[llvm] 022fdfa - AMDGPU: Mark the carry-out of the scalar abs expansion dead (#226046)

[via llvm-commits]

Author: Matt Arsenault
Date: 2026-09-24T10:37:43+02:00
New Revision: 022fdfa1e5a0b85dc785d30b17be3880768b22dc

URL: https://github.com/llvm/llvm-project/commit/022fdfa1e5a0b85dc785d30b17be3880768b22dc
DIFF: https://github.com/llvm/llvm-project/commit/022fdfa1e5a0b85dc785d30b17be3880768b22dc.diff

LOG: AMDGPU: Mark the carry-out of the scalar abs expansion dead (#226046)

Added: 
    

Modified: 
    llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
    llvm/test/CodeGen/AMDGPU/move-to-valu-absdiff.mir

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index b042a7c164a37..16ee5f5698680 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -9119,12 +9119,14 @@ void SIInstrInfo::lowerScalarAbs(SIInstrWorklist &Worklist,
   Register TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
   Register ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
 
-  unsigned SubOp = ST.hasAddNoCarryInsts() ? AMDGPU::V_SUB_U32_e32
-                                           : AMDGPU::V_SUB_CO_U32_e32;
+  bool HasCarryOut = !ST.hasAddNoCarryInsts();
+  unsigned SubOp =
+      HasCarryOut ? AMDGPU::V_SUB_CO_U32_e32 : AMDGPU::V_SUB_U32_e32;
 
-  BuildMI(MBB, MII, DL, get(SubOp), TmpReg)
-    .addImm(0)
-    .addReg(Src.getReg());
+  MachineInstrBuilder Sub =
+      BuildMI(MBB, MII, DL, get(SubOp), TmpReg).addImm(0).addReg(Src.getReg());
+  if (HasCarryOut)
+    Sub.setOperandDead(3); // Dead vcc
 
   BuildMI(MBB, MII, DL, get(AMDGPU::V_MAX_I32_e64), ResultReg)
     .addReg(Src.getReg())
@@ -9148,14 +9150,21 @@ void SIInstrInfo::lowerScalarAbsDiff(SIInstrWorklist &Worklist,
   Register TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
   Register ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
 
-  unsigned SubOp = ST.hasAddNoCarryInsts() ? AMDGPU::V_SUB_U32_e32
-                                           : AMDGPU::V_SUB_CO_U32_e32;
+  bool HasCarryOut = !ST.hasAddNoCarryInsts();
+  unsigned SubOp =
+      HasCarryOut ? AMDGPU::V_SUB_CO_U32_e32 : AMDGPU::V_SUB_U32_e32;
 
-  BuildMI(MBB, MII, DL, get(SubOp), SubResultReg)
-      .addReg(Src1.getReg())
-      .addReg(Src2.getReg());
+  MachineInstrBuilder Sub1 = BuildMI(MBB, MII, DL, get(SubOp), SubResultReg)
+                                 .addReg(Src1.getReg())
+                                 .addReg(Src2.getReg());
 
-  BuildMI(MBB, MII, DL, get(SubOp), TmpReg).addImm(0).addReg(SubResultReg);
+  MachineInstrBuilder Sub2 =
+      BuildMI(MBB, MII, DL, get(SubOp), TmpReg).addImm(0).addReg(SubResultReg);
+
+  if (HasCarryOut) {
+    Sub1.setOperandDead(3); // Dead vcc
+    Sub2.setOperandDead(3); // Dead vcc
+  }
 
   BuildMI(MBB, MII, DL, get(AMDGPU::V_MAX_I32_e64), ResultReg)
       .addReg(SubResultReg)

diff  --git a/llvm/test/CodeGen/AMDGPU/move-to-valu-abs
diff .mir b/llvm/test/CodeGen/AMDGPU/move-to-valu-abs
diff .mir
index a73d3a2c2d1e2..39a7948600d98 100644
--- a/llvm/test/CodeGen/AMDGPU/move-to-valu-abs
diff .mir
+++ b/llvm/test/CodeGen/AMDGPU/move-to-valu-abs
diff .mir
@@ -12,8 +12,8 @@ body:             |
     ; GFX8-NEXT: {{  $}}
     ; GFX8-NEXT: [[V_LSHL_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_LSHL_ADD_U32_e64 $vgpr0, $vgpr1, $vgpr2, 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_SUB_CO_U32_e32_:%[0-9]+]]:vgpr_32 = V_SUB_CO_U32_e32 [[S_MOV_B32_]], [[V_LSHL_ADD_U32_e64_]], implicit-def dead $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 dead $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: abs
diff _i32
@@ -29,3 +29,25 @@ body:             |
     %2:sreg_32 = S_MOV_B32 10
     %3:sreg_32 = S_ABSDIFF_I32 killed %2:sreg_32, %1:sreg_32, implicit-def dead $scc
 ...
+---
+name:            abs_i32
+body:             |
+  bb.0:
+    liveins: $vgpr0, $vgpr1, $vgpr2
+    ; GFX8-LABEL: name: abs_i32
+    ; GFX8: liveins: $vgpr0, $vgpr1, $vgpr2
+    ; GFX8-NEXT: {{  $}}
+    ; GFX8-NEXT: [[V_LSHL_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_LSHL_ADD_U32_e64 $vgpr0, $vgpr1, $vgpr2, implicit $exec
+    ; GFX8-NEXT: [[V_SUB_CO_U32_e32_:%[0-9]+]]:vgpr_32 = V_SUB_CO_U32_e32 0, [[V_LSHL_ADD_U32_e64_]], implicit-def dead $vcc, implicit $exec
+    ; GFX8-NEXT: [[V_MAX_I32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_I32_e64 [[V_LSHL_ADD_U32_e64_]], [[V_SUB_CO_U32_e32_]], implicit $exec
+    ;
+    ; GFX12-LABEL: name: abs_i32
+    ; GFX12: liveins: $vgpr0, $vgpr1, $vgpr2
+    ; GFX12-NEXT: {{  $}}
+    ; GFX12-NEXT: [[V_LSHL_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_LSHL_ADD_U32_e64 $vgpr0, $vgpr1, $vgpr2, implicit $exec
+    ; GFX12-NEXT: [[V_SUB_U32_e32_:%[0-9]+]]:vgpr_32 = V_SUB_U32_e32 0, [[V_LSHL_ADD_U32_e64_]], implicit $exec
+    ; GFX12-NEXT: [[V_MAX_I32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_I32_e64 [[V_LSHL_ADD_U32_e64_]], [[V_SUB_U32_e32_]], implicit $exec
+    %0:vgpr_32 = V_LSHL_ADD_U32_e64 $vgpr0, $vgpr1, $vgpr2, implicit $exec
+    %1:sreg_32 = COPY %0:vgpr_32
+    %2:sreg_32 = S_ABS_I32 %1:sreg_32, implicit-def dead $scc
+...