BasicAliasAnalysis.cpp revision 201d1e56bb7535802c70d5eb46601afcc325045d
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/AliasAnalysis.h" 17#include "llvm/Analysis/Passes.h" 18#include "llvm/Constants.h" 19#include "llvm/DerivedTypes.h" 20#include "llvm/Function.h" 21#include "llvm/GlobalAlias.h" 22#include "llvm/GlobalVariable.h" 23#include "llvm/Instructions.h" 24#include "llvm/IntrinsicInst.h" 25#include "llvm/LLVMContext.h" 26#include "llvm/Operator.h" 27#include "llvm/Pass.h" 28#include "llvm/Analysis/CaptureTracking.h" 29#include "llvm/Analysis/MemoryBuiltins.h" 30#include "llvm/Analysis/ValueTracking.h" 31#include "llvm/Target/TargetData.h" 32#include "llvm/ADT/SmallPtrSet.h" 33#include "llvm/ADT/SmallVector.h" 34#include "llvm/Support/ErrorHandling.h" 35#include "llvm/Support/GetElementPtrTypeIterator.h" 36#include <algorithm> 37using namespace llvm; 38 39//===----------------------------------------------------------------------===// 40// Useful predicates 41//===----------------------------------------------------------------------===// 42 43/// isKnownNonNull - Return true if we know that the specified value is never 44/// null. 45static bool isKnownNonNull(const Value *V) { 46 // Alloca never returns null, malloc might. 47 if (isa<AllocaInst>(V)) return true; 48 49 // A byval argument is never null. 50 if (const Argument *A = dyn_cast<Argument>(V)) 51 return A->hasByValAttr(); 52 53 // Global values are not null unless extern weak. 54 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) 55 return !GV->hasExternalWeakLinkage(); 56 return false; 57} 58 59/// isNonEscapingLocalObject - Return true if the pointer is to a function-local 60/// object that never escapes from the function. 61static bool isNonEscapingLocalObject(const Value *V) { 62 // If this is a local allocation, check to see if it escapes. 63 if (isa<AllocaInst>(V) || isNoAliasCall(V)) 64 // Set StoreCaptures to True so that we can assume in our callers that the 65 // pointer is not the result of a load instruction. Currently 66 // PointerMayBeCaptured doesn't have any special analysis for the 67 // StoreCaptures=false case; if it did, our callers could be refined to be 68 // more precise. 69 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true); 70 71 // If this is an argument that corresponds to a byval or noalias argument, 72 // then it has not escaped before entering the function. Check if it escapes 73 // inside the function. 74 if (const Argument *A = dyn_cast<Argument>(V)) 75 if (A->hasByValAttr() || A->hasNoAliasAttr()) { 76 // Don't bother analyzing arguments already known not to escape. 77 if (A->hasNoCaptureAttr()) 78 return true; 79 return !PointerMayBeCaptured(V, false, /*StoreCaptures=*/true); 80 } 81 return false; 82} 83 84/// isEscapeSource - Return true if the pointer is one which would have 85/// been considered an escape by isNonEscapingLocalObject. 86static bool isEscapeSource(const Value *V) { 87 if (isa<CallInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V)) 88 return true; 89 90 // The load case works because isNonEscapingLocalObject considers all 91 // stores to be escapes (it passes true for the StoreCaptures argument 92 // to PointerMayBeCaptured). 93 if (isa<LoadInst>(V)) 94 return true; 95 96 return false; 97} 98 99/// isObjectSmallerThan - Return true if we can prove that the object specified 100/// by V is smaller than Size. 101static bool isObjectSmallerThan(const Value *V, uint64_t Size, 102 const TargetData &TD) { 103 const Type *AccessTy; 104 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) { 105 AccessTy = GV->getType()->getElementType(); 106 } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 107 if (!AI->isArrayAllocation()) 108 AccessTy = AI->getType()->getElementType(); 109 else 110 return false; 111 } else if (const CallInst* CI = extractMallocCall(V)) { 112 if (!isArrayMalloc(V, &TD)) 113 // The size is the argument to the malloc call. 114 if (const ConstantInt* C = dyn_cast<ConstantInt>(CI->getArgOperand(0))) 115 return (C->getZExtValue() < Size); 116 return false; 117 } else if (const Argument *A = dyn_cast<Argument>(V)) { 118 if (A->hasByValAttr()) 119 AccessTy = cast<PointerType>(A->getType())->getElementType(); 120 else 121 return false; 122 } else { 123 return false; 124 } 125 126 if (AccessTy->isSized()) 127 return TD.getTypeAllocSize(AccessTy) < Size; 128 return false; 129} 130 131//===----------------------------------------------------------------------===// 132// GetElementPtr Instruction Decomposition and Analysis 133//===----------------------------------------------------------------------===// 134 135namespace { 136 enum ExtensionKind { 137 EK_NotExtended, 138 EK_SignExt, 139 EK_ZeroExt 140 }; 141 142 struct VariableGEPIndex { 143 const Value *V; 144 ExtensionKind Extension; 145 int64_t Scale; 146 }; 147} 148 149 150/// GetLinearExpression - Analyze the specified value as a linear expression: 151/// "A*V + B", where A and B are constant integers. Return the scale and offset 152/// values as APInts and return V as a Value*, and return whether we looked 153/// through any sign or zero extends. The incoming Value is known to have 154/// IntegerType and it may already be sign or zero extended. 155/// 156/// Note that this looks through extends, so the high bits may not be 157/// represented in the result. 158static Value *GetLinearExpression(Value *V, APInt &Scale, APInt &Offset, 159 ExtensionKind &Extension, 160 const TargetData &TD, unsigned Depth) { 161 assert(V->getType()->isIntegerTy() && "Not an integer value"); 162 163 // Limit our recursion depth. 164 if (Depth == 6) { 165 Scale = 1; 166 Offset = 0; 167 return V; 168 } 169 170 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(V)) { 171 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(BOp->getOperand(1))) { 172 switch (BOp->getOpcode()) { 173 default: break; 174 case Instruction::Or: 175 // X|C == X+C if all the bits in C are unset in X. Otherwise we can't 176 // analyze it. 177 if (!MaskedValueIsZero(BOp->getOperand(0), RHSC->getValue(), &TD)) 178 break; 179 // FALL THROUGH. 180 case Instruction::Add: 181 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension, 182 TD, Depth+1); 183 Offset += RHSC->getValue(); 184 return V; 185 case Instruction::Mul: 186 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension, 187 TD, Depth+1); 188 Offset *= RHSC->getValue(); 189 Scale *= RHSC->getValue(); 190 return V; 191 case Instruction::Shl: 192 V = GetLinearExpression(BOp->getOperand(0), Scale, Offset, Extension, 193 TD, Depth+1); 194 Offset <<= RHSC->getValue().getLimitedValue(); 195 Scale <<= RHSC->getValue().getLimitedValue(); 196 return V; 197 } 198 } 199 } 200 201 // Since GEP indices are sign extended anyway, we don't care about the high 202 // bits of a sign or zero extended value - just scales and offsets. The 203 // extensions have to be consistent though. 204 if ((isa<SExtInst>(V) && Extension != EK_ZeroExt) || 205 (isa<ZExtInst>(V) && Extension != EK_SignExt)) { 206 Value *CastOp = cast<CastInst>(V)->getOperand(0); 207 unsigned OldWidth = Scale.getBitWidth(); 208 unsigned SmallWidth = CastOp->getType()->getPrimitiveSizeInBits(); 209 Scale.trunc(SmallWidth); 210 Offset.trunc(SmallWidth); 211 Extension = isa<SExtInst>(V) ? EK_SignExt : EK_ZeroExt; 212 213 Value *Result = GetLinearExpression(CastOp, Scale, Offset, Extension, 214 TD, Depth+1); 215 Scale.zext(OldWidth); 216 Offset.zext(OldWidth); 217 218 return Result; 219 } 220 221 Scale = 1; 222 Offset = 0; 223 return V; 224} 225 226/// DecomposeGEPExpression - If V is a symbolic pointer expression, decompose it 227/// into a base pointer with a constant offset and a number of scaled symbolic 228/// offsets. 229/// 230/// The scaled symbolic offsets (represented by pairs of a Value* and a scale in 231/// the VarIndices vector) are Value*'s that are known to be scaled by the 232/// specified amount, but which may have other unrepresented high bits. As such, 233/// the gep cannot necessarily be reconstructed from its decomposed form. 234/// 235/// When TargetData is around, this function is capable of analyzing everything 236/// that Value::getUnderlyingObject() can look through. When not, it just looks 237/// through pointer casts. 238/// 239static const Value * 240DecomposeGEPExpression(const Value *V, int64_t &BaseOffs, 241 SmallVectorImpl<VariableGEPIndex> &VarIndices, 242 const TargetData *TD) { 243 // Limit recursion depth to limit compile time in crazy cases. 244 unsigned MaxLookup = 6; 245 246 BaseOffs = 0; 247 do { 248 // See if this is a bitcast or GEP. 249 const Operator *Op = dyn_cast<Operator>(V); 250 if (Op == 0) { 251 // The only non-operator case we can handle are GlobalAliases. 252 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 253 if (!GA->mayBeOverridden()) { 254 V = GA->getAliasee(); 255 continue; 256 } 257 } 258 return V; 259 } 260 261 if (Op->getOpcode() == Instruction::BitCast) { 262 V = Op->getOperand(0); 263 continue; 264 } 265 266 const GEPOperator *GEPOp = dyn_cast<GEPOperator>(Op); 267 if (GEPOp == 0) 268 return V; 269 270 // Don't attempt to analyze GEPs over unsized objects. 271 if (!cast<PointerType>(GEPOp->getOperand(0)->getType()) 272 ->getElementType()->isSized()) 273 return V; 274 275 // If we are lacking TargetData information, we can't compute the offets of 276 // elements computed by GEPs. However, we can handle bitcast equivalent 277 // GEPs. 278 if (TD == 0) { 279 if (!GEPOp->hasAllZeroIndices()) 280 return V; 281 V = GEPOp->getOperand(0); 282 continue; 283 } 284 285 // Walk the indices of the GEP, accumulating them into BaseOff/VarIndices. 286 gep_type_iterator GTI = gep_type_begin(GEPOp); 287 for (User::const_op_iterator I = GEPOp->op_begin()+1, 288 E = GEPOp->op_end(); I != E; ++I) { 289 Value *Index = *I; 290 // Compute the (potentially symbolic) offset in bytes for this index. 291 if (const StructType *STy = dyn_cast<StructType>(*GTI++)) { 292 // For a struct, add the member offset. 293 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue(); 294 if (FieldNo == 0) continue; 295 296 BaseOffs += TD->getStructLayout(STy)->getElementOffset(FieldNo); 297 continue; 298 } 299 300 // For an array/pointer, add the element offset, explicitly scaled. 301 if (ConstantInt *CIdx = dyn_cast<ConstantInt>(Index)) { 302 if (CIdx->isZero()) continue; 303 BaseOffs += TD->getTypeAllocSize(*GTI)*CIdx->getSExtValue(); 304 continue; 305 } 306 307 uint64_t Scale = TD->getTypeAllocSize(*GTI); 308 ExtensionKind Extension = EK_NotExtended; 309 310 // If the integer type is smaller than the pointer size, it is implicitly 311 // sign extended to pointer size. 312 unsigned Width = cast<IntegerType>(Index->getType())->getBitWidth(); 313 if (TD->getPointerSizeInBits() > Width) 314 Extension = EK_SignExt; 315 316 // Use GetLinearExpression to decompose the index into a C1*V+C2 form. 317 APInt IndexScale(Width, 0), IndexOffset(Width, 0); 318 Index = GetLinearExpression(Index, IndexScale, IndexOffset, Extension, 319 *TD, 0); 320 321 // The GEP index scale ("Scale") scales C1*V+C2, yielding (C1*V+C2)*Scale. 322 // This gives us an aggregate computation of (C1*Scale)*V + C2*Scale. 323 BaseOffs += IndexOffset.getSExtValue()*Scale; 324 Scale *= IndexScale.getSExtValue(); 325 326 327 // If we already had an occurrance of this index variable, merge this 328 // scale into it. For example, we want to handle: 329 // A[x][x] -> x*16 + x*4 -> x*20 330 // This also ensures that 'x' only appears in the index list once. 331 for (unsigned i = 0, e = VarIndices.size(); i != e; ++i) { 332 if (VarIndices[i].V == Index && 333 VarIndices[i].Extension == Extension) { 334 Scale += VarIndices[i].Scale; 335 VarIndices.erase(VarIndices.begin()+i); 336 break; 337 } 338 } 339 340 // Make sure that we have a scale that makes sense for this target's 341 // pointer size. 342 if (unsigned ShiftBits = 64-TD->getPointerSizeInBits()) { 343 Scale <<= ShiftBits; 344 Scale = (int64_t)Scale >> ShiftBits; 345 } 346 347 if (Scale) { 348 VariableGEPIndex Entry = {Index, Extension, Scale}; 349 VarIndices.push_back(Entry); 350 } 351 } 352 353 // Analyze the base pointer next. 354 V = GEPOp->getOperand(0); 355 } while (--MaxLookup); 356 357 // If the chain of expressions is too deep, just return early. 358 return V; 359} 360 361/// GetIndexDifference - Dest and Src are the variable indices from two 362/// decomposed GetElementPtr instructions GEP1 and GEP2 which have common base 363/// pointers. Subtract the GEP2 indices from GEP1 to find the symbolic 364/// difference between the two pointers. 365static void GetIndexDifference(SmallVectorImpl<VariableGEPIndex> &Dest, 366 const SmallVectorImpl<VariableGEPIndex> &Src) { 367 if (Src.empty()) return; 368 369 for (unsigned i = 0, e = Src.size(); i != e; ++i) { 370 const Value *V = Src[i].V; 371 ExtensionKind Extension = Src[i].Extension; 372 int64_t Scale = Src[i].Scale; 373 374 // Find V in Dest. This is N^2, but pointer indices almost never have more 375 // than a few variable indexes. 376 for (unsigned j = 0, e = Dest.size(); j != e; ++j) { 377 if (Dest[j].V != V || Dest[j].Extension != Extension) continue; 378 379 // If we found it, subtract off Scale V's from the entry in Dest. If it 380 // goes to zero, remove the entry. 381 if (Dest[j].Scale != Scale) 382 Dest[j].Scale -= Scale; 383 else 384 Dest.erase(Dest.begin()+j); 385 Scale = 0; 386 break; 387 } 388 389 // If we didn't consume this entry, add it to the end of the Dest list. 390 if (Scale) { 391 VariableGEPIndex Entry = { V, Extension, -Scale }; 392 Dest.push_back(Entry); 393 } 394 } 395} 396 397//===----------------------------------------------------------------------===// 398// BasicAliasAnalysis Pass 399//===----------------------------------------------------------------------===// 400 401#ifndef NDEBUG 402static const Function *getParent(const Value *V) { 403 if (const Instruction *inst = dyn_cast<Instruction>(V)) 404 return inst->getParent()->getParent(); 405 406 if (const Argument *arg = dyn_cast<Argument>(V)) 407 return arg->getParent(); 408 409 return NULL; 410} 411 412static bool notDifferentParent(const Value *O1, const Value *O2) { 413 414 const Function *F1 = getParent(O1); 415 const Function *F2 = getParent(O2); 416 417 return !F1 || !F2 || F1 == F2; 418} 419#endif 420 421namespace { 422 /// BasicAliasAnalysis - This is the primary alias analysis implementation. 423 struct BasicAliasAnalysis : public ImmutablePass, public AliasAnalysis { 424 static char ID; // Class identification, replacement for typeinfo 425 BasicAliasAnalysis() : ImmutablePass(ID) { 426 initializeBasicAliasAnalysisPass(*PassRegistry::getPassRegistry()); 427 } 428 429 virtual void initializePass() { 430 InitializeAliasAnalysis(this); 431 } 432 433 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 434 AU.addRequired<AliasAnalysis>(); 435 } 436 437 virtual AliasResult alias(const Location &LocA, 438 const Location &LocB) { 439 assert(Visited.empty() && "Visited must be cleared after use!"); 440 assert(notDifferentParent(LocA.Ptr, LocB.Ptr) && 441 "BasicAliasAnalysis doesn't support interprocedural queries."); 442 AliasResult Alias = aliasCheck(LocA.Ptr, LocA.Size, LocA.TBAATag, 443 LocB.Ptr, LocB.Size, LocB.TBAATag); 444 Visited.clear(); 445 return Alias; 446 } 447 448 virtual ModRefResult getModRefInfo(ImmutableCallSite CS, 449 const Location &Loc); 450 451 virtual ModRefResult getModRefInfo(ImmutableCallSite CS1, 452 ImmutableCallSite CS2) { 453 // The AliasAnalysis base class has some smarts, lets use them. 454 return AliasAnalysis::getModRefInfo(CS1, CS2); 455 } 456 457 /// pointsToConstantMemory - Chase pointers until we find a (constant 458 /// global) or not. 459 virtual bool pointsToConstantMemory(const Location &Loc, bool OrLocal); 460 461 /// getModRefBehavior - Return the behavior when calling the given 462 /// call site. 463 virtual ModRefBehavior getModRefBehavior(ImmutableCallSite CS); 464 465 /// getModRefBehavior - Return the behavior when calling the given function. 466 /// For use when the call site is not known. 467 virtual ModRefBehavior getModRefBehavior(const Function *F); 468 469 /// getAdjustedAnalysisPointer - This method is used when a pass implements 470 /// an analysis interface through multiple inheritance. If needed, it 471 /// should override this to adjust the this pointer as needed for the 472 /// specified pass info. 473 virtual void *getAdjustedAnalysisPointer(const void *ID) { 474 if (ID == &AliasAnalysis::ID) 475 return (AliasAnalysis*)this; 476 return this; 477 } 478 479 private: 480 // Visited - Track instructions visited by a aliasPHI, aliasSelect(), and aliasGEP(). 481 SmallPtrSet<const Value*, 16> Visited; 482 483 // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP 484 // instruction against another. 485 AliasResult aliasGEP(const GEPOperator *V1, uint64_t V1Size, 486 const Value *V2, uint64_t V2Size, 487 const MDNode *V2TBAAInfo, 488 const Value *UnderlyingV1, const Value *UnderlyingV2); 489 490 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI 491 // instruction against another. 492 AliasResult aliasPHI(const PHINode *PN, uint64_t PNSize, 493 const MDNode *PNTBAAInfo, 494 const Value *V2, uint64_t V2Size, 495 const MDNode *V2TBAAInfo); 496 497 /// aliasSelect - Disambiguate a Select instruction against another value. 498 AliasResult aliasSelect(const SelectInst *SI, uint64_t SISize, 499 const MDNode *SITBAAInfo, 500 const Value *V2, uint64_t V2Size, 501 const MDNode *V2TBAAInfo); 502 503 AliasResult aliasCheck(const Value *V1, uint64_t V1Size, 504 const MDNode *V1TBAATag, 505 const Value *V2, uint64_t V2Size, 506 const MDNode *V2TBAATag); 507 }; 508} // End of anonymous namespace 509 510// Register this pass... 511char BasicAliasAnalysis::ID = 0; 512INITIALIZE_AG_PASS(BasicAliasAnalysis, AliasAnalysis, "basicaa", 513 "Basic Alias Analysis (stateless AA impl)", 514 false, true, false) 515 516ImmutablePass *llvm::createBasicAliasAnalysisPass() { 517 return new BasicAliasAnalysis(); 518} 519 520/// pointsToConstantMemory - Returns whether the given pointer value 521/// points to memory that is local to the function, with global constants being 522/// considered local to all functions. 523bool 524BasicAliasAnalysis::pointsToConstantMemory(const Location &Loc, bool OrLocal) { 525 assert(Visited.empty() && "Visited must be cleared after use!"); 526 527 unsigned MaxLookup = 8; 528 SmallVector<const Value *, 16> Worklist; 529 Worklist.push_back(Loc.Ptr); 530 do { 531 const Value *V = Worklist.pop_back_val()->getUnderlyingObject(); 532 if (!Visited.insert(V)) { 533 Visited.clear(); 534 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal); 535 } 536 537 // An alloca instruction defines local memory. 538 if (OrLocal && isa<AllocaInst>(V)) 539 continue; 540 541 // A global constant counts as local memory for our purposes. 542 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) { 543 // Note: this doesn't require GV to be "ODR" because it isn't legal for a 544 // global to be marked constant in some modules and non-constant in 545 // others. GV may even be a declaration, not a definition. 546 if (!GV->isConstant()) { 547 Visited.clear(); 548 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal); 549 } 550 continue; 551 } 552 553 // If both select values point to local memory, then so does the select. 554 if (const SelectInst *SI = dyn_cast<SelectInst>(V)) { 555 Worklist.push_back(SI->getTrueValue()); 556 Worklist.push_back(SI->getFalseValue()); 557 continue; 558 } 559 560 // If all values incoming to a phi node point to local memory, then so does 561 // the phi. 562 if (const PHINode *PN = dyn_cast<PHINode>(V)) { 563 // Don't bother inspecting phi nodes with many operands. 564 if (PN->getNumIncomingValues() > MaxLookup) { 565 Visited.clear(); 566 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal); 567 } 568 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 569 Worklist.push_back(PN->getIncomingValue(i)); 570 continue; 571 } 572 573 // Otherwise be conservative. 574 Visited.clear(); 575 return AliasAnalysis::pointsToConstantMemory(Loc, OrLocal); 576 577 } while (!Worklist.empty() && --MaxLookup); 578 579 Visited.clear(); 580 return Worklist.empty(); 581} 582 583/// getModRefBehavior - Return the behavior when calling the given call site. 584AliasAnalysis::ModRefBehavior 585BasicAliasAnalysis::getModRefBehavior(ImmutableCallSite CS) { 586 if (CS.doesNotAccessMemory()) 587 // Can't do better than this. 588 return DoesNotAccessMemory; 589 590 ModRefBehavior Min = UnknownModRefBehavior; 591 592 // If the callsite knows it only reads memory, don't return worse 593 // than that. 594 if (CS.onlyReadsMemory()) 595 Min = OnlyReadsMemory; 596 597 // The AliasAnalysis base class has some smarts, lets use them. 598 return ModRefBehavior(AliasAnalysis::getModRefBehavior(CS) & Min); 599} 600 601/// getModRefBehavior - Return the behavior when calling the given function. 602/// For use when the call site is not known. 603AliasAnalysis::ModRefBehavior 604BasicAliasAnalysis::getModRefBehavior(const Function *F) { 605 // If the function declares it doesn't access memory, we can't do better. 606 if (F->doesNotAccessMemory()) 607 return DoesNotAccessMemory; 608 609 // For intrinsics, we can check the table. 610 if (unsigned iid = F->getIntrinsicID()) { 611#define GET_INTRINSIC_MODREF_BEHAVIOR 612#include "llvm/Intrinsics.gen" 613#undef GET_INTRINSIC_MODREF_BEHAVIOR 614 } 615 616 ModRefBehavior Min = UnknownModRefBehavior; 617 618 // If the function declares it only reads memory, go with that. 619 if (F->onlyReadsMemory()) 620 Min = OnlyReadsMemory; 621 622 // Otherwise be conservative. 623 return ModRefBehavior(AliasAnalysis::getModRefBehavior(F) & Min); 624} 625 626/// getModRefInfo - Check to see if the specified callsite can clobber the 627/// specified memory object. Since we only look at local properties of this 628/// function, we really can't say much about this query. We do, however, use 629/// simple "address taken" analysis on local objects. 630AliasAnalysis::ModRefResult 631BasicAliasAnalysis::getModRefInfo(ImmutableCallSite CS, 632 const Location &Loc) { 633 assert(notDifferentParent(CS.getInstruction(), Loc.Ptr) && 634 "AliasAnalysis query involving multiple functions!"); 635 636 const Value *Object = Loc.Ptr->getUnderlyingObject(); 637 638 // If this is a tail call and Loc.Ptr points to a stack location, we know that 639 // the tail call cannot access or modify the local stack. 640 // We cannot exclude byval arguments here; these belong to the caller of 641 // the current function not to the current function, and a tail callee 642 // may reference them. 643 if (isa<AllocaInst>(Object)) 644 if (const CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) 645 if (CI->isTailCall()) 646 return NoModRef; 647 648 // If the pointer is to a locally allocated object that does not escape, 649 // then the call can not mod/ref the pointer unless the call takes the pointer 650 // as an argument, and itself doesn't capture it. 651 if (!isa<Constant>(Object) && CS.getInstruction() != Object && 652 isNonEscapingLocalObject(Object)) { 653 bool PassedAsArg = false; 654 unsigned ArgNo = 0; 655 for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end(); 656 CI != CE; ++CI, ++ArgNo) { 657 // Only look at the no-capture pointer arguments. 658 if (!(*CI)->getType()->isPointerTy() || 659 !CS.paramHasAttr(ArgNo+1, Attribute::NoCapture)) 660 continue; 661 662 // If this is a no-capture pointer argument, see if we can tell that it 663 // is impossible to alias the pointer we're checking. If not, we have to 664 // assume that the call could touch the pointer, even though it doesn't 665 // escape. 666 if (!isNoAlias(Location(cast<Value>(CI)), Loc)) { 667 PassedAsArg = true; 668 break; 669 } 670 } 671 672 if (!PassedAsArg) 673 return NoModRef; 674 } 675 676 ModRefResult Min = ModRef; 677 678 // Finally, handle specific knowledge of intrinsics. 679 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction()); 680 if (II != 0) 681 switch (II->getIntrinsicID()) { 682 default: break; 683 case Intrinsic::memcpy: 684 case Intrinsic::memmove: { 685 uint64_t Len = UnknownSize; 686 if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2))) 687 Len = LenCI->getZExtValue(); 688 Value *Dest = II->getArgOperand(0); 689 Value *Src = II->getArgOperand(1); 690 if (isNoAlias(Location(Dest, Len), Loc)) { 691 if (isNoAlias(Location(Src, Len), Loc)) 692 return NoModRef; 693 Min = Ref; 694 } 695 break; 696 } 697 case Intrinsic::memset: 698 // Since memset is 'accesses arguments' only, the AliasAnalysis base class 699 // will handle it for the variable length case. 700 if (ConstantInt *LenCI = dyn_cast<ConstantInt>(II->getArgOperand(2))) { 701 uint64_t Len = LenCI->getZExtValue(); 702 Value *Dest = II->getArgOperand(0); 703 if (isNoAlias(Location(Dest, Len), Loc)) 704 return NoModRef; 705 } 706 // We know that memset doesn't load anything. 707 Min = Mod; 708 break; 709 case Intrinsic::atomic_cmp_swap: 710 case Intrinsic::atomic_swap: 711 case Intrinsic::atomic_load_add: 712 case Intrinsic::atomic_load_sub: 713 case Intrinsic::atomic_load_and: 714 case Intrinsic::atomic_load_nand: 715 case Intrinsic::atomic_load_or: 716 case Intrinsic::atomic_load_xor: 717 case Intrinsic::atomic_load_max: 718 case Intrinsic::atomic_load_min: 719 case Intrinsic::atomic_load_umax: 720 case Intrinsic::atomic_load_umin: 721 if (TD) { 722 Value *Op1 = II->getArgOperand(0); 723 uint64_t Op1Size = TD->getTypeStoreSize(Op1->getType()); 724 MDNode *Tag = II->getMetadata(LLVMContext::MD_tbaa); 725 if (isNoAlias(Location(Op1, Op1Size, Tag), Loc)) 726 return NoModRef; 727 } 728 break; 729 case Intrinsic::lifetime_start: 730 case Intrinsic::lifetime_end: 731 case Intrinsic::invariant_start: { 732 uint64_t PtrSize = 733 cast<ConstantInt>(II->getArgOperand(0))->getZExtValue(); 734 if (isNoAlias(Location(II->getArgOperand(1), 735 PtrSize, 736 II->getMetadata(LLVMContext::MD_tbaa)), 737 Loc)) 738 return NoModRef; 739 break; 740 } 741 case Intrinsic::invariant_end: { 742 uint64_t PtrSize = 743 cast<ConstantInt>(II->getArgOperand(1))->getZExtValue(); 744 if (isNoAlias(Location(II->getArgOperand(2), 745 PtrSize, 746 II->getMetadata(LLVMContext::MD_tbaa)), 747 Loc)) 748 return NoModRef; 749 break; 750 } 751 } 752 753 // The AliasAnalysis base class has some smarts, lets use them. 754 return ModRefResult(AliasAnalysis::getModRefInfo(CS, Loc) & Min); 755} 756 757/// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction 758/// against another pointer. We know that V1 is a GEP, but we don't know 759/// anything about V2. UnderlyingV1 is GEP1->getUnderlyingObject(), 760/// UnderlyingV2 is the same for V2. 761/// 762AliasAnalysis::AliasResult 763BasicAliasAnalysis::aliasGEP(const GEPOperator *GEP1, uint64_t V1Size, 764 const Value *V2, uint64_t V2Size, 765 const MDNode *V2TBAAInfo, 766 const Value *UnderlyingV1, 767 const Value *UnderlyingV2) { 768 // If this GEP has been visited before, we're on a use-def cycle. 769 // Such cycles are only valid when PHI nodes are involved or in unreachable 770 // code. The visitPHI function catches cycles containing PHIs, but there 771 // could still be a cycle without PHIs in unreachable code. 772 if (!Visited.insert(GEP1)) 773 return MayAlias; 774 775 int64_t GEP1BaseOffset; 776 SmallVector<VariableGEPIndex, 4> GEP1VariableIndices; 777 778 // If we have two gep instructions with must-alias'ing base pointers, figure 779 // out if the indexes to the GEP tell us anything about the derived pointer. 780 if (const GEPOperator *GEP2 = dyn_cast<GEPOperator>(V2)) { 781 // Do the base pointers alias? 782 AliasResult BaseAlias = aliasCheck(UnderlyingV1, UnknownSize, 0, 783 UnderlyingV2, UnknownSize, 0); 784 785 // If we get a No or May, then return it immediately, no amount of analysis 786 // will improve this situation. 787 if (BaseAlias != MustAlias) return BaseAlias; 788 789 // Otherwise, we have a MustAlias. Since the base pointers alias each other 790 // exactly, see if the computed offset from the common pointer tells us 791 // about the relation of the resulting pointer. 792 const Value *GEP1BasePtr = 793 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD); 794 795 int64_t GEP2BaseOffset; 796 SmallVector<VariableGEPIndex, 4> GEP2VariableIndices; 797 const Value *GEP2BasePtr = 798 DecomposeGEPExpression(GEP2, GEP2BaseOffset, GEP2VariableIndices, TD); 799 800 // If DecomposeGEPExpression isn't able to look all the way through the 801 // addressing operation, we must not have TD and this is too complex for us 802 // to handle without it. 803 if (GEP1BasePtr != UnderlyingV1 || GEP2BasePtr != UnderlyingV2) { 804 assert(TD == 0 && 805 "DecomposeGEPExpression and getUnderlyingObject disagree!"); 806 return MayAlias; 807 } 808 809 // Subtract the GEP2 pointer from the GEP1 pointer to find out their 810 // symbolic difference. 811 GEP1BaseOffset -= GEP2BaseOffset; 812 GetIndexDifference(GEP1VariableIndices, GEP2VariableIndices); 813 814 } else { 815 // Check to see if these two pointers are related by the getelementptr 816 // instruction. If one pointer is a GEP with a non-zero index of the other 817 // pointer, we know they cannot alias. 818 819 // If both accesses are unknown size, we can't do anything useful here. 820 if (V1Size == UnknownSize && V2Size == UnknownSize) 821 return MayAlias; 822 823 AliasResult R = aliasCheck(UnderlyingV1, UnknownSize, 0, 824 V2, V2Size, V2TBAAInfo); 825 if (R != MustAlias) 826 // If V2 may alias GEP base pointer, conservatively returns MayAlias. 827 // If V2 is known not to alias GEP base pointer, then the two values 828 // cannot alias per GEP semantics: "A pointer value formed from a 829 // getelementptr instruction is associated with the addresses associated 830 // with the first operand of the getelementptr". 831 return R; 832 833 const Value *GEP1BasePtr = 834 DecomposeGEPExpression(GEP1, GEP1BaseOffset, GEP1VariableIndices, TD); 835 836 // If DecomposeGEPExpression isn't able to look all the way through the 837 // addressing operation, we must not have TD and this is too complex for us 838 // to handle without it. 839 if (GEP1BasePtr != UnderlyingV1) { 840 assert(TD == 0 && 841 "DecomposeGEPExpression and getUnderlyingObject disagree!"); 842 return MayAlias; 843 } 844 } 845 846 // In the two GEP Case, if there is no difference in the offsets of the 847 // computed pointers, the resultant pointers are a must alias. This 848 // hapens when we have two lexically identical GEP's (for example). 849 // 850 // In the other case, if we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 851 // must aliases the GEP, the end result is a must alias also. 852 if (GEP1BaseOffset == 0 && GEP1VariableIndices.empty()) 853 return MustAlias; 854 855 // If we have a known constant offset, see if this offset is larger than the 856 // access size being queried. If so, and if no variable indices can remove 857 // pieces of this constant, then we know we have a no-alias. For example, 858 // &A[100] != &A. 859 860 // In order to handle cases like &A[100][i] where i is an out of range 861 // subscript, we have to ignore all constant offset pieces that are a multiple 862 // of a scaled index. Do this by removing constant offsets that are a 863 // multiple of any of our variable indices. This allows us to transform 864 // things like &A[i][1] because i has a stride of (e.g.) 8 bytes but the 1 865 // provides an offset of 4 bytes (assuming a <= 4 byte access). 866 for (unsigned i = 0, e = GEP1VariableIndices.size(); 867 i != e && GEP1BaseOffset;++i) 868 if (int64_t RemovedOffset = GEP1BaseOffset/GEP1VariableIndices[i].Scale) 869 GEP1BaseOffset -= RemovedOffset*GEP1VariableIndices[i].Scale; 870 871 // If our known offset is bigger than the access size, we know we don't have 872 // an alias. 873 if (GEP1BaseOffset) { 874 if (GEP1BaseOffset >= 0 ? 875 (V2Size != UnknownSize && (uint64_t)GEP1BaseOffset >= V2Size) : 876 (V1Size != UnknownSize && -(uint64_t)GEP1BaseOffset >= V1Size && 877 GEP1BaseOffset != INT64_MIN)) 878 return NoAlias; 879 } 880 881 return MayAlias; 882} 883 884/// aliasSelect - Provide a bunch of ad-hoc rules to disambiguate a Select 885/// instruction against another. 886AliasAnalysis::AliasResult 887BasicAliasAnalysis::aliasSelect(const SelectInst *SI, uint64_t SISize, 888 const MDNode *SITBAAInfo, 889 const Value *V2, uint64_t V2Size, 890 const MDNode *V2TBAAInfo) { 891 // If this select has been visited before, we're on a use-def cycle. 892 // Such cycles are only valid when PHI nodes are involved or in unreachable 893 // code. The visitPHI function catches cycles containing PHIs, but there 894 // could still be a cycle without PHIs in unreachable code. 895 if (!Visited.insert(SI)) 896 return MayAlias; 897 898 // If the values are Selects with the same condition, we can do a more precise 899 // check: just check for aliases between the values on corresponding arms. 900 if (const SelectInst *SI2 = dyn_cast<SelectInst>(V2)) 901 if (SI->getCondition() == SI2->getCondition()) { 902 AliasResult Alias = 903 aliasCheck(SI->getTrueValue(), SISize, SITBAAInfo, 904 SI2->getTrueValue(), V2Size, V2TBAAInfo); 905 if (Alias == MayAlias) 906 return MayAlias; 907 AliasResult ThisAlias = 908 aliasCheck(SI->getFalseValue(), SISize, SITBAAInfo, 909 SI2->getFalseValue(), V2Size, V2TBAAInfo); 910 if (ThisAlias != Alias) 911 return MayAlias; 912 return Alias; 913 } 914 915 // If both arms of the Select node NoAlias or MustAlias V2, then returns 916 // NoAlias / MustAlias. Otherwise, returns MayAlias. 917 AliasResult Alias = 918 aliasCheck(V2, V2Size, V2TBAAInfo, SI->getTrueValue(), SISize, SITBAAInfo); 919 if (Alias == MayAlias) 920 return MayAlias; 921 922 // If V2 is visited, the recursive case will have been caught in the 923 // above aliasCheck call, so these subsequent calls to aliasCheck 924 // don't need to assume that V2 is being visited recursively. 925 Visited.erase(V2); 926 927 AliasResult ThisAlias = 928 aliasCheck(V2, V2Size, V2TBAAInfo, SI->getFalseValue(), SISize, SITBAAInfo); 929 if (ThisAlias != Alias) 930 return MayAlias; 931 return Alias; 932} 933 934// aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction 935// against another. 936AliasAnalysis::AliasResult 937BasicAliasAnalysis::aliasPHI(const PHINode *PN, uint64_t PNSize, 938 const MDNode *PNTBAAInfo, 939 const Value *V2, uint64_t V2Size, 940 const MDNode *V2TBAAInfo) { 941 // The PHI node has already been visited, avoid recursion any further. 942 if (!Visited.insert(PN)) 943 return MayAlias; 944 945 // If the values are PHIs in the same block, we can do a more precise 946 // as well as efficient check: just check for aliases between the values 947 // on corresponding edges. 948 if (const PHINode *PN2 = dyn_cast<PHINode>(V2)) 949 if (PN2->getParent() == PN->getParent()) { 950 AliasResult Alias = 951 aliasCheck(PN->getIncomingValue(0), PNSize, PNTBAAInfo, 952 PN2->getIncomingValueForBlock(PN->getIncomingBlock(0)), 953 V2Size, V2TBAAInfo); 954 if (Alias == MayAlias) 955 return MayAlias; 956 for (unsigned i = 1, e = PN->getNumIncomingValues(); i != e; ++i) { 957 AliasResult ThisAlias = 958 aliasCheck(PN->getIncomingValue(i), PNSize, PNTBAAInfo, 959 PN2->getIncomingValueForBlock(PN->getIncomingBlock(i)), 960 V2Size, V2TBAAInfo); 961 if (ThisAlias != Alias) 962 return MayAlias; 963 } 964 return Alias; 965 } 966 967 SmallPtrSet<Value*, 4> UniqueSrc; 968 SmallVector<Value*, 4> V1Srcs; 969 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 970 Value *PV1 = PN->getIncomingValue(i); 971 if (isa<PHINode>(PV1)) 972 // If any of the source itself is a PHI, return MayAlias conservatively 973 // to avoid compile time explosion. The worst possible case is if both 974 // sides are PHI nodes. In which case, this is O(m x n) time where 'm' 975 // and 'n' are the number of PHI sources. 976 return MayAlias; 977 if (UniqueSrc.insert(PV1)) 978 V1Srcs.push_back(PV1); 979 } 980 981 AliasResult Alias = aliasCheck(V2, V2Size, V2TBAAInfo, 982 V1Srcs[0], PNSize, PNTBAAInfo); 983 // Early exit if the check of the first PHI source against V2 is MayAlias. 984 // Other results are not possible. 985 if (Alias == MayAlias) 986 return MayAlias; 987 988 // If all sources of the PHI node NoAlias or MustAlias V2, then returns 989 // NoAlias / MustAlias. Otherwise, returns MayAlias. 990 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) { 991 Value *V = V1Srcs[i]; 992 993 // If V2 is visited, the recursive case will have been caught in the 994 // above aliasCheck call, so these subsequent calls to aliasCheck 995 // don't need to assume that V2 is being visited recursively. 996 Visited.erase(V2); 997 998 AliasResult ThisAlias = aliasCheck(V2, V2Size, V2TBAAInfo, 999 V, PNSize, PNTBAAInfo); 1000 if (ThisAlias != Alias || ThisAlias == MayAlias) 1001 return MayAlias; 1002 } 1003 1004 return Alias; 1005} 1006 1007// aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases, 1008// such as array references. 1009// 1010AliasAnalysis::AliasResult 1011BasicAliasAnalysis::aliasCheck(const Value *V1, uint64_t V1Size, 1012 const MDNode *V1TBAAInfo, 1013 const Value *V2, uint64_t V2Size, 1014 const MDNode *V2TBAAInfo) { 1015 // If either of the memory references is empty, it doesn't matter what the 1016 // pointer values are. 1017 if (V1Size == 0 || V2Size == 0) 1018 return NoAlias; 1019 1020 // Strip off any casts if they exist. 1021 V1 = V1->stripPointerCasts(); 1022 V2 = V2->stripPointerCasts(); 1023 1024 // Are we checking for alias of the same value? 1025 if (V1 == V2) return MustAlias; 1026 1027 if (!V1->getType()->isPointerTy() || !V2->getType()->isPointerTy()) 1028 return NoAlias; // Scalars cannot alias each other 1029 1030 // Figure out what objects these things are pointing to if we can. 1031 const Value *O1 = V1->getUnderlyingObject(); 1032 const Value *O2 = V2->getUnderlyingObject(); 1033 1034 // Null values in the default address space don't point to any object, so they 1035 // don't alias any other pointer. 1036 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O1)) 1037 if (CPN->getType()->getAddressSpace() == 0) 1038 return NoAlias; 1039 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(O2)) 1040 if (CPN->getType()->getAddressSpace() == 0) 1041 return NoAlias; 1042 1043 if (O1 != O2) { 1044 // If V1/V2 point to two different objects we know that we have no alias. 1045 if (isIdentifiedObject(O1) && isIdentifiedObject(O2)) 1046 return NoAlias; 1047 1048 // Constant pointers can't alias with non-const isIdentifiedObject objects. 1049 if ((isa<Constant>(O1) && isIdentifiedObject(O2) && !isa<Constant>(O2)) || 1050 (isa<Constant>(O2) && isIdentifiedObject(O1) && !isa<Constant>(O1))) 1051 return NoAlias; 1052 1053 // Arguments can't alias with local allocations or noalias calls 1054 // in the same function. 1055 if (((isa<Argument>(O1) && (isa<AllocaInst>(O2) || isNoAliasCall(O2))) || 1056 (isa<Argument>(O2) && (isa<AllocaInst>(O1) || isNoAliasCall(O1))))) 1057 return NoAlias; 1058 1059 // Most objects can't alias null. 1060 if ((isa<ConstantPointerNull>(O2) && isKnownNonNull(O1)) || 1061 (isa<ConstantPointerNull>(O1) && isKnownNonNull(O2))) 1062 return NoAlias; 1063 1064 // If one pointer is the result of a call/invoke or load and the other is a 1065 // non-escaping local object within the same function, then we know the 1066 // object couldn't escape to a point where the call could return it. 1067 // 1068 // Note that if the pointers are in different functions, there are a 1069 // variety of complications. A call with a nocapture argument may still 1070 // temporary store the nocapture argument's value in a temporary memory 1071 // location if that memory location doesn't escape. Or it may pass a 1072 // nocapture value to other functions as long as they don't capture it. 1073 if (isEscapeSource(O1) && isNonEscapingLocalObject(O2)) 1074 return NoAlias; 1075 if (isEscapeSource(O2) && isNonEscapingLocalObject(O1)) 1076 return NoAlias; 1077 } 1078 1079 // If the size of one access is larger than the entire object on the other 1080 // side, then we know such behavior is undefined and can assume no alias. 1081 if (TD) 1082 if ((V1Size != UnknownSize && isObjectSmallerThan(O2, V1Size, *TD)) || 1083 (V2Size != UnknownSize && isObjectSmallerThan(O1, V2Size, *TD))) 1084 return NoAlias; 1085 1086 // FIXME: This isn't aggressively handling alias(GEP, PHI) for example: if the 1087 // GEP can't simplify, we don't even look at the PHI cases. 1088 if (!isa<GEPOperator>(V1) && isa<GEPOperator>(V2)) { 1089 std::swap(V1, V2); 1090 std::swap(V1Size, V2Size); 1091 std::swap(O1, O2); 1092 } 1093 if (const GEPOperator *GV1 = dyn_cast<GEPOperator>(V1)) { 1094 AliasResult Result = aliasGEP(GV1, V1Size, V2, V2Size, V2TBAAInfo, O1, O2); 1095 if (Result != MayAlias) return Result; 1096 } 1097 1098 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) { 1099 std::swap(V1, V2); 1100 std::swap(V1Size, V2Size); 1101 } 1102 if (const PHINode *PN = dyn_cast<PHINode>(V1)) { 1103 AliasResult Result = aliasPHI(PN, V1Size, V1TBAAInfo, 1104 V2, V2Size, V2TBAAInfo); 1105 if (Result != MayAlias) return Result; 1106 } 1107 1108 if (isa<SelectInst>(V2) && !isa<SelectInst>(V1)) { 1109 std::swap(V1, V2); 1110 std::swap(V1Size, V2Size); 1111 } 1112 if (const SelectInst *S1 = dyn_cast<SelectInst>(V1)) { 1113 AliasResult Result = aliasSelect(S1, V1Size, V1TBAAInfo, 1114 V2, V2Size, V2TBAAInfo); 1115 if (Result != MayAlias) return Result; 1116 } 1117 1118 return AliasAnalysis::alias(Location(V1, V1Size, V1TBAAInfo), 1119 Location(V2, V2Size, V2TBAAInfo)); 1120} 1121