Perform optimizeMasksToEVL as late as possible

So that #149706 doesn't need to worry about EVL recipes

Author: lukel97

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -8237,6 +8237,13 @@ void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF,
  }
  VF = SubRange.End;
  }
+
+ if (CM.foldTailWithEVL()) {
+ for (auto &Plan : VPlans) {
+ VPlanTransforms::runPass(VPlanTransforms::optimizeMasksToEVL, *Plan);
+ assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid");
+ }
+ }
 }
 
 VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -2334,7 +2334,6 @@ void VPlanTransforms::optimize(VPlan &Plan) {
  runPass(removeRedundantExpandSCEVRecipes, Plan);
  runPass(simplifyRecipes, Plan);
  runPass(removeBranchOnConst, Plan);
- runPass(optimizeMasksToEVL, Plan);
  runPass(removeDeadRecipes, Plan);
 
  runPass(createAndOptimizeReplicateRegions, Plan);
@@ -2654,6 +2653,8 @@ void VPlanTransforms::optimizeMasksToEVL(VPlan &Plan) {
  R->eraseFromParent();
  }
  }
+
+ removeDeadRecipes(Plan);
 }
 
 /// After replacing the IV with a EVL-based IV, fixup recipes that use VF to use