[RFC] Names for Flang-RT libraries

Subprojects Flang
1

Petr Hosek started a discussion on the naming of the flang runtime libraries.

Since it was deeply embedded into the PR, probably not many have seen it. I am opening the RFC for more visibility.

The current library names are (before #122341):

  • libFortranRuntime{.a/.so}
  • libFortranFloat128Math.a (static-only by design)
  • libCufRuntime_cuda_${CUDAToolkit_VERSION_MAJOR}{.a/.so} (e.g libCufRuntime_cuda_12.so)

After #122341, the names will be:

  • libflang_rt{.a/.so}
  • libFortranFloat128Math.a
  • libCufRuntime_cuda_${CUDAToolkit_VERSION_MAJOR}{.a/.so}.

The proposal is to use names:

  • libflang_rt.runtime{.a/.so}
  • libflang_rt.quadmath.a
  • libflang_rt.cuda_${CUDAToolkit_VERSION_MAJOR}{.a/.so}.

The idea is consistency and to resemble more the pattern used by compiler-rt. Without concensus, I will continue with the names from #122341.

1 Like
2

For my two cents, I like the suffixed names best. Seems clearer and better aligned with our sister projects.

1 Like
3

I also tend to prefer the names in this proposal so it is more consistent. At least ok for me for the cuda part.

4

And you would still insist on the libflang_rt.cuda_${CUDAToolkit_VERSION_MAJOR} sources to reside in a subdirectory of “flang_rt”? That is, the following directory layout?

  • lib/runtime and include/flang-rt/runtime for libflang_rt.runtime{.a/.so}
  • lib/quadmath for libflang_rt.quadmath.a
  • lib/runtime/CUDA and include/flang-rt/runtime/CUDA for libflang_rt.cuda_${CUDAToolkit_VERSION_MAJOR}{.a/.so}

Whouldn’t it be more consistent to do the following?

  • lib/runtime and include/flang-rt/runtime for libflang_rt.runtime{.a/.so}
  • lib/quadmath for libflang_rt.quadmath{.a/.so}
  • lib/cuda and include/flang-rt/cuda for libflang_rt.cuda_${CUDAToolkit_VERSION_MAJOR}{.a/.so}
5

I’m ok to move them into lib/cuda.

1 Like
6

+1 for consistency with compiler_rt since flang shares a lot of the same driver code. The CUDA libraries should probably go somewhere special, long-term I’d like to use standard paths for the device build with the GPU triple.

7

Since there were only agreements to the idea, I will make that change.