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/EquivalenceClasses.h" 18 #include "llvm/ADT/PointerIntPair.h" 19 #include "llvm/ADT/STLExtras.h" 20 #include "llvm/ADT/SetVector.h" 21 #include "llvm/ADT/SmallPtrSet.h" 22 #include "llvm/ADT/SmallSet.h" 23 #include "llvm/ADT/SmallVector.h" 24 #include "llvm/Analysis/AliasAnalysis.h" 25 #include "llvm/Analysis/AliasSetTracker.h" 26 #include "llvm/Analysis/LoopAnalysisManager.h" 27 #include "llvm/Analysis/LoopInfo.h" 28 #include "llvm/Analysis/LoopIterator.h" 29 #include "llvm/Analysis/MemoryLocation.h" 30 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 31 #include "llvm/Analysis/ScalarEvolution.h" 32 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 33 #include "llvm/Analysis/TargetLibraryInfo.h" 34 #include "llvm/Analysis/ValueTracking.h" 35 #include "llvm/Analysis/VectorUtils.h" 36 #include "llvm/IR/BasicBlock.h" 37 #include "llvm/IR/Constants.h" 38 #include "llvm/IR/DataLayout.h" 39 #include "llvm/IR/DebugLoc.h" 40 #include "llvm/IR/DerivedTypes.h" 41 #include "llvm/IR/DiagnosticInfo.h" 42 #include "llvm/IR/Dominators.h" 43 #include "llvm/IR/Function.h" 44 #include "llvm/IR/GetElementPtrTypeIterator.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/PatternMatch.h" 51 #include "llvm/IR/Type.h" 52 #include "llvm/IR/Value.h" 53 #include "llvm/IR/ValueHandle.h" 54 #include "llvm/Support/Casting.h" 55 #include "llvm/Support/CommandLine.h" 56 #include "llvm/Support/Debug.h" 57 #include "llvm/Support/ErrorHandling.h" 58 #include "llvm/Support/raw_ostream.h" 59 #include <algorithm> 60 #include <cassert> 61 #include <cstdint> 62 #include <iterator> 63 #include <utility> 64 #include <variant> 65 #include <vector> 66 67 using namespace llvm; 68 using namespace llvm::PatternMatch; 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 static cl::opt<unsigned> MaxForkedSCEVDepth( 133 "max-forked-scev-depth", cl::Hidden, 134 cl::desc("Maximum recursion depth when finding forked SCEVs (default = 5)"), 135 cl::init(5)); 136 137 static cl::opt<bool> SpeculateUnitStride( 138 "laa-speculate-unit-stride", cl::Hidden, 139 cl::desc("Speculate that non-constant strides are unit in LAA"), 140 cl::init(true)); 141 142 static cl::opt<bool, true> HoistRuntimeChecks( 143 "hoist-runtime-checks", cl::Hidden, 144 cl::desc( 145 "Hoist inner loop runtime memory checks to outer loop if possible"), 146 cl::location(VectorizerParams::HoistRuntimeChecks), cl::init(true)); 147 bool VectorizerParams::HoistRuntimeChecks; 148 149 bool VectorizerParams::isInterleaveForced() { 150 return ::VectorizationInterleave.getNumOccurrences() > 0; 151 } 152 153 const SCEV *llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE, 154 const DenseMap<Value *, const SCEV *> &PtrToStride, 155 Value *Ptr) { 156 const SCEV *OrigSCEV = PSE.getSCEV(Ptr); 157 158 // If there is an entry in the map return the SCEV of the pointer with the 159 // symbolic stride replaced by one. 160 DenseMap<Value *, const SCEV *>::const_iterator SI = PtrToStride.find(Ptr); 161 if (SI == PtrToStride.end()) 162 // For a non-symbolic stride, just return the original expression. 163 return OrigSCEV; 164 165 const SCEV *StrideSCEV = SI->second; 166 // Note: This assert is both overly strong and overly weak. The actual 167 // invariant here is that StrideSCEV should be loop invariant. The only 168 // such invariant strides we happen to speculate right now are unknowns 169 // and thus this is a reasonable proxy of the actual invariant. 170 assert(isa<SCEVUnknown>(StrideSCEV) && "shouldn't be in map"); 171 172 ScalarEvolution *SE = PSE.getSE(); 173 const auto *CT = SE->getOne(StrideSCEV->getType()); 174 PSE.addPredicate(*SE->getEqualPredicate(StrideSCEV, CT)); 175 auto *Expr = PSE.getSCEV(Ptr); 176 177 LLVM_DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV 178 << " by: " << *Expr << "\n"); 179 return Expr; 180 } 181 182 RuntimeCheckingPtrGroup::RuntimeCheckingPtrGroup( 183 unsigned Index, RuntimePointerChecking &RtCheck) 184 : High(RtCheck.Pointers[Index].End), Low(RtCheck.Pointers[Index].Start), 185 AddressSpace(RtCheck.Pointers[Index] 186 .PointerValue->getType() 187 ->getPointerAddressSpace()), 188 NeedsFreeze(RtCheck.Pointers[Index].NeedsFreeze) { 189 Members.push_back(Index); 190 } 191 192 /// Calculate Start and End points of memory access. 193 /// Let's assume A is the first access and B is a memory access on N-th loop 194 /// iteration. Then B is calculated as: 195 /// B = A + Step*N . 196 /// Step value may be positive or negative. 197 /// N is a calculated back-edge taken count: 198 /// N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0 199 /// Start and End points are calculated in the following way: 200 /// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt, 201 /// where SizeOfElt is the size of single memory access in bytes. 202 /// 203 /// There is no conflict when the intervals are disjoint: 204 /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End) 205 void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, const SCEV *PtrExpr, 206 Type *AccessTy, bool WritePtr, 207 unsigned DepSetId, unsigned ASId, 208 PredicatedScalarEvolution &PSE, 209 bool NeedsFreeze) { 210 ScalarEvolution *SE = PSE.getSE(); 211 212 const SCEV *ScStart; 213 const SCEV *ScEnd; 214 215 if (SE->isLoopInvariant(PtrExpr, Lp)) { 216 ScStart = ScEnd = PtrExpr; 217 } else { 218 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrExpr); 219 assert(AR && "Invalid addrec expression"); 220 const SCEV *Ex = PSE.getBackedgeTakenCount(); 221 222 ScStart = AR->getStart(); 223 ScEnd = AR->evaluateAtIteration(Ex, *SE); 224 const SCEV *Step = AR->getStepRecurrence(*SE); 225 226 // For expressions with negative step, the upper bound is ScStart and the 227 // lower bound is ScEnd. 228 if (const auto *CStep = dyn_cast<SCEVConstant>(Step)) { 229 if (CStep->getValue()->isNegative()) 230 std::swap(ScStart, ScEnd); 231 } else { 232 // Fallback case: the step is not constant, but we can still 233 // get the upper and lower bounds of the interval by using min/max 234 // expressions. 235 ScStart = SE->getUMinExpr(ScStart, ScEnd); 236 ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd); 237 } 238 } 239 assert(SE->isLoopInvariant(ScStart, Lp) && "ScStart needs to be invariant"); 240 assert(SE->isLoopInvariant(ScEnd, Lp)&& "ScEnd needs to be invariant"); 241 242 // Add the size of the pointed element to ScEnd. 243 auto &DL = Lp->getHeader()->getModule()->getDataLayout(); 244 Type *IdxTy = DL.getIndexType(Ptr->getType()); 245 const SCEV *EltSizeSCEV = SE->getStoreSizeOfExpr(IdxTy, AccessTy); 246 ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV); 247 248 Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, PtrExpr, 249 NeedsFreeze); 250 } 251 252 void RuntimePointerChecking::tryToCreateDiffCheck( 253 const RuntimeCheckingPtrGroup &CGI, const RuntimeCheckingPtrGroup &CGJ) { 254 if (!CanUseDiffCheck) 255 return; 256 257 // If either group contains multiple different pointers, bail out. 258 // TODO: Support multiple pointers by using the minimum or maximum pointer, 259 // depending on src & sink. 260 if (CGI.Members.size() != 1 || CGJ.Members.size() != 1) { 261 CanUseDiffCheck = false; 262 return; 263 } 264 265 PointerInfo *Src = &Pointers[CGI.Members[0]]; 266 PointerInfo *Sink = &Pointers[CGJ.Members[0]]; 267 268 // If either pointer is read and written, multiple checks may be needed. Bail 269 // out. 270 if (!DC.getOrderForAccess(Src->PointerValue, !Src->IsWritePtr).empty() || 271 !DC.getOrderForAccess(Sink->PointerValue, !Sink->IsWritePtr).empty()) { 272 CanUseDiffCheck = false; 273 return; 274 } 275 276 ArrayRef<unsigned> AccSrc = 277 DC.getOrderForAccess(Src->PointerValue, Src->IsWritePtr); 278 ArrayRef<unsigned> AccSink = 279 DC.getOrderForAccess(Sink->PointerValue, Sink->IsWritePtr); 280 // If either pointer is accessed multiple times, there may not be a clear 281 // src/sink relation. Bail out for now. 282 if (AccSrc.size() != 1 || AccSink.size() != 1) { 283 CanUseDiffCheck = false; 284 return; 285 } 286 // If the sink is accessed before src, swap src/sink. 287 if (AccSink[0] < AccSrc[0]) 288 std::swap(Src, Sink); 289 290 auto *SrcAR = dyn_cast<SCEVAddRecExpr>(Src->Expr); 291 auto *SinkAR = dyn_cast<SCEVAddRecExpr>(Sink->Expr); 292 if (!SrcAR || !SinkAR || SrcAR->getLoop() != DC.getInnermostLoop() || 293 SinkAR->getLoop() != DC.getInnermostLoop()) { 294 CanUseDiffCheck = false; 295 return; 296 } 297 298 SmallVector<Instruction *, 4> SrcInsts = 299 DC.getInstructionsForAccess(Src->PointerValue, Src->IsWritePtr); 300 SmallVector<Instruction *, 4> SinkInsts = 301 DC.getInstructionsForAccess(Sink->PointerValue, Sink->IsWritePtr); 302 Type *SrcTy = getLoadStoreType(SrcInsts[0]); 303 Type *DstTy = getLoadStoreType(SinkInsts[0]); 304 if (isa<ScalableVectorType>(SrcTy) || isa<ScalableVectorType>(DstTy)) { 305 CanUseDiffCheck = false; 306 return; 307 } 308 const DataLayout &DL = 309 SinkAR->getLoop()->getHeader()->getModule()->getDataLayout(); 310 unsigned AllocSize = 311 std::max(DL.getTypeAllocSize(SrcTy), DL.getTypeAllocSize(DstTy)); 312 313 // Only matching constant steps matching the AllocSize are supported at the 314 // moment. This simplifies the difference computation. Can be extended in the 315 // future. 316 auto *Step = dyn_cast<SCEVConstant>(SinkAR->getStepRecurrence(*SE)); 317 if (!Step || Step != SrcAR->getStepRecurrence(*SE) || 318 Step->getAPInt().abs() != AllocSize) { 319 CanUseDiffCheck = false; 320 return; 321 } 322 323 IntegerType *IntTy = 324 IntegerType::get(Src->PointerValue->getContext(), 325 DL.getPointerSizeInBits(CGI.AddressSpace)); 326 327 // When counting down, the dependence distance needs to be swapped. 328 if (Step->getValue()->isNegative()) 329 std::swap(SinkAR, SrcAR); 330 331 const SCEV *SinkStartInt = SE->getPtrToIntExpr(SinkAR->getStart(), IntTy); 332 const SCEV *SrcStartInt = SE->getPtrToIntExpr(SrcAR->getStart(), IntTy); 333 if (isa<SCEVCouldNotCompute>(SinkStartInt) || 334 isa<SCEVCouldNotCompute>(SrcStartInt)) { 335 CanUseDiffCheck = false; 336 return; 337 } 338 339 const Loop *InnerLoop = SrcAR->getLoop(); 340 // If the start values for both Src and Sink also vary according to an outer 341 // loop, then it's probably better to avoid creating diff checks because 342 // they may not be hoisted. We should instead let llvm::addRuntimeChecks 343 // do the expanded full range overlap checks, which can be hoisted. 344 if (HoistRuntimeChecks && InnerLoop->getParentLoop() && 345 isa<SCEVAddRecExpr>(SinkStartInt) && isa<SCEVAddRecExpr>(SrcStartInt)) { 346 auto *SrcStartAR = cast<SCEVAddRecExpr>(SrcStartInt); 347 auto *SinkStartAR = cast<SCEVAddRecExpr>(SinkStartInt); 348 const Loop *StartARLoop = SrcStartAR->getLoop(); 349 if (StartARLoop == SinkStartAR->getLoop() && 350 StartARLoop == InnerLoop->getParentLoop() && 351 // If the diff check would already be loop invariant (due to the 352 // recurrences being the same), then we prefer to keep the diff checks 353 // because they are cheaper. 354 SrcStartAR->getStepRecurrence(*SE) != 355 SinkStartAR->getStepRecurrence(*SE)) { 356 LLVM_DEBUG(dbgs() << "LAA: Not creating diff runtime check, since these " 357 "cannot be hoisted out of the outer loop\n"); 358 CanUseDiffCheck = false; 359 return; 360 } 361 } 362 363 LLVM_DEBUG(dbgs() << "LAA: Creating diff runtime check for:\n" 364 << "SrcStart: " << *SrcStartInt << '\n' 365 << "SinkStartInt: " << *SinkStartInt << '\n'); 366 DiffChecks.emplace_back(SrcStartInt, SinkStartInt, AllocSize, 367 Src->NeedsFreeze || Sink->NeedsFreeze); 368 } 369 370 SmallVector<RuntimePointerCheck, 4> RuntimePointerChecking::generateChecks() { 371 SmallVector<RuntimePointerCheck, 4> Checks; 372 373 for (unsigned I = 0; I < CheckingGroups.size(); ++I) { 374 for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) { 375 const RuntimeCheckingPtrGroup &CGI = CheckingGroups[I]; 376 const RuntimeCheckingPtrGroup &CGJ = CheckingGroups[J]; 377 378 if (needsChecking(CGI, CGJ)) { 379 tryToCreateDiffCheck(CGI, CGJ); 380 Checks.push_back(std::make_pair(&CGI, &CGJ)); 381 } 382 } 383 } 384 return Checks; 385 } 386 387 void RuntimePointerChecking::generateChecks( 388 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) { 389 assert(Checks.empty() && "Checks is not empty"); 390 groupChecks(DepCands, UseDependencies); 391 Checks = generateChecks(); 392 } 393 394 bool RuntimePointerChecking::needsChecking( 395 const RuntimeCheckingPtrGroup &M, const RuntimeCheckingPtrGroup &N) const { 396 for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I) 397 for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J) 398 if (needsChecking(M.Members[I], N.Members[J])) 399 return true; 400 return false; 401 } 402 403 /// Compare \p I and \p J and return the minimum. 404 /// Return nullptr in case we couldn't find an answer. 405 static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J, 406 ScalarEvolution *SE) { 407 const SCEV *Diff = SE->getMinusSCEV(J, I); 408 const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff); 409 410 if (!C) 411 return nullptr; 412 if (C->getValue()->isNegative()) 413 return J; 414 return I; 415 } 416 417 bool RuntimeCheckingPtrGroup::addPointer(unsigned Index, 418 RuntimePointerChecking &RtCheck) { 419 return addPointer( 420 Index, RtCheck.Pointers[Index].Start, RtCheck.Pointers[Index].End, 421 RtCheck.Pointers[Index].PointerValue->getType()->getPointerAddressSpace(), 422 RtCheck.Pointers[Index].NeedsFreeze, *RtCheck.SE); 423 } 424 425 bool RuntimeCheckingPtrGroup::addPointer(unsigned Index, const SCEV *Start, 426 const SCEV *End, unsigned AS, 427 bool NeedsFreeze, 428 ScalarEvolution &SE) { 429 assert(AddressSpace == AS && 430 "all pointers in a checking group must be in the same address space"); 431 432 // Compare the starts and ends with the known minimum and maximum 433 // of this set. We need to know how we compare against the min/max 434 // of the set in order to be able to emit memchecks. 435 const SCEV *Min0 = getMinFromExprs(Start, Low, &SE); 436 if (!Min0) 437 return false; 438 439 const SCEV *Min1 = getMinFromExprs(End, High, &SE); 440 if (!Min1) 441 return false; 442 443 // Update the low bound expression if we've found a new min value. 444 if (Min0 == Start) 445 Low = Start; 446 447 // Update the high bound expression if we've found a new max value. 448 if (Min1 != End) 449 High = End; 450 451 Members.push_back(Index); 452 this->NeedsFreeze |= NeedsFreeze; 453 return true; 454 } 455 456 void RuntimePointerChecking::groupChecks( 457 MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) { 458 // We build the groups from dependency candidates equivalence classes 459 // because: 460 // - We know that pointers in the same equivalence class share 461 // the same underlying object and therefore there is a chance 462 // that we can compare pointers 463 // - We wouldn't be able to merge two pointers for which we need 464 // to emit a memcheck. The classes in DepCands are already 465 // conveniently built such that no two pointers in the same 466 // class need checking against each other. 467 468 // We use the following (greedy) algorithm to construct the groups 469 // For every pointer in the equivalence class: 470 // For each existing group: 471 // - if the difference between this pointer and the min/max bounds 472 // of the group is a constant, then make the pointer part of the 473 // group and update the min/max bounds of that group as required. 474 475 CheckingGroups.clear(); 476 477 // If we need to check two pointers to the same underlying object 478 // with a non-constant difference, we shouldn't perform any pointer 479 // grouping with those pointers. This is because we can easily get 480 // into cases where the resulting check would return false, even when 481 // the accesses are safe. 482 // 483 // The following example shows this: 484 // for (i = 0; i < 1000; ++i) 485 // a[5000 + i * m] = a[i] + a[i + 9000] 486 // 487 // Here grouping gives a check of (5000, 5000 + 1000 * m) against 488 // (0, 10000) which is always false. However, if m is 1, there is no 489 // dependence. Not grouping the checks for a[i] and a[i + 9000] allows 490 // us to perform an accurate check in this case. 491 // 492 // The above case requires that we have an UnknownDependence between 493 // accesses to the same underlying object. This cannot happen unless 494 // FoundNonConstantDistanceDependence is set, and therefore UseDependencies 495 // is also false. In this case we will use the fallback path and create 496 // separate checking groups for all pointers. 497 498 // If we don't have the dependency partitions, construct a new 499 // checking pointer group for each pointer. This is also required 500 // for correctness, because in this case we can have checking between 501 // pointers to the same underlying object. 502 if (!UseDependencies) { 503 for (unsigned I = 0; I < Pointers.size(); ++I) 504 CheckingGroups.push_back(RuntimeCheckingPtrGroup(I, *this)); 505 return; 506 } 507 508 unsigned TotalComparisons = 0; 509 510 DenseMap<Value *, SmallVector<unsigned>> PositionMap; 511 for (unsigned Index = 0; Index < Pointers.size(); ++Index) { 512 auto Iter = PositionMap.insert({Pointers[Index].PointerValue, {}}); 513 Iter.first->second.push_back(Index); 514 } 515 516 // We need to keep track of what pointers we've already seen so we 517 // don't process them twice. 518 SmallSet<unsigned, 2> Seen; 519 520 // Go through all equivalence classes, get the "pointer check groups" 521 // and add them to the overall solution. We use the order in which accesses 522 // appear in 'Pointers' to enforce determinism. 523 for (unsigned I = 0; I < Pointers.size(); ++I) { 524 // We've seen this pointer before, and therefore already processed 525 // its equivalence class. 526 if (Seen.count(I)) 527 continue; 528 529 MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue, 530 Pointers[I].IsWritePtr); 531 532 SmallVector<RuntimeCheckingPtrGroup, 2> Groups; 533 auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access)); 534 535 // Because DepCands is constructed by visiting accesses in the order in 536 // which they appear in alias sets (which is deterministic) and the 537 // iteration order within an equivalence class member is only dependent on 538 // the order in which unions and insertions are performed on the 539 // equivalence class, the iteration order is deterministic. 540 for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end(); 541 MI != ME; ++MI) { 542 auto PointerI = PositionMap.find(MI->getPointer()); 543 assert(PointerI != PositionMap.end() && 544 "pointer in equivalence class not found in PositionMap"); 545 for (unsigned Pointer : PointerI->second) { 546 bool Merged = false; 547 // Mark this pointer as seen. 548 Seen.insert(Pointer); 549 550 // Go through all the existing sets and see if we can find one 551 // which can include this pointer. 552 for (RuntimeCheckingPtrGroup &Group : Groups) { 553 // Don't perform more than a certain amount of comparisons. 554 // This should limit the cost of grouping the pointers to something 555 // reasonable. If we do end up hitting this threshold, the algorithm 556 // will create separate groups for all remaining pointers. 557 if (TotalComparisons > MemoryCheckMergeThreshold) 558 break; 559 560 TotalComparisons++; 561 562 if (Group.addPointer(Pointer, *this)) { 563 Merged = true; 564 break; 565 } 566 } 567 568 if (!Merged) 569 // We couldn't add this pointer to any existing set or the threshold 570 // for the number of comparisons has been reached. Create a new group 571 // to hold the current pointer. 572 Groups.push_back(RuntimeCheckingPtrGroup(Pointer, *this)); 573 } 574 } 575 576 // We've computed the grouped checks for this partition. 577 // Save the results and continue with the next one. 578 llvm::copy(Groups, std::back_inserter(CheckingGroups)); 579 } 580 } 581 582 bool RuntimePointerChecking::arePointersInSamePartition( 583 const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1, 584 unsigned PtrIdx2) { 585 return (PtrToPartition[PtrIdx1] != -1 && 586 PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]); 587 } 588 589 bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const { 590 const PointerInfo &PointerI = Pointers[I]; 591 const PointerInfo &PointerJ = Pointers[J]; 592 593 // No need to check if two readonly pointers intersect. 594 if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr) 595 return false; 596 597 // Only need to check pointers between two different dependency sets. 598 if (PointerI.DependencySetId == PointerJ.DependencySetId) 599 return false; 600 601 // Only need to check pointers in the same alias set. 602 if (PointerI.AliasSetId != PointerJ.AliasSetId) 603 return false; 604 605 return true; 606 } 607 608 void RuntimePointerChecking::printChecks( 609 raw_ostream &OS, const SmallVectorImpl<RuntimePointerCheck> &Checks, 610 unsigned Depth) const { 611 unsigned N = 0; 612 for (const auto &Check : Checks) { 613 const auto &First = Check.first->Members, &Second = Check.second->Members; 614 615 OS.indent(Depth) << "Check " << N++ << ":\n"; 616 617 OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n"; 618 for (unsigned K = 0; K < First.size(); ++K) 619 OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n"; 620 621 OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n"; 622 for (unsigned K = 0; K < Second.size(); ++K) 623 OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n"; 624 } 625 } 626 627 void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const { 628 629 OS.indent(Depth) << "Run-time memory checks:\n"; 630 printChecks(OS, Checks, Depth); 631 632 OS.indent(Depth) << "Grouped accesses:\n"; 633 for (unsigned I = 0; I < CheckingGroups.size(); ++I) { 634 const auto &CG = CheckingGroups[I]; 635 636 OS.indent(Depth + 2) << "Group " << &CG << ":\n"; 637 OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High 638 << ")\n"; 639 for (unsigned J = 0; J < CG.Members.size(); ++J) { 640 OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr 641 << "\n"; 642 } 643 } 644 } 645 646 namespace { 647 648 /// Analyses memory accesses in a loop. 649 /// 650 /// Checks whether run time pointer checks are needed and builds sets for data 651 /// dependence checking. 652 class AccessAnalysis { 653 public: 654 /// Read or write access location. 655 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo; 656 typedef SmallVector<MemAccessInfo, 8> MemAccessInfoList; 657 658 AccessAnalysis(Loop *TheLoop, AAResults *AA, LoopInfo *LI, 659 MemoryDepChecker::DepCandidates &DA, 660 PredicatedScalarEvolution &PSE, 661 SmallPtrSetImpl<MDNode *> &LoopAliasScopes) 662 : TheLoop(TheLoop), BAA(*AA), AST(BAA), LI(LI), DepCands(DA), PSE(PSE), 663 LoopAliasScopes(LoopAliasScopes) { 664 // We're analyzing dependences across loop iterations. 665 BAA.enableCrossIterationMode(); 666 } 667 668 /// Register a load and whether it is only read from. 669 void addLoad(MemoryLocation &Loc, Type *AccessTy, bool IsReadOnly) { 670 Value *Ptr = const_cast<Value *>(Loc.Ptr); 671 AST.add(adjustLoc(Loc)); 672 Accesses[MemAccessInfo(Ptr, false)].insert(AccessTy); 673 if (IsReadOnly) 674 ReadOnlyPtr.insert(Ptr); 675 } 676 677 /// Register a store. 678 void addStore(MemoryLocation &Loc, Type *AccessTy) { 679 Value *Ptr = const_cast<Value *>(Loc.Ptr); 680 AST.add(adjustLoc(Loc)); 681 Accesses[MemAccessInfo(Ptr, true)].insert(AccessTy); 682 } 683 684 /// Check if we can emit a run-time no-alias check for \p Access. 685 /// 686 /// Returns true if we can emit a run-time no alias check for \p Access. 687 /// If we can check this access, this also adds it to a dependence set and 688 /// adds a run-time to check for it to \p RtCheck. If \p Assume is true, 689 /// we will attempt to use additional run-time checks in order to get 690 /// the bounds of the pointer. 691 bool createCheckForAccess(RuntimePointerChecking &RtCheck, 692 MemAccessInfo Access, Type *AccessTy, 693 const DenseMap<Value *, const SCEV *> &Strides, 694 DenseMap<Value *, unsigned> &DepSetId, 695 Loop *TheLoop, unsigned &RunningDepId, 696 unsigned ASId, bool ShouldCheckStride, bool Assume); 697 698 /// Check whether we can check the pointers at runtime for 699 /// non-intersection. 700 /// 701 /// Returns true if we need no check or if we do and we can generate them 702 /// (i.e. the pointers have computable bounds). 703 bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE, 704 Loop *TheLoop, const DenseMap<Value *, const SCEV *> &Strides, 705 Value *&UncomputablePtr, bool ShouldCheckWrap = false); 706 707 /// Goes over all memory accesses, checks whether a RT check is needed 708 /// and builds sets of dependent accesses. 709 void buildDependenceSets() { 710 processMemAccesses(); 711 } 712 713 /// Initial processing of memory accesses determined that we need to 714 /// perform dependency checking. 715 /// 716 /// Note that this can later be cleared if we retry memcheck analysis without 717 /// dependency checking (i.e. FoundNonConstantDistanceDependence). 718 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); } 719 720 /// We decided that no dependence analysis would be used. Reset the state. 721 void resetDepChecks(MemoryDepChecker &DepChecker) { 722 CheckDeps.clear(); 723 DepChecker.clearDependences(); 724 } 725 726 MemAccessInfoList &getDependenciesToCheck() { return CheckDeps; } 727 728 const DenseMap<Value *, SmallVector<const Value *, 16>> & 729 getUnderlyingObjects() { 730 return UnderlyingObjects; 731 } 732 733 private: 734 typedef MapVector<MemAccessInfo, SmallSetVector<Type *, 1>> PtrAccessMap; 735 736 /// Adjust the MemoryLocation so that it represents accesses to this 737 /// location across all iterations, rather than a single one. 738 MemoryLocation adjustLoc(MemoryLocation Loc) const { 739 // The accessed location varies within the loop, but remains within the 740 // underlying object. 741 Loc.Size = LocationSize::beforeOrAfterPointer(); 742 Loc.AATags.Scope = adjustAliasScopeList(Loc.AATags.Scope); 743 Loc.AATags.NoAlias = adjustAliasScopeList(Loc.AATags.NoAlias); 744 return Loc; 745 } 746 747 /// Drop alias scopes that are only valid within a single loop iteration. 748 MDNode *adjustAliasScopeList(MDNode *ScopeList) const { 749 if (!ScopeList) 750 return nullptr; 751 752 // For the sake of simplicity, drop the whole scope list if any scope is 753 // iteration-local. 754 if (any_of(ScopeList->operands(), [&](Metadata *Scope) { 755 return LoopAliasScopes.contains(cast<MDNode>(Scope)); 756 })) 757 return nullptr; 758 759 return ScopeList; 760 } 761 762 /// Go over all memory access and check whether runtime pointer checks 763 /// are needed and build sets of dependency check candidates. 764 void processMemAccesses(); 765 766 /// Map of all accesses. Values are the types used to access memory pointed to 767 /// by the pointer. 768 PtrAccessMap Accesses; 769 770 /// The loop being checked. 771 const Loop *TheLoop; 772 773 /// List of accesses that need a further dependence check. 774 MemAccessInfoList CheckDeps; 775 776 /// Set of pointers that are read only. 777 SmallPtrSet<Value*, 16> ReadOnlyPtr; 778 779 /// Batched alias analysis results. 780 BatchAAResults BAA; 781 782 /// An alias set tracker to partition the access set by underlying object and 783 //intrinsic property (such as TBAA metadata). 784 AliasSetTracker AST; 785 786 LoopInfo *LI; 787 788 /// Sets of potentially dependent accesses - members of one set share an 789 /// underlying pointer. The set "CheckDeps" identfies which sets really need a 790 /// dependence check. 791 MemoryDepChecker::DepCandidates &DepCands; 792 793 /// Initial processing of memory accesses determined that we may need 794 /// to add memchecks. Perform the analysis to determine the necessary checks. 795 /// 796 /// Note that, this is different from isDependencyCheckNeeded. When we retry 797 /// memcheck analysis without dependency checking 798 /// (i.e. FoundNonConstantDistanceDependence), isDependencyCheckNeeded is 799 /// cleared while this remains set if we have potentially dependent accesses. 800 bool IsRTCheckAnalysisNeeded = false; 801 802 /// The SCEV predicate containing all the SCEV-related assumptions. 803 PredicatedScalarEvolution &PSE; 804 805 DenseMap<Value *, SmallVector<const Value *, 16>> UnderlyingObjects; 806 807 /// Alias scopes that are declared inside the loop, and as such not valid 808 /// across iterations. 809 SmallPtrSetImpl<MDNode *> &LoopAliasScopes; 810 }; 811 812 } // end anonymous namespace 813 814 /// Check whether a pointer can participate in a runtime bounds check. 815 /// If \p Assume, try harder to prove that we can compute the bounds of \p Ptr 816 /// by adding run-time checks (overflow checks) if necessary. 817 static bool hasComputableBounds(PredicatedScalarEvolution &PSE, Value *Ptr, 818 const SCEV *PtrScev, Loop *L, bool Assume) { 819 // The bounds for loop-invariant pointer is trivial. 820 if (PSE.getSE()->isLoopInvariant(PtrScev, L)) 821 return true; 822 823 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); 824 825 if (!AR && Assume) 826 AR = PSE.getAsAddRec(Ptr); 827 828 if (!AR) 829 return false; 830 831 return AR->isAffine(); 832 } 833 834 /// Check whether a pointer address cannot wrap. 835 static bool isNoWrap(PredicatedScalarEvolution &PSE, 836 const DenseMap<Value *, const SCEV *> &Strides, Value *Ptr, Type *AccessTy, 837 Loop *L) { 838 const SCEV *PtrScev = PSE.getSCEV(Ptr); 839 if (PSE.getSE()->isLoopInvariant(PtrScev, L)) 840 return true; 841 842 int64_t Stride = getPtrStride(PSE, AccessTy, Ptr, L, Strides).value_or(0); 843 if (Stride == 1 || PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW)) 844 return true; 845 846 return false; 847 } 848 849 static void visitPointers(Value *StartPtr, const Loop &InnermostLoop, 850 function_ref<void(Value *)> AddPointer) { 851 SmallPtrSet<Value *, 8> Visited; 852 SmallVector<Value *> WorkList; 853 WorkList.push_back(StartPtr); 854 855 while (!WorkList.empty()) { 856 Value *Ptr = WorkList.pop_back_val(); 857 if (!Visited.insert(Ptr).second) 858 continue; 859 auto *PN = dyn_cast<PHINode>(Ptr); 860 // SCEV does not look through non-header PHIs inside the loop. Such phis 861 // can be analyzed by adding separate accesses for each incoming pointer 862 // value. 863 if (PN && InnermostLoop.contains(PN->getParent()) && 864 PN->getParent() != InnermostLoop.getHeader()) { 865 for (const Use &Inc : PN->incoming_values()) 866 WorkList.push_back(Inc); 867 } else 868 AddPointer(Ptr); 869 } 870 } 871 872 // Walk back through the IR for a pointer, looking for a select like the 873 // following: 874 // 875 // %offset = select i1 %cmp, i64 %a, i64 %b 876 // %addr = getelementptr double, double* %base, i64 %offset 877 // %ld = load double, double* %addr, align 8 878 // 879 // We won't be able to form a single SCEVAddRecExpr from this since the 880 // address for each loop iteration depends on %cmp. We could potentially 881 // produce multiple valid SCEVAddRecExprs, though, and check all of them for 882 // memory safety/aliasing if needed. 883 // 884 // If we encounter some IR we don't yet handle, or something obviously fine 885 // like a constant, then we just add the SCEV for that term to the list passed 886 // in by the caller. If we have a node that may potentially yield a valid 887 // SCEVAddRecExpr then we decompose it into parts and build the SCEV terms 888 // ourselves before adding to the list. 889 static void findForkedSCEVs( 890 ScalarEvolution *SE, const Loop *L, Value *Ptr, 891 SmallVectorImpl<PointerIntPair<const SCEV *, 1, bool>> &ScevList, 892 unsigned Depth) { 893 // If our Value is a SCEVAddRecExpr, loop invariant, not an instruction, or 894 // we've exceeded our limit on recursion, just return whatever we have 895 // regardless of whether it can be used for a forked pointer or not, along 896 // with an indication of whether it might be a poison or undef value. 897 const SCEV *Scev = SE->getSCEV(Ptr); 898 if (isa<SCEVAddRecExpr>(Scev) || L->isLoopInvariant(Ptr) || 899 !isa<Instruction>(Ptr) || Depth == 0) { 900 ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(Ptr)); 901 return; 902 } 903 904 Depth--; 905 906 auto UndefPoisonCheck = [](PointerIntPair<const SCEV *, 1, bool> S) { 907 return get<1>(S); 908 }; 909 910 auto GetBinOpExpr = [&SE](unsigned Opcode, const SCEV *L, const SCEV *R) { 911 switch (Opcode) { 912 case Instruction::Add: 913 return SE->getAddExpr(L, R); 914 case Instruction::Sub: 915 return SE->getMinusSCEV(L, R); 916 default: 917 llvm_unreachable("Unexpected binary operator when walking ForkedPtrs"); 918 } 919 }; 920 921 Instruction *I = cast<Instruction>(Ptr); 922 unsigned Opcode = I->getOpcode(); 923 switch (Opcode) { 924 case Instruction::GetElementPtr: { 925 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I); 926 Type *SourceTy = GEP->getSourceElementType(); 927 // We only handle base + single offset GEPs here for now. 928 // Not dealing with preexisting gathers yet, so no vectors. 929 if (I->getNumOperands() != 2 || SourceTy->isVectorTy()) { 930 ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(GEP)); 931 break; 932 } 933 SmallVector<PointerIntPair<const SCEV *, 1, bool>, 2> BaseScevs; 934 SmallVector<PointerIntPair<const SCEV *, 1, bool>, 2> OffsetScevs; 935 findForkedSCEVs(SE, L, I->getOperand(0), BaseScevs, Depth); 936 findForkedSCEVs(SE, L, I->getOperand(1), OffsetScevs, Depth); 937 938 // See if we need to freeze our fork... 939 bool NeedsFreeze = any_of(BaseScevs, UndefPoisonCheck) || 940 any_of(OffsetScevs, UndefPoisonCheck); 941 942 // Check that we only have a single fork, on either the base or the offset. 943 // Copy the SCEV across for the one without a fork in order to generate 944 // the full SCEV for both sides of the GEP. 945 if (OffsetScevs.size() == 2 && BaseScevs.size() == 1) 946 BaseScevs.push_back(BaseScevs[0]); 947 else if (BaseScevs.size() == 2 && OffsetScevs.size() == 1) 948 OffsetScevs.push_back(OffsetScevs[0]); 949 else { 950 ScevList.emplace_back(Scev, NeedsFreeze); 951 break; 952 } 953 954 // Find the pointer type we need to extend to. 955 Type *IntPtrTy = SE->getEffectiveSCEVType( 956 SE->getSCEV(GEP->getPointerOperand())->getType()); 957 958 // Find the size of the type being pointed to. We only have a single 959 // index term (guarded above) so we don't need to index into arrays or 960 // structures, just get the size of the scalar value. 961 const SCEV *Size = SE->getSizeOfExpr(IntPtrTy, SourceTy); 962 963 // Scale up the offsets by the size of the type, then add to the bases. 964 const SCEV *Scaled1 = SE->getMulExpr( 965 Size, SE->getTruncateOrSignExtend(get<0>(OffsetScevs[0]), IntPtrTy)); 966 const SCEV *Scaled2 = SE->getMulExpr( 967 Size, SE->getTruncateOrSignExtend(get<0>(OffsetScevs[1]), IntPtrTy)); 968 ScevList.emplace_back(SE->getAddExpr(get<0>(BaseScevs[0]), Scaled1), 969 NeedsFreeze); 970 ScevList.emplace_back(SE->getAddExpr(get<0>(BaseScevs[1]), Scaled2), 971 NeedsFreeze); 972 break; 973 } 974 case Instruction::Select: { 975 SmallVector<PointerIntPair<const SCEV *, 1, bool>, 2> ChildScevs; 976 // A select means we've found a forked pointer, but we currently only 977 // support a single select per pointer so if there's another behind this 978 // then we just bail out and return the generic SCEV. 979 findForkedSCEVs(SE, L, I->getOperand(1), ChildScevs, Depth); 980 findForkedSCEVs(SE, L, I->getOperand(2), ChildScevs, Depth); 981 if (ChildScevs.size() == 2) { 982 ScevList.push_back(ChildScevs[0]); 983 ScevList.push_back(ChildScevs[1]); 984 } else 985 ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(Ptr)); 986 break; 987 } 988 case Instruction::PHI: { 989 SmallVector<PointerIntPair<const SCEV *, 1, bool>, 2> ChildScevs; 990 // A phi means we've found a forked pointer, but we currently only 991 // support a single phi per pointer so if there's another behind this 992 // then we just bail out and return the generic SCEV. 993 if (I->getNumOperands() == 2) { 994 findForkedSCEVs(SE, L, I->getOperand(0), ChildScevs, Depth); 995 findForkedSCEVs(SE, L, I->getOperand(1), ChildScevs, Depth); 996 } 997 if (ChildScevs.size() == 2) { 998 ScevList.push_back(ChildScevs[0]); 999 ScevList.push_back(ChildScevs[1]); 1000 } else 1001 ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(Ptr)); 1002 break; 1003 } 1004 case Instruction::Add: 1005 case Instruction::Sub: { 1006 SmallVector<PointerIntPair<const SCEV *, 1, bool>> LScevs; 1007 SmallVector<PointerIntPair<const SCEV *, 1, bool>> RScevs; 1008 findForkedSCEVs(SE, L, I->getOperand(0), LScevs, Depth); 1009 findForkedSCEVs(SE, L, I->getOperand(1), RScevs, Depth); 1010 1011 // See if we need to freeze our fork... 1012 bool NeedsFreeze = 1013 any_of(LScevs, UndefPoisonCheck) || any_of(RScevs, UndefPoisonCheck); 1014 1015 // Check that we only have a single fork, on either the left or right side. 1016 // Copy the SCEV across for the one without a fork in order to generate 1017 // the full SCEV for both sides of the BinOp. 1018 if (LScevs.size() == 2 && RScevs.size() == 1) 1019 RScevs.push_back(RScevs[0]); 1020 else if (RScevs.size() == 2 && LScevs.size() == 1) 1021 LScevs.push_back(LScevs[0]); 1022 else { 1023 ScevList.emplace_back(Scev, NeedsFreeze); 1024 break; 1025 } 1026 1027 ScevList.emplace_back( 1028 GetBinOpExpr(Opcode, get<0>(LScevs[0]), get<0>(RScevs[0])), 1029 NeedsFreeze); 1030 ScevList.emplace_back( 1031 GetBinOpExpr(Opcode, get<0>(LScevs[1]), get<0>(RScevs[1])), 1032 NeedsFreeze); 1033 break; 1034 } 1035 default: 1036 // Just return the current SCEV if we haven't handled the instruction yet. 1037 LLVM_DEBUG(dbgs() << "ForkedPtr unhandled instruction: " << *I << "\n"); 1038 ScevList.emplace_back(Scev, !isGuaranteedNotToBeUndefOrPoison(Ptr)); 1039 break; 1040 } 1041 } 1042 1043 static SmallVector<PointerIntPair<const SCEV *, 1, bool>> 1044 findForkedPointer(PredicatedScalarEvolution &PSE, 1045 const DenseMap<Value *, const SCEV *> &StridesMap, Value *Ptr, 1046 const Loop *L) { 1047 ScalarEvolution *SE = PSE.getSE(); 1048 assert(SE->isSCEVable(Ptr->getType()) && "Value is not SCEVable!"); 1049 SmallVector<PointerIntPair<const SCEV *, 1, bool>> Scevs; 1050 findForkedSCEVs(SE, L, Ptr, Scevs, MaxForkedSCEVDepth); 1051 1052 // For now, we will only accept a forked pointer with two possible SCEVs 1053 // that are either SCEVAddRecExprs or loop invariant. 1054 if (Scevs.size() == 2 && 1055 (isa<SCEVAddRecExpr>(get<0>(Scevs[0])) || 1056 SE->isLoopInvariant(get<0>(Scevs[0]), L)) && 1057 (isa<SCEVAddRecExpr>(get<0>(Scevs[1])) || 1058 SE->isLoopInvariant(get<0>(Scevs[1]), L))) { 1059 LLVM_DEBUG(dbgs() << "LAA: Found forked pointer: " << *Ptr << "\n"); 1060 LLVM_DEBUG(dbgs() << "\t(1) " << *get<0>(Scevs[0]) << "\n"); 1061 LLVM_DEBUG(dbgs() << "\t(2) " << *get<0>(Scevs[1]) << "\n"); 1062 return Scevs; 1063 } 1064 1065 return {{replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr), false}}; 1066 } 1067 1068 bool AccessAnalysis::createCheckForAccess(RuntimePointerChecking &RtCheck, 1069 MemAccessInfo Access, Type *AccessTy, 1070 const DenseMap<Value *, const SCEV *> &StridesMap, 1071 DenseMap<Value *, unsigned> &DepSetId, 1072 Loop *TheLoop, unsigned &RunningDepId, 1073 unsigned ASId, bool ShouldCheckWrap, 1074 bool Assume) { 1075 Value *Ptr = Access.getPointer(); 1076 1077 SmallVector<PointerIntPair<const SCEV *, 1, bool>> TranslatedPtrs = 1078 findForkedPointer(PSE, StridesMap, Ptr, TheLoop); 1079 1080 for (auto &P : TranslatedPtrs) { 1081 const SCEV *PtrExpr = get<0>(P); 1082 if (!hasComputableBounds(PSE, Ptr, PtrExpr, TheLoop, Assume)) 1083 return false; 1084 1085 // When we run after a failing dependency check we have to make sure 1086 // we don't have wrapping pointers. 1087 if (ShouldCheckWrap) { 1088 // Skip wrap checking when translating pointers. 1089 if (TranslatedPtrs.size() > 1) 1090 return false; 1091 1092 if (!isNoWrap(PSE, StridesMap, Ptr, AccessTy, TheLoop)) { 1093 auto *Expr = PSE.getSCEV(Ptr); 1094 if (!Assume || !isa<SCEVAddRecExpr>(Expr)) 1095 return false; 1096 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); 1097 } 1098 } 1099 // If there's only one option for Ptr, look it up after bounds and wrap 1100 // checking, because assumptions might have been added to PSE. 1101 if (TranslatedPtrs.size() == 1) 1102 TranslatedPtrs[0] = {replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr), 1103 false}; 1104 } 1105 1106 for (auto [PtrExpr, NeedsFreeze] : TranslatedPtrs) { 1107 // The id of the dependence set. 1108 unsigned DepId; 1109 1110 if (isDependencyCheckNeeded()) { 1111 Value *Leader = DepCands.getLeaderValue(Access).getPointer(); 1112 unsigned &LeaderId = DepSetId[Leader]; 1113 if (!LeaderId) 1114 LeaderId = RunningDepId++; 1115 DepId = LeaderId; 1116 } else 1117 // Each access has its own dependence set. 1118 DepId = RunningDepId++; 1119 1120 bool IsWrite = Access.getInt(); 1121 RtCheck.insert(TheLoop, Ptr, PtrExpr, AccessTy, IsWrite, DepId, ASId, PSE, 1122 NeedsFreeze); 1123 LLVM_DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n'); 1124 } 1125 1126 return true; 1127 } 1128 1129 bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck, 1130 ScalarEvolution *SE, Loop *TheLoop, 1131 const DenseMap<Value *, const SCEV *> &StridesMap, 1132 Value *&UncomputablePtr, bool ShouldCheckWrap) { 1133 // Find pointers with computable bounds. We are going to use this information 1134 // to place a runtime bound check. 1135 bool CanDoRT = true; 1136 1137 bool MayNeedRTCheck = false; 1138 if (!IsRTCheckAnalysisNeeded) return true; 1139 1140 bool IsDepCheckNeeded = isDependencyCheckNeeded(); 1141 1142 // We assign a consecutive id to access from different alias sets. 1143 // Accesses between different groups doesn't need to be checked. 1144 unsigned ASId = 0; 1145 for (auto &AS : AST) { 1146 int NumReadPtrChecks = 0; 1147 int NumWritePtrChecks = 0; 1148 bool CanDoAliasSetRT = true; 1149 ++ASId; 1150 auto ASPointers = AS.getPointers(); 1151 1152 // We assign consecutive id to access from different dependence sets. 1153 // Accesses within the same set don't need a runtime check. 1154 unsigned RunningDepId = 1; 1155 DenseMap<Value *, unsigned> DepSetId; 1156 1157 SmallVector<std::pair<MemAccessInfo, Type *>, 4> Retries; 1158 1159 // First, count how many write and read accesses are in the alias set. Also 1160 // collect MemAccessInfos for later. 1161 SmallVector<MemAccessInfo, 4> AccessInfos; 1162 for (const Value *Ptr_ : ASPointers) { 1163 Value *Ptr = const_cast<Value *>(Ptr_); 1164 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true)); 1165 if (IsWrite) 1166 ++NumWritePtrChecks; 1167 else 1168 ++NumReadPtrChecks; 1169 AccessInfos.emplace_back(Ptr, IsWrite); 1170 } 1171 1172 // We do not need runtime checks for this alias set, if there are no writes 1173 // or a single write and no reads. 1174 if (NumWritePtrChecks == 0 || 1175 (NumWritePtrChecks == 1 && NumReadPtrChecks == 0)) { 1176 assert((ASPointers.size() <= 1 || 1177 all_of(ASPointers, 1178 [this](const Value *Ptr) { 1179 MemAccessInfo AccessWrite(const_cast<Value *>(Ptr), 1180 true); 1181 return DepCands.findValue(AccessWrite) == DepCands.end(); 1182 })) && 1183 "Can only skip updating CanDoRT below, if all entries in AS " 1184 "are reads or there is at most 1 entry"); 1185 continue; 1186 } 1187 1188 for (auto &Access : AccessInfos) { 1189 for (const auto &AccessTy : Accesses[Access]) { 1190 if (!createCheckForAccess(RtCheck, Access, AccessTy, StridesMap, 1191 DepSetId, TheLoop, RunningDepId, ASId, 1192 ShouldCheckWrap, false)) { 1193 LLVM_DEBUG(dbgs() << "LAA: Can't find bounds for ptr:" 1194 << *Access.getPointer() << '\n'); 1195 Retries.push_back({Access, AccessTy}); 1196 CanDoAliasSetRT = false; 1197 } 1198 } 1199 } 1200 1201 // Note that this function computes CanDoRT and MayNeedRTCheck 1202 // independently. For example CanDoRT=false, MayNeedRTCheck=false means that 1203 // we have a pointer for which we couldn't find the bounds but we don't 1204 // actually need to emit any checks so it does not matter. 1205 // 1206 // We need runtime checks for this alias set, if there are at least 2 1207 // dependence sets (in which case RunningDepId > 2) or if we need to re-try 1208 // any bound checks (because in that case the number of dependence sets is 1209 // incomplete). 1210 bool NeedsAliasSetRTCheck = RunningDepId > 2 || !Retries.empty(); 1211 1212 // We need to perform run-time alias checks, but some pointers had bounds 1213 // that couldn't be checked. 1214 if (NeedsAliasSetRTCheck && !CanDoAliasSetRT) { 1215 // Reset the CanDoSetRt flag and retry all accesses that have failed. 1216 // We know that we need these checks, so we can now be more aggressive 1217 // and add further checks if required (overflow checks). 1218 CanDoAliasSetRT = true; 1219 for (auto Retry : Retries) { 1220 MemAccessInfo Access = Retry.first; 1221 Type *AccessTy = Retry.second; 1222 if (!createCheckForAccess(RtCheck, Access, AccessTy, StridesMap, 1223 DepSetId, TheLoop, RunningDepId, ASId, 1224 ShouldCheckWrap, /*Assume=*/true)) { 1225 CanDoAliasSetRT = false; 1226 UncomputablePtr = Access.getPointer(); 1227 break; 1228 } 1229 } 1230 } 1231 1232 CanDoRT &= CanDoAliasSetRT; 1233 MayNeedRTCheck |= NeedsAliasSetRTCheck; 1234 ++ASId; 1235 } 1236 1237 // If the pointers that we would use for the bounds comparison have different 1238 // address spaces, assume the values aren't directly comparable, so we can't 1239 // use them for the runtime check. We also have to assume they could 1240 // overlap. In the future there should be metadata for whether address spaces 1241 // are disjoint. 1242 unsigned NumPointers = RtCheck.Pointers.size(); 1243 for (unsigned i = 0; i < NumPointers; ++i) { 1244 for (unsigned j = i + 1; j < NumPointers; ++j) { 1245 // Only need to check pointers between two different dependency sets. 1246 if (RtCheck.Pointers[i].DependencySetId == 1247 RtCheck.Pointers[j].DependencySetId) 1248 continue; 1249 // Only need to check pointers in the same alias set. 1250 if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId) 1251 continue; 1252 1253 Value *PtrI = RtCheck.Pointers[i].PointerValue; 1254 Value *PtrJ = RtCheck.Pointers[j].PointerValue; 1255 1256 unsigned ASi = PtrI->getType()->getPointerAddressSpace(); 1257 unsigned ASj = PtrJ->getType()->getPointerAddressSpace(); 1258 if (ASi != ASj) { 1259 LLVM_DEBUG( 1260 dbgs() << "LAA: Runtime check would require comparison between" 1261 " different address spaces\n"); 1262 return false; 1263 } 1264 } 1265 } 1266 1267 if (MayNeedRTCheck && CanDoRT) 1268 RtCheck.generateChecks(DepCands, IsDepCheckNeeded); 1269 1270 LLVM_DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks() 1271 << " pointer comparisons.\n"); 1272 1273 // If we can do run-time checks, but there are no checks, no runtime checks 1274 // are needed. This can happen when all pointers point to the same underlying 1275 // object for example. 1276 RtCheck.Need = CanDoRT ? RtCheck.getNumberOfChecks() != 0 : MayNeedRTCheck; 1277 1278 bool CanDoRTIfNeeded = !RtCheck.Need || CanDoRT; 1279 if (!CanDoRTIfNeeded) 1280 RtCheck.reset(); 1281 return CanDoRTIfNeeded; 1282 } 1283 1284 void AccessAnalysis::processMemAccesses() { 1285 // We process the set twice: first we process read-write pointers, last we 1286 // process read-only pointers. This allows us to skip dependence tests for 1287 // read-only pointers. 1288 1289 LLVM_DEBUG(dbgs() << "LAA: Processing memory accesses...\n"); 1290 LLVM_DEBUG(dbgs() << " AST: "; AST.dump()); 1291 LLVM_DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n"); 1292 LLVM_DEBUG({ 1293 for (auto A : Accesses) 1294 dbgs() << "\t" << *A.first.getPointer() << " (" 1295 << (A.first.getInt() 1296 ? "write" 1297 : (ReadOnlyPtr.count(A.first.getPointer()) ? "read-only" 1298 : "read")) 1299 << ")\n"; 1300 }); 1301 1302 // The AliasSetTracker has nicely partitioned our pointers by metadata 1303 // compatibility and potential for underlying-object overlap. As a result, we 1304 // only need to check for potential pointer dependencies within each alias 1305 // set. 1306 for (const auto &AS : AST) { 1307 // Note that both the alias-set tracker and the alias sets themselves used 1308 // ordered collections internally and so the iteration order here is 1309 // deterministic. 1310 auto ASPointers = AS.getPointers(); 1311 1312 bool SetHasWrite = false; 1313 1314 // Map of pointers to last access encountered. 1315 typedef DenseMap<const Value*, MemAccessInfo> UnderlyingObjToAccessMap; 1316 UnderlyingObjToAccessMap ObjToLastAccess; 1317 1318 // Set of access to check after all writes have been processed. 1319 PtrAccessMap DeferredAccesses; 1320 1321 // Iterate over each alias set twice, once to process read/write pointers, 1322 // and then to process read-only pointers. 1323 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) { 1324 bool UseDeferred = SetIteration > 0; 1325 PtrAccessMap &S = UseDeferred ? DeferredAccesses : Accesses; 1326 1327 for (const Value *Ptr_ : ASPointers) { 1328 Value *Ptr = const_cast<Value *>(Ptr_); 1329 1330 // For a single memory access in AliasSetTracker, Accesses may contain 1331 // both read and write, and they both need to be handled for CheckDeps. 1332 for (const auto &AC : S) { 1333 if (AC.first.getPointer() != Ptr) 1334 continue; 1335 1336 bool IsWrite = AC.first.getInt(); 1337 1338 // If we're using the deferred access set, then it contains only 1339 // reads. 1340 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite; 1341 if (UseDeferred && !IsReadOnlyPtr) 1342 continue; 1343 // Otherwise, the pointer must be in the PtrAccessSet, either as a 1344 // read or a write. 1345 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite || 1346 S.count(MemAccessInfo(Ptr, false))) && 1347 "Alias-set pointer not in the access set?"); 1348 1349 MemAccessInfo Access(Ptr, IsWrite); 1350 DepCands.insert(Access); 1351 1352 // Memorize read-only pointers for later processing and skip them in 1353 // the first round (they need to be checked after we have seen all 1354 // write pointers). Note: we also mark pointer that are not 1355 // consecutive as "read-only" pointers (so that we check 1356 // "a[b[i]] +="). Hence, we need the second check for "!IsWrite". 1357 if (!UseDeferred && IsReadOnlyPtr) { 1358 // We only use the pointer keys, the types vector values don't 1359 // matter. 1360 DeferredAccesses.insert({Access, {}}); 1361 continue; 1362 } 1363 1364 // If this is a write - check other reads and writes for conflicts. If 1365 // this is a read only check other writes for conflicts (but only if 1366 // there is no other write to the ptr - this is an optimization to 1367 // catch "a[i] = a[i] + " without having to do a dependence check). 1368 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) { 1369 CheckDeps.push_back(Access); 1370 IsRTCheckAnalysisNeeded = true; 1371 } 1372 1373 if (IsWrite) 1374 SetHasWrite = true; 1375 1376 // Create sets of pointers connected by a shared alias set and 1377 // underlying object. 1378 typedef SmallVector<const Value *, 16> ValueVector; 1379 ValueVector TempObjects; 1380 1381 UnderlyingObjects[Ptr] = {}; 1382 SmallVector<const Value *, 16> &UOs = UnderlyingObjects[Ptr]; 1383 ::getUnderlyingObjects(Ptr, UOs, LI); 1384 LLVM_DEBUG(dbgs() 1385 << "Underlying objects for pointer " << *Ptr << "\n"); 1386 for (const Value *UnderlyingObj : UOs) { 1387 // nullptr never alias, don't join sets for pointer that have "null" 1388 // in their UnderlyingObjects list. 1389 if (isa<ConstantPointerNull>(UnderlyingObj) && 1390 !NullPointerIsDefined( 1391 TheLoop->getHeader()->getParent(), 1392 UnderlyingObj->getType()->getPointerAddressSpace())) 1393 continue; 1394 1395 UnderlyingObjToAccessMap::iterator Prev = 1396 ObjToLastAccess.find(UnderlyingObj); 1397 if (Prev != ObjToLastAccess.end()) 1398 DepCands.unionSets(Access, Prev->second); 1399 1400 ObjToLastAccess[UnderlyingObj] = Access; 1401 LLVM_DEBUG(dbgs() << " " << *UnderlyingObj << "\n"); 1402 } 1403 } 1404 } 1405 } 1406 } 1407 } 1408 1409 /// Return true if an AddRec pointer \p Ptr is unsigned non-wrapping, 1410 /// i.e. monotonically increasing/decreasing. 1411 static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR, 1412 PredicatedScalarEvolution &PSE, const Loop *L) { 1413 1414 // FIXME: This should probably only return true for NUW. 1415 if (AR->getNoWrapFlags(SCEV::NoWrapMask)) 1416 return true; 1417 1418 if (PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW)) 1419 return true; 1420 1421 // Scalar evolution does not propagate the non-wrapping flags to values that 1422 // are derived from a non-wrapping induction variable because non-wrapping 1423 // could be flow-sensitive. 1424 // 1425 // Look through the potentially overflowing instruction to try to prove 1426 // non-wrapping for the *specific* value of Ptr. 1427 1428 // The arithmetic implied by an inbounds GEP can't overflow. 1429 auto *GEP = dyn_cast<GetElementPtrInst>(Ptr); 1430 if (!GEP || !GEP->isInBounds()) 1431 return false; 1432 1433 // Make sure there is only one non-const index and analyze that. 1434 Value *NonConstIndex = nullptr; 1435 for (Value *Index : GEP->indices()) 1436 if (!isa<ConstantInt>(Index)) { 1437 if (NonConstIndex) 1438 return false; 1439 NonConstIndex = Index; 1440 } 1441 if (!NonConstIndex) 1442 // The recurrence is on the pointer, ignore for now. 1443 return false; 1444 1445 // The index in GEP is signed. It is non-wrapping if it's derived from a NSW 1446 // AddRec using a NSW operation. 1447 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex)) 1448 if (OBO->hasNoSignedWrap() && 1449 // Assume constant for other the operand so that the AddRec can be 1450 // easily found. 1451 isa<ConstantInt>(OBO->getOperand(1))) { 1452 auto *OpScev = PSE.getSCEV(OBO->getOperand(0)); 1453 1454 if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev)) 1455 return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW); 1456 } 1457 1458 return false; 1459 } 1460 1461 /// Check whether the access through \p Ptr has a constant stride. 1462 std::optional<int64_t> llvm::getPtrStride(PredicatedScalarEvolution &PSE, 1463 Type *AccessTy, Value *Ptr, 1464 const Loop *Lp, 1465 const DenseMap<Value *, const SCEV *> &StridesMap, 1466 bool Assume, bool ShouldCheckWrap) { 1467 Type *Ty = Ptr->getType(); 1468 assert(Ty->isPointerTy() && "Unexpected non-ptr"); 1469 1470 if (isa<ScalableVectorType>(AccessTy)) { 1471 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Scalable object: " << *AccessTy 1472 << "\n"); 1473 return std::nullopt; 1474 } 1475 1476 const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr); 1477 1478 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); 1479 if (Assume && !AR) 1480 AR = PSE.getAsAddRec(Ptr); 1481 1482 if (!AR) { 1483 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr 1484 << " SCEV: " << *PtrScev << "\n"); 1485 return std::nullopt; 1486 } 1487 1488 // The access function must stride over the innermost loop. 1489 if (Lp != AR->getLoop()) { 1490 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop " 1491 << *Ptr << " SCEV: " << *AR << "\n"); 1492 return std::nullopt; 1493 } 1494 1495 // Check the step is constant. 1496 const SCEV *Step = AR->getStepRecurrence(*PSE.getSE()); 1497 1498 // Calculate the pointer stride and check if it is constant. 1499 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); 1500 if (!C) { 1501 LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr 1502 << " SCEV: " << *AR << "\n"); 1503 return std::nullopt; 1504 } 1505 1506 auto &DL = Lp->getHeader()->getModule()->getDataLayout(); 1507 TypeSize AllocSize = DL.getTypeAllocSize(AccessTy); 1508 int64_t Size = AllocSize.getFixedValue(); 1509 const APInt &APStepVal = C->getAPInt(); 1510 1511 // Huge step value - give up. 1512 if (APStepVal.getBitWidth() > 64) 1513 return std::nullopt; 1514 1515 int64_t StepVal = APStepVal.getSExtValue(); 1516 1517 // Strided access. 1518 int64_t Stride = StepVal / Size; 1519 int64_t Rem = StepVal % Size; 1520 if (Rem) 1521 return std::nullopt; 1522 1523 if (!ShouldCheckWrap) 1524 return Stride; 1525 1526 // The address calculation must not wrap. Otherwise, a dependence could be 1527 // inverted. 1528 if (isNoWrapAddRec(Ptr, AR, PSE, Lp)) 1529 return Stride; 1530 1531 // An inbounds getelementptr that is a AddRec with a unit stride 1532 // cannot wrap per definition. If it did, the result would be poison 1533 // and any memory access dependent on it would be immediate UB 1534 // when executed. 1535 if (auto *GEP = dyn_cast<GetElementPtrInst>(Ptr); 1536 GEP && GEP->isInBounds() && (Stride == 1 || Stride == -1)) 1537 return Stride; 1538 1539 // If the null pointer is undefined, then a access sequence which would 1540 // otherwise access it can be assumed not to unsigned wrap. Note that this 1541 // assumes the object in memory is aligned to the natural alignment. 1542 unsigned AddrSpace = Ty->getPointerAddressSpace(); 1543 if (!NullPointerIsDefined(Lp->getHeader()->getParent(), AddrSpace) && 1544 (Stride == 1 || Stride == -1)) 1545 return Stride; 1546 1547 if (Assume) { 1548 PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); 1549 LLVM_DEBUG(dbgs() << "LAA: Pointer may wrap:\n" 1550 << "LAA: Pointer: " << *Ptr << "\n" 1551 << "LAA: SCEV: " << *AR << "\n" 1552 << "LAA: Added an overflow assumption\n"); 1553 return Stride; 1554 } 1555 LLVM_DEBUG( 1556 dbgs() << "LAA: Bad stride - Pointer may wrap in the address space " 1557 << *Ptr << " SCEV: " << *AR << "\n"); 1558 return std::nullopt; 1559 } 1560 1561 std::optional<int> llvm::getPointersDiff(Type *ElemTyA, Value *PtrA, 1562 Type *ElemTyB, Value *PtrB, 1563 const DataLayout &DL, 1564 ScalarEvolution &SE, bool StrictCheck, 1565 bool CheckType) { 1566 assert(PtrA && PtrB && "Expected non-nullptr pointers."); 1567 1568 // Make sure that A and B are different pointers. 1569 if (PtrA == PtrB) 1570 return 0; 1571 1572 // Make sure that the element types are the same if required. 1573 if (CheckType && ElemTyA != ElemTyB) 1574 return std::nullopt; 1575 1576 unsigned ASA = PtrA->getType()->getPointerAddressSpace(); 1577 unsigned ASB = PtrB->getType()->getPointerAddressSpace(); 1578 1579 // Check that the address spaces match. 1580 if (ASA != ASB) 1581 return std::nullopt; 1582 unsigned IdxWidth = DL.getIndexSizeInBits(ASA); 1583 1584 APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0); 1585 Value *PtrA1 = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA); 1586 Value *PtrB1 = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB); 1587 1588 int Val; 1589 if (PtrA1 == PtrB1) { 1590 // Retrieve the address space again as pointer stripping now tracks through 1591 // `addrspacecast`. 1592 ASA = cast<PointerType>(PtrA1->getType())->getAddressSpace(); 1593 ASB = cast<PointerType>(PtrB1->getType())->getAddressSpace(); 1594 // Check that the address spaces match and that the pointers are valid. 1595 if (ASA != ASB) 1596 return std::nullopt; 1597 1598 IdxWidth = DL.getIndexSizeInBits(ASA); 1599 OffsetA = OffsetA.sextOrTrunc(IdxWidth); 1600 OffsetB = OffsetB.sextOrTrunc(IdxWidth); 1601 1602 OffsetB -= OffsetA; 1603 Val = OffsetB.getSExtValue(); 1604 } else { 1605 // Otherwise compute the distance with SCEV between the base pointers. 1606 const SCEV *PtrSCEVA = SE.getSCEV(PtrA); 1607 const SCEV *PtrSCEVB = SE.getSCEV(PtrB); 1608 const auto *Diff = 1609 dyn_cast<SCEVConstant>(SE.getMinusSCEV(PtrSCEVB, PtrSCEVA)); 1610 if (!Diff) 1611 return std::nullopt; 1612 Val = Diff->getAPInt().getSExtValue(); 1613 } 1614 int Size = DL.getTypeStoreSize(ElemTyA); 1615 int Dist = Val / Size; 1616 1617 // Ensure that the calculated distance matches the type-based one after all 1618 // the bitcasts removal in the provided pointers. 1619 if (!StrictCheck || Dist * Size == Val) 1620 return Dist; 1621 return std::nullopt; 1622 } 1623 1624 bool llvm::sortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy, 1625 const DataLayout &DL, ScalarEvolution &SE, 1626 SmallVectorImpl<unsigned> &SortedIndices) { 1627 assert(llvm::all_of( 1628 VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && 1629 "Expected list of pointer operands."); 1630 // Walk over the pointers, and map each of them to an offset relative to 1631 // first pointer in the array. 1632 Value *Ptr0 = VL[0]; 1633 1634 using DistOrdPair = std::pair<int64_t, int>; 1635 auto Compare = llvm::less_first(); 1636 std::set<DistOrdPair, decltype(Compare)> Offsets(Compare); 1637 Offsets.emplace(0, 0); 1638 int Cnt = 1; 1639 bool IsConsecutive = true; 1640 for (auto *Ptr : VL.drop_front()) { 1641 std::optional<int> Diff = getPointersDiff(ElemTy, Ptr0, ElemTy, Ptr, DL, SE, 1642 /*StrictCheck=*/true); 1643 if (!Diff) 1644 return false; 1645 1646 // Check if the pointer with the same offset is found. 1647 int64_t Offset = *Diff; 1648 auto Res = Offsets.emplace(Offset, Cnt); 1649 if (!Res.second) 1650 return false; 1651 // Consecutive order if the inserted element is the last one. 1652 IsConsecutive = IsConsecutive && std::next(Res.first) == Offsets.end(); 1653 ++Cnt; 1654 } 1655 SortedIndices.clear(); 1656 if (!IsConsecutive) { 1657 // Fill SortedIndices array only if it is non-consecutive. 1658 SortedIndices.resize(VL.size()); 1659 Cnt = 0; 1660 for (const std::pair<int64_t, int> &Pair : Offsets) { 1661 SortedIndices[Cnt] = Pair.second; 1662 ++Cnt; 1663 } 1664 } 1665 return true; 1666 } 1667 1668 /// Returns true if the memory operations \p A and \p B are consecutive. 1669 bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL, 1670 ScalarEvolution &SE, bool CheckType) { 1671 Value *PtrA = getLoadStorePointerOperand(A); 1672 Value *PtrB = getLoadStorePointerOperand(B); 1673 if (!PtrA || !PtrB) 1674 return false; 1675 Type *ElemTyA = getLoadStoreType(A); 1676 Type *ElemTyB = getLoadStoreType(B); 1677 std::optional<int> Diff = 1678 getPointersDiff(ElemTyA, PtrA, ElemTyB, PtrB, DL, SE, 1679 /*StrictCheck=*/true, CheckType); 1680 return Diff && *Diff == 1; 1681 } 1682 1683 void MemoryDepChecker::addAccess(StoreInst *SI) { 1684 visitPointers(SI->getPointerOperand(), *InnermostLoop, 1685 [this, SI](Value *Ptr) { 1686 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx); 1687 InstMap.push_back(SI); 1688 ++AccessIdx; 1689 }); 1690 } 1691 1692 void MemoryDepChecker::addAccess(LoadInst *LI) { 1693 visitPointers(LI->getPointerOperand(), *InnermostLoop, 1694 [this, LI](Value *Ptr) { 1695 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx); 1696 InstMap.push_back(LI); 1697 ++AccessIdx; 1698 }); 1699 } 1700 1701 MemoryDepChecker::VectorizationSafetyStatus 1702 MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) { 1703 switch (Type) { 1704 case NoDep: 1705 case Forward: 1706 case BackwardVectorizable: 1707 return VectorizationSafetyStatus::Safe; 1708 1709 case Unknown: 1710 return VectorizationSafetyStatus::PossiblySafeWithRtChecks; 1711 case ForwardButPreventsForwarding: 1712 case Backward: 1713 case BackwardVectorizableButPreventsForwarding: 1714 case IndirectUnsafe: 1715 return VectorizationSafetyStatus::Unsafe; 1716 } 1717 llvm_unreachable("unexpected DepType!"); 1718 } 1719 1720 bool MemoryDepChecker::Dependence::isBackward() const { 1721 switch (Type) { 1722 case NoDep: 1723 case Forward: 1724 case ForwardButPreventsForwarding: 1725 case Unknown: 1726 case IndirectUnsafe: 1727 return false; 1728 1729 case BackwardVectorizable: 1730 case Backward: 1731 case BackwardVectorizableButPreventsForwarding: 1732 return true; 1733 } 1734 llvm_unreachable("unexpected DepType!"); 1735 } 1736 1737 bool MemoryDepChecker::Dependence::isPossiblyBackward() const { 1738 return isBackward() || Type == Unknown; 1739 } 1740 1741 bool MemoryDepChecker::Dependence::isForward() const { 1742 switch (Type) { 1743 case Forward: 1744 case ForwardButPreventsForwarding: 1745 return true; 1746 1747 case NoDep: 1748 case Unknown: 1749 case BackwardVectorizable: 1750 case Backward: 1751 case BackwardVectorizableButPreventsForwarding: 1752 case IndirectUnsafe: 1753 return false; 1754 } 1755 llvm_unreachable("unexpected DepType!"); 1756 } 1757 1758 bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance, 1759 uint64_t TypeByteSize) { 1760 // If loads occur at a distance that is not a multiple of a feasible vector 1761 // factor store-load forwarding does not take place. 1762 // Positive dependences might cause troubles because vectorizing them might 1763 // prevent store-load forwarding making vectorized code run a lot slower. 1764 // a[i] = a[i-3] ^ a[i-8]; 1765 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and 1766 // hence on your typical architecture store-load forwarding does not take 1767 // place. Vectorizing in such cases does not make sense. 1768 // Store-load forwarding distance. 1769 1770 // After this many iterations store-to-load forwarding conflicts should not 1771 // cause any slowdowns. 1772 const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize; 1773 // Maximum vector factor. 1774 uint64_t MaxVFWithoutSLForwardIssues = std::min( 1775 VectorizerParams::MaxVectorWidth * TypeByteSize, MinDepDistBytes); 1776 1777 // Compute the smallest VF at which the store and load would be misaligned. 1778 for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues; 1779 VF *= 2) { 1780 // If the number of vector iteration between the store and the load are 1781 // small we could incur conflicts. 1782 if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) { 1783 MaxVFWithoutSLForwardIssues = (VF >> 1); 1784 break; 1785 } 1786 } 1787 1788 if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) { 1789 LLVM_DEBUG( 1790 dbgs() << "LAA: Distance " << Distance 1791 << " that could cause a store-load forwarding conflict\n"); 1792 return true; 1793 } 1794 1795 if (MaxVFWithoutSLForwardIssues < MinDepDistBytes && 1796 MaxVFWithoutSLForwardIssues != 1797 VectorizerParams::MaxVectorWidth * TypeByteSize) 1798 MinDepDistBytes = MaxVFWithoutSLForwardIssues; 1799 return false; 1800 } 1801 1802 void MemoryDepChecker::mergeInStatus(VectorizationSafetyStatus S) { 1803 if (Status < S) 1804 Status = S; 1805 } 1806 1807 /// Given a dependence-distance \p Dist between two 1808 /// memory accesses, that have the same stride whose absolute value is given 1809 /// in \p Stride, and that have the same type size \p TypeByteSize, 1810 /// in a loop whose takenCount is \p BackedgeTakenCount, check if it is 1811 /// possible to prove statically that the dependence distance is larger 1812 /// than the range that the accesses will travel through the execution of 1813 /// the loop. If so, return true; false otherwise. This is useful for 1814 /// example in loops such as the following (PR31098): 1815 /// for (i = 0; i < D; ++i) { 1816 /// = out[i]; 1817 /// out[i+D] = 1818 /// } 1819 static bool isSafeDependenceDistance(const DataLayout &DL, ScalarEvolution &SE, 1820 const SCEV &BackedgeTakenCount, 1821 const SCEV &Dist, uint64_t Stride, 1822 uint64_t TypeByteSize) { 1823 1824 // If we can prove that 1825 // (**) |Dist| > BackedgeTakenCount * Step 1826 // where Step is the absolute stride of the memory accesses in bytes, 1827 // then there is no dependence. 1828 // 1829 // Rationale: 1830 // We basically want to check if the absolute distance (|Dist/Step|) 1831 // is >= the loop iteration count (or > BackedgeTakenCount). 1832 // This is equivalent to the Strong SIV Test (Practical Dependence Testing, 1833 // Section 4.2.1); Note, that for vectorization it is sufficient to prove 1834 // that the dependence distance is >= VF; This is checked elsewhere. 1835 // But in some cases we can prune dependence distances early, and 1836 // even before selecting the VF, and without a runtime test, by comparing 1837 // the distance against the loop iteration count. Since the vectorized code 1838 // will be executed only if LoopCount >= VF, proving distance >= LoopCount 1839 // also guarantees that distance >= VF. 1840 // 1841 const uint64_t ByteStride = Stride * TypeByteSize; 1842 const SCEV *Step = SE.getConstant(BackedgeTakenCount.getType(), ByteStride); 1843 const SCEV *Product = SE.getMulExpr(&BackedgeTakenCount, Step); 1844 1845 const SCEV *CastedDist = &Dist; 1846 const SCEV *CastedProduct = Product; 1847 uint64_t DistTypeSizeBits = DL.getTypeSizeInBits(Dist.getType()); 1848 uint64_t ProductTypeSizeBits = DL.getTypeSizeInBits(Product->getType()); 1849 1850 // The dependence distance can be positive/negative, so we sign extend Dist; 1851 // The multiplication of the absolute stride in bytes and the 1852 // backedgeTakenCount is non-negative, so we zero extend Product. 1853 if (DistTypeSizeBits > ProductTypeSizeBits) 1854 CastedProduct = SE.getZeroExtendExpr(Product, Dist.getType()); 1855 else 1856 CastedDist = SE.getNoopOrSignExtend(&Dist, Product->getType()); 1857 1858 // Is Dist - (BackedgeTakenCount * Step) > 0 ? 1859 // (If so, then we have proven (**) because |Dist| >= Dist) 1860 const SCEV *Minus = SE.getMinusSCEV(CastedDist, CastedProduct); 1861 if (SE.isKnownPositive(Minus)) 1862 return true; 1863 1864 // Second try: Is -Dist - (BackedgeTakenCount * Step) > 0 ? 1865 // (If so, then we have proven (**) because |Dist| >= -1*Dist) 1866 const SCEV *NegDist = SE.getNegativeSCEV(CastedDist); 1867 Minus = SE.getMinusSCEV(NegDist, CastedProduct); 1868 if (SE.isKnownPositive(Minus)) 1869 return true; 1870 1871 return false; 1872 } 1873 1874 /// Check the dependence for two accesses with the same stride \p Stride. 1875 /// \p Distance is the positive distance and \p TypeByteSize is type size in 1876 /// bytes. 1877 /// 1878 /// \returns true if they are independent. 1879 static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride, 1880 uint64_t TypeByteSize) { 1881 assert(Stride > 1 && "The stride must be greater than 1"); 1882 assert(TypeByteSize > 0 && "The type size in byte must be non-zero"); 1883 assert(Distance > 0 && "The distance must be non-zero"); 1884 1885 // Skip if the distance is not multiple of type byte size. 1886 if (Distance % TypeByteSize) 1887 return false; 1888 1889 uint64_t ScaledDist = Distance / TypeByteSize; 1890 1891 // No dependence if the scaled distance is not multiple of the stride. 1892 // E.g. 1893 // for (i = 0; i < 1024 ; i += 4) 1894 // A[i+2] = A[i] + 1; 1895 // 1896 // Two accesses in memory (scaled distance is 2, stride is 4): 1897 // | A[0] | | | | A[4] | | | | 1898 // | | | A[2] | | | | A[6] | | 1899 // 1900 // E.g. 1901 // for (i = 0; i < 1024 ; i += 3) 1902 // A[i+4] = A[i] + 1; 1903 // 1904 // Two accesses in memory (scaled distance is 4, stride is 3): 1905 // | A[0] | | | A[3] | | | A[6] | | | 1906 // | | | | | A[4] | | | A[7] | | 1907 return ScaledDist % Stride; 1908 } 1909 1910 /// Returns true if any of the underlying objects has a loop varying address, 1911 /// i.e. may change in \p L. 1912 static bool 1913 isLoopVariantIndirectAddress(ArrayRef<const Value *> UnderlyingObjects, 1914 ScalarEvolution &SE, const Loop *L) { 1915 return any_of(UnderlyingObjects, [&SE, L](const Value *UO) { 1916 return !SE.isLoopInvariant(SE.getSCEV(const_cast<Value *>(UO)), L); 1917 }); 1918 } 1919 1920 // Get the dependence distance, stride, type size in whether i is a write for 1921 // the dependence between A and B. Returns a DepType, if we can prove there's 1922 // no dependence or the analysis fails. Outlined to lambda to limit he scope 1923 // of various temporary variables, like A/BPtr, StrideA/BPtr and others. 1924 // Returns either the dependence result, if it could already be determined, or a 1925 // tuple with (Distance, Stride, TypeSize, AIsWrite, BIsWrite). 1926 static std::variant<MemoryDepChecker::Dependence::DepType, 1927 std::tuple<const SCEV *, uint64_t, uint64_t, bool, bool>> 1928 getDependenceDistanceStrideAndSize( 1929 const AccessAnalysis::MemAccessInfo &A, Instruction *AInst, 1930 const AccessAnalysis::MemAccessInfo &B, Instruction *BInst, 1931 const DenseMap<Value *, const SCEV *> &Strides, 1932 const DenseMap<Value *, SmallVector<const Value *, 16>> &UnderlyingObjects, 1933 PredicatedScalarEvolution &PSE, const Loop *InnermostLoop) { 1934 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout(); 1935 auto &SE = *PSE.getSE(); 1936 auto [APtr, AIsWrite] = A; 1937 auto [BPtr, BIsWrite] = B; 1938 1939 // Two reads are independent. 1940 if (!AIsWrite && !BIsWrite) 1941 return MemoryDepChecker::Dependence::NoDep; 1942 1943 Type *ATy = getLoadStoreType(AInst); 1944 Type *BTy = getLoadStoreType(BInst); 1945 1946 // We cannot check pointers in different address spaces. 1947 if (APtr->getType()->getPointerAddressSpace() != 1948 BPtr->getType()->getPointerAddressSpace()) 1949 return MemoryDepChecker::Dependence::Unknown; 1950 1951 int64_t StrideAPtr = 1952 getPtrStride(PSE, ATy, APtr, InnermostLoop, Strides, true).value_or(0); 1953 int64_t StrideBPtr = 1954 getPtrStride(PSE, BTy, BPtr, InnermostLoop, Strides, true).value_or(0); 1955 1956 const SCEV *Src = PSE.getSCEV(APtr); 1957 const SCEV *Sink = PSE.getSCEV(BPtr); 1958 1959 // If the induction step is negative we have to invert source and sink of the 1960 // dependence when measuring the distance between them. We should not swap 1961 // AIsWrite with BIsWrite, as their uses expect them in program order. 1962 if (StrideAPtr < 0) { 1963 std::swap(Src, Sink); 1964 std::swap(AInst, BInst); 1965 } 1966 1967 const SCEV *Dist = SE.getMinusSCEV(Sink, Src); 1968 1969 LLVM_DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink 1970 << "(Induction step: " << StrideAPtr << ")\n"); 1971 LLVM_DEBUG(dbgs() << "LAA: Distance for " << *AInst << " to " << *BInst 1972 << ": " << *Dist << "\n"); 1973 1974 // Needs accesses where the addresses of the accessed underlying objects do 1975 // not change within the loop. 1976 if (isLoopVariantIndirectAddress(UnderlyingObjects.find(APtr)->second, SE, 1977 InnermostLoop) || 1978 isLoopVariantIndirectAddress(UnderlyingObjects.find(BPtr)->second, SE, 1979 InnermostLoop)) 1980 return MemoryDepChecker::Dependence::IndirectUnsafe; 1981 1982 // Need accesses with constant stride. We don't want to vectorize 1983 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap 1984 // in the address space. 1985 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr) { 1986 LLVM_DEBUG(dbgs() << "Pointer access with non-constant stride\n"); 1987 return MemoryDepChecker::Dependence::Unknown; 1988 } 1989 1990 uint64_t TypeByteSize = DL.getTypeAllocSize(ATy); 1991 bool HasSameSize = 1992 DL.getTypeStoreSizeInBits(ATy) == DL.getTypeStoreSizeInBits(BTy); 1993 if (!HasSameSize) 1994 TypeByteSize = 0; 1995 uint64_t Stride = std::abs(StrideAPtr); 1996 return std::make_tuple(Dist, Stride, TypeByteSize, AIsWrite, BIsWrite); 1997 } 1998 1999 MemoryDepChecker::Dependence::DepType MemoryDepChecker::isDependent( 2000 const MemAccessInfo &A, unsigned AIdx, const MemAccessInfo &B, 2001 unsigned BIdx, const DenseMap<Value *, const SCEV *> &Strides, 2002 const DenseMap<Value *, SmallVector<const Value *, 16>> 2003 &UnderlyingObjects) { 2004 assert(AIdx < BIdx && "Must pass arguments in program order"); 2005 2006 // Get the dependence distance, stride, type size and what access writes for 2007 // the dependence between A and B. 2008 auto Res = getDependenceDistanceStrideAndSize( 2009 A, InstMap[AIdx], B, InstMap[BIdx], Strides, UnderlyingObjects, PSE, 2010 InnermostLoop); 2011 if (std::holds_alternative<Dependence::DepType>(Res)) 2012 return std::get<Dependence::DepType>(Res); 2013 2014 const auto &[Dist, Stride, TypeByteSize, AIsWrite, BIsWrite] = 2015 std::get<std::tuple<const SCEV *, uint64_t, uint64_t, bool, bool>>(Res); 2016 bool HasSameSize = TypeByteSize > 0; 2017 2018 ScalarEvolution &SE = *PSE.getSE(); 2019 auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout(); 2020 if (!isa<SCEVCouldNotCompute>(Dist) && HasSameSize && 2021 isSafeDependenceDistance(DL, SE, *(PSE.getBackedgeTakenCount()), *Dist, 2022 Stride, TypeByteSize)) 2023 return Dependence::NoDep; 2024 2025 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist); 2026 if (!C) { 2027 LLVM_DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n"); 2028 FoundNonConstantDistanceDependence = true; 2029 return Dependence::Unknown; 2030 } 2031 2032 const APInt &Val = C->getAPInt(); 2033 int64_t Distance = Val.getSExtValue(); 2034 2035 // Attempt to prove strided accesses independent. 2036 if (std::abs(Distance) > 0 && Stride > 1 && HasSameSize && 2037 areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) { 2038 LLVM_DEBUG(dbgs() << "LAA: Strided accesses are independent\n"); 2039 return Dependence::NoDep; 2040 } 2041 2042 // Negative distances are not plausible dependencies. 2043 if (Val.isNegative()) { 2044 bool IsTrueDataDependence = (AIsWrite && !BIsWrite); 2045 // There is no need to update MaxSafeVectorWidthInBits after call to 2046 // couldPreventStoreLoadForward, even if it changed MinDepDistBytes, 2047 // since a forward dependency will allow vectorization using any width. 2048 if (IsTrueDataDependence && EnableForwardingConflictDetection && 2049 (!HasSameSize || couldPreventStoreLoadForward(Val.abs().getZExtValue(), 2050 TypeByteSize))) { 2051 LLVM_DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n"); 2052 return Dependence::ForwardButPreventsForwarding; 2053 } 2054 2055 LLVM_DEBUG(dbgs() << "LAA: Dependence is negative\n"); 2056 return Dependence::Forward; 2057 } 2058 2059 // Write to the same location with the same size. 2060 if (Val == 0) { 2061 if (HasSameSize) 2062 return Dependence::Forward; 2063 LLVM_DEBUG( 2064 dbgs() << "LAA: Zero dependence difference but different type sizes\n"); 2065 return Dependence::Unknown; 2066 } 2067 2068 assert(Val.isStrictlyPositive() && "Expect a positive value"); 2069 2070 if (!HasSameSize) { 2071 LLVM_DEBUG(dbgs() << "LAA: ReadWrite-Write positive dependency with " 2072 "different type sizes\n"); 2073 return Dependence::Unknown; 2074 } 2075 2076 // Bail out early if passed-in parameters make vectorization not feasible. 2077 unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ? 2078 VectorizerParams::VectorizationFactor : 1); 2079 unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ? 2080 VectorizerParams::VectorizationInterleave : 1); 2081 // The minimum number of iterations for a vectorized/unrolled version. 2082 unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U); 2083 2084 // It's not vectorizable if the distance is smaller than the minimum distance 2085 // needed for a vectroized/unrolled version. Vectorizing one iteration in 2086 // front needs TypeByteSize * Stride. Vectorizing the last iteration needs 2087 // TypeByteSize (No need to plus the last gap distance). 2088 // 2089 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes. 2090 // foo(int *A) { 2091 // int *B = (int *)((char *)A + 14); 2092 // for (i = 0 ; i < 1024 ; i += 2) 2093 // B[i] = A[i] + 1; 2094 // } 2095 // 2096 // Two accesses in memory (stride is 2): 2097 // | A[0] | | A[2] | | A[4] | | A[6] | | 2098 // | B[0] | | B[2] | | B[4] | 2099 // 2100 // Distance needs for vectorizing iterations except the last iteration: 2101 // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4. 2102 // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4. 2103 // 2104 // If MinNumIter is 2, it is vectorizable as the minimum distance needed is 2105 // 12, which is less than distance. 2106 // 2107 // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4), 2108 // the minimum distance needed is 28, which is greater than distance. It is 2109 // not safe to do vectorization. 2110 uint64_t MinDistanceNeeded = 2111 TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize; 2112 if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) { 2113 LLVM_DEBUG(dbgs() << "LAA: Failure because of positive distance " 2114 << Distance << '\n'); 2115 return Dependence::Backward; 2116 } 2117 2118 // Unsafe if the minimum distance needed is greater than smallest dependence 2119 // distance distance. 2120 if (MinDistanceNeeded > MinDepDistBytes) { 2121 LLVM_DEBUG(dbgs() << "LAA: Failure because it needs at least " 2122 << MinDistanceNeeded << " size in bytes\n"); 2123 return Dependence::Backward; 2124 } 2125 2126 // Positive distance bigger than max vectorization factor. 2127 // FIXME: Should use max factor instead of max distance in bytes, which could 2128 // not handle different types. 2129 // E.g. Assume one char is 1 byte in memory and one int is 4 bytes. 2130 // void foo (int *A, char *B) { 2131 // for (unsigned i = 0; i < 1024; i++) { 2132 // A[i+2] = A[i] + 1; 2133 // B[i+2] = B[i] + 1; 2134 // } 2135 // } 2136 // 2137 // This case is currently unsafe according to the max safe distance. If we 2138 // analyze the two accesses on array B, the max safe dependence distance 2139 // is 2. Then we analyze the accesses on array A, the minimum distance needed 2140 // is 8, which is less than 2 and forbidden vectorization, But actually 2141 // both A and B could be vectorized by 2 iterations. 2142 MinDepDistBytes = 2143 std::min(static_cast<uint64_t>(Distance), MinDepDistBytes); 2144 2145 bool IsTrueDataDependence = (!AIsWrite && BIsWrite); 2146 uint64_t MinDepDistBytesOld = MinDepDistBytes; 2147 if (IsTrueDataDependence && EnableForwardingConflictDetection && 2148 couldPreventStoreLoadForward(Distance, TypeByteSize)) { 2149 // Sanity check that we didn't update MinDepDistBytes when calling 2150 // couldPreventStoreLoadForward 2151 assert(MinDepDistBytes == MinDepDistBytesOld && 2152 "An update to MinDepDistBytes requires an update to " 2153 "MaxSafeVectorWidthInBits"); 2154 (void)MinDepDistBytesOld; 2155 return Dependence::BackwardVectorizableButPreventsForwarding; 2156 } 2157 2158 // An update to MinDepDistBytes requires an update to MaxSafeVectorWidthInBits 2159 // since there is a backwards dependency. 2160 uint64_t MaxVF = MinDepDistBytes / (TypeByteSize * Stride); 2161 LLVM_DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue() 2162 << " with max VF = " << MaxVF << '\n'); 2163 uint64_t MaxVFInBits = MaxVF * TypeByteSize * 8; 2164 MaxSafeVectorWidthInBits = std::min(MaxSafeVectorWidthInBits, MaxVFInBits); 2165 return Dependence::BackwardVectorizable; 2166 } 2167 2168 bool MemoryDepChecker::areDepsSafe( 2169 DepCandidates &AccessSets, MemAccessInfoList &CheckDeps, 2170 const DenseMap<Value *, const SCEV *> &Strides, 2171 const DenseMap<Value *, SmallVector<const Value *, 16>> 2172 &UnderlyingObjects) { 2173 2174 MinDepDistBytes = -1; 2175 SmallPtrSet<MemAccessInfo, 8> Visited; 2176 for (MemAccessInfo CurAccess : CheckDeps) { 2177 if (Visited.count(CurAccess)) 2178 continue; 2179 2180 // Get the relevant memory access set. 2181 EquivalenceClasses<MemAccessInfo>::iterator I = 2182 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess)); 2183 2184 // Check accesses within this set. 2185 EquivalenceClasses<MemAccessInfo>::member_iterator AI = 2186 AccessSets.member_begin(I); 2187 EquivalenceClasses<MemAccessInfo>::member_iterator AE = 2188 AccessSets.member_end(); 2189 2190 // Check every access pair. 2191 while (AI != AE) { 2192 Visited.insert(*AI); 2193 bool AIIsWrite = AI->getInt(); 2194 // Check loads only against next equivalent class, but stores also against 2195 // other stores in the same equivalence class - to the same address. 2196 EquivalenceClasses<MemAccessInfo>::member_iterator OI = 2197 (AIIsWrite ? AI : std::next(AI)); 2198 while (OI != AE) { 2199 // Check every accessing instruction pair in program order. 2200 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(), 2201 I1E = Accesses[*AI].end(); I1 != I1E; ++I1) 2202 // Scan all accesses of another equivalence class, but only the next 2203 // accesses of the same equivalent class. 2204 for (std::vector<unsigned>::iterator 2205 I2 = (OI == AI ? std::next(I1) : Accesses[*OI].begin()), 2206 I2E = (OI == AI ? I1E : Accesses[*OI].end()); 2207 I2 != I2E; ++I2) { 2208 auto A = std::make_pair(&*AI, *I1); 2209 auto B = std::make_pair(&*OI, *I2); 2210 2211 assert(*I1 != *I2); 2212 if (*I1 > *I2) 2213 std::swap(A, B); 2214 2215 Dependence::DepType Type = 2216 isDependent(*A.first, A.second, *B.first, B.second, Strides, 2217 UnderlyingObjects); 2218 mergeInStatus(Dependence::isSafeForVectorization(Type)); 2219 2220 // Gather dependences unless we accumulated MaxDependences 2221 // dependences. In that case return as soon as we find the first 2222 // unsafe dependence. This puts a limit on this quadratic 2223 // algorithm. 2224 if (RecordDependences) { 2225 if (Type != Dependence::NoDep) 2226 Dependences.push_back(Dependence(A.second, B.second, Type)); 2227 2228 if (Dependences.size() >= MaxDependences) { 2229 RecordDependences = false; 2230 Dependences.clear(); 2231 LLVM_DEBUG(dbgs() 2232 << "Too many dependences, stopped recording\n"); 2233 } 2234 } 2235 if (!RecordDependences && !isSafeForVectorization()) 2236 return false; 2237 } 2238 ++OI; 2239 } 2240 AI++; 2241 } 2242 } 2243 2244 LLVM_DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n"); 2245 return isSafeForVectorization(); 2246 } 2247 2248 SmallVector<Instruction *, 4> 2249 MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const { 2250 MemAccessInfo Access(Ptr, isWrite); 2251 auto &IndexVector = Accesses.find(Access)->second; 2252 2253 SmallVector<Instruction *, 4> Insts; 2254 transform(IndexVector, 2255 std::back_inserter(Insts), 2256 [&](unsigned Idx) { return this->InstMap[Idx]; }); 2257 return Insts; 2258 } 2259 2260 const char *MemoryDepChecker::Dependence::DepName[] = { 2261 "NoDep", 2262 "Unknown", 2263 "IndidrectUnsafe", 2264 "Forward", 2265 "ForwardButPreventsForwarding", 2266 "Backward", 2267 "BackwardVectorizable", 2268 "BackwardVectorizableButPreventsForwarding"}; 2269 2270 void MemoryDepChecker::Dependence::print( 2271 raw_ostream &OS, unsigned Depth, 2272 const SmallVectorImpl<Instruction *> &Instrs) const { 2273 OS.indent(Depth) << DepName[Type] << ":\n"; 2274 OS.indent(Depth + 2) << *Instrs[Source] << " -> \n"; 2275 OS.indent(Depth + 2) << *Instrs[Destination] << "\n"; 2276 } 2277 2278 bool LoopAccessInfo::canAnalyzeLoop() { 2279 // We need to have a loop header. 2280 LLVM_DEBUG(dbgs() << "LAA: Found a loop in " 2281 << TheLoop->getHeader()->getParent()->getName() << ": " 2282 << TheLoop->getHeader()->getName() << '\n'); 2283 2284 // We can only analyze innermost loops. 2285 if (!TheLoop->isInnermost()) { 2286 LLVM_DEBUG(dbgs() << "LAA: loop is not the innermost loop\n"); 2287 recordAnalysis("NotInnerMostLoop") << "loop is not the innermost loop"; 2288 return false; 2289 } 2290 2291 // We must have a single backedge. 2292 if (TheLoop->getNumBackEdges() != 1) { 2293 LLVM_DEBUG( 2294 dbgs() << "LAA: loop control flow is not understood by analyzer\n"); 2295 recordAnalysis("CFGNotUnderstood") 2296 << "loop control flow is not understood by analyzer"; 2297 return false; 2298 } 2299 2300 // ScalarEvolution needs to be able to find the exit count. 2301 const SCEV *ExitCount = PSE->getBackedgeTakenCount(); 2302 if (isa<SCEVCouldNotCompute>(ExitCount)) { 2303 recordAnalysis("CantComputeNumberOfIterations") 2304 << "could not determine number of loop iterations"; 2305 LLVM_DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n"); 2306 return false; 2307 } 2308 2309 return true; 2310 } 2311 2312 void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI, 2313 const TargetLibraryInfo *TLI, 2314 DominatorTree *DT) { 2315 // Holds the Load and Store instructions. 2316 SmallVector<LoadInst *, 16> Loads; 2317 SmallVector<StoreInst *, 16> Stores; 2318 SmallPtrSet<MDNode *, 8> LoopAliasScopes; 2319 2320 // Holds all the different accesses in the loop. 2321 unsigned NumReads = 0; 2322 unsigned NumReadWrites = 0; 2323 2324 bool HasComplexMemInst = false; 2325 2326 // A runtime check is only legal to insert if there are no convergent calls. 2327 HasConvergentOp = false; 2328 2329 PtrRtChecking->Pointers.clear(); 2330 PtrRtChecking->Need = false; 2331 2332 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel(); 2333 2334 const bool EnableMemAccessVersioningOfLoop = 2335 EnableMemAccessVersioning && 2336 !TheLoop->getHeader()->getParent()->hasOptSize(); 2337 2338 // Traverse blocks in fixed RPOT order, regardless of their storage in the 2339 // loop info, as it may be arbitrary. 2340 LoopBlocksRPO RPOT(TheLoop); 2341 RPOT.perform(LI); 2342 for (BasicBlock *BB : RPOT) { 2343 // Scan the BB and collect legal loads and stores. Also detect any 2344 // convergent instructions. 2345 for (Instruction &I : *BB) { 2346 if (auto *Call = dyn_cast<CallBase>(&I)) { 2347 if (Call->isConvergent()) 2348 HasConvergentOp = true; 2349 } 2350 2351 // With both a non-vectorizable memory instruction and a convergent 2352 // operation, found in this loop, no reason to continue the search. 2353 if (HasComplexMemInst && HasConvergentOp) { 2354 CanVecMem = false; 2355 return; 2356 } 2357 2358 // Avoid hitting recordAnalysis multiple times. 2359 if (HasComplexMemInst) 2360 continue; 2361 2362 // Record alias scopes defined inside the loop. 2363 if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&I)) 2364 for (Metadata *Op : Decl->getScopeList()->operands()) 2365 LoopAliasScopes.insert(cast<MDNode>(Op)); 2366 2367 // Many math library functions read the rounding mode. We will only 2368 // vectorize a loop if it contains known function calls that don't set 2369 // the flag. Therefore, it is safe to ignore this read from memory. 2370 auto *Call = dyn_cast<CallInst>(&I); 2371 if (Call && getVectorIntrinsicIDForCall(Call, TLI)) 2372 continue; 2373 2374 // If this is a load, save it. If this instruction can read from memory 2375 // but is not a load, then we quit. Notice that we don't handle function 2376 // calls that read or write. 2377 if (I.mayReadFromMemory()) { 2378 // If the function has an explicit vectorized counterpart, we can safely 2379 // assume that it can be vectorized. 2380 if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() && 2381 !VFDatabase::getMappings(*Call).empty()) 2382 continue; 2383 2384 auto *Ld = dyn_cast<LoadInst>(&I); 2385 if (!Ld) { 2386 recordAnalysis("CantVectorizeInstruction", Ld) 2387 << "instruction cannot be vectorized"; 2388 HasComplexMemInst = true; 2389 continue; 2390 } 2391 if (!Ld->isSimple() && !IsAnnotatedParallel) { 2392 recordAnalysis("NonSimpleLoad", Ld) 2393 << "read with atomic ordering or volatile read"; 2394 LLVM_DEBUG(dbgs() << "LAA: Found a non-simple load.\n"); 2395 HasComplexMemInst = true; 2396 continue; 2397 } 2398 NumLoads++; 2399 Loads.push_back(Ld); 2400 DepChecker->addAccess(Ld); 2401 if (EnableMemAccessVersioningOfLoop) 2402 collectStridedAccess(Ld); 2403 continue; 2404 } 2405 2406 // Save 'store' instructions. Abort if other instructions write to memory. 2407 if (I.mayWriteToMemory()) { 2408 auto *St = dyn_cast<StoreInst>(&I); 2409 if (!St) { 2410 recordAnalysis("CantVectorizeInstruction", St) 2411 << "instruction cannot be vectorized"; 2412 HasComplexMemInst = true; 2413 continue; 2414 } 2415 if (!St->isSimple() && !IsAnnotatedParallel) { 2416 recordAnalysis("NonSimpleStore", St) 2417 << "write with atomic ordering or volatile write"; 2418 LLVM_DEBUG(dbgs() << "LAA: Found a non-simple store.\n"); 2419 HasComplexMemInst = true; 2420 continue; 2421 } 2422 NumStores++; 2423 Stores.push_back(St); 2424 DepChecker->addAccess(St); 2425 if (EnableMemAccessVersioningOfLoop) 2426 collectStridedAccess(St); 2427 } 2428 } // Next instr. 2429 } // Next block. 2430 2431 if (HasComplexMemInst) { 2432 CanVecMem = false; 2433 return; 2434 } 2435 2436 // Now we have two lists that hold the loads and the stores. 2437 // Next, we find the pointers that they use. 2438 2439 // Check if we see any stores. If there are no stores, then we don't 2440 // care if the pointers are *restrict*. 2441 if (!Stores.size()) { 2442 LLVM_DEBUG(dbgs() << "LAA: Found a read-only loop!\n"); 2443 CanVecMem = true; 2444 return; 2445 } 2446 2447 MemoryDepChecker::DepCandidates DependentAccesses; 2448 AccessAnalysis Accesses(TheLoop, AA, LI, DependentAccesses, *PSE, 2449 LoopAliasScopes); 2450 2451 // Holds the analyzed pointers. We don't want to call getUnderlyingObjects 2452 // multiple times on the same object. If the ptr is accessed twice, once 2453 // for read and once for write, it will only appear once (on the write 2454 // list). This is okay, since we are going to check for conflicts between 2455 // writes and between reads and writes, but not between reads and reads. 2456 SmallSet<std::pair<Value *, Type *>, 16> Seen; 2457 2458 // Record uniform store addresses to identify if we have multiple stores 2459 // to the same address. 2460 SmallPtrSet<Value *, 16> UniformStores; 2461 2462 for (StoreInst *ST : Stores) { 2463 Value *Ptr = ST->getPointerOperand(); 2464 2465 if (isInvariant(Ptr)) { 2466 // Record store instructions to loop invariant addresses 2467 StoresToInvariantAddresses.push_back(ST); 2468 HasDependenceInvolvingLoopInvariantAddress |= 2469 !UniformStores.insert(Ptr).second; 2470 } 2471 2472 // If we did *not* see this pointer before, insert it to the read-write 2473 // list. At this phase it is only a 'write' list. 2474 Type *AccessTy = getLoadStoreType(ST); 2475 if (Seen.insert({Ptr, AccessTy}).second) { 2476 ++NumReadWrites; 2477 2478 MemoryLocation Loc = MemoryLocation::get(ST); 2479 // The TBAA metadata could have a control dependency on the predication 2480 // condition, so we cannot rely on it when determining whether or not we 2481 // need runtime pointer checks. 2482 if (blockNeedsPredication(ST->getParent(), TheLoop, DT)) 2483 Loc.AATags.TBAA = nullptr; 2484 2485 visitPointers(const_cast<Value *>(Loc.Ptr), *TheLoop, 2486 [&Accesses, AccessTy, Loc](Value *Ptr) { 2487 MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr); 2488 Accesses.addStore(NewLoc, AccessTy); 2489 }); 2490 } 2491 } 2492 2493 if (IsAnnotatedParallel) { 2494 LLVM_DEBUG( 2495 dbgs() << "LAA: A loop annotated parallel, ignore memory dependency " 2496 << "checks.\n"); 2497 CanVecMem = true; 2498 return; 2499 } 2500 2501 for (LoadInst *LD : Loads) { 2502 Value *Ptr = LD->getPointerOperand(); 2503 // If we did *not* see this pointer before, insert it to the 2504 // read list. If we *did* see it before, then it is already in 2505 // the read-write list. This allows us to vectorize expressions 2506 // such as A[i] += x; Because the address of A[i] is a read-write 2507 // pointer. This only works if the index of A[i] is consecutive. 2508 // If the address of i is unknown (for example A[B[i]]) then we may 2509 // read a few words, modify, and write a few words, and some of the 2510 // words may be written to the same address. 2511 bool IsReadOnlyPtr = false; 2512 Type *AccessTy = getLoadStoreType(LD); 2513 if (Seen.insert({Ptr, AccessTy}).second || 2514 !getPtrStride(*PSE, LD->getType(), Ptr, TheLoop, SymbolicStrides).value_or(0)) { 2515 ++NumReads; 2516 IsReadOnlyPtr = true; 2517 } 2518 2519 // See if there is an unsafe dependency between a load to a uniform address and 2520 // store to the same uniform address. 2521 if (UniformStores.count(Ptr)) { 2522 LLVM_DEBUG(dbgs() << "LAA: Found an unsafe dependency between a uniform " 2523 "load and uniform store to the same address!\n"); 2524 HasDependenceInvolvingLoopInvariantAddress = true; 2525 } 2526 2527 MemoryLocation Loc = MemoryLocation::get(LD); 2528 // The TBAA metadata could have a control dependency on the predication 2529 // condition, so we cannot rely on it when determining whether or not we 2530 // need runtime pointer checks. 2531 if (blockNeedsPredication(LD->getParent(), TheLoop, DT)) 2532 Loc.AATags.TBAA = nullptr; 2533 2534 visitPointers(const_cast<Value *>(Loc.Ptr), *TheLoop, 2535 [&Accesses, AccessTy, Loc, IsReadOnlyPtr](Value *Ptr) { 2536 MemoryLocation NewLoc = Loc.getWithNewPtr(Ptr); 2537 Accesses.addLoad(NewLoc, AccessTy, IsReadOnlyPtr); 2538 }); 2539 } 2540 2541 // If we write (or read-write) to a single destination and there are no 2542 // other reads in this loop then is it safe to vectorize. 2543 if (NumReadWrites == 1 && NumReads == 0) { 2544 LLVM_DEBUG(dbgs() << "LAA: Found a write-only loop!\n"); 2545 CanVecMem = true; 2546 return; 2547 } 2548 2549 // Build dependence sets and check whether we need a runtime pointer bounds 2550 // check. 2551 Accesses.buildDependenceSets(); 2552 2553 // Find pointers with computable bounds. We are going to use this information 2554 // to place a runtime bound check. 2555 Value *UncomputablePtr = nullptr; 2556 bool CanDoRTIfNeeded = 2557 Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(), TheLoop, 2558 SymbolicStrides, UncomputablePtr, false); 2559 if (!CanDoRTIfNeeded) { 2560 auto *I = dyn_cast_or_null<Instruction>(UncomputablePtr); 2561 recordAnalysis("CantIdentifyArrayBounds", I) 2562 << "cannot identify array bounds"; 2563 LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because we can't find " 2564 << "the array bounds.\n"); 2565 CanVecMem = false; 2566 return; 2567 } 2568 2569 LLVM_DEBUG( 2570 dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n"); 2571 2572 CanVecMem = true; 2573 if (Accesses.isDependencyCheckNeeded()) { 2574 LLVM_DEBUG(dbgs() << "LAA: Checking memory dependencies\n"); 2575 CanVecMem = DepChecker->areDepsSafe( 2576 DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides, 2577 Accesses.getUnderlyingObjects()); 2578 2579 if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) { 2580 LLVM_DEBUG(dbgs() << "LAA: Retrying with memory checks\n"); 2581 2582 // Clear the dependency checks. We assume they are not needed. 2583 Accesses.resetDepChecks(*DepChecker); 2584 2585 PtrRtChecking->reset(); 2586 PtrRtChecking->Need = true; 2587 2588 auto *SE = PSE->getSE(); 2589 UncomputablePtr = nullptr; 2590 CanDoRTIfNeeded = Accesses.canCheckPtrAtRT( 2591 *PtrRtChecking, SE, TheLoop, SymbolicStrides, UncomputablePtr, true); 2592 2593 // Check that we found the bounds for the pointer. 2594 if (!CanDoRTIfNeeded) { 2595 auto *I = dyn_cast_or_null<Instruction>(UncomputablePtr); 2596 recordAnalysis("CantCheckMemDepsAtRunTime", I) 2597 << "cannot check memory dependencies at runtime"; 2598 LLVM_DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n"); 2599 CanVecMem = false; 2600 return; 2601 } 2602 2603 CanVecMem = true; 2604 } 2605 } 2606 2607 if (HasConvergentOp) { 2608 recordAnalysis("CantInsertRuntimeCheckWithConvergent") 2609 << "cannot add control dependency to convergent operation"; 2610 LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because a runtime check " 2611 "would be needed with a convergent operation\n"); 2612 CanVecMem = false; 2613 return; 2614 } 2615 2616 if (CanVecMem) 2617 LLVM_DEBUG( 2618 dbgs() << "LAA: No unsafe dependent memory operations in loop. We" 2619 << (PtrRtChecking->Need ? "" : " don't") 2620 << " need runtime memory checks.\n"); 2621 else 2622 emitUnsafeDependenceRemark(); 2623 } 2624 2625 void LoopAccessInfo::emitUnsafeDependenceRemark() { 2626 auto Deps = getDepChecker().getDependences(); 2627 if (!Deps) 2628 return; 2629 auto Found = llvm::find_if(*Deps, [](const MemoryDepChecker::Dependence &D) { 2630 return MemoryDepChecker::Dependence::isSafeForVectorization(D.Type) != 2631 MemoryDepChecker::VectorizationSafetyStatus::Safe; 2632 }); 2633 if (Found == Deps->end()) 2634 return; 2635 MemoryDepChecker::Dependence Dep = *Found; 2636 2637 LLVM_DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n"); 2638 2639 // Emit remark for first unsafe dependence 2640 bool HasForcedDistribution = false; 2641 std::optional<const MDOperand *> Value = 2642 findStringMetadataForLoop(TheLoop, "llvm.loop.distribute.enable"); 2643 if (Value) { 2644 const MDOperand *Op = *Value; 2645 assert(Op && mdconst::hasa<ConstantInt>(*Op) && "invalid metadata"); 2646 HasForcedDistribution = mdconst::extract<ConstantInt>(*Op)->getZExtValue(); 2647 } 2648 2649 const std::string Info = 2650 HasForcedDistribution 2651 ? "unsafe dependent memory operations in loop." 2652 : "unsafe dependent memory operations in loop. Use " 2653 "#pragma clang loop distribute(enable) to allow loop distribution " 2654 "to attempt to isolate the offending operations into a separate " 2655 "loop"; 2656 OptimizationRemarkAnalysis &R = 2657 recordAnalysis("UnsafeDep", Dep.getDestination(*this)) << Info; 2658 2659 switch (Dep.Type) { 2660 case MemoryDepChecker::Dependence::NoDep: 2661 case MemoryDepChecker::Dependence::Forward: 2662 case MemoryDepChecker::Dependence::BackwardVectorizable: 2663 llvm_unreachable("Unexpected dependence"); 2664 case MemoryDepChecker::Dependence::Backward: 2665 R << "\nBackward loop carried data dependence."; 2666 break; 2667 case MemoryDepChecker::Dependence::ForwardButPreventsForwarding: 2668 R << "\nForward loop carried data dependence that prevents " 2669 "store-to-load forwarding."; 2670 break; 2671 case MemoryDepChecker::Dependence::BackwardVectorizableButPreventsForwarding: 2672 R << "\nBackward loop carried data dependence that prevents " 2673 "store-to-load forwarding."; 2674 break; 2675 case MemoryDepChecker::Dependence::IndirectUnsafe: 2676 R << "\nUnsafe indirect dependence."; 2677 break; 2678 case MemoryDepChecker::Dependence::Unknown: 2679 R << "\nUnknown data dependence."; 2680 break; 2681 } 2682 2683 if (Instruction *I = Dep.getSource(*this)) { 2684 DebugLoc SourceLoc = I->getDebugLoc(); 2685 if (auto *DD = dyn_cast_or_null<Instruction>(getPointerOperand(I))) 2686 SourceLoc = DD->getDebugLoc(); 2687 if (SourceLoc) 2688 R << " Memory location is the same as accessed at " 2689 << ore::NV("Location", SourceLoc); 2690 } 2691 } 2692 2693 bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop, 2694 DominatorTree *DT) { 2695 assert(TheLoop->contains(BB) && "Unknown block used"); 2696 2697 // Blocks that do not dominate the latch need predication. 2698 BasicBlock* Latch = TheLoop->getLoopLatch(); 2699 return !DT->dominates(BB, Latch); 2700 } 2701 2702 OptimizationRemarkAnalysis &LoopAccessInfo::recordAnalysis(StringRef RemarkName, 2703 Instruction *I) { 2704 assert(!Report && "Multiple reports generated"); 2705 2706 Value *CodeRegion = TheLoop->getHeader(); 2707 DebugLoc DL = TheLoop->getStartLoc(); 2708 2709 if (I) { 2710 CodeRegion = I->getParent(); 2711 // If there is no debug location attached to the instruction, revert back to 2712 // using the loop's. 2713 if (I->getDebugLoc()) 2714 DL = I->getDebugLoc(); 2715 } 2716 2717 Report = std::make_unique<OptimizationRemarkAnalysis>(DEBUG_TYPE, RemarkName, DL, 2718 CodeRegion); 2719 return *Report; 2720 } 2721 2722 bool LoopAccessInfo::isInvariant(Value *V) const { 2723 auto *SE = PSE->getSE(); 2724 // TODO: Is this really what we want? Even without FP SCEV, we may want some 2725 // trivially loop-invariant FP values to be considered invariant. 2726 if (!SE->isSCEVable(V->getType())) 2727 return false; 2728 const SCEV *S = SE->getSCEV(V); 2729 return SE->isLoopInvariant(S, TheLoop); 2730 } 2731 2732 /// Find the operand of the GEP that should be checked for consecutive 2733 /// stores. This ignores trailing indices that have no effect on the final 2734 /// pointer. 2735 static unsigned getGEPInductionOperand(const GetElementPtrInst *Gep) { 2736 const DataLayout &DL = Gep->getModule()->getDataLayout(); 2737 unsigned LastOperand = Gep->getNumOperands() - 1; 2738 TypeSize GEPAllocSize = DL.getTypeAllocSize(Gep->getResultElementType()); 2739 2740 // Walk backwards and try to peel off zeros. 2741 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) { 2742 // Find the type we're currently indexing into. 2743 gep_type_iterator GEPTI = gep_type_begin(Gep); 2744 std::advance(GEPTI, LastOperand - 2); 2745 2746 // If it's a type with the same allocation size as the result of the GEP we 2747 // can peel off the zero index. 2748 TypeSize ElemSize = GEPTI.isStruct() 2749 ? DL.getTypeAllocSize(GEPTI.getIndexedType()) 2750 : GEPTI.getSequentialElementStride(DL); 2751 if (ElemSize != GEPAllocSize) 2752 break; 2753 --LastOperand; 2754 } 2755 2756 return LastOperand; 2757 } 2758 2759 /// If the argument is a GEP, then returns the operand identified by 2760 /// getGEPInductionOperand. However, if there is some other non-loop-invariant 2761 /// operand, it returns that instead. 2762 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE, Loop *Lp) { 2763 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr); 2764 if (!GEP) 2765 return Ptr; 2766 2767 unsigned InductionOperand = getGEPInductionOperand(GEP); 2768 2769 // Check that all of the gep indices are uniform except for our induction 2770 // operand. 2771 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) 2772 if (i != InductionOperand && 2773 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp)) 2774 return Ptr; 2775 return GEP->getOperand(InductionOperand); 2776 } 2777 2778 /// If a value has only one user that is a CastInst, return it. 2779 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) { 2780 Value *UniqueCast = nullptr; 2781 for (User *U : Ptr->users()) { 2782 CastInst *CI = dyn_cast<CastInst>(U); 2783 if (CI && CI->getType() == Ty) { 2784 if (!UniqueCast) 2785 UniqueCast = CI; 2786 else 2787 return nullptr; 2788 } 2789 } 2790 return UniqueCast; 2791 } 2792 2793 /// Get the stride of a pointer access in a loop. Looks for symbolic 2794 /// strides "a[i*stride]". Returns the symbolic stride, or null otherwise. 2795 static const SCEV *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE, Loop *Lp) { 2796 auto *PtrTy = dyn_cast<PointerType>(Ptr->getType()); 2797 if (!PtrTy || PtrTy->isAggregateType()) 2798 return nullptr; 2799 2800 // Try to remove a gep instruction to make the pointer (actually index at this 2801 // point) easier analyzable. If OrigPtr is equal to Ptr we are analyzing the 2802 // pointer, otherwise, we are analyzing the index. 2803 Value *OrigPtr = Ptr; 2804 2805 // The size of the pointer access. 2806 int64_t PtrAccessSize = 1; 2807 2808 Ptr = stripGetElementPtr(Ptr, SE, Lp); 2809 const SCEV *V = SE->getSCEV(Ptr); 2810 2811 if (Ptr != OrigPtr) 2812 // Strip off casts. 2813 while (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(V)) 2814 V = C->getOperand(); 2815 2816 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V); 2817 if (!S) 2818 return nullptr; 2819 2820 // If the pointer is invariant then there is no stride and it makes no 2821 // sense to add it here. 2822 if (Lp != S->getLoop()) 2823 return nullptr; 2824 2825 V = S->getStepRecurrence(*SE); 2826 if (!V) 2827 return nullptr; 2828 2829 // Strip off the size of access multiplication if we are still analyzing the 2830 // pointer. 2831 if (OrigPtr == Ptr) { 2832 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) { 2833 if (M->getOperand(0)->getSCEVType() != scConstant) 2834 return nullptr; 2835 2836 const APInt &APStepVal = cast<SCEVConstant>(M->getOperand(0))->getAPInt(); 2837 2838 // Huge step value - give up. 2839 if (APStepVal.getBitWidth() > 64) 2840 return nullptr; 2841 2842 int64_t StepVal = APStepVal.getSExtValue(); 2843 if (PtrAccessSize != StepVal) 2844 return nullptr; 2845 V = M->getOperand(1); 2846 } 2847 } 2848 2849 // Note that the restriction after this loop invariant check are only 2850 // profitability restrictions. 2851 if (!SE->isLoopInvariant(V, Lp)) 2852 return nullptr; 2853 2854 // Look for the loop invariant symbolic value. 2855 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V); 2856 if (!U) { 2857 const auto *C = dyn_cast<SCEVIntegralCastExpr>(V); 2858 if (!C) 2859 return nullptr; 2860 U = dyn_cast<SCEVUnknown>(C->getOperand()); 2861 if (!U) 2862 return nullptr; 2863 2864 // Match legacy behavior - this is not needed for correctness 2865 if (!getUniqueCastUse(U->getValue(), Lp, V->getType())) 2866 return nullptr; 2867 } 2868 2869 return V; 2870 } 2871 2872 void LoopAccessInfo::collectStridedAccess(Value *MemAccess) { 2873 Value *Ptr = getLoadStorePointerOperand(MemAccess); 2874 if (!Ptr) 2875 return; 2876 2877 // Note: getStrideFromPointer is a *profitability* heuristic. We 2878 // could broaden the scope of values returned here - to anything 2879 // which happens to be loop invariant and contributes to the 2880 // computation of an interesting IV - but we chose not to as we 2881 // don't have a cost model here, and broadening the scope exposes 2882 // far too many unprofitable cases. 2883 const SCEV *StrideExpr = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop); 2884 if (!StrideExpr) 2885 return; 2886 2887 LLVM_DEBUG(dbgs() << "LAA: Found a strided access that is a candidate for " 2888 "versioning:"); 2889 LLVM_DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *StrideExpr << "\n"); 2890 2891 if (!SpeculateUnitStride) { 2892 LLVM_DEBUG(dbgs() << " Chose not to due to -laa-speculate-unit-stride\n"); 2893 return; 2894 } 2895 2896 // Avoid adding the "Stride == 1" predicate when we know that 2897 // Stride >= Trip-Count. Such a predicate will effectively optimize a single 2898 // or zero iteration loop, as Trip-Count <= Stride == 1. 2899 // 2900 // TODO: We are currently not making a very informed decision on when it is 2901 // beneficial to apply stride versioning. It might make more sense that the 2902 // users of this analysis (such as the vectorizer) will trigger it, based on 2903 // their specific cost considerations; For example, in cases where stride 2904 // versioning does not help resolving memory accesses/dependences, the 2905 // vectorizer should evaluate the cost of the runtime test, and the benefit 2906 // of various possible stride specializations, considering the alternatives 2907 // of using gather/scatters (if available). 2908 2909 const SCEV *BETakenCount = PSE->getBackedgeTakenCount(); 2910 2911 // Match the types so we can compare the stride and the BETakenCount. 2912 // The Stride can be positive/negative, so we sign extend Stride; 2913 // The backedgeTakenCount is non-negative, so we zero extend BETakenCount. 2914 const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout(); 2915 uint64_t StrideTypeSizeBits = DL.getTypeSizeInBits(StrideExpr->getType()); 2916 uint64_t BETypeSizeBits = DL.getTypeSizeInBits(BETakenCount->getType()); 2917 const SCEV *CastedStride = StrideExpr; 2918 const SCEV *CastedBECount = BETakenCount; 2919 ScalarEvolution *SE = PSE->getSE(); 2920 if (BETypeSizeBits >= StrideTypeSizeBits) 2921 CastedStride = SE->getNoopOrSignExtend(StrideExpr, BETakenCount->getType()); 2922 else 2923 CastedBECount = SE->getZeroExtendExpr(BETakenCount, StrideExpr->getType()); 2924 const SCEV *StrideMinusBETaken = SE->getMinusSCEV(CastedStride, CastedBECount); 2925 // Since TripCount == BackEdgeTakenCount + 1, checking: 2926 // "Stride >= TripCount" is equivalent to checking: 2927 // Stride - BETakenCount > 0 2928 if (SE->isKnownPositive(StrideMinusBETaken)) { 2929 LLVM_DEBUG( 2930 dbgs() << "LAA: Stride>=TripCount; No point in versioning as the " 2931 "Stride==1 predicate will imply that the loop executes " 2932 "at most once.\n"); 2933 return; 2934 } 2935 LLVM_DEBUG(dbgs() << "LAA: Found a strided access that we can version.\n"); 2936 2937 // Strip back off the integer cast, and check that our result is a 2938 // SCEVUnknown as we expect. 2939 const SCEV *StrideBase = StrideExpr; 2940 if (const auto *C = dyn_cast<SCEVIntegralCastExpr>(StrideBase)) 2941 StrideBase = C->getOperand(); 2942 SymbolicStrides[Ptr] = cast<SCEVUnknown>(StrideBase); 2943 } 2944 2945 LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE, 2946 const TargetLibraryInfo *TLI, AAResults *AA, 2947 DominatorTree *DT, LoopInfo *LI) 2948 : PSE(std::make_unique<PredicatedScalarEvolution>(*SE, *L)), 2949 PtrRtChecking(nullptr), 2950 DepChecker(std::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L) { 2951 PtrRtChecking = std::make_unique<RuntimePointerChecking>(*DepChecker, SE); 2952 if (canAnalyzeLoop()) { 2953 analyzeLoop(AA, LI, TLI, DT); 2954 } 2955 } 2956 2957 void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const { 2958 if (CanVecMem) { 2959 OS.indent(Depth) << "Memory dependences are safe"; 2960 const MemoryDepChecker &DC = getDepChecker(); 2961 if (!DC.isSafeForAnyVectorWidth()) 2962 OS << " with a maximum safe vector width of " 2963 << DC.getMaxSafeVectorWidthInBits() << " bits"; 2964 if (PtrRtChecking->Need) 2965 OS << " with run-time checks"; 2966 OS << "\n"; 2967 } 2968 2969 if (HasConvergentOp) 2970 OS.indent(Depth) << "Has convergent operation in loop\n"; 2971 2972 if (Report) 2973 OS.indent(Depth) << "Report: " << Report->getMsg() << "\n"; 2974 2975 if (auto *Dependences = DepChecker->getDependences()) { 2976 OS.indent(Depth) << "Dependences:\n"; 2977 for (const auto &Dep : *Dependences) { 2978 Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions()); 2979 OS << "\n"; 2980 } 2981 } else 2982 OS.indent(Depth) << "Too many dependences, not recorded\n"; 2983 2984 // List the pair of accesses need run-time checks to prove independence. 2985 PtrRtChecking->print(OS, Depth); 2986 OS << "\n"; 2987 2988 OS.indent(Depth) << "Non vectorizable stores to invariant address were " 2989 << (HasDependenceInvolvingLoopInvariantAddress ? "" : "not ") 2990 << "found in loop.\n"; 2991 2992 OS.indent(Depth) << "SCEV assumptions:\n"; 2993 PSE->getPredicate().print(OS, Depth); 2994 2995 OS << "\n"; 2996 2997 OS.indent(Depth) << "Expressions re-written:\n"; 2998 PSE->print(OS, Depth); 2999 } 3000 3001 const LoopAccessInfo &LoopAccessInfoManager::getInfo(Loop &L) { 3002 auto I = LoopAccessInfoMap.insert({&L, nullptr}); 3003 3004 if (I.second) 3005 I.first->second = 3006 std::make_unique<LoopAccessInfo>(&L, &SE, TLI, &AA, &DT, &LI); 3007 3008 return *I.first->second; 3009 } 3010 3011 bool LoopAccessInfoManager::invalidate( 3012 Function &F, const PreservedAnalyses &PA, 3013 FunctionAnalysisManager::Invalidator &Inv) { 3014 // Check whether our analysis is preserved. 3015 auto PAC = PA.getChecker<LoopAccessAnalysis>(); 3016 if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Function>>()) 3017 // If not, give up now. 3018 return true; 3019 3020 // Check whether the analyses we depend on became invalid for any reason. 3021 // Skip checking TargetLibraryAnalysis as it is immutable and can't become 3022 // invalid. 3023 return Inv.invalidate<AAManager>(F, PA) || 3024 Inv.invalidate<ScalarEvolutionAnalysis>(F, PA) || 3025 Inv.invalidate<LoopAnalysis>(F, PA) || 3026 Inv.invalidate<DominatorTreeAnalysis>(F, PA); 3027 } 3028 3029 LoopAccessInfoManager LoopAccessAnalysis::run(Function &F, 3030 FunctionAnalysisManager &FAM) { 3031 auto &SE = FAM.getResult<ScalarEvolutionAnalysis>(F); 3032 auto &AA = FAM.getResult<AAManager>(F); 3033 auto &DT = FAM.getResult<DominatorTreeAnalysis>(F); 3034 auto &LI = FAM.getResult<LoopAnalysis>(F); 3035 auto &TLI = FAM.getResult<TargetLibraryAnalysis>(F); 3036 return LoopAccessInfoManager(SE, AA, DT, LI, &TLI); 3037 } 3038 3039 AnalysisKey LoopAccessAnalysis::Key; 3040