| File: | llvm/lib/Analysis/LoopAccessAnalysis.cpp |
| Warning: | line 1341, column 38 Although the value stored to 'VF' is used in the enclosing expression, the value is never actually read from 'VF' |
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 { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *Expr << "\n"; } } while (false) |
| 167 | << " by: " << *Expr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Replacing SCEV: " << *OrigSCEV << " by: " << *Expr << "\n"; } } while (false); |
| 168 | return Expr; |
| 169 | } |
| 170 | |
| 171 | RuntimeCheckingPtrGroup::RuntimeCheckingPtrGroup( |
| 172 | unsigned Index, RuntimePointerChecking &RtCheck) |
| 173 | : RtCheck(RtCheck), High(RtCheck.Pointers[Index].End), |
| 174 | Low(RtCheck.Pointers[Index].Start) { |
| 175 | Members.push_back(Index); |
| 176 | } |
| 177 | |
| 178 | /// Calculate Start and End points of memory access. |
| 179 | /// Let's assume A is the first access and B is a memory access on N-th loop |
| 180 | /// iteration. Then B is calculated as: |
| 181 | /// B = A + Step*N . |
| 182 | /// Step value may be positive or negative. |
| 183 | /// N is a calculated back-edge taken count: |
| 184 | /// N = (TripCount > 0) ? RoundDown(TripCount -1 , VF) : 0 |
| 185 | /// Start and End points are calculated in the following way: |
| 186 | /// Start = UMIN(A, B) ; End = UMAX(A, B) + SizeOfElt, |
| 187 | /// where SizeOfElt is the size of single memory access in bytes. |
| 188 | /// |
| 189 | /// There is no conflict when the intervals are disjoint: |
| 190 | /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End) |
| 191 | void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, bool WritePtr, |
| 192 | unsigned DepSetId, unsigned ASId, |
| 193 | const ValueToValueMap &Strides, |
| 194 | PredicatedScalarEvolution &PSE) { |
| 195 | // Get the stride replaced scev. |
| 196 | const SCEV *Sc = replaceSymbolicStrideSCEV(PSE, Strides, Ptr); |
| 197 | ScalarEvolution *SE = PSE.getSE(); |
| 198 | |
| 199 | const SCEV *ScStart; |
| 200 | const SCEV *ScEnd; |
| 201 | |
| 202 | if (SE->isLoopInvariant(Sc, Lp)) |
| 203 | ScStart = ScEnd = Sc; |
| 204 | else { |
| 205 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc); |
| 206 | assert(AR && "Invalid addrec expression")((AR && "Invalid addrec expression") ? static_cast< void> (0) : __assert_fail ("AR && \"Invalid addrec expression\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 206, __PRETTY_FUNCTION__)); |
| 207 | const SCEV *Ex = PSE.getBackedgeTakenCount(); |
| 208 | |
| 209 | ScStart = AR->getStart(); |
| 210 | ScEnd = AR->evaluateAtIteration(Ex, *SE); |
| 211 | const SCEV *Step = AR->getStepRecurrence(*SE); |
| 212 | |
| 213 | // For expressions with negative step, the upper bound is ScStart and the |
| 214 | // lower bound is ScEnd. |
| 215 | if (const auto *CStep = dyn_cast<SCEVConstant>(Step)) { |
| 216 | if (CStep->getValue()->isNegative()) |
| 217 | std::swap(ScStart, ScEnd); |
| 218 | } else { |
| 219 | // Fallback case: the step is not constant, but we can still |
| 220 | // get the upper and lower bounds of the interval by using min/max |
| 221 | // expressions. |
| 222 | ScStart = SE->getUMinExpr(ScStart, ScEnd); |
| 223 | ScEnd = SE->getUMaxExpr(AR->getStart(), ScEnd); |
| 224 | } |
| 225 | // Add the size of the pointed element to ScEnd. |
| 226 | auto &DL = Lp->getHeader()->getModule()->getDataLayout(); |
| 227 | Type *IdxTy = DL.getIndexType(Ptr->getType()); |
| 228 | const SCEV *EltSizeSCEV = |
| 229 | SE->getStoreSizeOfExpr(IdxTy, Ptr->getType()->getPointerElementType()); |
| 230 | ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV); |
| 231 | } |
| 232 | |
| 233 | Pointers.emplace_back(Ptr, ScStart, ScEnd, WritePtr, DepSetId, ASId, Sc); |
| 234 | } |
| 235 | |
| 236 | SmallVector<RuntimePointerCheck, 4> |
| 237 | RuntimePointerChecking::generateChecks() const { |
| 238 | SmallVector<RuntimePointerCheck, 4> Checks; |
| 239 | |
| 240 | for (unsigned I = 0; I < CheckingGroups.size(); ++I) { |
| 241 | for (unsigned J = I + 1; J < CheckingGroups.size(); ++J) { |
| 242 | const RuntimeCheckingPtrGroup &CGI = CheckingGroups[I]; |
| 243 | const RuntimeCheckingPtrGroup &CGJ = CheckingGroups[J]; |
| 244 | |
| 245 | if (needsChecking(CGI, CGJ)) |
| 246 | Checks.push_back(std::make_pair(&CGI, &CGJ)); |
| 247 | } |
| 248 | } |
| 249 | return Checks; |
| 250 | } |
| 251 | |
| 252 | void RuntimePointerChecking::generateChecks( |
| 253 | MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) { |
| 254 | assert(Checks.empty() && "Checks is not empty")((Checks.empty() && "Checks is not empty") ? static_cast <void> (0) : __assert_fail ("Checks.empty() && \"Checks is not empty\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 254, __PRETTY_FUNCTION__)); |
| 255 | groupChecks(DepCands, UseDependencies); |
| 256 | Checks = generateChecks(); |
| 257 | } |
| 258 | |
| 259 | bool RuntimePointerChecking::needsChecking( |
| 260 | const RuntimeCheckingPtrGroup &M, const RuntimeCheckingPtrGroup &N) const { |
| 261 | for (unsigned I = 0, EI = M.Members.size(); EI != I; ++I) |
| 262 | for (unsigned J = 0, EJ = N.Members.size(); EJ != J; ++J) |
| 263 | if (needsChecking(M.Members[I], N.Members[J])) |
| 264 | return true; |
| 265 | return false; |
| 266 | } |
| 267 | |
| 268 | /// Compare \p I and \p J and return the minimum. |
| 269 | /// Return nullptr in case we couldn't find an answer. |
| 270 | static const SCEV *getMinFromExprs(const SCEV *I, const SCEV *J, |
| 271 | ScalarEvolution *SE) { |
| 272 | const SCEV *Diff = SE->getMinusSCEV(J, I); |
| 273 | const SCEVConstant *C = dyn_cast<const SCEVConstant>(Diff); |
| 274 | |
| 275 | if (!C) |
| 276 | return nullptr; |
| 277 | if (C->getValue()->isNegative()) |
| 278 | return J; |
| 279 | return I; |
| 280 | } |
| 281 | |
| 282 | bool RuntimeCheckingPtrGroup::addPointer(unsigned Index) { |
| 283 | const SCEV *Start = RtCheck.Pointers[Index].Start; |
| 284 | const SCEV *End = RtCheck.Pointers[Index].End; |
| 285 | |
| 286 | // Compare the starts and ends with the known minimum and maximum |
| 287 | // of this set. We need to know how we compare against the min/max |
| 288 | // of the set in order to be able to emit memchecks. |
| 289 | const SCEV *Min0 = getMinFromExprs(Start, Low, RtCheck.SE); |
| 290 | if (!Min0) |
| 291 | return false; |
| 292 | |
| 293 | const SCEV *Min1 = getMinFromExprs(End, High, RtCheck.SE); |
| 294 | if (!Min1) |
| 295 | return false; |
| 296 | |
| 297 | // Update the low bound expression if we've found a new min value. |
| 298 | if (Min0 == Start) |
| 299 | Low = Start; |
| 300 | |
| 301 | // Update the high bound expression if we've found a new max value. |
| 302 | if (Min1 != End) |
| 303 | High = End; |
| 304 | |
| 305 | Members.push_back(Index); |
| 306 | return true; |
| 307 | } |
| 308 | |
| 309 | void RuntimePointerChecking::groupChecks( |
| 310 | MemoryDepChecker::DepCandidates &DepCands, bool UseDependencies) { |
| 311 | // We build the groups from dependency candidates equivalence classes |
| 312 | // because: |
| 313 | // - We know that pointers in the same equivalence class share |
| 314 | // the same underlying object and therefore there is a chance |
| 315 | // that we can compare pointers |
| 316 | // - We wouldn't be able to merge two pointers for which we need |
| 317 | // to emit a memcheck. The classes in DepCands are already |
| 318 | // conveniently built such that no two pointers in the same |
| 319 | // class need checking against each other. |
| 320 | |
| 321 | // We use the following (greedy) algorithm to construct the groups |
| 322 | // For every pointer in the equivalence class: |
| 323 | // For each existing group: |
| 324 | // - if the difference between this pointer and the min/max bounds |
| 325 | // of the group is a constant, then make the pointer part of the |
| 326 | // group and update the min/max bounds of that group as required. |
| 327 | |
| 328 | CheckingGroups.clear(); |
| 329 | |
| 330 | // If we need to check two pointers to the same underlying object |
| 331 | // with a non-constant difference, we shouldn't perform any pointer |
| 332 | // grouping with those pointers. This is because we can easily get |
| 333 | // into cases where the resulting check would return false, even when |
| 334 | // the accesses are safe. |
| 335 | // |
| 336 | // The following example shows this: |
| 337 | // for (i = 0; i < 1000; ++i) |
| 338 | // a[5000 + i * m] = a[i] + a[i + 9000] |
| 339 | // |
| 340 | // Here grouping gives a check of (5000, 5000 + 1000 * m) against |
| 341 | // (0, 10000) which is always false. However, if m is 1, there is no |
| 342 | // dependence. Not grouping the checks for a[i] and a[i + 9000] allows |
| 343 | // us to perform an accurate check in this case. |
| 344 | // |
| 345 | // The above case requires that we have an UnknownDependence between |
| 346 | // accesses to the same underlying object. This cannot happen unless |
| 347 | // FoundNonConstantDistanceDependence is set, and therefore UseDependencies |
| 348 | // is also false. In this case we will use the fallback path and create |
| 349 | // separate checking groups for all pointers. |
| 350 | |
| 351 | // If we don't have the dependency partitions, construct a new |
| 352 | // checking pointer group for each pointer. This is also required |
| 353 | // for correctness, because in this case we can have checking between |
| 354 | // pointers to the same underlying object. |
| 355 | if (!UseDependencies) { |
| 356 | for (unsigned I = 0; I < Pointers.size(); ++I) |
| 357 | CheckingGroups.push_back(RuntimeCheckingPtrGroup(I, *this)); |
| 358 | return; |
| 359 | } |
| 360 | |
| 361 | unsigned TotalComparisons = 0; |
| 362 | |
| 363 | DenseMap<Value *, unsigned> PositionMap; |
| 364 | for (unsigned Index = 0; Index < Pointers.size(); ++Index) |
| 365 | PositionMap[Pointers[Index].PointerValue] = Index; |
| 366 | |
| 367 | // We need to keep track of what pointers we've already seen so we |
| 368 | // don't process them twice. |
| 369 | SmallSet<unsigned, 2> Seen; |
| 370 | |
| 371 | // Go through all equivalence classes, get the "pointer check groups" |
| 372 | // and add them to the overall solution. We use the order in which accesses |
| 373 | // appear in 'Pointers' to enforce determinism. |
| 374 | for (unsigned I = 0; I < Pointers.size(); ++I) { |
| 375 | // We've seen this pointer before, and therefore already processed |
| 376 | // its equivalence class. |
| 377 | if (Seen.count(I)) |
| 378 | continue; |
| 379 | |
| 380 | MemoryDepChecker::MemAccessInfo Access(Pointers[I].PointerValue, |
| 381 | Pointers[I].IsWritePtr); |
| 382 | |
| 383 | SmallVector<RuntimeCheckingPtrGroup, 2> Groups; |
| 384 | auto LeaderI = DepCands.findValue(DepCands.getLeaderValue(Access)); |
| 385 | |
| 386 | // Because DepCands is constructed by visiting accesses in the order in |
| 387 | // which they appear in alias sets (which is deterministic) and the |
| 388 | // iteration order within an equivalence class member is only dependent on |
| 389 | // the order in which unions and insertions are performed on the |
| 390 | // equivalence class, the iteration order is deterministic. |
| 391 | for (auto MI = DepCands.member_begin(LeaderI), ME = DepCands.member_end(); |
| 392 | MI != ME; ++MI) { |
| 393 | auto PointerI = PositionMap.find(MI->getPointer()); |
| 394 | assert(PointerI != PositionMap.end() &&((PointerI != PositionMap.end() && "pointer in equivalence class not found in PositionMap" ) ? static_cast<void> (0) : __assert_fail ("PointerI != PositionMap.end() && \"pointer in equivalence class not found in PositionMap\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 395, __PRETTY_FUNCTION__)) |
| 395 | "pointer in equivalence class not found in PositionMap")((PointerI != PositionMap.end() && "pointer in equivalence class not found in PositionMap" ) ? static_cast<void> (0) : __assert_fail ("PointerI != PositionMap.end() && \"pointer in equivalence class not found in PositionMap\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 395, __PRETTY_FUNCTION__)); |
| 396 | unsigned Pointer = PointerI->second; |
| 397 | bool Merged = false; |
| 398 | // Mark this pointer as seen. |
| 399 | Seen.insert(Pointer); |
| 400 | |
| 401 | // Go through all the existing sets and see if we can find one |
| 402 | // which can include this pointer. |
| 403 | for (RuntimeCheckingPtrGroup &Group : Groups) { |
| 404 | // Don't perform more than a certain amount of comparisons. |
| 405 | // This should limit the cost of grouping the pointers to something |
| 406 | // reasonable. If we do end up hitting this threshold, the algorithm |
| 407 | // will create separate groups for all remaining pointers. |
| 408 | if (TotalComparisons > MemoryCheckMergeThreshold) |
| 409 | break; |
| 410 | |
| 411 | TotalComparisons++; |
| 412 | |
| 413 | if (Group.addPointer(Pointer)) { |
| 414 | Merged = true; |
| 415 | break; |
| 416 | } |
| 417 | } |
| 418 | |
| 419 | if (!Merged) |
| 420 | // We couldn't add this pointer to any existing set or the threshold |
| 421 | // for the number of comparisons has been reached. Create a new group |
| 422 | // to hold the current pointer. |
| 423 | Groups.push_back(RuntimeCheckingPtrGroup(Pointer, *this)); |
| 424 | } |
| 425 | |
| 426 | // We've computed the grouped checks for this partition. |
| 427 | // Save the results and continue with the next one. |
| 428 | llvm::copy(Groups, std::back_inserter(CheckingGroups)); |
| 429 | } |
| 430 | } |
| 431 | |
| 432 | bool RuntimePointerChecking::arePointersInSamePartition( |
| 433 | const SmallVectorImpl<int> &PtrToPartition, unsigned PtrIdx1, |
| 434 | unsigned PtrIdx2) { |
| 435 | return (PtrToPartition[PtrIdx1] != -1 && |
| 436 | PtrToPartition[PtrIdx1] == PtrToPartition[PtrIdx2]); |
| 437 | } |
| 438 | |
| 439 | bool RuntimePointerChecking::needsChecking(unsigned I, unsigned J) const { |
| 440 | const PointerInfo &PointerI = Pointers[I]; |
| 441 | const PointerInfo &PointerJ = Pointers[J]; |
| 442 | |
| 443 | // No need to check if two readonly pointers intersect. |
| 444 | if (!PointerI.IsWritePtr && !PointerJ.IsWritePtr) |
| 445 | return false; |
| 446 | |
| 447 | // Only need to check pointers between two different dependency sets. |
| 448 | if (PointerI.DependencySetId == PointerJ.DependencySetId) |
| 449 | return false; |
| 450 | |
| 451 | // Only need to check pointers in the same alias set. |
| 452 | if (PointerI.AliasSetId != PointerJ.AliasSetId) |
| 453 | return false; |
| 454 | |
| 455 | return true; |
| 456 | } |
| 457 | |
| 458 | void RuntimePointerChecking::printChecks( |
| 459 | raw_ostream &OS, const SmallVectorImpl<RuntimePointerCheck> &Checks, |
| 460 | unsigned Depth) const { |
| 461 | unsigned N = 0; |
| 462 | for (const auto &Check : Checks) { |
| 463 | const auto &First = Check.first->Members, &Second = Check.second->Members; |
| 464 | |
| 465 | OS.indent(Depth) << "Check " << N++ << ":\n"; |
| 466 | |
| 467 | OS.indent(Depth + 2) << "Comparing group (" << Check.first << "):\n"; |
| 468 | for (unsigned K = 0; K < First.size(); ++K) |
| 469 | OS.indent(Depth + 2) << *Pointers[First[K]].PointerValue << "\n"; |
| 470 | |
| 471 | OS.indent(Depth + 2) << "Against group (" << Check.second << "):\n"; |
| 472 | for (unsigned K = 0; K < Second.size(); ++K) |
| 473 | OS.indent(Depth + 2) << *Pointers[Second[K]].PointerValue << "\n"; |
| 474 | } |
| 475 | } |
| 476 | |
| 477 | void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const { |
| 478 | |
| 479 | OS.indent(Depth) << "Run-time memory checks:\n"; |
| 480 | printChecks(OS, Checks, Depth); |
| 481 | |
| 482 | OS.indent(Depth) << "Grouped accesses:\n"; |
| 483 | for (unsigned I = 0; I < CheckingGroups.size(); ++I) { |
| 484 | const auto &CG = CheckingGroups[I]; |
| 485 | |
| 486 | OS.indent(Depth + 2) << "Group " << &CG << ":\n"; |
| 487 | OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High |
| 488 | << ")\n"; |
| 489 | for (unsigned J = 0; J < CG.Members.size(); ++J) { |
| 490 | OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr |
| 491 | << "\n"; |
| 492 | } |
| 493 | } |
| 494 | } |
| 495 | |
| 496 | namespace { |
| 497 | |
| 498 | /// Analyses memory accesses in a loop. |
| 499 | /// |
| 500 | /// Checks whether run time pointer checks are needed and builds sets for data |
| 501 | /// dependence checking. |
| 502 | class AccessAnalysis { |
| 503 | public: |
| 504 | /// Read or write access location. |
| 505 | typedef PointerIntPair<Value *, 1, bool> MemAccessInfo; |
| 506 | typedef SmallVector<MemAccessInfo, 8> MemAccessInfoList; |
| 507 | |
| 508 | AccessAnalysis(Loop *TheLoop, AAResults *AA, LoopInfo *LI, |
| 509 | MemoryDepChecker::DepCandidates &DA, |
| 510 | PredicatedScalarEvolution &PSE) |
| 511 | : TheLoop(TheLoop), AST(*AA), LI(LI), DepCands(DA), |
| 512 | IsRTCheckAnalysisNeeded(false), PSE(PSE) {} |
| 513 | |
| 514 | /// Register a load and whether it is only read from. |
| 515 | void addLoad(MemoryLocation &Loc, bool IsReadOnly) { |
| 516 | Value *Ptr = const_cast<Value*>(Loc.Ptr); |
| 517 | AST.add(Ptr, LocationSize::beforeOrAfterPointer(), Loc.AATags); |
| 518 | Accesses.insert(MemAccessInfo(Ptr, false)); |
| 519 | if (IsReadOnly) |
| 520 | ReadOnlyPtr.insert(Ptr); |
| 521 | } |
| 522 | |
| 523 | /// Register a store. |
| 524 | void addStore(MemoryLocation &Loc) { |
| 525 | Value *Ptr = const_cast<Value*>(Loc.Ptr); |
| 526 | AST.add(Ptr, LocationSize::beforeOrAfterPointer(), Loc.AATags); |
| 527 | Accesses.insert(MemAccessInfo(Ptr, true)); |
| 528 | } |
| 529 | |
| 530 | /// Check if we can emit a run-time no-alias check for \p Access. |
| 531 | /// |
| 532 | /// Returns true if we can emit a run-time no alias check for \p Access. |
| 533 | /// If we can check this access, this also adds it to a dependence set and |
| 534 | /// adds a run-time to check for it to \p RtCheck. If \p Assume is true, |
| 535 | /// we will attempt to use additional run-time checks in order to get |
| 536 | /// the bounds of the pointer. |
| 537 | bool createCheckForAccess(RuntimePointerChecking &RtCheck, |
| 538 | MemAccessInfo Access, |
| 539 | const ValueToValueMap &Strides, |
| 540 | DenseMap<Value *, unsigned> &DepSetId, |
| 541 | Loop *TheLoop, unsigned &RunningDepId, |
| 542 | unsigned ASId, bool ShouldCheckStride, |
| 543 | bool Assume); |
| 544 | |
| 545 | /// Check whether we can check the pointers at runtime for |
| 546 | /// non-intersection. |
| 547 | /// |
| 548 | /// Returns true if we need no check or if we do and we can generate them |
| 549 | /// (i.e. the pointers have computable bounds). |
| 550 | bool canCheckPtrAtRT(RuntimePointerChecking &RtCheck, ScalarEvolution *SE, |
| 551 | Loop *TheLoop, const ValueToValueMap &Strides, |
| 552 | bool ShouldCheckWrap = false); |
| 553 | |
| 554 | /// Goes over all memory accesses, checks whether a RT check is needed |
| 555 | /// and builds sets of dependent accesses. |
| 556 | void buildDependenceSets() { |
| 557 | processMemAccesses(); |
| 558 | } |
| 559 | |
| 560 | /// Initial processing of memory accesses determined that we need to |
| 561 | /// perform dependency checking. |
| 562 | /// |
| 563 | /// Note that this can later be cleared if we retry memcheck analysis without |
| 564 | /// dependency checking (i.e. FoundNonConstantDistanceDependence). |
| 565 | bool isDependencyCheckNeeded() { return !CheckDeps.empty(); } |
| 566 | |
| 567 | /// We decided that no dependence analysis would be used. Reset the state. |
| 568 | void resetDepChecks(MemoryDepChecker &DepChecker) { |
| 569 | CheckDeps.clear(); |
| 570 | DepChecker.clearDependences(); |
| 571 | } |
| 572 | |
| 573 | MemAccessInfoList &getDependenciesToCheck() { return CheckDeps; } |
| 574 | |
| 575 | private: |
| 576 | typedef SetVector<MemAccessInfo> PtrAccessSet; |
| 577 | |
| 578 | /// Go over all memory access and check whether runtime pointer checks |
| 579 | /// are needed and build sets of dependency check candidates. |
| 580 | void processMemAccesses(); |
| 581 | |
| 582 | /// Set of all accesses. |
| 583 | PtrAccessSet Accesses; |
| 584 | |
| 585 | /// The loop being checked. |
| 586 | const Loop *TheLoop; |
| 587 | |
| 588 | /// List of accesses that need a further dependence check. |
| 589 | MemAccessInfoList CheckDeps; |
| 590 | |
| 591 | /// Set of pointers that are read only. |
| 592 | SmallPtrSet<Value*, 16> ReadOnlyPtr; |
| 593 | |
| 594 | /// An alias set tracker to partition the access set by underlying object and |
| 595 | //intrinsic property (such as TBAA metadata). |
| 596 | AliasSetTracker AST; |
| 597 | |
| 598 | LoopInfo *LI; |
| 599 | |
| 600 | /// Sets of potentially dependent accesses - members of one set share an |
| 601 | /// underlying pointer. The set "CheckDeps" identfies which sets really need a |
| 602 | /// dependence check. |
| 603 | MemoryDepChecker::DepCandidates &DepCands; |
| 604 | |
| 605 | /// Initial processing of memory accesses determined that we may need |
| 606 | /// to add memchecks. Perform the analysis to determine the necessary checks. |
| 607 | /// |
| 608 | /// Note that, this is different from isDependencyCheckNeeded. When we retry |
| 609 | /// memcheck analysis without dependency checking |
| 610 | /// (i.e. FoundNonConstantDistanceDependence), isDependencyCheckNeeded is |
| 611 | /// cleared while this remains set if we have potentially dependent accesses. |
| 612 | bool IsRTCheckAnalysisNeeded; |
| 613 | |
| 614 | /// The SCEV predicate containing all the SCEV-related assumptions. |
| 615 | PredicatedScalarEvolution &PSE; |
| 616 | }; |
| 617 | |
| 618 | } // end anonymous namespace |
| 619 | |
| 620 | /// Check whether a pointer can participate in a runtime bounds check. |
| 621 | /// If \p Assume, try harder to prove that we can compute the bounds of \p Ptr |
| 622 | /// by adding run-time checks (overflow checks) if necessary. |
| 623 | static bool hasComputableBounds(PredicatedScalarEvolution &PSE, |
| 624 | const ValueToValueMap &Strides, Value *Ptr, |
| 625 | Loop *L, bool Assume) { |
| 626 | const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, Strides, Ptr); |
| 627 | |
| 628 | // The bounds for loop-invariant pointer is trivial. |
| 629 | if (PSE.getSE()->isLoopInvariant(PtrScev, L)) |
| 630 | return true; |
| 631 | |
| 632 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); |
| 633 | |
| 634 | if (!AR && Assume) |
| 635 | AR = PSE.getAsAddRec(Ptr); |
| 636 | |
| 637 | if (!AR) |
| 638 | return false; |
| 639 | |
| 640 | return AR->isAffine(); |
| 641 | } |
| 642 | |
| 643 | /// Check whether a pointer address cannot wrap. |
| 644 | static bool isNoWrap(PredicatedScalarEvolution &PSE, |
| 645 | const ValueToValueMap &Strides, Value *Ptr, Loop *L) { |
| 646 | const SCEV *PtrScev = PSE.getSCEV(Ptr); |
| 647 | if (PSE.getSE()->isLoopInvariant(PtrScev, L)) |
| 648 | return true; |
| 649 | |
| 650 | int64_t Stride = getPtrStride(PSE, Ptr, L, Strides); |
| 651 | if (Stride == 1 || PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW)) |
| 652 | return true; |
| 653 | |
| 654 | return false; |
| 655 | } |
| 656 | |
| 657 | bool AccessAnalysis::createCheckForAccess(RuntimePointerChecking &RtCheck, |
| 658 | MemAccessInfo Access, |
| 659 | const ValueToValueMap &StridesMap, |
| 660 | DenseMap<Value *, unsigned> &DepSetId, |
| 661 | Loop *TheLoop, unsigned &RunningDepId, |
| 662 | unsigned ASId, bool ShouldCheckWrap, |
| 663 | bool Assume) { |
| 664 | Value *Ptr = Access.getPointer(); |
| 665 | |
| 666 | if (!hasComputableBounds(PSE, StridesMap, Ptr, TheLoop, Assume)) |
| 667 | return false; |
| 668 | |
| 669 | // When we run after a failing dependency check we have to make sure |
| 670 | // we don't have wrapping pointers. |
| 671 | if (ShouldCheckWrap && !isNoWrap(PSE, StridesMap, Ptr, TheLoop)) { |
| 672 | auto *Expr = PSE.getSCEV(Ptr); |
| 673 | if (!Assume || !isa<SCEVAddRecExpr>(Expr)) |
| 674 | return false; |
| 675 | PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); |
| 676 | } |
| 677 | |
| 678 | // The id of the dependence set. |
| 679 | unsigned DepId; |
| 680 | |
| 681 | if (isDependencyCheckNeeded()) { |
| 682 | Value *Leader = DepCands.getLeaderValue(Access).getPointer(); |
| 683 | unsigned &LeaderId = DepSetId[Leader]; |
| 684 | if (!LeaderId) |
| 685 | LeaderId = RunningDepId++; |
| 686 | DepId = LeaderId; |
| 687 | } else |
| 688 | // Each access has its own dependence set. |
| 689 | DepId = RunningDepId++; |
| 690 | |
| 691 | bool IsWrite = Access.getInt(); |
| 692 | RtCheck.insert(TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap, PSE); |
| 693 | LLVM_DEBUG(dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a runtime check ptr:" << *Ptr << '\n'; } } while (false); |
| 694 | |
| 695 | return true; |
| 696 | } |
| 697 | |
| 698 | bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck, |
| 699 | ScalarEvolution *SE, Loop *TheLoop, |
| 700 | const ValueToValueMap &StridesMap, |
| 701 | bool ShouldCheckWrap) { |
| 702 | // Find pointers with computable bounds. We are going to use this information |
| 703 | // to place a runtime bound check. |
| 704 | bool CanDoRT = true; |
| 705 | |
| 706 | bool MayNeedRTCheck = false; |
| 707 | if (!IsRTCheckAnalysisNeeded) return true; |
| 708 | |
| 709 | bool IsDepCheckNeeded = isDependencyCheckNeeded(); |
| 710 | |
| 711 | // We assign a consecutive id to access from different alias sets. |
| 712 | // Accesses between different groups doesn't need to be checked. |
| 713 | unsigned ASId = 0; |
| 714 | for (auto &AS : AST) { |
| 715 | int NumReadPtrChecks = 0; |
| 716 | int NumWritePtrChecks = 0; |
| 717 | bool CanDoAliasSetRT = true; |
| 718 | ++ASId; |
| 719 | |
| 720 | // We assign consecutive id to access from different dependence sets. |
| 721 | // Accesses within the same set don't need a runtime check. |
| 722 | unsigned RunningDepId = 1; |
| 723 | DenseMap<Value *, unsigned> DepSetId; |
| 724 | |
| 725 | SmallVector<MemAccessInfo, 4> Retries; |
| 726 | |
| 727 | // First, count how many write and read accesses are in the alias set. Also |
| 728 | // collect MemAccessInfos for later. |
| 729 | SmallVector<MemAccessInfo, 4> AccessInfos; |
| 730 | for (const auto &A : AS) { |
| 731 | Value *Ptr = A.getValue(); |
| 732 | bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true)); |
| 733 | |
| 734 | if (IsWrite) |
| 735 | ++NumWritePtrChecks; |
| 736 | else |
| 737 | ++NumReadPtrChecks; |
| 738 | AccessInfos.emplace_back(Ptr, IsWrite); |
| 739 | } |
| 740 | |
| 741 | // We do not need runtime checks for this alias set, if there are no writes |
| 742 | // or a single write and no reads. |
| 743 | if (NumWritePtrChecks == 0 || |
| 744 | (NumWritePtrChecks == 1 && NumReadPtrChecks == 0)) { |
| 745 | assert((AS.size() <= 1 ||(((AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue( AccessWrite) == DepCands.end(); })) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? static_cast<void > (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 752, __PRETTY_FUNCTION__)) |
| 746 | all_of(AS,(((AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue( AccessWrite) == DepCands.end(); })) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? static_cast<void > (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 752, __PRETTY_FUNCTION__)) |
| 747 | [this](auto AC) {(((AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue( AccessWrite) == DepCands.end(); })) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? static_cast<void > (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 752, __PRETTY_FUNCTION__)) |
| 748 | MemAccessInfo AccessWrite(AC.getValue(), true);(((AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue( AccessWrite) == DepCands.end(); })) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? static_cast<void > (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 752, __PRETTY_FUNCTION__)) |
| 749 | return DepCands.findValue(AccessWrite) == DepCands.end();(((AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue( AccessWrite) == DepCands.end(); })) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? static_cast<void > (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 752, __PRETTY_FUNCTION__)) |
| 750 | })) &&(((AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue( AccessWrite) == DepCands.end(); })) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? static_cast<void > (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 752, __PRETTY_FUNCTION__)) |
| 751 | "Can only skip updating CanDoRT below, if all entries in AS "(((AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue( AccessWrite) == DepCands.end(); })) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? static_cast<void > (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 752, __PRETTY_FUNCTION__)) |
| 752 | "are reads or there is at most 1 entry")(((AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue( AccessWrite) == DepCands.end(); })) && "Can only skip updating CanDoRT below, if all entries in AS " "are reads or there is at most 1 entry") ? static_cast<void > (0) : __assert_fail ("(AS.size() <= 1 || all_of(AS, [this](auto AC) { MemAccessInfo AccessWrite(AC.getValue(), true); return DepCands.findValue(AccessWrite) == DepCands.end(); })) && \"Can only skip updating CanDoRT below, if all entries in AS \" \"are reads or there is at most 1 entry\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 752, __PRETTY_FUNCTION__)); |
| 753 | continue; |
| 754 | } |
| 755 | |
| 756 | for (auto &Access : AccessInfos) { |
| 757 | if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId, TheLoop, |
| 758 | RunningDepId, ASId, ShouldCheckWrap, false)) { |
| 759 | LLVM_DEBUG(dbgs() << "LAA: Can't find bounds for ptr:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Can't find bounds for ptr:" << *Access.getPointer() << '\n'; } } while (false ) |
| 760 | << *Access.getPointer() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Can't find bounds for ptr:" << *Access.getPointer() << '\n'; } } while (false ); |
| 761 | Retries.push_back(Access); |
| 762 | CanDoAliasSetRT = false; |
| 763 | } |
| 764 | } |
| 765 | |
| 766 | // Note that this function computes CanDoRT and MayNeedRTCheck |
| 767 | // independently. For example CanDoRT=false, MayNeedRTCheck=false means that |
| 768 | // we have a pointer for which we couldn't find the bounds but we don't |
| 769 | // actually need to emit any checks so it does not matter. |
| 770 | // |
| 771 | // We need runtime checks for this alias set, if there are at least 2 |
| 772 | // dependence sets (in which case RunningDepId > 2) or if we need to re-try |
| 773 | // any bound checks (because in that case the number of dependence sets is |
| 774 | // incomplete). |
| 775 | bool NeedsAliasSetRTCheck = RunningDepId > 2 || !Retries.empty(); |
| 776 | |
| 777 | // We need to perform run-time alias checks, but some pointers had bounds |
| 778 | // that couldn't be checked. |
| 779 | if (NeedsAliasSetRTCheck && !CanDoAliasSetRT) { |
| 780 | // Reset the CanDoSetRt flag and retry all accesses that have failed. |
| 781 | // We know that we need these checks, so we can now be more aggressive |
| 782 | // and add further checks if required (overflow checks). |
| 783 | CanDoAliasSetRT = true; |
| 784 | for (auto Access : Retries) |
| 785 | if (!createCheckForAccess(RtCheck, Access, StridesMap, DepSetId, |
| 786 | TheLoop, RunningDepId, ASId, |
| 787 | ShouldCheckWrap, /*Assume=*/true)) { |
| 788 | CanDoAliasSetRT = false; |
| 789 | break; |
| 790 | } |
| 791 | } |
| 792 | |
| 793 | CanDoRT &= CanDoAliasSetRT; |
| 794 | MayNeedRTCheck |= NeedsAliasSetRTCheck; |
| 795 | ++ASId; |
| 796 | } |
| 797 | |
| 798 | // If the pointers that we would use for the bounds comparison have different |
| 799 | // address spaces, assume the values aren't directly comparable, so we can't |
| 800 | // use them for the runtime check. We also have to assume they could |
| 801 | // overlap. In the future there should be metadata for whether address spaces |
| 802 | // are disjoint. |
| 803 | unsigned NumPointers = RtCheck.Pointers.size(); |
| 804 | for (unsigned i = 0; i < NumPointers; ++i) { |
| 805 | for (unsigned j = i + 1; j < NumPointers; ++j) { |
| 806 | // Only need to check pointers between two different dependency sets. |
| 807 | if (RtCheck.Pointers[i].DependencySetId == |
| 808 | RtCheck.Pointers[j].DependencySetId) |
| 809 | continue; |
| 810 | // Only need to check pointers in the same alias set. |
| 811 | if (RtCheck.Pointers[i].AliasSetId != RtCheck.Pointers[j].AliasSetId) |
| 812 | continue; |
| 813 | |
| 814 | Value *PtrI = RtCheck.Pointers[i].PointerValue; |
| 815 | Value *PtrJ = RtCheck.Pointers[j].PointerValue; |
| 816 | |
| 817 | unsigned ASi = PtrI->getType()->getPointerAddressSpace(); |
| 818 | unsigned ASj = PtrJ->getType()->getPointerAddressSpace(); |
| 819 | if (ASi != ASj) { |
| 820 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Runtime check would require comparison between" " different address spaces\n"; } } while (false) |
| 821 | dbgs() << "LAA: Runtime check would require comparison between"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Runtime check would require comparison between" " different address spaces\n"; } } while (false) |
| 822 | " different address spaces\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Runtime check would require comparison between" " different address spaces\n"; } } while (false); |
| 823 | return false; |
| 824 | } |
| 825 | } |
| 826 | } |
| 827 | |
| 828 | if (MayNeedRTCheck && CanDoRT) |
| 829 | RtCheck.generateChecks(DepCands, IsDepCheckNeeded); |
| 830 | |
| 831 | LLVM_DEBUG(dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks() << " pointer comparisons.\n" ; } } while (false) |
| 832 | << " pointer comparisons.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We need to do " << RtCheck.getNumberOfChecks() << " pointer comparisons.\n" ; } } while (false); |
| 833 | |
| 834 | // If we can do run-time checks, but there are no checks, no runtime checks |
| 835 | // are needed. This can happen when all pointers point to the same underlying |
| 836 | // object for example. |
| 837 | RtCheck.Need = CanDoRT ? RtCheck.getNumberOfChecks() != 0 : MayNeedRTCheck; |
| 838 | |
| 839 | bool CanDoRTIfNeeded = !RtCheck.Need || CanDoRT; |
| 840 | if (!CanDoRTIfNeeded) |
| 841 | RtCheck.reset(); |
| 842 | return CanDoRTIfNeeded; |
| 843 | } |
| 844 | |
| 845 | void AccessAnalysis::processMemAccesses() { |
| 846 | // We process the set twice: first we process read-write pointers, last we |
| 847 | // process read-only pointers. This allows us to skip dependence tests for |
| 848 | // read-only pointers. |
| 849 | |
| 850 | LLVM_DEBUG(dbgs() << "LAA: Processing memory accesses...\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Processing memory accesses...\n" ; } } while (false); |
| 851 | LLVM_DEBUG(dbgs() << " AST: "; AST.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << " AST: "; AST.dump(); } } while (false); |
| 852 | LLVM_DEBUG(dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Accesses(" << Accesses.size() << "):\n"; } } while (false); |
| 853 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.getPointer() << " (" << (A.getInt () ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only" : "read")) << ")\n"; }; } } while (false) |
| 854 | for (auto A : Accesses)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.getPointer() << " (" << (A.getInt () ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only" : "read")) << ")\n"; }; } } while (false) |
| 855 | dbgs() << "\t" << *A.getPointer() << " (" <<do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.getPointer() << " (" << (A.getInt () ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only" : "read")) << ")\n"; }; } } while (false) |
| 856 | (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.getPointer() << " (" << (A.getInt () ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only" : "read")) << ")\n"; }; } } while (false) |
| 857 | "read-only" : "read")) << ")\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.getPointer() << " (" << (A.getInt () ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only" : "read")) << ")\n"; }; } } while (false) |
| 858 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { { for (auto A : Accesses) dbgs() << "\t" << *A.getPointer() << " (" << (A.getInt () ? "write" : (ReadOnlyPtr.count(A.getPointer()) ? "read-only" : "read")) << ")\n"; }; } } while (false); |
| 859 | |
| 860 | // The AliasSetTracker has nicely partitioned our pointers by metadata |
| 861 | // compatibility and potential for underlying-object overlap. As a result, we |
| 862 | // only need to check for potential pointer dependencies within each alias |
| 863 | // set. |
| 864 | for (const auto &AS : AST) { |
| 865 | // Note that both the alias-set tracker and the alias sets themselves used |
| 866 | // linked lists internally and so the iteration order here is deterministic |
| 867 | // (matching the original instruction order within each set). |
| 868 | |
| 869 | bool SetHasWrite = false; |
| 870 | |
| 871 | // Map of pointers to last access encountered. |
| 872 | typedef DenseMap<const Value*, MemAccessInfo> UnderlyingObjToAccessMap; |
| 873 | UnderlyingObjToAccessMap ObjToLastAccess; |
| 874 | |
| 875 | // Set of access to check after all writes have been processed. |
| 876 | PtrAccessSet DeferredAccesses; |
| 877 | |
| 878 | // Iterate over each alias set twice, once to process read/write pointers, |
| 879 | // and then to process read-only pointers. |
| 880 | for (int SetIteration = 0; SetIteration < 2; ++SetIteration) { |
| 881 | bool UseDeferred = SetIteration > 0; |
| 882 | PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses; |
| 883 | |
| 884 | for (const auto &AV : AS) { |
| 885 | Value *Ptr = AV.getValue(); |
| 886 | |
| 887 | // For a single memory access in AliasSetTracker, Accesses may contain |
| 888 | // both read and write, and they both need to be handled for CheckDeps. |
| 889 | for (const auto &AC : S) { |
| 890 | if (AC.getPointer() != Ptr) |
| 891 | continue; |
| 892 | |
| 893 | bool IsWrite = AC.getInt(); |
| 894 | |
| 895 | // If we're using the deferred access set, then it contains only |
| 896 | // reads. |
| 897 | bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite; |
| 898 | if (UseDeferred && !IsReadOnlyPtr) |
| 899 | continue; |
| 900 | // Otherwise, the pointer must be in the PtrAccessSet, either as a |
| 901 | // read or a write. |
| 902 | assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||((((IsReadOnlyPtr && UseDeferred) || IsWrite || S.count (MemAccessInfo(Ptr, false))) && "Alias-set pointer not in the access set?" ) ? static_cast<void> (0) : __assert_fail ("((IsReadOnlyPtr && UseDeferred) || IsWrite || S.count(MemAccessInfo(Ptr, false))) && \"Alias-set pointer not in the access set?\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 904, __PRETTY_FUNCTION__)) |
| 903 | S.count(MemAccessInfo(Ptr, false))) &&((((IsReadOnlyPtr && UseDeferred) || IsWrite || S.count (MemAccessInfo(Ptr, false))) && "Alias-set pointer not in the access set?" ) ? static_cast<void> (0) : __assert_fail ("((IsReadOnlyPtr && UseDeferred) || IsWrite || S.count(MemAccessInfo(Ptr, false))) && \"Alias-set pointer not in the access set?\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 904, __PRETTY_FUNCTION__)) |
| 904 | "Alias-set pointer not in the access set?")((((IsReadOnlyPtr && UseDeferred) || IsWrite || S.count (MemAccessInfo(Ptr, false))) && "Alias-set pointer not in the access set?" ) ? static_cast<void> (0) : __assert_fail ("((IsReadOnlyPtr && UseDeferred) || IsWrite || S.count(MemAccessInfo(Ptr, false))) && \"Alias-set pointer not in the access set?\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 904, __PRETTY_FUNCTION__)); |
| 905 | |
| 906 | MemAccessInfo Access(Ptr, IsWrite); |
| 907 | DepCands.insert(Access); |
| 908 | |
| 909 | // Memorize read-only pointers for later processing and skip them in |
| 910 | // the first round (they need to be checked after we have seen all |
| 911 | // write pointers). Note: we also mark pointer that are not |
| 912 | // consecutive as "read-only" pointers (so that we check |
| 913 | // "a[b[i]] +="). Hence, we need the second check for "!IsWrite". |
| 914 | if (!UseDeferred && IsReadOnlyPtr) { |
| 915 | DeferredAccesses.insert(Access); |
| 916 | continue; |
| 917 | } |
| 918 | |
| 919 | // If this is a write - check other reads and writes for conflicts. If |
| 920 | // this is a read only check other writes for conflicts (but only if |
| 921 | // there is no other write to the ptr - this is an optimization to |
| 922 | // catch "a[i] = a[i] + " without having to do a dependence check). |
| 923 | if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) { |
| 924 | CheckDeps.push_back(Access); |
| 925 | IsRTCheckAnalysisNeeded = true; |
| 926 | } |
| 927 | |
| 928 | if (IsWrite) |
| 929 | SetHasWrite = true; |
| 930 | |
| 931 | // Create sets of pointers connected by a shared alias set and |
| 932 | // underlying object. |
| 933 | typedef SmallVector<const Value *, 16> ValueVector; |
| 934 | ValueVector TempObjects; |
| 935 | |
| 936 | getUnderlyingObjects(Ptr, TempObjects, LI); |
| 937 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Underlying objects for pointer " << *Ptr << "\n"; } } while (false) |
| 938 | << "Underlying objects for pointer " << *Ptr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Underlying objects for pointer " << *Ptr << "\n"; } } while (false); |
| 939 | for (const Value *UnderlyingObj : TempObjects) { |
| 940 | // nullptr never alias, don't join sets for pointer that have "null" |
| 941 | // in their UnderlyingObjects list. |
| 942 | if (isa<ConstantPointerNull>(UnderlyingObj) && |
| 943 | !NullPointerIsDefined( |
| 944 | TheLoop->getHeader()->getParent(), |
| 945 | UnderlyingObj->getType()->getPointerAddressSpace())) |
| 946 | continue; |
| 947 | |
| 948 | UnderlyingObjToAccessMap::iterator Prev = |
| 949 | ObjToLastAccess.find(UnderlyingObj); |
| 950 | if (Prev != ObjToLastAccess.end()) |
| 951 | DepCands.unionSets(Access, Prev->second); |
| 952 | |
| 953 | ObjToLastAccess[UnderlyingObj] = Access; |
| 954 | LLVM_DEBUG(dbgs() << " " << *UnderlyingObj << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << " " << *UnderlyingObj << "\n"; } } while (false); |
| 955 | } |
| 956 | } |
| 957 | } |
| 958 | } |
| 959 | } |
| 960 | } |
| 961 | |
| 962 | static bool isInBoundsGep(Value *Ptr) { |
| 963 | if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) |
| 964 | return GEP->isInBounds(); |
| 965 | return false; |
| 966 | } |
| 967 | |
| 968 | /// Return true if an AddRec pointer \p Ptr is unsigned non-wrapping, |
| 969 | /// i.e. monotonically increasing/decreasing. |
| 970 | static bool isNoWrapAddRec(Value *Ptr, const SCEVAddRecExpr *AR, |
| 971 | PredicatedScalarEvolution &PSE, const Loop *L) { |
| 972 | // FIXME: This should probably only return true for NUW. |
| 973 | if (AR->getNoWrapFlags(SCEV::NoWrapMask)) |
| 974 | return true; |
| 975 | |
| 976 | // Scalar evolution does not propagate the non-wrapping flags to values that |
| 977 | // are derived from a non-wrapping induction variable because non-wrapping |
| 978 | // could be flow-sensitive. |
| 979 | // |
| 980 | // Look through the potentially overflowing instruction to try to prove |
| 981 | // non-wrapping for the *specific* value of Ptr. |
| 982 | |
| 983 | // The arithmetic implied by an inbounds GEP can't overflow. |
| 984 | auto *GEP = dyn_cast<GetElementPtrInst>(Ptr); |
| 985 | if (!GEP || !GEP->isInBounds()) |
| 986 | return false; |
| 987 | |
| 988 | // Make sure there is only one non-const index and analyze that. |
| 989 | Value *NonConstIndex = nullptr; |
| 990 | for (Value *Index : make_range(GEP->idx_begin(), GEP->idx_end())) |
| 991 | if (!isa<ConstantInt>(Index)) { |
| 992 | if (NonConstIndex) |
| 993 | return false; |
| 994 | NonConstIndex = Index; |
| 995 | } |
| 996 | if (!NonConstIndex) |
| 997 | // The recurrence is on the pointer, ignore for now. |
| 998 | return false; |
| 999 | |
| 1000 | // The index in GEP is signed. It is non-wrapping if it's derived from a NSW |
| 1001 | // AddRec using a NSW operation. |
| 1002 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(NonConstIndex)) |
| 1003 | if (OBO->hasNoSignedWrap() && |
| 1004 | // Assume constant for other the operand so that the AddRec can be |
| 1005 | // easily found. |
| 1006 | isa<ConstantInt>(OBO->getOperand(1))) { |
| 1007 | auto *OpScev = PSE.getSCEV(OBO->getOperand(0)); |
| 1008 | |
| 1009 | if (auto *OpAR = dyn_cast<SCEVAddRecExpr>(OpScev)) |
| 1010 | return OpAR->getLoop() == L && OpAR->getNoWrapFlags(SCEV::FlagNSW); |
| 1011 | } |
| 1012 | |
| 1013 | return false; |
| 1014 | } |
| 1015 | |
| 1016 | /// Check whether the access through \p Ptr has a constant stride. |
| 1017 | int64_t llvm::getPtrStride(PredicatedScalarEvolution &PSE, Value *Ptr, |
| 1018 | const Loop *Lp, const ValueToValueMap &StridesMap, |
| 1019 | bool Assume, bool ShouldCheckWrap) { |
| 1020 | Type *Ty = Ptr->getType(); |
| 1021 | assert(Ty->isPointerTy() && "Unexpected non-ptr")((Ty->isPointerTy() && "Unexpected non-ptr") ? static_cast <void> (0) : __assert_fail ("Ty->isPointerTy() && \"Unexpected non-ptr\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1021, __PRETTY_FUNCTION__)); |
| 1022 | |
| 1023 | // Make sure that the pointer does not point to aggregate types. |
| 1024 | auto *PtrTy = cast<PointerType>(Ty); |
| 1025 | if (PtrTy->getElementType()->isAggregateType()) { |
| 1026 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a pointer to a scalar type"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not a pointer to a scalar type" << *Ptr << "\n"; } } while (false) |
| 1027 | << *Ptr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not a pointer to a scalar type" << *Ptr << "\n"; } } while (false); |
| 1028 | return 0; |
| 1029 | } |
| 1030 | |
| 1031 | const SCEV *PtrScev = replaceSymbolicStrideSCEV(PSE, StridesMap, Ptr); |
| 1032 | |
| 1033 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev); |
| 1034 | if (Assume && !AR) |
| 1035 | AR = PSE.getAsAddRec(Ptr); |
| 1036 | |
| 1037 | if (!AR) { |
| 1038 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr << " SCEV: " << *PtrScev << "\n" ; } } while (false) |
| 1039 | << " SCEV: " << *PtrScev << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not an AddRecExpr pointer " << *Ptr << " SCEV: " << *PtrScev << "\n" ; } } while (false); |
| 1040 | return 0; |
| 1041 | } |
| 1042 | |
| 1043 | // The access function must stride over the innermost loop. |
| 1044 | if (Lp != AR->getLoop()) { |
| 1045 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not striding over innermost loop "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not striding over innermost loop " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false) |
| 1046 | << *Ptr << " SCEV: " << *AR << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not striding over innermost loop " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false); |
| 1047 | return 0; |
| 1048 | } |
| 1049 | |
| 1050 | // The address calculation must not wrap. Otherwise, a dependence could be |
| 1051 | // inverted. |
| 1052 | // An inbounds getelementptr that is a AddRec with a unit stride |
| 1053 | // cannot wrap per definition. The unit stride requirement is checked later. |
| 1054 | // An getelementptr without an inbounds attribute and unit stride would have |
| 1055 | // to access the pointer value "0" which is undefined behavior in address |
| 1056 | // space 0, therefore we can also vectorize this case. |
| 1057 | bool IsInBoundsGEP = isInBoundsGep(Ptr); |
| 1058 | bool IsNoWrapAddRec = !ShouldCheckWrap || |
| 1059 | PSE.hasNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW) || |
| 1060 | isNoWrapAddRec(Ptr, AR, PSE, Lp); |
| 1061 | if (!IsNoWrapAddRec && !IsInBoundsGEP && |
| 1062 | NullPointerIsDefined(Lp->getHeader()->getParent(), |
| 1063 | PtrTy->getAddressSpace())) { |
| 1064 | if (Assume) { |
| 1065 | PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); |
| 1066 | IsNoWrapAddRec = true; |
| 1067 | LLVM_DEBUG(dbgs() << "LAA: Pointer may wrap in the address space:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Pointer may wrap in the address space:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) |
| 1068 | << "LAA: Pointer: " << *Ptr << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Pointer may wrap in the address space:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) |
| 1069 | << "LAA: SCEV: " << *AR << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Pointer may wrap in the address space:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) |
| 1070 | << "LAA: Added an overflow assumption\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Pointer may wrap in the address space:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false); |
| 1071 | } else { |
| 1072 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Pointer may wrap in the address space " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false) |
| 1073 | dbgs() << "LAA: Bad stride - Pointer may wrap in the address space "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Pointer may wrap in the address space " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false) |
| 1074 | << *Ptr << " SCEV: " << *AR << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Pointer may wrap in the address space " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false); |
| 1075 | return 0; |
| 1076 | } |
| 1077 | } |
| 1078 | |
| 1079 | // Check the step is constant. |
| 1080 | const SCEV *Step = AR->getStepRecurrence(*PSE.getSE()); |
| 1081 | |
| 1082 | // Calculate the pointer stride and check if it is constant. |
| 1083 | const SCEVConstant *C = dyn_cast<SCEVConstant>(Step); |
| 1084 | if (!C) { |
| 1085 | LLVM_DEBUG(dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false) |
| 1086 | << " SCEV: " << *AR << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Bad stride - Not a constant strided " << *Ptr << " SCEV: " << *AR << "\n"; } } while (false); |
| 1087 | return 0; |
| 1088 | } |
| 1089 | |
| 1090 | auto &DL = Lp->getHeader()->getModule()->getDataLayout(); |
| 1091 | int64_t Size = DL.getTypeAllocSize(PtrTy->getElementType()); |
| 1092 | const APInt &APStepVal = C->getAPInt(); |
| 1093 | |
| 1094 | // Huge step value - give up. |
| 1095 | if (APStepVal.getBitWidth() > 64) |
| 1096 | return 0; |
| 1097 | |
| 1098 | int64_t StepVal = APStepVal.getSExtValue(); |
| 1099 | |
| 1100 | // Strided access. |
| 1101 | int64_t Stride = StepVal / Size; |
| 1102 | int64_t Rem = StepVal % Size; |
| 1103 | if (Rem) |
| 1104 | return 0; |
| 1105 | |
| 1106 | // If the SCEV could wrap but we have an inbounds gep with a unit stride we |
| 1107 | // know we can't "wrap around the address space". In case of address space |
| 1108 | // zero we know that this won't happen without triggering undefined behavior. |
| 1109 | if (!IsNoWrapAddRec && Stride != 1 && Stride != -1 && |
| 1110 | (IsInBoundsGEP || !NullPointerIsDefined(Lp->getHeader()->getParent(), |
| 1111 | PtrTy->getAddressSpace()))) { |
| 1112 | if (Assume) { |
| 1113 | // We can avoid this case by adding a run-time check. |
| 1114 | LLVM_DEBUG(dbgs() << "LAA: Non unit strided pointer which is not either "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Non unit strided pointer which is not either " << "inbounds or in address space 0 may wrap:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) |
| 1115 | << "inbounds or in address space 0 may wrap:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Non unit strided pointer which is not either " << "inbounds or in address space 0 may wrap:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) |
| 1116 | << "LAA: Pointer: " << *Ptr << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Non unit strided pointer which is not either " << "inbounds or in address space 0 may wrap:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) |
| 1117 | << "LAA: SCEV: " << *AR << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Non unit strided pointer which is not either " << "inbounds or in address space 0 may wrap:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false) |
| 1118 | << "LAA: Added an overflow assumption\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Non unit strided pointer which is not either " << "inbounds or in address space 0 may wrap:\n" << "LAA: Pointer: " << *Ptr << "\n" << "LAA: SCEV: " << *AR << "\n" << "LAA: Added an overflow assumption\n" ; } } while (false); |
| 1119 | PSE.setNoOverflow(Ptr, SCEVWrapPredicate::IncrementNUSW); |
| 1120 | } else |
| 1121 | return 0; |
| 1122 | } |
| 1123 | |
| 1124 | return Stride; |
| 1125 | } |
| 1126 | |
| 1127 | bool llvm::sortPtrAccesses(ArrayRef<Value *> VL, const DataLayout &DL, |
| 1128 | ScalarEvolution &SE, |
| 1129 | SmallVectorImpl<unsigned> &SortedIndices) { |
| 1130 | assert(llvm::all_of(((llvm::all_of( VL, [](const Value *V) { return V->getType ()->isPointerTy(); }) && "Expected list of pointer operands." ) ? static_cast<void> (0) : __assert_fail ("llvm::all_of( VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && \"Expected list of pointer operands.\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1132, __PRETTY_FUNCTION__)) |
| 1131 | VL, [](const Value *V) { return V->getType()->isPointerTy(); }) &&((llvm::all_of( VL, [](const Value *V) { return V->getType ()->isPointerTy(); }) && "Expected list of pointer operands." ) ? static_cast<void> (0) : __assert_fail ("llvm::all_of( VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && \"Expected list of pointer operands.\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1132, __PRETTY_FUNCTION__)) |
| 1132 | "Expected list of pointer operands.")((llvm::all_of( VL, [](const Value *V) { return V->getType ()->isPointerTy(); }) && "Expected list of pointer operands." ) ? static_cast<void> (0) : __assert_fail ("llvm::all_of( VL, [](const Value *V) { return V->getType()->isPointerTy(); }) && \"Expected list of pointer operands.\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1132, __PRETTY_FUNCTION__)); |
| 1133 | SmallVector<std::pair<int64_t, Value *>, 4> OffValPairs; |
| 1134 | OffValPairs.reserve(VL.size()); |
| 1135 | |
| 1136 | // Walk over the pointers, and map each of them to an offset relative to |
| 1137 | // first pointer in the array. |
| 1138 | Value *Ptr0 = VL[0]; |
| 1139 | const SCEV *Scev0 = SE.getSCEV(Ptr0); |
| 1140 | Value *Obj0 = getUnderlyingObject(Ptr0); |
| 1141 | |
| 1142 | llvm::SmallSet<int64_t, 4> Offsets; |
| 1143 | for (auto *Ptr : VL) { |
| 1144 | // TODO: Outline this code as a special, more time consuming, version of |
| 1145 | // computeConstantDifference() function. |
| 1146 | if (Ptr->getType()->getPointerAddressSpace() != |
| 1147 | Ptr0->getType()->getPointerAddressSpace()) |
| 1148 | return false; |
| 1149 | // If a pointer refers to a different underlying object, bail - the |
| 1150 | // pointers are by definition incomparable. |
| 1151 | Value *CurrObj = getUnderlyingObject(Ptr); |
| 1152 | if (CurrObj != Obj0) |
| 1153 | return false; |
| 1154 | |
| 1155 | const SCEV *Scev = SE.getSCEV(Ptr); |
| 1156 | const auto *Diff = dyn_cast<SCEVConstant>(SE.getMinusSCEV(Scev, Scev0)); |
| 1157 | // The pointers may not have a constant offset from each other, or SCEV |
| 1158 | // may just not be smart enough to figure out they do. Regardless, |
| 1159 | // there's nothing we can do. |
| 1160 | if (!Diff) |
| 1161 | return false; |
| 1162 | |
| 1163 | // Check if the pointer with the same offset is found. |
| 1164 | int64_t Offset = Diff->getAPInt().getSExtValue(); |
| 1165 | if (!Offsets.insert(Offset).second) |
| 1166 | return false; |
| 1167 | OffValPairs.emplace_back(Offset, Ptr); |
| 1168 | } |
| 1169 | SortedIndices.clear(); |
| 1170 | SortedIndices.resize(VL.size()); |
| 1171 | std::iota(SortedIndices.begin(), SortedIndices.end(), 0); |
| 1172 | |
| 1173 | // Sort the memory accesses and keep the order of their uses in UseOrder. |
| 1174 | llvm::stable_sort(SortedIndices, [&](unsigned Left, unsigned Right) { |
| 1175 | return OffValPairs[Left].first < OffValPairs[Right].first; |
| 1176 | }); |
| 1177 | |
| 1178 | // Check if the order is consecutive already. |
| 1179 | if (llvm::all_of(SortedIndices, [&SortedIndices](const unsigned I) { |
| 1180 | return I == SortedIndices[I]; |
| 1181 | })) |
| 1182 | SortedIndices.clear(); |
| 1183 | |
| 1184 | return true; |
| 1185 | } |
| 1186 | |
| 1187 | /// Take the address space operand from the Load/Store instruction. |
| 1188 | /// Returns -1 if this is not a valid Load/Store instruction. |
| 1189 | static unsigned getAddressSpaceOperand(Value *I) { |
| 1190 | if (LoadInst *L = dyn_cast<LoadInst>(I)) |
| 1191 | return L->getPointerAddressSpace(); |
| 1192 | if (StoreInst *S = dyn_cast<StoreInst>(I)) |
| 1193 | return S->getPointerAddressSpace(); |
| 1194 | return -1; |
| 1195 | } |
| 1196 | |
| 1197 | /// Returns true if the memory operations \p A and \p B are consecutive. |
| 1198 | bool llvm::isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL, |
| 1199 | ScalarEvolution &SE, bool CheckType) { |
| 1200 | Value *PtrA = getLoadStorePointerOperand(A); |
| 1201 | Value *PtrB = getLoadStorePointerOperand(B); |
| 1202 | unsigned ASA = getAddressSpaceOperand(A); |
| 1203 | unsigned ASB = getAddressSpaceOperand(B); |
| 1204 | |
| 1205 | // Check that the address spaces match and that the pointers are valid. |
| 1206 | if (!PtrA || !PtrB || (ASA != ASB)) |
| 1207 | return false; |
| 1208 | |
| 1209 | // Make sure that A and B are different pointers. |
| 1210 | if (PtrA == PtrB) |
| 1211 | return false; |
| 1212 | |
| 1213 | // Make sure that A and B have the same type if required. |
| 1214 | if (CheckType && PtrA->getType() != PtrB->getType()) |
| 1215 | return false; |
| 1216 | |
| 1217 | unsigned IdxWidth = DL.getIndexSizeInBits(ASA); |
| 1218 | Type *Ty = cast<PointerType>(PtrA->getType())->getElementType(); |
| 1219 | |
| 1220 | APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0); |
| 1221 | PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA); |
| 1222 | PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB); |
| 1223 | |
| 1224 | // Retrieve the address space again as pointer stripping now tracks through |
| 1225 | // `addrspacecast`. |
| 1226 | ASA = cast<PointerType>(PtrA->getType())->getAddressSpace(); |
| 1227 | ASB = cast<PointerType>(PtrB->getType())->getAddressSpace(); |
| 1228 | // Check that the address spaces match and that the pointers are valid. |
| 1229 | if (ASA != ASB) |
| 1230 | return false; |
| 1231 | |
| 1232 | IdxWidth = DL.getIndexSizeInBits(ASA); |
| 1233 | OffsetA = OffsetA.sextOrTrunc(IdxWidth); |
| 1234 | OffsetB = OffsetB.sextOrTrunc(IdxWidth); |
| 1235 | |
| 1236 | APInt Size(IdxWidth, DL.getTypeStoreSize(Ty)); |
| 1237 | |
| 1238 | // OffsetDelta = OffsetB - OffsetA; |
| 1239 | const SCEV *OffsetSCEVA = SE.getConstant(OffsetA); |
| 1240 | const SCEV *OffsetSCEVB = SE.getConstant(OffsetB); |
| 1241 | const SCEV *OffsetDeltaSCEV = SE.getMinusSCEV(OffsetSCEVB, OffsetSCEVA); |
| 1242 | const APInt &OffsetDelta = cast<SCEVConstant>(OffsetDeltaSCEV)->getAPInt(); |
| 1243 | |
| 1244 | // Check if they are based on the same pointer. That makes the offsets |
| 1245 | // sufficient. |
| 1246 | if (PtrA == PtrB) |
| 1247 | return OffsetDelta == Size; |
| 1248 | |
| 1249 | // Compute the necessary base pointer delta to have the necessary final delta |
| 1250 | // equal to the size. |
| 1251 | // BaseDelta = Size - OffsetDelta; |
| 1252 | const SCEV *SizeSCEV = SE.getConstant(Size); |
| 1253 | const SCEV *BaseDelta = SE.getMinusSCEV(SizeSCEV, OffsetDeltaSCEV); |
| 1254 | |
| 1255 | // Otherwise compute the distance with SCEV between the base pointers. |
| 1256 | const SCEV *PtrSCEVA = SE.getSCEV(PtrA); |
| 1257 | const SCEV *PtrSCEVB = SE.getSCEV(PtrB); |
| 1258 | const SCEV *X = SE.getAddExpr(PtrSCEVA, BaseDelta); |
| 1259 | return X == PtrSCEVB; |
| 1260 | } |
| 1261 | |
| 1262 | MemoryDepChecker::VectorizationSafetyStatus |
| 1263 | MemoryDepChecker::Dependence::isSafeForVectorization(DepType Type) { |
| 1264 | switch (Type) { |
| 1265 | case NoDep: |
| 1266 | case Forward: |
| 1267 | case BackwardVectorizable: |
| 1268 | return VectorizationSafetyStatus::Safe; |
| 1269 | |
| 1270 | case Unknown: |
| 1271 | return VectorizationSafetyStatus::PossiblySafeWithRtChecks; |
| 1272 | case ForwardButPreventsForwarding: |
| 1273 | case Backward: |
| 1274 | case BackwardVectorizableButPreventsForwarding: |
| 1275 | return VectorizationSafetyStatus::Unsafe; |
| 1276 | } |
| 1277 | llvm_unreachable("unexpected DepType!")::llvm::llvm_unreachable_internal("unexpected DepType!", "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1277); |
| 1278 | } |
| 1279 | |
| 1280 | bool MemoryDepChecker::Dependence::isBackward() const { |
| 1281 | switch (Type) { |
| 1282 | case NoDep: |
| 1283 | case Forward: |
| 1284 | case ForwardButPreventsForwarding: |
| 1285 | case Unknown: |
| 1286 | return false; |
| 1287 | |
| 1288 | case BackwardVectorizable: |
| 1289 | case Backward: |
| 1290 | case BackwardVectorizableButPreventsForwarding: |
| 1291 | return true; |
| 1292 | } |
| 1293 | llvm_unreachable("unexpected DepType!")::llvm::llvm_unreachable_internal("unexpected DepType!", "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1293); |
| 1294 | } |
| 1295 | |
| 1296 | bool MemoryDepChecker::Dependence::isPossiblyBackward() const { |
| 1297 | return isBackward() || Type == Unknown; |
| 1298 | } |
| 1299 | |
| 1300 | bool MemoryDepChecker::Dependence::isForward() const { |
| 1301 | switch (Type) { |
| 1302 | case Forward: |
| 1303 | case ForwardButPreventsForwarding: |
| 1304 | return true; |
| 1305 | |
| 1306 | case NoDep: |
| 1307 | case Unknown: |
| 1308 | case BackwardVectorizable: |
| 1309 | case Backward: |
| 1310 | case BackwardVectorizableButPreventsForwarding: |
| 1311 | return false; |
| 1312 | } |
| 1313 | llvm_unreachable("unexpected DepType!")::llvm::llvm_unreachable_internal("unexpected DepType!", "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1313); |
| 1314 | } |
| 1315 | |
| 1316 | bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance, |
| 1317 | uint64_t TypeByteSize) { |
| 1318 | // If loads occur at a distance that is not a multiple of a feasible vector |
| 1319 | // factor store-load forwarding does not take place. |
| 1320 | // Positive dependences might cause troubles because vectorizing them might |
| 1321 | // prevent store-load forwarding making vectorized code run a lot slower. |
| 1322 | // a[i] = a[i-3] ^ a[i-8]; |
| 1323 | // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and |
| 1324 | // hence on your typical architecture store-load forwarding does not take |
| 1325 | // place. Vectorizing in such cases does not make sense. |
| 1326 | // Store-load forwarding distance. |
| 1327 | |
| 1328 | // After this many iterations store-to-load forwarding conflicts should not |
| 1329 | // cause any slowdowns. |
| 1330 | const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize; |
| 1331 | // Maximum vector factor. |
| 1332 | uint64_t MaxVFWithoutSLForwardIssues = std::min( |
| 1333 | VectorizerParams::MaxVectorWidth * TypeByteSize, MaxSafeDepDistBytes); |
| 1334 | |
| 1335 | // Compute the smallest VF at which the store and load would be misaligned. |
| 1336 | for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues; |
| 1337 | VF *= 2) { |
| 1338 | // If the number of vector iteration between the store and the load are |
| 1339 | // small we could incur conflicts. |
| 1340 | if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) { |
| 1341 | MaxVFWithoutSLForwardIssues = (VF >>= 1); |
Although the value stored to 'VF' is used in the enclosing expression, the value is never actually read from 'VF' | |
| 1342 | break; |
| 1343 | } |
| 1344 | } |
| 1345 | |
| 1346 | if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) { |
| 1347 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Distance " << Distance << " that could cause a store-load forwarding conflict\n" ; } } while (false) |
| 1348 | dbgs() << "LAA: Distance " << Distancedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Distance " << Distance << " that could cause a store-load forwarding conflict\n" ; } } while (false) |
| 1349 | << " that could cause a store-load forwarding conflict\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Distance " << Distance << " that could cause a store-load forwarding conflict\n" ; } } while (false); |
| 1350 | return true; |
| 1351 | } |
| 1352 | |
| 1353 | if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes && |
| 1354 | MaxVFWithoutSLForwardIssues != |
| 1355 | VectorizerParams::MaxVectorWidth * TypeByteSize) |
| 1356 | MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues; |
| 1357 | return false; |
| 1358 | } |
| 1359 | |
| 1360 | void MemoryDepChecker::mergeInStatus(VectorizationSafetyStatus S) { |
| 1361 | if (Status < S) |
| 1362 | Status = S; |
| 1363 | } |
| 1364 | |
| 1365 | /// Given a non-constant (unknown) dependence-distance \p Dist between two |
| 1366 | /// memory accesses, that have the same stride whose absolute value is given |
| 1367 | /// in \p Stride, and that have the same type size \p TypeByteSize, |
| 1368 | /// in a loop whose takenCount is \p BackedgeTakenCount, check if it is |
| 1369 | /// possible to prove statically that the dependence distance is larger |
| 1370 | /// than the range that the accesses will travel through the execution of |
| 1371 | /// the loop. If so, return true; false otherwise. This is useful for |
| 1372 | /// example in loops such as the following (PR31098): |
| 1373 | /// for (i = 0; i < D; ++i) { |
| 1374 | /// = out[i]; |
| 1375 | /// out[i+D] = |
| 1376 | /// } |
| 1377 | static bool isSafeDependenceDistance(const DataLayout &DL, ScalarEvolution &SE, |
| 1378 | const SCEV &BackedgeTakenCount, |
| 1379 | const SCEV &Dist, uint64_t Stride, |
| 1380 | uint64_t TypeByteSize) { |
| 1381 | |
| 1382 | // If we can prove that |
| 1383 | // (**) |Dist| > BackedgeTakenCount * Step |
| 1384 | // where Step is the absolute stride of the memory accesses in bytes, |
| 1385 | // then there is no dependence. |
| 1386 | // |
| 1387 | // Rationale: |
| 1388 | // We basically want to check if the absolute distance (|Dist/Step|) |
| 1389 | // is >= the loop iteration count (or > BackedgeTakenCount). |
| 1390 | // This is equivalent to the Strong SIV Test (Practical Dependence Testing, |
| 1391 | // Section 4.2.1); Note, that for vectorization it is sufficient to prove |
| 1392 | // that the dependence distance is >= VF; This is checked elsewhere. |
| 1393 | // But in some cases we can prune unknown dependence distances early, and |
| 1394 | // even before selecting the VF, and without a runtime test, by comparing |
| 1395 | // the distance against the loop iteration count. Since the vectorized code |
| 1396 | // will be executed only if LoopCount >= VF, proving distance >= LoopCount |
| 1397 | // also guarantees that distance >= VF. |
| 1398 | // |
| 1399 | const uint64_t ByteStride = Stride * TypeByteSize; |
| 1400 | const SCEV *Step = SE.getConstant(BackedgeTakenCount.getType(), ByteStride); |
| 1401 | const SCEV *Product = SE.getMulExpr(&BackedgeTakenCount, Step); |
| 1402 | |
| 1403 | const SCEV *CastedDist = &Dist; |
| 1404 | const SCEV *CastedProduct = Product; |
| 1405 | uint64_t DistTypeSize = DL.getTypeAllocSize(Dist.getType()); |
| 1406 | uint64_t ProductTypeSize = DL.getTypeAllocSize(Product->getType()); |
| 1407 | |
| 1408 | // The dependence distance can be positive/negative, so we sign extend Dist; |
| 1409 | // The multiplication of the absolute stride in bytes and the |
| 1410 | // backedgeTakenCount is non-negative, so we zero extend Product. |
| 1411 | if (DistTypeSize > ProductTypeSize) |
| 1412 | CastedProduct = SE.getZeroExtendExpr(Product, Dist.getType()); |
| 1413 | else |
| 1414 | CastedDist = SE.getNoopOrSignExtend(&Dist, Product->getType()); |
| 1415 | |
| 1416 | // Is Dist - (BackedgeTakenCount * Step) > 0 ? |
| 1417 | // (If so, then we have proven (**) because |Dist| >= Dist) |
| 1418 | const SCEV *Minus = SE.getMinusSCEV(CastedDist, CastedProduct); |
| 1419 | if (SE.isKnownPositive(Minus)) |
| 1420 | return true; |
| 1421 | |
| 1422 | // Second try: Is -Dist - (BackedgeTakenCount * Step) > 0 ? |
| 1423 | // (If so, then we have proven (**) because |Dist| >= -1*Dist) |
| 1424 | const SCEV *NegDist = SE.getNegativeSCEV(CastedDist); |
| 1425 | Minus = SE.getMinusSCEV(NegDist, CastedProduct); |
| 1426 | if (SE.isKnownPositive(Minus)) |
| 1427 | return true; |
| 1428 | |
| 1429 | return false; |
| 1430 | } |
| 1431 | |
| 1432 | /// Check the dependence for two accesses with the same stride \p Stride. |
| 1433 | /// \p Distance is the positive distance and \p TypeByteSize is type size in |
| 1434 | /// bytes. |
| 1435 | /// |
| 1436 | /// \returns true if they are independent. |
| 1437 | static bool areStridedAccessesIndependent(uint64_t Distance, uint64_t Stride, |
| 1438 | uint64_t TypeByteSize) { |
| 1439 | assert(Stride > 1 && "The stride must be greater than 1")((Stride > 1 && "The stride must be greater than 1" ) ? static_cast<void> (0) : __assert_fail ("Stride > 1 && \"The stride must be greater than 1\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1439, __PRETTY_FUNCTION__)); |
| 1440 | assert(TypeByteSize > 0 && "The type size in byte must be non-zero")((TypeByteSize > 0 && "The type size in byte must be non-zero" ) ? static_cast<void> (0) : __assert_fail ("TypeByteSize > 0 && \"The type size in byte must be non-zero\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1440, __PRETTY_FUNCTION__)); |
| 1441 | assert(Distance > 0 && "The distance must be non-zero")((Distance > 0 && "The distance must be non-zero") ? static_cast<void> (0) : __assert_fail ("Distance > 0 && \"The distance must be non-zero\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1441, __PRETTY_FUNCTION__)); |
| 1442 | |
| 1443 | // Skip if the distance is not multiple of type byte size. |
| 1444 | if (Distance % TypeByteSize) |
| 1445 | return false; |
| 1446 | |
| 1447 | uint64_t ScaledDist = Distance / TypeByteSize; |
| 1448 | |
| 1449 | // No dependence if the scaled distance is not multiple of the stride. |
| 1450 | // E.g. |
| 1451 | // for (i = 0; i < 1024 ; i += 4) |
| 1452 | // A[i+2] = A[i] + 1; |
| 1453 | // |
| 1454 | // Two accesses in memory (scaled distance is 2, stride is 4): |
| 1455 | // | A[0] | | | | A[4] | | | | |
| 1456 | // | | | A[2] | | | | A[6] | | |
| 1457 | // |
| 1458 | // E.g. |
| 1459 | // for (i = 0; i < 1024 ; i += 3) |
| 1460 | // A[i+4] = A[i] + 1; |
| 1461 | // |
| 1462 | // Two accesses in memory (scaled distance is 4, stride is 3): |
| 1463 | // | A[0] | | | A[3] | | | A[6] | | | |
| 1464 | // | | | | | A[4] | | | A[7] | | |
| 1465 | return ScaledDist % Stride; |
| 1466 | } |
| 1467 | |
| 1468 | MemoryDepChecker::Dependence::DepType |
| 1469 | MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx, |
| 1470 | const MemAccessInfo &B, unsigned BIdx, |
| 1471 | const ValueToValueMap &Strides) { |
| 1472 | assert (AIdx < BIdx && "Must pass arguments in program order")((AIdx < BIdx && "Must pass arguments in program order" ) ? static_cast<void> (0) : __assert_fail ("AIdx < BIdx && \"Must pass arguments in program order\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1472, __PRETTY_FUNCTION__)); |
| 1473 | |
| 1474 | Value *APtr = A.getPointer(); |
| 1475 | Value *BPtr = B.getPointer(); |
| 1476 | bool AIsWrite = A.getInt(); |
| 1477 | bool BIsWrite = B.getInt(); |
| 1478 | |
| 1479 | // Two reads are independent. |
| 1480 | if (!AIsWrite && !BIsWrite) |
| 1481 | return Dependence::NoDep; |
| 1482 | |
| 1483 | // We cannot check pointers in different address spaces. |
| 1484 | if (APtr->getType()->getPointerAddressSpace() != |
| 1485 | BPtr->getType()->getPointerAddressSpace()) |
| 1486 | return Dependence::Unknown; |
| 1487 | |
| 1488 | int64_t StrideAPtr = getPtrStride(PSE, APtr, InnermostLoop, Strides, true); |
| 1489 | int64_t StrideBPtr = getPtrStride(PSE, BPtr, InnermostLoop, Strides, true); |
| 1490 | |
| 1491 | const SCEV *Src = PSE.getSCEV(APtr); |
| 1492 | const SCEV *Sink = PSE.getSCEV(BPtr); |
| 1493 | |
| 1494 | // If the induction step is negative we have to invert source and sink of the |
| 1495 | // dependence. |
| 1496 | if (StrideAPtr < 0) { |
| 1497 | std::swap(APtr, BPtr); |
| 1498 | std::swap(Src, Sink); |
| 1499 | std::swap(AIsWrite, BIsWrite); |
| 1500 | std::swap(AIdx, BIdx); |
| 1501 | std::swap(StrideAPtr, StrideBPtr); |
| 1502 | } |
| 1503 | |
| 1504 | const SCEV *Dist = PSE.getSE()->getMinusSCEV(Sink, Src); |
| 1505 | |
| 1506 | LLVM_DEBUG(dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sinkdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink << "(Induction step: " << StrideAPtr << ")\n"; } } while (false) |
| 1507 | << "(Induction step: " << StrideAPtr << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Src Scev: " << *Src << "Sink Scev: " << *Sink << "(Induction step: " << StrideAPtr << ")\n"; } } while (false); |
| 1508 | LLVM_DEBUG(dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to " << *InstMap[BIdx] << ": " << *Dist << "\n"; } } while (false) |
| 1509 | << *InstMap[BIdx] << ": " << *Dist << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Distance for " << *InstMap[AIdx] << " to " << *InstMap[BIdx] << ": " << *Dist << "\n"; } } while (false); |
| 1510 | |
| 1511 | // Need accesses with constant stride. We don't want to vectorize |
| 1512 | // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in |
| 1513 | // the address space. |
| 1514 | if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){ |
| 1515 | LLVM_DEBUG(dbgs() << "Pointer access with non-constant stride\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Pointer access with non-constant stride\n" ; } } while (false); |
| 1516 | return Dependence::Unknown; |
| 1517 | } |
| 1518 | |
| 1519 | Type *ATy = APtr->getType()->getPointerElementType(); |
| 1520 | Type *BTy = BPtr->getType()->getPointerElementType(); |
| 1521 | auto &DL = InnermostLoop->getHeader()->getModule()->getDataLayout(); |
| 1522 | uint64_t TypeByteSize = DL.getTypeAllocSize(ATy); |
| 1523 | uint64_t Stride = std::abs(StrideAPtr); |
| 1524 | const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist); |
| 1525 | if (!C) { |
| 1526 | if (TypeByteSize == DL.getTypeAllocSize(BTy) && |
| 1527 | isSafeDependenceDistance(DL, *(PSE.getSE()), |
| 1528 | *(PSE.getBackedgeTakenCount()), *Dist, Stride, |
| 1529 | TypeByteSize)) |
| 1530 | return Dependence::NoDep; |
| 1531 | |
| 1532 | LLVM_DEBUG(dbgs() << "LAA: Dependence because of non-constant distance\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Dependence because of non-constant distance\n" ; } } while (false); |
| 1533 | FoundNonConstantDistanceDependence = true; |
| 1534 | return Dependence::Unknown; |
| 1535 | } |
| 1536 | |
| 1537 | const APInt &Val = C->getAPInt(); |
| 1538 | int64_t Distance = Val.getSExtValue(); |
| 1539 | |
| 1540 | // Attempt to prove strided accesses independent. |
| 1541 | if (std::abs(Distance) > 0 && Stride > 1 && ATy == BTy && |
| 1542 | areStridedAccessesIndependent(std::abs(Distance), Stride, TypeByteSize)) { |
| 1543 | LLVM_DEBUG(dbgs() << "LAA: Strided accesses are independent\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Strided accesses are independent\n" ; } } while (false); |
| 1544 | return Dependence::NoDep; |
| 1545 | } |
| 1546 | |
| 1547 | // Negative distances are not plausible dependencies. |
| 1548 | if (Val.isNegative()) { |
| 1549 | bool IsTrueDataDependence = (AIsWrite && !BIsWrite); |
| 1550 | if (IsTrueDataDependence && EnableForwardingConflictDetection && |
| 1551 | (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) || |
| 1552 | ATy != BTy)) { |
| 1553 | LLVM_DEBUG(dbgs() << "LAA: Forward but may prevent st->ld forwarding\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Forward but may prevent st->ld forwarding\n" ; } } while (false); |
| 1554 | return Dependence::ForwardButPreventsForwarding; |
| 1555 | } |
| 1556 | |
| 1557 | LLVM_DEBUG(dbgs() << "LAA: Dependence is negative\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Dependence is negative\n" ; } } while (false); |
| 1558 | return Dependence::Forward; |
| 1559 | } |
| 1560 | |
| 1561 | // Write to the same location with the same size. |
| 1562 | // Could be improved to assert type sizes are the same (i32 == float, etc). |
| 1563 | if (Val == 0) { |
| 1564 | if (ATy == BTy) |
| 1565 | return Dependence::Forward; |
| 1566 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Zero dependence difference but different types\n" ; } } while (false) |
| 1567 | dbgs() << "LAA: Zero dependence difference but different types\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Zero dependence difference but different types\n" ; } } while (false); |
| 1568 | return Dependence::Unknown; |
| 1569 | } |
| 1570 | |
| 1571 | assert(Val.isStrictlyPositive() && "Expect a positive value")((Val.isStrictlyPositive() && "Expect a positive value" ) ? static_cast<void> (0) : __assert_fail ("Val.isStrictlyPositive() && \"Expect a positive value\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1571, __PRETTY_FUNCTION__)); |
| 1572 | |
| 1573 | if (ATy != BTy) { |
| 1574 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: ReadWrite-Write positive dependency with different types\n" ; } } while (false) |
| 1575 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: ReadWrite-Write positive dependency with different types\n" ; } } while (false) |
| 1576 | << "LAA: ReadWrite-Write positive dependency with different types\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: ReadWrite-Write positive dependency with different types\n" ; } } while (false); |
| 1577 | return Dependence::Unknown; |
| 1578 | } |
| 1579 | |
| 1580 | // Bail out early if passed-in parameters make vectorization not feasible. |
| 1581 | unsigned ForcedFactor = (VectorizerParams::VectorizationFactor ? |
| 1582 | VectorizerParams::VectorizationFactor : 1); |
| 1583 | unsigned ForcedUnroll = (VectorizerParams::VectorizationInterleave ? |
| 1584 | VectorizerParams::VectorizationInterleave : 1); |
| 1585 | // The minimum number of iterations for a vectorized/unrolled version. |
| 1586 | unsigned MinNumIter = std::max(ForcedFactor * ForcedUnroll, 2U); |
| 1587 | |
| 1588 | // It's not vectorizable if the distance is smaller than the minimum distance |
| 1589 | // needed for a vectroized/unrolled version. Vectorizing one iteration in |
| 1590 | // front needs TypeByteSize * Stride. Vectorizing the last iteration needs |
| 1591 | // TypeByteSize (No need to plus the last gap distance). |
| 1592 | // |
| 1593 | // E.g. Assume one char is 1 byte in memory and one int is 4 bytes. |
| 1594 | // foo(int *A) { |
| 1595 | // int *B = (int *)((char *)A + 14); |
| 1596 | // for (i = 0 ; i < 1024 ; i += 2) |
| 1597 | // B[i] = A[i] + 1; |
| 1598 | // } |
| 1599 | // |
| 1600 | // Two accesses in memory (stride is 2): |
| 1601 | // | A[0] | | A[2] | | A[4] | | A[6] | | |
| 1602 | // | B[0] | | B[2] | | B[4] | |
| 1603 | // |
| 1604 | // Distance needs for vectorizing iterations except the last iteration: |
| 1605 | // 4 * 2 * (MinNumIter - 1). Distance needs for the last iteration: 4. |
| 1606 | // So the minimum distance needed is: 4 * 2 * (MinNumIter - 1) + 4. |
| 1607 | // |
| 1608 | // If MinNumIter is 2, it is vectorizable as the minimum distance needed is |
| 1609 | // 12, which is less than distance. |
| 1610 | // |
| 1611 | // If MinNumIter is 4 (Say if a user forces the vectorization factor to be 4), |
| 1612 | // the minimum distance needed is 28, which is greater than distance. It is |
| 1613 | // not safe to do vectorization. |
| 1614 | uint64_t MinDistanceNeeded = |
| 1615 | TypeByteSize * Stride * (MinNumIter - 1) + TypeByteSize; |
| 1616 | if (MinDistanceNeeded > static_cast<uint64_t>(Distance)) { |
| 1617 | LLVM_DEBUG(dbgs() << "LAA: Failure because of positive distance "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Failure because of positive distance " << Distance << '\n'; } } while (false) |
| 1618 | << Distance << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Failure because of positive distance " << Distance << '\n'; } } while (false); |
| 1619 | return Dependence::Backward; |
| 1620 | } |
| 1621 | |
| 1622 | // Unsafe if the minimum distance needed is greater than max safe distance. |
| 1623 | if (MinDistanceNeeded > MaxSafeDepDistBytes) { |
| 1624 | LLVM_DEBUG(dbgs() << "LAA: Failure because it needs at least "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Failure because it needs at least " << MinDistanceNeeded << " size in bytes"; } } while (false) |
| 1625 | << MinDistanceNeeded << " size in bytes")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Failure because it needs at least " << MinDistanceNeeded << " size in bytes"; } } while (false); |
| 1626 | return Dependence::Backward; |
| 1627 | } |
| 1628 | |
| 1629 | // Positive distance bigger than max vectorization factor. |
| 1630 | // FIXME: Should use max factor instead of max distance in bytes, which could |
| 1631 | // not handle different types. |
| 1632 | // E.g. Assume one char is 1 byte in memory and one int is 4 bytes. |
| 1633 | // void foo (int *A, char *B) { |
| 1634 | // for (unsigned i = 0; i < 1024; i++) { |
| 1635 | // A[i+2] = A[i] + 1; |
| 1636 | // B[i+2] = B[i] + 1; |
| 1637 | // } |
| 1638 | // } |
| 1639 | // |
| 1640 | // This case is currently unsafe according to the max safe distance. If we |
| 1641 | // analyze the two accesses on array B, the max safe dependence distance |
| 1642 | // is 2. Then we analyze the accesses on array A, the minimum distance needed |
| 1643 | // is 8, which is less than 2 and forbidden vectorization, But actually |
| 1644 | // both A and B could be vectorized by 2 iterations. |
| 1645 | MaxSafeDepDistBytes = |
| 1646 | std::min(static_cast<uint64_t>(Distance), MaxSafeDepDistBytes); |
| 1647 | |
| 1648 | bool IsTrueDataDependence = (!AIsWrite && BIsWrite); |
| 1649 | if (IsTrueDataDependence && EnableForwardingConflictDetection && |
| 1650 | couldPreventStoreLoadForward(Distance, TypeByteSize)) |
| 1651 | return Dependence::BackwardVectorizableButPreventsForwarding; |
| 1652 | |
| 1653 | uint64_t MaxVF = MaxSafeDepDistBytes / (TypeByteSize * Stride); |
| 1654 | LLVM_DEBUG(dbgs() << "LAA: Positive distance " << Val.getSExtValue()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Positive distance " << Val.getSExtValue() << " with max VF = " << MaxVF << '\n'; } } while (false) |
| 1655 | << " with max VF = " << MaxVF << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Positive distance " << Val.getSExtValue() << " with max VF = " << MaxVF << '\n'; } } while (false); |
| 1656 | uint64_t MaxVFInBits = MaxVF * TypeByteSize * 8; |
| 1657 | MaxSafeVectorWidthInBits = std::min(MaxSafeVectorWidthInBits, MaxVFInBits); |
| 1658 | return Dependence::BackwardVectorizable; |
| 1659 | } |
| 1660 | |
| 1661 | bool MemoryDepChecker::areDepsSafe(DepCandidates &AccessSets, |
| 1662 | MemAccessInfoList &CheckDeps, |
| 1663 | const ValueToValueMap &Strides) { |
| 1664 | |
| 1665 | MaxSafeDepDistBytes = -1; |
| 1666 | SmallPtrSet<MemAccessInfo, 8> Visited; |
| 1667 | for (MemAccessInfo CurAccess : CheckDeps) { |
| 1668 | if (Visited.count(CurAccess)) |
| 1669 | continue; |
| 1670 | |
| 1671 | // Get the relevant memory access set. |
| 1672 | EquivalenceClasses<MemAccessInfo>::iterator I = |
| 1673 | AccessSets.findValue(AccessSets.getLeaderValue(CurAccess)); |
| 1674 | |
| 1675 | // Check accesses within this set. |
| 1676 | EquivalenceClasses<MemAccessInfo>::member_iterator AI = |
| 1677 | AccessSets.member_begin(I); |
| 1678 | EquivalenceClasses<MemAccessInfo>::member_iterator AE = |
| 1679 | AccessSets.member_end(); |
| 1680 | |
| 1681 | // Check every access pair. |
| 1682 | while (AI != AE) { |
| 1683 | Visited.insert(*AI); |
| 1684 | bool AIIsWrite = AI->getInt(); |
| 1685 | // Check loads only against next equivalent class, but stores also against |
| 1686 | // other stores in the same equivalence class - to the same address. |
| 1687 | EquivalenceClasses<MemAccessInfo>::member_iterator OI = |
| 1688 | (AIIsWrite ? AI : std::next(AI)); |
| 1689 | while (OI != AE) { |
| 1690 | // Check every accessing instruction pair in program order. |
| 1691 | for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(), |
| 1692 | I1E = Accesses[*AI].end(); I1 != I1E; ++I1) |
| 1693 | // Scan all accesses of another equivalence class, but only the next |
| 1694 | // accesses of the same equivalent class. |
| 1695 | for (std::vector<unsigned>::iterator |
| 1696 | I2 = (OI == AI ? std::next(I1) : Accesses[*OI].begin()), |
| 1697 | I2E = (OI == AI ? I1E : Accesses[*OI].end()); |
| 1698 | I2 != I2E; ++I2) { |
| 1699 | auto A = std::make_pair(&*AI, *I1); |
| 1700 | auto B = std::make_pair(&*OI, *I2); |
| 1701 | |
| 1702 | assert(*I1 != *I2)((*I1 != *I2) ? static_cast<void> (0) : __assert_fail ( "*I1 != *I2", "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 1702, __PRETTY_FUNCTION__)); |
| 1703 | if (*I1 > *I2) |
| 1704 | std::swap(A, B); |
| 1705 | |
| 1706 | Dependence::DepType Type = |
| 1707 | isDependent(*A.first, A.second, *B.first, B.second, Strides); |
| 1708 | mergeInStatus(Dependence::isSafeForVectorization(Type)); |
| 1709 | |
| 1710 | // Gather dependences unless we accumulated MaxDependences |
| 1711 | // dependences. In that case return as soon as we find the first |
| 1712 | // unsafe dependence. This puts a limit on this quadratic |
| 1713 | // algorithm. |
| 1714 | if (RecordDependences) { |
| 1715 | if (Type != Dependence::NoDep) |
| 1716 | Dependences.push_back(Dependence(A.second, B.second, Type)); |
| 1717 | |
| 1718 | if (Dependences.size() >= MaxDependences) { |
| 1719 | RecordDependences = false; |
| 1720 | Dependences.clear(); |
| 1721 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Too many dependences, stopped recording\n" ; } } while (false) |
| 1722 | << "Too many dependences, stopped recording\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Too many dependences, stopped recording\n" ; } } while (false); |
| 1723 | } |
| 1724 | } |
| 1725 | if (!RecordDependences && !isSafeForVectorization()) |
| 1726 | return false; |
| 1727 | } |
| 1728 | ++OI; |
| 1729 | } |
| 1730 | AI++; |
| 1731 | } |
| 1732 | } |
| 1733 | |
| 1734 | LLVM_DEBUG(dbgs() << "Total Dependences: " << Dependences.size() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "Total Dependences: " << Dependences.size() << "\n"; } } while (false); |
| 1735 | return isSafeForVectorization(); |
| 1736 | } |
| 1737 | |
| 1738 | SmallVector<Instruction *, 4> |
| 1739 | MemoryDepChecker::getInstructionsForAccess(Value *Ptr, bool isWrite) const { |
| 1740 | MemAccessInfo Access(Ptr, isWrite); |
| 1741 | auto &IndexVector = Accesses.find(Access)->second; |
| 1742 | |
| 1743 | SmallVector<Instruction *, 4> Insts; |
| 1744 | transform(IndexVector, |
| 1745 | std::back_inserter(Insts), |
| 1746 | [&](unsigned Idx) { return this->InstMap[Idx]; }); |
| 1747 | return Insts; |
| 1748 | } |
| 1749 | |
| 1750 | const char *MemoryDepChecker::Dependence::DepName[] = { |
| 1751 | "NoDep", "Unknown", "Forward", "ForwardButPreventsForwarding", "Backward", |
| 1752 | "BackwardVectorizable", "BackwardVectorizableButPreventsForwarding"}; |
| 1753 | |
| 1754 | void MemoryDepChecker::Dependence::print( |
| 1755 | raw_ostream &OS, unsigned Depth, |
| 1756 | const SmallVectorImpl<Instruction *> &Instrs) const { |
| 1757 | OS.indent(Depth) << DepName[Type] << ":\n"; |
| 1758 | OS.indent(Depth + 2) << *Instrs[Source] << " -> \n"; |
| 1759 | OS.indent(Depth + 2) << *Instrs[Destination] << "\n"; |
| 1760 | } |
| 1761 | |
| 1762 | bool LoopAccessInfo::canAnalyzeLoop() { |
| 1763 | // We need to have a loop header. |
| 1764 | LLVM_DEBUG(dbgs() << "LAA: Found a loop in "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a loop in " << TheLoop->getHeader()->getParent()->getName() << ": " << TheLoop->getHeader()->getName() << '\n'; } } while (false) |
| 1765 | << TheLoop->getHeader()->getParent()->getName() << ": "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a loop in " << TheLoop->getHeader()->getParent()->getName() << ": " << TheLoop->getHeader()->getName() << '\n'; } } while (false) |
| 1766 | << TheLoop->getHeader()->getName() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a loop in " << TheLoop->getHeader()->getParent()->getName() << ": " << TheLoop->getHeader()->getName() << '\n'; } } while (false); |
| 1767 | |
| 1768 | // We can only analyze innermost loops. |
| 1769 | if (!TheLoop->isInnermost()) { |
| 1770 | LLVM_DEBUG(dbgs() << "LAA: loop is not the innermost loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: loop is not the innermost loop\n" ; } } while (false); |
| 1771 | recordAnalysis("NotInnerMostLoop") << "loop is not the innermost loop"; |
| 1772 | return false; |
| 1773 | } |
| 1774 | |
| 1775 | // We must have a single backedge. |
| 1776 | if (TheLoop->getNumBackEdges() != 1) { |
| 1777 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: loop control flow is not understood by analyzer\n" ; } } while (false) |
| 1778 | dbgs() << "LAA: loop control flow is not understood by analyzer\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: loop control flow is not understood by analyzer\n" ; } } while (false); |
| 1779 | recordAnalysis("CFGNotUnderstood") |
| 1780 | << "loop control flow is not understood by analyzer"; |
| 1781 | return false; |
| 1782 | } |
| 1783 | |
| 1784 | // ScalarEvolution needs to be able to find the exit count. |
| 1785 | const SCEV *ExitCount = PSE->getBackedgeTakenCount(); |
| 1786 | if (isa<SCEVCouldNotCompute>(ExitCount)) { |
| 1787 | recordAnalysis("CantComputeNumberOfIterations") |
| 1788 | << "could not determine number of loop iterations"; |
| 1789 | LLVM_DEBUG(dbgs() << "LAA: SCEV could not compute the loop exit count.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: SCEV could not compute the loop exit count.\n" ; } } while (false); |
| 1790 | return false; |
| 1791 | } |
| 1792 | |
| 1793 | return true; |
| 1794 | } |
| 1795 | |
| 1796 | void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI, |
| 1797 | const TargetLibraryInfo *TLI, |
| 1798 | DominatorTree *DT) { |
| 1799 | typedef SmallPtrSet<Value*, 16> ValueSet; |
| 1800 | |
| 1801 | // Holds the Load and Store instructions. |
| 1802 | SmallVector<LoadInst *, 16> Loads; |
| 1803 | SmallVector<StoreInst *, 16> Stores; |
| 1804 | |
| 1805 | // Holds all the different accesses in the loop. |
| 1806 | unsigned NumReads = 0; |
| 1807 | unsigned NumReadWrites = 0; |
| 1808 | |
| 1809 | bool HasComplexMemInst = false; |
| 1810 | |
| 1811 | // A runtime check is only legal to insert if there are no convergent calls. |
| 1812 | HasConvergentOp = false; |
| 1813 | |
| 1814 | PtrRtChecking->Pointers.clear(); |
| 1815 | PtrRtChecking->Need = false; |
| 1816 | |
| 1817 | const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel(); |
| 1818 | |
| 1819 | const bool EnableMemAccessVersioningOfLoop = |
| 1820 | EnableMemAccessVersioning && |
| 1821 | !TheLoop->getHeader()->getParent()->hasOptSize(); |
| 1822 | |
| 1823 | // For each block. |
| 1824 | for (BasicBlock *BB : TheLoop->blocks()) { |
| 1825 | // Scan the BB and collect legal loads and stores. Also detect any |
| 1826 | // convergent instructions. |
| 1827 | for (Instruction &I : *BB) { |
| 1828 | if (auto *Call = dyn_cast<CallBase>(&I)) { |
| 1829 | if (Call->isConvergent()) |
| 1830 | HasConvergentOp = true; |
| 1831 | } |
| 1832 | |
| 1833 | // With both a non-vectorizable memory instruction and a convergent |
| 1834 | // operation, found in this loop, no reason to continue the search. |
| 1835 | if (HasComplexMemInst && HasConvergentOp) { |
| 1836 | CanVecMem = false; |
| 1837 | return; |
| 1838 | } |
| 1839 | |
| 1840 | // Avoid hitting recordAnalysis multiple times. |
| 1841 | if (HasComplexMemInst) |
| 1842 | continue; |
| 1843 | |
| 1844 | // If this is a load, save it. If this instruction can read from memory |
| 1845 | // but is not a load, then we quit. Notice that we don't handle function |
| 1846 | // calls that read or write. |
| 1847 | if (I.mayReadFromMemory()) { |
| 1848 | // Many math library functions read the rounding mode. We will only |
| 1849 | // vectorize a loop if it contains known function calls that don't set |
| 1850 | // the flag. Therefore, it is safe to ignore this read from memory. |
| 1851 | auto *Call = dyn_cast<CallInst>(&I); |
| 1852 | if (Call && getVectorIntrinsicIDForCall(Call, TLI)) |
| 1853 | continue; |
| 1854 | |
| 1855 | // If the function has an explicit vectorized counterpart, we can safely |
| 1856 | // assume that it can be vectorized. |
| 1857 | if (Call && !Call->isNoBuiltin() && Call->getCalledFunction() && |
| 1858 | !VFDatabase::getMappings(*Call).empty()) |
| 1859 | continue; |
| 1860 | |
| 1861 | auto *Ld = dyn_cast<LoadInst>(&I); |
| 1862 | if (!Ld) { |
| 1863 | recordAnalysis("CantVectorizeInstruction", Ld) |
| 1864 | << "instruction cannot be vectorized"; |
| 1865 | HasComplexMemInst = true; |
| 1866 | continue; |
| 1867 | } |
| 1868 | if (!Ld->isSimple() && !IsAnnotatedParallel) { |
| 1869 | recordAnalysis("NonSimpleLoad", Ld) |
| 1870 | << "read with atomic ordering or volatile read"; |
| 1871 | LLVM_DEBUG(dbgs() << "LAA: Found a non-simple load.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a non-simple load.\n" ; } } while (false); |
| 1872 | HasComplexMemInst = true; |
| 1873 | continue; |
| 1874 | } |
| 1875 | NumLoads++; |
| 1876 | Loads.push_back(Ld); |
| 1877 | DepChecker->addAccess(Ld); |
| 1878 | if (EnableMemAccessVersioningOfLoop) |
| 1879 | collectStridedAccess(Ld); |
| 1880 | continue; |
| 1881 | } |
| 1882 | |
| 1883 | // Save 'store' instructions. Abort if other instructions write to memory. |
| 1884 | if (I.mayWriteToMemory()) { |
| 1885 | auto *St = dyn_cast<StoreInst>(&I); |
| 1886 | if (!St) { |
| 1887 | recordAnalysis("CantVectorizeInstruction", St) |
| 1888 | << "instruction cannot be vectorized"; |
| 1889 | HasComplexMemInst = true; |
| 1890 | continue; |
| 1891 | } |
| 1892 | if (!St->isSimple() && !IsAnnotatedParallel) { |
| 1893 | recordAnalysis("NonSimpleStore", St) |
| 1894 | << "write with atomic ordering or volatile write"; |
| 1895 | LLVM_DEBUG(dbgs() << "LAA: Found a non-simple store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a non-simple store.\n" ; } } while (false); |
| 1896 | HasComplexMemInst = true; |
| 1897 | continue; |
| 1898 | } |
| 1899 | NumStores++; |
| 1900 | Stores.push_back(St); |
| 1901 | DepChecker->addAccess(St); |
| 1902 | if (EnableMemAccessVersioningOfLoop) |
| 1903 | collectStridedAccess(St); |
| 1904 | } |
| 1905 | } // Next instr. |
| 1906 | } // Next block. |
| 1907 | |
| 1908 | if (HasComplexMemInst) { |
| 1909 | CanVecMem = false; |
| 1910 | return; |
| 1911 | } |
| 1912 | |
| 1913 | // Now we have two lists that hold the loads and the stores. |
| 1914 | // Next, we find the pointers that they use. |
| 1915 | |
| 1916 | // Check if we see any stores. If there are no stores, then we don't |
| 1917 | // care if the pointers are *restrict*. |
| 1918 | if (!Stores.size()) { |
| 1919 | LLVM_DEBUG(dbgs() << "LAA: Found a read-only loop!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a read-only loop!\n" ; } } while (false); |
| 1920 | CanVecMem = true; |
| 1921 | return; |
| 1922 | } |
| 1923 | |
| 1924 | MemoryDepChecker::DepCandidates DependentAccesses; |
| 1925 | AccessAnalysis Accesses(TheLoop, AA, LI, DependentAccesses, *PSE); |
| 1926 | |
| 1927 | // Holds the analyzed pointers. We don't want to call getUnderlyingObjects |
| 1928 | // multiple times on the same object. If the ptr is accessed twice, once |
| 1929 | // for read and once for write, it will only appear once (on the write |
| 1930 | // list). This is okay, since we are going to check for conflicts between |
| 1931 | // writes and between reads and writes, but not between reads and reads. |
| 1932 | ValueSet Seen; |
| 1933 | |
| 1934 | // Record uniform store addresses to identify if we have multiple stores |
| 1935 | // to the same address. |
| 1936 | ValueSet UniformStores; |
| 1937 | |
| 1938 | for (StoreInst *ST : Stores) { |
| 1939 | Value *Ptr = ST->getPointerOperand(); |
| 1940 | |
| 1941 | if (isUniform(Ptr)) |
| 1942 | HasDependenceInvolvingLoopInvariantAddress |= |
| 1943 | !UniformStores.insert(Ptr).second; |
| 1944 | |
| 1945 | // If we did *not* see this pointer before, insert it to the read-write |
| 1946 | // list. At this phase it is only a 'write' list. |
| 1947 | if (Seen.insert(Ptr).second) { |
| 1948 | ++NumReadWrites; |
| 1949 | |
| 1950 | MemoryLocation Loc = MemoryLocation::get(ST); |
| 1951 | // The TBAA metadata could have a control dependency on the predication |
| 1952 | // condition, so we cannot rely on it when determining whether or not we |
| 1953 | // need runtime pointer checks. |
| 1954 | if (blockNeedsPredication(ST->getParent(), TheLoop, DT)) |
| 1955 | Loc.AATags.TBAA = nullptr; |
| 1956 | |
| 1957 | Accesses.addStore(Loc); |
| 1958 | } |
| 1959 | } |
| 1960 | |
| 1961 | if (IsAnnotatedParallel) { |
| 1962 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: A loop annotated parallel, ignore memory dependency " << "checks.\n"; } } while (false) |
| 1963 | dbgs() << "LAA: A loop annotated parallel, ignore memory dependency "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: A loop annotated parallel, ignore memory dependency " << "checks.\n"; } } while (false) |
| 1964 | << "checks.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: A loop annotated parallel, ignore memory dependency " << "checks.\n"; } } while (false); |
| 1965 | CanVecMem = true; |
| 1966 | return; |
| 1967 | } |
| 1968 | |
| 1969 | for (LoadInst *LD : Loads) { |
| 1970 | Value *Ptr = LD->getPointerOperand(); |
| 1971 | // If we did *not* see this pointer before, insert it to the |
| 1972 | // read list. If we *did* see it before, then it is already in |
| 1973 | // the read-write list. This allows us to vectorize expressions |
| 1974 | // such as A[i] += x; Because the address of A[i] is a read-write |
| 1975 | // pointer. This only works if the index of A[i] is consecutive. |
| 1976 | // If the address of i is unknown (for example A[B[i]]) then we may |
| 1977 | // read a few words, modify, and write a few words, and some of the |
| 1978 | // words may be written to the same address. |
| 1979 | bool IsReadOnlyPtr = false; |
| 1980 | if (Seen.insert(Ptr).second || |
| 1981 | !getPtrStride(*PSE, Ptr, TheLoop, SymbolicStrides)) { |
| 1982 | ++NumReads; |
| 1983 | IsReadOnlyPtr = true; |
| 1984 | } |
| 1985 | |
| 1986 | // See if there is an unsafe dependency between a load to a uniform address and |
| 1987 | // store to the same uniform address. |
| 1988 | if (UniformStores.count(Ptr)) { |
| 1989 | LLVM_DEBUG(dbgs() << "LAA: Found an unsafe dependency between a uniform "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found an unsafe dependency between a uniform " "load and uniform store to the same address!\n"; } } while ( false) |
| 1990 | "load and uniform store to the same address!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found an unsafe dependency between a uniform " "load and uniform store to the same address!\n"; } } while ( false); |
| 1991 | HasDependenceInvolvingLoopInvariantAddress = true; |
| 1992 | } |
| 1993 | |
| 1994 | MemoryLocation Loc = MemoryLocation::get(LD); |
| 1995 | // The TBAA metadata could have a control dependency on the predication |
| 1996 | // condition, so we cannot rely on it when determining whether or not we |
| 1997 | // need runtime pointer checks. |
| 1998 | if (blockNeedsPredication(LD->getParent(), TheLoop, DT)) |
| 1999 | Loc.AATags.TBAA = nullptr; |
| 2000 | |
| 2001 | Accesses.addLoad(Loc, IsReadOnlyPtr); |
| 2002 | } |
| 2003 | |
| 2004 | // If we write (or read-write) to a single destination and there are no |
| 2005 | // other reads in this loop then is it safe to vectorize. |
| 2006 | if (NumReadWrites == 1 && NumReads == 0) { |
| 2007 | LLVM_DEBUG(dbgs() << "LAA: Found a write-only loop!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a write-only loop!\n" ; } } while (false); |
| 2008 | CanVecMem = true; |
| 2009 | return; |
| 2010 | } |
| 2011 | |
| 2012 | // Build dependence sets and check whether we need a runtime pointer bounds |
| 2013 | // check. |
| 2014 | Accesses.buildDependenceSets(); |
| 2015 | |
| 2016 | // Find pointers with computable bounds. We are going to use this information |
| 2017 | // to place a runtime bound check. |
| 2018 | bool CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, PSE->getSE(), |
| 2019 | TheLoop, SymbolicStrides); |
| 2020 | if (!CanDoRTIfNeeded) { |
| 2021 | recordAnalysis("CantIdentifyArrayBounds") << "cannot identify array bounds"; |
| 2022 | LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because we can't find "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We can't vectorize because we can't find " << "the array bounds.\n"; } } while (false) |
| 2023 | << "the array bounds.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We can't vectorize because we can't find " << "the array bounds.\n"; } } while (false); |
| 2024 | CanVecMem = false; |
| 2025 | return; |
| 2026 | } |
| 2027 | |
| 2028 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n" ; } } while (false) |
| 2029 | dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n" ; } } while (false); |
| 2030 | |
| 2031 | CanVecMem = true; |
| 2032 | if (Accesses.isDependencyCheckNeeded()) { |
| 2033 | LLVM_DEBUG(dbgs() << "LAA: Checking memory dependencies\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Checking memory dependencies\n" ; } } while (false); |
| 2034 | CanVecMem = DepChecker->areDepsSafe( |
| 2035 | DependentAccesses, Accesses.getDependenciesToCheck(), SymbolicStrides); |
| 2036 | MaxSafeDepDistBytes = DepChecker->getMaxSafeDepDistBytes(); |
| 2037 | |
| 2038 | if (!CanVecMem && DepChecker->shouldRetryWithRuntimeCheck()) { |
| 2039 | LLVM_DEBUG(dbgs() << "LAA: Retrying with memory checks\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Retrying with memory checks\n" ; } } while (false); |
| 2040 | |
| 2041 | // Clear the dependency checks. We assume they are not needed. |
| 2042 | Accesses.resetDepChecks(*DepChecker); |
| 2043 | |
| 2044 | PtrRtChecking->reset(); |
| 2045 | PtrRtChecking->Need = true; |
| 2046 | |
| 2047 | auto *SE = PSE->getSE(); |
| 2048 | CanDoRTIfNeeded = Accesses.canCheckPtrAtRT(*PtrRtChecking, SE, TheLoop, |
| 2049 | SymbolicStrides, true); |
| 2050 | |
| 2051 | // Check that we found the bounds for the pointer. |
| 2052 | if (!CanDoRTIfNeeded) { |
| 2053 | recordAnalysis("CantCheckMemDepsAtRunTime") |
| 2054 | << "cannot check memory dependencies at runtime"; |
| 2055 | LLVM_DEBUG(dbgs() << "LAA: Can't vectorize with memory checks\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Can't vectorize with memory checks\n" ; } } while (false); |
| 2056 | CanVecMem = false; |
| 2057 | return; |
| 2058 | } |
| 2059 | |
| 2060 | CanVecMem = true; |
| 2061 | } |
| 2062 | } |
| 2063 | |
| 2064 | if (HasConvergentOp) { |
| 2065 | recordAnalysis("CantInsertRuntimeCheckWithConvergent") |
| 2066 | << "cannot add control dependency to convergent operation"; |
| 2067 | LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because a runtime check "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We can't vectorize because a runtime check " "would be needed with a convergent operation\n"; } } while ( false) |
| 2068 | "would be needed with a convergent operation\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: We can't vectorize because a runtime check " "would be needed with a convergent operation\n"; } } while ( false); |
| 2069 | CanVecMem = false; |
| 2070 | return; |
| 2071 | } |
| 2072 | |
| 2073 | if (CanVecMem) |
| 2074 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: No unsafe dependent memory operations in loop. We" << (PtrRtChecking->Need ? "" : " don't") << " need runtime memory checks.\n" ; } } while (false) |
| 2075 | dbgs() << "LAA: No unsafe dependent memory operations in loop. We"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: No unsafe dependent memory operations in loop. We" << (PtrRtChecking->Need ? "" : " don't") << " need runtime memory checks.\n" ; } } while (false) |
| 2076 | << (PtrRtChecking->Need ? "" : " don't")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: No unsafe dependent memory operations in loop. We" << (PtrRtChecking->Need ? "" : " don't") << " need runtime memory checks.\n" ; } } while (false) |
| 2077 | << " need runtime memory checks.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: No unsafe dependent memory operations in loop. We" << (PtrRtChecking->Need ? "" : " don't") << " need runtime memory checks.\n" ; } } while (false); |
| 2078 | else { |
| 2079 | recordAnalysis("UnsafeMemDep") |
| 2080 | << "unsafe dependent memory operations in loop. Use " |
| 2081 | "#pragma loop distribute(enable) to allow loop distribution " |
| 2082 | "to attempt to isolate the offending operations into a separate " |
| 2083 | "loop"; |
| 2084 | LLVM_DEBUG(dbgs() << "LAA: unsafe dependent memory operations in loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: unsafe dependent memory operations in loop\n" ; } } while (false); |
| 2085 | } |
| 2086 | } |
| 2087 | |
| 2088 | bool LoopAccessInfo::blockNeedsPredication(BasicBlock *BB, Loop *TheLoop, |
| 2089 | DominatorTree *DT) { |
| 2090 | assert(TheLoop->contains(BB) && "Unknown block used")((TheLoop->contains(BB) && "Unknown block used") ? static_cast<void> (0) : __assert_fail ("TheLoop->contains(BB) && \"Unknown block used\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 2090, __PRETTY_FUNCTION__)); |
| 2091 | |
| 2092 | // Blocks that do not dominate the latch need predication. |
| 2093 | BasicBlock* Latch = TheLoop->getLoopLatch(); |
| 2094 | return !DT->dominates(BB, Latch); |
| 2095 | } |
| 2096 | |
| 2097 | OptimizationRemarkAnalysis &LoopAccessInfo::recordAnalysis(StringRef RemarkName, |
| 2098 | Instruction *I) { |
| 2099 | assert(!Report && "Multiple reports generated")((!Report && "Multiple reports generated") ? static_cast <void> (0) : __assert_fail ("!Report && \"Multiple reports generated\"" , "/build/llvm-toolchain-snapshot-12~++20210107111113+c9154e8fa377/llvm/lib/Analysis/LoopAccessAnalysis.cpp" , 2099, __PRETTY_FUNCTION__)); |
| 2100 | |
| 2101 | Value *CodeRegion = TheLoop->getHeader(); |
| 2102 | DebugLoc DL = TheLoop->getStartLoc(); |
| 2103 | |
| 2104 | if (I) { |
| 2105 | CodeRegion = I->getParent(); |
| 2106 | // If there is no debug location attached to the instruction, revert back to |
| 2107 | // using the loop's. |
| 2108 | if (I->getDebugLoc()) |
| 2109 | DL = I->getDebugLoc(); |
| 2110 | } |
| 2111 | |
| 2112 | Report = std::make_unique<OptimizationRemarkAnalysis>(DEBUG_TYPE"loop-accesses", RemarkName, DL, |
| 2113 | CodeRegion); |
| 2114 | return *Report; |
| 2115 | } |
| 2116 | |
| 2117 | bool LoopAccessInfo::isUniform(Value *V) const { |
| 2118 | auto *SE = PSE->getSE(); |
| 2119 | // Since we rely on SCEV for uniformity, if the type is not SCEVable, it is |
| 2120 | // never considered uniform. |
| 2121 | // TODO: Is this really what we want? Even without FP SCEV, we may want some |
| 2122 | // trivially loop-invariant FP values to be considered uniform. |
| 2123 | if (!SE->isSCEVable(V->getType())) |
| 2124 | return false; |
| 2125 | return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); |
| 2126 | } |
| 2127 | |
| 2128 | void LoopAccessInfo::collectStridedAccess(Value *MemAccess) { |
| 2129 | Value *Ptr = getLoadStorePointerOperand(MemAccess); |
| 2130 | if (!Ptr) |
| 2131 | return; |
| 2132 | |
| 2133 | Value *Stride = getStrideFromPointer(Ptr, PSE->getSE(), TheLoop); |
| 2134 | if (!Stride) |
| 2135 | return; |
| 2136 | |
| 2137 | LLVM_DEBUG(dbgs() << "LAA: Found a strided access that is a candidate for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a strided access that is a candidate for " "versioning:"; } } while (false) |
| 2138 | "versioning:")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a strided access that is a candidate for " "versioning:"; } } while (false); |
| 2139 | LLVM_DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n"; } } while (false ); |
| 2140 | |
| 2141 | // Avoid adding the "Stride == 1" predicate when we know that |
| 2142 | // Stride >= Trip-Count. Such a predicate will effectively optimize a single |
| 2143 | // or zero iteration loop, as Trip-Count <= Stride == 1. |
| 2144 | // |
| 2145 | // TODO: We are currently not making a very informed decision on when it is |
| 2146 | // beneficial to apply stride versioning. It might make more sense that the |
| 2147 | // users of this analysis (such as the vectorizer) will trigger it, based on |
| 2148 | // their specific cost considerations; For example, in cases where stride |
| 2149 | // versioning does not help resolving memory accesses/dependences, the |
| 2150 | // vectorizer should evaluate the cost of the runtime test, and the benefit |
| 2151 | // of various possible stride specializations, considering the alternatives |
| 2152 | // of using gather/scatters (if available). |
| 2153 | |
| 2154 | const SCEV *StrideExpr = PSE->getSCEV(Stride); |
| 2155 | const SCEV *BETakenCount = PSE->getBackedgeTakenCount(); |
| 2156 | |
| 2157 | // Match the types so we can compare the stride and the BETakenCount. |
| 2158 | // The Stride can be positive/negative, so we sign extend Stride; |
| 2159 | // The backedgeTakenCount is non-negative, so we zero extend BETakenCount. |
| 2160 | const DataLayout &DL = TheLoop->getHeader()->getModule()->getDataLayout(); |
| 2161 | uint64_t StrideTypeSize = DL.getTypeAllocSize(StrideExpr->getType()); |
| 2162 | uint64_t BETypeSize = DL.getTypeAllocSize(BETakenCount->getType()); |
| 2163 | const SCEV *CastedStride = StrideExpr; |
| 2164 | const SCEV *CastedBECount = BETakenCount; |
| 2165 | ScalarEvolution *SE = PSE->getSE(); |
| 2166 | if (BETypeSize >= StrideTypeSize) |
| 2167 | CastedStride = SE->getNoopOrSignExtend(StrideExpr, BETakenCount->getType()); |
| 2168 | else |
| 2169 | CastedBECount = SE->getZeroExtendExpr(BETakenCount, StrideExpr->getType()); |
| 2170 | const SCEV *StrideMinusBETaken = SE->getMinusSCEV(CastedStride, CastedBECount); |
| 2171 | // Since TripCount == BackEdgeTakenCount + 1, checking: |
| 2172 | // "Stride >= TripCount" is equivalent to checking: |
| 2173 | // Stride - BETakenCount > 0 |
| 2174 | if (SE->isKnownPositive(StrideMinusBETaken)) { |
| 2175 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Stride>=TripCount; No point in versioning as the " "Stride==1 predicate will imply that the loop executes " "at most once.\n" ; } } while (false) |
| 2176 | dbgs() << "LAA: Stride>=TripCount; No point in versioning as the "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Stride>=TripCount; No point in versioning as the " "Stride==1 predicate will imply that the loop executes " "at most once.\n" ; } } while (false) |
| 2177 | "Stride==1 predicate will imply that the loop executes "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Stride>=TripCount; No point in versioning as the " "Stride==1 predicate will imply that the loop executes " "at most once.\n" ; } } while (false) |
| 2178 | "at most once.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Stride>=TripCount; No point in versioning as the " "Stride==1 predicate will imply that the loop executes " "at most once.\n" ; } } while (false); |
| 2179 | return; |
| 2180 | } |
| 2181 | LLVM_DEBUG(dbgs() << "LAA: Found a strided access that we can version.")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-accesses")) { dbgs() << "LAA: Found a strided access that we can version." ; } } while (false); |
| 2182 | |
| 2183 | SymbolicStrides[Ptr] = Stride; |
| 2184 | StrideSet.insert(Stride); |
| 2185 | } |
| 2186 | |
| 2187 | LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE, |
| 2188 | const TargetLibraryInfo *TLI, AAResults *AA, |
| 2189 | DominatorTree *DT, LoopInfo *LI) |
| 2190 | : PSE(std::make_unique<PredicatedScalarEvolution>(*SE, *L)), |
| 2191 | PtrRtChecking(std::make_unique<RuntimePointerChecking>(SE)), |
| 2192 | DepChecker(std::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L), |
| 2193 | NumLoads(0), NumStores(0), MaxSafeDepDistBytes(-1), CanVecMem(false), |
| 2194 | HasConvergentOp(false), |
| 2195 | HasDependenceInvolvingLoopInvariantAddress(false) { |
| 2196 | if (canAnalyzeLoop()) |
| 2197 | analyzeLoop(AA, LI, TLI, DT); |
| 2198 | } |
| 2199 | |
| 2200 | void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const { |
| 2201 | if (CanVecMem) { |
| 2202 | OS.indent(Depth) << "Memory dependences are safe"; |
| 2203 | if (MaxSafeDepDistBytes != -1ULL) |
| 2204 | OS << " with a maximum dependence distance of " << MaxSafeDepDistBytes |
| 2205 | << " bytes"; |
| 2206 | if (PtrRtChecking->Need) |
| 2207 | OS << " with run-time checks"; |
| 2208 | OS << "\n"; |
| 2209 | } |
| 2210 | |
| 2211 | if (HasConvergentOp) |
| 2212 | OS.indent(Depth) << "Has convergent operation in loop\n"; |
| 2213 | |
| 2214 | if (Report) |
| 2215 | OS.indent(Depth) << "Report: " << Report->getMsg() << "\n"; |
| 2216 | |
| 2217 | if (auto *Dependences = DepChecker->getDependences()) { |
| 2218 | OS.indent(Depth) << "Dependences:\n"; |
| 2219 | for (auto &Dep : *Dependences) { |
| 2220 | Dep.print(OS, Depth + 2, DepChecker->getMemoryInstructions()); |
| 2221 | OS << "\n"; |
| 2222 | } |
| 2223 | } else |
| 2224 | OS.indent(Depth) << "Too many dependences, not recorded\n"; |
| 2225 | |
| 2226 | // List the pair of accesses need run-time checks to prove independence. |
| 2227 | PtrRtChecking->print(OS, Depth); |
| 2228 | OS << "\n"; |
| 2229 | |
| 2230 | OS.indent(Depth) << "Non vectorizable stores to invariant address were " |
| 2231 | << (HasDependenceInvolvingLoopInvariantAddress ? "" : "not ") |
| 2232 | << "found in loop.\n"; |
| 2233 | |
| 2234 | OS.indent(Depth) << "SCEV assumptions:\n"; |
| 2235 | PSE->getUnionPredicate().print(OS, Depth); |
| 2236 | |
| 2237 | OS << "\n"; |
| 2238 | |
| 2239 | OS.indent(Depth) << "Expressions re-written:\n"; |
| 2240 | PSE->print(OS, Depth); |
| 2241 | } |
| 2242 | |
| 2243 | LoopAccessLegacyAnalysis::LoopAccessLegacyAnalysis() : FunctionPass(ID) { |
| 2244 | initializeLoopAccessLegacyAnalysisPass(*PassRegistry::getPassRegistry()); |
| 2245 | } |
| 2246 | |
| 2247 | const LoopAccessInfo &LoopAccessLegacyAnalysis::getInfo(Loop *L) { |
| 2248 | auto &LAI = LoopAccessInfoMap[L]; |
| 2249 | |
| 2250 | if (!LAI) |
| 2251 | LAI = std::make_unique<LoopAccessInfo>(L, SE, TLI, AA, DT, LI); |
| 2252 | |
| 2253 | return *LAI.get(); |
| 2254 | } |
| 2255 | |
| 2256 | void LoopAccessLegacyAnalysis::print(raw_ostream &OS, const Module *M) const { |
| 2257 | LoopAccessLegacyAnalysis &LAA = *const_cast<LoopAccessLegacyAnalysis *>(this); |
| 2258 | |
| 2259 | for (Loop *TopLevelLoop : *LI) |
| 2260 | for (Loop *L : depth_first(TopLevelLoop)) { |
| 2261 | OS.indent(2) << L->getHeader()->getName() << ":\n"; |
| 2262 | auto &LAI = LAA.getInfo(L); |
| 2263 | LAI.print(OS, 4); |
| 2264 | } |
| 2265 | } |
| 2266 | |
| 2267 | bool LoopAccessLegacyAnalysis::runOnFunction(Function &F) { |
| 2268 | SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
| 2269 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); |
| 2270 | TLI = TLIP ? &TLIP->getTLI(F) : nullptr; |
| 2271 | AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); |
| 2272 | DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
| 2273 | LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
| 2274 | |
| 2275 | return false; |
| 2276 | } |
| 2277 | |
| 2278 | void LoopAccessLegacyAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { |
| 2279 | AU.addRequired<ScalarEvolutionWrapperPass>(); |
| 2280 | AU.addRequired<AAResultsWrapperPass>(); |
| 2281 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 2282 | AU.addRequired<LoopInfoWrapperPass>(); |
| 2283 | |
| 2284 | AU.setPreservesAll(); |
| 2285 | } |
| 2286 | |
| 2287 | char LoopAccessLegacyAnalysis::ID = 0; |
| 2288 | static const char laa_name[] = "Loop Access Analysis"; |
| 2289 | #define LAA_NAME"loop-accesses" "loop-accesses" |
| 2290 | |
| 2291 | INITIALIZE_PASS_BEGIN(LoopAccessLegacyAnalysis, LAA_NAME, laa_name, false, true)static void *initializeLoopAccessLegacyAnalysisPassOnce(PassRegistry &Registry) { |
| 2292 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); |
| 2293 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); |
| 2294 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); |
| 2295 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); |
| 2296 | 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 )); } |
| 2297 | |
| 2298 | AnalysisKey LoopAccessAnalysis::Key; |
| 2299 | |
| 2300 | LoopAccessInfo LoopAccessAnalysis::run(Loop &L, LoopAnalysisManager &AM, |
| 2301 | LoopStandardAnalysisResults &AR) { |
| 2302 | return LoopAccessInfo(&L, &AR.SE, &AR.TLI, &AR.AA, &AR.DT, &AR.LI); |
| 2303 | } |
| 2304 | |
| 2305 | namespace llvm { |
| 2306 | |
| 2307 | Pass *createLAAPass() { |
| 2308 | return new LoopAccessLegacyAnalysis(); |
| 2309 | } |
| 2310 | |
| 2311 | } // end namespace llvm |