[InstCombine] Preserve NSW/NUW flags when folding const BOp with min/max #143471

AlexMaclean · 2025-06-10T02:38:58Z

When folding X Pred C2 ? X BOp C1 : C2 BOp C1 to min/max(X, C2) BOp C1, if NUW/NSW flags are present on X BOp C1 and could be safely applied to C2 BOp C1, then they may be added on the BOp after the fold is complete. https://alive2.llvm.org/ce/z/n_3aNJ

Preserving these flags can allow subsequent transforms to re-order the min/max and BOp, which in the case of NVPTX would allow for some potential future transformations which would improve instruction-selection.

llvmbot · 2025-06-10T02:39:38Z

@llvm/pr-subscribers-llvm-transforms

Author: Alex MacLean (AlexMaclean)

Changes

When folding X Pred C2 ? X BOp C1 : C2 BOp C1 to min/max(X, C2) BOp C1, if NUW/NSW flags are present on X BOp C1 and could be safely applied to C2 BOp C1, then they may be added on the BOp after the fold is complete. https://alive2.llvm.org/ce/z/n_3aNJ

Preserving these flags can allow subsequent transforms to re-order the min/max and BOp, which in the case of NVPTX would allow for some potential future transformations which would improve instruction-selection.

Full diff: https://github.com/llvm/llvm-project/pull/143471.diff

3 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineInternal.h (+2)
(modified) llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp (+27-9)
(modified) llvm/test/Transforms/InstCombine/canonicalize-const-to-bop.ll (+81-2)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
index 334462d715f95..48f718bae29af 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
+++ b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
@@ -767,6 +767,8 @@ class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final
                             Value *A, Value *B, Instruction &Outer,
                             SelectPatternFlavor SPF2, Value *C);
   Instruction *foldSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
+  Value *foldSelectWithConstOpToBinOp(ICmpInst *Cmp, Value *TrueVal,
+                                      Value *FalseVal);
   Instruction *foldSelectValueEquivalence(SelectInst &SI, CmpInst &CI);
   bool replaceInInstruction(Value *V, Value *Old, Value *New,
                             unsigned Depth = 0);
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
index 8f46ae304353d..062b576369ce2 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp
@@ -1823,9 +1823,9 @@ static Instruction *foldSelectICmpEq(SelectInst &SI, ICmpInst *ICI,
 
 /// Fold `X Pred C1 ? X BOp C2 : C1 BOp C2` to `min/max(X, C1) BOp C2`.
 /// This allows for better canonicalization.
-static Value *foldSelectWithConstOpToBinOp(ICmpInst *Cmp, Value *TrueVal,
-                                           Value *FalseVal,
-                                           IRBuilderBase &Builder) {
+Value *InstCombinerImpl::foldSelectWithConstOpToBinOp(ICmpInst *Cmp,
+                                                      Value *TrueVal,
+                                                      Value *FalseVal) {
   Constant *C1, *C2, *C3;
   Value *X;
   CmpPredicate Predicate;
@@ -1889,11 +1889,29 @@ static Value *foldSelectWithConstOpToBinOp(ICmpInst *Cmp, Value *TrueVal,
     return nullptr;
   }
 
-  Intrinsic::ID IntrinsicID = getMinMaxIntrinsic(SPF);
-  Value *Intrinsic = Builder.CreateBinaryIntrinsic(IntrinsicID, X, RHS);
-  return IsIntrinsic ? Builder.CreateBinaryIntrinsic(Opcode, Intrinsic, C2)
-                     : Builder.CreateBinOp(Instruction::BinaryOps(Opcode),
-                                           Intrinsic, C2);
+  Intrinsic::ID MinMaxID = getMinMaxIntrinsic(SPF);
+  Value *MinMax = Builder.CreateBinaryIntrinsic(MinMaxID, X, RHS);
+  if (IsIntrinsic)
+    return Builder.CreateBinaryIntrinsic(Opcode, MinMax, C2);
+
+  const auto BinOpc = Instruction::BinaryOps(Opcode);
+  Value *BinOp = Builder.CreateBinOp(BinOpc, MinMax, C2);
+
+  // If we can attach no-wrap flags to the new instruction, do so if the
+  // old instruction had them and C1 BinOp C2 does not overflow.
+  if (Instruction *BinOpInst = dyn_cast<Instruction>(BinOp)) {
+    if (BinOpc == Instruction::Add || BinOpc == Instruction::Sub ||
+        BinOpc == Instruction::Mul) {
+      Instruction *OldBinOp = cast<BinaryOperator>(TrueVal);
+      if (OldBinOp->hasNoSignedWrap() &&
+          willNotOverflow(BinOpc, RHS, C2, *BinOpInst, /*IsSigned*/ true))
+        BinOpInst->setHasNoSignedWrap();
+      if (OldBinOp->hasNoUnsignedWrap() &&
+          willNotOverflow(BinOpc, RHS, C2, *BinOpInst, /*IsSigned*/ false))
+        BinOpInst->setHasNoUnsignedWrap();
+    }
+  }
+  return BinOp;
 }
 
 /// Visit a SelectInst that has an ICmpInst as its first operand.
@@ -1968,7 +1986,7 @@ Instruction *InstCombinerImpl::foldSelectInstWithICmp(SelectInst &SI,
   if (Value *V = foldAbsDiff(ICI, TrueVal, FalseVal, Builder))
     return replaceInstUsesWith(SI, V);
 
-  if (Value *V = foldSelectWithConstOpToBinOp(ICI, TrueVal, FalseVal, Builder))
+  if (Value *V = foldSelectWithConstOpToBinOp(ICI, TrueVal, FalseVal))
     return replaceInstUsesWith(SI, V);
 
   return Changed ? &SI : nullptr;
diff --git a/llvm/test/Transforms/InstCombine/canonicalize-const-to-bop.ll b/llvm/test/Transforms/InstCombine/canonicalize-const-to-bop.ll
index c08ec1bb7de0d..b3093a92624ae 100644
--- a/llvm/test/Transforms/InstCombine/canonicalize-const-to-bop.ll
+++ b/llvm/test/Transforms/InstCombine/canonicalize-const-to-bop.ll
@@ -5,7 +5,7 @@ define i8 @add_and_sgt(i8 %x) {
 ; CHECK-LABEL: define i8 @add_and_sgt(
 ; CHECK-SAME: i8 [[X:%.*]]) {
 ; CHECK-NEXT:    [[TMP1:%.*]] = call i8 @llvm.smax.i8(i8 [[X]], i8 8)
-; CHECK-NEXT:    [[S:%.*]] = add nuw i8 [[TMP1]], 16
+; CHECK-NEXT:    [[S:%.*]] = add nuw nsw i8 [[TMP1]], 16
 ; CHECK-NEXT:    ret i8 [[S]]
 ;
   %add = add nsw i8 %x, 16
@@ -155,7 +155,7 @@ define i8 @multi_use_cond_and_sel(i8 %x) {
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[X]], 8
 ; CHECK-NEXT:    call void @use(i1 [[CMP]])
 ; CHECK-NEXT:    [[TMP1:%.*]] = call i8 @llvm.smax.i8(i8 [[X]], i8 8)
-; CHECK-NEXT:    [[S:%.*]] = add nuw i8 [[TMP1]], 16
+; CHECK-NEXT:    [[S:%.*]] = add nuw nsw i8 [[TMP1]], 16
 ; CHECK-NEXT:    call void @use_byte(i8 [[S]])
 ; CHECK-NEXT:    ret i8 [[S]]
 ;
@@ -450,3 +450,82 @@ define i8 @umax_sgt(i8 %x) {
   %s = select i1 %cmp, i8 100, i8 %umax
   ret i8 %s
 }
+
+define i8 @add_sgt_nuw_nsw_safe(i8 %x) {
+; CHECK-LABEL: define i8 @add_sgt_nuw_nsw_safe(
+; CHECK-SAME: i8 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i8 @llvm.smin.i8(i8 [[X]], i8 100)
+; CHECK-NEXT:    [[S:%.*]] = add nuw nsw i8 [[TMP1]], 1
+; CHECK-NEXT:    ret i8 [[S]]
+;
+  %add = add nuw nsw i8 %x, 1
+  %cmp = icmp sgt i8 %x, 100
+  %s = select i1 %cmp, i8 101, i8 %add
+  ret i8 %s
+}
+
+define i8 @add_sgt_nuw_only(i8 %x) {
+; CHECK-LABEL: define i8 @add_sgt_nuw_only(
+; CHECK-SAME: i8 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i8 @llvm.smin.i8(i8 [[X]], i8 100)
+; CHECK-NEXT:    [[S:%.*]] = add nuw i8 [[TMP1]], 50
+; CHECK-NEXT:    ret i8 [[S]]
+;
+  %add = add nuw nsw i8 %x, 50
+  %cmp = icmp sgt i8 %x, 100
+  %s = select i1 %cmp, i8 150, i8 %add
+  ret i8 %s
+}
+
+define i8 @add_sgt_nsw_only(i8 %x) {
+; CHECK-LABEL: define i8 @add_sgt_nsw_only(
+; CHECK-SAME: i8 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i8 @llvm.smin.i8(i8 [[X]], i8 100)
+; CHECK-NEXT:    [[S:%.*]] = add nsw i8 [[TMP1]], -99
+; CHECK-NEXT:    ret i8 [[S]]
+;
+  %add = add nuw nsw i8 %x, -99
+  %cmp = icmp sgt i8 %x, 100
+  %s = select i1 %cmp, i8 1, i8 %add
+  ret i8 %s
+}
+
+
+define i8 @mul_ult_nuw_nsw_safe(i8 %x) {
+; CHECK-LABEL: define i8 @mul_ult_nuw_nsw_safe(
+; CHECK-SAME: i8 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i8 @llvm.umax.i8(i8 [[X]], i8 10)
+; CHECK-NEXT:    [[S:%.*]] = mul nuw nsw i8 [[TMP1]], 3
+; CHECK-NEXT:    ret i8 [[S]]
+;
+  %mul = mul nuw nsw i8 %x, 3
+  %cmp = icmp ult i8 %x, 10
+  %s = select i1 %cmp, i8 30, i8 %mul
+  ret i8 %s
+}
+
+define i8 @mul_ult_nuw_only(i8 %x) {
+; CHECK-LABEL: define i8 @mul_ult_nuw_only(
+; CHECK-SAME: i8 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i8 @llvm.umax.i8(i8 [[X]], i8 10)
+; CHECK-NEXT:    [[S:%.*]] = mul nuw i8 [[TMP1]], 25
+; CHECK-NEXT:    ret i8 [[S]]
+;
+  %mul = mul nuw nsw i8 %x, 25
+  %cmp = icmp ult i8 %x, 10
+  %s = select i1 %cmp, i8 250, i8 %mul
+  ret i8 %s
+}
+
+define i8 @mul_ult_nsw_only(i8 %x) {
+; CHECK-LABEL: define i8 @mul_ult_nsw_only(
+; CHECK-SAME: i8 [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i8 @llvm.umax.i8(i8 [[X]], i8 40)
+; CHECK-NEXT:    [[S:%.*]] = mul nsw i8 [[TMP1]], -2
+; CHECK-NEXT:    ret i8 [[S]]
+;
+  %mul = mul nuw nsw i8 %x, -2
+  %cmp = icmp ult i8 %x, 40
+  %s = select i1 %cmp, i8 -80, i8 %mul
+  ret i8 %s
+}

nikic

LGTM

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

Co-authored-by: Nikita Popov <[email protected]>

dtcxzyw

LG

…max (llvm#143471) When folding `X Pred C2 ? X BOp C1 : C2 BOp C1` to `min/max(X, C2) BOp C1`, if NUW/NSW flags are present on `X BOp C1` and could be safely applied to `C2 BOp C1`, then they may be added on the BOp after the fold is complete. https://alive2.llvm.org/ce/z/n_3aNJ Preserving these flags can allow subsequent transforms to re-order the min/max and BOp, which in the case of NVPTX would allow for some potential future transformations which would improve instruction-selection.

AlexMaclean added 2 commits June 9, 2025 15:23

pre-commit tests

6bee710

[InstCombine] Preserve NSW/NUW flags when folding const BOp with min/max

16bc374

AlexMaclean requested a review from dtcxzyw June 10, 2025 02:38

AlexMaclean self-assigned this Jun 10, 2025

AlexMaclean requested a review from nikic as a code owner June 10, 2025 02:38

llvmbot added llvm:instcombine Covers the InstCombine, InstSimplify and AggressiveInstCombine passes llvm:transforms labels Jun 10, 2025

AlexMaclean mentioned this pull request Jun 10, 2025

[InstCombine] Allow min/max in constant BOp min/max folding #142878

Merged

This was referenced Jun 10, 2025

Task submission dtcxzyw/llvm-opt-benchmark#1312

Open

pre-commit: PR143471 dtcxzyw/llvm-opt-benchmark#2414

Closed

nikic approved these changes Jun 10, 2025

View reviewed changes

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp Outdated Show resolved Hide resolved

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp Outdated Show resolved Hide resolved

AlexMaclean and others added 2 commits June 10, 2025 11:09

Update llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

8d6be44

Co-authored-by: Nikita Popov <[email protected]>

Update llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

7940b46

Co-authored-by: Nikita Popov <[email protected]>

dtcxzyw mentioned this pull request Jun 13, 2025

Fuzz PR143471 dtcxzyw/llvm-mutation-based-fuzz-service#60

Closed

dtcxzyw approved these changes Jun 13, 2025

View reviewed changes

AlexMaclean merged commit 59388fb into llvm:main Jun 13, 2025
7 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[InstCombine] Preserve NSW/NUW flags when folding const BOp with min/max #143471

[InstCombine] Preserve NSW/NUW flags when folding const BOp with min/max #143471

Uh oh!

AlexMaclean commented Jun 10, 2025

Uh oh!

llvmbot commented Jun 10, 2025

Uh oh!

nikic left a comment

Uh oh!

Uh oh!

Uh oh!

dtcxzyw left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

[InstCombine] Preserve NSW/NUW flags when folding const BOp with min/max #143471

[InstCombine] Preserve NSW/NUW flags when folding const BOp with min/max #143471

Uh oh!

Conversation

AlexMaclean commented Jun 10, 2025

Uh oh!

llvmbot commented Jun 10, 2025

Uh oh!

nikic left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

dtcxzyw left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants