[llvm] [AMDGPU] Fix GFX11 WMMA intrinsic lowering regression for compute kernels (PR #164036)

[Jay Foad via llvm-commits]

jayfoad wrote:

> Add explicit high-priority (AddedComplexity=10000) patterns that match bare intrinsic calls directly without requiring VOP3PMods wrappers. These patterns provide default zero modifiers to the instruction format and override the broken patterns.

I'm not a big fan of this approach, because generally AddedComplexity should only be used as a cost model to prefer one pattern over another, it should not be used to fix correctness issues. "Broken" patterns should rather be disabled using predicates.

So maybe the ROCm-style fix is preferable? But I have not looked at it closely, and to be honest I do not understand the root cause of the problem you are fixing.

https://github.com/llvm/llvm-project/pull/164036