[llvm-branch-commits] [llvm] [AMDGPU][True16] Add legalization/selection handling for G_MERGE_VALUES of 2 s16 -> s32 (PR #200082)

[Domenic Nutile via llvm-branch-commits]

saxlungs wrote:

@petar-avramovic @kosarev @Sisyph @broxigarchen Context for this PR (and the stacked test PR): There's a change downstream to reorder the legalizer rules for G_MERGE_VALUES and G_UNMERGE_VALUES to avoid merging two s16s into an s32. When attempting to upstream this change, we discovered it causes lots of regressions. So, the plan is to revert that change downstream. However, there are some downstream tests that are broken by doing this.

This change instead implements a method of legalizing and selecting the G_MERGE_VALUES pseudo in this case. If this goes in, then the downstream change can be safely reverted.

https://github.com/llvm/llvm-project/pull/200082