[flang][OpenMP] real(10) prevents compilation

[hakostra]

This issue is the same as issue #128015. I have a recent flang built from sources with OpenMP offload support:

$ flang --version
flang version 21.0.0git (https://github.com/llvm/llvm-project.git b68e8f1de71a65b21910aff4aea75fc1ca6bf6e0)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /opt/llvm/llvm-project/install/bin

I have found issue #128015 still to prevent targeting my Nvidia GPU (I have no AMD GPU to test with atm). The same small reproducer from PR #129742 still applies:

program real10
  use ieee_arithmetic
  implicit none
end program

Produces:

$ flang -fopenmp -fopenmp-targets=nvptx64 real10.F90
error: Semantic errors in real10.F90
/opt/llvm/llvm-project/install/bin/../include/flang/ieee_arithmetic.mod:143:1: error: REAL(KIND=10) is not an enabled type for this target
  real(10),intent(in)::x
  ^^^^^^^^^^^^^^^^^^^^^^
/opt/llvm/llvm-project/install/bin/../include/flang/ieee_arithmetic.mod:183:1: error: REAL(KIND=10) is not an enabled type for this target
  real(10),intent(in)::y

and so on for many, many lines. This error is also preventing me from using the HDF5 library in any projects with OpenMP, even if no HDF5 related functionality is ever touching the GPU...

If it makes any difference, my LLVM build is compiled with the following CMake command:

cmake -GNinja \
    -DCMAKE_BUILD_TYPE="Release" \
    -DCMAKE_INSTALL_PREFIX="/opt/llvm/llvm-project/install" \
    -DCMAKE_CUDA_COMPILER=${CUDACXX} \
    -DCMAKE_CUDA_HOST_COMPILER="clang++-19" \
    -DCMAKE_CUDA_ARCHITECTURES="89;120" \
    -DLLVM_ENABLE_PROJECTS="clang;lld;flang;lldb" \
    -DLLVM_ENABLE_RUNTIMES="compiler-rt;openmp;offload;flang-rt" \
    -DLLVM_TARGETS_TO_BUILD="X86;NVPTX" \
    -DLLVM_ENABLE_RTTI="ON" \
    -DOPENMP_ENABLE_LIBOMPTARGET="ON" \
    -DFLANG_RT_EXPERIMENTAL_OFFLOAD_SUPPORT="OpenMP" \
    -DFLANG_RT_DEVICE_ARCHITECTURES="sm_89;sm_120" \
    -DFLANG_EXPERIMENTAL_OMP_OFFLOAD_BUILD="host_device" \
    ../llvm

[eugeneepshteyn]

I ran the following command to see where the error happens: flang -fopenmp -fopenmp-targets=nvptx64 kind10.f90 -save-temps -###

The error happens during this step:

"flang" "-fc1" "-triple" "nvptx64-nvidia-cuda" "-emit-llvm-bc" "-ffixed-form" "-fcolor-diagnostics" "-mrelocation-model" "pic" "-pic-level" "2" "-target-cpu" "sm_80" "-fopenmp" "-resource-dir" ".../lib/clang/21" "-fopenmp-host-ir-file-path" "kind10-host-x86_64-unknown-linux-gnu.bc" "-fopenmp-is-target-device" "-mframe-pointer=all" "-o" "kind10-openmp-nvptx64-nvidia-cuda-sm_80.bc" "-save-temps=cwd" "-x" "f95-cpp-input" "kind10-openmp-nvptx64-nvidia-cuda-sm_80.i"

It seems that ieee_arithmetic module is only processed for x86_64 target, and thus includes real(10) type, which is disabled for GPU target. Hence the errors.

[hakostra]

Thanks for looking into this.

It's not only about ieee_aritmethic, but all modules that contain a real(10). If I compile the HDF5 library with my LLVM clang/flang compiler, I get by default module files with real(10) inside. If I want to make a program that use OpenMP offloading, I am prevented from including those module file anywhere, even if I have no desire to use real(10) in any place in my program at all, and even if the HDF5 library is never used anywhere near the GPU... For HDF5 I have managed to make a patch to the build scripts to prevent real(10) support from being built, mitigating that problem locally on my development workstation right now, but the principle problem persists.

[hakostra]

I built a recent flang after PR #137828 was merged:

$ flang --version
flang version 22.0.0git (https://github.com/llvm/llvm-project.git 86fbaef99a53c5a0d3d5b96011797215296ec478)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /opt/llvm/llvm-project/install/bin

The error message is now:

$ flang -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda real10.F90
error: Semantic errors in real10.F90
./real10.F90:2:7: error: Cannot parse module file for module 'ieee_arithmetic': Source file 'ieee_arithmetic.mod' was not found
    use ieee_arithmetic
        ^^^^^^^^^^^^^^^

I cannot find any other module files for any other targets than x86_64 in my install directories. If it is of interest, this is how I configure LLVM:

cmake -GNinja \
    -DCMAKE_BUILD_TYPE="Release" \
    -DCMAKE_INSTALL_PREFIX="/opt/llvm/llvm-project/install" \
    -DCMAKE_CUDA_COMPILER=${CUDACXX} \
    -DCMAKE_CUDA_HOST_COMPILER="clang++-19" \
    -DCMAKE_CUDA_ARCHITECTURES="89;120" \
    -DLLVM_ENABLE_PROJECTS="clang;lld;flang;lldb" \
    -DLLVM_ENABLE_RUNTIMES="compiler-rt;openmp;offload;flang-rt" \
    -DLLVM_TARGETS_TO_BUILD="X86;NVPTX" \
    -DLLVM_ENABLE_RTTI="ON" \
    -DLLVM_RUNTIME_TARGETS='default;nvptx64-nvidia-cuda' \
    -DRUNTIMES_nvptx64-nvidia-cuda_LLVM_ENABLE_RUNTIMES=openmp \
    -DOPENMP_ENABLE_LIBOMPTARGET="ON" \
    -DFLANG_RT_EXPERIMENTAL_OFFLOAD_SUPPORT="OpenMP" \
    -DFLANG_RT_DEVICE_ARCHITECTURES="sm_89;sm_120" \
    -DFLANG_EXPERIMENTAL_OMP_OFFLOAD_BUILD="host_device" \
    ../llvm 2>&1 | tee cmake.log

I'm not sure if I miss something in my build or not. Any ideas?

[llvmbot]

@llvm/issue-subscribers-flang-frontend

Author: Håkon Strandenes (hakostra)

This issue is the same as issue #128015. I have a recent `flang` built from sources with OpenMP offload support:

$ flang --version
flang version 21.0.0git (https://github.com/llvm/llvm-project.git b68e8f1de71a65b21910aff4aea75fc1ca6bf6e0)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /opt/llvm/llvm-project/install/bin

I have found issue #128015 still to prevent targeting my Nvidia GPU (I have no AMD GPU to test with atm). The same small reproducer from PR #129742 still applies:

program real10
  use ieee_arithmetic
  implicit none
end program

Produces:

$ flang -fopenmp -fopenmp-targets=nvptx64 real10.F90
error: Semantic errors in real10.F90
/opt/llvm/llvm-project/install/bin/../include/flang/ieee_arithmetic.mod:143:1: error: REAL(KIND=10) is not an enabled type for this target
  real(10),intent(in)::x
  ^^^^^^^^^^^^^^^^^^^^^^
/opt/llvm/llvm-project/install/bin/../include/flang/ieee_arithmetic.mod:183:1: error: REAL(KIND=10) is not an enabled type for this target
  real(10),intent(in)::y

and so on for many, many lines. This error is also preventing me from using the HDF5 library in any projects with OpenMP, even if no HDF5 related functionality is ever touching the GPU...

If it makes any difference, my LLVM build is compiled with the following CMake command:

cmake -GNinja \
    -DCMAKE_BUILD_TYPE="Release" \
    -DCMAKE_INSTALL_PREFIX="/opt/llvm/llvm-project/install" \
    -DCMAKE_CUDA_COMPILER=${CUDACXX} \
    -DCMAKE_CUDA_HOST_COMPILER="clang++-19" \
    -DCMAKE_CUDA_ARCHITECTURES="89;120" \
    -DLLVM_ENABLE_PROJECTS="clang;lld;flang;lldb" \
    -DLLVM_ENABLE_RUNTIMES="compiler-rt;openmp;offload;flang-rt" \
    -DLLVM_TARGETS_TO_BUILD="X86;NVPTX" \
    -DLLVM_ENABLE_RTTI="ON" \
    -DOPENMP_ENABLE_LIBOMPTARGET="ON" \
    -DFLANG_RT_EXPERIMENTAL_OFFLOAD_SUPPORT="OpenMP" \
    -DFLANG_RT_DEVICE_ARCHITECTURES="sm_89;sm_120" \
    -DFLANG_EXPERIMENTAL_OMP_OFFLOAD_BUILD="host_device" \
    ../llvm