[VPLAN] Determine Vector Width programmatically.

[fpetrogalli]

Summary

With this change, the VPlan native path is triggered with the directive:

#pragma clang loop vectorize(enable)

There is no need to specify the vectorize_width(N) clause.

[hsaito]

While we are at this, let's talk about downstream dependency, if any, for allowing more than one candidate VF along the native path. At least we can write down a list of TODOs so that we'll be aware of the things we need to improve at a time in the future.

This may be a bit more than what you anticipated for this patch, but I think this will help everyone.

lib/Transforms/Vectorize/LoopVectorize.cpp

7123

This is in some sense reinventing the wheel, and I kind of imagine this can quickly get out of control if different people start trying to add different constraints of their interest.

Let's start talking about how we can get to the point of being able to use unified computeMaxVF(). For this patch, however, I'd be happy if we get to the point of being able to use unified getSmallestAndWidestTypes().

7146

Let's try not do this here. Please do this kind of thing inside planInVPlanNativePath().

[fhahn]

Thanks for working on this Francesco! +1 to Hideki's point about moving to LoopVectorizationPlanner.

lib/Transforms/Vectorize/LoopVectorize.cpp

1005

Is this related to the patch? I suppose guessVectorizationFactor could set UserVF to 1, which could reach this code. It would be better to not vectorize with VF == 1.

3796

Related to the patch?

test/Transforms/LoopVectorize/outer_loop_test1_no_explicit_vect_width.ll

91

Looks like we are just using a single store width in this test. Maybe it would be worth adding loads/stores to a different type as well?

[npanchen]

lib/Transforms/Vectorize/LoopVectorize.cpp

1005

For VF = 1 CM is not called, thus this change looks as unrelated to this commit. If it does, what is a reason and why other places in CM are not modified ?

7123

Agree with Hideki. Ideally, this function has to return a range [MinVF, MaxVF], which can be used in buildVPlans() to minimize number of built VPlans.

Please, remember that target can have different vector registers.

7128

For a cases when load and store instructions were hoisted out of the loop, this function will return WidestVectorRegBits, which could be >= 128.

7131

getMemInstValueType(I)

[fpetrogalli]

@hsaito , @fhahn , @npanchen ,

Thank you for looking into this.

I have responded to all comments about the design. I will work on a new patch an submit it according to your suggestions. For now, I have ignored those comments on trivial changes (for example, @fhahn comment on adding more test coverage). I will fix such comments and update them along the way with the re-implementation of the patch.

Thank you,

Francesco

lib/Transforms/Vectorize/LoopVectorize.cpp

1005

Yes, this is related to this patch. By guessing the number of lanes with the algorithm i proposed, all the target independent invocations of opt that use VPLAN end up reporting vector registers of size 32-bit (TTI->getRegisterBitWidth(true /* Vector*/)), which means the vectorization performed in some of the tests are done with a vectorization factor of 1. I am not saying this is the right thing to do, I was actually hacking something that worked to initiate the discussion.

1005

This assertion was firing when testing test/Transforms/LoopVectorize/explicit_outer_detection.ll. Because the test is run with no explicit vector with (this patch removes such need) and without specifying a target for opt, the vectorizer was generating a vector loop with one lane, for the same reason explained in my previsous comment (no target implies TTI->getRegisterBitWidth(true /* Vector*/) returning 32).

3796

Yes. Again, this was needed because of an assertion firing when building a binary operator in the process of vectorizing case2 in test/Transforms/LoopVectorize/explicit_outer_detection.ll. The loop in case2 get's vectorized with this patch, with a vectorization factor of 1, but the code generation in the inner loop vectorizer is trying to build a binary operator between a scalar phi and a vector consisting of one lane, which is of course not possible. This change makes sure that the phi is generated not as a scalar but a one lane vector to prevent the failure when building the binary operator.

7123

Responding to both comments from @hsaito and @npanchen here:

> This is in some sense reinventing the wheel

I fully agree :). I needed a starting point to be able to discuss this with you.

Let's start talking about how we can get to the point of being able to use unified computeMaxVF(). For this patch, however, I'd be happy if we get to the point of being able to use unified getSmallestAndWidestTypes().

OK, I will look into getting access to getSmallestAndWidestTypes here.

Please, remember that target can have different vector registers.

Yes, I understand that using the widest ones is not ideal, as you miss for example 2-lane vectorization on floating point data on machines that support both 64-bit and 128-bit vector registers (for example aarch64).

Just to make sure we are on the same page here. This could be solved by considering all possible power of two vector widths up to the maximum one. For example, if the loop is processing 32-bit data, and the target has 128, 256, and 512-bit wide registers, we should ask VPlan to consider 4, 8 and 16 lanes vectorizations. Is my interpretation correct here?

7128

Sorry I don't understand this comment, could you please explain it with and example? Also, given that my understanding is that you want me to use getSmallestAndWidestTypes, is the comment still valid? Because I think that by using getSmallestAndWidestTypes I essentially will remove my custom code.

[dcaballe]

Thanks Francesco for helping us remove some of the constraints we have in VPlan native path!
I agree with the idea of using getSmallestAndWidestTypes() as a starting point whereas we don't have the proper cost model.

One more comment below.
Thanks,
Diego

[dcaballe]

lib/Transforms/Vectorize/LoopVectorize.cpp

7149

Regarding the issue with VF = 1, we are using VF = 1 to denote no vectorization and I think we should preserve that behavior. This basically means we shouldn't try to generate vector code (shouldn't invoke getWideningDecision et al.) with VF = 1.

I think the problem happens because we are setting UserVF here, instead of VF, and probably 1 is an unexpected value for UserVF (just guessing, I haven't checked it out). Maybe we should leave UserVF to values actually coming from the user and set VF instead? If you move this code to planInVPlanNativePath, as suggested, I think the VF = 1 problem would be fixed.

[npanchen]

lib/Transforms/Vectorize/LoopVectorize.cpp

7128

The example I was thinking about is pretty simple:

#pragma clang loop vectorize(enable)
for (i = 0; i < n; ++i) {
  red += i;
}

After some optimization the loop body can look like:

loop_body:
%1 = phi i64     %i_init,              %i
%2 = phi i32     %red_init,          %red
%red = %2 + %1
%i = %1 + 1
cmp %i, %n

In this case guessVPlanVF() will not find any LD/ST instruction, thus Max will be equal to 1 and WidestVectorRegBits will be returned.
For example, in case of SKX WidestVectorRegBits = 512.

[fpetrogalli]

Hi all,

I finally addressed your comments and update the patch. Let me know what you think.

Francesco

lib/Transforms/Vectorize/LoopVectorize.cpp

1005

I restored this assertion.

3796

I restored this code.

7123

I have a version that does [MinVF, MaxVF] computation, but this triggers the assertion in this method:

void LoopVectorizationPlanner::setBestPlan(unsigned VF, unsigned UF) {
  LLVM_DEBUG(dbgs() << "Setting best plan to VF=" << VF << ", UF=" << UF
                    << '\n');
  BestVF = VF;
  BestUF = UF;

  erase_if(VPlans, [VF](const VPlanPtr &Plan) {
    return !Plan->hasVF(VF);
  });
  assert(VPlans.size() == 1 && "Best VF has not a single VPlan.");
}

Am I correct thinking that to be able to choose among different plans, we need to provide VPLAN with a cost model that is able to evaluate all the options? If that's the case, I think that generating Max and Min VF goes beyond the scope of this patch.

7128

The new version doesn't use my custom code but getSmallestAndWidestTypes. I could add a test that does the hoisting on outer loops, but I am not sure how to construct this example. Can you provide a C example to start with? Or are you happy for me to skip this check?

7131

This code is not used anymore in the last version.

7149

The VF = 1 case is not generated anymore after the last change set, marking this comment as done.

test/Transforms/LoopVectorize/outer_loop_test1_no_explicit_vect_width.ll

91

The patch doesn't rely anymore on the load/store types, but uses getSmallestAndWidestTypes. Do you still want me to add such test, or can we trust getSmallestAndWidestTypes of doing the right job here?

[fpetrogalli]

I mostly changes to code to use the infrastructure that LLVM already provides to determine the vectorization factor.

[hsaito]

In D57598#1428062, @fpetrogalli wrote:

I mostly changes to code to use the infrastructure that LLVM already provides to determine the vectorization factor.

Haven't gone through the LIT test yet, but the code addressed my concerns.

lib/Transforms/Vectorize/LoopVectorize.cpp

6104

If the return value is always 2 or greater, we should replace lines 6099-6103 with assert for VF >= 2.
Else, we should move lines 6099-6103 below line 6108 and adjust the code and the comment accordingly.

6111

This is no longer user VF. Should be something along the lines of

"LV: Using " << UserVF ? "user VF " : "computed VF " << VF

[fpetrogalli]

Addressed second round of comments from @hsaito.

[hsaito]

lib/Transforms/Vectorize/LoopVectorize.cpp

6105

(VPlanBuildStressTest && VF < 2) ?

[fpetrogalli]

lib/Transforms/Vectorize/LoopVectorize.cpp

6105

Sure, but why? If we are in VPlanBuildStressTest, why would you care about the value of VF?

Isn't it better to have full control on the stress tests and make sure that we always vectorize with VF = 4?

[hsaito]

lib/Transforms/Vectorize/LoopVectorize.cpp

6105

I then suggest going back to (VPlanBuildStressTest && !UserVF) --- override only if the programmer doesn't explicitly set it.

For the spirit of stress testing, I think it makes sense to have the ability to choose any legal VF (and in the future extend it to "scalable"). Maybe, we should take compiler option instead of hard-coded 4. So, something along like the following?

if (!UserVF)

if (VPlanBuildStressTest)
     VF = 4
else
     VF = determineVplanVF()

[fpetrogalli]

lib/Transforms/Vectorize/LoopVectorize.cpp

6105

Thank you for explaining. I have re-worked the if statement as requested.

[hsaito]

LGTM. Please wait for a few days to give others a chance to go over your updated patch.

[Meinersbur]

lib/Transforms/Vectorize/LoopVectorizationPlanner.h

233–234 ↗

(On Diff #190562)

[nit] It it somewhat unusual to pass by-value with const.

test/Transforms/LoopVectorize/outer_loop_test1_no_explicit_vect_width.ll

17–18

The RUN lines are usually at the top of the file

[fpetrogalli]

lib/Transforms/Vectorize/LoopVectorizationPlanner.h

233–234 ↗

(On Diff #190562)

I see, but I run into problems while developing this caused by the fact that I could modify UserVF inside the function. Would a const ref be better? I can try that, if it doesn't work and you don't like this const value, I will revert the interface to the original one.

[Meinersbur]

lib/Transforms/Vectorize/LoopVectorizationPlanner.h

233–234 ↗

(On Diff #190562)

It should be a coding standard question. It may help against accidental assignment in the implementation to have a by-value const. Unfortunately, this function implementation detail leaks into the function signature.

Since I don't see this anywhere else in the LLVM code base, so I'd prefer to not do it. Otherwise, if we want to apply this consequently, we'd have to add const to by-value parameters to many functions (e.g. setBestPlan below).

Const ref would be worse.

[fpetrogalli]

lib/Transforms/Vectorize/LoopVectorizationPlanner.h

233–234 ↗

(On Diff #190562)

Thank you for explaining. I have removed the const and restored the original interface..

[dcaballe]

Thanks, Francesco. LGTM!

[fpetrogalli]

@npanchen , @fhahn , gentle ping :)

Francesco

[hsaito]

In D57598#1440282, @fpetrogalli wrote:

@npanchen , @fhahn , gentle ping :)

Francesco

I think you waited long enough already. I suggest proceeding to commit. Any further comments can be addressed post-commit.

[fpetrogalli]

In D57598#1440284, @hsaito wrote:

I think you waited long enough already. I suggest proceeding to commit. Any further comments can be addressed post-commit.

OK.

@hsaito, I don't have commit access, could you commit this change for me?

Thank you!

[hsaito]

@hsaito, I don't have commit access, could you commit this change for me?

Will do. I had a day off today. Will take care of it, hopefully tomorrow.

[fhahn]

lib/Transforms/Vectorize/LoopVectorize.cpp

6111

Is the lambda necessary? You can just have << (UserVF ? "user VF" : "computed VF") << inline, I think

6114

I think it would be clearer to say something along the lines of "LV: Using .. VF to build VPlans", to make it a bit clearer that VF is not necessarily what will be used for vectorization (e.g. VF == 1 means no vectorization).

7157

There is VectorizationFactor::Disabled(), which is used in other places. Could you use this here instead of VF.Widht == 1?

test/Transforms/LoopVectorize/explicit_outer_detection.ll

71

Comment needs updating, we now analyze loops without user VF too.

test/Transforms/LoopVectorize/outer_loop_test1_no_explicit_vect_width.ll

1

This test will fail on targets not built with X86/AArch64 targets. The X86 version should go in test/Transforms/LoopVectorize/X86/ and the AArch64 one in test/Transforms/LoopVectorize/AArch64/

[fhahn]

lib/Transforms/Vectorize/LoopVectorize.cpp

6103

I think as a follow up, we can drop VPlanBuildStressTest, now that we do not require a UserVF to build VPlans.

[fpetrogalli]

I have addressed last round of comments from @fhahn .

[fpetrogalli]

lib/Transforms/Vectorize/LoopVectorize.cpp

6103

Should I address this in a separate patch or this one?

7157

struct VectorizationFactor did not have a "==" operator. I have added it, and adapted this condition.

[fpetrogalli]

I forgot to update the comment in the test as requested from @fhahn . Now it is done.

[fhahn]

Thanks Francesco! I'll commit the change tomorrow, unless @hsaito does it today :)

lib/Transforms/Vectorize/LoopVectorize.cpp

6103

Yep that would be great!

[fpetrogalli]

In D57598#1445180, @fhahn wrote:

Thanks Francesco! I'll commit the change tomorrow, unless @hsaito does it today :)

Thank you @fhahn !

[fhahn]

lib/Transforms/Vectorize/LoopVectorizationPlanner.h

180 ↗

(On Diff #192498)

No braces needed, I'll run clang-format on your patch before committing.

[hsaito]

In D57598#1445180, @fhahn wrote:

Thanks Francesco! I'll commit the change tomorrow, unless @hsaito does it today :)

@fhahn, I'm trying to figure out the appropriate proxy setting for external git.
You may be quicker. First commit attempt after SVN is gone. Some learning curve here.

[fhahn]

Thanks Francesco!

[hsaito]

In D57598#1445741, @fhahn wrote:

Thanks Francesco!

Thanks, Francesco and Florian.

lib/Transforms/Vectorize/LoopVectorize.cpp

for (unsigned i = 0; i < Grp->getFactor(); ++i) {
       }
     }
   }
 
   /// Return the cost model decision for the given instruction \p I and vector
   /// width \p VF. Return CM_Unknown if this instruction did not pass
   /// through the cost modeling.
   InstWidening getWideningDecision(Instruction *I, unsigned VF) {
     assert(VF >= 2 && "Expected VF >=2");
 
     // Cost model is not run in the VPlan-native path - return conservative
     // result until this changes.
     if (EnableVPlanNativePath)
       return CM_GatherScatter;
 
     std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
     auto Itr = WideningDecisions.find(InstOnVF);
static bool isExplicitVecOuterLoop(Loop *OuterLp,
 
   Function *Fn = OuterLp->getHeader()->getParent();
   if (!Hints.allowVectorization(Fn, OuterLp,
                                 true /*VectorizeOnlyWhenForced*/)) {
     LLVM_DEBUG(dbgs() << "LV: Loop hints prevent outer loop vectorization.\n");
     return false;
   }
 
-  if (!Hints.getWidth()) {
-    LLVM_DEBUG(dbgs() << "LV: Not vectorizing: No user vector width.\n");
-    Hints.emitRemarkWithHints();
-    return false;
-  }
-
   if (Hints.getInterleave() > 1) {
     // TODO: Interleave support is future work.
     LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Interleave is not supported for "
                          "outer loops.\n");
     Hints.emitRemarkWithHints();
     return false;
   }
 
void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, unsigned UF,
                                               unsigned VF) {
   PHINode *P = cast<PHINode>(PN);
   if (EnableVPlanNativePath) {
     // Currently we enter here in the VPlan-native path for non-induction
     // PHIs where all control flow is uniform. We simply widen these PHIs.
     // Create a vector phi with no operands - the vector phi operands will be
     // set at the end of vector code generation.
     Type *VecTy =
         (VF == 1) ? PN->getType() : VectorType::get(PN->getType(), VF);
     Value *VecPhi = Builder.CreatePHI(VecTy, PN->getNumOperands(), "vec.phi");
     VectorLoopValueMap.setVectorValue(P, 0, VecPhi);
     OrigPHIsToFix.push_back(P);
 
     return;
   }
 
   assert(PN->getParent() == OrigLoop->getHeader() &&
void LoopVectorizationCostModel::collectValuesToIgnore() {
   // detection.
   for (auto &Induction : *Legal->getInductionVars()) {
     InductionDescriptor &IndDes = Induction.second;
     const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts();
     VecValuesToIgnore.insert(Casts.begin(), Casts.end());
   }
 }
 
+// TODO: we could return a pair of values that specify the max VF and
+// min VF, to be used in `buildVPlans(MinVF, MaxVF)` instead of
+// `buildVPlans(VF, VF)`. We cannot do it because VPLAN at the moment
+// doesn't have a cost model that can choose which plan to execute if
+// more than one is generated.
+unsigned determineVPlanVF(const unsigned WidestVectorRegBits,
+                          LoopVectorizationCostModel &CM) {
+  unsigned WidestType;
+  std::tie(std::ignore, WidestType) = CM.getSmallestAndWidestTypes();
+  return WidestVectorRegBits / WidestType;
+}
+
 VectorizationFactor
 LoopVectorizationPlanner::planInVPlanNativePath(bool OptForSize,
                                                 unsigned UserVF) {
+  unsigned VF = UserVF;
   // Outer loop handling: They may require CFG and instruction level
   // transformations before even evaluating whether vectorization is profitable.
   // Since we cannot modify the incoming IR, we need to build VPlan upfront in
   // the vectorization pipeline.
   if (!OrigLoop->empty()) {
-    // TODO: If UserVF is not provided, we set UserVF to 4 for stress testing.
-    // This won't be necessary when UserVF is not required in the VPlan-native
-    // path.
-    if (VPlanBuildStressTest && !UserVF)
-      UserVF = 4;
+    // If the user doesn't provide a vectorization factor, determine a
+    // reasonable one.
+    if (!UserVF) {
+      // We set VF to 4 for stress testing.
+      if (VPlanBuildStressTest)
+        VF = 4;
+      else
+        VF = determineVPlanVF(TTI->getRegisterBitWidth(true /* Vector*/), CM);
+    }
 
     assert(EnableVPlanNativePath && "VPlan-native path is not enabled.");
-    assert(UserVF && "Expected UserVF for outer loop vectorization.");
-    assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
-    LLVM_DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
-    buildVPlans(UserVF, UserVF);
+    assert(isPowerOf2_32(VF) && "VF needs to be a power of two");
+    const auto Msg = [](const bool IsUserVF) -> const char * {
+      return IsUserVF ? "user VF " : "computed VF ";
+    };
+    LLVM_DEBUG(dbgs() << "LV: Using " << Msg(UserVF) << VF << ".\n");
+    buildVPlans(VF, VF);
 
     // For VPlan build stress testing, we bail out after VPlan construction.
     if (VPlanBuildStressTest)
       return VectorizationFactor::Disabled();
 
-    return {UserVF, 0};
+    return {VF, 0};
   }
 
   LLVM_DEBUG(
       dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the "
                 "VPlan-native path.\n");
   return VectorizationFactor::Disabled();
 }
 
void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) {
   // Last (and currently only) operand is a mask.
   InnerLoopVectorizer::VectorParts MaskValues(State.UF);
   VPValue *Mask = User->getOperand(User->getNumOperands() - 1);
   for (unsigned Part = 0; Part < State.UF; ++Part)
     MaskValues[Part] = State.get(Mask, Part);
   State.ILV->vectorizeMemoryInstruction(&Instr, &MaskValues);
 }
 
 // Process the loop in the VPlan-native vectorization path. This path builds
 // VPlan upfront in the vectorization pipeline, which allows to apply
 // VPlan-to-VPlan transformations from the very beginning without modifying the
 // input LLVM IR.
 static bool processLoopInVPlanNativePath(
     Loop *L, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT,
     LoopVectorizationLegality *LVL, TargetTransformInfo *TTI,
     TargetLibraryInfo *TLI, DemandedBits *DB, AssumptionCache *AC,
     OptimizationRemarkEmitter *ORE, LoopVectorizeHints &Hints) {
 
   assert(EnableVPlanNativePath && "VPlan-native path is disabled.");
   Function *F = L->getHeader()->getParent();
   InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL->getLAI());
   LoopVectorizationCostModel CM(L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE, F,
                                 &Hints, IAI);
   // Use the planner for outer loop vectorization.
   // TODO: CM is not used at this point inside the planner. Turn CM into an
   // optional argument if we don't need it in the future.
   LoopVectorizationPlanner LVP(L, LI, TLI, TTI, LVL, CM);
 
   // Get user vectorization factor.
-  unsigned UserVF = Hints.getWidth();
+  const unsigned UserVF = Hints.getWidth();
 
   // Check the function attributes to find out if this function should be
   // optimized for size.
   bool OptForSize =
       Hints.getForce() != LoopVectorizeHints::FK_Enabled && F->optForSize();
 
   // Plan how to best vectorize, return the best VF and its cost.
-  VectorizationFactor VF = LVP.planInVPlanNativePath(OptForSize, UserVF);
+  const VectorizationFactor VF = LVP.planInVPlanNativePath(OptForSize, UserVF);
 
   // If we are stress testing VPlan builds, do not attempt to generate vector
   // code. Masked vector code generation support will follow soon.
-  if (VPlanBuildStressTest || EnableVPlanPredication)
+  // Also, do not attempt to vectorize if no vector code will be produced.
+  if (VPlanBuildStressTest || EnableVPlanPredication || VF.Width == 1)
     return false;
 
   LVP.setBestPlan(VF.Width, 1);
 
-  InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, UserVF, 1, LVL,
+  InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, 1, LVL,
                          &CM);
   LLVM_DEBUG(dbgs() << "Vectorizing outer loop in \""
                     << L->getHeader()->getParent()->getName() << "\"\n");
   LVP.executePlan(LB, DT);
 
   // Mark the loop as already vectorized to avoid vectorizing again.
   Hints.setAlreadyVectorized();
 

test/Transforms/LoopVectorize/explicit_outer_detection.ll

outer.inc:                                        ; preds = %inner.body, %outer.body
   %indvars.iv.next36 = add nuw nsw i64 %indvars.iv35, 1
   %exitcond39 = icmp eq i64 %indvars.iv.next36, %wide.trip.count38
   br i1 %exitcond39, label %for.end15, label %outer.body, !llvm.loop !6
 
 for.end15:                                        ; preds = %outer.inc, %entry
   ret void
 }
 
 ; Case 2: Annotated outer loop WITHOUT vector width information doesn't have to
 ; be collected.
 
 ; CHECK-LABEL: case2
-; CHECK-NOT: LV: Loop hints: force=enabled
-; CHECK-NOT: LV: We can vectorize this outer loop!
-; CHECK: LV: Loop hints: force=?
-; CHECK: LV: Found a loop: inner.body
+; CHECK: LV: Loop hints: force=enabled width=0 unroll=0
+; CHECK: LV: We can vectorize this outer loop!
+; CHECK: LV: Using computed VF 1.
 
 define void @case2(i32* nocapture %a, i32* nocapture readonly %b, i32 %N, i32 %M) local_unnamed_addr {
 entry:
   %cmp32 = icmp sgt i32 %N, 0
   br i1 %cmp32, label %outer.ph, label %for.end15
 
 outer.ph:                                          ; preds = %entry
   %cmp230 = icmp sgt i32 %M, 0

test/Transforms/LoopVectorize/outer_loop_test1_no_explicit_vect_width.ll

+; RUN: opt -S -loop-vectorize -enable-vplan-native-path -mtriple aarch64-gnu-linux < %s | FileCheck %s
+; RUN: opt -S -loop-vectorize -enable-vplan-native-path -mtriple x86_64  < %s | FileCheck %s
+; RUN: opt -S -loop-vectorize -enable-vplan-native-path -mtriple x86_64  -mattr=+avx < %s | FileCheck %s --check-prefix=AVX
+; RUN: opt -S -loop-vectorize -enable-vplan-native-path -mtriple x86_64  -mattr=+avx2 < %s | FileCheck %s --check-prefix=AVX
+
+; extern int arr[8][8];
+; extern int arr2[8];
+;
+; void foo(int n)
+; {
+;   int i1, i2;
+;
+; #pragma clang loop vectorize(enable)
+;   for (i1 = 0; i1 < 8; i1++) {
+;     arr2[i1] = i1;
+;     for (i2 = 0; i2 < 8; i2++)
+;       arr[i2][i1] = i1 + n;
+;   }
+; }
+;
+
+; CHECK-LABEL: vector.ph:
+; CHECK: %[[SplatVal:.*]] = insertelement <4 x i32> undef, i32 %n, i32 0
+; CHECK: %[[Splat:.*]] = shufflevector <4 x i32> %[[SplatVal]], <4 x i32> undef, <4 x i32> zeroinitializer
+
+; CHECK-LABEL: vector.body:
+; CHECK: %[[Ind:.*]] = phi i64 [ 0, %vector.ph ], [ %[[IndNext:.*]], %[[ForInc:.*]] ]
+; CHECK: %[[VecInd:.*]] = phi <4 x i64> [ <i64 0, i64 1, i64 2, i64 3>, %vector.ph ], [ %[[VecIndNext:.*]], %[[ForInc]] ]
+; CHECK: %[[AAddr:.*]] = getelementptr inbounds [8 x i32], [8 x i32]* @arr2, i64 0, <4 x i64> %[[VecInd]]
+; CHECK: %[[VecIndTr:.*]] = trunc <4 x i64> %[[VecInd]] to <4 x i32>
+; CHECK: call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %[[VecIndTr]], <4 x i32*> %[[AAddr]], i32 4, <4 x i1> <i1 true, i1 true, i1 true, i1 true>)
+; CHECK: %[[VecIndTr2:.*]] = trunc <4 x i64> %[[VecInd]] to <4 x i32>
+; CHECK: %[[StoreVal:.*]] = add nsw <4 x i32> %[[VecIndTr2]], %[[Splat]]
+; CHECK: br label %[[InnerLoop:.+]]
+
+; CHECK: [[InnerLoop]]:
+; CHECK: %[[InnerPhi:.*]] = phi <4 x i64> [ %[[InnerPhiNext:.*]], %[[InnerLoop]] ], [ zeroinitializer, %vector.body ]
+; CHECK: %[[AAddr2:.*]] = getelementptr inbounds [8 x [8 x i32]], [8 x [8 x i32]]* @arr, i64 0, <4 x i64> %[[InnerPhi]], <4 x i64> %[[VecInd]]
+; CHECK: call void @llvm.masked.scatter.v4i32.v4p0i32(<4 x i32> %[[StoreVal]], <4 x i32*> %[[AAddr2]], i32 4, <4 x i1> <i1 true, i1 true, i1 true
+; CHECK: %[[InnerPhiNext]] = add nuw nsw <4 x i64> %[[InnerPhi]], <i64 1, i64 1, i64 1, i64 1>
+; CHECK: %[[VecCond:.*]] = icmp eq <4 x i64> %[[InnerPhiNext]], <i64 8, i64 8, i64 8, i64 8>
+; CHECK: %[[InnerCond:.*]] = extractelement <4 x i1> %[[VecCond]], i32 0
+; CHECK: br i1 %[[InnerCond]], label %[[ForInc]], label %[[InnerLoop]]
+
+; CHECK: [[ForInc]]:
+; CHECK: %[[IndNext]] = add i64 %[[Ind]], 4
+; CHECK: %[[VecIndNext]] = add <4 x i64> %[[VecInd]], <i64 4, i64 4, i64 4, i64 4>
+; CHECK: %[[Cmp:.*]] = icmp eq i64 %[[IndNext]], 8
+; CHECK: br i1 %[[Cmp]], label %middle.block, label %vector.body
+
+; AVX-LABEL: vector.ph:
+; AVX: %[[SplatVal:.*]] = insertelement <8 x i32> undef, i32 %n, i32 0
+; AVX: %[[Splat:.*]] = shufflevector <8 x i32> %[[SplatVal]], <8 x i32> undef, <8 x i32> zeroinitializer
+
+; AVX-LABEL: vector.body:
+; AVX: %[[Ind:.*]] = phi i64 [ 0, %vector.ph ], [ %[[IndNext:.*]], %[[ForInc:.*]] ]
+; AVX: %[[VecInd:.*]] = phi <8 x i64> [ <i64 0, i64 1, i64 2, i64 3, i64 4, i64 5, i64 6, i64 7>, %vector.ph ], [ %[[VecIndNext:.*]], %[[ForInc]] ]
+; AVX: %[[AAddr:.*]] = getelementptr inbounds [8 x i32], [8 x i32]* @arr2, i64 0, <8 x i64> %[[VecInd]]
+; AVX: %[[VecIndTr:.*]] = trunc <8 x i64> %[[VecInd]] to <8 x i32>
+; AVX: call void @llvm.masked.scatter.v8i32.v8p0i32(<8 x i32> %[[VecIndTr]], <8 x i32*> %[[AAddr]], i32 4, <8 x i1> <i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true>)
+; AVX: %[[VecIndTr2:.*]] = trunc <8 x i64> %[[VecInd]] to <8 x i32>
+; AVX: %[[StoreVal:.*]] = add nsw <8 x i32> %[[VecIndTr2]], %[[Splat]]
+; AVX: br label %[[InnerLoop:.+]]
+
+; AVX: [[InnerLoop]]:
+; AVX: %[[InnerPhi:.*]] = phi <8 x i64> [ %[[InnerPhiNext:.*]], %[[InnerLoop]] ], [ zeroinitializer, %vector.body ]
+; AVX: %[[AAddr2:.*]] = getelementptr inbounds [8 x [8 x i32]], [8 x [8 x i32]]* @arr, i64 0, <8 x i64> %[[InnerPhi]], <8 x i64> %[[VecInd]]
+; AVX: call void @llvm.masked.scatter.v8i32.v8p0i32(<8 x i32> %[[StoreVal]], <8 x i32*> %[[AAddr2]], i32 4, <8 x i1> <i1 true, i1 true, i1 true
+; AVX: %[[InnerPhiNext]] = add nuw nsw <8 x i64> %[[InnerPhi]], <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>
+; AVX: %[[VecCond:.*]] = icmp eq <8 x i64> %[[InnerPhiNext]], <i64 8, i64 8, i64 8, i64 8, i64 8, i64 8, i64 8, i64 8>
+; AVX: %[[InnerCond:.*]] = extractelement <8 x i1> %[[VecCond]], i32 0
+; AVX: br i1 %[[InnerCond]], label %[[ForInc]], label %[[InnerLoop]]
+
+; AVX: [[ForInc]]:
+; AVX: %[[IndNext]] = add i64 %[[Ind]], 8
+; AVX: %[[VecIndNext]] = add <8 x i64> %[[VecInd]], <i64 8, i64 8, i64 8, i64 8, i64 8, i64 8, i64 8, i64 8>
+; AVX: %[[Cmp:.*]] = icmp eq i64 %[[IndNext]], 8
+; AVX: br i1 %[[Cmp]], label %middle.block, label %vector.body
+
+@arr2 = external global [8 x i32], align 16
+@arr = external global [8 x [8 x i32]], align 16
+
+; Function Attrs: norecurse nounwind uwtable
+define void @foo(i32 %n) {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.inc8, %entry
+  %indvars.iv21 = phi i64 [ 0, %entry ], [ %indvars.iv.next22, %for.inc8 ]
+  %arrayidx = getelementptr inbounds [8 x i32], [8 x i32]* @arr2, i64 0, i64 %indvars.iv21
+  %0 = trunc i64 %indvars.iv21 to i32
+  store i32 %0, i32* %arrayidx, align 4
+  %1 = trunc i64 %indvars.iv21 to i32
+  %add = add nsw i32 %1, %n
+  br label %for.body3
+
+for.body3:                                        ; preds = %for.body3, %for.body
+  %indvars.iv = phi i64 [ 0, %for.body ], [ %indvars.iv.next, %for.body3 ]
+  %arrayidx7 = getelementptr inbounds [8 x [8 x i32]], [8 x [8 x i32]]* @arr, i64 0, i64 %indvars.iv, i64 %indvars.iv21
+  store i32 %add, i32* %arrayidx7, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp eq i64 %indvars.iv.next, 8
+  br i1 %exitcond, label %for.inc8, label %for.body3
+
+for.inc8:                                         ; preds = %for.body3
+  %indvars.iv.next22 = add nuw nsw i64 %indvars.iv21, 1
+  %exitcond23 = icmp eq i64 %indvars.iv.next22, 8
+  br i1 %exitcond23, label %for.end10, label %for.body, !llvm.loop !1
+
+for.end10:                                        ; preds = %for.inc8
+  ret void
+}
+
+!1 = distinct !{!1, !2}
+!2 = !{!"llvm.loop.vectorize.enable", i1 true}