[VPlan] Flatten the CFG separately after creating wide recipes (NFC).

Move flattening of the CFG out of the loop that creates the wide
recipes. This simplifies the already large loop and prepares for moving
flattening to a separate transform.

Author: fhahn

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -9464,7 +9464,6 @@ LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(VFRange &Range,
       HeaderVPBB);
 
   VPBasicBlock::iterator MBIP = MiddleVPBB->getFirstNonPhi();
-  VPBlockBase *PrevVPBB = nullptr;
   for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) {
     // Create mask based on the IR BB corresponding to VPBB.
     // TODO: Predicate directly based on VPlan.
@@ -9548,7 +9547,10 @@ LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(VFRange &Range,
                "Unexpected multidef recipe");
       R.eraseFromParent();
     }
+  }
 
+  VPBlockBase *PrevVPBB = nullptr;
+  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) {
     // Flatten the CFG in the loop. Masks for blocks have already been generated
     // and added to recipes as needed. To do so, first disconnect VPBB from its
     // successors. Then connect VPBB to the previously visited VPBB.