[LV] Legalize scalable VF hints

[c-rhodes]

Summary

In the following loop:

void foo(int *a, int *b, int N) {
  for (int i=0; i<N; ++i)
    a[i + 4] = a[i] + b[i];
}

The loop dependence constrains the VF to a maximum of (4, fixed), which
would mean using <4 x i32> as the vector type in vectorization.
Extending this to scalable vectorization, a VF of (4, scalable) implies
a vector type of <vscale x 4 x i32>. To determine if this is legal
vscale must be taken into account. For this example, unless
max(vscale)=1, it's unsafe to vectorize.

For SVE, the number of bits in an SVE register is architecturally
defined to be a multiple of 128 bits with a maximum of 2048 bits, thus
the maximum vscale is 16. In the loop above it is therefore unfeasible
to vectorize with SVE. However, in this loop:

void foo(int *a, int *b, int N) {
  #pragma clang loop vectorize_width(X, scalable)
  for (int i=0; i<N; ++i)
    a[i + 32] = a[i] + b[i];
}

As long as max(vscale) multiplied by the number of lanes 'X' doesn't
exceed the dependence distance, it is safe to vectorize. For SVE a VF of
(2, scalable) is within this constraint, since a vector of <16 x 2 x 32>
will have no dependencies between lanes. For any number of lanes larger
than this it would be unsafe to vectorize.

This patch extends 'computeFeasibleMaxVF' to legalize scalable VFs
specified as loop hints, implementing the following behaviour:

• If the backend does not support scalable vectors, ignore the hint.
• If scalable vectorization is unfeasible given the loop dependence, like in the first example above for SVE, then use a fixed VF.
• Accept scalable VFs if it's safe to do so.
• Otherwise, clamp scalable VFs that exceed the maximum safe VF.

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5580

nit: MaxSafeElements?

5582

Instead of returning an Optional<ElementCount>, I'd prefer the code here to just clamp to MaxFixedVF,
in this order:

• first try to see if it can use the requested scalable VF.
• if not, then try to see if it can use the requested VF with vscale = 1 (i.e. fixed width)
• if not, then try if it can use a clamped fixed VF.

[fhahn]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5577

This check seems a bit odd. I think we should at least use a named constant to make things clearer and ensure that LAA & LV are kept in sync on the meaning. But would it be possible to instead compute the maximum width of UserVF using MaxVScale (something like UserVF * WidestType * MaxVScale)?

[c-rhodes]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5577

This check seems a bit odd. I think we should at least use a named constant to make things clearer and ensure that LAA & LV are kept in sync on the meaning. But would it be possible to instead compute the maximum width of UserVF using MaxVScale (something like UserVF * WidestType * MaxVScale)?

I don't see how this is related, the reason I added this check is this block is validating a user VF and clamping it when it exceeds what is safe in terms of dependencies, but if there are no dependencies there's nothing to do here. I'm not sure if there's a better way of querying if there are no dependencies? MaxSafeRegisterWidth is initialized as -1U in LAA so I compared it against UINT_MAX, maybe it would be a little clearer if MaxSafeRegisterWidth was initialized as UINT_MAX. It would probably have made sense to introduced this in D90687 although I didn't consider it, it became obvious with this patch since large max safe VFs for loops with no dependencies were being emitted.

5580

nit: MaxSafeElements?

Cheers that makes more sense, will update it

5582

Instead of returning an Optional<ElementCount>, I'd prefer the code here to just clamp to MaxFixedVF,
in this order:

• first try to see if it can use the requested scalable VF.
• if not, then try to see if it can use the requested VF with vscale = 1 (i.e. fixed width)
• if not, then try if it can use a clamped fixed VF.

Deferring to fixed-width so Optional isn't necessary is a good point, the first 2 cases should already be handled, I'll add the latter

[c-rhodes]

Changes:

• s/FixedVF/MaxSafeElements/
• Rather than bail out, clamp to fixed VF is scalable is unfeasible. Remove change to return Optional<ElementCount> from LoopVectorizationCostModel::computeFeasibleMaxVF.
• In LoopVectorizationPlanner::plan, if a UserVF is specified but isn't legal then use MaxVF (earlier).
• Simplified the IR in the tests.

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5595

This does something different than what I suggested though.

For the following example, your code would set the MaxSafeVF to 32 elements instead of using a fixed VF of 4 elements (fixed):

#pragma clang loop vectorize_width(4, scalable)
for (int i=0; i<N; ++i)
  a[i + 32] = a[i] + 1;   // for some int* a

I think a VF of 4 would be preferred, because otherwise selectiondag will need to expand <32 x i32> to 8 x <4 x i32> fixed-width vectors (assuming 128bit vectors), which would be somewhat similar to requesting an interleave-count of 8.

llvm/test/Transforms/LoopVectorize/scalable-vf-hint.ll

21

At least until we add lowering of scalable vectors for targets that don't support it, I would rather the code explicitly checks if the target can handle them and ignore the hint if it doesn't (a similar suggestion was made on D88962).

[c-rhodes]

Changes:

• Rebase since D88962 and D91077 have now landed.
• Address @sdesmalen's comments. Added a target hook supportsScalableVectors in D93060 which is used in this patch to bail out of vectorization if unsupported. Also changed the behavior when scalable vectorization is unfeasible given a dependency, it now changes the hint to fixed, where existing behavior applies (i.e. clamping).
• Updated tests.

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5574

MaxSafeVectorWidthInBits can't exceed UINT_MAX. If this this early bail out isn't strictly necessary, I'd suggest removing it.

5579–5585

Instead of returning 'None' (and thus not allowing any vectorziation), isn't it better to ignore the UserVF and continuing by allowing a VF to be chosen as normal?

i.e.

bool IgnoreUserVF = UserVF.isScalable() && TTI.supportsScalableVectors();
if (IgnoreUserVF)
  ORE->emit([&]() {
     return OptimizationRemark(....) };

if (UserVF.isNonZero() && !IgnoreUserVF) {
  ...
}

// otherwise, let the LoopVectorizer choose a VF itself.

I guess that can be split out into D93060 as well.

5604

Should this be an OptRemark instead?

5609

Can it know that ElementCount::getFixed(UserVF.getKnownMinValue()) would be a valid VF? (that check is below on line 5614).
Probably good to have a test for that as well.

7433

We still want to clamp to a smaller scalable VF if UserVF.isScalable() && MaxVF.isScalable() && UserVF > MaxVF. Because the loop below doesn't support scalable VFs, it is best to single that case out.

That means:

1. UserVF = Scalable && MaxVF = Scalable && UserVF > MaxVF => Pick MaxVF
2. UserVF = Scalable && MaxVF = Fixed => Fall through into loop below.
3. UserVF = Fixed && MaxVF = Fixed && UserVF > MaxVF => Fall through into loop below.

Where the clamping in 1. is temporary until the loop below works on scalable vectors, so that it can choose the most profitable VF based on the cost-model.

llvm/test/Transforms/LoopVectorize/AArch64/scalable-vf-hint.ll

1

Why does it require asserts?

[fhahn]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5577

My point was that the original code handled the case where there are no dependencies naturally without an extra check, right? Is there a reason the logic for scalable vectors cannot do the same?

5579–5585

Agreed, if the user requests scalable vectorization through a hint, I probably should just drop the hint and proceed with normal cost-modeling.

(if there's a reason to keep the bail-out, can we just return a VF of 1? this is how other places in the function indicate that vectorization should be avoided)

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5609

It seems I misread this line/statement, so please ignore this comment.

[c-rhodes]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5574

MaxSafeVectorWidthInBits can't exceed UINT_MAX. If this this early bail out isn't strictly necessary, I'd suggest removing it.

There's a couple of existing tests:

• Transforms/LoopVectorize/scalable-loop-unpredicated-body-scalar-tail.ll
• Transforms/LoopVectorize/metadata-width.ll

that will fall over if the hint is ignored when scalable vectorization isn't supported. I suppose those could become target-specific tests.

5579–5585

Instead of returning 'None' (and thus not allowing any vectorziation), isn't it better to ignore the UserVF and continuing by allowing a VF to be chosen as normal?

...

I guess that can be split out into D93060 as well.

Ok np, I'll split that out then.

5604

Should this be an OptRemark instead?

Good point, I'll add that

7433

Ok, I'll add a temporary workaround for scalable VFs to select MaxVF above if UserVF is unsafe.

llvm/test/Transforms/LoopVectorize/AArch64/scalable-vf-hint.ll

1

Why does it require asserts?

Because of the debug flags

[c-rhodes]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5577

My point was that the original code handled the case where there are no dependencies naturally without an extra check, right? Is there a reason the logic for scalable vectors cannot do the same?

Sorry I missed your comment. One issue is for non-target specific tests with scalable VFs such as:

• Transforms/LoopVectorize/scalable-loop-unpredicated-body-scalar-tail.ll
• Transforms/LoopVectorize/metadata-width.ll

the hint will be ignored since scalable vectorization isn't supported, and also getMaxVScale now returns None. I suppose there's a couple of options, we could add new loop vectorizer flags or those tests could become become target-specific, I'm not sure what the best option is.

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5574

If the test has a loop-carried distance that may invalidate auto-vectorization that needs checking, I think it's fair enough for the test to be moved to be a target-specific test (because that will require getMaxVScale).

For a test like metadata-width.ll it may be useful to add some -force-target-supports-scalable-vectors=true flag to avoid making all scalable-vector tests for the LoopVectorizer target-specific.

@fhahn do you have any strong feelings about this?

[c-rhodes]

• Rather than disable vectorization if scalable VF isn't supported, let the LV pick a VF as it normally does and process with cost-modelling.
• Added a new LV flag -force-target-supports-scalable-vectors to support writing non-target specific tests with scalable VFs.
• Added a temporary workaround to use MaxVF in place of UserVF until cost modelling loop works for scalable VFs.
• Emit opt remark.

[c-rhodes]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5579–5585

Instead of returning 'None' (and thus not allowing any vectorziation), isn't it better to ignore the UserVF and continuing by allowing a VF to be chosen as normal?

...

I guess that can be split out into D93060 as well.

Ok np, I'll split that out then.

Done, although I've kept it as part of this patch. Ignoring the hint if scalable vectors aren't supported breaks existing tests so I've implemented the flag you suggested -force-target-supports-scalable-vectors which can be used for non-target specific tests. Although this only works for loops with no dependencies since getMaxVScale is None by default, so I've moved scalable-loop-unpredicated-body-scalar-tail.ll to AArch64 and enabled SVE.

[david-arm]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

7424

Is this additional check necessary here? It looks like MaxVF came from computeMaXVF, which makes use of your clamping code above. If you revert this change do your tests still pass?

[david-arm]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

7424

Ah I think I misunderstood what the clamping above is doing. It only clamps MaxVF, not UserVF. I wonder if it's worth emitting a remark for this case to let the user know something happened? It might look a bit odd if we downgrade their VF without telling them. I guess it depends upon how quick we get a fix for this.

[c-rhodes]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

7424

Is this additional check necessary here?

Until the loop below supports scalable VFs yes, with this check we always fall into this block for a scalable VF, this handles the case where the UserVF isn't legal but a scalable VF is feasible so it clamps to the scalable MaxVF. Otherwise the assert below would be triggered, this is changed from the previous revision to implement @sdesmalen suggestion of: UserVF = Scalable && MaxVF = Scalable && UserVF > MaxVF => Pick MaxVF, rather than change MaxVF from scalable -> fixed.

[c-rhodes]

Handle scalable VF in a single isScalable block in computeFeasibleMaxVF.

[david-arm]

LGTM. I agree the FIXME in the code isn't ideal, but I understand why it's there and is only a temporary measure until we can fix the loop to support scalable vectors.

llvm/test/Transforms/LoopVectorize/AArch64/scalable-loop-unpredicated-body-scalar-tail.ll

1 ↗

(On Diff #312451)

nit: Perhaps it's better to run llvm/utils/update_llc_test_checks.py on these tests rather than hand-write the CHECK lines? This seems to be the accepted way of doing it in LLVM these days. :)

[fhahn]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

278

It would be good to be clear in the message that this is just for testing (same as the previous option). Perhaps something like Pretend that scalable vectors are supported, even if the target does not support them. The flag should only be used for testing.

5573–5574

This is just about ignoring scalable UserVFs, right? Might be good to make that clear as part of the variable name.

5574

Yes that sounds reasonable to me.

5577

Ah right, if getMaxVScale can return None, then my suggestion doesn't work as expected. In that case, I think it would be good to encapsulate the check in LAA and have something like areAllVectorWidthsSafe or something. This should be more robust to future changes in LAA.

[c-rhodes]

Address @fhahn's comments, changes:

• Clarify help message in -force-target-supports-scalable-vectors flag.
• s/IgnoreUserVF/IgnoreScalableUserVF/g
• Added isSafeForAnyVectorWidth to LAA that checks if the number of elements that can be operated on simultaneously is not bound. This is used in computeFeasibleMaxVF to skip the checks for whether the VF is safe given the dependence distance.

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5583–5584

It should be fine for MaxVScale to be undefined for a target architecture. Based on the logic below where it tries to pick a suitable alternative VF, it should probably pick a different (non-scalable) VF, instead of crashing.

[c-rhodes]

Changes:

• Rebased. Target hook for max vscale has been removed from this patch since it landed in D93030.
• Address Sander's comment w.r.t. falling back to fixed vectorization if scalable vectors are supported, but max vscale is undefined.

[fhahn]

LGTM, thanks. Some small comments left inline.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5573–5574

This early exit could also be on the top of the function, before we compute the register widths & co and is not specific to scalable vectors, right? Might be good to move, although the 'scalable ignored' message might also needs to move.

llvm/test/Transforms/LoopVectorize/AArch64/scalable-loop-unpredicated-body-scalar-tail.ll

100 ↗

(On Diff #314648)

-force-vector-width should also work, right? In that case, it would be good to use that at least for some tests.

Also, could this be kept target-independent with -force-target-supports-scalable-vectors=true?

llvm/test/Transforms/LoopVectorize/optimal-epilog-vectorization-limitations.ll

2 ↗

(On Diff #314648)

is that necessary? I think it would be better to just add a separate test specifically checking that epilogue vectorization is disabled with scalable vectors.

[c-rhodes]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5573–5574

This early exit could also be on the top of the function, before we compute the register widths & co and is not specific to scalable vectors, right? Might be good to move, although the 'scalable ignored' message might also needs to move.

That's right it's not specific to scalable vectors, as long as User.isNonZero() it applies. I think you're right, if this is moved to the top IgnoreScalableUserVF will need to be part of the check, so:

// Nothing to do if there are no dependencies.
if (UserVF.isNonZero() && !IgnoreScalableUserVF &&
    Legal->isSafeForAnyVectorWidth())
  return UserVF

Otherwise scalable VFs will be accepted for targets without scalable vector support, where it currently falls into the existing code to pick a VF. It's fine with me to move this to the top with that in mind.

llvm/test/Transforms/LoopVectorize/AArch64/scalable-loop-unpredicated-body-scalar-tail.ll

100 ↗

(On Diff #314648)

-force-vector-width should also work, right? In that case, it would be good to use that at least for some tests.

Yeah good idea, not sure we test the combination of that flag and the metadata, I'll change that.

Also, could this be kept target-independent with -force-target-supports-scalable-vectors=true?

It can for this test since there's no dependencies so max vscale isn't required to determine if the scalable VF is valid. I'll implement your suggestion and keep this target-indepedent.

llvm/test/Transforms/LoopVectorize/optimal-epilog-vectorization-limitations.ll

2 ↗

(On Diff #314648)

is that necessary? I think it would be better to just add a separate test specifically checking that epilogue vectorization is disabled with scalable vectors.

The final test here checks epilogue vectorization is disabled for scalable vectors, I can split that out but it'll still require this flag to target-independent. Or it could become an AArch64 test with SVE.

[c-rhodes]

Address @fhahn comments.

[c-rhodes]

@fhahn thanks for reviewing, I think I've addressed all comments. I'll probably land this over the coming days unless there's any objections between then, cheers.

llvm/test/Transforms/LoopVectorize/optimal-epilog-vectorization-limitations.ll

2 ↗

(On Diff #314648)

is that necessary? I think it would be better to just add a separate test specifically checking that epilogue vectorization is disabled with scalable vectors.

The final test here checks epilogue vectorization is disabled for scalable vectors, I can split that out but it'll still require this flag to target-independent. Or it could become an AArch64 test with SVE.

I've split this into a separate test: llvm/test/Transforms/LoopVectorize/optimal-epilog-vectorization-scalable.ll.

[sdesmalen]

LGTM, thanks for your work on this patch @c-rhodes!

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5599

nit: is it useful to print MaxSafeElements or MaxSafeVectorWidthInBits here?

[c-rhodes]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

5599

nit: is it useful to print MaxSafeElements or MaxSafeVectorWidthInBits here?

MaxSafeElements is undefined here but LAA emits the max VF anyway which should be sufficient?

llvm/include/llvm/Analysis/TargetTransformInfo.h

public:
   const char *getRegisterClassName(unsigned ClassID) const;
 
   /// \return The width of the largest scalar or vector register type.
   unsigned getRegisterBitWidth(bool Vector) const;
 
   /// \return The width of the smallest vector register type.
   unsigned getMinVectorRegisterBitWidth() const;
 
+  /// \return The maximum value for vscale in scalable vectors such as
+  /// <vscale x 4 x i32>. Default is None.
+  Optional<unsigned> getMaxVScale() const;
+
   /// \return True if the vectorization factor should be chosen to
   /// make the vector of the smallest element type match the size of a
   /// vector register. For wider element types, this could result in
   /// creating vectors that span multiple vector registers.
   /// If false, the vectorization factor will be chosen based on the
   /// size of the widest element type.
   bool shouldMaximizeVectorBandwidth(bool OptSize) const;
 
virtual int getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
                                   const APInt &Imm, Type *Ty,
                                   TargetCostKind CostKind) = 0;
   virtual unsigned getNumberOfRegisters(unsigned ClassID) const = 0;
   virtual unsigned getRegisterClassForType(bool Vector,
                                            Type *Ty = nullptr) const = 0;
   virtual const char *getRegisterClassName(unsigned ClassID) const = 0;
   virtual unsigned getRegisterBitWidth(bool Vector) const = 0;
   virtual unsigned getMinVectorRegisterBitWidth() = 0;
+  virtual Optional<unsigned> getMaxVScale() const = 0;
   virtual bool shouldMaximizeVectorBandwidth(bool OptSize) const = 0;
   virtual unsigned getMinimumVF(unsigned ElemWidth) const = 0;
   virtual bool shouldConsiderAddressTypePromotion(
       const Instruction &I, bool &AllowPromotionWithoutCommonHeader) = 0;
   virtual unsigned getCacheLineSize() const = 0;
   virtual Optional<unsigned> getCacheSize(CacheLevel Level) const = 0;
   virtual Optional<unsigned> getCacheAssociativity(CacheLevel Level) const = 0;
 
const char *getRegisterClassName(unsigned ClassID) const override {
     return Impl.getRegisterClassName(ClassID);
   }
   unsigned getRegisterBitWidth(bool Vector) const override {
     return Impl.getRegisterBitWidth(Vector);
   }
   unsigned getMinVectorRegisterBitWidth() override {
     return Impl.getMinVectorRegisterBitWidth();
   }
+  Optional<unsigned> getMaxVScale() const override {
+    return Impl.getMaxVScale();
+  }
   bool shouldMaximizeVectorBandwidth(bool OptSize) const override {
     return Impl.shouldMaximizeVectorBandwidth(OptSize);
   }
   unsigned getMinimumVF(unsigned ElemWidth) const override {
     return Impl.getMinimumVF(ElemWidth);
   }
   bool shouldConsiderAddressTypePromotion(
       const Instruction &I, bool &AllowPromotionWithoutCommonHeader) override {

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

case 1:
       return "Generic::VectorRC";
     }
   }
 
   unsigned getRegisterBitWidth(bool Vector) const { return 32; }
 
   unsigned getMinVectorRegisterBitWidth() { return 128; }
 
+  llvm::Optional<unsigned> getMaxVScale() const { return None; }
+
   bool shouldMaximizeVectorBandwidth(bool OptSize) const { return false; }
 
   unsigned getMinimumVF(unsigned ElemWidth) const { return 0; }
 
   bool
   shouldConsiderAddressTypePromotion(const Instruction &I,
                                      bool &AllowPromotionWithoutCommonHeader) {
     AllowPromotionWithoutCommonHeader = false;

llvm/include/llvm/CodeGen/BasicTTIImpl.h

public:
 
   /// @}
 
   /// \name Vector TTI Implementations
   /// @{
 
   unsigned getRegisterBitWidth(bool Vector) const { return 32; }
 
+  Optional<unsigned> getMaxVScale() const { return None; }
+
   /// Estimate the overhead of scalarizing an instruction. Insert and Extract
   /// are set if the demanded result elements need to be inserted and/or
   /// extracted from vectors.
   unsigned getScalarizationOverhead(VectorType *InTy, const APInt &DemandedElts,
                                     bool Insert, bool Extract) {
     /// FIXME: a bitfield is not a reasonable abstraction for talking about
     /// which elements are needed from a scalable vector
     auto *Ty = cast<FixedVectorType>(InTy);

llvm/lib/Analysis/TargetTransformInfo.cpp

 unsigned TargetTransformInfo::getRegisterBitWidth(bool Vector) const {
   return TTIImpl->getRegisterBitWidth(Vector);
 }
 
 unsigned TargetTransformInfo::getMinVectorRegisterBitWidth() const {
   return TTIImpl->getMinVectorRegisterBitWidth();
 }
 
+llvm::Optional<unsigned> TargetTransformInfo::getMaxVScale() const {
+  return TTIImpl->getMaxVScale();
+}
+
 bool TargetTransformInfo::shouldMaximizeVectorBandwidth(bool OptSize) const {
   return TTIImpl->shouldMaximizeVectorBandwidth(OptSize);
 }
 
 unsigned TargetTransformInfo::getMinimumVF(unsigned ElemWidth) const {
   return TTIImpl->getMinimumVF(ElemWidth);
 }
 

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.h

unsigned getRegisterBitWidth(bool Vector) const {
     }
     return 64;
   }
 
   unsigned getMinVectorRegisterBitWidth() {
     return ST->getMinVectorRegisterBitWidth();
   }
 
+  Optional<unsigned> getMaxVScale() const {
+    if (ST->hasSVE())
+      return AArch64::SVEMaxBitsPerVector / AArch64::SVEBitsPerBlock;
+    return BaseT::getMaxVScale();
+  }
+
   unsigned getMaxInterleaveFactor(unsigned VF);
 
   int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
                        TTI::CastContextHint CCH, TTI::TargetCostKind CostKind,
                        const Instruction *I = nullptr);
 
   int getExtractWithExtendCost(unsigned Opcode, Type *Dst, VectorType *VecTy,
                                unsigned Index);

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

static cl::opt<unsigned> ForceTargetInstructionCost(
     "force-target-instruction-cost", cl::init(0), cl::Hidden,
     cl::desc("A flag that overrides the target's expected cost for "
              "an instruction to a single constant value. Mostly "
              "useful for getting consistent testing."));
 
 static cl::opt<unsigned> SmallLoopCost(
     "small-loop-cost", cl::init(20), cl::Hidden,
     cl::desc(
         "The cost of a loop that is considered 'small' by the interleaver."));
 
 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
     "loop-vectorize-with-block-frequency", cl::init(true), cl::Hidden,
     cl::desc("Enable the use of the block frequency analysis to access PGO "
              "heuristics minimizing code growth in cold regions and being more "
              "aggressive in hot regions."));
 
 // Runtime interleave loops for load/store throughput.
public:
   }
 
 private:
   unsigned NumPredStores = 0;
 
   /// \return An upper bound for the vectorization factor, a power-of-2 larger
   /// than zero. One is returned if vectorization should best be avoided due
   /// to cost.
-  ElementCount computeFeasibleMaxVF(unsigned ConstTripCount,
+  Optional<ElementCount> computeFeasibleMaxVF(unsigned ConstTripCount,
                                               ElementCount UserVF);
 
   /// 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).
LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) {
   LLVM_DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
   if (TC == 1) {
     reportVectorizationFailure("Single iteration (non) loop",
         "loop trip count is one, irrelevant for vectorization",
         "SingleIterationLoop", ORE, TheLoop);
     return None;
   }
 
-  ElementCount MaxVF = computeFeasibleMaxVF(TC, UserVF);
+  Optional<ElementCount> MaybeMaxVF = computeFeasibleMaxVF(TC, UserVF);
+  if (!MaybeMaxVF)
+    return None;
+
+  ElementCount MaxVF = MaybeMaxVF.getValue();
 
   switch (ScalarEpilogueStatus) {
   case CM_ScalarEpilogueAllowed:
     return MaxVF;
   case CM_ScalarEpilogueNotNeededUsePredicate:
     LLVM_DEBUG(
         dbgs() << "LV: vector predicate hint/switch found.\n"
                << "LV: Not allowing scalar epilogue, creating predicated "
reportVectorizationFailure(
       "Cannot optimize for size and vectorize at the same time.",
       "cannot optimize for size and vectorize at the same time. "
       "Enable vectorization of this loop with '#pragma clang loop "
       "vectorize(enable)' when compiling with -Os/-Oz",
       "NoTailLoopWithOptForSize", ORE, TheLoop);
   return None;
 }
 
-ElementCount
+Optional<ElementCount>
 LoopVectorizationCostModel::computeFeasibleMaxVF(unsigned ConstTripCount,
                                                  ElementCount UserVF) {
   MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);
   unsigned SmallestType, WidestType;
   std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes();
   unsigned WidestRegister = TTI.getRegisterBitWidth(true);
 
   // Get the maximum safe dependence distance in bits computed by LAA.
   // It is computed by MaxVF * sizeOf(type) * 8, where type is taken from
   // the memory accesses that is most restrictive (involved in the smallest
   // dependence distance).
   unsigned MaxSafeVectorWidthInBits = Legal->getMaxSafeVectorWidthInBits();
 
+  // If the user vectorization factor is legally unsafe, clamp it to a safe
+  // value. Otherwise, return as is.
   if (UserVF.isNonZero()) {
-    // For now, don't verify legality of scalable vectors.
-    // This will be addressed properly in https://reviews.llvm.org/D91718.
-    if (UserVF.isScalable())
+    // Nothing to do if there are no dependencies.
+    if (MaxSafeVectorWidthInBits >= UINT_MAX)
       return UserVF;
 
-    // If legally unsafe, clamp the user vectorization factor to a safe value.
-    unsigned MaxSafeVF = PowerOf2Floor(MaxSafeVectorWidthInBits / WidestType);
-    if (UserVF.getFixedValue() <= MaxSafeVF)
+    unsigned MaxVScale;
+    if (UserVF.isScalable()) {
+      if (!TTI.supportsScalableVectors()) {
+        reportVectorizationFailure(
+            "Ignoring scalable VF because target does not support scalable vectors",
+            "Ignoring scalable VF because target does not support scalable vectors.",
+            "NoScalableVectorSupport", ORE, TheLoop);
+        return None;
+      }
+
+      Optional<unsigned> MaybeMaxVScale = TTI.getMaxVScale();
+      assert(MaybeMaxVScale &&
+             "max vscale undefined for target that supports scalable vectors!");
+      MaxVScale = MaybeMaxVScale.getValue();
+    }
+
+    unsigned MaxSafeElements =
+        PowerOf2Floor(MaxSafeVectorWidthInBits / WidestType);
+    // If scalable, scale VF by vscale before checking if it's safe.
+    ElementCount MaxSafeVF =
+        UserVF.isScalable()
+            ? ElementCount::getScalable(MaxSafeElements / MaxVScale)
+            : ElementCount::getFixed(MaxSafeElements);
+
+    if (UserVF.isScalable() && MaxSafeVF.isZero()) {
+      // Dependence distance too small to use scalable vectors, use fixed VF
+      // instead.
+      LLVM_DEBUG(
+          dbgs()
+          << "LV: Max legal vector width too small, scalable vectorization "
+             "unfeasible. Using fixed-width vectorization instead.\n");
+      return computeFeasibleMaxVF(
+          ConstTripCount, ElementCount::getFixed(UserVF.getKnownMinValue()));
+    }
+
+    LLVM_DEBUG(dbgs() << "LV: The max safe VF is: " << MaxSafeVF << ".\n");
+
+    if (ElementCount::isKnownLE(UserVF, MaxSafeVF))
       return UserVF;
 
     LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVF
                       << " is unsafe, clamping to max safe VF=" << MaxSafeVF
                       << ".\n");
     ORE->emit([&]() {
       return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationFactor",
                                         TheLoop->getStartLoc(),
                                         TheLoop->getHeader())
              << "User-specified vectorization factor "
              << ore::NV("UserVectorizationFactor", UserVF)
              << " is unsafe, clamping to maximum safe vectorization factor "
              << ore::NV("VectorizationFactor", MaxSafeVF);
     });
-    return ElementCount::getFixed(MaxSafeVF);
+    return MaxSafeVF;
   }
 
   WidestRegister = std::min(WidestRegister, MaxSafeVectorWidthInBits);
 
   // Ensure MaxVF is a power of 2; the dependence distance bound may not be.
   // Note that both WidestRegister and WidestType may not be a powers of 2.
   unsigned MaxVectorSize = PowerOf2Floor(WidestRegister / WidestType);
 
LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) {
 
   ElementCount MaxVF = MaybeMaxVF.getValue();
   assert(MaxVF.isNonZero() && "MaxVF is zero.");
 
   if (!UserVF.isZero() && ElementCount::isKnownLE(UserVF, MaxVF)) {
     LLVM_DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
     assert(isPowerOf2_32(UserVF.getKnownMinValue()) &&
            "VF needs to be a power of two");
     // Collect the instructions (and their associated costs) that will be more
     // profitable to scalarize.
     CM.selectUserVectorizationFactor(UserVF);
     CM.collectInLoopReductions();
     buildVPlansWithVPRecipes(UserVF, UserVF);
     LLVM_DEBUG(printPlans(dbgs()));
     return {{UserVF, 0}};
   }
 
-  assert(!MaxVF.isScalable() &&
-         "Scalable vectors not yet supported beyond this point");
+  if (MaxVF.isScalable()) {
+    // TODO: Use scalable MaxVF once we've added support for scalable
+    // vectorization beyond this point. For now we convert it to fixed width,
+    // but this will be removed in a later patch.
+    MaxVF = ElementCount::getFixed(MaxVF.getKnownMinValue());
+  }
 
   for (ElementCount VF = ElementCount::getFixed(1);
        ElementCount::isKnownLE(VF, MaxVF); VF *= 2) {
     // Collect Uniform and Scalar instructions after vectorization with VF.
     CM.collectUniformsAndScalars(VF);
 
     // Collect the instructions (and their associated costs) that will be more
     // profitable to scalarize.

llvm/test/Transforms/LoopVectorize/AArch64/scalable-vf-hint.ll

+; REQUIRES: asserts
+; RUN: opt -mtriple=aarch64-none-linux-gnu -mattr=+sve -loop-vectorize -S < %s 2>&1 | FileCheck %s
+; RUN: opt -mtriple=aarch64-none-linux-gnu -mattr=+sve -loop-vectorize -pass-remarks-analysis=loop-vectorize -debug-only=loop-vectorize -S < %s 2>&1 | FileCheck --check-prefix=CHECK-DBG %s
+; RUN: opt -mtriple=aarch64-none-linux-gnu -loop-vectorize -pass-remarks-analysis=loop-vectorize -debug-only=loop-vectorize -S < %s 2>&1 | FileCheck --check-prefix=CHECK-NO-SVE %s
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+
+; These tests validate the behaviour of scalable vectorization factor hints,
+; where the following applies:
+;
+; * If the backend does not support scalable vectors, ignore the hint.
+; * If there are no dependencies and assuming the VF is a power of 2 the VF
+;   should be accepted. This applies to both fixed and scalable VFs.
+; * If the dependency is too small to use scalable vectors, change the VF to
+;   fixed, where existing behavior applies (clamping).
+; * If scalable vectorization is feasible given the dependency and the VF is
+;   valid, accept it. Otherwise, clamp to the max scalable VF.
+
+; test1
+;
+; Scalable vectorization unfeasible, clamp VF from (4, scalable) -> (4, fixed).
+;
+; The pragma applied to this loop implies a scalable vector <vscale x 4 x i32>
+; be used for vectorization. For fixed vectors the MaxVF=8, otherwise there
+; would be a dependence between vector lanes for vectors greater than 256 bits.
+;
+; void test1(int *a, int *b, int N) {
+;   #pragma clang loop vectorize(enable) vectorize_width(4, scalable)
+;   for (int i=0; i<N; ++i) {
+;     a[i + 8] = a[i] + b[i];
+;   }
+; }
+;
+; For scalable vectorization 'vscale' has to be considered, for this example
+; unless max(vscale)=2 it's unsafe to vectorize. For SVE max(vscale)=16, check
+; fixed-width vectorization is used instead.
+
+; CHECK-DBG: LV: Max legal vector width too small, scalable vectorization unfeasible. Using fixed-width vectorization instead.
+; CHECK-DBG: LV: The max safe VF is: 8.
+; CHECK-DBG: LV: Selecting VF: 4.
+; CHECK-LABEL: @test1
+; CHECK: <4 x i32>
+define void @test1(i32* %a, i32* %b) {
+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
+  %add = add nsw i32 %1, %0
+  %2 = add nuw nsw i64 %iv, 8
+  %arrayidx5 = getelementptr inbounds i32, i32* %a, i64 %2
+  store i32 %add, 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
+}
+
+!0 = !{!0, !1, !2}
+!1 = !{!"llvm.loop.vectorize.width", i32 4}
+!2 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
+
+; test2
+;
+; Scalable vectorization unfeasible, clamp VF from (8, scalable) -> (4, fixed).
+;
+; void test2(int *a, int *b, int N) {
+;   #pragma clang loop vectorize(enable) vectorize_width(8, scalable)
+;   for (int i=0; i<N; ++i) {
+;     a[i + 4] = a[i] + b[i];
+;   }
+; }
+
+; CHECK-DBG: LV: Max legal vector width too small, scalable vectorization unfeasible. Using fixed-width vectorization instead.
+; CHECK-DBG: LV: The max safe VF is: 4.
+; CHECK-DBG: LV: User VF=8 is unsafe, clamping to max safe VF=4.
+; CHECK-DBG: LV: Selecting VF: 4.
+; CHECK-LABEL: @test2
+; CHECK: <4 x i32>
+define void @test2(i32* %a, i32* %b) {
+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
+  %add = add nsw i32 %1, %0
+  %2 = add nuw nsw i64 %iv, 4
+  %arrayidx5 = getelementptr inbounds i32, i32* %a, i64 %2
+  store i32 %add, 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 !3
+
+exit:
+  ret void
+}
+
+!3 = !{!3, !4, !5}
+!4 = !{!"llvm.loop.vectorize.width", i32 8}
+!5 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
+
+; test3
+;
+; Scalable vectorization feasible and the VF is valid.
+;
+; Specifies a vector of <vscale x 2 x i32>, i.e. maximum of 32 x i32 with 2
+; words per 128-bits (unpacked).
+;
+; void test3(int *a, int *b, int N) {
+;   #pragma clang loop vectorize(enable) vectorize_width(2, scalable)
+;   for (int i=0; i<N; ++i) {
+;     a[i + 32] = a[i] + b[i];
+;   }
+; }
+;
+; Max fixed VF=32, Max scalable VF=2, safe to vectorize.
+
+; CHECK-DBG-LABEL: LV: Checking a loop in "test3"
+; CHECK-DBG: LV: The max safe VF is: vscale x 2.
+; CHECK-DBG: LV: Using user VF vscale x 2.
+; CHECK-LABEL: @test3
+; CHECK: <vscale x 2 x i32>
+define void @test3(i32* %a, i32* %b) {
+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
+  %add = add nsw i32 %1, %0
+  %2 = add nuw nsw i64 %iv, 32
+  %arrayidx5 = getelementptr inbounds i32, i32* %a, i64 %2
+  store i32 %add, 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 !6
+
+exit:
+  ret void
+}
+
+!6 = !{!6, !7, !8}
+!7 = !{!"llvm.loop.vectorize.width", i32 2}
+!8 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
+
+; test4
+;
+; Scalable vectorization feasible, but the VF is unsafe. Should clamp.
+;
+; Specifies a vector of <vscale x 4 x i32>, i.e. maximum of 64 x i32 with 4
+; words per 128-bits (packed).
+;
+; void test4(int *a, int *b, int N) {
+;   #pragma clang loop vectorize(enable) vectorize_width(4, scalable)
+;   for (int i=0; i<N; ++i) {
+;     a[i + 32] = a[i] + b[i];
+;   }
+; }
+;
+; Max fixed VF=32, Max scalable VF=2, unsafe to vectorize. Should clamp to 2.
+
+; CHECK-DBG-LABEL: LV: Checking a loop in "test4"
+; CHECK-DBG: LV: The max safe VF is: vscale x 2.
+; CHECK-DBG: LV: User VF=vscale x 4 is unsafe, clamping to max safe VF=vscale x 2.
+; CHECK-DBG: remark: <unknown>:0:0: User-specified vectorization factor vscale x 4 is unsafe, clamping to maximum safe vectorization factor vscale x 2
+; CHECK-DBG: LV: Selecting VF: 2.
+; CHECK-LABEL: @test4
+; CHECK: <2 x i32>
+define void @test4(i32* %a, i32* %b) {
+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
+  %add = add nsw i32 %1, %0
+  %2 = add nuw nsw i64 %iv, 32
+  %arrayidx5 = getelementptr inbounds i32, i32* %a, i64 %2
+  store i32 %add, 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 !9
+
+exit:
+  ret void
+}
+
+!9 = !{!9, !10, !11}
+!10 = !{!"llvm.loop.vectorize.width", i32 4}
+!11 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
+
+; test5
+;
+; Scalable vectorization feasible and the VF is valid.
+;
+; Specifies a vector of <vscale x 4 x i32>, i.e. maximum of 64 x i32 with 4
+; words per 128-bits (packed).
+;
+; void test5(int *a, int *b, int N) {
+;   #pragma clang loop vectorize(enable) vectorize_width(4, scalable)
+;   for (int i=0; i<N; ++i) {
+;     a[i + 128] = a[i] + b[i];
+;   }
+; }
+;
+; Max fixed VF=128, Max scalable VF=8, safe to vectorize.
+
+; CHECK-DBG-LABEL: LV: Checking a loop in "test5"
+; CHECK-DBG: LV: The max safe VF is: vscale x 8.
+; CHECK-DBG: LV: Using user VF vscale x 4
+; CHECK-LABEL: @test5
+; CHECK: <vscale x 4 x i32>
+define void @test5(i32* %a, i32* %b) {
+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
+  %add = add nsw i32 %1, %0
+  %2 = add nuw nsw i64 %iv, 128
+  %arrayidx5 = getelementptr inbounds i32, i32* %a, i64 %2
+  store i32 %add, 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 !12
+
+exit:
+  ret void
+}
+
+!12 = !{!12, !13, !14}
+!13 = !{!"llvm.loop.vectorize.width", i32 4}
+!14 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
+
+; test6
+;
+; Scalable vectorization feasible, but the VF is unsafe. Should clamp.
+;
+; Specifies a vector of <vscale x 16 x i32>, i.e. maximum of 256 x i32.
+;
+; void test6(int *a, int *b, int N) {
+;   #pragma clang loop vectorize(enable) vectorize_width(16, scalable)
+;   for (int i=0; i<N; ++i) {
+;     a[i + 128] = a[i] + b[i];
+;   }
+; }
+;
+; Max fixed VF=128, Max scalable VF=8, unsafe to vectorize. Should clamp to 8.
+
+; CHECK-DBG-LABEL: LV: Checking a loop in "test6"
+; CHECK-DBG: LV: The max safe VF is: vscale x 8.
+; CHECK-DBG: LV: User VF=vscale x 16 is unsafe, clamping to max safe VF=vscale x 8.
+; CHECK-DBG: remark: <unknown>:0:0: User-specified vectorization factor vscale x 16 is unsafe, clamping to maximum safe vectorization factor vscale x 8
+; CHECK-DBG: LV: Selecting VF: 8.
+; CHECK-LABEL: @test6
+; CHECK: <8 x i32>
+define void @test6(i32* %a, i32* %b) {
+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
+  %add = add nsw i32 %1, %0
+  %2 = add nuw nsw i64 %iv, 128
+  %arrayidx5 = getelementptr inbounds i32, i32* %a, i64 %2
+  store i32 %add, 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 !15
+
+exit:
+  ret void
+}
+
+!15 = !{!15, !16, !17}
+!16 = !{!"llvm.loop.vectorize.width", i32 16}
+!17 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
+
+; CHECK-NO-SVE: LV: Not vectorizing: Ignoring scalable VF because target does not support scalable vectors.
+; CHECK-NO-SVE: remark: <unknown>:0:0: loop not vectorized: Ignoring scalable VF because target does not support scalable vectors.
+define void @test_no_sve(i32* %a, i32* %b) {
+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
+  %add = add nsw i32 %1, %0
+  %2 = add nuw nsw i64 %iv, 4
+  %arrayidx5 = getelementptr inbounds i32, i32* %a, i64 %2
+  store i32 %add, 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 !18
+
+exit:
+  ret void
+}
+
+!18 = !{!18, !19, !20}
+!19 = !{!"llvm.loop.vectorize.width", i32 4}
+!20 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}

llvm/test/Transforms/LoopVectorize/scalable-vf-hint.ll

+; REQUIRES: asserts
+; RUN: opt -loop-vectorize -pass-remarks-analysis=loop-vectorize -debug-only=loop-vectorize -S < %s 2>&1 | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+
+; CHECK: LV: Not vectorizing: Ignoring scalable VF because target does not support scalable vectors.
+; CHECK: remark: <unknown>:0:0: loop not vectorized: Ignoring scalable VF because target does not support scalable vectors.
+define void @test1(i32* %a, i32* %b) {
+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
+  %add = add nsw i32 %1, %0
+  %2 = add nuw nsw i64 %iv, 4
+  %arrayidx5 = getelementptr inbounds i32, i32* %a, i64 %2
+  store i32 %add, 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
+}
+
+!0 = !{!0, !1, !2}
+!1 = !{!"llvm.loop.vectorize.width", i32 4}
+!2 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}