[llvm] [LV] Prefer fixed VFs for epilogue vectorization (PR #211814)

[Benjamin Maxwell via llvm-commits]

================
@@ -706,6 +710,23 @@ bool LoopVectorizationPlanner::isMoreProfitable(const VectorizationFactor &A,
         !B.Width.isScalar())
       return true;
 
+  // Favor fixed VFs for epilogue loops by scaling the costs of scalable VFs
+  // 'ScalableEpilogueVFCostScaleFactor' (default 2.0). This is intended to
+  // model that fixed VFs are more likely to be fully unrolled (or optimized
+  // out) post vectorization. TODO: Reconsider this restriction for predicated
+  // epilogues (once supported).
+  if (IsEpilogue && A.Width.isScalable() != B.Width.isScalable() &&
+      A.Cost.isValid() && B.Cost.isValid()) {
+    auto [FixedCost, ScalableCost] = std::make_pair(CostA, CostB);
+    if (B.Width.isFixed())
+      std::swap(FixedCost, ScalableCost);
+
+    ScalableCost *= ScalableEpilogueVFCostScaleFactor.getValue();
----------------
MacDue wrote:

Removed :+1: 

https://github.com/llvm/llvm-project/pull/211814