Add a loop vectorizer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@166112 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/InitializePasses.h

@@ -261,6 +261,7 @@ void initializeVirtRegRewriterPass(PassRegistry&);
 void initializeInstSimplifierPass(PassRegistry&);
 void initializeUnpackMachineBundlesPass(PassRegistry&);
 void initializeFinalizeMachineBundlesPass(PassRegistry&);
+void initializeLoopVectorizePass(PassRegistry&);
 void initializeBBVectorizePass(PassRegistry&);
 void initializeMachineFunctionPrinterPassPass(PassRegistry&);
 }

include/llvm/LinkAllPasses.h

@@ -156,6 +156,7 @@ namespace {
       (void) llvm::createCorrelatedValuePropagationPass();
       (void) llvm::createMemDepPrinter();
       (void) llvm::createInstructionSimplifierPass();
+      (void) llvm::createLoopVectorizePass();
       (void) llvm::createBBVectorizePass();
 
       (void)new llvm::IntervalPartition();

include/llvm/Transforms/Vectorize.h

@@ -106,6 +106,12 @@ struct VectorizeConfig {
 BasicBlockPass *
 createBBVectorizePass(const VectorizeConfig &C = VectorizeConfig());
 
+//===----------------------------------------------------------------------===//
+//
+// LoopVectorize - Create a loop vectorization pass.
+//
+Pass * createLoopVectorizePass();
+
 //===----------------------------------------------------------------------===//
 /// @brief Vectorize the BasicBlock.
 ///

lib/Transforms/IPO/PassManagerBuilder.cpp

@@ -176,6 +176,12 @@ void PassManagerBuilder::populateModulePassManager(PassManagerBase &MPM) {
   MPM.add(createIndVarSimplifyPass());        // Canonicalize indvars
   MPM.add(createLoopIdiomPass());             // Recognize idioms like memset.
   MPM.add(createLoopDeletionPass());          // Delete dead loops
+
+  if (Vectorize) {
+    MPM.add(createLoopVectorizePass());
+    MPM.add(createLICMPass());
+  }
+
   if (!DisableUnrollLoops)
     MPM.add(createLoopUnrollPass());          // Unroll small loops
   addExtensionsToPM(EP_LoopOptimizerEnd, MPM);

lib/Transforms/Vectorize/CMakeLists.txt

@@ -1,6 +1,7 @@
 add_llvm_library(LLVMVectorize
   BBVectorize.cpp
   Vectorize.cpp
+  LoopVectorize.cpp
   )
 
 add_dependencies(LLVMVectorize intrinsics_gen)

lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -0,0 +1,801 @@
+//===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This is a simple loop vectorizer. We currently only support single block
+// loops. We have a very simple and restrictive legality check: we need to read
+// and write from disjoint memory locations. We still don't have a cost model.
+// This pass has three parts:
+// 1. The main loop pass that drives the different parts.
+// 2. LoopVectorizationLegality - A helper class that checks for the legality
+//    of the vectorization.
+// 3. SingleBlockLoopVectorizer - A helper class that performs the actual
+//    widening of instructions.
+//
+//===----------------------------------------------------------------------===//
+#define LV_NAME "loop-vectorize"
+#define DEBUG_TYPE LV_NAME
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Instructions.h"
+#include "llvm/LLVMContext.h"
+#include "llvm/Pass.h"
+#include "llvm/Analysis/LoopPass.h"
+#include "llvm/Value.h"
+#include "llvm/Function.h"
+#include "llvm/Module.h"
+#include "llvm/Type.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/StringExtras.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/AliasSetTracker.h"
+#include "llvm/Transforms/Scalar.h"
+#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Analysis/ScalarEvolutionExpander.h"
+#include "llvm/Transforms/Utils/BasicBlockUtils.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/DataLayout.h"
+#include "llvm/Transforms/Utils/Local.h"
+#include <algorithm>
+using namespace llvm;
+
+static cl::opt<unsigned>
+DefaultVectorizationFactor("default-loop-vectorize-width",
+                          cl::init(4), cl::Hidden,
+                          cl::desc("Set the default loop vectorization width"));
+
+namespace {
+
+/// Vectorize a simple loop. This class performs the widening of simple single
+/// basic block loops into vectors. It does not perform any
+/// vectorization-legality checks, and just does it.  It widens the vectors
+/// to a given vectorization factor (VF).
+class SingleBlockLoopVectorizer {
+public:
+
+  /// Ctor.
+  SingleBlockLoopVectorizer(Loop *OrigLoop, ScalarEvolution *Se, LoopInfo *Li,
+                            unsigned VecWidth):
+  Orig(OrigLoop), SE(Se), LI(Li), VF(VecWidth),
+   Builder(0), Induction(0), OldInduction(0) { }
+
+  ~SingleBlockLoopVectorizer() {
+    delete Builder;
+  }
+
+  // Perform the actual loop widening (vectorization).
+  void vectorize() {
+    ///Create a new empty loop. Unlink the old loop and connect the new one.
+    copyEmptyLoop();
+    /// Widen each instruction in the old loop to a new one in the new loop.
+    vectorizeLoop();
+    // Delete the old loop.
+    deleteOldLoop();
+ }
+
+private:
+  /// Create an empty loop, based on the loop ranges of the old loop.
+  void copyEmptyLoop();
+  /// Copy and widen the instructions from the old loop.
+  void vectorizeLoop();
+  /// Delete the old loop.
+  void deleteOldLoop();
+
+  /// This instruction is un-vectorizable. Implement it as a sequence
+  /// of scalars.
+  void scalarizeInstruction(Instruction *Instr);
+
+  /// Create a broadcast instruction. This method generates a broadcast
+  /// instruction (shuffle) for loop invariant values and for the induction
+  /// value. If this is the induction variable then we extend it to N, N+1, ...
+  /// this is needed because each iteration in the loop corresponds to a SIMD
+  /// element.
+  Value *getBroadcastInstrs(Value *V);
+
+  /// This is a helper function used by getBroadcastInstrs. It adds 0, 1, 2 ..
+  /// for each element in the vector. Starting from zero.
+  Value *getConsecutiveVector(Value* Val);
+
+  /// Check that the GEP operands are all uniform except for the last index
+  /// which has to be the induction variable.
+  bool isConsecutiveGep(GetElementPtrInst *Gep);
+
+  /// When we go over instructions in the basic block we rely on previous
+  /// values within the current basic block or on loop invariant values.
+  /// When we widen (vectorize) values we place them in the map. If the values
+  /// are not within the map, they have to be loop invariant, so we simply
+  /// broadcast them into a vector.
+  Value *getVectorValue(Value *V);
+
+  /// The original loop.
+  Loop *Orig;
+  // Scev analysis to use.
+  ScalarEvolution *SE;
+  // Loop Info.
+  LoopInfo *LI;
+  // The vectorization factor to use.
+  unsigned VF;
+
+  // The builder that we use
+  IRBuilder<> *Builder;
+
+  // --- Vectorization state ---
+
+  /// The new Induction variable which was added to the new block.
+  Instruction *Induction;
+  /// The induction variable of the old basic block.
+  Instruction *OldInduction;
+  // Maps scalars to widened vectors.
+  DenseMap<Value*, Value*> WidenMap;
+};
+
+
+/// Perform the vectorization legality check. This class does not look at the
+/// profitability of vectorization, only the legality. At the moment the checks
+/// are very simple and focus on single basic block loops with a constant
+/// iteration count and no reductions.
+class LoopVectorizationLegality {
+public:
+  LoopVectorizationLegality(Loop *Lp, ScalarEvolution *Se, DataLayout *Dl):
+  TheLoop(Lp), SE(Se), DL(Dl) { }
+
+  /// Returns the maximum vectorization factor that we *can* use to vectorize
+  /// this loop. This does not mean that it is profitable to vectorize this
+  /// loop, only that it is legal to do so. This may be a large number. We
+  /// can vectorize to any SIMD width below this number.
+  unsigned getLoopMaxVF();
+
+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 canVectorizeBlock(BasicBlock &BB);
+
+  // Check if a pointer value is known to be disjoint.
+  // Example: Alloca, Global, NoAlias.
+  bool isKnownDisjoint(Value* Val);
+
+  /// The loop that we evaluate.
+  Loop *TheLoop;
+  /// Scev analysis.
+  ScalarEvolution *SE;
+  /// DataLayout analysis.
+  DataLayout *DL;
+};
+
+struct LoopVectorize : public LoopPass {
+  static char ID; // Pass identification, replacement for typeid
+
+  LoopVectorize() : LoopPass(ID) {
+    initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
+  }
+
+  AliasAnalysis *AA;
+  ScalarEvolution *SE;
+  DataLayout *DL;
+  LoopInfo *LI;
+
+  virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
+    // Only vectorize innermost loops.
+    if (!L->empty())
+      return false;
+
+    AA = &getAnalysis<AliasAnalysis>();
+    SE = &getAnalysis<ScalarEvolution>();
+    DL = getAnalysisIfAvailable<DataLayout>();
+    LI = &getAnalysis<LoopInfo>();
+
+    BasicBlock *Header = L->getHeader();
+    DEBUG(dbgs() << "LV: Checking a loop in \"" <<
+          Header->getParent()->getName() << "\"\n");
+
+    // Check if it is legal to vectorize the loop.
+    LoopVectorizationLegality LVL(L, SE, DL);
+    unsigned MaxVF = LVL.getLoopMaxVF();
+
+    // Check that we can vectorize using the chosen vectorization width.
+    if ((MaxVF < DefaultVectorizationFactor) ||
+        (MaxVF % DefaultVectorizationFactor)) {
+      DEBUG(dbgs() << "LV: non-vectorizable MaxVF ("<< MaxVF << ").\n");
+      return false;
+    }
+
+    DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< MaxVF << ").\n");
+
+    // If we decided that is is *legal* to vectorizer the loop. Do it.
+    SingleBlockLoopVectorizer LB(L, SE, LI, DefaultVectorizationFactor);
+    LB.vectorize();
+
+    // The loop is now vectorized. Remove it from LMP.
+    LPM.deleteLoopFromQueue(L);
+    return true;
+  }
+
+  virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+    LoopPass::getAnalysisUsage(AU);
+    AU.addRequiredID(LoopSimplifyID);
+    AU.addRequired<AliasAnalysis>();
+    AU.addRequired<LoopInfo>();
+    AU.addRequired<ScalarEvolution>();
+  }
+
+};
+
+Value *SingleBlockLoopVectorizer::getBroadcastInstrs(Value *V) {
+  // Instructions that access the old induction variable
+  // actually want to get the new one.
+  if (V == OldInduction)
+    V = Induction;
+  // Create the types.
+  LLVMContext &C = V->getContext();
+  Type *VTy = VectorType::get(V->getType(), VF);
+  Type *I32 = IntegerType::getInt32Ty(C);
+  Constant *Zero = ConstantInt::get(I32, 0);
+  Value *Zeros = ConstantAggregateZero::get(VectorType::get(I32, VF));
+  Value *UndefVal = UndefValue::get(VTy);
+  // Insert the value into a new vector.
+  Value *SingleElem = Builder->CreateInsertElement(UndefVal, V, Zero);
+  // Broadcast the scalar into all locations in the vector.
+  Value *Shuf = Builder->CreateShuffleVector(SingleElem, UndefVal, Zeros,
+                                             "broadcast");
+  // We are accessing the induction variable. Make sure to promote the
+  // index for each consecutive SIMD lane. This adds 0,1,2 ... to all lanes.
+  if (V == Induction)
+    return getConsecutiveVector(Shuf);
+  return Shuf;
+}
+
+Value *SingleBlockLoopVectorizer::getConsecutiveVector(Value* Val) {
+  assert(Val->getType()->isVectorTy() && "Must be a vector");
+  assert(Val->getType()->getScalarType()->isIntegerTy() &&
+         "Elem must be an integer");
+  // Create the types.
+  Type *ITy = Val->getType()->getScalarType();
+  VectorType *Ty = cast<VectorType>(Val->getType());
+  unsigned VLen = Ty->getNumElements();
+  SmallVector<Constant*, 8> Indices;
+
+  // Create a vector of consecutive numbers from zero to VF.
+  for (unsigned i = 0; i < VLen; ++i)
+    Indices.push_back(ConstantInt::get(ITy, i));
+
+  // Add the consecutive indices to the vector value.
+  Constant *Cv = ConstantVector::get(Indices);
+  assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
+  return Builder->CreateAdd(Val, Cv, "induction");
+}
+
+
+bool SingleBlockLoopVectorizer::isConsecutiveGep(GetElementPtrInst *Gep) {
+  if (!Gep)
+    return false;
+
+  unsigned NumOperands = Gep->getNumOperands();
+  Value *LastIndex = Gep->getOperand(NumOperands - 1);
+
+  // Check that all of the gep indices are uniform except for the last.
+  for (unsigned i = 0; i < NumOperands - 1; ++i)
+    if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), Orig))
+      return false;
+
+  // The last operand has to be the induction in order to emit
+  // a wide load/store.
+  const SCEV *Last = SE->getSCEV(LastIndex);
+  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
+    const SCEV *Step = AR->getStepRecurrence(*SE);
+
+    // The memory is consecutive because the last index is consecutive
+    // and all other indices are loop invariant.
+    if (Step->isOne())
+      return true;
+  }
+
+  return false;
+}
+
+Value *SingleBlockLoopVectorizer::getVectorValue(Value *V) {
+  if (WidenMap.count(V))
+    return WidenMap[V];
+  return getBroadcastInstrs(V);
+}
+
+void SingleBlockLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
+  assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
+  // Holds vector parameters or scalars, in case of uniform vals.
+  SmallVector<Value*, 8> Params;
+
+  // Find all of the vectorized parameters.
+  for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
+    Value *SrcOp = Instr->getOperand(op);
+
+    // If we are accessing the old induction variable, use the new one.
+    if (SrcOp == OldInduction) {
+      Params.push_back(getBroadcastInstrs(Induction));
+      continue;
+    }
+
+    // Try using previously calculated values.
+    Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
+
+    // If the src is an instruction that appeared earlier in the basic block
+    // then it should already be vectorized. 
+    if (SrcInst && SrcInst->getParent() == Instr->getParent()) {
+      assert(WidenMap.count(SrcInst) && "Source operand is unavailable");
+      // The parameter is a vector value from earlier.
+      Params.push_back(WidenMap[SrcInst]);
+    } else {
+      // The parameter is a scalar from outside the loop. Maybe even a constant.
+      Params.push_back(SrcOp);
+    }
+  }
+
+  assert(Params.size() == Instr->getNumOperands() &&
+         "Invalid number of operands");
+
+  // Does this instruction return a value ?
+  bool IsVoidRetTy = Instr->getType()->isVoidTy();
+  Value *VecResults = 0;
+
+  // If we have a return value, create an empty vector. We place the scalarized
+  // instructions in this vector.
+  if (!IsVoidRetTy)
+    VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF));
+
+  // For each scalar that we create.
+  for (unsigned i = 0; i < VF; ++i) {
+    Instruction *Cloned = Instr->clone();
+    if (!IsVoidRetTy)
+      Cloned->setName(Instr->getName() + ".cloned");
+    // Replace the operands of the cloned instrucions with extracted scalars.
+    for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
+      Value *Op = Params[op];
+      // Param is a vector. Need to extract the right lane.
+      if (Op->getType()->isVectorTy())
+        Op = Builder->CreateExtractElement(Op, Builder->getInt32(i));
+      Cloned->setOperand(op, Op);
+    }
+
+    // Place the clonsed scalar in the new loop.
+    Builder->Insert(Cloned);
+
+    // If the original scalar returns a value we need to place it in a vector
+    // so that future users will be able to use it.
+    if (!IsVoidRetTy)
+      VecResults = Builder->CreateInsertElement(VecResults, Cloned,
+                                               Builder->getInt32(i));
+  }
+
+  if (!IsVoidRetTy)
+    WidenMap[Instr] = VecResults;
+}
+
+void SingleBlockLoopVectorizer::copyEmptyLoop() {
+  assert(Orig->getNumBlocks() == 1 && "Invalid loop");
+  BasicBlock *PH = Orig->getLoopPreheader();
+  BasicBlock *ExitBlock = Orig->getExitBlock();
+  assert(ExitBlock && "Invalid loop exit");
+
+  // Create a new single-basic block loop.
+  BasicBlock *BB = BasicBlock::Create(PH->getContext(), "vectorizedloop",
+                                      PH->getParent(), ExitBlock);
+
+  // Find the induction variable.
+  BasicBlock *OldBasicBlock = Orig->getHeader();
+  PHINode *OldInd = dyn_cast<PHINode>(OldBasicBlock->begin());
+  assert(OldInd && "We must have a single phi node.");
+  Type *IdxTy = OldInd->getType();
+
+  // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
+  // inside the loop.
+  Builder = new IRBuilder<>(BB);
+  Builder->SetInsertPoint(BB);
+
+  // Generate the induction variable.
+  PHINode *Phi = Builder->CreatePHI(IdxTy, 2, "index");
+  Constant *Zero = ConstantInt::get(IdxTy, 0);
+  Constant *Step = ConstantInt::get(IdxTy, VF);
+
+  // Find the loop boundaries.
+  const SCEV *ExitCount = SE->getExitCount(Orig, Orig->getHeader());
+  assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
+
+  // Get the trip count from the count by adding 1.
+  ExitCount = SE->getAddExpr(ExitCount,
+                             SE->getConstant(ExitCount->getType(), 1));
+
+  // Expand the trip count and place the new instructions in the preheader.
+  // Notice that the pre-header does not change, only the loop body.
+  SCEVExpander Exp(*SE, "induction");
+  Instruction *Loc = Orig->getLoopPreheader()->getTerminator();
+  if (ExitCount->getType() != Phi->getType())
+    ExitCount = SE->getSignExtendExpr(ExitCount, Phi->getType());
+  Value *Count = Exp.expandCodeFor(ExitCount, Phi->getType(), Loc);
+  
+  // Create i+1 and fill the PHINode.
+  Value *Next = Builder->CreateAdd(Phi, Step, "index.next");
+  Phi->addIncoming(Zero, PH);
+  Phi->addIncoming(Next, BB);
+  // Create the compare.
+  Value *ICmp = Builder->CreateICmpEQ(Next, Count);
+  Builder->CreateCondBr(ICmp, ExitBlock, BB);
+  // Fix preheader.
+  PH->getTerminator()->setSuccessor(0, BB);
+  Builder->SetInsertPoint(BB->getFirstInsertionPt());
+
+  // Save the indiction variables.
+  Induction = Phi;
+  OldInduction = OldInd;
+}
+
+void SingleBlockLoopVectorizer::vectorizeLoop() {
+  BasicBlock &BB = *Orig->getHeader();
+
+  // For each instruction in the old loop.
+  for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
+  Instruction *Inst = it;
+
+    switch (Inst->getOpcode()) {
+      case Instruction::PHI:
+      case Instruction::Br:
+        // Nothing to do for PHIs and BR, since we already took care of the
+        // loop control flow instructions.
+        continue;
+
+      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: {
+        // Just widen binops.
+        BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
+        Value *A = getVectorValue(Inst->getOperand(0));
+        Value *B = getVectorValue(Inst->getOperand(1));
+        // Use this vector value for all users of the original instruction.
+        WidenMap[Inst] = Builder->CreateBinOp(BinOp->getOpcode(), A, B);
+        break;
+      }
+      case Instruction::Select: {
+        // Widen selects.
+        Value *A = getVectorValue(Inst->getOperand(0));
+        Value *B = getVectorValue(Inst->getOperand(1));
+        Value *C = getVectorValue(Inst->getOperand(2));
+        WidenMap[Inst] = Builder->CreateSelect(A, B, C);
+        break;
+      }
+
+      case Instruction::ICmp:
+      case Instruction::FCmp: {
+        // Widen compares. Generate vector compares.
+        bool FCmp = (Inst->getOpcode() == Instruction::FCmp);
+        CmpInst *Cmp = dyn_cast<CmpInst>(Inst);
+        Value *A = getVectorValue(Inst->getOperand(0));
+        Value *B = getVectorValue(Inst->getOperand(1));
+        if (FCmp)
+          WidenMap[Inst] = Builder->CreateFCmp(Cmp->getPredicate(), A, B);
+        else
+          WidenMap[Inst] = Builder->CreateICmp(Cmp->getPredicate(), A, B);
+        break;
+      }
+
+      case Instruction::Store: {
+        // Attempt to issue a wide store.
+        StoreInst *SI = dyn_cast<StoreInst>(Inst);
+        Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
+        Value *Ptr = SI->getPointerOperand();
+        unsigned Alignment = SI->getAlignment();
+        GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
+        // This store does not use GEPs.
+        if (!isConsecutiveGep(Gep)) {
+          scalarizeInstruction(Inst);
+          break;
+        }
+
+        // Create the new GEP with the new induction variable.
+        GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
+        unsigned NumOperands = Gep->getNumOperands();
+        Gep2->setOperand(NumOperands - 1, Induction);
+        Ptr = Builder->Insert(Gep2);
+        Ptr = Builder->CreateBitCast(Ptr, StTy->getPointerTo());
+        Value *Val = getVectorValue(SI->getValueOperand());
+        Builder->CreateStore(Val, Ptr)->setAlignment(Alignment);
+        break;
+      }
+      case Instruction::Load: {
+        // Attempt to issue a wide load.
+        LoadInst *LI = dyn_cast<LoadInst>(Inst);
+        Type *RetTy = VectorType::get(LI->getType(), VF);
+        Value *Ptr = LI->getPointerOperand();
+        unsigned Alignment = LI->getAlignment();
+        GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
+
+        // We don't have a gep. Scalarize the load.
+        if (!isConsecutiveGep(Gep)) {
+          scalarizeInstruction(Inst);
+          break;
+        }
+
+        // Create the new GEP with the new induction variable.
+        GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
+        unsigned NumOperands = Gep->getNumOperands();
+        Gep2->setOperand(NumOperands - 1, Induction);
+        Ptr = Builder->Insert(Gep2);
+        Ptr = Builder->CreateBitCast(Ptr, RetTy->getPointerTo());
+        LI = Builder->CreateLoad(Ptr);
+        LI->setAlignment(Alignment);
+        // Use this vector value for all users of the load.
+        WidenMap[Inst] = LI;
+        break;
+      }
+      case Instruction::ZExt:
+      case Instruction::SExt:
+      case Instruction::FPToUI:
+      case Instruction::FPToSI:
+      case Instruction::FPExt:
+      case Instruction::PtrToInt:
+      case Instruction::IntToPtr:
+      case Instruction::SIToFP:
+      case Instruction::UIToFP:
+      case Instruction::Trunc:
+      case Instruction::FPTrunc:
+      case Instruction::BitCast: {
+        /// Vectorize bitcasts.
+        CastInst *CI = dyn_cast<CastInst>(Inst);
+        Value *A = getVectorValue(Inst->getOperand(0));
+        Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
+        WidenMap[Inst] = Builder->CreateCast(CI->getOpcode(), A, DestTy);
+        break;
+      }
+
+      default:
+        /// All other instructions are unsupported. Scalarize them.
+        scalarizeInstruction(Inst);
+        break;
+    }// end of switch.
+  }// end of for_each instr.
+}
+
+void SingleBlockLoopVectorizer::deleteOldLoop() {
+  // The original basic block.
+  BasicBlock *BB = Orig->getHeader();
+  SE->forgetLoop(Orig);
+
+  LI->removeBlock(BB);
+  Orig->addBasicBlockToLoop(Induction->getParent(), LI->getBase());
+
+  // Remove the old loop block.
+  DeleteDeadBlock(BB);
+}
+
+unsigned LoopVectorizationLegality::getLoopMaxVF() {
+  if (!TheLoop->getLoopPreheader()) {
+    assert(false && "No preheader!!");
+    DEBUG(dbgs() << "LV: Loop not normalized." << "\n");
+    return  1;
+  }
+
+  // We can only vectorize single basic block loops.
+  unsigned NumBlocks = TheLoop->getNumBlocks();
+  if (NumBlocks != 1) {
+    DEBUG(dbgs() << "LV: Too many blocks:" << NumBlocks << "\n");
+    return 1;
+  }
+
+  // We need to have a loop header.
+  BasicBlock *BB = TheLoop->getHeader();
+  DEBUG(dbgs() << "LV: Found a loop: " << BB->getName() << "\n");
+
+  // Find the max vectorization factor.
+  unsigned MaxVF = SE->getSmallConstantTripMultiple(TheLoop, BB);
+
+
+  // Perform an early check. Do not scan the block if we did not find a loop.
+  if (MaxVF < 2) {
+    DEBUG(dbgs() << "LV: Can't find a vectorizable loop structure\n");
+    return 1;
+  }
+
+  // Go over each instruction and look at memory deps.
+  if (!canVectorizeBlock(*BB)) {
+    DEBUG(dbgs() << "LV: Can't vectorize this loop header\n");
+    return 1;
+  }
+
+  DEBUG(dbgs() << "LV: We can vectorize this loop! VF="<<MaxVF<<"\n");
+  
+  // Okay! We can vectorize. Return the max trip multiple.
+  return MaxVF;
+}
+
+bool LoopVectorizationLegality::canVectorizeBlock(BasicBlock &BB) {
+  // Holds the read and write pointers that we find.
+  typedef SmallVector<Value*, 10> ValueVector;
+  ValueVector Reads;
+  ValueVector Writes;
+
+  unsigned NumPhis = 0;
+  for (BasicBlock::iterator it = BB.begin(), e = BB.end(); it != e; ++it) {
+    Instruction *I = it;
+
+    PHINode *Phi = dyn_cast<PHINode>(I);
+    if (Phi) {
+      NumPhis++;
+      // We only look at integer phi nodes.
+      if (!Phi->getType()->isIntegerTy()) {
+        DEBUG(dbgs() << "LV: Found an non-int PHI.\n");
+        return false;
+      }
+
+      // If we found an induction variable.
+      if (NumPhis > 1) {
+        DEBUG(dbgs() << "LV: Found more than one PHI.\n");
+        return false;
+      }
+
+      // This should not happen because the loop should be normalized.
+      if (Phi->getNumIncomingValues() != 2) {
+        DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
+        return false;
+      }
+
+      // Check that the PHI is consecutive and starts at zero.
+      const SCEV *PhiScev = SE->getSCEV(Phi);
+      const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
+      if (!AR) {
+        DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
+        return false;
+      }
+
+      const SCEV *Step = AR->getStepRecurrence(*SE);
+      const SCEV *Start = AR->getStart();
+
+      if (!Step->isOne() || !Start->isZero()) {
+        DEBUG(dbgs() << "LV: PHI does not start at zero or steps by one.\n");
+        return false;
+      }
+    }
+
+    // IF this is a load, record its pointer. If it is not a load, abort.
+    // Notice that we don't handle function calls that read or write.
+    if (I->mayReadFromMemory()) {
+      LoadInst *Ld = dyn_cast<LoadInst>(I);
+      if (!Ld) return false;
+      if (!Ld->isSimple()) {
+        DEBUG(dbgs() << "LV: Found a non-simple load.\n");
+        return false;
+      }
+      GetUnderlyingObjects(Ld->getPointerOperand(), Reads, DL);
+    }
+
+    // Record store pointers. Abort on all other instructions that write to
+    // memory.
+    if (I->mayWriteToMemory()) {
+      StoreInst *St = dyn_cast<StoreInst>(I);
+      if (!St) return false;
+      if (!St->isSimple()) {
+        DEBUG(dbgs() << "LV: Found a non-simple store.\n");
+        return false;
+      }
+      GetUnderlyingObjects(St->getPointerOperand(), Writes, DL);
+    }
+
+    // We still don't handle functions.
+    CallInst *CI = dyn_cast<CallInst>(I);
+    if (CI) {
+      DEBUG(dbgs() << "LV: Found a call site:"<<
+            CI->getCalledFunction()->getName() << "\n");
+      return false;
+    }
+
+    // We do not re-vectorize vectors.
+    if (!VectorType::isValidElementType(I->getType()) &&
+        !I->getType()->isVoidTy()) {
+      DEBUG(dbgs() << "LV: Found unvectorizable type." << "\n");
+      return false;
+    }
+    //Check that all of the users of the loop are inside the BB.
+    for (Value::use_iterator it = I->use_begin(), e = I->use_end();
+         it != e; ++it) {
+      Instruction *U = cast<Instruction>(*it);
+      BasicBlock *Parent = U->getParent();
+      if (Parent != &BB) {
+        DEBUG(dbgs() << "LV: Found an outside user for : "<< *U << "\n");
+        return false;
+      }
+    }
+  } // next instr.
+
+  // Check that the underlying objects of the reads and writes are either
+  // disjoint memory locations, or that they are no-alias arguments.
+  ValueVector::iterator r, re, w, we;
+  for (r = Reads.begin(), re = Reads.end(); r != re; ++r) {
+    if (!isKnownDisjoint(*r)) {
+      DEBUG(dbgs() << "LV: Found a bad read Ptr: "<< **r << "\n");
+      return false;
+    }
+  }
+
+  for (w = Writes.begin(), we = Writes.end(); w != we; ++w) {
+    if (!isKnownDisjoint(*w)) {
+      DEBUG(dbgs() << "LV: Found a bad write Ptr: "<< **w << "\n");
+      return false;
+    }
+  }
+
+  // Check that there are no multiple write locations to the same pointer.
+  SmallPtrSet<Value*, 8> BasePointers;
+  for (w = Writes.begin(), we = Writes.end(); w != we; ++w) {
+    if (BasePointers.count(*w)) {
+      DEBUG(dbgs() << "LV: Multiple writes to the same index :"<< **w << "\n");
+      return false;
+    }
+    BasePointers.insert(*w);
+  }
+
+  // Sort the writes vector so that we can use a binary search.
+  std::sort(Writes.begin(), Writes.end());
+  // Check that the reads and the writes are disjoint.
+  for (r = Reads.begin(), re = Reads.end(); r != re; ++r) {
+    if (std::binary_search(Writes.begin(), Writes.end(), *r)) {
+      DEBUG(dbgs() << "Vectorizer: Found a read/write ptr:"<< **r << "\n");
+      return false;
+    }
+  }
+
+  // All is okay.
+  return true;
+}
+
+/// Checks if the value is a Global variable or if it is an Arguments
+/// marked with the NoAlias attribute.
+bool LoopVectorizationLegality::isKnownDisjoint(Value* Val) {
+  assert(Val && "Invalid value");
+  if (dyn_cast<GlobalValue>(Val))
+    return true;
+  if (dyn_cast<AllocaInst>(Val))
+    return true;
+  Argument *A = dyn_cast<Argument>(Val);
+  if (!A)
+    return false;
+  return A->hasNoAliasAttr();
+}
+
+} // namespace
+
+char LoopVectorize::ID = 0;
+static const char lv_name[] = "Loop Vectorization";
+INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
+INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
+INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
+INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
+
+namespace llvm {
+  Pass *createLoopVectorizePass() {
+    return new LoopVectorize();
+  }
+
+}
+

lib/Transforms/Vectorize/Vectorize.cpp

@@ -7,7 +7,7 @@
 //
 //===----------------------------------------------------------------------===//
 //
-// This file implements common infrastructure for libLLVMVectorizeOpts.a, which 
+// This file implements common infrastructure for libLLVMVectorizeOpts.a, which
 // implements several vectorization transformations over the LLVM intermediate
 // representation, including the C bindings for that library.
 //
@@ -23,10 +23,11 @@
 
 using namespace llvm;
 
-/// initializeVectorizationPasses - Initialize all passes linked into the 
+/// initializeVectorizationPasses - Initialize all passes linked into the
 /// Vectorization library.
 void llvm::initializeVectorization(PassRegistry &Registry) {
   initializeBBVectorizePass(Registry);
+  initializeLoopVectorizePass(Registry);
 }
 
 void LLVMInitializeVectorization(LLVMPassRegistryRef R) {
@@ -37,3 +38,6 @@ void LLVMAddBBVectorizePass(LLVMPassManagerRef PM) {
   unwrap(PM)->add(createBBVectorizePass());
 }
 
+void LLVMAddLoopVectorizePass(LLVMPassManagerRef PM) {
+  unwrap(PM)->add(createLoopVectorizePass());
+}

test/Transforms/LoopVectorize/gcc-examples.ll

@@ -0,0 +1,651 @@
+; RUN: opt < %s  -loop-vectorize -dce -instcombine -licm -S | 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"
+
+@b = common global [2048 x i32] zeroinitializer, align 16
+@c = common global [2048 x i32] zeroinitializer, align 16
+@a = common global [2048 x i32] zeroinitializer, align 16
+@G = common global [32 x [1024 x i32]] zeroinitializer, align 16
+@ub = common global [1024 x i32] zeroinitializer, align 16
+@uc = common global [1024 x i32] zeroinitializer, align 16
+@d = common global [2048 x i32] zeroinitializer, align 16
+@fa = common global [1024 x float] zeroinitializer, align 16
+@fb = common global [1024 x float] zeroinitializer, align 16
+@ic = common global [1024 x i32] zeroinitializer, align 16
+@da = common global [1024 x float] zeroinitializer, align 16
+@db = common global [1024 x float] zeroinitializer, align 16
+@dc = common global [1024 x float] zeroinitializer, align 16
+@dd = common global [1024 x float] zeroinitializer, align 16
+@dj = common global [1024 x i32] zeroinitializer, align 16
+
+;CHECK: @example1
+;CHECK: load <4 x i32>
+;CHECK: add <4 x i32>
+;CHECK: store <4 x i32>
+;CHECK: ret void
+define void @example1() nounwind uwtable ssp {
+  br label %1
+
+; <label>:1                                       ; preds = %1, %0
+  %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+  %2 = getelementptr inbounds [2048 x i32]* @b, i64 0, i64 %indvars.iv
+  %3 = load i32* %2, align 4
+  %4 = getelementptr inbounds [2048 x i32]* @c, i64 0, i64 %indvars.iv
+  %5 = load i32* %4, align 4
+  %6 = add nsw i32 %5, %3
+  %7 = getelementptr inbounds [2048 x i32]* @a, i64 0, i64 %indvars.iv
+  store i32 %6, i32* %7, align 4
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 256
+  br i1 %exitcond, label %8, label %1
+
+; <label>:8                                       ; preds = %1
+  ret void
+}
+
+; We can't vectorize this loop because it has non constant loop bounds.
+;CHECK: @example2
+;CHECK-NOT: <4 x i32>
+;CHECK: ret void
+define void @example2(i32 %n, i32 %x) nounwind uwtable ssp {
+  %1 = icmp sgt i32 %n, 0
+  br i1 %1, label %.lr.ph5, label %.preheader
+
+..preheader_crit_edge:                            ; preds = %.lr.ph5
+  %phitmp = sext i32 %n to i64
+  br label %.preheader
+
+.preheader:                                       ; preds = %..preheader_crit_edge, %0
+  %i.0.lcssa = phi i64 [ %phitmp, %..preheader_crit_edge ], [ 0, %0 ]
+  %2 = icmp eq i32 %n, 0
+  br i1 %2, label %._crit_edge, label %.lr.ph
+
+.lr.ph5:                                          ; preds = %0, %.lr.ph5
+  %indvars.iv6 = phi i64 [ %indvars.iv.next7, %.lr.ph5 ], [ 0, %0 ]
+  %3 = getelementptr inbounds [2048 x i32]* @b, i64 0, i64 %indvars.iv6
+  store i32 %x, i32* %3, align 4
+  %indvars.iv.next7 = add i64 %indvars.iv6, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next7 to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, %n
+  br i1 %exitcond, label %..preheader_crit_edge, label %.lr.ph5
+
+.lr.ph:                                           ; preds = %.preheader, %.lr.ph
+  %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ %i.0.lcssa, %.preheader ]
+  %.02 = phi i32 [ %4, %.lr.ph ], [ %n, %.preheader ]
+  %4 = add nsw i32 %.02, -1
+  %5 = getelementptr inbounds [2048 x i32]* @b, i64 0, i64 %indvars.iv
+  %6 = load i32* %5, align 4
+  %7 = getelementptr inbounds [2048 x i32]* @c, i64 0, i64 %indvars.iv
+  %8 = load i32* %7, align 4
+  %9 = and i32 %8, %6
+  %10 = getelementptr inbounds [2048 x i32]* @a, i64 0, i64 %indvars.iv
+  store i32 %9, i32* %10, align 4
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %11 = icmp eq i32 %4, 0
+  br i1 %11, label %._crit_edge, label %.lr.ph
+
+._crit_edge:                                      ; preds = %.lr.ph, %.preheader
+  ret void
+}
+
+; We can't vectorize this loop because it has non constant loop bounds.
+;CHECK: @example3
+;CHECK-NOT: <4 x i32>
+;CHECK: ret void
+define void @example3(i32 %n, i32* noalias nocapture %p, i32* noalias nocapture %q) nounwind uwtable ssp {
+  %1 = icmp eq i32 %n, 0
+  br i1 %1, label %._crit_edge, label %.lr.ph
+
+.lr.ph:                                           ; preds = %0, %.lr.ph
+  %.05 = phi i32 [ %2, %.lr.ph ], [ %n, %0 ]
+  %.014 = phi i32* [ %5, %.lr.ph ], [ %p, %0 ]
+  %.023 = phi i32* [ %3, %.lr.ph ], [ %q, %0 ]
+  %2 = add nsw i32 %.05, -1
+  %3 = getelementptr inbounds i32* %.023, i64 1
+  %4 = load i32* %.023, align 16
+  %5 = getelementptr inbounds i32* %.014, i64 1
+  store i32 %4, i32* %.014, align 16
+  %6 = icmp eq i32 %2, 0
+  br i1 %6, label %._crit_edge, label %.lr.ph
+
+._crit_edge:                                      ; preds = %.lr.ph, %0
+  ret void
+}
+
+; We can't vectorize this loop because it has non constant loop bounds.
+;CHECK: @example4
+;CHECK-NOT: <4 x i32>
+;CHECK: ret void
+define void @example4(i32 %n, i32* noalias nocapture %p, i32* noalias nocapture %q) nounwind uwtable ssp {
+  %1 = add nsw i32 %n, -1
+  %2 = icmp eq i32 %n, 0
+  br i1 %2, label %.preheader4, label %.lr.ph10
+
+.preheader4:                                      ; preds = %0
+  %3 = icmp sgt i32 %1, 0
+  br i1 %3, label %.lr.ph6, label %._crit_edge
+
+.lr.ph10:                                         ; preds = %0, %.lr.ph10
+  %4 = phi i32 [ %9, %.lr.ph10 ], [ %1, %0 ]
+  %.018 = phi i32* [ %8, %.lr.ph10 ], [ %p, %0 ]
+  %.027 = phi i32* [ %5, %.lr.ph10 ], [ %q, %0 ]
+  %5 = getelementptr inbounds i32* %.027, i64 1
+  %6 = load i32* %.027, align 16
+  %7 = add nsw i32 %6, 5
+  %8 = getelementptr inbounds i32* %.018, i64 1
+  store i32 %7, i32* %.018, align 16
+  %9 = add nsw i32 %4, -1
+  %10 = icmp eq i32 %4, 0
+  br i1 %10, label %._crit_edge, label %.lr.ph10
+
+.preheader:                                       ; preds = %.lr.ph6
+  br i1 %3, label %.lr.ph, label %._crit_edge
+
+.lr.ph6:                                          ; preds = %.preheader4, %.lr.ph6
+  %indvars.iv11 = phi i64 [ %indvars.iv.next12, %.lr.ph6 ], [ 0, %.preheader4 ]
+  %indvars.iv.next12 = add i64 %indvars.iv11, 1
+  %11 = getelementptr inbounds [2048 x i32]* @b, i64 0, i64 %indvars.iv.next12
+  %12 = load i32* %11, align 4
+  %13 = add nsw i64 %indvars.iv11, 3
+  %14 = getelementptr inbounds [2048 x i32]* @c, i64 0, i64 %13
+  %15 = load i32* %14, align 4
+  %16 = add nsw i32 %15, %12
+  %17 = getelementptr inbounds [2048 x i32]* @a, i64 0, i64 %indvars.iv11
+  store i32 %16, i32* %17, align 4
+  %lftr.wideiv13 = trunc i64 %indvars.iv.next12 to i32
+  %exitcond14 = icmp eq i32 %lftr.wideiv13, %1
+  br i1 %exitcond14, label %.preheader, label %.lr.ph6
+
+.lr.ph:                                           ; preds = %.preheader, %.lr.ph
+  %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %.preheader ]
+  %18 = getelementptr inbounds [2048 x i32]* @a, i64 0, i64 %indvars.iv
+  %19 = load i32* %18, align 4
+  %20 = icmp sgt i32 %19, 4
+  %21 = select i1 %20, i32 4, i32 0
+  %22 = getelementptr inbounds [2048 x i32]* @b, i64 0, i64 %indvars.iv
+  store i32 %21, i32* %22, align 4
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, %1
+  br i1 %exitcond, label %._crit_edge, label %.lr.ph
+
+._crit_edge:                                      ; preds = %.lr.ph10, %.preheader4, %.lr.ph, %.preheader
+  ret void
+}
+
+;CHECK: @example8
+;CHECK: store <4 x i32>
+;CHECK: ret void
+define void @example8(i32 %x) nounwind uwtable ssp {
+  br label %.preheader
+
+.preheader:                                       ; preds = %3, %0
+  %indvars.iv3 = phi i64 [ 0, %0 ], [ %indvars.iv.next4, %3 ]
+  br label %1
+
+; <label>:1                                       ; preds = %1, %.preheader
+  %indvars.iv = phi i64 [ 0, %.preheader ], [ %indvars.iv.next, %1 ]
+  %2 = getelementptr inbounds [32 x [1024 x i32]]* @G, i64 0, i64 %indvars.iv3, i64 %indvars.iv
+  store i32 %x, i32* %2, align 4
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 1024
+  br i1 %exitcond, label %3, label %1
+
+; <label>:3                                       ; preds = %1
+  %indvars.iv.next4 = add i64 %indvars.iv3, 1
+  %lftr.wideiv5 = trunc i64 %indvars.iv.next4 to i32
+  %exitcond6 = icmp eq i32 %lftr.wideiv5, 32
+  br i1 %exitcond6, label %4, label %.preheader
+
+; <label>:4                                       ; preds = %3
+  ret void
+}
+
+; We can't vectorize because it has a reduction variable.
+;CHECK: @example9
+;CHECK-NOT: <4 x i32>
+;CHECK: ret i32
+define i32 @example9() nounwind uwtable readonly ssp {
+  br label %1
+
+; <label>:1                                       ; preds = %1, %0
+  %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+  %diff.01 = phi i32 [ 0, %0 ], [ %7, %1 ]
+  %2 = getelementptr inbounds [1024 x i32]* @ub, i64 0, i64 %indvars.iv
+  %3 = load i32* %2, align 4
+  %4 = getelementptr inbounds [1024 x i32]* @uc, i64 0, i64 %indvars.iv
+  %5 = load i32* %4, align 4
+  %6 = add i32 %3, %diff.01
+  %7 = sub i32 %6, %5
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 1024
+  br i1 %exitcond, label %8, label %1
+
+; <label>:8                                       ; preds = %1
+  ret i32 %7
+}
+
+;CHECK: @example10a
+;CHECK: load <4 x i16>
+;CHECK: add <4 x i16>
+;CHECK: store <4 x i16>
+;CHECK: ret void
+define void @example10a(i16* noalias nocapture %sa, i16* noalias nocapture %sb, i16* noalias nocapture %sc, i32* noalias nocapture %ia, i32* noalias nocapture %ib, i32* noalias nocapture %ic) nounwind uwtable ssp {
+  br label %1
+
+; <label>:1                                       ; preds = %1, %0
+  %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+  %2 = getelementptr inbounds i32* %ib, i64 %indvars.iv
+  %3 = load i32* %2, align 4
+  %4 = getelementptr inbounds i32* %ic, i64 %indvars.iv
+  %5 = load i32* %4, align 4
+  %6 = add nsw i32 %5, %3
+  %7 = getelementptr inbounds i32* %ia, i64 %indvars.iv
+  store i32 %6, i32* %7, align 4
+  %8 = getelementptr inbounds i16* %sb, i64 %indvars.iv
+  %9 = load i16* %8, align 2
+  %10 = getelementptr inbounds i16* %sc, i64 %indvars.iv
+  %11 = load i16* %10, align 2
+  %12 = add i16 %11, %9
+  %13 = getelementptr inbounds i16* %sa, i64 %indvars.iv
+  store i16 %12, i16* %13, align 2
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 1024
+  br i1 %exitcond, label %14, label %1
+
+; <label>:14                                      ; preds = %1
+  ret void
+}
+
+;CHECK: @example10b
+;CHECK: load <4 x i16>
+;CHECK: sext <4 x i16>
+;CHECK: store <4 x i32>
+;CHECK: ret void
+define void @example10b(i16* noalias nocapture %sa, i16* noalias nocapture %sb, i16* noalias nocapture %sc, i32* noalias nocapture %ia, i32* noalias nocapture %ib, i32* noalias nocapture %ic) nounwind uwtable ssp {
+  br label %1
+
+; <label>:1                                       ; preds = %1, %0
+  %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+  %2 = getelementptr inbounds i16* %sb, i64 %indvars.iv
+  %3 = load i16* %2, align 2
+  %4 = sext i16 %3 to i32
+  %5 = getelementptr inbounds i32* %ia, i64 %indvars.iv
+  store i32 %4, i32* %5, align 4
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 1024
+  br i1 %exitcond, label %6, label %1
+
+; <label>:6                                       ; preds = %1
+  ret void
+}
+
+;CHECK: @example11
+;CHECK: load i32
+;CHECK: load i32
+;CHECK: load i32
+;CHECK: load i32
+;CHECK: insertelement
+;CHECK: insertelement
+;CHECK: insertelement
+;CHECK: insertelement
+;CHECK: ret void
+define void @example11() nounwind uwtable ssp {
+  br label %1
+
+; <label>:1                                       ; preds = %1, %0
+  %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+  %2 = shl nsw i64 %indvars.iv, 1
+  %3 = or i64 %2, 1
+  %4 = getelementptr inbounds [2048 x i32]* @b, i64 0, i64 %3
+  %5 = load i32* %4, align 4
+  %6 = getelementptr inbounds [2048 x i32]* @c, i64 0, i64 %3
+  %7 = load i32* %6, align 4
+  %8 = mul nsw i32 %7, %5
+  %9 = getelementptr inbounds [2048 x i32]* @b, i64 0, i64 %2
+  %10 = load i32* %9, align 8
+  %11 = getelementptr inbounds [2048 x i32]* @c, i64 0, i64 %2
+  %12 = load i32* %11, align 8
+  %13 = mul nsw i32 %12, %10
+  %14 = sub nsw i32 %8, %13
+  %15 = getelementptr inbounds [2048 x i32]* @a, i64 0, i64 %indvars.iv
+  store i32 %14, i32* %15, align 4
+  %16 = mul nsw i32 %7, %10
+  %17 = mul nsw i32 %12, %5
+  %18 = add nsw i32 %17, %16
+  %19 = getelementptr inbounds [2048 x i32]* @d, i64 0, i64 %indvars.iv
+  store i32 %18, i32* %19, align 4
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 512
+  br i1 %exitcond, label %20, label %1
+
+; <label>:20                                      ; preds = %1
+  ret void
+}
+
+;CHECK: @example12
+;CHECK: trunc <4 x i64>
+;CHECK: store <4 x i32>
+;CHECK: ret void
+define void @example12() nounwind uwtable ssp {
+  br label %1
+
+; <label>:1                                       ; preds = %1, %0
+  %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+  %2 = getelementptr inbounds [2048 x i32]* @a, i64 0, i64 %indvars.iv
+  %3 = trunc i64 %indvars.iv to i32
+  store i32 %3, i32* %2, align 4
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 1024
+  br i1 %exitcond, label %4, label %1
+
+; <label>:4                                       ; preds = %1
+  ret void
+}
+
+; Can't vectorize because of reductions.
+;CHECK: @example13
+;CHECK-NOT: <4 x i32>
+;CHECK: ret void
+define void @example13(i32** nocapture %A, i32** nocapture %B, i32* nocapture %out) nounwind uwtable ssp {
+  br label %.preheader
+
+.preheader:                                       ; preds = %14, %0
+  %indvars.iv4 = phi i64 [ 0, %0 ], [ %indvars.iv.next5, %14 ]
+  %1 = getelementptr inbounds i32** %A, i64 %indvars.iv4
+  %2 = load i32** %1, align 8
+  %3 = getelementptr inbounds i32** %B, i64 %indvars.iv4
+  %4 = load i32** %3, align 8
+  br label %5
+
+; <label>:5                                       ; preds = %.preheader, %5
+  %indvars.iv = phi i64 [ 0, %.preheader ], [ %indvars.iv.next, %5 ]
+  %diff.02 = phi i32 [ 0, %.preheader ], [ %11, %5 ]
+  %6 = getelementptr inbounds i32* %2, i64 %indvars.iv
+  %7 = load i32* %6, align 4
+  %8 = getelementptr inbounds i32* %4, i64 %indvars.iv
+  %9 = load i32* %8, align 4
+  %10 = add i32 %7, %diff.02
+  %11 = sub i32 %10, %9
+  %indvars.iv.next = add i64 %indvars.iv, 8
+  %12 = trunc i64 %indvars.iv.next to i32
+  %13 = icmp slt i32 %12, 1024
+  br i1 %13, label %5, label %14
+
+; <label>:14                                      ; preds = %5
+  %15 = getelementptr inbounds i32* %out, i64 %indvars.iv4
+  store i32 %11, i32* %15, align 4
+  %indvars.iv.next5 = add i64 %indvars.iv4, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next5 to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 32
+  br i1 %exitcond, label %16, label %.preheader
+
+; <label>:16                                      ; preds = %14
+  ret void
+}
+
+; Can't vectorize because of reductions.
+;CHECK: @example14
+;CHECK-NOT: <4 x i32>
+;CHECK: ret void
+define void @example14(i32** nocapture %in, i32** nocapture %coeff, i32* nocapture %out) nounwind uwtable ssp {
+.preheader3:
+  br label %.preheader
+
+.preheader:                                       ; preds = %11, %.preheader3
+  %indvars.iv7 = phi i64 [ 0, %.preheader3 ], [ %indvars.iv.next8, %11 ]
+  %sum.05 = phi i32 [ 0, %.preheader3 ], [ %10, %11 ]
+  br label %0
+
+; <label>:0                                       ; preds = %0, %.preheader
+  %indvars.iv = phi i64 [ 0, %.preheader ], [ %indvars.iv.next, %0 ]
+  %sum.12 = phi i32 [ %sum.05, %.preheader ], [ %10, %0 ]
+  %1 = getelementptr inbounds i32** %in, i64 %indvars.iv
+  %2 = load i32** %1, align 8
+  %3 = getelementptr inbounds i32* %2, i64 %indvars.iv7
+  %4 = load i32* %3, align 4
+  %5 = getelementptr inbounds i32** %coeff, i64 %indvars.iv
+  %6 = load i32** %5, align 8
+  %7 = getelementptr inbounds i32* %6, i64 %indvars.iv7
+  %8 = load i32* %7, align 4
+  %9 = mul nsw i32 %8, %4
+  %10 = add nsw i32 %9, %sum.12
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 1024
+  br i1 %exitcond, label %11, label %0
+
+; <label>:11                                      ; preds = %0
+  %indvars.iv.next8 = add i64 %indvars.iv7, 1
+  %lftr.wideiv9 = trunc i64 %indvars.iv.next8 to i32
+  %exitcond10 = icmp eq i32 %lftr.wideiv9, 32
+  br i1 %exitcond10, label %.preheader3.1, label %.preheader
+
+.preheader3.1:                                    ; preds = %11
+  store i32 %10, i32* %out, align 4
+  br label %.preheader.1
+
+.preheader.1:                                     ; preds = %24, %.preheader3.1
+  %indvars.iv7.1 = phi i64 [ 0, %.preheader3.1 ], [ %indvars.iv.next8.1, %24 ]
+  %sum.05.1 = phi i32 [ 0, %.preheader3.1 ], [ %23, %24 ]
+  br label %12
+
+; <label>:12                                      ; preds = %12, %.preheader.1
+  %indvars.iv.1 = phi i64 [ 0, %.preheader.1 ], [ %13, %12 ]
+  %sum.12.1 = phi i32 [ %sum.05.1, %.preheader.1 ], [ %23, %12 ]
+  %13 = add nsw i64 %indvars.iv.1, 1
+  %14 = getelementptr inbounds i32** %in, i64 %13
+  %15 = load i32** %14, align 8
+  %16 = getelementptr inbounds i32* %15, i64 %indvars.iv7.1
+  %17 = load i32* %16, align 4
+  %18 = getelementptr inbounds i32** %coeff, i64 %indvars.iv.1
+  %19 = load i32** %18, align 8
+  %20 = getelementptr inbounds i32* %19, i64 %indvars.iv7.1
+  %21 = load i32* %20, align 4
+  %22 = mul nsw i32 %21, %17
+  %23 = add nsw i32 %22, %sum.12.1
+  %lftr.wideiv.1 = trunc i64 %13 to i32
+  %exitcond.1 = icmp eq i32 %lftr.wideiv.1, 1024
+  br i1 %exitcond.1, label %24, label %12
+
+; <label>:24                                      ; preds = %12
+  %indvars.iv.next8.1 = add i64 %indvars.iv7.1, 1
+  %lftr.wideiv9.1 = trunc i64 %indvars.iv.next8.1 to i32
+  %exitcond10.1 = icmp eq i32 %lftr.wideiv9.1, 32
+  br i1 %exitcond10.1, label %.preheader3.2, label %.preheader.1
+
+.preheader3.2:                                    ; preds = %24
+  %25 = getelementptr inbounds i32* %out, i64 1
+  store i32 %23, i32* %25, align 4
+  br label %.preheader.2
+
+.preheader.2:                                     ; preds = %38, %.preheader3.2
+  %indvars.iv7.2 = phi i64 [ 0, %.preheader3.2 ], [ %indvars.iv.next8.2, %38 ]
+  %sum.05.2 = phi i32 [ 0, %.preheader3.2 ], [ %37, %38 ]
+  br label %26
+
+; <label>:26                                      ; preds = %26, %.preheader.2
+  %indvars.iv.2 = phi i64 [ 0, %.preheader.2 ], [ %indvars.iv.next.2, %26 ]
+  %sum.12.2 = phi i32 [ %sum.05.2, %.preheader.2 ], [ %37, %26 ]
+  %27 = add nsw i64 %indvars.iv.2, 2
+  %28 = getelementptr inbounds i32** %in, i64 %27
+  %29 = load i32** %28, align 8
+  %30 = getelementptr inbounds i32* %29, i64 %indvars.iv7.2
+  %31 = load i32* %30, align 4
+  %32 = getelementptr inbounds i32** %coeff, i64 %indvars.iv.2
+  %33 = load i32** %32, align 8
+  %34 = getelementptr inbounds i32* %33, i64 %indvars.iv7.2
+  %35 = load i32* %34, align 4
+  %36 = mul nsw i32 %35, %31
+  %37 = add nsw i32 %36, %sum.12.2
+  %indvars.iv.next.2 = add i64 %indvars.iv.2, 1
+  %lftr.wideiv.2 = trunc i64 %indvars.iv.next.2 to i32
+  %exitcond.2 = icmp eq i32 %lftr.wideiv.2, 1024
+  br i1 %exitcond.2, label %38, label %26
+
+; <label>:38                                      ; preds = %26
+  %indvars.iv.next8.2 = add i64 %indvars.iv7.2, 1
+  %lftr.wideiv9.2 = trunc i64 %indvars.iv.next8.2 to i32
+  %exitcond10.2 = icmp eq i32 %lftr.wideiv9.2, 32
+  br i1 %exitcond10.2, label %.preheader3.3, label %.preheader.2
+
+.preheader3.3:                                    ; preds = %38
+  %39 = getelementptr inbounds i32* %out, i64 2
+  store i32 %37, i32* %39, align 4
+  br label %.preheader.3
+
+.preheader.3:                                     ; preds = %52, %.preheader3.3
+  %indvars.iv7.3 = phi i64 [ 0, %.preheader3.3 ], [ %indvars.iv.next8.3, %52 ]
+  %sum.05.3 = phi i32 [ 0, %.preheader3.3 ], [ %51, %52 ]
+  br label %40
+
+; <label>:40                                      ; preds = %40, %.preheader.3
+  %indvars.iv.3 = phi i64 [ 0, %.preheader.3 ], [ %indvars.iv.next.3, %40 ]
+  %sum.12.3 = phi i32 [ %sum.05.3, %.preheader.3 ], [ %51, %40 ]
+  %41 = add nsw i64 %indvars.iv.3, 3
+  %42 = getelementptr inbounds i32** %in, i64 %41
+  %43 = load i32** %42, align 8
+  %44 = getelementptr inbounds i32* %43, i64 %indvars.iv7.3
+  %45 = load i32* %44, align 4
+  %46 = getelementptr inbounds i32** %coeff, i64 %indvars.iv.3
+  %47 = load i32** %46, align 8
+  %48 = getelementptr inbounds i32* %47, i64 %indvars.iv7.3
+  %49 = load i32* %48, align 4
+  %50 = mul nsw i32 %49, %45
+  %51 = add nsw i32 %50, %sum.12.3
+  %indvars.iv.next.3 = add i64 %indvars.iv.3, 1
+  %lftr.wideiv.3 = trunc i64 %indvars.iv.next.3 to i32
+  %exitcond.3 = icmp eq i32 %lftr.wideiv.3, 1024
+  br i1 %exitcond.3, label %52, label %40
+
+; <label>:52                                      ; preds = %40
+  %indvars.iv.next8.3 = add i64 %indvars.iv7.3, 1
+  %lftr.wideiv9.3 = trunc i64 %indvars.iv.next8.3 to i32
+  %exitcond10.3 = icmp eq i32 %lftr.wideiv9.3, 32
+  br i1 %exitcond10.3, label %53, label %.preheader.3
+
+; <label>:53                                      ; preds = %52
+  %54 = getelementptr inbounds i32* %out, i64 3
+  store i32 %51, i32* %54, align 4
+  ret void
+}
+
+; Can't vectorize because the src and dst pointers are not disjoint.
+;CHECK: @example21
+;CHECK-NOT: <4 x i32>
+;CHECK: ret i32
+define i32 @example21(i32* nocapture %b, i32 %n) nounwind uwtable readonly ssp {
+  %1 = icmp sgt i32 %n, 0
+  br i1 %1, label %.lr.ph, label %._crit_edge
+
+.lr.ph:                                           ; preds = %0
+  %2 = sext i32 %n to i64
+  br label %3
+
+; <label>:3                                       ; preds = %.lr.ph, %3
+  %indvars.iv = phi i64 [ %2, %.lr.ph ], [ %indvars.iv.next, %3 ]
+  %a.02 = phi i32 [ 0, %.lr.ph ], [ %6, %3 ]
+  %indvars.iv.next = add i64 %indvars.iv, -1
+  %4 = getelementptr inbounds i32* %b, i64 %indvars.iv.next
+  %5 = load i32* %4, align 4
+  %6 = add nsw i32 %5, %a.02
+  %7 = trunc i64 %indvars.iv.next to i32
+  %8 = icmp sgt i32 %7, 0
+  br i1 %8, label %3, label %._crit_edge
+
+._crit_edge:                                      ; preds = %3, %0
+  %a.0.lcssa = phi i32 [ 0, %0 ], [ %6, %3 ]
+  ret i32 %a.0.lcssa
+}
+
+; Can't vectorize because there are multiple PHIs.
+;CHECK: @example23
+;CHECK-NOT: <4 x i32>
+;CHECK: ret void
+define void @example23(i16* nocapture %src, i32* nocapture %dst) nounwind uwtable ssp {
+  br label %1
+
+; <label>:1                                       ; preds = %1, %0
+  %.04 = phi i16* [ %src, %0 ], [ %2, %1 ]
+  %.013 = phi i32* [ %dst, %0 ], [ %6, %1 ]
+  %i.02 = phi i32 [ 0, %0 ], [ %7, %1 ]
+  %2 = getelementptr inbounds i16* %.04, i64 1
+  %3 = load i16* %.04, align 2
+  %4 = zext i16 %3 to i32
+  %5 = shl nuw nsw i32 %4, 7
+  %6 = getelementptr inbounds i32* %.013, i64 1
+  store i32 %5, i32* %.013, align 4
+  %7 = add nsw i32 %i.02, 1
+  %exitcond = icmp eq i32 %7, 256
+  br i1 %exitcond, label %8, label %1
+
+; <label>:8                                       ; preds = %1
+  ret void
+}
+
+;CHECK: @example24
+;CHECK: shufflevector <4 x i16>
+;CHECK: ret void
+define void @example24(i16 signext %x, i16 signext %y) nounwind uwtable ssp {
+  br label %1
+
+; <label>:1                                       ; preds = %1, %0
+  %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+  %2 = getelementptr inbounds [1024 x float]* @fa, i64 0, i64 %indvars.iv
+  %3 = load float* %2, align 4
+  %4 = getelementptr inbounds [1024 x float]* @fb, i64 0, i64 %indvars.iv
+  %5 = load float* %4, align 4
+  %6 = fcmp olt float %3, %5
+  %x.y = select i1 %6, i16 %x, i16 %y
+  %7 = sext i16 %x.y to i32
+  %8 = getelementptr inbounds [1024 x i32]* @ic, i64 0, i64 %indvars.iv
+  store i32 %7, i32* %8, align 4
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 1024
+  br i1 %exitcond, label %9, label %1
+
+; <label>:9                                       ; preds = %1
+  ret void
+}
+
+;CHECK: @example25
+;CHECK: and <4 x i1>
+;CHECK: zext <4 x i1>
+;CHECK: ret void
+define void @example25() nounwind uwtable ssp {
+  br label %1
+
+; <label>:1                                       ; preds = %1, %0
+  %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+  %2 = getelementptr inbounds [1024 x float]* @da, i64 0, i64 %indvars.iv
+  %3 = load float* %2, align 4
+  %4 = getelementptr inbounds [1024 x float]* @db, i64 0, i64 %indvars.iv
+  %5 = load float* %4, align 4
+  %6 = fcmp olt float %3, %5
+  %7 = getelementptr inbounds [1024 x float]* @dc, i64 0, i64 %indvars.iv
+  %8 = load float* %7, align 4
+  %9 = getelementptr inbounds [1024 x float]* @dd, i64 0, i64 %indvars.iv
+  %10 = load float* %9, align 4
+  %11 = fcmp olt float %8, %10
+  %12 = and i1 %6, %11
+  %13 = zext i1 %12 to i32
+  %14 = getelementptr inbounds [1024 x i32]* @dj, i64 0, i64 %indvars.iv
+  store i32 %13, i32* %14, align 4
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, 1024
+  br i1 %exitcond, label %15, label %1
+
+; <label>:15                                      ; preds = %1
+  ret void
+}
+

test/Transforms/LoopVectorize/lit.local.cfg

@@ -0,0 +1 @@
+config.suffixes = ['.ll', '.c', '.cpp']

test/Transforms/LoopVectorize/non-const-n.ll

@@ -0,0 +1,38 @@
+; RUN: opt < %s  -loop-vectorize -dce -instcombine -licm -S | 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"
+
+@b = common global [2048 x i32] zeroinitializer, align 16
+@c = common global [2048 x i32] zeroinitializer, align 16
+@a = common global [2048 x i32] zeroinitializer, align 16
+
+;CHECK: @example1
+;CHECK: shl i32
+;CHECK: sext i32
+;CHECK: load <4 x i32>
+;CHECK: add <4 x i32>
+;CHECK: store <4 x i32>
+;CHECK: ret void
+define void @example1(i32 %n) nounwind uwtable ssp {
+  %n4 = shl i32 %n, 2
+  br label %1
+
+; <label>:1                                       ; preds = %1, %0
+  %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %1 ]
+  %2 = getelementptr inbounds [2048 x i32]* @b, i64 0, i64 %indvars.iv
+  %3 = load i32* %2, align 4
+  %4 = getelementptr inbounds [2048 x i32]* @c, i64 0, i64 %indvars.iv
+  %5 = load i32* %4, align 4
+  %6 = add nsw i32 %5, %3
+  %7 = getelementptr inbounds [2048 x i32]* @a, i64 0, i64 %indvars.iv
+  store i32 %6, i32* %7, align 4
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, %n4
+  br i1 %exitcond, label %8, label %1
+
+; <label>:8                                       ; preds = %1
+  ret void
+}
+