1 //===- LoopAccessAnalysis.cpp - Loop Access Analysis Implementation --------==// 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 // The implementation for the loop memory dependence that was originally 10 // developed for the loop vectorizer. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Analysis/LoopAccessAnalysis.h" 15 #include "llvm/ADT/APInt.h" 16 #include "llvm/ADT/DenseMap.h" 17 #include "llvm/ADT/DepthFirstIterator.h" 18 #include "llvm/ADT/EquivalenceClasses.h" 19 #include "llvm/ADT/PointerIntPair.h" 20 #include "llvm/ADT/STLExtras.h" 21 #include "llvm/ADT/SetVector.h" 22 #include "llvm/ADT/SmallPtrSet.h" 23 #include "llvm/ADT/SmallSet.h" 24 #include "llvm/ADT/SmallVector.h" 25 #include "llvm/ADT/iterator_range.h" 26 #include "llvm/Analysis/AliasAnalysis.h" 27 #include "llvm/Analysis/AliasSetTracker.h" 28 #include "llvm/Analysis/LoopAnalysisManager.h" 29 #include "llvm/Analysis/LoopInfo.h" 30 #include "llvm/Analysis/MemoryLocation.h" 31 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 32 #include "llvm/Analysis/ScalarEvolution.h" 33 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 34 #include "llvm/Analysis/TargetLibraryInfo.h" 35 #include "llvm/Analysis/ValueTracking.h" 36 #include "llvm/Analysis/VectorUtils.h" 37 #include "llvm/IR/BasicBlock.h" 38 #include "llvm/IR/Constants.h" 39 #include "llvm/IR/DataLayout.h" 40 #include "llvm/IR/DebugLoc.h" 41 #include "llvm/IR/DerivedTypes.h" 42 #include "llvm/IR/DiagnosticInfo.h" 43 #include "llvm/IR/Dominators.h" 44 #include "llvm/IR/Function.h" 45 #include "llvm/IR/InstrTypes.h" 46 #include "llvm/IR/Instruction.h" 47 #include "llvm/IR/Instructions.h" 48 #include "llvm/IR/Operator.h" 49 #include "llvm/IR/PassManager.h" 50 #include "llvm/IR/Type.h" 51 #include "llvm/IR/Value.h" 52 #include "llvm/IR/ValueHandle.h" 53 #include "llvm/InitializePasses.h" 54 #include "llvm/Pass.h" 55 #include "llvm/Support/Casting.h" 56 #include "llvm/Support/CommandLine.h" 57 #include "llvm/Support/Debug.h" 58 #include "llvm/Support/ErrorHandling.h" 59 #include "llvm/Support/raw_ostream.h" 60 #include <algorithm> 61 #include <cassert> 62 #include <cstdint> 63 #include <cstdlib> 64 #include <iterator> 65 #include <utility> 66 #include <vector> 67 68 using namespace llvm; 69 70 #define DEBUG_TYPE "loop-accesses" 71 72 static cl::opt<unsigned, true> 73 VectorizationFactor("force-vector-width", cl::Hidden, 74 cl::desc("Sets the SIMD width. Zero is autoselect."), 75 cl::location(VectorizerParams::VectorizationFactor)); 76 unsigned VectorizerParams::VectorizationFactor; 77 78 static cl::opt<unsigned, true> 79 VectorizationInterleave("force-vector-interleave", cl::Hidden, 80 cl::desc("Sets the vectorization interleave count. " 81 "Zero is autoselect."), 82 cl::location( 83 VectorizerParams::VectorizationInterleave)); 84 unsigned VectorizerParams::VectorizationInterleave; 85 86 static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold( 87 "runtime-memory-check-threshold", cl::Hidden, 88 cl::desc("When performing memory disambiguation checks at runtime do not " 89 "generate more than this number of comparisons (default = 8)."), 90 cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8)); 91 unsigned VectorizerParams::RuntimeMemoryCheckThreshold; 92 93 /// The maximum iterations used to merge memory checks 94 static cl::opt<unsigned> MemoryCheckMergeThreshold( 95 "memory-check-merge-threshold", cl::Hidden, 96 cl::desc("Maximum number of comparisons done when trying to merge " 97 "runtime memory checks. (default = 100)"), 98 cl::init(100)); 99 100 /// Maximum SIMD width. 101 const unsigned VectorizerParams::MaxVectorWidth = 64; 102 103 /// We collect dependences up to this threshold. 104 static cl::opt<unsigned> 105 MaxDependences("max-dependences", cl::Hidden, 106 cl::desc("Maximum number of dependences collected by " 107 "loop-access analysis (default = 100)"), 108 cl::init(100)); 109 110 /// This enables versioning on the strides of symbolically striding memory 111 /// accesses in code like the following. 112 /// for (i = 0; i < N; ++i) 113 /// A[i * Stride1] += B[i * Stride2] ... 114 /// 115 /// Will be roughly translated to 116 /// if (Stride1 == 1 && Stride2 == 1) { 117 /// for (i = 0; i < N; i+=4) 118 /// A[i:i+3] += ... 119 /// } else 120 /// ... 121 static cl::opt<bool> EnableMemAccessVersioning( 122 "enable-mem-access-versioning", cl::init(true), cl::Hidden, 123 cl::desc("Enable symbolic stride memory access versioning")); 124 125 /// Enable store-to-load forwarding conflict detection. This option can 126 /// be disabled for correctness testing. 127 static cl::opt<bool> EnableForwardingConflictDetection( 128 "store-to-load-forwarding-conflict-detection", cl::Hidden, 129 cl::desc("Enable conflict detection in loop-access analysis"), 130 cl::init(true)); 131 132 bool VectorizerParams::isInterleaveForced() { 133 return ::VectorizationInterleave.getNumOccurrences() > 0; 134 } 135 136 Value *llvm::stripIntegerCast(Value *V) { 137 if (auto *CI = dyn_cast<CastInst>(V)) 138 if (CI->getOperand(0)->getType()->isIntegerTy()) 139 return CI->getOperand(0); 140 return V; 141 } 142 143 const SCEV *llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE, 144 const ValueToValueMap &PtrToStride, 145 Value *Ptr, Value *OrigPtr) { 146 const SCEV *OrigSCEV = PSE.getSCEV(Ptr); 147 148 // If there is an entry in the map return the SCEV of the pointer with the 149 // symbolic stride replaced by one. 150 ValueToValueMap::const_iterator SI = 151 PtrToStride.find(OrigPtr ? OrigPtr : Ptr); 152 if (SI != PtrToStride.end()) { 153 Value *StrideVal = SI->second; 154 155 // Strip casts. 156 StrideVal = stripIntegerCast(StrideVal); 157 158 ScalarEvolution *SE = PSE.getSE(); 159 const auto *U = cast<SCEVUnknown>(SE->getSCEV(StrideVal)); 160 const auto *CT = 161 static_cast<const SCEVConstant *>(SE->getOne(StrideVal->getType())); 162 163 PSE.addPredicate(*SE->getEqualPredicate(U, CT)); 164 auto *Expr = PSE.getSCEV(Ptr); 165 166 LLVM_DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV 167 << " by: " << *Expr << "\n"); 168 return Expr; 169 } 170 171 // Otherwise, just return the SCEV of the original pointer. 172 return OrigSCEV; 173 } 174 175 RuntimeCheckingPtrGroup::RuntimeCheckingPtrGroup( 176 unsigned Index, RuntimePointerChecking &RtCheck) 177 : RtCheck(RtCheck), High(RtCheck.Pointers[Index].End), 178 Low(RtCheck.Pointers[Index].Start) { 179 Members.push_back(Index); 180 } 181 182 /// Calculate Start and End points of memory access. 183 /// Let's assume A is the first access and B is a memory access on N-th loop 184 /// iteration. Then B is calculated as: 185 /// B = A + Step*N . 186 /// Step value may be positive or negative. 187 /// N is a calculated back-edge taken count: 188 /// N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0 189 /// Start and End points are calculated in the following way: 190 /// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt, 191 /// where SizeOfElt is the size of single memory access in bytes. 192 /// 193 /// There is no conflict when the intervals are disjoint: 194 /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End) 195 void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr, 196 unsigned DepSetId, unsigned ASId, 197 const ValueToValueMap &Strides, 198 PredicatedScalarEvolution &PSE) { 199 // Get the stride replaced scev. 200 const SCEV *Sc = replaceSymbolicStrideSCEV(PSE, Strides, Ptr); 201 ScalarEvolution *SE = PSE.getSE(); 202 203 const SCEV *ScStart; 204 const SCEV *ScEnd; 205 206 if (SE->isLoopInvariant(Sc, Lp)) 207 ScStart = ScEnd = Sc; 208 else { 209 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc); 210 assert(AR && "Invalid addrec expression"); 211 const SCEV *Ex = PSE.getBackedgeTakenCount(); 212 213 ScStart = AR->getStart(); 214 ScEnd = AR->evaluateAtIteration(Ex, *SE); 215 const SCEV *Step = AR->getStepRecurrence(*SE); 216 217 // For expressions with negative step, the upper bound is ScStart and the 218 // lower bound is ScEnd. 219 if (const auto *CStep = dyn_cast<SCEVConstant>(Step)) { 220 if (CStep->getValue()->isNegative()) 221 std::swap(ScStart, ScEnd); 222 } else { 223 // Fallback case: the step is not constant, but we can still 224 // get the upper and lower bounds of the interval by using min/max 225 // expressions. 226 ScStart = SE->getUMinExpr(ScStart, ScEnd); 227 ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd); 228 } 229 // Add the size of the pointed element to ScEnd. 230 unsigned EltSize = 231 Ptr->getType()->getPointerElementType()->getScalarSizeInBits() / 8; 232 const SCEV *EltSizeSCEV = SE->getConstant(ScEnd->getType(), EltSize); 233 ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV); 234 } 235 236 Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc); 237 } 238 239 SmallVector<RuntimePointerCheck, 4> 240 RuntimePointerChecking::generateChecks() const { 241 SmallVector<RuntimePointerCheck, 4> Checks; 242 243 for (unsigned I = 0; I < CheckingGroups.size(); ++I) { 244 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) { 245 const RuntimeCheckingPtrGroup &CGI = CheckingGroups[I]; 246 const RuntimeCheckingPtrGroup &CGJ = CheckingGroups[J]; 247 248 if (needsChecking(CGI, CGJ)) 249 Checks.push_back(std::make_pair(&CGI, &CGJ)); 250 } 251 } 252 return Checks; 253 } 254 255 void RuntimePointerChecking::generateChecks( 256 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) { 257 assert(Checks.empty() && "Checks is not empty"); 258 groupChecks(DepCands, UseDependencies); 259 Checks = generateChecks(); 260 } 261 262 bool RuntimePointerChecking::needsChecking( 263 const RuntimeCheckingPtrGroup &M, const RuntimeCheckingPtrGroup &N) const { 264 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I) 265 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J) 266 if (needsChecking(M.Members[I], N.Members[J])) 267 return true; 268 return false; 269 } 270 271 /// Compare \p I and \p J and return the minimum. 272 /// Return nullptr in case we couldn't find an answer. 273 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J, 274 ScalarEvolution *SE) { 275 const SCEV *Diff = SE->getMinusSCEV(J, I); 276 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff); 277 278 if (!C) 279 return nullptr; 280 if (C->getValue()->isNegative()) 281 return J; 282 return I; 283 } 284 285 bool RuntimeCheckingPtrGroup::addPointer(unsigned Index) { 286 const SCEV *Start = RtCheck.Pointers[Index].Start; 287 const SCEV *End = RtCheck.Pointers[Index].End; 288 289 // Compare the starts and ends with the known minimum and maximum 290 // of this set. We need to know how we compare against the min/max 291 // of the set in order to be able to emit memchecks. 292 const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE); 293 if (!Min0) 294 return false; 295 296 const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE); 297 if (!Min1) 298 return false; 299 300 // Update the low bound expression if we've found a new min value. 301 if (Min0 == Start) 302 Low = Start; 303 304 // Update the high bound expression if we've found a new max value. 305 if (Min1 != End) 306 High = End; 307 308 Members.push_back(Index); 309 return true; 310 } 311 312 void RuntimePointerChecking::groupChecks( 313 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) { 314 // We build the groups from dependency candidates equivalence classes 315 // because: 316 // - We know that pointers in the same equivalence class share 317 // the same underlying object and therefore there is a chance 318 // that we can compare pointers 319 // - We wouldn't be able to merge two pointers for which we need 320 // to emit a memcheck. The classes in DepCands are already 321 // conveniently built such that no two pointers in the same 322 // class need checking against each other. 323 324 // We use the following (greedy) algorithm to construct the groups 325 // For every pointer in the equivalence class: 326 // For each existing group: 327 // - if the difference between this pointer and the min/max bounds 328 // of the group is a constant, then make the pointer part of the 329 // group and update the min/max bounds of that group as required. 330 331 CheckingGroups.clear(); 332 333 // If we need to check two pointers to the same underlying object 334 // with a non-constant difference, we shouldn't perform any pointer 335 // grouping with those pointers. This is because we can easily get 336 // into cases where the resulting check would return false, even when 337 // the accesses are safe. 338 // 339 // The following example shows this: 340 // for (i = 0; i < 1000; ++i) 341 // a[5000 + i * m] = a[i] + a[i + 9000] 342 // 343 // Here grouping gives a check of (5000, 5000 + 1000 * m) against 344 // (0, 10000) which is always false. However, if m is 1, there is no 345 // dependence. Not grouping the checks for a[i] and a[i + 9000] allows 346 // us to perform an accurate check in this case. 347 // 348 // The above case requires that we have an UnknownDependence between 349 // accesses to the same underlying object. This cannot happen unless 350 // FoundNonConstantDistanceDependence is set, and therefore UseDependencies 351 // is also false. In this case we will use the fallback path and create 352 // separate checking groups for all pointers. 353 354 // If we don't have the dependency partitions, construct a new 355 // checking pointer group for each pointer. This is also required 356 // for correctness, because in this case we can have checking between 357 // pointers to the same underlying object. 358 if (!UseDependencies) { 359 for (unsigned I = 0; I < Pointers.size(); ++I) 360 CheckingGroups.push_back(RuntimeCheckingPtrGroup(I, *this)); 361 return; 362 } 363 364 unsigned TotalComparisons = 0; 365 366 DenseMap<Value *, unsigned> PositionMap; 367 for (unsigned Index = 0; Index < Pointers.size(); ++Index) 368 PositionMap[Pointers[Index].PointerValue] = Index; 369 370 // We need to keep track of what pointers we've already seen so we 371 // don't process them twice. 372 SmallSet<unsigned, 2> Seen; 373 374 // Go through all equivalence classes, get the "pointer check groups" 375 // and add them to the overall solution. We use the order in which accesses 376 // appear in 'Pointers' to enforce determinism. 377 for (unsigned I = 0; I < Pointers.size(); ++I) { 378 // We've seen this pointer before, and therefore already processed 379 // its equivalence class. 380 if (Seen.count(I)) 381 continue; 382 383 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue, 384 Pointers[I].IsWritePtr); 385 386 SmallVector<RuntimeCheckingPtrGroup, 2> Groups; 387 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access)); 388 389 // Because DepCands is constructed by visiting accesses in the order in 390 // which they appear in alias sets (which is deterministic) and the 391 // iteration order within an equivalence class member is only dependent on 392 // the order in which unions and insertions are performed on the 393 // equivalence class, the iteration order is deterministic. 394 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end(); 395 MI != ME; ++MI) { 396 auto PointerI = PositionMap.find(MI->getPointer()); 397 assert(PointerI != PositionMap.end() && 398 "pointer in equivalence class not found in PositionMap"); 399 unsigned Pointer = PointerI->second; 400 bool Merged = false; 401 // Mark this pointer as seen. 402 Seen.insert(Pointer); 403 404 // Go through all the existing sets and see if we can find one 405 // which can include this pointer. 406 for (RuntimeCheckingPtrGroup &Group : Groups) { 407 // Don't perform more than a certain amount of comparisons. 408 // This should limit the cost of grouping the pointers to something 409 // reasonable. If we do end up hitting this threshold, the algorithm 410 // will create separate groups for all remaining pointers. 411 if (TotalComparisons > MemoryCheckMergeThreshold) 412 break; 413 414 TotalComparisons++; 415 416 if (Group.addPointer(Pointer)) { 417 Merged = true; 418 break; 419 } 420 } 421 422 if (!Merged) 423 // We couldn't add this pointer to any existing set or the threshold 424 // for the number of comparisons has been reached. Create a new group 425 // to hold the current pointer. 426 Groups.push_back(RuntimeCheckingPtrGroup(Pointer, *this)); 427 } 428 429 // We've computed the grouped checks for this partition. 430 // Save the results and continue with the next one. 431 llvm::copy(Groups, std::back_inserter(CheckingGroups)); 432 } 433 } 434 435 bool RuntimePointerChecking::arePointersInSamePartition( 436 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1, 437 unsigned PtrIdx2) { 438 return (PtrToPartition[PtrIdx1] != -1 && 439 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]); 440 } 441 442 bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const { 443 const PointerInfo &PointerI = Pointers[I]; 444 const PointerInfo &PointerJ = Pointers[J]; 445 446 // No need to check if two readonly pointers intersect. 447 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr) 448 return false; 449 450 // Only need to check pointers between two different dependency sets. 451 if (PointerI.DependencySetId == PointerJ.DependencySetId) 452 return false; 453 454 // Only need to check pointers in the same alias set. 455 if (PointerI.AliasSetId != PointerJ.AliasSetId) 456 return false; 457 458 return true; 459 } 460 461 void RuntimePointerChecking::printChecks( 462 raw_ostream &OS, const SmallVectorImpl<RuntimePointerCheck> &Checks, 463 unsigned Depth) const { 464 unsigned N = 0; 465 for (const auto &Check : Checks) { 466 const auto &First = Check.first->Members, &Second = Check.second->Members; 467 468 OS.indent(Depth) << "Check " << N++ << ":\n"; 469 470 OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n"; 471 for (unsigned K = 0; K < First.size(); ++K) 472 OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n"; 473 474 OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n"; 475 for (unsigned K = 0; K < Second.size(); ++K) 476 OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n"; 477 } 478 } 479 480 void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const { 481 482 OS.indent(Depth) << "Run-time memory checks:\n"; 483 printChecks(OS, Checks, Depth); 484 485 OS.indent(Depth) << "Grouped accesses:\n"; 486 for (unsigned I = 0; I < CheckingGroups.size(); ++I) { 487 const auto &CG = CheckingGroups[I]; 488 489 OS.indent(Depth + 2) << "Group " << &CG << ":\n"; 490 OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High 491 << ")\n"; 492 for (unsigned J = 0; J < CG.Members.size(); ++J) { 493 OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr 494 << "\n"; 495 } 496 } 497 } 498 499 namespace { 500 501 /// Analyses memory accesses in a loop. 502 /// 503 /// Checks whether run time pointer checks are needed and builds sets for data 504 /// dependence checking. 505 class AccessAnalysis { 506 public: 507 /// Read or write access location. 508 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo; 509 typedef SmallVector<MemAccessInfo, 8> MemAccessInfoList; 510 511 AccessAnalysis(const DataLayout &Dl, Loop *TheLoop, AAResults *AA, 512 LoopInfo *LI, MemoryDepChecker::DepCandidates &DA, 513 PredicatedScalarEvolution &PSE) 514 : DL(Dl), TheLoop(TheLoop), AST(*AA), LI(LI), DepCands(DA), 515 IsRTCheckAnalysisNeeded(false), PSE(PSE) {} 516 517 /// Register a load and whether it is only read from. 518 void addLoad(MemoryLocation &Loc, bool IsReadOnly) { 519 Value *Ptr = const_cast<Value*>(Loc.Ptr); 520 AST.add(Ptr, LocationSize::unknown(), Loc.AATags); 521 Accesses.insert(MemAccessInfo(Ptr, false)); 522 if (IsReadOnly) 523 ReadOnlyPtr.insert(Ptr); 524 } 525 526 /// Register a store. 527 void addStore(MemoryLocation &Loc) { 528 Value *Ptr = const_cast<Value*>(Loc.Ptr); 529 AST.add(Ptr, LocationSize::unknown(), Loc.AATags); 530 Accesses.insert(MemAccessInfo(Ptr, true)); 531 } 532 533 /// Check if we can emit a run-time no-alias check for \p Access. 534 /// 535 /// Returns true if we can emit a run-time no alias check for \p Access. 536 /// If we can check this access, this also adds it to a dependence set and 537 /// adds a run-time to check for it to \p RtCheck. If \p Assume is true, 538 /// we will attempt to use additional run-time checks in order to get 539 /// the bounds of the pointer. 540 bool createCheckForAccess(RuntimePointerChecking &RtCheck, 541 MemAccessInfo Access, 542 const ValueToValueMap &Strides, 543 DenseMap<Value *, unsigned> &DepSetId, 544 Loop *TheLoop, unsigned &RunningDepId, 545 unsigned ASId, bool ShouldCheckStride, 546 bool Assume); 547 548 /// Check whether we can check the pointers at runtime for 549 /// non-intersection. 550 /// 551 /// Returns true if we need no check or if we do and we can generate them 552 /// (i.e. the pointers have computable bounds). 553 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE, 554 Loop *TheLoop, const ValueToValueMap &Strides, 555 bool ShouldCheckWrap = false); 556 557 /// Goes over all memory accesses, checks whether a RT check is needed 558 /// and builds sets of dependent accesses. 559 void buildDependenceSets() { 560 processMemAccesses(); 561 } 562 563 /// Initial processing of memory accesses determined that we need to 564 /// perform dependency checking. 565 /// 566 /// Note that this can later be cleared if we retry memcheck analysis without 567 /// dependency checking (i.e. FoundNonConstantDistanceDependence). 568 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); } 569 570 /// We decided that no dependence analysis would be used. Reset the state. 571 void resetDepChecks(MemoryDepChecker &DepChecker) { 572 CheckDeps.clear(); 573 DepChecker.clearDependences(); 574 } 575 576 MemAccessInfoList &getDependenciesToCheck() { return CheckDeps; } 577 578 private: 579 typedef SetVector<MemAccessInfo> PtrAccessSet; 580 581 /// Go over all memory access and check whether runtime pointer checks 582 /// are needed and build sets of dependency check candidates. 583 void processMemAccesses(); 584 585 /// Set of all accesses. 586 PtrAccessSet Accesses; 587 588 const DataLayout &DL; 589 590 /// The loop being checked. 591 const Loop *TheLoop; 592 593 /// List of accesses that need a further dependence check. 594 MemAccessInfoList CheckDeps; 595 596 /// Set of pointers that are read only. 597 SmallPtrSet<Value*, 16> ReadOnlyPtr; 598 599 /// An alias set tracker to partition the access set by underlying object and 600 //intrinsic property (such as TBAA metadata). 601 AliasSetTracker AST; 602 603 LoopInfo *LI; 604 605 /// Sets of potentially dependent accesses - members of one set share an 606 /// underlying pointer. The set "CheckDeps" identfies which sets really need a 607 /// dependence check. 608 MemoryDepChecker::DepCandidates &DepCands; 609 610 /// Initial processing of memory accesses determined that we may need 611 /// to add memchecks. Perform the analysis to determine the necessary checks. 612 /// 613 /// Note that, this is different from isDependencyCheckNeeded. When we retry 614 /// memcheck analysis without dependency checking 615 /// (i.e. FoundNonConstantDistanceDependence), isDependencyCheckNeeded is 616 /// cleared while this remains set if we have potentially dependent accesses. 617 bool IsRTCheckAnalysisNeeded; 618 619 /// The SCEV predicate containing all the SCEV-related assumptions. 620 PredicatedScalarEvolution &PSE; 621 }; 622 623 } // end anonymous namespace 624 625 /// Check whether a pointer can participate in a runtime bounds check. 626 /// If \p Assume, try harder to prove that we can compute the bounds of \p Ptr 627 /// by adding run-time checks (overflow checks) if necessary. 628 static bool hasComputableBounds(PredicatedScalarEvolution &PSE, 629 const ValueToValueMap &Strides, Value *Ptr, 630 Loop *L, bool Assume) { 631 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr); 632 633 // The bounds for loop-invariant pointer is trivial. 634 if (PSE.getSE()->isLoopInvariant(PtrScev, L)) 635 return true; 636 637 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); 638 639 if (!AR && Assume) 640 AR = PSE.getAsAddRec(Ptr); 641 642 if (!AR) 643 return false; 644 645 return AR->isAffine(); 646 } 647 648 /// Check whether a pointer address cannot wrap. 649 static bool isNoWrap(PredicatedScalarEvolution &PSE, 650 const ValueToValueMap &Strides, Value *Ptr, Loop *L) { 651 const SCEV *PtrScev = PSE.getSCEV(Ptr); 652 if (PSE.getSE()->isLoopInvariant(PtrScev, L)) 653 return true; 654 655 int64_t Stride = getPtrStride(PSE, Ptr, L, Strides); 656 if (Stride == 1 || PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW)) 657 return true; 658 659 return false; 660 } 661 662 bool AccessAnalysis::createCheckForAccess(RuntimePointerChecking &RtCheck, 663 MemAccessInfo Access, 664 const ValueToValueMap &StridesMap, 665 DenseMap<Value *, unsigned> &DepSetId, 666 Loop *TheLoop, unsigned &RunningDepId, 667 unsigned ASId, bool ShouldCheckWrap, 668 bool Assume) { 669 Value *Ptr = Access.getPointer(); 670 671 if (!hasComputableBounds(PSE, StridesMap, Ptr, TheLoop, Assume)) 672 return false; 673 674 // When we run after a failing dependency check we have to make sure 675 // we don't have wrapping pointers. 676 if (ShouldCheckWrap && !isNoWrap(PSE, StridesMap, Ptr, TheLoop)) { 677 auto *Expr = PSE.getSCEV(Ptr); 678 if (!Assume || !isa<SCEVAddRecExpr>(Expr)) 679 return false; 680 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); 681 } 682 683 // The id of the dependence set. 684 unsigned DepId; 685 686 if (isDependencyCheckNeeded()) { 687 Value *Leader = DepCands.getLeaderValue(Access).getPointer(); 688 unsigned &LeaderId = DepSetId[Leader]; 689 if (!LeaderId) 690 LeaderId = RunningDepId++; 691 DepId = LeaderId; 692 } else 693 // Each access has its own dependence set. 694 DepId = RunningDepId++; 695 696 bool IsWrite = Access.getInt(); 697 RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap, PSE); 698 LLVM_DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n'); 699 700 return true; 701 } 702 703 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck, 704 ScalarEvolution *SE, Loop *TheLoop, 705 const ValueToValueMap &StridesMap, 706 bool ShouldCheckWrap) { 707 // Find pointers with computable bounds. We are going to use this information 708 // to place a runtime bound check. 709 bool CanDoRT = true; 710 711 bool MayNeedRTCheck = false; 712 if (!IsRTCheckAnalysisNeeded) return true; 713 714 bool IsDepCheckNeeded = isDependencyCheckNeeded(); 715 716 // We assign a consecutive id to access from different alias sets. 717 // Accesses between different groups doesn't need to be checked. 718 unsigned ASId = 0; 719 for (auto &AS : AST) { 720 int NumReadPtrChecks = 0; 721 int NumWritePtrChecks = 0; 722 bool CanDoAliasSetRT = true; 723 ++ASId; 724 725 // We assign consecutive id to access from different dependence sets. 726 // Accesses within the same set don't need a runtime check. 727 unsigned RunningDepId = 1; 728 DenseMap<Value *, unsigned> DepSetId; 729 730 SmallVector<MemAccessInfo, 4> Retries; 731 732 // First, count how many write and read accesses are in the alias set. Also 733 // collect MemAccessInfos for later. 734 SmallVector<MemAccessInfo, 4> AccessInfos; 735 for (auto A : AS) { 736 Value *Ptr = A.getValue(); 737 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true)); 738 739 if (IsWrite) 740 ++NumWritePtrChecks; 741 else 742 ++NumReadPtrChecks; 743 AccessInfos.emplace_back(Ptr, IsWrite); 744 } 745 746 // We do not need runtime checks for this alias set, if there are no writes 747 // or a single write and no reads. 748 if (NumWritePtrChecks == 0 || 749 (NumWritePtrChecks == 1 && NumReadPtrChecks == 0)) { 750 assert((AS.size() <= 1 || 751 all_of(AS, 752 [this](auto AC) { 753 MemAccessInfo AccessWrite(AC.getValue(), true); 754 return DepCands.findValue(AccessWrite) == DepCands.end(); 755 })) && 756 "Can only skip updating CanDoRT below, if all entries in AS " 757 "are reads or there is at most 1 entry"); 758 continue; 759 } 760 761 for (auto &Access : AccessInfos) { 762 if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId, TheLoop, 763 RunningDepId, ASId, ShouldCheckWrap, false)) { 764 LLVM_DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" 765 << *Access.getPointer() << '\n'); 766 Retries.push_back(Access); 767 CanDoAliasSetRT = false; 768 } 769 } 770 771 // Note that this function computes CanDoRT and MayNeedRTCheck 772 // independently. For example CanDoRT=false, MayNeedRTCheck=false means that 773 // we have a pointer for which we couldn't find the bounds but we don't 774 // actually need to emit any checks so it does not matter. 775 // 776 // We need runtime checks for this alias set, if there are at least 2 777 // dependence sets (in which case RunningDepId > 2) or if we need to re-try 778 // any bound checks (because in that case the number of dependence sets is 779 // incomplete). 780 bool NeedsAliasSetRTCheck = RunningDepId > 2 || !Retries.empty(); 781 782 // We need to perform run-time alias checks, but some pointers had bounds 783 // that couldn't be checked. 784 if (NeedsAliasSetRTCheck && !CanDoAliasSetRT) { 785 // Reset the CanDoSetRt flag and retry all accesses that have failed. 786 // We know that we need these checks, so we can now be more aggressive 787 // and add further checks if required (overflow checks). 788 CanDoAliasSetRT = true; 789 for (auto Access : Retries) 790 if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId, 791 TheLoop, RunningDepId, ASId, 792 ShouldCheckWrap, /*Assume=*/true)) { 793 CanDoAliasSetRT = false; 794 break; 795 } 796 } 797 798 CanDoRT &= CanDoAliasSetRT; 799 MayNeedRTCheck |= NeedsAliasSetRTCheck; 800 ++ASId; 801 } 802 803 // If the pointers that we would use for the bounds comparison have different 804 // address spaces, assume the values aren't directly comparable, so we can't 805 // use them for the runtime check. We also have to assume they could 806 // overlap. In the future there should be metadata for whether address spaces 807 // are disjoint. 808 unsigned NumPointers = RtCheck.Pointers.size(); 809 for (unsigned i = 0; i < NumPointers; ++i) { 810 for (unsigned j = i + 1; j < NumPointers; ++j) { 811 // Only need to check pointers between two different dependency sets. 812 if (RtCheck.Pointers[i].DependencySetId == 813 RtCheck.Pointers[j].DependencySetId) 814 continue; 815 // Only need to check pointers in the same alias set. 816 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId) 817 continue; 818 819 Value *PtrI = RtCheck.Pointers[i].PointerValue; 820 Value *PtrJ = RtCheck.Pointers[j].PointerValue; 821 822 unsigned ASi = PtrI->getType()->getPointerAddressSpace(); 823 unsigned ASj = PtrJ->getType()->getPointerAddressSpace(); 824 if (ASi != ASj) { 825 LLVM_DEBUG( 826 dbgs() << "LAA: Runtime check would require comparison between" 827 " different address spaces\n"); 828 return false; 829 } 830 } 831 } 832 833 if (MayNeedRTCheck && CanDoRT) 834 RtCheck.generateChecks(DepCands, IsDepCheckNeeded); 835 836 LLVM_DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks() 837 << " pointer comparisons.\n"); 838 839 // If we can do run-time checks, but there are no checks, no runtime checks 840 // are needed. This can happen when all pointers point to the same underlying 841 // object for example. 842 RtCheck.Need = CanDoRT ? RtCheck.getNumberOfChecks() != 0 : MayNeedRTCheck; 843 844 bool CanDoRTIfNeeded = !RtCheck.Need || CanDoRT; 845 if (!CanDoRTIfNeeded) 846 RtCheck.reset(); 847 return CanDoRTIfNeeded; 848 } 849 850 void AccessAnalysis::processMemAccesses() { 851 // We process the set twice: first we process read-write pointers, last we 852 // process read-only pointers. This allows us to skip dependence tests for 853 // read-only pointers. 854 855 LLVM_DEBUG(dbgs() << "LAA: Processing memory accesses...\n"); 856 LLVM_DEBUG(dbgs() << " AST: "; AST.dump()); 857 LLVM_DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n"); 858 LLVM_DEBUG({ 859 for (auto A : Accesses) 860 dbgs() << "\t" << *A.getPointer() << " (" << 861 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? 862 "read-only" : "read")) << ")\n"; 863 }); 864 865 // The AliasSetTracker has nicely partitioned our pointers by metadata 866 // compatibility and potential for underlying-object overlap. As a result, we 867 // only need to check for potential pointer dependencies within each alias 868 // set. 869 for (auto &AS : AST) { 870 // Note that both the alias-set tracker and the alias sets themselves used 871 // linked lists internally and so the iteration order here is deterministic 872 // (matching the original instruction order within each set). 873 874 bool SetHasWrite = false; 875 876 // Map of pointers to last access encountered. 877 typedef DenseMap<const Value*, MemAccessInfo> UnderlyingObjToAccessMap; 878 UnderlyingObjToAccessMap ObjToLastAccess; 879 880 // Set of access to check after all writes have been processed. 881 PtrAccessSet DeferredAccesses; 882 883 // Iterate over each alias set twice, once to process read/write pointers, 884 // and then to process read-only pointers. 885 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) { 886 bool UseDeferred = SetIteration > 0; 887 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses; 888 889 for (auto AV : AS) { 890 Value *Ptr = AV.getValue(); 891 892 // For a single memory access in AliasSetTracker, Accesses may contain 893 // both read and write, and they both need to be handled for CheckDeps. 894 for (auto AC : S) { 895 if (AC.getPointer() != Ptr) 896 continue; 897 898 bool IsWrite = AC.getInt(); 899 900 // If we're using the deferred access set, then it contains only 901 // reads. 902 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite; 903 if (UseDeferred && !IsReadOnlyPtr) 904 continue; 905 // Otherwise, the pointer must be in the PtrAccessSet, either as a 906 // read or a write. 907 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite || 908 S.count(MemAccessInfo(Ptr, false))) && 909 "Alias-set pointer not in the access set?"); 910 911 MemAccessInfo Access(Ptr, IsWrite); 912 DepCands.insert(Access); 913 914 // Memorize read-only pointers for later processing and skip them in 915 // the first round (they need to be checked after we have seen all 916 // write pointers). Note: we also mark pointer that are not 917 // consecutive as "read-only" pointers (so that we check 918 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite". 919 if (!UseDeferred && IsReadOnlyPtr) { 920 DeferredAccesses.insert(Access); 921 continue; 922 } 923 924 // If this is a write - check other reads and writes for conflicts. If 925 // this is a read only check other writes for conflicts (but only if 926 // there is no other write to the ptr - this is an optimization to 927 // catch "a[i] = a[i] + " without having to do a dependence check). 928 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) { 929 CheckDeps.push_back(Access); 930 IsRTCheckAnalysisNeeded = true; 931 } 932 933 if (IsWrite) 934 SetHasWrite = true; 935 936 // Create sets of pointers connected by a shared alias set and 937 // underlying object. 938 typedef SmallVector<const Value *, 16> ValueVector; 939 ValueVector TempObjects; 940 941 GetUnderlyingObjects(Ptr, TempObjects, DL, LI); 942 LLVM_DEBUG(dbgs() 943 << "Underlying objects for pointer " << *Ptr << "\n"); 944 for (const Value *UnderlyingObj : TempObjects) { 945 // nullptr never alias, don't join sets for pointer that have "null" 946 // in their UnderlyingObjects list. 947 if (isa<ConstantPointerNull>(UnderlyingObj) && 948 !NullPointerIsDefined( 949 TheLoop->getHeader()->getParent(), 950 UnderlyingObj->getType()->getPointerAddressSpace())) 951 continue; 952 953 UnderlyingObjToAccessMap::iterator Prev = 954 ObjToLastAccess.find(UnderlyingObj); 955 if (Prev != ObjToLastAccess.end()) 956 DepCands.unionSets(Access, Prev->second); 957 958 ObjToLastAccess[UnderlyingObj] = Access; 959 LLVM_DEBUG(dbgs() << " " << *UnderlyingObj << "\n"); 960 } 961 } 962 } 963 } 964 } 965 } 966 967 static bool isInBoundsGep(Value *Ptr) { 968 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) 969 return GEP->isInBounds(); 970 return false; 971 } 972 973 /// Return true if an AddRec pointer \p Ptr is unsigned non-wrapping, 974 /// i.e. monotonically increasing/decreasing. 975 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR, 976 PredicatedScalarEvolution &PSE, const Loop *L) { 977 // FIXME: This should probably only return true for NUW. 978 if (AR->getNoWrapFlags(SCEV::NoWrapMask)) 979 return true; 980 981 // Scalar evolution does not propagate the non-wrapping flags to values that 982 // are derived from a non-wrapping induction variable because non-wrapping 983 // could be flow-sensitive. 984 // 985 // Look through the potentially overflowing instruction to try to prove 986 // non-wrapping for the *specific* value of Ptr. 987 988 // The arithmetic implied by an inbounds GEP can't overflow. 989 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr); 990 if (!GEP || !GEP->isInBounds()) 991 return false; 992 993 // Make sure there is only one non-const index and analyze that. 994 Value *NonConstIndex = nullptr; 995 for (Value *Index : make_range(GEP->idx_begin(), GEP->idx_end())) 996 if (!isa<ConstantInt>(Index)) { 997 if (NonConstIndex) 998 return false; 999 NonConstIndex = Index; 1000 } 1001 if (!NonConstIndex) 1002 // The recurrence is on the pointer, ignore for now. 1003 return false; 1004 1005 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW 1006 // AddRec using a NSW operation. 1007 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex)) 1008 if (OBO->hasNoSignedWrap() && 1009 // Assume constant for other the operand so that the AddRec can be 1010 // easily found. 1011 isa<ConstantInt>(OBO->getOperand(1))) { 1012 auto *OpScev = PSE.getSCEV(OBO->getOperand(0)); 1013 1014 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev)) 1015 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW); 1016 } 1017 1018 return false; 1019 } 1020 1021 /// Check whether the access through \p Ptr has a constant stride. 1022 int64_t llvm::getPtrStride(PredicatedScalarEvolution &PSE, Value *Ptr, 1023 const Loop *Lp, const ValueToValueMap &StridesMap, 1024 bool Assume, bool ShouldCheckWrap) { 1025 Type *Ty = Ptr->getType(); 1026 assert(Ty->isPointerTy() && "Unexpected non-ptr"); 1027 1028 // Make sure that the pointer does not point to aggregate types. 1029 auto *PtrTy = cast<PointerType>(Ty); 1030 if (PtrTy->getElementType()->isAggregateType()) { 1031 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type" 1032 << *Ptr << "\n"); 1033 return 0; 1034 } 1035 1036 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr); 1037 1038 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); 1039 if (Assume && !AR) 1040 AR = PSE.getAsAddRec(Ptr); 1041 1042 if (!AR) { 1043 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr 1044 << " SCEV: " << *PtrScev << "\n"); 1045 return 0; 1046 } 1047 1048 // The access function must stride over the innermost loop. 1049 if (Lp != AR->getLoop()) { 1050 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " 1051 << *Ptr << " SCEV: " << *AR << "\n"); 1052 return 0; 1053 } 1054 1055 // The address calculation must not wrap. Otherwise, a dependence could be 1056 // inverted. 1057 // An inbounds getelementptr that is a AddRec with a unit stride 1058 // cannot wrap per definition. The unit stride requirement is checked later. 1059 // An getelementptr without an inbounds attribute and unit stride would have 1060 // to access the pointer value "0" which is undefined behavior in address 1061 // space 0, therefore we can also vectorize this case. 1062 bool IsInBoundsGEP = isInBoundsGep(Ptr); 1063 bool IsNoWrapAddRec = !ShouldCheckWrap || 1064 PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW) || 1065 isNoWrapAddRec(Ptr, AR, PSE, Lp); 1066 if (!IsNoWrapAddRec && !IsInBoundsGEP && 1067 NullPointerIsDefined(Lp->getHeader()->getParent(), 1068 PtrTy->getAddressSpace())) { 1069 if (Assume) { 1070 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); 1071 IsNoWrapAddRec = true; 1072 LLVM_DEBUG(dbgs() << "LAA: Pointer may wrap in the address space:\n" 1073 << "LAA: Pointer: " << *Ptr << "\n" 1074 << "LAA: SCEV: " << *AR << "\n" 1075 << "LAA: Added an overflow assumption\n"); 1076 } else { 1077 LLVM_DEBUG( 1078 dbgs() << "LAA: Bad stride - Pointer may wrap in the address space " 1079 << *Ptr << " SCEV: " << *AR << "\n"); 1080 return 0; 1081 } 1082 } 1083 1084 // Check the step is constant. 1085 const SCEV *Step = AR->getStepRecurrence(*PSE.getSE()); 1086 1087 // Calculate the pointer stride and check if it is constant. 1088 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 1089 if (!C) { 1090 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr 1091 << " SCEV: " << *AR << "\n"); 1092 return 0; 1093 } 1094 1095 auto &DL = Lp->getHeader()->getModule()->getDataLayout(); 1096 int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType()); 1097 const APInt &APStepVal = C->getAPInt(); 1098 1099 // Huge step value - give up. 1100 if (APStepVal.getBitWidth() > 64) 1101 return 0; 1102 1103 int64_t StepVal = APStepVal.getSExtValue(); 1104 1105 // Strided access. 1106 int64_t Stride = StepVal / Size; 1107 int64_t Rem = StepVal % Size; 1108 if (Rem) 1109 return 0; 1110 1111 // If the SCEV could wrap but we have an inbounds gep with a unit stride we 1112 // know we can't "wrap around the address space". In case of address space 1113 // zero we know that this won't happen without triggering undefined behavior. 1114 if (!IsNoWrapAddRec && Stride != 1 && Stride != -1 && 1115 (IsInBoundsGEP || !NullPointerIsDefined(Lp->getHeader()->getParent(), 1116 PtrTy->getAddressSpace()))) { 1117 if (Assume) { 1118 // We can avoid this case by adding a run-time check. 1119 LLVM_DEBUG(dbgs() << "LAA: Non unit strided pointer which is not either " 1120 << "inbounds or in address space 0 may wrap:\n" 1121 << "LAA: Pointer: " << *Ptr << "\n" 1122 << "LAA: SCEV: " << *AR << "\n" 1123 << "LAA: Added an overflow assumption\n"); 1124 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); 1125 } else 1126 return 0; 1127 } 1128 1129 return Stride; 1130 } 1131 1132 bool llvm::sortPtrAccesses(ArrayRef<Value *> VL, const DataLayout &DL, 1133 ScalarEvolution &SE, 1134 SmallVectorImpl<unsigned> &SortedIndices) { 1135 assert(llvm::all_of( 1136 VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && 1137 "Expected list of pointer operands."); 1138 SmallVector<std::pair<int64_t, Value *>, 4> OffValPairs; 1139 OffValPairs.reserve(VL.size()); 1140 1141 // Walk over the pointers, and map each of them to an offset relative to 1142 // first pointer in the array. 1143 Value *Ptr0 = VL[0]; 1144 const SCEV *Scev0 = SE.getSCEV(Ptr0); 1145 Value *Obj0 = GetUnderlyingObject(Ptr0, DL); 1146 1147 llvm::SmallSet<int64_t, 4> Offsets; 1148 for (auto *Ptr : VL) { 1149 // TODO: Outline this code as a special, more time consuming, version of 1150 // computeConstantDifference() function. 1151 if (Ptr->getType()->getPointerAddressSpace() != 1152 Ptr0->getType()->getPointerAddressSpace()) 1153 return false; 1154 // If a pointer refers to a different underlying object, bail - the 1155 // pointers are by definition incomparable. 1156 Value *CurrObj = GetUnderlyingObject(Ptr, DL); 1157 if (CurrObj != Obj0) 1158 return false; 1159 1160 const SCEV *Scev = SE.getSCEV(Ptr); 1161 const auto *Diff = dyn_cast<SCEVConstant>(SE.getMinusSCEV(Scev, Scev0)); 1162 // The pointers may not have a constant offset from each other, or SCEV 1163 // may just not be smart enough to figure out they do. Regardless, 1164 // there's nothing we can do. 1165 if (!Diff) 1166 return false; 1167 1168 // Check if the pointer with the same offset is found. 1169 int64_t Offset = Diff->getAPInt().getSExtValue(); 1170 if (!Offsets.insert(Offset).second) 1171 return false; 1172 OffValPairs.emplace_back(Offset, Ptr); 1173 } 1174 SortedIndices.clear(); 1175 SortedIndices.resize(VL.size()); 1176 std::iota(SortedIndices.begin(), SortedIndices.end(), 0); 1177 1178 // Sort the memory accesses and keep the order of their uses in UseOrder. 1179 llvm::stable_sort(SortedIndices, [&](unsigned Left, unsigned Right) { 1180 return OffValPairs[Left].first < OffValPairs[Right].first; 1181 }); 1182 1183 // Check if the order is consecutive already. 1184 if (llvm::all_of(SortedIndices, [&SortedIndices](const unsigned I) { 1185 return I == SortedIndices[I]; 1186 })) 1187 SortedIndices.clear(); 1188 1189 return true; 1190 } 1191 1192 /// Take the address space operand from the Load/Store instruction. 1193 /// Returns -1 if this is not a valid Load/Store instruction. 1194 static unsigned getAddressSpaceOperand(Value *I) { 1195 if (LoadInst *L = dyn_cast<LoadInst>(I)) 1196 return L->getPointerAddressSpace(); 1197 if (StoreInst *S = dyn_cast<StoreInst>(I)) 1198 return S->getPointerAddressSpace(); 1199 return -1; 1200 } 1201 1202 /// Returns true if the memory operations \p A and \p B are consecutive. 1203 bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL, 1204 ScalarEvolution &SE, bool CheckType) { 1205 Value *PtrA = getLoadStorePointerOperand(A); 1206 Value *PtrB = getLoadStorePointerOperand(B); 1207 unsigned ASA = getAddressSpaceOperand(A); 1208 unsigned ASB = getAddressSpaceOperand(B); 1209 1210 // Check that the address spaces match and that the pointers are valid. 1211 if (!PtrA || !PtrB || (ASA != ASB)) 1212 return false; 1213 1214 // Make sure that A and B are different pointers. 1215 if (PtrA == PtrB) 1216 return false; 1217 1218 // Make sure that A and B have the same type if required. 1219 if (CheckType && PtrA->getType() != PtrB->getType()) 1220 return false; 1221 1222 unsigned IdxWidth = DL.getIndexSizeInBits(ASA); 1223 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType(); 1224 1225 APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0); 1226 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA); 1227 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB); 1228 1229 // Retrieve the address space again as pointer stripping now tracks through 1230 // `addrspacecast`. 1231 ASA = cast<PointerType>(PtrA->getType())->getAddressSpace(); 1232 ASB = cast<PointerType>(PtrB->getType())->getAddressSpace(); 1233 // Check that the address spaces match and that the pointers are valid. 1234 if (ASA != ASB) 1235 return false; 1236 1237 IdxWidth = DL.getIndexSizeInBits(ASA); 1238 OffsetA = OffsetA.sextOrTrunc(IdxWidth); 1239 OffsetB = OffsetB.sextOrTrunc(IdxWidth); 1240 1241 APInt Size(IdxWidth, DL.getTypeStoreSize(Ty)); 1242 1243 // OffsetDelta = OffsetB - OffsetA; 1244 const SCEV *OffsetSCEVA = SE.getConstant(OffsetA); 1245 const SCEV *OffsetSCEVB = SE.getConstant(OffsetB); 1246 const SCEV *OffsetDeltaSCEV = SE.getMinusSCEV(OffsetSCEVB, OffsetSCEVA); 1247 const APInt &OffsetDelta = cast<SCEVConstant>(OffsetDeltaSCEV)->getAPInt(); 1248 1249 // Check if they are based on the same pointer. That makes the offsets 1250 // sufficient. 1251 if (PtrA == PtrB) 1252 return OffsetDelta == Size; 1253 1254 // Compute the necessary base pointer delta to have the necessary final delta 1255 // equal to the size. 1256 // BaseDelta = Size - OffsetDelta; 1257 const SCEV *SizeSCEV = SE.getConstant(Size); 1258 const SCEV *BaseDelta = SE.getMinusSCEV(SizeSCEV, OffsetDeltaSCEV); 1259 1260 // Otherwise compute the distance with SCEV between the base pointers. 1261 const SCEV *PtrSCEVA = SE.getSCEV(PtrA); 1262 const SCEV *PtrSCEVB = SE.getSCEV(PtrB); 1263 const SCEV *X = SE.getAddExpr(PtrSCEVA, BaseDelta); 1264 return X == PtrSCEVB; 1265 } 1266 1267 MemoryDepChecker::VectorizationSafetyStatus 1268 MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) { 1269 switch (Type) { 1270 case NoDep: 1271 case Forward: 1272 case BackwardVectorizable: 1273 return VectorizationSafetyStatus::Safe; 1274 1275 case Unknown: 1276 return VectorizationSafetyStatus::PossiblySafeWithRtChecks; 1277 case ForwardButPreventsForwarding: 1278 case Backward: 1279 case BackwardVectorizableButPreventsForwarding: 1280 return VectorizationSafetyStatus::Unsafe; 1281 } 1282 llvm_unreachable("unexpected DepType!"); 1283 } 1284 1285 bool MemoryDepChecker::Dependence::isBackward() const { 1286 switch (Type) { 1287 case NoDep: 1288 case Forward: 1289 case ForwardButPreventsForwarding: 1290 case Unknown: 1291 return false; 1292 1293 case BackwardVectorizable: 1294 case Backward: 1295 case BackwardVectorizableButPreventsForwarding: 1296 return true; 1297 } 1298 llvm_unreachable("unexpected DepType!"); 1299 } 1300 1301 bool MemoryDepChecker::Dependence::isPossiblyBackward() const { 1302 return isBackward() || Type == Unknown; 1303 } 1304 1305 bool MemoryDepChecker::Dependence::isForward() const { 1306 switch (Type) { 1307 case Forward: 1308 case ForwardButPreventsForwarding: 1309 return true; 1310 1311 case NoDep: 1312 case Unknown: 1313 case BackwardVectorizable: 1314 case Backward: 1315 case BackwardVectorizableButPreventsForwarding: 1316 return false; 1317 } 1318 llvm_unreachable("unexpected DepType!"); 1319 } 1320 1321 bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance, 1322 uint64_t TypeByteSize) { 1323 // If loads occur at a distance that is not a multiple of a feasible vector 1324 // factor store-load forwarding does not take place. 1325 // Positive dependences might cause troubles because vectorizing them might 1326 // prevent store-load forwarding making vectorized code run a lot slower. 1327 // a[i] = a[i-3] ^ a[i-8]; 1328 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and 1329 // hence on your typical architecture store-load forwarding does not take 1330 // place. Vectorizing in such cases does not make sense. 1331 // Store-load forwarding distance. 1332 1333 // After this many iterations store-to-load forwarding conflicts should not 1334 // cause any slowdowns. 1335 const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize; 1336 // Maximum vector factor. 1337 uint64_t MaxVFWithoutSLForwardIssues = std::min( 1338 VectorizerParams::MaxVectorWidth * TypeByteSize, MaxSafeDepDistBytes); 1339 1340 // Compute the smallest VF at which the store and load would be misaligned. 1341 for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues; 1342 VF *= 2) { 1343 // If the number of vector iteration between the store and the load are 1344 // small we could incur conflicts. 1345 if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) { 1346 MaxVFWithoutSLForwardIssues = (VF >>= 1); 1347 break; 1348 } 1349 } 1350 1351 if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) { 1352 LLVM_DEBUG( 1353 dbgs() << "LAA: Distance " << Distance 1354 << " that could cause a store-load forwarding conflict\n"); 1355 return true; 1356 } 1357 1358 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes && 1359 MaxVFWithoutSLForwardIssues != 1360 VectorizerParams::MaxVectorWidth * TypeByteSize) 1361 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues; 1362 return false; 1363 } 1364 1365 void MemoryDepChecker::mergeInStatus(VectorizationSafetyStatus S) { 1366 if (Status < S) 1367 Status = S; 1368 } 1369 1370 /// Given a non-constant (unknown) dependence-distance \p Dist between two 1371 /// memory accesses, that have the same stride whose absolute value is given 1372 /// in \p Stride, and that have the same type size \p TypeByteSize, 1373 /// in a loop whose takenCount is \p BackedgeTakenCount, check if it is 1374 /// possible to prove statically that the dependence distance is larger 1375 /// than the range that the accesses will travel through the execution of 1376 /// the loop. If so, return true; false otherwise. This is useful for 1377 /// example in loops such as the following (PR31098): 1378 /// for (i = 0; i < D; ++i) { 1379 /// = out[i]; 1380 /// out[i+D] = 1381 /// } 1382 static bool isSafeDependenceDistance(const DataLayout &DL, ScalarEvolution &SE, 1383 const SCEV &BackedgeTakenCount, 1384 const SCEV &Dist, uint64_t Stride, 1385 uint64_t TypeByteSize) { 1386 1387 // If we can prove that 1388 // (**) |Dist| > BackedgeTakenCount * Step 1389 // where Step is the absolute stride of the memory accesses in bytes, 1390 // then there is no dependence. 1391 // 1392 // Rationale: 1393 // We basically want to check if the absolute distance (|Dist/Step|) 1394 // is >= the loop iteration count (or > BackedgeTakenCount). 1395 // This is equivalent to the Strong SIV Test (Practical Dependence Testing, 1396 // Section 4.2.1); Note, that for vectorization it is sufficient to prove 1397 // that the dependence distance is >= VF; This is checked elsewhere. 1398 // But in some cases we can prune unknown dependence distances early, and 1399 // even before selecting the VF, and without a runtime test, by comparing 1400 // the distance against the loop iteration count. Since the vectorized code 1401 // will be executed only if LoopCount >= VF, proving distance >= LoopCount 1402 // also guarantees that distance >= VF. 1403 // 1404 const uint64_t ByteStride = Stride * TypeByteSize; 1405 const SCEV *Step = SE.getConstant(BackedgeTakenCount.getType(), ByteStride); 1406 const SCEV *Product = SE.getMulExpr(&BackedgeTakenCount, Step); 1407 1408 const SCEV *CastedDist = &Dist; 1409 const SCEV *CastedProduct = Product; 1410 uint64_t DistTypeSize = DL.getTypeAllocSize(Dist.getType()); 1411 uint64_t ProductTypeSize = DL.getTypeAllocSize(Product->getType()); 1412 1413 // The dependence distance can be positive/negative, so we sign extend Dist; 1414 // The multiplication of the absolute stride in bytes and the 1415 // backedgeTakenCount is non-negative, so we zero extend Product. 1416 if (DistTypeSize > ProductTypeSize) 1417 CastedProduct = SE.getZeroExtendExpr(Product, Dist.getType()); 1418 else 1419 CastedDist = SE.getNoopOrSignExtend(&Dist, Product->getType()); 1420 1421 // Is Dist - (BackedgeTakenCount * Step) > 0 ? 1422 // (If so, then we have proven (**) because |Dist| >= Dist) 1423 const SCEV *Minus = SE.getMinusSCEV(CastedDist, CastedProduct); 1424 if (SE.isKnownPositive(Minus)) 1425 return true; 1426 1427 // Second try: Is -Dist - (BackedgeTakenCount * Step) > 0 ? 1428 // (If so, then we have proven (**) because |Dist| >= -1*Dist) 1429 const SCEV *NegDist = SE.getNegativeSCEV(CastedDist); 1430 Minus = SE.getMinusSCEV(NegDist, CastedProduct); 1431 if (SE.isKnownPositive(Minus)) 1432 return true; 1433 1434 return false; 1435 } 1436 1437 /// Check the dependence for two accesses with the same stride \p Stride. 1438 /// \p Distance is the positive distance and \p TypeByteSize is type size in 1439 /// bytes. 1440 /// 1441 /// \returns true if they are independent. 1442 static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride, 1443 uint64_t TypeByteSize) { 1444 assert(Stride > 1 && "The stride must be greater than 1"); 1445 assert(TypeByteSize > 0 && "The type size in byte must be non-zero"); 1446 assert(Distance > 0 && "The distance must be non-zero"); 1447 1448 // Skip if the distance is not multiple of type byte size. 1449 if (Distance % TypeByteSize) 1450 return false; 1451 1452 uint64_t ScaledDist = Distance / TypeByteSize; 1453 1454 // No dependence if the scaled distance is not multiple of the stride. 1455 // E.g. 1456 // for (i = 0; i < 1024 ; i += 4) 1457 // A[i+2] = A[i] + 1; 1458 // 1459 // Two accesses in memory (scaled distance is 2, stride is 4): 1460 // | A[0] | | | | A[4] | | | | 1461 // | | | A[2] | | | | A[6] | | 1462 // 1463 // E.g. 1464 // for (i = 0; i < 1024 ; i += 3) 1465 // A[i+4] = A[i] + 1; 1466 // 1467 // Two accesses in memory (scaled distance is 4, stride is 3): 1468 // | A[0] | | | A[3] | | | A[6] | | | 1469 // | | | | | A[4] | | | A[7] | | 1470 return ScaledDist % Stride; 1471 } 1472 1473 MemoryDepChecker::Dependence::DepType 1474 MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx, 1475 const MemAccessInfo &B, unsigned BIdx, 1476 const ValueToValueMap &Strides) { 1477 assert (AIdx < BIdx && "Must pass arguments in program order"); 1478 1479 Value *APtr = A.getPointer(); 1480 Value *BPtr = B.getPointer(); 1481 bool AIsWrite = A.getInt(); 1482 bool BIsWrite = B.getInt(); 1483 1484 // Two reads are independent. 1485 if (!AIsWrite && !BIsWrite) 1486 return Dependence::NoDep; 1487 1488 // We cannot check pointers in different address spaces. 1489 if (APtr->getType()->getPointerAddressSpace() != 1490 BPtr->getType()->getPointerAddressSpace()) 1491 return Dependence::Unknown; 1492 1493 int64_t StrideAPtr = getPtrStride(PSE, APtr, InnermostLoop, Strides, true); 1494 int64_t StrideBPtr = getPtrStride(PSE, BPtr, InnermostLoop, Strides, true); 1495 1496 const SCEV *Src = PSE.getSCEV(APtr); 1497 const SCEV *Sink = PSE.getSCEV(BPtr); 1498 1499 // If the induction step is negative we have to invert source and sink of the 1500 // dependence. 1501 if (StrideAPtr < 0) { 1502 std::swap(APtr, BPtr); 1503 std::swap(Src, Sink); 1504 std::swap(AIsWrite, BIsWrite); 1505 std::swap(AIdx, BIdx); 1506 std::swap(StrideAPtr, StrideBPtr); 1507 } 1508 1509 const SCEV *Dist = PSE.getSE()->getMinusSCEV(Sink, Src); 1510 1511 LLVM_DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink 1512 << "(Induction step: " << StrideAPtr << ")\n"); 1513 LLVM_DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to " 1514 << *InstMap[BIdx] << ": " << *Dist << "\n"); 1515 1516 // Need accesses with constant stride. We don't want to vectorize 1517 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in 1518 // the address space. 1519 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){ 1520 LLVM_DEBUG(dbgs() << "Pointer access with non-constant stride\n"); 1521 return Dependence::Unknown; 1522 } 1523 1524 Type *ATy = APtr->getType()->getPointerElementType(); 1525 Type *BTy = BPtr->getType()->getPointerElementType(); 1526 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout(); 1527 uint64_t TypeByteSize = DL.getTypeAllocSize(ATy); 1528 uint64_t Stride = std::abs(StrideAPtr); 1529 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist); 1530 if (!C) { 1531 if (TypeByteSize == DL.getTypeAllocSize(BTy) && 1532 isSafeDependenceDistance(DL, *(PSE.getSE()), 1533 *(PSE.getBackedgeTakenCount()), *Dist, Stride, 1534 TypeByteSize)) 1535 return Dependence::NoDep; 1536 1537 LLVM_DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n"); 1538 FoundNonConstantDistanceDependence = true; 1539 return Dependence::Unknown; 1540 } 1541 1542 const APInt &Val = C->getAPInt(); 1543 int64_t Distance = Val.getSExtValue(); 1544 1545 // Attempt to prove strided accesses independent. 1546 if (std::abs(Distance) > 0 && Stride > 1 && ATy == BTy && 1547 areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) { 1548 LLVM_DEBUG(dbgs() << "LAA: Strided accesses are independent\n"); 1549 return Dependence::NoDep; 1550 } 1551 1552 // Negative distances are not plausible dependencies. 1553 if (Val.isNegative()) { 1554 bool IsTrueDataDependence = (AIsWrite && !BIsWrite); 1555 if (IsTrueDataDependence && EnableForwardingConflictDetection && 1556 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) || 1557 ATy != BTy)) { 1558 LLVM_DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n"); 1559 return Dependence::ForwardButPreventsForwarding; 1560 } 1561 1562 LLVM_DEBUG(dbgs() << "LAA: Dependence is negative\n"); 1563 return Dependence::Forward; 1564 } 1565 1566 // Write to the same location with the same size. 1567 // Could be improved to assert type sizes are the same (i32 == float, etc). 1568 if (Val == 0) { 1569 if (ATy == BTy) 1570 return Dependence::Forward; 1571 LLVM_DEBUG( 1572 dbgs() << "LAA: Zero dependence difference but different types\n"); 1573 return Dependence::Unknown; 1574 } 1575 1576 assert(Val.isStrictlyPositive() && "Expect a positive value"); 1577 1578 if (ATy != BTy) { 1579 LLVM_DEBUG( 1580 dbgs() 1581 << "LAA: ReadWrite-Write positive dependency with different types\n"); 1582 return Dependence::Unknown; 1583 } 1584 1585 // Bail out early if passed-in parameters make vectorization not feasible. 1586 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ? 1587 VectorizerParams::VectorizationFactor : 1); 1588 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ? 1589 VectorizerParams::VectorizationInterleave : 1); 1590 // The minimum number of iterations for a vectorized/unrolled version. 1591 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U); 1592 1593 // It's not vectorizable if the distance is smaller than the minimum distance 1594 // needed for a vectroized/unrolled version. Vectorizing one iteration in 1595 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs 1596 // TypeByteSize (No need to plus the last gap distance). 1597 // 1598 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes. 1599 // foo(int *A) { 1600 // int *B = (int *)((char *)A + 14); 1601 // for (i = 0 ; i < 1024 ; i += 2) 1602 // B[i] = A[i] + 1; 1603 // } 1604 // 1605 // Two accesses in memory (stride is 2): 1606 // | A[0] | | A[2] | | A[4] | | A[6] | | 1607 // | B[0] | | B[2] | | B[4] | 1608 // 1609 // Distance needs for vectorizing iterations except the last iteration: 1610 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4. 1611 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4. 1612 // 1613 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is 1614 // 12, which is less than distance. 1615 // 1616 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4), 1617 // the minimum distance needed is 28, which is greater than distance. It is 1618 // not safe to do vectorization. 1619 uint64_t MinDistanceNeeded = 1620 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize; 1621 if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) { 1622 LLVM_DEBUG(dbgs() << "LAA: Failure because of positive distance " 1623 << Distance << '\n'); 1624 return Dependence::Backward; 1625 } 1626 1627 // Unsafe if the minimum distance needed is greater than max safe distance. 1628 if (MinDistanceNeeded > MaxSafeDepDistBytes) { 1629 LLVM_DEBUG(dbgs() << "LAA: Failure because it needs at least " 1630 << MinDistanceNeeded << " size in bytes"); 1631 return Dependence::Backward; 1632 } 1633 1634 // Positive distance bigger than max vectorization factor. 1635 // FIXME: Should use max factor instead of max distance in bytes, which could 1636 // not handle different types. 1637 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes. 1638 // void foo (int *A, char *B) { 1639 // for (unsigned i = 0; i < 1024; i++) { 1640 // A[i+2] = A[i] + 1; 1641 // B[i+2] = B[i] + 1; 1642 // } 1643 // } 1644 // 1645 // This case is currently unsafe according to the max safe distance. If we 1646 // analyze the two accesses on array B, the max safe dependence distance 1647 // is 2. Then we analyze the accesses on array A, the minimum distance needed 1648 // is 8, which is less than 2 and forbidden vectorization, But actually 1649 // both A and B could be vectorized by 2 iterations. 1650 MaxSafeDepDistBytes = 1651 std::min(static_cast<uint64_t>(Distance), MaxSafeDepDistBytes); 1652 1653 bool IsTrueDataDependence = (!AIsWrite && BIsWrite); 1654 if (IsTrueDataDependence && EnableForwardingConflictDetection && 1655 couldPreventStoreLoadForward(Distance, TypeByteSize)) 1656 return Dependence::BackwardVectorizableButPreventsForwarding; 1657 1658 uint64_t MaxVF = MaxSafeDepDistBytes / (TypeByteSize * Stride); 1659 LLVM_DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue() 1660 << " with max VF = " << MaxVF << '\n'); 1661 uint64_t MaxVFInBits = MaxVF * TypeByteSize * 8; 1662 MaxSafeRegisterWidth = std::min(MaxSafeRegisterWidth, MaxVFInBits); 1663 return Dependence::BackwardVectorizable; 1664 } 1665 1666 bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets, 1667 MemAccessInfoList &CheckDeps, 1668 const ValueToValueMap &Strides) { 1669 1670 MaxSafeDepDistBytes = -1; 1671 SmallPtrSet<MemAccessInfo, 8> Visited; 1672 for (MemAccessInfo CurAccess : CheckDeps) { 1673 if (Visited.count(CurAccess)) 1674 continue; 1675 1676 // Get the relevant memory access set. 1677 EquivalenceClasses<MemAccessInfo>::iterator I = 1678 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess)); 1679 1680 // Check accesses within this set. 1681 EquivalenceClasses<MemAccessInfo>::member_iterator AI = 1682 AccessSets.member_begin(I); 1683 EquivalenceClasses<MemAccessInfo>::member_iterator AE = 1684 AccessSets.member_end(); 1685 1686 // Check every access pair. 1687 while (AI != AE) { 1688 Visited.insert(*AI); 1689 bool AIIsWrite = AI->getInt(); 1690 // Check loads only against next equivalent class, but stores also against 1691 // other stores in the same equivalence class - to the same address. 1692 EquivalenceClasses<MemAccessInfo>::member_iterator OI = 1693 (AIIsWrite ? AI : std::next(AI)); 1694 while (OI != AE) { 1695 // Check every accessing instruction pair in program order. 1696 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(), 1697 I1E = Accesses[*AI].end(); I1 != I1E; ++I1) 1698 // Scan all accesses of another equivalence class, but only the next 1699 // accesses of the same equivalent class. 1700 for (std::vector<unsigned>::iterator 1701 I2 = (OI == AI ? std::next(I1) : Accesses[*OI].begin()), 1702 I2E = (OI == AI ? I1E : Accesses[*OI].end()); 1703 I2 != I2E; ++I2) { 1704 auto A = std::make_pair(&*AI, *I1); 1705 auto B = std::make_pair(&*OI, *I2); 1706 1707 assert(*I1 != *I2); 1708 if (*I1 > *I2) 1709 std::swap(A, B); 1710 1711 Dependence::DepType Type = 1712 isDependent(*A.first, A.second, *B.first, B.second, Strides); 1713 mergeInStatus(Dependence::isSafeForVectorization(Type)); 1714 1715 // Gather dependences unless we accumulated MaxDependences 1716 // dependences. In that case return as soon as we find the first 1717 // unsafe dependence. This puts a limit on this quadratic 1718 // algorithm. 1719 if (RecordDependences) { 1720 if (Type != Dependence::NoDep) 1721 Dependences.push_back(Dependence(A.second, B.second, Type)); 1722 1723 if (Dependences.size() >= MaxDependences) { 1724 RecordDependences = false; 1725 Dependences.clear(); 1726 LLVM_DEBUG(dbgs() 1727 << "Too many dependences, stopped recording\n"); 1728 } 1729 } 1730 if (!RecordDependences && !isSafeForVectorization()) 1731 return false; 1732 } 1733 ++OI; 1734 } 1735 AI++; 1736 } 1737 } 1738 1739 LLVM_DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n"); 1740 return isSafeForVectorization(); 1741 } 1742 1743 SmallVector<Instruction *, 4> 1744 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const { 1745 MemAccessInfo Access(Ptr, isWrite); 1746 auto &IndexVector = Accesses.find(Access)->second; 1747 1748 SmallVector<Instruction *, 4> Insts; 1749 transform(IndexVector, 1750 std::back_inserter(Insts), 1751 [&](unsigned Idx) { return this->InstMap[Idx]; }); 1752 return Insts; 1753 } 1754 1755 const char *MemoryDepChecker::Dependence::DepName[] = { 1756 "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward", 1757 "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"}; 1758 1759 void MemoryDepChecker::Dependence::print( 1760 raw_ostream &OS, unsigned Depth, 1761 const SmallVectorImpl<Instruction *> &Instrs) const { 1762 OS.indent(Depth) << DepName[Type] << ":\n"; 1763 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n"; 1764 OS.indent(Depth + 2) << *Instrs[Destination] << "\n"; 1765 } 1766 1767 bool LoopAccessInfo::canAnalyzeLoop() { 1768 // We need to have a loop header. 1769 LLVM_DEBUG(dbgs() << "LAA: Found a loop in " 1770 << TheLoop->getHeader()->getParent()->getName() << ": " 1771 << TheLoop->getHeader()->getName() << '\n'); 1772 1773 // We can only analyze innermost loops. 1774 if (!TheLoop->empty()) { 1775 LLVM_DEBUG(dbgs() << "LAA: loop is not the innermost loop\n"); 1776 recordAnalysis("NotInnerMostLoop") << "loop is not the innermost loop"; 1777 return false; 1778 } 1779 1780 // We must have a single backedge. 1781 if (TheLoop->getNumBackEdges() != 1) { 1782 LLVM_DEBUG( 1783 dbgs() << "LAA: loop control flow is not understood by analyzer\n"); 1784 recordAnalysis("CFGNotUnderstood") 1785 << "loop control flow is not understood by analyzer"; 1786 return false; 1787 } 1788 1789 // We must have a single exiting block. 1790 if (!TheLoop->getExitingBlock()) { 1791 LLVM_DEBUG( 1792 dbgs() << "LAA: loop control flow is not understood by analyzer\n"); 1793 recordAnalysis("CFGNotUnderstood") 1794 << "loop control flow is not understood by analyzer"; 1795 return false; 1796 } 1797 1798 // We only handle bottom-tested loops, i.e. loop in which the condition is 1799 // checked at the end of each iteration. With that we can assume that all 1800 // instructions in the loop are executed the same number of times. 1801 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) { 1802 LLVM_DEBUG( 1803 dbgs() << "LAA: loop control flow is not understood by analyzer\n"); 1804 recordAnalysis("CFGNotUnderstood") 1805 << "loop control flow is not understood by analyzer"; 1806 return false; 1807 } 1808 1809 // ScalarEvolution needs to be able to find the exit count. 1810 const SCEV *ExitCount = PSE->getBackedgeTakenCount(); 1811 if (ExitCount == PSE->getSE()->getCouldNotCompute()) { 1812 recordAnalysis("CantComputeNumberOfIterations") 1813 << "could not determine number of loop iterations"; 1814 LLVM_DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n"); 1815 return false; 1816 } 1817 1818 return true; 1819 } 1820 1821 void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI, 1822 const TargetLibraryInfo *TLI, 1823 DominatorTree *DT) { 1824 typedef SmallPtrSet<Value*, 16> ValueSet; 1825 1826 // Holds the Load and Store instructions. 1827 SmallVector<LoadInst *, 16> Loads; 1828 SmallVector<StoreInst *, 16> Stores; 1829 1830 // Holds all the different accesses in the loop. 1831 unsigned NumReads = 0; 1832 unsigned NumReadWrites = 0; 1833 1834 bool HasComplexMemInst = false; 1835 1836 // A runtime check is only legal to insert if there are no convergent calls. 1837 HasConvergentOp = false; 1838 1839 PtrRtChecking->Pointers.clear(); 1840 PtrRtChecking->Need = false; 1841 1842 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel(); 1843 1844 const bool EnableMemAccessVersioningOfLoop = 1845 EnableMemAccessVersioning && 1846 !TheLoop->getHeader()->getParent()->hasOptSize(); 1847 1848 // For each block. 1849 for (BasicBlock *BB : TheLoop->blocks()) { 1850 // Scan the BB and collect legal loads and stores. Also detect any 1851 // convergent instructions. 1852 for (Instruction &I : *BB) { 1853 if (auto *Call = dyn_cast<CallBase>(&I)) { 1854 if (Call->isConvergent()) 1855 HasConvergentOp = true; 1856 } 1857 1858 // With both a non-vectorizable memory instruction and a convergent 1859 // operation, found in this loop, no reason to continue the search. 1860 if (HasComplexMemInst && HasConvergentOp) { 1861 CanVecMem = false; 1862 return; 1863 } 1864 1865 // Avoid hitting recordAnalysis multiple times. 1866 if (HasComplexMemInst) 1867 continue; 1868 1869 // If this is a load, save it. If this instruction can read from memory 1870 // but is not a load, then we quit. Notice that we don't handle function 1871 // calls that read or write. 1872 if (I.mayReadFromMemory()) { 1873 // Many math library functions read the rounding mode. We will only 1874 // vectorize a loop if it contains known function calls that don't set 1875 // the flag. Therefore, it is safe to ignore this read from memory. 1876 auto *Call = dyn_cast<CallInst>(&I); 1877 if (Call && getVectorIntrinsicIDForCall(Call, TLI)) 1878 continue; 1879 1880 // If the function has an explicit vectorized counterpart, we can safely 1881 // assume that it can be vectorized. 1882 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() && 1883 !VFDatabase::getMappings(*Call).empty()) 1884 continue; 1885 1886 auto *Ld = dyn_cast<LoadInst>(&I); 1887 if (!Ld) { 1888 recordAnalysis("CantVectorizeInstruction", Ld) 1889 << "instruction cannot be vectorized"; 1890 HasComplexMemInst = true; 1891 continue; 1892 } 1893 if (!Ld->isSimple() && !IsAnnotatedParallel) { 1894 recordAnalysis("NonSimpleLoad", Ld) 1895 << "read with atomic ordering or volatile read"; 1896 LLVM_DEBUG(dbgs() << "LAA: Found a non-simple load.\n"); 1897 HasComplexMemInst = true; 1898 continue; 1899 } 1900 NumLoads++; 1901 Loads.push_back(Ld); 1902 DepChecker->addAccess(Ld); 1903 if (EnableMemAccessVersioningOfLoop) 1904 collectStridedAccess(Ld); 1905 continue; 1906 } 1907 1908 // Save 'store' instructions. Abort if other instructions write to memory. 1909 if (I.mayWriteToMemory()) { 1910 auto *St = dyn_cast<StoreInst>(&I); 1911 if (!St) { 1912 recordAnalysis("CantVectorizeInstruction", St) 1913 << "instruction cannot be vectorized"; 1914 HasComplexMemInst = true; 1915 continue; 1916 } 1917 if (!St->isSimple() && !IsAnnotatedParallel) { 1918 recordAnalysis("NonSimpleStore", St) 1919 << "write with atomic ordering or volatile write"; 1920 LLVM_DEBUG(dbgs() << "LAA: Found a non-simple store.\n"); 1921 HasComplexMemInst = true; 1922 continue; 1923 } 1924 NumStores++; 1925 Stores.push_back(St); 1926 DepChecker->addAccess(St); 1927 if (EnableMemAccessVersioningOfLoop) 1928 collectStridedAccess(St); 1929 } 1930 } // Next instr. 1931 } // Next block. 1932 1933 if (HasComplexMemInst) { 1934 CanVecMem = false; 1935 return; 1936 } 1937 1938 // Now we have two lists that hold the loads and the stores. 1939 // Next, we find the pointers that they use. 1940 1941 // Check if we see any stores. If there are no stores, then we don't 1942 // care if the pointers are *restrict*. 1943 if (!Stores.size()) { 1944 LLVM_DEBUG(dbgs() << "LAA: Found a read-only loop!\n"); 1945 CanVecMem = true; 1946 return; 1947 } 1948 1949 MemoryDepChecker::DepCandidates DependentAccesses; 1950 AccessAnalysis Accesses(TheLoop->getHeader()->getModule()->getDataLayout(), 1951 TheLoop, AA, LI, DependentAccesses, *PSE); 1952 1953 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects 1954 // multiple times on the same object. If the ptr is accessed twice, once 1955 // for read and once for write, it will only appear once (on the write 1956 // list). This is okay, since we are going to check for conflicts between 1957 // writes and between reads and writes, but not between reads and reads. 1958 ValueSet Seen; 1959 1960 // Record uniform store addresses to identify if we have multiple stores 1961 // to the same address. 1962 ValueSet UniformStores; 1963 1964 for (StoreInst *ST : Stores) { 1965 Value *Ptr = ST->getPointerOperand(); 1966 1967 if (isUniform(Ptr)) 1968 HasDependenceInvolvingLoopInvariantAddress |= 1969 !UniformStores.insert(Ptr).second; 1970 1971 // If we did *not* see this pointer before, insert it to the read-write 1972 // list. At this phase it is only a 'write' list. 1973 if (Seen.insert(Ptr).second) { 1974 ++NumReadWrites; 1975 1976 MemoryLocation Loc = MemoryLocation::get(ST); 1977 // The TBAA metadata could have a control dependency on the predication 1978 // condition, so we cannot rely on it when determining whether or not we 1979 // need runtime pointer checks. 1980 if (blockNeedsPredication(ST->getParent(), TheLoop, DT)) 1981 Loc.AATags.TBAA = nullptr; 1982 1983 Accesses.addStore(Loc); 1984 } 1985 } 1986 1987 if (IsAnnotatedParallel) { 1988 LLVM_DEBUG( 1989 dbgs() << "LAA: A loop annotated parallel, ignore memory dependency " 1990 << "checks.\n"); 1991 CanVecMem = true; 1992 return; 1993 } 1994 1995 for (LoadInst *LD : Loads) { 1996 Value *Ptr = LD->getPointerOperand(); 1997 // If we did *not* see this pointer before, insert it to the 1998 // read list. If we *did* see it before, then it is already in 1999 // the read-write list. This allows us to vectorize expressions 2000 // such as A[i] += x; Because the address of A[i] is a read-write 2001 // pointer. This only works if the index of A[i] is consecutive. 2002 // If the address of i is unknown (for example A[B[i]]) then we may 2003 // read a few words, modify, and write a few words, and some of the 2004 // words may be written to the same address. 2005 bool IsReadOnlyPtr = false; 2006 if (Seen.insert(Ptr).second || 2007 !getPtrStride(*PSE, Ptr, TheLoop, SymbolicStrides)) { 2008 ++NumReads; 2009 IsReadOnlyPtr = true; 2010 } 2011 2012 // See if there is an unsafe dependency between a load to a uniform address and 2013 // store to the same uniform address. 2014 if (UniformStores.count(Ptr)) { 2015 LLVM_DEBUG(dbgs() << "LAA: Found an unsafe dependency between a uniform " 2016 "load and uniform store to the same address!\n"); 2017 HasDependenceInvolvingLoopInvariantAddress = true; 2018 } 2019 2020 MemoryLocation Loc = MemoryLocation::get(LD); 2021 // The TBAA metadata could have a control dependency on the predication 2022 // condition, so we cannot rely on it when determining whether or not we 2023 // need runtime pointer checks. 2024 if (blockNeedsPredication(LD->getParent(), TheLoop, DT)) 2025 Loc.AATags.TBAA = nullptr; 2026 2027 Accesses.addLoad(Loc, IsReadOnlyPtr); 2028 } 2029 2030 // If we write (or read-write) to a single destination and there are no 2031 // other reads in this loop then is it safe to vectorize. 2032 if (NumReadWrites == 1 && NumReads == 0) { 2033 LLVM_DEBUG(dbgs() << "LAA: Found a write-only loop!\n"); 2034 CanVecMem = true; 2035 return; 2036 } 2037 2038 // Build dependence sets and check whether we need a runtime pointer bounds 2039 // check. 2040 Accesses.buildDependenceSets(); 2041 2042 // Find pointers with computable bounds. We are going to use this information 2043 // to place a runtime bound check. 2044 bool CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(), 2045 TheLoop, SymbolicStrides); 2046 if (!CanDoRTIfNeeded) { 2047 recordAnalysis("CantIdentifyArrayBounds") << "cannot identify array bounds"; 2048 LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because we can't find " 2049 << "the array bounds.\n"); 2050 CanVecMem = false; 2051 return; 2052 } 2053 2054 LLVM_DEBUG( 2055 dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n"); 2056 2057 CanVecMem = true; 2058 if (Accesses.isDependencyCheckNeeded()) { 2059 LLVM_DEBUG(dbgs() << "LAA: Checking memory dependencies\n"); 2060 CanVecMem = DepChecker->areDepsSafe( 2061 DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides); 2062 MaxSafeDepDistBytes = DepChecker->getMaxSafeDepDistBytes(); 2063 2064 if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) { 2065 LLVM_DEBUG(dbgs() << "LAA: Retrying with memory checks\n"); 2066 2067 // Clear the dependency checks. We assume they are not needed. 2068 Accesses.resetDepChecks(*DepChecker); 2069 2070 PtrRtChecking->reset(); 2071 PtrRtChecking->Need = true; 2072 2073 auto *SE = PSE->getSE(); 2074 CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, SE, TheLoop, 2075 SymbolicStrides, true); 2076 2077 // Check that we found the bounds for the pointer. 2078 if (!CanDoRTIfNeeded) { 2079 recordAnalysis("CantCheckMemDepsAtRunTime") 2080 << "cannot check memory dependencies at runtime"; 2081 LLVM_DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n"); 2082 CanVecMem = false; 2083 return; 2084 } 2085 2086 CanVecMem = true; 2087 } 2088 } 2089 2090 if (HasConvergentOp) { 2091 recordAnalysis("CantInsertRuntimeCheckWithConvergent") 2092 << "cannot add control dependency to convergent operation"; 2093 LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because a runtime check " 2094 "would be needed with a convergent operation\n"); 2095 CanVecMem = false; 2096 return; 2097 } 2098 2099 if (CanVecMem) 2100 LLVM_DEBUG( 2101 dbgs() << "LAA: No unsafe dependent memory operations in loop. We" 2102 << (PtrRtChecking->Need ? "" : " don't") 2103 << " need runtime memory checks.\n"); 2104 else { 2105 recordAnalysis("UnsafeMemDep") 2106 << "unsafe dependent memory operations in loop. Use " 2107 "#pragma loop distribute(enable) to allow loop distribution " 2108 "to attempt to isolate the offending operations into a separate " 2109 "loop"; 2110 LLVM_DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n"); 2111 } 2112 } 2113 2114 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop, 2115 DominatorTree *DT) { 2116 assert(TheLoop->contains(BB) && "Unknown block used"); 2117 2118 // Blocks that do not dominate the latch need predication. 2119 BasicBlock* Latch = TheLoop->getLoopLatch(); 2120 return !DT->dominates(BB, Latch); 2121 } 2122 2123 OptimizationRemarkAnalysis &LoopAccessInfo::recordAnalysis(StringRef RemarkName, 2124 Instruction *I) { 2125 assert(!Report && "Multiple reports generated"); 2126 2127 Value *CodeRegion = TheLoop->getHeader(); 2128 DebugLoc DL = TheLoop->getStartLoc(); 2129 2130 if (I) { 2131 CodeRegion = I->getParent(); 2132 // If there is no debug location attached to the instruction, revert back to 2133 // using the loop's. 2134 if (I->getDebugLoc()) 2135 DL = I->getDebugLoc(); 2136 } 2137 2138 Report = std::make_unique<OptimizationRemarkAnalysis>(DEBUG_TYPE, RemarkName, DL, 2139 CodeRegion); 2140 return *Report; 2141 } 2142 2143 bool LoopAccessInfo::isUniform(Value *V) const { 2144 auto *SE = PSE->getSE(); 2145 // Since we rely on SCEV for uniformity, if the type is not SCEVable, it is 2146 // never considered uniform. 2147 // TODO: Is this really what we want? Even without FP SCEV, we may want some 2148 // trivially loop-invariant FP values to be considered uniform. 2149 if (!SE->isSCEVable(V->getType())) 2150 return false; 2151 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); 2152 } 2153 2154 void LoopAccessInfo::collectStridedAccess(Value *MemAccess) { 2155 Value *Ptr = nullptr; 2156 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess)) 2157 Ptr = LI->getPointerOperand(); 2158 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess)) 2159 Ptr = SI->getPointerOperand(); 2160 else 2161 return; 2162 2163 Value *Stride = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop); 2164 if (!Stride) 2165 return; 2166 2167 LLVM_DEBUG(dbgs() << "LAA: Found a strided access that is a candidate for " 2168 "versioning:"); 2169 LLVM_DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n"); 2170 2171 // Avoid adding the "Stride == 1" predicate when we know that 2172 // Stride >= Trip-Count. Such a predicate will effectively optimize a single 2173 // or zero iteration loop, as Trip-Count <= Stride == 1. 2174 // 2175 // TODO: We are currently not making a very informed decision on when it is 2176 // beneficial to apply stride versioning. It might make more sense that the 2177 // users of this analysis (such as the vectorizer) will trigger it, based on 2178 // their specific cost considerations; For example, in cases where stride 2179 // versioning does not help resolving memory accesses/dependences, the 2180 // vectorizer should evaluate the cost of the runtime test, and the benefit 2181 // of various possible stride specializations, considering the alternatives 2182 // of using gather/scatters (if available). 2183 2184 const SCEV *StrideExpr = PSE->getSCEV(Stride); 2185 const SCEV *BETakenCount = PSE->getBackedgeTakenCount(); 2186 2187 // Match the types so we can compare the stride and the BETakenCount. 2188 // The Stride can be positive/negative, so we sign extend Stride; 2189 // The backedgeTakenCount is non-negative, so we zero extend BETakenCount. 2190 const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout(); 2191 uint64_t StrideTypeSize = DL.getTypeAllocSize(StrideExpr->getType()); 2192 uint64_t BETypeSize = DL.getTypeAllocSize(BETakenCount->getType()); 2193 const SCEV *CastedStride = StrideExpr; 2194 const SCEV *CastedBECount = BETakenCount; 2195 ScalarEvolution *SE = PSE->getSE(); 2196 if (BETypeSize >= StrideTypeSize) 2197 CastedStride = SE->getNoopOrSignExtend(StrideExpr, BETakenCount->getType()); 2198 else 2199 CastedBECount = SE->getZeroExtendExpr(BETakenCount, StrideExpr->getType()); 2200 const SCEV *StrideMinusBETaken = SE->getMinusSCEV(CastedStride, CastedBECount); 2201 // Since TripCount == BackEdgeTakenCount + 1, checking: 2202 // "Stride >= TripCount" is equivalent to checking: 2203 // Stride - BETakenCount > 0 2204 if (SE->isKnownPositive(StrideMinusBETaken)) { 2205 LLVM_DEBUG( 2206 dbgs() << "LAA: Stride>=TripCount; No point in versioning as the " 2207 "Stride==1 predicate will imply that the loop executes " 2208 "at most once.\n"); 2209 return; 2210 } 2211 LLVM_DEBUG(dbgs() << "LAA: Found a strided access that we can version."); 2212 2213 SymbolicStrides[Ptr] = Stride; 2214 StrideSet.insert(Stride); 2215 } 2216 2217 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE, 2218 const TargetLibraryInfo *TLI, AAResults *AA, 2219 DominatorTree *DT, LoopInfo *LI) 2220 : PSE(std::make_unique<PredicatedScalarEvolution>(*SE, *L)), 2221 PtrRtChecking(std::make_unique<RuntimePointerChecking>(SE)), 2222 DepChecker(std::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L), 2223 NumLoads(0), NumStores(0), MaxSafeDepDistBytes(-1), CanVecMem(false), 2224 HasConvergentOp(false), 2225 HasDependenceInvolvingLoopInvariantAddress(false) { 2226 if (canAnalyzeLoop()) 2227 analyzeLoop(AA, LI, TLI, DT); 2228 } 2229 2230 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const { 2231 if (CanVecMem) { 2232 OS.indent(Depth) << "Memory dependences are safe"; 2233 if (MaxSafeDepDistBytes != -1ULL) 2234 OS << " with a maximum dependence distance of " << MaxSafeDepDistBytes 2235 << " bytes"; 2236 if (PtrRtChecking->Need) 2237 OS << " with run-time checks"; 2238 OS << "\n"; 2239 } 2240 2241 if (HasConvergentOp) 2242 OS.indent(Depth) << "Has convergent operation in loop\n"; 2243 2244 if (Report) 2245 OS.indent(Depth) << "Report: " << Report->getMsg() << "\n"; 2246 2247 if (auto *Dependences = DepChecker->getDependences()) { 2248 OS.indent(Depth) << "Dependences:\n"; 2249 for (auto &Dep : *Dependences) { 2250 Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions()); 2251 OS << "\n"; 2252 } 2253 } else 2254 OS.indent(Depth) << "Too many dependences, not recorded\n"; 2255 2256 // List the pair of accesses need run-time checks to prove independence. 2257 PtrRtChecking->print(OS, Depth); 2258 OS << "\n"; 2259 2260 OS.indent(Depth) << "Non vectorizable stores to invariant address were " 2261 << (HasDependenceInvolvingLoopInvariantAddress ? "" : "not ") 2262 << "found in loop.\n"; 2263 2264 OS.indent(Depth) << "SCEV assumptions:\n"; 2265 PSE->getUnionPredicate().print(OS, Depth); 2266 2267 OS << "\n"; 2268 2269 OS.indent(Depth) << "Expressions re-written:\n"; 2270 PSE->print(OS, Depth); 2271 } 2272 2273 LoopAccessLegacyAnalysis::LoopAccessLegacyAnalysis() : FunctionPass(ID) { 2274 initializeLoopAccessLegacyAnalysisPass(*PassRegistry::getPassRegistry()); 2275 } 2276 2277 const LoopAccessInfo &LoopAccessLegacyAnalysis::getInfo(Loop *L) { 2278 auto &LAI = LoopAccessInfoMap[L]; 2279 2280 if (!LAI) 2281 LAI = std::make_unique<LoopAccessInfo>(L, SE, TLI, AA, DT, LI); 2282 2283 return *LAI.get(); 2284 } 2285 2286 void LoopAccessLegacyAnalysis::print(raw_ostream &OS, const Module *M) const { 2287 LoopAccessLegacyAnalysis &LAA = *const_cast<LoopAccessLegacyAnalysis *>(this); 2288 2289 for (Loop *TopLevelLoop : *LI) 2290 for (Loop *L : depth_first(TopLevelLoop)) { 2291 OS.indent(2) << L->getHeader()->getName() << ":\n"; 2292 auto &LAI = LAA.getInfo(L); 2293 LAI.print(OS, 4); 2294 } 2295 } 2296 2297 bool LoopAccessLegacyAnalysis::runOnFunction(Function &F) { 2298 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 2299 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); 2300 TLI = TLIP ? &TLIP->getTLI(F) : nullptr; 2301 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 2302 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2303 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 2304 2305 return false; 2306 } 2307 2308 void LoopAccessLegacyAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 2309 AU.addRequired<ScalarEvolutionWrapperPass>(); 2310 AU.addRequired<AAResultsWrapperPass>(); 2311 AU.addRequired<DominatorTreeWrapperPass>(); 2312 AU.addRequired<LoopInfoWrapperPass>(); 2313 2314 AU.setPreservesAll(); 2315 } 2316 2317 char LoopAccessLegacyAnalysis::ID = 0; 2318 static const char laa_name[] = "Loop Access Analysis"; 2319 #define LAA_NAME "loop-accesses" 2320 2321 INITIALIZE_PASS_BEGIN(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true) 2322 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 2323 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 2324 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 2325 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 2326 INITIALIZE_PASS_END(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true) 2327 2328 AnalysisKey LoopAccessAnalysis::Key; 2329 2330 LoopAccessInfo LoopAccessAnalysis::run(Loop &L, LoopAnalysisManager &AM, 2331 LoopStandardAnalysisResults &AR) { 2332 return LoopAccessInfo(&L, &AR.SE, &AR.TLI, &AR.AA, &AR.DT, &AR.LI); 2333 } 2334 2335 namespace llvm { 2336 2337 Pass *createLAAPass() { 2338 return new LoopAccessLegacyAnalysis(); 2339 } 2340 2341 } // end namespace llvm 2342