[Instcombine]: llvm.ucmp and llvm.scmp recognition
#168505
Open
+250
−0
Conversation
This file contains hidden or 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
llvmbot
added
llvm:instcombine
Member
|
@llvm/pr-subscribers-llvm-transforms Author: None (kper) ChangesCreated a pattern to recognize Closes #166579 Full diff: https://github.com/llvm/llvm-project/pull/168505.diff 2 Files Affected:
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
index 9572f9d702e1b..5c8008700e181 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
@@ -1955,6 +1955,48 @@ static Instruction *foldSelectICmpEq(SelectInst &SI, ICmpInst *ICI,
return nullptr;
}
+/// Transform
+///
+/// select(icmp(eq, X, Y), Z, select(icmp(ult, X, Y), -1, 1))
+/// into select(icmp(eq, X, Y), Z, llvm.ucmp(freeze(X), freeze(Y)))
+///
+/// or
+///
+/// select(icmp(eq, X, Y), Z, select(icmp(slt, X, Y), -1, 1))
+/// into select(icmp(eq, X, Y), Z, llvm.scmp(freeze(X), freeze(Y)))
+static Value *foldSelectToInstrincCmp(SelectInst &SI, const ICmpInst *ICI,
+ Value *TrueVal, Value *FalseVal,
+ InstCombiner::BuilderTy &Builder) {
+ ICmpInst::Predicate Pred = ICI->getPredicate();
+
+ if (Pred != ICmpInst::ICMP_EQ)
+ return nullptr;
+
+ CmpPredicate IPred;
+ if (match(FalseVal, m_Select(m_ICmp(IPred, m_Specific(ICI->getOperand(0)),
+ m_Specific(ICI->getOperand(1))),
+ m_AllOnes(), m_One())) &&
+ (IPred == ICmpInst::ICMP_ULT || IPred == ICmpInst::ICMP_SLT)) {
+ Value *X = ICI->getOperand(0);
+ Value *Y = ICI->getOperand(1);
+ Builder.SetInsertPoint(&SI);
+ auto IID = IPred == ICmpInst::ICMP_ULT ? Intrinsic::ucmp : Intrinsic::scmp;
+
+ // Edge Case: if Z is the constant 0 then the select can be folded
+ // to just the instrinsic comparison.
+ if (match(TrueVal, m_Zero()))
+ return Builder.CreateIntrinsic(SI.getType(), IID, {X, Y});
+
+ Value *FrozenX = Builder.CreateFreeze(X, X->getName() + ".frz");
+ Value *FrozenY = Builder.CreateFreeze(Y, Y->getName() + ".frz");
+ Value *Cmp =
+ Builder.CreateIntrinsic(FrozenX->getType(), IID, {FrozenX, FrozenY});
+ return Builder.CreateSelect(SI.getCondition(), TrueVal, Cmp, "select.ucmp");
+ }
+
+ return nullptr;
+}
+
/// Fold `X Pred C1 ? X BOp C2 : C1 BOp C2` to `min/max(X, C1) BOp C2`.
/// This allows for better canonicalization.
Value *InstCombinerImpl::foldSelectWithConstOpToBinOp(ICmpInst *Cmp,
@@ -2186,6 +2228,9 @@ Instruction *InstCombinerImpl::foldSelectInstWithICmp(SelectInst &SI,
if (Value *V = foldSelectWithConstOpToBinOp(ICI, TrueVal, FalseVal))
return replaceInstUsesWith(SI, V);
+ if (Value *V = foldSelectToInstrincCmp(SI, ICI, TrueVal, FalseVal, Builder))
+ return replaceInstUsesWith(SI, V);
+
return Changed ? &SI : nullptr;
}
diff --git a/llvm/test/Transforms/InstCombine/select-cmp.ll b/llvm/test/Transforms/InstCombine/select-cmp.ll
index b1bd7a0ecc8ac..bf1a6cb047c37 100644
--- a/llvm/test/Transforms/InstCombine/select-cmp.ll
+++ b/llvm/test/Transforms/InstCombine/select-cmp.ll
@@ -808,5 +808,119 @@ define i1 @icmp_lt_slt(i1 %c, i32 %arg) {
ret i1 %select
}
+define i16 @icmp_fold_to_llvm_ucmp_when_eq(i16 %x, i16 %y) {
+; CHECK-LABEL: @icmp_fold_to_llvm_ucmp_when_eq(
+; CHECK-NEXT: [[Y_FRZ:%.*]] = freeze i16 [[Y:%.*]]
+; CHECK-NEXT: [[X_FRZ:%.*]] = freeze i16 [[X:%.*]]
+; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i16 [[X_FRZ]], [[Y_FRZ]]
+; CHECK-NEXT: [[TMP2:%.*]] = call i16 @llvm.ucmp.i16.i16(i16 [[X_FRZ]], i16 [[Y_FRZ]])
+; CHECK-NEXT: [[SELECT_UCMP:%.*]] = select i1 [[TMP1]], i16 42, i16 [[TMP2]]
+; CHECK-NEXT: ret i16 [[SELECT_UCMP]]
+;
+ %3 = icmp eq i16 %x, %y
+ %4 = icmp ult i16 %x, %y
+ %5 = select i1 %4, i16 -1, i16 1
+ %6 = select i1 %3, i16 42, i16 %5
+ ret i16 %6
+}
+
+define i16 @icmp_fold_to_llvm_ucmp_when_ult_and_Z_zero(i16 %x, i16 %y) {
+; CHECK-LABEL: @icmp_fold_to_llvm_ucmp_when_ult_and_Z_zero(
+; CHECK-NEXT: [[TMP1:%.*]] = call i16 @llvm.ucmp.i16.i16(i16 [[X:%.*]], i16 [[Y:%.*]])
+; CHECK-NEXT: ret i16 [[TMP1]]
+;
+ %3 = icmp eq i16 %x, %y
+ %4 = icmp ult i16 %x, %y
+ %5 = select i1 %4, i16 -1, i16 1
+ %6 = select i1 %3, i16 0, i16 %5
+ ret i16 %6
+}
+
+define i16 @icmp_fold_to_llvm_ucmp_when_slt_and_Z_zero(i16 %x, i16 %y) {
+; CHECK-LABEL: @icmp_fold_to_llvm_ucmp_when_slt_and_Z_zero(
+; CHECK-NEXT: [[TMP1:%.*]] = call i16 @llvm.scmp.i16.i16(i16 [[X:%.*]], i16 [[Y:%.*]])
+; CHECK-NEXT: ret i16 [[TMP1]]
+;
+ %3 = icmp eq i16 %x, %y
+ %4 = icmp slt i16 %x, %y
+ %5 = select i1 %4, i16 -1, i16 1
+ %6 = select i1 %3, i16 0, i16 %5
+ ret i16 %6
+}
+
+define i16 @icmp_fold_to_llvm_ucmp_when_cmp_slt(i16 %x, i16 %y) {
+; CHECK-LABEL: @icmp_fold_to_llvm_ucmp_when_cmp_slt(
+; CHECK-NEXT: [[Y_FRZ:%.*]] = freeze i16 [[Y:%.*]]
+; CHECK-NEXT: [[X_FRZ:%.*]] = freeze i16 [[X:%.*]]
+; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i16 [[X_FRZ]], [[Y_FRZ]]
+; CHECK-NEXT: [[TMP2:%.*]] = call i16 @llvm.scmp.i16.i16(i16 [[X_FRZ]], i16 [[Y_FRZ]])
+; CHECK-NEXT: [[SELECT_UCMP:%.*]] = select i1 [[TMP1]], i16 42, i16 [[TMP2]]
+; CHECK-NEXT: ret i16 [[SELECT_UCMP]]
+;
+ %3 = icmp eq i16 %x, %y
+ %4 = icmp slt i16 %x, %y ; here "ult" changed to "slt"
+ %5 = select i1 %4, i16 -1, i16 1
+ %6 = select i1 %3, i16 42, i16 %5
+ ret i16 %6
+}
+
+define i16 @icmp_fold_to_llvm_ucmp_when_value(i16 %x, i16 %y, i16 %Z) {
+; CHECK-LABEL: @icmp_fold_to_llvm_ucmp_when_value(
+; CHECK-NEXT: [[Y_FRZ:%.*]] = freeze i16 [[Y:%.*]]
+; CHECK-NEXT: [[X_FRZ:%.*]] = freeze i16 [[X:%.*]]
+; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i16 [[X_FRZ]], [[Y_FRZ]]
+; CHECK-NEXT: [[TMP2:%.*]] = call i16 @llvm.ucmp.i16.i16(i16 [[X_FRZ]], i16 [[Y_FRZ]])
+; CHECK-NEXT: [[SELECT_UCMP:%.*]] = select i1 [[TMP1]], i16 [[Z:%.*]], i16 [[TMP2]]
+; CHECK-NEXT: ret i16 [[SELECT_UCMP]]
+;
+ %3 = icmp eq i16 %x, %y
+ %4 = icmp ult i16 %x, %y
+ %5 = select i1 %4, i16 -1, i16 1
+ %6 = select i1 %3, i16 %Z, i16 %5
+ ret i16 %6
+}
+
+define i16 @icmp_fold_to_llvm_ucmp_when_ne(i16 %x, i16 %y) {
+; CHECK-LABEL: @icmp_fold_to_llvm_ucmp_when_ne(
+; CHECK-NEXT: [[Y_FRZ:%.*]] = freeze i16 [[Y:%.*]]
+; CHECK-NEXT: [[X_FRZ:%.*]] = freeze i16 [[X:%.*]]
+; CHECK-NEXT: [[DOTNOT:%.*]] = icmp eq i16 [[X_FRZ]], [[Y_FRZ]]
+; CHECK-NEXT: [[TMP1:%.*]] = call i16 @llvm.ucmp.i16.i16(i16 [[X_FRZ]], i16 [[Y_FRZ]])
+; CHECK-NEXT: [[SELECT_UCMP:%.*]] = select i1 [[DOTNOT]], i16 42, i16 [[TMP1]]
+; CHECK-NEXT: ret i16 [[SELECT_UCMP]]
+;
+ %3 = icmp ne i16 %x, %y
+ %4 = icmp ult i16 %x, %y
+ %5 = select i1 %4, i16 -1, i16 1
+ %6 = select i1 %3, i16 %5, i16 42
+ ret i16 %6
+}
+
+define i16 @icmp_fold_to_llvm_ucmp_negative_test_invalid_constant_1(i16 %x, i16 %y, i16 %Z) {
+; CHECK-LABEL: @icmp_fold_to_llvm_ucmp_negative_test_invalid_constant_1(
+; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i16 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], i16 [[Z:%.*]], i16 1
+; CHECK-NEXT: ret i16 [[TMP2]]
+;
+ %3 = icmp eq i16 %x, %y
+ %4 = icmp ult i16 %x, %y
+ %5 = select i1 %4, i16 1, i16 1 ; invalid constant
+ %6 = select i1 %3, i16 %Z, i16 %5
+ ret i16 %6
+}
+
+define i16 @icmp_fold_to_llvm_ucmp_negative_test_invalid_constant_2(i16 %x, i16 %y, i16 %Z) {
+; CHECK-LABEL: @icmp_fold_to_llvm_ucmp_negative_test_invalid_constant_2(
+; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i16 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], i16 [[Z:%.*]], i16 -1
+; CHECK-NEXT: ret i16 [[TMP2]]
+;
+ %3 = icmp eq i16 %x, %y
+ %4 = icmp ult i16 %x, %y
+ %5 = select i1 %4, i16 -1, i16 -1 ; invalid constant
+ %6 = select i1 %3, i16 %Z, i16 %5
+ ret i16 %6
+}
+
declare void @use(i1)
declare void @use.i8(i8)
|
Contributor
Author
|
@dtcxzyw could you run the benchmarks to see whether this transformation is profitable? |
🐧 Linux x64 Test Results
|
Contributor
Author
|
@dtcxzyw thanks, I fixed two bugs. First, the return type of the intrinsic wasn't correct. Second, the fold must not be applied to pointer comparisons. |
Contributor
Author
dtcxzyw
reviewed
Nov 21, 2025
Member
dtcxzyw
left a comment
dtcxzyw
left a comment
There was a problem hiding this comment.
For the first pattern, s/ucmp is less profitable than select + icmp because the former one needs two icmp + two selects/one sub (See also TargetLowering::expandCMP). It doesn't simplify the IR since X and Y still have two uses after the transformation.
For the second pattern select(icmp(eq, X, Y), 0, llvm.cmp(X, Y)) -> llvm.cmp(X, Y), absorbing the equality test into s/ucmp is interesting. Unfortunately we haven't seen its existence in real-world programs.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Created a pattern to recognize
llvm.ucmpandllvm.scmp.Alive Proof: https://alive2.llvm.org/ce/z/BYRyu-
Closes #166579