[llvm] [LV] Always try to use fixed VF for low TC loops if no epilogue allowed. (PR #226318)

[Florian Hahn via llvm-commits]

================
@@ -7813,8 +7812,12 @@ bool LoopVectorizePass::processLoop(Loop *L) {
       // `CM_EpilogueNotAllowedLowTripLoop` prevents vectorizing loops
       // with runtime checks. It's more effective to let
       // `isOutsideLoopWorkProfitable` determine if vectorization is
-      // beneficial for the loop.
-      if (SEL != CM_EpilogueNotNeededFoldTail)
+      // beneficial for the loop. If the trip count is below the target's
+      // minimum for tail-folding, the tail cannot be folded, so treat it like
+      // any other low trip count loop.
+      if (SEL != CM_EpilogueNotNeededFoldTail ||
----------------
fhahn wrote:

Both changes are needed for the test changes. Without that,`SEL` is incorrectly set to `CM_EpilogueNotNeededFoldTail` and the logic around line 3062 does not trigger. Alternatively we could check both, but it seems cleaner to not set to `CM_EpilogueNotNeededFoldTail` when tail folding is not allowed for the trip count

https://github.com/llvm/llvm-project/pull/226318