[llvm-branch-commits] [llvm] AMDGPU/GlobalISel: Use integer as MMO type for loads and stores lowering (PR #210096)

[Matt Arsenault via llvm-branch-commits]

================
@@ -144,7 +144,7 @@ static LegalizeMutation moreEltsToNext32Bit(unsigned TypeIdx) {
 static LegalizeMutation getScalarTypeFromMemDesc(unsigned TypeIdx) {
   return [=](const LegalityQuery &Query) {
     unsigned MemSize = Query.MMODescrs[0].MemoryTy.getSizeInBits();
----------------
arsenm wrote:

Most of the time we probably should be taking the raw MemoryTy? The use of this function is the abuse of the widen/narrow functions to legalize the memory type rather than just the register values like I want to fix 

https://github.com/llvm/llvm-project/pull/210096