[llvm] [CVP] #114820 add 'samesign' for 'icmp' (PR #115642)

[Nikita Popov via llvm-commits]

================
@@ -288,24 +288,36 @@ static bool processICmp(ICmpInst *Cmp, LazyValueInfo *LVI) {
   if (!Cmp->getOperand(0)->getType()->isIntOrIntVectorTy())
     return false;
 
-  if (!Cmp->isSigned())
+  if (!(Cmp->isSigned() || (Cmp->isUnsigned() && !Cmp->hasSameSign())))
     return false;
 
-  ICmpInst::Predicate UnsignedPred =
-      ConstantRange::getEquivalentPredWithFlippedSignedness(
-          Cmp->getPredicate(),
-          LVI->getConstantRangeAtUse(Cmp->getOperandUse(0),
-                                     /*UndefAllowed*/ true),
-          LVI->getConstantRangeAtUse(Cmp->getOperandUse(1),
-                                     /*UndefAllowed*/ true));
+  bool Changed = false;
 
-  if (UnsignedPred == ICmpInst::Predicate::BAD_ICMP_PREDICATE)
-    return false;
+  ConstantRange CR1 = LVI->getConstantRangeAtUse(Cmp->getOperandUse(0),
+                                                 /*UndefAllowed=*/ true),
+                CR2 = LVI->getConstantRangeAtUse(Cmp->getOperandUse(1),
+                                                 /*UndefAllowed=*/ true);
----------------
nikic wrote:

I don't think that UndefAllowed=true is correct for this optimization. We should set UndefAllowed=false.

https://github.com/llvm/llvm-project/pull/115642