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                             const TargetTransformInfo *TTI)
00585       : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
00586         DT(DT), TLI(TLI), AA(AA), TheFunction(F), TTI(TTI), Induction(nullptr),
00587         WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
00588   }
00589 
00590   /// This enum represents the kinds of reductions that we support.
00591   enum ReductionKind {
00592     RK_NoReduction, ///< Not a reduction.
00593     RK_IntegerAdd,  ///< Sum of integers.
00594     RK_IntegerMult, ///< Product of integers.
00595     RK_IntegerOr,   ///< Bitwise or logical OR of numbers.
00596     RK_IntegerAnd,  ///< Bitwise or logical AND of numbers.
00597     RK_IntegerXor,  ///< Bitwise or logical XOR of numbers.
00598     RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
00599     RK_FloatAdd,    ///< Sum of floats.
00600     RK_FloatMult,   ///< Product of floats.
00601     RK_FloatMinMax  ///< Min/max implemented in terms of select(cmp()).
00602   };
00603 
00604   /// This enum represents the kinds of inductions that we support.
00605   enum InductionKind {
00606     IK_NoInduction,         ///< Not an induction variable.
00607     IK_IntInduction,        ///< Integer induction variable. Step = 1.
00608     IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
00609     IK_PtrInduction,        ///< Pointer induction var. Step = sizeof(elem).
00610     IK_ReversePtrInduction  ///< Reverse ptr indvar. Step = - sizeof(elem).
00611   };
00612 
00613   // This enum represents the kind of minmax reduction.
00614   enum MinMaxReductionKind {
00615     MRK_Invalid,
00616     MRK_UIntMin,
00617     MRK_UIntMax,
00618     MRK_SIntMin,
00619     MRK_SIntMax,
00620     MRK_FloatMin,
00621     MRK_FloatMax
00622   };
00623 
00624   /// This struct holds information about reduction variables.
00625   struct ReductionDescriptor {
00626     ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
00627       Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
00628 
00629     ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
00630                         MinMaxReductionKind MK)
00631         : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
00632 
00633     // The starting value of the reduction.
00634     // It does not have to be zero!
00635     TrackingVH<Value> StartValue;
00636     // The instruction who's value is used outside the loop.
00637     Instruction *LoopExitInstr;
00638     // The kind of the reduction.
00639     ReductionKind Kind;
00640     // If this a min/max reduction the kind of reduction.
00641     MinMaxReductionKind MinMaxKind;
00642   };
00643 
00644   /// This POD struct holds information about a potential reduction operation.
00645   struct ReductionInstDesc {
00646     ReductionInstDesc(bool IsRedux, Instruction *I) :
00647       IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
00648 
00649     ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
00650       IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
00651 
00652     // Is this instruction a reduction candidate.
00653     bool IsReduction;
00654     // The last instruction in a min/max pattern (select of the select(icmp())
00655     // pattern), or the current reduction instruction otherwise.
00656     Instruction *PatternLastInst;
00657     // If this is a min/max pattern the comparison predicate.
00658     MinMaxReductionKind MinMaxKind;
00659   };
00660 
00661   /// This struct holds information about the memory runtime legality
00662   /// check that a group of pointers do not overlap.
00663   struct RuntimePointerCheck {
00664     RuntimePointerCheck() : Need(false) {}
00665 
00666     /// Reset the state of the pointer runtime information.
00667     void reset() {
00668       Need = false;
00669       Pointers.clear();
00670       Starts.clear();
00671       Ends.clear();
00672       IsWritePtr.clear();
00673       DependencySetId.clear();
00674       AliasSetId.clear();
00675     }
00676 
00677     /// Insert a pointer and calculate the start and end SCEVs.
00678     void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
00679                 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
00680 
00681     /// This flag indicates if we need to add the runtime check.
00682     bool Need;
00683     /// Holds the pointers that we need to check.
00684     SmallVector<TrackingVH<Value>, 2> Pointers;
00685     /// Holds the pointer value at the beginning of the loop.
00686     SmallVector<const SCEV*, 2> Starts;
00687     /// Holds the pointer value at the end of the loop.
00688     SmallVector<const SCEV*, 2> Ends;
00689     /// Holds the information if this pointer is used for writing to memory.
00690     SmallVector<bool, 2> IsWritePtr;
00691     /// Holds the id of the set of pointers that could be dependent because of a
00692     /// shared underlying object.
00693     SmallVector<unsigned, 2> DependencySetId;
00694     /// Holds the id of the disjoint alias set to which this pointer belongs.
00695     SmallVector<unsigned, 2> AliasSetId;
00696   };
00697 
00698   /// A struct for saving information about induction variables.
00699   struct InductionInfo {
00700     InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
00701     InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
00702     /// Start value.
00703     TrackingVH<Value> StartValue;
00704     /// Induction kind.
00705     InductionKind IK;
00706   };
00707 
00708   /// ReductionList contains the reduction descriptors for all
00709   /// of the reductions that were found in the loop.
00710   typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
00711 
00712   /// InductionList saves induction variables and maps them to the
00713   /// induction descriptor.
00714   typedef MapVector<PHINode*, InductionInfo> InductionList;
00715 
00716   /// Returns true if it is legal to vectorize this loop.
00717   /// This does not mean that it is profitable to vectorize this
00718   /// loop, only that it is legal to do so.
00719   bool canVectorize();
00720 
00721   /// Returns the Induction variable.
00722   PHINode *getInduction() { return Induction; }
00723 
00724   /// Returns the reduction variables found in the loop.
00725   ReductionList *getReductionVars() { return &Reductions; }
00726 
00727   /// Returns the induction variables found in the loop.
00728   InductionList *getInductionVars() { return &Inductions; }
00729 
00730   /// Returns the widest induction type.
00731   Type *getWidestInductionType() { return WidestIndTy; }
00732 
00733   /// Returns True if V is an induction variable in this loop.
00734   bool isInductionVariable(const Value *V);
00735 
00736   /// Return true if the block BB needs to be predicated in order for the loop
00737   /// to be vectorized.
00738   bool blockNeedsPredication(BasicBlock *BB);
00739 
00740   /// Check if this  pointer is consecutive when vectorizing. This happens
00741   /// when the last index of the GEP is the induction variable, or that the
00742   /// pointer itself is an induction variable.
00743   /// This check allows us to vectorize A[idx] into a wide load/store.
00744   /// Returns:
00745   /// 0 - Stride is unknown or non-consecutive.
00746   /// 1 - Address is consecutive.
00747   /// -1 - Address is consecutive, and decreasing.
00748   int isConsecutivePtr(Value *Ptr);
00749 
00750   /// Returns true if the value V is uniform within the loop.
00751   bool isUniform(Value *V);
00752 
00753   /// Returns true if this instruction will remain scalar after vectorization.
00754   bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
00755 
00756   /// Returns the information that we collected about runtime memory check.
00757   RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
00758 
00759   /// This function returns the identity element (or neutral element) for
00760   /// the operation K.
00761   static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
00762 
00763   unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
00764 
00765   bool hasStride(Value *V) { return StrideSet.count(V); }
00766   bool mustCheckStrides() { return !StrideSet.empty(); }
00767   SmallPtrSet<Value *, 8>::iterator strides_begin() {
00768     return StrideSet.begin();
00769   }
00770   SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
00771 
00772   bool canPredicateStore(Type *DataType, Value *Ptr) {
00773     return TTI->isLegalPredicatedStore(DataType, isConsecutivePtr(Ptr));
00774   }
00775   bool canPredicateLoad(Type *DataType, Value *Ptr) {
00776     return TTI->isLegalPredicatedLoad(DataType, isConsecutivePtr(Ptr));
00777   }
00778   bool setMaskedOp(const Instruction* I) {
00779     return (MaskedOp.find(I) != MaskedOp.end());
00780   }
00781 private:
00782   /// Check if a single basic block loop is vectorizable.
00783   /// At this point we know that this is a loop with a constant trip count
00784   /// and we only need to check individual instructions.
00785   bool canVectorizeInstrs();
00786 
00787   /// When we vectorize loops we may change the order in which
00788   /// we read and write from memory. This method checks if it is
00789   /// legal to vectorize the code, considering only memory constrains.
00790   /// Returns true if the loop is vectorizable
00791   bool canVectorizeMemory();
00792 
00793   /// Return true if we can vectorize this loop using the IF-conversion
00794   /// transformation.
00795   bool canVectorizeWithIfConvert();
00796 
00797   /// Collect the variables that need to stay uniform after vectorization.
00798   void collectLoopUniforms();
00799 
00800   /// Return true if all of the instructions in the block can be speculatively
00801   /// executed. \p SafePtrs is a list of addresses that are known to be legal
00802   /// and we know that we can read from them without segfault.
00803   bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
00804 
00805   /// Returns True, if 'Phi' is the kind of reduction variable for type
00806   /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
00807   bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
00808   /// Returns a struct describing if the instruction 'I' can be a reduction
00809   /// variable of type 'Kind'. If the reduction is a min/max pattern of
00810   /// select(icmp()) this function advances the instruction pointer 'I' from the
00811   /// compare instruction to the select instruction and stores this pointer in
00812   /// 'PatternLastInst' member of the returned struct.
00813   ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
00814                                      ReductionInstDesc &Desc);
00815   /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
00816   /// pattern corresponding to a min(X, Y) or max(X, Y).
00817   static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
00818                                                     ReductionInstDesc &Prev);
00819   /// Returns the induction kind of Phi. This function may return NoInduction
00820   /// if the PHI is not an induction variable.
00821   InductionKind isInductionVariable(PHINode *Phi);
00822 
00823   /// \brief Collect memory access with loop invariant strides.
00824   ///
00825   /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
00826   /// invariant.
00827   void collectStridedAcccess(Value *LoadOrStoreInst);
00828 
00829   /// Report an analysis message to assist the user in diagnosing loops that are
00830   /// not vectorized.
00831   void emitAnalysis(Report &Message) {
00832     DebugLoc DL = TheLoop->getStartLoc();
00833     if (Instruction *I = Message.getInstr())
00834       DL = I->getDebugLoc();
00835     emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
00836                                    *TheFunction, DL, Message.str());
00837   }
00838 
00839   /// The loop that we evaluate.
00840   Loop *TheLoop;
00841   /// Scev analysis.
00842   ScalarEvolution *SE;
00843   /// DataLayout analysis.
00844   const DataLayout *DL;
00845   /// Dominators.
00846   DominatorTree *DT;
00847   /// Target Library Info.
00848   TargetLibraryInfo *TLI;
00849   /// Alias analysis.
00850   AliasAnalysis *AA;
00851   /// Parent function
00852   Function *TheFunction;
00853   /// Target Transform Info
00854   const TargetTransformInfo *TTI;
00855 
00856   //  ---  vectorization state --- //
00857 
00858   /// Holds the integer induction variable. This is the counter of the
00859   /// loop.
00860   PHINode *Induction;
00861   /// Holds the reduction variables.
00862   ReductionList Reductions;
00863   /// Holds all of the induction variables that we found in the loop.
00864   /// Notice that inductions don't need to start at zero and that induction
00865   /// variables can be pointers.
00866   InductionList Inductions;
00867   /// Holds the widest induction type encountered.
00868   Type *WidestIndTy;
00869 
00870   /// Allowed outside users. This holds the reduction
00871   /// vars which can be accessed from outside the loop.
00872   SmallPtrSet<Value*, 4> AllowedExit;
00873   /// This set holds the variables which are known to be uniform after
00874   /// vectorization.
00875   SmallPtrSet<Instruction*, 4> Uniforms;
00876   /// We need to check that all of the pointers in this list are disjoint
00877   /// at runtime.
00878   RuntimePointerCheck PtrRtCheck;
00879   /// Can we assume the absence of NaNs.
00880   bool HasFunNoNaNAttr;
00881 
00882   unsigned MaxSafeDepDistBytes;
00883 
00884   ValueToValueMap Strides;
00885   SmallPtrSet<Value *, 8> StrideSet;
00886   
00887   /// While vectorizing these instructions we have to generate a
00888   /// call to an appropriate masked intrinsic
00889   std::set<const Instruction*> MaskedOp;
00890 };
00891 
00892 /// LoopVectorizationCostModel - estimates the expected speedups due to
00893 /// vectorization.
00894 /// In many cases vectorization is not profitable. This can happen because of
00895 /// a number of reasons. In this class we mainly attempt to predict the
00896 /// expected speedup/slowdowns due to the supported instruction set. We use the
00897 /// TargetTransformInfo to query the different backends for the cost of
00898 /// different operations.
00899 class LoopVectorizationCostModel {
00900 public:
00901   LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
00902                              LoopVectorizationLegality *Legal,
00903                              const TargetTransformInfo &TTI,
00904                              const DataLayout *DL, const TargetLibraryInfo *TLI,
00905                              AssumptionTracker *AT, const Function *F,
00906                              const LoopVectorizeHints *Hints)
00907       : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
00908         TheFunction(F), Hints(Hints) {
00909     CodeMetrics::collectEphemeralValues(L, AT, EphValues);
00910   }
00911 
00912   /// Information about vectorization costs
00913   struct VectorizationFactor {
00914     unsigned Width; // Vector width with best cost
00915     unsigned Cost; // Cost of the loop with that width
00916   };
00917   /// \return The most profitable vectorization factor and the cost of that VF.
00918   /// This method checks every power of two up to VF. If UserVF is not ZERO
00919   /// then this vectorization factor will be selected if vectorization is
00920   /// possible.
00921   VectorizationFactor selectVectorizationFactor(bool OptForSize);
00922 
00923   /// \return The size (in bits) of the widest type in the code that
00924   /// needs to be vectorized. We ignore values that remain scalar such as
00925   /// 64 bit loop indices.
00926   unsigned getWidestType();
00927 
00928   /// \return The most profitable unroll factor.
00929   /// If UserUF is non-zero then this method finds the best unroll-factor
00930   /// based on register pressure and other parameters.
00931   /// VF and LoopCost are the selected vectorization factor and the cost of the
00932   /// selected VF.
00933   unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
00934 
00935   /// \brief A struct that represents some properties of the register usage
00936   /// of a loop.
00937   struct RegisterUsage {
00938     /// Holds the number of loop invariant values that are used in the loop.
00939     unsigned LoopInvariantRegs;
00940     /// Holds the maximum number of concurrent live intervals in the loop.
00941     unsigned MaxLocalUsers;
00942     /// Holds the number of instructions in the loop.
00943     unsigned NumInstructions;
00944   };
00945 
00946   /// \return  information about the register usage of the loop.
00947   RegisterUsage calculateRegisterUsage();
00948 
00949 private:
00950   /// Returns the expected execution cost. The unit of the cost does
00951   /// not matter because we use the 'cost' units to compare different
00952   /// vector widths. The cost that is returned is *not* normalized by
00953   /// the factor width.
00954   unsigned expectedCost(unsigned VF);
00955 
00956   /// Returns the execution time cost of an instruction for a given vector
00957   /// width. Vector width of one means scalar.
00958   unsigned getInstructionCost(Instruction *I, unsigned VF);
00959 
00960   /// A helper function for converting Scalar types to vector types.
00961   /// If the incoming type is void, we return void. If the VF is 1, we return
00962   /// the scalar type.
00963   static Type* ToVectorTy(Type *Scalar, unsigned VF);
00964 
00965   /// Returns whether the instruction is a load or store and will be a emitted
00966   /// as a vector operation.
00967   bool isConsecutiveLoadOrStore(Instruction *I);
00968 
00969   /// Report an analysis message to assist the user in diagnosing loops that are
00970   /// not vectorized.
00971   void emitAnalysis(Report &Message) {
00972     DebugLoc DL = TheLoop->getStartLoc();
00973     if (Instruction *I = Message.getInstr())
00974       DL = I->getDebugLoc();
00975     emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
00976                                    *TheFunction, DL, Message.str());
00977   }
00978 
00979   /// Values used only by @llvm.assume calls.
00980   SmallPtrSet<const Value *, 32> EphValues;
00981 
00982   /// The loop that we evaluate.
00983   Loop *TheLoop;
00984   /// Scev analysis.
00985   ScalarEvolution *SE;
00986   /// Loop Info analysis.
00987   LoopInfo *LI;
00988   /// Vectorization legality.
00989   LoopVectorizationLegality *Legal;
00990   /// Vector target information.
00991   const TargetTransformInfo &TTI;
00992   /// Target data layout information.
00993   const DataLayout *DL;
00994   /// Target Library Info.
00995   const TargetLibraryInfo *TLI;
00996   const Function *TheFunction;
00997   // Loop Vectorize Hint.
00998   const LoopVectorizeHints *Hints;
00999 };
01000 
01001 /// Utility class for getting and setting loop vectorizer hints in the form
01002 /// of loop metadata.
01003 /// This class keeps a number of loop annotations locally (as member variables)
01004 /// and can, upon request, write them back as metadata on the loop. It will
01005 /// initially scan the loop for existing metadata, and will update the local
01006 /// values based on information in the loop.
01007 /// We cannot write all values to metadata, as the mere presence of some info,
01008 /// for example 'force', means a decision has been made. So, we need to be
01009 /// careful NOT to add them if the user hasn't specifically asked so.
01010 class LoopVectorizeHints {
01011   enum HintKind {
01012     HK_WIDTH,
01013     HK_UNROLL,
01014     HK_FORCE
01015   };
01016 
01017   /// Hint - associates name and validation with the hint value.
01018   struct Hint {
01019     const char * Name;
01020     unsigned Value; // This may have to change for non-numeric values.
01021     HintKind Kind;
01022 
01023     Hint(const char * Name, unsigned Value, HintKind Kind)
01024       : Name(Name), Value(Value), Kind(Kind) { }
01025 
01026     bool validate(unsigned Val) {
01027       switch (Kind) {
01028       case HK_WIDTH:
01029         return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
01030       case HK_UNROLL:
01031         return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
01032       case HK_FORCE:
01033         return (Val <= 1);
01034       }
01035       return false;
01036     }
01037   };
01038 
01039   /// Vectorization width.
01040   Hint Width;
01041   /// Vectorization interleave factor.
01042   Hint Interleave;
01043   /// Vectorization forced
01044   Hint Force;
01045 
01046   /// Return the loop metadata prefix.
01047   static StringRef Prefix() { return "llvm.loop."; }
01048 
01049 public:
01050   enum ForceKind {
01051     FK_Undefined = -1, ///< Not selected.
01052     FK_Disabled = 0,   ///< Forcing disabled.
01053     FK_Enabled = 1,    ///< Forcing enabled.
01054   };
01055 
01056   LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
01057       : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
01058         Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
01059         Force("vectorize.enable", FK_Undefined, HK_FORCE),
01060         TheLoop(L) {
01061     // Populate values with existing loop metadata.
01062     getHintsFromMetadata();
01063 
01064     // force-vector-interleave overrides DisableInterleaving.
01065     if (VectorizationInterleave.getNumOccurrences() > 0)
01066       Interleave.Value = VectorizationInterleave;
01067 
01068     DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
01069           << "LV: Interleaving disabled by the pass manager\n");
01070   }
01071 
01072   /// Mark the loop L as already vectorized by setting the width to 1.
01073   void setAlreadyVectorized() {
01074     Width.Value = Interleave.Value = 1;
01075     Hint Hints[] = {Width, Interleave};
01076     writeHintsToMetadata(Hints);
01077   }
01078 
01079   /// Dumps all the hint information.
01080   std::string emitRemark() const {
01081     Report R;
01082     if (Force.Value == LoopVectorizeHints::FK_Disabled)
01083       R << "vectorization is explicitly disabled";
01084     else {
01085       R << "use -Rpass-analysis=loop-vectorize for more info";
01086       if (Force.Value == LoopVectorizeHints::FK_Enabled) {
01087         R << " (Force=true";
01088         if (Width.Value != 0)
01089           R << ", Vector Width=" << Width.Value;
01090         if (Interleave.Value != 0)
01091           R << ", Interleave Count=" << Interleave.Value;
01092         R << ")";
01093       }
01094     }
01095 
01096     return R.str();
01097   }
01098 
01099   unsigned getWidth() const { return Width.Value; }
01100   unsigned getInterleave() const { return Interleave.Value; }
01101   enum ForceKind getForce() const { return (ForceKind)Force.Value; }
01102 
01103 private:
01104   /// Find hints specified in the loop metadata and update local values.
01105   void getHintsFromMetadata() {
01106     MDNode *LoopID = TheLoop->getLoopID();
01107     if (!LoopID)
01108       return;
01109 
01110     // First operand should refer to the loop id itself.
01111     assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
01112     assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
01113 
01114     for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
01115       const MDString *S = nullptr;
01116       SmallVector<Value*, 4> Args;
01117 
01118       // The expected hint is either a MDString or a MDNode with the first
01119       // operand a MDString.
01120       if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
01121         if (!MD || MD->getNumOperands() == 0)
01122           continue;
01123         S = dyn_cast<MDString>(MD->getOperand(0));
01124         for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
01125           Args.push_back(MD->getOperand(i));
01126       } else {
01127         S = dyn_cast<MDString>(LoopID->getOperand(i));
01128         assert(Args.size() == 0 && "too many arguments for MDString");
01129       }
01130 
01131       if (!S)
01132         continue;
01133 
01134       // Check if the hint starts with the loop metadata prefix.
01135       StringRef Name = S->getString();
01136       if (Args.size() == 1)
01137         setHint(Name, Args[0]);
01138     }
01139   }
01140 
01141   /// Checks string hint with one operand and set value if valid.
01142   void setHint(StringRef Name, Value *Arg) {
01143     if (!Name.startswith(Prefix()))
01144       return;
01145     Name = Name.substr(Prefix().size(), StringRef::npos);
01146 
01147     const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
01148     if (!C) return;
01149     unsigned Val = C->getZExtValue();
01150 
01151     Hint *Hints[] = {&Width, &Interleave, &Force};
01152     for (auto H : Hints) {
01153       if (Name == H->Name) {
01154         if (H->validate(Val))
01155           H->Value = Val;
01156         else
01157           DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
01158         break;
01159       }
01160     }
01161   }
01162 
01163   /// Create a new hint from name / value pair.
01164   MDNode *createHintMetadata(StringRef Name, unsigned V) const {
01165     LLVMContext &Context = TheLoop->getHeader()->getContext();
01166     Value *Vals[] = {MDString::get(Context, Name),
01167                      ConstantInt::get(Type::getInt32Ty(Context), V)};
01168     return MDNode::get(Context, Vals);
01169   }
01170 
01171   /// Matches metadata with hint name.
01172   bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
01173     MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
01174     if (!Name)
01175       return false;
01176 
01177     for (auto H : HintTypes)
01178       if (Name->getString().endswith(H.Name))
01179         return true;
01180     return false;
01181   }
01182 
01183   /// Sets current hints into loop metadata, keeping other values intact.
01184   void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
01185     if (HintTypes.size() == 0)
01186       return;
01187 
01188     // Reserve the first element to LoopID (see below).
01189     SmallVector<Value*, 4> Vals(1);
01190     // If the loop already has metadata, then ignore the existing operands.
01191     MDNode *LoopID = TheLoop->getLoopID();
01192     if (LoopID) {
01193       for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
01194         MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
01195         // If node in update list, ignore old value.
01196         if (!matchesHintMetadataName(Node, HintTypes))
01197           Vals.push_back(Node);
01198       }
01199     }
01200 
01201     // Now, add the missing hints.
01202     for (auto H : HintTypes)
01203       Vals.push_back(
01204           createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
01205 
01206     // Replace current metadata node with new one.
01207     LLVMContext &Context = TheLoop->getHeader()->getContext();
01208     MDNode *NewLoopID = MDNode::get(Context, Vals);
01209     // Set operand 0 to refer to the loop id itself.
01210     NewLoopID->replaceOperandWith(0, NewLoopID);
01211 
01212     TheLoop->setLoopID(NewLoopID);
01213     if (LoopID)
01214       LoopID->replaceAllUsesWith(NewLoopID);
01215     LoopID = NewLoopID;
01216   }
01217 
01218   /// The loop these hints belong to.
01219   const Loop *TheLoop;
01220 };
01221 
01222 static void emitMissedWarning(Function *F, Loop *L,
01223                               const LoopVectorizeHints &LH) {
01224   emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
01225                                L->getStartLoc(), LH.emitRemark());
01226 
01227   if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
01228     if (LH.getWidth() != 1)
01229       emitLoopVectorizeWarning(
01230           F->getContext(), *F, L->getStartLoc(),
01231           "failed explicitly specified loop vectorization");
01232     else if (LH.getInterleave() != 1)
01233       emitLoopInterleaveWarning(
01234           F->getContext(), *F, L->getStartLoc(),
01235           "failed explicitly specified loop interleaving");
01236   }
01237 }
01238 
01239 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
01240   if (L.empty())
01241     return V.push_back(&L);
01242 
01243   for (Loop *InnerL : L)
01244     addInnerLoop(*InnerL, V);
01245 }
01246 
01247 /// The LoopVectorize Pass.
01248 struct LoopVectorize : public FunctionPass {
01249   /// Pass identification, replacement for typeid
01250   static char ID;
01251 
01252   explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
01253     : FunctionPass(ID),
01254       DisableUnrolling(NoUnrolling),
01255       AlwaysVectorize(AlwaysVectorize) {
01256     initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
01257   }
01258 
01259   ScalarEvolution *SE;
01260   const DataLayout *DL;
01261   LoopInfo *LI;
01262   TargetTransformInfo *TTI;
01263   DominatorTree *DT;
01264   BlockFrequencyInfo *BFI;
01265   TargetLibraryInfo *TLI;
01266   AliasAnalysis *AA;
01267   AssumptionTracker *AT;
01268   bool DisableUnrolling;
01269   bool AlwaysVectorize;
01270 
01271   BlockFrequency ColdEntryFreq;
01272 
01273   bool runOnFunction(Function &F) override {
01274     SE = &getAnalysis<ScalarEvolution>();
01275     DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
01276     DL = DLP ? &DLP->getDataLayout() : nullptr;
01277     LI = &getAnalysis<LoopInfo>();
01278     TTI = &getAnalysis<TargetTransformInfo>();
01279     DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
01280     BFI = &getAnalysis<BlockFrequencyInfo>();
01281     TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
01282     AA = &getAnalysis<AliasAnalysis>();
01283     AT = &getAnalysis<AssumptionTracker>();
01284 
01285     // Compute some weights outside of the loop over the loops. Compute this
01286     // using a BranchProbability to re-use its scaling math.
01287     const BranchProbability ColdProb(1, 5); // 20%
01288     ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
01289 
01290     // If the target claims to have no vector registers don't attempt
01291     // vectorization.
01292     if (!TTI->getNumberOfRegisters(true))
01293       return false;
01294 
01295     if (!DL) {
01296       DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
01297                    << ": Missing data layout\n");
01298       return false;
01299     }
01300 
01301     // Build up a worklist of inner-loops to vectorize. This is necessary as
01302     // the act of vectorizing or partially unrolling a loop creates new loops
01303     // and can invalidate iterators across the loops.
01304     SmallVector<Loop *, 8> Worklist;
01305 
01306     for (Loop *L : *LI)
01307       addInnerLoop(*L, Worklist);
01308 
01309     LoopsAnalyzed += Worklist.size();
01310 
01311     // Now walk the identified inner loops.
01312     bool Changed = false;
01313     while (!Worklist.empty())
01314       Changed |= processLoop(Worklist.pop_back_val());
01315 
01316     // Process each loop nest in the function.
01317     return Changed;
01318   }
01319 
01320   bool processLoop(Loop *L) {
01321     assert(L->empty() && "Only process inner loops.");
01322 
01323 #ifndef NDEBUG
01324     const std::string DebugLocStr = getDebugLocString(L);
01325 #endif /* NDEBUG */
01326 
01327     DEBUG(dbgs() << "\nLV: Checking a loop in \""
01328                  << L->getHeader()->getParent()->getName() << "\" from "
01329                  << DebugLocStr << "\n");
01330 
01331     LoopVectorizeHints Hints(L, DisableUnrolling);
01332 
01333     DEBUG(dbgs() << "LV: Loop hints:"
01334                  << " force="
01335                  << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
01336                          ? "disabled"
01337                          : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
01338                                 ? "enabled"
01339                                 : "?")) << " width=" << Hints.getWidth()
01340                  << " unroll=" << Hints.getInterleave() << "\n");
01341 
01342     // Function containing loop
01343     Function *F = L->getHeader()->getParent();
01344 
01345     // Looking at the diagnostic output is the only way to determine if a loop
01346     // was vectorized (other than looking at the IR or machine code), so it
01347     // is important to generate an optimization remark for each loop. Most of
01348     // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
01349     // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
01350     // less verbose reporting vectorized loops and unvectorized loops that may
01351     // benefit from vectorization, respectively.
01352 
01353     if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
01354       DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
01355       emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
01356                                      L->getStartLoc(), Hints.emitRemark());
01357       return false;
01358     }
01359 
01360     if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
01361       DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
01362       emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
01363                                      L->getStartLoc(), Hints.emitRemark());
01364       return false;
01365     }
01366 
01367     if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
01368       DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
01369       emitOptimizationRemarkAnalysis(
01370           F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01371           "loop not vectorized: vector width and interleave count are "
01372           "explicitly set to 1");
01373       return false;
01374     }
01375 
01376     // Check the loop for a trip count threshold:
01377     // do not vectorize loops with a tiny trip count.
01378     const unsigned TC = SE->getSmallConstantTripCount(L);
01379     if (TC > 0u && TC < TinyTripCountVectorThreshold) {
01380       DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
01381                    << "This loop is not worth vectorizing.");
01382       if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
01383         DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
01384       else {
01385         DEBUG(dbgs() << "\n");
01386         emitOptimizationRemarkAnalysis(
01387             F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01388             "vectorization is not beneficial and is not explicitly forced");
01389         return false;
01390       }
01391     }
01392 
01393     // Check if it is legal to vectorize the loop.
01394     LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
01395     if (!LVL.canVectorize()) {
01396       DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
01397       emitMissedWarning(F, L, Hints);
01398       return false;
01399     }
01400 
01401     // Use the cost model.
01402     LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AT, F,
01403                                   &Hints);
01404 
01405     // Check the function attributes to find out if this function should be
01406     // optimized for size.
01407     bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
01408                       F->hasFnAttribute(Attribute::OptimizeForSize);
01409 
01410     // Compute the weighted frequency of this loop being executed and see if it
01411     // is less than 20% of the function entry baseline frequency. Note that we
01412     // always have a canonical loop here because we think we *can* vectoriez.
01413     // FIXME: This is hidden behind a flag due to pervasive problems with
01414     // exactly what block frequency models.
01415     if (LoopVectorizeWithBlockFrequency) {
01416       BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
01417       if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
01418           LoopEntryFreq < ColdEntryFreq)
01419         OptForSize = true;
01420     }
01421 
01422     // Check the function attributes to see if implicit floats are allowed.a
01423     // FIXME: This check doesn't seem possibly correct -- what if the loop is
01424     // an integer loop and the vector instructions selected are purely integer
01425     // vector instructions?
01426     if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
01427       DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
01428             "attribute is used.\n");
01429       emitOptimizationRemarkAnalysis(
01430           F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01431           "loop not vectorized due to NoImplicitFloat attribute");
01432       emitMissedWarning(F, L, Hints);
01433       return false;
01434     }
01435 
01436     // Select the optimal vectorization factor.
01437     const LoopVectorizationCostModel::VectorizationFactor VF =
01438         CM.selectVectorizationFactor(OptForSize);
01439 
01440     // Select the unroll factor.
01441     const unsigned UF =
01442         CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
01443 
01444     DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
01445                  << DebugLocStr << '\n');
01446     DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
01447 
01448     if (VF.Width == 1) {
01449       DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
01450 
01451       if (UF == 1) {
01452         emitOptimizationRemarkAnalysis(
01453             F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01454             "not beneficial to vectorize and user disabled interleaving");
01455         return false;
01456       }
01457       DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
01458 
01459       // Report the unrolling decision.
01460       emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01461                              Twine("unrolled with interleaving factor " +
01462                                    Twine(UF) +
01463                                    " (vectorization not beneficial)"));
01464 
01465       // We decided not to vectorize, but we may want to unroll.
01466 
01467       InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
01468       Unroller.vectorize(&LVL);
01469     } else {
01470       // If we decided that it is *legal* to vectorize the loop then do it.
01471       InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
01472       LB.vectorize(&LVL);
01473       ++LoopsVectorized;
01474 
01475       // Report the vectorization decision.
01476       emitOptimizationRemark(
01477           F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01478           Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
01479               ", unrolling interleave factor: " + Twine(UF) + ")");
01480     }
01481 
01482     // Mark the loop as already vectorized to avoid vectorizing again.
01483     Hints.setAlreadyVectorized();
01484 
01485     DEBUG(verifyFunction(*L->getHeader()->getParent()));
01486     return true;
01487   }
01488 
01489   void getAnalysisUsage(AnalysisUsage &AU) const override {
01490     AU.addRequired<AssumptionTracker>();
01491     AU.addRequiredID(LoopSimplifyID);
01492     AU.addRequiredID(LCSSAID);
01493     AU.addRequired<BlockFrequencyInfo>();
01494     AU.addRequired<DominatorTreeWrapperPass>();
01495     AU.addRequired<LoopInfo>();
01496     AU.addRequired<ScalarEvolution>();
01497     AU.addRequired<TargetTransformInfo>();
01498     AU.addRequired<AliasAnalysis>();
01499     AU.addPreserved<LoopInfo>();
01500     AU.addPreserved<DominatorTreeWrapperPass>();
01501     AU.addPreserved<AliasAnalysis>();
01502   }
01503 
01504 };
01505 
01506 } // end anonymous namespace
01507 
01508 //===----------------------------------------------------------------------===//
01509 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
01510 // LoopVectorizationCostModel.
01511 //===----------------------------------------------------------------------===//
01512 
01513 static Value *stripIntegerCast(Value *V) {
01514   if (CastInst *CI = dyn_cast<CastInst>(V))
01515     if (CI->getOperand(0)->getType()->isIntegerTy())
01516       return CI->getOperand(0);
01517   return V;
01518 }
01519 
01520 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
01521 ///
01522 /// If \p OrigPtr is not null, use it to look up the stride value instead of
01523 /// \p Ptr.
01524 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
01525                                              ValueToValueMap &PtrToStride,
01526                                              Value *Ptr, Value *OrigPtr = nullptr) {
01527 
01528   const SCEV *OrigSCEV = SE->getSCEV(Ptr);
01529 
01530   // If there is an entry in the map return the SCEV of the pointer with the
01531   // symbolic stride replaced by one.
01532   ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
01533   if (SI != PtrToStride.end()) {
01534     Value *StrideVal = SI->second;
01535 
01536     // Strip casts.
01537     StrideVal = stripIntegerCast(StrideVal);
01538 
01539     // Replace symbolic stride by one.
01540     Value *One = ConstantInt::get(StrideVal->getType(), 1);
01541     ValueToValueMap RewriteMap;
01542     RewriteMap[StrideVal] = One;
01543 
01544     const SCEV *ByOne =
01545         SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
01546     DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
01547                  << "\n");
01548     return ByOne;
01549   }
01550 
01551   // Otherwise, just return the SCEV of the original pointer.
01552   return SE->getSCEV(Ptr);
01553 }
01554 
01555 void LoopVectorizationLegality::RuntimePointerCheck::insert(
01556     ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
01557     unsigned ASId, ValueToValueMap &Strides) {
01558   // Get the stride replaced scev.
01559   const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
01560   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
01561   assert(AR && "Invalid addrec expression");
01562   const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
01563   const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
01564   Pointers.push_back(Ptr);
01565   Starts.push_back(AR->getStart());
01566   Ends.push_back(ScEnd);
01567   IsWritePtr.push_back(WritePtr);
01568   DependencySetId.push_back(DepSetId);
01569   AliasSetId.push_back(ASId);
01570 }
01571 
01572 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
01573   // We need to place the broadcast of invariant variables outside the loop.
01574   Instruction *Instr = dyn_cast<Instruction>(V);
01575   bool NewInstr =
01576       (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
01577                           Instr->getParent()) != LoopVectorBody.end());
01578   bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
01579 
01580   // Place the code for broadcasting invariant variables in the new preheader.
01581   IRBuilder<>::InsertPointGuard Guard(Builder);
01582   if (Invariant)
01583     Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
01584 
01585   // Broadcast the scalar into all locations in the vector.
01586   Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
01587 
01588   return Shuf;
01589 }
01590 
01591 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
01592                                                  bool Negate) {
01593   assert(Val->getType()->isVectorTy() && "Must be a vector");
01594   assert(Val->getType()->getScalarType()->isIntegerTy() &&
01595          "Elem must be an integer");
01596   // Create the types.
01597   Type *ITy = Val->getType()->getScalarType();
01598   VectorType *Ty = cast<VectorType>(Val->getType());
01599   int VLen = Ty->getNumElements();
01600   SmallVector<Constant*, 8> Indices;
01601 
01602   // Create a vector of consecutive numbers from zero to VF.
01603   for (int i = 0; i < VLen; ++i) {
01604     int64_t Idx = Negate ? (-i) : i;
01605     Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
01606   }
01607 
01608   // Add the consecutive indices to the vector value.
01609   Constant *Cv = ConstantVector::get(Indices);
01610   assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
01611   return Builder.CreateAdd(Val, Cv, "induction");
01612 }
01613 
01614 /// \brief Find the operand of the GEP that should be checked for consecutive
01615 /// stores. This ignores trailing indices that have no effect on the final
01616 /// pointer.
01617 static unsigned getGEPInductionOperand(const DataLayout *DL,
01618                                        const GetElementPtrInst *Gep) {
01619   unsigned LastOperand = Gep->getNumOperands() - 1;
01620   unsigned GEPAllocSize = DL->getTypeAllocSize(
01621       cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
01622 
01623   // Walk backwards and try to peel off zeros.
01624   while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
01625     // Find the type we're currently indexing into.
01626     gep_type_iterator GEPTI = gep_type_begin(Gep);
01627     std::advance(GEPTI, LastOperand - 1);
01628 
01629     // If it's a type with the same allocation size as the result of the GEP we
01630     // can peel off the zero index.
01631     if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
01632       break;
01633     --LastOperand;
01634   }
01635 
01636   return LastOperand;
01637 }
01638 
01639 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
01640   assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
01641   // Make sure that the pointer does not point to structs.
01642   if (Ptr->getType()->getPointerElementType()->isAggregateType())
01643     return 0;
01644 
01645   // If this value is a pointer induction variable we know it is consecutive.
01646   PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
01647   if (Phi && Inductions.count(Phi)) {
01648     InductionInfo II = Inductions[Phi];
01649     if (IK_PtrInduction == II.IK)
01650       return 1;
01651     else if (IK_ReversePtrInduction == II.IK)
01652       return -1;
01653   }
01654 
01655   GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
01656   if (!Gep)
01657     return 0;
01658 
01659   unsigned NumOperands = Gep->getNumOperands();
01660   Value *GpPtr = Gep->getPointerOperand();
01661   // If this GEP value is a consecutive pointer induction variable and all of
01662   // the indices are constant then we know it is consecutive. We can
01663   Phi = dyn_cast<PHINode>(GpPtr);
01664   if (Phi && Inductions.count(Phi)) {
01665 
01666     // Make sure that the pointer does not point to structs.
01667     PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
01668     if (GepPtrType->getElementType()->isAggregateType())
01669       return 0;
01670 
01671     // Make sure that all of the index operands are loop invariant.
01672     for (unsigned i = 1; i < NumOperands; ++i)
01673       if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
01674         return 0;
01675 
01676     InductionInfo II = Inductions[Phi];
01677     if (IK_PtrInduction == II.IK)
01678       return 1;
01679     else if (IK_ReversePtrInduction == II.IK)
01680       return -1;
01681   }
01682 
01683   unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
01684 
01685   // Check that all of the gep indices are uniform except for our induction
01686   // operand.
01687   for (unsigned i = 0; i != NumOperands; ++i)
01688     if (i != InductionOperand &&
01689         !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
01690       return 0;
01691 
01692   // We can emit wide load/stores only if the last non-zero index is the
01693   // induction variable.
01694   const SCEV *Last = nullptr;
01695   if (!Strides.count(Gep))
01696     Last = SE->getSCEV(Gep->getOperand(InductionOperand));
01697   else {
01698     // Because of the multiplication by a stride we can have a s/zext cast.
01699     // We are going to replace this stride by 1 so the cast is safe to ignore.
01700     //
01701     //  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
01702     //  %0 = trunc i64 %indvars.iv to i32
01703     //  %mul = mul i32 %0, %Stride1
01704     //  %idxprom = zext i32 %mul to i64  << Safe cast.
01705     //  %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
01706     //
01707     Last = replaceSymbolicStrideSCEV(SE, Strides,
01708                                      Gep->getOperand(InductionOperand), Gep);
01709     if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
01710       Last =
01711           (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
01712               ? C->getOperand()
01713               : Last;
01714   }
01715   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
01716     const SCEV *Step = AR->getStepRecurrence(*SE);
01717 
01718     // The memory is consecutive because the last index is consecutive
01719     // and all other indices are loop invariant.
01720     if (Step->isOne())
01721       return 1;
01722     if (Step->isAllOnesValue())
01723       return -1;
01724   }
01725 
01726   return 0;
01727 }
01728 
01729 bool LoopVectorizationLegality::isUniform(Value *V) {
01730   return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
01731 }
01732 
01733 InnerLoopVectorizer::VectorParts&
01734 InnerLoopVectorizer::getVectorValue(Value *V) {
01735   assert(V != Induction && "The new induction variable should not be used.");
01736   assert(!V->getType()->isVectorTy() && "Can't widen a vector");
01737 
01738   // If we have a stride that is replaced by one, do it here.
01739   if (Legal->hasStride(V))
01740     V = ConstantInt::get(V->getType(), 1);
01741 
01742   // If we have this scalar in the map, return it.
01743   if (WidenMap.has(V))
01744     return WidenMap.get(V);
01745 
01746   // If this scalar is unknown, assume that it is a constant or that it is
01747   // loop invariant. Broadcast V and save the value for future uses.
01748   Value *B = getBroadcastInstrs(V);
01749   return WidenMap.splat(V, B);
01750 }
01751 
01752 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
01753   assert(Vec->getType()->isVectorTy() && "Invalid type");
01754   SmallVector<Constant*, 8> ShuffleMask;
01755   for (unsigned i = 0; i < VF; ++i)
01756     ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
01757 
01758   return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
01759                                      ConstantVector::get(ShuffleMask),
01760                                      "reverse");
01761 }
01762 
01763 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
01764   // Attempt to issue a wide load.
01765   LoadInst *LI = dyn_cast<LoadInst>(Instr);
01766   StoreInst *SI = dyn_cast<StoreInst>(Instr);
01767 
01768   assert((LI || SI) && "Invalid Load/Store instruction");
01769 
01770   Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
01771   Type *DataTy = VectorType::get(ScalarDataTy, VF);
01772   Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
01773   unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
01774   // An alignment of 0 means target abi alignment. We need to use the scalar's
01775   // target abi alignment in such a case.
01776   if (!Alignment)
01777     Alignment = DL->getABITypeAlignment(ScalarDataTy);
01778   unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
01779   unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
01780   unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
01781 
01782   if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
01783       !Legal->setMaskedOp(SI))
01784     return scalarizeInstruction(Instr, true);
01785 
01786   if (ScalarAllocatedSize != VectorElementSize)
01787     return scalarizeInstruction(Instr);
01788 
01789   // If the pointer is loop invariant or if it is non-consecutive,
01790   // scalarize the load.
01791   int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
01792   bool Reverse = ConsecutiveStride < 0;
01793   bool UniformLoad = LI && Legal->isUniform(Ptr);
01794   if (!ConsecutiveStride || UniformLoad)
01795     return scalarizeInstruction(Instr);
01796 
01797   Constant *Zero = Builder.getInt32(0);
01798   VectorParts &Entry = WidenMap.get(Instr);
01799 
01800   // Handle consecutive loads/stores.
01801   GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
01802   if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
01803     setDebugLocFromInst(Builder, Gep);
01804     Value *PtrOperand = Gep->getPointerOperand();
01805     Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
01806     FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
01807 
01808     // Create the new GEP with the new induction variable.
01809     GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
01810     Gep2->setOperand(0, FirstBasePtr);
01811     Gep2->setName("gep.indvar.base");
01812     Ptr = Builder.Insert(Gep2);
01813   } else if (Gep) {
01814     setDebugLocFromInst(Builder, Gep);
01815     assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
01816                                OrigLoop) && "Base ptr must be invariant");
01817 
01818     // The last index does not have to be the induction. It can be
01819     // consecutive and be a function of the index. For example A[I+1];
01820     unsigned NumOperands = Gep->getNumOperands();
01821     unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
01822     // Create the new GEP with the new induction variable.
01823     GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
01824 
01825     for (unsigned i = 0; i < NumOperands; ++i) {
01826       Value *GepOperand = Gep->getOperand(i);
01827       Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
01828 
01829       // Update last index or loop invariant instruction anchored in loop.
01830       if (i == InductionOperand ||
01831           (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
01832         assert((i == InductionOperand ||
01833                SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
01834                "Must be last index or loop invariant");
01835 
01836         VectorParts &GEPParts = getVectorValue(GepOperand);
01837         Value *Index = GEPParts[0];
01838         Index = Builder.CreateExtractElement(Index, Zero);
01839         Gep2->setOperand(i, Index);
01840         Gep2->setName("gep.indvar.idx");
01841       }
01842     }
01843     Ptr = Builder.Insert(Gep2);
01844   } else {
01845     // Use the induction element ptr.
01846     assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
01847     setDebugLocFromInst(Builder, Ptr);
01848     VectorParts &PtrVal = getVectorValue(Ptr);
01849     Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
01850   }
01851 
01852   // Handle Stores:
01853   if (SI) {
01854     assert(!Legal->isUniform(SI->getPointerOperand()) &&
01855            "We do not allow storing to uniform addresses");
01856     setDebugLocFromInst(Builder, SI);
01857     // We don't want to update the value in the map as it might be used in
01858     // another expression. So don't use a reference type for "StoredVal".
01859     VectorParts StoredVal = getVectorValue(SI->getValueOperand());
01860 
01861     for (unsigned Part = 0; Part < UF; ++Part) {
01862       // Calculate the pointer for the specific unroll-part.
01863       Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
01864 
01865       if (Reverse) {
01866         // If we store to reverse consecutive memory locations then we need
01867         // to reverse the order of elements in the stored value.
01868         StoredVal[Part] = reverseVector(StoredVal[Part]);
01869         // If the address is consecutive but reversed, then the
01870         // wide store needs to start at the last vector element.
01871         PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
01872         PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
01873       }
01874 
01875       Value *VecPtr = Builder.CreateBitCast(PartPtr,
01876                                             DataTy->getPointerTo(AddressSpace));
01877 
01878       Instruction *NewSI;
01879       if (Legal->setMaskedOp(SI)) {
01880         Type *I8PtrTy =
01881         Builder.getInt8PtrTy(PartPtr->getType()->getPointerAddressSpace());
01882 
01883         Value *I8Ptr = Builder.CreateBitCast(PartPtr, I8PtrTy);
01884 
01885         VectorParts Cond = createEdgeMask(SI->getParent()->getSinglePredecessor(),
01886                                           SI->getParent());
01887         SmallVector <Value *, 8> Ops;
01888         Ops.push_back(I8Ptr);
01889         Ops.push_back(StoredVal[Part]);
01890         Ops.push_back(Builder.getInt32(Alignment));
01891         Ops.push_back(Cond[Part]);
01892         NewSI = Builder.CreateMaskedStore(Ops);
01893       }
01894       else 
01895         NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
01896       propagateMetadata(NewSI, SI);
01897     }
01898     return;
01899   }
01900 
01901   // Handle loads.
01902   assert(LI && "Must have a load instruction");
01903   setDebugLocFromInst(Builder, LI);
01904   for (unsigned Part = 0; Part < UF; ++Part) {
01905     // Calculate the pointer for the specific unroll-part.
01906     Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
01907 
01908     if (Reverse) {
01909       // If the address is consecutive but reversed, then the
01910       // wide load needs to start at the last vector element.
01911       PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
01912       PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
01913     }
01914 
01915     Instruction* NewLI;
01916     if (Legal->setMaskedOp(LI)) {
01917       Type *I8PtrTy =
01918         Builder.getInt8PtrTy(PartPtr->getType()->getPointerAddressSpace());
01919 
01920       Value *I8Ptr = Builder.CreateBitCast(PartPtr, I8PtrTy);
01921 
01922       VectorParts SrcMask = createBlockInMask(LI->getParent());
01923       SmallVector <Value *, 8> Ops;
01924       Ops.push_back(I8Ptr);
01925       Ops.push_back(UndefValue::get(DataTy));
01926       Ops.push_back(Builder.getInt32(Alignment));
01927       Ops.push_back(SrcMask[Part]);
01928       NewLI = Builder.CreateMaskedLoad(Ops);
01929     }
01930     else {
01931       Value *VecPtr = Builder.CreateBitCast(PartPtr,
01932                                             DataTy->getPointerTo(AddressSpace));
01933       NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
01934     }
01935     propagateMetadata(NewLI, LI);
01936     Entry[Part] = Reverse ? reverseVector(NewLI) :  NewLI;
01937   }
01938 }
01939 
01940 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
01941   assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
01942   // Holds vector parameters or scalars, in case of uniform vals.
01943   SmallVector<VectorParts, 4> Params;
01944 
01945   setDebugLocFromInst(Builder, Instr);
01946 
01947   // Find all of the vectorized parameters.
01948   for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
01949     Value *SrcOp = Instr->getOperand(op);
01950 
01951     // If we are accessing the old induction variable, use the new one.
01952     if (SrcOp == OldInduction) {
01953       Params.push_back(getVectorValue(SrcOp));
01954       continue;
01955     }
01956 
01957     // Try using previously calculated values.
01958     Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
01959 
01960     // If the src is an instruction that appeared earlier in the basic block
01961     // then it should already be vectorized.
01962     if (SrcInst && OrigLoop->contains(SrcInst)) {
01963       assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
01964       // The parameter is a vector value from earlier.
01965       Params.push_back(WidenMap.get(SrcInst));
01966     } else {
01967       // The parameter is a scalar from outside the loop. Maybe even a constant.
01968       VectorParts Scalars;
01969       Scalars.append(UF, SrcOp);
01970       Params.push_back(Scalars);
01971     }
01972   }
01973 
01974   assert(Params.size() == Instr->getNumOperands() &&
01975          "Invalid number of operands");
01976 
01977   // Does this instruction return a value ?
01978   bool IsVoidRetTy = Instr->getType()->isVoidTy();
01979 
01980   Value *UndefVec = IsVoidRetTy ? nullptr :
01981     UndefValue::get(VectorType::get(Instr->getType(), VF));
01982   // Create a new entry in the WidenMap and initialize it to Undef or Null.
01983   VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
01984 
01985   Instruction *InsertPt = Builder.GetInsertPoint();
01986   BasicBlock *IfBlock = Builder.GetInsertBlock();
01987   BasicBlock *CondBlock = nullptr;
01988 
01989   VectorParts Cond;
01990   Loop *VectorLp = nullptr;
01991   if (IfPredicateStore) {
01992     assert(Instr->getParent()->getSinglePredecessor() &&
01993            "Only support single predecessor blocks");
01994     Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
01995                           Instr->getParent());
01996     VectorLp = LI->getLoopFor(IfBlock);
01997     assert(VectorLp && "Must have a loop for this block");
01998   }
01999 
02000   // For each vector unroll 'part':
02001   for (unsigned Part = 0; Part < UF; ++Part) {
02002     // For each scalar that we create:
02003     for (unsigned Width = 0; Width < VF; ++Width) {
02004 
02005       // Start if-block.
02006       Value *Cmp = nullptr;
02007       if (IfPredicateStore) {
02008         Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
02009         Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
02010         CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
02011         LoopVectorBody.push_back(CondBlock);
02012         VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
02013         // Update Builder with newly created basic block.
02014         Builder.SetInsertPoint(InsertPt);
02015       }
02016 
02017       Instruction *Cloned = Instr->clone();
02018       if (!IsVoidRetTy)
02019         Cloned->setName(Instr->getName() + ".cloned");
02020       // Replace the operands of the cloned instructions with extracted scalars.
02021       for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
02022         Value *Op = Params[op][Part];
02023         // Param is a vector. Need to extract the right lane.
02024         if (Op->getType()->isVectorTy())
02025           Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
02026         Cloned->setOperand(op, Op);
02027       }
02028 
02029       // Place the cloned scalar in the new loop.
02030       Builder.Insert(Cloned);
02031 
02032       // If the original scalar returns a value we need to place it in a vector
02033       // so that future users will be able to use it.
02034       if (!IsVoidRetTy)
02035         VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
02036                                                        Builder.getInt32(Width));
02037       // End if-block.
02038       if (IfPredicateStore) {
02039          BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
02040          LoopVectorBody.push_back(NewIfBlock);
02041          VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
02042          Builder.SetInsertPoint(InsertPt);
02043          Instruction *OldBr = IfBlock->getTerminator();
02044          BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
02045          OldBr->eraseFromParent();
02046          IfBlock = NewIfBlock;
02047       }
02048     }
02049   }
02050 }
02051 
02052 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
02053                                  Instruction *Loc) {
02054   if (FirstInst)
02055     return FirstInst;
02056   if (Instruction *I = dyn_cast<Instruction>(V))
02057     return I->getParent() == Loc->getParent() ? I : nullptr;
02058   return nullptr;
02059 }
02060 
02061 std::pair<Instruction *, Instruction *>
02062 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
02063   Instruction *tnullptr = nullptr;
02064   if (!Legal->mustCheckStrides())
02065     return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
02066 
02067   IRBuilder<> ChkBuilder(Loc);
02068 
02069   // Emit checks.
02070   Value *Check = nullptr;
02071   Instruction *FirstInst = nullptr;
02072   for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
02073                                          SE = Legal->strides_end();
02074        SI != SE; ++SI) {
02075     Value *Ptr = stripIntegerCast(*SI);
02076     Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
02077                                        "stride.chk");
02078     // Store the first instruction we create.
02079     FirstInst = getFirstInst(FirstInst, C, Loc);
02080     if (Check)
02081       Check = ChkBuilder.CreateOr(Check, C);
02082     else
02083       Check = C;
02084   }
02085 
02086   // We have to do this trickery because the IRBuilder might fold the check to a
02087   // constant expression in which case there is no Instruction anchored in a
02088   // the block.
02089   LLVMContext &Ctx = Loc->getContext();
02090   Instruction *TheCheck =
02091       BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
02092   ChkBuilder.Insert(TheCheck, "stride.not.one");
02093   FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
02094 
02095   return std::make_pair(FirstInst, TheCheck);
02096 }
02097 
02098 std::pair<Instruction *, Instruction *>
02099 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
02100   LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
02101   Legal->getRuntimePointerCheck();
02102 
02103   Instruction *tnullptr = nullptr;
02104   if (!PtrRtCheck->Need)
02105     return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
02106 
02107   unsigned NumPointers = PtrRtCheck->Pointers.size();
02108   SmallVector<TrackingVH<Value> , 2> Starts;
02109   SmallVector<TrackingVH<Value> , 2> Ends;
02110 
02111   LLVMContext &Ctx = Loc->getContext();
02112   SCEVExpander Exp(*SE, "induction");
02113   Instruction *FirstInst = nullptr;
02114 
02115   for (unsigned i = 0; i < NumPointers; ++i) {
02116     Value *Ptr = PtrRtCheck->Pointers[i];
02117     const SCEV *Sc = SE->getSCEV(Ptr);
02118 
02119     if (SE->isLoopInvariant(Sc, OrigLoop)) {
02120       DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
02121             *Ptr <<"\n");
02122       Starts.push_back(Ptr);
02123       Ends.push_back(Ptr);
02124     } else {
02125       DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
02126       unsigned AS = Ptr->getType()->getPointerAddressSpace();
02127 
02128       // Use this type for pointer arithmetic.
02129       Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
02130 
02131       Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
02132       Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
02133       Starts.push_back(Start);
02134       Ends.push_back(End);
02135     }
02136   }
02137 
02138   IRBuilder<> ChkBuilder(Loc);
02139   // Our instructions might fold to a constant.
02140   Value *MemoryRuntimeCheck = nullptr;
02141   for (unsigned i = 0; i < NumPointers; ++i) {
02142     for (unsigned j = i+1; j < NumPointers; ++j) {
02143       // No need to check if two readonly pointers intersect.
02144       if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
02145         continue;
02146 
02147       // Only need to check pointers between two different dependency sets.
02148       if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
02149        continue;
02150       // Only need to check pointers in the same alias set.
02151       if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
02152         continue;
02153 
02154       unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
02155       unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
02156 
02157       assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
02158              (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
02159              "Trying to bounds check pointers with different address spaces");
02160 
02161       Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
02162       Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
02163 
02164       Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
02165       Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
02166       Value *End0 =   ChkBuilder.CreateBitCast(Ends[i],   PtrArithTy1, "bc");
02167       Value *End1 =   ChkBuilder.CreateBitCast(Ends[j],   PtrArithTy0, "bc");
02168 
02169       Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
02170       FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
02171       Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
02172       FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
02173       Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
02174       FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
02175       if (MemoryRuntimeCheck) {
02176         IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
02177                                          "conflict.rdx");
02178         FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
02179       }
02180       MemoryRuntimeCheck = IsConflict;
02181     }
02182   }
02183 
02184   // We have to do this trickery because the IRBuilder might fold the check to a
02185   // constant expression in which case there is no Instruction anchored in a
02186   // the block.
02187   Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
02188                                                  ConstantInt::getTrue(Ctx));
02189   ChkBuilder.Insert(Check, "memcheck.conflict");
02190   FirstInst = getFirstInst(FirstInst, Check, Loc);
02191   return std::make_pair(FirstInst, Check);
02192 }
02193 
02194 void InnerLoopVectorizer::createEmptyLoop() {
02195   /*
02196    In this function we generate a new loop. The new loop will contain
02197    the vectorized instructions while the old loop will continue to run the
02198    scalar remainder.
02199 
02200        [ ] <-- Back-edge taken count overflow check.
02201     /   |
02202    /    v
02203   |    [ ] <-- vector loop bypass (may consist of multiple blocks).
02204   |  /  |
02205   | /   v
02206   ||   [ ]     <-- vector pre header.
02207   ||    |
02208   ||    v
02209   ||   [  ] \
02210   ||   [  ]_|   <-- vector loop.
02211   ||    |
02212   | \   v
02213   |   >[ ]   <--- middle-block.
02214   |  /  |
02215   | /   v
02216   -|- >[ ]     <--- new preheader.
02217    |    |
02218    |    v
02219    |   [ ] \
02220    |   [ ]_|   <-- old scalar loop to handle remainder.
02221     \   |
02222      \  v
02223       >[ ]     <-- exit block.
02224    ...
02225    */
02226 
02227   BasicBlock *OldBasicBlock = OrigLoop->getHeader();
02228   BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
02229   BasicBlock *ExitBlock = OrigLoop->getExitBlock();
02230   assert(BypassBlock && "Invalid loop structure");
02231   assert(ExitBlock && "Must have an exit block");
02232 
02233   // Some loops have a single integer induction variable, while other loops
02234   // don't. One example is c++ iterators that often have multiple pointer
02235   // induction variables. In the code below we also support a case where we
02236   // don't have a single induction variable.
02237   OldInduction = Legal->getInduction();
02238   Type *IdxTy = Legal->getWidestInductionType();
02239 
02240   // Find the loop boundaries.
02241   const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
02242   assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
02243 
02244   // The exit count might have the type of i64 while the phi is i32. This can
02245   // happen if we have an induction variable that is sign extended before the
02246   // compare. The only way that we get a backedge taken count is that the
02247   // induction variable was signed and as such will not overflow. In such a case
02248   // truncation is legal.
02249   if (ExitCount->getType()->getPrimitiveSizeInBits() >
02250       IdxTy->getPrimitiveSizeInBits())
02251     ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
02252 
02253   const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
02254   // Get the total trip count from the count by adding 1.
02255   ExitCount = SE->getAddExpr(BackedgeTakeCount,
02256                              SE->getConstant(BackedgeTakeCount->getType(), 1));
02257 
02258   // Expand the trip count and place the new instructions in the preheader.
02259   // Notice that the pre-header does not change, only the loop body.
02260   SCEVExpander Exp(*SE, "induction");
02261 
02262   // We need to test whether the backedge-taken count is uint##_max. Adding one
02263   // to it will cause overflow and an incorrect loop trip count in the vector
02264   // body. In case of overflow we want to directly jump to the scalar remainder
02265   // loop.
02266   Value *BackedgeCount =
02267       Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
02268                         BypassBlock->getTerminator());
02269   if (BackedgeCount->getType()->isPointerTy())
02270     BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
02271                                                 "backedge.ptrcnt.to.int",
02272                                                 BypassBlock->getTerminator());
02273   Instruction *CheckBCOverflow =
02274       CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
02275                       Constant::getAllOnesValue(BackedgeCount->getType()),
02276                       "backedge.overflow", BypassBlock->getTerminator());
02277 
02278   // The loop index does not have to start at Zero. Find the original start
02279   // value from the induction PHI node. If we don't have an induction variable
02280   // then we know that it starts at zero.
02281   Builder.SetInsertPoint(BypassBlock->getTerminator());
02282   Value *StartIdx = ExtendedIdx = OldInduction ?
02283     Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
02284                        IdxTy):
02285     ConstantInt::get(IdxTy, 0);
02286 
02287   // We need an instruction to anchor the overflow check on. StartIdx needs to
02288   // be defined before the overflow check branch. Because the scalar preheader
02289   // is going to merge the start index and so the overflow branch block needs to
02290   // contain a definition of the start index.
02291   Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
02292       StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
02293       BypassBlock->getTerminator());
02294 
02295   // Count holds the overall loop count (N).
02296   Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
02297                                    BypassBlock->getTerminator());
02298 
02299   LoopBypassBlocks.push_back(BypassBlock);
02300 
02301   // Split the single block loop into the two loop structure described above.
02302   BasicBlock *VectorPH =
02303   BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
02304   BasicBlock *VecBody =
02305   VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
02306   BasicBlock *MiddleBlock =
02307   VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
02308   BasicBlock *ScalarPH =
02309   MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
02310 
02311   // Create and register the new vector loop.
02312   Loop* Lp = new Loop();
02313   Loop *ParentLoop = OrigLoop->getParentLoop();
02314 
02315   // Insert the new loop into the loop nest and register the new basic blocks
02316   // before calling any utilities such as SCEV that require valid LoopInfo.
02317   if (ParentLoop) {
02318     ParentLoop->addChildLoop(Lp);
02319     ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
02320     ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
02321     ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
02322   } else {
02323     LI->addTopLevelLoop(Lp);
02324   }
02325   Lp->addBasicBlockToLoop(VecBody, LI->getBase());
02326 
02327   // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
02328   // inside the loop.
02329   Builder.SetInsertPoint(VecBody->getFirstNonPHI());
02330 
02331   // Generate the induction variable.
02332   setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
02333   Induction = Builder.CreatePHI(IdxTy, 2, "index");
02334   // The loop step is equal to the vectorization factor (num of SIMD elements)
02335   // times the unroll factor (num of SIMD instructions).
02336   Constant *Step = ConstantInt::get(IdxTy, VF * UF);
02337 
02338   // This is the IR builder that we use to add all of the logic for bypassing
02339   // the new vector loop.
02340   IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
02341   setDebugLocFromInst(BypassBuilder,
02342                       getDebugLocFromInstOrOperands(OldInduction));
02343 
02344   // We may need to extend the index in case there is a type mismatch.
02345   // We know that the count starts at zero and does not overflow.
02346   if (Count->getType() != IdxTy) {
02347     // The exit count can be of pointer type. Convert it to the correct
02348     // integer type.
02349     if (ExitCount->getType()->isPointerTy())
02350       Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
02351     else
02352       Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
02353   }
02354 
02355   // Add the start index to the loop count to get the new end index.
02356   Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
02357 
02358   // Now we need to generate the expression for N - (N % VF), which is
02359   // the part that the vectorized body will execute.
02360   Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
02361   Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
02362   Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
02363                                                      "end.idx.rnd.down");
02364 
02365   // Now, compare the new count to zero. If it is zero skip the vector loop and
02366   // jump to the scalar loop.
02367   Value *Cmp =
02368       BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
02369 
02370   BasicBlock *LastBypassBlock = BypassBlock;
02371 
02372   // Generate code to check that the loops trip count that we computed by adding
02373   // one to the backedge-taken count will not overflow.
02374   {
02375     auto PastOverflowCheck =
02376         std::next(BasicBlock::iterator(OverflowCheckAnchor));
02377     BasicBlock *CheckBlock =
02378       LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
02379     if (ParentLoop)
02380       ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
02381     LoopBypassBlocks.push_back(CheckBlock);
02382     Instruction *OldTerm = LastBypassBlock->getTerminator();
02383     BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
02384     OldTerm->eraseFromParent();
02385     LastBypassBlock = CheckBlock;
02386   }
02387 
02388   // Generate the code to check that the strides we assumed to be one are really
02389   // one. We want the new basic block to start at the first instruction in a
02390   // sequence of instructions that form a check.
02391   Instruction *StrideCheck;
02392   Instruction *FirstCheckInst;
02393   std::tie(FirstCheckInst, StrideCheck) =
02394       addStrideCheck(LastBypassBlock->getTerminator());
02395   if (StrideCheck) {
02396     // Create a new block containing the stride check.
02397     BasicBlock *CheckBlock =
02398         LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
02399     if (ParentLoop)
02400       ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
02401     LoopBypassBlocks.push_back(CheckBlock);
02402 
02403     // Replace the branch into the memory check block with a conditional branch
02404     // for the "few elements case".
02405     Instruction *OldTerm = LastBypassBlock->getTerminator();
02406     BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
02407     OldTerm->eraseFromParent();
02408 
02409     Cmp = StrideCheck;
02410     LastBypassBlock = CheckBlock;
02411   }
02412 
02413   // Generate the code that checks in runtime if arrays overlap. We put the
02414   // checks into a separate block to make the more common case of few elements
02415   // faster.
02416   Instruction *MemRuntimeCheck;
02417   std::tie(FirstCheckInst, MemRuntimeCheck) =
02418       addRuntimeCheck(LastBypassBlock->getTerminator());
02419   if (MemRuntimeCheck) {
02420     // Create a new block containing the memory check.
02421     BasicBlock *CheckBlock =
02422         LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
02423     if (ParentLoop)
02424       ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
02425     LoopBypassBlocks.push_back(CheckBlock);
02426 
02427     // Replace the branch into the memory check block with a conditional branch
02428     // for the "few elements case".
02429     Instruction *OldTerm = LastBypassBlock->getTerminator();
02430     BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
02431     OldTerm->eraseFromParent();
02432 
02433     Cmp = MemRuntimeCheck;
02434     LastBypassBlock = CheckBlock;
02435   }
02436 
02437   LastBypassBlock->getTerminator()->eraseFromParent();
02438   BranchInst::Create(MiddleBlock, VectorPH, Cmp,
02439                      LastBypassBlock);
02440 
02441   // We are going to resume the execution of the scalar loop.
02442   // Go over all of the induction variables that we found and fix the
02443   // PHIs that are left in the scalar version of the loop.
02444   // The starting values of PHI nodes depend on the counter of the last
02445   // iteration in the vectorized loop.
02446   // If we come from a bypass edge then we need to start from the original
02447   // start value.
02448 
02449   // This variable saves the new starting index for the scalar loop.
02450   PHINode *ResumeIndex = nullptr;
02451   LoopVectorizationLegality::InductionList::iterator I, E;
02452   LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
02453   // Set builder to point to last bypass block.
02454   BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
02455   for (I = List->begin(), E = List->end(); I != E; ++I) {
02456     PHINode *OrigPhi = I->first;
02457     LoopVectorizationLegality::InductionInfo II = I->second;
02458 
02459     Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
02460     PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
02461                                          MiddleBlock->getTerminator());
02462     // We might have extended the type of the induction variable but we need a
02463     // truncated version for the scalar loop.
02464     PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
02465       PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
02466                       MiddleBlock->getTerminator()) : nullptr;
02467 
02468     // Create phi nodes to merge from the  backedge-taken check block.
02469     PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
02470                                            ScalarPH->getTerminator());
02471     BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
02472 
02473     PHINode *BCTruncResumeVal = nullptr;
02474     if (OrigPhi == OldInduction) {
02475       BCTruncResumeVal =
02476           PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
02477                           ScalarPH->getTerminator());
02478       BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
02479     }
02480 
02481     Value *EndValue = nullptr;
02482     switch (II.IK) {
02483     case LoopVectorizationLegality::IK_NoInduction:
02484       llvm_unreachable("Unknown induction");
02485     case LoopVectorizationLegality::IK_IntInduction: {
02486       // Handle the integer induction counter.
02487       assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
02488 
02489       // We have the canonical induction variable.
02490       if (OrigPhi == OldInduction) {
02491         // Create a truncated version of the resume value for the scalar loop,
02492         // we might have promoted the type to a larger width.
02493         EndValue =
02494           BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
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           TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
02499         TruncResumeVal->addIncoming(EndValue, VecBody);
02500 
02501         BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
02502 
02503         // We know what the end value is.
02504         EndValue = IdxEndRoundDown;
02505         // We also know which PHI node holds it.
02506         ResumeIndex = ResumeVal;
02507         break;
02508       }
02509 
02510       // Not the canonical induction variable - add the vector loop count to the
02511       // start value.
02512       Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
02513                                                    II.StartValue->getType(),
02514                                                    "cast.crd");
02515       EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
02516       break;
02517     }
02518     case LoopVectorizationLegality::IK_ReverseIntInduction: {
02519       // Convert the CountRoundDown variable to the PHI size.
02520       Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
02521                                                    II.StartValue->getType(),
02522                                                    "cast.crd");
02523       // Handle reverse integer induction counter.
02524       EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
02525       break;
02526     }
02527     case LoopVectorizationLegality::IK_PtrInduction: {
02528       // For pointer induction variables, calculate the offset using
02529       // the end index.
02530       EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
02531                                          "ptr.ind.end");
02532       break;
02533     }
02534     case LoopVectorizationLegality::IK_ReversePtrInduction: {
02535       // The value at the end of the loop for the reverse pointer is calculated
02536       // by creating a GEP with a negative index starting from the start value.
02537       Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
02538       Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
02539                                               "rev.ind.end");
02540       EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
02541                                          "rev.ptr.ind.end");
02542       break;
02543     }
02544     }// end of case
02545 
02546     // The new PHI merges the original incoming value, in case of a bypass,
02547     // or the value at the end of the vectorized loop.
02548     for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
02549       if (OrigPhi == OldInduction)
02550         ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
02551       else
02552         ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
02553     }
02554     ResumeVal->addIncoming(EndValue, VecBody);
02555 
02556     // Fix the scalar body counter (PHI node).
02557     unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
02558 
02559     // The old induction's phi node in the scalar body needs the truncated
02560     // value.
02561     if (OrigPhi == OldInduction) {
02562       BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
02563       OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
02564     } else {
02565       BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
02566       OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
02567     }
02568   }
02569 
02570   // If we are generating a new induction variable then we also need to
02571   // generate the code that calculates the exit value. This value is not
02572   // simply the end of the counter because we may skip the vectorized body
02573   // in case of a runtime check.
02574   if (!OldInduction){
02575     assert(!ResumeIndex && "Unexpected resume value found");
02576     ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
02577                                   MiddleBlock->getTerminator());
02578     for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
02579       ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
02580     ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
02581   }
02582 
02583   // Make sure that we found the index where scalar loop needs to continue.
02584   assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
02585          "Invalid resume Index");
02586 
02587   // Add a check in the middle block to see if we have completed
02588   // all of the iterations in the first vector loop.
02589   // If (N - N%VF) == N, then we *don't* need to run the remainder.
02590   Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
02591                                 ResumeIndex, "cmp.n",
02592                                 MiddleBlock->getTerminator());
02593 
02594   BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
02595   // Remove the old terminator.
02596   MiddleBlock->getTerminator()->eraseFromParent();
02597 
02598   // Create i+1 and fill the PHINode.
02599   Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
02600   Induction->addIncoming(StartIdx, VectorPH);
02601   Induction->addIncoming(NextIdx, VecBody);
02602   // Create the compare.
02603   Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
02604   Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
02605 
02606   // Now we have two terminators. Remove the old one from the block.
02607   VecBody->getTerminator()->eraseFromParent();
02608 
02609   // Get ready to start creating new instructions into the vectorized body.
02610   Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
02611 
02612   // Save the state.
02613   LoopVectorPreHeader = VectorPH;
02614   LoopScalarPreHeader = ScalarPH;
02615   LoopMiddleBlock = MiddleBlock;
02616   LoopExitBlock = ExitBlock;
02617   LoopVectorBody.push_back(VecBody);
02618   LoopScalarBody = OldBasicBlock;
02619 
02620   LoopVectorizeHints Hints(Lp, true);
02621   Hints.setAlreadyVectorized();
02622 }
02623 
02624 /// This function returns the identity element (or neutral element) for
02625 /// the operation K.
02626 Constant*
02627 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
02628   switch (K) {
02629   case RK_IntegerXor:
02630   case RK_IntegerAdd:
02631   case RK_IntegerOr:
02632     // Adding, Xoring, Oring zero to a number does not change it.
02633     return ConstantInt::get(Tp, 0);
02634   case RK_IntegerMult:
02635     // Multiplying a number by 1 does not change it.
02636     return ConstantInt::get(Tp, 1);
02637   case RK_IntegerAnd:
02638     // AND-ing a number with an all-1 value does not change it.
02639     return ConstantInt::get(Tp, -1, true);
02640   case  RK_FloatMult:
02641     // Multiplying a number by 1 does not change it.
02642     return ConstantFP::get(Tp, 1.0L);
02643   case  RK_FloatAdd:
02644     // Adding zero to a number does not change it.
02645     return ConstantFP::get(Tp, 0.0L);
02646   default:
02647     llvm_unreachable("Unknown reduction kind");
02648   }
02649 }
02650 
02651 /// This function translates the reduction kind to an LLVM binary operator.
02652 static unsigned
02653 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
02654   switch (Kind) {
02655     case LoopVectorizationLegality::RK_IntegerAdd:
02656       return Instruction::Add;
02657     case LoopVectorizationLegality::RK_IntegerMult:
02658       return Instruction::Mul;
02659     case LoopVectorizationLegality::RK_IntegerOr:
02660       return Instruction::Or;
02661     case LoopVectorizationLegality::RK_IntegerAnd:
02662       return Instruction::And;
02663     case LoopVectorizationLegality::RK_IntegerXor:
02664       return Instruction::Xor;
02665     case LoopVectorizationLegality::RK_FloatMult:
02666       return Instruction::FMul;
02667     case LoopVectorizationLegality::RK_FloatAdd:
02668       return Instruction::FAdd;
02669     case LoopVectorizationLegality::RK_IntegerMinMax:
02670       return Instruction::ICmp;
02671     case LoopVectorizationLegality::RK_FloatMinMax:
02672       return Instruction::FCmp;
02673     default:
02674       llvm_unreachable("Unknown reduction operation");
02675   }
02676 }
02677 
02678 Value *createMinMaxOp(IRBuilder<> &Builder,
02679                       LoopVectorizationLegality::MinMaxReductionKind RK,
02680                       Value *Left,
02681                       Value *Right) {
02682   CmpInst::Predicate P = CmpInst::ICMP_NE;
02683   switch (RK) {
02684   default:
02685     llvm_unreachable("Unknown min/max reduction kind");
02686   case LoopVectorizationLegality::MRK_UIntMin:
02687     P = CmpInst::ICMP_ULT;
02688     break;
02689   case LoopVectorizationLegality::MRK_UIntMax:
02690     P = CmpInst::ICMP_UGT;
02691     break;
02692   case LoopVectorizationLegality::MRK_SIntMin:
02693     P = CmpInst::ICMP_SLT;
02694     break;
02695   case LoopVectorizationLegality::MRK_SIntMax:
02696     P = CmpInst::ICMP_SGT;
02697     break;
02698   case LoopVectorizationLegality::MRK_FloatMin:
02699     P = CmpInst::FCMP_OLT;
02700     break;
02701   case LoopVectorizationLegality::MRK_FloatMax:
02702     P = CmpInst::FCMP_OGT;
02703     break;
02704   }
02705 
02706   Value *Cmp;
02707   if (RK == LoopVectorizationLegality::MRK_FloatMin ||
02708       RK == LoopVectorizationLegality::MRK_FloatMax)
02709     Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
02710   else
02711     Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
02712 
02713   Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
02714   return Select;
02715 }
02716 
02717 namespace {
02718 struct CSEDenseMapInfo {
02719   static bool canHandle(Instruction *I) {
02720     return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
02721            isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
02722   }
02723   static inline Instruction *getEmptyKey() {
02724     return DenseMapInfo<Instruction *>::getEmptyKey();
02725   }
02726   static inline Instruction *getTombstoneKey() {
02727     return DenseMapInfo<Instruction *>::getTombstoneKey();
02728   }
02729   static unsigned getHashValue(Instruction *I) {
02730     assert(canHandle(I) && "Unknown instruction!");
02731     return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
02732                                                            I->value_op_end()));
02733   }
02734   static bool isEqual(Instruction *LHS, Instruction *RHS) {
02735     if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
02736         LHS == getTombstoneKey() || RHS == getTombstoneKey())
02737       return LHS == RHS;
02738     return LHS->isIdenticalTo(RHS);
02739   }
02740 };
02741 }
02742 
02743 /// \brief Check whether this block is a predicated block.
02744 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
02745 /// = ...;  " blocks. We start with one vectorized basic block. For every
02746 /// conditional block we split this vectorized block. Therefore, every second
02747 /// block will be a predicated one.
02748 static bool isPredicatedBlock(unsigned BlockNum) {
02749   return BlockNum % 2;
02750 }
02751 
02752 ///\brief Perform cse of induction variable instructions.
02753 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
02754   // Perform simple cse.
02755   SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
02756   for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
02757     BasicBlock *BB = BBs[i];
02758     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
02759       Instruction *In = I++;
02760 
02761       if (!CSEDenseMapInfo::canHandle(In))
02762         continue;
02763 
02764       // Check if we can replace this instruction with any of the
02765       // visited instructions.
02766       if (Instruction *V = CSEMap.lookup(In)) {
02767         In->replaceAllUsesWith(V);
02768         In->eraseFromParent();
02769         continue;
02770       }
02771       // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
02772       // ...;" blocks for predicated stores. Every second block is a predicated
02773       // block.
02774       if (isPredicatedBlock(i))
02775         continue;
02776 
02777       CSEMap[In] = In;
02778     }
02779   }
02780 }
02781 
02782 /// \brief Adds a 'fast' flag to floating point operations.
02783 static Value *addFastMathFlag(Value *V) {
02784   if (isa<FPMathOperator>(V)){
02785     FastMathFlags Flags;
02786     Flags.setUnsafeAlgebra();
02787     cast<Instruction>(V)->setFastMathFlags(Flags);
02788   }
02789   return V;
02790 }
02791 
02792 void InnerLoopVectorizer::vectorizeLoop() {
02793   //===------------------------------------------------===//
02794   //
02795   // Notice: any optimization or new instruction that go
02796   // into the code below should be also be implemented in
02797   // the cost-model.
02798   //
02799   //===------------------------------------------------===//
02800   Constant *Zero = Builder.getInt32(0);
02801 
02802   // In order to support reduction variables we need to be able to vectorize
02803   // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
02804   // stages. First, we create a new vector PHI node with no incoming edges.
02805   // We use this value when we vectorize all of the instructions that use the
02806   // PHI. Next, after all of the instructions in the block are complete we
02807   // add the new incoming edges to the PHI. At this point all of the
02808   // instructions in the basic block are vectorized, so we can use them to
02809   // construct the PHI.
02810   PhiVector RdxPHIsToFix;
02811 
02812   // Scan the loop in a topological order to ensure that defs are vectorized
02813   // before users.
02814   LoopBlocksDFS DFS(OrigLoop);
02815   DFS.perform(LI);
02816 
02817   // Vectorize all of the blocks in the original loop.
02818   for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
02819        be = DFS.endRPO(); bb != be; ++bb)
02820     vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
02821 
02822   // At this point every instruction in the original loop is widened to
02823   // a vector form. We are almost done. Now, we need to fix the PHI nodes
02824   // that we vectorized. The PHI nodes are currently empty because we did
02825   // not want to introduce cycles. Notice that the remaining PHI nodes
02826   // that we need to fix are reduction variables.
02827 
02828   // Create the 'reduced' values for each of the induction vars.
02829   // The reduced values are the vector values that we scalarize and combine
02830   // after the loop is finished.
02831   for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
02832        it != e; ++it) {
02833     PHINode *RdxPhi = *it;
02834     assert(RdxPhi && "Unable to recover vectorized PHI");
02835 
02836     // Find the reduction variable descriptor.
02837     assert(Legal->getReductionVars()->count(RdxPhi) &&
02838            "Unable to find the reduction variable");
02839     LoopVectorizationLegality::ReductionDescriptor RdxDesc =
02840     (*Legal->getReductionVars())[RdxPhi];
02841 
02842     setDebugLocFromInst(Builder, RdxDesc.StartValue);
02843 
02844     // We need to generate a reduction vector from the incoming scalar.
02845     // To do so, we need to generate the 'identity' vector and override
02846     // one of the elements with the incoming scalar reduction. We need
02847     // to do it in the vector-loop preheader.
02848     Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
02849 
02850     // This is the vector-clone of the value that leaves the loop.
02851     VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
02852     Type *VecTy = VectorExit[0]->getType();
02853 
02854     // Find the reduction identity variable. Zero for addition, or, xor,
02855     // one for multiplication, -1 for And.
02856     Value *Identity;
02857     Value *VectorStart;
02858     if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
02859         RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
02860       // MinMax reduction have the start value as their identify.
02861       if (VF == 1) {
02862         VectorStart = Identity = RdxDesc.StartValue;
02863       } else {
02864         VectorStart = Identity = Builder.CreateVectorSplat(VF,
02865                                                            RdxDesc.StartValue,
02866                                                            "minmax.ident");
02867       }
02868     } else {
02869       // Handle other reduction kinds:
02870       Constant *Iden =
02871       LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
02872                                                       VecTy->getScalarType());
02873       if (VF == 1) {
02874         Identity = Iden;
02875         // This vector is the Identity vector where the first element is the
02876         // incoming scalar reduction.
02877         VectorStart = RdxDesc.StartValue;
02878       } else {
02879         Identity = ConstantVector::getSplat(VF, Iden);
02880 
02881         // This vector is the Identity vector where the first element is the
02882         // incoming scalar reduction.
02883         VectorStart = Builder.CreateInsertElement(Identity,
02884                                                   RdxDesc.StartValue, Zero);
02885       }
02886     }
02887 
02888     // Fix the vector-loop phi.
02889     // We created the induction variable so we know that the
02890     // preheader is the first entry.
02891     BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
02892 
02893     // Reductions do not have to start at zero. They can start with
02894     // any loop invariant values.
02895     VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
02896     BasicBlock *Latch = OrigLoop->getLoopLatch();
02897     Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
02898     VectorParts &Val = getVectorValue(LoopVal);
02899     for (unsigned part = 0; part < UF; ++part) {
02900       // Make sure to add the reduction stat value only to the
02901       // first unroll part.
02902       Value *StartVal = (part == 0) ? VectorStart : Identity;
02903       cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
02904       cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
02905                                                   LoopVectorBody.back());
02906     }
02907 
02908     // Before each round, move the insertion point right between
02909     // the PHIs and the values we are going to write.
02910     // This allows us to write both PHINodes and the extractelement
02911     // instructions.
02912     Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
02913 
02914     VectorParts RdxParts;
02915     setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
02916     for (unsigned part = 0; part < UF; ++part) {
02917       // This PHINode contains the vectorized reduction variable, or
02918       // the initial value vector, if we bypass the vector loop.
02919       VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
02920       PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
02921       Value *StartVal = (part == 0) ? VectorStart : Identity;
02922       for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
02923         NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
02924       NewPhi->addIncoming(RdxExitVal[part],
02925                           LoopVectorBody.back());
02926       RdxParts.push_back(NewPhi);
02927     }
02928 
02929     // Reduce all of the unrolled parts into a single vector.
02930     Value *ReducedPartRdx = RdxParts[0];
02931     unsigned Op = getReductionBinOp(RdxDesc.Kind);
02932     setDebugLocFromInst(Builder, ReducedPartRdx);
02933     for (unsigned part = 1; part < UF; ++part) {
02934       if (Op != Instruction::ICmp && Op != Instruction::FCmp)
02935         // Floating point operations had to be 'fast' to enable the reduction.
02936         ReducedPartRdx = addFastMathFlag(
02937             Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
02938                                 ReducedPartRdx, "bin.rdx"));
02939       else
02940         ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
02941                                         ReducedPartRdx, RdxParts[part]);
02942     }
02943 
02944     if (VF > 1) {
02945       // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
02946       // and vector ops, reducing the set of values being computed by half each
02947       // round.
02948       assert(isPowerOf2_32(VF) &&
02949              "Reduction emission only supported for pow2 vectors!");
02950       Value *TmpVec = ReducedPartRdx;
02951       SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
02952       for (unsigned i = VF; i != 1; i >>= 1) {
02953         // Move the upper half of the vector to the lower half.
02954         for (unsigned j = 0; j != i/2; ++j)
02955           ShuffleMask[j] = Builder.getInt32(i/2 + j);
02956 
02957         // Fill the rest of the mask with undef.
02958         std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
02959                   UndefValue::get(Builder.getInt32Ty()));
02960 
02961         Value *Shuf =
02962         Builder.CreateShuffleVector(TmpVec,
02963                                     UndefValue::get(TmpVec->getType()),
02964                                     ConstantVector::get(ShuffleMask),
02965                                     "rdx.shuf");
02966 
02967         if (Op != Instruction::ICmp && Op != Instruction::FCmp)
02968           // Floating point operations had to be 'fast' to enable the reduction.
02969           TmpVec = addFastMathFlag(Builder.CreateBinOp(
02970               (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
02971         else
02972           TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
02973       }
02974 
02975       // The result is in the first element of the vector.
02976       ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
02977                                                     Builder.getInt32(0));
02978     }
02979 
02980     // Create a phi node that merges control-flow from the backedge-taken check
02981     // block and the middle block.
02982     PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
02983                                           LoopScalarPreHeader->getTerminator());
02984     BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
02985     BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
02986 
02987     // Now, we need to fix the users of the reduction variable
02988     // inside and outside of the scalar remainder loop.
02989     // We know that the loop is in LCSSA form. We need to update the
02990     // PHI nodes in the exit blocks.
02991     for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
02992          LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
02993       PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
02994       if (!LCSSAPhi) break;
02995 
02996       // All PHINodes need to have a single entry edge, or two if
02997       // we already fixed them.
02998       assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
02999 
03000       // We found our reduction value exit-PHI. Update it with the
03001       // incoming bypass edge.
03002       if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
03003         // Add an edge coming from the bypass.
03004         LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
03005         break;
03006       }
03007     }// end of the LCSSA phi scan.
03008 
03009     // Fix the scalar loop reduction variable with the incoming reduction sum
03010     // from the vector body and from the backedge value.
03011     int IncomingEdgeBlockIdx =
03012     (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
03013     assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
03014     // Pick the other block.
03015     int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
03016     (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
03017     (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
03018   }// end of for each redux variable.
03019 
03020   fixLCSSAPHIs();
03021 
03022   // Remove redundant induction instructions.
03023   cse(LoopVectorBody);
03024 }
03025 
03026 void InnerLoopVectorizer::fixLCSSAPHIs() {
03027   for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
03028        LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
03029     PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
03030     if (!LCSSAPhi) break;
03031     if (LCSSAPhi->getNumIncomingValues() == 1)
03032       LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
03033                             LoopMiddleBlock);
03034   }
03035 }
03036 
03037 InnerLoopVectorizer::VectorParts
03038 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
03039   assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
03040          "Invalid edge");
03041 
03042   // Look for cached value.
03043   std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
03044   EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
03045   if (ECEntryIt != MaskCache.end())
03046     return ECEntryIt->second;
03047 
03048   VectorParts SrcMask = createBlockInMask(Src);
03049 
03050   // The terminator has to be a branch inst!
03051   BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
03052   assert(BI && "Unexpected terminator found");
03053 
03054   if (BI->isConditional()) {
03055     VectorParts EdgeMask = getVectorValue(BI->getCondition());
03056 
03057     if (BI->getSuccessor(0) != Dst)
03058       for (unsigned part = 0; part < UF; ++part)
03059         EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
03060 
03061     for (unsigned part = 0; part < UF; ++part)
03062       EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
03063 
03064     MaskCache[Edge] = EdgeMask;
03065     return EdgeMask;
03066   }
03067 
03068   MaskCache[Edge] = SrcMask;
03069   return SrcMask;
03070 }
03071 
03072 InnerLoopVectorizer::VectorParts
03073 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
03074   assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
03075 
03076   // Loop incoming mask is all-one.
03077   if (OrigLoop->getHeader() == BB) {
03078     Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
03079     return getVectorValue(C);
03080   }
03081 
03082   // This is the block mask. We OR all incoming edges, and with zero.
03083   Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
03084   VectorParts BlockMask = getVectorValue(Zero);
03085 
03086   // For each pred:
03087   for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
03088     VectorParts EM = createEdgeMask(*it, BB);
03089     for (unsigned part = 0; part < UF; ++part)
03090       BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
03091   }
03092 
03093   return BlockMask;
03094 }
03095 
03096 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
03097                                               InnerLoopVectorizer::VectorParts &Entry,
03098                                               unsigned UF, unsigned VF, PhiVector *PV) {
03099   PHINode* P = cast<PHINode>(PN);
03100   // Handle reduction variables:
03101   if (Legal->getReductionVars()->count(P)) {
03102     for (unsigned part = 0; part < UF; ++part) {
03103       // This is phase one of vectorizing PHIs.
03104       Type *VecTy = (VF == 1) ? PN->getType() :
03105       VectorType::get(PN->getType(), VF);
03106       Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
03107                                     LoopVectorBody.back()-> getFirstInsertionPt());
03108     }
03109     PV->push_back(P);
03110     return;
03111   }
03112 
03113   setDebugLocFromInst(Builder, P);
03114   // Check for PHI nodes that are lowered to vector selects.
03115   if (P->getParent() != OrigLoop->getHeader()) {
03116     // We know that all PHIs in non-header blocks are converted into
03117     // selects, so we don't have to worry about the insertion order and we
03118     // can just use the builder.
03119     // At this point we generate the predication tree. There may be
03120     // duplications since this is a simple recursive scan, but future
03121     // optimizations will clean it up.
03122 
03123     unsigned NumIncoming = P->getNumIncomingValues();
03124 
03125     // Generate a sequence of selects of the form:
03126     // SELECT(Mask3, In3,
03127     //      SELECT(Mask2, In2,
03128     //                   ( ...)))
03129     for (unsigned In = 0; In < NumIncoming; In++) {
03130       VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
03131                                         P->getParent());
03132       VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
03133 
03134       for (unsigned part = 0; part < UF; ++part) {
03135         // We might have single edge PHIs (blocks) - use an identity
03136         // 'select' for the first PHI operand.
03137         if (In == 0)
03138           Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
03139                                              In0[part]);
03140         else
03141           // Select between the current value and the previous incoming edge
03142           // based on the incoming mask.
03143           Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
03144                                              Entry[part], "predphi");
03145       }
03146     }
03147     return;
03148   }
03149 
03150   // This PHINode must be an induction variable.
03151   // Make sure that we know about it.
03152   assert(Legal->getInductionVars()->count(P) &&
03153          "Not an induction variable");
03154 
03155   LoopVectorizationLegality::InductionInfo II =
03156   Legal->getInductionVars()->lookup(P);
03157 
03158   switch (II.IK) {
03159     case LoopVectorizationLegality::IK_NoInduction:
03160       llvm_unreachable("Unknown induction");
03161     case LoopVectorizationLegality::IK_IntInduction: {
03162       assert(P->getType() == II.StartValue->getType() && "Types must match");
03163       Type *PhiTy = P->getType();
03164       Value *Broadcasted;
03165       if (P == OldInduction) {
03166         // Handle the canonical induction variable. We might have had to
03167         // extend the type.
03168         Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
03169       } else {
03170         // Handle other induction variables that are now based on the
03171         // canonical one.
03172         Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
03173                                                  "normalized.idx");
03174         NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
03175         Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
03176                                         "offset.idx");
03177       }
03178       Broadcasted = getBroadcastInstrs(Broadcasted);
03179       // After broadcasting the induction variable we need to make the vector
03180       // consecutive by adding 0, 1, 2, etc.
03181       for (unsigned part = 0; part < UF; ++part)
03182         Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
03183       return;
03184     }
03185     case LoopVectorizationLegality::IK_ReverseIntInduction:
03186     case LoopVectorizationLegality::IK_PtrInduction:
03187     case LoopVectorizationLegality::IK_ReversePtrInduction:
03188       // Handle reverse integer and pointer inductions.
03189       Value *StartIdx = ExtendedIdx;
03190       // This is the normalized GEP that starts counting at zero.
03191       Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
03192                                                "normalized.idx");
03193 
03194       // Handle the reverse integer induction variable case.
03195       if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
03196         IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
03197         Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
03198                                                "resize.norm.idx");
03199         Value *ReverseInd  = Builder.CreateSub(II.StartValue, CNI,
03200                                                "reverse.idx");
03201 
03202         // This is a new value so do not hoist it out.
03203         Value *Broadcasted = getBroadcastInstrs(ReverseInd);
03204         // After broadcasting the induction variable we need to make the
03205         // vector consecutive by adding  ... -3, -2, -1, 0.
03206         for (unsigned part = 0; part < UF; ++part)
03207           Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
03208                                              true);
03209         return;
03210       }
03211 
03212       // Handle the pointer induction variable case.
03213       assert(P->getType()->isPointerTy() && "Unexpected type.");
03214 
03215       // Is this a reverse induction ptr or a consecutive induction ptr.
03216       bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
03217                       II.IK);
03218 
03219       // This is the vector of results. Notice that we don't generate
03220       // vector geps because scalar geps result in better code.
03221       for (unsigned part = 0; part < UF; ++part) {
03222         if (VF == 1) {
03223           int EltIndex = (part) * (Reverse ? -1 : 1);
03224           Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
03225           Value *GlobalIdx;
03226           if (Reverse)
03227             GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
03228           else
03229             GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
03230 
03231           Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
03232                                              "next.gep");
03233           Entry[part] = SclrGep;
03234           continue;
03235         }
03236 
03237         Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
03238         for (unsigned int i = 0; i < VF; ++i) {
03239           int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
03240           Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
03241           Value *GlobalIdx;
03242           if (!Reverse)
03243             GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
03244           else
03245             GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
03246 
03247           Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
03248                                              "next.gep");
03249           VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
03250                                                Builder.getInt32(i),
03251                                                "insert.gep");
03252         }
03253         Entry[part] = VecVal;
03254       }
03255       return;
03256   }
03257 }
03258 
03259 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
03260   // For each instruction in the old loop.
03261   for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
03262     VectorParts &Entry = WidenMap.get(it);
03263     switch (it->getOpcode()) {
03264     case Instruction::Br:
03265       // Nothing to do for PHIs and BR, since we already took care of the
03266       // loop control flow instructions.
03267       continue;
03268     case Instruction::PHI:{
03269       // Vectorize PHINodes.
03270       widenPHIInstruction(it, Entry, UF, VF, PV);
03271       continue;
03272     }// End of PHI.
03273 
03274     case Instruction::Add:
03275     case Instruction::FAdd:
03276     case Instruction::Sub:
03277     case Instruction::FSub:
03278     case Instruction::Mul:
03279     case Instruction::FMul:
03280     case Instruction::UDiv:
03281     case Instruction::SDiv:
03282     case Instruction::FDiv:
03283     case Instruction::URem:
03284     case Instruction::SRem:
03285     case Instruction::FRem:
03286     case Instruction::Shl:
03287     case Instruction::LShr:
03288     case Instruction::AShr:
03289     case Instruction::And:
03290     case Instruction::Or:
03291     case Instruction::Xor: {
03292       // Just widen binops.
03293       BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
03294       setDebugLocFromInst(Builder, BinOp);
03295       VectorParts &A = getVectorValue(it->getOperand(0));
03296       VectorParts &B = getVectorValue(it->getOperand(1));
03297 
03298       // Use this vector value for all users of the original instruction.
03299       for (unsigned Part = 0; Part < UF; ++Part) {
03300         Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
03301 
03302         if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
03303           VecOp->copyIRFlags(BinOp);
03304 
03305         Entry[Part] = V;
03306       }
03307 
03308       propagateMetadata(Entry, it);
03309       break;
03310     }
03311     case Instruction::Select: {
03312       // Widen selects.
03313       // If the selector is loop invariant we can create a select
03314       // instruction with a scalar condition. Otherwise, use vector-select.
03315       bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
03316                                                OrigLoop);
03317       setDebugLocFromInst(Builder, it);
03318 
03319       // The condition can be loop invariant  but still defined inside the
03320       // loop. This means that we can't just use the original 'cond' value.
03321       // We have to take the 'vectorized' value and pick the first lane.
03322       // Instcombine will make this a no-op.
03323       VectorParts &Cond = getVectorValue(it->getOperand(0));
03324       VectorParts &Op0  = getVectorValue(it->getOperand(1));
03325       VectorParts &Op1  = getVectorValue(it->getOperand(2));
03326 
03327       Value *ScalarCond = (VF == 1) ? Cond[0] :
03328         Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
03329 
03330       for (unsigned Part = 0; Part < UF; ++Part) {
03331         Entry[Part] = Builder.CreateSelect(
03332           InvariantCond ? ScalarCond : Cond[Part],
03333           Op0[Part],
03334           Op1[Part]);
03335       }
03336 
03337       propagateMetadata(Entry, it);
03338       break;
03339     }
03340 
03341     case Instruction::ICmp:
03342     case Instruction::FCmp: {
03343       // Widen compares. Generate vector compares.
03344       bool FCmp = (it->getOpcode() == Instruction::FCmp);
03345       CmpInst *Cmp = dyn_cast<CmpInst>(it);
03346       setDebugLocFromInst(Builder, it);
03347       VectorParts &A = getVectorValue(it->getOperand(0));
03348       VectorParts &B = getVectorValue(it->getOperand(1));
03349       for (unsigned Part = 0; Part < UF; ++Part) {
03350         Value *C = nullptr;
03351         if (FCmp)
03352           C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
03353         else
03354           C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
03355         Entry[Part] = C;
03356       }
03357 
03358       propagateMetadata(Entry, it);
03359       break;
03360     }
03361 
03362     case Instruction::Store:
03363     case Instruction::Load:
03364       vectorizeMemoryInstruction(it);
03365         break;
03366     case Instruction::ZExt:
03367     case Instruction::SExt:
03368     case Instruction::FPToUI:
03369     case Instruction::FPToSI:
03370     case Instruction::FPExt:
03371     case Instruction::PtrToInt:
03372     case Instruction::IntToPtr:
03373     case Instruction::SIToFP:
03374     case Instruction::UIToFP:
03375     case Instruction::Trunc:
03376     case Instruction::FPTrunc:
03377     case Instruction::BitCast: {
03378       CastInst *CI = dyn_cast<CastInst>(it);
03379       setDebugLocFromInst(Builder, it);
03380       /// Optimize the special case where the source is the induction
03381       /// variable. Notice that we can only optimize the 'trunc' case
03382       /// because: a. FP conversions lose precision, b. sext/zext may wrap,
03383       /// c. other casts depend on pointer size.
03384       if (CI->getOperand(0) == OldInduction &&
03385           it->getOpcode() == Instruction::Trunc) {
03386         Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
03387                                                CI->getType());
03388         Value *Broadcasted = getBroadcastInstrs(ScalarCast);
03389         for (unsigned Part = 0; Part < UF; ++Part)
03390           Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
03391         propagateMetadata(Entry, it);
03392         break;
03393       }
03394       /// Vectorize casts.
03395       Type *DestTy = (VF == 1) ? CI->getType() :
03396                                  VectorType::get(CI->getType(), VF);
03397 
03398       VectorParts &A = getVectorValue(it->getOperand(0));
03399       for (unsigned Part = 0; Part < UF; ++Part)
03400         Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
03401       propagateMetadata(Entry, it);
03402       break;
03403     }
03404 
03405     case Instruction::Call: {
03406       // Ignore dbg intrinsics.
03407       if (isa<DbgInfoIntrinsic>(it))
03408         break;
03409       setDebugLocFromInst(Builder, it);
03410 
03411       Module *M = BB->getParent()->getParent();
03412       CallInst *CI = cast<CallInst>(it);
03413       Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
03414       assert(ID && "Not an intrinsic call!");
03415       switch (ID) {
03416       case Intrinsic::assume:
03417       case Intrinsic::lifetime_end:
03418       case Intrinsic::lifetime_start:
03419         scalarizeInstruction(it);
03420         break;
03421       default:
03422         bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
03423         for (unsigned Part = 0; Part < UF; ++Part) {
03424           SmallVector<Value *, 4> Args;
03425           for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
03426             if (HasScalarOpd && i == 1) {
03427               Args.push_back(CI->getArgOperand(i));
03428               continue;
03429             }
03430             VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
03431             Args.push_back(Arg[Part]);
03432           }
03433           Type *Tys[] = {CI->getType()};
03434           if (VF > 1)
03435             Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
03436 
03437           Function *F = Intrinsic::getDeclaration(M, ID, Tys);
03438           Entry[Part] = Builder.CreateCall(F, Args);
03439         }
03440 
03441         propagateMetadata(Entry, it);
03442         break;
03443       }
03444       break;
03445     }
03446 
03447     default:
03448       // All other instructions are unsupported. Scalarize them.
03449       scalarizeInstruction(it);
03450       break;
03451     }// end of switch.
03452   }// end of for_each instr.
03453 }
03454 
03455 void InnerLoopVectorizer::updateAnalysis() {
03456   // Forget the original basic block.
03457   SE->forgetLoop(OrigLoop);
03458 
03459   // Update the dominator tree information.
03460   assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
03461          "Entry does not dominate exit.");
03462 
03463   for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
03464     DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
03465   DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
03466 
03467   // Due to if predication of stores we might create a sequence of "if(pred)
03468   // a[i] = ...;  " blocks.
03469   for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
03470     if (i == 0)
03471       DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
03472     else if (isPredicatedBlock(i)) {
03473       DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
03474     } else {
03475       DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
03476     }
03477   }
03478 
03479   DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
03480   DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
03481   DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
03482   DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
03483 
03484   DEBUG(DT->verifyDomTree());
03485 }
03486 
03487 /// \brief Check whether it is safe to if-convert this phi node.
03488 ///
03489 /// Phi nodes with constant expressions that can trap are not safe to if
03490 /// convert.
03491 static bool canIfConvertPHINodes(BasicBlock *BB) {
03492   for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
03493     PHINode *Phi = dyn_cast<PHINode>(I);
03494     if (!Phi)
03495       return true;
03496     for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
03497       if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
03498         if (C->canTrap())
03499           return false;
03500   }
03501   return true;
03502 }
03503 
03504 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
03505   if (!EnableIfConversion) {
03506     emitAnalysis(Report() << "if-conversion is disabled");
03507     return false;
03508   }
03509 
03510   assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
03511 
03512   // A list of pointers that we can safely read and write to.
03513   SmallPtrSet<Value *, 8> SafePointes;
03514 
03515   // Collect safe addresses.
03516   for (Loop::block_iterator BI = TheLoop->block_begin(),
03517          BE = TheLoop->block_end(); BI != BE; ++BI) {
03518     BasicBlock *BB = *BI;
03519 
03520     if (blockNeedsPredication(BB))
03521       continue;
03522 
03523     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
03524       if (LoadInst *LI = dyn_cast<LoadInst>(I))
03525         SafePointes.insert(LI->getPointerOperand());
03526       else if (StoreInst *SI = dyn_cast<StoreInst>(I))
03527         SafePointes.insert(SI->getPointerOperand());
03528     }
03529   }
03530 
03531   // Collect the blocks that need predication.
03532   BasicBlock *Header = TheLoop->getHeader();
03533   for (Loop::block_iterator BI = TheLoop->block_begin(),
03534          BE = TheLoop->block_end(); BI != BE; ++BI) {
03535     BasicBlock *BB = *BI;
03536 
03537     // We don't support switch statements inside loops.
03538     if (!isa<BranchInst>(BB->getTerminator())) {
03539       emitAnalysis(Report(BB->getTerminator())
03540                    << "loop contains a switch statement");
03541       return false;
03542     }
03543 
03544     // We must be able to predicate all blocks that need to be predicated.
03545     if (blockNeedsPredication(BB)) {
03546       if (!blockCanBePredicated(BB, SafePointes)) {
03547         emitAnalysis(Report(BB->getTerminator())
03548                      << "control flow cannot be substituted for a select");
03549         return false;
03550       }
03551     } else if (BB != Header && !canIfConvertPHINodes(BB)) {
03552       emitAnalysis(Report(BB->getTerminator())
03553                    << "control flow cannot be substituted for a select");
03554       return false;
03555     }
03556   }
03557 
03558   // We can if-convert this loop.
03559   return true;
03560 }
03561 
03562 bool LoopVectorizationLegality::canVectorize() {
03563   // We must have a loop in canonical form. Loops with indirectbr in them cannot
03564   // be canonicalized.
03565   if (!TheLoop->getLoopPreheader()) {
03566     emitAnalysis(
03567         Report() << "loop control flow is not understood by vectorizer");
03568     return false;
03569   }
03570 
03571   // We can only vectorize innermost loops.
03572   if (TheLoop->getSubLoopsVector().size()) {
03573     emitAnalysis(Report() << "loop is not the innermost loop");
03574     return false;
03575   }
03576 
03577   // We must have a single backedge.
03578   if (TheLoop->getNumBackEdges() != 1) {
03579     emitAnalysis(
03580         Report() << "loop control flow is not understood by vectorizer");
03581     return false;
03582   }
03583 
03584   // We must have a single exiting block.
03585   if (!TheLoop->getExitingBlock()) {
03586     emitAnalysis(
03587         Report() << "loop control flow is not understood by vectorizer");
03588     return false;
03589   }
03590 
03591   // We need to have a loop header.
03592   DEBUG(dbgs() << "LV: Found a loop: " <<
03593         TheLoop->getHeader()->getName() << '\n');
03594 
03595   // Check if we can if-convert non-single-bb loops.
03596   unsigned NumBlocks = TheLoop->getNumBlocks();
03597   if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
03598     DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
03599     return false;
03600   }
03601 
03602   // ScalarEvolution needs to be able to find the exit count.
03603   const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
03604   if (ExitCount == SE->getCouldNotCompute()) {
03605     emitAnalysis(Report() << "could not determine number of loop iterations");
03606     DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
03607     return false;
03608   }
03609 
03610   // Check if we can vectorize the instructions and CFG in this loop.
03611   if (!canVectorizeInstrs()) {
03612     DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
03613     return false;
03614   }
03615 
03616   // Go over each instruction and look at memory deps.
03617   if (!canVectorizeMemory()) {
03618     DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
03619     return false;
03620   }
03621 
03622   // Collect all of the variables that remain uniform after vectorization.
03623   collectLoopUniforms();
03624 
03625   DEBUG(dbgs() << "LV: We can vectorize this loop" <<
03626         (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
03627         <<"!\n");
03628 
03629   // Okay! We can vectorize. At this point we don't have any other mem analysis
03630   // which may limit our maximum vectorization factor, so just return true with
03631   // no restrictions.
03632   return true;
03633 }
03634 
03635 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
03636   if (Ty->isPointerTy())
03637     return DL.getIntPtrType(Ty);
03638 
03639   // It is possible that char's or short's overflow when we ask for the loop's
03640   // trip count, work around this by changing the type size.
03641   if (Ty->getScalarSizeInBits() < 32)
03642     return Type::getInt32Ty(Ty->getContext());
03643 
03644   return Ty;
03645 }
03646 
03647 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
03648   Ty0 = convertPointerToIntegerType(DL, Ty0);
03649   Ty1 = convertPointerToIntegerType(DL, Ty1);
03650   if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
03651     return Ty0;
03652   return Ty1;
03653 }
03654 
03655 /// \brief Check that the instruction has outside loop users and is not an
03656 /// identified reduction variable.
03657 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
03658                                SmallPtrSetImpl<Value *> &Reductions) {
03659   // Reduction instructions are allowed to have exit users. All other
03660   // instructions must not have external users.
03661   if (!Reductions.count(Inst))
03662     //Check that all of the users of the loop are inside the BB.
03663     for (User *U : Inst->users()) {
03664       Instruction *UI = cast<Instruction>(U);
03665       // This user may be a reduction exit value.
03666       if (!TheLoop->contains(UI)) {
03667         DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
03668         return true;
03669       }
03670     }
03671   return false;
03672 }
03673 
03674 bool LoopVectorizationLegality::canVectorizeInstrs() {
03675   BasicBlock *PreHeader = TheLoop->getLoopPreheader();
03676   BasicBlock *Header = TheLoop->getHeader();
03677 
03678   // Look for the attribute signaling the absence of NaNs.
03679   Function &F = *Header->getParent();
03680   if (F.hasFnAttribute("no-nans-fp-math"))
03681     HasFunNoNaNAttr = F.getAttributes().getAttribute(
03682       AttributeSet::FunctionIndex,
03683       "no-nans-fp-math").getValueAsString() == "true";
03684 
03685   // For each block in the loop.
03686   for (Loop::block_iterator bb = TheLoop->block_begin(),
03687        be = TheLoop->block_end(); bb != be; ++bb) {
03688 
03689     // Scan the instructions in the block and look for hazards.
03690     for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
03691          ++it) {
03692 
03693       if (PHINode *Phi = dyn_cast<PHINode>(it)) {
03694         Type *PhiTy = Phi->getType();
03695         // Check that this PHI type is allowed.
03696         if (!PhiTy->isIntegerTy() &&
03697             !PhiTy->isFloatingPointTy() &&
03698             !PhiTy->isPointerTy()) {
03699           emitAnalysis(Report(it)
03700                        << "loop control flow is not understood by vectorizer");
03701           DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
03702           return false;
03703         }
03704 
03705         // If this PHINode is not in the header block, then we know that we
03706         // can convert it to select during if-conversion. No need to check if
03707         // the PHIs in this block are induction or reduction variables.
03708         if (*bb != Header) {
03709           // Check that this instruction has no outside users or is an
03710           // identified reduction value with an outside user.
03711           if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
03712             continue;
03713           emitAnalysis(Report(it) << "value could not be identified as "
03714                                      "an induction or reduction variable");
03715           return false;
03716         }
03717 
03718         // We only allow if-converted PHIs with more than two incoming values.
03719         if (Phi->getNumIncomingValues() != 2) {
03720           emitAnalysis(Report(it)
03721                        << "control flow not understood by vectorizer");
03722           DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
03723           return false;
03724         }
03725 
03726         // This is the value coming from the preheader.
03727         Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
03728         // Check if this is an induction variable.
03729         InductionKind IK = isInductionVariable(Phi);
03730 
03731         if (IK_NoInduction != IK) {
03732           // Get the widest type.
03733           if (!WidestIndTy)
03734             WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
03735           else
03736             WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
03737 
03738           // Int inductions are special because we only allow one IV.
03739           if (IK == IK_IntInduction) {
03740             // Use the phi node with the widest type as induction. Use the last
03741             // one if there are multiple (no good reason for doing this other
03742             // than it is expedient).
03743             if (!Induction || PhiTy == WidestIndTy)
03744               Induction = Phi;
03745           }
03746 
03747           DEBUG(dbgs() << "LV: Found an induction variable.\n");
03748           Inductions[Phi] = InductionInfo(StartValue, IK);
03749 
03750           // Until we explicitly handle the case of an induction variable with
03751           // an outside loop user we have to give up vectorizing this loop.
03752           if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
03753             emitAnalysis(Report(it) << "use of induction value outside of the "
03754                                        "loop is not handled by vectorizer");
03755             return false;
03756           }
03757 
03758           continue;
03759         }
03760 
03761         if (AddReductionVar(Phi, RK_IntegerAdd)) {
03762           DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
03763           continue;
03764         }
03765         if (AddReductionVar(Phi, RK_IntegerMult)) {
03766           DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
03767           continue;
03768         }
03769         if (AddReductionVar(Phi, RK_IntegerOr)) {
03770           DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
03771           continue;
03772         }
03773         if (AddReductionVar(Phi, RK_IntegerAnd)) {
03774           DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
03775           continue;
03776         }
03777         if (AddReductionVar(Phi, RK_IntegerXor)) {
03778           DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
03779           continue;
03780         }
03781         if (AddReductionVar(Phi, RK_IntegerMinMax)) {
03782           DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
03783           continue;
03784         }
03785         if (AddReductionVar(Phi, RK_FloatMult)) {
03786           DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
03787           continue;
03788         }
03789         if (AddReductionVar(Phi, RK_FloatAdd)) {
03790           DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
03791           continue;
03792         }
03793         if (AddReductionVar(Phi, RK_FloatMinMax)) {
03794           DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
03795                 "\n");
03796           continue;
03797         }
03798 
03799         emitAnalysis(Report(it) << "value that could not be identified as "
03800                                    "reduction is used outside the loop");
03801         DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
03802         return false;
03803       }// end of PHI handling
03804 
03805       // We still don't handle functions. However, we can ignore dbg intrinsic
03806       // calls and we do handle certain intrinsic and libm functions.
03807       CallInst *CI = dyn_cast<CallInst>(it);
03808       if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
03809         emitAnalysis(Report(it) << "call instruction cannot be vectorized");
03810         DEBUG(dbgs() << "LV: Found a call site.\n");
03811         return false;
03812       }
03813 
03814       // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
03815       // second argument is the same (i.e. loop invariant)
03816       if (CI &&
03817           hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
03818         if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
03819           emitAnalysis(Report(it)
03820                        << "intrinsic instruction cannot be vectorized");
03821           DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
03822           return false;
03823         }
03824       }
03825 
03826       // Check that the instruction return type is vectorizable.
03827       // Also, we can't vectorize extractelement instructions.
03828       if ((!VectorType::isValidElementType(it->getType()) &&
03829            !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
03830         emitAnalysis(Report(it)
03831                      << "instruction return type cannot be vectorized");
03832         DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
03833         return false;
03834       }
03835 
03836       // Check that the stored type is vectorizable.
03837       if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
03838         Type *T = ST->getValueOperand()->getType();
03839         if (!VectorType::isValidElementType(T)) {
03840           emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
03841           return false;
03842         }
03843         if (EnableMemAccessVersioning)
03844           collectStridedAcccess(ST);
03845       }
03846 
03847       if (EnableMemAccessVersioning)
03848         if (LoadInst *LI = dyn_cast<LoadInst>(it))
03849           collectStridedAcccess(LI);
03850 
03851       // Reduction instructions are allowed to have exit users.
03852       // All other instructions must not have external users.
03853       if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
03854         emitAnalysis(Report(it) << "value cannot be used outside the loop");
03855         return false;
03856       }
03857 
03858     } // next instr.
03859 
03860   }
03861 
03862   if (!Induction) {
03863     DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
03864     if (Inductions.empty()) {
03865       emitAnalysis(Report()
03866                    << "loop induction variable could not be identified");
03867       return false;
03868     }
03869   }
03870 
03871   return true;
03872 }
03873 
03874 ///\brief Remove GEPs whose indices but the last one are loop invariant and
03875 /// return the induction operand of the gep pointer.
03876 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
03877                                  const DataLayout *DL, Loop *Lp) {
03878   GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
03879   if (!GEP)
03880     return Ptr;
03881 
03882   unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
03883 
03884   // Check that all of the gep indices are uniform except for our induction
03885   // operand.
03886   for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
03887     if (i != InductionOperand &&
03888         !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
03889       return Ptr;
03890   return GEP->getOperand(InductionOperand);
03891 }
03892 
03893 ///\brief Look for a cast use of the passed value.
03894 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
03895   Value *UniqueCast = nullptr;
03896   for (User *U : Ptr->users()) {
03897     CastInst *CI = dyn_cast<CastInst>(U);
03898     if (CI && CI->getType() == Ty) {
03899       if (!UniqueCast)
03900         UniqueCast = CI;
03901       else
03902         return nullptr;
03903     }
03904   }
03905   return UniqueCast;
03906 }
03907 
03908 ///\brief Get the stride of a pointer access in a loop.
03909 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
03910 /// pointer to the Value, or null otherwise.
03911 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
03912                                    const DataLayout *DL, Loop *Lp) {
03913   const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
03914   if (!PtrTy || PtrTy->isAggregateType())
03915     return nullptr;
03916 
03917   // Try to remove a gep instruction to make the pointer (actually index at this
03918   // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
03919   // pointer, otherwise, we are analyzing the index.
03920   Value *OrigPtr = Ptr;
03921 
03922   // The size of the pointer access.
03923   int64_t PtrAccessSize = 1;
03924 
03925   Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
03926   const SCEV *V = SE->getSCEV(Ptr);
03927 
03928   if (Ptr != OrigPtr)
03929     // Strip off casts.
03930     while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
03931       V = C->getOperand();
03932 
03933   const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
03934   if (!S)
03935     return nullptr;
03936 
03937   V = S->getStepRecurrence(*SE);
03938   if (!V)
03939     return nullptr;
03940 
03941   // Strip off the size of access multiplication if we are still analyzing the
03942   // pointer.
03943   if (OrigPtr == Ptr) {
03944     DL->getTypeAllocSize(PtrTy->getElementType());
03945     if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
03946       if (M->getOperand(0)->getSCEVType() != scConstant)
03947         return nullptr;
03948 
03949       const APInt &APStepVal =
03950           cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
03951 
03952       // Huge step value - give up.
03953       if (APStepVal.getBitWidth() > 64)
03954         return nullptr;
03955 
03956       int64_t StepVal = APStepVal.getSExtValue();
03957       if (PtrAccessSize != StepVal)
03958         return nullptr;
03959       V = M->getOperand(1);
03960     }
03961   }
03962 
03963   // Strip off casts.
03964   Type *StripedOffRecurrenceCast = nullptr;
03965   if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
03966     StripedOffRecurrenceCast = C->getType();
03967     V = C->getOperand();
03968   }
03969 
03970   // Look for the loop invariant symbolic value.
03971   const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
03972   if (!U)
03973     return nullptr;
03974 
03975   Value *Stride = U->getValue();
03976   if (!Lp->isLoopInvariant(Stride))
03977     return nullptr;
03978 
03979   // If we have stripped off the recurrence cast we have to make sure that we
03980   // return the value that is used in this loop so that we can replace it later.
03981   if (StripedOffRecurrenceCast)
03982     Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
03983 
03984   return Stride;
03985 }
03986 
03987 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
03988   Value *Ptr = nullptr;
03989   if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
03990     Ptr = LI->getPointerOperand();
03991   else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
03992     Ptr = SI->getPointerOperand();
03993   else
03994     return;
03995 
03996   Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
03997   if (!Stride)
03998     return;
03999 
04000   DEBUG(dbgs() << "LV: Found a strided access that we can version");
04001   DEBUG(dbgs() << "  Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
04002   Strides[Ptr] = Stride;
04003   StrideSet.insert(Stride);
04004 }
04005 
04006 void LoopVectorizationLegality::collectLoopUniforms() {
04007   // We now know that the loop is vectorizable!
04008   // Collect variables that will remain uniform after vectorization.
04009   std::vector<Value*> Worklist;
04010   BasicBlock *Latch = TheLoop->getLoopLatch();
04011 
04012   // Start with the conditional branch and walk up the block.
04013   Worklist.push_back(Latch->getTerminator()->getOperand(0));
04014 
04015   // Also add all consecutive pointer values; these values will be uniform
04016   // after vectorization (and subsequent cleanup) and, until revectorization is
04017   // supported, all dependencies must also be uniform.
04018   for (Loop::block_iterator B = TheLoop->block_begin(),
04019        BE = TheLoop->block_end(); B != BE; ++B)
04020     for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
04021          I != IE; ++I)
04022       if (I->getType()->isPointerTy() && isConsecutivePtr(I))
04023         Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
04024 
04025   while (Worklist.size()) {
04026     Instruction *I = dyn_cast<Instruction>(Worklist.back());
04027     Worklist.pop_back();
04028 
04029     // Look at instructions inside this loop.
04030     // Stop when reaching PHI nodes.
04031     // TODO: we need to follow values all over the loop, not only in this block.
04032     if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
04033       continue;
04034 
04035     // This is a known uniform.
04036     Uniforms.insert(I);
04037 
04038     // Insert all operands.
04039     Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
04040   }
04041 }
04042 
04043 namespace {
04044 /// \brief Analyses memory accesses in a loop.
04045 ///
04046 /// Checks whether run time pointer checks are needed and builds sets for data
04047 /// dependence checking.
04048 class AccessAnalysis {
04049 public:
04050   /// \brief Read or write access location.
04051   typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
04052   typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
04053 
04054   /// \brief Set of potential dependent memory accesses.
04055   typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
04056 
04057   AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
04058     DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
04059 
04060   /// \brief Register a load  and whether it is only read from.
04061   void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
04062     Value *Ptr = const_cast<Value*>(Loc.Ptr);
04063     AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
04064     Accesses.insert(MemAccessInfo(Ptr, false));
04065     if (IsReadOnly)
04066       ReadOnlyPtr.insert(Ptr);
04067   }
04068 
04069   /// \brief Register a store.
04070   void addStore(AliasAnalysis::Location &Loc) {
04071     Value *Ptr = const_cast<Value*>(Loc.Ptr);
04072     AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
04073     Accesses.insert(MemAccessInfo(Ptr, true));
04074   }
04075 
04076   /// \brief Check whether we can check the pointers at runtime for
04077   /// non-intersection.
04078   bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
04079                        unsigned &NumComparisons, ScalarEvolution *SE,
04080                        Loop *TheLoop, ValueToValueMap &Strides,
04081                        bool ShouldCheckStride = false);
04082 
04083   /// \brief Goes over all memory accesses, checks whether a RT check is needed
04084   /// and builds sets of dependent accesses.
04085   void buildDependenceSets() {
04086     processMemAccesses();
04087   }
04088 
04089   bool isRTCheckNeeded() { return IsRTCheckNeeded; }
04090 
04091   bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
04092   void resetDepChecks() { CheckDeps.clear(); }
04093 
04094   MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
04095 
04096 private:
04097   typedef SetVector<MemAccessInfo> PtrAccessSet;
04098 
04099   /// \brief Go over all memory access and check whether runtime pointer checks
04100   /// are needed /// and build sets of dependency check candidates.
04101   void processMemAccesses();
04102 
04103   /// Set of all accesses.
04104   PtrAccessSet Accesses;
04105 
04106   /// Set of accesses that need a further dependence check.
04107   MemAccessInfoSet CheckDeps;
04108 
04109   /// Set of pointers that are read only.
04110   SmallPtrSet<Value*, 16> ReadOnlyPtr;
04111 
04112   const DataLayout *DL;
04113 
04114   /// An alias set tracker to partition the access set by underlying object and
04115   //intrinsic property (such as TBAA metadata).
04116   AliasSetTracker AST;
04117 
04118   /// Sets of potentially dependent accesses - members of one set share an
04119   /// underlying pointer. The set "CheckDeps" identfies which sets really need a
04120   /// dependence check.
04121   DepCandidates &DepCands;
04122 
04123   bool IsRTCheckNeeded;
04124 };
04125 
04126 } // end anonymous namespace
04127 
04128 /// \brief Check whether a pointer can participate in a runtime bounds check.
04129 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
04130                                 Value *Ptr) {
04131   const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
04132   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
04133   if (!AR)
04134     return false;
04135 
04136   return AR->isAffine();
04137 }
04138 
04139 /// \brief Check the stride of the pointer and ensure that it does not wrap in
04140 /// the address space.
04141 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
04142                         const Loop *Lp, ValueToValueMap &StridesMap);
04143 
04144 bool AccessAnalysis::canCheckPtrAtRT(
04145     LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
04146     unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
04147     ValueToValueMap &StridesMap, bool ShouldCheckStride) {
04148   // Find pointers with computable bounds. We are going to use this information
04149   // to place a runtime bound check.
04150   bool CanDoRT = true;
04151 
04152   bool IsDepCheckNeeded = isDependencyCheckNeeded();
04153   NumComparisons = 0;
04154 
04155   // We assign a consecutive id to access from different alias sets.
04156   // Accesses between different groups doesn't need to be checked.
04157   unsigned ASId = 1;
04158   for (auto &AS : AST) {
04159     unsigned NumReadPtrChecks = 0;
04160     unsigned NumWritePtrChecks = 0;
04161 
04162     // We assign consecutive id to access from different dependence sets.
04163     // Accesses within the same set don't need a runtime check.
04164     unsigned RunningDepId = 1;
04165     DenseMap<Value *, unsigned> DepSetId;
04166 
04167     for (auto A : AS) {
04168       Value *Ptr = A.getValue();
04169       bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
04170       MemAccessInfo Access(Ptr, IsWrite);
04171 
04172       if (IsWrite)
04173         ++NumWritePtrChecks;
04174       else
04175         ++NumReadPtrChecks;
04176 
04177       if (hasComputableBounds(SE, StridesMap, Ptr) &&
04178           // When we run after a failing dependency check we have to make sure we
04179           // don't have wrapping pointers.
04180           (!ShouldCheckStride ||
04181            isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
04182         // The id of the dependence set.
04183         unsigned DepId;
04184 
04185         if (IsDepCheckNeeded) {
04186           Value *Leader = DepCands.getLeaderValue(Access).getPointer();
04187           unsigned &LeaderId = DepSetId[Leader];
04188           if (!LeaderId)
04189             LeaderId = RunningDepId++;
04190           DepId = LeaderId;
04191         } else
04192           // Each access has its own dependence set.
04193           DepId = RunningDepId++;
04194 
04195         RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
04196 
04197         DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
04198       } else {
04199         CanDoRT = false;
04200       }
04201     }
04202 
04203     if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
04204       NumComparisons += 0; // Only one dependence set.
04205     else {
04206       NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
04207                                               NumWritePtrChecks - 1));
04208     }
04209 
04210     ++ASId;
04211   }
04212 
04213   // If the pointers that we would use for the bounds comparison have different
04214   // address spaces, assume the values aren't directly comparable, so we can't
04215   // use them for the runtime check. We also have to assume they could
04216   // overlap. In the future there should be metadata for whether address spaces
04217   // are disjoint.
04218   unsigned NumPointers = RtCheck.Pointers.size();
04219   for (unsigned i = 0; i < NumPointers; ++i) {
04220     for (unsigned j = i + 1; j < NumPointers; ++j) {
04221       // Only need to check pointers between two different dependency sets.
04222       if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
04223        continue;
04224       // Only need to check pointers in the same alias set.
04225       if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
04226         continue;
04227 
04228       Value *PtrI = RtCheck.Pointers[i];
04229       Value *PtrJ = RtCheck.Pointers[j];
04230 
04231       unsigned ASi = PtrI->getType()->getPointerAddressSpace();
04232       unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
04233       if (ASi != ASj) {
04234         DEBUG(dbgs() << "LV: Runtime check would require comparison between"
04235                        " different address spaces\n");
04236         return false;
04237       }
04238     }
04239   }
04240 
04241   return CanDoRT;
04242 }
04243 
04244 void AccessAnalysis::processMemAccesses() {
04245   // We process the set twice: first we process read-write pointers, last we
04246   // process read-only pointers. This allows us to skip dependence tests for
04247   // read-only pointers.
04248 
04249   DEBUG(dbgs() << "LV: Processing memory accesses...\n");
04250   DEBUG(dbgs() << "  AST: "; AST.dump());
04251   DEBUG(dbgs() << "LV:   Accesses:\n");
04252   DEBUG({
04253     for (auto A : Accesses)
04254       dbgs() << "\t" << *A.getPointer() << " (" <<
04255                 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
04256                                          "read-only" : "read")) << ")\n";
04257   });
04258 
04259   // The AliasSetTracker has nicely partitioned our pointers by metadata
04260   // compatibility and potential for underlying-object overlap. As a result, we
04261   // only need to check for potential pointer dependencies within each alias
04262   // set.
04263   for (auto &AS : AST) {
04264     // Note that both the alias-set tracker and the alias sets themselves used
04265     // linked lists internally and so the iteration order here is deterministic
04266     // (matching the original instruction order within each set).
04267 
04268     bool SetHasWrite = false;
04269 
04270     // Map of pointers to last access encountered.
04271     typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
04272     UnderlyingObjToAccessMap ObjToLastAccess;
04273 
04274     // Set of access to check after all writes have been processed.
04275     PtrAccessSet DeferredAccesses;
04276 
04277     // Iterate over each alias set twice, once to process read/write pointers,
04278     // and then to process read-only pointers.
04279     for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
04280       bool UseDeferred = SetIteration > 0;
04281       PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
04282 
04283       for (auto A : AS) {
04284         Value *Ptr = A.getValue();
04285         bool IsWrite = S.count(MemAccessInfo(Ptr, true));
04286 
04287         // If we're using the deferred access set, then it contains only reads.
04288         bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
04289         if (UseDeferred && !IsReadOnlyPtr)
04290           continue;
04291         // Otherwise, the pointer must be in the PtrAccessSet, either as a read
04292         // or a write.
04293         assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
04294                  S.count(MemAccessInfo(Ptr, false))) &&
04295                "Alias-set pointer not in the access set?");
04296 
04297         MemAccessInfo Access(Ptr, IsWrite);
04298         DepCands.insert(Access);
04299 
04300         // Memorize read-only pointers for later processing and skip them in the
04301         // first round (they need to be checked after we have seen all write
04302         // pointers). Note: we also mark pointer that are not consecutive as
04303         // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
04304         // the second check for "!IsWrite".
04305         if (!UseDeferred && IsReadOnlyPtr) {
04306           DeferredAccesses.insert(Access);
04307           continue;
04308         }
04309 
04310         // If this is a write - check other reads and writes for conflicts.  If
04311         // this is a read only check other writes for conflicts (but only if
04312         // there is no other write to the ptr - this is an optimization to
04313         // catch "a[i] = a[i] + " without having to do a dependence check).
04314         if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
04315           CheckDeps.insert(Access);
04316           IsRTCheckNeeded = true;
04317         }
04318 
04319         if (IsWrite)
04320           SetHasWrite = true;
04321 
04322         // Create sets of pointers connected by a shared alias set and
04323         // underlying object.
04324         typedef SmallVector<Value *, 16> ValueVector;
04325         ValueVector TempObjects;
04326         GetUnderlyingObjects(Ptr, TempObjects, DL);
04327         for (Value *UnderlyingObj : TempObjects) {
04328           UnderlyingObjToAccessMap::iterator Prev =
04329             ObjToLastAccess.find(UnderlyingObj);
04330           if (Prev != ObjToLastAccess.end())
04331             DepCands.unionSets(Access, Prev->second);
04332 
04333           ObjToLastAccess[UnderlyingObj] = Access;
04334         }
04335       }
04336     }
04337   }
04338 }
04339 
04340 namespace {
04341 /// \brief Checks memory dependences among accesses to the same underlying
04342 /// object to determine whether there vectorization is legal or not (and at
04343 /// which vectorization factor).
04344 ///
04345 /// This class works under the assumption that we already checked that memory
04346 /// locations with different underlying pointers are "must-not alias".
04347 /// We use the ScalarEvolution framework to symbolically evalutate access
04348 /// functions pairs. Since we currently don't restructure the loop we can rely
04349 /// on the program order of memory accesses to determine their safety.
04350 /// At the moment we will only deem accesses as safe for:
04351 ///  * A negative constant distance assuming program order.
04352 ///
04353 ///      Safe: tmp = a[i + 1];     OR     a[i + 1] = x;
04354 ///            a[i] = tmp;                y = a[i];
04355 ///
04356 ///   The latter case is safe because later checks guarantuee that there can't
04357 ///   be a cycle through a phi node (that is, we check that "x" and "y" is not
04358 ///   the same variable: a header phi can only be an induction or a reduction, a
04359 ///   reduction can't have a memory sink, an induction can't have a memory
04360 ///   source). This is important and must not be violated (or we have to
04361 ///   resort to checking for cycles through memory).
04362 ///
04363 ///  * A positive constant distance assuming program order that is bigger
04364 ///    than the biggest memory access.
04365 ///
04366 ///     tmp = a[i]        OR              b[i] = x
04367 ///     a[i+2] = tmp                      y = b[i+2];
04368 ///
04369 ///     Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
04370 ///
04371 ///  * Zero distances and all accesses have the same size.
04372 ///
04373 class MemoryDepChecker {
04374 public:
04375   typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
04376   typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
04377 
04378   MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
04379       : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
04380         ShouldRetryWithRuntimeCheck(false) {}
04381 
04382   /// \brief Register the location (instructions are given increasing numbers)
04383   /// of a write access.
04384   void addAccess(StoreInst *SI) {
04385     Value *Ptr = SI->getPointerOperand();
04386     Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
04387     InstMap.push_back(SI);
04388     ++AccessIdx;
04389   }
04390 
04391   /// \brief Register the location (instructions are given increasing numbers)
04392   /// of a write access.
04393   void addAccess(LoadInst *LI) {
04394     Value *Ptr = LI->getPointerOperand();
04395     Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
04396     InstMap.push_back(LI);
04397     ++AccessIdx;
04398   }
04399 
04400   /// \brief Check whether the dependencies between the accesses are safe.
04401   ///
04402   /// Only checks sets with elements in \p CheckDeps.
04403   bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
04404                    MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
04405 
04406   /// \brief The maximum number of bytes of a vector register we can vectorize
04407   /// the accesses safely with.
04408   unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
04409 
04410   /// \brief In same cases when the dependency check fails we can still
04411   /// vectorize the loop with a dynamic array access check.
04412   bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
04413 
04414 private:
04415   ScalarEvolution *SE;
04416   const DataLayout *DL;
04417   const Loop *InnermostLoop;
04418 
04419   /// \brief Maps access locations (ptr, read/write) to program order.
04420   DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
04421 
04422   /// \brief Memory access instructions in program order.
04423   SmallVector<Instruction *, 16> InstMap;
04424 
04425   /// \brief The program order index to be used for the next instruction.
04426   unsigned AccessIdx;
04427 
04428   // We can access this many bytes in parallel safely.
04429   unsigned MaxSafeDepDistBytes;
04430 
04431   /// \brief If we see a non-constant dependence distance we can still try to
04432   /// vectorize this loop with runtime checks.
04433   bool ShouldRetryWithRuntimeCheck;
04434 
04435   /// \brief Check whether there is a plausible dependence between the two
04436   /// accesses.
04437   ///
04438   /// Access \p A must happen before \p B in program order. The two indices
04439   /// identify the index into the program order map.
04440   ///
04441   /// This function checks  whether there is a plausible dependence (or the
04442   /// absence of such can't be proved) between the two accesses. If there is a
04443   /// plausible dependence but the dependence distance is bigger than one
04444   /// element access it records this distance in \p MaxSafeDepDistBytes (if this
04445   /// distance is smaller than any other distance encountered so far).
04446   /// Otherwise, this function returns true signaling a possible dependence.
04447   bool isDependent(const MemAccessInfo &A, unsigned AIdx,
04448                    const MemAccessInfo &B, unsigned BIdx,
04449                    ValueToValueMap &Strides);
04450 
04451   /// \brief Check whether the data dependence could prevent store-load
04452   /// forwarding.
04453   bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
04454 };
04455 
04456 } // end anonymous namespace
04457 
04458 static bool isInBoundsGep(Value *Ptr) {
04459   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
04460     return GEP->isInBounds();
04461   return false;
04462 }
04463 
04464 /// \brief Check whether the access through \p Ptr has a constant stride.
04465 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
04466                         const Loop *Lp, ValueToValueMap &StridesMap) {
04467   const Type *Ty = Ptr->getType();
04468   assert(Ty->isPointerTy() && "Unexpected non-ptr");
04469 
04470   // Make sure that the pointer does not point to aggregate types.
04471   const PointerType *PtrTy = cast<PointerType>(Ty);
04472   if (PtrTy->getElementType()->isAggregateType()) {
04473     DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
04474           "\n");
04475     return 0;
04476   }
04477 
04478   const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
04479 
04480   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
04481   if (!AR) {
04482     DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
04483           << *Ptr << " SCEV: " << *PtrScev << "\n");
04484     return 0;
04485   }
04486 
04487   // The accesss function must stride over the innermost loop.
04488   if (Lp != AR->getLoop()) {
04489     DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
04490           *Ptr << " SCEV: " << *PtrScev << "\n");
04491   }
04492 
04493   // The address calculation must not wrap. Otherwise, a dependence could be
04494   // inverted.
04495   // An inbounds getelementptr that is a AddRec with a unit stride
04496   // cannot wrap per definition. The unit stride requirement is checked later.
04497   // An getelementptr without an inbounds attribute and unit stride would have
04498   // to access the pointer value "0" which is undefined behavior in address
04499   // space 0, therefore we can also vectorize this case.
04500   bool IsInBoundsGEP = isInBoundsGep(Ptr);
04501   bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
04502   bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
04503   if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
04504     DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
04505           << *Ptr << " SCEV: " << *PtrScev << "\n");
04506     return 0;
04507   }
04508 
04509   // Check the step is constant.
04510   const SCEV *Step = AR->getStepRecurrence(*SE);
04511 
04512   // Calculate the pointer stride and check if it is consecutive.
04513   const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
04514   if (!C) {
04515     DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
04516           " SCEV: " << *PtrScev << "\n");
04517     return 0;
04518   }
04519 
04520   int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
04521   const APInt &APStepVal = C->getValue()->getValue();
04522 
04523   // Huge step value - give up.
04524   if (APStepVal.getBitWidth() > 64)
04525     return 0;
04526 
04527   int64_t StepVal = APStepVal.getSExtValue();
04528 
04529   // Strided access.
04530   int64_t Stride = StepVal / Size;
04531   int64_t Rem = StepVal % Size;
04532   if (Rem)
04533     return 0;
04534 
04535   // If the SCEV could wrap but we have an inbounds gep with a unit stride we
04536   // know we can't "wrap around the address space". In case of address space
04537   // zero we know that this won't happen without triggering undefined behavior.
04538   if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
04539       Stride != 1 && Stride != -1)
04540     return 0;
04541 
04542   return Stride;
04543 }
04544 
04545 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
04546                                                     unsigned TypeByteSize) {
04547   // If loads occur at a distance that is not a multiple of a feasible vector
04548   // factor store-load forwarding does not take place.
04549   // Positive dependences might cause troubles because vectorizing them might
04550   // prevent store-load forwarding making vectorized code run a lot slower.
04551   //   a[i] = a[i-3] ^ a[i-8];
04552   //   The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
04553   //   hence on your typical architecture store-load forwarding does not take
04554   //   place. Vectorizing in such cases does not make sense.
04555   // Store-load forwarding distance.
04556   const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
04557   // Maximum vector factor.
04558   unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
04559   if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
04560     MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
04561 
04562   for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
04563        vf *= 2) {
04564     if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
04565       MaxVFWithoutSLForwardIssues = (vf >>=1);
04566       break;
04567     }
04568   }
04569 
04570   if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
04571     DEBUG(dbgs() << "LV: Distance " << Distance <<
04572           " that could cause a store-load forwarding conflict\n");
04573     return true;
04574   }
04575 
04576   if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
04577       MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
04578     MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
04579   return false;
04580 }
04581 
04582 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
04583                                    const MemAccessInfo &B, unsigned BIdx,
04584                                    ValueToValueMap &Strides) {
04585   assert (AIdx < BIdx && "Must pass arguments in program order");
04586 
04587   Value *APtr = A.getPointer();
04588   Value *BPtr = B.getPointer();
04589   bool AIsWrite = A.getInt();
04590   bool BIsWrite = B.getInt();
04591 
04592   // Two reads are independent.
04593   if (!AIsWrite && !BIsWrite)
04594     return false;
04595 
04596   // We cannot check pointers in different address spaces.
04597   if (APtr->getType()->getPointerAddressSpace() !=
04598       BPtr->getType()->getPointerAddressSpace())
04599     return true;
04600 
04601   const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
04602   const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
04603 
04604   int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
04605   int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
04606 
04607   const SCEV *Src = AScev;
04608   const SCEV *Sink = BScev;
04609 
04610   // If the induction step is negative we have to invert source and sink of the
04611   // dependence.
04612   if (StrideAPtr < 0) {
04613     //Src = BScev;
04614     //Sink = AScev;
04615     std::swap(APtr, BPtr);
04616     std::swap(Src, Sink);
04617     std::swap(AIsWrite, BIsWrite);
04618     std::swap(AIdx, BIdx);
04619     std::swap(StrideAPtr, StrideBPtr);
04620   }
04621 
04622   const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
04623 
04624   DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
04625         << "(Induction step: " << StrideAPtr <<  ")\n");
04626   DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
04627         << *InstMap[BIdx] << ": " << *Dist << "\n");
04628 
04629   // Need consecutive accesses. We don't want to vectorize
04630   // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
04631   // the address space.
04632   if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
04633     DEBUG(dbgs() << "Non-consecutive pointer access\n");
04634     return true;
04635   }
04636 
04637   const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
04638   if (!C) {
04639     DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
04640     ShouldRetryWithRuntimeCheck = true;
04641     return true;
04642   }
04643 
04644   Type *ATy = APtr->getType()->getPointerElementType();
04645   Type *BTy = BPtr->getType()->getPointerElementType();
04646   unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
04647 
04648   // Negative distances are not plausible dependencies.
04649   const APInt &Val = C->getValue()->getValue();
04650   if (Val.isNegative()) {
04651     bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
04652     if (IsTrueDataDependence &&
04653         (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
04654          ATy != BTy))
04655       return true;
04656 
04657     DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
04658     return false;
04659   }
04660 
04661   // Write to the same location with the same size.
04662   // Could be improved to assert type sizes are the same (i32 == float, etc).
04663   if (Val == 0) {
04664     if (ATy == BTy)
04665       return false;
04666     DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
04667     return true;
04668   }
04669 
04670   assert(Val.isStrictlyPositive() && "Expect a positive value");
04671 
04672   // Positive distance bigger than max vectorization factor.
04673   if (ATy != BTy) {
04674     DEBUG(dbgs() <<
04675           "LV: ReadWrite-Write positive dependency with different types\n");
04676     return false;
04677   }
04678 
04679   unsigned Distance = (unsigned) Val.getZExtValue();
04680 
04681   // Bail out early if passed-in parameters make vectorization not feasible.
04682   unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
04683   unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
04684 
04685   // The distance must be bigger than the size needed for a vectorized version
04686   // of the operation and the size of the vectorized operation must not be
04687   // bigger than the currrent maximum size.
04688   if (Distance < 2*TypeByteSize ||
04689       2*TypeByteSize > MaxSafeDepDistBytes ||
04690       Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
04691     DEBUG(dbgs() << "LV: Failure because of Positive distance "
04692         << Val.getSExtValue() << '\n');
04693     return true;
04694   }
04695 
04696   MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
04697     Distance : MaxSafeDepDistBytes;
04698 
04699   bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
04700   if (IsTrueDataDependence &&
04701       couldPreventStoreLoadForward(Distance, TypeByteSize))
04702      return true;
04703 
04704   DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
04705         " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
04706 
04707   return false;
04708 }
04709 
04710 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
04711                                    MemAccessInfoSet &CheckDeps,
04712                                    ValueToValueMap &Strides) {
04713 
04714   MaxSafeDepDistBytes = -1U;
04715   while (!CheckDeps.empty()) {
04716     MemAccessInfo CurAccess = *CheckDeps.begin();
04717 
04718     // Get the relevant memory access set.
04719     EquivalenceClasses<MemAccessInfo>::iterator I =
04720       AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
04721 
04722     // Check accesses within this set.
04723     EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
04724     AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
04725 
04726     // Check every access pair.
04727     while (AI != AE) {
04728       CheckDeps.erase(*AI);
04729       EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
04730       while (OI != AE) {
04731         // Check every accessing instruction pair in program order.
04732         for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
04733              I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
04734           for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
04735                I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
04736             if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
04737               return false;
04738             if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
04739               return false;
04740           }
04741         ++OI;
04742       }
04743       AI++;
04744     }
04745   }
04746   return true;
04747 }
04748 
04749 bool LoopVectorizationLegality::canVectorizeMemory() {
04750 
04751   typedef SmallVector<Value*, 16> ValueVector;
04752   typedef SmallPtrSet<Value*, 16> ValueSet;
04753 
04754   // Holds the Load and Store *instructions*.
04755   ValueVector Loads;
04756   ValueVector Stores;
04757 
04758   // Holds all the different accesses in the loop.
04759   unsigned NumReads = 0;
04760   unsigned NumReadWrites = 0;
04761 
04762   PtrRtCheck.Pointers.clear();
04763   PtrRtCheck.Need = false;
04764 
04765   const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
04766   MemoryDepChecker DepChecker(SE, DL, TheLoop);
04767 
04768   // For each block.
04769   for (Loop::block_iterator bb = TheLoop->block_begin(),
04770        be = TheLoop->block_end(); bb != be; ++bb) {
04771 
04772     // Scan the BB and collect legal loads and stores.
04773     for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
04774          ++it) {
04775 
04776       // If this is a load, save it. If this instruction can read from memory
04777       // but is not a load, then we quit. Notice that we don't handle function
04778       // calls that read or write.
04779       if (it->mayReadFromMemory()) {
04780         // Many math library functions read the rounding mode. We will only
04781         // vectorize a loop if it contains known function calls that don't set
04782         // the flag. Therefore, it is safe to ignore this read from memory.
04783         CallInst *Call = dyn_cast<CallInst>(it);
04784         if (Call && getIntrinsicIDForCall(Call, TLI))
04785           continue;
04786 
04787         LoadInst *Ld = dyn_cast<LoadInst>(it);
04788         if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
04789           emitAnalysis(Report(Ld)
04790                        << "read with atomic ordering or volatile read");
04791           DEBUG(dbgs() << "LV: Found a non-simple load.\n");
04792           return false;
04793         }
04794         NumLoads++;
04795         Loads.push_back(Ld);
04796         DepChecker.addAccess(Ld);
04797         continue;
04798       }
04799 
04800       // Save 'store' instructions. Abort if other instructions write to memory.
04801       if (it->mayWriteToMemory()) {
04802         StoreInst *St = dyn_cast<StoreInst>(it);
04803         if (!St) {
04804           emitAnalysis(Report(it) << "instruction cannot be vectorized");
04805           return false;
04806         }
04807         if (!St->isSimple() && !IsAnnotatedParallel) {
04808           emitAnalysis(Report(St)
04809                        << "write with atomic ordering or volatile write");
04810           DEBUG(dbgs() << "LV: Found a non-simple store.\n");
04811           return false;
04812         }
04813         NumStores++;
04814         Stores.push_back(St);
04815         DepChecker.addAccess(St);
04816       }
04817     } // Next instr.
04818   } // Next block.
04819 
04820   // Now we have two lists that hold the loads and the stores.
04821   // Next, we find the pointers that they use.
04822 
04823   // Check if we see any stores. If there are no stores, then we don't
04824   // care if the pointers are *restrict*.
04825   if (!Stores.size()) {
04826     DEBUG(dbgs() << "LV: Found a read-only loop!\n");
04827     return true;
04828   }
04829 
04830   AccessAnalysis::DepCandidates DependentAccesses;
04831   AccessAnalysis Accesses(DL, AA, DependentAccesses);
04832 
04833   // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
04834   // multiple times on the same object. If the ptr is accessed twice, once
04835   // for read and once for write, it will only appear once (on the write
04836   // list). This is okay, since we are going to check for conflicts between
04837   // writes and between reads and writes, but not between reads and reads.
04838   ValueSet Seen;
04839 
04840   ValueVector::iterator I, IE;
04841   for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
04842     StoreInst *ST = cast<StoreInst>(*I);
04843     Value* Ptr = ST->getPointerOperand();
04844 
04845     if (isUniform(Ptr)) {
04846       emitAnalysis(
04847           Report(ST)
04848           << "write to a loop invariant address could not be vectorized");
04849       DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
04850       return false;
04851     }
04852 
04853     // If we did *not* see this pointer before, insert it to  the read-write
04854     // list. At this phase it is only a 'write' list.
04855     if (Seen.insert(Ptr).second) {
04856       ++NumReadWrites;
04857 
04858       AliasAnalysis::Location Loc = AA->getLocation(ST);
04859       // The TBAA metadata could have a control dependency on the predication
04860       // condition, so we cannot rely on it when determining whether or not we
04861       // need runtime pointer checks.
04862       if (blockNeedsPredication(ST->getParent()))
04863         Loc.AATags.TBAA = nullptr;
04864 
04865       Accesses.addStore(Loc);
04866     }
04867   }
04868 
04869   if (IsAnnotatedParallel) {
04870     DEBUG(dbgs()
04871           << "LV: A loop annotated parallel, ignore memory dependency "
04872           << "checks.\n");
04873     return true;
04874   }
04875 
04876   for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
04877     LoadInst *LD = cast<LoadInst>(*I);
04878     Value* Ptr = LD->getPointerOperand();
04879     // If we did *not* see this pointer before, insert it to the
04880     // read list. If we *did* see it before, then it is already in
04881     // the read-write list. This allows us to vectorize expressions
04882     // such as A[i] += x;  Because the address of A[i] is a read-write
04883     // pointer. This only works if the index of A[i] is consecutive.
04884     // If the address of i is unknown (for example A[B[i]]) then we may
04885     // read a few words, modify, and write a few words, and some of the
04886     // words may be written to the same address.
04887     bool IsReadOnlyPtr = false;
04888     if (Seen.insert(Ptr).second ||
04889         !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
04890       ++NumReads;
04891       IsReadOnlyPtr = true;
04892     }
04893 
04894     AliasAnalysis::Location Loc = AA->getLocation(LD);
04895     // The TBAA metadata could have a control dependency on the predication
04896     // condition, so we cannot rely on it when determining whether or not we
04897     // need runtime pointer checks.
04898     if (blockNeedsPredication(LD->getParent()))
04899       Loc.AATags.TBAA = nullptr;
04900 
04901     Accesses.addLoad(Loc, IsReadOnlyPtr);
04902   }
04903 
04904   // If we write (or read-write) to a single destination and there are no
04905   // other reads in this loop then is it safe to vectorize.
04906   if (NumReadWrites == 1 && NumReads == 0) {
04907     DEBUG(dbgs() << "LV: Found a write-only loop!\n");
04908     return true;
04909   }
04910 
04911   // Build dependence sets and check whether we need a runtime pointer bounds
04912   // check.
04913   Accesses.buildDependenceSets();
04914   bool NeedRTCheck = Accesses.isRTCheckNeeded();
04915 
04916   // Find pointers with computable bounds. We are going to use this information
04917   // to place a runtime bound check.
04918   unsigned NumComparisons = 0;
04919   bool CanDoRT = false;
04920   if (NeedRTCheck)
04921     CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
04922                                        Strides);
04923 
04924   DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
04925         " pointer comparisons.\n");
04926 
04927   // If we only have one set of dependences to check pointers among we don't
04928   // need a runtime check.
04929   if (NumComparisons == 0 && NeedRTCheck)
04930     NeedRTCheck = false;
04931 
04932   // Check that we did not collect too many pointers or found an unsizeable
04933   // pointer.
04934   if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
04935     PtrRtCheck.reset();
04936     CanDoRT = false;
04937   }
04938 
04939   if (CanDoRT) {
04940     DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
04941   }
04942 
04943   if (NeedRTCheck && !CanDoRT) {
04944     emitAnalysis(Report() << "cannot identify array bounds");
04945     DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
04946           "the array bounds.\n");
04947     PtrRtCheck.reset();
04948     return false;
04949   }
04950 
04951   PtrRtCheck.Need = NeedRTCheck;
04952 
04953   bool CanVecMem = true;
04954   if (Accesses.isDependencyCheckNeeded()) {
04955     DEBUG(dbgs() << "LV: Checking memory dependencies\n");
04956     CanVecMem = DepChecker.areDepsSafe(
04957         DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
04958     MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
04959 
04960     if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
04961       DEBUG(dbgs() << "LV: Retrying with memory checks\n");
04962       NeedRTCheck = true;
04963 
04964       // Clear the dependency checks. We assume they are not needed.
04965       Accesses.resetDepChecks();
04966 
04967       PtrRtCheck.reset();
04968       PtrRtCheck.Need = true;
04969 
04970       CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
04971                                          TheLoop, Strides, true);
04972       // Check that we did not collect too many pointers or found an unsizeable
04973       // pointer.
04974       if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
04975         if (!CanDoRT && NumComparisons > 0)
04976           emitAnalysis(Report()
04977                        << "cannot check memory dependencies at runtime");
04978         else
04979           emitAnalysis(Report()
04980                        << NumComparisons << " exceeds limit of "
04981                        << RuntimeMemoryCheckThreshold
04982                        << " dependent memory operations checked at runtime");
04983         DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
04984         PtrRtCheck.reset();
04985         return false;
04986       }
04987 
04988       CanVecMem = true;
04989     }
04990   }
04991 
04992   if (!CanVecMem)
04993     emitAnalysis(Report() << "unsafe dependent memory operations in loop");
04994 
04995   DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
04996         " need a runtime memory check.\n");
04997 
04998   return CanVecMem;
04999 }
05000 
05001 static bool hasMultipleUsesOf(Instruction *I,
05002                               SmallPtrSetImpl<Instruction *> &Insts) {
05003   unsigned NumUses = 0;
05004   for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
05005     if (Insts.count(dyn_cast<Instruction>(*Use)))
05006       ++NumUses;
05007     if (NumUses > 1)
05008       return true;
05009   }
05010 
05011   return false;
05012 }
05013 
05014 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
05015   for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
05016     if (!Set.count(dyn_cast<Instruction>(*Use)))
05017       return false;
05018   return true;
05019 }
05020 
05021 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
05022                                                 ReductionKind Kind) {
05023   if (Phi->getNumIncomingValues() != 2)
05024     return false;
05025 
05026   // Reduction variables are only found in the loop header block.
05027   if (Phi->getParent() != TheLoop->getHeader())
05028     return false;
05029 
05030   // Obtain the reduction start value from the value that comes from the loop
05031   // preheader.
05032   Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
05033 
05034   // ExitInstruction is the single value which is used outside the loop.
05035   // We only allow for a single reduction value to be used outside the loop.
05036   // This includes users of the reduction, variables (which form a cycle
05037   // which ends in the phi node).
05038   Instruction *ExitInstruction = nullptr;
05039   // Indicates that we found a reduction operation in our scan.
05040   bool FoundReduxOp = false;
05041 
05042   // We start with the PHI node and scan for all of the users of this
05043   // instruction. All users must be instructions that can be used as reduction
05044   // variables (such as ADD). We must have a single out-of-block user. The cycle
05045   // must include the original PHI.
05046   bool FoundStartPHI = false;
05047 
05048   // To recognize min/max patterns formed by a icmp select sequence, we store
05049   // the number of instruction we saw from the recognized min/max pattern,
05050   //  to make sure we only see exactly the two instructions.
05051   unsigned NumCmpSelectPatternInst = 0;
05052   ReductionInstDesc ReduxDesc(false, nullptr);
05053 
05054   SmallPtrSet<Instruction *, 8> VisitedInsts;
05055   SmallVector<Instruction *, 8> Worklist;
05056   Worklist.push_back(Phi);
05057   VisitedInsts.insert(Phi);
05058 
05059   // A value in the reduction can be used:
05060   //  - By the reduction:
05061   //      - Reduction operation:
05062   //        - One use of reduction value (safe).
05063   //        - Multiple use of reduction value (not safe).
05064   //      - PHI:
05065   //        - All uses of the PHI must be the reduction (safe).
05066   //        - Otherwise, not safe.
05067   //  - By one instruction outside of the loop (safe).
05068   //  - By further instructions outside of the loop (not safe).
05069   //  - By an instruction that is not part of the reduction (not safe).
05070   //    This is either:
05071   //      * An instruction type other than PHI or the reduction operation.
05072   //      * A PHI in the header other than the initial PHI.
05073   while (!Worklist.empty()) {
05074     Instruction *Cur = Worklist.back();
05075     Worklist.pop_back();
05076 
05077     // No Users.
05078     // If the instruction has no users then this is a broken chain and can't be
05079     // a reduction variable.
05080     if (Cur->use_empty())
05081       return false;
05082 
05083     bool IsAPhi = isa<PHINode>(Cur);
05084 
05085     // A header PHI use other than the original PHI.
05086     if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
05087       return false;
05088 
05089     // Reductions of instructions such as Div, and Sub is only possible if the
05090     // LHS is the reduction variable.
05091     if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
05092         !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
05093         !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
05094       return false;
05095 
05096     // Any reduction instruction must be of one of the allowed kinds.
05097     ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
05098     if (!ReduxDesc.IsReduction)
05099       return false;
05100 
05101     // A reduction operation must only have one use of the reduction value.
05102     if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
05103         hasMultipleUsesOf(Cur, VisitedInsts))
05104       return false;
05105 
05106     // All inputs to a PHI node must be a reduction value.
05107     if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
05108       return false;
05109 
05110     if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
05111                                      isa<SelectInst>(Cur)))
05112       ++NumCmpSelectPatternInst;
05113     if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
05114                                    isa<SelectInst>(Cur)))
05115       ++NumCmpSelectPatternInst;
05116 
05117     // Check  whether we found a reduction operator.
05118     FoundReduxOp |= !IsAPhi;
05119 
05120     // Process users of current instruction. Push non-PHI nodes after PHI nodes
05121     // onto the stack. This way we are going to have seen all inputs to PHI
05122     // nodes once we get to them.
05123     SmallVector<Instruction *, 8> NonPHIs;
05124     SmallVector<Instruction *, 8> PHIs;
05125     for (User *U : Cur->users()) {
05126       Instruction *UI = cast<Instruction>(U);
05127 
05128       // Check if we found the exit user.
05129       BasicBlock *Parent = UI->getParent();
05130       if (!TheLoop->contains(Parent)) {
05131         // Exit if you find multiple outside users or if the header phi node is
05132         // being used. In this case the user uses the value of the previous
05133         // iteration, in which case we would loose "VF-1" iterations of the
05134         // reduction operation if we vectorize.
05135         if (ExitInstruction != nullptr || Cur == Phi)
05136           return false;
05137 
05138         // The instruction used by an outside user must be the last instruction
05139         // before we feed back to the reduction phi. Otherwise, we loose VF-1
05140         // operations on the value.
05141         if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
05142          return false;
05143 
05144         ExitInstruction = Cur;
05145         continue;
05146       }
05147 
05148       // Process instructions only once (termination). Each reduction cycle
05149       // value must only be used once, except by phi nodes and min/max
05150       // reductions which are represented as a cmp followed by a select.
05151       ReductionInstDesc IgnoredVal(false, nullptr);
05152       if (VisitedInsts.insert(UI).second) {
05153         if (isa<PHINode>(UI))
05154           PHIs.push_back(UI);
05155         else
05156           NonPHIs.push_back(UI);
05157       } else if (!isa<PHINode>(UI) &&
05158                  ((!isa<FCmpInst>(UI) &&
05159                    !isa<ICmpInst>(UI) &&
05160                    !isa<SelectInst>(UI)) ||
05161                   !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
05162         return false;
05163 
05164       // Remember that we completed the cycle.
05165       if (UI == Phi)
05166         FoundStartPHI = true;
05167     }
05168     Worklist.append(PHIs.begin(), PHIs.end());
05169     Worklist.append(NonPHIs.begin(), NonPHIs.end());
05170   }
05171 
05172   // This means we have seen one but not the other instruction of the
05173   // pattern or more than just a select and cmp.
05174   if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
05175       NumCmpSelectPatternInst != 2)
05176     return false;
05177 
05178   if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
05179     return false;
05180 
05181   // We found a reduction var if we have reached the original phi node and we
05182   // only have a single instruction with out-of-loop users.
05183 
05184   // This instruction is allowed to have out-of-loop users.
05185   AllowedExit.insert(ExitInstruction);
05186 
05187   // Save the description of this reduction variable.
05188   ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
05189                          ReduxDesc.MinMaxKind);
05190   Reductions[Phi] = RD;
05191   // We've ended the cycle. This is a reduction variable if we have an
05192   // outside user and it has a binary op.
05193 
05194   return true;
05195 }
05196 
05197 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
05198 /// pattern corresponding to a min(X, Y) or max(X, Y).
05199 LoopVectorizationLegality::ReductionInstDesc
05200 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
05201                                                     ReductionInstDesc &Prev) {
05202 
05203   assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
05204          "Expect a select instruction");
05205   Instruction *Cmp = nullptr;
05206   SelectInst *Select = nullptr;
05207 
05208   // We must handle the select(cmp()) as a single instruction. Advance to the
05209   // select.
05210   if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
05211     if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
05212       return ReductionInstDesc(false, I);
05213     return ReductionInstDesc(Select, Prev.MinMaxKind);
05214   }
05215 
05216   // Only handle single use cases for now.
05217   if (!(Select = dyn_cast<SelectInst>(I)))
05218     return ReductionInstDesc(false, I);
05219   if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
05220       !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
05221     return ReductionInstDesc(false, I);
05222   if (!Cmp->hasOneUse())
05223     return ReductionInstDesc(false, I);
05224 
05225   Value *CmpLeft;
05226   Value *CmpRight;
05227 
05228   // Look for a min/max pattern.
05229   if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05230     return ReductionInstDesc(Select, MRK_UIntMin);
05231   else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05232     return ReductionInstDesc(Select, MRK_UIntMax);
05233   else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05234     return ReductionInstDesc(Select, MRK_SIntMax);
05235   else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05236     return ReductionInstDesc(Select, MRK_SIntMin);
05237   else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05238     return ReductionInstDesc(Select, MRK_FloatMin);
05239   else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05240     return ReductionInstDesc(Select, MRK_FloatMax);
05241   else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05242     return ReductionInstDesc(Select, MRK_FloatMin);
05243   else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
05244     return ReductionInstDesc(Select, MRK_FloatMax);
05245 
05246   return ReductionInstDesc(false, I);
05247 }
05248 
05249 LoopVectorizationLegality::ReductionInstDesc
05250 LoopVectorizationLegality::isReductionInstr(Instruction *I,
05251                                             ReductionKind Kind,
05252                                             ReductionInstDesc &Prev) {
05253   bool FP = I->getType()->isFloatingPointTy();
05254   bool FastMath = FP && I->hasUnsafeAlgebra();
05255   switch (I->getOpcode()) {
05256   default:
05257     return ReductionInstDesc(false, I);
05258   case Instruction::PHI:
05259       if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
05260                  Kind != RK_FloatMinMax))
05261         return ReductionInstDesc(false, I);
05262     return ReductionInstDesc(I, Prev.MinMaxKind);
05263   case Instruction::Sub:
05264   case Instruction::Add:
05265     return ReductionInstDesc(Kind == RK_IntegerAdd, I);
05266   case Instruction::Mul:
05267     return ReductionInstDesc(Kind == RK_IntegerMult, I);
05268   case Instruction::And:
05269     return ReductionInstDesc(Kind == RK_IntegerAnd, I);
05270   case Instruction::Or:
05271     return ReductionInstDesc(Kind == RK_IntegerOr, I);
05272   case Instruction::Xor:
05273     return ReductionInstDesc(Kind == RK_IntegerXor, I);
05274   case Instruction::FMul:
05275     return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
05276   case Instruction::FSub:
05277   case Instruction::FAdd:
05278     return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
05279   case Instruction::FCmp:
05280   case Instruction::ICmp:
05281   case Instruction::Select:
05282     if (Kind != RK_IntegerMinMax &&
05283         (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
05284       return ReductionInstDesc(false, I);
05285     return isMinMaxSelectCmpPattern(I, Prev);
05286   }
05287 }
05288 
05289 LoopVectorizationLegality::InductionKind
05290 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
05291   Type *PhiTy = Phi->getType();
05292   // We only handle integer and pointer inductions variables.
05293   if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
05294     return IK_NoInduction;
05295 
05296   // Check that the PHI is consecutive.
05297   const SCEV *PhiScev = SE->getSCEV(Phi);
05298   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
05299   if (!AR) {
05300     DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
05301     return IK_NoInduction;
05302   }
05303   const SCEV *Step = AR->getStepRecurrence(*SE);
05304 
05305   // Integer inductions need to have a stride of one.
05306   if (PhiTy->isIntegerTy()) {
05307     if (Step->isOne())
05308       return IK_IntInduction;
05309     if (Step->isAllOnesValue())
05310       return IK_ReverseIntInduction;
05311     return IK_NoInduction;
05312   }
05313 
05314   // Calculate the pointer stride and check if it is consecutive.
05315   const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
05316   if (!C)
05317     return IK_NoInduction;
05318 
05319   assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
05320   Type *PointerElementType = PhiTy->getPointerElementType();
05321   // The pointer stride cannot be determined if the pointer element type is not
05322   // sized.
05323   if (!PointerElementType->isSized())
05324     return IK_NoInduction;
05325 
05326   uint64_t Size = DL->getTypeAllocSize(PointerElementType);
05327   if (C->getValue()->equalsInt(Size))
05328     return IK_PtrInduction;
05329   else if (C->getValue()->equalsInt(0 - Size))
05330     return IK_ReversePtrInduction;
05331 
05332   return IK_NoInduction;
05333 }
05334 
05335 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
05336   Value *In0 = const_cast<Value*>(V);
05337   PHINode *PN = dyn_cast_or_null<PHINode>(In0);
05338   if (!PN)
05339     return false;
05340 
05341   return Inductions.count(PN);
05342 }
05343 
05344 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB)  {
05345   assert(TheLoop->contains(BB) && "Unknown block used");
05346 
05347   // Blocks that do not dominate the latch need predication.
05348   BasicBlock* Latch = TheLoop->getLoopLatch();
05349   return !DT->dominates(BB, Latch);
05350 }
05351 
05352 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
05353                                            SmallPtrSetImpl<Value *> &SafePtrs) {
05354   for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
05355     // We might be able to hoist the load.
05356     if (it->mayReadFromMemory()) {
05357       LoadInst *LI = dyn_cast<LoadInst>(it);
05358       if (!LI)
05359         return false;
05360       if (!SafePtrs.count(LI->getPointerOperand())) {
05361         if (canPredicateLoad(LI->getType(), LI->getPointerOperand())) {
05362           MaskedOp.insert(LI);
05363           continue;
05364         }
05365         return false;
05366       }
05367     }
05368 
05369     // We don't predicate stores at the moment.
05370     if (it->mayWriteToMemory()) {
05371       StoreInst *SI = dyn_cast<StoreInst>(it);
05372       // We only support predication of stores in basic blocks with one
05373       // predecessor.
05374       if (!SI)
05375         return false;
05376 
05377       if (++NumPredStores > NumberOfStoresToPredicate ||
05378           !SafePtrs.count(SI->getPointerOperand()) ||
05379           !SI->getParent()->getSinglePredecessor()) {
05380         if (canPredicateStore(SI->getValueOperand()->getType(),
05381                                   SI->getPointerOperand())) {
05382           MaskedOp.insert(SI);
05383           --NumPredStores;
05384           continue;
05385         }
05386         return false;
05387       }        
05388     }
05389     if (it->mayThrow())
05390       return false;
05391 
05392     // Check that we don't have a constant expression that can trap as operand.
05393     for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
05394          OI != OE; ++OI) {
05395       if (Constant *C = dyn_cast<Constant>(*OI))
05396         if (C->canTrap())
05397           return false;
05398     }
05399 
05400     // The instructions below can trap.
05401     switch (it->getOpcode()) {
05402     default: continue;
05403     case Instruction::UDiv:
05404     case Instruction::SDiv:
05405     case Instruction::URem:
05406     case Instruction::SRem:
05407              return false;
05408     }
05409   }
05410 
05411   return true;
05412 }
05413 
05414 LoopVectorizationCostModel::VectorizationFactor
05415 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
05416   // Width 1 means no vectorize
05417   VectorizationFactor Factor = { 1U, 0U };
05418   if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
05419     emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
05420     DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
05421     return Factor;
05422   }
05423 
05424   if (!EnableCondStoresVectorization && Legal->NumPredStores) {
05425     emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
05426     DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
05427     return Factor;
05428   }
05429 
05430   // Find the trip count.
05431   unsigned TC = SE->getSmallConstantTripCount(TheLoop);
05432   DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
05433 
05434   unsigned WidestType = getWidestType();
05435   unsigned WidestRegister = TTI.getRegisterBitWidth(true);
05436   unsigned MaxSafeDepDist = -1U;
05437   if (Legal->getMaxSafeDepDistBytes() != -1U)
05438     MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
05439   WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
05440                     WidestRegister : MaxSafeDepDist);
05441   unsigned MaxVectorSize = WidestRegister / WidestType;
05442   DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
05443   DEBUG(dbgs() << "LV: The Widest register is: "
05444           << WidestRegister << " bits.\n");
05445 
05446   if (MaxVectorSize == 0) {
05447     DEBUG(dbgs() << "LV: The target has no vector registers.\n");
05448     MaxVectorSize = 1;
05449   }
05450 
05451   assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
05452          " into one vector!");
05453 
05454   unsigned VF = MaxVectorSize;
05455 
05456   // If we optimize the program for size, avoid creating the tail loop.
05457   if (OptForSize) {
05458     // If we are unable to calculate the trip count then don't try to vectorize.
05459     if (TC < 2) {
05460       emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
05461       DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
05462       return Factor;
05463     }
05464 
05465     // Find the maximum SIMD width that can fit within the trip count.
05466     VF = TC % MaxVectorSize;
05467 
05468     if (VF == 0)
05469       VF = MaxVectorSize;
05470 
05471     // If the trip count that we found modulo the vectorization factor is not
05472     // zero then we require a tail.
05473     if (VF < 2) {
05474       emitAnalysis(Report() << "cannot optimize for size and vectorize at the "
05475                                "same time. Enable vectorization of this loop "
05476                                "with '#pragma clang loop vectorize(enable)' "
05477                                "when compiling with -Os");
05478       DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
05479       return Factor;
05480     }
05481   }
05482 
05483   int UserVF = Hints->getWidth();
05484   if (UserVF != 0) {
05485     assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
05486     DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
05487 
05488     Factor.Width = UserVF;
05489     return Factor;
05490   }
05491 
05492   float Cost = expectedCost(1);
05493 #ifndef NDEBUG
05494   const float ScalarCost = Cost;
05495 #endif /* NDEBUG */
05496   unsigned Width = 1;
05497   DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
05498 
05499   bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
05500   // Ignore scalar width, because the user explicitly wants vectorization.
05501   if (ForceVectorization && VF > 1) {
05502     Width = 2;
05503     Cost = expectedCost(Width) / (float)Width;
05504   }
05505 
05506   for (unsigned i=2; i <= VF; i*=2) {
05507     // Notice that the vector loop needs to be executed less times, so
05508     // we need to divide the cost of the vector loops by the width of
05509     // the vector elements.
05510     float VectorCost = expectedCost(i) / (float)i;
05511     DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
05512           VectorCost << ".\n");
05513     if (VectorCost < Cost) {
05514       Cost = VectorCost;
05515       Width = i;
05516     }
05517   }
05518 
05519   DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
05520         << "LV: Vectorization seems to be not beneficial, "
05521         << "but was forced by a user.\n");
05522   DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
05523   Factor.Width = Width;
05524   Factor.Cost = Width * Cost;
05525   return Factor;
05526 }
05527 
05528 unsigned LoopVectorizationCostModel::getWidestType() {
05529   unsigned MaxWidth = 8;
05530 
05531   // For each block.
05532   for (Loop::block_iterator bb = TheLoop->block_begin(),
05533        be = TheLoop->block_end(); bb != be; ++bb) {
05534     BasicBlock *BB = *bb;
05535 
05536     // For each instruction in the loop.
05537     for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
05538       Type *T = it->getType();
05539 
05540       // Ignore ephemeral values.
05541       if (EphValues.count(it))
05542         continue;
05543 
05544       // Only examine Loads, Stores and PHINodes.
05545       if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
05546         continue;
05547 
05548       // Examine PHI nodes that are reduction variables.
05549       if (PHINode *PN = dyn_cast<PHINode>(it))
05550         if (!Legal->getReductionVars()->count(PN))
05551           continue;
05552 
05553       // Examine the stored values.
05554       if (StoreInst *ST = dyn_cast<StoreInst>(it))
05555         T = ST->getValueOperand()->getType();
05556 
05557       // Ignore loaded pointer types and stored pointer types that are not
05558       // consecutive. However, we do want to take consecutive stores/loads of
05559       // pointer vectors into account.
05560       if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
05561         continue;
05562 
05563       MaxWidth = std::max(MaxWidth,
05564                           (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
05565     }
05566   }
05567 
05568   return MaxWidth;
05569 }
05570 
05571 unsigned
05572 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
05573                                                unsigned VF,
05574                                                unsigned LoopCost) {
05575 
05576   // -- The unroll heuristics --
05577   // We unroll the loop in order to expose ILP and reduce the loop overhead.
05578   // There are many micro-architectural considerations that we can't predict
05579   // at this level. For example, frontend pressure (on decode or fetch) due to
05580   // code size, or the number and capabilities of the execution ports.
05581   //
05582   // We use the following heuristics to select the unroll factor:
05583   // 1. If the code has reductions, then we unroll in order to break the cross
05584   // iteration dependency.
05585   // 2. If the loop is really small, then we unroll in order to reduce the loop
05586   // overhead.
05587   // 3. We don't unroll if we think that we will spill registers to memory due
05588   // to the increased register pressure.
05589 
05590   // Use the user preference, unless 'auto' is selected.
05591   int UserUF = Hints->getInterleave();
05592   if (UserUF != 0)
05593     return UserUF;
05594 
05595   // When we optimize for size, we don't unroll.
05596   if (OptForSize)
05597     return 1;
05598 
05599   // We used the distance for the unroll factor.
05600   if (Legal->getMaxSafeDepDistBytes() != -1U)
05601     return 1;
05602 
05603   // Do not unroll loops with a relatively small trip count.
05604   unsigned TC = SE->getSmallConstantTripCount(TheLoop);
05605   if (TC > 1 && TC < TinyTripCountUnrollThreshold)
05606     return 1;
05607 
05608   unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
05609   DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
05610         " registers\n");
05611 
05612   if (VF == 1) {
05613     if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
05614       TargetNumRegisters = ForceTargetNumScalarRegs;
05615   } else {
05616     if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
05617       TargetNumRegisters = ForceTargetNumVectorRegs;
05618   }
05619 
05620   LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
05621   // We divide by these constants so assume that we have at least one
05622   // instruction that uses at least one register.
05623   R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
05624   R.NumInstructions = std::max(R.NumInstructions, 1U);
05625 
05626   // We calculate the unroll factor using the following formula.
05627   // Subtract the number of loop invariants from the number of available
05628   // registers. These registers are used by all of the unrolled instances.
05629   // Next, divide the remaining registers by the number of registers that is
05630   // required by the loop, in order to estimate how many parallel instances
05631   // fit without causing spills. All of this is rounded down if necessary to be
05632   // a power of two. We want power of two unroll factors to simplify any
05633   // addressing operations or alignment considerations.
05634   unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
05635                               R.MaxLocalUsers);
05636 
05637   // Don't count the induction variable as unrolled.
05638   if (EnableIndVarRegisterHeur)
05639     UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
05640                        std::max(1U, (R.MaxLocalUsers - 1)));
05641 
05642   // Clamp the unroll factor ranges to reasonable factors.
05643   unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
05644 
05645   // Check if the user has overridden the unroll max.
05646   if (VF == 1) {
05647     if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
05648       MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
05649   } else {
05650     if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
05651       MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
05652   }
05653 
05654   // If we did not calculate the cost for VF (because the user selected the VF)
05655   // then we calculate the cost of VF here.
05656   if (LoopCost == 0)
05657     LoopCost = expectedCost(VF);
05658 
05659   // Clamp the calculated UF to be between the 1 and the max unroll factor
05660   // that the target allows.
05661   if (UF > MaxInterleaveSize)
05662     UF = MaxInterleaveSize;
05663   else if (UF < 1)
05664     UF = 1;
05665 
05666   // Unroll if we vectorized this loop and there is a reduction that could
05667   // benefit from unrolling.
05668   if (VF > 1 && Legal->getReductionVars()->size()) {
05669     DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
05670     return UF;
05671   }
05672 
05673   // Note that if we've already vectorized the loop we will have done the
05674   // runtime check and so unrolling won't require further checks.
05675   bool UnrollingRequiresRuntimePointerCheck =
05676       (VF == 1 && Legal->getRuntimePointerCheck()->Need);
05677 
05678   // We want to unroll small loops in order to reduce the loop overhead and
05679   // potentially expose ILP opportunities.
05680   DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
05681   if (!UnrollingRequiresRuntimePointerCheck &&
05682       LoopCost < SmallLoopCost) {
05683     // We assume that the cost overhead is 1 and we use the cost model
05684     // to estimate the cost of the loop and unroll until the cost of the
05685     // loop overhead is about 5% of the cost of the loop.
05686     unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
05687 
05688     // Unroll until store/load ports (estimated by max unroll factor) are
05689     // saturated.
05690     unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
05691     unsigned LoadsUF = UF /  (Legal->NumLoads ? Legal->NumLoads : 1);
05692 
05693     // If we have a scalar reduction (vector reductions are already dealt with
05694     // by this point), we can increase the critical path length if the loop
05695     // we're unrolling is inside another loop. Limit, by default to 2, so the
05696     // critical path only gets increased by one reduction operation.
05697     if (Legal->getReductionVars()->size() &&
05698         TheLoop->getLoopDepth() > 1) {
05699       unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
05700       SmallUF = std::min(SmallUF, F);
05701       StoresUF = std::min(StoresUF, F);
05702       LoadsUF = std::min(LoadsUF, F);
05703     }
05704 
05705     if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
05706       DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
05707       return std::max(StoresUF, LoadsUF);
05708     }
05709 
05710     DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
05711     return SmallUF;
05712   }
05713 
05714   DEBUG(dbgs() << "LV: Not Unrolling.\n");
05715   return 1;
05716 }
05717 
05718 LoopVectorizationCostModel::RegisterUsage
05719 LoopVectorizationCostModel::calculateRegisterUsage() {
05720   // This function calculates the register usage by measuring the highest number
05721   // of values that are alive at a single location. Obviously, this is a very
05722   // rough estimation. We scan the loop in a topological order in order and
05723   // assign a number to each instruction. We use RPO to ensure that defs are
05724   // met before their users. We assume that each instruction that has in-loop
05725   // users starts an interval. We record every time that an in-loop value is
05726   // used, so we have a list of the first and last occurrences of each
05727   // instruction. Next, we transpose this data structure into a multi map that
05728   // holds the list of intervals that *end* at a specific location. This multi
05729   // map allows us to perform a linear search. We scan the instructions linearly
05730   // and record each time that a new interval starts, by placing it in a set.
05731   // If we find this value in the multi-map then we remove it from the set.
05732   // The max register usage is the maximum size of the set.
05733   // We also search for instructions that are defined outside the loop, but are
05734   // used inside the loop. We need this number separately from the max-interval
05735   // usage number because when we unroll, loop-invariant values do not take
05736   // more register.
05737   LoopBlocksDFS DFS(TheLoop);
05738   DFS.perform(LI);
05739 
05740   RegisterUsage R;
05741   R.NumInstructions = 0;
05742 
05743   // Each 'key' in the map opens a new interval. The values
05744   // of the map are the index of the 'last seen' usage of the
05745   // instruction that is the key.
05746   typedef DenseMap<Instruction*, unsigned> IntervalMap;
05747   // Maps instruction to its index.
05748   DenseMap<unsigned, Instruction*> IdxToInstr;
05749   // Marks the end of each interval.
05750   IntervalMap EndPoint;
05751   // Saves the list of instruction indices that are used in the loop.
05752   SmallSet<Instruction*, 8> Ends;
05753   // Saves the list of values that are used in the loop but are
05754   // defined outside the loop, such as arguments and constants.
05755   SmallPtrSet<Value*, 8> LoopInvariants;
05756 
05757   unsigned Index = 0;
05758   for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
05759        be = DFS.endRPO(); bb != be; ++bb) {
05760     R.NumInstructions += (*bb)->size();
05761     for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
05762          ++it) {
05763       Instruction *I = it;
05764       IdxToInstr[Index++] = I;
05765 
05766       // Save the end location of each USE.
05767       for (unsigned i = 0; i < I->getNumOperands(); ++i) {
05768         Value *U = I->getOperand(i);
05769         Instruction *Instr = dyn_cast<Instruction>(U);
05770 
05771         // Ignore non-instruction values such as arguments, constants, etc.
05772         if (!Instr) continue;
05773 
05774         // If this instruction is outside the loop then record it and continue.
05775         if (!TheLoop->contains(Instr)) {
05776           LoopInvariants.insert(Instr);
05777           continue;
05778         }
05779 
05780         // Overwrite previous end points.
05781         EndPoint[Instr] = Index;
05782         Ends.insert(Instr);
05783       }
05784     }
05785   }
05786 
05787   // Saves the list of intervals that end with the index in 'key'.
05788   typedef SmallVector<Instruction*, 2> InstrList;
05789   DenseMap<unsigned, InstrList> TransposeEnds;
05790 
05791   // Transpose the EndPoints to a list of values that end at each index.
05792   for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
05793        it != e; ++it)
05794     TransposeEnds[it->second].push_back(it->first);
05795 
05796   SmallSet<Instruction*, 8> OpenIntervals;
05797   unsigned MaxUsage = 0;
05798 
05799 
05800   DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
05801   for (unsigned int i = 0; i < Index; ++i) {
05802     Instruction *I = IdxToInstr[i];
05803     // Ignore instructions that are never used within the loop.
05804     if (!Ends.count(I)) continue;
05805 
05806     // Ignore ephemeral values.
05807     if (EphValues.count(I))
05808       continue;
05809 
05810     // Remove all of the instructions that end at this location.
05811     InstrList &List = TransposeEnds[i];
05812     for (unsigned int j=0, e = List.size(); j < e; ++j)
05813       OpenIntervals.erase(List[j]);
05814 
05815     // Count the number of live interals.
05816     MaxUsage = std::max(MaxUsage, OpenIntervals.size());
05817 
05818     DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
05819           OpenIntervals.size() << '\n');
05820 
05821     // Add the current instruction to the list of open intervals.
05822     OpenIntervals.insert(I);
05823   }
05824 
05825   unsigned Invariant = LoopInvariants.size();
05826   DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
05827   DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
05828   DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
05829 
05830   R.LoopInvariantRegs = Invariant;
05831   R.MaxLocalUsers = MaxUsage;
05832   return R;
05833 }
05834 
05835 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
05836   unsigned Cost = 0;
05837 
05838   // For each block.
05839   for (Loop::block_iterator bb = TheLoop->block_begin(),
05840        be = TheLoop->block_end(); bb != be; ++bb) {
05841     unsigned BlockCost = 0;
05842     BasicBlock *BB = *bb;
05843 
05844     // For each instruction in the old loop.
05845     for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
05846       // Skip dbg intrinsics.
05847       if (isa<DbgInfoIntrinsic>(it))
05848         continue;
05849 
05850       // Ignore ephemeral values.
05851       if (EphValues.count(it))
05852         continue;
05853 
05854       unsigned C = getInstructionCost(it, VF);
05855 
05856       // Check if we should override the cost.
05857       if (ForceTargetInstructionCost.getNumOccurrences() > 0)
05858         C = ForceTargetInstructionCost;
05859 
05860       BlockCost += C;
05861       DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
05862             VF << " For instruction: " << *it << '\n');
05863     }
05864 
05865     // We assume that if-converted blocks have a 50% chance of being executed.
05866     // When the code is scalar then some of the blocks are avoided due to CF.
05867     // When the code is vectorized we execute all code paths.
05868     if (VF == 1 && Legal->blockNeedsPredication(*bb))
05869       BlockCost /= 2;
05870 
05871     Cost += BlockCost;
05872   }
05873 
05874   return Cost;
05875 }
05876 
05877 /// \brief Check whether the address computation for a non-consecutive memory
05878 /// access looks like an unlikely candidate for being merged into the indexing
05879 /// mode.
05880 ///
05881 /// We look for a GEP which has one index that is an induction variable and all
05882 /// other indices are loop invariant. If the stride of this access is also
05883 /// within a small bound we decide that this address computation can likely be
05884 /// merged into the addressing mode.
05885 /// In all other cases, we identify the address computation as complex.
05886 static bool isLikelyComplexAddressComputation(Value *Ptr,
05887                                               LoopVectorizationLegality *Legal,
05888                                               ScalarEvolution *SE,
05889                                               const Loop *TheLoop) {
05890   GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
05891   if (!Gep)
05892     return true;
05893 
05894   // We are looking for a gep with all loop invariant indices except for one
05895   // which should be an induction variable.
05896   unsigned NumOperands = Gep->getNumOperands();
05897   for (unsigned i = 1; i < NumOperands; ++i) {
05898     Value *Opd = Gep->getOperand(i);
05899     if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
05900         !Legal->isInductionVariable(Opd))
05901       return true;
05902   }
05903 
05904   // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
05905   // can likely be merged into the address computation.
05906   unsigned MaxMergeDistance = 64;
05907 
05908   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
05909   if (!AddRec)
05910     return true;
05911 
05912   // Check the step is constant.
05913   const SCEV *Step = AddRec->getStepRecurrence(*SE);
05914   // Calculate the pointer stride and check if it is consecutive.
05915   const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
05916   if (!C)
05917     return true;
05918 
05919   const APInt &APStepVal = C->getValue()->getValue();
05920 
05921   // Huge step value - give up.
05922   if (APStepVal.getBitWidth() > 64)
05923     return true;
05924 
05925   int64_t StepVal = APStepVal.getSExtValue();
05926 
05927   return StepVal > MaxMergeDistance;
05928 }
05929 
05930 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
05931   if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
05932     return true;
05933   return false;
05934 }
05935 
05936 unsigned
05937 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
05938   // If we know that this instruction will remain uniform, check the cost of
05939   // the scalar version.
05940   if (Legal->isUniformAfterVectorization(I))
05941     VF = 1;
05942 
05943   Type *RetTy = I->getType();
05944   Type *VectorTy = ToVectorTy(RetTy, VF);
05945 
05946   // TODO: We need to estimate the cost of intrinsic calls.
05947   switch (I->getOpcode()) {
05948   case Instruction::GetElementPtr:
05949     // We mark this instruction as zero-cost because the cost of GEPs in
05950     // vectorized code depends on whether the corresponding memory instruction
05951     // is scalarized or not. Therefore, we handle GEPs with the memory
05952     // instruction cost.
05953     return 0;
05954   case Instruction::Br: {
05955     return TTI.getCFInstrCost(I->getOpcode());
05956   }
05957   case Instruction::PHI:
05958     //TODO: IF-converted IFs become selects.
05959     return 0;
05960   case Instruction::Add:
05961   case Instruction::FAdd:
05962   case Instruction::Sub:
05963   case Instruction::FSub:
05964   case Instruction::Mul:
05965   case Instruction::FMul:
05966   case Instruction::UDiv:
05967   case Instruction::SDiv:
05968   case Instruction::FDiv:
05969   case Instruction::URem:
05970   case Instruction::SRem:
05971   case Instruction::FRem:
05972   case Instruction::Shl:
05973   case Instruction::LShr:
05974   case Instruction::AShr:
05975   case Instruction::And:
05976   case Instruction::Or:
05977   case Instruction::Xor: {
05978     // Since we will replace the stride by 1 the multiplication should go away.
05979     if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
05980       return 0;
05981     // Certain instructions can be cheaper to vectorize if they have a constant
05982     // second vector operand. One example of this are shifts on x86.
05983     TargetTransformInfo::OperandValueKind Op1VK =
05984       TargetTransformInfo::OK_AnyValue;
05985     TargetTransformInfo::OperandValueKind Op2VK =
05986       TargetTransformInfo::OK_AnyValue;
05987     TargetTransformInfo::OperandValueProperties Op1VP =
05988         TargetTransformInfo::OP_None;
05989     TargetTransformInfo::OperandValueProperties Op2VP =
05990         TargetTransformInfo::OP_None;
05991     Value *Op2 = I->getOperand(1);
05992 
05993     // Check for a splat of a constant or for a non uniform vector of constants.
05994     if (isa<ConstantInt>(Op2)) {
05995       ConstantInt *CInt = cast<ConstantInt>(Op2);
05996       if (CInt && CInt->getValue().isPowerOf2())
05997         Op2VP = TargetTransformInfo::OP_PowerOf2;
05998       Op2VK = TargetTransformInfo::OK_UniformConstantValue;
05999     } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
06000       Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
06001       Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
06002       if (SplatValue) {
06003         ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
06004         if (CInt && CInt->getValue().isPowerOf2())
06005           Op2VP = TargetTransformInfo::OP_PowerOf2;
06006         Op2VK = TargetTransformInfo::OK_UniformConstantValue;
06007       }
06008     }
06009 
06010     return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
06011                                       Op1VP, Op2VP);
06012   }
06013   case Instruction::Select: {
06014     SelectInst *SI = cast<SelectInst>(I);
06015     const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
06016     bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
06017     Type *CondTy = SI->getCondition()->getType();
06018     if (!ScalarCond)
06019       CondTy = VectorType::get(CondTy, VF);
06020 
06021     return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
06022   }
06023   case Instruction::ICmp:
06024   case Instruction::FCmp: {
06025     Type *ValTy = I->getOperand(0)->getType();
06026     VectorTy = ToVectorTy(ValTy, VF);
06027     return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
06028   }
06029   case Instruction::Store:
06030   case Instruction::Load: {
06031     StoreInst *SI = dyn_cast<StoreInst>(I);
06032     LoadInst *LI = dyn_cast<LoadInst>(I);
06033     Type *ValTy = (SI ? SI->getValueOperand()->getType() :
06034                    LI->getType());
06035     VectorTy = ToVectorTy(ValTy, VF);
06036 
06037     unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
06038     unsigned AS = SI ? SI->getPointerAddressSpace() :
06039       LI->getPointerAddressSpace();
06040     Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
06041     // We add the cost of address computation here instead of with the gep
06042     // instruction because only here we know whether the operation is
06043     // scalarized.
06044     if (VF == 1)
06045       return TTI.getAddressComputationCost(VectorTy) +
06046         TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
06047 
06048     // Scalarized loads/stores.
06049     int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
06050     bool Reverse = ConsecutiveStride < 0;
06051     unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
06052     unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
06053     if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
06054       bool IsComplexComputation =
06055         isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
06056       unsigned Cost = 0;
06057       // The cost of extracting from the value vector and pointer vector.
06058       Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
06059       for (unsigned i = 0; i < VF; ++i) {
06060         //  The cost of extracting the pointer operand.
06061         Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
06062         // In case of STORE, the cost of ExtractElement from the vector.
06063         // In case of LOAD, the cost of InsertElement into the returned
06064         // vector.
06065         Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
06066                                             Instruction::InsertElement,
06067                                             VectorTy, i);
06068       }
06069 
06070       // The cost of the scalar loads/stores.
06071       Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
06072       Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
06073                                        Alignment, AS);
06074       return Cost;
06075     }
06076 
06077     // Wide load/stores.
06078     unsigned Cost = TTI.getAddressComputationCost(VectorTy);
06079     Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
06080 
06081     if (Reverse)
06082       Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
06083                                   VectorTy, 0);
06084     return Cost;
06085   }
06086   case Instruction::ZExt:
06087   case Instruction::SExt:
06088   case Instruction::FPToUI:
06089   case Instruction::FPToSI:
06090   case Instruction::FPExt:
06091   case Instruction::PtrToInt:
06092   case Instruction::IntToPtr:
06093   case Instruction::SIToFP:
06094   case Instruction::UIToFP:
06095   case Instruction::Trunc:
06096   case Instruction::FPTrunc:
06097   case Instruction::BitCast: {
06098     // We optimize the truncation of induction variable.
06099     // The cost of these is the same as the scalar operation.
06100     if (I->getOpcode() == Instruction::Trunc &&
06101         Legal->isInductionVariable(I->getOperand(0)))
06102       return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
06103                                   I->getOperand(0)->getType());
06104 
06105     Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
06106     return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
06107   }
06108   case Instruction::Call: {
06109     CallInst *CI = cast<CallInst>(I);
06110     Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
06111     assert(ID && "Not an intrinsic call!");
06112     Type *RetTy = ToVectorTy(CI->getType(), VF);
06113     SmallVector<Type*, 4> Tys;
06114     for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
06115       Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
06116     return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
06117   }
06118   default: {
06119     // We are scalarizing the instruction. Return the cost of the scalar
06120     // instruction, plus the cost of insert and extract into vector
06121     // elements, times the vector width.
06122     unsigned Cost = 0;
06123 
06124     if (!RetTy->isVoidTy() && VF != 1) {
06125       unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
06126                                                 VectorTy);
06127       unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
06128                                                 VectorTy);
06129 
06130       // The cost of inserting the results plus extracting each one of the
06131       // operands.
06132       Cost += VF * (InsCost + ExtCost * I->getNumOperands());
06133     }
06134 
06135     // The cost of executing VF copies of the scalar instruction. This opcode
06136     // is unknown. Assume that it is the same as 'mul'.
06137     Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
06138     return Cost;
06139   }
06140   }// end of switch.
06141 }
06142 
06143 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
06144   if (Scalar->isVoidTy() || VF == 1)
06145     return Scalar;
06146   return VectorType::get(Scalar, VF);
06147 }
06148 
06149 char LoopVectorize::ID = 0;
06150 static const char lv_name[] = "Loop Vectorization";
06151 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
06152 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
06153 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
06154 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
06155 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
06156 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
06157 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
06158 INITIALIZE_PASS_DEPENDENCY(LCSSA)
06159 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
06160 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
06161 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
06162 
06163 namespace llvm {
06164   Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
06165     return new LoopVectorize(NoUnrolling, AlwaysVectorize);
06166   }
06167 }
06168 
06169 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
06170   // Check for a store.
06171   if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
06172     return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
06173 
06174   // Check for a load.
06175   if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
06176     return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
06177 
06178   return false;
06179 }
06180 
06181 
06182 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
06183                                              bool IfPredicateStore) {
06184   assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
06185   // Holds vector parameters or scalars, in case of uniform vals.
06186   SmallVector<VectorParts, 4> Params;
06187 
06188   setDebugLocFromInst(Builder, Instr);
06189 
06190   // Find all of the vectorized parameters.
06191   for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
06192     Value *SrcOp = Instr->getOperand(op);
06193 
06194     // If we are accessing the old induction variable, use the new one.
06195     if (SrcOp == OldInduction) {
06196       Params.push_back(getVectorValue(SrcOp));
06197       continue;
06198     }
06199 
06200     // Try using previously calculated values.
06201     Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
06202 
06203     // If the src is an instruction that appeared earlier in the basic block
06204     // then it should already be vectorized.
06205     if (SrcInst && OrigLoop->contains(SrcInst)) {
06206       assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
06207       // The parameter is a vector value from earlier.
06208       Params.push_back(WidenMap.get(SrcInst));
06209     } else {
06210       // The parameter is a scalar from outside the loop. Maybe even a constant.
06211       VectorParts Scalars;
06212       Scalars.append(UF, SrcOp);
06213       Params.push_back(Scalars);
06214     }
06215   }
06216 
06217   assert(Params.size() == Instr->getNumOperands() &&
06218          "Invalid number of operands");
06219 
06220   // Does this instruction return a value ?
06221   bool IsVoidRetTy = Instr->getType()->isVoidTy();
06222 
06223   Value *UndefVec = IsVoidRetTy ? nullptr :
06224   UndefValue::get(Instr->getType());
06225   // Create a new entry in the WidenMap and initialize it to Undef or Null.
06226   VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
06227 
06228   Instruction *InsertPt = Builder.GetInsertPoint();
06229   BasicBlock *IfBlock = Builder.GetInsertBlock();
06230   BasicBlock *CondBlock = nullptr;
06231 
06232   VectorParts Cond;
06233   Loop *VectorLp = nullptr;
06234   if (IfPredicateStore) {
06235     assert(Instr->getParent()->getSinglePredecessor() &&
06236            "Only support single predecessor blocks");
06237     Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
06238                           Instr->getParent());
06239     VectorLp = LI->getLoopFor(IfBlock);
06240     assert(VectorLp && "Must have a loop for this block");
06241   }
06242 
06243   // For each vector unroll 'part':
06244   for (unsigned Part = 0; Part < UF; ++Part) {
06245     // For each scalar that we create:
06246 
06247     // Start an "if (pred) a[i] = ..." block.
06248     Value *Cmp = nullptr;
06249     if (IfPredicateStore) {
06250       if (Cond[Part]->getType()->isVectorTy())
06251         Cond[Part] =
06252             Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
06253       Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
06254                                ConstantInt::get(Cond[Part]->getType(), 1));
06255       CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
06256       LoopVectorBody.push_back(CondBlock);
06257       VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
06258       // Update Builder with newly created basic block.
06259       Builder.SetInsertPoint(InsertPt);
06260     }
06261 
06262     Instruction *Cloned = Instr->clone();
06263       if (!IsVoidRetTy)
06264         Cloned->setName(Instr->getName() + ".cloned");
06265       // Replace the operands of the cloned instructions with extracted scalars.
06266       for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
06267         Value *Op = Params[op][Part];
06268         Cloned->setOperand(op, Op);
06269       }
06270 
06271       // Place the cloned scalar in the new loop.
06272       Builder.Insert(Cloned);
06273 
06274       // If the original scalar returns a value we need to place it in a vector
06275       // so that future users will be able to use it.
06276       if (!IsVoidRetTy)
06277         VecResults[Part] = Cloned;
06278 
06279     // End if-block.
06280       if (IfPredicateStore) {
06281         BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
06282         LoopVectorBody.push_back(NewIfBlock);
06283         VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
06284         Builder.SetInsertPoint(InsertPt);
06285         Instruction *OldBr = IfBlock->getTerminator();
06286         BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
06287         OldBr->eraseFromParent();
06288         IfBlock = NewIfBlock;
06289       }
06290   }
06291 }
06292 
06293 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
06294   StoreInst *SI = dyn_cast<StoreInst>(Instr);
06295   bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
06296 
06297   return scalarizeInstruction(Instr, IfPredicateStore);
06298 }
06299 
06300 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
06301   return Vec;
06302 }
06303 
06304 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
06305   return V;
06306 }
06307 
06308 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
06309                                                bool Negate) {
06310   // When unrolling and the VF is 1, we only need to add a simple scalar.
06311   Type *ITy = Val->getType();
06312   assert(!ITy->isVectorTy() && "Val must be a scalar");
06313   Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
06314   return Builder.CreateAdd(Val, C, "induction");
06315 }