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

[Jay Foad via llvm-commits]

jayfoad wrote:

> ### Why I Chose the Explicit Pattern Approach
> While ROCm's approach is more comprehensive, I believe the explicit pattern approach is better **for upstreaming** because:
> 
> 1. **Minimal Surface Area**: ~60 lines of TableGen patterns vs touching C++ pattern selection infrastructure
> 2. **Surgical Fix**: Targets exactly the broken case (GFX11 compute kernels) without affecting:
>    
>    * GFX12+ code generation
>    * Graphics shader paths
>    * Other VOP3P instruction selection
> 3. **Explicit Intent**: The high-priority patterns with comments make it immediately obvious what's being fixed and why
> 4. **Easier Review**: Reviewers can verify pattern matching logic directly in TableGen rather than tracing through C++ selection code
> 5. **Lower Risk**: Modifying `AMDGPUISelDAGToDAG.cpp` affects all VOP3P instruction selection; explicit patterns only affect GFX11 WMMA intrinsics

I feel like I'm dealing with an AI.

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