[llvm] [AMDGPU] Lower S_ABSDIFF_I32 to VALU instructions (PR #167691)

[Jay Foad via llvm-commits]

================
@@ -0,0 +1,35 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=amdgcn -mcpu=fiji -run-pass=si-fix-sgpr-copies -verify-machineinstrs -o - %s | FileCheck --check-prefix=GFX8 %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -run-pass=si-fix-sgpr-copies -verify-machineinstrs -o - %s | FileCheck --check-prefix=GFX12 %s
+
+---
+name:            absdiff_i32
+body:             |
+  bb.0:
+    ; GFX8-LABEL: name: absdiff_i32
+    ; GFX8: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GFX8-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GFX8-NEXT: [[DEF2:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GFX8-NEXT: [[V_LSHL_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_LSHL_ADD_U32_e64 [[DEF]], [[DEF1]], [[DEF2]], implicit $exec
+    ; GFX8-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 10
+    ; GFX8-NEXT: [[V_SUB_CO_U32_e32_:%[0-9]+]]:vgpr_32 = V_SUB_CO_U32_e32 [[S_MOV_B32_]], [[V_LSHL_ADD_U32_e64_]], implicit-def $vcc, implicit $exec
+    ; GFX8-NEXT: [[V_SUB_CO_U32_e32_1:%[0-9]+]]:vgpr_32 = V_SUB_CO_U32_e32 0, [[V_SUB_CO_U32_e32_]], implicit-def $vcc, implicit $exec
+    ; GFX8-NEXT: [[V_MAX_I32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_I32_e64 [[V_SUB_CO_U32_e32_]], [[V_SUB_CO_U32_e32_1]], implicit $exec
+    ;
+    ; GFX12-LABEL: name: absdiff_i32
+    ; GFX12: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GFX12-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GFX12-NEXT: [[DEF2:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GFX12-NEXT: [[V_LSHL_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_LSHL_ADD_U32_e64 [[DEF]], [[DEF1]], [[DEF2]], implicit $exec
+    ; GFX12-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 10
+    ; GFX12-NEXT: [[V_SUB_U32_e32_:%[0-9]+]]:vgpr_32 = V_SUB_U32_e32 [[S_MOV_B32_]], [[V_LSHL_ADD_U32_e64_]], implicit $exec
+    ; GFX12-NEXT: [[V_SUB_U32_e32_1:%[0-9]+]]:vgpr_32 = V_SUB_U32_e32 0, [[V_SUB_U32_e32_]], implicit $exec
+    ; GFX12-NEXT: [[V_MAX_I32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_I32_e64 [[V_SUB_U32_e32_]], [[V_SUB_U32_e32_1]], implicit $exec
+    %0:vgpr_32 = IMPLICIT_DEF
+    %1:vgpr_32 = IMPLICIT_DEF
+    %2:vgpr_32 = IMPLICIT_DEF
+    %3:vgpr_32 = V_LSHL_ADD_U32_e64 %0:vgpr_32, %1:vgpr_32, %2:vgpr_32, implicit $exec
----------------
jayfoad wrote:

I think you could probably remove this and just use a vgpr that is live-in to the basic block?

https://github.com/llvm/llvm-project/pull/167691