[llvm] [LoopVectorizer][AArch64] Move getMinTripCountTailFoldingThreshold later. (PR #132170)

[David Green via llvm-commits]

================
@@ -4105,13 +4102,40 @@ LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) {
     const SCEV *Rem = SE->getURemExpr(
         SE->applyLoopGuards(ExitCount, TheLoop),
         SE->getConstant(BackedgeTakenCount->getType(), MaxVFtimesIC));
-    if (Rem->isZero()) {
+    return Rem->isZero();
+  };
+
+  if (MaxPowerOf2RuntimeVF && *MaxPowerOf2RuntimeVF > 0) {
+    assert((UserVF.isNonZero() || isPowerOf2_32(*MaxPowerOf2RuntimeVF)) &&
+           "MaxFixedVF must be a power of 2");
+    if (IsKnownModTripCountZero(*MaxPowerOf2RuntimeVF)) {
       // Accept MaxFixedVF if we do not have a tail.
       LLVM_DEBUG(dbgs() << "LV: No tail will remain for any chosen VF.\n");
       return MaxFactors;
     }
   }
 
+  if (MaxTC && MaxTC <= TTI.getMinTripCountTailFoldingThreshold()) {
----------------
davemgreen wrote:

This getMinTripCountTailFoldingThreshold appears to have been added to fix a particular difficult benchmark. It would be better if it was able to cost model correctly, with known overheads. I've added a test case.

https://github.com/llvm/llvm-project/pull/132170