[InstCombine] Add (A & B) - (A & ~B) --> B - (A ^ B)

[AtariDreams]

(A & B) - (A & ~B) --> B - (A ^ B), and (A & ~B) - (A & B) --> (A ^ B) - B

Alive2 Proof:
https://alive2.llvm.org/ce/z/DWVWYy

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: AtariDreams (AtariDreams)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/79717.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp (+18)
• (modified) llvm/test/Transforms/InstCombine/sub.ll (+26)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index 8a00b75a1f7404..aceb178263811c 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -2371,6 +2371,24 @@ Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
       return BinaryOperator::CreateNeg(Y);
   }
 
+  {
+    Value *A, *B;
+    // (A & ~B) - (A & B) => (A ^ B) - B
+    if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
+        match(Op1, m_c_And(m_Specific(A), m_Specific(B)))) {
+      return BinaryOperator::CreateSub(Builder.CreateXor(A, B), B);
+    }
+  }
+
+  {
+    Value *A, *B;
+    // (A & B) - (A & ~B) => B - (A ^ B)
+    if (match(Op1, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
+        match(Op0, m_c_And(m_Specific(A), m_Specific(B)))) {
+      return BinaryOperator::CreateSub(Builder.CreateXor(A, B), B);
+    }
+  }
+
   // (sub (or A, B) (and A, B)) --> (xor A, B)
   {
     Value *A, *B;
diff --git a/llvm/test/Transforms/InstCombine/sub.ll b/llvm/test/Transforms/InstCombine/sub.ll
index 494cdf62c7975e..65142d4e200633 100644
--- a/llvm/test/Transforms/InstCombine/sub.ll
+++ b/llvm/test/Transforms/InstCombine/sub.ll
@@ -2626,3 +2626,29 @@ define i8 @sub_of_adds_2xc(i8 %x, i8 %y) {
   %r = sub i8 %xc, %yc
   ret i8 %r
 }
+
+define i32 @sub_and_xor(i32 %A, i32 %B) {
+; CHECK-LABEL: @sub_and_xor(
+; CHECK-NEXT:    [[TMP1:%.*]] = xor i32 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[RES:%.*]] = sub i32 [[TMP1]], [[B]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %xor = xor i32 %B, -1
+  %and1 = and i32 %xor, %A
+  %and2 = and i32 %B, %A
+  %res = sub nsw i32 %and1, %and2
+  ret i32 %res
+}
+
+define i32 @sub_xor_and(i32 %A, i32 %B) {
+; CHECK-LABEL: @sub_xor_and(
+; CHECK-NEXT:    [[TMP1:%.*]] = xor i32 [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[RES:%.*]] = sub i32 [[TMP1]], [[B]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %and1 = and i32 %B, %A
+  %xor = xor i32 %B, -1
+  %and2 = and i32 %xor, %A
+  %res = sub nsw i32 %and1, %and2
+  ret i32 %res
+}

[nikic]

Did you encounter this pattern in real-world code?

[AtariDreams]

Not this in particular, but I've seen code that folds to (A+~B)-(A+B) for example.

[AtariDreams]

_Bytes = _Bytes & ~(_Align - 1); Is a common idiom for rounding up to the nearest alignment.
Difference = Bytes & (_Align - 1) is used to get the modulo.

subtracting these values is useful if you want to know the difference if say, _Bytes were to be the size of a buffer for example.

[topperc]

Please don't force push to PRs https://llvm.org/docs/GitHub.html#rebasing-pull-requests-and-force-pushes

[dtcxzyw]

Please don't force push to PRs https://llvm.org/docs/GitHub.html#rebasing-pull-requests-and-force-pushes

TBH I always force-push my PRs since I have a bunch of concurrent PRs with different base commits. Without rebasing and force-pushing, ccache works poorly, and I have to waste about an hour every time I switch to another branch and rebuild LLVM :(

[dtcxzyw]

_Bytes = _Bytes & ~(_Align - 1); Is a common idiom for rounding up to the nearest alignment.

Should be (_Bytes + _Align - 1) & ~(_Align - 1).

[nikic]

_Bytes = _Bytes & ~(_Align - 1); Is a common idiom for rounding up to the nearest alignment. Difference = Bytes & (_Align - 1) is used to get the modulo.

subtracting these values is useful if you want to know the difference if say, _Bytes were to be the size of a buffer for example.

~(_Align - 1) will become -_Align and your pattern will not match. If _Align is constant it will also not match.

Please don't invent hypothetical scenarios where it could apply -- provide positive proof where it does apply.

Besides, this pattern is not undef-safe. Your proof specified noundef, and if we drop it we get this: https://alive2.llvm.org/ce/z/-ajcoF It would be necessary to insert freeze instructions to make it correct.

Given that no real-world motivation was provided, I'm closing this.