[VPlan] Switch to checking sinking legality for recurrences in VPlan.

[fhahn]

Summary

Building on D142885 and D142589, retire the SinkAfter map from the
recurrence handling code. It is replaced by checking whether it is
possible to sink all users of a recurrence directly in VPlan. This
results in simpler code overall and allows to handle additional cases
(see the improvements in @test_crash).

Depends on D142885.
Depends on D142589.

[fhahn]

Rebase and ping :)

[Ayal]

llvm/include/llvm/Analysis/IVDescriptors.h

189–191

Leave a comment that this optimistically assumes instructions may be reordered successfully if needed?

llvm/lib/Analysis/IVDescriptors.cpp

967

(Relying here on Previous is fragile as it may have been considered for sinking.)

985

nit: such a set can be used to store "Seen" recipes avoiding revisiting recipes as they are added to the worklist and then iterating over them in dominance order to sink them.

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

564–565

nit: is this consideration of replicate regions is still needed here, or can be replaced by an assert?

615–621

Is this change to support multiple def'd recipes independent of the rest of this patch?

708

Is SeenPrevious really needed, to make sure chained FORs are processed in order, or can it be removed - given that every FOR takes care of its own sinkings followed by introducing its splice hooking it up to defs and uses? Potentially sinking candidates multiple times.

llvm/lib/Transforms/Vectorize/VPlanTransforms.h

74–75

Explain return value?

llvm/test/Transforms/LoopVectorize/first-order-recurrence-chains.ll

641–642

nit: update name of test, and its comment?

[fhahn]

Address latest comments, thanks!

llvm/include/llvm/Analysis/IVDescriptors.h

189–191

Added, thanks!

llvm/lib/Analysis/IVDescriptors.cpp

967

Agreed, but I think until we fully check legality in VPlan I think it is needed for now here (and I think it is not different to the current handling)

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

564–565

Thanks, adjusted in c18bc7f7feac.

615–621

Yep that's independent of the patch, I restored the original code here.

708

I checked and it is not needed in the latest version indeed! It also yields a nice improvement in @test_pr54223_sink_after_insertion_order! Updated to remove it.

llvm/lib/Transforms/Vectorize/VPlanTransforms.h

74–75

Added a comment, thanks!

llvm/test/Transforms/LoopVectorize/first-order-recurrence-chains.ll

641–642

Added, thanks!

[Ayal]

LGTM

llvm/lib/Analysis/IVDescriptors.cpp

967

(Above comment refers to ignoring if SinkAfter(Previous), which is the source of fragility/optimism.)

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

8982

TODO (Independent of this patch): Consider moving this potentially bailing-out check earlier to save compile-time for bailed-out cases, which would avoid vectorizing altogether.

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

569

Add meaning of return value?

591–594

"cannot" >> "must not"?
If SinkCandidate equals Previous a cycle is closed and phi is not a fixed order recurrence.

708

Very well! A direct consequence of VPlan's design to rely on updated def-use IR rather than recorded maps and sets.

[fhahn]

Address latest comments and rebase, thanks! I am planning on landing this shortly.

[manojgupta]

Heads up: This change is causing a clang hang when building xz utils (https://github.com/tukaani-project/xz). I'll file a bug and attach a repro asap.

[manojgupta]

Opened https://github.com/llvm/llvm-project/issues/62187 with a test case.

[fhahn]

Thanks for the report! I recommitted the change with a fix for the compile-time issue in 6b8d19d2b57ecd45eb2ff1964798f3cd8c907ac4. The new version of the patch adds a set to avoid duplicating work in isFixedOrderRecurrence, which was previously done through the removed SinkAfter map.

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

569

Thanks, documented!

591–594

Adjusted the comment, thanks!

llvm/include/llvm/Analysis/IVDescriptors.h

isReductionPHI(PHINode *Phi, Loop *TheLoop, RecurrenceDescriptor &RedDes,
                  DominatorTree *DT = nullptr, ScalarEvolution *SE = nullptr);
 
   /// Returns true if Phi is a fixed-order recurrence. A fixed-order recurrence
   /// is a non-reduction recurrence relation in which the value of the
   /// recurrence in the current loop iteration equals a value defined in a
   /// previous iteration (e.g. if the value is defined in the previous
   /// iteration, we refer to it as first-order recurrence, if it is defined in
   /// the iteration before the previous, we refer to it as second-order
-  /// recurrence and so on). \p SinkAfter includes pairs of instructions where
-  /// the first will be rescheduled to appear after the second if/when the loop
-  /// is vectorized. It may be augmented with additional pairs if needed in
-  /// order to handle Phi as a first-order recurrence.
-  static bool
-  isFixedOrderRecurrence(PHINode *Phi, Loop *TheLoop,
-                         MapVector<Instruction *, Instruction *> &SinkAfter,
+  /// recurrence and so on).
+  static bool isFixedOrderRecurrence(PHINode *Phi, Loop *TheLoop,
                                      DominatorTree *DT);
 
   RecurKind getRecurrenceKind() const { return Kind; }
 
   unsigned getOpcode() const { return getOpcode(getRecurrenceKind()); }
 
   FastMathFlags getFastMathFlags() const { return FMF; }
 
   TrackingVH<Value> getRecurrenceStartValue() const { return StartValue; }

llvm/include/llvm/Transforms/Vectorize/LoopVectorizationLegality.h

private:
   /// variables, and that have been proven to be redundant (possibly under a
   /// runtime guard). These casts can be ignored when creating the vectorized
   /// loop body.
   SmallPtrSet<Instruction *, 4> InductionCastsToIgnore;
 
   /// Holds the phi nodes that are fixed-order recurrences.
   RecurrenceSet FixedOrderRecurrences;
 
-  /// Holds instructions that need to sink past other instructions to handle
-  /// fixed-order recurrences.
-  MapVector<Instruction *, Instruction *> SinkAfter;
-
   /// Holds the widest induction type encountered.
   Type *WidestIndTy = nullptr;
 
   /// Allowed outside users. This holds the variables that can be accessed from
   /// outside the loop.
   SmallPtrSet<Value *, 4> AllowedExit;
 
   /// Vectorization requirements that will go through late-evaluation.

llvm/lib/Analysis/IVDescriptors.cpp

if (AddReductionVar(Phi, RecurKind::FMulAdd, TheLoop, FMF, RedDes, DB, AC, DT,
                       SE)) {
     LLVM_DEBUG(dbgs() << "Found an FMulAdd reduction PHI." << *Phi << "\n");
     return true;
   }
   // Not a reduction of known type.
   return false;
 }
 
-bool RecurrenceDescriptor::isFixedOrderRecurrence(
-    PHINode *Phi, Loop *TheLoop,
-    MapVector<Instruction *, Instruction *> &SinkAfter, DominatorTree *DT) {
+bool RecurrenceDescriptor::isFixedOrderRecurrence(PHINode *Phi, Loop *TheLoop,
+                                                  DominatorTree *DT) {
 
   // Ensure the phi node is in the loop header and has two incoming values.
   if (Phi->getParent() != TheLoop->getHeader() ||
       Phi->getNumIncomingValues() != 2)
     return false;
 
   // Ensure the loop has a preheader and a single latch block. The loop
   // vectorizer will need the latch to set up the next iteration of the loop.
bool RecurrenceDescriptor::isFixedOrderRecurrence(PHINode *Phi, Loop *TheLoop,
   while (auto *PrevPhi = dyn_cast_or_null<PHINode>(Previous)) {
     if (PrevPhi->getParent() != Phi->getParent())
       return false;
     if (!SeenPhis.insert(PrevPhi).second)
       return false;
     Previous = dyn_cast<Instruction>(PrevPhi->getIncomingValueForBlock(Latch));
   }
 
-  if (!Previous || !TheLoop->contains(Previous) || isa<PHINode>(Previous) ||
-      SinkAfter.count(Previous)) // Cannot rely on dominance due to motion.
+  if (!Previous || !TheLoop->contains(Previous) || isa<PHINode>(Previous))
     return false;
 
   // Ensure every user of the phi node (recursively) is dominated by the
   // previous value. The dominance requirement ensures the loop vectorizer will
   // not need to vectorize the initial value prior to the first iteration of the
   // loop.
   // TODO: Consider extending this sinking to handle memory instructions.
 
-  // We optimistically assume we can sink all users after Previous. Keep a set
-  // of instructions to sink after Previous ordered by dominance in the common
-  // basic block. It will be applied to SinkAfter if all users can be sunk.
-  auto CompareByComesBefore = [](const Instruction *A, const Instruction *B) {
-    return A->comesBefore(B);
-  };
-  std::set<Instruction *, decltype(CompareByComesBefore)> InstrsToSink(
-      CompareByComesBefore);
-
   BasicBlock *PhiBB = Phi->getParent();
   SmallVector<Instruction *, 8> WorkList;
   auto TryToPushSinkCandidate = [&](Instruction *SinkCandidate) {
-    // Already sunk SinkCandidate.
-    if (SinkCandidate->getParent() == PhiBB &&
-        InstrsToSink.find(SinkCandidate) != InstrsToSink.end())
-      return true;
-
     // Cyclic dependence.
     if (Previous == SinkCandidate)
       return false;
 
     if (DT->dominates(Previous,
                       SinkCandidate)) // We already are good w/o sinking.
       return true;
 
     if (SinkCandidate->getParent() != PhiBB ||
         SinkCandidate->mayHaveSideEffects() ||
         SinkCandidate->mayReadFromMemory() || SinkCandidate->isTerminator())
       return false;
-
-    // Avoid sinking an instruction multiple times (if multiple operands are
-    // fixed order recurrences) by sinking once - after the latest 'previous'
-    // instruction.
-    auto It = SinkAfter.find(SinkCandidate);
-    if (It != SinkAfter.end()) {
-      auto *OtherPrev = It->second;
-      // Find the earliest entry in the 'sink-after' chain. The last entry in
-      // the chain is the original 'Previous' for a recurrence handled earlier.
-      auto EarlierIt = SinkAfter.find(OtherPrev);
-      while (EarlierIt != SinkAfter.end()) {
-        Instruction *EarlierInst = EarlierIt->second;
-        EarlierIt = SinkAfter.find(EarlierInst);
-        // Bail out if order has not been preserved.
-        if (EarlierIt != SinkAfter.end() &&
-            !DT->dominates(EarlierInst, OtherPrev))
-          return false;
-        OtherPrev = EarlierInst;
-      }
-      // Bail out if order has not been preserved.
-      if (OtherPrev != It->second && !DT->dominates(It->second, OtherPrev))
-        return false;
-
-      // SinkCandidate is already being sunk after an instruction after
-      // Previous. Nothing left to do.
-      if (DT->dominates(Previous, OtherPrev) || Previous == OtherPrev)
-        return true;
-
-      // If there are other instructions to be sunk after SinkCandidate, remove
-      // and re-insert SinkCandidate can break those instructions. Bail out for
-      // simplicity.
-      if (any_of(SinkAfter,
-          [SinkCandidate](const std::pair<Instruction *, Instruction *> &P) {
-            return P.second == SinkCandidate;
-          }))
-        return false;
-
-      // Otherwise, Previous comes after OtherPrev and SinkCandidate needs to be
-      // re-sunk to Previous, instead of sinking to OtherPrev. Remove
-      // SinkCandidate from SinkAfter to ensure it's insert position is updated.
-      SinkAfter.erase(SinkCandidate);
-    }
-
+    //
     // If we reach a PHI node that is not dominated by Previous, we reached a
     // header PHI. No need for sinking.
     if (isa<PHINode>(SinkCandidate))
       return true;
 
     // Sink User tentatively and check its users
-    InstrsToSink.insert(SinkCandidate);
     WorkList.push_back(SinkCandidate);
     return true;
   };
 
   WorkList.push_back(Phi);
   // Try to recursively sink instructions and their users after Previous.
   while (!WorkList.empty()) {
     Instruction *Current = WorkList.pop_back_val();
     for (User *User : Current->users()) {
       if (!TryToPushSinkCandidate(cast<Instruction>(User)))
         return false;
     }
   }
 
-  // We can sink all users of Phi. Update the mapping.
-  for (Instruction *I : InstrsToSink) {
-    SinkAfter[I] = Previous;
-    Previous = I;
-  }
   return true;
 }
 
 /// This function returns the identity element (or neutral element) for
 /// the operation K.
 Value *RecurrenceDescriptor::getRecurrenceIdentity(RecurKind K, Type *Tp,
                                                    FastMathFlags FMF) const {
   switch (K) {

llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp

for (Instruction &I : *BB) {
         // of these NotAllowedExit.
         InductionDescriptor ID;
         if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID)) {
           addInductionPhi(Phi, ID, AllowedExit);
           Requirements->addExactFPMathInst(ID.getExactFPMathInst());
           continue;
         }
 
-        if (RecurrenceDescriptor::isFixedOrderRecurrence(Phi, TheLoop,
-                                                         SinkAfter, DT)) {
+        if (RecurrenceDescriptor::isFixedOrderRecurrence(Phi, TheLoop, DT)) {
           AllowedExit.insert(Phi);
           FixedOrderRecurrences.insert(Phi);
           continue;
         }
 
         // As a last resort, coerce the PHI to a AddRec expression
         // and re-try classifying it a an induction PHI.
         if (InductionDescriptor::isInductionPHI(Phi, TheLoop, PSE, ID, true)) {
if (Inductions.empty()) {
           "integer loop induction variable could not be identified",
           "NoIntegerInductionVariable", ORE, TheLoop);
       return false;
     } else {
       LLVM_DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
     }
   }
 
-  // For fixed order recurrences, we use the previous value (incoming value from
-  // the latch) to check if it dominates all users of the recurrence. Bail out
-  // if we have to sink such an instruction for another recurrence, as the
-  // dominance requirement may not hold after sinking.
-  BasicBlock *LoopLatch = TheLoop->getLoopLatch();
-  if (any_of(FixedOrderRecurrences, [LoopLatch, this](const PHINode *Phi) {
-        Instruction *V =
-            cast<Instruction>(Phi->getIncomingValueForBlock(LoopLatch));
-        return SinkAfter.find(V) != SinkAfter.end();
-      }))
-    return false;
-
   // Now we know the widest induction type, check if our found induction
   // is the same size. If it's not, unset it here and InnerLoopVectorizer
   // will create another.
   if (PrimaryInduction && WidestIndTy != PrimaryInduction->getType())
     PrimaryInduction = nullptr;
 
   return true;
 }

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

std::optional<VPlanPtr> LoopVectorizationPlanner::buildVPlanWithVPRecipes(
 
   VPlanTransforms::removeRedundantCanonicalIVs(*Plan);
   VPlanTransforms::removeRedundantInductionCasts(*Plan);
 
   // Adjust the recipes for any inloop reductions.
   adjustRecipesForReductions(cast<VPBasicBlock>(TopRegion->getExiting()), Plan,
                              RecipeBuilder, Range.Start);
 
-  VPlanTransforms::fixFirstOrderRecurrences(*Plan, Builder);
+  if (!VPlanTransforms::fixFirstOrderRecurrences(*Plan, Builder))
+    return std::nullopt;
 
   // Interleave memory: for each Interleave Group we marked earlier as relevant
   // for this VPlan, replace the Recipes widening its memory instructions with a
   // single VPInterleaveRecipe at its insertion point.
   for (const auto *IG : InterleaveGroups) {
     auto *Recipe = cast<VPWidenMemoryInstructionRecipe>(
         RecipeBuilder.getRecipe(IG->getInsertPos()));
     SmallVector<VPValue *, 4> StoredValues;

llvm/lib/Transforms/Vectorize/VPlanTransforms.h

struct VPlanTransforms {
   /// provide them by building scalar steps off of the canonical scalar IV and
   /// update the original IV's users. This is an optional optimization to reduce
   /// the needs of vector extracts.
   static void optimizeInductions(VPlan &Plan, ScalarEvolution &SE);
 
   /// Remove redundant EpxandSCEVRecipes in \p Plan's entry block by replacing
   /// them with already existing recipes expanding the same SCEV expression.
   static void removeRedundantExpandSCEVRecipes(VPlan &Plan);
 
   /// Optimize \p Plan based on \p BestVF and \p BestUF. This may restrict the
   /// resulting plan to \p BestVF and \p BestUF.
   static void optimizeForVFAndUF(VPlan &Plan, ElementCount BestVF,
                                  unsigned BestUF,
                                  PredicatedScalarEvolution &PSE);
 
-  static void fixFirstOrderRecurrences(VPlan &Plan, VPBuilder &Builder);
+  static bool fixFirstOrderRecurrences(VPlan &Plan, VPBuilder &Builder);
 };
 
 } // namespace llvm
 
 #endif // LLVM_TRANSFORMS_VECTORIZE_VPLANTRANSFORMS_H

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

for (auto &R : *A->getParent()) {
       if (&R == A)
         return true;
       if (&R == B)
         return false;
     }
     llvm_unreachable("recipe not found");
   };
   if (A->getParent() == B->getParent())
     return LocalComesBefore(A, B);
   return VPDT.dominates(A->getParent(), B->getParent());
 }
 
 // Sink users of \p FOR after the recipe defining the previous value \p Previous
 // of the recurrence.
-static void
-sinkRecurrenceUsersAfterPrevious(VPFirstOrderRecurrencePHIRecipe *FOR,
-                                 VPRecipeBase *Previous,
-                                 VPDominatorTree &VPDT) {
+static bool sinkRecurrenceUsersAfterPrevious(
+    VPFirstOrderRecurrencePHIRecipe *FOR, VPRecipeBase *Previous,
+    SmallPtrSetImpl<VPRecipeBase *> &SeenPrevious, VPDominatorTree &VPDT) {
   auto CompareByComesBefore = [&](const VPRecipeBase *A,
                                   const VPRecipeBase *B) {
     auto *RegionA = GetReplicateRegion(const_cast<VPRecipeBase *>(A));
     auto *RegionB = GetReplicateRegion(const_cast<VPRecipeBase *>(B));
     if (RegionA && RegionB)
       return VPDT.dominates(RegionA, RegionB);
 
     return recipeComesBefore(A, B, VPDT);
   };
 
   // Collect recipes that need sinking.
   SmallVector<VPRecipeBase *> WorkList;
   SmallPtrSet<VPRecipeBase *, 8> Seen;
   Seen.insert(Previous);
   // Keep recipes to sink ordered by dominance so earlier instructions are
   // processed first.
   std::set<VPRecipeBase *, decltype(CompareByComesBefore)> InstrsToSink(
       CompareByComesBefore);
   auto TryToPushSinkCandidate = [&](VPRecipeBase *SinkCandidate) {
+    if (SinkCandidate == Previous)
+      return false;
     if (isa<VPHeaderPHIRecipe>(SinkCandidate) ||
         !Seen.insert(SinkCandidate).second ||
         recipeComesBefore(Previous, SinkCandidate, VPDT))
-      return;
+      return true;
+
+    if (SeenPrevious.contains(SinkCandidate))
+      return false;
 
     WorkList.push_back(SinkCandidate);
 
     // VPPredInstPHIRecipes don't need sinking, because they will be sunk when
     // sinking the containing replicate region.
     if (!isa<VPPredInstPHIRecipe>(SinkCandidate))
       InstrsToSink.insert(SinkCandidate);
+    return true;
   };
 
   // Recursively sink users of FOR after Previous.
   WorkList.push_back(FOR);
   for (unsigned I = 0; I != WorkList.size(); ++I) {
     VPRecipeBase *Current = WorkList[I];
-    for (VPUser *User : Current->getVPSingleValue()->users()) {
+
+    for (VPValue *Val : Current->definedValues()) {
+      for (VPUser *User : Val->users()) {
         if (auto *R = dyn_cast<VPRecipeBase>(User))
-        TryToPushSinkCandidate(R);
+          if (!TryToPushSinkCandidate(R))
+            return false;
+      }
     }
   }
   for (VPRecipeBase *SinkCandidate : InstrsToSink) {
     VPRecipeBase *Target = Previous;
     for (VPValue *Op : SinkCandidate->operands()) {
       VPRecipeBase *Curr = Op->getDefiningRecipe();
       if (!Curr)
         continue;
if (TargetRegion) {
 
       VPBlockUtils::disconnectBlocks(SplitPred, SplitBlock);
       VPBlockUtils::connectBlocks(SplitPred, SinkRegion);
       VPBlockUtils::connectBlocks(SinkRegion, SplitBlock);
     }
     // We modified the CFG, update dominator tree.
     VPDT.recalculate(*SinkRegion->getPlan());
   }
+  return true;
 }
 
-void VPlanTransforms::fixFirstOrderRecurrences(VPlan &Plan,
+bool VPlanTransforms::fixFirstOrderRecurrences(VPlan &Plan,
                                                VPBuilder &Builder) {
   VPDominatorTree VPDT;
   VPDT.recalculate(Plan);
 
+  SmallPtrSet<VPRecipeBase *, 4> SeenPrevious;
   for (VPRecipeBase &R :
        Plan.getVectorLoopRegion()->getEntry()->getEntryBasicBlock()->phis()) {
     auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R);
     if (!FOR)
       continue;
 
     SmallPtrSet<VPFirstOrderRecurrencePHIRecipe *, 4> SeenPhis;
     VPRecipeBase *Previous = FOR->getPreviousValue()->getDefiningRecipe();
     // Fixed-order recurrences do not contain cycles, so this loop is guaranteed
     // to terminate.
     while (auto *PrevPhi =
                dyn_cast_or_null<VPFirstOrderRecurrencePHIRecipe>(Previous)) {
       assert(PrevPhi->getParent() == FOR->getParent());
       assert(SeenPhis.insert(PrevPhi).second);
       Previous = PrevPhi->getPreviousValue()->getDefiningRecipe();
     }
 
-    sinkRecurrenceUsersAfterPrevious(FOR, Previous, VPDT);
+    if (!sinkRecurrenceUsersAfterPrevious(FOR, Previous, SeenPrevious, VPDT))
+      return false;
+    SeenPrevious.insert(Previous);
 
     // Introduce a recipe to combine the incoming and previous values of a
     // fixed-order recurrence.
     VPBasicBlock *InsertBlock = Previous->getParent();
     auto *Region = GetReplicateRegion(Previous);
     if (Region)
       InsertBlock = dyn_cast<VPBasicBlock>(Region->getSingleSuccessor());
     if (!InsertBlock) {
auto *RecurSplice = cast<VPInstruction>(
         Builder.createNaryOp(VPInstruction::FirstOrderRecurrenceSplice,
                              {FOR, FOR->getBackedgeValue()}));
 
     FOR->replaceAllUsesWith(RecurSplice);
     // Set the first operand of RecurSplice to FOR again, after replacing
     // all users.
     RecurSplice->setOperand(0, FOR);
   }
+  return true;
 }

llvm/test/Transforms/LoopVectorize/first-order-recurrence-chains.ll

loop:
   %ptr.iv.next = getelementptr i32, ptr %ptr.iv, i64 1
   %exitcond.not = icmp eq i64 %iv.next, 1000
   br i1 %exitcond.not, label %exit, label %loop
 
 exit:
   ret void
 }
 
-; Make sure LLVM doesn't generate wrong data in SinkAfter, and causes crash in
-; loop vectorizer.
 define void @test_crash(ptr %p) {
-; CHECK-LABEL: @test_crash
-; CHECK-NOT:   vector.body:
-; CHECK:       ret
+; CHECK-LABEL: @test_crash(
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT:%.*]], %vector.body ]
+; CHECK-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x double> [ <double poison, double poison, double poison, double 0.000000e+00>, %vector.ph ], [ [[BROADCAST_SPLAT:%.*]], %vector.body ]
+; CHECK-NEXT:    [[VECTOR_RECUR1:%.*]] = phi <4 x double> [ <double poison, double poison, double poison, double 0.000000e+00>, %vector.ph ], [ [[BROADCAST_SPLAT4:%.*]], %vector.body ]
+; CHECK-NEXT:    [[VECTOR_RECUR2:%.*]] = phi <4 x double> [ <double poison, double poison, double poison, double 0.000000e+00>, %vector.ph ], [ [[TMP4:%.*]], %vector.body ]
+; CHECK-NEXT:    [[TMP0:%.*]] = load double, ptr null, align 8
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x double> poison, double [[TMP0]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT]] = shufflevector <4 x double> [[BROADCAST_SPLATINSERT]], <4 x double> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <4 x double> [[VECTOR_RECUR]], <4 x double> [[BROADCAST_SPLAT]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
+; CHECK-NEXT:    [[TMP2:%.*]] = fdiv <4 x double> zeroinitializer, [[TMP1]]
+; CHECK-NEXT:    [[TMP3:%.*]] = load double, ptr null, align 8
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <4 x double> poison, double [[TMP3]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT4]] = shufflevector <4 x double> [[BROADCAST_SPLATINSERT3]], <4 x double> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP4]] = shufflevector <4 x double> [[VECTOR_RECUR1]], <4 x double> [[BROADCAST_SPLAT4]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
+; CHECK-NEXT:    [[TMP5:%.*]] = shufflevector <4 x double> [[VECTOR_RECUR2]], <4 x double> [[TMP4]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
+; CHECK-NEXT:    [[TMP6:%.*]] = extractelement <4 x double> [[TMP2]], i32 3
+; CHECK-NEXT:    store double [[TMP6]], ptr [[P:%.*]], align 8
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 0
+; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP28:![0-9]+]]
+; CHECK:       middle.block:
+; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 0, 0
+; CHECK-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x double> [[BROADCAST_SPLAT]], i32 3
+; CHECK-NEXT:    [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x double> [[BROADCAST_SPLAT]], i32 2
+; CHECK-NEXT:    [[VECTOR_RECUR_EXTRACT5:%.*]] = extractelement <4 x double> [[BROADCAST_SPLAT4]], i32 3
+; CHECK-NEXT:    [[VECTOR_RECUR_EXTRACT_FOR_PHI6:%.*]] = extractelement <4 x double> [[BROADCAST_SPLAT4]], i32 2
+; CHECK-NEXT:    [[VECTOR_RECUR_EXTRACT9:%.*]] = extractelement <4 x double> [[TMP4]], i32 3
+; CHECK-NEXT:    [[VECTOR_RECUR_EXTRACT_FOR_PHI10:%.*]] = extractelement <4 x double> [[TMP4]], i32 2
+; CHECK-NEXT:    br i1 [[CMP_N]], label %End, label %scalar.ph
+;
 Entry:
   br label %Loop
 
 Loop:
   %for.1 = phi double [ %iv1, %Loop ], [ 0.000000e+00, %Entry ]
   %for.2 = phi double [ %iv2, %Loop ], [ 0.000000e+00, %Entry ]
   %for.3 = phi double [ %for.2, %Loop ], [ 0.000000e+00, %Entry ]
   %for.4 = phi i64 [ %count, %Loop ], [ 0, %Entry ]