Home | History | Annotate | Download | only in Vectorize
      1 //===- BBVectorize.cpp - A Basic-Block Vectorizer -------------------------===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 //
     10 // This file implements a basic-block vectorization pass. The algorithm was
     11 // inspired by that used by the Vienna MAP Vectorizor by Franchetti and Kral,
     12 // et al. It works by looking for chains of pairable operations and then
     13 // pairing them.
     14 //
     15 //===----------------------------------------------------------------------===//
     16 
     17 #define BBV_NAME "bb-vectorize"
     18 #define DEBUG_TYPE BBV_NAME
     19 #include "llvm/Transforms/Vectorize.h"
     20 #include "llvm/ADT/DenseMap.h"
     21 #include "llvm/ADT/DenseSet.h"
     22 #include "llvm/ADT/STLExtras.h"
     23 #include "llvm/ADT/SmallSet.h"
     24 #include "llvm/ADT/SmallVector.h"
     25 #include "llvm/ADT/Statistic.h"
     26 #include "llvm/ADT/StringExtras.h"
     27 #include "llvm/Analysis/AliasAnalysis.h"
     28 #include "llvm/Analysis/AliasSetTracker.h"
     29 #include "llvm/Analysis/Dominators.h"
     30 #include "llvm/Analysis/ScalarEvolution.h"
     31 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
     32 #include "llvm/Analysis/TargetTransformInfo.h"
     33 #include "llvm/Analysis/ValueTracking.h"
     34 #include "llvm/IR/Constants.h"
     35 #include "llvm/IR/DataLayout.h"
     36 #include "llvm/IR/DerivedTypes.h"
     37 #include "llvm/IR/Function.h"
     38 #include "llvm/IR/Instructions.h"
     39 #include "llvm/IR/IntrinsicInst.h"
     40 #include "llvm/IR/Intrinsics.h"
     41 #include "llvm/IR/LLVMContext.h"
     42 #include "llvm/IR/Metadata.h"
     43 #include "llvm/IR/Type.h"
     44 #include "llvm/Pass.h"
     45 #include "llvm/Support/CommandLine.h"
     46 #include "llvm/Support/Debug.h"
     47 #include "llvm/Support/ValueHandle.h"
     48 #include "llvm/Support/raw_ostream.h"
     49 #include "llvm/Transforms/Utils/Local.h"
     50 #include <algorithm>
     51 using namespace llvm;
     52 
     53 static cl::opt<bool>
     54 IgnoreTargetInfo("bb-vectorize-ignore-target-info",  cl::init(false),
     55   cl::Hidden, cl::desc("Ignore target information"));
     56 
     57 static cl::opt<unsigned>
     58 ReqChainDepth("bb-vectorize-req-chain-depth", cl::init(6), cl::Hidden,
     59   cl::desc("The required chain depth for vectorization"));
     60 
     61 static cl::opt<bool>
     62 UseChainDepthWithTI("bb-vectorize-use-chain-depth",  cl::init(false),
     63   cl::Hidden, cl::desc("Use the chain depth requirement with"
     64                        " target information"));
     65 
     66 static cl::opt<unsigned>
     67 SearchLimit("bb-vectorize-search-limit", cl::init(400), cl::Hidden,
     68   cl::desc("The maximum search distance for instruction pairs"));
     69 
     70 static cl::opt<bool>
     71 SplatBreaksChain("bb-vectorize-splat-breaks-chain", cl::init(false), cl::Hidden,
     72   cl::desc("Replicating one element to a pair breaks the chain"));
     73 
     74 static cl::opt<unsigned>
     75 VectorBits("bb-vectorize-vector-bits", cl::init(128), cl::Hidden,
     76   cl::desc("The size of the native vector registers"));
     77 
     78 static cl::opt<unsigned>
     79 MaxIter("bb-vectorize-max-iter", cl::init(0), cl::Hidden,
     80   cl::desc("The maximum number of pairing iterations"));
     81 
     82 static cl::opt<bool>
     83 Pow2LenOnly("bb-vectorize-pow2-len-only", cl::init(false), cl::Hidden,
     84   cl::desc("Don't try to form non-2^n-length vectors"));
     85 
     86 static cl::opt<unsigned>
     87 MaxInsts("bb-vectorize-max-instr-per-group", cl::init(500), cl::Hidden,
     88   cl::desc("The maximum number of pairable instructions per group"));
     89 
     90 static cl::opt<unsigned>
     91 MaxPairs("bb-vectorize-max-pairs-per-group", cl::init(3000), cl::Hidden,
     92   cl::desc("The maximum number of candidate instruction pairs per group"));
     93 
     94 static cl::opt<unsigned>
     95 MaxCandPairsForCycleCheck("bb-vectorize-max-cycle-check-pairs", cl::init(200),
     96   cl::Hidden, cl::desc("The maximum number of candidate pairs with which to use"
     97                        " a full cycle check"));
     98 
     99 static cl::opt<bool>
    100 NoBools("bb-vectorize-no-bools", cl::init(false), cl::Hidden,
    101   cl::desc("Don't try to vectorize boolean (i1) values"));
    102 
    103 static cl::opt<bool>
    104 NoInts("bb-vectorize-no-ints", cl::init(false), cl::Hidden,
    105   cl::desc("Don't try to vectorize integer values"));
    106 
    107 static cl::opt<bool>
    108 NoFloats("bb-vectorize-no-floats", cl::init(false), cl::Hidden,
    109   cl::desc("Don't try to vectorize floating-point values"));
    110 
    111 // FIXME: This should default to false once pointer vector support works.
    112 static cl::opt<bool>
    113 NoPointers("bb-vectorize-no-pointers", cl::init(/*false*/ true), cl::Hidden,
    114   cl::desc("Don't try to vectorize pointer values"));
    115 
    116 static cl::opt<bool>
    117 NoCasts("bb-vectorize-no-casts", cl::init(false), cl::Hidden,
    118   cl::desc("Don't try to vectorize casting (conversion) operations"));
    119 
    120 static cl::opt<bool>
    121 NoMath("bb-vectorize-no-math", cl::init(false), cl::Hidden,
    122   cl::desc("Don't try to vectorize floating-point math intrinsics"));
    123 
    124 static cl::opt<bool>
    125 NoFMA("bb-vectorize-no-fma", cl::init(false), cl::Hidden,
    126   cl::desc("Don't try to vectorize the fused-multiply-add intrinsic"));
    127 
    128 static cl::opt<bool>
    129 NoSelect("bb-vectorize-no-select", cl::init(false), cl::Hidden,
    130   cl::desc("Don't try to vectorize select instructions"));
    131 
    132 static cl::opt<bool>
    133 NoCmp("bb-vectorize-no-cmp", cl::init(false), cl::Hidden,
    134   cl::desc("Don't try to vectorize comparison instructions"));
    135 
    136 static cl::opt<bool>
    137 NoGEP("bb-vectorize-no-gep", cl::init(false), cl::Hidden,
    138   cl::desc("Don't try to vectorize getelementptr instructions"));
    139 
    140 static cl::opt<bool>
    141 NoMemOps("bb-vectorize-no-mem-ops", cl::init(false), cl::Hidden,
    142   cl::desc("Don't try to vectorize loads and stores"));
    143 
    144 static cl::opt<bool>
    145 AlignedOnly("bb-vectorize-aligned-only", cl::init(false), cl::Hidden,
    146   cl::desc("Only generate aligned loads and stores"));
    147 
    148 static cl::opt<bool>
    149 NoMemOpBoost("bb-vectorize-no-mem-op-boost",
    150   cl::init(false), cl::Hidden,
    151   cl::desc("Don't boost the chain-depth contribution of loads and stores"));
    152 
    153 static cl::opt<bool>
    154 FastDep("bb-vectorize-fast-dep", cl::init(false), cl::Hidden,
    155   cl::desc("Use a fast instruction dependency analysis"));
    156 
    157 #ifndef NDEBUG
    158 static cl::opt<bool>
    159 DebugInstructionExamination("bb-vectorize-debug-instruction-examination",
    160   cl::init(false), cl::Hidden,
    161   cl::desc("When debugging is enabled, output information on the"
    162            " instruction-examination process"));
    163 static cl::opt<bool>
    164 DebugCandidateSelection("bb-vectorize-debug-candidate-selection",
    165   cl::init(false), cl::Hidden,
    166   cl::desc("When debugging is enabled, output information on the"
    167            " candidate-selection process"));
    168 static cl::opt<bool>
    169 DebugPairSelection("bb-vectorize-debug-pair-selection",
    170   cl::init(false), cl::Hidden,
    171   cl::desc("When debugging is enabled, output information on the"
    172            " pair-selection process"));
    173 static cl::opt<bool>
    174 DebugCycleCheck("bb-vectorize-debug-cycle-check",
    175   cl::init(false), cl::Hidden,
    176   cl::desc("When debugging is enabled, output information on the"
    177            " cycle-checking process"));
    178 
    179 static cl::opt<bool>
    180 PrintAfterEveryPair("bb-vectorize-debug-print-after-every-pair",
    181   cl::init(false), cl::Hidden,
    182   cl::desc("When debugging is enabled, dump the basic block after"
    183            " every pair is fused"));
    184 #endif
    185 
    186 STATISTIC(NumFusedOps, "Number of operations fused by bb-vectorize");
    187 
    188 namespace {
    189   struct BBVectorize : public BasicBlockPass {
    190     static char ID; // Pass identification, replacement for typeid
    191 
    192     const VectorizeConfig Config;
    193 
    194     BBVectorize(const VectorizeConfig &C = VectorizeConfig())
    195       : BasicBlockPass(ID), Config(C) {
    196       initializeBBVectorizePass(*PassRegistry::getPassRegistry());
    197     }
    198 
    199     BBVectorize(Pass *P, const VectorizeConfig &C)
    200       : BasicBlockPass(ID), Config(C) {
    201       AA = &P->getAnalysis<AliasAnalysis>();
    202       DT = &P->getAnalysis<DominatorTree>();
    203       SE = &P->getAnalysis<ScalarEvolution>();
    204       TD = P->getAnalysisIfAvailable<DataLayout>();
    205       TTI = IgnoreTargetInfo ? 0 : &P->getAnalysis<TargetTransformInfo>();
    206     }
    207 
    208     typedef std::pair<Value *, Value *> ValuePair;
    209     typedef std::pair<ValuePair, int> ValuePairWithCost;
    210     typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
    211     typedef std::pair<ValuePair, ValuePair> VPPair; // A ValuePair pair
    212     typedef std::pair<VPPair, unsigned> VPPairWithType;
    213 
    214     AliasAnalysis *AA;
    215     DominatorTree *DT;
    216     ScalarEvolution *SE;
    217     DataLayout *TD;
    218     const TargetTransformInfo *TTI;
    219 
    220     // FIXME: const correct?
    221 
    222     bool vectorizePairs(BasicBlock &BB, bool NonPow2Len = false);
    223 
    224     bool getCandidatePairs(BasicBlock &BB,
    225                        BasicBlock::iterator &Start,
    226                        DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
    227                        DenseSet<ValuePair> &FixedOrderPairs,
    228                        DenseMap<ValuePair, int> &CandidatePairCostSavings,
    229                        std::vector<Value *> &PairableInsts, bool NonPow2Len);
    230 
    231     // FIXME: The current implementation does not account for pairs that
    232     // are connected in multiple ways. For example:
    233     //   C1 = A1 / A2; C2 = A2 / A1 (which may be both direct and a swap)
    234     enum PairConnectionType {
    235       PairConnectionDirect,
    236       PairConnectionSwap,
    237       PairConnectionSplat
    238     };
    239 
    240     void computeConnectedPairs(
    241              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
    242              DenseSet<ValuePair> &CandidatePairsSet,
    243              std::vector<Value *> &PairableInsts,
    244              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
    245              DenseMap<VPPair, unsigned> &PairConnectionTypes);
    246 
    247     void buildDepMap(BasicBlock &BB,
    248              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
    249              std::vector<Value *> &PairableInsts,
    250              DenseSet<ValuePair> &PairableInstUsers);
    251 
    252     void choosePairs(DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
    253              DenseSet<ValuePair> &CandidatePairsSet,
    254              DenseMap<ValuePair, int> &CandidatePairCostSavings,
    255              std::vector<Value *> &PairableInsts,
    256              DenseSet<ValuePair> &FixedOrderPairs,
    257              DenseMap<VPPair, unsigned> &PairConnectionTypes,
    258              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
    259              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
    260              DenseSet<ValuePair> &PairableInstUsers,
    261              DenseMap<Value *, Value *>& ChosenPairs);
    262 
    263     void fuseChosenPairs(BasicBlock &BB,
    264              std::vector<Value *> &PairableInsts,
    265              DenseMap<Value *, Value *>& ChosenPairs,
    266              DenseSet<ValuePair> &FixedOrderPairs,
    267              DenseMap<VPPair, unsigned> &PairConnectionTypes,
    268              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
    269              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps);
    270 
    271 
    272     bool isInstVectorizable(Instruction *I, bool &IsSimpleLoadStore);
    273 
    274     bool areInstsCompatible(Instruction *I, Instruction *J,
    275                        bool IsSimpleLoadStore, bool NonPow2Len,
    276                        int &CostSavings, int &FixedOrder);
    277 
    278     bool trackUsesOfI(DenseSet<Value *> &Users,
    279                       AliasSetTracker &WriteSet, Instruction *I,
    280                       Instruction *J, bool UpdateUsers = true,
    281                       DenseSet<ValuePair> *LoadMoveSetPairs = 0);
    282 
    283   void computePairsConnectedTo(
    284              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
    285              DenseSet<ValuePair> &CandidatePairsSet,
    286              std::vector<Value *> &PairableInsts,
    287              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
    288              DenseMap<VPPair, unsigned> &PairConnectionTypes,
    289              ValuePair P);
    290 
    291     bool pairsConflict(ValuePair P, ValuePair Q,
    292              DenseSet<ValuePair> &PairableInstUsers,
    293              DenseMap<ValuePair, std::vector<ValuePair> >
    294                *PairableInstUserMap = 0,
    295              DenseSet<VPPair> *PairableInstUserPairSet = 0);
    296 
    297     bool pairWillFormCycle(ValuePair P,
    298              DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUsers,
    299              DenseSet<ValuePair> &CurrentPairs);
    300 
    301     void pruneDAGFor(
    302              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
    303              std::vector<Value *> &PairableInsts,
    304              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
    305              DenseSet<ValuePair> &PairableInstUsers,
    306              DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
    307              DenseSet<VPPair> &PairableInstUserPairSet,
    308              DenseMap<Value *, Value *> &ChosenPairs,
    309              DenseMap<ValuePair, size_t> &DAG,
    310              DenseSet<ValuePair> &PrunedDAG, ValuePair J,
    311              bool UseCycleCheck);
    312 
    313     void buildInitialDAGFor(
    314              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
    315              DenseSet<ValuePair> &CandidatePairsSet,
    316              std::vector<Value *> &PairableInsts,
    317              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
    318              DenseSet<ValuePair> &PairableInstUsers,
    319              DenseMap<Value *, Value *> &ChosenPairs,
    320              DenseMap<ValuePair, size_t> &DAG, ValuePair J);
    321 
    322     void findBestDAGFor(
    323              DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
    324              DenseSet<ValuePair> &CandidatePairsSet,
    325              DenseMap<ValuePair, int> &CandidatePairCostSavings,
    326              std::vector<Value *> &PairableInsts,
    327              DenseSet<ValuePair> &FixedOrderPairs,
    328              DenseMap<VPPair, unsigned> &PairConnectionTypes,
    329              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
    330              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
    331              DenseSet<ValuePair> &PairableInstUsers,
    332              DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
    333              DenseSet<VPPair> &PairableInstUserPairSet,
    334              DenseMap<Value *, Value *> &ChosenPairs,
    335              DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
    336              int &BestEffSize, Value *II, std::vector<Value *>&JJ,
    337              bool UseCycleCheck);
    338 
    339     Value *getReplacementPointerInput(LLVMContext& Context, Instruction *I,
    340                      Instruction *J, unsigned o);
    341 
    342     void fillNewShuffleMask(LLVMContext& Context, Instruction *J,
    343                      unsigned MaskOffset, unsigned NumInElem,
    344                      unsigned NumInElem1, unsigned IdxOffset,
    345                      std::vector<Constant*> &Mask);
    346 
    347     Value *getReplacementShuffleMask(LLVMContext& Context, Instruction *I,
    348                      Instruction *J);
    349 
    350     bool expandIEChain(LLVMContext& Context, Instruction *I, Instruction *J,
    351                        unsigned o, Value *&LOp, unsigned numElemL,
    352                        Type *ArgTypeL, Type *ArgTypeR, bool IBeforeJ,
    353                        unsigned IdxOff = 0);
    354 
    355     Value *getReplacementInput(LLVMContext& Context, Instruction *I,
    356                      Instruction *J, unsigned o, bool IBeforeJ);
    357 
    358     void getReplacementInputsForPair(LLVMContext& Context, Instruction *I,
    359                      Instruction *J, SmallVectorImpl<Value *> &ReplacedOperands,
    360                      bool IBeforeJ);
    361 
    362     void replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
    363                      Instruction *J, Instruction *K,
    364                      Instruction *&InsertionPt, Instruction *&K1,
    365                      Instruction *&K2);
    366 
    367     void collectPairLoadMoveSet(BasicBlock &BB,
    368                      DenseMap<Value *, Value *> &ChosenPairs,
    369                      DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
    370                      DenseSet<ValuePair> &LoadMoveSetPairs,
    371                      Instruction *I);
    372 
    373     void collectLoadMoveSet(BasicBlock &BB,
    374                      std::vector<Value *> &PairableInsts,
    375                      DenseMap<Value *, Value *> &ChosenPairs,
    376                      DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
    377                      DenseSet<ValuePair> &LoadMoveSetPairs);
    378 
    379     bool canMoveUsesOfIAfterJ(BasicBlock &BB,
    380                      DenseSet<ValuePair> &LoadMoveSetPairs,
    381                      Instruction *I, Instruction *J);
    382 
    383     void moveUsesOfIAfterJ(BasicBlock &BB,
    384                      DenseSet<ValuePair> &LoadMoveSetPairs,
    385                      Instruction *&InsertionPt,
    386                      Instruction *I, Instruction *J);
    387 
    388     void combineMetadata(Instruction *K, const Instruction *J);
    389 
    390     bool vectorizeBB(BasicBlock &BB) {
    391       if (!DT->isReachableFromEntry(&BB)) {
    392         DEBUG(dbgs() << "BBV: skipping unreachable " << BB.getName() <<
    393               " in " << BB.getParent()->getName() << "\n");
    394         return false;
    395       }
    396 
    397       DEBUG(if (TTI) dbgs() << "BBV: using target information\n");
    398 
    399       bool changed = false;
    400       // Iterate a sufficient number of times to merge types of size 1 bit,
    401       // then 2 bits, then 4, etc. up to half of the target vector width of the
    402       // target vector register.
    403       unsigned n = 1;
    404       for (unsigned v = 2;
    405            (TTI || v <= Config.VectorBits) &&
    406            (!Config.MaxIter || n <= Config.MaxIter);
    407            v *= 2, ++n) {
    408         DEBUG(dbgs() << "BBV: fusing loop #" << n <<
    409               " for " << BB.getName() << " in " <<
    410               BB.getParent()->getName() << "...\n");
    411         if (vectorizePairs(BB))
    412           changed = true;
    413         else
    414           break;
    415       }
    416 
    417       if (changed && !Pow2LenOnly) {
    418         ++n;
    419         for (; !Config.MaxIter || n <= Config.MaxIter; ++n) {
    420           DEBUG(dbgs() << "BBV: fusing for non-2^n-length vectors loop #: " <<
    421                 n << " for " << BB.getName() << " in " <<
    422                 BB.getParent()->getName() << "...\n");
    423           if (!vectorizePairs(BB, true)) break;
    424         }
    425       }
    426 
    427       DEBUG(dbgs() << "BBV: done!\n");
    428       return changed;
    429     }
    430 
    431     virtual bool runOnBasicBlock(BasicBlock &BB) {
    432       AA = &getAnalysis<AliasAnalysis>();
    433       DT = &getAnalysis<DominatorTree>();
    434       SE = &getAnalysis<ScalarEvolution>();
    435       TD = getAnalysisIfAvailable<DataLayout>();
    436       TTI = IgnoreTargetInfo ? 0 : &getAnalysis<TargetTransformInfo>();
    437 
    438       return vectorizeBB(BB);
    439     }
    440 
    441     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
    442       BasicBlockPass::getAnalysisUsage(AU);
    443       AU.addRequired<AliasAnalysis>();
    444       AU.addRequired<DominatorTree>();
    445       AU.addRequired<ScalarEvolution>();
    446       AU.addRequired<TargetTransformInfo>();
    447       AU.addPreserved<AliasAnalysis>();
    448       AU.addPreserved<DominatorTree>();
    449       AU.addPreserved<ScalarEvolution>();
    450       AU.setPreservesCFG();
    451     }
    452 
    453     static inline VectorType *getVecTypeForPair(Type *ElemTy, Type *Elem2Ty) {
    454       assert(ElemTy->getScalarType() == Elem2Ty->getScalarType() &&
    455              "Cannot form vector from incompatible scalar types");
    456       Type *STy = ElemTy->getScalarType();
    457 
    458       unsigned numElem;
    459       if (VectorType *VTy = dyn_cast<VectorType>(ElemTy)) {
    460         numElem = VTy->getNumElements();
    461       } else {
    462         numElem = 1;
    463       }
    464 
    465       if (VectorType *VTy = dyn_cast<VectorType>(Elem2Ty)) {
    466         numElem += VTy->getNumElements();
    467       } else {
    468         numElem += 1;
    469       }
    470 
    471       return VectorType::get(STy, numElem);
    472     }
    473 
    474     static inline void getInstructionTypes(Instruction *I,
    475                                            Type *&T1, Type *&T2) {
    476       if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
    477         // For stores, it is the value type, not the pointer type that matters
    478         // because the value is what will come from a vector register.
    479 
    480         Value *IVal = SI->getValueOperand();
    481         T1 = IVal->getType();
    482       } else {
    483         T1 = I->getType();
    484       }
    485 
    486       if (CastInst *CI = dyn_cast<CastInst>(I))
    487         T2 = CI->getSrcTy();
    488       else
    489         T2 = T1;
    490 
    491       if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
    492         T2 = SI->getCondition()->getType();
    493       } else if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(I)) {
    494         T2 = SI->getOperand(0)->getType();
    495       } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
    496         T2 = CI->getOperand(0)->getType();
    497       }
    498     }
    499 
    500     // Returns the weight associated with the provided value. A chain of
    501     // candidate pairs has a length given by the sum of the weights of its
    502     // members (one weight per pair; the weight of each member of the pair
    503     // is assumed to be the same). This length is then compared to the
    504     // chain-length threshold to determine if a given chain is significant
    505     // enough to be vectorized. The length is also used in comparing
    506     // candidate chains where longer chains are considered to be better.
    507     // Note: when this function returns 0, the resulting instructions are
    508     // not actually fused.
    509     inline size_t getDepthFactor(Value *V) {
    510       // InsertElement and ExtractElement have a depth factor of zero. This is
    511       // for two reasons: First, they cannot be usefully fused. Second, because
    512       // the pass generates a lot of these, they can confuse the simple metric
    513       // used to compare the dags in the next iteration. Thus, giving them a
    514       // weight of zero allows the pass to essentially ignore them in
    515       // subsequent iterations when looking for vectorization opportunities
    516       // while still tracking dependency chains that flow through those
    517       // instructions.
    518       if (isa<InsertElementInst>(V) || isa<ExtractElementInst>(V))
    519         return 0;
    520 
    521       // Give a load or store half of the required depth so that load/store
    522       // pairs will vectorize.
    523       if (!Config.NoMemOpBoost && (isa<LoadInst>(V) || isa<StoreInst>(V)))
    524         return Config.ReqChainDepth/2;
    525 
    526       return 1;
    527     }
    528 
    529     // Returns the cost of the provided instruction using TTI.
    530     // This does not handle loads and stores.
    531     unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) {
    532       switch (Opcode) {
    533       default: break;
    534       case Instruction::GetElementPtr:
    535         // We mark this instruction as zero-cost because scalar GEPs are usually
    536         // lowered to the intruction addressing mode. At the moment we don't
    537         // generate vector GEPs.
    538         return 0;
    539       case Instruction::Br:
    540         return TTI->getCFInstrCost(Opcode);
    541       case Instruction::PHI:
    542         return 0;
    543       case Instruction::Add:
    544       case Instruction::FAdd:
    545       case Instruction::Sub:
    546       case Instruction::FSub:
    547       case Instruction::Mul:
    548       case Instruction::FMul:
    549       case Instruction::UDiv:
    550       case Instruction::SDiv:
    551       case Instruction::FDiv:
    552       case Instruction::URem:
    553       case Instruction::SRem:
    554       case Instruction::FRem:
    555       case Instruction::Shl:
    556       case Instruction::LShr:
    557       case Instruction::AShr:
    558       case Instruction::And:
    559       case Instruction::Or:
    560       case Instruction::Xor:
    561         return TTI->getArithmeticInstrCost(Opcode, T1);
    562       case Instruction::Select:
    563       case Instruction::ICmp:
    564       case Instruction::FCmp:
    565         return TTI->getCmpSelInstrCost(Opcode, T1, T2);
    566       case Instruction::ZExt:
    567       case Instruction::SExt:
    568       case Instruction::FPToUI:
    569       case Instruction::FPToSI:
    570       case Instruction::FPExt:
    571       case Instruction::PtrToInt:
    572       case Instruction::IntToPtr:
    573       case Instruction::SIToFP:
    574       case Instruction::UIToFP:
    575       case Instruction::Trunc:
    576       case Instruction::FPTrunc:
    577       case Instruction::BitCast:
    578       case Instruction::ShuffleVector:
    579         return TTI->getCastInstrCost(Opcode, T1, T2);
    580       }
    581 
    582       return 1;
    583     }
    584 
    585     // This determines the relative offset of two loads or stores, returning
    586     // true if the offset could be determined to be some constant value.
    587     // For example, if OffsetInElmts == 1, then J accesses the memory directly
    588     // after I; if OffsetInElmts == -1 then I accesses the memory
    589     // directly after J.
    590     bool getPairPtrInfo(Instruction *I, Instruction *J,
    591         Value *&IPtr, Value *&JPtr, unsigned &IAlignment, unsigned &JAlignment,
    592         unsigned &IAddressSpace, unsigned &JAddressSpace,
    593         int64_t &OffsetInElmts, bool ComputeOffset = true) {
    594       OffsetInElmts = 0;
    595       if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
    596         LoadInst *LJ = cast<LoadInst>(J);
    597         IPtr = LI->getPointerOperand();
    598         JPtr = LJ->getPointerOperand();
    599         IAlignment = LI->getAlignment();
    600         JAlignment = LJ->getAlignment();
    601         IAddressSpace = LI->getPointerAddressSpace();
    602         JAddressSpace = LJ->getPointerAddressSpace();
    603       } else {
    604         StoreInst *SI = cast<StoreInst>(I), *SJ = cast<StoreInst>(J);
    605         IPtr = SI->getPointerOperand();
    606         JPtr = SJ->getPointerOperand();
    607         IAlignment = SI->getAlignment();
    608         JAlignment = SJ->getAlignment();
    609         IAddressSpace = SI->getPointerAddressSpace();
    610         JAddressSpace = SJ->getPointerAddressSpace();
    611       }
    612 
    613       if (!ComputeOffset)
    614         return true;
    615 
    616       const SCEV *IPtrSCEV = SE->getSCEV(IPtr);
    617       const SCEV *JPtrSCEV = SE->getSCEV(JPtr);
    618 
    619       // If this is a trivial offset, then we'll get something like
    620       // 1*sizeof(type). With target data, which we need anyway, this will get
    621       // constant folded into a number.
    622       const SCEV *OffsetSCEV = SE->getMinusSCEV(JPtrSCEV, IPtrSCEV);
    623       if (const SCEVConstant *ConstOffSCEV =
    624             dyn_cast<SCEVConstant>(OffsetSCEV)) {
    625         ConstantInt *IntOff = ConstOffSCEV->getValue();
    626         int64_t Offset = IntOff->getSExtValue();
    627 
    628         Type *VTy = cast<PointerType>(IPtr->getType())->getElementType();
    629         int64_t VTyTSS = (int64_t) TD->getTypeStoreSize(VTy);
    630 
    631         Type *VTy2 = cast<PointerType>(JPtr->getType())->getElementType();
    632         if (VTy != VTy2 && Offset < 0) {
    633           int64_t VTy2TSS = (int64_t) TD->getTypeStoreSize(VTy2);
    634           OffsetInElmts = Offset/VTy2TSS;
    635           return (abs64(Offset) % VTy2TSS) == 0;
    636         }
    637 
    638         OffsetInElmts = Offset/VTyTSS;
    639         return (abs64(Offset) % VTyTSS) == 0;
    640       }
    641 
    642       return false;
    643     }
    644 
    645     // Returns true if the provided CallInst represents an intrinsic that can
    646     // be vectorized.
    647     bool isVectorizableIntrinsic(CallInst* I) {
    648       Function *F = I->getCalledFunction();
    649       if (!F) return false;
    650 
    651       Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
    652       if (!IID) return false;
    653 
    654       switch(IID) {
    655       default:
    656         return false;
    657       case Intrinsic::sqrt:
    658       case Intrinsic::powi:
    659       case Intrinsic::sin:
    660       case Intrinsic::cos:
    661       case Intrinsic::log:
    662       case Intrinsic::log2:
    663       case Intrinsic::log10:
    664       case Intrinsic::exp:
    665       case Intrinsic::exp2:
    666       case Intrinsic::pow:
    667         return Config.VectorizeMath;
    668       case Intrinsic::fma:
    669       case Intrinsic::fmuladd:
    670         return Config.VectorizeFMA;
    671       }
    672     }
    673 
    674     bool isPureIEChain(InsertElementInst *IE) {
    675       InsertElementInst *IENext = IE;
    676       do {
    677         if (!isa<UndefValue>(IENext->getOperand(0)) &&
    678             !isa<InsertElementInst>(IENext->getOperand(0))) {
    679           return false;
    680         }
    681       } while ((IENext =
    682                  dyn_cast<InsertElementInst>(IENext->getOperand(0))));
    683 
    684       return true;
    685     }
    686   };
    687 
    688   // This function implements one vectorization iteration on the provided
    689   // basic block. It returns true if the block is changed.
    690   bool BBVectorize::vectorizePairs(BasicBlock &BB, bool NonPow2Len) {
    691     bool ShouldContinue;
    692     BasicBlock::iterator Start = BB.getFirstInsertionPt();
    693 
    694     std::vector<Value *> AllPairableInsts;
    695     DenseMap<Value *, Value *> AllChosenPairs;
    696     DenseSet<ValuePair> AllFixedOrderPairs;
    697     DenseMap<VPPair, unsigned> AllPairConnectionTypes;
    698     DenseMap<ValuePair, std::vector<ValuePair> > AllConnectedPairs,
    699                                                  AllConnectedPairDeps;
    700 
    701     do {
    702       std::vector<Value *> PairableInsts;
    703       DenseMap<Value *, std::vector<Value *> > CandidatePairs;
    704       DenseSet<ValuePair> FixedOrderPairs;
    705       DenseMap<ValuePair, int> CandidatePairCostSavings;
    706       ShouldContinue = getCandidatePairs(BB, Start, CandidatePairs,
    707                                          FixedOrderPairs,
    708                                          CandidatePairCostSavings,
    709                                          PairableInsts, NonPow2Len);
    710       if (PairableInsts.empty()) continue;
    711 
    712       // Build the candidate pair set for faster lookups.
    713       DenseSet<ValuePair> CandidatePairsSet;
    714       for (DenseMap<Value *, std::vector<Value *> >::iterator I =
    715            CandidatePairs.begin(), E = CandidatePairs.end(); I != E; ++I)
    716         for (std::vector<Value *>::iterator J = I->second.begin(),
    717              JE = I->second.end(); J != JE; ++J)
    718           CandidatePairsSet.insert(ValuePair(I->first, *J));
    719 
    720       // Now we have a map of all of the pairable instructions and we need to
    721       // select the best possible pairing. A good pairing is one such that the
    722       // users of the pair are also paired. This defines a (directed) forest
    723       // over the pairs such that two pairs are connected iff the second pair
    724       // uses the first.
    725 
    726       // Note that it only matters that both members of the second pair use some
    727       // element of the first pair (to allow for splatting).
    728 
    729       DenseMap<ValuePair, std::vector<ValuePair> > ConnectedPairs,
    730                                                    ConnectedPairDeps;
    731       DenseMap<VPPair, unsigned> PairConnectionTypes;
    732       computeConnectedPairs(CandidatePairs, CandidatePairsSet,
    733                             PairableInsts, ConnectedPairs, PairConnectionTypes);
    734       if (ConnectedPairs.empty()) continue;
    735 
    736       for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
    737            I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
    738            I != IE; ++I)
    739         for (std::vector<ValuePair>::iterator J = I->second.begin(),
    740              JE = I->second.end(); J != JE; ++J)
    741           ConnectedPairDeps[*J].push_back(I->first);
    742 
    743       // Build the pairable-instruction dependency map
    744       DenseSet<ValuePair> PairableInstUsers;
    745       buildDepMap(BB, CandidatePairs, PairableInsts, PairableInstUsers);
    746 
    747       // There is now a graph of the connected pairs. For each variable, pick
    748       // the pairing with the largest dag meeting the depth requirement on at
    749       // least one branch. Then select all pairings that are part of that dag
    750       // and remove them from the list of available pairings and pairable
    751       // variables.
    752 
    753       DenseMap<Value *, Value *> ChosenPairs;
    754       choosePairs(CandidatePairs, CandidatePairsSet,
    755         CandidatePairCostSavings,
    756         PairableInsts, FixedOrderPairs, PairConnectionTypes,
    757         ConnectedPairs, ConnectedPairDeps,
    758         PairableInstUsers, ChosenPairs);
    759 
    760       if (ChosenPairs.empty()) continue;
    761       AllPairableInsts.insert(AllPairableInsts.end(), PairableInsts.begin(),
    762                               PairableInsts.end());
    763       AllChosenPairs.insert(ChosenPairs.begin(), ChosenPairs.end());
    764 
    765       // Only for the chosen pairs, propagate information on fixed-order pairs,
    766       // pair connections, and their types to the data structures used by the
    767       // pair fusion procedures.
    768       for (DenseMap<Value *, Value *>::iterator I = ChosenPairs.begin(),
    769            IE = ChosenPairs.end(); I != IE; ++I) {
    770         if (FixedOrderPairs.count(*I))
    771           AllFixedOrderPairs.insert(*I);
    772         else if (FixedOrderPairs.count(ValuePair(I->second, I->first)))
    773           AllFixedOrderPairs.insert(ValuePair(I->second, I->first));
    774 
    775         for (DenseMap<Value *, Value *>::iterator J = ChosenPairs.begin();
    776              J != IE; ++J) {
    777           DenseMap<VPPair, unsigned>::iterator K =
    778             PairConnectionTypes.find(VPPair(*I, *J));
    779           if (K != PairConnectionTypes.end()) {
    780             AllPairConnectionTypes.insert(*K);
    781           } else {
    782             K = PairConnectionTypes.find(VPPair(*J, *I));
    783             if (K != PairConnectionTypes.end())
    784               AllPairConnectionTypes.insert(*K);
    785           }
    786         }
    787       }
    788 
    789       for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator
    790            I = ConnectedPairs.begin(), IE = ConnectedPairs.end();
    791            I != IE; ++I)
    792         for (std::vector<ValuePair>::iterator J = I->second.begin(),
    793           JE = I->second.end(); J != JE; ++J)
    794           if (AllPairConnectionTypes.count(VPPair(I->first, *J))) {
    795             AllConnectedPairs[I->first].push_back(*J);
    796             AllConnectedPairDeps[*J].push_back(I->first);
    797           }
    798     } while (ShouldContinue);
    799 
    800     if (AllChosenPairs.empty()) return false;
    801     NumFusedOps += AllChosenPairs.size();
    802 
    803     // A set of pairs has now been selected. It is now necessary to replace the
    804     // paired instructions with vector instructions. For this procedure each
    805     // operand must be replaced with a vector operand. This vector is formed
    806     // by using build_vector on the old operands. The replaced values are then
    807     // replaced with a vector_extract on the result.  Subsequent optimization
    808     // passes should coalesce the build/extract combinations.
    809 
    810     fuseChosenPairs(BB, AllPairableInsts, AllChosenPairs, AllFixedOrderPairs,
    811                     AllPairConnectionTypes,
    812                     AllConnectedPairs, AllConnectedPairDeps);
    813 
    814     // It is important to cleanup here so that future iterations of this
    815     // function have less work to do.
    816     (void) SimplifyInstructionsInBlock(&BB, TD, AA->getTargetLibraryInfo());
    817     return true;
    818   }
    819 
    820   // This function returns true if the provided instruction is capable of being
    821   // fused into a vector instruction. This determination is based only on the
    822   // type and other attributes of the instruction.
    823   bool BBVectorize::isInstVectorizable(Instruction *I,
    824                                          bool &IsSimpleLoadStore) {
    825     IsSimpleLoadStore = false;
    826 
    827     if (CallInst *C = dyn_cast<CallInst>(I)) {
    828       if (!isVectorizableIntrinsic(C))
    829         return false;
    830     } else if (LoadInst *L = dyn_cast<LoadInst>(I)) {
    831       // Vectorize simple loads if possbile:
    832       IsSimpleLoadStore = L->isSimple();
    833       if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
    834         return false;
    835     } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {
    836       // Vectorize simple stores if possbile:
    837       IsSimpleLoadStore = S->isSimple();
    838       if (!IsSimpleLoadStore || !Config.VectorizeMemOps)
    839         return false;
    840     } else if (CastInst *C = dyn_cast<CastInst>(I)) {
    841       // We can vectorize casts, but not casts of pointer types, etc.
    842       if (!Config.VectorizeCasts)
    843         return false;
    844 
    845       Type *SrcTy = C->getSrcTy();
    846       if (!SrcTy->isSingleValueType())
    847         return false;
    848 
    849       Type *DestTy = C->getDestTy();
    850       if (!DestTy->isSingleValueType())
    851         return false;
    852     } else if (isa<SelectInst>(I)) {
    853       if (!Config.VectorizeSelect)
    854         return false;
    855     } else if (isa<CmpInst>(I)) {
    856       if (!Config.VectorizeCmp)
    857         return false;
    858     } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(I)) {
    859       if (!Config.VectorizeGEP)
    860         return false;
    861 
    862       // Currently, vector GEPs exist only with one index.
    863       if (G->getNumIndices() != 1)
    864         return false;
    865     } else if (!(I->isBinaryOp() || isa<ShuffleVectorInst>(I) ||
    866         isa<ExtractElementInst>(I) || isa<InsertElementInst>(I))) {
    867       return false;
    868     }
    869 
    870     // We can't vectorize memory operations without target data
    871     if (TD == 0 && IsSimpleLoadStore)
    872       return false;
    873 
    874     Type *T1, *T2;
    875     getInstructionTypes(I, T1, T2);
    876 
    877     // Not every type can be vectorized...
    878     if (!(VectorType::isValidElementType(T1) || T1->isVectorTy()) ||
    879         !(VectorType::isValidElementType(T2) || T2->isVectorTy()))
    880       return false;
    881 
    882     if (T1->getScalarSizeInBits() == 1) {
    883       if (!Config.VectorizeBools)
    884         return false;
    885     } else {
    886       if (!Config.VectorizeInts && T1->isIntOrIntVectorTy())
    887         return false;
    888     }
    889 
    890     if (T2->getScalarSizeInBits() == 1) {
    891       if (!Config.VectorizeBools)
    892         return false;
    893     } else {
    894       if (!Config.VectorizeInts && T2->isIntOrIntVectorTy())
    895         return false;
    896     }
    897 
    898     if (!Config.VectorizeFloats
    899         && (T1->isFPOrFPVectorTy() || T2->isFPOrFPVectorTy()))
    900       return false;
    901 
    902     // Don't vectorize target-specific types.
    903     if (T1->isX86_FP80Ty() || T1->isPPC_FP128Ty() || T1->isX86_MMXTy())
    904       return false;
    905     if (T2->isX86_FP80Ty() || T2->isPPC_FP128Ty() || T2->isX86_MMXTy())
    906       return false;
    907 
    908     if ((!Config.VectorizePointers || TD == 0) &&
    909         (T1->getScalarType()->isPointerTy() ||
    910          T2->getScalarType()->isPointerTy()))
    911       return false;
    912 
    913     if (!TTI && (T1->getPrimitiveSizeInBits() >= Config.VectorBits ||
    914                  T2->getPrimitiveSizeInBits() >= Config.VectorBits))
    915       return false;
    916 
    917     return true;
    918   }
    919 
    920   // This function returns true if the two provided instructions are compatible
    921   // (meaning that they can be fused into a vector instruction). This assumes
    922   // that I has already been determined to be vectorizable and that J is not
    923   // in the use dag of I.
    924   bool BBVectorize::areInstsCompatible(Instruction *I, Instruction *J,
    925                        bool IsSimpleLoadStore, bool NonPow2Len,
    926                        int &CostSavings, int &FixedOrder) {
    927     DEBUG(if (DebugInstructionExamination) dbgs() << "BBV: looking at " << *I <<
    928                      " <-> " << *J << "\n");
    929 
    930     CostSavings = 0;
    931     FixedOrder = 0;
    932 
    933     // Loads and stores can be merged if they have different alignments,
    934     // but are otherwise the same.
    935     if (!J->isSameOperationAs(I, Instruction::CompareIgnoringAlignment |
    936                       (NonPow2Len ? Instruction::CompareUsingScalarTypes : 0)))
    937       return false;
    938 
    939     Type *IT1, *IT2, *JT1, *JT2;
    940     getInstructionTypes(I, IT1, IT2);
    941     getInstructionTypes(J, JT1, JT2);
    942     unsigned MaxTypeBits = std::max(
    943       IT1->getPrimitiveSizeInBits() + JT1->getPrimitiveSizeInBits(),
    944       IT2->getPrimitiveSizeInBits() + JT2->getPrimitiveSizeInBits());
    945     if (!TTI && MaxTypeBits > Config.VectorBits)
    946       return false;
    947 
    948     // FIXME: handle addsub-type operations!
    949 
    950     if (IsSimpleLoadStore) {
    951       Value *IPtr, *JPtr;
    952       unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
    953       int64_t OffsetInElmts = 0;
    954       if (getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
    955             IAddressSpace, JAddressSpace,
    956             OffsetInElmts) && abs64(OffsetInElmts) == 1) {
    957         FixedOrder = (int) OffsetInElmts;
    958         unsigned BottomAlignment = IAlignment;
    959         if (OffsetInElmts < 0) BottomAlignment = JAlignment;
    960 
    961         Type *aTypeI = isa<StoreInst>(I) ?
    962           cast<StoreInst>(I)->getValueOperand()->getType() : I->getType();
    963         Type *aTypeJ = isa<StoreInst>(J) ?
    964           cast<StoreInst>(J)->getValueOperand()->getType() : J->getType();
    965         Type *VType = getVecTypeForPair(aTypeI, aTypeJ);
    966 
    967         if (Config.AlignedOnly) {
    968           // An aligned load or store is possible only if the instruction
    969           // with the lower offset has an alignment suitable for the
    970           // vector type.
    971 
    972           unsigned VecAlignment = TD->getPrefTypeAlignment(VType);
    973           if (BottomAlignment < VecAlignment)
    974             return false;
    975         }
    976 
    977         if (TTI) {
    978           unsigned ICost = TTI->getMemoryOpCost(I->getOpcode(), aTypeI,
    979                                                 IAlignment, IAddressSpace);
    980           unsigned JCost = TTI->getMemoryOpCost(J->getOpcode(), aTypeJ,
    981                                                 JAlignment, JAddressSpace);
    982           unsigned VCost = TTI->getMemoryOpCost(I->getOpcode(), VType,
    983                                                 BottomAlignment,
    984                                                 IAddressSpace);
    985 
    986           ICost += TTI->getAddressComputationCost(aTypeI);
    987           JCost += TTI->getAddressComputationCost(aTypeJ);
    988           VCost += TTI->getAddressComputationCost(VType);
    989 
    990           if (VCost > ICost + JCost)
    991             return false;
    992 
    993           // We don't want to fuse to a type that will be split, even
    994           // if the two input types will also be split and there is no other
    995           // associated cost.
    996           unsigned VParts = TTI->getNumberOfParts(VType);
    997           if (VParts > 1)
    998             return false;
    999           else if (!VParts && VCost == ICost + JCost)
   1000             return false;
   1001 
   1002           CostSavings = ICost + JCost - VCost;
   1003         }
   1004       } else {
   1005         return false;
   1006       }
   1007     } else if (TTI) {
   1008       unsigned ICost = getInstrCost(I->getOpcode(), IT1, IT2);
   1009       unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2);
   1010       Type *VT1 = getVecTypeForPair(IT1, JT1),
   1011            *VT2 = getVecTypeForPair(IT2, JT2);
   1012 
   1013       // Note that this procedure is incorrect for insert and extract element
   1014       // instructions (because combining these often results in a shuffle),
   1015       // but this cost is ignored (because insert and extract element
   1016       // instructions are assigned a zero depth factor and are not really
   1017       // fused in general).
   1018       unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2);
   1019 
   1020       if (VCost > ICost + JCost)
   1021         return false;
   1022 
   1023       // We don't want to fuse to a type that will be split, even
   1024       // if the two input types will also be split and there is no other
   1025       // associated cost.
   1026       unsigned VParts1 = TTI->getNumberOfParts(VT1),
   1027                VParts2 = TTI->getNumberOfParts(VT2);
   1028       if (VParts1 > 1 || VParts2 > 1)
   1029         return false;
   1030       else if ((!VParts1 || !VParts2) && VCost == ICost + JCost)
   1031         return false;
   1032 
   1033       CostSavings = ICost + JCost - VCost;
   1034     }
   1035 
   1036     // The powi intrinsic is special because only the first argument is
   1037     // vectorized, the second arguments must be equal.
   1038     CallInst *CI = dyn_cast<CallInst>(I);
   1039     Function *FI;
   1040     if (CI && (FI = CI->getCalledFunction())) {
   1041       Intrinsic::ID IID = (Intrinsic::ID) FI->getIntrinsicID();
   1042       if (IID == Intrinsic::powi) {
   1043         Value *A1I = CI->getArgOperand(1),
   1044               *A1J = cast<CallInst>(J)->getArgOperand(1);
   1045         const SCEV *A1ISCEV = SE->getSCEV(A1I),
   1046                    *A1JSCEV = SE->getSCEV(A1J);
   1047         return (A1ISCEV == A1JSCEV);
   1048       }
   1049 
   1050       if (IID && TTI) {
   1051         SmallVector<Type*, 4> Tys;
   1052         for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
   1053           Tys.push_back(CI->getArgOperand(i)->getType());
   1054         unsigned ICost = TTI->getIntrinsicInstrCost(IID, IT1, Tys);
   1055 
   1056         Tys.clear();
   1057         CallInst *CJ = cast<CallInst>(J);
   1058         for (unsigned i = 0, ie = CJ->getNumArgOperands(); i != ie; ++i)
   1059           Tys.push_back(CJ->getArgOperand(i)->getType());
   1060         unsigned JCost = TTI->getIntrinsicInstrCost(IID, JT1, Tys);
   1061 
   1062         Tys.clear();
   1063         assert(CI->getNumArgOperands() == CJ->getNumArgOperands() &&
   1064                "Intrinsic argument counts differ");
   1065         for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
   1066           if (IID == Intrinsic::powi && i == 1)
   1067             Tys.push_back(CI->getArgOperand(i)->getType());
   1068           else
   1069             Tys.push_back(getVecTypeForPair(CI->getArgOperand(i)->getType(),
   1070                                             CJ->getArgOperand(i)->getType()));
   1071         }
   1072 
   1073         Type *RetTy = getVecTypeForPair(IT1, JT1);
   1074         unsigned VCost = TTI->getIntrinsicInstrCost(IID, RetTy, Tys);
   1075 
   1076         if (VCost > ICost + JCost)
   1077           return false;
   1078 
   1079         // We don't want to fuse to a type that will be split, even
   1080         // if the two input types will also be split and there is no other
   1081         // associated cost.
   1082         unsigned RetParts = TTI->getNumberOfParts(RetTy);
   1083         if (RetParts > 1)
   1084           return false;
   1085         else if (!RetParts && VCost == ICost + JCost)
   1086           return false;
   1087 
   1088         for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
   1089           if (!Tys[i]->isVectorTy())
   1090             continue;
   1091 
   1092           unsigned NumParts = TTI->getNumberOfParts(Tys[i]);
   1093           if (NumParts > 1)
   1094             return false;
   1095           else if (!NumParts && VCost == ICost + JCost)
   1096             return false;
   1097         }
   1098 
   1099         CostSavings = ICost + JCost - VCost;
   1100       }
   1101     }
   1102 
   1103     return true;
   1104   }
   1105 
   1106   // Figure out whether or not J uses I and update the users and write-set
   1107   // structures associated with I. Specifically, Users represents the set of
   1108   // instructions that depend on I. WriteSet represents the set
   1109   // of memory locations that are dependent on I. If UpdateUsers is true,
   1110   // and J uses I, then Users is updated to contain J and WriteSet is updated
   1111   // to contain any memory locations to which J writes. The function returns
   1112   // true if J uses I. By default, alias analysis is used to determine
   1113   // whether J reads from memory that overlaps with a location in WriteSet.
   1114   // If LoadMoveSet is not null, then it is a previously-computed map
   1115   // where the key is the memory-based user instruction and the value is
   1116   // the instruction to be compared with I. So, if LoadMoveSet is provided,
   1117   // then the alias analysis is not used. This is necessary because this
   1118   // function is called during the process of moving instructions during
   1119   // vectorization and the results of the alias analysis are not stable during
   1120   // that process.
   1121   bool BBVectorize::trackUsesOfI(DenseSet<Value *> &Users,
   1122                        AliasSetTracker &WriteSet, Instruction *I,
   1123                        Instruction *J, bool UpdateUsers,
   1124                        DenseSet<ValuePair> *LoadMoveSetPairs) {
   1125     bool UsesI = false;
   1126 
   1127     // This instruction may already be marked as a user due, for example, to
   1128     // being a member of a selected pair.
   1129     if (Users.count(J))
   1130       UsesI = true;
   1131 
   1132     if (!UsesI)
   1133       for (User::op_iterator JU = J->op_begin(), JE = J->op_end();
   1134            JU != JE; ++JU) {
   1135         Value *V = *JU;
   1136         if (I == V || Users.count(V)) {
   1137           UsesI = true;
   1138           break;
   1139         }
   1140       }
   1141     if (!UsesI && J->mayReadFromMemory()) {
   1142       if (LoadMoveSetPairs) {
   1143         UsesI = LoadMoveSetPairs->count(ValuePair(J, I));
   1144       } else {
   1145         for (AliasSetTracker::iterator W = WriteSet.begin(),
   1146              WE = WriteSet.end(); W != WE; ++W) {
   1147           if (W->aliasesUnknownInst(J, *AA)) {
   1148             UsesI = true;
   1149             break;
   1150           }
   1151         }
   1152       }
   1153     }
   1154 
   1155     if (UsesI && UpdateUsers) {
   1156       if (J->mayWriteToMemory()) WriteSet.add(J);
   1157       Users.insert(J);
   1158     }
   1159 
   1160     return UsesI;
   1161   }
   1162 
   1163   // This function iterates over all instruction pairs in the provided
   1164   // basic block and collects all candidate pairs for vectorization.
   1165   bool BBVectorize::getCandidatePairs(BasicBlock &BB,
   1166                        BasicBlock::iterator &Start,
   1167                        DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
   1168                        DenseSet<ValuePair> &FixedOrderPairs,
   1169                        DenseMap<ValuePair, int> &CandidatePairCostSavings,
   1170                        std::vector<Value *> &PairableInsts, bool NonPow2Len) {
   1171     size_t TotalPairs = 0;
   1172     BasicBlock::iterator E = BB.end();
   1173     if (Start == E) return false;
   1174 
   1175     bool ShouldContinue = false, IAfterStart = false;
   1176     for (BasicBlock::iterator I = Start++; I != E; ++I) {
   1177       if (I == Start) IAfterStart = true;
   1178 
   1179       bool IsSimpleLoadStore;
   1180       if (!isInstVectorizable(I, IsSimpleLoadStore)) continue;
   1181 
   1182       // Look for an instruction with which to pair instruction *I...
   1183       DenseSet<Value *> Users;
   1184       AliasSetTracker WriteSet(*AA);
   1185       bool JAfterStart = IAfterStart;
   1186       BasicBlock::iterator J = llvm::next(I);
   1187       for (unsigned ss = 0; J != E && ss <= Config.SearchLimit; ++J, ++ss) {
   1188         if (J == Start) JAfterStart = true;
   1189 
   1190         // Determine if J uses I, if so, exit the loop.
   1191         bool UsesI = trackUsesOfI(Users, WriteSet, I, J, !Config.FastDep);
   1192         if (Config.FastDep) {
   1193           // Note: For this heuristic to be effective, independent operations
   1194           // must tend to be intermixed. This is likely to be true from some
   1195           // kinds of grouped loop unrolling (but not the generic LLVM pass),
   1196           // but otherwise may require some kind of reordering pass.
   1197 
   1198           // When using fast dependency analysis,
   1199           // stop searching after first use:
   1200           if (UsesI) break;
   1201         } else {
   1202           if (UsesI) continue;
   1203         }
   1204 
   1205         // J does not use I, and comes before the first use of I, so it can be
   1206         // merged with I if the instructions are compatible.
   1207         int CostSavings, FixedOrder;
   1208         if (!areInstsCompatible(I, J, IsSimpleLoadStore, NonPow2Len,
   1209             CostSavings, FixedOrder)) continue;
   1210 
   1211         // J is a candidate for merging with I.
   1212         if (!PairableInsts.size() ||
   1213              PairableInsts[PairableInsts.size()-1] != I) {
   1214           PairableInsts.push_back(I);
   1215         }
   1216 
   1217         CandidatePairs[I].push_back(J);
   1218         ++TotalPairs;
   1219         if (TTI)
   1220           CandidatePairCostSavings.insert(ValuePairWithCost(ValuePair(I, J),
   1221                                                             CostSavings));
   1222 
   1223         if (FixedOrder == 1)
   1224           FixedOrderPairs.insert(ValuePair(I, J));
   1225         else if (FixedOrder == -1)
   1226           FixedOrderPairs.insert(ValuePair(J, I));
   1227 
   1228         // The next call to this function must start after the last instruction
   1229         // selected during this invocation.
   1230         if (JAfterStart) {
   1231           Start = llvm::next(J);
   1232           IAfterStart = JAfterStart = false;
   1233         }
   1234 
   1235         DEBUG(if (DebugCandidateSelection) dbgs() << "BBV: candidate pair "
   1236                      << *I << " <-> " << *J << " (cost savings: " <<
   1237                      CostSavings << ")\n");
   1238 
   1239         // If we have already found too many pairs, break here and this function
   1240         // will be called again starting after the last instruction selected
   1241         // during this invocation.
   1242         if (PairableInsts.size() >= Config.MaxInsts ||
   1243             TotalPairs >= Config.MaxPairs) {
   1244           ShouldContinue = true;
   1245           break;
   1246         }
   1247       }
   1248 
   1249       if (ShouldContinue)
   1250         break;
   1251     }
   1252 
   1253     DEBUG(dbgs() << "BBV: found " << PairableInsts.size()
   1254            << " instructions with candidate pairs\n");
   1255 
   1256     return ShouldContinue;
   1257   }
   1258 
   1259   // Finds candidate pairs connected to the pair P = <PI, PJ>. This means that
   1260   // it looks for pairs such that both members have an input which is an
   1261   // output of PI or PJ.
   1262   void BBVectorize::computePairsConnectedTo(
   1263                   DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
   1264                   DenseSet<ValuePair> &CandidatePairsSet,
   1265                   std::vector<Value *> &PairableInsts,
   1266                   DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
   1267                   DenseMap<VPPair, unsigned> &PairConnectionTypes,
   1268                   ValuePair P) {
   1269     StoreInst *SI, *SJ;
   1270 
   1271     // For each possible pairing for this variable, look at the uses of
   1272     // the first value...
   1273     for (Value::use_iterator I = P.first->use_begin(),
   1274          E = P.first->use_end(); I != E; ++I) {
   1275       if (isa<LoadInst>(*I)) {
   1276         // A pair cannot be connected to a load because the load only takes one
   1277         // operand (the address) and it is a scalar even after vectorization.
   1278         continue;
   1279       } else if ((SI = dyn_cast<StoreInst>(*I)) &&
   1280                  P.first == SI->getPointerOperand()) {
   1281         // Similarly, a pair cannot be connected to a store through its
   1282         // pointer operand.
   1283         continue;
   1284       }
   1285 
   1286       // For each use of the first variable, look for uses of the second
   1287       // variable...
   1288       for (Value::use_iterator J = P.second->use_begin(),
   1289            E2 = P.second->use_end(); J != E2; ++J) {
   1290         if ((SJ = dyn_cast<StoreInst>(*J)) &&
   1291             P.second == SJ->getPointerOperand())
   1292           continue;
   1293 
   1294         // Look for <I, J>:
   1295         if (CandidatePairsSet.count(ValuePair(*I, *J))) {
   1296           VPPair VP(P, ValuePair(*I, *J));
   1297           ConnectedPairs[VP.first].push_back(VP.second);
   1298           PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionDirect));
   1299         }
   1300 
   1301         // Look for <J, I>:
   1302         if (CandidatePairsSet.count(ValuePair(*J, *I))) {
   1303           VPPair VP(P, ValuePair(*J, *I));
   1304           ConnectedPairs[VP.first].push_back(VP.second);
   1305           PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSwap));
   1306         }
   1307       }
   1308 
   1309       if (Config.SplatBreaksChain) continue;
   1310       // Look for cases where just the first value in the pair is used by
   1311       // both members of another pair (splatting).
   1312       for (Value::use_iterator J = P.first->use_begin(); J != E; ++J) {
   1313         if ((SJ = dyn_cast<StoreInst>(*J)) &&
   1314             P.first == SJ->getPointerOperand())
   1315           continue;
   1316 
   1317         if (CandidatePairsSet.count(ValuePair(*I, *J))) {
   1318           VPPair VP(P, ValuePair(*I, *J));
   1319           ConnectedPairs[VP.first].push_back(VP.second);
   1320           PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
   1321         }
   1322       }
   1323     }
   1324 
   1325     if (Config.SplatBreaksChain) return;
   1326     // Look for cases where just the second value in the pair is used by
   1327     // both members of another pair (splatting).
   1328     for (Value::use_iterator I = P.second->use_begin(),
   1329          E = P.second->use_end(); I != E; ++I) {
   1330       if (isa<LoadInst>(*I))
   1331         continue;
   1332       else if ((SI = dyn_cast<StoreInst>(*I)) &&
   1333                P.second == SI->getPointerOperand())
   1334         continue;
   1335 
   1336       for (Value::use_iterator J = P.second->use_begin(); J != E; ++J) {
   1337         if ((SJ = dyn_cast<StoreInst>(*J)) &&
   1338             P.second == SJ->getPointerOperand())
   1339           continue;
   1340 
   1341         if (CandidatePairsSet.count(ValuePair(*I, *J))) {
   1342           VPPair VP(P, ValuePair(*I, *J));
   1343           ConnectedPairs[VP.first].push_back(VP.second);
   1344           PairConnectionTypes.insert(VPPairWithType(VP, PairConnectionSplat));
   1345         }
   1346       }
   1347     }
   1348   }
   1349 
   1350   // This function figures out which pairs are connected.  Two pairs are
   1351   // connected if some output of the first pair forms an input to both members
   1352   // of the second pair.
   1353   void BBVectorize::computeConnectedPairs(
   1354                   DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
   1355                   DenseSet<ValuePair> &CandidatePairsSet,
   1356                   std::vector<Value *> &PairableInsts,
   1357                   DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
   1358                   DenseMap<VPPair, unsigned> &PairConnectionTypes) {
   1359     for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
   1360          PE = PairableInsts.end(); PI != PE; ++PI) {
   1361       DenseMap<Value *, std::vector<Value *> >::iterator PP =
   1362         CandidatePairs.find(*PI);
   1363       if (PP == CandidatePairs.end())
   1364         continue;
   1365 
   1366       for (std::vector<Value *>::iterator P = PP->second.begin(),
   1367            E = PP->second.end(); P != E; ++P)
   1368         computePairsConnectedTo(CandidatePairs, CandidatePairsSet,
   1369                                 PairableInsts, ConnectedPairs,
   1370                                 PairConnectionTypes, ValuePair(*PI, *P));
   1371     }
   1372 
   1373     DEBUG(size_t TotalPairs = 0;
   1374           for (DenseMap<ValuePair, std::vector<ValuePair> >::iterator I =
   1375                ConnectedPairs.begin(), IE = ConnectedPairs.end(); I != IE; ++I)
   1376             TotalPairs += I->second.size();
   1377           dbgs() << "BBV: found " << TotalPairs
   1378                  << " pair connections.\n");
   1379   }
   1380 
   1381   // This function builds a set of use tuples such that <A, B> is in the set
   1382   // if B is in the use dag of A. If B is in the use dag of A, then B
   1383   // depends on the output of A.
   1384   void BBVectorize::buildDepMap(
   1385                       BasicBlock &BB,
   1386                       DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
   1387                       std::vector<Value *> &PairableInsts,
   1388                       DenseSet<ValuePair> &PairableInstUsers) {
   1389     DenseSet<Value *> IsInPair;
   1390     for (DenseMap<Value *, std::vector<Value *> >::iterator C =
   1391          CandidatePairs.begin(), E = CandidatePairs.end(); C != E; ++C) {
   1392       IsInPair.insert(C->first);
   1393       IsInPair.insert(C->second.begin(), C->second.end());
   1394     }
   1395 
   1396     // Iterate through the basic block, recording all users of each
   1397     // pairable instruction.
   1398 
   1399     BasicBlock::iterator E = BB.end(), EL =
   1400       BasicBlock::iterator(cast<Instruction>(PairableInsts.back()));
   1401     for (BasicBlock::iterator I = BB.getFirstInsertionPt(); I != E; ++I) {
   1402       if (IsInPair.find(I) == IsInPair.end()) continue;
   1403 
   1404       DenseSet<Value *> Users;
   1405       AliasSetTracker WriteSet(*AA);
   1406       for (BasicBlock::iterator J = llvm::next(I); J != E; ++J) {
   1407         (void) trackUsesOfI(Users, WriteSet, I, J);
   1408 
   1409         if (J == EL)
   1410           break;
   1411       }
   1412 
   1413       for (DenseSet<Value *>::iterator U = Users.begin(), E = Users.end();
   1414            U != E; ++U) {
   1415         if (IsInPair.find(*U) == IsInPair.end()) continue;
   1416         PairableInstUsers.insert(ValuePair(I, *U));
   1417       }
   1418 
   1419       if (I == EL)
   1420         break;
   1421     }
   1422   }
   1423 
   1424   // Returns true if an input to pair P is an output of pair Q and also an
   1425   // input of pair Q is an output of pair P. If this is the case, then these
   1426   // two pairs cannot be simultaneously fused.
   1427   bool BBVectorize::pairsConflict(ValuePair P, ValuePair Q,
   1428              DenseSet<ValuePair> &PairableInstUsers,
   1429              DenseMap<ValuePair, std::vector<ValuePair> > *PairableInstUserMap,
   1430              DenseSet<VPPair> *PairableInstUserPairSet) {
   1431     // Two pairs are in conflict if they are mutual Users of eachother.
   1432     bool QUsesP = PairableInstUsers.count(ValuePair(P.first,  Q.first))  ||
   1433                   PairableInstUsers.count(ValuePair(P.first,  Q.second)) ||
   1434                   PairableInstUsers.count(ValuePair(P.second, Q.first))  ||
   1435                   PairableInstUsers.count(ValuePair(P.second, Q.second));
   1436     bool PUsesQ = PairableInstUsers.count(ValuePair(Q.first,  P.first))  ||
   1437                   PairableInstUsers.count(ValuePair(Q.first,  P.second)) ||
   1438                   PairableInstUsers.count(ValuePair(Q.second, P.first))  ||
   1439                   PairableInstUsers.count(ValuePair(Q.second, P.second));
   1440     if (PairableInstUserMap) {
   1441       // FIXME: The expensive part of the cycle check is not so much the cycle
   1442       // check itself but this edge insertion procedure. This needs some
   1443       // profiling and probably a different data structure.
   1444       if (PUsesQ) {
   1445         if (PairableInstUserPairSet->insert(VPPair(Q, P)).second)
   1446           (*PairableInstUserMap)[Q].push_back(P);
   1447       }
   1448       if (QUsesP) {
   1449         if (PairableInstUserPairSet->insert(VPPair(P, Q)).second)
   1450           (*PairableInstUserMap)[P].push_back(Q);
   1451       }
   1452     }
   1453 
   1454     return (QUsesP && PUsesQ);
   1455   }
   1456 
   1457   // This function walks the use graph of current pairs to see if, starting
   1458   // from P, the walk returns to P.
   1459   bool BBVectorize::pairWillFormCycle(ValuePair P,
   1460              DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
   1461              DenseSet<ValuePair> &CurrentPairs) {
   1462     DEBUG(if (DebugCycleCheck)
   1463             dbgs() << "BBV: starting cycle check for : " << *P.first << " <-> "
   1464                    << *P.second << "\n");
   1465     // A lookup table of visisted pairs is kept because the PairableInstUserMap
   1466     // contains non-direct associations.
   1467     DenseSet<ValuePair> Visited;
   1468     SmallVector<ValuePair, 32> Q;
   1469     // General depth-first post-order traversal:
   1470     Q.push_back(P);
   1471     do {
   1472       ValuePair QTop = Q.pop_back_val();
   1473       Visited.insert(QTop);
   1474 
   1475       DEBUG(if (DebugCycleCheck)
   1476               dbgs() << "BBV: cycle check visiting: " << *QTop.first << " <-> "
   1477                      << *QTop.second << "\n");
   1478       DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
   1479         PairableInstUserMap.find(QTop);
   1480       if (QQ == PairableInstUserMap.end())
   1481         continue;
   1482 
   1483       for (std::vector<ValuePair>::iterator C = QQ->second.begin(),
   1484            CE = QQ->second.end(); C != CE; ++C) {
   1485         if (*C == P) {
   1486           DEBUG(dbgs()
   1487                  << "BBV: rejected to prevent non-trivial cycle formation: "
   1488                  << QTop.first << " <-> " << C->second << "\n");
   1489           return true;
   1490         }
   1491 
   1492         if (CurrentPairs.count(*C) && !Visited.count(*C))
   1493           Q.push_back(*C);
   1494       }
   1495     } while (!Q.empty());
   1496 
   1497     return false;
   1498   }
   1499 
   1500   // This function builds the initial dag of connected pairs with the
   1501   // pair J at the root.
   1502   void BBVectorize::buildInitialDAGFor(
   1503                   DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
   1504                   DenseSet<ValuePair> &CandidatePairsSet,
   1505                   std::vector<Value *> &PairableInsts,
   1506                   DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
   1507                   DenseSet<ValuePair> &PairableInstUsers,
   1508                   DenseMap<Value *, Value *> &ChosenPairs,
   1509                   DenseMap<ValuePair, size_t> &DAG, ValuePair J) {
   1510     // Each of these pairs is viewed as the root node of a DAG. The DAG
   1511     // is then walked (depth-first). As this happens, we keep track of
   1512     // the pairs that compose the DAG and the maximum depth of the DAG.
   1513     SmallVector<ValuePairWithDepth, 32> Q;
   1514     // General depth-first post-order traversal:
   1515     Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
   1516     do {
   1517       ValuePairWithDepth QTop = Q.back();
   1518 
   1519       // Push each child onto the queue:
   1520       bool MoreChildren = false;
   1521       size_t MaxChildDepth = QTop.second;
   1522       DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
   1523         ConnectedPairs.find(QTop.first);
   1524       if (QQ != ConnectedPairs.end())
   1525         for (std::vector<ValuePair>::iterator k = QQ->second.begin(),
   1526              ke = QQ->second.end(); k != ke; ++k) {
   1527           // Make sure that this child pair is still a candidate:
   1528           if (CandidatePairsSet.count(*k)) {
   1529             DenseMap<ValuePair, size_t>::iterator C = DAG.find(*k);
   1530             if (C == DAG.end()) {
   1531               size_t d = getDepthFactor(k->first);
   1532               Q.push_back(ValuePairWithDepth(*k, QTop.second+d));
   1533               MoreChildren = true;
   1534             } else {
   1535               MaxChildDepth = std::max(MaxChildDepth, C->second);
   1536             }
   1537           }
   1538         }
   1539 
   1540       if (!MoreChildren) {
   1541         // Record the current pair as part of the DAG:
   1542         DAG.insert(ValuePairWithDepth(QTop.first, MaxChildDepth));
   1543         Q.pop_back();
   1544       }
   1545     } while (!Q.empty());
   1546   }
   1547 
   1548   // Given some initial dag, prune it by removing conflicting pairs (pairs
   1549   // that cannot be simultaneously chosen for vectorization).
   1550   void BBVectorize::pruneDAGFor(
   1551               DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
   1552               std::vector<Value *> &PairableInsts,
   1553               DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
   1554               DenseSet<ValuePair> &PairableInstUsers,
   1555               DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
   1556               DenseSet<VPPair> &PairableInstUserPairSet,
   1557               DenseMap<Value *, Value *> &ChosenPairs,
   1558               DenseMap<ValuePair, size_t> &DAG,
   1559               DenseSet<ValuePair> &PrunedDAG, ValuePair J,
   1560               bool UseCycleCheck) {
   1561     SmallVector<ValuePairWithDepth, 32> Q;
   1562     // General depth-first post-order traversal:
   1563     Q.push_back(ValuePairWithDepth(J, getDepthFactor(J.first)));
   1564     do {
   1565       ValuePairWithDepth QTop = Q.pop_back_val();
   1566       PrunedDAG.insert(QTop.first);
   1567 
   1568       // Visit each child, pruning as necessary...
   1569       SmallVector<ValuePairWithDepth, 8> BestChildren;
   1570       DenseMap<ValuePair, std::vector<ValuePair> >::iterator QQ =
   1571         ConnectedPairs.find(QTop.first);
   1572       if (QQ == ConnectedPairs.end())
   1573         continue;
   1574 
   1575       for (std::vector<ValuePair>::iterator K = QQ->second.begin(),
   1576            KE = QQ->second.end(); K != KE; ++K) {
   1577         DenseMap<ValuePair, size_t>::iterator C = DAG.find(*K);
   1578         if (C == DAG.end()) continue;
   1579 
   1580         // This child is in the DAG, now we need to make sure it is the
   1581         // best of any conflicting children. There could be multiple
   1582         // conflicting children, so first, determine if we're keeping
   1583         // this child, then delete conflicting children as necessary.
   1584 
   1585         // It is also necessary to guard against pairing-induced
   1586         // dependencies. Consider instructions a .. x .. y .. b
   1587         // such that (a,b) are to be fused and (x,y) are to be fused
   1588         // but a is an input to x and b is an output from y. This
   1589         // means that y cannot be moved after b but x must be moved
   1590         // after b for (a,b) to be fused. In other words, after
   1591         // fusing (a,b) we have y .. a/b .. x where y is an input
   1592         // to a/b and x is an output to a/b: x and y can no longer
   1593         // be legally fused. To prevent this condition, we must
   1594         // make sure that a child pair added to the DAG is not
   1595         // both an input and output of an already-selected pair.
   1596 
   1597         // Pairing-induced dependencies can also form from more complicated
   1598         // cycles. The pair vs. pair conflicts are easy to check, and so
   1599         // that is done explicitly for "fast rejection", and because for
   1600         // child vs. child conflicts, we may prefer to keep the current
   1601         // pair in preference to the already-selected child.
   1602         DenseSet<ValuePair> CurrentPairs;
   1603 
   1604         bool CanAdd = true;
   1605         for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
   1606               = BestChildren.begin(), E2 = BestChildren.end();
   1607              C2 != E2; ++C2) {
   1608           if (C2->first.first == C->first.first ||
   1609               C2->first.first == C->first.second ||
   1610               C2->first.second == C->first.first ||
   1611               C2->first.second == C->first.second ||
   1612               pairsConflict(C2->first, C->first, PairableInstUsers,
   1613                             UseCycleCheck ? &PairableInstUserMap : 0,
   1614                             UseCycleCheck ? &PairableInstUserPairSet : 0)) {
   1615             if (C2->second >= C->second) {
   1616               CanAdd = false;
   1617               break;
   1618             }
   1619 
   1620             CurrentPairs.insert(C2->first);
   1621           }
   1622         }
   1623         if (!CanAdd) continue;
   1624 
   1625         // Even worse, this child could conflict with another node already
   1626         // selected for the DAG. If that is the case, ignore this child.
   1627         for (DenseSet<ValuePair>::iterator T = PrunedDAG.begin(),
   1628              E2 = PrunedDAG.end(); T != E2; ++T) {
   1629           if (T->first == C->first.first ||
   1630               T->first == C->first.second ||
   1631               T->second == C->first.first ||
   1632               T->second == C->first.second ||
   1633               pairsConflict(*T, C->first, PairableInstUsers,
   1634                             UseCycleCheck ? &PairableInstUserMap : 0,
   1635                             UseCycleCheck ? &PairableInstUserPairSet : 0)) {
   1636             CanAdd = false;
   1637             break;
   1638           }
   1639 
   1640           CurrentPairs.insert(*T);
   1641         }
   1642         if (!CanAdd) continue;
   1643 
   1644         // And check the queue too...
   1645         for (SmallVectorImpl<ValuePairWithDepth>::iterator C2 = Q.begin(),
   1646              E2 = Q.end(); C2 != E2; ++C2) {
   1647           if (C2->first.first == C->first.first ||
   1648               C2->first.first == C->first.second ||
   1649               C2->first.second == C->first.first ||
   1650               C2->first.second == C->first.second ||
   1651               pairsConflict(C2->first, C->first, PairableInstUsers,
   1652                             UseCycleCheck ? &PairableInstUserMap : 0,
   1653                             UseCycleCheck ? &PairableInstUserPairSet : 0)) {
   1654             CanAdd = false;
   1655             break;
   1656           }
   1657 
   1658           CurrentPairs.insert(C2->first);
   1659         }
   1660         if (!CanAdd) continue;
   1661 
   1662         // Last but not least, check for a conflict with any of the
   1663         // already-chosen pairs.
   1664         for (DenseMap<Value *, Value *>::iterator C2 =
   1665               ChosenPairs.begin(), E2 = ChosenPairs.end();
   1666              C2 != E2; ++C2) {
   1667           if (pairsConflict(*C2, C->first, PairableInstUsers,
   1668                             UseCycleCheck ? &PairableInstUserMap : 0,
   1669                             UseCycleCheck ? &PairableInstUserPairSet : 0)) {
   1670             CanAdd = false;
   1671             break;
   1672           }
   1673 
   1674           CurrentPairs.insert(*C2);
   1675         }
   1676         if (!CanAdd) continue;
   1677 
   1678         // To check for non-trivial cycles formed by the addition of the
   1679         // current pair we've formed a list of all relevant pairs, now use a
   1680         // graph walk to check for a cycle. We start from the current pair and
   1681         // walk the use dag to see if we again reach the current pair. If we
   1682         // do, then the current pair is rejected.
   1683 
   1684         // FIXME: It may be more efficient to use a topological-ordering
   1685         // algorithm to improve the cycle check. This should be investigated.
   1686         if (UseCycleCheck &&
   1687             pairWillFormCycle(C->first, PairableInstUserMap, CurrentPairs))
   1688           continue;
   1689 
   1690         // This child can be added, but we may have chosen it in preference
   1691         // to an already-selected child. Check for this here, and if a
   1692         // conflict is found, then remove the previously-selected child
   1693         // before adding this one in its place.
   1694         for (SmallVectorImpl<ValuePairWithDepth>::iterator C2
   1695               = BestChildren.begin(); C2 != BestChildren.end();) {
   1696           if (C2->first.first == C->first.first ||
   1697               C2->first.first == C->first.second ||
   1698               C2->first.second == C->first.first ||
   1699               C2->first.second == C->first.second ||
   1700               pairsConflict(C2->first, C->first, PairableInstUsers))
   1701             C2 = BestChildren.erase(C2);
   1702           else
   1703             ++C2;
   1704         }
   1705 
   1706         BestChildren.push_back(ValuePairWithDepth(C->first, C->second));
   1707       }
   1708 
   1709       for (SmallVectorImpl<ValuePairWithDepth>::iterator C
   1710             = BestChildren.begin(), E2 = BestChildren.end();
   1711            C != E2; ++C) {
   1712         size_t DepthF = getDepthFactor(C->first.first);
   1713         Q.push_back(ValuePairWithDepth(C->first, QTop.second+DepthF));
   1714       }
   1715     } while (!Q.empty());
   1716   }
   1717 
   1718   // This function finds the best dag of mututally-compatible connected
   1719   // pairs, given the choice of root pairs as an iterator range.
   1720   void BBVectorize::findBestDAGFor(
   1721               DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
   1722               DenseSet<ValuePair> &CandidatePairsSet,
   1723               DenseMap<ValuePair, int> &CandidatePairCostSavings,
   1724               std::vector<Value *> &PairableInsts,
   1725               DenseSet<ValuePair> &FixedOrderPairs,
   1726               DenseMap<VPPair, unsigned> &PairConnectionTypes,
   1727               DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
   1728               DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
   1729               DenseSet<ValuePair> &PairableInstUsers,
   1730               DenseMap<ValuePair, std::vector<ValuePair> > &PairableInstUserMap,
   1731               DenseSet<VPPair> &PairableInstUserPairSet,
   1732               DenseMap<Value *, Value *> &ChosenPairs,
   1733               DenseSet<ValuePair> &BestDAG, size_t &BestMaxDepth,
   1734               int &BestEffSize, Value *II, std::vector<Value *>&JJ,
   1735               bool UseCycleCheck) {
   1736     for (std::vector<Value *>::iterator J = JJ.begin(), JE = JJ.end();
   1737          J != JE; ++J) {
   1738       ValuePair IJ(II, *J);
   1739       if (!CandidatePairsSet.count(IJ))
   1740         continue;
   1741 
   1742       // Before going any further, make sure that this pair does not
   1743       // conflict with any already-selected pairs (see comment below
   1744       // near the DAG pruning for more details).
   1745       DenseSet<ValuePair> ChosenPairSet;
   1746       bool DoesConflict = false;
   1747       for (DenseMap<Value *, Value *>::iterator C = ChosenPairs.begin(),
   1748            E = ChosenPairs.end(); C != E; ++C) {
   1749         if (pairsConflict(*C, IJ, PairableInstUsers,
   1750                           UseCycleCheck ? &PairableInstUserMap : 0,
   1751                           UseCycleCheck ? &PairableInstUserPairSet : 0)) {
   1752           DoesConflict = true;
   1753           break;
   1754         }
   1755 
   1756         ChosenPairSet.insert(*C);
   1757       }
   1758       if (DoesConflict) continue;
   1759 
   1760       if (UseCycleCheck &&
   1761           pairWillFormCycle(IJ, PairableInstUserMap, ChosenPairSet))
   1762         continue;
   1763 
   1764       DenseMap<ValuePair, size_t> DAG;
   1765       buildInitialDAGFor(CandidatePairs, CandidatePairsSet,
   1766                           PairableInsts, ConnectedPairs,
   1767                           PairableInstUsers, ChosenPairs, DAG, IJ);
   1768 
   1769       // Because we'll keep the child with the largest depth, the largest
   1770       // depth is still the same in the unpruned DAG.
   1771       size_t MaxDepth = DAG.lookup(IJ);
   1772 
   1773       DEBUG(if (DebugPairSelection) dbgs() << "BBV: found DAG for pair {"
   1774                    << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
   1775                    MaxDepth << " and size " << DAG.size() << "\n");
   1776 
   1777       // At this point the DAG has been constructed, but, may contain
   1778       // contradictory children (meaning that different children of
   1779       // some dag node may be attempting to fuse the same instruction).
   1780       // So now we walk the dag again, in the case of a conflict,
   1781       // keep only the child with the largest depth. To break a tie,
   1782       // favor the first child.
   1783 
   1784       DenseSet<ValuePair> PrunedDAG;
   1785       pruneDAGFor(CandidatePairs, PairableInsts, ConnectedPairs,
   1786                    PairableInstUsers, PairableInstUserMap,
   1787                    PairableInstUserPairSet,
   1788                    ChosenPairs, DAG, PrunedDAG, IJ, UseCycleCheck);
   1789 
   1790       int EffSize = 0;
   1791       if (TTI) {
   1792         DenseSet<Value *> PrunedDAGInstrs;
   1793         for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
   1794              E = PrunedDAG.end(); S != E; ++S) {
   1795           PrunedDAGInstrs.insert(S->first);
   1796           PrunedDAGInstrs.insert(S->second);
   1797         }
   1798 
   1799         // The set of pairs that have already contributed to the total cost.
   1800         DenseSet<ValuePair> IncomingPairs;
   1801 
   1802         // If the cost model were perfect, this might not be necessary; but we
   1803         // need to make sure that we don't get stuck vectorizing our own
   1804         // shuffle chains.
   1805         bool HasNontrivialInsts = false;
   1806 
   1807         // The node weights represent the cost savings associated with
   1808         // fusing the pair of instructions.
   1809         for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
   1810              E = PrunedDAG.end(); S != E; ++S) {
   1811           if (!isa<ShuffleVectorInst>(S->first) &&
   1812               !isa<InsertElementInst>(S->first) &&
   1813               !isa<ExtractElementInst>(S->first))
   1814             HasNontrivialInsts = true;
   1815 
   1816           bool FlipOrder = false;
   1817 
   1818           if (getDepthFactor(S->first)) {
   1819             int ESContrib = CandidatePairCostSavings.find(*S)->second;
   1820             DEBUG(if (DebugPairSelection) dbgs() << "\tweight {"
   1821                    << *S->first << " <-> " << *S->second << "} = " <<
   1822                    ESContrib << "\n");
   1823             EffSize += ESContrib;
   1824           }
   1825 
   1826           // The edge weights contribute in a negative sense: they represent
   1827           // the cost of shuffles.
   1828           DenseMap<ValuePair, std::vector<ValuePair> >::iterator SS =
   1829             ConnectedPairDeps.find(*S);
   1830           if (SS != ConnectedPairDeps.end()) {
   1831             unsigned NumDepsDirect = 0, NumDepsSwap = 0;
   1832             for (std::vector<ValuePair>::iterator T = SS->second.begin(),
   1833                  TE = SS->second.end(); T != TE; ++T) {
   1834               VPPair Q(*S, *T);
   1835               if (!PrunedDAG.count(Q.second))
   1836                 continue;
   1837               DenseMap<VPPair, unsigned>::iterator R =
   1838                 PairConnectionTypes.find(VPPair(Q.second, Q.first));
   1839               assert(R != PairConnectionTypes.end() &&
   1840                      "Cannot find pair connection type");
   1841               if (R->second == PairConnectionDirect)
   1842                 ++NumDepsDirect;
   1843               else if (R->second == PairConnectionSwap)
   1844                 ++NumDepsSwap;
   1845             }
   1846 
   1847             // If there are more swaps than direct connections, then
   1848             // the pair order will be flipped during fusion. So the real
   1849             // number of swaps is the minimum number.
   1850             FlipOrder = !FixedOrderPairs.count(*S) &&
   1851               ((NumDepsSwap > NumDepsDirect) ||
   1852                 FixedOrderPairs.count(ValuePair(S->second, S->first)));
   1853 
   1854             for (std::vector<ValuePair>::iterator T = SS->second.begin(),
   1855                  TE = SS->second.end(); T != TE; ++T) {
   1856               VPPair Q(*S, *T);
   1857               if (!PrunedDAG.count(Q.second))
   1858                 continue;
   1859               DenseMap<VPPair, unsigned>::iterator R =
   1860                 PairConnectionTypes.find(VPPair(Q.second, Q.first));
   1861               assert(R != PairConnectionTypes.end() &&
   1862                      "Cannot find pair connection type");
   1863               Type *Ty1 = Q.second.first->getType(),
   1864                    *Ty2 = Q.second.second->getType();
   1865               Type *VTy = getVecTypeForPair(Ty1, Ty2);
   1866               if ((R->second == PairConnectionDirect && FlipOrder) ||
   1867                   (R->second == PairConnectionSwap && !FlipOrder)  ||
   1868                   R->second == PairConnectionSplat) {
   1869                 int ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
   1870                                                    VTy, VTy);
   1871 
   1872                 if (VTy->getVectorNumElements() == 2) {
   1873                   if (R->second == PairConnectionSplat)
   1874                     ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
   1875                       TargetTransformInfo::SK_Broadcast, VTy));
   1876                   else
   1877                     ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
   1878                       TargetTransformInfo::SK_Reverse, VTy));
   1879                 }
   1880 
   1881                 DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
   1882                   *Q.second.first << " <-> " << *Q.second.second <<
   1883                     "} -> {" <<
   1884                   *S->first << " <-> " << *S->second << "} = " <<
   1885                    ESContrib << "\n");
   1886                 EffSize -= ESContrib;
   1887               }
   1888             }
   1889           }
   1890 
   1891           // Compute the cost of outgoing edges. We assume that edges outgoing
   1892           // to shuffles, inserts or extracts can be merged, and so contribute
   1893           // no additional cost.
   1894           if (!S->first->getType()->isVoidTy()) {
   1895             Type *Ty1 = S->first->getType(),
   1896                  *Ty2 = S->second->getType();
   1897             Type *VTy = getVecTypeForPair(Ty1, Ty2);
   1898 
   1899             bool NeedsExtraction = false;
   1900             for (Value::use_iterator I = S->first->use_begin(),
   1901                  IE = S->first->use_end(); I != IE; ++I) {
   1902               if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
   1903                 // Shuffle can be folded if it has no other input
   1904                 if (isa<UndefValue>(SI->getOperand(1)))
   1905                   continue;
   1906               }
   1907               if (isa<ExtractElementInst>(*I))
   1908                 continue;
   1909               if (PrunedDAGInstrs.count(*I))
   1910                 continue;
   1911               NeedsExtraction = true;
   1912               break;
   1913             }
   1914 
   1915             if (NeedsExtraction) {
   1916               int ESContrib;
   1917               if (Ty1->isVectorTy()) {
   1918                 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
   1919                                                Ty1, VTy);
   1920                 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
   1921                   TargetTransformInfo::SK_ExtractSubvector, VTy, 0, Ty1));
   1922               } else
   1923                 ESContrib = (int) TTI->getVectorInstrCost(
   1924                                     Instruction::ExtractElement, VTy, 0);
   1925 
   1926               DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
   1927                 *S->first << "} = " << ESContrib << "\n");
   1928               EffSize -= ESContrib;
   1929             }
   1930 
   1931             NeedsExtraction = false;
   1932             for (Value::use_iterator I = S->second->use_begin(),
   1933                  IE = S->second->use_end(); I != IE; ++I) {
   1934               if (ShuffleVectorInst *SI = dyn_cast<ShuffleVectorInst>(*I)) {
   1935                 // Shuffle can be folded if it has no other input
   1936                 if (isa<UndefValue>(SI->getOperand(1)))
   1937                   continue;
   1938               }
   1939               if (isa<ExtractElementInst>(*I))
   1940                 continue;
   1941               if (PrunedDAGInstrs.count(*I))
   1942                 continue;
   1943               NeedsExtraction = true;
   1944               break;
   1945             }
   1946 
   1947             if (NeedsExtraction) {
   1948               int ESContrib;
   1949               if (Ty2->isVectorTy()) {
   1950                 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
   1951                                                Ty2, VTy);
   1952                 ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
   1953                   TargetTransformInfo::SK_ExtractSubvector, VTy,
   1954                   Ty1->isVectorTy() ? Ty1->getVectorNumElements() : 1, Ty2));
   1955               } else
   1956                 ESContrib = (int) TTI->getVectorInstrCost(
   1957                                     Instruction::ExtractElement, VTy, 1);
   1958               DEBUG(if (DebugPairSelection) dbgs() << "\tcost {" <<
   1959                 *S->second << "} = " << ESContrib << "\n");
   1960               EffSize -= ESContrib;
   1961             }
   1962           }
   1963 
   1964           // Compute the cost of incoming edges.
   1965           if (!isa<LoadInst>(S->first) && !isa<StoreInst>(S->first)) {
   1966             Instruction *S1 = cast<Instruction>(S->first),
   1967                         *S2 = cast<Instruction>(S->second);
   1968             for (unsigned o = 0; o < S1->getNumOperands(); ++o) {
   1969               Value *O1 = S1->getOperand(o), *O2 = S2->getOperand(o);
   1970 
   1971               // Combining constants into vector constants (or small vector
   1972               // constants into larger ones are assumed free).
   1973               if (isa<Constant>(O1) && isa<Constant>(O2))
   1974                 continue;
   1975 
   1976               if (FlipOrder)
   1977                 std::swap(O1, O2);
   1978 
   1979               ValuePair VP  = ValuePair(O1, O2);
   1980               ValuePair VPR = ValuePair(O2, O1);
   1981 
   1982               // Internal edges are not handled here.
   1983               if (PrunedDAG.count(VP) || PrunedDAG.count(VPR))
   1984                 continue;
   1985 
   1986               Type *Ty1 = O1->getType(),
   1987                    *Ty2 = O2->getType();
   1988               Type *VTy = getVecTypeForPair(Ty1, Ty2);
   1989 
   1990               // Combining vector operations of the same type is also assumed
   1991               // folded with other operations.
   1992               if (Ty1 == Ty2) {
   1993                 // If both are insert elements, then both can be widened.
   1994                 InsertElementInst *IEO1 = dyn_cast<InsertElementInst>(O1),
   1995                                   *IEO2 = dyn_cast<InsertElementInst>(O2);
   1996                 if (IEO1 && IEO2 && isPureIEChain(IEO1) && isPureIEChain(IEO2))
   1997                   continue;
   1998                 // If both are extract elements, and both have the same input
   1999                 // type, then they can be replaced with a shuffle
   2000                 ExtractElementInst *EIO1 = dyn_cast<ExtractElementInst>(O1),
   2001                                    *EIO2 = dyn_cast<ExtractElementInst>(O2);
   2002                 if (EIO1 && EIO2 &&
   2003                     EIO1->getOperand(0)->getType() ==
   2004                       EIO2->getOperand(0)->getType())
   2005                   continue;
   2006                 // If both are a shuffle with equal operand types and only two
   2007                 // unqiue operands, then they can be replaced with a single
   2008                 // shuffle
   2009                 ShuffleVectorInst *SIO1 = dyn_cast<ShuffleVectorInst>(O1),
   2010                                   *SIO2 = dyn_cast<ShuffleVectorInst>(O2);
   2011                 if (SIO1 && SIO2 &&
   2012                     SIO1->getOperand(0)->getType() ==
   2013                       SIO2->getOperand(0)->getType()) {
   2014                   SmallSet<Value *, 4> SIOps;
   2015                   SIOps.insert(SIO1->getOperand(0));
   2016                   SIOps.insert(SIO1->getOperand(1));
   2017                   SIOps.insert(SIO2->getOperand(0));
   2018                   SIOps.insert(SIO2->getOperand(1));
   2019                   if (SIOps.size() <= 2)
   2020                     continue;
   2021                 }
   2022               }
   2023 
   2024               int ESContrib;
   2025               // This pair has already been formed.
   2026               if (IncomingPairs.count(VP)) {
   2027                 continue;
   2028               } else if (IncomingPairs.count(VPR)) {
   2029                 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
   2030                                                VTy, VTy);
   2031 
   2032                 if (VTy->getVectorNumElements() == 2)
   2033                   ESContrib = std::min(ESContrib, (int) TTI->getShuffleCost(
   2034                     TargetTransformInfo::SK_Reverse, VTy));
   2035               } else if (!Ty1->isVectorTy() && !Ty2->isVectorTy()) {
   2036                 ESContrib = (int) TTI->getVectorInstrCost(
   2037                                     Instruction::InsertElement, VTy, 0);
   2038                 ESContrib += (int) TTI->getVectorInstrCost(
   2039                                      Instruction::InsertElement, VTy, 1);
   2040               } else if (!Ty1->isVectorTy()) {
   2041                 // O1 needs to be inserted into a vector of size O2, and then
   2042                 // both need to be shuffled together.
   2043                 ESContrib = (int) TTI->getVectorInstrCost(
   2044                                     Instruction::InsertElement, Ty2, 0);
   2045                 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
   2046                                                 VTy, Ty2);
   2047               } else if (!Ty2->isVectorTy()) {
   2048                 // O2 needs to be inserted into a vector of size O1, and then
   2049                 // both need to be shuffled together.
   2050                 ESContrib = (int) TTI->getVectorInstrCost(
   2051                                     Instruction::InsertElement, Ty1, 0);
   2052                 ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
   2053                                                 VTy, Ty1);
   2054               } else {
   2055                 Type *TyBig = Ty1, *TySmall = Ty2;
   2056                 if (Ty2->getVectorNumElements() > Ty1->getVectorNumElements())
   2057                   std::swap(TyBig, TySmall);
   2058 
   2059                 ESContrib = (int) getInstrCost(Instruction::ShuffleVector,
   2060                                                VTy, TyBig);
   2061                 if (TyBig != TySmall)
   2062                   ESContrib += (int) getInstrCost(Instruction::ShuffleVector,
   2063                                                   TyBig, TySmall);
   2064               }
   2065 
   2066               DEBUG(if (DebugPairSelection) dbgs() << "\tcost {"
   2067                      << *O1 << " <-> " << *O2 << "} = " <<
   2068                      ESContrib << "\n");
   2069               EffSize -= ESContrib;
   2070               IncomingPairs.insert(VP);
   2071             }
   2072           }
   2073         }
   2074 
   2075         if (!HasNontrivialInsts) {
   2076           DEBUG(if (DebugPairSelection) dbgs() <<
   2077                 "\tNo non-trivial instructions in DAG;"
   2078                 " override to zero effective size\n");
   2079           EffSize = 0;
   2080         }
   2081       } else {
   2082         for (DenseSet<ValuePair>::iterator S = PrunedDAG.begin(),
   2083              E = PrunedDAG.end(); S != E; ++S)
   2084           EffSize += (int) getDepthFactor(S->first);
   2085       }
   2086 
   2087       DEBUG(if (DebugPairSelection)
   2088              dbgs() << "BBV: found pruned DAG for pair {"
   2089              << *IJ.first << " <-> " << *IJ.second << "} of depth " <<
   2090              MaxDepth << " and size " << PrunedDAG.size() <<
   2091             " (effective size: " << EffSize << ")\n");
   2092       if (((TTI && !UseChainDepthWithTI) ||
   2093             MaxDepth >= Config.ReqChainDepth) &&
   2094           EffSize > 0 && EffSize > BestEffSize) {
   2095         BestMaxDepth = MaxDepth;
   2096         BestEffSize = EffSize;
   2097         BestDAG = PrunedDAG;
   2098       }
   2099     }
   2100   }
   2101 
   2102   // Given the list of candidate pairs, this function selects those
   2103   // that will be fused into vector instructions.
   2104   void BBVectorize::choosePairs(
   2105                 DenseMap<Value *, std::vector<Value *> > &CandidatePairs,
   2106                 DenseSet<ValuePair> &CandidatePairsSet,
   2107                 DenseMap<ValuePair, int> &CandidatePairCostSavings,
   2108                 std::vector<Value *> &PairableInsts,
   2109                 DenseSet<ValuePair> &FixedOrderPairs,
   2110                 DenseMap<VPPair, unsigned> &PairConnectionTypes,
   2111                 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
   2112                 DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps,
   2113                 DenseSet<ValuePair> &PairableInstUsers,
   2114                 DenseMap<Value *, Value *>& ChosenPairs) {
   2115     bool UseCycleCheck =
   2116      CandidatePairsSet.size() <= Config.MaxCandPairsForCycleCheck;
   2117 
   2118     DenseMap<Value *, std::vector<Value *> > CandidatePairs2;
   2119     for (DenseSet<ValuePair>::iterator I = CandidatePairsSet.begin(),
   2120          E = CandidatePairsSet.end(); I != E; ++I) {
   2121       std::vector<Value *> &JJ = CandidatePairs2[I->second];
   2122       if (JJ.empty()) JJ.reserve(32);
   2123       JJ.push_back(I->first);
   2124     }
   2125 
   2126     DenseMap<ValuePair, std::vector<ValuePair> > PairableInstUserMap;
   2127     DenseSet<VPPair> PairableInstUserPairSet;
   2128     for (std::vector<Value *>::iterator I = PairableInsts.begin(),
   2129          E = PairableInsts.end(); I != E; ++I) {
   2130       // The number of possible pairings for this variable:
   2131       size_t NumChoices = CandidatePairs.lookup(*I).size();
   2132       if (!NumChoices) continue;
   2133 
   2134       std::vector<Value *> &JJ = CandidatePairs[*I];
   2135 
   2136       // The best pair to choose and its dag:
   2137       size_t BestMaxDepth = 0;
   2138       int BestEffSize = 0;
   2139       DenseSet<ValuePair> BestDAG;
   2140       findBestDAGFor(CandidatePairs, CandidatePairsSet,
   2141                       CandidatePairCostSavings,
   2142                       PairableInsts, FixedOrderPairs, PairConnectionTypes,
   2143                       ConnectedPairs, ConnectedPairDeps,
   2144                       PairableInstUsers, PairableInstUserMap,
   2145                       PairableInstUserPairSet, ChosenPairs,
   2146                       BestDAG, BestMaxDepth, BestEffSize, *I, JJ,
   2147                       UseCycleCheck);
   2148 
   2149       if (BestDAG.empty())
   2150         continue;
   2151 
   2152       // A dag has been chosen (or not) at this point. If no dag was
   2153       // chosen, then this instruction, I, cannot be paired (and is no longer
   2154       // considered).
   2155 
   2156       DEBUG(dbgs() << "BBV: selected pairs in the best DAG for: "
   2157                    << *cast<Instruction>(*I) << "\n");
   2158 
   2159       for (DenseSet<ValuePair>::iterator S = BestDAG.begin(),
   2160            SE2 = BestDAG.end(); S != SE2; ++S) {
   2161         // Insert the members of this dag into the list of chosen pairs.
   2162         ChosenPairs.insert(ValuePair(S->first, S->second));
   2163         DEBUG(dbgs() << "BBV: selected pair: " << *S->first << " <-> " <<
   2164                *S->second << "\n");
   2165 
   2166         // Remove all candidate pairs that have values in the chosen dag.
   2167         std::vector<Value *> &KK = CandidatePairs[S->first];
   2168         for (std::vector<Value *>::iterator K = KK.begin(), KE = KK.end();
   2169              K != KE; ++K) {
   2170           if (*K == S->second)
   2171             continue;
   2172 
   2173           CandidatePairsSet.erase(ValuePair(S->first, *K));
   2174         }
   2175 
   2176         std::vector<Value *> &LL = CandidatePairs2[S->second];
   2177         for (std::vector<Value *>::iterator L = LL.begin(), LE = LL.end();
   2178              L != LE; ++L) {
   2179           if (*L == S->first)
   2180             continue;
   2181 
   2182           CandidatePairsSet.erase(ValuePair(*L, S->second));
   2183         }
   2184 
   2185         std::vector<Value *> &MM = CandidatePairs[S->second];
   2186         for (std::vector<Value *>::iterator M = MM.begin(), ME = MM.end();
   2187              M != ME; ++M) {
   2188           assert(*M != S->first && "Flipped pair in candidate list?");
   2189           CandidatePairsSet.erase(ValuePair(S->second, *M));
   2190         }
   2191 
   2192         std::vector<Value *> &NN = CandidatePairs2[S->first];
   2193         for (std::vector<Value *>::iterator N = NN.begin(), NE = NN.end();
   2194              N != NE; ++N) {
   2195           assert(*N != S->second && "Flipped pair in candidate list?");
   2196           CandidatePairsSet.erase(ValuePair(*N, S->first));
   2197         }
   2198       }
   2199     }
   2200 
   2201     DEBUG(dbgs() << "BBV: selected " << ChosenPairs.size() << " pairs.\n");
   2202   }
   2203 
   2204   std::string getReplacementName(Instruction *I, bool IsInput, unsigned o,
   2205                      unsigned n = 0) {
   2206     if (!I->hasName())
   2207       return "";
   2208 
   2209     return (I->getName() + (IsInput ? ".v.i" : ".v.r") + utostr(o) +
   2210              (n > 0 ? "." + utostr(n) : "")).str();
   2211   }
   2212 
   2213   // Returns the value that is to be used as the pointer input to the vector
   2214   // instruction that fuses I with J.
   2215   Value *BBVectorize::getReplacementPointerInput(LLVMContext& Context,
   2216                      Instruction *I, Instruction *J, unsigned o) {
   2217     Value *IPtr, *JPtr;
   2218     unsigned IAlignment, JAlignment, IAddressSpace, JAddressSpace;
   2219     int64_t OffsetInElmts;
   2220 
   2221     // Note: the analysis might fail here, that is why the pair order has
   2222     // been precomputed (OffsetInElmts must be unused here).
   2223     (void) getPairPtrInfo(I, J, IPtr, JPtr, IAlignment, JAlignment,
   2224                           IAddressSpace, JAddressSpace,
   2225                           OffsetInElmts, false);
   2226 
   2227     // The pointer value is taken to be the one with the lowest offset.
   2228     Value *VPtr = IPtr;
   2229 
   2230     Type *ArgTypeI = cast<PointerType>(IPtr->getType())->getElementType();
   2231     Type *ArgTypeJ = cast<PointerType>(JPtr->getType())->getElementType();
   2232     Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
   2233     Type *VArgPtrType = PointerType::get(VArgType,
   2234       cast<PointerType>(IPtr->getType())->getAddressSpace());
   2235     return new BitCastInst(VPtr, VArgPtrType, getReplacementName(I, true, o),
   2236                         /* insert before */ I);
   2237   }
   2238 
   2239   void BBVectorize::fillNewShuffleMask(LLVMContext& Context, Instruction *J,
   2240                      unsigned MaskOffset, unsigned NumInElem,
   2241                      unsigned NumInElem1, unsigned IdxOffset,
   2242                      std::vector<Constant*> &Mask) {
   2243     unsigned NumElem1 = cast<VectorType>(J->getType())->getNumElements();
   2244     for (unsigned v = 0; v < NumElem1; ++v) {
   2245       int m = cast<ShuffleVectorInst>(J)->getMaskValue(v);
   2246       if (m < 0) {
   2247         Mask[v+MaskOffset] = UndefValue::get(Type::getInt32Ty(Context));
   2248       } else {
   2249         unsigned mm = m + (int) IdxOffset;
   2250         if (m >= (int) NumInElem1)
   2251           mm += (int) NumInElem;
   2252 
   2253         Mask[v+MaskOffset] =
   2254           ConstantInt::get(Type::getInt32Ty(Context), mm);
   2255       }
   2256     }
   2257   }
   2258 
   2259   // Returns the value that is to be used as the vector-shuffle mask to the
   2260   // vector instruction that fuses I with J.
   2261   Value *BBVectorize::getReplacementShuffleMask(LLVMContext& Context,
   2262                      Instruction *I, Instruction *J) {
   2263     // This is the shuffle mask. We need to append the second
   2264     // mask to the first, and the numbers need to be adjusted.
   2265 
   2266     Type *ArgTypeI = I->getType();
   2267     Type *ArgTypeJ = J->getType();
   2268     Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
   2269 
   2270     unsigned NumElemI = cast<VectorType>(ArgTypeI)->getNumElements();
   2271 
   2272     // Get the total number of elements in the fused vector type.
   2273     // By definition, this must equal the number of elements in
   2274     // the final mask.
   2275     unsigned NumElem = cast<VectorType>(VArgType)->getNumElements();
   2276     std::vector<Constant*> Mask(NumElem);
   2277 
   2278     Type *OpTypeI = I->getOperand(0)->getType();
   2279     unsigned NumInElemI = cast<VectorType>(OpTypeI)->getNumElements();
   2280     Type *OpTypeJ = J->getOperand(0)->getType();
   2281     unsigned NumInElemJ = cast<VectorType>(OpTypeJ)->getNumElements();
   2282 
   2283     // The fused vector will be:
   2284     // -----------------------------------------------------
   2285     // | NumInElemI | NumInElemJ | NumInElemI | NumInElemJ |
   2286     // -----------------------------------------------------
   2287     // from which we'll extract NumElem total elements (where the first NumElemI
   2288     // of them come from the mask in I and the remainder come from the mask
   2289     // in J.
   2290 
   2291     // For the mask from the first pair...
   2292     fillNewShuffleMask(Context, I, 0,        NumInElemJ, NumInElemI,
   2293                        0,          Mask);
   2294 
   2295     // For the mask from the second pair...
   2296     fillNewShuffleMask(Context, J, NumElemI, NumInElemI, NumInElemJ,
   2297                        NumInElemI, Mask);
   2298 
   2299     return ConstantVector::get(Mask);
   2300   }
   2301 
   2302   bool BBVectorize::expandIEChain(LLVMContext& Context, Instruction *I,
   2303                                   Instruction *J, unsigned o, Value *&LOp,
   2304                                   unsigned numElemL,
   2305                                   Type *ArgTypeL, Type *ArgTypeH,
   2306                                   bool IBeforeJ, unsigned IdxOff) {
   2307     bool ExpandedIEChain = false;
   2308     if (InsertElementInst *LIE = dyn_cast<InsertElementInst>(LOp)) {
   2309       // If we have a pure insertelement chain, then this can be rewritten
   2310       // into a chain that directly builds the larger type.
   2311       if (isPureIEChain(LIE)) {
   2312         SmallVector<Value *, 8> VectElemts(numElemL,
   2313           UndefValue::get(ArgTypeL->getScalarType()));
   2314         InsertElementInst *LIENext = LIE;
   2315         do {
   2316           unsigned Idx =
   2317             cast<ConstantInt>(LIENext->getOperand(2))->getSExtValue();
   2318           VectElemts[Idx] = LIENext->getOperand(1);
   2319         } while ((LIENext =
   2320                    dyn_cast<InsertElementInst>(LIENext->getOperand(0))));
   2321 
   2322         LIENext = 0;
   2323         Value *LIEPrev = UndefValue::get(ArgTypeH);
   2324         for (unsigned i = 0; i < numElemL; ++i) {
   2325           if (isa<UndefValue>(VectElemts[i])) continue;
   2326           LIENext = InsertElementInst::Create(LIEPrev, VectElemts[i],
   2327                              ConstantInt::get(Type::getInt32Ty(Context),
   2328                                               i + IdxOff),
   2329                              getReplacementName(IBeforeJ ? I : J,
   2330                                                 true, o, i+1));
   2331           LIENext->insertBefore(IBeforeJ ? J : I);
   2332           LIEPrev = LIENext;
   2333         }
   2334 
   2335         LOp = LIENext ? (Value*) LIENext : UndefValue::get(ArgTypeH);
   2336         ExpandedIEChain = true;
   2337       }
   2338     }
   2339 
   2340     return ExpandedIEChain;
   2341   }
   2342 
   2343   // Returns the value to be used as the specified operand of the vector
   2344   // instruction that fuses I with J.
   2345   Value *BBVectorize::getReplacementInput(LLVMContext& Context, Instruction *I,
   2346                      Instruction *J, unsigned o, bool IBeforeJ) {
   2347     Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
   2348     Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), 1);
   2349 
   2350     // Compute the fused vector type for this operand
   2351     Type *ArgTypeI = I->getOperand(o)->getType();
   2352     Type *ArgTypeJ = J->getOperand(o)->getType();
   2353     VectorType *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
   2354 
   2355     Instruction *L = I, *H = J;
   2356     Type *ArgTypeL = ArgTypeI, *ArgTypeH = ArgTypeJ;
   2357 
   2358     unsigned numElemL;
   2359     if (ArgTypeL->isVectorTy())
   2360       numElemL = cast<VectorType>(ArgTypeL)->getNumElements();
   2361     else
   2362       numElemL = 1;
   2363 
   2364     unsigned numElemH;
   2365     if (ArgTypeH->isVectorTy())
   2366       numElemH = cast<VectorType>(ArgTypeH)->getNumElements();
   2367     else
   2368       numElemH = 1;
   2369 
   2370     Value *LOp = L->getOperand(o);
   2371     Value *HOp = H->getOperand(o);
   2372     unsigned numElem = VArgType->getNumElements();
   2373 
   2374     // First, we check if we can reuse the "original" vector outputs (if these
   2375     // exist). We might need a shuffle.
   2376     ExtractElementInst *LEE = dyn_cast<ExtractElementInst>(LOp);
   2377     ExtractElementInst *HEE = dyn_cast<ExtractElementInst>(HOp);
   2378     ShuffleVectorInst *LSV = dyn_cast<ShuffleVectorInst>(LOp);
   2379     ShuffleVectorInst *HSV = dyn_cast<ShuffleVectorInst>(HOp);
   2380 
   2381     // FIXME: If we're fusing shuffle instructions, then we can't apply this
   2382     // optimization. The input vectors to the shuffle might be a different
   2383     // length from the shuffle outputs. Unfortunately, the replacement
   2384     // shuffle mask has already been formed, and the mask entries are sensitive
   2385     // to the sizes of the inputs.
   2386     bool IsSizeChangeShuffle =
   2387       isa<ShuffleVectorInst>(L) &&
   2388         (LOp->getType() != L->getType() || HOp->getType() != H->getType());
   2389 
   2390     if ((LEE || LSV) && (HEE || HSV) && !IsSizeChangeShuffle) {
   2391       // We can have at most two unique vector inputs.
   2392       bool CanUseInputs = true;
   2393       Value *I1, *I2 = 0;
   2394       if (LEE) {
   2395         I1 = LEE->getOperand(0);
   2396       } else {
   2397         I1 = LSV->getOperand(0);
   2398         I2 = LSV->getOperand(1);
   2399         if (I2 == I1 || isa<UndefValue>(I2))
   2400           I2 = 0;
   2401       }
   2402 
   2403       if (HEE) {
   2404         Value *I3 = HEE->getOperand(0);
   2405         if (!I2 && I3 != I1)
   2406           I2 = I3;
   2407         else if (I3 != I1 && I3 != I2)
   2408           CanUseInputs = false;
   2409       } else {
   2410         Value *I3 = HSV->getOperand(0);
   2411         if (!I2 && I3 != I1)
   2412           I2 = I3;
   2413         else if (I3 != I1 && I3 != I2)
   2414           CanUseInputs = false;
   2415 
   2416         if (CanUseInputs) {
   2417           Value *I4 = HSV->getOperand(1);
   2418           if (!isa<UndefValue>(I4)) {
   2419             if (!I2 && I4 != I1)
   2420               I2 = I4;
   2421             else if (I4 != I1 && I4 != I2)
   2422               CanUseInputs = false;
   2423           }
   2424         }
   2425       }
   2426 
   2427       if (CanUseInputs) {
   2428         unsigned LOpElem =
   2429           cast<VectorType>(cast<Instruction>(LOp)->getOperand(0)->getType())
   2430             ->getNumElements();
   2431         unsigned HOpElem =
   2432           cast<VectorType>(cast<Instruction>(HOp)->getOperand(0)->getType())
   2433             ->getNumElements();
   2434 
   2435         // We have one or two input vectors. We need to map each index of the
   2436         // operands to the index of the original vector.
   2437         SmallVector<std::pair<int, int>, 8>  II(numElem);
   2438         for (unsigned i = 0; i < numElemL; ++i) {
   2439           int Idx, INum;
   2440           if (LEE) {
   2441             Idx =
   2442               cast<ConstantInt>(LEE->getOperand(1))->getSExtValue();
   2443             INum = LEE->getOperand(0) == I1 ? 0 : 1;
   2444           } else {
   2445             Idx = LSV->getMaskValue(i);
   2446             if (Idx < (int) LOpElem) {
   2447               INum = LSV->getOperand(0) == I1 ? 0 : 1;
   2448             } else {
   2449               Idx -= LOpElem;
   2450               INum = LSV->getOperand(1) == I1 ? 0 : 1;
   2451             }
   2452           }
   2453 
   2454           II[i] = std::pair<int, int>(Idx, INum);
   2455         }
   2456         for (unsigned i = 0; i < numElemH; ++i) {
   2457           int Idx, INum;
   2458           if (HEE) {
   2459             Idx =
   2460               cast<ConstantInt>(HEE->getOperand(1))->getSExtValue();
   2461             INum = HEE->getOperand(0) == I1 ? 0 : 1;
   2462           } else {
   2463             Idx = HSV->getMaskValue(i);
   2464             if (Idx < (int) HOpElem) {
   2465               INum = HSV->getOperand(0) == I1 ? 0 : 1;
   2466             } else {
   2467               Idx -= HOpElem;
   2468               INum = HSV->getOperand(1) == I1 ? 0 : 1;
   2469             }
   2470           }
   2471 
   2472           II[i + numElemL] = std::pair<int, int>(Idx, INum);
   2473         }
   2474 
   2475         // We now have an array which tells us from which index of which
   2476         // input vector each element of the operand comes.
   2477         VectorType *I1T = cast<VectorType>(I1->getType());
   2478         unsigned I1Elem = I1T->getNumElements();
   2479 
   2480         if (!I2) {
   2481           // In this case there is only one underlying vector input. Check for
   2482           // the trivial case where we can use the input directly.
   2483           if (I1Elem == numElem) {
   2484             bool ElemInOrder = true;
   2485             for (unsigned i = 0; i < numElem; ++i) {
   2486               if (II[i].first != (int) i && II[i].first != -1) {
   2487                 ElemInOrder = false;
   2488                 break;
   2489               }
   2490             }
   2491 
   2492             if (ElemInOrder)
   2493               return I1;
   2494           }
   2495 
   2496           // A shuffle is needed.
   2497           std::vector<Constant *> Mask(numElem);
   2498           for (unsigned i = 0; i < numElem; ++i) {
   2499             int Idx = II[i].first;
   2500             if (Idx == -1)
   2501               Mask[i] = UndefValue::get(Type::getInt32Ty(Context));
   2502             else
   2503               Mask[i] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
   2504           }
   2505 
   2506           Instruction *S =
   2507             new ShuffleVectorInst(I1, UndefValue::get(I1T),
   2508                                   ConstantVector::get(Mask),
   2509                                   getReplacementName(IBeforeJ ? I : J,
   2510                                                      true, o));
   2511           S->insertBefore(IBeforeJ ? J : I);
   2512           return S;
   2513         }
   2514 
   2515         VectorType *I2T = cast<VectorType>(I2->getType());
   2516         unsigned I2Elem = I2T->getNumElements();
   2517 
   2518         // This input comes from two distinct vectors. The first step is to
   2519         // make sure that both vectors are the same length. If not, the
   2520         // smaller one will need to grow before they can be shuffled together.
   2521         if (I1Elem < I2Elem) {
   2522           std::vector<Constant *> Mask(I2Elem);
   2523           unsigned v = 0;
   2524           for (; v < I1Elem; ++v)
   2525             Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
   2526           for (; v < I2Elem; ++v)
   2527             Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
   2528 
   2529           Instruction *NewI1 =
   2530             new ShuffleVectorInst(I1, UndefValue::get(I1T),
   2531                                   ConstantVector::get(Mask),
   2532                                   getReplacementName(IBeforeJ ? I : J,
   2533                                                      true, o, 1));
   2534           NewI1->insertBefore(IBeforeJ ? J : I);
   2535           I1 = NewI1;
   2536           I1T = I2T;
   2537           I1Elem = I2Elem;
   2538         } else if (I1Elem > I2Elem) {
   2539           std::vector<Constant *> Mask(I1Elem);
   2540           unsigned v = 0;
   2541           for (; v < I2Elem; ++v)
   2542             Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
   2543           for (; v < I1Elem; ++v)
   2544             Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
   2545 
   2546           Instruction *NewI2 =
   2547             new ShuffleVectorInst(I2, UndefValue::get(I2T),
   2548                                   ConstantVector::get(Mask),
   2549                                   getReplacementName(IBeforeJ ? I : J,
   2550                                                      true, o, 1));
   2551           NewI2->insertBefore(IBeforeJ ? J : I);
   2552           I2 = NewI2;
   2553           I2T = I1T;
   2554           I2Elem = I1Elem;
   2555         }
   2556 
   2557         // Now that both I1 and I2 are the same length we can shuffle them
   2558         // together (and use the result).
   2559         std::vector<Constant *> Mask(numElem);
   2560         for (unsigned v = 0; v < numElem; ++v) {
   2561           if (II[v].first == -1) {
   2562             Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
   2563           } else {
   2564             int Idx = II[v].first + II[v].second * I1Elem;
   2565             Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
   2566           }
   2567         }
   2568 
   2569         Instruction *NewOp =
   2570           new ShuffleVectorInst(I1, I2, ConstantVector::get(Mask),
   2571                                 getReplacementName(IBeforeJ ? I : J, true, o));
   2572         NewOp->insertBefore(IBeforeJ ? J : I);
   2573         return NewOp;
   2574       }
   2575     }
   2576 
   2577     Type *ArgType = ArgTypeL;
   2578     if (numElemL < numElemH) {
   2579       if (numElemL == 1 && expandIEChain(Context, I, J, o, HOp, numElemH,
   2580                                          ArgTypeL, VArgType, IBeforeJ, 1)) {
   2581         // This is another short-circuit case: we're combining a scalar into
   2582         // a vector that is formed by an IE chain. We've just expanded the IE
   2583         // chain, now insert the scalar and we're done.
   2584 
   2585         Instruction *S = InsertElementInst::Create(HOp, LOp, CV0,
   2586                            getReplacementName(IBeforeJ ? I : J, true, o));
   2587         S->insertBefore(IBeforeJ ? J : I);
   2588         return S;
   2589       } else if (!expandIEChain(Context, I, J, o, LOp, numElemL, ArgTypeL,
   2590                                 ArgTypeH, IBeforeJ)) {
   2591         // The two vector inputs to the shuffle must be the same length,
   2592         // so extend the smaller vector to be the same length as the larger one.
   2593         Instruction *NLOp;
   2594         if (numElemL > 1) {
   2595 
   2596           std::vector<Constant *> Mask(numElemH);
   2597           unsigned v = 0;
   2598           for (; v < numElemL; ++v)
   2599             Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
   2600           for (; v < numElemH; ++v)
   2601             Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
   2602 
   2603           NLOp = new ShuffleVectorInst(LOp, UndefValue::get(ArgTypeL),
   2604                                        ConstantVector::get(Mask),
   2605                                        getReplacementName(IBeforeJ ? I : J,
   2606                                                           true, o, 1));
   2607         } else {
   2608           NLOp = InsertElementInst::Create(UndefValue::get(ArgTypeH), LOp, CV0,
   2609                                            getReplacementName(IBeforeJ ? I : J,
   2610                                                               true, o, 1));
   2611         }
   2612 
   2613         NLOp->insertBefore(IBeforeJ ? J : I);
   2614         LOp = NLOp;
   2615       }
   2616 
   2617       ArgType = ArgTypeH;
   2618     } else if (numElemL > numElemH) {
   2619       if (numElemH == 1 && expandIEChain(Context, I, J, o, LOp, numElemL,
   2620                                          ArgTypeH, VArgType, IBeforeJ)) {
   2621         Instruction *S =
   2622           InsertElementInst::Create(LOp, HOp,
   2623                                     ConstantInt::get(Type::getInt32Ty(Context),
   2624                                                      numElemL),
   2625                                     getReplacementName(IBeforeJ ? I : J,
   2626                                                        true, o));
   2627         S->insertBefore(IBeforeJ ? J : I);
   2628         return S;
   2629       } else if (!expandIEChain(Context, I, J, o, HOp, numElemH, ArgTypeH,
   2630                                 ArgTypeL, IBeforeJ)) {
   2631         Instruction *NHOp;
   2632         if (numElemH > 1) {
   2633           std::vector<Constant *> Mask(numElemL);
   2634           unsigned v = 0;
   2635           for (; v < numElemH; ++v)
   2636             Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
   2637           for (; v < numElemL; ++v)
   2638             Mask[v] = UndefValue::get(Type::getInt32Ty(Context));
   2639 
   2640           NHOp = new ShuffleVectorInst(HOp, UndefValue::get(ArgTypeH),
   2641                                        ConstantVector::get(Mask),
   2642                                        getReplacementName(IBeforeJ ? I : J,
   2643                                                           true, o, 1));
   2644         } else {
   2645           NHOp = InsertElementInst::Create(UndefValue::get(ArgTypeL), HOp, CV0,
   2646                                            getReplacementName(IBeforeJ ? I : J,
   2647                                                               true, o, 1));
   2648         }
   2649 
   2650         NHOp->insertBefore(IBeforeJ ? J : I);
   2651         HOp = NHOp;
   2652       }
   2653     }
   2654 
   2655     if (ArgType->isVectorTy()) {
   2656       unsigned numElem = cast<VectorType>(VArgType)->getNumElements();
   2657       std::vector<Constant*> Mask(numElem);
   2658       for (unsigned v = 0; v < numElem; ++v) {
   2659         unsigned Idx = v;
   2660         // If the low vector was expanded, we need to skip the extra
   2661         // undefined entries.
   2662         if (v >= numElemL && numElemH > numElemL)
   2663           Idx += (numElemH - numElemL);
   2664         Mask[v] = ConstantInt::get(Type::getInt32Ty(Context), Idx);
   2665       }
   2666 
   2667       Instruction *BV = new ShuffleVectorInst(LOp, HOp,
   2668                           ConstantVector::get(Mask),
   2669                           getReplacementName(IBeforeJ ? I : J, true, o));
   2670       BV->insertBefore(IBeforeJ ? J : I);
   2671       return BV;
   2672     }
   2673 
   2674     Instruction *BV1 = InsertElementInst::Create(
   2675                                           UndefValue::get(VArgType), LOp, CV0,
   2676                                           getReplacementName(IBeforeJ ? I : J,
   2677                                                              true, o, 1));
   2678     BV1->insertBefore(IBeforeJ ? J : I);
   2679     Instruction *BV2 = InsertElementInst::Create(BV1, HOp, CV1,
   2680                                           getReplacementName(IBeforeJ ? I : J,
   2681                                                              true, o, 2));
   2682     BV2->insertBefore(IBeforeJ ? J : I);
   2683     return BV2;
   2684   }
   2685 
   2686   // This function creates an array of values that will be used as the inputs
   2687   // to the vector instruction that fuses I with J.
   2688   void BBVectorize::getReplacementInputsForPair(LLVMContext& Context,
   2689                      Instruction *I, Instruction *J,
   2690                      SmallVectorImpl<Value *> &ReplacedOperands,
   2691                      bool IBeforeJ) {
   2692     unsigned NumOperands = I->getNumOperands();
   2693 
   2694     for (unsigned p = 0, o = NumOperands-1; p < NumOperands; ++p, --o) {
   2695       // Iterate backward so that we look at the store pointer
   2696       // first and know whether or not we need to flip the inputs.
   2697 
   2698       if (isa<LoadInst>(I) || (o == 1 && isa<StoreInst>(I))) {
   2699         // This is the pointer for a load/store instruction.
   2700         ReplacedOperands[o] = getReplacementPointerInput(Context, I, J, o);
   2701         continue;
   2702       } else if (isa<CallInst>(I)) {
   2703         Function *F = cast<CallInst>(I)->getCalledFunction();
   2704         Intrinsic::ID IID = (Intrinsic::ID) F->getIntrinsicID();
   2705         if (o == NumOperands-1) {
   2706           BasicBlock &BB = *I->getParent();
   2707 
   2708           Module *M = BB.getParent()->getParent();
   2709           Type *ArgTypeI = I->getType();
   2710           Type *ArgTypeJ = J->getType();
   2711           Type *VArgType = getVecTypeForPair(ArgTypeI, ArgTypeJ);
   2712 
   2713           ReplacedOperands[o] = Intrinsic::getDeclaration(M, IID, VArgType);
   2714           continue;
   2715         } else if (IID == Intrinsic::powi && o == 1) {
   2716           // The second argument of powi is a single integer and we've already
   2717           // checked that both arguments are equal. As a result, we just keep
   2718           // I's second argument.
   2719           ReplacedOperands[o] = I->getOperand(o);
   2720           continue;
   2721         }
   2722       } else if (isa<ShuffleVectorInst>(I) && o == NumOperands-1) {
   2723         ReplacedOperands[o] = getReplacementShuffleMask(Context, I, J);
   2724         continue;
   2725       }
   2726 
   2727       ReplacedOperands[o] = getReplacementInput(Context, I, J, o, IBeforeJ);
   2728     }
   2729   }
   2730 
   2731   // This function creates two values that represent the outputs of the
   2732   // original I and J instructions. These are generally vector shuffles
   2733   // or extracts. In many cases, these will end up being unused and, thus,
   2734   // eliminated by later passes.
   2735   void BBVectorize::replaceOutputsOfPair(LLVMContext& Context, Instruction *I,
   2736                      Instruction *J, Instruction *K,
   2737                      Instruction *&InsertionPt,
   2738                      Instruction *&K1, Instruction *&K2) {
   2739     if (isa<StoreInst>(I)) {
   2740       AA->replaceWithNewValue(I, K);
   2741       AA->replaceWithNewValue(J, K);
   2742     } else {
   2743       Type *IType = I->getType();
   2744       Type *JType = J->getType();
   2745 
   2746       VectorType *VType = getVecTypeForPair(IType, JType);
   2747       unsigned numElem = VType->getNumElements();
   2748 
   2749       unsigned numElemI, numElemJ;
   2750       if (IType->isVectorTy())
   2751         numElemI = cast<VectorType>(IType)->getNumElements();
   2752       else
   2753         numElemI = 1;
   2754 
   2755       if (JType->isVectorTy())
   2756         numElemJ = cast<VectorType>(JType)->getNumElements();
   2757       else
   2758         numElemJ = 1;
   2759 
   2760       if (IType->isVectorTy()) {
   2761         std::vector<Constant*> Mask1(numElemI), Mask2(numElemI);
   2762         for (unsigned v = 0; v < numElemI; ++v) {
   2763           Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
   2764           Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemJ+v);
   2765         }
   2766 
   2767         K1 = new ShuffleVectorInst(K, UndefValue::get(VType),
   2768                                    ConstantVector::get( Mask1),
   2769                                    getReplacementName(K, false, 1));
   2770       } else {
   2771         Value *CV0 = ConstantInt::get(Type::getInt32Ty(Context), 0);
   2772         K1 = ExtractElementInst::Create(K, CV0,
   2773                                           getReplacementName(K, false, 1));
   2774       }
   2775 
   2776       if (JType->isVectorTy()) {
   2777         std::vector<Constant*> Mask1(numElemJ), Mask2(numElemJ);
   2778         for (unsigned v = 0; v < numElemJ; ++v) {
   2779           Mask1[v] = ConstantInt::get(Type::getInt32Ty(Context), v);
   2780           Mask2[v] = ConstantInt::get(Type::getInt32Ty(Context), numElemI+v);
   2781         }
   2782 
   2783         K2 = new ShuffleVectorInst(K, UndefValue::get(VType),
   2784                                    ConstantVector::get( Mask2),
   2785                                    getReplacementName(K, false, 2));
   2786       } else {
   2787         Value *CV1 = ConstantInt::get(Type::getInt32Ty(Context), numElem-1);
   2788         K2 = ExtractElementInst::Create(K, CV1,
   2789                                           getReplacementName(K, false, 2));
   2790       }
   2791 
   2792       K1->insertAfter(K);
   2793       K2->insertAfter(K1);
   2794       InsertionPt = K2;
   2795     }
   2796   }
   2797 
   2798   // Move all uses of the function I (including pairing-induced uses) after J.
   2799   bool BBVectorize::canMoveUsesOfIAfterJ(BasicBlock &BB,
   2800                      DenseSet<ValuePair> &LoadMoveSetPairs,
   2801                      Instruction *I, Instruction *J) {
   2802     // Skip to the first instruction past I.
   2803     BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
   2804 
   2805     DenseSet<Value *> Users;
   2806     AliasSetTracker WriteSet(*AA);
   2807     for (; cast<Instruction>(L) != J; ++L)
   2808       (void) trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs);
   2809 
   2810     assert(cast<Instruction>(L) == J &&
   2811       "Tracking has not proceeded far enough to check for dependencies");
   2812     // If J is now in the use set of I, then trackUsesOfI will return true
   2813     // and we have a dependency cycle (and the fusing operation must abort).
   2814     return !trackUsesOfI(Users, WriteSet, I, J, true, &LoadMoveSetPairs);
   2815   }
   2816 
   2817   // Move all uses of the function I (including pairing-induced uses) after J.
   2818   void BBVectorize::moveUsesOfIAfterJ(BasicBlock &BB,
   2819                      DenseSet<ValuePair> &LoadMoveSetPairs,
   2820                      Instruction *&InsertionPt,
   2821                      Instruction *I, Instruction *J) {
   2822     // Skip to the first instruction past I.
   2823     BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
   2824 
   2825     DenseSet<Value *> Users;
   2826     AliasSetTracker WriteSet(*AA);
   2827     for (; cast<Instruction>(L) != J;) {
   2828       if (trackUsesOfI(Users, WriteSet, I, L, true, &LoadMoveSetPairs)) {
   2829         // Move this instruction
   2830         Instruction *InstToMove = L; ++L;
   2831 
   2832         DEBUG(dbgs() << "BBV: moving: " << *InstToMove <<
   2833                         " to after " << *InsertionPt << "\n");
   2834         InstToMove->removeFromParent();
   2835         InstToMove->insertAfter(InsertionPt);
   2836         InsertionPt = InstToMove;
   2837       } else {
   2838         ++L;
   2839       }
   2840     }
   2841   }
   2842 
   2843   // Collect all load instruction that are in the move set of a given first
   2844   // pair member.  These loads depend on the first instruction, I, and so need
   2845   // to be moved after J (the second instruction) when the pair is fused.
   2846   void BBVectorize::collectPairLoadMoveSet(BasicBlock &BB,
   2847                      DenseMap<Value *, Value *> &ChosenPairs,
   2848                      DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
   2849                      DenseSet<ValuePair> &LoadMoveSetPairs,
   2850                      Instruction *I) {
   2851     // Skip to the first instruction past I.
   2852     BasicBlock::iterator L = llvm::next(BasicBlock::iterator(I));
   2853 
   2854     DenseSet<Value *> Users;
   2855     AliasSetTracker WriteSet(*AA);
   2856 
   2857     // Note: We cannot end the loop when we reach J because J could be moved
   2858     // farther down the use chain by another instruction pairing. Also, J
   2859     // could be before I if this is an inverted input.
   2860     for (BasicBlock::iterator E = BB.end(); cast<Instruction>(L) != E; ++L) {
   2861       if (trackUsesOfI(Users, WriteSet, I, L)) {
   2862         if (L->mayReadFromMemory()) {
   2863           LoadMoveSet[L].push_back(I);
   2864           LoadMoveSetPairs.insert(ValuePair(L, I));
   2865         }
   2866       }
   2867     }
   2868   }
   2869 
   2870   // In cases where both load/stores and the computation of their pointers
   2871   // are chosen for vectorization, we can end up in a situation where the
   2872   // aliasing analysis starts returning different query results as the
   2873   // process of fusing instruction pairs continues. Because the algorithm
   2874   // relies on finding the same use dags here as were found earlier, we'll
   2875   // need to precompute the necessary aliasing information here and then
   2876   // manually update it during the fusion process.
   2877   void BBVectorize::collectLoadMoveSet(BasicBlock &BB,
   2878                      std::vector<Value *> &PairableInsts,
   2879                      DenseMap<Value *, Value *> &ChosenPairs,
   2880                      DenseMap<Value *, std::vector<Value *> > &LoadMoveSet,
   2881                      DenseSet<ValuePair> &LoadMoveSetPairs) {
   2882     for (std::vector<Value *>::iterator PI = PairableInsts.begin(),
   2883          PIE = PairableInsts.end(); PI != PIE; ++PI) {
   2884       DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(*PI);
   2885       if (P == ChosenPairs.end()) continue;
   2886 
   2887       Instruction *I = cast<Instruction>(P->first);
   2888       collectPairLoadMoveSet(BB, ChosenPairs, LoadMoveSet,
   2889                              LoadMoveSetPairs, I);
   2890     }
   2891   }
   2892 
   2893   // When the first instruction in each pair is cloned, it will inherit its
   2894   // parent's metadata. This metadata must be combined with that of the other
   2895   // instruction in a safe way.
   2896   void BBVectorize::combineMetadata(Instruction *K, const Instruction *J) {
   2897     SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
   2898     K->getAllMetadataOtherThanDebugLoc(Metadata);
   2899     for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
   2900       unsigned Kind = Metadata[i].first;
   2901       MDNode *JMD = J->getMetadata(Kind);
   2902       MDNode *KMD = Metadata[i].second;
   2903 
   2904       switch (Kind) {
   2905       default:
   2906         K->setMetadata(Kind, 0); // Remove unknown metadata
   2907         break;
   2908       case LLVMContext::MD_tbaa:
   2909         K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
   2910         break;
   2911       case LLVMContext::MD_fpmath:
   2912         K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
   2913         break;
   2914       }
   2915     }
   2916   }
   2917 
   2918   // This function fuses the chosen instruction pairs into vector instructions,
   2919   // taking care preserve any needed scalar outputs and, then, it reorders the
   2920   // remaining instructions as needed (users of the first member of the pair
   2921   // need to be moved to after the location of the second member of the pair
   2922   // because the vector instruction is inserted in the location of the pair's
   2923   // second member).
   2924   void BBVectorize::fuseChosenPairs(BasicBlock &BB,
   2925              std::vector<Value *> &PairableInsts,
   2926              DenseMap<Value *, Value *> &ChosenPairs,
   2927              DenseSet<ValuePair> &FixedOrderPairs,
   2928              DenseMap<VPPair, unsigned> &PairConnectionTypes,
   2929              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairs,
   2930              DenseMap<ValuePair, std::vector<ValuePair> > &ConnectedPairDeps) {
   2931     LLVMContext& Context = BB.getContext();
   2932 
   2933     // During the vectorization process, the order of the pairs to be fused
   2934     // could be flipped. So we'll add each pair, flipped, into the ChosenPairs
   2935     // list. After a pair is fused, the flipped pair is removed from the list.
   2936     DenseSet<ValuePair> FlippedPairs;
   2937     for (DenseMap<Value *, Value *>::iterator P = ChosenPairs.begin(),
   2938          E = ChosenPairs.end(); P != E; ++P)
   2939       FlippedPairs.insert(ValuePair(P->second, P->first));
   2940     for (DenseSet<ValuePair>::iterator P = FlippedPairs.begin(),
   2941          E = FlippedPairs.end(); P != E; ++P)
   2942       ChosenPairs.insert(*P);
   2943 
   2944     DenseMap<Value *, std::vector<Value *> > LoadMoveSet;
   2945     DenseSet<ValuePair> LoadMoveSetPairs;
   2946     collectLoadMoveSet(BB, PairableInsts, ChosenPairs,
   2947                        LoadMoveSet, LoadMoveSetPairs);
   2948 
   2949     DEBUG(dbgs() << "BBV: initial: \n" << BB << "\n");
   2950 
   2951     for (BasicBlock::iterator PI = BB.getFirstInsertionPt(); PI != BB.end();) {
   2952       DenseMap<Value *, Value *>::iterator P = ChosenPairs.find(PI);
   2953       if (P == ChosenPairs.end()) {
   2954         ++PI;
   2955         continue;
   2956       }
   2957 
   2958       if (getDepthFactor(P->first) == 0) {
   2959         // These instructions are not really fused, but are tracked as though
   2960         // they are. Any case in which it would be interesting to fuse them
   2961         // will be taken care of by InstCombine.
   2962         --NumFusedOps;
   2963         ++PI;
   2964         continue;
   2965       }
   2966 
   2967       Instruction *I = cast<Instruction>(P->first),
   2968         *J = cast<Instruction>(P->second);
   2969 
   2970       DEBUG(dbgs() << "BBV: fusing: " << *I <<
   2971              " <-> " << *J << "\n");
   2972 
   2973       // Remove the pair and flipped pair from the list.
   2974       DenseMap<Value *, Value *>::iterator FP = ChosenPairs.find(P->second);
   2975       assert(FP != ChosenPairs.end() && "Flipped pair not found in list");
   2976       ChosenPairs.erase(FP);
   2977       ChosenPairs.erase(P);
   2978 
   2979       if (!canMoveUsesOfIAfterJ(BB, LoadMoveSetPairs, I, J)) {
   2980         DEBUG(dbgs() << "BBV: fusion of: " << *I <<
   2981                " <-> " << *J <<
   2982                " aborted because of non-trivial dependency cycle\n");
   2983         --NumFusedOps;
   2984         ++PI;
   2985         continue;
   2986       }
   2987 
   2988       // If the pair must have the other order, then flip it.
   2989       bool FlipPairOrder = FixedOrderPairs.count(ValuePair(J, I));
   2990       if (!FlipPairOrder && !FixedOrderPairs.count(ValuePair(I, J))) {
   2991         // This pair does not have a fixed order, and so we might want to
   2992         // flip it if that will yield fewer shuffles. We count the number
   2993         // of dependencies connected via swaps, and those directly connected,
   2994         // and flip the order if the number of swaps is greater.
   2995         bool OrigOrder = true;
   2996         DenseMap<ValuePair, std::vector<ValuePair> >::iterator IJ =
   2997           ConnectedPairDeps.find(ValuePair(I, J));
   2998         if (IJ == ConnectedPairDeps.end()) {
   2999           IJ = ConnectedPairDeps.find(ValuePair(J, I));
   3000           OrigOrder = false;
   3001         }
   3002 
   3003         if (IJ != ConnectedPairDeps.end()) {
   3004           unsigned NumDepsDirect = 0, NumDepsSwap = 0;
   3005           for (std::vector<ValuePair>::iterator T = IJ->second.begin(),
   3006                TE = IJ->second.end(); T != TE; ++T) {
   3007             VPPair Q(IJ->first, *T);
   3008             DenseMap<VPPair, unsigned>::iterator R =
   3009               PairConnectionTypes.find(VPPair(Q.second, Q.first));
   3010             assert(R != PairConnectionTypes.end() &&
   3011                    "Cannot find pair connection type");
   3012             if (R->second == PairConnectionDirect)
   3013               ++NumDepsDirect;
   3014             else if (R->second == PairConnectionSwap)
   3015               ++NumDepsSwap;
   3016           }
   3017 
   3018           if (!OrigOrder)
   3019             std::swap(NumDepsDirect, NumDepsSwap);
   3020 
   3021           if (NumDepsSwap > NumDepsDirect) {
   3022             FlipPairOrder = true;
   3023             DEBUG(dbgs() << "BBV: reordering pair: " << *I <<
   3024                             " <-> " << *J << "\n");
   3025           }
   3026         }
   3027       }
   3028 
   3029       Instruction *L = I, *H = J;
   3030       if (FlipPairOrder)
   3031         std::swap(H, L);
   3032 
   3033       // If the pair being fused uses the opposite order from that in the pair
   3034       // connection map, then we need to flip the types.
   3035       DenseMap<ValuePair, std::vector<ValuePair> >::iterator HL =
   3036         ConnectedPairs.find(ValuePair(H, L));
   3037       if (HL != ConnectedPairs.end())
   3038         for (std::vector<ValuePair>::iterator T = HL->second.begin(),
   3039              TE = HL->second.end(); T != TE; ++T) {
   3040           VPPair Q(HL->first, *T);
   3041           DenseMap<VPPair, unsigned>::iterator R = PairConnectionTypes.find(Q);
   3042           assert(R != PairConnectionTypes.end() &&
   3043                  "Cannot find pair connection type");
   3044           if (R->second == PairConnectionDirect)
   3045             R->second = PairConnectionSwap;
   3046           else if (R->second == PairConnectionSwap)
   3047             R->second = PairConnectionDirect;
   3048         }
   3049 
   3050       bool LBeforeH = !FlipPairOrder;
   3051       unsigned NumOperands = I->getNumOperands();
   3052       SmallVector<Value *, 3> ReplacedOperands(NumOperands);
   3053       getReplacementInputsForPair(Context, L, H, ReplacedOperands,
   3054                                   LBeforeH);
   3055 
   3056       // Make a copy of the original operation, change its type to the vector
   3057       // type and replace its operands with the vector operands.
   3058       Instruction *K = L->clone();
   3059       if (L->hasName())
   3060         K->takeName(L);
   3061       else if (H->hasName())
   3062         K->takeName(H);
   3063 
   3064       if (!isa<StoreInst>(K))
   3065         K->mutateType(getVecTypeForPair(L->getType(), H->getType()));
   3066 
   3067       combineMetadata(K, H);
   3068       K->intersectOptionalDataWith(H);
   3069 
   3070       for (unsigned o = 0; o < NumOperands; ++o)
   3071         K->setOperand(o, ReplacedOperands[o]);
   3072 
   3073       K->insertAfter(J);
   3074 
   3075       // Instruction insertion point:
   3076       Instruction *InsertionPt = K;
   3077       Instruction *K1 = 0, *K2 = 0;
   3078       replaceOutputsOfPair(Context, L, H, K, InsertionPt, K1, K2);
   3079 
   3080       // The use dag of the first original instruction must be moved to after
   3081       // the location of the second instruction. The entire use dag of the
   3082       // first instruction is disjoint from the input dag of the second
   3083       // (by definition), and so commutes with it.
   3084 
   3085       moveUsesOfIAfterJ(BB, LoadMoveSetPairs, InsertionPt, I, J);
   3086 
   3087       if (!isa<StoreInst>(I)) {
   3088         L->replaceAllUsesWith(K1);
   3089         H->replaceAllUsesWith(K2);
   3090         AA->replaceWithNewValue(L, K1);
   3091         AA->replaceWithNewValue(H, K2);
   3092       }
   3093 
   3094       // Instructions that may read from memory may be in the load move set.
   3095       // Once an instruction is fused, we no longer need its move set, and so
   3096       // the values of the map never need to be updated. However, when a load
   3097       // is fused, we need to merge the entries from both instructions in the
   3098       // pair in case those instructions were in the move set of some other
   3099       // yet-to-be-fused pair. The loads in question are the keys of the map.
   3100       if (I->mayReadFromMemory()) {
   3101         std::vector<ValuePair> NewSetMembers;
   3102         DenseMap<Value *, std::vector<Value *> >::iterator II =
   3103           LoadMoveSet.find(I);
   3104         if (II != LoadMoveSet.end())
   3105           for (std::vector<Value *>::iterator N = II->second.begin(),
   3106                NE = II->second.end(); N != NE; ++N)
   3107             NewSetMembers.push_back(ValuePair(K, *N));
   3108         DenseMap<Value *, std::vector<Value *> >::iterator JJ =
   3109           LoadMoveSet.find(J);
   3110         if (JJ != LoadMoveSet.end())
   3111           for (std::vector<Value *>::iterator N = JJ->second.begin(),
   3112                NE = JJ->second.end(); N != NE; ++N)
   3113             NewSetMembers.push_back(ValuePair(K, *N));
   3114         for (std::vector<ValuePair>::iterator A = NewSetMembers.begin(),
   3115              AE = NewSetMembers.end(); A != AE; ++A) {
   3116           LoadMoveSet[A->first].push_back(A->second);
   3117           LoadMoveSetPairs.insert(*A);
   3118         }
   3119       }
   3120 
   3121       // Before removing I, set the iterator to the next instruction.
   3122       PI = llvm::next(BasicBlock::iterator(I));
   3123       if (cast<Instruction>(PI) == J)
   3124         ++PI;
   3125 
   3126       SE->forgetValue(I);
   3127       SE->forgetValue(J);
   3128       I->eraseFromParent();
   3129       J->eraseFromParent();
   3130 
   3131       DEBUG(if (PrintAfterEveryPair) dbgs() << "BBV: block is now: \n" <<
   3132                                                BB << "\n");
   3133     }
   3134 
   3135     DEBUG(dbgs() << "BBV: final: \n" << BB << "\n");
   3136   }
   3137 }
   3138 
   3139 char BBVectorize::ID = 0;
   3140 static const char bb_vectorize_name[] = "Basic-Block Vectorization";
   3141 INITIALIZE_PASS_BEGIN(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
   3142 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
   3143 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
   3144 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
   3145 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
   3146 INITIALIZE_PASS_END(BBVectorize, BBV_NAME, bb_vectorize_name, false, false)
   3147 
   3148 BasicBlockPass *llvm::createBBVectorizePass(const VectorizeConfig &C) {
   3149   return new BBVectorize(C);
   3150 }
   3151 
   3152 bool
   3153 llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB, const VectorizeConfig &C) {
   3154   BBVectorize BBVectorizer(P, C);
   3155   return BBVectorizer.vectorizeBB(BB);
   3156 }
   3157 
   3158 //===----------------------------------------------------------------------===//
   3159 VectorizeConfig::VectorizeConfig() {
   3160   VectorBits = ::VectorBits;
   3161   VectorizeBools = !::NoBools;
   3162   VectorizeInts = !::NoInts;
   3163   VectorizeFloats = !::NoFloats;
   3164   VectorizePointers = !::NoPointers;
   3165   VectorizeCasts = !::NoCasts;
   3166   VectorizeMath = !::NoMath;
   3167   VectorizeFMA = !::NoFMA;
   3168   VectorizeSelect = !::NoSelect;
   3169   VectorizeCmp = !::NoCmp;
   3170   VectorizeGEP = !::NoGEP;
   3171   VectorizeMemOps = !::NoMemOps;
   3172   AlignedOnly = ::AlignedOnly;
   3173   ReqChainDepth= ::ReqChainDepth;
   3174   SearchLimit = ::SearchLimit;
   3175   MaxCandPairsForCycleCheck = ::MaxCandPairsForCycleCheck;
   3176   SplatBreaksChain = ::SplatBreaksChain;
   3177   MaxInsts = ::MaxInsts;
   3178   MaxPairs = ::MaxPairs;
   3179   MaxIter = ::MaxIter;
   3180   Pow2LenOnly = ::Pow2LenOnly;
   3181   NoMemOpBoost = ::NoMemOpBoost;
   3182   FastDep = ::FastDep;
   3183 }
   3184