[llvm] [AMDGPU] Fold frame offset into immediate operand when possible (PR #220521)

[Arseniy Obolenskiy via llvm-commits]

================
@@ -885,9 +885,8 @@ body:             |
     ; MUBUFW64-NEXT: frame-setup CFI_INSTRUCTION llvm_def_aspace_cfa $sgpr32, 0, 6
     ; MUBUFW64-NEXT: frame-setup CFI_INSTRUCTION llvm_register_pair $pc_reg, $sgpr30, 32, $sgpr31, 32
     ; MUBUFW64-NEXT: frame-setup CFI_INSTRUCTION undefined $sgpr7
-    ; MUBUFW64-NEXT: renamable $sgpr7 = S_LSHR_B32 $sgpr32, 6, implicit-def dead $scc
-    ; MUBUFW64-NEXT: renamable $sgpr7 = S_ADD_I32 killed $sgpr7, $sgpr8, implicit-def $scc
-    ; MUBUFW64-NEXT: renamable $sgpr7 = S_ADD_I32 killed renamable $sgpr7, 0, implicit-def $scc
+    ; MUBUFW64-NEXT: $sgpr4 = S_LSHR_B32 $sgpr32, 6, implicit-def dead $scc
----------------
aobolensk wrote:

It seems that only the 32-bit wrap check was unreached, I added some cases folding a wrapping immediate into a nonzero frame offset

https://github.com/llvm/llvm-project/pull/220521