Add support for bottom-up SLP vectorization infrastructure.

This commit adds the infrastructure for performing bottom-up SLP vectorization (and other optimizations) on parallel computations.
The infrastructure has three potential users:

  1. The loop vectorizer needs to be able to vectorize AOS data structures such as (sum += A[i] + A[i+1]).

  2. The BB-vectorizer needs this infrastructure for bottom-up SLP vectorization, because bottom-up vectorization is faster to compute.

  3. A loop-roller needs to be able to analyze consecutive chains and roll them into a loop, in order to reduce code size. A loop roller does not need to create vector instructions, and this infrastructure separates the chain analysis from the vectorization.

This patch also includes a simple (100 LOC) bottom up SLP vectorizer that uses the infrastructure, and can vectorize this code:

void SAXPY(int *x, int *y, int a, int i) {
  x[i]   = a * x[i]   + y[i];
  x[i+1] = a * x[i+1] + y[i+1];
  x[i+2] = a * x[i+2] + y[i+2];
  x[i+3] = a * x[i+3] + y[i+3];
}



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179117 91177308-0d34-0410-b5e6-96231b3b80d8

15 files changed, 995 insertions, 0 deletions

diff --git a/include/llvm/InitializePasses.h b/include/llvm/InitializePasses.h
index 9cc194b..5b2cd60 100644
--- a/include/llvm/InitializePasses.h
+++ b/include/llvm/InitializePasses.h
@@ -271,6 +271,7 @@ void initializeInstSimplifierPass(PassRegistry&);
 void initializeUnpackMachineBundlesPass(PassRegistry&);
 void initializeFinalizeMachineBundlesPass(PassRegistry&);
 void initializeLoopVectorizePass(PassRegistry&);
+void initializeSLPVectorizerPass(PassRegistry&);
 void initializeBBVectorizePass(PassRegistry&);
 void initializeMachineFunctionPrinterPassPass(PassRegistry&);
 }
diff --git a/include/llvm/LinkAllPasses.h b/include/llvm/LinkAllPasses.h
index 1f017e4..ca1c139 100644
--- a/include/llvm/LinkAllPasses.h
+++ b/include/llvm/LinkAllPasses.h
@@ -161,6 +161,7 @@ namespace {
       (void) llvm::createMemDepPrinter();
       (void) llvm::createInstructionSimplifierPass();
       (void) llvm::createLoopVectorizePass();
+      (void) llvm::createSLPVectorizerPass();
       (void) llvm::createBBVectorizePass();
 
       (void)new llvm::IntervalPartition();
diff --git a/include/llvm/Transforms/Vectorize.h b/include/llvm/Transforms/Vectorize.h
index d205dbd..8d0db16 100644
--- a/include/llvm/Transforms/Vectorize.h
+++ b/include/llvm/Transforms/Vectorize.h
@@ -117,6 +117,12 @@ createBBVectorizePass(const VectorizeConfig &C = VectorizeConfig());
 Pass *createLoopVectorizePass();
 
 //===----------------------------------------------------------------------===//
+//
+// SLPVectorizer - Create a bottom-up SLP vectorizer pass.
+//
+Pass *createSLPVectorizerPass();
+
+//===----------------------------------------------------------------------===//
 /// @brief Vectorize the BasicBlock.
 ///
 /// @param BB The BasicBlock to be vectorized
diff --git a/lib/Transforms/Vectorize/CMakeLists.txt b/lib/Transforms/Vectorize/CMakeLists.txt
index e64034a..7ae082f 100644
--- a/lib/Transforms/Vectorize/CMakeLists.txt
+++ b/lib/Transforms/Vectorize/CMakeLists.txt
@@ -2,6 +2,8 @@ add_llvm_library(LLVMVectorize
   BBVectorize.cpp
   Vectorize.cpp
   LoopVectorize.cpp
+  SLPVectorizer.cpp
+  VecUtils.cpp
   )
 
 add_dependencies(LLVMVectorize intrinsics_gen)
diff --git a/lib/Transforms/Vectorize/SLPVectorizer.cpp b/lib/Transforms/Vectorize/SLPVectorizer.cpp
new file mode 100644
index 0000000..4b61dc9
--- /dev/null
+++ b/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -0,0 +1,153 @@
+//===- SLPVectorizer.cpp - A bottom up SLP Vectorizer ---------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+// This pass implements the Bottom Up SLP vectorizer. It detects consecutive
+// stores that can be put together into vector-stores. Next, it attempts to
+// construct vectorizable tree using the use-def chains. If a profitable tree
+// was found, the SLP vectorizer performs vectorization on the tree.
+//
+// The pass is inspired by the work described in the paper:
+//  "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks.
+//
+//===----------------------------------------------------------------------===//
+#define SV_NAME "slp-vectorizer"
+#define DEBUG_TYPE SV_NAME
+
+#include "VecUtils.h"
+#include "llvm/Transforms/Vectorize.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/Analysis/Verifier.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Type.h"
+#include "llvm/IR/Value.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include <map>
+
+using namespace llvm;
+
+static cl::opt<int>
+SLPCostThreshold("slp-threshold", cl::init(1), cl::Hidden,
+                 cl::desc("Only vectorize trees if the gain is above this "
+                          "number. (gain = -cost of vectorization)"));
+namespace {
+
+/// The SLPVectorizer Pass.
+struct SLPVectorizer : public BasicBlockPass {
+  typedef std::map<Value*, BoUpSLP::StoreList> StoreListMap;
+
+  /// Pass identification, replacement for typeid
+  static char ID;
+
+  explicit SLPVectorizer() : BasicBlockPass(ID) {
+    initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
+  }
+
+  ScalarEvolution *SE;
+  DataLayout *DL;
+  TargetTransformInfo *TTI;
+  AliasAnalysis *AA;
+
+  /// \brief Collect memory references and sort them according to their base
+  /// object. We sort the stores to their base objects to reduce the cost of the
+  /// quadratic search on the stores. TODO: We can further reduce this cost
+  /// if we flush the chain creation every time we run into a memory barrier.
+  bool CollectStores(BasicBlock *BB, BoUpSLP &R) {
+    for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
+      // Can't vectorize instructions with side effects.
+      if (it->mayThrow())
+        return false;
+
+      StoreInst *SI = dyn_cast<StoreInst>(it);
+      if (!SI)
+        continue;
+
+      // Check that the pointer points to scalars.
+      if (SI->getValueOperand()->getType()->isAggregateType())
+        return false;
+
+      // Find the base of the GEP.
+      Value *Ptr = SI->getPointerOperand();
+      if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
+        Ptr = GEP->getPointerOperand();
+
+      // Save the store locations.
+      StoreRefs[Ptr].push_back(SI);
+    }
+    return true;
+  }
+
+  bool RollStoreChains(BoUpSLP &R) {
+    bool Changed = false;
+    // Attempt to sort and vectorize each of the store-groups.
+    for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
+         it != e; ++it) {
+      if (it->second.size() < 2)
+        continue;
+      Changed |= R.vectorizeStores(it->second, -SLPCostThreshold);
+    }
+    return Changed;
+  }
+
+  virtual bool runOnBasicBlock(BasicBlock &BB) {
+    SE = &getAnalysis<ScalarEvolution>();
+    DL = getAnalysisIfAvailable<DataLayout>();
+    TTI = &getAnalysis<TargetTransformInfo>();
+    AA = &getAnalysis<AliasAnalysis>();
+    StoreRefs.clear();
+
+    // Use the bollom up slp vectorizer to construct chains that start with
+    // he store instructions.
+    BoUpSLP R(&BB, SE, DL, TTI, AA);
+
+    if (!CollectStores(&BB, R))
+      return false;
+
+    bool Changed = RollStoreChains(R);
+    if (Changed) {
+      DEBUG(dbgs()<<"Rolled chains in \""<<BB.getParent()->getName()<<"\"\n");
+      DEBUG(verifyFunction(*BB.getParent()));
+    }
+
+    return Changed;
+  }
+
+  virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+    BasicBlockPass::getAnalysisUsage(AU);
+    AU.addRequired<ScalarEvolution>();
+    AU.addRequired<AliasAnalysis>();
+    AU.addRequired<TargetTransformInfo>();
+  }
+
+private:
+  StoreListMap StoreRefs;
+};
+
+} // end anonymous namespace
+
+char SLPVectorizer::ID = 0;
+static const char lv_name[] = "SLP Vectorizer";
+INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
+INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
+INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
+INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
+INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
+
+namespace llvm {
+  Pass *createSLPVectorizerPass() {
+    return new SLPVectorizer();
+  }
+}
+
diff --git a/lib/Transforms/Vectorize/VecUtils.cpp b/lib/Transforms/Vectorize/VecUtils.cpp
new file mode 100644
index 0000000..7e9f12d
--- /dev/null
+++ b/lib/Transforms/Vectorize/VecUtils.cpp
@@ -0,0 +1,439 @@
+//===- VecUtils.h --- Vectorization Utilities -----------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "VecUtils"
+
+#include "VecUtils.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/Analysis/Verifier.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Type.h"
+#include "llvm/IR/Value.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetLibraryInfo.h"
+#include "llvm/Transforms/Scalar.h"
+#include "llvm/Transforms/Utils/Local.h"
+#include <algorithm>
+#include <map>
+
+using namespace llvm;
+
+namespace llvm {
+
+BoUpSLP::BoUpSLP(BasicBlock *Bb, ScalarEvolution *S, DataLayout *Dl,
+                             TargetTransformInfo *Tti, AliasAnalysis *Aa) :
+                             BB(Bb), SE(S), DL(Dl), TTI(Tti), AA(Aa) {
+  numberInstructions();
+}
+
+void BoUpSLP::numberInstructions() {
+  int Loc = 0;
+  InstrIdx.clear();
+  InstrVec.clear();
+  // Number the instructions in the block.
+  for (BasicBlock::iterator it=BB->begin(), e=BB->end(); it != e; ++it) {
+    InstrIdx[it] = Loc++;
+    InstrVec.push_back(it);
+    assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation");
+  }
+}
+
+Value *BoUpSLP::getPointerOperand(Value *I) {
+  if (LoadInst *LI = dyn_cast<LoadInst>(I)) return LI->getPointerOperand();
+  if (StoreInst *SI = dyn_cast<StoreInst>(I)) return SI->getPointerOperand();
+  return 0;
+}
+
+unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
+  if (LoadInst *L=dyn_cast<LoadInst>(I)) return L->getPointerAddressSpace();
+  if (StoreInst *S=dyn_cast<StoreInst>(I)) return S->getPointerAddressSpace();
+  return -1;
+}
+
+bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
+  Value *PtrA = getPointerOperand(A);
+  Value *PtrB = getPointerOperand(B);
+  unsigned ASA = getAddressSpaceOperand(A);
+  unsigned ASB = getAddressSpaceOperand(B);
+
+  // Check that the address spaces match and that the pointers are valid.
+  if (!PtrA || !PtrB || (ASA != ASB)) return false;
+
+  // Check that A and B are of the same type.
+  if (PtrA->getType() != PtrB->getType()) return false;
+
+  // Calculate the distance.
+  const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
+  const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
+  const SCEV *OffsetSCEV = SE->getMinusSCEV(PtrSCEVA, PtrSCEVB);
+  const SCEVConstant *ConstOffSCEV = dyn_cast<SCEVConstant>(OffsetSCEV);
+
+  // Non constant distance.
+  if (!ConstOffSCEV) return false;
+
+  unsigned Offset = ConstOffSCEV->getValue()->getSExtValue();
+  Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
+  // The Instructions are connsecutive if the size of the first load/store is
+  // the same as the offset.
+  unsigned Sz = (DL ? DL->getTypeStoreSize(Ty) : Ty->getScalarSizeInBits()/8);
+  return ((-Offset) == Sz);
+}
+
+bool BoUpSLP::vectorizeStores(StoreList &Stores, int costThreshold) {
+  ValueSet Heads, Tails;
+  SmallDenseMap<Value*, Value*> ConsecutiveChain;
+  bool Changed = false;
+
+  // Do a quadratic search on all of the given stores and find
+  // all of the pairs of loads that follow each other.
+  for (unsigned i = 0, e = Stores.size(); i < e; ++i)
+    for (unsigned j = 0; j < e; ++j) {
+      if (i == j) continue;
+      if (isConsecutiveAccess(Stores[i], Stores[j])) {
+        Tails.insert(Stores[j]);
+        Heads.insert(Stores[i]);
+        ConsecutiveChain[Stores[i]] = Stores[j];
+      }
+    }
+
+  // For stores that start but don't end a link in the chain:
+  for (ValueSet::iterator it = Heads.begin(), e = Heads.end();it != e; ++it) {
+    if (Tails.count(*it)) continue;
+
+    // We found a store instr that starts a chain. Now follow the chain and try
+    // to vectorize it.
+    ValueList Operands;
+    Value *I = *it;
+    int MinCost = 0, MinVF = 0;
+    while (Tails.count(I) || Heads.count(I)) {
+      Operands.push_back(I);
+      unsigned VF = Operands.size();
+      if (isPowerOf2_32(VF) && VF > 1) {
+        int cost = getTreeRollCost(Operands, 0);
+        DEBUG(dbgs() << "Found cost=" << cost << " for VF=" << VF << "\n");
+        if (cost < MinCost) { MinCost = cost; MinVF = VF; }
+      }
+      // Move to the next value in the chain.
+      I = ConsecutiveChain[I];
+    }
+
+    if (MinCost <= costThreshold && MinVF > 1) {
+      DEBUG(dbgs() << "Decided to vectorize cost=" << MinCost << "\n");
+      vectorizeTree(Operands, MinVF);
+      Stores.clear();
+      // The current numbering is invalid because we added and removed instrs.
+      numberInstructions();
+      Changed = true;
+    }
+  }
+
+  return Changed;
+}
+
+int BoUpSLP::getScalarizationCost(Type *Ty) {
+  int Cost = 0;
+  for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
+    Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
+  return Cost;
+}
+
+AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
+  if (StoreInst *SI = dyn_cast<StoreInst>(I)) return AA->getLocation(SI);
+  if (LoadInst *LI = dyn_cast<LoadInst>(I)) return AA->getLocation(LI);
+  return AliasAnalysis::Location();
+}
+
+Value *BoUpSLP::isUnsafeToSink(Instruction *Src, Instruction *Dst) {
+  assert(Src->getParent() == Dst->getParent() && "Not the same BB");
+  BasicBlock::iterator I = Src, E = Dst;
+  /// Scan all of the instruction from SRC to DST and check if
+  /// the source may alias.
+  for (++I; I != E; ++I) {
+    // Ignore store instructions that are marked as 'ignore'.
+    if (MemBarrierIgnoreList.count(I)) continue;
+    if (Src->mayWriteToMemory()) /* Write */ {
+      if (!I->mayReadOrWriteMemory()) continue;
+    } else /* Read */ {
+      if (!I->mayWriteToMemory()) continue;
+    }
+    AliasAnalysis::Location A = getLocation(&*I);
+    AliasAnalysis::Location B = getLocation(Src);
+
+    if (!A.Ptr || !B.Ptr || AA->alias(A, B))
+      return I;
+  }
+  return 0;
+}
+
+int BoUpSLP::getTreeRollCost(ValueList &VL, unsigned Depth) {
+  if (Depth == 6) return max_cost;
+  Type *ScalarTy = VL[0]->getType();
+
+  if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
+    ScalarTy = SI->getValueOperand()->getType();
+
+  /// Don't mess with vectors.
+  if (ScalarTy->isVectorTy()) return max_cost;
+
+  VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
+
+  // Check if all of the operands are constants.
+  bool AllConst = true;
+  bool AllSameScalar = true;
+  for (unsigned i = 0, e = VL.size(); i < e; ++i) {
+    AllConst &= isa<Constant>(VL[i]);
+    AllSameScalar &= (VL[0] == VL[i]);
+    // Must have a single use.
+    Instruction *I = dyn_cast<Instruction>(VL[i]);
+    // Need to scalarize instructions with multiple users or from other BBs.
+    if (I && ((I->getNumUses() > 1) || (I->getParent() != BB)))
+      return getScalarizationCost(VecTy);
+  }
+
+  // Is this a simple vector constant.
+  if (AllConst) return 0;
+
+  // If all of the operands are identical we can broadcast them.
+  if (AllSameScalar)
+    return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
+
+  // Scalarize unknown structures.
+  Instruction *VL0 = dyn_cast<Instruction>(VL[0]);
+  if (!VL0) return getScalarizationCost(VecTy);
+  assert(VL0->getParent() == BB && "Wrong BB");
+
+  unsigned Opcode = VL0->getOpcode();
+  for (unsigned i = 0, e = VL.size(); i < e; ++i) {
+    Instruction *I = dyn_cast<Instruction>(VL[i]);
+    // If not all of the instructions are identical then we have to scalarize.
+    if (!I || Opcode != I->getOpcode()) return getScalarizationCost(VecTy);
+  }
+
+  // Check if it is safe to sink the loads or the stores.
+  if (Opcode == Instruction::Load || Opcode == Instruction::Store) {
+    int MaxIdx = InstrIdx[VL0];
+    for (unsigned i = 1, e = VL.size(); i < e; ++i )
+      MaxIdx = std::max(MaxIdx, InstrIdx[VL[i]]);
+
+    Instruction *Last = InstrVec[MaxIdx];
+    for (unsigned i = 0, e = VL.size(); i < e; ++i ) {
+      if (VL[i] == Last) continue;
+      Value *Barrier = isUnsafeToSink(cast<Instruction>(VL[i]), Last);
+      if (Barrier) {
+        DEBUG(dbgs() << "LR: Can't sink " << *VL[i] << "\n down to " <<
+              *Last << "\n because of " << *Barrier << "\n");
+        return max_cost;
+      }
+    }
+  }
+
+  switch (Opcode) {
+  case Instruction::Add:
+  case Instruction::FAdd:
+  case Instruction::Sub:
+  case Instruction::FSub:
+  case Instruction::Mul:
+  case Instruction::FMul:
+  case Instruction::UDiv:
+  case Instruction::SDiv:
+  case Instruction::FDiv:
+  case Instruction::URem:
+  case Instruction::SRem:
+  case Instruction::FRem:
+  case Instruction::Shl:
+  case Instruction::LShr:
+  case Instruction::AShr:
+  case Instruction::And:
+  case Instruction::Or:
+  case Instruction::Xor: {
+    ValueList Operands;
+    int Cost = 0;
+    // Calculate the cost of all of the operands.
+    for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
+      // Prepare the operand vector.
+      for (unsigned j = 0; j < VL.size(); ++j)
+        Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
+      Cost += getTreeRollCost(Operands, Depth+1);
+      Operands.clear();
+    }
+
+    // Calculate the cost of this instruction.
+    int ScalarCost = VecTy->getNumElements() *
+      TTI->getArithmeticInstrCost(Opcode, ScalarTy);
+    int VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy);
+    Cost += (VecCost - ScalarCost);
+    return Cost;
+  }
+  case Instruction::Load: {
+    // If we are scalarize the loads, add the cost of forming the vector.
+    for (unsigned i = 0, e = VL.size()-1; i < e; ++i)
+      if (!isConsecutiveAccess(VL[i], VL[i+1]))
+        return getScalarizationCost(VecTy);
+
+    // Cost of wide load - cost of scalar loads.
+    int ScalarLdCost = VecTy->getNumElements() *
+      TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
+    int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
+    return VecLdCost - ScalarLdCost;
+  }
+  case Instruction::Store: {
+    // We know that we can merge the stores. Calculate the cost.
+    int ScalarStCost = VecTy->getNumElements() *
+      TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
+    int VecStCost = TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1,0);
+    int StoreCost = VecStCost - ScalarStCost;
+
+    ValueList Operands;
+    for (unsigned j = 0; j < VL.size(); ++j) {
+      Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
+      MemBarrierIgnoreList.insert(VL[j]);
+    }
+
+    int TotalCost =  StoreCost + getTreeRollCost(Operands, Depth + 1);
+    MemBarrierIgnoreList.clear();
+    return TotalCost;
+  }
+  default:
+    // Unable to vectorize unknown instructions.
+    return getScalarizationCost(VecTy);
+  }
+}
+
+Instruction *BoUpSLP::GetLastInstr(ValueList &VL, unsigned VF) {
+  int MaxIdx = InstrIdx[BB->getFirstNonPHI()];
+  for (unsigned i = 0; i < VF; ++i )
+    MaxIdx = std::max(MaxIdx, InstrIdx[VL[i]]);
+  return InstrVec[MaxIdx + 1];
+}
+
+Value *BoUpSLP::Scalarize(ValueList &VL, VectorType *Ty) {
+  IRBuilder<> Builder(GetLastInstr(VL, Ty->getNumElements()));
+  Value *Vec = UndefValue::get(Ty);
+  for (unsigned i=0; i < Ty->getNumElements(); ++i)
+    Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
+  return Vec;
+}
+
+Value *BoUpSLP::vectorizeTree(ValueList &VL, int VF) {
+  Type *ScalarTy = VL[0]->getType();
+  if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
+    ScalarTy = SI->getValueOperand()->getType();
+  VectorType *VecTy = VectorType::get(ScalarTy, VF);
+
+  // Check if all of the operands are constants or identical.
+  bool AllConst = true;
+  bool AllSameScalar = true;
+  for (unsigned i = 0, e = VF; i < e; ++i) {
+    AllConst &= !!dyn_cast<Constant>(VL[i]);
+    AllSameScalar &= (VL[0] == VL[i]);
+    // Must have a single use.
+    Instruction *I = dyn_cast<Instruction>(VL[i]);
+    if (I && (I->getNumUses() > 1 || I->getParent() != BB))
+      return Scalarize(VL, VecTy);
+  }
+
+  // Is this a simple vector constant.
+  if (AllConst || AllSameScalar) return Scalarize(VL, VecTy);
+
+  // Scalarize unknown structures.
+  Instruction *VL0 = dyn_cast<Instruction>(VL[0]);
+  if (!VL0) return Scalarize(VL, VecTy);
+
+  unsigned Opcode = VL0->getOpcode();
+  for (unsigned i = 0, e = VF; i < e; ++i) {
+    Instruction *I = dyn_cast<Instruction>(VL[i]);
+    // If not all of the instructions are identical then we have to scalarize.
+    if (!I || Opcode != I->getOpcode()) return Scalarize(VL, VecTy);
+  }
+
+  switch (Opcode) {
+  case Instruction::Add:
+  case Instruction::FAdd:
+  case Instruction::Sub:
+  case Instruction::FSub:
+  case Instruction::Mul:
+  case Instruction::FMul:
+  case Instruction::UDiv:
+  case Instruction::SDiv:
+  case Instruction::FDiv:
+  case Instruction::URem:
+  case Instruction::SRem:
+  case Instruction::FRem:
+  case Instruction::Shl:
+  case Instruction::LShr:
+  case Instruction::AShr:
+  case Instruction::And:
+  case Instruction::Or:
+  case Instruction::Xor: {
+    ValueList LHSVL, RHSVL;
+    for (int i = 0; i < VF; ++i) {
+      RHSVL.push_back(cast<Instruction>(VL[i])->getOperand(0));
+      LHSVL.push_back(cast<Instruction>(VL[i])->getOperand(1));
+    }
+
+    Value *RHS = vectorizeTree(RHSVL, VF);
+    Value *LHS = vectorizeTree(LHSVL, VF);
+    IRBuilder<> Builder(GetLastInstr(VL, VF));
+    BinaryOperator *BinOp = dyn_cast<BinaryOperator>(VL0);
+    return Builder.CreateBinOp(BinOp->getOpcode(), RHS,LHS);
+  }
+  case Instruction::Load: {
+    LoadInst *LI = dyn_cast<LoadInst>(VL0);
+    unsigned Alignment = LI->getAlignment();
+
+    // Check if all of the loads are consecutive.
+    for (unsigned i = 1, e = VF; i < e; ++i)
+      if (!isConsecutiveAccess(VL[i-1], VL[i]))
+        return Scalarize(VL, VecTy);
+
+    IRBuilder<> Builder(GetLastInstr(VL, VF));
+    Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
+                                          VecTy->getPointerTo());
+    LI = Builder.CreateLoad(VecPtr);
+    LI->setAlignment(Alignment);
+    return LI;
+  }
+  case Instruction::Store: {
+    StoreInst *SI = dyn_cast<StoreInst>(VL0);
+    unsigned Alignment = SI->getAlignment();
+
+    ValueList ValueOp;
+    for (int i = 0; i < VF; ++i)
+      ValueOp.push_back(cast<StoreInst>(VL[i])->getValueOperand());
+
+    Value *VecValue = vectorizeTree(ValueOp, VF);
+
+    IRBuilder<> Builder(GetLastInstr(VL, VF));
+    Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
+                                          VecTy->getPointerTo());
+    Builder.CreateStore(VecValue, VecPtr)->setAlignment(Alignment);
+
+    for (int i = 0; i < VF; ++i)
+      cast<Instruction>(VL[i])->eraseFromParent();
+    return 0;
+  }
+  default:
+    return Scalarize(VL, VecTy);
+  }
+}
+
+} // end of namespace
diff --git a/lib/Transforms/Vectorize/VecUtils.h b/lib/Transforms/Vectorize/VecUtils.h
new file mode 100644
index 0000000..b1c2b0c
--- /dev/null
+++ b/lib/Transforms/Vectorize/VecUtils.h
@@ -0,0 +1,108 @@
+//===- VecUtils.cpp - Vectorization Utilities -----------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This family of classes and functions manipulate vectors and chains of
+// vectors.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_TRANSFORMS_VECTORIZE_AOSVECTORIZER_H
+#define  LLVM_TRANSFORMS_VECTORIZE_AOSVECTORIZER_H
+
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include <vector>
+
+using namespace llvm;
+
+namespace llvm {
+
+class BasicBlock; class Instruction; class Type;
+class VectorType; class StoreInst; class Value;
+class ScalarEvolution; class DataLayout;
+class TargetTransformInfo; class AliasAnalysis;
+
+/// Bottom Up SLP vectorization utility class.
+struct BoUpSLP  {
+  typedef SmallVector<Value*, 8> ValueList;
+  typedef SmallPtrSet<Value*, 16> ValueSet;
+  typedef SmallVector<StoreInst*, 8> StoreList;
+  static const int max_cost = 1<<20;
+
+  // \brief C'tor.
+  BoUpSLP(BasicBlock *Bb, ScalarEvolution *Se, DataLayout *Dl,
+         TargetTransformInfo *Tti, AliasAnalysis *Aa);
+
+  /// \returns true if the memory operations A and B are consecutive.
+  bool isConsecutiveAccess(Value *A, Value *B);
+
+  /// \brief Vectorize the tree that starts with the elements in \p VL.
+  /// \returns the vectorized value.
+  Value *vectorizeTree(ValueList &VL, int VF);
+
+  /// \returns the vectorization cost of the subtree that starts at \p VL.
+  /// A negative number means that this is profitable.
+  int getTreeRollCost(ValueList &VL, unsigned Depth);
+
+  /// \brief Take the pointer operand from the Load/Store instruction.
+  /// \returns NULL if this is not a valid Load/Store instruction.
+  static Value *getPointerOperand(Value *I);
+
+  /// \brief Take the address space operand from the Load/Store instruction.
+  /// \returns -1 if this is not a valid Load/Store instruction.
+  static unsigned getAddressSpaceOperand(Value *I);
+
+  /// \brief Attempts to order and vectorize a sequence of stores. This
+  /// function does a quadratic scan of the given stores.
+  /// \returns true if the basic block was modified.
+  bool vectorizeStores(StoreList &Stores, int costThreshold);
+
+  /// \brief Number all of the instructions in the block.
+  void numberInstructions();
+
+private:
+  /// \returns the scalarization cost for this type. Scalarization in this
+  /// context means the creation of vectors from a group of scalars.
+  int getScalarizationCost(Type *Ty);
+
+  /// \returns the AA location that is being access by the instruction.
+  AliasAnalysis::Location getLocation(Instruction *I);
+
+  /// \brief Checks if it is possible to sink an instruction from
+  /// \p Src to \p Dst.
+  /// \returns the pointer to the barrier instruction if we can't sink.
+  Value *isUnsafeToSink(Instruction *Src, Instruction *Dst);
+
+  /// \returns the instruction that appears last in the BB from \p VL.
+  /// Only consider the first \p VF elements.
+  Instruction *GetLastInstr(ValueList &VL, unsigned VF);
+
+  /// \returns a vector from a collection of scalars in \p VL.
+  Value *Scalarize(ValueList &VL, VectorType *Ty);
+
+  // Maps instructions to numbers and back.
+  SmallDenseMap<Value*, int> InstrIdx;
+  std::vector<Instruction*> InstrVec;
+  // A list of instructions to ignore while sinking
+  // memory instructions.
+  SmallSet<Value*, 8> MemBarrierIgnoreList;
+  // Analysis and block reference.
+  BasicBlock *BB;
+  ScalarEvolution *SE;
+  DataLayout *DL;
+  TargetTransformInfo *TTI;
+  AliasAnalysis *AA;
+};
+
+} // end of namespace
+# endif  //LLVM_TRANSFORMS_VECTORIZE_AOSVECTORIZER_H
+
diff --git a/lib/Transforms/Vectorize/Vectorize.cpp b/lib/Transforms/Vectorize/Vectorize.cpp
index 19eefd2..3aff636 100644
--- a/lib/Transforms/Vectorize/Vectorize.cpp
+++ b/lib/Transforms/Vectorize/Vectorize.cpp
@@ -28,6 +28,7 @@ using namespace llvm;
 void llvm::initializeVectorization(PassRegistry &Registry) {
   initializeBBVectorizePass(Registry);
   initializeLoopVectorizePass(Registry);
+  initializeSLPVectorizerPass(Registry);
 }
 
 void LLVMInitializeVectorization(LLVMPassRegistryRef R) {
@@ -41,3 +42,7 @@ void LLVMAddBBVectorizePass(LLVMPassManagerRef PM) {
 void LLVMAddLoopVectorizePass(LLVMPassManagerRef PM) {
   unwrap(PM)->add(createLoopVectorizePass());
 }
+
+void LLVMAddLoopRollerPass(LLVMPassManagerRef PM) {
+  unwrap(PM)->add(createSLPVectorizerPass());
+}
diff --git a/test/Transforms/SLPVectorizer/X86/flag.ll b/test/Transforms/SLPVectorizer/X86/flag.ll
new file mode 100644
index 0000000..a76ebd7
--- /dev/null
+++ b/test/Transforms/SLPVectorizer/X86/flag.ll
@@ -0,0 +1,51 @@
+; RUN: opt < %s -basicaa -slp-vectorizer -slp-threshold=1000 -dce -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.8.0"
+
+; Check that the command line flag works.
+;CHECK:rollable
+;CHECK-NOT:load <4 x i32>
+;CHECK: ret
+
+define i32 @rollable(i32* noalias nocapture %in, i32* noalias nocapture %out, i64 %n) nounwind ssp uwtable {
+  %1 = icmp eq i64 %n, 0
+  br i1 %1, label %._crit_edge, label %.lr.ph
+
+.lr.ph:                                           ; preds = %0, %.lr.ph
+  %i.019 = phi i64 [ %26, %.lr.ph ], [ 0, %0 ]
+  %2 = shl i64 %i.019, 2
+  %3 = getelementptr inbounds i32* %in, i64 %2
+  %4 = load i32* %3, align 4
+  %5 = or i64 %2, 1
+  %6 = getelementptr inbounds i32* %in, i64 %5
+  %7 = load i32* %6, align 4
+  %8 = or i64 %2, 2
+  %9 = getelementptr inbounds i32* %in, i64 %8
+  %10 = load i32* %9, align 4
+  %11 = or i64 %2, 3
+  %12 = getelementptr inbounds i32* %in, i64 %11
+  %13 = load i32* %12, align 4
+  %14 = mul i32 %4, 7
+  %15 = add i32 %14, 7
+  %16 = mul i32 %7, 7
+  %17 = add i32 %16, 14
+  %18 = mul i32 %10, 7
+  %19 = add i32 %18, 21
+  %20 = mul i32 %13, 7
+  %21 = add i32 %20, 28
+  %22 = getelementptr inbounds i32* %out, i64 %2
+  store i32 %15, i32* %22, align 4
+  %23 = getelementptr inbounds i32* %out, i64 %5
+  store i32 %17, i32* %23, align 4
+  %24 = getelementptr inbounds i32* %out, i64 %8
+  store i32 %19, i32* %24, align 4
+  %25 = getelementptr inbounds i32* %out, i64 %11
+  store i32 %21, i32* %25, align 4
+  %26 = add i64 %i.019, 1
+  %exitcond = icmp eq i64 %26, %n
+  br i1 %exitcond, label %._crit_edge, label %.lr.ph
+
+._crit_edge:                                      ; preds = %.lr.ph, %0
+  ret i32 undef
+}
diff --git a/test/Transforms/SLPVectorizer/X86/lit.local.cfg b/test/Transforms/SLPVectorizer/X86/lit.local.cfg
new file mode 100644
index 0000000..a8ad0f1
--- /dev/null
+++ b/test/Transforms/SLPVectorizer/X86/lit.local.cfg
@@ -0,0 +1,6 @@
+config.suffixes = ['.ll', '.c', '.cpp']
+
+targets = set(config.root.targets_to_build.split())
+if not 'X86' in targets:
+    config.unsupported = True
+
diff --git a/test/Transforms/SLPVectorizer/X86/multi_user.ll b/test/Transforms/SLPVectorizer/X86/multi_user.ll
new file mode 100644
index 0000000..cd47eb3
--- /dev/null
+++ b/test/Transforms/SLPVectorizer/X86/multi_user.ll
@@ -0,0 +1,47 @@
+; RUN: opt < %s -basicaa -slp-vectorizer -dce -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.7.0"
+
+;int foo (int *A, int n) {
+;  A[0] += n * 5 + 7;
+;  A[1] += n * 5 + 8;
+;  A[2] += n * 5 + 9;
+;  A[3] += n * 5 + 10;
+;  A[4] += n * 5 + 11;
+;}
+
+;CHECK: @foo
+;CHECK: insertelement <4 x i32>
+;CHECK: load <4 x i32>
+;CHECK: add <4 x i32>
+;CHECK: store <4 x i32>
+;CHECK: ret
+define i32 @foo(i32* nocapture %A, i32 %n) nounwind ssp uwtable {
+  %1 = mul nsw i32 %n, 5
+  %2 = add nsw i32 %1, 7
+  %3 = load i32* %A, align 4
+  %4 = add nsw i32 %2, %3
+  store i32 %4, i32* %A, align 4
+  %5 = add nsw i32 %1, 8
+  %6 = getelementptr inbounds i32* %A, i64 1
+  %7 = load i32* %6, align 4
+  %8 = add nsw i32 %5, %7
+  store i32 %8, i32* %6, align 4
+  %9 = add nsw i32 %1, 9
+  %10 = getelementptr inbounds i32* %A, i64 2
+  %11 = load i32* %10, align 4
+  %12 = add nsw i32 %9, %11
+  store i32 %12, i32* %10, align 4
+  %13 = add nsw i32 %1, 10
+  %14 = getelementptr inbounds i32* %A, i64 3
+  %15 = load i32* %14, align 4
+  %16 = add nsw i32 %13, %15
+  store i32 %16, i32* %14, align 4
+  %17 = add nsw i32 %1, 11
+  %18 = getelementptr inbounds i32* %A, i64 4
+  %19 = load i32* %18, align 4
+  %20 = add nsw i32 %17, %19
+  store i32 %20, i32* %18, align 4
+  ret i32 undef
+}
diff --git a/test/Transforms/SLPVectorizer/X86/saxpy.ll b/test/Transforms/SLPVectorizer/X86/saxpy.ll
new file mode 100644
index 0000000..5f06f9f
--- /dev/null
+++ b/test/Transforms/SLPVectorizer/X86/saxpy.ll
@@ -0,0 +1,50 @@
+; RUN: opt < %s -basicaa -slp-vectorizer -dce -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.8.0"
+
+; SLP vectorization example from http://cs.stanford.edu/people/eschkufz/research/asplos291-schkufza.pdf
+;CHECK: SAXPY
+;CHECK: mul <4 x i32>
+;CHECK: ret
+
+define void @SAXPY(i32* noalias nocapture %x, i32* noalias nocapture %y, i32 %a, i64 %i) #0 {
+  %1 = getelementptr inbounds i32* %x, i64 %i
+  %2 = load i32* %1, align 4, !tbaa !0
+  %3 = mul nsw i32 %2, %a
+  %4 = getelementptr inbounds i32* %y, i64 %i
+  %5 = load i32* %4, align 4, !tbaa !0
+  %6 = add nsw i32 %3, %5
+  store i32 %6, i32* %1, align 4, !tbaa !0
+  %7 = add i64 %i, 1
+  %8 = getelementptr inbounds i32* %x, i64 %7
+  %9 = load i32* %8, align 4, !tbaa !0
+  %10 = mul nsw i32 %9, %a
+  %11 = getelementptr inbounds i32* %y, i64 %7
+  %12 = load i32* %11, align 4, !tbaa !0
+  %13 = add nsw i32 %10, %12
+  store i32 %13, i32* %8, align 4, !tbaa !0
+  %14 = add i64 %i, 2
+  %15 = getelementptr inbounds i32* %x, i64 %14
+  %16 = load i32* %15, align 4, !tbaa !0
+  %17 = mul nsw i32 %16, %a
+  %18 = getelementptr inbounds i32* %y, i64 %14
+  %19 = load i32* %18, align 4, !tbaa !0
+  %20 = add nsw i32 %17, %19
+  store i32 %20, i32* %15, align 4, !tbaa !0
+  %21 = add i64 %i, 3
+  %22 = getelementptr inbounds i32* %x, i64 %21
+  %23 = load i32* %22, align 4, !tbaa !0
+  %24 = mul nsw i32 %23, %a
+  %25 = getelementptr inbounds i32* %y, i64 %21
+  %26 = load i32* %25, align 4, !tbaa !0
+  %27 = add nsw i32 %24, %26
+  store i32 %27, i32* %22, align 4, !tbaa !0
+  ret void
+}
+
+attributes #0 = { nounwind ssp uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf"="true" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "unsafe-fp-math"="false" "use-soft-float"="false" }
+
+!0 = metadata !{metadata !"int", metadata !1}
+!1 = metadata !{metadata !"omnipotent char", metadata !2}
+!2 = metadata !{metadata !"Simple C/C++ TBAA"}
diff --git a/test/Transforms/SLPVectorizer/X86/simple-loop.ll b/test/Transforms/SLPVectorizer/X86/simple-loop.ll
new file mode 100644
index 0000000..4c15ed0
--- /dev/null
+++ b/test/Transforms/SLPVectorizer/X86/simple-loop.ll
@@ -0,0 +1,100 @@
+; RUN: opt < %s -basicaa -slp-vectorizer -dce -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.8.0"
+
+;CHECK:rollable
+define i32 @rollable(i32* noalias nocapture %in, i32* noalias nocapture %out, i64 %n) nounwind ssp uwtable {
+  %1 = icmp eq i64 %n, 0
+  br i1 %1, label %._crit_edge, label %.lr.ph
+
+.lr.ph:                                           ; preds = %0, %.lr.ph
+  %i.019 = phi i64 [ %26, %.lr.ph ], [ 0, %0 ]
+  %2 = shl i64 %i.019, 2
+  %3 = getelementptr inbounds i32* %in, i64 %2
+;CHECK:load <4 x i32>
+  %4 = load i32* %3, align 4
+  %5 = or i64 %2, 1
+  %6 = getelementptr inbounds i32* %in, i64 %5
+  %7 = load i32* %6, align 4
+  %8 = or i64 %2, 2
+  %9 = getelementptr inbounds i32* %in, i64 %8
+  %10 = load i32* %9, align 4
+  %11 = or i64 %2, 3
+  %12 = getelementptr inbounds i32* %in, i64 %11
+  %13 = load i32* %12, align 4
+;CHECK:mul <4 x i32>
+  %14 = mul i32 %4, 7
+;CHECK:add <4 x i32>
+  %15 = add i32 %14, 7
+  %16 = mul i32 %7, 7
+  %17 = add i32 %16, 14
+  %18 = mul i32 %10, 7
+  %19 = add i32 %18, 21
+  %20 = mul i32 %13, 7
+  %21 = add i32 %20, 28
+  %22 = getelementptr inbounds i32* %out, i64 %2
+;CHECK:store <4 x i32>
+  store i32 %15, i32* %22, align 4
+  %23 = getelementptr inbounds i32* %out, i64 %5
+  store i32 %17, i32* %23, align 4
+  %24 = getelementptr inbounds i32* %out, i64 %8
+  store i32 %19, i32* %24, align 4
+  %25 = getelementptr inbounds i32* %out, i64 %11
+  store i32 %21, i32* %25, align 4
+  %26 = add i64 %i.019, 1
+  %exitcond = icmp eq i64 %26, %n
+  br i1 %exitcond, label %._crit_edge, label %.lr.ph
+
+._crit_edge:                                      ; preds = %.lr.ph, %0
+;CHECK: ret
+  ret i32 undef
+}
+
+;CHECK:unrollable
+;CHECK-NOT: <4 x i32>
+;CHECK: ret
+define i32 @unrollable(i32* %in, i32* %out, i64 %n) nounwind ssp uwtable {
+  %1 = icmp eq i64 %n, 0
+  br i1 %1, label %._crit_edge, label %.lr.ph
+
+.lr.ph:                                           ; preds = %0, %.lr.ph
+  %i.019 = phi i64 [ %26, %.lr.ph ], [ 0, %0 ]
+  %2 = shl i64 %i.019, 2
+  %3 = getelementptr inbounds i32* %in, i64 %2
+  %4 = load i32* %3, align 4
+  %5 = or i64 %2, 1
+  %6 = getelementptr inbounds i32* %in, i64 %5
+  %7 = load i32* %6, align 4
+  %8 = or i64 %2, 2
+  %9 = getelementptr inbounds i32* %in, i64 %8
+  %10 = load i32* %9, align 4
+  %11 = or i64 %2, 3
+  %12 = getelementptr inbounds i32* %in, i64 %11
+  %13 = load i32* %12, align 4
+  %14 = mul i32 %4, 7
+  %15 = add i32 %14, 7
+  %16 = mul i32 %7, 7
+  %17 = add i32 %16, 14
+  %18 = mul i32 %10, 7
+  %19 = add i32 %18, 21
+  %20 = mul i32 %13, 7
+  %21 = add i32 %20, 28
+  %22 = getelementptr inbounds i32* %out, i64 %2
+  store i32 %15, i32* %22, align 4
+  %23 = getelementptr inbounds i32* %out, i64 %5
+  store i32 %17, i32* %23, align 4
+  %barrier = call i32 @goo(i32 0)                      ; <---------------- memory barrier.
+  %24 = getelementptr inbounds i32* %out, i64 %8
+  store i32 %19, i32* %24, align 4
+  %25 = getelementptr inbounds i32* %out, i64 %11
+  store i32 %21, i32* %25, align 4
+  %26 = add i64 %i.019, 1
+  %exitcond = icmp eq i64 %26, %n
+  br i1 %exitcond, label %._crit_edge, label %.lr.ph
+
+._crit_edge:                                      ; preds = %.lr.ph, %0
+  ret i32 undef
+}
+
+declare i32 @goo(i32)
diff --git a/test/Transforms/SLPVectorizer/X86/simplebb.ll b/test/Transforms/SLPVectorizer/X86/simplebb.ll
new file mode 100644
index 0000000..0af30ab
--- /dev/null
+++ b/test/Transforms/SLPVectorizer/X86/simplebb.ll
@@ -0,0 +1,25 @@
+; RUN: opt < %s -basicaa -slp-vectorizer -dce -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.8.0"
+
+; Simple 3-pair chain with loads and stores
+; CHECK: test1
+; CHECK: store <2 x double>
+; CHECK: ret
+define void @test1(double* %a, double* %b, double* %c) nounwind uwtable readonly {
+entry:
+  %i0 = load double* %a, align 8
+  %i1 = load double* %b, align 8
+  %mul = fmul double %i0, %i1
+  %arrayidx3 = getelementptr inbounds double* %a, i64 1
+  %i3 = load double* %arrayidx3, align 8
+  %arrayidx4 = getelementptr inbounds double* %b, i64 1
+  %i4 = load double* %arrayidx4, align 8
+  %mul5 = fmul double %i3, %i4
+  store double %mul, double* %c, align 8
+  %arrayidx5 = getelementptr inbounds double* %c, i64 1
+  store double %mul5, double* %arrayidx5, align 8
+  ret void
+}
+
diff --git a/test/Transforms/SLPVectorizer/lit.local.cfg b/test/Transforms/SLPVectorizer/lit.local.cfg
new file mode 100644
index 0000000..19eebc0
--- /dev/null
+++ b/test/Transforms/SLPVectorizer/lit.local.cfg
@@ -0,0 +1 @@
+config.suffixes = ['.ll', '.c', '.cpp']