[llvm] [AArch64][InstCombine] Fold zext cmpne umin boolean trees (PR #209234)
Matthew Devereau via llvm-commits
llvm-commits at lists.llvm.org
Wed Jul 22 02:29:50 PDT 2026
https://github.com/MDevereau updated https://github.com/llvm/llvm-project/pull/209234
>From 2723e7b1ba2ef5850f6a558aba250b93b8beebba Mon Sep 17 00:00:00 2001
From: Matt Devereau <matthew.devereau at arm.com>
Date: Wed, 8 Jul 2026 16:00:59 +0000
Subject: [PATCH 1/3] [AArch64][InstCombine] Fold zext cmpne umin boolean trees
Fold an all-active umin tree of zext(cmpne(0, umin)) nodes:
zext(cmpne(0, umin(zext(cmpne(0, umin(%a, %b)))
zext(cmpne(0, umin(%c, %d))))))
->
zext(cmpne(0, umin(umin(%a, %b),
umin(%c, %d))))
---
.../AArch64/AArch64TargetTransformInfo.cpp | 66 ++++++++
.../AArch64/sve-intrinsic-opts-cmpne.ll | 154 ++++++++++++++++++
2 files changed, 220 insertions(+)
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index 9e1a0960617a5..024f9efaf18f2 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -2205,6 +2205,69 @@ static std::optional<Instruction *> instCombineZExtSVECmpNE(InstCombiner &IC,
return std::nullopt;
}
+static std::optional<Instruction *> isZextedCmpneUMin(Value *V) {
+ Instruction *CmpNE;
+ if (!match(V, m_ZExt(m_Instruction(CmpNE))))
+ return std::nullopt;
+
+ Value *Pg;
+ Instruction *UMin;
+ if (!match(CmpNE, m_Intrinsic<Intrinsic::aarch64_sve_cmpne>(
+ m_Value(Pg), m_Zero(), m_Instruction(UMin))) ||
+ !isAllActivePredicate(Pg) || V->getType() != UMin->getType())
+ return std::nullopt;
+
+ Value *UMinPg;
+ if (!match(UMin, m_Intrinsic<Intrinsic::aarch64_sve_umin_u>(
+ m_Value(UMinPg), m_Value(), m_Value())) ||
+ !isAllActivePredicate(UMinPg))
+ return std::nullopt;
+
+ return UMin;
+}
+
+// InstCombine:
+// zext(cmpne(0, umin(zext(cmpne(0, umin(%a, %b)))
+// zext(cmpne(0, umin(%c, %d))))))
+// ->
+// zext(cmpne(0, umin(umin(%a, %b),
+// umin(%c, %d))))
+//
+// To remove redundant ZExt and CmpNE intrinsics.
+static std::optional<Instruction *>
+instCombineZExtCmpNEUMinTree(InstCombiner &IC, IntrinsicInst &II) {
+ Value *Pg = II.getOperand(0);
+ if (!isAllActivePredicate(Pg) || !II.hasOneUse())
+ return std::nullopt;
+
+ auto ZExt = cast<Instruction>(*II.user_begin());
+ auto _UMin = isZextedCmpneUMin(ZExt);
+ if (!_UMin)
+ return std::nullopt;
+ Instruction *UMin = *_UMin;
+ // Check for outer UMin single use to avoid generating an extra UMin
+ if (!UMin->hasOneUse())
+ return std::nullopt;
+
+ auto LhsUMin = isZextedCmpneUMin(UMin->getOperand(1));
+ auto RhsUMin = isZextedCmpneUMin(UMin->getOperand(2));
+ if (!LhsUMin || !RhsUMin)
+ return std::nullopt;
+
+ IC.Builder.SetInsertPoint(ZExt);
+ Type *Ty = UMin->getType();
+ Value *NewUMin = IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_umin_u, Ty,
+ {Pg, *LhsUMin, *RhsUMin});
+ Value *NewCmpne =
+ IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_cmpne, Ty,
+ {Pg, ConstantInt::getNullValue(Ty), NewUMin});
+ Value *NewZext = IC.Builder.CreateZExt(NewCmpne, Ty);
+
+ IC.replaceInstUsesWith(*ZExt, NewZext);
+ IC.eraseInstFromFunction(*ZExt);
+ return &II;
+}
+
static std::optional<Instruction *> instCombineSVECmpNE(InstCombiner &IC,
IntrinsicInst &II) {
LLVMContext &Ctx = II.getContext();
@@ -2215,6 +2278,9 @@ static std::optional<Instruction *> instCombineSVECmpNE(InstCombiner &IC,
if (auto Res = instCombineZExtSVECmpNE(IC, II))
return Res;
+ if (auto Res = instCombineZExtCmpNEUMinTree(IC, II))
+ return Res;
+
if (!isAllActivePredicate(II.getArgOperand(0)))
return std::nullopt;
diff --git a/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-opts-cmpne.ll b/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-opts-cmpne.ll
index 11e6e0089293a..a434ad086d891 100644
--- a/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-opts-cmpne.ll
+++ b/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-opts-cmpne.ll
@@ -573,6 +573,160 @@ define <vscale x 16 x i1> @dupq_b_idx(i64 %idx) #0 {
ret <vscale x 16 x i1> %5
}
+define <vscale x 16 x i8> @logical_bool_tree_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c, <vscale x 16 x i8> %d) #0 {
+; CHECK-LABEL: define <vscale x 16 x i8> @logical_bool_tree_i8(
+; CHECK-SAME: <vscale x 16 x i8> [[A:%.*]], <vscale x 16 x i8> [[B:%.*]], <vscale x 16 x i8> [[C:%.*]], <vscale x 16 x i8> [[D:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT: [[AB:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[A]], <vscale x 16 x i8> [[B]])
+; CHECK-NEXT: [[CD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[C]], <vscale x 16 x i8> [[D]])
+; CHECK-NEXT: [[ABCD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[AB]], <vscale x 16 x i8> [[CD]])
+; CHECK-NEXT: [[ABCD_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> zeroinitializer, <vscale x 16 x i8> [[ABCD]])
+; CHECK-NEXT: [[ABCD_BOOL:%.*]] = zext <vscale x 16 x i1> [[ABCD_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: ret <vscale x 16 x i8> [[ABCD_BOOL]]
+;
+ %ab = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %a, <vscale x 16 x i8> %b)
+ %ab.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> zeroinitializer, <vscale x 16 x i8> %ab)
+ %ab.bool = zext <vscale x 16 x i1> %ab.cmp to <vscale x 16 x i8>
+ %cd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %c, <vscale x 16 x i8> %d)
+ %cd.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> zeroinitializer, <vscale x 16 x i8> %cd)
+ %cd.bool = zext <vscale x 16 x i1> %cd.cmp to <vscale x 16 x i8>
+ %abcd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %ab.bool, <vscale x 16 x i8> %cd.bool)
+ %abcd.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> zeroinitializer, <vscale x 16 x i8> %abcd)
+ %abcd.bool = zext <vscale x 16 x i1> %abcd.cmp to <vscale x 16 x i8>
+ ret <vscale x 16 x i8> %abcd.bool
+}
+
+define <vscale x 16 x i8> @logical_bool_tree_outer_umin_extra_use_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c, <vscale x 16 x i8> %d) #0 {
+; CHECK-LABEL: define <vscale x 16 x i8> @logical_bool_tree_outer_umin_extra_use_i8(
+; CHECK-SAME: <vscale x 16 x i8> [[A:%.*]], <vscale x 16 x i8> [[B:%.*]], <vscale x 16 x i8> [[C:%.*]], <vscale x 16 x i8> [[D:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT: [[AB:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[A]], <vscale x 16 x i8> [[B]])
+; CHECK-NEXT: [[AB_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> zeroinitializer, <vscale x 16 x i8> [[AB]])
+; CHECK-NEXT: [[AB_BOOL:%.*]] = zext <vscale x 16 x i1> [[AB_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: [[CD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[C]], <vscale x 16 x i8> [[D]])
+; CHECK-NEXT: [[CD_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> zeroinitializer, <vscale x 16 x i8> [[CD]])
+; CHECK-NEXT: [[CD_BOOL:%.*]] = zext <vscale x 16 x i1> [[CD_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: [[TMP1:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[AB_BOOL]], <vscale x 16 x i8> [[CD_BOOL]])
+; CHECK-NEXT: [[ABCD_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> zeroinitializer, <vscale x 16 x i8> [[TMP1]])
+; CHECK-NEXT: [[ABCD_BOOL:%.*]] = zext <vscale x 16 x i1> [[ABCD_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: call void (...) @llvm.fake.use(<vscale x 16 x i8> [[TMP1]])
+; CHECK-NEXT: ret <vscale x 16 x i8> [[ABCD_BOOL]]
+;
+ %ab = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %a, <vscale x 16 x i8> %b)
+ %ab.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> zeroinitializer, <vscale x 16 x i8> %ab)
+ %ab.bool = zext <vscale x 16 x i1> %ab.cmp to <vscale x 16 x i8>
+ %cd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %c, <vscale x 16 x i8> %d)
+ %cd.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> zeroinitializer, <vscale x 16 x i8> %cd)
+ %cd.bool = zext <vscale x 16 x i1> %cd.cmp to <vscale x 16 x i8>
+ %abcd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %ab.bool, <vscale x 16 x i8> %cd.bool)
+ %abcd.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> zeroinitializer, <vscale x 16 x i8> %abcd)
+ %abcd.bool = zext <vscale x 16 x i1> %abcd.cmp to <vscale x 16 x i8>
+ call void (...) @llvm.fake.use(<vscale x 16 x i8> %abcd)
+ ret <vscale x 16 x i8> %abcd.bool
+}
+
+define <vscale x 16 x i8> @logical_bool_tree_nonzero_outer_cmp_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c, <vscale x 16 x i8> %d, <vscale x 16 x i8> %mask) #0 {
+; CHECK-LABEL: define <vscale x 16 x i8> @logical_bool_tree_nonzero_outer_cmp_i8(
+; CHECK-SAME: <vscale x 16 x i8> [[A:%.*]], <vscale x 16 x i8> [[B:%.*]], <vscale x 16 x i8> [[C:%.*]], <vscale x 16 x i8> [[D:%.*]], <vscale x 16 x i8> [[MASK:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT: [[AB:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[A]], <vscale x 16 x i8> [[B]])
+; CHECK-NEXT: [[AB_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> zeroinitializer, <vscale x 16 x i8> [[AB]])
+; CHECK-NEXT: [[AB_BOOL:%.*]] = zext <vscale x 16 x i1> [[AB_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: [[CD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[C]], <vscale x 16 x i8> [[D]])
+; CHECK-NEXT: [[CD_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> zeroinitializer, <vscale x 16 x i8> [[CD]])
+; CHECK-NEXT: [[CD_BOOL:%.*]] = zext <vscale x 16 x i1> [[CD_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: [[ABCD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[AB_BOOL]], <vscale x 16 x i8> [[CD_BOOL]])
+; CHECK-NEXT: [[ABCD_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> [[MASK]], <vscale x 16 x i8> [[ABCD]])
+; CHECK-NEXT: [[ABCD_BOOL:%.*]] = zext <vscale x 16 x i1> [[ABCD_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: ret <vscale x 16 x i8> [[ABCD_BOOL]]
+;
+ %ab = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %a, <vscale x 16 x i8> %b)
+ %ab.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> zeroinitializer, <vscale x 16 x i8> %ab)
+ %ab.bool = zext <vscale x 16 x i1> %ab.cmp to <vscale x 16 x i8>
+ %cd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %c, <vscale x 16 x i8> %d)
+ %cd.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> zeroinitializer, <vscale x 16 x i8> %cd)
+ %cd.bool = zext <vscale x 16 x i1> %cd.cmp to <vscale x 16 x i8>
+ %abcd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %ab.bool, <vscale x 16 x i8> %cd.bool)
+ %abcd.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> %mask, <vscale x 16 x i8> %abcd)
+ %abcd.bool = zext <vscale x 16 x i1> %abcd.cmp to <vscale x 16 x i8>
+ ret <vscale x 16 x i8> %abcd.bool
+}
+
+define <vscale x 16 x i8> @logical_bool_tree_inactive_predicate_i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c, <vscale x 16 x i8> %d) #0 {
+; CHECK-LABEL: define <vscale x 16 x i8> @logical_bool_tree_inactive_predicate_i8(
+; CHECK-SAME: <vscale x 16 x i1> [[PG:%.*]], <vscale x 16 x i8> [[A:%.*]], <vscale x 16 x i8> [[B:%.*]], <vscale x 16 x i8> [[C:%.*]], <vscale x 16 x i8> [[D:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT: [[AB:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> [[PG]], <vscale x 16 x i8> [[A]], <vscale x 16 x i8> [[B]])
+; CHECK-NEXT: [[AB_CMP:%.*]] = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpne.nxv16i8(<vscale x 16 x i1> [[PG]], <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8> [[AB]])
+; CHECK-NEXT: [[AB_BOOL:%.*]] = zext <vscale x 16 x i1> [[AB_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: [[CD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> [[PG]], <vscale x 16 x i8> [[C]], <vscale x 16 x i8> [[D]])
+; CHECK-NEXT: [[CD_CMP:%.*]] = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpne.nxv16i8(<vscale x 16 x i1> [[PG]], <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8> [[CD]])
+; CHECK-NEXT: [[CD_BOOL:%.*]] = zext <vscale x 16 x i1> [[CD_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: [[ABCD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> [[PG]], <vscale x 16 x i8> [[AB_BOOL]], <vscale x 16 x i8> [[CD_BOOL]])
+; CHECK-NEXT: [[ABCD_CMP:%.*]] = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpne.nxv16i8(<vscale x 16 x i1> [[PG]], <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8> [[ABCD]])
+; CHECK-NEXT: [[ABCD_BOOL:%.*]] = zext <vscale x 16 x i1> [[ABCD_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: ret <vscale x 16 x i8> [[ABCD_BOOL]]
+;
+ %ab = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> %a, <vscale x 16 x i8> %b)
+ %ab.cmp = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpne.nxv16i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8> %ab)
+ %ab.bool = zext <vscale x 16 x i1> %ab.cmp to <vscale x 16 x i8>
+ %cd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> %c, <vscale x 16 x i8> %d)
+ %cd.cmp = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpne.nxv16i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8> %cd)
+ %cd.bool = zext <vscale x 16 x i1> %cd.cmp to <vscale x 16 x i8>
+ %abcd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> %ab.bool, <vscale x 16 x i8> %cd.bool)
+ %abcd.cmp = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpne.nxv16i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> zeroinitializer, <vscale x 16 x i8> %abcd)
+ %abcd.bool = zext <vscale x 16 x i1> %abcd.cmp to <vscale x 16 x i8>
+ ret <vscale x 16 x i8> %abcd.bool
+}
+
+define <vscale x 16 x i16> @logical_bool_tree_wide_final_zext_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c, <vscale x 16 x i8> %d) #0 {
+; CHECK-LABEL: define <vscale x 16 x i16> @logical_bool_tree_wide_final_zext_i8(
+; CHECK-SAME: <vscale x 16 x i8> [[A:%.*]], <vscale x 16 x i8> [[B:%.*]], <vscale x 16 x i8> [[C:%.*]], <vscale x 16 x i8> [[D:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT: [[AB:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[A]], <vscale x 16 x i8> [[B]])
+; CHECK-NEXT: [[AB_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> zeroinitializer, <vscale x 16 x i8> [[AB]])
+; CHECK-NEXT: [[AB_BOOL:%.*]] = zext <vscale x 16 x i1> [[AB_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: [[CD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[C]], <vscale x 16 x i8> [[D]])
+; CHECK-NEXT: [[CD_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> zeroinitializer, <vscale x 16 x i8> [[CD]])
+; CHECK-NEXT: [[CD_BOOL:%.*]] = zext <vscale x 16 x i1> [[CD_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: [[ABCD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[AB_BOOL]], <vscale x 16 x i8> [[CD_BOOL]])
+; CHECK-NEXT: [[ABCD_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> zeroinitializer, <vscale x 16 x i8> [[ABCD]])
+; CHECK-NEXT: [[ABCD_BOOL:%.*]] = zext <vscale x 16 x i1> [[ABCD_CMP]] to <vscale x 16 x i16>
+; CHECK-NEXT: ret <vscale x 16 x i16> [[ABCD_BOOL]]
+;
+ %ab = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %a, <vscale x 16 x i8> %b)
+ %ab.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> zeroinitializer, <vscale x 16 x i8> %ab)
+ %ab.bool = zext <vscale x 16 x i1> %ab.cmp to <vscale x 16 x i8>
+ %cd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %c, <vscale x 16 x i8> %d)
+ %cd.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> zeroinitializer, <vscale x 16 x i8> %cd)
+ %cd.bool = zext <vscale x 16 x i1> %cd.cmp to <vscale x 16 x i8>
+ %abcd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %ab.bool, <vscale x 16 x i8> %cd.bool)
+ %abcd.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> zeroinitializer, <vscale x 16 x i8> %abcd)
+ %abcd.bool = zext <vscale x 16 x i1> %abcd.cmp to <vscale x 16 x i16>
+ ret <vscale x 16 x i16> %abcd.bool
+}
+
+define <vscale x 16 x i16> @logical_bool_tree_wide_inner_zext_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c, <vscale x 16 x i8> %d) #0 {
+; CHECK-LABEL: define <vscale x 16 x i16> @logical_bool_tree_wide_inner_zext_i8(
+; CHECK-SAME: <vscale x 16 x i8> [[A:%.*]], <vscale x 16 x i8> [[B:%.*]], <vscale x 16 x i8> [[C:%.*]], <vscale x 16 x i8> [[D:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT: [[AB:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[A]], <vscale x 16 x i8> [[B]])
+; CHECK-NEXT: [[AB_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> zeroinitializer, <vscale x 16 x i8> [[AB]])
+; CHECK-NEXT: [[AB_BOOL:%.*]] = zext <vscale x 16 x i1> [[AB_CMP]] to <vscale x 16 x i16>
+; CHECK-NEXT: [[CD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[C]], <vscale x 16 x i8> [[D]])
+; CHECK-NEXT: [[CD_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> zeroinitializer, <vscale x 16 x i8> [[CD]])
+; CHECK-NEXT: [[CD_BOOL:%.*]] = zext <vscale x 16 x i1> [[CD_CMP]] to <vscale x 16 x i16>
+; CHECK-NEXT: [[ABCD:%.*]] = call <vscale x 16 x i16> @llvm.aarch64.sve.umin.u.nxv16i16(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i16> [[AB_BOOL]], <vscale x 16 x i16> [[CD_BOOL]])
+; CHECK-NEXT: [[ABCD_CMP:%.*]] = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpne.nxv16i16(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i16> zeroinitializer, <vscale x 16 x i16> [[ABCD]])
+; CHECK-NEXT: [[ABCD_BOOL:%.*]] = zext <vscale x 16 x i1> [[ABCD_CMP]] to <vscale x 16 x i16>
+; CHECK-NEXT: ret <vscale x 16 x i16> [[ABCD_BOOL]]
+;
+ %ab = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %a, <vscale x 16 x i8> %b)
+ %ab.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> zeroinitializer, <vscale x 16 x i8> %ab)
+ %ab.bool = zext <vscale x 16 x i1> %ab.cmp to <vscale x 16 x i16>
+ %cd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %c, <vscale x 16 x i8> %d)
+ %cd.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> zeroinitializer, <vscale x 16 x i8> %cd)
+ %cd.bool = zext <vscale x 16 x i1> %cd.cmp to <vscale x 16 x i16>
+ %abcd = call <vscale x 16 x i16> @llvm.aarch64.sve.umin.u.nxv16i16(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i16> %ab.bool, <vscale x 16 x i16> %cd.bool)
+ %abcd.cmp = call <vscale x 16 x i1> @llvm.aarch64.sve.cmpne.nxv16i16(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i16> zeroinitializer, <vscale x 16 x i16> %abcd)
+ %abcd.bool = zext <vscale x 16 x i1> %abcd.cmp to <vscale x 16 x i16>
+ ret <vscale x 16 x i16> %abcd.bool
+}
+
declare <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32)
declare <vscale x 16 x i8> @llvm.vector.insert.nxv16i8.v16i8(<vscale x 16 x i8>, <16 x i8>, i64)
>From fc74ead324e8ac0281db5c09dc4f2dca3e90ed05 Mon Sep 17 00:00:00 2001
From: Matt Devereau <matthew.devereau at arm.com>
Date: Mon, 20 Jul 2026 13:39:17 +0000
Subject: [PATCH 2/3] Add commutativity for cmpne
---
.../AArch64/AArch64TargetTransformInfo.cpp | 10 +++++----
.../AArch64/sve-intrinsic-opts-cmpne.ll | 22 +++++++++++++++++++
2 files changed, 28 insertions(+), 4 deletions(-)
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index 024f9efaf18f2..23b190ccc3df1 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -2212,8 +2212,10 @@ static std::optional<Instruction *> isZextedCmpneUMin(Value *V) {
Value *Pg;
Instruction *UMin;
- if (!match(CmpNE, m_Intrinsic<Intrinsic::aarch64_sve_cmpne>(
- m_Value(Pg), m_Zero(), m_Instruction(UMin))) ||
+ if ((!match(CmpNE, m_Intrinsic<Intrinsic::aarch64_sve_cmpne>(
+ m_Value(Pg), m_Zero(), m_Instruction(UMin))) &&
+ !match(CmpNE, m_Intrinsic<Intrinsic::aarch64_sve_cmpne>(
+ m_Value(Pg), m_Instruction(UMin), m_Zero()))) ||
!isAllActivePredicate(Pg) || V->getType() != UMin->getType())
return std::nullopt;
@@ -2275,10 +2277,10 @@ static std::optional<Instruction *> instCombineSVECmpNE(InstCombiner &IC,
if (auto Res = instCombineXorSVECmpCC(IC, II))
return Res;
- if (auto Res = instCombineZExtSVECmpNE(IC, II))
+ if (auto Res = instCombineZExtCmpNEUMinTree(IC, II))
return Res;
- if (auto Res = instCombineZExtCmpNEUMinTree(IC, II))
+ if (auto Res = instCombineZExtSVECmpNE(IC, II))
return Res;
if (!isAllActivePredicate(II.getArgOperand(0)))
diff --git a/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-opts-cmpne.ll b/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-opts-cmpne.ll
index a434ad086d891..40bee22f9c263 100644
--- a/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-opts-cmpne.ll
+++ b/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-opts-cmpne.ll
@@ -595,6 +595,28 @@ define <vscale x 16 x i8> @logical_bool_tree_i8(<vscale x 16 x i8> %a, <vscale x
ret <vscale x 16 x i8> %abcd.bool
}
+define <vscale x 16 x i8> @logical_bool_tree_commuted_outer_cmp_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c, <vscale x 16 x i8> %d) #0 {
+; CHECK-LABEL: define <vscale x 16 x i8> @logical_bool_tree_commuted_outer_cmp_i8(
+; CHECK-SAME: <vscale x 16 x i8> [[A:%.*]], <vscale x 16 x i8> [[B:%.*]], <vscale x 16 x i8> [[C:%.*]], <vscale x 16 x i8> [[D:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT: [[AB:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[A]], <vscale x 16 x i8> [[B]])
+; CHECK-NEXT: [[CD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[C]], <vscale x 16 x i8> [[D]])
+; CHECK-NEXT: [[ABCD:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> [[AB]], <vscale x 16 x i8> [[CD]])
+; CHECK-NEXT: [[ABCD_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> zeroinitializer, <vscale x 16 x i8> [[ABCD]])
+; CHECK-NEXT: [[ABCD_BOOL:%.*]] = zext <vscale x 16 x i1> [[ABCD_CMP]] to <vscale x 16 x i8>
+; CHECK-NEXT: ret <vscale x 16 x i8> [[ABCD_BOOL]]
+;
+ %ab = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %a, <vscale x 16 x i8> %b)
+ %ab.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> zeroinitializer, <vscale x 16 x i8> %ab)
+ %ab.bool = zext <vscale x 16 x i1> %ab.cmp to <vscale x 16 x i8>
+ %cd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %c, <vscale x 16 x i8> %d)
+ %cd.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> zeroinitializer, <vscale x 16 x i8> %cd)
+ %cd.bool = zext <vscale x 16 x i1> %cd.cmp to <vscale x 16 x i8>
+ %abcd = call <vscale x 16 x i8> @llvm.aarch64.sve.umin.u.nxv16i8(<vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> %ab.bool, <vscale x 16 x i8> %cd.bool)
+ %abcd.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> %abcd, <vscale x 16 x i8> zeroinitializer)
+ %abcd.bool = zext <vscale x 16 x i1> %abcd.cmp to <vscale x 16 x i8>
+ ret <vscale x 16 x i8> %abcd.bool
+}
+
define <vscale x 16 x i8> @logical_bool_tree_outer_umin_extra_use_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b, <vscale x 16 x i8> %c, <vscale x 16 x i8> %d) #0 {
; CHECK-LABEL: define <vscale x 16 x i8> @logical_bool_tree_outer_umin_extra_use_i8(
; CHECK-SAME: <vscale x 16 x i8> [[A:%.*]], <vscale x 16 x i8> [[B:%.*]], <vscale x 16 x i8> [[C:%.*]], <vscale x 16 x i8> [[D:%.*]]) #[[ATTR0]] {
>From 40e723b5d61d396b1b0b4eab81641f02c4feb412 Mon Sep 17 00:00:00 2001
From: Matt Devereau <matthew.devereau at arm.com>
Date: Wed, 22 Jul 2026 09:16:46 +0000
Subject: [PATCH 3/3] remove std::optional
---
.../AArch64/AArch64TargetTransformInfo.cpp | 23 ++++++++-----------
1 file changed, 10 insertions(+), 13 deletions(-)
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index 23b190ccc3df1..d3c61a8da9b32 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -2205,10 +2205,10 @@ static std::optional<Instruction *> instCombineZExtSVECmpNE(InstCombiner &IC,
return std::nullopt;
}
-static std::optional<Instruction *> isZextedCmpneUMin(Value *V) {
+static Instruction *getZextedCmpneUMin(Value *V) {
Instruction *CmpNE;
if (!match(V, m_ZExt(m_Instruction(CmpNE))))
- return std::nullopt;
+ return nullptr;
Value *Pg;
Instruction *UMin;
@@ -2217,13 +2217,13 @@ static std::optional<Instruction *> isZextedCmpneUMin(Value *V) {
!match(CmpNE, m_Intrinsic<Intrinsic::aarch64_sve_cmpne>(
m_Value(Pg), m_Instruction(UMin), m_Zero()))) ||
!isAllActivePredicate(Pg) || V->getType() != UMin->getType())
- return std::nullopt;
+ return nullptr;
Value *UMinPg;
if (!match(UMin, m_Intrinsic<Intrinsic::aarch64_sve_umin_u>(
m_Value(UMinPg), m_Value(), m_Value())) ||
!isAllActivePredicate(UMinPg))
- return std::nullopt;
+ return nullptr;
return UMin;
}
@@ -2243,23 +2243,20 @@ instCombineZExtCmpNEUMinTree(InstCombiner &IC, IntrinsicInst &II) {
return std::nullopt;
auto ZExt = cast<Instruction>(*II.user_begin());
- auto _UMin = isZextedCmpneUMin(ZExt);
- if (!_UMin)
- return std::nullopt;
- Instruction *UMin = *_UMin;
- // Check for outer UMin single use to avoid generating an extra UMin
- if (!UMin->hasOneUse())
+ Instruction *UMin = getZextedCmpneUMin(ZExt);
+ // Avoid generating an extra UMin
+ if (!UMin || !UMin->hasOneUse())
return std::nullopt;
- auto LhsUMin = isZextedCmpneUMin(UMin->getOperand(1));
- auto RhsUMin = isZextedCmpneUMin(UMin->getOperand(2));
+ Instruction *LhsUMin = getZextedCmpneUMin(UMin->getOperand(1));
+ Instruction *RhsUMin = getZextedCmpneUMin(UMin->getOperand(2));
if (!LhsUMin || !RhsUMin)
return std::nullopt;
IC.Builder.SetInsertPoint(ZExt);
Type *Ty = UMin->getType();
Value *NewUMin = IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_umin_u, Ty,
- {Pg, *LhsUMin, *RhsUMin});
+ {Pg, LhsUMin, RhsUMin});
Value *NewCmpne =
IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_cmpne, Ty,
{Pg, ConstantInt::getNullValue(Ty), NewUMin});
More information about the llvm-commits
mailing list