[InstCombine] Fold (sub nsw X, umin(X, C)) s< C2 to (X s< C + C2) #170422
+144
−3
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
This patch implements a fold for the pattern icmp slt (sub nsw %x, umin(%x, %c)), %c2. Typically, the sub nsw ... umin sequence is canonicalized to @llvm.usub.sat. However, that transformation discards the nsw constraint, which provides valuable range information (guaranteeing no signed underflow). By detecting this pattern before canonicalization, we can merge the range check into a single comparison This optimization is valid under the following constraints: - %c is non-negative (C >= 0). - %c2 is strictly positive (C2 > 0). - The sum %c + %c2 does not cause a signed overflow. Alive Proof : https://alive2.llvm.org/ce/z/JG3SnP Fixed: llvm#167055
llvmbot
added
llvm:instcombine
Member
|
@llvm/pr-subscribers-llvm-transforms Author: hanbeom (ParkHanbum) ChangesThis patch implements a fold for the pattern Typically, the sub nsw ... umin sequence is canonicalized to This optimization is valid under the following constraints:
Alive Proof : https://alive2.llvm.org/ce/z/JG3SnP Full diff: https://github.com/llvm/llvm-project/pull/170422.diff 2 Files Affected:
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index 9bee523c7b7e5..df096e7b806ad 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -2857,9 +2857,33 @@ Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {X, Op1}));
// Op0 - umin(X, Op0) --> usub.sat(Op0, X)
- if (match(Op1, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op0)))))
- return replaceInstUsesWith(
- I, Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op0, X}));
+ if (match(Op1, m_OneUse(m_c_UMin(m_Value(X), m_Specific(Op0))))) {
+ // Op0 - umin(Op0, C) s< C2 --> Op0 s< C2 + C
+ // Constraints:
+ // 1. C >= 0: Range [0, C) must be valid.
+ // 2. C2 > 0: Range [C, C+C2) must imply extension.
+ // 3. No Signed Overflow: Merged upper bound (C + C2) must be safe.
+ const APInt *C;
+ if (I.hasOneUse() && I.hasNoSignedWrap() &&
+ match(X, m_APInt(C)) && C->isNonNegative()) {
+ const APInt *C2;
+ CmpPredicate Pred;
+ Instruction *MustICmp = cast<Instruction>(I.user_back());
+ if (match(MustICmp, m_ICmp(Pred, m_Specific(&I), m_APInt(C2))) &&
+ Pred == CmpInst::ICMP_SLT) {
+ bool Overflow;
+ APInt Sum = C->sadd_ov(*C2, Overflow);
+ if (C2->isStrictlyPositive() && !Overflow) {
+ Value *NewCmp = Builder.CreateICmpSLT(Op0, ConstantInt::get(Ty, Sum));
+ eraseInstFromFunction(*replaceInstUsesWith(*MustICmp, NewCmp));
+ // return eraseInstFromFunction(I);
+ return nullptr;
+ }
+ }
+ }
+ Value *USub = Builder.CreateIntrinsic(Intrinsic::usub_sat, {Ty}, {Op0, X});
+ return replaceInstUsesWith(I, USub);
+ }
// Op0 - umax(X, Op0) --> 0 - usub.sat(X, Op0)
if (match(Op1, m_OneUse(m_c_UMax(m_Value(X), m_Specific(Op0))))) {
diff --git a/llvm/test/Transforms/InstCombine/icmp-sub.ll b/llvm/test/Transforms/InstCombine/icmp-sub.ll
index 13ed7ba0c1703..e261e38697c69 100644
--- a/llvm/test/Transforms/InstCombine/icmp-sub.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-sub.ll
@@ -3,6 +3,125 @@
declare void @use(i32)
declare void @use_vec(<2 x i8>)
+declare i8 @llvm.umin.i8(i8, i8)
+
+; 1. Positive Tests
+; Basic valid case
+define i1 @test_basic_opt(i8 %x) {
+; CHECK-LABEL: @test_basic_opt(
+; CHECK-NEXT: [[TMP1:%.*]] = add i8 [[X:%.*]], 118
+; CHECK-NEXT: [[CMP:%.*]] = icmp ult i8 [[TMP1]], -108
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+ %min = call i8 @llvm.umin.i8(i8 %x, i8 10)
+ %sub = sub nsw i8 %x, %min
+ %cmp = icmp slt i8 %sub, 20
+ ret i1 %cmp
+}
+
+; Boundary case - C is 0
+define i1 @test_c_is_zero(i8 %x) {
+; CHECK-LABEL: @test_c_is_zero(
+; CHECK-NEXT: [[CMP:%.*]] = icmp slt i8 [[X:%.*]], 10
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+ %min = call i8 @llvm.umin.i8(i8 %x, i8 0)
+ %sub = sub nsw i8 %x, %min
+ %cmp = icmp slt i8 %sub, 10
+ ret i1 %cmp
+}
+
+; Boundary case - Sum is exactly SMAX (127)
+define i1 @test_sum_is_smax(i8 %x) {
+; CHECK-LABEL: @test_sum_is_smax(
+; CHECK-NEXT: [[TMP1:%.*]] = add i8 [[X:%.*]], 28
+; CHECK-NEXT: [[CMP:%.*]] = icmp ult i8 [[TMP1]], -101
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+ %min = call i8 @llvm.umin.i8(i8 %x, i8 100)
+ %sub = sub nsw i8 %x, %min
+ %cmp = icmp slt i8 %sub, 27
+ ret i1 %cmp
+}
+
+; 2. Negative Tests
+; Missing 'nsw' flag on sub
+define i1 @fail_no_nsw(i8 %x) {
+; CHECK-LABEL: @fail_no_nsw(
+; CHECK-NEXT: [[TMP1:%.*]] = add i8 [[X:%.*]], 118
+; CHECK-NEXT: [[CMP:%.*]] = icmp ult i8 [[TMP1]], -108
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+ %min = call i8 @llvm.umin.i8(i8 %x, i8 10)
+ %sub = sub i8 %x, %min
+ %cmp = icmp slt i8 %sub, 20
+ ret i1 %cmp
+}
+
+; C is Negative (Constraint 1 Violation)
+define i1 @fail_c_negative(i8 %x) {
+; CHECK-LABEL: @fail_c_negative(
+; CHECK-NEXT: ret i1 true
+;
+ %min = call i8 @llvm.umin.i8(i8 %x, i8 -10)
+ %sub = sub nsw i8 %x, %min
+ %cmp = icmp slt i8 %sub, 20
+ ret i1 %cmp
+}
+
+; C2 is Zero (Constraint 2 Violation)
+define i1 @fail_c2_zero(i8 %x) {
+; CHECK-LABEL: @fail_c2_zero(
+; CHECK-NEXT: [[CMP:%.*]] = icmp ugt i8 [[X:%.*]], -119
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+ %min = call i8 @llvm.umin.i8(i8 %x, i8 10)
+ %sub = sub nsw i8 %x, %min
+ %cmp = icmp slt i8 %sub, 0
+ ret i1 %cmp
+}
+
+; C2 is Negative (Constraint 2 Violation)
+define i1 @fail_c2_negative(i8 %x) {
+; CHECK-LABEL: @fail_c2_negative(
+; CHECK-NEXT: [[CMP:%.*]] = icmp ugt i8 [[X:%.*]], -119
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+ %min = call i8 @llvm.umin.i8(i8 %x, i8 10)
+ %sub = sub nsw i8 %x, %min
+ %cmp = icmp slt i8 %sub, -5
+ ret i1 %cmp
+}
+
+; Signed Overflow in C + C2 (Constraint 3 Violation)
+define i1 @fail_sum_overflow(i8 %x) {
+; CHECK-LABEL: @fail_sum_overflow(
+; CHECK-NEXT: [[TMP1:%.*]] = add i8 [[X:%.*]], 28
+; CHECK-NEXT: [[CMP:%.*]] = icmp ult i8 [[TMP1]], -98
+; CHECK-NEXT: ret i1 [[CMP]]
+;
+ %min = call i8 @llvm.umin.i8(i8 %x, i8 100)
+ %sub = sub nsw i8 %x, %min
+ %cmp = icmp slt i8 %sub, 30
+ ret i1 %cmp
+}
+
+; Multi-use of sub instruction
+define i8 @fail_multi_use(i8 %x) {
+; CHECK-LABEL: @fail_multi_use(
+; CHECK-NEXT: [[SUB:%.*]] = call i8 @llvm.usub.sat.i8(i8 [[X:%.*]], i8 10)
+; CHECK-NEXT: [[CMP:%.*]] = icmp slt i8 [[SUB]], 20
+; CHECK-NEXT: [[RES:%.*]] = zext i1 [[CMP]] to i8
+; CHECK-NEXT: [[RET:%.*]] = add nuw i8 [[SUB]], [[RES]]
+; CHECK-NEXT: ret i8 [[RET]]
+;
+ %min = call i8 @llvm.umin.i8(i8 %x, i8 10)
+ %sub = sub nsw i8 %x, %min
+ %cmp = icmp slt i8 %sub, 20
+ %res = zext i1 %cmp to i8
+ %ret = add i8 %sub, %res
+ ret i8 %ret
+}
define i1 @test_nuw_and_unsigned_pred(i64 %x) {
; CHECK-LABEL: @test_nuw_and_unsigned_pred(
@@ -290,7 +409,8 @@ define i1 @subC_nsw_ne(i32 %x) {
; CHECK-LABEL: @subC_nsw_ne(
; CHECK-NEXT: [[SUBX:%.*]] = sub nsw i32 -2147483647, [[X:%.*]]
; CHECK-NEXT: call void @use(i32 [[SUBX]])
-; CHECK-NEXT: ret i1 true
+; CHECK-NEXT: [[R:%.*]] = icmp ne i32 [[X]], 2147483603
+; CHECK-NEXT: ret i1 [[R]]
;
%subx = sub nsw i32 -2147483647, %x
call void @use(i32 %subx)
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
Contributor
|
Doesn't this go against the optimization approach suggested in #168880? |
Contributor
Author
|
I attempted to implement it using the approach discussed in that issue. it could find no alternative method besides adding multiple instructions and removing them from the worklist. Since it hadn't seen this approach implemented elsewhere, I changed the implementation method. |
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.
This patch implements a fold for the pattern
icmp slt (sub nsw %x, umin(%x, %c)), %c2.
Typically, the sub nsw ... umin sequence is canonicalized to
@llvm.usub.sat. However, that transformation discards the nsw
constraint, which provides valuable range information
(guaranteeing no signed underflow). By detecting this pattern before
canonicalization, we can merge the range check into a single
comparison
This optimization is valid under the following constraints:
Alive Proof : https://alive2.llvm.org/ce/z/JG3SnP
Fixed: #167055