[LV][nearly NFC] Move isLegalMasked* functions from Legality to CostModel

[hsaito]

Summary

Please see http://lists.llvm.org/pipermail/llvm-dev/2018-January/120164.html for RFC and the discussions.

All SIMD architectures can emulate masked load/store/gather/scatter through element-wise condition check, scalar load/store, and insert/extract. Therefore, bailing out of vectorization as legality failure, when they return false, is incorrect. We should proceed to cost model and determine profitability.

This patch is to address the vectorizer's architectural limitation described above. As such, I tried to keep the cost model and vectorize/don't-vectorize behavior nearly unchanged. Cost model tuning should be done separately.

[hsaito]

lib/Transforms/Vectorize/LoopVectorize.cpp

2042–2047

dyn_cast<> abuse fixed. Code simplified a bit. Should be NFC.

6202

Do we even want to keep this bail out?

test/Transforms/LoopVectorize/conditional-assignment.ll

4

I think the remark emission needs to be fixed so that it includes "loop not vectorized:". If so, I'll do that in a separate patch.

[rengolin]

lib/Transforms/Vectorize/LoopVectorize.cpp

6202

I think this is done more as diagnosis than bail out.

6586

What if this was a hook into the TTI, it could use the targets' information to get a slightly more accurate cost estimate (ie, how hacky is the emulation, or even if one is available through its back-end)

6588

high / high

[hsaito]

lib/Transforms/Vectorize/LoopVectorize.cpp

6202

In my opinion, this is a bail out in the sense that we blindly set VF=1 and say "go scalar". What's special about conditional stores compared to other "less than optimal" vector operations? We don't have such "enable" flags for every single one of them. Do we still consider the code to support conditional stores semi-"experimental"?

I just wanted to point that out. If anyone wants to keep this, I won't insist.

6586

Thanks. That's actually my preferred solution (http://lists.llvm.org/pipermail/llvm-dev/2018-January/120271.html) and Hal Finkel agrees (http://lists.llvm.org/pipermail/llvm-dev/2018-January/120273.html). If there are others feeling the same way (or opposite way), please add your voice here.

Having said that, is it okay for me to take the TTI approach as a separate patch submission? I'm sure TTI discussion will involve more stakeholders and the bulk of that discussion is tangent from fixing vectorizer's architectural issue related to isLegalMasked* this patch is trying to address.

This is especially so with the hacky NumberOfStoresToPredicate getting exposed outside of LoopVectorize code. I could imagine some reviewers would be asking us vectorizer developers to "tune the cost first so that the hacky NumberOfStoresToPredicate doesn't have to be leaked out of LoopVectorize". I don't think we should wait fixing the architectural problem until tuning is done.

6588

Will fix. Thanks.

[rengolin]

lib/Transforms/Vectorize/LoopVectorize.cpp

6202

Can you measure any difference? Like, running with and without on the test-suite to see if there are any hits that the new cost model is not picking it up?

6586

The way we've done in the past was to add the function we expect in TTI and implement it in the base class with a big FIXME, explaining the problem. In this way, targets can "FIXIT" in different ways at different times.

But I'm not against the idea of this hack, if that's the consensus (just had't followed the original thread :).

[hsaito]

lib/Transforms/Vectorize/LoopVectorize.cpp

6202

EnableCondStoresVectorization is true by default. So, this IF is no-op unless someone is explicitly setting it to false.

There are a few LIT tests checking for letting at least one emulated masked store vectorized.

I saw lots of perf regressions when I ran tests prior to adding cost model hooks (i.e., same cost as the first emulated masked store). On the opposite side of spectrum, if I remember correctly, I didn't see any of our perf tests regressing when I had all emulated masked load/store super high cost ---- that is a very easy way to get the code bit-rottened and thus I won't recommend it.

6586

You are the third person saying that the cost hack should be ideally coming from TTI (the other two are Hal and myself). No other opinions explicitly mentioned yet.

The basic idea of the hack (whether we implement in LoopVectorize or in TTI) is inevitable to make this patch "almost NFC". The TTI version I'm thinking about would have the same basic idea (i.e., return low cost for the first <NumberOfStoresToPredicate> emulated masked stores and super high cost for the rest) ---- because that's the essence of the code prior to this patch.

For those who are reviewing the TTI version, it would be a lot easier if the patch does not include isLegalMasked* change (from Legality to CostModel). That's another reason why I'd like to do it as a separate patch.

[hsaito]

high high ==> high

[rengolin]

LGTM, thanks!

lib/Transforms/Vectorize/LoopVectorize.cpp

6202

I agree, better to drop this check, then.

6586

Makes sense.

[hsaito]

In D43208#1018661, @rengolin wrote:

LGTM, thanks!

Thanks a lot. I do not have a commit-after-review right. Would be great if you'd commit this for me.

[rengolin]

r326079, thanks!

[hsaito]

In D43208#1019126, @rengolin wrote:

r326079, thanks!

Thanks!

lib/Transforms/Vectorize/LoopVectorize.cpp

public:
   unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
 
   uint64_t getMaxSafeRegisterWidth() const {
     return LAI->getDepChecker().getMaxSafeRegisterWidth();
   }
 
   bool hasStride(Value *V) { return LAI->hasStride(V); }
 
-  /// Returns true if the target machine supports masked store operation
-  /// for the given \p DataType and kind of access to \p Ptr.
-  bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
-    return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
-  }
-
-  /// Returns true if the target machine supports masked load operation
-  /// for the given \p DataType and kind of access to \p Ptr.
-  bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
-    return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
-  }
-
-  /// Returns true if the target machine supports masked scatter operation
-  /// for the given \p DataType.
-  bool isLegalMaskedScatter(Type *DataType) {
-    return TTI->isLegalMaskedScatter(DataType);
-  }
-
-  /// Returns true if the target machine supports masked gather operation
-  /// for the given \p DataType.
-  bool isLegalMaskedGather(Type *DataType) {
-    return TTI->isLegalMaskedGather(DataType);
-  }
-
-  /// Returns true if the target machine can represent \p V as a masked gather
-  /// or scatter operation.
-  bool isLegalGatherOrScatter(Value *V) {
-    auto *LI = dyn_cast<LoadInst>(V);
-    auto *SI = dyn_cast<StoreInst>(V);
-    if (!LI && !SI)
-      return false;
-    auto *Ptr = getPointerOperand(V);
-    auto *Ty = cast<PointerType>(Ptr->getType())->getElementType();
-    return (LI && isLegalMaskedGather(Ty)) || (SI && isLegalMaskedScatter(Ty));
-  }
-
   /// Returns true if vector representation of the instruction \p I
   /// requires mask.
   bool isMaskRequired(const Instruction *I) { return (MaskedOp.count(I) != 0); }
 
   unsigned getNumStores() const { return LAI->getNumStores(); }
   unsigned getNumLoads() const { return LAI->getNumLoads(); }
-  unsigned getNumPredStores() const { return NumPredStores; }
-
-  /// Returns true if \p I is an instruction that will be scalarized with
-  /// predication. Such instructions include conditional stores and
-  /// instructions that may divide by zero.
-  bool isScalarWithPredication(Instruction *I);
-
-  /// Returns true if \p I is a memory instruction with consecutive memory
-  /// access that can be widened.
-  bool memoryInstructionCanBeWidened(Instruction *I, unsigned VF = 1);
 
   // Returns true if the NoNaN attribute is set on the function.
   bool hasFunNoNaNAttr() const { return HasFunNoNaNAttr; }
 
 private:
   /// Check if a single basic block loop is vectorizable.
   /// At this point we know that this is a loop with a constant trip count
   /// and we only need to check individual instructions.
private:
   const ValueToValueMap *getSymbolicStrides() {
     // FIXME: Currently, the set of symbolic strides is sometimes queried before
     // it's collected.  This happens from canVectorizeWithIfConvert, when the
     // pointer is checked to reference consecutive elements suitable for a
     // masked access.
     return LAI ? &LAI->getSymbolicStrides() : nullptr;
   }
 
-  unsigned NumPredStores = 0;
-
   /// The loop that we evaluate.
   Loop *TheLoop;
 
   /// A wrapper around ScalarEvolution used to add runtime SCEV checks.
   /// Applies dynamic knowledge to simplify SCEV expressions in the context
   /// of existing SCEV assumptions. The analysis will also add a minimal set
   /// of new predicates if this is required to enable vectorization and
   /// unrolling.
void collectUniformsAndScalars(unsigned VF) {
     // Do the analysis once.
     if (VF == 1 || Uniforms.count(VF))
       return;
     setCostBasedWideningDecision(VF);
     collectLoopUniforms(VF);
     collectLoopScalars(VF);
   }
 
+  /// Returns true if the target machine supports masked store operation
+  /// for the given \p DataType and kind of access to \p Ptr.
+  bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
+    return Legal->isConsecutivePtr(Ptr) && TTI.isLegalMaskedStore(DataType);
+  }
+
+  /// Returns true if the target machine supports masked load operation
+  /// for the given \p DataType and kind of access to \p Ptr.
+  bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
+    return Legal->isConsecutivePtr(Ptr) && TTI.isLegalMaskedLoad(DataType);
+  }
+
+  /// Returns true if the target machine supports masked scatter operation
+  /// for the given \p DataType.
+  bool isLegalMaskedScatter(Type *DataType) {
+    return TTI.isLegalMaskedScatter(DataType);
+  }
+
+  /// Returns true if the target machine supports masked gather operation
+  /// for the given \p DataType.
+  bool isLegalMaskedGather(Type *DataType) {
+    return TTI.isLegalMaskedGather(DataType);
+  }
+
+  /// Returns true if the target machine can represent \p V as a masked gather
+  /// or scatter operation.
+  bool isLegalGatherOrScatter(Value *V) {
+    bool LI = isa<LoadInst>(V);
+    bool SI = isa<StoreInst>(V);
+    if (!LI && !SI)
+      return false;
+    auto *Ty = getMemInstValueType(V);
+    return (LI && isLegalMaskedGather(Ty)) || (SI && isLegalMaskedScatter(Ty));
+  }
+
+  /// Returns true if \p I is an instruction that will be scalarized with
+  /// predication. Such instructions include conditional stores and
+  /// instructions that may divide by zero.
+  bool isScalarWithPredication(Instruction *I);
+
+  /// Returns true if \p I is a memory instruction with consecutive memory
+  /// access that can be widened.
+  bool memoryInstructionCanBeWidened(Instruction *I, unsigned VF = 1);
+
 private:
+  unsigned NumPredStores = 0;
+
   /// \return An upper bound for the vectorization factor, larger than zero.
   /// One is returned if vectorization should best be avoided due to cost.
   unsigned computeFeasibleMaxVF(bool OptForSize, unsigned ConstTripCount);
 
   /// 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
private:
   /// The cost calculation for Load instruction \p I with uniform pointer -
   /// scalar load + broadcast.
   unsigned getUniformMemOpCost(Instruction *I, unsigned VF);
 
   /// Returns whether the instruction is a load or store and will be a emitted
   /// as a vector operation.
   bool isConsecutiveLoadOrStore(Instruction *I);
 
+  /// Returns true if an artificially high cost for emulated masked memrefs
+  /// should be used.
+  bool useEmulatedMaskMemRefHack(Instruction *I);
+
   /// Create an analysis remark that explains why vectorization failed
   ///
   /// \p RemarkName is the identifier for the remark.  \return the remark object
   /// that can be streamed to.
   OptimizationRemarkAnalysis createMissedAnalysis(StringRef RemarkName) {
     return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
                                   RemarkName, TheLoop);
   }
for (auto &Induction : *Legal->getInductionVars()) {
     Worklist.insert(IndUpdate);
     DEBUG(dbgs() << "LV: Found scalar instruction: " << *Ind << "\n");
     DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n");
   }
 
   Scalars[VF].insert(Worklist.begin(), Worklist.end());
 }
 
-bool LoopVectorizationLegality::isScalarWithPredication(Instruction *I) {
-  if (!blockNeedsPredication(I->getParent()))
+bool LoopVectorizationCostModel::isScalarWithPredication(Instruction *I) {
+  if (!Legal->blockNeedsPredication(I->getParent()))
     return false;
   switch(I->getOpcode()) {
   default:
     break;
-  case Instruction::Store:
-    return !isMaskRequired(I);
+  case Instruction::Load:
+  case Instruction::Store: {
+    if (!Legal->isMaskRequired(I))
+      return false;
+    auto *Ptr = getPointerOperand(I);
+    auto *Ty = getMemInstValueType(I);
+    return isa<LoadInst>(I) ?
+        !(isLegalMaskedLoad(Ty, Ptr)  || isLegalMaskedGather(Ty))
+      : !(isLegalMaskedStore(Ty, Ptr) || isLegalMaskedScatter(Ty));
+  }
   case Instruction::UDiv:
   case Instruction::SDiv:
   case Instruction::SRem:
   case Instruction::URem:
     return mayDivideByZero(*I);
   }
   return false;
 }
 
-bool LoopVectorizationLegality::memoryInstructionCanBeWidened(Instruction *I,
+bool LoopVectorizationCostModel::memoryInstructionCanBeWidened(Instruction *I,
                                                                unsigned VF) {
   // Get and ensure we have a valid memory instruction.
   LoadInst *LI = dyn_cast<LoadInst>(I);
   StoreInst *SI = dyn_cast<StoreInst>(I);
   assert((LI || SI) && "Invalid memory instruction");
 
   auto *Ptr = getPointerOperand(I);
 
   // In order to be widened, the pointer should be consecutive, first of all.
-  if (!isConsecutivePtr(Ptr))
+  if (!Legal->isConsecutivePtr(Ptr))
     return false;
 
   // If the instruction is a store located in a predicated block, it will be
   // scalarized.
   if (isScalarWithPredication(I))
     return false;
 
   // If the instruction's allocated size doesn't equal it's type size, it
for (Value *Operand : I.operands()) {
           return false;
     }
     // We might be able to hoist the load.
     if (I.mayReadFromMemory()) {
       auto *LI = dyn_cast<LoadInst>(&I);
       if (!LI)
         return false;
       if (!SafePtrs.count(LI->getPointerOperand())) {
-        if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand()) ||
-            isLegalMaskedGather(LI->getType())) {
-          MaskedOp.insert(LI);
-          continue;
-        }
         // !llvm.mem.parallel_loop_access implies if-conversion safety.
-        if (IsAnnotatedParallel)
-          continue;
-        return false;
+        // Otherwise, record that the load needs (real or emulated) masking
+        // and let the cost model decide.
+        if (!IsAnnotatedParallel)
+          MaskedOp.insert(LI);
       }
     }
 
     if (I.mayWriteToMemory()) {
       auto *SI = dyn_cast<StoreInst>(&I);
-      // We only support predication of stores in basic blocks with one
-      // predecessor.
       if (!SI)
         return false;
-
-      // Build a masked store if it is legal for the target.
-      if (isLegalMaskedStore(SI->getValueOperand()->getType(),
-                             SI->getPointerOperand()) ||
-          isLegalMaskedScatter(SI->getValueOperand()->getType())) {
-        MaskedOp.insert(SI);
-        continue;
-      }
-
-      bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
-      bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
-
-      if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
-          !isSinglePredecessor)
-        return false;
-    }
+      // Predicated store requires some form of masking:
+      // 1) masked store HW instruction,
+      // 2) emulation via load-blend-store (only if safe and legal to do so,
+      //    be aware on the race conditions), or
+      // 3) element-by-element predicate check and scalar store.
     if (I.mayThrow())
       return false;
   }
 
   return true;
 }
 
 void InterleavedAccessInfo::collectConstStrideAccesses(
if (LastMember) {
       }
       DEBUG(dbgs() << "LV: Interleaved group requires epilogue iteration.\n");
       RequiresScalarEpilogue = true;
     }
   }
 }
 
 Optional<unsigned> LoopVectorizationCostModel::computeMaxVF(bool OptForSize) {
-  if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
-    ORE->emit(createMissedAnalysis("ConditionalStore")
-              << "store that is conditionally executed prevents vectorization");
-    DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
-    return None;
-  }
-
   if (Legal->getRuntimePointerChecking()->Need && TTI.hasBranchDivergence()) {
     // TODO: It may by useful to do since it's still likely to be dynamically
     // uniform if the target can skip.
     DEBUG(dbgs() << "LV: Not inserting runtime ptr check for divergent target");
 
     ORE->emit(
       createMissedAnalysis("CantVersionLoopWithDivergentTarget")
       << "runtime pointer checks needed. Not enabled for divergent target");
if (MaximizeBandwidth && !OptForSize) {
     }
   }
   return MaxVF;
 }
 
 VectorizationFactor
 LoopVectorizationCostModel::selectVectorizationFactor(unsigned MaxVF) {
   float Cost = expectedCost(1).first;
-#ifndef NDEBUG
   const float ScalarCost = Cost;
-#endif /* NDEBUG */
   unsigned Width = 1;
   DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
 
   bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
   // Ignore scalar width, because the user explicitly wants vectorization.
   if (ForceVectorization && MaxVF > 1) {
     Width = 2;
     Cost = expectedCost(Width).first / (float)Width;
if (!C.second && !ForceVectorization) {
       continue;
     }
     if (VectorCost < Cost) {
       Cost = VectorCost;
       Width = i;
     }
   }
 
+  if (!EnableCondStoresVectorization && NumPredStores) {
+    ORE->emit(createMissedAnalysis("ConditionalStore")
+              << "store that is conditionally executed prevents vectorization");
+    DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
+    Width = 1;
+    Cost = ScalarCost;
+  }
+
   DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
         << "LV: Vectorization seems to be not beneficial, "
         << "but was forced by a user.\n");
   DEBUG(dbgs() << "LV: Selecting VF: " << Width << ".\n");
   VectorizationFactor Factor = {Width, (unsigned)(Width * Cost)};
   return Factor;
 }
 
for (Instruction &I : *BB) {
       //
       // FIXME: The check here attempts to predict whether a load or store will
       //        be vectorized. We only know this for certain after a VF has
       //        been selected. Here, we assume that if an access can be
       //        vectorized, it will be. We should also look at extending this
       //        optimization to non-pointer types.
       //
       if (T->isPointerTy() && !isConsecutiveLoadOrStore(&I) &&
-          !Legal->isAccessInterleaved(&I) && !Legal->isLegalGatherOrScatter(&I))
+          !Legal->isAccessInterleaved(&I) && !isLegalGatherOrScatter(&I))
         continue;
 
       MinWidth = std::min(MinWidth,
                           (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
       MaxWidth = std::max(MaxWidth,
                           (unsigned)DL.getTypeSizeInBits(T->getScalarType()));
     }
   }
for (unsigned i = 0, e = VFs.size(); i < e; ++i) {
     RU.LoopInvariantRegs = Invariant;
     RU.MaxLocalUsers = MaxUsages[i];
     RUs[i] = RU;
   }
 
   return RUs;
 }
 
+bool LoopVectorizationCostModel::useEmulatedMaskMemRefHack(Instruction *I){
+  // TODO: Cost model for emulated masked load/store is completely
+  // broken. This hack guides the cost model to use an artificially high
+  // high enough value to practically disable vectorization with such
+  // operations, except where previously deployed legality hack allowed
+  // using very low cost values. This is to avoid regressions coming simply
+  // from moving "masked load/store" check from legality to cost model. 
+  // Masked Load/Gather emulation was previously never allowed.
+  // Limited number of Masked Store/Scatter emulation was allowed.
+  assert(isScalarWithPredication(I) &&
+         "Expecting a scalar emulated instruction");
+  return isa<LoadInst>(I) ||
+         (isa<StoreInst>(I) &&
+          NumPredStores > NumberOfStoresToPredicate);
+}
+
 void LoopVectorizationCostModel::collectInstsToScalarize(unsigned VF) {
   // If we aren't vectorizing the loop, or if we've already collected the
   // instructions to scalarize, there's nothing to do. Collection may already
   // have occurred if we have a user-selected VF and are now computing the
   // expected cost for interleaving.
   if (VF < 2 || InstsToScalarize.count(VF))
     return;
 
   // Initialize a mapping for VF in InstsToScalalarize. If we find that it's
   // not profitable to scalarize any instructions, the presence of VF in the
   // map will indicate that we've analyzed it already.
   ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF];
 
   // Find all the instructions that are scalar with predication in the loop and
   // 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 (!Legal->blockNeedsPredication(BB))
       continue;
     for (Instruction &I : *BB)
-      if (Legal->isScalarWithPredication(&I)) {
+      if (isScalarWithPredication(&I)) {
         ScalarCostsTy ScalarCosts;
-        if (computePredInstDiscount(&I, ScalarCosts, VF) >= 0)
+        // Do not apply discount logic if hacked cost is needed
+        // for emulated masked memrefs.
+        if (!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, DenseMap<Instruction *, unsigned> &ScalarCosts,
auto canBeScalarized = [&](Instruction *I) -> bool {
     // already be scalar to avoid traversing chains that are unlikely to be
     // beneficial.
     if (!I->hasOneUse() || PredInst->getParent() != I->getParent() ||
         isScalarAfterVectorization(I, VF))
       return false;
 
     // If the instruction is scalar with predication, it will be analyzed
     // separately. We ignore it within the context of PredInst.
-    if (Legal->isScalarWithPredication(I))
+    if (isScalarWithPredication(I))
       return false;
 
     // If any of the instruction's operands are uniform after vectorization,
     // the instruction cannot be scalarized. This prevents, for example, a
     // masked load from being scalarized.
     //
     // We assume we will only emit a value for lane zero of an instruction
     // marked uniform after vectorization, rather than VF identical values.
while (!Worklist.empty()) {
     // 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.
     unsigned ScalarCost = VF * getInstructionCost(I, 1).first;
 
     // Compute the scalarization overhead of needed insertelement instructions
     // and phi nodes.
-    if (Legal->isScalarWithPredication(I) && !I->getType()->isVoidTy()) {
+    if (isScalarWithPredication(I) && !I->getType()->isVoidTy()) {
       ScalarCost += TTI.getScalarizationOverhead(ToVectorTy(I->getType(), VF),
                                                  true, false);
       ScalarCost += VF * TTI.getCFInstrCost(Instruction::PHI);
     }
 
     // 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
unsigned LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I,
 
   // Get the overhead of the extractelement and insertelement instructions
   // we might create due to scalarization.
   Cost += getScalarizationOverhead(I, VF, TTI);
 
   // If we have a predicated store, it may not be executed for each vector
   // lane. Scale the cost by the probability of executing the predicated
   // block.
-  if (Legal->isScalarWithPredication(I))
+  if (isScalarWithPredication(I)) {
     Cost /= getReciprocalPredBlockProb();
 
+    if (useEmulatedMaskMemRefHack(I))
+      // Artificially setting to a high enough value to practically disable
+      // vectorization with such operations.
+      Cost = 3000000;
+  }
+
   return Cost;
 }
 
 unsigned LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I,
                                                              unsigned VF) {
   Type *ValTy = getMemInstValueType(I);
   Type *VectorTy = ToVectorTy(ValTy, VF);
   unsigned Alignment = getMemInstAlignment(I);
LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
   bool TypeNotScalarized =
       VF > 1 && VectorTy->isVectorTy() && TTI.getNumberOfParts(VectorTy) < VF;
   return VectorizationCostTy(C, TypeNotScalarized);
 }
 
 void LoopVectorizationCostModel::setCostBasedWideningDecision(unsigned VF) {
   if (VF == 1)
     return;
+  NumPredStores = 0;
   for (BasicBlock *BB : TheLoop->blocks()) {
     // For each instruction in the old loop.
     for (Instruction &I : *BB) {
       Value *Ptr = getPointerOperand(&I);
       if (!Ptr)
         continue;
 
+      if (isa<StoreInst>(&I) && isScalarWithPredication(&I))
+        NumPredStores++;
       if (isa<LoadInst>(&I) && Legal->isUniform(Ptr)) {
         // Scalar load + broadcast
         unsigned Cost = getUniformMemOpCost(&I, VF);
         setWideningDecision(&I, VF, CM_Scalarize, Cost);
         continue;
       }
 
       // We assume that widening is the best solution when possible.
-      if (Legal->memoryInstructionCanBeWidened(&I, VF)) {
+      if (memoryInstructionCanBeWidened(&I, VF)) {
         unsigned Cost = getConsecutiveMemOpCost(&I, VF);
         int ConsecutiveStride = Legal->isConsecutivePtr(getPointerOperand(&I));
         assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&
                "Expected consecutive stride.");
         InstWidening Decision =
             ConsecutiveStride == 1 ? CM_Widen : CM_Widen_Reverse;
         setWideningDecision(&I, VF, Decision, Cost);
         continue;
for (Instruction &I : *BB) {
         if (getWideningDecision(&I, VF) != CM_Unknown)
           continue;
 
         NumAccesses = Group->getNumMembers();
         InterleaveCost = getInterleaveGroupCost(&I, VF);
       }
 
       unsigned GatherScatterCost =
-          Legal->isLegalGatherOrScatter(&I)
+          isLegalGatherOrScatter(&I)
               ? getGatherScatterCost(&I, VF) * NumAccesses
               : std::numeric_limits<unsigned>::max();
 
       unsigned ScalarizationCost =
           getMemInstScalarizationCost(&I, VF) * NumAccesses;
 
       // Choose better solution for the current VF,
       // write down this decision and use it during vectorization.
unsigned LoopVectorizationCostModel::getInstructionCost(Instruction *I,
   case Instruction::UDiv:
   case Instruction::SDiv:
   case Instruction::URem:
   case Instruction::SRem:
     // If we have a predicated instruction, it may not be executed for each
     // vector lane. Get the scalarization cost and scale this amount by the
     // probability of executing the predicated block. If the instruction is not
     // predicated, we fall through to the next case.
-    if (VF > 1 && Legal->isScalarWithPredication(I)) {
+    if (VF > 1 && isScalarWithPredication(I)) {
       unsigned Cost = 0;
 
       // These instructions have a non-void type, so account for the phi nodes
       // that we will create. This cost is likely to be zero. The phi node
       // cost, if any, should be scaled by the block probability because it
       // models a copy at the end of each predicated block.
       Cost += VF * TTI.getCFInstrCost(Instruction::PHI);
 
for (unsigned In = 0; In < NumIncoming; In++) {
     if (EdgeMask)
       Masks.push_back(EdgeMask);
   }
   return new VPBlendRecipe(Phi, Masks);
 }
 
 bool LoopVectorizationPlanner::tryToWiden(Instruction *I, VPBasicBlock *VPBB,
                                           VFRange &Range) {
-  if (Legal->isScalarWithPredication(I))
+  if (CM.isScalarWithPredication(I))
     return false;
 
   auto IsVectorizableOpcode = [](unsigned Opcode) {
     switch (Opcode) {
     case Instruction::Add:
     case Instruction::And:
     case Instruction::AShr:
     case Instruction::BitCast:

 VPBasicBlock *LoopVectorizationPlanner::handleReplication(
     Instruction *I, VFRange &Range, VPBasicBlock *VPBB,
     DenseMap<Instruction *, VPReplicateRecipe *> &PredInst2Recipe,
     VPlanPtr &Plan) {
   bool IsUniform = getDecisionAndClampRange(
       [&](unsigned VF) { return CM.isUniformAfterVectorization(I, VF); },
       Range);
 
-  bool IsPredicated = Legal->isScalarWithPredication(I);
+  bool IsPredicated = CM.isScalarWithPredication(I);
   auto *Recipe = new VPReplicateRecipe(I, IsUniform, IsPredicated);
 
   // Find if I uses a predicated instruction. If so, it will use its scalar
   // value. Avoid hoisting the insert-element which packs the scalar value into
   // a vector value, as that happens iff all users use the vector value.
   for (auto &Op : I->operands())
     if (auto *PredInst = dyn_cast<Instruction>(Op))
       if (PredInst2Recipe.find(PredInst) != PredInst2Recipe.end())

test/Transforms/LoopVectorize/conditional-assignment.ll

 ; RUN: opt < %s -enable-cond-stores-vec=false -loop-vectorize -S -pass-remarks-missed='loop-vectorize' -pass-remarks-analysis='loop-vectorize' 2>&1 | FileCheck %s
 ; RUN: opt < %s -enable-cond-stores-vec=false -passes=loop-vectorize -S -pass-remarks-missed='loop-vectorize' -pass-remarks-analysis='loop-vectorize' 2>&1 | FileCheck %s
 
-; CHECK: remark: source.c:2:8: loop not vectorized: store that is conditionally executed prevents vectorization
+; CHECK: remark: source.c:2:8: the cost-model indicates that vectorization is not beneficial
 
 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
 
 ; Function Attrs: nounwind ssp uwtable
 define void @conditional_store(i32* noalias nocapture %indices) #0 !dbg !4 {
 entry:
   br label %for.body, !dbg !10
 

test/Transforms/LoopVectorize/hoist-loads.ll

if.end9:
   %exitcond = icmp ne i32 %lftr.wideiv, 1024
   br i1 %exitcond, label %for.body, label %for.end
 
 for.end:
   ret void
 }
 
 ; However, we can't hoist loads whose address we have not seen unconditionally
-; accessed.
+; accessed. One wide load is fine, but not the second.
 ; CHECK-LABEL: @dont_hoist_cond_load(
+; CHECK: load <2 x float>
 ; CHECK-NOT: load <2 x float>
 
 define void @dont_hoist_cond_load() {
 entry:
   br label %for.body
 for.body:
   %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %if.end9 ]
   %arrayidx = getelementptr inbounds [1024 x float], [1024 x float]* @A, i64 0, i64 %indvars.iv