[llvm] AMDGPU: Fix using illegal VOP3 literal in frame index elimination (PR #115747)

[Jay Foad via llvm-commits]

================
@@ -20,8 +20,9 @@ define amdgpu_kernel void @local_stack_offset_uses_sp(ptr addrspace(1) %out) {
 ; MUBUF-LABEL: local_stack_offset_uses_sp:
 ; MUBUF:       ; %bb.0: ; %entry
 ; MUBUF-NEXT:    s_add_u32 s0, s0, s17
+; MUBUF-NEXT:    v_mov_b32_e32 v1, 0x3000
 ; MUBUF-NEXT:    s_addc_u32 s1, s1, 0
-; MUBUF-NEXT:    v_mov_b32_e32 v0, 0x3040
+; MUBUF-NEXT:    v_add_u32_e32 v0, 64, v1
----------------
jayfoad wrote:

Would your "TODO: Handle VOP3->VOP2 shrink to support the fold" fix this regression?

https://github.com/llvm/llvm-project/pull/115747