[llvm] [ValueTracking] Add support for non-splat vecs in cmpExcludesZero (PR #68331)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Oct 6 09:57:38 PDT 2023
https://github.com/goldsteinn updated https://github.com/llvm/llvm-project/pull/68331
>From b6c37124ebe0ca9ac9820582a62078d405e26703 Mon Sep 17 00:00:00 2001
From: Noah Goldstein <goldstein.w.n at gmail.com>
Date: Thu, 5 Oct 2023 11:17:14 -0500
Subject: [PATCH 1/2] [ValueTracking] Add tests for `cmpExcludesZero` for
non-splat vecs; NFC
---
.../Analysis/ValueTracking/known-non-zero.ll | 62 +++++++++++++++++++
1 file changed, 62 insertions(+)
diff --git a/llvm/test/Analysis/ValueTracking/known-non-zero.ll b/llvm/test/Analysis/ValueTracking/known-non-zero.ll
index 6dce6e528165ea6..f64303f17301504 100644
--- a/llvm/test/Analysis/ValueTracking/known-non-zero.ll
+++ b/llvm/test/Analysis/ValueTracking/known-non-zero.ll
@@ -1160,3 +1160,65 @@ define i1 @sdiv_known_non_zero_fail(i8 %x, i8 %y) {
%nz = icmp ne i8 %xy, 0
ret i1 %nz
}
+
+define <2 x i1> @cmp_excludes_zero_with_nonsplat_vec(<2 x i8> %a, <2 x i8> %b) {
+; CHECK-LABEL: @cmp_excludes_zero_with_nonsplat_vec(
+; CHECK-NEXT: [[C:%.*]] = icmp sge <2 x i8> [[A:%.*]], <i8 1, i8 4>
+; CHECK-NEXT: [[S:%.*]] = select <2 x i1> [[C]], <2 x i8> [[A]], <2 x i8> <i8 4, i8 5>
+; CHECK-NEXT: [[AND:%.*]] = or <2 x i8> [[S]], [[B:%.*]]
+; CHECK-NEXT: [[R:%.*]] = icmp eq <2 x i8> [[AND]], zeroinitializer
+; CHECK-NEXT: ret <2 x i1> [[R]]
+;
+ %c = icmp sge <2 x i8> %a, <i8 1, i8 4>
+ %s = select <2 x i1> %c, <2 x i8> %a, <2 x i8> <i8 4, i8 5>
+ %and = or <2 x i8> %s, %b
+ %r = icmp eq <2 x i8> %and, zeroinitializer
+ ret <2 x i1> %r
+}
+
+define <2 x i1> @cmp_excludes_zero_with_nonsplat_vec_wundef(<2 x i8> %a, <2 x i8> %b) {
+; CHECK-LABEL: @cmp_excludes_zero_with_nonsplat_vec_wundef(
+; CHECK-NEXT: [[C:%.*]] = icmp sge <2 x i8> [[A:%.*]], <i8 1, i8 undef>
+; CHECK-NEXT: [[S:%.*]] = select <2 x i1> [[C]], <2 x i8> [[A]], <2 x i8> <i8 4, i8 5>
+; CHECK-NEXT: [[AND:%.*]] = or <2 x i8> [[S]], [[B:%.*]]
+; CHECK-NEXT: [[R:%.*]] = icmp eq <2 x i8> [[AND]], zeroinitializer
+; CHECK-NEXT: ret <2 x i1> [[R]]
+;
+ %c = icmp sge <2 x i8> %a, <i8 1, i8 undef>
+ %s = select <2 x i1> %c, <2 x i8> %a, <2 x i8> <i8 4, i8 5>
+ %and = or <2 x i8> %s, %b
+ %r = icmp eq <2 x i8> %and, zeroinitializer
+ ret <2 x i1> %r
+}
+
+define <2 x i1> @cmp_excludes_zero_with_nonsplat_vec_wpoison(<2 x i8> %a, <2 x i8> %b) {
+; CHECK-LABEL: @cmp_excludes_zero_with_nonsplat_vec_wpoison(
+; CHECK-NEXT: [[C:%.*]] = icmp sge <2 x i8> [[A:%.*]], <i8 1, i8 poison>
+; CHECK-NEXT: [[S:%.*]] = select <2 x i1> [[C]], <2 x i8> [[A]], <2 x i8> <i8 4, i8 5>
+; CHECK-NEXT: [[AND:%.*]] = or <2 x i8> [[S]], [[B:%.*]]
+; CHECK-NEXT: [[R:%.*]] = icmp eq <2 x i8> [[AND]], zeroinitializer
+; CHECK-NEXT: ret <2 x i1> [[R]]
+;
+ %c = icmp sge <2 x i8> %a, <i8 1, i8 poison>
+ %s = select <2 x i1> %c, <2 x i8> %a, <2 x i8> <i8 4, i8 5>
+ %and = or <2 x i8> %s, %b
+ %r = icmp eq <2 x i8> %and, zeroinitializer
+ ret <2 x i1> %r
+}
+
+
+define <2 x i1> @cmp_excludes_zero_with_nonsplat_vec_fail(<2 x i8> %a, <2 x i8> %b) {
+; CHECK-LABEL: @cmp_excludes_zero_with_nonsplat_vec_fail(
+; CHECK-NEXT: [[C:%.*]] = icmp sge <2 x i8> [[A:%.*]], <i8 0, i8 4>
+; CHECK-NEXT: [[S:%.*]] = select <2 x i1> [[C]], <2 x i8> [[A]], <2 x i8> <i8 4, i8 5>
+; CHECK-NEXT: [[AND:%.*]] = or <2 x i8> [[S]], [[B:%.*]]
+; CHECK-NEXT: [[R:%.*]] = icmp eq <2 x i8> [[AND]], zeroinitializer
+; CHECK-NEXT: ret <2 x i1> [[R]]
+;
+ %c = icmp sge <2 x i8> %a, <i8 0, i8 4>
+ %s = select <2 x i1> %c, <2 x i8> %a, <2 x i8> <i8 4, i8 5>
+ %and = or <2 x i8> %s, %b
+ %r = icmp eq <2 x i8> %and, zeroinitializer
+ ret <2 x i1> %r
+}
+
>From 30119963f13026b4d086f64ee583f27fe87d8822 Mon Sep 17 00:00:00 2001
From: Noah Goldstein <goldstein.w.n at gmail.com>
Date: Thu, 5 Oct 2023 21:16:34 -0500
Subject: [PATCH 2/2] [ValueTracking] Add support for non-splat vecs in
cmpExcludesZero Just a small QOL change.
---
llvm/lib/Analysis/ValueTracking.cpp | 21 +++++++++++++++----
.../Analysis/ValueTracking/known-non-zero.ll | 6 +-----
2 files changed, 18 insertions(+), 9 deletions(-)
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index b76becf24d10fc9..f4de1c4a1f22ad0 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -558,7 +558,7 @@ bool llvm::isValidAssumeForContext(const Instruction *Inv,
// example Pred=EQ, RHS=isKnownNonZero. cmpExcludesZero is called in loops
// so the extra compile time may not be worth it, but possibly a second API
// should be created for use outside of loops.
-static bool cmpExcludesZero(CmpInst::Predicate Pred, const Value *RHS) {
+static bool cmpExcludesZero(CmpInst::Predicate Pred, Value *RHS) {
// v u> y implies v != 0.
if (Pred == ICmpInst::ICMP_UGT)
return true;
@@ -569,11 +569,24 @@ static bool cmpExcludesZero(CmpInst::Predicate Pred, const Value *RHS) {
// All other predicates - rely on generic ConstantRange handling.
const APInt *C;
- if (!match(RHS, m_APInt(C)))
+ auto Zero = APInt::getZero(RHS->getType()->getScalarSizeInBits());
+ if (match(RHS, m_APInt(C))) {
+ ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(Pred, *C);
+ return !TrueValues.contains(Zero);
+ }
+
+ auto *VC = dyn_cast<ConstantDataVector>(RHS);
+ if (VC == nullptr)
return false;
- ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(Pred, *C);
- return !TrueValues.contains(APInt::getZero(C->getBitWidth()));
+ for (unsigned EleIdx = 0, NEle = VC->getNumElements(); EleIdx < NEle;
+ ++EleIdx) {
+ ConstantRange TrueValues =
+ ConstantRange::makeExactICmpRegion(Pred, VC->getElementAsAPInt(EleIdx));
+ if (TrueValues.contains(Zero))
+ return false;
+ }
+ return true;
}
static bool isKnownNonZeroFromAssume(const Value *V, const SimplifyQuery &Q) {
diff --git a/llvm/test/Analysis/ValueTracking/known-non-zero.ll b/llvm/test/Analysis/ValueTracking/known-non-zero.ll
index f64303f17301504..dbec47ea0ae261a 100644
--- a/llvm/test/Analysis/ValueTracking/known-non-zero.ll
+++ b/llvm/test/Analysis/ValueTracking/known-non-zero.ll
@@ -1163,11 +1163,7 @@ define i1 @sdiv_known_non_zero_fail(i8 %x, i8 %y) {
define <2 x i1> @cmp_excludes_zero_with_nonsplat_vec(<2 x i8> %a, <2 x i8> %b) {
; CHECK-LABEL: @cmp_excludes_zero_with_nonsplat_vec(
-; CHECK-NEXT: [[C:%.*]] = icmp sge <2 x i8> [[A:%.*]], <i8 1, i8 4>
-; CHECK-NEXT: [[S:%.*]] = select <2 x i1> [[C]], <2 x i8> [[A]], <2 x i8> <i8 4, i8 5>
-; CHECK-NEXT: [[AND:%.*]] = or <2 x i8> [[S]], [[B:%.*]]
-; CHECK-NEXT: [[R:%.*]] = icmp eq <2 x i8> [[AND]], zeroinitializer
-; CHECK-NEXT: ret <2 x i1> [[R]]
+; CHECK-NEXT: ret <2 x i1> zeroinitializer
;
%c = icmp sge <2 x i8> %a, <i8 1, i8 4>
%s = select <2 x i1> %c, <2 x i8> %a, <2 x i8> <i8 4, i8 5>
More information about the llvm-commits
mailing list