[llvm] [VPlan] Remove loop region in optimizeForVFAndUF. (PR #108378)

[via llvm-commits]

================
@@ -1069,19 +1069,28 @@ InstructionCost VPlan::cost(ElementCount VF, VPCostContext &Ctx) {
   return getVectorLoopRegion()->cost(VF, Ctx);
 }
 
+VPBasicBlock *VPlan::getVectorPreheader() {
+  VPBlockBase *Current = getEntry()->getSuccessors().back();
+  while (Current->getNumSuccessors() == 2)
+    Current = Current->getSuccessors().back();
+  return cast<VPBasicBlock>(Current);
+}
+
+VPBasicBlock *VPlan::getVectorPreheader() const {
+  VPBlockBase *Current = getEntry()->getSuccessors().back();
+  while (Current->getNumSuccessors() == 2)
+    Current = Current->getSuccessors().back();
+  return cast<VPBasicBlock>(Current);
+}
+
 VPRegionBlock *VPlan::getVectorLoopRegion() {
   // TODO: Cache if possible.
-  for (VPBlockBase *B : vp_depth_first_shallow(getEntry()))
-    if (auto *R = dyn_cast<VPRegionBlock>(B))
-      return R;
-  return nullptr;
+  return dyn_cast<VPRegionBlock>(getVectorPreheader()->getSingleSuccessor());
----------------
ayalz wrote:

Return null if this successor is a replicating region?

Can retain existing implementation which uses a shallow scan from entry, provided it also returns null if the first region found is replicating.

https://github.com/llvm/llvm-project/pull/108378