[LoopVectorize] Fix crash for predicated instruction with scalable VF

[CarolineConcatto]

Summary

This patch avoids computing discounts for predicated instructions when the
VF is scalable.
There is no support for vectorization of loops with division because the
vectorizer cannot guarantee that zero divisions will not happen.

This loop now does not use VF scalable

for (long long i = 0; i < n; i++)
    if (cond[i])
      a[i] /= b[i];

[CarolineConcatto]

-s/crach/crash/

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

1282

nit: when the loop contains a predicated instruction that requires scalarization.

1283

nit: loopHasScalarWithPredication

7882–7884

is this condition necessary? I would expect any instructions that explicitly need predication+scalarization, to never be values that can be ignored.

7885

There is a problem with this approach, in that isScalarWithPredication calls blockNeedsPredication, which in turn calls foldTailByMasking, whose return value is determined after calling computeFeasibleMaxVF, which ends up calling this function in order to determine the min/max VF. This may lead to unexpected results when we want to use scalable vectors for tail-folded (predicated) loops, so you'll need to figure out a better place to call loopHasPredicatedScalar.

Additionally, it might be useful to add a mechanism that records if FoldTailByMasking already has its final value, and then assert that in isScalarWithPredication.

[david-arm]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

7885

OK that's a good point. It's tricky because we also need to ensure we call this before attempting to calculate any costs as we'll hit asserts in computePredInstDiscount again. If we didn't have to worry about blockNeedsPredication, this seems like a sensible place to bail out because we have to also worry about the general case without hints, i.e. we should really be telling the computeMaxVF code the max legal scalable VF = 0 as I believe this is how you've been recently structuring the code to work?

I realise this is not an ideal solution, but what if for now we could avoid concerns about tail folding by somehow marking tail folding as not an option for scalable vectors, since we don't support it anyway for scalable vectors?

[CarolineConcatto]

• address review's comment about circular dependency with foldTailByMasking
• add dependency in D102437 because of memory cost model

[CarolineConcatto]

-remove old comments in loopHasScalarWithPredication

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5602

Should this be !isScalarEpilogueAllowed()? i.e. if a scalar epilogue loop is not allowed, then we know predication is required.

[CarolineConcatto]

Thank you @david-arm and @sdesmalen for the review.
I believe I've addressed all your comments.
I've updated the patch based on D102437 because there was a dependency between the cost model and the scalarization.
I could only use predicateWithScalar if I was able to compute the cost model, but the cost model was only able to be computed if it was able to be scalarized, and in this case, that was not possible because of sdiv.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5602

Hey Sander,
I may be wrong, but if the loop allows scalar epilogue then it should check if the instruction has a !mayDivideByZero(*I). If it does not allows scalar epilogue then we do not need to check the instruction, because the loop will not scalarize the epilogue.

// If predication is used for this block and the operation would otherwise
// be guarded, then this requires scalarizing.
if (blockNeedsPredication(I->getParent()) || VectorizeWithPredication)
  return !mayDivideByZero(*I);

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5602

If the loop requires no predication, then we know that if the instruction divides by zero then that is an issue caused by the user, who should have avoided this case. If the loop requires predication - either because the user added some if (condition) around the divide, or because the compiler has chosen to fold the tail loop into the vector body and decided to use predication to enable/disable the lanes - then the LV must guarantee the program does not cause different behaviour after vectorizing the code. Because at the moment the LV cannot handle this case yet, it has to fall back on scalarization.

So, if the LV decided to use predication where no predication was needed before and the instruction may divide by zero, then it requires scalarization in order not to change the behaviour of the original program.

The question to ask is "does the loop have instructions that require predication, given that we need predication to handle the tail loop". The "do we need predication to handle the tail loop" part is only true, when the scalar epilogue is not allowed, because then the tail loop is folded into the main vector body.

e.g. 10 iterations with VF=4 without predication (and scalar tail):

1st vector iteration handles 0..3
2nd vector iteration handles 4..7
scalar tail loop handles 8..9

Alternatively, 10 iterations with VF=4 and predication

1st vector iteration handles 0..3, with predicate <true, true, true, true>
2nd vector iteration handles 4..7, with predicate <true, true, true, true>
3rd vector iteration handles 8, 9, xx, xx with predicate <true, true, false, false>

[CarolineConcatto]

• Address review's comment. Only check the loop body, that is when !isScalarEpilogueAllowed()

[CarolineConcatto]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5602

Thank you @sdesmalen.
It took me a while to understand, but if I understood correct, now I agree with you.
The test here should only check instructions in the loop body and not in the tail of the loop. I added a comment in case I pass through here again and ask the same question.
I guess this same check(!Legal->canVectorizeWithPredication) will be needed when checking if the epilogue scalar is allowed.

[david-arm]

LGTM! Looks like you've addressed all the review comments. :)

llvm/test/Transforms/LoopVectorize/scalable-predicate-instruction.ll

6 ↗

(On Diff #346104)

nit: Maybe "This test corresponds to the following function:"?

13 ↗

(On Diff #346104)

nit: Simple typo - it should be guaranteed.

[fhahn]

llvm/test/Transforms/LoopVectorize/scalable-predicate-instruction.ll

13 ↗

(On Diff #346104)

Would it be more accurate to say this can only be vectorized by scalarizing the division for now which cannot be done for scalable vectors at the moment?

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5604

In the previous iteration you changed this condition, but this didn't affect the test. Can you add a test that exercises this change?

llvm/test/Transforms/LoopVectorize/scalable-predicate-instruction.ll

16 ↗

(On Diff #346104)

Instead of checking the output to be exactly this, is it sufficient to just check no vector type is being used?
e.g. CHECK-NOT: <vscale x 4>

[CarolineConcatto]

• add test when the loop does not need predication and epilogue not allowed

[CarolineConcatto]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5604

Thank you @sdesmalen
I added the test as you suggested for when the loop does not need predication and the epilogue is not allowed.
This path is only triggered with the flag:

prefer-predicate-over-epilogue=predicate-dont-vectorize

Otherwise isScalarEpilogueAllowed() is true and the loop can vectorize. That is the reason the second test has

; CHECK:  sdiv <vscale x 4 x i32>

llvm/test/Transforms/LoopVectorize/scalable-predicate-instruction.ll

13 ↗

(On Diff #346104)

Thank you @fhahn,
Is the explanation better now?
I added that it will be possible when llvm.vp is implemented.

16 ↗

(On Diff #346104)

Thank you @sdesmalen,
There is no problem. I have changed to check only SDIV.
Is that better?

[CarolineConcatto]

• Fix and simplify llvm-ir test when the epilogue is not allowed

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

7878

I think it'd be good to give this a more generic name, so that we can use it to test for other cases as well. How about loopCanBeWidenedWithScalableVectors which implies the conditions should be negated, and you may want to return some message of why it did not vectorize, e.g.

LoopVectorizationCostModel::loopCanBeWidenedWithScalableVectors(ElementCount MaxVF, std::string &Message)

llvm/test/Transforms/LoopVectorize/scalable-predicate-instruction.ll

1 ↗

(On Diff #346428)

This is no longer true

17 ↗

(On Diff #346428)

nit: predication_in_loop

63 ↗

(On Diff #346428)

nit: unpredicated_loop_predication_through_tailfolding

90 ↗

(On Diff #346428)

unnecessary?

[david-arm]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

7878

This is similar to https://reviews.llvm.org/D102394 where I moved the canVectorizeReductions call in there too.

[CarolineConcatto]

• Rebase patch bases on D102394 to use canWidenLoopWithScalableVectors
• Address reviewer's comment on the test

[CarolineConcatto]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

7878

Thank you @david-arm and @sdesmalen
In order to not create the same function twice I have rebased this patch based on D102394.
So there is only one canWidenLoopWithScalableVectors.
This patch already has one dependence in D102437.
However, we can remove dependence on D102394, and copy the function if this patch and D102437 are accepted first

[david-arm]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5568–5569

nit: Hi @CarolineConcatto, this is just a suggestion but I think it might be clearer here to say something like:

// Check we are able to vectorize predicated instructions using scalable
// vectors. Such predication may occur if the scalar epilogue is folded into
// the vector loop body.

Not sure what others think?

5572

I wonder if this is potentially misleading as we're not guaranteed to fall back on fixed-width vectorization, particularly if the user applied a C/C++ pragma to the loop?

[CarolineConcatto]

• Address comments

[CarolineConcatto]

Thank you @david-arm for the review.
It is quite difficult to get the comments right, but I think that your comment is better than mine
So I've replaced.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5572

Ok, I removed the part that used to say that it would rely on a fixed-width vector.
Now it has the same message as the others

[CarolineConcatto]

• Rebase patch
• Remove loopHasPredicatedScalar and add test for scalable in collectInstsToScalarize

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

6751–6753

Can you extend the comment here explaining that the discount logic is not applied when the VF is scalable, because that would lead to invalid scalarization costs.

6754–6755

Can you merge these conditions, i.e. if (!VFScalable() && !useEmulatedMaskMemRefHack(&I) && ..

6838

I think this assert can be removed in favour of using VF.getFixedValue() instead of VF.getKnownMinValue() everywhere in this function.

llvm/test/Transforms/LoopVectorize/scalable-predicate-instruction.ll

1 ↗

(On Diff #359826)

The RUN lines don't actually enable scalable vectorization (=off by default), so this test always passes.
You'll need to add -scalable-vectorization=preferred.

I'd also recommend moving this function to llvm/test/Transforms/LoopVectorize/AArch64. The -force-target-supports-scalable-vectors flag doesn't work properly, because the BasicTTIImpl costmodel implementation has defaults that must be overridden for scalable auto-vec to work.

[CarolineConcatto]

• Add scalable-vectorization=on in the llvm-ir test.
• Move llvm-ir test to be in AArch64 folder.
• Remove assertions in computePredInstDiscount and getInstructionCost.
• Replace getKnownMinValue by getFixedValue() in computePredInstDiscount and getInstructionCost.

[sdesmalen]

Thanks for the changes! Just a few little nits left.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

6755

nit: this can just be part of the same expression, i.e.

if (!VF.isScalable() && !useEmulatedMaskMemRefHack(&I) &&
    computePredInstDiscount(&I, ScalarCosts, VF) >= 0)

6830

Can you split out all the changes to use getFixedValue() from this patch and commit those as a separate NFC patch?

llvm/test/Transforms/LoopVectorize/AArch64/scalable-predicate-instruction.ll

2

nit: Can you remove the -mattr=+sve -mtriple aarch64-unknown-linux-gnu and instead encode it in the IR as:

target triple = aarch64-unknown-linux-gnu

define void @predication_in_loop(...) #0 {
  ..
}

attributes #0 = { "target-features"="+sve" }

[CarolineConcatto]

Hey @sdesmalen,
Thank you for pointing out the missing flag on the test.
I have moved the test to be inside the AArch64 folder.
One thing I am not sure is the fix I added in getinstructionCost for scalable vector with BR instruction.
I believe it makes sense to not compute the cost of scalarized predicated instruction when VF is scalable, but not sure if the cost is correct.
Or if it should be invalid.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

6754–6755

Is it ok if I do:

if (!VF.isScalable() && !useEmulatedMaskMemRefHack(&I))
  if (computePredInstDiscount(&I, ScalarCosts, VF) >= 0)

Because computePredInstDiscount should not be called for scalable vector.

[CarolineConcatto]

• Remove the changes from collectInstsToScalarize and getInstructionCost(now in D106359)
• Add attribute in the llvm-ir test

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

7546

The cost of the Br should also be invalid, because it tries to calculate the cost as if it's trying to scalarize the operation if ScalarPredicatedBB == true.

[CarolineConcatto]

• Make cost invalid for predicated instruction when VF is scalable in getInstructionCost for Instruction::Br

[sdesmalen]

LGTM, thanks @CarolineConcatto

Can you please also update the commit message to reflect the new approach and remove the 'Depends on <other patches>' since it no longer depends on either of those?

[CarolineConcatto]

• Update the message with --verbatim

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

public:
   std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
 
   /// \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(ElementCount 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.
   /// The calculated cost is saved with widening decision in order to
   /// avoid redundant calculations.
   void setCostBasedWideningDecision(ElementCount VF);
 
   /// A struct that represents some properties of the register usage
reportVectorizationFailure("Runtime stride check for small trip count",
         "this loop without such check by compiling with -Os/-Oz",
         "CantVersionLoopWithOptForSize", ORE, TheLoop);
     return true;
   }
 
   return false;
 }
 
 ElementCount
 LoopVectorizationCostModel::getMaxLegalScalableVF(unsigned MaxSafeElements) {
   if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) {
     reportVectorizationInfo(
         "Disabling scalable vectorization, because target does not "
         "support scalable vectors.",
         "ScalableVectorsUnsupported", ORE, TheLoop);
     return ElementCount::getScalable(0);
   }
 
   if (Hints->isScalableVectorizationDisabled()) {
     reportVectorizationInfo("Scalable vectorization is explicitly disabled",
                             "ScalableVectorizationDisabled", ORE, TheLoop);
LoopVectorizationCostModel::getMaxLegalScalableVF(unsigned MaxSafeElements) {
   if (!canVectorizeReductions(MaxScalableVF)) {
     reportVectorizationInfo(
         "Scalable vectorization not supported for the reduction "
         "operations found in this loop.",
         "ScalableVFUnfeasible", ORE, TheLoop);
     return ElementCount::getScalable(0);
   }
 
   // Disable scalable vectorization if the loop contains any instructions
   // with element types not supported for scalable vectors.
   if (any_of(ElementTypesInLoop, [&](Type *Ty) {
         return !Ty->isVoidTy() &&
                !this->TTI.isElementTypeLegalForScalableVector(Ty);
       })) {
     reportVectorizationInfo("Scalable vectorization is not supported "
                             "for all element types found in this loop.",
                             "ScalableVFUnfeasible", ORE, TheLoop);
     return ElementCount::getScalable(0);
   }
void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) {
   // determine if it would be better to not if-convert the blocks they are in.
   // If so, we also record the instructions to scalarize.
   for (BasicBlock *BB : TheLoop->blocks()) {
     if (!blockNeedsPredication(BB))
       continue;
     for (Instruction &I : *BB)
       if (isScalarWithPredication(&I)) {
         ScalarCostsTy ScalarCosts;
+        // Do not apply discount if scalable, because that would lead to
+        // invalid scalarization costs.
         // Do not apply discount logic if hacked cost is needed
         // for emulated masked memrefs.
-        if (!useEmulatedMaskMemRefHack(&I) &&
+        if (!VF.isScalable() && !useEmulatedMaskMemRefHack(&I) &&
             computePredInstDiscount(&I, ScalarCosts, VF) >= 0)
           ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end());
         // Remember that BB will remain after vectorization.
         PredicatedBBsAfterVectorization.insert(BB);
       }
   }
 }
 
 int LoopVectorizationCostModel::computePredInstDiscount(
     Instruction *PredInst, ScalarCostsTy &ScalarCosts, ElementCount VF) {
   assert(!isUniformAfterVectorization(PredInst, VF) &&
while (!Worklist.empty()) {
     // includes the scalarization overhead of the predicated instruction.
     InstructionCost VectorCost = getInstructionCost(I, VF).first;
 
     // Compute the cost of the scalarized instruction. This cost is the cost of
     // the instruction as if it wasn't if-converted and instead remained in the
     // predicated block. We will scale this cost by block probability after
     // computing the scalarization overhead.
     InstructionCost ScalarCost =
         VF.getFixedValue() *
         getInstructionCost(I, ElementCount::getFixed(1)).first;
 
     // Compute the scalarization overhead of needed insertelement instructions
     // and phi nodes.
     if (isScalarWithPredication(I) && !I->getType()->isVoidTy()) {
       ScalarCost += TTI.getScalarizationOverhead(
           cast<VectorType>(ToVectorTy(I->getType(), VF)),
           APInt::getAllOnesValue(VF.getFixedValue()), true, false);
       ScalarCost +=
           VF.getFixedValue() *
           TTI.getCFInstrCost(Instruction::PHI, TTI::TCK_RecipThroughput);
     }
 
     // Compute the scalarization overhead of needed extractelement
     // instructions. For each of the instruction's operands, if the operand can
     // be scalarized, add it to the worklist; otherwise, account for the
case Instruction::GetElementPtr:
     // instruction cost.
     return 0;
   case Instruction::Br: {
     // In cases of scalarized and predicated instructions, there will be VF
     // predicated blocks in the vectorized loop. Each branch around these
     // blocks requires also an extract of its vector compare i1 element.
     bool ScalarPredicatedBB = false;
     BranchInst *BI = cast<BranchInst>(I);
     if (VF.isVector() && BI->isConditional() &&
         (PredicatedBBsAfterVectorization.count(BI->getSuccessor(0)) ||
          PredicatedBBsAfterVectorization.count(BI->getSuccessor(1))))
       ScalarPredicatedBB = true;
 
     if (ScalarPredicatedBB) {
+      // Not possible to scalarize scalable vector with predicated instructions.
+      if (VF.isScalable())
+        return InstructionCost::getInvalid();
       // Return cost for branches around scalarized and predicated blocks.
       auto *Vec_i1Ty =
           VectorType::get(IntegerType::getInt1Ty(RetTy->getContext()), VF);
       return (
           TTI.getScalarizationOverhead(
               Vec_i1Ty, APInt::getAllOnesValue(VF.getFixedValue()), false,
               true) +
           (TTI.getCFInstrCost(Instruction::Br, CostKind) * VF.getFixedValue()));

 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
   // Check if the pointer operand of a load or store instruction is
   // consecutive.
   if (auto *Ptr = getLoadStorePointerOperand(Inst))
     return Legal->isConsecutivePtr(Ptr);
   return false;
 }
 
 void LoopVectorizationCostModel::collectValuesToIgnore() {
   // Ignore ephemeral values.
   CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore);
 
   // Ignore type-promoting instructions we identified during reduction
   // detection.
   for (auto &Reduction : Legal->getReductionVars()) {
     RecurrenceDescriptor &RedDes = Reduction.second;
     const SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();
     VecValuesToIgnore.insert(Casts.begin(), Casts.end());
   }
   // Ignore type-casting instructions we identified during induction
   // detection.
   for (auto &Induction : Legal->getInductionVars()) {
     InductionDescriptor &IndDes = Induction.second;
     const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts();

llvm/test/Transforms/LoopVectorize/AArch64/scalable-predicate-instruction.ll

+; RUN: opt < %s -loop-vectorize -scalable-vectorization=on -S | FileCheck %s
+; RUN: opt < %s -loop-vectorize -scalable-vectorization=on -prefer-predicate-over-epilogue=predicate-dont-vectorize -S | FileCheck %s
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64-unknown-linux-gnu"
+
+; The test predication_in_loop corresponds
+; to the following function
+;   for (long long i = 0; i < 1024; i++) {
+;    if (cond[i])
+;      a[i] /= b[i];
+;  }
+
+; Scalarizing the division cannot be done for scalable vectors at the moment
+; when the loop needs predication
+; Future implementation of llvm.vp could allow this to happen
+
+define void  @predication_in_loop(i32* %a, i32* %b, i32* %cond) #0 {
+; CHECK-LABEL: @predication_in_loop
+; CHECK-NOT:  sdiv <vscale x 4 x i32>
+;
+entry:
+  br label %for.body
+
+for.cond.cleanup:                                 ; preds = %for.inc, %entry
+  ret void
+
+for.body:                                         ; preds = %entry, %for.inc
+  %i.09 = phi i64 [ %inc, %for.inc ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i32, i32* %cond, i64 %i.09
+  %0 = load i32, i32* %arrayidx, align 4
+  %tobool.not = icmp eq i32 %0, 0
+  br i1 %tobool.not, label %for.inc, label %if.then
+
+if.then:                                          ; preds = %for.body
+  %arrayidx1 = getelementptr inbounds i32, i32* %b, i64 %i.09
+  %1 = load i32, i32* %arrayidx1, align 4
+  %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %i.09
+  %2 = load i32, i32* %arrayidx2, align 4
+  %div = sdiv i32 %2, %1
+  store i32 %div, i32* %arrayidx2, align 4
+  br label %for.inc
+
+for.inc:                                          ; preds = %for.body, %if.then
+  %inc = add nuw nsw i64 %i.09, 1
+  %exitcond.not = icmp eq i64 %inc, 1024
+  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !0
+}
+
+
+;
+; This test unpredicated_loop_predication_through_tailfolding corresponds
+; to the following function
+;   for (long long i = 0; i < 1024; i++) {
+;      a[i] /= b[i];
+;  }
+
+; Scalarization not possible in the main loop when there is no predication and
+; epilogue should not be able to allow scalarization
+; otherwise it could  be able to vectorize, but will not because
+; "Max legal vector width too small, scalable vectorization unfeasible.."
+
+define void @unpredicated_loop_predication_through_tailfolding(i32* %a, i32* %b) #0 {
+; CHECK-LABEL: @unpredicated_loop_predication_through_tailfolding
+; CHECK-NOT:  sdiv <vscale x 4 x i32>
+
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+  %arrayidx = getelementptr inbounds i32, i32* %a, i64 %iv
+  %0 = load i32, i32* %arrayidx, align 4
+  %arrayidx2 = getelementptr inbounds i32, i32* %b, i64 %iv
+  %1 = load i32, i32* %arrayidx2, align 4
+  %sdiv = sdiv i32 %1, %0
+  %2 = add nuw nsw i64 %iv, 8
+  %arrayidx5 = getelementptr inbounds i32, i32* %a, i64 %2
+  store i32 %sdiv, i32* %arrayidx5, align 4
+  %iv.next = add nuw nsw i64 %iv, 1
+  %exitcond.not = icmp eq i64 %iv.next, 1024
+  br i1 %exitcond.not, label %exit, label %loop, !llvm.loop !0
+
+exit:
+  ret void
+
+}
+
+attributes #0 = { "target-features"="+sve" }
+
+!0 = distinct !{!0, !1, !2, !3, !4}
+!1 = !{!"llvm.loop.vectorize.width", i32 4}
+!2 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
+!3 = !{!"llvm.loop.interleave.count", i32 1}
+!4 = !{!"llvm.loop.vectorize.enable", i1 true}