Skip to content
Merged
Show file tree
Hide file tree
Changes from 4 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
36 changes: 32 additions & 4 deletions llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -4721,9 +4721,6 @@ bool X86DAGToDAGISel::tryVPTERNLOG(SDNode *N) {
if (!(Subtarget->hasVLX() || NVT.is512BitVector()))
return false;

SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);

auto getFoldableLogicOp = [](SDValue Op) {
// Peek through single use bitcast.
if (Op.getOpcode() == ISD::BITCAST && Op.hasOneUse())
Expand All @@ -4740,6 +4737,34 @@ bool X86DAGToDAGISel::tryVPTERNLOG(SDNode *N) {
return SDValue();
};

// Identify and (optionally) peel an outer NOT that wraps a pure logic tree
auto tryPeelOuterNotWrappingLogic = [&](SDNode *Op) {
if (Op->getOpcode() == ISD::XOR && Op->hasOneUse() &&
ISD::isBuildVectorAllOnes(Op->getOperand(1).getNode())) {
SDValue InnerOp = Op->getOperand(0);

if (!getFoldableLogicOp(InnerOp)) {
return SDValue();
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@yichi170 Should this be:

SDValue InnerOp = getFoldableLogicOp(Op->getOperand(0));
if (!InnerOp)
  return SDValue();

getFoldableLogicOp might peek through a bitcast so the original InnerOp might not be a logic binop

Copy link
Contributor Author

@yichi170 yichi170 Nov 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, you're right. Thanks for pointing it out! Should I submit another PR to fix it?

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yes please - I've managed to get a test case to crash here, but you might have ideas for a better one:

define <32 x i8> @crashme(<32 x i8> %a0) {
  %cmp = icmp ne <32 x i8> %a0, zeroinitializer
  %sext = sext <32 x i1> %cmp to <32 x i8>
  %xor = xor <32 x i8> %a0, %sext
  %shuffle = shufflevector <32 x i8> %xor, <32 x i8> zeroinitializer, <32 x i32> <i32 3, i32 1, i32 3, i32 2, i32 3, i32 2, i32 0, i32 1, i32 0, i32 1, i32 1, i32 2, i32 3, i32 3, i32 1, i32 3, i32 1, i32 3, i32 2, i32 2, i32 0, i32 3, i32 1, i32 1, i32 3, i32 0, i32 0, i32 3, i32 3, i32 0, i32 0, i32 0>
  ret <32 x i8> %shuffle
}

}

SDValue InnerN0 = InnerOp.getOperand(0);
SDValue InnerN1 = InnerOp.getOperand(1);
if (getFoldableLogicOp(InnerN1) || getFoldableLogicOp(InnerN0))
return InnerOp;
}
return SDValue();
};

bool PeeledOuterNot = false;
SDNode *OriN = N;
if (SDValue InnerOp = tryPeelOuterNotWrappingLogic(N)) {
PeeledOuterNot = true;
N = InnerOp.getNode();
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
N = InnerOp.getNode();
ParentA = InnerOp.getNode();

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Unfortunately, we still need N in line 4813...

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If we peel the not out, the immediate (Imm) should be computed using the inner node (InnerOp), since the negation will be applied afterward. Therefore, we should not update this line here.

}

SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);

SDValue A, FoldableOp;
if ((FoldableOp = getFoldableLogicOp(N1))) {
A = N0;
Expand Down Expand Up @@ -4798,7 +4823,10 @@ bool X86DAGToDAGISel::tryVPTERNLOG(SDNode *N) {
case ISD::XOR: Imm ^= TernlogMagicA; break;
}

return matchVPTERNLOG(N, ParentA, ParentB, ParentC, A, B, C, Imm);
if (PeeledOuterNot)
Imm = ~Imm;

return matchVPTERNLOG(OriN, ParentA, ParentB, ParentC, A, B, C, Imm);
}

/// If the high bits of an 'and' operand are known zero, try setting the
Expand Down
13 changes: 13 additions & 0 deletions llvm/test/CodeGen/X86/issue163738.ll
Original file line number Diff line number Diff line change
@@ -0,0 +1,13 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f,+avx512vl | FileCheck %s --check-prefixes=CHECK

define <8 x i64> @foo(<8 x i64> %a, <8 x i64> %b, <8 x i64> %c) {
; CHECK-LABEL: foo:
; CHECK: # %bb.0:
; CHECK-NEXT: vpternlogq {{.*#+}} zmm0 = ~(zmm0 | zmm2 | zmm1)
; CHECK-NEXT: retq
%and.demorgan = or <8 x i64> %b, %a
%and3.demorgan = or <8 x i64> %and.demorgan, %c
%and3 = xor <8 x i64> %and3.demorgan, splat (i64 -1)
ret <8 x i64> %and3
}