[llvm] [LV] Ignore some costs when loop gets fully unrolled (PR #106699)

[David Sherwood via llvm-commits]

================
@@ -5557,14 +5576,23 @@ InstructionCost LoopVectorizationCostModel::computePredInstDiscount(
 InstructionCost LoopVectorizationCostModel::expectedCost(ElementCount VF) {
   InstructionCost Cost;
 
+  // If the vector loop gets executed exactly once with the given VF, ignore the
+  // costs of comparison and induction instructions, as they'll get simplified
+  // away.
+  SmallPtrSet<Instruction *, 2> ValuesToIgnoreForVF;
+  auto TC = PSE.getSE()->getSmallConstantTripCount(TheLoop);
+  if (VF.isFixed() && TC == VF.getFixedValue() && !foldTailByMasking())
----------------
david-arm wrote:

This is because in `computeMaxVF` the variable `MaxPowerOf2RuntimeVF` is bigger than the trip count and so the remainder of `MaxPowerOf2RuntimeVF/TC` is non-zero. Basically we only disable tail-folding if we can prove for every single possible VF there will not be a tail. So I believe @igogo-x86 is right and we need an explicit check.

https://github.com/llvm/llvm-project/pull/106699