[llvm] [InstCombine] Fold zext(icmp (A, xxx)) == shr(A, BW - 1) => not(trunc(xor(zext(icmp), shl))) (PR #68244)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Oct 6 05:41:07 PDT 2023
https://github.com/XChy updated https://github.com/llvm/llvm-project/pull/68244
>From 7d5d9f51ac67054bb8a820657c7229c8f4a03bf8 Mon Sep 17 00:00:00 2001
From: XChy <xxs_chy at outlook.com>
Date: Wed, 4 Oct 2023 14:03:06 +0800
Subject: [PATCH 1/2] [InstCombine] Tests for the fold A < 0 == icmp(A)
---
.../InstCombine/icmp-xor-signbit.ll | 90 +++++++++++++++++++
1 file changed, 90 insertions(+)
diff --git a/llvm/test/Transforms/InstCombine/icmp-xor-signbit.ll b/llvm/test/Transforms/InstCombine/icmp-xor-signbit.ll
index 29a18ebbdd94e16..fa820b3981b699e 100644
--- a/llvm/test/Transforms/InstCombine/icmp-xor-signbit.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-xor-signbit.ll
@@ -217,3 +217,93 @@ define <2 x i1> @negative_simplify_splat(<4 x i8> %x) {
ret <2 x i1> %c
}
+
+define i1 @slt_zero_eq_ne_0(i32 %a) {
+; CHECK-LABEL: @slt_zero_eq_ne_0(
+; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 0
+; CHECK-NEXT: [[CONV:%.*]] = zext i1 [[CMP]] to i32
+; CHECK-NEXT: [[CMP1:%.*]] = lshr i32 [[A]], 31
+; CHECK-NEXT: [[CMP2:%.*]] = icmp eq i32 [[CMP1]], [[CONV]]
+; CHECK-NEXT: ret i1 [[CMP2]]
+;
+ %cmp = icmp ne i32 %a, 0
+ %conv = zext i1 %cmp to i32
+ %cmp1 = lshr i32 %a, 31
+ %cmp2 = icmp eq i32 %conv, %cmp1
+ ret i1 %cmp2
+}
+
+define i1 @slt_zero_ne_ne_0(i32 %a) {
+; CHECK-LABEL: @slt_zero_ne_ne_0(
+; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 0
+; CHECK-NEXT: [[CONV:%.*]] = zext i1 [[CMP]] to i32
+; CHECK-NEXT: [[CMP1:%.*]] = lshr i32 [[A]], 31
+; CHECK-NEXT: [[CMP2:%.*]] = icmp ne i32 [[CMP1]], [[CONV]]
+; CHECK-NEXT: ret i1 [[CMP2]]
+;
+ %cmp = icmp ne i32 %a, 0
+ %conv = zext i1 %cmp to i32
+ %cmp1 = lshr i32 %a, 31
+ %cmp2 = icmp ne i32 %conv, %cmp1
+ ret i1 %cmp2
+}
+
+define <4 x i1> @slt_zero_eq_ne_0_vec(<4 x i32> %a) {
+; CHECK-LABEL: @slt_zero_eq_ne_0_vec(
+; CHECK-NEXT: [[CMP:%.*]] = icmp ne <4 x i32> [[A:%.*]], zeroinitializer
+; CHECK-NEXT: [[CONV:%.*]] = zext <4 x i1> [[CMP]] to <4 x i32>
+; CHECK-NEXT: [[CMP1:%.*]] = lshr <4 x i32> [[A]], <i32 31, i32 31, i32 31, i32 31>
+; CHECK-NEXT: [[CMP2:%.*]] = icmp eq <4 x i32> [[CMP1]], [[CONV]]
+; CHECK-NEXT: ret <4 x i1> [[CMP2]]
+;
+ %cmp = icmp ne <4 x i32> %a, zeroinitializer
+ %conv = zext <4 x i1> %cmp to <4 x i32>
+ %cmp1 = lshr <4 x i32> %a, <i32 31, i32 31, i32 31, i32 31>
+ %cmp2 = icmp eq <4 x i32> %conv, %cmp1
+ ret <4 x i1> %cmp2
+}
+
+define i1 @slt_zero_ne_ne_b(i32 %a, i32 %b) {
+; CHECK-LABEL: @slt_zero_ne_ne_b(
+; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT: [[CONV:%.*]] = zext i1 [[CMP]] to i32
+; CHECK-NEXT: [[CMP1:%.*]] = lshr i32 [[A]], 31
+; CHECK-NEXT: [[CMP2:%.*]] = icmp ne i32 [[CMP1]], [[CONV]]
+; CHECK-NEXT: ret i1 [[CMP2]]
+;
+ %cmp = icmp ne i32 %a, %b
+ %conv = zext i1 %cmp to i32
+ %cmp1 = lshr i32 %a, 31
+ %cmp2 = icmp ne i32 %conv, %cmp1
+ ret i1 %cmp2
+}
+
+define i1 @slt_zero_eq_ne_0_fail1(i32 %a) {
+; CHECK-LABEL: @slt_zero_eq_ne_0_fail1(
+; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 0
+; CHECK-NEXT: [[CONV:%.*]] = zext i1 [[CMP]] to i32
+; CHECK-NEXT: [[CMP1:%.*]] = ashr i32 [[A]], 31
+; CHECK-NEXT: [[CMP2:%.*]] = icmp eq i32 [[CMP1]], [[CONV]]
+; CHECK-NEXT: ret i1 [[CMP2]]
+;
+ %cmp = icmp ne i32 %a, 0
+ %conv = zext i1 %cmp to i32
+ %cmp1 = ashr i32 %a, 31
+ %cmp2 = icmp eq i32 %conv, %cmp1
+ ret i1 %cmp2
+}
+
+define i1 @slt_zero_eq_ne_0_fail2(i32 %a) {
+; CHECK-LABEL: @slt_zero_eq_ne_0_fail2(
+; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 0
+; CHECK-NEXT: [[CONV:%.*]] = zext i1 [[CMP]] to i32
+; CHECK-NEXT: [[CMP1:%.*]] = lshr i32 [[A]], 30
+; CHECK-NEXT: [[CMP2:%.*]] = icmp eq i32 [[CMP1]], [[CONV]]
+; CHECK-NEXT: ret i1 [[CMP2]]
+;
+ %cmp = icmp ne i32 %a, 0
+ %conv = zext i1 %cmp to i32
+ %cmp1 = lshr i32 %a, 30
+ %cmp2 = icmp eq i32 %conv, %cmp1
+ ret i1 %cmp2
+}
>From f4c691e5fc873f53be53b711580593df57d9e820 Mon Sep 17 00:00:00 2001
From: XChy <xxs_chy at outlook.com>
Date: Tue, 3 Oct 2023 23:27:02 +0800
Subject: [PATCH 2/2] [InstCombine] Fold A<0 == icmp(A) into
not(xor(A<0,icmp(A)))
---
.../InstCombine/InstCombineCompares.cpp | 44 +++++++++++++------
llvm/test/Transforms/InstCombine/icmp-shr.ll | 7 ++-
.../InstCombine/icmp-xor-signbit.ll | 28 ++++--------
3 files changed, 42 insertions(+), 37 deletions(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index 9f034aba874a8c4..0c75bfdb2f26564 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -5311,11 +5311,7 @@ Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
return new ICmpInst(Pred, A, Builder.CreateTrunc(B, A->getType()));
}
- // Test if 2 values have different or same signbits:
- // (X u>> BitWidth - 1) == zext (Y s> -1) --> (X ^ Y) < 0
- // (X u>> BitWidth - 1) != zext (Y s> -1) --> (X ^ Y) > -1
- // (X s>> BitWidth - 1) == sext (Y s> -1) --> (X ^ Y) < 0
- // (X s>> BitWidth - 1) != sext (Y s> -1) --> (X ^ Y) > -1
+ // Signbit test folds
Instruction *ExtI;
if (match(Op1, m_CombineAnd(m_Instruction(ExtI), m_ZExtOrSExt(m_Value(A)))) &&
(Op0->hasOneUse() || Op1->hasOneUse())) {
@@ -5325,17 +5321,37 @@ Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
ICmpInst::Predicate Pred2;
if (match(Op0, m_CombineAnd(m_Instruction(ShiftI),
m_Shr(m_Value(X),
- m_SpecificIntAllowUndef(OpWidth - 1)))) &&
- match(A, m_ICmp(Pred2, m_Value(Y), m_AllOnes())) &&
- Pred2 == ICmpInst::ICMP_SGT && X->getType() == Y->getType()) {
+ m_SpecificIntAllowUndef(OpWidth - 1))))) {
+ // Test if 2 values have different or same signbits:
+ // (X u>> BitWidth - 1) == zext (Y s> -1) --> (X ^ Y) < 0
+ // (X u>> BitWidth - 1) != zext (Y s> -1) --> (X ^ Y) > -1
+ // (X s>> BitWidth - 1) == sext (Y s> -1) --> (X ^ Y) < 0
+ // (X s>> BitWidth - 1) != sext (Y s> -1) --> (X ^ Y) > -1
unsigned ExtOpc = ExtI->getOpcode();
unsigned ShiftOpc = ShiftI->getOpcode();
- if ((ExtOpc == Instruction::ZExt && ShiftOpc == Instruction::LShr) ||
- (ExtOpc == Instruction::SExt && ShiftOpc == Instruction::AShr)) {
- Value *Xor = Builder.CreateXor(X, Y, "xor.signbits");
- Value *R = (Pred == ICmpInst::ICMP_EQ) ? Builder.CreateIsNeg(Xor)
- : Builder.CreateIsNotNeg(Xor);
- return replaceInstUsesWith(I, R);
+
+ if (match(A, m_ICmp(Pred2, m_Value(Y), m_AllOnes())) &&
+ Pred2 == ICmpInst::ICMP_SGT && X->getType() == Y->getType()) {
+ if ((ExtOpc == Instruction::ZExt && ShiftOpc == Instruction::LShr) ||
+ (ExtOpc == Instruction::SExt && ShiftOpc == Instruction::AShr)) {
+ Value *Xor = Builder.CreateXor(X, Y, "xor.signbits");
+ Value *R = (Pred == ICmpInst::ICMP_EQ) ? Builder.CreateIsNeg(Xor)
+ : Builder.CreateIsNotNeg(Xor);
+ return replaceInstUsesWith(I, R);
+ }
+ }
+
+ // Transform (X u>> BitWidth - 1 == zext(i1)) into X s< 0 == i1
+ // Transform (X u>> BitWidth - 1 != zext(i1)) into X s< 0 != i1
+ if (A->getType()->getScalarSizeInBits() == 1 && Op0->hasOneUse() &&
+ Op1->hasOneUse() && ExtOpc == Instruction::ZExt &&
+ ShiftOpc == Instruction::LShr) {
+
+ Value *SLTZero =
+ Builder.CreateICmpSLT(X, Constant::getNullValue(X->getType()));
+ SLTZero->takeName(Op0);
+ Value *Eq = Builder.CreateICmp(Pred, SLTZero, A, I.getName());
+ return replaceInstUsesWith(I, Eq);
}
}
}
diff --git a/llvm/test/Transforms/InstCombine/icmp-shr.ll b/llvm/test/Transforms/InstCombine/icmp-shr.ll
index f4dfa2edfa17710..a8466202c4a8554 100644
--- a/llvm/test/Transforms/InstCombine/icmp-shr.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-shr.ll
@@ -1397,11 +1397,10 @@ define <2 x i1> @same_signbit_poison_elts(<2 x i8> %x, <2 x i8> %y) {
define i1 @same_signbit_wrong_type(i8 %x, i32 %y) {
; CHECK-LABEL: @same_signbit_wrong_type(
-; CHECK-NEXT: [[XSIGN:%.*]] = lshr i8 [[X:%.*]], 7
; CHECK-NEXT: [[YPOS:%.*]] = icmp sgt i32 [[Y:%.*]], -1
-; CHECK-NEXT: [[YPOSZ:%.*]] = zext i1 [[YPOS]] to i8
-; CHECK-NEXT: [[R:%.*]] = icmp ne i8 [[XSIGN]], [[YPOSZ]]
-; CHECK-NEXT: ret i1 [[R]]
+; CHECK-NEXT: [[XSIGN:%.*]] = icmp slt i8 [[X:%.*]], 0
+; CHECK-NEXT: [[R1:%.*]] = xor i1 [[XSIGN]], [[YPOS]]
+; CHECK-NEXT: ret i1 [[R1]]
;
%xsign = lshr i8 %x, 7
%ypos = icmp sgt i32 %y, -1
diff --git a/llvm/test/Transforms/InstCombine/icmp-xor-signbit.ll b/llvm/test/Transforms/InstCombine/icmp-xor-signbit.ll
index fa820b3981b699e..652d82f2c4e3bcb 100644
--- a/llvm/test/Transforms/InstCombine/icmp-xor-signbit.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-xor-signbit.ll
@@ -220,11 +220,8 @@ define <2 x i1> @negative_simplify_splat(<4 x i8> %x) {
define i1 @slt_zero_eq_ne_0(i32 %a) {
; CHECK-LABEL: @slt_zero_eq_ne_0(
-; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 0
-; CHECK-NEXT: [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT: [[CMP1:%.*]] = lshr i32 [[A]], 31
-; CHECK-NEXT: [[CMP2:%.*]] = icmp eq i32 [[CMP1]], [[CONV]]
-; CHECK-NEXT: ret i1 [[CMP2]]
+; CHECK-NEXT: [[TMP1:%.*]] = icmp slt i32 [[A:%.*]], 1
+; CHECK-NEXT: ret i1 [[TMP1]]
;
%cmp = icmp ne i32 %a, 0
%conv = zext i1 %cmp to i32
@@ -235,11 +232,8 @@ define i1 @slt_zero_eq_ne_0(i32 %a) {
define i1 @slt_zero_ne_ne_0(i32 %a) {
; CHECK-LABEL: @slt_zero_ne_ne_0(
-; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[A:%.*]], 0
-; CHECK-NEXT: [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT: [[CMP1:%.*]] = lshr i32 [[A]], 31
-; CHECK-NEXT: [[CMP2:%.*]] = icmp ne i32 [[CMP1]], [[CONV]]
-; CHECK-NEXT: ret i1 [[CMP2]]
+; CHECK-NEXT: [[CMP21:%.*]] = icmp sgt i32 [[A:%.*]], 0
+; CHECK-NEXT: ret i1 [[CMP21]]
;
%cmp = icmp ne i32 %a, 0
%conv = zext i1 %cmp to i32
@@ -250,11 +244,8 @@ define i1 @slt_zero_ne_ne_0(i32 %a) {
define <4 x i1> @slt_zero_eq_ne_0_vec(<4 x i32> %a) {
; CHECK-LABEL: @slt_zero_eq_ne_0_vec(
-; CHECK-NEXT: [[CMP:%.*]] = icmp ne <4 x i32> [[A:%.*]], zeroinitializer
-; CHECK-NEXT: [[CONV:%.*]] = zext <4 x i1> [[CMP]] to <4 x i32>
-; CHECK-NEXT: [[CMP1:%.*]] = lshr <4 x i32> [[A]], <i32 31, i32 31, i32 31, i32 31>
-; CHECK-NEXT: [[CMP2:%.*]] = icmp eq <4 x i32> [[CMP1]], [[CONV]]
-; CHECK-NEXT: ret <4 x i1> [[CMP2]]
+; CHECK-NEXT: [[TMP1:%.*]] = icmp slt <4 x i32> [[A:%.*]], <i32 1, i32 1, i32 1, i32 1>
+; CHECK-NEXT: ret <4 x i1> [[TMP1]]
;
%cmp = icmp ne <4 x i32> %a, zeroinitializer
%conv = zext <4 x i1> %cmp to <4 x i32>
@@ -266,10 +257,9 @@ define <4 x i1> @slt_zero_eq_ne_0_vec(<4 x i32> %a) {
define i1 @slt_zero_ne_ne_b(i32 %a, i32 %b) {
; CHECK-LABEL: @slt_zero_ne_ne_b(
; CHECK-NEXT: [[CMP:%.*]] = icmp ne i32 [[A:%.*]], [[B:%.*]]
-; CHECK-NEXT: [[CONV:%.*]] = zext i1 [[CMP]] to i32
-; CHECK-NEXT: [[CMP1:%.*]] = lshr i32 [[A]], 31
-; CHECK-NEXT: [[CMP2:%.*]] = icmp ne i32 [[CMP1]], [[CONV]]
-; CHECK-NEXT: ret i1 [[CMP2]]
+; CHECK-NEXT: [[CMP1:%.*]] = icmp slt i32 [[A]], 0
+; CHECK-NEXT: [[CMP21:%.*]] = xor i1 [[CMP1]], [[CMP]]
+; CHECK-NEXT: ret i1 [[CMP21]]
;
%cmp = icmp ne i32 %a, %b
%conv = zext i1 %cmp to i32
More information about the llvm-commits
mailing list