[InstCombine] When -A + B both have nsw flag, set nsw flag.

[Z572]

Fixes #72119

https://alive2.llvm.org/ce/z/5f_QuC

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: None (Z572)

Changes

Fixes #72119

https://alive2.llvm.org/ce/z/5f_QuC

---

Full diff: https://github.com/llvm/llvm-project/pull/72127.diff

3 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp (+5-1)
• (modified) llvm/test/Transforms/InstCombine/add.ll (+30)
• (modified) llvm/test/Transforms/InstCombine/pr14365.ll (+2-2)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index 22fd3edc39acb03..9ff1a57b6be93ba 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -1487,7 +1487,11 @@ Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
       return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));
 
     // -A + B --> B - A
-    return BinaryOperator::CreateSub(RHS, A);
+    auto *Sub = BinaryOperator::CreateSub(RHS, A);
+    if (auto *Op1 = dyn_cast<BinaryOperator>(LHS)) {
+      Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && Op1->hasNoSignedWrap());
+    }
+    return Sub;
   }
 
   // A + -B  -->  A - B
diff --git a/llvm/test/Transforms/InstCombine/add.ll b/llvm/test/Transforms/InstCombine/add.ll
index 7234fb39e566ebf..fff839c2d89a138 100644
--- a/llvm/test/Transforms/InstCombine/add.ll
+++ b/llvm/test/Transforms/InstCombine/add.ll
@@ -120,6 +120,36 @@ define i32 @test5(i32 %A, i32 %B) {
   ret i32 %D
 }
 
+define i32 @test5_2(i32 %A, i32 %B) {
+; CHECK-LABEL: @test5_2(
+; CHECK-NEXT:    [[D:%.*]] = sub nsw i32 [[B:%.*]], [[A:%.*]]
+; CHECK-NEXT:    ret i32 [[D]]
+;
+  %C = sub nsw i32 0, %A
+  %D = add nsw i32 %C, %B
+  ret i32 %D
+}
+
+define i32 @test5_3(i32 %A, i32 %B) {
+; CHECK-LABEL: @test5_3(
+; CHECK-NEXT:    [[D:%.*]] = sub i32 [[B:%.*]], [[A:%.*]]
+; CHECK-NEXT:    ret i32 [[D]]
+;
+  %C = sub nsw i32 0, %A
+  %D = add i32 %C, %B
+  ret i32 %D
+}
+
+define i32 @test5_4(i32 %A, i32 %B) {
+; CHECK-LABEL: @test5_4(
+; CHECK-NEXT:    [[D:%.*]] = sub i32 [[B:%.*]], [[A:%.*]]
+; CHECK-NEXT:    ret i32 [[D]]
+;
+  %C = sub i32 0, %A
+  %D = add nsw i32 %C, %B
+  ret i32 %D
+}
+
 define <2 x i8> @neg_op0_vec_undef_elt(<2 x i8> %a, <2 x i8> %b) {
 ; CHECK-LABEL: @neg_op0_vec_undef_elt(
 ; CHECK-NEXT:    [[R:%.*]] = sub <2 x i8> [[B:%.*]], [[A:%.*]]
diff --git a/llvm/test/Transforms/InstCombine/pr14365.ll b/llvm/test/Transforms/InstCombine/pr14365.ll
index 5e8dca13fa1b4d8..3a09b55aba309d9 100644
--- a/llvm/test/Transforms/InstCombine/pr14365.ll
+++ b/llvm/test/Transforms/InstCombine/pr14365.ll
@@ -31,7 +31,7 @@ define i32 @test1(i32 %a0) {
 ; CHECK-LABEL: @test1(
 ; CHECK-NEXT:    [[TMP1:%.*]] = lshr i32 [[A0:%.*]], 1
 ; CHECK-NEXT:    [[TMP2:%.*]] = and i32 [[TMP1]], 1431655765
-; CHECK-NEXT:    [[TMP3:%.*]] = sub i32 [[A0]], [[TMP2]]
+; CHECK-NEXT:    [[TMP3:%.*]] = sub nsw i32 [[A0]], [[TMP2]]
 ; CHECK-NEXT:    ret i32 [[TMP3]]
 ;
   %1 = ashr i32 %a0, 1
@@ -46,7 +46,7 @@ define <4 x i32> @test1_vec(<4 x i32> %a0) {
 ; CHECK-LABEL: @test1_vec(
 ; CHECK-NEXT:    [[TMP1:%.*]] = lshr <4 x i32> [[A0:%.*]], <i32 1, i32 1, i32 1, i32 1>
 ; CHECK-NEXT:    [[TMP2:%.*]] = and <4 x i32> [[TMP1]], <i32 1431655765, i32 1431655765, i32 1431655765, i32 1431655765>
-; CHECK-NEXT:    [[TMP3:%.*]] = sub <4 x i32> [[A0]], [[TMP2]]
+; CHECK-NEXT:    [[TMP3:%.*]] = sub nsw <4 x i32> [[A0]], [[TMP2]]
 ; CHECK-NEXT:    ret <4 x i32> [[TMP3]]
 ;
   %1 = ashr <4 x i32> %a0, <i32 1, i32 1, i32 1, i32 1>

[nikic]

Suggested change

	if (auto *Op1 = dyn_cast<BinaryOperator>(LHS)) {
	Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && Op1->hasNoSignedWrap());
	}
	auto *OBO = cast<OverflowingBinaryOperator>(LHS));
	Sub->setHasNoSignedWrap(I.hasNoSignedWrap() && OBO->hasNoSignedWrap());

[arichardson]

This was not fixed in the committed version, it's still using dyn_cast.

[Z572]

#72282

[nikic]

LGTM