Build fails for openmp

[almishra]

When trying to build openmp I'm getting the following error:

[ 42%] Building CXX object libomptarget/plugins/cuda/CMakeFiles/omptarget.rtl.cuda.dir/src/rtl.cpp.o
/gpfs/projects/ChapmanGroup/alok/git/llvm-project-build/openmp/libomptarget/plugins/cuda/src/rtl.cpp:262:18: error: no member named 'make_unique' in namespace 'std'
      Ptr = std::make_unique<std::atomic_uint>(0);
            ~~~~~^
/gpfs/projects/ChapmanGroup/alok/git/llvm-project-build/openmp/libomptarget/plugins/cuda/src/rtl.cpp:262:46: error: expected '(' for function-style cast or type construction
      Ptr = std::make_unique<std::atomic_uint>(0);
                             ~~~~~~~~~~~~~~~~^
2 errors generated.
make[2]: *** [libomptarget/plugins/cuda/CMakeFiles/omptarget.rtl.cuda.dir/src/rtl.cpp.o] Error 1
make[1]: *** [libomptarget/plugins/cuda/CMakeFiles/omptarget.rtl.cuda.dir/all] Error 2
make: *** [all] Error 2

I never faced this issue till a couple of weeks back.
I always build my llvm with clang first and then use the clang to build openmp with offloading support.

To build llvm/clang:

cmake -G "Unix Makefiles" -DLLVM_ENABLE_PROJECTS="clang" -DCMAKE_BUILD_TYPE=Release -DCLANG_BUILD_EXAMPLES=1 -DCMAKE_INSTALL_PREFIX=$LLVM_PATH -DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DLLVM_TARGETS_TO_BUILD="X86;NVPTX" $LLVM_SRC/llvm
make -j16
make install

To build openmp:

cmake -G "Unix Makefiles" -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=$LLVM_PATH -DCMAKE_C_COMPILER=$LLVM_PATH/bin/clang -DCMAKE_CXX_COMPILER=$LLVM_PATH/bin/clang++ -DLIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES=35,60,70 $LLVM_SRC/openmp
make -j16
make install

My gcc version is 6.5.0

$ gcc --version
gcc (GCC) 6.5.0
Copyright (C) 2017 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

[JonChesterfield]

Stale cmake files? They should be setting C++ 14 as of a couple of weeks ago

edit: though I can't find where we set the c++ version for the plugins (not currently at a desk) so it maybe we're just using the default, which is 11 and won't work any more. If so you can work around by specifying c++14

[almishra]

I see that patch - 3ff4e2e

And I am still able to build openmp if I checkout the commit before this patch.

[almishra]

I have tried passing c++14 as option to cmake. It didn't work

-DCMAKE_CXX_FLAGS="-std=c++14"

I get the same error.

[shiltian]

Finally I can reproduced it now, but I didn’t find anything wrong in the dry run:

$ make -n
cd /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build && /gpfs/software/cmake-3.16.2/bin/cmake -S/gpfs/scratch/shiltian/Documents/llvm-project/openmp -B/gpfs/scratch/shiltian/Documents/llvm-project/openmp/build --check-build-system CMakeFiles/Makefile.cmake 0
cd /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build && /gpfs/software/cmake-3.16.2/bin/cmake -E cmake_progress_start /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/CMakeFiles /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/libomptarget/plugins/cuda/CMakeFiles/progress.marks
cd /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build && make -f CMakeFiles/Makefile2 libomptarget/plugins/cuda/all
make -f libomptarget/plugins/cuda/CMakeFiles/omptarget.rtl.cuda.dir/build.make libomptarget/plugins/cuda/CMakeFiles/omptarget.rtl.cuda.dir/depend
cd /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build && /gpfs/software/cmake-3.16.2/bin/cmake -E cmake_depends "Unix Makefiles" /gpfs/scratch/shiltian/Documents/llvm-project/openmp /gpfs/scratch/shiltian/Documents/llvm-project/openmp/libomptarget/plugins/cuda /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/libomptarget/plugins/cuda /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/libomptarget/plugins/cuda/CMakeFiles/omptarget.rtl.cuda.dir/DependInfo.cmake --color=
make -f libomptarget/plugins/cuda/CMakeFiles/omptarget.rtl.cuda.dir/build.make libomptarget/plugins/cuda/CMakeFiles/omptarget.rtl.cuda.dir/build
/gpfs/software/cmake-3.16.2/bin/cmake -E cmake_echo_color --switch= --green --progress-dir=/gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/CMakeFiles --progress-num=79 "Building CXX object libomptarget/plugins/cuda/CMakeFiles/omptarget.rtl.cuda.dir/src/rtl.cpp.o"
cd /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/libomptarget/plugins/cuda && /gpfs/scratch/shiltian/Documents/llvm-project/build/bin/clang++  -DTARGET_NAME=CUDA -Domptarget_rtl_cuda_EXPORTS -I/gpfs/scratch/shiltian/Documents/llvm-project/openmp/libomptarget/include -I/gpfs/software/cuda/10.0/include  -Wall -Wcast-qual -Wformat-pedantic -Wimplicit-fallthrough -Wsign-compare -Wno-extra -Wno-pedantic -std=gnu++14 -O3 -DNDEBUG -fPIC   -o CMakeFiles/omptarget.rtl.cuda.dir/src/rtl.cpp.o -c /gpfs/scratch/shiltian/Documents/llvm-project/openmp/libomptarget/plugins/cuda/src/rtl.cpp
/gpfs/software/cmake-3.16.2/bin/cmake -E cmake_echo_color --switch= --green --bold --progress-dir=/gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/CMakeFiles --progress-num=80 "Linking CXX shared library ../../libomptarget.rtl.cuda.so"
cd /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/libomptarget/plugins/cuda && /gpfs/software/cmake-3.16.2/bin/cmake -E cmake_link_script CMakeFiles/omptarget.rtl.cuda.dir/link.txt --verbose=
/gpfs/software/cmake-3.16.2/bin/cmake -E cmake_echo_color --switch= --progress-dir=/gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/CMakeFiles --progress-num=79,80 "Built target omptarget.rtl.cuda"
/gpfs/software/cmake-3.16.2/bin/cmake -E cmake_progress_start /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/CMakeFiles 0

As you can see -std=gnu++14 has already there, so it is of course no use to append it yourself. I’ll try to figure out what is wrong on this system.

[shiltian]

I have root cause the issue, if you append --gcc-toolchain=/gpfs/software/gcc/6.5.0, problem solved. The reason is clang by default will use system provided gcc, even if another gcc has already been in PATH, etc. If you try to print the compilation command appending -###, you can see:

-bash-4.2$ cd /gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/libomptarget/plugins/cuda && /gpfs/scratch/shiltian/Documents/llvm-project/build/bin/clang++  -DTARGET_NAME=CUDA -Domptarget_rtl_cuda_EXPORTS -I/gpfs/scratch/shiltian/Documents/llvm-project/openmp/libomptarget/include -I/gpfs/software/cuda/10.0/include  -Wall -Wcast-qual -Wformat-pedantic -Wimplicit-fallthrough -Wsign-compare -Wno-extra -Wno-pedantic -std=gnu++14 -O3 -DNDEBUG -fPIC   -o CMakeFiles/omptarget.rtl.cuda.dir/src/rtl.cpp.o -c /gpfs/scratch/shiltian/Documents/llvm-project/openmp/libomptarget/plugins/cuda/src/rtl.cpp -###
clang version 11.0.0 (git@github.com:llvm/llvm-project.git 4fa63fd4524c9818dab2fc0d3ea3a8d65d3bc22b)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /gpfs/scratch/shiltian/Documents/llvm-project/build/bin
 (in-process)
 "/gpfs/scratch/shiltian/Documents/llvm-project/build/bin/clang-11" "-cc1" "-triple" "x86_64-unknown-linux-gnu" "-emit-obj" "-disable-free" "-disable-llvm-verifier" "-discard-value-names" "-main-file-name" "rtl.cpp" "-mrelocation-model" "pic" "-pic-level" "2" "-mthread-model" "posix" "-mframe-pointer=none" "-fmath-errno" "-fdenormal-fp-math=ieee,ieee" "-fdenormal-fp-math-f32=ieee,ieee" "-fno-rounding-math" "-masm-verbose" "-mconstructor-aliases" "-munwind-tables" "-target-cpu" "x86-64" "-dwarf-column-info" "-fno-split-dwarf-inlining" "-debugger-tuning=gdb" "-resource-dir" "/gpfs/scratch/shiltian/Documents/llvm-project/build/lib/clang/11.0.0" "-D" "TARGET_NAME=CUDA" "-D" "omptarget_rtl_cuda_EXPORTS" "-I" "/gpfs/scratch/shiltian/Documents/llvm-project/openmp/libomptarget/include" "-I" "/gpfs/software/cuda/10.0/include" "-D" "NDEBUG" "-I/cm/shared/apps/cuda100/sdk/10.0/common/inc" "-I/gpfs/software/cuda/10.0/include" "-I/gpfs/software/cmake-3.16.2/share/cmake-3.16.2/include" "-I/gpfs/software/gcc/6.5.0/include" "-c-isystem" "/gpfs/software/cmake-3.16.2/share/cmake-3.16.2/include" "-c-isystem" "/gpfs/software/gcc/6.5.0/include" "-cxx-isystem" "/gpfs/software/cmake-3.16.2/share/cmake-3.16.2/include" "-cxx-isystem" "/gpfs/software/gcc/6.5.0/include" "-objc-isystem" "/gpfs/software/cmake-3.16.2/share/cmake-3.16.2/include" "-objc-isystem" "/gpfs/software/gcc/6.5.0/include" "-internal-isystem" "/usr/lib/gcc/x86_64-redhat-linux/4.8.5/../../../../include/c++/4.8.5" "-internal-isystem" "/usr/lib/gcc/x86_64-redhat-linux/4.8.5/../../../../include/c++/4.8.5/x86_64-redhat-linux" "-internal-isystem" "/usr/lib/gcc/x86_64-redhat-linux/4.8.5/../../../../include/c++/4.8.5/backward" "-internal-isystem" "/usr/local/include" "-internal-isystem" "/gpfs/scratch/shiltian/Documents/llvm-project/build/lib/clang/11.0.0/include" "-internal-externc-isystem" "/include" "-internal-externc-isystem" "/usr/include" "-O3" "-Wall" "-Wcast-qual" "-Wformat-pedantic" "-Wimplicit-fallthrough" "-Wsign-compare" "-Wno-extra" "-Wno-pedantic" "-std=gnu++14" "-fdeprecated-macro" "-fdebug-compilation-dir" "/gpfs/scratch/shiltian/Documents/llvm-project/openmp/build/libomptarget/plugins/cuda" "-ferror-limit" "19" "-fmessage-length" "0" "-fgnuc-version=4.2.1" "-fobjc-runtime=gcc" "-fcxx-exceptions" "-fexceptions" "-fdiagnostics-show-option" "-fcolor-diagnostics" "-vectorize-loops" "-vectorize-slp" "-faddrsig" "-o" "CMakeFiles/omptarget.rtl.cuda.dir/src/rtl.cpp.o" "-x" "c++" "/gpfs/scratch/shiltian/Documents/llvm-project/openmp/libomptarget/plugins/cuda/src/rtl.cpp"

Still using head files from system which is 4.8.5. GCC 4.8.5 has partially support for C++14 so that’s why that function is not found.

[almishra]

And how to set this option in cmake?

[shiltian]

Try CMAKE_<LANG>_COMPILER_EXTERNAL_TOOLCHAIN.

[almishra]

This worked. Thanks Shilei for the help.

[almishra]

I still don't understand one thing. Why does clang use system provided gcc, even if we specified our own compiler with the CMAKE_CXX_COMPILER option?

[shiltian]

That is different stories. CMAKE_<LANG>_COMPILER is telling CMake to use specific compiler. In your case, you want to use clang. You did right for this "story". Another story is only for clang, has nothing to do with CMake. Clang doesn't provide every C/C++ header, and it should not actually. Then it must use system provided one. The search path for system headers is hard coded into clang if you do not specify GCC toolchain. module only sets PATH, LIBRARY_PATH, C_INCLUDE_PATH, etc. It doesn't affect how clang searches for system headers actually.

[llvmbot]

@llvm/issue-subscribers-openmp