[InstCombine] Fold negation of calls to ucmp/scmp by swapping its operands

[Poseydon42]

Proofs: https://alive2.llvm.org/ce/z/cp_a36

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: None (Poseydon42)

Changes

Proofs: https://alive2.llvm.org/ce/z/cp_a36

---

Full diff: https://github.com/llvm/llvm-project/pull/98360.diff

3 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineNegator.cpp (+5)
• (modified) llvm/test/Transforms/InstCombine/scmp.ll (+12)
• (modified) llvm/test/Transforms/InstCombine/ucmp.ll (+12)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineNegator.cpp b/llvm/lib/Transforms/InstCombine/InstCombineNegator.cpp
index b3426562a4d87..9ad594a6d7d61 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineNegator.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineNegator.cpp
@@ -222,6 +222,11 @@ std::array<Value *, 2> Negator::getSortedOperandsOfBinOp(Instruction *I) {
     }
     break;
   }
+  case Instruction::Call:
+    if (auto *CI = dyn_cast<CmpIntrinsic>(I))
+      return Builder.CreateIntrinsic(CI->getType(), CI->getIntrinsicID(),
+                                     {CI->getRHS(), CI->getLHS()});
+    break;
   default:
     break; // Other instructions require recursive reasoning.
   }
diff --git a/llvm/test/Transforms/InstCombine/scmp.ll b/llvm/test/Transforms/InstCombine/scmp.ll
index 4f903a79afd5d..80b32e73a757d 100644
--- a/llvm/test/Transforms/InstCombine/scmp.ll
+++ b/llvm/test/Transforms/InstCombine/scmp.ll
@@ -154,3 +154,15 @@ define i1 @scmp_sle_neg_1(i32 %x, i32 %y) {
   %2 = icmp sle i8 %1, -1
   ret i1 %2
 }
+
+; ========== Fold -scmp(x, y) => scmp(y, x) ==========
+define i8 @scmp_negated(i32 %x, i32 %y) {
+; CHECK-LABEL: define i8 @scmp_negated(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[TMP2:%.*]] = call i8 @llvm.scmp.i8.i32(i32 [[Y]], i32 [[X]])
+; CHECK-NEXT:    ret i8 [[TMP2]]
+;
+  %1 = call i8 @llvm.scmp(i32 %x, i32 %y)
+  %2 = sub i8 0, %1
+  ret i8 %2
+}
diff --git a/llvm/test/Transforms/InstCombine/ucmp.ll b/llvm/test/Transforms/InstCombine/ucmp.ll
index 9ab67560c9117..74313403d7c0c 100644
--- a/llvm/test/Transforms/InstCombine/ucmp.ll
+++ b/llvm/test/Transforms/InstCombine/ucmp.ll
@@ -154,3 +154,15 @@ define i1 @ucmp_sle_neg_1(i32 %x, i32 %y) {
   %2 = icmp sle i8 %1, -1
   ret i1 %2
 }
+
+; ========== Fold -ucmp(x, y) => ucmp(y, x) ==========
+define i8 @ucmp_negated(i32 %x, i32 %y) {
+; CHECK-LABEL: define i8 @ucmp_negated(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[TMP2:%.*]] = call i8 @llvm.ucmp.i8.i32(i32 [[Y]], i32 [[X]])
+; CHECK-NEXT:    ret i8 [[TMP2]]
+;
+  %1 = call i8 @llvm.ucmp(i32 %x, i32 %y)
+  %2 = sub i8 0, %1
+  ret i8 %2
+}

[nikic]

Could you please add multi-use tests? I don't think we'd want to fold in that case, as replacing a negation with another ucmp/scmp doesn't seem profitable.

[Poseydon42]

Fixed.

[nikic]

LGTM