[Loop Vectorizer] Interleave vs Gather - in some cases Gather is better.

[delena]

Summary

The bug is described in PR31426.
The cost of Load instruction is calculated in the following order isConsecultive - isInterleave - isGather - scalar.
When a Load instruction belongs to Interleave group, the "Gather" option is not checked at all. But when the interleave factor exceeds the maximum, the cost is high and the "Gather" is preferred in this case. The following loop is not vectorized on AVX-512 due to this bug:
for (i=0; i<N; ++i)

B[i] = A[i*5]

[delena]

Added some comments.

[mssimpso]

../../ver4/lib/Transforms/Vectorize/LoopVectorize.cpp

7047–7049 ↗

(On Diff #81949)

Why don't you just compare the costs? You wouldn't need to make this assumption anymore.

[mkuper]

../../ver4/lib/Transforms/Vectorize/LoopVectorize.cpp

7052 ↗

(On Diff #81949)

Aren't we already checking this in selectInterleaveCount()?
How do we end up with interleave factors above MaxInterleaveFactor in the first place?

[mssimpso]

../../ver4/lib/Transforms/Vectorize/LoopVectorize.cpp

7052 ↗

(On Diff #81949)

Ah, our naming conventions have confused this somewhat, I think. TTI.getMaxInterleaveFacor is the hook for the max unroll factor ("interleaving"). I think Elana was wanting TLI.getMaxSupportedInterleaveFactor instead. This is the hook for determining the max factor of the interleaved access groups.

[mkuper]

../../ver4/lib/Transforms/Vectorize/LoopVectorize.cpp

7052 ↗

(On Diff #81949)

Argh, right.
Sorry for the stupid question, I just looked at this and went "this looks odd" without actually reading the context. But yes, Elena, you want the other function.

Regardless, can we maybe change the naming to something sane? As a separate patch, of course. (I don't have any good ideas, though.)

[delena]

../../ver4/lib/Transforms/Vectorize/LoopVectorize.cpp

7047–7049 ↗

(On Diff #81949)

The cost that we provide for interleaved access is incorrect, specially for AVX-512. AVX-512 has 3-src shuffles and the real cost is much lower . I can't compare it to Gather - the Gather cost wins today, even for small stride, but it is not true. So there are 2 bugs:

(1) The loop is scalarized and Gather/Scatter option is not considered at all
(2) Incorrect cost for interleaving

I can start from providing a correct cost for interleaving on AVX-512.
Or I can fix the (1) first of all. I'll retrieve the proper "MaxSupportedInterleaveFactor".

What do you think?

[mkuper]

../../ver4/lib/Transforms/Vectorize/LoopVectorize.cpp

7047–7049 ↗

(On Diff #81949)

I think it would be better to fix the cost model first.

It's very pessimistic for x86 in general, not AVX-512, but you're right, it's even worse for AVX-512, because the real cost is lower. But I thought Farhana was already working on that. Am I confused?

[delena]

Now, when the interleaving cost calculation is correct, I compare the cost of 3 possible options - interleave, gather and scalar and choose the better option.
The decision made by cost model should be saved and latter used when we generate the vector code.

The code is refactored in order to allow this comparison, but the comparison is the *only* functional change in the code. The rest of the logic remains the same.

I added a test case that demonstrates gather - vs - interleave case.

[delena]

Merged the patch with the latest changes in LV.

[delena]

Ping*

[mssimpso]

Hi Elena,

This patch causes a crash in spec2006/povray on AArch64. I've pasted a test case over at P7951. The problem has to do with the analysis in collectLoopUniforms and the new decision to scalarize. collectLoopUniforms is very conservative about what instructions remain uniform after vectorization. If a memory access has the possibility of being scalarized (even though it may not be), it's pointer operand is not marked uniform. What's happening here is that you've introduced a new scalarization decision based on the cost model that collectLoopUniforms doesn't know about (and likely can't know about). In the test case, collectLoopUniforms marks the pointer operands of the loads uniform even though the loads are now scalarized, which causes the pointer operands to be non-uniform.

I'm not really sure what the best way to fix this would be. Any thoughts? One way would be to remove scalarization from your widening decision.

[mkuper]

In D27919#644255, @mssimpso wrote:

Hi Elena,

I'm not really sure what the best way to fix this would be. Any thoughts? One way would be to remove scalarization from your widening decision.

That sounds a bit unfortunate. I think we'd rather like to the cost model to be able to say "please scalarize this".

../lib/Transforms/Vectorize/LoopVectorize.cpp

1696

I'm not a fan of the name of this enum, but don't really have any good ideas. Anyone else?

1719

This looks like a suboptimal way to iterate over InterleaveGroup, because it requires a lookup for every index, instead of just iterating above the Members collection. (we don't care about the order here, right?)
Perhaps add a better interface?

1725

Is this really a per-instruction thing, or do we just always end up with LV_NONE when the cost model is not in use (e.g. explicit user-provided VF).

If the former, when does this happen?
If the latter, then I think the comment is a bit misleading - and it would be good to be able to assert on this happening only when there's no cost model.

Edit: Oh, I think I see, you care about temporary LV_NONE values for Interleave groups. I still think it'd be nice if we could somehow distinguish between the cases, so we could assert on not having a cost when we should.

2911

Extra space after =.

2915

assert(!Legal->memoryInstructionMustBeScalarized(Instr, VF)), maybe?

Because now we leave the door open for the cost model decision being to widen, even though memoryInstructionMustBeScalarized().
Another option is to replace the if with:

if (Decision == LoopVectorizationLegality::LV_SCALARIZE ||     
    Legal->memoryInstructionMustBeScalarized(Instr, VF))

But I'm not sure it makes sense, because there's no good reason to always call memoryInstructionMustBeScalarized() in a non-asserts build - in theory, the cost model should have already checked this.

7352

Wouldn't you calculate it several times per group if it's unprofitable?

../test/Analysis/CostModel/X86/interleave-load-i32.ll

13 ↗

(On Diff #83607)

What happened to the VF = 1 cost?

[delena]

In D27919#644255, @mssimpso wrote:

Hi Elena,

This patch causes a crash in spec2006/povray on AArch64. I've pasted a test case over at P7951. The problem has to do with the analysis in collectLoopUniforms and the new decision to scalarize. collectLoopUniforms is very conservative about what instructions remain uniform after vectorization. If a memory access has the possibility of being scalarized (even though it may not be), it's pointer operand is not marked uniform. What's happening here is that you've introduced a new scalarization decision based on the cost model that collectLoopUniforms doesn't know about (and likely can't know about). In the test case, collectLoopUniforms marks the pointer operands of the loads uniform even though the loads are now scalarized, which causes the pointer operands to be non-uniform.

I'm not really sure what the best way to fix this would be. Any thoughts? One way would be to remove scalarization from your widening decision.

I'm trying to revisit collections of Uniforms and Scalars after cost estimation and remove GEPs and induction variables if we decided to scalarize. Not finished yet..

../lib/Transforms/Vectorize/LoopVectorize.cpp

1696

CM_DECISION_NONE,
CM_DECISION_WIDEN .. ?

1719

I agree, but iteration inside Factor is not a big overhead and it used in one more place, at least. This patch is complex enough, I'd postpone unrelated changes to another patch.

2915

The form I used just prevents a redundant call to memoryInstructionMustBeScalarized(). Because we may have a decision (INTERLEAVE, WIDEN or GATHER_SCATTER) at this stage.

7352

LV_NONE means "not calculated yet". If we are here, each memory instruction will have " a decision".

../test/Analysis/CostModel/X86/interleave-load-i32.ll

13 ↗

(On Diff #83607)

There is no group for VF 1. Instruction cost is still there.

[mssimpso]

In D27919#647975, @delena wrote:

I'm trying to revisit collections of Uniforms and Scalars after cost estimation and remove GEPs and induction variables if we decided to scalarize. Not finished yet..

That sounds like a good idea to me, and I hope we can separate the uniform/scalar collection from the cost estimation in a way that makes sense. When I looked at doing that a while back, the complication I ran into was that the cost estimates depend on knowing the uniforms. But if we use the cost estimates to help determine the uniforms, then we end up with a weird circular dependence. If we mark something uniform/scalar after computing the costs, we will need to update the costs somehow. That in turn might change what we would like to scalarize, etc.

[delena]

I'm revisiting the set of Scalars and Uniforms after cost modeling. Taking into account that the cost model uses Uniforms and the Uniforms are changed after the cost modeling, I still think that there is no circular dependency here.
(1) I remove uniform GEP (and the corresponding induction) when we decide to scalarize memory instruction.
(2) I do not remove GEP from scalars if it will not be used in Gather/Scatter.

I added a test that Matthew sent me.

[mssimpso]

../lib/Transforms/Vectorize/LoopVectorize.cpp

7007–7008

Hi Elena,

I had been thinking about the use of isUniformAfterVectorization() here in getInstructionCost(). Wouldn't it now be possible for the set of uniforms to differ from the first collection (before VF selection) and the second collection (after VF selection)? So we would choose a VF based on costs assuming an instruction may or may not be uniform. Then we could later reverse our initial decision about the instruction's uniformity after VF selection, making the total cost on which we based our VF decision inaccurate. Or am I missing something? I haven't yet thought through the implications of this in enough detail to know whether this would matter much or not.

[delena]

../lib/Transforms/Vectorize/LoopVectorize.cpp

7007–7008

About the list of Uniforms. We insert and then remove only GEPs and Induction variables. We do not calculate cost for them anyway. All other Uniform values stay in place. So, the cost is accurate at the end. There is no circular dependency here.

[mssimpso]

../lib/Transforms/Vectorize/LoopVectorize.cpp

7007–7008

I don't think this is true in general. We mark an instruction uniform if all its users are uniform. So for example, if we have a uniform GEP whose index is some computation, that computation is also uniform if it's only used by the GEP. I think we have some examples in induction.ll, but something like this:

%i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
%sum = add i64 %i, %x
%idx = getelementptr inbounds float, float* %a, i64 %sum
load float, float* %idx, align 4

The GEP is consecutive, so it will be marked uniform. %sum will aslo be marked uniform because it's only used by the GEP. If we later decide to scalarize the load, the GEP, the IV, and %sum will all no longer be uniform. So the cost for %sum will have been wrong.

[mssimpso]

../lib/Transforms/Vectorize/LoopVectorize.cpp

7007–7008

Just a thought - why not recompute and cache the uniforms (and possibly scalars) for each VF we compute costs for? That would avoid any potential logical inconsistencies. I think the compile-time overhead would probably be minimal (and you're already computing these sets twice anyway).

[delena]

../lib/Transforms/Vectorize/LoopVectorize.cpp

7007–7008

Just talked with Ayal about this.

I can collect Uniforms after making decision about Load/Store intructions. And the decision is based on cost. The decision affects another instructions inside the loop, as you've pointed before. Theoretically, if I have N variants of representing all memory instructions inside the loop, I should examine 2**N combinations per VF.

Ayal proposed the following sequence, which should be done on CM stage, after legality is finished:
Per VF:

1. Go through all memory insts and make CM decision
2. Build Uniforms and Scalars per VF (that's what you say now)
3. Calculate cost for VF, based on Uniforms and Scalars

It is still not ideal, but, probably better than what we have.

[mssimpso]

../lib/Transforms/Vectorize/LoopVectorize.cpp

7007–7008

Ayal's sequence makes sense to me. We should probably also try to move Uniforms/Scalars and related functions over to LoopVectorizationCostModel rather than keep them in LoopVectorizationLegality, as they will now be a function of Cost/VF and would be more appropriate there in my opinion. This could probably be a separate patch.

[delena]

I moved Uniforms and Scalars from Legality to the Cost Model. Now we collect Uniforms and Scalars per VF and these collections depend on widening decisions that CM takes for Load/Store instructions.

The patch is big, but I did not see how to split it into separate patches.

[delena]

Ping *
We are in sync with Ayal and Gil working on VPlan.

[mssimpso]

Hi Elena,

I'll take a look at this again today. Thanks for the reminder!

[mssimpso]

Hi Elena,

Thanks for your patience. I haven't yet looked in detail at the widening decision selection, but here are some comments around the uniforms/scalars.

Matt.

../lib/Transforms/Vectorize/LoopVectorize.cpp

5616–5619

This is not true now, right? An interleaved access may be scalarized based on the cost model.

5625–5626

Can we change this to something like:

if (Uniforms.count(VF))
  return;

auto &UniformsVF = Uniforms[VF];

We want to distinguish the case that (1) Uniforms have not been computed for VF from (2) Uniforms have been computed for VF but there aren't any, so we don't need to compute them again. We can end up calling this twice for the same VF if we have a user-selected VF and then compute the expected cost for interleaving. This is similar to the way we do this check in collectInstsToScalarize.

This will also apply to collectLoopScalars.

5687

Why not roll this check into memoryInstructionMustBeScalarized? Either way, I think this check and the check below in isVectorizedMemAccessUse that calls memoryInstructionMustBeScalarized, should be the same.

7007

The VF > 1 check is not needed because you check that condition in isUniformAfterVectorization.

7022

This should be VF == 1, since we can't have a zero VF.

7466–7479

Can we just delete this in favor of a helper function that checks VecValuesToIgnore and IsScalarAfterVectorization for a given VF? Something like:

bool LoopVectorizationCostModel::shouldIgnoreVecValueInCostModel(Instruction *I, unsigned VF) {
  return VecValuesToIgnore.count(I) || isScalarAfterVectorization(I, VF);
}

This way we won't have to be imprecise.

[delena]

../lib/Transforms/Vectorize/LoopVectorize.cpp

5616–5619

It is still true. Instruction must be scalarized if there is no any other option. This is "legality" check, cost model comes later.

5625–5626

Yes. I've changed.

5687

Yes, I don't need to check legality any more. It is already implied in CM decision.

7007

Yes. You are right, thanks!

7466–7479

Yes, it is possible.

[delena]

Updated, following Matthew's comments.

[mssimpso]

Hi Elena,

Here are some more inline comments. Also, please clang-format the patch if you haven't already done so to make the review easier. Thanks!

../lib/Transforms/Vectorize/LoopVectorize.cpp

1932

How are we using CM_DECISION_NONE? Aren't we forced to make some sort of decision? I would think the default would be CM_DECISION_WIDEN unless the cost model points to one of the others.

2056–2057

The VF > 2 comment can be removed now.

2064–2065

The VF > 2 comment can be removed now.

5502–5503

Since you've added a new check in collectUnifomsAndScalars to ensure the analysis is performed only once, the check here and in collectLoopUniforms is redundant. Should these be asserts now?

5556–5563

I don't think a see a real use for this function anymore. Please see my related comment about memoryAccessMustBeScalarized. This function was only ever used in collectLoopUniforms to help determine how a memory access would be vectorized. I think you can probably greatly simplify the logic in collectLoopUniforms and remove it. Now, we know what the vectorizer will do based the the cost model decision.

For a GEP to remain uniform, I think we just need to know that all its users are CM_DECISION_INTERLEAVE or CM_DECISION_WIDEN. Is this right? If so, please work it into the GEP part of collectLoopUniforms.

5616–5619

I think the cost vs legality distinction is not important here. My original intent with this function was just to consolidate all the scalarization conditions for a given access. That way it could be called when collecting uniforms, computing costs, and vectorizing and they all would agree on what would happen. Perhaps the choice of name was misleading?

In any case, unless I'm missing something, I don't think you actually use this function anymore. You've replaced all uses with

getWideningDecision(I, VF) == CM_DECISION_SCALARIZE

This makes sense because I think you've moved all the non-cost related scalarization decisions into the inverse: memoryInstructionCanBeWidened. Can this function be deleted now?

[delena]

In D27919#662246, @mssimpso wrote:

Hi Elena,

Here are some more inline comments. Also, please clang-format the patch if you haven't already done so to make the review easier. Thanks!

Done.

../lib/Transforms/Vectorize/LoopVectorize.cpp

1932

I used it while capturing decisions for an interleave group. At this point I can get rid of it. I return "NONE" if no decision, and it is convenient.

InstWidening getWideningDecision(Instruction *I, unsigned VF) {
    assert(VF >= 2 && "Expected VF >=2");
    std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
    if (!WideningDecisions.count(InstOnVF))
      return CM_DECISION_NONE;
    return WideningDecisions[InstOnVF];
  }

Then I use it in assertion, to make sure that decision is made.

5502–5503

May be. Until somebody will decide to call these functions separately. Right now I don't see a reason. I'll put an "assert" and add a comment.

5556–5563

I removed mustBeScalarized(). (I suppose you meant this function, the lines are mixed up)

[mssimpso]

../lib/Transforms/Vectorize/LoopVectorize.cpp

5556–5563

In this comment I was talking about hasConsecutiveLikePtrOperand. I don't think it's needed anymore and can probably be deleted. It's only used in collectLoopUniforms to help determine if a GEP will remain uniform. But can't we now determine that much easier by just checking the CM_DECISION of the accesses? Does that make sense?

[delena]

I removed hasConsecutiveLikePtrOperand and simplified the code.

[mssimpso]

Hi Elena,

I like the direction this patch is going. Thanks for all your work. Here are some more comments inline.

../lib/Transforms/Vectorize/LoopVectorize.cpp

7068

Can we avoid the divisions here and below and use multiplication instead? We won't have any round-off issues that way. What about something like:

unsigned Accesses = 1;
if (Legal->isAccessInterleaved(&I)) {
  Accesses = Group->getNumMembers();
  ...
}
...
ScalarizationCost *= Accesses;
...

7087–7095

I think we're fairly consistent in other places where we compare costs, that we prefer the scalar version if there's no benefit for vectorization. So if the scalarization cost is <= the other costs, I think we should go with that.

7276–7305

Can't this all be replaced by a call to getInterleaveGroupCost()?

7312–7335

Can't this all be replaced by a call to getMemInstScalarizationCost()?

7344–7348

Similar comment to the above. I don't think you have a helper for computing gather/scatter cost, but I think it would be nice. It would be easier to keep getInstructionCost in sync with setCostBasedWideningDecision.

[delena]

../lib/Transforms/Vectorize/LoopVectorize.cpp

7087–7095

Agree. I've changed the comparison.

7276–7305

I reached the conclusion that we don't need to calculate the cost again. We can keep it together with widening decision.

[delena]

The memory instruction cost that was calculated during widening decision is saved in WideningDecisions map in order to avoid recalculation.

[mssimpso]

Hi Elena,

I have one comment about getWideningCost; otherwise, the patch looks fine to me now. But please let Michael have one more pass at review since he's been quiet for a while. Thanks!

../lib/Transforms/Vectorize/LoopVectorize.cpp

1981–1982

This seems weird to me. All instructions are supposed to have a cost computed by the cost model. I would much rather us assert that I is in WideningDecisions. Isn't it true that if I is a load or store, we would have already computed a cost and saved it in WideningDecisions?

7058

Just assert inside getWideningCost() that "I" is present in the mapping and return the cost. The int to unsigned max conversion is unnecessary.

[delena]

Updated according to the resent Matthew's comments.
Matthew, thanks a lot for your review.

Michael, could you, please, take a look?

[mkuper]

The high-level structure of this looks good to me, thanks Elena!

Some minor/style comments inline.

../lib/Transforms/Vectorize/LoopVectorize.cpp

330

We probably already have several clones of those functions around the code-base. And they are probably all slightly different.
LSR has getAccessType(), LAA has getAddressSpaceOperand(), LoadStoreVectorizer has getPointerAddressSpace(), and I'm sure there are more.
I don't want to make merging them a precondition of this patch, but can you please at least add a FIXME here?

1909

Why not just "return Uniforms[VF]->second.count(I);"? I don't think the verbosity helps here, and we don't actually care about the difference between find() and operator[] due to the assert.
Or, to save a lookup in an asserts build, you could find() and then assert on the result of the find().

1917

Same as above.

1932

Maybe rename NONE to UNKNOWN, then? But I'm fine with None if you think that's clearer.
Also, I looked up the naming convention for enums in the coding standard, and I think it should be something like "CM_Widen, CM_Interleave", etc, not all caps.

1970

Here, on the other hand, I think find() would be better - that way you don't need two lookups.

2072

I'm still not sure I understand why this gets called twice for a user-provided VF. Could you explain again?

5592

Do we still want this check here?
I mean:

1. An instruction with a uniform pointer *can* be widended.
2. That ties in with the way this is used - we check for the uniform case before the widening case, as we should. So, IIUC, we should never actually hit this.

7046

UINT_MAX

7060

Why the NumAccesses * 2 cut-off?

7071

UINT_MAX

7084

Why do you need the "GatherScatterCost < InterleaveCost" check here?

[delena]

../lib/Transforms/Vectorize/LoopVectorize.cpp

1909

The "const" qualifier does not allow the Uniforms[VF] form.

2072

We call this function from multiple places. From calculateRegisterUsage(), selectVectorizationFactor() - user-defined-VF and from expectedCost().
I just prevent recalculation.

7060

I consider a cost per instruction. In this case InterleaveCost / NumAccesses == 1. (Matthew asked to avoid divisions).
About 1 inst per access is good enough. I added more comments.

[delena]

Some code improvements, addressed Michael's comments.

[mkuper]

../lib/Transforms/Vectorize/LoopVectorize.cpp

1909

Ohh, right, missed it's const, sorry.

2072

I understand, I'm just wondering whether we really need to do that.
Anyway, it's not a new problem, we don't have to solve it here.

5592

I think you may have missed this comment.

7060

Well, this isn't really "about 1", this is "below 2".
I'd be more conservative here (InerelaveCost <= NumAccesses ? Can this even happen? Or are you trying to catch the cases where the ratio is ~1.1-1.2?). Or maybe remove this altogether. Is getGatherScatterCost() expensive in terms of compile time?

7085

So, in case the costs are equal, you prefer scalarization to interleaving, and interleaving to scatter/gather?
(In theory it shouldn't matter what happens when the costs are equal, just making sure this I understand.)

[mssimpso]

../lib/Transforms/Vectorize/LoopVectorize.cpp

7060

The TTI estimates are supposed to be cheap to compute. I think it makes sense to remove this altogether in favor of greater simplicity.

[delena]

../lib/Transforms/Vectorize/LoopVectorize.cpp

2072

In the current version, before my changes, we calculate Uniforms and Scalars once and do this in Legality. In this patch, I moved Uniforms and Scalars from Legality to the Cost Model and calculate them per VF. I did not find a single place to put the call, I'm calling the collectUniformsAndScalars() from multiple places. Doing that, I want to prevent the data recalculation.

I suppose that finding a right single place for calling collectUniformsAndScalars() is possible, but it will require additional movements in selectVectorizationFactor(). I think it can be done in a separate patch.

5592

I'll fix.

7060

ok.

7085

Matthew asked to change:

"I think we're fairly consistent in other places where we compare costs, that we prefer the scalar version if there's no benefit for vectorization. So if the scalarization cost is <= the other costs, I think we should go with that."

[delena]

Some fixes following Michael's recent comments.

[mkuper]

LGTM

../lib/Transforms/Vectorize/LoopVectorize.cpp

2072

Ah, ok, got it. Sure, sounds good as a follow-up.

7085

Ah, ok, missed that. Could you please add this as an explicit comment?

../lib/Transforms/Vectorize/LoopVectorize.cpp

 static Value *getPointerOperand(Value *I) {
   if (auto *LI = dyn_cast<LoadInst>(I))
     return LI->getPointerOperand();
   if (auto *SI = dyn_cast<StoreInst>(I))
     return SI->getPointerOperand();
   return nullptr;
 }
 
 /// A helper function that returns true if the given type is irregular. The
 /// type is irregular if its allocated size doesn't equal the store size of an
 /// element of the corresponding vector type at the given vectorization factor.
 static bool hasIrregularType(Type *Ty, const DataLayout &DL, unsigned VF) {
 
   // Determine if an array of VF elements of type Ty is "bitcast compatible"
   // with a <VF x Ty> vector.
   if (VF > 1) {
     auto *VectorTy = VectorType::get(Ty, VF);
public:
   /// 0 - Stride is unknown or non-consecutive.
   /// 1 - Address is consecutive.
   /// -1 - Address is consecutive, and decreasing.
   int isConsecutivePtr(Value *Ptr);
 
   /// Returns true if the value V is uniform within the loop.
   bool isUniform(Value *V);
 
-  /// Returns true if \p I is known to be uniform after vectorization.
-  bool isUniformAfterVectorization(Instruction *I) { return Uniforms.count(I); }
-
-  /// Returns true if \p I is known to be scalar after vectorization.
-  bool isScalarAfterVectorization(Instruction *I) { return Scalars.count(I); }
-
   /// Returns the information that we collected about runtime memory check.
   const RuntimePointerChecking *getRuntimePointerChecking() const {
     return LAI->getRuntimePointerChecking();
   }
 
   const LoopAccessInfo *getLAI() const { return LAI; }
 
   /// \brief Check if \p Instr belongs to any interleaved access group.
public:
   unsigned getNumLoads() const { return LAI->getNumLoads(); }
   unsigned getNumPredStores() const { return NumPredStores; }
 
   /// Returns true if \p I is an instruction that will be scalarized with
   /// predication. Such instructions include conditional stores and
   /// instructions that may divide by zero.
   bool isScalarWithPredication(Instruction *I);
 
-  /// Returns true if \p I is a memory instruction that has a consecutive or
-  /// consecutive-like pointer operand. Consecutive-like pointers are pointers
-  /// that are treated like consecutive pointers during vectorization. The
-  /// pointer operands of interleaved accesses are an example.
-  bool hasConsecutiveLikePtrOperand(Instruction *I);
-
   /// Returns true if \p I is a memory instruction that must be scalarized
   /// during vectorization.
   bool memoryInstructionMustBeScalarized(Instruction *I, unsigned VF = 1);
 
+  /// Returns true if \p I is a memory instruction with consecutive memory
+  ///  access that can be widened.
+  bool memoryInstructionCanBeWidened(Instruction *I, unsigned VF = 1);
+
 private:
   /// Check if a single basic block loop is vectorizable.
   /// At this point we know that this is a loop with a constant trip count
   /// and we only need to check individual instructions.
   bool canVectorizeInstrs();
 
   /// When we vectorize loops we may change the order in which
   /// we read and write from memory. This method checks if it is
   /// legal to vectorize the code, considering only memory constrains.
   /// Returns true if the loop is vectorizable
   bool canVectorizeMemory();
 
   /// Return true if we can vectorize this loop using the IF-conversion
   /// transformation.
   bool canVectorizeWithIfConvert();
 
-  /// Collect the instructions that are uniform after vectorization. An
-  /// instruction is uniform if we represent it with a single scalar value in
-  /// the vectorized loop corresponding to each vector iteration. Examples of
-  /// uniform instructions include pointer operands of consecutive or
-  /// interleaved memory accesses. Note that although uniformity implies an
-  /// instruction will be scalar, the reverse is not true. In general, a
-  /// scalarized instruction will be represented by VF scalar values in the
-  /// vectorized loop, each corresponding to an iteration of the original
-  /// scalar loop.
-  void collectLoopUniforms();
-
-  /// Collect the instructions that are scalar after vectorization. An
-  /// instruction is scalar if it is known to be uniform or will be scalarized
-  /// during vectorization. Non-uniform scalarized instructions will be
-  /// represented by VF values in the vectorized loop, each corresponding to an
-  /// iteration of the original scalar loop.
-  void collectLoopScalars();
-
   /// Return true if all of the instructions in the block can be speculatively
   /// executed. \p SafePtrs is a list of addresses that are known to be legal
   /// and we know that we can read from them without segfault.
   bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
 
   /// Updates the vectorization state by adding \p Phi to the inductions list.
   /// This can set \p Phi as the main induction of the loop if \p Phi is a
   /// better choice for the main induction than the existing one.
   void addInductionPhi(PHINode *Phi, const InductionDescriptor &ID,
                        SmallPtrSetImpl<Value *> &AllowedExit);
 
   /// Report an analysis message to assist the user in diagnosing loops that are
   /// not vectorized.  These are handled as LoopAccessReport rather than
   /// VectorizationReport because the << operator of VectorizationReport returns
   /// LoopAccessReport.
   void emitAnalysis(const LoopAccessReport &Message) const {
     emitAnalysisDiag(TheLoop, *Hints, *ORE, Message);
   }
 
   /// Create an analysis remark that explains why vectorization failed
   ///
   /// \p RemarkName is the identifier for the remark.  If \p I is passed it is
   /// an instruction that prevents vectorization.  Otherwise the loop is used
private:
   RecurrenceSet FirstOrderRecurrences;
   /// Holds the widest induction type encountered.
   Type *WidestIndTy;
 
   /// Allowed outside users. This holds the induction and reduction
   /// vars which can be accessed from outside the loop.
   SmallPtrSet<Value *, 4> AllowedExit;
 
-  /// Holds the instructions known to be uniform after vectorization.
-  SmallPtrSet<Instruction *, 4> Uniforms;
-
-  /// Holds the instructions known to be scalar after vectorization.
-  SmallPtrSet<Instruction *, 4> Scalars;
-
   /// Can we assume the absence of NaNs.
   bool HasFunNoNaNAttr;
 
   /// Vectorization requirements that will go through late-evaluation.
   LoopVectorizationRequirements *Requirements;
 
   /// Used to emit an analysis of any legality issues.
   LoopVectorizeHints *Hints;
public:
 
   /// \return The desired interleave count.
   /// If interleave count has been specified by metadata it will be returned.
   /// Otherwise, the interleave count is computed and returned. VF and LoopCost
   /// are the selected vectorization factor and the cost of the selected VF.
   unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
                                  unsigned LoopCost);
 
+  /// Memory access instruction may be vectorized in more than one way.
+  /// Form of instruction after vectorization depends on cost.
+  /// This function takes cost-based decisions for Load/Store instructions
+  /// and collects them in a map. This decisions map is used for building
+  /// the lists of loop-uniform and loop-scalar instructions.
+  void setCostBasedWideningDecision(unsigned VF);
+
+  /// Returns true if \p I is a memory instruction that has a consecutive or
+  /// consecutive-like pointer operand. Consecutive-like pointers are pointers
+  /// that are treated like consecutive pointers during vectorization. The
+  /// pointer operands of interleaved accesses are an example.
+  bool hasConsecutiveLikePtrOperand(Instruction *I, unsigned VF);
+
   /// \brief A struct that represents some properties of the register usage
   /// of a loop.
   struct RegisterUsage {
     /// Holds the number of loop invariant values that are used in the loop.
     unsigned LoopInvariantRegs;
     /// Holds the maximum number of concurrent live intervals in the loop.
     unsigned MaxLocalUsers;
     /// Holds the number of instructions in the loop.
public:
   /// vectorization factor \p VF.
   bool isProfitableToScalarize(Instruction *I, unsigned VF) const {
     auto Scalars = InstsToScalarize.find(VF);
     assert(Scalars != InstsToScalarize.end() &&
            "VF not yet analyzed for scalarization profitability");
     return Scalars->second.count(I);
   }
 
+  /// Returns true if \p I is known to be uniform after vectorization.
+  bool isUniformAfterVectorization(Instruction *I, unsigned VF) const {
+    if (VF == 1)
+      return true;
+    auto UniformsPerVF = Uniforms.find(VF);
+    assert(UniformsPerVF != Uniforms.end() &&
+           "VF not yet analyzed for uniformity");
+    return UniformsPerVF->second.count(I);
+  }
+
+  /// Returns true if \p I is known to be scalar after vectorization.
+  bool isScalarAfterVectorization(Instruction *I, unsigned VF) const {
+    if (VF == 1)
+      return true;
+    auto ScalarsPerVF = Scalars.find(VF);
+    assert(ScalarsPerVF != Scalars.end() &&
+           "Scalar values are not calculated for VF");
+    return ScalarsPerVF->second.count(I);
+  }
+
   /// \returns True if instruction \p I can be truncated to a smaller bitwidth
   /// for vectorization factor \p VF.
   bool canTruncateToMinimalBitwidth(Instruction *I, unsigned VF) const {
     return VF > 1 && MinBWs.count(I) && !isProfitableToScalarize(I, VF) &&
-           !Legal->isScalarAfterVectorization(I);
+           !isScalarAfterVectorization(I, VF);
+  }
+
+  /// Decision that was taken during cost calculation for memory instruction.
+  enum InstWidening {
+    CM_DECISION_NONE,
+    CM_DECISION_WIDEN,
+    CM_DECISION_INTERLEAVE,
+    CM_DECISION_GATHER_SCATTER,
+    CM_DECISION_SCALARIZE
+  };
+
+  /// Save vectorization decision \p W taken by the cost model for 
+  /// instruction \p I and vector width \p VF.
+  void setWideningDecision(Instruction *I, unsigned VF, InstWidening W) {
+    assert(VF >= 2 && "Expected VF >=2");
+    WideningDecisions[std::make_pair(I, VF)] = W;
+  }
+
+  /// Save vectorization decision \p W taken by the cost model for 
+  /// interleaving group \p Grp and vector width \p VF.
+  void setWideningDecision(const InterleaveGroup *Grp, unsigned VF,
+                            InstWidening W) {
+    assert(VF >= 2 && "Expected VF >=2");
+    /// Broadcast this decicion to all instructions inside the group.
+    for (unsigned i = 0; i < Grp->getFactor(); ++i) {
+      if (auto *I = Grp->getMember(i))
+        WideningDecisions[std::make_pair(I, VF)] = W;
+    }
+  }
+
+  /// Return the cost model decision for the given instruction \p I and vector
+  /// width \p VF. Return CM_DECISION_NONE if this instruction did not pass
+  /// through the cost modeling.
+  InstWidening getWideningDecision(Instruction *I, unsigned VF) {
+    assert(VF >= 2 && "Expected VF >=2");
+    std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
+    if (!WideningDecisions.count(InstOnVF))
+      return CM_DECISION_NONE;
+    return WideningDecisions[InstOnVF];
   }
 
 private:
   /// The vectorization cost is a combination of the cost itself and a boolean
   /// indicating whether any of the contributing operations will actually
   /// operate on
   /// vector values after type legalization in the backend. If this latter value
   /// is
   /// false, then all operations will be scalarized (i.e. no vectorization has
   /// actually taken place).
   typedef std::pair<unsigned, bool> VectorizationCostTy;
 
   /// Returns the expected execution cost. The unit of the cost does
   /// not matter because we use the 'cost' units to compare different
   /// vector widths. The cost that is returned is *not* normalized by
   /// the factor width.
   VectorizationCostTy expectedCost(unsigned VF);
 
   /// Returns the execution time cost of an instruction for a given vector
   /// width. Vector width of one means scalar.
   VectorizationCostTy getInstructionCost(Instruction *I, unsigned VF);
 
   /// The cost-computation logic from getInstructionCost which provides
   /// the vector type as an output parameter.
   unsigned getInstructionCost(Instruction *I, unsigned VF, Type *&VectorTy);
 
+  /// The cost computation for scalarized memory instruction.
+  unsigned getMemInstScalarizationCost(Instruction *I, unsigned VF);
+
+  /// The cost computation for interleaving group of memory instructions.
+  unsigned getInterleaveGroupCost(Instruction *I, unsigned VF);
+
   /// Returns whether the instruction is a load or store and will be a emitted
   /// as a vector operation.
   bool isConsecutiveLoadOrStore(Instruction *I);
 
   /// Create an analysis remark that explains why vectorization failed
   ///
   /// \p RemarkName is the identifier for the remark.  \return the remark object
   /// that can be streamed to.
private:
   typedef DenseMap<Instruction *, unsigned> ScalarCostsTy;
 
   /// A map holding scalar costs for different vectorization factors. The
   /// presence of a cost for an instruction in the mapping indicates that the
   /// instruction will be scalarized when vectorizing with the associated
   /// vectorization factor. The entries are VF-ScalarCostTy pairs.
   DenseMap<unsigned, ScalarCostsTy> InstsToScalarize;
 
+  /// Holds the instructions known to be uniform after vectorization.
+  /// The data is collected per VF.
+  DenseMap<unsigned, SmallPtrSet<Instruction *, 4> >Uniforms;
+
+  /// Holds the instructions known to be scalar after vectorization.
+  /// The data is collected per VF.
+  DenseMap<unsigned, SmallPtrSet<Instruction *, 4> > Scalars;
+
   /// Returns the expected difference in cost from scalarizing the expression
   /// feeding a predicated instruction \p PredInst. The instructions to
   /// scalarize and their scalar costs are collected in \p ScalarCosts. A
   /// non-negative return value implies the expression will be scalarized.
   /// Currently, only single-use chains are considered for scalarization.
   int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts,
                               unsigned VF);
 
   /// Collects the instructions to scalarize for each predicated instruction in
   /// the loop.
   void collectInstsToScalarize(unsigned VF);
 
+  /// Collect the instructions that are uniform after vectorization. An
+  /// instruction is uniform if we represent it with a single scalar value in
+  /// the vectorized loop corresponding to each vector iteration. Examples of
+  /// uniform instructions include pointer operands of consecutive or
+  /// interleaved memory accesses. Note that although uniformity implies an
+  /// instruction will be scalar, the reverse is not true. In general, a
+  /// scalarized instruction will be represented by VF scalar values in the
+  /// vectorized loop, each corresponding to an iteration of the original
+  /// scalar loop. When the \p VF is > 2 we take cost model decisions into
+  /// consideration. 
+  void collectLoopUniforms(unsigned VF);
+
+  /// Collect the instructions that are scalar after vectorization. An
+  /// instruction is scalar if it is known to be uniform or will be scalarized
+  /// during vectorization. Non-uniform scalarized instructions will be
+  /// represented by VF values in the vectorized loop, each corresponding to an
+  /// iteration of the original scalar loop. When the \p VF is > 2 
+  /// we take cost model decisions into consideration.
+  void collectLoopScalars(unsigned VF);
+
+  /// Collect Uniform and Scalar values for the given \p VF.
+  /// The sets depend on CM decision for Load/Store instructions
+  /// that may be vectorized as interleave, gather-scatter or scalarized.
+  void collectUniformsAndScalars(unsigned VF) {
+    setCostBasedWideningDecision(VF);
+    collectLoopUniforms(VF);
+    collectLoopScalars(VF);
+  }
+
+  /// Keeps cost model decisions for instructions.
+  /// Right now it is used for memory instructions only.
+  typedef DenseMap<std::pair<Instruction *, unsigned>, InstWidening> DecisionList;
+
+  DecisionList WideningDecisions;
+
 public:
   /// The loop that we evaluate.
   Loop *TheLoop;
   /// Predicated scalar evolution analysis.
   PredicatedScalarEvolution &PSE;
   /// Loop Info analysis.
   LoopInfo *LI;
   /// Vectorization legality.
void InnerLoopVectorizer::createVectorIntInductionPHI(
   LastInduction->moveBefore(ICmp);
   LastInduction->setName("vec.ind.next");
 
   VecInd->addIncoming(SteppedStart, LoopVectorPreHeader);
   VecInd->addIncoming(LastInduction, LoopVectorLatch);
 }
 
 bool InnerLoopVectorizer::shouldScalarizeInstruction(Instruction *I) const {
-  return Legal->isScalarAfterVectorization(I) ||
+  return Cost->isScalarAfterVectorization(I, VF) ||
          Cost->isProfitableToScalarize(I, VF);
 }
 
 bool InnerLoopVectorizer::needsScalarInduction(Instruction *IV) const {
   if (shouldScalarizeInstruction(IV))
     return true;
   auto isScalarInst = [&](User *U) -> bool {
     auto *I = cast<Instruction>(U);
void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step,
   Type *ScalarIVTy = ScalarIV->getType()->getScalarType();
   assert(ScalarIVTy->isIntegerTy() && ScalarIVTy == Step->getType() &&
          "Val and Step should have the same integer type");
 
   // Determine the number of scalars we need to generate for each unroll
   // iteration. If EntryVal is uniform, we only need to generate the first
   // lane. Otherwise, we generate all VF values.
   unsigned Lanes =
-      Legal->isUniformAfterVectorization(cast<Instruction>(EntryVal)) ? 1 : VF;
+    Cost->isUniformAfterVectorization(cast<Instruction>(EntryVal), VF) ? 1 : VF;
 
   // Compute the scalar steps and save the results in VectorLoopValueMap.
   ScalarParts Entry(UF);
   for (unsigned Part = 0; Part < UF; ++Part) {
     Entry[Part].resize(VF);
     for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
       auto *StartIdx = ConstantInt::get(ScalarIVTy, VF * Part + Lane);
       auto *Mul = Builder.CreateMul(StartIdx, Step);
if (VF == 1) {
         Entry[Part] = getScalarValue(V, Part, 0);
       return VectorLoopValueMap.initVector(V, Entry);
     }
 
     // Get the last scalar instruction we generated for V. If the value is
     // known to be uniform after vectorization, this corresponds to lane zero
     // of the last unroll iteration. Otherwise, the last instruction is the one
     // we created for the last vector lane of the last unroll iteration.
-    unsigned LastLane = Legal->isUniformAfterVectorization(I) ? 0 : VF - 1;
+    unsigned LastLane = Cost->isUniformAfterVectorization(I, VF) ? 0 : VF - 1;
     auto *LastInst = cast<Instruction>(getScalarValue(V, UF - 1, LastLane));
 
     // Set the insert point after the last scalarized instruction. This ensures
     // the insertelement sequence will directly follow the scalar definitions.
     auto OldIP = Builder.saveIP();
     auto NewIP = std::next(BasicBlock::iterator(LastInst));
     Builder.SetInsertPoint(&*NewIP);
 
     // However, if we are vectorizing, we need to construct the vector values.
     // If the value is known to be uniform after vectorization, we can just
     // broadcast the scalar value corresponding to lane zero for each unroll
     // iteration. Otherwise, we construct the vector values using insertelement
     // instructions. Since the resulting vectors are stored in
     // VectorLoopValueMap, we will only generate the insertelements once.
     for (unsigned Part = 0; Part < UF; ++Part) {
       Value *VectorValue = nullptr;
-      if (Legal->isUniformAfterVectorization(I)) {
+      if (Cost->isUniformAfterVectorization(I, VF)) {
         VectorValue = getBroadcastInstrs(getScalarValue(V, Part, 0));
       } else {
         VectorValue = UndefValue::get(VectorType::get(V->getType(), VF));
         for (unsigned Lane = 0; Lane < VF; ++Lane)
           VectorValue = Builder.CreateInsertElement(
               VectorValue, getScalarValue(V, Part, Lane),
               Builder.getInt32(Lane));
       }

 Value *InnerLoopVectorizer::getScalarValue(Value *V, unsigned Part,
                                            unsigned Lane) {
 
   // If the value is not an instruction contained in the loop, it should
   // already be scalar.
   if (OrigLoop->isLoopInvariant(V))
     return V;
 
-  assert(Lane > 0 ? !Legal->isUniformAfterVectorization(cast<Instruction>(V))
-                  : true && "Uniform values only have lane zero");
+  assert(Lane > 0 ?
+         !Cost->isUniformAfterVectorization(cast<Instruction>(V), VF)
 
   // If the value from the original loop has not been vectorized, it is
   // represented by UF x VF scalar values in the new loop. Return the requested
   // scalar value.
   if (VectorLoopValueMap.hasScalar(V))
     return VectorLoopValueMap.ScalarMapStorage[V][Part][Lane];
 
   // If the value has not been scalarized, get its entry in VectorLoopValueMap

 
 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
   // Attempt to issue a wide load.
   LoadInst *LI = dyn_cast<LoadInst>(Instr);
   StoreInst *SI = dyn_cast<StoreInst>(Instr);
 
   assert((LI || SI) && "Invalid Load/Store instruction");
 
-  // Try to vectorize the interleave group if this access is interleaved.
-  if (Legal->isAccessInterleaved(Instr))
+  LoopVectorizationCostModel::InstWidening Decision =
+    Cost->getWideningDecision(Instr, VF);
+  assert(Decision != LoopVectorizationCostModel::CM_DECISION_NONE &&
+         "CM decision should be taken at this point");
+  if (Decision == LoopVectorizationCostModel::CM_DECISION_INTERLEAVE)
     return vectorizeInterleaveGroup(Instr);
 
   Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
   Type *DataTy = VectorType::get(ScalarDataTy, VF);
   Value *Ptr = getPointerOperand(Instr);
   unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
   // An alignment of 0 means target abi alignment. We need to use the scalar's
   // target abi alignment in such a case.
   const DataLayout &DL = Instr->getModule()->getDataLayout();
   if (!Alignment)
     Alignment = DL.getABITypeAlignment(ScalarDataTy);
   unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
 
   // Scalarize the memory instruction if necessary.
-  if (Legal->memoryInstructionMustBeScalarized(Instr, VF))
+  if (Decision == LoopVectorizationCostModel::CM_DECISION_SCALARIZE)
     return scalarizeInstruction(Instr, Legal->isScalarWithPredication(Instr));
 
   // Determine if the pointer operand of the access is either consecutive or
   // reverse consecutive.
   int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
   bool Reverse = ConsecutiveStride < 0;
-
-  // Determine if either a gather or scatter operation is legal.
   bool CreateGatherScatter =
-      !ConsecutiveStride && Legal->isLegalGatherOrScatter(Instr);
+    (Decision == LoopVectorizationCostModel::CM_DECISION_GATHER_SCATTER);
 
   VectorParts VectorGep;
 
   // Handle consecutive loads/stores.
   GetElementPtrInst *Gep = getGEPInstruction(Ptr);
   if (ConsecutiveStride) {
     if (Gep) {
       unsigned NumOperands = Gep->getNumOperands();
void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr,
 
   VectorParts Cond;
   if (IfPredicateInstr)
     Cond = createBlockInMask(Instr->getParent());
 
   // Determine the number of scalars we need to generate for each unroll
   // iteration. If the instruction is uniform, we only need to generate the
   // first lane. Otherwise, we generate all VF values.
-  unsigned Lanes = Legal->isUniformAfterVectorization(Instr) ? 1 : VF;
+  unsigned Lanes = Cost->isUniformAfterVectorization(Instr, VF) ? 1 : VF;
 
   // For each vector unroll 'part':
   for (unsigned Part = 0; Part < UF; ++Part) {
     Entry[Part].resize(VF);
     // For each scalar that we create:
     for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
 
       // Start if-block.
case InductionDescriptor::IK_PtrInduction: {
     // Handle the pointer induction variable case.
     assert(P->getType()->isPointerTy() && "Unexpected type.");
     // This is the normalized GEP that starts counting at zero.
     Value *PtrInd = Induction;
     PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStep()->getType());
     // Determine the number of scalars we need to generate for each unroll
     // iteration. If the instruction is uniform, we only need to generate the
     // first lane. Otherwise, we generate all VF values.
-    unsigned Lanes = Legal->isUniformAfterVectorization(P) ? 1 : VF;
+    unsigned Lanes = Cost->isUniformAfterVectorization(P, VF) ? 1 : VF;
     // These are the scalar results. Notice that we don't generate vector GEPs
     // because scalar GEPs result in better code.
     ScalarParts Entry(UF);
     for (unsigned Part = 0; Part < UF; ++Part) {
       Entry[Part].resize(VF);
       for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
         Constant *Idx = ConstantInt::get(PtrInd->getType(), Lane + Part * VF);
         Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
bool LoopVectorizationLegality::canVectorize() {
   // If an override option has been passed in for interleaved accesses, use it.
   if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
     UseInterleaved = EnableInterleavedMemAccesses;
 
   // Analyze interleaved memory accesses.
   if (UseInterleaved)
     InterleaveInfo.analyzeInterleaving(*getSymbolicStrides());
 
-  // Collect all instructions that are known to be uniform after vectorization.
-  collectLoopUniforms();
-
-  // Collect all instructions that are known to be scalar after vectorization.
-  collectLoopScalars();
-
   unsigned SCEVThreshold = VectorizeSCEVCheckThreshold;
   if (Hints->getForce() == LoopVectorizeHints::FK_Enabled)
     SCEVThreshold = PragmaVectorizeSCEVCheckThreshold;
 
   if (PSE.getUnionPredicate().getComplexity() > SCEVThreshold) {
     ORE->emit(createMissedAnalysis("TooManySCEVRunTimeChecks")
               << "Too many SCEV assumptions need to be made and checked "
               << "at runtime");
bool LoopVectorizationLegality::canVectorizeInstrs() {
   // is the same size. If it's not, unset it here and InnerLoopVectorizer
   // will create another.
   if (Induction && WidestIndTy != Induction->getType())
     Induction = nullptr;
 
   return true;
 }
 
-void LoopVectorizationLegality::collectLoopScalars() {
+void LoopVectorizationCostModel::collectLoopScalars(unsigned VF) {
 
+  if (!Scalars[VF].empty())
+    return;
   // If an instruction is uniform after vectorization, it will remain scalar.
-  Scalars.insert(Uniforms.begin(), Uniforms.end());
+  Scalars[VF].insert(Uniforms[VF].begin(), Uniforms[VF].end());
 
   // Collect the getelementptr instructions that will not be vectorized. A
   // getelementptr instruction is only vectorized if it is used for a legal
   // gather or scatter operation.
   for (auto *BB : TheLoop->blocks())
     for (auto &I : *BB) {
       if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
-        Scalars.insert(GEP);
+        Scalars[VF].insert(GEP);
         continue;
       }
       auto *Ptr = getPointerOperand(&I);
       if (!Ptr)
         continue;
       auto *GEP = getGEPInstruction(Ptr);
-      if (GEP && isLegalGatherOrScatter(&I))
-        Scalars.erase(GEP);
+      if (GEP && VF > 1 &&
+          getWideningDecision(&I, VF) == CM_DECISION_GATHER_SCATTER)
+        Scalars[VF].erase(GEP);
     }
 
   // An induction variable will remain scalar if all users of the induction
   // variable and induction variable update remain scalar.
   auto *Latch = TheLoop->getLoopLatch();
-  for (auto &Induction : *getInductionVars()) {
+  for (auto &Induction : *Legal->getInductionVars()) {
     auto *Ind = Induction.first;
     auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
 
     // Determine if all users of the induction variable are scalar after
     // vectorization.
     auto ScalarInd = all_of(Ind->users(), [&](User *U) -> bool {
       auto *I = cast<Instruction>(U);
-      return I == IndUpdate || !TheLoop->contains(I) || Scalars.count(I);
+      return I == IndUpdate || !TheLoop->contains(I) || Scalars[VF].count(I);
     });
     if (!ScalarInd)
       continue;
 
     // Determine if all users of the induction variable update instruction are
     // scalar after vectorization.
     auto ScalarIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {
       auto *I = cast<Instruction>(U);
-      return I == Ind || !TheLoop->contains(I) || Scalars.count(I);
+      return I == Ind || !TheLoop->contains(I) || Scalars[VF].count(I);
     });
     if (!ScalarIndUpdate)
       continue;
 
     // The induction variable and its update instruction will remain scalar.
-    Scalars.insert(Ind);
-    Scalars.insert(IndUpdate);
+    Scalars[VF].insert(Ind);
+    Scalars[VF].insert(IndUpdate);
   }
 }
 
-bool LoopVectorizationLegality::hasConsecutiveLikePtrOperand(Instruction *I) {
-  if (isAccessInterleaved(I))
+bool LoopVectorizationCostModel::hasConsecutiveLikePtrOperand(Instruction *I,
+                                                              unsigned VF) {
+  if (VF > 1 && getWideningDecision(I, VF) == CM_DECISION_INTERLEAVE)
     return true;
   if (auto *Ptr = getPointerOperand(I))
-    return isConsecutivePtr(Ptr);
+    return Legal->isConsecutivePtr(Ptr);
   return false;
 }
 
 bool LoopVectorizationLegality::isScalarWithPredication(Instruction *I) {
   if (!blockNeedsPredication(I->getParent()))
     return false;
   switch(I->getOpcode()) {
   default:
     break;
   case Instruction::Store:
     return !isMaskRequired(I);
   case Instruction::UDiv:
   case Instruction::SDiv:
   case Instruction::SRem:
   case Instruction::URem:
     return mayDivideByZero(*I);
   }
   return false;
 }
 
-bool LoopVectorizationLegality::memoryInstructionMustBeScalarized(
-    Instruction *I, unsigned VF) {
-
-  // If the memory instruction is in an interleaved group, it will be
-  // vectorized and its pointer will remain uniform.
-  if (isAccessInterleaved(I))
-    return false;
-
+bool LoopVectorizationLegality::memoryInstructionCanBeWidened(Instruction *I,
+                                                              unsigned VF) {
   // Get and ensure we have a valid memory instruction.
   LoadInst *LI = dyn_cast<LoadInst>(I);
   StoreInst *SI = dyn_cast<StoreInst>(I);
   assert((LI || SI) && "Invalid memory instruction");
 
   // If the pointer operand is uniform (loop invariant), the memory instruction
   // will be scalarized.
   auto *Ptr = getPointerOperand(I);
   if (LI && isUniform(Ptr))
-    return true;
+    return false;
 
-  // If the pointer operand is non-consecutive and neither a gather nor a
-  // scatter operation is legal, the memory instruction will be scalarized.
-  if (!isConsecutivePtr(Ptr) && !isLegalGatherOrScatter(I))
-    return true;
+  if (!isConsecutivePtr(Ptr))
+    return false;
 
   // If the instruction is a store located in a predicated block, it will be
   // scalarized.
   if (isScalarWithPredication(I))
-    return true;
+    return false;
 
   // If the instruction's allocated size doesn't equal it's type size, it
   // requires padding and will be scalarized.
   auto &DL = I->getModule()->getDataLayout();
   auto *ScalarTy = LI ? LI->getType() : SI->getValueOperand()->getType();
   if (hasIrregularType(ScalarTy, DL, VF))
-    return true;
+    return false;
+
+}
+
+bool LoopVectorizationLegality::memoryInstructionMustBeScalarized(
+  Instruction *I, unsigned VF) {
 
-  // Otherwise, the memory instruction should be vectorized if the rest of the
-  // loop is.
+  // If the memory instruction is in an interleaved group, it will be
+  // vectorized and its pointer will remain uniform.
+  if (isAccessInterleaved(I) || isLegalGatherOrScatter(I))
     return false;
+
+  return !memoryInstructionCanBeWidened(I, VF);
 }
 
-void LoopVectorizationLegality::collectLoopUniforms() {
+void LoopVectorizationCostModel::collectLoopUniforms(unsigned VF) {
+  if (!Uniforms[VF].empty())
+    return;
   // We now know that the loop is vectorizable!
   // Collect instructions inside the loop that will remain uniform after
   // vectorization.
 
   // Global values, params and instructions outside of current loop are out of
   // scope.
   auto isOutOfScope = [&](Value *V) -> bool {
     Instruction *I = dyn_cast<Instruction>(V);
for (auto &I : *BB) {
         return getPointerOperand(U) == Ptr;
       });
 
       // Ensure the memory instruction will not be scalarized, making its
       // pointer operand non-uniform. If the pointer operand is used by some
       // instruction other than a memory access, we're not going to check if
       // that other instruction may be scalarized here. Thus, conservatively
       // assume the pointer operand may be non-uniform.
-      if (!UsersAreMemAccesses || memoryInstructionMustBeScalarized(&I))
+      if (!UsersAreMemAccesses || Legal->memoryInstructionMustBeScalarized(&I) ||
+          (VF >= 2 && getWideningDecision(&I, VF) == CM_DECISION_SCALARIZE))
         PossibleNonUniformPtrs.insert(Ptr);
 
       // If the memory instruction will be vectorized and its pointer operand
       // is consecutive-like, the pointer operand should remain uniform.
-      else if (hasConsecutiveLikePtrOperand(&I))
+      else if (hasConsecutiveLikePtrOperand(&I, VF))
         ConsecutiveLikePtrs.insert(Ptr);
 
       // Otherwise, if the memory instruction will be vectorized and its
       // pointer operand is non-consecutive-like, the memory instruction should
       // be a gather or scatter operation. Its pointer operand will be
       // non-uniform.
       else
         PossibleNonUniformPtrs.insert(Ptr);
for (auto OV : I->operand_values()) {
         DEBUG(dbgs() << "LV: Found uniform instruction: " << *OI << "\n");
       }
     }
   }
 
   // Returns true if Ptr is the pointer operand of a memory access instruction
   // I, and I is known to not require scalarization.
   auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool {
-    return getPointerOperand(I) == Ptr && !memoryInstructionMustBeScalarized(I);
+    return getPointerOperand(I) == Ptr &&
+      !Legal->memoryInstructionMustBeScalarized(I);
   };
 
   // For an instruction to be added into Worklist above, all its users inside
   // the loop should also be in Worklist. However, this condition cannot be
   // true for phi nodes that form a cyclic dependence. We must process phi
   // nodes separately. An induction variable will remain uniform if all users
   // of the induction variable and induction variable update remain uniform.
   // The code below handles both pointer and non-pointer induction variables.
-  for (auto &Induction : Inductions) {
+  for (auto &Induction : *Legal->getInductionVars()) {
     auto *Ind = Induction.first;
     auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));
 
     // Determine if all users of the induction variable are uniform after
     // vectorization.
     auto UniformInd = all_of(Ind->users(), [&](User *U) -> bool {
       auto *I = cast<Instruction>(U);
       return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) ||
for (auto &Induction : *Legal->getInductionVars()) {
 
     // The induction variable and its update instruction will remain uniform.
     Worklist.insert(Ind);
     Worklist.insert(IndUpdate);
     DEBUG(dbgs() << "LV: Found uniform instruction: " << *Ind << "\n");
     DEBUG(dbgs() << "LV: Found uniform instruction: " << *IndUpdate << "\n");
   }
 
-  Uniforms.insert(Worklist.begin(), Worklist.end());
+  Uniforms[VF].insert(Worklist.begin(), Worklist.end());
 }
 
 bool LoopVectorizationLegality::canVectorizeMemory() {
   LAI = &(*GetLAA)(*TheLoop);
   InterleaveInfo.setLAI(LAI);
   const OptimizationRemarkAnalysis *LAR = LAI->getReport();
   if (LAR) {
     OptimizationRemarkAnalysis VR(Hints->vectorizeAnalysisPassName(),
LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
   }
 
   int UserVF = Hints->getWidth();
   if (UserVF != 0) {
     assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
     DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
 
     Factor.Width = UserVF;
+
+    collectUniformsAndScalars(UserVF);
     collectInstsToScalarize(UserVF);
     return Factor;
   }
 
   float Cost = expectedCost(1).first;
 #ifndef NDEBUG
   const float ScalarCost = Cost;
 #endif /* NDEBUG */
for (Instruction &I : *BB)
       }
   }
 }
 
 int LoopVectorizationCostModel::computePredInstDiscount(
     Instruction *PredInst, DenseMap<Instruction *, unsigned> &ScalarCosts,
     unsigned VF) {
 
-  assert(!Legal->isUniformAfterVectorization(PredInst) &&
+  assert(!isUniformAfterVectorization(PredInst, VF) &&
          "Instruction marked uniform-after-vectorization will be predicated");
 
   // Initialize the discount to zero, meaning that the scalar version and the
   // vector version cost the same.
   int Discount = 0;
 
   // Holds instructions to analyze. The instructions we visit are mapped in
   // ScalarCosts. Those instructions are the ones that would be scalarized if
   // we find that the scalar version costs less.
   SmallVector<Instruction *, 8> Worklist;
 
   // Returns true if the given instruction can be scalarized.
   auto canBeScalarized = [&](Instruction *I) -> bool {
 
     // We only attempt to scalarize instructions forming a single-use chain
     // from the original predicated block that would otherwise be vectorized.
     // Although not strictly necessary, we give up on instructions we know will
     // already be scalar to avoid traversing chains that are unlikely to be
     // beneficial.
     if (!I->hasOneUse() || PredInst->getParent() != I->getParent() ||
-        Legal->isScalarAfterVectorization(I))
+        isScalarAfterVectorization(I, VF))
       return false;
 
     // If the instruction is scalar with predication, it will be analyzed
     // separately. We ignore it within the context of PredInst.
     if (Legal->isScalarWithPredication(I))
       return false;
 
     // If any of the instruction's operands are uniform after vectorization,
     // the instruction cannot be scalarized. This prevents, for example, a
     // masked load from being scalarized.
     //
     // We assume we will only emit a value for lane zero of an instruction
     // marked uniform after vectorization, rather than VF identical values.
     // Thus, if we scalarize an instruction that uses a uniform, we would
     // create uses of values corresponding to the lanes we aren't emitting code
     // for. This behavior can be changed by allowing getScalarValue to clone
     // the lane zero values for uniforms rather than asserting.
     for (Use &U : I->operands())
       if (auto *J = dyn_cast<Instruction>(U.get()))
-        if (Legal->isUniformAfterVectorization(J))
+        if (isUniformAfterVectorization(J, VF))
           return false;
 
     // Otherwise, we can scalarize the instruction.
     return true;
   };
 
   // Returns true if an operand that cannot be scalarized must be extracted
   // from a vector. We will account for this scalarization overhead below. Note
   // that the non-void predicated instructions are placed in their own blocks,
   // and their return values are inserted into vectors. Thus, an extract would
   // still be required.
   auto needsExtract = [&](Instruction *I) -> bool {
-    return TheLoop->contains(I) && !Legal->isScalarAfterVectorization(I);
+    return TheLoop->contains(I) && !isScalarAfterVectorization(I, VF);
   };
 
   // Compute the expected cost discount from scalarizing the entire expression
   // feeding the predicated instruction. We currently only consider expressions
   // that are single-use instruction chains.
   Worklist.push_back(PredInst);
   while (!Worklist.empty()) {
     Instruction *I = Worklist.pop_back_val();
int LoopVectorizationCostModel::computePredInstDiscount(
 
   return Discount;
 }
 
 LoopVectorizationCostModel::VectorizationCostTy
 LoopVectorizationCostModel::expectedCost(unsigned VF) {
   VectorizationCostTy Cost;
 
+  // Collect Uniform and Scalar instructions after vectorization with VF.
+  collectUniformsAndScalars(VF);
+
   // Collect the instructions (and their associated costs) that will be more
   // profitable to scalarize.
   collectInstsToScalarize(VF);
 
   // For each block.
   for (BasicBlock *BB : TheLoop->blocks()) {
     VectorizationCostTy BlockCost;
 
static const SCEV *getAddressAccessSCEV(
   return SE->getSCEV(Ptr);
 }
 
 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
   return Legal->hasStride(I->getOperand(0)) ||
          Legal->hasStride(I->getOperand(1));
 }
 
+unsigned 
+LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I,
+                                                        unsigned VF) {
+  StoreInst *SI = dyn_cast<StoreInst>(I);
+  LoadInst *LI = dyn_cast<LoadInst>(I);
+  Type *ValTy = (SI ? SI->getValueOperand()->getType() : LI->getType());
+  auto SE = PSE.getSE();
+
+  unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
+  unsigned AS =
+    SI ? SI->getPointerAddressSpace() : LI->getPointerAddressSpace();
+  Value *Ptr = getPointerOperand(I);
+  Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
+
+  // Figure out whether the access is strided and get the stride value
+  // if it's known in compile time
+  const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, SE, TheLoop); 
+
+  // Get the cost of the scalar memory instruction and address computation.
+  unsigned Cost = VF * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV);
+
+  Cost += VF *
+    TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
+                        Alignment, AS);
+
+  // Get the overhead of the extractelement and insertelement instructions
+  // we might create due to scalarization.
+  Cost += getScalarizationOverhead(I, VF, TTI);
+
+  // If we have a predicated store, it may not be executed for each vector
+  // lane. Scale the cost by the probability of executing the predicated
+  // block.
+  if (Legal->isScalarWithPredication(I))
+    Cost /= getReciprocalPredBlockProb();
+
+  return Cost;
+}
+
+unsigned
+LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I, unsigned VF) {
+  StoreInst *SI = dyn_cast<StoreInst>(I);
+  LoadInst *LI = dyn_cast<LoadInst>(I);
+  Type *ValTy = (SI ? SI->getValueOperand()->getType() : LI->getType());
+  Type *VectorTy = ToVectorTy(ValTy, VF);
+  unsigned AS =
+    SI ? SI->getPointerAddressSpace() : LI->getPointerAddressSpace();
+
+  auto Group = Legal->getInterleavedAccessGroup(I);
+  assert(Group && "Fail to get an interleaved access group.");
+
+  unsigned InterleaveFactor = Group->getFactor();
+  Type *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor);
+
+  // Holds the indices of existing members in an interleaved load group.
+  // An interleaved store group doesn't need this as it doesn't allow gaps.
+  SmallVector<unsigned, 4> Indices;
+  if (LI) {
+    for (unsigned i = 0; i < InterleaveFactor; i++)
+      if (Group->getMember(i))
+        Indices.push_back(i);
+  }
+
+  // Calculate the cost of the whole interleaved group.
+  unsigned Cost = TTI.getInterleavedMemoryOpCost(
+    I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
+    Group->getAlignment(), AS);
+
+  if (Group->isReverse())
+    Cost += Group->getNumMembers() *
+      TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
+  return Cost;
+}
+
 LoopVectorizationCostModel::VectorizationCostTy
 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
   // If we know that this instruction will remain uniform, check the cost of
   // the scalar version.
-  if (Legal->isUniformAfterVectorization(I))
+  if (VF > 1 && isUniformAfterVectorization(I, VF))
     VF = 1;
 
   if (VF > 1 && isProfitableToScalarize(I, VF))
     return VectorizationCostTy(InstsToScalarize[VF][I], false);
 
   Type *VectorTy;
   unsigned C = getInstructionCost(I, VF, VectorTy);
 
   bool TypeNotScalarized =
       VF > 1 && !VectorTy->isVoidTy() && TTI.getNumberOfParts(VectorTy) < VF;
   return VectorizationCostTy(C, TypeNotScalarized);
 }
 
+void LoopVectorizationCostModel::setCostBasedWideningDecision(unsigned VF) {
+  if (VF <= 1)
+    return;
+  for (BasicBlock *BB : TheLoop->blocks()) {
+    VectorizationCostTy BlockCost;
+
+    // For each instruction in the old loop.
+    for (Instruction &I : *BB) {
+      if (I.getOpcode() != Instruction::Store &&
+          I.getOpcode() != Instruction::Load)
+        continue;
+
+      StoreInst *SI = dyn_cast<StoreInst>(&I);
+      LoadInst *LI = dyn_cast<LoadInst>(&I);
+
+      Type *ValTy = (SI ? SI->getValueOperand()->getType() : LI->getType());
+      Type *VectorTy = ToVectorTy(ValTy, VF);
+
+      unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
+      Value *Ptr = getPointerOperand(&I);
+
+      if (LI && Legal->isUniform(Ptr)) {
+        // Scalar load + broadcast
+        setWideningDecision(&I, VF, CM_DECISION_SCALARIZE);
+        continue;
+      }
+
+      // We assume that widening is the best solution when possible.
+      if (Legal->memoryInstructionCanBeWidened(&I, VF)) {
+        setWideningDecision(&I, VF, CM_DECISION_WIDEN);
+        continue;
+      }
+
+      // Choose between Interleaving, Gather/Scatter or Scalarization.
+      unsigned InterleaveGrpCost = 0;
+      unsigned InterleaveInstCost = (unsigned)(-1);
+      unsigned GatherScatterCost = (unsigned)(-1);
+      if (Legal->isAccessInterleaved(&I)) {
+        auto Group = Legal->getInterleavedAccessGroup(&I);
+        assert(Group && "Fail to get an interleaved access group.");
+
+        InstWidening Decision = getWideningDecision(&I, VF);
+        if (Decision == CM_DECISION_NONE) {
+          // Only calculate the cost once for the group.
+          InterleaveGrpCost = getInterleaveGroupCost(&I, VF);
+
+          if (InterleaveGrpCost / Group->getNumMembers() <= 1) {
+            // The interleaving cost is good enough.
+            setWideningDecision(Group, VF, CM_DECISION_INTERLEAVE);
+            continue;
+          }
+          // We are going to check all other options.
+          InterleaveInstCost = InterleaveGrpCost / Group->getNumMembers();
+
+        } else if (Decision == CM_DECISION_INTERLEAVE)
+          // This instructions belongs to interleaving group and will be 
+          // vectorized inside this group.
+          continue;
+        // If we are here it means that the instruction belongs to an 
+        // interleaving group, but interleaving is not profitable for the current 
+        // VF. We'll need to choose another option.
+      }
+
+      if (Legal->isLegalGatherOrScatter(&I)) {
+        GatherScatterCost = TTI.getAddressComputationCost(VectorTy) +
+          TTI.getGatherScatterOpCost(I.getOpcode(), VectorTy, Ptr,
+                                     Legal->isMaskRequired(&I), Alignment);
+      }
+      unsigned ScalarizationCost = getMemInstScalarizationCost(&I, VF);
+
+      // Choose better solution for the current VF,
+      // write down this decision and use it during vectorization.
+      InstWidening Decision;
+      if (InterleaveInstCost <= GatherScatterCost &&
+          InterleaveInstCost <= ScalarizationCost)
+        Decision = CM_DECISION_INTERLEAVE;
+      else if (GatherScatterCost < InterleaveInstCost &&
+               GatherScatterCost < ScalarizationCost)
+        Decision = CM_DECISION_GATHER_SCATTER;
+      else
+        Decision = CM_DECISION_SCALARIZE;
+
+      // If the instructions belongs to an interleave group, the whole group
+      // receives the same decision.
+      if (auto Group = Legal->getInterleavedAccessGroup(&I))
+        setWideningDecision(Group, VF, Decision);
+      else
+        setWideningDecision(&I, VF, Decision);
+    }
+  }
+}
+
 unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I,
                                                         unsigned VF,
                                                         Type *&VectorTy) {
   Type *RetTy = I->getType();
   if (canTruncateToMinimalBitwidth(I, VF))
     RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);
   VectorTy = ToVectorTy(RetTy, VF);
   auto SE = PSE.getSE();
if (VF == 1)
              TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
 
     if (LI && Legal->isUniform(Ptr)) {
       // Scalar load + broadcast
       unsigned Cost = TTI.getAddressComputationCost(ValTy->getScalarType());
       Cost += TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
                                   Alignment, AS);
       return Cost +
-             TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, ValTy);
+             TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy);
     }
 
     // For an interleaved access, calculate the total cost of the whole
     // interleave group.
-    if (Legal->isAccessInterleaved(I)) {
+    InstWidening Decision = getWideningDecision(I, VF);
+    assert(Decision != CM_DECISION_NONE &&
+           "Widening decision should be taken at this moment");
+    if (Decision == CM_DECISION_INTERLEAVE) {
       auto Group = Legal->getInterleavedAccessGroup(I);
       assert(Group && "Fail to get an interleaved access group.");
 
       // Only calculate the cost once at the insert position.
       if (Group->getInsertPos() != I)
         return 0;
 
       unsigned InterleaveFactor = Group->getFactor();
       Type *WideVecTy =
           VectorType::get(VectorTy->getVectorElementType(),
                           VectorTy->getVectorNumElements() * InterleaveFactor);
 
       // Holds the indices of existing members in an interleaved load group.
       // An interleaved store group doesn't need this as it doesn't allow gaps.
       SmallVector<unsigned, 4> Indices;
       if (LI) {
         for (unsigned i = 0; i < InterleaveFactor; i++)
           if (Group->getMember(i))
             Indices.push_back(i);
       }
 
       // Calculate the cost of the whole interleaved group.
       unsigned Cost = TTI.getInterleavedMemoryOpCost(
           I->getOpcode(), WideVecTy, Group->getFactor(), Indices,
           Group->getAlignment(), AS);
 
       if (Group->isReverse())
         Cost +=
             Group->getNumMembers() *
             TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
 
-      // FIXME: The interleaved load group with a huge gap could be even more
-      // expensive than scalar operations. Then we could ignore such group and
-      // use scalar operations instead.
       return Cost;
     }
 
     // Check if the memory instruction will be scalarized.
-    if (Legal->memoryInstructionMustBeScalarized(I, VF)) {
+    if (Decision == CM_DECISION_SCALARIZE) {
       unsigned Cost = 0;
       Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
 
       // Figure out whether the access is strided and get the stride value
       // if it's known in compile time
       const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, SE, TheLoop); 
 
       // Get the cost of the scalar memory instruction and address computation.
       Cost += VF * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV);
       Cost += VF *
               TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
                                   Alignment, AS);
 
       // Get the overhead of the extractelement and insertelement instructions
       // we might create due to scalarization.
       Cost += getScalarizationOverhead(I, VF, TTI);
 
       // If we have a predicated store, it may not be executed for each vector
       // lane. Scale the cost by the probability of executing the predicated
       // block.
       if (Legal->isScalarWithPredication(I))
         Cost /= getReciprocalPredBlockProb();
 
       return Cost;
     }
 
     // Determine if the pointer operand of the access is either consecutive or
     // reverse consecutive.
     int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
-    bool Reverse = ConsecutiveStride < 0;
-
-    // Determine if either a gather or scatter operation is legal.
-    bool UseGatherOrScatter =
-        !ConsecutiveStride && Legal->isLegalGatherOrScatter(I);
 
     unsigned Cost = TTI.getAddressComputationCost(VectorTy);
-    if (UseGatherOrScatter) {
+    if (Decision == CM_DECISION_GATHER_SCATTER) {
       assert(ConsecutiveStride == 0 &&
              "Gather/Scatter are not used for consecutive stride");
       return Cost +
              TTI.getGatherScatterOpCost(I->getOpcode(), VectorTy, Ptr,
                                         Legal->isMaskRequired(I), Alignment);
     }
     // Wide load/stores.
+    assert(Decision == CM_DECISION_WIDEN && ConsecutiveStride != 0 &&
+           "Unexpected CM decision to handle");
     if (Legal->isMaskRequired(I))
       Cost +=
           TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
     else
       Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
 
+    bool Reverse = ConsecutiveStride < 0;
     if (Reverse)
       Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, 0);
     return Cost;
   }
   case Instruction::ZExt:
   case Instruction::SExt:
   case Instruction::FPToUI:
   case Instruction::FPToSI:
void LoopVectorizationCostModel::collectValuesToIgnore() {
   // Ignore type-promoting instructions we identified during reduction
   // detection.
   for (auto &Reduction : *Legal->getReductionVars()) {
     RecurrenceDescriptor &RedDes = Reduction.second;
     SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
     VecValuesToIgnore.insert(Casts.begin(), Casts.end());
   }
 
   // Insert values known to be scalar into VecValuesToIgnore. This is a
   // conservative estimation of the values that will later be scalarized.
   //
   // FIXME: Even though an instruction is not scalar-after-vectoriztion, it may
   //        still be scalarized. For example, we may find an instruction to be
   //        more profitable for a given vectorization factor if it were to be
   //        scalarized. But at this point, we haven't yet computed the
-  //        vectorization factor.
+  //        vectorization factor. Taking VF=2 for estimation.
+  unsigned VFForEstimation = 2;
+  collectUniformsAndScalars(VFForEstimation);
   for (auto *BB : TheLoop->getBlocks())
     for (auto &I : *BB)
-      if (Legal->isScalarAfterVectorization(&I))
+      if (isScalarAfterVectorization(&I, VFForEstimation))
          VecValuesToIgnore.insert(&I);
 }
 
 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
                                              bool IfPredicateInstr) {
   assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
   // Holds vector parameters or scalars, in case of uniform vals.
   SmallVector<VectorParts, 4> Params;
 

../test/Transforms/LoopVectorize/AArch64/interleaved-vs-scalar.ll

+; REQUIRES: asserts
+; RUN: opt < %s -force-vector-width=2 -force-vector-interleave=1 -loop-vectorize -S --debug-only=loop-vectorize 2>&1 | FileCheck %s
+
+; This test shows extremely high interleaving cost that, probably, should be fixed.
+; Due to the high cost, interleaving is not beneficial and the cost model chooses to scalarize
+; the load instructions.
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64--linux-gnu"
+
+%pair = type { i8, i8 }
+
+; CHECK-LABEL: test
+; CHECK: Found an estimated cost of 10 for VF 2 For instruction:   {{.*}} load i8
+; CHECK: vector.body
+; CHECK: load i8
+; CHECK: load i8
+; CHECK: br i1 {{.*}}, label %middle.block, label %vector.body
+
+define void @test(%pair* %p, i64 %n) {
+entry:
+  br label %for.body
+
+for.body:
+  %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+  %tmp0 = getelementptr %pair, %pair* %p, i64 %i, i32 0
+  %tmp1 = load i8, i8* %tmp0, align 1
+  %tmp2 = getelementptr %pair, %pair* %p, i64 %i, i32 1
+  %tmp3 = load i8, i8* %tmp2, align 1
+  %i.next = add nuw nsw i64 %i, 1
+  %cond = icmp eq i64 %i.next, %n
+  br i1 %cond, label %for.end, label %for.body
+
+for.end:
+  ret void
+}
+

../test/Transforms/LoopVectorize/X86/consecutive-ptr-uniforms.ll

 ; scatter operation. %tmp3 (and the induction variable) should not be marked
 ; uniform-after-vectorization.
 ;
 ; CHECK:     LV: Found uniform instruction: %tmp0 = getelementptr inbounds %data, %data* %d, i64 0, i32 3, i64 %i
 ; CHECK-NOT: LV: Found uniform instruction: %tmp3 = getelementptr inbounds %data, %data* %d, i64 0, i32 0, i64 %i
 ; CHECK-NOT: LV: Found uniform instruction: %i = phi i64 [ %i.next, %for.body ], [ 0, %entry ]
 ; CHECK-NOT: LV: Found uniform instruction: %i.next = add nuw nsw i64 %i, 5
 ; CHECK:     vector.body:
+; CHECK:       %index = phi i64
 ; CHECK:       %vec.ind = phi <16 x i64>
-; CHECK:       %[[T0:.+]] = extractelement <16 x i64> %vec.ind, i32 0
+; CHECK:       %[[T0:.+]] = mul i64 %index, 5
 ; CHECK:       %[[T1:.+]] = getelementptr inbounds %data, %data* %d, i64 0, i32 3, i64 %[[T0]]
 ; CHECK:       %[[T2:.+]] = bitcast float* %[[T1]] to <80 x float>*
 ; CHECK:       load <80 x float>, <80 x float>* %[[T2]], align 4
 ; CHECK:       %[[T3:.+]] = getelementptr inbounds %data, %data* %d, i64 0, i32 0, i64 %[[T0]]
 ; CHECK:       %[[T4:.+]] = bitcast float* %[[T3]] to <80 x float>*
 ; CHECK:       load <80 x float>, <80 x float>* %[[T4]], align 4
 ; CHECK:       %VectorGep = getelementptr inbounds %data, %data* %d, i64 0, i32 0, <16 x i64> %vec.ind
 ; CHECK:       call void @llvm.masked.scatter.v16f32({{.*}}, <16 x float*> %VectorGep, {{.*}})

../test/Transforms/LoopVectorize/X86/gather-vs-interleave.ll

+; RUN: opt -loop-vectorize -S -mcpu=skylake-avx512  < %s | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; This test checks that "gather" operation is choosen since it's cost is better
+; than interleaving pattern.
+;
+;unsigned long A[SIZE];
+;unsigned long B[SIZE];
+;
+;void foo() {
+;  for (int i=0; i<N; i+=8) {
+;    B[i] = A[i] + 5;
+;  }
+;}
+
+@A = global [10240 x i64] zeroinitializer, align 16
+@B = global [10240 x i64] zeroinitializer, align 16
+
+
+; CHECK_LABEL: strided_load_i64
+; CHECK: masked.gather
+define void @strided_load_i64() {
+  br label %1
+
+; <label>:1:                                      ; preds = %0, %1
+  %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+  %2 = getelementptr inbounds [10240 x i64], [10240 x i64]* @A, i64 0, i64 %indvars.iv
+  %3 = load i64, i64* %2, align 16
+  %4 = add i64 %3, 5
+  %5 = getelementptr inbounds [10240 x i64], [10240 x i64]* @B, i64 0, i64 %indvars.iv
+  store i64 %4, i64* %5, align 16
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 8
+  %6 = icmp slt i64 %indvars.iv.next, 1024
+  br i1 %6, label %1, label %7
+
+; <label>:7:                                      ; preds = %1
+  ret void
+}
+