[llvm] [AArch64][SVE] Fold predicate SETCC negations (PR #206931)
Matthew Devereau via llvm-commits
llvm-commits at lists.llvm.org
Wed Jul 1 02:56:47 PDT 2026
https://github.com/MDevereau updated https://github.com/llvm/llvm-project/pull/206931
>From b8ce0b78b756de005dab4cd587990fa71ccd12a6 Mon Sep 17 00:00:00 2001
From: Matthew Devereau <matthew.devereau at arm.com>
Date: Wed, 1 Jul 2026 09:37:45 +0000
Subject: [PATCH] [AArch64][SVE] Fold predicate SETCC negations
Fold XORs of an SVE SETCC_MERGE_ZERO result with its governing
predicate into the inverse SETCC_MERGE_ZERO condition. This lets
patterns like `xor (cmpne pg, ...), pg` lower directly to `cmpeq`,
and likewise folds `cmpeq` negations to `cmpne`.
---
.../Target/AArch64/AArch64ISelLowering.cpp | 29 ++++++++++++
llvm/test/CodeGen/AArch64/sve-setcc.ll | 44 +++++++++++++++++++
2 files changed, 73 insertions(+)
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index e883c8bb5e96e..d1ce401527870 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -20480,12 +20480,41 @@ static SDValue performBICiCombine(SDNode *N, SelectionDAG &DAG,
return SDValue();
}
+static SDValue performPredicateSetCCNotCombine(SDNode *N, SelectionDAG &DAG) {
+ assert(N->getOpcode() == ISD::XOR && "Unexpected opcode!");
+
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ if (N1->getOpcode() == AArch64ISD::SETCC_MERGE_ZERO)
+ std::swap(N0, N1);
+
+ if (N0->getOpcode() != AArch64ISD::SETCC_MERGE_ZERO || !N0.hasOneUse())
+ return SDValue();
+
+ SDValue Pred = N0->getOperand(0);
+ if (N1 != Pred)
+ return SDValue();
+
+ SDLoc DL(N);
+ SDValue LHS = N0->getOperand(1);
+ SDValue RHS = N0->getOperand(2);
+ ISD::CondCode CC = cast<CondCodeSDNode>(N0->getOperand(3))->get();
+ if (CC != ISD::CondCode::SETNE && CC != ISD::CondCode::SETEQ)
+ return SDValue();
+ ISD::CondCode InvCC = ISD::getSetCCInverse(CC, LHS.getValueType());
+ return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, DL, N->getValueType(0), Pred,
+ LHS, RHS, DAG.getCondCode(InvCC));
+}
+
static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const AArch64Subtarget *Subtarget) {
if (DCI.isBeforeLegalizeOps())
return SDValue();
+ if (SDValue V = performPredicateSetCCNotCombine(N, DAG))
+ return V;
+
return foldVectorXorShiftIntoCmp(N, DAG, Subtarget);
}
diff --git a/llvm/test/CodeGen/AArch64/sve-setcc.ll b/llvm/test/CodeGen/AArch64/sve-setcc.ll
index 30e2a77f32120..4f79224d24ac2 100644
--- a/llvm/test/CodeGen/AArch64/sve-setcc.ll
+++ b/llvm/test/CodeGen/AArch64/sve-setcc.ll
@@ -116,6 +116,50 @@ define <vscale x 16 x i1> @sve_cmpne_setcc_different_pred(<vscale x 16 x i8> %ve
ret <vscale x 16 x i1> %cmp2
}
+define <vscale x 16 x i1> @sve_not_cmpne_setcc(<vscale x 16 x i8> %vec, <vscale x 16 x i1> %pg) {
+; CHECK-LABEL: sve_not_cmpne_setcc:
+; CHECK: // %bb.0:
+; CHECK-NEXT: cmpeq p0.b, p0/z, z0.b, #0
+; CHECK-NEXT: ret
+ %cmp = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpne.nxv16i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> %vec, <vscale x 16 x i8> zeroinitializer)
+ %not = xor <vscale x 16 x i1> %cmp, %pg
+ ret <vscale x 16 x i1> %not
+}
+
+define <vscale x 16 x i1> @sve_not_cmpne_setcc_true(<vscale x 16 x i8> %vec) {
+; CHECK-LABEL: sve_not_cmpne_setcc_true:
+; CHECK: // %bb.0:
+; CHECK-NEXT: ptrue p0.b
+; CHECK-NEXT: cmpeq p0.b, p0/z, z0.b, #0
+; CHECK-NEXT: ret
+ %cmp = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpne.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %vec, <vscale x 16 x i8> zeroinitializer)
+ %not = xor <vscale x 16 x i1> %cmp, splat (i1 true)
+ ret <vscale x 16 x i1> %not
+}
+
+define <vscale x 16 x i1> @sve_not_cmpeq_setcc(<vscale x 16 x i8> %vec, <vscale x 16 x i1> %pg) {
+; CHECK-LABEL: sve_not_cmpeq_setcc:
+; CHECK: // %bb.0:
+; CHECK-NEXT: cmpne p0.b, p0/z, z0.b, #0
+; CHECK-NEXT: ret
+ %cmp = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpeq.nxv16i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> %vec, <vscale x 16 x i8> zeroinitializer)
+ %not = xor <vscale x 16 x i1> %cmp, %pg
+ ret <vscale x 16 x i1> %not
+}
+
+define <vscale x 16 x i1> @sve_not_cmpeq_setcc_true(<vscale x 16 x i8> %vec) {
+; CHECK-LABEL: sve_not_cmpeq_setcc_true:
+; CHECK: // %bb.0:
+; CHECK-NEXT: ptrue p0.b
+; CHECK-NEXT: cmpne p0.b, p0/z, z0.b, #0
+; CHECK-NEXT: ret
+ %cmp = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpeq.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %vec, <vscale x 16 x i8> zeroinitializer)
+ %not = xor <vscale x 16 x i1> %cmp, splat (i1 true)
+ ret <vscale x 16 x i1> %not
+}
+
+
+
declare <vscale x 16 x i1> @llvm.aarch64.sve.cmpne.nxv16i8(<vscale x 16 x i1>, <vscale x 16 x i8>, <vscale x 16 x i8>)
declare i1 @llvm.aarch64.sve.ptest.any.nxv8i1(<vscale x 8 x i1>, <vscale x 8 x i1>)
More information about the llvm-commits
mailing list