[VPlan] Model branch cond to enter scalar epilogue in VPlan.

This patch moves branch condition creation to enter the scalar epilogue
loop to VPlan. Modeling the branch in the middle block also requires
modeling the successor blocks. To do so, this patch introduces a new
VPBlockBase sub-type that simply wraps an existing IR block:
VPIRWrapperBlock (name subject to better suggestions).

This allows allows connecting blocks naturally to leave nodes. It can
also be used to transition modeling of more bits of the skeleton to
VPlan gradually.

Note that the middle.block is still created as part of the skeleton and
then patched in during VPlan execution. Unfortunately the skeleton needs
to create the middle.block early on, as it is also used for induction
resume value creation and is also needed to properly update the
dominator tree during skeleton creation.

After this patch lands, I plan to move induction resume value and phi
node creation in the scalar preheader to VPlan. Once that is done, we
should be able to create the middle.block in VPlan directly.

At the moment, VPIRWrapperBlocks only wrap an original IR block and
don't allow any additions. We could allow IR wrapper blocks to also
contain recipes that get added at the beginning of the block. Then the
middle.block could also be wrapped. Something like that will also be
needed to place the induction/reduction resume phi nodes in the scalar
preheader.

This is a re-worked version based on the earlier
https://reviews.llvm.org/D150398 and the main change is the introduction
and use of VPIRWrapperBlock.

Note that this patch adds and uses a new helper reorderIncomingBlocks to
preserve the original order of incoming blocks in created phi nodes; as
this patch changes the order branches are created, the order of the
predecessors changes, which would cause different order of incoming
blocks in some phis we create. The helper ensures the created IR doesn't
change (modulo some minor difference in name numbering). After this
change landed, the tests should be updated and the helper removed
separately.

Depends on #92525

Author: fhahn

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -59,6 +59,7 @@
 #include "VPlan.h"
 #include "VPlanAnalysis.h"
 #include "VPlanHCFGBuilder.h"
+#include "VPlanPatternMatch.h"
 #include "VPlanTransforms.h"
 #include "VPlanVerifier.h"
 #include "llvm/ADT/APInt.h"
@@ -2972,22 +2973,7 @@ void InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) {
       SplitBlock(LoopMiddleBlock, LoopMiddleBlock->getTerminator(), DT, LI,
                  nullptr, Twine(Prefix) + "scalar.ph");
 
-  auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator();
-
-  // Set up the middle block terminator.  Two cases:
-  // 1) If we know that we must execute the scalar epilogue, emit an
-  //    unconditional branch.
-  // 2) Otherwise, we must have a single unique exit block (due to how we
-  //    implement the multiple exit case).  In this case, set up a conditional
-  //    branch from the middle block to the loop scalar preheader, and the
-  //    exit block.  completeLoopSkeleton will update the condition to use an
-  //    iteration check, if required to decide whether to execute the remainder.
-  BranchInst *BrInst =
-      Cost->requiresScalarEpilogue(VF.isVector())
-          ? BranchInst::Create(LoopScalarPreHeader)
-          : BranchInst::Create(LoopExitBlock, LoopScalarPreHeader,
-                               Builder.getTrue());
-  BrInst->setDebugLoc(ScalarLatchTerm->getDebugLoc());
+  auto *BrInst = new UnreachableInst(LoopMiddleBlock->getContext());
   ReplaceInstWithInst(LoopMiddleBlock->getTerminator(), BrInst);
 
   // Update dominator for loop exit. During skeleton creation, only the vector
@@ -3094,50 +3080,7 @@ void InnerLoopVectorizer::createInductionResumeValues(
   }
 }
 
-BasicBlock *InnerLoopVectorizer::completeLoopSkeleton() {
-  // The trip counts should be cached by now.
-  Value *Count = getTripCount();
-  Value *VectorTripCount = getOrCreateVectorTripCount(LoopVectorPreHeader);
-
-  auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator();
-
-  // Add a check in the middle block to see if we have completed
-  // all of the iterations in the first vector loop.  Three cases:
-  // 1) If we require a scalar epilogue, there is no conditional branch as
-  //    we unconditionally branch to the scalar preheader.  Do nothing.
-  // 2) If (N - N%VF) == N, then we *don't* need to run the remainder.
-  //    Thus if tail is to be folded, we know we don't need to run the
-  //    remainder and we can use the previous value for the condition (true).
-  // 3) Otherwise, construct a runtime check.
-  if (!Cost->requiresScalarEpilogue(VF.isVector()) &&
-      !Cost->foldTailByMasking()) {
-    // Here we use the same DebugLoc as the scalar loop latch terminator instead
-    // of the corresponding compare because they may have ended up with
-    // different line numbers and we want to avoid awkward line stepping while
-    // debugging. Eg. if the compare has got a line number inside the loop.
-    // TODO: At the moment, CreateICmpEQ will simplify conditions with constant
-    // operands. Perform simplification directly on VPlan once the branch is
-    // modeled there.
-    IRBuilder<> B(LoopMiddleBlock->getTerminator());
-    B.SetCurrentDebugLocation(ScalarLatchTerm->getDebugLoc());
-    Value *CmpN = B.CreateICmpEQ(Count, VectorTripCount, "cmp.n");
-    BranchInst &BI = *cast<BranchInst>(LoopMiddleBlock->getTerminator());
-    BI.setCondition(CmpN);
-    if (hasBranchWeightMD(*ScalarLatchTerm)) {
-      // Assume that `Count % VectorTripCount` is equally distributed.
-      unsigned TripCount = UF * VF.getKnownMinValue();
-      assert(TripCount > 0 && "trip count should not be zero");
-      const uint32_t Weights[] = {1, TripCount - 1};
-      setBranchWeights(BI, Weights);
-    }
-  }
-
-#ifdef EXPENSIVE_CHECKS
-  assert(DT->verify(DominatorTree::VerificationLevel::Fast));
-#endif
 
-  return LoopVectorPreHeader;
-}
 
 std::pair<BasicBlock *, Value *>
 InnerLoopVectorizer::createVectorizedLoopSkeleton(
@@ -3198,7 +3141,7 @@ InnerLoopVectorizer::createVectorizedLoopSkeleton(
   // Emit phis for the new starting index of the scalar loop.
   createInductionResumeValues(ExpandedSCEVs);
 
-  return {completeLoopSkeleton(), nullptr};
+  return {LoopVectorPreHeader, nullptr};
 }
 
 // Fix up external users of the induction variable. At this point, we are
@@ -3481,6 +3424,18 @@ void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State,
                                VF.getKnownMinValue() * UF);
 }
 
+// Helper to reorder blocks so they match the original order even after the
+// order of the predecessors changes. This is only used to avoid a number of
+// test changes due to reordering of incoming blocks in phi nodes and should be
+// removed soon, with the tests being updated.
+static void reorderIncomingBlocks(SmallVectorImpl<BasicBlock *> &Blocks,
+                                  BasicBlock *LoopMiddleBlock) {
+  if (Blocks.front() == LoopMiddleBlock)
+    std::swap(Blocks.front(), Blocks.back());
+  if (Blocks.size() == 3)
+    std::swap(Blocks[0], Blocks[1]);
+}
+
 void InnerLoopVectorizer::fixFixedOrderRecurrence(
     VPFirstOrderRecurrencePHIRecipe *PhiR, VPTransformState &State) {
   // This is the second phase of vectorizing first-order recurrences. An
@@ -3591,7 +3546,9 @@ void InnerLoopVectorizer::fixFixedOrderRecurrence(
   PHINode *Phi = cast<PHINode>(PhiR->getUnderlyingValue());
   auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init");
   auto *ScalarInit = PhiR->getStartValue()->getLiveInIRValue();
-  for (auto *BB : predecessors(LoopScalarPreHeader)) {
+  SmallVector<BasicBlock *> Blocks(predecessors(LoopScalarPreHeader));
+  reorderIncomingBlocks(Blocks, LoopMiddleBlock);
+  for (auto *BB : Blocks) {
     auto *Incoming = BB == LoopMiddleBlock ? ExtractForScalar : ScalarInit;
     Start->addIncoming(Incoming, BB);
   }
@@ -7480,7 +7437,9 @@ static void createAndCollectMergePhiForReduction(
   // If we are fixing reductions in the epilogue loop then we should already
   // have created a bc.merge.rdx Phi after the main vector body. Ensure that
   // we carry over the incoming values correctly.
-  for (auto *Incoming : predecessors(LoopScalarPreHeader)) {
+  SmallVector<BasicBlock *> Blocks(predecessors(LoopScalarPreHeader));
+  reorderIncomingBlocks(Blocks, LoopMiddleBlock);
+  for (auto *Incoming : Blocks) {
     if (Incoming == LoopMiddleBlock)
       BCBlockPhi->addIncoming(FinalValue, Incoming);
     else if (ResumePhi && is_contained(ResumePhi->blocks(), Incoming))
@@ -7551,6 +7510,21 @@ LoopVectorizationPlanner::executePlan(
   std::tie(State.CFG.PrevBB, CanonicalIVStartValue) =
       ILV.createVectorizedLoopSkeleton(ExpandedSCEVs ? *ExpandedSCEVs
                                                      : State.ExpandedSCEVs);
+#ifdef EXPENSIVE_CHECKS
+  assert(DT->verify(DominatorTree::VerificationLevel::Fast));
+#endif
+
+  VPBasicBlock *MiddleVPBB =
+      cast<VPBasicBlock>(BestVPlan.getVectorLoopRegion()->getSingleSuccessor());
+
+  using namespace llvm::VPlanPatternMatch;
+  if (MiddleVPBB->begin() != MiddleVPBB->end() &&
+      match(&MiddleVPBB->back(), m_BranchOnCond(m_VPValue()))) {
+    cast<VPIRWrapperBlock>(MiddleVPBB->getSuccessors()[1])
+        ->resetBlock(OrigLoop->getLoopPreheader());
+  } else
+    cast<VPIRWrapperBlock>(MiddleVPBB->getSuccessors()[0])
+        ->resetBlock(OrigLoop->getLoopPreheader());
 
   // Only use noalias metadata when using memory checks guaranteeing no overlap
   // across all iterations.
@@ -7687,7 +7661,7 @@ EpilogueVectorizerMainLoop::createEpilogueVectorizedLoopSkeleton(
   // inductions in the epilogue loop are created before executing the plan for
   // the epilogue loop.
 
-  return {completeLoopSkeleton(), nullptr};
+  return {LoopVectorPreHeader, nullptr};
 }
 
 void EpilogueVectorizerMainLoop::printDebugTracesAtStart() {
@@ -7811,8 +7785,11 @@ EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton(
       VecEpilogueIterationCountCheck,
       VecEpilogueIterationCountCheck->getSinglePredecessor());
 
-  DT->changeImmediateDominator(LoopScalarPreHeader,
-                               EPI.EpilogueIterationCountCheck);
+  if (auto *N = DT->getNode(LoopScalarPreHeader))
+    DT->changeImmediateDominator(LoopScalarPreHeader,
+                                 EPI.EpilogueIterationCountCheck);
+  else
+    DT->addNewBlock(LoopScalarPreHeader, EPI.EpilogueIterationCountCheck);
   if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF.isVector()))
     // If there is an epilogue which must run, there's no edge from the
     // middle block to exit blocks  and thus no need to update the immediate
@@ -7876,7 +7853,7 @@ EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton(
                               {VecEpilogueIterationCountCheck,
                                EPI.VectorTripCount} /* AdditionalBypass */);
 
-  return {completeLoopSkeleton(), EPResumeVal};
+  return {LoopVectorPreHeader, EPResumeVal};
 }
 
 BasicBlock *
@@ -8625,9 +8602,25 @@ LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(VFRange &Range) {
   // modified; a basic block for the vector pre-header, followed by a region for
   // the vector loop, followed by the middle basic block. The skeleton vector
   // loop region contains a header and latch basic blocks.
+
+  // Add a check in the middle block to see if we have completed
+  // all of the iterations in the first vector loop.  Three cases:
+  // 1) If we require a scalar epilogue, there is no conditional branch as
+  //    we unconditionally branch to the scalar preheader.  Do nothing.
+  // 2) If (N - N%VF) == N, then we *don't* need to run the remainder.
+  //    Thus if tail is to be folded, we know we don't need to run the
+  //    remainder and we can use the previous value for the condition (true).
+  // 3) Otherwise, construct a runtime check.
+  bool RequiresScalarEpilogueCheck =
+      LoopVectorizationPlanner::getDecisionAndClampRange(
+          [this](ElementCount VF) {
+            return !CM.requiresScalarEpilogue(VF.isVector());
+          },
+          Range);
   VPlanPtr Plan = VPlan::createInitialVPlan(
       createTripCountSCEV(Legal->getWidestInductionType(), PSE, OrigLoop),
-      *PSE.getSE());
+      *PSE.getSE(), RequiresScalarEpilogueCheck, CM.foldTailByMasking(),
+      OrigLoop);
   VPBasicBlock *HeaderVPBB = new VPBasicBlock("vector.body");
   VPBasicBlock *LatchVPBB = new VPBasicBlock("vector.latch");
   VPBlockUtils::insertBlockAfter(LatchVPBB, HeaderVPBB);
@@ -8875,7 +8868,7 @@ VPlanPtr LoopVectorizationPlanner::buildVPlan(VFRange &Range) {
   // Create new empty VPlan
   auto Plan = VPlan::createInitialVPlan(
       createTripCountSCEV(Legal->getWidestInductionType(), PSE, OrigLoop),
-      *PSE.getSE());
+      *PSE.getSE(), true, false, OrigLoop);
 
   // Build hierarchical CFG
   VPlanHCFGBuilder HCFGBuilder(OrigLoop, LI, *Plan);
@@ -9084,6 +9077,9 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
     }
   }
   Builder.setInsertPoint(&*LatchVPBB->begin());
+  VPBasicBlock *MiddleVPBB =
+      cast<VPBasicBlock>(VectorLoopRegion->getSingleSuccessor());
+  VPBasicBlock::iterator IP = MiddleVPBB->begin();
   for (VPRecipeBase &R :
        Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) {
     VPReductionPHIRecipe *PhiR = dyn_cast<VPReductionPHIRecipe>(&R);
@@ -9192,8 +9188,8 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
     // also modeled in VPlan.
     auto *FinalReductionResult = new VPInstruction(
         VPInstruction::ComputeReductionResult, {PhiR, NewExitingVPV}, ExitDL);
-    cast<VPBasicBlock>(VectorLoopRegion->getSingleSuccessor())
-        ->appendRecipe(FinalReductionResult);
+    FinalReductionResult->insertBefore(*MiddleVPBB, IP);
+    IP = std::next(FinalReductionResult->getIterator());
     OrigExitingVPV->replaceUsesWithIf(
         FinalReductionResult,
         [](VPUser &User, unsigned) { return isa<VPLiveOut>(&User); });