| File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Analysis/LoopAccessAnalysis.cpp |
| Warning: | line 211, column 15 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- LoopAccessAnalysis.cpp - Loop Access Analysis Implementation --------==// | ||||
| 2 | // | ||||
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | ||||
| 4 | // See https://llvm.org/LICENSE.txt for license information. | ||||
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||||
| 6 | // | ||||
| 7 | //===----------------------------------------------------------------------===// | ||||
| 8 | // | ||||
| 9 | // The implementation for the loop memory dependence that was originally | ||||
| 10 | // developed for the loop vectorizer. | ||||
| 11 | // | ||||
| 12 | //===----------------------------------------------------------------------===// | ||||
| 13 | |||||
| 14 | #include "llvm/Analysis/LoopAccessAnalysis.h" | ||||
| 15 | #include "llvm/ADT/APInt.h" | ||||
| 16 | #include "llvm/ADT/DenseMap.h" | ||||
| 17 | #include "llvm/ADT/DepthFirstIterator.h" | ||||
| 18 | #include "llvm/ADT/EquivalenceClasses.h" | ||||
| 19 | #include "llvm/ADT/PointerIntPair.h" | ||||
| 20 | #include "llvm/ADT/STLExtras.h" | ||||
| 21 | #include "llvm/ADT/SetVector.h" | ||||
| 22 | #include "llvm/ADT/SmallPtrSet.h" | ||||
| 23 | #include "llvm/ADT/SmallSet.h" | ||||
| 24 | #include "llvm/ADT/SmallVector.h" | ||||
| 25 | #include "llvm/ADT/iterator_range.h" | ||||
| 26 | #include "llvm/Analysis/AliasAnalysis.h" | ||||
| 27 | #include "llvm/Analysis/AliasSetTracker.h" | ||||
| 28 | #include "llvm/Analysis/LoopAnalysisManager.h" | ||||
| 29 | #include "llvm/Analysis/LoopInfo.h" | ||||
| 30 | #include "llvm/Analysis/MemoryLocation.h" | ||||
| 31 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||
| 32 | #include "llvm/Analysis/ScalarEvolution.h" | ||||
| 33 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | ||||
| 34 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||
| 35 | #include "llvm/Analysis/ValueTracking.h" | ||||
| 36 | #include "llvm/Analysis/VectorUtils.h" | ||||
| 37 | #include "llvm/IR/BasicBlock.h" | ||||
| 38 | #include "llvm/IR/Constants.h" | ||||
| 39 | #include "llvm/IR/DataLayout.h" | ||||
| 40 | #include "llvm/IR/DebugLoc.h" | ||||
| 41 | #include "llvm/IR/DerivedTypes.h" | ||||
| 42 | #include "llvm/IR/DiagnosticInfo.h" | ||||
| 43 | #include "llvm/IR/Dominators.h" | ||||
| 44 | #include "llvm/IR/Function.h" | ||||
| 45 | #include "llvm/IR/InstrTypes.h" | ||||
| 46 | #include "llvm/IR/Instruction.h" | ||||
| 47 | #include "llvm/IR/Instructions.h" | ||||
| 48 | #include "llvm/IR/Operator.h" | ||||
| 49 | #include "llvm/IR/PassManager.h" | ||||
| 50 | #include "llvm/IR/Type.h" | ||||
| 51 | #include "llvm/IR/Value.h" | ||||
| 52 | #include "llvm/IR/ValueHandle.h" | ||||
| 53 | #include "llvm/InitializePasses.h" | ||||
| 54 | #include "llvm/Pass.h" | ||||
| 55 | #include "llvm/Support/Casting.h" | ||||
| 56 | #include "llvm/Support/CommandLine.h" | ||||
| 57 | #include "llvm/Support/Debug.h" | ||||
| 58 | #include "llvm/Support/ErrorHandling.h" | ||||
| 59 | #include "llvm/Support/raw_ostream.h" | ||||
| 60 | #include <algorithm> | ||||
| 61 | #include <cassert> | ||||
| 62 | #include <cstdint> | ||||
| 63 | #include <cstdlib> | ||||
| 64 | #include <iterator> | ||||
| 65 | #include <utility> | ||||
| 66 | #include <vector> | ||||
| 67 | |||||
| 68 | using namespace llvm; | ||||
| 69 | |||||
| 70 | #define DEBUG_TYPE"loop-accesses" "loop-accesses" | ||||
| 71 | |||||
| 72 | static cl::opt<unsigned, true> | ||||
| 73 | VectorizationFactor("force-vector-width", cl::Hidden, | ||||
| 74 | cl::desc("Sets the SIMD width. Zero is autoselect."), | ||||
| 75 | cl::location(VectorizerParams::VectorizationFactor)); | ||||
| 76 | unsigned VectorizerParams::VectorizationFactor; | ||||
| 77 | |||||
| 78 | static cl::opt<unsigned, true> | ||||
| 79 | VectorizationInterleave("force-vector-interleave", cl::Hidden, | ||||
| 80 | cl::desc("Sets the vectorization interleave count. " | ||||
| 81 | "Zero is autoselect."), | ||||
| 82 | cl::location( | ||||
| 83 | VectorizerParams::VectorizationInterleave)); | ||||
| 84 | unsigned VectorizerParams::VectorizationInterleave; | ||||
| 85 | |||||
| 86 | static cl::opt<unsigned, true> RuntimeMemoryCheckThreshold( | ||||
| 87 | "runtime-memory-check-threshold", cl::Hidden, | ||||
| 88 | cl::desc("When performing memory disambiguation checks at runtime do not " | ||||
| 89 | "generate more than this number of comparisons (default = 8)."), | ||||
| 90 | cl::location(VectorizerParams::RuntimeMemoryCheckThreshold), cl::init(8)); | ||||
| 91 | unsigned VectorizerParams::RuntimeMemoryCheckThreshold; | ||||
| 92 | |||||
| 93 | /// The maximum iterations used to merge memory checks | ||||
| 94 | static cl::opt<unsigned> MemoryCheckMergeThreshold( | ||||
| 95 | "memory-check-merge-threshold", cl::Hidden, | ||||
| 96 | cl::desc("Maximum number of comparisons done when trying to merge " | ||||
| 97 | "runtime memory checks. (default = 100)"), | ||||
| 98 | cl::init(100)); | ||||
| 99 | |||||
| 100 | /// Maximum SIMD width. | ||||
| 101 | const unsigned VectorizerParams::MaxVectorWidth = 64; | ||||
| 102 | |||||
| 103 | /// We collect dependences up to this threshold. | ||||
| 104 | static cl::opt<unsigned> | ||||
| 105 | MaxDependences("max-dependences", cl::Hidden, | ||||
| 106 | cl::desc("Maximum number of dependences collected by " | ||||
| 107 | "loop-access analysis (default = 100)"), | ||||
| 108 | cl::init(100)); | ||||
| 109 | |||||
| 110 | /// This enables versioning on the strides of symbolically striding memory | ||||
| 111 | /// accesses in code like the following. | ||||
| 112 | /// for (i = 0; i < N; ++i) | ||||
| 113 | /// A[i * Stride1] += B[i * Stride2] ... | ||||
| 114 | /// | ||||
| 115 | /// Will be roughly translated to | ||||
| 116 | /// if (Stride1 == 1 && Stride2 == 1) { | ||||
| 117 | /// for (i = 0; i < N; i+=4) | ||||
| 118 | /// A[i:i+3] += ... | ||||
| 119 | /// } else | ||||
| 120 | /// ... | ||||
| 121 | static cl::opt<bool> EnableMemAccessVersioning( | ||||
| 122 | "enable-mem-access-versioning", cl::init(true), cl::Hidden, | ||||
| 123 | cl::desc("Enable symbolic stride memory access versioning")); | ||||
| 124 | |||||
| 125 | /// Enable store-to-load forwarding conflict detection. This option can | ||||
| 126 | /// be disabled for correctness testing. | ||||
| 127 | static cl::opt<bool> EnableForwardingConflictDetection( | ||||
| 128 | "store-to-load-forwarding-conflict-detection", cl::Hidden, | ||||
| 129 | cl::desc("Enable conflict detection in loop-access analysis"), | ||||
| 130 | cl::init(true)); | ||||
| 131 | |||||
| 132 | bool VectorizerParams::isInterleaveForced() { | ||||
| 133 | return ::VectorizationInterleave.getNumOccurrences() > 0; | ||||
| 134 | } | ||||
| 135 | |||||
| 136 | Value *llvm::stripIntegerCast(Value *V) { | ||||
| 137 | if (auto *CI = dyn_cast<CastInst>(V)) | ||||
| 138 | if (CI->getOperand(0)->getType()->isIntegerTy()) | ||||
| 139 | return CI->getOperand(0); | ||||
| 140 | return V; | ||||
| 141 | } | ||||
| 142 | |||||
| 143 | const SCEV *llvm::replaceSymbolicStrideSCEV(PredicatedScalarEvolution &PSE, | ||||
| 144 | const ValueToValueMap &PtrToStride, | ||||
| 145 | Value *Ptr, Value *OrigPtr) { | ||||
| 146 | const SCEV *OrigSCEV = PSE.getSCEV(Ptr); | ||||
| 147 | |||||
| 148 | // If there is an entry in the map return the SCEV of the pointer with the | ||||
| 149 | // symbolic stride replaced by one. | ||||
| 150 | ValueToValueMap::const_iterator SI = | ||||
| 151 | PtrToStride.find(OrigPtr ? OrigPtr : Ptr); | ||||
| 152 | if (SI == PtrToStride.end()) | ||||
| 153 | // For a non-symbolic stride, just return the original expression. | ||||
| 154 | return OrigSCEV; | ||||
| 155 | |||||
| 156 | Value *StrideVal = stripIntegerCast(SI->second); | ||||
| 157 | |||||
| 158 | ScalarEvolution *SE = PSE.getSE(); | ||||
| 159 | const auto *U = cast<SCEVUnknown>(SE->getSCEV(StrideVal)); | ||||
| 160 | const auto *CT = | ||||
| 161 | static_cast<const SCEVConstant *>(SE->getOne(StrideVal->getType())); | ||||
| 162 | |||||
| 163 | PSE.addPredicate(*SE->getEqualPredicate(U, CT)); | ||||
| 164 | auto *Expr = PSE.getSCEV(Ptr); | ||||
| 165 | |||||
| 166 | LLVM_DEBUG(dbgs() << "LAA: Replacing SCEV: " << *OrigSCEVdo { } while (false) | ||||
| 167 | << " by: " << *Expr << "\n")do { } while (false); | ||||
| 168 | return Expr; | ||||
| 169 | } | ||||
| 170 | |||||
| 171 | RuntimeCheckingPtrGroup::RuntimeCheckingPtrGroup( | ||||
| 172 | unsigned Index, RuntimePointerChecking &RtCheck) | ||||
| 173 | : High(RtCheck.Pointers[Index].End), Low(RtCheck.Pointers[Index].Start), | ||||
| 174 | AddressSpace(RtCheck.Pointers[Index] | ||||
| 175 | .PointerValue->getType() | ||||
| 176 | ->getPointerAddressSpace()) { | ||||
| 177 | Members.push_back(Index); | ||||
| 178 | } | ||||
| 179 | |||||
| 180 | /// Calculate Start and End points of memory access. | ||||
| 181 | /// Let's assume A is the first access and B is a memory access on N-th loop | ||||
| 182 | /// iteration. Then B is calculated as: | ||||
| 183 | /// B = A + Step*N . | ||||
| 184 | /// Step value may be positive or negative. | ||||
| 185 | /// N is a calculated back-edge taken count: | ||||
| 186 | /// N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0 | ||||
| 187 | /// Start and End points are calculated in the following way: | ||||
| 188 | /// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt, | ||||
| 189 | /// where SizeOfElt is the size of single memory access in bytes. | ||||
| 190 | /// | ||||
| 191 | /// There is no conflict when the intervals are disjoint: | ||||
| 192 | /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End) | ||||
| 193 | void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr, | ||||
| 194 | unsigned DepSetId, unsigned ASId, | ||||
| 195 | const ValueToValueMap &Strides, | ||||
| 196 | PredicatedScalarEvolution &PSE) { | ||||
| 197 | // Get the stride replaced scev. | ||||
| 198 | const SCEV *Sc = replaceSymbolicStrideSCEV(PSE, Strides, Ptr); | ||||
| 199 | ScalarEvolution *SE = PSE.getSE(); | ||||
| 200 | |||||
| 201 | const SCEV *ScStart; | ||||
| 202 | const SCEV *ScEnd; | ||||
| 203 | |||||
| 204 | if (SE->isLoopInvariant(Sc, Lp)) { | ||||
| 205 | ScStart = ScEnd = Sc; | ||||
| 206 | } else { | ||||
| 207 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc); | ||||
| 208 | assert(AR && "Invalid addrec expression")((void)0); | ||||
| 209 | const SCEV *Ex = PSE.getBackedgeTakenCount(); | ||||
| 210 | |||||
| 211 | ScStart = AR->getStart(); | ||||
| |||||
| 212 | ScEnd = AR->evaluateAtIteration(Ex, *SE); | ||||
| 213 | const SCEV *Step = AR->getStepRecurrence(*SE); | ||||
| 214 | |||||
| 215 | // For expressions with negative step, the upper bound is ScStart and the | ||||
| 216 | // lower bound is ScEnd. | ||||
| 217 | if (const auto *CStep = dyn_cast<SCEVConstant>(Step)) { | ||||
| 218 | if (CStep->getValue()->isNegative()) | ||||
| 219 | std::swap(ScStart, ScEnd); | ||||
| 220 | } else { | ||||
| 221 | // Fallback case: the step is not constant, but we can still | ||||
| 222 | // get the upper and lower bounds of the interval by using min/max | ||||
| 223 | // expressions. | ||||
| 224 | ScStart = SE->getUMinExpr(ScStart, ScEnd); | ||||
| 225 | ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd); | ||||
| 226 | } | ||||
| 227 | } | ||||
| 228 | // Add the size of the pointed element to ScEnd. | ||||
| 229 | auto &DL = Lp->getHeader()->getModule()->getDataLayout(); | ||||
| 230 | Type *IdxTy = DL.getIndexType(Ptr->getType()); | ||||
| 231 | const SCEV *EltSizeSCEV = | ||||
| 232 | SE->getStoreSizeOfExpr(IdxTy, Ptr->getType()->getPointerElementType()); | ||||
| 233 | ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV); | ||||
| 234 | |||||
| 235 | Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc); | ||||
| 236 | } | ||||
| 237 | |||||
| 238 | SmallVector<RuntimePointerCheck, 4> | ||||
| 239 | RuntimePointerChecking::generateChecks() const { | ||||
| 240 | SmallVector<RuntimePointerCheck, 4> Checks; | ||||
| 241 | |||||
| 242 | for (unsigned I = 0; I < CheckingGroups.size(); ++I) { | ||||
| 243 | for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) { | ||||
| 244 | const RuntimeCheckingPtrGroup &CGI = CheckingGroups[I]; | ||||
| 245 | const RuntimeCheckingPtrGroup &CGJ = CheckingGroups[J]; | ||||
| 246 | |||||
| 247 | if (needsChecking(CGI, CGJ)) | ||||
| 248 | Checks.push_back(std::make_pair(&CGI, &CGJ)); | ||||
| 249 | } | ||||
| 250 | } | ||||
| 251 | return Checks; | ||||
| 252 | } | ||||
| 253 | |||||
| 254 | void RuntimePointerChecking::generateChecks( | ||||
| 255 | MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) { | ||||
| 256 | assert(Checks.empty() && "Checks is not empty")((void)0); | ||||
| 257 | groupChecks(DepCands, UseDependencies); | ||||
| 258 | Checks = generateChecks(); | ||||
| 259 | } | ||||
| 260 | |||||
| 261 | bool RuntimePointerChecking::needsChecking( | ||||
| 262 | const RuntimeCheckingPtrGroup &M, const RuntimeCheckingPtrGroup &N) const { | ||||
| 263 | for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I) | ||||
| 264 | for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J) | ||||
| 265 | if (needsChecking(M.Members[I], N.Members[J])) | ||||
| 266 | return true; | ||||
| 267 | return false; | ||||
| 268 | } | ||||
| 269 | |||||
| 270 | /// Compare \p I and \p J and return the minimum. | ||||
| 271 | /// Return nullptr in case we couldn't find an answer. | ||||
| 272 | static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J, | ||||
| 273 | ScalarEvolution *SE) { | ||||
| 274 | const SCEV *Diff = SE->getMinusSCEV(J, I); | ||||
| 275 | const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff); | ||||
| 276 | |||||
| 277 | if (!C) | ||||
| 278 | return nullptr; | ||||
| 279 | if (C->getValue()->isNegative()) | ||||
| 280 | return J; | ||||
| 281 | return I; | ||||
| 282 | } | ||||
| 283 | |||||
| 284 | bool RuntimeCheckingPtrGroup::addPointer(unsigned Index, | ||||
| 285 | RuntimePointerChecking &RtCheck) { | ||||
| 286 | return addPointer( | ||||
| 287 | Index, RtCheck.Pointers[Index].Start, RtCheck.Pointers[Index].End, | ||||
| 288 | RtCheck.Pointers[Index].PointerValue->getType()->getPointerAddressSpace(), | ||||
| 289 | *RtCheck.SE); | ||||
| 290 | } | ||||
| 291 | |||||
| 292 | bool RuntimeCheckingPtrGroup::addPointer(unsigned Index, const SCEV *Start, | ||||
| 293 | const SCEV *End, unsigned AS, | ||||
| 294 | ScalarEvolution &SE) { | ||||
| 295 | assert(AddressSpace == AS &&((void)0) | ||||
| 296 | "all pointers in a checking group must be in the same address space")((void)0); | ||||
| 297 | |||||
| 298 | // Compare the starts and ends with the known minimum and maximum | ||||
| 299 | // of this set. We need to know how we compare against the min/max | ||||
| 300 | // of the set in order to be able to emit memchecks. | ||||
| 301 | const SCEV *Min0 = getMinFromExprs(Start, Low, &SE); | ||||
| 302 | if (!Min0) | ||||
| 303 | return false; | ||||
| 304 | |||||
| 305 | const SCEV *Min1 = getMinFromExprs(End, High, &SE); | ||||
| 306 | if (!Min1) | ||||
| 307 | return false; | ||||
| 308 | |||||
| 309 | // Update the low bound expression if we've found a new min value. | ||||
| 310 | if (Min0 == Start) | ||||
| 311 | Low = Start; | ||||
| 312 | |||||
| 313 | // Update the high bound expression if we've found a new max value. | ||||
| 314 | if (Min1 != End) | ||||
| 315 | High = End; | ||||
| 316 | |||||
| 317 | Members.push_back(Index); | ||||
| 318 | return true; | ||||
| 319 | } | ||||
| 320 | |||||
| 321 | void RuntimePointerChecking::groupChecks( | ||||
| 322 | MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) { | ||||
| 323 | // We build the groups from dependency candidates equivalence classes | ||||
| 324 | // because: | ||||
| 325 | // - We know that pointers in the same equivalence class share | ||||
| 326 | // the same underlying object and therefore there is a chance | ||||
| 327 | // that we can compare pointers | ||||
| 328 | // - We wouldn't be able to merge two pointers for which we need | ||||
| 329 | // to emit a memcheck. The classes in DepCands are already | ||||
| 330 | // conveniently built such that no two pointers in the same | ||||
| 331 | // class need checking against each other. | ||||
| 332 | |||||
| 333 | // We use the following (greedy) algorithm to construct the groups | ||||
| 334 | // For every pointer in the equivalence class: | ||||
| 335 | // For each existing group: | ||||
| 336 | // - if the difference between this pointer and the min/max bounds | ||||
| 337 | // of the group is a constant, then make the pointer part of the | ||||
| 338 | // group and update the min/max bounds of that group as required. | ||||
| 339 | |||||
| 340 | CheckingGroups.clear(); | ||||
| 341 | |||||
| 342 | // If we need to check two pointers to the same underlying object | ||||
| 343 | // with a non-constant difference, we shouldn't perform any pointer | ||||
| 344 | // grouping with those pointers. This is because we can easily get | ||||
| 345 | // into cases where the resulting check would return false, even when | ||||
| 346 | // the accesses are safe. | ||||
| 347 | // | ||||
| 348 | // The following example shows this: | ||||
| 349 | // for (i = 0; i < 1000; ++i) | ||||
| 350 | // a[5000 + i * m] = a[i] + a[i + 9000] | ||||
| 351 | // | ||||
| 352 | // Here grouping gives a check of (5000, 5000 + 1000 * m) against | ||||
| 353 | // (0, 10000) which is always false. However, if m is 1, there is no | ||||
| 354 | // dependence. Not grouping the checks for a[i] and a[i + 9000] allows | ||||
| 355 | // us to perform an accurate check in this case. | ||||
| 356 | // | ||||
| 357 | // The above case requires that we have an UnknownDependence between | ||||
| 358 | // accesses to the same underlying object. This cannot happen unless | ||||
| 359 | // FoundNonConstantDistanceDependence is set, and therefore UseDependencies | ||||
| 360 | // is also false. In this case we will use the fallback path and create | ||||
| 361 | // separate checking groups for all pointers. | ||||
| 362 | |||||
| 363 | // If we don't have the dependency partitions, construct a new | ||||
| 364 | // checking pointer group for each pointer. This is also required | ||||
| 365 | // for correctness, because in this case we can have checking between | ||||
| 366 | // pointers to the same underlying object. | ||||
| 367 | if (!UseDependencies) { | ||||
| 368 | for (unsigned I = 0; I < Pointers.size(); ++I) | ||||
| 369 | CheckingGroups.push_back(RuntimeCheckingPtrGroup(I, *this)); | ||||
| 370 | return; | ||||
| 371 | } | ||||
| 372 | |||||
| 373 | unsigned TotalComparisons = 0; | ||||
| 374 | |||||
| 375 | DenseMap<Value *, unsigned> PositionMap; | ||||
| 376 | for (unsigned Index = 0; Index < Pointers.size(); ++Index) | ||||
| 377 | PositionMap[Pointers[Index].PointerValue] = Index; | ||||
| 378 | |||||
| 379 | // We need to keep track of what pointers we've already seen so we | ||||
| 380 | // don't process them twice. | ||||
| 381 | SmallSet<unsigned, 2> Seen; | ||||
| 382 | |||||
| 383 | // Go through all equivalence classes, get the "pointer check groups" | ||||
| 384 | // and add them to the overall solution. We use the order in which accesses | ||||
| 385 | // appear in 'Pointers' to enforce determinism. | ||||
| 386 | for (unsigned I = 0; I < Pointers.size(); ++I) { | ||||
| 387 | // We've seen this pointer before, and therefore already processed | ||||
| 388 | // its equivalence class. | ||||
| 389 | if (Seen.count(I)) | ||||
| 390 | continue; | ||||
| 391 | |||||
| 392 | MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue, | ||||
| 393 | Pointers[I].IsWritePtr); | ||||
| 394 | |||||
| 395 | SmallVector<RuntimeCheckingPtrGroup, 2> Groups; | ||||
| 396 | auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access)); | ||||
| 397 | |||||
| 398 | // Because DepCands is constructed by visiting accesses in the order in | ||||
| 399 | // which they appear in alias sets (which is deterministic) and the | ||||
| 400 | // iteration order within an equivalence class member is only dependent on | ||||
| 401 | // the order in which unions and insertions are performed on the | ||||
| 402 | // equivalence class, the iteration order is deterministic. | ||||
| 403 | for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end(); | ||||
| 404 | MI != ME; ++MI) { | ||||
| 405 | auto PointerI = PositionMap.find(MI->getPointer()); | ||||
| 406 | assert(PointerI != PositionMap.end() &&((void)0) | ||||
| 407 | "pointer in equivalence class not found in PositionMap")((void)0); | ||||
| 408 | unsigned Pointer = PointerI->second; | ||||
| 409 | bool Merged = false; | ||||
| 410 | // Mark this pointer as seen. | ||||
| 411 | Seen.insert(Pointer); | ||||
| 412 | |||||
| 413 | // Go through all the existing sets and see if we can find one | ||||
| 414 | // which can include this pointer. | ||||
| 415 | for (RuntimeCheckingPtrGroup &Group : Groups) { | ||||
| 416 | // Don't perform more than a certain amount of comparisons. | ||||
| 417 | // This should limit the cost of grouping the pointers to something | ||||
| 418 | // reasonable. If we do end up hitting this threshold, the algorithm | ||||
| 419 | // will create separate groups for all remaining pointers. | ||||
| 420 | if (TotalComparisons > MemoryCheckMergeThreshold) | ||||
| 421 | break; | ||||
| 422 | |||||
| 423 | TotalComparisons++; | ||||
| 424 | |||||
| 425 | if (Group.addPointer(Pointer, *this)) { | ||||
| 426 | Merged = true; | ||||
| 427 | break; | ||||
| 428 | } | ||||
| 429 | } | ||||
| 430 | |||||
| 431 | if (!Merged) | ||||
| 432 | // We couldn't add this pointer to any existing set or the threshold | ||||
| 433 | // for the number of comparisons has been reached. Create a new group | ||||
| 434 | // to hold the current pointer. | ||||
| 435 | Groups.push_back(RuntimeCheckingPtrGroup(Pointer, *this)); | ||||
| 436 | } | ||||
| 437 | |||||
| 438 | // We've computed the grouped checks for this partition. | ||||
| 439 | // Save the results and continue with the next one. | ||||
| 440 | llvm::copy(Groups, std::back_inserter(CheckingGroups)); | ||||
| 441 | } | ||||
| 442 | } | ||||
| 443 | |||||
| 444 | bool RuntimePointerChecking::arePointersInSamePartition( | ||||
| 445 | const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1, | ||||
| 446 | unsigned PtrIdx2) { | ||||
| 447 | return (PtrToPartition[PtrIdx1] != -1 && | ||||
| 448 | PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]); | ||||
| 449 | } | ||||
| 450 | |||||
| 451 | bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const { | ||||
| 452 | const PointerInfo &PointerI = Pointers[I]; | ||||
| 453 | const PointerInfo &PointerJ = Pointers[J]; | ||||
| 454 | |||||
| 455 | // No need to check if two readonly pointers intersect. | ||||
| 456 | if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr) | ||||
| 457 | return false; | ||||
| 458 | |||||
| 459 | // Only need to check pointers between two different dependency sets. | ||||
| 460 | if (PointerI.DependencySetId == PointerJ.DependencySetId) | ||||
| 461 | return false; | ||||
| 462 | |||||
| 463 | // Only need to check pointers in the same alias set. | ||||
| 464 | if (PointerI.AliasSetId != PointerJ.AliasSetId) | ||||
| 465 | return false; | ||||
| 466 | |||||
| 467 | return true; | ||||
| 468 | } | ||||
| 469 | |||||
| 470 | void RuntimePointerChecking::printChecks( | ||||
| 471 | raw_ostream &OS, const SmallVectorImpl<RuntimePointerCheck> &Checks, | ||||
| 472 | unsigned Depth) const { | ||||
| 473 | unsigned N = 0; | ||||
| 474 | for (const auto &Check : Checks) { | ||||
| 475 | const auto &First = Check.first->Members, &Second = Check.second->Members; | ||||
| 476 | |||||
| 477 | OS.indent(Depth) << "Check " << N++ << ":\n"; | ||||
| 478 | |||||
| 479 | OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n"; | ||||
| 480 | for (unsigned K = 0; K < First.size(); ++K) | ||||
| 481 | OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n"; | ||||
| 482 | |||||
| 483 | OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n"; | ||||
| 484 | for (unsigned K = 0; K < Second.size(); ++K) | ||||
| 485 | OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n"; | ||||
| 486 | } | ||||
| 487 | } | ||||
| 488 | |||||
| 489 | void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const { | ||||
| 490 | |||||
| 491 | OS.indent(Depth) << "Run-time memory checks:\n"; | ||||
| 492 | printChecks(OS, Checks, Depth); | ||||
| 493 | |||||
| 494 | OS.indent(Depth) << "Grouped accesses:\n"; | ||||
| 495 | for (unsigned I = 0; I < CheckingGroups.size(); ++I) { | ||||
| 496 | const auto &CG = CheckingGroups[I]; | ||||
| 497 | |||||
| 498 | OS.indent(Depth + 2) << "Group " << &CG << ":\n"; | ||||
| 499 | OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High | ||||
| 500 | << ")\n"; | ||||
| 501 | for (unsigned J = 0; J < CG.Members.size(); ++J) { | ||||
| 502 | OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr | ||||
| 503 | << "\n"; | ||||
| 504 | } | ||||
| 505 | } | ||||
| 506 | } | ||||
| 507 | |||||
| 508 | namespace { | ||||
| 509 | |||||
| 510 | /// Analyses memory accesses in a loop. | ||||
| 511 | /// | ||||
| 512 | /// Checks whether run time pointer checks are needed and builds sets for data | ||||
| 513 | /// dependence checking. | ||||
| 514 | class AccessAnalysis { | ||||
| 515 | public: | ||||
| 516 | /// Read or write access location. | ||||
| 517 | typedef PointerIntPair<Value *, 1, bool> MemAccessInfo; | ||||
| 518 | typedef SmallVector<MemAccessInfo, 8> MemAccessInfoList; | ||||
| 519 | |||||
| 520 | AccessAnalysis(Loop *TheLoop, AAResults *AA, LoopInfo *LI, | ||||
| 521 | MemoryDepChecker::DepCandidates &DA, | ||||
| 522 | PredicatedScalarEvolution &PSE) | ||||
| 523 | : TheLoop(TheLoop), AST(*AA), LI(LI), DepCands(DA), | ||||
| 524 | IsRTCheckAnalysisNeeded(false), PSE(PSE) {} | ||||
| 525 | |||||
| 526 | /// Register a load and whether it is only read from. | ||||
| 527 | void addLoad(MemoryLocation &Loc, bool IsReadOnly) { | ||||
| 528 | Value *Ptr = const_cast<Value*>(Loc.Ptr); | ||||
| 529 | AST.add(Ptr, LocationSize::beforeOrAfterPointer(), Loc.AATags); | ||||
| 530 | Accesses.insert(MemAccessInfo(Ptr, false)); | ||||
| 531 | if (IsReadOnly) | ||||
| 532 | ReadOnlyPtr.insert(Ptr); | ||||
| 533 | } | ||||
| 534 | |||||
| 535 | /// Register a store. | ||||
| 536 | void addStore(MemoryLocation &Loc) { | ||||
| 537 | Value *Ptr = const_cast<Value*>(Loc.Ptr); | ||||
| 538 | AST.add(Ptr, LocationSize::beforeOrAfterPointer(), Loc.AATags); | ||||
| 539 | Accesses.insert(MemAccessInfo(Ptr, true)); | ||||
| 540 | } | ||||
| 541 | |||||
| 542 | /// Check if we can emit a run-time no-alias check for \p Access. | ||||
| 543 | /// | ||||
| 544 | /// Returns true if we can emit a run-time no alias check for \p Access. | ||||
| 545 | /// If we can check this access, this also adds it to a dependence set and | ||||
| 546 | /// adds a run-time to check for it to \p RtCheck. If \p Assume is true, | ||||
| 547 | /// we will attempt to use additional run-time checks in order to get | ||||
| 548 | /// the bounds of the pointer. | ||||
| 549 | bool createCheckForAccess(RuntimePointerChecking &RtCheck, | ||||
| 550 | MemAccessInfo Access, | ||||
| 551 | const ValueToValueMap &Strides, | ||||
| 552 | DenseMap<Value *, unsigned> &DepSetId, | ||||
| 553 | Loop *TheLoop, unsigned &RunningDepId, | ||||
| 554 | unsigned ASId, bool ShouldCheckStride, | ||||
| 555 | bool Assume); | ||||
| 556 | |||||
| 557 | /// Check whether we can check the pointers at runtime for | ||||
| 558 | /// non-intersection. | ||||
| 559 | /// | ||||
| 560 | /// Returns true if we need no check or if we do and we can generate them | ||||
| 561 | /// (i.e. the pointers have computable bounds). | ||||
| 562 | bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE, | ||||
| 563 | Loop *TheLoop, const ValueToValueMap &Strides, | ||||
| 564 | bool ShouldCheckWrap = false); | ||||
| 565 | |||||
| 566 | /// Goes over all memory accesses, checks whether a RT check is needed | ||||
| 567 | /// and builds sets of dependent accesses. | ||||
| 568 | void buildDependenceSets() { | ||||
| 569 | processMemAccesses(); | ||||
| 570 | } | ||||
| 571 | |||||
| 572 | /// Initial processing of memory accesses determined that we need to | ||||
| 573 | /// perform dependency checking. | ||||
| 574 | /// | ||||
| 575 | /// Note that this can later be cleared if we retry memcheck analysis without | ||||
| 576 | /// dependency checking (i.e. FoundNonConstantDistanceDependence). | ||||
| 577 | bool isDependencyCheckNeeded() { return !CheckDeps.empty(); } | ||||
| 578 | |||||
| 579 | /// We decided that no dependence analysis would be used. Reset the state. | ||||
| 580 | void resetDepChecks(MemoryDepChecker &DepChecker) { | ||||
| 581 | CheckDeps.clear(); | ||||
| 582 | DepChecker.clearDependences(); | ||||
| 583 | } | ||||
| 584 | |||||
| 585 | MemAccessInfoList &getDependenciesToCheck() { return CheckDeps; } | ||||
| 586 | |||||
| 587 | private: | ||||
| 588 | typedef SetVector<MemAccessInfo> PtrAccessSet; | ||||
| 589 | |||||
| 590 | /// Go over all memory access and check whether runtime pointer checks | ||||
| 591 | /// are needed and build sets of dependency check candidates. | ||||
| 592 | void processMemAccesses(); | ||||
| 593 | |||||
| 594 | /// Set of all accesses. | ||||
| 595 | PtrAccessSet Accesses; | ||||
| 596 | |||||
| 597 | /// The loop being checked. | ||||
| 598 | const Loop *TheLoop; | ||||
| 599 | |||||
| 600 | /// List of accesses that need a further dependence check. | ||||
| 601 | MemAccessInfoList CheckDeps; | ||||
| 602 | |||||
| 603 | /// Set of pointers that are read only. | ||||
| 604 | SmallPtrSet<Value*, 16> ReadOnlyPtr; | ||||
| 605 | |||||
| 606 | /// An alias set tracker to partition the access set by underlying object and | ||||
| 607 | //intrinsic property (such as TBAA metadata). | ||||
| 608 | AliasSetTracker AST; | ||||
| 609 | |||||
| 610 | LoopInfo *LI; | ||||
| 611 | |||||
| 612 | /// Sets of potentially dependent accesses - members of one set share an | ||||
| 613 | /// underlying pointer. The set "CheckDeps" identfies which sets really need a | ||||
| 614 | /// dependence check. | ||||
| 615 | MemoryDepChecker::DepCandidates &DepCands; | ||||
| 616 | |||||
| 617 | /// Initial processing of memory accesses determined that we may need | ||||
| 618 | /// to add memchecks. Perform the analysis to determine the necessary checks. | ||||
| 619 | /// | ||||
| 620 | /// Note that, this is different from isDependencyCheckNeeded. When we retry | ||||
| 621 | /// memcheck analysis without dependency checking | ||||
| 622 | /// (i.e. FoundNonConstantDistanceDependence), isDependencyCheckNeeded is | ||||
| 623 | /// cleared while this remains set if we have potentially dependent accesses. | ||||
| 624 | bool IsRTCheckAnalysisNeeded; | ||||
| 625 | |||||
| 626 | /// The SCEV predicate containing all the SCEV-related assumptions. | ||||
| 627 | PredicatedScalarEvolution &PSE; | ||||
| 628 | }; | ||||
| 629 | |||||
| 630 | } // end anonymous namespace | ||||
| 631 | |||||
| 632 | /// Check whether a pointer can participate in a runtime bounds check. | ||||
| 633 | /// If \p Assume, try harder to prove that we can compute the bounds of \p Ptr | ||||
| 634 | /// by adding run-time checks (overflow checks) if necessary. | ||||
| 635 | static bool hasComputableBounds(PredicatedScalarEvolution &PSE, | ||||
| 636 | const ValueToValueMap &Strides, Value *Ptr, | ||||
| 637 | Loop *L, bool Assume) { | ||||
| 638 | const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr); | ||||
| 639 | |||||
| 640 | // The bounds for loop-invariant pointer is trivial. | ||||
| 641 | if (PSE.getSE()->isLoopInvariant(PtrScev, L)) | ||||
| 642 | return true; | ||||
| 643 | |||||
| 644 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); | ||||
| 645 | |||||
| 646 | if (!AR && Assume) | ||||
| 647 | AR = PSE.getAsAddRec(Ptr); | ||||
| 648 | |||||
| 649 | if (!AR) | ||||
| 650 | return false; | ||||
| 651 | |||||
| 652 | return AR->isAffine(); | ||||
| 653 | } | ||||
| 654 | |||||
| 655 | /// Check whether a pointer address cannot wrap. | ||||
| 656 | static bool isNoWrap(PredicatedScalarEvolution &PSE, | ||||
| 657 | const ValueToValueMap &Strides, Value *Ptr, Loop *L) { | ||||
| 658 | const SCEV *PtrScev = PSE.getSCEV(Ptr); | ||||
| 659 | if (PSE.getSE()->isLoopInvariant(PtrScev, L)) | ||||
| 660 | return true; | ||||
| 661 | |||||
| 662 | int64_t Stride = getPtrStride(PSE, Ptr, L, Strides); | ||||
| 663 | if (Stride == 1 || PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW)) | ||||
| 664 | return true; | ||||
| 665 | |||||
| 666 | return false; | ||||
| 667 | } | ||||
| 668 | |||||
| 669 | bool AccessAnalysis::createCheckForAccess(RuntimePointerChecking &RtCheck, | ||||
| 670 | MemAccessInfo Access, | ||||
| 671 | const ValueToValueMap &StridesMap, | ||||
| 672 | DenseMap<Value *, unsigned> &DepSetId, | ||||
| 673 | Loop *TheLoop, unsigned &RunningDepId, | ||||
| 674 | unsigned ASId, bool ShouldCheckWrap, | ||||
| 675 | bool Assume) { | ||||
| 676 | Value *Ptr = Access.getPointer(); | ||||
| 677 | |||||
| 678 | if (!hasComputableBounds(PSE, StridesMap, Ptr, TheLoop, Assume)) | ||||
| 679 | return false; | ||||
| 680 | |||||
| 681 | // When we run after a failing dependency check we have to make sure | ||||
| 682 | // we don't have wrapping pointers. | ||||
| 683 | if (ShouldCheckWrap
| ||||
| 684 | auto *Expr = PSE.getSCEV(Ptr); | ||||
| 685 | if (!Assume || !isa<SCEVAddRecExpr>(Expr)) | ||||
| 686 | return false; | ||||
| 687 | PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); | ||||
| 688 | } | ||||
| 689 | |||||
| 690 | // The id of the dependence set. | ||||
| 691 | unsigned DepId; | ||||
| 692 | |||||
| 693 | if (isDependencyCheckNeeded()) { | ||||
| 694 | Value *Leader = DepCands.getLeaderValue(Access).getPointer(); | ||||
| 695 | unsigned &LeaderId = DepSetId[Leader]; | ||||
| 696 | if (!LeaderId) | ||||
| 697 | LeaderId = RunningDepId++; | ||||
| 698 | DepId = LeaderId; | ||||
| 699 | } else | ||||
| 700 | // Each access has its own dependence set. | ||||
| 701 | DepId = RunningDepId++; | ||||
| 702 | |||||
| 703 | bool IsWrite = Access.getInt(); | ||||
| 704 | RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap, PSE); | ||||
| 705 | LLVM_DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n')do { } while (false); | ||||
| 706 | |||||
| 707 | return true; | ||||
| 708 | } | ||||
| 709 | |||||
| 710 | bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck, | ||||
| 711 | ScalarEvolution *SE, Loop *TheLoop, | ||||
| 712 | const ValueToValueMap &StridesMap, | ||||
| 713 | bool ShouldCheckWrap) { | ||||
| 714 | // Find pointers with computable bounds. We are going to use this information | ||||
| 715 | // to place a runtime bound check. | ||||
| 716 | bool CanDoRT = true; | ||||
| 717 | |||||
| 718 | bool MayNeedRTCheck = false; | ||||
| 719 | if (!IsRTCheckAnalysisNeeded) return true; | ||||
| 720 | |||||
| 721 | bool IsDepCheckNeeded = isDependencyCheckNeeded(); | ||||
| 722 | |||||
| 723 | // We assign a consecutive id to access from different alias sets. | ||||
| 724 | // Accesses between different groups doesn't need to be checked. | ||||
| 725 | unsigned ASId = 0; | ||||
| 726 | for (auto &AS : AST) { | ||||
| 727 | int NumReadPtrChecks = 0; | ||||
| 728 | int NumWritePtrChecks = 0; | ||||
| 729 | bool CanDoAliasSetRT = true; | ||||
| 730 | ++ASId; | ||||
| 731 | |||||
| 732 | // We assign consecutive id to access from different dependence sets. | ||||
| 733 | // Accesses within the same set don't need a runtime check. | ||||
| 734 | unsigned RunningDepId = 1; | ||||
| 735 | DenseMap<Value *, unsigned> DepSetId; | ||||
| 736 | |||||
| 737 | SmallVector<MemAccessInfo, 4> Retries; | ||||
| 738 | |||||
| 739 | // First, count how many write and read accesses are in the alias set. Also | ||||
| 740 | // collect MemAccessInfos for later. | ||||
| 741 | SmallVector<MemAccessInfo, 4> AccessInfos; | ||||
| 742 | for (const auto &A : AS) { | ||||
| 743 | Value *Ptr = A.getValue(); | ||||
| 744 | bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true)); | ||||
| 745 | |||||
| 746 | if (IsWrite) | ||||
| 747 | ++NumWritePtrChecks; | ||||
| 748 | else | ||||
| 749 | ++NumReadPtrChecks; | ||||
| 750 | AccessInfos.emplace_back(Ptr, IsWrite); | ||||
| 751 | } | ||||
| 752 | |||||
| 753 | // We do not need runtime checks for this alias set, if there are no writes | ||||
| 754 | // or a single write and no reads. | ||||
| 755 | if (NumWritePtrChecks
| ||||
| 756 | (NumWritePtrChecks
| ||||
| 757 | assert((AS.size() <= 1 ||((void)0) | ||||
| 758 | all_of(AS,((void)0) | ||||
| 759 | [this](auto AC) {((void)0) | ||||
| 760 | MemAccessInfo AccessWrite(AC.getValue(), true);((void)0) | ||||
| 761 | return DepCands.findValue(AccessWrite) == DepCands.end();((void)0) | ||||
| 762 | })) &&((void)0) | ||||
| 763 | "Can only skip updating CanDoRT below, if all entries in AS "((void)0) | ||||
| 764 | "are reads or there is at most 1 entry")((void)0); | ||||
| 765 | continue; | ||||
| 766 | } | ||||
| 767 | |||||
| 768 | for (auto &Access : AccessInfos) { | ||||
| 769 | if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId, TheLoop, | ||||
| 770 | RunningDepId, ASId, ShouldCheckWrap, false)) { | ||||
| 771 | LLVM_DEBUG(dbgs() << "LAA: Can't find bounds for ptr:"do { } while (false) | ||||
| 772 | << *Access.getPointer() << '\n')do { } while (false); | ||||
| 773 | Retries.push_back(Access); | ||||
| 774 | CanDoAliasSetRT = false; | ||||
| 775 | } | ||||
| 776 | } | ||||
| 777 | |||||
| 778 | // Note that this function computes CanDoRT and MayNeedRTCheck | ||||
| 779 | // independently. For example CanDoRT=false, MayNeedRTCheck=false means that | ||||
| 780 | // we have a pointer for which we couldn't find the bounds but we don't | ||||
| 781 | // actually need to emit any checks so it does not matter. | ||||
| 782 | // | ||||
| 783 | // We need runtime checks for this alias set, if there are at least 2 | ||||
| 784 | // dependence sets (in which case RunningDepId > 2) or if we need to re-try | ||||
| 785 | // any bound checks (because in that case the number of dependence sets is | ||||
| 786 | // incomplete). | ||||
| 787 | bool NeedsAliasSetRTCheck = RunningDepId > 2 || !Retries.empty(); | ||||
| 788 | |||||
| 789 | // We need to perform run-time alias checks, but some pointers had bounds | ||||
| 790 | // that couldn't be checked. | ||||
| 791 | if (NeedsAliasSetRTCheck && !CanDoAliasSetRT) { | ||||
| 792 | // Reset the CanDoSetRt flag and retry all accesses that have failed. | ||||
| 793 | // We know that we need these checks, so we can now be more aggressive | ||||
| 794 | // and add further checks if required (overflow checks). | ||||
| 795 | CanDoAliasSetRT = true; | ||||
| 796 | for (auto Access : Retries) | ||||
| 797 | if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId, | ||||
| 798 | TheLoop, RunningDepId, ASId, | ||||
| 799 | ShouldCheckWrap, /*Assume=*/true)) { | ||||
| 800 | CanDoAliasSetRT = false; | ||||
| 801 | break; | ||||
| 802 | } | ||||
| 803 | } | ||||
| 804 | |||||
| 805 | CanDoRT &= CanDoAliasSetRT; | ||||
| 806 | MayNeedRTCheck |= NeedsAliasSetRTCheck; | ||||
| 807 | ++ASId; | ||||
| 808 | } | ||||
| 809 | |||||
| 810 | // If the pointers that we would use for the bounds comparison have different | ||||
| 811 | // address spaces, assume the values aren't directly comparable, so we can't | ||||
| 812 | // use them for the runtime check. We also have to assume they could | ||||
| 813 | // overlap. In the future there should be metadata for whether address spaces | ||||
| 814 | // are disjoint. | ||||
| 815 | unsigned NumPointers = RtCheck.Pointers.size(); | ||||
| 816 | for (unsigned i = 0; i < NumPointers; ++i) { | ||||
| 817 | for (unsigned j = i + 1; j < NumPointers; ++j) { | ||||
| 818 | // Only need to check pointers between two different dependency sets. | ||||
| 819 | if (RtCheck.Pointers[i].DependencySetId == | ||||
| 820 | RtCheck.Pointers[j].DependencySetId) | ||||
| 821 | continue; | ||||
| 822 | // Only need to check pointers in the same alias set. | ||||
| 823 | if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId) | ||||
| 824 | continue; | ||||
| 825 | |||||
| 826 | Value *PtrI = RtCheck.Pointers[i].PointerValue; | ||||
| 827 | Value *PtrJ = RtCheck.Pointers[j].PointerValue; | ||||
| 828 | |||||
| 829 | unsigned ASi = PtrI->getType()->getPointerAddressSpace(); | ||||
| 830 | unsigned ASj = PtrJ->getType()->getPointerAddressSpace(); | ||||
| 831 | if (ASi != ASj) { | ||||
| 832 | LLVM_DEBUG(do { } while (false) | ||||
| 833 | dbgs() << "LAA: Runtime check would require comparison between"do { } while (false) | ||||
| 834 | " different address spaces\n")do { } while (false); | ||||
| 835 | return false; | ||||
| 836 | } | ||||
| 837 | } | ||||
| 838 | } | ||||
| 839 | |||||
| 840 | if (MayNeedRTCheck && CanDoRT) | ||||
| 841 | RtCheck.generateChecks(DepCands, IsDepCheckNeeded); | ||||
| 842 | |||||
| 843 | LLVM_DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks()do { } while (false) | ||||
| 844 | << " pointer comparisons.\n")do { } while (false); | ||||
| 845 | |||||
| 846 | // If we can do run-time checks, but there are no checks, no runtime checks | ||||
| 847 | // are needed. This can happen when all pointers point to the same underlying | ||||
| 848 | // object for example. | ||||
| 849 | RtCheck.Need = CanDoRT ? RtCheck.getNumberOfChecks() != 0 : MayNeedRTCheck; | ||||
| 850 | |||||
| 851 | bool CanDoRTIfNeeded = !RtCheck.Need || CanDoRT; | ||||
| 852 | if (!CanDoRTIfNeeded) | ||||
| 853 | RtCheck.reset(); | ||||
| 854 | return CanDoRTIfNeeded; | ||||
| 855 | } | ||||
| 856 | |||||
| 857 | void AccessAnalysis::processMemAccesses() { | ||||
| 858 | // We process the set twice: first we process read-write pointers, last we | ||||
| 859 | // process read-only pointers. This allows us to skip dependence tests for | ||||
| 860 | // read-only pointers. | ||||
| 861 | |||||
| 862 | LLVM_DEBUG(dbgs() << "LAA: Processing memory accesses...\n")do { } while (false); | ||||
| 863 | LLVM_DEBUG(dbgs() << " AST: "; AST.dump())do { } while (false); | ||||
| 864 | LLVM_DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n")do { } while (false); | ||||
| 865 | LLVM_DEBUG({do { } while (false) | ||||
| 866 | for (auto A : Accesses)do { } while (false) | ||||
| 867 | dbgs() << "\t" << *A.getPointer() << " (" <<do { } while (false) | ||||
| 868 | (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?do { } while (false) | ||||
| 869 | "read-only" : "read")) << ")\n";do { } while (false) | ||||
| 870 | })do { } while (false); | ||||
| 871 | |||||
| 872 | // The AliasSetTracker has nicely partitioned our pointers by metadata | ||||
| 873 | // compatibility and potential for underlying-object overlap. As a result, we | ||||
| 874 | // only need to check for potential pointer dependencies within each alias | ||||
| 875 | // set. | ||||
| 876 | for (const auto &AS : AST) { | ||||
| 877 | // Note that both the alias-set tracker and the alias sets themselves used | ||||
| 878 | // linked lists internally and so the iteration order here is deterministic | ||||
| 879 | // (matching the original instruction order within each set). | ||||
| 880 | |||||
| 881 | bool SetHasWrite = false; | ||||
| 882 | |||||
| 883 | // Map of pointers to last access encountered. | ||||
| 884 | typedef DenseMap<const Value*, MemAccessInfo> UnderlyingObjToAccessMap; | ||||
| 885 | UnderlyingObjToAccessMap ObjToLastAccess; | ||||
| 886 | |||||
| 887 | // Set of access to check after all writes have been processed. | ||||
| 888 | PtrAccessSet DeferredAccesses; | ||||
| 889 | |||||
| 890 | // Iterate over each alias set twice, once to process read/write pointers, | ||||
| 891 | // and then to process read-only pointers. | ||||
| 892 | for (int SetIteration = 0; SetIteration < 2; ++SetIteration) { | ||||
| 893 | bool UseDeferred = SetIteration > 0; | ||||
| 894 | PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses; | ||||
| 895 | |||||
| 896 | for (const auto &AV : AS) { | ||||
| 897 | Value *Ptr = AV.getValue(); | ||||
| 898 | |||||
| 899 | // For a single memory access in AliasSetTracker, Accesses may contain | ||||
| 900 | // both read and write, and they both need to be handled for CheckDeps. | ||||
| 901 | for (const auto &AC : S) { | ||||
| 902 | if (AC.getPointer() != Ptr) | ||||
| 903 | continue; | ||||
| 904 | |||||
| 905 | bool IsWrite = AC.getInt(); | ||||
| 906 | |||||
| 907 | // If we're using the deferred access set, then it contains only | ||||
| 908 | // reads. | ||||
| 909 | bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite; | ||||
| 910 | if (UseDeferred && !IsReadOnlyPtr) | ||||
| 911 | continue; | ||||
| 912 | // Otherwise, the pointer must be in the PtrAccessSet, either as a | ||||
| 913 | // read or a write. | ||||
| 914 | assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||((void)0) | ||||
| 915 | S.count(MemAccessInfo(Ptr, false))) &&((void)0) | ||||
| 916 | "Alias-set pointer not in the access set?")((void)0); | ||||
| 917 | |||||
| 918 | MemAccessInfo Access(Ptr, IsWrite); | ||||
| 919 | DepCands.insert(Access); | ||||
| 920 | |||||
| 921 | // Memorize read-only pointers for later processing and skip them in | ||||
| 922 | // the first round (they need to be checked after we have seen all | ||||
| 923 | // write pointers). Note: we also mark pointer that are not | ||||
| 924 | // consecutive as "read-only" pointers (so that we check | ||||
| 925 | // "a[b[i]] +="). Hence, we need the second check for "!IsWrite". | ||||
| 926 | if (!UseDeferred && IsReadOnlyPtr) { | ||||
| 927 | DeferredAccesses.insert(Access); | ||||
| 928 | continue; | ||||
| 929 | } | ||||
| 930 | |||||
| 931 | // If this is a write - check other reads and writes for conflicts. If | ||||
| 932 | // this is a read only check other writes for conflicts (but only if | ||||
| 933 | // there is no other write to the ptr - this is an optimization to | ||||
| 934 | // catch "a[i] = a[i] + " without having to do a dependence check). | ||||
| 935 | if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) { | ||||
| 936 | CheckDeps.push_back(Access); | ||||
| 937 | IsRTCheckAnalysisNeeded = true; | ||||
| 938 | } | ||||
| 939 | |||||
| 940 | if (IsWrite) | ||||
| 941 | SetHasWrite = true; | ||||
| 942 | |||||
| 943 | // Create sets of pointers connected by a shared alias set and | ||||
| 944 | // underlying object. | ||||
| 945 | typedef SmallVector<const Value *, 16> ValueVector; | ||||
| 946 | ValueVector TempObjects; | ||||
| 947 | |||||
| 948 | getUnderlyingObjects(Ptr, TempObjects, LI); | ||||
| 949 | LLVM_DEBUG(dbgs()do { } while (false) | ||||
| 950 | << "Underlying objects for pointer " << *Ptr << "\n")do { } while (false); | ||||
| 951 | for (const Value *UnderlyingObj : TempObjects) { | ||||
| 952 | // nullptr never alias, don't join sets for pointer that have "null" | ||||
| 953 | // in their UnderlyingObjects list. | ||||
| 954 | if (isa<ConstantPointerNull>(UnderlyingObj) && | ||||
| 955 | !NullPointerIsDefined( | ||||
| 956 | TheLoop->getHeader()->getParent(), | ||||
| 957 | UnderlyingObj->getType()->getPointerAddressSpace())) | ||||
| 958 | continue; | ||||
| 959 | |||||
| 960 | UnderlyingObjToAccessMap::iterator Prev = | ||||
| 961 | ObjToLastAccess.find(UnderlyingObj); | ||||
| 962 | if (Prev != ObjToLastAccess.end()) | ||||
| 963 | DepCands.unionSets(Access, Prev->second); | ||||
| 964 | |||||
| 965 | ObjToLastAccess[UnderlyingObj] = Access; | ||||
| 966 | LLVM_DEBUG(dbgs() << " " << *UnderlyingObj << "\n")do { } while (false); | ||||
| 967 | } | ||||
| 968 | } | ||||
| 969 | } | ||||
| 970 | } | ||||
| 971 | } | ||||
| 972 | } | ||||
| 973 | |||||
| 974 | static bool isInBoundsGep(Value *Ptr) { | ||||
| 975 | if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) | ||||
| 976 | return GEP->isInBounds(); | ||||
| 977 | return false; | ||||
| 978 | } | ||||
| 979 | |||||
| 980 | /// Return true if an AddRec pointer \p Ptr is unsigned non-wrapping, | ||||
| 981 | /// i.e. monotonically increasing/decreasing. | ||||
| 982 | static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR, | ||||
| 983 | PredicatedScalarEvolution &PSE, const Loop *L) { | ||||
| 984 | // FIXME: This should probably only return true for NUW. | ||||
| 985 | if (AR->getNoWrapFlags(SCEV::NoWrapMask)) | ||||
| 986 | return true; | ||||
| 987 | |||||
| 988 | // Scalar evolution does not propagate the non-wrapping flags to values that | ||||
| 989 | // are derived from a non-wrapping induction variable because non-wrapping | ||||
| 990 | // could be flow-sensitive. | ||||
| 991 | // | ||||
| 992 | // Look through the potentially overflowing instruction to try to prove | ||||
| 993 | // non-wrapping for the *specific* value of Ptr. | ||||
| 994 | |||||
| 995 | // The arithmetic implied by an inbounds GEP can't overflow. | ||||
| 996 | auto *GEP = dyn_cast<GetElementPtrInst>(Ptr); | ||||
| 997 | if (!GEP || !GEP->isInBounds()) | ||||
| 998 | return false; | ||||
| 999 | |||||
| 1000 | // Make sure there is only one non-const index and analyze that. | ||||
| 1001 | Value *NonConstIndex = nullptr; | ||||
| 1002 | for (Value *Index : GEP->indices()) | ||||
| 1003 | if (!isa<ConstantInt>(Index)) { | ||||
| 1004 | if (NonConstIndex) | ||||
| 1005 | return false; | ||||
| 1006 | NonConstIndex = Index; | ||||
| 1007 | } | ||||
| 1008 | if (!NonConstIndex) | ||||
| 1009 | // The recurrence is on the pointer, ignore for now. | ||||
| 1010 | return false; | ||||
| 1011 | |||||
| 1012 | // The index in GEP is signed. It is non-wrapping if it's derived from a NSW | ||||
| 1013 | // AddRec using a NSW operation. | ||||
| 1014 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex)) | ||||
| 1015 | if (OBO->hasNoSignedWrap() && | ||||
| 1016 | // Assume constant for other the operand so that the AddRec can be | ||||
| 1017 | // easily found. | ||||
| 1018 | isa<ConstantInt>(OBO->getOperand(1))) { | ||||
| 1019 | auto *OpScev = PSE.getSCEV(OBO->getOperand(0)); | ||||
| 1020 | |||||
| 1021 | if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev)) | ||||
| 1022 | return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW); | ||||
| 1023 | } | ||||
| 1024 | |||||
| 1025 | return false; | ||||
| 1026 | } | ||||
| 1027 | |||||
| 1028 | /// Check whether the access through \p Ptr has a constant stride. | ||||
| 1029 | int64_t llvm::getPtrStride(PredicatedScalarEvolution &PSE, Value *Ptr, | ||||
| 1030 | const Loop *Lp, const ValueToValueMap &StridesMap, | ||||
| 1031 | bool Assume, bool ShouldCheckWrap) { | ||||
| 1032 | Type *Ty = Ptr->getType(); | ||||
| 1033 | assert(Ty->isPointerTy() && "Unexpected non-ptr")((void)0); | ||||
| 1034 | |||||
| 1035 | // Make sure that the pointer does not point to aggregate types. | ||||
| 1036 | auto *PtrTy = cast<PointerType>(Ty); | ||||
| 1037 | if (PtrTy->getElementType()->isAggregateType()) { | ||||
| 1038 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"do { } while (false) | ||||
| 1039 | << *Ptr << "\n")do { } while (false); | ||||
| 1040 | return 0; | ||||
| 1041 | } | ||||
| 1042 | |||||
| 1043 | const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr); | ||||
| 1044 | |||||
| 1045 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); | ||||
| 1046 | if (Assume && !AR) | ||||
| 1047 | AR = PSE.getAsAddRec(Ptr); | ||||
| 1048 | |||||
| 1049 | if (!AR) { | ||||
| 1050 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptrdo { } while (false) | ||||
| 1051 | << " SCEV: " << *PtrScev << "\n")do { } while (false); | ||||
| 1052 | return 0; | ||||
| 1053 | } | ||||
| 1054 | |||||
| 1055 | // The access function must stride over the innermost loop. | ||||
| 1056 | if (Lp != AR->getLoop()) { | ||||
| 1057 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop "do { } while (false) | ||||
| 1058 | << *Ptr << " SCEV: " << *AR << "\n")do { } while (false); | ||||
| 1059 | return 0; | ||||
| 1060 | } | ||||
| 1061 | |||||
| 1062 | // The address calculation must not wrap. Otherwise, a dependence could be | ||||
| 1063 | // inverted. | ||||
| 1064 | // An inbounds getelementptr that is a AddRec with a unit stride | ||||
| 1065 | // cannot wrap per definition. The unit stride requirement is checked later. | ||||
| 1066 | // An getelementptr without an inbounds attribute and unit stride would have | ||||
| 1067 | // to access the pointer value "0" which is undefined behavior in address | ||||
| 1068 | // space 0, therefore we can also vectorize this case. | ||||
| 1069 | bool IsInBoundsGEP = isInBoundsGep(Ptr); | ||||
| 1070 | bool IsNoWrapAddRec = !ShouldCheckWrap || | ||||
| 1071 | PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW) || | ||||
| 1072 | isNoWrapAddRec(Ptr, AR, PSE, Lp); | ||||
| 1073 | if (!IsNoWrapAddRec && !IsInBoundsGEP && | ||||
| 1074 | NullPointerIsDefined(Lp->getHeader()->getParent(), | ||||
| 1075 | PtrTy->getAddressSpace())) { | ||||
| 1076 | if (Assume) { | ||||
| 1077 | PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); | ||||
| 1078 | IsNoWrapAddRec = true; | ||||
| 1079 | LLVM_DEBUG(dbgs() << "LAA: Pointer may wrap in the address space:\n"do { } while (false) | ||||
| 1080 | << "LAA: Pointer: " << *Ptr << "\n"do { } while (false) | ||||
| 1081 | << "LAA: SCEV: " << *AR << "\n"do { } while (false) | ||||
| 1082 | << "LAA: Added an overflow assumption\n")do { } while (false); | ||||
| 1083 | } else { | ||||
| 1084 | LLVM_DEBUG(do { } while (false) | ||||
| 1085 | dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "do { } while (false) | ||||
| 1086 | << *Ptr << " SCEV: " << *AR << "\n")do { } while (false); | ||||
| 1087 | return 0; | ||||
| 1088 | } | ||||
| 1089 | } | ||||
| 1090 | |||||
| 1091 | // Check the step is constant. | ||||
| 1092 | const SCEV *Step = AR->getStepRecurrence(*PSE.getSE()); | ||||
| 1093 | |||||
| 1094 | // Calculate the pointer stride and check if it is constant. | ||||
| 1095 | const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); | ||||
| 1096 | if (!C) { | ||||
| 1097 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptrdo { } while (false) | ||||
| 1098 | << " SCEV: " << *AR << "\n")do { } while (false); | ||||
| 1099 | return 0; | ||||
| 1100 | } | ||||
| 1101 | |||||
| 1102 | auto &DL = Lp->getHeader()->getModule()->getDataLayout(); | ||||
| 1103 | int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType()); | ||||
| 1104 | const APInt &APStepVal = C->getAPInt(); | ||||
| 1105 | |||||
| 1106 | // Huge step value - give up. | ||||
| 1107 | if (APStepVal.getBitWidth() > 64) | ||||
| 1108 | return 0; | ||||
| 1109 | |||||
| 1110 | int64_t StepVal = APStepVal.getSExtValue(); | ||||
| 1111 | |||||
| 1112 | // Strided access. | ||||
| 1113 | int64_t Stride = StepVal / Size; | ||||
| 1114 | int64_t Rem = StepVal % Size; | ||||
| 1115 | if (Rem) | ||||
| 1116 | return 0; | ||||
| 1117 | |||||
| 1118 | // If the SCEV could wrap but we have an inbounds gep with a unit stride we | ||||
| 1119 | // know we can't "wrap around the address space". In case of address space | ||||
| 1120 | // zero we know that this won't happen without triggering undefined behavior. | ||||
| 1121 | if (!IsNoWrapAddRec && Stride != 1 && Stride != -1 && | ||||
| 1122 | (IsInBoundsGEP || !NullPointerIsDefined(Lp->getHeader()->getParent(), | ||||
| 1123 | PtrTy->getAddressSpace()))) { | ||||
| 1124 | if (Assume) { | ||||
| 1125 | // We can avoid this case by adding a run-time check. | ||||
| 1126 | LLVM_DEBUG(dbgs() << "LAA: Non unit strided pointer which is not either "do { } while (false) | ||||
| 1127 | << "inbounds or in address space 0 may wrap:\n"do { } while (false) | ||||
| 1128 | << "LAA: Pointer: " << *Ptr << "\n"do { } while (false) | ||||
| 1129 | << "LAA: SCEV: " << *AR << "\n"do { } while (false) | ||||
| 1130 | << "LAA: Added an overflow assumption\n")do { } while (false); | ||||
| 1131 | PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); | ||||
| 1132 | } else | ||||
| 1133 | return 0; | ||||
| 1134 | } | ||||
| 1135 | |||||
| 1136 | return Stride; | ||||
| 1137 | } | ||||
| 1138 | |||||
| 1139 | Optional<int> llvm::getPointersDiff(Type *ElemTyA, Value *PtrA, Type *ElemTyB, | ||||
| 1140 | Value *PtrB, const DataLayout &DL, | ||||
| 1141 | ScalarEvolution &SE, bool StrictCheck, | ||||
| 1142 | bool CheckType) { | ||||
| 1143 | assert(PtrA && PtrB && "Expected non-nullptr pointers.")((void)0); | ||||
| 1144 | assert(cast<PointerType>(PtrA->getType())((void)0) | ||||
| 1145 | ->isOpaqueOrPointeeTypeMatches(ElemTyA) && "Wrong PtrA type")((void)0); | ||||
| 1146 | assert(cast<PointerType>(PtrB->getType())((void)0) | ||||
| 1147 | ->isOpaqueOrPointeeTypeMatches(ElemTyB) && "Wrong PtrB type")((void)0); | ||||
| 1148 | |||||
| 1149 | // Make sure that A and B are different pointers. | ||||
| 1150 | if (PtrA == PtrB) | ||||
| 1151 | return 0; | ||||
| 1152 | |||||
| 1153 | // Make sure that the element types are the same if required. | ||||
| 1154 | if (CheckType && ElemTyA != ElemTyB) | ||||
| 1155 | return None; | ||||
| 1156 | |||||
| 1157 | unsigned ASA = PtrA->getType()->getPointerAddressSpace(); | ||||
| 1158 | unsigned ASB = PtrB->getType()->getPointerAddressSpace(); | ||||
| 1159 | |||||
| 1160 | // Check that the address spaces match. | ||||
| 1161 | if (ASA != ASB) | ||||
| 1162 | return None; | ||||
| 1163 | unsigned IdxWidth = DL.getIndexSizeInBits(ASA); | ||||
| 1164 | |||||
| 1165 | APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0); | ||||
| 1166 | Value *PtrA1 = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA); | ||||
| 1167 | Value *PtrB1 = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB); | ||||
| 1168 | |||||
| 1169 | int Val; | ||||
| 1170 | if (PtrA1 == PtrB1) { | ||||
| 1171 | // Retrieve the address space again as pointer stripping now tracks through | ||||
| 1172 | // `addrspacecast`. | ||||
| 1173 | ASA = cast<PointerType>(PtrA1->getType())->getAddressSpace(); | ||||
| 1174 | ASB = cast<PointerType>(PtrB1->getType())->getAddressSpace(); | ||||
| 1175 | // Check that the address spaces match and that the pointers are valid. | ||||
| 1176 | if (ASA != ASB) | ||||
| 1177 | return None; | ||||
| 1178 | |||||
| 1179 | IdxWidth = DL.getIndexSizeInBits(ASA); | ||||
| 1180 | OffsetA = OffsetA.sextOrTrunc(IdxWidth); | ||||
| 1181 | OffsetB = OffsetB.sextOrTrunc(IdxWidth); | ||||
| 1182 | |||||
| 1183 | OffsetB -= OffsetA; | ||||
| 1184 | Val = OffsetB.getSExtValue(); | ||||
| 1185 | } else { | ||||
| 1186 | // Otherwise compute the distance with SCEV between the base pointers. | ||||
| 1187 | const SCEV *PtrSCEVA = SE.getSCEV(PtrA); | ||||
| 1188 | const SCEV *PtrSCEVB = SE.getSCEV(PtrB); | ||||
| 1189 | const auto *Diff = | ||||
| 1190 | dyn_cast<SCEVConstant>(SE.getMinusSCEV(PtrSCEVB, PtrSCEVA)); | ||||
| 1191 | if (!Diff) | ||||
| 1192 | return None; | ||||
| 1193 | Val = Diff->getAPInt().getSExtValue(); | ||||
| 1194 | } | ||||
| 1195 | int Size = DL.getTypeStoreSize(ElemTyA); | ||||
| 1196 | int Dist = Val / Size; | ||||
| 1197 | |||||
| 1198 | // Ensure that the calculated distance matches the type-based one after all | ||||
| 1199 | // the bitcasts removal in the provided pointers. | ||||
| 1200 | if (!StrictCheck || Dist * Size == Val) | ||||
| 1201 | return Dist; | ||||
| 1202 | return None; | ||||
| 1203 | } | ||||
| 1204 | |||||
| 1205 | bool llvm::sortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy, | ||||
| 1206 | const DataLayout &DL, ScalarEvolution &SE, | ||||
| 1207 | SmallVectorImpl<unsigned> &SortedIndices) { | ||||
| 1208 | assert(llvm::all_of(((void)0) | ||||
| 1209 | VL, [](const Value *V) { return V->getType()->isPointerTy(); }) &&((void)0) | ||||
| 1210 | "Expected list of pointer operands.")((void)0); | ||||
| 1211 | // Walk over the pointers, and map each of them to an offset relative to | ||||
| 1212 | // first pointer in the array. | ||||
| 1213 | Value *Ptr0 = VL[0]; | ||||
| 1214 | |||||
| 1215 | using DistOrdPair = std::pair<int64_t, int>; | ||||
| 1216 | auto Compare = [](const DistOrdPair &L, const DistOrdPair &R) { | ||||
| 1217 | return L.first < R.first; | ||||
| 1218 | }; | ||||
| 1219 | std::set<DistOrdPair, decltype(Compare)> Offsets(Compare); | ||||
| 1220 | Offsets.emplace(0, 0); | ||||
| 1221 | int Cnt = 1; | ||||
| 1222 | bool IsConsecutive = true; | ||||
| 1223 | for (auto *Ptr : VL.drop_front()) { | ||||
| 1224 | Optional<int> Diff = getPointersDiff(ElemTy, Ptr0, ElemTy, Ptr, DL, SE, | ||||
| 1225 | /*StrictCheck=*/true); | ||||
| 1226 | if (!Diff) | ||||
| 1227 | return false; | ||||
| 1228 | |||||
| 1229 | // Check if the pointer with the same offset is found. | ||||
| 1230 | int64_t Offset = *Diff; | ||||
| 1231 | auto Res = Offsets.emplace(Offset, Cnt); | ||||
| 1232 | if (!Res.second) | ||||
| 1233 | return false; | ||||
| 1234 | // Consecutive order if the inserted element is the last one. | ||||
| 1235 | IsConsecutive = IsConsecutive && std::next(Res.first) == Offsets.end(); | ||||
| 1236 | ++Cnt; | ||||
| 1237 | } | ||||
| 1238 | SortedIndices.clear(); | ||||
| 1239 | if (!IsConsecutive) { | ||||
| 1240 | // Fill SortedIndices array only if it is non-consecutive. | ||||
| 1241 | SortedIndices.resize(VL.size()); | ||||
| 1242 | Cnt = 0; | ||||
| 1243 | for (const std::pair<int64_t, int> &Pair : Offsets) { | ||||
| 1244 | SortedIndices[Cnt] = Pair.second; | ||||
| 1245 | ++Cnt; | ||||
| 1246 | } | ||||
| 1247 | } | ||||
| 1248 | return true; | ||||
| 1249 | } | ||||
| 1250 | |||||
| 1251 | /// Returns true if the memory operations \p A and \p B are consecutive. | ||||
| 1252 | bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL, | ||||
| 1253 | ScalarEvolution &SE, bool CheckType) { | ||||
| 1254 | Value *PtrA = getLoadStorePointerOperand(A); | ||||
| 1255 | Value *PtrB = getLoadStorePointerOperand(B); | ||||
| 1256 | if (!PtrA || !PtrB) | ||||
| 1257 | return false; | ||||
| 1258 | Type *ElemTyA = getLoadStoreType(A); | ||||
| 1259 | Type *ElemTyB = getLoadStoreType(B); | ||||
| 1260 | Optional<int> Diff = getPointersDiff(ElemTyA, PtrA, ElemTyB, PtrB, DL, SE, | ||||
| 1261 | /*StrictCheck=*/true, CheckType); | ||||
| 1262 | return Diff && *Diff == 1; | ||||
| 1263 | } | ||||
| 1264 | |||||
| 1265 | MemoryDepChecker::VectorizationSafetyStatus | ||||
| 1266 | MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) { | ||||
| 1267 | switch (Type) { | ||||
| 1268 | case NoDep: | ||||
| 1269 | case Forward: | ||||
| 1270 | case BackwardVectorizable: | ||||
| 1271 | return VectorizationSafetyStatus::Safe; | ||||
| 1272 | |||||
| 1273 | case Unknown: | ||||
| 1274 | return VectorizationSafetyStatus::PossiblySafeWithRtChecks; | ||||
| 1275 | case ForwardButPreventsForwarding: | ||||
| 1276 | case Backward: | ||||
| 1277 | case BackwardVectorizableButPreventsForwarding: | ||||
| 1278 | return VectorizationSafetyStatus::Unsafe; | ||||
| 1279 | } | ||||
| 1280 | llvm_unreachable("unexpected DepType!")__builtin_unreachable(); | ||||
| 1281 | } | ||||
| 1282 | |||||
| 1283 | bool MemoryDepChecker::Dependence::isBackward() const { | ||||
| 1284 | switch (Type) { | ||||
| 1285 | case NoDep: | ||||
| 1286 | case Forward: | ||||
| 1287 | case ForwardButPreventsForwarding: | ||||
| 1288 | case Unknown: | ||||
| 1289 | return false; | ||||
| 1290 | |||||
| 1291 | case BackwardVectorizable: | ||||
| 1292 | case Backward: | ||||
| 1293 | case BackwardVectorizableButPreventsForwarding: | ||||
| 1294 | return true; | ||||
| 1295 | } | ||||
| 1296 | llvm_unreachable("unexpected DepType!")__builtin_unreachable(); | ||||
| 1297 | } | ||||
| 1298 | |||||
| 1299 | bool MemoryDepChecker::Dependence::isPossiblyBackward() const { | ||||
| 1300 | return isBackward() || Type == Unknown; | ||||
| 1301 | } | ||||
| 1302 | |||||
| 1303 | bool MemoryDepChecker::Dependence::isForward() const { | ||||
| 1304 | switch (Type) { | ||||
| 1305 | case Forward: | ||||
| 1306 | case ForwardButPreventsForwarding: | ||||
| 1307 | return true; | ||||
| 1308 | |||||
| 1309 | case NoDep: | ||||
| 1310 | case Unknown: | ||||
| 1311 | case BackwardVectorizable: | ||||
| 1312 | case Backward: | ||||
| 1313 | case BackwardVectorizableButPreventsForwarding: | ||||
| 1314 | return false; | ||||
| 1315 | } | ||||
| 1316 | llvm_unreachable("unexpected DepType!")__builtin_unreachable(); | ||||
| 1317 | } | ||||
| 1318 | |||||
| 1319 | bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance, | ||||
| 1320 | uint64_t TypeByteSize) { | ||||
| 1321 | // If loads occur at a distance that is not a multiple of a feasible vector | ||||
| 1322 | // factor store-load forwarding does not take place. | ||||
| 1323 | // Positive dependences might cause troubles because vectorizing them might | ||||
| 1324 | // prevent store-load forwarding making vectorized code run a lot slower. | ||||
| 1325 | // a[i] = a[i-3] ^ a[i-8]; | ||||
| 1326 | // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and | ||||
| 1327 | // hence on your typical architecture store-load forwarding does not take | ||||
| 1328 | // place. Vectorizing in such cases does not make sense. | ||||
| 1329 | // Store-load forwarding distance. | ||||
| 1330 | |||||
| 1331 | // After this many iterations store-to-load forwarding conflicts should not | ||||
| 1332 | // cause any slowdowns. | ||||
| 1333 | const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize; | ||||
| 1334 | // Maximum vector factor. | ||||
| 1335 | uint64_t MaxVFWithoutSLForwardIssues = std::min( | ||||
| 1336 | VectorizerParams::MaxVectorWidth * TypeByteSize, MaxSafeDepDistBytes); | ||||
| 1337 | |||||
| 1338 | // Compute the smallest VF at which the store and load would be misaligned. | ||||
| 1339 | for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues; | ||||
| 1340 | VF *= 2) { | ||||
| 1341 | // If the number of vector iteration between the store and the load are | ||||
| 1342 | // small we could incur conflicts. | ||||
| 1343 | if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) { | ||||
| 1344 | MaxVFWithoutSLForwardIssues = (VF >> 1); | ||||
| 1345 | break; | ||||
| 1346 | } | ||||
| 1347 | } | ||||
| 1348 | |||||
| 1349 | if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) { | ||||
| 1350 | LLVM_DEBUG(do { } while (false) | ||||
| 1351 | dbgs() << "LAA: Distance " << Distancedo { } while (false) | ||||
| 1352 | << " that could cause a store-load forwarding conflict\n")do { } while (false); | ||||
| 1353 | return true; | ||||
| 1354 | } | ||||
| 1355 | |||||
| 1356 | if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes && | ||||
| 1357 | MaxVFWithoutSLForwardIssues != | ||||
| 1358 | VectorizerParams::MaxVectorWidth * TypeByteSize) | ||||
| 1359 | MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues; | ||||
| 1360 | return false; | ||||
| 1361 | } | ||||
| 1362 | |||||
| 1363 | void MemoryDepChecker::mergeInStatus(VectorizationSafetyStatus S) { | ||||
| 1364 | if (Status < S) | ||||
| 1365 | Status = S; | ||||
| 1366 | } | ||||
| 1367 | |||||
| 1368 | /// Given a non-constant (unknown) dependence-distance \p Dist between two | ||||
| 1369 | /// memory accesses, that have the same stride whose absolute value is given | ||||
| 1370 | /// in \p Stride, and that have the same type size \p TypeByteSize, | ||||
| 1371 | /// in a loop whose takenCount is \p BackedgeTakenCount, check if it is | ||||
| 1372 | /// possible to prove statically that the dependence distance is larger | ||||
| 1373 | /// than the range that the accesses will travel through the execution of | ||||
| 1374 | /// the loop. If so, return true; false otherwise. This is useful for | ||||
| 1375 | /// example in loops such as the following (PR31098): | ||||
| 1376 | /// for (i = 0; i < D; ++i) { | ||||
| 1377 | /// = out[i]; | ||||
| 1378 | /// out[i+D] = | ||||
| 1379 | /// } | ||||
| 1380 | static bool isSafeDependenceDistance(const DataLayout &DL, ScalarEvolution &SE, | ||||
| 1381 | const SCEV &BackedgeTakenCount, | ||||
| 1382 | const SCEV &Dist, uint64_t Stride, | ||||
| 1383 | uint64_t TypeByteSize) { | ||||
| 1384 | |||||
| 1385 | // If we can prove that | ||||
| 1386 | // (**) |Dist| > BackedgeTakenCount * Step | ||||
| 1387 | // where Step is the absolute stride of the memory accesses in bytes, | ||||
| 1388 | // then there is no dependence. | ||||
| 1389 | // | ||||
| 1390 | // Rationale: | ||||
| 1391 | // We basically want to check if the absolute distance (|Dist/Step|) | ||||
| 1392 | // is >= the loop iteration count (or > BackedgeTakenCount). | ||||
| 1393 | // This is equivalent to the Strong SIV Test (Practical Dependence Testing, | ||||
| 1394 | // Section 4.2.1); Note, that for vectorization it is sufficient to prove | ||||
| 1395 | // that the dependence distance is >= VF; This is checked elsewhere. | ||||
| 1396 | // But in some cases we can prune unknown dependence distances early, and | ||||
| 1397 | // even before selecting the VF, and without a runtime test, by comparing | ||||
| 1398 | // the distance against the loop iteration count. Since the vectorized code | ||||
| 1399 | // will be executed only if LoopCount >= VF, proving distance >= LoopCount | ||||
| 1400 | // also guarantees that distance >= VF. | ||||
| 1401 | // | ||||
| 1402 | const uint64_t ByteStride = Stride * TypeByteSize; | ||||
| 1403 | const SCEV *Step = SE.getConstant(BackedgeTakenCount.getType(), ByteStride); | ||||
| 1404 | const SCEV *Product = SE.getMulExpr(&BackedgeTakenCount, Step); | ||||
| 1405 | |||||
| 1406 | const SCEV *CastedDist = &Dist; | ||||
| 1407 | const SCEV *CastedProduct = Product; | ||||
| 1408 | uint64_t DistTypeSize = DL.getTypeAllocSize(Dist.getType()); | ||||
| 1409 | uint64_t ProductTypeSize = DL.getTypeAllocSize(Product->getType()); | ||||
| 1410 | |||||
| 1411 | // The dependence distance can be positive/negative, so we sign extend Dist; | ||||
| 1412 | // The multiplication of the absolute stride in bytes and the | ||||
| 1413 | // backedgeTakenCount is non-negative, so we zero extend Product. | ||||
| 1414 | if (DistTypeSize > ProductTypeSize) | ||||
| 1415 | CastedProduct = SE.getZeroExtendExpr(Product, Dist.getType()); | ||||
| 1416 | else | ||||
| 1417 | CastedDist = SE.getNoopOrSignExtend(&Dist, Product->getType()); | ||||
| 1418 | |||||
| 1419 | // Is Dist - (BackedgeTakenCount * Step) > 0 ? | ||||
| 1420 | // (If so, then we have proven (**) because |Dist| >= Dist) | ||||
| 1421 | const SCEV *Minus = SE.getMinusSCEV(CastedDist, CastedProduct); | ||||
| 1422 | if (SE.isKnownPositive(Minus)) | ||||
| 1423 | return true; | ||||
| 1424 | |||||
| 1425 | // Second try: Is -Dist - (BackedgeTakenCount * Step) > 0 ? | ||||
| 1426 | // (If so, then we have proven (**) because |Dist| >= -1*Dist) | ||||
| 1427 | const SCEV *NegDist = SE.getNegativeSCEV(CastedDist); | ||||
| 1428 | Minus = SE.getMinusSCEV(NegDist, CastedProduct); | ||||
| 1429 | if (SE.isKnownPositive(Minus)) | ||||
| 1430 | return true; | ||||
| 1431 | |||||
| 1432 | return false; | ||||
| 1433 | } | ||||
| 1434 | |||||
| 1435 | /// Check the dependence for two accesses with the same stride \p Stride. | ||||
| 1436 | /// \p Distance is the positive distance and \p TypeByteSize is type size in | ||||
| 1437 | /// bytes. | ||||
| 1438 | /// | ||||
| 1439 | /// \returns true if they are independent. | ||||
| 1440 | static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride, | ||||
| 1441 | uint64_t TypeByteSize) { | ||||
| 1442 | assert(Stride > 1 && "The stride must be greater than 1")((void)0); | ||||
| 1443 | assert(TypeByteSize > 0 && "The type size in byte must be non-zero")((void)0); | ||||
| 1444 | assert(Distance > 0 && "The distance must be non-zero")((void)0); | ||||
| 1445 | |||||
| 1446 | // Skip if the distance is not multiple of type byte size. | ||||
| 1447 | if (Distance % TypeByteSize) | ||||
| 1448 | return false; | ||||
| 1449 | |||||
| 1450 | uint64_t ScaledDist = Distance / TypeByteSize; | ||||
| 1451 | |||||
| 1452 | // No dependence if the scaled distance is not multiple of the stride. | ||||
| 1453 | // E.g. | ||||
| 1454 | // for (i = 0; i < 1024 ; i += 4) | ||||
| 1455 | // A[i+2] = A[i] + 1; | ||||
| 1456 | // | ||||
| 1457 | // Two accesses in memory (scaled distance is 2, stride is 4): | ||||
| 1458 | // | A[0] | | | | A[4] | | | | | ||||
| 1459 | // | | | A[2] | | | | A[6] | | | ||||
| 1460 | // | ||||
| 1461 | // E.g. | ||||
| 1462 | // for (i = 0; i < 1024 ; i += 3) | ||||
| 1463 | // A[i+4] = A[i] + 1; | ||||
| 1464 | // | ||||
| 1465 | // Two accesses in memory (scaled distance is 4, stride is 3): | ||||
| 1466 | // | A[0] | | | A[3] | | | A[6] | | | | ||||
| 1467 | // | | | | | A[4] | | | A[7] | | | ||||
| 1468 | return ScaledDist % Stride; | ||||
| 1469 | } | ||||
| 1470 | |||||
| 1471 | MemoryDepChecker::Dependence::DepType | ||||
| 1472 | MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx, | ||||
| 1473 | const MemAccessInfo &B, unsigned BIdx, | ||||
| 1474 | const ValueToValueMap &Strides) { | ||||
| 1475 | assert (AIdx < BIdx && "Must pass arguments in program order")((void)0); | ||||
| 1476 | |||||
| 1477 | Value *APtr = A.getPointer(); | ||||
| 1478 | Value *BPtr = B.getPointer(); | ||||
| 1479 | bool AIsWrite = A.getInt(); | ||||
| 1480 | bool BIsWrite = B.getInt(); | ||||
| 1481 | |||||
| 1482 | // Two reads are independent. | ||||
| 1483 | if (!AIsWrite && !BIsWrite) | ||||
| 1484 | return Dependence::NoDep; | ||||
| 1485 | |||||
| 1486 | // We cannot check pointers in different address spaces. | ||||
| 1487 | if (APtr->getType()->getPointerAddressSpace() != | ||||
| 1488 | BPtr->getType()->getPointerAddressSpace()) | ||||
| 1489 | return Dependence::Unknown; | ||||
| 1490 | |||||
| 1491 | int64_t StrideAPtr = getPtrStride(PSE, APtr, InnermostLoop, Strides, true); | ||||
| 1492 | int64_t StrideBPtr = getPtrStride(PSE, BPtr, InnermostLoop, Strides, true); | ||||
| 1493 | |||||
| 1494 | const SCEV *Src = PSE.getSCEV(APtr); | ||||
| 1495 | const SCEV *Sink = PSE.getSCEV(BPtr); | ||||
| 1496 | |||||
| 1497 | // If the induction step is negative we have to invert source and sink of the | ||||
| 1498 | // dependence. | ||||
| 1499 | if (StrideAPtr < 0) { | ||||
| 1500 | std::swap(APtr, BPtr); | ||||
| 1501 | std::swap(Src, Sink); | ||||
| 1502 | std::swap(AIsWrite, BIsWrite); | ||||
| 1503 | std::swap(AIdx, BIdx); | ||||
| 1504 | std::swap(StrideAPtr, StrideBPtr); | ||||
| 1505 | } | ||||
| 1506 | |||||
| 1507 | const SCEV *Dist = PSE.getSE()->getMinusSCEV(Sink, Src); | ||||
| 1508 | |||||
| 1509 | LLVM_DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sinkdo { } while (false) | ||||
| 1510 | << "(Induction step: " << StrideAPtr << ")\n")do { } while (false); | ||||
| 1511 | LLVM_DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "do { } while (false) | ||||
| 1512 | << *InstMap[BIdx] << ": " << *Dist << "\n")do { } while (false); | ||||
| 1513 | |||||
| 1514 | // Need accesses with constant stride. We don't want to vectorize | ||||
| 1515 | // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in | ||||
| 1516 | // the address space. | ||||
| 1517 | if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){ | ||||
| 1518 | LLVM_DEBUG(dbgs() << "Pointer access with non-constant stride\n")do { } while (false); | ||||
| 1519 | return Dependence::Unknown; | ||||
| 1520 | } | ||||
| 1521 | |||||
| 1522 | Type *ATy = APtr->getType()->getPointerElementType(); | ||||
| 1523 | Type *BTy = BPtr->getType()->getPointerElementType(); | ||||
| 1524 | auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout(); | ||||
| 1525 | uint64_t TypeByteSize = DL.getTypeAllocSize(ATy); | ||||
| 1526 | uint64_t Stride = std::abs(StrideAPtr); | ||||
| 1527 | const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist); | ||||
| 1528 | if (!C) { | ||||
| 1529 | if (!isa<SCEVCouldNotCompute>(Dist) && | ||||
| 1530 | TypeByteSize == DL.getTypeAllocSize(BTy) && | ||||
| 1531 | isSafeDependenceDistance(DL, *(PSE.getSE()), | ||||
| 1532 | *(PSE.getBackedgeTakenCount()), *Dist, Stride, | ||||
| 1533 | TypeByteSize)) | ||||
| 1534 | return Dependence::NoDep; | ||||
| 1535 | |||||
| 1536 | LLVM_DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n")do { } while (false); | ||||
| 1537 | FoundNonConstantDistanceDependence = true; | ||||
| 1538 | return Dependence::Unknown; | ||||
| 1539 | } | ||||
| 1540 | |||||
| 1541 | const APInt &Val = C->getAPInt(); | ||||
| 1542 | int64_t Distance = Val.getSExtValue(); | ||||
| 1543 | |||||
| 1544 | // Attempt to prove strided accesses independent. | ||||
| 1545 | if (std::abs(Distance) > 0 && Stride > 1 && ATy == BTy && | ||||
| 1546 | areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) { | ||||
| 1547 | LLVM_DEBUG(dbgs() << "LAA: Strided accesses are independent\n")do { } while (false); | ||||
| 1548 | return Dependence::NoDep; | ||||
| 1549 | } | ||||
| 1550 | |||||
| 1551 | // Negative distances are not plausible dependencies. | ||||
| 1552 | if (Val.isNegative()) { | ||||
| 1553 | bool IsTrueDataDependence = (AIsWrite && !BIsWrite); | ||||
| 1554 | if (IsTrueDataDependence && EnableForwardingConflictDetection && | ||||
| 1555 | (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) || | ||||
| 1556 | ATy != BTy)) { | ||||
| 1557 | LLVM_DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n")do { } while (false); | ||||
| 1558 | return Dependence::ForwardButPreventsForwarding; | ||||
| 1559 | } | ||||
| 1560 | |||||
| 1561 | LLVM_DEBUG(dbgs() << "LAA: Dependence is negative\n")do { } while (false); | ||||
| 1562 | return Dependence::Forward; | ||||
| 1563 | } | ||||
| 1564 | |||||
| 1565 | // Write to the same location with the same size. | ||||
| 1566 | // Could be improved to assert type sizes are the same (i32 == float, etc). | ||||
| 1567 | if (Val == 0) { | ||||
| 1568 | if (ATy == BTy) | ||||
| 1569 | return Dependence::Forward; | ||||
| 1570 | LLVM_DEBUG(do { } while (false) | ||||
| 1571 | dbgs() << "LAA: Zero dependence difference but different types\n")do { } while (false); | ||||
| 1572 | return Dependence::Unknown; | ||||
| 1573 | } | ||||
| 1574 | |||||
| 1575 | assert(Val.isStrictlyPositive() && "Expect a positive value")((void)0); | ||||
| 1576 | |||||
| 1577 | if (ATy != BTy) { | ||||
| 1578 | LLVM_DEBUG(do { } while (false) | ||||
| 1579 | dbgs()do { } while (false) | ||||
| 1580 | << "LAA: ReadWrite-Write positive dependency with different types\n")do { } while (false); | ||||
| 1581 | return Dependence::Unknown; | ||||
| 1582 | } | ||||
| 1583 | |||||
| 1584 | // Bail out early if passed-in parameters make vectorization not feasible. | ||||
| 1585 | unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ? | ||||
| 1586 | VectorizerParams::VectorizationFactor : 1); | ||||
| 1587 | unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ? | ||||
| 1588 | VectorizerParams::VectorizationInterleave : 1); | ||||
| 1589 | // The minimum number of iterations for a vectorized/unrolled version. | ||||
| 1590 | unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U); | ||||
| 1591 | |||||
| 1592 | // It's not vectorizable if the distance is smaller than the minimum distance | ||||
| 1593 | // needed for a vectroized/unrolled version. Vectorizing one iteration in | ||||
| 1594 | // front needs TypeByteSize * Stride. Vectorizing the last iteration needs | ||||
| 1595 | // TypeByteSize (No need to plus the last gap distance). | ||||
| 1596 | // | ||||
| 1597 | // E.g. Assume one char is 1 byte in memory and one int is 4 bytes. | ||||
| 1598 | // foo(int *A) { | ||||
| 1599 | // int *B = (int *)((char *)A + 14); | ||||
| 1600 | // for (i = 0 ; i < 1024 ; i += 2) | ||||
| 1601 | // B[i] = A[i] + 1; | ||||
| 1602 | // } | ||||
| 1603 | // | ||||
| 1604 | // Two accesses in memory (stride is 2): | ||||
| 1605 | // | A[0] | | A[2] | | A[4] | | A[6] | | | ||||
| 1606 | // | B[0] | | B[2] | | B[4] | | ||||
| 1607 | // | ||||
| 1608 | // Distance needs for vectorizing iterations except the last iteration: | ||||
| 1609 | // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4. | ||||
| 1610 | // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4. | ||||
| 1611 | // | ||||
| 1612 | // If MinNumIter is 2, it is vectorizable as the minimum distance needed is | ||||
| 1613 | // 12, which is less than distance. | ||||
| 1614 | // | ||||
| 1615 | // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4), | ||||
| 1616 | // the minimum distance needed is 28, which is greater than distance. It is | ||||
| 1617 | // not safe to do vectorization. | ||||
| 1618 | uint64_t MinDistanceNeeded = | ||||
| 1619 | TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize; | ||||
| 1620 | if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) { | ||||
| 1621 | LLVM_DEBUG(dbgs() << "LAA: Failure because of positive distance "do { } while (false) | ||||
| 1622 | << Distance << '\n')do { } while (false); | ||||
| 1623 | return Dependence::Backward; | ||||
| 1624 | } | ||||
| 1625 | |||||
| 1626 | // Unsafe if the minimum distance needed is greater than max safe distance. | ||||
| 1627 | if (MinDistanceNeeded > MaxSafeDepDistBytes) { | ||||
| 1628 | LLVM_DEBUG(dbgs() << "LAA: Failure because it needs at least "do { } while (false) | ||||
| 1629 | << MinDistanceNeeded << " size in bytes")do { } while (false); | ||||
| 1630 | return Dependence::Backward; | ||||
| 1631 | } | ||||
| 1632 | |||||
| 1633 | // Positive distance bigger than max vectorization factor. | ||||
| 1634 | // FIXME: Should use max factor instead of max distance in bytes, which could | ||||
| 1635 | // not handle different types. | ||||
| 1636 | // E.g. Assume one char is 1 byte in memory and one int is 4 bytes. | ||||
| 1637 | // void foo (int *A, char *B) { | ||||
| 1638 | // for (unsigned i = 0; i < 1024; i++) { | ||||
| 1639 | // A[i+2] = A[i] + 1; | ||||
| 1640 | // B[i+2] = B[i] + 1; | ||||
| 1641 | // } | ||||
| 1642 | // } | ||||
| 1643 | // | ||||
| 1644 | // This case is currently unsafe according to the max safe distance. If we | ||||
| 1645 | // analyze the two accesses on array B, the max safe dependence distance | ||||
| 1646 | // is 2. Then we analyze the accesses on array A, the minimum distance needed | ||||
| 1647 | // is 8, which is less than 2 and forbidden vectorization, But actually | ||||
| 1648 | // both A and B could be vectorized by 2 iterations. | ||||
| 1649 | MaxSafeDepDistBytes = | ||||
| 1650 | std::min(static_cast<uint64_t>(Distance), MaxSafeDepDistBytes); | ||||
| 1651 | |||||
| 1652 | bool IsTrueDataDependence = (!AIsWrite && BIsWrite); | ||||
| 1653 | if (IsTrueDataDependence && EnableForwardingConflictDetection && | ||||
| 1654 | couldPreventStoreLoadForward(Distance, TypeByteSize)) | ||||
| 1655 | return Dependence::BackwardVectorizableButPreventsForwarding; | ||||
| 1656 | |||||
| 1657 | uint64_t MaxVF = MaxSafeDepDistBytes / (TypeByteSize * Stride); | ||||
| 1658 | LLVM_DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()do { } while (false) | ||||
| 1659 | << " with max VF = " << MaxVF << '\n')do { } while (false); | ||||
| 1660 | uint64_t MaxVFInBits = MaxVF * TypeByteSize * 8; | ||||
| 1661 | MaxSafeVectorWidthInBits = std::min(MaxSafeVectorWidthInBits, MaxVFInBits); | ||||
| 1662 | return Dependence::BackwardVectorizable; | ||||
| 1663 | } | ||||
| 1664 | |||||
| 1665 | bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets, | ||||
| 1666 | MemAccessInfoList &CheckDeps, | ||||
| 1667 | const ValueToValueMap &Strides) { | ||||
| 1668 | |||||
| 1669 | MaxSafeDepDistBytes = -1; | ||||
| 1670 | SmallPtrSet<MemAccessInfo, 8> Visited; | ||||
| 1671 | for (MemAccessInfo CurAccess : CheckDeps) { | ||||
| 1672 | if (Visited.count(CurAccess)) | ||||
| 1673 | continue; | ||||
| 1674 | |||||
| 1675 | // Get the relevant memory access set. | ||||
| 1676 | EquivalenceClasses<MemAccessInfo>::iterator I = | ||||
| 1677 | AccessSets.findValue(AccessSets.getLeaderValue(CurAccess)); | ||||
| 1678 | |||||
| 1679 | // Check accesses within this set. | ||||
| 1680 | EquivalenceClasses<MemAccessInfo>::member_iterator AI = | ||||
| 1681 | AccessSets.member_begin(I); | ||||
| 1682 | EquivalenceClasses<MemAccessInfo>::member_iterator AE = | ||||
| 1683 | AccessSets.member_end(); | ||||
| 1684 | |||||
| 1685 | // Check every access pair. | ||||
| 1686 | while (AI != AE) { | ||||
| 1687 | Visited.insert(*AI); | ||||
| 1688 | bool AIIsWrite = AI->getInt(); | ||||
| 1689 | // Check loads only against next equivalent class, but stores also against | ||||
| 1690 | // other stores in the same equivalence class - to the same address. | ||||
| 1691 | EquivalenceClasses<MemAccessInfo>::member_iterator OI = | ||||
| 1692 | (AIIsWrite ? AI : std::next(AI)); | ||||
| 1693 | while (OI != AE) { | ||||
| 1694 | // Check every accessing instruction pair in program order. | ||||
| 1695 | for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(), | ||||
| 1696 | I1E = Accesses[*AI].end(); I1 != I1E; ++I1) | ||||
| 1697 | // Scan all accesses of another equivalence class, but only the next | ||||
| 1698 | // accesses of the same equivalent class. | ||||
| 1699 | for (std::vector<unsigned>::iterator | ||||
| 1700 | I2 = (OI == AI ? std::next(I1) : Accesses[*OI].begin()), | ||||
| 1701 | I2E = (OI == AI ? I1E : Accesses[*OI].end()); | ||||
| 1702 | I2 != I2E; ++I2) { | ||||
| 1703 | auto A = std::make_pair(&*AI, *I1); | ||||
| 1704 | auto B = std::make_pair(&*OI, *I2); | ||||
| 1705 | |||||
| 1706 | assert(*I1 != *I2)((void)0); | ||||
| 1707 | if (*I1 > *I2) | ||||
| 1708 | std::swap(A, B); | ||||
| 1709 | |||||
| 1710 | Dependence::DepType Type = | ||||
| 1711 | isDependent(*A.first, A.second, *B.first, B.second, Strides); | ||||
| 1712 | mergeInStatus(Dependence::isSafeForVectorization(Type)); | ||||
| 1713 | |||||
| 1714 | // Gather dependences unless we accumulated MaxDependences | ||||
| 1715 | // dependences. In that case return as soon as we find the first | ||||
| 1716 | // unsafe dependence. This puts a limit on this quadratic | ||||
| 1717 | // algorithm. | ||||
| 1718 | if (RecordDependences) { | ||||
| 1719 | if (Type != Dependence::NoDep) | ||||
| 1720 | Dependences.push_back(Dependence(A.second, B.second, Type)); | ||||
| 1721 | |||||
| 1722 | if (Dependences.size() >= MaxDependences) { | ||||
| 1723 | RecordDependences = false; | ||||
| 1724 | Dependences.clear(); | ||||
| 1725 | LLVM_DEBUG(dbgs()do { } while (false) | ||||
| 1726 | << "Too many dependences, stopped recording\n")do { } while (false); | ||||
| 1727 | } | ||||
| 1728 | } | ||||
| 1729 | if (!RecordDependences && !isSafeForVectorization()) | ||||
| 1730 | return false; | ||||
| 1731 | } | ||||
| 1732 | ++OI; | ||||
| 1733 | } | ||||
| 1734 | AI++; | ||||
| 1735 | } | ||||
| 1736 | } | ||||
| 1737 | |||||
| 1738 | LLVM_DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n")do { } while (false); | ||||
| 1739 | return isSafeForVectorization(); | ||||
| 1740 | } | ||||
| 1741 | |||||
| 1742 | SmallVector<Instruction *, 4> | ||||
| 1743 | MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const { | ||||
| 1744 | MemAccessInfo Access(Ptr, isWrite); | ||||
| 1745 | auto &IndexVector = Accesses.find(Access)->second; | ||||
| 1746 | |||||
| 1747 | SmallVector<Instruction *, 4> Insts; | ||||
| 1748 | transform(IndexVector, | ||||
| 1749 | std::back_inserter(Insts), | ||||
| 1750 | [&](unsigned Idx) { return this->InstMap[Idx]; }); | ||||
| 1751 | return Insts; | ||||
| 1752 | } | ||||
| 1753 | |||||
| 1754 | const char *MemoryDepChecker::Dependence::DepName[] = { | ||||
| 1755 | "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward", | ||||
| 1756 | "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"}; | ||||
| 1757 | |||||
| 1758 | void MemoryDepChecker::Dependence::print( | ||||
| 1759 | raw_ostream &OS, unsigned Depth, | ||||
| 1760 | const SmallVectorImpl<Instruction *> &Instrs) const { | ||||
| 1761 | OS.indent(Depth) << DepName[Type] << ":\n"; | ||||
| 1762 | OS.indent(Depth + 2) << *Instrs[Source] << " -> \n"; | ||||
| 1763 | OS.indent(Depth + 2) << *Instrs[Destination] << "\n"; | ||||
| 1764 | } | ||||
| 1765 | |||||
| 1766 | bool LoopAccessInfo::canAnalyzeLoop() { | ||||
| 1767 | // We need to have a loop header. | ||||
| 1768 | LLVM_DEBUG(dbgs() << "LAA: Found a loop in "do { } while (false) | ||||
| 1769 | << TheLoop->getHeader()->getParent()->getName() << ": "do { } while (false) | ||||
| 1770 | << TheLoop->getHeader()->getName() << '\n')do { } while (false); | ||||
| 1771 | |||||
| 1772 | // We can only analyze innermost loops. | ||||
| 1773 | if (!TheLoop->isInnermost()) { | ||||
| 1774 | LLVM_DEBUG(dbgs() << "LAA: loop is not the innermost loop\n")do { } while (false); | ||||
| 1775 | recordAnalysis("NotInnerMostLoop") << "loop is not the innermost loop"; | ||||
| 1776 | return false; | ||||
| 1777 | } | ||||
| 1778 | |||||
| 1779 | // We must have a single backedge. | ||||
| 1780 | if (TheLoop->getNumBackEdges() != 1) { | ||||
| 1781 | LLVM_DEBUG(do { } while (false) | ||||
| 1782 | dbgs() << "LAA: loop control flow is not understood by analyzer\n")do { } while (false); | ||||
| 1783 | recordAnalysis("CFGNotUnderstood") | ||||
| 1784 | << "loop control flow is not understood by analyzer"; | ||||
| 1785 | return false; | ||||
| 1786 | } | ||||
| 1787 | |||||
| 1788 | // ScalarEvolution needs to be able to find the exit count. | ||||
| 1789 | const SCEV *ExitCount = PSE->getBackedgeTakenCount(); | ||||
| 1790 | if (isa<SCEVCouldNotCompute>(ExitCount)) { | ||||
| 1791 | recordAnalysis("CantComputeNumberOfIterations") | ||||
| 1792 | << "could not determine number of loop iterations"; | ||||
| 1793 | LLVM_DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n")do { } while (false); | ||||
| 1794 | return false; | ||||
| 1795 | } | ||||
| 1796 | |||||
| 1797 | return true; | ||||
| 1798 | } | ||||
| 1799 | |||||
| 1800 | void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI, | ||||
| 1801 | const TargetLibraryInfo *TLI, | ||||
| 1802 | DominatorTree *DT) { | ||||
| 1803 | typedef SmallPtrSet<Value*, 16> ValueSet; | ||||
| 1804 | |||||
| 1805 | // Holds the Load and Store instructions. | ||||
| 1806 | SmallVector<LoadInst *, 16> Loads; | ||||
| 1807 | SmallVector<StoreInst *, 16> Stores; | ||||
| 1808 | |||||
| 1809 | // Holds all the different accesses in the loop. | ||||
| 1810 | unsigned NumReads = 0; | ||||
| 1811 | unsigned NumReadWrites = 0; | ||||
| 1812 | |||||
| 1813 | bool HasComplexMemInst = false; | ||||
| 1814 | |||||
| 1815 | // A runtime check is only legal to insert if there are no convergent calls. | ||||
| 1816 | HasConvergentOp = false; | ||||
| 1817 | |||||
| 1818 | PtrRtChecking->Pointers.clear(); | ||||
| 1819 | PtrRtChecking->Need = false; | ||||
| 1820 | |||||
| 1821 | const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel(); | ||||
| 1822 | |||||
| 1823 | const bool EnableMemAccessVersioningOfLoop = | ||||
| 1824 | EnableMemAccessVersioning && | ||||
| |||||
| 1825 | !TheLoop->getHeader()->getParent()->hasOptSize(); | ||||
| 1826 | |||||
| 1827 | // For each block. | ||||
| 1828 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||
| 1829 | // Scan the BB and collect legal loads and stores. Also detect any | ||||
| 1830 | // convergent instructions. | ||||
| 1831 | for (Instruction &I : *BB) { | ||||
| 1832 | if (auto *Call = dyn_cast<CallBase>(&I)) { | ||||
| 1833 | if (Call->isConvergent()) | ||||
| 1834 | HasConvergentOp = true; | ||||
| 1835 | } | ||||
| 1836 | |||||
| 1837 | // With both a non-vectorizable memory instruction and a convergent | ||||
| 1838 | // operation, found in this loop, no reason to continue the search. | ||||
| 1839 | if (HasComplexMemInst && HasConvergentOp) { | ||||
| 1840 | CanVecMem = false; | ||||
| 1841 | return; | ||||
| 1842 | } | ||||
| 1843 | |||||
| 1844 | // Avoid hitting recordAnalysis multiple times. | ||||
| 1845 | if (HasComplexMemInst) | ||||
| 1846 | continue; | ||||
| 1847 | |||||
| 1848 | // If this is a load, save it. If this instruction can read from memory | ||||
| 1849 | // but is not a load, then we quit. Notice that we don't handle function | ||||
| 1850 | // calls that read or write. | ||||
| 1851 | if (I.mayReadFromMemory()) { | ||||
| 1852 | // Many math library functions read the rounding mode. We will only | ||||
| 1853 | // vectorize a loop if it contains known function calls that don't set | ||||
| 1854 | // the flag. Therefore, it is safe to ignore this read from memory. | ||||
| 1855 | auto *Call = dyn_cast<CallInst>(&I); | ||||
| 1856 | if (Call && getVectorIntrinsicIDForCall(Call, TLI)) | ||||
| 1857 | continue; | ||||
| 1858 | |||||
| 1859 | // If the function has an explicit vectorized counterpart, we can safely | ||||
| 1860 | // assume that it can be vectorized. | ||||
| 1861 | if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() && | ||||
| 1862 | !VFDatabase::getMappings(*Call).empty()) | ||||
| 1863 | continue; | ||||
| 1864 | |||||
| 1865 | auto *Ld = dyn_cast<LoadInst>(&I); | ||||
| 1866 | if (!Ld) { | ||||
| 1867 | recordAnalysis("CantVectorizeInstruction", Ld) | ||||
| 1868 | << "instruction cannot be vectorized"; | ||||
| 1869 | HasComplexMemInst = true; | ||||
| 1870 | continue; | ||||
| 1871 | } | ||||
| 1872 | if (!Ld->isSimple() && !IsAnnotatedParallel) { | ||||
| 1873 | recordAnalysis("NonSimpleLoad", Ld) | ||||
| 1874 | << "read with atomic ordering or volatile read"; | ||||
| 1875 | LLVM_DEBUG(dbgs() << "LAA: Found a non-simple load.\n")do { } while (false); | ||||
| 1876 | HasComplexMemInst = true; | ||||
| 1877 | continue; | ||||
| 1878 | } | ||||
| 1879 | NumLoads++; | ||||
| 1880 | Loads.push_back(Ld); | ||||
| 1881 | DepChecker->addAccess(Ld); | ||||
| 1882 | if (EnableMemAccessVersioningOfLoop) | ||||
| 1883 | collectStridedAccess(Ld); | ||||
| 1884 | continue; | ||||
| 1885 | } | ||||
| 1886 | |||||
| 1887 | // Save 'store' instructions. Abort if other instructions write to memory. | ||||
| 1888 | if (I.mayWriteToMemory()) { | ||||
| 1889 | auto *St = dyn_cast<StoreInst>(&I); | ||||
| 1890 | if (!St) { | ||||
| 1891 | recordAnalysis("CantVectorizeInstruction", St) | ||||
| 1892 | << "instruction cannot be vectorized"; | ||||
| 1893 | HasComplexMemInst = true; | ||||
| 1894 | continue; | ||||
| 1895 | } | ||||
| 1896 | if (!St->isSimple() && !IsAnnotatedParallel) { | ||||
| 1897 | recordAnalysis("NonSimpleStore", St) | ||||
| 1898 | << "write with atomic ordering or volatile write"; | ||||
| 1899 | LLVM_DEBUG(dbgs() << "LAA: Found a non-simple store.\n")do { } while (false); | ||||
| 1900 | HasComplexMemInst = true; | ||||
| 1901 | continue; | ||||
| 1902 | } | ||||
| 1903 | NumStores++; | ||||
| 1904 | Stores.push_back(St); | ||||
| 1905 | DepChecker->addAccess(St); | ||||
| 1906 | if (EnableMemAccessVersioningOfLoop) | ||||
| 1907 | collectStridedAccess(St); | ||||
| 1908 | } | ||||
| 1909 | } // Next instr. | ||||
| 1910 | } // Next block. | ||||
| 1911 | |||||
| 1912 | if (HasComplexMemInst
| ||||
| 1913 | CanVecMem = false; | ||||
| 1914 | return; | ||||
| 1915 | } | ||||
| 1916 | |||||
| 1917 | // Now we have two lists that hold the loads and the stores. | ||||
| 1918 | // Next, we find the pointers that they use. | ||||
| 1919 | |||||
| 1920 | // Check if we see any stores. If there are no stores, then we don't | ||||
| 1921 | // care if the pointers are *restrict*. | ||||
| 1922 | if (!Stores.size()) { | ||||
| 1923 | LLVM_DEBUG(dbgs() << "LAA: Found a read-only loop!\n")do { } while (false); | ||||
| 1924 | CanVecMem = true; | ||||
| 1925 | return; | ||||
| 1926 | } | ||||
| 1927 | |||||
| 1928 | MemoryDepChecker::DepCandidates DependentAccesses; | ||||
| 1929 | AccessAnalysis Accesses(TheLoop, AA, LI, DependentAccesses, *PSE); | ||||
| 1930 | |||||
| 1931 | // Holds the analyzed pointers. We don't want to call getUnderlyingObjects | ||||
| 1932 | // multiple times on the same object. If the ptr is accessed twice, once | ||||
| 1933 | // for read and once for write, it will only appear once (on the write | ||||
| 1934 | // list). This is okay, since we are going to check for conflicts between | ||||
| 1935 | // writes and between reads and writes, but not between reads and reads. | ||||
| 1936 | ValueSet Seen; | ||||
| 1937 | |||||
| 1938 | // Record uniform store addresses to identify if we have multiple stores | ||||
| 1939 | // to the same address. | ||||
| 1940 | ValueSet UniformStores; | ||||
| 1941 | |||||
| 1942 | for (StoreInst *ST : Stores) { | ||||
| 1943 | Value *Ptr = ST->getPointerOperand(); | ||||
| 1944 | |||||
| 1945 | if (isUniform(Ptr)) | ||||
| 1946 | HasDependenceInvolvingLoopInvariantAddress |= | ||||
| 1947 | !UniformStores.insert(Ptr).second; | ||||
| 1948 | |||||
| 1949 | // If we did *not* see this pointer before, insert it to the read-write | ||||
| 1950 | // list. At this phase it is only a 'write' list. | ||||
| 1951 | if (Seen.insert(Ptr).second) { | ||||
| 1952 | ++NumReadWrites; | ||||
| 1953 | |||||
| 1954 | MemoryLocation Loc = MemoryLocation::get(ST); | ||||
| 1955 | // The TBAA metadata could have a control dependency on the predication | ||||
| 1956 | // condition, so we cannot rely on it when determining whether or not we | ||||
| 1957 | // need runtime pointer checks. | ||||
| 1958 | if (blockNeedsPredication(ST->getParent(), TheLoop, DT)) | ||||
| 1959 | Loc.AATags.TBAA = nullptr; | ||||
| 1960 | |||||
| 1961 | Accesses.addStore(Loc); | ||||
| 1962 | } | ||||
| 1963 | } | ||||
| 1964 | |||||
| 1965 | if (IsAnnotatedParallel) { | ||||
| 1966 | LLVM_DEBUG(do { } while (false) | ||||
| 1967 | dbgs() << "LAA: A loop annotated parallel, ignore memory dependency "do { } while (false) | ||||
| 1968 | << "checks.\n")do { } while (false); | ||||
| 1969 | CanVecMem = true; | ||||
| 1970 | return; | ||||
| 1971 | } | ||||
| 1972 | |||||
| 1973 | for (LoadInst *LD : Loads) { | ||||
| 1974 | Value *Ptr = LD->getPointerOperand(); | ||||
| 1975 | // If we did *not* see this pointer before, insert it to the | ||||
| 1976 | // read list. If we *did* see it before, then it is already in | ||||
| 1977 | // the read-write list. This allows us to vectorize expressions | ||||
| 1978 | // such as A[i] += x; Because the address of A[i] is a read-write | ||||
| 1979 | // pointer. This only works if the index of A[i] is consecutive. | ||||
| 1980 | // If the address of i is unknown (for example A[B[i]]) then we may | ||||
| 1981 | // read a few words, modify, and write a few words, and some of the | ||||
| 1982 | // words may be written to the same address. | ||||
| 1983 | bool IsReadOnlyPtr = false; | ||||
| 1984 | if (Seen.insert(Ptr).second || | ||||
| 1985 | !getPtrStride(*PSE, Ptr, TheLoop, SymbolicStrides)) { | ||||
| 1986 | ++NumReads; | ||||
| 1987 | IsReadOnlyPtr = true; | ||||
| 1988 | } | ||||
| 1989 | |||||
| 1990 | // See if there is an unsafe dependency between a load to a uniform address and | ||||
| 1991 | // store to the same uniform address. | ||||
| 1992 | if (UniformStores.count(Ptr)) { | ||||
| 1993 | LLVM_DEBUG(dbgs() << "LAA: Found an unsafe dependency between a uniform "do { } while (false) | ||||
| 1994 | "load and uniform store to the same address!\n")do { } while (false); | ||||
| 1995 | HasDependenceInvolvingLoopInvariantAddress = true; | ||||
| 1996 | } | ||||
| 1997 | |||||
| 1998 | MemoryLocation Loc = MemoryLocation::get(LD); | ||||
| 1999 | // The TBAA metadata could have a control dependency on the predication | ||||
| 2000 | // condition, so we cannot rely on it when determining whether or not we | ||||
| 2001 | // need runtime pointer checks. | ||||
| 2002 | if (blockNeedsPredication(LD->getParent(), TheLoop, DT)) | ||||
| 2003 | Loc.AATags.TBAA = nullptr; | ||||
| 2004 | |||||
| 2005 | Accesses.addLoad(Loc, IsReadOnlyPtr); | ||||
| 2006 | } | ||||
| 2007 | |||||
| 2008 | // If we write (or read-write) to a single destination and there are no | ||||
| 2009 | // other reads in this loop then is it safe to vectorize. | ||||
| 2010 | if (NumReadWrites
| ||||
| 2011 | LLVM_DEBUG(dbgs() << "LAA: Found a write-only loop!\n")do { } while (false); | ||||
| 2012 | CanVecMem = true; | ||||
| 2013 | return; | ||||
| 2014 | } | ||||
| 2015 | |||||
| 2016 | // Build dependence sets and check whether we need a runtime pointer bounds | ||||
| 2017 | // check. | ||||
| 2018 | Accesses.buildDependenceSets(); | ||||
| 2019 | |||||
| 2020 | // Find pointers with computable bounds. We are going to use this information | ||||
| 2021 | // to place a runtime bound check. | ||||
| 2022 | bool CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(), | ||||
| 2023 | TheLoop, SymbolicStrides); | ||||
| 2024 | if (!CanDoRTIfNeeded) { | ||||
| 2025 | recordAnalysis("CantIdentifyArrayBounds") << "cannot identify array bounds"; | ||||
| 2026 | LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "do { } while (false) | ||||
| 2027 | << "the array bounds.\n")do { } while (false); | ||||
| 2028 | CanVecMem = false; | ||||
| 2029 | return; | ||||
| 2030 | } | ||||
| 2031 | |||||
| 2032 | LLVM_DEBUG(do { } while (false) | ||||
| 2033 | dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n")do { } while (false); | ||||
| 2034 | |||||
| 2035 | CanVecMem = true; | ||||
| 2036 | if (Accesses.isDependencyCheckNeeded()) { | ||||
| 2037 | LLVM_DEBUG(dbgs() << "LAA: Checking memory dependencies\n")do { } while (false); | ||||
| 2038 | CanVecMem = DepChecker->areDepsSafe( | ||||
| 2039 | DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides); | ||||
| 2040 | MaxSafeDepDistBytes = DepChecker->getMaxSafeDepDistBytes(); | ||||
| 2041 | |||||
| 2042 | if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) { | ||||
| 2043 | LLVM_DEBUG(dbgs() << "LAA: Retrying with memory checks\n")do { } while (false); | ||||
| 2044 | |||||
| 2045 | // Clear the dependency checks. We assume they are not needed. | ||||
| 2046 | Accesses.resetDepChecks(*DepChecker); | ||||
| 2047 | |||||
| 2048 | PtrRtChecking->reset(); | ||||
| 2049 | PtrRtChecking->Need = true; | ||||
| 2050 | |||||
| 2051 | auto *SE = PSE->getSE(); | ||||
| 2052 | CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, SE, TheLoop, | ||||
| 2053 | SymbolicStrides, true); | ||||
| 2054 | |||||
| 2055 | // Check that we found the bounds for the pointer. | ||||
| 2056 | if (!CanDoRTIfNeeded) { | ||||
| 2057 | recordAnalysis("CantCheckMemDepsAtRunTime") | ||||
| 2058 | << "cannot check memory dependencies at runtime"; | ||||
| 2059 | LLVM_DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n")do { } while (false); | ||||
| 2060 | CanVecMem = false; | ||||
| 2061 | return; | ||||
| 2062 | } | ||||
| 2063 | |||||
| 2064 | CanVecMem = true; | ||||
| 2065 | } | ||||
| 2066 | } | ||||
| 2067 | |||||
| 2068 | if (HasConvergentOp) { | ||||
| 2069 | recordAnalysis("CantInsertRuntimeCheckWithConvergent") | ||||
| 2070 | << "cannot add control dependency to convergent operation"; | ||||
| 2071 | LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because a runtime check "do { } while (false) | ||||
| 2072 | "would be needed with a convergent operation\n")do { } while (false); | ||||
| 2073 | CanVecMem = false; | ||||
| 2074 | return; | ||||
| 2075 | } | ||||
| 2076 | |||||
| 2077 | if (CanVecMem) | ||||
| 2078 | LLVM_DEBUG(do { } while (false) | ||||
| 2079 | dbgs() << "LAA: No unsafe dependent memory operations in loop. We"do { } while (false) | ||||
| 2080 | << (PtrRtChecking->Need ? "" : " don't")do { } while (false) | ||||
| 2081 | << " need runtime memory checks.\n")do { } while (false); | ||||
| 2082 | else { | ||||
| 2083 | recordAnalysis("UnsafeMemDep") | ||||
| 2084 | << "unsafe dependent memory operations in loop. Use " | ||||
| 2085 | "#pragma loop distribute(enable) to allow loop distribution " | ||||
| 2086 | "to attempt to isolate the offending operations into a separate " | ||||
| 2087 | "loop"; | ||||
| 2088 | LLVM_DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n")do { } while (false); | ||||
| 2089 | } | ||||
| 2090 | } | ||||
| 2091 | |||||
| 2092 | bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop, | ||||
| 2093 | DominatorTree *DT) { | ||||
| 2094 | assert(TheLoop->contains(BB) && "Unknown block used")((void)0); | ||||
| 2095 | |||||
| 2096 | // Blocks that do not dominate the latch need predication. | ||||
| 2097 | BasicBlock* Latch = TheLoop->getLoopLatch(); | ||||
| 2098 | return !DT->dominates(BB, Latch); | ||||
| 2099 | } | ||||
| 2100 | |||||
| 2101 | OptimizationRemarkAnalysis &LoopAccessInfo::recordAnalysis(StringRef RemarkName, | ||||
| 2102 | Instruction *I) { | ||||
| 2103 | assert(!Report && "Multiple reports generated")((void)0); | ||||
| 2104 | |||||
| 2105 | Value *CodeRegion = TheLoop->getHeader(); | ||||
| 2106 | DebugLoc DL = TheLoop->getStartLoc(); | ||||
| 2107 | |||||
| 2108 | if (I) { | ||||
| 2109 | CodeRegion = I->getParent(); | ||||
| 2110 | // If there is no debug location attached to the instruction, revert back to | ||||
| 2111 | // using the loop's. | ||||
| 2112 | if (I->getDebugLoc()) | ||||
| 2113 | DL = I->getDebugLoc(); | ||||
| 2114 | } | ||||
| 2115 | |||||
| 2116 | Report = std::make_unique<OptimizationRemarkAnalysis>(DEBUG_TYPE"loop-accesses", RemarkName, DL, | ||||
| 2117 | CodeRegion); | ||||
| 2118 | return *Report; | ||||
| 2119 | } | ||||
| 2120 | |||||
| 2121 | bool LoopAccessInfo::isUniform(Value *V) const { | ||||
| 2122 | auto *SE = PSE->getSE(); | ||||
| 2123 | // Since we rely on SCEV for uniformity, if the type is not SCEVable, it is | ||||
| 2124 | // never considered uniform. | ||||
| 2125 | // TODO: Is this really what we want? Even without FP SCEV, we may want some | ||||
| 2126 | // trivially loop-invariant FP values to be considered uniform. | ||||
| 2127 | if (!SE->isSCEVable(V->getType())) | ||||
| 2128 | return false; | ||||
| 2129 | return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); | ||||
| 2130 | } | ||||
| 2131 | |||||
| 2132 | void LoopAccessInfo::collectStridedAccess(Value *MemAccess) { | ||||
| 2133 | Value *Ptr = getLoadStorePointerOperand(MemAccess); | ||||
| 2134 | if (!Ptr) | ||||
| 2135 | return; | ||||
| 2136 | |||||
| 2137 | Value *Stride = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop); | ||||
| 2138 | if (!Stride) | ||||
| 2139 | return; | ||||
| 2140 | |||||
| 2141 | LLVM_DEBUG(dbgs() << "LAA: Found a strided access that is a candidate for "do { } while (false) | ||||
| 2142 | "versioning:")do { } while (false); | ||||
| 2143 | LLVM_DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n")do { } while (false); | ||||
| 2144 | |||||
| 2145 | // Avoid adding the "Stride == 1" predicate when we know that | ||||
| 2146 | // Stride >= Trip-Count. Such a predicate will effectively optimize a single | ||||
| 2147 | // or zero iteration loop, as Trip-Count <= Stride == 1. | ||||
| 2148 | // | ||||
| 2149 | // TODO: We are currently not making a very informed decision on when it is | ||||
| 2150 | // beneficial to apply stride versioning. It might make more sense that the | ||||
| 2151 | // users of this analysis (such as the vectorizer) will trigger it, based on | ||||
| 2152 | // their specific cost considerations; For example, in cases where stride | ||||
| 2153 | // versioning does not help resolving memory accesses/dependences, the | ||||
| 2154 | // vectorizer should evaluate the cost of the runtime test, and the benefit | ||||
| 2155 | // of various possible stride specializations, considering the alternatives | ||||
| 2156 | // of using gather/scatters (if available). | ||||
| 2157 | |||||
| 2158 | const SCEV *StrideExpr = PSE->getSCEV(Stride); | ||||
| 2159 | const SCEV *BETakenCount = PSE->getBackedgeTakenCount(); | ||||
| 2160 | |||||
| 2161 | // Match the types so we can compare the stride and the BETakenCount. | ||||
| 2162 | // The Stride can be positive/negative, so we sign extend Stride; | ||||
| 2163 | // The backedgeTakenCount is non-negative, so we zero extend BETakenCount. | ||||
| 2164 | const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout(); | ||||
| 2165 | uint64_t StrideTypeSize = DL.getTypeAllocSize(StrideExpr->getType()); | ||||
| 2166 | uint64_t BETypeSize = DL.getTypeAllocSize(BETakenCount->getType()); | ||||
| 2167 | const SCEV *CastedStride = StrideExpr; | ||||
| 2168 | const SCEV *CastedBECount = BETakenCount; | ||||
| 2169 | ScalarEvolution *SE = PSE->getSE(); | ||||
| 2170 | if (BETypeSize >= StrideTypeSize) | ||||
| 2171 | CastedStride = SE->getNoopOrSignExtend(StrideExpr, BETakenCount->getType()); | ||||
| 2172 | else | ||||
| 2173 | CastedBECount = SE->getZeroExtendExpr(BETakenCount, StrideExpr->getType()); | ||||
| 2174 | const SCEV *StrideMinusBETaken = SE->getMinusSCEV(CastedStride, CastedBECount); | ||||
| 2175 | // Since TripCount == BackEdgeTakenCount + 1, checking: | ||||
| 2176 | // "Stride >= TripCount" is equivalent to checking: | ||||
| 2177 | // Stride - BETakenCount > 0 | ||||
| 2178 | if (SE->isKnownPositive(StrideMinusBETaken)) { | ||||
| 2179 | LLVM_DEBUG(do { } while (false) | ||||
| 2180 | dbgs() << "LAA: Stride>=TripCount; No point in versioning as the "do { } while (false) | ||||
| 2181 | "Stride==1 predicate will imply that the loop executes "do { } while (false) | ||||
| 2182 | "at most once.\n")do { } while (false); | ||||
| 2183 | return; | ||||
| 2184 | } | ||||
| 2185 | LLVM_DEBUG(dbgs() << "LAA: Found a strided access that we can version.")do { } while (false); | ||||
| 2186 | |||||
| 2187 | SymbolicStrides[Ptr] = Stride; | ||||
| 2188 | StrideSet.insert(Stride); | ||||
| 2189 | } | ||||
| 2190 | |||||
| 2191 | LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE, | ||||
| 2192 | const TargetLibraryInfo *TLI, AAResults *AA, | ||||
| 2193 | DominatorTree *DT, LoopInfo *LI) | ||||
| 2194 | : PSE(std::make_unique<PredicatedScalarEvolution>(*SE, *L)), | ||||
| 2195 | PtrRtChecking(std::make_unique<RuntimePointerChecking>(SE)), | ||||
| 2196 | DepChecker(std::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L), | ||||
| 2197 | NumLoads(0), NumStores(0), MaxSafeDepDistBytes(-1), CanVecMem(false), | ||||
| 2198 | HasConvergentOp(false), | ||||
| 2199 | HasDependenceInvolvingLoopInvariantAddress(false) { | ||||
| 2200 | if (canAnalyzeLoop()) | ||||
| 2201 | analyzeLoop(AA, LI, TLI, DT); | ||||
| 2202 | } | ||||
| 2203 | |||||
| 2204 | void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const { | ||||
| 2205 | if (CanVecMem) { | ||||
| 2206 | OS.indent(Depth) << "Memory dependences are safe"; | ||||
| 2207 | if (MaxSafeDepDistBytes != -1ULL) | ||||
| 2208 | OS << " with a maximum dependence distance of " << MaxSafeDepDistBytes | ||||
| 2209 | << " bytes"; | ||||
| 2210 | if (PtrRtChecking->Need) | ||||
| 2211 | OS << " with run-time checks"; | ||||
| 2212 | OS << "\n"; | ||||
| 2213 | } | ||||
| 2214 | |||||
| 2215 | if (HasConvergentOp) | ||||
| 2216 | OS.indent(Depth) << "Has convergent operation in loop\n"; | ||||
| 2217 | |||||
| 2218 | if (Report) | ||||
| 2219 | OS.indent(Depth) << "Report: " << Report->getMsg() << "\n"; | ||||
| 2220 | |||||
| 2221 | if (auto *Dependences = DepChecker->getDependences()) { | ||||
| 2222 | OS.indent(Depth) << "Dependences:\n"; | ||||
| 2223 | for (auto &Dep : *Dependences) { | ||||
| 2224 | Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions()); | ||||
| 2225 | OS << "\n"; | ||||
| 2226 | } | ||||
| 2227 | } else | ||||
| 2228 | OS.indent(Depth) << "Too many dependences, not recorded\n"; | ||||
| 2229 | |||||
| 2230 | // List the pair of accesses need run-time checks to prove independence. | ||||
| 2231 | PtrRtChecking->print(OS, Depth); | ||||
| 2232 | OS << "\n"; | ||||
| 2233 | |||||
| 2234 | OS.indent(Depth) << "Non vectorizable stores to invariant address were " | ||||
| 2235 | << (HasDependenceInvolvingLoopInvariantAddress ? "" : "not ") | ||||
| 2236 | << "found in loop.\n"; | ||||
| 2237 | |||||
| 2238 | OS.indent(Depth) << "SCEV assumptions:\n"; | ||||
| 2239 | PSE->getUnionPredicate().print(OS, Depth); | ||||
| 2240 | |||||
| 2241 | OS << "\n"; | ||||
| 2242 | |||||
| 2243 | OS.indent(Depth) << "Expressions re-written:\n"; | ||||
| 2244 | PSE->print(OS, Depth); | ||||
| 2245 | } | ||||
| 2246 | |||||
| 2247 | LoopAccessLegacyAnalysis::LoopAccessLegacyAnalysis() : FunctionPass(ID) { | ||||
| 2248 | initializeLoopAccessLegacyAnalysisPass(*PassRegistry::getPassRegistry()); | ||||
| 2249 | } | ||||
| 2250 | |||||
| 2251 | const LoopAccessInfo &LoopAccessLegacyAnalysis::getInfo(Loop *L) { | ||||
| 2252 | auto &LAI = LoopAccessInfoMap[L]; | ||||
| 2253 | |||||
| 2254 | if (!LAI) | ||||
| 2255 | LAI = std::make_unique<LoopAccessInfo>(L, SE, TLI, AA, DT, LI); | ||||
| 2256 | |||||
| 2257 | return *LAI.get(); | ||||
| 2258 | } | ||||
| 2259 | |||||
| 2260 | void LoopAccessLegacyAnalysis::print(raw_ostream &OS, const Module *M) const { | ||||
| 2261 | LoopAccessLegacyAnalysis &LAA = *const_cast<LoopAccessLegacyAnalysis *>(this); | ||||
| 2262 | |||||
| 2263 | for (Loop *TopLevelLoop : *LI) | ||||
| 2264 | for (Loop *L : depth_first(TopLevelLoop)) { | ||||
| 2265 | OS.indent(2) << L->getHeader()->getName() << ":\n"; | ||||
| 2266 | auto &LAI = LAA.getInfo(L); | ||||
| 2267 | LAI.print(OS, 4); | ||||
| 2268 | } | ||||
| 2269 | } | ||||
| 2270 | |||||
| 2271 | bool LoopAccessLegacyAnalysis::runOnFunction(Function &F) { | ||||
| 2272 | SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | ||||
| 2273 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | ||||
| 2274 | TLI = TLIP ? &TLIP->getTLI(F) : nullptr; | ||||
| 2275 | AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | ||||
| 2276 | DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||
| 2277 | LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||
| 2278 | |||||
| 2279 | return false; | ||||
| 2280 | } | ||||
| 2281 | |||||
| 2282 | void LoopAccessLegacyAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { | ||||
| 2283 | AU.addRequiredTransitive<ScalarEvolutionWrapperPass>(); | ||||
| 2284 | AU.addRequiredTransitive<AAResultsWrapperPass>(); | ||||
| 2285 | AU.addRequiredTransitive<DominatorTreeWrapperPass>(); | ||||
| 2286 | AU.addRequiredTransitive<LoopInfoWrapperPass>(); | ||||
| 2287 | |||||
| 2288 | AU.setPreservesAll(); | ||||
| 2289 | } | ||||
| 2290 | |||||
| 2291 | char LoopAccessLegacyAnalysis::ID = 0; | ||||
| 2292 | static const char laa_name[] = "Loop Access Analysis"; | ||||
| 2293 | #define LAA_NAME"loop-accesses" "loop-accesses" | ||||
| 2294 | |||||
| 2295 | INITIALIZE_PASS_BEGIN(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)static void *initializeLoopAccessLegacyAnalysisPassOnce(PassRegistry &Registry) { | ||||
| 2296 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | ||||
| 2297 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | ||||
| 2298 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | ||||
| 2299 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); | ||||
| 2300 | INITIALIZE_PASS_END(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)PassInfo *PI = new PassInfo( laa_name, "loop-accesses", & LoopAccessLegacyAnalysis::ID, PassInfo::NormalCtor_t(callDefaultCtor <LoopAccessLegacyAnalysis>), false, true); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeLoopAccessLegacyAnalysisPassFlag ; void llvm::initializeLoopAccessLegacyAnalysisPass(PassRegistry &Registry) { llvm::call_once(InitializeLoopAccessLegacyAnalysisPassFlag , initializeLoopAccessLegacyAnalysisPassOnce, std::ref(Registry )); } | ||||
| 2301 | |||||
| 2302 | AnalysisKey LoopAccessAnalysis::Key; | ||||
| 2303 | |||||
| 2304 | LoopAccessInfo LoopAccessAnalysis::run(Loop &L, LoopAnalysisManager &AM, | ||||
| 2305 | LoopStandardAnalysisResults &AR) { | ||||
| 2306 | return LoopAccessInfo(&L, &AR.SE, &AR.TLI, &AR.AA, &AR.DT, &AR.LI); | ||||
| 2307 | } | ||||
| 2308 | |||||
| 2309 | namespace llvm { | ||||
| 2310 | |||||
| 2311 | Pass *createLAAPass() { | ||||
| 2312 | return new LoopAccessLegacyAnalysis(); | ||||
| 2313 | } | ||||
| 2314 | |||||
| 2315 | } // end namespace llvm |