[llvm] cc4f462 - [X86][ISel] Improve VPTERNLOG matching for negated logic trees (#164863)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Oct 24 18:48:39 PDT 2025
Author: Yi-Chi Lee
Date: 2025-10-25T09:48:35+08:00
New Revision: cc4f462ba46c288dc26aa155775444cd9a7c4d0e
URL: https://github.com/llvm/llvm-project/commit/cc4f462ba46c288dc26aa155775444cd9a7c4d0e
DIFF: https://github.com/llvm/llvm-project/commit/cc4f462ba46c288dc26aa155775444cd9a7c4d0e.diff
LOG: [X86][ISel] Improve VPTERNLOG matching for negated logic trees (#164863)
This patch extends VPTERNLOG pattern matching to handle cases where an
outer NOT wraps a pure logical tree, such as `~(A | B | C)`. By
recognizing these negated logic trees, the instruction selector can now
emit a single vpternlog instruction.
The change preserves the match for patterns like `(x != C1) & (x !=
C2)`, which also have the xor-with-all-ones pattern outside. The patch
conservatively peels the outer XOR-with-all-ones only when it directly
wraps a foldable logical operator (AND, OR, XOR, or ANDNP).
Resolves #163738
Added:
llvm/test/CodeGen/X86/issue163738.ll
Modified:
llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
index 62073ec125e8f..4393f6ecaa033 100644
--- a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -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())
@@ -4740,13 +4737,47 @@ bool X86DAGToDAGISel::tryVPTERNLOG(SDNode *N) {
return SDValue();
};
- SDValue A, FoldableOp;
- if ((FoldableOp = getFoldableLogicOp(N1))) {
- A = N0;
- } else if ((FoldableOp = getFoldableLogicOp(N0))) {
- A = N1;
- } else
- return false;
+ SDValue N0, N1, A, FoldableOp;
+
+ // 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();
+
+ N0 = InnerOp.getOperand(0);
+ N1 = InnerOp.getOperand(1);
+ if ((FoldableOp = getFoldableLogicOp(N1))) {
+ A = N0;
+ return InnerOp;
+ }
+ if ((FoldableOp = getFoldableLogicOp(N0))) {
+ A = N1;
+ return InnerOp;
+ }
+ }
+ return SDValue();
+ };
+
+ bool PeeledOuterNot = false;
+ SDNode *OriN = N;
+ if (SDValue InnerOp = tryPeelOuterNotWrappingLogic(N)) {
+ PeeledOuterNot = true;
+ N = InnerOp.getNode();
+ } else {
+ N0 = N->getOperand(0);
+ N1 = N->getOperand(1);
+
+ if ((FoldableOp = getFoldableLogicOp(N1)))
+ A = N0;
+ else if ((FoldableOp = getFoldableLogicOp(N0)))
+ A = N1;
+ else
+ return false;
+ }
SDValue B = FoldableOp.getOperand(0);
SDValue C = FoldableOp.getOperand(1);
@@ -4798,7 +4829,10 @@ 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
diff --git a/llvm/test/CodeGen/X86/issue163738.ll b/llvm/test/CodeGen/X86/issue163738.ll
new file mode 100644
index 0000000000000..61fe043a970dd
--- /dev/null
+++ b/llvm/test/CodeGen/X86/issue163738.ll
@@ -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
+}
More information about the llvm-commits
mailing list