[llvm] [SelectionDAG] Improve isGuaranteedNotToBeUndefOrPoison (PR #147019)
Björn Pettersson via llvm-commits
llvm-commits at lists.llvm.org
Fri Jul 4 01:56:28 PDT 2025
https://github.com/bjope created https://github.com/llvm/llvm-project/pull/147019
Add special handling of EXTRACT_SUBVECTOR, INSERT_SUBVECTOR, EXTRACT_VECTOR_ELT, INSERT_VECTOR_ELT and SCALAR_TO_VECTOR in isGuaranteedNotToBeUndefOrPoison. Make use of DemandedElts to improve the analysis and only check relevant elements for each operand.
Also start using DemandedElts in the recursive calls that check isGuaranteedNotToBeUndefOrPoison for all operands for operations that do not create undef/poison. We can do that for a number of elementwise operations for which the DemandedElts can be applied to every operand (e.g. ADD, OR, BITREVERSE, TRUNCATE).
>From 047637140f4348939f56c87274d82b597bcc6b79 Mon Sep 17 00:00:00 2001
From: Bjorn Pettersson <bjorn.a.pettersson at ericsson.com>
Date: Tue, 3 Jun 2025 10:01:01 +0200
Subject: [PATCH] [SelectionDAG] Improve isGuaranteedNotToBeUndefOrPoison
Add special handling of EXTRACT_SUBVECTOR, INSERT_SUBVECTOR,
EXTRACT_VECTOR_ELT, INSERT_VECTOR_ELT and SCALAR_TO_VECTOR in
isGuaranteedNotToBeUndefOrPoison. Make use of DemandedElts to
improve the analysis and only check relevant elements for each
operand.
Also start using DemandedElts in the recursive calls that
check isGuaranteedNotToBeUndefOrPoison for all operands for
operations that do not create undef/poison. We can do that for
a number of elementwise operations for which the DemandedElts
can be applied to every operand (e.g. ADD, OR, BITREVERSE, TRUNCATE).
---
llvm/include/llvm/CodeGen/SelectionDAGNodes.h | 6 +
.../lib/CodeGen/SelectionDAG/SelectionDAG.cpp | 142 ++++++++++++++++++
.../X86/div-rem-pair-recomposition-signed.ll | 31 ++--
llvm/test/CodeGen/X86/pr62286.ll | 14 +-
4 files changed, 171 insertions(+), 22 deletions(-)
diff --git a/llvm/include/llvm/CodeGen/SelectionDAGNodes.h b/llvm/include/llvm/CodeGen/SelectionDAGNodes.h
index a3675eecfea3f..08db31c63367d 100644
--- a/llvm/include/llvm/CodeGen/SelectionDAGNodes.h
+++ b/llvm/include/llvm/CodeGen/SelectionDAGNodes.h
@@ -1889,6 +1889,12 @@ LLVM_ABI SDValue peekThroughExtractSubvectors(SDValue V);
/// If \p V is not a truncation, it is returned as-is.
LLVM_ABI SDValue peekThroughTruncates(SDValue V);
+/// Recursively peek through INSERT_VECTOR_ELT nodes, returning the source
+/// vector operand of \p V, as long as \p V is an INSERT_VECTOR_ELT operation
+/// that do not insert into any of the demanded vector elts.
+LLVM_ABI SDValue peekThroughInsertVectorElt(SDValue V,
+ const APInt &DemandedElts);
+
/// Returns true if \p V is a bitwise not operation. Assumes that an all ones
/// constant is canonicalized to be operand 1.
LLVM_ABI bool isBitwiseNot(SDValue V, bool AllowUndefs = false);
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 6df21b624137f..ebeda8d86e3c9 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -5456,6 +5456,85 @@ bool SelectionDAG::isGuaranteedNotToBeUndefOrPoison(SDValue Op,
}
return true;
+ case ISD::EXTRACT_SUBVECTOR: {
+ SDValue Src = Op.getOperand(0);
+ if (Src.getValueType().isScalableVector())
+ break;
+ uint64_t Idx = Op.getConstantOperandVal(1);
+ unsigned NumSrcElts = Src.getValueType().getVectorNumElements();
+ APInt DemandedSrcElts = DemandedElts.zext(NumSrcElts).shl(Idx);
+ return isGuaranteedNotToBeUndefOrPoison(Src, DemandedSrcElts, PoisonOnly,
+ Depth + 1);
+ }
+
+ case ISD::INSERT_SUBVECTOR: {
+ if (Op.getValueType().isScalableVector())
+ break;
+ SDValue Src = Op.getOperand(0);
+ SDValue Sub = Op.getOperand(1);
+ uint64_t Idx = Op.getConstantOperandVal(2);
+ unsigned NumSubElts = Sub.getValueType().getVectorNumElements();
+ APInt DemandedSubElts = DemandedElts.extractBits(NumSubElts, Idx);
+ APInt DemandedSrcElts = DemandedElts;
+ DemandedSrcElts.clearBits(Idx, Idx + NumSubElts);
+
+ if (!!DemandedSubElts && !isGuaranteedNotToBeUndefOrPoison(
+ Sub, DemandedSubElts, PoisonOnly, Depth + 1))
+ return false;
+ if (!!DemandedSrcElts && !isGuaranteedNotToBeUndefOrPoison(
+ Src, DemandedSrcElts, PoisonOnly, Depth + 1))
+ return false;
+ return true;
+ }
+
+ case ISD::EXTRACT_VECTOR_ELT: {
+ SDValue Src = Op.getOperand(0);
+ auto *IndexC = dyn_cast<ConstantSDNode>(Op.getOperand(1));
+ EVT SrcVT = Src.getValueType();
+ if (SrcVT.isFixedLengthVector() && IndexC &&
+ IndexC->getAPIntValue().ult(SrcVT.getVectorNumElements())) {
+ APInt DemandedSrcElts = APInt::getOneBitSet(SrcVT.getVectorNumElements(),
+ IndexC->getZExtValue());
+ return isGuaranteedNotToBeUndefOrPoison(Src, DemandedSrcElts, PoisonOnly,
+ Depth + 1);
+ }
+ break;
+ }
+
+ case ISD::INSERT_VECTOR_ELT: {
+ SDValue InVec = Op.getOperand(0);
+ SDValue InVal = Op.getOperand(1);
+ SDValue EltNo = Op.getOperand(2);
+ EVT VT = InVec.getValueType();
+ auto *IndexC = dyn_cast<ConstantSDNode>(EltNo);
+ if (IndexC && VT.isFixedLengthVector() &&
+ IndexC->getZExtValue() < VT.getVectorNumElements()) {
+ if (DemandedElts[IndexC->getZExtValue()] &&
+ !isGuaranteedNotToBeUndefOrPoison(InVal, PoisonOnly, Depth + 1))
+ return false;
+ APInt InVecDemandedElts = DemandedElts;
+ InVecDemandedElts.clearBit(IndexC->getZExtValue());
+ if (!!InVecDemandedElts &&
+ !isGuaranteedNotToBeUndefOrPoison(
+ peekThroughInsertVectorElt(InVec, InVecDemandedElts),
+ InVecDemandedElts, PoisonOnly, Depth + 1))
+ return false;
+ return true;
+ }
+ break;
+ }
+
+ case ISD::SCALAR_TO_VECTOR:
+ // Check upper (known undef) elements.
+ if (DemandedElts.ugt(1) && !PoisonOnly)
+ return false;
+ // Check element zero.
+ if (DemandedElts[0] && !isGuaranteedNotToBeUndefOrPoison(Op.getOperand(0),
+ PoisonOnly,
+ Depth + 1))
+ return false;
+ return true;
+
case ISD::SPLAT_VECTOR:
return isGuaranteedNotToBeUndefOrPoison(Op.getOperand(0), PoisonOnly,
Depth + 1);
@@ -5478,6 +5557,52 @@ bool SelectionDAG::isGuaranteedNotToBeUndefOrPoison(SDValue Op,
return true;
}
+ case ISD::SHL:
+ case ISD::SRL:
+ case ISD::SRA:
+ // Shift amount operand is checked by canCreateUndefOrPoison. So it is
+ // enough to check operand 0 if Op can't create undef/poison.
+ return !canCreateUndefOrPoison(Op, DemandedElts, PoisonOnly,
+ /*ConsiderFlags*/ true, Depth) &&
+ isGuaranteedNotToBeUndefOrPoison(Op.getOperand(0), DemandedElts,
+ PoisonOnly, Depth + 1);
+
+ case ISD::BSWAP:
+ case ISD::CTPOP:
+ case ISD::BITREVERSE:
+ case ISD::AND:
+ case ISD::OR:
+ case ISD::XOR:
+ case ISD::ADD:
+ case ISD::SUB:
+ case ISD::MUL:
+ case ISD::SADDSAT:
+ case ISD::UADDSAT:
+ case ISD::SSUBSAT:
+ case ISD::USUBSAT:
+ case ISD::SSHLSAT:
+ case ISD::USHLSAT:
+ case ISD::SMIN:
+ case ISD::SMAX:
+ case ISD::UMIN:
+ case ISD::UMAX:
+ case ISD::ZERO_EXTEND:
+ case ISD::SIGN_EXTEND:
+ case ISD::ANY_EXTEND:
+ case ISD::TRUNCATE:
+ case ISD::VSELECT: {
+ // If Op can't create undef/poison and none of its operands are undef/poison
+ // then Op is never undef/poison. A difference from the more common check
+ // below, outside the switch, is that we handle elementwise operations for
+ // which the DemandedElts mask is valid for all operands here.
+ return !canCreateUndefOrPoison(Op, DemandedElts, PoisonOnly,
+ /*ConsiderFlags*/ true, Depth) &&
+ all_of(Op->ops(), [&](SDValue V) {
+ return isGuaranteedNotToBeUndefOrPoison(V, DemandedElts,
+ PoisonOnly, Depth + 1);
+ });
+ }
+
// TODO: Search for noundef attributes from library functions.
// TODO: Pointers dereferenced by ISD::LOAD/STORE ops are noundef.
@@ -12463,6 +12588,23 @@ SDValue llvm::peekThroughTruncates(SDValue V) {
return V;
}
+SDValue llvm::peekThroughInsertVectorElt(SDValue V, const APInt &DemandedElts) {
+ while (V.getOpcode() == ISD::INSERT_VECTOR_ELT) {
+ SDValue InVec = V.getOperand(0);
+ SDValue EltNo = V.getOperand(2);
+ EVT VT = InVec.getValueType();
+ auto *IndexC = dyn_cast<ConstantSDNode>(EltNo);
+ if (IndexC && VT.isFixedLengthVector() &&
+ IndexC->getZExtValue() < VT.getVectorNumElements() &&
+ !DemandedElts[IndexC->getZExtValue()]) {
+ V = InVec;
+ continue;
+ }
+ break;
+ }
+ return V;
+}
+
bool llvm::isBitwiseNot(SDValue V, bool AllowUndefs) {
if (V.getOpcode() != ISD::XOR)
return false;
diff --git a/llvm/test/CodeGen/X86/div-rem-pair-recomposition-signed.ll b/llvm/test/CodeGen/X86/div-rem-pair-recomposition-signed.ll
index d869f8ec01a5a..4586fce615487 100644
--- a/llvm/test/CodeGen/X86/div-rem-pair-recomposition-signed.ll
+++ b/llvm/test/CodeGen/X86/div-rem-pair-recomposition-signed.ll
@@ -172,10 +172,9 @@ define i128 @scalar_i128(i128 %x, i128 %y, ptr %divdst) nounwind {
; X86-NEXT: movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
; X86-NEXT: sbbl %eax, %edi
; X86-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
-; X86-NEXT: movl 40(%ebp), %ecx
-; X86-NEXT: movl %ecx, %edx
+; X86-NEXT: movl 40(%ebp), %esi
+; X86-NEXT: movl %esi, %edx
; X86-NEXT: sarl $31, %edx
-; X86-NEXT: movl %ecx, %esi
; X86-NEXT: xorl %edx, %esi
; X86-NEXT: movl 36(%ebp), %ecx
; X86-NEXT: xorl %edx, %ecx
@@ -204,14 +203,14 @@ define i128 @scalar_i128(i128 %x, i128 %y, ptr %divdst) nounwind {
; X86-NEXT: sete %al
; X86-NEXT: orb %cl, %al
; X86-NEXT: movb %al, {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Spill
-; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
-; X86-NEXT: bsrl %eax, %edx
+; X86-NEXT: bsrl %esi, %edx
; X86-NEXT: xorl $31, %edx
-; X86-NEXT: addl $32, %edx
-; X86-NEXT: bsrl %esi, %ecx
+; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
+; X86-NEXT: bsrl %eax, %ecx
; X86-NEXT: xorl $31, %ecx
+; X86-NEXT: addl $32, %ecx
; X86-NEXT: testl %esi, %esi
-; X86-NEXT: cmovel %edx, %ecx
+; X86-NEXT: cmovnel %edx, %ecx
; X86-NEXT: bsrl %ebx, %edx
; X86-NEXT: xorl $31, %edx
; X86-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
@@ -226,14 +225,14 @@ define i128 @scalar_i128(i128 %x, i128 %y, ptr %divdst) nounwind {
; X86-NEXT: movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
; X86-NEXT: orl %esi, %edx
; X86-NEXT: cmovnel %ecx, %edi
-; X86-NEXT: bsrl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
-; X86-NEXT: xorl $31, %edx
-; X86-NEXT: addl $32, %edx
; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
-; X86-NEXT: bsrl %eax, %ecx
+; X86-NEXT: bsrl %eax, %edx
+; X86-NEXT: xorl $31, %edx
+; X86-NEXT: bsrl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
; X86-NEXT: xorl $31, %ecx
+; X86-NEXT: addl $32, %ecx
; X86-NEXT: testl %eax, %eax
-; X86-NEXT: cmovel %edx, %ecx
+; X86-NEXT: cmovnel %edx, %ecx
; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
; X86-NEXT: bsrl %ebx, %esi
; X86-NEXT: xorl $31, %esi
@@ -380,9 +379,9 @@ define i128 @scalar_i128(i128 %x, i128 %y, ptr %divdst) nounwind {
; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
; X86-NEXT: adcl $-1, %eax
; X86-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
-; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
-; X86-NEXT: adcl $-1, %eax
-; X86-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
+; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
+; X86-NEXT: adcl $-1, %ecx
+; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
; X86-NEXT: adcl $-1, %ecx
; X86-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
diff --git a/llvm/test/CodeGen/X86/pr62286.ll b/llvm/test/CodeGen/X86/pr62286.ll
index 2d1b7fcbf0239..ce03f8fad4a19 100644
--- a/llvm/test/CodeGen/X86/pr62286.ll
+++ b/llvm/test/CodeGen/X86/pr62286.ll
@@ -28,8 +28,9 @@ define i64 @PR62286(i32 %a) {
; AVX1-NEXT: vmovd %edi, %xmm0
; AVX1-NEXT: vpslldq {{.*#+}} xmm1 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm0[0,1,2,3]
; AVX1-NEXT: vpaddd %xmm0, %xmm0, %xmm0
+; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm2[2,3,4,5,6,7]
; AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1
-; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
; AVX1-NEXT: vpmovsxdq %xmm0, %xmm0
; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
; AVX1-NEXT: vpmovsxdq %xmm1, %xmm1
@@ -58,12 +59,13 @@ define i64 @PR62286(i32 %a) {
; AVX512-LABEL: PR62286:
; AVX512: # %bb.0:
; AVX512-NEXT: vmovd %edi, %xmm0
-; AVX512-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,1,0]
-; AVX512-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm1
-; AVX512-NEXT: movw $4369, %ax # imm = 0x1111
+; AVX512-NEXT: movb $8, %al
; AVX512-NEXT: kmovd %eax, %k1
-; AVX512-NEXT: vpaddd %zmm0, %zmm0, %zmm1 {%k1}
-; AVX512-NEXT: vpmovsxdq %ymm1, %zmm0
+; AVX512-NEXT: vpexpandd %ymm0, %ymm1 {%k1} {z}
+; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; AVX512-NEXT: vpaddd %ymm0, %ymm0, %ymm0
+; AVX512-NEXT: vpblendd {{.*#+}} ymm0 = ymm0[0],ymm1[1,2,3],ymm0[4],ymm1[5,6,7]
+; AVX512-NEXT: vpmovsxdq %ymm0, %zmm0
; AVX512-NEXT: vextracti64x4 $1, %zmm0, %ymm1
; AVX512-NEXT: vpaddq %zmm1, %zmm0, %zmm0
; AVX512-NEXT: vextracti128 $1, %ymm0, %xmm1
More information about the llvm-commits
mailing list