[llvm] AMDGPU/GlobalISel: Switch to extended LLTs (PR #196522)

[Chinmay Deshpande via llvm-commits]

chinmaydd wrote:

Hi Petar, sorry for the delayed response. Thinking more about `s32` vs `i32/f32` mixing feels like going down the wrong road. The incremental path makes more sense to me.

We can land the AMDGPU-only changes and mark the affected tests `XFAIL: *` in the interim, then bring the fixes back one group at a time (similar in spirit to how `new-regbankselect` was enabled incrementally ?). For the common-code changes that are "correct in all cases," two options:

1. pre-commit a small AArch64 test that exercises the path (AArch64 already has extended LLTs on), then land the change; or
2. if the path isn't reachable from AArch64, cover it by un-XFAIL'ing the relevant AMDGPU test in the same change.

The tradeoff with (1) is it may be slower if it needs AArch64 reviewer time. But I'm happy to help with the re-enablement work wherever needed.

https://github.com/llvm/llvm-project/pull/196522