1 //===- BasicAliasAnalysis.cpp - Stateless Alias Analysis Impl -------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file defines the primary stateless implementation of the 11 // Alias Analysis interface that implements identities (two different 12 // globals cannot alias, etc), but does no stateful analysis. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "llvm/Analysis/BasicAliasAnalysis.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/ADT/Statistic.h" 19 #include "llvm/Analysis/AliasAnalysis.h" 20 #include "llvm/Analysis/CFG.h" 21 #include "llvm/Analysis/CaptureTracking.h" 22 #include "llvm/Analysis/InstructionSimplify.h" 23 #include "llvm/Analysis/LoopInfo.h" 24 #include "llvm/Analysis/MemoryBuiltins.h" 25 #include "llvm/Analysis/ValueTracking.h" 26 #include "llvm/Analysis/AssumptionCache.h" 27 #include "llvm/IR/Constants.h" 28 #include "llvm/IR/DataLayout.h" 29 #include "llvm/IR/DerivedTypes.h" 30 #include "llvm/IR/Dominators.h" 31 #include "llvm/IR/GlobalAlias.h" 32 #include "llvm/IR/GlobalVariable.h" 33 #include "llvm/IR/Instructions.h" 34 #include "llvm/IR/IntrinsicInst.h" 35 #include "llvm/IR/LLVMContext.h" 36 #include "llvm/IR/Operator.h" 37 #include "llvm/Pass.h" 38 #include "llvm/Support/ErrorHandling.h" 39 #include <algorithm> 40 41 #define DEBUG_TYPE "basicaa" 42 43 using namespace llvm; 44 45 /// Enable analysis of recursive PHI nodes. 46 static cl::opt<bool> EnableRecPhiAnalysis("basicaa-recphi", cl::Hidden, 47 cl::init(false)); 48 /// SearchLimitReached / SearchTimes shows how often the limit of 49 /// to decompose GEPs is reached. It will affect the precision 50 /// of basic alias analysis. 51 STATISTIC(SearchLimitReached, "Number of times the limit to " 52 "decompose GEPs is reached"); 53 STATISTIC(SearchTimes, "Number of times a GEP is decomposed"); 54 55 /// Cutoff after which to stop analysing a set of phi nodes potentially involved 56 /// in a cycle. Because we are analysing 'through' phi nodes, we need to be 57 /// careful with value equivalence. We use reachability to make sure a value 58 /// cannot be involved in a cycle. 59 const unsigned MaxNumPhiBBsValueReachabilityCheck = 20; 60 61 // The max limit of the search depth in DecomposeGEPExpression() and 62 // GetUnderlyingObject(), both functions need to use the same search 63 // depth otherwise the algorithm in aliasGEP will assert. 64 static const unsigned MaxLookupSearchDepth = 6; 65 66 bool BasicAAResult::invalidate(Function &F, const PreservedAnalyses &PA, 67 FunctionAnalysisManager::Invalidator &Inv) { 68 // We don't care if this analysis itself is preserved, it has no state. But 69 // we need to check that the analyses it depends on have been. Note that we 70 // may be created without handles to some analyses and in that case don't 71 // depend on them. 72 if (Inv.invalidate<AssumptionAnalysis>(F, PA) || 73 (DT && Inv.invalidate<DominatorTreeAnalysis>(F, PA)) || 74 (LI && Inv.invalidate<LoopAnalysis>(F, PA))) 75 return true; 76 77 // Otherwise this analysis result remains valid. 78 return false; 79 } 80 81 //===----------------------------------------------------------------------===// 82 // Useful predicates 83 //===----------------------------------------------------------------------===// 84 85 /// Returns true if the pointer is to a function-local object that never 86 /// escapes from the function. 87 static bool isNonEscapingLocalObject(const Value *V) { 88 // If this is a local allocation, check to see if it escapes. 89 if (isa<AllocaInst>(V) || isNoAliasCall(V)) 90 // Set StoreCaptures to True so that we can assume in our callers that the 91 // pointer is not the result of a load instruction. Currently 92 // PointerMayBeCaptured doesn't have any special analysis for the 93 // StoreCaptures=false case; if it did, our callers could be refined to be 94 // more precise. 95 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true); 96 97 // If this is an argument that corresponds to a byval or noalias argument, 98 // then it has not escaped before entering the function. Check if it escapes 99 // inside the function. 100 if (const Argument *A = dyn_cast<Argument>(V)) 101 if (A->hasByValAttr() || A->hasNoAliasAttr()) 102 // Note even if the argument is marked nocapture, we still need to check 103 // for copies made inside the function. The nocapture attribute only 104 // specifies that there are no copies made that outlive the function. 105 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true); 106 107 return false; 108 } 109 110 /// Returns true if the pointer is one which would have been considered an 111 /// escape by isNonEscapingLocalObject. 112 static bool isEscapeSource(const Value *V) { 113 if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V)) 114 return true; 115 116 // The load case works because isNonEscapingLocalObject considers all 117 // stores to be escapes (it passes true for the StoreCaptures argument 118 // to PointerMayBeCaptured). 119 if (isa<LoadInst>(V)) 120 return true; 121 122 return false; 123 } 124 125 /// Returns the size of the object specified by V or UnknownSize if unknown. 126 static uint64_t getObjectSize(const Value *V, const DataLayout &DL, 127 const TargetLibraryInfo &TLI, 128 bool RoundToAlign = false) { 129 uint64_t Size; 130 if (getObjectSize(V, Size, DL, &TLI, RoundToAlign)) 131 return Size; 132 return MemoryLocation::UnknownSize; 133 } 134 135 /// Returns true if we can prove that the object specified by V is smaller than 136 /// Size. 137 static bool isObjectSmallerThan(const Value *V, uint64_t Size, 138 const DataLayout &DL, 139 const TargetLibraryInfo &TLI) { 140 // Note that the meanings of the "object" are slightly different in the 141 // following contexts: 142 // c1: llvm::getObjectSize() 143 // c2: llvm.objectsize() intrinsic 144 // c3: isObjectSmallerThan() 145 // c1 and c2 share the same meaning; however, the meaning of "object" in c3 146 // refers to the "entire object". 147 // 148 // Consider this example: 149 // char *p = (char*)malloc(100) 150 // char *q = p+80; 151 // 152 // In the context of c1 and c2, the "object" pointed by q refers to the 153 // stretch of memory of q[0:19]. So, getObjectSize(q) should return 20. 154 // 155 // However, in the context of c3, the "object" refers to the chunk of memory 156 // being allocated. So, the "object" has 100 bytes, and q points to the middle 157 // the "object". In case q is passed to isObjectSmallerThan() as the 1st 158 // parameter, before the llvm::getObjectSize() is called to get the size of 159 // entire object, we should: 160 // - either rewind the pointer q to the base-address of the object in 161 // question (in this case rewind to p), or 162 // - just give up. It is up to caller to make sure the pointer is pointing 163 // to the base address the object. 164 // 165 // We go for 2nd option for simplicity. 166 if (!isIdentifiedObject(V)) 167 return false; 168 169 // This function needs to use the aligned object size because we allow 170 // reads a bit past the end given sufficient alignment. 171 uint64_t ObjectSize = getObjectSize(V, DL, TLI, /*RoundToAlign*/ true); 172 173 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize < Size; 174 } 175 176 /// Returns true if we can prove that the object specified by V has size Size. 177 static bool isObjectSize(const Value *V, uint64_t Size, const DataLayout &DL, 178 const TargetLibraryInfo &TLI) { 179 uint64_t ObjectSize = getObjectSize(V, DL, TLI); 180 return ObjectSize != MemoryLocation::UnknownSize && ObjectSize == Size; 181 } 182 183 //===----------------------------------------------------------------------===// 184 // GetElementPtr Instruction Decomposition and Analysis 185 //===----------------------------------------------------------------------===// 186 187 /// Analyzes the specified value as a linear expression: "A*V + B", where A and 188 /// B are constant integers. 189 /// 190 /// Returns the scale and offset values as APInts and return V as a Value*, and 191 /// return whether we looked through any sign or zero extends. The incoming 192 /// Value is known to have IntegerType, and it may already be sign or zero 193 /// extended. 194 /// 195 /// Note that this looks through extends, so the high bits may not be 196 /// represented in the result. 197 /*static*/ const Value *BasicAAResult::GetLinearExpression( 198 const Value *V, APInt &Scale, APInt &Offset, unsigned &ZExtBits, 199 unsigned &SExtBits, const DataLayout &DL, unsigned Depth, 200 AssumptionCache *AC, DominatorTree *DT, bool &NSW, bool &NUW) { 201 assert(V->getType()->isIntegerTy() && "Not an integer value"); 202 203 // Limit our recursion depth. 204 if (Depth == 6) { 205 Scale = 1; 206 Offset = 0; 207 return V; 208 } 209 210 if (const ConstantInt *Const = dyn_cast<ConstantInt>(V)) { 211 // If it's a constant, just convert it to an offset and remove the variable. 212 // If we've been called recursively, the Offset bit width will be greater 213 // than the constant's (the Offset's always as wide as the outermost call), 214 // so we'll zext here and process any extension in the isa<SExtInst> & 215 // isa<ZExtInst> cases below. 216 Offset += Const->getValue().zextOrSelf(Offset.getBitWidth()); 217 assert(Scale == 0 && "Constant values don't have a scale"); 218 return V; 219 } 220 221 if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) { 222 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) { 223 224 // If we've been called recursively, then Offset and Scale will be wider 225 // than the BOp operands. We'll always zext it here as we'll process sign 226 // extensions below (see the isa<SExtInst> / isa<ZExtInst> cases). 227 APInt RHS = RHSC->getValue().zextOrSelf(Offset.getBitWidth()); 228 229 switch (BOp->getOpcode()) { 230 default: 231 // We don't understand this instruction, so we can't decompose it any 232 // further. 233 Scale = 1; 234 Offset = 0; 235 return V; 236 case Instruction::Or: 237 // X|C == X+C if all the bits in C are unset in X. Otherwise we can't 238 // analyze it. 239 if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), DL, 0, AC, 240 BOp, DT)) { 241 Scale = 1; 242 Offset = 0; 243 return V; 244 } 245 LLVM_FALLTHROUGH; 246 case Instruction::Add: 247 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits, 248 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW); 249 Offset += RHS; 250 break; 251 case Instruction::Sub: 252 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits, 253 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW); 254 Offset -= RHS; 255 break; 256 case Instruction::Mul: 257 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits, 258 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW); 259 Offset *= RHS; 260 Scale *= RHS; 261 break; 262 case Instruction::Shl: 263 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, ZExtBits, 264 SExtBits, DL, Depth + 1, AC, DT, NSW, NUW); 265 Offset <<= RHS.getLimitedValue(); 266 Scale <<= RHS.getLimitedValue(); 267 // the semantics of nsw and nuw for left shifts don't match those of 268 // multiplications, so we won't propagate them. 269 NSW = NUW = false; 270 return V; 271 } 272 273 if (isa<OverflowingBinaryOperator>(BOp)) { 274 NUW &= BOp->hasNoUnsignedWrap(); 275 NSW &= BOp->hasNoSignedWrap(); 276 } 277 return V; 278 } 279 } 280 281 // Since GEP indices are sign extended anyway, we don't care about the high 282 // bits of a sign or zero extended value - just scales and offsets. The 283 // extensions have to be consistent though. 284 if (isa<SExtInst>(V) || isa<ZExtInst>(V)) { 285 Value *CastOp = cast<CastInst>(V)->getOperand(0); 286 unsigned NewWidth = V->getType()->getPrimitiveSizeInBits(); 287 unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits(); 288 unsigned OldZExtBits = ZExtBits, OldSExtBits = SExtBits; 289 const Value *Result = 290 GetLinearExpression(CastOp, Scale, Offset, ZExtBits, SExtBits, DL, 291 Depth + 1, AC, DT, NSW, NUW); 292 293 // zext(zext(%x)) == zext(%x), and similarly for sext; we'll handle this 294 // by just incrementing the number of bits we've extended by. 295 unsigned ExtendedBy = NewWidth - SmallWidth; 296 297 if (isa<SExtInst>(V) && ZExtBits == 0) { 298 // sext(sext(%x, a), b) == sext(%x, a + b) 299 300 if (NSW) { 301 // We haven't sign-wrapped, so it's valid to decompose sext(%x + c) 302 // into sext(%x) + sext(c). We'll sext the Offset ourselves: 303 unsigned OldWidth = Offset.getBitWidth(); 304 Offset = Offset.trunc(SmallWidth).sext(NewWidth).zextOrSelf(OldWidth); 305 } else { 306 // We may have signed-wrapped, so don't decompose sext(%x + c) into 307 // sext(%x) + sext(c) 308 Scale = 1; 309 Offset = 0; 310 Result = CastOp; 311 ZExtBits = OldZExtBits; 312 SExtBits = OldSExtBits; 313 } 314 SExtBits += ExtendedBy; 315 } else { 316 // sext(zext(%x, a), b) = zext(zext(%x, a), b) = zext(%x, a + b) 317 318 if (!NUW) { 319 // We may have unsigned-wrapped, so don't decompose zext(%x + c) into 320 // zext(%x) + zext(c) 321 Scale = 1; 322 Offset = 0; 323 Result = CastOp; 324 ZExtBits = OldZExtBits; 325 SExtBits = OldSExtBits; 326 } 327 ZExtBits += ExtendedBy; 328 } 329 330 return Result; 331 } 332 333 Scale = 1; 334 Offset = 0; 335 return V; 336 } 337 338 /// To ensure a pointer offset fits in an integer of size PointerSize 339 /// (in bits) when that size is smaller than 64. This is an issue in 340 /// particular for 32b programs with negative indices that rely on two's 341 /// complement wrap-arounds for precise alias information. 342 static int64_t adjustToPointerSize(int64_t Offset, unsigned PointerSize) { 343 assert(PointerSize <= 64 && "Invalid PointerSize!"); 344 unsigned ShiftBits = 64 - PointerSize; 345 return (int64_t)((uint64_t)Offset << ShiftBits) >> ShiftBits; 346 } 347 348 /// If V is a symbolic pointer expression, decompose it into a base pointer 349 /// with a constant offset and a number of scaled symbolic offsets. 350 /// 351 /// The scaled symbolic offsets (represented by pairs of a Value* and a scale 352 /// in the VarIndices vector) are Value*'s that are known to be scaled by the 353 /// specified amount, but which may have other unrepresented high bits. As 354 /// such, the gep cannot necessarily be reconstructed from its decomposed form. 355 /// 356 /// When DataLayout is around, this function is capable of analyzing everything 357 /// that GetUnderlyingObject can look through. To be able to do that 358 /// GetUnderlyingObject and DecomposeGEPExpression must use the same search 359 /// depth (MaxLookupSearchDepth). When DataLayout not is around, it just looks 360 /// through pointer casts. 361 bool BasicAAResult::DecomposeGEPExpression(const Value *V, 362 DecomposedGEP &Decomposed, const DataLayout &DL, AssumptionCache *AC, 363 DominatorTree *DT) { 364 // Limit recursion depth to limit compile time in crazy cases. 365 unsigned MaxLookup = MaxLookupSearchDepth; 366 SearchTimes++; 367 368 Decomposed.StructOffset = 0; 369 Decomposed.OtherOffset = 0; 370 Decomposed.VarIndices.clear(); 371 do { 372 // See if this is a bitcast or GEP. 373 const Operator *Op = dyn_cast<Operator>(V); 374 if (!Op) { 375 // The only non-operator case we can handle are GlobalAliases. 376 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 377 if (!GA->isInterposable()) { 378 V = GA->getAliasee(); 379 continue; 380 } 381 } 382 Decomposed.Base = V; 383 return false; 384 } 385 386 if (Op->getOpcode() == Instruction::BitCast || 387 Op->getOpcode() == Instruction::AddrSpaceCast) { 388 V = Op->getOperand(0); 389 continue; 390 } 391 392 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op); 393 if (!GEPOp) { 394 if (auto CS = ImmutableCallSite(V)) 395 if (const Value *RV = CS.getReturnedArgOperand()) { 396 V = RV; 397 continue; 398 } 399 400 // If it's not a GEP, hand it off to SimplifyInstruction to see if it 401 // can come up with something. This matches what GetUnderlyingObject does. 402 if (const Instruction *I = dyn_cast<Instruction>(V)) 403 // TODO: Get a DominatorTree and AssumptionCache and use them here 404 // (these are both now available in this function, but this should be 405 // updated when GetUnderlyingObject is updated). TLI should be 406 // provided also. 407 if (const Value *Simplified = 408 SimplifyInstruction(const_cast<Instruction *>(I), DL)) { 409 V = Simplified; 410 continue; 411 } 412 413 Decomposed.Base = V; 414 return false; 415 } 416 417 // Don't attempt to analyze GEPs over unsized objects. 418 if (!GEPOp->getSourceElementType()->isSized()) { 419 Decomposed.Base = V; 420 return false; 421 } 422 423 unsigned AS = GEPOp->getPointerAddressSpace(); 424 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices. 425 gep_type_iterator GTI = gep_type_begin(GEPOp); 426 unsigned PointerSize = DL.getPointerSizeInBits(AS); 427 // Assume all GEP operands are constants until proven otherwise. 428 bool GepHasConstantOffset = true; 429 for (User::const_op_iterator I = GEPOp->op_begin() + 1, E = GEPOp->op_end(); 430 I != E; ++I, ++GTI) { 431 const Value *Index = *I; 432 // Compute the (potentially symbolic) offset in bytes for this index. 433 if (StructType *STy = GTI.getStructTypeOrNull()) { 434 // For a struct, add the member offset. 435 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue(); 436 if (FieldNo == 0) 437 continue; 438 439 Decomposed.StructOffset += 440 DL.getStructLayout(STy)->getElementOffset(FieldNo); 441 continue; 442 } 443 444 // For an array/pointer, add the element offset, explicitly scaled. 445 if (const ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) { 446 if (CIdx->isZero()) 447 continue; 448 Decomposed.OtherOffset += 449 DL.getTypeAllocSize(GTI.getIndexedType()) * CIdx->getSExtValue(); 450 continue; 451 } 452 453 GepHasConstantOffset = false; 454 455 uint64_t Scale = DL.getTypeAllocSize(GTI.getIndexedType()); 456 unsigned ZExtBits = 0, SExtBits = 0; 457 458 // If the integer type is smaller than the pointer size, it is implicitly 459 // sign extended to pointer size. 460 unsigned Width = Index->getType()->getIntegerBitWidth(); 461 if (PointerSize > Width) 462 SExtBits += PointerSize - Width; 463 464 // Use GetLinearExpression to decompose the index into a C1*V+C2 form. 465 APInt IndexScale(Width, 0), IndexOffset(Width, 0); 466 bool NSW = true, NUW = true; 467 Index = GetLinearExpression(Index, IndexScale, IndexOffset, ZExtBits, 468 SExtBits, DL, 0, AC, DT, NSW, NUW); 469 470 // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale. 471 // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale. 472 Decomposed.OtherOffset += IndexOffset.getSExtValue() * Scale; 473 Scale *= IndexScale.getSExtValue(); 474 475 // If we already had an occurrence of this index variable, merge this 476 // scale into it. For example, we want to handle: 477 // A[x][x] -> x*16 + x*4 -> x*20 478 // This also ensures that 'x' only appears in the index list once. 479 for (unsigned i = 0, e = Decomposed.VarIndices.size(); i != e; ++i) { 480 if (Decomposed.VarIndices[i].V == Index && 481 Decomposed.VarIndices[i].ZExtBits == ZExtBits && 482 Decomposed.VarIndices[i].SExtBits == SExtBits) { 483 Scale += Decomposed.VarIndices[i].Scale; 484 Decomposed.VarIndices.erase(Decomposed.VarIndices.begin() + i); 485 break; 486 } 487 } 488 489 // Make sure that we have a scale that makes sense for this target's 490 // pointer size. 491 Scale = adjustToPointerSize(Scale, PointerSize); 492 493 if (Scale) { 494 VariableGEPIndex Entry = {Index, ZExtBits, SExtBits, 495 static_cast<int64_t>(Scale)}; 496 Decomposed.VarIndices.push_back(Entry); 497 } 498 } 499 500 // Take care of wrap-arounds 501 if (GepHasConstantOffset) { 502 Decomposed.StructOffset = 503 adjustToPointerSize(Decomposed.StructOffset, PointerSize); 504 Decomposed.OtherOffset = 505 adjustToPointerSize(Decomposed.OtherOffset, PointerSize); 506 } 507 508 // Analyze the base pointer next. 509 V = GEPOp->getOperand(0); 510 } while (--MaxLookup); 511 512 // If the chain of expressions is too deep, just return early. 513 Decomposed.Base = V; 514 SearchLimitReached++; 515 return true; 516 } 517 518 /// Returns whether the given pointer value points to memory that is local to 519 /// the function, with global constants being considered local to all 520 /// functions. 521 bool BasicAAResult::pointsToConstantMemory(const MemoryLocation &Loc, 522 bool OrLocal) { 523 assert(Visited.empty() && "Visited must be cleared after use!"); 524 525 unsigned MaxLookup = 8; 526 SmallVector<const Value *, 16> Worklist; 527 Worklist.push_back(Loc.Ptr); 528 do { 529 const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), DL); 530 if (!Visited.insert(V).second) { 531 Visited.clear(); 532 return AAResultBase::pointsToConstantMemory(Loc, OrLocal); 533 } 534 535 // An alloca instruction defines local memory. 536 if (OrLocal && isa<AllocaInst>(V)) 537 continue; 538 539 // A global constant counts as local memory for our purposes. 540 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) { 541 // Note: this doesn't require GV to be "ODR" because it isn't legal for a 542 // global to be marked constant in some modules and non-constant in 543 // others. GV may even be a declaration, not a definition. 544 if (!GV->isConstant()) { 545 Visited.clear(); 546 return AAResultBase::pointsToConstantMemory(Loc, OrLocal); 547 } 548 continue; 549 } 550 551 // If both select values point to local memory, then so does the select. 552 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) { 553 Worklist.push_back(SI->getTrueValue()); 554 Worklist.push_back(SI->getFalseValue()); 555 continue; 556 } 557 558 // If all values incoming to a phi node point to local memory, then so does 559 // the phi. 560 if (const PHINode *PN = dyn_cast<PHINode>(V)) { 561 // Don't bother inspecting phi nodes with many operands. 562 if (PN->getNumIncomingValues() > MaxLookup) { 563 Visited.clear(); 564 return AAResultBase::pointsToConstantMemory(Loc, OrLocal); 565 } 566 for (Value *IncValue : PN->incoming_values()) 567 Worklist.push_back(IncValue); 568 continue; 569 } 570 571 // Otherwise be conservative. 572 Visited.clear(); 573 return AAResultBase::pointsToConstantMemory(Loc, OrLocal); 574 575 } while (!Worklist.empty() && --MaxLookup); 576 577 Visited.clear(); 578 return Worklist.empty(); 579 } 580 581 /// Returns the behavior when calling the given call site. 582 FunctionModRefBehavior BasicAAResult::getModRefBehavior(ImmutableCallSite CS) { 583 if (CS.doesNotAccessMemory()) 584 // Can't do better than this. 585 return FMRB_DoesNotAccessMemory; 586 587 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior; 588 589 // If the callsite knows it only reads memory, don't return worse 590 // than that. 591 if (CS.onlyReadsMemory()) 592 Min = FMRB_OnlyReadsMemory; 593 else if (CS.doesNotReadMemory()) 594 Min = FMRB_DoesNotReadMemory; 595 596 if (CS.onlyAccessesArgMemory()) 597 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees); 598 599 // If CS has operand bundles then aliasing attributes from the function it 600 // calls do not directly apply to the CallSite. This can be made more 601 // precise in the future. 602 if (!CS.hasOperandBundles()) 603 if (const Function *F = CS.getCalledFunction()) 604 Min = 605 FunctionModRefBehavior(Min & getBestAAResults().getModRefBehavior(F)); 606 607 return Min; 608 } 609 610 /// Returns the behavior when calling the given function. For use when the call 611 /// site is not known. 612 FunctionModRefBehavior BasicAAResult::getModRefBehavior(const Function *F) { 613 // If the function declares it doesn't access memory, we can't do better. 614 if (F->doesNotAccessMemory()) 615 return FMRB_DoesNotAccessMemory; 616 617 FunctionModRefBehavior Min = FMRB_UnknownModRefBehavior; 618 619 // If the function declares it only reads memory, go with that. 620 if (F->onlyReadsMemory()) 621 Min = FMRB_OnlyReadsMemory; 622 else if (F->doesNotReadMemory()) 623 Min = FMRB_DoesNotReadMemory; 624 625 if (F->onlyAccessesArgMemory()) 626 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesArgumentPointees); 627 else if (F->onlyAccessesInaccessibleMemory()) 628 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleMem); 629 else if (F->onlyAccessesInaccessibleMemOrArgMem()) 630 Min = FunctionModRefBehavior(Min & FMRB_OnlyAccessesInaccessibleOrArgMem); 631 632 return Min; 633 } 634 635 /// Returns true if this is a writeonly (i.e Mod only) parameter. 636 static bool isWriteOnlyParam(ImmutableCallSite CS, unsigned ArgIdx, 637 const TargetLibraryInfo &TLI) { 638 if (CS.paramHasAttr(ArgIdx + 1, Attribute::WriteOnly)) 639 return true; 640 641 // We can bound the aliasing properties of memset_pattern16 just as we can 642 // for memcpy/memset. This is particularly important because the 643 // LoopIdiomRecognizer likes to turn loops into calls to memset_pattern16 644 // whenever possible. 645 // FIXME Consider handling this in InferFunctionAttr.cpp together with other 646 // attributes. 647 LibFunc::Func F; 648 if (CS.getCalledFunction() && TLI.getLibFunc(*CS.getCalledFunction(), F) && 649 F == LibFunc::memset_pattern16 && TLI.has(F)) 650 if (ArgIdx == 0) 651 return true; 652 653 // TODO: memset_pattern4, memset_pattern8 654 // TODO: _chk variants 655 // TODO: strcmp, strcpy 656 657 return false; 658 } 659 660 ModRefInfo BasicAAResult::getArgModRefInfo(ImmutableCallSite CS, 661 unsigned ArgIdx) { 662 663 // Checking for known builtin intrinsics and target library functions. 664 if (isWriteOnlyParam(CS, ArgIdx, TLI)) 665 return MRI_Mod; 666 667 if (CS.paramHasAttr(ArgIdx + 1, Attribute::ReadOnly)) 668 return MRI_Ref; 669 670 if (CS.paramHasAttr(ArgIdx + 1, Attribute::ReadNone)) 671 return MRI_NoModRef; 672 673 return AAResultBase::getArgModRefInfo(CS, ArgIdx); 674 } 675 676 static bool isIntrinsicCall(ImmutableCallSite CS, Intrinsic::ID IID) { 677 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction()); 678 return II && II->getIntrinsicID() == IID; 679 } 680 681 #ifndef NDEBUG 682 static const Function *getParent(const Value *V) { 683 if (const Instruction *inst = dyn_cast<Instruction>(V)) 684 return inst->getParent()->getParent(); 685 686 if (const Argument *arg = dyn_cast<Argument>(V)) 687 return arg->getParent(); 688 689 return nullptr; 690 } 691 692 static bool notDifferentParent(const Value *O1, const Value *O2) { 693 694 const Function *F1 = getParent(O1); 695 const Function *F2 = getParent(O2); 696 697 return !F1 || !F2 || F1 == F2; 698 } 699 #endif 700 701 AliasResult BasicAAResult::alias(const MemoryLocation &LocA, 702 const MemoryLocation &LocB) { 703 assert(notDifferentParent(LocA.Ptr, LocB.Ptr) && 704 "BasicAliasAnalysis doesn't support interprocedural queries."); 705 706 // If we have a directly cached entry for these locations, we have recursed 707 // through this once, so just return the cached results. Notably, when this 708 // happens, we don't clear the cache. 709 auto CacheIt = AliasCache.find(LocPair(LocA, LocB)); 710 if (CacheIt != AliasCache.end()) 711 return CacheIt->second; 712 713 AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.AATags, LocB.Ptr, 714 LocB.Size, LocB.AATags); 715 // AliasCache rarely has more than 1 or 2 elements, always use 716 // shrink_and_clear so it quickly returns to the inline capacity of the 717 // SmallDenseMap if it ever grows larger. 718 // FIXME: This should really be shrink_to_inline_capacity_and_clear(). 719 AliasCache.shrink_and_clear(); 720 VisitedPhiBBs.clear(); 721 return Alias; 722 } 723 724 /// Checks to see if the specified callsite can clobber the specified memory 725 /// object. 726 /// 727 /// Since we only look at local properties of this function, we really can't 728 /// say much about this query. We do, however, use simple "address taken" 729 /// analysis on local objects. 730 ModRefInfo BasicAAResult::getModRefInfo(ImmutableCallSite CS, 731 const MemoryLocation &Loc) { 732 assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) && 733 "AliasAnalysis query involving multiple functions!"); 734 735 const Value *Object = GetUnderlyingObject(Loc.Ptr, DL); 736 737 // If this is a tail call and Loc.Ptr points to a stack location, we know that 738 // the tail call cannot access or modify the local stack. 739 // We cannot exclude byval arguments here; these belong to the caller of 740 // the current function not to the current function, and a tail callee 741 // may reference them. 742 if (isa<AllocaInst>(Object)) 743 if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) 744 if (CI->isTailCall()) 745 return MRI_NoModRef; 746 747 // If the pointer is to a locally allocated object that does not escape, 748 // then the call can not mod/ref the pointer unless the call takes the pointer 749 // as an argument, and itself doesn't capture it. 750 if (!isa<Constant>(Object) && CS.getInstruction() != Object && 751 isNonEscapingLocalObject(Object)) { 752 bool PassedAsArg = false; 753 unsigned OperandNo = 0; 754 for (auto CI = CS.data_operands_begin(), CE = CS.data_operands_end(); 755 CI != CE; ++CI, ++OperandNo) { 756 // Only look at the no-capture or byval pointer arguments. If this 757 // pointer were passed to arguments that were neither of these, then it 758 // couldn't be no-capture. 759 if (!(*CI)->getType()->isPointerTy() || 760 (!CS.doesNotCapture(OperandNo) && 761 OperandNo < CS.getNumArgOperands() && !CS.isByValArgument(OperandNo))) 762 continue; 763 764 // If this is a no-capture pointer argument, see if we can tell that it 765 // is impossible to alias the pointer we're checking. If not, we have to 766 // assume that the call could touch the pointer, even though it doesn't 767 // escape. 768 AliasResult AR = 769 getBestAAResults().alias(MemoryLocation(*CI), MemoryLocation(Object)); 770 if (AR) { 771 PassedAsArg = true; 772 break; 773 } 774 } 775 776 if (!PassedAsArg) 777 return MRI_NoModRef; 778 } 779 780 // If the CallSite is to malloc or calloc, we can assume that it doesn't 781 // modify any IR visible value. This is only valid because we assume these 782 // routines do not read values visible in the IR. TODO: Consider special 783 // casing realloc and strdup routines which access only their arguments as 784 // well. Or alternatively, replace all of this with inaccessiblememonly once 785 // that's implemented fully. 786 auto *Inst = CS.getInstruction(); 787 if (isMallocLikeFn(Inst, &TLI) || isCallocLikeFn(Inst, &TLI)) { 788 // Be conservative if the accessed pointer may alias the allocation - 789 // fallback to the generic handling below. 790 if (getBestAAResults().alias(MemoryLocation(Inst), Loc) == NoAlias) 791 return MRI_NoModRef; 792 } 793 794 // The semantics of memcpy intrinsics forbid overlap between their respective 795 // operands, i.e., source and destination of any given memcpy must no-alias. 796 // If Loc must-aliases either one of these two locations, then it necessarily 797 // no-aliases the other. 798 if (auto *Inst = dyn_cast<MemCpyInst>(CS.getInstruction())) { 799 AliasResult SrcAA, DestAA; 800 801 if ((SrcAA = getBestAAResults().alias(MemoryLocation::getForSource(Inst), 802 Loc)) == MustAlias) 803 // Loc is exactly the memcpy source thus disjoint from memcpy dest. 804 return MRI_Ref; 805 if ((DestAA = getBestAAResults().alias(MemoryLocation::getForDest(Inst), 806 Loc)) == MustAlias) 807 // The converse case. 808 return MRI_Mod; 809 810 // It's also possible for Loc to alias both src and dest, or neither. 811 ModRefInfo rv = MRI_NoModRef; 812 if (SrcAA != NoAlias) 813 rv = static_cast<ModRefInfo>(rv | MRI_Ref); 814 if (DestAA != NoAlias) 815 rv = static_cast<ModRefInfo>(rv | MRI_Mod); 816 return rv; 817 } 818 819 // While the assume intrinsic is marked as arbitrarily writing so that 820 // proper control dependencies will be maintained, it never aliases any 821 // particular memory location. 822 if (isIntrinsicCall(CS, Intrinsic::assume)) 823 return MRI_NoModRef; 824 825 // Like assumes, guard intrinsics are also marked as arbitrarily writing so 826 // that proper control dependencies are maintained but they never mods any 827 // particular memory location. 828 // 829 // *Unlike* assumes, guard intrinsics are modeled as reading memory since the 830 // heap state at the point the guard is issued needs to be consistent in case 831 // the guard invokes the "deopt" continuation. 832 if (isIntrinsicCall(CS, Intrinsic::experimental_guard)) 833 return MRI_Ref; 834 835 // Like assumes, invariant.start intrinsics were also marked as arbitrarily 836 // writing so that proper control dependencies are maintained but they never 837 // mod any particular memory location visible to the IR. 838 // *Unlike* assumes (which are now modeled as NoModRef), invariant.start 839 // intrinsic is now modeled as reading memory. This prevents hoisting the 840 // invariant.start intrinsic over stores. Consider: 841 // *ptr = 40; 842 // *ptr = 50; 843 // invariant_start(ptr) 844 // int val = *ptr; 845 // print(val); 846 // 847 // This cannot be transformed to: 848 // 849 // *ptr = 40; 850 // invariant_start(ptr) 851 // *ptr = 50; 852 // int val = *ptr; 853 // print(val); 854 // 855 // The transformation will cause the second store to be ignored (based on 856 // rules of invariant.start) and print 40, while the first program always 857 // prints 50. 858 if (isIntrinsicCall(CS, Intrinsic::invariant_start)) 859 return MRI_Ref; 860 861 // The AAResultBase base class has some smarts, lets use them. 862 return AAResultBase::getModRefInfo(CS, Loc); 863 } 864 865 ModRefInfo BasicAAResult::getModRefInfo(ImmutableCallSite CS1, 866 ImmutableCallSite CS2) { 867 // While the assume intrinsic is marked as arbitrarily writing so that 868 // proper control dependencies will be maintained, it never aliases any 869 // particular memory location. 870 if (isIntrinsicCall(CS1, Intrinsic::assume) || 871 isIntrinsicCall(CS2, Intrinsic::assume)) 872 return MRI_NoModRef; 873 874 // Like assumes, guard intrinsics are also marked as arbitrarily writing so 875 // that proper control dependencies are maintained but they never mod any 876 // particular memory location. 877 // 878 // *Unlike* assumes, guard intrinsics are modeled as reading memory since the 879 // heap state at the point the guard is issued needs to be consistent in case 880 // the guard invokes the "deopt" continuation. 881 882 // NB! This function is *not* commutative, so we specical case two 883 // possibilities for guard intrinsics. 884 885 if (isIntrinsicCall(CS1, Intrinsic::experimental_guard)) 886 return getModRefBehavior(CS2) & MRI_Mod ? MRI_Ref : MRI_NoModRef; 887 888 if (isIntrinsicCall(CS2, Intrinsic::experimental_guard)) 889 return getModRefBehavior(CS1) & MRI_Mod ? MRI_Mod : MRI_NoModRef; 890 891 // The AAResultBase base class has some smarts, lets use them. 892 return AAResultBase::getModRefInfo(CS1, CS2); 893 } 894 895 /// Provide ad-hoc rules to disambiguate accesses through two GEP operators, 896 /// both having the exact same pointer operand. 897 static AliasResult aliasSameBasePointerGEPs(const GEPOperator *GEP1, 898 uint64_t V1Size, 899 const GEPOperator *GEP2, 900 uint64_t V2Size, 901 const DataLayout &DL) { 902 903 assert(GEP1->getPointerOperand()->stripPointerCasts() == 904 GEP2->getPointerOperand()->stripPointerCasts() && 905 GEP1->getPointerOperand()->getType() == 906 GEP2->getPointerOperand()->getType() && 907 "Expected GEPs with the same pointer operand"); 908 909 // Try to determine whether GEP1 and GEP2 index through arrays, into structs, 910 // such that the struct field accesses provably cannot alias. 911 // We also need at least two indices (the pointer, and the struct field). 912 if (GEP1->getNumIndices() != GEP2->getNumIndices() || 913 GEP1->getNumIndices() < 2) 914 return MayAlias; 915 916 // If we don't know the size of the accesses through both GEPs, we can't 917 // determine whether the struct fields accessed can't alias. 918 if (V1Size == MemoryLocation::UnknownSize || 919 V2Size == MemoryLocation::UnknownSize) 920 return MayAlias; 921 922 ConstantInt *C1 = 923 dyn_cast<ConstantInt>(GEP1->getOperand(GEP1->getNumOperands() - 1)); 924 ConstantInt *C2 = 925 dyn_cast<ConstantInt>(GEP2->getOperand(GEP2->getNumOperands() - 1)); 926 927 // If the last (struct) indices are constants and are equal, the other indices 928 // might be also be dynamically equal, so the GEPs can alias. 929 if (C1 && C2 && C1->getSExtValue() == C2->getSExtValue()) 930 return MayAlias; 931 932 // Find the last-indexed type of the GEP, i.e., the type you'd get if 933 // you stripped the last index. 934 // On the way, look at each indexed type. If there's something other 935 // than an array, different indices can lead to different final types. 936 SmallVector<Value *, 8> IntermediateIndices; 937 938 // Insert the first index; we don't need to check the type indexed 939 // through it as it only drops the pointer indirection. 940 assert(GEP1->getNumIndices() > 1 && "Not enough GEP indices to examine"); 941 IntermediateIndices.push_back(GEP1->getOperand(1)); 942 943 // Insert all the remaining indices but the last one. 944 // Also, check that they all index through arrays. 945 for (unsigned i = 1, e = GEP1->getNumIndices() - 1; i != e; ++i) { 946 if (!isa<ArrayType>(GetElementPtrInst::getIndexedType( 947 GEP1->getSourceElementType(), IntermediateIndices))) 948 return MayAlias; 949 IntermediateIndices.push_back(GEP1->getOperand(i + 1)); 950 } 951 952 auto *Ty = GetElementPtrInst::getIndexedType( 953 GEP1->getSourceElementType(), IntermediateIndices); 954 StructType *LastIndexedStruct = dyn_cast<StructType>(Ty); 955 956 if (isa<SequentialType>(Ty)) { 957 // We know that: 958 // - both GEPs begin indexing from the exact same pointer; 959 // - the last indices in both GEPs are constants, indexing into a sequential 960 // type (array or pointer); 961 // - both GEPs only index through arrays prior to that. 962 // 963 // Because array indices greater than the number of elements are valid in 964 // GEPs, unless we know the intermediate indices are identical between 965 // GEP1 and GEP2 we cannot guarantee that the last indexed arrays don't 966 // partially overlap. We also need to check that the loaded size matches 967 // the element size, otherwise we could still have overlap. 968 const uint64_t ElementSize = 969 DL.getTypeStoreSize(cast<SequentialType>(Ty)->getElementType()); 970 if (V1Size != ElementSize || V2Size != ElementSize) 971 return MayAlias; 972 973 for (unsigned i = 0, e = GEP1->getNumIndices() - 1; i != e; ++i) 974 if (GEP1->getOperand(i + 1) != GEP2->getOperand(i + 1)) 975 return MayAlias; 976 977 // Now we know that the array/pointer that GEP1 indexes into and that 978 // that GEP2 indexes into must either precisely overlap or be disjoint. 979 // Because they cannot partially overlap and because fields in an array 980 // cannot overlap, if we can prove the final indices are different between 981 // GEP1 and GEP2, we can conclude GEP1 and GEP2 don't alias. 982 983 // If the last indices are constants, we've already checked they don't 984 // equal each other so we can exit early. 985 if (C1 && C2) 986 return NoAlias; 987 if (isKnownNonEqual(GEP1->getOperand(GEP1->getNumOperands() - 1), 988 GEP2->getOperand(GEP2->getNumOperands() - 1), 989 DL)) 990 return NoAlias; 991 return MayAlias; 992 } else if (!LastIndexedStruct || !C1 || !C2) { 993 return MayAlias; 994 } 995 996 // We know that: 997 // - both GEPs begin indexing from the exact same pointer; 998 // - the last indices in both GEPs are constants, indexing into a struct; 999 // - said indices are different, hence, the pointed-to fields are different; 1000 // - both GEPs only index through arrays prior to that. 1001 // 1002 // This lets us determine that the struct that GEP1 indexes into and the 1003 // struct that GEP2 indexes into must either precisely overlap or be 1004 // completely disjoint. Because they cannot partially overlap, indexing into 1005 // different non-overlapping fields of the struct will never alias. 1006 1007 // Therefore, the only remaining thing needed to show that both GEPs can't 1008 // alias is that the fields are not overlapping. 1009 const StructLayout *SL = DL.getStructLayout(LastIndexedStruct); 1010 const uint64_t StructSize = SL->getSizeInBytes(); 1011 const uint64_t V1Off = SL->getElementOffset(C1->getZExtValue()); 1012 const uint64_t V2Off = SL->getElementOffset(C2->getZExtValue()); 1013 1014 auto EltsDontOverlap = [StructSize](uint64_t V1Off, uint64_t V1Size, 1015 uint64_t V2Off, uint64_t V2Size) { 1016 return V1Off < V2Off && V1Off + V1Size <= V2Off && 1017 ((V2Off + V2Size <= StructSize) || 1018 (V2Off + V2Size - StructSize <= V1Off)); 1019 }; 1020 1021 if (EltsDontOverlap(V1Off, V1Size, V2Off, V2Size) || 1022 EltsDontOverlap(V2Off, V2Size, V1Off, V1Size)) 1023 return NoAlias; 1024 1025 return MayAlias; 1026 } 1027 1028 // If a we have (a) a GEP and (b) a pointer based on an alloca, and the 1029 // beginning of the object the GEP points would have a negative offset with 1030 // repsect to the alloca, that means the GEP can not alias pointer (b). 1031 // Note that the pointer based on the alloca may not be a GEP. For 1032 // example, it may be the alloca itself. 1033 // The same applies if (b) is based on a GlobalVariable. Note that just being 1034 // based on isIdentifiedObject() is not enough - we need an identified object 1035 // that does not permit access to negative offsets. For example, a negative 1036 // offset from a noalias argument or call can be inbounds w.r.t the actual 1037 // underlying object. 1038 // 1039 // For example, consider: 1040 // 1041 // struct { int f0, int f1, ...} foo; 1042 // foo alloca; 1043 // foo* random = bar(alloca); 1044 // int *f0 = &alloca.f0 1045 // int *f1 = &random->f1; 1046 // 1047 // Which is lowered, approximately, to: 1048 // 1049 // %alloca = alloca %struct.foo 1050 // %random = call %struct.foo* @random(%struct.foo* %alloca) 1051 // %f0 = getelementptr inbounds %struct, %struct.foo* %alloca, i32 0, i32 0 1052 // %f1 = getelementptr inbounds %struct, %struct.foo* %random, i32 0, i32 1 1053 // 1054 // Assume %f1 and %f0 alias. Then %f1 would point into the object allocated 1055 // by %alloca. Since the %f1 GEP is inbounds, that means %random must also 1056 // point into the same object. But since %f0 points to the beginning of %alloca, 1057 // the highest %f1 can be is (%alloca + 3). This means %random can not be higher 1058 // than (%alloca - 1), and so is not inbounds, a contradiction. 1059 bool BasicAAResult::isGEPBaseAtNegativeOffset(const GEPOperator *GEPOp, 1060 const DecomposedGEP &DecompGEP, const DecomposedGEP &DecompObject, 1061 uint64_t ObjectAccessSize) { 1062 // If the object access size is unknown, or the GEP isn't inbounds, bail. 1063 if (ObjectAccessSize == MemoryLocation::UnknownSize || !GEPOp->isInBounds()) 1064 return false; 1065 1066 // We need the object to be an alloca or a globalvariable, and want to know 1067 // the offset of the pointer from the object precisely, so no variable 1068 // indices are allowed. 1069 if (!(isa<AllocaInst>(DecompObject.Base) || 1070 isa<GlobalVariable>(DecompObject.Base)) || 1071 !DecompObject.VarIndices.empty()) 1072 return false; 1073 1074 int64_t ObjectBaseOffset = DecompObject.StructOffset + 1075 DecompObject.OtherOffset; 1076 1077 // If the GEP has no variable indices, we know the precise offset 1078 // from the base, then use it. If the GEP has variable indices, we're in 1079 // a bit more trouble: we can't count on the constant offsets that come 1080 // from non-struct sources, since these can be "rewound" by a negative 1081 // variable offset. So use only offsets that came from structs. 1082 int64_t GEPBaseOffset = DecompGEP.StructOffset; 1083 if (DecompGEP.VarIndices.empty()) 1084 GEPBaseOffset += DecompGEP.OtherOffset; 1085 1086 return (GEPBaseOffset >= ObjectBaseOffset + (int64_t)ObjectAccessSize); 1087 } 1088 1089 /// Provides a bunch of ad-hoc rules to disambiguate a GEP instruction against 1090 /// another pointer. 1091 /// 1092 /// We know that V1 is a GEP, but we don't know anything about V2. 1093 /// UnderlyingV1 is GetUnderlyingObject(GEP1, DL), UnderlyingV2 is the same for 1094 /// V2. 1095 AliasResult BasicAAResult::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size, 1096 const AAMDNodes &V1AAInfo, const Value *V2, 1097 uint64_t V2Size, const AAMDNodes &V2AAInfo, 1098 const Value *UnderlyingV1, 1099 const Value *UnderlyingV2) { 1100 DecomposedGEP DecompGEP1, DecompGEP2; 1101 bool GEP1MaxLookupReached = 1102 DecomposeGEPExpression(GEP1, DecompGEP1, DL, &AC, DT); 1103 bool GEP2MaxLookupReached = 1104 DecomposeGEPExpression(V2, DecompGEP2, DL, &AC, DT); 1105 1106 int64_t GEP1BaseOffset = DecompGEP1.StructOffset + DecompGEP1.OtherOffset; 1107 int64_t GEP2BaseOffset = DecompGEP2.StructOffset + DecompGEP2.OtherOffset; 1108 1109 assert(DecompGEP1.Base == UnderlyingV1 && DecompGEP2.Base == UnderlyingV2 && 1110 "DecomposeGEPExpression returned a result different from " 1111 "GetUnderlyingObject"); 1112 1113 // If the GEP's offset relative to its base is such that the base would 1114 // fall below the start of the object underlying V2, then the GEP and V2 1115 // cannot alias. 1116 if (!GEP1MaxLookupReached && !GEP2MaxLookupReached && 1117 isGEPBaseAtNegativeOffset(GEP1, DecompGEP1, DecompGEP2, V2Size)) 1118 return NoAlias; 1119 // If we have two gep instructions with must-alias or not-alias'ing base 1120 // pointers, figure out if the indexes to the GEP tell us anything about the 1121 // derived pointer. 1122 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) { 1123 // Check for the GEP base being at a negative offset, this time in the other 1124 // direction. 1125 if (!GEP1MaxLookupReached && !GEP2MaxLookupReached && 1126 isGEPBaseAtNegativeOffset(GEP2, DecompGEP2, DecompGEP1, V1Size)) 1127 return NoAlias; 1128 // Do the base pointers alias? 1129 AliasResult BaseAlias = 1130 aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize, AAMDNodes(), 1131 UnderlyingV2, MemoryLocation::UnknownSize, AAMDNodes()); 1132 1133 // Check for geps of non-aliasing underlying pointers where the offsets are 1134 // identical. 1135 if ((BaseAlias == MayAlias) && V1Size == V2Size) { 1136 // Do the base pointers alias assuming type and size. 1137 AliasResult PreciseBaseAlias = aliasCheck(UnderlyingV1, V1Size, V1AAInfo, 1138 UnderlyingV2, V2Size, V2AAInfo); 1139 if (PreciseBaseAlias == NoAlias) { 1140 // See if the computed offset from the common pointer tells us about the 1141 // relation of the resulting pointer. 1142 // If the max search depth is reached the result is undefined 1143 if (GEP2MaxLookupReached || GEP1MaxLookupReached) 1144 return MayAlias; 1145 1146 // Same offsets. 1147 if (GEP1BaseOffset == GEP2BaseOffset && 1148 DecompGEP1.VarIndices == DecompGEP2.VarIndices) 1149 return NoAlias; 1150 } 1151 } 1152 1153 // If we get a No or May, then return it immediately, no amount of analysis 1154 // will improve this situation. 1155 if (BaseAlias != MustAlias) 1156 return BaseAlias; 1157 1158 // Otherwise, we have a MustAlias. Since the base pointers alias each other 1159 // exactly, see if the computed offset from the common pointer tells us 1160 // about the relation of the resulting pointer. 1161 // If we know the two GEPs are based off of the exact same pointer (and not 1162 // just the same underlying object), see if that tells us anything about 1163 // the resulting pointers. 1164 if (GEP1->getPointerOperand()->stripPointerCasts() == 1165 GEP2->getPointerOperand()->stripPointerCasts() && 1166 GEP1->getPointerOperand()->getType() == 1167 GEP2->getPointerOperand()->getType()) { 1168 AliasResult R = aliasSameBasePointerGEPs(GEP1, V1Size, GEP2, V2Size, DL); 1169 // If we couldn't find anything interesting, don't abandon just yet. 1170 if (R != MayAlias) 1171 return R; 1172 } 1173 1174 // If the max search depth is reached, the result is undefined 1175 if (GEP2MaxLookupReached || GEP1MaxLookupReached) 1176 return MayAlias; 1177 1178 // Subtract the GEP2 pointer from the GEP1 pointer to find out their 1179 // symbolic difference. 1180 GEP1BaseOffset -= GEP2BaseOffset; 1181 GetIndexDifference(DecompGEP1.VarIndices, DecompGEP2.VarIndices); 1182 1183 } else { 1184 // Check to see if these two pointers are related by the getelementptr 1185 // instruction. If one pointer is a GEP with a non-zero index of the other 1186 // pointer, we know they cannot alias. 1187 1188 // If both accesses are unknown size, we can't do anything useful here. 1189 if (V1Size == MemoryLocation::UnknownSize && 1190 V2Size == MemoryLocation::UnknownSize) 1191 return MayAlias; 1192 1193 AliasResult R = aliasCheck(UnderlyingV1, MemoryLocation::UnknownSize, 1194 AAMDNodes(), V2, V2Size, V2AAInfo, 1195 nullptr, UnderlyingV2); 1196 if (R != MustAlias) 1197 // If V2 may alias GEP base pointer, conservatively returns MayAlias. 1198 // If V2 is known not to alias GEP base pointer, then the two values 1199 // cannot alias per GEP semantics: "A pointer value formed from a 1200 // getelementptr instruction is associated with the addresses associated 1201 // with the first operand of the getelementptr". 1202 return R; 1203 1204 // If the max search depth is reached the result is undefined 1205 if (GEP1MaxLookupReached) 1206 return MayAlias; 1207 } 1208 1209 // In the two GEP Case, if there is no difference in the offsets of the 1210 // computed pointers, the resultant pointers are a must alias. This 1211 // happens when we have two lexically identical GEP's (for example). 1212 // 1213 // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 1214 // must aliases the GEP, the end result is a must alias also. 1215 if (GEP1BaseOffset == 0 && DecompGEP1.VarIndices.empty()) 1216 return MustAlias; 1217 1218 // If there is a constant difference between the pointers, but the difference 1219 // is less than the size of the associated memory object, then we know 1220 // that the objects are partially overlapping. If the difference is 1221 // greater, we know they do not overlap. 1222 if (GEP1BaseOffset != 0 && DecompGEP1.VarIndices.empty()) { 1223 if (GEP1BaseOffset >= 0) { 1224 if (V2Size != MemoryLocation::UnknownSize) { 1225 if ((uint64_t)GEP1BaseOffset < V2Size) 1226 return PartialAlias; 1227 return NoAlias; 1228 } 1229 } else { 1230 // We have the situation where: 1231 // + + 1232 // | BaseOffset | 1233 // ---------------->| 1234 // |-->V1Size |-------> V2Size 1235 // GEP1 V2 1236 // We need to know that V2Size is not unknown, otherwise we might have 1237 // stripped a gep with negative index ('gep <ptr>, -1, ...). 1238 if (V1Size != MemoryLocation::UnknownSize && 1239 V2Size != MemoryLocation::UnknownSize) { 1240 if (-(uint64_t)GEP1BaseOffset < V1Size) 1241 return PartialAlias; 1242 return NoAlias; 1243 } 1244 } 1245 } 1246 1247 if (!DecompGEP1.VarIndices.empty()) { 1248 uint64_t Modulo = 0; 1249 bool AllPositive = true; 1250 for (unsigned i = 0, e = DecompGEP1.VarIndices.size(); i != e; ++i) { 1251 1252 // Try to distinguish something like &A[i][1] against &A[42][0]. 1253 // Grab the least significant bit set in any of the scales. We 1254 // don't need std::abs here (even if the scale's negative) as we'll 1255 // be ^'ing Modulo with itself later. 1256 Modulo |= (uint64_t)DecompGEP1.VarIndices[i].Scale; 1257 1258 if (AllPositive) { 1259 // If the Value could change between cycles, then any reasoning about 1260 // the Value this cycle may not hold in the next cycle. We'll just 1261 // give up if we can't determine conditions that hold for every cycle: 1262 const Value *V = DecompGEP1.VarIndices[i].V; 1263 1264 bool SignKnownZero, SignKnownOne; 1265 ComputeSignBit(const_cast<Value *>(V), SignKnownZero, SignKnownOne, DL, 1266 0, &AC, nullptr, DT); 1267 1268 // Zero-extension widens the variable, and so forces the sign 1269 // bit to zero. 1270 bool IsZExt = DecompGEP1.VarIndices[i].ZExtBits > 0 || isa<ZExtInst>(V); 1271 SignKnownZero |= IsZExt; 1272 SignKnownOne &= !IsZExt; 1273 1274 // If the variable begins with a zero then we know it's 1275 // positive, regardless of whether the value is signed or 1276 // unsigned. 1277 int64_t Scale = DecompGEP1.VarIndices[i].Scale; 1278 AllPositive = 1279 (SignKnownZero && Scale >= 0) || (SignKnownOne && Scale < 0); 1280 } 1281 } 1282 1283 Modulo = Modulo ^ (Modulo & (Modulo - 1)); 1284 1285 // We can compute the difference between the two addresses 1286 // mod Modulo. Check whether that difference guarantees that the 1287 // two locations do not alias. 1288 uint64_t ModOffset = (uint64_t)GEP1BaseOffset & (Modulo - 1); 1289 if (V1Size != MemoryLocation::UnknownSize && 1290 V2Size != MemoryLocation::UnknownSize && ModOffset >= V2Size && 1291 V1Size <= Modulo - ModOffset) 1292 return NoAlias; 1293 1294 // If we know all the variables are positive, then GEP1 >= GEP1BasePtr. 1295 // If GEP1BasePtr > V2 (GEP1BaseOffset > 0) then we know the pointers 1296 // don't alias if V2Size can fit in the gap between V2 and GEP1BasePtr. 1297 if (AllPositive && GEP1BaseOffset > 0 && V2Size <= (uint64_t)GEP1BaseOffset) 1298 return NoAlias; 1299 1300 if (constantOffsetHeuristic(DecompGEP1.VarIndices, V1Size, V2Size, 1301 GEP1BaseOffset, &AC, DT)) 1302 return NoAlias; 1303 } 1304 1305 // Statically, we can see that the base objects are the same, but the 1306 // pointers have dynamic offsets which we can't resolve. And none of our 1307 // little tricks above worked. 1308 // 1309 // TODO: Returning PartialAlias instead of MayAlias is a mild hack; the 1310 // practical effect of this is protecting TBAA in the case of dynamic 1311 // indices into arrays of unions or malloc'd memory. 1312 return PartialAlias; 1313 } 1314 1315 static AliasResult MergeAliasResults(AliasResult A, AliasResult B) { 1316 // If the results agree, take it. 1317 if (A == B) 1318 return A; 1319 // A mix of PartialAlias and MustAlias is PartialAlias. 1320 if ((A == PartialAlias && B == MustAlias) || 1321 (B == PartialAlias && A == MustAlias)) 1322 return PartialAlias; 1323 // Otherwise, we don't know anything. 1324 return MayAlias; 1325 } 1326 1327 /// Provides a bunch of ad-hoc rules to disambiguate a Select instruction 1328 /// against another. 1329 AliasResult BasicAAResult::aliasSelect(const SelectInst *SI, uint64_t SISize, 1330 const AAMDNodes &SIAAInfo, 1331 const Value *V2, uint64_t V2Size, 1332 const AAMDNodes &V2AAInfo, 1333 const Value *UnderV2) { 1334 // If the values are Selects with the same condition, we can do a more precise 1335 // check: just check for aliases between the values on corresponding arms. 1336 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2)) 1337 if (SI->getCondition() == SI2->getCondition()) { 1338 AliasResult Alias = aliasCheck(SI->getTrueValue(), SISize, SIAAInfo, 1339 SI2->getTrueValue(), V2Size, V2AAInfo); 1340 if (Alias == MayAlias) 1341 return MayAlias; 1342 AliasResult ThisAlias = 1343 aliasCheck(SI->getFalseValue(), SISize, SIAAInfo, 1344 SI2->getFalseValue(), V2Size, V2AAInfo); 1345 return MergeAliasResults(ThisAlias, Alias); 1346 } 1347 1348 // If both arms of the Select node NoAlias or MustAlias V2, then returns 1349 // NoAlias / MustAlias. Otherwise, returns MayAlias. 1350 AliasResult Alias = 1351 aliasCheck(V2, V2Size, V2AAInfo, SI->getTrueValue(), 1352 SISize, SIAAInfo, UnderV2); 1353 if (Alias == MayAlias) 1354 return MayAlias; 1355 1356 AliasResult ThisAlias = 1357 aliasCheck(V2, V2Size, V2AAInfo, SI->getFalseValue(), SISize, SIAAInfo, 1358 UnderV2); 1359 return MergeAliasResults(ThisAlias, Alias); 1360 } 1361 1362 /// Provide a bunch of ad-hoc rules to disambiguate a PHI instruction against 1363 /// another. 1364 AliasResult BasicAAResult::aliasPHI(const PHINode *PN, uint64_t PNSize, 1365 const AAMDNodes &PNAAInfo, const Value *V2, 1366 uint64_t V2Size, const AAMDNodes &V2AAInfo, 1367 const Value *UnderV2) { 1368 // Track phi nodes we have visited. We use this information when we determine 1369 // value equivalence. 1370 VisitedPhiBBs.insert(PN->getParent()); 1371 1372 // If the values are PHIs in the same block, we can do a more precise 1373 // as well as efficient check: just check for aliases between the values 1374 // on corresponding edges. 1375 if (const PHINode *PN2 = dyn_cast<PHINode>(V2)) 1376 if (PN2->getParent() == PN->getParent()) { 1377 LocPair Locs(MemoryLocation(PN, PNSize, PNAAInfo), 1378 MemoryLocation(V2, V2Size, V2AAInfo)); 1379 if (PN > V2) 1380 std::swap(Locs.first, Locs.second); 1381 // Analyse the PHIs' inputs under the assumption that the PHIs are 1382 // NoAlias. 1383 // If the PHIs are May/MustAlias there must be (recursively) an input 1384 // operand from outside the PHIs' cycle that is MayAlias/MustAlias or 1385 // there must be an operation on the PHIs within the PHIs' value cycle 1386 // that causes a MayAlias. 1387 // Pretend the phis do not alias. 1388 AliasResult Alias = NoAlias; 1389 assert(AliasCache.count(Locs) && 1390 "There must exist an entry for the phi node"); 1391 AliasResult OrigAliasResult = AliasCache[Locs]; 1392 AliasCache[Locs] = NoAlias; 1393 1394 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1395 AliasResult ThisAlias = 1396 aliasCheck(PN->getIncomingValue(i), PNSize, PNAAInfo, 1397 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)), 1398 V2Size, V2AAInfo); 1399 Alias = MergeAliasResults(ThisAlias, Alias); 1400 if (Alias == MayAlias) 1401 break; 1402 } 1403 1404 // Reset if speculation failed. 1405 if (Alias != NoAlias) 1406 AliasCache[Locs] = OrigAliasResult; 1407 1408 return Alias; 1409 } 1410 1411 SmallPtrSet<Value *, 4> UniqueSrc; 1412 SmallVector<Value *, 4> V1Srcs; 1413 bool isRecursive = false; 1414 for (Value *PV1 : PN->incoming_values()) { 1415 if (isa<PHINode>(PV1)) 1416 // If any of the source itself is a PHI, return MayAlias conservatively 1417 // to avoid compile time explosion. The worst possible case is if both 1418 // sides are PHI nodes. In which case, this is O(m x n) time where 'm' 1419 // and 'n' are the number of PHI sources. 1420 return MayAlias; 1421 1422 if (EnableRecPhiAnalysis) 1423 if (GEPOperator *PV1GEP = dyn_cast<GEPOperator>(PV1)) { 1424 // Check whether the incoming value is a GEP that advances the pointer 1425 // result of this PHI node (e.g. in a loop). If this is the case, we 1426 // would recurse and always get a MayAlias. Handle this case specially 1427 // below. 1428 if (PV1GEP->getPointerOperand() == PN && PV1GEP->getNumIndices() == 1 && 1429 isa<ConstantInt>(PV1GEP->idx_begin())) { 1430 isRecursive = true; 1431 continue; 1432 } 1433 } 1434 1435 if (UniqueSrc.insert(PV1).second) 1436 V1Srcs.push_back(PV1); 1437 } 1438 1439 // If this PHI node is recursive, set the size of the accessed memory to 1440 // unknown to represent all the possible values the GEP could advance the 1441 // pointer to. 1442 if (isRecursive) 1443 PNSize = MemoryLocation::UnknownSize; 1444 1445 AliasResult Alias = 1446 aliasCheck(V2, V2Size, V2AAInfo, V1Srcs[0], 1447 PNSize, PNAAInfo, UnderV2); 1448 1449 // Early exit if the check of the first PHI source against V2 is MayAlias. 1450 // Other results are not possible. 1451 if (Alias == MayAlias) 1452 return MayAlias; 1453 1454 // If all sources of the PHI node NoAlias or MustAlias V2, then returns 1455 // NoAlias / MustAlias. Otherwise, returns MayAlias. 1456 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) { 1457 Value *V = V1Srcs[i]; 1458 1459 AliasResult ThisAlias = 1460 aliasCheck(V2, V2Size, V2AAInfo, V, PNSize, PNAAInfo, UnderV2); 1461 Alias = MergeAliasResults(ThisAlias, Alias); 1462 if (Alias == MayAlias) 1463 break; 1464 } 1465 1466 return Alias; 1467 } 1468 1469 /// Provides a bunch of ad-hoc rules to disambiguate in common cases, such as 1470 /// array references. 1471 AliasResult BasicAAResult::aliasCheck(const Value *V1, uint64_t V1Size, 1472 AAMDNodes V1AAInfo, const Value *V2, 1473 uint64_t V2Size, AAMDNodes V2AAInfo, 1474 const Value *O1, const Value *O2) { 1475 // If either of the memory references is empty, it doesn't matter what the 1476 // pointer values are. 1477 if (V1Size == 0 || V2Size == 0) 1478 return NoAlias; 1479 1480 // Strip off any casts if they exist. 1481 V1 = V1->stripPointerCasts(); 1482 V2 = V2->stripPointerCasts(); 1483 1484 // If V1 or V2 is undef, the result is NoAlias because we can always pick a 1485 // value for undef that aliases nothing in the program. 1486 if (isa<UndefValue>(V1) || isa<UndefValue>(V2)) 1487 return NoAlias; 1488 1489 // Are we checking for alias of the same value? 1490 // Because we look 'through' phi nodes, we could look at "Value" pointers from 1491 // different iterations. We must therefore make sure that this is not the 1492 // case. The function isValueEqualInPotentialCycles ensures that this cannot 1493 // happen by looking at the visited phi nodes and making sure they cannot 1494 // reach the value. 1495 if (isValueEqualInPotentialCycles(V1, V2)) 1496 return MustAlias; 1497 1498 if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy()) 1499 return NoAlias; // Scalars cannot alias each other 1500 1501 // Figure out what objects these things are pointing to if we can. 1502 if (O1 == nullptr) 1503 O1 = GetUnderlyingObject(V1, DL, MaxLookupSearchDepth); 1504 1505 if (O2 == nullptr) 1506 O2 = GetUnderlyingObject(V2, DL, MaxLookupSearchDepth); 1507 1508 // Null values in the default address space don't point to any object, so they 1509 // don't alias any other pointer. 1510 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1)) 1511 if (CPN->getType()->getAddressSpace() == 0) 1512 return NoAlias; 1513 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2)) 1514 if (CPN->getType()->getAddressSpace() == 0) 1515 return NoAlias; 1516 1517 if (O1 != O2) { 1518 // If V1/V2 point to two different objects, we know that we have no alias. 1519 if (isIdentifiedObject(O1) && isIdentifiedObject(O2)) 1520 return NoAlias; 1521 1522 // Constant pointers can't alias with non-const isIdentifiedObject objects. 1523 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) || 1524 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1))) 1525 return NoAlias; 1526 1527 // Function arguments can't alias with things that are known to be 1528 // unambigously identified at the function level. 1529 if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) || 1530 (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1))) 1531 return NoAlias; 1532 1533 // Most objects can't alias null. 1534 if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) || 1535 (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2))) 1536 return NoAlias; 1537 1538 // If one pointer is the result of a call/invoke or load and the other is a 1539 // non-escaping local object within the same function, then we know the 1540 // object couldn't escape to a point where the call could return it. 1541 // 1542 // Note that if the pointers are in different functions, there are a 1543 // variety of complications. A call with a nocapture argument may still 1544 // temporary store the nocapture argument's value in a temporary memory 1545 // location if that memory location doesn't escape. Or it may pass a 1546 // nocapture value to other functions as long as they don't capture it. 1547 if (isEscapeSource(O1) && isNonEscapingLocalObject(O2)) 1548 return NoAlias; 1549 if (isEscapeSource(O2) && isNonEscapingLocalObject(O1)) 1550 return NoAlias; 1551 } 1552 1553 // If the size of one access is larger than the entire object on the other 1554 // side, then we know such behavior is undefined and can assume no alias. 1555 if ((V1Size != MemoryLocation::UnknownSize && 1556 isObjectSmallerThan(O2, V1Size, DL, TLI)) || 1557 (V2Size != MemoryLocation::UnknownSize && 1558 isObjectSmallerThan(O1, V2Size, DL, TLI))) 1559 return NoAlias; 1560 1561 // Check the cache before climbing up use-def chains. This also terminates 1562 // otherwise infinitely recursive queries. 1563 LocPair Locs(MemoryLocation(V1, V1Size, V1AAInfo), 1564 MemoryLocation(V2, V2Size, V2AAInfo)); 1565 if (V1 > V2) 1566 std::swap(Locs.first, Locs.second); 1567 std::pair<AliasCacheTy::iterator, bool> Pair = 1568 AliasCache.insert(std::make_pair(Locs, MayAlias)); 1569 if (!Pair.second) 1570 return Pair.first->second; 1571 1572 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the 1573 // GEP can't simplify, we don't even look at the PHI cases. 1574 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) { 1575 std::swap(V1, V2); 1576 std::swap(V1Size, V2Size); 1577 std::swap(O1, O2); 1578 std::swap(V1AAInfo, V2AAInfo); 1579 } 1580 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) { 1581 AliasResult Result = 1582 aliasGEP(GV1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O1, O2); 1583 if (Result != MayAlias) 1584 return AliasCache[Locs] = Result; 1585 } 1586 1587 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) { 1588 std::swap(V1, V2); 1589 std::swap(O1, O2); 1590 std::swap(V1Size, V2Size); 1591 std::swap(V1AAInfo, V2AAInfo); 1592 } 1593 if (const PHINode *PN = dyn_cast<PHINode>(V1)) { 1594 AliasResult Result = aliasPHI(PN, V1Size, V1AAInfo, 1595 V2, V2Size, V2AAInfo, O2); 1596 if (Result != MayAlias) 1597 return AliasCache[Locs] = Result; 1598 } 1599 1600 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) { 1601 std::swap(V1, V2); 1602 std::swap(O1, O2); 1603 std::swap(V1Size, V2Size); 1604 std::swap(V1AAInfo, V2AAInfo); 1605 } 1606 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) { 1607 AliasResult Result = 1608 aliasSelect(S1, V1Size, V1AAInfo, V2, V2Size, V2AAInfo, O2); 1609 if (Result != MayAlias) 1610 return AliasCache[Locs] = Result; 1611 } 1612 1613 // If both pointers are pointing into the same object and one of them 1614 // accesses the entire object, then the accesses must overlap in some way. 1615 if (O1 == O2) 1616 if ((V1Size != MemoryLocation::UnknownSize && 1617 isObjectSize(O1, V1Size, DL, TLI)) || 1618 (V2Size != MemoryLocation::UnknownSize && 1619 isObjectSize(O2, V2Size, DL, TLI))) 1620 return AliasCache[Locs] = PartialAlias; 1621 1622 // Recurse back into the best AA results we have, potentially with refined 1623 // memory locations. We have already ensured that BasicAA has a MayAlias 1624 // cache result for these, so any recursion back into BasicAA won't loop. 1625 AliasResult Result = getBestAAResults().alias(Locs.first, Locs.second); 1626 return AliasCache[Locs] = Result; 1627 } 1628 1629 /// Check whether two Values can be considered equivalent. 1630 /// 1631 /// In addition to pointer equivalence of \p V1 and \p V2 this checks whether 1632 /// they can not be part of a cycle in the value graph by looking at all 1633 /// visited phi nodes an making sure that the phis cannot reach the value. We 1634 /// have to do this because we are looking through phi nodes (That is we say 1635 /// noalias(V, phi(VA, VB)) if noalias(V, VA) and noalias(V, VB). 1636 bool BasicAAResult::isValueEqualInPotentialCycles(const Value *V, 1637 const Value *V2) { 1638 if (V != V2) 1639 return false; 1640 1641 const Instruction *Inst = dyn_cast<Instruction>(V); 1642 if (!Inst) 1643 return true; 1644 1645 if (VisitedPhiBBs.empty()) 1646 return true; 1647 1648 if (VisitedPhiBBs.size() > MaxNumPhiBBsValueReachabilityCheck) 1649 return false; 1650 1651 // Make sure that the visited phis cannot reach the Value. This ensures that 1652 // the Values cannot come from different iterations of a potential cycle the 1653 // phi nodes could be involved in. 1654 for (auto *P : VisitedPhiBBs) 1655 if (isPotentiallyReachable(&P->front(), Inst, DT, LI)) 1656 return false; 1657 1658 return true; 1659 } 1660 1661 /// Computes the symbolic difference between two de-composed GEPs. 1662 /// 1663 /// Dest and Src are the variable indices from two decomposed GetElementPtr 1664 /// instructions GEP1 and GEP2 which have common base pointers. 1665 void BasicAAResult::GetIndexDifference( 1666 SmallVectorImpl<VariableGEPIndex> &Dest, 1667 const SmallVectorImpl<VariableGEPIndex> &Src) { 1668 if (Src.empty()) 1669 return; 1670 1671 for (unsigned i = 0, e = Src.size(); i != e; ++i) { 1672 const Value *V = Src[i].V; 1673 unsigned ZExtBits = Src[i].ZExtBits, SExtBits = Src[i].SExtBits; 1674 int64_t Scale = Src[i].Scale; 1675 1676 // Find V in Dest. This is N^2, but pointer indices almost never have more 1677 // than a few variable indexes. 1678 for (unsigned j = 0, e = Dest.size(); j != e; ++j) { 1679 if (!isValueEqualInPotentialCycles(Dest[j].V, V) || 1680 Dest[j].ZExtBits != ZExtBits || Dest[j].SExtBits != SExtBits) 1681 continue; 1682 1683 // If we found it, subtract off Scale V's from the entry in Dest. If it 1684 // goes to zero, remove the entry. 1685 if (Dest[j].Scale != Scale) 1686 Dest[j].Scale -= Scale; 1687 else 1688 Dest.erase(Dest.begin() + j); 1689 Scale = 0; 1690 break; 1691 } 1692 1693 // If we didn't consume this entry, add it to the end of the Dest list. 1694 if (Scale) { 1695 VariableGEPIndex Entry = {V, ZExtBits, SExtBits, -Scale}; 1696 Dest.push_back(Entry); 1697 } 1698 } 1699 } 1700 1701 bool BasicAAResult::constantOffsetHeuristic( 1702 const SmallVectorImpl<VariableGEPIndex> &VarIndices, uint64_t V1Size, 1703 uint64_t V2Size, int64_t BaseOffset, AssumptionCache *AC, 1704 DominatorTree *DT) { 1705 if (VarIndices.size() != 2 || V1Size == MemoryLocation::UnknownSize || 1706 V2Size == MemoryLocation::UnknownSize) 1707 return false; 1708 1709 const VariableGEPIndex &Var0 = VarIndices[0], &Var1 = VarIndices[1]; 1710 1711 if (Var0.ZExtBits != Var1.ZExtBits || Var0.SExtBits != Var1.SExtBits || 1712 Var0.Scale != -Var1.Scale) 1713 return false; 1714 1715 unsigned Width = Var1.V->getType()->getIntegerBitWidth(); 1716 1717 // We'll strip off the Extensions of Var0 and Var1 and do another round 1718 // of GetLinearExpression decomposition. In the example above, if Var0 1719 // is zext(%x + 1) we should get V1 == %x and V1Offset == 1. 1720 1721 APInt V0Scale(Width, 0), V0Offset(Width, 0), V1Scale(Width, 0), 1722 V1Offset(Width, 0); 1723 bool NSW = true, NUW = true; 1724 unsigned V0ZExtBits = 0, V0SExtBits = 0, V1ZExtBits = 0, V1SExtBits = 0; 1725 const Value *V0 = GetLinearExpression(Var0.V, V0Scale, V0Offset, V0ZExtBits, 1726 V0SExtBits, DL, 0, AC, DT, NSW, NUW); 1727 NSW = true; 1728 NUW = true; 1729 const Value *V1 = GetLinearExpression(Var1.V, V1Scale, V1Offset, V1ZExtBits, 1730 V1SExtBits, DL, 0, AC, DT, NSW, NUW); 1731 1732 if (V0Scale != V1Scale || V0ZExtBits != V1ZExtBits || 1733 V0SExtBits != V1SExtBits || !isValueEqualInPotentialCycles(V0, V1)) 1734 return false; 1735 1736 // We have a hit - Var0 and Var1 only differ by a constant offset! 1737 1738 // If we've been sext'ed then zext'd the maximum difference between Var0 and 1739 // Var1 is possible to calculate, but we're just interested in the absolute 1740 // minimum difference between the two. The minimum distance may occur due to 1741 // wrapping; consider "add i3 %i, 5": if %i == 7 then 7 + 5 mod 8 == 4, and so 1742 // the minimum distance between %i and %i + 5 is 3. 1743 APInt MinDiff = V0Offset - V1Offset, Wrapped = -MinDiff; 1744 MinDiff = APIntOps::umin(MinDiff, Wrapped); 1745 uint64_t MinDiffBytes = MinDiff.getZExtValue() * std::abs(Var0.Scale); 1746 1747 // We can't definitely say whether GEP1 is before or after V2 due to wrapping 1748 // arithmetic (i.e. for some values of GEP1 and V2 GEP1 < V2, and for other 1749 // values GEP1 > V2). We'll therefore only declare NoAlias if both V1Size and 1750 // V2Size can fit in the MinDiffBytes gap. 1751 return V1Size + std::abs(BaseOffset) <= MinDiffBytes && 1752 V2Size + std::abs(BaseOffset) <= MinDiffBytes; 1753 } 1754 1755 //===----------------------------------------------------------------------===// 1756 // BasicAliasAnalysis Pass 1757 //===----------------------------------------------------------------------===// 1758 1759 AnalysisKey BasicAA::Key; 1760 1761 BasicAAResult BasicAA::run(Function &F, FunctionAnalysisManager &AM) { 1762 return BasicAAResult(F.getParent()->getDataLayout(), 1763 AM.getResult<TargetLibraryAnalysis>(F), 1764 AM.getResult<AssumptionAnalysis>(F), 1765 &AM.getResult<DominatorTreeAnalysis>(F), 1766 AM.getCachedResult<LoopAnalysis>(F)); 1767 } 1768 1769 BasicAAWrapperPass::BasicAAWrapperPass() : FunctionPass(ID) { 1770 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry()); 1771 } 1772 1773 char BasicAAWrapperPass::ID = 0; 1774 void BasicAAWrapperPass::anchor() {} 1775 1776 INITIALIZE_PASS_BEGIN(BasicAAWrapperPass, "basicaa", 1777 "Basic Alias Analysis (stateless AA impl)", true, true) 1778 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 1779 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 1780 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1781 INITIALIZE_PASS_END(BasicAAWrapperPass, "basicaa", 1782 "Basic Alias Analysis (stateless AA impl)", true, true) 1783 1784 FunctionPass *llvm::createBasicAAWrapperPass() { 1785 return new BasicAAWrapperPass(); 1786 } 1787 1788 bool BasicAAWrapperPass::runOnFunction(Function &F) { 1789 auto &ACT = getAnalysis<AssumptionCacheTracker>(); 1790 auto &TLIWP = getAnalysis<TargetLibraryInfoWrapperPass>(); 1791 auto &DTWP = getAnalysis<DominatorTreeWrapperPass>(); 1792 auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>(); 1793 1794 Result.reset(new BasicAAResult(F.getParent()->getDataLayout(), TLIWP.getTLI(), 1795 ACT.getAssumptionCache(F), &DTWP.getDomTree(), 1796 LIWP ? &LIWP->getLoopInfo() : nullptr)); 1797 1798 return false; 1799 } 1800 1801 void BasicAAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 1802 AU.setPreservesAll(); 1803 AU.addRequired<AssumptionCacheTracker>(); 1804 AU.addRequired<DominatorTreeWrapperPass>(); 1805 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1806 } 1807 1808 BasicAAResult llvm::createLegacyPMBasicAAResult(Pass &P, Function &F) { 1809 return BasicAAResult( 1810 F.getParent()->getDataLayout(), 1811 P.getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(), 1812 P.getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F)); 1813 } 1814