[llvm] [CorrelatedValuePropagation] Fold calls to UCMP/SCMP when we know that ranges of operands do not overlap (PR #97235)

[Nikita Popov via llvm-commits]

================
@@ -548,6 +548,27 @@ static bool processAbsIntrinsic(IntrinsicInst *II, LazyValueInfo *LVI) {
   return false;
 }
 
+static bool processCmpIntrinsic(IntrinsicInst *II, LazyValueInfo *LVI) {
+  bool IsSigned = II->getIntrinsicID() == Intrinsic::scmp;
+  ConstantRange LHS_CR = LVI->getConstantRangeAtUse(II->getOperandUse(0),
+                                                    /*UndefAllowed*/ false);
+  ConstantRange RHS_CR = LVI->getConstantRangeAtUse(II->getOperandUse(1),
+                                                    /*UndefAllowed*/ false);
+
+  if (LHS_CR.icmp(IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT, RHS_CR)) {
+    II->replaceAllUsesWith(ConstantInt::get(II->getType(), 1));
+    II->eraseFromParent();
+    return true;
+  }
+  if (LHS_CR.icmp(IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, RHS_CR)) {
+    II->replaceAllUsesWith(ConstantInt::getSigned(II->getType(), -1));
+    II->eraseFromParent();
+    return true;
+  }
----------------
nikic wrote:

Should probably also handle the EQ case for completeness? This would mostly get handled by other case already, but it might come up if the range folds to a constant through some complex pathway, and doesn't really cost us anything to handle here.

https://github.com/llvm/llvm-project/pull/97235