[llvm] [InstCombine] Fold (X << Y) / (X << Z) -> 1 << Y >> Z (PR #68863)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Oct 13 07:47:36 PDT 2023
https://github.com/XChy updated https://github.com/llvm/llvm-project/pull/68863
>From a9b9f4046e1cd021351be56c6966429e0710c4af Mon Sep 17 00:00:00 2001
From: XChy <xxs_chy at outlook.com>
Date: Thu, 12 Oct 2023 15:42:21 +0800
Subject: [PATCH 1/2] [InstCombine][NFC] Precommit tests for PR68857
---
llvm/test/Transforms/InstCombine/div-shift.ll | 251 ++++++++++++++++++
1 file changed, 251 insertions(+)
diff --git a/llvm/test/Transforms/InstCombine/div-shift.ll b/llvm/test/Transforms/InstCombine/div-shift.ll
index 76c5328dc8499e0..883cc742e013be4 100644
--- a/llvm/test/Transforms/InstCombine/div-shift.ll
+++ b/llvm/test/Transforms/InstCombine/div-shift.ll
@@ -2,6 +2,7 @@
; RUN: opt < %s -passes=instcombine -S | FileCheck %s
declare void @use(i8)
+declare void @use32(i32)
declare i8 @llvm.umin.i8(i8, i8)
declare i8 @llvm.umax.i8(i8, i8)
@@ -1025,3 +1026,253 @@ define i8 @udiv_shl_no_overflow(i8 %x, i8 %y) {
%mul = udiv i8 %x, %min
ret i8 %mul
}
+
+; (X<<Y) / (X<<Z) -> 1 << Y >> Z
+
+define i32 @sdiv_shl_pair_const(i32 %a) {
+; CHECK-LABEL: @sdiv_shl_pair_const(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nsw i32 [[A:%.*]], 2
+; CHECK-NEXT: [[RHS:%.*]] = shl nsw i32 [[A]], 1
+; CHECK-NEXT: [[DIV:%.*]] = sdiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nsw i32 %a, 2
+ %rhs = shl nsw i32 %a, 1
+ %div = sdiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @udiv_shl_pair_const(i32 %a) {
+; CHECK-LABEL: @udiv_shl_pair_const(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nuw i32 [[A:%.*]], 2
+; CHECK-NEXT: [[RHS:%.*]] = shl nuw i32 [[A]], 1
+; CHECK-NEXT: [[DIV:%.*]] = udiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nuw i32 %a, 2
+ %rhs = shl nuw i32 %a, 1
+ %div = udiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @sdiv_shl_pair1(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @sdiv_shl_pair1(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nsw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl nuw nsw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = sdiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nsw i32 %a, %x
+ %rhs = shl nuw nsw i32 %a, %y
+ %div = sdiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @sdiv_shl_pair2(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @sdiv_shl_pair2(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nuw nsw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl nsw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = sdiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nuw nsw i32 %a, %x
+ %rhs = shl nsw i32 %a, %y
+ %div = sdiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @sdiv_shl_pair3(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @sdiv_shl_pair3(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nsw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl nsw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = sdiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nsw i32 %a, %x
+ %rhs = shl nsw i32 %a, %y
+ %div = sdiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @udiv_shl_pair1(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @udiv_shl_pair1(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nuw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl nuw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = udiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nuw i32 %a, %x
+ %rhs = shl nuw i32 %a, %y
+ %div = udiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @udiv_shl_pair2(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @udiv_shl_pair2(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nuw nsw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl nuw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = udiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nuw nsw i32 %a, %x
+ %rhs = shl nuw i32 %a, %y
+ %div = udiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @udiv_shl_pair3(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @udiv_shl_pair3(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nuw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl nuw nsw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = udiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nuw i32 %a, %x
+ %rhs = shl nuw nsw i32 %a, %y
+ %div = udiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @sdiv_shl_pair_overflow_fail1(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @sdiv_shl_pair_overflow_fail1(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl nsw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = sdiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl i32 %a, %x
+ %rhs = shl nsw i32 %a, %y
+ %div = sdiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @sdiv_shl_pair_overflow_fail2(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @sdiv_shl_pair_overflow_fail2(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nsw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl nuw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = sdiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nsw i32 %a, %x
+ %rhs = shl nuw i32 %a, %y
+ %div = sdiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @udiv_shl_pair_overflow_fail1(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @udiv_shl_pair_overflow_fail1(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nsw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl nuw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = udiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nsw i32 %a, %x
+ %rhs = shl nuw i32 %a, %y
+ %div = udiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @udiv_shl_pair_overflow_fail2(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @udiv_shl_pair_overflow_fail2(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nsw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = udiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nsw i32 %a, %x
+ %rhs = shl i32 %a, %y
+ %div = udiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @udiv_shl_pair_overflow_fail3(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @udiv_shl_pair_overflow_fail3(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nuw nsw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = udiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nuw nsw i32 %a, %x
+ %rhs = shl i32 %a, %y
+ %div = udiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @sdiv_shl_pair_multiuse1(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @sdiv_shl_pair_multiuse1(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nuw nsw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: call void @use32(i32 [[LHS]])
+; CHECK-NEXT: [[RHS:%.*]] = shl nsw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: [[DIV:%.*]] = sdiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nuw nsw i32 %a, %x
+ call void @use32(i32 %lhs)
+ %rhs = shl nsw i32 %a, %y
+ %div = sdiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @sdiv_shl_pair_multiuse2(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @sdiv_shl_pair_multiuse2(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nuw nsw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl nsw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: call void @use32(i32 [[RHS]])
+; CHECK-NEXT: [[DIV:%.*]] = sdiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nuw nsw i32 %a, %x
+ %rhs = shl nsw i32 %a, %y
+ call void @use32(i32 %rhs)
+ %div = sdiv i32 %lhs, %rhs
+ ret i32 %div
+}
+
+define i32 @sdiv_shl_pair_multiuse3(i32 %a, i32 %x, i32 %y) {
+; CHECK-LABEL: @sdiv_shl_pair_multiuse3(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[LHS:%.*]] = shl nuw nsw i32 [[A:%.*]], [[X:%.*]]
+; CHECK-NEXT: [[RHS:%.*]] = shl nsw i32 [[A]], [[Y:%.*]]
+; CHECK-NEXT: call void @use32(i32 [[LHS]])
+; CHECK-NEXT: call void @use32(i32 [[RHS]])
+; CHECK-NEXT: [[DIV:%.*]] = sdiv i32 [[LHS]], [[RHS]]
+; CHECK-NEXT: ret i32 [[DIV]]
+;
+entry:
+ %lhs = shl nuw nsw i32 %a, %x
+ %rhs = shl nsw i32 %a, %y
+ call void @use32(i32 %lhs)
+ call void @use32(i32 %rhs)
+ %div = sdiv i32 %lhs, %rhs
+ ret i32 %div
+}
>From b6b2d5095329374dfd250c1844ddb46f623a695e Mon Sep 17 00:00:00 2001
From: XChy <xxs_chy at outlook.com>
Date: Thu, 12 Oct 2023 15:12:11 +0800
Subject: [PATCH 2/2] [InstCombine] Fold (A<<X)/(A<<Y) -> (1<<X)/(1<<Y)
---
.../InstCombine/InstCombineMulDivRem.cpp | 22 +++++++++++++++++++
1 file changed, 22 insertions(+)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
index 560c87b6efa7038..d727b517bb90e97 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -980,6 +980,28 @@ static Instruction *foldIDivShl(BinaryOperator &I,
Ret = BinaryOperator::CreateSDiv(X, Y);
}
+ // If X << Y and X << Z does not overflow, then:
+ // (X << Y) / (X << Z) -> (1 << Y) / (1 << Z) -> 1 << Y >> Z
+ if (match(Op0, m_Shl(m_Value(X), m_Value(Y))) &&
+ match(Op1, m_Shl(m_Specific(X), m_Value(Z)))) {
+ auto *Shl0 = cast<OverflowingBinaryOperator>(Op0);
+ auto *Shl1 = cast<OverflowingBinaryOperator>(Op1);
+
+ if ((!IsSigned && Shl0->hasNoUnsignedWrap() && Shl1->hasNoUnsignedWrap()) ||
+ (IsSigned && Shl0->hasNoSignedWrap() && Shl1->hasNoSignedWrap())) {
+ Constant *One = ConstantInt::get(X->getType(), 1);
+ // Only preserve the nsw flag if dividend has nsw
+ // or divisor has nsw and operator is sdiv.
+ Value *Dividend = Builder.CreateShl(
+ One, Y, "shl.dividend",
+ /*HasNUW*/ true,
+ /*HasNSW*/
+ IsSigned ? (Shl0->hasNoUnsignedWrap() || Shl1->hasNoUnsignedWrap())
+ : Shl0->hasNoSignedWrap());
+ Ret = BinaryOperator::CreateLShr(Dividend, Z);
+ }
+ }
+
if (!Ret)
return nullptr;
More information about the llvm-commits
mailing list