[llvm] [VPlan] Use ResumePhi to create reduction resume phis. (PR #110004)

[via llvm-commits]

================
@@ -7494,67 +7494,49 @@ static void addRuntimeUnrollDisableMetaData(Loop *L) {
   }
 }
 
-// Check if \p RedResult is a ComputeReductionResult instruction, and if it is
-// create a merge phi node for it.
-static void createAndCollectMergePhiForReduction(
-    VPInstruction *RedResult,
-    VPTransformState &State, Loop *OrigLoop, BasicBlock *LoopMiddleBlock,
-    bool VectorizingEpilogue) {
+// If \p RedResult is a ComputeReductionResult when vectorizing the epilog loop,
+// update the reduction's scalar PHI node by adding the incoming value from the
+// main vector loop.
+static void updateMergePhiForReductionForEpilogueVectorization(
+    VPInstruction *RedResult, VPTransformState &State, Loop *OrigLoop,
+    BasicBlock *LoopMiddleBlock) {
   if (!RedResult ||
       RedResult->getOpcode() != VPInstruction::ComputeReductionResult)
     return;
 
+  using namespace VPlanPatternMatch;
+  VPValue *ResumePhiVPV =
+      cast<VPInstruction>(*find_if(RedResult->users(), [](VPUser *U) {
+        return match(U, m_VPInstruction<VPInstruction::ResumePhi>(m_VPValue(),
+                                                                  m_VPValue()));
+      }));
+  assert(ResumePhiVPV->getNumUsers() == 1 &&
+         "ResumePhi must have a single user");
+  auto *BCBlockPhi = cast<PHINode>(State.get(ResumePhiVPV, true));
   auto *PhiR = cast<VPReductionPHIRecipe>(RedResult->getOperand(0));
   const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor();
-
-  Value *FinalValue = State.get(RedResult, VPLane(VPLane::getFirstLane()));
   auto *ResumePhi =
----------------
ayalz wrote:

It's somewhat confusing to first locate a ResumePhi recipe in scalar preheader, and now to define a ResumePhi PHINode as the starting value of the reduction header phi (merging bypasses with main loop result), possibly going through a compare for AnyOf reductions (boolean reduction's start value?)... some explanation and/or more distinct names may be helpful.

https://github.com/llvm/llvm-project/pull/110004