[AMDGPU] Lower S_ABSDIFF_I32 to VALU instructions#167691
Merged
mariusz-sikora-at-amd merged 3 commits intollvm:mainfrom Nov 13, 2025
Merged
[AMDGPU] Lower S_ABSDIFF_I32 to VALU instructions#167691mariusz-sikora-at-amd merged 3 commits intollvm:mainfrom
mariusz-sikora-at-amd merged 3 commits intollvm:mainfrom
Conversation
Added support for lowering the scalar S_ABSDIFF_I32 instruction to equivalent VALU operations.
Member
|
@llvm/pr-subscribers-backend-amdgpu Author: Mariusz Sikora (mariusz-sikora-at-amd) ChangesAdded support for lowering the scalar S_ABSDIFF_I32 instruction to equivalent VALU operations. Full diff: https://github.com/llvm/llvm-project/pull/167691.diff 4 Files Affected:
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 6e39a2de9b805..f7c4fb2351b69 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -7775,6 +7775,11 @@ void SIInstrInfo::moveToVALUImpl(SIInstrWorklist &Worklist,
Inst.eraseFromParent();
return;
+ case AMDGPU::S_ABSDIFF_I32:
+ lowerScalarAbsDiff(Worklist, Inst);
+ Inst.eraseFromParent();
+ return;
+
case AMDGPU::S_CBRANCH_SCC0:
case AMDGPU::S_CBRANCH_SCC1: {
// Clear unused bits of vcc
@@ -8424,6 +8429,37 @@ void SIInstrInfo::lowerScalarAbs(SIInstrWorklist &Worklist,
addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
}
+void SIInstrInfo::lowerScalarAbsDiff(SIInstrWorklist &Worklist,
+ MachineInstr &Inst) const {
+ MachineBasicBlock &MBB = *Inst.getParent();
+ MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
+ MachineBasicBlock::iterator MII = Inst;
+ DebugLoc DL = Inst.getDebugLoc();
+
+ MachineOperand &Dest = Inst.getOperand(0);
+ MachineOperand &Src1 = Inst.getOperand(1);
+ MachineOperand &Src2 = Inst.getOperand(2);
+ Register SubResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+ Register TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+ Register ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+
+ unsigned SubOp =
+ ST.hasAddNoCarry() ? AMDGPU::V_SUB_U32_e32 : AMDGPU::V_SUB_CO_U32_e32;
+
+ BuildMI(MBB, MII, DL, get(SubOp), SubResultReg)
+ .addReg(Src1.getReg())
+ .addReg(Src2.getReg());
+
+ BuildMI(MBB, MII, DL, get(SubOp), TmpReg).addImm(0).addReg(SubResultReg);
+
+ BuildMI(MBB, MII, DL, get(AMDGPU::V_MAX_I32_e64), ResultReg)
+ .addReg(SubResultReg)
+ .addReg(TmpReg);
+
+ MRI.replaceRegWith(Dest.getReg(), ResultReg);
+ addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
+}
+
void SIInstrInfo::lowerScalarXnor(SIInstrWorklist &Worklist,
MachineInstr &Inst) const {
MachineBasicBlock &MBB = *Inst.getParent();
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.h b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
index 2ecd94186e1e0..d24dfd657ddcc 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
@@ -136,6 +136,8 @@ class SIInstrInfo final : public AMDGPUGenInstrInfo {
void lowerScalarAbs(SIInstrWorklist &Worklist, MachineInstr &Inst) const;
+ void lowerScalarAbsDiff(SIInstrWorklist &Worklist, MachineInstr &Inst) const;
+
void lowerScalarXnor(SIInstrWorklist &Worklist, MachineInstr &Inst) const;
void splitScalarNotBinop(SIInstrWorklist &Worklist, MachineInstr &Inst,
diff --git a/llvm/test/CodeGen/AMDGPU/absdiff.ll b/llvm/test/CodeGen/AMDGPU/absdiff.ll
index 9cb397fb9d1c6..5e4947ea5e0b5 100644
--- a/llvm/test/CodeGen/AMDGPU/absdiff.ll
+++ b/llvm/test/CodeGen/AMDGPU/absdiff.ll
@@ -1,6 +1,44 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck %s
+
+define amdgpu_gs float @absdiff_valu_input_regression() {
+; CHECK-LABEL: absdiff_valu_input_regression:
+; CHECK: ; %bb.0: ; %bb
+; CHECK-NEXT: s_mov_b32 s0, 0
+; CHECK-NEXT: .LBB0_1: ; %bb1
+; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: s_mov_b32 s1, s0
+; CHECK-NEXT: s_or_b32 s0, s0, 1
+; CHECK-NEXT: s_cmp_gt_i32 s1, 0
+; CHECK-NEXT: s_cbranch_scc1 .LBB0_1
+; CHECK-NEXT: ; %bb.2: ; %bb11
+; CHECK-NEXT: v_med3_i32 v0, s1, 0, 1
+; CHECK-NEXT: v_sub_u32_e32 v0, 0, v0
+; CHECK-NEXT: v_sub_u32_e32 v1, 0, v0
+; CHECK-NEXT: v_max_i32_e32 v0, v0, v1
+; CHECK-NEXT: ; return to shader part epilog
+bb:
+ br label %bb1
+
+bb1: ; preds = %bb1, %bb
+ %i = phi i32 [ 0, %bb ], [ %i9, %bb1 ]
+ %i2 = phi i32 [ 0, %bb ], [ %i5, %bb1 ]
+ %i3 = or i32 %i2, 1
+ %i4 = or i32 %i3, 0
+ %i5 = call i32 @llvm.smax.i32(i32 %i, i32 0)
+ %i6 = call i32 @llvm.umin.i32(i32 %i5, i32 1)
+ %i7 = sub i32 0, %i6
+ %i8 = call i32 @llvm.abs.i32(i32 %i7, i1 false)
+ %i9 = or i32 %i, 1
+ %i10 = icmp sgt i32 %i, 0
+ br i1 %i10, label %bb1, label %bb11
+
+bb11: ; preds = %bb1
+ %i12 = bitcast i32 %i8 to float
+ ret float %i12
+}
+
define amdgpu_ps i16 @absdiff_i16_false(i16 inreg %arg0, i16 inreg %arg1) {
; CHECK-LABEL: absdiff_i16_false:
; CHECK: ; %bb.0:
diff --git a/llvm/test/CodeGen/AMDGPU/move-to-valu-absdiff.mir b/llvm/test/CodeGen/AMDGPU/move-to-valu-absdiff.mir
new file mode 100644
index 0000000000000..f4c99881170d1
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/move-to-valu-absdiff.mir
@@ -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
+ %4:sreg_32 = COPY %3:vgpr_32
+ %5:sreg_32 = S_MOV_B32 10
+ %6:sreg_32 = S_ABSDIFF_I32 killed %5:sreg_32, %4:sreg_32, implicit-def dead $scc
+...
|
jayfoad
reviewed
Nov 12, 2025
jayfoad
reviewed
Nov 12, 2025
Comment on lines
+17
to
+19
| ; CHECK-NEXT: v_sub_u32_e32 v0, 0, v0 | ||
| ; CHECK-NEXT: v_sub_u32_e32 v1, 0, v0 | ||
| ; CHECK-NEXT: v_max_i32_e32 v0, v0, v1 |
Contributor
There was a problem hiding this comment.
This is a confusing example because I think it is calculating absdiff(0, x) which should really have been simplified to abs(x) already. But I think the code is correct.
Contributor
Author
There was a problem hiding this comment.
This example is a reduce version of CTS shader that started to fail after #164835
Contributor
There was a problem hiding this comment.
I think instcombine is missing a lot of patterns for these newer intrinsics
jayfoad
approved these changes
Nov 12, 2025
Contributor
jayfoad
left a comment
jayfoad
left a comment
There was a problem hiding this comment.
Looks OK to me. Maybe wait a day in case there are other comments.
Contributor
|
LGTM |
arsenm
reviewed
Nov 12, 2025
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Added support for lowering the scalar S_ABSDIFF_I32 instruction to equivalent VALU operations.