1 //===- llvm/Analysis/LoopAccessAnalysis.h -----------------------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file defines the interface for the loop memory dependence framework that 11 // was originally developed for the Loop Vectorizer. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #ifndef LLVM_ANALYSIS_LOOPACCESSANALYSIS_H 16 #define LLVM_ANALYSIS_LOOPACCESSANALYSIS_H 17 18 #include "llvm/ADT/EquivalenceClasses.h" 19 #include "llvm/ADT/Optional.h" 20 #include "llvm/ADT/SetVector.h" 21 #include "llvm/Analysis/AliasAnalysis.h" 22 #include "llvm/Analysis/AliasSetTracker.h" 23 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 24 #include "llvm/IR/ValueHandle.h" 25 #include "llvm/Pass.h" 26 #include "llvm/Support/raw_ostream.h" 27 28 namespace llvm { 29 30 class Value; 31 class DataLayout; 32 class ScalarEvolution; 33 class Loop; 34 class SCEV; 35 class SCEVUnionPredicate; 36 class LoopAccessInfo; 37 38 /// Optimization analysis message produced during vectorization. Messages inform 39 /// the user why vectorization did not occur. 40 class LoopAccessReport { 41 std::string Message; 42 const Instruction *Instr; 43 44 protected: 45 LoopAccessReport(const Twine &Message, const Instruction *I) 46 : Message(Message.str()), Instr(I) {} 47 48 public: 49 LoopAccessReport(const Instruction *I = nullptr) : Instr(I) {} 50 51 template <typename A> LoopAccessReport &operator<<(const A &Value) { 52 raw_string_ostream Out(Message); 53 Out << Value; 54 return *this; 55 } 56 57 const Instruction *getInstr() const { return Instr; } 58 59 std::string &str() { return Message; } 60 const std::string &str() const { return Message; } 61 operator Twine() { return Message; } 62 63 /// \brief Emit an analysis note for \p PassName with the debug location from 64 /// the instruction in \p Message if available. Otherwise use the location of 65 /// \p TheLoop. 66 static void emitAnalysis(const LoopAccessReport &Message, 67 const Function *TheFunction, 68 const Loop *TheLoop, 69 const char *PassName); 70 }; 71 72 /// \brief Collection of parameters shared beetween the Loop Vectorizer and the 73 /// Loop Access Analysis. 74 struct VectorizerParams { 75 /// \brief Maximum SIMD width. 76 static const unsigned MaxVectorWidth; 77 78 /// \brief VF as overridden by the user. 79 static unsigned VectorizationFactor; 80 /// \brief Interleave factor as overridden by the user. 81 static unsigned VectorizationInterleave; 82 /// \brief True if force-vector-interleave was specified by the user. 83 static bool isInterleaveForced(); 84 85 /// \\brief When performing memory disambiguation checks at runtime do not 86 /// make more than this number of comparisons. 87 static unsigned RuntimeMemoryCheckThreshold; 88 }; 89 90 /// \brief Checks memory dependences among accesses to the same underlying 91 /// object to determine whether there vectorization is legal or not (and at 92 /// which vectorization factor). 93 /// 94 /// Note: This class will compute a conservative dependence for access to 95 /// different underlying pointers. Clients, such as the loop vectorizer, will 96 /// sometimes deal these potential dependencies by emitting runtime checks. 97 /// 98 /// We use the ScalarEvolution framework to symbolically evalutate access 99 /// functions pairs. Since we currently don't restructure the loop we can rely 100 /// on the program order of memory accesses to determine their safety. 101 /// At the moment we will only deem accesses as safe for: 102 /// * A negative constant distance assuming program order. 103 /// 104 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x; 105 /// a[i] = tmp; y = a[i]; 106 /// 107 /// The latter case is safe because later checks guarantuee that there can't 108 /// be a cycle through a phi node (that is, we check that "x" and "y" is not 109 /// the same variable: a header phi can only be an induction or a reduction, a 110 /// reduction can't have a memory sink, an induction can't have a memory 111 /// source). This is important and must not be violated (or we have to 112 /// resort to checking for cycles through memory). 113 /// 114 /// * A positive constant distance assuming program order that is bigger 115 /// than the biggest memory access. 116 /// 117 /// tmp = a[i] OR b[i] = x 118 /// a[i+2] = tmp y = b[i+2]; 119 /// 120 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively. 121 /// 122 /// * Zero distances and all accesses have the same size. 123 /// 124 class MemoryDepChecker { 125 public: 126 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo; 127 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet; 128 /// \brief Set of potential dependent memory accesses. 129 typedef EquivalenceClasses<MemAccessInfo> DepCandidates; 130 131 /// \brief Dependece between memory access instructions. 132 struct Dependence { 133 /// \brief The type of the dependence. 134 enum DepType { 135 // No dependence. 136 NoDep, 137 // We couldn't determine the direction or the distance. 138 Unknown, 139 // Lexically forward. 140 // 141 // FIXME: If we only have loop-independent forward dependences (e.g. a 142 // read and write of A[i]), LAA will locally deem the dependence "safe" 143 // without querying the MemoryDepChecker. Therefore we can miss 144 // enumerating loop-independent forward dependences in 145 // getDependences. Note that as soon as there are different 146 // indices used to access the same array, the MemoryDepChecker *is* 147 // queried and the dependence list is complete. 148 Forward, 149 // Forward, but if vectorized, is likely to prevent store-to-load 150 // forwarding. 151 ForwardButPreventsForwarding, 152 // Lexically backward. 153 Backward, 154 // Backward, but the distance allows a vectorization factor of 155 // MaxSafeDepDistBytes. 156 BackwardVectorizable, 157 // Same, but may prevent store-to-load forwarding. 158 BackwardVectorizableButPreventsForwarding 159 }; 160 161 /// \brief String version of the types. 162 static const char *DepName[]; 163 164 /// \brief Index of the source of the dependence in the InstMap vector. 165 unsigned Source; 166 /// \brief Index of the destination of the dependence in the InstMap vector. 167 unsigned Destination; 168 /// \brief The type of the dependence. 169 DepType Type; 170 171 Dependence(unsigned Source, unsigned Destination, DepType Type) 172 : Source(Source), Destination(Destination), Type(Type) {} 173 174 /// \brief Return the source instruction of the dependence. 175 Instruction *getSource(const LoopAccessInfo &LAI) const; 176 /// \brief Return the destination instruction of the dependence. 177 Instruction *getDestination(const LoopAccessInfo &LAI) const; 178 179 /// \brief Dependence types that don't prevent vectorization. 180 static bool isSafeForVectorization(DepType Type); 181 182 /// \brief Lexically forward dependence. 183 bool isForward() const; 184 /// \brief Lexically backward dependence. 185 bool isBackward() const; 186 187 /// \brief May be a lexically backward dependence type (includes Unknown). 188 bool isPossiblyBackward() const; 189 190 /// \brief Print the dependence. \p Instr is used to map the instruction 191 /// indices to instructions. 192 void print(raw_ostream &OS, unsigned Depth, 193 const SmallVectorImpl<Instruction *> &Instrs) const; 194 }; 195 196 MemoryDepChecker(PredicatedScalarEvolution &PSE, const Loop *L) 197 : PSE(PSE), InnermostLoop(L), AccessIdx(0), 198 ShouldRetryWithRuntimeCheck(false), SafeForVectorization(true), 199 RecordDependences(true) {} 200 201 /// \brief Register the location (instructions are given increasing numbers) 202 /// of a write access. 203 void addAccess(StoreInst *SI) { 204 Value *Ptr = SI->getPointerOperand(); 205 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx); 206 InstMap.push_back(SI); 207 ++AccessIdx; 208 } 209 210 /// \brief Register the location (instructions are given increasing numbers) 211 /// of a write access. 212 void addAccess(LoadInst *LI) { 213 Value *Ptr = LI->getPointerOperand(); 214 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx); 215 InstMap.push_back(LI); 216 ++AccessIdx; 217 } 218 219 /// \brief Check whether the dependencies between the accesses are safe. 220 /// 221 /// Only checks sets with elements in \p CheckDeps. 222 bool areDepsSafe(DepCandidates &AccessSets, MemAccessInfoSet &CheckDeps, 223 const ValueToValueMap &Strides); 224 225 /// \brief No memory dependence was encountered that would inhibit 226 /// vectorization. 227 bool isSafeForVectorization() const { return SafeForVectorization; } 228 229 /// \brief The maximum number of bytes of a vector register we can vectorize 230 /// the accesses safely with. 231 uint64_t getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; } 232 233 /// \brief In same cases when the dependency check fails we can still 234 /// vectorize the loop with a dynamic array access check. 235 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; } 236 237 /// \brief Returns the memory dependences. If null is returned we exceeded 238 /// the MaxDependences threshold and this information is not 239 /// available. 240 const SmallVectorImpl<Dependence> *getDependences() const { 241 return RecordDependences ? &Dependences : nullptr; 242 } 243 244 void clearDependences() { Dependences.clear(); } 245 246 /// \brief The vector of memory access instructions. The indices are used as 247 /// instruction identifiers in the Dependence class. 248 const SmallVectorImpl<Instruction *> &getMemoryInstructions() const { 249 return InstMap; 250 } 251 252 /// \brief Generate a mapping between the memory instructions and their 253 /// indices according to program order. 254 DenseMap<Instruction *, unsigned> generateInstructionOrderMap() const { 255 DenseMap<Instruction *, unsigned> OrderMap; 256 257 for (unsigned I = 0; I < InstMap.size(); ++I) 258 OrderMap[InstMap[I]] = I; 259 260 return OrderMap; 261 } 262 263 /// \brief Find the set of instructions that read or write via \p Ptr. 264 SmallVector<Instruction *, 4> getInstructionsForAccess(Value *Ptr, 265 bool isWrite) const; 266 267 private: 268 /// A wrapper around ScalarEvolution, used to add runtime SCEV checks, and 269 /// applies dynamic knowledge to simplify SCEV expressions and convert them 270 /// to a more usable form. We need this in case assumptions about SCEV 271 /// expressions need to be made in order to avoid unknown dependences. For 272 /// example we might assume a unit stride for a pointer in order to prove 273 /// that a memory access is strided and doesn't wrap. 274 PredicatedScalarEvolution &PSE; 275 const Loop *InnermostLoop; 276 277 /// \brief Maps access locations (ptr, read/write) to program order. 278 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses; 279 280 /// \brief Memory access instructions in program order. 281 SmallVector<Instruction *, 16> InstMap; 282 283 /// \brief The program order index to be used for the next instruction. 284 unsigned AccessIdx; 285 286 // We can access this many bytes in parallel safely. 287 uint64_t MaxSafeDepDistBytes; 288 289 /// \brief If we see a non-constant dependence distance we can still try to 290 /// vectorize this loop with runtime checks. 291 bool ShouldRetryWithRuntimeCheck; 292 293 /// \brief No memory dependence was encountered that would inhibit 294 /// vectorization. 295 bool SafeForVectorization; 296 297 //// \brief True if Dependences reflects the dependences in the 298 //// loop. If false we exceeded MaxDependences and 299 //// Dependences is invalid. 300 bool RecordDependences; 301 302 /// \brief Memory dependences collected during the analysis. Only valid if 303 /// RecordDependences is true. 304 SmallVector<Dependence, 8> Dependences; 305 306 /// \brief Check whether there is a plausible dependence between the two 307 /// accesses. 308 /// 309 /// Access \p A must happen before \p B in program order. The two indices 310 /// identify the index into the program order map. 311 /// 312 /// This function checks whether there is a plausible dependence (or the 313 /// absence of such can't be proved) between the two accesses. If there is a 314 /// plausible dependence but the dependence distance is bigger than one 315 /// element access it records this distance in \p MaxSafeDepDistBytes (if this 316 /// distance is smaller than any other distance encountered so far). 317 /// Otherwise, this function returns true signaling a possible dependence. 318 Dependence::DepType isDependent(const MemAccessInfo &A, unsigned AIdx, 319 const MemAccessInfo &B, unsigned BIdx, 320 const ValueToValueMap &Strides); 321 322 /// \brief Check whether the data dependence could prevent store-load 323 /// forwarding. 324 /// 325 /// \return false if we shouldn't vectorize at all or avoid larger 326 /// vectorization factors by limiting MaxSafeDepDistBytes. 327 bool couldPreventStoreLoadForward(uint64_t Distance, uint64_t TypeByteSize); 328 }; 329 330 /// \brief Holds information about the memory runtime legality checks to verify 331 /// that a group of pointers do not overlap. 332 class RuntimePointerChecking { 333 public: 334 struct PointerInfo { 335 /// Holds the pointer value that we need to check. 336 TrackingVH<Value> PointerValue; 337 /// Holds the pointer value at the beginning of the loop. 338 const SCEV *Start; 339 /// Holds the pointer value at the end of the loop. 340 const SCEV *End; 341 /// Holds the information if this pointer is used for writing to memory. 342 bool IsWritePtr; 343 /// Holds the id of the set of pointers that could be dependent because of a 344 /// shared underlying object. 345 unsigned DependencySetId; 346 /// Holds the id of the disjoint alias set to which this pointer belongs. 347 unsigned AliasSetId; 348 /// SCEV for the access. 349 const SCEV *Expr; 350 351 PointerInfo(Value *PointerValue, const SCEV *Start, const SCEV *End, 352 bool IsWritePtr, unsigned DependencySetId, unsigned AliasSetId, 353 const SCEV *Expr) 354 : PointerValue(PointerValue), Start(Start), End(End), 355 IsWritePtr(IsWritePtr), DependencySetId(DependencySetId), 356 AliasSetId(AliasSetId), Expr(Expr) {} 357 }; 358 359 RuntimePointerChecking(ScalarEvolution *SE) : Need(false), SE(SE) {} 360 361 /// Reset the state of the pointer runtime information. 362 void reset() { 363 Need = false; 364 Pointers.clear(); 365 Checks.clear(); 366 } 367 368 /// Insert a pointer and calculate the start and end SCEVs. 369 /// We need \p PSE in order to compute the SCEV expression of the pointer 370 /// according to the assumptions that we've made during the analysis. 371 /// The method might also version the pointer stride according to \p Strides, 372 /// and add new predicates to \p PSE. 373 void insert(Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId, 374 unsigned ASId, const ValueToValueMap &Strides, 375 PredicatedScalarEvolution &PSE); 376 377 /// \brief No run-time memory checking is necessary. 378 bool empty() const { return Pointers.empty(); } 379 380 /// A grouping of pointers. A single memcheck is required between 381 /// two groups. 382 struct CheckingPtrGroup { 383 /// \brief Create a new pointer checking group containing a single 384 /// pointer, with index \p Index in RtCheck. 385 CheckingPtrGroup(unsigned Index, RuntimePointerChecking &RtCheck) 386 : RtCheck(RtCheck), High(RtCheck.Pointers[Index].End), 387 Low(RtCheck.Pointers[Index].Start) { 388 Members.push_back(Index); 389 } 390 391 /// \brief Tries to add the pointer recorded in RtCheck at index 392 /// \p Index to this pointer checking group. We can only add a pointer 393 /// to a checking group if we will still be able to get 394 /// the upper and lower bounds of the check. Returns true in case 395 /// of success, false otherwise. 396 bool addPointer(unsigned Index); 397 398 /// Constitutes the context of this pointer checking group. For each 399 /// pointer that is a member of this group we will retain the index 400 /// at which it appears in RtCheck. 401 RuntimePointerChecking &RtCheck; 402 /// The SCEV expression which represents the upper bound of all the 403 /// pointers in this group. 404 const SCEV *High; 405 /// The SCEV expression which represents the lower bound of all the 406 /// pointers in this group. 407 const SCEV *Low; 408 /// Indices of all the pointers that constitute this grouping. 409 SmallVector<unsigned, 2> Members; 410 }; 411 412 /// \brief A memcheck which made up of a pair of grouped pointers. 413 /// 414 /// These *have* to be const for now, since checks are generated from 415 /// CheckingPtrGroups in LAI::addRuntimeChecks which is a const member 416 /// function. FIXME: once check-generation is moved inside this class (after 417 /// the PtrPartition hack is removed), we could drop const. 418 typedef std::pair<const CheckingPtrGroup *, const CheckingPtrGroup *> 419 PointerCheck; 420 421 /// \brief Generate the checks and store it. This also performs the grouping 422 /// of pointers to reduce the number of memchecks necessary. 423 void generateChecks(MemoryDepChecker::DepCandidates &DepCands, 424 bool UseDependencies); 425 426 /// \brief Returns the checks that generateChecks created. 427 const SmallVector<PointerCheck, 4> &getChecks() const { return Checks; } 428 429 /// \brief Decide if we need to add a check between two groups of pointers, 430 /// according to needsChecking. 431 bool needsChecking(const CheckingPtrGroup &M, 432 const CheckingPtrGroup &N) const; 433 434 /// \brief Returns the number of run-time checks required according to 435 /// needsChecking. 436 unsigned getNumberOfChecks() const { return Checks.size(); } 437 438 /// \brief Print the list run-time memory checks necessary. 439 void print(raw_ostream &OS, unsigned Depth = 0) const; 440 441 /// Print \p Checks. 442 void printChecks(raw_ostream &OS, const SmallVectorImpl<PointerCheck> &Checks, 443 unsigned Depth = 0) const; 444 445 /// This flag indicates if we need to add the runtime check. 446 bool Need; 447 448 /// Information about the pointers that may require checking. 449 SmallVector<PointerInfo, 2> Pointers; 450 451 /// Holds a partitioning of pointers into "check groups". 452 SmallVector<CheckingPtrGroup, 2> CheckingGroups; 453 454 /// \brief Check if pointers are in the same partition 455 /// 456 /// \p PtrToPartition contains the partition number for pointers (-1 if the 457 /// pointer belongs to multiple partitions). 458 static bool 459 arePointersInSamePartition(const SmallVectorImpl<int> &PtrToPartition, 460 unsigned PtrIdx1, unsigned PtrIdx2); 461 462 /// \brief Decide whether we need to issue a run-time check for pointer at 463 /// index \p I and \p J to prove their independence. 464 bool needsChecking(unsigned I, unsigned J) const; 465 466 /// \brief Return PointerInfo for pointer at index \p PtrIdx. 467 const PointerInfo &getPointerInfo(unsigned PtrIdx) const { 468 return Pointers[PtrIdx]; 469 } 470 471 private: 472 /// \brief Groups pointers such that a single memcheck is required 473 /// between two different groups. This will clear the CheckingGroups vector 474 /// and re-compute it. We will only group dependecies if \p UseDependencies 475 /// is true, otherwise we will create a separate group for each pointer. 476 void groupChecks(MemoryDepChecker::DepCandidates &DepCands, 477 bool UseDependencies); 478 479 /// Generate the checks and return them. 480 SmallVector<PointerCheck, 4> 481 generateChecks() const; 482 483 /// Holds a pointer to the ScalarEvolution analysis. 484 ScalarEvolution *SE; 485 486 /// \brief Set of run-time checks required to establish independence of 487 /// otherwise may-aliasing pointers in the loop. 488 SmallVector<PointerCheck, 4> Checks; 489 }; 490 491 /// \brief Drive the analysis of memory accesses in the loop 492 /// 493 /// This class is responsible for analyzing the memory accesses of a loop. It 494 /// collects the accesses and then its main helper the AccessAnalysis class 495 /// finds and categorizes the dependences in buildDependenceSets. 496 /// 497 /// For memory dependences that can be analyzed at compile time, it determines 498 /// whether the dependence is part of cycle inhibiting vectorization. This work 499 /// is delegated to the MemoryDepChecker class. 500 /// 501 /// For memory dependences that cannot be determined at compile time, it 502 /// generates run-time checks to prove independence. This is done by 503 /// AccessAnalysis::canCheckPtrAtRT and the checks are maintained by the 504 /// RuntimePointerCheck class. 505 /// 506 /// If pointers can wrap or can't be expressed as affine AddRec expressions by 507 /// ScalarEvolution, we will generate run-time checks by emitting a 508 /// SCEVUnionPredicate. 509 /// 510 /// Checks for both memory dependences and the SCEV predicates contained in the 511 /// PSE must be emitted in order for the results of this analysis to be valid. 512 class LoopAccessInfo { 513 public: 514 LoopAccessInfo(Loop *L, ScalarEvolution *SE, const TargetLibraryInfo *TLI, 515 AliasAnalysis *AA, DominatorTree *DT, LoopInfo *LI); 516 517 // FIXME: 518 // Hack for MSVC 2013 which sems like it can't synthesize this even 519 // with default keyword: 520 // LoopAccessInfo(LoopAccessInfo &&LAI) = default; 521 LoopAccessInfo(LoopAccessInfo &&LAI) 522 : PSE(std::move(LAI.PSE)), PtrRtChecking(std::move(LAI.PtrRtChecking)), 523 DepChecker(std::move(LAI.DepChecker)), TheLoop(LAI.TheLoop), 524 NumLoads(LAI.NumLoads), NumStores(LAI.NumStores), 525 MaxSafeDepDistBytes(LAI.MaxSafeDepDistBytes), CanVecMem(LAI.CanVecMem), 526 StoreToLoopInvariantAddress(LAI.StoreToLoopInvariantAddress), 527 Report(std::move(LAI.Report)), 528 SymbolicStrides(std::move(LAI.SymbolicStrides)), 529 StrideSet(std::move(LAI.StrideSet)) {} 530 // LoopAccessInfo &operator=(LoopAccessInfo &&LAI) = default; 531 LoopAccessInfo &operator=(LoopAccessInfo &&LAI) { 532 assert(this != &LAI); 533 534 PSE = std::move(LAI.PSE); 535 PtrRtChecking = std::move(LAI.PtrRtChecking); 536 DepChecker = std::move(LAI.DepChecker); 537 TheLoop = LAI.TheLoop; 538 NumLoads = LAI.NumLoads; 539 NumStores = LAI.NumStores; 540 MaxSafeDepDistBytes = LAI.MaxSafeDepDistBytes; 541 CanVecMem = LAI.CanVecMem; 542 StoreToLoopInvariantAddress = LAI.StoreToLoopInvariantAddress; 543 Report = std::move(LAI.Report); 544 SymbolicStrides = std::move(LAI.SymbolicStrides); 545 StrideSet = std::move(LAI.StrideSet); 546 return *this; 547 } 548 549 /// Return true we can analyze the memory accesses in the loop and there are 550 /// no memory dependence cycles. 551 bool canVectorizeMemory() const { return CanVecMem; } 552 553 const RuntimePointerChecking *getRuntimePointerChecking() const { 554 return PtrRtChecking.get(); 555 } 556 557 /// \brief Number of memchecks required to prove independence of otherwise 558 /// may-alias pointers. 559 unsigned getNumRuntimePointerChecks() const { 560 return PtrRtChecking->getNumberOfChecks(); 561 } 562 563 /// Return true if the block BB needs to be predicated in order for the loop 564 /// to be vectorized. 565 static bool blockNeedsPredication(BasicBlock *BB, Loop *TheLoop, 566 DominatorTree *DT); 567 568 /// Returns true if the value V is uniform within the loop. 569 bool isUniform(Value *V) const; 570 571 uint64_t getMaxSafeDepDistBytes() const { return MaxSafeDepDistBytes; } 572 unsigned getNumStores() const { return NumStores; } 573 unsigned getNumLoads() const { return NumLoads;} 574 575 /// \brief Add code that checks at runtime if the accessed arrays overlap. 576 /// 577 /// Returns a pair of instructions where the first element is the first 578 /// instruction generated in possibly a sequence of instructions and the 579 /// second value is the final comparator value or NULL if no check is needed. 580 std::pair<Instruction *, Instruction *> 581 addRuntimeChecks(Instruction *Loc) const; 582 583 /// \brief Generete the instructions for the checks in \p PointerChecks. 584 /// 585 /// Returns a pair of instructions where the first element is the first 586 /// instruction generated in possibly a sequence of instructions and the 587 /// second value is the final comparator value or NULL if no check is needed. 588 std::pair<Instruction *, Instruction *> 589 addRuntimeChecks(Instruction *Loc, 590 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> 591 &PointerChecks) const; 592 593 /// \brief The diagnostics report generated for the analysis. E.g. why we 594 /// couldn't analyze the loop. 595 const Optional<LoopAccessReport> &getReport() const { return Report; } 596 597 /// \brief the Memory Dependence Checker which can determine the 598 /// loop-independent and loop-carried dependences between memory accesses. 599 const MemoryDepChecker &getDepChecker() const { return *DepChecker; } 600 601 /// \brief Return the list of instructions that use \p Ptr to read or write 602 /// memory. 603 SmallVector<Instruction *, 4> getInstructionsForAccess(Value *Ptr, 604 bool isWrite) const { 605 return DepChecker->getInstructionsForAccess(Ptr, isWrite); 606 } 607 608 /// \brief If an access has a symbolic strides, this maps the pointer value to 609 /// the stride symbol. 610 const ValueToValueMap &getSymbolicStrides() const { return SymbolicStrides; } 611 612 /// \brief Pointer has a symbolic stride. 613 bool hasStride(Value *V) const { return StrideSet.count(V); } 614 615 /// \brief Print the information about the memory accesses in the loop. 616 void print(raw_ostream &OS, unsigned Depth = 0) const; 617 618 /// \brief Checks existence of store to invariant address inside loop. 619 /// If the loop has any store to invariant address, then it returns true, 620 /// else returns false. 621 bool hasStoreToLoopInvariantAddress() const { 622 return StoreToLoopInvariantAddress; 623 } 624 625 /// Used to add runtime SCEV checks. Simplifies SCEV expressions and converts 626 /// them to a more usable form. All SCEV expressions during the analysis 627 /// should be re-written (and therefore simplified) according to PSE. 628 /// A user of LoopAccessAnalysis will need to emit the runtime checks 629 /// associated with this predicate. 630 const PredicatedScalarEvolution &getPSE() const { return *PSE; } 631 632 private: 633 /// \brief Analyze the loop. 634 void analyzeLoop(AliasAnalysis *AA, LoopInfo *LI, 635 const TargetLibraryInfo *TLI, DominatorTree *DT); 636 637 /// \brief Check if the structure of the loop allows it to be analyzed by this 638 /// pass. 639 bool canAnalyzeLoop(); 640 641 void emitAnalysis(LoopAccessReport &Message); 642 643 /// \brief Collect memory access with loop invariant strides. 644 /// 645 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop 646 /// invariant. 647 void collectStridedAccess(Value *LoadOrStoreInst); 648 649 std::unique_ptr<PredicatedScalarEvolution> PSE; 650 651 /// We need to check that all of the pointers in this list are disjoint 652 /// at runtime. Using std::unique_ptr to make using move ctor simpler. 653 std::unique_ptr<RuntimePointerChecking> PtrRtChecking; 654 655 /// \brief the Memory Dependence Checker which can determine the 656 /// loop-independent and loop-carried dependences between memory accesses. 657 std::unique_ptr<MemoryDepChecker> DepChecker; 658 659 Loop *TheLoop; 660 661 unsigned NumLoads; 662 unsigned NumStores; 663 664 uint64_t MaxSafeDepDistBytes; 665 666 /// \brief Cache the result of analyzeLoop. 667 bool CanVecMem; 668 669 /// \brief Indicator for storing to uniform addresses. 670 /// If a loop has write to a loop invariant address then it should be true. 671 bool StoreToLoopInvariantAddress; 672 673 /// \brief The diagnostics report generated for the analysis. E.g. why we 674 /// couldn't analyze the loop. 675 Optional<LoopAccessReport> Report; 676 677 /// \brief If an access has a symbolic strides, this maps the pointer value to 678 /// the stride symbol. 679 ValueToValueMap SymbolicStrides; 680 681 /// \brief Set of symbolic strides values. 682 SmallPtrSet<Value *, 8> StrideSet; 683 }; 684 685 Value *stripIntegerCast(Value *V); 686 687 /// \brief Return the SCEV corresponding to a pointer with the symbolic stride 688 /// replaced with constant one, assuming the SCEV predicate associated with 689 /// \p PSE is true. 690 /// 691 /// If necessary this method will version the stride of the pointer according 692 /// to \p PtrToStride and therefore add further predicates to \p PSE. 693 /// 694 /// If \p OrigPtr is not null, use it to look up the stride value instead of \p 695 /// Ptr. \p PtrToStride provides the mapping between the pointer value and its 696 /// stride as collected by LoopVectorizationLegality::collectStridedAccess. 697 const SCEV *replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE, 698 const ValueToValueMap &PtrToStride, 699 Value *Ptr, Value *OrigPtr = nullptr); 700 701 /// \brief If the pointer has a constant stride return it in units of its 702 /// element size. Otherwise return zero. 703 /// 704 /// Ensure that it does not wrap in the address space, assuming the predicate 705 /// associated with \p PSE is true. 706 /// 707 /// If necessary this method will version the stride of the pointer according 708 /// to \p PtrToStride and therefore add further predicates to \p PSE. 709 /// The \p Assume parameter indicates if we are allowed to make additional 710 /// run-time assumptions. 711 int64_t getPtrStride(PredicatedScalarEvolution &PSE, Value *Ptr, const Loop *Lp, 712 const ValueToValueMap &StridesMap = ValueToValueMap(), 713 bool Assume = false); 714 715 /// \brief Returns true if the memory operations \p A and \p B are consecutive. 716 /// This is a simple API that does not depend on the analysis pass. 717 bool isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL, 718 ScalarEvolution &SE, bool CheckType = true); 719 720 /// \brief This analysis provides dependence information for the memory accesses 721 /// of a loop. 722 /// 723 /// It runs the analysis for a loop on demand. This can be initiated by 724 /// querying the loop access info via LAA::getInfo. getInfo return a 725 /// LoopAccessInfo object. See this class for the specifics of what information 726 /// is provided. 727 class LoopAccessLegacyAnalysis : public FunctionPass { 728 public: 729 static char ID; 730 731 LoopAccessLegacyAnalysis() : FunctionPass(ID) { 732 initializeLoopAccessLegacyAnalysisPass(*PassRegistry::getPassRegistry()); 733 } 734 735 bool runOnFunction(Function &F) override; 736 737 void getAnalysisUsage(AnalysisUsage &AU) const override; 738 739 /// \brief Query the result of the loop access information for the loop \p L. 740 /// 741 /// If there is no cached result available run the analysis. 742 const LoopAccessInfo &getInfo(Loop *L); 743 744 void releaseMemory() override { 745 // Invalidate the cache when the pass is freed. 746 LoopAccessInfoMap.clear(); 747 } 748 749 /// \brief Print the result of the analysis when invoked with -analyze. 750 void print(raw_ostream &OS, const Module *M = nullptr) const override; 751 752 private: 753 /// \brief The cache. 754 DenseMap<Loop *, std::unique_ptr<LoopAccessInfo>> LoopAccessInfoMap; 755 756 // The used analysis passes. 757 ScalarEvolution *SE; 758 const TargetLibraryInfo *TLI; 759 AliasAnalysis *AA; 760 DominatorTree *DT; 761 LoopInfo *LI; 762 }; 763 764 /// \brief This analysis provides dependence information for the memory 765 /// accesses of a loop. 766 /// 767 /// It runs the analysis for a loop on demand. This can be initiated by 768 /// querying the loop access info via AM.getResult<LoopAccessAnalysis>. 769 /// getResult return a LoopAccessInfo object. See this class for the 770 /// specifics of what information is provided. 771 class LoopAccessAnalysis 772 : public AnalysisInfoMixin<LoopAccessAnalysis> { 773 friend AnalysisInfoMixin<LoopAccessAnalysis>; 774 static char PassID; 775 776 public: 777 typedef LoopAccessInfo Result; 778 Result run(Loop &, AnalysisManager<Loop> &); 779 static StringRef name() { return "LoopAccessAnalysis"; } 780 }; 781 782 /// \brief Printer pass for the \c LoopAccessInfo results. 783 class LoopAccessInfoPrinterPass 784 : public PassInfoMixin<LoopAccessInfoPrinterPass> { 785 raw_ostream &OS; 786 787 public: 788 explicit LoopAccessInfoPrinterPass(raw_ostream &OS) : OS(OS) {} 789 PreservedAnalyses run(Loop &L, AnalysisManager<Loop> &AM); 790 }; 791 792 inline Instruction *MemoryDepChecker::Dependence::getSource( 793 const LoopAccessInfo &LAI) const { 794 return LAI.getDepChecker().getMemoryInstructions()[Source]; 795 } 796 797 inline Instruction *MemoryDepChecker::Dependence::getDestination( 798 const LoopAccessInfo &LAI) const { 799 return LAI.getDepChecker().getMemoryInstructions()[Destination]; 800 } 801 802 } // End llvm namespace 803 804 #endif 805