[InstCombine] Fold trunc nuw/nsw (x xor y) to i1 to x != y

[YanWQ-monad]

Fold:

define i1 @src(i8 %x, i8 %y) {
  %xor = xor i8 %x, %y
  %r = trunc nuw i8 %xor to i1
  ret i1 %r
}

define i1 @tgt(i8 %x, i8 %y) {
  %r = icmp ne i8 %x, %y
  ret i1 %r
}

Proof: https://alive2.llvm.org/ce/z/dcuHmn

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Monad (YanWQ-monad)

Changes

Fold:

define i1 @<!-- -->src(i64 %x, i64 %y) {
entry:
  %xor = xor i64 %x, %y
  %r = trunc nuw i64 %xor to i1
  ret i1 %r
}

define i1 @<!-- -->tgt(i64 %x, i64 %y) {
  %r = icmp ne i64 %x, %y
  ret i1 %r
}

Proof: https://alive2.llvm.org/ce/z/RSWpbS

---

Full diff: https://github.com/llvm/llvm-project/pull/90408.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp (+7)
• (modified) llvm/test/Transforms/InstCombine/trunc.ll (+31)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
index 567b27b4630439..c769c28c90c1a9 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -770,6 +770,13 @@ Instruction *InstCombinerImpl::visitTrunc(TruncInst &Trunc) {
         return new ICmpInst(ICmpInst::Predicate::ICMP_EQ, X, Zero);
       }
     }
+
+    if (Trunc.hasNoUnsignedWrap()) {
+      Value *X, *Y;
+      if (match(Src, m_Xor(m_Value(X), m_Value(Y)))) {
+        return new ICmpInst(ICmpInst::ICMP_NE, X, Y);
+      }
+    }
   }
 
   Value *A, *B;
diff --git a/llvm/test/Transforms/InstCombine/trunc.ll b/llvm/test/Transforms/InstCombine/trunc.ll
index e59b2bea6684c0..d4e4b1866349c0 100644
--- a/llvm/test/Transforms/InstCombine/trunc.ll
+++ b/llvm/test/Transforms/InstCombine/trunc.ll
@@ -1054,3 +1054,34 @@ define i8 @drop_both_trunc(i16 %x, i16 %y) {
   %res = trunc nuw nsw i16 %and2 to i8
   ret i8 %res
 }
+
+define i1 @trunc_xor(i8 %x, i8 %y) {
+; CHECK-LABEL: @trunc_xor(
+; CHECK-NEXT:    [[XOR:%.*]] = xor i8 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = trunc i8 [[XOR]] to i1
+; CHECK-NEXT:    ret i1 [[R]]
+;
+  %xor = xor i8 %x, %y
+  %r = trunc i8 %xor to i1
+  ret i1 %r
+}
+
+define i1 @trunc_nuw_xor(i8 %x, i8 %y) {
+; CHECK-LABEL: @trunc_nuw_xor(
+; CHECK-NEXT:    [[R:%.*]] = icmp ne i8 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret i1 [[R]]
+;
+  %xor = xor i8 %x, %y
+  %r = trunc nuw i8 %xor to i1
+  ret i1 %r
+}
+
+define <2 x i1> @trunc_nuw_xor_vector(<2 x i8> %x, <2 x i8> %y) {
+; CHECK-LABEL: @trunc_nuw_xor_vector(
+; CHECK-NEXT:    [[R:%.*]] = icmp ne <2 x i8> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    ret <2 x i1> [[R]]
+;
+  %xor = xor <2 x i8> %x, %y
+  %r = trunc nuw <2 x i8> %xor to <2 x i1>
+  ret <2 x i1> %r
+}

[dtcxzyw]

LGTM.

[vfdff]

LGTM

[asmok-g]

Heads-up: we're seeing a probable mis-compile after this commit in a google internal test. The function that changes behavior with and without optimization is: https://github.com/google/arolla/blob/65c8389f271b9dce973801074217afb0e35f850c/arolla/memory/simple_buffer.h#L336.

Still checking if this is some UB in the code or a real mis-compile.

[asmok-g]

The "probable" mis-compile is only happening on arm.

[alexfh]

I suppose, this particular function is only involved in what might be a miscompile, when it's inlined into another function, but disabling the optimization just on it using #pragma clang optimize off/on is enough to fix the test we see failing. The problem is that the whole setup leading to the problem is quite convoluted, and there's no isolated reproducer so far.

We're also actively investigating other possibilities like a preexisting UB in the code.

[alexfh]

False alarm, it seems. At least, we found UB in the code, fixing which resolves the problem. Apologies for the noise.