[llvm] [LoopVectorize] Improve Vectorization of Low Trip Count Loops (PR #195823)

[Sander de Smalen via llvm-commits]

================
@@ -5895,6 +5961,19 @@ LoopVectorizationPlanner::computeBestVF() {
           cost(*P, VF, ConsiderRegPressure ? &RUs[I] : nullptr);
       VectorizationFactor CurrentFactor(VF, Cost, ScalarCost);
 
+      unsigned int UserIC =
+          Hints.getInterleave() != 0 ? Hints.getInterleave() : 1;
+      if (!ForceVectorization && P->hasScalarTail() &&
+          shouldConsiderOneScalarTail(CurrentFactor, ExactTC, UserIC) &&
+          isProfitableOneScalarTail(CurrentFactor, ExactTC, UserIC)) {
+        // If we have identified a case where a small trip count loop can
+        // be Vectorized as one Vector Iteration and, if needed, one scalar
+        // iteration, use this VF and break.
+        BestFactor = CurrentFactor;
+        PlanForBestVF = P.get();
+        break;
+      }
----------------
sdesmalen-arm wrote:

Given that the only supported VF's for this case is MaxVF (see my comment above on `isProfitableOneScalarTail`), you could also prune `VFs` to contain only the maximum VF (i.e. pick the last one), when this is a very-low-tripcount loop and the plan has a tail.

https://github.com/llvm/llvm-project/pull/195823