[PATCH] D128090: [Clang][OpenMP] Process multi-arch compilation options given via -march

Wed Jul 13 11:02:56 PDT 2022

jhuber6 added a comment.

In D128090#3648984 <https://reviews.llvm.org/D128090#3648984>, @tra wrote:

> At some point we should start consolidating the ways we can specify an offload target and try to avoid adding new ones until then.

Agreed, that was my intention with making `--offload-arch` work for everything the same way in the new driver. The difference with CUDA / HIP and OpenMP right now is  the default behavior if nothing was given. For CUDA / HIP we just default the bound architecture to something like sm_35 and gfx803 I believe. For OpenMP we keep the bound architecture empty which signals us to check the value of `-march=` and use that if present, or default to something more intelligent. Right now there's `CLANG_OPENMP_NVPTX_DEFAULT_ARCH`, which is defined by CMake to be the architecture of the system used to build clang, and `amdgpu-arch` which is just a program that runs at compile time. I'm not sure if there would be a desire to make CUDA / HIP adhere to this as well with the new driver.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D128090/new/

https://reviews.llvm.org/D128090