test child process.
+  // Must not be called before a call to InitGoogleTest.
+  const InternalRunDeathTestFlag* internal_run_death_test_flag() const {
+    return internal_run_death_test_flag_.get();
+  }
+
+  // Returns a pointer to the current death test factory.
+  internal::DeathTestFactory* death_test_factory() {
+    return death_test_factory_.get();
+  }
+
+  void SuppressTestEventsIfInSubprocess();
+
+  friend class ReplaceDeathTestFactory;
+#endif  // GTEST_HAS_DEATH_TEST
+
+  // Initializes the event listener performing XML output as specified by
+  // UnitTestOptions. Must not be called before InitGoogleTest.
+  void ConfigureXmlOutput();
+
+#if GTEST_CAN_STREAM_RESULTS_
+  // Initializes the event listener for streaming test results to a socket.
+  // Must not be called before InitGoogleTest.
+  void ConfigureStreamingOutput();
+#endif
+
+  // Performs initialization dependent upon flag values obtained in
+  // ParseGoogleTestFlagsOnly.  Is called from InitGoogleTest after the call to
+  // ParseGoogleTestFlagsOnly.  In case a user neglects to call InitGoogleTest
+  // this function is also called from RunAllTests.  Since this function can be
+  // called more than once, it has to be idempotent.
+  void PostFlagParsingInit();
+
+  // Gets the random seed used at the start of the current test iteration.
+  int random_seed() const { return random_seed_; }
+
+  // Gets the random number generator.
+  internal::Random* random() { return &random_; }
+
+  // Shuffles all test cases, and the tests within each test case,
+  // making sure that death tests are still run first.
+  void ShuffleTests();
+
+  // Restores the test cases and tests to their order before the first shuffle.
+  void UnshuffleTests();
+
+  // Returns the value of GTEST_FLAG(catch_exceptions) at the moment
+  // UnitTest::Run() starts.
+  bool catch_exceptions() const { return catch_exceptions_; }
+
+ private:
+  friend class ::testing::UnitTest;
+
+  // Used by UnitTest::Run() to capture the state of
+  // GTEST_FLAG(catch_exceptions) at the moment it starts.
+  void set_catch_exceptions(bool value) { catch_exceptions_ = value; }
+
+  // The UnitTest object that owns this implementation object.
+  UnitTest* const parent_;
+
+  // The working directory when the first TEST() or TEST_F() was
+  // executed.
+  internal::FilePath original_working_dir_;
+
+  // The default test part result reporters.
+  DefaultGlobalTestPartResultReporter default_global_test_part_result_reporter_;
+  DefaultPerThreadTestPartResultReporter
+      default_per_thread_test_part_result_reporter_;
+
+  // Points to (but doesn't own) the global test part result reporter.
+  TestPartResultReporterInterface* global_test_part_result_repoter_;
+
+  // Protects read and write access to global_test_part_result_reporter_.
+  internal::Mutex global_test_part_result_reporter_mutex_;
+
+  // Points to (but doesn't own) the per-thread test part result reporter.
+  internal::ThreadLocal
+      per_thread_test_part_result_reporter_;
+
+  // The vector of environments that need to be set-up/torn-down
+  // before/after the tests are run.
+  std::vector environments_;
+
+  // The vector of TestCases in their original order.  It owns the
+  // elements in the vector.
+  std::vector test_cases_;
+
+  // Provides a level of indirection for the test case list to allow
+  // easy shuffling and restoring the test case order.  The i-th
+  // element of this vector is the index of the i-th test case in the
+  // shuffled order.
+  std::vector test_case_indices_;
+
+#if GTEST_HAS_PARAM_TEST
+  // ParameterizedTestRegistry object used to register value-parameterized
+  // tests.
+  internal::ParameterizedTestCaseRegistry parameterized_test_registry_;
+
+  // Indicates whether RegisterParameterizedTests() has been called already.
+  bool parameterized_tests_registered_;
+#endif  // GTEST_HAS_PARAM_TEST
+
+  // Index of the last death test case registered.  Initially -1.
+  int last_death_test_case_;
+
+  // This points to the TestCase for the currently running test.  It
+  // changes as Google Test goes through one test case after another.
+  // When no test is running, this is set to NULL and Google Test
+  // stores assertion results in ad_hoc_test_result_.  Initially NULL.
+  TestCase* current_test_case_;
+
+  // This points to the TestInfo for the currently running test.  It
+  // changes as Google Test goes through one test after another.  When
+  // no test is running, this is set to NULL and Google Test stores
+  // assertion results in ad_hoc_test_result_.  Initially NULL.
+  TestInfo* current_test_info_;
+
+  // Normally, a user only writes assertions inside a TEST or TEST_F,
+  // or inside a function called by a TEST or TEST_F.  Since Google
+  // Test keeps track of which test is current running, it can
+  // associate such an assertion with the test it belongs to.
+  //
+  // If an assertion is encountered when no TEST or TEST_F is running,
+  // Google Test attributes the assertion result to an imaginary "ad hoc"
+  // test, and records the result in ad_hoc_test_result_.
+  TestResult ad_hoc_test_result_;
+
+  // The list of event listeners that can be used to track events inside
+  // Google Test.
+  TestEventListeners listeners_;
+
+  // The OS stack trace getter.  Will be deleted when the UnitTest
+  // object is destructed.  By default, an OsStackTraceGetter is used,
+  // but the user can set this field to use a custom getter if that is
+  // desired.
+  OsStackTraceGetterInterface* os_stack_trace_getter_;
+
+  // True iff PostFlagParsingInit() has been called.
+  bool post_flag_parse_init_performed_;
+
+  // The random number seed used at the beginning of the test run.
+  int random_seed_;
+
+  // Our random number generator.
+  internal::Random random_;
+
+  // The time of the test program start, in ms from the start of the
+  // UNIX epoch.
+  TimeInMillis start_timestamp_;
+
+  // How long the test took to run, in milliseconds.
+  TimeInMillis elapsed_time_;
+
+#if GTEST_HAS_DEATH_TEST
+  // The decomposed components of the gtest_internal_run_death_test flag,
+  // parsed when RUN_ALL_TESTS is called.
+  internal::scoped_ptr internal_run_death_test_flag_;
+  internal::scoped_ptr death_test_factory_;
+#endif  // GTEST_HAS_DEATH_TEST
+
+  // A per-thread stack of traces created by the SCOPED_TRACE() macro.
+  internal::ThreadLocal > gtest_trace_stack_;
+
+  // The value of GTEST_FLAG(catch_exceptions) at the moment RunAllTests()
+  // starts.
+  bool catch_exceptions_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(UnitTestImpl);
+};  // class UnitTestImpl
+
+// Convenience function for accessing the global UnitTest
+// implementation object.
+inline UnitTestImpl* GetUnitTestImpl() {
+  return UnitTest::GetInstance()->impl();
+}
+
+#if GTEST_USES_SIMPLE_RE
+
+// Internal helper functions for implementing the simple regular
+// expression matcher.
+GTEST_API_ bool IsInSet(char ch, const char* str);
+GTEST_API_ bool IsAsciiDigit(char ch);
+GTEST_API_ bool IsAsciiPunct(char ch);
+GTEST_API_ bool IsRepeat(char ch);
+GTEST_API_ bool IsAsciiWhiteSpace(char ch);
+GTEST_API_ bool IsAsciiWordChar(char ch);
+GTEST_API_ bool IsValidEscape(char ch);
+GTEST_API_ bool AtomMatchesChar(bool escaped, char pattern, char ch);
+GTEST_API_ bool ValidateRegex(const char* regex);
+GTEST_API_ bool MatchRegexAtHead(const char* regex, const char* str);
+GTEST_API_ bool MatchRepetitionAndRegexAtHead(
+    bool escaped, char ch, char repeat, const char* regex, const char* str);
+GTEST_API_ bool MatchRegexAnywhere(const char* regex, const char* str);
+
+#endif  // GTEST_USES_SIMPLE_RE
+
+// Parses the command line for Google Test flags, without initializing
+// other parts of Google Test.
+GTEST_API_ void ParseGoogleTestFlagsOnly(int* argc, char** argv);
+GTEST_API_ void ParseGoogleTestFlagsOnly(int* argc, wchar_t** argv);
+
+#if GTEST_HAS_DEATH_TEST
+
+// Returns the message describing the last system error, regardless of the
+// platform.
+GTEST_API_ std::string GetLastErrnoDescription();
+
+# if GTEST_OS_WINDOWS
+// Provides leak-safe Windows kernel handle ownership.
+class AutoHandle {
+ public:
+  AutoHandle() : handle_(INVALID_HANDLE_VALUE) {}
+  explicit AutoHandle(HANDLE handle) : handle_(handle) {}
+
+  ~AutoHandle() { Reset(); }
+
+  HANDLE Get() const { return handle_; }
+  void Reset() { Reset(INVALID_HANDLE_VALUE); }
+  void Reset(HANDLE handle) {
+    if (handle != handle_) {
+      if (handle_ != INVALID_HANDLE_VALUE)
+        ::CloseHandle(handle_);
+      handle_ = handle;
+    }
+  }
+
+ private:
+  HANDLE handle_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(AutoHandle);
+};
+# endif  // GTEST_OS_WINDOWS
+
+// Attempts to parse a string into a positive integer pointed to by the
+// number parameter.  Returns true if that is possible.
+// GTEST_HAS_DEATH_TEST implies that we have ::std::string, so we can use
+// it here.
+template 
+bool ParseNaturalNumber(const ::std::string& str, Integer* number) {
+  // Fail fast if the given string does not begin with a digit;
+  // this bypasses strtoXXX's "optional leading whitespace and plus
+  // or minus sign" semantics, which are undesirable here.
+  if (str.empty() || !IsDigit(str[0])) {
+    return false;
+  }
+  errno = 0;
+
+  char* end;
+  // BiggestConvertible is the largest integer type that system-provided
+  // string-to-number conversion routines can return.
+
+# if GTEST_OS_WINDOWS && !defined(__GNUC__)
+
+  // MSVC and C++ Builder define __int64 instead of the standard long long.
+  typedef unsigned __int64 BiggestConvertible;
+  const BiggestConvertible parsed = _strtoui64(str.c_str(), &end, 10);
+
+# else
+
+  typedef unsigned long long BiggestConvertible;  // NOLINT
+  const BiggestConvertible parsed = strtoull(str.c_str(), &end, 10);
+
+# endif  // GTEST_OS_WINDOWS && !defined(__GNUC__)
+
+  const bool parse_success = *end == '\0' && errno == 0;
+
+  // TODO(vladl@google.com): Convert this to compile time assertion when it is
+  // available.
+  GTEST_CHECK_(sizeof(Integer) <= sizeof(parsed));
+
+  const Integer result = static_cast(parsed);
+  if (parse_success && static_cast(result) == parsed) {
+    *number = result;
+    return true;
+  }
+  return false;
+}
+#endif  // GTEST_HAS_DEATH_TEST
+
+// TestResult contains some private methods that should be hidden from
+// Google Test user but are required for testing. This class allow our tests
+// to access them.
+//
+// This class is supplied only for the purpose of testing Google Test's own
+// constructs. Do not use it in user tests, either directly or indirectly.
+class TestResultAccessor {
+ public:
+  static void RecordProperty(TestResult* test_result,
+                             const std::string& xml_element,
+                             const TestProperty& property) {
+    test_result->RecordProperty(xml_element, property);
+  }
+
+  static void ClearTestPartResults(TestResult* test_result) {
+    test_result->ClearTestPartResults();
+  }
+
+  static const std::vector& test_part_results(
+      const TestResult& test_result) {
+    return test_result.test_part_results();
+  }
+};
+
+#if GTEST_CAN_STREAM_RESULTS_
+
+// Streams test results to the given port on the given host machine.
+class StreamingListener : public EmptyTestEventListener {
+ public:
+  // Abstract base class for writing strings to a socket.
+  class AbstractSocketWriter {
+   public:
+    virtual ~AbstractSocketWriter() {}
+
+    // Sends a string to the socket.
+    virtual void Send(const string& message) = 0;
+
+    // Closes the socket.
+    virtual void CloseConnection() {}
+
+    // Sends a string and a newline to the socket.
+    void SendLn(const string& message) {
+      Send(message + "\n");
+    }
+  };
+
+  // Concrete class for actually writing strings to a socket.
+  class SocketWriter : public AbstractSocketWriter {
+   public:
+    SocketWriter(const string& host, const string& port)
+        : sockfd_(-1), host_name_(host), port_num_(port) {
+      MakeConnection();
+    }
+
+    virtual ~SocketWriter() {
+      if (sockfd_ != -1)
+        CloseConnection();
+    }
+
+    // Sends a string to the socket.
+    virtual void Send(const string& message) {
+      GTEST_CHECK_(sockfd_ != -1)
+          << "Send() can be called only when there is a connection.";
+
+      const int len = static_cast(message.length());
+      if (write(sockfd_, message.c_str(), len) != len) {
+        GTEST_LOG_(WARNING)
+            << "stream_result_to: failed to stream to "
+            << host_name_ << ":" << port_num_;
+      }
+    }
+
+   private:
+    // Creates a client socket and connects to the server.
+    void MakeConnection();
+
+    // Closes the socket.
+    void CloseConnection() {
+      GTEST_CHECK_(sockfd_ != -1)
+          << "CloseConnection() can be called only when there is a connection.";
+
+      close(sockfd_);
+      sockfd_ = -1;
+    }
+
+    int sockfd_;  // socket file descriptor
+    const string host_name_;
+    const string port_num_;
+
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(SocketWriter);
+  };  // class SocketWriter
+
+  // Escapes '=', '&', '%', and '\n' characters in str as "%xx".
+  static string UrlEncode(const char* str);
+
+  StreamingListener(const string& host, const string& port)
+      : socket_writer_(new SocketWriter(host, port)) { Start(); }
+
+  explicit StreamingListener(AbstractSocketWriter* socket_writer)
+      : socket_writer_(socket_writer) { Start(); }
+
+  void OnTestProgramStart(const UnitTest& /* unit_test */) {
+    SendLn("event=TestProgramStart");
+  }
+
+  void OnTestProgramEnd(const UnitTest& unit_test) {
+    // Note that Google Test current only report elapsed time for each
+    // test iteration, not for the entire test program.
+    SendLn("event=TestProgramEnd&passed=" + FormatBool(unit_test.Passed()));
+
+    // Notify the streaming server to stop.
+    socket_writer_->CloseConnection();
+  }
+
+  void OnTestIterationStart(const UnitTest& /* unit_test */, int iteration) {
+    SendLn("event=TestIterationStart&iteration=" +
+           StreamableToString(iteration));
+  }
+
+  void OnTestIterationEnd(const UnitTest& unit_test, int /* iteration */) {
+    SendLn("event=TestIterationEnd&passed=" +
+           FormatBool(unit_test.Passed()) + "&elapsed_time=" +
+           StreamableToString(unit_test.elapsed_time()) + "ms");
+  }
+
+  void OnTestCaseStart(const TestCase& test_case) {
+    SendLn(std::string("event=TestCaseStart&name=") + test_case.name());
+  }
+
+  void OnTestCaseEnd(const TestCase& test_case) {
+    SendLn("event=TestCaseEnd&passed=" + FormatBool(test_case.Passed())
+           + "&elapsed_time=" + StreamableToString(test_case.elapsed_time())
+           + "ms");
+  }
+
+  void OnTestStart(const TestInfo& test_info) {
+    SendLn(std::string("event=TestStart&name=") + test_info.name());
+  }
+
+  void OnTestEnd(const TestInfo& test_info) {
+    SendLn("event=TestEnd&passed=" +
+           FormatBool((test_info.result())->Passed()) +
+           "&elapsed_time=" +
+           StreamableToString((test_info.result())->elapsed_time()) + "ms");
+  }
+
+  void OnTestPartResult(const TestPartResult& test_part_result) {
+    const char* file_name = test_part_result.file_name();
+    if (file_name == NULL)
+      file_name = "";
+    SendLn("event=TestPartResult&file=" + UrlEncode(file_name) +
+           "&line=" + StreamableToString(test_part_result.line_number()) +
+           "&message=" + UrlEncode(test_part_result.message()));
+  }
+
+ private:
+  // Sends the given message and a newline to the socket.
+  void SendLn(const string& message) { socket_writer_->SendLn(message); }
+
+  // Called at the start of streaming to notify the receiver what
+  // protocol we are using.
+  void Start() { SendLn("gtest_streaming_protocol_version=1.0"); }
+
+  string FormatBool(bool value) { return value ? "1" : "0"; }
+
+  const scoped_ptr socket_writer_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(StreamingListener);
+};  // class StreamingListener
+
+#endif  // GTEST_CAN_STREAM_RESULTS_
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_SRC_GTEST_INTERNAL_INL_H_
+#undef GTEST_IMPLEMENTATION_
+
+#if GTEST_OS_WINDOWS
+# define vsnprintf _vsnprintf
+#endif  // GTEST_OS_WINDOWS
+
+namespace testing {
+
+using internal::CountIf;
+using internal::ForEach;
+using internal::GetElementOr;
+using internal::Shuffle;
+
+// Constants.
+
+// A test whose test case name or test name matches this filter is
+// disabled and not run.
+static const char kDisableTestFilter[] = "DISABLED_*:*/DISABLED_*";
+
+// A test case whose name matches this filter is considered a death
+// test case and will be run before test cases whose name doesn't
+// match this filter.
+static const char kDeathTestCaseFilter[] = "*DeathTest:*DeathTest/*";
+
+// A test filter that matches everything.
+static const char kUniversalFilter[] = "*";
+
+// The default output file for XML output.
+static const char kDefaultOutputFile[] = "test_detail.xml";
+
+// The environment variable name for the test shard index.
+static const char kTestShardIndex[] = "GTEST_SHARD_INDEX";
+// The environment variable name for the total number of test shards.
+static const char kTestTotalShards[] = "GTEST_TOTAL_SHARDS";
+// The environment variable name for the test shard status file.
+static const char kTestShardStatusFile[] = "GTEST_SHARD_STATUS_FILE";
+
+namespace internal {
+
+// The text used in failure messages to indicate the start of the
+// stack trace.
+const char kStackTraceMarker[] = "\nStack trace:\n";
+
+// g_help_flag is true iff the --help flag or an equivalent form is
+// specified on the command line.
+bool g_help_flag = false;
+
+}  // namespace internal
+
+static const char* GetDefaultFilter() {
+  return kUniversalFilter;
+}
+
+GTEST_DEFINE_bool_(
+    also_run_disabled_tests,
+    internal::BoolFromGTestEnv("also_run_disabled_tests", false),
+    "Run disabled tests too, in addition to the tests normally being run.");
+
+GTEST_DEFINE_bool_(
+    break_on_failure,
+    internal::BoolFromGTestEnv("break_on_failure", false),
+    "True iff a failed assertion should be a debugger break-point.");
+
+GTEST_DEFINE_bool_(
+    catch_exceptions,
+    internal::BoolFromGTestEnv("catch_exceptions", true),
+    "True iff " GTEST_NAME_
+    " should catch exceptions and treat them as test failures.");
+
+GTEST_DEFINE_string_(
+    color,
+    internal::StringFromGTestEnv("color", "auto"),
+    "Whether to use colors in the output.  Valid values: yes, no, "
+    "and auto.  'auto' means to use colors if the output is "
+    "being sent to a terminal and the TERM environment variable "
+    "is set to a terminal type that supports colors.");
+
+GTEST_DEFINE_string_(
+    filter,
+    internal::StringFromGTestEnv("filter", GetDefaultFilter()),
+    "A colon-separated list of glob (not regex) patterns "
+    "for filtering the tests to run, optionally followed by a "
+    "'-' and a : separated list of negative patterns (tests to "
+    "exclude).  A test is run if it matches one of the positive "
+    "patterns and does not match any of the negative patterns.");
+
+GTEST_DEFINE_bool_(list_tests, false,
+                   "List all tests without running them.");
+
+GTEST_DEFINE_string_(
+    output,
+    internal::StringFromGTestEnv("output", ""),
+    "A format (currently must be \"xml\"), optionally followed "
+    "by a colon and an output file name or directory. A directory "
+    "is indicated by a trailing pathname separator. "
+    "Examples: \"xml:filename.xml\", \"xml::directoryname/\". "
+    "If a directory is specified, output files will be created "
+    "within that directory, with file-names based on the test "
+    "executable's name and, if necessary, made unique by adding "
+    "digits.");
+
+GTEST_DEFINE_bool_(
+    print_time,
+    internal::BoolFromGTestEnv("print_time", true),
+    "True iff " GTEST_NAME_
+    " should display elapsed time in text output.");
+
+GTEST_DEFINE_int32_(
+    random_seed,
+    internal::Int32FromGTestEnv("random_seed", 0),
+    "Random number seed to use when shuffling test orders.  Must be in range "
+    "[1, 99999], or 0 to use a seed based on the current time.");
+
+GTEST_DEFINE_int32_(
+    repeat,
+    internal::Int32FromGTestEnv("repeat", 1),
+    "How many times to repeat each test.  Specify a negative number "
+    "for repeating forever.  Useful for shaking out flaky tests.");
+
+GTEST_DEFINE_bool_(
+    show_internal_stack_frames, false,
+    "True iff " GTEST_NAME_ " should include internal stack frames when "
+    "printing test failure stack traces.");
+
+GTEST_DEFINE_bool_(
+    shuffle,
+    internal::BoolFromGTestEnv("shuffle", false),
+    "True iff " GTEST_NAME_
+    " should randomize tests' order on every run.");
+
+GTEST_DEFINE_int32_(
+    stack_trace_depth,
+    internal::Int32FromGTestEnv("stack_trace_depth", kMaxStackTraceDepth),
+    "The maximum number of stack frames to print when an "
+    "assertion fails.  The valid range is 0 through 100, inclusive.");
+
+GTEST_DEFINE_string_(
+    stream_result_to,
+    internal::StringFromGTestEnv("stream_result_to", ""),
+    "This flag specifies the host name and the port number on which to stream "
+    "test results. Example: \"localhost:555\". The flag is effective only on "
+    "Linux.");
+
+GTEST_DEFINE_bool_(
+    throw_on_failure,
+    internal::BoolFromGTestEnv("throw_on_failure", false),
+    "When this flag is specified, a failed assertion will throw an exception "
+    "if exceptions are enabled or exit the program with a non-zero code "
+    "otherwise.");
+
+namespace internal {
+
+// Generates a random number from [0, range), using a Linear
+// Congruential Generator (LCG).  Crashes if 'range' is 0 or greater
+// than kMaxRange.
+UInt32 Random::Generate(UInt32 range) {
+  // These constants are the same as are used in glibc's rand(3).
+  state_ = (1103515245U*state_ + 12345U) % kMaxRange;
+
+  GTEST_CHECK_(range > 0)
+      << "Cannot generate a number in the range [0, 0).";
+  GTEST_CHECK_(range <= kMaxRange)
+      << "Generation of a number in [0, " << range << ") was requested, "
+      << "but this can only generate numbers in [0, " << kMaxRange << ").";
+
+  // Converting via modulus introduces a bit of downward bias, but
+  // it's simple, and a linear congruential generator isn't too good
+  // to begin with.
+  return state_ % range;
+}
+
+// GTestIsInitialized() returns true iff the user has initialized
+// Google Test.  Useful for catching the user mistake of not initializing
+// Google Test before calling RUN_ALL_TESTS().
+//
+// A user must call testing::InitGoogleTest() to initialize Google
+// Test.  g_init_gtest_count is set to the number of times
+// InitGoogleTest() has been called.  We don't protect this variable
+// under a mutex as it is only accessed in the main thread.
+GTEST_API_ int g_init_gtest_count = 0;
+static bool GTestIsInitialized() { return g_init_gtest_count != 0; }
+
+// Iterates over a vector of TestCases, keeping a running sum of the
+// results of calling a given int-returning method on each.
+// Returns the sum.
+static int SumOverTestCaseList(const std::vector& case_list,
+                               int (TestCase::*method)() const) {
+  int sum = 0;
+  for (size_t i = 0; i < case_list.size(); i++) {
+    sum += (case_list[i]->*method)();
+  }
+  return sum;
+}
+
+// Returns true iff the test case passed.
+static bool TestCasePassed(const TestCase* test_case) {
+  return test_case->should_run() && test_case->Passed();
+}
+
+// Returns true iff the test case failed.
+static bool TestCaseFailed(const TestCase* test_case) {
+  return test_case->should_run() && test_case->Failed();
+}
+
+// Returns true iff test_case contains at least one test that should
+// run.
+static bool ShouldRunTestCase(const TestCase* test_case) {
+  return test_case->should_run();
+}
+
+// AssertHelper constructor.
+AssertHelper::AssertHelper(TestPartResult::Type type,
+                           const char* file,
+                           int line,
+                           const char* message)
+    : data_(new AssertHelperData(type, file, line, message)) {
+}
+
+AssertHelper::~AssertHelper() {
+  delete data_;
+}
+
+// Message assignment, for assertion streaming support.
+void AssertHelper::operator=(const Message& message) const {
+  UnitTest::GetInstance()->
+    AddTestPartResult(data_->type, data_->file, data_->line,
+                      AppendUserMessage(data_->message, message),
+                      UnitTest::GetInstance()->impl()
+                      ->CurrentOsStackTraceExceptTop(1)
+                      // Skips the stack frame for this function itself.
+                      );  // NOLINT
+}
+
+// Mutex for linked pointers.
+GTEST_API_ GTEST_DEFINE_STATIC_MUTEX_(g_linked_ptr_mutex);
+
+// Application pathname gotten in InitGoogleTest.
+std::string g_executable_path;
+
+// Returns the current application's name, removing directory path if that
+// is present.
+FilePath GetCurrentExecutableName() {
+  FilePath result;
+
+#if GTEST_OS_WINDOWS
+  result.Set(FilePath(g_executable_path).RemoveExtension("exe"));
+#else
+  result.Set(FilePath(g_executable_path));
+#endif  // GTEST_OS_WINDOWS
+
+  return result.RemoveDirectoryName();
+}
+
+// Functions for processing the gtest_output flag.
+
+// Returns the output format, or "" for normal printed output.
+std::string UnitTestOptions::GetOutputFormat() {
+  const char* const gtest_output_flag = GTEST_FLAG(output).c_str();
+  if (gtest_output_flag == NULL) return std::string("");
+
+  const char* const colon = strchr(gtest_output_flag, ':');
+  return (colon == NULL) ?
+      std::string(gtest_output_flag) :
+      std::string(gtest_output_flag, colon - gtest_output_flag);
+}
+
+// Returns the name of the requested output file, or the default if none
+// was explicitly specified.
+std::string UnitTestOptions::GetAbsolutePathToOutputFile() {
+  const char* const gtest_output_flag = GTEST_FLAG(output).c_str();
+  if (gtest_output_flag == NULL)
+    return "";
+
+  const char* const colon = strchr(gtest_output_flag, ':');
+  if (colon == NULL)
+    return internal::FilePath::ConcatPaths(
+        internal::FilePath(
+            UnitTest::GetInstance()->original_working_dir()),
+        internal::FilePath(kDefaultOutputFile)).string();
+
+  internal::FilePath output_name(colon + 1);
+  if (!output_name.IsAbsolutePath())
+    // TODO(wan@google.com): on Windows \some\path is not an absolute
+    // path (as its meaning depends on the current drive), yet the
+    // following logic for turning it into an absolute path is wrong.
+    // Fix it.
+    output_name = internal::FilePath::ConcatPaths(
+        internal::FilePath(UnitTest::GetInstance()->original_working_dir()),
+        internal::FilePath(colon + 1));
+
+  if (!output_name.IsDirectory())
+    return output_name.string();
+
+  internal::FilePath result(internal::FilePath::GenerateUniqueFileName(
+      output_name, internal::GetCurrentExecutableName(),
+      GetOutputFormat().c_str()));
+  return result.string();
+}
+
+// Returns true iff the wildcard pattern matches the string.  The
+// first ':' or '\0' character in pattern marks the end of it.
+//
+// This recursive algorithm isn't very efficient, but is clear and
+// works well enough for matching test names, which are short.
+bool UnitTestOptions::PatternMatchesString(const char *pattern,
+                                           const char *str) {
+  switch (*pattern) {
+    case '\0':
+    case ':':  // Either ':' or '\0' marks the end of the pattern.
+      return *str == '\0';
+    case '?':  // Matches any single character.
+      return *str != '\0' && PatternMatchesString(pattern + 1, str + 1);
+    case '*':  // Matches any string (possibly empty) of characters.
+      return (*str != '\0' && PatternMatchesString(pattern, str + 1)) ||
+          PatternMatchesString(pattern + 1, str);
+    default:  // Non-special character.  Matches itself.
+      return *pattern == *str &&
+          PatternMatchesString(pattern + 1, str + 1);
+  }
+}
+
+bool UnitTestOptions::MatchesFilter(
+    const std::string& name, const char* filter) {
+  const char *cur_pattern = filter;
+  for (;;) {
+    if (PatternMatchesString(cur_pattern, name.c_str())) {
+      return true;
+    }
+
+    // Finds the next pattern in the filter.
+    cur_pattern = strchr(cur_pattern, ':');
+
+    // Returns if no more pattern can be found.
+    if (cur_pattern == NULL) {
+      return false;
+    }
+
+    // Skips the pattern separater (the ':' character).
+    cur_pattern++;
+  }
+}
+
+// Returns true iff the user-specified filter matches the test case
+// name and the test name.
+bool UnitTestOptions::FilterMatchesTest(const std::string &test_case_name,
+                                        const std::string &test_name) {
+  const std::string& full_name = test_case_name + "." + test_name.c_str();
+
+  // Split --gtest_filter at '-', if there is one, to separate into
+  // positive filter and negative filter portions
+  const char* const p = GTEST_FLAG(filter).c_str();
+  const char* const dash = strchr(p, '-');
+  std::string positive;
+  std::string negative;
+  if (dash == NULL) {
+    positive = GTEST_FLAG(filter).c_str();  // Whole string is a positive filter
+    negative = "";
+  } else {
+    positive = std::string(p, dash);   // Everything up to the dash
+    negative = std::string(dash + 1);  // Everything after the dash
+    if (positive.empty()) {
+      // Treat '-test1' as the same as '*-test1'
+      positive = kUniversalFilter;
+    }
+  }
+
+  // A filter is a colon-separated list of patterns.  It matches a
+  // test if any pattern in it matches the test.
+  return (MatchesFilter(full_name, positive.c_str()) &&
+          !MatchesFilter(full_name, negative.c_str()));
+}
+
+#if GTEST_HAS_SEH
+// Returns EXCEPTION_EXECUTE_HANDLER if Google Test should handle the
+// given SEH exception, or EXCEPTION_CONTINUE_SEARCH otherwise.
+// This function is useful as an __except condition.
+int UnitTestOptions::GTestShouldProcessSEH(DWORD exception_code) {
+  // Google Test should handle a SEH exception if:
+  //   1. the user wants it to, AND
+  //   2. this is not a breakpoint exception, AND
+  //   3. this is not a C++ exception (VC++ implements them via SEH,
+  //      apparently).
+  //
+  // SEH exception code for C++ exceptions.
+  // (see http://support.microsoft.com/kb/185294 for more information).
+  const DWORD kCxxExceptionCode = 0xe06d7363;
+
+  bool should_handle = true;
+
+  if (!GTEST_FLAG(catch_exceptions))
+    should_handle = false;
+  else if (exception_code == EXCEPTION_BREAKPOINT)
+    should_handle = false;
+  else if (exception_code == kCxxExceptionCode)
+    should_handle = false;
+
+  return should_handle ? EXCEPTION_EXECUTE_HANDLER : EXCEPTION_CONTINUE_SEARCH;
+}
+#endif  // GTEST_HAS_SEH
+
+}  // namespace internal
+
+// The c'tor sets this object as the test part result reporter used by
+// Google Test.  The 'result' parameter specifies where to report the
+// results. Intercepts only failures from the current thread.
+ScopedFakeTestPartResultReporter::ScopedFakeTestPartResultReporter(
+    TestPartResultArray* result)
+    : intercept_mode_(INTERCEPT_ONLY_CURRENT_THREAD),
+      result_(result) {
+  Init();
+}
+
+// The c'tor sets this object as the test part result reporter used by
+// Google Test.  The 'result' parameter specifies where to report the
+// results.
+ScopedFakeTestPartResultReporter::ScopedFakeTestPartResultReporter(
+    InterceptMode intercept_mode, TestPartResultArray* result)
+    : intercept_mode_(intercept_mode),
+      result_(result) {
+  Init();
+}
+
+void ScopedFakeTestPartResultReporter::Init() {
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  if (intercept_mode_ == INTERCEPT_ALL_THREADS) {
+    old_reporter_ = impl->GetGlobalTestPartResultReporter();
+    impl->SetGlobalTestPartResultReporter(this);
+  } else {
+    old_reporter_ = impl->GetTestPartResultReporterForCurrentThread();
+    impl->SetTestPartResultReporterForCurrentThread(this);
+  }
+}
+
+// The d'tor restores the test part result reporter used by Google Test
+// before.
+ScopedFakeTestPartResultReporter::~ScopedFakeTestPartResultReporter() {
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  if (intercept_mode_ == INTERCEPT_ALL_THREADS) {
+    impl->SetGlobalTestPartResultReporter(old_reporter_);
+  } else {
+    impl->SetTestPartResultReporterForCurrentThread(old_reporter_);
+  }
+}
+
+// Increments the test part result count and remembers the result.
+// This method is from the TestPartResultReporterInterface interface.
+void ScopedFakeTestPartResultReporter::ReportTestPartResult(
+    const TestPartResult& result) {
+  result_->Append(result);
+}
+
+namespace internal {
+
+// Returns the type ID of ::testing::Test.  We should always call this
+// instead of GetTypeId< ::testing::Test>() to get the type ID of
+// testing::Test.  This is to work around a suspected linker bug when
+// using Google Test as a framework on Mac OS X.  The bug causes
+// GetTypeId< ::testing::Test>() to return different values depending
+// on whether the call is from the Google Test framework itself or
+// from user test code.  GetTestTypeId() is guaranteed to always
+// return the same value, as it always calls GetTypeId<>() from the
+// gtest.cc, which is within the Google Test framework.
+TypeId GetTestTypeId() {
+  return GetTypeId();
+}
+
+// The value of GetTestTypeId() as seen from within the Google Test
+// library.  This is solely for testing GetTestTypeId().
+extern const TypeId kTestTypeIdInGoogleTest = GetTestTypeId();
+
+// This predicate-formatter checks that 'results' contains a test part
+// failure of the given type and that the failure message contains the
+// given substring.
+AssertionResult HasOneFailure(const char* /* results_expr */,
+                              const char* /* type_expr */,
+                              const char* /* substr_expr */,
+                              const TestPartResultArray& results,
+                              TestPartResult::Type type,
+                              const string& substr) {
+  const std::string expected(type == TestPartResult::kFatalFailure ?
+                        "1 fatal failure" :
+                        "1 non-fatal failure");
+  Message msg;
+  if (results.size() != 1) {
+    msg << "Expected: " << expected << "\n"
+        << "  Actual: " << results.size() << " failures";
+    for (int i = 0; i < results.size(); i++) {
+      msg << "\n" << results.GetTestPartResult(i);
+    }
+    return AssertionFailure() << msg;
+  }
+
+  const TestPartResult& r = results.GetTestPartResult(0);
+  if (r.type() != type) {
+    return AssertionFailure() << "Expected: " << expected << "\n"
+                              << "  Actual:\n"
+                              << r;
+  }
+
+  if (strstr(r.message(), substr.c_str()) == NULL) {
+    return AssertionFailure() << "Expected: " << expected << " containing \""
+                              << substr << "\"\n"
+                              << "  Actual:\n"
+                              << r;
+  }
+
+  return AssertionSuccess();
+}
+
+// The constructor of SingleFailureChecker remembers where to look up
+// test part results, what type of failure we expect, and what
+// substring the failure message should contain.
+SingleFailureChecker:: SingleFailureChecker(
+    const TestPartResultArray* results,
+    TestPartResult::Type type,
+    const string& substr)
+    : results_(results),
+      type_(type),
+      substr_(substr) {}
+
+// The destructor of SingleFailureChecker verifies that the given
+// TestPartResultArray contains exactly one failure that has the given
+// type and contains the given substring.  If that's not the case, a
+// non-fatal failure will be generated.
+SingleFailureChecker::~SingleFailureChecker() {
+  EXPECT_PRED_FORMAT3(HasOneFailure, *results_, type_, substr_);
+}
+
+DefaultGlobalTestPartResultReporter::DefaultGlobalTestPartResultReporter(
+    UnitTestImpl* unit_test) : unit_test_(unit_test) {}
+
+void DefaultGlobalTestPartResultReporter::ReportTestPartResult(
+    const TestPartResult& result) {
+  unit_test_->current_test_result()->AddTestPartResult(result);
+  unit_test_->listeners()->repeater()->OnTestPartResult(result);
+}
+
+DefaultPerThreadTestPartResultReporter::DefaultPerThreadTestPartResultReporter(
+    UnitTestImpl* unit_test) : unit_test_(unit_test) {}
+
+void DefaultPerThreadTestPartResultReporter::ReportTestPartResult(
+    const TestPartResult& result) {
+  unit_test_->GetGlobalTestPartResultReporter()->ReportTestPartResult(result);
+}
+
+// Returns the global test part result reporter.
+TestPartResultReporterInterface*
+UnitTestImpl::GetGlobalTestPartResultReporter() {
+  internal::MutexLock lock(&global_test_part_result_reporter_mutex_);
+  return global_test_part_result_repoter_;
+}
+
+// Sets the global test part result reporter.
+void UnitTestImpl::SetGlobalTestPartResultReporter(
+    TestPartResultReporterInterface* reporter) {
+  internal::MutexLock lock(&global_test_part_result_reporter_mutex_);
+  global_test_part_result_repoter_ = reporter;
+}
+
+// Returns the test part result reporter for the current thread.
+TestPartResultReporterInterface*
+UnitTestImpl::GetTestPartResultReporterForCurrentThread() {
+  return per_thread_test_part_result_reporter_.get();
+}
+
+// Sets the test part result reporter for the current thread.
+void UnitTestImpl::SetTestPartResultReporterForCurrentThread(
+    TestPartResultReporterInterface* reporter) {
+  per_thread_test_part_result_reporter_.set(reporter);
+}
+
+// Gets the number of successful test cases.
+int UnitTestImpl::successful_test_case_count() const {
+  return CountIf(test_cases_, TestCasePassed);
+}
+
+// Gets the number of failed test cases.
+int UnitTestImpl::failed_test_case_count() const {
+  return CountIf(test_cases_, TestCaseFailed);
+}
+
+// Gets the number of all test cases.
+int UnitTestImpl::total_test_case_count() const {
+  return static_cast(test_cases_.size());
+}
+
+// Gets the number of all test cases that contain at least one test
+// that should run.
+int UnitTestImpl::test_case_to_run_count() const {
+  return CountIf(test_cases_, ShouldRunTestCase);
+}
+
+// Gets the number of successful tests.
+int UnitTestImpl::successful_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::successful_test_count);
+}
+
+// Gets the number of failed tests.
+int UnitTestImpl::failed_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::failed_test_count);
+}
+
+// Gets the number of disabled tests that will be reported in the XML report.
+int UnitTestImpl::reportable_disabled_test_count() const {
+  return SumOverTestCaseList(test_cases_,
+                             &TestCase::reportable_disabled_test_count);
+}
+
+// Gets the number of disabled tests.
+int UnitTestImpl::disabled_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::disabled_test_count);
+}
+
+// Gets the number of tests to be printed in the XML report.
+int UnitTestImpl::reportable_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::reportable_test_count);
+}
+
+// Gets the number of all tests.
+int UnitTestImpl::total_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::total_test_count);
+}
+
+// Gets the number of tests that should run.
+int UnitTestImpl::test_to_run_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::test_to_run_count);
+}
+
+// Returns the current OS stack trace as an std::string.
+//
+// The maximum number of stack frames to be included is specified by
+// the gtest_stack_trace_depth flag.  The skip_count parameter
+// specifies the number of top frames to be skipped, which doesn't
+// count against the number of frames to be included.
+//
+// For example, if Foo() calls Bar(), which in turn calls
+// CurrentOsStackTraceExceptTop(1), Foo() will be included in the
+// trace but Bar() and CurrentOsStackTraceExceptTop() won't.
+std::string UnitTestImpl::CurrentOsStackTraceExceptTop(int skip_count) {
+  (void)skip_count;
+  return "";
+}
+
+// Returns the current time in milliseconds.
+TimeInMillis GetTimeInMillis() {
+#if GTEST_OS_WINDOWS_MOBILE || defined(__BORLANDC__)
+  // Difference between 1970-01-01 and 1601-01-01 in milliseconds.
+  // http://analogous.blogspot.com/2005/04/epoch.html
+  const TimeInMillis kJavaEpochToWinFileTimeDelta =
+    static_cast(116444736UL) * 100000UL;
+  const DWORD kTenthMicrosInMilliSecond = 10000;
+
+  SYSTEMTIME now_systime;
+  FILETIME now_filetime;
+  ULARGE_INTEGER now_int64;
+  // TODO(kenton@google.com): Shouldn't this just use
+  //   GetSystemTimeAsFileTime()?
+  GetSystemTime(&now_systime);
+  if (SystemTimeToFileTime(&now_systime, &now_filetime)) {
+    now_int64.LowPart = now_filetime.dwLowDateTime;
+    now_int64.HighPart = now_filetime.dwHighDateTime;
+    now_int64.QuadPart = (now_int64.QuadPart / kTenthMicrosInMilliSecond) -
+      kJavaEpochToWinFileTimeDelta;
+    return now_int64.QuadPart;
+  }
+  return 0;
+#elif GTEST_OS_WINDOWS && !GTEST_HAS_GETTIMEOFDAY_
+  __timeb64 now;
+
+# ifdef _MSC_VER
+
+  // MSVC 8 deprecates _ftime64(), so we want to suppress warning 4996
+  // (deprecated function) there.
+  // TODO(kenton@google.com): Use GetTickCount()?  Or use
+  //   SystemTimeToFileTime()
+#  pragma warning(push)          // Saves the current warning state.
+#  pragma warning(disable:4996)  // Temporarily disables warning 4996.
+  _ftime64(&now);
+#  pragma warning(pop)           // Restores the warning state.
+# else
+
+  _ftime64(&now);
+
+# endif  // _MSC_VER
+
+  return static_cast(now.time) * 1000 + now.millitm;
+#elif GTEST_HAS_GETTIMEOFDAY_
+  struct timeval now;
+  gettimeofday(&now, NULL);
+  return static_cast(now.tv_sec) * 1000 + now.tv_usec / 1000;
+#else
+# error "Don't know how to get the current time on your system."
+#endif
+}
+
+// Utilities
+
+// class String.
+
+#if GTEST_OS_WINDOWS_MOBILE
+// Creates a UTF-16 wide string from the given ANSI string, allocating
+// memory using new. The caller is responsible for deleting the return
+// value using delete[]. Returns the wide string, or NULL if the
+// input is NULL.
+LPCWSTR String::AnsiToUtf16(const char* ansi) {
+  if (!ansi) return NULL;
+  const int length = strlen(ansi);
+  const int unicode_length =
+      MultiByteToWideChar(CP_ACP, 0, ansi, length,
+                          NULL, 0);
+  WCHAR* unicode = new WCHAR[unicode_length + 1];
+  MultiByteToWideChar(CP_ACP, 0, ansi, length,
+                      unicode, unicode_length);
+  unicode[unicode_length] = 0;
+  return unicode;
+}
+
+// Creates an ANSI string from the given wide string, allocating
+// memory using new. The caller is responsible for deleting the return
+// value using delete[]. Returns the ANSI string, or NULL if the
+// input is NULL.
+const char* String::Utf16ToAnsi(LPCWSTR utf16_str)  {
+  if (!utf16_str) return NULL;
+  const int ansi_length =
+      WideCharToMultiByte(CP_ACP, 0, utf16_str, -1,
+                          NULL, 0, NULL, NULL);
+  char* ansi = new char[ansi_length + 1];
+  WideCharToMultiByte(CP_ACP, 0, utf16_str, -1,
+                      ansi, ansi_length, NULL, NULL);
+  ansi[ansi_length] = 0;
+  return ansi;
+}
+
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+// Compares two C strings.  Returns true iff they have the same content.
+//
+// Unlike strcmp(), this function can handle NULL argument(s).  A NULL
+// C string is considered different to any non-NULL C string,
+// including the empty string.
+bool String::CStringEquals(const char * lhs, const char * rhs) {
+  if ( lhs == NULL ) return rhs == NULL;
+
+  if ( rhs == NULL ) return false;
+
+  return strcmp(lhs, rhs) == 0;
+}
+
+#if GTEST_HAS_STD_WSTRING || GTEST_HAS_GLOBAL_WSTRING
+
+// Converts an array of wide chars to a narrow string using the UTF-8
+// encoding, and streams the result to the given Message object.
+static void StreamWideCharsToMessage(const wchar_t* wstr, size_t length,
+                                     Message* msg) {
+  for (size_t i = 0; i != length; ) {  // NOLINT
+    if (wstr[i] != L'\0') {
+      *msg << WideStringToUtf8(wstr + i, static_cast(length - i));
+      while (i != length && wstr[i] != L'\0')
+        i++;
+    } else {
+      *msg << '\0';
+      i++;
+    }
+  }
+}
+
+#endif  // GTEST_HAS_STD_WSTRING || GTEST_HAS_GLOBAL_WSTRING
+
+}  // namespace internal
+
+// Constructs an empty Message.
+// We allocate the stringstream separately because otherwise each use of
+// ASSERT/EXPECT in a procedure adds over 200 bytes to the procedure's
+// stack frame leading to huge stack frames in some cases; gcc does not reuse
+// the stack space.
+Message::Message() : ss_(new ::std::stringstream) {
+  // By default, we want there to be enough precision when printing
+  // a double to a Message.
+  *ss_ << std::setprecision(std::numeric_limits::digits10 + 2);
+}
+
+// These two overloads allow streaming a wide C string to a Message
+// using the UTF-8 encoding.
+Message& Message::operator <<(const wchar_t* wide_c_str) {
+  return *this << internal::String::ShowWideCString(wide_c_str);
+}
+Message& Message::operator <<(wchar_t* wide_c_str) {
+  return *this << internal::String::ShowWideCString(wide_c_str);
+}
+
+#if GTEST_HAS_STD_WSTRING
+// Converts the given wide string to a narrow string using the UTF-8
+// encoding, and streams the result to this Message object.
+Message& Message::operator <<(const ::std::wstring& wstr) {
+  internal::StreamWideCharsToMessage(wstr.c_str(), wstr.length(), this);
+  return *this;
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+#if GTEST_HAS_GLOBAL_WSTRING
+// Converts the given wide string to a narrow string using the UTF-8
+// encoding, and streams the result to this Message object.
+Message& Message::operator <<(const ::wstring& wstr) {
+  internal::StreamWideCharsToMessage(wstr.c_str(), wstr.length(), this);
+  return *this;
+}
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+// Gets the text streamed to this object so far as an std::string.
+// Each '\0' character in the buffer is replaced with "\\0".
+std::string Message::GetString() const {
+  return internal::StringStreamToString(ss_.get());
+}
+
+// AssertionResult constructors.
+// Used in EXPECT_TRUE/FALSE(assertion_result).
+AssertionResult::AssertionResult(const AssertionResult& other)
+    : success_(other.success_),
+      message_(other.message_.get() != NULL ?
+               new ::std::string(*other.message_) :
+               static_cast< ::std::string*>(NULL)) {
+}
+
+// Returns the assertion's negation. Used with EXPECT/ASSERT_FALSE.
+AssertionResult AssertionResult::operator!() const {
+  AssertionResult negation(!success_);
+  if (message_.get() != NULL)
+    negation << *message_;
+  return negation;
+}
+
+// Makes a successful assertion result.
+AssertionResult AssertionSuccess() {
+  return AssertionResult(true);
+}
+
+// Makes a failed assertion result.
+AssertionResult AssertionFailure() {
+  return AssertionResult(false);
+}
+
+// Makes a failed assertion result with the given failure message.
+// Deprecated; use AssertionFailure() << message.
+AssertionResult AssertionFailure(const Message& message) {
+  return AssertionFailure() << message;
+}
+
+namespace internal {
+
+// Constructs and returns the message for an equality assertion
+// (e.g. ASSERT_EQ, EXPECT_STREQ, etc) failure.
+//
+// The first four parameters are the expressions used in the assertion
+// and their values, as strings.  For example, for ASSERT_EQ(foo, bar)
+// where foo is 5 and bar is 6, we have:
+//
+//   expected_expression: "foo"
+//   actual_expression:   "bar"
+//   expected_value:      "5"
+//   actual_value:        "6"
+//
+// The ignoring_case parameter is true iff the assertion is a
+// *_STRCASEEQ*.  When it's true, the string " (ignoring case)" will
+// be inserted into the message.
+AssertionResult EqFailure(const char* expected_expression,
+                          const char* actual_expression,
+                          const std::string& expected_value,
+                          const std::string& actual_value,
+                          bool ignoring_case) {
+  Message msg;
+  msg << "Value of: " << actual_expression;
+  if (actual_value != actual_expression) {
+    msg << "\n  Actual: " << actual_value;
+  }
+
+  msg << "\nExpected: " << expected_expression;
+  if (ignoring_case) {
+    msg << " (ignoring case)";
+  }
+  if (expected_value != expected_expression) {
+    msg << "\nWhich is: " << expected_value;
+  }
+
+  return AssertionFailure() << msg;
+}
+
+// Constructs a failure message for Boolean assertions such as EXPECT_TRUE.
+std::string GetBoolAssertionFailureMessage(
+    const AssertionResult& assertion_result,
+    const char* expression_text,
+    const char* actual_predicate_value,
+    const char* expected_predicate_value) {
+  const char* actual_message = assertion_result.message();
+  Message msg;
+  msg << "Value of: " << expression_text
+      << "\n  Actual: " << actual_predicate_value;
+  if (actual_message[0] != '\0')
+    msg << " (" << actual_message << ")";
+  msg << "\nExpected: " << expected_predicate_value;
+  return msg.GetString();
+}
+
+// Helper function for implementing ASSERT_NEAR.
+AssertionResult DoubleNearPredFormat(const char* expr1,
+                                     const char* expr2,
+                                     const char* abs_error_expr,
+                                     double val1,
+                                     double val2,
+                                     double abs_error) {
+  const double diff = fabs(val1 - val2);
+  if (diff <= abs_error) return AssertionSuccess();
+
+  // TODO(wan): do not print the value of an expression if it's
+  // already a literal.
+  return AssertionFailure()
+      << "The difference between " << expr1 << " and " << expr2
+      << " is " << diff << ", which exceeds " << abs_error_expr << ", where\n"
+      << expr1 << " evaluates to " << val1 << ",\n"
+      << expr2 << " evaluates to " << val2 << ", and\n"
+      << abs_error_expr << " evaluates to " << abs_error << ".";
+}
+
+
+// Helper template for implementing FloatLE() and DoubleLE().
+template 
+AssertionResult FloatingPointLE(const char* expr1,
+                                const char* expr2,
+                                RawType val1,
+                                RawType val2) {
+  // Returns success if val1 is less than val2,
+  if (val1 < val2) {
+    return AssertionSuccess();
+  }
+
+  // or if val1 is almost equal to val2.
+  const FloatingPoint lhs(val1), rhs(val2);
+  if (lhs.AlmostEquals(rhs)) {
+    return AssertionSuccess();
+  }
+
+  // Note that the above two checks will both fail if either val1 or
+  // val2 is NaN, as the IEEE floating-point standard requires that
+  // any predicate involving a NaN must return false.
+
+  ::std::stringstream val1_ss;
+  val1_ss << std::setprecision(std::numeric_limits::digits10 + 2)
+          << val1;
+
+  ::std::stringstream val2_ss;
+  val2_ss << std::setprecision(std::numeric_limits::digits10 + 2)
+          << val2;
+
+  return AssertionFailure()
+      << "Expected: (" << expr1 << ") <= (" << expr2 << ")\n"
+      << "  Actual: " << StringStreamToString(&val1_ss) << " vs "
+      << StringStreamToString(&val2_ss);
+}
+
+}  // namespace internal
+
+// Asserts that val1 is less than, or almost equal to, val2.  Fails
+// otherwise.  In particular, it fails if either val1 or val2 is NaN.
+AssertionResult FloatLE(const char* expr1, const char* expr2,
+                        float val1, float val2) {
+  return internal::FloatingPointLE(expr1, expr2, val1, val2);
+}
+
+// Asserts that val1 is less than, or almost equal to, val2.  Fails
+// otherwise.  In particular, it fails if either val1 or val2 is NaN.
+AssertionResult DoubleLE(const char* expr1, const char* expr2,
+                         double val1, double val2) {
+  return internal::FloatingPointLE(expr1, expr2, val1, val2);
+}
+
+namespace internal {
+
+// The helper function for {ASSERT|EXPECT}_EQ with int or enum
+// arguments.
+AssertionResult CmpHelperEQ(const char* expected_expression,
+                            const char* actual_expression,
+                            BiggestInt expected,
+                            BiggestInt actual) {
+  if (expected == actual) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   FormatForComparisonFailureMessage(expected, actual),
+                   FormatForComparisonFailureMessage(actual, expected),
+                   false);
+}
+
+// A macro for implementing the helper functions needed to implement
+// ASSERT_?? and EXPECT_?? with integer or enum arguments.  It is here
+// just to avoid copy-and-paste of similar code.
+#define GTEST_IMPL_CMP_HELPER_(op_name, op)\
+AssertionResult CmpHelper##op_name(const char* expr1, const char* expr2, \
+                                   BiggestInt val1, BiggestInt val2) {\
+  if (val1 op val2) {\
+    return AssertionSuccess();\
+  } else {\
+    return AssertionFailure() \
+        << "Expected: (" << expr1 << ") " #op " (" << expr2\
+        << "), actual: " << FormatForComparisonFailureMessage(val1, val2)\
+        << " vs " << FormatForComparisonFailureMessage(val2, val1);\
+  }\
+}
+
+// Implements the helper function for {ASSERT|EXPECT}_NE with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(NE, !=)
+// Implements the helper function for {ASSERT|EXPECT}_LE with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(LE, <=)
+// Implements the helper function for {ASSERT|EXPECT}_LT with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(LT, < )
+// Implements the helper function for {ASSERT|EXPECT}_GE with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(GE, >=)
+// Implements the helper function for {ASSERT|EXPECT}_GT with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(GT, > )
+
+#undef GTEST_IMPL_CMP_HELPER_
+
+// The helper function for {ASSERT|EXPECT}_STREQ.
+AssertionResult CmpHelperSTREQ(const char* expected_expression,
+                               const char* actual_expression,
+                               const char* expected,
+                               const char* actual) {
+  if (String::CStringEquals(expected, actual)) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   PrintToString(expected),
+                   PrintToString(actual),
+                   false);
+}
+
+// The helper function for {ASSERT|EXPECT}_STRCASEEQ.
+AssertionResult CmpHelperSTRCASEEQ(const char* expected_expression,
+                                   const char* actual_expression,
+                                   const char* expected,
+                                   const char* actual) {
+  if (String::CaseInsensitiveCStringEquals(expected, actual)) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   PrintToString(expected),
+                   PrintToString(actual),
+                   true);
+}
+
+// The helper function for {ASSERT|EXPECT}_STRNE.
+AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                               const char* s2_expression,
+                               const char* s1,
+                               const char* s2) {
+  if (!String::CStringEquals(s1, s2)) {
+    return AssertionSuccess();
+  } else {
+    return AssertionFailure() << "Expected: (" << s1_expression << ") != ("
+                              << s2_expression << "), actual: \""
+                              << s1 << "\" vs \"" << s2 << "\"";
+  }
+}
+
+// The helper function for {ASSERT|EXPECT}_STRCASENE.
+AssertionResult CmpHelperSTRCASENE(const char* s1_expression,
+                                   const char* s2_expression,
+                                   const char* s1,
+                                   const char* s2) {
+  if (!String::CaseInsensitiveCStringEquals(s1, s2)) {
+    return AssertionSuccess();
+  } else {
+    return AssertionFailure()
+        << "Expected: (" << s1_expression << ") != ("
+        << s2_expression << ") (ignoring case), actual: \""
+        << s1 << "\" vs \"" << s2 << "\"";
+  }
+}
+
+}  // namespace internal
+
+namespace {
+
+// Helper functions for implementing IsSubString() and IsNotSubstring().
+
+// This group of overloaded functions return true iff needle is a
+// substring of haystack.  NULL is considered a substring of itself
+// only.
+
+bool IsSubstringPred(const char* needle, const char* haystack) {
+  if (needle == NULL || haystack == NULL)
+    return needle == haystack;
+
+  return strstr(haystack, needle) != NULL;
+}
+
+bool IsSubstringPred(const wchar_t* needle, const wchar_t* haystack) {
+  if (needle == NULL || haystack == NULL)
+    return needle == haystack;
+
+  return wcsstr(haystack, needle) != NULL;
+}
+
+// StringType here can be either ::std::string or ::std::wstring.
+template 
+bool IsSubstringPred(const StringType& needle,
+                     const StringType& haystack) {
+  return haystack.find(needle) != StringType::npos;
+}
+
+// This function implements either IsSubstring() or IsNotSubstring(),
+// depending on the value of the expected_to_be_substring parameter.
+// StringType here can be const char*, const wchar_t*, ::std::string,
+// or ::std::wstring.
+template 
+AssertionResult IsSubstringImpl(
+    bool expected_to_be_substring,
+    const char* needle_expr, const char* haystack_expr,
+    const StringType& needle, const StringType& haystack) {
+  if (IsSubstringPred(needle, haystack) == expected_to_be_substring)
+    return AssertionSuccess();
+
+  const bool is_wide_string = sizeof(needle[0]) > 1;
+  const char* const begin_string_quote = is_wide_string ? "L\"" : "\"";
+  return AssertionFailure()
+      << "Value of: " << needle_expr << "\n"
+      << "  Actual: " << begin_string_quote << needle << "\"\n"
+      << "Expected: " << (expected_to_be_substring ? "" : "not ")
+      << "a substring of " << haystack_expr << "\n"
+      << "Which is: " << begin_string_quote << haystack << "\"";
+}
+
+}  // namespace
+
+// IsSubstring() and IsNotSubstring() check whether needle is a
+// substring of haystack (NULL is considered a substring of itself
+// only), and return an appropriate error message when they fail.
+
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+
+#if GTEST_HAS_STD_WSTRING
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+namespace internal {
+
+#if GTEST_OS_WINDOWS
+
+namespace {
+
+// Helper function for IsHRESULT{SuccessFailure} predicates
+AssertionResult HRESULTFailureHelper(const char* expr,
+                                     const char* expected,
+                                     long hr) {  // NOLINT
+# if GTEST_OS_WINDOWS_MOBILE
+
+  // Windows CE doesn't support FormatMessage.
+  const char error_text[] = "";
+
+# else
+
+  // Looks up the human-readable system message for the HRESULT code
+  // and since we're not passing any params to FormatMessage, we don't
+  // want inserts expanded.
+  const DWORD kFlags = FORMAT_MESSAGE_FROM_SYSTEM |
+                       FORMAT_MESSAGE_IGNORE_INSERTS;
+  const DWORD kBufSize = 4096;
+  // Gets the system's human readable message string for this HRESULT.
+  char error_text[kBufSize] = { '\0' };
+  DWORD message_length = ::FormatMessageA(kFlags,
+                                          0,  // no source, we're asking system
+                                          hr,  // the error
+                                          0,  // no line width restrictions
+                                          error_text,  // output buffer
+                                          kBufSize,  // buf size
+                                          NULL);  // no arguments for inserts
+  // Trims tailing white space (FormatMessage leaves a trailing CR-LF)
+  for (; message_length && IsSpace(error_text[message_length - 1]);
+          --message_length) {
+    error_text[message_length - 1] = '\0';
+  }
+
+# endif  // GTEST_OS_WINDOWS_MOBILE
+
+  const std::string error_hex("0x" + String::FormatHexInt(hr));
+  return ::testing::AssertionFailure()
+      << "Expected: " << expr << " " << expected << ".\n"
+      << "  Actual: " << error_hex << " " << error_text << "\n";
+}
+
+}  // namespace
+
+AssertionResult IsHRESULTSuccess(const char* expr, long hr) {  // NOLINT
+  if (SUCCEEDED(hr)) {
+    return AssertionSuccess();
+  }
+  return HRESULTFailureHelper(expr, "succeeds", hr);
+}
+
+AssertionResult IsHRESULTFailure(const char* expr, long hr) {  // NOLINT
+  if (FAILED(hr)) {
+    return AssertionSuccess();
+  }
+  return HRESULTFailureHelper(expr, "fails", hr);
+}
+
+#endif  // GTEST_OS_WINDOWS
+
+// Utility functions for encoding Unicode text (wide strings) in
+// UTF-8.
+
+// A Unicode code-point can have upto 21 bits, and is encoded in UTF-8
+// like this:
+//
+// Code-point length   Encoding
+//   0 -  7 bits       0xxxxxxx
+//   8 - 11 bits       110xxxxx 10xxxxxx
+//  12 - 16 bits       1110xxxx 10xxxxxx 10xxxxxx
+//  17 - 21 bits       11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+
+// The maximum code-point a one-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint1 = (static_cast(1) <<  7) - 1;
+
+// The maximum code-point a two-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint2 = (static_cast(1) << (5 + 6)) - 1;
+
+// The maximum code-point a three-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint3 = (static_cast(1) << (4 + 2*6)) - 1;
+
+// The maximum code-point a four-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint4 = (static_cast(1) << (3 + 3*6)) - 1;
+
+// Chops off the n lowest bits from a bit pattern.  Returns the n
+// lowest bits.  As a side effect, the original bit pattern will be
+// shifted to the right by n bits.
+inline UInt32 ChopLowBits(UInt32* bits, int n) {
+  const UInt32 low_bits = *bits & ((static_cast(1) << n) - 1);
+  *bits >>= n;
+  return low_bits;
+}
+
+// Converts a Unicode code point to a narrow string in UTF-8 encoding.
+// code_point parameter is of type UInt32 because wchar_t may not be
+// wide enough to contain a code point.
+// If the code_point is not a valid Unicode code point
+// (i.e. outside of Unicode range U+0 to U+10FFFF) it will be converted
+// to "(Invalid Unicode 0xXXXXXXXX)".
+std::string CodePointToUtf8(UInt32 code_point) {
+  if (code_point > kMaxCodePoint4) {
+    return "(Invalid Unicode 0x" + String::FormatHexInt(code_point) + ")";
+  }
+
+  char str[5];  // Big enough for the largest valid code point.
+  if (code_point <= kMaxCodePoint1) {
+    str[1] = '\0';
+    str[0] = static_cast(code_point);                          // 0xxxxxxx
+  } else if (code_point <= kMaxCodePoint2) {
+    str[2] = '\0';
+    str[1] = static_cast(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[0] = static_cast(0xC0 | code_point);                   // 110xxxxx
+  } else if (code_point <= kMaxCodePoint3) {
+    str[3] = '\0';
+    str[2] = static_cast(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[1] = static_cast(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[0] = static_cast(0xE0 | code_point);                   // 1110xxxx
+  } else {  // code_point <= kMaxCodePoint4
+    str[4] = '\0';
+    str[3] = static_cast(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[2] = static_cast(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[1] = static_cast(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[0] = static_cast(0xF0 | code_point);                   // 11110xxx
+  }
+  return str;
+}
+
+// The following two functions only make sense if the the system
+// uses UTF-16 for wide string encoding. All supported systems
+// with 16 bit wchar_t (Windows, Cygwin, Symbian OS) do use UTF-16.
+
+// Determines if the arguments constitute UTF-16 surrogate pair
+// and thus should be combined into a single Unicode code point
+// using CreateCodePointFromUtf16SurrogatePair.
+inline bool IsUtf16SurrogatePair(wchar_t first, wchar_t second) {
+  return sizeof(wchar_t) == 2 &&
+      (first & 0xFC00) == 0xD800 && (second & 0xFC00) == 0xDC00;
+}
+
+// Creates a Unicode code point from UTF16 surrogate pair.
+inline UInt32 CreateCodePointFromUtf16SurrogatePair(wchar_t first,
+                                                    wchar_t second) {
+  const UInt32 mask = (1 << 10) - 1;
+  return (sizeof(wchar_t) == 2) ?
+      (((first & mask) << 10) | (second & mask)) + 0x10000 :
+      // This function should not be called when the condition is
+      // false, but we provide a sensible default in case it is.
+      static_cast(first);
+}
+
+// Converts a wide string to a narrow string in UTF-8 encoding.
+// The wide string is assumed to have the following encoding:
+//   UTF-16 if sizeof(wchar_t) == 2 (on Windows, Cygwin, Symbian OS)
+//   UTF-32 if sizeof(wchar_t) == 4 (on Linux)
+// Parameter str points to a null-terminated wide string.
+// Parameter num_chars may additionally limit the number
+// of wchar_t characters processed. -1 is used when the entire string
+// should be processed.
+// If the string contains code points that are not valid Unicode code points
+// (i.e. outside of Unicode range U+0 to U+10FFFF) they will be output
+// as '(Invalid Unicode 0xXXXXXXXX)'. If the string is in UTF16 encoding
+// and contains invalid UTF-16 surrogate pairs, values in those pairs
+// will be encoded as individual Unicode characters from Basic Normal Plane.
+std::string WideStringToUtf8(const wchar_t* str, int num_chars) {
+  if (num_chars == -1)
+    num_chars = static_cast(wcslen(str));
+
+  ::std::stringstream stream;
+  for (int i = 0; i < num_chars; ++i) {
+    UInt32 unicode_code_point;
+
+    if (str[i] == L'\0') {
+      break;
+    } else if (i + 1 < num_chars && IsUtf16SurrogatePair(str[i], str[i + 1])) {
+      unicode_code_point = CreateCodePointFromUtf16SurrogatePair(str[i],
+                                                                 str[i + 1]);
+      i++;
+    } else {
+      unicode_code_point = static_cast(str[i]);
+    }
+
+    stream << CodePointToUtf8(unicode_code_point);
+  }
+  return StringStreamToString(&stream);
+}
+
+// Converts a wide C string to an std::string using the UTF-8 encoding.
+// NULL will be converted to "(null)".
+std::string String::ShowWideCString(const wchar_t * wide_c_str) {
+  if (wide_c_str == NULL)  return "(null)";
+
+  return internal::WideStringToUtf8(wide_c_str, -1);
+}
+
+// Compares two wide C strings.  Returns true iff they have the same
+// content.
+//
+// Unlike wcscmp(), this function can handle NULL argument(s).  A NULL
+// C string is considered different to any non-NULL C string,
+// including the empty string.
+bool String::WideCStringEquals(const wchar_t * lhs, const wchar_t * rhs) {
+  if (lhs == NULL) return rhs == NULL;
+
+  if (rhs == NULL) return false;
+
+  return wcscmp(lhs, rhs) == 0;
+}
+
+// Helper function for *_STREQ on wide strings.
+AssertionResult CmpHelperSTREQ(const char* expected_expression,
+                               const char* actual_expression,
+                               const wchar_t* expected,
+                               const wchar_t* actual) {
+  if (String::WideCStringEquals(expected, actual)) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   PrintToString(expected),
+                   PrintToString(actual),
+                   false);
+}
+
+// Helper function for *_STRNE on wide strings.
+AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                               const char* s2_expression,
+                               const wchar_t* s1,
+                               const wchar_t* s2) {
+  if (!String::WideCStringEquals(s1, s2)) {
+    return AssertionSuccess();
+  }
+
+  return AssertionFailure() << "Expected: (" << s1_expression << ") != ("
+                            << s2_expression << "), actual: "
+                            << PrintToString(s1)
+                            << " vs " << PrintToString(s2);
+}
+
+// Compares two C strings, ignoring case.  Returns true iff they have
+// the same content.
+//
+// Unlike strcasecmp(), this function can handle NULL argument(s).  A
+// NULL C string is considered different to any non-NULL C string,
+// including the empty string.
+bool String::CaseInsensitiveCStringEquals(const char * lhs, const char * rhs) {
+  if (lhs == NULL)
+    return rhs == NULL;
+  if (rhs == NULL)
+    return false;
+  return posix::StrCaseCmp(lhs, rhs) == 0;
+}
+
+  // Compares two wide C strings, ignoring case.  Returns true iff they
+  // have the same content.
+  //
+  // Unlike wcscasecmp(), this function can handle NULL argument(s).
+  // A NULL C string is considered different to any non-NULL wide C string,
+  // including the empty string.
+  // NB: The implementations on different platforms slightly differ.
+  // On windows, this method uses _wcsicmp which compares according to LC_CTYPE
+  // environment variable. On GNU platform this method uses wcscasecmp
+  // which compares according to LC_CTYPE category of the current locale.
+  // On MacOS X, it uses towlower, which also uses LC_CTYPE category of the
+  // current locale.
+bool String::CaseInsensitiveWideCStringEquals(const wchar_t* lhs,
+                                              const wchar_t* rhs) {
+  if (lhs == NULL) return rhs == NULL;
+
+  if (rhs == NULL) return false;
+
+#if GTEST_OS_WINDOWS
+  return _wcsicmp(lhs, rhs) == 0;
+#elif GTEST_OS_LINUX && !GTEST_OS_LINUX_ANDROID
+  return wcscasecmp(lhs, rhs) == 0;
+#else
+  // Android, Mac OS X and Cygwin don't define wcscasecmp.
+  // Other unknown OSes may not define it either.
+  wint_t left, right;
+  do {
+    left = towlower(*lhs++);
+    right = towlower(*rhs++);
+  } while (left && left == right);
+  return left == right;
+#endif  // OS selector
+}
+
+// Returns true iff str ends with the given suffix, ignoring case.
+// Any string is considered to end with an empty suffix.
+bool String::EndsWithCaseInsensitive(
+    const std::string& str, const std::string& suffix) {
+  const size_t str_len = str.length();
+  const size_t suffix_len = suffix.length();
+  return (str_len >= suffix_len) &&
+         CaseInsensitiveCStringEquals(str.c_str() + str_len - suffix_len,
+                                      suffix.c_str());
+}
+
+// Formats an int value as "%02d".
+std::string String::FormatIntWidth2(int value) {
+  std::stringstream ss;
+  ss << std::setfill('0') << std::setw(2) << value;
+  return ss.str();
+}
+
+// Formats an int value as "%X".
+std::string String::FormatHexInt(int value) {
+  std::stringstream ss;
+  ss << std::hex << std::uppercase << value;
+  return ss.str();
+}
+
+// Formats a byte as "%02X".
+std::string String::FormatByte(unsigned char value) {
+  std::stringstream ss;
+  ss << std::setfill('0') << std::setw(2) << std::hex << std::uppercase
+     << static_cast(value);
+  return ss.str();
+}
+
+// Converts the buffer in a stringstream to an std::string, converting NUL
+// bytes to "\\0" along the way.
+std::string StringStreamToString(::std::stringstream* ss) {
+  const ::std::string& str = ss->str();
+  const char* const start = str.c_str();
+  const char* const end = start + str.length();
+
+  std::string result;
+  result.reserve(2 * (end - start));
+  for (const char* ch = start; ch != end; ++ch) {
+    if (*ch == '\0') {
+      result += "\\0";  // Replaces NUL with "\\0";
+    } else {
+      result += *ch;
+    }
+  }
+
+  return result;
+}
+
+// Appends the user-supplied message to the Google-Test-generated message.
+std::string AppendUserMessage(const std::string& gtest_msg,
+                              const Message& user_msg) {
+  // Appends the user message if it's non-empty.
+  const std::string user_msg_string = user_msg.GetString();
+  if (user_msg_string.empty()) {
+    return gtest_msg;
+  }
+
+  return gtest_msg + "\n" + user_msg_string;
+}
+
+}  // namespace internal
+
+// class TestResult
+
+// Creates an empty TestResult.
+TestResult::TestResult()
+    : death_test_count_(0),
+      elapsed_time_(0) {
+}
+
+// D'tor.
+TestResult::~TestResult() {
+}
+
+// Returns the i-th test part result among all the results. i can
+// range from 0 to total_part_count() - 1. If i is not in that range,
+// aborts the program.
+const TestPartResult& TestResult::GetTestPartResult(int i) const {
+  if (i < 0 || i >= total_part_count())
+    internal::posix::Abort();
+  return test_part_results_.at(i);
+}
+
+// Returns the i-th test property. i can range from 0 to
+// test_property_count() - 1. If i is not in that range, aborts the
+// program.
+const TestProperty& TestResult::GetTestProperty(int i) const {
+  if (i < 0 || i >= test_property_count())
+    internal::posix::Abort();
+  return test_properties_.at(i);
+}
+
+// Clears the test part results.
+void TestResult::ClearTestPartResults() {
+  test_part_results_.clear();
+}
+
+// Adds a test part result to the list.
+void TestResult::AddTestPartResult(const TestPartResult& test_part_result) {
+  test_part_results_.push_back(test_part_result);
+}
+
+// Adds a test property to the list. If a property with the same key as the
+// supplied property is already represented, the value of this test_property
+// replaces the old value for that key.
+void TestResult::RecordProperty(const std::string& xml_element,
+                                const TestProperty& test_property) {
+  if (!ValidateTestProperty(xml_element, test_property)) {
+    return;
+  }
+  internal::MutexLock lock(&test_properites_mutex_);
+  const std::vector::iterator property_with_matching_key =
+      std::find_if(test_properties_.begin(), test_properties_.end(),
+                   internal::TestPropertyKeyIs(test_property.key()));
+  if (property_with_matching_key == test_properties_.end()) {
+    test_properties_.push_back(test_property);
+    return;
+  }
+  property_with_matching_key->SetValue(test_property.value());
+}
+
+// The list of reserved attributes used in the  element of XML
+// output.
+static const char* const kReservedTestSuitesAttributes[] = {
+  "disabled",
+  "errors",
+  "failures",
+  "name",
+  "random_seed",
+  "tests",
+  "time",
+  "timestamp"
+};
+
+// The list of reserved attributes used in the  element of XML
+// output.
+static const char* const kReservedTestSuiteAttributes[] = {
+  "disabled",
+  "errors",
+  "failures",
+  "name",
+  "tests",
+  "time"
+};
+
+// The list of reserved attributes used in the  element of XML output.
+static const char* const kReservedTestCaseAttributes[] = {
+  "classname",
+  "name",
+  "status",
+  "time",
+  "type_param",
+  "value_param"
+};
+
+template 
+std::vector ArrayAsVector(const char* const (&array)[kSize]) {
+  return std::vector(array, array + kSize);
+}
+
+static std::vector GetReservedAttributesForElement(
+    const std::string& xml_element) {
+  if (xml_element == "testsuites") {
+    return ArrayAsVector(kReservedTestSuitesAttributes);
+  } else if (xml_element == "testsuite") {
+    return ArrayAsVector(kReservedTestSuiteAttributes);
+  } else if (xml_element == "testcase") {
+    return ArrayAsVector(kReservedTestCaseAttributes);
+  } else {
+    GTEST_CHECK_(false) << "Unrecognized xml_element provided: " << xml_element;
+  }
+  // This code is unreachable but some compilers may not realizes that.
+  return std::vector();
+}
+
+static std::string FormatWordList(const std::vector& words) {
+  Message word_list;
+  for (size_t i = 0; i < words.size(); ++i) {
+    if (i > 0 && words.size() > 2) {
+      word_list << ", ";
+    }
+    if (i == words.size() - 1) {
+      word_list << "and ";
+    }
+    word_list << "'" << words[i] << "'";
+  }
+  return word_list.GetString();
+}
+
+bool ValidateTestPropertyName(const std::string& property_name,
+                              const std::vector& reserved_names) {
+  if (std::find(reserved_names.begin(), reserved_names.end(), property_name) !=
+          reserved_names.end()) {
+    ADD_FAILURE() << "Reserved key used in RecordProperty(): " << property_name
+                  << " (" << FormatWordList(reserved_names)
+                  << " are reserved by " << GTEST_NAME_ << ")";
+    return false;
+  }
+  return true;
+}
+
+// Adds a failure if the key is a reserved attribute of the element named
+// xml_element.  Returns true if the property is valid.
+bool TestResult::ValidateTestProperty(const std::string& xml_element,
+                                      const TestProperty& test_property) {
+  return ValidateTestPropertyName(test_property.key(),
+                                  GetReservedAttributesForElement(xml_element));
+}
+
+// Clears the object.
+void TestResult::Clear() {
+  test_part_results_.clear();
+  test_properties_.clear();
+  death_test_count_ = 0;
+  elapsed_time_ = 0;
+}
+
+// Returns true iff the test failed.
+bool TestResult::Failed() const {
+  for (int i = 0; i < total_part_count(); ++i) {
+    if (GetTestPartResult(i).failed())
+      return true;
+  }
+  return false;
+}
+
+// Returns true iff the test part fatally failed.
+static bool TestPartFatallyFailed(const TestPartResult& result) {
+  return result.fatally_failed();
+}
+
+// Returns true iff the test fatally failed.
+bool TestResult::HasFatalFailure() const {
+  return CountIf(test_part_results_, TestPartFatallyFailed) > 0;
+}
+
+// Returns true iff the test part non-fatally failed.
+static bool TestPartNonfatallyFailed(const TestPartResult& result) {
+  return result.nonfatally_failed();
+}
+
+// Returns true iff the test has a non-fatal failure.
+bool TestResult::HasNonfatalFailure() const {
+  return CountIf(test_part_results_, TestPartNonfatallyFailed) > 0;
+}
+
+// Gets the number of all test parts.  This is the sum of the number
+// of successful test parts and the number of failed test parts.
+int TestResult::total_part_count() const {
+  return static_cast(test_part_results_.size());
+}
+
+// Returns the number of the test properties.
+int TestResult::test_property_count() const {
+  return static_cast(test_properties_.size());
+}
+
+// class Test
+
+// Creates a Test object.
+
+// The c'tor saves the values of all Google Test flags.
+Test::Test()
+    : gtest_flag_saver_(new internal::GTestFlagSaver) {
+}
+
+// The d'tor restores the values of all Google Test flags.
+Test::~Test() {
+  delete gtest_flag_saver_;
+}
+
+// Sets up the test fixture.
+//
+// A sub-class may override this.
+void Test::SetUp() {
+}
+
+// Tears down the test fixture.
+//
+// A sub-class may override this.
+void Test::TearDown() {
+}
+
+// Allows user supplied key value pairs to be recorded for later output.
+void Test::RecordProperty(const std::string& key, const std::string& value) {
+  UnitTest::GetInstance()->RecordProperty(key, value);
+}
+
+// Allows user supplied key value pairs to be recorded for later output.
+void Test::RecordProperty(const std::string& key, int value) {
+  Message value_message;
+  value_message << value;
+  RecordProperty(key, value_message.GetString().c_str());
+}
+
+namespace internal {
+
+void ReportFailureInUnknownLocation(TestPartResult::Type result_type,
+                                    const std::string& message) {
+  // This function is a friend of UnitTest and as such has access to
+  // AddTestPartResult.
+  UnitTest::GetInstance()->AddTestPartResult(
+      result_type,
+      NULL,  // No info about the source file where the exception occurred.
+      -1,    // We have no info on which line caused the exception.
+      message,
+      "");   // No stack trace, either.
+}
+
+}  // namespace internal
+
+// Google Test requires all tests in the same test case to use the same test
+// fixture class.  This function checks if the current test has the
+// same fixture class as the first test in the current test case.  If
+// yes, it returns true; otherwise it generates a Google Test failure and
+// returns false.
+bool Test::HasSameFixtureClass() {
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  const TestCase* const test_case = impl->current_test_case();
+
+  // Info about the first test in the current test case.
+  const TestInfo* const first_test_info = test_case->test_info_list()[0];
+  const internal::TypeId first_fixture_id = first_test_info->fixture_class_id_;
+  const char* const first_test_name = first_test_info->name();
+
+  // Info about the current test.
+  const TestInfo* const this_test_info = impl->current_test_info();
+  const internal::TypeId this_fixture_id = this_test_info->fixture_class_id_;
+  const char* const this_test_name = this_test_info->name();
+
+  if (this_fixture_id != first_fixture_id) {
+    // Is the first test defined using TEST?
+    const bool first_is_TEST = first_fixture_id == internal::GetTestTypeId();
+    // Is this test defined using TEST?
+    const bool this_is_TEST = this_fixture_id == internal::GetTestTypeId();
+
+    if (first_is_TEST || this_is_TEST) {
+      // The user mixed TEST and TEST_F in this test case - we'll tell
+      // him/her how to fix it.
+
+      // Gets the name of the TEST and the name of the TEST_F.  Note
+      // that first_is_TEST and this_is_TEST cannot both be true, as
+      // the fixture IDs are different for the two tests.
+      const char* const TEST_name =
+          first_is_TEST ? first_test_name : this_test_name;
+      const char* const TEST_F_name =
+          first_is_TEST ? this_test_name : first_test_name;
+
+      ADD_FAILURE()
+          << "All tests in the same test case must use the same test fixture\n"
+          << "class, so mixing TEST_F and TEST in the same test case is\n"
+          << "illegal.  In test case " << this_test_info->test_case_name()
+          << ",\n"
+          << "test " << TEST_F_name << " is defined using TEST_F but\n"
+          << "test " << TEST_name << " is defined using TEST.  You probably\n"
+          << "want to change the TEST to TEST_F or move it to another test\n"
+          << "case.";
+    } else {
+      // The user defined two fixture classes with the same name in
+      // two namespaces - we'll tell him/her how to fix it.
+      ADD_FAILURE()
+          << "All tests in the same test case must use the same test fixture\n"
+          << "class.  However, in test case "
+          << this_test_info->test_case_name() << ",\n"
+          << "you defined test " << first_test_name
+          << " and test " << this_test_name << "\n"
+          << "using two different test fixture classes.  This can happen if\n"
+          << "the two classes are from different namespaces or translation\n"
+          << "units and have the same name.  You should probably rename one\n"
+          << "of the classes to put the tests into different test cases.";
+    }
+    return false;
+  }
+
+  return true;
+}
+
+#if GTEST_HAS_SEH
+
+// Adds an "exception thrown" fatal failure to the current test.  This
+// function returns its result via an output parameter pointer because VC++
+// prohibits creation of objects with destructors on stack in functions
+// using __try (see error C2712).
+static std::string* FormatSehExceptionMessage(DWORD exception_code,
+                                              const char* location) {
+  Message message;
+  message << "SEH exception with code 0x" << std::setbase(16) <<
+    exception_code << std::setbase(10) << " thrown in " << location << ".";
+
+  return new std::string(message.GetString());
+}
+
+#endif  // GTEST_HAS_SEH
+
+namespace internal {
+
+#if GTEST_HAS_EXCEPTIONS
+
+// Adds an "exception thrown" fatal failure to the current test.
+static std::string FormatCxxExceptionMessage(const char* description,
+                                             const char* location) {
+  Message message;
+  if (description != NULL) {
+    message << "C++ exception with description \"" << description << "\"";
+  } else {
+    message << "Unknown C++ exception";
+  }
+  message << " thrown in " << location << ".";
+
+  return message.GetString();
+}
+
+static std::string PrintTestPartResultToString(
+    const TestPartResult& test_part_result);
+
+GoogleTestFailureException::GoogleTestFailureException(
+    const TestPartResult& failure)
+    : ::std::runtime_error(PrintTestPartResultToString(failure).c_str()) {}
+
+#endif  // GTEST_HAS_EXCEPTIONS
+
+// We put these helper functions in the internal namespace as IBM's xlC
+// compiler rejects the code if they were declared static.
+
+// Runs the given method and handles SEH exceptions it throws, when
+// SEH is supported; returns the 0-value for type Result in case of an
+// SEH exception.  (Microsoft compilers cannot handle SEH and C++
+// exceptions in the same function.  Therefore, we provide a separate
+// wrapper function for handling SEH exceptions.)
+template 
+Result HandleSehExceptionsInMethodIfSupported(
+    T* object, Result (T::*method)(), const char* location) {
+#if GTEST_HAS_SEH
+  __try {
+    return (object->*method)();
+  } __except (internal::UnitTestOptions::GTestShouldProcessSEH(  // NOLINT
+      GetExceptionCode())) {
+    // We create the exception message on the heap because VC++ prohibits
+    // creation of objects with destructors on stack in functions using __try
+    // (see error C2712).
+    std::string* exception_message = FormatSehExceptionMessage(
+        GetExceptionCode(), location);
+    internal::ReportFailureInUnknownLocation(TestPartResult::kFatalFailure,
+                                             *exception_message);
+    delete exception_message;
+    return static_cast(0);
+  }
+#else
+  (void)location;
+  return (object->*method)();
+#endif  // GTEST_HAS_SEH
+}
+
+// Runs the given method and catches and reports C++ and/or SEH-style
+// exceptions, if they are supported; returns the 0-value for type
+// Result in case of an SEH exception.
+template 
+Result HandleExceptionsInMethodIfSupported(
+    T* object, Result (T::*method)(), const char* location) {
+  // NOTE: The user code can affect the way in which Google Test handles
+  // exceptions by setting GTEST_FLAG(catch_exceptions), but only before
+  // RUN_ALL_TESTS() starts. It is technically possible to check the flag
+  // after the exception is caught and either report or re-throw the
+  // exception based on the flag's value:
+  //
+  // try {
+  //   // Perform the test method.
+  // } catch (...) {
+  //   if (GTEST_FLAG(catch_exceptions))
+  //     // Report the exception as failure.
+  //   else
+  //     throw;  // Re-throws the original exception.
+  // }
+  //
+  // However, the purpose of this flag is to allow the program to drop into
+  // the debugger when the exception is thrown. On most platforms, once the
+  // control enters the catch block, the exception origin information is
+  // lost and the debugger will stop the program at the point of the
+  // re-throw in this function -- instead of at the point of the original
+  // throw statement in the code under test.  For this reason, we perform
+  // the check early, sacrificing the ability to affect Google Test's
+  // exception handling in the method where the exception is thrown.
+  if (internal::GetUnitTestImpl()->catch_exceptions()) {
+#if GTEST_HAS_EXCEPTIONS
+    try {
+      return HandleSehExceptionsInMethodIfSupported(object, method, location);
+    } catch (const internal::GoogleTestFailureException&) {  // NOLINT
+      // This exception type can only be thrown by a failed Google
+      // Test assertion with the intention of letting another testing
+      // framework catch it.  Therefore we just re-throw it.
+      throw;
+    } catch (const std::exception& e) {  // NOLINT
+      internal::ReportFailureInUnknownLocation(
+          TestPartResult::kFatalFailure,
+          FormatCxxExceptionMessage(e.what(), location));
+    } catch (...) {  // NOLINT
+      internal::ReportFailureInUnknownLocation(
+          TestPartResult::kFatalFailure,
+          FormatCxxExceptionMessage(NULL, location));
+    }
+    return static_cast(0);
+#else
+    return HandleSehExceptionsInMethodIfSupported(object, method, location);
+#endif  // GTEST_HAS_EXCEPTIONS
+  } else {
+    return (object->*method)();
+  }
+}
+
+}  // namespace internal
+
+// Runs the test and updates the test result.
+void Test::Run() {
+  if (!HasSameFixtureClass()) return;
+
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(this, &Test::SetUp, "SetUp()");
+  // We will run the test only if SetUp() was successful.
+  if (!HasFatalFailure()) {
+    impl->os_stack_trace_getter()->UponLeavingGTest();
+    internal::HandleExceptionsInMethodIfSupported(
+        this, &Test::TestBody, "the test body");
+  }
+
+  // However, we want to clean up as much as possible.  Hence we will
+  // always call TearDown(), even if SetUp() or the test body has
+  // failed.
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      this, &Test::TearDown, "TearDown()");
+}
+
+// Returns true iff the current test has a fatal failure.
+bool Test::HasFatalFailure() {
+  return internal::GetUnitTestImpl()->current_test_result()->HasFatalFailure();
+}
+
+// Returns true iff the current test has a non-fatal failure.
+bool Test::HasNonfatalFailure() {
+  return internal::GetUnitTestImpl()->current_test_result()->
+      HasNonfatalFailure();
+}
+
+// class TestInfo
+
+// Constructs a TestInfo object. It assumes ownership of the test factory
+// object.
+TestInfo::TestInfo(const std::string& a_test_case_name,
+                   const std::string& a_name,
+                   const char* a_type_param,
+                   const char* a_value_param,
+                   internal::TypeId fixture_class_id,
+                   internal::TestFactoryBase* factory)
+    : test_case_name_(a_test_case_name),
+      name_(a_name),
+      type_param_(a_type_param ? new std::string(a_type_param) : NULL),
+      value_param_(a_value_param ? new std::string(a_value_param) : NULL),
+      fixture_class_id_(fixture_class_id),
+      should_run_(false),
+      is_disabled_(false),
+      matches_filter_(false),
+      factory_(factory),
+      result_() {}
+
+// Destructs a TestInfo object.
+TestInfo::~TestInfo() { delete factory_; }
+
+namespace internal {
+
+// Creates a new TestInfo object and registers it with Google Test;
+// returns the created object.
+//
+// Arguments:
+//
+//   test_case_name:   name of the test case
+//   name:             name of the test
+//   type_param:       the name of the test's type parameter, or NULL if
+//                     this is not a typed or a type-parameterized test.
+//   value_param:      text representation of the test's value parameter,
+//                     or NULL if this is not a value-parameterized test.
+//   fixture_class_id: ID of the test fixture class
+//   set_up_tc:        pointer to the function that sets up the test case
+//   tear_down_tc:     pointer to the function that tears down the test case
+//   factory:          pointer to the factory that creates a test object.
+//                     The newly created TestInfo instance will assume
+//                     ownership of the factory object.
+TestInfo* MakeAndRegisterTestInfo(
+    const char* test_case_name,
+    const char* name,
+    const char* type_param,
+    const char* value_param,
+    TypeId fixture_class_id,
+    SetUpTestCaseFunc set_up_tc,
+    TearDownTestCaseFunc tear_down_tc,
+    TestFactoryBase* factory) {
+  TestInfo* const test_info =
+      new TestInfo(test_case_name, name, type_param, value_param,
+                   fixture_class_id, factory);
+  GetUnitTestImpl()->AddTestInfo(set_up_tc, tear_down_tc, test_info);
+  return test_info;
+}
+
+#if GTEST_HAS_PARAM_TEST
+void ReportInvalidTestCaseType(const char* test_case_name,
+                               const char* file, int line) {
+  Message errors;
+  errors
+      << "Attempted redefinition of test case " << test_case_name << ".\n"
+      << "All tests in the same test case must use the same test fixture\n"
+      << "class.  However, in test case " << test_case_name << ", you tried\n"
+      << "to define a test using a fixture class different from the one\n"
+      << "used earlier. This can happen if the two fixture classes are\n"
+      << "from different namespaces and have the same name. You should\n"
+      << "probably rename one of the classes to put the tests into different\n"
+      << "test cases.";
+
+  fprintf(stderr, "%s %s", FormatFileLocation(file, line).c_str(),
+          errors.GetString().c_str());
+}
+#endif  // GTEST_HAS_PARAM_TEST
+
+}  // namespace internal
+
+namespace {
+
+// A predicate that checks the test name of a TestInfo against a known
+// value.
+//
+// This is used for implementation of the TestCase class only.  We put
+// it in the anonymous namespace to prevent polluting the outer
+// namespace.
+//
+// TestNameIs is copyable.
+
+//Commenting out this class since its not used and wherefor produces warnings
+// class TestNameIs {
+// public:
+//  // Constructor.
+//  //
+//  // TestNameIs has NO default constructor.
+//  explicit TestNameIs(const char* name)
+//      : name_(name) {}
+//
+//  // Returns true iff the test name of test_info matches name_.
+//  bool operator()(const TestInfo * test_info) const {
+//    return test_info && test_info->name() == name_;
+//  }
+//
+// private:
+//  std::string name_;
+//};
+
+}  // namespace
+
+namespace internal {
+
+// This method expands all parameterized tests registered with macros TEST_P
+// and INSTANTIATE_TEST_CASE_P into regular tests and registers those.
+// This will be done just once during the program runtime.
+void UnitTestImpl::RegisterParameterizedTests() {
+#if GTEST_HAS_PARAM_TEST
+  if (!parameterized_tests_registered_) {
+    parameterized_test_registry_.RegisterTests();
+    parameterized_tests_registered_ = true;
+  }
+#endif
+}
+
+}  // namespace internal
+
+// Creates the test object, runs it, records its result, and then
+// deletes it.
+void TestInfo::Run() {
+  if (!should_run_) return;
+
+  // Tells UnitTest where to store test result.
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  impl->set_current_test_info(this);
+
+  TestEventListener* repeater = UnitTest::GetInstance()->listeners().repeater();
+
+  // Notifies the unit test event listeners that a test is about to start.
+  repeater->OnTestStart(*this);
+
+  const TimeInMillis start = internal::GetTimeInMillis();
+
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+
+  // Creates the test object.
+  Test* const test = internal::HandleExceptionsInMethodIfSupported(
+      factory_, &internal::TestFactoryBase::CreateTest,
+      "the test fixture's constructor");
+
+  // Runs the test only if the test object was created and its
+  // constructor didn't generate a fatal failure.
+  if ((test != NULL) && !Test::HasFatalFailure()) {
+    // This doesn't throw as all user code that can throw are wrapped into
+    // exception handling code.
+    test->Run();
+  }
+
+  // Deletes the test object.
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      test, &Test::DeleteSelf_, "the test fixture's destructor");
+
+  result_.set_elapsed_time(internal::GetTimeInMillis() - start);
+
+  // Notifies the unit test event listener that a test has just finished.
+  repeater->OnTestEnd(*this);
+
+  // Tells UnitTest to stop associating assertion results to this
+  // test.
+  impl->set_current_test_info(NULL);
+}
+
+// class TestCase
+
+// Gets the number of successful tests in this test case.
+int TestCase::successful_test_count() const {
+  return CountIf(test_info_list_, TestPassed);
+}
+
+// Gets the number of failed tests in this test case.
+int TestCase::failed_test_count() const {
+  return CountIf(test_info_list_, TestFailed);
+}
+
+// Gets the number of disabled tests that will be reported in the XML report.
+int TestCase::reportable_disabled_test_count() const {
+  return CountIf(test_info_list_, TestReportableDisabled);
+}
+
+// Gets the number of disabled tests in this test case.
+int TestCase::disabled_test_count() const {
+  return CountIf(test_info_list_, TestDisabled);
+}
+
+// Gets the number of tests to be printed in the XML report.
+int TestCase::reportable_test_count() const {
+  return CountIf(test_info_list_, TestReportable);
+}
+
+// Get the number of tests in this test case that should run.
+int TestCase::test_to_run_count() const {
+  return CountIf(test_info_list_, ShouldRunTest);
+}
+
+// Gets the number of all tests.
+int TestCase::total_test_count() const {
+  return static_cast(test_info_list_.size());
+}
+
+// Creates a TestCase with the given name.
+//
+// Arguments:
+//
+//   name:         name of the test case
+//   a_type_param: the name of the test case's type parameter, or NULL if
+//                 this is not a typed or a type-parameterized test case.
+//   set_up_tc:    pointer to the function that sets up the test case
+//   tear_down_tc: pointer to the function that tears down the test case
+TestCase::TestCase(const char* a_name, const char* a_type_param,
+                   Test::SetUpTestCaseFunc set_up_tc,
+                   Test::TearDownTestCaseFunc tear_down_tc)
+    : name_(a_name),
+      type_param_(a_type_param ? new std::string(a_type_param) : NULL),
+      set_up_tc_(set_up_tc),
+      tear_down_tc_(tear_down_tc),
+      should_run_(false),
+      elapsed_time_(0) {
+}
+
+// Destructor of TestCase.
+TestCase::~TestCase() {
+  // Deletes every Test in the collection.
+  ForEach(test_info_list_, internal::Delete);
+}
+
+// Returns the i-th test among all the tests. i can range from 0 to
+// total_test_count() - 1. If i is not in that range, returns NULL.
+const TestInfo* TestCase::GetTestInfo(int i) const {
+  const int index = GetElementOr(test_indices_, i, -1);
+  return index < 0 ? NULL : test_info_list_[index];
+}
+
+// Returns the i-th test among all the tests. i can range from 0 to
+// total_test_count() - 1. If i is not in that range, returns NULL.
+TestInfo* TestCase::GetMutableTestInfo(int i) {
+  const int index = GetElementOr(test_indices_, i, -1);
+  return index < 0 ? NULL : test_info_list_[index];
+}
+
+// Adds a test to this test case.  Will delete the test upon
+// destruction of the TestCase object.
+void TestCase::AddTestInfo(TestInfo * test_info) {
+  test_info_list_.push_back(test_info);
+  test_indices_.push_back(static_cast(test_indices_.size()));
+}
+
+// Runs every test in this TestCase.
+void TestCase::Run() {
+  if (!should_run_) return;
+
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  impl->set_current_test_case(this);
+
+  TestEventListener* repeater = UnitTest::GetInstance()->listeners().repeater();
+
+  repeater->OnTestCaseStart(*this);
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      this, &TestCase::RunSetUpTestCase, "SetUpTestCase()");
+
+  const internal::TimeInMillis start = internal::GetTimeInMillis();
+  for (int i = 0; i < total_test_count(); i++) {
+    GetMutableTestInfo(i)->Run();
+  }
+  elapsed_time_ = internal::GetTimeInMillis() - start;
+
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      this, &TestCase::RunTearDownTestCase, "TearDownTestCase()");
+
+  repeater->OnTestCaseEnd(*this);
+  impl->set_current_test_case(NULL);
+}
+
+// Clears the results of all tests in this test case.
+void TestCase::ClearResult() {
+  ad_hoc_test_result_.Clear();
+  ForEach(test_info_list_, TestInfo::ClearTestResult);
+}
+
+// Shuffles the tests in this test case.
+void TestCase::ShuffleTests(internal::Random* random) {
+  Shuffle(random, &test_indices_);
+}
+
+// Restores the test order to before the first shuffle.
+void TestCase::UnshuffleTests() {
+  for (size_t i = 0; i < test_indices_.size(); i++) {
+    test_indices_[i] = static_cast(i);
+  }
+}
+
+// Formats a countable noun.  Depending on its quantity, either the
+// singular form or the plural form is used. e.g.
+//
+// FormatCountableNoun(1, "formula", "formuli") returns "1 formula".
+// FormatCountableNoun(5, "book", "books") returns "5 books".
+static std::string FormatCountableNoun(int count,
+                                       const char * singular_form,
+                                       const char * plural_form) {
+  return internal::StreamableToString(count) + " " +
+      (count == 1 ? singular_form : plural_form);
+}
+
+// Formats the count of tests.
+static std::string FormatTestCount(int test_count) {
+  return FormatCountableNoun(test_count, "test", "tests");
+}
+
+// Formats the count of test cases.
+static std::string FormatTestCaseCount(int test_case_count) {
+  return FormatCountableNoun(test_case_count, "test case", "test cases");
+}
+
+// Converts a TestPartResult::Type enum to human-friendly string
+// representation.  Both kNonFatalFailure and kFatalFailure are translated
+// to "Failure", as the user usually doesn't care about the difference
+// between the two when viewing the test result.
+static const char * TestPartResultTypeToString(TestPartResult::Type type) {
+  switch (type) {
+    case TestPartResult::kSuccess:
+      return "Success";
+
+    case TestPartResult::kNonFatalFailure:
+    case TestPartResult::kFatalFailure:
+#ifdef _MSC_VER
+      return "error: ";
+#else
+      return "Failure\n";
+#endif
+    default:
+      return "Unknown result type";
+  }
+}
+
+namespace internal {
+
+// Prints a TestPartResult to an std::string.
+static std::string PrintTestPartResultToString(
+    const TestPartResult& test_part_result) {
+  return (Message()
+          << internal::FormatFileLocation(test_part_result.file_name(),
+                                          test_part_result.line_number())
+          << " " << TestPartResultTypeToString(test_part_result.type())
+          << test_part_result.message()).GetString();
+}
+
+// Prints a TestPartResult.
+static void PrintTestPartResult(const TestPartResult& test_part_result) {
+  const std::string& result =
+      PrintTestPartResultToString(test_part_result);
+  printf("%s\n", result.c_str());
+  fflush(stdout);
+  // If the test program runs in Visual Studio or a debugger, the
+  // following statements add the test part result message to the Output
+  // window such that the user can double-click on it to jump to the
+  // corresponding source code location; otherwise they do nothing.
+#if GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+  // We don't call OutputDebugString*() on Windows Mobile, as printing
+  // to stdout is done by OutputDebugString() there already - we don't
+  // want the same message printed twice.
+  ::OutputDebugStringA(result.c_str());
+  ::OutputDebugStringA("\n");
+#endif
+}
+
+// class PrettyUnitTestResultPrinter
+
+enum GTestColor {
+  COLOR_DEFAULT,
+  COLOR_RED,
+  COLOR_GREEN,
+  COLOR_YELLOW
+};
+
+#if GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+
+// Returns the character attribute for the given color.
+WORD GetColorAttribute(GTestColor color) {
+  switch (color) {
+    case COLOR_RED:    return FOREGROUND_RED;
+    case COLOR_GREEN:  return FOREGROUND_GREEN;
+    case COLOR_YELLOW: return FOREGROUND_RED | FOREGROUND_GREEN;
+    default:           return 0;
+  }
+}
+
+#else
+
+// Returns the ANSI color code for the given color.  COLOR_DEFAULT is
+// an invalid input.
+const char* GetAnsiColorCode(GTestColor color) {
+  switch (color) {
+    case COLOR_RED:     return "1";
+    case COLOR_GREEN:   return "2";
+    case COLOR_YELLOW:  return "3";
+    default:            return NULL;
+  };
+}
+
+#endif  // GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+
+// Returns true iff Google Test should use colors in the output.
+bool ShouldUseColor(bool stdout_is_tty) {
+  const char* const gtest_color = GTEST_FLAG(color).c_str();
+
+  if (String::CaseInsensitiveCStringEquals(gtest_color, "auto")) {
+#if GTEST_OS_WINDOWS
+    // On Windows the TERM variable is usually not set, but the
+    // console there does support colors.
+    return stdout_is_tty;
+#else
+    // On non-Windows platforms, we rely on the TERM variable.
+    const char* const term = posix::GetEnv("TERM");
+    const bool term_supports_color =
+        String::CStringEquals(term, "xterm") ||
+        String::CStringEquals(term, "xterm-color") ||
+        String::CStringEquals(term, "xterm-256color") ||
+        String::CStringEquals(term, "screen") ||
+        String::CStringEquals(term, "screen-256color") ||
+        String::CStringEquals(term, "linux") ||
+        String::CStringEquals(term, "cygwin");
+    return stdout_is_tty && term_supports_color;
+#endif  // GTEST_OS_WINDOWS
+  }
+
+  return String::CaseInsensitiveCStringEquals(gtest_color, "yes") ||
+      String::CaseInsensitiveCStringEquals(gtest_color, "true") ||
+      String::CaseInsensitiveCStringEquals(gtest_color, "t") ||
+      String::CStringEquals(gtest_color, "1");
+  // We take "yes", "true", "t", and "1" as meaning "yes".  If the
+  // value is neither one of these nor "auto", we treat it as "no" to
+  // be conservative.
+}
+
+// Helpers for printing colored strings to stdout. Note that on Windows, we
+// cannot simply emit special characters and have the terminal change colors.
+// This routine must actually emit the characters rather than return a string
+// that would be colored when printed, as can be done on Linux.
+void ColoredPrintf(GTestColor color, const char* fmt, ...) {
+  va_list args;
+  va_start(args, fmt);
+
+#if GTEST_OS_WINDOWS_MOBILE || GTEST_OS_SYMBIAN || GTEST_OS_ZOS || GTEST_OS_IOS
+  const bool use_color = false;
+#else
+  static const bool in_color_mode =
+      ShouldUseColor(posix::IsATTY(posix::FileNo(stdout)) != 0);
+  const bool use_color = in_color_mode && (color != COLOR_DEFAULT);
+#endif  // GTEST_OS_WINDOWS_MOBILE || GTEST_OS_SYMBIAN || GTEST_OS_ZOS
+  // The '!= 0' comparison is necessary to satisfy MSVC 7.1.
+
+  if (!use_color) {
+    vprintf(fmt, args);
+    va_end(args);
+    return;
+  }
+
+#if GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+  const HANDLE stdout_handle = GetStdHandle(STD_OUTPUT_HANDLE);
+
+  // Gets the current text color.
+  CONSOLE_SCREEN_BUFFER_INFO buffer_info;
+  GetConsoleScreenBufferInfo(stdout_handle, &buffer_info);
+  const WORD old_color_attrs = buffer_info.wAttributes;
+
+  // We need to flush the stream buffers into the console before each
+  // SetConsoleTextAttribute call lest it affect the text that is already
+  // printed but has not yet reached the console.
+  fflush(stdout);
+  SetConsoleTextAttribute(stdout_handle,
+                          GetColorAttribute(color) | FOREGROUND_INTENSITY);
+  vprintf(fmt, args);
+
+  fflush(stdout);
+  // Restores the text color.
+  SetConsoleTextAttribute(stdout_handle, old_color_attrs);
+#else
+  printf("\033[0;3%sm", GetAnsiColorCode(color));
+  vprintf(fmt, args);
+  printf("\033[m");  // Resets the terminal to default.
+#endif  // GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+  va_end(args);
+}
+
+// Text printed in Google Test's text output and --gunit_list_tests
+// output to label the type parameter and value parameter for a test.
+static const char kTypeParamLabel[] = "TypeParam";
+static const char kValueParamLabel[] = "GetParam()";
+
+void PrintFullTestCommentIfPresent(const TestInfo& test_info) {
+  const char* const type_param = test_info.type_param();
+  const char* const value_param = test_info.value_param();
+
+  if (type_param != NULL || value_param != NULL) {
+    printf(", where ");
+    if (type_param != NULL) {
+      printf("%s = %s", kTypeParamLabel, type_param);
+      if (value_param != NULL)
+        printf(" and ");
+    }
+    if (value_param != NULL) {
+      printf("%s = %s", kValueParamLabel, value_param);
+    }
+  }
+}
+
+// This class implements the TestEventListener interface.
+//
+// Class PrettyUnitTestResultPrinter is copyable.
+class PrettyUnitTestResultPrinter : public TestEventListener {
+ public:
+  PrettyUnitTestResultPrinter() {}
+  static void PrintTestName(const char * test_case, const char * test) {
+    printf("%s.%s", test_case, test);
+  }
+
+  // The following methods override what's in the TestEventListener class.
+  virtual void OnTestProgramStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationStart(const UnitTest& unit_test, int iteration);
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestCaseStart(const TestCase& test_case);
+  virtual void OnTestStart(const TestInfo& test_info);
+  virtual void OnTestPartResult(const TestPartResult& result);
+  virtual void OnTestEnd(const TestInfo& test_info);
+  virtual void OnTestCaseEnd(const TestCase& test_case);
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationEnd(const UnitTest& unit_test, int iteration);
+  virtual void OnTestProgramEnd(const UnitTest& /*unit_test*/) {}
+
+ private:
+  static void PrintFailedTests(const UnitTest& unit_test);
+};
+
+  // Fired before each iteration of tests starts.
+void PrettyUnitTestResultPrinter::OnTestIterationStart(
+    const UnitTest& unit_test, int iteration) {
+  if (GTEST_FLAG(repeat) != 1)
+    printf("\nRepeating all tests (iteration %d) . . .\n\n", iteration + 1);
+
+  const char* const filter = GTEST_FLAG(filter).c_str();
+
+  // Prints the filter if it's not *.  This reminds the user that some
+  // tests may be skipped.
+  if (!String::CStringEquals(filter, kUniversalFilter)) {
+    ColoredPrintf(COLOR_YELLOW,
+                  "Note: %s filter = %s\n", GTEST_NAME_, filter);
+  }
+
+  if (internal::ShouldShard(kTestTotalShards, kTestShardIndex, false)) {
+    const Int32 shard_index = Int32FromEnvOrDie(kTestShardIndex, -1);
+    ColoredPrintf(COLOR_YELLOW,
+                  "Note: This is test shard %d of %s.\n",
+                  static_cast(shard_index) + 1,
+                  internal::posix::GetEnv(kTestTotalShards));
+  }
+
+  if (GTEST_FLAG(shuffle)) {
+    ColoredPrintf(COLOR_YELLOW,
+                  "Note: Randomizing tests' orders with a seed of %d .\n",
+                  unit_test.random_seed());
+  }
+
+  ColoredPrintf(COLOR_GREEN,  "[==========] ");
+  printf("Running %s from %s.\n",
+         FormatTestCount(unit_test.test_to_run_count()).c_str(),
+         FormatTestCaseCount(unit_test.test_case_to_run_count()).c_str());
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnEnvironmentsSetUpStart(
+    const UnitTest& /*unit_test*/) {
+  ColoredPrintf(COLOR_GREEN,  "[----------] ");
+  printf("Global test environment set-up.\n");
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestCaseStart(const TestCase& test_case) {
+  const std::string counts =
+      FormatCountableNoun(test_case.test_to_run_count(), "test", "tests");
+  ColoredPrintf(COLOR_GREEN, "[----------] ");
+  printf("%s from %s", counts.c_str(), test_case.name());
+  if (test_case.type_param() == NULL) {
+    printf("\n");
+  } else {
+    printf(", where %s = %s\n", kTypeParamLabel, test_case.type_param());
+  }
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestStart(const TestInfo& test_info) {
+  ColoredPrintf(COLOR_GREEN,  "[ RUN      ] ");
+  PrintTestName(test_info.test_case_name(), test_info.name());
+  printf("\n");
+  fflush(stdout);
+}
+
+// Called after an assertion failure.
+void PrettyUnitTestResultPrinter::OnTestPartResult(
+    const TestPartResult& result) {
+  // If the test part succeeded, we don't need to do anything.
+  if (result.type() == TestPartResult::kSuccess)
+    return;
+
+  // Print failure message from the assertion (e.g. expected this and got that).
+  PrintTestPartResult(result);
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestEnd(const TestInfo& test_info) {
+  if (test_info.result()->Passed()) {
+    ColoredPrintf(COLOR_GREEN, "[       OK ] ");
+  } else {
+    ColoredPrintf(COLOR_RED, "[  FAILED  ] ");
+  }
+  PrintTestName(test_info.test_case_name(), test_info.name());
+  if (test_info.result()->Failed())
+    PrintFullTestCommentIfPresent(test_info);
+
+  if (GTEST_FLAG(print_time)) {
+    printf(" (%s ms)\n", internal::StreamableToString(
+           test_info.result()->elapsed_time()).c_str());
+  } else {
+    printf("\n");
+  }
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestCaseEnd(const TestCase& test_case) {
+  if (!GTEST_FLAG(print_time)) return;
+
+  const std::string counts =
+      FormatCountableNoun(test_case.test_to_run_count(), "test", "tests");
+  ColoredPrintf(COLOR_GREEN, "[----------] ");
+  printf("%s from %s (%s ms total)\n\n",
+         counts.c_str(), test_case.name(),
+         internal::StreamableToString(test_case.elapsed_time()).c_str());
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnEnvironmentsTearDownStart(
+    const UnitTest& /*unit_test*/) {
+  ColoredPrintf(COLOR_GREEN,  "[----------] ");
+  printf("Global test environment tear-down\n");
+  fflush(stdout);
+}
+
+// Internal helper for printing the list of failed tests.
+void PrettyUnitTestResultPrinter::PrintFailedTests(const UnitTest& unit_test) {
+  const int failed_test_count = unit_test.failed_test_count();
+  if (failed_test_count == 0) {
+    return;
+  }
+
+  for (int i = 0; i < unit_test.total_test_case_count(); ++i) {
+    const TestCase& test_case = *unit_test.GetTestCase(i);
+    if (!test_case.should_run() || (test_case.failed_test_count() == 0)) {
+      continue;
+    }
+    for (int j = 0; j < test_case.total_test_count(); ++j) {
+      const TestInfo& test_info = *test_case.GetTestInfo(j);
+      if (!test_info.should_run() || test_info.result()->Passed()) {
+        continue;
+      }
+      ColoredPrintf(COLOR_RED, "[  FAILED  ] ");
+      printf("%s.%s", test_case.name(), test_info.name());
+      PrintFullTestCommentIfPresent(test_info);
+      printf("\n");
+    }
+  }
+}
+
+void PrettyUnitTestResultPrinter::OnTestIterationEnd(const UnitTest& unit_test,
+                                                     int /*iteration*/) {
+  ColoredPrintf(COLOR_GREEN,  "[==========] ");
+  printf("%s from %s ran.",
+         FormatTestCount(unit_test.test_to_run_count()).c_str(),
+         FormatTestCaseCount(unit_test.test_case_to_run_count()).c_str());
+  if (GTEST_FLAG(print_time)) {
+    printf(" (%s ms total)",
+           internal::StreamableToString(unit_test.elapsed_time()).c_str());
+  }
+  printf("\n");
+  ColoredPrintf(COLOR_GREEN,  "[  PASSED  ] ");
+  printf("%s.\n", FormatTestCount(unit_test.successful_test_count()).c_str());
+
+  int num_failures = unit_test.failed_test_count();
+  if (!unit_test.Passed()) {
+    const int failed_test_count = unit_test.failed_test_count();
+    ColoredPrintf(COLOR_RED,  "[  FAILED  ] ");
+    printf("%s, listed below:\n", FormatTestCount(failed_test_count).c_str());
+    PrintFailedTests(unit_test);
+    printf("\n%2d FAILED %s\n", num_failures,
+                        num_failures == 1 ? "TEST" : "TESTS");
+  }
+
+  int num_disabled = unit_test.reportable_disabled_test_count();
+  if (num_disabled && !GTEST_FLAG(also_run_disabled_tests)) {
+    if (!num_failures) {
+      printf("\n");  // Add a spacer if no FAILURE banner is displayed.
+    }
+    ColoredPrintf(COLOR_YELLOW,
+                  "  YOU HAVE %d DISABLED %s\n\n",
+                  num_disabled,
+                  num_disabled == 1 ? "TEST" : "TESTS");
+  }
+  // Ensure that Google Test output is printed before, e.g., heapchecker output.
+  fflush(stdout);
+}
+
+// End PrettyUnitTestResultPrinter
+
+// class TestEventRepeater
+//
+// This class forwards events to other event listeners.
+class TestEventRepeater : public TestEventListener {
+ public:
+  TestEventRepeater() : forwarding_enabled_(true) {}
+  virtual ~TestEventRepeater();
+  void Append(TestEventListener *listener);
+  TestEventListener* Release(TestEventListener* listener);
+
+  // Controls whether events will be forwarded to listeners_. Set to false
+  // in death test child processes.
+  bool forwarding_enabled() const { return forwarding_enabled_; }
+  void set_forwarding_enabled(bool enable) { forwarding_enabled_ = enable; }
+
+  virtual void OnTestProgramStart(const UnitTest& unit_test);
+  virtual void OnTestIterationStart(const UnitTest& unit_test, int iteration);
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& unit_test);
+  virtual void OnTestCaseStart(const TestCase& test_case);
+  virtual void OnTestStart(const TestInfo& test_info);
+  virtual void OnTestPartResult(const TestPartResult& result);
+  virtual void OnTestEnd(const TestInfo& test_info);
+  virtual void OnTestCaseEnd(const TestCase& test_case);
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& unit_test);
+  virtual void OnTestIterationEnd(const UnitTest& unit_test, int iteration);
+  virtual void OnTestProgramEnd(const UnitTest& unit_test);
+
+ private:
+  // Controls whether events will be forwarded to listeners_. Set to false
+  // in death test child processes.
+  bool forwarding_enabled_;
+  // The list of listeners that receive events.
+  std::vector listeners_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestEventRepeater);
+};
+
+TestEventRepeater::~TestEventRepeater() {
+  ForEach(listeners_, Delete);
+}
+
+void TestEventRepeater::Append(TestEventListener *listener) {
+  listeners_.push_back(listener);
+}
+
+// TODO(vladl@google.com): Factor the search functionality into Vector::Find.
+TestEventListener* TestEventRepeater::Release(TestEventListener *listener) {
+  for (size_t i = 0; i < listeners_.size(); ++i) {
+    if (listeners_[i] == listener) {
+      listeners_.erase(listeners_.begin() + i);
+      return listener;
+    }
+  }
+
+  return NULL;
+}
+
+// Since most methods are very similar, use macros to reduce boilerplate.
+// This defines a member that forwards the call to all listeners.
+#define GTEST_REPEATER_METHOD_(Name, Type) \
+void TestEventRepeater::Name(const Type& parameter) { \
+  if (forwarding_enabled_) { \
+    for (size_t i = 0; i < listeners_.size(); i++) { \
+      listeners_[i]->Name(parameter); \
+    } \
+  } \
+}
+// This defines a member that forwards the call to all listeners in reverse
+// order.
+#define GTEST_REVERSE_REPEATER_METHOD_(Name, Type) \
+void TestEventRepeater::Name(const Type& parameter) { \
+  if (forwarding_enabled_) { \
+    for (int i = static_cast(listeners_.size()) - 1; i >= 0; i--) { \
+      listeners_[i]->Name(parameter); \
+    } \
+  } \
+}
+
+GTEST_REPEATER_METHOD_(OnTestProgramStart, UnitTest)
+GTEST_REPEATER_METHOD_(OnEnvironmentsSetUpStart, UnitTest)
+GTEST_REPEATER_METHOD_(OnTestCaseStart, TestCase)
+GTEST_REPEATER_METHOD_(OnTestStart, TestInfo)
+GTEST_REPEATER_METHOD_(OnTestPartResult, TestPartResult)
+GTEST_REPEATER_METHOD_(OnEnvironmentsTearDownStart, UnitTest)
+GTEST_REVERSE_REPEATER_METHOD_(OnEnvironmentsSetUpEnd, UnitTest)
+GTEST_REVERSE_REPEATER_METHOD_(OnEnvironmentsTearDownEnd, UnitTest)
+GTEST_REVERSE_REPEATER_METHOD_(OnTestEnd, TestInfo)
+GTEST_REVERSE_REPEATER_METHOD_(OnTestCaseEnd, TestCase)
+GTEST_REVERSE_REPEATER_METHOD_(OnTestProgramEnd, UnitTest)
+
+#undef GTEST_REPEATER_METHOD_
+#undef GTEST_REVERSE_REPEATER_METHOD_
+
+void TestEventRepeater::OnTestIterationStart(const UnitTest& unit_test,
+                                             int iteration) {
+  if (forwarding_enabled_) {
+    for (size_t i = 0; i < listeners_.size(); i++) {
+      listeners_[i]->OnTestIterationStart(unit_test, iteration);
+    }
+  }
+}
+
+void TestEventRepeater::OnTestIterationEnd(const UnitTest& unit_test,
+                                           int iteration) {
+  if (forwarding_enabled_) {
+    for (int i = static_cast(listeners_.size()) - 1; i >= 0; i--) {
+      listeners_[i]->OnTestIterationEnd(unit_test, iteration);
+    }
+  }
+}
+
+// End TestEventRepeater
+
+// This class generates an XML output file.
+class XmlUnitTestResultPrinter : public EmptyTestEventListener {
+ public:
+  explicit XmlUnitTestResultPrinter(const char* output_file);
+
+  virtual void OnTestIterationEnd(const UnitTest& unit_test, int iteration);
+
+ private:
+  // Is c a whitespace character that is normalized to a space character
+  // when it appears in an XML attribute value?
+  static bool IsNormalizableWhitespace(char c) {
+    return c == 0x9 || c == 0xA || c == 0xD;
+  }
+
+  // May c appear in a well-formed XML document?
+  static bool IsValidXmlCharacter(char c) {
+    return IsNormalizableWhitespace(c) || c >= 0x20;
+  }
+
+  // Returns an XML-escaped copy of the input string str.  If
+  // is_attribute is true, the text is meant to appear as an attribute
+  // value, and normalizable whitespace is preserved by replacing it
+  // with character references.
+  static std::string EscapeXml(const std::string& str, bool is_attribute);
+
+  // Returns the given string with all characters invalid in XML removed.
+  static std::string RemoveInvalidXmlCharacters(const std::string& str);
+
+  // Convenience wrapper around EscapeXml when str is an attribute value.
+  static std::string EscapeXmlAttribute(const std::string& str) {
+    return EscapeXml(str, true);
+  }
+
+  // Convenience wrapper around EscapeXml when str is not an attribute value.
+  static std::string EscapeXmlText(const char* str) {
+    return EscapeXml(str, false);
+  }
+
+  // Verifies that the given attribute belongs to the given element and
+  // streams the attribute as XML.
+  static void OutputXmlAttribute(std::ostream* stream,
+                                 const std::string& element_name,
+                                 const std::string& name,
+                                 const std::string& value);
+
+  // Streams an XML CDATA section, escaping invalid CDATA sequences as needed.
+  static void OutputXmlCDataSection(::std::ostream* stream, const char* data);
+
+  // Streams an XML representation of a TestInfo object.
+  static void OutputXmlTestInfo(::std::ostream* stream,
+                                const char* test_case_name,
+                                const TestInfo& test_info);
+
+  // Prints an XML representation of a TestCase object
+  static void PrintXmlTestCase(::std::ostream* stream,
+                               const TestCase& test_case);
+
+  // Prints an XML summary of unit_test to output stream out.
+  static void PrintXmlUnitTest(::std::ostream* stream,
+                               const UnitTest& unit_test);
+
+  // Produces a string representing the test properties in a result as space
+  // delimited XML attributes based on the property key="value" pairs.
+  // When the std::string is not empty, it includes a space at the beginning,
+  // to delimit this attribute from prior attributes.
+  static std::string TestPropertiesAsXmlAttributes(const TestResult& result);
+
+  // The output file.
+  const std::string output_file_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(XmlUnitTestResultPrinter);
+};
+
+// Creates a new XmlUnitTestResultPrinter.
+XmlUnitTestResultPrinter::XmlUnitTestResultPrinter(const char* output_file)
+    : output_file_(output_file) {
+  if (output_file_.c_str() == NULL || output_file_.empty()) {
+    fprintf(stderr, "XML output file may not be null\n");
+    fflush(stderr);
+    exit(EXIT_FAILURE);
+  }
+}
+
+// Called after the unit test ends.
+void XmlUnitTestResultPrinter::OnTestIterationEnd(const UnitTest& unit_test,
+                                                  int /*iteration*/) {
+  FILE* xmlout = NULL;
+  FilePath output_file(output_file_);
+  FilePath output_dir(output_file.RemoveFileName());
+
+  if (output_dir.CreateDirectoriesRecursively()) {
+    xmlout = posix::FOpen(output_file_.c_str(), "w");
+  }
+  if (xmlout == NULL) {
+    // TODO(wan): report the reason of the failure.
+    //
+    // We don't do it for now as:
+    //
+    //   1. There is no urgent need for it.
+    //   2. It's a bit involved to make the errno variable thread-safe on
+    //      all three operating systems (Linux, Windows, and Mac OS).
+    //   3. To interpret the meaning of errno in a thread-safe way,
+    //      we need the strerror_r() function, which is not available on
+    //      Windows.
+    fprintf(stderr,
+            "Unable to open file \"%s\"\n",
+            output_file_.c_str());
+    fflush(stderr);
+    exit(EXIT_FAILURE);
+  }
+  std::stringstream stream;
+  PrintXmlUnitTest(&stream, unit_test);
+  fprintf(xmlout, "%s", StringStreamToString(&stream).c_str());
+  fclose(xmlout);
+}
+
+// Returns an XML-escaped copy of the input string str.  If is_attribute
+// is true, the text is meant to appear as an attribute value, and
+// normalizable whitespace is preserved by replacing it with character
+// references.
+//
+// Invalid XML characters in str, if any, are stripped from the output.
+// It is expected that most, if not all, of the text processed by this
+// module will consist of ordinary English text.
+// If this module is ever modified to produce version 1.1 XML output,
+// most invalid characters can be retained using character references.
+// TODO(wan): It might be nice to have a minimally invasive, human-readable
+// escaping scheme for invalid characters, rather than dropping them.
+std::string XmlUnitTestResultPrinter::EscapeXml(
+    const std::string& str, bool is_attribute) {
+  Message m;
+
+  for (size_t i = 0; i < str.size(); ++i) {
+    const char ch = str[i];
+    switch (ch) {
+      case '<':
+        m << "<";
+        break;
+      case '>':
+        m << ">";
+        break;
+      case '&':
+        m << "&";
+        break;
+      case '\'':
+        if (is_attribute)
+          m << "'";
+        else
+          m << '\'';
+        break;
+      case '"':
+        if (is_attribute)
+          m << """;
+        else
+          m << '"';
+        break;
+      default:
+        if (IsValidXmlCharacter(ch)) {
+          if (is_attribute && IsNormalizableWhitespace(ch))
+            m << "&#x" << String::FormatByte(static_cast(ch))
+              << ";";
+          else
+            m << ch;
+        }
+        break;
+    }
+  }
+
+  return m.GetString();
+}
+
+// Returns the given string with all characters invalid in XML removed.
+// Currently invalid characters are dropped from the string. An
+// alternative is to replace them with certain characters such as . or ?.
+std::string XmlUnitTestResultPrinter::RemoveInvalidXmlCharacters(
+    const std::string& str) {
+  std::string output;
+  output.reserve(str.size());
+  for (std::string::const_iterator it = str.begin(); it != str.end(); ++it)
+    if (IsValidXmlCharacter(*it))
+      output.push_back(*it);
+
+  return output;
+}
+
+// The following routines generate an XML representation of a UnitTest
+// object.
+//
+// This is how Google Test concepts map to the DTD:
+//
+//         <-- corresponds to a UnitTest object
+//     <-- corresponds to a TestCase object
+//          <-- corresponds to a TestInfo object
+//       ...
+//       ...
+//       ...
+//                                     <-- individual assertion failures
+//     
+//   
+// 
+
+// Formats the given time in milliseconds as seconds.
+std::string FormatTimeInMillisAsSeconds(TimeInMillis ms) {
+  ::std::stringstream ss;
+  ss << ms/1000.0;
+  return ss.str();
+}
+
+// Converts the given epoch time in milliseconds to a date string in the ISO
+// 8601 format, without the timezone information.
+std::string FormatEpochTimeInMillisAsIso8601(TimeInMillis ms) {
+  // Using non-reentrant version as localtime_r is not portable.
+  time_t seconds = static_cast(ms / 1000);
+#ifdef _MSC_VER
+# pragma warning(push)          // Saves the current warning state.
+# pragma warning(disable:4996)  // Temporarily disables warning 4996
+                                // (function or variable may be unsafe).
+  const struct tm* const time_struct = localtime(&seconds);  // NOLINT
+# pragma warning(pop)           // Restores the warning state again.
+#else
+  const struct tm* const time_struct = localtime(&seconds);  // NOLINT
+#endif
+  if (time_struct == NULL)
+    return "";  // Invalid ms value
+
+  // YYYY-MM-DDThh:mm:ss
+  return StreamableToString(time_struct->tm_year + 1900) + "-" +
+      String::FormatIntWidth2(time_struct->tm_mon + 1) + "-" +
+      String::FormatIntWidth2(time_struct->tm_mday) + "T" +
+      String::FormatIntWidth2(time_struct->tm_hour) + ":" +
+      String::FormatIntWidth2(time_struct->tm_min) + ":" +
+      String::FormatIntWidth2(time_struct->tm_sec);
+}
+
+// Streams an XML CDATA section, escaping invalid CDATA sequences as needed.
+void XmlUnitTestResultPrinter::OutputXmlCDataSection(::std::ostream* stream,
+                                                     const char* data) {
+  const char* segment = data;
+  *stream << "");
+    if (next_segment != NULL) {
+      stream->write(
+          segment, static_cast(next_segment - segment));
+      *stream << "]]>]]>");
+    } else {
+      *stream << segment;
+      break;
+    }
+  }
+  *stream << "]]>";
+}
+
+void XmlUnitTestResultPrinter::OutputXmlAttribute(
+    std::ostream* stream,
+    const std::string& element_name,
+    const std::string& name,
+    const std::string& value) {
+  const std::vector& allowed_names =
+      GetReservedAttributesForElement(element_name);
+
+  GTEST_CHECK_(std::find(allowed_names.begin(), allowed_names.end(), name) !=
+                   allowed_names.end())
+      << "Attribute " << name << " is not allowed for element <" << element_name
+      << ">.";
+
+  *stream << " " << name << "=\"" << EscapeXmlAttribute(value) << "\"";
+}
+
+// Prints an XML representation of a TestInfo object.
+// TODO(wan): There is also value in printing properties with the plain printer.
+void XmlUnitTestResultPrinter::OutputXmlTestInfo(::std::ostream* stream,
+                                                 const char* test_case_name,
+                                                 const TestInfo& test_info) {
+  const TestResult& result = *test_info.result();
+  const std::string kTestcase = "testcase";
+
+  *stream << "    \n";
+      }
+      const string location = internal::FormatCompilerIndependentFileLocation(
+          part.file_name(), part.line_number());
+      const string summary = location + "\n" + part.summary();
+      *stream << "      ";
+      const string detail = location + "\n" + part.message();
+      OutputXmlCDataSection(stream, RemoveInvalidXmlCharacters(detail).c_str());
+      *stream << "\n";
+    }
+  }
+
+  if (failures == 0)
+    *stream << " />\n";
+  else
+    *stream << "    \n";
+}
+
+// Prints an XML representation of a TestCase object
+void XmlUnitTestResultPrinter::PrintXmlTestCase(std::ostream* stream,
+                                                const TestCase& test_case) {
+  const std::string kTestsuite = "testsuite";
+  *stream << "  <" << kTestsuite;
+  OutputXmlAttribute(stream, kTestsuite, "name", test_case.name());
+  OutputXmlAttribute(stream, kTestsuite, "tests",
+                     StreamableToString(test_case.reportable_test_count()));
+  OutputXmlAttribute(stream, kTestsuite, "failures",
+                     StreamableToString(test_case.failed_test_count()));
+  OutputXmlAttribute(
+      stream, kTestsuite, "disabled",
+      StreamableToString(test_case.reportable_disabled_test_count()));
+  OutputXmlAttribute(stream, kTestsuite, "errors", "0");
+  OutputXmlAttribute(stream, kTestsuite, "time",
+                     FormatTimeInMillisAsSeconds(test_case.elapsed_time()));
+  *stream << TestPropertiesAsXmlAttributes(test_case.ad_hoc_test_result())
+          << ">\n";
+
+  for (int i = 0; i < test_case.total_test_count(); ++i) {
+    if (test_case.GetTestInfo(i)->is_reportable())
+      OutputXmlTestInfo(stream, test_case.name(), *test_case.GetTestInfo(i));
+  }
+  *stream << "  \n";
+}
+
+// Prints an XML summary of unit_test to output stream out.
+void XmlUnitTestResultPrinter::PrintXmlUnitTest(std::ostream* stream,
+                                                const UnitTest& unit_test) {
+  const std::string kTestsuites = "testsuites";
+
+  *stream << "\n";
+  *stream << "<" << kTestsuites;
+
+  OutputXmlAttribute(stream, kTestsuites, "tests",
+                     StreamableToString(unit_test.reportable_test_count()));
+  OutputXmlAttribute(stream, kTestsuites, "failures",
+                     StreamableToString(unit_test.failed_test_count()));
+  OutputXmlAttribute(
+      stream, kTestsuites, "disabled",
+      StreamableToString(unit_test.reportable_disabled_test_count()));
+  OutputXmlAttribute(stream, kTestsuites, "errors", "0");
+  OutputXmlAttribute(
+      stream, kTestsuites, "timestamp",
+      FormatEpochTimeInMillisAsIso8601(unit_test.start_timestamp()));
+  OutputXmlAttribute(stream, kTestsuites, "time",
+                     FormatTimeInMillisAsSeconds(unit_test.elapsed_time()));
+
+  if (GTEST_FLAG(shuffle)) {
+    OutputXmlAttribute(stream, kTestsuites, "random_seed",
+                       StreamableToString(unit_test.random_seed()));
+  }
+
+  *stream << TestPropertiesAsXmlAttributes(unit_test.ad_hoc_test_result());
+
+  OutputXmlAttribute(stream, kTestsuites, "name", "AllTests");
+  *stream << ">\n";
+
+  for (int i = 0; i < unit_test.total_test_case_count(); ++i) {
+    if (unit_test.GetTestCase(i)->reportable_test_count() > 0)
+      PrintXmlTestCase(stream, *unit_test.GetTestCase(i));
+  }
+  *stream << "\n";
+}
+
+// Produces a string representing the test properties in a result as space
+// delimited XML attributes based on the property key="value" pairs.
+std::string XmlUnitTestResultPrinter::TestPropertiesAsXmlAttributes(
+    const TestResult& result) {
+  Message attributes;
+  for (int i = 0; i < result.test_property_count(); ++i) {
+    const TestProperty& property = result.GetTestProperty(i);
+    attributes << " " << property.key() << "="
+        << "\"" << EscapeXmlAttribute(property.value()) << "\"";
+  }
+  return attributes.GetString();
+}
+
+// End XmlUnitTestResultPrinter
+
+#if GTEST_CAN_STREAM_RESULTS_
+
+// Checks if str contains '=', '&', '%' or '\n' characters. If yes,
+// replaces them by "%xx" where xx is their hexadecimal value. For
+// example, replaces "=" with "%3D".  This algorithm is O(strlen(str))
+// in both time and space -- important as the input str may contain an
+// arbitrarily long test failure message and stack trace.
+string StreamingListener::UrlEncode(const char* str) {
+  string result;
+  result.reserve(strlen(str) + 1);
+  for (char ch = *str; ch != '\0'; ch = *++str) {
+    switch (ch) {
+      case '%':
+      case '=':
+      case '&':
+      case '\n':
+        result.append("%" + String::FormatByte(static_cast(ch)));
+        break;
+      default:
+        result.push_back(ch);
+        break;
+    }
+  }
+  return result;
+}
+
+void StreamingListener::SocketWriter::MakeConnection() {
+  GTEST_CHECK_(sockfd_ == -1)
+      << "MakeConnection() can't be called when there is already a connection.";
+
+  addrinfo hints;
+  memset(&hints, 0, sizeof(hints));
+  hints.ai_family = AF_UNSPEC;    // To allow both IPv4 and IPv6 addresses.
+  hints.ai_socktype = SOCK_STREAM;
+  addrinfo* servinfo = NULL;
+
+  // Use the getaddrinfo() to get a linked list of IP addresses for
+  // the given host name.
+  const int error_num = getaddrinfo(
+      host_name_.c_str(), port_num_.c_str(), &hints, &servinfo);
+  if (error_num != 0) {
+    GTEST_LOG_(WARNING) << "stream_result_to: getaddrinfo() failed: "
+                        << gai_strerror(error_num);
+  }
+
+  // Loop through all the results and connect to the first we can.
+  for (addrinfo* cur_addr = servinfo; sockfd_ == -1 && cur_addr != NULL;
+       cur_addr = cur_addr->ai_next) {
+    sockfd_ = socket(
+        cur_addr->ai_family, cur_addr->ai_socktype, cur_addr->ai_protocol);
+    if (sockfd_ != -1) {
+      // Connect the client socket to the server socket.
+      if (connect(sockfd_, cur_addr->ai_addr, cur_addr->ai_addrlen) == -1) {
+        close(sockfd_);
+        sockfd_ = -1;
+      }
+    }
+  }
+
+  freeaddrinfo(servinfo);  // all done with this structure
+
+  if (sockfd_ == -1) {
+    GTEST_LOG_(WARNING) << "stream_result_to: failed to connect to "
+                        << host_name_ << ":" << port_num_;
+  }
+}
+
+// End of class Streaming Listener
+#endif  // GTEST_CAN_STREAM_RESULTS__
+
+// Class ScopedTrace
+
+// Pushes the given source file location and message onto a per-thread
+// trace stack maintained by Google Test.
+ScopedTrace::ScopedTrace(const char* file, int line, const Message& message)
+    GTEST_LOCK_EXCLUDED_(&UnitTest::mutex_) {
+  TraceInfo trace;
+  trace.file = file;
+  trace.line = line;
+  trace.message = message.GetString();
+
+  UnitTest::GetInstance()->PushGTestTrace(trace);
+}
+
+// Pops the info pushed by the c'tor.
+ScopedTrace::~ScopedTrace()
+    GTEST_LOCK_EXCLUDED_(&UnitTest::mutex_) {
+  UnitTest::GetInstance()->PopGTestTrace();
+}
+
+
+// class OsStackTraceGetter
+
+// Returns the current OS stack trace as an std::string.  Parameters:
+//
+//   max_depth  - the maximum number of stack frames to be included
+//                in the trace.
+//   skip_count - the number of top frames to be skipped; doesn't count
+//                against max_depth.
+//
+string OsStackTraceGetter::CurrentStackTrace(int /* max_depth */,
+                                             int /* skip_count */)
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  return "";
+}
+
+void OsStackTraceGetter::UponLeavingGTest()
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+}
+
+const char* const
+OsStackTraceGetter::kElidedFramesMarker =
+    "... " GTEST_NAME_ " internal frames ...";
+
+// A helper class that creates the premature-exit file in its
+// constructor and deletes the file in its destructor.
+class ScopedPrematureExitFile {
+ public:
+  explicit ScopedPrematureExitFile(const char* premature_exit_filepath)
+      : premature_exit_filepath_(premature_exit_filepath) {
+    // If a path to the premature-exit file is specified...
+    if (premature_exit_filepath != NULL && *premature_exit_filepath != '\0') {
+      // create the file with a single "0" character in it.  I/O
+      // errors are ignored as there's nothing better we can do and we
+      // don't want to fail the test because of this.
+      FILE* pfile = posix::FOpen(premature_exit_filepath, "w");
+      fwrite("0", 1, 1, pfile);
+      fclose(pfile);
+    }
+  }
+
+  ~ScopedPrematureExitFile() {
+    if (premature_exit_filepath_ != NULL && *premature_exit_filepath_ != '\0') {
+      remove(premature_exit_filepath_);
+    }
+  }
+
+ private:
+  const char* const premature_exit_filepath_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ScopedPrematureExitFile);
+};
+
+}  // namespace internal
+
+// class TestEventListeners
+
+TestEventListeners::TestEventListeners()
+    : repeater_(new internal::TestEventRepeater()),
+      default_result_printer_(NULL),
+      default_xml_generator_(NULL) {
+}
+
+TestEventListeners::~TestEventListeners() { delete repeater_; }
+
+// Returns the standard listener responsible for the default console
+// output.  Can be removed from the listeners list to shut down default
+// console output.  Note that removing this object from the listener list
+// with Release transfers its ownership to the user.
+void TestEventListeners::Append(TestEventListener* listener) {
+  repeater_->Append(listener);
+}
+
+// Removes the given event listener from the list and returns it.  It then
+// becomes the caller's responsibility to delete the listener. Returns
+// NULL if the listener is not found in the list.
+TestEventListener* TestEventListeners::Release(TestEventListener* listener) {
+  if (listener == default_result_printer_)
+    default_result_printer_ = NULL;
+  else if (listener == default_xml_generator_)
+    default_xml_generator_ = NULL;
+  return repeater_->Release(listener);
+}
+
+// Returns repeater that broadcasts the TestEventListener events to all
+// subscribers.
+TestEventListener* TestEventListeners::repeater() { return repeater_; }
+
+// Sets the default_result_printer attribute to the provided listener.
+// The listener is also added to the listener list and previous
+// default_result_printer is removed from it and deleted. The listener can
+// also be NULL in which case it will not be added to the list. Does
+// nothing if the previous and the current listener objects are the same.
+void TestEventListeners::SetDefaultResultPrinter(TestEventListener* listener) {
+  if (default_result_printer_ != listener) {
+    // It is an error to pass this method a listener that is already in the
+    // list.
+    delete Release(default_result_printer_);
+    default_result_printer_ = listener;
+    if (listener != NULL)
+      Append(listener);
+  }
+}
+
+// Sets the default_xml_generator attribute to the provided listener.  The
+// listener is also added to the listener list and previous
+// default_xml_generator is removed from it and deleted. The listener can
+// also be NULL in which case it will not be added to the list. Does
+// nothing if the previous and the current listener objects are the same.
+void TestEventListeners::SetDefaultXmlGenerator(TestEventListener* listener) {
+  if (default_xml_generator_ != listener) {
+    // It is an error to pass this method a listener that is already in the
+    // list.
+    delete Release(default_xml_generator_);
+    default_xml_generator_ = listener;
+    if (listener != NULL)
+      Append(listener);
+  }
+}
+
+// Controls whether events will be forwarded by the repeater to the
+// listeners in the list.
+bool TestEventListeners::EventForwardingEnabled() const {
+  return repeater_->forwarding_enabled();
+}
+
+void TestEventListeners::SuppressEventForwarding() {
+  repeater_->set_forwarding_enabled(false);
+}
+
+// class UnitTest
+
+// Gets the singleton UnitTest object.  The first time this method is
+// called, a UnitTest object is constructed and returned.  Consecutive
+// calls will return the same object.
+//
+// We don't protect this under mutex_ as a user is not supposed to
+// call this before main() starts, from which point on the return
+// value will never change.
+UnitTest* UnitTest::GetInstance() {
+  // When compiled with MSVC 7.1 in optimized mode, destroying the
+  // UnitTest object upon exiting the program messes up the exit code,
+  // causing successful tests to appear failed.  We have to use a
+  // different implementation in this case to bypass the compiler bug.
+  // This implementation makes the compiler happy, at the cost of
+  // leaking the UnitTest object.
+
+  // CodeGear C++Builder insists on a public destructor for the
+  // default implementation.  Use this implementation to keep good OO
+  // design with private destructor.
+
+#if (_MSC_VER == 1310 && !defined(_DEBUG)) || defined(__BORLANDC__)
+  static UnitTest* const instance = new UnitTest;
+  return instance;
+#else
+  static UnitTest instance;
+  return &instance;
+#endif  // (_MSC_VER == 1310 && !defined(_DEBUG)) || defined(__BORLANDC__)
+}
+
+// Gets the number of successful test cases.
+int UnitTest::successful_test_case_count() const {
+  return impl()->successful_test_case_count();
+}
+
+// Gets the number of failed test cases.
+int UnitTest::failed_test_case_count() const {
+  return impl()->failed_test_case_count();
+}
+
+// Gets the number of all test cases.
+int UnitTest::total_test_case_count() const {
+  return impl()->total_test_case_count();
+}
+
+// Gets the number of all test cases that contain at least one test
+// that should run.
+int UnitTest::test_case_to_run_count() const {
+  return impl()->test_case_to_run_count();
+}
+
+// Gets the number of successful tests.
+int UnitTest::successful_test_count() const {
+  return impl()->successful_test_count();
+}
+
+// Gets the number of failed tests.
+int UnitTest::failed_test_count() const { return impl()->failed_test_count(); }
+
+// Gets the number of disabled tests that will be reported in the XML report.
+int UnitTest::reportable_disabled_test_count() const {
+  return impl()->reportable_disabled_test_count();
+}
+
+// Gets the number of disabled tests.
+int UnitTest::disabled_test_count() const {
+  return impl()->disabled_test_count();
+}
+
+// Gets the number of tests to be printed in the XML report.
+int UnitTest::reportable_test_count() const {
+  return impl()->reportable_test_count();
+}
+
+// Gets the number of all tests.
+int UnitTest::total_test_count() const { return impl()->total_test_count(); }
+
+// Gets the number of tests that should run.
+int UnitTest::test_to_run_count() const { return impl()->test_to_run_count(); }
+
+// Gets the time of the test program start, in ms from the start of the
+// UNIX epoch.
+internal::TimeInMillis UnitTest::start_timestamp() const {
+    return impl()->start_timestamp();
+}
+
+// Gets the elapsed time, in milliseconds.
+internal::TimeInMillis UnitTest::elapsed_time() const {
+  return impl()->elapsed_time();
+}
+
+// Returns true iff the unit test passed (i.e. all test cases passed).
+bool UnitTest::Passed() const { return impl()->Passed(); }
+
+// Returns true iff the unit test failed (i.e. some test case failed
+// or something outside of all tests failed).
+bool UnitTest::Failed() const { return impl()->Failed(); }
+
+// Gets the i-th test case among all the test cases. i can range from 0 to
+// total_test_case_count() - 1. If i is not in that range, returns NULL.
+const TestCase* UnitTest::GetTestCase(int i) const {
+  return impl()->GetTestCase(i);
+}
+
+// Returns the TestResult containing information on test failures and
+// properties logged outside of individual test cases.
+const TestResult& UnitTest::ad_hoc_test_result() const {
+  return *impl()->ad_hoc_test_result();
+}
+
+// Gets the i-th test case among all the test cases. i can range from 0 to
+// total_test_case_count() - 1. If i is not in that range, returns NULL.
+TestCase* UnitTest::GetMutableTestCase(int i) {
+  return impl()->GetMutableTestCase(i);
+}
+
+// Returns the list of event listeners that can be used to track events
+// inside Google Test.
+TestEventListeners& UnitTest::listeners() {
+  return *impl()->listeners();
+}
+
+// Registers and returns a global test environment.  When a test
+// program is run, all global test environments will be set-up in the
+// order they were registered.  After all tests in the program have
+// finished, all global test environments will be torn-down in the
+// *reverse* order they were registered.
+//
+// The UnitTest object takes ownership of the given environment.
+//
+// We don't protect this under mutex_, as we only support calling it
+// from the main thread.
+Environment* UnitTest::AddEnvironment(Environment* env) {
+  if (env == NULL) {
+    return NULL;
+  }
+
+  impl_->environments().push_back(env);
+  return env;
+}
+
+// Adds a TestPartResult to the current TestResult object.  All Google Test
+// assertion macros (e.g. ASSERT_TRUE, EXPECT_EQ, etc) eventually call
+// this to report their results.  The user code should use the
+// assertion macros instead of calling this directly.
+void UnitTest::AddTestPartResult(
+    TestPartResult::Type result_type,
+    const char* file_name,
+    int line_number,
+    const std::string& message,
+    const std::string& os_stack_trace) GTEST_LOCK_EXCLUDED_(mutex_) {
+  Message msg;
+  msg << message;
+
+  internal::MutexLock lock(&mutex_);
+  if (impl_->gtest_trace_stack().size() > 0) {
+    msg << "\n" << GTEST_NAME_ << " trace:";
+
+    for (int i = static_cast(impl_->gtest_trace_stack().size());
+         i > 0; --i) {
+      const internal::TraceInfo& trace = impl_->gtest_trace_stack()[i - 1];
+      msg << "\n" << internal::FormatFileLocation(trace.file, trace.line)
+          << " " << trace.message;
+    }
+  }
+
+  if (os_stack_trace.c_str() != NULL && !os_stack_trace.empty()) {
+    msg << internal::kStackTraceMarker << os_stack_trace;
+  }
+
+  const TestPartResult result =
+    TestPartResult(result_type, file_name, line_number,
+                   msg.GetString().c_str());
+  impl_->GetTestPartResultReporterForCurrentThread()->
+      ReportTestPartResult(result);
+
+  if (result_type != TestPartResult::kSuccess) {
+    // gtest_break_on_failure takes precedence over
+    // gtest_throw_on_failure.  This allows a user to set the latter
+    // in the code (perhaps in order to use Google Test assertions
+    // with another testing framework) and specify the former on the
+    // command line for debugging.
+    if (GTEST_FLAG(break_on_failure)) {
+#if GTEST_OS_WINDOWS
+      // Using DebugBreak on Windows allows gtest to still break into a debugger
+      // when a failure happens and both the --gtest_break_on_failure and
+      // the --gtest_catch_exceptions flags are specified.
+      DebugBreak();
+#else
+      // Dereference NULL through a volatile pointer to prevent the compiler
+      // from removing. We use this rather than abort() or __builtin_trap() for
+      // portability: Symbian doesn't implement abort() well, and some debuggers
+      // don't correctly trap abort().
+      *static_cast(NULL) = 1;
+#endif  // GTEST_OS_WINDOWS
+    } else if (GTEST_FLAG(throw_on_failure)) {
+#if GTEST_HAS_EXCEPTIONS
+      throw internal::GoogleTestFailureException(result);
+#else
+      // We cannot call abort() as it generates a pop-up in debug mode
+      // that cannot be suppressed in VC 7.1 or below.
+      exit(1);
+#endif
+    }
+  }
+}
+
+// Adds a TestProperty to the current TestResult object when invoked from
+// inside a test, to current TestCase's ad_hoc_test_result_ when invoked
+// from SetUpTestCase or TearDownTestCase, or to the global property set
+// when invoked elsewhere.  If the result already contains a property with
+// the same key, the value will be updated.
+void UnitTest::RecordProperty(const std::string& key,
+                              const std::string& value) {
+  impl_->RecordProperty(TestProperty(key, value));
+}
+
+// Runs all tests in this UnitTest object and prints the result.
+// Returns 0 if successful, or 1 otherwise.
+//
+// We don't protect this under mutex_, as we only support calling it
+// from the main thread.
+int UnitTest::Run() {
+  const bool in_death_test_child_process =
+      internal::GTEST_FLAG(internal_run_death_test).length() > 0;
+
+  // Google Test implements this protocol for catching that a test
+  // program exits before returning control to Google Test:
+  //
+  //   1. Upon start, Google Test creates a file whose absolute path
+  //      is specified by the environment variable
+  //      TEST_PREMATURE_EXIT_FILE.
+  //   2. When Google Test has finished its work, it deletes the file.
+  //
+  // This allows a test runner to set TEST_PREMATURE_EXIT_FILE before
+  // running a Google-Test-based test program and check the existence
+  // of the file at the end of the test execution to see if it has
+  // exited prematurely.
+
+  // If we are in the child process of a death test, don't
+  // create/delete the premature exit file, as doing so is unnecessary
+  // and will confuse the parent process.  Otherwise, create/delete
+  // the file upon entering/leaving this function.  If the program
+  // somehow exits before this function has a chance to return, the
+  // premature-exit file will be left undeleted, causing a test runner
+  // that understands the premature-exit-file protocol to report the
+  // test as having failed.
+  const internal::ScopedPrematureExitFile premature_exit_file(
+      in_death_test_child_process ?
+      NULL : internal::posix::GetEnv("TEST_PREMATURE_EXIT_FILE"));
+
+  // Captures the value of GTEST_FLAG(catch_exceptions).  This value will be
+  // used for the duration of the program.
+  impl()->set_catch_exceptions(GTEST_FLAG(catch_exceptions));
+
+#if GTEST_HAS_SEH
+  // Either the user wants Google Test to catch exceptions thrown by the
+  // tests or this is executing in the context of death test child
+  // process. In either case the user does not want to see pop-up dialogs
+  // about crashes - they are expected.
+  if (impl()->catch_exceptions() || in_death_test_child_process) {
+# if !GTEST_OS_WINDOWS_MOBILE
+    // SetErrorMode doesn't exist on CE.
+    SetErrorMode(SEM_FAILCRITICALERRORS | SEM_NOALIGNMENTFAULTEXCEPT |
+                 SEM_NOGPFAULTERRORBOX | SEM_NOOPENFILEERRORBOX);
+# endif  // !GTEST_OS_WINDOWS_MOBILE
+
+# if (defined(_MSC_VER) || GTEST_OS_WINDOWS_MINGW) && !GTEST_OS_WINDOWS_MOBILE
+    // Death test children can be terminated with _abort().  On Windows,
+    // _abort() can show a dialog with a warning message.  This forces the
+    // abort message to go to stderr instead.
+    _set_error_mode(_OUT_TO_STDERR);
+# endif
+
+# if _MSC_VER >= 1400 && !GTEST_OS_WINDOWS_MOBILE
+    // In the debug version, Visual Studio pops up a separate dialog
+    // offering a choice to debug the aborted program. We need to suppress
+    // this dialog or it will pop up for every EXPECT/ASSERT_DEATH statement
+    // executed. Google Test will notify the user of any unexpected
+    // failure via stderr.
+    //
+    // VC++ doesn't define _set_abort_behavior() prior to the version 8.0.
+    // Users of prior VC versions shall suffer the agony and pain of
+    // clicking through the countless debug dialogs.
+    // TODO(vladl@google.com): find a way to suppress the abort dialog() in the
+    // debug mode when compiled with VC 7.1 or lower.
+    if (!GTEST_FLAG(break_on_failure))
+      _set_abort_behavior(
+          0x0,                                    // Clear the following flags:
+          _WRITE_ABORT_MSG | _CALL_REPORTFAULT);  // pop-up window, core dump.
+# endif
+  }
+#endif  // GTEST_HAS_SEH
+
+  return internal::HandleExceptionsInMethodIfSupported(
+      impl(),
+      &internal::UnitTestImpl::RunAllTests,
+      "auxiliary test code (environments or event listeners)") ? 0 : 1;
+}
+
+// Returns the working directory when the first TEST() or TEST_F() was
+// executed.
+const char* UnitTest::original_working_dir() const {
+  return impl_->original_working_dir_.c_str();
+}
+
+// Returns the TestCase object for the test that's currently running,
+// or NULL if no test is running.
+const TestCase* UnitTest::current_test_case() const
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  return impl_->current_test_case();
+}
+
+// Returns the TestInfo object for the test that's currently running,
+// or NULL if no test is running.
+const TestInfo* UnitTest::current_test_info() const
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  return impl_->current_test_info();
+}
+
+// Returns the random seed used at the start of the current test run.
+int UnitTest::random_seed() const { return impl_->random_seed(); }
+
+#if GTEST_HAS_PARAM_TEST
+// Returns ParameterizedTestCaseRegistry object used to keep track of
+// value-parameterized tests and instantiate and register them.
+internal::ParameterizedTestCaseRegistry&
+    UnitTest::parameterized_test_registry()
+        GTEST_LOCK_EXCLUDED_(mutex_) {
+  return impl_->parameterized_test_registry();
+}
+#endif  // GTEST_HAS_PARAM_TEST
+
+// Creates an empty UnitTest.
+UnitTest::UnitTest() {
+  impl_ = new internal::UnitTestImpl(this);
+}
+
+// Destructor of UnitTest.
+UnitTest::~UnitTest() {
+  delete impl_;
+}
+
+// Pushes a trace defined by SCOPED_TRACE() on to the per-thread
+// Google Test trace stack.
+void UnitTest::PushGTestTrace(const internal::TraceInfo& trace)
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  impl_->gtest_trace_stack().push_back(trace);
+}
+
+// Pops a trace from the per-thread Google Test trace stack.
+void UnitTest::PopGTestTrace()
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  impl_->gtest_trace_stack().pop_back();
+}
+
+namespace internal {
+
+UnitTestImpl::UnitTestImpl(UnitTest* parent)
+    : parent_(parent),
+#ifdef _MSC_VER
+# pragma warning(push)                    // Saves the current warning state.
+# pragma warning(disable:4355)            // Temporarily disables warning 4355
+                                         // (using this in initializer).
+      default_global_test_part_result_reporter_(this),
+      default_per_thread_test_part_result_reporter_(this),
+# pragma warning(pop)                     // Restores the warning state again.
+#else
+      default_global_test_part_result_reporter_(this),
+      default_per_thread_test_part_result_reporter_(this),
+#endif  // _MSC_VER
+      global_test_part_result_repoter_(
+          &default_global_test_part_result_reporter_),
+      per_thread_test_part_result_reporter_(
+          &default_per_thread_test_part_result_reporter_),
+#if GTEST_HAS_PARAM_TEST
+      parameterized_test_registry_(),
+      parameterized_tests_registered_(false),
+#endif  // GTEST_HAS_PARAM_TEST
+      last_death_test_case_(-1),
+      current_test_case_(NULL),
+      current_test_info_(NULL),
+      ad_hoc_test_result_(),
+      os_stack_trace_getter_(NULL),
+      post_flag_parse_init_performed_(false),
+      random_seed_(0),  // Will be overridden by the flag before first use.
+      random_(0),  // Will be reseeded before first use.
+      start_timestamp_(0),
+      elapsed_time_(0),
+#if GTEST_HAS_DEATH_TEST
+      death_test_factory_(new DefaultDeathTestFactory),
+#endif
+      // Will be overridden by the flag before first use.
+      catch_exceptions_(false) {
+  listeners()->SetDefaultResultPrinter(new PrettyUnitTestResultPrinter);
+}
+
+UnitTestImpl::~UnitTestImpl() {
+  // Deletes every TestCase.
+  ForEach(test_cases_, internal::Delete);
+
+  // Deletes every Environment.
+  ForEach(environments_, internal::Delete);
+
+  delete os_stack_trace_getter_;
+}
+
+// Adds a TestProperty to the current TestResult object when invoked in a
+// context of a test, to current test case's ad_hoc_test_result when invoke
+// from SetUpTestCase/TearDownTestCase, or to the global property set
+// otherwise.  If the result already contains a property with the same key,
+// the value will be updated.
+void UnitTestImpl::RecordProperty(const TestProperty& test_property) {
+  std::string xml_element;
+  TestResult* test_result;  // TestResult appropriate for property recording.
+
+  if (current_test_info_ != NULL) {
+    xml_element = "testcase";
+    test_result = &(current_test_info_->result_);
+  } else if (current_test_case_ != NULL) {
+    xml_element = "testsuite";
+    test_result = &(current_test_case_->ad_hoc_test_result_);
+  } else {
+    xml_element = "testsuites";
+    test_result = &ad_hoc_test_result_;
+  }
+  test_result->RecordProperty(xml_element, test_property);
+}
+
+#if GTEST_HAS_DEATH_TEST
+// Disables event forwarding if the control is currently in a death test
+// subprocess. Must not be called before InitGoogleTest.
+void UnitTestImpl::SuppressTestEventsIfInSubprocess() {
+  if (internal_run_death_test_flag_.get() != NULL)
+    listeners()->SuppressEventForwarding();
+}
+#endif  // GTEST_HAS_DEATH_TEST
+
+// Initializes event listeners performing XML output as specified by
+// UnitTestOptions. Must not be called before InitGoogleTest.
+void UnitTestImpl::ConfigureXmlOutput() {
+  const std::string& output_format = UnitTestOptions::GetOutputFormat();
+  if (output_format == "xml") {
+    listeners()->SetDefaultXmlGenerator(new XmlUnitTestResultPrinter(
+        UnitTestOptions::GetAbsolutePathToOutputFile().c_str()));
+  } else if (output_format != "") {
+    printf("WARNING: unrecognized output format \"%s\" ignored.\n",
+           output_format.c_str());
+    fflush(stdout);
+  }
+}
+
+#if GTEST_CAN_STREAM_RESULTS_
+// Initializes event listeners for streaming test results in string form.
+// Must not be called before InitGoogleTest.
+void UnitTestImpl::ConfigureStreamingOutput() {
+  const std::string& target = GTEST_FLAG(stream_result_to);
+  if (!target.empty()) {
+    const size_t pos = target.find(':');
+    if (pos != std::string::npos) {
+      listeners()->Append(new StreamingListener(target.substr(0, pos),
+                                                target.substr(pos+1)));
+    } else {
+      printf("WARNING: unrecognized streaming target \"%s\" ignored.\n",
+             target.c_str());
+      fflush(stdout);
+    }
+  }
+}
+#endif  // GTEST_CAN_STREAM_RESULTS_
+
+// Performs initialization dependent upon flag values obtained in
+// ParseGoogleTestFlagsOnly.  Is called from InitGoogleTest after the call to
+// ParseGoogleTestFlagsOnly.  In case a user neglects to call InitGoogleTest
+// this function is also called from RunAllTests.  Since this function can be
+// called more than once, it has to be idempotent.
+void UnitTestImpl::PostFlagParsingInit() {
+  // Ensures that this function does not execute more than once.
+  if (!post_flag_parse_init_performed_) {
+    post_flag_parse_init_performed_ = true;
+
+#if GTEST_HAS_DEATH_TEST
+    InitDeathTestSubprocessControlInfo();
+    SuppressTestEventsIfInSubprocess();
+#endif  // GTEST_HAS_DEATH_TEST
+
+    // Registers parameterized tests. This makes parameterized tests
+    // available to the UnitTest reflection API without running
+    // RUN_ALL_TESTS.
+    RegisterParameterizedTests();
+
+    // Configures listeners for XML output. This makes it possible for users
+    // to shut down the default XML output before invoking RUN_ALL_TESTS.
+    ConfigureXmlOutput();
+
+#if GTEST_CAN_STREAM_RESULTS_
+    // Configures listeners for streaming test results to the specified server.
+    ConfigureStreamingOutput();
+#endif  // GTEST_CAN_STREAM_RESULTS_
+  }
+}
+
+// A predicate that checks the name of a TestCase against a known
+// value.
+//
+// This is used for implementation of the UnitTest class only.  We put
+// it in the anonymous namespace to prevent polluting the outer
+// namespace.
+//
+// TestCaseNameIs is copyable.
+class TestCaseNameIs {
+ public:
+  // Constructor.
+  explicit TestCaseNameIs(const std::string& name)
+      : name_(name) {}
+
+  // Returns true iff the name of test_case matches name_.
+  bool operator()(const TestCase* test_case) const {
+    return test_case != NULL && strcmp(test_case->name(), name_.c_str()) == 0;
+  }
+
+ private:
+  std::string name_;
+};
+
+// Finds and returns a TestCase with the given name.  If one doesn't
+// exist, creates one and returns it.  It's the CALLER'S
+// RESPONSIBILITY to ensure that this function is only called WHEN THE
+// TESTS ARE NOT SHUFFLED.
+//
+// Arguments:
+//
+//   test_case_name: name of the test case
+//   type_param:     the name of the test case's type parameter, or NULL if
+//                   this is not a typed or a type-parameterized test case.
+//   set_up_tc:      pointer to the function that sets up the test case
+//   tear_down_tc:   pointer to the function that tears down the test case
+TestCase* UnitTestImpl::GetTestCase(const char* test_case_name,
+                                    const char* type_param,
+                                    Test::SetUpTestCaseFunc set_up_tc,
+                                    Test::TearDownTestCaseFunc tear_down_tc) {
+  // Can we find a TestCase with the given name?
+  const std::vector::const_iterator test_case =
+      std::find_if(test_cases_.begin(), test_cases_.end(),
+                   TestCaseNameIs(test_case_name));
+
+  if (test_case != test_cases_.end())
+    return *test_case;
+
+  // No.  Let's create one.
+  TestCase* const new_test_case =
+      new TestCase(test_case_name, type_param, set_up_tc, tear_down_tc);
+
+  // Is this a death test case?
+  if (internal::UnitTestOptions::MatchesFilter(test_case_name,
+                                               kDeathTestCaseFilter)) {
+    // Yes.  Inserts the test case after the last death test case
+    // defined so far.  This only works when the test cases haven't
+    // been shuffled.  Otherwise we may end up running a death test
+    // after a non-death test.
+    ++last_death_test_case_;
+    test_cases_.insert(test_cases_.begin() + last_death_test_case_,
+                       new_test_case);
+  } else {
+    // No.  Appends to the end of the list.
+    test_cases_.push_back(new_test_case);
+  }
+
+  test_case_indices_.push_back(static_cast(test_case_indices_.size()));
+  return new_test_case;
+}
+
+// Helpers for setting up / tearing down the given environment.  They
+// are for use in the ForEach() function.
+static void SetUpEnvironment(Environment* env) { env->SetUp(); }
+static void TearDownEnvironment(Environment* env) { env->TearDown(); }
+
+// Runs all tests in this UnitTest object, prints the result, and
+// returns true if all tests are successful.  If any exception is
+// thrown during a test, the test is considered to be failed, but the
+// rest of the tests will still be run.
+//
+// When parameterized tests are enabled, it expands and registers
+// parameterized tests first in RegisterParameterizedTests().
+// All other functions called from RunAllTests() may safely assume that
+// parameterized tests are ready to be counted and run.
+bool UnitTestImpl::RunAllTests() {
+  // Makes sure InitGoogleTest() was called.
+  if (!GTestIsInitialized()) {
+    printf("%s",
+           "\nThis test program did NOT call ::testing::InitGoogleTest "
+           "before calling RUN_ALL_TESTS().  Please fix it.\n");
+    return false;
+  }
+
+  // Do not run any test if the --help flag was specified.
+  if (g_help_flag)
+    return true;
+
+  // Repeats the call to the post-flag parsing initialization in case the
+  // user didn't call InitGoogleTest.
+  PostFlagParsingInit();
+
+  // Even if sharding is not on, test runners may want to use the
+  // GTEST_SHARD_STATUS_FILE to query whether the test supports the sharding
+  // protocol.
+  internal::WriteToShardStatusFileIfNeeded();
+
+  // True iff we are in a subprocess for running a thread-safe-style
+  // death test.
+  bool in_subprocess_for_death_test = false;
+
+#if GTEST_HAS_DEATH_TEST
+  in_subprocess_for_death_test = (internal_run_death_test_flag_.get() != NULL);
+#endif  // GTEST_HAS_DEATH_TEST
+
+  const bool should_shard = ShouldShard(kTestTotalShards, kTestShardIndex,
+                                        in_subprocess_for_death_test);
+
+  // Compares the full test names with the filter to decide which
+  // tests to run.
+  const bool has_tests_to_run = FilterTests(should_shard
+                                              ? HONOR_SHARDING_PROTOCOL
+                                              : IGNORE_SHARDING_PROTOCOL) > 0;
+
+  // Lists the tests and exits if the --gtest_list_tests flag was specified.
+  if (GTEST_FLAG(list_tests)) {
+    // This must be called *after* FilterTests() has been called.
+    ListTestsMatchingFilter();
+    return true;
+  }
+
+  random_seed_ = GTEST_FLAG(shuffle) ?
+      GetRandomSeedFromFlag(GTEST_FLAG(random_seed)) : 0;
+
+  // True iff at least one test has failed.
+  bool failed = false;
+
+  TestEventListener* repeater = listeners()->repeater();
+
+  start_timestamp_ = GetTimeInMillis();
+  repeater->OnTestProgramStart(*parent_);
+
+  // How many times to repeat the tests?  We don't want to repeat them
+  // when we are inside the subprocess of a death test.
+  const int repeat = in_subprocess_for_death_test ? 1 : GTEST_FLAG(repeat);
+  // Repeats forever if the repeat count is negative.
+  const bool forever = repeat < 0;
+  for (int i = 0; forever || i != repeat; i++) {
+    // We want to preserve failures generated by ad-hoc test
+    // assertions executed before RUN_ALL_TESTS().
+    ClearNonAdHocTestResult();
+
+    const TimeInMillis start = GetTimeInMillis();
+
+    // Shuffles test cases and tests if requested.
+    if (has_tests_to_run && GTEST_FLAG(shuffle)) {
+      random()->Reseed(random_seed_);
+      // This should be done before calling OnTestIterationStart(),
+      // such that a test event listener can see the actual test order
+      // in the event.
+      ShuffleTests();
+    }
+
+    // Tells the unit test event listeners that the tests are about to start.
+    repeater->OnTestIterationStart(*parent_, i);
+
+    // Runs each test case if there is at least one test to run.
+    if (has_tests_to_run) {
+      // Sets up all environments beforehand.
+      repeater->OnEnvironmentsSetUpStart(*parent_);
+      ForEach(environments_, SetUpEnvironment);
+      repeater->OnEnvironmentsSetUpEnd(*parent_);
+
+      // Runs the tests only if there was no fatal failure during global
+      // set-up.
+      if (!Test::HasFatalFailure()) {
+        for (int test_index = 0; test_index < total_test_case_count();
+             test_index++) {
+          GetMutableTestCase(test_index)->Run();
+        }
+      }
+
+      // Tears down all environments in reverse order afterwards.
+      repeater->OnEnvironmentsTearDownStart(*parent_);
+      std::for_each(environments_.rbegin(), environments_.rend(),
+                    TearDownEnvironment);
+      repeater->OnEnvironmentsTearDownEnd(*parent_);
+    }
+
+    elapsed_time_ = GetTimeInMillis() - start;
+
+    // Tells the unit test event listener that the tests have just finished.
+    repeater->OnTestIterationEnd(*parent_, i);
+
+    // Gets the result and clears it.
+    if (!Passed()) {
+      failed = true;
+    }
+
+    // Restores the original test order after the iteration.  This
+    // allows the user to quickly repro a failure that happens in the
+    // N-th iteration without repeating the first (N - 1) iterations.
+    // This is not enclosed in "if (GTEST_FLAG(shuffle)) { ... }", in
+    // case the user somehow changes the value of the flag somewhere
+    // (it's always safe to unshuffle the tests).
+    UnshuffleTests();
+
+    if (GTEST_FLAG(shuffle)) {
+      // Picks a new random seed for each iteration.
+      random_seed_ = GetNextRandomSeed(random_seed_);
+    }
+  }
+
+  repeater->OnTestProgramEnd(*parent_);
+
+  return !failed;
+}
+
+// Reads the GTEST_SHARD_STATUS_FILE environment variable, and creates the file
+// if the variable is present. If a file already exists at this location, this
+// function will write over it. If the variable is present, but the file cannot
+// be created, prints an error and exits.
+void WriteToShardStatusFileIfNeeded() {
+  const char* const test_shard_file = posix::GetEnv(kTestShardStatusFile);
+  if (test_shard_file != NULL) {
+    FILE* const file = posix::FOpen(test_shard_file, "w");
+    if (file == NULL) {
+      ColoredPrintf(COLOR_RED,
+                    "Could not write to the test shard status file \"%s\" "
+                    "specified by the %s environment variable.\n",
+                    test_shard_file, kTestShardStatusFile);
+      fflush(stdout);
+      exit(EXIT_FAILURE);
+    }
+    fclose(file);
+  }
+}
+
+// Checks whether sharding is enabled by examining the relevant
+// environment variable values. If the variables are present,
+// but inconsistent (i.e., shard_index >= total_shards), prints
+// an error and exits. If in_subprocess_for_death_test, sharding is
+// disabled because it must only be applied to the original test
+// process. Otherwise, we could filter out death tests we intended to execute.
+bool ShouldShard(const char* total_shards_env,
+                 const char* shard_index_env,
+                 bool in_subprocess_for_death_test) {
+  if (in_subprocess_for_death_test) {
+    return false;
+  }
+
+  const Int32 total_shards = Int32FromEnvOrDie(total_shards_env, -1);
+  const Int32 shard_index = Int32FromEnvOrDie(shard_index_env, -1);
+
+  if (total_shards == -1 && shard_index == -1) {
+    return false;
+  } else if (total_shards == -1 && shard_index != -1) {
+    const Message msg = Message()
+      << "Invalid environment variables: you have "
+      << kTestShardIndex << " = " << shard_index
+      << ", but have left " << kTestTotalShards << " unset.\n";
+    ColoredPrintf(COLOR_RED, msg.GetString().c_str());
+    fflush(stdout);
+    exit(EXIT_FAILURE);
+  } else if (total_shards != -1 && shard_index == -1) {
+    const Message msg = Message()
+      << "Invalid environment variables: you have "
+      << kTestTotalShards << " = " << total_shards
+      << ", but have left " << kTestShardIndex << " unset.\n";
+    ColoredPrintf(COLOR_RED, msg.GetString().c_str());
+    fflush(stdout);
+    exit(EXIT_FAILURE);
+  } else if (shard_index < 0 || shard_index >= total_shards) {
+    const Message msg = Message()
+      << "Invalid environment variables: we require 0 <= "
+      << kTestShardIndex << " < " << kTestTotalShards
+      << ", but you have " << kTestShardIndex << "=" << shard_index
+      << ", " << kTestTotalShards << "=" << total_shards << ".\n";
+    ColoredPrintf(COLOR_RED, msg.GetString().c_str());
+    fflush(stdout);
+    exit(EXIT_FAILURE);
+  }
+
+  return total_shards > 1;
+}
+
+// Parses the environment variable var as an Int32. If it is unset,
+// returns default_val. If it is not an Int32, prints an error
+// and aborts.
+Int32 Int32FromEnvOrDie(const char* var, Int32 default_val) {
+  const char* str_val = posix::GetEnv(var);
+  if (str_val == NULL) {
+    return default_val;
+  }
+
+  Int32 result;
+  if (!ParseInt32(Message() << "The value of environment variable " << var,
+                  str_val, &result)) {
+    exit(EXIT_FAILURE);
+  }
+  return result;
+}
+
+// Given the total number of shards, the shard index, and the test id,
+// returns true iff the test should be run on this shard. The test id is
+// some arbitrary but unique non-negative integer assigned to each test
+// method. Assumes that 0 <= shard_index < total_shards.
+bool ShouldRunTestOnShard(int total_shards, int shard_index, int test_id) {
+  return (test_id % total_shards) == shard_index;
+}
+
+// Compares the name of each test with the user-specified filter to
+// decide whether the test should be run, then records the result in
+// each TestCase and TestInfo object.
+// If shard_tests == true, further filters tests based on sharding
+// variables in the environment - see
+// http://code.google.com/p/googletest/wiki/GoogleTestAdvancedGuide.
+// Returns the number of tests that should run.
+int UnitTestImpl::FilterTests(ReactionToSharding shard_tests) {
+  const Int32 total_shards = shard_tests == HONOR_SHARDING_PROTOCOL ?
+      Int32FromEnvOrDie(kTestTotalShards, -1) : -1;
+  const Int32 shard_index = shard_tests == HONOR_SHARDING_PROTOCOL ?
+      Int32FromEnvOrDie(kTestShardIndex, -1) : -1;
+
+  // num_runnable_tests are the number of tests that will
+  // run across all shards (i.e., match filter and are not disabled).
+  // num_selected_tests are the number of tests to be run on
+  // this shard.
+  int num_runnable_tests = 0;
+  int num_selected_tests = 0;
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    TestCase* const test_case = test_cases_[i];
+    const std::string &test_case_name = test_case->name();
+    test_case->set_should_run(false);
+
+    for (size_t j = 0; j < test_case->test_info_list().size(); j++) {
+      TestInfo* const test_info = test_case->test_info_list()[j];
+      const std::string test_name(test_info->name());
+      // A test is disabled if test case name or test name matches
+      // kDisableTestFilter.
+      const bool is_disabled =
+          internal::UnitTestOptions::MatchesFilter(test_case_name,
+                                                   kDisableTestFilter) ||
+          internal::UnitTestOptions::MatchesFilter(test_name,
+                                                   kDisableTestFilter);
+      test_info->is_disabled_ = is_disabled;
+
+      const bool matches_filter =
+          internal::UnitTestOptions::FilterMatchesTest(test_case_name,
+                                                       test_name);
+      test_info->matches_filter_ = matches_filter;
+
+      const bool is_runnable =
+          (GTEST_FLAG(also_run_disabled_tests) || !is_disabled) &&
+          matches_filter;
+
+      const bool is_selected = is_runnable &&
+          (shard_tests == IGNORE_SHARDING_PROTOCOL ||
+           ShouldRunTestOnShard(total_shards, shard_index,
+                                num_runnable_tests));
+
+      num_runnable_tests += is_runnable;
+      num_selected_tests += is_selected;
+
+      test_info->should_run_ = is_selected;
+      test_case->set_should_run(test_case->should_run() || is_selected);
+    }
+  }
+  return num_selected_tests;
+}
+
+// Prints the given C-string on a single line by replacing all '\n'
+// characters with string "\\n".  If the output takes more than
+// max_length characters, only prints the first max_length characters
+// and "...".
+static void PrintOnOneLine(const char* str, int max_length) {
+  if (str != NULL) {
+    for (int i = 0; *str != '\0'; ++str) {
+      if (i >= max_length) {
+        printf("...");
+        break;
+      }
+      if (*str == '\n') {
+        printf("\\n");
+        i += 2;
+      } else {
+        printf("%c", *str);
+        ++i;
+      }
+    }
+  }
+}
+
+// Prints the names of the tests matching the user-specified filter flag.
+void UnitTestImpl::ListTestsMatchingFilter() {
+  // Print at most this many characters for each type/value parameter.
+  const int kMaxParamLength = 250;
+
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    const TestCase* const test_case = test_cases_[i];
+    bool printed_test_case_name = false;
+
+    for (size_t j = 0; j < test_case->test_info_list().size(); j++) {
+      const TestInfo* const test_info =
+          test_case->test_info_list()[j];
+      if (test_info->matches_filter_) {
+        if (!printed_test_case_name) {
+          printed_test_case_name = true;
+          printf("%s.", test_case->name());
+          if (test_case->type_param() != NULL) {
+            printf("  # %s = ", kTypeParamLabel);
+            // We print the type parameter on a single line to make
+            // the output easy to parse by a program.
+            PrintOnOneLine(test_case->type_param(), kMaxParamLength);
+          }
+          printf("\n");
+        }
+        printf("  %s", test_info->name());
+        if (test_info->value_param() != NULL) {
+          printf("  # %s = ", kValueParamLabel);
+          // We print the value parameter on a single line to make the
+          // output easy to parse by a program.
+          PrintOnOneLine(test_info->value_param(), kMaxParamLength);
+        }
+        printf("\n");
+      }
+    }
+  }
+  fflush(stdout);
+}
+
+// Sets the OS stack trace getter.
+//
+// Does nothing if the input and the current OS stack trace getter are
+// the same; otherwise, deletes the old getter and makes the input the
+// current getter.
+void UnitTestImpl::set_os_stack_trace_getter(
+    OsStackTraceGetterInterface* getter) {
+  if (os_stack_trace_getter_ != getter) {
+    delete os_stack_trace_getter_;
+    os_stack_trace_getter_ = getter;
+  }
+}
+
+// Returns the current OS stack trace getter if it is not NULL;
+// otherwise, creates an OsStackTraceGetter, makes it the current
+// getter, and returns it.
+OsStackTraceGetterInterface* UnitTestImpl::os_stack_trace_getter() {
+  if (os_stack_trace_getter_ == NULL) {
+    os_stack_trace_getter_ = new OsStackTraceGetter;
+  }
+
+  return os_stack_trace_getter_;
+}
+
+// Returns the TestResult for the test that's currently running, or
+// the TestResult for the ad hoc test if no test is running.
+TestResult* UnitTestImpl::current_test_result() {
+  return current_test_info_ ?
+      &(current_test_info_->result_) : &ad_hoc_test_result_;
+}
+
+// Shuffles all test cases, and the tests within each test case,
+// making sure that death tests are still run first.
+void UnitTestImpl::ShuffleTests() {
+  // Shuffles the death test cases.
+  ShuffleRange(random(), 0, last_death_test_case_ + 1, &test_case_indices_);
+
+  // Shuffles the non-death test cases.
+  ShuffleRange(random(), last_death_test_case_ + 1,
+               static_cast(test_cases_.size()), &test_case_indices_);
+
+  // Shuffles the tests inside each test case.
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    test_cases_[i]->ShuffleTests(random());
+  }
+}
+
+// Restores the test cases and tests to their order before the first shuffle.
+void UnitTestImpl::UnshuffleTests() {
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    // Unshuffles the tests in each test case.
+    test_cases_[i]->UnshuffleTests();
+    // Resets the index of each test case.
+    test_case_indices_[i] = static_cast(i);
+  }
+}
+
+// Returns the current OS stack trace as an std::string.
+//
+// The maximum number of stack frames to be included is specified by
+// the gtest_stack_trace_depth flag.  The skip_count parameter
+// specifies the number of top frames to be skipped, which doesn't
+// count against the number of frames to be included.
+//
+// For example, if Foo() calls Bar(), which in turn calls
+// GetCurrentOsStackTraceExceptTop(..., 1), Foo() will be included in
+// the trace but Bar() and GetCurrentOsStackTraceExceptTop() won't.
+std::string GetCurrentOsStackTraceExceptTop(UnitTest* /*unit_test*/,
+                                            int skip_count) {
+  // We pass skip_count + 1 to skip this wrapper function in addition
+  // to what the user really wants to skip.
+  return GetUnitTestImpl()->CurrentOsStackTraceExceptTop(skip_count + 1);
+}
+
+// Used by the GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_ macro to
+// suppress unreachable code warnings.
+namespace {
+class ClassUniqueToAlwaysTrue {};
+}
+
+bool IsTrue(bool condition) { return condition; }
+
+bool AlwaysTrue() {
+#if GTEST_HAS_EXCEPTIONS
+  // This condition is always false so AlwaysTrue() never actually throws,
+  // but it makes the compiler think that it may throw.
+  if (IsTrue(false))
+    throw ClassUniqueToAlwaysTrue();
+#endif  // GTEST_HAS_EXCEPTIONS
+  return true;
+}
+
+// If *pstr starts with the given prefix, modifies *pstr to be right
+// past the prefix and returns true; otherwise leaves *pstr unchanged
+// and returns false.  None of pstr, *pstr, and prefix can be NULL.
+bool SkipPrefix(const char* prefix, const char** pstr) {
+  const size_t prefix_len = strlen(prefix);
+  if (strncmp(*pstr, prefix, prefix_len) == 0) {
+    *pstr += prefix_len;
+    return true;
+  }
+  return false;
+}
+
+// Parses a string as a command line flag.  The string should have
+// the format "--flag=value".  When def_optional is true, the "=value"
+// part can be omitted.
+//
+// Returns the value of the flag, or NULL if the parsing failed.
+const char* ParseFlagValue(const char* str,
+                           const char* flag,
+                           bool def_optional) {
+  // str and flag must not be NULL.
+  if (str == NULL || flag == NULL) return NULL;
+
+  // The flag must start with "--" followed by GTEST_FLAG_PREFIX_.
+  const std::string flag_str = std::string("--") + GTEST_FLAG_PREFIX_ + flag;
+  const size_t flag_len = flag_str.length();
+  if (strncmp(str, flag_str.c_str(), flag_len) != 0) return NULL;
+
+  // Skips the flag name.
+  const char* flag_end = str + flag_len;
+
+  // When def_optional is true, it's OK to not have a "=value" part.
+  if (def_optional && (flag_end[0] == '\0')) {
+    return flag_end;
+  }
+
+  // If def_optional is true and there are more characters after the
+  // flag name, or if def_optional is false, there must be a '=' after
+  // the flag name.
+  if (flag_end[0] != '=') return NULL;
+
+  // Returns the string after "=".
+  return flag_end + 1;
+}
+
+// Parses a string for a bool flag, in the form of either
+// "--flag=value" or "--flag".
+//
+// In the former case, the value is taken as true as long as it does
+// not start with '0', 'f', or 'F'.
+//
+// In the latter case, the value is taken as true.
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseBoolFlag(const char* str, const char* flag, bool* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, true);
+
+  // Aborts if the parsing failed.
+  if (value_str == NULL) return false;
+
+  // Converts the string value to a bool.
+  *value = !(*value_str == '0' || *value_str == 'f' || *value_str == 'F');
+  return true;
+}
+
+// Parses a string for an Int32 flag, in the form of
+// "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseInt32Flag(const char* str, const char* flag, Int32* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, false);
+
+  // Aborts if the parsing failed.
+  if (value_str == NULL) return false;
+
+  // Sets *value to the value of the flag.
+  return ParseInt32(Message() << "The value of flag --" << flag,
+                    value_str, value);
+}
+
+// Parses a string for a string flag, in the form of
+// "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseStringFlag(const char* str, const char* flag, std::string* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, false);
+
+  // Aborts if the parsing failed.
+  if (value_str == NULL) return false;
+
+  // Sets *value to the value of the flag.
+  *value = value_str;
+  return true;
+}
+
+// Determines whether a string has a prefix that Google Test uses for its
+// flags, i.e., starts with GTEST_FLAG_PREFIX_ or GTEST_FLAG_PREFIX_DASH_.
+// If Google Test detects that a command line flag has its prefix but is not
+// recognized, it will print its help message. Flags starting with
+// GTEST_INTERNAL_PREFIX_ followed by "internal_" are considered Google Test
+// internal flags and do not trigger the help message.
+static bool HasGoogleTestFlagPrefix(const char* str) {
+  return (SkipPrefix("--", &str) ||
+          SkipPrefix("-", &str) ||
+          SkipPrefix("/", &str)) &&
+         !SkipPrefix(GTEST_FLAG_PREFIX_ "internal_", &str) &&
+         (SkipPrefix(GTEST_FLAG_PREFIX_, &str) ||
+          SkipPrefix(GTEST_FLAG_PREFIX_DASH_, &str));
+}
+
+// Prints a string containing code-encoded text.  The following escape
+// sequences can be used in the string to control the text color:
+//
+//   @@    prints a single '@' character.
+//   @R    changes the color to red.
+//   @G    changes the color to green.
+//   @Y    changes the color to yellow.
+//   @D    changes to the default terminal text color.
+//
+// TODO(wan@google.com): Write tests for this once we add stdout
+// capturing to Google Test.
+static void PrintColorEncoded(const char* str) {
+  GTestColor color = COLOR_DEFAULT;  // The current color.
+
+  // Conceptually, we split the string into segments divided by escape
+  // sequences.  Then we print one segment at a time.  At the end of
+  // each iteration, the str pointer advances to the beginning of the
+  // next segment.
+  for (;;) {
+    const char* p = strchr(str, '@');
+    if (p == NULL) {
+      ColoredPrintf(color, "%s", str);
+      return;
+    }
+
+    ColoredPrintf(color, "%s", std::string(str, p).c_str());
+
+    const char ch = p[1];
+    str = p + 2;
+    if (ch == '@') {
+      ColoredPrintf(color, "@");
+    } else if (ch == 'D') {
+      color = COLOR_DEFAULT;
+    } else if (ch == 'R') {
+      color = COLOR_RED;
+    } else if (ch == 'G') {
+      color = COLOR_GREEN;
+    } else if (ch == 'Y') {
+      color = COLOR_YELLOW;
+    } else {
+      --str;
+    }
+  }
+}
+
+static const char kColorEncodedHelpMessage[] =
+"This program contains tests written using " GTEST_NAME_ ". You can use the\n"
+"following command line flags to control its behavior:\n"
+"\n"
+"Test Selection:\n"
+"  @G--" GTEST_FLAG_PREFIX_ "list_tests@D\n"
+"      List the names of all tests instead of running them. The name of\n"
+"      TEST(Foo, Bar) is \"Foo.Bar\".\n"
+"  @G--" GTEST_FLAG_PREFIX_ "filter=@YPOSTIVE_PATTERNS"
+    "[@G-@YNEGATIVE_PATTERNS]@D\n"
+"      Run only the tests whose name matches one of the positive patterns but\n"
+"      none of the negative patterns. '?' matches any single character; '*'\n"
+"      matches any substring; ':' separates two patterns.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "also_run_disabled_tests@D\n"
+"      Run all disabled tests too.\n"
+"\n"
+"Test Execution:\n"
+"  @G--" GTEST_FLAG_PREFIX_ "repeat=@Y[COUNT]@D\n"
+"      Run the tests repeatedly; use a negative count to repeat forever.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "shuffle@D\n"
+"      Randomize tests' orders on every iteration.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "random_seed=@Y[NUMBER]@D\n"
+"      Random number seed to use for shuffling test orders (between 1 and\n"
+"      99999, or 0 to use a seed based on the current time).\n"
+"\n"
+"Test Output:\n"
+"  @G--" GTEST_FLAG_PREFIX_ "color=@Y(@Gyes@Y|@Gno@Y|@Gauto@Y)@D\n"
+"      Enable/disable colored output. The default is @Gauto@D.\n"
+"  -@G-" GTEST_FLAG_PREFIX_ "print_time=0@D\n"
+"      Don't print the elapsed time of each test.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "output=xml@Y[@G:@YDIRECTORY_PATH@G"
+    GTEST_PATH_SEP_ "@Y|@G:@YFILE_PATH]@D\n"
+"      Generate an XML report in the given directory or with the given file\n"
+"      name. @YFILE_PATH@D defaults to @Gtest_details.xml@D.\n"
+#if GTEST_CAN_STREAM_RESULTS_
+"  @G--" GTEST_FLAG_PREFIX_ "stream_result_to=@YHOST@G:@YPORT@D\n"
+"      Stream test results to the given server.\n"
+#endif  // GTEST_CAN_STREAM_RESULTS_
+"\n"
+"Assertion Behavior:\n"
+#if GTEST_HAS_DEATH_TEST && !GTEST_OS_WINDOWS
+"  @G--" GTEST_FLAG_PREFIX_ "death_test_style=@Y(@Gfast@Y|@Gthreadsafe@Y)@D\n"
+"      Set the default death test style.\n"
+#endif  // GTEST_HAS_DEATH_TEST && !GTEST_OS_WINDOWS
+"  @G--" GTEST_FLAG_PREFIX_ "break_on_failure@D\n"
+"      Turn assertion failures into debugger break-points.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "throw_on_failure@D\n"
+"      Turn assertion failures into C++ exceptions.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "catch_exceptions=0@D\n"
+"      Do not report exceptions as test failures. Instead, allow them\n"
+"      to crash the program or throw a pop-up (on Windows).\n"
+"\n"
+"Except for @G--" GTEST_FLAG_PREFIX_ "list_tests@D, you can alternatively set "
+    "the corresponding\n"
+"environment variable of a flag (all letters in upper-case). For example, to\n"
+"disable colored text output, you can either specify @G--" GTEST_FLAG_PREFIX_
+    "color=no@D or set\n"
+"the @G" GTEST_FLAG_PREFIX_UPPER_ "COLOR@D environment variable to @Gno@D.\n"
+"\n"
+"For more information, please read the " GTEST_NAME_ " documentation at\n"
+"@G" GTEST_PROJECT_URL_ "@D. If you find a bug in " GTEST_NAME_ "\n"
+"(not one in your own code or tests), please report it to\n"
+"@G<" GTEST_DEV_EMAIL_ ">@D.\n";
+
+// Parses the command line for Google Test flags, without initializing
+// other parts of Google Test.  The type parameter CharType can be
+// instantiated to either char or wchar_t.
+template 
+void ParseGoogleTestFlagsOnlyImpl(int* argc, CharType** argv) {
+  for (int i = 1; i < *argc; i++) {
+    const std::string arg_string = StreamableToString(argv[i]);
+    const char* const arg = arg_string.c_str();
+
+    using internal::ParseBoolFlag;
+    using internal::ParseInt32Flag;
+    using internal::ParseStringFlag;
+
+    // Do we see a Google Test flag?
+    if (ParseBoolFlag(arg, kAlsoRunDisabledTestsFlag,
+                      >EST_FLAG(also_run_disabled_tests)) ||
+        ParseBoolFlag(arg, kBreakOnFailureFlag,
+                      >EST_FLAG(break_on_failure)) ||
+        ParseBoolFlag(arg, kCatchExceptionsFlag,
+                      >EST_FLAG(catch_exceptions)) ||
+        ParseStringFlag(arg, kColorFlag, >EST_FLAG(color)) ||
+        ParseStringFlag(arg, kDeathTestStyleFlag,
+                        >EST_FLAG(death_test_style)) ||
+        ParseBoolFlag(arg, kDeathTestUseFork,
+                      >EST_FLAG(death_test_use_fork)) ||
+        ParseStringFlag(arg, kFilterFlag, >EST_FLAG(filter)) ||
+        ParseStringFlag(arg, kInternalRunDeathTestFlag,
+                        >EST_FLAG(internal_run_death_test)) ||
+        ParseBoolFlag(arg, kListTestsFlag, >EST_FLAG(list_tests)) ||
+        ParseStringFlag(arg, kOutputFlag, >EST_FLAG(output)) ||
+        ParseBoolFlag(arg, kPrintTimeFlag, >EST_FLAG(print_time)) ||
+        ParseInt32Flag(arg, kRandomSeedFlag, >EST_FLAG(random_seed)) ||
+        ParseInt32Flag(arg, kRepeatFlag, >EST_FLAG(repeat)) ||
+        ParseBoolFlag(arg, kShuffleFlag, >EST_FLAG(shuffle)) ||
+        ParseInt32Flag(arg, kStackTraceDepthFlag,
+                       >EST_FLAG(stack_trace_depth)) ||
+        ParseStringFlag(arg, kStreamResultToFlag,
+                        >EST_FLAG(stream_result_to)) ||
+        ParseBoolFlag(arg, kThrowOnFailureFlag,
+                      >EST_FLAG(throw_on_failure))
+        ) {
+      // Yes.  Shift the remainder of the argv list left by one.  Note
+      // that argv has (*argc + 1) elements, the last one always being
+      // NULL.  The following loop moves the trailing NULL element as
+      // well.
+      for (int j = i; j != *argc; j++) {
+        argv[j] = argv[j + 1];
+      }
+
+      // Decrements the argument count.
+      (*argc)--;
+
+      // We also need to decrement the iterator as we just removed
+      // an element.
+      i--;
+    } else if (arg_string == "--help" || arg_string == "-h" ||
+               arg_string == "-?" || arg_string == "/?" ||
+               HasGoogleTestFlagPrefix(arg)) {
+      // Both help flag and unrecognized Google Test flags (excluding
+      // internal ones) trigger help display.
+      g_help_flag = true;
+    }
+  }
+
+  if (g_help_flag) {
+    // We print the help here instead of in RUN_ALL_TESTS(), as the
+    // latter may not be called at all if the user is using Google
+    // Test with another testing framework.
+    PrintColorEncoded(kColorEncodedHelpMessage);
+  }
+}
+
+// Parses the command line for Google Test flags, without initializing
+// other parts of Google Test.
+void ParseGoogleTestFlagsOnly(int* argc, char** argv) {
+  ParseGoogleTestFlagsOnlyImpl(argc, argv);
+}
+void ParseGoogleTestFlagsOnly(int* argc, wchar_t** argv) {
+  ParseGoogleTestFlagsOnlyImpl(argc, argv);
+}
+
+// The internal implementation of InitGoogleTest().
+//
+// The type parameter CharType can be instantiated to either char or
+// wchar_t.
+template 
+void InitGoogleTestImpl(int* argc, CharType** argv) {
+  g_init_gtest_count++;
+
+  // We don't want to run the initialization code twice.
+  if (g_init_gtest_count != 1) return;
+
+  if (*argc <= 0) return;
+
+  internal::g_executable_path = internal::StreamableToString(argv[0]);
+
+#if GTEST_HAS_DEATH_TEST
+
+  g_argvs.clear();
+  for (int i = 0; i != *argc; i++) {
+    g_argvs.push_back(StreamableToString(argv[i]));
+  }
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+  ParseGoogleTestFlagsOnly(argc, argv);
+  GetUnitTestImpl()->PostFlagParsingInit();
+}
+
+}  // namespace internal
+
+// Initializes Google Test.  This must be called before calling
+// RUN_ALL_TESTS().  In particular, it parses a command line for the
+// flags that Google Test recognizes.  Whenever a Google Test flag is
+// seen, it is removed from argv, and *argc is decremented.
+//
+// No value is returned.  Instead, the Google Test flag variables are
+// updated.
+//
+// Calling the function for the second time has no user-visible effect.
+void InitGoogleTest(int* argc, char** argv) {
+  internal::InitGoogleTestImpl(argc, argv);
+}
+
+// This overloaded version can be used in Windows programs compiled in
+// UNICODE mode.
+void InitGoogleTest(int* argc, wchar_t** argv) {
+  internal::InitGoogleTestImpl(argc, argv);
+}
+
+}  // namespace testing
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan), vladl@google.com (Vlad Losev)
+//
+// This file implements death tests.
+
+
+#if GTEST_HAS_DEATH_TEST
+
+# if GTEST_OS_MAC
+#  include 
+# endif  // GTEST_OS_MAC
+
+# include 
+# include 
+# include 
+
+# if GTEST_OS_LINUX
+#  include 
+# endif  // GTEST_OS_LINUX
+
+# include 
+
+# if GTEST_OS_WINDOWS
+#  include 
+# else
+#  include 
+#  include 
+# endif  // GTEST_OS_WINDOWS
+
+# if GTEST_OS_QNX
+#  include 
+# endif  // GTEST_OS_QNX
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick is to
+// prevent a user from accidentally including gtest-internal-inl.h in
+// his code.
+#define GTEST_IMPLEMENTATION_ 1
+#undef GTEST_IMPLEMENTATION_
+
+namespace testing {
+
+// Constants.
+
+// The default death test style.
+static const char kDefaultDeathTestStyle[] = "fast";
+
+GTEST_DEFINE_string_(
+    death_test_style,
+    internal::StringFromGTestEnv("death_test_style", kDefaultDeathTestStyle),
+    "Indicates how to run a death test in a forked child process: "
+    "\"threadsafe\" (child process re-executes the test binary "
+    "from the beginning, running only the specific death test) or "
+    "\"fast\" (child process runs the death test immediately "
+    "after forking).");
+
+GTEST_DEFINE_bool_(
+    death_test_use_fork,
+    internal::BoolFromGTestEnv("death_test_use_fork", false),
+    "Instructs to use fork()/_exit() instead of clone() in death tests. "
+    "Ignored and always uses fork() on POSIX systems where clone() is not "
+    "implemented. Useful when running under valgrind or similar tools if "
+    "those do not support clone(). Valgrind 3.3.1 will just fail if "
+    "it sees an unsupported combination of clone() flags. "
+    "It is not recommended to use this flag w/o valgrind though it will "
+    "work in 99% of the cases. Once valgrind is fixed, this flag will "
+    "most likely be removed.");
+
+namespace internal {
+GTEST_DEFINE_string_(
+    internal_run_death_test, "",
+    "Indicates the file, line number, temporal index of "
+    "the single death test to run, and a file descriptor to "
+    "which a success code may be sent, all separated by "
+    "the '|' characters.  This flag is specified if and only if the current "
+    "process is a sub-process launched for running a thread-safe "
+    "death test.  FOR INTERNAL USE ONLY.");
+}  // namespace internal
+
+#if GTEST_HAS_DEATH_TEST
+
+namespace internal {
+
+// Valid only for fast death tests. Indicates the code is running in the
+// child process of a fast style death test.
+static bool g_in_fast_death_test_child = false;
+
+// Returns a Boolean value indicating whether the caller is currently
+// executing in the context of the death test child process.  Tools such as
+// Valgrind heap checkers may need this to modify their behavior in death
+// tests.  IMPORTANT: This is an internal utility.  Using it may break the
+// implementation of death tests.  User code MUST NOT use it.
+bool InDeathTestChild() {
+# if GTEST_OS_WINDOWS
+
+  // On Windows, death tests are thread-safe regardless of the value of the
+  // death_test_style flag.
+  return !GTEST_FLAG(internal_run_death_test).empty();
+
+# else
+
+  if (GTEST_FLAG(death_test_style) == "threadsafe")
+    return !GTEST_FLAG(internal_run_death_test).empty();
+  else
+    return g_in_fast_death_test_child;
+#endif
+}
+
+}  // namespace internal
+
+// ExitedWithCode constructor.
+ExitedWithCode::ExitedWithCode(int exit_code) : exit_code_(exit_code) {
+}
+
+// ExitedWithCode function-call operator.
+bool ExitedWithCode::operator()(int exit_status) const {
+# if GTEST_OS_WINDOWS
+
+  return exit_status == exit_code_;
+
+# else
+
+  return WIFEXITED(exit_status) && WEXITSTATUS(exit_status) == exit_code_;
+
+# endif  // GTEST_OS_WINDOWS
+}
+
+# if !GTEST_OS_WINDOWS
+// KilledBySignal constructor.
+KilledBySignal::KilledBySignal(int signum) : signum_(signum) {
+}
+
+// KilledBySignal function-call operator.
+bool KilledBySignal::operator()(int exit_status) const {
+  return WIFSIGNALED(exit_status) && WTERMSIG(exit_status) == signum_;
+}
+# endif  // !GTEST_OS_WINDOWS
+
+namespace internal {
+
+// Utilities needed for death tests.
+
+// Generates a textual description of a given exit code, in the format
+// specified by wait(2).
+static std::string ExitSummary(int exit_code) {
+  Message m;
+
+# if GTEST_OS_WINDOWS
+
+  m << "Exited with exit status " << exit_code;
+
+# else
+
+  if (WIFEXITED(exit_code)) {
+    m << "Exited with exit status " << WEXITSTATUS(exit_code);
+  } else if (WIFSIGNALED(exit_code)) {
+    m << "Terminated by signal " << WTERMSIG(exit_code);
+  }
+#  ifdef WCOREDUMP
+  if (WCOREDUMP(exit_code)) {
+    m << " (core dumped)";
+  }
+#  endif
+# endif  // GTEST_OS_WINDOWS
+
+  return m.GetString();
+}
+
+// Returns true if exit_status describes a process that was terminated
+// by a signal, or exited normally with a nonzero exit code.
+bool ExitedUnsuccessfully(int exit_status) {
+  return !ExitedWithCode(0)(exit_status);
+}
+
+# if !GTEST_OS_WINDOWS
+// Generates a textual failure message when a death test finds more than
+// one thread running, or cannot determine the number of threads, prior
+// to executing the given statement.  It is the responsibility of the
+// caller not to pass a thread_count of 1.
+static std::string DeathTestThreadWarning(size_t thread_count) {
+  Message msg;
+  msg << "Death tests use fork(), which is unsafe particularly"
+      << " in a threaded context. For this test, " << GTEST_NAME_ << " ";
+  if (thread_count == 0)
+    msg << "couldn't detect the number of threads.";
+  else
+    msg << "detected " << thread_count << " threads.";
+  return msg.GetString();
+}
+# endif  // !GTEST_OS_WINDOWS
+
+// Flag characters for reporting a death test that did not die.
+static const char kDeathTestLived = 'L';
+static const char kDeathTestReturned = 'R';
+static const char kDeathTestThrew = 'T';
+static const char kDeathTestInternalError = 'I';
+
+// An enumeration describing all of the possible ways that a death test can
+// conclude.  DIED means that the process died while executing the test
+// code; LIVED means that process lived beyond the end of the test code;
+// RETURNED means that the test statement attempted to execute a return
+// statement, which is not allowed; THREW means that the test statement
+// returned control by throwing an exception.  IN_PROGRESS means the test
+// has not yet concluded.
+// TODO(vladl@google.com): Unify names and possibly values for
+// AbortReason, DeathTestOutcome, and flag characters above.
+enum DeathTestOutcome { IN_PROGRESS, DIED, LIVED, RETURNED, THREW };
+
+// Routine for aborting the program which is safe to call from an
+// exec-style death test child process, in which case the error
+// message is propagated back to the parent process.  Otherwise, the
+// message is simply printed to stderr.  In either case, the program
+// then exits with status 1.
+void DeathTestAbort(const std::string& message) {
+  // On a POSIX system, this function may be called from a threadsafe-style
+  // death test child process, which operates on a very small stack.  Use
+  // the heap for any additional non-minuscule memory requirements.
+  const InternalRunDeathTestFlag* const flag =
+      GetUnitTestImpl()->internal_run_death_test_flag();
+  if (flag != NULL) {
+    FILE* parent = posix::FDOpen(flag->write_fd(), "w");
+    fputc(kDeathTestInternalError, parent);
+    fprintf(parent, "%s", message.c_str());
+    fflush(parent);
+    _exit(1);
+  } else {
+    fprintf(stderr, "%s", message.c_str());
+    fflush(stderr);
+    posix::Abort();
+  }
+}
+
+// A replacement for CHECK that calls DeathTestAbort if the assertion
+// fails.
+# define GTEST_DEATH_TEST_CHECK_(expression) \
+  do { \
+    if (!::testing::internal::IsTrue(expression)) { \
+      DeathTestAbort( \
+          ::std::string("CHECK failed: File ") + __FILE__ +  ", line " \
+          + ::testing::internal::StreamableToString(__LINE__) + ": " \
+          + #expression); \
+    } \
+  } while (::testing::internal::AlwaysFalse())
+
+// This macro is similar to GTEST_DEATH_TEST_CHECK_, but it is meant for
+// evaluating any system call that fulfills two conditions: it must return
+// -1 on failure, and set errno to EINTR when it is interrupted and
+// should be tried again.  The macro expands to a loop that repeatedly
+// evaluates the expression as long as it evaluates to -1 and sets
+// errno to EINTR.  If the expression evaluates to -1 but errno is
+// something other than EINTR, DeathTestAbort is called.
+# define GTEST_DEATH_TEST_CHECK_SYSCALL_(expression) \
+  do { \
+    int gtest_retval; \
+    do { \
+      gtest_retval = (expression); \
+    } while (gtest_retval == -1 && errno == EINTR); \
+    if (gtest_retval == -1) { \
+      DeathTestAbort( \
+          ::std::string("CHECK failed: File ") + __FILE__ + ", line " \
+          + ::testing::internal::StreamableToString(__LINE__) + ": " \
+          + #expression + " != -1"); \
+    } \
+  } while (::testing::internal::AlwaysFalse())
+
+// Returns the message describing the last system error in errno.
+std::string GetLastErrnoDescription() {
+    return errno == 0 ? "" : posix::StrError(errno);
+}
+
+// This is called from a death test parent process to read a failure
+// message from the death test child process and log it with the FATAL
+// severity. On Windows, the message is read from a pipe handle. On other
+// platforms, it is read from a file descriptor.
+static void FailFromInternalError(int fd) {
+  Message error;
+  char buffer[256];
+  int num_read;
+
+  do {
+    while ((num_read = posix::Read(fd, buffer, 255)) > 0) {
+      buffer[num_read] = '\0';
+      error << buffer;
+    }
+  } while (num_read == -1 && errno == EINTR);
+
+  if (num_read == 0) {
+    GTEST_LOG_(FATAL) << error.GetString();
+  } else {
+    const int last_error = errno;
+    GTEST_LOG_(FATAL) << "Error while reading death test internal: "
+                      << GetLastErrnoDescription() << " [" << last_error << "]";
+  }
+}
+
+// Death test constructor.  Increments the running death test count
+// for the current test.
+DeathTest::DeathTest() {
+  TestInfo* const info = GetUnitTestImpl()->current_test_info();
+  if (info == NULL) {
+    DeathTestAbort("Cannot run a death test outside of a TEST or "
+                   "TEST_F construct");
+  }
+}
+
+// Creates and returns a death test by dispatching to the current
+// death test factory.
+bool DeathTest::Create(const char* statement, const RE* regex,
+                       const char* file, int line, DeathTest** test) {
+  return GetUnitTestImpl()->death_test_factory()->Create(
+      statement, regex, file, line, test);
+}
+
+const char* DeathTest::LastMessage() {
+  return last_death_test_message_.c_str();
+}
+
+void DeathTest::set_last_death_test_message(const std::string& message) {
+  last_death_test_message_ = message;
+}
+
+std::string DeathTest::last_death_test_message_;
+
+// Provides cross platform implementation for some death functionality.
+class DeathTestImpl : public DeathTest {
+ protected:
+  DeathTestImpl(const char* a_statement, const RE* a_regex)
+      : statement_(a_statement),
+        regex_(a_regex),
+        spawned_(false),
+        status_(-1),
+        outcome_(IN_PROGRESS),
+        read_fd_(-1),
+        write_fd_(-1) {}
+
+  // read_fd_ is expected to be closed and cleared by a derived class.
+  ~DeathTestImpl() { GTEST_DEATH_TEST_CHECK_(read_fd_ == -1); }
+
+  void Abort(AbortReason reason);
+  virtual bool Passed(bool status_ok);
+
+  const char* statement() const { return statement_; }
+  const RE* regex() const { return regex_; }
+  bool spawned() const { return spawned_; }
+  void set_spawned(bool is_spawned) { spawned_ = is_spawned; }
+  int status() const { return status_; }
+  void set_status(int a_status) { status_ = a_status; }
+  DeathTestOutcome outcome() const { return outcome_; }
+  void set_outcome(DeathTestOutcome an_outcome) { outcome_ = an_outcome; }
+  int read_fd() const { return read_fd_; }
+  void set_read_fd(int fd) { read_fd_ = fd; }
+  int write_fd() const { return write_fd_; }
+  void set_write_fd(int fd) { write_fd_ = fd; }
+
+  // Called in the parent process only. Reads the result code of the death
+  // test child process via a pipe, interprets it to set the outcome_
+  // member, and closes read_fd_.  Outputs diagnostics and terminates in
+  // case of unexpected codes.
+  void ReadAndInterpretStatusByte();
+
+ private:
+  // The textual content of the code this object is testing.  This class
+  // doesn't own this string and should not attempt to delete it.
+  const char* const statement_;
+  // The regular expression which test output must match.  DeathTestImpl
+  // doesn't own this object and should not attempt to delete it.
+  const RE* const regex_;
+  // True if the death test child process has been successfully spawned.
+  bool spawned_;
+  // The exit status of the child process.
+  int status_;
+  // How the death test concluded.
+  DeathTestOutcome outcome_;
+  // Descriptor to the read end of the pipe to the child process.  It is
+  // always -1 in the child process.  The child keeps its write end of the
+  // pipe in write_fd_.
+  int read_fd_;
+  // Descriptor to the child's write end of the pipe to the parent process.
+  // It is always -1 in the parent process.  The parent keeps its end of the
+  // pipe in read_fd_.
+  int write_fd_;
+};
+
+// Called in the parent process only. Reads the result code of the death
+// test child process via a pipe, interprets it to set the outcome_
+// member, and closes read_fd_.  Outputs diagnostics and terminates in
+// case of unexpected codes.
+void DeathTestImpl::ReadAndInterpretStatusByte() {
+  char flag;
+  int bytes_read;
+
+  // The read() here blocks until data is available (signifying the
+  // failure of the death test) or until the pipe is closed (signifying
+  // its success), so it's okay to call this in the parent before
+  // the child process has exited.
+  do {
+    bytes_read = posix::Read(read_fd(), &flag, 1);
+  } while (bytes_read == -1 && errno == EINTR);
+
+  if (bytes_read == 0) {
+    set_outcome(DIED);
+  } else if (bytes_read == 1) {
+    switch (flag) {
+      case kDeathTestReturned:
+        set_outcome(RETURNED);
+        break;
+      case kDeathTestThrew:
+        set_outcome(THREW);
+        break;
+      case kDeathTestLived:
+        set_outcome(LIVED);
+        break;
+      case kDeathTestInternalError:
+        FailFromInternalError(read_fd());  // Does not return.
+        break;
+      default:
+        GTEST_LOG_(FATAL) << "Death test child process reported "
+                          << "unexpected status byte ("
+                          << static_cast(flag) << ")";
+    }
+  } else {
+    GTEST_LOG_(FATAL) << "Read from death test child process failed: "
+                      << GetLastErrnoDescription();
+  }
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(posix::Close(read_fd()));
+  set_read_fd(-1);
+}
+
+// Signals that the death test code which should have exited, didn't.
+// Should be called only in a death test child process.
+// Writes a status byte to the child's status file descriptor, then
+// calls _exit(1).
+void DeathTestImpl::Abort(AbortReason reason) {
+  // The parent process considers the death test to be a failure if
+  // it finds any data in our pipe.  So, here we write a single flag byte
+  // to the pipe, then exit.
+  const char status_ch =
+      reason == TEST_DID_NOT_DIE ? kDeathTestLived :
+      reason == TEST_THREW_EXCEPTION ? kDeathTestThrew : kDeathTestReturned;
+
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(posix::Write(write_fd(), &status_ch, 1));
+  // We are leaking the descriptor here because on some platforms (i.e.,
+  // when built as Windows DLL), destructors of global objects will still
+  // run after calling _exit(). On such systems, write_fd_ will be
+  // indirectly closed from the destructor of UnitTestImpl, causing double
+  // close if it is also closed here. On debug configurations, double close
+  // may assert. As there are no in-process buffers to flush here, we are
+  // relying on the OS to close the descriptor after the process terminates
+  // when the destructors are not run.
+  _exit(1);  // Exits w/o any normal exit hooks (we were supposed to crash)
+}
+
+// Returns an indented copy of stderr output for a death test.
+// This makes distinguishing death test output lines from regular log lines
+// much easier.
+static ::std::string FormatDeathTestOutput(const ::std::string& output) {
+  ::std::string ret;
+  for (size_t at = 0; ; ) {
+    const size_t line_end = output.find('\n', at);
+    ret += "[  DEATH   ] ";
+    if (line_end == ::std::string::npos) {
+      ret += output.substr(at);
+      break;
+    }
+    ret += output.substr(at, line_end + 1 - at);
+    at = line_end + 1;
+  }
+  return ret;
+}
+
+// Assesses the success or failure of a death test, using both private
+// members which have previously been set, and one argument:
+//
+// Private data members:
+//   outcome:  An enumeration describing how the death test
+//             concluded: DIED, LIVED, THREW, or RETURNED.  The death test
+//             fails in the latter three cases.
+//   status:   The exit status of the child process. On *nix, it is in the
+//             in the format specified by wait(2). On Windows, this is the
+//             value supplied to the ExitProcess() API or a numeric code
+//             of the exception that terminated the program.
+//   regex:    A regular expression object to be applied to
+//             the test's captured standard error output; the death test
+//             fails if it does not match.
+//
+// Argument:
+//   status_ok: true if exit_status is acceptable in the context of
+//              this particular death test, which fails if it is false
+//
+// Returns true iff all of the above conditions are met.  Otherwise, the
+// first failing condition, in the order given above, is the one that is
+// reported. Also sets the last death test message string.
+bool DeathTestImpl::Passed(bool status_ok) {
+  if (!spawned())
+    return false;
+
+  const std::string error_message = GetCapturedStderr();
+
+  bool success = false;
+  Message buffer;
+
+  buffer << "Death test: " << statement() << "\n";
+  switch (outcome()) {
+    case LIVED:
+      buffer << "    Result: failed to die.\n"
+             << " Error msg:\n" << FormatDeathTestOutput(error_message);
+      break;
+    case THREW:
+      buffer << "    Result: threw an exception.\n"
+             << " Error msg:\n" << FormatDeathTestOutput(error_message);
+      break;
+    case RETURNED:
+      buffer << "    Result: illegal return in test statement.\n"
+             << " Error msg:\n" << FormatDeathTestOutput(error_message);
+      break;
+    case DIED:
+      if (status_ok) {
+        const bool matched = RE::PartialMatch(error_message.c_str(), *regex());
+        if (matched) {
+          success = true;
+        } else {
+          buffer << "    Result: died but not with expected error.\n"
+                 << "  Expected: " << regex()->pattern() << "\n"
+                 << "Actual msg:\n" << FormatDeathTestOutput(error_message);
+        }
+      } else {
+        buffer << "    Result: died but not with expected exit code:\n"
+               << "            " << ExitSummary(status()) << "\n"
+               << "Actual msg:\n" << FormatDeathTestOutput(error_message);
+      }
+      break;
+    case IN_PROGRESS:
+    default:
+      GTEST_LOG_(FATAL)
+          << "DeathTest::Passed somehow called before conclusion of test";
+  }
+
+  DeathTest::set_last_death_test_message(buffer.GetString());
+  return success;
+}
+
+# if GTEST_OS_WINDOWS
+// WindowsDeathTest implements death tests on Windows. Due to the
+// specifics of starting new processes on Windows, death tests there are
+// always threadsafe, and Google Test considers the
+// --gtest_death_test_style=fast setting to be equivalent to
+// --gtest_death_test_style=threadsafe there.
+//
+// A few implementation notes:  Like the Linux version, the Windows
+// implementation uses pipes for child-to-parent communication. But due to
+// the specifics of pipes on Windows, some extra steps are required:
+//
+// 1. The parent creates a communication pipe and stores handles to both
+//    ends of it.
+// 2. The parent starts the child and provides it with the information
+//    necessary to acquire the handle to the write end of the pipe.
+// 3. The child acquires the write end of the pipe and signals the parent
+//    using a Windows event.
+// 4. Now the parent can release the write end of the pipe on its side. If
+//    this is done before step 3, the object's reference count goes down to
+//    0 and it is destroyed, preventing the child from acquiring it. The
+//    parent now has to release it, or read operations on the read end of
+//    the pipe will not return when the child terminates.
+// 5. The parent reads child's output through the pipe (outcome code and
+//    any possible error messages) from the pipe, and its stderr and then
+//    determines whether to fail the test.
+//
+// Note: to distinguish Win32 API calls from the local method and function
+// calls, the former are explicitly resolved in the global namespace.
+//
+class WindowsDeathTest : public DeathTestImpl {
+ public:
+  WindowsDeathTest(const char* a_statement,
+                   const RE* a_regex,
+                   const char* file,
+                   int line)
+      : DeathTestImpl(a_statement, a_regex), file_(file), line_(line) {}
+
+  // All of these virtual functions are inherited from DeathTest.
+  virtual int Wait();
+  virtual TestRole AssumeRole();
+
+ private:
+  // The name of the file in which the death test is located.
+  const char* const file_;
+  // The line number on which the death test is located.
+  const int line_;
+  // Handle to the write end of the pipe to the child process.
+  AutoHandle write_handle_;
+  // Child process handle.
+  AutoHandle child_handle_;
+  // Event the child process uses to signal the parent that it has
+  // acquired the handle to the write end of the pipe. After seeing this
+  // event the parent can release its own handles to make sure its
+  // ReadFile() calls return when the child terminates.
+  AutoHandle event_handle_;
+};
+
+// Waits for the child in a death test to exit, returning its exit
+// status, or 0 if no child process exists.  As a side effect, sets the
+// outcome data member.
+int WindowsDeathTest::Wait() {
+  if (!spawned())
+    return 0;
+
+  // Wait until the child either signals that it has acquired the write end
+  // of the pipe or it dies.
+  const HANDLE wait_handles[2] = { child_handle_.Get(), event_handle_.Get() };
+  switch (::WaitForMultipleObjects(2,
+                                   wait_handles,
+                                   FALSE,  // Waits for any of the handles.
+                                   INFINITE)) {
+    case WAIT_OBJECT_0:
+    case WAIT_OBJECT_0 + 1:
+      break;
+    default:
+      GTEST_DEATH_TEST_CHECK_(false);  // Should not get here.
+  }
+
+  // The child has acquired the write end of the pipe or exited.
+  // We release the handle on our side and continue.
+  write_handle_.Reset();
+  event_handle_.Reset();
+
+  ReadAndInterpretStatusByte();
+
+  // Waits for the child process to exit if it haven't already. This
+  // returns immediately if the child has already exited, regardless of
+  // whether previous calls to WaitForMultipleObjects synchronized on this
+  // handle or not.
+  GTEST_DEATH_TEST_CHECK_(
+      WAIT_OBJECT_0 == ::WaitForSingleObject(child_handle_.Get(),
+                                             INFINITE));
+  DWORD status_code;
+  GTEST_DEATH_TEST_CHECK_(
+      ::GetExitCodeProcess(child_handle_.Get(), &status_code) != FALSE);
+  child_handle_.Reset();
+  set_status(static_cast(status_code));
+  return status();
+}
+
+// The AssumeRole process for a Windows death test.  It creates a child
+// process with the same executable as the current process to run the
+// death test.  The child process is given the --gtest_filter and
+// --gtest_internal_run_death_test flags such that it knows to run the
+// current death test only.
+DeathTest::TestRole WindowsDeathTest::AssumeRole() {
+  const UnitTestImpl* const impl = GetUnitTestImpl();
+  const InternalRunDeathTestFlag* const flag =
+      impl->internal_run_death_test_flag();
+  const TestInfo* const info = impl->current_test_info();
+  const int death_test_index = info->result()->death_test_count();
+
+  if (flag != NULL) {
+    // ParseInternalRunDeathTestFlag() has performed all the necessary
+    // processing.
+    set_write_fd(flag->write_fd());
+    return EXECUTE_TEST;
+  }
+
+  // WindowsDeathTest uses an anonymous pipe to communicate results of
+  // a death test.
+  SECURITY_ATTRIBUTES handles_are_inheritable = {
+    sizeof(SECURITY_ATTRIBUTES), NULL, TRUE };
+  HANDLE read_handle, write_handle;
+  GTEST_DEATH_TEST_CHECK_(
+      ::CreatePipe(&read_handle, &write_handle, &handles_are_inheritable,
+                   0)  // Default buffer size.
+      != FALSE);
+  set_read_fd(::_open_osfhandle(reinterpret_cast(read_handle),
+                                O_RDONLY));
+  write_handle_.Reset(write_handle);
+  event_handle_.Reset(::CreateEvent(
+      &handles_are_inheritable,
+      TRUE,    // The event will automatically reset to non-signaled state.
+      FALSE,   // The initial state is non-signalled.
+      NULL));  // The even is unnamed.
+  GTEST_DEATH_TEST_CHECK_(event_handle_.Get() != NULL);
+  const std::string filter_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kFilterFlag + "=" +
+      info->test_case_name() + "." + info->name();
+  const std::string internal_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kInternalRunDeathTestFlag +
+      "=" + file_ + "|" + StreamableToString(line_) + "|" +
+      StreamableToString(death_test_index) + "|" +
+      StreamableToString(static_cast(::GetCurrentProcessId())) +
+      // size_t has the same width as pointers on both 32-bit and 64-bit
+      // Windows platforms.
+      // See http://msdn.microsoft.com/en-us/library/tcxf1dw6.aspx.
+      "|" + StreamableToString(reinterpret_cast(write_handle)) +
+      "|" + StreamableToString(reinterpret_cast(event_handle_.Get()));
+
+  char executable_path[_MAX_PATH + 1];  // NOLINT
+  GTEST_DEATH_TEST_CHECK_(
+      _MAX_PATH + 1 != ::GetModuleFileNameA(NULL,
+                                            executable_path,
+                                            _MAX_PATH));
+
+  std::string command_line =
+      std::string(::GetCommandLineA()) + " " + filter_flag + " \"" +
+      internal_flag + "\"";
+
+  DeathTest::set_last_death_test_message("");
+
+  CaptureStderr();
+  // Flush the log buffers since the log streams are shared with the child.
+  FlushInfoLog();
+
+  // The child process will share the standard handles with the parent.
+  STARTUPINFOA startup_info;
+  memset(&startup_info, 0, sizeof(STARTUPINFO));
+  startup_info.dwFlags = STARTF_USESTDHANDLES;
+  startup_info.hStdInput = ::GetStdHandle(STD_INPUT_HANDLE);
+  startup_info.hStdOutput = ::GetStdHandle(STD_OUTPUT_HANDLE);
+  startup_info.hStdError = ::GetStdHandle(STD_ERROR_HANDLE);
+
+  PROCESS_INFORMATION process_info;
+  GTEST_DEATH_TEST_CHECK_(::CreateProcessA(
+      executable_path,
+      const_cast(command_line.c_str()),
+      NULL,   // Retuned process handle is not inheritable.
+      NULL,   // Retuned thread handle is not inheritable.
+      TRUE,   // Child inherits all inheritable handles (for write_handle_).
+      0x0,    // Default creation flags.
+      NULL,   // Inherit the parent's environment.
+      UnitTest::GetInstance()->original_working_dir(),
+      &startup_info,
+      &process_info) != FALSE);
+  child_handle_.Reset(process_info.hProcess);
+  ::CloseHandle(process_info.hThread);
+  set_spawned(true);
+  return OVERSEE_TEST;
+}
+# else  // We are not on Windows.
+
+// ForkingDeathTest provides implementations for most of the abstract
+// methods of the DeathTest interface.  Only the AssumeRole method is
+// left undefined.
+class ForkingDeathTest : public DeathTestImpl {
+ public:
+  ForkingDeathTest(const char* statement, const RE* regex);
+
+  // All of these virtual functions are inherited from DeathTest.
+  virtual int Wait();
+
+ protected:
+  void set_child_pid(pid_t child_pid) { child_pid_ = child_pid; }
+
+ private:
+  // PID of child process during death test; 0 in the child process itself.
+  pid_t child_pid_;
+};
+
+// Constructs a ForkingDeathTest.
+ForkingDeathTest::ForkingDeathTest(const char* a_statement, const RE* a_regex)
+    : DeathTestImpl(a_statement, a_regex),
+      child_pid_(-1) {}
+
+// Waits for the child in a death test to exit, returning its exit
+// status, or 0 if no child process exists.  As a side effect, sets the
+// outcome data member.
+int ForkingDeathTest::Wait() {
+  if (!spawned())
+    return 0;
+
+  ReadAndInterpretStatusByte();
+
+  int status_value;
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(waitpid(child_pid_, &status_value, 0));
+  set_status(status_value);
+  return status_value;
+}
+
+// A concrete death test class that forks, then immediately runs the test
+// in the child process.
+class NoExecDeathTest : public ForkingDeathTest {
+ public:
+  NoExecDeathTest(const char* a_statement, const RE* a_regex) :
+      ForkingDeathTest(a_statement, a_regex) { }
+  virtual TestRole AssumeRole();
+};
+
+// The AssumeRole process for a fork-and-run death test.  It implements a
+// straightforward fork, with a simple pipe to transmit the status byte.
+DeathTest::TestRole NoExecDeathTest::AssumeRole() {
+  const size_t thread_count = GetThreadCount();
+  if (thread_count != 1) {
+    GTEST_LOG_(WARNING) << DeathTestThreadWarning(thread_count);
+  }
+
+  int pipe_fd[2];
+  GTEST_DEATH_TEST_CHECK_(pipe(pipe_fd) != -1);
+
+  DeathTest::set_last_death_test_message("");
+  CaptureStderr();
+  // When we fork the process below, the log file buffers are copied, but the
+  // file descriptors are shared.  We flush all log files here so that closing
+  // the file descriptors in the child process doesn't throw off the
+  // synchronization between descriptors and buffers in the parent process.
+  // This is as close to the fork as possible to avoid a race condition in case
+  // there are multiple threads running before the death test, and another
+  // thread writes to the log file.
+  FlushInfoLog();
+
+  const pid_t child_pid = fork();
+  GTEST_DEATH_TEST_CHECK_(child_pid != -1);
+  set_child_pid(child_pid);
+  if (child_pid == 0) {
+    GTEST_DEATH_TEST_CHECK_SYSCALL_(close(pipe_fd[0]));
+    set_write_fd(pipe_fd[1]);
+    // Redirects all logging to stderr in the child process to prevent
+    // concurrent writes to the log files.  We capture stderr in the parent
+    // process and append the child process' output to a log.
+    LogToStderr();
+    // Event forwarding to the listeners of event listener API mush be shut
+    // down in death test subprocesses.
+    GetUnitTestImpl()->listeners()->SuppressEventForwarding();
+    g_in_fast_death_test_child = true;
+    return EXECUTE_TEST;
+  } else {
+    GTEST_DEATH_TEST_CHECK_SYSCALL_(close(pipe_fd[1]));
+    set_read_fd(pipe_fd[0]);
+    set_spawned(true);
+    return OVERSEE_TEST;
+  }
+}
+
+// A concrete death test class that forks and re-executes the main
+// program from the beginning, with command-line flags set that cause
+// only this specific death test to be run.
+class ExecDeathTest : public ForkingDeathTest {
+ public:
+  ExecDeathTest(const char* a_statement, const RE* a_regex,
+                const char* file, int line) :
+      ForkingDeathTest(a_statement, a_regex), file_(file), line_(line) { }
+  virtual TestRole AssumeRole();
+ private:
+  static ::std::vector
+  GetArgvsForDeathTestChildProcess() {
+    ::std::vector args = GetInjectableArgvs();
+    return args;
+  }
+  // The name of the file in which the death test is located.
+  const char* const file_;
+  // The line number on which the death test is located.
+  const int line_;
+};
+
+// Utility class for accumulating command-line arguments.
+class Arguments {
+ public:
+  Arguments() {
+    args_.push_back(NULL);
+  }
+
+  ~Arguments() {
+    for (std::vector::iterator i = args_.begin(); i != args_.end();
+         ++i) {
+      free(*i);
+    }
+  }
+  void AddArgument(const char* argument) {
+    args_.insert(args_.end() - 1, posix::StrDup(argument));
+  }
+
+  template 
+  void AddArguments(const ::std::vector& arguments) {
+    for (typename ::std::vector::const_iterator i = arguments.begin();
+         i != arguments.end();
+         ++i) {
+      args_.insert(args_.end() - 1, posix::StrDup(i->c_str()));
+    }
+  }
+  char* const* Argv() {
+    return &args_[0];
+  }
+
+ private:
+  std::vector args_;
+};
+
+// A struct that encompasses the arguments to the child process of a
+// threadsafe-style death test process.
+struct ExecDeathTestArgs {
+  char* const* argv;  // Command-line arguments for the child's call to exec
+  int close_fd;       // File descriptor to close; the read end of a pipe
+};
+
+#  if GTEST_OS_MAC
+inline char** GetEnviron() {
+  // When Google Test is built as a framework on MacOS X, the environ variable
+  // is unavailable. Apple's documentation (man environ) recommends using
+  // _NSGetEnviron() instead.
+  return *_NSGetEnviron();
+}
+#  else
+// Some POSIX platforms expect you to declare environ. extern "C" makes
+// it reside in the global namespace.
+extern "C" char** environ;
+inline char** GetEnviron() { return environ; }
+#  endif  // GTEST_OS_MAC
+
+#  if !GTEST_OS_QNX
+// The main function for a threadsafe-style death test child process.
+// This function is called in a clone()-ed process and thus must avoid
+// any potentially unsafe operations like malloc or libc functions.
+static int ExecDeathTestChildMain(void* child_arg) {
+  ExecDeathTestArgs* const args = static_cast(child_arg);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(close(args->close_fd));
+
+  // We need to execute the test program in the same environment where
+  // it was originally invoked.  Therefore we change to the original
+  // working directory first.
+  const char* const original_dir =
+      UnitTest::GetInstance()->original_working_dir();
+  // We can safely call chdir() as it's a direct system call.
+  if (chdir(original_dir) != 0) {
+    DeathTestAbort(std::string("chdir(\"") + original_dir + "\") failed: " +
+                   GetLastErrnoDescription());
+    return EXIT_FAILURE;
+  }
+
+  // We can safely call execve() as it's a direct system call.  We
+  // cannot use execvp() as it's a libc function and thus potentially
+  // unsafe.  Since execve() doesn't search the PATH, the user must
+  // invoke the test program via a valid path that contains at least
+  // one path separator.
+  execve(args->argv[0], args->argv, GetEnviron());
+  DeathTestAbort(std::string("execve(") + args->argv[0] + ", ...) in " +
+                 original_dir + " failed: " +
+                 GetLastErrnoDescription());
+  return EXIT_FAILURE;
+}
+#  endif  // !GTEST_OS_QNX
+
+// Two utility routines that together determine the direction the stack
+// grows.
+// This could be accomplished more elegantly by a single recursive
+// function, but we want to guard against the unlikely possibility of
+// a smart compiler optimizing the recursion away.
+//
+// GTEST_NO_INLINE_ is required to prevent GCC 4.6 from inlining
+// StackLowerThanAddress into StackGrowsDown, which then doesn't give
+// correct answer.
+void StackLowerThanAddress(const void* ptr, bool* result) GTEST_NO_INLINE_;
+void StackLowerThanAddress(const void* ptr, bool* result) {
+  int dummy;
+  *result = (&dummy < ptr);
+}
+
+bool StackGrowsDown() {
+  int dummy;
+  bool result;
+  StackLowerThanAddress(&dummy, &result);
+  return result;
+}
+
+// Spawns a child process with the same executable as the current process in
+// a thread-safe manner and instructs it to run the death test.  The
+// implementation uses fork(2) + exec.  On systems where clone(2) is
+// available, it is used instead, being slightly more thread-safe.  On QNX,
+// fork supports only single-threaded environments, so this function uses
+// spawn(2) there instead.  The function dies with an error message if
+// anything goes wrong.
+static pid_t ExecDeathTestSpawnChild(char* const* argv, int close_fd) {
+  ExecDeathTestArgs args = { argv, close_fd };
+  pid_t child_pid = -1;
+
+#  if GTEST_OS_QNX
+  // Obtains the current directory and sets it to be closed in the child
+  // process.
+  const int cwd_fd = open(".", O_RDONLY);
+  GTEST_DEATH_TEST_CHECK_(cwd_fd != -1);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(fcntl(cwd_fd, F_SETFD, FD_CLOEXEC));
+  // We need to execute the test program in the same environment where
+  // it was originally invoked.  Therefore we change to the original
+  // working directory first.
+  const char* const original_dir =
+      UnitTest::GetInstance()->original_working_dir();
+  // We can safely call chdir() as it's a direct system call.
+  if (chdir(original_dir) != 0) {
+    DeathTestAbort(std::string("chdir(\"") + original_dir + "\") failed: " +
+                   GetLastErrnoDescription());
+    return EXIT_FAILURE;
+  }
+
+  int fd_flags;
+  // Set close_fd to be closed after spawn.
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(fd_flags = fcntl(close_fd, F_GETFD));
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(fcntl(close_fd, F_SETFD,
+                                        fd_flags | FD_CLOEXEC));
+  struct inheritance inherit = {0};
+  // spawn is a system call.
+  child_pid = spawn(args.argv[0], 0, NULL, &inherit, args.argv, GetEnviron());
+  // Restores the current working directory.
+  GTEST_DEATH_TEST_CHECK_(fchdir(cwd_fd) != -1);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(close(cwd_fd));
+
+#  else   // GTEST_OS_QNX
+#   if GTEST_OS_LINUX
+  // When a SIGPROF signal is received while fork() or clone() are executing,
+  // the process may hang. To avoid this, we ignore SIGPROF here and re-enable
+  // it after the call to fork()/clone() is complete.
+  struct sigaction saved_sigprof_action;
+  struct sigaction ignore_sigprof_action;
+  memset(&ignore_sigprof_action, 0, sizeof(ignore_sigprof_action));
+  sigemptyset(&ignore_sigprof_action.sa_mask);
+  ignore_sigprof_action.sa_handler = SIG_IGN;
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(sigaction(
+      SIGPROF, &ignore_sigprof_action, &saved_sigprof_action));
+#   endif  // GTEST_OS_LINUX
+
+#   if GTEST_HAS_CLONE
+  const bool use_fork = GTEST_FLAG(death_test_use_fork);
+
+  if (!use_fork) {
+    static const bool stack_grows_down = StackGrowsDown();
+    const size_t stack_size = getpagesize();
+    // MMAP_ANONYMOUS is not defined on Mac, so we use MAP_ANON instead.
+    void* const stack = mmap(NULL, stack_size, PROT_READ | PROT_WRITE,
+                             MAP_ANON | MAP_PRIVATE, -1, 0);
+    GTEST_DEATH_TEST_CHECK_(stack != MAP_FAILED);
+
+    // Maximum stack alignment in bytes:  For a downward-growing stack, this
+    // amount is subtracted from size of the stack space to get an address
+    // that is within the stack space and is aligned on all systems we care
+    // about.  As far as I know there is no ABI with stack alignment greater
+    // than 64.  We assume stack and stack_size already have alignment of
+    // kMaxStackAlignment.
+    const size_t kMaxStackAlignment = 64;
+    void* const stack_top =
+        static_cast(stack) +
+            (stack_grows_down ? stack_size - kMaxStackAlignment : 0);
+    GTEST_DEATH_TEST_CHECK_(stack_size > kMaxStackAlignment &&
+        reinterpret_cast(stack_top) % kMaxStackAlignment == 0);
+
+    child_pid = clone(&ExecDeathTestChildMain, stack_top, SIGCHLD, &args);
+
+    GTEST_DEATH_TEST_CHECK_(munmap(stack, stack_size) != -1);
+  }
+#   else
+  const bool use_fork = true;
+#   endif  // GTEST_HAS_CLONE
+
+  if (use_fork && (child_pid = fork()) == 0) {
+      ExecDeathTestChildMain(&args);
+      _exit(0);
+  }
+#  endif  // GTEST_OS_QNX
+#  if GTEST_OS_LINUX
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(
+      sigaction(SIGPROF, &saved_sigprof_action, NULL));
+#  endif  // GTEST_OS_LINUX
+
+  GTEST_DEATH_TEST_CHECK_(child_pid != -1);
+  return child_pid;
+}
+
+// The AssumeRole process for a fork-and-exec death test.  It re-executes the
+// main program from the beginning, setting the --gtest_filter
+// and --gtest_internal_run_death_test flags to cause only the current
+// death test to be re-run.
+DeathTest::TestRole ExecDeathTest::AssumeRole() {
+  const UnitTestImpl* const impl = GetUnitTestImpl();
+  const InternalRunDeathTestFlag* const flag =
+      impl->internal_run_death_test_flag();
+  const TestInfo* const info = impl->current_test_info();
+  const int death_test_index = info->result()->death_test_count();
+
+  if (flag != NULL) {
+    set_write_fd(flag->write_fd());
+    return EXECUTE_TEST;
+  }
+
+  int pipe_fd[2];
+  GTEST_DEATH_TEST_CHECK_(pipe(pipe_fd) != -1);
+  // Clear the close-on-exec flag on the write end of the pipe, lest
+  // it be closed when the child process does an exec:
+  GTEST_DEATH_TEST_CHECK_(fcntl(pipe_fd[1], F_SETFD, 0) != -1);
+
+  const std::string filter_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kFilterFlag + "="
+      + info->test_case_name() + "." + info->name();
+  const std::string internal_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kInternalRunDeathTestFlag + "="
+      + file_ + "|" + StreamableToString(line_) + "|"
+      + StreamableToString(death_test_index) + "|"
+      + StreamableToString(pipe_fd[1]);
+  Arguments args;
+  args.AddArguments(GetArgvsForDeathTestChildProcess());
+  args.AddArgument(filter_flag.c_str());
+  args.AddArgument(internal_flag.c_str());
+
+  DeathTest::set_last_death_test_message("");
+
+  CaptureStderr();
+  // See the comment in NoExecDeathTest::AssumeRole for why the next line
+  // is necessary.
+  FlushInfoLog();
+
+  const pid_t child_pid = ExecDeathTestSpawnChild(args.Argv(), pipe_fd[0]);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(close(pipe_fd[1]));
+  set_child_pid(child_pid);
+  set_read_fd(pipe_fd[0]);
+  set_spawned(true);
+  return OVERSEE_TEST;
+}
+
+# endif  // !GTEST_OS_WINDOWS
+
+// Creates a concrete DeathTest-derived class that depends on the
+// --gtest_death_test_style flag, and sets the pointer pointed to
+// by the "test" argument to its address.  If the test should be
+// skipped, sets that pointer to NULL.  Returns true, unless the
+// flag is set to an invalid value.
+bool DefaultDeathTestFactory::Create(const char* statement, const RE* regex,
+                                     const char* file, int line,
+                                     DeathTest** test) {
+  UnitTestImpl* const impl = GetUnitTestImpl();
+  const InternalRunDeathTestFlag* const flag =
+      impl->internal_run_death_test_flag();
+  const int death_test_index = impl->current_test_info()
+      ->increment_death_test_count();
+
+  if (flag != NULL) {
+    if (death_test_index > flag->index()) {
+      DeathTest::set_last_death_test_message(
+          "Death test count (" + StreamableToString(death_test_index)
+          + ") somehow exceeded expected maximum ("
+          + StreamableToString(flag->index()) + ")");
+      return false;
+    }
+
+    if (!(flag->file() == file && flag->line() == line &&
+          flag->index() == death_test_index)) {
+      *test = NULL;
+      return true;
+    }
+  }
+
+# if GTEST_OS_WINDOWS
+
+  if (GTEST_FLAG(death_test_style) == "threadsafe" ||
+      GTEST_FLAG(death_test_style) == "fast") {
+    *test = new WindowsDeathTest(statement, regex, file, line);
+  }
+
+# else
+
+  if (GTEST_FLAG(death_test_style) == "threadsafe") {
+    *test = new ExecDeathTest(statement, regex, file, line);
+  } else if (GTEST_FLAG(death_test_style) == "fast") {
+    *test = new NoExecDeathTest(statement, regex);
+  }
+
+# endif  // GTEST_OS_WINDOWS
+
+  else {  // NOLINT - this is more readable than unbalanced brackets inside #if.
+    DeathTest::set_last_death_test_message(
+        "Unknown death test style \"" + GTEST_FLAG(death_test_style)
+        + "\" encountered");
+    return false;
+  }
+
+  return true;
+}
+
+// Splits a given string on a given delimiter, populating a given
+// vector with the fields.  GTEST_HAS_DEATH_TEST implies that we have
+// ::std::string, so we can use it here.
+static void SplitString(const ::std::string& str, char delimiter,
+                        ::std::vector< ::std::string>* dest) {
+  ::std::vector< ::std::string> parsed;
+  ::std::string::size_type pos = 0;
+  while (::testing::internal::AlwaysTrue()) {
+    const ::std::string::size_type colon = str.find(delimiter, pos);
+    if (colon == ::std::string::npos) {
+      parsed.push_back(str.substr(pos));
+      break;
+    } else {
+      parsed.push_back(str.substr(pos, colon - pos));
+      pos = colon + 1;
+    }
+  }
+  dest->swap(parsed);
+}
+
+# if GTEST_OS_WINDOWS
+// Recreates the pipe and event handles from the provided parameters,
+// signals the event, and returns a file descriptor wrapped around the pipe
+// handle. This function is called in the child process only.
+int GetStatusFileDescriptor(unsigned int parent_process_id,
+                            size_t write_handle_as_size_t,
+                            size_t event_handle_as_size_t) {
+  AutoHandle parent_process_handle(::OpenProcess(PROCESS_DUP_HANDLE,
+                                                   FALSE,  // Non-inheritable.
+                                                   parent_process_id));
+  if (parent_process_handle.Get() == INVALID_HANDLE_VALUE) {
+    DeathTestAbort("Unable to open parent process " +
+                   StreamableToString(parent_process_id));
+  }
+
+  // TODO(vladl@google.com): Replace the following check with a
+  // compile-time assertion when available.
+  GTEST_CHECK_(sizeof(HANDLE) <= sizeof(size_t));
+
+  const HANDLE write_handle =
+      reinterpret_cast(write_handle_as_size_t);
+  HANDLE dup_write_handle;
+
+  // The newly initialized handle is accessible only in in the parent
+  // process. To obtain one accessible within the child, we need to use
+  // DuplicateHandle.
+  if (!::DuplicateHandle(parent_process_handle.Get(), write_handle,
+                         ::GetCurrentProcess(), &dup_write_handle,
+                         0x0,    // Requested privileges ignored since
+                                 // DUPLICATE_SAME_ACCESS is used.
+                         FALSE,  // Request non-inheritable handler.
+                         DUPLICATE_SAME_ACCESS)) {
+    DeathTestAbort("Unable to duplicate the pipe handle " +
+                   StreamableToString(write_handle_as_size_t) +
+                   " from the parent process " +
+                   StreamableToString(parent_process_id));
+  }
+
+  const HANDLE event_handle = reinterpret_cast(event_handle_as_size_t);
+  HANDLE dup_event_handle;
+
+  if (!::DuplicateHandle(parent_process_handle.Get(), event_handle,
+                         ::GetCurrentProcess(), &dup_event_handle,
+                         0x0,
+                         FALSE,
+                         DUPLICATE_SAME_ACCESS)) {
+    DeathTestAbort("Unable to duplicate the event handle " +
+                   StreamableToString(event_handle_as_size_t) +
+                   " from the parent process " +
+                   StreamableToString(parent_process_id));
+  }
+
+  const int write_fd =
+      ::_open_osfhandle(reinterpret_cast(dup_write_handle), O_APPEND);
+  if (write_fd == -1) {
+    DeathTestAbort("Unable to convert pipe handle " +
+                   StreamableToString(write_handle_as_size_t) +
+                   " to a file descriptor");
+  }
+
+  // Signals the parent that the write end of the pipe has been acquired
+  // so the parent can release its own write end.
+  ::SetEvent(dup_event_handle);
+
+  return write_fd;
+}
+# endif  // GTEST_OS_WINDOWS
+
+// Returns a newly created InternalRunDeathTestFlag object with fields
+// initialized from the GTEST_FLAG(internal_run_death_test) flag if
+// the flag is specified; otherwise returns NULL.
+InternalRunDeathTestFlag* ParseInternalRunDeathTestFlag() {
+  if (GTEST_FLAG(internal_run_death_test) == "") return NULL;
+
+  // GTEST_HAS_DEATH_TEST implies that we have ::std::string, so we
+  // can use it here.
+  int line = -1;
+  int index = -1;
+  ::std::vector< ::std::string> fields;
+  SplitString(GTEST_FLAG(internal_run_death_test).c_str(), '|', &fields);
+  int write_fd = -1;
+
+# if GTEST_OS_WINDOWS
+
+  unsigned int parent_process_id = 0;
+  size_t write_handle_as_size_t = 0;
+  size_t event_handle_as_size_t = 0;
+
+  if (fields.size() != 6
+      || !ParseNaturalNumber(fields[1], &line)
+      || !ParseNaturalNumber(fields[2], &index)
+      || !ParseNaturalNumber(fields[3], &parent_process_id)
+      || !ParseNaturalNumber(fields[4], &write_handle_as_size_t)
+      || !ParseNaturalNumber(fields[5], &event_handle_as_size_t)) {
+    DeathTestAbort("Bad --gtest_internal_run_death_test flag: " +
+                   GTEST_FLAG(internal_run_death_test));
+  }
+  write_fd = GetStatusFileDescriptor(parent_process_id,
+                                     write_handle_as_size_t,
+                                     event_handle_as_size_t);
+# else
+
+  if (fields.size() != 4
+      || !ParseNaturalNumber(fields[1], &line)
+      || !ParseNaturalNumber(fields[2], &index)
+      || !ParseNaturalNumber(fields[3], &write_fd)) {
+    DeathTestAbort("Bad --gtest_internal_run_death_test flag: "
+        + GTEST_FLAG(internal_run_death_test));
+  }
+
+# endif  // GTEST_OS_WINDOWS
+
+  return new InternalRunDeathTestFlag(fields[0], line, index, write_fd);
+}
+
+}  // namespace internal
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+}  // namespace testing
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: keith.ray@gmail.com (Keith Ray)
+
+
+#include 
+
+#if GTEST_OS_WINDOWS_MOBILE
+# include 
+#elif GTEST_OS_WINDOWS
+# include 
+# include 
+#elif GTEST_OS_SYMBIAN
+// Symbian OpenC has PATH_MAX in sys/syslimits.h
+# include 
+#else
+# include 
+# include   // Some Linux distributions define PATH_MAX here.
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+#if GTEST_OS_WINDOWS
+# define GTEST_PATH_MAX_ _MAX_PATH
+#elif defined(PATH_MAX)
+# define GTEST_PATH_MAX_ PATH_MAX
+#elif defined(_XOPEN_PATH_MAX)
+# define GTEST_PATH_MAX_ _XOPEN_PATH_MAX
+#else
+# define GTEST_PATH_MAX_ _POSIX_PATH_MAX
+#endif  // GTEST_OS_WINDOWS
+
+
+namespace testing {
+namespace internal {
+
+#if GTEST_OS_WINDOWS
+// On Windows, '\\' is the standard path separator, but many tools and the
+// Windows API also accept '/' as an alternate path separator. Unless otherwise
+// noted, a file path can contain either kind of path separators, or a mixture
+// of them.
+const char kPathSeparator = '\\';
+const char kAlternatePathSeparator = '/';
+//const char kPathSeparatorString[] = "\\";
+const char kAlternatePathSeparatorString[] = "/";
+# if GTEST_OS_WINDOWS_MOBILE
+// Windows CE doesn't have a current directory. You should not use
+// the current directory in tests on Windows CE, but this at least
+// provides a reasonable fallback.
+const char kCurrentDirectoryString[] = "\\";
+// Windows CE doesn't define INVALID_FILE_ATTRIBUTES
+const DWORD kInvalidFileAttributes = 0xffffffff;
+# else
+const char kCurrentDirectoryString[] = ".\\";
+# endif  // GTEST_OS_WINDOWS_MOBILE
+#else
+const char kPathSeparator = '/';
+//const char kPathSeparatorString[] = "/";
+const char kCurrentDirectoryString[] = "./";
+#endif  // GTEST_OS_WINDOWS
+
+// Returns whether the given character is a valid path separator.
+static bool IsPathSeparator(char c) {
+#if GTEST_HAS_ALT_PATH_SEP_
+  return (c == kPathSeparator) || (c == kAlternatePathSeparator);
+#else
+  return c == kPathSeparator;
+#endif
+}
+
+// Returns the current working directory, or "" if unsuccessful.
+FilePath FilePath::GetCurrentDir() {
+#if GTEST_OS_WINDOWS_MOBILE
+  // Windows CE doesn't have a current directory, so we just return
+  // something reasonable.
+  return FilePath(kCurrentDirectoryString);
+#elif GTEST_OS_WINDOWS
+  char cwd[GTEST_PATH_MAX_ + 1] = { '\0' };
+  return FilePath(_getcwd(cwd, sizeof(cwd)) == NULL ? "" : cwd);
+#else
+  char cwd[GTEST_PATH_MAX_ + 1] = { '\0' };
+  return FilePath(getcwd(cwd, sizeof(cwd)) == NULL ? "" : cwd);
+#endif  // GTEST_OS_WINDOWS_MOBILE
+}
+
+// Returns a copy of the FilePath with the case-insensitive extension removed.
+// Example: FilePath("dir/file.exe").RemoveExtension("EXE") returns
+// FilePath("dir/file"). If a case-insensitive extension is not
+// found, returns a copy of the original FilePath.
+FilePath FilePath::RemoveExtension(const char* extension) const {
+  const std::string dot_extension = std::string(".") + extension;
+  if (String::EndsWithCaseInsensitive(pathname_, dot_extension)) {
+    return FilePath(pathname_.substr(
+        0, pathname_.length() - dot_extension.length()));
+  }
+  return *this;
+}
+
+// Returns a pointer to the last occurence of a valid path separator in
+// the FilePath. On Windows, for example, both '/' and '\' are valid path
+// separators. Returns NULL if no path separator was found.
+const char* FilePath::FindLastPathSeparator() const {
+  const char* const last_sep = strrchr(c_str(), kPathSeparator);
+#if GTEST_HAS_ALT_PATH_SEP_
+  const char* const last_alt_sep = strrchr(c_str(), kAlternatePathSeparator);
+  // Comparing two pointers of which only one is NULL is undefined.
+  if (last_alt_sep != NULL &&
+      (last_sep == NULL || last_alt_sep > last_sep)) {
+    return last_alt_sep;
+  }
+#endif
+  return last_sep;
+}
+
+// Returns a copy of the FilePath with the directory part removed.
+// Example: FilePath("path/to/file").RemoveDirectoryName() returns
+// FilePath("file"). If there is no directory part ("just_a_file"), it returns
+// the FilePath unmodified. If there is no file part ("just_a_dir/") it
+// returns an empty FilePath ("").
+// On Windows platform, '\' is the path separator, otherwise it is '/'.
+FilePath FilePath::RemoveDirectoryName() const {
+  const char* const last_sep = FindLastPathSeparator();
+  return last_sep ? FilePath(last_sep + 1) : *this;
+}
+
+// RemoveFileName returns the directory path with the filename removed.
+// Example: FilePath("path/to/file").RemoveFileName() returns "path/to/".
+// If the FilePath is "a_file" or "/a_file", RemoveFileName returns
+// FilePath("./") or, on Windows, FilePath(".\\"). If the filepath does
+// not have a file, like "just/a/dir/", it returns the FilePath unmodified.
+// On Windows platform, '\' is the path separator, otherwise it is '/'.
+FilePath FilePath::RemoveFileName() const {
+  const char* const last_sep = FindLastPathSeparator();
+  std::string dir;
+  if (last_sep) {
+    dir = std::string(c_str(), last_sep + 1 - c_str());
+  } else {
+    dir = kCurrentDirectoryString;
+  }
+  return FilePath(dir);
+}
+
+// Helper functions for naming files in a directory for xml output.
+
+// Given directory = "dir", base_name = "test", number = 0,
+// extension = "xml", returns "dir/test.xml". If number is greater
+// than zero (e.g., 12), returns "dir/test_12.xml".
+// On Windows platform, uses \ as the separator rather than /.
+FilePath FilePath::MakeFileName(const FilePath& directory,
+                                const FilePath& base_name,
+                                int number,
+                                const char* extension) {
+  std::string file;
+  if (number == 0) {
+    file = base_name.string() + "." + extension;
+  } else {
+    file = base_name.string() + "_" + StreamableToString(number)
+        + "." + extension;
+  }
+  return ConcatPaths(directory, FilePath(file));
+}
+
+// Given directory = "dir", relative_path = "test.xml", returns "dir/test.xml".
+// On Windows, uses \ as the separator rather than /.
+FilePath FilePath::ConcatPaths(const FilePath& directory,
+                               const FilePath& relative_path) {
+  if (directory.IsEmpty())
+    return relative_path;
+  const FilePath dir(directory.RemoveTrailingPathSeparator());
+  return FilePath(dir.string() + kPathSeparator + relative_path.string());
+}
+
+// Returns true if pathname describes something findable in the file-system,
+// either a file, directory, or whatever.
+bool FilePath::FileOrDirectoryExists() const {
+#if GTEST_OS_WINDOWS_MOBILE
+  LPCWSTR unicode = String::AnsiToUtf16(pathname_.c_str());
+  const DWORD attributes = GetFileAttributes(unicode);
+  delete [] unicode;
+  return attributes != kInvalidFileAttributes;
+#else
+  posix::StatStruct file_stat;
+  return posix::Stat(pathname_.c_str(), &file_stat) == 0;
+#endif  // GTEST_OS_WINDOWS_MOBILE
+}
+
+// Returns true if pathname describes a directory in the file-system
+// that exists.
+bool FilePath::DirectoryExists() const {
+  bool result = false;
+#if GTEST_OS_WINDOWS
+  // Don't strip off trailing separator if path is a root directory on
+  // Windows (like "C:\\").
+  const FilePath& path(IsRootDirectory() ? *this :
+                                           RemoveTrailingPathSeparator());
+#else
+  const FilePath& path(*this);
+#endif
+
+#if GTEST_OS_WINDOWS_MOBILE
+  LPCWSTR unicode = String::AnsiToUtf16(path.c_str());
+  const DWORD attributes = GetFileAttributes(unicode);
+  delete [] unicode;
+  if ((attributes != kInvalidFileAttributes) &&
+      (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+    result = true;
+  }
+#else
+  posix::StatStruct file_stat;
+  result = posix::Stat(path.c_str(), &file_stat) == 0 &&
+      posix::IsDir(file_stat);
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+  return result;
+}
+
+// Returns true if pathname describes a root directory. (Windows has one
+// root directory per disk drive.)
+bool FilePath::IsRootDirectory() const {
+#if GTEST_OS_WINDOWS
+  // TODO(wan@google.com): on Windows a network share like
+  // \\server\share can be a root directory, although it cannot be the
+  // current directory.  Handle this properly.
+  return pathname_.length() == 3 && IsAbsolutePath();
+#else
+  return pathname_.length() == 1 && IsPathSeparator(pathname_.c_str()[0]);
+#endif
+}
+
+// Returns true if pathname describes an absolute path.
+bool FilePath::IsAbsolutePath() const {
+  const char* const name = pathname_.c_str();
+#if GTEST_OS_WINDOWS
+  return pathname_.length() >= 3 &&
+     ((name[0] >= 'a' && name[0] <= 'z') ||
+      (name[0] >= 'A' && name[0] <= 'Z')) &&
+     name[1] == ':' &&
+     IsPathSeparator(name[2]);
+#else
+  return IsPathSeparator(name[0]);
+#endif
+}
+
+// Returns a pathname for a file that does not currently exist. The pathname
+// will be directory/base_name.extension or
+// directory/base_name_.extension if directory/base_name.extension
+// already exists. The number will be incremented until a pathname is found
+// that does not already exist.
+// Examples: 'dir/foo_test.xml' or 'dir/foo_test_1.xml'.
+// There could be a race condition if two or more processes are calling this
+// function at the same time -- they could both pick the same filename.
+FilePath FilePath::GenerateUniqueFileName(const FilePath& directory,
+                                          const FilePath& base_name,
+                                          const char* extension) {
+  FilePath full_pathname;
+  int number = 0;
+  do {
+    full_pathname.Set(MakeFileName(directory, base_name, number++, extension));
+  } while (full_pathname.FileOrDirectoryExists());
+  return full_pathname;
+}
+
+// Returns true if FilePath ends with a path separator, which indicates that
+// it is intended to represent a directory. Returns false otherwise.
+// This does NOT check that a directory (or file) actually exists.
+bool FilePath::IsDirectory() const {
+  return !pathname_.empty() &&
+         IsPathSeparator(pathname_.c_str()[pathname_.length() - 1]);
+}
+
+// Create directories so that path exists. Returns true if successful or if
+// the directories already exist; returns false if unable to create directories
+// for any reason.
+bool FilePath::CreateDirectoriesRecursively() const {
+  if (!this->IsDirectory()) {
+    return false;
+  }
+
+  if (pathname_.length() == 0 || this->DirectoryExists()) {
+    return true;
+  }
+
+  const FilePath parent(this->RemoveTrailingPathSeparator().RemoveFileName());
+  return parent.CreateDirectoriesRecursively() && this->CreateFolder();
+}
+
+// Create the directory so that path exists. Returns true if successful or
+// if the directory already exists; returns false if unable to create the
+// directory for any reason, including if the parent directory does not
+// exist. Not named "CreateDirectory" because that's a macro on Windows.
+bool FilePath::CreateFolder() const {
+#if GTEST_OS_WINDOWS_MOBILE
+  FilePath removed_sep(this->RemoveTrailingPathSeparator());
+  LPCWSTR unicode = String::AnsiToUtf16(removed_sep.c_str());
+  int result = CreateDirectory(unicode, NULL) ? 0 : -1;
+  delete [] unicode;
+#elif GTEST_OS_WINDOWS
+  int result = _mkdir(pathname_.c_str());
+#else
+  int result = mkdir(pathname_.c_str(), 0777);
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+  if (result == -1) {
+    return this->DirectoryExists();  // An error is OK if the directory exists.
+  }
+  return true;  // No error.
+}
+
+// If input name has a trailing separator character, remove it and return the
+// name, otherwise return the name string unmodified.
+// On Windows platform, uses \ as the separator, other platforms use /.
+FilePath FilePath::RemoveTrailingPathSeparator() const {
+  return IsDirectory()
+      ? FilePath(pathname_.substr(0, pathname_.length() - 1))
+      : *this;
+}
+
+// Removes any redundant separators that might be in the pathname.
+// For example, "bar///foo" becomes "bar/foo". Does not eliminate other
+// redundancies that might be in a pathname involving "." or "..".
+// TODO(wan@google.com): handle Windows network shares (e.g. \\server\share).
+void FilePath::Normalize() {
+  if (pathname_.c_str() == NULL) {
+    pathname_ = "";
+    return;
+  }
+  const char* src = pathname_.c_str();
+  char* const dest = new char[pathname_.length() + 1];
+  char* dest_ptr = dest;
+  memset(dest_ptr, 0, pathname_.length() + 1);
+
+  while (*src != '\0') {
+    *dest_ptr = *src;
+    if (!IsPathSeparator(*src)) {
+      src++;
+    } else {
+#if GTEST_HAS_ALT_PATH_SEP_
+      if (*dest_ptr == kAlternatePathSeparator) {
+        *dest_ptr = kPathSeparator;
+      }
+#endif
+      while (IsPathSeparator(*src))
+        src++;
+    }
+    dest_ptr++;
+  }
+  *dest_ptr = '\0';
+  pathname_ = dest;
+  delete[] dest;
+}
+
+}  // namespace internal
+}  // namespace testing
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+
+#include 
+#include 
+#include 
+#include 
+
+#if GTEST_OS_WINDOWS_MOBILE
+# include   // For TerminateProcess()
+#elif GTEST_OS_WINDOWS
+# include 
+# include 
+#else
+# include 
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+#if GTEST_OS_MAC
+# include 
+# include 
+# include 
+#endif  // GTEST_OS_MAC
+
+#if GTEST_OS_QNX
+# include 
+# include 
+#endif  // GTEST_OS_QNX
+
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick is to
+// prevent a user from accidentally including gtest-internal-inl.h in
+// his code.
+#define GTEST_IMPLEMENTATION_ 1
+#undef GTEST_IMPLEMENTATION_
+
+namespace testing {
+namespace internal {
+
+#if defined(_MSC_VER) || defined(__BORLANDC__)
+// MSVC and C++Builder do not provide a definition of STDERR_FILENO.
+const int kStdOutFileno = 1;
+const int kStdErrFileno = 2;
+#else
+const int kStdOutFileno = STDOUT_FILENO;
+const int kStdErrFileno = STDERR_FILENO;
+#endif  // _MSC_VER
+
+#if GTEST_OS_MAC
+
+// Returns the number of threads running in the process, or 0 to indicate that
+// we cannot detect it.
+size_t GetThreadCount() {
+  const task_t task = mach_task_self();
+  mach_msg_type_number_t thread_count;
+  thread_act_array_t thread_list;
+  const kern_return_t status = task_threads(task, &thread_list, &thread_count);
+  if (status == KERN_SUCCESS) {
+    // task_threads allocates resources in thread_list and we need to free them
+    // to avoid leaks.
+    vm_deallocate(task,
+                  reinterpret_cast(thread_list),
+                  sizeof(thread_t) * thread_count);
+    return static_cast(thread_count);
+  } else {
+    return 0;
+  }
+}
+
+#elif GTEST_OS_QNX
+
+// Returns the number of threads running in the process, or 0 to indicate that
+// we cannot detect it.
+size_t GetThreadCount() {
+  const int fd = open("/proc/self/as", O_RDONLY);
+  if (fd < 0) {
+    return 0;
+  }
+  procfs_info process_info;
+  const int status =
+      devctl(fd, DCMD_PROC_INFO, &process_info, sizeof(process_info), NULL);
+  close(fd);
+  if (status == EOK) {
+    return static_cast(process_info.num_threads);
+  } else {
+    return 0;
+  }
+}
+
+#else
+
+size_t GetThreadCount() {
+  // There's no portable way to detect the number of threads, so we just
+  // return 0 to indicate that we cannot detect it.
+  return 0;
+}
+
+#endif  // GTEST_OS_MAC
+
+#if GTEST_USES_POSIX_RE
+
+// Implements RE.  Currently only needed for death tests.
+
+RE::~RE() {
+  if (is_valid_) {
+    // regfree'ing an invalid regex might crash because the content
+    // of the regex is undefined. Since the regex's are essentially
+    // the same, one cannot be valid (or invalid) without the other
+    // being so too.
+    regfree(&partial_regex_);
+    regfree(&full_regex_);
+  }
+  free(const_cast(pattern_));
+}
+
+// Returns true iff regular expression re matches the entire str.
+bool RE::FullMatch(const char* str, const RE& re) {
+  if (!re.is_valid_) return false;
+
+  regmatch_t match;
+  return regexec(&re.full_regex_, str, 1, &match, 0) == 0;
+}
+
+// Returns true iff regular expression re matches a substring of str
+// (including str itself).
+bool RE::PartialMatch(const char* str, const RE& re) {
+  if (!re.is_valid_) return false;
+
+  regmatch_t match;
+  return regexec(&re.partial_regex_, str, 1, &match, 0) == 0;
+}
+
+// Initializes an RE from its string representation.
+void RE::Init(const char* regex) {
+  pattern_ = posix::StrDup(regex);
+
+  // Reserves enough bytes to hold the regular expression used for a
+  // full match.
+  const size_t full_regex_len = strlen(regex) + 10;
+  char* const full_pattern = new char[full_regex_len];
+
+  snprintf(full_pattern, full_regex_len, "^(%s)$", regex);
+  is_valid_ = regcomp(&full_regex_, full_pattern, REG_EXTENDED) == 0;
+  // We want to call regcomp(&partial_regex_, ...) even if the
+  // previous expression returns false.  Otherwise partial_regex_ may
+  // not be properly initialized can may cause trouble when it's
+  // freed.
+  //
+  // Some implementation of POSIX regex (e.g. on at least some
+  // versions of Cygwin) doesn't accept the empty string as a valid
+  // regex.  We change it to an equivalent form "()" to be safe.
+  if (is_valid_) {
+    const char* const partial_regex = (*regex == '\0') ? "()" : regex;
+    is_valid_ = regcomp(&partial_regex_, partial_regex, REG_EXTENDED) == 0;
+  }
+  EXPECT_TRUE(is_valid_)
+      << "Regular expression \"" << regex
+      << "\" is not a valid POSIX Extended regular expression.";
+
+  delete[] full_pattern;
+}
+
+#elif GTEST_USES_SIMPLE_RE
+
+// Returns true iff ch appears anywhere in str (excluding the
+// terminating '\0' character).
+bool IsInSet(char ch, const char* str) {
+  return ch != '\0' && strchr(str, ch) != NULL;
+}
+
+// Returns true iff ch belongs to the given classification.  Unlike
+// similar functions in , these aren't affected by the
+// current locale.
+bool IsAsciiDigit(char ch) { return '0' <= ch && ch <= '9'; }
+bool IsAsciiPunct(char ch) {
+  return IsInSet(ch, "^-!\"#$%&'()*+,./:;<=>?@[\\]_`{|}~");
+}
+bool IsRepeat(char ch) { return IsInSet(ch, "?*+"); }
+bool IsAsciiWhiteSpace(char ch) { return IsInSet(ch, " \f\n\r\t\v"); }
+bool IsAsciiWordChar(char ch) {
+  return ('a' <= ch && ch <= 'z') || ('A' <= ch && ch <= 'Z') ||
+      ('0' <= ch && ch <= '9') || ch == '_';
+}
+
+// Returns true iff "\\c" is a supported escape sequence.
+bool IsValidEscape(char c) {
+  return (IsAsciiPunct(c) || IsInSet(c, "dDfnrsStvwW"));
+}
+
+// Returns true iff the given atom (specified by escaped and pattern)
+// matches ch.  The result is undefined if the atom is invalid.
+bool AtomMatchesChar(bool escaped, char pattern_char, char ch) {
+  if (escaped) {  // "\\p" where p is pattern_char.
+    switch (pattern_char) {
+      case 'd': return IsAsciiDigit(ch);
+      case 'D': return !IsAsciiDigit(ch);
+      case 'f': return ch == '\f';
+      case 'n': return ch == '\n';
+      case 'r': return ch == '\r';
+      case 's': return IsAsciiWhiteSpace(ch);
+      case 'S': return !IsAsciiWhiteSpace(ch);
+      case 't': return ch == '\t';
+      case 'v': return ch == '\v';
+      case 'w': return IsAsciiWordChar(ch);
+      case 'W': return !IsAsciiWordChar(ch);
+    }
+    return IsAsciiPunct(pattern_char) && pattern_char == ch;
+  }
+
+  return (pattern_char == '.' && ch != '\n') || pattern_char == ch;
+}
+
+// Helper function used by ValidateRegex() to format error messages.
+std::string FormatRegexSyntaxError(const char* regex, int index) {
+  return (Message() << "Syntax error at index " << index
+          << " in simple regular expression \"" << regex << "\": ").GetString();
+}
+
+// Generates non-fatal failures and returns false if regex is invalid;
+// otherwise returns true.
+bool ValidateRegex(const char* regex) {
+  if (regex == NULL) {
+    // TODO(wan@google.com): fix the source file location in the
+    // assertion failures to match where the regex is used in user
+    // code.
+    ADD_FAILURE() << "NULL is not a valid simple regular expression.";
+    return false;
+  }
+
+  bool is_valid = true;
+
+  // True iff ?, *, or + can follow the previous atom.
+  bool prev_repeatable = false;
+  for (int i = 0; regex[i]; i++) {
+    if (regex[i] == '\\') {  // An escape sequence
+      i++;
+      if (regex[i] == '\0') {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i - 1)
+                      << "'\\' cannot appear at the end.";
+        return false;
+      }
+
+      if (!IsValidEscape(regex[i])) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i - 1)
+                      << "invalid escape sequence \"\\" << regex[i] << "\".";
+        is_valid = false;
+      }
+      prev_repeatable = true;
+    } else {  // Not an escape sequence.
+      const char ch = regex[i];
+
+      if (ch == '^' && i > 0) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'^' can only appear at the beginning.";
+        is_valid = false;
+      } else if (ch == '$' && regex[i + 1] != '\0') {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'$' can only appear at the end.";
+        is_valid = false;
+      } else if (IsInSet(ch, "()[]{}|")) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'" << ch << "' is unsupported.";
+        is_valid = false;
+      } else if (IsRepeat(ch) && !prev_repeatable) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'" << ch << "' can only follow a repeatable token.";
+        is_valid = false;
+      }
+
+      prev_repeatable = !IsInSet(ch, "^$?*+");
+    }
+  }
+
+  return is_valid;
+}
+
+// Matches a repeated regex atom followed by a valid simple regular
+// expression.  The regex atom is defined as c if escaped is false,
+// or \c otherwise.  repeat is the repetition meta character (?, *,
+// or +).  The behavior is undefined if str contains too many
+// characters to be indexable by size_t, in which case the test will
+// probably time out anyway.  We are fine with this limitation as
+// std::string has it too.
+bool MatchRepetitionAndRegexAtHead(
+    bool escaped, char c, char repeat, const char* regex,
+    const char* str) {
+  const size_t min_count = (repeat == '+') ? 1 : 0;
+  const size_t max_count = (repeat == '?') ? 1 :
+      static_cast(-1) - 1;
+  // We cannot call numeric_limits::max() as it conflicts with the
+  // max() macro on Windows.
+
+  for (size_t i = 0; i <= max_count; ++i) {
+    // We know that the atom matches each of the first i characters in str.
+    if (i >= min_count && MatchRegexAtHead(regex, str + i)) {
+      // We have enough matches at the head, and the tail matches too.
+      // Since we only care about *whether* the pattern matches str
+      // (as opposed to *how* it matches), there is no need to find a
+      // greedy match.
+      return true;
+    }
+    if (str[i] == '\0' || !AtomMatchesChar(escaped, c, str[i]))
+      return false;
+  }
+  return false;
+}
+
+// Returns true iff regex matches a prefix of str.  regex must be a
+// valid simple regular expression and not start with "^", or the
+// result is undefined.
+bool MatchRegexAtHead(const char* regex, const char* str) {
+  if (*regex == '\0')  // An empty regex matches a prefix of anything.
+    return true;
+
+  // "$" only matches the end of a string.  Note that regex being
+  // valid guarantees that there's nothing after "$" in it.
+  if (*regex == '$')
+    return *str == '\0';
+
+  // Is the first thing in regex an escape sequence?
+  const bool escaped = *regex == '\\';
+  if (escaped)
+    ++regex;
+  if (IsRepeat(regex[1])) {
+    // MatchRepetitionAndRegexAtHead() calls MatchRegexAtHead(), so
+    // here's an indirect recursion.  It terminates as the regex gets
+    // shorter in each recursion.
+    return MatchRepetitionAndRegexAtHead(
+        escaped, regex[0], regex[1], regex + 2, str);
+  } else {
+    // regex isn't empty, isn't "$", and doesn't start with a
+    // repetition.  We match the first atom of regex with the first
+    // character of str and recurse.
+    return (*str != '\0') && AtomMatchesChar(escaped, *regex, *str) &&
+        MatchRegexAtHead(regex + 1, str + 1);
+  }
+}
+
+// Returns true iff regex matches any substring of str.  regex must be
+// a valid simple regular expression, or the result is undefined.
+//
+// The algorithm is recursive, but the recursion depth doesn't exceed
+// the regex length, so we won't need to worry about running out of
+// stack space normally.  In rare cases the time complexity can be
+// exponential with respect to the regex length + the string length,
+// but usually it's must faster (often close to linear).
+bool MatchRegexAnywhere(const char* regex, const char* str) {
+  if (regex == NULL || str == NULL)
+    return false;
+
+  if (*regex == '^')
+    return MatchRegexAtHead(regex + 1, str);
+
+  // A successful match can be anywhere in str.
+  do {
+    if (MatchRegexAtHead(regex, str))
+      return true;
+  } while (*str++ != '\0');
+  return false;
+}
+
+// Implements the RE class.
+
+RE::~RE() {
+  free(const_cast(pattern_));
+  free(const_cast(full_pattern_));
+}
+
+// Returns true iff regular expression re matches the entire str.
+bool RE::FullMatch(const char* str, const RE& re) {
+  return re.is_valid_ && MatchRegexAnywhere(re.full_pattern_, str);
+}
+
+// Returns true iff regular expression re matches a substring of str
+// (including str itself).
+bool RE::PartialMatch(const char* str, const RE& re) {
+  return re.is_valid_ && MatchRegexAnywhere(re.pattern_, str);
+}
+
+// Initializes an RE from its string representation.
+void RE::Init(const char* regex) {
+  pattern_ = full_pattern_ = NULL;
+  if (regex != NULL) {
+    pattern_ = posix::StrDup(regex);
+  }
+
+  is_valid_ = ValidateRegex(regex);
+  if (!is_valid_) {
+    // No need to calculate the full pattern when the regex is invalid.
+    return;
+  }
+
+  const size_t len = strlen(regex);
+  // Reserves enough bytes to hold the regular expression used for a
+  // full match: we need space to prepend a '^', append a '$', and
+  // terminate the string with '\0'.
+  char* buffer = static_cast(malloc(len + 3));
+  full_pattern_ = buffer;
+
+  if (*regex != '^')
+    *buffer++ = '^';  // Makes sure full_pattern_ starts with '^'.
+
+  // We don't use snprintf or strncpy, as they trigger a warning when
+  // compiled with VC++ 8.0.
+  memcpy(buffer, regex, len);
+  buffer += len;
+
+  if (len == 0 || regex[len - 1] != '$')
+    *buffer++ = '$';  // Makes sure full_pattern_ ends with '$'.
+
+  *buffer = '\0';
+}
+
+#endif  // GTEST_USES_POSIX_RE
+
+const char kUnknownFile[] = "unknown file";
+
+// Formats a source file path and a line number as they would appear
+// in an error message from the compiler used to compile this code.
+GTEST_API_ ::std::string FormatFileLocation(const char* file, int line) {
+  const std::string file_name(file == NULL ? kUnknownFile : file);
+
+  if (line < 0) {
+    return file_name + ":";
+  }
+#ifdef _MSC_VER
+  return file_name + "(" + StreamableToString(line) + "):";
+#else
+  return file_name + ":" + StreamableToString(line) + ":";
+#endif  // _MSC_VER
+}
+
+// Formats a file location for compiler-independent XML output.
+// Although this function is not platform dependent, we put it next to
+// FormatFileLocation in order to contrast the two functions.
+// Note that FormatCompilerIndependentFileLocation() does NOT append colon
+// to the file location it produces, unlike FormatFileLocation().
+GTEST_API_ ::std::string FormatCompilerIndependentFileLocation(
+    const char* file, int line) {
+  const std::string file_name(file == NULL ? kUnknownFile : file);
+
+  if (line < 0)
+    return file_name;
+  else
+    return file_name + ":" + StreamableToString(line);
+}
+
+
+GTestLog::GTestLog(GTestLogSeverity severity, const char* file, int line)
+    : severity_(severity) {
+  const char* const marker =
+      severity == GTEST_INFO ?    "[  INFO ]" :
+      severity == GTEST_WARNING ? "[WARNING]" :
+      severity == GTEST_ERROR ?   "[ ERROR ]" : "[ FATAL ]";
+  GetStream() << ::std::endl << marker << " "
+              << FormatFileLocation(file, line).c_str() << ": ";
+}
+
+// Flushes the buffers and, if severity is GTEST_FATAL, aborts the program.
+GTestLog::~GTestLog() {
+  GetStream() << ::std::endl;
+  if (severity_ == GTEST_FATAL) {
+    fflush(stderr);
+    posix::Abort();
+  }
+}
+// Disable Microsoft deprecation warnings for POSIX functions called from
+// this class (creat, dup, dup2, and close)
+#ifdef _MSC_VER
+# pragma warning(push)
+# pragma warning(disable: 4996)
+#endif  // _MSC_VER
+
+#if GTEST_HAS_STREAM_REDIRECTION
+
+// Object that captures an output stream (stdout/stderr).
+class CapturedStream {
+ public:
+  // The ctor redirects the stream to a temporary file.
+  explicit CapturedStream(int fd) : fd_(fd), uncaptured_fd_(dup(fd)) {
+# if GTEST_OS_WINDOWS
+    char temp_dir_path[MAX_PATH + 1] = { '\0' };  // NOLINT
+    char temp_file_path[MAX_PATH + 1] = { '\0' };  // NOLINT
+
+    ::GetTempPathA(sizeof(temp_dir_path), temp_dir_path);
+    const UINT success = ::GetTempFileNameA(temp_dir_path,
+                                            "gtest_redir",
+                                            0,  // Generate unique file name.
+                                            temp_file_path);
+    GTEST_CHECK_(success != 0)
+        << "Unable to create a temporary file in " << temp_dir_path;
+    const int captured_fd = creat(temp_file_path, _S_IREAD | _S_IWRITE);
+    GTEST_CHECK_(captured_fd != -1) << "Unable to open temporary file "
+                                    << temp_file_path;
+    filename_ = temp_file_path;
+# else
+    // There's no guarantee that a test has write access to the current
+    // directory, so we create the temporary file in the /tmp directory
+    // instead. We use /tmp on most systems, and /sdcard on Android.
+    // That's because Android doesn't have /tmp.
+#  if GTEST_OS_LINUX_ANDROID
+    // Note: Android applications are expected to call the framework's
+    // Context.getExternalStorageDirectory() method through JNI to get
+    // the location of the world-writable SD Card directory. However,
+    // this requires a Context handle, which cannot be retrieved
+    // globally from native code. Doing so also precludes running the
+    // code as part of a regular standalone executable, which doesn't
+    // run in a Dalvik process (e.g. when running it through 'adb shell').
+    //
+    // The location /sdcard is directly accessible from native code
+    // and is the only location (unofficially) supported by the Android
+    // team. It's generally a symlink to the real SD Card mount point
+    // which can be /mnt/sdcard, /mnt/sdcard0, /system/media/sdcard, or
+    // other OEM-customized locations. Never rely on these, and always
+    // use /sdcard.
+    char name_template[] = "/sdcard/gtest_captured_stream.XXXXXX";
+#  else
+    char name_template[] = "/tmp/captured_stream.XXXXXX";
+#  endif  // GTEST_OS_LINUX_ANDROID
+    const int captured_fd = mkstemp(name_template);
+    filename_ = name_template;
+# endif  // GTEST_OS_WINDOWS
+    fflush(NULL);
+    dup2(captured_fd, fd_);
+    close(captured_fd);
+  }
+
+  ~CapturedStream() {
+    remove(filename_.c_str());
+  }
+
+  std::string GetCapturedString() {
+    if (uncaptured_fd_ != -1) {
+      // Restores the original stream.
+      fflush(NULL);
+      dup2(uncaptured_fd_, fd_);
+      close(uncaptured_fd_);
+      uncaptured_fd_ = -1;
+    }
+
+    FILE* const file = posix::FOpen(filename_.c_str(), "r");
+    const std::string content = ReadEntireFile(file);
+    posix::FClose(file);
+    return content;
+  }
+
+ private:
+  // Reads the entire content of a file as an std::string.
+  static std::string ReadEntireFile(FILE* file);
+
+  // Returns the size (in bytes) of a file.
+  static size_t GetFileSize(FILE* file);
+
+  const int fd_;  // A stream to capture.
+  int uncaptured_fd_;
+  // Name of the temporary file holding the stderr output.
+  ::std::string filename_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(CapturedStream);
+};
+
+// Returns the size (in bytes) of a file.
+size_t CapturedStream::GetFileSize(FILE* file) {
+  fseek(file, 0, SEEK_END);
+  return static_cast(ftell(file));
+}
+
+// Reads the entire content of a file as a string.
+std::string CapturedStream::ReadEntireFile(FILE* file) {
+  const size_t file_size = GetFileSize(file);
+  char* const buffer = new char[file_size];
+
+  size_t bytes_last_read = 0;  // # of bytes read in the last fread()
+  size_t bytes_read = 0;       // # of bytes read so far
+
+  fseek(file, 0, SEEK_SET);
+
+  // Keeps reading the file until we cannot read further or the
+  // pre-determined file size is reached.
+  do {
+    bytes_last_read = fread(buffer+bytes_read, 1, file_size-bytes_read, file);
+    bytes_read += bytes_last_read;
+  } while (bytes_last_read > 0 && bytes_read < file_size);
+
+  const std::string content(buffer, bytes_read);
+  delete[] buffer;
+
+  return content;
+}
+
+# ifdef _MSC_VER
+#  pragma warning(pop)
+# endif  // _MSC_VER
+
+static CapturedStream* g_captured_stderr = NULL;
+static CapturedStream* g_captured_stdout = NULL;
+
+// Starts capturing an output stream (stdout/stderr).
+void CaptureStream(int fd, const char* stream_name, CapturedStream** stream) {
+  if (*stream != NULL) {
+    GTEST_LOG_(FATAL) << "Only one " << stream_name
+                      << " capturer can exist at a time.";
+  }
+  *stream = new CapturedStream(fd);
+}
+
+// Stops capturing the output stream and returns the captured string.
+std::string GetCapturedStream(CapturedStream** captured_stream) {
+  const std::string content = (*captured_stream)->GetCapturedString();
+
+  delete *captured_stream;
+  *captured_stream = NULL;
+
+  return content;
+}
+
+// Starts capturing stdout.
+void CaptureStdout() {
+  CaptureStream(kStdOutFileno, "stdout", &g_captured_stdout);
+}
+
+// Starts capturing stderr.
+void CaptureStderr() {
+  CaptureStream(kStdErrFileno, "stderr", &g_captured_stderr);
+}
+
+// Stops capturing stdout and returns the captured string.
+std::string GetCapturedStdout() {
+  return GetCapturedStream(&g_captured_stdout);
+}
+
+// Stops capturing stderr and returns the captured string.
+std::string GetCapturedStderr() {
+  return GetCapturedStream(&g_captured_stderr);
+}
+
+#endif  // GTEST_HAS_STREAM_REDIRECTION
+
+#if GTEST_HAS_DEATH_TEST
+
+// A copy of all command line arguments.  Set by InitGoogleTest().
+::std::vector g_argvs;
+
+static const ::std::vector* g_injected_test_argvs =
+                                        NULL;  // Owned.
+
+void SetInjectableArgvs(const ::std::vector* argvs) {
+  if (g_injected_test_argvs != argvs)
+    delete g_injected_test_argvs;
+  g_injected_test_argvs = argvs;
+}
+
+const ::std::vector& GetInjectableArgvs() {
+  if (g_injected_test_argvs != NULL) {
+    return *g_injected_test_argvs;
+  }
+  return g_argvs;
+}
+#endif  // GTEST_HAS_DEATH_TEST
+
+#if GTEST_OS_WINDOWS_MOBILE
+namespace posix {
+void Abort() {
+  DebugBreak();
+  TerminateProcess(GetCurrentProcess(), 1);
+}
+}  // namespace posix
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+// Returns the name of the environment variable corresponding to the
+// given flag.  For example, FlagToEnvVar("foo") will return
+// "GTEST_FOO" in the open-source version.
+static std::string FlagToEnvVar(const char* flag) {
+  const std::string full_flag =
+      (Message() << GTEST_FLAG_PREFIX_ << flag).GetString();
+
+  Message env_var;
+  for (size_t i = 0; i != full_flag.length(); i++) {
+    env_var << ToUpper(full_flag.c_str()[i]);
+  }
+
+  return env_var.GetString();
+}
+
+// Parses 'str' for a 32-bit signed integer.  If successful, writes
+// the result to *value and returns true; otherwise leaves *value
+// unchanged and returns false.
+bool ParseInt32(const Message& src_text, const char* str, Int32* value) {
+  // Parses the environment variable as a decimal integer.
+  char* end = NULL;
+  const long long_value = strtol(str, &end, 10);  // NOLINT
+
+  // Has strtol() consumed all characters in the string?
+  if (*end != '\0') {
+    // No - an invalid character was encountered.
+    Message msg;
+    msg << "WARNING: " << src_text
+        << " is expected to be a 32-bit integer, but actually"
+        << " has value \"" << str << "\".\n";
+    printf("%s", msg.GetString().c_str());
+    fflush(stdout);
+    return false;
+  }
+
+  // Is the parsed value in the range of an Int32?
+  const Int32 result = static_cast(long_value);
+  if (long_value == LONG_MAX || long_value == LONG_MIN ||
+      // The parsed value overflows as a long.  (strtol() returns
+      // LONG_MAX or LONG_MIN when the input overflows.)
+      result != long_value
+      // The parsed value overflows as an Int32.
+      ) {
+    Message msg;
+    msg << "WARNING: " << src_text
+        << " is expected to be a 32-bit integer, but actually"
+        << " has value " << str << ", which overflows.\n";
+    printf("%s", msg.GetString().c_str());
+    fflush(stdout);
+    return false;
+  }
+
+  *value = result;
+  return true;
+}
+
+// Reads and returns the Boolean environment variable corresponding to
+// the given flag; if it's not set, returns default_value.
+//
+// The value is considered true iff it's not "0".
+bool BoolFromGTestEnv(const char* flag, bool default_value) {
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* const string_value = posix::GetEnv(env_var.c_str());
+  return string_value == NULL ?
+      default_value : strcmp(string_value, "0") != 0;
+}
+
+// Reads and returns a 32-bit integer stored in the environment
+// variable corresponding to the given flag; if it isn't set or
+// doesn't represent a valid 32-bit integer, returns default_value.
+Int32 Int32FromGTestEnv(const char* flag, Int32 default_value) {
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* const string_value = posix::GetEnv(env_var.c_str());
+  if (string_value == NULL) {
+    // The environment variable is not set.
+    return default_value;
+  }
+
+  Int32 result = default_value;
+  if (!ParseInt32(Message() << "Environment variable " << env_var,
+                  string_value, &result)) {
+    printf("The default value %s is used.\n",
+           (Message() << default_value).GetString().c_str());
+    fflush(stdout);
+    return default_value;
+  }
+
+  return result;
+}
+
+// Reads and returns the string environment variable corresponding to
+// the given flag; if it's not set, returns default_value.
+const char* StringFromGTestEnv(const char* flag, const char* default_value) {
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* const value = posix::GetEnv(env_var.c_str());
+  return value == NULL ? default_value : value;
+}
+
+}  // namespace internal
+}  // namespace testing
+// Copyright 2007, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Google Test - The Google C++ Testing Framework
+//
+// This file implements a universal value printer that can print a
+// value of any type T:
+//
+//   void ::testing::internal::UniversalPrinter::Print(value, ostream_ptr);
+//
+// It uses the << operator when possible, and prints the bytes in the
+// object otherwise.  A user can override its behavior for a class
+// type Foo by defining either operator<<(::std::ostream&, const Foo&)
+// or void PrintTo(const Foo&, ::std::ostream*) in the namespace that
+// defines Foo.
+
+#include 
+#include 
+#include   // NOLINT
+#include 
+
+namespace testing {
+
+namespace {
+
+using ::std::ostream;
+
+// Prints a segment of bytes in the given object.
+void PrintByteSegmentInObjectTo(const unsigned char* obj_bytes, size_t start,
+                                size_t count, ostream* os) {
+  char text[5] = "";
+  for (size_t i = 0; i != count; i++) {
+    const size_t j = start + i;
+    if (i != 0) {
+      // Organizes the bytes into groups of 2 for easy parsing by
+      // human.
+      if ((j % 2) == 0)
+        *os << ' ';
+      else
+        *os << '-';
+    }
+    GTEST_SNPRINTF_(text, sizeof(text), "%02X", obj_bytes[j]);
+    *os << text;
+  }
+}
+
+// Prints the bytes in the given value to the given ostream.
+void PrintBytesInObjectToImpl(const unsigned char* obj_bytes, size_t count,
+                              ostream* os) {
+  // Tells the user how big the object is.
+  *os << count << "-byte object <";
+
+  const size_t kThreshold = 132;
+  const size_t kChunkSize = 64;
+  // If the object size is bigger than kThreshold, we'll have to omit
+  // some details by printing only the first and the last kChunkSize
+  // bytes.
+  // TODO(wan): let the user control the threshold using a flag.
+  if (count < kThreshold) {
+    PrintByteSegmentInObjectTo(obj_bytes, 0, count, os);
+  } else {
+    PrintByteSegmentInObjectTo(obj_bytes, 0, kChunkSize, os);
+    *os << " ... ";
+    // Rounds up to 2-byte boundary.
+    const size_t resume_pos = (count - kChunkSize + 1)/2*2;
+    PrintByteSegmentInObjectTo(obj_bytes, resume_pos, count - resume_pos, os);
+  }
+  *os << ">";
+}
+
+}  // namespace
+
+namespace internal2 {
+
+// Delegates to PrintBytesInObjectToImpl() to print the bytes in the
+// given object.  The delegation simplifies the implementation, which
+// uses the << operator and thus is easier done outside of the
+// ::testing::internal namespace, which contains a << operator that
+// sometimes conflicts with the one in STL.
+void PrintBytesInObjectTo(const unsigned char* obj_bytes, size_t count,
+                          ostream* os) {
+  PrintBytesInObjectToImpl(obj_bytes, count, os);
+}
+
+}  // namespace internal2
+
+namespace internal {
+
+// Depending on the value of a char (or wchar_t), we print it in one
+// of three formats:
+//   - as is if it's a printable ASCII (e.g. 'a', '2', ' '),
+//   - as a hexidecimal escape sequence (e.g. '\x7F'), or
+//   - as a special escape sequence (e.g. '\r', '\n').
+enum CharFormat {
+  kAsIs,
+  kHexEscape,
+  kSpecialEscape
+};
+
+// Returns true if c is a printable ASCII character.  We test the
+// value of c directly instead of calling isprint(), which is buggy on
+// Windows Mobile.
+inline bool IsPrintableAscii(wchar_t c) {
+  return 0x20 <= c && c <= 0x7E;
+}
+
+// Prints a wide or narrow char c as a character literal without the
+// quotes, escaping it when necessary; returns how c was formatted.
+// The template argument UnsignedChar is the unsigned version of Char,
+// which is the type of c.
+template 
+static CharFormat PrintAsCharLiteralTo(Char c, ostream* os) {
+  switch (static_cast(c)) {
+    case L'\0':
+      *os << "\\0";
+      break;
+    case L'\'':
+      *os << "\\'";
+      break;
+    case L'\\':
+      *os << "\\\\";
+      break;
+    case L'\a':
+      *os << "\\a";
+      break;
+    case L'\b':
+      *os << "\\b";
+      break;
+    case L'\f':
+      *os << "\\f";
+      break;
+    case L'\n':
+      *os << "\\n";
+      break;
+    case L'\r':
+      *os << "\\r";
+      break;
+    case L'\t':
+      *os << "\\t";
+      break;
+    case L'\v':
+      *os << "\\v";
+      break;
+    default:
+      if (IsPrintableAscii(c)) {
+        *os << static_cast(c);
+        return kAsIs;
+      } else {
+        *os << "\\x" + String::FormatHexInt(static_cast(c));
+        return kHexEscape;
+      }
+  }
+  return kSpecialEscape;
+}
+
+// Prints a wchar_t c as if it's part of a string literal, escaping it when
+// necessary; returns how c was formatted.
+static CharFormat PrintAsStringLiteralTo(wchar_t c, ostream* os) {
+  switch (c) {
+    case L'\'':
+      *os << "'";
+      return kAsIs;
+    case L'"':
+      *os << "\\\"";
+      return kSpecialEscape;
+    default:
+      return PrintAsCharLiteralTo(c, os);
+  }
+}
+
+// Prints a char c as if it's part of a string literal, escaping it when
+// necessary; returns how c was formatted.
+static CharFormat PrintAsStringLiteralTo(char c, ostream* os) {
+  return PrintAsStringLiteralTo(
+      static_cast(static_cast(c)), os);
+}
+
+// Prints a wide or narrow character c and its code.  '\0' is printed
+// as "'\\0'", other unprintable characters are also properly escaped
+// using the standard C++ escape sequence.  The template argument
+// UnsignedChar is the unsigned version of Char, which is the type of c.
+template 
+void PrintCharAndCodeTo(Char c, ostream* os) {
+  // First, print c as a literal in the most readable form we can find.
+  *os << ((sizeof(c) > 1) ? "L'" : "'");
+  const CharFormat format = PrintAsCharLiteralTo(c, os);
+  *os << "'";
+
+  // To aid user debugging, we also print c's code in decimal, unless
+  // it's 0 (in which case c was printed as '\\0', making the code
+  // obvious).
+  if (c == 0)
+    return;
+  *os << " (" << static_cast(c);
+
+  // For more convenience, we print c's code again in hexidecimal,
+  // unless c was already printed in the form '\x##' or the code is in
+  // [1, 9].
+  if (format == kHexEscape || (1 <= c && c <= 9)) {
+    // Do nothing.
+  } else {
+    *os << ", 0x" << String::FormatHexInt(static_cast(c));
+  }
+  *os << ")";
+}
+
+void PrintTo(unsigned char c, ::std::ostream* os) {
+  PrintCharAndCodeTo(c, os);
+}
+void PrintTo(signed char c, ::std::ostream* os) {
+  PrintCharAndCodeTo(c, os);
+}
+
+// Prints a wchar_t as a symbol if it is printable or as its internal
+// code otherwise and also as its code.  L'\0' is printed as "L'\\0'".
+void PrintTo(wchar_t wc, ostream* os) {
+  PrintCharAndCodeTo(wc, os);
+}
+
+// Prints the given array of characters to the ostream.  CharType must be either
+// char or wchar_t.
+// The array starts at begin, the length is len, it may include '\0' characters
+// and may not be NUL-terminated.
+template 
+static void PrintCharsAsStringTo(
+    const CharType* begin, size_t len, ostream* os) {
+  const char* const kQuoteBegin = sizeof(CharType) == 1 ? "\"" : "L\"";
+  *os << kQuoteBegin;
+  bool is_previous_hex = false;
+  for (size_t index = 0; index < len; ++index) {
+    const CharType cur = begin[index];
+    if (is_previous_hex && IsXDigit(cur)) {
+      // Previous character is of '\x..' form and this character can be
+      // interpreted as another hexadecimal digit in its number. Break string to
+      // disambiguate.
+      *os << "\" " << kQuoteBegin;
+    }
+    is_previous_hex = PrintAsStringLiteralTo(cur, os) == kHexEscape;
+  }
+  *os << "\"";
+}
+
+// Prints a (const) char/wchar_t array of 'len' elements, starting at address
+// 'begin'.  CharType must be either char or wchar_t.
+template 
+static void UniversalPrintCharArray(
+    const CharType* begin, size_t len, ostream* os) {
+  // The code
+  //   const char kFoo[] = "foo";
+  // generates an array of 4, not 3, elements, with the last one being '\0'.
+  //
+  // Therefore when printing a char array, we don't print the last element if
+  // it's '\0', such that the output matches the string literal as it's
+  // written in the source code.
+  if (len > 0 && begin[len - 1] == '\0') {
+    PrintCharsAsStringTo(begin, len - 1, os);
+    return;
+  }
+
+  // If, however, the last element in the array is not '\0', e.g.
+  //    const char kFoo[] = { 'f', 'o', 'o' };
+  // we must print the entire array.  We also print a message to indicate
+  // that the array is not NUL-terminated.
+  PrintCharsAsStringTo(begin, len, os);
+  *os << " (no terminating NUL)";
+}
+
+// Prints a (const) char array of 'len' elements, starting at address 'begin'.
+void UniversalPrintArray(const char* begin, size_t len, ostream* os) {
+  UniversalPrintCharArray(begin, len, os);
+}
+
+// Prints a (const) wchar_t array of 'len' elements, starting at address
+// 'begin'.
+void UniversalPrintArray(const wchar_t* begin, size_t len, ostream* os) {
+  UniversalPrintCharArray(begin, len, os);
+}
+
+// Prints the given C string to the ostream.
+void PrintTo(const char* s, ostream* os) {
+  if (s == NULL) {
+    *os << "NULL";
+  } else {
+    *os << ImplicitCast_(s) << " pointing to ";
+    PrintCharsAsStringTo(s, strlen(s), os);
+  }
+}
+
+// MSVC compiler can be configured to define whar_t as a typedef
+// of unsigned short. Defining an overload for const wchar_t* in that case
+// would cause pointers to unsigned shorts be printed as wide strings,
+// possibly accessing more memory than intended and causing invalid
+// memory accesses. MSVC defines _NATIVE_WCHAR_T_DEFINED symbol when
+// wchar_t is implemented as a native type.
+#if !defined(_MSC_VER) || defined(_NATIVE_WCHAR_T_DEFINED)
+// Prints the given wide C string to the ostream.
+void PrintTo(const wchar_t* s, ostream* os) {
+  if (s == NULL) {
+    *os << "NULL";
+  } else {
+    *os << ImplicitCast_(s) << " pointing to ";
+    PrintCharsAsStringTo(s, wcslen(s), os);
+  }
+}
+#endif  // wchar_t is native
+
+// Prints a ::string object.
+#if GTEST_HAS_GLOBAL_STRING
+void PrintStringTo(const ::string& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+void PrintStringTo(const ::std::string& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+
+// Prints a ::wstring object.
+#if GTEST_HAS_GLOBAL_WSTRING
+void PrintWideStringTo(const ::wstring& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+#if GTEST_HAS_STD_WSTRING
+void PrintWideStringTo(const ::std::wstring& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+}  // namespace internal
+
+}  // namespace testing
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: mheule@google.com (Markus Heule)
+//
+// The Google C++ Testing Framework (Google Test)
+
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick is to
+// prevent a user from accidentally including gtest-internal-inl.h in
+// his code.
+#define GTEST_IMPLEMENTATION_ 1
+#undef GTEST_IMPLEMENTATION_
+
+namespace testing {
+
+using internal::GetUnitTestImpl;
+
+// Gets the summary of the failure message by omitting the stack trace
+// in it.
+std::string TestPartResult::ExtractSummary(const char* message) {
+  const char* const stack_trace = strstr(message, internal::kStackTraceMarker);
+  return stack_trace == NULL ? message :
+      std::string(message, stack_trace);
+}
+
+// Prints a TestPartResult object.
+std::ostream& operator<<(std::ostream& os, const TestPartResult& result) {
+  return os
+      << result.file_name() << ":" << result.line_number() << ": "
+      << (result.type() == TestPartResult::kSuccess ? "Success" :
+          result.type() == TestPartResult::kFatalFailure ? "Fatal failure" :
+          "Non-fatal failure") << ":\n"
+      << result.message() << std::endl;
+}
+
+// Appends a TestPartResult to the array.
+void TestPartResultArray::Append(const TestPartResult& result) {
+  array_.push_back(result);
+}
+
+// Returns the TestPartResult at the given index (0-based).
+const TestPartResult& TestPartResultArray::GetTestPartResult(int index) const {
+  if (index < 0 || index >= size()) {
+    printf("\nInvalid index (%d) into TestPartResultArray.\n", index);
+    internal::posix::Abort();
+  }
+
+  return array_[index];
+}
+
+// Returns the number of TestPartResult objects in the array.
+int TestPartResultArray::size() const {
+  return static_cast(array_.size());
+}
+
+namespace internal {
+
+HasNewFatalFailureHelper::HasNewFatalFailureHelper()
+    : has_new_fatal_failure_(false),
+      original_reporter_(GetUnitTestImpl()->
+                         GetTestPartResultReporterForCurrentThread()) {
+  GetUnitTestImpl()->SetTestPartResultReporterForCurrentThread(this);
+}
+
+HasNewFatalFailureHelper::~HasNewFatalFailureHelper() {
+  GetUnitTestImpl()->SetTestPartResultReporterForCurrentThread(
+      original_reporter_);
+}
+
+void HasNewFatalFailureHelper::ReportTestPartResult(
+    const TestPartResult& result) {
+  if (result.fatally_failed())
+    has_new_fatal_failure_ = true;
+  original_reporter_->ReportTestPartResult(result);
+}
+
+}  // namespace internal
+
+}  // namespace testing
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+
+namespace testing {
+namespace internal {
+
+#if GTEST_HAS_TYPED_TEST_P
+
+// Skips to the first non-space char in str. Returns an empty string if str
+// contains only whitespace characters.
+static const char* SkipSpaces(const char* str) {
+  while (IsSpace(*str))
+    str++;
+  return str;
+}
+
+// Verifies that registered_tests match the test names in
+// defined_test_names_; returns registered_tests if successful, or
+// aborts the program otherwise.
+const char* TypedTestCasePState::VerifyRegisteredTestNames(
+    const char* file, int line, const char* registered_tests) {
+  typedef ::std::set::const_iterator DefinedTestIter;
+  registered_ = true;
+
+  // Skip initial whitespace in registered_tests since some
+  // preprocessors prefix stringizied literals with whitespace.
+  registered_tests = SkipSpaces(registered_tests);
+
+  Message errors;
+  ::std::set tests;
+  for (const char* names = registered_tests; names != NULL;
+       names = SkipComma(names)) {
+    const std::string name = GetPrefixUntilComma(names);
+    if (tests.count(name) != 0) {
+      errors << "Test " << name << " is listed more than once.\n";
+      continue;
+    }
+
+    bool found = false;
+    for (DefinedTestIter it = defined_test_names_.begin();
+         it != defined_test_names_.end();
+         ++it) {
+      if (name == *it) {
+        found = true;
+        break;
+      }
+    }
+
+    if (found) {
+      tests.insert(name);
+    } else {
+      errors << "No test named " << name
+             << " can be found in this test case.\n";
+    }
+  }
+
+  for (DefinedTestIter it = defined_test_names_.begin();
+       it != defined_test_names_.end();
+       ++it) {
+    if (tests.count(*it) == 0) {
+      errors << "You forgot to list test " << *it << ".\n";
+    }
+  }
+
+  const std::string& errors_str = errors.GetString();
+  if (errors_str != "") {
+    fprintf(stderr, "%s %s", FormatFileLocation(file, line).c_str(),
+            errors_str.c_str());
+    fflush(stderr);
+    posix::Abort();
+  }
+
+  return registered_tests;
+}
+
+#endif  // GTEST_HAS_TYPED_TEST_P
+
+}  // namespace internal
+}  // namespace testing
diff --git a/lib/kokkos/tpls/gtest/gtest/gtest-test-part.h b/lib/kokkos/tpls/gtest/gtest/gtest-test-part.h
new file mode 120000
index 0000000000..48d39090f1
--- /dev/null
+++ b/lib/kokkos/tpls/gtest/gtest/gtest-test-part.h
@@ -0,0 +1 @@
+gtest.h
\ No newline at end of file
diff --git a/lib/kokkos/tpls/gtest/gtest/gtest.h b/lib/kokkos/tpls/gtest/gtest/gtest.h
new file mode 100644
index 0000000000..c74d098fa9
--- /dev/null
+++ b/lib/kokkos/tpls/gtest/gtest/gtest.h
@@ -0,0 +1,20065 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines the public API for Google Test.  It should be
+// included by any test program that uses Google Test.
+//
+// IMPORTANT NOTE: Due to limitation of the C++ language, we have to
+// leave some internal implementation details in this header file.
+// They are clearly marked by comments like this:
+//
+//   // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+//
+// Such code is NOT meant to be used by a user directly, and is subject
+// to CHANGE WITHOUT NOTICE.  Therefore DO NOT DEPEND ON IT in a user
+// program!
+//
+// Acknowledgment: Google Test borrowed the idea of automatic test
+// registration from Barthelemy Dagenais' (barthelemy@prologique.com)
+// easyUnit framework.
+
+#ifdef __GNUC__
+#pragma GCC system_header
+#endif
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_H_
+
+#include 
+#include 
+#include 
+
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan), eefacm@gmail.com (Sean Mcafee)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file declares functions and macros used internally by
+// Google Test.  They are subject to change without notice.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_INTERNAL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_INTERNAL_H_
+
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan)
+//
+// Low-level types and utilities for porting Google Test to various
+// platforms.  They are subject to change without notice.  DO NOT USE
+// THEM IN USER CODE.
+//
+// This file is fundamental to Google Test.  All other Google Test source
+// files are expected to #include this.  Therefore, it cannot #include
+// any other Google Test header.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_H_
+
+// The user can define the following macros in the build script to
+// control Google Test's behavior.  If the user doesn't define a macro
+// in this list, Google Test will define it.
+//
+//   GTEST_HAS_CLONE          - Define it to 1/0 to indicate that clone(2)
+//                              is/isn't available.
+//   GTEST_HAS_EXCEPTIONS     - Define it to 1/0 to indicate that exceptions
+//                              are enabled.
+//   GTEST_HAS_GLOBAL_STRING  - Define it to 1/0 to indicate that ::string
+//                              is/isn't available (some systems define
+//                              ::string, which is different to std::string).
+//   GTEST_HAS_GLOBAL_WSTRING - Define it to 1/0 to indicate that ::string
+//                              is/isn't available (some systems define
+//                              ::wstring, which is different to std::wstring).
+//   GTEST_HAS_POSIX_RE       - Define it to 1/0 to indicate that POSIX regular
+//                              expressions are/aren't available.
+//   GTEST_HAS_PTHREAD        - Define it to 1/0 to indicate that 
+//                              is/isn't available.
+//   GTEST_HAS_RTTI           - Define it to 1/0 to indicate that RTTI is/isn't
+//                              enabled.
+//   GTEST_HAS_STD_WSTRING    - Define it to 1/0 to indicate that
+//                              std::wstring does/doesn't work (Google Test can
+//                              be used where std::wstring is unavailable).
+//   GTEST_HAS_TR1_TUPLE      - Define it to 1/0 to indicate tr1::tuple
+//                              is/isn't available.
+//   GTEST_HAS_SEH            - Define it to 1/0 to indicate whether the
+//                              compiler supports Microsoft's "Structured
+//                              Exception Handling".
+//   GTEST_HAS_STREAM_REDIRECTION
+//                            - Define it to 1/0 to indicate whether the
+//                              platform supports I/O stream redirection using
+//                              dup() and dup2().
+//   GTEST_USE_OWN_TR1_TUPLE  - Define it to 1/0 to indicate whether Google
+//                              Test's own tr1 tuple implementation should be
+//                              used.  Unused when the user sets
+//                              GTEST_HAS_TR1_TUPLE to 0.
+//   GTEST_LANG_CXX11         - Define it to 1/0 to indicate that Google Test
+//                              is building in C++11/C++98 mode.
+//   GTEST_LINKED_AS_SHARED_LIBRARY
+//                            - Define to 1 when compiling tests that use
+//                              Google Test as a shared library (known as
+//                              DLL on Windows).
+//   GTEST_CREATE_SHARED_LIBRARY
+//                            - Define to 1 when compiling Google Test itself
+//                              as a shared library.
+
+// This header defines the following utilities:
+//
+// Macros indicating the current platform (defined to 1 if compiled on
+// the given platform; otherwise undefined):
+//   GTEST_OS_AIX      - IBM AIX
+//   GTEST_OS_CYGWIN   - Cygwin
+//   GTEST_OS_HPUX     - HP-UX
+//   GTEST_OS_LINUX    - Linux
+//     GTEST_OS_LINUX_ANDROID - Google Android
+//   GTEST_OS_MAC      - Mac OS X
+//     GTEST_OS_IOS    - iOS
+//       GTEST_OS_IOS_SIMULATOR - iOS simulator
+//   GTEST_OS_NACL     - Google Native Client (NaCl)
+//   GTEST_OS_OPENBSD  - OpenBSD
+//   GTEST_OS_QNX      - QNX
+//   GTEST_OS_SOLARIS  - Sun Solaris
+//   GTEST_OS_SYMBIAN  - Symbian
+//   GTEST_OS_WINDOWS  - Windows (Desktop, MinGW, or Mobile)
+//     GTEST_OS_WINDOWS_DESKTOP  - Windows Desktop
+//     GTEST_OS_WINDOWS_MINGW    - MinGW
+//     GTEST_OS_WINDOWS_MOBILE   - Windows Mobile
+//   GTEST_OS_ZOS      - z/OS
+//
+// Among the platforms, Cygwin, Linux, Max OS X, and Windows have the
+// most stable support.  Since core members of the Google Test project
+// don't have access to other platforms, support for them may be less
+// stable.  If you notice any problems on your platform, please notify
+// googletestframework@googlegroups.com (patches for fixing them are
+// even more welcome!).
+//
+// Note that it is possible that none of the GTEST_OS_* macros are defined.
+//
+// Macros indicating available Google Test features (defined to 1 if
+// the corresponding feature is supported; otherwise undefined):
+//   GTEST_HAS_COMBINE      - the Combine() function (for value-parameterized
+//                            tests)
+//   GTEST_HAS_DEATH_TEST   - death tests
+//   GTEST_HAS_PARAM_TEST   - value-parameterized tests
+//   GTEST_HAS_TYPED_TEST   - typed tests
+//   GTEST_HAS_TYPED_TEST_P - type-parameterized tests
+//   GTEST_USES_POSIX_RE    - enhanced POSIX regex is used. Do not confuse with
+//                            GTEST_HAS_POSIX_RE (see above) which users can
+//                            define themselves.
+//   GTEST_USES_SIMPLE_RE   - our own simple regex is used;
+//                            the above two are mutually exclusive.
+//   GTEST_CAN_COMPARE_NULL - accepts untyped NULL in EXPECT_EQ().
+//
+// Macros for basic C++ coding:
+//   GTEST_AMBIGUOUS_ELSE_BLOCKER_ - for disabling a gcc warning.
+//   GTEST_ATTRIBUTE_UNUSED_  - declares that a class' instances or a
+//                              variable don't have to be used.
+//   GTEST_DISALLOW_ASSIGN_   - disables operator=.
+//   GTEST_DISALLOW_COPY_AND_ASSIGN_ - disables copy ctor and operator=.
+//   GTEST_MUST_USE_RESULT_   - declares that a function's result must be used.
+//
+// Synchronization:
+//   Mutex, MutexLock, ThreadLocal, GetThreadCount()
+//                  - synchronization primitives.
+//   GTEST_IS_THREADSAFE - defined to 1 to indicate that the above
+//                         synchronization primitives have real implementations
+//                         and Google Test is thread-safe; or 0 otherwise.
+//
+// Template meta programming:
+//   is_pointer     - as in TR1; needed on Symbian and IBM XL C/C++ only.
+//   IteratorTraits - partial implementation of std::iterator_traits, which
+//                    is not available in libCstd when compiled with Sun C++.
+//
+// Smart pointers:
+//   scoped_ptr     - as in TR2.
+//
+// Regular expressions:
+//   RE             - a simple regular expression class using the POSIX
+//                    Extended Regular Expression syntax on UNIX-like
+//                    platforms, or a reduced regular exception syntax on
+//                    other platforms, including Windows.
+//
+// Logging:
+//   GTEST_LOG_()   - logs messages at the specified severity level.
+//   LogToStderr()  - directs all log messages to stderr.
+//   FlushInfoLog() - flushes informational log messages.
+//
+// Stdout and stderr capturing:
+//   CaptureStdout()     - starts capturing stdout.
+//   GetCapturedStdout() - stops capturing stdout and returns the captured
+//                         string.
+//   CaptureStderr()     - starts capturing stderr.
+//   GetCapturedStderr() - stops capturing stderr and returns the captured
+//                         string.
+//
+// Integer types:
+//   TypeWithSize   - maps an integer to a int type.
+//   Int32, UInt32, Int64, UInt64, TimeInMillis
+//                  - integers of known sizes.
+//   BiggestInt     - the biggest signed integer type.
+//
+// Command-line utilities:
+//   GTEST_FLAG()       - references a flag.
+//   GTEST_DECLARE_*()  - declares a flag.
+//   GTEST_DEFINE_*()   - defines a flag.
+//   GetInjectableArgvs() - returns the command line as a vector of strings.
+//
+// Environment variable utilities:
+//   GetEnv()             - gets the value of an environment variable.
+//   BoolFromGTestEnv()   - parses a bool environment variable.
+//   Int32FromGTestEnv()  - parses an Int32 environment variable.
+//   StringFromGTestEnv() - parses a string environment variable.
+
+#include    // for isspace, etc
+#include   // for ptrdiff_t
+#include 
+#include 
+#include 
+#ifndef _WIN32_WCE
+# include 
+# include 
+#endif  // !_WIN32_WCE
+
+#if defined __APPLE__
+# include 
+# include 
+#endif
+
+#include   // NOLINT
+#include   // NOLINT
+#include   // NOLINT
+
+#define GTEST_DEV_EMAIL_ "googletestframework@@googlegroups.com"
+#define GTEST_FLAG_PREFIX_ "gtest_"
+#define GTEST_FLAG_PREFIX_DASH_ "gtest-"
+#define GTEST_FLAG_PREFIX_UPPER_ "GTEST_"
+#define GTEST_NAME_ "Google Test"
+#define GTEST_PROJECT_URL_ "http://code.google.com/p/googletest/"
+
+// Determines the version of gcc that is used to compile this.
+#ifdef __GNUC__
+// 40302 means version 4.3.2.
+# define GTEST_GCC_VER_ \
+    (__GNUC__*10000 + __GNUC_MINOR__*100 + __GNUC_PATCHLEVEL__)
+#endif  // __GNUC__
+
+// Determines the platform on which Google Test is compiled.
+#ifdef __CYGWIN__
+# define GTEST_OS_CYGWIN 1
+#elif defined __SYMBIAN32__
+# define GTEST_OS_SYMBIAN 1
+#elif defined _WIN32
+# define GTEST_OS_WINDOWS 1
+# ifdef _WIN32_WCE
+#  define GTEST_OS_WINDOWS_MOBILE 1
+# elif defined(__MINGW__) || defined(__MINGW32__)
+#  define GTEST_OS_WINDOWS_MINGW 1
+# else
+#  define GTEST_OS_WINDOWS_DESKTOP 1
+# endif  // _WIN32_WCE
+#elif defined __APPLE__
+# define GTEST_OS_MAC 1
+# if TARGET_OS_IPHONE
+#  define GTEST_OS_IOS 1
+#  if TARGET_IPHONE_SIMULATOR
+#   define GTEST_OS_IOS_SIMULATOR 1
+#  endif
+# endif
+#elif defined __linux__
+# define GTEST_OS_LINUX 1
+# if defined __ANDROID__
+#  define GTEST_OS_LINUX_ANDROID 1
+# endif
+#elif defined __MVS__
+# define GTEST_OS_ZOS 1
+#elif defined(__sun) && defined(__SVR4)
+# define GTEST_OS_SOLARIS 1
+#elif defined(_AIX)
+# define GTEST_OS_AIX 1
+#elif defined(__hpux)
+# define GTEST_OS_HPUX 1
+#elif defined __native_client__
+# define GTEST_OS_NACL 1
+#elif defined __OpenBSD__
+# define GTEST_OS_OPENBSD 1
+#elif defined __QNX__
+# define GTEST_OS_QNX 1
+#endif  // __CYGWIN__
+
+#ifndef GTEST_LANG_CXX11
+// gcc and clang define __GXX_EXPERIMENTAL_CXX0X__ when
+// -std={c,gnu}++{0x,11} is passed.  The C++11 standard specifies a
+// value for __cplusplus, and recent versions of clang, gcc, and
+// probably other compilers set that too in C++11 mode.
+# if __GXX_EXPERIMENTAL_CXX0X__ || __cplusplus >= 201103L
+// Compiling in at least C++11 mode.
+#  define GTEST_LANG_CXX11 1
+# else
+#  define GTEST_LANG_CXX11 0
+# endif
+#endif
+
+// Brings in definitions for functions used in the testing::internal::posix
+// namespace (read, write, close, chdir, isatty, stat). We do not currently
+// use them on Windows Mobile.
+#if !GTEST_OS_WINDOWS
+// This assumes that non-Windows OSes provide unistd.h. For OSes where this
+// is not the case, we need to include headers that provide the functions
+// mentioned above.
+# include 
+# include 
+#elif !GTEST_OS_WINDOWS_MOBILE
+# include 
+# include 
+#endif
+
+#if GTEST_OS_LINUX_ANDROID
+// Used to define __ANDROID_API__ matching the target NDK API level.
+#  include   // NOLINT
+#endif
+
+// Defines this to true iff Google Test can use POSIX regular expressions.
+#ifndef GTEST_HAS_POSIX_RE
+# if GTEST_OS_LINUX_ANDROID
+// On Android,  is only available starting with Gingerbread.
+#  define GTEST_HAS_POSIX_RE (__ANDROID_API__ >= 9)
+# else
+#  define GTEST_HAS_POSIX_RE (!GTEST_OS_WINDOWS)
+# endif
+#endif
+
+#if GTEST_HAS_POSIX_RE
+
+// On some platforms,  needs someone to define size_t, and
+// won't compile otherwise.  We can #include it here as we already
+// included , which is guaranteed to define size_t through
+// .
+# include   // NOLINT
+
+# define GTEST_USES_POSIX_RE 1
+
+#elif GTEST_OS_WINDOWS
+
+//  is not available on Windows.  Use our own simple regex
+// implementation instead.
+# define GTEST_USES_SIMPLE_RE 1
+
+#else
+
+//  may not be available on this platform.  Use our own
+// simple regex implementation instead.
+# define GTEST_USES_SIMPLE_RE 1
+
+#endif  // GTEST_HAS_POSIX_RE
+
+#ifndef GTEST_HAS_EXCEPTIONS
+// The user didn't tell us whether exceptions are enabled, so we need
+// to figure it out.
+# if defined(_MSC_VER) || defined(__BORLANDC__)
+// MSVC's and C++Builder's implementations of the STL use the _HAS_EXCEPTIONS
+// macro to enable exceptions, so we'll do the same.
+// Assumes that exceptions are enabled by default.
+#  ifndef _HAS_EXCEPTIONS
+#   define _HAS_EXCEPTIONS 1
+#  endif  // _HAS_EXCEPTIONS
+#  define GTEST_HAS_EXCEPTIONS _HAS_EXCEPTIONS
+# elif defined(__GNUC__) && __EXCEPTIONS
+// gcc defines __EXCEPTIONS to 1 iff exceptions are enabled.
+#  define GTEST_HAS_EXCEPTIONS 1
+# elif defined(__SUNPRO_CC)
+// Sun Pro CC supports exceptions.  However, there is no compile-time way of
+// detecting whether they are enabled or not.  Therefore, we assume that
+// they are enabled unless the user tells us otherwise.
+#  define GTEST_HAS_EXCEPTIONS 1
+# elif defined(__IBMCPP__) && __EXCEPTIONS
+// xlC defines __EXCEPTIONS to 1 iff exceptions are enabled.
+#  define GTEST_HAS_EXCEPTIONS 1
+# elif defined(__HP_aCC)
+// Exception handling is in effect by default in HP aCC compiler. It has to
+// be turned of by +noeh compiler option if desired.
+#  define GTEST_HAS_EXCEPTIONS 1
+# else
+// For other compilers, we assume exceptions are disabled to be
+// conservative.
+#  define GTEST_HAS_EXCEPTIONS 0
+# endif  // defined(_MSC_VER) || defined(__BORLANDC__)
+#endif  // GTEST_HAS_EXCEPTIONS
+
+#if !defined(GTEST_HAS_STD_STRING)
+// Even though we don't use this macro any longer, we keep it in case
+// some clients still depend on it.
+# define GTEST_HAS_STD_STRING 1
+#elif !GTEST_HAS_STD_STRING
+// The user told us that ::std::string isn't available.
+# error "Google Test cannot be used where ::std::string isn't available."
+#endif  // !defined(GTEST_HAS_STD_STRING)
+
+#ifndef GTEST_HAS_GLOBAL_STRING
+// The user didn't tell us whether ::string is available, so we need
+// to figure it out.
+
+# define GTEST_HAS_GLOBAL_STRING 0
+
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+#ifndef GTEST_HAS_STD_WSTRING
+// The user didn't tell us whether ::std::wstring is available, so we need
+// to figure it out.
+// TODO(wan@google.com): uses autoconf to detect whether ::std::wstring
+//   is available.
+
+// Cygwin 1.7 and below doesn't support ::std::wstring.
+// Solaris' libc++ doesn't support it either.  Android has
+// no support for it at least as recent as Froyo (2.2).
+# define GTEST_HAS_STD_WSTRING \
+    (!(GTEST_OS_LINUX_ANDROID || GTEST_OS_CYGWIN || GTEST_OS_SOLARIS))
+
+#endif  // GTEST_HAS_STD_WSTRING
+
+#ifndef GTEST_HAS_GLOBAL_WSTRING
+// The user didn't tell us whether ::wstring is available, so we need
+// to figure it out.
+# define GTEST_HAS_GLOBAL_WSTRING \
+    (GTEST_HAS_STD_WSTRING && GTEST_HAS_GLOBAL_STRING)
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+// Determines whether RTTI is available.
+#ifndef GTEST_HAS_RTTI
+// The user didn't tell us whether RTTI is enabled, so we need to
+// figure it out.
+
+# ifdef _MSC_VER
+
+#  ifdef _CPPRTTI  // MSVC defines this macro iff RTTI is enabled.
+#   define GTEST_HAS_RTTI 1
+#  else
+#   define GTEST_HAS_RTTI 0
+#  endif
+
+// Starting with version 4.3.2, gcc defines __GXX_RTTI iff RTTI is enabled.
+# elif defined(__GNUC__) && (GTEST_GCC_VER_ >= 40302)
+
+#  ifdef __GXX_RTTI
+// When building against STLport with the Android NDK and with
+// -frtti -fno-exceptions, the build fails at link time with undefined
+// references to __cxa_bad_typeid. Note sure if STL or toolchain bug,
+// so disable RTTI when detected.
+#   if GTEST_OS_LINUX_ANDROID && defined(_STLPORT_MAJOR) && \
+       !defined(__EXCEPTIONS)
+#    define GTEST_HAS_RTTI 0
+#   else
+#    define GTEST_HAS_RTTI 1
+#   endif  // GTEST_OS_LINUX_ANDROID && __STLPORT_MAJOR && !__EXCEPTIONS
+#  else
+#   define GTEST_HAS_RTTI 0
+#  endif  // __GXX_RTTI
+
+// Clang defines __GXX_RTTI starting with version 3.0, but its manual recommends
+// using has_feature instead. has_feature(cxx_rtti) is supported since 2.7, the
+// first version with C++ support.
+# elif defined(__clang__)
+
+#  define GTEST_HAS_RTTI __has_feature(cxx_rtti)
+
+// Starting with version 9.0 IBM Visual Age defines __RTTI_ALL__ to 1 if
+// both the typeid and dynamic_cast features are present.
+# elif defined(__IBMCPP__) && (__IBMCPP__ >= 900)
+
+#  ifdef __RTTI_ALL__
+#   define GTEST_HAS_RTTI 1
+#  else
+#   define GTEST_HAS_RTTI 0
+#  endif
+
+# else
+
+// For all other compilers, we assume RTTI is enabled.
+#  define GTEST_HAS_RTTI 1
+
+# endif  // _MSC_VER
+
+#endif  // GTEST_HAS_RTTI
+
+// It's this header's responsibility to #include  when RTTI
+// is enabled.
+#if GTEST_HAS_RTTI
+# include 
+#endif
+
+// Determines whether Google Test can use the pthreads library.
+#ifndef GTEST_HAS_PTHREAD
+// The user didn't tell us explicitly, so we assume pthreads support is
+// available on Linux and Mac.
+//
+// To disable threading support in Google Test, add -DGTEST_HAS_PTHREAD=0
+// to your compiler flags.
+# define GTEST_HAS_PTHREAD (GTEST_OS_LINUX || GTEST_OS_MAC || GTEST_OS_HPUX \
+    || GTEST_OS_QNX)
+#endif  // GTEST_HAS_PTHREAD
+
+#if GTEST_HAS_PTHREAD
+// gtest-port.h guarantees to #include  when GTEST_HAS_PTHREAD is
+// true.
+# include   // NOLINT
+
+// For timespec and nanosleep, used below.
+# include   // NOLINT
+#endif
+
+// Determines whether Google Test can use tr1/tuple.  You can define
+// this macro to 0 to prevent Google Test from using tuple (any
+// feature depending on tuple with be disabled in this mode).
+#ifndef GTEST_HAS_TR1_TUPLE
+# if GTEST_OS_LINUX_ANDROID && defined(_STLPORT_MAJOR)
+// STLport, provided with the Android NDK, has neither  or .
+#  define GTEST_HAS_TR1_TUPLE 0
+# else
+// The user didn't tell us not to do it, so we assume it's OK.
+#  define GTEST_HAS_TR1_TUPLE 1
+# endif
+#endif  // GTEST_HAS_TR1_TUPLE
+
+// Determines whether Google Test's own tr1 tuple implementation
+// should be used.
+#ifndef GTEST_USE_OWN_TR1_TUPLE
+// The user didn't tell us, so we need to figure it out.
+
+// We use our own TR1 tuple if we aren't sure the user has an
+// implementation of it already.  At this time, libstdc++ 4.0.0+ and
+// MSVC 2010 are the only mainstream standard libraries that come
+// with a TR1 tuple implementation.  NVIDIA's CUDA NVCC compiler
+// pretends to be GCC by defining __GNUC__ and friends, but cannot
+// compile GCC's tuple implementation.  MSVC 2008 (9.0) provides TR1
+// tuple in a 323 MB Feature Pack download, which we cannot assume the
+// user has.  QNX's QCC compiler is a modified GCC but it doesn't
+// support TR1 tuple.  libc++ only provides std::tuple, in C++11 mode,
+// and it can be used with some compilers that define __GNUC__.
+# if (defined(__GNUC__) && !defined(__CUDACC__) && (GTEST_GCC_VER_ >= 40000) \
+      && !GTEST_OS_QNX && !defined(_LIBCPP_VERSION)) || _MSC_VER >= 1600
+#  define GTEST_ENV_HAS_TR1_TUPLE_ 1
+# endif
+
+// C++11 specifies that  provides std::tuple. Use that if gtest is used
+// in C++11 mode and libstdc++ isn't very old (binaries targeting OS X 10.6
+// can build with clang but need to use gcc4.2's libstdc++).
+# if GTEST_LANG_CXX11 && (!defined(__GLIBCXX__) || __GLIBCXX__ > 20110325)
+#  define GTEST_ENV_HAS_STD_TUPLE_ 1
+# endif
+
+# if GTEST_ENV_HAS_TR1_TUPLE_ || GTEST_ENV_HAS_STD_TUPLE_
+#  define GTEST_USE_OWN_TR1_TUPLE 0
+# else
+#  define GTEST_USE_OWN_TR1_TUPLE 1
+# endif
+
+#endif  // GTEST_USE_OWN_TR1_TUPLE
+
+// To avoid conditional compilation everywhere, we make it
+// gtest-port.h's responsibility to #include the header implementing
+// tr1/tuple.
+#if GTEST_HAS_TR1_TUPLE
+
+# if GTEST_USE_OWN_TR1_TUPLE
+// This file was GENERATED by command:
+//     pump.py gtest-tuple.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2009 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Implements a subset of TR1 tuple needed by Google Test and Google Mock.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
+
+#include   // For ::std::pair.
+
+// The compiler used in Symbian has a bug that prevents us from declaring the
+// tuple template as a friend (it complains that tuple is redefined).  This
+// hack bypasses the bug by declaring the members that should otherwise be
+// private as public.
+// Sun Studio versions < 12 also have the above bug.
+#if defined(__SYMBIAN32__) || (defined(__SUNPRO_CC) && __SUNPRO_CC < 0x590)
+# define GTEST_DECLARE_TUPLE_AS_FRIEND_ public:
+#else
+# define GTEST_DECLARE_TUPLE_AS_FRIEND_ \
+    template  friend class tuple; \
+   private:
+#endif
+
+// GTEST_n_TUPLE_(T) is the type of an n-tuple.
+#define GTEST_0_TUPLE_(T) tuple<>
+#define GTEST_1_TUPLE_(T) tuple
+#define GTEST_2_TUPLE_(T) tuple
+#define GTEST_3_TUPLE_(T) tuple
+#define GTEST_4_TUPLE_(T) tuple
+#define GTEST_5_TUPLE_(T) tuple
+#define GTEST_6_TUPLE_(T) tuple
+#define GTEST_7_TUPLE_(T) tuple
+#define GTEST_8_TUPLE_(T) tuple
+#define GTEST_9_TUPLE_(T) tuple
+#define GTEST_10_TUPLE_(T) tuple
+
+// GTEST_n_TYPENAMES_(T) declares a list of n typenames.
+#define GTEST_0_TYPENAMES_(T)
+#define GTEST_1_TYPENAMES_(T) typename T##0
+#define GTEST_2_TYPENAMES_(T) typename T##0, typename T##1
+#define GTEST_3_TYPENAMES_(T) typename T##0, typename T##1, typename T##2
+#define GTEST_4_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3
+#define GTEST_5_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4
+#define GTEST_6_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5
+#define GTEST_7_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6
+#define GTEST_8_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6, typename T##7
+#define GTEST_9_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6, \
+    typename T##7, typename T##8
+#define GTEST_10_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6, \
+    typename T##7, typename T##8, typename T##9
+
+// In theory, defining stuff in the ::std namespace is undefined
+// behavior.  We can do this as we are playing the role of a standard
+// library vendor.
+namespace std {
+namespace tr1 {
+
+template 
+class tuple;
+
+// Anything in namespace gtest_internal is Google Test's INTERNAL
+// IMPLEMENTATION DETAIL and MUST NOT BE USED DIRECTLY in user code.
+namespace gtest_internal {
+
+// ByRef::type is T if T is a reference; otherwise it's const T&.
+template 
+struct ByRef { typedef const T& type; };  // NOLINT
+template 
+struct ByRef { typedef T& type; };  // NOLINT
+
+// A handy wrapper for ByRef.
+#define GTEST_BY_REF_(T) typename ::std::tr1::gtest_internal::ByRef::type
+
+// AddRef::type is T if T is a reference; otherwise it's T&.  This
+// is the same as tr1::add_reference::type.
+template 
+struct AddRef { typedef T& type; };  // NOLINT
+template 
+struct AddRef { typedef T& type; };  // NOLINT
+
+// A handy wrapper for AddRef.
+#define GTEST_ADD_REF_(T) typename ::std::tr1::gtest_internal::AddRef::type
+
+// A helper for implementing get().
+template  class Get;
+
+// A helper for implementing tuple_element.  kIndexValid is true
+// iff k < the number of fields in tuple type T.
+template 
+struct TupleElement;
+
+template 
+struct TupleElement {
+  typedef T0 type;
+};
+
+template 
+struct TupleElement {
+  typedef T1 type;
+};
+
+template 
+struct TupleElement {
+  typedef T2 type;
+};
+
+template 
+struct TupleElement {
+  typedef T3 type;
+};
+
+template 
+struct TupleElement {
+  typedef T4 type;
+};
+
+template 
+struct TupleElement {
+  typedef T5 type;
+};
+
+template 
+struct TupleElement {
+  typedef T6 type;
+};
+
+template 
+struct TupleElement {
+  typedef T7 type;
+};
+
+template 
+struct TupleElement {
+  typedef T8 type;
+};
+
+template 
+struct TupleElement {
+  typedef T9 type;
+};
+
+}  // namespace gtest_internal
+
+template <>
+class tuple<> {
+ public:
+  tuple() {}
+  tuple(const tuple& /* t */)  {}
+  tuple& operator=(const tuple& /* t */) { return *this; }
+};
+
+template 
+class GTEST_1_TUPLE_(T) {
+ public:
+  template  friend class gtest_internal::Get;
+
+  tuple() : f0_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0) : f0_(f0) {}
+
+  tuple(const tuple& t) : f0_(t.f0_) {}
+
+  template 
+  tuple(const GTEST_1_TUPLE_(U)& t) : f0_(t.f0_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template 
+  tuple& operator=(const GTEST_1_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template 
+  tuple& CopyFrom(const GTEST_1_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    return *this;
+  }
+
+  T0 f0_;
+};
+
+template 
+class GTEST_2_TUPLE_(T) {
+ public:
+  template  friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1) : f0_(f0),
+      f1_(f1) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_) {}
+
+  template 
+  tuple(const GTEST_2_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_) {}
+  template 
+  tuple(const ::std::pair& p) : f0_(p.first), f1_(p.second) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template 
+  tuple& operator=(const GTEST_2_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+  template 
+  tuple& operator=(const ::std::pair& p) {
+    f0_ = p.first;
+    f1_ = p.second;
+    return *this;
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template 
+  tuple& CopyFrom(const GTEST_2_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+};
+
+template 
+class GTEST_3_TUPLE_(T) {
+ public:
+  template  friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2) : f0_(f0), f1_(f1), f2_(f2) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_) {}
+
+  template 
+  tuple(const GTEST_3_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template 
+  tuple& operator=(const GTEST_3_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template 
+  tuple& CopyFrom(const GTEST_3_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+};
+
+template 
+class GTEST_4_TUPLE_(T) {
+ public:
+  template  friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3) : f0_(f0), f1_(f1), f2_(f2),
+      f3_(f3) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_) {}
+
+  template 
+  tuple(const GTEST_4_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template 
+  tuple& operator=(const GTEST_4_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template 
+  tuple& CopyFrom(const GTEST_4_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+};
+
+template 
+class GTEST_5_TUPLE_(T) {
+ public:
+  template  friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3,
+      GTEST_BY_REF_(T4) f4) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_) {}
+
+  template 
+  tuple(const GTEST_5_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template 
+  tuple& operator=(const GTEST_5_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template 
+  tuple& CopyFrom(const GTEST_5_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+};
+
+template 
+class GTEST_6_TUPLE_(T) {
+ public:
+  template  friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4),
+      f5_(f5) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_) {}
+
+  template 
+  tuple(const GTEST_6_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template 
+  tuple& operator=(const GTEST_6_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template 
+  tuple& CopyFrom(const GTEST_6_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+};
+
+template 
+class GTEST_7_TUPLE_(T) {
+ public:
+  template  friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6) : f0_(f0), f1_(f1), f2_(f2),
+      f3_(f3), f4_(f4), f5_(f5), f6_(f6) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_) {}
+
+  template 
+  tuple(const GTEST_7_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template 
+  tuple& operator=(const GTEST_7_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template 
+  tuple& CopyFrom(const GTEST_7_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+};
+
+template 
+class GTEST_8_TUPLE_(T) {
+ public:
+  template  friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_(), f7_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6,
+      GTEST_BY_REF_(T7) f7) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4),
+      f5_(f5), f6_(f6), f7_(f7) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_) {}
+
+  template 
+  tuple(const GTEST_8_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template 
+  tuple& operator=(const GTEST_8_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template 
+  tuple& CopyFrom(const GTEST_8_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    f7_ = t.f7_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+  T7 f7_;
+};
+
+template 
+class GTEST_9_TUPLE_(T) {
+ public:
+  template  friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_(), f7_(), f8_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6, GTEST_BY_REF_(T7) f7,
+      GTEST_BY_REF_(T8) f8) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4),
+      f5_(f5), f6_(f6), f7_(f7), f8_(f8) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_) {}
+
+  template 
+  tuple(const GTEST_9_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template 
+  tuple& operator=(const GTEST_9_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template 
+  tuple& CopyFrom(const GTEST_9_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    f7_ = t.f7_;
+    f8_ = t.f8_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+  T7 f7_;
+  T8 f8_;
+};
+
+template 
+class tuple {
+ public:
+  template  friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_(), f7_(), f8_(),
+      f9_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6, GTEST_BY_REF_(T7) f7,
+      GTEST_BY_REF_(T8) f8, GTEST_BY_REF_(T9) f9) : f0_(f0), f1_(f1), f2_(f2),
+      f3_(f3), f4_(f4), f5_(f5), f6_(f6), f7_(f7), f8_(f8), f9_(f9) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_), f9_(t.f9_) {}
+
+  template 
+  tuple(const GTEST_10_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_),
+      f9_(t.f9_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template 
+  tuple& operator=(const GTEST_10_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template 
+  tuple& CopyFrom(const GTEST_10_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    f7_ = t.f7_;
+    f8_ = t.f8_;
+    f9_ = t.f9_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+  T7 f7_;
+  T8 f8_;
+  T9 f9_;
+};
+
+// 6.1.3.2 Tuple creation functions.
+
+// Known limitations: we don't support passing an
+// std::tr1::reference_wrapper to make_tuple().  And we don't
+// implement tie().
+
+inline tuple<> make_tuple() { return tuple<>(); }
+
+template 
+inline GTEST_1_TUPLE_(T) make_tuple(const T0& f0) {
+  return GTEST_1_TUPLE_(T)(f0);
+}
+
+template 
+inline GTEST_2_TUPLE_(T) make_tuple(const T0& f0, const T1& f1) {
+  return GTEST_2_TUPLE_(T)(f0, f1);
+}
+
+template 
+inline GTEST_3_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2) {
+  return GTEST_3_TUPLE_(T)(f0, f1, f2);
+}
+
+template 
+inline GTEST_4_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3) {
+  return GTEST_4_TUPLE_(T)(f0, f1, f2, f3);
+}
+
+template 
+inline GTEST_5_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4) {
+  return GTEST_5_TUPLE_(T)(f0, f1, f2, f3, f4);
+}
+
+template 
+inline GTEST_6_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5) {
+  return GTEST_6_TUPLE_(T)(f0, f1, f2, f3, f4, f5);
+}
+
+template 
+inline GTEST_7_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6) {
+  return GTEST_7_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6);
+}
+
+template 
+inline GTEST_8_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6, const T7& f7) {
+  return GTEST_8_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6, f7);
+}
+
+template 
+inline GTEST_9_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6, const T7& f7,
+    const T8& f8) {
+  return GTEST_9_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6, f7, f8);
+}
+
+template 
+inline GTEST_10_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6, const T7& f7,
+    const T8& f8, const T9& f9) {
+  return GTEST_10_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6, f7, f8, f9);
+}
+
+// 6.1.3.3 Tuple helper classes.
+
+template  struct tuple_size;
+
+template 
+struct tuple_size {
+  static const int value = 0;
+};
+
+template 
+struct tuple_size {
+  static const int value = 1;
+};
+
+template 
+struct tuple_size {
+  static const int value = 2;
+};
+
+template 
+struct tuple_size {
+  static const int value = 3;
+};
+
+template 
+struct tuple_size {
+  static const int value = 4;
+};
+
+template 
+struct tuple_size {
+  static const int value = 5;
+};
+
+template 
+struct tuple_size {
+  static const int value = 6;
+};
+
+template 
+struct tuple_size {
+  static const int value = 7;
+};
+
+template 
+struct tuple_size {
+  static const int value = 8;
+};
+
+template 
+struct tuple_size {
+  static const int value = 9;
+};
+
+template 
+struct tuple_size {
+  static const int value = 10;
+};
+
+template 
+struct tuple_element {
+  typedef typename gtest_internal::TupleElement<
+      k < (tuple_size::value), k, Tuple>::type type;
+};
+
+#define GTEST_TUPLE_ELEMENT_(k, Tuple) typename tuple_element::type
+
+// 6.1.3.4 Element access.
+
+namespace gtest_internal {
+
+template <>
+class Get<0> {
+ public:
+  template 
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(0, Tuple))
+  Field(Tuple& t) { return t.f0_; }  // NOLINT
+
+  template 
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(0, Tuple))
+  ConstField(const Tuple& t) { return t.f0_; }
+};
+
+template <>
+class Get<1> {
+ public:
+  template 
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(1, Tuple))
+  Field(Tuple& t) { return t.f1_; }  // NOLINT
+
+  template 
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(1, Tuple))
+  ConstField(const Tuple& t) { return t.f1_; }
+};
+
+template <>
+class Get<2> {
+ public:
+  template 
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(2, Tuple))
+  Field(Tuple& t) { return t.f2_; }  // NOLINT
+
+  template 
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(2, Tuple))
+  ConstField(const Tuple& t) { return t.f2_; }
+};
+
+template <>
+class Get<3> {
+ public:
+  template 
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(3, Tuple))
+  Field(Tuple& t) { return t.f3_; }  // NOLINT
+
+  template 
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(3, Tuple))
+  ConstField(const Tuple& t) { return t.f3_; }
+};
+
+template <>
+class Get<4> {
+ public:
+  template 
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(4, Tuple))
+  Field(Tuple& t) { return t.f4_; }  // NOLINT
+
+  template 
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(4, Tuple))
+  ConstField(const Tuple& t) { return t.f4_; }
+};
+
+template <>
+class Get<5> {
+ public:
+  template 
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(5, Tuple))
+  Field(Tuple& t) { return t.f5_; }  // NOLINT
+
+  template 
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(5, Tuple))
+  ConstField(const Tuple& t) { return t.f5_; }
+};
+
+template <>
+class Get<6> {
+ public:
+  template 
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(6, Tuple))
+  Field(Tuple& t) { return t.f6_; }  // NOLINT
+
+  template 
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(6, Tuple))
+  ConstField(const Tuple& t) { return t.f6_; }
+};
+
+template <>
+class Get<7> {
+ public:
+  template 
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(7, Tuple))
+  Field(Tuple& t) { return t.f7_; }  // NOLINT
+
+  template 
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(7, Tuple))
+  ConstField(const Tuple& t) { return t.f7_; }
+};
+
+template <>
+class Get<8> {
+ public:
+  template 
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(8, Tuple))
+  Field(Tuple& t) { return t.f8_; }  // NOLINT
+
+  template 
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(8, Tuple))
+  ConstField(const Tuple& t) { return t.f8_; }
+};
+
+template <>
+class Get<9> {
+ public:
+  template 
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(9, Tuple))
+  Field(Tuple& t) { return t.f9_; }  // NOLINT
+
+  template 
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(9, Tuple))
+  ConstField(const Tuple& t) { return t.f9_; }
+};
+
+}  // namespace gtest_internal
+
+template 
+GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(k, GTEST_10_TUPLE_(T)))
+get(GTEST_10_TUPLE_(T)& t) {
+  return gtest_internal::Get::Field(t);
+}
+
+template 
+GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(k,  GTEST_10_TUPLE_(T)))
+get(const GTEST_10_TUPLE_(T)& t) {
+  return gtest_internal::Get::ConstField(t);
+}
+
+// 6.1.3.5 Relational operators
+
+// We only implement == and !=, as we don't have a need for the rest yet.
+
+namespace gtest_internal {
+
+// SameSizeTuplePrefixComparator::Eq(t1, t2) returns true if the
+// first k fields of t1 equals the first k fields of t2.
+// SameSizeTuplePrefixComparator(k1, k2) would be a compiler error if
+// k1 != k2.
+template 
+struct SameSizeTuplePrefixComparator;
+
+template <>
+struct SameSizeTuplePrefixComparator<0, 0> {
+  template 
+  static bool Eq(const Tuple1& /* t1 */, const Tuple2& /* t2 */) {
+    return true;
+  }
+};
+
+template 
+struct SameSizeTuplePrefixComparator {
+  template 
+  static bool Eq(const Tuple1& t1, const Tuple2& t2) {
+    return SameSizeTuplePrefixComparator::Eq(t1, t2) &&
+        ::std::tr1::get(t1) == ::std::tr1::get(t2);
+  }
+};
+
+}  // namespace gtest_internal
+
+template 
+inline bool operator==(const GTEST_10_TUPLE_(T)& t,
+                       const GTEST_10_TUPLE_(U)& u) {
+  return gtest_internal::SameSizeTuplePrefixComparator<
+      tuple_size::value,
+      tuple_size::value>::Eq(t, u);
+}
+
+template 
+inline bool operator!=(const GTEST_10_TUPLE_(T)& t,
+                       const GTEST_10_TUPLE_(U)& u) { return !(t == u); }
+
+// 6.1.4 Pairs.
+// Unimplemented.
+
+}  // namespace tr1
+}  // namespace std
+
+#undef GTEST_0_TUPLE_
+#undef GTEST_1_TUPLE_
+#undef GTEST_2_TUPLE_
+#undef GTEST_3_TUPLE_
+#undef GTEST_4_TUPLE_
+#undef GTEST_5_TUPLE_
+#undef GTEST_6_TUPLE_
+#undef GTEST_7_TUPLE_
+#undef GTEST_8_TUPLE_
+#undef GTEST_9_TUPLE_
+#undef GTEST_10_TUPLE_
+
+#undef GTEST_0_TYPENAMES_
+#undef GTEST_1_TYPENAMES_
+#undef GTEST_2_TYPENAMES_
+#undef GTEST_3_TYPENAMES_
+#undef GTEST_4_TYPENAMES_
+#undef GTEST_5_TYPENAMES_
+#undef GTEST_6_TYPENAMES_
+#undef GTEST_7_TYPENAMES_
+#undef GTEST_8_TYPENAMES_
+#undef GTEST_9_TYPENAMES_
+#undef GTEST_10_TYPENAMES_
+
+#undef GTEST_DECLARE_TUPLE_AS_FRIEND_
+#undef GTEST_BY_REF_
+#undef GTEST_ADD_REF_
+#undef GTEST_TUPLE_ELEMENT_
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
+# elif GTEST_ENV_HAS_STD_TUPLE_
+#  include 
+// C++11 puts its tuple into the ::std namespace rather than
+// ::std::tr1.  gtest expects tuple to live in ::std::tr1, so put it there.
+// This causes undefined behavior, but supported compilers react in
+// the way we intend.
+namespace std {
+namespace tr1 {
+using ::std::get;
+using ::std::make_tuple;
+using ::std::tuple;
+using ::std::tuple_element;
+using ::std::tuple_size;
+}
+}
+
+# elif GTEST_OS_SYMBIAN
+
+// On Symbian, BOOST_HAS_TR1_TUPLE causes Boost's TR1 tuple library to
+// use STLport's tuple implementation, which unfortunately doesn't
+// work as the copy of STLport distributed with Symbian is incomplete.
+// By making sure BOOST_HAS_TR1_TUPLE is undefined, we force Boost to
+// use its own tuple implementation.
+#  ifdef BOOST_HAS_TR1_TUPLE
+#   undef BOOST_HAS_TR1_TUPLE
+#  endif  // BOOST_HAS_TR1_TUPLE
+
+// This prevents , which defines
+// BOOST_HAS_TR1_TUPLE, from being #included by Boost's .
+#  define BOOST_TR1_DETAIL_CONFIG_HPP_INCLUDED
+#  include 
+
+# elif defined(__GNUC__) && (GTEST_GCC_VER_ >= 40000)
+// GCC 4.0+ implements tr1/tuple in the  header.  This does
+// not conform to the TR1 spec, which requires the header to be .
+
+#  if !GTEST_HAS_RTTI && GTEST_GCC_VER_ < 40302
+// Until version 4.3.2, gcc has a bug that causes ,
+// which is #included by , to not compile when RTTI is
+// disabled.  _TR1_FUNCTIONAL is the header guard for
+// .  Hence the following #define is a hack to prevent
+//  from being included.
+#   define _TR1_FUNCTIONAL 1
+#   include 
+#   undef _TR1_FUNCTIONAL  // Allows the user to #include
+                        //  if he chooses to.
+#  else
+#   include   // NOLINT
+#  endif  // !GTEST_HAS_RTTI && GTEST_GCC_VER_ < 40302
+
+# else
+// If the compiler is not GCC 4.0+, we assume the user is using a
+// spec-conforming TR1 implementation.
+#  include   // NOLINT
+# endif  // GTEST_USE_OWN_TR1_TUPLE
+
+#endif  // GTEST_HAS_TR1_TUPLE
+
+// Determines whether clone(2) is supported.
+// Usually it will only be available on Linux, excluding
+// Linux on the Itanium architecture.
+// Also see http://linux.die.net/man/2/clone.
+#ifndef GTEST_HAS_CLONE
+// The user didn't tell us, so we need to figure it out.
+
+# if GTEST_OS_LINUX && !defined(__ia64__)
+#  if GTEST_OS_LINUX_ANDROID
+// On Android, clone() is only available on ARM starting with Gingerbread.
+#    if defined(__arm__) && __ANDROID_API__ >= 9
+#     define GTEST_HAS_CLONE 1
+#    else
+#     define GTEST_HAS_CLONE 0
+#    endif
+#  else
+#   define GTEST_HAS_CLONE 1
+#  endif
+# else
+#  define GTEST_HAS_CLONE 0
+# endif  // GTEST_OS_LINUX && !defined(__ia64__)
+
+#endif  // GTEST_HAS_CLONE
+
+// Determines whether to support stream redirection. This is used to test
+// output correctness and to implement death tests.
+#ifndef GTEST_HAS_STREAM_REDIRECTION
+// By default, we assume that stream redirection is supported on all
+// platforms except known mobile ones.
+# if GTEST_OS_WINDOWS_MOBILE || GTEST_OS_SYMBIAN
+#  define GTEST_HAS_STREAM_REDIRECTION 0
+# else
+#  define GTEST_HAS_STREAM_REDIRECTION 1
+# endif  // !GTEST_OS_WINDOWS_MOBILE && !GTEST_OS_SYMBIAN
+#endif  // GTEST_HAS_STREAM_REDIRECTION
+
+// Determines whether to support death tests.
+// Google Test does not support death tests for VC 7.1 and earlier as
+// abort() in a VC 7.1 application compiled as GUI in debug config
+// pops up a dialog window that cannot be suppressed programmatically.
+#if (GTEST_OS_LINUX || GTEST_OS_CYGWIN || GTEST_OS_SOLARIS || \
+     (GTEST_OS_MAC && !GTEST_OS_IOS) || GTEST_OS_IOS_SIMULATOR || \
+     (GTEST_OS_WINDOWS_DESKTOP && _MSC_VER >= 1400) || \
+     GTEST_OS_WINDOWS_MINGW || GTEST_OS_AIX || GTEST_OS_HPUX || \
+     GTEST_OS_OPENBSD || GTEST_OS_QNX)
+# define GTEST_HAS_DEATH_TEST 1
+# include   // NOLINT
+#endif
+
+// We don't support MSVC 7.1 with exceptions disabled now.  Therefore
+// all the compilers we care about are adequate for supporting
+// value-parameterized tests.
+#define GTEST_HAS_PARAM_TEST 1
+
+// Determines whether to support type-driven tests.
+
+// Typed tests need  and variadic macros, which GCC, VC++ 8.0,
+// Sun Pro CC, IBM Visual Age, and HP aCC support.
+#if defined(__GNUC__) || (_MSC_VER >= 1400) || defined(__SUNPRO_CC) || \
+    defined(__IBMCPP__) || defined(__HP_aCC)
+# define GTEST_HAS_TYPED_TEST 1
+# define GTEST_HAS_TYPED_TEST_P 1
+#endif
+
+// Determines whether to support Combine(). This only makes sense when
+// value-parameterized tests are enabled.  The implementation doesn't
+// work on Sun Studio since it doesn't understand templated conversion
+// operators.
+#if GTEST_HAS_PARAM_TEST && GTEST_HAS_TR1_TUPLE && !defined(__SUNPRO_CC)
+# define GTEST_HAS_COMBINE 1
+#endif
+
+// Determines whether the system compiler uses UTF-16 for encoding wide strings.
+#define GTEST_WIDE_STRING_USES_UTF16_ \
+    (GTEST_OS_WINDOWS || GTEST_OS_CYGWIN || GTEST_OS_SYMBIAN || GTEST_OS_AIX)
+
+// Determines whether test results can be streamed to a socket.
+#if GTEST_OS_LINUX
+# define GTEST_CAN_STREAM_RESULTS_ 1
+#endif
+
+// Defines some utility macros.
+
+// The GNU compiler emits a warning if nested "if" statements are followed by
+// an "else" statement and braces are not used to explicitly disambiguate the
+// "else" binding.  This leads to problems with code like:
+//
+//   if (gate)
+//     ASSERT_*(condition) << "Some message";
+//
+// The "switch (0) case 0:" idiom is used to suppress this.
+#ifdef __INTEL_COMPILER
+# define GTEST_AMBIGUOUS_ELSE_BLOCKER_
+#else
+# define GTEST_AMBIGUOUS_ELSE_BLOCKER_ switch (0) case 0: default:  // NOLINT
+#endif
+
+// Use this annotation at the end of a struct/class definition to
+// prevent the compiler from optimizing away instances that are never
+// used.  This is useful when all interesting logic happens inside the
+// c'tor and / or d'tor.  Example:
+//
+//   struct Foo {
+//     Foo() { ... }
+//   } GTEST_ATTRIBUTE_UNUSED_;
+//
+// Also use it after a variable or parameter declaration to tell the
+// compiler the variable/parameter does not have to be used.
+#if defined(__GNUC__) && !defined(COMPILER_ICC)
+# define GTEST_ATTRIBUTE_UNUSED_ __attribute__ ((unused))
+#else
+# define GTEST_ATTRIBUTE_UNUSED_
+#endif
+
+// A macro to disallow operator=
+// This should be used in the private: declarations for a class.
+#define GTEST_DISALLOW_ASSIGN_(type)\
+  void operator=(type const &)
+
+// A macro to disallow copy constructor and operator=
+// This should be used in the private: declarations for a class.
+#define GTEST_DISALLOW_COPY_AND_ASSIGN_(type)\
+  type(type const &);\
+  GTEST_DISALLOW_ASSIGN_(type)
+
+// Tell the compiler to warn about unused return values for functions declared
+// with this macro.  The macro should be used on function declarations
+// following the argument list:
+//
+//   Sprocket* AllocateSprocket() GTEST_MUST_USE_RESULT_;
+#if defined(__GNUC__) && (GTEST_GCC_VER_ >= 30400) && !defined(COMPILER_ICC)
+# define GTEST_MUST_USE_RESULT_ __attribute__ ((warn_unused_result))
+#else
+# define GTEST_MUST_USE_RESULT_
+#endif  // __GNUC__ && (GTEST_GCC_VER_ >= 30400) && !COMPILER_ICC
+
+// Determine whether the compiler supports Microsoft's Structured Exception
+// Handling.  This is supported by several Windows compilers but generally
+// does not exist on any other system.
+#ifndef GTEST_HAS_SEH
+// The user didn't tell us, so we need to figure it out.
+
+# if defined(_MSC_VER) || defined(__BORLANDC__)
+// These two compilers are known to support SEH.
+#  define GTEST_HAS_SEH 1
+# else
+// Assume no SEH.
+#  define GTEST_HAS_SEH 0
+# endif
+
+#endif  // GTEST_HAS_SEH
+
+#ifdef _MSC_VER
+
+# if GTEST_LINKED_AS_SHARED_LIBRARY
+#  define GTEST_API_ __declspec(dllimport)
+# elif GTEST_CREATE_SHARED_LIBRARY
+#  define GTEST_API_ __declspec(dllexport)
+# endif
+
+#endif  // _MSC_VER
+
+#ifndef GTEST_API_
+# define GTEST_API_
+#endif
+
+#ifdef __GNUC__
+// Ask the compiler to never inline a given function.
+# define GTEST_NO_INLINE_ __attribute__((noinline))
+#else
+# define GTEST_NO_INLINE_
+#endif
+
+// _LIBCPP_VERSION is defined by the libc++ library from the LLVM project.
+#if defined(__GLIBCXX__) || defined(_LIBCPP_VERSION)
+# define GTEST_HAS_CXXABI_H_ 1
+#else
+# define GTEST_HAS_CXXABI_H_ 0
+#endif
+
+namespace testing {
+
+class Message;
+
+namespace internal {
+
+// A secret type that Google Test users don't know about.  It has no
+// definition on purpose.  Therefore it's impossible to create a
+// Secret object, which is what we want.
+class Secret;
+
+// The GTEST_COMPILE_ASSERT_ macro can be used to verify that a compile time
+// expression is true. For example, you could use it to verify the
+// size of a static array:
+//
+//   GTEST_COMPILE_ASSERT_(ARRAYSIZE(content_type_names) == CONTENT_NUM_TYPES,
+//                         content_type_names_incorrect_size);
+//
+// or to make sure a struct is smaller than a certain size:
+//
+//   GTEST_COMPILE_ASSERT_(sizeof(foo) < 128, foo_too_large);
+//
+// The second argument to the macro is the name of the variable. If
+// the expression is false, most compilers will issue a warning/error
+// containing the name of the variable.
+
+template 
+struct CompileAssert {
+};
+
+#define GTEST_COMPILE_ASSERT_(expr, msg) \
+  typedef ::testing::internal::CompileAssert<(static_cast(expr))> \
+      msg[static_cast(expr) ? 1 : -1] GTEST_ATTRIBUTE_UNUSED_
+
+// Implementation details of GTEST_COMPILE_ASSERT_:
+//
+// - GTEST_COMPILE_ASSERT_ works by defining an array type that has -1
+//   elements (and thus is invalid) when the expression is false.
+//
+// - The simpler definition
+//
+//    #define GTEST_COMPILE_ASSERT_(expr, msg) typedef char msg[(expr) ? 1 : -1]
+//
+//   does not work, as gcc supports variable-length arrays whose sizes
+//   are determined at run-time (this is gcc's extension and not part
+//   of the C++ standard).  As a result, gcc fails to reject the
+//   following code with the simple definition:
+//
+//     int foo;
+//     GTEST_COMPILE_ASSERT_(foo, msg); // not supposed to compile as foo is
+//                                      // not a compile-time constant.
+//
+// - By using the type CompileAssert<(bool(expr))>, we ensures that
+//   expr is a compile-time constant.  (Template arguments must be
+//   determined at compile-time.)
+//
+// - The outter parentheses in CompileAssert<(bool(expr))> are necessary
+//   to work around a bug in gcc 3.4.4 and 4.0.1.  If we had written
+//
+//     CompileAssert
+//
+//   instead, these compilers will refuse to compile
+//
+//     GTEST_COMPILE_ASSERT_(5 > 0, some_message);
+//
+//   (They seem to think the ">" in "5 > 0" marks the end of the
+//   template argument list.)
+//
+// - The array size is (bool(expr) ? 1 : -1), instead of simply
+//
+//     ((expr) ? 1 : -1).
+//
+//   This is to avoid running into a bug in MS VC 7.1, which
+//   causes ((0.0) ? 1 : -1) to incorrectly evaluate to 1.
+
+// StaticAssertTypeEqHelper is used by StaticAssertTypeEq defined in gtest.h.
+//
+// This template is declared, but intentionally undefined.
+template 
+struct StaticAssertTypeEqHelper;
+
+template 
+struct StaticAssertTypeEqHelper {};
+
+#if GTEST_HAS_GLOBAL_STRING
+typedef ::string string;
+#else
+typedef ::std::string string;
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+#if GTEST_HAS_GLOBAL_WSTRING
+typedef ::wstring wstring;
+#elif GTEST_HAS_STD_WSTRING
+typedef ::std::wstring wstring;
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+// A helper for suppressing warnings on constant condition.  It just
+// returns 'condition'.
+GTEST_API_ bool IsTrue(bool condition);
+
+// Defines scoped_ptr.
+
+// This implementation of scoped_ptr is PARTIAL - it only contains
+// enough stuff to satisfy Google Test's need.
+template 
+class scoped_ptr {
+ public:
+  typedef T element_type;
+
+  explicit scoped_ptr(T* p = NULL) : ptr_(p) {}
+  ~scoped_ptr() { reset(); }
+
+  T& operator*() const { return *ptr_; }
+  T* operator->() const { return ptr_; }
+  T* get() const { return ptr_; }
+
+  T* release() {
+    T* const ptr = ptr_;
+    ptr_ = NULL;
+    return ptr;
+  }
+
+  void reset(T* p = NULL) {
+    if (p != ptr_) {
+      if (IsTrue(sizeof(T) > 0)) {  // Makes sure T is a complete type.
+        delete ptr_;
+      }
+      ptr_ = p;
+    }
+  }
+
+ private:
+  T* ptr_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(scoped_ptr);
+};
+
+// Defines RE.
+
+// A simple C++ wrapper for .  It uses the POSIX Extended
+// Regular Expression syntax.
+class GTEST_API_ RE {
+ public:
+  // A copy constructor is required by the Standard to initialize object
+  // references from r-values.
+  RE(const RE& other) { Init(other.pattern()); }
+
+  // Constructs an RE from a string.
+  RE(const ::std::string& regex) { Init(regex.c_str()); }  // NOLINT
+
+#if GTEST_HAS_GLOBAL_STRING
+
+  RE(const ::string& regex) { Init(regex.c_str()); }  // NOLINT
+
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+  RE(const char* regex) { Init(regex); }  // NOLINT
+  ~RE();
+
+  // Returns the string representation of the regex.
+  const char* pattern() const { return pattern_; }
+
+  // FullMatch(str, re) returns true iff regular expression re matches
+  // the entire str.
+  // PartialMatch(str, re) returns true iff regular expression re
+  // matches a substring of str (including str itself).
+  //
+  // TODO(wan@google.com): make FullMatch() and PartialMatch() work
+  // when str contains NUL characters.
+  static bool FullMatch(const ::std::string& str, const RE& re) {
+    return FullMatch(str.c_str(), re);
+  }
+  static bool PartialMatch(const ::std::string& str, const RE& re) {
+    return PartialMatch(str.c_str(), re);
+  }
+
+#if GTEST_HAS_GLOBAL_STRING
+
+  static bool FullMatch(const ::string& str, const RE& re) {
+    return FullMatch(str.c_str(), re);
+  }
+  static bool PartialMatch(const ::string& str, const RE& re) {
+    return PartialMatch(str.c_str(), re);
+  }
+
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+  static bool FullMatch(const char* str, const RE& re);
+  static bool PartialMatch(const char* str, const RE& re);
+
+ private:
+  void Init(const char* regex);
+
+  // We use a const char* instead of an std::string, as Google Test used to be
+  // used where std::string is not available.  TODO(wan@google.com): change to
+  // std::string.
+  const char* pattern_;
+  bool is_valid_;
+
+#if GTEST_USES_POSIX_RE
+
+  regex_t full_regex_;     // For FullMatch().
+  regex_t partial_regex_;  // For PartialMatch().
+
+#else  // GTEST_USES_SIMPLE_RE
+
+  const char* full_pattern_;  // For FullMatch();
+
+#endif
+
+  GTEST_DISALLOW_ASSIGN_(RE);
+};
+
+// Formats a source file path and a line number as they would appear
+// in an error message from the compiler used to compile this code.
+GTEST_API_ ::std::string FormatFileLocation(const char* file, int line);
+
+// Formats a file location for compiler-independent XML output.
+// Although this function is not platform dependent, we put it next to
+// FormatFileLocation in order to contrast the two functions.
+GTEST_API_ ::std::string FormatCompilerIndependentFileLocation(const char* file,
+                                                               int line);
+
+// Defines logging utilities:
+//   GTEST_LOG_(severity) - logs messages at the specified severity level. The
+//                          message itself is streamed into the macro.
+//   LogToStderr()  - directs all log messages to stderr.
+//   FlushInfoLog() - flushes informational log messages.
+
+enum GTestLogSeverity {
+  GTEST_INFO,
+  GTEST_WARNING,
+  GTEST_ERROR,
+  GTEST_FATAL
+};
+
+// Formats log entry severity, provides a stream object for streaming the
+// log message, and terminates the message with a newline when going out of
+// scope.
+class GTEST_API_ GTestLog {
+ public:
+  GTestLog(GTestLogSeverity severity, const char* file, int line);
+
+  // Flushes the buffers and, if severity is GTEST_FATAL, aborts the program.
+  ~GTestLog();
+
+  ::std::ostream& GetStream() { return ::std::cerr; }
+
+ private:
+  const GTestLogSeverity severity_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(GTestLog);
+};
+
+#define GTEST_LOG_(severity) \
+    ::testing::internal::GTestLog(::testing::internal::GTEST_##severity, \
+                                  __FILE__, __LINE__).GetStream()
+
+inline void LogToStderr() {}
+inline void FlushInfoLog() { fflush(NULL); }
+
+// INTERNAL IMPLEMENTATION - DO NOT USE.
+//
+// GTEST_CHECK_ is an all-mode assert. It aborts the program if the condition
+// is not satisfied.
+//  Synopsys:
+//    GTEST_CHECK_(boolean_condition);
+//     or
+//    GTEST_CHECK_(boolean_condition) << "Additional message";
+//
+//    This checks the condition and if the condition is not satisfied
+//    it prints message about the condition violation, including the
+//    condition itself, plus additional message streamed into it, if any,
+//    and then it aborts the program. It aborts the program irrespective of
+//    whether it is built in the debug mode or not.
+#define GTEST_CHECK_(condition) \
+    GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+    if (::testing::internal::IsTrue(condition)) \
+      ; \
+    else \
+      GTEST_LOG_(FATAL) << "Condition " #condition " failed. "
+
+// An all-mode assert to verify that the given POSIX-style function
+// call returns 0 (indicating success).  Known limitation: this
+// doesn't expand to a balanced 'if' statement, so enclose the macro
+// in {} if you need to use it as the only statement in an 'if'
+// branch.
+#define GTEST_CHECK_POSIX_SUCCESS_(posix_call) \
+  if (const int gtest_error = (posix_call)) \
+    GTEST_LOG_(FATAL) << #posix_call << "failed with error " \
+                      << gtest_error
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Use ImplicitCast_ as a safe version of static_cast for upcasting in
+// the type hierarchy (e.g. casting a Foo* to a SuperclassOfFoo* or a
+// const Foo*).  When you use ImplicitCast_, the compiler checks that
+// the cast is safe.  Such explicit ImplicitCast_s are necessary in
+// surprisingly many situations where C++ demands an exact type match
+// instead of an argument type convertable to a target type.
+//
+// The syntax for using ImplicitCast_ is the same as for static_cast:
+//
+//   ImplicitCast_(expr)
+//
+// ImplicitCast_ would have been part of the C++ standard library,
+// but the proposal was submitted too late.  It will probably make
+// its way into the language in the future.
+//
+// This relatively ugly name is intentional. It prevents clashes with
+// similar functions users may have (e.g., implicit_cast). The internal
+// namespace alone is not enough because the function can be found by ADL.
+template
+inline To ImplicitCast_(To x) { return x; }
+
+// When you upcast (that is, cast a pointer from type Foo to type
+// SuperclassOfFoo), it's fine to use ImplicitCast_<>, since upcasts
+// always succeed.  When you downcast (that is, cast a pointer from
+// type Foo to type SubclassOfFoo), static_cast<> isn't safe, because
+// how do you know the pointer is really of type SubclassOfFoo?  It
+// could be a bare Foo, or of type DifferentSubclassOfFoo.  Thus,
+// when you downcast, you should use this macro.  In debug mode, we
+// use dynamic_cast<> to double-check the downcast is legal (we die
+// if it's not).  In normal mode, we do the efficient static_cast<>
+// instead.  Thus, it's important to test in debug mode to make sure
+// the cast is legal!
+//    This is the only place in the code we should use dynamic_cast<>.
+// In particular, you SHOULDN'T be using dynamic_cast<> in order to
+// do RTTI (eg code like this:
+//    if (dynamic_cast(foo)) HandleASubclass1Object(foo);
+//    if (dynamic_cast(foo)) HandleASubclass2Object(foo);
+// You should design the code some other way not to need this.
+//
+// This relatively ugly name is intentional. It prevents clashes with
+// similar functions users may have (e.g., down_cast). The internal
+// namespace alone is not enough because the function can be found by ADL.
+template  // use like this: DownCast_(foo);
+inline To DownCast_(From* f) {  // so we only accept pointers
+  // Ensures that To is a sub-type of From *.  This test is here only
+  // for compile-time type checking, and has no overhead in an
+  // optimized build at run-time, as it will be optimized away
+  // completely.
+  if (false) {
+    const To to = NULL;
+    ::testing::internal::ImplicitCast_(to);
+  }
+
+#if GTEST_HAS_RTTI
+  // RTTI: debug mode only!
+  GTEST_CHECK_(f == NULL || dynamic_cast(f) != NULL);
+#endif
+  return static_cast(f);
+}
+
+// Downcasts the pointer of type Base to Derived.
+// Derived must be a subclass of Base. The parameter MUST
+// point to a class of type Derived, not any subclass of it.
+// When RTTI is available, the function performs a runtime
+// check to enforce this.
+template 
+Derived* CheckedDowncastToActualType(Base* base) {
+#if GTEST_HAS_RTTI
+  GTEST_CHECK_(typeid(*base) == typeid(Derived));
+  return dynamic_cast(base);  // NOLINT
+#else
+  return static_cast(base);  // Poor man's downcast.
+#endif
+}
+
+#if GTEST_HAS_STREAM_REDIRECTION
+
+// Defines the stderr capturer:
+//   CaptureStdout     - starts capturing stdout.
+//   GetCapturedStdout - stops capturing stdout and returns the captured string.
+//   CaptureStderr     - starts capturing stderr.
+//   GetCapturedStderr - stops capturing stderr and returns the captured string.
+//
+GTEST_API_ void CaptureStdout();
+GTEST_API_ std::string GetCapturedStdout();
+GTEST_API_ void CaptureStderr();
+GTEST_API_ std::string GetCapturedStderr();
+
+#endif  // GTEST_HAS_STREAM_REDIRECTION
+
+
+#if GTEST_HAS_DEATH_TEST
+
+const ::std::vector& GetInjectableArgvs();
+void SetInjectableArgvs(const ::std::vector*
+                             new_argvs);
+
+// A copy of all command line arguments.  Set by InitGoogleTest().
+extern ::std::vector g_argvs;
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+// Defines synchronization primitives.
+
+#if GTEST_HAS_PTHREAD
+
+// Sleeps for (roughly) n milli-seconds.  This function is only for
+// testing Google Test's own constructs.  Don't use it in user tests,
+// either directly or indirectly.
+inline void SleepMilliseconds(int n) {
+  const timespec time = {
+    0,                  // 0 seconds.
+    n * 1000L * 1000L,  // And n ms.
+  };
+  nanosleep(&time, NULL);
+}
+
+// Allows a controller thread to pause execution of newly created
+// threads until notified.  Instances of this class must be created
+// and destroyed in the controller thread.
+//
+// This class is only for testing Google Test's own constructs. Do not
+// use it in user tests, either directly or indirectly.
+class Notification {
+ public:
+  Notification() : notified_(false) {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_init(&mutex_, NULL));
+  }
+  ~Notification() {
+    pthread_mutex_destroy(&mutex_);
+  }
+
+  // Notifies all threads created with this notification to start. Must
+  // be called from the controller thread.
+  void Notify() {
+    pthread_mutex_lock(&mutex_);
+    notified_ = true;
+    pthread_mutex_unlock(&mutex_);
+  }
+
+  // Blocks until the controller thread notifies. Must be called from a test
+  // thread.
+  void WaitForNotification() {
+    for (;;) {
+      pthread_mutex_lock(&mutex_);
+      const bool notified = notified_;
+      pthread_mutex_unlock(&mutex_);
+      if (notified)
+        break;
+      SleepMilliseconds(10);
+    }
+  }
+
+ private:
+  pthread_mutex_t mutex_;
+  bool notified_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Notification);
+};
+
+// As a C-function, ThreadFuncWithCLinkage cannot be templated itself.
+// Consequently, it cannot select a correct instantiation of ThreadWithParam
+// in order to call its Run(). Introducing ThreadWithParamBase as a
+// non-templated base class for ThreadWithParam allows us to bypass this
+// problem.
+class ThreadWithParamBase {
+ public:
+  virtual ~ThreadWithParamBase() {}
+  virtual void Run() = 0;
+};
+
+// pthread_create() accepts a pointer to a function type with the C linkage.
+// According to the Standard (7.5/1), function types with different linkages
+// are different even if they are otherwise identical.  Some compilers (for
+// example, SunStudio) treat them as different types.  Since class methods
+// cannot be defined with C-linkage we need to define a free C-function to
+// pass into pthread_create().
+extern "C" inline void* ThreadFuncWithCLinkage(void* thread) {
+  static_cast(thread)->Run();
+  return NULL;
+}
+
+// Helper class for testing Google Test's multi-threading constructs.
+// To use it, write:
+//
+//   void ThreadFunc(int param) { /* Do things with param */ }
+//   Notification thread_can_start;
+//   ...
+//   // The thread_can_start parameter is optional; you can supply NULL.
+//   ThreadWithParam thread(&ThreadFunc, 5, &thread_can_start);
+//   thread_can_start.Notify();
+//
+// These classes are only for testing Google Test's own constructs. Do
+// not use them in user tests, either directly or indirectly.
+template 
+class ThreadWithParam : public ThreadWithParamBase {
+ public:
+  typedef void (*UserThreadFunc)(T);
+
+  ThreadWithParam(
+      UserThreadFunc func, T param, Notification* thread_can_start)
+      : func_(func),
+        param_(param),
+        thread_can_start_(thread_can_start),
+        finished_(false) {
+    ThreadWithParamBase* const base = this;
+    // The thread can be created only after all fields except thread_
+    // have been initialized.
+    GTEST_CHECK_POSIX_SUCCESS_(
+        pthread_create(&thread_, 0, &ThreadFuncWithCLinkage, base));
+  }
+  ~ThreadWithParam() { Join(); }
+
+  void Join() {
+    if (!finished_) {
+      GTEST_CHECK_POSIX_SUCCESS_(pthread_join(thread_, 0));
+      finished_ = true;
+    }
+  }
+
+  virtual void Run() {
+    if (thread_can_start_ != NULL)
+      thread_can_start_->WaitForNotification();
+    func_(param_);
+  }
+
+ private:
+  const UserThreadFunc func_;  // User-supplied thread function.
+  const T param_;  // User-supplied parameter to the thread function.
+  // When non-NULL, used to block execution until the controller thread
+  // notifies.
+  Notification* const thread_can_start_;
+  bool finished_;  // true iff we know that the thread function has finished.
+  pthread_t thread_;  // The native thread object.
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ThreadWithParam);
+};
+
+// MutexBase and Mutex implement mutex on pthreads-based platforms. They
+// are used in conjunction with class MutexLock:
+//
+//   Mutex mutex;
+//   ...
+//   MutexLock lock(&mutex);  // Acquires the mutex and releases it at the end
+//                            // of the current scope.
+//
+// MutexBase implements behavior for both statically and dynamically
+// allocated mutexes.  Do not use MutexBase directly.  Instead, write
+// the following to define a static mutex:
+//
+//   GTEST_DEFINE_STATIC_MUTEX_(g_some_mutex);
+//
+// You can forward declare a static mutex like this:
+//
+//   GTEST_DECLARE_STATIC_MUTEX_(g_some_mutex);
+//
+// To create a dynamic mutex, just define an object of type Mutex.
+class MutexBase {
+ public:
+  // Acquires this mutex.
+  void Lock() {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_lock(&mutex_));
+    owner_ = pthread_self();
+    has_owner_ = true;
+  }
+
+  // Releases this mutex.
+  void Unlock() {
+    // Since the lock is being released the owner_ field should no longer be
+    // considered valid. We don't protect writing to has_owner_ here, as it's
+    // the caller's responsibility to ensure that the current thread holds the
+    // mutex when this is called.
+    has_owner_ = false;
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_unlock(&mutex_));
+  }
+
+  // Does nothing if the current thread holds the mutex. Otherwise, crashes
+  // with high probability.
+  void AssertHeld() const {
+    GTEST_CHECK_(has_owner_ && pthread_equal(owner_, pthread_self()))
+        << "The current thread is not holding the mutex @" << this;
+  }
+
+  // A static mutex may be used before main() is entered.  It may even
+  // be used before the dynamic initialization stage.  Therefore we
+  // must be able to initialize a static mutex object at link time.
+  // This means MutexBase has to be a POD and its member variables
+  // have to be public.
+ public:
+  pthread_mutex_t mutex_;  // The underlying pthread mutex.
+  // has_owner_ indicates whether the owner_ field below contains a valid thread
+  // ID and is therefore safe to inspect (e.g., to use in pthread_equal()). All
+  // accesses to the owner_ field should be protected by a check of this field.
+  // An alternative might be to memset() owner_ to all zeros, but there's no
+  // guarantee that a zero'd pthread_t is necessarily invalid or even different
+  // from pthread_self().
+  bool has_owner_;
+  pthread_t owner_;  // The thread holding the mutex.
+};
+
+// Forward-declares a static mutex.
+# define GTEST_DECLARE_STATIC_MUTEX_(mutex) \
+    extern ::testing::internal::MutexBase mutex
+
+// Defines and statically (i.e. at link time) initializes a static mutex.
+// The initialization list here does not explicitly initialize each field,
+// instead relying on default initialization for the unspecified fields. In
+// particular, the owner_ field (a pthread_t) is not explicitly initialized.
+// This allows initialization to work whether pthread_t is a scalar or struct.
+// The flag -Wmissing-field-initializers must not be specified for this to work.
+# define GTEST_DEFINE_STATIC_MUTEX_(mutex) \
+    ::testing::internal::MutexBase mutex = { PTHREAD_MUTEX_INITIALIZER, false }
+
+// The Mutex class can only be used for mutexes created at runtime. It
+// shares its API with MutexBase otherwise.
+class Mutex : public MutexBase {
+ public:
+  Mutex() {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_init(&mutex_, NULL));
+    has_owner_ = false;
+  }
+  ~Mutex() {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_destroy(&mutex_));
+  }
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Mutex);
+};
+
+// We cannot name this class MutexLock as the ctor declaration would
+// conflict with a macro named MutexLock, which is defined on some
+// platforms.  Hence the typedef trick below.
+class GTestMutexLock {
+ public:
+  explicit GTestMutexLock(MutexBase* mutex)
+      : mutex_(mutex) { mutex_->Lock(); }
+
+  ~GTestMutexLock() { mutex_->Unlock(); }
+
+ private:
+  MutexBase* const mutex_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(GTestMutexLock);
+};
+
+typedef GTestMutexLock MutexLock;
+
+// Helpers for ThreadLocal.
+
+// pthread_key_create() requires DeleteThreadLocalValue() to have
+// C-linkage.  Therefore it cannot be templatized to access
+// ThreadLocal.  Hence the need for class
+// ThreadLocalValueHolderBase.
+class ThreadLocalValueHolderBase {
+ public:
+  virtual ~ThreadLocalValueHolderBase() {}
+};
+
+// Called by pthread to delete thread-local data stored by
+// pthread_setspecific().
+extern "C" inline void DeleteThreadLocalValue(void* value_holder) {
+  delete static_cast(value_holder);
+}
+
+// Implements thread-local storage on pthreads-based systems.
+//
+//   // Thread 1
+//   ThreadLocal tl(100);  // 100 is the default value for each thread.
+//
+//   // Thread 2
+//   tl.set(150);  // Changes the value for thread 2 only.
+//   EXPECT_EQ(150, tl.get());
+//
+//   // Thread 1
+//   EXPECT_EQ(100, tl.get());  // In thread 1, tl has the original value.
+//   tl.set(200);
+//   EXPECT_EQ(200, tl.get());
+//
+// The template type argument T must have a public copy constructor.
+// In addition, the default ThreadLocal constructor requires T to have
+// a public default constructor.
+//
+// An object managed for a thread by a ThreadLocal instance is deleted
+// when the thread exits.  Or, if the ThreadLocal instance dies in
+// that thread, when the ThreadLocal dies.  It's the user's
+// responsibility to ensure that all other threads using a ThreadLocal
+// have exited when it dies, or the per-thread objects for those
+// threads will not be deleted.
+//
+// Google Test only uses global ThreadLocal objects.  That means they
+// will die after main() has returned.  Therefore, no per-thread
+// object managed by Google Test will be leaked as long as all threads
+// using Google Test have exited when main() returns.
+template 
+class ThreadLocal {
+ public:
+  ThreadLocal() : key_(CreateKey()),
+                  default_() {}
+  explicit ThreadLocal(const T& value) : key_(CreateKey()),
+                                         default_(value) {}
+
+  ~ThreadLocal() {
+    // Destroys the managed object for the current thread, if any.
+    DeleteThreadLocalValue(pthread_getspecific(key_));
+
+    // Releases resources associated with the key.  This will *not*
+    // delete managed objects for other threads.
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_key_delete(key_));
+  }
+
+  T* pointer() { return GetOrCreateValue(); }
+  const T* pointer() const { return GetOrCreateValue(); }
+  const T& get() const { return *pointer(); }
+  void set(const T& value) { *pointer() = value; }
+
+ private:
+  // Holds a value of type T.
+  class ValueHolder : public ThreadLocalValueHolderBase {
+   public:
+    explicit ValueHolder(const T& value) : value_(value) {}
+
+    T* pointer() { return &value_; }
+
+   private:
+    T value_;
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(ValueHolder);
+  };
+
+  static pthread_key_t CreateKey() {
+    pthread_key_t key;
+    // When a thread exits, DeleteThreadLocalValue() will be called on
+    // the object managed for that thread.
+    GTEST_CHECK_POSIX_SUCCESS_(
+        pthread_key_create(&key, &DeleteThreadLocalValue));
+    return key;
+  }
+
+  T* GetOrCreateValue() const {
+    ThreadLocalValueHolderBase* const holder =
+        static_cast(pthread_getspecific(key_));
+    if (holder != NULL) {
+      return CheckedDowncastToActualType(holder)->pointer();
+    }
+
+    ValueHolder* const new_holder = new ValueHolder(default_);
+    ThreadLocalValueHolderBase* const holder_base = new_holder;
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_setspecific(key_, holder_base));
+    return new_holder->pointer();
+  }
+
+  // A key pthreads uses for looking up per-thread values.
+  const pthread_key_t key_;
+  const T default_;  // The default value for each thread.
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ThreadLocal);
+};
+
+# define GTEST_IS_THREADSAFE 1
+
+#else  // GTEST_HAS_PTHREAD
+
+// A dummy implementation of synchronization primitives (mutex, lock,
+// and thread-local variable).  Necessary for compiling Google Test where
+// mutex is not supported - using Google Test in multiple threads is not
+// supported on such platforms.
+
+class Mutex {
+ public:
+  Mutex() {}
+  void Lock() {}
+  void Unlock() {}
+  void AssertHeld() const {}
+};
+
+# define GTEST_DECLARE_STATIC_MUTEX_(mutex) \
+  extern ::testing::internal::Mutex mutex
+
+# define GTEST_DEFINE_STATIC_MUTEX_(mutex) ::testing::internal::Mutex mutex
+
+class GTestMutexLock {
+ public:
+  explicit GTestMutexLock(Mutex*) {}  // NOLINT
+};
+
+typedef GTestMutexLock MutexLock;
+
+template 
+class ThreadLocal {
+ public:
+  ThreadLocal() : value_() {}
+  explicit ThreadLocal(const T& value) : value_(value) {}
+  T* pointer() { return &value_; }
+  const T* pointer() const { return &value_; }
+  const T& get() const { return value_; }
+  void set(const T& value) { value_ = value; }
+ private:
+  T value_;
+};
+
+// The above synchronization primitives have dummy implementations.
+// Therefore Google Test is not thread-safe.
+# define GTEST_IS_THREADSAFE 0
+
+#endif  // GTEST_HAS_PTHREAD
+
+// Returns the number of threads running in the process, or 0 to indicate that
+// we cannot detect it.
+GTEST_API_ size_t GetThreadCount();
+
+// Passing non-POD classes through ellipsis (...) crashes the ARM
+// compiler and generates a warning in Sun Studio.  The Nokia Symbian
+// and the IBM XL C/C++ compiler try to instantiate a copy constructor
+// for objects passed through ellipsis (...), failing for uncopyable
+// objects.  We define this to ensure that only POD is passed through
+// ellipsis on these systems.
+#if defined(__SYMBIAN32__) || defined(__IBMCPP__) || defined(__SUNPRO_CC)
+// We lose support for NULL detection where the compiler doesn't like
+// passing non-POD classes through ellipsis (...).
+# define GTEST_ELLIPSIS_NEEDS_POD_ 1
+#else
+# define GTEST_CAN_COMPARE_NULL 1
+#endif
+
+// The Nokia Symbian and IBM XL C/C++ compilers cannot decide between
+// const T& and const T* in a function template.  These compilers
+// _can_ decide between class template specializations for T and T*,
+// so a tr1::type_traits-like is_pointer works.
+#if defined(__SYMBIAN32__) || defined(__IBMCPP__)
+# define GTEST_NEEDS_IS_POINTER_ 1
+#endif
+
+template 
+struct bool_constant {
+  typedef bool_constant type;
+  static const bool value = bool_value;
+};
+template  const bool bool_constant::value;
+
+typedef bool_constant false_type;
+typedef bool_constant true_type;
+
+template 
+struct is_pointer : public false_type {};
+
+template 
+struct is_pointer : public true_type {};
+
+template 
+struct IteratorTraits {
+  typedef typename Iterator::value_type value_type;
+};
+
+template 
+struct IteratorTraits {
+  typedef T value_type;
+};
+
+template 
+struct IteratorTraits {
+  typedef T value_type;
+};
+
+#if GTEST_OS_WINDOWS
+# define GTEST_PATH_SEP_ "\\"
+# define GTEST_HAS_ALT_PATH_SEP_ 1
+// The biggest signed integer type the compiler supports.
+typedef __int64 BiggestInt;
+#else
+# define GTEST_PATH_SEP_ "/"
+# define GTEST_HAS_ALT_PATH_SEP_ 0
+typedef long long BiggestInt;  // NOLINT
+#endif  // GTEST_OS_WINDOWS
+
+// Utilities for char.
+
+// isspace(int ch) and friends accept an unsigned char or EOF.  char
+// may be signed, depending on the compiler (or compiler flags).
+// Therefore we need to cast a char to unsigned char before calling
+// isspace(), etc.
+
+inline bool IsAlpha(char ch) {
+  return isalpha(static_cast(ch)) != 0;
+}
+inline bool IsAlNum(char ch) {
+  return isalnum(static_cast(ch)) != 0;
+}
+inline bool IsDigit(char ch) {
+  return isdigit(static_cast(ch)) != 0;
+}
+inline bool IsLower(char ch) {
+  return islower(static_cast(ch)) != 0;
+}
+inline bool IsSpace(char ch) {
+  return isspace(static_cast(ch)) != 0;
+}
+inline bool IsUpper(char ch) {
+  return isupper(static_cast(ch)) != 0;
+}
+inline bool IsXDigit(char ch) {
+  return isxdigit(static_cast(ch)) != 0;
+}
+inline bool IsXDigit(wchar_t ch) {
+  const unsigned char low_byte = static_cast(ch);
+  return ch == low_byte && isxdigit(low_byte) != 0;
+}
+
+inline char ToLower(char ch) {
+  return static_cast(tolower(static_cast(ch)));
+}
+inline char ToUpper(char ch) {
+  return static_cast(toupper(static_cast(ch)));
+}
+
+// The testing::internal::posix namespace holds wrappers for common
+// POSIX functions.  These wrappers hide the differences between
+// Windows/MSVC and POSIX systems.  Since some compilers define these
+// standard functions as macros, the wrapper cannot have the same name
+// as the wrapped function.
+
+namespace posix {
+
+// Functions with a different name on Windows.
+
+#if GTEST_OS_WINDOWS
+
+typedef struct _stat StatStruct;
+
+# ifdef __BORLANDC__
+inline int IsATTY(int fd) { return isatty(fd); }
+inline int StrCaseCmp(const char* s1, const char* s2) {
+  return stricmp(s1, s2);
+}
+inline char* StrDup(const char* src) { return strdup(src); }
+# else  // !__BORLANDC__
+#  if GTEST_OS_WINDOWS_MOBILE
+inline int IsATTY(int /* fd */) { return 0; }
+#  else
+inline int IsATTY(int fd) { return _isatty(fd); }
+#  endif  // GTEST_OS_WINDOWS_MOBILE
+inline int StrCaseCmp(const char* s1, const char* s2) {
+  return _stricmp(s1, s2);
+}
+inline char* StrDup(const char* src) { return _strdup(src); }
+# endif  // __BORLANDC__
+
+# if GTEST_OS_WINDOWS_MOBILE
+inline int FileNo(FILE* file) { return reinterpret_cast(_fileno(file)); }
+// Stat(), RmDir(), and IsDir() are not needed on Windows CE at this
+// time and thus not defined there.
+# else
+inline int FileNo(FILE* file) { return _fileno(file); }
+inline int Stat(const char* path, StatStruct* buf) { return _stat(path, buf); }
+inline int RmDir(const char* dir) { return _rmdir(dir); }
+inline bool IsDir(const StatStruct& st) {
+  return (_S_IFDIR & st.st_mode) != 0;
+}
+# endif  // GTEST_OS_WINDOWS_MOBILE
+
+#else
+
+typedef struct stat StatStruct;
+
+inline int FileNo(FILE* file) { return fileno(file); }
+inline int IsATTY(int fd) { return isatty(fd); }
+inline int Stat(const char* path, StatStruct* buf) { return stat(path, buf); }
+inline int StrCaseCmp(const char* s1, const char* s2) {
+  return strcasecmp(s1, s2);
+}
+inline char* StrDup(const char* src) { return strdup(src); }
+inline int RmDir(const char* dir) { return rmdir(dir); }
+inline bool IsDir(const StatStruct& st) { return S_ISDIR(st.st_mode); }
+
+#endif  // GTEST_OS_WINDOWS
+
+// Functions deprecated by MSVC 8.0.
+
+#ifdef _MSC_VER
+// Temporarily disable warning 4996 (deprecated function).
+# pragma warning(push)
+# pragma warning(disable:4996)
+#endif
+
+inline const char* StrNCpy(char* dest, const char* src, size_t n) {
+  return strncpy(dest, src, n);
+}
+
+// ChDir(), FReopen(), FDOpen(), Read(), Write(), Close(), and
+// StrError() aren't needed on Windows CE at this time and thus not
+// defined there.
+
+#if !GTEST_OS_WINDOWS_MOBILE
+inline int ChDir(const char* dir) { return chdir(dir); }
+#endif
+inline FILE* FOpen(const char* path, const char* mode) {
+  return fopen(path, mode);
+}
+#if !GTEST_OS_WINDOWS_MOBILE
+inline FILE *FReopen(const char* path, const char* mode, FILE* stream) {
+  return freopen(path, mode, stream);
+}
+inline FILE* FDOpen(int fd, const char* mode) { return fdopen(fd, mode); }
+#endif
+inline int FClose(FILE* fp) { return fclose(fp); }
+#if !GTEST_OS_WINDOWS_MOBILE
+inline int Read(int fd, void* buf, unsigned int count) {
+  return static_cast(read(fd, buf, count));
+}
+inline int Write(int fd, const void* buf, unsigned int count) {
+  return static_cast(write(fd, buf, count));
+}
+inline int Close(int fd) { return close(fd); }
+inline const char* StrError(int errnum) { return strerror(errnum); }
+#endif
+inline const char* GetEnv(const char* name) {
+#if GTEST_OS_WINDOWS_MOBILE
+  // We are on Windows CE, which has no environment variables.
+  return NULL;
+#elif defined(__BORLANDC__) || defined(__SunOS_5_8) || defined(__SunOS_5_9)
+  // Environment variables which we programmatically clear will be set to the
+  // empty string rather than unset (NULL).  Handle that case.
+  const char* const env = getenv(name);
+  return (env != NULL && env[0] != '\0') ? env : NULL;
+#else
+  return getenv(name);
+#endif
+}
+
+#ifdef _MSC_VER
+# pragma warning(pop)  // Restores the warning state.
+#endif
+
+#if GTEST_OS_WINDOWS_MOBILE
+// Windows CE has no C library. The abort() function is used in
+// several places in Google Test. This implementation provides a reasonable
+// imitation of standard behaviour.
+void Abort();
+#else
+inline void Abort() { abort(); }
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+}  // namespace posix
+
+// MSVC "deprecates" snprintf and issues warnings wherever it is used.  In
+// order to avoid these warnings, we need to use _snprintf or _snprintf_s on
+// MSVC-based platforms.  We map the GTEST_SNPRINTF_ macro to the appropriate
+// function in order to achieve that.  We use macro definition here because
+// snprintf is a variadic function.
+#if _MSC_VER >= 1400 && !GTEST_OS_WINDOWS_MOBILE
+// MSVC 2005 and above support variadic macros.
+# define GTEST_SNPRINTF_(buffer, size, format, ...) \
+     _snprintf_s(buffer, size, size, format, __VA_ARGS__)
+#elif defined(_MSC_VER)
+// Windows CE does not define _snprintf_s and MSVC prior to 2005 doesn't
+// complain about _snprintf.
+# define GTEST_SNPRINTF_ _snprintf
+#else
+# define GTEST_SNPRINTF_ snprintf
+#endif
+
+// The maximum number a BiggestInt can represent.  This definition
+// works no matter BiggestInt is represented in one's complement or
+// two's complement.
+//
+// We cannot rely on numeric_limits in STL, as __int64 and long long
+// are not part of standard C++ and numeric_limits doesn't need to be
+// defined for them.
+const BiggestInt kMaxBiggestInt =
+    ~(static_cast(1) << (8*sizeof(BiggestInt) - 1));
+
+// This template class serves as a compile-time function from size to
+// type.  It maps a size in bytes to a primitive type with that
+// size. e.g.
+//
+//   TypeWithSize<4>::UInt
+//
+// is typedef-ed to be unsigned int (unsigned integer made up of 4
+// bytes).
+//
+// Such functionality should belong to STL, but I cannot find it
+// there.
+//
+// Google Test uses this class in the implementation of floating-point
+// comparison.
+//
+// For now it only handles UInt (unsigned int) as that's all Google Test
+// needs.  Other types can be easily added in the future if need
+// arises.
+template 
+class TypeWithSize {
+ public:
+  // This prevents the user from using TypeWithSize with incorrect
+  // values of N.
+  typedef void UInt;
+};
+
+// The specialization for size 4.
+template <>
+class TypeWithSize<4> {
+ public:
+  // unsigned int has size 4 in both gcc and MSVC.
+  //
+  // As base/basictypes.h doesn't compile on Windows, we cannot use
+  // uint32, uint64, and etc here.
+  typedef int Int;
+  typedef unsigned int UInt;
+};
+
+// The specialization for size 8.
+template <>
+class TypeWithSize<8> {
+ public:
+#if GTEST_OS_WINDOWS
+  typedef __int64 Int;
+  typedef unsigned __int64 UInt;
+#else
+  typedef long long Int;  // NOLINT
+  typedef unsigned long long UInt;  // NOLINT
+#endif  // GTEST_OS_WINDOWS
+};
+
+// Integer types of known sizes.
+typedef TypeWithSize<4>::Int Int32;
+typedef TypeWithSize<4>::UInt UInt32;
+typedef TypeWithSize<8>::Int Int64;
+typedef TypeWithSize<8>::UInt UInt64;
+typedef TypeWithSize<8>::Int TimeInMillis;  // Represents time in milliseconds.
+
+// Utilities for command line flags and environment variables.
+
+// Macro for referencing flags.
+#define GTEST_FLAG(name) FLAGS_gtest_##name
+
+// Macros for declaring flags.
+#define GTEST_DECLARE_bool_(name) GTEST_API_ extern bool GTEST_FLAG(name)
+#define GTEST_DECLARE_int32_(name) \
+    GTEST_API_ extern ::testing::internal::Int32 GTEST_FLAG(name)
+#define GTEST_DECLARE_string_(name) \
+    GTEST_API_ extern ::std::string GTEST_FLAG(name)
+
+// Macros for defining flags.
+#define GTEST_DEFINE_bool_(name, default_val, doc) \
+    GTEST_API_ bool GTEST_FLAG(name) = (default_val)
+#define GTEST_DEFINE_int32_(name, default_val, doc) \
+    GTEST_API_ ::testing::internal::Int32 GTEST_FLAG(name) = (default_val)
+#define GTEST_DEFINE_string_(name, default_val, doc) \
+    GTEST_API_ ::std::string GTEST_FLAG(name) = (default_val)
+
+// Thread annotations
+#define GTEST_EXCLUSIVE_LOCK_REQUIRED_(locks)
+#define GTEST_LOCK_EXCLUDED_(locks)
+
+// Parses 'str' for a 32-bit signed integer.  If successful, writes the result
+// to *value and returns true; otherwise leaves *value unchanged and returns
+// false.
+// TODO(chandlerc): Find a better way to refactor flag and environment parsing
+// out of both gtest-port.cc and gtest.cc to avoid exporting this utility
+// function.
+bool ParseInt32(const Message& src_text, const char* str, Int32* value);
+
+// Parses a bool/Int32/string from the environment variable
+// corresponding to the given Google Test flag.
+bool BoolFromGTestEnv(const char* flag, bool default_val);
+GTEST_API_ Int32 Int32FromGTestEnv(const char* flag, Int32 default_val);
+const char* StringFromGTestEnv(const char* flag, const char* default_val);
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_H_
+
+#if GTEST_OS_LINUX
+# include 
+# include 
+# include 
+# include 
+#endif  // GTEST_OS_LINUX
+
+#if GTEST_HAS_EXCEPTIONS
+# include 
+#endif
+
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines the Message class.
+//
+// IMPORTANT NOTE: Due to limitation of the C++ language, we have to
+// leave some internal implementation details in this header file.
+// They are clearly marked by comments like this:
+//
+//   // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+//
+// Such code is NOT meant to be used by a user directly, and is subject
+// to CHANGE WITHOUT NOTICE.  Therefore DO NOT DEPEND ON IT in a user
+// program!
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_MESSAGE_H_
+#define GTEST_INCLUDE_GTEST_GTEST_MESSAGE_H_
+
+#include 
+
+
+// Ensures that there is at least one operator<< in the global namespace.
+// See Message& operator<<(...) below for why.
+void operator<<(const testing::internal::Secret&, int);
+
+namespace testing {
+
+// The Message class works like an ostream repeater.
+//
+// Typical usage:
+//
+//   1. You stream a bunch of values to a Message object.
+//      It will remember the text in a stringstream.
+//   2. Then you stream the Message object to an ostream.
+//      This causes the text in the Message to be streamed
+//      to the ostream.
+//
+// For example;
+//
+//   testing::Message foo;
+//   foo << 1 << " != " << 2;
+//   std::cout << foo;
+//
+// will print "1 != 2".
+//
+// Message is not intended to be inherited from.  In particular, its
+// destructor is not virtual.
+//
+// Note that stringstream behaves differently in gcc and in MSVC.  You
+// can stream a NULL char pointer to it in the former, but not in the
+// latter (it causes an access violation if you do).  The Message
+// class hides this difference by treating a NULL char pointer as
+// "(null)".
+class GTEST_API_ Message {
+ private:
+  // The type of basic IO manipulators (endl, ends, and flush) for
+  // narrow streams.
+  typedef std::ostream& (*BasicNarrowIoManip)(std::ostream&);
+
+ public:
+  // Constructs an empty Message.
+  Message();
+
+  // Copy constructor.
+  Message(const Message& msg) : ss_(new ::std::stringstream) {  // NOLINT
+    *ss_ << msg.GetString();
+  }
+
+  // Constructs a Message from a C-string.
+  explicit Message(const char* str) : ss_(new ::std::stringstream) {
+    *ss_ << str;
+  }
+
+#if GTEST_OS_SYMBIAN
+  // Streams a value (either a pointer or not) to this object.
+  template 
+  inline Message& operator <<(const T& value) {
+    StreamHelper(typename internal::is_pointer::type(), value);
+    return *this;
+  }
+#else
+  // Streams a non-pointer value to this object.
+  template 
+  inline Message& operator <<(const T& val) {
+    // Some libraries overload << for STL containers.  These
+    // overloads are defined in the global namespace instead of ::std.
+    //
+    // C++'s symbol lookup rule (i.e. Koenig lookup) says that these
+    // overloads are visible in either the std namespace or the global
+    // namespace, but not other namespaces, including the testing
+    // namespace which Google Test's Message class is in.
+    //
+    // To allow STL containers (and other types that has a << operator
+    // defined in the global namespace) to be used in Google Test
+    // assertions, testing::Message must access the custom << operator
+    // from the global namespace.  With this using declaration,
+    // overloads of << defined in the global namespace and those
+    // visible via Koenig lookup are both exposed in this function.
+    using ::operator <<;
+    *ss_ << val;
+    return *this;
+  }
+
+  // Streams a pointer value to this object.
+  //
+  // This function is an overload of the previous one.  When you
+  // stream a pointer to a Message, this definition will be used as it
+  // is more specialized.  (The C++ Standard, section
+  // [temp.func.order].)  If you stream a non-pointer, then the
+  // previous definition will be used.
+  //
+  // The reason for this overload is that streaming a NULL pointer to
+  // ostream is undefined behavior.  Depending on the compiler, you
+  // may get "0", "(nil)", "(null)", or an access violation.  To
+  // ensure consistent result across compilers, we always treat NULL
+  // as "(null)".
+  template 
+  inline Message& operator <<(T* const& pointer) {  // NOLINT
+    if (pointer == NULL) {
+      *ss_ << "(null)";
+    } else {
+      *ss_ << pointer;
+    }
+    return *this;
+  }
+#endif  // GTEST_OS_SYMBIAN
+
+  // Since the basic IO manipulators are overloaded for both narrow
+  // and wide streams, we have to provide this specialized definition
+  // of operator <<, even though its body is the same as the
+  // templatized version above.  Without this definition, streaming
+  // endl or other basic IO manipulators to Message will confuse the
+  // compiler.
+  Message& operator <<(BasicNarrowIoManip val) {
+    *ss_ << val;
+    return *this;
+  }
+
+  // Instead of 1/0, we want to see true/false for bool values.
+  Message& operator <<(bool b) {
+    return *this << (b ? "true" : "false");
+  }
+
+  // These two overloads allow streaming a wide C string to a Message
+  // using the UTF-8 encoding.
+  Message& operator <<(const wchar_t* wide_c_str);
+  Message& operator <<(wchar_t* wide_c_str);
+
+#if GTEST_HAS_STD_WSTRING
+  // Converts the given wide string to a narrow string using the UTF-8
+  // encoding, and streams the result to this Message object.
+  Message& operator <<(const ::std::wstring& wstr);
+#endif  // GTEST_HAS_STD_WSTRING
+
+#if GTEST_HAS_GLOBAL_WSTRING
+  // Converts the given wide string to a narrow string using the UTF-8
+  // encoding, and streams the result to this Message object.
+  Message& operator <<(const ::wstring& wstr);
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+  // Gets the text streamed to this object so far as an std::string.
+  // Each '\0' character in the buffer is replaced with "\\0".
+  //
+  // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+  std::string GetString() const;
+
+ private:
+
+#if GTEST_OS_SYMBIAN
+  // These are needed as the Nokia Symbian Compiler cannot decide between
+  // const T& and const T* in a function template. The Nokia compiler _can_
+  // decide between class template specializations for T and T*, so a
+  // tr1::type_traits-like is_pointer works, and we can overload on that.
+  template 
+  inline void StreamHelper(internal::true_type /*is_pointer*/, T* pointer) {
+    if (pointer == NULL) {
+      *ss_ << "(null)";
+    } else {
+      *ss_ << pointer;
+    }
+  }
+  template 
+  inline void StreamHelper(internal::false_type /*is_pointer*/,
+                           const T& value) {
+    // See the comments in Message& operator <<(const T&) above for why
+    // we need this using statement.
+    using ::operator <<;
+    *ss_ << value;
+  }
+#endif  // GTEST_OS_SYMBIAN
+
+  // We'll hold the text streamed to this object here.
+  const internal::scoped_ptr< ::std::stringstream> ss_;
+
+  // We declare (but don't implement) this to prevent the compiler
+  // from implementing the assignment operator.
+  void operator=(const Message&);
+};
+
+// Streams a Message to an ostream.
+inline std::ostream& operator <<(std::ostream& os, const Message& sb) {
+  return os << sb.GetString();
+}
+
+namespace internal {
+
+// Converts a streamable value to an std::string.  A NULL pointer is
+// converted to "(null)".  When the input value is a ::string,
+// ::std::string, ::wstring, or ::std::wstring object, each NUL
+// character in it is replaced with "\\0".
+template 
+std::string StreamableToString(const T& streamable) {
+  return (Message() << streamable).GetString();
+}
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_MESSAGE_H_
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan), eefacm@gmail.com (Sean Mcafee)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file declares the String class and functions used internally by
+// Google Test.  They are subject to change without notice. They should not used
+// by code external to Google Test.
+//
+// This header file is #included by .
+// It should not be #included by other files.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_STRING_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_STRING_H_
+
+#ifdef __BORLANDC__
+// string.h is not guaranteed to provide strcpy on C++ Builder.
+# include 
+#endif
+
+#include 
+#include 
+
+
+namespace testing {
+namespace internal {
+
+// String - an abstract class holding static string utilities.
+class GTEST_API_ String {
+ public:
+  // Static utility methods
+
+  // Clones a 0-terminated C string, allocating memory using new.  The
+  // caller is responsible for deleting the return value using
+  // delete[].  Returns the cloned string, or NULL if the input is
+  // NULL.
+  //
+  // This is different from strdup() in string.h, which allocates
+  // memory using malloc().
+  static const char* CloneCString(const char* c_str);
+
+#if GTEST_OS_WINDOWS_MOBILE
+  // Windows CE does not have the 'ANSI' versions of Win32 APIs. To be
+  // able to pass strings to Win32 APIs on CE we need to convert them
+  // to 'Unicode', UTF-16.
+
+  // Creates a UTF-16 wide string from the given ANSI string, allocating
+  // memory using new. The caller is responsible for deleting the return
+  // value using delete[]. Returns the wide string, or NULL if the
+  // input is NULL.
+  //
+  // The wide string is created using the ANSI codepage (CP_ACP) to
+  // match the behaviour of the ANSI versions of Win32 calls and the
+  // C runtime.
+  static LPCWSTR AnsiToUtf16(const char* c_str);
+
+  // Creates an ANSI string from the given wide string, allocating
+  // memory using new. The caller is responsible for deleting the return
+  // value using delete[]. Returns the ANSI string, or NULL if the
+  // input is NULL.
+  //
+  // The returned string is created using the ANSI codepage (CP_ACP) to
+  // match the behaviour of the ANSI versions of Win32 calls and the
+  // C runtime.
+  static const char* Utf16ToAnsi(LPCWSTR utf16_str);
+#endif
+
+  // Compares two C strings.  Returns true iff they have the same content.
+  //
+  // Unlike strcmp(), this function can handle NULL argument(s).  A
+  // NULL C string is considered different to any non-NULL C string,
+  // including the empty string.
+  static bool CStringEquals(const char* lhs, const char* rhs);
+
+  // Converts a wide C string to a String using the UTF-8 encoding.
+  // NULL will be converted to "(null)".  If an error occurred during
+  // the conversion, "(failed to convert from wide string)" is
+  // returned.
+  static std::string ShowWideCString(const wchar_t* wide_c_str);
+
+  // Compares two wide C strings.  Returns true iff they have the same
+  // content.
+  //
+  // Unlike wcscmp(), this function can handle NULL argument(s).  A
+  // NULL C string is considered different to any non-NULL C string,
+  // including the empty string.
+  static bool WideCStringEquals(const wchar_t* lhs, const wchar_t* rhs);
+
+  // Compares two C strings, ignoring case.  Returns true iff they
+  // have the same content.
+  //
+  // Unlike strcasecmp(), this function can handle NULL argument(s).
+  // A NULL C string is considered different to any non-NULL C string,
+  // including the empty string.
+  static bool CaseInsensitiveCStringEquals(const char* lhs,
+                                           const char* rhs);
+
+  // Compares two wide C strings, ignoring case.  Returns true iff they
+  // have the same content.
+  //
+  // Unlike wcscasecmp(), this function can handle NULL argument(s).
+  // A NULL C string is considered different to any non-NULL wide C string,
+  // including the empty string.
+  // NB: The implementations on different platforms slightly differ.
+  // On windows, this method uses _wcsicmp which compares according to LC_CTYPE
+  // environment variable. On GNU platform this method uses wcscasecmp
+  // which compares according to LC_CTYPE category of the current locale.
+  // On MacOS X, it uses towlower, which also uses LC_CTYPE category of the
+  // current locale.
+  static bool CaseInsensitiveWideCStringEquals(const wchar_t* lhs,
+                                               const wchar_t* rhs);
+
+  // Returns true iff the given string ends with the given suffix, ignoring
+  // case. Any string is considered to end with an empty suffix.
+  static bool EndsWithCaseInsensitive(
+      const std::string& str, const std::string& suffix);
+
+  // Formats an int value as "%02d".
+  static std::string FormatIntWidth2(int value);  // "%02d" for width == 2
+
+  // Formats an int value as "%X".
+  static std::string FormatHexInt(int value);
+
+  // Formats a byte as "%02X".
+  static std::string FormatByte(unsigned char value);
+
+ private:
+  String();  // Not meant to be instantiated.
+};  // class String
+
+// Gets the content of the stringstream's buffer as an std::string.  Each '\0'
+// character in the buffer is replaced with "\\0".
+GTEST_API_ std::string StringStreamToString(::std::stringstream* stream);
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_STRING_H_
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: keith.ray@gmail.com (Keith Ray)
+//
+// Google Test filepath utilities
+//
+// This header file declares classes and functions used internally by
+// Google Test.  They are subject to change without notice.
+//
+// This file is #included in .
+// Do not include this header file separately!
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_FILEPATH_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_FILEPATH_H_
+
+
+namespace testing {
+namespace internal {
+
+// FilePath - a class for file and directory pathname manipulation which
+// handles platform-specific conventions (like the pathname separator).
+// Used for helper functions for naming files in a directory for xml output.
+// Except for Set methods, all methods are const or static, which provides an
+// "immutable value object" -- useful for peace of mind.
+// A FilePath with a value ending in a path separator ("like/this/") represents
+// a directory, otherwise it is assumed to represent a file. In either case,
+// it may or may not represent an actual file or directory in the file system.
+// Names are NOT checked for syntax correctness -- no checking for illegal
+// characters, malformed paths, etc.
+
+class GTEST_API_ FilePath {
+ public:
+  FilePath() : pathname_("") { }
+  FilePath(const FilePath& rhs) : pathname_(rhs.pathname_) { }
+
+  explicit FilePath(const std::string& pathname) : pathname_(pathname) {
+    Normalize();
+  }
+
+  FilePath& operator=(const FilePath& rhs) {
+    Set(rhs);
+    return *this;
+  }
+
+  void Set(const FilePath& rhs) {
+    pathname_ = rhs.pathname_;
+  }
+
+  const std::string& string() const { return pathname_; }
+  const char* c_str() const { return pathname_.c_str(); }
+
+  // Returns the current working directory, or "" if unsuccessful.
+  static FilePath GetCurrentDir();
+
+  // Given directory = "dir", base_name = "test", number = 0,
+  // extension = "xml", returns "dir/test.xml". If number is greater
+  // than zero (e.g., 12), returns "dir/test_12.xml".
+  // On Windows platform, uses \ as the separator rather than /.
+  static FilePath MakeFileName(const FilePath& directory,
+                               const FilePath& base_name,
+                               int number,
+                               const char* extension);
+
+  // Given directory = "dir", relative_path = "test.xml",
+  // returns "dir/test.xml".
+  // On Windows, uses \ as the separator rather than /.
+  static FilePath ConcatPaths(const FilePath& directory,
+                              const FilePath& relative_path);
+
+  // Returns a pathname for a file that does not currently exist. The pathname
+  // will be directory/base_name.extension or
+  // directory/base_name_.extension if directory/base_name.extension
+  // already exists. The number will be incremented until a pathname is found
+  // that does not already exist.
+  // Examples: 'dir/foo_test.xml' or 'dir/foo_test_1.xml'.
+  // There could be a race condition if two or more processes are calling this
+  // function at the same time -- they could both pick the same filename.
+  static FilePath GenerateUniqueFileName(const FilePath& directory,
+                                         const FilePath& base_name,
+                                         const char* extension);
+
+  // Returns true iff the path is "".
+  bool IsEmpty() const { return pathname_.empty(); }
+
+  // If input name has a trailing separator character, removes it and returns
+  // the name, otherwise return the name string unmodified.
+  // On Windows platform, uses \ as the separator, other platforms use /.
+  FilePath RemoveTrailingPathSeparator() const;
+
+  // Returns a copy of the FilePath with the directory part removed.
+  // Example: FilePath("path/to/file").RemoveDirectoryName() returns
+  // FilePath("file"). If there is no directory part ("just_a_file"), it returns
+  // the FilePath unmodified. If there is no file part ("just_a_dir/") it
+  // returns an empty FilePath ("").
+  // On Windows platform, '\' is the path separator, otherwise it is '/'.
+  FilePath RemoveDirectoryName() const;
+
+  // RemoveFileName returns the directory path with the filename removed.
+  // Example: FilePath("path/to/file").RemoveFileName() returns "path/to/".
+  // If the FilePath is "a_file" or "/a_file", RemoveFileName returns
+  // FilePath("./") or, on Windows, FilePath(".\\"). If the filepath does
+  // not have a file, like "just/a/dir/", it returns the FilePath unmodified.
+  // On Windows platform, '\' is the path separator, otherwise it is '/'.
+  FilePath RemoveFileName() const;
+
+  // Returns a copy of the FilePath with the case-insensitive extension removed.
+  // Example: FilePath("dir/file.exe").RemoveExtension("EXE") returns
+  // FilePath("dir/file"). If a case-insensitive extension is not
+  // found, returns a copy of the original FilePath.
+  FilePath RemoveExtension(const char* extension) const;
+
+  // Creates directories so that path exists. Returns true if successful or if
+  // the directories already exist; returns false if unable to create
+  // directories for any reason. Will also return false if the FilePath does
+  // not represent a directory (that is, it doesn't end with a path separator).
+  bool CreateDirectoriesRecursively() const;
+
+  // Create the directory so that path exists. Returns true if successful or
+  // if the directory already exists; returns false if unable to create the
+  // directory for any reason, including if the parent directory does not
+  // exist. Not named "CreateDirectory" because that's a macro on Windows.
+  bool CreateFolder() const;
+
+  // Returns true if FilePath describes something in the file-system,
+  // either a file, directory, or whatever, and that something exists.
+  bool FileOrDirectoryExists() const;
+
+  // Returns true if pathname describes a directory in the file-system
+  // that exists.
+  bool DirectoryExists() const;
+
+  // Returns true if FilePath ends with a path separator, which indicates that
+  // it is intended to represent a directory. Returns false otherwise.
+  // This does NOT check that a directory (or file) actually exists.
+  bool IsDirectory() const;
+
+  // Returns true if pathname describes a root directory. (Windows has one
+  // root directory per disk drive.)
+  bool IsRootDirectory() const;
+
+  // Returns true if pathname describes an absolute path.
+  bool IsAbsolutePath() const;
+
+ private:
+  // Replaces multiple consecutive separators with a single separator.
+  // For example, "bar///foo" becomes "bar/foo". Does not eliminate other
+  // redundancies that might be in a pathname involving "." or "..".
+  //
+  // A pathname with multiple consecutive separators may occur either through
+  // user error or as a result of some scripts or APIs that generate a pathname
+  // with a trailing separator. On other platforms the same API or script
+  // may NOT generate a pathname with a trailing "/". Then elsewhere that
+  // pathname may have another "/" and pathname components added to it,
+  // without checking for the separator already being there.
+  // The script language and operating system may allow paths like "foo//bar"
+  // but some of the functions in FilePath will not handle that correctly. In
+  // particular, RemoveTrailingPathSeparator() only removes one separator, and
+  // it is called in CreateDirectoriesRecursively() assuming that it will change
+  // a pathname from directory syntax (trailing separator) to filename syntax.
+  //
+  // On Windows this method also replaces the alternate path separator '/' with
+  // the primary path separator '\\', so that for example "bar\\/\\foo" becomes
+  // "bar\\foo".
+
+  void Normalize();
+
+  // Returns a pointer to the last occurence of a valid path separator in
+  // the FilePath. On Windows, for example, both '/' and '\' are valid path
+  // separators. Returns NULL if no path separator was found.
+  const char* FindLastPathSeparator() const;
+
+  std::string pathname_;
+};  // class FilePath
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_FILEPATH_H_
+// This file was GENERATED by command:
+//     pump.py gtest-type-util.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Type utilities needed for implementing typed and type-parameterized
+// tests.  This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+// Currently we support at most 50 types in a list, and at most 50
+// type-parameterized tests in one type-parameterized test case.
+// Please contact googletestframework@googlegroups.com if you need
+// more.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
+
+
+// #ifdef __GNUC__ is too general here.  It is possible to use gcc without using
+// libstdc++ (which is where cxxabi.h comes from).
+# if GTEST_HAS_CXXABI_H_
+#  include 
+# elif defined(__HP_aCC)
+#  include 
+# endif  // GTEST_HASH_CXXABI_H_
+
+namespace testing {
+namespace internal {
+
+// GetTypeName() returns a human-readable name of type T.
+// NB: This function is also used in Google Mock, so don't move it inside of
+// the typed-test-only section below.
+template 
+std::string GetTypeName() {
+# if GTEST_HAS_RTTI
+
+  const char* const name = typeid(T).name();
+#  if GTEST_HAS_CXXABI_H_ || defined(__HP_aCC)
+  int status = 0;
+  // gcc's implementation of typeid(T).name() mangles the type name,
+  // so we have to demangle it.
+#   if GTEST_HAS_CXXABI_H_
+  using abi::__cxa_demangle;
+#   endif  // GTEST_HAS_CXXABI_H_
+  char* const readable_name = __cxa_demangle(name, 0, 0, &status);
+  const std::string name_str(status == 0 ? readable_name : name);
+  free(readable_name);
+  return name_str;
+#  else
+  return name;
+#  endif  // GTEST_HAS_CXXABI_H_ || __HP_aCC
+
+# else
+
+  return "";
+
+# endif  // GTEST_HAS_RTTI
+}
+
+#if GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+// AssertyTypeEq::type is defined iff T1 and T2 are the same
+// type.  This can be used as a compile-time assertion to ensure that
+// two types are equal.
+
+template 
+struct AssertTypeEq;
+
+template 
+struct AssertTypeEq {
+  typedef bool type;
+};
+
+// A unique type used as the default value for the arguments of class
+// template Types.  This allows us to simulate variadic templates
+// (e.g. Types, Type, and etc), which C++ doesn't
+// support directly.
+struct None {};
+
+// The following family of struct and struct templates are used to
+// represent type lists.  In particular, TypesN
+// represents a type list with N types (T1, T2, ..., and TN) in it.
+// Except for Types0, every struct in the family has two member types:
+// Head for the first type in the list, and Tail for the rest of the
+// list.
+
+// The empty type list.
+struct Types0 {};
+
+// Type lists of length 1, 2, 3, and so on.
+
+template 
+struct Types1 {
+  typedef T1 Head;
+  typedef Types0 Tail;
+};
+template 
+struct Types2 {
+  typedef T1 Head;
+  typedef Types1 Tail;
+};
+
+template 
+struct Types3 {
+  typedef T1 Head;
+  typedef Types2 Tail;
+};
+
+template 
+struct Types4 {
+  typedef T1 Head;
+  typedef Types3 Tail;
+};
+
+template 
+struct Types5 {
+  typedef T1 Head;
+  typedef Types4 Tail;
+};
+
+template 
+struct Types6 {
+  typedef T1 Head;
+  typedef Types5 Tail;
+};
+
+template 
+struct Types7 {
+  typedef T1 Head;
+  typedef Types6 Tail;
+};
+
+template 
+struct Types8 {
+  typedef T1 Head;
+  typedef Types7 Tail;
+};
+
+template 
+struct Types9 {
+  typedef T1 Head;
+  typedef Types8 Tail;
+};
+
+template 
+struct Types10 {
+  typedef T1 Head;
+  typedef Types9 Tail;
+};
+
+template 
+struct Types11 {
+  typedef T1 Head;
+  typedef Types10 Tail;
+};
+
+template 
+struct Types12 {
+  typedef T1 Head;
+  typedef Types11 Tail;
+};
+
+template 
+struct Types13 {
+  typedef T1 Head;
+  typedef Types12 Tail;
+};
+
+template 
+struct Types14 {
+  typedef T1 Head;
+  typedef Types13 Tail;
+};
+
+template 
+struct Types15 {
+  typedef T1 Head;
+  typedef Types14 Tail;
+};
+
+template 
+struct Types16 {
+  typedef T1 Head;
+  typedef Types15 Tail;
+};
+
+template 
+struct Types17 {
+  typedef T1 Head;
+  typedef Types16 Tail;
+};
+
+template 
+struct Types18 {
+  typedef T1 Head;
+  typedef Types17 Tail;
+};
+
+template 
+struct Types19 {
+  typedef T1 Head;
+  typedef Types18 Tail;
+};
+
+template 
+struct Types20 {
+  typedef T1 Head;
+  typedef Types19 Tail;
+};
+
+template 
+struct Types21 {
+  typedef T1 Head;
+  typedef Types20 Tail;
+};
+
+template 
+struct Types22 {
+  typedef T1 Head;
+  typedef Types21 Tail;
+};
+
+template 
+struct Types23 {
+  typedef T1 Head;
+  typedef Types22 Tail;
+};
+
+template 
+struct Types24 {
+  typedef T1 Head;
+  typedef Types23 Tail;
+};
+
+template 
+struct Types25 {
+  typedef T1 Head;
+  typedef Types24 Tail;
+};
+
+template 
+struct Types26 {
+  typedef T1 Head;
+  typedef Types25 Tail;
+};
+
+template 
+struct Types27 {
+  typedef T1 Head;
+  typedef Types26 Tail;
+};
+
+template 
+struct Types28 {
+  typedef T1 Head;
+  typedef Types27 Tail;
+};
+
+template 
+struct Types29 {
+  typedef T1 Head;
+  typedef Types28 Tail;
+};
+
+template 
+struct Types30 {
+  typedef T1 Head;
+  typedef Types29 Tail;
+};
+
+template 
+struct Types31 {
+  typedef T1 Head;
+  typedef Types30 Tail;
+};
+
+template 
+struct Types32 {
+  typedef T1 Head;
+  typedef Types31 Tail;
+};
+
+template 
+struct Types33 {
+  typedef T1 Head;
+  typedef Types32 Tail;
+};
+
+template 
+struct Types34 {
+  typedef T1 Head;
+  typedef Types33 Tail;
+};
+
+template 
+struct Types35 {
+  typedef T1 Head;
+  typedef Types34 Tail;
+};
+
+template 
+struct Types36 {
+  typedef T1 Head;
+  typedef Types35 Tail;
+};
+
+template 
+struct Types37 {
+  typedef T1 Head;
+  typedef Types36 Tail;
+};
+
+template 
+struct Types38 {
+  typedef T1 Head;
+  typedef Types37 Tail;
+};
+
+template 
+struct Types39 {
+  typedef T1 Head;
+  typedef Types38 Tail;
+};
+
+template 
+struct Types40 {
+  typedef T1 Head;
+  typedef Types39 Tail;
+};
+
+template 
+struct Types41 {
+  typedef T1 Head;
+  typedef Types40 Tail;
+};
+
+template 
+struct Types42 {
+  typedef T1 Head;
+  typedef Types41 Tail;
+};
+
+template 
+struct Types43 {
+  typedef T1 Head;
+  typedef Types42 Tail;
+};
+
+template 
+struct Types44 {
+  typedef T1 Head;
+  typedef Types43 Tail;
+};
+
+template 
+struct Types45 {
+  typedef T1 Head;
+  typedef Types44 Tail;
+};
+
+template 
+struct Types46 {
+  typedef T1 Head;
+  typedef Types45 Tail;
+};
+
+template 
+struct Types47 {
+  typedef T1 Head;
+  typedef Types46 Tail;
+};
+
+template 
+struct Types48 {
+  typedef T1 Head;
+  typedef Types47 Tail;
+};
+
+template 
+struct Types49 {
+  typedef T1 Head;
+  typedef Types48 Tail;
+};
+
+template 
+struct Types50 {
+  typedef T1 Head;
+  typedef Types49 Tail;
+};
+
+
+}  // namespace internal
+
+// We don't want to require the users to write TypesN<...> directly,
+// as that would require them to count the length.  Types<...> is much
+// easier to write, but generates horrible messages when there is a
+// compiler error, as gcc insists on printing out each template
+// argument, even if it has the default value (this means Types
+// will appear as Types in the compiler
+// errors).
+//
+// Our solution is to combine the best part of the two approaches: a
+// user would write Types, and Google Test will translate
+// that to TypesN internally to make error messages
+// readable.  The translation is done by the 'type' member of the
+// Types template.
+template 
+struct Types {
+  typedef internal::Types50 type;
+};
+
+template <>
+struct Types {
+  typedef internal::Types0 type;
+};
+template 
+struct Types {
+  typedef internal::Types1 type;
+};
+template 
+struct Types {
+  typedef internal::Types2 type;
+};
+template 
+struct Types {
+  typedef internal::Types3 type;
+};
+template 
+struct Types {
+  typedef internal::Types4 type;
+};
+template 
+struct Types {
+  typedef internal::Types5 type;
+};
+template 
+struct Types {
+  typedef internal::Types6 type;
+};
+template 
+struct Types {
+  typedef internal::Types7 type;
+};
+template 
+struct Types {
+  typedef internal::Types8 type;
+};
+template 
+struct Types {
+  typedef internal::Types9 type;
+};
+template 
+struct Types {
+  typedef internal::Types10 type;
+};
+template 
+struct Types {
+  typedef internal::Types11 type;
+};
+template 
+struct Types {
+  typedef internal::Types12 type;
+};
+template 
+struct Types {
+  typedef internal::Types13 type;
+};
+template 
+struct Types {
+  typedef internal::Types14 type;
+};
+template 
+struct Types {
+  typedef internal::Types15 type;
+};
+template 
+struct Types {
+  typedef internal::Types16 type;
+};
+template 
+struct Types {
+  typedef internal::Types17 type;
+};
+template 
+struct Types {
+  typedef internal::Types18 type;
+};
+template 
+struct Types {
+  typedef internal::Types19 type;
+};
+template 
+struct Types {
+  typedef internal::Types20 type;
+};
+template 
+struct Types {
+  typedef internal::Types21 type;
+};
+template 
+struct Types {
+  typedef internal::Types22 type;
+};
+template 
+struct Types {
+  typedef internal::Types23 type;
+};
+template 
+struct Types {
+  typedef internal::Types24 type;
+};
+template 
+struct Types {
+  typedef internal::Types25 type;
+};
+template 
+struct Types {
+  typedef internal::Types26 type;
+};
+template 
+struct Types {
+  typedef internal::Types27 type;
+};
+template 
+struct Types {
+  typedef internal::Types28 type;
+};
+template 
+struct Types {
+  typedef internal::Types29 type;
+};
+template 
+struct Types {
+  typedef internal::Types30 type;
+};
+template 
+struct Types {
+  typedef internal::Types31 type;
+};
+template 
+struct Types {
+  typedef internal::Types32 type;
+};
+template 
+struct Types {
+  typedef internal::Types33 type;
+};
+template 
+struct Types {
+  typedef internal::Types34 type;
+};
+template 
+struct Types {
+  typedef internal::Types35 type;
+};
+template 
+struct Types {
+  typedef internal::Types36 type;
+};
+template 
+struct Types {
+  typedef internal::Types37 type;
+};
+template 
+struct Types {
+  typedef internal::Types38 type;
+};
+template 
+struct Types {
+  typedef internal::Types39 type;
+};
+template 
+struct Types {
+  typedef internal::Types40 type;
+};
+template 
+struct Types {
+  typedef internal::Types41 type;
+};
+template 
+struct Types {
+  typedef internal::Types42 type;
+};
+template 
+struct Types {
+  typedef internal::Types43 type;
+};
+template 
+struct Types {
+  typedef internal::Types44 type;
+};
+template 
+struct Types {
+  typedef internal::Types45 type;
+};
+template 
+struct Types {
+  typedef internal::Types46 type;
+};
+template 
+struct Types {
+  typedef internal::Types47 type;
+};
+template 
+struct Types {
+  typedef internal::Types48 type;
+};
+template 
+struct Types {
+  typedef internal::Types49 type;
+};
+
+namespace internal {
+
+# define GTEST_TEMPLATE_ template  class
+
+// The template "selector" struct TemplateSel is used to
+// represent Tmpl, which must be a class template with one type
+// parameter, as a type.  TemplateSel::Bind::type is defined
+// as the type Tmpl.  This allows us to actually instantiate the
+// template "selected" by TemplateSel.
+//
+// This trick is necessary for simulating typedef for class templates,
+// which C++ doesn't support directly.
+template 
+struct TemplateSel {
+  template 
+  struct Bind {
+    typedef Tmpl type;
+  };
+};
+
+# define GTEST_BIND_(TmplSel, T) \
+  TmplSel::template Bind::type
+
+// A unique struct template used as the default value for the
+// arguments of class template Templates.  This allows us to simulate
+// variadic templates (e.g. Templates, Templates,
+// and etc), which C++ doesn't support directly.
+template 
+struct NoneT {};
+
+// The following family of struct and struct templates are used to
+// represent template lists.  In particular, TemplatesN represents a list of N templates (T1, T2, ..., and TN).  Except
+// for Templates0, every struct in the family has two member types:
+// Head for the selector of the first template in the list, and Tail
+// for the rest of the list.
+
+// The empty template list.
+struct Templates0 {};
+
+// Template lists of length 1, 2, 3, and so on.
+
+template 
+struct Templates1 {
+  typedef TemplateSel Head;
+  typedef Templates0 Tail;
+};
+template 
+struct Templates2 {
+  typedef TemplateSel Head;
+  typedef Templates1 Tail;
+};
+
+template 
+struct Templates3 {
+  typedef TemplateSel Head;
+  typedef Templates2 Tail;
+};
+
+template 
+struct Templates4 {
+  typedef TemplateSel Head;
+  typedef Templates3 Tail;
+};
+
+template 
+struct Templates5 {
+  typedef TemplateSel Head;
+  typedef Templates4 Tail;
+};
+
+template 
+struct Templates6 {
+  typedef TemplateSel Head;
+  typedef Templates5 Tail;
+};
+
+template 
+struct Templates7 {
+  typedef TemplateSel Head;
+  typedef Templates6 Tail;
+};
+
+template 
+struct Templates8 {
+  typedef TemplateSel Head;
+  typedef Templates7 Tail;
+};
+
+template 
+struct Templates9 {
+  typedef TemplateSel Head;
+  typedef Templates8 Tail;
+};
+
+template 
+struct Templates10 {
+  typedef TemplateSel Head;
+  typedef Templates9 Tail;
+};
+
+template 
+struct Templates11 {
+  typedef TemplateSel Head;
+  typedef Templates10 Tail;
+};
+
+template 
+struct Templates12 {
+  typedef TemplateSel Head;
+  typedef Templates11 Tail;
+};
+
+template 
+struct Templates13 {
+  typedef TemplateSel Head;
+  typedef Templates12 Tail;
+};
+
+template 
+struct Templates14 {
+  typedef TemplateSel Head;
+  typedef Templates13 Tail;
+};
+
+template 
+struct Templates15 {
+  typedef TemplateSel Head;
+  typedef Templates14 Tail;
+};
+
+template 
+struct Templates16 {
+  typedef TemplateSel Head;
+  typedef Templates15 Tail;
+};
+
+template 
+struct Templates17 {
+  typedef TemplateSel Head;
+  typedef Templates16 Tail;
+};
+
+template 
+struct Templates18 {
+  typedef TemplateSel Head;
+  typedef Templates17 Tail;
+};
+
+template 
+struct Templates19 {
+  typedef TemplateSel Head;
+  typedef Templates18 Tail;
+};
+
+template 
+struct Templates20 {
+  typedef TemplateSel Head;
+  typedef Templates19 Tail;
+};
+
+template 
+struct Templates21 {
+  typedef TemplateSel Head;
+  typedef Templates20 Tail;
+};
+
+template 
+struct Templates22 {
+  typedef TemplateSel Head;
+  typedef Templates21 Tail;
+};
+
+template 
+struct Templates23 {
+  typedef TemplateSel Head;
+  typedef Templates22 Tail;
+};
+
+template 
+struct Templates24 {
+  typedef TemplateSel Head;
+  typedef Templates23 Tail;
+};
+
+template 
+struct Templates25 {
+  typedef TemplateSel Head;
+  typedef Templates24 Tail;
+};
+
+template 
+struct Templates26 {
+  typedef TemplateSel Head;
+  typedef Templates25 Tail;
+};
+
+template 
+struct Templates27 {
+  typedef TemplateSel Head;
+  typedef Templates26 Tail;
+};
+
+template 
+struct Templates28 {
+  typedef TemplateSel Head;
+  typedef Templates27 Tail;
+};
+
+template 
+struct Templates29 {
+  typedef TemplateSel Head;
+  typedef Templates28 Tail;
+};
+
+template 
+struct Templates30 {
+  typedef TemplateSel Head;
+  typedef Templates29 Tail;
+};
+
+template 
+struct Templates31 {
+  typedef TemplateSel Head;
+  typedef Templates30 Tail;
+};
+
+template 
+struct Templates32 {
+  typedef TemplateSel Head;
+  typedef Templates31 Tail;
+};
+
+template 
+struct Templates33 {
+  typedef TemplateSel Head;
+  typedef Templates32 Tail;
+};
+
+template 
+struct Templates34 {
+  typedef TemplateSel Head;
+  typedef Templates33 Tail;
+};
+
+template 
+struct Templates35 {
+  typedef TemplateSel Head;
+  typedef Templates34 Tail;
+};
+
+template 
+struct Templates36 {
+  typedef TemplateSel Head;
+  typedef Templates35 Tail;
+};
+
+template 
+struct Templates37 {
+  typedef TemplateSel Head;
+  typedef Templates36 Tail;
+};
+
+template 
+struct Templates38 {
+  typedef TemplateSel Head;
+  typedef Templates37 Tail;
+};
+
+template 
+struct Templates39 {
+  typedef TemplateSel Head;
+  typedef Templates38 Tail;
+};
+
+template 
+struct Templates40 {
+  typedef TemplateSel Head;
+  typedef Templates39 Tail;
+};
+
+template 
+struct Templates41 {
+  typedef TemplateSel Head;
+  typedef Templates40 Tail;
+};
+
+template 
+struct Templates42 {
+  typedef TemplateSel Head;
+  typedef Templates41 Tail;
+};
+
+template 
+struct Templates43 {
+  typedef TemplateSel Head;
+  typedef Templates42 Tail;
+};
+
+template 
+struct Templates44 {
+  typedef TemplateSel Head;
+  typedef Templates43 Tail;
+};
+
+template 
+struct Templates45 {
+  typedef TemplateSel Head;
+  typedef Templates44 Tail;
+};
+
+template 
+struct Templates46 {
+  typedef TemplateSel Head;
+  typedef Templates45 Tail;
+};
+
+template 
+struct Templates47 {
+  typedef TemplateSel Head;
+  typedef Templates46 Tail;
+};
+
+template 
+struct Templates48 {
+  typedef TemplateSel Head;
+  typedef Templates47 Tail;
+};
+
+template 
+struct Templates49 {
+  typedef TemplateSel Head;
+  typedef Templates48 Tail;
+};
+
+template 
+struct Templates50 {
+  typedef TemplateSel Head;
+  typedef Templates49 Tail;
+};
+
+
+// We don't want to require the users to write TemplatesN<...> directly,
+// as that would require them to count the length.  Templates<...> is much
+// easier to write, but generates horrible messages when there is a
+// compiler error, as gcc insists on printing out each template
+// argument, even if it has the default value (this means Templates
+// will appear as Templates in the compiler
+// errors).
+//
+// Our solution is to combine the best part of the two approaches: a
+// user would write Templates, and Google Test will translate
+// that to TemplatesN internally to make error messages
+// readable.  The translation is done by the 'type' member of the
+// Templates template.
+template 
+struct Templates {
+  typedef Templates50 type;
+};
+
+template <>
+struct Templates {
+  typedef Templates0 type;
+};
+template 
+struct Templates {
+  typedef Templates1 type;
+};
+template 
+struct Templates {
+  typedef Templates2 type;
+};
+template 
+struct Templates {
+  typedef Templates3 type;
+};
+template 
+struct Templates {
+  typedef Templates4 type;
+};
+template 
+struct Templates {
+  typedef Templates5 type;
+};
+template 
+struct Templates {
+  typedef Templates6 type;
+};
+template 
+struct Templates {
+  typedef Templates7 type;
+};
+template 
+struct Templates {
+  typedef Templates8 type;
+};
+template 
+struct Templates {
+  typedef Templates9 type;
+};
+template 
+struct Templates {
+  typedef Templates10 type;
+};
+template 
+struct Templates {
+  typedef Templates11 type;
+};
+template 
+struct Templates {
+  typedef Templates12 type;
+};
+template 
+struct Templates {
+  typedef Templates13 type;
+};
+template 
+struct Templates {
+  typedef Templates14 type;
+};
+template 
+struct Templates {
+  typedef Templates15 type;
+};
+template 
+struct Templates {
+  typedef Templates16 type;
+};
+template 
+struct Templates {
+  typedef Templates17 type;
+};
+template 
+struct Templates {
+  typedef Templates18 type;
+};
+template 
+struct Templates {
+  typedef Templates19 type;
+};
+template 
+struct Templates {
+  typedef Templates20 type;
+};
+template 
+struct Templates {
+  typedef Templates21 type;
+};
+template 
+struct Templates {
+  typedef Templates22 type;
+};
+template 
+struct Templates {
+  typedef Templates23 type;
+};
+template 
+struct Templates {
+  typedef Templates24 type;
+};
+template 
+struct Templates {
+  typedef Templates25 type;
+};
+template 
+struct Templates {
+  typedef Templates26 type;
+};
+template 
+struct Templates {
+  typedef Templates27 type;
+};
+template 
+struct Templates {
+  typedef Templates28 type;
+};
+template 
+struct Templates {
+  typedef Templates29 type;
+};
+template 
+struct Templates {
+  typedef Templates30 type;
+};
+template 
+struct Templates {
+  typedef Templates31 type;
+};
+template 
+struct Templates {
+  typedef Templates32 type;
+};
+template 
+struct Templates {
+  typedef Templates33 type;
+};
+template 
+struct Templates {
+  typedef Templates34 type;
+};
+template 
+struct Templates {
+  typedef Templates35 type;
+};
+template 
+struct Templates {
+  typedef Templates36 type;
+};
+template 
+struct Templates {
+  typedef Templates37 type;
+};
+template 
+struct Templates {
+  typedef Templates38 type;
+};
+template 
+struct Templates {
+  typedef Templates39 type;
+};
+template 
+struct Templates {
+  typedef Templates40 type;
+};
+template 
+struct Templates {
+  typedef Templates41 type;
+};
+template 
+struct Templates {
+  typedef Templates42 type;
+};
+template 
+struct Templates {
+  typedef Templates43 type;
+};
+template 
+struct Templates {
+  typedef Templates44 type;
+};
+template 
+struct Templates {
+  typedef Templates45 type;
+};
+template 
+struct Templates {
+  typedef Templates46 type;
+};
+template 
+struct Templates {
+  typedef Templates47 type;
+};
+template 
+struct Templates {
+  typedef Templates48 type;
+};
+template 
+struct Templates {
+  typedef Templates49 type;
+};
+
+// The TypeList template makes it possible to use either a single type
+// or a Types<...> list in TYPED_TEST_CASE() and
+// INSTANTIATE_TYPED_TEST_CASE_P().
+
+template 
+struct TypeList {
+  typedef Types1 type;
+};
+
+template 
+struct TypeList > {
+  typedef typename Types::type type;
+};
+
+#endif  // GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
+
+// Due to C++ preprocessor weirdness, we need double indirection to
+// concatenate two tokens when one of them is __LINE__.  Writing
+//
+//   foo ## __LINE__
+//
+// will result in the token foo__LINE__, instead of foo followed by
+// the current line number.  For more details, see
+// http://www.parashift.com/c++-faq-lite/misc-technical-issues.html#faq-39.6
+#define GTEST_CONCAT_TOKEN_(foo, bar) GTEST_CONCAT_TOKEN_IMPL_(foo, bar)
+#define GTEST_CONCAT_TOKEN_IMPL_(foo, bar) foo ## bar
+
+class ProtocolMessage;
+namespace proto2 { class Message; }
+
+namespace testing {
+
+// Forward declarations.
+
+class AssertionResult;                 // Result of an assertion.
+class Message;                         // Represents a failure message.
+class Test;                            // Represents a test.
+class TestInfo;                        // Information about a test.
+class TestPartResult;                  // Result of a test part.
+class UnitTest;                        // A collection of test cases.
+
+template 
+::std::string PrintToString(const T& value);
+
+namespace internal {
+
+struct TraceInfo;                      // Information about a trace point.
+class ScopedTrace;                     // Implements scoped trace.
+class TestInfoImpl;                    // Opaque implementation of TestInfo
+class UnitTestImpl;                    // Opaque implementation of UnitTest
+
+// How many times InitGoogleTest() has been called.
+GTEST_API_ extern int g_init_gtest_count;
+
+// The text used in failure messages to indicate the start of the
+// stack trace.
+GTEST_API_ extern const char kStackTraceMarker[];
+
+// Two overloaded helpers for checking at compile time whether an
+// expression is a null pointer literal (i.e. NULL or any 0-valued
+// compile-time integral constant).  Their return values have
+// different sizes, so we can use sizeof() to test which version is
+// picked by the compiler.  These helpers have no implementations, as
+// we only need their signatures.
+//
+// Given IsNullLiteralHelper(x), the compiler will pick the first
+// version if x can be implicitly converted to Secret*, and pick the
+// second version otherwise.  Since Secret is a secret and incomplete
+// type, the only expression a user can write that has type Secret* is
+// a null pointer literal.  Therefore, we know that x is a null
+// pointer literal if and only if the first version is picked by the
+// compiler.
+char IsNullLiteralHelper(Secret* p);
+char (&IsNullLiteralHelper(...))[2];  // NOLINT
+
+// A compile-time bool constant that is true if and only if x is a
+// null pointer literal (i.e. NULL or any 0-valued compile-time
+// integral constant).
+#ifdef GTEST_ELLIPSIS_NEEDS_POD_
+// We lose support for NULL detection where the compiler doesn't like
+// passing non-POD classes through ellipsis (...).
+# define GTEST_IS_NULL_LITERAL_(x) false
+#else
+# define GTEST_IS_NULL_LITERAL_(x) \
+    (sizeof(::testing::internal::IsNullLiteralHelper(x)) == 1)
+#endif  // GTEST_ELLIPSIS_NEEDS_POD_
+
+// Appends the user-supplied message to the Google-Test-generated message.
+GTEST_API_ std::string AppendUserMessage(
+    const std::string& gtest_msg, const Message& user_msg);
+
+#if GTEST_HAS_EXCEPTIONS
+
+// This exception is thrown by (and only by) a failed Google Test
+// assertion when GTEST_FLAG(throw_on_failure) is true (if exceptions
+// are enabled).  We derive it from std::runtime_error, which is for
+// errors presumably detectable only at run time.  Since
+// std::runtime_error inherits from std::exception, many testing
+// frameworks know how to extract and print the message inside it.
+class GTEST_API_ GoogleTestFailureException : public ::std::runtime_error {
+ public:
+  explicit GoogleTestFailureException(const TestPartResult& failure);
+};
+
+#endif  // GTEST_HAS_EXCEPTIONS
+
+// A helper class for creating scoped traces in user programs.
+class GTEST_API_ ScopedTrace {
+ public:
+  // The c'tor pushes the given source file location and message onto
+  // a trace stack maintained by Google Test.
+  ScopedTrace(const char* file, int line, const Message& message);
+
+  // The d'tor pops the info pushed by the c'tor.
+  //
+  // Note that the d'tor is not virtual in order to be efficient.
+  // Don't inherit from ScopedTrace!
+  ~ScopedTrace();
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ScopedTrace);
+} GTEST_ATTRIBUTE_UNUSED_;  // A ScopedTrace object does its job in its
+                            // c'tor and d'tor.  Therefore it doesn't
+                            // need to be used otherwise.
+
+// Constructs and returns the message for an equality assertion
+// (e.g. ASSERT_EQ, EXPECT_STREQ, etc) failure.
+//
+// The first four parameters are the expressions used in the assertion
+// and their values, as strings.  For example, for ASSERT_EQ(foo, bar)
+// where foo is 5 and bar is 6, we have:
+//
+//   expected_expression: "foo"
+//   actual_expression:   "bar"
+//   expected_value:      "5"
+//   actual_value:        "6"
+//
+// The ignoring_case parameter is true iff the assertion is a
+// *_STRCASEEQ*.  When it's true, the string " (ignoring case)" will
+// be inserted into the message.
+GTEST_API_ AssertionResult EqFailure(const char* expected_expression,
+                                     const char* actual_expression,
+                                     const std::string& expected_value,
+                                     const std::string& actual_value,
+                                     bool ignoring_case);
+
+// Constructs a failure message for Boolean assertions such as EXPECT_TRUE.
+GTEST_API_ std::string GetBoolAssertionFailureMessage(
+    const AssertionResult& assertion_result,
+    const char* expression_text,
+    const char* actual_predicate_value,
+    const char* expected_predicate_value);
+
+// This template class represents an IEEE floating-point number
+// (either single-precision or double-precision, depending on the
+// template parameters).
+//
+// The purpose of this class is to do more sophisticated number
+// comparison.  (Due to round-off error, etc, it's very unlikely that
+// two floating-points will be equal exactly.  Hence a naive
+// comparison by the == operation often doesn't work.)
+//
+// Format of IEEE floating-point:
+//
+//   The most-significant bit being the leftmost, an IEEE
+//   floating-point looks like
+//
+//     sign_bit exponent_bits fraction_bits
+//
+//   Here, sign_bit is a single bit that designates the sign of the
+//   number.
+//
+//   For float, there are 8 exponent bits and 23 fraction bits.
+//
+//   For double, there are 11 exponent bits and 52 fraction bits.
+//
+//   More details can be found at
+//   http://en.wikipedia.org/wiki/IEEE_floating-point_standard.
+//
+// Template parameter:
+//
+//   RawType: the raw floating-point type (either float or double)
+template 
+class FloatingPoint {
+ public:
+  // Defines the unsigned integer type that has the same size as the
+  // floating point number.
+  typedef typename TypeWithSize::UInt Bits;
+
+  // Constants.
+
+  // # of bits in a number.
+  static const size_t kBitCount = 8*sizeof(RawType);
+
+  // # of fraction bits in a number.
+  static const size_t kFractionBitCount =
+    std::numeric_limits::digits - 1;
+
+  // # of exponent bits in a number.
+  static const size_t kExponentBitCount = kBitCount - 1 - kFractionBitCount;
+
+  // The mask for the sign bit.
+  static const Bits kSignBitMask = static_cast(1) << (kBitCount - 1);
+
+  // The mask for the fraction bits.
+  static const Bits kFractionBitMask =
+    ~static_cast(0) >> (kExponentBitCount + 1);
+
+  // The mask for the exponent bits.
+  static const Bits kExponentBitMask = ~(kSignBitMask | kFractionBitMask);
+
+  // How many ULP's (Units in the Last Place) we want to tolerate when
+  // comparing two numbers.  The larger the value, the more error we
+  // allow.  A 0 value means that two numbers must be exactly the same
+  // to be considered equal.
+  //
+  // The maximum error of a single floating-point operation is 0.5
+  // units in the last place.  On Intel CPU's, all floating-point
+  // calculations are done with 80-bit precision, while double has 64
+  // bits.  Therefore, 4 should be enough for ordinary use.
+  //
+  // See the following article for more details on ULP:
+  // http://randomascii.wordpress.com/2012/02/25/comparing-floating-point-numbers-2012-edition/
+  static const size_t kMaxUlps = 4;
+
+  // Constructs a FloatingPoint from a raw floating-point number.
+  //
+  // On an Intel CPU, passing a non-normalized NAN (Not a Number)
+  // around may change its bits, although the new value is guaranteed
+  // to be also a NAN.  Therefore, don't expect this constructor to
+  // preserve the bits in x when x is a NAN.
+  explicit FloatingPoint(const RawType& x) { u_.value_ = x; }
+
+  // Static methods
+
+  // Reinterprets a bit pattern as a floating-point number.
+  //
+  // This function is needed to test the AlmostEquals() method.
+  static RawType ReinterpretBits(const Bits bits) {
+    FloatingPoint fp(0);
+    fp.u_.bits_ = bits;
+    return fp.u_.value_;
+  }
+
+  // Returns the floating-point number that represent positive infinity.
+  static RawType Infinity() {
+    return ReinterpretBits(kExponentBitMask);
+  }
+
+  // Returns the maximum representable finite floating-point number.
+  static RawType Max();
+
+  // Non-static methods
+
+  // Returns the bits that represents this number.
+  const Bits &bits() const { return u_.bits_; }
+
+  // Returns the exponent bits of this number.
+  Bits exponent_bits() const { return kExponentBitMask & u_.bits_; }
+
+  // Returns the fraction bits of this number.
+  Bits fraction_bits() const { return kFractionBitMask & u_.bits_; }
+
+  // Returns the sign bit of this number.
+  Bits sign_bit() const { return kSignBitMask & u_.bits_; }
+
+  // Returns true iff this is NAN (not a number).
+  bool is_nan() const {
+    // It's a NAN if the exponent bits are all ones and the fraction
+    // bits are not entirely zeros.
+    return (exponent_bits() == kExponentBitMask) && (fraction_bits() != 0);
+  }
+
+  // Returns true iff this number is at most kMaxUlps ULP's away from
+  // rhs.  In particular, this function:
+  //
+  //   - returns false if either number is (or both are) NAN.
+  //   - treats really large numbers as almost equal to infinity.
+  //   - thinks +0.0 and -0.0 are 0 DLP's apart.
+  bool AlmostEquals(const FloatingPoint& rhs) const {
+    // The IEEE standard says that any comparison operation involving
+    // a NAN must return false.
+    if (is_nan() || rhs.is_nan()) return false;
+
+    return DistanceBetweenSignAndMagnitudeNumbers(u_.bits_, rhs.u_.bits_)
+        <= kMaxUlps;
+  }
+
+ private:
+  // The data type used to store the actual floating-point number.
+  union FloatingPointUnion {
+    RawType value_;  // The raw floating-point number.
+    Bits bits_;      // The bits that represent the number.
+  };
+
+  // Converts an integer from the sign-and-magnitude representation to
+  // the biased representation.  More precisely, let N be 2 to the
+  // power of (kBitCount - 1), an integer x is represented by the
+  // unsigned number x + N.
+  //
+  // For instance,
+  //
+  //   -N + 1 (the most negative number representable using
+  //          sign-and-magnitude) is represented by 1;
+  //   0      is represented by N; and
+  //   N - 1  (the biggest number representable using
+  //          sign-and-magnitude) is represented by 2N - 1.
+  //
+  // Read http://en.wikipedia.org/wiki/Signed_number_representations
+  // for more details on signed number representations.
+  static Bits SignAndMagnitudeToBiased(const Bits &sam) {
+    if (kSignBitMask & sam) {
+      // sam represents a negative number.
+      return ~sam + 1;
+    } else {
+      // sam represents a positive number.
+      return kSignBitMask | sam;
+    }
+  }
+
+  // Given two numbers in the sign-and-magnitude representation,
+  // returns the distance between them as an unsigned number.
+  static Bits DistanceBetweenSignAndMagnitudeNumbers(const Bits &sam1,
+                                                     const Bits &sam2) {
+    const Bits biased1 = SignAndMagnitudeToBiased(sam1);
+    const Bits biased2 = SignAndMagnitudeToBiased(sam2);
+    return (biased1 >= biased2) ? (biased1 - biased2) : (biased2 - biased1);
+  }
+
+  FloatingPointUnion u_;
+};
+
+// We cannot use std::numeric_limits::max() as it clashes with the max()
+// macro defined by .
+template <>
+inline float FloatingPoint::Max() { return FLT_MAX; }
+template <>
+inline double FloatingPoint::Max() { return DBL_MAX; }
+
+// Typedefs the instances of the FloatingPoint template class that we
+// care to use.
+typedef FloatingPoint Float;
+typedef FloatingPoint Double;
+
+// In order to catch the mistake of putting tests that use different
+// test fixture classes in the same test case, we need to assign
+// unique IDs to fixture classes and compare them.  The TypeId type is
+// used to hold such IDs.  The user should treat TypeId as an opaque
+// type: the only operation allowed on TypeId values is to compare
+// them for equality using the == operator.
+typedef const void* TypeId;
+
+template 
+class TypeIdHelper {
+ public:
+  // dummy_ must not have a const type.  Otherwise an overly eager
+  // compiler (e.g. MSVC 7.1 & 8.0) may try to merge
+  // TypeIdHelper::dummy_ for different Ts as an "optimization".
+  static bool dummy_;
+};
+
+template 
+bool TypeIdHelper::dummy_ = false;
+
+// GetTypeId() returns the ID of type T.  Different values will be
+// returned for different types.  Calling the function twice with the
+// same type argument is guaranteed to return the same ID.
+template 
+TypeId GetTypeId() {
+  // The compiler is required to allocate a different
+  // TypeIdHelper::dummy_ variable for each T used to instantiate
+  // the template.  Therefore, the address of dummy_ is guaranteed to
+  // be unique.
+  return &(TypeIdHelper::dummy_);
+}
+
+// Returns the type ID of ::testing::Test.  Always call this instead
+// of GetTypeId< ::testing::Test>() to get the type ID of
+// ::testing::Test, as the latter may give the wrong result due to a
+// suspected linker bug when compiling Google Test as a Mac OS X
+// framework.
+GTEST_API_ TypeId GetTestTypeId();
+
+// Defines the abstract factory interface that creates instances
+// of a Test object.
+class TestFactoryBase {
+ public:
+  virtual ~TestFactoryBase() {}
+
+  // Creates a test instance to run. The instance is both created and destroyed
+  // within TestInfoImpl::Run()
+  virtual Test* CreateTest() = 0;
+
+ protected:
+  TestFactoryBase() {}
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestFactoryBase);
+};
+
+// This class provides implementation of TeastFactoryBase interface.
+// It is used in TEST and TEST_F macros.
+template 
+class TestFactoryImpl : public TestFactoryBase {
+ public:
+  virtual Test* CreateTest() { return new TestClass; }
+};
+
+#if GTEST_OS_WINDOWS
+
+// Predicate-formatters for implementing the HRESULT checking macros
+// {ASSERT|EXPECT}_HRESULT_{SUCCEEDED|FAILED}
+// We pass a long instead of HRESULT to avoid causing an
+// include dependency for the HRESULT type.
+GTEST_API_ AssertionResult IsHRESULTSuccess(const char* expr,
+                                            long hr);  // NOLINT
+GTEST_API_ AssertionResult IsHRESULTFailure(const char* expr,
+                                            long hr);  // NOLINT
+
+#endif  // GTEST_OS_WINDOWS
+
+// Types of SetUpTestCase() and TearDownTestCase() functions.
+typedef void (*SetUpTestCaseFunc)();
+typedef void (*TearDownTestCaseFunc)();
+
+// Creates a new TestInfo object and registers it with Google Test;
+// returns the created object.
+//
+// Arguments:
+//
+//   test_case_name:   name of the test case
+//   name:             name of the test
+//   type_param        the name of the test's type parameter, or NULL if
+//                     this is not a typed or a type-parameterized test.
+//   value_param       text representation of the test's value parameter,
+//                     or NULL if this is not a type-parameterized test.
+//   fixture_class_id: ID of the test fixture class
+//   set_up_tc:        pointer to the function that sets up the test case
+//   tear_down_tc:     pointer to the function that tears down the test case
+//   factory:          pointer to the factory that creates a test object.
+//                     The newly created TestInfo instance will assume
+//                     ownership of the factory object.
+GTEST_API_ TestInfo* MakeAndRegisterTestInfo(
+    const char* test_case_name,
+    const char* name,
+    const char* type_param,
+    const char* value_param,
+    TypeId fixture_class_id,
+    SetUpTestCaseFunc set_up_tc,
+    TearDownTestCaseFunc tear_down_tc,
+    TestFactoryBase* factory);
+
+// If *pstr starts with the given prefix, modifies *pstr to be right
+// past the prefix and returns true; otherwise leaves *pstr unchanged
+// and returns false.  None of pstr, *pstr, and prefix can be NULL.
+GTEST_API_ bool SkipPrefix(const char* prefix, const char** pstr);
+
+#if GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+// State of the definition of a type-parameterized test case.
+class GTEST_API_ TypedTestCasePState {
+ public:
+  TypedTestCasePState() : registered_(false) {}
+
+  // Adds the given test name to defined_test_names_ and return true
+  // if the test case hasn't been registered; otherwise aborts the
+  // program.
+  bool AddTestName(const char* file, int line, const char* case_name,
+                   const char* test_name) {
+    if (registered_) {
+      fprintf(stderr, "%s Test %s must be defined before "
+              "REGISTER_TYPED_TEST_CASE_P(%s, ...).\n",
+              FormatFileLocation(file, line).c_str(), test_name, case_name);
+      fflush(stderr);
+      posix::Abort();
+    }
+    defined_test_names_.insert(test_name);
+    return true;
+  }
+
+  // Verifies that registered_tests match the test names in
+  // defined_test_names_; returns registered_tests if successful, or
+  // aborts the program otherwise.
+  const char* VerifyRegisteredTestNames(
+      const char* file, int line, const char* registered_tests);
+
+ private:
+  bool registered_;
+  ::std::set defined_test_names_;
+};
+
+// Skips to the first non-space char after the first comma in 'str';
+// returns NULL if no comma is found in 'str'.
+inline const char* SkipComma(const char* str) {
+  const char* comma = strchr(str, ',');
+  if (comma == NULL) {
+    return NULL;
+  }
+  while (IsSpace(*(++comma))) {}
+  return comma;
+}
+
+// Returns the prefix of 'str' before the first comma in it; returns
+// the entire string if it contains no comma.
+inline std::string GetPrefixUntilComma(const char* str) {
+  const char* comma = strchr(str, ',');
+  return comma == NULL ? str : std::string(str, comma);
+}
+
+// TypeParameterizedTest::Register()
+// registers a list of type-parameterized tests with Google Test.  The
+// return value is insignificant - we just need to return something
+// such that we can call this function in a namespace scope.
+//
+// Implementation note: The GTEST_TEMPLATE_ macro declares a template
+// template parameter.  It's defined in gtest-type-util.h.
+template 
+class TypeParameterizedTest {
+ public:
+  // 'index' is the index of the test in the type list 'Types'
+  // specified in INSTANTIATE_TYPED_TEST_CASE_P(Prefix, TestCase,
+  // Types).  Valid values for 'index' are [0, N - 1] where N is the
+  // length of Types.
+  static bool Register(const char* prefix, const char* case_name,
+                       const char* test_names, int index) {
+    typedef typename Types::Head Type;
+    typedef Fixture FixtureClass;
+    typedef typename GTEST_BIND_(TestSel, Type) TestClass;
+
+    // First, registers the first type-parameterized test in the type
+    // list.
+    MakeAndRegisterTestInfo(
+        (std::string(prefix) + (prefix[0] == '\0' ? "" : "/") + case_name + "/"
+         + StreamableToString(index)).c_str(),
+        GetPrefixUntilComma(test_names).c_str(),
+        GetTypeName().c_str(),
+        NULL,  // No value parameter.
+        GetTypeId(),
+        TestClass::SetUpTestCase,
+        TestClass::TearDownTestCase,
+        new TestFactoryImpl);
+
+    // Next, recurses (at compile time) with the tail of the type list.
+    return TypeParameterizedTest
+        ::Register(prefix, case_name, test_names, index + 1);
+  }
+};
+
+// The base case for the compile time recursion.
+template 
+class TypeParameterizedTest {
+ public:
+  static bool Register(const char* /*prefix*/, const char* /*case_name*/,
+                       const char* /*test_names*/, int /*index*/) {
+    return true;
+  }
+};
+
+// TypeParameterizedTestCase::Register()
+// registers *all combinations* of 'Tests' and 'Types' with Google
+// Test.  The return value is insignificant - we just need to return
+// something such that we can call this function in a namespace scope.
+template 
+class TypeParameterizedTestCase {
+ public:
+  static bool Register(const char* prefix, const char* case_name,
+                       const char* test_names) {
+    typedef typename Tests::Head Head;
+
+    // First, register the first test in 'Test' for each type in 'Types'.
+    TypeParameterizedTest::Register(
+        prefix, case_name, test_names, 0);
+
+    // Next, recurses (at compile time) with the tail of the test list.
+    return TypeParameterizedTestCase
+        ::Register(prefix, case_name, SkipComma(test_names));
+  }
+};
+
+// The base case for the compile time recursion.
+template 
+class TypeParameterizedTestCase {
+ public:
+  static bool Register(const char* /*prefix*/, const char* /*case_name*/,
+                       const char* /*test_names*/) {
+    return true;
+  }
+};
+
+#endif  // GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+// Returns the current OS stack trace as an std::string.
+//
+// The maximum number of stack frames to be included is specified by
+// the gtest_stack_trace_depth flag.  The skip_count parameter
+// specifies the number of top frames to be skipped, which doesn't
+// count against the number of frames to be included.
+//
+// For example, if Foo() calls Bar(), which in turn calls
+// GetCurrentOsStackTraceExceptTop(..., 1), Foo() will be included in
+// the trace but Bar() and GetCurrentOsStackTraceExceptTop() won't.
+GTEST_API_ std::string GetCurrentOsStackTraceExceptTop(
+    UnitTest* unit_test, int skip_count);
+
+// Helpers for suppressing warnings on unreachable code or constant
+// condition.
+
+// Always returns true.
+GTEST_API_ bool AlwaysTrue();
+
+// Always returns false.
+inline bool AlwaysFalse() { return !AlwaysTrue(); }
+
+// Helper for suppressing false warning from Clang on a const char*
+// variable declared in a conditional expression always being NULL in
+// the else branch.
+struct GTEST_API_ ConstCharPtr {
+  ConstCharPtr(const char* str) : value(str) {}
+  operator bool() const { return true; }
+  const char* value;
+};
+
+// A simple Linear Congruential Generator for generating random
+// numbers with a uniform distribution.  Unlike rand() and srand(), it
+// doesn't use global state (and therefore can't interfere with user
+// code).  Unlike rand_r(), it's portable.  An LCG isn't very random,
+// but it's good enough for our purposes.
+class GTEST_API_ Random {
+ public:
+  static const UInt32 kMaxRange = 1u << 31;
+
+  explicit Random(UInt32 seed) : state_(seed) {}
+
+  void Reseed(UInt32 seed) { state_ = seed; }
+
+  // Generates a random number from [0, range).  Crashes if 'range' is
+  // 0 or greater than kMaxRange.
+  UInt32 Generate(UInt32 range);
+
+ private:
+  UInt32 state_;
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Random);
+};
+
+// Defining a variable of type CompileAssertTypesEqual will cause a
+// compiler error iff T1 and T2 are different types.
+template 
+struct CompileAssertTypesEqual;
+
+template 
+struct CompileAssertTypesEqual {
+};
+
+// Removes the reference from a type if it is a reference type,
+// otherwise leaves it unchanged.  This is the same as
+// tr1::remove_reference, which is not widely available yet.
+template 
+struct RemoveReference { typedef T type; };  // NOLINT
+template 
+struct RemoveReference { typedef T type; };  // NOLINT
+
+// A handy wrapper around RemoveReference that works when the argument
+// T depends on template parameters.
+#define GTEST_REMOVE_REFERENCE_(T) \
+    typename ::testing::internal::RemoveReference::type
+
+// Removes const from a type if it is a const type, otherwise leaves
+// it unchanged.  This is the same as tr1::remove_const, which is not
+// widely available yet.
+template 
+struct RemoveConst { typedef T type; };  // NOLINT
+template 
+struct RemoveConst { typedef T type; };  // NOLINT
+
+// MSVC 8.0, Sun C++, and IBM XL C++ have a bug which causes the above
+// definition to fail to remove the const in 'const int[3]' and 'const
+// char[3][4]'.  The following specialization works around the bug.
+template 
+struct RemoveConst {
+  typedef typename RemoveConst::type type[N];
+};
+
+#if defined(_MSC_VER) && _MSC_VER < 1400
+// This is the only specialization that allows VC++ 7.1 to remove const in
+// 'const int[3] and 'const int[3][4]'.  However, it causes trouble with GCC
+// and thus needs to be conditionally compiled.
+template 
+struct RemoveConst {
+  typedef typename RemoveConst::type type[N];
+};
+#endif
+
+// A handy wrapper around RemoveConst that works when the argument
+// T depends on template parameters.
+#define GTEST_REMOVE_CONST_(T) \
+    typename ::testing::internal::RemoveConst::type
+
+// Turns const U&, U&, const U, and U all into U.
+#define GTEST_REMOVE_REFERENCE_AND_CONST_(T) \
+    GTEST_REMOVE_CONST_(GTEST_REMOVE_REFERENCE_(T))
+
+// Adds reference to a type if it is not a reference type,
+// otherwise leaves it unchanged.  This is the same as
+// tr1::add_reference, which is not widely available yet.
+template 
+struct AddReference { typedef T& type; };  // NOLINT
+template 
+struct AddReference { typedef T& type; };  // NOLINT
+
+// A handy wrapper around AddReference that works when the argument T
+// depends on template parameters.
+#define GTEST_ADD_REFERENCE_(T) \
+    typename ::testing::internal::AddReference::type
+
+// Adds a reference to const on top of T as necessary.  For example,
+// it transforms
+//
+//   char         ==> const char&
+//   const char   ==> const char&
+//   char&        ==> const char&
+//   const char&  ==> const char&
+//
+// The argument T must depend on some template parameters.
+#define GTEST_REFERENCE_TO_CONST_(T) \
+    GTEST_ADD_REFERENCE_(const GTEST_REMOVE_REFERENCE_(T))
+
+// ImplicitlyConvertible::value is a compile-time bool
+// constant that's true iff type From can be implicitly converted to
+// type To.
+template 
+class ImplicitlyConvertible {
+ private:
+  // We need the following helper functions only for their types.
+  // They have no implementations.
+
+  // MakeFrom() is an expression whose type is From.  We cannot simply
+  // use From(), as the type From may not have a public default
+  // constructor.
+  static From MakeFrom();
+
+  // These two functions are overloaded.  Given an expression
+  // Helper(x), the compiler will pick the first version if x can be
+  // implicitly converted to type To; otherwise it will pick the
+  // second version.
+  //
+  // The first version returns a value of size 1, and the second
+  // version returns a value of size 2.  Therefore, by checking the
+  // size of Helper(x), which can be done at compile time, we can tell
+  // which version of Helper() is used, and hence whether x can be
+  // implicitly converted to type To.
+  static char Helper(To);
+  static char (&Helper(...))[2];  // NOLINT
+
+  // We have to put the 'public' section after the 'private' section,
+  // or MSVC refuses to compile the code.
+ public:
+  // MSVC warns about implicitly converting from double to int for
+  // possible loss of data, so we need to temporarily disable the
+  // warning.
+#ifdef _MSC_VER
+# pragma warning(push)          // Saves the current warning state.
+# pragma warning(disable:4244)  // Temporarily disables warning 4244.
+
+  static const bool value =
+      sizeof(Helper(ImplicitlyConvertible::MakeFrom())) == 1;
+# pragma warning(pop)           // Restores the warning state.
+#elif defined(__BORLANDC__)
+  // C++Builder cannot use member overload resolution during template
+  // instantiation.  The simplest workaround is to use its C++0x type traits
+  // functions (C++Builder 2009 and above only).
+  static const bool value = __is_convertible(From, To);
+#else
+  static const bool value =
+      sizeof(Helper(ImplicitlyConvertible::MakeFrom())) == 1;
+#endif  // _MSV_VER
+};
+template 
+const bool ImplicitlyConvertible::value;
+
+// IsAProtocolMessage::value is a compile-time bool constant that's
+// true iff T is type ProtocolMessage, proto2::Message, or a subclass
+// of those.
+template 
+struct IsAProtocolMessage
+    : public bool_constant<
+  ImplicitlyConvertible::value ||
+  ImplicitlyConvertible::value> {
+};
+
+// When the compiler sees expression IsContainerTest(0), if C is an
+// STL-style container class, the first overload of IsContainerTest
+// will be viable (since both C::iterator* and C::const_iterator* are
+// valid types and NULL can be implicitly converted to them).  It will
+// be picked over the second overload as 'int' is a perfect match for
+// the type of argument 0.  If C::iterator or C::const_iterator is not
+// a valid type, the first overload is not viable, and the second
+// overload will be picked.  Therefore, we can determine whether C is
+// a container class by checking the type of IsContainerTest(0).
+// The value of the expression is insignificant.
+//
+// Note that we look for both C::iterator and C::const_iterator.  The
+// reason is that C++ injects the name of a class as a member of the
+// class itself (e.g. you can refer to class iterator as either
+// 'iterator' or 'iterator::iterator').  If we look for C::iterator
+// only, for example, we would mistakenly think that a class named
+// iterator is an STL container.
+//
+// Also note that the simpler approach of overloading
+// IsContainerTest(typename C::const_iterator*) and
+// IsContainerTest(...) doesn't work with Visual Age C++ and Sun C++.
+typedef int IsContainer;
+template 
+IsContainer IsContainerTest(int /* dummy */,
+                            typename C::iterator* /* it */ = NULL,
+                            typename C::const_iterator* /* const_it */ = NULL) {
+  return 0;
+}
+
+typedef char IsNotContainer;
+template 
+IsNotContainer IsContainerTest(long /* dummy */) { return '\0'; }
+
+// EnableIf::type is void when 'Cond' is true, and
+// undefined when 'Cond' is false.  To use SFINAE to make a function
+// overload only apply when a particular expression is true, add
+// "typename EnableIf::type* = 0" as the last parameter.
+template struct EnableIf;
+template<> struct EnableIf { typedef void type; };  // NOLINT
+
+// Utilities for native arrays.
+
+// ArrayEq() compares two k-dimensional native arrays using the
+// elements' operator==, where k can be any integer >= 0.  When k is
+// 0, ArrayEq() degenerates into comparing a single pair of values.
+
+template 
+bool ArrayEq(const T* lhs, size_t size, const U* rhs);
+
+// This generic version is used when k is 0.
+template 
+inline bool ArrayEq(const T& lhs, const U& rhs) { return lhs == rhs; }
+
+// This overload is used when k >= 1.
+template 
+inline bool ArrayEq(const T(&lhs)[N], const U(&rhs)[N]) {
+  return internal::ArrayEq(lhs, N, rhs);
+}
+
+// This helper reduces code bloat.  If we instead put its logic inside
+// the previous ArrayEq() function, arrays with different sizes would
+// lead to different copies of the template code.
+template 
+bool ArrayEq(const T* lhs, size_t size, const U* rhs) {
+  for (size_t i = 0; i != size; i++) {
+    if (!internal::ArrayEq(lhs[i], rhs[i]))
+      return false;
+  }
+  return true;
+}
+
+// Finds the first element in the iterator range [begin, end) that
+// equals elem.  Element may be a native array type itself.
+template 
+Iter ArrayAwareFind(Iter begin, Iter end, const Element& elem) {
+  for (Iter it = begin; it != end; ++it) {
+    if (internal::ArrayEq(*it, elem))
+      return it;
+  }
+  return end;
+}
+
+// CopyArray() copies a k-dimensional native array using the elements'
+// operator=, where k can be any integer >= 0.  When k is 0,
+// CopyArray() degenerates into copying a single value.
+
+template 
+void CopyArray(const T* from, size_t size, U* to);
+
+// This generic version is used when k is 0.
+template 
+inline void CopyArray(const T& from, U* to) { *to = from; }
+
+// This overload is used when k >= 1.
+template 
+inline void CopyArray(const T(&from)[N], U(*to)[N]) {
+  internal::CopyArray(from, N, *to);
+}
+
+// This helper reduces code bloat.  If we instead put its logic inside
+// the previous CopyArray() function, arrays with different sizes
+// would lead to different copies of the template code.
+template 
+void CopyArray(const T* from, size_t size, U* to) {
+  for (size_t i = 0; i != size; i++) {
+    internal::CopyArray(from[i], to + i);
+  }
+}
+
+// The relation between an NativeArray object (see below) and the
+// native array it represents.
+enum RelationToSource {
+  kReference,  // The NativeArray references the native array.
+  kCopy        // The NativeArray makes a copy of the native array and
+               // owns the copy.
+};
+
+// Adapts a native array to a read-only STL-style container.  Instead
+// of the complete STL container concept, this adaptor only implements
+// members useful for Google Mock's container matchers.  New members
+// should be added as needed.  To simplify the implementation, we only
+// support Element being a raw type (i.e. having no top-level const or
+// reference modifier).  It's the client's responsibility to satisfy
+// this requirement.  Element can be an array type itself (hence
+// multi-dimensional arrays are supported).
+template 
+class NativeArray {
+ public:
+  // STL-style container typedefs.
+  typedef Element value_type;
+  typedef Element* iterator;
+  typedef const Element* const_iterator;
+
+  // Constructs from a native array.
+  NativeArray(const Element* array, size_t count, RelationToSource relation) {
+    Init(array, count, relation);
+  }
+
+  // Copy constructor.
+  NativeArray(const NativeArray& rhs) {
+    Init(rhs.array_, rhs.size_, rhs.relation_to_source_);
+  }
+
+  ~NativeArray() {
+    // Ensures that the user doesn't instantiate NativeArray with a
+    // const or reference type.
+    static_cast(StaticAssertTypeEqHelper());
+    if (relation_to_source_ == kCopy)
+      delete[] array_;
+  }
+
+  // STL-style container methods.
+  size_t size() const { return size_; }
+  const_iterator begin() const { return array_; }
+  const_iterator end() const { return array_ + size_; }
+  bool operator==(const NativeArray& rhs) const {
+    return size() == rhs.size() &&
+        ArrayEq(begin(), size(), rhs.begin());
+  }
+
+ private:
+  // Initializes this object; makes a copy of the input array if
+  // 'relation' is kCopy.
+  void Init(const Element* array, size_t a_size, RelationToSource relation) {
+    if (relation == kReference) {
+      array_ = array;
+    } else {
+      Element* const copy = new Element[a_size];
+      CopyArray(array, a_size, copy);
+      array_ = copy;
+    }
+    size_ = a_size;
+    relation_to_source_ = relation;
+  }
+
+  const Element* array_;
+  size_t size_;
+  RelationToSource relation_to_source_;
+
+  GTEST_DISALLOW_ASSIGN_(NativeArray);
+};
+
+}  // namespace internal
+}  // namespace testing
+
+#define GTEST_MESSAGE_AT_(file, line, message, result_type) \
+  ::testing::internal::AssertHelper(result_type, file, line, message) \
+    = ::testing::Message()
+
+#define GTEST_MESSAGE_(message, result_type) \
+  GTEST_MESSAGE_AT_(__FILE__, __LINE__, message, result_type)
+
+#define GTEST_FATAL_FAILURE_(message) \
+  return GTEST_MESSAGE_(message, ::testing::TestPartResult::kFatalFailure)
+
+#define GTEST_NONFATAL_FAILURE_(message) \
+  GTEST_MESSAGE_(message, ::testing::TestPartResult::kNonFatalFailure)
+
+#define GTEST_SUCCESS_(message) \
+  GTEST_MESSAGE_(message, ::testing::TestPartResult::kSuccess)
+
+// Suppresses MSVC warnings 4072 (unreachable code) for the code following
+// statement if it returns or throws (or doesn't return or throw in some
+// situations).
+#define GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement) \
+  if (::testing::internal::AlwaysTrue()) { statement; }
+
+#define GTEST_TEST_THROW_(statement, expected_exception, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::ConstCharPtr gtest_msg = "") { \
+    bool gtest_caught_expected = false; \
+    try { \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    } \
+    catch (expected_exception const&) { \
+      gtest_caught_expected = true; \
+    } \
+    catch (...) { \
+      gtest_msg.value = \
+          "Expected: " #statement " throws an exception of type " \
+          #expected_exception ".\n  Actual: it throws a different type."; \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testthrow_, __LINE__); \
+    } \
+    if (!gtest_caught_expected) { \
+      gtest_msg.value = \
+          "Expected: " #statement " throws an exception of type " \
+          #expected_exception ".\n  Actual: it throws nothing."; \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testthrow_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testthrow_, __LINE__): \
+      fail(gtest_msg.value)
+
+#define GTEST_TEST_NO_THROW_(statement, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    try { \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    } \
+    catch (...) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testnothrow_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testnothrow_, __LINE__): \
+      fail("Expected: " #statement " doesn't throw an exception.\n" \
+           "  Actual: it throws.")
+
+#define GTEST_TEST_ANY_THROW_(statement, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    bool gtest_caught_any = false; \
+    try { \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    } \
+    catch (...) { \
+      gtest_caught_any = true; \
+    } \
+    if (!gtest_caught_any) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testanythrow_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testanythrow_, __LINE__): \
+      fail("Expected: " #statement " throws an exception.\n" \
+           "  Actual: it doesn't.")
+
+
+// Implements Boolean test assertions such as EXPECT_TRUE. expression can be
+// either a boolean expression or an AssertionResult. text is a textual
+// represenation of expression as it was passed into the EXPECT_TRUE.
+#define GTEST_TEST_BOOLEAN_(expression, text, actual, expected, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (const ::testing::AssertionResult gtest_ar_ = \
+      ::testing::AssertionResult(expression)) \
+    ; \
+  else \
+    fail(::testing::internal::GetBoolAssertionFailureMessage(\
+        gtest_ar_, text, #actual, #expected).c_str())
+
+#define GTEST_TEST_NO_FATAL_FAILURE_(statement, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    ::testing::internal::HasNewFatalFailureHelper gtest_fatal_failure_checker; \
+    GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    if (gtest_fatal_failure_checker.has_new_fatal_failure()) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testnofatal_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testnofatal_, __LINE__): \
+      fail("Expected: " #statement " doesn't generate new fatal " \
+           "failures in the current thread.\n" \
+           "  Actual: it does.")
+
+// Expands to the name of the class that implements the given test.
+#define GTEST_TEST_CLASS_NAME_(test_case_name, test_name) \
+  test_case_name##_##test_name##_Test
+
+// Helper macro for defining tests.
+#define GTEST_TEST_(test_case_name, test_name, parent_class, parent_id)\
+class GTEST_TEST_CLASS_NAME_(test_case_name, test_name) : public parent_class {\
+ public:\
+  GTEST_TEST_CLASS_NAME_(test_case_name, test_name)() {}\
+ private:\
+  virtual void TestBody();\
+  static ::testing::TestInfo* const test_info_ GTEST_ATTRIBUTE_UNUSED_;\
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(\
+      GTEST_TEST_CLASS_NAME_(test_case_name, test_name));\
+};\
+\
+::testing::TestInfo* const GTEST_TEST_CLASS_NAME_(test_case_name, test_name)\
+  ::test_info_ =\
+    ::testing::internal::MakeAndRegisterTestInfo(\
+        #test_case_name, #test_name, NULL, NULL, \
+        (parent_id), \
+        parent_class::SetUpTestCase, \
+        parent_class::TearDownTestCase, \
+        new ::testing::internal::TestFactoryImpl<\
+            GTEST_TEST_CLASS_NAME_(test_case_name, test_name)>);\
+void GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::TestBody()
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_INTERNAL_H_
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines the public API for death tests.  It is
+// #included by gtest.h so a user doesn't need to include this
+// directly.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_DEATH_TEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_DEATH_TEST_H_
+
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan), eefacm@gmail.com (Sean Mcafee)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines internal utilities needed for implementing
+// death tests.  They are subject to change without notice.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_DEATH_TEST_INTERNAL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_DEATH_TEST_INTERNAL_H_
+
+
+#include 
+
+namespace testing {
+namespace internal {
+
+GTEST_DECLARE_string_(internal_run_death_test);
+
+// Names of the flags (needed for parsing Google Test flags).
+const char kDeathTestStyleFlag[] = "death_test_style";
+const char kDeathTestUseFork[] = "death_test_use_fork";
+const char kInternalRunDeathTestFlag[] = "internal_run_death_test";
+
+#if GTEST_HAS_DEATH_TEST
+
+// DeathTest is a class that hides much of the complexity of the
+// GTEST_DEATH_TEST_ macro.  It is abstract; its static Create method
+// returns a concrete class that depends on the prevailing death test
+// style, as defined by the --gtest_death_test_style and/or
+// --gtest_internal_run_death_test flags.
+
+// In describing the results of death tests, these terms are used with
+// the corresponding definitions:
+//
+// exit status:  The integer exit information in the format specified
+//               by wait(2)
+// exit code:    The integer code passed to exit(3), _exit(2), or
+//               returned from main()
+class GTEST_API_ DeathTest {
+ public:
+  // Create returns false if there was an error determining the
+  // appropriate action to take for the current death test; for example,
+  // if the gtest_death_test_style flag is set to an invalid value.
+  // The LastMessage method will return a more detailed message in that
+  // case.  Otherwise, the DeathTest pointer pointed to by the "test"
+  // argument is set.  If the death test should be skipped, the pointer
+  // is set to NULL; otherwise, it is set to the address of a new concrete
+  // DeathTest object that controls the execution of the current test.
+  static bool Create(const char* statement, const RE* regex,
+                     const char* file, int line, DeathTest** test);
+  DeathTest();
+  virtual ~DeathTest() { }
+
+  // A helper class that aborts a death test when it's deleted.
+  class ReturnSentinel {
+   public:
+    explicit ReturnSentinel(DeathTest* test) : test_(test) { }
+    ~ReturnSentinel() { test_->Abort(TEST_ENCOUNTERED_RETURN_STATEMENT); }
+   private:
+    DeathTest* const test_;
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(ReturnSentinel);
+  } GTEST_ATTRIBUTE_UNUSED_;
+
+  // An enumeration of possible roles that may be taken when a death
+  // test is encountered.  EXECUTE means that the death test logic should
+  // be executed immediately.  OVERSEE means that the program should prepare
+  // the appropriate environment for a child process to execute the death
+  // test, then wait for it to complete.
+  enum TestRole { OVERSEE_TEST, EXECUTE_TEST };
+
+  // An enumeration of the three reasons that a test might be aborted.
+  enum AbortReason {
+    TEST_ENCOUNTERED_RETURN_STATEMENT,
+    TEST_THREW_EXCEPTION,
+    TEST_DID_NOT_DIE
+  };
+
+  // Assumes one of the above roles.
+  virtual TestRole AssumeRole() = 0;
+
+  // Waits for the death test to finish and returns its status.
+  virtual int Wait() = 0;
+
+  // Returns true if the death test passed; that is, the test process
+  // exited during the test, its exit status matches a user-supplied
+  // predicate, and its stderr output matches a user-supplied regular
+  // expression.
+  // The user-supplied predicate may be a macro expression rather
+  // than a function pointer or functor, or else Wait and Passed could
+  // be combined.
+  virtual bool Passed(bool exit_status_ok) = 0;
+
+  // Signals that the death test did not die as expected.
+  virtual void Abort(AbortReason reason) = 0;
+
+  // Returns a human-readable outcome message regarding the outcome of
+  // the last death test.
+  static const char* LastMessage();
+
+  static void set_last_death_test_message(const std::string& message);
+
+ private:
+  // A string containing a description of the outcome of the last death test.
+  static std::string last_death_test_message_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(DeathTest);
+};
+
+// Factory interface for death tests.  May be mocked out for testing.
+class DeathTestFactory {
+ public:
+  virtual ~DeathTestFactory() { }
+  virtual bool Create(const char* statement, const RE* regex,
+                      const char* file, int line, DeathTest** test) = 0;
+};
+
+// A concrete DeathTestFactory implementation for normal use.
+class DefaultDeathTestFactory : public DeathTestFactory {
+ public:
+  virtual bool Create(const char* statement, const RE* regex,
+                      const char* file, int line, DeathTest** test);
+};
+
+// Returns true if exit_status describes a process that was terminated
+// by a signal, or exited normally with a nonzero exit code.
+GTEST_API_ bool ExitedUnsuccessfully(int exit_status);
+
+// Traps C++ exceptions escaping statement and reports them as test
+// failures. Note that trapping SEH exceptions is not implemented here.
+# if GTEST_HAS_EXCEPTIONS
+#  define GTEST_EXECUTE_DEATH_TEST_STATEMENT_(statement, death_test) \
+  try { \
+    GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+  } catch (const ::std::exception& gtest_exception) { \
+    fprintf(\
+        stderr, \
+        "\n%s: Caught std::exception-derived exception escaping the " \
+        "death test statement. Exception message: %s\n", \
+        ::testing::internal::FormatFileLocation(__FILE__, __LINE__).c_str(), \
+        gtest_exception.what()); \
+    fflush(stderr); \
+    death_test->Abort(::testing::internal::DeathTest::TEST_THREW_EXCEPTION); \
+  } catch (...) { \
+    death_test->Abort(::testing::internal::DeathTest::TEST_THREW_EXCEPTION); \
+  }
+
+# else
+#  define GTEST_EXECUTE_DEATH_TEST_STATEMENT_(statement, death_test) \
+  GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement)
+
+# endif
+
+// This macro is for implementing ASSERT_DEATH*, EXPECT_DEATH*,
+// ASSERT_EXIT*, and EXPECT_EXIT*.
+# define GTEST_DEATH_TEST_(statement, predicate, regex, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    const ::testing::internal::RE& gtest_regex = (regex); \
+    ::testing::internal::DeathTest* gtest_dt; \
+    if (!::testing::internal::DeathTest::Create(#statement, >est_regex, \
+        __FILE__, __LINE__, >est_dt)) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_, __LINE__); \
+    } \
+    if (gtest_dt != NULL) { \
+      ::testing::internal::scoped_ptr< ::testing::internal::DeathTest> \
+          gtest_dt_ptr(gtest_dt); \
+      switch (gtest_dt->AssumeRole()) { \
+        case ::testing::internal::DeathTest::OVERSEE_TEST: \
+          if (!gtest_dt->Passed(predicate(gtest_dt->Wait()))) { \
+            goto GTEST_CONCAT_TOKEN_(gtest_label_, __LINE__); \
+          } \
+          break; \
+        case ::testing::internal::DeathTest::EXECUTE_TEST: { \
+          ::testing::internal::DeathTest::ReturnSentinel \
+              gtest_sentinel(gtest_dt); \
+          GTEST_EXECUTE_DEATH_TEST_STATEMENT_(statement, gtest_dt); \
+          gtest_dt->Abort(::testing::internal::DeathTest::TEST_DID_NOT_DIE); \
+          break; \
+        } \
+        default: \
+          break; \
+      } \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_, __LINE__): \
+      fail(::testing::internal::DeathTest::LastMessage())
+// The symbol "fail" here expands to something into which a message
+// can be streamed.
+
+// This macro is for implementing ASSERT/EXPECT_DEBUG_DEATH when compiled in
+// NDEBUG mode. In this case we need the statements to be executed, the regex is
+// ignored, and the macro must accept a streamed message even though the message
+// is never printed.
+# define GTEST_EXECUTE_STATEMENT_(statement, regex) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+     GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+  } else \
+    ::testing::Message()
+
+// A class representing the parsed contents of the
+// --gtest_internal_run_death_test flag, as it existed when
+// RUN_ALL_TESTS was called.
+class InternalRunDeathTestFlag {
+ public:
+  InternalRunDeathTestFlag(const std::string& a_file,
+                           int a_line,
+                           int an_index,
+                           int a_write_fd)
+      : file_(a_file), line_(a_line), index_(an_index),
+        write_fd_(a_write_fd) {}
+
+  ~InternalRunDeathTestFlag() {
+    if (write_fd_ >= 0)
+      posix::Close(write_fd_);
+  }
+
+  const std::string& file() const { return file_; }
+  int line() const { return line_; }
+  int index() const { return index_; }
+  int write_fd() const { return write_fd_; }
+
+ private:
+  std::string file_;
+  int line_;
+  int index_;
+  int write_fd_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(InternalRunDeathTestFlag);
+};
+
+// Returns a newly created InternalRunDeathTestFlag object with fields
+// initialized from the GTEST_FLAG(internal_run_death_test) flag if
+// the flag is specified; otherwise returns NULL.
+InternalRunDeathTestFlag* ParseInternalRunDeathTestFlag();
+
+#else  // GTEST_HAS_DEATH_TEST
+
+// This macro is used for implementing macros such as
+// EXPECT_DEATH_IF_SUPPORTED and ASSERT_DEATH_IF_SUPPORTED on systems where
+// death tests are not supported. Those macros must compile on such systems
+// iff EXPECT_DEATH and ASSERT_DEATH compile with the same parameters on
+// systems that support death tests. This allows one to write such a macro
+// on a system that does not support death tests and be sure that it will
+// compile on a death-test supporting system.
+//
+// Parameters:
+//   statement -  A statement that a macro such as EXPECT_DEATH would test
+//                for program termination. This macro has to make sure this
+//                statement is compiled but not executed, to ensure that
+//                EXPECT_DEATH_IF_SUPPORTED compiles with a certain
+//                parameter iff EXPECT_DEATH compiles with it.
+//   regex     -  A regex that a macro such as EXPECT_DEATH would use to test
+//                the output of statement.  This parameter has to be
+//                compiled but not evaluated by this macro, to ensure that
+//                this macro only accepts expressions that a macro such as
+//                EXPECT_DEATH would accept.
+//   terminator - Must be an empty statement for EXPECT_DEATH_IF_SUPPORTED
+//                and a return statement for ASSERT_DEATH_IF_SUPPORTED.
+//                This ensures that ASSERT_DEATH_IF_SUPPORTED will not
+//                compile inside functions where ASSERT_DEATH doesn't
+//                compile.
+//
+//  The branch that has an always false condition is used to ensure that
+//  statement and regex are compiled (and thus syntactically correct) but
+//  never executed. The unreachable code macro protects the terminator
+//  statement from generating an 'unreachable code' warning in case
+//  statement unconditionally returns or throws. The Message constructor at
+//  the end allows the syntax of streaming additional messages into the
+//  macro, for compilational compatibility with EXPECT_DEATH/ASSERT_DEATH.
+# define GTEST_UNSUPPORTED_DEATH_TEST_(statement, regex, terminator) \
+    GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+    if (::testing::internal::AlwaysTrue()) { \
+      GTEST_LOG_(WARNING) \
+          << "Death tests are not supported on this platform.\n" \
+          << "Statement '" #statement "' cannot be verified."; \
+    } else if (::testing::internal::AlwaysFalse()) { \
+      ::testing::internal::RE::PartialMatch(".*", (regex)); \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+      terminator; \
+    } else \
+      ::testing::Message()
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_DEATH_TEST_INTERNAL_H_
+
+namespace testing {
+
+// This flag controls the style of death tests.  Valid values are "threadsafe",
+// meaning that the death test child process will re-execute the test binary
+// from the start, running only a single death test, or "fast",
+// meaning that the child process will execute the test logic immediately
+// after forking.
+GTEST_DECLARE_string_(death_test_style);
+
+#if GTEST_HAS_DEATH_TEST
+
+namespace internal {
+
+// Returns a Boolean value indicating whether the caller is currently
+// executing in the context of the death test child process.  Tools such as
+// Valgrind heap checkers may need this to modify their behavior in death
+// tests.  IMPORTANT: This is an internal utility.  Using it may break the
+// implementation of death tests.  User code MUST NOT use it.
+GTEST_API_ bool InDeathTestChild();
+
+}  // namespace internal
+
+// The following macros are useful for writing death tests.
+
+// Here's what happens when an ASSERT_DEATH* or EXPECT_DEATH* is
+// executed:
+//
+//   1. It generates a warning if there is more than one active
+//   thread.  This is because it's safe to fork() or clone() only
+//   when there is a single thread.
+//
+//   2. The parent process clone()s a sub-process and runs the death
+//   test in it; the sub-process exits with code 0 at the end of the
+//   death test, if it hasn't exited already.
+//
+//   3. The parent process waits for the sub-process to terminate.
+//
+//   4. The parent process checks the exit code and error message of
+//   the sub-process.
+//
+// Examples:
+//
+//   ASSERT_DEATH(server.SendMessage(56, "Hello"), "Invalid port number");
+//   for (int i = 0; i < 5; i++) {
+//     EXPECT_DEATH(server.ProcessRequest(i),
+//                  "Invalid request .* in ProcessRequest()")
+//                  << "Failed to die on request " << i;
+//   }
+//
+//   ASSERT_EXIT(server.ExitNow(), ::testing::ExitedWithCode(0), "Exiting");
+//
+//   bool KilledBySIGHUP(int exit_code) {
+//     return WIFSIGNALED(exit_code) && WTERMSIG(exit_code) == SIGHUP;
+//   }
+//
+//   ASSERT_EXIT(client.HangUpServer(), KilledBySIGHUP, "Hanging up!");
+//
+// On the regular expressions used in death tests:
+//
+//   On POSIX-compliant systems (*nix), we use the  library,
+//   which uses the POSIX extended regex syntax.
+//
+//   On other platforms (e.g. Windows), we only support a simple regex
+//   syntax implemented as part of Google Test.  This limited
+//   implementation should be enough most of the time when writing
+//   death tests; though it lacks many features you can find in PCRE
+//   or POSIX extended regex syntax.  For example, we don't support
+//   union ("x|y"), grouping ("(xy)"), brackets ("[xy]"), and
+//   repetition count ("x{5,7}"), among others.
+//
+//   Below is the syntax that we do support.  We chose it to be a
+//   subset of both PCRE and POSIX extended regex, so it's easy to
+//   learn wherever you come from.  In the following: 'A' denotes a
+//   literal character, period (.), or a single \\ escape sequence;
+//   'x' and 'y' denote regular expressions; 'm' and 'n' are for
+//   natural numbers.
+//
+//     c     matches any literal character c
+//     \\d   matches any decimal digit
+//     \\D   matches any character that's not a decimal digit
+//     \\f   matches \f
+//     \\n   matches \n
+//     \\r   matches \r
+//     \\s   matches any ASCII whitespace, including \n
+//     \\S   matches any character that's not a whitespace
+//     \\t   matches \t
+//     \\v   matches \v
+//     \\w   matches any letter, _, or decimal digit
+//     \\W   matches any character that \\w doesn't match
+//     \\c   matches any literal character c, which must be a punctuation
+//     .     matches any single character except \n
+//     A?    matches 0 or 1 occurrences of A
+//     A*    matches 0 or many occurrences of A
+//     A+    matches 1 or many occurrences of A
+//     ^     matches the beginning of a string (not that of each line)
+//     $     matches the end of a string (not that of each line)
+//     xy    matches x followed by y
+//
+//   If you accidentally use PCRE or POSIX extended regex features
+//   not implemented by us, you will get a run-time failure.  In that
+//   case, please try to rewrite your regular expression within the
+//   above syntax.
+//
+//   This implementation is *not* meant to be as highly tuned or robust
+//   as a compiled regex library, but should perform well enough for a
+//   death test, which already incurs significant overhead by launching
+//   a child process.
+//
+// Known caveats:
+//
+//   A "threadsafe" style death test obtains the path to the test
+//   program from argv[0] and re-executes it in the sub-process.  For
+//   simplicity, the current implementation doesn't search the PATH
+//   when launching the sub-process.  This means that the user must
+//   invoke the test program via a path that contains at least one
+//   path separator (e.g. path/to/foo_test and
+//   /absolute/path/to/bar_test are fine, but foo_test is not).  This
+//   is rarely a problem as people usually don't put the test binary
+//   directory in PATH.
+//
+// TODO(wan@google.com): make thread-safe death tests search the PATH.
+
+// Asserts that a given statement causes the program to exit, with an
+// integer exit status that satisfies predicate, and emitting error output
+// that matches regex.
+# define ASSERT_EXIT(statement, predicate, regex) \
+    GTEST_DEATH_TEST_(statement, predicate, regex, GTEST_FATAL_FAILURE_)
+
+// Like ASSERT_EXIT, but continues on to successive tests in the
+// test case, if any:
+# define EXPECT_EXIT(statement, predicate, regex) \
+    GTEST_DEATH_TEST_(statement, predicate, regex, GTEST_NONFATAL_FAILURE_)
+
+// Asserts that a given statement causes the program to exit, either by
+// explicitly exiting with a nonzero exit code or being killed by a
+// signal, and emitting error output that matches regex.
+# define ASSERT_DEATH(statement, regex) \
+    ASSERT_EXIT(statement, ::testing::internal::ExitedUnsuccessfully, regex)
+
+// Like ASSERT_DEATH, but continues on to successive tests in the
+// test case, if any:
+# define EXPECT_DEATH(statement, regex) \
+    EXPECT_EXIT(statement, ::testing::internal::ExitedUnsuccessfully, regex)
+
+// Two predicate classes that can be used in {ASSERT,EXPECT}_EXIT*:
+
+// Tests that an exit code describes a normal exit with a given exit code.
+class GTEST_API_ ExitedWithCode {
+ public:
+  explicit ExitedWithCode(int exit_code);
+  bool operator()(int exit_status) const;
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ExitedWithCode& other);
+
+  const int exit_code_;
+};
+
+# if !GTEST_OS_WINDOWS
+// Tests that an exit code describes an exit due to termination by a
+// given signal.
+class GTEST_API_ KilledBySignal {
+ public:
+  explicit KilledBySignal(int signum);
+  bool operator()(int exit_status) const;
+ private:
+  const int signum_;
+};
+# endif  // !GTEST_OS_WINDOWS
+
+// EXPECT_DEBUG_DEATH asserts that the given statements die in debug mode.
+// The death testing framework causes this to have interesting semantics,
+// since the sideeffects of the call are only visible in opt mode, and not
+// in debug mode.
+//
+// In practice, this can be used to test functions that utilize the
+// LOG(DFATAL) macro using the following style:
+//
+// int DieInDebugOr12(int* sideeffect) {
+//   if (sideeffect) {
+//     *sideeffect = 12;
+//   }
+//   LOG(DFATAL) << "death";
+//   return 12;
+// }
+//
+// TEST(TestCase, TestDieOr12WorksInDgbAndOpt) {
+//   int sideeffect = 0;
+//   // Only asserts in dbg.
+//   EXPECT_DEBUG_DEATH(DieInDebugOr12(&sideeffect), "death");
+//
+// #ifdef NDEBUG
+//   // opt-mode has sideeffect visible.
+//   EXPECT_EQ(12, sideeffect);
+// #else
+//   // dbg-mode no visible sideeffect.
+//   EXPECT_EQ(0, sideeffect);
+// #endif
+// }
+//
+// This will assert that DieInDebugReturn12InOpt() crashes in debug
+// mode, usually due to a DCHECK or LOG(DFATAL), but returns the
+// appropriate fallback value (12 in this case) in opt mode. If you
+// need to test that a function has appropriate side-effects in opt
+// mode, include assertions against the side-effects.  A general
+// pattern for this is:
+//
+// EXPECT_DEBUG_DEATH({
+//   // Side-effects here will have an effect after this statement in
+//   // opt mode, but none in debug mode.
+//   EXPECT_EQ(12, DieInDebugOr12(&sideeffect));
+// }, "death");
+//
+# ifdef NDEBUG
+
+#  define EXPECT_DEBUG_DEATH(statement, regex) \
+  GTEST_EXECUTE_STATEMENT_(statement, regex)
+
+#  define ASSERT_DEBUG_DEATH(statement, regex) \
+  GTEST_EXECUTE_STATEMENT_(statement, regex)
+
+# else
+
+#  define EXPECT_DEBUG_DEATH(statement, regex) \
+  EXPECT_DEATH(statement, regex)
+
+#  define ASSERT_DEBUG_DEATH(statement, regex) \
+  ASSERT_DEATH(statement, regex)
+
+# endif  // NDEBUG for EXPECT_DEBUG_DEATH
+#endif  // GTEST_HAS_DEATH_TEST
+
+// EXPECT_DEATH_IF_SUPPORTED(statement, regex) and
+// ASSERT_DEATH_IF_SUPPORTED(statement, regex) expand to real death tests if
+// death tests are supported; otherwise they just issue a warning.  This is
+// useful when you are combining death test assertions with normal test
+// assertions in one test.
+#if GTEST_HAS_DEATH_TEST
+# define EXPECT_DEATH_IF_SUPPORTED(statement, regex) \
+    EXPECT_DEATH(statement, regex)
+# define ASSERT_DEATH_IF_SUPPORTED(statement, regex) \
+    ASSERT_DEATH(statement, regex)
+#else
+# define EXPECT_DEATH_IF_SUPPORTED(statement, regex) \
+    GTEST_UNSUPPORTED_DEATH_TEST_(statement, regex, )
+# define ASSERT_DEATH_IF_SUPPORTED(statement, regex) \
+    GTEST_UNSUPPORTED_DEATH_TEST_(statement, regex, return)
+#endif
+
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_DEATH_TEST_H_
+// This file was GENERATED by command:
+//     pump.py gtest-param-test.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: vladl@google.com (Vlad Losev)
+//
+// Macros and functions for implementing parameterized tests
+// in Google C++ Testing Framework (Google Test)
+//
+// This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
+
+
+// Value-parameterized tests allow you to test your code with different
+// parameters without writing multiple copies of the same test.
+//
+// Here is how you use value-parameterized tests:
+
+#if 0
+
+// To write value-parameterized tests, first you should define a fixture
+// class. It is usually derived from testing::TestWithParam (see below for
+// another inheritance scheme that's sometimes useful in more complicated
+// class hierarchies), where the type of your parameter values.
+// TestWithParam is itself derived from testing::Test. T can be any
+// copyable type. If it's a raw pointer, you are responsible for managing the
+// lifespan of the pointed values.
+
+class FooTest : public ::testing::TestWithParam {
+  // You can implement all the usual class fixture members here.
+};
+
+// Then, use the TEST_P macro to define as many parameterized tests
+// for this fixture as you want. The _P suffix is for "parameterized"
+// or "pattern", whichever you prefer to think.
+
+TEST_P(FooTest, DoesBlah) {
+  // Inside a test, access the test parameter with the GetParam() method
+  // of the TestWithParam class:
+  EXPECT_TRUE(foo.Blah(GetParam()));
+  ...
+}
+
+TEST_P(FooTest, HasBlahBlah) {
+  ...
+}
+
+// Finally, you can use INSTANTIATE_TEST_CASE_P to instantiate the test
+// case with any set of parameters you want. Google Test defines a number
+// of functions for generating test parameters. They return what we call
+// (surprise!) parameter generators. Here is a  summary of them, which
+// are all in the testing namespace:
+//
+//
+//  Range(begin, end [, step]) - Yields values {begin, begin+step,
+//                               begin+step+step, ...}. The values do not
+//                               include end. step defaults to 1.
+//  Values(v1, v2, ..., vN)    - Yields values {v1, v2, ..., vN}.
+//  ValuesIn(container)        - Yields values from a C-style array, an STL
+//  ValuesIn(begin,end)          container, or an iterator range [begin, end).
+//  Bool()                     - Yields sequence {false, true}.
+//  Combine(g1, g2, ..., gN)   - Yields all combinations (the Cartesian product
+//                               for the math savvy) of the values generated
+//                               by the N generators.
+//
+// For more details, see comments at the definitions of these functions below
+// in this file.
+//
+// The following statement will instantiate tests from the FooTest test case
+// each with parameter values "meeny", "miny", and "moe".
+
+INSTANTIATE_TEST_CASE_P(InstantiationName,
+                        FooTest,
+                        Values("meeny", "miny", "moe"));
+
+// To distinguish different instances of the pattern, (yes, you
+// can instantiate it more then once) the first argument to the
+// INSTANTIATE_TEST_CASE_P macro is a prefix that will be added to the
+// actual test case name. Remember to pick unique prefixes for different
+// instantiations. The tests from the instantiation above will have
+// these names:
+//
+//    * InstantiationName/FooTest.DoesBlah/0 for "meeny"
+//    * InstantiationName/FooTest.DoesBlah/1 for "miny"
+//    * InstantiationName/FooTest.DoesBlah/2 for "moe"
+//    * InstantiationName/FooTest.HasBlahBlah/0 for "meeny"
+//    * InstantiationName/FooTest.HasBlahBlah/1 for "miny"
+//    * InstantiationName/FooTest.HasBlahBlah/2 for "moe"
+//
+// You can use these names in --gtest_filter.
+//
+// This statement will instantiate all tests from FooTest again, each
+// with parameter values "cat" and "dog":
+
+const char* pets[] = {"cat", "dog"};
+INSTANTIATE_TEST_CASE_P(AnotherInstantiationName, FooTest, ValuesIn(pets));
+
+// The tests from the instantiation above will have these names:
+//
+//    * AnotherInstantiationName/FooTest.DoesBlah/0 for "cat"
+//    * AnotherInstantiationName/FooTest.DoesBlah/1 for "dog"
+//    * AnotherInstantiationName/FooTest.HasBlahBlah/0 for "cat"
+//    * AnotherInstantiationName/FooTest.HasBlahBlah/1 for "dog"
+//
+// Please note that INSTANTIATE_TEST_CASE_P will instantiate all tests
+// in the given test case, whether their definitions come before or
+// AFTER the INSTANTIATE_TEST_CASE_P statement.
+//
+// Please also note that generator expressions (including parameters to the
+// generators) are evaluated in InitGoogleTest(), after main() has started.
+// This allows the user on one hand, to adjust generator parameters in order
+// to dynamically determine a set of tests to run and on the other hand,
+// give the user a chance to inspect the generated tests with Google Test
+// reflection API before RUN_ALL_TESTS() is executed.
+//
+// You can see samples/sample7_unittest.cc and samples/sample8_unittest.cc
+// for more examples.
+//
+// In the future, we plan to publish the API for defining new parameter
+// generators. But for now this interface remains part of the internal
+// implementation and is subject to change.
+//
+//
+// A parameterized test fixture must be derived from testing::Test and from
+// testing::WithParamInterface, where T is the type of the parameter
+// values. Inheriting from TestWithParam satisfies that requirement because
+// TestWithParam inherits from both Test and WithParamInterface. In more
+// complicated hierarchies, however, it is occasionally useful to inherit
+// separately from Test and WithParamInterface. For example:
+
+class BaseTest : public ::testing::Test {
+  // You can inherit all the usual members for a non-parameterized test
+  // fixture here.
+};
+
+class DerivedTest : public BaseTest, public ::testing::WithParamInterface {
+  // The usual test fixture members go here too.
+};
+
+TEST_F(BaseTest, HasFoo) {
+  // This is an ordinary non-parameterized test.
+}
+
+TEST_P(DerivedTest, DoesBlah) {
+  // GetParam works just the same here as if you inherit from TestWithParam.
+  EXPECT_TRUE(foo.Blah(GetParam()));
+}
+
+#endif  // 0
+
+
+#if !GTEST_OS_SYMBIAN
+# include 
+#endif
+
+// scripts/fuse_gtest.py depends on gtest's own header being #included
+// *unconditionally*.  Therefore these #includes cannot be moved
+// inside #if GTEST_HAS_PARAM_TEST.
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: vladl@google.com (Vlad Losev)
+
+// Type and function utilities for implementing parameterized tests.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_H_
+
+#include 
+#include 
+#include 
+
+// scripts/fuse_gtest.py depends on gtest's own header being #included
+// *unconditionally*.  Therefore these #includes cannot be moved
+// inside #if GTEST_HAS_PARAM_TEST.
+// Copyright 2003 Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: Dan Egnor (egnor@google.com)
+//
+// A "smart" pointer type with reference tracking.  Every pointer to a
+// particular object is kept on a circular linked list.  When the last pointer
+// to an object is destroyed or reassigned, the object is deleted.
+//
+// Used properly, this deletes the object when the last reference goes away.
+// There are several caveats:
+// - Like all reference counting schemes, cycles lead to leaks.
+// - Each smart pointer is actually two pointers (8 bytes instead of 4).
+// - Every time a pointer is assigned, the entire list of pointers to that
+//   object is traversed.  This class is therefore NOT SUITABLE when there
+//   will often be more than two or three pointers to a particular object.
+// - References are only tracked as long as linked_ptr<> objects are copied.
+//   If a linked_ptr<> is converted to a raw pointer and back, BAD THINGS
+//   will happen (double deletion).
+//
+// A good use of this class is storing object references in STL containers.
+// You can safely put linked_ptr<> in a vector<>.
+// Other uses may not be as good.
+//
+// Note: If you use an incomplete type with linked_ptr<>, the class
+// *containing* linked_ptr<> must have a constructor and destructor (even
+// if they do nothing!).
+//
+// Bill Gibbons suggested we use something like this.
+//
+// Thread Safety:
+//   Unlike other linked_ptr implementations, in this implementation
+//   a linked_ptr object is thread-safe in the sense that:
+//     - it's safe to copy linked_ptr objects concurrently,
+//     - it's safe to copy *from* a linked_ptr and read its underlying
+//       raw pointer (e.g. via get()) concurrently, and
+//     - it's safe to write to two linked_ptrs that point to the same
+//       shared object concurrently.
+// TODO(wan@google.com): rename this to safe_linked_ptr to avoid
+// confusion with normal linked_ptr.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_LINKED_PTR_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_LINKED_PTR_H_
+
+#include 
+#include 
+
+
+namespace testing {
+namespace internal {
+
+// Protects copying of all linked_ptr objects.
+GTEST_API_ GTEST_DECLARE_STATIC_MUTEX_(g_linked_ptr_mutex);
+
+// This is used internally by all instances of linked_ptr<>.  It needs to be
+// a non-template class because different types of linked_ptr<> can refer to
+// the same object (linked_ptr(obj) vs linked_ptr(obj)).
+// So, it needs to be possible for different types of linked_ptr to participate
+// in the same circular linked list, so we need a single class type here.
+//
+// DO NOT USE THIS CLASS DIRECTLY YOURSELF.  Use linked_ptr.
+class linked_ptr_internal {
+ public:
+  // Create a new circle that includes only this instance.
+  void join_new() {
+    next_ = this;
+  }
+
+  // Many linked_ptr operations may change p.link_ for some linked_ptr
+  // variable p in the same circle as this object.  Therefore we need
+  // to prevent two such operations from occurring concurrently.
+  //
+  // Note that different types of linked_ptr objects can coexist in a
+  // circle (e.g. linked_ptr, linked_ptr, and
+  // linked_ptr).  Therefore we must use a single mutex to
+  // protect all linked_ptr objects.  This can create serious
+  // contention in production code, but is acceptable in a testing
+  // framework.
+
+  // Join an existing circle.
+  void join(linked_ptr_internal const* ptr)
+      GTEST_LOCK_EXCLUDED_(g_linked_ptr_mutex) {
+    MutexLock lock(&g_linked_ptr_mutex);
+
+    linked_ptr_internal const* p = ptr;
+    while (p->next_ != ptr) p = p->next_;
+    p->next_ = this;
+    next_ = ptr;
+  }
+
+  // Leave whatever circle we're part of.  Returns true if we were the
+  // last member of the circle.  Once this is done, you can join() another.
+  bool depart()
+      GTEST_LOCK_EXCLUDED_(g_linked_ptr_mutex) {
+    MutexLock lock(&g_linked_ptr_mutex);
+
+    if (next_ == this) return true;
+    linked_ptr_internal const* p = next_;
+    while (p->next_ != this) p = p->next_;
+    p->next_ = next_;
+    return false;
+  }
+
+ private:
+  mutable linked_ptr_internal const* next_;
+};
+
+template 
+class linked_ptr {
+ public:
+  typedef T element_type;
+
+  // Take over ownership of a raw pointer.  This should happen as soon as
+  // possible after the object is created.
+  explicit linked_ptr(T* ptr = NULL) { capture(ptr); }
+  ~linked_ptr() { depart(); }
+
+  // Copy an existing linked_ptr<>, adding ourselves to the list of references.
+  template  linked_ptr(linked_ptr const& ptr) { copy(&ptr); }
+  linked_ptr(linked_ptr const& ptr) {  // NOLINT
+    assert(&ptr != this);
+    copy(&ptr);
+  }
+
+  // Assignment releases the old value and acquires the new.
+  template  linked_ptr& operator=(linked_ptr const& ptr) {
+    depart();
+    copy(&ptr);
+    return *this;
+  }
+
+  linked_ptr& operator=(linked_ptr const& ptr) {
+    if (&ptr != this) {
+      depart();
+      copy(&ptr);
+    }
+    return *this;
+  }
+
+  // Smart pointer members.
+  void reset(T* ptr = NULL) {
+    depart();
+    capture(ptr);
+  }
+  T* get() const { return value_; }
+  T* operator->() const { return value_; }
+  T& operator*() const { return *value_; }
+
+  bool operator==(T* p) const { return value_ == p; }
+  bool operator!=(T* p) const { return value_ != p; }
+  template 
+  bool operator==(linked_ptr const& ptr) const {
+    return value_ == ptr.get();
+  }
+  template 
+  bool operator!=(linked_ptr const& ptr) const {
+    return value_ != ptr.get();
+  }
+
+ private:
+  template 
+  friend class linked_ptr;
+
+  T* value_;
+  linked_ptr_internal link_;
+
+  void depart() {
+    if (link_.depart()) delete value_;
+  }
+
+  void capture(T* ptr) {
+    value_ = ptr;
+    link_.join_new();
+  }
+
+  template  void copy(linked_ptr const* ptr) {
+    value_ = ptr->get();
+    if (value_)
+      link_.join(&ptr->link_);
+    else
+      link_.join_new();
+  }
+};
+
+template inline
+bool operator==(T* ptr, const linked_ptr& x) {
+  return ptr == x.get();
+}
+
+template inline
+bool operator!=(T* ptr, const linked_ptr& x) {
+  return ptr != x.get();
+}
+
+// A function to convert T* into linked_ptr
+// Doing e.g. make_linked_ptr(new FooBarBaz(arg)) is a shorter notation
+// for linked_ptr >(new FooBarBaz(arg))
+template 
+linked_ptr make_linked_ptr(T* ptr) {
+  return linked_ptr(ptr);
+}
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_LINKED_PTR_H_
+// Copyright 2007, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Google Test - The Google C++ Testing Framework
+//
+// This file implements a universal value printer that can print a
+// value of any type T:
+//
+//   void ::testing::internal::UniversalPrinter::Print(value, ostream_ptr);
+//
+// A user can teach this function how to print a class type T by
+// defining either operator<<() or PrintTo() in the namespace that
+// defines T.  More specifically, the FIRST defined function in the
+// following list will be used (assuming T is defined in namespace
+// foo):
+//
+//   1. foo::PrintTo(const T&, ostream*)
+//   2. operator<<(ostream&, const T&) defined in either foo or the
+//      global namespace.
+//
+// If none of the above is defined, it will print the debug string of
+// the value if it is a protocol buffer, or print the raw bytes in the
+// value otherwise.
+//
+// To aid debugging: when T is a reference type, the address of the
+// value is also printed; when T is a (const) char pointer, both the
+// pointer value and the NUL-terminated string it points to are
+// printed.
+//
+// We also provide some convenient wrappers:
+//
+//   // Prints a value to a string.  For a (const or not) char
+//   // pointer, the NUL-terminated string (but not the pointer) is
+//   // printed.
+//   std::string ::testing::PrintToString(const T& value);
+//
+//   // Prints a value tersely: for a reference type, the referenced
+//   // value (but not the address) is printed; for a (const or not) char
+//   // pointer, the NUL-terminated string (but not the pointer) is
+//   // printed.
+//   void ::testing::internal::UniversalTersePrint(const T& value, ostream*);
+//
+//   // Prints value using the type inferred by the compiler.  The difference
+//   // from UniversalTersePrint() is that this function prints both the
+//   // pointer and the NUL-terminated string for a (const or not) char pointer.
+//   void ::testing::internal::UniversalPrint(const T& value, ostream*);
+//
+//   // Prints the fields of a tuple tersely to a string vector, one
+//   // element for each field. Tuple support must be enabled in
+//   // gtest-port.h.
+//   std::vector UniversalTersePrintTupleFieldsToStrings(
+//       const Tuple& value);
+//
+// Known limitation:
+//
+// The print primitives print the elements of an STL-style container
+// using the compiler-inferred type of *iter where iter is a
+// const_iterator of the container.  When const_iterator is an input
+// iterator but not a forward iterator, this inferred type may not
+// match value_type, and the print output may be incorrect.  In
+// practice, this is rarely a problem as for most containers
+// const_iterator is a forward iterator.  We'll fix this if there's an
+// actual need for it.  Note that this fix cannot rely on value_type
+// being defined as many user-defined container types don't have
+// value_type.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PRINTERS_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PRINTERS_H_
+
+#include   // NOLINT
+#include 
+#include 
+#include 
+#include 
+
+namespace testing {
+
+// Definitions in the 'internal' and 'internal2' name spaces are
+// subject to change without notice.  DO NOT USE THEM IN USER CODE!
+namespace internal2 {
+
+// Prints the given number of bytes in the given object to the given
+// ostream.
+GTEST_API_ void PrintBytesInObjectTo(const unsigned char* obj_bytes,
+                                     size_t count,
+                                     ::std::ostream* os);
+
+// For selecting which printer to use when a given type has neither <<
+// nor PrintTo().
+enum TypeKind {
+  kProtobuf,              // a protobuf type
+  kConvertibleToInteger,  // a type implicitly convertible to BiggestInt
+                          // (e.g. a named or unnamed enum type)
+  kOtherType              // anything else
+};
+
+// TypeWithoutFormatter::PrintValue(value, os) is called
+// by the universal printer to print a value of type T when neither
+// operator<< nor PrintTo() is defined for T, where kTypeKind is the
+// "kind" of T as defined by enum TypeKind.
+template 
+class TypeWithoutFormatter {
+ public:
+  // This default version is called when kTypeKind is kOtherType.
+  static void PrintValue(const T& value, ::std::ostream* os) {
+    PrintBytesInObjectTo(reinterpret_cast(&value),
+                         sizeof(value), os);
+  }
+};
+
+// We print a protobuf using its ShortDebugString() when the string
+// doesn't exceed this many characters; otherwise we print it using
+// DebugString() for better readability.
+const size_t kProtobufOneLinerMaxLength = 50;
+
+template 
+class TypeWithoutFormatter {
+ public:
+  static void PrintValue(const T& value, ::std::ostream* os) {
+    const ::testing::internal::string short_str = value.ShortDebugString();
+    const ::testing::internal::string pretty_str =
+        short_str.length() <= kProtobufOneLinerMaxLength ?
+        short_str : ("\n" + value.DebugString());
+    *os << ("<" + pretty_str + ">");
+  }
+};
+
+template 
+class TypeWithoutFormatter {
+ public:
+  // Since T has no << operator or PrintTo() but can be implicitly
+  // converted to BiggestInt, we print it as a BiggestInt.
+  //
+  // Most likely T is an enum type (either named or unnamed), in which
+  // case printing it as an integer is the desired behavior.  In case
+  // T is not an enum, printing it as an integer is the best we can do
+  // given that it has no user-defined printer.
+  static void PrintValue(const T& value, ::std::ostream* os) {
+    const internal::BiggestInt kBigInt = value;
+    *os << kBigInt;
+  }
+};
+
+// Prints the given value to the given ostream.  If the value is a
+// protocol message, its debug string is printed; if it's an enum or
+// of a type implicitly convertible to BiggestInt, it's printed as an
+// integer; otherwise the bytes in the value are printed.  This is
+// what UniversalPrinter::Print() does when it knows nothing about
+// type T and T has neither << operator nor PrintTo().
+//
+// A user can override this behavior for a class type Foo by defining
+// a << operator in the namespace where Foo is defined.
+//
+// We put this operator in namespace 'internal2' instead of 'internal'
+// to simplify the implementation, as much code in 'internal' needs to
+// use << in STL, which would conflict with our own << were it defined
+// in 'internal'.
+//
+// Note that this operator<< takes a generic std::basic_ostream type instead of the more restricted std::ostream.  If
+// we define it to take an std::ostream instead, we'll get an
+// "ambiguous overloads" compiler error when trying to print a type
+// Foo that supports streaming to std::basic_ostream, as the compiler cannot tell whether
+// operator<<(std::ostream&, const T&) or
+// operator<<(std::basic_stream, const Foo&) is more
+// specific.
+template 
+::std::basic_ostream& operator<<(
+    ::std::basic_ostream& os, const T& x) {
+  TypeWithoutFormatter::value ? kProtobuf :
+       internal::ImplicitlyConvertible::value ?
+       kConvertibleToInteger : kOtherType)>::PrintValue(x, &os);
+  return os;
+}
+
+}  // namespace internal2
+}  // namespace testing
+
+// This namespace MUST NOT BE NESTED IN ::testing, or the name look-up
+// magic needed for implementing UniversalPrinter won't work.
+namespace testing_internal {
+
+// Used to print a value that is not an STL-style container when the
+// user doesn't define PrintTo() for it.
+template 
+void DefaultPrintNonContainerTo(const T& value, ::std::ostream* os) {
+  // With the following statement, during unqualified name lookup,
+  // testing::internal2::operator<< appears as if it was declared in
+  // the nearest enclosing namespace that contains both
+  // ::testing_internal and ::testing::internal2, i.e. the global
+  // namespace.  For more details, refer to the C++ Standard section
+  // 7.3.4-1 [namespace.udir].  This allows us to fall back onto
+  // testing::internal2::operator<< in case T doesn't come with a <<
+  // operator.
+  //
+  // We cannot write 'using ::testing::internal2::operator<<;', which
+  // gcc 3.3 fails to compile due to a compiler bug.
+  using namespace ::testing::internal2;  // NOLINT
+
+  // Assuming T is defined in namespace foo, in the next statement,
+  // the compiler will consider all of:
+  //
+  //   1. foo::operator<< (thanks to Koenig look-up),
+  //   2. ::operator<< (as the current namespace is enclosed in ::),
+  //   3. testing::internal2::operator<< (thanks to the using statement above).
+  //
+  // The operator<< whose type matches T best will be picked.
+  //
+  // We deliberately allow #2 to be a candidate, as sometimes it's
+  // impossible to define #1 (e.g. when foo is ::std, defining
+  // anything in it is undefined behavior unless you are a compiler
+  // vendor.).
+  *os << value;
+}
+
+}  // namespace testing_internal
+
+namespace testing {
+namespace internal {
+
+// UniversalPrinter::Print(value, ostream_ptr) prints the given
+// value to the given ostream.  The caller must ensure that
+// 'ostream_ptr' is not NULL, or the behavior is undefined.
+//
+// We define UniversalPrinter as a class template (as opposed to a
+// function template), as we need to partially specialize it for
+// reference types, which cannot be done with function templates.
+template 
+class UniversalPrinter;
+
+template 
+void UniversalPrint(const T& value, ::std::ostream* os);
+
+// Used to print an STL-style container when the user doesn't define
+// a PrintTo() for it.
+template 
+void DefaultPrintTo(IsContainer /* dummy */,
+                    false_type /* is not a pointer */,
+                    const C& container, ::std::ostream* os) {
+  const size_t kMaxCount = 32;  // The maximum number of elements to print.
+  *os << '{';
+  size_t count = 0;
+  for (typename C::const_iterator it = container.begin();
+       it != container.end(); ++it, ++count) {
+    if (count > 0) {
+      *os << ',';
+      if (count == kMaxCount) {  // Enough has been printed.
+        *os << " ...";
+        break;
+      }
+    }
+    *os << ' ';
+    // We cannot call PrintTo(*it, os) here as PrintTo() doesn't
+    // handle *it being a native array.
+    internal::UniversalPrint(*it, os);
+  }
+
+  if (count > 0) {
+    *os << ' ';
+  }
+  *os << '}';
+}
+
+// Used to print a pointer that is neither a char pointer nor a member
+// pointer, when the user doesn't define PrintTo() for it.  (A member
+// variable pointer or member function pointer doesn't really point to
+// a location in the address space.  Their representation is
+// implementation-defined.  Therefore they will be printed as raw
+// bytes.)
+template 
+void DefaultPrintTo(IsNotContainer /* dummy */,
+                    true_type /* is a pointer */,
+                    T* p, ::std::ostream* os) {
+  if (p == NULL) {
+    *os << "NULL";
+  } else {
+    // C++ doesn't allow casting from a function pointer to any object
+    // pointer.
+    //
+    // IsTrue() silences warnings: "Condition is always true",
+    // "unreachable code".
+    if (IsTrue(ImplicitlyConvertible::value)) {
+      // T is not a function type.  We just call << to print p,
+      // relying on ADL to pick up user-defined << for their pointer
+      // types, if any.
+      *os << p;
+    } else {
+      // T is a function type, so '*os << p' doesn't do what we want
+      // (it just prints p as bool).  We want to print p as a const
+      // void*.  However, we cannot cast it to const void* directly,
+      // even using reinterpret_cast, as earlier versions of gcc
+      // (e.g. 3.4.5) cannot compile the cast when p is a function
+      // pointer.  Casting to UInt64 first solves the problem.
+      *os << reinterpret_cast(
+          reinterpret_cast(p));
+    }
+  }
+}
+
+// Used to print a non-container, non-pointer value when the user
+// doesn't define PrintTo() for it.
+template 
+void DefaultPrintTo(IsNotContainer /* dummy */,
+                    false_type /* is not a pointer */,
+                    const T& value, ::std::ostream* os) {
+  ::testing_internal::DefaultPrintNonContainerTo(value, os);
+}
+
+// Prints the given value using the << operator if it has one;
+// otherwise prints the bytes in it.  This is what
+// UniversalPrinter::Print() does when PrintTo() is not specialized
+// or overloaded for type T.
+//
+// A user can override this behavior for a class type Foo by defining
+// an overload of PrintTo() in the namespace where Foo is defined.  We
+// give the user this option as sometimes defining a << operator for
+// Foo is not desirable (e.g. the coding style may prevent doing it,
+// or there is already a << operator but it doesn't do what the user
+// wants).
+template 
+void PrintTo(const T& value, ::std::ostream* os) {
+  // DefaultPrintTo() is overloaded.  The type of its first two
+  // arguments determine which version will be picked.  If T is an
+  // STL-style container, the version for container will be called; if
+  // T is a pointer, the pointer version will be called; otherwise the
+  // generic version will be called.
+  //
+  // Note that we check for container types here, prior to we check
+  // for protocol message types in our operator<<.  The rationale is:
+  //
+  // For protocol messages, we want to give people a chance to
+  // override Google Mock's format by defining a PrintTo() or
+  // operator<<.  For STL containers, other formats can be
+  // incompatible with Google Mock's format for the container
+  // elements; therefore we check for container types here to ensure
+  // that our format is used.
+  //
+  // The second argument of DefaultPrintTo() is needed to bypass a bug
+  // in Symbian's C++ compiler that prevents it from picking the right
+  // overload between:
+  //
+  //   PrintTo(const T& x, ...);
+  //   PrintTo(T* x, ...);
+  DefaultPrintTo(IsContainerTest(0), is_pointer(), value, os);
+}
+
+// The following list of PrintTo() overloads tells
+// UniversalPrinter::Print() how to print standard types (built-in
+// types, strings, plain arrays, and pointers).
+
+// Overloads for various char types.
+GTEST_API_ void PrintTo(unsigned char c, ::std::ostream* os);
+GTEST_API_ void PrintTo(signed char c, ::std::ostream* os);
+inline void PrintTo(char c, ::std::ostream* os) {
+  // When printing a plain char, we always treat it as unsigned.  This
+  // way, the output won't be affected by whether the compiler thinks
+  // char is signed or not.
+  PrintTo(static_cast(c), os);
+}
+
+// Overloads for other simple built-in types.
+inline void PrintTo(bool x, ::std::ostream* os) {
+  *os << (x ? "true" : "false");
+}
+
+// Overload for wchar_t type.
+// Prints a wchar_t as a symbol if it is printable or as its internal
+// code otherwise and also as its decimal code (except for L'\0').
+// The L'\0' char is printed as "L'\\0'". The decimal code is printed
+// as signed integer when wchar_t is implemented by the compiler
+// as a signed type and is printed as an unsigned integer when wchar_t
+// is implemented as an unsigned type.
+GTEST_API_ void PrintTo(wchar_t wc, ::std::ostream* os);
+
+// Overloads for C strings.
+GTEST_API_ void PrintTo(const char* s, ::std::ostream* os);
+inline void PrintTo(char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_(s), os);
+}
+
+// signed/unsigned char is often used for representing binary data, so
+// we print pointers to it as void* to be safe.
+inline void PrintTo(const signed char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_(s), os);
+}
+inline void PrintTo(signed char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_(s), os);
+}
+inline void PrintTo(const unsigned char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_(s), os);
+}
+inline void PrintTo(unsigned char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_(s), os);
+}
+
+// MSVC can be configured to define wchar_t as a typedef of unsigned
+// short.  It defines _NATIVE_WCHAR_T_DEFINED when wchar_t is a native
+// type.  When wchar_t is a typedef, defining an overload for const
+// wchar_t* would cause unsigned short* be printed as a wide string,
+// possibly causing invalid memory accesses.
+#if !defined(_MSC_VER) || defined(_NATIVE_WCHAR_T_DEFINED)
+// Overloads for wide C strings
+GTEST_API_ void PrintTo(const wchar_t* s, ::std::ostream* os);
+inline void PrintTo(wchar_t* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_(s), os);
+}
+#endif
+
+// Overload for C arrays.  Multi-dimensional arrays are printed
+// properly.
+
+// Prints the given number of elements in an array, without printing
+// the curly braces.
+template 
+void PrintRawArrayTo(const T a[], size_t count, ::std::ostream* os) {
+  UniversalPrint(a[0], os);
+  for (size_t i = 1; i != count; i++) {
+    *os << ", ";
+    UniversalPrint(a[i], os);
+  }
+}
+
+// Overloads for ::string and ::std::string.
+#if GTEST_HAS_GLOBAL_STRING
+GTEST_API_ void PrintStringTo(const ::string&s, ::std::ostream* os);
+inline void PrintTo(const ::string& s, ::std::ostream* os) {
+  PrintStringTo(s, os);
+}
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+GTEST_API_ void PrintStringTo(const ::std::string&s, ::std::ostream* os);
+inline void PrintTo(const ::std::string& s, ::std::ostream* os) {
+  PrintStringTo(s, os);
+}
+
+// Overloads for ::wstring and ::std::wstring.
+#if GTEST_HAS_GLOBAL_WSTRING
+GTEST_API_ void PrintWideStringTo(const ::wstring&s, ::std::ostream* os);
+inline void PrintTo(const ::wstring& s, ::std::ostream* os) {
+  PrintWideStringTo(s, os);
+}
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+#if GTEST_HAS_STD_WSTRING
+GTEST_API_ void PrintWideStringTo(const ::std::wstring&s, ::std::ostream* os);
+inline void PrintTo(const ::std::wstring& s, ::std::ostream* os) {
+  PrintWideStringTo(s, os);
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+#if GTEST_HAS_TR1_TUPLE
+// Overload for ::std::tr1::tuple.  Needed for printing function arguments,
+// which are packed as tuples.
+
+// Helper function for printing a tuple.  T must be instantiated with
+// a tuple type.
+template 
+void PrintTupleTo(const T& t, ::std::ostream* os);
+
+// Overloaded PrintTo() for tuples of various arities.  We support
+// tuples of up-to 10 fields.  The following implementation works
+// regardless of whether tr1::tuple is implemented using the
+// non-standard variadic template feature or not.
+
+inline void PrintTo(const ::std::tr1::tuple<>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template 
+void PrintTo(const ::std::tr1::tuple& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template 
+void PrintTo(const ::std::tr1::tuple& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template 
+void PrintTo(const ::std::tr1::tuple& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template 
+void PrintTo(const ::std::tr1::tuple& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template 
+void PrintTo(const ::std::tr1::tuple& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template 
+void PrintTo(const ::std::tr1::tuple& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template 
+void PrintTo(const ::std::tr1::tuple& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template 
+void PrintTo(const ::std::tr1::tuple& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template 
+void PrintTo(const ::std::tr1::tuple& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template 
+void PrintTo(
+    const ::std::tr1::tuple& t,
+    ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+#endif  // GTEST_HAS_TR1_TUPLE
+
+// Overload for std::pair.
+template 
+void PrintTo(const ::std::pair& value, ::std::ostream* os) {
+  *os << '(';
+  // We cannot use UniversalPrint(value.first, os) here, as T1 may be
+  // a reference type.  The same for printing value.second.
+  UniversalPrinter::Print(value.first, os);
+  *os << ", ";
+  UniversalPrinter::Print(value.second, os);
+  *os << ')';
+}
+
+// Implements printing a non-reference type T by letting the compiler
+// pick the right overload of PrintTo() for T.
+template 
+class UniversalPrinter {
+ public:
+  // MSVC warns about adding const to a function type, so we want to
+  // disable the warning.
+#ifdef _MSC_VER
+# pragma warning(push)          // Saves the current warning state.
+# pragma warning(disable:4180)  // Temporarily disables warning 4180.
+#endif  // _MSC_VER
+
+  // Note: we deliberately don't call this PrintTo(), as that name
+  // conflicts with ::testing::internal::PrintTo in the body of the
+  // function.
+  static void Print(const T& value, ::std::ostream* os) {
+    // By default, ::testing::internal::PrintTo() is used for printing
+    // the value.
+    //
+    // Thanks to Koenig look-up, if T is a class and has its own
+    // PrintTo() function defined in its namespace, that function will
+    // be visible here.  Since it is more specific than the generic ones
+    // in ::testing::internal, it will be picked by the compiler in the
+    // following statement - exactly what we want.
+    PrintTo(value, os);
+  }
+
+#ifdef _MSC_VER
+# pragma warning(pop)           // Restores the warning state.
+#endif  // _MSC_VER
+};
+
+// UniversalPrintArray(begin, len, os) prints an array of 'len'
+// elements, starting at address 'begin'.
+template 
+void UniversalPrintArray(const T* begin, size_t len, ::std::ostream* os) {
+  if (len == 0) {
+    *os << "{}";
+  } else {
+    *os << "{ ";
+    const size_t kThreshold = 18;
+    const size_t kChunkSize = 8;
+    // If the array has more than kThreshold elements, we'll have to
+    // omit some details by printing only the first and the last
+    // kChunkSize elements.
+    // TODO(wan@google.com): let the user control the threshold using a flag.
+    if (len <= kThreshold) {
+      PrintRawArrayTo(begin, len, os);
+    } else {
+      PrintRawArrayTo(begin, kChunkSize, os);
+      *os << ", ..., ";
+      PrintRawArrayTo(begin + len - kChunkSize, kChunkSize, os);
+    }
+    *os << " }";
+  }
+}
+// This overload prints a (const) char array compactly.
+GTEST_API_ void UniversalPrintArray(
+    const char* begin, size_t len, ::std::ostream* os);
+
+// This overload prints a (const) wchar_t array compactly.
+GTEST_API_ void UniversalPrintArray(
+    const wchar_t* begin, size_t len, ::std::ostream* os);
+
+// Implements printing an array type T[N].
+template 
+class UniversalPrinter {
+ public:
+  // Prints the given array, omitting some elements when there are too
+  // many.
+  static void Print(const T (&a)[N], ::std::ostream* os) {
+    UniversalPrintArray(a, N, os);
+  }
+};
+
+// Implements printing a reference type T&.
+template 
+class UniversalPrinter {
+ public:
+  // MSVC warns about adding const to a function type, so we want to
+  // disable the warning.
+#ifdef _MSC_VER
+# pragma warning(push)          // Saves the current warning state.
+# pragma warning(disable:4180)  // Temporarily disables warning 4180.
+#endif  // _MSC_VER
+
+  static void Print(const T& value, ::std::ostream* os) {
+    // Prints the address of the value.  We use reinterpret_cast here
+    // as static_cast doesn't compile when T is a function type.
+    *os << "@" << reinterpret_cast(&value) << " ";
+
+    // Then prints the value itself.
+    UniversalPrint(value, os);
+  }
+
+#ifdef _MSC_VER
+# pragma warning(pop)           // Restores the warning state.
+#endif  // _MSC_VER
+};
+
+// Prints a value tersely: for a reference type, the referenced value
+// (but not the address) is printed; for a (const) char pointer, the
+// NUL-terminated string (but not the pointer) is printed.
+
+template 
+class UniversalTersePrinter {
+ public:
+  static void Print(const T& value, ::std::ostream* os) {
+    UniversalPrint(value, os);
+  }
+};
+template 
+class UniversalTersePrinter {
+ public:
+  static void Print(const T& value, ::std::ostream* os) {
+    UniversalPrint(value, os);
+  }
+};
+template 
+class UniversalTersePrinter {
+ public:
+  static void Print(const T (&value)[N], ::std::ostream* os) {
+    UniversalPrinter::Print(value, os);
+  }
+};
+template <>
+class UniversalTersePrinter {
+ public:
+  static void Print(const char* str, ::std::ostream* os) {
+    if (str == NULL) {
+      *os << "NULL";
+    } else {
+      UniversalPrint(string(str), os);
+    }
+  }
+};
+template <>
+class UniversalTersePrinter {
+ public:
+  static void Print(char* str, ::std::ostream* os) {
+    UniversalTersePrinter::Print(str, os);
+  }
+};
+
+#if GTEST_HAS_STD_WSTRING
+template <>
+class UniversalTersePrinter {
+ public:
+  static void Print(const wchar_t* str, ::std::ostream* os) {
+    if (str == NULL) {
+      *os << "NULL";
+    } else {
+      UniversalPrint(::std::wstring(str), os);
+    }
+  }
+};
+#endif
+
+template <>
+class UniversalTersePrinter {
+ public:
+  static void Print(wchar_t* str, ::std::ostream* os) {
+    UniversalTersePrinter::Print(str, os);
+  }
+};
+
+template 
+void UniversalTersePrint(const T& value, ::std::ostream* os) {
+  UniversalTersePrinter::Print(value, os);
+}
+
+// Prints a value using the type inferred by the compiler.  The
+// difference between this and UniversalTersePrint() is that for a
+// (const) char pointer, this prints both the pointer and the
+// NUL-terminated string.
+template 
+void UniversalPrint(const T& value, ::std::ostream* os) {
+  // A workarond for the bug in VC++ 7.1 that prevents us from instantiating
+  // UniversalPrinter with T directly.
+  typedef T T1;
+  UniversalPrinter::Print(value, os);
+}
+
+#if GTEST_HAS_TR1_TUPLE
+typedef ::std::vector Strings;
+
+// This helper template allows PrintTo() for tuples and
+// UniversalTersePrintTupleFieldsToStrings() to be defined by
+// induction on the number of tuple fields.  The idea is that
+// TuplePrefixPrinter::PrintPrefixTo(t, os) prints the first N
+// fields in tuple t, and can be defined in terms of
+// TuplePrefixPrinter.
+
+// The inductive case.
+template 
+struct TuplePrefixPrinter {
+  // Prints the first N fields of a tuple.
+  template 
+  static void PrintPrefixTo(const Tuple& t, ::std::ostream* os) {
+    TuplePrefixPrinter::PrintPrefixTo(t, os);
+    *os << ", ";
+    UniversalPrinter::type>
+        ::Print(::std::tr1::get(t), os);
+  }
+
+  // Tersely prints the first N fields of a tuple to a string vector,
+  // one element for each field.
+  template 
+  static void TersePrintPrefixToStrings(const Tuple& t, Strings* strings) {
+    TuplePrefixPrinter::TersePrintPrefixToStrings(t, strings);
+    ::std::stringstream ss;
+    UniversalTersePrint(::std::tr1::get(t), &ss);
+    strings->push_back(ss.str());
+  }
+};
+
+// Base cases.
+template <>
+struct TuplePrefixPrinter<0> {
+  template 
+  static void PrintPrefixTo(const Tuple&, ::std::ostream*) {}
+
+  template 
+  static void TersePrintPrefixToStrings(const Tuple&, Strings*) {}
+};
+// We have to specialize the entire TuplePrefixPrinter<> class
+// template here, even though the definition of
+// TersePrintPrefixToStrings() is the same as the generic version, as
+// Embarcadero (formerly CodeGear, formerly Borland) C++ doesn't
+// support specializing a method template of a class template.
+template <>
+struct TuplePrefixPrinter<1> {
+  template 
+  static void PrintPrefixTo(const Tuple& t, ::std::ostream* os) {
+    UniversalPrinter::type>::
+        Print(::std::tr1::get<0>(t), os);
+  }
+
+  template 
+  static void TersePrintPrefixToStrings(const Tuple& t, Strings* strings) {
+    ::std::stringstream ss;
+    UniversalTersePrint(::std::tr1::get<0>(t), &ss);
+    strings->push_back(ss.str());
+  }
+};
+
+// Helper function for printing a tuple.  T must be instantiated with
+// a tuple type.
+template 
+void PrintTupleTo(const T& t, ::std::ostream* os) {
+  *os << "(";
+  TuplePrefixPrinter< ::std::tr1::tuple_size::value>::
+      PrintPrefixTo(t, os);
+  *os << ")";
+}
+
+// Prints the fields of a tuple tersely to a string vector, one
+// element for each field.  See the comment before
+// UniversalTersePrint() for how we define "tersely".
+template 
+Strings UniversalTersePrintTupleFieldsToStrings(const Tuple& value) {
+  Strings result;
+  TuplePrefixPrinter< ::std::tr1::tuple_size::value>::
+      TersePrintPrefixToStrings(value, &result);
+  return result;
+}
+#endif  // GTEST_HAS_TR1_TUPLE
+
+}  // namespace internal
+
+template 
+::std::string PrintToString(const T& value) {
+  ::std::stringstream ss;
+  internal::UniversalTersePrinter::Print(value, &ss);
+  return ss.str();
+}
+
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PRINTERS_H_
+
+#if GTEST_HAS_PARAM_TEST
+
+namespace testing {
+namespace internal {
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Outputs a message explaining invalid registration of different
+// fixture class for the same test case. This may happen when
+// TEST_P macro is used to define two tests with the same name
+// but in different namespaces.
+GTEST_API_ void ReportInvalidTestCaseType(const char* test_case_name,
+                                          const char* file, int line);
+
+template  class ParamGeneratorInterface;
+template  class ParamGenerator;
+
+// Interface for iterating over elements provided by an implementation
+// of ParamGeneratorInterface.
+template 
+class ParamIteratorInterface {
+ public:
+  virtual ~ParamIteratorInterface() {}
+  // A pointer to the base generator instance.
+  // Used only for the purposes of iterator comparison
+  // to make sure that two iterators belong to the same generator.
+  virtual const ParamGeneratorInterface* BaseGenerator() const = 0;
+  // Advances iterator to point to the next element
+  // provided by the generator. The caller is responsible
+  // for not calling Advance() on an iterator equal to
+  // BaseGenerator()->End().
+  virtual void Advance() = 0;
+  // Clones the iterator object. Used for implementing copy semantics
+  // of ParamIterator.
+  virtual ParamIteratorInterface* Clone() const = 0;
+  // Dereferences the current iterator and provides (read-only) access
+  // to the pointed value. It is the caller's responsibility not to call
+  // Current() on an iterator equal to BaseGenerator()->End().
+  // Used for implementing ParamGenerator::operator*().
+  virtual const T* Current() const = 0;
+  // Determines whether the given iterator and other point to the same
+  // element in the sequence generated by the generator.
+  // Used for implementing ParamGenerator::operator==().
+  virtual bool Equals(const ParamIteratorInterface& other) const = 0;
+};
+
+// Class iterating over elements provided by an implementation of
+// ParamGeneratorInterface. It wraps ParamIteratorInterface
+// and implements the const forward iterator concept.
+template 
+class ParamIterator {
+ public:
+  typedef T value_type;
+  typedef const T& reference;
+  typedef ptrdiff_t difference_type;
+
+  // ParamIterator assumes ownership of the impl_ pointer.
+  ParamIterator(const ParamIterator& other) : impl_(other.impl_->Clone()) {}
+  ParamIterator& operator=(const ParamIterator& other) {
+    if (this != &other)
+      impl_.reset(other.impl_->Clone());
+    return *this;
+  }
+
+  const T& operator*() const { return *impl_->Current(); }
+  const T* operator->() const { return impl_->Current(); }
+  // Prefix version of operator++.
+  ParamIterator& operator++() {
+    impl_->Advance();
+    return *this;
+  }
+  // Postfix version of operator++.
+  ParamIterator operator++(int /*unused*/) {
+    ParamIteratorInterface* clone = impl_->Clone();
+    impl_->Advance();
+    return ParamIterator(clone);
+  }
+  bool operator==(const ParamIterator& other) const {
+    return impl_.get() == other.impl_.get() || impl_->Equals(*other.impl_);
+  }
+  bool operator!=(const ParamIterator& other) const {
+    return !(*this == other);
+  }
+
+ private:
+  friend class ParamGenerator;
+  explicit ParamIterator(ParamIteratorInterface* impl) : impl_(impl) {}
+  scoped_ptr > impl_;
+};
+
+// ParamGeneratorInterface is the binary interface to access generators
+// defined in other translation units.
+template 
+class ParamGeneratorInterface {
+ public:
+  typedef T ParamType;
+
+  virtual ~ParamGeneratorInterface() {}
+
+  // Generator interface definition
+  virtual ParamIteratorInterface* Begin() const = 0;
+  virtual ParamIteratorInterface* End() const = 0;
+};
+
+// Wraps ParamGeneratorInterface and provides general generator syntax
+// compatible with the STL Container concept.
+// This class implements copy initialization semantics and the contained
+// ParamGeneratorInterface instance is shared among all copies
+// of the original object. This is possible because that instance is immutable.
+template
+class ParamGenerator {
+ public:
+  typedef ParamIterator iterator;
+
+  explicit ParamGenerator(ParamGeneratorInterface* impl) : impl_(impl) {}
+  ParamGenerator(const ParamGenerator& other) : impl_(other.impl_) {}
+
+  ParamGenerator& operator=(const ParamGenerator& other) {
+    impl_ = other.impl_;
+    return *this;
+  }
+
+  iterator begin() const { return iterator(impl_->Begin()); }
+  iterator end() const { return iterator(impl_->End()); }
+
+ private:
+  linked_ptr > impl_;
+};
+
+// Generates values from a range of two comparable values. Can be used to
+// generate sequences of user-defined types that implement operator+() and
+// operator<().
+// This class is used in the Range() function.
+template 
+class RangeGenerator : public ParamGeneratorInterface {
+ public:
+  RangeGenerator(T begin, T end, IncrementT step)
+      : begin_(begin), end_(end),
+        step_(step), end_index_(CalculateEndIndex(begin, end, step)) {}
+  virtual ~RangeGenerator() {}
+
+  virtual ParamIteratorInterface* Begin() const {
+    return new Iterator(this, begin_, 0, step_);
+  }
+  virtual ParamIteratorInterface* End() const {
+    return new Iterator(this, end_, end_index_, step_);
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface {
+   public:
+    Iterator(const ParamGeneratorInterface* base, T value, int index,
+             IncrementT step)
+        : base_(base), value_(value), index_(index), step_(step) {}
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface* BaseGenerator() const {
+      return base_;
+    }
+    virtual void Advance() {
+      value_ = value_ + step_;
+      index_++;
+    }
+    virtual ParamIteratorInterface* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const T* Current() const { return &value_; }
+    virtual bool Equals(const ParamIteratorInterface& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const int other_index =
+          CheckedDowncastToActualType(&other)->index_;
+      return index_ == other_index;
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : ParamIteratorInterface(),
+          base_(other.base_), value_(other.value_), index_(other.index_),
+          step_(other.step_) {}
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface* const base_;
+    T value_;
+    int index_;
+    const IncrementT step_;
+  };  // class RangeGenerator::Iterator
+
+  static int CalculateEndIndex(const T& begin,
+                               const T& end,
+                               const IncrementT& step) {
+    int end_index = 0;
+    for (T i = begin; i < end; i = i + step)
+      end_index++;
+    return end_index;
+  }
+
+  // No implementation - assignment is unsupported.
+  void operator=(const RangeGenerator& other);
+
+  const T begin_;
+  const T end_;
+  const IncrementT step_;
+  // The index for the end() iterator. All the elements in the generated
+  // sequence are indexed (0-based) to aid iterator comparison.
+  const int end_index_;
+};  // class RangeGenerator
+
+
+// Generates values from a pair of STL-style iterators. Used in the
+// ValuesIn() function. The elements are copied from the source range
+// since the source can be located on the stack, and the generator
+// is likely to persist beyond that stack frame.
+template 
+class ValuesInIteratorRangeGenerator : public ParamGeneratorInterface {
+ public:
+  template 
+  ValuesInIteratorRangeGenerator(ForwardIterator begin, ForwardIterator end)
+      : container_(begin, end) {}
+  virtual ~ValuesInIteratorRangeGenerator() {}
+
+  virtual ParamIteratorInterface* Begin() const {
+    return new Iterator(this, container_.begin());
+  }
+  virtual ParamIteratorInterface* End() const {
+    return new Iterator(this, container_.end());
+  }
+
+ private:
+  typedef typename ::std::vector ContainerType;
+
+  class Iterator : public ParamIteratorInterface {
+   public:
+    Iterator(const ParamGeneratorInterface* base,
+             typename ContainerType::const_iterator iterator)
+        : base_(base), iterator_(iterator) {}
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface* BaseGenerator() const {
+      return base_;
+    }
+    virtual void Advance() {
+      ++iterator_;
+      value_.reset();
+    }
+    virtual ParamIteratorInterface* Clone() const {
+      return new Iterator(*this);
+    }
+    // We need to use cached value referenced by iterator_ because *iterator_
+    // can return a temporary object (and of type other then T), so just
+    // having "return &*iterator_;" doesn't work.
+    // value_ is updated here and not in Advance() because Advance()
+    // can advance iterator_ beyond the end of the range, and we cannot
+    // detect that fact. The client code, on the other hand, is
+    // responsible for not calling Current() on an out-of-range iterator.
+    virtual const T* Current() const {
+      if (value_.get() == NULL)
+        value_.reset(new T(*iterator_));
+      return value_.get();
+    }
+    virtual bool Equals(const ParamIteratorInterface& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      return iterator_ ==
+          CheckedDowncastToActualType(&other)->iterator_;
+    }
+
+   private:
+    Iterator(const Iterator& other)
+          // The explicit constructor call suppresses a false warning
+          // emitted by gcc when supplied with the -Wextra option.
+        : ParamIteratorInterface(),
+          base_(other.base_),
+          iterator_(other.iterator_) {}
+
+    const ParamGeneratorInterface* const base_;
+    typename ContainerType::const_iterator iterator_;
+    // A cached value of *iterator_. We keep it here to allow access by
+    // pointer in the wrapping iterator's operator->().
+    // value_ needs to be mutable to be accessed in Current().
+    // Use of scoped_ptr helps manage cached value's lifetime,
+    // which is bound by the lifespan of the iterator itself.
+    mutable scoped_ptr value_;
+  };  // class ValuesInIteratorRangeGenerator::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const ValuesInIteratorRangeGenerator& other);
+
+  const ContainerType container_;
+};  // class ValuesInIteratorRangeGenerator
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Stores a parameter value and later creates tests parameterized with that
+// value.
+template 
+class ParameterizedTestFactory : public TestFactoryBase {
+ public:
+  typedef typename TestClass::ParamType ParamType;
+  explicit ParameterizedTestFactory(ParamType parameter) :
+      parameter_(parameter) {}
+  virtual Test* CreateTest() {
+    TestClass::SetParam(¶meter_);
+    return new TestClass();
+  }
+
+ private:
+  const ParamType parameter_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestFactory);
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// TestMetaFactoryBase is a base class for meta-factories that create
+// test factories for passing into MakeAndRegisterTestInfo function.
+template 
+class TestMetaFactoryBase {
+ public:
+  virtual ~TestMetaFactoryBase() {}
+
+  virtual TestFactoryBase* CreateTestFactory(ParamType parameter) = 0;
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// TestMetaFactory creates test factories for passing into
+// MakeAndRegisterTestInfo function. Since MakeAndRegisterTestInfo receives
+// ownership of test factory pointer, same factory object cannot be passed
+// into that method twice. But ParameterizedTestCaseInfo is going to call
+// it for each Test/Parameter value combination. Thus it needs meta factory
+// creator class.
+template 
+class TestMetaFactory
+    : public TestMetaFactoryBase {
+ public:
+  typedef typename TestCase::ParamType ParamType;
+
+  TestMetaFactory() {}
+
+  virtual TestFactoryBase* CreateTestFactory(ParamType parameter) {
+    return new ParameterizedTestFactory(parameter);
+  }
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestMetaFactory);
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// ParameterizedTestCaseInfoBase is a generic interface
+// to ParameterizedTestCaseInfo classes. ParameterizedTestCaseInfoBase
+// accumulates test information provided by TEST_P macro invocations
+// and generators provided by INSTANTIATE_TEST_CASE_P macro invocations
+// and uses that information to register all resulting test instances
+// in RegisterTests method. The ParameterizeTestCaseRegistry class holds
+// a collection of pointers to the ParameterizedTestCaseInfo objects
+// and calls RegisterTests() on each of them when asked.
+class ParameterizedTestCaseInfoBase {
+ public:
+  virtual ~ParameterizedTestCaseInfoBase() {}
+
+  // Base part of test case name for display purposes.
+  virtual const string& GetTestCaseName() const = 0;
+  // Test case id to verify identity.
+  virtual TypeId GetTestCaseTypeId() const = 0;
+  // UnitTest class invokes this method to register tests in this
+  // test case right before running them in RUN_ALL_TESTS macro.
+  // This method should not be called more then once on any single
+  // instance of a ParameterizedTestCaseInfoBase derived class.
+  virtual void RegisterTests() = 0;
+
+ protected:
+  ParameterizedTestCaseInfoBase() {}
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestCaseInfoBase);
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// ParameterizedTestCaseInfo accumulates tests obtained from TEST_P
+// macro invocations for a particular test case and generators
+// obtained from INSTANTIATE_TEST_CASE_P macro invocations for that
+// test case. It registers tests with all values generated by all
+// generators when asked.
+template 
+class ParameterizedTestCaseInfo : public ParameterizedTestCaseInfoBase {
+ public:
+  // ParamType and GeneratorCreationFunc are private types but are required
+  // for declarations of public methods AddTestPattern() and
+  // AddTestCaseInstantiation().
+  typedef typename TestCase::ParamType ParamType;
+  // A function that returns an instance of appropriate generator type.
+  typedef ParamGenerator(GeneratorCreationFunc)();
+
+  explicit ParameterizedTestCaseInfo(const char* name)
+      : test_case_name_(name) {}
+
+  // Test case base name for display purposes.
+  virtual const string& GetTestCaseName() const { return test_case_name_; }
+  // Test case id to verify identity.
+  virtual TypeId GetTestCaseTypeId() const { return GetTypeId(); }
+  // TEST_P macro uses AddTestPattern() to record information
+  // about a single test in a LocalTestInfo structure.
+  // test_case_name is the base name of the test case (without invocation
+  // prefix). test_base_name is the name of an individual test without
+  // parameter index. For the test SequenceA/FooTest.DoBar/1 FooTest is
+  // test case base name and DoBar is test base name.
+  void AddTestPattern(const char* test_case_name,
+                      const char* test_base_name,
+                      TestMetaFactoryBase* meta_factory) {
+    tests_.push_back(linked_ptr(new TestInfo(test_case_name,
+                                                       test_base_name,
+                                                       meta_factory)));
+  }
+  // INSTANTIATE_TEST_CASE_P macro uses AddGenerator() to record information
+  // about a generator.
+  int AddTestCaseInstantiation(const string& instantiation_name,
+                               GeneratorCreationFunc* func,
+                               const char* /* file */,
+                               int /* line */) {
+    instantiations_.push_back(::std::make_pair(instantiation_name, func));
+    return 0;  // Return value used only to run this method in namespace scope.
+  }
+  // UnitTest class invokes this method to register tests in this test case
+  // test cases right before running tests in RUN_ALL_TESTS macro.
+  // This method should not be called more then once on any single
+  // instance of a ParameterizedTestCaseInfoBase derived class.
+  // UnitTest has a guard to prevent from calling this method more then once.
+  virtual void RegisterTests() {
+    for (typename TestInfoContainer::iterator test_it = tests_.begin();
+         test_it != tests_.end(); ++test_it) {
+      linked_ptr test_info = *test_it;
+      for (typename InstantiationContainer::iterator gen_it =
+               instantiations_.begin(); gen_it != instantiations_.end();
+               ++gen_it) {
+        const string& instantiation_name = gen_it->first;
+        ParamGenerator generator((*gen_it->second)());
+
+        string test_case_name;
+        if ( !instantiation_name.empty() )
+          test_case_name = instantiation_name + "/";
+        test_case_name += test_info->test_case_base_name;
+
+        int i = 0;
+        for (typename ParamGenerator::iterator param_it =
+                 generator.begin();
+             param_it != generator.end(); ++param_it, ++i) {
+          Message test_name_stream;
+          test_name_stream << test_info->test_base_name << "/" << i;
+          MakeAndRegisterTestInfo(
+              test_case_name.c_str(),
+              test_name_stream.GetString().c_str(),
+              NULL,  // No type parameter.
+              PrintToString(*param_it).c_str(),
+              GetTestCaseTypeId(),
+              TestCase::SetUpTestCase,
+              TestCase::TearDownTestCase,
+              test_info->test_meta_factory->CreateTestFactory(*param_it));
+        }  // for param_it
+      }  // for gen_it
+    }  // for test_it
+  }  // RegisterTests
+
+ private:
+  // LocalTestInfo structure keeps information about a single test registered
+  // with TEST_P macro.
+  struct TestInfo {
+    TestInfo(const char* a_test_case_base_name,
+             const char* a_test_base_name,
+             TestMetaFactoryBase* a_test_meta_factory) :
+        test_case_base_name(a_test_case_base_name),
+        test_base_name(a_test_base_name),
+        test_meta_factory(a_test_meta_factory) {}
+
+    const string test_case_base_name;
+    const string test_base_name;
+    const scoped_ptr > test_meta_factory;
+  };
+  typedef ::std::vector > TestInfoContainer;
+  // Keeps pairs of 
+  // received from INSTANTIATE_TEST_CASE_P macros.
+  typedef ::std::vector >
+      InstantiationContainer;
+
+  const string test_case_name_;
+  TestInfoContainer tests_;
+  InstantiationContainer instantiations_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestCaseInfo);
+};  // class ParameterizedTestCaseInfo
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// ParameterizedTestCaseRegistry contains a map of ParameterizedTestCaseInfoBase
+// classes accessed by test case names. TEST_P and INSTANTIATE_TEST_CASE_P
+// macros use it to locate their corresponding ParameterizedTestCaseInfo
+// descriptors.
+class ParameterizedTestCaseRegistry {
+ public:
+  ParameterizedTestCaseRegistry() {}
+  ~ParameterizedTestCaseRegistry() {
+    for (TestCaseInfoContainer::iterator it = test_case_infos_.begin();
+         it != test_case_infos_.end(); ++it) {
+      delete *it;
+    }
+  }
+
+  // Looks up or creates and returns a structure containing information about
+  // tests and instantiations of a particular test case.
+  template 
+  ParameterizedTestCaseInfo* GetTestCasePatternHolder(
+      const char* test_case_name,
+      const char* file,
+      int line) {
+    ParameterizedTestCaseInfo* typed_test_info = NULL;
+    for (TestCaseInfoContainer::iterator it = test_case_infos_.begin();
+         it != test_case_infos_.end(); ++it) {
+      if ((*it)->GetTestCaseName() == test_case_name) {
+        if ((*it)->GetTestCaseTypeId() != GetTypeId()) {
+          // Complain about incorrect usage of Google Test facilities
+          // and terminate the program since we cannot guaranty correct
+          // test case setup and tear-down in this case.
+          ReportInvalidTestCaseType(test_case_name,  file, line);
+          posix::Abort();
+        } else {
+          // At this point we are sure that the object we found is of the same
+          // type we are looking for, so we downcast it to that type
+          // without further checks.
+          typed_test_info = CheckedDowncastToActualType<
+              ParameterizedTestCaseInfo >(*it);
+        }
+        break;
+      }
+    }
+    if (typed_test_info == NULL) {
+      typed_test_info = new ParameterizedTestCaseInfo(test_case_name);
+      test_case_infos_.push_back(typed_test_info);
+    }
+    return typed_test_info;
+  }
+  void RegisterTests() {
+    for (TestCaseInfoContainer::iterator it = test_case_infos_.begin();
+         it != test_case_infos_.end(); ++it) {
+      (*it)->RegisterTests();
+    }
+  }
+
+ private:
+  typedef ::std::vector TestCaseInfoContainer;
+
+  TestCaseInfoContainer test_case_infos_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestCaseRegistry);
+};
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  //  GTEST_HAS_PARAM_TEST
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_H_
+// This file was GENERATED by command:
+//     pump.py gtest-param-util-generated.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: vladl@google.com (Vlad Losev)
+
+// Type and function utilities for implementing parameterized tests.
+// This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+// Currently Google Test supports at most 50 arguments in Values,
+// and at most 10 arguments in Combine. Please contact
+// googletestframework@googlegroups.com if you need more.
+// Please note that the number of arguments to Combine is limited
+// by the maximum arity of the implementation of tr1::tuple which is
+// currently set at 10.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
+
+// scripts/fuse_gtest.py depends on gtest's own header being #included
+// *unconditionally*.  Therefore these #includes cannot be moved
+// inside #if GTEST_HAS_PARAM_TEST.
+
+#if GTEST_HAS_PARAM_TEST
+
+namespace testing {
+
+// Forward declarations of ValuesIn(), which is implemented in
+// include/gtest/gtest-param-test.h.
+template 
+internal::ParamGenerator<
+  typename ::testing::internal::IteratorTraits::value_type>
+ValuesIn(ForwardIterator begin, ForwardIterator end);
+
+template 
+internal::ParamGenerator ValuesIn(const T (&array)[N]);
+
+template 
+internal::ParamGenerator ValuesIn(
+    const Container& container);
+
+namespace internal {
+
+// Used in the Values() function to provide polymorphic capabilities.
+template 
+class ValueArray1 {
+ public:
+  explicit ValueArray1(T1 v1) : v1_(v1) {}
+
+  template 
+  operator ParamGenerator() const { return ValuesIn(&v1_, &v1_ + 1); }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray1& other);
+
+  const T1 v1_;
+};
+
+template 
+class ValueArray2 {
+ public:
+  ValueArray2(T1 v1, T2 v2) : v1_(v1), v2_(v2) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray2& other);
+
+  const T1 v1_;
+  const T2 v2_;
+};
+
+template 
+class ValueArray3 {
+ public:
+  ValueArray3(T1 v1, T2 v2, T3 v3) : v1_(v1), v2_(v2), v3_(v3) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray3& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+};
+
+template 
+class ValueArray4 {
+ public:
+  ValueArray4(T1 v1, T2 v2, T3 v3, T4 v4) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray4& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+};
+
+template 
+class ValueArray5 {
+ public:
+  ValueArray5(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray5& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+};
+
+template 
+class ValueArray6 {
+ public:
+  ValueArray6(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray6& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+};
+
+template 
+class ValueArray7 {
+ public:
+  ValueArray7(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray7& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+};
+
+template 
+class ValueArray8 {
+ public:
+  ValueArray8(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+      T8 v8) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray8& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+};
+
+template 
+class ValueArray9 {
+ public:
+  ValueArray9(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+      T9 v9) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray9& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+};
+
+template 
+class ValueArray10 {
+ public:
+  ValueArray10(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray10& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+};
+
+template 
+class ValueArray11 {
+ public:
+  ValueArray11(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray11& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+};
+
+template 
+class ValueArray12 {
+ public:
+  ValueArray12(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray12& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+};
+
+template 
+class ValueArray13 {
+ public:
+  ValueArray13(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray13& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+};
+
+template 
+class ValueArray14 {
+ public:
+  ValueArray14(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray14& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+};
+
+template 
+class ValueArray15 {
+ public:
+  ValueArray15(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray15& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+};
+
+template 
+class ValueArray16 {
+ public:
+  ValueArray16(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray16& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+};
+
+template 
+class ValueArray17 {
+ public:
+  ValueArray17(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16,
+      T17 v17) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray17& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+};
+
+template 
+class ValueArray18 {
+ public:
+  ValueArray18(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray18& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+};
+
+template 
+class ValueArray19 {
+ public:
+  ValueArray19(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray19& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+};
+
+template 
+class ValueArray20 {
+ public:
+  ValueArray20(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray20& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+};
+
+template 
+class ValueArray21 {
+ public:
+  ValueArray21(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray21& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+};
+
+template 
+class ValueArray22 {
+ public:
+  ValueArray22(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray22& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+};
+
+template 
+class ValueArray23 {
+ public:
+  ValueArray23(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray23& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+};
+
+template 
+class ValueArray24 {
+ public:
+  ValueArray24(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray24& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+};
+
+template 
+class ValueArray25 {
+ public:
+  ValueArray25(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24,
+      T25 v25) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray25& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+};
+
+template 
+class ValueArray26 {
+ public:
+  ValueArray26(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray26& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+};
+
+template 
+class ValueArray27 {
+ public:
+  ValueArray27(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19),
+      v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25),
+      v26_(v26), v27_(v27) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray27& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+};
+
+template 
+class ValueArray28 {
+ public:
+  ValueArray28(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24),
+      v25_(v25), v26_(v26), v27_(v27), v28_(v28) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray28& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+};
+
+template 
+class ValueArray29 {
+ public:
+  ValueArray29(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23),
+      v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray29& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+};
+
+template 
+class ValueArray30 {
+ public:
+  ValueArray30(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray30& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+};
+
+template 
+class ValueArray31 {
+ public:
+  ValueArray31(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray31& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+};
+
+template 
+class ValueArray32 {
+ public:
+  ValueArray32(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27),
+      v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray32& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+};
+
+template 
+class ValueArray33 {
+ public:
+  ValueArray33(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32,
+      T33 v33) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray33& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+};
+
+template 
+class ValueArray34 {
+ public:
+  ValueArray34(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray34& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+};
+
+template 
+class ValueArray35 {
+ public:
+  ValueArray35(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19),
+      v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25),
+      v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31),
+      v32_(v32), v33_(v33), v34_(v34), v35_(v35) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray35& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+};
+
+template 
+class ValueArray36 {
+ public:
+  ValueArray36(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24),
+      v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30),
+      v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35), v36_(v36) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray36& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+};
+
+template 
+class ValueArray37 {
+ public:
+  ValueArray37(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23),
+      v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29),
+      v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35),
+      v36_(v36), v37_(v37) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray37& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+};
+
+template 
+class ValueArray38 {
+ public:
+  ValueArray38(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray38& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+};
+
+template 
+class ValueArray39 {
+ public:
+  ValueArray39(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray39& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+};
+
+template 
+class ValueArray40 {
+ public:
+  ValueArray40(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27),
+      v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33),
+      v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39),
+      v40_(v40) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_), static_cast(v40_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray40& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+};
+
+template 
+class ValueArray41 {
+ public:
+  ValueArray41(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40,
+      T41 v41) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_), static_cast(v40_), static_cast(v41_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray41& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+};
+
+template 
+class ValueArray42 {
+ public:
+  ValueArray42(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41), v42_(v42) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_), static_cast(v40_), static_cast(v41_),
+        static_cast(v42_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray42& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+};
+
+template 
+class ValueArray43 {
+ public:
+  ValueArray43(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19),
+      v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25),
+      v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31),
+      v32_(v32), v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37),
+      v38_(v38), v39_(v39), v40_(v40), v41_(v41), v42_(v42), v43_(v43) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_), static_cast(v40_), static_cast(v41_),
+        static_cast(v42_), static_cast(v43_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray43& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+};
+
+template 
+class ValueArray44 {
+ public:
+  ValueArray44(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24),
+      v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30),
+      v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35), v36_(v36),
+      v37_(v37), v38_(v38), v39_(v39), v40_(v40), v41_(v41), v42_(v42),
+      v43_(v43), v44_(v44) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_), static_cast(v40_), static_cast(v41_),
+        static_cast(v42_), static_cast(v43_), static_cast(v44_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray44& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+};
+
+template 
+class ValueArray45 {
+ public:
+  ValueArray45(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23),
+      v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29),
+      v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35),
+      v36_(v36), v37_(v37), v38_(v38), v39_(v39), v40_(v40), v41_(v41),
+      v42_(v42), v43_(v43), v44_(v44), v45_(v45) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_), static_cast(v40_), static_cast(v41_),
+        static_cast(v42_), static_cast(v43_), static_cast(v44_),
+        static_cast(v45_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray45& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+};
+
+template 
+class ValueArray46 {
+ public:
+  ValueArray46(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39), v40_(v40),
+      v41_(v41), v42_(v42), v43_(v43), v44_(v44), v45_(v45), v46_(v46) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_), static_cast(v40_), static_cast(v41_),
+        static_cast(v42_), static_cast(v43_), static_cast(v44_),
+        static_cast(v45_), static_cast(v46_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray46& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+};
+
+template 
+class ValueArray47 {
+ public:
+  ValueArray47(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39), v40_(v40),
+      v41_(v41), v42_(v42), v43_(v43), v44_(v44), v45_(v45), v46_(v46),
+      v47_(v47) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_), static_cast(v40_), static_cast(v41_),
+        static_cast(v42_), static_cast(v43_), static_cast(v44_),
+        static_cast(v45_), static_cast(v46_), static_cast(v47_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray47& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+};
+
+template 
+class ValueArray48 {
+ public:
+  ValueArray48(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47, T48 v48) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27),
+      v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33),
+      v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39),
+      v40_(v40), v41_(v41), v42_(v42), v43_(v43), v44_(v44), v45_(v45),
+      v46_(v46), v47_(v47), v48_(v48) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_), static_cast(v40_), static_cast(v41_),
+        static_cast(v42_), static_cast(v43_), static_cast(v44_),
+        static_cast(v45_), static_cast(v46_), static_cast(v47_),
+        static_cast(v48_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray48& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+  const T48 v48_;
+};
+
+template 
+class ValueArray49 {
+ public:
+  ValueArray49(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47, T48 v48,
+      T49 v49) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41), v42_(v42), v43_(v43), v44_(v44),
+      v45_(v45), v46_(v46), v47_(v47), v48_(v48), v49_(v49) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_), static_cast(v40_), static_cast(v41_),
+        static_cast(v42_), static_cast(v43_), static_cast(v44_),
+        static_cast(v45_), static_cast(v46_), static_cast(v47_),
+        static_cast(v48_), static_cast(v49_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray49& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+  const T48 v48_;
+  const T49 v49_;
+};
+
+template 
+class ValueArray50 {
+ public:
+  ValueArray50(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47, T48 v48, T49 v49,
+      T50 v50) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41), v42_(v42), v43_(v43), v44_(v44),
+      v45_(v45), v46_(v46), v47_(v47), v48_(v48), v49_(v49), v50_(v50) {}
+
+  template 
+  operator ParamGenerator() const {
+    const T array[] = {static_cast(v1_), static_cast(v2_),
+        static_cast(v3_), static_cast(v4_), static_cast(v5_),
+        static_cast(v6_), static_cast(v7_), static_cast(v8_),
+        static_cast(v9_), static_cast(v10_), static_cast(v11_),
+        static_cast(v12_), static_cast(v13_), static_cast(v14_),
+        static_cast(v15_), static_cast(v16_), static_cast(v17_),
+        static_cast(v18_), static_cast(v19_), static_cast(v20_),
+        static_cast(v21_), static_cast(v22_), static_cast(v23_),
+        static_cast(v24_), static_cast(v25_), static_cast(v26_),
+        static_cast(v27_), static_cast(v28_), static_cast(v29_),
+        static_cast(v30_), static_cast(v31_), static_cast(v32_),
+        static_cast(v33_), static_cast(v34_), static_cast(v35_),
+        static_cast(v36_), static_cast(v37_), static_cast(v38_),
+        static_cast(v39_), static_cast(v40_), static_cast(v41_),
+        static_cast(v42_), static_cast(v43_), static_cast(v44_),
+        static_cast(v45_), static_cast(v46_), static_cast(v47_),
+        static_cast(v48_), static_cast(v49_), static_cast(v50_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray50& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+  const T48 v48_;
+  const T49 v49_;
+  const T50 v50_;
+};
+
+# if GTEST_HAS_COMBINE
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Generates values from the Cartesian product of values produced
+// by the argument generators.
+//
+template 
+class CartesianProductGenerator2
+    : public ParamGeneratorInterface< ::std::tr1::tuple > {
+ public:
+  typedef ::std::tr1::tuple ParamType;
+
+  CartesianProductGenerator2(const ParamGenerator& g1,
+      const ParamGenerator& g2)
+      : g1_(g1), g2_(g2) {}
+  virtual ~CartesianProductGenerator2() {}
+
+  virtual ParamIteratorInterface* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin());
+  }
+  virtual ParamIteratorInterface* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface {
+   public:
+    Iterator(const ParamGeneratorInterface* base,
+      const ParamGenerator& g1,
+      const typename ParamGenerator::iterator& current1,
+      const ParamGenerator& g2,
+      const typename ParamGenerator::iterator& current2)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current2_;
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return ¤t_value_; }
+    virtual bool Equals(const ParamIteratorInterface& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator::iterator begin1_;
+    const typename ParamGenerator::iterator end1_;
+    typename ParamGenerator::iterator current1_;
+    const typename ParamGenerator::iterator begin2_;
+    const typename ParamGenerator::iterator end2_;
+    typename ParamGenerator::iterator current2_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator2::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator2& other);
+
+  const ParamGenerator g1_;
+  const ParamGenerator g2_;
+};  // class CartesianProductGenerator2
+
+
+template 
+class CartesianProductGenerator3
+    : public ParamGeneratorInterface< ::std::tr1::tuple > {
+ public:
+  typedef ::std::tr1::tuple ParamType;
+
+  CartesianProductGenerator3(const ParamGenerator& g1,
+      const ParamGenerator& g2, const ParamGenerator& g3)
+      : g1_(g1), g2_(g2), g3_(g3) {}
+  virtual ~CartesianProductGenerator3() {}
+
+  virtual ParamIteratorInterface* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin());
+  }
+  virtual ParamIteratorInterface* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface {
+   public:
+    Iterator(const ParamGeneratorInterface* base,
+      const ParamGenerator& g1,
+      const typename ParamGenerator::iterator& current1,
+      const ParamGenerator& g2,
+      const typename ParamGenerator::iterator& current2,
+      const ParamGenerator& g3,
+      const typename ParamGenerator::iterator& current3)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current3_;
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return ¤t_value_; }
+    virtual bool Equals(const ParamIteratorInterface& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator::iterator begin1_;
+    const typename ParamGenerator::iterator end1_;
+    typename ParamGenerator::iterator current1_;
+    const typename ParamGenerator::iterator begin2_;
+    const typename ParamGenerator::iterator end2_;
+    typename ParamGenerator::iterator current2_;
+    const typename ParamGenerator::iterator begin3_;
+    const typename ParamGenerator::iterator end3_;
+    typename ParamGenerator::iterator current3_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator3::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator3& other);
+
+  const ParamGenerator g1_;
+  const ParamGenerator g2_;
+  const ParamGenerator g3_;
+};  // class CartesianProductGenerator3
+
+
+template 
+class CartesianProductGenerator4
+    : public ParamGeneratorInterface< ::std::tr1::tuple > {
+ public:
+  typedef ::std::tr1::tuple ParamType;
+
+  CartesianProductGenerator4(const ParamGenerator& g1,
+      const ParamGenerator& g2, const ParamGenerator& g3,
+      const ParamGenerator& g4)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4) {}
+  virtual ~CartesianProductGenerator4() {}
+
+  virtual ParamIteratorInterface* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin());
+  }
+  virtual ParamIteratorInterface* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface {
+   public:
+    Iterator(const ParamGeneratorInterface* base,
+      const ParamGenerator& g1,
+      const typename ParamGenerator::iterator& current1,
+      const ParamGenerator& g2,
+      const typename ParamGenerator::iterator& current2,
+      const ParamGenerator& g3,
+      const typename ParamGenerator::iterator& current3,
+      const ParamGenerator& g4,
+      const typename ParamGenerator::iterator& current4)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current4_;
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return ¤t_value_; }
+    virtual bool Equals(const ParamIteratorInterface& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator::iterator begin1_;
+    const typename ParamGenerator::iterator end1_;
+    typename ParamGenerator::iterator current1_;
+    const typename ParamGenerator::iterator begin2_;
+    const typename ParamGenerator::iterator end2_;
+    typename ParamGenerator::iterator current2_;
+    const typename ParamGenerator::iterator begin3_;
+    const typename ParamGenerator::iterator end3_;
+    typename ParamGenerator::iterator current3_;
+    const typename ParamGenerator::iterator begin4_;
+    const typename ParamGenerator::iterator end4_;
+    typename ParamGenerator::iterator current4_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator4::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator4& other);
+
+  const ParamGenerator g1_;
+  const ParamGenerator g2_;
+  const ParamGenerator g3_;
+  const ParamGenerator g4_;
+};  // class CartesianProductGenerator4
+
+
+template 
+class CartesianProductGenerator5
+    : public ParamGeneratorInterface< ::std::tr1::tuple > {
+ public:
+  typedef ::std::tr1::tuple ParamType;
+
+  CartesianProductGenerator5(const ParamGenerator& g1,
+      const ParamGenerator& g2, const ParamGenerator& g3,
+      const ParamGenerator& g4, const ParamGenerator& g5)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5) {}
+  virtual ~CartesianProductGenerator5() {}
+
+  virtual ParamIteratorInterface* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin());
+  }
+  virtual ParamIteratorInterface* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface {
+   public:
+    Iterator(const ParamGeneratorInterface* base,
+      const ParamGenerator& g1,
+      const typename ParamGenerator::iterator& current1,
+      const ParamGenerator& g2,
+      const typename ParamGenerator::iterator& current2,
+      const ParamGenerator& g3,
+      const typename ParamGenerator::iterator& current3,
+      const ParamGenerator& g4,
+      const typename ParamGenerator::iterator& current4,
+      const ParamGenerator& g5,
+      const typename ParamGenerator::iterator& current5)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current5_;
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return ¤t_value_; }
+    virtual bool Equals(const ParamIteratorInterface& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator::iterator begin1_;
+    const typename ParamGenerator::iterator end1_;
+    typename ParamGenerator::iterator current1_;
+    const typename ParamGenerator::iterator begin2_;
+    const typename ParamGenerator::iterator end2_;
+    typename ParamGenerator::iterator current2_;
+    const typename ParamGenerator::iterator begin3_;
+    const typename ParamGenerator::iterator end3_;
+    typename ParamGenerator::iterator current3_;
+    const typename ParamGenerator::iterator begin4_;
+    const typename ParamGenerator::iterator end4_;
+    typename ParamGenerator::iterator current4_;
+    const typename ParamGenerator::iterator begin5_;
+    const typename ParamGenerator::iterator end5_;
+    typename ParamGenerator::iterator current5_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator5::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator5& other);
+
+  const ParamGenerator g1_;
+  const ParamGenerator g2_;
+  const ParamGenerator g3_;
+  const ParamGenerator g4_;
+  const ParamGenerator g5_;
+};  // class CartesianProductGenerator5
+
+
+template 
+class CartesianProductGenerator6
+    : public ParamGeneratorInterface< ::std::tr1::tuple > {
+ public:
+  typedef ::std::tr1::tuple ParamType;
+
+  CartesianProductGenerator6(const ParamGenerator& g1,
+      const ParamGenerator& g2, const ParamGenerator& g3,
+      const ParamGenerator& g4, const ParamGenerator& g5,
+      const ParamGenerator& g6)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6) {}
+  virtual ~CartesianProductGenerator6() {}
+
+  virtual ParamIteratorInterface* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin());
+  }
+  virtual ParamIteratorInterface* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface {
+   public:
+    Iterator(const ParamGeneratorInterface* base,
+      const ParamGenerator& g1,
+      const typename ParamGenerator::iterator& current1,
+      const ParamGenerator& g2,
+      const typename ParamGenerator::iterator& current2,
+      const ParamGenerator& g3,
+      const typename ParamGenerator::iterator& current3,
+      const ParamGenerator& g4,
+      const typename ParamGenerator::iterator& current4,
+      const ParamGenerator& g5,
+      const typename ParamGenerator::iterator& current5,
+      const ParamGenerator& g6,
+      const typename ParamGenerator::iterator& current6)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current6_;
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return ¤t_value_; }
+    virtual bool Equals(const ParamIteratorInterface& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator::iterator begin1_;
+    const typename ParamGenerator::iterator end1_;
+    typename ParamGenerator::iterator current1_;
+    const typename ParamGenerator::iterator begin2_;
+    const typename ParamGenerator::iterator end2_;
+    typename ParamGenerator::iterator current2_;
+    const typename ParamGenerator::iterator begin3_;
+    const typename ParamGenerator::iterator end3_;
+    typename ParamGenerator::iterator current3_;
+    const typename ParamGenerator::iterator begin4_;
+    const typename ParamGenerator::iterator end4_;
+    typename ParamGenerator::iterator current4_;
+    const typename ParamGenerator::iterator begin5_;
+    const typename ParamGenerator::iterator end5_;
+    typename ParamGenerator::iterator current5_;
+    const typename ParamGenerator::iterator begin6_;
+    const typename ParamGenerator::iterator end6_;
+    typename ParamGenerator::iterator current6_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator6::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator6& other);
+
+  const ParamGenerator g1_;
+  const ParamGenerator g2_;
+  const ParamGenerator g3_;
+  const ParamGenerator g4_;
+  const ParamGenerator g5_;
+  const ParamGenerator g6_;
+};  // class CartesianProductGenerator6
+
+
+template 
+class CartesianProductGenerator7
+    : public ParamGeneratorInterface< ::std::tr1::tuple > {
+ public:
+  typedef ::std::tr1::tuple ParamType;
+
+  CartesianProductGenerator7(const ParamGenerator& g1,
+      const ParamGenerator& g2, const ParamGenerator& g3,
+      const ParamGenerator& g4, const ParamGenerator& g5,
+      const ParamGenerator& g6, const ParamGenerator& g7)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7) {}
+  virtual ~CartesianProductGenerator7() {}
+
+  virtual ParamIteratorInterface* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin());
+  }
+  virtual ParamIteratorInterface* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface {
+   public:
+    Iterator(const ParamGeneratorInterface* base,
+      const ParamGenerator& g1,
+      const typename ParamGenerator::iterator& current1,
+      const ParamGenerator& g2,
+      const typename ParamGenerator::iterator& current2,
+      const ParamGenerator& g3,
+      const typename ParamGenerator::iterator& current3,
+      const ParamGenerator& g4,
+      const typename ParamGenerator::iterator& current4,
+      const ParamGenerator& g5,
+      const typename ParamGenerator::iterator& current5,
+      const ParamGenerator& g6,
+      const typename ParamGenerator::iterator& current6,
+      const ParamGenerator& g7,
+      const typename ParamGenerator::iterator& current7)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current7_;
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return ¤t_value_; }
+    virtual bool Equals(const ParamIteratorInterface& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator::iterator begin1_;
+    const typename ParamGenerator::iterator end1_;
+    typename ParamGenerator::iterator current1_;
+    const typename ParamGenerator::iterator begin2_;
+    const typename ParamGenerator::iterator end2_;
+    typename ParamGenerator::iterator current2_;
+    const typename ParamGenerator::iterator begin3_;
+    const typename ParamGenerator::iterator end3_;
+    typename ParamGenerator::iterator current3_;
+    const typename ParamGenerator::iterator begin4_;
+    const typename ParamGenerator::iterator end4_;
+    typename ParamGenerator::iterator current4_;
+    const typename ParamGenerator::iterator begin5_;
+    const typename ParamGenerator::iterator end5_;
+    typename ParamGenerator::iterator current5_;
+    const typename ParamGenerator::iterator begin6_;
+    const typename ParamGenerator::iterator end6_;
+    typename ParamGenerator::iterator current6_;
+    const typename ParamGenerator::iterator begin7_;
+    const typename ParamGenerator::iterator end7_;
+    typename ParamGenerator::iterator current7_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator7::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator7& other);
+
+  const ParamGenerator g1_;
+  const ParamGenerator g2_;
+  const ParamGenerator g3_;
+  const ParamGenerator g4_;
+  const ParamGenerator g5_;
+  const ParamGenerator g6_;
+  const ParamGenerator g7_;
+};  // class CartesianProductGenerator7
+
+
+template 
+class CartesianProductGenerator8
+    : public ParamGeneratorInterface< ::std::tr1::tuple > {
+ public:
+  typedef ::std::tr1::tuple ParamType;
+
+  CartesianProductGenerator8(const ParamGenerator& g1,
+      const ParamGenerator& g2, const ParamGenerator& g3,
+      const ParamGenerator& g4, const ParamGenerator& g5,
+      const ParamGenerator& g6, const ParamGenerator& g7,
+      const ParamGenerator& g8)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7),
+          g8_(g8) {}
+  virtual ~CartesianProductGenerator8() {}
+
+  virtual ParamIteratorInterface* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin(), g8_, g8_.begin());
+  }
+  virtual ParamIteratorInterface* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end(), g8_,
+        g8_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface {
+   public:
+    Iterator(const ParamGeneratorInterface* base,
+      const ParamGenerator& g1,
+      const typename ParamGenerator::iterator& current1,
+      const ParamGenerator& g2,
+      const typename ParamGenerator::iterator& current2,
+      const ParamGenerator& g3,
+      const typename ParamGenerator::iterator& current3,
+      const ParamGenerator& g4,
+      const typename ParamGenerator::iterator& current4,
+      const ParamGenerator& g5,
+      const typename ParamGenerator::iterator& current5,
+      const ParamGenerator& g6,
+      const typename ParamGenerator::iterator& current6,
+      const ParamGenerator& g7,
+      const typename ParamGenerator::iterator& current7,
+      const ParamGenerator& g8,
+      const typename ParamGenerator::iterator& current8)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7),
+          begin8_(g8.begin()), end8_(g8.end()), current8_(current8)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current8_;
+      if (current8_ == end8_) {
+        current8_ = begin8_;
+        ++current7_;
+      }
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return ¤t_value_; }
+    virtual bool Equals(const ParamIteratorInterface& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_ &&
+          current8_ == typed_other->current8_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_),
+        begin8_(other.begin8_),
+        end8_(other.end8_),
+        current8_(other.current8_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_, *current8_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_ ||
+          current8_ == end8_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator::iterator begin1_;
+    const typename ParamGenerator::iterator end1_;
+    typename ParamGenerator::iterator current1_;
+    const typename ParamGenerator::iterator begin2_;
+    const typename ParamGenerator::iterator end2_;
+    typename ParamGenerator::iterator current2_;
+    const typename ParamGenerator::iterator begin3_;
+    const typename ParamGenerator::iterator end3_;
+    typename ParamGenerator::iterator current3_;
+    const typename ParamGenerator::iterator begin4_;
+    const typename ParamGenerator::iterator end4_;
+    typename ParamGenerator::iterator current4_;
+    const typename ParamGenerator::iterator begin5_;
+    const typename ParamGenerator::iterator end5_;
+    typename ParamGenerator::iterator current5_;
+    const typename ParamGenerator::iterator begin6_;
+    const typename ParamGenerator::iterator end6_;
+    typename ParamGenerator::iterator current6_;
+    const typename ParamGenerator::iterator begin7_;
+    const typename ParamGenerator::iterator end7_;
+    typename ParamGenerator::iterator current7_;
+    const typename ParamGenerator::iterator begin8_;
+    const typename ParamGenerator::iterator end8_;
+    typename ParamGenerator::iterator current8_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator8::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator8& other);
+
+  const ParamGenerator g1_;
+  const ParamGenerator g2_;
+  const ParamGenerator g3_;
+  const ParamGenerator g4_;
+  const ParamGenerator g5_;
+  const ParamGenerator g6_;
+  const ParamGenerator g7_;
+  const ParamGenerator g8_;
+};  // class CartesianProductGenerator8
+
+
+template 
+class CartesianProductGenerator9
+    : public ParamGeneratorInterface< ::std::tr1::tuple > {
+ public:
+  typedef ::std::tr1::tuple ParamType;
+
+  CartesianProductGenerator9(const ParamGenerator& g1,
+      const ParamGenerator& g2, const ParamGenerator& g3,
+      const ParamGenerator& g4, const ParamGenerator& g5,
+      const ParamGenerator& g6, const ParamGenerator& g7,
+      const ParamGenerator& g8, const ParamGenerator& g9)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9) {}
+  virtual ~CartesianProductGenerator9() {}
+
+  virtual ParamIteratorInterface* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin(), g8_, g8_.begin(), g9_, g9_.begin());
+  }
+  virtual ParamIteratorInterface* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end(), g8_,
+        g8_.end(), g9_, g9_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface {
+   public:
+    Iterator(const ParamGeneratorInterface* base,
+      const ParamGenerator& g1,
+      const typename ParamGenerator::iterator& current1,
+      const ParamGenerator& g2,
+      const typename ParamGenerator::iterator& current2,
+      const ParamGenerator& g3,
+      const typename ParamGenerator::iterator& current3,
+      const ParamGenerator& g4,
+      const typename ParamGenerator::iterator& current4,
+      const ParamGenerator& g5,
+      const typename ParamGenerator::iterator& current5,
+      const ParamGenerator& g6,
+      const typename ParamGenerator::iterator& current6,
+      const ParamGenerator& g7,
+      const typename ParamGenerator::iterator& current7,
+      const ParamGenerator& g8,
+      const typename ParamGenerator::iterator& current8,
+      const ParamGenerator& g9,
+      const typename ParamGenerator::iterator& current9)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7),
+          begin8_(g8.begin()), end8_(g8.end()), current8_(current8),
+          begin9_(g9.begin()), end9_(g9.end()), current9_(current9)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current9_;
+      if (current9_ == end9_) {
+        current9_ = begin9_;
+        ++current8_;
+      }
+      if (current8_ == end8_) {
+        current8_ = begin8_;
+        ++current7_;
+      }
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return ¤t_value_; }
+    virtual bool Equals(const ParamIteratorInterface& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_ &&
+          current8_ == typed_other->current8_ &&
+          current9_ == typed_other->current9_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_),
+        begin8_(other.begin8_),
+        end8_(other.end8_),
+        current8_(other.current8_),
+        begin9_(other.begin9_),
+        end9_(other.end9_),
+        current9_(other.current9_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_, *current8_,
+            *current9_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_ ||
+          current8_ == end8_ ||
+          current9_ == end9_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator::iterator begin1_;
+    const typename ParamGenerator::iterator end1_;
+    typename ParamGenerator::iterator current1_;
+    const typename ParamGenerator::iterator begin2_;
+    const typename ParamGenerator::iterator end2_;
+    typename ParamGenerator::iterator current2_;
+    const typename ParamGenerator::iterator begin3_;
+    const typename ParamGenerator::iterator end3_;
+    typename ParamGenerator::iterator current3_;
+    const typename ParamGenerator::iterator begin4_;
+    const typename ParamGenerator::iterator end4_;
+    typename ParamGenerator::iterator current4_;
+    const typename ParamGenerator::iterator begin5_;
+    const typename ParamGenerator::iterator end5_;
+    typename ParamGenerator::iterator current5_;
+    const typename ParamGenerator::iterator begin6_;
+    const typename ParamGenerator::iterator end6_;
+    typename ParamGenerator::iterator current6_;
+    const typename ParamGenerator::iterator begin7_;
+    const typename ParamGenerator::iterator end7_;
+    typename ParamGenerator::iterator current7_;
+    const typename ParamGenerator::iterator begin8_;
+    const typename ParamGenerator::iterator end8_;
+    typename ParamGenerator::iterator current8_;
+    const typename ParamGenerator::iterator begin9_;
+    const typename ParamGenerator::iterator end9_;
+    typename ParamGenerator::iterator current9_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator9::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator9& other);
+
+  const ParamGenerator g1_;
+  const ParamGenerator g2_;
+  const ParamGenerator g3_;
+  const ParamGenerator g4_;
+  const ParamGenerator g5_;
+  const ParamGenerator g6_;
+  const ParamGenerator g7_;
+  const ParamGenerator g8_;
+  const ParamGenerator g9_;
+};  // class CartesianProductGenerator9
+
+
+template 
+class CartesianProductGenerator10
+    : public ParamGeneratorInterface< ::std::tr1::tuple > {
+ public:
+  typedef ::std::tr1::tuple ParamType;
+
+  CartesianProductGenerator10(const ParamGenerator& g1,
+      const ParamGenerator& g2, const ParamGenerator& g3,
+      const ParamGenerator& g4, const ParamGenerator& g5,
+      const ParamGenerator& g6, const ParamGenerator& g7,
+      const ParamGenerator& g8, const ParamGenerator& g9,
+      const ParamGenerator& g10)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9), g10_(g10) {}
+  virtual ~CartesianProductGenerator10() {}
+
+  virtual ParamIteratorInterface* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin(), g8_, g8_.begin(), g9_, g9_.begin(), g10_, g10_.begin());
+  }
+  virtual ParamIteratorInterface* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end(), g8_,
+        g8_.end(), g9_, g9_.end(), g10_, g10_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface {
+   public:
+    Iterator(const ParamGeneratorInterface* base,
+      const ParamGenerator& g1,
+      const typename ParamGenerator::iterator& current1,
+      const ParamGenerator& g2,
+      const typename ParamGenerator::iterator& current2,
+      const ParamGenerator& g3,
+      const typename ParamGenerator::iterator& current3,
+      const ParamGenerator& g4,
+      const typename ParamGenerator::iterator& current4,
+      const ParamGenerator& g5,
+      const typename ParamGenerator::iterator& current5,
+      const ParamGenerator& g6,
+      const typename ParamGenerator::iterator& current6,
+      const ParamGenerator& g7,
+      const typename ParamGenerator::iterator& current7,
+      const ParamGenerator& g8,
+      const typename ParamGenerator::iterator& current8,
+      const ParamGenerator& g9,
+      const typename ParamGenerator::iterator& current9,
+      const ParamGenerator& g10,
+      const typename ParamGenerator::iterator& current10)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7),
+          begin8_(g8.begin()), end8_(g8.end()), current8_(current8),
+          begin9_(g9.begin()), end9_(g9.end()), current9_(current9),
+          begin10_(g10.begin()), end10_(g10.end()), current10_(current10)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current10_;
+      if (current10_ == end10_) {
+        current10_ = begin10_;
+        ++current9_;
+      }
+      if (current9_ == end9_) {
+        current9_ = begin9_;
+        ++current8_;
+      }
+      if (current8_ == end8_) {
+        current8_ = begin8_;
+        ++current7_;
+      }
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return ¤t_value_; }
+    virtual bool Equals(const ParamIteratorInterface& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_ &&
+          current8_ == typed_other->current8_ &&
+          current9_ == typed_other->current9_ &&
+          current10_ == typed_other->current10_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_),
+        begin8_(other.begin8_),
+        end8_(other.end8_),
+        current8_(other.current8_),
+        begin9_(other.begin9_),
+        end9_(other.end9_),
+        current9_(other.current9_),
+        begin10_(other.begin10_),
+        end10_(other.end10_),
+        current10_(other.current10_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_, *current8_,
+            *current9_, *current10_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_ ||
+          current8_ == end8_ ||
+          current9_ == end9_ ||
+          current10_ == end10_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator::iterator begin1_;
+    const typename ParamGenerator::iterator end1_;
+    typename ParamGenerator::iterator current1_;
+    const typename ParamGenerator::iterator begin2_;
+    const typename ParamGenerator::iterator end2_;
+    typename ParamGenerator::iterator current2_;
+    const typename ParamGenerator::iterator begin3_;
+    const typename ParamGenerator::iterator end3_;
+    typename ParamGenerator::iterator current3_;
+    const typename ParamGenerator::iterator begin4_;
+    const typename ParamGenerator::iterator end4_;
+    typename ParamGenerator::iterator current4_;
+    const typename ParamGenerator::iterator begin5_;
+    const typename ParamGenerator::iterator end5_;
+    typename ParamGenerator::iterator current5_;
+    const typename ParamGenerator::iterator begin6_;
+    const typename ParamGenerator::iterator end6_;
+    typename ParamGenerator::iterator current6_;
+    const typename ParamGenerator::iterator begin7_;
+    const typename ParamGenerator::iterator end7_;
+    typename ParamGenerator::iterator current7_;
+    const typename ParamGenerator::iterator begin8_;
+    const typename ParamGenerator::iterator end8_;
+    typename ParamGenerator::iterator current8_;
+    const typename ParamGenerator::iterator begin9_;
+    const typename ParamGenerator::iterator end9_;
+    typename ParamGenerator::iterator current9_;
+    const typename ParamGenerator::iterator begin10_;
+    const typename ParamGenerator::iterator end10_;
+    typename ParamGenerator::iterator current10_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator10::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator10& other);
+
+  const ParamGenerator g1_;
+  const ParamGenerator g2_;
+  const ParamGenerator g3_;
+  const ParamGenerator g4_;
+  const ParamGenerator g5_;
+  const ParamGenerator g6_;
+  const ParamGenerator g7_;
+  const ParamGenerator g8_;
+  const ParamGenerator g9_;
+  const ParamGenerator g10_;
+};  // class CartesianProductGenerator10
+
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Helper classes providing Combine() with polymorphic features. They allow
+// casting CartesianProductGeneratorN to ParamGenerator if T is
+// convertible to U.
+//
+template 
+class CartesianProductHolder2 {
+ public:
+CartesianProductHolder2(const Generator1& g1, const Generator2& g2)
+      : g1_(g1), g2_(g2) {}
+  template 
+  operator ParamGenerator< ::std::tr1::tuple >() const {
+    return ParamGenerator< ::std::tr1::tuple >(
+        new CartesianProductGenerator2(
+        static_cast >(g1_),
+        static_cast >(g2_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder2& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+};  // class CartesianProductHolder2
+
+template 
+class CartesianProductHolder3 {
+ public:
+CartesianProductHolder3(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3)
+      : g1_(g1), g2_(g2), g3_(g3) {}
+  template 
+  operator ParamGenerator< ::std::tr1::tuple >() const {
+    return ParamGenerator< ::std::tr1::tuple >(
+        new CartesianProductGenerator3(
+        static_cast >(g1_),
+        static_cast >(g2_),
+        static_cast >(g3_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder3& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+};  // class CartesianProductHolder3
+
+template 
+class CartesianProductHolder4 {
+ public:
+CartesianProductHolder4(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4) {}
+  template 
+  operator ParamGenerator< ::std::tr1::tuple >() const {
+    return ParamGenerator< ::std::tr1::tuple >(
+        new CartesianProductGenerator4(
+        static_cast >(g1_),
+        static_cast >(g2_),
+        static_cast >(g3_),
+        static_cast >(g4_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder4& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+};  // class CartesianProductHolder4
+
+template 
+class CartesianProductHolder5 {
+ public:
+CartesianProductHolder5(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5) {}
+  template 
+  operator ParamGenerator< ::std::tr1::tuple >() const {
+    return ParamGenerator< ::std::tr1::tuple >(
+        new CartesianProductGenerator5(
+        static_cast >(g1_),
+        static_cast >(g2_),
+        static_cast >(g3_),
+        static_cast >(g4_),
+        static_cast >(g5_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder5& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+};  // class CartesianProductHolder5
+
+template 
+class CartesianProductHolder6 {
+ public:
+CartesianProductHolder6(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6) {}
+  template 
+  operator ParamGenerator< ::std::tr1::tuple >() const {
+    return ParamGenerator< ::std::tr1::tuple >(
+        new CartesianProductGenerator6(
+        static_cast >(g1_),
+        static_cast >(g2_),
+        static_cast >(g3_),
+        static_cast >(g4_),
+        static_cast >(g5_),
+        static_cast >(g6_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder6& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+};  // class CartesianProductHolder6
+
+template 
+class CartesianProductHolder7 {
+ public:
+CartesianProductHolder7(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7) {}
+  template 
+  operator ParamGenerator< ::std::tr1::tuple >() const {
+    return ParamGenerator< ::std::tr1::tuple >(
+        new CartesianProductGenerator7(
+        static_cast >(g1_),
+        static_cast >(g2_),
+        static_cast >(g3_),
+        static_cast >(g4_),
+        static_cast >(g5_),
+        static_cast >(g6_),
+        static_cast >(g7_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder7& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+};  // class CartesianProductHolder7
+
+template 
+class CartesianProductHolder8 {
+ public:
+CartesianProductHolder8(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7, const Generator8& g8)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7),
+          g8_(g8) {}
+  template 
+  operator ParamGenerator< ::std::tr1::tuple >() const {
+    return ParamGenerator< ::std::tr1::tuple >(
+        new CartesianProductGenerator8(
+        static_cast >(g1_),
+        static_cast >(g2_),
+        static_cast >(g3_),
+        static_cast >(g4_),
+        static_cast >(g5_),
+        static_cast >(g6_),
+        static_cast >(g7_),
+        static_cast >(g8_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder8& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+  const Generator8 g8_;
+};  // class CartesianProductHolder8
+
+template 
+class CartesianProductHolder9 {
+ public:
+CartesianProductHolder9(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7, const Generator8& g8,
+    const Generator9& g9)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9) {}
+  template 
+  operator ParamGenerator< ::std::tr1::tuple >() const {
+    return ParamGenerator< ::std::tr1::tuple >(
+        new CartesianProductGenerator9(
+        static_cast >(g1_),
+        static_cast >(g2_),
+        static_cast >(g3_),
+        static_cast >(g4_),
+        static_cast >(g5_),
+        static_cast >(g6_),
+        static_cast >(g7_),
+        static_cast >(g8_),
+        static_cast >(g9_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder9& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+  const Generator8 g8_;
+  const Generator9 g9_;
+};  // class CartesianProductHolder9
+
+template 
+class CartesianProductHolder10 {
+ public:
+CartesianProductHolder10(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7, const Generator8& g8,
+    const Generator9& g9, const Generator10& g10)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9), g10_(g10) {}
+  template 
+  operator ParamGenerator< ::std::tr1::tuple >() const {
+    return ParamGenerator< ::std::tr1::tuple >(
+        new CartesianProductGenerator10(
+        static_cast >(g1_),
+        static_cast >(g2_),
+        static_cast >(g3_),
+        static_cast >(g4_),
+        static_cast >(g5_),
+        static_cast >(g6_),
+        static_cast >(g7_),
+        static_cast >(g8_),
+        static_cast >(g9_),
+        static_cast >(g10_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder10& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+  const Generator8 g8_;
+  const Generator9 g9_;
+  const Generator10 g10_;
+};  // class CartesianProductHolder10
+
+# endif  // GTEST_HAS_COMBINE
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  //  GTEST_HAS_PARAM_TEST
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
+
+#if GTEST_HAS_PARAM_TEST
+
+namespace testing {
+
+// Functions producing parameter generators.
+//
+// Google Test uses these generators to produce parameters for value-
+// parameterized tests. When a parameterized test case is instantiated
+// with a particular generator, Google Test creates and runs tests
+// for each element in the sequence produced by the generator.
+//
+// In the following sample, tests from test case FooTest are instantiated
+// each three times with parameter values 3, 5, and 8:
+//
+// class FooTest : public TestWithParam { ... };
+//
+// TEST_P(FooTest, TestThis) {
+// }
+// TEST_P(FooTest, TestThat) {
+// }
+// INSTANTIATE_TEST_CASE_P(TestSequence, FooTest, Values(3, 5, 8));
+//
+
+// Range() returns generators providing sequences of values in a range.
+//
+// Synopsis:
+// Range(start, end)
+//   - returns a generator producing a sequence of values {start, start+1,
+//     start+2, ..., }.
+// Range(start, end, step)
+//   - returns a generator producing a sequence of values {start, start+step,
+//     start+step+step, ..., }.
+// Notes:
+//   * The generated sequences never include end. For example, Range(1, 5)
+//     returns a generator producing a sequence {1, 2, 3, 4}. Range(1, 9, 2)
+//     returns a generator producing {1, 3, 5, 7}.
+//   * start and end must have the same type. That type may be any integral or
+//     floating-point type or a user defined type satisfying these conditions:
+//     * It must be assignable (have operator=() defined).
+//     * It must have operator+() (operator+(int-compatible type) for
+//       two-operand version).
+//     * It must have operator<() defined.
+//     Elements in the resulting sequences will also have that type.
+//   * Condition start < end must be satisfied in order for resulting sequences
+//     to contain any elements.
+//
+template 
+internal::ParamGenerator Range(T start, T end, IncrementT step) {
+  return internal::ParamGenerator(
+      new internal::RangeGenerator(start, end, step));
+}
+
+template 
+internal::ParamGenerator Range(T start, T end) {
+  return Range(start, end, 1);
+}
+
+// ValuesIn() function allows generation of tests with parameters coming from
+// a container.
+//
+// Synopsis:
+// ValuesIn(const T (&array)[N])
+//   - returns a generator producing sequences with elements from
+//     a C-style array.
+// ValuesIn(const Container& container)
+//   - returns a generator producing sequences with elements from
+//     an STL-style container.
+// ValuesIn(Iterator begin, Iterator end)
+//   - returns a generator producing sequences with elements from
+//     a range [begin, end) defined by a pair of STL-style iterators. These
+//     iterators can also be plain C pointers.
+//
+// Please note that ValuesIn copies the values from the containers
+// passed in and keeps them to generate tests in RUN_ALL_TESTS().
+//
+// Examples:
+//
+// This instantiates tests from test case StringTest
+// each with C-string values of "foo", "bar", and "baz":
+//
+// const char* strings[] = {"foo", "bar", "baz"};
+// INSTANTIATE_TEST_CASE_P(StringSequence, SrtingTest, ValuesIn(strings));
+//
+// This instantiates tests from test case StlStringTest
+// each with STL strings with values "a" and "b":
+//
+// ::std::vector< ::std::string> GetParameterStrings() {
+//   ::std::vector< ::std::string> v;
+//   v.push_back("a");
+//   v.push_back("b");
+//   return v;
+// }
+//
+// INSTANTIATE_TEST_CASE_P(CharSequence,
+//                         StlStringTest,
+//                         ValuesIn(GetParameterStrings()));
+//
+//
+// This will also instantiate tests from CharTest
+// each with parameter values 'a' and 'b':
+//
+// ::std::list GetParameterChars() {
+//   ::std::list list;
+//   list.push_back('a');
+//   list.push_back('b');
+//   return list;
+// }
+// ::std::list l = GetParameterChars();
+// INSTANTIATE_TEST_CASE_P(CharSequence2,
+//                         CharTest,
+//                         ValuesIn(l.begin(), l.end()));
+//
+template 
+internal::ParamGenerator<
+  typename ::testing::internal::IteratorTraits::value_type>
+ValuesIn(ForwardIterator begin, ForwardIterator end) {
+  typedef typename ::testing::internal::IteratorTraits
+      ::value_type ParamType;
+  return internal::ParamGenerator(
+      new internal::ValuesInIteratorRangeGenerator(begin, end));
+}
+
+template 
+internal::ParamGenerator ValuesIn(const T (&array)[N]) {
+  return ValuesIn(array, array + N);
+}
+
+template 
+internal::ParamGenerator ValuesIn(
+    const Container& container) {
+  return ValuesIn(container.begin(), container.end());
+}
+
+// Values() allows generating tests from explicitly specified list of
+// parameters.
+//
+// Synopsis:
+// Values(T v1, T v2, ..., T vN)
+//   - returns a generator producing sequences with elements v1, v2, ..., vN.
+//
+// For example, this instantiates tests from test case BarTest each
+// with values "one", "two", and "three":
+//
+// INSTANTIATE_TEST_CASE_P(NumSequence, BarTest, Values("one", "two", "three"));
+//
+// This instantiates tests from test case BazTest each with values 1, 2, 3.5.
+// The exact type of values will depend on the type of parameter in BazTest.
+//
+// INSTANTIATE_TEST_CASE_P(FloatingNumbers, BazTest, Values(1, 2, 3.5));
+//
+// Currently, Values() supports from 1 to 50 parameters.
+//
+template 
+internal::ValueArray1 Values(T1 v1) {
+  return internal::ValueArray1(v1);
+}
+
+template 
+internal::ValueArray2 Values(T1 v1, T2 v2) {
+  return internal::ValueArray2(v1, v2);
+}
+
+template 
+internal::ValueArray3 Values(T1 v1, T2 v2, T3 v3) {
+  return internal::ValueArray3(v1, v2, v3);
+}
+
+template 
+internal::ValueArray4 Values(T1 v1, T2 v2, T3 v3, T4 v4) {
+  return internal::ValueArray4(v1, v2, v3, v4);
+}
+
+template 
+internal::ValueArray5 Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5) {
+  return internal::ValueArray5(v1, v2, v3, v4, v5);
+}
+
+template 
+internal::ValueArray6 Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6) {
+  return internal::ValueArray6(v1, v2, v3, v4, v5, v6);
+}
+
+template 
+internal::ValueArray7 Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6, T7 v7) {
+  return internal::ValueArray7(v1, v2, v3, v4, v5,
+      v6, v7);
+}
+
+template 
+internal::ValueArray8 Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8) {
+  return internal::ValueArray8(v1, v2, v3, v4,
+      v5, v6, v7, v8);
+}
+
+template 
+internal::ValueArray9 Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9) {
+  return internal::ValueArray9(v1, v2, v3,
+      v4, v5, v6, v7, v8, v9);
+}
+
+template 
+internal::ValueArray10 Values(T1 v1,
+    T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10) {
+  return internal::ValueArray10(v1,
+      v2, v3, v4, v5, v6, v7, v8, v9, v10);
+}
+
+template 
+internal::ValueArray11 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11) {
+  return internal::ValueArray11(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11);
+}
+
+template 
+internal::ValueArray12 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12) {
+  return internal::ValueArray12(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12);
+}
+
+template 
+internal::ValueArray13 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13) {
+  return internal::ValueArray13(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13);
+}
+
+template 
+internal::ValueArray14 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14) {
+  return internal::ValueArray14(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+      v14);
+}
+
+template 
+internal::ValueArray15 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+    T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15) {
+  return internal::ValueArray15(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+      v13, v14, v15);
+}
+
+template 
+internal::ValueArray16 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16) {
+  return internal::ValueArray16(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+      v12, v13, v14, v15, v16);
+}
+
+template 
+internal::ValueArray17 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17) {
+  return internal::ValueArray17(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10,
+      v11, v12, v13, v14, v15, v16, v17);
+}
+
+template 
+internal::ValueArray18 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6,
+    T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18) {
+  return internal::ValueArray18(v1, v2, v3, v4, v5, v6, v7, v8, v9,
+      v10, v11, v12, v13, v14, v15, v16, v17, v18);
+}
+
+template 
+internal::ValueArray19 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5,
+    T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14,
+    T15 v15, T16 v16, T17 v17, T18 v18, T19 v19) {
+  return internal::ValueArray19(v1, v2, v3, v4, v5, v6, v7, v8,
+      v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19);
+}
+
+template 
+internal::ValueArray20 Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20) {
+  return internal::ValueArray20(v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20);
+}
+
+template 
+internal::ValueArray21 Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21) {
+  return internal::ValueArray21(v1, v2, v3, v4, v5, v6,
+      v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21);
+}
+
+template 
+internal::ValueArray22 Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22) {
+  return internal::ValueArray22(v1, v2, v3, v4,
+      v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22);
+}
+
+template 
+internal::ValueArray23 Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23) {
+  return internal::ValueArray23(v1, v2, v3,
+      v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23);
+}
+
+template 
+internal::ValueArray24 Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24) {
+  return internal::ValueArray24(v1, v2,
+      v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18,
+      v19, v20, v21, v22, v23, v24);
+}
+
+template 
+internal::ValueArray25 Values(T1 v1,
+    T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11,
+    T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19,
+    T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25) {
+  return internal::ValueArray25(v1,
+      v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17,
+      v18, v19, v20, v21, v22, v23, v24, v25);
+}
+
+template 
+internal::ValueArray26 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26) {
+  return internal::ValueArray26(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15,
+      v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26);
+}
+
+template 
+internal::ValueArray27 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27) {
+  return internal::ValueArray27(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14,
+      v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27);
+}
+
+template 
+internal::ValueArray28 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28) {
+  return internal::ValueArray28(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+      v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27,
+      v28);
+}
+
+template 
+internal::ValueArray29 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29) {
+  return internal::ValueArray29(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+      v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26,
+      v27, v28, v29);
+}
+
+template 
+internal::ValueArray30 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+    T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16,
+    T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24,
+    T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30) {
+  return internal::ValueArray30(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+      v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25,
+      v26, v27, v28, v29, v30);
+}
+
+template 
+internal::ValueArray31 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31) {
+  return internal::ValueArray31(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10,
+      v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24,
+      v25, v26, v27, v28, v29, v30, v31);
+}
+
+template 
+internal::ValueArray32 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32) {
+  return internal::ValueArray32(v1, v2, v3, v4, v5, v6, v7, v8, v9,
+      v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32);
+}
+
+template 
+internal::ValueArray33 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6,
+    T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33) {
+  return internal::ValueArray33(v1, v2, v3, v4, v5, v6, v7, v8,
+      v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32, v33);
+}
+
+template 
+internal::ValueArray34 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5,
+    T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14,
+    T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22,
+    T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30,
+    T31 v31, T32 v32, T33 v33, T34 v34) {
+  return internal::ValueArray34(v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22,
+      v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34);
+}
+
+template 
+internal::ValueArray35 Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21,
+    T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29,
+    T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35) {
+  return internal::ValueArray35(v1, v2, v3, v4, v5, v6,
+      v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21,
+      v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35);
+}
+
+template 
+internal::ValueArray36 Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21,
+    T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29,
+    T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36) {
+  return internal::ValueArray36(v1, v2, v3, v4,
+      v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33,
+      v34, v35, v36);
+}
+
+template 
+internal::ValueArray37 Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28,
+    T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36,
+    T37 v37) {
+  return internal::ValueArray37(v1, v2, v3,
+      v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33,
+      v34, v35, v36, v37);
+}
+
+template 
+internal::ValueArray38 Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28,
+    T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36,
+    T37 v37, T38 v38) {
+  return internal::ValueArray38(v1, v2,
+      v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18,
+      v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32,
+      v33, v34, v35, v36, v37, v38);
+}
+
+template 
+internal::ValueArray39 Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28,
+    T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36,
+    T37 v37, T38 v38, T39 v39) {
+  return internal::ValueArray39(v1,
+      v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17,
+      v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31,
+      v32, v33, v34, v35, v36, v37, v38, v39);
+}
+
+template 
+internal::ValueArray40 Values(T1 v1,
+    T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11,
+    T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19,
+    T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27,
+    T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35,
+    T36 v36, T37 v37, T38 v38, T39 v39, T40 v40) {
+  return internal::ValueArray40(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15,
+      v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29,
+      v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40);
+}
+
+template 
+internal::ValueArray41 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41) {
+  return internal::ValueArray41(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14,
+      v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28,
+      v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41);
+}
+
+template 
+internal::ValueArray42 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+    T42 v42) {
+  return internal::ValueArray42(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+      v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27,
+      v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41,
+      v42);
+}
+
+template 
+internal::ValueArray43 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+    T42 v42, T43 v43) {
+  return internal::ValueArray43(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+      v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26,
+      v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40,
+      v41, v42, v43);
+}
+
+template 
+internal::ValueArray44 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+    T42 v42, T43 v43, T44 v44) {
+  return internal::ValueArray44(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+      v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25,
+      v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39,
+      v40, v41, v42, v43, v44);
+}
+
+template 
+internal::ValueArray45 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+    T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16,
+    T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24,
+    T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32,
+    T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40,
+    T41 v41, T42 v42, T43 v43, T44 v44, T45 v45) {
+  return internal::ValueArray45(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10,
+      v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24,
+      v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38,
+      v39, v40, v41, v42, v43, v44, v45);
+}
+
+template 
+internal::ValueArray46 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39,
+    T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46) {
+  return internal::ValueArray46(v1, v2, v3, v4, v5, v6, v7, v8, v9,
+      v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37,
+      v38, v39, v40, v41, v42, v43, v44, v45, v46);
+}
+
+template 
+internal::ValueArray47 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39,
+    T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47) {
+  return internal::ValueArray47(v1, v2, v3, v4, v5, v6, v7, v8,
+      v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37,
+      v38, v39, v40, v41, v42, v43, v44, v45, v46, v47);
+}
+
+template 
+internal::ValueArray48 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6,
+    T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39,
+    T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47,
+    T48 v48) {
+  return internal::ValueArray48(v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22,
+      v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36,
+      v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48);
+}
+
+template 
+internal::ValueArray49 Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5,
+    T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14,
+    T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22,
+    T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30,
+    T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38,
+    T39 v39, T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46,
+    T47 v47, T48 v48, T49 v49) {
+  return internal::ValueArray49(v1, v2, v3, v4, v5, v6,
+      v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21,
+      v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35,
+      v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49);
+}
+
+template 
+internal::ValueArray50 Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21,
+    T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29,
+    T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37,
+    T38 v38, T39 v39, T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45,
+    T46 v46, T47 v47, T48 v48, T49 v49, T50 v50) {
+  return internal::ValueArray50(v1, v2, v3, v4,
+      v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33,
+      v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47,
+      v48, v49, v50);
+}
+
+// Bool() allows generating tests with parameters in a set of (false, true).
+//
+// Synopsis:
+// Bool()
+//   - returns a generator producing sequences with elements {false, true}.
+//
+// It is useful when testing code that depends on Boolean flags. Combinations
+// of multiple flags can be tested when several Bool()'s are combined using
+// Combine() function.
+//
+// In the following example all tests in the test case FlagDependentTest
+// will be instantiated twice with parameters false and true.
+//
+// class FlagDependentTest : public testing::TestWithParam {
+//   virtual void SetUp() {
+//     external_flag = GetParam();
+//   }
+// }
+// INSTANTIATE_TEST_CASE_P(BoolSequence, FlagDependentTest, Bool());
+//
+inline internal::ParamGenerator Bool() {
+  return Values(false, true);
+}
+
+# if GTEST_HAS_COMBINE
+// Combine() allows the user to combine two or more sequences to produce
+// values of a Cartesian product of those sequences' elements.
+//
+// Synopsis:
+// Combine(gen1, gen2, ..., genN)
+//   - returns a generator producing sequences with elements coming from
+//     the Cartesian product of elements from the sequences generated by
+//     gen1, gen2, ..., genN. The sequence elements will have a type of
+//     tuple where T1, T2, ..., TN are the types
+//     of elements from sequences produces by gen1, gen2, ..., genN.
+//
+// Combine can have up to 10 arguments. This number is currently limited
+// by the maximum number of elements in the tuple implementation used by Google
+// Test.
+//
+// Example:
+//
+// This will instantiate tests in test case AnimalTest each one with
+// the parameter values tuple("cat", BLACK), tuple("cat", WHITE),
+// tuple("dog", BLACK), and tuple("dog", WHITE):
+//
+// enum Color { BLACK, GRAY, WHITE };
+// class AnimalTest
+//     : public testing::TestWithParam > {...};
+//
+// TEST_P(AnimalTest, AnimalLooksNice) {...}
+//
+// INSTANTIATE_TEST_CASE_P(AnimalVariations, AnimalTest,
+//                         Combine(Values("cat", "dog"),
+//                                 Values(BLACK, WHITE)));
+//
+// This will instantiate tests in FlagDependentTest with all variations of two
+// Boolean flags:
+//
+// class FlagDependentTest
+//     : public testing::TestWithParam > {
+//   virtual void SetUp() {
+//     // Assigns external_flag_1 and external_flag_2 values from the tuple.
+//     tie(external_flag_1, external_flag_2) = GetParam();
+//   }
+// };
+//
+// TEST_P(FlagDependentTest, TestFeature1) {
+//   // Test your code using external_flag_1 and external_flag_2 here.
+// }
+// INSTANTIATE_TEST_CASE_P(TwoBoolSequence, FlagDependentTest,
+//                         Combine(Bool(), Bool()));
+//
+template 
+internal::CartesianProductHolder2 Combine(
+    const Generator1& g1, const Generator2& g2) {
+  return internal::CartesianProductHolder2(
+      g1, g2);
+}
+
+template 
+internal::CartesianProductHolder3 Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3) {
+  return internal::CartesianProductHolder3(
+      g1, g2, g3);
+}
+
+template 
+internal::CartesianProductHolder4 Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4) {
+  return internal::CartesianProductHolder4(
+      g1, g2, g3, g4);
+}
+
+template 
+internal::CartesianProductHolder5 Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5) {
+  return internal::CartesianProductHolder5(
+      g1, g2, g3, g4, g5);
+}
+
+template 
+internal::CartesianProductHolder6 Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6) {
+  return internal::CartesianProductHolder6(
+      g1, g2, g3, g4, g5, g6);
+}
+
+template 
+internal::CartesianProductHolder7 Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7) {
+  return internal::CartesianProductHolder7(
+      g1, g2, g3, g4, g5, g6, g7);
+}
+
+template 
+internal::CartesianProductHolder8 Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7, const Generator8& g8) {
+  return internal::CartesianProductHolder8(
+      g1, g2, g3, g4, g5, g6, g7, g8);
+}
+
+template 
+internal::CartesianProductHolder9 Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7, const Generator8& g8, const Generator9& g9) {
+  return internal::CartesianProductHolder9(
+      g1, g2, g3, g4, g5, g6, g7, g8, g9);
+}
+
+template 
+internal::CartesianProductHolder10 Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7, const Generator8& g8, const Generator9& g9,
+        const Generator10& g10) {
+  return internal::CartesianProductHolder10(
+      g1, g2, g3, g4, g5, g6, g7, g8, g9, g10);
+}
+# endif  // GTEST_HAS_COMBINE
+
+
+
+# define TEST_P(test_case_name, test_name) \
+  class GTEST_TEST_CLASS_NAME_(test_case_name, test_name) \
+      : public test_case_name { \
+   public: \
+    GTEST_TEST_CLASS_NAME_(test_case_name, test_name)() {} \
+    virtual void TestBody(); \
+   private: \
+    static int AddToRegistry() { \
+      ::testing::UnitTest::GetInstance()->parameterized_test_registry(). \
+          GetTestCasePatternHolder(\
+              #test_case_name, __FILE__, __LINE__)->AddTestPattern(\
+                  #test_case_name, \
+                  #test_name, \
+                  new ::testing::internal::TestMetaFactory< \
+                      GTEST_TEST_CLASS_NAME_(test_case_name, test_name)>()); \
+      return 0; \
+    } \
+    static int gtest_registering_dummy_; \
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(\
+        GTEST_TEST_CLASS_NAME_(test_case_name, test_name)); \
+  }; \
+  int GTEST_TEST_CLASS_NAME_(test_case_name, \
+                             test_name)::gtest_registering_dummy_ = \
+      GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::AddToRegistry(); \
+  void GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::TestBody()
+
+# define INSTANTIATE_TEST_CASE_P(prefix, test_case_name, generator) \
+  ::testing::internal::ParamGenerator \
+      gtest_##prefix##test_case_name##_EvalGenerator_() { return generator; } \
+  int gtest_##prefix##test_case_name##_dummy_ = \
+      ::testing::UnitTest::GetInstance()->parameterized_test_registry(). \
+          GetTestCasePatternHolder(\
+              #test_case_name, __FILE__, __LINE__)->AddTestCaseInstantiation(\
+                  #prefix, \
+                  >est_##prefix##test_case_name##_EvalGenerator_, \
+                  __FILE__, __LINE__)
+
+}  // namespace testing
+
+#endif  // GTEST_HAS_PARAM_TEST
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
+// Copyright 2006, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// Google C++ Testing Framework definitions useful in production code.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PROD_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PROD_H_
+
+// When you need to test the private or protected members of a class,
+// use the FRIEND_TEST macro to declare your tests as friends of the
+// class.  For example:
+//
+// class MyClass {
+//  private:
+//   void MyMethod();
+//   FRIEND_TEST(MyClassTest, MyMethod);
+// };
+//
+// class MyClassTest : public testing::Test {
+//   // ...
+// };
+//
+// TEST_F(MyClassTest, MyMethod) {
+//   // Can call MyClass::MyMethod() here.
+// }
+
+#define FRIEND_TEST(test_case_name, test_name)\
+friend class test_case_name##_##test_name##_Test
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PROD_H_
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: mheule@google.com (Markus Heule)
+//
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_TEST_PART_H_
+#define GTEST_INCLUDE_GTEST_GTEST_TEST_PART_H_
+
+#include 
+#include 
+
+namespace testing {
+
+// A copyable object representing the result of a test part (i.e. an
+// assertion or an explicit FAIL(), ADD_FAILURE(), or SUCCESS()).
+//
+// Don't inherit from TestPartResult as its destructor is not virtual.
+class GTEST_API_ TestPartResult {
+ public:
+  // The possible outcomes of a test part (i.e. an assertion or an
+  // explicit SUCCEED(), FAIL(), or ADD_FAILURE()).
+  enum Type {
+    kSuccess,          // Succeeded.
+    kNonFatalFailure,  // Failed but the test can continue.
+    kFatalFailure      // Failed and the test should be terminated.
+  };
+
+  // C'tor.  TestPartResult does NOT have a default constructor.
+  // Always use this constructor (with parameters) to create a
+  // TestPartResult object.
+  TestPartResult(Type a_type,
+                 const char* a_file_name,
+                 int a_line_number,
+                 const char* a_message)
+      : type_(a_type),
+        file_name_(a_file_name == NULL ? "" : a_file_name),
+        line_number_(a_line_number),
+        summary_(ExtractSummary(a_message)),
+        message_(a_message) {
+  }
+
+  // Gets the outcome of the test part.
+  Type type() const { return type_; }
+
+  // Gets the name of the source file where the test part took place, or
+  // NULL if it's unknown.
+  const char* file_name() const {
+    return file_name_.empty() ? NULL : file_name_.c_str();
+  }
+
+  // Gets the line in the source file where the test part took place,
+  // or -1 if it's unknown.
+  int line_number() const { return line_number_; }
+
+  // Gets the summary of the failure message.
+  const char* summary() const { return summary_.c_str(); }
+
+  // Gets the message associated with the test part.
+  const char* message() const { return message_.c_str(); }
+
+  // Returns true iff the test part passed.
+  bool passed() const { return type_ == kSuccess; }
+
+  // Returns true iff the test part failed.
+  bool failed() const { return type_ != kSuccess; }
+
+  // Returns true iff the test part non-fatally failed.
+  bool nonfatally_failed() const { return type_ == kNonFatalFailure; }
+
+  // Returns true iff the test part fatally failed.
+  bool fatally_failed() const { return type_ == kFatalFailure; }
+
+ private:
+  Type type_;
+
+  // Gets the summary of the failure message by omitting the stack
+  // trace in it.
+  static std::string ExtractSummary(const char* message);
+
+  // The name of the source file where the test part took place, or
+  // "" if the source file is unknown.
+  std::string file_name_;
+  // The line in the source file where the test part took place, or -1
+  // if the line number is unknown.
+  int line_number_;
+  std::string summary_;  // The test failure summary.
+  std::string message_;  // The test failure message.
+};
+
+// Prints a TestPartResult object.
+std::ostream& operator<<(std::ostream& os, const TestPartResult& result);
+
+// An array of TestPartResult objects.
+//
+// Don't inherit from TestPartResultArray as its destructor is not
+// virtual.
+class GTEST_API_ TestPartResultArray {
+ public:
+  TestPartResultArray() {}
+
+  // Appends the given TestPartResult to the array.
+  void Append(const TestPartResult& result);
+
+  // Returns the TestPartResult at the given index (0-based).
+  const TestPartResult& GetTestPartResult(int index) const;
+
+  // Returns the number of TestPartResult objects in the array.
+  int size() const;
+
+ private:
+  std::vector array_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestPartResultArray);
+};
+
+// This interface knows how to report a test part result.
+class TestPartResultReporterInterface {
+ public:
+  virtual ~TestPartResultReporterInterface() {}
+
+  virtual void ReportTestPartResult(const TestPartResult& result) = 0;
+};
+
+namespace internal {
+
+// This helper class is used by {ASSERT|EXPECT}_NO_FATAL_FAILURE to check if a
+// statement generates new fatal failures. To do so it registers itself as the
+// current test part result reporter. Besides checking if fatal failures were
+// reported, it only delegates the reporting to the former result reporter.
+// The original result reporter is restored in the destructor.
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+class GTEST_API_ HasNewFatalFailureHelper
+    : public TestPartResultReporterInterface {
+ public:
+  HasNewFatalFailureHelper();
+  virtual ~HasNewFatalFailureHelper();
+  virtual void ReportTestPartResult(const TestPartResult& result);
+  bool has_new_fatal_failure() const { return has_new_fatal_failure_; }
+ private:
+  bool has_new_fatal_failure_;
+  TestPartResultReporterInterface* original_reporter_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(HasNewFatalFailureHelper);
+};
+
+}  // namespace internal
+
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_TEST_PART_H_
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_TYPED_TEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_TYPED_TEST_H_
+
+// This header implements typed tests and type-parameterized tests.
+
+// Typed (aka type-driven) tests repeat the same test for types in a
+// list.  You must know which types you want to test with when writing
+// typed tests. Here's how you do it:
+
+#if 0
+
+// First, define a fixture class template.  It should be parameterized
+// by a type.  Remember to derive it from testing::Test.
+template 
+class FooTest : public testing::Test {
+ public:
+  ...
+  typedef std::list List;
+  static T shared_;
+  T value_;
+};
+
+// Next, associate a list of types with the test case, which will be
+// repeated for each type in the list.  The typedef is necessary for
+// the macro to parse correctly.
+typedef testing::Types MyTypes;
+TYPED_TEST_CASE(FooTest, MyTypes);
+
+// If the type list contains only one type, you can write that type
+// directly without Types<...>:
+//   TYPED_TEST_CASE(FooTest, int);
+
+// Then, use TYPED_TEST() instead of TEST_F() to define as many typed
+// tests for this test case as you want.
+TYPED_TEST(FooTest, DoesBlah) {
+  // Inside a test, refer to TypeParam to get the type parameter.
+  // Since we are inside a derived class template, C++ requires use to
+  // visit the members of FooTest via 'this'.
+  TypeParam n = this->value_;
+
+  // To visit static members of the fixture, add the TestFixture::
+  // prefix.
+  n += TestFixture::shared_;
+
+  // To refer to typedefs in the fixture, add the "typename
+  // TestFixture::" prefix.
+  typename TestFixture::List values;
+  values.push_back(n);
+  ...
+}
+
+TYPED_TEST(FooTest, HasPropertyA) { ... }
+
+#endif  // 0
+
+// Type-parameterized tests are abstract test patterns parameterized
+// by a type.  Compared with typed tests, type-parameterized tests
+// allow you to define the test pattern without knowing what the type
+// parameters are.  The defined pattern can be instantiated with
+// different types any number of times, in any number of translation
+// units.
+//
+// If you are designing an interface or concept, you can define a
+// suite of type-parameterized tests to verify properties that any
+// valid implementation of the interface/concept should have.  Then,
+// each implementation can easily instantiate the test suite to verify
+// that it conforms to the requirements, without having to write
+// similar tests repeatedly.  Here's an example:
+
+#if 0
+
+// First, define a fixture class template.  It should be parameterized
+// by a type.  Remember to derive it from testing::Test.
+template 
+class FooTest : public testing::Test {
+  ...
+};
+
+// Next, declare that you will define a type-parameterized test case
+// (the _P suffix is for "parameterized" or "pattern", whichever you
+// prefer):
+TYPED_TEST_CASE_P(FooTest);
+
+// Then, use TYPED_TEST_P() to define as many type-parameterized tests
+// for this type-parameterized test case as you want.
+TYPED_TEST_P(FooTest, DoesBlah) {
+  // Inside a test, refer to TypeParam to get the type parameter.
+  TypeParam n = 0;
+  ...
+}
+
+TYPED_TEST_P(FooTest, HasPropertyA) { ... }
+
+// Now the tricky part: you need to register all test patterns before
+// you can instantiate them.  The first argument of the macro is the
+// test case name; the rest are the names of the tests in this test
+// case.
+REGISTER_TYPED_TEST_CASE_P(FooTest,
+                           DoesBlah, HasPropertyA);
+
+// Finally, you are free to instantiate the pattern with the types you
+// want.  If you put the above code in a header file, you can #include
+// it in multiple C++ source files and instantiate it multiple times.
+//
+// To distinguish different instances of the pattern, the first
+// argument to the INSTANTIATE_* macro is a prefix that will be added
+// to the actual test case name.  Remember to pick unique prefixes for
+// different instances.
+typedef testing::Types MyTypes;
+INSTANTIATE_TYPED_TEST_CASE_P(My, FooTest, MyTypes);
+
+// If the type list contains only one type, you can write that type
+// directly without Types<...>:
+//   INSTANTIATE_TYPED_TEST_CASE_P(My, FooTest, int);
+
+#endif  // 0
+
+
+// Implements typed tests.
+
+#if GTEST_HAS_TYPED_TEST
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Expands to the name of the typedef for the type parameters of the
+// given test case.
+# define GTEST_TYPE_PARAMS_(TestCaseName) gtest_type_params_##TestCaseName##_
+
+// The 'Types' template argument below must have spaces around it
+// since some compilers may choke on '>>' when passing a template
+// instance (e.g. Types)
+# define TYPED_TEST_CASE(CaseName, Types) \
+  typedef ::testing::internal::TypeList< Types >::type \
+      GTEST_TYPE_PARAMS_(CaseName)
+
+# define TYPED_TEST(CaseName, TestName) \
+  template  \
+  class GTEST_TEST_CLASS_NAME_(CaseName, TestName) \
+      : public CaseName { \
+   private: \
+    typedef CaseName TestFixture; \
+    typedef gtest_TypeParam_ TypeParam; \
+    virtual void TestBody(); \
+  }; \
+  bool gtest_##CaseName##_##TestName##_registered_ GTEST_ATTRIBUTE_UNUSED_ = \
+      ::testing::internal::TypeParameterizedTest< \
+          CaseName, \
+          ::testing::internal::TemplateSel< \
+              GTEST_TEST_CLASS_NAME_(CaseName, TestName)>, \
+          GTEST_TYPE_PARAMS_(CaseName)>::Register(\
+              "", #CaseName, #TestName, 0); \
+  template  \
+  void GTEST_TEST_CLASS_NAME_(CaseName, TestName)::TestBody()
+
+#endif  // GTEST_HAS_TYPED_TEST
+
+// Implements type-parameterized tests.
+
+#if GTEST_HAS_TYPED_TEST_P
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Expands to the namespace name that the type-parameterized tests for
+// the given type-parameterized test case are defined in.  The exact
+// name of the namespace is subject to change without notice.
+# define GTEST_CASE_NAMESPACE_(TestCaseName) \
+  gtest_case_##TestCaseName##_
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Expands to the name of the variable used to remember the names of
+// the defined tests in the given test case.
+# define GTEST_TYPED_TEST_CASE_P_STATE_(TestCaseName) \
+  gtest_typed_test_case_p_state_##TestCaseName##_
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE DIRECTLY.
+//
+// Expands to the name of the variable used to remember the names of
+// the registered tests in the given test case.
+# define GTEST_REGISTERED_TEST_NAMES_(TestCaseName) \
+  gtest_registered_test_names_##TestCaseName##_
+
+// The variables defined in the type-parameterized test macros are
+// static as typically these macros are used in a .h file that can be
+// #included in multiple translation units linked together.
+# define TYPED_TEST_CASE_P(CaseName) \
+  static ::testing::internal::TypedTestCasePState \
+      GTEST_TYPED_TEST_CASE_P_STATE_(CaseName)
+
+# define TYPED_TEST_P(CaseName, TestName) \
+  namespace GTEST_CASE_NAMESPACE_(CaseName) { \
+  template  \
+  class TestName : public CaseName { \
+   private: \
+    typedef CaseName TestFixture; \
+    typedef gtest_TypeParam_ TypeParam; \
+    virtual void TestBody(); \
+  }; \
+  static bool gtest_##TestName##_defined_ GTEST_ATTRIBUTE_UNUSED_ = \
+      GTEST_TYPED_TEST_CASE_P_STATE_(CaseName).AddTestName(\
+          __FILE__, __LINE__, #CaseName, #TestName); \
+  } \
+  template  \
+  void GTEST_CASE_NAMESPACE_(CaseName)::TestName::TestBody()
+
+# define REGISTER_TYPED_TEST_CASE_P(CaseName, ...) \
+  namespace GTEST_CASE_NAMESPACE_(CaseName) { \
+  typedef ::testing::internal::Templates<__VA_ARGS__>::type gtest_AllTests_; \
+  } \
+  static const char* const GTEST_REGISTERED_TEST_NAMES_(CaseName) = \
+      GTEST_TYPED_TEST_CASE_P_STATE_(CaseName).VerifyRegisteredTestNames(\
+          __FILE__, __LINE__, #__VA_ARGS__)
+
+// The 'Types' template argument below must have spaces around it
+// since some compilers may choke on '>>' when passing a template
+// instance (e.g. Types)
+# define INSTANTIATE_TYPED_TEST_CASE_P(Prefix, CaseName, Types) \
+  bool gtest_##Prefix##_##CaseName GTEST_ATTRIBUTE_UNUSED_ = \
+      ::testing::internal::TypeParameterizedTestCase::type>::Register(\
+              #Prefix, #CaseName, GTEST_REGISTERED_TEST_NAMES_(CaseName))
+
+#endif  // GTEST_HAS_TYPED_TEST_P
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_TYPED_TEST_H_
+
+// Depending on the platform, different string classes are available.
+// On Linux, in addition to ::std::string, Google also makes use of
+// class ::string, which has the same interface as ::std::string, but
+// has a different implementation.
+//
+// The user can define GTEST_HAS_GLOBAL_STRING to 1 to indicate that
+// ::string is available AND is a distinct type to ::std::string, or
+// define it to 0 to indicate otherwise.
+//
+// If the user's ::std::string and ::string are the same class due to
+// aliasing, he should define GTEST_HAS_GLOBAL_STRING to 0.
+//
+// If the user doesn't define GTEST_HAS_GLOBAL_STRING, it is defined
+// heuristically.
+
+namespace testing {
+
+// Declares the flags.
+
+// This flag temporary enables the disabled tests.
+GTEST_DECLARE_bool_(also_run_disabled_tests);
+
+// This flag brings the debugger on an assertion failure.
+GTEST_DECLARE_bool_(break_on_failure);
+
+// This flag controls whether Google Test catches all test-thrown exceptions
+// and logs them as failures.
+GTEST_DECLARE_bool_(catch_exceptions);
+
+// This flag enables using colors in terminal output. Available values are
+// "yes" to enable colors, "no" (disable colors), or "auto" (the default)
+// to let Google Test decide.
+GTEST_DECLARE_string_(color);
+
+// This flag sets up the filter to select by name using a glob pattern
+// the tests to run. If the filter is not given all tests are executed.
+GTEST_DECLARE_string_(filter);
+
+// This flag causes the Google Test to list tests. None of the tests listed
+// are actually run if the flag is provided.
+GTEST_DECLARE_bool_(list_tests);
+
+// This flag controls whether Google Test emits a detailed XML report to a file
+// in addition to its normal textual output.
+GTEST_DECLARE_string_(output);
+
+// This flags control whether Google Test prints the elapsed time for each
+// test.
+GTEST_DECLARE_bool_(print_time);
+
+// This flag specifies the random number seed.
+GTEST_DECLARE_int32_(random_seed);
+
+// This flag sets how many times the tests are repeated. The default value
+// is 1. If the value is -1 the tests are repeating forever.
+GTEST_DECLARE_int32_(repeat);
+
+// This flag controls whether Google Test includes Google Test internal
+// stack frames in failure stack traces.
+GTEST_DECLARE_bool_(show_internal_stack_frames);
+
+// When this flag is specified, tests' order is randomized on every iteration.
+GTEST_DECLARE_bool_(shuffle);
+
+// This flag specifies the maximum number of stack frames to be
+// printed in a failure message.
+GTEST_DECLARE_int32_(stack_trace_depth);
+
+// When this flag is specified, a failed assertion will throw an
+// exception if exceptions are enabled, or exit the program with a
+// non-zero code otherwise.
+GTEST_DECLARE_bool_(throw_on_failure);
+
+// When this flag is set with a "host:port" string, on supported
+// platforms test results are streamed to the specified port on
+// the specified host machine.
+GTEST_DECLARE_string_(stream_result_to);
+
+// The upper limit for valid stack trace depths.
+const int kMaxStackTraceDepth = 100;
+
+namespace internal {
+
+class AssertHelper;
+class DefaultGlobalTestPartResultReporter;
+class ExecDeathTest;
+class NoExecDeathTest;
+class FinalSuccessChecker;
+class GTestFlagSaver;
+class StreamingListenerTest;
+class TestResultAccessor;
+class TestEventListenersAccessor;
+class TestEventRepeater;
+class UnitTestRecordPropertyTestHelper;
+class WindowsDeathTest;
+class UnitTestImpl* GetUnitTestImpl();
+void ReportFailureInUnknownLocation(TestPartResult::Type result_type,
+                                    const std::string& message);
+
+}  // namespace internal
+
+// The friend relationship of some of these classes is cyclic.
+// If we don't forward declare them the compiler might confuse the classes
+// in friendship clauses with same named classes on the scope.
+class Test;
+class TestCase;
+class TestInfo;
+class UnitTest;
+
+// A class for indicating whether an assertion was successful.  When
+// the assertion wasn't successful, the AssertionResult object
+// remembers a non-empty message that describes how it failed.
+//
+// To create an instance of this class, use one of the factory functions
+// (AssertionSuccess() and AssertionFailure()).
+//
+// This class is useful for two purposes:
+//   1. Defining predicate functions to be used with Boolean test assertions
+//      EXPECT_TRUE/EXPECT_FALSE and their ASSERT_ counterparts
+//   2. Defining predicate-format functions to be
+//      used with predicate assertions (ASSERT_PRED_FORMAT*, etc).
+//
+// For example, if you define IsEven predicate:
+//
+//   testing::AssertionResult IsEven(int n) {
+//     if ((n % 2) == 0)
+//       return testing::AssertionSuccess();
+//     else
+//       return testing::AssertionFailure() << n << " is odd";
+//   }
+//
+// Then the failed expectation EXPECT_TRUE(IsEven(Fib(5)))
+// will print the message
+//
+//   Value of: IsEven(Fib(5))
+//     Actual: false (5 is odd)
+//   Expected: true
+//
+// instead of a more opaque
+//
+//   Value of: IsEven(Fib(5))
+//     Actual: false
+//   Expected: true
+//
+// in case IsEven is a simple Boolean predicate.
+//
+// If you expect your predicate to be reused and want to support informative
+// messages in EXPECT_FALSE and ASSERT_FALSE (negative assertions show up
+// about half as often as positive ones in our tests), supply messages for
+// both success and failure cases:
+//
+//   testing::AssertionResult IsEven(int n) {
+//     if ((n % 2) == 0)
+//       return testing::AssertionSuccess() << n << " is even";
+//     else
+//       return testing::AssertionFailure() << n << " is odd";
+//   }
+//
+// Then a statement EXPECT_FALSE(IsEven(Fib(6))) will print
+//
+//   Value of: IsEven(Fib(6))
+//     Actual: true (8 is even)
+//   Expected: false
+//
+// NB: Predicates that support negative Boolean assertions have reduced
+// performance in positive ones so be careful not to use them in tests
+// that have lots (tens of thousands) of positive Boolean assertions.
+//
+// To use this class with EXPECT_PRED_FORMAT assertions such as:
+//
+//   // Verifies that Foo() returns an even number.
+//   EXPECT_PRED_FORMAT1(IsEven, Foo());
+//
+// you need to define:
+//
+//   testing::AssertionResult IsEven(const char* expr, int n) {
+//     if ((n % 2) == 0)
+//       return testing::AssertionSuccess();
+//     else
+//       return testing::AssertionFailure()
+//         << "Expected: " << expr << " is even\n  Actual: it's " << n;
+//   }
+//
+// If Foo() returns 5, you will see the following message:
+//
+//   Expected: Foo() is even
+//     Actual: it's 5
+//
+class GTEST_API_ AssertionResult {
+ public:
+  // Copy constructor.
+  // Used in EXPECT_TRUE/FALSE(assertion_result).
+  AssertionResult(const AssertionResult& other);
+  // Used in the EXPECT_TRUE/FALSE(bool_expression).
+  explicit AssertionResult(bool success) : success_(success) {}
+
+  // Returns true iff the assertion succeeded.
+  operator bool() const { return success_; }  // NOLINT
+
+  // Returns the assertion's negation. Used with EXPECT/ASSERT_FALSE.
+  AssertionResult operator!() const;
+
+  // Returns the text streamed into this AssertionResult. Test assertions
+  // use it when they fail (i.e., the predicate's outcome doesn't match the
+  // assertion's expectation). When nothing has been streamed into the
+  // object, returns an empty string.
+  const char* message() const {
+    return message_.get() != NULL ?  message_->c_str() : "";
+  }
+  // TODO(vladl@google.com): Remove this after making sure no clients use it.
+  // Deprecated; please use message() instead.
+  const char* failure_message() const { return message(); }
+
+  // Streams a custom failure message into this object.
+  template  AssertionResult& operator<<(const T& value) {
+    AppendMessage(Message() << value);
+    return *this;
+  }
+
+  // Allows streaming basic output manipulators such as endl or flush into
+  // this object.
+  AssertionResult& operator<<(
+      ::std::ostream& (*basic_manipulator)(::std::ostream& stream)) {
+    AppendMessage(Message() << basic_manipulator);
+    return *this;
+  }
+
+ private:
+  // Appends the contents of message to message_.
+  void AppendMessage(const Message& a_message) {
+    if (message_.get() == NULL)
+      message_.reset(new ::std::string);
+    message_->append(a_message.GetString().c_str());
+  }
+
+  // Stores result of the assertion predicate.
+  bool success_;
+  // Stores the message describing the condition in case the expectation
+  // construct is not satisfied with the predicate's outcome.
+  // Referenced via a pointer to avoid taking too much stack frame space
+  // with test assertions.
+  internal::scoped_ptr< ::std::string> message_;
+
+  GTEST_DISALLOW_ASSIGN_(AssertionResult);
+};
+
+// Makes a successful assertion result.
+GTEST_API_ AssertionResult AssertionSuccess();
+
+// Makes a failed assertion result.
+GTEST_API_ AssertionResult AssertionFailure();
+
+// Makes a failed assertion result with the given failure message.
+// Deprecated; use AssertionFailure() << msg.
+GTEST_API_ AssertionResult AssertionFailure(const Message& msg);
+
+// The abstract class that all tests inherit from.
+//
+// In Google Test, a unit test program contains one or many TestCases, and
+// each TestCase contains one or many Tests.
+//
+// When you define a test using the TEST macro, you don't need to
+// explicitly derive from Test - the TEST macro automatically does
+// this for you.
+//
+// The only time you derive from Test is when defining a test fixture
+// to be used a TEST_F.  For example:
+//
+//   class FooTest : public testing::Test {
+//    protected:
+//     virtual void SetUp() { ... }
+//     virtual void TearDown() { ... }
+//     ...
+//   };
+//
+//   TEST_F(FooTest, Bar) { ... }
+//   TEST_F(FooTest, Baz) { ... }
+//
+// Test is not copyable.
+class GTEST_API_ Test {
+ public:
+  friend class TestInfo;
+
+  // Defines types for pointers to functions that set up and tear down
+  // a test case.
+  typedef internal::SetUpTestCaseFunc SetUpTestCaseFunc;
+  typedef internal::TearDownTestCaseFunc TearDownTestCaseFunc;
+
+  // The d'tor is virtual as we intend to inherit from Test.
+  virtual ~Test();
+
+  // Sets up the stuff shared by all tests in this test case.
+  //
+  // Google Test will call Foo::SetUpTestCase() before running the first
+  // test in test case Foo.  Hence a sub-class can define its own
+  // SetUpTestCase() method to shadow the one defined in the super
+  // class.
+  static void SetUpTestCase() {}
+
+  // Tears down the stuff shared by all tests in this test case.
+  //
+  // Google Test will call Foo::TearDownTestCase() after running the last
+  // test in test case Foo.  Hence a sub-class can define its own
+  // TearDownTestCase() method to shadow the one defined in the super
+  // class.
+  static void TearDownTestCase() {}
+
+  // Returns true iff the current test has a fatal failure.
+  static bool HasFatalFailure();
+
+  // Returns true iff the current test has a non-fatal failure.
+  static bool HasNonfatalFailure();
+
+  // Returns true iff the current test has a (either fatal or
+  // non-fatal) failure.
+  static bool HasFailure() { return HasFatalFailure() || HasNonfatalFailure(); }
+
+  // Logs a property for the current test, test case, or for the entire
+  // invocation of the test program when used outside of the context of a
+  // test case.  Only the last value for a given key is remembered.  These
+  // are public static so they can be called from utility functions that are
+  // not members of the test fixture.  Calls to RecordProperty made during
+  // lifespan of the test (from the moment its constructor starts to the
+  // moment its destructor finishes) will be output in XML as attributes of
+  // the  element.  Properties recorded from fixture's
+  // SetUpTestCase or TearDownTestCase are logged as attributes of the
+  // corresponding  element.  Calls to RecordProperty made in the
+  // global context (before or after invocation of RUN_ALL_TESTS and from
+  // SetUp/TearDown method of Environment objects registered with Google
+  // Test) will be output as attributes of the  element.
+  static void RecordProperty(const std::string& key, const std::string& value);
+  static void RecordProperty(const std::string& key, int value);
+
+ protected:
+  // Creates a Test object.
+  Test();
+
+  // Sets up the test fixture.
+  virtual void SetUp();
+
+  // Tears down the test fixture.
+  virtual void TearDown();
+
+ private:
+  // Returns true iff the current test has the same fixture class as
+  // the first test in the current test case.
+  static bool HasSameFixtureClass();
+
+  // Runs the test after the test fixture has been set up.
+  //
+  // A sub-class must implement this to define the test logic.
+  //
+  // DO NOT OVERRIDE THIS FUNCTION DIRECTLY IN A USER PROGRAM.
+  // Instead, use the TEST or TEST_F macro.
+  virtual void TestBody() = 0;
+
+  // Sets up, executes, and tears down the test.
+  void Run();
+
+  // Deletes self.  We deliberately pick an unusual name for this
+  // internal method to avoid clashing with names used in user TESTs.
+  void DeleteSelf_() { delete this; }
+
+  // Uses a GTestFlagSaver to save and restore all Google Test flags.
+  const internal::GTestFlagSaver* const gtest_flag_saver_;
+
+  // Often a user mis-spells SetUp() as Setup() and spends a long time
+  // wondering why it is never called by Google Test.  The declaration of
+  // the following method is solely for catching such an error at
+  // compile time:
+  //
+  //   - The return type is deliberately chosen to be not void, so it
+  //   will be a conflict if a user declares void Setup() in his test
+  //   fixture.
+  //
+  //   - This method is private, so it will be another compiler error
+  //   if a user calls it from his test fixture.
+  //
+  // DO NOT OVERRIDE THIS FUNCTION.
+  //
+  // If you see an error about overriding the following function or
+  // about it being private, you have mis-spelled SetUp() as Setup().
+  struct Setup_should_be_spelled_SetUp {};
+  virtual Setup_should_be_spelled_SetUp* Setup() { return NULL; }
+
+  // We disallow copying Tests.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Test);
+};
+
+typedef internal::TimeInMillis TimeInMillis;
+
+// A copyable object representing a user specified test property which can be
+// output as a key/value string pair.
+//
+// Don't inherit from TestProperty as its destructor is not virtual.
+class TestProperty {
+ public:
+  // C'tor.  TestProperty does NOT have a default constructor.
+  // Always use this constructor (with parameters) to create a
+  // TestProperty object.
+  TestProperty(const std::string& a_key, const std::string& a_value) :
+    key_(a_key), value_(a_value) {
+  }
+
+  // Gets the user supplied key.
+  const char* key() const {
+    return key_.c_str();
+  }
+
+  // Gets the user supplied value.
+  const char* value() const {
+    return value_.c_str();
+  }
+
+  // Sets a new value, overriding the one supplied in the constructor.
+  void SetValue(const std::string& new_value) {
+    value_ = new_value;
+  }
+
+ private:
+  // The key supplied by the user.
+  std::string key_;
+  // The value supplied by the user.
+  std::string value_;
+};
+
+// The result of a single Test.  This includes a list of
+// TestPartResults, a list of TestProperties, a count of how many
+// death tests there are in the Test, and how much time it took to run
+// the Test.
+//
+// TestResult is not copyable.
+class GTEST_API_ TestResult {
+ public:
+  // Creates an empty TestResult.
+  TestResult();
+
+  // D'tor.  Do not inherit from TestResult.
+  ~TestResult();
+
+  // Gets the number of all test parts.  This is the sum of the number
+  // of successful test parts and the number of failed test parts.
+  int total_part_count() const;
+
+  // Returns the number of the test properties.
+  int test_property_count() const;
+
+  // Returns true iff the test passed (i.e. no test part failed).
+  bool Passed() const { return !Failed(); }
+
+  // Returns true iff the test failed.
+  bool Failed() const;
+
+  // Returns true iff the test fatally failed.
+  bool HasFatalFailure() const;
+
+  // Returns true iff the test has a non-fatal failure.
+  bool HasNonfatalFailure() const;
+
+  // Returns the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const { return elapsed_time_; }
+
+  // Returns the i-th test part result among all the results. i can range
+  // from 0 to test_property_count() - 1. If i is not in that range, aborts
+  // the program.
+  const TestPartResult& GetTestPartResult(int i) const;
+
+  // Returns the i-th test property. i can range from 0 to
+  // test_property_count() - 1. If i is not in that range, aborts the
+  // program.
+  const TestProperty& GetTestProperty(int i) const;
+
+ private:
+  friend class TestInfo;
+  friend class TestCase;
+  friend class UnitTest;
+  friend class internal::DefaultGlobalTestPartResultReporter;
+  friend class internal::ExecDeathTest;
+  friend class internal::TestResultAccessor;
+  friend class internal::UnitTestImpl;
+  friend class internal::WindowsDeathTest;
+
+  // Gets the vector of TestPartResults.
+  const std::vector& test_part_results() const {
+    return test_part_results_;
+  }
+
+  // Gets the vector of TestProperties.
+  const std::vector& test_properties() const {
+    return test_properties_;
+  }
+
+  // Sets the elapsed time.
+  void set_elapsed_time(TimeInMillis elapsed) { elapsed_time_ = elapsed; }
+
+  // Adds a test property to the list. The property is validated and may add
+  // a non-fatal failure if invalid (e.g., if it conflicts with reserved
+  // key names). If a property is already recorded for the same key, the
+  // value will be updated, rather than storing multiple values for the same
+  // key.  xml_element specifies the element for which the property is being
+  // recorded and is used for validation.
+  void RecordProperty(const std::string& xml_element,
+                      const TestProperty& test_property);
+
+  // Adds a failure if the key is a reserved attribute of Google Test
+  // testcase tags.  Returns true if the property is valid.
+  // TODO(russr): Validate attribute names are legal and human readable.
+  static bool ValidateTestProperty(const std::string& xml_element,
+                                   const TestProperty& test_property);
+
+  // Adds a test part result to the list.
+  void AddTestPartResult(const TestPartResult& test_part_result);
+
+  // Returns the death test count.
+  int death_test_count() const { return death_test_count_; }
+
+  // Increments the death test count, returning the new count.
+  int increment_death_test_count() { return ++death_test_count_; }
+
+  // Clears the test part results.
+  void ClearTestPartResults();
+
+  // Clears the object.
+  void Clear();
+
+  // Protects mutable state of the property vector and of owned
+  // properties, whose values may be updated.
+  internal::Mutex test_properites_mutex_;
+
+  // The vector of TestPartResults
+  std::vector test_part_results_;
+  // The vector of TestProperties
+  std::vector test_properties_;
+  // Running count of death tests.
+  int death_test_count_;
+  // The elapsed time, in milliseconds.
+  TimeInMillis elapsed_time_;
+
+  // We disallow copying TestResult.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestResult);
+};  // class TestResult
+
+// A TestInfo object stores the following information about a test:
+//
+//   Test case name
+//   Test name
+//   Whether the test should be run
+//   A function pointer that creates the test object when invoked
+//   Test result
+//
+// The constructor of TestInfo registers itself with the UnitTest
+// singleton such that the RUN_ALL_TESTS() macro knows which tests to
+// run.
+class GTEST_API_ TestInfo {
+ public:
+  // Destructs a TestInfo object.  This function is not virtual, so
+  // don't inherit from TestInfo.
+  ~TestInfo();
+
+  // Returns the test case name.
+  const char* test_case_name() const { return test_case_name_.c_str(); }
+
+  // Returns the test name.
+  const char* name() const { return name_.c_str(); }
+
+  // Returns the name of the parameter type, or NULL if this is not a typed
+  // or a type-parameterized test.
+  const char* type_param() const {
+    if (type_param_.get() != NULL)
+      return type_param_->c_str();
+    return NULL;
+  }
+
+  // Returns the text representation of the value parameter, or NULL if this
+  // is not a value-parameterized test.
+  const char* value_param() const {
+    if (value_param_.get() != NULL)
+      return value_param_->c_str();
+    return NULL;
+  }
+
+  // Returns true if this test should run, that is if the test is not
+  // disabled (or it is disabled but the also_run_disabled_tests flag has
+  // been specified) and its full name matches the user-specified filter.
+  //
+  // Google Test allows the user to filter the tests by their full names.
+  // The full name of a test Bar in test case Foo is defined as
+  // "Foo.Bar".  Only the tests that match the filter will run.
+  //
+  // A filter is a colon-separated list of glob (not regex) patterns,
+  // optionally followed by a '-' and a colon-separated list of
+  // negative patterns (tests to exclude).  A test is run if it
+  // matches one of the positive patterns and does not match any of
+  // the negative patterns.
+  //
+  // For example, *A*:Foo.* is a filter that matches any string that
+  // contains the character 'A' or starts with "Foo.".
+  bool should_run() const { return should_run_; }
+
+  // Returns true iff this test will appear in the XML report.
+  bool is_reportable() const {
+    // For now, the XML report includes all tests matching the filter.
+    // In the future, we may trim tests that are excluded because of
+    // sharding.
+    return matches_filter_;
+  }
+
+  // Returns the result of the test.
+  const TestResult* result() const { return &result_; }
+
+ private:
+#if GTEST_HAS_DEATH_TEST
+  friend class internal::DefaultDeathTestFactory;
+#endif  // GTEST_HAS_DEATH_TEST
+  friend class Test;
+  friend class TestCase;
+  friend class internal::UnitTestImpl;
+  friend class internal::StreamingListenerTest;
+  friend TestInfo* internal::MakeAndRegisterTestInfo(
+      const char* test_case_name,
+      const char* name,
+      const char* type_param,
+      const char* value_param,
+      internal::TypeId fixture_class_id,
+      Test::SetUpTestCaseFunc set_up_tc,
+      Test::TearDownTestCaseFunc tear_down_tc,
+      internal::TestFactoryBase* factory);
+
+  // Constructs a TestInfo object. The newly constructed instance assumes
+  // ownership of the factory object.
+  TestInfo(const std::string& test_case_name,
+           const std::string& name,
+           const char* a_type_param,   // NULL if not a type-parameterized test
+           const char* a_value_param,  // NULL if not a value-parameterized test
+           internal::TypeId fixture_class_id,
+           internal::TestFactoryBase* factory);
+
+  // Increments the number of death tests encountered in this test so
+  // far.
+  int increment_death_test_count() {
+    return result_.increment_death_test_count();
+  }
+
+  // Creates the test object, runs it, records its result, and then
+  // deletes it.
+  void Run();
+
+  static void ClearTestResult(TestInfo* test_info) {
+    test_info->result_.Clear();
+  }
+
+  // These fields are immutable properties of the test.
+  const std::string test_case_name_;     // Test case name
+  const std::string name_;               // Test name
+  // Name of the parameter type, or NULL if this is not a typed or a
+  // type-parameterized test.
+  const internal::scoped_ptr type_param_;
+  // Text representation of the value parameter, or NULL if this is not a
+  // value-parameterized test.
+  const internal::scoped_ptr value_param_;
+  const internal::TypeId fixture_class_id_;   // ID of the test fixture class
+  bool should_run_;                 // True iff this test should run
+  bool is_disabled_;                // True iff this test is disabled
+  bool matches_filter_;             // True if this test matches the
+                                    // user-specified filter.
+  internal::TestFactoryBase* const factory_;  // The factory that creates
+                                              // the test object
+
+  // This field is mutable and needs to be reset before running the
+  // test for the second time.
+  TestResult result_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestInfo);
+};
+
+// A test case, which consists of a vector of TestInfos.
+//
+// TestCase is not copyable.
+class GTEST_API_ TestCase {
+ public:
+  // Creates a TestCase with the given name.
+  //
+  // TestCase does NOT have a default constructor.  Always use this
+  // constructor to create a TestCase object.
+  //
+  // Arguments:
+  //
+  //   name:         name of the test case
+  //   a_type_param: the name of the test's type parameter, or NULL if
+  //                 this is not a type-parameterized test.
+  //   set_up_tc:    pointer to the function that sets up the test case
+  //   tear_down_tc: pointer to the function that tears down the test case
+  TestCase(const char* name, const char* a_type_param,
+           Test::SetUpTestCaseFunc set_up_tc,
+           Test::TearDownTestCaseFunc tear_down_tc);
+
+  // Destructor of TestCase.
+  virtual ~TestCase();
+
+  // Gets the name of the TestCase.
+  const char* name() const { return name_.c_str(); }
+
+  // Returns the name of the parameter type, or NULL if this is not a
+  // type-parameterized test case.
+  const char* type_param() const {
+    if (type_param_.get() != NULL)
+      return type_param_->c_str();
+    return NULL;
+  }
+
+  // Returns true if any test in this test case should run.
+  bool should_run() const { return should_run_; }
+
+  // Gets the number of successful tests in this test case.
+  int successful_test_count() const;
+
+  // Gets the number of failed tests in this test case.
+  int failed_test_count() const;
+
+  // Gets the number of disabled tests that will be reported in the XML report.
+  int reportable_disabled_test_count() const;
+
+  // Gets the number of disabled tests in this test case.
+  int disabled_test_count() const;
+
+  // Gets the number of tests to be printed in the XML report.
+  int reportable_test_count() const;
+
+  // Get the number of tests in this test case that should run.
+  int test_to_run_count() const;
+
+  // Gets the number of all tests in this test case.
+  int total_test_count() const;
+
+  // Returns true iff the test case passed.
+  bool Passed() const { return !Failed(); }
+
+  // Returns true iff the test case failed.
+  bool Failed() const { return failed_test_count() > 0; }
+
+  // Returns the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const { return elapsed_time_; }
+
+  // Returns the i-th test among all the tests. i can range from 0 to
+  // total_test_count() - 1. If i is not in that range, returns NULL.
+  const TestInfo* GetTestInfo(int i) const;
+
+  // Returns the TestResult that holds test properties recorded during
+  // execution of SetUpTestCase and TearDownTestCase.
+  const TestResult& ad_hoc_test_result() const { return ad_hoc_test_result_; }
+
+ private:
+  friend class Test;
+  friend class internal::UnitTestImpl;
+
+  // Gets the (mutable) vector of TestInfos in this TestCase.
+  std::vector& test_info_list() { return test_info_list_; }
+
+  // Gets the (immutable) vector of TestInfos in this TestCase.
+  const std::vector& test_info_list() const {
+    return test_info_list_;
+  }
+
+  // Returns the i-th test among all the tests. i can range from 0 to
+  // total_test_count() - 1. If i is not in that range, returns NULL.
+  TestInfo* GetMutableTestInfo(int i);
+
+  // Sets the should_run member.
+  void set_should_run(bool should) { should_run_ = should; }
+
+  // Adds a TestInfo to this test case.  Will delete the TestInfo upon
+  // destruction of the TestCase object.
+  void AddTestInfo(TestInfo * test_info);
+
+  // Clears the results of all tests in this test case.
+  void ClearResult();
+
+  // Clears the results of all tests in the given test case.
+  static void ClearTestCaseResult(TestCase* test_case) {
+    test_case->ClearResult();
+  }
+
+  // Runs every test in this TestCase.
+  void Run();
+
+  // Runs SetUpTestCase() for this TestCase.  This wrapper is needed
+  // for catching exceptions thrown from SetUpTestCase().
+  void RunSetUpTestCase() { (*set_up_tc_)(); }
+
+  // Runs TearDownTestCase() for this TestCase.  This wrapper is
+  // needed for catching exceptions thrown from TearDownTestCase().
+  void RunTearDownTestCase() { (*tear_down_tc_)(); }
+
+  // Returns true iff test passed.
+  static bool TestPassed(const TestInfo* test_info) {
+    return test_info->should_run() && test_info->result()->Passed();
+  }
+
+  // Returns true iff test failed.
+  static bool TestFailed(const TestInfo* test_info) {
+    return test_info->should_run() && test_info->result()->Failed();
+  }
+
+  // Returns true iff the test is disabled and will be reported in the XML
+  // report.
+  static bool TestReportableDisabled(const TestInfo* test_info) {
+    return test_info->is_reportable() && test_info->is_disabled_;
+  }
+
+  // Returns true iff test is disabled.
+  static bool TestDisabled(const TestInfo* test_info) {
+    return test_info->is_disabled_;
+  }
+
+  // Returns true iff this test will appear in the XML report.
+  static bool TestReportable(const TestInfo* test_info) {
+    return test_info->is_reportable();
+  }
+
+  // Returns true if the given test should run.
+  static bool ShouldRunTest(const TestInfo* test_info) {
+    return test_info->should_run();
+  }
+
+  // Shuffles the tests in this test case.
+  void ShuffleTests(internal::Random* random);
+
+  // Restores the test order to before the first shuffle.
+  void UnshuffleTests();
+
+  // Name of the test case.
+  std::string name_;
+  // Name of the parameter type, or NULL if this is not a typed or a
+  // type-parameterized test.
+  const internal::scoped_ptr type_param_;
+  // The vector of TestInfos in their original order.  It owns the
+  // elements in the vector.
+  std::vector test_info_list_;
+  // Provides a level of indirection for the test list to allow easy
+  // shuffling and restoring the test order.  The i-th element in this
+  // vector is the index of the i-th test in the shuffled test list.
+  std::vector test_indices_;
+  // Pointer to the function that sets up the test case.
+  Test::SetUpTestCaseFunc set_up_tc_;
+  // Pointer to the function that tears down the test case.
+  Test::TearDownTestCaseFunc tear_down_tc_;
+  // True iff any test in this test case should run.
+  bool should_run_;
+  // Elapsed time, in milliseconds.
+  TimeInMillis elapsed_time_;
+  // Holds test properties recorded during execution of SetUpTestCase and
+  // TearDownTestCase.
+  TestResult ad_hoc_test_result_;
+
+  // We disallow copying TestCases.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestCase);
+};
+
+// An Environment object is capable of setting up and tearing down an
+// environment.  The user should subclass this to define his own
+// environment(s).
+//
+// An Environment object does the set-up and tear-down in virtual
+// methods SetUp() and TearDown() instead of the constructor and the
+// destructor, as:
+//
+//   1. You cannot safely throw from a destructor.  This is a problem
+//      as in some cases Google Test is used where exceptions are enabled, and
+//      we may want to implement ASSERT_* using exceptions where they are
+//      available.
+//   2. You cannot use ASSERT_* directly in a constructor or
+//      destructor.
+class Environment {
+ public:
+  // The d'tor is virtual as we need to subclass Environment.
+  virtual ~Environment() {}
+
+  // Override this to define how to set up the environment.
+  virtual void SetUp() {}
+
+  // Override this to define how to tear down the environment.
+  virtual void TearDown() {}
+ private:
+  // If you see an error about overriding the following function or
+  // about it being private, you have mis-spelled SetUp() as Setup().
+  struct Setup_should_be_spelled_SetUp {};
+  virtual Setup_should_be_spelled_SetUp* Setup() { return NULL; }
+};
+
+// The interface for tracing execution of tests. The methods are organized in
+// the order the corresponding events are fired.
+class TestEventListener {
+ public:
+  virtual ~TestEventListener() {}
+
+  // Fired before any test activity starts.
+  virtual void OnTestProgramStart(const UnitTest& unit_test) = 0;
+
+  // Fired before each iteration of tests starts.  There may be more than
+  // one iteration if GTEST_FLAG(repeat) is set. iteration is the iteration
+  // index, starting from 0.
+  virtual void OnTestIterationStart(const UnitTest& unit_test,
+                                    int iteration) = 0;
+
+  // Fired before environment set-up for each iteration of tests starts.
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& unit_test) = 0;
+
+  // Fired after environment set-up for each iteration of tests ends.
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& unit_test) = 0;
+
+  // Fired before the test case starts.
+  virtual void OnTestCaseStart(const TestCase& test_case) = 0;
+
+  // Fired before the test starts.
+  virtual void OnTestStart(const TestInfo& test_info) = 0;
+
+  // Fired after a failed assertion or a SUCCEED() invocation.
+  virtual void OnTestPartResult(const TestPartResult& test_part_result) = 0;
+
+  // Fired after the test ends.
+  virtual void OnTestEnd(const TestInfo& test_info) = 0;
+
+  // Fired after the test case ends.
+  virtual void OnTestCaseEnd(const TestCase& test_case) = 0;
+
+  // Fired before environment tear-down for each iteration of tests starts.
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& unit_test) = 0;
+
+  // Fired after environment tear-down for each iteration of tests ends.
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& unit_test) = 0;
+
+  // Fired after each iteration of tests finishes.
+  virtual void OnTestIterationEnd(const UnitTest& unit_test,
+                                  int iteration) = 0;
+
+  // Fired after all test activities have ended.
+  virtual void OnTestProgramEnd(const UnitTest& unit_test) = 0;
+};
+
+// The convenience class for users who need to override just one or two
+// methods and are not concerned that a possible change to a signature of
+// the methods they override will not be caught during the build.  For
+// comments about each method please see the definition of TestEventListener
+// above.
+class EmptyTestEventListener : public TestEventListener {
+ public:
+  virtual void OnTestProgramStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationStart(const UnitTest& /*unit_test*/,
+                                    int /*iteration*/) {}
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestCaseStart(const TestCase& /*test_case*/) {}
+  virtual void OnTestStart(const TestInfo& /*test_info*/) {}
+  virtual void OnTestPartResult(const TestPartResult& /*test_part_result*/) {}
+  virtual void OnTestEnd(const TestInfo& /*test_info*/) {}
+  virtual void OnTestCaseEnd(const TestCase& /*test_case*/) {}
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationEnd(const UnitTest& /*unit_test*/,
+                                  int /*iteration*/) {}
+  virtual void OnTestProgramEnd(const UnitTest& /*unit_test*/) {}
+};
+
+// TestEventListeners lets users add listeners to track events in Google Test.
+class GTEST_API_ TestEventListeners {
+ public:
+  TestEventListeners();
+  ~TestEventListeners();
+
+  // Appends an event listener to the end of the list. Google Test assumes
+  // the ownership of the listener (i.e. it will delete the listener when
+  // the test program finishes).
+  void Append(TestEventListener* listener);
+
+  // Removes the given event listener from the list and returns it.  It then
+  // becomes the caller's responsibility to delete the listener. Returns
+  // NULL if the listener is not found in the list.
+  TestEventListener* Release(TestEventListener* listener);
+
+  // Returns the standard listener responsible for the default console
+  // output.  Can be removed from the listeners list to shut down default
+  // console output.  Note that removing this object from the listener list
+  // with Release transfers its ownership to the caller and makes this
+  // function return NULL the next time.
+  TestEventListener* default_result_printer() const {
+    return default_result_printer_;
+  }
+
+  // Returns the standard listener responsible for the default XML output
+  // controlled by the --gtest_output=xml flag.  Can be removed from the
+  // listeners list by users who want to shut down the default XML output
+  // controlled by this flag and substitute it with custom one.  Note that
+  // removing this object from the listener list with Release transfers its
+  // ownership to the caller and makes this function return NULL the next
+  // time.
+  TestEventListener* default_xml_generator() const {
+    return default_xml_generator_;
+  }
+
+ private:
+  friend class TestCase;
+  friend class TestInfo;
+  friend class internal::DefaultGlobalTestPartResultReporter;
+  friend class internal::NoExecDeathTest;
+  friend class internal::TestEventListenersAccessor;
+  friend class internal::UnitTestImpl;
+
+  // Returns repeater that broadcasts the TestEventListener events to all
+  // subscribers.
+  TestEventListener* repeater();
+
+  // Sets the default_result_printer attribute to the provided listener.
+  // The listener is also added to the listener list and previous
+  // default_result_printer is removed from it and deleted. The listener can
+  // also be NULL in which case it will not be added to the list. Does
+  // nothing if the previous and the current listener objects are the same.
+  void SetDefaultResultPrinter(TestEventListener* listener);
+
+  // Sets the default_xml_generator attribute to the provided listener.  The
+  // listener is also added to the listener list and previous
+  // default_xml_generator is removed from it and deleted. The listener can
+  // also be NULL in which case it will not be added to the list. Does
+  // nothing if the previous and the current listener objects are the same.
+  void SetDefaultXmlGenerator(TestEventListener* listener);
+
+  // Controls whether events will be forwarded by the repeater to the
+  // listeners in the list.
+  bool EventForwardingEnabled() const;
+  void SuppressEventForwarding();
+
+  // The actual list of listeners.
+  internal::TestEventRepeater* repeater_;
+  // Listener responsible for the standard result output.
+  TestEventListener* default_result_printer_;
+  // Listener responsible for the creation of the XML output file.
+  TestEventListener* default_xml_generator_;
+
+  // We disallow copying TestEventListeners.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestEventListeners);
+};
+
+// A UnitTest consists of a vector of TestCases.
+//
+// This is a singleton class.  The only instance of UnitTest is
+// created when UnitTest::GetInstance() is first called.  This
+// instance is never deleted.
+//
+// UnitTest is not copyable.
+//
+// This class is thread-safe as long as the methods are called
+// according to their specification.
+class GTEST_API_ UnitTest {
+ public:
+  // Gets the singleton UnitTest object.  The first time this method
+  // is called, a UnitTest object is constructed and returned.
+  // Consecutive calls will return the same object.
+  static UnitTest* GetInstance();
+
+  // Runs all tests in this UnitTest object and prints the result.
+  // Returns 0 if successful, or 1 otherwise.
+  //
+  // This method can only be called from the main thread.
+  //
+  // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+  int Run() GTEST_MUST_USE_RESULT_;
+
+  // Returns the working directory when the first TEST() or TEST_F()
+  // was executed.  The UnitTest object owns the string.
+  const char* original_working_dir() const;
+
+  // Returns the TestCase object for the test that's currently running,
+  // or NULL if no test is running.
+  const TestCase* current_test_case() const
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Returns the TestInfo object for the test that's currently running,
+  // or NULL if no test is running.
+  const TestInfo* current_test_info() const
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Returns the random seed used at the start of the current test run.
+  int random_seed() const;
+
+#if GTEST_HAS_PARAM_TEST
+  // Returns the ParameterizedTestCaseRegistry object used to keep track of
+  // value-parameterized tests and instantiate and register them.
+  //
+  // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+  internal::ParameterizedTestCaseRegistry& parameterized_test_registry()
+      GTEST_LOCK_EXCLUDED_(mutex_);
+#endif  // GTEST_HAS_PARAM_TEST
+
+  // Gets the number of successful test cases.
+  int successful_test_case_count() const;
+
+  // Gets the number of failed test cases.
+  int failed_test_case_count() const;
+
+  // Gets the number of all test cases.
+  int total_test_case_count() const;
+
+  // Gets the number of all test cases that contain at least one test
+  // that should run.
+  int test_case_to_run_count() const;
+
+  // Gets the number of successful tests.
+  int successful_test_count() const;
+
+  // Gets the number of failed tests.
+  int failed_test_count() const;
+
+  // Gets the number of disabled tests that will be reported in the XML report.
+  int reportable_disabled_test_count() const;
+
+  // Gets the number of disabled tests.
+  int disabled_test_count() const;
+
+  // Gets the number of tests to be printed in the XML report.
+  int reportable_test_count() const;
+
+  // Gets the number of all tests.
+  int total_test_count() const;
+
+  // Gets the number of tests that should run.
+  int test_to_run_count() const;
+
+  // Gets the time of the test program start, in ms from the start of the
+  // UNIX epoch.
+  TimeInMillis start_timestamp() const;
+
+  // Gets the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const;
+
+  // Returns true iff the unit test passed (i.e. all test cases passed).
+  bool Passed() const;
+
+  // Returns true iff the unit test failed (i.e. some test case failed
+  // or something outside of all tests failed).
+  bool Failed() const;
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  const TestCase* GetTestCase(int i) const;
+
+  // Returns the TestResult containing information on test failures and
+  // properties logged outside of individual test cases.
+  const TestResult& ad_hoc_test_result() const;
+
+  // Returns the list of event listeners that can be used to track events
+  // inside Google Test.
+  TestEventListeners& listeners();
+
+ private:
+  // Registers and returns a global test environment.  When a test
+  // program is run, all global test environments will be set-up in
+  // the order they were registered.  After all tests in the program
+  // have finished, all global test environments will be torn-down in
+  // the *reverse* order they were registered.
+  //
+  // The UnitTest object takes ownership of the given environment.
+  //
+  // This method can only be called from the main thread.
+  Environment* AddEnvironment(Environment* env);
+
+  // Adds a TestPartResult to the current TestResult object.  All
+  // Google Test assertion macros (e.g. ASSERT_TRUE, EXPECT_EQ, etc)
+  // eventually call this to report their results.  The user code
+  // should use the assertion macros instead of calling this directly.
+  void AddTestPartResult(TestPartResult::Type result_type,
+                         const char* file_name,
+                         int line_number,
+                         const std::string& message,
+                         const std::string& os_stack_trace)
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Adds a TestProperty to the current TestResult object when invoked from
+  // inside a test, to current TestCase's ad_hoc_test_result_ when invoked
+  // from SetUpTestCase or TearDownTestCase, or to the global property set
+  // when invoked elsewhere.  If the result already contains a property with
+  // the same key, the value will be updated.
+  void RecordProperty(const std::string& key, const std::string& value);
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  TestCase* GetMutableTestCase(int i);
+
+  // Accessors for the implementation object.
+  internal::UnitTestImpl* impl() { return impl_; }
+  const internal::UnitTestImpl* impl() const { return impl_; }
+
+  // These classes and funcions are friends as they need to access private
+  // members of UnitTest.
+  friend class Test;
+  friend class internal::AssertHelper;
+  friend class internal::ScopedTrace;
+  friend class internal::StreamingListenerTest;
+  friend class internal::UnitTestRecordPropertyTestHelper;
+  friend Environment* AddGlobalTestEnvironment(Environment* env);
+  friend internal::UnitTestImpl* internal::GetUnitTestImpl();
+  friend void internal::ReportFailureInUnknownLocation(
+      TestPartResult::Type result_type,
+      const std::string& message);
+
+  // Creates an empty UnitTest.
+  UnitTest();
+
+  // D'tor
+  virtual ~UnitTest();
+
+  // Pushes a trace defined by SCOPED_TRACE() on to the per-thread
+  // Google Test trace stack.
+  void PushGTestTrace(const internal::TraceInfo& trace)
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Pops a trace from the per-thread Google Test trace stack.
+  void PopGTestTrace()
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Protects mutable state in *impl_.  This is mutable as some const
+  // methods need to lock it too.
+  mutable internal::Mutex mutex_;
+
+  // Opaque implementation object.  This field is never changed once
+  // the object is constructed.  We don't mark it as const here, as
+  // doing so will cause a warning in the constructor of UnitTest.
+  // Mutable state in *impl_ is protected by mutex_.
+  internal::UnitTestImpl* impl_;
+
+  // We disallow copying UnitTest.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(UnitTest);
+};
+
+// A convenient wrapper for adding an environment for the test
+// program.
+//
+// You should call this before RUN_ALL_TESTS() is called, probably in
+// main().  If you use gtest_main, you need to call this before main()
+// starts for it to take effect.  For example, you can define a global
+// variable like this:
+//
+//   testing::Environment* const foo_env =
+//       testing::AddGlobalTestEnvironment(new FooEnvironment);
+//
+// However, we strongly recommend you to write your own main() and
+// call AddGlobalTestEnvironment() there, as relying on initialization
+// of global variables makes the code harder to read and may cause
+// problems when you register multiple environments from different
+// translation units and the environments have dependencies among them
+// (remember that the compiler doesn't guarantee the order in which
+// global variables from different translation units are initialized).
+inline Environment* AddGlobalTestEnvironment(Environment* env) {
+  return UnitTest::GetInstance()->AddEnvironment(env);
+}
+
+// Initializes Google Test.  This must be called before calling
+// RUN_ALL_TESTS().  In particular, it parses a command line for the
+// flags that Google Test recognizes.  Whenever a Google Test flag is
+// seen, it is removed from argv, and *argc is decremented.
+//
+// No value is returned.  Instead, the Google Test flag variables are
+// updated.
+//
+// Calling the function for the second time has no user-visible effect.
+GTEST_API_ void InitGoogleTest(int* argc, char** argv);
+
+// This overloaded version can be used in Windows programs compiled in
+// UNICODE mode.
+GTEST_API_ void InitGoogleTest(int* argc, wchar_t** argv);
+
+namespace internal {
+
+// FormatForComparison::Format(value) formats a
+// value of type ToPrint that is an operand of a comparison assertion
+// (e.g. ASSERT_EQ).  OtherOperand is the type of the other operand in
+// the comparison, and is used to help determine the best way to
+// format the value.  In particular, when the value is a C string
+// (char pointer) and the other operand is an STL string object, we
+// want to format the C string as a string, since we know it is
+// compared by value with the string object.  If the value is a char
+// pointer but the other operand is not an STL string object, we don't
+// know whether the pointer is supposed to point to a NUL-terminated
+// string, and thus want to print it as a pointer to be safe.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+
+// The default case.
+template 
+class FormatForComparison {
+ public:
+  static ::std::string Format(const ToPrint& value) {
+    return ::testing::PrintToString(value);
+  }
+};
+
+// Array.
+template 
+class FormatForComparison {
+ public:
+  static ::std::string Format(const ToPrint* value) {
+    return FormatForComparison::Format(value);
+  }
+};
+
+// By default, print C string as pointers to be safe, as we don't know
+// whether they actually point to a NUL-terminated string.
+
+#define GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(CharType)                \
+  template                                       \
+  class FormatForComparison {                  \
+   public:                                                              \
+    static ::std::string Format(CharType* value) {                      \
+      return ::testing::PrintToString(static_cast(value)); \
+    }                                                                   \
+  }
+
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(char);
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(const char);
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(wchar_t);
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(const wchar_t);
+
+#undef GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_
+
+// If a C string is compared with an STL string object, we know it's meant
+// to point to a NUL-terminated string, and thus can print it as a string.
+
+#define GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(CharType, OtherStringType) \
+  template <>                                                           \
+  class FormatForComparison {               \
+   public:                                                              \
+    static ::std::string Format(CharType* value) {                      \
+      return ::testing::PrintToString(value);                           \
+    }                                                                   \
+  }
+
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(char, ::std::string);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const char, ::std::string);
+
+#if GTEST_HAS_GLOBAL_STRING
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(char, ::string);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const char, ::string);
+#endif
+
+#if GTEST_HAS_GLOBAL_WSTRING
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(wchar_t, ::wstring);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const wchar_t, ::wstring);
+#endif
+
+#if GTEST_HAS_STD_WSTRING
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(wchar_t, ::std::wstring);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const wchar_t, ::std::wstring);
+#endif
+
+#undef GTEST_IMPL_FORMAT_C_STRING_AS_STRING_
+
+// Formats a comparison assertion (e.g. ASSERT_EQ, EXPECT_LT, and etc)
+// operand to be used in a failure message.  The type (but not value)
+// of the other operand may affect the format.  This allows us to
+// print a char* as a raw pointer when it is compared against another
+// char* or void*, and print it as a C string when it is compared
+// against an std::string object, for example.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+template 
+std::string FormatForComparisonFailureMessage(
+    const T1& value, const T2& /* other_operand */) {
+  return FormatForComparison::Format(value);
+}
+
+// The helper function for {ASSERT|EXPECT}_EQ.
+template 
+AssertionResult CmpHelperEQ(const char* expected_expression,
+                            const char* actual_expression,
+                            const T1& expected,
+                            const T2& actual) {
+#ifdef _MSC_VER
+# pragma warning(push)          // Saves the current warning state.
+# pragma warning(disable:4389)  // Temporarily disables warning on
+                                // signed/unsigned mismatch.
+#endif
+
+  if (expected == actual) {
+    return AssertionSuccess();
+  }
+
+#ifdef _MSC_VER
+# pragma warning(pop)          // Restores the warning state.
+#endif
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   FormatForComparisonFailureMessage(expected, actual),
+                   FormatForComparisonFailureMessage(actual, expected),
+                   false);
+}
+
+// With this overloaded version, we allow anonymous enums to be used
+// in {ASSERT|EXPECT}_EQ when compiled with gcc 4, as anonymous enums
+// can be implicitly cast to BiggestInt.
+GTEST_API_ AssertionResult CmpHelperEQ(const char* expected_expression,
+                                       const char* actual_expression,
+                                       BiggestInt expected,
+                                       BiggestInt actual);
+
+// The helper class for {ASSERT|EXPECT}_EQ.  The template argument
+// lhs_is_null_literal is true iff the first argument to ASSERT_EQ()
+// is a null pointer literal.  The following default implementation is
+// for lhs_is_null_literal being false.
+template 
+class EqHelper {
+ public:
+  // This templatized version is for the general case.
+  template 
+  static AssertionResult Compare(const char* expected_expression,
+                                 const char* actual_expression,
+                                 const T1& expected,
+                                 const T2& actual) {
+    return CmpHelperEQ(expected_expression, actual_expression, expected,
+                       actual);
+  }
+
+  // With this overloaded version, we allow anonymous enums to be used
+  // in {ASSERT|EXPECT}_EQ when compiled with gcc 4, as anonymous
+  // enums can be implicitly cast to BiggestInt.
+  //
+  // Even though its body looks the same as the above version, we
+  // cannot merge the two, as it will make anonymous enums unhappy.
+  static AssertionResult Compare(const char* expected_expression,
+                                 const char* actual_expression,
+                                 BiggestInt expected,
+                                 BiggestInt actual) {
+    return CmpHelperEQ(expected_expression, actual_expression, expected,
+                       actual);
+  }
+};
+
+// This specialization is used when the first argument to ASSERT_EQ()
+// is a null pointer literal, like NULL, false, or 0.
+template <>
+class EqHelper {
+ public:
+  // We define two overloaded versions of Compare().  The first
+  // version will be picked when the second argument to ASSERT_EQ() is
+  // NOT a pointer, e.g. ASSERT_EQ(0, AnIntFunction()) or
+  // EXPECT_EQ(false, a_bool).
+  template 
+  static AssertionResult Compare(
+      const char* expected_expression,
+      const char* actual_expression,
+      const T1& expected,
+      const T2& actual,
+      // The following line prevents this overload from being considered if T2
+      // is not a pointer type.  We need this because ASSERT_EQ(NULL, my_ptr)
+      // expands to Compare("", "", NULL, my_ptr), which requires a conversion
+      // to match the Secret* in the other overload, which would otherwise make
+      // this template match better.
+      typename EnableIf::value>::type* = 0) {
+    return CmpHelperEQ(expected_expression, actual_expression, expected,
+                       actual);
+  }
+
+  // This version will be picked when the second argument to ASSERT_EQ() is a
+  // pointer, e.g. ASSERT_EQ(NULL, a_pointer).
+  template 
+  static AssertionResult Compare(
+      const char* expected_expression,
+      const char* actual_expression,
+      // We used to have a second template parameter instead of Secret*.  That
+      // template parameter would deduce to 'long', making this a better match
+      // than the first overload even without the first overload's EnableIf.
+      // Unfortunately, gcc with -Wconversion-null warns when "passing NULL to
+      // non-pointer argument" (even a deduced integral argument), so the old
+      // implementation caused warnings in user code.
+      Secret* /* expected (NULL) */,
+      T* actual) {
+    // We already know that 'expected' is a null pointer.
+    return CmpHelperEQ(expected_expression, actual_expression,
+                       static_cast(NULL), actual);
+  }
+};
+
+// A macro for implementing the helper functions needed to implement
+// ASSERT_?? and EXPECT_??.  It is here just to avoid copy-and-paste
+// of similar code.
+//
+// For each templatized helper function, we also define an overloaded
+// version for BiggestInt in order to reduce code bloat and allow
+// anonymous enums to be used with {ASSERT|EXPECT}_?? when compiled
+// with gcc 4.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+#define GTEST_IMPL_CMP_HELPER_(op_name, op)\
+template \
+AssertionResult CmpHelper##op_name(const char* expr1, const char* expr2, \
+                                   const T1& val1, const T2& val2) {\
+  if (val1 op val2) {\
+    return AssertionSuccess();\
+  } else {\
+    return AssertionFailure() \
+        << "Expected: (" << expr1 << ") " #op " (" << expr2\
+        << "), actual: " << FormatForComparisonFailureMessage(val1, val2)\
+        << " vs " << FormatForComparisonFailureMessage(val2, val1);\
+  }\
+}\
+GTEST_API_ AssertionResult CmpHelper##op_name(\
+    const char* expr1, const char* expr2, BiggestInt val1, BiggestInt val2)
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+
+// Implements the helper function for {ASSERT|EXPECT}_NE
+GTEST_IMPL_CMP_HELPER_(NE, !=);
+// Implements the helper function for {ASSERT|EXPECT}_LE
+GTEST_IMPL_CMP_HELPER_(LE, <=);
+// Implements the helper function for {ASSERT|EXPECT}_LT
+GTEST_IMPL_CMP_HELPER_(LT, <);
+// Implements the helper function for {ASSERT|EXPECT}_GE
+GTEST_IMPL_CMP_HELPER_(GE, >=);
+// Implements the helper function for {ASSERT|EXPECT}_GT
+GTEST_IMPL_CMP_HELPER_(GT, >);
+
+#undef GTEST_IMPL_CMP_HELPER_
+
+// The helper function for {ASSERT|EXPECT}_STREQ.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTREQ(const char* expected_expression,
+                                          const char* actual_expression,
+                                          const char* expected,
+                                          const char* actual);
+
+// The helper function for {ASSERT|EXPECT}_STRCASEEQ.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRCASEEQ(const char* expected_expression,
+                                              const char* actual_expression,
+                                              const char* expected,
+                                              const char* actual);
+
+// The helper function for {ASSERT|EXPECT}_STRNE.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                                          const char* s2_expression,
+                                          const char* s1,
+                                          const char* s2);
+
+// The helper function for {ASSERT|EXPECT}_STRCASENE.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRCASENE(const char* s1_expression,
+                                              const char* s2_expression,
+                                              const char* s1,
+                                              const char* s2);
+
+
+// Helper function for *_STREQ on wide strings.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTREQ(const char* expected_expression,
+                                          const char* actual_expression,
+                                          const wchar_t* expected,
+                                          const wchar_t* actual);
+
+// Helper function for *_STRNE on wide strings.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                                          const char* s2_expression,
+                                          const wchar_t* s1,
+                                          const wchar_t* s2);
+
+}  // namespace internal
+
+// IsSubstring() and IsNotSubstring() are intended to be used as the
+// first argument to {EXPECT,ASSERT}_PRED_FORMAT2(), not by
+// themselves.  They check whether needle is a substring of haystack
+// (NULL is considered a substring of itself only), and return an
+// appropriate error message when they fail.
+//
+// The {needle,haystack}_expr arguments are the stringified
+// expressions that generated the two real arguments.
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack);
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack);
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack);
+
+#if GTEST_HAS_STD_WSTRING
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack);
+#endif  // GTEST_HAS_STD_WSTRING
+
+namespace internal {
+
+// Helper template function for comparing floating-points.
+//
+// Template parameter:
+//
+//   RawType: the raw floating-point type (either float or double)
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+template 
+AssertionResult CmpHelperFloatingPointEQ(const char* expected_expression,
+                                         const char* actual_expression,
+                                         RawType expected,
+                                         RawType actual) {
+  const FloatingPoint lhs(expected), rhs(actual);
+
+  if (lhs.AlmostEquals(rhs)) {
+    return AssertionSuccess();
+  }
+
+  ::std::stringstream expected_ss;
+  expected_ss << std::setprecision(std::numeric_limits::digits10 + 2)
+              << expected;
+
+  ::std::stringstream actual_ss;
+  actual_ss << std::setprecision(std::numeric_limits::digits10 + 2)
+            << actual;
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   StringStreamToString(&expected_ss),
+                   StringStreamToString(&actual_ss),
+                   false);
+}
+
+// Helper function for implementing ASSERT_NEAR.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult DoubleNearPredFormat(const char* expr1,
+                                                const char* expr2,
+                                                const char* abs_error_expr,
+                                                double val1,
+                                                double val2,
+                                                double abs_error);
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+// A class that enables one to stream messages to assertion macros
+class GTEST_API_ AssertHelper {
+ public:
+  // Constructor.
+  AssertHelper(TestPartResult::Type type,
+               const char* file,
+               int line,
+               const char* message);
+  ~AssertHelper();
+
+  // Message assignment is a semantic trick to enable assertion
+  // streaming; see the GTEST_MESSAGE_ macro below.
+  void operator=(const Message& message) const;
+
+ private:
+  // We put our data in a struct so that the size of the AssertHelper class can
+  // be as small as possible.  This is important because gcc is incapable of
+  // re-using stack space even for temporary variables, so every EXPECT_EQ
+  // reserves stack space for another AssertHelper.
+  struct AssertHelperData {
+    AssertHelperData(TestPartResult::Type t,
+                     const char* srcfile,
+                     int line_num,
+                     const char* msg)
+        : type(t), file(srcfile), line(line_num), message(msg) { }
+
+    TestPartResult::Type const type;
+    const char* const file;
+    int const line;
+    std::string const message;
+
+   private:
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(AssertHelperData);
+  };
+
+  AssertHelperData* const data_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(AssertHelper);
+};
+
+}  // namespace internal
+
+#if GTEST_HAS_PARAM_TEST
+// The pure interface class that all value-parameterized tests inherit from.
+// A value-parameterized class must inherit from both ::testing::Test and
+// ::testing::WithParamInterface. In most cases that just means inheriting
+// from ::testing::TestWithParam, but more complicated test hierarchies
+// may need to inherit from Test and WithParamInterface at different levels.
+//
+// This interface has support for accessing the test parameter value via
+// the GetParam() method.
+//
+// Use it with one of the parameter generator defining functions, like Range(),
+// Values(), ValuesIn(), Bool(), and Combine().
+//
+// class FooTest : public ::testing::TestWithParam {
+//  protected:
+//   FooTest() {
+//     // Can use GetParam() here.
+//   }
+//   virtual ~FooTest() {
+//     // Can use GetParam() here.
+//   }
+//   virtual void SetUp() {
+//     // Can use GetParam() here.
+//   }
+//   virtual void TearDown {
+//     // Can use GetParam() here.
+//   }
+// };
+// TEST_P(FooTest, DoesBar) {
+//   // Can use GetParam() method here.
+//   Foo foo;
+//   ASSERT_TRUE(foo.DoesBar(GetParam()));
+// }
+// INSTANTIATE_TEST_CASE_P(OneToTenRange, FooTest, ::testing::Range(1, 10));
+
+template 
+class WithParamInterface {
+ public:
+  typedef T ParamType;
+  virtual ~WithParamInterface() {}
+
+  // The current parameter value. Is also available in the test fixture's
+  // constructor. This member function is non-static, even though it only
+  // references static data, to reduce the opportunity for incorrect uses
+  // like writing 'WithParamInterface::GetParam()' for a test that
+  // uses a fixture whose parameter type is int.
+  const ParamType& GetParam() const {
+    GTEST_CHECK_(parameter_ != NULL)
+        << "GetParam() can only be called inside a value-parameterized test "
+        << "-- did you intend to write TEST_P instead of TEST_F?";
+    return *parameter_;
+  }
+
+ private:
+  // Sets parameter value. The caller is responsible for making sure the value
+  // remains alive and unchanged throughout the current test.
+  static void SetParam(const ParamType* parameter) {
+    parameter_ = parameter;
+  }
+
+  // Static value used for accessing parameter during a test lifetime.
+  static const ParamType* parameter_;
+
+  // TestClass must be a subclass of WithParamInterface and Test.
+  template  friend class internal::ParameterizedTestFactory;
+};
+
+template 
+const T* WithParamInterface::parameter_ = NULL;
+
+// Most value-parameterized classes can ignore the existence of
+// WithParamInterface, and can just inherit from ::testing::TestWithParam.
+
+template 
+class TestWithParam : public Test, public WithParamInterface {
+};
+
+#endif  // GTEST_HAS_PARAM_TEST
+
+// Macros for indicating success/failure in test code.
+
+// ADD_FAILURE unconditionally adds a failure to the current test.
+// SUCCEED generates a success - it doesn't automatically make the
+// current test successful, as a test is only successful when it has
+// no failure.
+//
+// EXPECT_* verifies that a certain condition is satisfied.  If not,
+// it behaves like ADD_FAILURE.  In particular:
+//
+//   EXPECT_TRUE  verifies that a Boolean condition is true.
+//   EXPECT_FALSE verifies that a Boolean condition is false.
+//
+// FAIL and ASSERT_* are similar to ADD_FAILURE and EXPECT_*, except
+// that they will also abort the current function on failure.  People
+// usually want the fail-fast behavior of FAIL and ASSERT_*, but those
+// writing data-driven tests often find themselves using ADD_FAILURE
+// and EXPECT_* more.
+
+// Generates a nonfatal failure with a generic message.
+#define ADD_FAILURE() GTEST_NONFATAL_FAILURE_("Failed")
+
+// Generates a nonfatal failure at the given source file location with
+// a generic message.
+#define ADD_FAILURE_AT(file, line) \
+  GTEST_MESSAGE_AT_(file, line, "Failed", \
+                    ::testing::TestPartResult::kNonFatalFailure)
+
+// Generates a fatal failure with a generic message.
+#define GTEST_FAIL() GTEST_FATAL_FAILURE_("Failed")
+
+// Define this macro to 1 to omit the definition of FAIL(), which is a
+// generic name and clashes with some other libraries.
+#if !GTEST_DONT_DEFINE_FAIL
+# define FAIL() GTEST_FAIL()
+#endif
+
+// Generates a success with a generic message.
+#define GTEST_SUCCEED() GTEST_SUCCESS_("Succeeded")
+
+// Define this macro to 1 to omit the definition of SUCCEED(), which
+// is a generic name and clashes with some other libraries.
+#if !GTEST_DONT_DEFINE_SUCCEED
+# define SUCCEED() GTEST_SUCCEED()
+#endif
+
+// Macros for testing exceptions.
+//
+//    * {ASSERT|EXPECT}_THROW(statement, expected_exception):
+//         Tests that the statement throws the expected exception.
+//    * {ASSERT|EXPECT}_NO_THROW(statement):
+//         Tests that the statement doesn't throw any exception.
+//    * {ASSERT|EXPECT}_ANY_THROW(statement):
+//         Tests that the statement throws an exception.
+
+#define EXPECT_THROW(statement, expected_exception) \
+  GTEST_TEST_THROW_(statement, expected_exception, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_NO_THROW(statement) \
+  GTEST_TEST_NO_THROW_(statement, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_ANY_THROW(statement) \
+  GTEST_TEST_ANY_THROW_(statement, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_THROW(statement, expected_exception) \
+  GTEST_TEST_THROW_(statement, expected_exception, GTEST_FATAL_FAILURE_)
+#define ASSERT_NO_THROW(statement) \
+  GTEST_TEST_NO_THROW_(statement, GTEST_FATAL_FAILURE_)
+#define ASSERT_ANY_THROW(statement) \
+  GTEST_TEST_ANY_THROW_(statement, GTEST_FATAL_FAILURE_)
+
+// Boolean assertions. Condition can be either a Boolean expression or an
+// AssertionResult. For more information on how to use AssertionResult with
+// these macros see comments on that class.
+#define EXPECT_TRUE(condition) \
+  GTEST_TEST_BOOLEAN_(condition, #condition, false, true, \
+                      GTEST_NONFATAL_FAILURE_)
+#define EXPECT_FALSE(condition) \
+  GTEST_TEST_BOOLEAN_(!(condition), #condition, true, false, \
+                      GTEST_NONFATAL_FAILURE_)
+#define ASSERT_TRUE(condition) \
+  GTEST_TEST_BOOLEAN_(condition, #condition, false, true, \
+                      GTEST_FATAL_FAILURE_)
+#define ASSERT_FALSE(condition) \
+  GTEST_TEST_BOOLEAN_(!(condition), #condition, true, false, \
+                      GTEST_FATAL_FAILURE_)
+
+// Includes the auto-generated header that implements a family of
+// generic predicate assertion macros.
+// Copyright 2006, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// This file is AUTOMATICALLY GENERATED on 10/31/2011 by command
+// 'gen_gtest_pred_impl.py 5'.  DO NOT EDIT BY HAND!
+//
+// Implements a family of generic predicate assertion macros.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PRED_IMPL_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PRED_IMPL_H_
+
+// Makes sure this header is not included before gtest.h.
+#ifndef GTEST_INCLUDE_GTEST_GTEST_H_
+# error Do not include gtest_pred_impl.h directly.  Include gtest.h instead.
+#endif  // GTEST_INCLUDE_GTEST_GTEST_H_
+
+// This header implements a family of generic predicate assertion
+// macros:
+//
+//   ASSERT_PRED_FORMAT1(pred_format, v1)
+//   ASSERT_PRED_FORMAT2(pred_format, v1, v2)
+//   ...
+//
+// where pred_format is a function or functor that takes n (in the
+// case of ASSERT_PRED_FORMATn) values and their source expression
+// text, and returns a testing::AssertionResult.  See the definition
+// of ASSERT_EQ in gtest.h for an example.
+//
+// If you don't care about formatting, you can use the more
+// restrictive version:
+//
+//   ASSERT_PRED1(pred, v1)
+//   ASSERT_PRED2(pred, v1, v2)
+//   ...
+//
+// where pred is an n-ary function or functor that returns bool,
+// and the values v1, v2, ..., must support the << operator for
+// streaming to std::ostream.
+//
+// We also define the EXPECT_* variations.
+//
+// For now we only support predicates whose arity is at most 5.
+// Please email googletestframework@googlegroups.com if you need
+// support for higher arities.
+
+// GTEST_ASSERT_ is the basic statement to which all of the assertions
+// in this file reduce.  Don't use this in your code.
+
+#define GTEST_ASSERT_(expression, on_failure) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (const ::testing::AssertionResult gtest_ar = (expression)) \
+    ; \
+  else \
+    on_failure(gtest_ar.failure_message())
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED1.  Don't use
+// this in your code.
+template 
+AssertionResult AssertPred1Helper(const char* pred_text,
+                                  const char* e1,
+                                  Pred pred,
+                                  const T1& v1) {
+  if (pred(v1)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT1.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT1_(pred_format, v1, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, v1), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED1.  Don't use
+// this in your code.
+#define GTEST_PRED1_(pred, v1, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred1Helper(#pred, \
+                                             #v1, \
+                                             pred, \
+                                             v1), on_failure)
+
+// Unary predicate assertion macros.
+#define EXPECT_PRED_FORMAT1(pred_format, v1) \
+  GTEST_PRED_FORMAT1_(pred_format, v1, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED1(pred, v1) \
+  GTEST_PRED1_(pred, v1, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT1(pred_format, v1) \
+  GTEST_PRED_FORMAT1_(pred_format, v1, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED1(pred, v1) \
+  GTEST_PRED1_(pred, v1, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED2.  Don't use
+// this in your code.
+template 
+AssertionResult AssertPred2Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2) {
+  if (pred(v1, v2)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT2.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT2_(pred_format, v1, v2, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, v1, v2), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED2.  Don't use
+// this in your code.
+#define GTEST_PRED2_(pred, v1, v2, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred2Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             pred, \
+                                             v1, \
+                                             v2), on_failure)
+
+// Binary predicate assertion macros.
+#define EXPECT_PRED_FORMAT2(pred_format, v1, v2) \
+  GTEST_PRED_FORMAT2_(pred_format, v1, v2, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED2(pred, v1, v2) \
+  GTEST_PRED2_(pred, v1, v2, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT2(pred_format, v1, v2) \
+  GTEST_PRED_FORMAT2_(pred_format, v1, v2, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED2(pred, v1, v2) \
+  GTEST_PRED2_(pred, v1, v2, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED3.  Don't use
+// this in your code.
+template 
+AssertionResult AssertPred3Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  const char* e3,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2,
+                                  const T3& v3) {
+  if (pred(v1, v2, v3)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ", "
+                            << e3 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2
+                            << "\n" << e3 << " evaluates to " << v3;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT3.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT3_(pred_format, v1, v2, v3, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, #v3, v1, v2, v3), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED3.  Don't use
+// this in your code.
+#define GTEST_PRED3_(pred, v1, v2, v3, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred3Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             #v3, \
+                                             pred, \
+                                             v1, \
+                                             v2, \
+                                             v3), on_failure)
+
+// Ternary predicate assertion macros.
+#define EXPECT_PRED_FORMAT3(pred_format, v1, v2, v3) \
+  GTEST_PRED_FORMAT3_(pred_format, v1, v2, v3, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED3(pred, v1, v2, v3) \
+  GTEST_PRED3_(pred, v1, v2, v3, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT3(pred_format, v1, v2, v3) \
+  GTEST_PRED_FORMAT3_(pred_format, v1, v2, v3, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED3(pred, v1, v2, v3) \
+  GTEST_PRED3_(pred, v1, v2, v3, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED4.  Don't use
+// this in your code.
+template 
+AssertionResult AssertPred4Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  const char* e3,
+                                  const char* e4,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2,
+                                  const T3& v3,
+                                  const T4& v4) {
+  if (pred(v1, v2, v3, v4)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ", "
+                            << e3 << ", "
+                            << e4 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2
+                            << "\n" << e3 << " evaluates to " << v3
+                            << "\n" << e4 << " evaluates to " << v4;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT4.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT4_(pred_format, v1, v2, v3, v4, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, #v3, #v4, v1, v2, v3, v4), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED4.  Don't use
+// this in your code.
+#define GTEST_PRED4_(pred, v1, v2, v3, v4, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred4Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             #v3, \
+                                             #v4, \
+                                             pred, \
+                                             v1, \
+                                             v2, \
+                                             v3, \
+                                             v4), on_failure)
+
+// 4-ary predicate assertion macros.
+#define EXPECT_PRED_FORMAT4(pred_format, v1, v2, v3, v4) \
+  GTEST_PRED_FORMAT4_(pred_format, v1, v2, v3, v4, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED4(pred, v1, v2, v3, v4) \
+  GTEST_PRED4_(pred, v1, v2, v3, v4, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT4(pred_format, v1, v2, v3, v4) \
+  GTEST_PRED_FORMAT4_(pred_format, v1, v2, v3, v4, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED4(pred, v1, v2, v3, v4) \
+  GTEST_PRED4_(pred, v1, v2, v3, v4, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED5.  Don't use
+// this in your code.
+template 
+AssertionResult AssertPred5Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  const char* e3,
+                                  const char* e4,
+                                  const char* e5,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2,
+                                  const T3& v3,
+                                  const T4& v4,
+                                  const T5& v5) {
+  if (pred(v1, v2, v3, v4, v5)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ", "
+                            << e3 << ", "
+                            << e4 << ", "
+                            << e5 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2
+                            << "\n" << e3 << " evaluates to " << v3
+                            << "\n" << e4 << " evaluates to " << v4
+                            << "\n" << e5 << " evaluates to " << v5;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT5.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT5_(pred_format, v1, v2, v3, v4, v5, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, #v3, #v4, #v5, v1, v2, v3, v4, v5), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED5.  Don't use
+// this in your code.
+#define GTEST_PRED5_(pred, v1, v2, v3, v4, v5, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred5Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             #v3, \
+                                             #v4, \
+                                             #v5, \
+                                             pred, \
+                                             v1, \
+                                             v2, \
+                                             v3, \
+                                             v4, \
+                                             v5), on_failure)
+
+// 5-ary predicate assertion macros.
+#define EXPECT_PRED_FORMAT5(pred_format, v1, v2, v3, v4, v5) \
+  GTEST_PRED_FORMAT5_(pred_format, v1, v2, v3, v4, v5, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED5(pred, v1, v2, v3, v4, v5) \
+  GTEST_PRED5_(pred, v1, v2, v3, v4, v5, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT5(pred_format, v1, v2, v3, v4, v5) \
+  GTEST_PRED_FORMAT5_(pred_format, v1, v2, v3, v4, v5, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED5(pred, v1, v2, v3, v4, v5) \
+  GTEST_PRED5_(pred, v1, v2, v3, v4, v5, GTEST_FATAL_FAILURE_)
+
+
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PRED_IMPL_H_
+
+// Macros for testing equalities and inequalities.
+//
+//    * {ASSERT|EXPECT}_EQ(expected, actual): Tests that expected == actual
+//    * {ASSERT|EXPECT}_NE(v1, v2):           Tests that v1 != v2
+//    * {ASSERT|EXPECT}_LT(v1, v2):           Tests that v1 < v2
+//    * {ASSERT|EXPECT}_LE(v1, v2):           Tests that v1 <= v2
+//    * {ASSERT|EXPECT}_GT(v1, v2):           Tests that v1 > v2
+//    * {ASSERT|EXPECT}_GE(v1, v2):           Tests that v1 >= v2
+//
+// When they are not, Google Test prints both the tested expressions and
+// their actual values.  The values must be compatible built-in types,
+// or you will get a compiler error.  By "compatible" we mean that the
+// values can be compared by the respective operator.
+//
+// Note:
+//
+//   1. It is possible to make a user-defined type work with
+//   {ASSERT|EXPECT}_??(), but that requires overloading the
+//   comparison operators and is thus discouraged by the Google C++
+//   Usage Guide.  Therefore, you are advised to use the
+//   {ASSERT|EXPECT}_TRUE() macro to assert that two objects are
+//   equal.
+//
+//   2. The {ASSERT|EXPECT}_??() macros do pointer comparisons on
+//   pointers (in particular, C strings).  Therefore, if you use it
+//   with two C strings, you are testing how their locations in memory
+//   are related, not how their content is related.  To compare two C
+//   strings by content, use {ASSERT|EXPECT}_STR*().
+//
+//   3. {ASSERT|EXPECT}_EQ(expected, actual) is preferred to
+//   {ASSERT|EXPECT}_TRUE(expected == actual), as the former tells you
+//   what the actual value is when it fails, and similarly for the
+//   other comparisons.
+//
+//   4. Do not depend on the order in which {ASSERT|EXPECT}_??()
+//   evaluate their arguments, which is undefined.
+//
+//   5. These macros evaluate their arguments exactly once.
+//
+// Examples:
+//
+//   EXPECT_NE(5, Foo());
+//   EXPECT_EQ(NULL, a_pointer);
+//   ASSERT_LT(i, array_size);
+//   ASSERT_GT(records.size(), 0) << "There is no record left.";
+
+#define EXPECT_EQ(expected, actual) \
+  EXPECT_PRED_FORMAT2(::testing::internal:: \
+                      EqHelper::Compare, \
+                      expected, actual)
+#define EXPECT_NE(expected, actual) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperNE, expected, actual)
+#define EXPECT_LE(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperLE, val1, val2)
+#define EXPECT_LT(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperLT, val1, val2)
+#define EXPECT_GE(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperGE, val1, val2)
+#define EXPECT_GT(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperGT, val1, val2)
+
+#define GTEST_ASSERT_EQ(expected, actual) \
+  ASSERT_PRED_FORMAT2(::testing::internal:: \
+                      EqHelper::Compare, \
+                      expected, actual)
+#define GTEST_ASSERT_NE(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperNE, val1, val2)
+#define GTEST_ASSERT_LE(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperLE, val1, val2)
+#define GTEST_ASSERT_LT(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperLT, val1, val2)
+#define GTEST_ASSERT_GE(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperGE, val1, val2)
+#define GTEST_ASSERT_GT(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperGT, val1, val2)
+
+// Define macro GTEST_DONT_DEFINE_ASSERT_XY to 1 to omit the definition of
+// ASSERT_XY(), which clashes with some users' own code.
+
+#if !GTEST_DONT_DEFINE_ASSERT_EQ
+# define ASSERT_EQ(val1, val2) GTEST_ASSERT_EQ(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_NE
+# define ASSERT_NE(val1, val2) GTEST_ASSERT_NE(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_LE
+# define ASSERT_LE(val1, val2) GTEST_ASSERT_LE(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_LT
+# define ASSERT_LT(val1, val2) GTEST_ASSERT_LT(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_GE
+# define ASSERT_GE(val1, val2) GTEST_ASSERT_GE(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_GT
+# define ASSERT_GT(val1, val2) GTEST_ASSERT_GT(val1, val2)
+#endif
+
+// C-string Comparisons.  All tests treat NULL and any non-NULL string
+// as different.  Two NULLs are equal.
+//
+//    * {ASSERT|EXPECT}_STREQ(s1, s2):     Tests that s1 == s2
+//    * {ASSERT|EXPECT}_STRNE(s1, s2):     Tests that s1 != s2
+//    * {ASSERT|EXPECT}_STRCASEEQ(s1, s2): Tests that s1 == s2, ignoring case
+//    * {ASSERT|EXPECT}_STRCASENE(s1, s2): Tests that s1 != s2, ignoring case
+//
+// For wide or narrow string objects, you can use the
+// {ASSERT|EXPECT}_??() macros.
+//
+// Don't depend on the order in which the arguments are evaluated,
+// which is undefined.
+//
+// These macros evaluate their arguments exactly once.
+
+#define EXPECT_STREQ(expected, actual) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTREQ, expected, actual)
+#define EXPECT_STRNE(s1, s2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTRNE, s1, s2)
+#define EXPECT_STRCASEEQ(expected, actual) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASEEQ, expected, actual)
+#define EXPECT_STRCASENE(s1, s2)\
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASENE, s1, s2)
+
+#define ASSERT_STREQ(expected, actual) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTREQ, expected, actual)
+#define ASSERT_STRNE(s1, s2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTRNE, s1, s2)
+#define ASSERT_STRCASEEQ(expected, actual) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASEEQ, expected, actual)
+#define ASSERT_STRCASENE(s1, s2)\
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASENE, s1, s2)
+
+// Macros for comparing floating-point numbers.
+//
+//    * {ASSERT|EXPECT}_FLOAT_EQ(expected, actual):
+//         Tests that two float values are almost equal.
+//    * {ASSERT|EXPECT}_DOUBLE_EQ(expected, actual):
+//         Tests that two double values are almost equal.
+//    * {ASSERT|EXPECT}_NEAR(v1, v2, abs_error):
+//         Tests that v1 and v2 are within the given distance to each other.
+//
+// Google Test uses ULP-based comparison to automatically pick a default
+// error bound that is appropriate for the operands.  See the
+// FloatingPoint template class in gtest-internal.h if you are
+// interested in the implementation details.
+
+#define EXPECT_FLOAT_EQ(expected, actual)\
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ, \
+                      expected, actual)
+
+#define EXPECT_DOUBLE_EQ(expected, actual)\
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ, \
+                      expected, actual)
+
+#define ASSERT_FLOAT_EQ(expected, actual)\
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ, \
+                      expected, actual)
+
+#define ASSERT_DOUBLE_EQ(expected, actual)\
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ, \
+                      expected, actual)
+
+#define EXPECT_NEAR(val1, val2, abs_error)\
+  EXPECT_PRED_FORMAT3(::testing::internal::DoubleNearPredFormat, \
+                      val1, val2, abs_error)
+
+#define ASSERT_NEAR(val1, val2, abs_error)\
+  ASSERT_PRED_FORMAT3(::testing::internal::DoubleNearPredFormat, \
+                      val1, val2, abs_error)
+
+// These predicate format functions work on floating-point values, and
+// can be used in {ASSERT|EXPECT}_PRED_FORMAT2*(), e.g.
+//
+//   EXPECT_PRED_FORMAT2(testing::DoubleLE, Foo(), 5.0);
+
+// Asserts that val1 is less than, or almost equal to, val2.  Fails
+// otherwise.  In particular, it fails if either val1 or val2 is NaN.
+GTEST_API_ AssertionResult FloatLE(const char* expr1, const char* expr2,
+                                   float val1, float val2);
+GTEST_API_ AssertionResult DoubleLE(const char* expr1, const char* expr2,
+                                    double val1, double val2);
+
+
+#if GTEST_OS_WINDOWS
+
+// Macros that test for HRESULT failure and success, these are only useful
+// on Windows, and rely on Windows SDK macros and APIs to compile.
+//
+//    * {ASSERT|EXPECT}_HRESULT_{SUCCEEDED|FAILED}(expr)
+//
+// When expr unexpectedly fails or succeeds, Google Test prints the
+// expected result and the actual result with both a human-readable
+// string representation of the error, if available, as well as the
+// hex result code.
+# define EXPECT_HRESULT_SUCCEEDED(expr) \
+    EXPECT_PRED_FORMAT1(::testing::internal::IsHRESULTSuccess, (expr))
+
+# define ASSERT_HRESULT_SUCCEEDED(expr) \
+    ASSERT_PRED_FORMAT1(::testing::internal::IsHRESULTSuccess, (expr))
+
+# define EXPECT_HRESULT_FAILED(expr) \
+    EXPECT_PRED_FORMAT1(::testing::internal::IsHRESULTFailure, (expr))
+
+# define ASSERT_HRESULT_FAILED(expr) \
+    ASSERT_PRED_FORMAT1(::testing::internal::IsHRESULTFailure, (expr))
+
+#endif  // GTEST_OS_WINDOWS
+
+// Macros that execute statement and check that it doesn't generate new fatal
+// failures in the current thread.
+//
+//   * {ASSERT|EXPECT}_NO_FATAL_FAILURE(statement);
+//
+// Examples:
+//
+//   EXPECT_NO_FATAL_FAILURE(Process());
+//   ASSERT_NO_FATAL_FAILURE(Process()) << "Process() failed";
+//
+#define ASSERT_NO_FATAL_FAILURE(statement) \
+    GTEST_TEST_NO_FATAL_FAILURE_(statement, GTEST_FATAL_FAILURE_)
+#define EXPECT_NO_FATAL_FAILURE(statement) \
+    GTEST_TEST_NO_FATAL_FAILURE_(statement, GTEST_NONFATAL_FAILURE_)
+
+// Causes a trace (including the source file path, the current line
+// number, and the given message) to be included in every test failure
+// message generated by code in the current scope.  The effect is
+// undone when the control leaves the current scope.
+//
+// The message argument can be anything streamable to std::ostream.
+//
+// In the implementation, we include the current line number as part
+// of the dummy variable name, thus allowing multiple SCOPED_TRACE()s
+// to appear in the same block - as long as they are on different
+// lines.
+#define SCOPED_TRACE(message) \
+  ::testing::internal::ScopedTrace GTEST_CONCAT_TOKEN_(gtest_trace_, __LINE__)(\
+    __FILE__, __LINE__, ::testing::Message() << (message))
+
+// Compile-time assertion for type equality.
+// StaticAssertTypeEq() compiles iff type1 and type2 are
+// the same type.  The value it returns is not interesting.
+//
+// Instead of making StaticAssertTypeEq a class template, we make it a
+// function template that invokes a helper class template.  This
+// prevents a user from misusing StaticAssertTypeEq by
+// defining objects of that type.
+//
+// CAVEAT:
+//
+// When used inside a method of a class template,
+// StaticAssertTypeEq() is effective ONLY IF the method is
+// instantiated.  For example, given:
+//
+//   template  class Foo {
+//    public:
+//     void Bar() { testing::StaticAssertTypeEq(); }
+//   };
+//
+// the code:
+//
+//   void Test1() { Foo foo; }
+//
+// will NOT generate a compiler error, as Foo::Bar() is never
+// actually instantiated.  Instead, you need:
+//
+//   void Test2() { Foo foo; foo.Bar(); }
+//
+// to cause a compiler error.
+template 
+bool StaticAssertTypeEq() {
+  (void)internal::StaticAssertTypeEqHelper();
+  return true;
+}
+
+// Defines a test.
+//
+// The first parameter is the name of the test case, and the second
+// parameter is the name of the test within the test case.
+//
+// The convention is to end the test case name with "Test".  For
+// example, a test case for the Foo class can be named FooTest.
+//
+// The user should put his test code between braces after using this
+// macro.  Example:
+//
+//   TEST(FooTest, InitializesCorrectly) {
+//     Foo foo;
+//     EXPECT_TRUE(foo.StatusIsOK());
+//   }
+
+// Note that we call GetTestTypeId() instead of GetTypeId<
+// ::testing::Test>() here to get the type ID of testing::Test.  This
+// is to work around a suspected linker bug when using Google Test as
+// a framework on Mac OS X.  The bug causes GetTypeId<
+// ::testing::Test>() to return different values depending on whether
+// the call is from the Google Test framework itself or from user test
+// code.  GetTestTypeId() is guaranteed to always return the same
+// value, as it always calls GetTypeId<>() from the Google Test
+// framework.
+#define GTEST_TEST(test_case_name, test_name)\
+  GTEST_TEST_(test_case_name, test_name, \
+              ::testing::Test, ::testing::internal::GetTestTypeId())
+
+// Define this macro to 1 to omit the definition of TEST(), which
+// is a generic name and clashes with some other libraries.
+#if !GTEST_DONT_DEFINE_TEST
+# define TEST(test_case_name, test_name) GTEST_TEST(test_case_name, test_name)
+#endif
+
+// Defines a test that uses a test fixture.
+//
+// The first parameter is the name of the test fixture class, which
+// also doubles as the test case name.  The second parameter is the
+// name of the test within the test case.
+//
+// A test fixture class must be declared earlier.  The user should put
+// his test code between braces after using this macro.  Example:
+//
+//   class FooTest : public testing::Test {
+//    protected:
+//     virtual void SetUp() { b_.AddElement(3); }
+//
+//     Foo a_;
+//     Foo b_;
+//   };
+//
+//   TEST_F(FooTest, InitializesCorrectly) {
+//     EXPECT_TRUE(a_.StatusIsOK());
+//   }
+//
+//   TEST_F(FooTest, ReturnsElementCountCorrectly) {
+//     EXPECT_EQ(0, a_.size());
+//     EXPECT_EQ(1, b_.size());
+//   }
+
+#define TEST_F(test_fixture, test_name)\
+  GTEST_TEST_(test_fixture, test_name, test_fixture, \
+              ::testing::internal::GetTypeId())
+
+}  // namespace testing
+
+// Use this function in main() to run all tests.  It returns 0 if all
+// tests are successful, or 1 otherwise.
+//
+// RUN_ALL_TESTS() should be invoked after the command line has been
+// parsed by InitGoogleTest().
+//
+// This function was formerly a macro; thus, it is in the global
+// namespace and has an all-caps name.
+int RUN_ALL_TESTS() GTEST_MUST_USE_RESULT_;
+
+inline int RUN_ALL_TESTS() {
+  return ::testing::UnitTest::GetInstance()->Run();
+}
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_H_