[llvm] [AMDGPU] Select trunc(srl x, 16) as hi16 subregister on true16 (PR #221960)

[Guo Chen via llvm-commits]

broxigarchen wrote:

Thanks. There are some known issues on `extract_subreg hi16` that even with the `copy_to_regclass src, vgpr_32`, the `extract_subreg hi16` could still return a sgpr32_hi16 if the src node is uniform, and this would cause problem in the later pipeline.

My understanding is that this cannot get in by itself without some tweak on the regclass selection on extract_subreg uniform node. Can you also try to add a uniform test case for this pattern and see if this causes failure in the pipeline?

https://github.com/llvm/llvm-project/pull/221960