[llvm] AMDGPU: Fix temporal divergence introduced by machine-sink and performance regression introduced by D155343 (PR #67456)

[via llvm-commits]

================
@@ -860,9 +861,9 @@ body: |
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   [[PHI4:%[0-9]+]]:vgpr_32 = PHI %95, %bb.9
   ; CHECK-NEXT:   SI_END_CF %96, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
-  ; CHECK-NEXT:   [[S_ADD_I32_2:%[0-9]+]]:sreg_32 = S_ADD_I32 %108, 1, implicit-def dead $scc
-  ; CHECK-NEXT:   [[S_ADD_I32_3:%[0-9]+]]:sreg_32 = S_ADD_I32 [[S_ADD_I32_2]], 2, implicit-def dead $scc
-  ; CHECK-NEXT:   [[V_ADD_U32_e64_3:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[V_OR3_B32_e64_]], [[S_ADD_I32_3]], 0, implicit $exec
+  ; CHECK-NEXT:   [[V_ADD_U32_e64_3:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 %150, 1, 0, implicit $exec
+  ; CHECK-NEXT:   [[V_ADD_U32_e64_4:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[V_ADD_U32_e64_3]], 2, 0, implicit $exec
+  ; CHECK-NEXT:   [[V_ADD_U32_e64_5:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[V_OR3_B32_e64_]], [[V_ADD_U32_e64_4]], 0, implicit $exec
----------------
petar-avramovic wrote:

moveToVALU

https://github.com/llvm/llvm-project/pull/67456