[llvm] [LV] Remove selectUserVF, remove use of legacy expectedCost (NFCI) (PR #191216)

[David Sherwood via llvm-commits]

================
@@ -6653,20 +6617,26 @@ void LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) {
       // Collect the instructions (and their associated costs) that will be more
       // profitable to scalarize.
       CM.collectInLoopReductions();
-      if (CM.selectUserVectorizationFactor(UserVF)) {
-        LLVM_DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
-        ElementCount EpilogueUserVF =
-            ElementCount::getFixed(EpilogueVectorizationForceVF);
-        if (EpilogueUserVF.isVector() &&
-            ElementCount::isKnownLT(EpilogueUserVF, UserVF) &&
-            CM.selectUserVectorizationFactor(EpilogueUserVF)) {
-          // Build a separate plan for the forced epilogue VF.
-          buildVPlansWithVPRecipes(EpilogueUserVF, EpilogueUserVF);
+      CM.collectNonVectorizedAndSetWideningDecisions(UserVF);
+      ElementCount EpilogueUserVF =
+          ElementCount::getFixed(EpilogueVectorizationForceVF);
+      if (EpilogueUserVF.isVector() &&
+          ElementCount::isKnownLT(EpilogueUserVF, UserVF)) {
+        CM.collectNonVectorizedAndSetWideningDecisions(EpilogueUserVF);
+        buildVPlansWithVPRecipes(EpilogueUserVF, EpilogueUserVF);
+      }
+      buildVPlansWithVPRecipes(UserVF, UserVF);
+      if (!VPlans.empty()) {
+        // For scalar VF, skip VPlan cost check as VPlan cost is designed for
+        // vector VFs only.
+        if (UserVF.isScalar() ||
+            cost(*VPlans.back(), UserVF, /*RU=*/nullptr).isValid()) {
----------------
david-arm wrote:

I still think there is a potential problem here. We're assuming that `!VPlans.empty()` means the back plan is for the `UserVF`, which may not be true if the VPlan wasn't added. There is no guarantee that `buildVPlansWithVPRecipes` will actually add a plan.

https://github.com/llvm/llvm-project/pull/191216