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

[Jack Styles via llvm-commits]

================
@@ -6092,6 +6152,33 @@ LoopVectorizationPlanner::computeBestVF() {
     assert((FirstPlan.hasScalarVFOnly() || hasPlanWithVF(UserVF) ||
             FirstPlan.isOuterLoop()) &&
            "must have a single scalar VF, UserVF or an outer loop");
+    bool ForceVectorization =
+        Hints.getForce() == LoopVectorizeHints::FK_Enabled;
+    if (!FirstPlan.hasScalarVFOnly() && !FirstPlan.isOuterLoop() &&
+        hasPlanWithVF(UserVF) && UserVF.isVector() && !ForceVectorization) {
+      ElementCount ExactTC = getSmallConstantTripCount(PSE.getSE(), OrigLoop);
+      if (FirstPlan.hasScalarTail() && ExactTC.isFixed() && UserVF.isFixed()) {
+        unsigned TC = ExactTC.getFixedValue();
+        unsigned EstimatedWidth =
+            estimateElementCount(UserVF, Config.getVScaleForTuning());
+        if (TC != 0 && TC <= TTI.getMinTripCountTailFoldingThreshold() &&
+            TC == EstimatedWidth + 1) {
+          ElementCount ScalarVF = ElementCount::getFixed(1);
+          InstructionCost ScalarCost = CM.expectedCost(ScalarVF);
+          LLVM_DEBUG(dbgs()
+                     << "LV: Scalar loop costs: " << ScalarCost << ".\n");
+
+          InstructionCost Cost = cost(FirstPlan, UserVF, /*RU=*/nullptr);
+          VectorizationFactor UserFactor(UserVF, Cost, ScalarCost);
+          VectorizationFactor ScalarFactor(ScalarVF, ScalarCost, ScalarCost);
+          if (IsUnprofitableOneScalarTail(UserFactor, FirstPlan.hasScalarTail(),
----------------
Stylie777 wrote:

Agreed, Removed

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