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

[Luis Chamberlain via llvm-commits]

mcgrof wrote:

> I tried to test your commit but it looks to me it came from some downstream llvm branch, it fails ninja check (small difference when I regenerated them to pass). Also tried with and without your change and I get same result. For the pattern itself it is probably best to use WMMAPat.
> 
> Can you re-check if your tests fails with cannot select on llvm::main.

Jeesh, I suspect the issue was the assumptions on requirements I was using, I suspect now the size validation is all I need to modify, instead of (8,8,8,8) I should use (16,16,8,8). That would explain why I was not getting the correct constraints generated. I'll run some more tests.

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