[llvm] [AMDGPU] Update EXECZ retention in SIPreEmitPeephole for GFX11/12 (PR #97676)
Carl Ritson via llvm-commits
llvm-commits at lists.llvm.org
Thu Jul 4 02:58:08 PDT 2024
perlfu wrote:
> > Also add code to retain uniform jumps over barrier instructions.
>
> This seems like it should be a separate patch. Is it required for correctness? Why hasn't it caused problems before?
Can separate these. I don't think we commonly generate code that jumps a barrier, because it would create potential problems if some waves did not follow the same decisions across a workgroup. And/or our blocks with s_barrier are usually above the threshold.
But I think the risk of generating something where this matters maybe higher with newer barrier operations on GFX12.
https://github.com/llvm/llvm-project/pull/97676
More information about the llvm-commits
mailing list