[llvm] [LoopPeel] Avoid unprofitable last iteration peeling with equalitis (PR #143588)

[Florian Hahn via llvm-commits]

================
@@ -381,9 +381,23 @@ static bool shouldPeelLastIteration(Loop &L, CmpPredicate Pred,
   const SCEV *ValAtSecondToLastIter = LeftAR->evaluateAtIteration(
       SE.getMinusSCEV(BTC, SE.getOne(BTC->getType())), SE);
 
-  return SE.isKnownPredicate(ICmpInst::getInversePredicate(Pred), ValAtLastIter,
-                             RightSCEV) &&
-         SE.isKnownPredicate(Pred, ValAtSecondToLastIter, RightSCEV);
+  CmpPredicate InvPred = ICmpInst::getInversePredicate(Pred);
+  if (!SE.isKnownPredicate(InvPred, ValAtLastIter, RightSCEV) ||
+      !SE.isKnownPredicate(Pred, ValAtSecondToLastIter, RightSCEV))
+    return false;
+
+  // For a monotonic predicate, we've found the single transition point and
+  // don't need to check further.
+  if (!ICmpInst::isEquality(Pred))
+    return true;
----------------
fhahn wrote:

I *think* we may be missing test coverage for this early exit.

https://github.com/llvm/llvm-project/pull/143588