LLVM API Documentation

LoopVectorize.cpp
Go to the documentation of this file.
00001 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
00002 //
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file is distributed under the University of Illinois Open Source
00006 // License. See LICENSE.TXT for details.
00007 //
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
00011 // and generates target-independent LLVM-IR.
00012 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
00013 // of instructions in order to estimate the profitability of vectorization.
00014 //
00015 // The loop vectorizer combines consecutive loop iterations into a single
00016 // 'wide' iteration. After this transformation the index is incremented
00017 // by the SIMD vector width, and not by one.
00018 //
00019 // This pass has three parts:
00020 // 1. The main loop pass that drives the different parts.
00021 // 2. LoopVectorizationLegality - A unit that checks for the legality
00022 //    of the vectorization.
00023 // 3. InnerLoopVectorizer - A unit that performs the actual
00024 //    widening of instructions.
00025 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
00026 //    of vectorization. It decides on the optimal vector width, which
00027 //    can be one, if vectorization is not profitable.
00028 //
00029 //===----------------------------------------------------------------------===//
00030 //
00031 // The reduction-variable vectorization is based on the paper:
00032 //  D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
00033 //
00034 // Variable uniformity checks are inspired by:
00035 //  Karrenberg, R. and Hack, S. Whole Function Vectorization.
00036 //
00037 // Other ideas/concepts are from:
00038 //  A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
00039 //
00040 //  S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua.  An Evaluation of
00041 //  Vectorizing Compilers.
00042 //
00043 //===----------------------------------------------------------------------===//
00044 
00045 #include "llvm/Transforms/Vectorize.h"
00046 #include "llvm/ADT/DenseMap.h"
00047 #include "llvm/ADT/EquivalenceClasses.h"
00048 #include "llvm/ADT/Hashing.h"
00049 #include "llvm/ADT/MapVector.h"
00050 #include "llvm/ADT/SetVector.h"
00051 #include "llvm/ADT/SmallPtrSet.h"
00052 #include "llvm/ADT/SmallSet.h"
00053 #include "llvm/ADT/SmallVector.h"
00054 #include "llvm/ADT/Statistic.h"
00055 #include "llvm/ADT/StringExtras.h"
00056 #include "llvm/Analysis/AliasAnalysis.h"
00057 #include "llvm/Analysis/AliasSetTracker.h"
00058 #include "llvm/Analysis/AssumptionTracker.h"
00059 #include "llvm/Analysis/BlockFrequencyInfo.h"
00060 #include "llvm/Analysis/CodeMetrics.h"
00061 #include "llvm/Analysis/LoopInfo.h"
00062 #include "llvm/Analysis/LoopIterator.h"
00063 #include "llvm/Analysis/LoopPass.h"
00064 #include "llvm/Analysis/ScalarEvolution.h"
00065 #include "llvm/Analysis/ScalarEvolutionExpander.h"
00066 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
00067 #include "llvm/Analysis/TargetTransformInfo.h"
00068 #include "llvm/Analysis/ValueTracking.h"
00069 #include "llvm/IR/Constants.h"
00070 #include "llvm/IR/DataLayout.h"
00071 #include "llvm/IR/DebugInfo.h"
00072 #include "llvm/IR/DerivedTypes.h"
00073 #include "llvm/IR/DiagnosticInfo.h"
00074 #include "llvm/IR/Dominators.h"
00075 #include "llvm/IR/Function.h"
00076 #include "llvm/IR/IRBuilder.h"
00077 #include "llvm/IR/Instructions.h"
00078 #include "llvm/IR/IntrinsicInst.h"
00079 #include "llvm/IR/LLVMContext.h"
00080 #include "llvm/IR/Module.h"
00081 #include "llvm/IR/PatternMatch.h"
00082 #include "llvm/IR/Type.h"
00083 #include "llvm/IR/Value.h"
00084 #include "llvm/IR/ValueHandle.h"
00085 #include "llvm/IR/Verifier.h"
00086 #include "llvm/Pass.h"
00087 #include "llvm/Support/BranchProbability.h"
00088 #include "llvm/Support/CommandLine.h"
00089 #include "llvm/Support/Debug.h"
00090 #include "llvm/Support/raw_ostream.h"
00091 #include "llvm/Transforms/Scalar.h"
00092 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
00093 #include "llvm/Transforms/Utils/Local.h"
00094 #include "llvm/Transforms/Utils/VectorUtils.h"
00095 #include <algorithm>
00096 #include <map>
00097 #include <tuple>
00098 
00099 using namespace llvm;
00100 using namespace llvm::PatternMatch;
00101 
00102 #define LV_NAME "loop-vectorize"
00103 #define DEBUG_TYPE LV_NAME
00104 
00105 STATISTIC(LoopsVectorized, "Number of loops vectorized");
00106 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
00107 
00108 static cl::opt<unsigned>
00109 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
00110                     cl::desc("Sets the SIMD width. Zero is autoselect."));
00111 
00112 static cl::opt<unsigned>
00113 VectorizationInterleave("force-vector-interleave", cl::init(0), cl::Hidden,
00114                     cl::desc("Sets the vectorization interleave count. "
00115                              "Zero is autoselect."));
00116 
00117 static cl::opt<bool>
00118 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
00119                    cl::desc("Enable if-conversion during vectorization."));
00120 
00121 /// We don't vectorize loops with a known constant trip count below this number.
00122 static cl::opt<unsigned>
00123 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
00124                              cl::Hidden,
00125                              cl::desc("Don't vectorize loops with a constant "
00126                                       "trip count that is smaller than this "
00127                                       "value."));
00128 
00129 /// This enables versioning on the strides of symbolically striding memory
00130 /// accesses in code like the following.
00131 ///   for (i = 0; i < N; ++i)
00132 ///     A[i * Stride1] += B[i * Stride2] ...
00133 ///
00134 /// Will be roughly translated to
00135 ///    if (Stride1 == 1 && Stride2 == 1) {
00136 ///      for (i = 0; i < N; i+=4)
00137 ///       A[i:i+3] += ...
00138 ///    } else
00139 ///      ...
00140 static cl::opt<bool> EnableMemAccessVersioning(
00141     "enable-mem-access-versioning", cl::init(true), cl::Hidden,
00142     cl::desc("Enable symblic stride memory access versioning"));
00143 
00144 /// We don't unroll loops with a known constant trip count below this number.
00145 static const unsigned TinyTripCountUnrollThreshold = 128;
00146 
00147 /// When performing memory disambiguation checks at runtime do not make more
00148 /// than this number of comparisons.
00149 static const unsigned RuntimeMemoryCheckThreshold = 8;
00150 
00151 /// Maximum simd width.
00152 static const unsigned MaxVectorWidth = 64;
00153 
00154 static cl::opt<unsigned> ForceTargetNumScalarRegs(
00155     "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
00156     cl::desc("A flag that overrides the target's number of scalar registers."));
00157 
00158 static cl::opt<unsigned> ForceTargetNumVectorRegs(
00159     "force-target-num-vector-regs", cl::init(0), cl::Hidden,
00160     cl::desc("A flag that overrides the target's number of vector registers."));
00161 
00162 /// Maximum vectorization interleave count.
00163 static const unsigned MaxInterleaveFactor = 16;
00164 
00165 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
00166     "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
00167     cl::desc("A flag that overrides the target's max interleave factor for "
00168              "scalar loops."));
00169 
00170 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
00171     "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
00172     cl::desc("A flag that overrides the target's max interleave factor for "
00173              "vectorized loops."));
00174 
00175 static cl::opt<unsigned> ForceTargetInstructionCost(
00176     "force-target-instruction-cost", cl::init(0), cl::Hidden,
00177     cl::desc("A flag that overrides the target's expected cost for "
00178              "an instruction to a single constant value. Mostly "
00179              "useful for getting consistent testing."));
00180 
00181 static cl::opt<unsigned> SmallLoopCost(
00182     "small-loop-cost", cl::init(20), cl::Hidden,
00183     cl::desc("The cost of a loop that is considered 'small' by the unroller."));
00184 
00185 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
00186     "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
00187     cl::desc("Enable the use of the block frequency analysis to access PGO "
00188              "heuristics minimizing code growth in cold regions and being more "
00189              "aggressive in hot regions."));
00190 
00191 // Runtime unroll loops for load/store throughput.
00192 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
00193     "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
00194     cl::desc("Enable runtime unrolling until load/store ports are saturated"));
00195 
00196 /// The number of stores in a loop that are allowed to need predication.
00197 static cl::opt<unsigned> NumberOfStoresToPredicate(
00198     "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
00199     cl::desc("Max number of stores to be predicated behind an if."));
00200 
00201 static cl::opt<bool> EnableIndVarRegisterHeur(
00202     "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
00203     cl::desc("Count the induction variable only once when unrolling"));
00204 
00205 static cl::opt<bool> EnableCondStoresVectorization(
00206     "enable-cond-stores-vec", cl::init(false), cl::Hidden,
00207     cl::desc("Enable if predication of stores during vectorization."));
00208 
00209 static cl::opt<unsigned> MaxNestedScalarReductionUF(
00210     "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
00211     cl::desc("The maximum unroll factor to use when unrolling a scalar "
00212              "reduction in a nested loop."));
00213 
00214 namespace {
00215 
00216 // Forward declarations.
00217 class LoopVectorizationLegality;
00218 class LoopVectorizationCostModel;
00219 class LoopVectorizeHints;
00220 
00221 /// Optimization analysis message produced during vectorization. Messages inform
00222 /// the user why vectorization did not occur.
00223 class Report {
00224   std::string Message;
00225   raw_string_ostream Out;
00226   Instruction *Instr;
00227 
00228 public:
00229   Report(Instruction *I = nullptr) : Out(Message), Instr(I) {
00230     Out << "loop not vectorized: ";
00231   }
00232 
00233   template <typename A> Report &operator<<(const A &Value) {
00234     Out << Value;
00235     return *this;
00236   }
00237 
00238   Instruction *getInstr() { return Instr; }
00239 
00240   std::string &str() { return Out.str(); }
00241   operator Twine() { return Out.str(); }
00242 };
00243 
00244 /// InnerLoopVectorizer vectorizes loops which contain only one basic
00245 /// block to a specified vectorization factor (VF).
00246 /// This class performs the widening of scalars into vectors, or multiple
00247 /// scalars. This class also implements the following features:
00248 /// * It inserts an epilogue loop for handling loops that don't have iteration
00249 ///   counts that are known to be a multiple of the vectorization factor.
00250 /// * It handles the code generation for reduction variables.
00251 /// * Scalarization (implementation using scalars) of un-vectorizable
00252 ///   instructions.
00253 /// InnerLoopVectorizer does not perform any vectorization-legality
00254 /// checks, and relies on the caller to check for the different legality
00255 /// aspects. The InnerLoopVectorizer relies on the
00256 /// LoopVectorizationLegality class to provide information about the induction
00257 /// and reduction variables that were found to a given vectorization factor.
00258 class InnerLoopVectorizer {
00259 public:
00260   InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
00261                       DominatorTree *DT, const DataLayout *DL,
00262                       const TargetLibraryInfo *TLI, unsigned VecWidth,
00263                       unsigned UnrollFactor)
00264       : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
00265         VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
00266         Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
00267         Legal(nullptr) {}
00268 
00269   // Perform the actual loop widening (vectorization).
00270   void vectorize(LoopVectorizationLegality *L) {
00271     Legal = L;
00272     // Create a new empty loop. Unlink the old loop and connect the new one.
00273     createEmptyLoop();
00274     // Widen each instruction in the old loop to a new one in the new loop.
00275     // Use the Legality module to find the induction and reduction variables.
00276     vectorizeLoop();
00277     // Register the new loop and update the analysis passes.
00278     updateAnalysis();
00279   }
00280 
00281   virtual ~InnerLoopVectorizer() {}
00282 
00283 protected:
00284   /// A small list of PHINodes.
00285   typedef SmallVector<PHINode*, 4> PhiVector;
00286   /// When we unroll loops we have multiple vector values for each scalar.
00287   /// This data structure holds the unrolled and vectorized values that
00288   /// originated from one scalar instruction.
00289   typedef SmallVector<Value*, 2> VectorParts;
00290 
00291   // When we if-convert we need create edge masks. We have to cache values so
00292   // that we don't end up with exponential recursion/IR.
00293   typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
00294                    VectorParts> EdgeMaskCache;
00295 
00296   /// \brief Add code that checks at runtime if the accessed arrays overlap.
00297   ///
00298   /// Returns a pair of instructions where the first element is the first
00299   /// instruction generated in possibly a sequence of instructions and the
00300   /// second value is the final comparator value or NULL if no check is needed.
00301   std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
00302 
00303   /// \brief Add checks for strides that where assumed to be 1.
00304   ///
00305   /// Returns the last check instruction and the first check instruction in the
00306   /// pair as (first, last).
00307   std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
00308 
00309   /// Create an empty loop, based on the loop ranges of the old loop.
00310   void createEmptyLoop();
00311   /// Copy and widen the instructions from the old loop.
00312   virtual void vectorizeLoop();
00313 
00314   /// \brief The Loop exit block may have single value PHI nodes where the
00315   /// incoming value is 'Undef'. While vectorizing we only handled real values
00316   /// that were defined inside the loop. Here we fix the 'undef case'.
00317   /// See PR14725.
00318   void fixLCSSAPHIs();
00319 
00320   /// A helper function that computes the predicate of the block BB, assuming
00321   /// that the header block of the loop is set to True. It returns the *entry*
00322   /// mask for the block BB.
00323   VectorParts createBlockInMask(BasicBlock *BB);
00324   /// A helper function that computes the predicate of the edge between SRC
00325   /// and DST.
00326   VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
00327 
00328   /// A helper function to vectorize a single BB within the innermost loop.
00329   void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
00330 
00331   /// Vectorize a single PHINode in a block. This method handles the induction
00332   /// variable canonicalization. It supports both VF = 1 for unrolled loops and
00333   /// arbitrary length vectors.
00334   void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
00335                            unsigned UF, unsigned VF, PhiVector *PV);
00336 
00337   /// Insert the new loop to the loop hierarchy and pass manager
00338   /// and update the analysis passes.
00339   void updateAnalysis();
00340 
00341   /// This instruction is un-vectorizable. Implement it as a sequence
00342   /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
00343   /// scalarized instruction behind an if block predicated on the control
00344   /// dependence of the instruction.
00345   virtual void scalarizeInstruction(Instruction *Instr,
00346                                     bool IfPredicateStore=false);
00347 
00348   /// Vectorize Load and Store instructions,
00349   virtual void vectorizeMemoryInstruction(Instruction *Instr);
00350 
00351   /// Create a broadcast instruction. This method generates a broadcast
00352   /// instruction (shuffle) for loop invariant values and for the induction
00353   /// value. If this is the induction variable then we extend it to N, N+1, ...
00354   /// this is needed because each iteration in the loop corresponds to a SIMD
00355   /// element.
00356   virtual Value *getBroadcastInstrs(Value *V);
00357 
00358   /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
00359   /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
00360   /// The sequence starts at StartIndex.
00361   virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
00362 
00363   /// When we go over instructions in the basic block we rely on previous
00364   /// values within the current basic block or on loop invariant values.
00365   /// When we widen (vectorize) values we place them in the map. If the values
00366   /// are not within the map, they have to be loop invariant, so we simply
00367   /// broadcast them into a vector.
00368   VectorParts &getVectorValue(Value *V);
00369 
00370   /// Generate a shuffle sequence that will reverse the vector Vec.
00371   virtual Value *reverseVector(Value *Vec);
00372 
00373   /// This is a helper class that holds the vectorizer state. It maps scalar
00374   /// instructions to vector instructions. When the code is 'unrolled' then
00375   /// then a single scalar value is mapped to multiple vector parts. The parts
00376   /// are stored in the VectorPart type.
00377   struct ValueMap {
00378     /// C'tor.  UnrollFactor controls the number of vectors ('parts') that
00379     /// are mapped.
00380     ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
00381 
00382     /// \return True if 'Key' is saved in the Value Map.
00383     bool has(Value *Key) const { return MapStorage.count(Key); }
00384 
00385     /// Initializes a new entry in the map. Sets all of the vector parts to the
00386     /// save value in 'Val'.
00387     /// \return A reference to a vector with splat values.
00388     VectorParts &splat(Value *Key, Value *Val) {
00389       VectorParts &Entry = MapStorage[Key];
00390       Entry.assign(UF, Val);
00391       return Entry;
00392     }
00393 
00394     ///\return A reference to the value that is stored at 'Key'.
00395     VectorParts &get(Value *Key) {
00396       VectorParts &Entry = MapStorage[Key];
00397       if (Entry.empty())
00398         Entry.resize(UF);
00399       assert(Entry.size() == UF);
00400       return Entry;
00401     }
00402 
00403   private:
00404     /// The unroll factor. Each entry in the map stores this number of vector
00405     /// elements.
00406     unsigned UF;
00407 
00408     /// Map storage. We use std::map and not DenseMap because insertions to a
00409     /// dense map invalidates its iterators.
00410     std::map<Value *, VectorParts> MapStorage;
00411   };
00412 
00413   /// The original loop.
00414   Loop *OrigLoop;
00415   /// Scev analysis to use.
00416   ScalarEvolution *SE;
00417   /// Loop Info.
00418   LoopInfo *LI;
00419   /// Dominator Tree.
00420   DominatorTree *DT;
00421   /// Alias Analysis.
00422   AliasAnalysis *AA;
00423   /// Data Layout.
00424   const DataLayout *DL;
00425   /// Target Library Info.
00426   const TargetLibraryInfo *TLI;
00427 
00428   /// The vectorization SIMD factor to use. Each vector will have this many
00429   /// vector elements.
00430   unsigned VF;
00431 
00432 protected:
00433   /// The vectorization unroll factor to use. Each scalar is vectorized to this
00434   /// many different vector instructions.
00435   unsigned UF;
00436 
00437   /// The builder that we use
00438   IRBuilder<> Builder;
00439 
00440   // --- Vectorization state ---
00441 
00442   /// The vector-loop preheader.
00443   BasicBlock *LoopVectorPreHeader;
00444   /// The scalar-loop preheader.
00445   BasicBlock *LoopScalarPreHeader;
00446   /// Middle Block between the vector and the scalar.
00447   BasicBlock *LoopMiddleBlock;
00448   ///The ExitBlock of the scalar loop.
00449   BasicBlock *LoopExitBlock;
00450   ///The vector loop body.
00451   SmallVector<BasicBlock *, 4> LoopVectorBody;
00452   ///The scalar loop body.
00453   BasicBlock *LoopScalarBody;
00454   /// A list of all bypass blocks. The first block is the entry of the loop.
00455   SmallVector<BasicBlock *, 4> LoopBypassBlocks;
00456 
00457   /// The new Induction variable which was added to the new block.
00458   PHINode *Induction;
00459   /// The induction variable of the old basic block.
00460   PHINode *OldInduction;
00461   /// Holds the extended (to the widest induction type) start index.
00462   Value *ExtendedIdx;
00463   /// Maps scalars to widened vectors.
00464   ValueMap WidenMap;
00465   EdgeMaskCache MaskCache;
00466 
00467   LoopVectorizationLegality *Legal;
00468 };
00469 
00470 class InnerLoopUnroller : public InnerLoopVectorizer {
00471 public:
00472   InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
00473                     DominatorTree *DT, const DataLayout *DL,
00474                     const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
00475     InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
00476 
00477 private:
00478   void scalarizeInstruction(Instruction *Instr,
00479                             bool IfPredicateStore = false) override;
00480   void vectorizeMemoryInstruction(Instruction *Instr) override;
00481   Value *getBroadcastInstrs(Value *V) override;
00482   Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
00483   Value *reverseVector(Value *Vec) override;
00484 };
00485 
00486 /// \brief Look for a meaningful debug location on the instruction or it's
00487 /// operands.
00488 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
00489   if (!I)
00490     return I;
00491 
00492   DebugLoc Empty;
00493   if (I->getDebugLoc() != Empty)
00494     return I;
00495 
00496   for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
00497     if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
00498       if (OpInst->getDebugLoc() != Empty)
00499         return OpInst;
00500   }
00501 
00502   return I;
00503 }
00504 
00505 /// \brief Set the debug location in the builder using the debug location in the
00506 /// instruction.
00507 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
00508   if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
00509     B.SetCurrentDebugLocation(Inst->getDebugLoc());
00510   else
00511     B.SetCurrentDebugLocation(DebugLoc());
00512 }
00513 
00514 #ifndef NDEBUG
00515 /// \return string containing a file name and a line # for the given loop.
00516 static std::string getDebugLocString(const Loop *L) {
00517   std::string Result;
00518   if (L) {
00519     raw_string_ostream OS(Result);
00520     const DebugLoc LoopDbgLoc = L->getStartLoc();
00521     if (!LoopDbgLoc.isUnknown())
00522       LoopDbgLoc.print(L->getHeader()->getContext(), OS);
00523     else
00524       // Just print the module name.
00525       OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
00526     OS.flush();
00527   }
00528   return Result;
00529 }
00530 #endif
00531 
00532 /// \brief Propagate known metadata from one instruction to another.
00533 static void propagateMetadata(Instruction *To, const Instruction *From) {
00534   SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
00535   From->getAllMetadataOtherThanDebugLoc(Metadata);
00536 
00537   for (auto M : Metadata) {
00538     unsigned Kind = M.first;
00539 
00540     // These are safe to transfer (this is safe for TBAA, even when we
00541     // if-convert, because should that metadata have had a control dependency
00542     // on the condition, and thus actually aliased with some other
00543     // non-speculated memory access when the condition was false, this would be
00544     // caught by the runtime overlap checks).
00545     if (Kind != LLVMContext::MD_tbaa &&
00546         Kind != LLVMContext::MD_alias_scope &&
00547         Kind != LLVMContext::MD_noalias &&
00548         Kind != LLVMContext::MD_fpmath)
00549       continue;
00550 
00551     To->setMetadata(Kind, M.second);
00552   }
00553 }
00554 
00555 /// \brief Propagate known metadata from one instruction to a vector of others.
00556 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
00557   for (Value *V : To)
00558     if (Instruction *I = dyn_cast<Instruction>(V))
00559       propagateMetadata(I, From);
00560 }
00561 
00562 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
00563 /// to what vectorization factor.
00564 /// This class does not look at the profitability of vectorization, only the
00565 /// legality. This class has two main kinds of checks:
00566 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
00567 ///   will change the order of memory accesses in a way that will change the
00568 ///   correctness of the program.
00569 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
00570 /// checks for a number of different conditions, such as the availability of a
00571 /// single induction variable, that all types are supported and vectorize-able,
00572 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
00573 /// This class is also used by InnerLoopVectorizer for identifying
00574 /// induction variable and the different reduction variables.
00575 class LoopVectorizationLegality {
00576 public:
00577   unsigned NumLoads;
00578   unsigned NumStores;
00579   unsigned NumPredStores;
00580 
00581   LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
00582                             DominatorTree *DT, TargetLibraryInfo *TLI,
00583                             AliasAnalysis *AA, Function *F)
00584       : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
00585         DT(DT), TLI(TLI), AA(AA), TheFunction(F), Induction(nullptr),
00586         WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
00587   }
00588 
00589   /// This enum represents the kinds of reductions that we support.
00590   enum ReductionKind {
00591     RK_NoReduction, ///< Not a reduction.
00592     RK_IntegerAdd,  ///< Sum of integers.
00593     RK_IntegerMult, ///< Product of integers.
00594     RK_IntegerOr,   ///< Bitwise or logical OR of numbers.
00595     RK_IntegerAnd,  ///< Bitwise or logical AND of numbers.
00596     RK_IntegerXor,  ///< Bitwise or logical XOR of numbers.
00597     RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
00598     RK_FloatAdd,    ///< Sum of floats.
00599     RK_FloatMult,   ///< Product of floats.
00600     RK_FloatMinMax  ///< Min/max implemented in terms of select(cmp()).
00601   };
00602 
00603   /// This enum represents the kinds of inductions that we support.
00604   enum InductionKind {
00605     IK_NoInduction,         ///< Not an induction variable.
00606     IK_IntInduction,        ///< Integer induction variable. Step = 1.
00607     IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
00608     IK_PtrInduction,        ///< Pointer induction var. Step = sizeof(elem).
00609     IK_ReversePtrInduction  ///< Reverse ptr indvar. Step = - sizeof(elem).
00610   };
00611 
00612   // This enum represents the kind of minmax reduction.
00613   enum MinMaxReductionKind {
00614     MRK_Invalid,
00615     MRK_UIntMin,
00616     MRK_UIntMax,
00617     MRK_SIntMin,
00618     MRK_SIntMax,
00619     MRK_FloatMin,
00620     MRK_FloatMax
00621   };
00622 
00623   /// This struct holds information about reduction variables.
00624   struct ReductionDescriptor {
00625     ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
00626       Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
00627 
00628     ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
00629                         MinMaxReductionKind MK)
00630         : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
00631 
00632     // The starting value of the reduction.
00633     // It does not have to be zero!
00634     TrackingVH<Value> StartValue;
00635     // The instruction who's value is used outside the loop.
00636     Instruction *LoopExitInstr;
00637     // The kind of the reduction.
00638     ReductionKind Kind;
00639     // If this a min/max reduction the kind of reduction.
00640     MinMaxReductionKind MinMaxKind;
00641   };
00642 
00643   /// This POD struct holds information about a potential reduction operation.
00644   struct ReductionInstDesc {
00645     ReductionInstDesc(bool IsRedux, Instruction *I) :
00646       IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
00647 
00648     ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
00649       IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
00650 
00651     // Is this instruction a reduction candidate.
00652     bool IsReduction;
00653     // The last instruction in a min/max pattern (select of the select(icmp())
00654     // pattern), or the current reduction instruction otherwise.
00655     Instruction *PatternLastInst;
00656     // If this is a min/max pattern the comparison predicate.
00657     MinMaxReductionKind MinMaxKind;
00658   };
00659 
00660   /// This struct holds information about the memory runtime legality
00661   /// check that a group of pointers do not overlap.
00662   struct RuntimePointerCheck {
00663     RuntimePointerCheck() : Need(false) {}
00664 
00665     /// Reset the state of the pointer runtime information.
00666     void reset() {
00667       Need = false;
00668       Pointers.clear();
00669       Starts.clear();
00670       Ends.clear();
00671       IsWritePtr.clear();
00672       DependencySetId.clear();
00673       AliasSetId.clear();
00674     }
00675 
00676     /// Insert a pointer and calculate the start and end SCEVs.
00677     void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
00678                 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
00679 
00680     /// This flag indicates if we need to add the runtime check.
00681     bool Need;
00682     /// Holds the pointers that we need to check.
00683     SmallVector<TrackingVH<Value>, 2> Pointers;
00684     /// Holds the pointer value at the beginning of the loop.
00685     SmallVector<const SCEV*, 2> Starts;
00686     /// Holds the pointer value at the end of the loop.
00687     SmallVector<const SCEV*, 2> Ends;
00688     /// Holds the information if this pointer is used for writing to memory.
00689     SmallVector<bool, 2> IsWritePtr;
00690     /// Holds the id of the set of pointers that could be dependent because of a
00691     /// shared underlying object.
00692     SmallVector<unsigned, 2> DependencySetId;
00693     /// Holds the id of the disjoint alias set to which this pointer belongs.
00694     SmallVector<unsigned, 2> AliasSetId;
00695   };
00696 
00697   /// A struct for saving information about induction variables.
00698   struct InductionInfo {
00699     InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
00700     InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
00701     /// Start value.
00702     TrackingVH<Value> StartValue;
00703     /// Induction kind.
00704     InductionKind IK;
00705   };
00706 
00707   /// ReductionList contains the reduction descriptors for all
00708   /// of the reductions that were found in the loop.
00709   typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
00710 
00711   /// InductionList saves induction variables and maps them to the
00712   /// induction descriptor.
00713   typedef MapVector<PHINode*, InductionInfo> InductionList;
00714 
00715   /// Returns true if it is legal to vectorize this loop.
00716   /// This does not mean that it is profitable to vectorize this
00717   /// loop, only that it is legal to do so.
00718   bool canVectorize();
00719 
00720   /// Returns the Induction variable.
00721   PHINode *getInduction() { return Induction; }
00722 
00723   /// Returns the reduction variables found in the loop.
00724   ReductionList *getReductionVars() { return &Reductions; }
00725 
00726   /// Returns the induction variables found in the loop.
00727   InductionList *getInductionVars() { return &Inductions; }
00728 
00729   /// Returns the widest induction type.
00730   Type *getWidestInductionType() { return WidestIndTy; }
00731 
00732   /// Returns True if V is an induction variable in this loop.
00733   bool isInductionVariable(const Value *V);
00734 
00735   /// Return true if the block BB needs to be predicated in order for the loop
00736   /// to be vectorized.
00737   bool blockNeedsPredication(BasicBlock *BB);
00738 
00739   /// Check if this  pointer is consecutive when vectorizing. This happens
00740   /// when the last index of the GEP is the induction variable, or that the
00741   /// pointer itself is an induction variable.
00742   /// This check allows us to vectorize A[idx] into a wide load/store.
00743   /// Returns:
00744   /// 0 - Stride is unknown or non-consecutive.
00745   /// 1 - Address is consecutive.
00746   /// -1 - Address is consecutive, and decreasing.
00747   int isConsecutivePtr(Value *Ptr);
00748 
00749   /// Returns true if the value V is uniform within the loop.
00750   bool isUniform(Value *V);
00751 
00752   /// Returns true if this instruction will remain scalar after vectorization.
00753   bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
00754 
00755   /// Returns the information that we collected about runtime memory check.
00756   RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
00757 
00758   /// This function returns the identity element (or neutral element) for
00759   /// the operation K.
00760   static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
00761 
00762   unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
00763 
00764   bool hasStride(Value *V) { return StrideSet.count(V); }
00765   bool mustCheckStrides() { return !StrideSet.empty(); }
00766   SmallPtrSet<Value *, 8>::iterator strides_begin() {
00767     return StrideSet.begin();
00768   }
00769   SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
00770 
00771 private:
00772   /// Check if a single basic block loop is vectorizable.
00773   /// At this point we know that this is a loop with a constant trip count
00774   /// and we only need to check individual instructions.
00775   bool canVectorizeInstrs();
00776 
00777   /// When we vectorize loops we may change the order in which
00778   /// we read and write from memory. This method checks if it is
00779   /// legal to vectorize the code, considering only memory constrains.
00780   /// Returns true if the loop is vectorizable
00781   bool canVectorizeMemory();
00782 
00783   /// Return true if we can vectorize this loop using the IF-conversion
00784   /// transformation.
00785   bool canVectorizeWithIfConvert();
00786 
00787   /// Collect the variables that need to stay uniform after vectorization.
00788   void collectLoopUniforms();
00789 
00790   /// Return true if all of the instructions in the block can be speculatively
00791   /// executed. \p SafePtrs is a list of addresses that are known to be legal
00792   /// and we know that we can read from them without segfault.
00793   bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
00794 
00795   /// Returns True, if 'Phi' is the kind of reduction variable for type
00796   /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
00797   bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
00798   /// Returns a struct describing if the instruction 'I' can be a reduction
00799   /// variable of type 'Kind'. If the reduction is a min/max pattern of
00800   /// select(icmp()) this function advances the instruction pointer 'I' from the
00801   /// compare instruction to the select instruction and stores this pointer in
00802   /// 'PatternLastInst' member of the returned struct.
00803   ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
00804                                      ReductionInstDesc &Desc);
00805   /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
00806   /// pattern corresponding to a min(X, Y) or max(X, Y).
00807   static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
00808                                                     ReductionInstDesc &Prev);
00809   /// Returns the induction kind of Phi. This function may return NoInduction
00810   /// if the PHI is not an induction variable.
00811   InductionKind isInductionVariable(PHINode *Phi);
00812 
00813   /// \brief Collect memory access with loop invariant strides.
00814   ///
00815   /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
00816   /// invariant.
00817   void collectStridedAcccess(Value *LoadOrStoreInst);
00818 
00819   /// Report an analysis message to assist the user in diagnosing loops that are
00820   /// not vectorized.
00821   void emitAnalysis(Report &Message) {
00822     DebugLoc DL = TheLoop->getStartLoc();
00823     if (Instruction *I = Message.getInstr())
00824       DL = I->getDebugLoc();
00825     emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
00826                                    *TheFunction, DL, Message.str());
00827   }
00828 
00829   /// The loop that we evaluate.
00830   Loop *TheLoop;
00831   /// Scev analysis.
00832   ScalarEvolution *SE;
00833   /// DataLayout analysis.
00834   const DataLayout *DL;
00835   /// Dominators.
00836   DominatorTree *DT;
00837   /// Target Library Info.
00838   TargetLibraryInfo *TLI;
00839   /// Alias analysis.
00840   AliasAnalysis *AA;
00841   /// Parent function
00842   Function *TheFunction;
00843 
00844   //  ---  vectorization state --- //
00845 
00846   /// Holds the integer induction variable. This is the counter of the
00847   /// loop.
00848   PHINode *Induction;
00849   /// Holds the reduction variables.
00850   ReductionList Reductions;
00851   /// Holds all of the induction variables that we found in the loop.
00852   /// Notice that inductions don't need to start at zero and that induction
00853   /// variables can be pointers.
00854   InductionList Inductions;
00855   /// Holds the widest induction type encountered.
00856   Type *WidestIndTy;
00857 
00858   /// Allowed outside users. This holds the reduction
00859   /// vars which can be accessed from outside the loop.
00860   SmallPtrSet<Value*, 4> AllowedExit;
00861   /// This set holds the variables which are known to be uniform after
00862   /// vectorization.
00863   SmallPtrSet<Instruction*, 4> Uniforms;
00864   /// We need to check that all of the pointers in this list are disjoint
00865   /// at runtime.
00866   RuntimePointerCheck PtrRtCheck;
00867   /// Can we assume the absence of NaNs.
00868   bool HasFunNoNaNAttr;
00869 
00870   unsigned MaxSafeDepDistBytes;
00871 
00872   ValueToValueMap Strides;
00873   SmallPtrSet<Value *, 8> StrideSet;
00874 };
00875 
00876 /// LoopVectorizationCostModel - estimates the expected speedups due to
00877 /// vectorization.
00878 /// In many cases vectorization is not profitable. This can happen because of
00879 /// a number of reasons. In this class we mainly attempt to predict the
00880 /// expected speedup/slowdowns due to the supported instruction set. We use the
00881 /// TargetTransformInfo to query the different backends for the cost of
00882 /// different operations.
00883 class LoopVectorizationCostModel {
00884 public:
00885   LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
00886                              LoopVectorizationLegality *Legal,
00887                              const TargetTransformInfo &TTI,
00888                              const DataLayout *DL, const TargetLibraryInfo *TLI,
00889                              AssumptionTracker *AT, const Function *F,
00890                              const LoopVectorizeHints *Hints)
00891       : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
00892         TheFunction(F), Hints(Hints) {
00893     CodeMetrics::collectEphemeralValues(L, AT, EphValues);
00894   }
00895 
00896   /// Information about vectorization costs
00897   struct VectorizationFactor {
00898     unsigned Width; // Vector width with best cost
00899     unsigned Cost; // Cost of the loop with that width
00900   };
00901   /// \return The most profitable vectorization factor and the cost of that VF.
00902   /// This method checks every power of two up to VF. If UserVF is not ZERO
00903   /// then this vectorization factor will be selected if vectorization is
00904   /// possible.
00905   VectorizationFactor selectVectorizationFactor(bool OptForSize);
00906 
00907   /// \return The size (in bits) of the widest type in the code that
00908   /// needs to be vectorized. We ignore values that remain scalar such as
00909   /// 64 bit loop indices.
00910   unsigned getWidestType();
00911 
00912   /// \return The most profitable unroll factor.
00913   /// If UserUF is non-zero then this method finds the best unroll-factor
00914   /// based on register pressure and other parameters.
00915   /// VF and LoopCost are the selected vectorization factor and the cost of the
00916   /// selected VF.
00917   unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
00918 
00919   /// \brief A struct that represents some properties of the register usage
00920   /// of a loop.
00921   struct RegisterUsage {
00922     /// Holds the number of loop invariant values that are used in the loop.
00923     unsigned LoopInvariantRegs;
00924     /// Holds the maximum number of concurrent live intervals in the loop.
00925     unsigned MaxLocalUsers;
00926     /// Holds the number of instructions in the loop.
00927     unsigned NumInstructions;
00928   };
00929 
00930   /// \return  information about the register usage of the loop.
00931   RegisterUsage calculateRegisterUsage();
00932 
00933 private:
00934   /// Returns the expected execution cost. The unit of the cost does
00935   /// not matter because we use the 'cost' units to compare different
00936   /// vector widths. The cost that is returned is *not* normalized by
00937   /// the factor width.
00938   unsigned expectedCost(unsigned VF);
00939 
00940   /// Returns the execution time cost of an instruction for a given vector
00941   /// width. Vector width of one means scalar.
00942   unsigned getInstructionCost(Instruction *I, unsigned VF);
00943 
00944   /// A helper function for converting Scalar types to vector types.
00945   /// If the incoming type is void, we return void. If the VF is 1, we return
00946   /// the scalar type.
00947   static Type* ToVectorTy(Type *Scalar, unsigned VF);
00948 
00949   /// Returns whether the instruction is a load or store and will be a emitted
00950   /// as a vector operation.
00951   bool isConsecutiveLoadOrStore(Instruction *I);
00952 
00953   /// Report an analysis message to assist the user in diagnosing loops that are
00954   /// not vectorized.
00955   void emitAnalysis(Report &Message) {
00956     DebugLoc DL = TheLoop->getStartLoc();
00957     if (Instruction *I = Message.getInstr())
00958       DL = I->getDebugLoc();
00959     emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
00960                                    *TheFunction, DL, Message.str());
00961   }
00962 
00963   /// Values used only by @llvm.assume calls.
00964   SmallPtrSet<const Value *, 32> EphValues;
00965 
00966   /// The loop that we evaluate.
00967   Loop *TheLoop;
00968   /// Scev analysis.
00969   ScalarEvolution *SE;
00970   /// Loop Info analysis.
00971   LoopInfo *LI;
00972   /// Vectorization legality.
00973   LoopVectorizationLegality *Legal;
00974   /// Vector target information.
00975   const TargetTransformInfo &TTI;
00976   /// Target data layout information.
00977   const DataLayout *DL;
00978   /// Target Library Info.
00979   const TargetLibraryInfo *TLI;
00980   const Function *TheFunction;
00981   // Loop Vectorize Hint.
00982   const LoopVectorizeHints *Hints;
00983 };
00984 
00985 /// Utility class for getting and setting loop vectorizer hints in the form
00986 /// of loop metadata.
00987 /// This class keeps a number of loop annotations locally (as member variables)
00988 /// and can, upon request, write them back as metadata on the loop. It will
00989 /// initially scan the loop for existing metadata, and will update the local
00990 /// values based on information in the loop.
00991 /// We cannot write all values to metadata, as the mere presence of some info,
00992 /// for example 'force', means a decision has been made. So, we need to be
00993 /// careful NOT to add them if the user hasn't specifically asked so.
00994 class LoopVectorizeHints {
00995   enum HintKind {
00996     HK_WIDTH,
00997     HK_UNROLL,
00998     HK_FORCE
00999   };
01000 
01001   /// Hint - associates name and validation with the hint value.
01002   struct Hint {
01003     const char * Name;
01004     unsigned Value; // This may have to change for non-numeric values.
01005     HintKind Kind;
01006 
01007     Hint(const char * Name, unsigned Value, HintKind Kind)
01008       : Name(Name), Value(Value), Kind(Kind) { }
01009 
01010     bool validate(unsigned Val) {
01011       switch (Kind) {
01012       case HK_WIDTH:
01013         return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
01014       case HK_UNROLL:
01015         return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
01016       case HK_FORCE:
01017         return (Val <= 1);
01018       }
01019       return false;
01020     }
01021   };
01022 
01023   /// Vectorization width.
01024   Hint Width;
01025   /// Vectorization interleave factor.
01026   Hint Interleave;
01027   /// Vectorization forced
01028   Hint Force;
01029 
01030   /// Return the loop metadata prefix.
01031   static StringRef Prefix() { return "llvm.loop."; }
01032 
01033 public:
01034   enum ForceKind {
01035     FK_Undefined = -1, ///< Not selected.
01036     FK_Disabled = 0,   ///< Forcing disabled.
01037     FK_Enabled = 1,    ///< Forcing enabled.
01038   };
01039 
01040   LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
01041       : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
01042         Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
01043         Force("vectorize.enable", FK_Undefined, HK_FORCE),
01044         TheLoop(L) {
01045     // Populate values with existing loop metadata.
01046     getHintsFromMetadata();
01047 
01048     // force-vector-interleave overrides DisableInterleaving.
01049     if (VectorizationInterleave.getNumOccurrences() > 0)
01050       Interleave.Value = VectorizationInterleave;
01051 
01052     DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
01053           << "LV: Interleaving disabled by the pass manager\n");
01054   }
01055 
01056   /// Mark the loop L as already vectorized by setting the width to 1.
01057   void setAlreadyVectorized() {
01058     Width.Value = Interleave.Value = 1;
01059     Hint Hints[] = {Width, Interleave};
01060     writeHintsToMetadata(Hints);
01061   }
01062 
01063   /// Dumps all the hint information.
01064   std::string emitRemark() const {
01065     Report R;
01066     if (Force.Value == LoopVectorizeHints::FK_Disabled)
01067       R << "vectorization is explicitly disabled";
01068     else {
01069       R << "use -Rpass-analysis=loop-vectorize for more info";
01070       if (Force.Value == LoopVectorizeHints::FK_Enabled) {
01071         R << " (Force=true";
01072         if (Width.Value != 0)
01073           R << ", Vector Width=" << Width.Value;
01074         if (Interleave.Value != 0)
01075           R << ", Interleave Count=" << Interleave.Value;
01076         R << ")";
01077       }
01078     }
01079 
01080     return R.str();
01081   }
01082 
01083   unsigned getWidth() const { return Width.Value; }
01084   unsigned getInterleave() const { return Interleave.Value; }
01085   enum ForceKind getForce() const { return (ForceKind)Force.Value; }
01086 
01087 private:
01088   /// Find hints specified in the loop metadata and update local values.
01089   void getHintsFromMetadata() {
01090     MDNode *LoopID = TheLoop->getLoopID();
01091     if (!LoopID)
01092       return;
01093 
01094     // First operand should refer to the loop id itself.
01095     assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
01096     assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
01097 
01098     for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
01099       const MDString *S = nullptr;
01100       SmallVector<Value*, 4> Args;
01101 
01102       // The expected hint is either a MDString or a MDNode with the first
01103       // operand a MDString.
01104       if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
01105         if (!MD || MD->getNumOperands() == 0)
01106           continue;
01107         S = dyn_cast<MDString>(MD->getOperand(0));
01108         for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
01109           Args.push_back(MD->getOperand(i));
01110       } else {
01111         S = dyn_cast<MDString>(LoopID->getOperand(i));
01112         assert(Args.size() == 0 && "too many arguments for MDString");
01113       }
01114 
01115       if (!S)
01116         continue;
01117 
01118       // Check if the hint starts with the loop metadata prefix.
01119       StringRef Name = S->getString();
01120       if (Args.size() == 1)
01121         setHint(Name, Args[0]);
01122     }
01123   }
01124 
01125   /// Checks string hint with one operand and set value if valid.
01126   void setHint(StringRef Name, Value *Arg) {
01127     if (!Name.startswith(Prefix()))
01128       return;
01129     Name = Name.substr(Prefix().size(), StringRef::npos);
01130 
01131     const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
01132     if (!C) return;
01133     unsigned Val = C->getZExtValue();
01134 
01135     Hint *Hints[] = {&Width, &Interleave, &Force};
01136     for (auto H : Hints) {
01137       if (Name == H->Name) {
01138         if (H->validate(Val))
01139           H->Value = Val;
01140         else
01141           DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
01142         break;
01143       }
01144     }
01145   }
01146 
01147   /// Create a new hint from name / value pair.
01148   MDNode *createHintMetadata(StringRef Name, unsigned V) const {
01149     LLVMContext &Context = TheLoop->getHeader()->getContext();
01150     Value *Vals[] = {MDString::get(Context, Name),
01151                      ConstantInt::get(Type::getInt32Ty(Context), V)};
01152     return MDNode::get(Context, Vals);
01153   }
01154 
01155   /// Matches metadata with hint name.
01156   bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
01157     MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
01158     if (!Name)
01159       return false;
01160 
01161     for (auto H : HintTypes)
01162       if (Name->getName().endswith(H.Name))
01163         return true;
01164     return false;
01165   }
01166 
01167   /// Sets current hints into loop metadata, keeping other values intact.
01168   void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
01169     if (HintTypes.size() == 0)
01170       return;
01171 
01172     // Reserve the first element to LoopID (see below).
01173     SmallVector<Value*, 4> Vals(1);
01174     // If the loop already has metadata, then ignore the existing operands.
01175     MDNode *LoopID = TheLoop->getLoopID();
01176     if (LoopID) {
01177       for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
01178         MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
01179         // If node in update list, ignore old value.
01180         if (!matchesHintMetadataName(Node, HintTypes))
01181           Vals.push_back(Node);
01182       }
01183     }
01184 
01185     // Now, add the missing hints.
01186     for (auto H : HintTypes)
01187       Vals.push_back(
01188           createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
01189 
01190     // Replace current metadata node with new one.
01191     LLVMContext &Context = TheLoop->getHeader()->getContext();
01192     MDNode *NewLoopID = MDNode::get(Context, Vals);
01193     // Set operand 0 to refer to the loop id itself.
01194     NewLoopID->replaceOperandWith(0, NewLoopID);
01195 
01196     TheLoop->setLoopID(NewLoopID);
01197     if (LoopID)
01198       LoopID->replaceAllUsesWith(NewLoopID);
01199     LoopID = NewLoopID;
01200   }
01201 
01202   /// The loop these hints belong to.
01203   const Loop *TheLoop;
01204 };
01205 
01206 static void emitMissedWarning(Function *F, Loop *L,
01207                               const LoopVectorizeHints &LH) {
01208   emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
01209                                L->getStartLoc(), LH.emitRemark());
01210 
01211   if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
01212     if (LH.getWidth() != 1)
01213       emitLoopVectorizeWarning(
01214           F->getContext(), *F, L->getStartLoc(),
01215           "failed explicitly specified loop vectorization");
01216     else if (LH.getInterleave() != 1)
01217       emitLoopInterleaveWarning(
01218           F->getContext(), *F, L->getStartLoc(),
01219           "failed explicitly specified loop interleaving");
01220   }
01221 }
01222 
01223 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
01224   if (L.empty())
01225     return V.push_back(&L);
01226 
01227   for (Loop *InnerL : L)
01228     addInnerLoop(*InnerL, V);
01229 }
01230 
01231 /// The LoopVectorize Pass.
01232 struct LoopVectorize : public FunctionPass {
01233   /// Pass identification, replacement for typeid
01234   static char ID;
01235 
01236   explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
01237     : FunctionPass(ID),
01238       DisableUnrolling(NoUnrolling),
01239       AlwaysVectorize(AlwaysVectorize) {
01240     initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
01241   }
01242 
01243   ScalarEvolution *SE;
01244   const DataLayout *DL;
01245   LoopInfo *LI;
01246   TargetTransformInfo *TTI;
01247   DominatorTree *DT;
01248   BlockFrequencyInfo *BFI;
01249   TargetLibraryInfo *TLI;
01250   AliasAnalysis *AA;
01251   AssumptionTracker *AT;
01252   bool DisableUnrolling;
01253   bool AlwaysVectorize;
01254 
01255   BlockFrequency ColdEntryFreq;
01256 
01257   bool runOnFunction(Function &F) override {
01258     SE = &getAnalysis<ScalarEvolution>();
01259     DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
01260     DL = DLP ? &DLP->getDataLayout() : nullptr;
01261     LI = &getAnalysis<LoopInfo>();
01262     TTI = &getAnalysis<TargetTransformInfo>();
01263     DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
01264     BFI = &getAnalysis<BlockFrequencyInfo>();
01265     TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
01266     AA = &getAnalysis<AliasAnalysis>();
01267     AT = &getAnalysis<AssumptionTracker>();
01268 
01269     // Compute some weights outside of the loop over the loops. Compute this
01270     // using a BranchProbability to re-use its scaling math.
01271     const BranchProbability ColdProb(1, 5); // 20%
01272     ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
01273 
01274     // If the target claims to have no vector registers don't attempt
01275     // vectorization.
01276     if (!TTI->getNumberOfRegisters(true))
01277       return false;
01278 
01279     if (!DL) {
01280       DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
01281                    << ": Missing data layout\n");
01282       return false;
01283     }
01284 
01285     // Build up a worklist of inner-loops to vectorize. This is necessary as
01286     // the act of vectorizing or partially unrolling a loop creates new loops
01287     // and can invalidate iterators across the loops.
01288     SmallVector<Loop *, 8> Worklist;
01289 
01290     for (Loop *L : *LI)
01291       addInnerLoop(*L, Worklist);
01292 
01293     LoopsAnalyzed += Worklist.size();
01294 
01295     // Now walk the identified inner loops.
01296     bool Changed = false;
01297     while (!Worklist.empty())
01298       Changed |= processLoop(Worklist.pop_back_val());
01299 
01300     // Process each loop nest in the function.
01301     return Changed;
01302   }
01303 
01304   bool processLoop(Loop *L) {
01305     assert(L->empty() && "Only process inner loops.");
01306 
01307 #ifndef NDEBUG
01308     const std::string DebugLocStr = getDebugLocString(L);
01309 #endif /* NDEBUG */
01310 
01311     DEBUG(dbgs() << "\nLV: Checking a loop in \""
01312                  << L->getHeader()->getParent()->getName() << "\" from "
01313                  << DebugLocStr << "\n");
01314 
01315     LoopVectorizeHints Hints(L, DisableUnrolling);
01316 
01317     DEBUG(dbgs() << "LV: Loop hints:"
01318                  << " force="
01319                  << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
01320                          ? "disabled"
01321                          : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
01322                                 ? "enabled"
01323                                 : "?")) << " width=" << Hints.getWidth()
01324                  << " unroll=" << Hints.getInterleave() << "\n");
01325 
01326     // Function containing loop
01327     Function *F = L->getHeader()->getParent();
01328 
01329     // Looking at the diagnostic output is the only way to determine if a loop
01330     // was vectorized (other than looking at the IR or machine code), so it
01331     // is important to generate an optimization remark for each loop. Most of
01332     // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
01333     // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
01334     // less verbose reporting vectorized loops and unvectorized loops that may
01335     // benefit from vectorization, respectively.
01336 
01337     if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
01338       DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
01339       emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
01340                                      L->getStartLoc(), Hints.emitRemark());
01341       return false;
01342     }
01343 
01344     if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
01345       DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
01346       emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
01347                                      L->getStartLoc(), Hints.emitRemark());
01348       return false;
01349     }
01350 
01351     if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
01352       DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
01353       emitOptimizationRemarkAnalysis(
01354           F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01355           "loop not vectorized: vector width and interleave count are "
01356           "explicitly set to 1");
01357       return false;
01358     }
01359 
01360     // Check the loop for a trip count threshold:
01361     // do not vectorize loops with a tiny trip count.
01362     const unsigned TC = SE->getSmallConstantTripCount(L);
01363     if (TC > 0u && TC < TinyTripCountVectorThreshold) {
01364       DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
01365                    << "This loop is not worth vectorizing.");
01366       if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
01367         DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
01368       else {
01369         DEBUG(dbgs() << "\n");
01370         emitOptimizationRemarkAnalysis(
01371             F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01372             "vectorization is not beneficial and is not explicitly forced");
01373         return false;
01374       }
01375     }
01376 
01377     // Check if it is legal to vectorize the loop.
01378     LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F);
01379     if (!LVL.canVectorize()) {
01380       DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
01381       emitMissedWarning(F, L, Hints);
01382       return false;
01383     }
01384 
01385     // Use the cost model.
01386     LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AT, F,
01387                                   &Hints);
01388 
01389     // Check the function attributes to find out if this function should be
01390     // optimized for size.
01391     bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
01392                       F->hasFnAttribute(Attribute::OptimizeForSize);
01393 
01394     // Compute the weighted frequency of this loop being executed and see if it
01395     // is less than 20% of the function entry baseline frequency. Note that we
01396     // always have a canonical loop here because we think we *can* vectoriez.
01397     // FIXME: This is hidden behind a flag due to pervasive problems with
01398     // exactly what block frequency models.
01399     if (LoopVectorizeWithBlockFrequency) {
01400       BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
01401       if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
01402           LoopEntryFreq < ColdEntryFreq)
01403         OptForSize = true;
01404     }
01405 
01406     // Check the function attributes to see if implicit floats are allowed.a
01407     // FIXME: This check doesn't seem possibly correct -- what if the loop is
01408     // an integer loop and the vector instructions selected are purely integer
01409     // vector instructions?
01410     if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
01411       DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
01412             "attribute is used.\n");
01413       emitOptimizationRemarkAnalysis(
01414           F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01415           "loop not vectorized due to NoImplicitFloat attribute");
01416       emitMissedWarning(F, L, Hints);
01417       return false;
01418     }
01419 
01420     // Select the optimal vectorization factor.
01421     const LoopVectorizationCostModel::VectorizationFactor VF =
01422         CM.selectVectorizationFactor(OptForSize);
01423 
01424     // Select the unroll factor.
01425     const unsigned UF =
01426         CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
01427 
01428     DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
01429                  << DebugLocStr << '\n');
01430     DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
01431 
01432     if (VF.Width == 1) {
01433       DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
01434 
01435       if (UF == 1) {
01436         emitOptimizationRemarkAnalysis(
01437             F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01438             "not beneficial to vectorize and user disabled interleaving");
01439         return false;
01440       }
01441       DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
01442 
01443       // Report the unrolling decision.
01444       emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01445                              Twine("unrolled with interleaving factor " +
01446                                    Twine(UF) +
01447                                    " (vectorization not beneficial)"));
01448 
01449       // We decided not to vectorize, but we may want to unroll.
01450 
01451       InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
01452       Unroller.vectorize(&LVL);
01453     } else {
01454       // If we decided that it is *legal* to vectorize the loop then do it.
01455       InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
01456       LB.vectorize(&LVL);
01457       ++LoopsVectorized;
01458 
01459       // Report the vectorization decision.
01460       emitOptimizationRemark(
01461           F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01462           Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
01463               ", unrolling interleave factor: " + Twine(UF) + ")");
01464     }
01465 
01466     // Mark the loop as already vectorized to avoid vectorizing again.
01467     Hints.setAlreadyVectorized();
01468 
01469     DEBUG(verifyFunction(*L->getHeader()->getParent()));
01470     return true;
01471   }
01472 
01473   void getAnalysisUsage(AnalysisUsage &AU) const override {
01474     AU.addRequired<AssumptionTracker>();
01475     AU.addRequiredID(LoopSimplifyID);
01476     AU.addRequiredID(LCSSAID);
01477     AU.addRequired<BlockFrequencyInfo>();
01478     AU.addRequired<DominatorTreeWrapperPass>();
01479     AU.addRequired<LoopInfo>();
01480     AU.addRequired<ScalarEvolution>();
01481     AU.addRequired<TargetTransformInfo>();
01482     AU.addRequired<AliasAnalysis>();
01483     AU.addPreserved<LoopInfo>();
01484     AU.addPreserved<DominatorTreeWrapperPass>();
01485     AU.addPreserved<AliasAnalysis>();
01486   }
01487 
01488 };
01489 
01490 } // end anonymous namespace
01491 
01492 //===----------------------------------------------------------------------===//
01493 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
01494 // LoopVectorizationCostModel.
01495 //===----------------------------------------------------------------------===//
01496 
01497 static Value *stripIntegerCast(Value *V) {
01498   if (CastInst *CI = dyn_cast<CastInst>(V))
01499     if (CI->getOperand(0)->getType()->isIntegerTy())
01500       return CI->getOperand(0);
01501   return V;
01502 }
01503 
01504 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
01505 ///
01506 /// If \p OrigPtr is not null, use it to look up the stride value instead of
01507 /// \p Ptr.
01508 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
01509                                              ValueToValueMap &PtrToStride,
01510                                              Value *Ptr, Value *OrigPtr = nullptr) {
01511 
01512   const SCEV *OrigSCEV = SE->getSCEV(Ptr);
01513 
01514   // If there is an entry in the map return the SCEV of the pointer with the
01515   // symbolic stride replaced by one.
01516   ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
01517   if (SI != PtrToStride.end()) {
01518     Value *StrideVal = SI->second;
01519 
01520     // Strip casts.
01521     StrideVal = stripIntegerCast(StrideVal);
01522 
01523     // Replace symbolic stride by one.
01524     Value *One = ConstantInt::get(StrideVal->getType(), 1);
01525     ValueToValueMap RewriteMap;
01526     RewriteMap[StrideVal] = One;
01527 
01528     const SCEV *ByOne =
01529         SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
01530     DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
01531                  << "\n");
01532     return ByOne;
01533   }
01534 
01535   // Otherwise, just return the SCEV of the original pointer.
01536   return SE->getSCEV(Ptr);
01537 }
01538 
01539 void LoopVectorizationLegality::RuntimePointerCheck::insert(
01540     ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
01541     unsigned ASId, ValueToValueMap &Strides) {
01542   // Get the stride replaced scev.
01543   const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
01544   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
01545   assert(AR && "Invalid addrec expression");
01546   const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
01547   const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
01548   Pointers.push_back(Ptr);
01549   Starts.push_back(AR->getStart());
01550   Ends.push_back(ScEnd);
01551   IsWritePtr.push_back(WritePtr);
01552   DependencySetId.push_back(DepSetId);
01553   AliasSetId.push_back(ASId);
01554 }
01555 
01556 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
01557   // We need to place the broadcast of invariant variables outside the loop.
01558   Instruction *Instr = dyn_cast<Instruction>(V);
01559   bool NewInstr =
01560       (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
01561                           Instr->getParent()) != LoopVectorBody.end());
01562   bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
01563 
01564   // Place the code for broadcasting invariant variables in the new preheader.
01565   IRBuilder<>::InsertPointGuard Guard(Builder);
01566   if (Invariant)
01567     Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
01568 
01569   // Broadcast the scalar into all locations in the vector.
01570   Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
01571 
01572   return Shuf;
01573 }
01574 
01575 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
01576                                                  bool Negate) {
01577   assert(Val->getType()->isVectorTy() && "Must be a vector");
01578   assert(Val->getType()->getScalarType()->isIntegerTy() &&
01579          "Elem must be an integer");
01580   // Create the types.
01581   Type *ITy = Val->getType()->getScalarType();
01582   VectorType *Ty = cast<VectorType>(Val->getType());
01583   int VLen = Ty->getNumElements();
01584   SmallVector<Constant*, 8> Indices;
01585 
01586   // Create a vector of consecutive numbers from zero to VF.
01587   for (int i = 0; i < VLen; ++i) {
01588     int64_t Idx = Negate ? (-i) : i;
01589     Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
01590   }
01591 
01592   // Add the consecutive indices to the vector value.
01593   Constant *Cv = ConstantVector::get(Indices);
01594   assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
01595   return Builder.CreateAdd(Val, Cv, "induction");
01596 }
01597 
01598 /// \brief Find the operand of the GEP that should be checked for consecutive
01599 /// stores. This ignores trailing indices that have no effect on the final
01600 /// pointer.
01601 static unsigned getGEPInductionOperand(const DataLayout *DL,
01602                                        const GetElementPtrInst *Gep) {
01603   unsigned LastOperand = Gep->getNumOperands() - 1;
01604   unsigned GEPAllocSize = DL->getTypeAllocSize(
01605       cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
01606 
01607   // Walk backwards and try to peel off zeros.
01608   while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
01609     // Find the type we're currently indexing into.
01610     gep_type_iterator GEPTI = gep_type_begin(Gep);
01611     std::advance(GEPTI, LastOperand - 1);
01612 
01613     // If it's a type with the same allocation size as the result of the GEP we
01614     // can peel off the zero index.
01615     if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
01616       break;
01617     --LastOperand;
01618   }
01619 
01620   return LastOperand;
01621 }
01622 
01623 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
01624   assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
01625   // Make sure that the pointer does not point to structs.
01626   if (Ptr->getType()->getPointerElementType()->isAggregateType())
01627     return 0;
01628 
01629   // If this value is a pointer induction variable we know it is consecutive.
01630   PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
01631   if (Phi && Inductions.count(Phi)) {
01632     InductionInfo II = Inductions[Phi];
01633     if (IK_PtrInduction == II.IK)
01634       return 1;
01635     else if (IK_ReversePtrInduction == II.IK)
01636       return -1;
01637   }
01638 
01639   GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
01640   if (!Gep)
01641     return 0;
01642 
01643   unsigned NumOperands = Gep->getNumOperands();
01644   Value *GpPtr = Gep->getPointerOperand();
01645   // If this GEP value is a consecutive pointer induction variable and all of
01646   // the indices are constant then we know it is consecutive. We can
01647   Phi = dyn_cast<PHINode>(GpPtr);
01648   if (Phi && Inductions.count(Phi)) {
01649 
01650     // Make sure that the pointer does not point to structs.
01651     PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
01652     if (GepPtrType->getElementType()->isAggregateType())
01653       return 0;
01654 
01655     // Make sure that all of the index operands are loop invariant.
01656     for (unsigned i = 1; i < NumOperands; ++i)
01657       if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
01658         return 0;
01659 
01660     InductionInfo II = Inductions[Phi];
01661     if (IK_PtrInduction == II.IK)
01662       return 1;
01663     else if (IK_ReversePtrInduction == II.IK)
01664       return -1;
01665   }
01666 
01667   unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
01668 
01669   // Check that all of the gep indices are uniform except for our induction
01670   // operand.
01671   for (unsigned i = 0; i != NumOperands; ++i)
01672     if (i != InductionOperand &&
01673         !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
01674       return 0;
01675 
01676   // We can emit wide load/stores only if the last non-zero index is the
01677   // induction variable.
01678   const SCEV *Last = nullptr;
01679   if (!Strides.count(Gep))
01680     Last = SE->getSCEV(Gep->getOperand(InductionOperand));
01681   else {
01682     // Because of the multiplication by a stride we can have a s/zext cast.
01683     // We are going to replace this stride by 1 so the cast is safe to ignore.
01684     //
01685     //  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
01686     //  %0 = trunc i64 %indvars.iv to i32
01687     //  %mul = mul i32 %0, %Stride1
01688     //  %idxprom = zext i32 %mul to i64  << Safe cast.
01689     //  %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
01690     //
01691     Last = replaceSymbolicStrideSCEV(SE, Strides,
01692                                      Gep->getOperand(InductionOperand), Gep);
01693     if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
01694       Last =
01695           (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
01696               ? C->getOperand()
01697               : Last;
01698   }
01699   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
01700     const SCEV *Step = AR->getStepRecurrence(*SE);
01701 
01702     // The memory is consecutive because the last index is consecutive
01703     // and all other indices are loop invariant.
01704     if (Step->isOne())
01705       return 1;
01706     if (Step->isAllOnesValue())
01707       return -1;
01708   }
01709 
01710   return 0;
01711 }
01712 
01713 bool LoopVectorizationLegality::isUniform(Value *V) {
01714   return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
01715 }
01716 
01717 InnerLoopVectorizer::VectorParts&
01718 InnerLoopVectorizer::getVectorValue(Value *V) {
01719   assert(V != Induction && "The new induction variable should not be used.");
01720   assert(!V->getType()->isVectorTy() && "Can't widen a vector");
01721 
01722   // If we have a stride that is replaced by one, do it here.
01723   if (Legal->hasStride(V))
01724     V = ConstantInt::get(V->getType(), 1);
01725 
01726   // If we have this scalar in the map, return it.
01727   if (WidenMap.has(V))
01728     return WidenMap.get(V);
01729 
01730   // If this scalar is unknown, assume that it is a constant or that it is
01731   // loop invariant. Broadcast V and save the value for future uses.
01732   Value *B = getBroadcastInstrs(V);
01733   return WidenMap.splat(V, B);
01734 }
01735 
01736 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
01737   assert(Vec->getType()->isVectorTy() && "Invalid type");
01738   SmallVector<Constant*, 8> ShuffleMask;
01739   for (unsigned i = 0; i < VF; ++i)
01740     ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
01741 
01742   return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
01743                                      ConstantVector::get(ShuffleMask),
01744                                      "reverse");
01745 }
01746 
01747 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
01748   // Attempt to issue a wide load.
01749   LoadInst *LI = dyn_cast<LoadInst>(Instr);
01750   StoreInst *SI = dyn_cast<StoreInst>(Instr);
01751 
01752   assert((LI || SI) && "Invalid Load/Store instruction");
01753 
01754   Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
01755   Type *DataTy = VectorType::get(ScalarDataTy, VF);
01756   Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
01757   unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
01758   // An alignment of 0 means target abi alignment. We need to use the scalar's
01759   // target abi alignment in such a case.
01760   if (!Alignment)
01761     Alignment = DL->getABITypeAlignment(ScalarDataTy);
01762   unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
01763   unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
01764   unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
01765 
01766   if (SI && Legal->blockNeedsPredication(SI->getParent()))
01767     return scalarizeInstruction(Instr, true);
01768 
01769   if (ScalarAllocatedSize != VectorElementSize)
01770     return scalarizeInstruction(Instr);
01771 
01772   // If the pointer is loop invariant or if it is non-consecutive,
01773   // scalarize the load.
01774   int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
01775   bool Reverse = ConsecutiveStride < 0;
01776   bool UniformLoad = LI && Legal->isUniform(Ptr);
01777   if (!ConsecutiveStride || UniformLoad)
01778     return scalarizeInstruction(Instr);
01779 
01780   Constant *Zero = Builder.getInt32(0);
01781   VectorParts &Entry = WidenMap.get(Instr);
01782 
01783   // Handle consecutive loads/stores.
01784   GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
01785   if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
01786     setDebugLocFromInst(Builder, Gep);
01787     Value *PtrOperand = Gep->getPointerOperand();
01788     Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
01789     FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
01790 
01791     // Create the new GEP with the new induction variable.
01792     GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
01793     Gep2->setOperand(0, FirstBasePtr);
01794     Gep2->setName("gep.indvar.base");
01795     Ptr = Builder.Insert(Gep2);
01796   } else if (Gep) {
01797     setDebugLocFromInst(Builder, Gep);
01798     assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
01799                                OrigLoop) && "Base ptr must be invariant");
01800 
01801     // The last index does not have to be the induction. It can be
01802     // consecutive and be a function of the index. For example A[I+1];
01803     unsigned NumOperands = Gep->getNumOperands();
01804     unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
01805     // Create the new GEP with the new induction variable.
01806     GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
01807 
01808     for (unsigned i = 0; i < NumOperands; ++i) {
01809       Value *GepOperand = Gep->getOperand(i);
01810       Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
01811 
01812       // Update last index or loop invariant instruction anchored in loop.
01813       if (i == InductionOperand ||
01814           (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
01815         assert((i == InductionOperand ||
01816                SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
01817                "Must be last index or loop invariant");
01818 
01819         VectorParts &GEPParts = getVectorValue(GepOperand);
01820         Value *Index = GEPParts[0];
01821         Index = Builder.CreateExtractElement(Index, Zero);
01822         Gep2->setOperand(i, Index);
01823         Gep2->setName("gep.indvar.idx");
01824       }
01825     }
01826     Ptr = Builder.Insert(Gep2);
01827   } else {
01828     // Use the induction element ptr.
01829     assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
01830     setDebugLocFromInst(Builder, Ptr);
01831     VectorParts &PtrVal = getVectorValue(Ptr);
01832     Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
01833   }
01834 
01835   // Handle Stores:
01836   if (SI) {
01837     assert(!Legal->isUniform(SI->getPointerOperand()) &&
01838            "We do not allow storing to uniform addresses");
01839     setDebugLocFromInst(Builder, SI);
01840     // We don't want to update the value in the map as it might be used in
01841     // another expression. So don't use a reference type for "StoredVal".
01842     VectorParts StoredVal = getVectorValue(SI->getValueOperand());
01843 
01844     for (unsigned Part = 0; Part < UF; ++Part) {
01845       // Calculate the pointer for the specific unroll-part.
01846       Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
01847 
01848       if (Reverse) {
01849         // If we store to reverse consecutive memory locations then we need
01850         // to reverse the order of elements in the stored value.
01851         StoredVal[Part] = reverseVector(StoredVal[Part]);
01852         // If the address is consecutive but reversed, then the
01853         // wide store needs to start at the last vector element.
01854         PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
01855         PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
01856       }
01857 
01858       Value *VecPtr = Builder.CreateBitCast(PartPtr,
01859                                             DataTy->getPointerTo(AddressSpace));
01860       StoreInst *NewSI =
01861         Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
01862       propagateMetadata(NewSI, SI);
01863     }
01864     return;
01865   }
01866 
01867   // Handle loads.
01868   assert(LI && "Must have a load instruction");
01869   setDebugLocFromInst(Builder, LI);
01870   for (unsigned Part = 0; Part < UF; ++Part) {
01871     // Calculate the pointer for the specific unroll-part.
01872     Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
01873 
01874     if (Reverse) {
01875       // If the address is consecutive but reversed, then the
01876       // wide store needs to start at the last vector element.
01877       PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
01878       PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
01879     }
01880 
01881     Value *VecPtr = Builder.CreateBitCast(PartPtr,
01882                                           DataTy->getPointerTo(AddressSpace));
01883     LoadInst *NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
01884     propagateMetadata(NewLI, LI);
01885     Entry[Part] = Reverse ? reverseVector(NewLI) :  NewLI;
01886   }
01887 }
01888 
01889 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
01890   assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
01891   // Holds vector parameters or scalars, in case of uniform vals.
01892   SmallVector<VectorParts, 4> Params;
01893 
01894   setDebugLocFromInst(Builder, Instr);
01895 
01896   // Find all of the vectorized parameters.
01897   for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
01898     Value *SrcOp = Instr->getOperand(op);
01899 
01900     // If we are accessing the old induction variable, use the new one.
01901     if (SrcOp == OldInduction) {
01902       Params.push_back(getVectorValue(SrcOp));
01903       continue;
01904     }
01905 
01906     // Try using previously calculated values.
01907     Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
01908 
01909     // If the src is an instruction that appeared earlier in the basic block
01910     // then it should already be vectorized.
01911     if (SrcInst && OrigLoop->contains(SrcInst)) {
01912       assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
01913       // The parameter is a vector value from earlier.
01914       Params.push_back(WidenMap.get(SrcInst));
01915     } else {
01916       // The parameter is a scalar from outside the loop. Maybe even a constant.
01917       VectorParts Scalars;
01918       Scalars.append(UF, SrcOp);
01919       Params.push_back(Scalars);
01920     }
01921   }
01922 
01923   assert(Params.size() == Instr->getNumOperands() &&
01924          "Invalid number of operands");
01925 
01926   // Does this instruction return a value ?
01927   bool IsVoidRetTy = Instr->getType()->isVoidTy();
01928 
01929   Value *UndefVec = IsVoidRetTy ? nullptr :
01930     UndefValue::get(VectorType::get(Instr->getType(), VF));
01931   // Create a new entry in the WidenMap and initialize it to Undef or Null.
01932   VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
01933 
01934   Instruction *InsertPt = Builder.GetInsertPoint();
01935   BasicBlock *IfBlock = Builder.GetInsertBlock();
01936   BasicBlock *CondBlock = nullptr;
01937 
01938   VectorParts Cond;
01939   Loop *VectorLp = nullptr;
01940   if (IfPredicateStore) {
01941     assert(Instr->getParent()->getSinglePredecessor() &&
01942            "Only support single predecessor blocks");
01943     Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
01944                           Instr->getParent());
01945     VectorLp = LI->getLoopFor(IfBlock);
01946     assert(VectorLp && "Must have a loop for this block");
01947   }
01948 
01949   // For each vector unroll 'part':
01950   for (unsigned Part = 0; Part < UF; ++Part) {
01951     // For each scalar that we create:
01952     for (unsigned Width = 0; Width < VF; ++Width) {
01953 
01954       // Start if-block.
01955       Value *Cmp = nullptr;
01956       if (IfPredicateStore) {
01957         Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
01958         Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
01959         CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
01960         LoopVectorBody.push_back(CondBlock);
01961         VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
01962         // Update Builder with newly created basic block.
01963         Builder.SetInsertPoint(InsertPt);
01964       }
01965 
01966       Instruction *Cloned = Instr->clone();
01967       if (!IsVoidRetTy)
01968         Cloned->setName(Instr->getName() + ".cloned");
01969       // Replace the operands of the cloned instructions with extracted scalars.
01970       for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
01971         Value *Op = Params[op][Part];
01972         // Param is a vector. Need to extract the right lane.
01973         if (Op->getType()->isVectorTy())
01974           Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
01975         Cloned->setOperand(op, Op);
01976       }
01977 
01978       // Place the cloned scalar in the new loop.
01979       Builder.Insert(Cloned);
01980 
01981       // If the original scalar returns a value we need to place it in a vector
01982       // so that future users will be able to use it.
01983       if (!IsVoidRetTy)
01984         VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
01985                                                        Builder.getInt32(Width));
01986       // End if-block.
01987       if (IfPredicateStore) {
01988          BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
01989          LoopVectorBody.push_back(NewIfBlock);
01990          VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
01991          Builder.SetInsertPoint(InsertPt);
01992          Instruction *OldBr = IfBlock->getTerminator();
01993          BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
01994          OldBr->eraseFromParent();
01995          IfBlock = NewIfBlock;
01996       }
01997     }
01998   }
01999 }
02000 
02001 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
02002                                  Instruction *Loc) {
02003   if (FirstInst)
02004     return FirstInst;
02005   if (Instruction *I = dyn_cast<Instruction>(V))
02006     return I->getParent() == Loc->getParent() ? I : nullptr;
02007   return nullptr;
02008 }
02009 
02010 std::pair<Instruction *, Instruction *>
02011 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
02012   Instruction *tnullptr = nullptr;
02013   if (!Legal->mustCheckStrides())
02014     return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
02015 
02016   IRBuilder<> ChkBuilder(Loc);
02017 
02018   // Emit checks.
02019   Value *Check = nullptr;
02020   Instruction *FirstInst = nullptr;
02021   for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
02022                                          SE = Legal->strides_end();
02023        SI != SE; ++SI) {
02024     Value *Ptr = stripIntegerCast(*SI);
02025     Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
02026                                        "stride.chk");
02027     // Store the first instruction we create.
02028     FirstInst = getFirstInst(FirstInst, C, Loc);
02029     if (Check)
02030       Check = ChkBuilder.CreateOr(Check, C);
02031     else
02032       Check = C;
02033   }
02034 
02035   // We have to do this trickery because the IRBuilder might fold the check to a
02036   // constant expression in which case there is no Instruction anchored in a
02037   // the block.
02038   LLVMContext &Ctx = Loc->getContext();
02039   Instruction *TheCheck =
02040       BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
02041   ChkBuilder.Insert(TheCheck, "stride.not.one");
02042   FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
02043 
02044   return std::make_pair(FirstInst, TheCheck);
02045 }
02046 
02047 std::pair<Instruction *, Instruction *>
02048 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
02049   LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
02050   Legal->getRuntimePointerCheck();
02051 
02052   Instruction *tnullptr = nullptr;
02053   if (!PtrRtCheck->Need)
02054     return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
02055 
02056   unsigned NumPointers = PtrRtCheck->Pointers.size();
02057   SmallVector<TrackingVH<Value> , 2> Starts;
02058   SmallVector<TrackingVH<Value> , 2> Ends;
02059 
02060   LLVMContext &Ctx = Loc->getContext();
02061   SCEVExpander Exp(*SE, "induction");
02062   Instruction *FirstInst = nullptr;
02063 
02064   for (unsigned i = 0; i < NumPointers; ++i) {
02065     Value *Ptr = PtrRtCheck->Pointers[i];
02066     const SCEV *Sc = SE->getSCEV(Ptr);
02067 
02068     if (SE->isLoopInvariant(Sc, OrigLoop)) {
02069       DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
02070             *Ptr <<"\n");
02071       Starts.push_back(Ptr);
02072       Ends.push_back(Ptr);
02073     } else {
02074       DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
02075       unsigned AS = Ptr->getType()->getPointerAddressSpace();
02076 
02077       // Use this type for pointer arithmetic.
02078       Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
02079 
02080       Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
02081       Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
02082       Starts.push_back(Start);
02083       Ends.push_back(End);
02084     }
02085   }
02086 
02087   IRBuilder<> ChkBuilder(Loc);
02088   // Our instructions might fold to a constant.
02089   Value *MemoryRuntimeCheck = nullptr;
02090   for (unsigned i = 0; i < NumPointers; ++i) {
02091     for (unsigned j = i+1; j < NumPointers; ++j) {
02092       // No need to check if two readonly pointers intersect.
02093       if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
02094         continue;
02095 
02096       // Only need to check pointers between two different dependency sets.
02097       if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
02098        continue;
02099       // Only need to check pointers in the same alias set.
02100       if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
02101         continue;
02102 
02103       unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
02104       unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
02105 
02106       assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
02107              (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
02108              "Trying to bounds check pointers with different address spaces");
02109 
02110       Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
02111       Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
02112 
02113       Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
02114       Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
02115       Value *End0 =   ChkBuilder.CreateBitCast(Ends[i],   PtrArithTy1, "bc");
02116       Value *End1 =   ChkBuilder.CreateBitCast(Ends[j],   PtrArithTy0, "bc");
02117 
02118       Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
02119       FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
02120       Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
02121       FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
02122       Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
02123       FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
02124       if (MemoryRuntimeCheck) {
02125         IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
02126                                          "conflict.rdx");
02127         FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
02128       }
02129       MemoryRuntimeCheck = IsConflict;
02130     }
02131   }
02132 
02133   // We have to do this trickery because the IRBuilder might fold the check to a
02134   // constant expression in which case there is no Instruction anchored in a
02135   // the block.
02136   Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
02137                                                  ConstantInt::getTrue(Ctx));
02138   ChkBuilder.Insert(Check, "memcheck.conflict");
02139   FirstInst = getFirstInst(FirstInst, Check, Loc);
02140   return std::make_pair(FirstInst, Check);
02141 }
02142 
02143 void InnerLoopVectorizer::createEmptyLoop() {
02144   /*
02145    In this function we generate a new loop. The new loop will contain
02146    the vectorized instructions while the old loop will continue to run the
02147    scalar remainder.
02148 
02149        [ ] <-- Back-edge taken count overflow check.
02150     /   |
02151    /    v
02152   |    [ ] <-- vector loop bypass (may consist of multiple blocks).
02153   |  /  |
02154   | /   v
02155   ||   [ ]     <-- vector pre header.
02156   ||    |
02157   ||    v
02158   ||   [  ] \
02159   ||   [  ]_|   <-- vector loop.
02160   ||    |
02161   | \   v
02162   |   >[ ]   <--- middle-block.
02163   |  /  |
02164   | /   v
02165   -|- >[ ]     <--- new preheader.
02166    |    |
02167    |    v
02168    |   [ ] \
02169    |   [ ]_|   <-- old scalar loop to handle remainder.
02170     \   |
02171      \  v
02172       >[ ]     <-- exit block.
02173    ...
02174    */
02175 
02176   BasicBlock *OldBasicBlock = OrigLoop->getHeader();
02177   BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
02178   BasicBlock *ExitBlock = OrigLoop->getExitBlock();
02179   assert(BypassBlock && "Invalid loop structure");
02180   assert(ExitBlock && "Must have an exit block");
02181 
02182   // Some loops have a single integer induction variable, while other loops
02183   // don't. One example is c++ iterators that often have multiple pointer
02184   // induction variables. In the code below we also support a case where we
02185   // don't have a single induction variable.
02186   OldInduction = Legal->getInduction();
02187   Type *IdxTy = Legal->getWidestInductionType();
02188 
02189   // Find the loop boundaries.
02190   const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
02191   assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
02192 
02193   // The exit count might have the type of i64 while the phi is i32. This can
02194   // happen if we have an induction variable that is sign extended before the
02195   // compare. The only way that we get a backedge taken count is that the
02196   // induction variable was signed and as such will not overflow. In such a case
02197   // truncation is legal.
02198   if (ExitCount->getType()->getPrimitiveSizeInBits() >
02199       IdxTy->getPrimitiveSizeInBits())
02200     ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
02201 
02202   const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
02203   // Get the total trip count from the count by adding 1.
02204   ExitCount = SE->getAddExpr(BackedgeTakeCount,
02205                              SE->getConstant(BackedgeTakeCount->getType(), 1));
02206 
02207   // Expand the trip count and place the new instructions in the preheader.
02208   // Notice that the pre-header does not change, only the loop body.
02209   SCEVExpander Exp(*SE, "induction");
02210 
02211   // We need to test whether the backedge-taken count is uint##_max. Adding one
02212   // to it will cause overflow and an incorrect loop trip count in the vector
02213   // body. In case of overflow we want to directly jump to the scalar remainder
02214   // loop.
02215   Value *BackedgeCount =
02216       Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
02217                         BypassBlock->getTerminator());
02218   if (BackedgeCount->getType()->isPointerTy())
02219     BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
02220                                                 "backedge.ptrcnt.to.int",
02221                                                 BypassBlock->getTerminator());
02222   Instruction *CheckBCOverflow =
02223       CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
02224                       Constant::getAllOnesValue(BackedgeCount->getType()),
02225                       "backedge.overflow", BypassBlock->getTerminator());
02226 
02227   // The loop index does not have to start at Zero. Find the original start
02228   // value from the induction PHI node. If we don't have an induction variable
02229   // then we know that it starts at zero.
02230   Builder.SetInsertPoint(BypassBlock->getTerminator());
02231   Value *StartIdx = ExtendedIdx = OldInduction ?
02232     Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
02233                        IdxTy):
02234     ConstantInt::get(IdxTy, 0);
02235 
02236   // We need an instruction to anchor the overflow check on. StartIdx needs to
02237   // be defined before the overflow check branch. Because the scalar preheader
02238   // is going to merge the start index and so the overflow branch block needs to
02239   // contain a definition of the start index.
02240   Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
02241       StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
02242       BypassBlock->getTerminator());
02243 
02244   // Count holds the overall loop count (N).
02245   Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
02246                                    BypassBlock->getTerminator());
02247 
02248   LoopBypassBlocks.push_back(BypassBlock);
02249 
02250   // Split the single block loop into the two loop structure described above.
02251   BasicBlock *VectorPH =
02252   BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
02253   BasicBlock *VecBody =
02254   VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
02255   BasicBlock *MiddleBlock =
02256   VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
02257   BasicBlock *ScalarPH =
02258   MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
02259 
02260   // Create and register the new vector loop.
02261   Loop* Lp = new Loop();
02262   Loop *ParentLoop = OrigLoop->getParentLoop();
02263 
02264   // Insert the new loop into the loop nest and register the new basic blocks
02265   // before calling any utilities such as SCEV that require valid LoopInfo.
02266   if (ParentLoop) {
02267     ParentLoop->addChildLoop(Lp);
02268     ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
02269     ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
02270     ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
02271   } else {
02272     LI->addTopLevelLoop(Lp);
02273   }
02274   Lp->addBasicBlockToLoop(VecBody, LI->getBase());
02275 
02276   // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
02277   // inside the loop.
02278   Builder.SetInsertPoint(VecBody->getFirstNonPHI());
02279 
02280   // Generate the induction variable.
02281   setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
02282   Induction = Builder.CreatePHI(IdxTy, 2, "index");
02283   // The loop step is equal to the vectorization factor (num of SIMD elements)
02284   // times the unroll factor (num of SIMD instructions).
02285   Constant *Step = ConstantInt::get(IdxTy, VF * UF);
02286 
02287   // This is the IR builder that we use to add all of the logic for bypassing
02288   // the new vector loop.
02289   IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
02290   setDebugLocFromInst(BypassBuilder,
02291                       getDebugLocFromInstOrOperands(OldInduction));
02292 
02293   // We may need to extend the index in case there is a type mismatch.
02294   // We know that the count starts at zero and does not overflow.
02295   if (Count->getType() != IdxTy) {
02296     // The exit count can be of pointer type. Convert it to the correct
02297     // integer type.
02298     if (ExitCount->getType()->isPointerTy())
02299       Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
02300     else
02301       Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
02302   }
02303 
02304   // Add the start index to the loop count to get the new end index.
02305   Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
02306 
02307   // Now we need to generate the expression for N - (N % VF), which is
02308   // the part that the vectorized body will execute.
02309   Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
02310   Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
02311   Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
02312                                                      "end.idx.rnd.down");
02313 
02314   // Now, compare the new count to zero. If it is zero skip the vector loop and
02315   // jump to the scalar loop.
02316   Value *Cmp =
02317       BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
02318 
02319   BasicBlock *LastBypassBlock = BypassBlock;
02320 
02321   // Generate code to check that the loops trip count that we computed by adding
02322   // one to the backedge-taken count will not overflow.
02323   {
02324     auto PastOverflowCheck =
02325         std::next(BasicBlock::iterator(OverflowCheckAnchor));
02326     BasicBlock *CheckBlock =
02327       LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
02328     if (ParentLoop)
02329       ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
02330     LoopBypassBlocks.push_back(CheckBlock);
02331     Instruction *OldTerm = LastBypassBlock->getTerminator();
02332     BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
02333     OldTerm->eraseFromParent();
02334     LastBypassBlock = CheckBlock;
02335   }
02336 
02337   // Generate the code to check that the strides we assumed to be one are really
02338   // one. We want the new basic block to start at the first instruction in a
02339   // sequence of instructions that form a check.
02340   Instruction *StrideCheck;
02341   Instruction *FirstCheckInst;
02342   std::tie(FirstCheckInst, StrideCheck) =
02343       addStrideCheck(LastBypassBlock->getTerminator());
02344   if (StrideCheck) {
02345     // Create a new block containing the stride check.
02346     BasicBlock *CheckBlock =
02347         LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
02348     if (ParentLoop)
02349       ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
02350     LoopBypassBlocks.push_back(CheckBlock);
02351 
02352     // Replace the branch into the memory check block with a conditional branch
02353     // for the "few elements case".
02354     Instruction *OldTerm = LastBypassBlock->getTerminator();
02355     BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
02356     OldTerm->eraseFromParent();
02357 
02358     Cmp = StrideCheck;
02359     LastBypassBlock = CheckBlock;
02360   }
02361 
02362   // Generate the code that checks in runtime if arrays overlap. We put the
02363   // checks into a separate block to make the more common case of few elements
02364   // faster.
02365   Instruction *MemRuntimeCheck;
02366   std::tie(FirstCheckInst, MemRuntimeCheck) =
02367       addRuntimeCheck(LastBypassBlock->getTerminator());
02368   if (MemRuntimeCheck) {
02369     // Create a new block containing the memory check.
02370     BasicBlock *CheckBlock =
02371         LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
02372     if (ParentLoop)
02373       ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
02374     LoopBypassBlocks.push_back(CheckBlock);
02375 
02376     // Replace the branch into the memory check block with a conditional branch
02377     // for the "few elements case".
02378     Instruction *OldTerm = LastBypassBlock->getTerminator();
02379     BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
02380     OldTerm->eraseFromParent();
02381 
02382     Cmp = MemRuntimeCheck;
02383     LastBypassBlock = CheckBlock;
02384   }
02385 
02386   LastBypassBlock->getTerminator()->eraseFromParent();
02387   BranchInst::Create(MiddleBlock, VectorPH, Cmp,
02388                      LastBypassBlock);
02389 
02390   // We are going to resume the execution of the scalar loop.
02391   // Go over all of the induction variables that we found and fix the
02392   // PHIs that are left in the scalar version of the loop.
02393   // The starting values of PHI nodes depend on the counter of the last
02394   // iteration in the vectorized loop.
02395   // If we come from a bypass edge then we need to start from the original
02396   // start value.
02397 
02398   // This variable saves the new starting index for the scalar loop.
02399   PHINode *ResumeIndex = nullptr;
02400   LoopVectorizationLegality::InductionList::iterator I, E;
02401   LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
02402   // Set builder to point to last bypass block.
02403   BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
02404   for (I = List->begin(), E = List->end(); I != E; ++I) {
02405     PHINode *OrigPhi = I->first;
02406     LoopVectorizationLegality::InductionInfo II = I->second;
02407 
02408     Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
02409     PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
02410                                          MiddleBlock->getTerminator());
02411     // We might have extended the type of the induction variable but we need a
02412     // truncated version for the scalar loop.
02413     PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
02414       PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
02415                       MiddleBlock->getTerminator()) : nullptr;
02416 
02417     // Create phi nodes to merge from the  backedge-taken check block.
02418     PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
02419                                            ScalarPH->getTerminator());
02420     BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
02421 
02422     PHINode *BCTruncResumeVal = nullptr;
02423     if (OrigPhi == OldInduction) {
02424       BCTruncResumeVal =
02425           PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
02426                           ScalarPH->getTerminator());
02427       BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
02428     }
02429 
02430     Value *EndValue = nullptr;
02431     switch (II.IK) {
02432     case LoopVectorizationLegality::IK_NoInduction:
02433       llvm_unreachable("Unknown induction");
02434     case LoopVectorizationLegality::IK_IntInduction: {
02435       // Handle the integer induction counter.
02436       assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
02437 
02438       // We have the canonical induction variable.
02439       if (OrigPhi == OldInduction) {
02440         // Create a truncated version of the resume value for the scalar loop,
02441         // we might have promoted the type to a larger width.
02442         EndValue =
02443           BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
02444         // The new PHI merges the original incoming value, in case of a bypass,
02445         // or the value at the end of the vectorized loop.
02446         for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
02447           TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
02448         TruncResumeVal->addIncoming(EndValue, VecBody);
02449 
02450         BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
02451 
02452         // We know what the end value is.
02453         EndValue = IdxEndRoundDown;
02454         // We also know which PHI node holds it.
02455         ResumeIndex = ResumeVal;
02456         break;
02457       }
02458 
02459       // Not the canonical induction variable - add the vector loop count to the
02460       // start value.
02461       Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
02462                                                    II.StartValue->getType(),
02463                                                    "cast.crd");
02464       EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
02465       break;
02466     }
02467     case LoopVectorizationLegality::IK_ReverseIntInduction: {
02468       // Convert the CountRoundDown variable to the PHI size.
02469       Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
02470                                                    II.StartValue->getType(),
02471                                                    "cast.crd");
02472       // Handle reverse integer induction counter.
02473       EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
02474       break;
02475     }
02476     case LoopVectorizationLegality::IK_PtrInduction: {
02477       // For pointer induction variables, calculate the offset using
02478       // the end index.
02479       EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
02480                                          "ptr.ind.end");
02481       break;
02482     }
02483     case LoopVectorizationLegality::IK_ReversePtrInduction: {
02484       // The value at the end of the loop for the reverse pointer is calculated
02485       // by creating a GEP with a negative index starting from the start value.
02486       Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
02487       Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
02488                                               "rev.ind.end");
02489       EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
02490                                          "rev.ptr.ind.end");
02491       break;
02492     }
02493     }// end of case
02494 
02495     // The new PHI merges the original incoming value, in case of a bypass,
02496     // or the value at the end of the vectorized loop.
02497     for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
02498       if (OrigPhi == OldInduction)
02499         ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
02500       else
02501         ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
02502     }
02503     ResumeVal->addIncoming(EndValue, VecBody);
02504 
02505     // Fix the scalar body counter (PHI node).
02506     unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
02507 
02508     // The old induction's phi node in the scalar body needs the truncated
02509     // value.
02510     if (OrigPhi == OldInduction) {
02511       BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
02512       OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
02513     } else {
02514       BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
02515       OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
02516     }
02517   }
02518 
02519   // If we are generating a new induction variable then we also need to
02520   // generate the code that calculates the exit value. This value is not
02521   // simply the end of the counter because we may skip the vectorized body
02522   // in case of a runtime check.
02523   if (!OldInduction){
02524     assert(!ResumeIndex && "Unexpected resume value found");
02525     ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
02526                                   MiddleBlock->getTerminator());
02527     for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
02528       ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
02529     ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
02530   }
02531 
02532   // Make sure that we found the index where scalar loop needs to continue.
02533   assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
02534          "Invalid resume Index");
02535 
02536   // Add a check in the middle block to see if we have completed
02537   // all of the iterations in the first vector loop.
02538   // If (N - N%VF) == N, then we *don't* need to run the remainder.
02539   Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
02540                                 ResumeIndex, "cmp.n",
02541                                 MiddleBlock->getTerminator());
02542 
02543   BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
02544   // Remove the old terminator.
02545   MiddleBlock->getTerminator()->eraseFromParent();
02546 
02547   // Create i+1 and fill the PHINode.
02548   Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
02549   Induction->addIncoming(StartIdx, VectorPH);
02550   Induction->addIncoming(NextIdx, VecBody);
02551   // Create the compare.
02552   Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
02553   Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
02554 
02555   // Now we have two terminators. Remove the old one from the block.
02556   VecBody->getTerminator()->eraseFromParent();
02557 
02558   // Get ready to start creating new instructions into the vectorized body.
02559   Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
02560 
02561   // Save the state.
02562   LoopVectorPreHeader = VectorPH;
02563   LoopScalarPreHeader = ScalarPH;
02564   LoopMiddleBlock = MiddleBlock;
02565   LoopExitBlock = ExitBlock;
02566   LoopVectorBody.push_back(VecBody);
02567   LoopScalarBody = OldBasicBlock;
02568 
02569   LoopVectorizeHints Hints(Lp, true);
02570   Hints.setAlreadyVectorized();
02571 }
02572 
02573 /// This function returns the identity element (or neutral element) for
02574 /// the operation K.
02575 Constant*
02576 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
02577   switch (K) {
02578   case RK_IntegerXor:
02579   case RK_IntegerAdd:
02580   case RK_IntegerOr:
02581     // Adding, Xoring, Oring zero to a number does not change it.
02582     return ConstantInt::get(Tp, 0);
02583   case RK_IntegerMult:
02584     // Multiplying a number by 1 does not change it.
02585     return ConstantInt::get(Tp, 1);
02586   case RK_IntegerAnd:
02587     // AND-ing a number with an all-1 value does not change it.
02588     return ConstantInt::get(Tp, -1, true);
02589   case  RK_FloatMult:
02590     // Multiplying a number by 1 does not change it.
02591     return ConstantFP::get(Tp, 1.0L);
02592   case  RK_FloatAdd:
02593     // Adding zero to a number does not change it.
02594     return ConstantFP::get(Tp, 0.0L);
02595   default:
02596     llvm_unreachable("Unknown reduction kind");
02597   }
02598 }
02599 
02600 /// This function translates the reduction kind to an LLVM binary operator.
02601 static unsigned
02602 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
02603   switch (Kind) {
02604     case LoopVectorizationLegality::RK_IntegerAdd:
02605       return Instruction::Add;
02606     case LoopVectorizationLegality::RK_IntegerMult:
02607       return Instruction::Mul;
02608     case LoopVectorizationLegality::RK_IntegerOr:
02609       return Instruction::Or;
02610     case LoopVectorizationLegality::RK_IntegerAnd:
02611       return Instruction::And;
02612     case LoopVectorizationLegality::RK_IntegerXor:
02613       return Instruction::Xor;
02614     case LoopVectorizationLegality::RK_FloatMult:
02615       return Instruction::FMul;
02616     case LoopVectorizationLegality::RK_FloatAdd:
02617       return Instruction::FAdd;
02618     case LoopVectorizationLegality::RK_IntegerMinMax:
02619       return Instruction::ICmp;
02620     case LoopVectorizationLegality::RK_FloatMinMax:
02621       return Instruction::FCmp;
02622     default:
02623       llvm_unreachable("Unknown reduction operation");
02624   }
02625 }
02626 
02627 Value *createMinMaxOp(IRBuilder<> &Builder,
02628                       LoopVectorizationLegality::MinMaxReductionKind RK,
02629                       Value *Left,
02630                       Value *Right) {
02631   CmpInst::Predicate P = CmpInst::ICMP_NE;
02632   switch (RK) {
02633   default:
02634     llvm_unreachable("Unknown min/max reduction kind");
02635   case LoopVectorizationLegality::MRK_UIntMin:
02636     P = CmpInst::ICMP_ULT;
02637     break;
02638   case LoopVectorizationLegality::MRK_UIntMax:
02639     P = CmpInst::ICMP_UGT;
02640     break;
02641   case LoopVectorizationLegality::MRK_SIntMin:
02642     P = CmpInst::ICMP_SLT;
02643     break;
02644   case LoopVectorizationLegality::MRK_SIntMax:
02645     P = CmpInst::ICMP_SGT;
02646     break;
02647   case LoopVectorizationLegality::MRK_FloatMin:
02648     P = CmpInst::FCMP_OLT;
02649     break;
02650   case LoopVectorizationLegality::MRK_FloatMax:
02651     P = CmpInst::FCMP_OGT;
02652     break;
02653   }
02654 
02655   Value *Cmp;
02656   if (RK == LoopVectorizationLegality::MRK_FloatMin ||
02657       RK == LoopVectorizationLegality::MRK_FloatMax)
02658     Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
02659   else
02660     Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
02661 
02662   Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
02663   return Select;
02664 }
02665 
02666 namespace {
02667 struct CSEDenseMapInfo {
02668   static bool canHandle(Instruction *I) {
02669     return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
02670            isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
02671   }
02672   static inline Instruction *getEmptyKey() {
02673     return DenseMapInfo<Instruction *>::getEmptyKey();
02674   }
02675   static inline Instruction *getTombstoneKey() {
02676     return DenseMapInfo<Instruction *>::getTombstoneKey();
02677   }
02678   static unsigned getHashValue(Instruction *I) {
02679     assert(canHandle(I) && "Unknown instruction!");
02680     return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
02681                                                            I->value_op_end()));
02682   }
02683   static bool isEqual(Instruction *LHS, Instruction *RHS) {
02684     if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
02685         LHS == getTombstoneKey() || RHS == getTombstoneKey())
02686       return LHS == RHS;
02687     return LHS->isIdenticalTo(RHS);
02688   }
02689 };
02690 }
02691 
02692 /// \brief Check whether this block is a predicated block.
02693 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
02694 /// = ...;  " blocks. We start with one vectorized basic block. For every
02695 /// conditional block we split this vectorized block. Therefore, every second
02696 /// block will be a predicated one.
02697 static bool isPredicatedBlock(unsigned BlockNum) {
02698   return BlockNum % 2;
02699 }
02700 
02701 ///\brief Perform cse of induction variable instructions.
02702 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
02703   // Perform simple cse.
02704   SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
02705   for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
02706     BasicBlock *BB = BBs[i];
02707     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
02708       Instruction *In = I++;
02709 
02710       if (!CSEDenseMapInfo::canHandle(In))
02711         continue;
02712 
02713       // Check if we can replace this instruction with any of the
02714       // visited instructions.
02715       if (Instruction *V = CSEMap.lookup(In)) {
02716         In->replaceAllUsesWith(V);
02717         In->eraseFromParent();
02718         continue;
02719       }
02720       // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
02721       // ...;" blocks for predicated stores. Every second block is a predicated
02722       // block.
02723       if (isPredicatedBlock(i))
02724         continue;
02725 
02726       CSEMap[In] = In;
02727     }
02728   }
02729 }
02730 
02731 /// \brief Adds a 'fast' flag to floating point operations.
02732 static Value *addFastMathFlag(Value *V) {
02733   if (isa<FPMathOperator>(V)){
02734     FastMathFlags Flags;
02735     Flags.setUnsafeAlgebra();
02736     cast<Instruction>(V)->setFastMathFlags(Flags);
02737   }
02738   return V;
02739 }
02740 
02741 void InnerLoopVectorizer::vectorizeLoop() {
02742   //===------------------------------------------------===//
02743   //
02744   // Notice: any optimization or new instruction that go
02745   // into the code below should be also be implemented in
02746   // the cost-model.
02747   //
02748   //===------------------------------------------------===//
02749   Constant *Zero = Builder.getInt32(0);
02750 
02751   // In order to support reduction variables we need to be able to vectorize
02752   // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
02753   // stages. First, we create a new vector PHI node with no incoming edges.
02754   // We use this value when we vectorize all of the instructions that use the
02755   // PHI. Next, after all of the instructions in the block are complete we
02756   // add the new incoming edges to the PHI. At this point all of the
02757   // instructions in the basic block are vectorized, so we can use them to
02758   // construct the PHI.
02759   PhiVector RdxPHIsToFix;
02760 
02761   // Scan the loop in a topological order to ensure that defs are vectorized
02762   // before users.
02763   LoopBlocksDFS DFS(OrigLoop);
02764   DFS.perform(LI);
02765 
02766   // Vectorize all of the blocks in the original loop.
02767   for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
02768        be = DFS.endRPO(); bb != be; ++bb)
02769     vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
02770 
02771   // At this point every instruction in the original loop is widened to
02772   // a vector form. We are almost done. Now, we need to fix the PHI nodes
02773   // that we vectorized. The PHI nodes are currently empty because we did
02774   // not want to introduce cycles. Notice that the remaining PHI nodes
02775   // that we need to fix are reduction variables.
02776 
02777   // Create the 'reduced' values for each of the induction vars.
02778   // The reduced values are the vector values that we scalarize and combine
02779   // after the loop is finished.
02780   for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
02781        it != e; ++it) {
02782     PHINode *RdxPhi = *it;
02783     assert(RdxPhi && "Unable to recover vectorized PHI");
02784 
02785     // Find the reduction variable descriptor.
02786     assert(Legal->getReductionVars()->count(RdxPhi) &&
02787            "Unable to find the reduction variable");
02788     LoopVectorizationLegality::ReductionDescriptor RdxDesc =
02789     (*Legal->getReductionVars())[RdxPhi];
02790 
02791     setDebugLocFromInst(Builder, RdxDesc.StartValue);
02792 
02793     // We need to generate a reduction vector from the incoming scalar.
02794     // To do so, we need to generate the 'identity' vector and override
02795     // one of the elements with the incoming scalar reduction. We need
02796     // to do it in the vector-loop preheader.
02797     Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
02798 
02799     // This is the vector-clone of the value that leaves the loop.
02800     VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
02801     Type *VecTy = VectorExit[0]->getType();
02802 
02803     // Find the reduction identity variable. Zero for addition, or, xor,
02804     // one for multiplication, -1 for And.
02805     Value *Identity;
02806     Value *VectorStart;
02807     if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
02808         RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
02809       // MinMax reduction have the start value as their identify.
02810       if (VF == 1) {
02811         VectorStart = Identity = RdxDesc.StartValue;
02812       } else {
02813         VectorStart = Identity = Builder.CreateVectorSplat(VF,
02814                                                            RdxDesc.StartValue,
02815                                                            "minmax.ident");
02816       }
02817     } else {
02818       // Handle other reduction kinds:
02819       Constant *Iden =
02820       LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
02821                                                       VecTy->getScalarType());
02822       if (VF == 1) {
02823         Identity = Iden;
02824         // This vector is the Identity vector where the first element is the
02825         // incoming scalar reduction.
02826         VectorStart = RdxDesc.StartValue;
02827       } else {
02828         Identity = ConstantVector::getSplat(VF, Iden);
02829 
02830         // This vector is the Identity vector where the first element is the
02831         // incoming scalar reduction.
02832         VectorStart = Builder.CreateInsertElement(Identity,
02833                                                   RdxDesc.StartValue, Zero);
02834       }
02835     }
02836 
02837     // Fix the vector-loop phi.
02838     // We created the induction variable so we know that the
02839     // preheader is the first entry.
02840     BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
02841 
02842     // Reductions do not have to start at zero. They can start with
02843     // any loop invariant values.
02844     VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
02845     BasicBlock *Latch = OrigLoop->getLoopLatch();
02846     Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
02847     VectorParts &Val = getVectorValue(LoopVal);
02848     for (unsigned part = 0; part < UF; ++part) {
02849       // Make sure to add the reduction stat value only to the
02850       // first unroll part.
02851       Value *StartVal = (part == 0) ? VectorStart : Identity;
02852       cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
02853       cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
02854                                                   LoopVectorBody.back());
02855     }
02856 
02857     // Before each round, move the insertion point right between
02858     // the PHIs and the values we are going to write.
02859     // This allows us to write both PHINodes and the extractelement
02860     // instructions.
02861     Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
02862 
02863     VectorParts RdxParts;
02864     setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
02865     for (unsigned part = 0; part < UF; ++part) {
02866       // This PHINode contains the vectorized reduction variable, or
02867       // the initial value vector, if we bypass the vector loop.
02868       VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
02869       PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
02870       Value *StartVal = (part == 0) ? VectorStart : Identity;
02871       for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
02872         NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
02873       NewPhi->addIncoming(RdxExitVal[part],
02874                           LoopVectorBody.back());
02875       RdxParts.push_back(NewPhi);
02876     }
02877 
02878     // Reduce all of the unrolled parts into a single vector.
02879     Value *ReducedPartRdx = RdxParts[0];
02880     unsigned Op = getReductionBinOp(RdxDesc.Kind);
02881     setDebugLocFromInst(Builder, ReducedPartRdx);
02882     for (unsigned part = 1; part < UF; ++part) {
02883       if (Op != Instruction::ICmp && Op != Instruction::FCmp)
02884         // Floating point operations had to be 'fast' to enable the reduction.
02885         ReducedPartRdx = addFastMathFlag(
02886             Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
02887                                 ReducedPartRdx, "bin.rdx"));
02888       else
02889         ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
02890                                         ReducedPartRdx, RdxParts[part]);
02891     }
02892 
02893     if (VF > 1) {
02894       // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
02895       // and vector ops, reducing the set of values being computed by half each
02896       // round.
02897       assert(isPowerOf2_32(VF) &&
02898              "Reduction emission only supported for pow2 vectors!");
02899       Value *TmpVec = ReducedPartRdx;
02900       SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
02901       for (unsigned i = VF; i != 1; i >>= 1) {
02902         // Move the upper half of the vector to the lower half.
02903         for (unsigned j = 0; j != i/2; ++j)
02904           ShuffleMask[j] = Builder.getInt32(i/2 + j);
02905 
02906         // Fill the rest of the mask with undef.
02907         std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
02908                   UndefValue::get(Builder.getInt32Ty()));
02909 
02910         Value *Shuf =
02911         Builder.CreateShuffleVector(TmpVec,
02912                                     UndefValue::get(TmpVec->getType()),
02913                                     ConstantVector::get(ShuffleMask),
02914                                     "rdx.shuf");
02915 
02916         if (Op != Instruction::ICmp && Op != Instruction::FCmp)
02917           // Floating point operations had to be 'fast' to enable the reduction.
02918           TmpVec = addFastMathFlag(Builder.CreateBinOp(
02919               (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
02920         else
02921           TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
02922       }
02923 
02924       // The result is in the first element of the vector.
02925       ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
02926                                                     Builder.getInt32(0));
02927     }
02928 
02929     // Create a phi node that merges control-flow from the backedge-taken check
02930     // block and the middle block.
02931     PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
02932                                           LoopScalarPreHeader->getTerminator());
02933     BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
02934     BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
02935 
02936     // Now, we need to fix the users of the reduction variable
02937     // inside and outside of the scalar remainder loop.
02938     // We know that the loop is in LCSSA form. We need to update the
02939     // PHI nodes in the exit blocks.
02940     for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
02941          LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
02942       PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
02943       if (!LCSSAPhi) break;
02944 
02945       // All PHINodes need to have a single entry edge, or two if
02946       // we already fixed them.
02947       assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
02948 
02949       // We found our reduction value exit-PHI. Update it with the
02950       // incoming bypass edge.
02951       if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
02952         // Add an edge coming from the bypass.
02953         LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
02954         break;
02955       }
02956     }// end of the LCSSA phi scan.
02957 
02958     // Fix the scalar loop reduction variable with the incoming reduction sum
02959     // from the vector body and from the backedge value.
02960     int IncomingEdgeBlockIdx =
02961     (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
02962     assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
02963     // Pick the other block.
02964     int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
02965     (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
02966     (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
02967   }// end of for each redux variable.
02968 
02969   fixLCSSAPHIs();
02970 
02971   // Remove redundant induction instructions.
02972   cse(LoopVectorBody);
02973 }
02974 
02975 void InnerLoopVectorizer::fixLCSSAPHIs() {
02976   for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
02977        LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
02978     PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
02979     if (!LCSSAPhi) break;
02980     if (LCSSAPhi->getNumIncomingValues() == 1)
02981       LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
02982                             LoopMiddleBlock);
02983   }
02984 }
02985 
02986 InnerLoopVectorizer::VectorParts
02987 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
02988   assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
02989          "Invalid edge");
02990 
02991   // Look for cached value.
02992   std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
02993   EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
02994   if (ECEntryIt != MaskCache.end())
02995     return ECEntryIt->second;
02996 
02997   VectorParts SrcMask = createBlockInMask(Src);
02998 
02999   // The terminator has to be a branch inst!
03000   BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
03001   assert(BI && "Unexpected terminator found");
03002 
03003   if (BI->isConditional()) {
03004     VectorParts EdgeMask = getVectorValue(BI->getCondition());
03005 
03006     if (BI->getSuccessor(0) != Dst)
03007       for (unsigned part = 0; part < UF; ++part)
03008         EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
03009 
03010     for (unsigned part = 0; part < UF; ++part)
03011       EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
03012 
03013     MaskCache[Edge] = EdgeMask;
03014     return EdgeMask;
03015   }
03016 
03017   MaskCache[Edge] = SrcMask;
03018   return SrcMask;
03019 }
03020 
03021 InnerLoopVectorizer::VectorParts
03022 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
03023   assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
03024 
03025   // Loop incoming mask is all-one.
03026   if (OrigLoop->getHeader() == BB) {
03027     Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
03028     return getVectorValue(C);
03029   }
03030 
03031   // This is the block mask. We OR all incoming edges, and with zero.
03032   Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
03033   VectorParts BlockMask = getVectorValue(Zero);
03034 
03035   // For each pred:
03036   for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
03037     VectorParts EM = createEdgeMask(*it, BB);
03038     for (unsigned part = 0; part < UF; ++part)
03039       BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
03040   }
03041 
03042   return BlockMask;
03043 }
03044 
03045 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
03046                                               InnerLoopVectorizer::VectorParts &Entry,
03047                                               unsigned UF, unsigned VF, PhiVector *PV) {
03048   PHINode* P = cast<PHINode>(PN);
03049   // Handle reduction variables:
03050   if (Legal->getReductionVars()->count(P)) {
03051     for (unsigned part = 0; part < UF; ++part) {
03052       // This is phase one of vectorizing PHIs.
03053       Type *VecTy = (VF == 1) ? PN->getType() :
03054       VectorType::get(PN->getType(), VF);
03055       Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
03056                                     LoopVectorBody.back()-> getFirstInsertionPt());
03057     }
03058     PV->push_back(P);
03059     return;
03060   }
03061 
03062   setDebugLocFromInst(Builder, P);
03063   // Check for PHI nodes that are lowered to vector selects.
03064   if (P->getParent() != OrigLoop->getHeader()) {
03065     // We know that all PHIs in non-header blocks are converted into
03066     // selects, so we don't have to worry about the insertion order and we
03067     // can just use the builder.
03068     // At this point we generate the predication tree. There may be
03069     // duplications since this is a simple recursive scan, but future
03070     // optimizations will clean it up.
03071 
03072     unsigned NumIncoming = P->getNumIncomingValues();
03073 
03074     // Generate a sequence of selects of the form:
03075     // SELECT(Mask3, In3,
03076     //      SELECT(Mask2, In2,
03077     //                   ( ...)))
03078     for (unsigned In = 0; In < NumIncoming; In++) {
03079       VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
03080                                         P->getParent());
03081       VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
03082 
03083       for (unsigned part = 0; part < UF; ++part) {
03084         // We might have single edge PHIs (blocks) - use an identity
03085         // 'select' for the first PHI operand.
03086         if (In == 0)
03087           Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
03088                                              In0[part]);
03089         else
03090           // Select between the current value and the previous incoming edge
03091           // based on the incoming mask.
03092           Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
03093                                              Entry[part], "predphi");
03094       }
03095     }
03096     return;
03097   }
03098 
03099   // This PHINode must be an induction variable.
03100   // Make sure that we know about it.
03101   assert(Legal->getInductionVars()->count(P) &&
03102          "Not an induction variable");
03103 
03104   LoopVectorizationLegality::InductionInfo II =
03105   Legal->getInductionVars()->lookup(P);
03106 
03107   switch (II.IK) {
03108     case LoopVectorizationLegality::IK_NoInduction:
03109       llvm_unreachable("Unknown induction");
03110     case LoopVectorizationLegality::IK_IntInduction: {
03111       assert(P->getType() == II.StartValue->getType() && "Types must match");
03112       Type *PhiTy = P->getType();
03113       Value *Broadcasted;
03114       if (P == OldInduction) {
03115         // Handle the canonical induction variable. We might have had to
03116         // extend the type.
03117         Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
03118       } else {
03119         // Handle other induction variables that are now based on the
03120         // canonical one.
03121         Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
03122                                                  "normalized.idx");
03123         NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
03124         Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
03125                                         "offset.idx");
03126       }
03127       Broadcasted = getBroadcastInstrs(Broadcasted);
03128       // After broadcasting the induction variable we need to make the vector
03129       // consecutive by adding 0, 1, 2, etc.
03130       for (unsigned part = 0; part < UF; ++part)
03131         Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
03132       return;
03133     }
03134     case LoopVectorizationLegality::IK_ReverseIntInduction:
03135     case LoopVectorizationLegality::IK_PtrInduction:
03136     case LoopVectorizationLegality::IK_ReversePtrInduction:
03137       // Handle reverse integer and pointer inductions.
03138       Value *StartIdx = ExtendedIdx;
03139       // This is the normalized GEP that starts counting at zero.
03140       Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
03141                                                "normalized.idx");
03142 
03143       // Handle the reverse integer induction variable case.
03144       if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
03145         IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
03146         Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
03147                                                "resize.norm.idx");
03148         Value *ReverseInd  = Builder.CreateSub(II.StartValue, CNI,
03149                                                "reverse.idx");
03150 
03151         // This is a new value so do not hoist it out.
03152         Value *Broadcasted = getBroadcastInstrs(ReverseInd);
03153         // After broadcasting the induction variable we need to make the
03154         // vector consecutive by adding  ... -3, -2, -1, 0.
03155         for (unsigned part = 0; part < UF; ++part)
03156           Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
03157                                              true);
03158         return;
03159       }
03160 
03161       // Handle the pointer induction variable case.
03162       assert(P->getType()->isPointerTy() && "Unexpected type.");
03163 
03164       // Is this a reverse induction ptr or a consecutive induction ptr.
03165       bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
03166                       II.IK);
03167 
03168       // This is the vector of results. Notice that we don't generate
03169       // vector geps because scalar geps result in better code.
03170       for (unsigned part = 0; part < UF; ++part) {
03171         if (VF == 1) {
03172           int EltIndex = (part) * (Reverse ? -1 : 1);
03173           Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
03174           Value *GlobalIdx;
03175           if (Reverse)
03176             GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
03177           else
03178             GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
03179 
03180           Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
03181                                              "next.gep");
03182           Entry[part] = SclrGep;
03183           continue;
03184         }
03185 
03186         Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
03187         for (unsigned int i = 0; i < VF; ++i) {
03188           int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
03189           Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
03190           Value *GlobalIdx;
03191           if (!Reverse)
03192             GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
03193           else
03194             GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
03195 
03196           Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
03197                                              "next.gep");
03198           VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
03199                                                Builder.getInt32(i),
03200                                                "insert.gep");
03201         }
03202         Entry[part] = VecVal;
03203       }
03204       return;
03205   }
03206 }
03207 
03208 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
03209   // For each instruction in the old loop.
03210   for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
03211     VectorParts &Entry = WidenMap.get(it);
03212     switch (it->getOpcode()) {
03213     case Instruction::Br:
03214       // Nothing to do for PHIs and BR, since we already took care of the
03215       // loop control flow instructions.
03216       continue;
03217     case Instruction::PHI:{
03218       // Vectorize PHINodes.
03219       widenPHIInstruction(it, Entry, UF, VF, PV);
03220       continue;
03221     }// End of PHI.
03222 
03223     case Instruction::Add:
03224     case Instruction::FAdd:
03225     case Instruction::Sub:
03226     case Instruction::FSub:
03227     case Instruction::Mul:
03228     case Instruction::FMul:
03229     case Instruction::UDiv:
03230     case Instruction::SDiv:
03231     case Instruction::FDiv:
03232     case Instruction::URem:
03233     case Instruction::SRem:
03234     case Instruction::FRem:
03235     case Instruction::Shl:
03236     case Instruction::LShr:
03237     case Instruction::AShr:
03238     case Instruction::And:
03239     case Instruction::Or:
03240     case Instruction::Xor: {
03241       // Just widen binops.
03242       BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
03243       setDebugLocFromInst(Builder, BinOp);
03244       VectorParts &A = getVectorValue(it->getOperand(0));
03245       VectorParts &B = getVectorValue(it->getOperand(1));
03246 
03247       // Use this vector value for all users of the original instruction.
03248       for (unsigned Part = 0; Part < UF; ++Part) {
03249         Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
03250 
03251         if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
03252           VecOp->copyIRFlags(BinOp);
03253 
03254         Entry[Part] = V;
03255       }
03256 
03257       propagateMetadata(Entry, it);
03258       break;
03259     }
03260     case Instruction::Select: {
03261       // Widen selects.
03262       // If the selector is loop invariant we can create a select
03263       // instruction with a scalar condition. Otherwise, use vector-select.
03264       bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
03265                                                OrigLoop);
03266       setDebugLocFromInst(Builder, it);
03267 
03268       // The condition can be loop invariant  but still defined inside the
03269       // loop. This means that we can't just use the original 'cond' value.
03270       // We have to take the 'vectorized' value and pick the first lane.
03271       // Instcombine will make this a no-op.
03272       VectorParts &Cond = getVectorValue(it->getOperand(0));
03273       VectorParts &Op0  = getVectorValue(it->getOperand(1));
03274       VectorParts &Op1  = getVectorValue(it->getOperand(2));
03275 
03276       Value *ScalarCond = (VF == 1) ? Cond[0] :
03277         Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
03278 
03279       for (unsigned Part = 0; Part < UF; ++Part) {
03280         Entry[Part] = Builder.CreateSelect(
03281           InvariantCond ? ScalarCond : Cond[Part],
03282           Op0[Part],
03283           Op1[Part]);
03284       }
03285 
03286       propagateMetadata(Entry, it);
03287       break;
03288     }
03289 
03290     case Instruction::ICmp:
03291     case Instruction::FCmp: {
03292       // Widen compares. Generate vector compares.
03293       bool FCmp = (it->getOpcode() == Instruction::FCmp);
03294       CmpInst *Cmp = dyn_cast<CmpInst>(it);
03295       setDebugLocFromInst(Builder, it);
03296       VectorParts &A = getVectorValue(it->getOperand(0));
03297       VectorParts &B = getVectorValue(it->getOperand(1));
03298       for (unsigned Part = 0; Part < UF; ++Part) {
03299         Value *C = nullptr;
03300         if (FCmp)
03301           C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
03302         else
03303           C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
03304         Entry[Part] = C;
03305       }
03306 
03307       propagateMetadata(Entry, it);
03308       break;
03309     }
03310 
03311     case Instruction::Store:
03312     case Instruction::Load:
03313       vectorizeMemoryInstruction(it);
03314         break;
03315     case Instruction::ZExt:
03316     case Instruction::SExt:
03317     case Instruction::FPToUI:
03318     case Instruction::FPToSI:
03319     case Instruction::FPExt:
03320     case Instruction::PtrToInt:
03321     case Instruction::IntToPtr:
03322     case Instruction::SIToFP:
03323     case Instruction::UIToFP:
03324     case Instruction::Trunc:
03325     case Instruction::FPTrunc:
03326     case Instruction::BitCast: {
03327       CastInst *CI = dyn_cast<CastInst>(it);
03328       setDebugLocFromInst(Builder, it);
03329       /// Optimize the special case where the source is the induction
03330       /// variable. Notice that we can only optimize the 'trunc' case
03331       /// because: a. FP conversions lose precision, b. sext/zext may wrap,
03332       /// c. other casts depend on pointer size.
03333       if (CI->getOperand(0) == OldInduction &&
03334           it->getOpcode() == Instruction::Trunc) {
03335         Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
03336                                                CI->getType());
03337         Value *Broadcasted = getBroadcastInstrs(ScalarCast);
03338         for (unsigned Part = 0; Part < UF; ++Part)
03339           Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
03340         propagateMetadata(Entry, it);
03341         break;
03342       }
03343       /// Vectorize casts.
03344       Type *DestTy = (VF == 1) ? CI->getType() :
03345                                  VectorType::get(CI->getType(), VF);
03346 
03347       VectorParts &A = getVectorValue(it->getOperand(0));
03348       for (unsigned Part = 0; Part < UF; ++Part)
03349         Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
03350       propagateMetadata(Entry, it);
03351       break;
03352     }
03353 
03354     case Instruction::Call: {
03355       // Ignore dbg intrinsics.
03356       if (isa<DbgInfoIntrinsic>(it))
03357         break;
03358       setDebugLocFromInst(Builder, it);
03359 
03360       Module *M = BB->getParent()->getParent();
03361       CallInst *CI = cast<CallInst>(it);
03362       Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
03363       assert(ID && "Not an intrinsic call!");
03364       switch (ID) {
03365       case Intrinsic::assume:
03366       case Intrinsic::lifetime_end:
03367       case Intrinsic::lifetime_start:
03368         scalarizeInstruction(it);
03369         break;
03370       default:
03371         bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
03372         for (unsigned Part = 0; Part < UF; ++Part) {
03373           SmallVector<Value *, 4> Args;
03374           for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
03375             if (HasScalarOpd && i == 1) {
03376               Args.push_back(CI->getArgOperand(i));
03377               continue;
03378             }
03379             VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
03380             Args.push_back(Arg[Part]);
03381           }
03382           Type *Tys[] = {CI->getType()};
03383           if (VF > 1)
03384             Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
03385 
03386           Function *F = Intrinsic::getDeclaration(M, ID, Tys);
03387           Entry[Part] = Builder.CreateCall(F, Args);
03388         }
03389 
03390         propagateMetadata(Entry, it);
03391         break;
03392       }
03393       break;
03394     }
03395 
03396     default:
03397       // All other instructions are unsupported. Scalarize them.
03398       scalarizeInstruction(it);
03399       break;
03400     }// end of switch.
03401   }// end of for_each instr.
03402 }
03403 
03404 void InnerLoopVectorizer::updateAnalysis() {
03405   // Forget the original basic block.
03406   SE->forgetLoop(OrigLoop);
03407 
03408   // Update the dominator tree information.
03409   assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
03410          "Entry does not dominate exit.");
03411 
03412   for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
03413     DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
03414   DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
03415 
03416   // Due to if predication of stores we might create a sequence of "if(pred)
03417   // a[i] = ...;  " blocks.
03418   for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
03419     if (i == 0)
03420       DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
03421     else if (isPredicatedBlock(i)) {
03422       DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
03423     } else {
03424       DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
03425     }
03426   }
03427 
03428   DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
03429   DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
03430   DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
03431   DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
03432 
03433   DEBUG(DT->verifyDomTree());
03434 }
03435 
03436 /// \brief Check whether it is safe to if-convert this phi node.
03437 ///
03438 /// Phi nodes with constant expressions that can trap are not safe to if
03439 /// convert.
03440 static bool canIfConvertPHINodes(BasicBlock *BB) {
03441   for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
03442     PHINode *Phi = dyn_cast<PHINode>(I);
03443     if (!Phi)
03444       return true;
03445     for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
03446       if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
03447         if (C->canTrap())
03448           return false;
03449   }
03450   return true;
03451 }
03452 
03453 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
03454   if (!EnableIfConversion) {
03455     emitAnalysis(Report() << "if-conversion is disabled");
03456     return false;
03457   }
03458 
03459   assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
03460 
03461   // A list of pointers that we can safely read and write to.
03462   SmallPtrSet<Value *, 8> SafePointes;
03463 
03464   // Collect safe addresses.
03465   for (Loop::block_iterator BI = TheLoop->block_begin(),
03466          BE = TheLoop->block_end(); BI != BE; ++BI) {
03467     BasicBlock *BB = *BI;
03468 
03469     if (blockNeedsPredication(BB))
03470       continue;
03471 
03472     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
03473       if (LoadInst *LI = dyn_cast<LoadInst>(I))
03474         SafePointes.insert(LI->getPointerOperand());
03475       else if (StoreInst *SI = dyn_cast<StoreInst>(I))
03476         SafePointes.insert(SI->getPointerOperand());
03477     }
03478   }
03479 
03480   // Collect the blocks that need predication.
03481   BasicBlock *Header = TheLoop->getHeader();
03482   for (Loop::block_iterator BI = TheLoop->block_begin(),
03483          BE = TheLoop->block_end(); BI != BE; ++BI) {
03484     BasicBlock *BB = *BI;
03485 
03486     // We don't support switch statements inside loops.
03487     if (!isa<BranchInst>(BB->getTerminator())) {
03488       emitAnalysis(Report(BB->getTerminator())
03489                    << "loop contains a switch statement");
03490       return false;
03491     }
03492 
03493     // We must be able to predicate all blocks that need to be predicated.
03494     if (blockNeedsPredication(BB)) {
03495       if (!blockCanBePredicated(BB, SafePointes)) {
03496         emitAnalysis(Report(BB->getTerminator())
03497                      << "control flow cannot be substituted for a select");
03498         return false;
03499       }
03500     } else if (BB != Header && !canIfConvertPHINodes(BB)) {
03501       emitAnalysis(Report(BB->getTerminator())
03502                    << "control flow cannot be substituted for a select");
03503       return false;
03504     }
03505   }
03506 
03507   // We can if-convert this loop.
03508   return true;
03509 }
03510 
03511 bool LoopVectorizationLegality::canVectorize() {
03512   // We must have a loop in canonical form. Loops with indirectbr in them cannot
03513   // be canonicalized.
03514   if (!TheLoop->getLoopPreheader()) {
03515     emitAnalysis(
03516         Report() << "loop control flow is not understood by vectorizer");
03517     return false;
03518   }
03519 
03520   // We can only vectorize innermost loops.
03521   if (TheLoop->getSubLoopsVector().size()) {
03522     emitAnalysis(Report() << "loop is not the innermost loop");
03523     return false;
03524   }
03525 
03526   // We must have a single backedge.
03527   if (TheLoop->getNumBackEdges() != 1) {
03528     emitAnalysis(
03529         Report() << "loop control flow is not understood by vectorizer");
03530     return false;
03531   }
03532 
03533   // We must have a single exiting block.
03534   if (!TheLoop->getExitingBlock()) {
03535     emitAnalysis(
03536         Report() << "loop control flow is not understood by vectorizer");
03537     return false;
03538   }
03539 
03540   // We need to have a loop header.
03541   DEBUG(dbgs() << "LV: Found a loop: " <<
03542         TheLoop->getHeader()->getName() << '\n');
03543 
03544   // Check if we can if-convert non-single-bb loops.
03545   unsigned NumBlocks = TheLoop->getNumBlocks();
03546   if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
03547     DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
03548     return false;
03549   }
03550 
03551   // ScalarEvolution needs to be able to find the exit count.
03552   const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
03553   if (ExitCount == SE->getCouldNotCompute()) {
03554     emitAnalysis(Report() << "could not determine number of loop iterations");
03555     DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
03556     return false;
03557   }
03558 
03559   // Check if we can vectorize the instructions and CFG in this loop.
03560   if (!canVectorizeInstrs()) {
03561     DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
03562     return false;
03563   }
03564 
03565   // Go over each instruction and look at memory deps.
03566   if (!canVectorizeMemory()) {
03567     DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
03568     return false;
03569   }
03570 
03571   // Collect all of the variables that remain uniform after vectorization.
03572   collectLoopUniforms();
03573 
03574   DEBUG(dbgs() << "LV: We can vectorize this loop" <<
03575         (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
03576         <<"!\n");
03577 
03578   // Okay! We can vectorize. At this point we don't have any other mem analysis
03579   // which may limit our maximum vectorization factor, so just return true with
03580   // no restrictions.
03581   return true;
03582 }
03583 
03584 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
03585   if (Ty->isPointerTy())
03586     return DL.getIntPtrType(Ty);
03587 
03588   // It is possible that char's or short's overflow when we ask for the loop's
03589   // trip count, work around this by changing the type size.
03590   if (Ty->getScalarSizeInBits() < 32)
03591     return Type::getInt32Ty(Ty->getContext());
03592 
03593   return Ty;
03594 }
03595 
03596 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
03597   Ty0 = convertPointerToIntegerType(DL, Ty0);
03598   Ty1 = convertPointerToIntegerType(DL, Ty1);
03599   if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
03600     return Ty0;
03601   return Ty1;
03602 }
03603 
03604 /// \brief Check that the instruction has outside loop users and is not an
03605 /// identified reduction variable.
03606 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
03607                                SmallPtrSetImpl<Value *> &Reductions) {
03608   // Reduction instructions are allowed to have exit users. All other
03609   // instructions must not have external users.
03610   if (!Reductions.count(Inst))
03611     //Check that all of the users of the loop are inside the BB.
03612     for (User *U : Inst->users()) {
03613       Instruction *UI = cast<Instruction>(U);
03614       // This user may be a reduction exit value.
03615       if (!TheLoop->contains(UI)) {
03616         DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
03617         return true;
03618       }
03619     }
03620   return false;
03621 }
03622 
03623 bool LoopVectorizationLegality::canVectorizeInstrs() {
03624   BasicBlock *PreHeader = TheLoop->getLoopPreheader();
03625   BasicBlock *Header = TheLoop->getHeader();
03626 
03627   // Look for the attribute signaling the absence of NaNs.
03628   Function &F = *Header->getParent();
03629   if (F.hasFnAttribute("no-nans-fp-math"))
03630     HasFunNoNaNAttr = F.getAttributes().getAttribute(
03631       AttributeSet::FunctionIndex,
03632       "no-nans-fp-math").getValueAsString() == "true";
03633 
03634   // For each block in the loop.
03635   for (Loop::block_iterator bb = TheLoop->block_begin(),
03636        be = TheLoop->block_end(); bb != be; ++bb) {
03637 
03638     // Scan the instructions in the block and look for hazards.
03639     for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
03640          ++it) {
03641 
03642       if (PHINode *Phi = dyn_cast<PHINode>(it)) {
03643         Type *PhiTy = Phi->getType();
03644         // Check that this PHI type is allowed.
03645         if (!PhiTy->isIntegerTy() &&
03646             !PhiTy->isFloatingPointTy() &&
03647             !PhiTy->isPointerTy()) {
03648           emitAnalysis(Report(it)
03649                        << "loop control flow is not understood by vectorizer");
03650           DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
03651           return false;
03652         }
03653 
03654         // If this PHINode is not in the header block, then we know that we
03655         // can convert it to select during if-conversion. No need to check if
03656         // the PHIs in this block are induction or reduction variables.
03657         if (*bb != Header) {
03658           // Check that this instruction has no outside users or is an
03659           // identified reduction value with an outside user.
03660           if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
03661             continue;
03662           emitAnalysis(Report(it) << "value could not be identified as "
03663                                      "an induction or reduction variable");
03664           return false;
03665         }
03666 
03667         // We only allow if-converted PHIs with more than two incoming values.
03668         if (Phi->getNumIncomingValues() != 2) {
03669           emitAnalysis(Report(it)
03670                        << "control flow not understood by vectorizer");
03671           DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
03672           return false;
03673         }
03674 
03675         // This is the value coming from the preheader.
03676         Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
03677         // Check if this is an induction variable.
03678         InductionKind IK = isInductionVariable(Phi);
03679 
03680         if (IK_NoInduction != IK) {
03681           // Get the widest type.
03682           if (!WidestIndTy)
03683             WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
03684           else
03685             WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
03686 
03687           // Int inductions are special because we only allow one IV.
03688           if (IK == IK_IntInduction) {
03689             // Use the phi node with the widest type as induction. Use the last
03690             // one if there are multiple (no good reason for doing this other
03691             // than it is expedient).
03692             if (!Induction || PhiTy == WidestIndTy)
03693               Induction = Phi;
03694           }
03695 
03696           DEBUG(dbgs() << "LV: Found an induction variable.\n");
03697           Inductions[Phi] = InductionInfo(StartValue, IK);
03698 
03699           // Until we explicitly handle the case of an induction variable with
03700           // an outside loop user we have to give up vectorizing this loop.
03701           if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
03702             emitAnalysis(Report(it) << "use of induction value outside of the "
03703                                        "loop is not handled by vectorizer");
03704             return false;
03705           }
03706 
03707           continue;
03708         }
03709 
03710         if (AddReductionVar(Phi, RK_IntegerAdd)) {
03711           DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
03712           continue;
03713         }
03714         if (AddReductionVar(Phi, RK_IntegerMult)) {
03715           DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
03716           continue;
03717         }
03718         if (AddReductionVar(Phi, RK_IntegerOr)) {
03719           DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
03720           continue;
03721         }
03722         if (AddReductionVar(Phi, RK_IntegerAnd)) {
03723           DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
03724           continue;
03725         }
03726         if (AddReductionVar(Phi, RK_IntegerXor)) {
03727           DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
03728           continue;
03729         }
03730         if (AddReductionVar(Phi, RK_IntegerMinMax)) {
03731           DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
03732           continue;
03733         }
03734         if (AddReductionVar(Phi, RK_FloatMult)) {
03735           DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
03736           continue;
03737         }
03738         if (AddReductionVar(Phi, RK_FloatAdd)) {
03739           DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
03740           continue;
03741         }
03742         if (AddReductionVar(Phi, RK_FloatMinMax)) {
03743           DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
03744                 "\n");
03745           continue;
03746         }
03747 
03748         emitAnalysis(Report(it) << "value that could not be identified as "
03749                                    "reduction is used outside the loop");
03750         DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
03751         return false;
03752       }// end of PHI handling
03753 
03754       // We still don't handle functions. However, we can ignore dbg intrinsic
03755       // calls and we do handle certain intrinsic and libm functions.
03756       CallInst *CI = dyn_cast<CallInst>(it);
03757       if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
03758         emitAnalysis(Report(it) << "call instruction cannot be vectorized");
03759         DEBUG(dbgs() << "LV: Found a call site.\n");
03760         return false;
03761       }
03762 
03763       // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
03764       // second argument is the same (i.e. loop invariant)
03765       if (CI &&
03766           hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
03767         if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
03768           emitAnalysis(Report(it)
03769                        << "intrinsic instruction cannot be vectorized");
03770           DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
03771           return false;
03772         }
03773       }
03774 
03775       // Check that the instruction return type is vectorizable.
03776       // Also, we can't vectorize extractelement instructions.
03777       if ((!VectorType::isValidElementType(it->getType()) &&
03778            !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
03779         emitAnalysis(Report(it)
03780                      << "instruction return type cannot be vectorized");
03781         DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
03782         return false;
03783       }
03784 
03785       // Check that the stored type is vectorizable.
03786       if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
03787         Type *T = ST->getValueOperand()->getType();
03788         if (!VectorType::isValidElementType(T)) {
03789           emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
03790           return false;
03791         }
03792         if (EnableMemAccessVersioning)
03793           collectStridedAcccess(ST);
03794       }
03795 
03796       if (EnableMemAccessVersioning)
03797         if (LoadInst *LI = dyn_cast<LoadInst>(it))
03798           collectStridedAcccess(LI);
03799 
03800       // Reduction instructions are allowed to have exit users.
03801       // All other instructions must not have external users.
03802       if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
03803         emitAnalysis(Report(it) << "value cannot be used outside the loop");
03804         return false;
03805       }
03806 
03807     } // next instr.
03808 
03809   }
03810 
03811   if (!Induction) {
03812     DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
03813     if (Inductions.empty()) {
03814       emitAnalysis(Report()
03815                    << "loop induction variable could not be identified");
03816       return false;
03817     }
03818   }
03819 
03820   return true;
03821 }
03822 
03823 ///\brief Remove GEPs whose indices but the last one are loop invariant and
03824 /// return the induction operand of the gep pointer.
03825 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
03826                                  const DataLayout *DL, Loop *Lp) {
03827   GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
03828   if (!GEP)
03829     return Ptr;
03830 
03831   unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
03832 
03833   // Check that all of the gep indices are uniform except for our induction
03834   // operand.
03835   for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
03836     if (i != InductionOperand &&
03837         !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
03838       return Ptr;
03839   return GEP->getOperand(InductionOperand);
03840 }
03841 
03842 ///\brief Look for a cast use of the passed value.
03843 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
03844   Value *UniqueCast = nullptr;
03845   for (User *U : Ptr->users()) {
03846     CastInst *CI = dyn_cast<CastInst>(U);
03847     if (CI && CI->getType() == Ty) {
03848       if (!UniqueCast)
03849         UniqueCast = CI;
03850       else
03851         return nullptr;
03852     }
03853   }
03854   return UniqueCast;
03855 }
03856 
03857 ///\brief Get the stride of a pointer access in a loop.
03858 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
03859 /// pointer to the Value, or null otherwise.
03860 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
03861                                    const DataLayout *DL, Loop *Lp) {
03862   const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
03863   if (!PtrTy || PtrTy->isAggregateType())
03864     return nullptr;
03865 
03866   // Try to remove a gep instruction to make the pointer (actually index at this
03867   // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
03868   // pointer, otherwise, we are analyzing the index.
03869   Value *OrigPtr = Ptr;
03870 
03871   // The size of the pointer access.
03872   int64_t PtrAccessSize = 1;
03873 
03874   Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
03875   const SCEV *V = SE->getSCEV(Ptr);
03876 
03877   if (Ptr != OrigPtr)
03878     // Strip off casts.
03879     while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
03880       V = C->getOperand();
03881 
03882   const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
03883   if (!S)
03884     return nullptr;
03885 
03886   V = S->getStepRecurrence(*SE);
03887   if (!V)
03888     return nullptr;
03889 
03890   // Strip off the size of access multiplication if we are still analyzing the
03891   // pointer.
03892   if (OrigPtr == Ptr) {
03893     DL->getTypeAllocSize(PtrTy->getElementType());
03894     if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
03895       if (M->getOperand(0)->getSCEVType() != scConstant)
03896         return nullptr;
03897 
03898       const APInt &APStepVal =
03899           cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
03900 
03901       // Huge step value - give up.
03902       if (APStepVal.getBitWidth() > 64)
03903         return nullptr;
03904 
03905       int64_t StepVal = APStepVal.getSExtValue();
03906       if (PtrAccessSize != StepVal)
03907         return nullptr;
03908       V = M->getOperand(1);
03909     }
03910   }
03911 
03912   // Strip off casts.
03913   Type *StripedOffRecurrenceCast = nullptr;
03914   if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
03915     StripedOffRecurrenceCast = C->getType();
03916     V = C->getOperand();
03917   }
03918 
03919   // Look for the loop invariant symbolic value.
03920   const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
03921   if (!U)
03922     return nullptr;
03923 
03924   Value *Stride = U->getValue();
03925   if (!Lp->isLoopInvariant(Stride))
03926     return nullptr;
03927 
03928   // If we have stripped off the recurrence cast we have to make sure that we
03929   // return the value that is used in this loop so that we can replace it later.
03930   if (StripedOffRecurrenceCast)
03931     Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
03932 
03933   return Stride;
03934 }
03935 
03936 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
03937   Value *Ptr = nullptr;
03938   if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
03939     Ptr = LI->getPointerOperand();
03940   else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
03941     Ptr = SI->getPointerOperand();
03942   else
03943     return;
03944 
03945   Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
03946   if (!Stride)
03947     return;
03948 
03949   DEBUG(dbgs() << "LV: Found a strided access that we can version");
03950   DEBUG(dbgs() << "  Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
03951   Strides[Ptr] = Stride;
03952   StrideSet.insert(Stride);
03953 }
03954 
03955 void LoopVectorizationLegality::collectLoopUniforms() {
03956   // We now know that the loop is vectorizable!
03957   // Collect variables that will remain uniform after vectorization.
03958   std::vector<Value*> Worklist;
03959   BasicBlock *Latch = TheLoop->getLoopLatch();
03960 
03961   // Start with the conditional branch and walk up the block.
03962   Worklist.push_back(Latch->getTerminator()->getOperand(0));
03963 
03964   // Also add all consecutive pointer values; these values will be uniform
03965   // after vectorization (and subsequent cleanup) and, until revectorization is
03966   // supported, all dependencies must also be uniform.
03967   for (Loop::block_iterator B = TheLoop->block_begin(),
03968        BE = TheLoop->block_end(); B != BE; ++B)
03969     for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
03970          I != IE; ++I)
03971       if (I->getType()->isPointerTy() && isConsecutivePtr(I))
03972         Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
03973 
03974   while (Worklist.size()) {
03975     Instruction *I = dyn_cast<Instruction>(Worklist.back());
03976     Worklist.pop_back();
03977 
03978     // Look at instructions inside this loop.
03979     // Stop when reaching PHI nodes.
03980     // TODO: we need to follow values all over the loop, not only in this block.
03981     if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
03982       continue;
03983 
03984     // This is a known uniform.
03985     Uniforms.insert(I);
03986 
03987     // Insert all operands.
03988     Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
03989   }
03990 }
03991 
03992 namespace {
03993 /// \brief Analyses memory accesses in a loop.
03994 ///
03995 /// Checks whether run time pointer checks are needed and builds sets for data
03996 /// dependence checking.
03997 class AccessAnalysis {
03998 public:
03999   /// \brief Read or write access location.
04000   typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
04001   typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
04002 
04003   /// \brief Set of potential dependent memory accesses.
04004   typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
04005 
04006   AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
04007     DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
04008 
04009   /// \brief Register a load  and whether it is only read from.
04010   void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
04011     Value *Ptr = const_cast<Value*>(Loc.Ptr);
04012     AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
04013     Accesses.insert(MemAccessInfo(Ptr, false));
04014     if (IsReadOnly)
04015       ReadOnlyPtr.insert(Ptr);
04016   }
04017 
04018   /// \brief Register a store.
04019   void addStore(AliasAnalysis::Location &Loc) {
04020     Value *Ptr = const_cast<Value*>(Loc.Ptr);
04021     AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
04022     Accesses.insert(MemAccessInfo(Ptr, true));
04023   }
04024 
04025   /// \brief Check whether we can check the pointers at runtime for
04026   /// non-intersection.
04027   bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
04028                        unsigned &NumComparisons, ScalarEvolution *SE,
04029                        Loop *TheLoop, ValueToValueMap &Strides,
04030                        bool ShouldCheckStride = false);
04031 
04032   /// \brief Goes over all memory accesses, checks whether a RT check is needed
04033   /// and builds sets of dependent accesses.
04034   void buildDependenceSets() {
04035     processMemAccesses();
04036   }
04037 
04038   bool isRTCheckNeeded() { return IsRTCheckNeeded; }
04039 
04040   bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
04041   void resetDepChecks() { CheckDeps.clear(); }
04042 
04043   MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
04044 
04045 private:
04046   typedef SetVector<MemAccessInfo> PtrAccessSet;
04047 
04048   /// \brief Go over all memory access and check whether runtime pointer checks
04049   /// are needed /// and build sets of dependency check candidates.
04050   void processMemAccesses();
04051 
04052   /// Set of all accesses.
04053   PtrAccessSet Accesses;
04054 
04055   /// Set of accesses that need a further dependence check.
04056   MemAccessInfoSet CheckDeps;
04057 
04058   /// Set of pointers that are read only.
04059   SmallPtrSet<Value*, 16> ReadOnlyPtr;
04060 
04061   const DataLayout *DL;
04062 
04063   /// An alias set tracker to partition the access set by underlying object and
04064   //intrinsic property (such as TBAA metadata).
04065   AliasSetTracker AST;
04066 
04067   /// Sets of potentially dependent accesses - members of one set share an
04068   /// underlying pointer. The set "CheckDeps" identfies which sets really need a
04069   /// dependence check.
04070   DepCandidates &DepCands;
04071 
04072   bool IsRTCheckNeeded;
04073 };
04074 
04075 } // end anonymous namespace
04076 
04077 /// \brief Check whether a pointer can participate in a runtime bounds check.
04078 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
04079                                 Value *Ptr) {
04080   const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
04081   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
04082   if (!AR)
04083     return false;
04084 
04085   return AR->isAffine();
04086 }
04087 
04088 /// \brief Check the stride of the pointer and ensure that it does not wrap in
04089 /// the address space.
04090 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
04091                         const Loop *Lp, ValueToValueMap &StridesMap);
04092 
04093 bool AccessAnalysis::canCheckPtrAtRT(
04094     LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
04095     unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
04096     ValueToValueMap &StridesMap, bool ShouldCheckStride) {
04097   // Find pointers with computable bounds. We are going to use this information
04098   // to place a runtime bound check.
04099   bool CanDoRT = true;
04100 
04101   bool IsDepCheckNeeded = isDependencyCheckNeeded();
04102   NumComparisons = 0;
04103 
04104   // We assign a consecutive id to access from different alias sets.
04105   // Accesses between different groups doesn't need to be checked.
04106   unsigned ASId = 1;
04107   for (auto &AS : AST) {
04108     unsigned NumReadPtrChecks = 0;
04109     unsigned NumWritePtrChecks = 0;
04110 
04111     // We assign consecutive id to access from different dependence sets.
04112     // Accesses within the same set don't need a runtime check.
04113     unsigned RunningDepId = 1;
04114     DenseMap<Value *, unsigned> DepSetId;
04115 
04116     for (auto A : AS) {
04117       Value *Ptr = A.getValue();
04118       bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
04119       MemAccessInfo Access(Ptr, IsWrite);
04120 
04121       if (IsWrite)
04122         ++NumWritePtrChecks;
04123       else
04124         ++NumReadPtrChecks;
04125 
04126       if (hasComputableBounds(SE, StridesMap, Ptr) &&
04127           // When we run after a failing dependency check we have to make sure we
04128           // don't have wrapping pointers.
04129           (!ShouldCheckStride ||
04130            isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
04131         // The id of the dependence set.
04132         unsigned DepId;
04133 
04134         if (IsDepCheckNeeded) {
04135           Value *Leader = DepCands.getLeaderValue(Access).getPointer();
04136           unsigned &LeaderId = DepSetId[Leader];
04137           if (!LeaderId)
04138             LeaderId = RunningDepId++;
04139           DepId = LeaderId;
04140         } else
04141           // Each access has its own dependence set.
04142           DepId = RunningDepId++;
04143 
04144         RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
04145 
04146         DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
04147       } else {
04148         CanDoRT = false;
04149       }
04150     }
04151 
04152     if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
04153       NumComparisons += 0; // Only one dependence set.
04154     else {
04155       NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
04156                                               NumWritePtrChecks - 1));
04157     }
04158 
04159     ++ASId;
04160   }
04161 
04162   // If the pointers that we would use for the bounds comparison have different
04163   // address spaces, assume the values aren't directly comparable, so we can't
04164   // use them for the runtime check. We also have to assume they could
04165   // overlap. In the future there should be metadata for whether address spaces
04166   // are disjoint.
04167   unsigned NumPointers = RtCheck.Pointers.size();
04168   for (unsigned i = 0; i < NumPointers; ++i) {
04169     for (unsigned j = i + 1; j < NumPointers; ++j) {
04170       // Only need to check pointers between two different dependency sets.
04171       if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
04172        continue;
04173       // Only need to check pointers in the same alias set.
04174       if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
04175         continue;
04176 
04177       Value *PtrI = RtCheck.Pointers[i];
04178       Value *PtrJ = RtCheck.Pointers[j];
04179 
04180       unsigned ASi = PtrI->getType()->getPointerAddressSpace();
04181       unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
04182       if (ASi != ASj) {
04183         DEBUG(dbgs() << "LV: Runtime check would require comparison between"
04184                        " different address spaces\n");
04185         return false;
04186       }
04187     }
04188   }
04189 
04190   return CanDoRT;
04191 }
04192 
04193 void AccessAnalysis::processMemAccesses() {
04194   // We process the set twice: first we process read-write pointers, last we
04195   // process read-only pointers. This allows us to skip dependence tests for
04196   // read-only pointers.
04197 
04198   DEBUG(dbgs() << "LV: Processing memory accesses...\n");
04199   DEBUG(dbgs() << "  AST: "; AST.dump());
04200   DEBUG(dbgs() << "LV:   Accesses:\n");
04201   DEBUG({
04202     for (auto A : Accesses)
04203       dbgs() << "\t" << *A.getPointer() << " (" <<
04204                 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
04205                                          "read-only" : "read")) << ")\n";
04206   });
04207 
04208   // The AliasSetTracker has nicely partitioned our pointers by metadata
04209   // compatibility and potential for underlying-object overlap. As a result, we
04210   // only need to check for potential pointer dependencies within each alias
04211   // set.
04212   for (auto &AS : AST) {
04213     // Note that both the alias-set tracker and the alias sets themselves used
04214     // linked lists internally and so the iteration order here is deterministic
04215     // (matching the original instruction order within each set).
04216 
04217     bool SetHasWrite = false;
04218 
04219     // Map of pointers to last access encountered.
04220     typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
04221     UnderlyingObjToAccessMap ObjToLastAccess;
04222 
04223     // Set of access to check after all writes have been processed.
04224     PtrAccessSet DeferredAccesses;
04225 
04226     // Iterate over each alias set twice, once to process read/write pointers,
04227     // and then to process read-only pointers.
04228     for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
04229       bool UseDeferred = SetIteration > 0;
04230       PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
04231 
04232       for (auto A : AS) {
04233         Value *Ptr = A.getValue();
04234         bool IsWrite = S.count(MemAccessInfo(Ptr, true));
04235 
04236         // If we're using the deferred access set, then it contains only reads.
04237         bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
04238         if (UseDeferred && !IsReadOnlyPtr)
04239           continue;
04240         // Otherwise, the pointer must be in the PtrAccessSet, either as a read
04241         // or a write.
04242         assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
04243                  S.count(MemAccessInfo(Ptr, false))) &&
04244                "Alias-set pointer not in the access set?");
04245 
04246         MemAccessInfo Access(Ptr, IsWrite);
04247         DepCands.insert(Access);
04248 
04249         // Memorize read-only pointers for later processing and skip them in the
04250         // first round (they need to be checked after we have seen all write
04251         // pointers). Note: we also mark pointer that are not consecutive as
04252         // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
04253         // the second check for "!IsWrite".
04254         if (!UseDeferred && IsReadOnlyPtr) {
04255           DeferredAccesses.insert(Access);
04256           continue;
04257         }
04258 
04259         // If this is a write - check other reads and writes for conflicts.  If
04260         // this is a read only check other writes for conflicts (but only if
04261         // there is no other write to the ptr - this is an optimization to
04262         // catch "a[i] = a[i] + " without having to do a dependence check).
04263         if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
04264           CheckDeps.insert(Access);
04265           IsRTCheckNeeded = true;
04266         }
04267 
04268         if (IsWrite)
04269           SetHasWrite = true;
04270 
04271         // Create sets of pointers connected by a shared alias set and
04272         // underlying object.
04273         typedef SmallVector<Value *, 16> ValueVector;
04274         ValueVector TempObjects;
04275         GetUnderlyingObjects(Ptr, TempObjects, DL);
04276         for (Value *UnderlyingObj : TempObjects) {
04277           UnderlyingObjToAccessMap::iterator Prev =
04278             ObjToLastAccess.find(UnderlyingObj);
04279           if (Prev != ObjToLastAccess.end())
04280             DepCands.unionSets(Access, Prev->second);
04281 
04282           ObjToLastAccess[UnderlyingObj] = Access;
04283         }
04284       }
04285     }
04286   }
04287 }
04288 
04289 namespace {
04290 /// \brief Checks memory dependences among accesses to the same underlying
04291 /// object to determine whether there vectorization is legal or not (and at
04292 /// which vectorization factor).
04293 ///
04294 /// This class works under the assumption that we already checked that memory
04295 /// locations with different underlying pointers are "must-not alias".
04296 /// We use the ScalarEvolution framework to symbolically evalutate access
04297 /// functions pairs. Since we currently don't restructure the loop we can rely
04298 /// on the program order of memory accesses to determine their safety.
04299 /// At the moment we will only deem accesses as safe for:
04300 ///  * A negative constant distance assuming program order.
04301 ///
04302 ///      Safe: tmp = a[i + 1];     OR     a[i + 1] = x;
04303 ///            a[i] = tmp;                y = a[i];
04304 ///
04305 ///   The latter case is safe because later checks guarantuee that there can't
04306 ///   be a cycle through a phi node (that is, we check that "x" and "y" is not
04307 ///   the same variable: a header phi can only be an induction or a reduction, a
04308 ///   reduction can't have a memory sink, an induction can't have a memory
04309 ///   source). This is important and must not be violated (or we have to
04310 ///   resort to checking for cycles through memory).
04311 ///
04312 ///  * A positive constant distance assuming program order that is bigger
04313 ///    than the biggest memory access.
04314 ///
04315 ///     tmp = a[i]        OR              b[i] = x
04316 ///     a[i+2] = tmp                      y = b[i+2];
04317 ///
04318 ///     Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
04319 ///
04320 ///  * Zero distances and all accesses have the same size.
04321 ///
04322 class MemoryDepChecker {
04323 public:
04324   typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
04325   typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
04326 
04327   MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
04328       : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
04329         ShouldRetryWithRuntimeCheck(false) {}
04330 
04331   /// \brief Register the location (instructions are given increasing numbers)
04332   /// of a write access.
04333   void addAccess(StoreInst *SI) {
04334     Value *Ptr = SI->getPointerOperand();
04335     Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
04336     InstMap.push_back(SI);
04337     ++AccessIdx;
04338   }
04339 
04340   /// \brief Register the location (instructions are given increasing numbers)
04341   /// of a write access.
04342   void addAccess(LoadInst *LI) {
04343     Value *Ptr = LI->getPointerOperand();
04344     Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
04345     InstMap.push_back(LI);
04346     ++AccessIdx;
04347   }
04348 
04349   /// \brief Check whether the dependencies between the accesses are safe.
04350   ///
04351   /// Only checks sets with elements in \p CheckDeps.
04352   bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
04353                    MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
04354 
04355   /// \brief The maximum number of bytes of a vector register we can vectorize
04356   /// the accesses safely with.
04357   unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
04358 
04359   /// \brief In same cases when the dependency check fails we can still
04360   /// vectorize the loop with a dynamic array access check.
04361   bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
04362 
04363 private:
04364   ScalarEvolution *SE;
04365   const DataLayout *DL;
04366   const Loop *InnermostLoop;
04367 
04368   /// \brief Maps access locations (ptr, read/write) to program order.
04369   DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
04370 
04371   /// \brief Memory access instructions in program order.
04372   SmallVector<Instruction *, 16> InstMap;
04373 
04374   /// \brief The program order index to be used for the next instruction.
04375   unsigned AccessIdx;
04376 
04377   // We can access this many bytes in parallel safely.
04378   unsigned MaxSafeDepDistBytes;
04379 
04380   /// \brief If we see a non-constant dependence distance we can still try to
04381   /// vectorize this loop with runtime checks.
04382   bool ShouldRetryWithRuntimeCheck;
04383 
04384   /// \brief Check whether there is a plausible dependence between the two
04385   /// accesses.
04386   ///
04387   /// Access \p A must happen before \p B in program order. The two indices
04388   /// identify the index into the program order map.
04389   ///
04390   /// This function checks  whether there is a plausible dependence (or the
04391   /// absence of such can't be proved) between the two accesses. If there is a
04392   /// plausible dependence but the dependence distance is bigger than one
04393   /// element access it records this distance in \p MaxSafeDepDistBytes (if this
04394   /// distance is smaller than any other distance encountered so far).
04395   /// Otherwise, this function returns true signaling a possible dependence.
04396   bool isDependent(const MemAccessInfo &A, unsigned AIdx,
04397                    const MemAccessInfo &B, unsigned BIdx,
04398                    ValueToValueMap &Strides);
04399 
04400   /// \brief Check whether the data dependence could prevent store-load
04401   /// forwarding.
04402   bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
04403 };
04404 
04405 } // end anonymous namespace
04406 
04407 static bool isInBoundsGep(Value *Ptr) {
04408   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
04409     return GEP->isInBounds();
04410   return false;
04411 }
04412 
04413 /// \brief Check whether the access through \p Ptr has a constant stride.
04414 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
04415                         const Loop *Lp, ValueToValueMap &StridesMap) {
04416   const Type *Ty = Ptr->getType();
04417   assert(Ty->isPointerTy() && "Unexpected non-ptr");
04418 
04419   // Make sure that the pointer does not point to aggregate types.
04420   const PointerType *PtrTy = cast<PointerType>(Ty);
04421   if (PtrTy->getElementType()->isAggregateType()) {
04422     DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
04423           "\n");
04424     return 0;
04425   }
04426 
04427   const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
04428 
04429   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
04430   if (!AR) {
04431     DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
04432           << *Ptr << " SCEV: " << *PtrScev << "\n");
04433     return 0;
04434   }
04435 
04436   // The accesss function must stride over the innermost loop.
04437   if (Lp != AR->getLoop()) {
04438     DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
04439           *Ptr << " SCEV: " << *PtrScev << "\n");
04440   }
04441 
04442   // The address calculation must not wrap. Otherwise, a dependence could be
04443   // inverted.
04444   // An inbounds getelementptr that is a AddRec with a unit stride
04445   // cannot wrap per definition. The unit stride requirement is checked later.
04446   // An getelementptr without an inbounds attribute and unit stride would have
04447   // to access the pointer value "0" which is undefined behavior in address
04448   // space 0, therefore we can also vectorize this case.
04449   bool IsInBoundsGEP = isInBoundsGep(Ptr);
04450   bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
04451   bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
04452   if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
04453     DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
04454           << *Ptr << " SCEV: " << *PtrScev << "\n");
04455     return 0;
04456   }
04457 
04458   // Check the step is constant.
04459   const SCEV *Step = AR->getStepRecurrence(*SE);
04460 
04461   // Calculate the pointer stride and check if it is consecutive.
04462   const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
04463   if (!C) {
04464     DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
04465           " SCEV: " << *PtrScev << "\n");
04466     return 0;
04467   }
04468 
04469   int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
04470   const APInt &APStepVal = C->getValue()->getValue();
04471 
04472   // Huge step value - give up.
04473   if (APStepVal.getBitWidth() > 64)
04474     return 0;
04475 
04476   int64_t StepVal = APStepVal.getSExtValue();
04477 
04478   // Strided access.
04479   int64_t Stride = StepVal / Size;
04480   int64_t Rem = StepVal % Size;
04481   if (Rem)
04482     return 0;
04483 
04484   // If the SCEV could wrap but we have an inbounds gep with a unit stride we
04485   // know we can't "wrap around the address space". In case of address space
04486   // zero we know that this won't happen without triggering undefined behavior.
04487   if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
04488       Stride != 1 && Stride != -1)
04489     return 0;
04490 
04491   return Stride;
04492 }
04493 
04494 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
04495                                                     unsigned TypeByteSize) {
04496   // If loads occur at a distance that is not a multiple of a feasible vector
04497   // factor store-load forwarding does not take place.
04498   // Positive dependences might cause troubles because vectorizing them might
04499   // prevent store-load forwarding making vectorized code run a lot slower.
04500   //   a[i] = a[i-3] ^ a[i-8];
04501   //   The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
04502   //   hence on your typical architecture store-load forwarding does not take
04503   //   place. Vectorizing in such cases does not make sense.
04504   // Store-load forwarding distance.
04505   const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
04506   // Maximum vector factor.
04507   unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
04508   if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
04509     MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
04510 
04511   for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
04512        vf *= 2) {
04513     if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
04514       MaxVFWithoutSLForwardIssues = (vf >>=1);
04515       break;
04516     }
04517   }
04518 
04519   if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
04520     DEBUG(dbgs() << "LV: Distance " << Distance <<
04521           " that could cause a store-load forwarding conflict\n");
04522     return true;
04523   }
04524 
04525   if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
04526       MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
04527     MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
04528   return false;
04529 }
04530 
04531 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
04532                                    const MemAccessInfo &B, unsigned BIdx,
04533                                    ValueToValueMap &Strides) {
04534   assert (AIdx < BIdx && "Must pass arguments in program order");
04535 
04536   Value *APtr = A.getPointer();
04537   Value *BPtr = B.getPointer();
04538   bool AIsWrite = A.getInt();
04539   bool BIsWrite = B.getInt();
04540 
04541   // Two reads are independent.
04542   if (!AIsWrite && !BIsWrite)
04543     return false;
04544 
04545   // We cannot check pointers in different address spaces.
04546   if (APtr->getType()->getPointerAddressSpace() !=
04547       BPtr->getType()->getPointerAddressSpace())
04548     return true;
04549 
04550   const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
04551   const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
04552 
04553   int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
04554   int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
04555 
04556   const SCEV *Src = AScev;
04557   const SCEV *Sink = BScev;
04558 
04559   // If the induction step is negative we have to invert source and sink of the
04560   // dependence.
04561   if (StrideAPtr < 0) {
04562     //Src = BScev;
04563     //Sink = AScev;
04564     std::swap(APtr, BPtr);
04565     std::swap(Src, Sink);
04566     std::swap(AIsWrite, BIsWrite);
04567     std::swap(AIdx, BIdx);
04568     std::swap(StrideAPtr, StrideBPtr);
04569   }
04570 
04571   const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
04572 
04573   DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
04574         << "(Induction step: " << StrideAPtr <<  ")\n");
04575   DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
04576         << *InstMap[BIdx] << ": " << *Dist << "\n");
04577 
04578   // Need consecutive accesses. We don't want to vectorize
04579   // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
04580   // the address space.
04581   if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
04582     DEBUG(dbgs() << "Non-consecutive pointer access\n");
04583     return true;
04584   }
04585 
04586   const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
04587   if (!C) {
04588     DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
04589     ShouldRetryWithRuntimeCheck = true;
04590     return true;
04591   }
04592 
04593   Type *ATy = APtr->getType()->getPointerElementType();
04594   Type *BTy = BPtr->getType()->getPointerElementType();
04595   unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
04596 
04597   // Negative distances are not plausible dependencies.
04598   const APInt &Val = C->getValue()->getValue();
04599   if (Val.isNegative()) {
04600     bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
04601     if (IsTrueDataDependence &&
04602         (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
04603          ATy != BTy))
04604       return true;
04605 
04606     DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
04607     return false;
04608   }
04609 
04610   // Write to the same location with the same size.
04611   // Could be improved to assert type sizes are the same (i32 == float, etc).
04612   if (Val == 0) {
04613     if (ATy == BTy)
04614       return false;
04615     DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
04616     return true;
04617   }
04618 
04619   assert(Val.isStrictlyPositive() && "Expect a positive value");
04620 
04621   // Positive distance bigger than max vectorization factor.
04622   if (ATy != BTy) {
04623     DEBUG(dbgs() <<
04624           "LV: ReadWrite-Write positive dependency with different types\n");
04625     return false;
04626   }
04627 
04628   unsigned Distance = (unsigned) Val.getZExtValue();
04629 
04630   // Bail out early if passed-in parameters make vectorization not feasible.
04631   unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
04632   unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
04633 
04634   // The distance must be bigger than the size needed for a vectorized version
04635   // of the operation and the size of the vectorized operation must not be
04636   // bigger than the currrent maximum size.
04637   if (Distance < 2*TypeByteSize ||
04638       2*TypeByteSize > MaxSafeDepDistBytes ||
04639       Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
04640     DEBUG(dbgs() << "LV: Failure because of Positive distance "
04641         << Val.getSExtValue() << '\n');
04642     return true;
04643   }
04644 
04645   MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
04646     Distance : MaxSafeDepDistBytes;
04647 
04648   bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
04649   if (IsTrueDataDependence &&
04650       couldPreventStoreLoadForward(Distance, TypeByteSize))
04651      return true;
04652 
04653   DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
04654         " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
04655 
04656   return false;
04657 }
04658 
04659 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
04660                                    MemAccessInfoSet &CheckDeps,
04661                                    ValueToValueMap &Strides) {
04662 
04663   MaxSafeDepDistBytes = -1U;
04664   while (!CheckDeps.empty()) {
04665     MemAccessInfo CurAccess = *CheckDeps.begin();
04666 
04667     // Get the relevant memory access set.
04668     EquivalenceClasses<MemAccessInfo>::iterator I =
04669       AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
04670 
04671     // Check accesses within this set.
04672     EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
04673     AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
04674 
04675     // Check every access pair.
04676     while (AI != AE) {
04677       CheckDeps.erase(*AI);
04678       EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
04679       while (OI != AE) {
04680         // Check every accessing instruction pair in program order.
04681         for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
04682              I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
04683           for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
04684                I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
04685             if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
04686               return false;
04687             if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
04688               return false;
04689           }
04690         ++OI;
04691       }
04692       AI++;
04693     }
04694   }
04695   return true;
04696 }
04697 
04698 bool LoopVectorizationLegality::canVectorizeMemory() {
04699 
04700   typedef SmallVector<Value*, 16> ValueVector;
04701   typedef SmallPtrSet<Value*, 16> ValueSet;
04702 
04703   // Holds the Load and Store *instructions*.
04704   ValueVector Loads;
04705   ValueVector Stores;
04706 
04707   // Holds all the different accesses in the loop.
04708   unsigned NumReads = 0;
04709   unsigned NumReadWrites = 0;
04710 
04711   PtrRtCheck.Pointers.clear();
04712   PtrRtCheck.Need = false;
04713 
04714   const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
04715   MemoryDepChecker DepChecker(SE, DL, TheLoop);
04716 
04717   // For each block.
04718   for (Loop::block_iterator bb = TheLoop->block_begin(),
04719        be = TheLoop->block_end(); bb != be; ++bb) {
04720 
04721     // Scan the BB and collect legal loads and stores.
04722     for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
04723          ++it) {
04724 
04725       // If this is a load, save it. If this instruction can read from memory
04726       // but is not a load, then we quit. Notice that we don't handle function
04727       // calls that read or write.
04728       if (it->mayReadFromMemory()) {
04729         // Many math library functions read the rounding mode. We will only
04730         // vectorize a loop if it contains known function calls that don't set
04731         // the flag. Therefore, it is safe to ignore this read from memory.
04732         CallInst *Call = dyn_cast<CallInst>(it);
04733         if (Call && getIntrinsicIDForCall(Call, TLI))
04734           continue;
04735 
04736         LoadInst *Ld = dyn_cast<LoadInst>(it);
04737         if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
04738           emitAnalysis(Report(Ld)
04739                        << "read with atomic ordering or volatile read");
04740           DEBUG(dbgs() << "LV: Found a non-simple load.\n");
04741           return false;
04742         }
04743         NumLoads++;
04744         Loads.push_back(Ld);
04745         DepChecker.addAccess(Ld);
04746         continue;
04747       }
04748 
04749       // Save 'store' instructions. Abort if other instructions write to memory.
04750       if (it->mayWriteToMemory()) {
04751         StoreInst *St = dyn_cast<StoreInst>(it);
04752         if (!St) {
04753           emitAnalysis(Report(it) << "instruction cannot be vectorized");
04754           return false;
04755         }
04756         if (!St->isSimple() && !IsAnnotatedParallel) {
04757           emitAnalysis(Report(St)
04758                        << "write with atomic ordering or volatile write");
04759           DEBUG(dbgs() << "LV: Found a non-simple store.\n");
04760           return false;
04761         }
04762         NumStores++;
04763         Stores.push_back(St);
04764         DepChecker.addAccess(St);
04765       }
04766     } // Next instr.
04767   } // Next block.
04768 
04769   // Now we have two lists that hold the loads and the stores.
04770   // Next, we find the pointers that they use.
04771 
04772   // Check if we see any stores. If there are no stores, then we don't
04773   // care if the pointers are *restrict*.
04774   if (!Stores.size()) {
04775     DEBUG(dbgs() << "LV: Found a read-only loop!\n");
04776     return true;
04777   }
04778 
04779   AccessAnalysis::DepCandidates DependentAccesses;
04780   AccessAnalysis Accesses(DL, AA, DependentAccesses);
04781 
04782   // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
04783   // multiple times on the same object. If the ptr is accessed twice, once
04784   // for read and once for write, it will only appear once (on the write
04785   // list). This is okay, since we are going to check for conflicts between
04786   // writes and between reads and writes, but not between reads and reads.
04787   ValueSet Seen;
04788 
04789   ValueVector::iterator I, IE;
04790   for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
04791     StoreInst *ST = cast<StoreInst>(*I);
04792     Value* Ptr = ST->getPointerOperand();
04793 
04794     if (isUniform(Ptr)) {
04795       emitAnalysis(
04796           Report(ST)
04797           << "write to a loop invariant address could not be vectorized");
04798       DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
04799       return false;
04800     }
04801 
04802     // If we did *not* see this pointer before, insert it to  the read-write
04803     // list. At this phase it is only a 'write' list.
04804     if (Seen.insert(Ptr)) {
04805       ++NumReadWrites;
04806 
04807       AliasAnalysis::Location Loc = AA->getLocation(ST);
04808       // The TBAA metadata could have a control dependency on the predication
04809       // condition, so we cannot rely on it when determining whether or not we
04810       // need runtime pointer checks.
04811       if (blockNeedsPredication(ST->getParent()))
04812         Loc.AATags.TBAA = nullptr;
04813 
04814       Accesses.addStore(Loc);
04815     }
04816   }
04817 
04818   if (IsAnnotatedParallel) {
04819     DEBUG(dbgs()
04820           << "LV: A loop annotated parallel, ignore memory dependency "
04821           << "checks.\n");
04822     return true;
04823   }
04824 
04825   for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
04826     LoadInst *LD = cast<LoadInst>(*I);
04827     Value* Ptr = LD->getPointerOperand();
04828     // If we did *not* see this pointer before, insert it to the
04829     // read list. If we *did* see it before, then it is already in
04830     // the read-write list. This allows us to vectorize expressions
04831     // such as A[i] += x;  Because the address of A[i] is a read-write
04832     // pointer. This only works if the index of A[i] is consecutive.
04833     // If the address of i is unknown (for example A[B[i]]) then we may
04834     // read a few words, modify, and write a few words, and some of the
04835     // words may be written to the same address.
04836     bool IsReadOnlyPtr = false;
04837     if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
04838       ++NumReads;
04839       IsReadOnlyPtr = true;
04840     }
04841 
04842     AliasAnalysis::Location Loc = AA->getLocation(LD);
04843     // The TBAA metadata could have a control dependency on the predication
04844     // condition, so we cannot rely on it when determining whether or not we
04845     // need runtime pointer checks.
04846     if (blockNeedsPredication(LD->getParent()))
04847       Loc.AATags.TBAA = nullptr;
04848 
04849     Accesses.addLoad(Loc, IsReadOnlyPtr);
04850   }
04851 
04852   // If we write (or read-write) to a single destination and there are no
04853   // other reads in this loop then is it safe to vectorize.
04854   if (NumReadWrites == 1 && NumReads == 0) {
04855     DEBUG(dbgs() << "LV: Found a write-only loop!\n");
04856     return true;
04857   }
04858 
04859   // Build dependence sets and check whether we need a runtime pointer bounds
04860   // check.
04861   Accesses.buildDependenceSets();
04862   bool NeedRTCheck = Accesses.isRTCheckNeeded();
04863 
04864   // Find pointers with computable bounds. We are going to use this information
04865   // to place a runtime bound check.
04866   unsigned NumComparisons = 0;
04867   bool CanDoRT = false;
04868   if (NeedRTCheck)
04869     CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
04870                                        Strides);
04871 
04872   DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
04873         " pointer comparisons.\n");
04874 
04875   // If we only have one set of dependences to check pointers among we don't
04876   // need a runtime check.
04877   if (NumComparisons == 0 && NeedRTCheck)
04878     NeedRTCheck = false;
04879 
04880   // Check that we did not collect too many pointers or found an unsizeable
04881   // pointer.
04882   if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
04883     PtrRtCheck.reset();
04884     CanDoRT = false;
04885   }
04886 
04887   if (CanDoRT) {
04888     DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
04889   }
04890 
04891   if (NeedRTCheck && !CanDoRT) {
04892     emitAnalysis(Report() << "cannot identify array bounds");
04893     DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
04894           "the array bounds.\n");
04895     PtrRtCheck.reset();
04896     return false;
04897   }
04898 
04899   PtrRtCheck.Need = NeedRTCheck;
04900 
04901   bool CanVecMem = true;
04902   if (Accesses.isDependencyCheckNeeded()) {
04903     DEBUG(dbgs() << "LV: Checking memory dependencies\n");
04904     CanVecMem = DepChecker.areDepsSafe(
04905         DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
04906     MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
04907 
04908     if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
04909       DEBUG(dbgs() << "LV: Retrying with memory checks\n");
04910       NeedRTCheck = true;
04911 
04912       // Clear the dependency checks. We assume they are not needed.
04913       Accesses.resetDepChecks();
04914 
04915       PtrRtCheck.reset();
04916       PtrRtCheck.Need = true;
04917 
04918       CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
04919                                          TheLoop, Strides, true);
04920       // Check that we did not collect too many pointers or found an unsizeable
04921       // pointer.
04922       if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
04923         if (!CanDoRT && NumComparisons > 0)
04924           emitAnalysis(Report()
04925                        << "cannot check memory dependencies at runtime");
04926         else
04927           emitAnalysis(Report()
04928                        << NumComparisons << " exceeds limit of "
04929                        << RuntimeMemoryCheckThreshold
04930                        << " dependent memory operations checked at runtime");
04931         DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
04932         PtrRtCheck.reset();
04933         return false;
04934       }
04935 
04936       CanVecMem = true;
04937     }
04938   }
04939 
04940   if (!CanVecMem)
04941     emitAnalysis(Report() << "unsafe dependent memory operations in loop");
04942 
04943   DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
04944         " need a runtime memory check.\n");
04945 
04946   return CanVecMem;
04947 }
04948 
04949 static bool hasMultipleUsesOf(Instruction *I,
04950                               SmallPtrSetImpl<Instruction *> &Insts) {
04951   unsigned NumUses = 0;
04952   for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
04953     if (Insts.count(dyn_cast<Instruction>(*Use)))
04954       ++NumUses;
04955     if (NumUses > 1)
04956       return true;
04957   }
04958 
04959   return false;
04960 }
04961 
04962 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
04963   for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
04964     if (!Set.count(dyn_cast<Instruction>(*Use)))
04965       return false;
04966   return true;
04967 }
04968 
04969 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
04970                                                 ReductionKind Kind) {
04971   if (Phi->getNumIncomingValues() != 2)
04972     return false;
04973 
04974   // Reduction variables are only found in the loop header block.
04975   if (Phi->getParent() != TheLoop->getHeader())
04976     return false;
04977 
04978   // Obtain the reduction start value from the value that comes from the loop
04979   // preheader.
04980   Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
04981 
04982   // ExitInstruction is the single value which is used outside the loop.
04983   // We only allow for a single reduction value to be used outside the loop.
04984   // This includes users of the reduction, variables (which form a cycle
04985   // which ends in the phi node).
04986   Instruction *ExitInstruction = nullptr;
04987   // Indicates that we found a reduction operation in our scan.
04988   bool FoundReduxOp = false;
04989 
04990   // We start with the PHI node and scan for all of the users of this
04991   // instruction. All users must be instructions that can be used as reduction
04992   // variables (such as ADD). We must have a single out-of-block user. The cycle
04993   // must include the original PHI.
04994   bool FoundStartPHI = false;
04995 
04996   // To recognize min/max patterns formed by a icmp select sequence, we store
04997   // the number of instruction we saw from the recognized min/max pattern,
04998   //  to make sure we only see exactly the two instructions.
04999   unsigned NumCmpSelectPatternInst = 0;
05000   ReductionInstDesc ReduxDesc(false, nullptr);
05001 
05002   SmallPtrSet<Instruction *, 8> VisitedInsts;
05003   SmallVector<Instruction *, 8> Worklist;
05004   Worklist.push_back(Phi);
05005   VisitedInsts.insert(Phi);
05006 
05007   // A value in the reduction can be used:
05008   //  - By the reduction:
05009   //      - Reduction operation:
05010   //        - One use of reduction value (safe).
05011   //        - Multiple use of reduction value (not safe).
05012   //      - PHI:
05013   //        - All uses of the PHI must be the reduction (safe).
05014   //        - Otherwise, not safe.
05015   //  - By one instruction outside of the loop (safe).
05016   //  - By further instructions outside of the loop (not safe).
05017   //  - By an instruction that is not part of the reduction (not safe).
05018   //    This is either:
05019   //      * An instruction type other than PHI or the reduction operation.
05020   //      * A PHI in the header other than the initial PHI.
05021   while (!Worklist.empty()) {
05022     Instruction *Cur = Worklist.back();
05023     Worklist.pop_back();
05024 
05025     // No Users.
05026     // If the instruction has no users then this is a broken chain and can't be
05027     // a reduction variable.
05028     if (Cur->use_empty())
05029       return false;
05030 
05031     bool IsAPhi = isa<PHINode>(Cur);
05032 
05033     // A header PHI use other than the original PHI.
05034     if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
05035       return false;
05036 
05037     // Reductions of instructions such as Div, and Sub is only possible if the
05038     // LHS is the reduction variable.
05039     if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
05040         !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
05041         !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
05042       return false;
05043 
05044     // Any reduction instruction must be of one of the allowed kinds.
05045     ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
05046     if (!ReduxDesc.IsReduction)
05047       return false;
05048 
05049     // A reduction operation must only have one use of the reduction value.
05050     if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
05051         hasMultipleUsesOf(Cur, VisitedInsts))
05052       return false;
05053 
05054     // All inputs to a PHI node must be a reduction value.
05055     if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
05056       return false;
05057 
05058     if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
05059                                      isa<SelectInst>(Cur)))
05060       ++NumCmpSelectPatternInst;
05061     if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
05062                                    isa<SelectInst>(Cur)))
05063       ++NumCmpSelectPatternInst;
05064 
05065     // Check  whether we found a reduction operator.
05066     FoundReduxOp |= !IsAPhi;
05067 
05068     // Process users of current instruction. Push non-PHI nodes after PHI nodes
05069     // onto the stack. This way we are going to have seen all inputs to PHI
05070     // nodes once we get to them.
05071     SmallVector<Instruction *, 8> NonPHIs;
05072     SmallVector<Instruction *, 8> PHIs;
05073     for (User *U : Cur->users()) {
05074       Instruction *UI = cast<Instruction>(U);
05075 
05076       // Check if we found the exit user.
05077       BasicBlock *Parent = UI->getParent();
05078       if (!TheLoop->contains(Parent)) {
05079         // Exit if you find multiple outside users or if the header phi node is
05080         // being used. In this case the user uses the value of the previous
05081         // iteration, in which case we would loose "VF-1" iterations of the
05082         // reduction operation if we vectorize.
05083         if (ExitInstruction != nullptr || Cur == Phi)
05084           return false;
05085 
05086         // The instruction used by an outside user must be the last instruction
05087         // before we feed back to the reduction phi. Otherwise, we loose VF-1
05088         // operations on the value.
05089         if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
05090          return false;
05091 
05092         ExitInstruction = Cur;
05093         continue;
05094       }
05095 
05096       // Process instructions only once (termination). Each reduction cycle
05097       // value must only be used once, except by phi nodes and min/max
05098       // reductions which are represented as a cmp followed by a select.
05099       ReductionInstDesc IgnoredVal(false, nullptr);
05100       if (VisitedInsts.insert(UI)) {
05101         if (isa<PHINode>(UI))
05102           PHIs.push_back(UI);
05103         else
05104           NonPHIs.push_back(UI);
05105       } else if (!isa<PHINode>(UI) &&
05106                  ((!isa<FCmpInst>(UI) &&
05107                    !isa<ICmpInst>(UI) &&
05108                    !isa<SelectInst>(UI)) ||
05109                   !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
05110         return false;
05111 
05112       // Remember that we completed the cycle.
05113       if (UI == Phi)
05114         FoundStartPHI = true;
05115     }
05116     Worklist.append(PHIs.begin(), PHIs.end());
05117     Worklist.append(NonPHIs.begin(), NonPHIs.end());
05118   }
05119 
05120   // This means we have seen one but not the other instruction of the
05121   // pattern or more than just a select and cmp.
05122   if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
05123       NumCmpSelectPatternInst != 2)
05124     return false;
05125 
05126   if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
05127     return false;
05128 
05129   // We found a reduction var if we have reached the original phi node and we
05130   // only have a single instruction with out-of-loop users.
05131 
05132   // This instruction is allowed to have out-of-loop users.
05133   AllowedExit.insert(ExitInstruction);
05134 
05135   // Save the description of this reduction variable.
05136   ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
05137                          ReduxDesc.MinMaxKind);
05138   Reductions[Phi] = RD;
05139   // We've ended the cycle. This is a reduction variable if we have an
05140   // outside user and it has a binary op.
05141 
05142   return true;
05143 }
05144 
05145 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
05146 /// pattern corresponding to a min(X, Y) or max(X, Y).
05147 LoopVectorizationLegality::ReductionInstDesc
05148 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
05149                                                     ReductionInstDesc &Prev) {
05150 
05151   assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
05152          "Expect a select instruction");
05153   Instruction *Cmp = nullptr;
05154   SelectInst *Select = nullptr;
05155 
05156   // We must handle the select(cmp()) as a single instruction. Advance to the
05157   // select.
05158   if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
05159     if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
05160       return ReductionInstDesc(false, I);
05161     return ReductionInstDesc(Select, Prev.MinMaxKind);
05162   }
05163 
05164   // Only handle single use cases for now.
05165   if (!(Select = dyn_cast<SelectInst>(I)))
05166     return ReductionInstDesc(false, I);
05167   if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
05168       !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
05169     return ReductionInstDesc(false, I);
05170   if (!Cmp->hasOneUse())
05171     return ReductionInstDesc(false, I);
05172 
05173   Value *CmpLeft;
05174   Value *CmpRight;
05175 
05176   // Look for a min/max pattern.
05177   if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05178     return ReductionInstDesc(Select, MRK_UIntMin);
05179   else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05180     return ReductionInstDesc(Select, MRK_UIntMax);
05181   else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05182     return ReductionInstDesc(Select, MRK_SIntMax);
05183   else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05184     return ReductionInstDesc(Select, MRK_SIntMin);
05185   else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05186     return ReductionInstDesc(Select, MRK_FloatMin);
05187   else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05188     return ReductionInstDesc(Select, MRK_FloatMax);
05189   else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05190     return ReductionInstDesc(Select, MRK_FloatMin);
05191   else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05192     return ReductionInstDesc(Select, MRK_FloatMax);
05193 
05194   return ReductionInstDesc(false, I);
05195 }
05196 
05197 LoopVectorizationLegality::ReductionInstDesc
05198 LoopVectorizationLegality::isReductionInstr(Instruction *I,
05199                                             ReductionKind Kind,
05200                                             ReductionInstDesc &Prev) {
05201   bool FP = I->getType()->isFloatingPointTy();
05202   bool FastMath = FP && I->hasUnsafeAlgebra();
05203   switch (I->getOpcode()) {
05204   default:
05205     return ReductionInstDesc(false, I);
05206   case Instruction::PHI:
05207       if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
05208                  Kind != RK_FloatMinMax))
05209         return ReductionInstDesc(false, I);
05210     return ReductionInstDesc(I, Prev.MinMaxKind);
05211   case Instruction::Sub:
05212   case Instruction::Add:
05213     return ReductionInstDesc(Kind == RK_IntegerAdd, I);
05214   case Instruction::Mul:
05215     return ReductionInstDesc(Kind == RK_IntegerMult, I);
05216   case Instruction::And:
05217     return ReductionInstDesc(Kind == RK_IntegerAnd, I);
05218   case Instruction::Or:
05219     return ReductionInstDesc(Kind == RK_IntegerOr, I);
05220   case Instruction::Xor:
05221     return ReductionInstDesc(Kind == RK_IntegerXor, I);
05222   case Instruction::FMul:
05223     return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
05224   case Instruction::FSub:
05225   case Instruction::FAdd:
05226     return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
05227   case Instruction::FCmp:
05228   case Instruction::ICmp:
05229   case Instruction::Select:
05230     if (Kind != RK_IntegerMinMax &&
05231         (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
05232       return ReductionInstDesc(false, I);
05233     return isMinMaxSelectCmpPattern(I, Prev);
05234   }
05235 }
05236 
05237 LoopVectorizationLegality::InductionKind
05238 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
05239   Type *PhiTy = Phi->getType();
05240   // We only handle integer and pointer inductions variables.
05241   if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
05242     return IK_NoInduction;
05243 
05244   // Check that the PHI is consecutive.
05245   const SCEV *PhiScev = SE->getSCEV(Phi);
05246   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
05247   if (!AR) {
05248     DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
05249     return IK_NoInduction;
05250   }
05251   const SCEV *Step = AR->getStepRecurrence(*SE);
05252 
05253   // Integer inductions need to have a stride of one.
05254   if (PhiTy->isIntegerTy()) {
05255     if (Step->isOne())
05256       return IK_IntInduction;
05257     if (Step->isAllOnesValue())
05258       return IK_ReverseIntInduction;
05259     return IK_NoInduction;
05260   }
05261 
05262   // Calculate the pointer stride and check if it is consecutive.
05263   const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
05264   if (!C)
05265     return IK_NoInduction;
05266 
05267   assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
05268   uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
05269   if (C->getValue()->equalsInt(Size))
05270     return IK_PtrInduction;
05271   else if (C->getValue()->equalsInt(0 - Size))
05272     return IK_ReversePtrInduction;
05273 
05274   return IK_NoInduction;
05275 }
05276 
05277 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
05278   Value *In0 = const_cast<Value*>(V);
05279   PHINode *PN = dyn_cast_or_null<PHINode>(In0);
05280   if (!PN)
05281     return false;
05282 
05283   return Inductions.count(PN);
05284 }
05285 
05286 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB)  {
05287   assert(TheLoop->contains(BB) && "Unknown block used");
05288 
05289   // Blocks that do not dominate the latch need predication.
05290   BasicBlock* Latch = TheLoop->getLoopLatch();
05291   return !DT->dominates(BB, Latch);
05292 }
05293 
05294 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
05295                                            SmallPtrSetImpl<Value *> &SafePtrs) {
05296   for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
05297     // We might be able to hoist the load.
05298     if (it->mayReadFromMemory()) {
05299       LoadInst *LI = dyn_cast<LoadInst>(it);
05300       if (!LI || !SafePtrs.count(LI->getPointerOperand()))
05301         return false;
05302     }
05303 
05304     // We don't predicate stores at the moment.
05305     if (it->mayWriteToMemory()) {
05306       StoreInst *SI = dyn_cast<StoreInst>(it);
05307       // We only support predication of stores in basic blocks with one
05308       // predecessor.
05309       if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
05310           !SafePtrs.count(SI->getPointerOperand()) ||
05311           !SI->getParent()->getSinglePredecessor())
05312         return false;
05313     }
05314     if (it->mayThrow())
05315       return false;
05316 
05317     // Check that we don't have a constant expression that can trap as operand.
05318     for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
05319          OI != OE; ++OI) {
05320       if (Constant *C = dyn_cast<Constant>(*OI))
05321         if (C->canTrap())
05322           return false;
05323     }
05324 
05325     // The instructions below can trap.
05326     switch (it->getOpcode()) {
05327     default: continue;
05328     case Instruction::UDiv:
05329     case Instruction::SDiv:
05330     case Instruction::URem:
05331     case Instruction::SRem:
05332              return false;
05333     }
05334   }
05335 
05336   return true;
05337 }
05338 
05339 LoopVectorizationCostModel::VectorizationFactor
05340 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
05341   // Width 1 means no vectorize
05342   VectorizationFactor Factor = { 1U, 0U };
05343   if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
05344     emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
05345     DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
05346     return Factor;
05347   }
05348 
05349   if (!EnableCondStoresVectorization && Legal->NumPredStores) {
05350     emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
05351     DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
05352     return Factor;
05353   }
05354 
05355   // Find the trip count.
05356   unsigned TC = SE->getSmallConstantTripCount(TheLoop);
05357   DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
05358 
05359   unsigned WidestType = getWidestType();
05360   unsigned WidestRegister = TTI.getRegisterBitWidth(true);
05361   unsigned MaxSafeDepDist = -1U;
05362   if (Legal->getMaxSafeDepDistBytes() != -1U)
05363     MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
05364   WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
05365                     WidestRegister : MaxSafeDepDist);
05366   unsigned MaxVectorSize = WidestRegister / WidestType;
05367   DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
05368   DEBUG(dbgs() << "LV: The Widest register is: "
05369           << WidestRegister << " bits.\n");
05370 
05371   if (MaxVectorSize == 0) {
05372     DEBUG(dbgs() << "LV: The target has no vector registers.\n");
05373     MaxVectorSize = 1;
05374   }
05375 
05376   assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
05377          " into one vector!");
05378 
05379   unsigned VF = MaxVectorSize;
05380 
05381   // If we optimize the program for size, avoid creating the tail loop.
05382   if (OptForSize) {
05383     // If we are unable to calculate the trip count then don't try to vectorize.
05384     if (TC < 2) {
05385       emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
05386       DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
05387       return Factor;
05388     }
05389 
05390     // Find the maximum SIMD width that can fit within the trip count.
05391     VF = TC % MaxVectorSize;
05392 
05393     if (VF == 0)
05394       VF = MaxVectorSize;
05395 
05396     // If the trip count that we found modulo the vectorization factor is not
05397     // zero then we require a tail.
05398     if (VF < 2) {
05399       emitAnalysis(Report() << "cannot optimize for size and vectorize at the "
05400                                "same time. Enable vectorization of this loop "
05401                                "with '#pragma clang loop vectorize(enable)' "
05402                                "when compiling with -Os");
05403       DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
05404       return Factor;
05405     }
05406   }
05407 
05408   int UserVF = Hints->getWidth();
05409   if (UserVF != 0) {
05410     assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
05411     DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
05412 
05413     Factor.Width = UserVF;
05414     return Factor;
05415   }
05416 
05417   float Cost = expectedCost(1);
05418 #ifndef NDEBUG
05419   const float ScalarCost = Cost;
05420 #endif /* NDEBUG */
05421   unsigned Width = 1;
05422   DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
05423 
05424   bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
05425   // Ignore scalar width, because the user explicitly wants vectorization.
05426   if (ForceVectorization && VF > 1) {
05427     Width = 2;
05428     Cost = expectedCost(Width) / (float)Width;
05429   }
05430 
05431   for (unsigned i=2; i <= VF; i*=2) {
05432     // Notice that the vector loop needs to be executed less times, so
05433     // we need to divide the cost of the vector loops by the width of
05434     // the vector elements.
05435     float VectorCost = expectedCost(i) / (float)i;
05436     DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
05437           (int)VectorCost << ".\n");
05438     if (VectorCost < Cost) {
05439       Cost = VectorCost;
05440       Width = i;
05441     }
05442   }
05443 
05444   DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
05445         << "LV: Vectorization seems to be not beneficial, "
05446         << "but was forced by a user.\n");
05447   DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
05448   Factor.Width = Width;
05449   Factor.Cost = Width * Cost;
05450   return Factor;
05451 }
05452 
05453 unsigned LoopVectorizationCostModel::getWidestType() {
05454   unsigned MaxWidth = 8;
05455 
05456   // For each block.
05457   for (Loop::block_iterator bb = TheLoop->block_begin(),
05458        be = TheLoop->block_end(); bb != be; ++bb) {
05459     BasicBlock *BB = *bb;
05460 
05461     // For each instruction in the loop.
05462     for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
05463       Type *T = it->getType();
05464 
05465       // Ignore ephemeral values.
05466       if (EphValues.count(it))
05467         continue;
05468 
05469       // Only examine Loads, Stores and PHINodes.
05470       if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
05471         continue;
05472 
05473       // Examine PHI nodes that are reduction variables.
05474       if (PHINode *PN = dyn_cast<PHINode>(it))
05475         if (!Legal->getReductionVars()->count(PN))
05476           continue;
05477 
05478       // Examine the stored values.
05479       if (StoreInst *ST = dyn_cast<StoreInst>(it))
05480         T = ST->getValueOperand()->getType();
05481 
05482       // Ignore loaded pointer types and stored pointer types that are not
05483       // consecutive. However, we do want to take consecutive stores/loads of
05484       // pointer vectors into account.
05485       if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
05486         continue;
05487 
05488       MaxWidth = std::max(MaxWidth,
05489                           (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
05490     }
05491   }
05492 
05493   return MaxWidth;
05494 }
05495 
05496 unsigned
05497 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
05498                                                unsigned VF,
05499                                                unsigned LoopCost) {
05500 
05501   // -- The unroll heuristics --
05502   // We unroll the loop in order to expose ILP and reduce the loop overhead.
05503   // There are many micro-architectural considerations that we can't predict
05504   // at this level. For example, frontend pressure (on decode or fetch) due to
05505   // code size, or the number and capabilities of the execution ports.
05506   //
05507   // We use the following heuristics to select the unroll factor:
05508   // 1. If the code has reductions, then we unroll in order to break the cross
05509   // iteration dependency.
05510   // 2. If the loop is really small, then we unroll in order to reduce the loop
05511   // overhead.
05512   // 3. We don't unroll if we think that we will spill registers to memory due
05513   // to the increased register pressure.
05514 
05515   // Use the user preference, unless 'auto' is selected.
05516   int UserUF = Hints->getInterleave();
05517   if (UserUF != 0)
05518     return UserUF;
05519 
05520   // When we optimize for size, we don't unroll.
05521   if (OptForSize)
05522     return 1;
05523 
05524   // We used the distance for the unroll factor.
05525   if (Legal->getMaxSafeDepDistBytes() != -1U)
05526     return 1;
05527 
05528   // Do not unroll loops with a relatively small trip count.
05529   unsigned TC = SE->getSmallConstantTripCount(TheLoop);
05530   if (TC > 1 && TC < TinyTripCountUnrollThreshold)
05531     return 1;
05532 
05533   unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
05534   DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
05535         " registers\n");
05536 
05537   if (VF == 1) {
05538     if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
05539       TargetNumRegisters = ForceTargetNumScalarRegs;
05540   } else {
05541     if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
05542       TargetNumRegisters = ForceTargetNumVectorRegs;
05543   }
05544 
05545   LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
05546   // We divide by these constants so assume that we have at least one
05547   // instruction that uses at least one register.
05548   R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
05549   R.NumInstructions = std::max(R.NumInstructions, 1U);
05550 
05551   // We calculate the unroll factor using the following formula.
05552   // Subtract the number of loop invariants from the number of available
05553   // registers. These registers are used by all of the unrolled instances.
05554   // Next, divide the remaining registers by the number of registers that is
05555   // required by the loop, in order to estimate how many parallel instances
05556   // fit without causing spills. All of this is rounded down if necessary to be
05557   // a power of two. We want power of two unroll factors to simplify any
05558   // addressing operations or alignment considerations.
05559   unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
05560                               R.MaxLocalUsers);
05561 
05562   // Don't count the induction variable as unrolled.
05563   if (EnableIndVarRegisterHeur)
05564     UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
05565                        std::max(1U, (R.MaxLocalUsers - 1)));
05566 
05567   // Clamp the unroll factor ranges to reasonable factors.
05568   unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
05569 
05570   // Check if the user has overridden the unroll max.
05571   if (VF == 1) {
05572     if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
05573       MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
05574   } else {
05575     if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
05576       MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
05577   }
05578 
05579   // If we did not calculate the cost for VF (because the user selected the VF)
05580   // then we calculate the cost of VF here.
05581   if (LoopCost == 0)
05582     LoopCost = expectedCost(VF);
05583 
05584   // Clamp the calculated UF to be between the 1 and the max unroll factor
05585   // that the target allows.
05586   if (UF > MaxInterleaveSize)
05587     UF = MaxInterleaveSize;
05588   else if (UF < 1)
05589     UF = 1;
05590 
05591   // Unroll if we vectorized this loop and there is a reduction that could
05592   // benefit from unrolling.
05593   if (VF > 1 && Legal->getReductionVars()->size()) {
05594     DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
05595     return UF;
05596   }
05597 
05598   // Note that if we've already vectorized the loop we will have done the
05599   // runtime check and so unrolling won't require further checks.
05600   bool UnrollingRequiresRuntimePointerCheck =
05601       (VF == 1 && Legal->getRuntimePointerCheck()->Need);
05602 
05603   // We want to unroll small loops in order to reduce the loop overhead and
05604   // potentially expose ILP opportunities.
05605   DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
05606   if (!UnrollingRequiresRuntimePointerCheck &&
05607       LoopCost < SmallLoopCost) {
05608     // We assume that the cost overhead is 1 and we use the cost model
05609     // to estimate the cost of the loop and unroll until the cost of the
05610     // loop overhead is about 5% of the cost of the loop.
05611     unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
05612 
05613     // Unroll until store/load ports (estimated by max unroll factor) are
05614     // saturated.
05615     unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
05616     unsigned LoadsUF = UF /  (Legal->NumLoads ? Legal->NumLoads : 1);
05617 
05618     // If we have a scalar reduction (vector reductions are already dealt with
05619     // by this point), we can increase the critical path length if the loop
05620     // we're unrolling is inside another loop. Limit, by default to 2, so the
05621     // critical path only gets increased by one reduction operation.
05622     if (Legal->getReductionVars()->size() &&
05623         TheLoop->getLoopDepth() > 1) {
05624       unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
05625       SmallUF = std::min(SmallUF, F);
05626       StoresUF = std::min(StoresUF, F);
05627       LoadsUF = std::min(LoadsUF, F);
05628     }
05629 
05630     if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
05631       DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
05632       return std::max(StoresUF, LoadsUF);
05633     }
05634 
05635     DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
05636     return SmallUF;
05637   }
05638 
05639   DEBUG(dbgs() << "LV: Not Unrolling.\n");
05640   return 1;
05641 }
05642 
05643 LoopVectorizationCostModel::RegisterUsage
05644 LoopVectorizationCostModel::calculateRegisterUsage() {
05645   // This function calculates the register usage by measuring the highest number
05646   // of values that are alive at a single location. Obviously, this is a very
05647   // rough estimation. We scan the loop in a topological order in order and
05648   // assign a number to each instruction. We use RPO to ensure that defs are
05649   // met before their users. We assume that each instruction that has in-loop
05650   // users starts an interval. We record every time that an in-loop value is
05651   // used, so we have a list of the first and last occurrences of each
05652   // instruction. Next, we transpose this data structure into a multi map that
05653   // holds the list of intervals that *end* at a specific location. This multi
05654   // map allows us to perform a linear search. We scan the instructions linearly
05655   // and record each time that a new interval starts, by placing it in a set.
05656   // If we find this value in the multi-map then we remove it from the set.
05657   // The max register usage is the maximum size of the set.
05658   // We also search for instructions that are defined outside the loop, but are
05659   // used inside the loop. We need this number separately from the max-interval
05660   // usage number because when we unroll, loop-invariant values do not take
05661   // more register.
05662   LoopBlocksDFS DFS(TheLoop);
05663   DFS.perform(LI);
05664 
05665   RegisterUsage R;
05666   R.NumInstructions = 0;
05667 
05668   // Each 'key' in the map opens a new interval. The values
05669   // of the map are the index of the 'last seen' usage of the
05670   // instruction that is the key.
05671   typedef DenseMap<Instruction*, unsigned> IntervalMap;
05672   // Maps instruction to its index.
05673   DenseMap<unsigned, Instruction*> IdxToInstr;
05674   // Marks the end of each interval.
05675   IntervalMap EndPoint;
05676   // Saves the list of instruction indices that are used in the loop.
05677   SmallSet<Instruction*, 8> Ends;
05678   // Saves the list of values that are used in the loop but are
05679   // defined outside the loop, such as arguments and constants.
05680   SmallPtrSet<Value*, 8> LoopInvariants;
05681 
05682   unsigned Index = 0;
05683   for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
05684        be = DFS.endRPO(); bb != be; ++bb) {
05685     R.NumInstructions += (*bb)->size();
05686     for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
05687          ++it) {
05688       Instruction *I = it;
05689       IdxToInstr[Index++] = I;
05690 
05691       // Save the end location of each USE.
05692       for (unsigned i = 0; i < I->getNumOperands(); ++i) {
05693         Value *U = I->getOperand(i);
05694         Instruction *Instr = dyn_cast<Instruction>(U);
05695 
05696         // Ignore non-instruction values such as arguments, constants, etc.
05697         if (!Instr) continue;
05698 
05699         // If this instruction is outside the loop then record it and continue.
05700         if (!TheLoop->contains(Instr)) {
05701           LoopInvariants.insert(Instr);
05702           continue;
05703         }
05704 
05705         // Overwrite previous end points.
05706         EndPoint[Instr] = Index;
05707         Ends.insert(Instr);
05708       }
05709     }
05710   }
05711 
05712   // Saves the list of intervals that end with the index in 'key'.
05713   typedef SmallVector<Instruction*, 2> InstrList;
05714   DenseMap<unsigned, InstrList> TransposeEnds;
05715 
05716   // Transpose the EndPoints to a list of values that end at each index.
05717   for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
05718        it != e; ++it)
05719     TransposeEnds[it->second].push_back(it->first);
05720 
05721   SmallSet<Instruction*, 8> OpenIntervals;
05722   unsigned MaxUsage = 0;
05723 
05724 
05725   DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
05726   for (unsigned int i = 0; i < Index; ++i) {
05727     Instruction *I = IdxToInstr[i];
05728     // Ignore instructions that are never used within the loop.
05729     if (!Ends.count(I)) continue;
05730 
05731     // Ignore ephemeral values.
05732     if (EphValues.count(I))
05733       continue;
05734 
05735     // Remove all of the instructions that end at this location.
05736     InstrList &List = TransposeEnds[i];
05737     for (unsigned int j=0, e = List.size(); j < e; ++j)
05738       OpenIntervals.erase(List[j]);
05739 
05740     // Count the number of live interals.
05741     MaxUsage = std::max(MaxUsage, OpenIntervals.size());
05742 
05743     DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
05744           OpenIntervals.size() << '\n');
05745 
05746     // Add the current instruction to the list of open intervals.
05747     OpenIntervals.insert(I);
05748   }
05749 
05750   unsigned Invariant = LoopInvariants.size();
05751   DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
05752   DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
05753   DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
05754 
05755   R.LoopInvariantRegs = Invariant;
05756   R.MaxLocalUsers = MaxUsage;
05757   return R;
05758 }
05759 
05760 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
05761   unsigned Cost = 0;
05762 
05763   // For each block.
05764   for (Loop::block_iterator bb = TheLoop->block_begin(),
05765        be = TheLoop->block_end(); bb != be; ++bb) {
05766     unsigned BlockCost = 0;
05767     BasicBlock *BB = *bb;
05768 
05769     // For each instruction in the old loop.
05770     for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
05771       // Skip dbg intrinsics.
05772       if (isa<DbgInfoIntrinsic>(it))
05773         continue;
05774 
05775       // Ignore ephemeral values.
05776       if (EphValues.count(it))
05777         continue;
05778 
05779       unsigned C = getInstructionCost(it, VF);
05780 
05781       // Check if we should override the cost.
05782       if (ForceTargetInstructionCost.getNumOccurrences() > 0)
05783         C = ForceTargetInstructionCost;
05784 
05785       BlockCost += C;
05786       DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
05787             VF << " For instruction: " << *it << '\n');
05788     }
05789 
05790     // We assume that if-converted blocks have a 50% chance of being executed.
05791     // When the code is scalar then some of the blocks are avoided due to CF.
05792     // When the code is vectorized we execute all code paths.
05793     if (VF == 1 && Legal->blockNeedsPredication(*bb))
05794       BlockCost /= 2;
05795 
05796     Cost += BlockCost;
05797   }
05798 
05799   return Cost;
05800 }
05801 
05802 /// \brief Check whether the address computation for a non-consecutive memory
05803 /// access looks like an unlikely candidate for being merged into the indexing
05804 /// mode.
05805 ///
05806 /// We look for a GEP which has one index that is an induction variable and all
05807 /// other indices are loop invariant. If the stride of this access is also
05808 /// within a small bound we decide that this address computation can likely be
05809 /// merged into the addressing mode.
05810 /// In all other cases, we identify the address computation as complex.
05811 static bool isLikelyComplexAddressComputation(Value *Ptr,
05812                                               LoopVectorizationLegality *Legal,
05813                                               ScalarEvolution *SE,
05814                                               const Loop *TheLoop) {
05815   GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
05816   if (!Gep)
05817     return true;
05818 
05819   // We are looking for a gep with all loop invariant indices except for one
05820   // which should be an induction variable.
05821   unsigned NumOperands = Gep->getNumOperands();
05822   for (unsigned i = 1; i < NumOperands; ++i) {
05823     Value *Opd = Gep->getOperand(i);
05824     if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
05825         !Legal->isInductionVariable(Opd))
05826       return true;
05827   }
05828 
05829   // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
05830   // can likely be merged into the address computation.
05831   unsigned MaxMergeDistance = 64;
05832 
05833   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
05834   if (!AddRec)
05835     return true;
05836 
05837   // Check the step is constant.
05838   const SCEV *Step = AddRec->getStepRecurrence(*SE);
05839   // Calculate the pointer stride and check if it is consecutive.
05840   const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
05841   if (!C)
05842     return true;
05843 
05844   const APInt &APStepVal = C->getValue()->getValue();
05845 
05846   // Huge step value - give up.
05847   if (APStepVal.getBitWidth() > 64)
05848     return true;
05849 
05850   int64_t StepVal = APStepVal.getSExtValue();
05851 
05852   return StepVal > MaxMergeDistance;
05853 }
05854 
05855 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
05856   if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
05857     return true;
05858   return false;
05859 }
05860 
05861 unsigned
05862 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
05863   // If we know that this instruction will remain uniform, check the cost of
05864   // the scalar version.
05865   if (Legal->isUniformAfterVectorization(I))
05866     VF = 1;
05867 
05868   Type *RetTy = I->getType();
05869   Type *VectorTy = ToVectorTy(RetTy, VF);
05870 
05871   // TODO: We need to estimate the cost of intrinsic calls.
05872   switch (I->getOpcode()) {
05873   case Instruction::GetElementPtr:
05874     // We mark this instruction as zero-cost because the cost of GEPs in
05875     // vectorized code depends on whether the corresponding memory instruction
05876     // is scalarized or not. Therefore, we handle GEPs with the memory
05877     // instruction cost.
05878     return 0;
05879   case Instruction::Br: {
05880     return TTI.getCFInstrCost(I->getOpcode());
05881   }
05882   case Instruction::PHI:
05883     //TODO: IF-converted IFs become selects.
05884     return 0;
05885   case Instruction::Add:
05886   case Instruction::FAdd:
05887   case Instruction::Sub:
05888   case Instruction::FSub:
05889   case Instruction::Mul:
05890   case Instruction::FMul:
05891   case Instruction::UDiv:
05892   case Instruction::SDiv:
05893   case Instruction::FDiv:
05894   case Instruction::URem:
05895   case Instruction::SRem:
05896   case Instruction::FRem:
05897   case Instruction::Shl:
05898   case Instruction::LShr:
05899   case Instruction::AShr:
05900   case Instruction::And:
05901   case Instruction::Or:
05902   case Instruction::Xor: {
05903     // Since we will replace the stride by 1 the multiplication should go away.
05904     if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
05905       return 0;
05906     // Certain instructions can be cheaper to vectorize if they have a constant
05907     // second vector operand. One example of this are shifts on x86.
05908     TargetTransformInfo::OperandValueKind Op1VK =
05909       TargetTransformInfo::OK_AnyValue;
05910     TargetTransformInfo::OperandValueKind Op2VK =
05911       TargetTransformInfo::OK_AnyValue;
05912     TargetTransformInfo::OperandValueProperties Op1VP =
05913         TargetTransformInfo::OP_None;
05914     TargetTransformInfo::OperandValueProperties Op2VP =
05915         TargetTransformInfo::OP_None;
05916     Value *Op2 = I->getOperand(1);
05917 
05918     // Check for a splat of a constant or for a non uniform vector of constants.
05919     if (isa<ConstantInt>(Op2)) {
05920       ConstantInt *CInt = cast<ConstantInt>(Op2);
05921       if (CInt && CInt->getValue().isPowerOf2())
05922         Op2VP = TargetTransformInfo::OP_PowerOf2;
05923       Op2VK = TargetTransformInfo::OK_UniformConstantValue;
05924     } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
05925       Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
05926       Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
05927       if (SplatValue) {
05928         ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
05929         if (CInt && CInt->getValue().isPowerOf2())
05930           Op2VP = TargetTransformInfo::OP_PowerOf2;
05931         Op2VK = TargetTransformInfo::OK_UniformConstantValue;
05932       }
05933     }
05934 
05935     return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
05936                                       Op1VP, Op2VP);
05937   }
05938   case Instruction::Select: {
05939     SelectInst *SI = cast<SelectInst>(I);
05940     const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
05941     bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
05942     Type *CondTy = SI->getCondition()->getType();
05943     if (!ScalarCond)
05944       CondTy = VectorType::get(CondTy, VF);
05945 
05946     return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
05947   }
05948   case Instruction::ICmp:
05949   case Instruction::FCmp: {
05950     Type *ValTy = I->getOperand(0)->getType();
05951     VectorTy = ToVectorTy(ValTy, VF);
05952     return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
05953   }
05954   case Instruction::Store:
05955   case Instruction::Load: {
05956     StoreInst *SI = dyn_cast<StoreInst>(I);
05957     LoadInst *LI = dyn_cast<LoadInst>(I);
05958     Type *ValTy = (SI ? SI->getValueOperand()->getType() :
05959                    LI->getType());
05960     VectorTy = ToVectorTy(ValTy, VF);
05961 
05962     unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
05963     unsigned AS = SI ? SI->getPointerAddressSpace() :
05964       LI->getPointerAddressSpace();
05965     Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
05966     // We add the cost of address computation here instead of with the gep
05967     // instruction because only here we know whether the operation is
05968     // scalarized.
05969     if (VF == 1)
05970       return TTI.getAddressComputationCost(VectorTy) +
05971         TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
05972 
05973     // Scalarized loads/stores.
05974     int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
05975     bool Reverse = ConsecutiveStride < 0;
05976     unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
05977     unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
05978     if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
05979       bool IsComplexComputation =
05980         isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
05981       unsigned Cost = 0;
05982       // The cost of extracting from the value vector and pointer vector.
05983       Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
05984       for (unsigned i = 0; i < VF; ++i) {
05985         //  The cost of extracting the pointer operand.
05986         Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
05987         // In case of STORE, the cost of ExtractElement from the vector.
05988         // In case of LOAD, the cost of InsertElement into the returned
05989         // vector.
05990         Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
05991                                             Instruction::InsertElement,
05992                                             VectorTy, i);
05993       }
05994 
05995       // The cost of the scalar loads/stores.
05996       Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
05997       Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
05998                                        Alignment, AS);
05999       return Cost;
06000     }
06001 
06002     // Wide load/stores.
06003     unsigned Cost = TTI.getAddressComputationCost(VectorTy);
06004     Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
06005 
06006     if (Reverse)
06007       Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
06008                                   VectorTy, 0);
06009     return Cost;
06010   }
06011   case Instruction::ZExt:
06012   case Instruction::SExt:
06013   case Instruction::FPToUI:
06014   case Instruction::FPToSI:
06015   case Instruction::FPExt:
06016   case Instruction::PtrToInt:
06017   case Instruction::IntToPtr:
06018   case Instruction::SIToFP:
06019   case Instruction::UIToFP:
06020   case Instruction::Trunc:
06021   case Instruction::FPTrunc:
06022   case Instruction::BitCast: {
06023     // We optimize the truncation of induction variable.
06024     // The cost of these is the same as the scalar operation.
06025     if (I->getOpcode() == Instruction::Trunc &&
06026         Legal->isInductionVariable(I->getOperand(0)))
06027       return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
06028                                   I->getOperand(0)->getType());
06029 
06030     Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
06031     return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
06032   }
06033   case Instruction::Call: {
06034     CallInst *CI = cast<CallInst>(I);
06035     Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
06036     assert(ID && "Not an intrinsic call!");
06037     Type *RetTy = ToVectorTy(CI->getType(), VF);
06038     SmallVector<Type*, 4> Tys;
06039     for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
06040       Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
06041     return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
06042   }
06043   default: {
06044     // We are scalarizing the instruction. Return the cost of the scalar
06045     // instruction, plus the cost of insert and extract into vector
06046     // elements, times the vector width.
06047     unsigned Cost = 0;
06048 
06049     if (!RetTy->isVoidTy() && VF != 1) {
06050       unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
06051                                                 VectorTy);
06052       unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
06053                                                 VectorTy);
06054 
06055       // The cost of inserting the results plus extracting each one of the
06056       // operands.
06057       Cost += VF * (InsCost + ExtCost * I->getNumOperands());
06058     }
06059 
06060     // The cost of executing VF copies of the scalar instruction. This opcode
06061     // is unknown. Assume that it is the same as 'mul'.
06062     Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
06063     return Cost;
06064   }
06065   }// end of switch.
06066 }
06067 
06068 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
06069   if (Scalar->isVoidTy() || VF == 1)
06070     return Scalar;
06071   return VectorType::get(Scalar, VF);
06072 }
06073 
06074 char LoopVectorize::ID = 0;
06075 static const char lv_name[] = "Loop Vectorization";
06076 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
06077 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
06078 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
06079 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
06080 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
06081 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
06082 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
06083 INITIALIZE_PASS_DEPENDENCY(LCSSA)
06084 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
06085 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
06086 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
06087 
06088 namespace llvm {
06089   Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
06090     return new LoopVectorize(NoUnrolling, AlwaysVectorize);
06091   }
06092 }
06093 
06094 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
06095   // Check for a store.
06096   if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
06097     return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
06098 
06099   // Check for a load.
06100   if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
06101     return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
06102 
06103   return false;
06104 }
06105 
06106 
06107 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
06108                                              bool IfPredicateStore) {
06109   assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
06110   // Holds vector parameters or scalars, in case of uniform vals.
06111   SmallVector<VectorParts, 4> Params;
06112 
06113   setDebugLocFromInst(Builder, Instr);
06114 
06115   // Find all of the vectorized parameters.
06116   for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
06117     Value *SrcOp = Instr->getOperand(op);
06118 
06119     // If we are accessing the old induction variable, use the new one.
06120     if (SrcOp == OldInduction) {
06121       Params.push_back(getVectorValue(SrcOp));
06122       continue;
06123     }
06124 
06125     // Try using previously calculated values.
06126     Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
06127 
06128     // If the src is an instruction that appeared earlier in the basic block
06129     // then it should already be vectorized.
06130     if (SrcInst && OrigLoop->contains(SrcInst)) {
06131       assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
06132       // The parameter is a vector value from earlier.
06133       Params.push_back(WidenMap.get(SrcInst));
06134     } else {
06135       // The parameter is a scalar from outside the loop. Maybe even a constant.
06136       VectorParts Scalars;
06137       Scalars.append(UF, SrcOp);
06138       Params.push_back(Scalars);
06139     }
06140   }
06141 
06142   assert(Params.size() == Instr->getNumOperands() &&
06143          "Invalid number of operands");
06144 
06145   // Does this instruction return a value ?
06146   bool IsVoidRetTy = Instr->getType()->isVoidTy();
06147 
06148   Value *UndefVec = IsVoidRetTy ? nullptr :
06149   UndefValue::get(Instr->getType());
06150   // Create a new entry in the WidenMap and initialize it to Undef or Null.
06151   VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
06152 
06153   Instruction *InsertPt = Builder.GetInsertPoint();
06154   BasicBlock *IfBlock = Builder.GetInsertBlock();
06155   BasicBlock *CondBlock = nullptr;
06156 
06157   VectorParts Cond;
06158   Loop *VectorLp = nullptr;
06159   if (IfPredicateStore) {
06160     assert(Instr->getParent()->getSinglePredecessor() &&
06161            "Only support single predecessor blocks");
06162     Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
06163                           Instr->getParent());
06164     VectorLp = LI->getLoopFor(IfBlock);
06165     assert(VectorLp && "Must have a loop for this block");
06166   }
06167 
06168   // For each vector unroll 'part':
06169   for (unsigned Part = 0; Part < UF; ++Part) {
06170     // For each scalar that we create:
06171 
06172     // Start an "if (pred) a[i] = ..." block.
06173     Value *Cmp = nullptr;
06174     if (IfPredicateStore) {
06175       if (Cond[Part]->getType()->isVectorTy())
06176         Cond[Part] =
06177             Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
06178       Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
06179                                ConstantInt::get(Cond[Part]->getType(), 1));
06180       CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
06181       LoopVectorBody.push_back(CondBlock);
06182       VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
06183       // Update Builder with newly created basic block.
06184       Builder.SetInsertPoint(InsertPt);
06185     }
06186 
06187     Instruction *Cloned = Instr->clone();
06188       if (!IsVoidRetTy)
06189         Cloned->setName(Instr->getName() + ".cloned");
06190       // Replace the operands of the cloned instructions with extracted scalars.
06191       for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
06192         Value *Op = Params[op][Part];
06193         Cloned->setOperand(op, Op);
06194       }
06195 
06196       // Place the cloned scalar in the new loop.
06197       Builder.Insert(Cloned);
06198 
06199       // If the original scalar returns a value we need to place it in a vector
06200       // so that future users will be able to use it.
06201       if (!IsVoidRetTy)
06202         VecResults[Part] = Cloned;
06203 
06204     // End if-block.
06205       if (IfPredicateStore) {
06206         BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
06207         LoopVectorBody.push_back(NewIfBlock);
06208         VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
06209         Builder.SetInsertPoint(InsertPt);
06210         Instruction *OldBr = IfBlock->getTerminator();
06211         BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
06212         OldBr->eraseFromParent();
06213         IfBlock = NewIfBlock;
06214       }
06215   }
06216 }
06217 
06218 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
06219   StoreInst *SI = dyn_cast<StoreInst>(Instr);
06220   bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
06221 
06222   return scalarizeInstruction(Instr, IfPredicateStore);
06223 }
06224 
06225 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
06226   return Vec;
06227 }
06228 
06229 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
06230   return V;
06231 }
06232 
06233 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
06234                                                bool Negate) {
06235   // When unrolling and the VF is 1, we only need to add a simple scalar.
06236   Type *ITy = Val->getType();
06237   assert(!ITy->isVectorTy() && "Val must be a scalar");
06238   Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
06239   return Builder.CreateAdd(Val, C, "induction");
06240 }