[LoopVectorizer] Lower uniform loads as a single load (instead of relying on CSE)

[reames]

Summary

A uniform load is one which loads from a uniform address across all lanes. As currently implemented, we cost model such loads as if we did a single scalar load + a broadcast, but the actual lowering replicates the load once per lane.

This change tweaks the lowering to use the REPLICATE strategy by marking such loads (and the computation leading to their memory operand) as uniform after vectorization. This is a useful change in itself, but it's real purpose is to pave the way for a following change which will generalize our uniformity logic.

[reames]

Rebase w/over landed tests.

As part of that, include a slight broadening of scope. For some reason, the code wasn't considering any operand which wasn't an instruction to be uniform. Since I happened to write my test with memory addresses as arguments, this fell out.

[reames]

Remove stray debug output and an extra bit of whitespace.

[fhahn]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

2667

The VPRecplicateRecipe contains a IsUniform flag. I think it should be possible to pass the flag through from the recipe to scalarizeInstruction. Ideally the recipes should contain all information required for code-generation to avoid having to tie code generation directly to the cost-model.

5079

nit: message for assert.

6272

nit: message for assert?

6451

this seems an unrelated refactoring, which could be split out and committed independently?

[fhahn]

There's D68831 which is potentially related and tries to extend the logic to consider loop-invariant ops as uniform. I suppose I need to rebase it again....

[reames]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

2667

I think you're misreading the code slightly. This isn't checking whether the recipe for the instruction being scalarized is uniform. It's checking whether the *input* to the instruction is uniform.

It seems like overkill to record a per operand uniform flag in the recipe?

6272

Aside from being pedantic, why? I'm happy to comply, but I don't really see any value in cases like this where it's obvious from context.

6451

Happy to do so since you seem to think the interface made sense. I'm very new to vectorizer code and didn't want to jump to that conclusion. :)

[fhahn]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

2667

I think you're misreading the code slightly. This isn't checking whether the recipe for the instruction being scalarized is uniform. It's checking whether the *input* to the instruction is uniform.

Indeed, I was thinking about something slightly differently. The underlying point about ideally not having this code depend on the cost-model, but the information in the corresponding VPlan still applies.

It seems like overkill to record a per operand uniform flag in the recipe?

Yes, but fortunately I don't think that is necessary, because I think all required information is already encoded in the operands in VPlan: they should be uniform VPReplicateRecipes. Making this information accessible here is currently work-in-progress. With D91500 which I just put up, you should be able to check if User.getOperand(op) is a VPReplicateRecipe and if so, IsUniform also needs to be true.

I think it would make sense to take it one step further, and use this logic directly in VPTransformState::get: when requesting a particular lane for a uniform VPValue, we should always be able to just return lane 0 and other callers potentially could also benefit. With D91501, the changes to scalarizeInstruction should not be needed.

6272

I agree this one is borderline and a message like "must be called with a uniform memory instruction" probably does not add too much. It might make it slightly more explicit why this assertion is here for people not familiar with the code. I don't really mind either way :)

[anna]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

2667

Frankly, I'm finding the code written here in review easier to read :) Mainly because my context with VPlan is limited. Do we need to block this change on two reviews landing? I'm okay either way - but we can also land this and then "clean this up" once D91500 and D91501 lands?

I think it would make sense to take it one step further, and use this logic directly in VPTransformState::get: when requesting a particular lane for a uniform VPValue, we should always be able to just return lane 0 and other callers potentially could also benefit.

@fhahn Could you pls clarify further what you mean by "other callers potentially could also benefit" ?

5039

Could this be turned into an assert and landed separately? I see all callers of addToWorklistIfAllowed either already checks for outOfScope or the instructions is already checked to be in loop.

[fhahn]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

2667

Frankly, I'm finding the code written here in review easier to read :) Mainly because my context with VPlan is limited. Do we need to block this change on two reviews landing? I'm okay either way - but we can also land this and then "clean this up" once D91500 and D91501 lands?

The linked patches are 'more' code yes, but most of the code is part of current in-progress improvements and the uniform handling is mostly a nice side benefit of VPlanization.

We don't necessarily need to block this change, I just wanted to note that this is a step backwards in terms of the general direction (quite a bit of work was spent on moving any cost-model references out of code generation). It's just a small step though and we have a clear path forward to resolve this.

@fhahn Could you pls clarify further what you mean by "other callers potentially could also benefit" ?

There might be other places that request individual lanes for uniform values. Those would also benefit from only getting the first lane.

[reames]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

2667

I've gone ahead and approved your D91501. I don't think it fully solves the problem, but I see little harm in it.

Your proposed approach creates a coupling between the lowering strategy chosen and the semantic property of uniformity. To some extend that might be unavoidable, but can we do better?

As an example case, consider an udiv(add X, Y), Z) where all inputs are loop invariant. If X and Y already have vector uses, we might reasonable decide that a widened vector add is the best lowering for this uniform expression. (Not sure if we do today, but it's a reasonable choice.) When considering the udiv, if we decide to replicate it (because we know it's uniform and expensive), we want to be able to extra lane 0 regardless of the lowering strategy we chose for the add.

Another approach is to directly skip the cost model and ask the legality question. That works for the example I just gave, but *doesn't* work with the way we currently handle GEPs feeding wideable loads - which is entirely handled in costing.

I suspect we need to move the GEP detection out of the cost model, but I don't fully understand the broader implications of that.

[reames]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5039

The check is needed for case where a uniform pointer is defined outside the loop but used inside of it. I could move the check to the caller, but the placement seems to make more sense here?

[reames]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

2667

We don't necessarily need to block this change, I just wanted to note that this is a step backwards in terms of the general direction (quite a bit of work was spent on moving any cost-model references out of code generation). It's just a small step though and we have a clear path forward to resolve this.

Could I get an LGTM then?

I'm happy to help evolve towards a good overall design, but it's really hard to make progress when the first patch which has a functional effect gets blocked in review.

[fhahn]

LGTM, thanks. I also added @Ayal, who might also have some thoughts. It would probably be best to wait a few days with committing, in case there are additional comments.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

2667

Your proposed approach creates a coupling between the lowering strategy chosen and the semantic property of uniformity. To some extend that might be unavoidable, but can we do better?

Yes, the VPlan changes I put up just ask the same question in a different way. I think we should be able to improve the code generated for your example, but that should probably happen before codegen, e.g. by widening.

Another approach is to directly skip the cost model and ask the legality question. That works for the example I just gave, but *doesn't* work with the way we currently handle GEPs feeding wideable loads - which is entirely handled in costing.

Asking legality here comes with the same problematic coupling which hinders modularizing transforms via VPlan I think.

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

public:
   /// -1 - Address is consecutive, and decreasing.
   /// NOTE: This method must only be used before modifying the original scalar
   /// loop. Do not use after invoking 'createVectorizedLoopSkeleton' (PR34965).
   int isConsecutivePtr(Value *Ptr);
 
   /// Returns true if the value V is uniform within the loop.
   bool isUniform(Value *V);
 
+  /// A uniform memory op is a load or store which accesses the same memory
+  /// location on all lanes.
+  bool isUniformMemOp(Instruction &I) {
+    Value *Ptr = getLoadStorePointerOperand(&I);
+    if (!Ptr)
+      return false;
+    // Note: There's nothing inherent which prevents predicated loads and
+    // stores from being uniform.  The current lowering simply doesn't handle
+    // it; in particular, the cost model distinguishes scatter/gather from
+    // scalar w/predication, and we currently rely on the scalar path.
+    return isUniform(Ptr) && !blockNeedsPredication(I.getParent());
+  }
+
   /// Returns the information that we collected about runtime memory check.
   const RuntimePointerChecking *getRuntimePointerChecking() const {
     return LAI->getRuntimePointerChecking();
   }
 
   const LoopAccessInfo *getLAI() const { return LAI; }
 
   unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, VPUser &User,
 
   Instruction *Cloned = Instr->clone();
   if (!IsVoidRetTy)
     Cloned->setName(Instr->getName() + ".cloned");
 
   // Replace the operands of the cloned instructions with their scalar
   // equivalents in the new loop.
   for (unsigned op = 0, e = User.getNumOperands(); op != e; ++op) {
-    auto *NewOp = State.get(User.getOperand(op), Instance);
+    auto *Operand = dyn_cast<Instruction>(Instr->getOperand(op));
+    auto InputInstance = Instance;
+    if (!Operand || !OrigLoop->contains(Operand) ||
+        (Cost->isUniformAfterVectorization(Operand, State.VF)))
+      InputInstance.Lane = 0;
+    auto *NewOp = State.get(User.getOperand(op), InputInstance);
     Cloned->setOperand(op, NewOp);
   }
   addNewMetadata(Cloned, Instr);
 
   // Place the cloned scalar in the new loop.
   Builder.Insert(Cloned);
 
   // Add the cloned scalar to the scalar map entry.
void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) {
   SetVector<Instruction *> Worklist;
   BasicBlock *Latch = TheLoop->getLoopLatch();
 
   // Instructions that are scalar with predication must not be considered
   // uniform after vectorization, because that would create an erroneous
   // replicating region where only a single instance out of VF should be formed.
   // TODO: optimize such seldom cases if found important, see PR40816.
   auto addToWorklistIfAllowed = [&](Instruction *I) -> void {
+    if (isOutOfScope(I)) {
+      LLVM_DEBUG(dbgs() << "LV: Found not uniform due to scope: "
+                        << *I << "\n");
+      return;
+    }
     if (isScalarWithPredication(I, VF)) {
       LLVM_DEBUG(dbgs() << "LV: Found not uniform being ScalarWithPredication: "
                         << *I << "\n");
       return;
     }
     LLVM_DEBUG(dbgs() << "LV: Found uniform instruction: " << *I << "\n");
     Worklist.insert(I);
   };
 
   // Start with the conditional branch. If the branch condition is an
   // instruction contained in the loop that is only used by the branch, it is
   // uniform.
   auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0));
   if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse())
     addToWorklistIfAllowed(Cmp);
 
   // Holds consecutive and consecutive-like pointers. Consecutive-like pointers
   // are pointers that are treated like consecutive pointers during
   // vectorization. The pointer operands of interleaved accesses are an
   // example.
-  SmallSetVector<Instruction *, 8> ConsecutiveLikePtrs;
+  SmallSetVector<Value *, 8> ConsecutiveLikePtrs;
 
   // Holds pointer operands of instructions that are possibly non-uniform.
-  SmallPtrSet<Instruction *, 8> PossibleNonUniformPtrs;
+  SmallPtrSet<Value *, 8> PossibleNonUniformPtrs;
 
   auto isUniformDecision = [&](Instruction *I, ElementCount VF) {
     InstWidening WideningDecision = getWideningDecision(I, VF);
     assert(WideningDecision != CM_Unknown &&
            "Widening decision should be ready at this moment");
 
+    // The address of a uniform mem op is itself uniform.  We exclude stores
+    // here as there's an assumption in the current code that all uses of
+    // uniform instructions are uniform and, as noted below, uniform stores are
+    // still handled via replication (i.e. aren't uniform after vectorization).
+    if (isa<LoadInst>(I) && Legal->isUniformMemOp(*I)) {
+      assert(WideningDecision == CM_Scalarize);
+      return true;
+    }
+
     return (WideningDecision == CM_Widen ||
             WideningDecision == CM_Widen_Reverse ||
             WideningDecision == CM_Interleave);
   };
   // Iterate over the instructions in the loop, and collect all
   // consecutive-like pointer operands in ConsecutiveLikePtrs. If it's possible
   // that a consecutive-like pointer operand will be scalarized, we collect it
   // in PossibleNonUniformPtrs instead. We use two sets here because a single
   // getelementptr instruction can be used by both vectorized and scalarized
   // memory instructions. For example, if a loop loads and stores from the same
   // location, but the store is conditional, the store will be scalarized, and
   // the getelementptr won't remain uniform.
   for (auto *BB : TheLoop->blocks())
     for (auto &I : *BB) {
       // If there's no pointer operand, there's nothing to do.
-      auto *Ptr = dyn_cast_or_null<Instruction>(getLoadStorePointerOperand(&I));
+      auto *Ptr = getLoadStorePointerOperand(&I);
       if (!Ptr)
         continue;
 
+      // For now, avoid walking use lists in other functions.
+      // TODO: Rewrite this algorithm from uses up.
+      if (!isa<Instruction>(Ptr) && !isa<Argument>(Ptr))
+        continue;
+
+      // A uniform memory op is itself uniform.  We exclude stores here as we
+      // haven't yet added dedicated logic in the CLONE path and rely on
+      // REPLICATE + DSE for correctness.
+      if (isa<LoadInst>(I) && Legal->isUniformMemOp(I))
+        addToWorklistIfAllowed(&I);
+
       // True if all users of Ptr are memory accesses that have Ptr as their
       // pointer operand.
       auto UsersAreMemAccesses =
           llvm::all_of(Ptr->users(), [&](User *U) -> bool {
             return getLoadStorePointerOperand(U) == Ptr;
           });
 
       // Ensure the memory instruction will not be scalarized or used by
for (auto &I : *BB) {
       else
         ConsecutiveLikePtrs.insert(Ptr);
     }
 
   // Add to the Worklist all consecutive and consecutive-like pointers that
   // aren't also identified as possibly non-uniform.
   for (auto *V : ConsecutiveLikePtrs)
     if (!PossibleNonUniformPtrs.count(V))
-      addToWorklistIfAllowed(V);
+      if (auto *I = dyn_cast<Instruction>(V))
+        addToWorklistIfAllowed(I);
 
   // Expand Worklist in topological order: whenever a new instruction
   // is added , its users should be already inside Worklist.  It ensures
   // a uniform instruction will only be used by uniform instructions.
   unsigned idx = 0;
   while (idx != Worklist.size()) {
     Instruction *I = Worklist[idx++];
 
unsigned LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I,
   bool Reverse = ConsecutiveStride < 0;
   if (Reverse)
     Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
   return Cost;
 }
 
 unsigned LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I,
                                                          ElementCount VF) {
+  assert(Legal->isUniformMemOp(*I));
+
   Type *ValTy = getMemInstValueType(I);
   auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF));
   const Align Alignment = getLoadStoreAlignment(I);
   unsigned AS = getLoadStoreAddressSpace(I);
   enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
   if (isa<LoadInst>(I)) {
     return TTI.getAddressComputationCost(ValTy) +
            TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS,
for (Instruction &I : *BB) {
 
       // TODO: We should generate better code and update the cost model for
       // predicated uniform stores. Today they are treated as any other
       // predicated store (see added test cases in
       // invariant-store-vectorization.ll).
       if (isa<StoreInst>(&I) && isScalarWithPredication(&I))
         NumPredStores++;
 
-      if (Legal->isUniform(Ptr) &&
-          // Conditional loads and stores should be scalarized and predicated.
-          // isScalarWithPredication cannot be used here since masked
-          // gather/scatters are not considered scalar with predication.
-          !Legal->blockNeedsPredication(I.getParent())) {
+      if (Legal->isUniformMemOp(I)) {
         // TODO: Avoid replicating loads and stores instead of
         // relying on instcombine to remove them.
         // Load: Scalar load + broadcast
         // Store: Scalar store + isLoopInvariantStoreValue ? 0 : extract
         unsigned Cost = getUniformMemOpCost(&I, VF);
         setWideningDecision(&I, VF, CM_Scalarize, Cost);
         continue;
       }

llvm/test/Transforms/LoopVectorize/X86/cost-model-assert.ll

 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
 ; CHECK-NEXT:    [[TMP1:%.*]] = mul i64 [[TMP0]], 4
 ; CHECK-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i8, i8* null, i64 [[TMP1]]
 ; CHECK-NEXT:    [[TMP2:%.*]] = zext <4 x i8> [[BROADCAST_SPLAT]] to <4 x i32>
 ; CHECK-NEXT:    [[TMP3:%.*]] = shl nuw <4 x i32> [[TMP2]], <i32 24, i32 24, i32 24, i32 24>
-; CHECK-NEXT:    [[TMP4:%.*]] = load i8, i8* [[P:%.*]], align 1, !tbaa !1
-; CHECK-NEXT:    [[TMP5:%.*]] = load i8, i8* [[P]], align 1, !tbaa !1
-; CHECK-NEXT:    [[TMP6:%.*]] = load i8, i8* [[P]], align 1, !tbaa !1
-; CHECK-NEXT:    [[TMP7:%.*]] = load i8, i8* [[P]], align 1, !tbaa !1
-; CHECK-NEXT:    [[TMP8:%.*]] = insertelement <4 x i8> undef, i8 [[TMP4]], i32 0
-; CHECK-NEXT:    [[TMP9:%.*]] = insertelement <4 x i8> [[TMP8]], i8 [[TMP5]], i32 1
-; CHECK-NEXT:    [[TMP10:%.*]] = insertelement <4 x i8> [[TMP9]], i8 [[TMP6]], i32 2
-; CHECK-NEXT:    [[TMP11:%.*]] = insertelement <4 x i8> [[TMP10]], i8 [[TMP7]], i32 3
-; CHECK-NEXT:    [[TMP12:%.*]] = zext <4 x i8> [[TMP11]] to <4 x i32>
-; CHECK-NEXT:    [[TMP13:%.*]] = shl nuw nsw <4 x i32> [[TMP12]], <i32 16, i32 16, i32 16, i32 16>
-; CHECK-NEXT:    [[TMP14:%.*]] = or <4 x i32> [[TMP13]], [[TMP3]]
-; CHECK-NEXT:    [[TMP15:%.*]] = load i8, i8* undef, align 1, !tbaa !1
-; CHECK-NEXT:    [[TMP16:%.*]] = load i8, i8* undef, align 1, !tbaa !1
-; CHECK-NEXT:    [[TMP17:%.*]] = load i8, i8* undef, align 1, !tbaa !1
-; CHECK-NEXT:    [[TMP18:%.*]] = load i8, i8* undef, align 1, !tbaa !1
-; CHECK-NEXT:    [[TMP19:%.*]] = or <4 x i32> [[TMP14]], zeroinitializer
-; CHECK-NEXT:    [[TMP20:%.*]] = or <4 x i32> [[TMP19]], zeroinitializer
-; CHECK-NEXT:    [[TMP21:%.*]] = extractelement <4 x i32> [[TMP20]], i32 0
-; CHECK-NEXT:    store i32 [[TMP21]], i32* undef, align 4, !tbaa !4
-; CHECK-NEXT:    [[TMP22:%.*]] = extractelement <4 x i32> [[TMP20]], i32 1
-; CHECK-NEXT:    store i32 [[TMP22]], i32* undef, align 4, !tbaa !4
-; CHECK-NEXT:    [[TMP23:%.*]] = extractelement <4 x i32> [[TMP20]], i32 2
-; CHECK-NEXT:    store i32 [[TMP23]], i32* undef, align 4, !tbaa !4
-; CHECK-NEXT:    [[TMP24:%.*]] = extractelement <4 x i32> [[TMP20]], i32 3
-; CHECK-NEXT:    store i32 [[TMP24]], i32* undef, align 4, !tbaa !4
+; CHECK-NEXT:    [[TMP4:%.*]] = load i8, i8* [[P:%.*]], align 1, [[TBAA1:!tbaa !.*]]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <4 x i8> undef, i8 [[TMP4]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <4 x i8> [[BROADCAST_SPLATINSERT1]], <4 x i8> undef, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = zext <4 x i8> [[BROADCAST_SPLAT2]] to <4 x i32>
+; CHECK-NEXT:    [[TMP6:%.*]] = shl nuw nsw <4 x i32> [[TMP5]], <i32 16, i32 16, i32 16, i32 16>
+; CHECK-NEXT:    [[TMP7:%.*]] = or <4 x i32> [[TMP6]], [[TMP3]]
+; CHECK-NEXT:    [[TMP8:%.*]] = load i8, i8* undef, align 1, [[TBAA1]]
+; CHECK-NEXT:    [[TMP9:%.*]] = load i8, i8* undef, align 1, [[TBAA1]]
+; CHECK-NEXT:    [[TMP10:%.*]] = load i8, i8* undef, align 1, [[TBAA1]]
+; CHECK-NEXT:    [[TMP11:%.*]] = load i8, i8* undef, align 1, [[TBAA1]]
+; CHECK-NEXT:    [[TMP12:%.*]] = or <4 x i32> [[TMP7]], zeroinitializer
+; CHECK-NEXT:    [[TMP13:%.*]] = or <4 x i32> [[TMP12]], zeroinitializer
+; CHECK-NEXT:    [[TMP14:%.*]] = extractelement <4 x i32> [[TMP13]], i32 0
+; CHECK-NEXT:    store i32 [[TMP14]], i32* undef, align 4, [[TBAA4:!tbaa !.*]]
+; CHECK-NEXT:    [[TMP15:%.*]] = extractelement <4 x i32> [[TMP13]], i32 1
+; CHECK-NEXT:    store i32 [[TMP15]], i32* undef, align 4, [[TBAA4]]
+; CHECK-NEXT:    [[TMP16:%.*]] = extractelement <4 x i32> [[TMP13]], i32 2
+; CHECK-NEXT:    store i32 [[TMP16]], i32* undef, align 4, [[TBAA4]]
+; CHECK-NEXT:    [[TMP17:%.*]] = extractelement <4 x i32> [[TMP13]], i32 3
+; CHECK-NEXT:    store i32 [[TMP17]], i32* undef, align 4, [[TBAA4]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP25:%.*]] = icmp eq i64 [[INDEX_NEXT]], 0
-; CHECK-NEXT:    br i1 [[TMP25]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !6
+; CHECK-NEXT:    [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], 0
+; CHECK-NEXT:    br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP6:!llvm.loop !.*]]
 ; CHECK:       middle.block:
 ; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 1, 0
 ; CHECK-NEXT:    br i1 [[CMP_N]], label [[SW_EPILOG:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i8* [ null, [[MIDDLE_BLOCK]] ], [ null, [[IF_THEN]] ]
 ; CHECK-NEXT:    br label [[FOR_BODY68:%.*]]
 ; CHECK:       for.body68:
 ; CHECK-NEXT:    [[P_359:%.*]] = phi i8* [ [[ADD_PTR86:%.*]], [[FOR_BODY68]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
 ; CHECK-NEXT:    [[CONV70:%.*]] = zext i8 [[X]] to i32
 ; CHECK-NEXT:    [[SHL71:%.*]] = shl nuw i32 [[CONV70]], 24
-; CHECK-NEXT:    [[TMP26:%.*]] = load i8, i8* [[P]], align 1, !tbaa !1
-; CHECK-NEXT:    [[CONV73:%.*]] = zext i8 [[TMP26]] to i32
+; CHECK-NEXT:    [[TMP19:%.*]] = load i8, i8* [[P]], align 1, [[TBAA1]]
+; CHECK-NEXT:    [[CONV73:%.*]] = zext i8 [[TMP19]] to i32
 ; CHECK-NEXT:    [[SHL74:%.*]] = shl nuw nsw i32 [[CONV73]], 16
 ; CHECK-NEXT:    [[OR75:%.*]] = or i32 [[SHL74]], [[SHL71]]
-; CHECK-NEXT:    [[TMP27:%.*]] = load i8, i8* undef, align 1, !tbaa !1
+; CHECK-NEXT:    [[TMP20:%.*]] = load i8, i8* undef, align 1, [[TBAA1]]
 ; CHECK-NEXT:    [[SHL78:%.*]] = shl nuw nsw i32 undef, 8
 ; CHECK-NEXT:    [[OR79:%.*]] = or i32 [[OR75]], [[SHL78]]
 ; CHECK-NEXT:    [[CONV81:%.*]] = zext i8 undef to i32
 ; CHECK-NEXT:    [[OR83:%.*]] = or i32 [[OR79]], [[CONV81]]
-; CHECK-NEXT:    store i32 [[OR83]], i32* undef, align 4, !tbaa !4
+; CHECK-NEXT:    store i32 [[OR83]], i32* undef, align 4, [[TBAA4]]
 ; CHECK-NEXT:    [[ADD_PTR86]] = getelementptr inbounds i8, i8* [[P_359]], i64 4
 ; CHECK-NEXT:    [[CMP66:%.*]] = icmp ult i8* [[ADD_PTR86]], undef
-; CHECK-NEXT:    br i1 [[CMP66]], label [[FOR_BODY68]], label [[SW_EPILOG]], !llvm.loop !8
+; CHECK-NEXT:    br i1 [[CMP66]], label [[FOR_BODY68]], label [[SW_EPILOG]], [[LOOP8:!llvm.loop !.*]]
 ; CHECK:       sw.epilog:
 ; CHECK-NEXT:    unreachable
 ; CHECK:       Exit:
 ; CHECK-NEXT:    ret void
 ;
 entry:
   br i1 %cond, label %if.then, label %Exit
 

llvm/test/Transforms/LoopVectorize/X86/uniform_mem_op.ll

 ; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
 ; CHECK-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = add i64 [[INDEX]], 8
 ; CHECK-NEXT:    [[TMP3:%.*]] = add i64 [[INDEX]], 12
 ; CHECK-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ADDR:%.*]], align 4
 ; CHECK-NEXT:    [[TMP5:%.*]] = load i32, i32* [[ADDR]], align 4
 ; CHECK-NEXT:    [[TMP6:%.*]] = load i32, i32* [[ADDR]], align 4
 ; CHECK-NEXT:    [[TMP7:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP8:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP9:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP10:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP11:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP12:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP13:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP14:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP15:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP16:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP17:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP18:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP19:%.*]] = load i32, i32* [[ADDR]], align 4
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 16
-; CHECK-NEXT:    [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4096
-; CHECK-NEXT:    br i1 [[TMP20]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP0:!llvm.loop !.*]]
+; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4096
+; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP0:!llvm.loop !.*]]
 ; CHECK:       middle.block:
 ; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 4097, 4096
 ; CHECK-NEXT:    br i1 [[CMP_N]], label [[LOOPEXIT:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 4096, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
 ; CHECK-NEXT:    [[LOAD:%.*]] = load i32, i32* [[ADDR]], align 4
 ; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
 ; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[IV]], 4096
 ; CHECK-NEXT:    br i1 [[EXITCOND]], label [[LOOPEXIT]], label [[FOR_BODY]], [[LOOP2:!llvm.loop !.*]]
 ; CHECK:       loopexit:
-; CHECK-NEXT:    [[LOAD_LCSSA:%.*]] = phi i32 [ [[LOAD]], [[FOR_BODY]] ], [ [[TMP19]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[LOAD_LCSSA:%.*]] = phi i32 [ [[LOAD]], [[FOR_BODY]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[LOAD_LCSSA]]
 ;
 entry:
   br label %for.body
 
 for.body:
   %iv = phi i64 [ %iv.next, %for.body ], [ 0, %entry ]
   %load = load i32, i32* %addr
   %iv.next = add nuw nsw i64 %iv, 1
   %exitcond = icmp eq i64 %iv, 4096
   br i1 %exitcond, label %loopexit, label %for.body
 
 loopexit:
   ret i32 %load
 }
 
 define i32 @uniform_load2(i32* align(4) %addr) {
 ; CHECK-LABEL: @uniform_load2(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP36:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP37:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP38:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI3:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP39:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP10:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI3:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP11:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
 ; CHECK-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = add i64 [[INDEX]], 8
 ; CHECK-NEXT:    [[TMP3:%.*]] = add i64 [[INDEX]], 12
 ; CHECK-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ADDR:%.*]], align 4
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> undef, i32 [[TMP4]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> undef, <4 x i32> zeroinitializer
 ; CHECK-NEXT:    [[TMP5:%.*]] = load i32, i32* [[ADDR]], align 4
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT4:%.*]] = insertelement <4 x i32> undef, i32 [[TMP5]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT5:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT4]], <4 x i32> undef, <4 x i32> zeroinitializer
 ; CHECK-NEXT:    [[TMP6:%.*]] = load i32, i32* [[ADDR]], align 4
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT6:%.*]] = insertelement <4 x i32> undef, i32 [[TMP6]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT7:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT6]], <4 x i32> undef, <4 x i32> zeroinitializer
 ; CHECK-NEXT:    [[TMP7:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP8:%.*]] = insertelement <4 x i32> undef, i32 [[TMP4]], i32 0
-; CHECK-NEXT:    [[TMP9:%.*]] = insertelement <4 x i32> [[TMP8]], i32 [[TMP5]], i32 1
-; CHECK-NEXT:    [[TMP10:%.*]] = insertelement <4 x i32> [[TMP9]], i32 [[TMP6]], i32 2
-; CHECK-NEXT:    [[TMP11:%.*]] = insertelement <4 x i32> [[TMP10]], i32 [[TMP7]], i32 3
-; CHECK-NEXT:    [[TMP12:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP13:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP14:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP15:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP16:%.*]] = insertelement <4 x i32> undef, i32 [[TMP12]], i32 0
-; CHECK-NEXT:    [[TMP17:%.*]] = insertelement <4 x i32> [[TMP16]], i32 [[TMP13]], i32 1
-; CHECK-NEXT:    [[TMP18:%.*]] = insertelement <4 x i32> [[TMP17]], i32 [[TMP14]], i32 2
-; CHECK-NEXT:    [[TMP19:%.*]] = insertelement <4 x i32> [[TMP18]], i32 [[TMP15]], i32 3
-; CHECK-NEXT:    [[TMP20:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP21:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP22:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP23:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP24:%.*]] = insertelement <4 x i32> undef, i32 [[TMP20]], i32 0
-; CHECK-NEXT:    [[TMP25:%.*]] = insertelement <4 x i32> [[TMP24]], i32 [[TMP21]], i32 1
-; CHECK-NEXT:    [[TMP26:%.*]] = insertelement <4 x i32> [[TMP25]], i32 [[TMP22]], i32 2
-; CHECK-NEXT:    [[TMP27:%.*]] = insertelement <4 x i32> [[TMP26]], i32 [[TMP23]], i32 3
-; CHECK-NEXT:    [[TMP28:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP29:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP30:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP31:%.*]] = load i32, i32* [[ADDR]], align 4
-; CHECK-NEXT:    [[TMP32:%.*]] = insertelement <4 x i32> undef, i32 [[TMP28]], i32 0
-; CHECK-NEXT:    [[TMP33:%.*]] = insertelement <4 x i32> [[TMP32]], i32 [[TMP29]], i32 1
-; CHECK-NEXT:    [[TMP34:%.*]] = insertelement <4 x i32> [[TMP33]], i32 [[TMP30]], i32 2
-; CHECK-NEXT:    [[TMP35:%.*]] = insertelement <4 x i32> [[TMP34]], i32 [[TMP31]], i32 3
-; CHECK-NEXT:    [[TMP36]] = add <4 x i32> [[VEC_PHI]], [[TMP11]]
-; CHECK-NEXT:    [[TMP37]] = add <4 x i32> [[VEC_PHI1]], [[TMP19]]
-; CHECK-NEXT:    [[TMP38]] = add <4 x i32> [[VEC_PHI2]], [[TMP27]]
-; CHECK-NEXT:    [[TMP39]] = add <4 x i32> [[VEC_PHI3]], [[TMP35]]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT8:%.*]] = insertelement <4 x i32> undef, i32 [[TMP7]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT9:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT8]], <4 x i32> undef, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP8]] = add <4 x i32> [[VEC_PHI]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP9]] = add <4 x i32> [[VEC_PHI1]], [[BROADCAST_SPLAT5]]
+; CHECK-NEXT:    [[TMP10]] = add <4 x i32> [[VEC_PHI2]], [[BROADCAST_SPLAT7]]
+; CHECK-NEXT:    [[TMP11]] = add <4 x i32> [[VEC_PHI3]], [[BROADCAST_SPLAT9]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 16
-; CHECK-NEXT:    [[TMP40:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4096
-; CHECK-NEXT:    br i1 [[TMP40]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP4:!llvm.loop !.*]]
+; CHECK-NEXT:    [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4096
+; CHECK-NEXT:    br i1 [[TMP12]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP4:!llvm.loop !.*]]
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[BIN_RDX:%.*]] = add <4 x i32> [[TMP37]], [[TMP36]]
-; CHECK-NEXT:    [[BIN_RDX4:%.*]] = add <4 x i32> [[TMP38]], [[BIN_RDX]]
-; CHECK-NEXT:    [[BIN_RDX5:%.*]] = add <4 x i32> [[TMP39]], [[BIN_RDX4]]
-; CHECK-NEXT:    [[TMP41:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX5]])
+; CHECK-NEXT:    [[BIN_RDX:%.*]] = add <4 x i32> [[TMP9]], [[TMP8]]
+; CHECK-NEXT:    [[BIN_RDX10:%.*]] = add <4 x i32> [[TMP10]], [[BIN_RDX]]
+; CHECK-NEXT:    [[BIN_RDX11:%.*]] = add <4 x i32> [[TMP11]], [[BIN_RDX10]]
+; CHECK-NEXT:    [[TMP13:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX11]])
 ; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 4097, 4096
 ; CHECK-NEXT:    br i1 [[CMP_N]], label [[LOOPEXIT:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 4096, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
-; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[TMP41]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[TMP13]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
 ; CHECK-NEXT:    [[ACCUM:%.*]] = phi i32 [ [[ACCUM_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
 ; CHECK-NEXT:    [[LOAD:%.*]] = load i32, i32* [[ADDR]], align 4
 ; CHECK-NEXT:    [[ACCUM_NEXT]] = add i32 [[ACCUM]], [[LOAD]]
 ; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
 ; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[IV]], 4096
 ; CHECK-NEXT:    br i1 [[EXITCOND]], label [[LOOPEXIT]], label [[FOR_BODY]], [[LOOP5:!llvm.loop !.*]]
 ; CHECK:       loopexit:
-; CHECK-NEXT:    [[ACCUM_NEXT_LCSSA:%.*]] = phi i32 [ [[ACCUM_NEXT]], [[FOR_BODY]] ], [ [[TMP41]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[ACCUM_NEXT_LCSSA:%.*]] = phi i32 [ [[ACCUM_NEXT]], [[FOR_BODY]] ], [ [[TMP13]], [[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i32 [[ACCUM_NEXT_LCSSA]]
 ;
 entry:
   br label %for.body
 
 for.body:
   %iv = phi i64 [ %iv.next, %for.body ], [ 0, %entry ]
   %accum = phi i32 [%accum.next, %for.body], [0, %entry]

 ; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
 ; CHECK-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = add i64 [[INDEX]], 8
 ; CHECK-NEXT:    [[TMP3:%.*]] = add i64 [[INDEX]], 12
 ; CHECK-NEXT:    [[TMP4:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
 ; CHECK-NEXT:    [[TMP5:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
 ; CHECK-NEXT:    [[TMP6:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
 ; CHECK-NEXT:    [[TMP7:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP8:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP9:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP10:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP11:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP12:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP13:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP14:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP15:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP16:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP17:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP18:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
-; CHECK-NEXT:    [[TMP19:%.*]] = load i32, i32* [[A]], align 4, !alias.scope !10
 ; CHECK-NEXT:    store i32 [[TMP4]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP4]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP4]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP4]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP5]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP5]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP5]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
 ; CHECK-NEXT:    store i32 [[TMP5]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
 ; CHECK-NEXT:    store i32 [[TMP6]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP6]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP6]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP6]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP7]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP7]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
+; CHECK-NEXT:    store i32 [[TMP7]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
 ; CHECK-NEXT:    store i32 [[TMP7]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP8]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP9]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP10]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP11]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP12]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP13]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP14]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP15]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP16]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP17]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP18]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
-; CHECK-NEXT:    store i32 [[TMP19]], i32* [[B]], align 4, !alias.scope !13, !noalias !10
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 16
-; CHECK-NEXT:    [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4096
-; CHECK-NEXT:    br i1 [[TMP20]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP15:!llvm.loop !.*]]
+; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4096
+; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP15:!llvm.loop !.*]]
 ; CHECK:       middle.block:
 ; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 4097, 4096
 ; CHECK-NEXT:    br i1 [[CMP_N]], label [[LOOPEXIT:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 4096, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ], [ 0, [[VECTOR_MEMCHECK]] ]
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]

llvm/test/Transforms/LoopVectorize/multiple-strides-vectorization.ll

 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP2:%.*]] = add i64 [[INDEX]], 0
 ; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 0, i64 [[TMP2]]
 ; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i32, i32* [[TMP3]], i32 0
 ; CHECK-NEXT:    [[TMP5:%.*]] = bitcast i32* [[TMP4]] to <4 x i32>*
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP5]], align 4, !alias.scope !0
 ; CHECK-NEXT:    [[TMP6:%.*]] = load i32, i32* [[TMP1]], align 4, !alias.scope !3
-; CHECK-NEXT:    [[TMP7:%.*]] = load i32, i32* [[TMP1]], align 4, !alias.scope !3
-; CHECK-NEXT:    [[TMP8:%.*]] = load i32, i32* [[TMP1]], align 4, !alias.scope !3
-; CHECK-NEXT:    [[TMP9:%.*]] = load i32, i32* [[TMP1]], align 4, !alias.scope !3
-; CHECK-NEXT:    [[TMP10:%.*]] = insertelement <4 x i32> undef, i32 [[TMP6]], i32 0
-; CHECK-NEXT:    [[TMP11:%.*]] = insertelement <4 x i32> [[TMP10]], i32 [[TMP7]], i32 1
-; CHECK-NEXT:    [[TMP12:%.*]] = insertelement <4 x i32> [[TMP11]], i32 [[TMP8]], i32 2
-; CHECK-NEXT:    [[TMP13:%.*]] = insertelement <4 x i32> [[TMP12]], i32 [[TMP9]], i32 3
-; CHECK-NEXT:    [[TMP14:%.*]] = add nsw <4 x i32> [[TMP13]], [[WIDE_LOAD]]
-; CHECK-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[TMP2]]
-; CHECK-NEXT:    [[TMP16:%.*]] = getelementptr inbounds i32, i32* [[TMP15]], i32 0
-; CHECK-NEXT:    [[TMP17:%.*]] = bitcast i32* [[TMP16]] to <4 x i32>*
-; CHECK-NEXT:    [[WIDE_LOAD12:%.*]] = load <4 x i32>, <4 x i32>* [[TMP17]], align 4, !alias.scope !5, !noalias !7
-; CHECK-NEXT:    [[TMP18:%.*]] = add nsw <4 x i32> [[TMP14]], [[WIDE_LOAD12]]
-; CHECK-NEXT:    [[TMP19:%.*]] = bitcast i32* [[TMP16]] to <4 x i32>*
-; CHECK-NEXT:    store <4 x i32> [[TMP18]], <4 x i32>* [[TMP19]], align 4, !alias.scope !5, !noalias !7
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> undef, i32 [[TMP6]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> undef, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP7:%.*]] = add nsw <4 x i32> [[BROADCAST_SPLAT]], [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[TMP2]]
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr inbounds i32, i32* [[TMP8]], i32 0
+; CHECK-NEXT:    [[TMP10:%.*]] = bitcast i32* [[TMP9]] to <4 x i32>*
+; CHECK-NEXT:    [[WIDE_LOAD12:%.*]] = load <4 x i32>, <4 x i32>* [[TMP10]], align 4, !alias.scope !5, !noalias !7
+; CHECK-NEXT:    [[TMP11:%.*]] = add nsw <4 x i32> [[TMP7]], [[WIDE_LOAD12]]
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast i32* [[TMP9]] to <4 x i32>*
+; CHECK-NEXT:    store <4 x i32> [[TMP11]], <4 x i32>* [[TMP12]], align 4, !alias.scope !5, !noalias !7
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
-; CHECK-NEXT:    br i1 [[TMP20]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !8
+; CHECK-NEXT:    [[TMP13:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP13]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP8:!llvm.loop !.*]]
 ; CHECK:       middle.block:
 ; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[Z]], [[N_VEC]]
 ; CHECK-NEXT:    br i1 [[CMP_N]], label [[DOTOUTER]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[DOTOUTER_PREHEADER]] ], [ 0, [[VECTOR_MEMCHECK]] ]
 ; CHECK-NEXT:    br label [[DOTINNER:%.*]]
 ; CHECK:       .exit:
 ; CHECK-NEXT:    ret void
 ; CHECK:       .outer:
 ; CHECK-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 1
 ; CHECK-NEXT:    [[EXITCOND_OUTER:%.*]] = icmp eq i64 [[I_NEXT]], 32
 ; CHECK-NEXT:    br i1 [[EXITCOND_OUTER]], label [[DOTEXIT:%.*]], label [[DOTOUTER_PREHEADER]]
 ; CHECK:       .inner:
 ; CHECK-NEXT:    [[J:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[J_NEXT:%.*]], [[DOTINNER]] ]
-; CHECK-NEXT:    [[TMP21:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 0, i64 [[J]]
-; CHECK-NEXT:    [[TMP22:%.*]] = load i32, i32* [[TMP21]], align 4
-; CHECK-NEXT:    [[TMP23:%.*]] = load i32, i32* [[TMP1]], align 4
-; CHECK-NEXT:    [[TMP24:%.*]] = add nsw i32 [[TMP23]], [[TMP22]]
-; CHECK-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[J]]
-; CHECK-NEXT:    [[TMP26:%.*]] = load i32, i32* [[TMP25]], align 4
-; CHECK-NEXT:    [[TMP27:%.*]] = add nsw i32 [[TMP24]], [[TMP26]]
-; CHECK-NEXT:    store i32 [[TMP27]], i32* [[TMP25]]
+; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 0, i64 [[J]]
+; CHECK-NEXT:    [[TMP15:%.*]] = load i32, i32* [[TMP14]], align 4
+; CHECK-NEXT:    [[TMP16:%.*]] = load i32, i32* [[TMP1]], align 4
+; CHECK-NEXT:    [[TMP17:%.*]] = add nsw i32 [[TMP16]], [[TMP15]]
+; CHECK-NEXT:    [[TMP18:%.*]] = getelementptr inbounds [[STRUCT_S]], %struct.s* [[OBJ]], i64 0, i32 2, i64 [[I]], i64 [[J]]
+; CHECK-NEXT:    [[TMP19:%.*]] = load i32, i32* [[TMP18]], align 4
+; CHECK-NEXT:    [[TMP20:%.*]] = add nsw i32 [[TMP17]], [[TMP19]]
+; CHECK-NEXT:    store i32 [[TMP20]], i32* [[TMP18]], align 4
 ; CHECK-NEXT:    [[J_NEXT]] = add nuw nsw i64 [[J]], 1
 ; CHECK-NEXT:    [[EXITCOND_INNER:%.*]] = icmp eq i64 [[J_NEXT]], [[Z]]
-; CHECK-NEXT:    br i1 [[EXITCOND_INNER]], label [[DOTOUTER]], label [[DOTINNER]], !llvm.loop !10
+; CHECK-NEXT:    br i1 [[EXITCOND_INNER]], label [[DOTOUTER]], label [[DOTINNER]], [[LOOP10:!llvm.loop !.*]]
 ;
   br label %.outer.preheader
 
 
 .outer.preheader:
   %i = phi i64 [ 0, %0 ], [ %i.next, %.outer ]
   %1 = getelementptr inbounds %struct.s, %struct.s* %obj, i64 0, i32 1, i64 %i
   br label %.inner