Using Masked Load / Store intrinsics in Loop Vectorizer

[delena]

Summary

The loop vectorizer optimizes loops containing conditional memory accesses by generating masked load and store intrinsics.
This decision is target dependent.

I already submitted the codegen changes for the intrinsics.

[aschwaighofer]

Very nice! I only see minor fixes.

Thanks!

lib/Transforms/Vectorize/LoopVectorize.cpp

778

This is a predicate. Could you rename it to isMaskedOperation?

895

Should this be a SmallPtrSet?

5467

Should this be in a separate patch? This seems like an unrelated change.

test/Transforms/LoopVectorize/X86/mask1.ll

1 ↗

(On Diff #16924)

Could the mask1.ll, mask2.ll, etc be in one file? We try to not create to many .ll files when test cases are related.

Could you use

CHECK-LABEL: @function_name
CHECK: ... // rest of the test for function_name test

CHECK-LABEL: @function_name2
CHECK: ... // rest of the test for function_name2 test

Could you modify the test so that it tests more of the expected IR?

[nadav]

Hi Elena,

Thank you for working on this.

+ bool canPredicateStore(Type *DataType, Value *Ptr) {
+ return TTI->isLegalPredicatedStore(DataType, isConsecutivePtr(Ptr));
+ }
+ bool canPredicateLoad(Type *DataType, Value *Ptr) {
+ return TTI->isLegalPredicatedLoad(DataType, isConsecutivePtr(Ptr));
+ }
+ bool setMaskedOp(const Instruction* I) {
+ return (MaskedOp.find(I) != MaskedOp.end());
+ }
private:

Can you please document these functions? The name setMaskedOp is confusing and Doxygen style comments could be useful here.

Thanks,
Nadav

[mzolotukhin]

Hi Elena,

Please see a question from me in inline comments. And thanks for doing this!

Michael

lib/Transforms/Vectorize/LoopVectorize.cpp

5393–5394

Could this condition be simplified? E.g. we can store result of canPredicateStore(..) in a bool variable and use it here.
Also, I think it might be incorrect - shouldn't we insert SI to MaskedOp even if NumPredStores hasn't exceeded NumberOfStoresToPredicate?

[delena]

I'm adding comments, changing function names, joining 4 tests in one.

lib/Transforms/Vectorize/LoopVectorize.cpp

778

do you mean isLegalMaskedLoad()? isLegalMaskedStore()?

895

What size should I give in this case? I don't know how many masked operations will be in this block.
Or it is not so important?

5393–5394

If NumPredStores hasn't exceeded NumberOfStoresToPredicate I'll not generate masked store. I want to keep the current behavior.

Something like this?

bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) !=0);
bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();

if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
    !isSinglePredecessor) {
  // Build a masked store if it is legal for the target, otherwise scalarize
  // the block.
  bool isLegalMaskedOp = canPredicateStore(SI->getValueOperand()->getType(),
                                           SI->getPointerOperand());
  if (isLegalMaskedOp) {
    --NumPredStores;
    MaskedOp.insert(SI);
    continue;
  }
  return false;
}

5467

My tests don't work without this patch because 512 / 8 = 64

test/Transforms/LoopVectorize/X86/mask1.ll

1 ↗

(On Diff #16924)

no problem

[delena]

Thank you all for reviewing this. I addressed all comments and applying the updated patch.

[nadav]

+ / Returns true if vector representation of the instruction \p I
+ / requires mask.
+ bool toBuildMaskedVectorInst(const Instruction* I) {
+ return (MaskedOp.count(I) != 0);
+ }
+ void SI();

Maybe a better name would be "isMaskRequired"?

[mzolotukhin]

Hi Elena,

Please see my comments inline.

lib/Transforms/Vectorize/LoopVectorize.cpp

5393–5394

Actually, I wanted to get rid of ++NumPredStores, followed by --NumPredStores.
I'd rewrite it like this:

if (!isSafePtr || !isSinglePredecessor) {
    if (isLegalMaskedOp) {
        // This store will be masked, no predication is needed for it.
        MaskedOp.insert(SI);
        continue;
    } else {
        // This store needs to be predicated.
        if (++NumPredStores > NumberOfStoresToPredicate)
            return false;
    }
}

Also, I wonder if we actually need a 'continue' statement in if(isLegalMaskedOp) - do you intentionally skip subsequent checks (e.g. it->mayThrow) for this case?

[delena]

• Elena

[delena]

To Michael:
Some tests fail in this case, I tried. Because NumberOfStoresToPredicate allows you to predicated some stores without masking.
I wanted to keep this knob as is.
I also do not check mayThrow() because it is for calls only.

To Nadav:

Maybe a better name would be "isMaskRequired"?

No problem

[delena]

Michael is right about potentially trapping constant expression. I addressed this issue and also added a test for it.

[nadav]

Some changes in the patch should go in separately. For example:

• return false;

+ return false;

  }
}

This is a whitespace fix. Just commit it as is.

• assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"

+ assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"

" into one vector!");

Also the assert change can go in.

+ void SI();
private:

What is SI() ??

I would also split the TTI changes into a separate patch. It can probably go in already.

[mzolotukhin]

Some more minor notes from me.

lib/Analysis/TargetTransformInfo.cpp

104–105 ↗

(On Diff #17150)

Please fix indentation here.

109–110 ↗

(On Diff #17150)

And here.

lib/Transforms/Vectorize/LoopVectorize.cpp

1913

This also might be committed separately.

[delena]

This is the diff after all minor changes went in.

[mzolotukhin]

Hi Elena,

Thank you for checking-in the first (small) part of the changes.

The code looks good to me except some minor nits (see inline), but I'd wait for a LGTM from Nadav or Arnold.

lib/Transforms/Vectorize/LoopVectorize.cpp

1884–1885

Please fix the indentation here.

1896–1897

'}' and 'else' should be on the same line.

1932–1933

And here too.

test/Transforms/LoopVectorize/X86/masked_load_store.ll

8

Do we really need the target triple, or could it be omitted?

12–19

Do we need a test for i64 case as well?

26

I believe this is unnecessary. We check for AVX2-LABEL here and later in other functions, so all AVX2-instructions should be between the two labels. Thus, the next label would designate the end of the current function, and there is no need to explicitly look for 'ret void'.

Also, I wonder if it's possible that basic blocks are printed in a different order - in this case we might get a fail even though the loop is correctly vectorized.

lib/Transforms/Vectorize/LoopVectorize.cpp

 class LoopVectorizationLegality {
 public:
   unsigned NumLoads;
   unsigned NumStores;
   unsigned NumPredStores;
 
   LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
                             DominatorTree *DT, TargetLibraryInfo *TLI,
-                            AliasAnalysis *AA, Function *F)
+                            AliasAnalysis *AA, Function *F,
+                            const TargetTransformInfo *TTI)
       : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
-        DT(DT), TLI(TLI), AA(AA), TheFunction(F), Induction(nullptr),
+        DT(DT), TLI(TLI), AA(AA), TheFunction(F), TTI(TTI), Induction(nullptr),
         WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
   }
 
   /// This enum represents the kinds of reductions that we support.
   enum ReductionKind {
     RK_NoReduction, ///< Not a reduction.
     RK_IntegerAdd,  ///< Sum of integers.
     RK_IntegerMult, ///< Product of integers.
public:
 
   bool hasStride(Value *V) { return StrideSet.count(V); }
   bool mustCheckStrides() { return !StrideSet.empty(); }
   SmallPtrSet<Value *, 8>::iterator strides_begin() {
     return StrideSet.begin();
   }
   SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
 
+  /// 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 TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
+  }
+  /// 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 TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
+  }
+  /// Returns true if vector representation of the instruction \p I
+  /// requires mask.
+  bool isMaskRequired(const Instruction* I) {
+    return (MaskedOp.count(I) != 0);
+  }
 private:
   /// Check if a single basic block loop is vectorizable.
   /// At this point we know that this is a loop with a constant trip count
   /// and we only need to check individual instructions.
   bool canVectorizeInstrs();
 
   /// When we vectorize loops we may change the order in which
   /// we read and write from memory. This method checks if it is
private:
   /// Dominators.
   DominatorTree *DT;
   /// Target Library Info.
   TargetLibraryInfo *TLI;
   /// Alias analysis.
   AliasAnalysis *AA;
   /// Parent function
   Function *TheFunction;
+  /// Target Transform Info
+  const TargetTransformInfo *TTI;
 
   //  ---  vectorization state --- //
 
   /// Holds the integer induction variable. This is the counter of the
   /// loop.
   PHINode *Induction;
   /// Holds the reduction variables.
   ReductionList Reductions;
private:
   RuntimePointerCheck PtrRtCheck;
   /// Can we assume the absence of NaNs.
   bool HasFunNoNaNAttr;
 
   unsigned MaxSafeDepDistBytes;
 
   ValueToValueMap Strides;
   SmallPtrSet<Value *, 8> StrideSet;
+  
+  /// While vectorizing these instructions we have to generate a
+  /// call to the appropriate masked intrinsic
+  SmallPtrSet<const Instruction*, 8> MaskedOp;
 };
 
 /// LoopVectorizationCostModel - estimates the expected speedups due to
 /// vectorization.
 /// In many cases vectorization is not profitable. This can happen because of
 /// a number of reasons. In this class we mainly attempt to predict the
 /// expected speedup/slowdowns due to the supported instruction set. We use the
 /// TargetTransformInfo to query the different backends for the cost of
if (TC > 0u && TC < TinyTripCountVectorThreshold) {
         emitOptimizationRemarkAnalysis(
             F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
             "vectorization is not beneficial and is not explicitly forced");
         return false;
       }
     }
 
     // Check if it is legal to vectorize the loop.
-    LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F);
+    LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
     if (!LVL.canVectorize()) {
       DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
       emitMissedWarning(F, L, Hints);
       return false;
     }
 
     // Use the cost model.
     LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AT, F,
void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
   // An alignment of 0 means target abi alignment. We need to use the scalar's
   // target abi alignment in such a case.
   if (!Alignment)
     Alignment = DL->getABITypeAlignment(ScalarDataTy);
   unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
   unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
   unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
 
-  if (SI && Legal->blockNeedsPredication(SI->getParent()))
+  if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
+      !Legal->isMaskRequired(SI))
     return scalarizeInstruction(Instr, true);
 
   if (ScalarAllocatedSize != VectorElementSize)
     return scalarizeInstruction(Instr);
 
   // If the pointer is loop invariant or if it is non-consecutive,
   // scalarize the load.
   int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
for (unsigned Part = 0; Part < UF; ++Part) {
         // If the address is consecutive but reversed, then the
         // wide store needs to start at the last vector element.
         PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
         PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
       }
 
       Value *VecPtr = Builder.CreateBitCast(PartPtr,
                                             DataTy->getPointerTo(AddressSpace));
-      StoreInst *NewSI =
-        Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
+
+      Instruction *NewSI;
+      if (Legal->isMaskRequired(SI)) {
+        Type *I8PtrTy =
+        Builder.getInt8PtrTy(PartPtr->getType()->getPointerAddressSpace());
+
+        Value *I8Ptr = Builder.CreateBitCast(PartPtr, I8PtrTy);
+
+        VectorParts Cond = createBlockInMask(SI->getParent());
+        SmallVector <Value *, 8> Ops;
+        Ops.push_back(I8Ptr);
+        Ops.push_back(StoredVal[Part]);
+        Ops.push_back(Builder.getInt32(Alignment));
+        Ops.push_back(Cond[Part]);
+        NewSI = Builder.CreateMaskedStore(Ops);
+      }
+      else 
+        NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
       propagateMetadata(NewSI, SI);
     }
     return;
   }
 
   // Handle loads.
   assert(LI && "Must have a load instruction");
   setDebugLocFromInst(Builder, LI);
   for (unsigned Part = 0; Part < UF; ++Part) {
     // Calculate the pointer for the specific unroll-part.
     Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
 
     if (Reverse) {
       // If the address is consecutive but reversed, then the
       // wide load needs to start at the last vector element.
       PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
       PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
     }
 
+    Instruction* NewLI;
+    if (Legal->isMaskRequired(LI)) {
+      Type *I8PtrTy =
+        Builder.getInt8PtrTy(PartPtr->getType()->getPointerAddressSpace());
+
+      Value *I8Ptr = Builder.CreateBitCast(PartPtr, I8PtrTy);
+
+      VectorParts SrcMask = createBlockInMask(LI->getParent());
+      SmallVector <Value *, 8> Ops;
+      Ops.push_back(I8Ptr);
+      Ops.push_back(UndefValue::get(DataTy));
+      Ops.push_back(Builder.getInt32(Alignment));
+      Ops.push_back(SrcMask[Part]);
+      NewLI = Builder.CreateMaskedLoad(Ops);
+    }
+    else {
       Value *VecPtr = Builder.CreateBitCast(PartPtr,
                                             DataTy->getPointerTo(AddressSpace));
-    LoadInst *NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
+      NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
+    }
     propagateMetadata(NewLI, LI);
     Entry[Part] = Reverse ? reverseVector(NewLI) :  NewLI;
   }
 }
 
 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
   assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
   // Holds vector parameters or scalars, in case of uniform vals.
bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB)  {
 
   // Blocks that do not dominate the latch need predication.
   BasicBlock* Latch = TheLoop->getLoopLatch();
   return !DT->dominates(BB, Latch);
 }
 
 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
                                            SmallPtrSetImpl<Value *> &SafePtrs) {
+  
   for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
+    // Check that we don't have a constant expression that can trap as operand.
+    for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
+         OI != OE; ++OI) {
+      if (Constant *C = dyn_cast<Constant>(*OI))
+        if (C->canTrap())
+          return false;
+    }
     // We might be able to hoist the load.
     if (it->mayReadFromMemory()) {
       LoadInst *LI = dyn_cast<LoadInst>(it);
-      if (!LI || !SafePtrs.count(LI->getPointerOperand()))
+      if (!LI)
         return false;
+      if (!SafePtrs.count(LI->getPointerOperand())) {
+        if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
+          MaskedOp.insert(LI);
+          continue;
+        }
+        return false;
+      }
     }
 
     // We don't predicate stores at the moment.
     if (it->mayWriteToMemory()) {
       StoreInst *SI = dyn_cast<StoreInst>(it);
       // We only support predication of stores in basic blocks with one
       // predecessor.
-      if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
-          !SafePtrs.count(SI->getPointerOperand()) ||
-          !SI->getParent()->getSinglePredecessor())
-        return false;
-    }
-    if (it->mayThrow())
+      if (!SI)
         return false;
 
-    // Check that we don't have a constant expression that can trap as operand.
-    for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
-         OI != OE; ++OI) {
-      if (Constant *C = dyn_cast<Constant>(*OI))
-        if (C->canTrap())
-          return false;
+      bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
+      bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
+      
+      if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
+          !isSinglePredecessor) {
+        // Build a masked store if it is legal for the target, otherwise scalarize
+        // the block.
+        bool isLegalMaskedOp =
+          isLegalMaskedStore(SI->getValueOperand()->getType(),
+                             SI->getPointerOperand());
+        if (isLegalMaskedOp) {
+          --NumPredStores;
+          MaskedOp.insert(SI);
+          continue;
+        }
       }
+    }
+    if (it->mayThrow())
+      return false;
 
     // The instructions below can trap.
     switch (it->getOpcode()) {
     default: continue;
     case Instruction::UDiv:
     case Instruction::SDiv:
     case Instruction::URem:
     case Instruction::SRem:
LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
   DEBUG(dbgs() << "LV: The Widest register is: "
           << WidestRegister << " bits.\n");
 
   if (MaxVectorSize == 0) {
     DEBUG(dbgs() << "LV: The target has no vector registers.\n");
     MaxVectorSize = 1;
   }
 
   assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
          " into one vector!");
 
   unsigned VF = MaxVectorSize;
 
   // If we optimize the program for size, avoid creating the tail loop.
   if (OptForSize) {
     // If we are unable to calculate the trip count then don't try to vectorize.
     if (TC < 2) {

test/Transforms/LoopVectorize/X86/masked_load_store.ll

+; RUN: opt < %s  -O3 -mcpu=corei7-avx -S | FileCheck %s -check-prefix=AVX1
+; RUN: opt < %s  -O3 -mcpu=core-avx2 -S | FileCheck %s -check-prefix=AVX2
+; RUN: opt < %s  -O3 -mcpu=knl -S | FileCheck %s -check-prefix=AVX512
+
+;AVX1-NOT: llvm.masked
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-pc_linux"
+
+; The source code:
+;
+;void foo1(int *A, int *B, int *trigger) {
+;
+;  for (int i=0; i<10000; i++) {
+;    if (trigger[i] < 100) {
+;          A[i] = B[i] + trigger[i];
+;    }
+;  }
+;}
+
+;AVX2-LABEL: @foo1
+;AVX2: icmp slt <8 x i32> %wide.load, <i32 100, i32 100, i32 100
+;AVX2: call <8 x i32> @llvm.masked.load.v8i32
+;AVX2: add nsw <8 x i32>
+;AVX2: call void @llvm.masked.store.v8i32
+;AVX2: ret void
+
+;AVX512-LABEL: @foo1
+;AVX512: icmp slt <16 x i32> %wide.load, <i32 100, i32 100, i32 100
+;AVX512: call <16 x i32> @llvm.masked.load.v16i32
+;AVX512: add nsw <16 x i32>
+;AVX512: call void @llvm.masked.store.v16i32
+;AVX512: ret void
+
+; Function Attrs: nounwind uwtable
+define void @foo1(i32* %A, i32* %B, i32* %trigger) {
+entry:
+  %A.addr = alloca i32*, align 8
+  %B.addr = alloca i32*, align 8
+  %trigger.addr = alloca i32*, align 8
+  %i = alloca i32, align 4
+  store i32* %A, i32** %A.addr, align 8
+  store i32* %B, i32** %B.addr, align 8
+  store i32* %trigger, i32** %trigger.addr, align 8
+  store i32 0, i32* %i, align 4
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %entry
+  %0 = load i32* %i, align 4
+  %cmp = icmp slt i32 %0, 10000
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %1 = load i32* %i, align 4
+  %idxprom = sext i32 %1 to i64
+  %2 = load i32** %trigger.addr, align 8
+  %arrayidx = getelementptr inbounds i32* %2, i64 %idxprom
+  %3 = load i32* %arrayidx, align 4
+  %cmp1 = icmp slt i32 %3, 100
+  br i1 %cmp1, label %if.then, label %if.end
+
+if.then:                                          ; preds = %for.body
+  %4 = load i32* %i, align 4
+  %idxprom2 = sext i32 %4 to i64
+  %5 = load i32** %B.addr, align 8
+  %arrayidx3 = getelementptr inbounds i32* %5, i64 %idxprom2
+  %6 = load i32* %arrayidx3, align 4
+  %7 = load i32* %i, align 4
+  %idxprom4 = sext i32 %7 to i64
+  %8 = load i32** %trigger.addr, align 8
+  %arrayidx5 = getelementptr inbounds i32* %8, i64 %idxprom4
+  %9 = load i32* %arrayidx5, align 4
+  %add = add nsw i32 %6, %9
+  %10 = load i32* %i, align 4
+  %idxprom6 = sext i32 %10 to i64
+  %11 = load i32** %A.addr, align 8
+  %arrayidx7 = getelementptr inbounds i32* %11, i64 %idxprom6
+  store i32 %add, i32* %arrayidx7, align 4
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %for.body
+  br label %for.inc
+
+for.inc:                                          ; preds = %if.end
+  %12 = load i32* %i, align 4
+  %inc = add nsw i32 %12, 1
+  store i32 %inc, i32* %i, align 4
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+; The source code:
+;
+;void foo2(float *A, float *B, int *trigger) {
+;
+;  for (int i=0; i<10000; i++) {
+;    if (trigger[i] < 100) {
+;          A[i] = B[i] + trigger[i];
+;    }
+;  }
+;}
+
+;AVX2-LABEL: @foo2
+;AVX2: icmp slt <8 x i32> %wide.load, <i32 100, i32 100, i32 100
+;AVX2: call <8 x float> @llvm.masked.load.v8f32
+;AVX2: fadd <8 x float>
+;AVX2: call void @llvm.masked.store.v8f32
+;AVX2: ret void
+
+;AVX512-LABEL: @foo2
+;AVX512: icmp slt <16 x i32> %wide.load, <i32 100, i32 100, i32 100
+;AVX512: call <16 x float> @llvm.masked.load.v16f32
+;AVX512: fadd <16 x float>
+;AVX512: call void @llvm.masked.store.v16f32
+;AVX512: ret void
+
+; Function Attrs: nounwind uwtable
+define void @foo2(float* %A, float* %B, i32* %trigger) {
+entry:
+  %A.addr = alloca float*, align 8
+  %B.addr = alloca float*, align 8
+  %trigger.addr = alloca i32*, align 8
+  %i = alloca i32, align 4
+  store float* %A, float** %A.addr, align 8
+  store float* %B, float** %B.addr, align 8
+  store i32* %trigger, i32** %trigger.addr, align 8
+  store i32 0, i32* %i, align 4
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %entry
+  %0 = load i32* %i, align 4
+  %cmp = icmp slt i32 %0, 10000
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %1 = load i32* %i, align 4
+  %idxprom = sext i32 %1 to i64
+  %2 = load i32** %trigger.addr, align 8
+  %arrayidx = getelementptr inbounds i32* %2, i64 %idxprom
+  %3 = load i32* %arrayidx, align 4
+  %cmp1 = icmp slt i32 %3, 100
+  br i1 %cmp1, label %if.then, label %if.end
+
+if.then:                                          ; preds = %for.body
+  %4 = load i32* %i, align 4
+  %idxprom2 = sext i32 %4 to i64
+  %5 = load float** %B.addr, align 8
+  %arrayidx3 = getelementptr inbounds float* %5, i64 %idxprom2
+  %6 = load float* %arrayidx3, align 4
+  %7 = load i32* %i, align 4
+  %idxprom4 = sext i32 %7 to i64
+  %8 = load i32** %trigger.addr, align 8
+  %arrayidx5 = getelementptr inbounds i32* %8, i64 %idxprom4
+  %9 = load i32* %arrayidx5, align 4
+  %conv = sitofp i32 %9 to float
+  %add = fadd float %6, %conv
+  %10 = load i32* %i, align 4
+  %idxprom6 = sext i32 %10 to i64
+  %11 = load float** %A.addr, align 8
+  %arrayidx7 = getelementptr inbounds float* %11, i64 %idxprom6
+  store float %add, float* %arrayidx7, align 4
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %for.body
+  br label %for.inc
+
+for.inc:                                          ; preds = %if.end
+  %12 = load i32* %i, align 4
+  %inc = add nsw i32 %12, 1
+  store i32 %inc, i32* %i, align 4
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+; The source code:
+;
+;void foo3(double *A, double *B, int *trigger) {
+;
+;  for (int i=0; i<10000; i++) {
+;    if (trigger[i] < 100) {
+;          A[i] = B[i] + trigger[i];
+;    }
+;  }
+;}
+
+;AVX2-LABEL: @foo3
+;AVX2: icmp slt <4 x i32> %wide.load, <i32 100, i32 100,
+;AVX2: call <4 x double> @llvm.masked.load.v4f64
+;AVX2: sitofp <4 x i32> %wide.load to <4 x double>
+;AVX2: fadd <4 x double>
+;AVX2: call void @llvm.masked.store.v4f64
+;AVX2: ret void
+
+;AVX512-LABEL: @foo3
+;AVX512: icmp slt <8 x i32> %wide.load, <i32 100, i32 100,
+;AVX512: call <8 x double> @llvm.masked.load.v8f64
+;AVX512: sitofp <8 x i32> %wide.load to <8 x double>
+;AVX512: fadd <8 x double>
+;AVX512: call void @llvm.masked.store.v8f64
+;AVX512: ret void
+
+
+; Function Attrs: nounwind uwtable
+define void @foo3(double* %A, double* %B, i32* %trigger) #0 {
+entry:
+  %A.addr = alloca double*, align 8
+  %B.addr = alloca double*, align 8
+  %trigger.addr = alloca i32*, align 8
+  %i = alloca i32, align 4
+  store double* %A, double** %A.addr, align 8
+  store double* %B, double** %B.addr, align 8
+  store i32* %trigger, i32** %trigger.addr, align 8
+  store i32 0, i32* %i, align 4
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %entry
+  %0 = load i32* %i, align 4
+  %cmp = icmp slt i32 %0, 10000
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %1 = load i32* %i, align 4
+  %idxprom = sext i32 %1 to i64
+  %2 = load i32** %trigger.addr, align 8
+  %arrayidx = getelementptr inbounds i32* %2, i64 %idxprom
+  %3 = load i32* %arrayidx, align 4
+  %cmp1 = icmp slt i32 %3, 100
+  br i1 %cmp1, label %if.then, label %if.end
+
+if.then:                                          ; preds = %for.body
+  %4 = load i32* %i, align 4
+  %idxprom2 = sext i32 %4 to i64
+  %5 = load double** %B.addr, align 8
+  %arrayidx3 = getelementptr inbounds double* %5, i64 %idxprom2
+  %6 = load double* %arrayidx3, align 8
+  %7 = load i32* %i, align 4
+  %idxprom4 = sext i32 %7 to i64
+  %8 = load i32** %trigger.addr, align 8
+  %arrayidx5 = getelementptr inbounds i32* %8, i64 %idxprom4
+  %9 = load i32* %arrayidx5, align 4
+  %conv = sitofp i32 %9 to double
+  %add = fadd double %6, %conv
+  %10 = load i32* %i, align 4
+  %idxprom6 = sext i32 %10 to i64
+  %11 = load double** %A.addr, align 8
+  %arrayidx7 = getelementptr inbounds double* %11, i64 %idxprom6
+  store double %add, double* %arrayidx7, align 8
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %for.body
+  br label %for.inc
+
+for.inc:                                          ; preds = %if.end
+  %12 = load i32* %i, align 4
+  %inc = add nsw i32 %12, 1
+  store i32 %inc, i32* %i, align 4
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+; The source code:
+;
+;void foo4(double *A, double *B, int *trigger) {
+;
+;  for (int i=0; i<10000; i++) {
+;    if (trigger[i] < 100) {
+;          A[i] = B[i*2] + trigger[i]; << non-cosecutive access
+;    }
+;  }
+;}
+
+;AVX2-LABEL: @foo4
+;AVX2-NOT: llvm.masked
+;AVX2: ret void
+
+;AVX512-LABEL: @foo4
+;AVX512-NOT: llvm.masked
+;AVX512: ret void
+
+; Function Attrs: nounwind uwtable
+define void @foo4(double* %A, double* %B, i32* %trigger)  {
+entry:
+  %A.addr = alloca double*, align 8
+  %B.addr = alloca double*, align 8
+  %trigger.addr = alloca i32*, align 8
+  %i = alloca i32, align 4
+  store double* %A, double** %A.addr, align 8
+  store double* %B, double** %B.addr, align 8
+  store i32* %trigger, i32** %trigger.addr, align 8
+  store i32 0, i32* %i, align 4
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %entry
+  %0 = load i32* %i, align 4
+  %cmp = icmp slt i32 %0, 10000
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %1 = load i32* %i, align 4
+  %idxprom = sext i32 %1 to i64
+  %2 = load i32** %trigger.addr, align 8
+  %arrayidx = getelementptr inbounds i32* %2, i64 %idxprom
+  %3 = load i32* %arrayidx, align 4
+  %cmp1 = icmp slt i32 %3, 100
+  br i1 %cmp1, label %if.then, label %if.end
+
+if.then:                                          ; preds = %for.body
+  %4 = load i32* %i, align 4
+  %mul = mul nsw i32 %4, 2
+  %idxprom2 = sext i32 %mul to i64
+  %5 = load double** %B.addr, align 8
+  %arrayidx3 = getelementptr inbounds double* %5, i64 %idxprom2
+  %6 = load double* %arrayidx3, align 8
+  %7 = load i32* %i, align 4
+  %idxprom4 = sext i32 %7 to i64
+  %8 = load i32** %trigger.addr, align 8
+  %arrayidx5 = getelementptr inbounds i32* %8, i64 %idxprom4
+  %9 = load i32* %arrayidx5, align 4
+  %conv = sitofp i32 %9 to double
+  %add = fadd double %6, %conv
+  %10 = load i32* %i, align 4
+  %idxprom6 = sext i32 %10 to i64
+  %11 = load double** %A.addr, align 8
+  %arrayidx7 = getelementptr inbounds double* %11, i64 %idxprom6
+  store double %add, double* %arrayidx7, align 8
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %for.body
+  br label %for.inc
+
+for.inc:                                          ; preds = %if.end
+  %12 = load i32* %i, align 4
+  %inc = add nsw i32 %12, 1
+  store i32 %inc, i32* %i, align 4
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+
+@a = common global [1 x i32*] zeroinitializer, align 8
+@c = common global i32* null, align 8
+
+; The loop here should not be vectorized due to trapping
+; constant expression
+;AVX2-LABEL: @foo5
+;AVX2-NOT: llvm.masked
+;AVX2: store i32 sdiv
+;AVX2: ret void
+
+;AVX512-LABEL: @foo5
+;AVX512-NOT: llvm.masked
+;AVX512: store i32 sdiv
+;AVX512: ret void
+
+; Function Attrs: nounwind uwtable
+define void @foo5(i32* %A, i32* %B, i32* %trigger) {
+entry:
+  %A.addr = alloca i32*, align 8
+  %B.addr = alloca i32*, align 8
+  %trigger.addr = alloca i32*, align 8
+  %i = alloca i32, align 4
+  store i32* %A, i32** %A.addr, align 8
+  store i32* %B, i32** %B.addr, align 8
+  store i32* %trigger, i32** %trigger.addr, align 8
+  store i32 0, i32* %i, align 4
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %entry
+  %0 = load i32* %i, align 4
+  %cmp = icmp slt i32 %0, 10000
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %1 = load i32* %i, align 4
+  %idxprom = sext i32 %1 to i64
+  %2 = load i32** %trigger.addr, align 8
+  %arrayidx = getelementptr inbounds i32* %2, i64 %idxprom
+  %3 = load i32* %arrayidx, align 4
+  %cmp1 = icmp slt i32 %3, 100
+  br i1 %cmp1, label %if.then, label %if.end
+
+if.then:                                          ; preds = %for.body
+  %4 = load i32* %i, align 4
+  %idxprom2 = sext i32 %4 to i64
+  %5 = load i32** %B.addr, align 8
+  %arrayidx3 = getelementptr inbounds i32* %5, i64 %idxprom2
+  %6 = load i32* %arrayidx3, align 4
+  %7 = load i32* %i, align 4
+  %idxprom4 = sext i32 %7 to i64
+  %8 = load i32** %trigger.addr, align 8
+  %arrayidx5 = getelementptr inbounds i32* %8, i64 %idxprom4
+  %9 = load i32* %arrayidx5, align 4
+  %add = add nsw i32 %6, %9
+  %10 = load i32* %i, align 4
+  %idxprom6 = sext i32 %10 to i64
+  %11 = load i32** %A.addr, align 8
+  %arrayidx7 = getelementptr inbounds i32* %11, i64 %idxprom6
+  store i32 sdiv (i32 1, i32 zext (i1 icmp eq (i32** getelementptr inbounds ([1 x i32*]* @a, i64 0, i64 1), i32** @c) to i32)), i32* %arrayidx7, align 4
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %for.body
+  br label %for.inc
+
+for.inc:                                          ; preds = %if.end
+  %12 = load i32* %i, align 4
+  %inc = add nsw i32 %12, 1
+  store i32 %inc, i32* %i, align 4
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}
+