[InstCombine] Fold umax(nuw_mul(x, C0), x + 1) into (x == 0 ? 1 : nuw_mul(x, C0)) #123468
+381
−0
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-llvm-transforms Author: None (Ruhung) ChangesThis PR introduces the following transformations:
Fixes : #122388 Full diff: https://github.com/llvm/llvm-project/pull/123468.diff 2 Files Affected:
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index c55c40c88bc845..0db7d8818fbd0b 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -1847,6 +1847,37 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
return CastInst::Create(Instruction::ZExt, NarrowMaxMin, II->getType());
}
}
+ // If C0 is not 0:
+ // umax(nuw_shl(x, C0), x + 1) -> x == 0 ? 1 : nuw_shl(x, C0)
+ // If C0 is not 0 or 1:
+ // umax(nuw_mul(x, C0), x + 1) -> x == 0 ? 1 : nuw_mul(x, C0)
+ ConstantInt *C0;
+ bool isShl = false;
+ BinaryOperator *Op = nullptr;
+ auto matchShiftOrMul = [&](Value *I) {
+ if (match(I, m_OneUse(m_NUWShl(m_Value(X), m_ConstantInt(C0))))) {
+ isShl = true;
+ return true;
+ } else if (match(I, m_OneUse(m_NUWMul(m_Value(X), m_ConstantInt(C0)))) &&
+ C0 && !C0->isOne()) {
+ isShl = false;
+ return true;
+ }
+ return false;
+ };
+ if (((matchShiftOrMul(I0) &&
+ match(I1, m_OneUse(m_Add(m_Specific(X), m_One())))) ||
+ (matchShiftOrMul(I1) &&
+ match(I0, m_OneUse(m_Add(m_Specific(X), m_One()))))) &&
+ C0 && !C0->isZero()) {
+ Op = isShl ? BinaryOperator::CreateNUWShl(X, C0)
+ : BinaryOperator::CreateNUWMul(X, C0);
+ Builder.Insert(Op);
+ Value *Cmp = Builder.CreateICmpEQ(X, ConstantInt::get(X->getType(), 0));
+ Value *NewSelect =
+ Builder.CreateSelect(Cmp, ConstantInt::get(X->getType(), 1), Op);
+ return replaceInstUsesWith(*II, NewSelect);
+ }
// If both operands of unsigned min/max are sign-extended, it is still ok
// to narrow the operation.
[[fallthrough]];
diff --git a/llvm/test/Transforms/InstCombine/add-shl-mul-umax.ll b/llvm/test/Transforms/InstCombine/add-shl-mul-umax.ll
new file mode 100644
index 00000000000000..86ce5dc06ce031
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/add-shl-mul-umax.ll
@@ -0,0 +1,300 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -passes=instcombine < %s | FileCheck %s
+
+; When C0 is neither 0 nor 1:
+; umax(nuw_mul(x, C0), x + 1) is optimized to:
+; x == 0 ? 1 : nuw_mul(x, C0)
+; When C0 is not 0:
+; umax(nuw_shl(x, C0), x + 1) is optimized to:
+; x == 0 ? 1 : nuw_shl(x, C0)
+
+; Positive Test Cases for `shl`
+
+define i64 @test_shl_by_2(i64 %x) {
+; CHECK-LABEL: define i64 @test_shl_by_2(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[TMP2:%.*]] = shl nuw i64 [[X]], 2
+; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[X]], 0
+; CHECK-NEXT: [[MAX:%.*]] = select i1 [[TMP3]], i64 1, i64 [[TMP2]]
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %shl = shl nuw i64 %x, 2
+ %max = call i64 @llvm.umax.i64(i64 %shl, i64 %x1)
+ ret i64 %max
+}
+
+define i64 @test_shl_by_5(i64 %x) {
+; CHECK-LABEL: define i64 @test_shl_by_5(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[TMP2:%.*]] = shl nuw i64 [[X]], 5
+; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[X]], 0
+; CHECK-NEXT: [[MAX:%.*]] = select i1 [[TMP3]], i64 1, i64 [[TMP2]]
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %shl = shl nuw i64 %x, 5
+ %max = call i64 @llvm.umax.i64(i64 %shl, i64 %x1)
+ ret i64 %max
+}
+
+; Commuted Test Cases for `shl`
+
+define i64 @test_shl_umax_commuted(i64 %x) {
+; CHECK-LABEL: define i64 @test_shl_umax_commuted(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[SHL:%.*]] = shl nuw i64 [[X]], 2
+; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i64 [[X]], 0
+; CHECK-NEXT: [[MAX:%.*]] = select i1 [[TMP2]], i64 1, i64 [[SHL]]
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %shl = shl nuw i64 %x, 2
+ %max = call i64 @llvm.umax.i64(i64 %x1, i64 %shl)
+ ret i64 %max
+}
+
+; Negative Test Cases for `shl`
+
+define i64 @test_shl_by_zero(i64 %x) {
+; CHECK-LABEL: define i64 @test_shl_by_zero(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[X1:%.*]] = add i64 [[X]], 1
+; CHECK-NEXT: [[MAX:%.*]] = call i64 @llvm.umax.i64(i64 [[X]], i64 [[X1]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %shl = shl nuw i64 %x, 0
+ %max = call i64 @llvm.umax.i64(i64 %shl, i64 %x1)
+ ret i64 %max
+}
+
+define i64 @test_shl_add_by_2(i64 %x) {
+; CHECK-LABEL: define i64 @test_shl_add_by_2(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[X1:%.*]] = add i64 [[X]], 2
+; CHECK-NEXT: [[SHL:%.*]] = shl nuw i64 [[X]], 2
+; CHECK-NEXT: [[MAX:%.*]] = call i64 @llvm.umax.i64(i64 [[SHL]], i64 [[X1]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 2
+ %shl = shl nuw i64 %x, 2
+ %max = call i64 @llvm.umax.i64(i64 %shl, i64 %x1)
+ ret i64 %max
+}
+
+define i64 @test_shl_without_nuw(i64 %x) {
+; CHECK-LABEL: define i64 @test_shl_without_nuw(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[X1:%.*]] = add i64 [[X]], 1
+; CHECK-NEXT: [[SHL:%.*]] = shl i64 [[X]], 2
+; CHECK-NEXT: [[MAX:%.*]] = call i64 @llvm.umax.i64(i64 [[SHL]], i64 [[X1]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %shl = shl i64 %x, 2
+ %max = call i64 @llvm.umax.i64(i64 %shl, i64 %x1)
+ ret i64 %max
+}
+
+; Multi-use Test Cases for `shl`
+declare void @use(i64)
+
+define i64 @test_shl_multi_use_add(i64 %x) {
+; CHECK-LABEL: define i64 @test_shl_multi_use_add(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[X1:%.*]] = add i64 [[X]], 1
+; CHECK-NEXT: call void @use(i64 [[X1]])
+; CHECK-NEXT: [[TMP2:%.*]] = shl nuw i64 [[X]], 3
+; CHECK-NEXT: [[MAX:%.*]] = call i64 @llvm.umax.i64(i64 [[TMP2]], i64 [[X1]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ call void @use(i64 %x1)
+ %shl = shl nuw i64 %x, 3
+ %max = call i64 @llvm.umax.i64(i64 %shl, i64 %x1)
+ ret i64 %max
+}
+
+define i64 @test_shl_multi_use_shl(i64 %x) {
+; CHECK-LABEL: define i64 @test_shl_multi_use_shl(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[X1:%.*]] = add i64 [[X]], 1
+; CHECK-NEXT: [[SHL:%.*]] = shl nuw i64 [[X]], 2
+; CHECK-NEXT: call void @use(i64 [[SHL]])
+; CHECK-NEXT: [[MAX:%.*]] = call i64 @llvm.umax.i64(i64 [[SHL]], i64 [[X1]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %shl = shl nuw i64 %x, 2
+ call void @use(i64 %shl)
+ %max = call i64 @llvm.umax.i64(i64 %shl, i64 %x1)
+ ret i64 %max
+}
+
+define i64 @test_shl_multi_use_max(i64 %x) {
+; CHECK-LABEL: define i64 @test_shl_multi_use_max(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[TMP2:%.*]] = shl nuw i64 [[X]], 3
+; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[X]], 0
+; CHECK-NEXT: [[MAX:%.*]] = select i1 [[TMP3]], i64 1, i64 [[TMP2]]
+; CHECK-NEXT: call void @use(i64 [[MAX]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %shl = shl nuw i64 %x, 3
+ %max = call i64 @llvm.umax.i64(i64 %shl, i64 %x1)
+ call void @use(i64 %max)
+ ret i64 %max
+}
+
+; Positive Test Cases for `mul`
+
+define i64 @test_mul_by_2(i64 %x) {
+; CHECK-LABEL: define i64 @test_mul_by_2(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[TMP2:%.*]] = shl nuw i64 [[X]], 1
+; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[X]], 0
+; CHECK-NEXT: [[MAX:%.*]] = select i1 [[TMP3]], i64 1, i64 [[TMP2]]
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %mul = mul nuw i64 %x, 2
+ %max = call i64 @llvm.umax.i64(i64 %mul, i64 %x1)
+ ret i64 %max
+}
+
+define i64 @test_mul_by_5(i64 %x) {
+; CHECK-LABEL: define i64 @test_mul_by_5(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[MUL:%.*]] = mul nuw i64 [[X]], 5
+; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i64 [[X]], 0
+; CHECK-NEXT: [[MAX:%.*]] = select i1 [[TMP2]], i64 1, i64 [[MUL]]
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %mul = mul nuw i64 %x, 5
+ %max = call i64 @llvm.umax.i64(i64 %mul, i64 %x1)
+ ret i64 %max
+}
+
+; Commuted Test Cases for `mul`
+
+define i64 @test_mul_max_commuted(i64 %x) {
+; CHECK-LABEL: define i64 @test_mul_max_commuted(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[MUL:%.*]] = shl nuw i64 [[X]], 1
+; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i64 [[X]], 0
+; CHECK-NEXT: [[MAX:%.*]] = select i1 [[TMP2]], i64 1, i64 [[MUL]]
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %mul = mul nuw i64 %x, 2
+ %max = call i64 @llvm.umax.i64(i64 %x1, i64 %mul)
+ ret i64 %max
+}
+
+; Negative Test Cases for `mul`
+
+define i64 @test_mul_by_zero(i64 %x) {
+; CHECK-LABEL: define i64 @test_mul_by_zero(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[X1:%.*]] = add i64 [[X]], 1
+; CHECK-NEXT: ret i64 [[X1]]
+;
+ %x1 = add i64 %x, 1
+ %mul = mul nuw i64 %x, 0
+ %max = call i64 @llvm.umax.i64(i64 %mul, i64 %x1)
+ ret i64 %max
+}
+
+define i64 @test_mul_by_1(i64 %x) {
+; CHECK-LABEL: define i64 @test_mul_by_1(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[X1:%.*]] = add i64 [[X]], 1
+; CHECK-NEXT: [[MAX:%.*]] = call i64 @llvm.umax.i64(i64 [[X]], i64 [[X1]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %mul = mul nuw i64 %x, 1
+ %max = call i64 @llvm.umax.i64(i64 %mul, i64 %x1)
+ ret i64 %max
+}
+
+define i64 @test_mul_add_by_2(i64 %x) {
+; CHECK-LABEL: define i64 @test_mul_add_by_2(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[X1:%.*]] = add i64 [[X]], 2
+; CHECK-NEXT: [[MUL:%.*]] = shl nuw i64 [[X]], 1
+; CHECK-NEXT: [[MAX:%.*]] = call i64 @llvm.umax.i64(i64 [[MUL]], i64 [[X1]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 2
+ %mul = mul nuw i64 %x, 2
+ %max = call i64 @llvm.umax.i64(i64 %mul, i64 %x1)
+ ret i64 %max
+}
+
+define i64 @test_mul_without_nuw(i64 %x) {
+; CHECK-LABEL: define i64 @test_mul_without_nuw(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[X1:%.*]] = add i64 [[X]], 1
+; CHECK-NEXT: [[MUL:%.*]] = shl i64 [[X]], 1
+; CHECK-NEXT: [[MAX:%.*]] = call i64 @llvm.umax.i64(i64 [[MUL]], i64 [[X1]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %mul = mul i64 %x, 2
+ %max = call i64 @llvm.umax.i64(i64 %mul, i64 %x1)
+ ret i64 %max
+}
+
+; Multi-use Test Cases for `mul`
+
+define i64 @test_mul_multi_use_add(i64 %x) {
+; CHECK-LABEL: define i64 @test_mul_multi_use_add(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[X1:%.*]] = add i64 [[X]], 1
+; CHECK-NEXT: call void @use(i64 [[X1]])
+; CHECK-NEXT: [[TMP2:%.*]] = shl nuw i64 [[X]], 1
+; CHECK-NEXT: [[MAX:%.*]] = call i64 @llvm.umax.i64(i64 [[TMP2]], i64 [[X1]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ call void @use(i64 %x1)
+ %mul = mul nuw i64 %x, 2
+ %max = call i64 @llvm.umax.i64(i64 %mul, i64 %x1)
+ ret i64 %max
+}
+
+define i64 @test_mul_multi_use_mul(i64 %x) {
+; CHECK-LABEL: define i64 @test_mul_multi_use_mul(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[X1:%.*]] = add i64 [[X]], 1
+; CHECK-NEXT: [[MUL:%.*]] = shl nuw i64 [[X]], 1
+; CHECK-NEXT: call void @use(i64 [[MUL]])
+; CHECK-NEXT: [[MAX:%.*]] = call i64 @llvm.umax.i64(i64 [[MUL]], i64 [[X1]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %mul = mul nuw i64 %x, 2
+ call void @use(i64 %mul)
+ %max = call i64 @llvm.umax.i64(i64 %mul, i64 %x1)
+ ret i64 %max
+}
+
+define i64 @test_mul_multi_use_max(i64 %x) {
+; CHECK-LABEL: define i64 @test_mul_multi_use_max(
+; CHECK-SAME: i64 [[X:%.*]]) {
+; CHECK-NEXT: [[TMP2:%.*]] = shl nuw i64 [[X]], 1
+; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[X]], 0
+; CHECK-NEXT: [[MAX:%.*]] = select i1 [[TMP3]], i64 1, i64 [[TMP2]]
+; CHECK-NEXT: call void @use(i64 [[MAX]])
+; CHECK-NEXT: ret i64 [[MAX]]
+;
+ %x1 = add i64 %x, 1
+ %mul = mul nuw i64 %x, 2
+ %max = call i64 @llvm.umax.i64(i64 %mul, i64 %x1)
+ call void @use(i64 %max)
+ ret i64 %max
+}
|
nikic
reviewed
Jan 18, 2025
| match(I0, m_OneUse(m_Add(m_Specific(X), m_One()))))) && | ||
| C0 && !C0->isZero()) { | ||
| Op = isShl ? BinaryOperator::CreateNUWShl(X, C0) | ||
| : BinaryOperator::CreateNUWMul(X, C0); |
There was a problem hiding this comment.
Can't you reuse the original mul/shl? Then you also don't need the one-use restriction. This will also preserve the nsw flag, which is valid per https://alive2.llvm.org/ce/z/BfYZmT. (There should be a test for the nuw nsw combination.)
nikic
reviewed
Jan 18, 2025
goldsteinn
reviewed
Jan 18, 2025
| match(I1, m_OneUse(m_Add(m_Specific(X), m_One())))) || | ||
| (matchShiftOrMul(I1) && | ||
| match(I0, m_OneUse(m_Add(m_Specific(X), m_One()))))) && | ||
| C0 && !C0->isZero()) { |
There was a problem hiding this comment.
Checking C0 is isn't nullptr is redundant here and above.
goldsteinn
reviewed
Jan 18, 2025
| %mul = mul nuw i64 %x, 2 | ||
| %max = call i64 @llvm.umax.i64(i64 %mul, i64 %x1) | ||
| call void @use(i64 %max) | ||
| ret i64 %max |
There was a problem hiding this comment.
The mul tests need to have non-power of two multipliers, otherwise they are actually going through the shl pass.
nikic
reviewed
Jan 18, 2025
| const APInt *C0; | ||
| auto matchShiftOrMul = [&](Value *I) { | ||
| return ((match(I, m_OneUse(m_NUWShl(m_Value(X), m_APInt(C0))))) || | ||
| (match(I, m_OneUse(m_NUWMul(m_Value(X), m_APInt(C0)))) && |
There was a problem hiding this comment.
The one-use restriction on mul/shl are unnecessary.
| (((matchI0 = matchShiftOrMul(I0)) && | ||
| match(I1, m_OneUse(m_Add(m_Specific(X), m_One())))) || | ||
| (matchShiftOrMul(I1) && | ||
| match(I0, m_OneUse(m_Add(m_Specific(X), m_One())))))) { |
There was a problem hiding this comment.
I think it would be cleaner if you structured the fold like moveNotAfterMinMax below (i.e. implement it in a lambda, call with commuted args).
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR introduces the following transformations:
umax(nuw_shl(x, C0), x + 1) -> x == 0 ? 1 : nuw_shl(x, C0)
umax(nuw_mul(x, C0), x + 1) -> x == 0 ? 1 : nuw_mul(x, C0)
Fixes : #122388
Alive2 proof : https://alive2.llvm.org/ce/z/rkp_8U