llvm-project

12b02727 - clang/AMDGPU: Respect __launch_bounds__ attribute (#215615)

clang/AMDGPU: Respect __launch_bounds__ attribute (#215615) Currently the HIP headers manually implement this with a macro setting amdgpu attributes, and the proper clang attribute is silently ignored. Directly map the proper attribute into the target IR attributes. The first argument sets "amdgpu-flat-work-group-size" and the second (reinterpreted by HIP as minimum waves per EU) sets "amdgpu-waves-per-eu". An explicit amdgpu_flat_work_group_size / amdgpu_waves_per_eu attribute takes precedence. This matches the launch_bounds macro in the HIP headers, which can now be dropped. The 3rd maxclusterrank argument is only handled for NVPTX, so restrict the sm_90 arch check to NVPTX targets and ignore the third argument on other targets. Fixes #91468 Co-authored-by: Claude (Claude-Opus-4.8) <noreply@anthropic.com>