[llvm] [LV] Refine tripcount estimate using minimum iteration count rt check. (PR #188135)

[Ramkumar Ramachandra via llvm-commits]

================
@@ -460,6 +460,26 @@ static ElementCount getSmallConstantTripCount(ScalarEvolution *SE,
   return ElementCount::getFixed(0);
 }
 
+/// Try to refine the maximum trip count for \p L using the unsigned range of
+/// the trip count, discarding the case we warp around to zero. The minimum
+/// iteration check ensures the vector loop won't get entered if computing the
+/// trip count wraps around to 0. Note that this only applies when not folding
+/// the tail. Returns 0 if the range cannot be determined.
+static unsigned tryToRefineConstantMaxTripCount(PredicatedScalarEvolution &PSE,
+                                                Loop *L) {
+  const SCEV *BTC = PSE.getBackedgeTakenCount();
+  if (isa<SCEVCouldNotCompute>(BTC))
+    return 0;
+  ScalarEvolution *SE = PSE.getSE();
+  const SCEV *TripCount = SE->getTripCountFromExitCount(BTC, BTC->getType(), L);
+  ConstantRange TCRange = SE->getUnsignedRange(TripCount);
+  APInt MaxTCFromRange = TCRange.getUnsignedMax();
+  if (MaxTCFromRange.ugt(0) &&
+      MaxTCFromRange.ule(std::numeric_limits<unsigned>::max()))
+    return MaxTCFromRange.getZExtValue();
+  return 0;
----------------
artagnon wrote:

Hm, if I understand correctly, this is a fallback when SCEV's getSmallConstantMaxTripCount fails? Wouldn't this be best absorbed into the CanUseConstantMax case of getSmallBestKnownTC? Also, "refineConstantMaxTripCount" would imply that we operate on the constant-max trip count, and perform computations to "refine" it, which is not the case here?

https://github.com/llvm/llvm-project/pull/188135