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/AssumptionCache.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   /// Returns true if the target machine supports masked store operation
00773   /// for the given \p DataType and kind of access to \p Ptr.
00774   bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
00775     return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
00776   }
00777   /// Returns true if the target machine supports masked load operation
00778   /// for the given \p DataType and kind of access to \p Ptr.
00779   bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
00780     return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
00781   }
00782   /// Returns true if vector representation of the instruction \p I
00783   /// requires mask.
00784   bool isMaskRequired(const Instruction* I) {
00785     return (MaskedOp.count(I) != 0);
00786   }
00787 private:
00788   /// Check if a single basic block loop is vectorizable.
00789   /// At this point we know that this is a loop with a constant trip count
00790   /// and we only need to check individual instructions.
00791   bool canVectorizeInstrs();
00792 
00793   /// When we vectorize loops we may change the order in which
00794   /// we read and write from memory. This method checks if it is
00795   /// legal to vectorize the code, considering only memory constrains.
00796   /// Returns true if the loop is vectorizable
00797   bool canVectorizeMemory();
00798 
00799   /// Return true if we can vectorize this loop using the IF-conversion
00800   /// transformation.
00801   bool canVectorizeWithIfConvert();
00802 
00803   /// Collect the variables that need to stay uniform after vectorization.
00804   void collectLoopUniforms();
00805 
00806   /// Return true if all of the instructions in the block can be speculatively
00807   /// executed. \p SafePtrs is a list of addresses that are known to be legal
00808   /// and we know that we can read from them without segfault.
00809   bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
00810 
00811   /// Returns True, if 'Phi' is the kind of reduction variable for type
00812   /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
00813   bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
00814   /// Returns a struct describing if the instruction 'I' can be a reduction
00815   /// variable of type 'Kind'. If the reduction is a min/max pattern of
00816   /// select(icmp()) this function advances the instruction pointer 'I' from the
00817   /// compare instruction to the select instruction and stores this pointer in
00818   /// 'PatternLastInst' member of the returned struct.
00819   ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
00820                                      ReductionInstDesc &Desc);
00821   /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
00822   /// pattern corresponding to a min(X, Y) or max(X, Y).
00823   static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
00824                                                     ReductionInstDesc &Prev);
00825   /// Returns the induction kind of Phi. This function may return NoInduction
00826   /// if the PHI is not an induction variable.
00827   InductionKind isInductionVariable(PHINode *Phi);
00828 
00829   /// \brief Collect memory access with loop invariant strides.
00830   ///
00831   /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
00832   /// invariant.
00833   void collectStridedAccess(Value *LoadOrStoreInst);
00834 
00835   /// Report an analysis message to assist the user in diagnosing loops that are
00836   /// not vectorized.
00837   void emitAnalysis(Report &Message) {
00838     DebugLoc DL = TheLoop->getStartLoc();
00839     if (Instruction *I = Message.getInstr())
00840       DL = I->getDebugLoc();
00841     emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
00842                                    *TheFunction, DL, Message.str());
00843   }
00844 
00845   /// The loop that we evaluate.
00846   Loop *TheLoop;
00847   /// Scev analysis.
00848   ScalarEvolution *SE;
00849   /// DataLayout analysis.
00850   const DataLayout *DL;
00851   /// Dominators.
00852   DominatorTree *DT;
00853   /// Target Library Info.
00854   TargetLibraryInfo *TLI;
00855   /// Alias analysis.
00856   AliasAnalysis *AA;
00857   /// Parent function
00858   Function *TheFunction;
00859   /// Target Transform Info
00860   const TargetTransformInfo *TTI;
00861 
00862   //  ---  vectorization state --- //
00863 
00864   /// Holds the integer induction variable. This is the counter of the
00865   /// loop.
00866   PHINode *Induction;
00867   /// Holds the reduction variables.
00868   ReductionList Reductions;
00869   /// Holds all of the induction variables that we found in the loop.
00870   /// Notice that inductions don't need to start at zero and that induction
00871   /// variables can be pointers.
00872   InductionList Inductions;
00873   /// Holds the widest induction type encountered.
00874   Type *WidestIndTy;
00875 
00876   /// Allowed outside users. This holds the reduction
00877   /// vars which can be accessed from outside the loop.
00878   SmallPtrSet<Value*, 4> AllowedExit;
00879   /// This set holds the variables which are known to be uniform after
00880   /// vectorization.
00881   SmallPtrSet<Instruction*, 4> Uniforms;
00882   /// We need to check that all of the pointers in this list are disjoint
00883   /// at runtime.
00884   RuntimePointerCheck PtrRtCheck;
00885   /// Can we assume the absence of NaNs.
00886   bool HasFunNoNaNAttr;
00887 
00888   unsigned MaxSafeDepDistBytes;
00889 
00890   ValueToValueMap Strides;
00891   SmallPtrSet<Value *, 8> StrideSet;
00892   
00893   /// While vectorizing these instructions we have to generate a
00894   /// call to the appropriate masked intrinsic
00895   SmallPtrSet<const Instruction*, 8> MaskedOp;
00896 };
00897 
00898 /// LoopVectorizationCostModel - estimates the expected speedups due to
00899 /// vectorization.
00900 /// In many cases vectorization is not profitable. This can happen because of
00901 /// a number of reasons. In this class we mainly attempt to predict the
00902 /// expected speedup/slowdowns due to the supported instruction set. We use the
00903 /// TargetTransformInfo to query the different backends for the cost of
00904 /// different operations.
00905 class LoopVectorizationCostModel {
00906 public:
00907   LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
00908                              LoopVectorizationLegality *Legal,
00909                              const TargetTransformInfo &TTI,
00910                              const DataLayout *DL, const TargetLibraryInfo *TLI,
00911                              AssumptionCache *AC, const Function *F,
00912                              const LoopVectorizeHints *Hints)
00913       : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
00914         TheFunction(F), Hints(Hints) {
00915     CodeMetrics::collectEphemeralValues(L, AC, EphValues);
00916   }
00917 
00918   /// Information about vectorization costs
00919   struct VectorizationFactor {
00920     unsigned Width; // Vector width with best cost
00921     unsigned Cost; // Cost of the loop with that width
00922   };
00923   /// \return The most profitable vectorization factor and the cost of that VF.
00924   /// This method checks every power of two up to VF. If UserVF is not ZERO
00925   /// then this vectorization factor will be selected if vectorization is
00926   /// possible.
00927   VectorizationFactor selectVectorizationFactor(bool OptForSize);
00928 
00929   /// \return The size (in bits) of the widest type in the code that
00930   /// needs to be vectorized. We ignore values that remain scalar such as
00931   /// 64 bit loop indices.
00932   unsigned getWidestType();
00933 
00934   /// \return The most profitable unroll factor.
00935   /// If UserUF is non-zero then this method finds the best unroll-factor
00936   /// based on register pressure and other parameters.
00937   /// VF and LoopCost are the selected vectorization factor and the cost of the
00938   /// selected VF.
00939   unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
00940 
00941   /// \brief A struct that represents some properties of the register usage
00942   /// of a loop.
00943   struct RegisterUsage {
00944     /// Holds the number of loop invariant values that are used in the loop.
00945     unsigned LoopInvariantRegs;
00946     /// Holds the maximum number of concurrent live intervals in the loop.
00947     unsigned MaxLocalUsers;
00948     /// Holds the number of instructions in the loop.
00949     unsigned NumInstructions;
00950   };
00951 
00952   /// \return  information about the register usage of the loop.
00953   RegisterUsage calculateRegisterUsage();
00954 
00955 private:
00956   /// Returns the expected execution cost. The unit of the cost does
00957   /// not matter because we use the 'cost' units to compare different
00958   /// vector widths. The cost that is returned is *not* normalized by
00959   /// the factor width.
00960   unsigned expectedCost(unsigned VF);
00961 
00962   /// Returns the execution time cost of an instruction for a given vector
00963   /// width. Vector width of one means scalar.
00964   unsigned getInstructionCost(Instruction *I, unsigned VF);
00965 
00966   /// A helper function for converting Scalar types to vector types.
00967   /// If the incoming type is void, we return void. If the VF is 1, we return
00968   /// the scalar type.
00969   static Type* ToVectorTy(Type *Scalar, unsigned VF);
00970 
00971   /// Returns whether the instruction is a load or store and will be a emitted
00972   /// as a vector operation.
00973   bool isConsecutiveLoadOrStore(Instruction *I);
00974 
00975   /// Report an analysis message to assist the user in diagnosing loops that are
00976   /// not vectorized.
00977   void emitAnalysis(Report &Message) {
00978     DebugLoc DL = TheLoop->getStartLoc();
00979     if (Instruction *I = Message.getInstr())
00980       DL = I->getDebugLoc();
00981     emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
00982                                    *TheFunction, DL, Message.str());
00983   }
00984 
00985   /// Values used only by @llvm.assume calls.
00986   SmallPtrSet<const Value *, 32> EphValues;
00987 
00988   /// The loop that we evaluate.
00989   Loop *TheLoop;
00990   /// Scev analysis.
00991   ScalarEvolution *SE;
00992   /// Loop Info analysis.
00993   LoopInfo *LI;
00994   /// Vectorization legality.
00995   LoopVectorizationLegality *Legal;
00996   /// Vector target information.
00997   const TargetTransformInfo &TTI;
00998   /// Target data layout information.
00999   const DataLayout *DL;
01000   /// Target Library Info.
01001   const TargetLibraryInfo *TLI;
01002   const Function *TheFunction;
01003   // Loop Vectorize Hint.
01004   const LoopVectorizeHints *Hints;
01005 };
01006 
01007 /// Utility class for getting and setting loop vectorizer hints in the form
01008 /// of loop metadata.
01009 /// This class keeps a number of loop annotations locally (as member variables)
01010 /// and can, upon request, write them back as metadata on the loop. It will
01011 /// initially scan the loop for existing metadata, and will update the local
01012 /// values based on information in the loop.
01013 /// We cannot write all values to metadata, as the mere presence of some info,
01014 /// for example 'force', means a decision has been made. So, we need to be
01015 /// careful NOT to add them if the user hasn't specifically asked so.
01016 class LoopVectorizeHints {
01017   enum HintKind {
01018     HK_WIDTH,
01019     HK_UNROLL,
01020     HK_FORCE
01021   };
01022 
01023   /// Hint - associates name and validation with the hint value.
01024   struct Hint {
01025     const char * Name;
01026     unsigned Value; // This may have to change for non-numeric values.
01027     HintKind Kind;
01028 
01029     Hint(const char * Name, unsigned Value, HintKind Kind)
01030       : Name(Name), Value(Value), Kind(Kind) { }
01031 
01032     bool validate(unsigned Val) {
01033       switch (Kind) {
01034       case HK_WIDTH:
01035         return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
01036       case HK_UNROLL:
01037         return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
01038       case HK_FORCE:
01039         return (Val <= 1);
01040       }
01041       return false;
01042     }
01043   };
01044 
01045   /// Vectorization width.
01046   Hint Width;
01047   /// Vectorization interleave factor.
01048   Hint Interleave;
01049   /// Vectorization forced
01050   Hint Force;
01051 
01052   /// Return the loop metadata prefix.
01053   static StringRef Prefix() { return "llvm.loop."; }
01054 
01055 public:
01056   enum ForceKind {
01057     FK_Undefined = -1, ///< Not selected.
01058     FK_Disabled = 0,   ///< Forcing disabled.
01059     FK_Enabled = 1,    ///< Forcing enabled.
01060   };
01061 
01062   LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
01063       : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
01064         Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
01065         Force("vectorize.enable", FK_Undefined, HK_FORCE),
01066         TheLoop(L) {
01067     // Populate values with existing loop metadata.
01068     getHintsFromMetadata();
01069 
01070     // force-vector-interleave overrides DisableInterleaving.
01071     if (VectorizationInterleave.getNumOccurrences() > 0)
01072       Interleave.Value = VectorizationInterleave;
01073 
01074     DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
01075           << "LV: Interleaving disabled by the pass manager\n");
01076   }
01077 
01078   /// Mark the loop L as already vectorized by setting the width to 1.
01079   void setAlreadyVectorized() {
01080     Width.Value = Interleave.Value = 1;
01081     Hint Hints[] = {Width, Interleave};
01082     writeHintsToMetadata(Hints);
01083   }
01084 
01085   /// Dumps all the hint information.
01086   std::string emitRemark() const {
01087     Report R;
01088     if (Force.Value == LoopVectorizeHints::FK_Disabled)
01089       R << "vectorization is explicitly disabled";
01090     else {
01091       R << "use -Rpass-analysis=loop-vectorize for more info";
01092       if (Force.Value == LoopVectorizeHints::FK_Enabled) {
01093         R << " (Force=true";
01094         if (Width.Value != 0)
01095           R << ", Vector Width=" << Width.Value;
01096         if (Interleave.Value != 0)
01097           R << ", Interleave Count=" << Interleave.Value;
01098         R << ")";
01099       }
01100     }
01101 
01102     return R.str();
01103   }
01104 
01105   unsigned getWidth() const { return Width.Value; }
01106   unsigned getInterleave() const { return Interleave.Value; }
01107   enum ForceKind getForce() const { return (ForceKind)Force.Value; }
01108 
01109 private:
01110   /// Find hints specified in the loop metadata and update local values.
01111   void getHintsFromMetadata() {
01112     MDNode *LoopID = TheLoop->getLoopID();
01113     if (!LoopID)
01114       return;
01115 
01116     // First operand should refer to the loop id itself.
01117     assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
01118     assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
01119 
01120     for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
01121       const MDString *S = nullptr;
01122       SmallVector<Metadata *, 4> Args;
01123 
01124       // The expected hint is either a MDString or a MDNode with the first
01125       // operand a MDString.
01126       if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
01127         if (!MD || MD->getNumOperands() == 0)
01128           continue;
01129         S = dyn_cast<MDString>(MD->getOperand(0));
01130         for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
01131           Args.push_back(MD->getOperand(i));
01132       } else {
01133         S = dyn_cast<MDString>(LoopID->getOperand(i));
01134         assert(Args.size() == 0 && "too many arguments for MDString");
01135       }
01136 
01137       if (!S)
01138         continue;
01139 
01140       // Check if the hint starts with the loop metadata prefix.
01141       StringRef Name = S->getString();
01142       if (Args.size() == 1)
01143         setHint(Name, Args[0]);
01144     }
01145   }
01146 
01147   /// Checks string hint with one operand and set value if valid.
01148   void setHint(StringRef Name, Metadata *Arg) {
01149     if (!Name.startswith(Prefix()))
01150       return;
01151     Name = Name.substr(Prefix().size(), StringRef::npos);
01152 
01153     const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
01154     if (!C) return;
01155     unsigned Val = C->getZExtValue();
01156 
01157     Hint *Hints[] = {&Width, &Interleave, &Force};
01158     for (auto H : Hints) {
01159       if (Name == H->Name) {
01160         if (H->validate(Val))
01161           H->Value = Val;
01162         else
01163           DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
01164         break;
01165       }
01166     }
01167   }
01168 
01169   /// Create a new hint from name / value pair.
01170   MDNode *createHintMetadata(StringRef Name, unsigned V) const {
01171     LLVMContext &Context = TheLoop->getHeader()->getContext();
01172     Metadata *MDs[] = {MDString::get(Context, Name),
01173                        ConstantAsMetadata::get(
01174                            ConstantInt::get(Type::getInt32Ty(Context), V))};
01175     return MDNode::get(Context, MDs);
01176   }
01177 
01178   /// Matches metadata with hint name.
01179   bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
01180     MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
01181     if (!Name)
01182       return false;
01183 
01184     for (auto H : HintTypes)
01185       if (Name->getString().endswith(H.Name))
01186         return true;
01187     return false;
01188   }
01189 
01190   /// Sets current hints into loop metadata, keeping other values intact.
01191   void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
01192     if (HintTypes.size() == 0)
01193       return;
01194 
01195     // Reserve the first element to LoopID (see below).
01196     SmallVector<Metadata *, 4> MDs(1);
01197     // If the loop already has metadata, then ignore the existing operands.
01198     MDNode *LoopID = TheLoop->getLoopID();
01199     if (LoopID) {
01200       for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
01201         MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
01202         // If node in update list, ignore old value.
01203         if (!matchesHintMetadataName(Node, HintTypes))
01204           MDs.push_back(Node);
01205       }
01206     }
01207 
01208     // Now, add the missing hints.
01209     for (auto H : HintTypes)
01210       MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
01211 
01212     // Replace current metadata node with new one.
01213     LLVMContext &Context = TheLoop->getHeader()->getContext();
01214     MDNode *NewLoopID = MDNode::get(Context, MDs);
01215     // Set operand 0 to refer to the loop id itself.
01216     NewLoopID->replaceOperandWith(0, NewLoopID);
01217 
01218     TheLoop->setLoopID(NewLoopID);
01219   }
01220 
01221   /// The loop these hints belong to.
01222   const Loop *TheLoop;
01223 };
01224 
01225 static void emitMissedWarning(Function *F, Loop *L,
01226                               const LoopVectorizeHints &LH) {
01227   emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
01228                                L->getStartLoc(), LH.emitRemark());
01229 
01230   if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
01231     if (LH.getWidth() != 1)
01232       emitLoopVectorizeWarning(
01233           F->getContext(), *F, L->getStartLoc(),
01234           "failed explicitly specified loop vectorization");
01235     else if (LH.getInterleave() != 1)
01236       emitLoopInterleaveWarning(
01237           F->getContext(), *F, L->getStartLoc(),
01238           "failed explicitly specified loop interleaving");
01239   }
01240 }
01241 
01242 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
01243   if (L.empty())
01244     return V.push_back(&L);
01245 
01246   for (Loop *InnerL : L)
01247     addInnerLoop(*InnerL, V);
01248 }
01249 
01250 /// The LoopVectorize Pass.
01251 struct LoopVectorize : public FunctionPass {
01252   /// Pass identification, replacement for typeid
01253   static char ID;
01254 
01255   explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
01256     : FunctionPass(ID),
01257       DisableUnrolling(NoUnrolling),
01258       AlwaysVectorize(AlwaysVectorize) {
01259     initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
01260   }
01261 
01262   ScalarEvolution *SE;
01263   const DataLayout *DL;
01264   LoopInfo *LI;
01265   TargetTransformInfo *TTI;
01266   DominatorTree *DT;
01267   BlockFrequencyInfo *BFI;
01268   TargetLibraryInfo *TLI;
01269   AliasAnalysis *AA;
01270   AssumptionCache *AC;
01271   bool DisableUnrolling;
01272   bool AlwaysVectorize;
01273 
01274   BlockFrequency ColdEntryFreq;
01275 
01276   bool runOnFunction(Function &F) override {
01277     SE = &getAnalysis<ScalarEvolution>();
01278     DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
01279     DL = DLP ? &DLP->getDataLayout() : nullptr;
01280     LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
01281     TTI = &getAnalysis<TargetTransformInfo>();
01282     DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
01283     BFI = &getAnalysis<BlockFrequencyInfo>();
01284     auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
01285     TLI = TLIP ? &TLIP->getTLI() : nullptr;
01286     AA = &getAnalysis<AliasAnalysis>();
01287     AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
01288 
01289     // Compute some weights outside of the loop over the loops. Compute this
01290     // using a BranchProbability to re-use its scaling math.
01291     const BranchProbability ColdProb(1, 5); // 20%
01292     ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
01293 
01294     // If the target claims to have no vector registers don't attempt
01295     // vectorization.
01296     if (!TTI->getNumberOfRegisters(true))
01297       return false;
01298 
01299     if (!DL) {
01300       DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
01301                    << ": Missing data layout\n");
01302       return false;
01303     }
01304 
01305     // Build up a worklist of inner-loops to vectorize. This is necessary as
01306     // the act of vectorizing or partially unrolling a loop creates new loops
01307     // and can invalidate iterators across the loops.
01308     SmallVector<Loop *, 8> Worklist;
01309 
01310     for (Loop *L : *LI)
01311       addInnerLoop(*L, Worklist);
01312 
01313     LoopsAnalyzed += Worklist.size();
01314 
01315     // Now walk the identified inner loops.
01316     bool Changed = false;
01317     while (!Worklist.empty())
01318       Changed |= processLoop(Worklist.pop_back_val());
01319 
01320     // Process each loop nest in the function.
01321     return Changed;
01322   }
01323 
01324   bool processLoop(Loop *L) {
01325     assert(L->empty() && "Only process inner loops.");
01326 
01327 #ifndef NDEBUG
01328     const std::string DebugLocStr = getDebugLocString(L);
01329 #endif /* NDEBUG */
01330 
01331     DEBUG(dbgs() << "\nLV: Checking a loop in \""
01332                  << L->getHeader()->getParent()->getName() << "\" from "
01333                  << DebugLocStr << "\n");
01334 
01335     LoopVectorizeHints Hints(L, DisableUnrolling);
01336 
01337     DEBUG(dbgs() << "LV: Loop hints:"
01338                  << " force="
01339                  << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
01340                          ? "disabled"
01341                          : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
01342                                 ? "enabled"
01343                                 : "?")) << " width=" << Hints.getWidth()
01344                  << " unroll=" << Hints.getInterleave() << "\n");
01345 
01346     // Function containing loop
01347     Function *F = L->getHeader()->getParent();
01348 
01349     // Looking at the diagnostic output is the only way to determine if a loop
01350     // was vectorized (other than looking at the IR or machine code), so it
01351     // is important to generate an optimization remark for each loop. Most of
01352     // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
01353     // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
01354     // less verbose reporting vectorized loops and unvectorized loops that may
01355     // benefit from vectorization, respectively.
01356 
01357     if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
01358       DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
01359       emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
01360                                      L->getStartLoc(), Hints.emitRemark());
01361       return false;
01362     }
01363 
01364     if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
01365       DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
01366       emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
01367                                      L->getStartLoc(), Hints.emitRemark());
01368       return false;
01369     }
01370 
01371     if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
01372       DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
01373       emitOptimizationRemarkAnalysis(
01374           F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01375           "loop not vectorized: vector width and interleave count are "
01376           "explicitly set to 1");
01377       return false;
01378     }
01379 
01380     // Check the loop for a trip count threshold:
01381     // do not vectorize loops with a tiny trip count.
01382     const unsigned TC = SE->getSmallConstantTripCount(L);
01383     if (TC > 0u && TC < TinyTripCountVectorThreshold) {
01384       DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
01385                    << "This loop is not worth vectorizing.");
01386       if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
01387         DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
01388       else {
01389         DEBUG(dbgs() << "\n");
01390         emitOptimizationRemarkAnalysis(
01391             F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01392             "vectorization is not beneficial and is not explicitly forced");
01393         return false;
01394       }
01395     }
01396 
01397     // Check if it is legal to vectorize the loop.
01398     LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI);
01399     if (!LVL.canVectorize()) {
01400       DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
01401       emitMissedWarning(F, L, Hints);
01402       return false;
01403     }
01404 
01405     // Use the cost model.
01406     LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
01407                                   &Hints);
01408 
01409     // Check the function attributes to find out if this function should be
01410     // optimized for size.
01411     bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
01412                       F->hasFnAttribute(Attribute::OptimizeForSize);
01413 
01414     // Compute the weighted frequency of this loop being executed and see if it
01415     // is less than 20% of the function entry baseline frequency. Note that we
01416     // always have a canonical loop here because we think we *can* vectoriez.
01417     // FIXME: This is hidden behind a flag due to pervasive problems with
01418     // exactly what block frequency models.
01419     if (LoopVectorizeWithBlockFrequency) {
01420       BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
01421       if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
01422           LoopEntryFreq < ColdEntryFreq)
01423         OptForSize = true;
01424     }
01425 
01426     // Check the function attributes to see if implicit floats are allowed.a
01427     // FIXME: This check doesn't seem possibly correct -- what if the loop is
01428     // an integer loop and the vector instructions selected are purely integer
01429     // vector instructions?
01430     if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
01431       DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
01432             "attribute is used.\n");
01433       emitOptimizationRemarkAnalysis(
01434           F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01435           "loop not vectorized due to NoImplicitFloat attribute");
01436       emitMissedWarning(F, L, Hints);
01437       return false;
01438     }
01439 
01440     // Select the optimal vectorization factor.
01441     const LoopVectorizationCostModel::VectorizationFactor VF =
01442         CM.selectVectorizationFactor(OptForSize);
01443 
01444     // Select the unroll factor.
01445     const unsigned UF =
01446         CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
01447 
01448     DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
01449                  << DebugLocStr << '\n');
01450     DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
01451 
01452     if (VF.Width == 1) {
01453       DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
01454 
01455       if (UF == 1) {
01456         emitOptimizationRemarkAnalysis(
01457             F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01458             "not beneficial to vectorize and user disabled interleaving");
01459         return false;
01460       }
01461       DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
01462 
01463       // Report the unrolling decision.
01464       emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01465                              Twine("unrolled with interleaving factor " +
01466                                    Twine(UF) +
01467                                    " (vectorization not beneficial)"));
01468 
01469       // We decided not to vectorize, but we may want to unroll.
01470 
01471       InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
01472       Unroller.vectorize(&LVL);
01473     } else {
01474       // If we decided that it is *legal* to vectorize the loop then do it.
01475       InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
01476       LB.vectorize(&LVL);
01477       ++LoopsVectorized;
01478 
01479       // Report the vectorization decision.
01480       emitOptimizationRemark(
01481           F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
01482           Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
01483               ", unrolling interleave factor: " + Twine(UF) + ")");
01484     }
01485 
01486     // Mark the loop as already vectorized to avoid vectorizing again.
01487     Hints.setAlreadyVectorized();
01488 
01489     DEBUG(verifyFunction(*L->getHeader()->getParent()));
01490     return true;
01491   }
01492 
01493   void getAnalysisUsage(AnalysisUsage &AU) const override {
01494     AU.addRequired<AssumptionCacheTracker>();
01495     AU.addRequiredID(LoopSimplifyID);
01496     AU.addRequiredID(LCSSAID);
01497     AU.addRequired<BlockFrequencyInfo>();
01498     AU.addRequired<DominatorTreeWrapperPass>();
01499     AU.addRequired<LoopInfoWrapperPass>();
01500     AU.addRequired<ScalarEvolution>();
01501     AU.addRequired<TargetTransformInfo>();
01502     AU.addRequired<AliasAnalysis>();
01503     AU.addPreserved<LoopInfoWrapperPass>();
01504     AU.addPreserved<DominatorTreeWrapperPass>();
01505     AU.addPreserved<AliasAnalysis>();
01506   }
01507 
01508 };
01509 
01510 } // end anonymous namespace
01511 
01512 //===----------------------------------------------------------------------===//
01513 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
01514 // LoopVectorizationCostModel.
01515 //===----------------------------------------------------------------------===//
01516 
01517 static Value *stripIntegerCast(Value *V) {
01518   if (CastInst *CI = dyn_cast<CastInst>(V))
01519     if (CI->getOperand(0)->getType()->isIntegerTy())
01520       return CI->getOperand(0);
01521   return V;
01522 }
01523 
01524 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
01525 ///
01526 /// If \p OrigPtr is not null, use it to look up the stride value instead of
01527 /// \p Ptr.
01528 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
01529                                              ValueToValueMap &PtrToStride,
01530                                              Value *Ptr, Value *OrigPtr = nullptr) {
01531 
01532   const SCEV *OrigSCEV = SE->getSCEV(Ptr);
01533 
01534   // If there is an entry in the map return the SCEV of the pointer with the
01535   // symbolic stride replaced by one.
01536   ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
01537   if (SI != PtrToStride.end()) {
01538     Value *StrideVal = SI->second;
01539 
01540     // Strip casts.
01541     StrideVal = stripIntegerCast(StrideVal);
01542 
01543     // Replace symbolic stride by one.
01544     Value *One = ConstantInt::get(StrideVal->getType(), 1);
01545     ValueToValueMap RewriteMap;
01546     RewriteMap[StrideVal] = One;
01547 
01548     const SCEV *ByOne =
01549         SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
01550     DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
01551                  << "\n");
01552     return ByOne;
01553   }
01554 
01555   // Otherwise, just return the SCEV of the original pointer.
01556   return SE->getSCEV(Ptr);
01557 }
01558 
01559 void LoopVectorizationLegality::RuntimePointerCheck::insert(
01560     ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
01561     unsigned ASId, ValueToValueMap &Strides) {
01562   // Get the stride replaced scev.
01563   const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
01564   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
01565   assert(AR && "Invalid addrec expression");
01566   const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
01567   const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
01568   Pointers.push_back(Ptr);
01569   Starts.push_back(AR->getStart());
01570   Ends.push_back(ScEnd);
01571   IsWritePtr.push_back(WritePtr);
01572   DependencySetId.push_back(DepSetId);
01573   AliasSetId.push_back(ASId);
01574 }
01575 
01576 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
01577   // We need to place the broadcast of invariant variables outside the loop.
01578   Instruction *Instr = dyn_cast<Instruction>(V);
01579   bool NewInstr =
01580       (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
01581                           Instr->getParent()) != LoopVectorBody.end());
01582   bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
01583 
01584   // Place the code for broadcasting invariant variables in the new preheader.
01585   IRBuilder<>::InsertPointGuard Guard(Builder);
01586   if (Invariant)
01587     Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
01588 
01589   // Broadcast the scalar into all locations in the vector.
01590   Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
01591 
01592   return Shuf;
01593 }
01594 
01595 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
01596                                                  bool Negate) {
01597   assert(Val->getType()->isVectorTy() && "Must be a vector");
01598   assert(Val->getType()->getScalarType()->isIntegerTy() &&
01599          "Elem must be an integer");
01600   // Create the types.
01601   Type *ITy = Val->getType()->getScalarType();
01602   VectorType *Ty = cast<VectorType>(Val->getType());
01603   int VLen = Ty->getNumElements();
01604   SmallVector<Constant*, 8> Indices;
01605 
01606   // Create a vector of consecutive numbers from zero to VF.
01607   for (int i = 0; i < VLen; ++i) {
01608     int64_t Idx = Negate ? (-i) : i;
01609     Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
01610   }
01611 
01612   // Add the consecutive indices to the vector value.
01613   Constant *Cv = ConstantVector::get(Indices);
01614   assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
01615   return Builder.CreateAdd(Val, Cv, "induction");
01616 }
01617 
01618 /// \brief Find the operand of the GEP that should be checked for consecutive
01619 /// stores. This ignores trailing indices that have no effect on the final
01620 /// pointer.
01621 static unsigned getGEPInductionOperand(const DataLayout *DL,
01622                                        const GetElementPtrInst *Gep) {
01623   unsigned LastOperand = Gep->getNumOperands() - 1;
01624   unsigned GEPAllocSize = DL->getTypeAllocSize(
01625       cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
01626 
01627   // Walk backwards and try to peel off zeros.
01628   while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
01629     // Find the type we're currently indexing into.
01630     gep_type_iterator GEPTI = gep_type_begin(Gep);
01631     std::advance(GEPTI, LastOperand - 1);
01632 
01633     // If it's a type with the same allocation size as the result of the GEP we
01634     // can peel off the zero index.
01635     if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
01636       break;
01637     --LastOperand;
01638   }
01639 
01640   return LastOperand;
01641 }
01642 
01643 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
01644   assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
01645   // Make sure that the pointer does not point to structs.
01646   if (Ptr->getType()->getPointerElementType()->isAggregateType())
01647     return 0;
01648 
01649   // If this value is a pointer induction variable we know it is consecutive.
01650   PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
01651   if (Phi && Inductions.count(Phi)) {
01652     InductionInfo II = Inductions[Phi];
01653     if (IK_PtrInduction == II.IK)
01654       return 1;
01655     else if (IK_ReversePtrInduction == II.IK)
01656       return -1;
01657   }
01658 
01659   GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
01660   if (!Gep)
01661     return 0;
01662 
01663   unsigned NumOperands = Gep->getNumOperands();
01664   Value *GpPtr = Gep->getPointerOperand();
01665   // If this GEP value is a consecutive pointer induction variable and all of
01666   // the indices are constant then we know it is consecutive. We can
01667   Phi = dyn_cast<PHINode>(GpPtr);
01668   if (Phi && Inductions.count(Phi)) {
01669 
01670     // Make sure that the pointer does not point to structs.
01671     PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
01672     if (GepPtrType->getElementType()->isAggregateType())
01673       return 0;
01674 
01675     // Make sure that all of the index operands are loop invariant.
01676     for (unsigned i = 1; i < NumOperands; ++i)
01677       if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
01678         return 0;
01679 
01680     InductionInfo II = Inductions[Phi];
01681     if (IK_PtrInduction == II.IK)
01682       return 1;
01683     else if (IK_ReversePtrInduction == II.IK)
01684       return -1;
01685   }
01686 
01687   unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
01688 
01689   // Check that all of the gep indices are uniform except for our induction
01690   // operand.
01691   for (unsigned i = 0; i != NumOperands; ++i)
01692     if (i != InductionOperand &&
01693         !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
01694       return 0;
01695 
01696   // We can emit wide load/stores only if the last non-zero index is the
01697   // induction variable.
01698   const SCEV *Last = nullptr;
01699   if (!Strides.count(Gep))
01700     Last = SE->getSCEV(Gep->getOperand(InductionOperand));
01701   else {
01702     // Because of the multiplication by a stride we can have a s/zext cast.
01703     // We are going to replace this stride by 1 so the cast is safe to ignore.
01704     //
01705     //  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
01706     //  %0 = trunc i64 %indvars.iv to i32
01707     //  %mul = mul i32 %0, %Stride1
01708     //  %idxprom = zext i32 %mul to i64  << Safe cast.
01709     //  %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
01710     //
01711     Last = replaceSymbolicStrideSCEV(SE, Strides,
01712                                      Gep->getOperand(InductionOperand), Gep);
01713     if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
01714       Last =
01715           (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
01716               ? C->getOperand()
01717               : Last;
01718   }
01719   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
01720     const SCEV *Step = AR->getStepRecurrence(*SE);
01721 
01722     // The memory is consecutive because the last index is consecutive
01723     // and all other indices are loop invariant.
01724     if (Step->isOne())
01725       return 1;
01726     if (Step->isAllOnesValue())
01727       return -1;
01728   }
01729 
01730   return 0;
01731 }
01732 
01733 bool LoopVectorizationLegality::isUniform(Value *V) {
01734   return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
01735 }
01736 
01737 InnerLoopVectorizer::VectorParts&
01738 InnerLoopVectorizer::getVectorValue(Value *V) {
01739   assert(V != Induction && "The new induction variable should not be used.");
01740   assert(!V->getType()->isVectorTy() && "Can't widen a vector");
01741 
01742   // If we have a stride that is replaced by one, do it here.
01743   if (Legal->hasStride(V))
01744     V = ConstantInt::get(V->getType(), 1);
01745 
01746   // If we have this scalar in the map, return it.
01747   if (WidenMap.has(V))
01748     return WidenMap.get(V);
01749 
01750   // If this scalar is unknown, assume that it is a constant or that it is
01751   // loop invariant. Broadcast V and save the value for future uses.
01752   Value *B = getBroadcastInstrs(V);
01753   return WidenMap.splat(V, B);
01754 }
01755 
01756 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
01757   assert(Vec->getType()->isVectorTy() && "Invalid type");
01758   SmallVector<Constant*, 8> ShuffleMask;
01759   for (unsigned i = 0; i < VF; ++i)
01760     ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
01761 
01762   return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
01763                                      ConstantVector::get(ShuffleMask),
01764                                      "reverse");
01765 }
01766 
01767 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
01768   // Attempt to issue a wide load.
01769   LoadInst *LI = dyn_cast<LoadInst>(Instr);
01770   StoreInst *SI = dyn_cast<StoreInst>(Instr);
01771 
01772   assert((LI || SI) && "Invalid Load/Store instruction");
01773 
01774   Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
01775   Type *DataTy = VectorType::get(ScalarDataTy, VF);
01776   Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
01777   unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
01778   // An alignment of 0 means target abi alignment. We need to use the scalar's
01779   // target abi alignment in such a case.
01780   if (!Alignment)
01781     Alignment = DL->getABITypeAlignment(ScalarDataTy);
01782   unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
01783   unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
01784   unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
01785 
01786   if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
01787       !Legal->isMaskRequired(SI))
01788     return scalarizeInstruction(Instr, true);
01789 
01790   if (ScalarAllocatedSize != VectorElementSize)
01791     return scalarizeInstruction(Instr);
01792 
01793   // If the pointer is loop invariant or if it is non-consecutive,
01794   // scalarize the load.
01795   int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
01796   bool Reverse = ConsecutiveStride < 0;
01797   bool UniformLoad = LI && Legal->isUniform(Ptr);
01798   if (!ConsecutiveStride || UniformLoad)
01799     return scalarizeInstruction(Instr);
01800 
01801   Constant *Zero = Builder.getInt32(0);
01802   VectorParts &Entry = WidenMap.get(Instr);
01803 
01804   // Handle consecutive loads/stores.
01805   GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
01806   if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
01807     setDebugLocFromInst(Builder, Gep);
01808     Value *PtrOperand = Gep->getPointerOperand();
01809     Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
01810     FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
01811 
01812     // Create the new GEP with the new induction variable.
01813     GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
01814     Gep2->setOperand(0, FirstBasePtr);
01815     Gep2->setName("gep.indvar.base");
01816     Ptr = Builder.Insert(Gep2);
01817   } else if (Gep) {
01818     setDebugLocFromInst(Builder, Gep);
01819     assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
01820                                OrigLoop) && "Base ptr must be invariant");
01821 
01822     // The last index does not have to be the induction. It can be
01823     // consecutive and be a function of the index. For example A[I+1];
01824     unsigned NumOperands = Gep->getNumOperands();
01825     unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
01826     // Create the new GEP with the new induction variable.
01827     GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
01828 
01829     for (unsigned i = 0; i < NumOperands; ++i) {
01830       Value *GepOperand = Gep->getOperand(i);
01831       Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
01832 
01833       // Update last index or loop invariant instruction anchored in loop.
01834       if (i == InductionOperand ||
01835           (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
01836         assert((i == InductionOperand ||
01837                SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
01838                "Must be last index or loop invariant");
01839 
01840         VectorParts &GEPParts = getVectorValue(GepOperand);
01841         Value *Index = GEPParts[0];
01842         Index = Builder.CreateExtractElement(Index, Zero);
01843         Gep2->setOperand(i, Index);
01844         Gep2->setName("gep.indvar.idx");
01845       }
01846     }
01847     Ptr = Builder.Insert(Gep2);
01848   } else {
01849     // Use the induction element ptr.
01850     assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
01851     setDebugLocFromInst(Builder, Ptr);
01852     VectorParts &PtrVal = getVectorValue(Ptr);
01853     Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
01854   }
01855 
01856   VectorParts Mask = createBlockInMask(Instr->getParent());
01857   // Handle Stores:
01858   if (SI) {
01859     assert(!Legal->isUniform(SI->getPointerOperand()) &&
01860            "We do not allow storing to uniform addresses");
01861     setDebugLocFromInst(Builder, SI);
01862     // We don't want to update the value in the map as it might be used in
01863     // another expression. So don't use a reference type for "StoredVal".
01864     VectorParts StoredVal = getVectorValue(SI->getValueOperand());
01865     
01866     for (unsigned Part = 0; Part < UF; ++Part) {
01867       // Calculate the pointer for the specific unroll-part.
01868       Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
01869 
01870       if (Reverse) {
01871         // If we store to reverse consecutive memory locations then we need
01872         // to reverse the order of elements in the stored value.
01873         StoredVal[Part] = reverseVector(StoredVal[Part]);
01874         // If the address is consecutive but reversed, then the
01875         // wide store needs to start at the last vector element.
01876         PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
01877         PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
01878         Mask[Part] = reverseVector(Mask[Part]);
01879       }
01880 
01881       Value *VecPtr = Builder.CreateBitCast(PartPtr,
01882                                             DataTy->getPointerTo(AddressSpace));
01883 
01884       Instruction *NewSI;
01885       if (Legal->isMaskRequired(SI))
01886         NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
01887                                           Mask[Part]);
01888       else 
01889         NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
01890       propagateMetadata(NewSI, SI);
01891     }
01892     return;
01893   }
01894 
01895   // Handle loads.
01896   assert(LI && "Must have a load instruction");
01897   setDebugLocFromInst(Builder, LI);
01898   for (unsigned Part = 0; Part < UF; ++Part) {
01899     // Calculate the pointer for the specific unroll-part.
01900     Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
01901 
01902     if (Reverse) {
01903       // If the address is consecutive but reversed, then the
01904       // wide load needs to start at the last vector element.
01905       PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
01906       PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
01907       Mask[Part] = reverseVector(Mask[Part]);
01908     }
01909 
01910     Instruction* NewLI;
01911     Value *VecPtr = Builder.CreateBitCast(PartPtr,
01912                                           DataTy->getPointerTo(AddressSpace));
01913     if (Legal->isMaskRequired(LI))
01914       NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
01915                                        UndefValue::get(DataTy),
01916                                        "wide.masked.load");
01917     else
01918       NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
01919     propagateMetadata(NewLI, LI);
01920     Entry[Part] = Reverse ? reverseVector(NewLI) :  NewLI;
01921   }
01922 }
01923 
01924 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
01925   assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
01926   // Holds vector parameters or scalars, in case of uniform vals.
01927   SmallVector<VectorParts, 4> Params;
01928 
01929   setDebugLocFromInst(Builder, Instr);
01930 
01931   // Find all of the vectorized parameters.
01932   for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
01933     Value *SrcOp = Instr->getOperand(op);
01934 
01935     // If we are accessing the old induction variable, use the new one.
01936     if (SrcOp == OldInduction) {
01937       Params.push_back(getVectorValue(SrcOp));
01938       continue;
01939     }
01940 
01941     // Try using previously calculated values.
01942     Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
01943 
01944     // If the src is an instruction that appeared earlier in the basic block
01945     // then it should already be vectorized.
01946     if (SrcInst && OrigLoop->contains(SrcInst)) {
01947       assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
01948       // The parameter is a vector value from earlier.
01949       Params.push_back(WidenMap.get(SrcInst));
01950     } else {
01951       // The parameter is a scalar from outside the loop. Maybe even a constant.
01952       VectorParts Scalars;
01953       Scalars.append(UF, SrcOp);
01954       Params.push_back(Scalars);
01955     }
01956   }
01957 
01958   assert(Params.size() == Instr->getNumOperands() &&
01959          "Invalid number of operands");
01960 
01961   // Does this instruction return a value ?
01962   bool IsVoidRetTy = Instr->getType()->isVoidTy();
01963 
01964   Value *UndefVec = IsVoidRetTy ? nullptr :
01965     UndefValue::get(VectorType::get(Instr->getType(), VF));
01966   // Create a new entry in the WidenMap and initialize it to Undef or Null.
01967   VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
01968 
01969   Instruction *InsertPt = Builder.GetInsertPoint();
01970   BasicBlock *IfBlock = Builder.GetInsertBlock();
01971   BasicBlock *CondBlock = nullptr;
01972 
01973   VectorParts Cond;
01974   Loop *VectorLp = nullptr;
01975   if (IfPredicateStore) {
01976     assert(Instr->getParent()->getSinglePredecessor() &&
01977            "Only support single predecessor blocks");
01978     Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
01979                           Instr->getParent());
01980     VectorLp = LI->getLoopFor(IfBlock);
01981     assert(VectorLp && "Must have a loop for this block");
01982   }
01983 
01984   // For each vector unroll 'part':
01985   for (unsigned Part = 0; Part < UF; ++Part) {
01986     // For each scalar that we create:
01987     for (unsigned Width = 0; Width < VF; ++Width) {
01988 
01989       // Start if-block.
01990       Value *Cmp = nullptr;
01991       if (IfPredicateStore) {
01992         Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
01993         Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
01994         CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
01995         LoopVectorBody.push_back(CondBlock);
01996         VectorLp->addBasicBlockToLoop(CondBlock, *LI);
01997         // Update Builder with newly created basic block.
01998         Builder.SetInsertPoint(InsertPt);
01999       }
02000 
02001       Instruction *Cloned = Instr->clone();
02002       if (!IsVoidRetTy)
02003         Cloned->setName(Instr->getName() + ".cloned");
02004       // Replace the operands of the cloned instructions with extracted scalars.
02005       for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
02006         Value *Op = Params[op][Part];
02007         // Param is a vector. Need to extract the right lane.
02008         if (Op->getType()->isVectorTy())
02009           Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
02010         Cloned->setOperand(op, Op);
02011       }
02012 
02013       // Place the cloned scalar in the new loop.
02014       Builder.Insert(Cloned);
02015 
02016       // If the original scalar returns a value we need to place it in a vector
02017       // so that future users will be able to use it.
02018       if (!IsVoidRetTy)
02019         VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
02020                                                        Builder.getInt32(Width));
02021       // End if-block.
02022       if (IfPredicateStore) {
02023          BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
02024          LoopVectorBody.push_back(NewIfBlock);
02025          VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
02026          Builder.SetInsertPoint(InsertPt);
02027          Instruction *OldBr = IfBlock->getTerminator();
02028          BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
02029          OldBr->eraseFromParent();
02030          IfBlock = NewIfBlock;
02031       }
02032     }
02033   }
02034 }
02035 
02036 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
02037                                  Instruction *Loc) {
02038   if (FirstInst)
02039     return FirstInst;
02040   if (Instruction *I = dyn_cast<Instruction>(V))
02041     return I->getParent() == Loc->getParent() ? I : nullptr;
02042   return nullptr;
02043 }
02044 
02045 std::pair<Instruction *, Instruction *>
02046 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
02047   Instruction *tnullptr = nullptr;
02048   if (!Legal->mustCheckStrides())
02049     return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
02050 
02051   IRBuilder<> ChkBuilder(Loc);
02052 
02053   // Emit checks.
02054   Value *Check = nullptr;
02055   Instruction *FirstInst = nullptr;
02056   for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
02057                                          SE = Legal->strides_end();
02058        SI != SE; ++SI) {
02059     Value *Ptr = stripIntegerCast(*SI);
02060     Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
02061                                        "stride.chk");
02062     // Store the first instruction we create.
02063     FirstInst = getFirstInst(FirstInst, C, Loc);
02064     if (Check)
02065       Check = ChkBuilder.CreateOr(Check, C);
02066     else
02067       Check = C;
02068   }
02069 
02070   // We have to do this trickery because the IRBuilder might fold the check to a
02071   // constant expression in which case there is no Instruction anchored in a
02072   // the block.
02073   LLVMContext &Ctx = Loc->getContext();
02074   Instruction *TheCheck =
02075       BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
02076   ChkBuilder.Insert(TheCheck, "stride.not.one");
02077   FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
02078 
02079   return std::make_pair(FirstInst, TheCheck);
02080 }
02081 
02082 std::pair<Instruction *, Instruction *>
02083 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
02084   LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
02085   Legal->getRuntimePointerCheck();
02086 
02087   Instruction *tnullptr = nullptr;
02088   if (!PtrRtCheck->Need)
02089     return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
02090 
02091   unsigned NumPointers = PtrRtCheck->Pointers.size();
02092   SmallVector<TrackingVH<Value> , 2> Starts;
02093   SmallVector<TrackingVH<Value> , 2> Ends;
02094 
02095   LLVMContext &Ctx = Loc->getContext();
02096   SCEVExpander Exp(*SE, "induction");
02097   Instruction *FirstInst = nullptr;
02098 
02099   for (unsigned i = 0; i < NumPointers; ++i) {
02100     Value *Ptr = PtrRtCheck->Pointers[i];
02101     const SCEV *Sc = SE->getSCEV(Ptr);
02102 
02103     if (SE->isLoopInvariant(Sc, OrigLoop)) {
02104       DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
02105             *Ptr <<"\n");
02106       Starts.push_back(Ptr);
02107       Ends.push_back(Ptr);
02108     } else {
02109       DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
02110       unsigned AS = Ptr->getType()->getPointerAddressSpace();
02111 
02112       // Use this type for pointer arithmetic.
02113       Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
02114 
02115       Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
02116       Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
02117       Starts.push_back(Start);
02118       Ends.push_back(End);
02119     }
02120   }
02121 
02122   IRBuilder<> ChkBuilder(Loc);
02123   // Our instructions might fold to a constant.
02124   Value *MemoryRuntimeCheck = nullptr;
02125   for (unsigned i = 0; i < NumPointers; ++i) {
02126     for (unsigned j = i+1; j < NumPointers; ++j) {
02127       // No need to check if two readonly pointers intersect.
02128       if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
02129         continue;
02130 
02131       // Only need to check pointers between two different dependency sets.
02132       if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
02133        continue;
02134       // Only need to check pointers in the same alias set.
02135       if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
02136         continue;
02137 
02138       unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
02139       unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
02140 
02141       assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
02142              (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
02143              "Trying to bounds check pointers with different address spaces");
02144 
02145       Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
02146       Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
02147 
02148       Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
02149       Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
02150       Value *End0 =   ChkBuilder.CreateBitCast(Ends[i],   PtrArithTy1, "bc");
02151       Value *End1 =   ChkBuilder.CreateBitCast(Ends[j],   PtrArithTy0, "bc");
02152 
02153       Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
02154       FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
02155       Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
02156       FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
02157       Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
02158       FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
02159       if (MemoryRuntimeCheck) {
02160         IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
02161                                          "conflict.rdx");
02162         FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
02163       }
02164       MemoryRuntimeCheck = IsConflict;
02165     }
02166   }
02167 
02168   // We have to do this trickery because the IRBuilder might fold the check to a
02169   // constant expression in which case there is no Instruction anchored in a
02170   // the block.
02171   Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
02172                                                  ConstantInt::getTrue(Ctx));
02173   ChkBuilder.Insert(Check, "memcheck.conflict");
02174   FirstInst = getFirstInst(FirstInst, Check, Loc);
02175   return std::make_pair(FirstInst, Check);
02176 }
02177 
02178 void InnerLoopVectorizer::createEmptyLoop() {
02179   /*
02180    In this function we generate a new loop. The new loop will contain
02181    the vectorized instructions while the old loop will continue to run the
02182    scalar remainder.
02183 
02184        [ ] <-- Back-edge taken count overflow check.
02185     /   |
02186    /    v
02187   |    [ ] <-- vector loop bypass (may consist of multiple blocks).
02188   |  /  |
02189   | /   v
02190   ||   [ ]     <-- vector pre header.
02191   ||    |
02192   ||    v
02193   ||   [  ] \
02194   ||   [  ]_|   <-- vector loop.
02195   ||    |
02196   | \   v
02197   |   >[ ]   <--- middle-block.
02198   |  /  |
02199   | /   v
02200   -|- >[ ]     <--- new preheader.
02201    |    |
02202    |    v
02203    |   [ ] \
02204    |   [ ]_|   <-- old scalar loop to handle remainder.
02205     \   |
02206      \  v
02207       >[ ]     <-- exit block.
02208    ...
02209    */
02210 
02211   BasicBlock *OldBasicBlock = OrigLoop->getHeader();
02212   BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
02213   BasicBlock *ExitBlock = OrigLoop->getExitBlock();
02214   assert(BypassBlock && "Invalid loop structure");
02215   assert(ExitBlock && "Must have an exit block");
02216 
02217   // Some loops have a single integer induction variable, while other loops
02218   // don't. One example is c++ iterators that often have multiple pointer
02219   // induction variables. In the code below we also support a case where we
02220   // don't have a single induction variable.
02221   OldInduction = Legal->getInduction();
02222   Type *IdxTy = Legal->getWidestInductionType();
02223 
02224   // Find the loop boundaries.
02225   const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
02226   assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
02227 
02228   // The exit count might have the type of i64 while the phi is i32. This can
02229   // happen if we have an induction variable that is sign extended before the
02230   // compare. The only way that we get a backedge taken count is that the
02231   // induction variable was signed and as such will not overflow. In such a case
02232   // truncation is legal.
02233   if (ExitCount->getType()->getPrimitiveSizeInBits() >
02234       IdxTy->getPrimitiveSizeInBits())
02235     ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
02236 
02237   const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
02238   // Get the total trip count from the count by adding 1.
02239   ExitCount = SE->getAddExpr(BackedgeTakeCount,
02240                              SE->getConstant(BackedgeTakeCount->getType(), 1));
02241 
02242   // Expand the trip count and place the new instructions in the preheader.
02243   // Notice that the pre-header does not change, only the loop body.
02244   SCEVExpander Exp(*SE, "induction");
02245 
02246   // We need to test whether the backedge-taken count is uint##_max. Adding one
02247   // to it will cause overflow and an incorrect loop trip count in the vector
02248   // body. In case of overflow we want to directly jump to the scalar remainder
02249   // loop.
02250   Value *BackedgeCount =
02251       Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
02252                         BypassBlock->getTerminator());
02253   if (BackedgeCount->getType()->isPointerTy())
02254     BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
02255                                                 "backedge.ptrcnt.to.int",
02256                                                 BypassBlock->getTerminator());
02257   Instruction *CheckBCOverflow =
02258       CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
02259                       Constant::getAllOnesValue(BackedgeCount->getType()),
02260                       "backedge.overflow", BypassBlock->getTerminator());
02261 
02262   // The loop index does not have to start at Zero. Find the original start
02263   // value from the induction PHI node. If we don't have an induction variable
02264   // then we know that it starts at zero.
02265   Builder.SetInsertPoint(BypassBlock->getTerminator());
02266   Value *StartIdx = ExtendedIdx = OldInduction ?
02267     Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
02268                        IdxTy):
02269     ConstantInt::get(IdxTy, 0);
02270 
02271   // We need an instruction to anchor the overflow check on. StartIdx needs to
02272   // be defined before the overflow check branch. Because the scalar preheader
02273   // is going to merge the start index and so the overflow branch block needs to
02274   // contain a definition of the start index.
02275   Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
02276       StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
02277       BypassBlock->getTerminator());
02278 
02279   // Count holds the overall loop count (N).
02280   Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
02281                                    BypassBlock->getTerminator());
02282 
02283   LoopBypassBlocks.push_back(BypassBlock);
02284 
02285   // Split the single block loop into the two loop structure described above.
02286   BasicBlock *VectorPH =
02287   BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
02288   BasicBlock *VecBody =
02289   VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
02290   BasicBlock *MiddleBlock =
02291   VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
02292   BasicBlock *ScalarPH =
02293   MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
02294 
02295   // Create and register the new vector loop.
02296   Loop* Lp = new Loop();
02297   Loop *ParentLoop = OrigLoop->getParentLoop();
02298 
02299   // Insert the new loop into the loop nest and register the new basic blocks
02300   // before calling any utilities such as SCEV that require valid LoopInfo.
02301   if (ParentLoop) {
02302     ParentLoop->addChildLoop(Lp);
02303     ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
02304     ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
02305     ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
02306   } else {
02307     LI->addTopLevelLoop(Lp);
02308   }
02309   Lp->addBasicBlockToLoop(VecBody, *LI);
02310 
02311   // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
02312   // inside the loop.
02313   Builder.SetInsertPoint(VecBody->getFirstNonPHI());
02314 
02315   // Generate the induction variable.
02316   setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
02317   Induction = Builder.CreatePHI(IdxTy, 2, "index");
02318   // The loop step is equal to the vectorization factor (num of SIMD elements)
02319   // times the unroll factor (num of SIMD instructions).
02320   Constant *Step = ConstantInt::get(IdxTy, VF * UF);
02321 
02322   // This is the IR builder that we use to add all of the logic for bypassing
02323   // the new vector loop.
02324   IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
02325   setDebugLocFromInst(BypassBuilder,
02326                       getDebugLocFromInstOrOperands(OldInduction));
02327 
02328   // We may need to extend the index in case there is a type mismatch.
02329   // We know that the count starts at zero and does not overflow.
02330   if (Count->getType() != IdxTy) {
02331     // The exit count can be of pointer type. Convert it to the correct
02332     // integer type.
02333     if (ExitCount->getType()->isPointerTy())
02334       Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
02335     else
02336       Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
02337   }
02338 
02339   // Add the start index to the loop count to get the new end index.
02340   Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
02341 
02342   // Now we need to generate the expression for N - (N % VF), which is
02343   // the part that the vectorized body will execute.
02344   Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
02345   Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
02346   Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
02347                                                      "end.idx.rnd.down");
02348 
02349   // Now, compare the new count to zero. If it is zero skip the vector loop and
02350   // jump to the scalar loop.
02351   Value *Cmp =
02352       BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
02353 
02354   BasicBlock *LastBypassBlock = BypassBlock;
02355 
02356   // Generate code to check that the loops trip count that we computed by adding
02357   // one to the backedge-taken count will not overflow.
02358   {
02359     auto PastOverflowCheck =
02360         std::next(BasicBlock::iterator(OverflowCheckAnchor));
02361     BasicBlock *CheckBlock =
02362       LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
02363     if (ParentLoop)
02364       ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
02365     LoopBypassBlocks.push_back(CheckBlock);
02366     Instruction *OldTerm = LastBypassBlock->getTerminator();
02367     BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
02368     OldTerm->eraseFromParent();
02369     LastBypassBlock = CheckBlock;
02370   }
02371 
02372   // Generate the code to check that the strides we assumed to be one are really
02373   // one. We want the new basic block to start at the first instruction in a
02374   // sequence of instructions that form a check.
02375   Instruction *StrideCheck;
02376   Instruction *FirstCheckInst;
02377   std::tie(FirstCheckInst, StrideCheck) =
02378       addStrideCheck(LastBypassBlock->getTerminator());
02379   if (StrideCheck) {
02380     // Create a new block containing the stride check.
02381     BasicBlock *CheckBlock =
02382         LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
02383     if (ParentLoop)
02384       ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
02385     LoopBypassBlocks.push_back(CheckBlock);
02386 
02387     // Replace the branch into the memory check block with a conditional branch
02388     // for the "few elements case".
02389     Instruction *OldTerm = LastBypassBlock->getTerminator();
02390     BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
02391     OldTerm->eraseFromParent();
02392 
02393     Cmp = StrideCheck;
02394     LastBypassBlock = CheckBlock;
02395   }
02396 
02397   // Generate the code that checks in runtime if arrays overlap. We put the
02398   // checks into a separate block to make the more common case of few elements
02399   // faster.
02400   Instruction *MemRuntimeCheck;
02401   std::tie(FirstCheckInst, MemRuntimeCheck) =
02402       addRuntimeCheck(LastBypassBlock->getTerminator());
02403   if (MemRuntimeCheck) {
02404     // Create a new block containing the memory check.
02405     BasicBlock *CheckBlock =
02406         LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
02407     if (ParentLoop)
02408       ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
02409     LoopBypassBlocks.push_back(CheckBlock);
02410 
02411     // Replace the branch into the memory check block with a conditional branch
02412     // for the "few elements case".
02413     Instruction *OldTerm = LastBypassBlock->getTerminator();
02414     BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
02415     OldTerm->eraseFromParent();
02416 
02417     Cmp = MemRuntimeCheck;
02418     LastBypassBlock = CheckBlock;
02419   }
02420 
02421   LastBypassBlock->getTerminator()->eraseFromParent();
02422   BranchInst::Create(MiddleBlock, VectorPH, Cmp,
02423                      LastBypassBlock);
02424 
02425   // We are going to resume the execution of the scalar loop.
02426   // Go over all of the induction variables that we found and fix the
02427   // PHIs that are left in the scalar version of the loop.
02428   // The starting values of PHI nodes depend on the counter of the last
02429   // iteration in the vectorized loop.
02430   // If we come from a bypass edge then we need to start from the original
02431   // start value.
02432 
02433   // This variable saves the new starting index for the scalar loop.
02434   PHINode *ResumeIndex = nullptr;
02435   LoopVectorizationLegality::InductionList::iterator I, E;
02436   LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
02437   // Set builder to point to last bypass block.
02438   BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
02439   for (I = List->begin(), E = List->end(); I != E; ++I) {
02440     PHINode *OrigPhi = I->first;
02441     LoopVectorizationLegality::InductionInfo II = I->second;
02442 
02443     Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
02444     PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
02445                                          MiddleBlock->getTerminator());
02446     // We might have extended the type of the induction variable but we need a
02447     // truncated version for the scalar loop.
02448     PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
02449       PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
02450                       MiddleBlock->getTerminator()) : nullptr;
02451 
02452     // Create phi nodes to merge from the  backedge-taken check block.
02453     PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
02454                                            ScalarPH->getTerminator());
02455     BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
02456 
02457     PHINode *BCTruncResumeVal = nullptr;
02458     if (OrigPhi == OldInduction) {
02459       BCTruncResumeVal =
02460           PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
02461                           ScalarPH->getTerminator());
02462       BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
02463     }
02464 
02465     Value *EndValue = nullptr;
02466     switch (II.IK) {
02467     case LoopVectorizationLegality::IK_NoInduction:
02468       llvm_unreachable("Unknown induction");
02469     case LoopVectorizationLegality::IK_IntInduction: {
02470       // Handle the integer induction counter.
02471       assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
02472 
02473       // We have the canonical induction variable.
02474       if (OrigPhi == OldInduction) {
02475         // Create a truncated version of the resume value for the scalar loop,
02476         // we might have promoted the type to a larger width.
02477         EndValue =
02478           BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
02479         // The new PHI merges the original incoming value, in case of a bypass,
02480         // or the value at the end of the vectorized loop.
02481         for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
02482           TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
02483         TruncResumeVal->addIncoming(EndValue, VecBody);
02484 
02485         BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
02486 
02487         // We know what the end value is.
02488         EndValue = IdxEndRoundDown;
02489         // We also know which PHI node holds it.
02490         ResumeIndex = ResumeVal;
02491         break;
02492       }
02493 
02494       // Not the canonical induction variable - add the vector loop count to the
02495       // start value.
02496       Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
02497                                                    II.StartValue->getType(),
02498                                                    "cast.crd");
02499       EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
02500       break;
02501     }
02502     case LoopVectorizationLegality::IK_ReverseIntInduction: {
02503       // Convert the CountRoundDown variable to the PHI size.
02504       Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
02505                                                    II.StartValue->getType(),
02506                                                    "cast.crd");
02507       // Handle reverse integer induction counter.
02508       EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
02509       break;
02510     }
02511     case LoopVectorizationLegality::IK_PtrInduction: {
02512       // For pointer induction variables, calculate the offset using
02513       // the end index.
02514       EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
02515                                          "ptr.ind.end");
02516       break;
02517     }
02518     case LoopVectorizationLegality::IK_ReversePtrInduction: {
02519       // The value at the end of the loop for the reverse pointer is calculated
02520       // by creating a GEP with a negative index starting from the start value.
02521       Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
02522       Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
02523                                               "rev.ind.end");
02524       EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
02525                                          "rev.ptr.ind.end");
02526       break;
02527     }
02528     }// end of case
02529 
02530     // The new PHI merges the original incoming value, in case of a bypass,
02531     // or the value at the end of the vectorized loop.
02532     for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
02533       if (OrigPhi == OldInduction)
02534         ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
02535       else
02536         ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
02537     }
02538     ResumeVal->addIncoming(EndValue, VecBody);
02539 
02540     // Fix the scalar body counter (PHI node).
02541     unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
02542 
02543     // The old induction's phi node in the scalar body needs the truncated
02544     // value.
02545     if (OrigPhi == OldInduction) {
02546       BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
02547       OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
02548     } else {
02549       BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
02550       OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
02551     }
02552   }
02553 
02554   // If we are generating a new induction variable then we also need to
02555   // generate the code that calculates the exit value. This value is not
02556   // simply the end of the counter because we may skip the vectorized body
02557   // in case of a runtime check.
02558   if (!OldInduction){
02559     assert(!ResumeIndex && "Unexpected resume value found");
02560     ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
02561                                   MiddleBlock->getTerminator());
02562     for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
02563       ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
02564     ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
02565   }
02566 
02567   // Make sure that we found the index where scalar loop needs to continue.
02568   assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
02569          "Invalid resume Index");
02570 
02571   // Add a check in the middle block to see if we have completed
02572   // all of the iterations in the first vector loop.
02573   // If (N - N%VF) == N, then we *don't* need to run the remainder.
02574   Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
02575                                 ResumeIndex, "cmp.n",
02576                                 MiddleBlock->getTerminator());
02577 
02578   BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
02579   // Remove the old terminator.
02580   MiddleBlock->getTerminator()->eraseFromParent();
02581 
02582   // Create i+1 and fill the PHINode.
02583   Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
02584   Induction->addIncoming(StartIdx, VectorPH);
02585   Induction->addIncoming(NextIdx, VecBody);
02586   // Create the compare.
02587   Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
02588   Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
02589 
02590   // Now we have two terminators. Remove the old one from the block.
02591   VecBody->getTerminator()->eraseFromParent();
02592 
02593   // Get ready to start creating new instructions into the vectorized body.
02594   Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
02595 
02596   // Save the state.
02597   LoopVectorPreHeader = VectorPH;
02598   LoopScalarPreHeader = ScalarPH;
02599   LoopMiddleBlock = MiddleBlock;
02600   LoopExitBlock = ExitBlock;
02601   LoopVectorBody.push_back(VecBody);
02602   LoopScalarBody = OldBasicBlock;
02603 
02604   LoopVectorizeHints Hints(Lp, true);
02605   Hints.setAlreadyVectorized();
02606 }
02607 
02608 /// This function returns the identity element (or neutral element) for
02609 /// the operation K.
02610 Constant*
02611 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
02612   switch (K) {
02613   case RK_IntegerXor:
02614   case RK_IntegerAdd:
02615   case RK_IntegerOr:
02616     // Adding, Xoring, Oring zero to a number does not change it.
02617     return ConstantInt::get(Tp, 0);
02618   case RK_IntegerMult:
02619     // Multiplying a number by 1 does not change it.
02620     return ConstantInt::get(Tp, 1);
02621   case RK_IntegerAnd:
02622     // AND-ing a number with an all-1 value does not change it.
02623     return ConstantInt::get(Tp, -1, true);
02624   case  RK_FloatMult:
02625     // Multiplying a number by 1 does not change it.
02626     return ConstantFP::get(Tp, 1.0L);
02627   case  RK_FloatAdd:
02628     // Adding zero to a number does not change it.
02629     return ConstantFP::get(Tp, 0.0L);
02630   default:
02631     llvm_unreachable("Unknown reduction kind");
02632   }
02633 }
02634 
02635 /// This function translates the reduction kind to an LLVM binary operator.
02636 static unsigned
02637 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
02638   switch (Kind) {
02639     case LoopVectorizationLegality::RK_IntegerAdd:
02640       return Instruction::Add;
02641     case LoopVectorizationLegality::RK_IntegerMult:
02642       return Instruction::Mul;
02643     case LoopVectorizationLegality::RK_IntegerOr:
02644       return Instruction::Or;
02645     case LoopVectorizationLegality::RK_IntegerAnd:
02646       return Instruction::And;
02647     case LoopVectorizationLegality::RK_IntegerXor:
02648       return Instruction::Xor;
02649     case LoopVectorizationLegality::RK_FloatMult:
02650       return Instruction::FMul;
02651     case LoopVectorizationLegality::RK_FloatAdd:
02652       return Instruction::FAdd;
02653     case LoopVectorizationLegality::RK_IntegerMinMax:
02654       return Instruction::ICmp;
02655     case LoopVectorizationLegality::RK_FloatMinMax:
02656       return Instruction::FCmp;
02657     default:
02658       llvm_unreachable("Unknown reduction operation");
02659   }
02660 }
02661 
02662 Value *createMinMaxOp(IRBuilder<> &Builder,
02663                       LoopVectorizationLegality::MinMaxReductionKind RK,
02664                       Value *Left,
02665                       Value *Right) {
02666   CmpInst::Predicate P = CmpInst::ICMP_NE;
02667   switch (RK) {
02668   default:
02669     llvm_unreachable("Unknown min/max reduction kind");
02670   case LoopVectorizationLegality::MRK_UIntMin:
02671     P = CmpInst::ICMP_ULT;
02672     break;
02673   case LoopVectorizationLegality::MRK_UIntMax:
02674     P = CmpInst::ICMP_UGT;
02675     break;
02676   case LoopVectorizationLegality::MRK_SIntMin:
02677     P = CmpInst::ICMP_SLT;
02678     break;
02679   case LoopVectorizationLegality::MRK_SIntMax:
02680     P = CmpInst::ICMP_SGT;
02681     break;
02682   case LoopVectorizationLegality::MRK_FloatMin:
02683     P = CmpInst::FCMP_OLT;
02684     break;
02685   case LoopVectorizationLegality::MRK_FloatMax:
02686     P = CmpInst::FCMP_OGT;
02687     break;
02688   }
02689 
02690   Value *Cmp;
02691   if (RK == LoopVectorizationLegality::MRK_FloatMin ||
02692       RK == LoopVectorizationLegality::MRK_FloatMax)
02693     Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
02694   else
02695     Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
02696 
02697   Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
02698   return Select;
02699 }
02700 
02701 namespace {
02702 struct CSEDenseMapInfo {
02703   static bool canHandle(Instruction *I) {
02704     return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
02705            isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
02706   }
02707   static inline Instruction *getEmptyKey() {
02708     return DenseMapInfo<Instruction *>::getEmptyKey();
02709   }
02710   static inline Instruction *getTombstoneKey() {
02711     return DenseMapInfo<Instruction *>::getTombstoneKey();
02712   }
02713   static unsigned getHashValue(Instruction *I) {
02714     assert(canHandle(I) && "Unknown instruction!");
02715     return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
02716                                                            I->value_op_end()));
02717   }
02718   static bool isEqual(Instruction *LHS, Instruction *RHS) {
02719     if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
02720         LHS == getTombstoneKey() || RHS == getTombstoneKey())
02721       return LHS == RHS;
02722     return LHS->isIdenticalTo(RHS);
02723   }
02724 };
02725 }
02726 
02727 /// \brief Check whether this block is a predicated block.
02728 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
02729 /// = ...;  " blocks. We start with one vectorized basic block. For every
02730 /// conditional block we split this vectorized block. Therefore, every second
02731 /// block will be a predicated one.
02732 static bool isPredicatedBlock(unsigned BlockNum) {
02733   return BlockNum % 2;
02734 }
02735 
02736 ///\brief Perform cse of induction variable instructions.
02737 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
02738   // Perform simple cse.
02739   SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
02740   for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
02741     BasicBlock *BB = BBs[i];
02742     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
02743       Instruction *In = I++;
02744 
02745       if (!CSEDenseMapInfo::canHandle(In))
02746         continue;
02747 
02748       // Check if we can replace this instruction with any of the
02749       // visited instructions.
02750       if (Instruction *V = CSEMap.lookup(In)) {
02751         In->replaceAllUsesWith(V);
02752         In->eraseFromParent();
02753         continue;
02754       }
02755       // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
02756       // ...;" blocks for predicated stores. Every second block is a predicated
02757       // block.
02758       if (isPredicatedBlock(i))
02759         continue;
02760 
02761       CSEMap[In] = In;
02762     }
02763   }
02764 }
02765 
02766 /// \brief Adds a 'fast' flag to floating point operations.
02767 static Value *addFastMathFlag(Value *V) {
02768   if (isa<FPMathOperator>(V)){
02769     FastMathFlags Flags;
02770     Flags.setUnsafeAlgebra();
02771     cast<Instruction>(V)->setFastMathFlags(Flags);
02772   }
02773   return V;
02774 }
02775 
02776 void InnerLoopVectorizer::vectorizeLoop() {
02777   //===------------------------------------------------===//
02778   //
02779   // Notice: any optimization or new instruction that go
02780   // into the code below should be also be implemented in
02781   // the cost-model.
02782   //
02783   //===------------------------------------------------===//
02784   Constant *Zero = Builder.getInt32(0);
02785 
02786   // In order to support reduction variables we need to be able to vectorize
02787   // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
02788   // stages. First, we create a new vector PHI node with no incoming edges.
02789   // We use this value when we vectorize all of the instructions that use the
02790   // PHI. Next, after all of the instructions in the block are complete we
02791   // add the new incoming edges to the PHI. At this point all of the
02792   // instructions in the basic block are vectorized, so we can use them to
02793   // construct the PHI.
02794   PhiVector RdxPHIsToFix;
02795 
02796   // Scan the loop in a topological order to ensure that defs are vectorized
02797   // before users.
02798   LoopBlocksDFS DFS(OrigLoop);
02799   DFS.perform(LI);
02800 
02801   // Vectorize all of the blocks in the original loop.
02802   for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
02803        be = DFS.endRPO(); bb != be; ++bb)
02804     vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
02805 
02806   // At this point every instruction in the original loop is widened to
02807   // a vector form. We are almost done. Now, we need to fix the PHI nodes
02808   // that we vectorized. The PHI nodes are currently empty because we did
02809   // not want to introduce cycles. Notice that the remaining PHI nodes
02810   // that we need to fix are reduction variables.
02811 
02812   // Create the 'reduced' values for each of the induction vars.
02813   // The reduced values are the vector values that we scalarize and combine
02814   // after the loop is finished.
02815   for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
02816        it != e; ++it) {
02817     PHINode *RdxPhi = *it;
02818     assert(RdxPhi && "Unable to recover vectorized PHI");
02819 
02820     // Find the reduction variable descriptor.
02821     assert(Legal->getReductionVars()->count(RdxPhi) &&
02822            "Unable to find the reduction variable");
02823     LoopVectorizationLegality::ReductionDescriptor RdxDesc =
02824     (*Legal->getReductionVars())[RdxPhi];
02825 
02826     setDebugLocFromInst(Builder, RdxDesc.StartValue);
02827 
02828     // We need to generate a reduction vector from the incoming scalar.
02829     // To do so, we need to generate the 'identity' vector and override
02830     // one of the elements with the incoming scalar reduction. We need
02831     // to do it in the vector-loop preheader.
02832     Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
02833 
02834     // This is the vector-clone of the value that leaves the loop.
02835     VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
02836     Type *VecTy = VectorExit[0]->getType();
02837 
02838     // Find the reduction identity variable. Zero for addition, or, xor,
02839     // one for multiplication, -1 for And.
02840     Value *Identity;
02841     Value *VectorStart;
02842     if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
02843         RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
02844       // MinMax reduction have the start value as their identify.
02845       if (VF == 1) {
02846         VectorStart = Identity = RdxDesc.StartValue;
02847       } else {
02848         VectorStart = Identity = Builder.CreateVectorSplat(VF,
02849                                                            RdxDesc.StartValue,
02850                                                            "minmax.ident");
02851       }
02852     } else {
02853       // Handle other reduction kinds:
02854       Constant *Iden =
02855       LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
02856                                                       VecTy->getScalarType());
02857       if (VF == 1) {
02858         Identity = Iden;
02859         // This vector is the Identity vector where the first element is the
02860         // incoming scalar reduction.
02861         VectorStart = RdxDesc.StartValue;
02862       } else {
02863         Identity = ConstantVector::getSplat(VF, Iden);
02864 
02865         // This vector is the Identity vector where the first element is the
02866         // incoming scalar reduction.
02867         VectorStart = Builder.CreateInsertElement(Identity,
02868                                                   RdxDesc.StartValue, Zero);
02869       }
02870     }
02871 
02872     // Fix the vector-loop phi.
02873 
02874     // Reductions do not have to start at zero. They can start with
02875     // any loop invariant values.
02876     VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
02877     BasicBlock *Latch = OrigLoop->getLoopLatch();
02878     Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
02879     VectorParts &Val = getVectorValue(LoopVal);
02880     for (unsigned part = 0; part < UF; ++part) {
02881       // Make sure to add the reduction stat value only to the
02882       // first unroll part.
02883       Value *StartVal = (part == 0) ? VectorStart : Identity;
02884       cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
02885                                                   LoopVectorPreHeader);
02886       cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
02887                                                   LoopVectorBody.back());
02888     }
02889 
02890     // Before each round, move the insertion point right between
02891     // the PHIs and the values we are going to write.
02892     // This allows us to write both PHINodes and the extractelement
02893     // instructions.
02894     Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
02895 
02896     VectorParts RdxParts;
02897     setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
02898     for (unsigned part = 0; part < UF; ++part) {
02899       // This PHINode contains the vectorized reduction variable, or
02900       // the initial value vector, if we bypass the vector loop.
02901       VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
02902       PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
02903       Value *StartVal = (part == 0) ? VectorStart : Identity;
02904       for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
02905         NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
02906       NewPhi->addIncoming(RdxExitVal[part],
02907                           LoopVectorBody.back());
02908       RdxParts.push_back(NewPhi);
02909     }
02910 
02911     // Reduce all of the unrolled parts into a single vector.
02912     Value *ReducedPartRdx = RdxParts[0];
02913     unsigned Op = getReductionBinOp(RdxDesc.Kind);
02914     setDebugLocFromInst(Builder, ReducedPartRdx);
02915     for (unsigned part = 1; part < UF; ++part) {
02916       if (Op != Instruction::ICmp && Op != Instruction::FCmp)
02917         // Floating point operations had to be 'fast' to enable the reduction.
02918         ReducedPartRdx = addFastMathFlag(
02919             Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
02920                                 ReducedPartRdx, "bin.rdx"));
02921       else
02922         ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
02923                                         ReducedPartRdx, RdxParts[part]);
02924     }
02925 
02926     if (VF > 1) {
02927       // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
02928       // and vector ops, reducing the set of values being computed by half each
02929       // round.
02930       assert(isPowerOf2_32(VF) &&
02931              "Reduction emission only supported for pow2 vectors!");
02932       Value *TmpVec = ReducedPartRdx;
02933       SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
02934       for (unsigned i = VF; i != 1; i >>= 1) {
02935         // Move the upper half of the vector to the lower half.
02936         for (unsigned j = 0; j != i/2; ++j)
02937           ShuffleMask[j] = Builder.getInt32(i/2 + j);
02938 
02939         // Fill the rest of the mask with undef.
02940         std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
02941                   UndefValue::get(Builder.getInt32Ty()));
02942 
02943         Value *Shuf =
02944         Builder.CreateShuffleVector(TmpVec,
02945                                     UndefValue::get(TmpVec->getType()),
02946                                     ConstantVector::get(ShuffleMask),
02947                                     "rdx.shuf");
02948 
02949         if (Op != Instruction::ICmp && Op != Instruction::FCmp)
02950           // Floating point operations had to be 'fast' to enable the reduction.
02951           TmpVec = addFastMathFlag(Builder.CreateBinOp(
02952               (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
02953         else
02954           TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
02955       }
02956 
02957       // The result is in the first element of the vector.
02958       ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
02959                                                     Builder.getInt32(0));
02960     }
02961 
02962     // Create a phi node that merges control-flow from the backedge-taken check
02963     // block and the middle block.
02964     PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
02965                                           LoopScalarPreHeader->getTerminator());
02966     BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
02967     BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
02968 
02969     // Now, we need to fix the users of the reduction variable
02970     // inside and outside of the scalar remainder loop.
02971     // We know that the loop is in LCSSA form. We need to update the
02972     // PHI nodes in the exit blocks.
02973     for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
02974          LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
02975       PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
02976       if (!LCSSAPhi) break;
02977 
02978       // All PHINodes need to have a single entry edge, or two if
02979       // we already fixed them.
02980       assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
02981 
02982       // We found our reduction value exit-PHI. Update it with the
02983       // incoming bypass edge.
02984       if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
02985         // Add an edge coming from the bypass.
02986         LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
02987         break;
02988       }
02989     }// end of the LCSSA phi scan.
02990 
02991     // Fix the scalar loop reduction variable with the incoming reduction sum
02992     // from the vector body and from the backedge value.
02993     int IncomingEdgeBlockIdx =
02994     (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
02995     assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
02996     // Pick the other block.
02997     int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
02998     (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
02999     (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
03000   }// end of for each redux variable.
03001 
03002   fixLCSSAPHIs();
03003 
03004   // Remove redundant induction instructions.
03005   cse(LoopVectorBody);
03006 }
03007 
03008 void InnerLoopVectorizer::fixLCSSAPHIs() {
03009   for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
03010        LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
03011     PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
03012     if (!LCSSAPhi) break;
03013     if (LCSSAPhi->getNumIncomingValues() == 1)
03014       LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
03015                             LoopMiddleBlock);
03016   }
03017 }
03018 
03019 InnerLoopVectorizer::VectorParts
03020 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
03021   assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
03022          "Invalid edge");
03023 
03024   // Look for cached value.
03025   std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
03026   EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
03027   if (ECEntryIt != MaskCache.end())
03028     return ECEntryIt->second;
03029 
03030   VectorParts SrcMask = createBlockInMask(Src);
03031 
03032   // The terminator has to be a branch inst!
03033   BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
03034   assert(BI && "Unexpected terminator found");
03035 
03036   if (BI->isConditional()) {
03037     VectorParts EdgeMask = getVectorValue(BI->getCondition());
03038 
03039     if (BI->getSuccessor(0) != Dst)
03040       for (unsigned part = 0; part < UF; ++part)
03041         EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
03042 
03043     for (unsigned part = 0; part < UF; ++part)
03044       EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
03045 
03046     MaskCache[Edge] = EdgeMask;
03047     return EdgeMask;
03048   }
03049 
03050   MaskCache[Edge] = SrcMask;
03051   return SrcMask;
03052 }
03053 
03054 InnerLoopVectorizer::VectorParts
03055 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
03056   assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
03057 
03058   // Loop incoming mask is all-one.
03059   if (OrigLoop->getHeader() == BB) {
03060     Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
03061     return getVectorValue(C);
03062   }
03063 
03064   // This is the block mask. We OR all incoming edges, and with zero.
03065   Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
03066   VectorParts BlockMask = getVectorValue(Zero);
03067 
03068   // For each pred:
03069   for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
03070     VectorParts EM = createEdgeMask(*it, BB);
03071     for (unsigned part = 0; part < UF; ++part)
03072       BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
03073   }
03074 
03075   return BlockMask;
03076 }
03077 
03078 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
03079                                               InnerLoopVectorizer::VectorParts &Entry,
03080                                               unsigned UF, unsigned VF, PhiVector *PV) {
03081   PHINode* P = cast<PHINode>(PN);
03082   // Handle reduction variables:
03083   if (Legal->getReductionVars()->count(P)) {
03084     for (unsigned part = 0; part < UF; ++part) {
03085       // This is phase one of vectorizing PHIs.
03086       Type *VecTy = (VF == 1) ? PN->getType() :
03087       VectorType::get(PN->getType(), VF);
03088       Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
03089                                     LoopVectorBody.back()-> getFirstInsertionPt());
03090     }
03091     PV->push_back(P);
03092     return;
03093   }
03094 
03095   setDebugLocFromInst(Builder, P);
03096   // Check for PHI nodes that are lowered to vector selects.
03097   if (P->getParent() != OrigLoop->getHeader()) {
03098     // We know that all PHIs in non-header blocks are converted into
03099     // selects, so we don't have to worry about the insertion order and we
03100     // can just use the builder.
03101     // At this point we generate the predication tree. There may be
03102     // duplications since this is a simple recursive scan, but future
03103     // optimizations will clean it up.
03104 
03105     unsigned NumIncoming = P->getNumIncomingValues();
03106 
03107     // Generate a sequence of selects of the form:
03108     // SELECT(Mask3, In3,
03109     //      SELECT(Mask2, In2,
03110     //                   ( ...)))
03111     for (unsigned In = 0; In < NumIncoming; In++) {
03112       VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
03113                                         P->getParent());
03114       VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
03115 
03116       for (unsigned part = 0; part < UF; ++part) {
03117         // We might have single edge PHIs (blocks) - use an identity
03118         // 'select' for the first PHI operand.
03119         if (In == 0)
03120           Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
03121                                              In0[part]);
03122         else
03123           // Select between the current value and the previous incoming edge
03124           // based on the incoming mask.
03125           Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
03126                                              Entry[part], "predphi");
03127       }
03128     }
03129     return;
03130   }
03131 
03132   // This PHINode must be an induction variable.
03133   // Make sure that we know about it.
03134   assert(Legal->getInductionVars()->count(P) &&
03135          "Not an induction variable");
03136 
03137   LoopVectorizationLegality::InductionInfo II =
03138   Legal->getInductionVars()->lookup(P);
03139 
03140   switch (II.IK) {
03141     case LoopVectorizationLegality::IK_NoInduction:
03142       llvm_unreachable("Unknown induction");
03143     case LoopVectorizationLegality::IK_IntInduction: {
03144       assert(P->getType() == II.StartValue->getType() && "Types must match");
03145       Type *PhiTy = P->getType();
03146       Value *Broadcasted;
03147       if (P == OldInduction) {
03148         // Handle the canonical induction variable. We might have had to
03149         // extend the type.
03150         Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
03151       } else {
03152         // Handle other induction variables that are now based on the
03153         // canonical one.
03154         Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
03155                                                  "normalized.idx");
03156         NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
03157         Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
03158                                         "offset.idx");
03159       }
03160       Broadcasted = getBroadcastInstrs(Broadcasted);
03161       // After broadcasting the induction variable we need to make the vector
03162       // consecutive by adding 0, 1, 2, etc.
03163       for (unsigned part = 0; part < UF; ++part)
03164         Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
03165       return;
03166     }
03167     case LoopVectorizationLegality::IK_ReverseIntInduction:
03168     case LoopVectorizationLegality::IK_PtrInduction:
03169     case LoopVectorizationLegality::IK_ReversePtrInduction:
03170       // Handle reverse integer and pointer inductions.
03171       Value *StartIdx = ExtendedIdx;
03172       // This is the normalized GEP that starts counting at zero.
03173       Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
03174                                                "normalized.idx");
03175 
03176       // Handle the reverse integer induction variable case.
03177       if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
03178         IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
03179         Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
03180                                                "resize.norm.idx");
03181         Value *ReverseInd  = Builder.CreateSub(II.StartValue, CNI,
03182                                                "reverse.idx");
03183 
03184         // This is a new value so do not hoist it out.
03185         Value *Broadcasted = getBroadcastInstrs(ReverseInd);
03186         // After broadcasting the induction variable we need to make the
03187         // vector consecutive by adding  ... -3, -2, -1, 0.
03188         for (unsigned part = 0; part < UF; ++part)
03189           Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
03190                                              true);
03191         return;
03192       }
03193 
03194       // Handle the pointer induction variable case.
03195       assert(P->getType()->isPointerTy() && "Unexpected type.");
03196 
03197       // Is this a reverse induction ptr or a consecutive induction ptr.
03198       bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
03199                       II.IK);
03200 
03201       // This is the vector of results. Notice that we don't generate
03202       // vector geps because scalar geps result in better code.
03203       for (unsigned part = 0; part < UF; ++part) {
03204         if (VF == 1) {
03205           int EltIndex = (part) * (Reverse ? -1 : 1);
03206           Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
03207           Value *GlobalIdx;
03208           if (Reverse)
03209             GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
03210           else
03211             GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
03212 
03213           Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
03214                                              "next.gep");
03215           Entry[part] = SclrGep;
03216           continue;
03217         }
03218 
03219         Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
03220         for (unsigned int i = 0; i < VF; ++i) {
03221           int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
03222           Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
03223           Value *GlobalIdx;
03224           if (!Reverse)
03225             GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
03226           else
03227             GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
03228 
03229           Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
03230                                              "next.gep");
03231           VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
03232                                                Builder.getInt32(i),
03233                                                "insert.gep");
03234         }
03235         Entry[part] = VecVal;
03236       }
03237       return;
03238   }
03239 }
03240 
03241 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
03242   // For each instruction in the old loop.
03243   for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
03244     VectorParts &Entry = WidenMap.get(it);
03245     switch (it->getOpcode()) {
03246     case Instruction::Br:
03247       // Nothing to do for PHIs and BR, since we already took care of the
03248       // loop control flow instructions.
03249       continue;
03250     case Instruction::PHI:{
03251       // Vectorize PHINodes.
03252       widenPHIInstruction(it, Entry, UF, VF, PV);
03253       continue;
03254     }// End of PHI.
03255 
03256     case Instruction::Add:
03257     case Instruction::FAdd:
03258     case Instruction::Sub:
03259     case Instruction::FSub:
03260     case Instruction::Mul:
03261     case Instruction::FMul:
03262     case Instruction::UDiv:
03263     case Instruction::SDiv:
03264     case Instruction::FDiv:
03265     case Instruction::URem:
03266     case Instruction::SRem:
03267     case Instruction::FRem:
03268     case Instruction::Shl:
03269     case Instruction::LShr:
03270     case Instruction::AShr:
03271     case Instruction::And:
03272     case Instruction::Or:
03273     case Instruction::Xor: {
03274       // Just widen binops.
03275       BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
03276       setDebugLocFromInst(Builder, BinOp);
03277       VectorParts &A = getVectorValue(it->getOperand(0));
03278       VectorParts &B = getVectorValue(it->getOperand(1));
03279 
03280       // Use this vector value for all users of the original instruction.
03281       for (unsigned Part = 0; Part < UF; ++Part) {
03282         Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
03283 
03284         if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
03285           VecOp->copyIRFlags(BinOp);
03286 
03287         Entry[Part] = V;
03288       }
03289 
03290       propagateMetadata(Entry, it);
03291       break;
03292     }
03293     case Instruction::Select: {
03294       // Widen selects.
03295       // If the selector is loop invariant we can create a select
03296       // instruction with a scalar condition. Otherwise, use vector-select.
03297       bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
03298                                                OrigLoop);
03299       setDebugLocFromInst(Builder, it);
03300 
03301       // The condition can be loop invariant  but still defined inside the
03302       // loop. This means that we can't just use the original 'cond' value.
03303       // We have to take the 'vectorized' value and pick the first lane.
03304       // Instcombine will make this a no-op.
03305       VectorParts &Cond = getVectorValue(it->getOperand(0));
03306       VectorParts &Op0  = getVectorValue(it->getOperand(1));
03307       VectorParts &Op1  = getVectorValue(it->getOperand(2));
03308 
03309       Value *ScalarCond = (VF == 1) ? Cond[0] :
03310         Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
03311 
03312       for (unsigned Part = 0; Part < UF; ++Part) {
03313         Entry[Part] = Builder.CreateSelect(
03314           InvariantCond ? ScalarCond : Cond[Part],
03315           Op0[Part],
03316           Op1[Part]);
03317       }
03318 
03319       propagateMetadata(Entry, it);
03320       break;
03321     }
03322 
03323     case Instruction::ICmp:
03324     case Instruction::FCmp: {
03325       // Widen compares. Generate vector compares.
03326       bool FCmp = (it->getOpcode() == Instruction::FCmp);
03327       CmpInst *Cmp = dyn_cast<CmpInst>(it);
03328       setDebugLocFromInst(Builder, it);
03329       VectorParts &A = getVectorValue(it->getOperand(0));
03330       VectorParts &B = getVectorValue(it->getOperand(1));
03331       for (unsigned Part = 0; Part < UF; ++Part) {
03332         Value *C = nullptr;
03333         if (FCmp)
03334           C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
03335         else
03336           C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
03337         Entry[Part] = C;
03338       }
03339 
03340       propagateMetadata(Entry, it);
03341       break;
03342     }
03343 
03344     case Instruction::Store:
03345     case Instruction::Load:
03346       vectorizeMemoryInstruction(it);
03347         break;
03348     case Instruction::ZExt:
03349     case Instruction::SExt:
03350     case Instruction::FPToUI:
03351     case Instruction::FPToSI:
03352     case Instruction::FPExt:
03353     case Instruction::PtrToInt:
03354     case Instruction::IntToPtr:
03355     case Instruction::SIToFP:
03356     case Instruction::UIToFP:
03357     case Instruction::Trunc:
03358     case Instruction::FPTrunc:
03359     case Instruction::BitCast: {
03360       CastInst *CI = dyn_cast<CastInst>(it);
03361       setDebugLocFromInst(Builder, it);
03362       /// Optimize the special case where the source is the induction
03363       /// variable. Notice that we can only optimize the 'trunc' case
03364       /// because: a. FP conversions lose precision, b. sext/zext may wrap,
03365       /// c. other casts depend on pointer size.
03366       if (CI->getOperand(0) == OldInduction &&
03367           it->getOpcode() == Instruction::Trunc) {
03368         Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
03369                                                CI->getType());
03370         Value *Broadcasted = getBroadcastInstrs(ScalarCast);
03371         for (unsigned Part = 0; Part < UF; ++Part)
03372           Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
03373         propagateMetadata(Entry, it);
03374         break;
03375       }
03376       /// Vectorize casts.
03377       Type *DestTy = (VF == 1) ? CI->getType() :
03378                                  VectorType::get(CI->getType(), VF);
03379 
03380       VectorParts &A = getVectorValue(it->getOperand(0));
03381       for (unsigned Part = 0; Part < UF; ++Part)
03382         Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
03383       propagateMetadata(Entry, it);
03384       break;
03385     }
03386 
03387     case Instruction::Call: {
03388       // Ignore dbg intrinsics.
03389       if (isa<DbgInfoIntrinsic>(it))
03390         break;
03391       setDebugLocFromInst(Builder, it);
03392 
03393       Module *M = BB->getParent()->getParent();
03394       CallInst *CI = cast<CallInst>(it);
03395       Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
03396       assert(ID && "Not an intrinsic call!");
03397       switch (ID) {
03398       case Intrinsic::assume:
03399       case Intrinsic::lifetime_end:
03400       case Intrinsic::lifetime_start:
03401         scalarizeInstruction(it);
03402         break;
03403       default:
03404         bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
03405         for (unsigned Part = 0; Part < UF; ++Part) {
03406           SmallVector<Value *, 4> Args;
03407           for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
03408             if (HasScalarOpd && i == 1) {
03409               Args.push_back(CI->getArgOperand(i));
03410               continue;
03411             }
03412             VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
03413             Args.push_back(Arg[Part]);
03414           }
03415           Type *Tys[] = {CI->getType()};
03416           if (VF > 1)
03417             Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
03418 
03419           Function *F = Intrinsic::getDeclaration(M, ID, Tys);
03420           Entry[Part] = Builder.CreateCall(F, Args);
03421         }
03422 
03423         propagateMetadata(Entry, it);
03424         break;
03425       }
03426       break;
03427     }
03428 
03429     default:
03430       // All other instructions are unsupported. Scalarize them.
03431       scalarizeInstruction(it);
03432       break;
03433     }// end of switch.
03434   }// end of for_each instr.
03435 }
03436 
03437 void InnerLoopVectorizer::updateAnalysis() {
03438   // Forget the original basic block.
03439   SE->forgetLoop(OrigLoop);
03440 
03441   // Update the dominator tree information.
03442   assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
03443          "Entry does not dominate exit.");
03444 
03445   for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
03446     DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
03447   DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
03448 
03449   // Due to if predication of stores we might create a sequence of "if(pred)
03450   // a[i] = ...;  " blocks.
03451   for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
03452     if (i == 0)
03453       DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
03454     else if (isPredicatedBlock(i)) {
03455       DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
03456     } else {
03457       DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
03458     }
03459   }
03460 
03461   DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
03462   DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
03463   DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
03464   DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
03465 
03466   DEBUG(DT->verifyDomTree());
03467 }
03468 
03469 /// \brief Check whether it is safe to if-convert this phi node.
03470 ///
03471 /// Phi nodes with constant expressions that can trap are not safe to if
03472 /// convert.
03473 static bool canIfConvertPHINodes(BasicBlock *BB) {
03474   for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
03475     PHINode *Phi = dyn_cast<PHINode>(I);
03476     if (!Phi)
03477       return true;
03478     for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
03479       if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
03480         if (C->canTrap())
03481           return false;
03482   }
03483   return true;
03484 }
03485 
03486 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
03487   if (!EnableIfConversion) {
03488     emitAnalysis(Report() << "if-conversion is disabled");
03489     return false;
03490   }
03491 
03492   assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
03493 
03494   // A list of pointers that we can safely read and write to.
03495   SmallPtrSet<Value *, 8> SafePointes;
03496 
03497   // Collect safe addresses.
03498   for (Loop::block_iterator BI = TheLoop->block_begin(),
03499          BE = TheLoop->block_end(); BI != BE; ++BI) {
03500     BasicBlock *BB = *BI;
03501 
03502     if (blockNeedsPredication(BB))
03503       continue;
03504 
03505     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
03506       if (LoadInst *LI = dyn_cast<LoadInst>(I))
03507         SafePointes.insert(LI->getPointerOperand());
03508       else if (StoreInst *SI = dyn_cast<StoreInst>(I))
03509         SafePointes.insert(SI->getPointerOperand());
03510     }
03511   }
03512 
03513   // Collect the blocks that need predication.
03514   BasicBlock *Header = TheLoop->getHeader();
03515   for (Loop::block_iterator BI = TheLoop->block_begin(),
03516          BE = TheLoop->block_end(); BI != BE; ++BI) {
03517     BasicBlock *BB = *BI;
03518 
03519     // We don't support switch statements inside loops.
03520     if (!isa<BranchInst>(BB->getTerminator())) {
03521       emitAnalysis(Report(BB->getTerminator())
03522                    << "loop contains a switch statement");
03523       return false;
03524     }
03525 
03526     // We must be able to predicate all blocks that need to be predicated.
03527     if (blockNeedsPredication(BB)) {
03528       if (!blockCanBePredicated(BB, SafePointes)) {
03529         emitAnalysis(Report(BB->getTerminator())
03530                      << "control flow cannot be substituted for a select");
03531         return false;
03532       }
03533     } else if (BB != Header && !canIfConvertPHINodes(BB)) {
03534       emitAnalysis(Report(BB->getTerminator())
03535                    << "control flow cannot be substituted for a select");
03536       return false;
03537     }
03538   }
03539 
03540   // We can if-convert this loop.
03541   return true;
03542 }
03543 
03544 bool LoopVectorizationLegality::canVectorize() {
03545   // We must have a loop in canonical form. Loops with indirectbr in them cannot
03546   // be canonicalized.
03547   if (!TheLoop->getLoopPreheader()) {
03548     emitAnalysis(
03549         Report() << "loop control flow is not understood by vectorizer");
03550     return false;
03551   }
03552 
03553   // We can only vectorize innermost loops.
03554   if (!TheLoop->getSubLoopsVector().empty()) {
03555     emitAnalysis(Report() << "loop is not the innermost loop");
03556     return false;
03557   }
03558 
03559   // We must have a single backedge.
03560   if (TheLoop->getNumBackEdges() != 1) {
03561     emitAnalysis(
03562         Report() << "loop control flow is not understood by vectorizer");
03563     return false;
03564   }
03565 
03566   // We must have a single exiting block.
03567   if (!TheLoop->getExitingBlock()) {
03568     emitAnalysis(
03569         Report() << "loop control flow is not understood by vectorizer");
03570     return false;
03571   }
03572 
03573   // We only handle bottom-tested loops, i.e. loop in which the condition is
03574   // checked at the end of each iteration. With that we can assume that all
03575   // instructions in the loop are executed the same number of times.
03576   if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
03577     emitAnalysis(
03578         Report() << "loop control flow is not understood by vectorizer");
03579     return false;
03580   }
03581 
03582   // We need to have a loop header.
03583   DEBUG(dbgs() << "LV: Found a loop: " <<
03584         TheLoop->getHeader()->getName() << '\n');
03585 
03586   // Check if we can if-convert non-single-bb loops.
03587   unsigned NumBlocks = TheLoop->getNumBlocks();
03588   if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
03589     DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
03590     return false;
03591   }
03592 
03593   // ScalarEvolution needs to be able to find the exit count.
03594   const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
03595   if (ExitCount == SE->getCouldNotCompute()) {
03596     emitAnalysis(Report() << "could not determine number of loop iterations");
03597     DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
03598     return false;
03599   }
03600 
03601   // Check if we can vectorize the instructions and CFG in this loop.
03602   if (!canVectorizeInstrs()) {
03603     DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
03604     return false;
03605   }
03606 
03607   // Go over each instruction and look at memory deps.
03608   if (!canVectorizeMemory()) {
03609     DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
03610     return false;
03611   }
03612 
03613   // Collect all of the variables that remain uniform after vectorization.
03614   collectLoopUniforms();
03615 
03616   DEBUG(dbgs() << "LV: We can vectorize this loop" <<
03617         (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
03618         <<"!\n");
03619 
03620   // Okay! We can vectorize. At this point we don't have any other mem analysis
03621   // which may limit our maximum vectorization factor, so just return true with
03622   // no restrictions.
03623   return true;
03624 }
03625 
03626 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
03627   if (Ty->isPointerTy())
03628     return DL.getIntPtrType(Ty);
03629 
03630   // It is possible that char's or short's overflow when we ask for the loop's
03631   // trip count, work around this by changing the type size.
03632   if (Ty->getScalarSizeInBits() < 32)
03633     return Type::getInt32Ty(Ty->getContext());
03634 
03635   return Ty;
03636 }
03637 
03638 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
03639   Ty0 = convertPointerToIntegerType(DL, Ty0);
03640   Ty1 = convertPointerToIntegerType(DL, Ty1);
03641   if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
03642     return Ty0;
03643   return Ty1;
03644 }
03645 
03646 /// \brief Check that the instruction has outside loop users and is not an
03647 /// identified reduction variable.
03648 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
03649                                SmallPtrSetImpl<Value *> &Reductions) {
03650   // Reduction instructions are allowed to have exit users. All other
03651   // instructions must not have external users.
03652   if (!Reductions.count(Inst))
03653     //Check that all of the users of the loop are inside the BB.
03654     for (User *U : Inst->users()) {
03655       Instruction *UI = cast<Instruction>(U);
03656       // This user may be a reduction exit value.
03657       if (!TheLoop->contains(UI)) {
03658         DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
03659         return true;
03660       }
03661     }
03662   return false;
03663 }
03664 
03665 bool LoopVectorizationLegality::canVectorizeInstrs() {
03666   BasicBlock *PreHeader = TheLoop->getLoopPreheader();
03667   BasicBlock *Header = TheLoop->getHeader();
03668 
03669   // Look for the attribute signaling the absence of NaNs.
03670   Function &F = *Header->getParent();
03671   if (F.hasFnAttribute("no-nans-fp-math"))
03672     HasFunNoNaNAttr = F.getAttributes().getAttribute(
03673       AttributeSet::FunctionIndex,
03674       "no-nans-fp-math").getValueAsString() == "true";
03675 
03676   // For each block in the loop.
03677   for (Loop::block_iterator bb = TheLoop->block_begin(),
03678        be = TheLoop->block_end(); bb != be; ++bb) {
03679 
03680     // Scan the instructions in the block and look for hazards.
03681     for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
03682          ++it) {
03683 
03684       if (PHINode *Phi = dyn_cast<PHINode>(it)) {
03685         Type *PhiTy = Phi->getType();
03686         // Check that this PHI type is allowed.
03687         if (!PhiTy->isIntegerTy() &&
03688             !PhiTy->isFloatingPointTy() &&
03689             !PhiTy->isPointerTy()) {
03690           emitAnalysis(Report(it)
03691                        << "loop control flow is not understood by vectorizer");
03692           DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
03693           return false;
03694         }
03695 
03696         // If this PHINode is not in the header block, then we know that we
03697         // can convert it to select during if-conversion. No need to check if
03698         // the PHIs in this block are induction or reduction variables.
03699         if (*bb != Header) {
03700           // Check that this instruction has no outside users or is an
03701           // identified reduction value with an outside user.
03702           if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
03703             continue;
03704           emitAnalysis(Report(it) << "value could not be identified as "
03705                                      "an induction or reduction variable");
03706           return false;
03707         }
03708 
03709         // We only allow if-converted PHIs with exactly two incoming values.
03710         if (Phi->getNumIncomingValues() != 2) {
03711           emitAnalysis(Report(it)
03712                        << "control flow not understood by vectorizer");
03713           DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
03714           return false;
03715         }
03716 
03717         // This is the value coming from the preheader.
03718         Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
03719         // Check if this is an induction variable.
03720         InductionKind IK = isInductionVariable(Phi);
03721 
03722         if (IK_NoInduction != IK) {
03723           // Get the widest type.
03724           if (!WidestIndTy)
03725             WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
03726           else
03727             WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
03728 
03729           // Int inductions are special because we only allow one IV.
03730           if (IK == IK_IntInduction) {
03731             // Use the phi node with the widest type as induction. Use the last
03732             // one if there are multiple (no good reason for doing this other
03733             // than it is expedient).
03734             if (!Induction || PhiTy == WidestIndTy)
03735               Induction = Phi;
03736           }
03737 
03738           DEBUG(dbgs() << "LV: Found an induction variable.\n");
03739           Inductions[Phi] = InductionInfo(StartValue, IK);
03740 
03741           // Until we explicitly handle the case of an induction variable with
03742           // an outside loop user we have to give up vectorizing this loop.
03743           if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
03744             emitAnalysis(Report(it) << "use of induction value outside of the "
03745                                        "loop is not handled by vectorizer");
03746             return false;
03747           }
03748 
03749           continue;
03750         }
03751 
03752         if (AddReductionVar(Phi, RK_IntegerAdd)) {
03753           DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
03754           continue;
03755         }
03756         if (AddReductionVar(Phi, RK_IntegerMult)) {
03757           DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
03758           continue;
03759         }
03760         if (AddReductionVar(Phi, RK_IntegerOr)) {
03761           DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
03762           continue;
03763         }
03764         if (AddReductionVar(Phi, RK_IntegerAnd)) {
03765           DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
03766           continue;
03767         }
03768         if (AddReductionVar(Phi, RK_IntegerXor)) {
03769           DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
03770           continue;
03771         }
03772         if (AddReductionVar(Phi, RK_IntegerMinMax)) {
03773           DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
03774           continue;
03775         }
03776         if (AddReductionVar(Phi, RK_FloatMult)) {
03777           DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
03778           continue;
03779         }
03780         if (AddReductionVar(Phi, RK_FloatAdd)) {
03781           DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
03782           continue;
03783         }
03784         if (AddReductionVar(Phi, RK_FloatMinMax)) {
03785           DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
03786                 "\n");
03787           continue;
03788         }
03789 
03790         emitAnalysis(Report(it) << "value that could not be identified as "
03791                                    "reduction is used outside the loop");
03792         DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
03793         return false;
03794       }// end of PHI handling
03795 
03796       // We still don't handle functions. However, we can ignore dbg intrinsic
03797       // calls and we do handle certain intrinsic and libm functions.
03798       CallInst *CI = dyn_cast<CallInst>(it);
03799       if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
03800         emitAnalysis(Report(it) << "call instruction cannot be vectorized");
03801         DEBUG(dbgs() << "LV: Found a call site.\n");
03802         return false;
03803       }
03804 
03805       // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
03806       // second argument is the same (i.e. loop invariant)
03807       if (CI &&
03808           hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
03809         if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
03810           emitAnalysis(Report(it)
03811                        << "intrinsic instruction cannot be vectorized");
03812           DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
03813           return false;
03814         }
03815       }
03816 
03817       // Check that the instruction return type is vectorizable.
03818       // Also, we can't vectorize extractelement instructions.
03819       if ((!VectorType::isValidElementType(it->getType()) &&
03820            !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
03821         emitAnalysis(Report(it)
03822                      << "instruction return type cannot be vectorized");
03823         DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
03824         return false;
03825       }
03826 
03827       // Check that the stored type is vectorizable.
03828       if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
03829         Type *T = ST->getValueOperand()->getType();
03830         if (!VectorType::isValidElementType(T)) {
03831           emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
03832           return false;
03833         }
03834         if (EnableMemAccessVersioning)
03835           collectStridedAccess(ST);
03836       }
03837 
03838       if (EnableMemAccessVersioning)
03839         if (LoadInst *LI = dyn_cast<LoadInst>(it))
03840           collectStridedAccess(LI);
03841 
03842       // Reduction instructions are allowed to have exit users.
03843       // All other instructions must not have external users.
03844       if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
03845         emitAnalysis(Report(it) << "value cannot be used outside the loop");
03846         return false;
03847       }
03848 
03849     } // next instr.
03850 
03851   }
03852 
03853   if (!Induction) {
03854     DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
03855     if (Inductions.empty()) {
03856       emitAnalysis(Report()
03857                    << "loop induction variable could not be identified");
03858       return false;
03859     }
03860   }
03861 
03862   return true;
03863 }
03864 
03865 ///\brief Remove GEPs whose indices but the last one are loop invariant and
03866 /// return the induction operand of the gep pointer.
03867 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
03868                                  const DataLayout *DL, Loop *Lp) {
03869   GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
03870   if (!GEP)
03871     return Ptr;
03872 
03873   unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
03874 
03875   // Check that all of the gep indices are uniform except for our induction
03876   // operand.
03877   for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
03878     if (i != InductionOperand &&
03879         !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
03880       return Ptr;
03881   return GEP->getOperand(InductionOperand);
03882 }
03883 
03884 ///\brief Look for a cast use of the passed value.
03885 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
03886   Value *UniqueCast = nullptr;
03887   for (User *U : Ptr->users()) {
03888     CastInst *CI = dyn_cast<CastInst>(U);
03889     if (CI && CI->getType() == Ty) {
03890       if (!UniqueCast)
03891         UniqueCast = CI;
03892       else
03893         return nullptr;
03894     }
03895   }
03896   return UniqueCast;
03897 }
03898 
03899 ///\brief Get the stride of a pointer access in a loop.
03900 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
03901 /// pointer to the Value, or null otherwise.
03902 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
03903                                    const DataLayout *DL, Loop *Lp) {
03904   const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
03905   if (!PtrTy || PtrTy->isAggregateType())
03906     return nullptr;
03907 
03908   // Try to remove a gep instruction to make the pointer (actually index at this
03909   // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
03910   // pointer, otherwise, we are analyzing the index.
03911   Value *OrigPtr = Ptr;
03912 
03913   // The size of the pointer access.
03914   int64_t PtrAccessSize = 1;
03915 
03916   Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
03917   const SCEV *V = SE->getSCEV(Ptr);
03918 
03919   if (Ptr != OrigPtr)
03920     // Strip off casts.
03921     while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
03922       V = C->getOperand();
03923 
03924   const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
03925   if (!S)
03926     return nullptr;
03927 
03928   V = S->getStepRecurrence(*SE);
03929   if (!V)
03930     return nullptr;
03931 
03932   // Strip off the size of access multiplication if we are still analyzing the
03933   // pointer.
03934   if (OrigPtr == Ptr) {
03935     DL->getTypeAllocSize(PtrTy->getElementType());
03936     if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
03937       if (M->getOperand(0)->getSCEVType() != scConstant)
03938         return nullptr;
03939 
03940       const APInt &APStepVal =
03941           cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
03942 
03943       // Huge step value - give up.
03944       if (APStepVal.getBitWidth() > 64)
03945         return nullptr;
03946 
03947       int64_t StepVal = APStepVal.getSExtValue();
03948       if (PtrAccessSize != StepVal)
03949         return nullptr;
03950       V = M->getOperand(1);
03951     }
03952   }
03953 
03954   // Strip off casts.
03955   Type *StripedOffRecurrenceCast = nullptr;
03956   if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
03957     StripedOffRecurrenceCast = C->getType();
03958     V = C->getOperand();
03959   }
03960 
03961   // Look for the loop invariant symbolic value.
03962   const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
03963   if (!U)
03964     return nullptr;
03965 
03966   Value *Stride = U->getValue();
03967   if (!Lp->isLoopInvariant(Stride))
03968     return nullptr;
03969 
03970   // If we have stripped off the recurrence cast we have to make sure that we
03971   // return the value that is used in this loop so that we can replace it later.
03972   if (StripedOffRecurrenceCast)
03973     Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
03974 
03975   return Stride;
03976 }
03977 
03978 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
03979   Value *Ptr = nullptr;
03980   if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
03981     Ptr = LI->getPointerOperand();
03982   else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
03983     Ptr = SI->getPointerOperand();
03984   else
03985     return;
03986 
03987   Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
03988   if (!Stride)
03989     return;
03990 
03991   DEBUG(dbgs() << "LV: Found a strided access that we can version");
03992   DEBUG(dbgs() << "  Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
03993   Strides[Ptr] = Stride;
03994   StrideSet.insert(Stride);
03995 }
03996 
03997 void LoopVectorizationLegality::collectLoopUniforms() {
03998   // We now know that the loop is vectorizable!
03999   // Collect variables that will remain uniform after vectorization.
04000   std::vector<Value*> Worklist;
04001   BasicBlock *Latch = TheLoop->getLoopLatch();
04002 
04003   // Start with the conditional branch and walk up the block.
04004   Worklist.push_back(Latch->getTerminator()->getOperand(0));
04005 
04006   // Also add all consecutive pointer values; these values will be uniform
04007   // after vectorization (and subsequent cleanup) and, until revectorization is
04008   // supported, all dependencies must also be uniform.
04009   for (Loop::block_iterator B = TheLoop->block_begin(),
04010        BE = TheLoop->block_end(); B != BE; ++B)
04011     for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
04012          I != IE; ++I)
04013       if (I->getType()->isPointerTy() && isConsecutivePtr(I))
04014         Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
04015 
04016   while (!Worklist.empty()) {
04017     Instruction *I = dyn_cast<Instruction>(Worklist.back());
04018     Worklist.pop_back();
04019 
04020     // Look at instructions inside this loop.
04021     // Stop when reaching PHI nodes.
04022     // TODO: we need to follow values all over the loop, not only in this block.
04023     if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
04024       continue;
04025 
04026     // This is a known uniform.
04027     Uniforms.insert(I);
04028 
04029     // Insert all operands.
04030     Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
04031   }
04032 }
04033 
04034 namespace {
04035 /// \brief Analyses memory accesses in a loop.
04036 ///
04037 /// Checks whether run time pointer checks are needed and builds sets for data
04038 /// dependence checking.
04039 class AccessAnalysis {
04040 public:
04041   /// \brief Read or write access location.
04042   typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
04043   typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
04044 
04045   /// \brief Set of potential dependent memory accesses.
04046   typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
04047 
04048   AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
04049     DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
04050 
04051   /// \brief Register a load  and whether it is only read from.
04052   void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
04053     Value *Ptr = const_cast<Value*>(Loc.Ptr);
04054     AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
04055     Accesses.insert(MemAccessInfo(Ptr, false));
04056     if (IsReadOnly)
04057       ReadOnlyPtr.insert(Ptr);
04058   }
04059 
04060   /// \brief Register a store.
04061   void addStore(AliasAnalysis::Location &Loc) {
04062     Value *Ptr = const_cast<Value*>(Loc.Ptr);
04063     AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
04064     Accesses.insert(MemAccessInfo(Ptr, true));
04065   }
04066 
04067   /// \brief Check whether we can check the pointers at runtime for
04068   /// non-intersection.
04069   bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
04070                        unsigned &NumComparisons, ScalarEvolution *SE,
04071                        Loop *TheLoop, ValueToValueMap &Strides,
04072                        bool ShouldCheckStride = false);
04073 
04074   /// \brief Goes over all memory accesses, checks whether a RT check is needed
04075   /// and builds sets of dependent accesses.
04076   void buildDependenceSets() {
04077     processMemAccesses();
04078   }
04079 
04080   bool isRTCheckNeeded() { return IsRTCheckNeeded; }
04081 
04082   bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
04083   void resetDepChecks() { CheckDeps.clear(); }
04084 
04085   MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
04086 
04087 private:
04088   typedef SetVector<MemAccessInfo> PtrAccessSet;
04089 
04090   /// \brief Go over all memory access and check whether runtime pointer checks
04091   /// are needed /// and build sets of dependency check candidates.
04092   void processMemAccesses();
04093 
04094   /// Set of all accesses.
04095   PtrAccessSet Accesses;
04096 
04097   /// Set of accesses that need a further dependence check.
04098   MemAccessInfoSet CheckDeps;
04099 
04100   /// Set of pointers that are read only.
04101   SmallPtrSet<Value*, 16> ReadOnlyPtr;
04102 
04103   const DataLayout *DL;
04104 
04105   /// An alias set tracker to partition the access set by underlying object and
04106   //intrinsic property (such as TBAA metadata).
04107   AliasSetTracker AST;
04108 
04109   /// Sets of potentially dependent accesses - members of one set share an
04110   /// underlying pointer. The set "CheckDeps" identfies which sets really need a
04111   /// dependence check.
04112   DepCandidates &DepCands;
04113 
04114   bool IsRTCheckNeeded;
04115 };
04116 
04117 } // end anonymous namespace
04118 
04119 /// \brief Check whether a pointer can participate in a runtime bounds check.
04120 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
04121                                 Value *Ptr) {
04122   const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
04123   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
04124   if (!AR)
04125     return false;
04126 
04127   return AR->isAffine();
04128 }
04129 
04130 /// \brief Check the stride of the pointer and ensure that it does not wrap in
04131 /// the address space.
04132 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
04133                         const Loop *Lp, ValueToValueMap &StridesMap);
04134 
04135 bool AccessAnalysis::canCheckPtrAtRT(
04136     LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
04137     unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
04138     ValueToValueMap &StridesMap, bool ShouldCheckStride) {
04139   // Find pointers with computable bounds. We are going to use this information
04140   // to place a runtime bound check.
04141   bool CanDoRT = true;
04142 
04143   bool IsDepCheckNeeded = isDependencyCheckNeeded();
04144   NumComparisons = 0;
04145 
04146   // We assign a consecutive id to access from different alias sets.
04147   // Accesses between different groups doesn't need to be checked.
04148   unsigned ASId = 1;
04149   for (auto &AS : AST) {
04150     unsigned NumReadPtrChecks = 0;
04151     unsigned NumWritePtrChecks = 0;
04152 
04153     // We assign consecutive id to access from different dependence sets.
04154     // Accesses within the same set don't need a runtime check.
04155     unsigned RunningDepId = 1;
04156     DenseMap<Value *, unsigned> DepSetId;
04157 
04158     for (auto A : AS) {
04159       Value *Ptr = A.getValue();
04160       bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
04161       MemAccessInfo Access(Ptr, IsWrite);
04162 
04163       if (IsWrite)
04164         ++NumWritePtrChecks;
04165       else
04166         ++NumReadPtrChecks;
04167 
04168       if (hasComputableBounds(SE, StridesMap, Ptr) &&
04169           // When we run after a failing dependency check we have to make sure we
04170           // don't have wrapping pointers.
04171           (!ShouldCheckStride ||
04172            isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
04173         // The id of the dependence set.
04174         unsigned DepId;
04175 
04176         if (IsDepCheckNeeded) {
04177           Value *Leader = DepCands.getLeaderValue(Access).getPointer();
04178           unsigned &LeaderId = DepSetId[Leader];
04179           if (!LeaderId)
04180             LeaderId = RunningDepId++;
04181           DepId = LeaderId;
04182         } else
04183           // Each access has its own dependence set.
04184           DepId = RunningDepId++;
04185 
04186         RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
04187 
04188         DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
04189       } else {
04190         CanDoRT = false;
04191       }
04192     }
04193 
04194     if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
04195       NumComparisons += 0; // Only one dependence set.
04196     else {
04197       NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
04198                                               NumWritePtrChecks - 1));
04199     }
04200 
04201     ++ASId;
04202   }
04203 
04204   // If the pointers that we would use for the bounds comparison have different
04205   // address spaces, assume the values aren't directly comparable, so we can't
04206   // use them for the runtime check. We also have to assume they could
04207   // overlap. In the future there should be metadata for whether address spaces
04208   // are disjoint.
04209   unsigned NumPointers = RtCheck.Pointers.size();
04210   for (unsigned i = 0; i < NumPointers; ++i) {
04211     for (unsigned j = i + 1; j < NumPointers; ++j) {
04212       // Only need to check pointers between two different dependency sets.
04213       if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
04214        continue;
04215       // Only need to check pointers in the same alias set.
04216       if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
04217         continue;
04218 
04219       Value *PtrI = RtCheck.Pointers[i];
04220       Value *PtrJ = RtCheck.Pointers[j];
04221 
04222       unsigned ASi = PtrI->getType()->getPointerAddressSpace();
04223       unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
04224       if (ASi != ASj) {
04225         DEBUG(dbgs() << "LV: Runtime check would require comparison between"
04226                        " different address spaces\n");
04227         return false;
04228       }
04229     }
04230   }
04231 
04232   return CanDoRT;
04233 }
04234 
04235 void AccessAnalysis::processMemAccesses() {
04236   // We process the set twice: first we process read-write pointers, last we
04237   // process read-only pointers. This allows us to skip dependence tests for
04238   // read-only pointers.
04239 
04240   DEBUG(dbgs() << "LV: Processing memory accesses...\n");
04241   DEBUG(dbgs() << "  AST: "; AST.dump());
04242   DEBUG(dbgs() << "LV:   Accesses:\n");
04243   DEBUG({
04244     for (auto A : Accesses)
04245       dbgs() << "\t" << *A.getPointer() << " (" <<
04246                 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
04247                                          "read-only" : "read")) << ")\n";
04248   });
04249 
04250   // The AliasSetTracker has nicely partitioned our pointers by metadata
04251   // compatibility and potential for underlying-object overlap. As a result, we
04252   // only need to check for potential pointer dependencies within each alias
04253   // set.
04254   for (auto &AS : AST) {
04255     // Note that both the alias-set tracker and the alias sets themselves used
04256     // linked lists internally and so the iteration order here is deterministic
04257     // (matching the original instruction order within each set).
04258 
04259     bool SetHasWrite = false;
04260 
04261     // Map of pointers to last access encountered.
04262     typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
04263     UnderlyingObjToAccessMap ObjToLastAccess;
04264 
04265     // Set of access to check after all writes have been processed.
04266     PtrAccessSet DeferredAccesses;
04267 
04268     // Iterate over each alias set twice, once to process read/write pointers,
04269     // and then to process read-only pointers.
04270     for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
04271       bool UseDeferred = SetIteration > 0;
04272       PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
04273 
04274       for (auto AV : AS) {
04275         Value *Ptr = AV.getValue();
04276 
04277         // For a single memory access in AliasSetTracker, Accesses may contain
04278         // both read and write, and they both need to be handled for CheckDeps.
04279         for (auto AC : S) {
04280           if (AC.getPointer() != Ptr)
04281             continue;
04282 
04283           bool IsWrite = AC.getInt();
04284 
04285           // If we're using the deferred access set, then it contains only
04286           // reads.
04287           bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
04288           if (UseDeferred && !IsReadOnlyPtr)
04289             continue;
04290           // Otherwise, the pointer must be in the PtrAccessSet, either as a
04291           // read or a write.
04292           assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
04293                   S.count(MemAccessInfo(Ptr, false))) &&
04294                  "Alias-set pointer not in the access set?");
04295 
04296           MemAccessInfo Access(Ptr, IsWrite);
04297           DepCands.insert(Access);
04298 
04299           // Memorize read-only pointers for later processing and skip them in
04300           // the first round (they need to be checked after we have seen all
04301           // write pointers). Note: we also mark pointer that are not
04302           // consecutive as "read-only" pointers (so that we check
04303           // "a[b[i]] +="). Hence, we need the second check for "!IsWrite".
04304           if (!UseDeferred && IsReadOnlyPtr) {
04305             DeferredAccesses.insert(Access);
04306             continue;
04307           }
04308 
04309           // If this is a write - check other reads and writes for conflicts. If
04310           // this is a read only check other writes for conflicts (but only if
04311           // there is no other write to the ptr - this is an optimization to
04312           // catch "a[i] = a[i] + " without having to do a dependence check).
04313           if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
04314             CheckDeps.insert(Access);
04315             IsRTCheckNeeded = true;
04316           }
04317 
04318           if (IsWrite)
04319             SetHasWrite = true;
04320 
04321           // Create sets of pointers connected by a shared alias set and
04322           // underlying object.
04323           typedef SmallVector<Value *, 16> ValueVector;
04324           ValueVector TempObjects;
04325           GetUnderlyingObjects(Ptr, TempObjects, DL);
04326           for (Value *UnderlyingObj : TempObjects) {
04327             UnderlyingObjToAccessMap::iterator Prev =
04328                 ObjToLastAccess.find(UnderlyingObj);
04329             if (Prev != ObjToLastAccess.end())
04330               DepCands.unionSets(Access, Prev->second);
04331 
04332             ObjToLastAccess[UnderlyingObj] = Access;
04333           }
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   
05355   for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
05356     // Check that we don't have a constant expression that can trap as operand.
05357     for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
05358          OI != OE; ++OI) {
05359       if (Constant *C = dyn_cast<Constant>(*OI))
05360         if (C->canTrap())
05361           return false;
05362     }
05363     // We might be able to hoist the load.
05364     if (it->mayReadFromMemory()) {
05365       LoadInst *LI = dyn_cast<LoadInst>(it);
05366       if (!LI)
05367         return false;
05368       if (!SafePtrs.count(LI->getPointerOperand())) {
05369         if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
05370           MaskedOp.insert(LI);
05371           continue;
05372         }
05373         return false;
05374       }
05375     }
05376 
05377     // We don't predicate stores at the moment.
05378     if (it->mayWriteToMemory()) {
05379       StoreInst *SI = dyn_cast<StoreInst>(it);
05380       // We only support predication of stores in basic blocks with one
05381       // predecessor.
05382       if (!SI)
05383         return false;
05384 
05385       bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
05386       bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
05387       
05388       if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
05389           !isSinglePredecessor) {
05390         // Build a masked store if it is legal for the target, otherwise scalarize
05391         // the block.
05392         bool isLegalMaskedOp =
05393           isLegalMaskedStore(SI->getValueOperand()->getType(),
05394                              SI->getPointerOperand());
05395         if (isLegalMaskedOp) {
05396           --NumPredStores;
05397           MaskedOp.insert(SI);
05398           continue;
05399         }
05400         return false;
05401       }
05402     }
05403     if (it->mayThrow())
05404       return false;
05405 
05406     // The instructions below can trap.
05407     switch (it->getOpcode()) {
05408     default: continue;
05409     case Instruction::UDiv:
05410     case Instruction::SDiv:
05411     case Instruction::URem:
05412     case Instruction::SRem:
05413       return false;
05414     }
05415   }
05416 
05417   return true;
05418 }
05419 
05420 LoopVectorizationCostModel::VectorizationFactor
05421 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
05422   // Width 1 means no vectorize
05423   VectorizationFactor Factor = { 1U, 0U };
05424   if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
05425     emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
05426     DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
05427     return Factor;
05428   }
05429 
05430   if (!EnableCondStoresVectorization && Legal->NumPredStores) {
05431     emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
05432     DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
05433     return Factor;
05434   }
05435 
05436   // Find the trip count.
05437   unsigned TC = SE->getSmallConstantTripCount(TheLoop);
05438   DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
05439 
05440   unsigned WidestType = getWidestType();
05441   unsigned WidestRegister = TTI.getRegisterBitWidth(true);
05442   unsigned MaxSafeDepDist = -1U;
05443   if (Legal->getMaxSafeDepDistBytes() != -1U)
05444     MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
05445   WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
05446                     WidestRegister : MaxSafeDepDist);
05447   unsigned MaxVectorSize = WidestRegister / WidestType;
05448   DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
05449   DEBUG(dbgs() << "LV: The Widest register is: "
05450           << WidestRegister << " bits.\n");
05451 
05452   if (MaxVectorSize == 0) {
05453     DEBUG(dbgs() << "LV: The target has no vector registers.\n");
05454     MaxVectorSize = 1;
05455   }
05456 
05457   assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
05458          " into one vector!");
05459 
05460   unsigned VF = MaxVectorSize;
05461 
05462   // If we optimize the program for size, avoid creating the tail loop.
05463   if (OptForSize) {
05464     // If we are unable to calculate the trip count then don't try to vectorize.
05465     if (TC < 2) {
05466       emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
05467       DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
05468       return Factor;
05469     }
05470 
05471     // Find the maximum SIMD width that can fit within the trip count.
05472     VF = TC % MaxVectorSize;
05473 
05474     if (VF == 0)
05475       VF = MaxVectorSize;
05476 
05477     // If the trip count that we found modulo the vectorization factor is not
05478     // zero then we require a tail.
05479     if (VF < 2) {
05480       emitAnalysis(Report() << "cannot optimize for size and vectorize at the "
05481                                "same time. Enable vectorization of this loop "
05482                                "with '#pragma clang loop vectorize(enable)' "
05483                                "when compiling with -Os");
05484       DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
05485       return Factor;
05486     }
05487   }
05488 
05489   int UserVF = Hints->getWidth();
05490   if (UserVF != 0) {
05491     assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
05492     DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
05493 
05494     Factor.Width = UserVF;
05495     return Factor;
05496   }
05497 
05498   float Cost = expectedCost(1);
05499 #ifndef NDEBUG
05500   const float ScalarCost = Cost;
05501 #endif /* NDEBUG */
05502   unsigned Width = 1;
05503   DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
05504 
05505   bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
05506   // Ignore scalar width, because the user explicitly wants vectorization.
05507   if (ForceVectorization && VF > 1) {
05508     Width = 2;
05509     Cost = expectedCost(Width) / (float)Width;
05510   }
05511 
05512   for (unsigned i=2; i <= VF; i*=2) {
05513     // Notice that the vector loop needs to be executed less times, so
05514     // we need to divide the cost of the vector loops by the width of
05515     // the vector elements.
05516     float VectorCost = expectedCost(i) / (float)i;
05517     DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
05518           (int)VectorCost << ".\n");
05519     if (VectorCost < Cost) {
05520       Cost = VectorCost;
05521       Width = i;
05522     }
05523   }
05524 
05525   DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
05526         << "LV: Vectorization seems to be not beneficial, "
05527         << "but was forced by a user.\n");
05528   DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
05529   Factor.Width = Width;
05530   Factor.Cost = Width * Cost;
05531   return Factor;
05532 }
05533 
05534 unsigned LoopVectorizationCostModel::getWidestType() {
05535   unsigned MaxWidth = 8;
05536 
05537   // For each block.
05538   for (Loop::block_iterator bb = TheLoop->block_begin(),
05539        be = TheLoop->block_end(); bb != be; ++bb) {
05540     BasicBlock *BB = *bb;
05541 
05542     // For each instruction in the loop.
05543     for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
05544       Type *T = it->getType();
05545 
05546       // Ignore ephemeral values.
05547       if (EphValues.count(it))
05548         continue;
05549 
05550       // Only examine Loads, Stores and PHINodes.
05551       if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
05552         continue;
05553 
05554       // Examine PHI nodes that are reduction variables.
05555       if (PHINode *PN = dyn_cast<PHINode>(it))
05556         if (!Legal->getReductionVars()->count(PN))
05557           continue;
05558 
05559       // Examine the stored values.
05560       if (StoreInst *ST = dyn_cast<StoreInst>(it))
05561         T = ST->getValueOperand()->getType();
05562 
05563       // Ignore loaded pointer types and stored pointer types that are not
05564       // consecutive. However, we do want to take consecutive stores/loads of
05565       // pointer vectors into account.
05566       if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
05567         continue;
05568 
05569       MaxWidth = std::max(MaxWidth,
05570                           (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
05571     }
05572   }
05573 
05574   return MaxWidth;
05575 }
05576 
05577 unsigned
05578 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
05579                                                unsigned VF,
05580                                                unsigned LoopCost) {
05581 
05582   // -- The unroll heuristics --
05583   // We unroll the loop in order to expose ILP and reduce the loop overhead.
05584   // There are many micro-architectural considerations that we can't predict
05585   // at this level. For example, frontend pressure (on decode or fetch) due to
05586   // code size, or the number and capabilities of the execution ports.
05587   //
05588   // We use the following heuristics to select the unroll factor:
05589   // 1. If the code has reductions, then we unroll in order to break the cross
05590   // iteration dependency.
05591   // 2. If the loop is really small, then we unroll in order to reduce the loop
05592   // overhead.
05593   // 3. We don't unroll if we think that we will spill registers to memory due
05594   // to the increased register pressure.
05595 
05596   // Use the user preference, unless 'auto' is selected.
05597   int UserUF = Hints->getInterleave();
05598   if (UserUF != 0)
05599     return UserUF;
05600 
05601   // When we optimize for size, we don't unroll.
05602   if (OptForSize)
05603     return 1;
05604 
05605   // We used the distance for the unroll factor.
05606   if (Legal->getMaxSafeDepDistBytes() != -1U)
05607     return 1;
05608 
05609   // Do not unroll loops with a relatively small trip count.
05610   unsigned TC = SE->getSmallConstantTripCount(TheLoop);
05611   if (TC > 1 && TC < TinyTripCountUnrollThreshold)
05612     return 1;
05613 
05614   unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
05615   DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
05616         " registers\n");
05617 
05618   if (VF == 1) {
05619     if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
05620       TargetNumRegisters = ForceTargetNumScalarRegs;
05621   } else {
05622     if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
05623       TargetNumRegisters = ForceTargetNumVectorRegs;
05624   }
05625 
05626   LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
05627   // We divide by these constants so assume that we have at least one
05628   // instruction that uses at least one register.
05629   R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
05630   R.NumInstructions = std::max(R.NumInstructions, 1U);
05631 
05632   // We calculate the unroll factor using the following formula.
05633   // Subtract the number of loop invariants from the number of available
05634   // registers. These registers are used by all of the unrolled instances.
05635   // Next, divide the remaining registers by the number of registers that is
05636   // required by the loop, in order to estimate how many parallel instances
05637   // fit without causing spills. All of this is rounded down if necessary to be
05638   // a power of two. We want power of two unroll factors to simplify any
05639   // addressing operations or alignment considerations.
05640   unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
05641                               R.MaxLocalUsers);
05642 
05643   // Don't count the induction variable as unrolled.
05644   if (EnableIndVarRegisterHeur)
05645     UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
05646                        std::max(1U, (R.MaxLocalUsers - 1)));
05647 
05648   // Clamp the unroll factor ranges to reasonable factors.
05649   unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
05650 
05651   // Check if the user has overridden the unroll max.
05652   if (VF == 1) {
05653     if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
05654       MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
05655   } else {
05656     if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
05657       MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
05658   }
05659 
05660   // If we did not calculate the cost for VF (because the user selected the VF)
05661   // then we calculate the cost of VF here.
05662   if (LoopCost == 0)
05663     LoopCost = expectedCost(VF);
05664 
05665   // Clamp the calculated UF to be between the 1 and the max unroll factor
05666   // that the target allows.
05667   if (UF > MaxInterleaveSize)
05668     UF = MaxInterleaveSize;
05669   else if (UF < 1)
05670     UF = 1;
05671 
05672   // Unroll if we vectorized this loop and there is a reduction that could
05673   // benefit from unrolling.
05674   if (VF > 1 && Legal->getReductionVars()->size()) {
05675     DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
05676     return UF;
05677   }
05678 
05679   // Note that if we've already vectorized the loop we will have done the
05680   // runtime check and so unrolling won't require further checks.
05681   bool UnrollingRequiresRuntimePointerCheck =
05682       (VF == 1 && Legal->getRuntimePointerCheck()->Need);
05683 
05684   // We want to unroll small loops in order to reduce the loop overhead and
05685   // potentially expose ILP opportunities.
05686   DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
05687   if (!UnrollingRequiresRuntimePointerCheck &&
05688       LoopCost < SmallLoopCost) {
05689     // We assume that the cost overhead is 1 and we use the cost model
05690     // to estimate the cost of the loop and unroll until the cost of the
05691     // loop overhead is about 5% of the cost of the loop.
05692     unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
05693 
05694     // Unroll until store/load ports (estimated by max unroll factor) are
05695     // saturated.
05696     unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
05697     unsigned LoadsUF = UF /  (Legal->NumLoads ? Legal->NumLoads : 1);
05698 
05699     // If we have a scalar reduction (vector reductions are already dealt with
05700     // by this point), we can increase the critical path length if the loop
05701     // we're unrolling is inside another loop. Limit, by default to 2, so the
05702     // critical path only gets increased by one reduction operation.
05703     if (Legal->getReductionVars()->size() &&
05704         TheLoop->getLoopDepth() > 1) {
05705       unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
05706       SmallUF = std::min(SmallUF, F);
05707       StoresUF = std::min(StoresUF, F);
05708       LoadsUF = std::min(LoadsUF, F);
05709     }
05710 
05711     if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
05712       DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
05713       return std::max(StoresUF, LoadsUF);
05714     }
05715 
05716     DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
05717     return SmallUF;
05718   }
05719 
05720   DEBUG(dbgs() << "LV: Not Unrolling.\n");
05721   return 1;
05722 }
05723 
05724 LoopVectorizationCostModel::RegisterUsage
05725 LoopVectorizationCostModel::calculateRegisterUsage() {
05726   // This function calculates the register usage by measuring the highest number
05727   // of values that are alive at a single location. Obviously, this is a very
05728   // rough estimation. We scan the loop in a topological order in order and
05729   // assign a number to each instruction. We use RPO to ensure that defs are
05730   // met before their users. We assume that each instruction that has in-loop
05731   // users starts an interval. We record every time that an in-loop value is
05732   // used, so we have a list of the first and last occurrences of each
05733   // instruction. Next, we transpose this data structure into a multi map that
05734   // holds the list of intervals that *end* at a specific location. This multi
05735   // map allows us to perform a linear search. We scan the instructions linearly
05736   // and record each time that a new interval starts, by placing it in a set.
05737   // If we find this value in the multi-map then we remove it from the set.
05738   // The max register usage is the maximum size of the set.
05739   // We also search for instructions that are defined outside the loop, but are
05740   // used inside the loop. We need this number separately from the max-interval
05741   // usage number because when we unroll, loop-invariant values do not take
05742   // more register.
05743   LoopBlocksDFS DFS(TheLoop);
05744   DFS.perform(LI);
05745 
05746   RegisterUsage R;
05747   R.NumInstructions = 0;
05748 
05749   // Each 'key' in the map opens a new interval. The values
05750   // of the map are the index of the 'last seen' usage of the
05751   // instruction that is the key.
05752   typedef DenseMap<Instruction*, unsigned> IntervalMap;
05753   // Maps instruction to its index.
05754   DenseMap<unsigned, Instruction*> IdxToInstr;
05755   // Marks the end of each interval.
05756   IntervalMap EndPoint;
05757   // Saves the list of instruction indices that are used in the loop.
05758   SmallSet<Instruction*, 8> Ends;
05759   // Saves the list of values that are used in the loop but are
05760   // defined outside the loop, such as arguments and constants.
05761   SmallPtrSet<Value*, 8> LoopInvariants;
05762 
05763   unsigned Index = 0;
05764   for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
05765        be = DFS.endRPO(); bb != be; ++bb) {
05766     R.NumInstructions += (*bb)->size();
05767     for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
05768          ++it) {
05769       Instruction *I = it;
05770       IdxToInstr[Index++] = I;
05771 
05772       // Save the end location of each USE.
05773       for (unsigned i = 0; i < I->getNumOperands(); ++i) {
05774         Value *U = I->getOperand(i);
05775         Instruction *Instr = dyn_cast<Instruction>(U);
05776 
05777         // Ignore non-instruction values such as arguments, constants, etc.
05778         if (!Instr) continue;
05779 
05780         // If this instruction is outside the loop then record it and continue.
05781         if (!TheLoop->contains(Instr)) {
05782           LoopInvariants.insert(Instr);
05783           continue;
05784         }
05785 
05786         // Overwrite previous end points.
05787         EndPoint[Instr] = Index;
05788         Ends.insert(Instr);
05789       }
05790     }
05791   }
05792 
05793   // Saves the list of intervals that end with the index in 'key'.
05794   typedef SmallVector<Instruction*, 2> InstrList;
05795   DenseMap<unsigned, InstrList> TransposeEnds;
05796 
05797   // Transpose the EndPoints to a list of values that end at each index.
05798   for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
05799        it != e; ++it)
05800     TransposeEnds[it->second].push_back(it->first);
05801 
05802   SmallSet<Instruction*, 8> OpenIntervals;
05803   unsigned MaxUsage = 0;
05804 
05805 
05806   DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
05807   for (unsigned int i = 0; i < Index; ++i) {
05808     Instruction *I = IdxToInstr[i];
05809     // Ignore instructions that are never used within the loop.
05810     if (!Ends.count(I)) continue;
05811 
05812     // Ignore ephemeral values.
05813     if (EphValues.count(I))
05814       continue;
05815 
05816     // Remove all of the instructions that end at this location.
05817     InstrList &List = TransposeEnds[i];
05818     for (unsigned int j=0, e = List.size(); j < e; ++j)
05819       OpenIntervals.erase(List[j]);
05820 
05821     // Count the number of live interals.
05822     MaxUsage = std::max(MaxUsage, OpenIntervals.size());
05823 
05824     DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
05825           OpenIntervals.size() << '\n');
05826 
05827     // Add the current instruction to the list of open intervals.
05828     OpenIntervals.insert(I);
05829   }
05830 
05831   unsigned Invariant = LoopInvariants.size();
05832   DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
05833   DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
05834   DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
05835 
05836   R.LoopInvariantRegs = Invariant;
05837   R.MaxLocalUsers = MaxUsage;
05838   return R;
05839 }
05840 
05841 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
05842   unsigned Cost = 0;
05843 
05844   // For each block.
05845   for (Loop::block_iterator bb = TheLoop->block_begin(),
05846        be = TheLoop->block_end(); bb != be; ++bb) {
05847     unsigned BlockCost = 0;
05848     BasicBlock *BB = *bb;
05849 
05850     // For each instruction in the old loop.
05851     for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
05852       // Skip dbg intrinsics.
05853       if (isa<DbgInfoIntrinsic>(it))
05854         continue;
05855 
05856       // Ignore ephemeral values.
05857       if (EphValues.count(it))
05858         continue;
05859 
05860       unsigned C = getInstructionCost(it, VF);
05861 
05862       // Check if we should override the cost.
05863       if (ForceTargetInstructionCost.getNumOccurrences() > 0)
05864         C = ForceTargetInstructionCost;
05865 
05866       BlockCost += C;
05867       DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
05868             VF << " For instruction: " << *it << '\n');
05869     }
05870 
05871     // We assume that if-converted blocks have a 50% chance of being executed.
05872     // When the code is scalar then some of the blocks are avoided due to CF.
05873     // When the code is vectorized we execute all code paths.
05874     if (VF == 1 && Legal->blockNeedsPredication(*bb))
05875       BlockCost /= 2;
05876 
05877     Cost += BlockCost;
05878   }
05879 
05880   return Cost;
05881 }
05882 
05883 /// \brief Check whether the address computation for a non-consecutive memory
05884 /// access looks like an unlikely candidate for being merged into the indexing
05885 /// mode.
05886 ///
05887 /// We look for a GEP which has one index that is an induction variable and all
05888 /// other indices are loop invariant. If the stride of this access is also
05889 /// within a small bound we decide that this address computation can likely be
05890 /// merged into the addressing mode.
05891 /// In all other cases, we identify the address computation as complex.
05892 static bool isLikelyComplexAddressComputation(Value *Ptr,
05893                                               LoopVectorizationLegality *Legal,
05894                                               ScalarEvolution *SE,
05895                                               const Loop *TheLoop) {
05896   GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
05897   if (!Gep)
05898     return true;
05899 
05900   // We are looking for a gep with all loop invariant indices except for one
05901   // which should be an induction variable.
05902   unsigned NumOperands = Gep->getNumOperands();
05903   for (unsigned i = 1; i < NumOperands; ++i) {
05904     Value *Opd = Gep->getOperand(i);
05905     if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
05906         !Legal->isInductionVariable(Opd))
05907       return true;
05908   }
05909 
05910   // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
05911   // can likely be merged into the address computation.
05912   unsigned MaxMergeDistance = 64;
05913 
05914   const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
05915   if (!AddRec)
05916     return true;
05917 
05918   // Check the step is constant.
05919   const SCEV *Step = AddRec->getStepRecurrence(*SE);
05920   // Calculate the pointer stride and check if it is consecutive.
05921   const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
05922   if (!C)
05923     return true;
05924 
05925   const APInt &APStepVal = C->getValue()->getValue();
05926 
05927   // Huge step value - give up.
05928   if (APStepVal.getBitWidth() > 64)
05929     return true;
05930 
05931   int64_t StepVal = APStepVal.getSExtValue();
05932 
05933   return StepVal > MaxMergeDistance;
05934 }
05935 
05936 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
05937   if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
05938     return true;
05939   return false;
05940 }
05941 
05942 unsigned
05943 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
05944   // If we know that this instruction will remain uniform, check the cost of
05945   // the scalar version.
05946   if (Legal->isUniformAfterVectorization(I))
05947     VF = 1;
05948 
05949   Type *RetTy = I->getType();
05950   Type *VectorTy = ToVectorTy(RetTy, VF);
05951 
05952   // TODO: We need to estimate the cost of intrinsic calls.
05953   switch (I->getOpcode()) {
05954   case Instruction::GetElementPtr:
05955     // We mark this instruction as zero-cost because the cost of GEPs in
05956     // vectorized code depends on whether the corresponding memory instruction
05957     // is scalarized or not. Therefore, we handle GEPs with the memory
05958     // instruction cost.
05959     return 0;
05960   case Instruction::Br: {
05961     return TTI.getCFInstrCost(I->getOpcode());
05962   }
05963   case Instruction::PHI:
05964     //TODO: IF-converted IFs become selects.
05965     return 0;
05966   case Instruction::Add:
05967   case Instruction::FAdd:
05968   case Instruction::Sub:
05969   case Instruction::FSub:
05970   case Instruction::Mul:
05971   case Instruction::FMul:
05972   case Instruction::UDiv:
05973   case Instruction::SDiv:
05974   case Instruction::FDiv:
05975   case Instruction::URem:
05976   case Instruction::SRem:
05977   case Instruction::FRem:
05978   case Instruction::Shl:
05979   case Instruction::LShr:
05980   case Instruction::AShr:
05981   case Instruction::And:
05982   case Instruction::Or:
05983   case Instruction::Xor: {
05984     // Since we will replace the stride by 1 the multiplication should go away.
05985     if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
05986       return 0;
05987     // Certain instructions can be cheaper to vectorize if they have a constant
05988     // second vector operand. One example of this are shifts on x86.
05989     TargetTransformInfo::OperandValueKind Op1VK =
05990       TargetTransformInfo::OK_AnyValue;
05991     TargetTransformInfo::OperandValueKind Op2VK =
05992       TargetTransformInfo::OK_AnyValue;
05993     TargetTransformInfo::OperandValueProperties Op1VP =
05994         TargetTransformInfo::OP_None;
05995     TargetTransformInfo::OperandValueProperties Op2VP =
05996         TargetTransformInfo::OP_None;
05997     Value *Op2 = I->getOperand(1);
05998 
05999     // Check for a splat of a constant or for a non uniform vector of constants.
06000     if (isa<ConstantInt>(Op2)) {
06001       ConstantInt *CInt = cast<ConstantInt>(Op2);
06002       if (CInt && CInt->getValue().isPowerOf2())
06003         Op2VP = TargetTransformInfo::OP_PowerOf2;
06004       Op2VK = TargetTransformInfo::OK_UniformConstantValue;
06005     } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
06006       Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
06007       Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
06008       if (SplatValue) {
06009         ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
06010         if (CInt && CInt->getValue().isPowerOf2())
06011           Op2VP = TargetTransformInfo::OP_PowerOf2;
06012         Op2VK = TargetTransformInfo::OK_UniformConstantValue;
06013       }
06014     }
06015 
06016     return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
06017                                       Op1VP, Op2VP);
06018   }
06019   case Instruction::Select: {
06020     SelectInst *SI = cast<SelectInst>(I);
06021     const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
06022     bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
06023     Type *CondTy = SI->getCondition()->getType();
06024     if (!ScalarCond)
06025       CondTy = VectorType::get(CondTy, VF);
06026 
06027     return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
06028   }
06029   case Instruction::ICmp:
06030   case Instruction::FCmp: {
06031     Type *ValTy = I->getOperand(0)->getType();
06032     VectorTy = ToVectorTy(ValTy, VF);
06033     return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
06034   }
06035   case Instruction::Store:
06036   case Instruction::Load: {
06037     StoreInst *SI = dyn_cast<StoreInst>(I);
06038     LoadInst *LI = dyn_cast<LoadInst>(I);
06039     Type *ValTy = (SI ? SI->getValueOperand()->getType() :
06040                    LI->getType());
06041     VectorTy = ToVectorTy(ValTy, VF);
06042 
06043     unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
06044     unsigned AS = SI ? SI->getPointerAddressSpace() :
06045       LI->getPointerAddressSpace();
06046     Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
06047     // We add the cost of address computation here instead of with the gep
06048     // instruction because only here we know whether the operation is
06049     // scalarized.
06050     if (VF == 1)
06051       return TTI.getAddressComputationCost(VectorTy) +
06052         TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
06053 
06054     // Scalarized loads/stores.
06055     int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
06056     bool Reverse = ConsecutiveStride < 0;
06057     unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
06058     unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
06059     if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
06060       bool IsComplexComputation =
06061         isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
06062       unsigned Cost = 0;
06063       // The cost of extracting from the value vector and pointer vector.
06064       Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
06065       for (unsigned i = 0; i < VF; ++i) {
06066         //  The cost of extracting the pointer operand.
06067         Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
06068         // In case of STORE, the cost of ExtractElement from the vector.
06069         // In case of LOAD, the cost of InsertElement into the returned
06070         // vector.
06071         Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
06072                                             Instruction::InsertElement,
06073                                             VectorTy, i);
06074       }
06075 
06076       // The cost of the scalar loads/stores.
06077       Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
06078       Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
06079                                        Alignment, AS);
06080       return Cost;
06081     }
06082 
06083     // Wide load/stores.
06084     unsigned Cost = TTI.getAddressComputationCost(VectorTy);
06085     Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
06086 
06087     if (Reverse)
06088       Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
06089                                   VectorTy, 0);
06090     return Cost;
06091   }
06092   case Instruction::ZExt:
06093   case Instruction::SExt:
06094   case Instruction::FPToUI:
06095   case Instruction::FPToSI:
06096   case Instruction::FPExt:
06097   case Instruction::PtrToInt:
06098   case Instruction::IntToPtr:
06099   case Instruction::SIToFP:
06100   case Instruction::UIToFP:
06101   case Instruction::Trunc:
06102   case Instruction::FPTrunc:
06103   case Instruction::BitCast: {
06104     // We optimize the truncation of induction variable.
06105     // The cost of these is the same as the scalar operation.
06106     if (I->getOpcode() == Instruction::Trunc &&
06107         Legal->isInductionVariable(I->getOperand(0)))
06108       return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
06109                                   I->getOperand(0)->getType());
06110 
06111     Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
06112     return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
06113   }
06114   case Instruction::Call: {
06115     CallInst *CI = cast<CallInst>(I);
06116     Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
06117     assert(ID && "Not an intrinsic call!");
06118     Type *RetTy = ToVectorTy(CI->getType(), VF);
06119     SmallVector<Type*, 4> Tys;
06120     for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
06121       Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
06122     return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
06123   }
06124   default: {
06125     // We are scalarizing the instruction. Return the cost of the scalar
06126     // instruction, plus the cost of insert and extract into vector
06127     // elements, times the vector width.
06128     unsigned Cost = 0;
06129 
06130     if (!RetTy->isVoidTy() && VF != 1) {
06131       unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
06132                                                 VectorTy);
06133       unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
06134                                                 VectorTy);
06135 
06136       // The cost of inserting the results plus extracting each one of the
06137       // operands.
06138       Cost += VF * (InsCost + ExtCost * I->getNumOperands());
06139     }
06140 
06141     // The cost of executing VF copies of the scalar instruction. This opcode
06142     // is unknown. Assume that it is the same as 'mul'.
06143     Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
06144     return Cost;
06145   }
06146   }// end of switch.
06147 }
06148 
06149 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
06150   if (Scalar->isVoidTy() || VF == 1)
06151     return Scalar;
06152   return VectorType::get(Scalar, VF);
06153 }
06154 
06155 char LoopVectorize::ID = 0;
06156 static const char lv_name[] = "Loop Vectorization";
06157 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
06158 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
06159 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
06160 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
06161 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
06162 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
06163 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
06164 INITIALIZE_PASS_DEPENDENCY(LCSSA)
06165 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
06166 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
06167 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
06168 
06169 namespace llvm {
06170   Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
06171     return new LoopVectorize(NoUnrolling, AlwaysVectorize);
06172   }
06173 }
06174 
06175 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
06176   // Check for a store.
06177   if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
06178     return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
06179 
06180   // Check for a load.
06181   if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
06182     return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
06183 
06184   return false;
06185 }
06186 
06187 
06188 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
06189                                              bool IfPredicateStore) {
06190   assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
06191   // Holds vector parameters or scalars, in case of uniform vals.
06192   SmallVector<VectorParts, 4> Params;
06193 
06194   setDebugLocFromInst(Builder, Instr);
06195 
06196   // Find all of the vectorized parameters.
06197   for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
06198     Value *SrcOp = Instr->getOperand(op);
06199 
06200     // If we are accessing the old induction variable, use the new one.
06201     if (SrcOp == OldInduction) {
06202       Params.push_back(getVectorValue(SrcOp));
06203       continue;
06204     }
06205 
06206     // Try using previously calculated values.
06207     Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
06208 
06209     // If the src is an instruction that appeared earlier in the basic block
06210     // then it should already be vectorized.
06211     if (SrcInst && OrigLoop->contains(SrcInst)) {
06212       assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
06213       // The parameter is a vector value from earlier.
06214       Params.push_back(WidenMap.get(SrcInst));
06215     } else {
06216       // The parameter is a scalar from outside the loop. Maybe even a constant.
06217       VectorParts Scalars;
06218       Scalars.append(UF, SrcOp);
06219       Params.push_back(Scalars);
06220     }
06221   }
06222 
06223   assert(Params.size() == Instr->getNumOperands() &&
06224          "Invalid number of operands");
06225 
06226   // Does this instruction return a value ?
06227   bool IsVoidRetTy = Instr->getType()->isVoidTy();
06228 
06229   Value *UndefVec = IsVoidRetTy ? nullptr :
06230   UndefValue::get(Instr->getType());
06231   // Create a new entry in the WidenMap and initialize it to Undef or Null.
06232   VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
06233 
06234   Instruction *InsertPt = Builder.GetInsertPoint();
06235   BasicBlock *IfBlock = Builder.GetInsertBlock();
06236   BasicBlock *CondBlock = nullptr;
06237 
06238   VectorParts Cond;
06239   Loop *VectorLp = nullptr;
06240   if (IfPredicateStore) {
06241     assert(Instr->getParent()->getSinglePredecessor() &&
06242            "Only support single predecessor blocks");
06243     Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
06244                           Instr->getParent());
06245     VectorLp = LI->getLoopFor(IfBlock);
06246     assert(VectorLp && "Must have a loop for this block");
06247   }
06248 
06249   // For each vector unroll 'part':
06250   for (unsigned Part = 0; Part < UF; ++Part) {
06251     // For each scalar that we create:
06252 
06253     // Start an "if (pred) a[i] = ..." block.
06254     Value *Cmp = nullptr;
06255     if (IfPredicateStore) {
06256       if (Cond[Part]->getType()->isVectorTy())
06257         Cond[Part] =
06258             Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
06259       Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
06260                                ConstantInt::get(Cond[Part]->getType(), 1));
06261       CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
06262       LoopVectorBody.push_back(CondBlock);
06263       VectorLp->addBasicBlockToLoop(CondBlock, *LI);
06264       // Update Builder with newly created basic block.
06265       Builder.SetInsertPoint(InsertPt);
06266     }
06267 
06268     Instruction *Cloned = Instr->clone();
06269       if (!IsVoidRetTy)
06270         Cloned->setName(Instr->getName() + ".cloned");
06271       // Replace the operands of the cloned instructions with extracted scalars.
06272       for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
06273         Value *Op = Params[op][Part];
06274         Cloned->setOperand(op, Op);
06275       }
06276 
06277       // Place the cloned scalar in the new loop.
06278       Builder.Insert(Cloned);
06279 
06280       // If the original scalar returns a value we need to place it in a vector
06281       // so that future users will be able to use it.
06282       if (!IsVoidRetTy)
06283         VecResults[Part] = Cloned;
06284 
06285     // End if-block.
06286       if (IfPredicateStore) {
06287         BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
06288         LoopVectorBody.push_back(NewIfBlock);
06289         VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
06290         Builder.SetInsertPoint(InsertPt);
06291         Instruction *OldBr = IfBlock->getTerminator();
06292         BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
06293         OldBr->eraseFromParent();
06294         IfBlock = NewIfBlock;
06295       }
06296   }
06297 }
06298 
06299 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
06300   StoreInst *SI = dyn_cast<StoreInst>(Instr);
06301   bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
06302 
06303   return scalarizeInstruction(Instr, IfPredicateStore);
06304 }
06305 
06306 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
06307   return Vec;
06308 }
06309 
06310 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
06311   return V;
06312 }
06313 
06314 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
06315                                                bool Negate) {
06316   // When unrolling and the VF is 1, we only need to add a simple scalar.
06317   Type *ITy = Val->getType();
06318   assert(!ITy->isVectorTy() && "Val must be a scalar");
06319   Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
06320   return Builder.CreateAdd(Val, C, "induction");
06321 }