1 //===- llvm/Analysis/LoopAccessAnalysis.h -----------------------*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file defines the interface for the loop memory dependence framework that 10 // was originally developed for the Loop Vectorizer. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef LLVM_ANALYSIS_LOOPACCESSANALYSIS_H 15 #define LLVM_ANALYSIS_LOOPACCESSANALYSIS_H 16 17 #include "llvm/ADT/EquivalenceClasses.h" 18 #include "llvm/Analysis/LoopAnalysisManager.h" 19 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 20 #include "llvm/IR/DiagnosticInfo.h" 21 #include "llvm/Pass.h" 22 #include <optional> 23 24 namespace llvm { 25 26 class AAResults; 27 class DataLayout; 28 class Loop; 29 class LoopAccessInfo; 30 class raw_ostream; 31 class SCEV; 32 class SCEVUnionPredicate; 33 class Value; 34 35 /// Collection of parameters shared beetween the Loop Vectorizer and the 36 /// Loop Access Analysis. 37 struct VectorizerParams { 38 /// Maximum SIMD width. 39 static const unsigned MaxVectorWidth; 40 41 /// VF as overridden by the user. 42 static unsigned VectorizationFactor; 43 /// Interleave factor as overridden by the user. 44 static unsigned VectorizationInterleave; 45 /// True if force-vector-interleave was specified by the user. 46 static bool isInterleaveForced(); 47 48 /// \When performing memory disambiguation checks at runtime do not 49 /// make more than this number of comparisons. 50 static unsigned RuntimeMemoryCheckThreshold; 51 }; 52 53 /// Checks memory dependences among accesses to the same underlying 54 /// object to determine whether there vectorization is legal or not (and at 55 /// which vectorization factor). 56 /// 57 /// Note: This class will compute a conservative dependence for access to 58 /// different underlying pointers. Clients, such as the loop vectorizer, will 59 /// sometimes deal these potential dependencies by emitting runtime checks. 60 /// 61 /// We use the ScalarEvolution framework to symbolically evalutate access 62 /// functions pairs. Since we currently don't restructure the loop we can rely 63 /// on the program order of memory accesses to determine their safety. 64 /// At the moment we will only deem accesses as safe for: 65 /// * A negative constant distance assuming program order. 66 /// 67 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x; 68 /// a[i] = tmp; y = a[i]; 69 /// 70 /// The latter case is safe because later checks guarantuee that there can't 71 /// be a cycle through a phi node (that is, we check that "x" and "y" is not 72 /// the same variable: a header phi can only be an induction or a reduction, a 73 /// reduction can't have a memory sink, an induction can't have a memory 74 /// source). This is important and must not be violated (or we have to 75 /// resort to checking for cycles through memory). 76 /// 77 /// * A positive constant distance assuming program order that is bigger 78 /// than the biggest memory access. 79 /// 80 /// tmp = a[i] OR b[i] = x 81 /// a[i+2] = tmp y = b[i+2]; 82 /// 83 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively. 84 /// 85 /// * Zero distances and all accesses have the same size. 86 /// 87 class MemoryDepChecker { 88 public: 89 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo; 90 typedef SmallVector<MemAccessInfo, 8> MemAccessInfoList; 91 /// Set of potential dependent memory accesses. 92 typedef EquivalenceClasses<MemAccessInfo> DepCandidates; 93 94 /// Type to keep track of the status of the dependence check. The order of 95 /// the elements is important and has to be from most permissive to least 96 /// permissive. 97 enum class VectorizationSafetyStatus { 98 // Can vectorize safely without RT checks. All dependences are known to be 99 // safe. 100 Safe, 101 // Can possibly vectorize with RT checks to overcome unknown dependencies. 102 PossiblySafeWithRtChecks, 103 // Cannot vectorize due to known unsafe dependencies. 104 Unsafe, 105 }; 106 107 /// Dependece between memory access instructions. 108 struct Dependence { 109 /// The type of the dependence. 110 enum DepType { 111 // No dependence. 112 NoDep, 113 // We couldn't determine the direction or the distance. 114 Unknown, 115 // Lexically forward. 116 // 117 // FIXME: If we only have loop-independent forward dependences (e.g. a 118 // read and write of A[i]), LAA will locally deem the dependence "safe" 119 // without querying the MemoryDepChecker. Therefore we can miss 120 // enumerating loop-independent forward dependences in 121 // getDependences. Note that as soon as there are different 122 // indices used to access the same array, the MemoryDepChecker *is* 123 // queried and the dependence list is complete. 124 Forward, 125 // Forward, but if vectorized, is likely to prevent store-to-load 126 // forwarding. 127 ForwardButPreventsForwarding, 128 // Lexically backward. 129 Backward, 130 // Backward, but the distance allows a vectorization factor of 131 // MaxSafeDepDistBytes. 132 BackwardVectorizable, 133 // Same, but may prevent store-to-load forwarding. 134 BackwardVectorizableButPreventsForwarding 135 }; 136 137 /// String version of the types. 138 static const char *DepName[]; 139 140 /// Index of the source of the dependence in the InstMap vector. 141 unsigned Source; 142 /// Index of the destination of the dependence in the InstMap vector. 143 unsigned Destination; 144 /// The type of the dependence. 145 DepType Type; 146 147 Dependence(unsigned Source, unsigned Destination, DepType Type) 148 : Source(Source), Destination(Destination), Type(Type) {} 149 150 /// Return the source instruction of the dependence. 151 Instruction *getSource(const LoopAccessInfo &LAI) const; 152 /// Return the destination instruction of the dependence. 153 Instruction *getDestination(const LoopAccessInfo &LAI) const; 154 155 /// Dependence types that don't prevent vectorization. 156 static VectorizationSafetyStatus isSafeForVectorization(DepType Type); 157 158 /// Lexically forward dependence. 159 bool isForward() const; 160 /// Lexically backward dependence. 161 bool isBackward() const; 162 163 /// May be a lexically backward dependence type (includes Unknown). 164 bool isPossiblyBackward() const; 165 166 /// Print the dependence. \p Instr is used to map the instruction 167 /// indices to instructions. 168 void print(raw_ostream &OS, unsigned Depth, 169 const SmallVectorImpl<Instruction *> &Instrs) const; 170 }; 171 172 MemoryDepChecker(PredicatedScalarEvolution &PSE, const Loop *L) 173 : PSE(PSE), InnermostLoop(L) {} 174 175 /// Register the location (instructions are given increasing numbers) 176 /// of a write access. 177 void addAccess(StoreInst *SI); 178 179 /// Register the location (instructions are given increasing numbers) 180 /// of a write access. 181 void addAccess(LoadInst *LI); 182 183 /// Check whether the dependencies between the accesses are safe. 184 /// 185 /// Only checks sets with elements in \p CheckDeps. 186 bool areDepsSafe(DepCandidates &AccessSets, MemAccessInfoList &CheckDeps, 187 const DenseMap<Value *, const SCEV *> &Strides); 188 189 /// No memory dependence was encountered that would inhibit 190 /// vectorization. 191 bool isSafeForVectorization() const { 192 return Status == VectorizationSafetyStatus::Safe; 193 } 194 195 /// Return true if the number of elements that are safe to operate on 196 /// simultaneously is not bounded. 197 bool isSafeForAnyVectorWidth() const { 198 return MaxSafeVectorWidthInBits == UINT_MAX; 199 } 200 201 /// The maximum number of bytes of a vector register we can vectorize 202 /// the accesses safely with. 203 uint64_t getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; } 204 205 /// Return the number of elements that are safe to operate on 206 /// simultaneously, multiplied by the size of the element in bits. 207 uint64_t getMaxSafeVectorWidthInBits() const { 208 return MaxSafeVectorWidthInBits; 209 } 210 211 /// In same cases when the dependency check fails we can still 212 /// vectorize the loop with a dynamic array access check. 213 bool shouldRetryWithRuntimeCheck() const { 214 return FoundNonConstantDistanceDependence && 215 Status == VectorizationSafetyStatus::PossiblySafeWithRtChecks; 216 } 217 218 /// Returns the memory dependences. If null is returned we exceeded 219 /// the MaxDependences threshold and this information is not 220 /// available. 221 const SmallVectorImpl<Dependence> *getDependences() const { 222 return RecordDependences ? &Dependences : nullptr; 223 } 224 225 void clearDependences() { Dependences.clear(); } 226 227 /// The vector of memory access instructions. The indices are used as 228 /// instruction identifiers in the Dependence class. 229 const SmallVectorImpl<Instruction *> &getMemoryInstructions() const { 230 return InstMap; 231 } 232 233 /// Generate a mapping between the memory instructions and their 234 /// indices according to program order. 235 DenseMap<Instruction *, unsigned> generateInstructionOrderMap() const { 236 DenseMap<Instruction *, unsigned> OrderMap; 237 238 for (unsigned I = 0; I < InstMap.size(); ++I) 239 OrderMap[InstMap[I]] = I; 240 241 return OrderMap; 242 } 243 244 /// Find the set of instructions that read or write via \p Ptr. 245 SmallVector<Instruction *, 4> getInstructionsForAccess(Value *Ptr, 246 bool isWrite) const; 247 248 /// Return the program order indices for the access location (Ptr, IsWrite). 249 /// Returns an empty ArrayRef if there are no accesses for the location. 250 ArrayRef<unsigned> getOrderForAccess(Value *Ptr, bool IsWrite) const { 251 auto I = Accesses.find({Ptr, IsWrite}); 252 if (I != Accesses.end()) 253 return I->second; 254 return {}; 255 } 256 257 const Loop *getInnermostLoop() const { return InnermostLoop; } 258 259 private: 260 /// A wrapper around ScalarEvolution, used to add runtime SCEV checks, and 261 /// applies dynamic knowledge to simplify SCEV expressions and convert them 262 /// to a more usable form. We need this in case assumptions about SCEV 263 /// expressions need to be made in order to avoid unknown dependences. For 264 /// example we might assume a unit stride for a pointer in order to prove 265 /// that a memory access is strided and doesn't wrap. 266 PredicatedScalarEvolution &PSE; 267 const Loop *InnermostLoop; 268 269 /// Maps access locations (ptr, read/write) to program order. 270 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses; 271 272 /// Memory access instructions in program order. 273 SmallVector<Instruction *, 16> InstMap; 274 275 /// The program order index to be used for the next instruction. 276 unsigned AccessIdx = 0; 277 278 // We can access this many bytes in parallel safely. 279 uint64_t MaxSafeDepDistBytes = 0; 280 281 /// Number of elements (from consecutive iterations) that are safe to 282 /// operate on simultaneously, multiplied by the size of the element in bits. 283 /// The size of the element is taken from the memory access that is most 284 /// restrictive. 285 uint64_t MaxSafeVectorWidthInBits = -1U; 286 287 /// If we see a non-constant dependence distance we can still try to 288 /// vectorize this loop with runtime checks. 289 bool FoundNonConstantDistanceDependence = false; 290 291 /// Result of the dependence checks, indicating whether the checked 292 /// dependences are safe for vectorization, require RT checks or are known to 293 /// be unsafe. 294 VectorizationSafetyStatus Status = VectorizationSafetyStatus::Safe; 295 296 //// True if Dependences reflects the dependences in the 297 //// loop. If false we exceeded MaxDependences and 298 //// Dependences is invalid. 299 bool RecordDependences = true; 300 301 /// Memory dependences collected during the analysis. Only valid if 302 /// RecordDependences is true. 303 SmallVector<Dependence, 8> Dependences; 304 305 /// Check whether there is a plausible dependence between the two 306 /// accesses. 307 /// 308 /// Access \p A must happen before \p B in program order. The two indices 309 /// identify the index into the program order map. 310 /// 311 /// This function checks whether there is a plausible dependence (or the 312 /// absence of such can't be proved) between the two accesses. If there is a 313 /// plausible dependence but the dependence distance is bigger than one 314 /// element access it records this distance in \p MaxSafeDepDistBytes (if this 315 /// distance is smaller than any other distance encountered so far). 316 /// Otherwise, this function returns true signaling a possible dependence. 317 Dependence::DepType isDependent(const MemAccessInfo &A, unsigned AIdx, 318 const MemAccessInfo &B, unsigned BIdx, 319 const DenseMap<Value *, const SCEV *> &Strides); 320 321 /// Check whether the data dependence could prevent store-load 322 /// forwarding. 323 /// 324 /// \return false if we shouldn't vectorize at all or avoid larger 325 /// vectorization factors by limiting MaxSafeDepDistBytes. 326 bool couldPreventStoreLoadForward(uint64_t Distance, uint64_t TypeByteSize); 327 328 /// Updates the current safety status with \p S. We can go from Safe to 329 /// either PossiblySafeWithRtChecks or Unsafe and from 330 /// PossiblySafeWithRtChecks to Unsafe. 331 void mergeInStatus(VectorizationSafetyStatus S); 332 }; 333 334 class RuntimePointerChecking; 335 /// A grouping of pointers. A single memcheck is required between 336 /// two groups. 337 struct RuntimeCheckingPtrGroup { 338 /// Create a new pointer checking group containing a single 339 /// pointer, with index \p Index in RtCheck. 340 RuntimeCheckingPtrGroup(unsigned Index, RuntimePointerChecking &RtCheck); 341 342 /// Tries to add the pointer recorded in RtCheck at index 343 /// \p Index to this pointer checking group. We can only add a pointer 344 /// to a checking group if we will still be able to get 345 /// the upper and lower bounds of the check. Returns true in case 346 /// of success, false otherwise. 347 bool addPointer(unsigned Index, RuntimePointerChecking &RtCheck); 348 bool addPointer(unsigned Index, const SCEV *Start, const SCEV *End, 349 unsigned AS, bool NeedsFreeze, ScalarEvolution &SE); 350 351 /// The SCEV expression which represents the upper bound of all the 352 /// pointers in this group. 353 const SCEV *High; 354 /// The SCEV expression which represents the lower bound of all the 355 /// pointers in this group. 356 const SCEV *Low; 357 /// Indices of all the pointers that constitute this grouping. 358 SmallVector<unsigned, 2> Members; 359 /// Address space of the involved pointers. 360 unsigned AddressSpace; 361 /// Whether the pointer needs to be frozen after expansion, e.g. because it 362 /// may be poison outside the loop. 363 bool NeedsFreeze = false; 364 }; 365 366 /// A memcheck which made up of a pair of grouped pointers. 367 typedef std::pair<const RuntimeCheckingPtrGroup *, 368 const RuntimeCheckingPtrGroup *> 369 RuntimePointerCheck; 370 371 struct PointerDiffInfo { 372 const SCEV *SrcStart; 373 const SCEV *SinkStart; 374 unsigned AccessSize; 375 bool NeedsFreeze; 376 377 PointerDiffInfo(const SCEV *SrcStart, const SCEV *SinkStart, 378 unsigned AccessSize, bool NeedsFreeze) 379 : SrcStart(SrcStart), SinkStart(SinkStart), AccessSize(AccessSize), 380 NeedsFreeze(NeedsFreeze) {} 381 }; 382 383 /// Holds information about the memory runtime legality checks to verify 384 /// that a group of pointers do not overlap. 385 class RuntimePointerChecking { 386 friend struct RuntimeCheckingPtrGroup; 387 388 public: 389 struct PointerInfo { 390 /// Holds the pointer value that we need to check. 391 TrackingVH<Value> PointerValue; 392 /// Holds the smallest byte address accessed by the pointer throughout all 393 /// iterations of the loop. 394 const SCEV *Start; 395 /// Holds the largest byte address accessed by the pointer throughout all 396 /// iterations of the loop, plus 1. 397 const SCEV *End; 398 /// Holds the information if this pointer is used for writing to memory. 399 bool IsWritePtr; 400 /// Holds the id of the set of pointers that could be dependent because of a 401 /// shared underlying object. 402 unsigned DependencySetId; 403 /// Holds the id of the disjoint alias set to which this pointer belongs. 404 unsigned AliasSetId; 405 /// SCEV for the access. 406 const SCEV *Expr; 407 /// True if the pointer expressions needs to be frozen after expansion. 408 bool NeedsFreeze; 409 410 PointerInfo(Value *PointerValue, const SCEV *Start, const SCEV *End, 411 bool IsWritePtr, unsigned DependencySetId, unsigned AliasSetId, 412 const SCEV *Expr, bool NeedsFreeze) 413 : PointerValue(PointerValue), Start(Start), End(End), 414 IsWritePtr(IsWritePtr), DependencySetId(DependencySetId), 415 AliasSetId(AliasSetId), Expr(Expr), NeedsFreeze(NeedsFreeze) {} 416 }; 417 418 RuntimePointerChecking(MemoryDepChecker &DC, ScalarEvolution *SE) 419 : DC(DC), SE(SE) {} 420 421 /// Reset the state of the pointer runtime information. 422 void reset() { 423 Need = false; 424 Pointers.clear(); 425 Checks.clear(); 426 } 427 428 /// Insert a pointer and calculate the start and end SCEVs. 429 /// We need \p PSE in order to compute the SCEV expression of the pointer 430 /// according to the assumptions that we've made during the analysis. 431 /// The method might also version the pointer stride according to \p Strides, 432 /// and add new predicates to \p PSE. 433 void insert(Loop *Lp, Value *Ptr, const SCEV *PtrExpr, Type *AccessTy, 434 bool WritePtr, unsigned DepSetId, unsigned ASId, 435 PredicatedScalarEvolution &PSE, bool NeedsFreeze); 436 437 /// No run-time memory checking is necessary. 438 bool empty() const { return Pointers.empty(); } 439 440 /// Generate the checks and store it. This also performs the grouping 441 /// of pointers to reduce the number of memchecks necessary. 442 void generateChecks(MemoryDepChecker::DepCandidates &DepCands, 443 bool UseDependencies); 444 445 /// Returns the checks that generateChecks created. They can be used to ensure 446 /// no read/write accesses overlap across all loop iterations. 447 const SmallVectorImpl<RuntimePointerCheck> &getChecks() const { 448 return Checks; 449 } 450 451 // Returns an optional list of (pointer-difference expressions, access size) 452 // pairs that can be used to prove that there are no vectorization-preventing 453 // dependencies at runtime. There are is a vectorization-preventing dependency 454 // if any pointer-difference is <u VF * InterleaveCount * access size. Returns 455 // std::nullopt if pointer-difference checks cannot be used. 456 std::optional<ArrayRef<PointerDiffInfo>> getDiffChecks() const { 457 if (!CanUseDiffCheck) 458 return std::nullopt; 459 return {DiffChecks}; 460 } 461 462 /// Decide if we need to add a check between two groups of pointers, 463 /// according to needsChecking. 464 bool needsChecking(const RuntimeCheckingPtrGroup &M, 465 const RuntimeCheckingPtrGroup &N) const; 466 467 /// Returns the number of run-time checks required according to 468 /// needsChecking. 469 unsigned getNumberOfChecks() const { return Checks.size(); } 470 471 /// Print the list run-time memory checks necessary. 472 void print(raw_ostream &OS, unsigned Depth = 0) const; 473 474 /// Print \p Checks. 475 void printChecks(raw_ostream &OS, 476 const SmallVectorImpl<RuntimePointerCheck> &Checks, 477 unsigned Depth = 0) const; 478 479 /// This flag indicates if we need to add the runtime check. 480 bool Need = false; 481 482 /// Information about the pointers that may require checking. 483 SmallVector<PointerInfo, 2> Pointers; 484 485 /// Holds a partitioning of pointers into "check groups". 486 SmallVector<RuntimeCheckingPtrGroup, 2> CheckingGroups; 487 488 /// Check if pointers are in the same partition 489 /// 490 /// \p PtrToPartition contains the partition number for pointers (-1 if the 491 /// pointer belongs to multiple partitions). 492 static bool 493 arePointersInSamePartition(const SmallVectorImpl<int> &PtrToPartition, 494 unsigned PtrIdx1, unsigned PtrIdx2); 495 496 /// Decide whether we need to issue a run-time check for pointer at 497 /// index \p I and \p J to prove their independence. 498 bool needsChecking(unsigned I, unsigned J) const; 499 500 /// Return PointerInfo for pointer at index \p PtrIdx. 501 const PointerInfo &getPointerInfo(unsigned PtrIdx) const { 502 return Pointers[PtrIdx]; 503 } 504 505 ScalarEvolution *getSE() const { return SE; } 506 507 private: 508 /// Groups pointers such that a single memcheck is required 509 /// between two different groups. This will clear the CheckingGroups vector 510 /// and re-compute it. We will only group dependecies if \p UseDependencies 511 /// is true, otherwise we will create a separate group for each pointer. 512 void groupChecks(MemoryDepChecker::DepCandidates &DepCands, 513 bool UseDependencies); 514 515 /// Generate the checks and return them. 516 SmallVector<RuntimePointerCheck, 4> generateChecks(); 517 518 /// Try to create add a new (pointer-difference, access size) pair to 519 /// DiffCheck for checking groups \p CGI and \p CGJ. If pointer-difference 520 /// checks cannot be used for the groups, set CanUseDiffCheck to false. 521 void tryToCreateDiffCheck(const RuntimeCheckingPtrGroup &CGI, 522 const RuntimeCheckingPtrGroup &CGJ); 523 524 MemoryDepChecker &DC; 525 526 /// Holds a pointer to the ScalarEvolution analysis. 527 ScalarEvolution *SE; 528 529 /// Set of run-time checks required to establish independence of 530 /// otherwise may-aliasing pointers in the loop. 531 SmallVector<RuntimePointerCheck, 4> Checks; 532 533 /// Flag indicating if pointer-difference checks can be used 534 bool CanUseDiffCheck = true; 535 536 /// A list of (pointer-difference, access size) pairs that can be used to 537 /// prove that there are no vectorization-preventing dependencies. 538 SmallVector<PointerDiffInfo> DiffChecks; 539 }; 540 541 /// Drive the analysis of memory accesses in the loop 542 /// 543 /// This class is responsible for analyzing the memory accesses of a loop. It 544 /// collects the accesses and then its main helper the AccessAnalysis class 545 /// finds and categorizes the dependences in buildDependenceSets. 546 /// 547 /// For memory dependences that can be analyzed at compile time, it determines 548 /// whether the dependence is part of cycle inhibiting vectorization. This work 549 /// is delegated to the MemoryDepChecker class. 550 /// 551 /// For memory dependences that cannot be determined at compile time, it 552 /// generates run-time checks to prove independence. This is done by 553 /// AccessAnalysis::canCheckPtrAtRT and the checks are maintained by the 554 /// RuntimePointerCheck class. 555 /// 556 /// If pointers can wrap or can't be expressed as affine AddRec expressions by 557 /// ScalarEvolution, we will generate run-time checks by emitting a 558 /// SCEVUnionPredicate. 559 /// 560 /// Checks for both memory dependences and the SCEV predicates contained in the 561 /// PSE must be emitted in order for the results of this analysis to be valid. 562 class LoopAccessInfo { 563 public: 564 LoopAccessInfo(Loop *L, ScalarEvolution *SE, const TargetLibraryInfo *TLI, 565 AAResults *AA, DominatorTree *DT, LoopInfo *LI); 566 567 /// Return true we can analyze the memory accesses in the loop and there are 568 /// no memory dependence cycles. 569 bool canVectorizeMemory() const { return CanVecMem; } 570 571 /// Return true if there is a convergent operation in the loop. There may 572 /// still be reported runtime pointer checks that would be required, but it is 573 /// not legal to insert them. 574 bool hasConvergentOp() const { return HasConvergentOp; } 575 576 const RuntimePointerChecking *getRuntimePointerChecking() const { 577 return PtrRtChecking.get(); 578 } 579 580 /// Number of memchecks required to prove independence of otherwise 581 /// may-alias pointers. 582 unsigned getNumRuntimePointerChecks() const { 583 return PtrRtChecking->getNumberOfChecks(); 584 } 585 586 /// Return true if the block BB needs to be predicated in order for the loop 587 /// to be vectorized. 588 static bool blockNeedsPredication(BasicBlock *BB, Loop *TheLoop, 589 DominatorTree *DT); 590 591 /// Returns true if value \p V is loop invariant. 592 bool isInvariant(Value *V) const; 593 594 unsigned getNumStores() const { return NumStores; } 595 unsigned getNumLoads() const { return NumLoads;} 596 597 /// The diagnostics report generated for the analysis. E.g. why we 598 /// couldn't analyze the loop. 599 const OptimizationRemarkAnalysis *getReport() const { return Report.get(); } 600 601 /// the Memory Dependence Checker which can determine the 602 /// loop-independent and loop-carried dependences between memory accesses. 603 const MemoryDepChecker &getDepChecker() const { return *DepChecker; } 604 605 /// Return the list of instructions that use \p Ptr to read or write 606 /// memory. 607 SmallVector<Instruction *, 4> getInstructionsForAccess(Value *Ptr, 608 bool isWrite) const { 609 return DepChecker->getInstructionsForAccess(Ptr, isWrite); 610 } 611 612 /// If an access has a symbolic strides, this maps the pointer value to 613 /// the stride symbol. 614 const DenseMap<Value *, const SCEV *> &getSymbolicStrides() const { 615 return SymbolicStrides; 616 } 617 618 /// Print the information about the memory accesses in the loop. 619 void print(raw_ostream &OS, unsigned Depth = 0) const; 620 621 /// If the loop has memory dependence involving an invariant address, i.e. two 622 /// stores or a store and a load, then return true, else return false. 623 bool hasDependenceInvolvingLoopInvariantAddress() const { 624 return HasDependenceInvolvingLoopInvariantAddress; 625 } 626 627 /// Return the list of stores to invariant addresses. 628 ArrayRef<StoreInst *> getStoresToInvariantAddresses() const { 629 return StoresToInvariantAddresses; 630 } 631 632 /// Used to add runtime SCEV checks. Simplifies SCEV expressions and converts 633 /// them to a more usable form. All SCEV expressions during the analysis 634 /// should be re-written (and therefore simplified) according to PSE. 635 /// A user of LoopAccessAnalysis will need to emit the runtime checks 636 /// associated with this predicate. 637 const PredicatedScalarEvolution &getPSE() const { return *PSE; } 638 639 private: 640 /// Analyze the loop. 641 void analyzeLoop(AAResults *AA, LoopInfo *LI, 642 const TargetLibraryInfo *TLI, DominatorTree *DT); 643 644 /// Check if the structure of the loop allows it to be analyzed by this 645 /// pass. 646 bool canAnalyzeLoop(); 647 648 /// Save the analysis remark. 649 /// 650 /// LAA does not directly emits the remarks. Instead it stores it which the 651 /// client can retrieve and presents as its own analysis 652 /// (e.g. -Rpass-analysis=loop-vectorize). 653 OptimizationRemarkAnalysis &recordAnalysis(StringRef RemarkName, 654 Instruction *Instr = nullptr); 655 656 /// Collect memory access with loop invariant strides. 657 /// 658 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop 659 /// invariant. 660 void collectStridedAccess(Value *LoadOrStoreInst); 661 662 // Emits the first unsafe memory dependence in a loop. 663 // Emits nothing if there are no unsafe dependences 664 // or if the dependences were not recorded. 665 void emitUnsafeDependenceRemark(); 666 667 std::unique_ptr<PredicatedScalarEvolution> PSE; 668 669 /// We need to check that all of the pointers in this list are disjoint 670 /// at runtime. Using std::unique_ptr to make using move ctor simpler. 671 std::unique_ptr<RuntimePointerChecking> PtrRtChecking; 672 673 /// the Memory Dependence Checker which can determine the 674 /// loop-independent and loop-carried dependences between memory accesses. 675 std::unique_ptr<MemoryDepChecker> DepChecker; 676 677 Loop *TheLoop; 678 679 unsigned NumLoads = 0; 680 unsigned NumStores = 0; 681 682 uint64_t MaxSafeDepDistBytes = -1; 683 684 /// Cache the result of analyzeLoop. 685 bool CanVecMem = false; 686 bool HasConvergentOp = false; 687 688 /// Indicator that there are non vectorizable stores to a uniform address. 689 bool HasDependenceInvolvingLoopInvariantAddress = false; 690 691 /// List of stores to invariant addresses. 692 SmallVector<StoreInst *> StoresToInvariantAddresses; 693 694 /// The diagnostics report generated for the analysis. E.g. why we 695 /// couldn't analyze the loop. 696 std::unique_ptr<OptimizationRemarkAnalysis> Report; 697 698 /// If an access has a symbolic strides, this maps the pointer value to 699 /// the stride symbol. 700 DenseMap<Value *, const SCEV *> SymbolicStrides; 701 }; 702 703 /// Return the SCEV corresponding to a pointer with the symbolic stride 704 /// replaced with constant one, assuming the SCEV predicate associated with 705 /// \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 /// 710 /// \p PtrToStride provides the mapping between the pointer value and its 711 /// stride as collected by LoopVectorizationLegality::collectStridedAccess. 712 const SCEV * 713 replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE, 714 const DenseMap<Value *, const SCEV *> &PtrToStride, 715 Value *Ptr); 716 717 /// If the pointer has a constant stride return it in units of the access type 718 /// size. Otherwise return std::nullopt. 719 /// 720 /// Ensure that it does not wrap in the address space, assuming the predicate 721 /// associated with \p PSE is true. 722 /// 723 /// If necessary this method will version the stride of the pointer according 724 /// to \p PtrToStride and therefore add further predicates to \p PSE. 725 /// The \p Assume parameter indicates if we are allowed to make additional 726 /// run-time assumptions. 727 /// 728 /// Note that the analysis results are defined if-and-only-if the original 729 /// memory access was defined. If that access was dead, or UB, then the 730 /// result of this function is undefined. 731 std::optional<int64_t> 732 getPtrStride(PredicatedScalarEvolution &PSE, Type *AccessTy, Value *Ptr, 733 const Loop *Lp, 734 const DenseMap<Value *, const SCEV *> &StridesMap = DenseMap<Value *, const SCEV *>(), 735 bool Assume = false, bool ShouldCheckWrap = true); 736 737 /// Returns the distance between the pointers \p PtrA and \p PtrB iff they are 738 /// compatible and it is possible to calculate the distance between them. This 739 /// is a simple API that does not depend on the analysis pass. 740 /// \param StrictCheck Ensure that the calculated distance matches the 741 /// type-based one after all the bitcasts removal in the provided pointers. 742 std::optional<int> getPointersDiff(Type *ElemTyA, Value *PtrA, Type *ElemTyB, 743 Value *PtrB, const DataLayout &DL, 744 ScalarEvolution &SE, 745 bool StrictCheck = false, 746 bool CheckType = true); 747 748 /// Attempt to sort the pointers in \p VL and return the sorted indices 749 /// in \p SortedIndices, if reordering is required. 750 /// 751 /// Returns 'true' if sorting is legal, otherwise returns 'false'. 752 /// 753 /// For example, for a given \p VL of memory accesses in program order, a[i+4], 754 /// a[i+0], a[i+1] and a[i+7], this function will sort the \p VL and save the 755 /// sorted indices in \p SortedIndices as a[i+0], a[i+1], a[i+4], a[i+7] and 756 /// saves the mask for actual memory accesses in program order in 757 /// \p SortedIndices as <1,2,0,3> 758 bool sortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy, const DataLayout &DL, 759 ScalarEvolution &SE, 760 SmallVectorImpl<unsigned> &SortedIndices); 761 762 /// Returns true if the memory operations \p A and \p B are consecutive. 763 /// This is a simple API that does not depend on the analysis pass. 764 bool isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL, 765 ScalarEvolution &SE, bool CheckType = true); 766 767 class LoopAccessInfoManager { 768 /// The cache. 769 DenseMap<Loop *, std::unique_ptr<LoopAccessInfo>> LoopAccessInfoMap; 770 771 // The used analysis passes. 772 ScalarEvolution &SE; 773 AAResults &AA; 774 DominatorTree &DT; 775 LoopInfo &LI; 776 const TargetLibraryInfo *TLI = nullptr; 777 778 public: 779 LoopAccessInfoManager(ScalarEvolution &SE, AAResults &AA, DominatorTree &DT, 780 LoopInfo &LI, const TargetLibraryInfo *TLI) 781 : SE(SE), AA(AA), DT(DT), LI(LI), TLI(TLI) {} 782 783 const LoopAccessInfo &getInfo(Loop &L); 784 785 void clear() { LoopAccessInfoMap.clear(); } 786 787 bool invalidate(Function &F, const PreservedAnalyses &PA, 788 FunctionAnalysisManager::Invalidator &Inv); 789 }; 790 791 /// This analysis provides dependence information for the memory 792 /// accesses of a loop. 793 /// 794 /// It runs the analysis for a loop on demand. This can be initiated by 795 /// querying the loop access info via AM.getResult<LoopAccessAnalysis>. 796 /// getResult return a LoopAccessInfo object. See this class for the 797 /// specifics of what information is provided. 798 class LoopAccessAnalysis 799 : public AnalysisInfoMixin<LoopAccessAnalysis> { 800 friend AnalysisInfoMixin<LoopAccessAnalysis>; 801 static AnalysisKey Key; 802 803 public: 804 typedef LoopAccessInfoManager Result; 805 806 Result run(Function &F, FunctionAnalysisManager &AM); 807 }; 808 809 inline Instruction *MemoryDepChecker::Dependence::getSource( 810 const LoopAccessInfo &LAI) const { 811 return LAI.getDepChecker().getMemoryInstructions()[Source]; 812 } 813 814 inline Instruction *MemoryDepChecker::Dependence::getDestination( 815 const LoopAccessInfo &LAI) const { 816 return LAI.getDepChecker().getMemoryInstructions()[Destination]; 817 } 818 819 } // End llvm namespace 820 821 #endif 822