This is an archive of the discontinued LLVM Phabricator instance.

[AMDGPU][GFX11] Add test coverage for 16-bit conversions, part 12.
ClosedPublic

Authored by kosarev on Jun 14 2023, 5:42 AM.

Diff Detail

Event Timeline

kosarev created this revision.Jun 14 2023, 5:42 AM
Herald added a project: Restricted Project. · View Herald TranscriptJun 14 2023, 5:42 AM
kosarev requested review of this revision.Jun 14 2023, 5:42 AM
Herald added a project: Restricted Project. · View Herald TranscriptJun 14 2023, 5:42 AM
Joe_Nash added inline comments.Jun 19 2023, 9:17 AM
llvm/test/CodeGen/AMDGPU/function-args.ll
89

The changes to llvm.amdgcn.raw.buffer.store.ll and llvm.amdgcn.struct.buffer.store.ll LGTM.

The changes to function-args.ll, I'm not so sure about. There are a lot of not lines. I don't know what properties these are ensuring, but I know something was being checked. When we remove those and go to auto-generated, there is no evidence some property was being checked. A use of v0 in this case.

kosarev added inline comments.Jun 23 2023, 5:11 AM
llvm/test/CodeGen/AMDGPU/function-args.ll
89

From what I see the -NOT lines just make sure there are no redefinitions or extra uses of the registers of interest and no extra s_waitcnt instructions, deviations from which in all cases should be easy to catch by the usual examination of changes in the autogenerated checks.

arsenm accepted this revision.Jun 23 2023, 5:20 AM

Aggressive -NEXT checks are usually safer then -NOT anyway. I think this was mostly looking for accidental clobbers.

This revision is now accepted and ready to land.Jun 23 2023, 5:20 AM