[LV][VPlan] Implement VPlan-based cost for exit condition.

This patch tried to model the cost of exit conditions through
vplan-based cost model.

* `BranchOnCount` will generate icmp + br.

The branch instruction is already implemented by the VPRegionBlock so
we only need to calculate the cost of icmp.

If the VF is same as the trip count of the loop, the cost of the
BranchOnCount is free.

This patch is not quite NFC for following reasons.

* Some of the BranchOnCount could be optimized to BranchOnCond, which is
free.
* Some of the instructions calculated in the exit condition in legacy
cost model will be optimized out.

Author: ElvisWang123

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -7447,8 +7447,10 @@ static bool planContainsAdditionalSimplifications(VPlan &Plan,
       if (isa<VPPartialReductionRecipe>(&R))
         return true;
 
-      // The legacy cost model will under estimate the cost of BranchOnCount if exit condition were explicit contructed in the vplan.
-      if (VPInstruction *VPI = dyn_cast<VPInstruction>(&R); VPI && VPI->getOpcode() == VPInstruction::BranchOnCount)
+      // The legacy cost model will under estimate the cost of BranchOnCount if
+      // exit condition were explicit contructed in the vplan.
+      if (VPInstruction *VPI = dyn_cast<VPInstruction>(&R);
+          VPI && VPI->getOpcode() == VPInstruction::BranchOnCount)
         return true;
 
       if (Instruction *UI = GetInstructionForCost(&R))

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -744,7 +744,6 @@ InstructionCost VPInstruction::computeCost(ElementCount VF,
         Instruction::Or, cast<VectorType>(VecTy), std::nullopt, Ctx.CostKind);
   }
   case VPInstruction::BranchOnCount: {
-    if (getUnderlyingValue())
     // BranchOnCount will genearte icmp_eq + br instructions and the
     // cost of branch will be calculated in VPRegionBlock.
     // If the vector loop only executed once, ignore the cost of the cmp.