BasicAliasAnalysis.cpp revision f0429fd1ca7f6d359a33c20617785d7572a0292d
1//===- BasicAliasAnalysis.cpp - Local 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 default implementation of the Alias Analysis interface 11// that simply implements a few identities (two different globals cannot alias, 12// etc), but otherwise does no analysis. 13// 14//===----------------------------------------------------------------------===// 15 16#include "llvm/Analysis/AliasAnalysis.h" 17#include "llvm/Analysis/CaptureTracking.h" 18#include "llvm/Analysis/MallocHelper.h" 19#include "llvm/Analysis/Passes.h" 20#include "llvm/Constants.h" 21#include "llvm/DerivedTypes.h" 22#include "llvm/Function.h" 23#include "llvm/GlobalVariable.h" 24#include "llvm/Instructions.h" 25#include "llvm/IntrinsicInst.h" 26#include "llvm/LLVMContext.h" 27#include "llvm/Operator.h" 28#include "llvm/Pass.h" 29#include "llvm/Target/TargetData.h" 30#include "llvm/ADT/SmallSet.h" 31#include "llvm/ADT/SmallVector.h" 32#include "llvm/ADT/STLExtras.h" 33#include "llvm/Support/Compiler.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 43static const GEPOperator *isGEP(const Value *V) { 44 return dyn_cast<GEPOperator>(V); 45} 46 47static const Value *GetGEPOperands(const Value *V, 48 SmallVector<Value*, 16> &GEPOps) { 49 assert(GEPOps.empty() && "Expect empty list to populate!"); 50 GEPOps.insert(GEPOps.end(), cast<User>(V)->op_begin()+1, 51 cast<User>(V)->op_end()); 52 53 // Accumulate all of the chained indexes into the operand array 54 V = cast<User>(V)->getOperand(0); 55 56 while (const User *G = isGEP(V)) { 57 if (!isa<Constant>(GEPOps[0]) || isa<GlobalValue>(GEPOps[0]) || 58 !cast<Constant>(GEPOps[0])->isNullValue()) 59 break; // Don't handle folding arbitrary pointer offsets yet... 60 GEPOps.erase(GEPOps.begin()); // Drop the zero index 61 GEPOps.insert(GEPOps.begin(), G->op_begin()+1, G->op_end()); 62 V = G->getOperand(0); 63 } 64 return V; 65} 66 67/// isKnownNonNull - Return true if we know that the specified value is never 68/// null. 69static bool isKnownNonNull(const Value *V) { 70 // Alloca never returns null, malloc might. 71 if (isa<AllocaInst>(V)) return true; 72 73 // A byval argument is never null. 74 if (const Argument *A = dyn_cast<Argument>(V)) 75 return A->hasByValAttr(); 76 77 // Global values are not null unless extern weak. 78 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) 79 return !GV->hasExternalWeakLinkage(); 80 return false; 81} 82 83/// isNonEscapingLocalObject - Return true if the pointer is to a function-local 84/// object that never escapes from the function. 85static bool isNonEscapingLocalObject(const Value *V) { 86 // If this is a local allocation, check to see if it escapes. 87 if (isa<AllocationInst>(V) || isNoAliasCall(V)) 88 return !PointerMayBeCaptured(V, false); 89 90 // If this is an argument that corresponds to a byval or noalias argument, 91 // then it has not escaped before entering the function. Check if it escapes 92 // inside the function. 93 if (const Argument *A = dyn_cast<Argument>(V)) 94 if (A->hasByValAttr() || A->hasNoAliasAttr()) { 95 // Don't bother analyzing arguments already known not to escape. 96 if (A->hasNoCaptureAttr()) 97 return true; 98 return !PointerMayBeCaptured(V, false); 99 } 100 return false; 101} 102 103 104/// isObjectSmallerThan - Return true if we can prove that the object specified 105/// by V is smaller than Size. 106static bool isObjectSmallerThan(const Value *V, unsigned Size, 107 LLVMContext &Context, const TargetData &TD) { 108 const Type *AccessTy; 109 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) { 110 AccessTy = GV->getType()->getElementType(); 111 } else if (const AllocationInst *AI = dyn_cast<AllocationInst>(V)) { 112 if (!AI->isArrayAllocation()) 113 AccessTy = AI->getType()->getElementType(); 114 else 115 return false; 116 } else if (const CallInst* CI = extractMallocCall(V)) { 117 if (!isArrayMalloc(V, Context, &TD)) 118 // The size is the argument to the malloc call. 119 if (const ConstantInt* C = dyn_cast<ConstantInt>(CI->getOperand(1))) 120 return (C->getZExtValue() < Size); 121 return false; 122 } else if (const Argument *A = dyn_cast<Argument>(V)) { 123 if (A->hasByValAttr()) 124 AccessTy = cast<PointerType>(A->getType())->getElementType(); 125 else 126 return false; 127 } else { 128 return false; 129 } 130 131 if (AccessTy->isSized()) 132 return TD.getTypeAllocSize(AccessTy) < Size; 133 return false; 134} 135 136//===----------------------------------------------------------------------===// 137// NoAA Pass 138//===----------------------------------------------------------------------===// 139 140namespace { 141 /// NoAA - This class implements the -no-aa pass, which always returns "I 142 /// don't know" for alias queries. NoAA is unlike other alias analysis 143 /// implementations, in that it does not chain to a previous analysis. As 144 /// such it doesn't follow many of the rules that other alias analyses must. 145 /// 146 struct VISIBILITY_HIDDEN NoAA : public ImmutablePass, public AliasAnalysis { 147 static char ID; // Class identification, replacement for typeinfo 148 NoAA() : ImmutablePass(&ID) {} 149 explicit NoAA(void *PID) : ImmutablePass(PID) { } 150 151 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 152 } 153 154 virtual void initializePass() { 155 TD = getAnalysisIfAvailable<TargetData>(); 156 } 157 158 virtual AliasResult alias(const Value *V1, unsigned V1Size, 159 const Value *V2, unsigned V2Size) { 160 return MayAlias; 161 } 162 163 virtual void getArgumentAccesses(Function *F, CallSite CS, 164 std::vector<PointerAccessInfo> &Info) { 165 llvm_unreachable("This method may not be called on this function!"); 166 } 167 168 virtual void getMustAliases(Value *P, std::vector<Value*> &RetVals) { } 169 virtual bool pointsToConstantMemory(const Value *P) { return false; } 170 virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) { 171 return ModRef; 172 } 173 virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) { 174 return ModRef; 175 } 176 virtual bool hasNoModRefInfoForCalls() const { return true; } 177 178 virtual void deleteValue(Value *V) {} 179 virtual void copyValue(Value *From, Value *To) {} 180 }; 181} // End of anonymous namespace 182 183// Register this pass... 184char NoAA::ID = 0; 185static RegisterPass<NoAA> 186U("no-aa", "No Alias Analysis (always returns 'may' alias)", true, true); 187 188// Declare that we implement the AliasAnalysis interface 189static RegisterAnalysisGroup<AliasAnalysis> V(U); 190 191ImmutablePass *llvm::createNoAAPass() { return new NoAA(); } 192 193//===----------------------------------------------------------------------===// 194// BasicAA Pass 195//===----------------------------------------------------------------------===// 196 197namespace { 198 /// BasicAliasAnalysis - This is the default alias analysis implementation. 199 /// Because it doesn't chain to a previous alias analysis (like -no-aa), it 200 /// derives from the NoAA class. 201 struct VISIBILITY_HIDDEN BasicAliasAnalysis : public NoAA { 202 static char ID; // Class identification, replacement for typeinfo 203 BasicAliasAnalysis() : NoAA(&ID) {} 204 AliasResult alias(const Value *V1, unsigned V1Size, 205 const Value *V2, unsigned V2Size) { 206 assert(VisitedPHIs.empty() && "VisitedPHIs must be cleared after use!"); 207 AliasResult Alias = aliasCheck(V1, V1Size, V2, V2Size); 208 VisitedPHIs.clear(); 209 return Alias; 210 } 211 212 ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size); 213 ModRefResult getModRefInfo(CallSite CS1, CallSite CS2); 214 215 /// hasNoModRefInfoForCalls - We can provide mod/ref information against 216 /// non-escaping allocations. 217 virtual bool hasNoModRefInfoForCalls() const { return false; } 218 219 /// pointsToConstantMemory - Chase pointers until we find a (constant 220 /// global) or not. 221 bool pointsToConstantMemory(const Value *P); 222 223 private: 224 // VisitedPHIs - Track PHI nodes visited by a aliasCheck() call. 225 SmallSet<const PHINode*, 16> VisitedPHIs; 226 227 // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction 228 // against another. 229 AliasResult aliasGEP(const Value *V1, unsigned V1Size, 230 const Value *V2, unsigned V2Size); 231 232 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction 233 // against another. 234 AliasResult aliasPHI(const PHINode *PN, unsigned PNSize, 235 const Value *V2, unsigned V2Size); 236 237 AliasResult aliasCheck(const Value *V1, unsigned V1Size, 238 const Value *V2, unsigned V2Size); 239 240 // CheckGEPInstructions - Check two GEP instructions with known 241 // must-aliasing base pointers. This checks to see if the index expressions 242 // preclude the pointers from aliasing... 243 AliasResult 244 CheckGEPInstructions(const Type* BasePtr1Ty, 245 Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1Size, 246 const Type *BasePtr2Ty, 247 Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size); 248 }; 249} // End of anonymous namespace 250 251// Register this pass... 252char BasicAliasAnalysis::ID = 0; 253static RegisterPass<BasicAliasAnalysis> 254X("basicaa", "Basic Alias Analysis (default AA impl)", false, true); 255 256// Declare that we implement the AliasAnalysis interface 257static RegisterAnalysisGroup<AliasAnalysis, true> Y(X); 258 259ImmutablePass *llvm::createBasicAliasAnalysisPass() { 260 return new BasicAliasAnalysis(); 261} 262 263 264/// pointsToConstantMemory - Chase pointers until we find a (constant 265/// global) or not. 266bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) { 267 if (const GlobalVariable *GV = 268 dyn_cast<GlobalVariable>(P->getUnderlyingObject())) 269 return GV->isConstant(); 270 return false; 271} 272 273 274// getModRefInfo - Check to see if the specified callsite can clobber the 275// specified memory object. Since we only look at local properties of this 276// function, we really can't say much about this query. We do, however, use 277// simple "address taken" analysis on local objects. 278// 279AliasAnalysis::ModRefResult 280BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) { 281 if (!isa<Constant>(P)) { 282 const Value *Object = P->getUnderlyingObject(); 283 284 // If this is a tail call and P points to a stack location, we know that 285 // the tail call cannot access or modify the local stack. 286 // We cannot exclude byval arguments here; these belong to the caller of 287 // the current function not to the current function, and a tail callee 288 // may reference them. 289 if (isa<AllocaInst>(Object)) 290 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) 291 if (CI->isTailCall()) 292 return NoModRef; 293 294 // If the pointer is to a locally allocated object that does not escape, 295 // then the call can not mod/ref the pointer unless the call takes the 296 // argument without capturing it. 297 if (isNonEscapingLocalObject(Object) && CS.getInstruction() != Object) { 298 bool passedAsArg = false; 299 // TODO: Eventually only check 'nocapture' arguments. 300 for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end(); 301 CI != CE; ++CI) 302 if (isa<PointerType>((*CI)->getType()) && 303 alias(cast<Value>(CI), ~0U, P, ~0U) != NoAlias) 304 passedAsArg = true; 305 306 if (!passedAsArg) 307 return NoModRef; 308 } 309 310 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) { 311 switch (II->getIntrinsicID()) { 312 default: break; 313 case Intrinsic::atomic_cmp_swap: 314 case Intrinsic::atomic_swap: 315 case Intrinsic::atomic_load_add: 316 case Intrinsic::atomic_load_sub: 317 case Intrinsic::atomic_load_and: 318 case Intrinsic::atomic_load_nand: 319 case Intrinsic::atomic_load_or: 320 case Intrinsic::atomic_load_xor: 321 case Intrinsic::atomic_load_max: 322 case Intrinsic::atomic_load_min: 323 case Intrinsic::atomic_load_umax: 324 case Intrinsic::atomic_load_umin: 325 if (alias(II->getOperand(1), Size, P, Size) == NoAlias) 326 return NoModRef; 327 break; 328 } 329 } 330 } 331 332 // The AliasAnalysis base class has some smarts, lets use them. 333 return AliasAnalysis::getModRefInfo(CS, P, Size); 334} 335 336 337AliasAnalysis::ModRefResult 338BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) { 339 // If CS1 or CS2 are readnone, they don't interact. 340 ModRefBehavior CS1B = AliasAnalysis::getModRefBehavior(CS1); 341 if (CS1B == DoesNotAccessMemory) return NoModRef; 342 343 ModRefBehavior CS2B = AliasAnalysis::getModRefBehavior(CS2); 344 if (CS2B == DoesNotAccessMemory) return NoModRef; 345 346 // If they both only read from memory, just return ref. 347 if (CS1B == OnlyReadsMemory && CS2B == OnlyReadsMemory) 348 return Ref; 349 350 // Otherwise, fall back to NoAA (mod+ref). 351 return NoAA::getModRefInfo(CS1, CS2); 352} 353 354// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction 355// against another. 356// 357AliasAnalysis::AliasResult 358BasicAliasAnalysis::aliasGEP(const Value *V1, unsigned V1Size, 359 const Value *V2, unsigned V2Size) { 360 // If we have two gep instructions with must-alias'ing base pointers, figure 361 // out if the indexes to the GEP tell us anything about the derived pointer. 362 // Note that we also handle chains of getelementptr instructions as well as 363 // constant expression getelementptrs here. 364 // 365 if (isGEP(V1) && isGEP(V2)) { 366 const User *GEP1 = cast<User>(V1); 367 const User *GEP2 = cast<User>(V2); 368 369 // If V1 and V2 are identical GEPs, just recurse down on both of them. 370 // This allows us to analyze things like: 371 // P = gep A, 0, i, 1 372 // Q = gep B, 0, i, 1 373 // by just analyzing A and B. This is even safe for variable indices. 374 if (GEP1->getType() == GEP2->getType() && 375 GEP1->getNumOperands() == GEP2->getNumOperands() && 376 GEP1->getOperand(0)->getType() == GEP2->getOperand(0)->getType() && 377 // All operands are the same, ignoring the base. 378 std::equal(GEP1->op_begin()+1, GEP1->op_end(), GEP2->op_begin()+1)) 379 return aliasCheck(GEP1->getOperand(0), V1Size, 380 GEP2->getOperand(0), V2Size); 381 382 // Drill down into the first non-gep value, to test for must-aliasing of 383 // the base pointers. 384 while (isGEP(GEP1->getOperand(0)) && 385 GEP1->getOperand(1) == 386 Constant::getNullValue(GEP1->getOperand(1)->getType())) 387 GEP1 = cast<User>(GEP1->getOperand(0)); 388 const Value *BasePtr1 = GEP1->getOperand(0); 389 390 while (isGEP(GEP2->getOperand(0)) && 391 GEP2->getOperand(1) == 392 Constant::getNullValue(GEP2->getOperand(1)->getType())) 393 GEP2 = cast<User>(GEP2->getOperand(0)); 394 const Value *BasePtr2 = GEP2->getOperand(0); 395 396 // Do the base pointers alias? 397 AliasResult BaseAlias = aliasCheck(BasePtr1, ~0U, BasePtr2, ~0U); 398 if (BaseAlias == NoAlias) return NoAlias; 399 if (BaseAlias == MustAlias) { 400 // If the base pointers alias each other exactly, check to see if we can 401 // figure out anything about the resultant pointers, to try to prove 402 // non-aliasing. 403 404 // Collect all of the chained GEP operands together into one simple place 405 SmallVector<Value*, 16> GEP1Ops, GEP2Ops; 406 BasePtr1 = GetGEPOperands(V1, GEP1Ops); 407 BasePtr2 = GetGEPOperands(V2, GEP2Ops); 408 409 // If GetGEPOperands were able to fold to the same must-aliased pointer, 410 // do the comparison. 411 if (BasePtr1 == BasePtr2) { 412 AliasResult GAlias = 413 CheckGEPInstructions(BasePtr1->getType(), 414 &GEP1Ops[0], GEP1Ops.size(), V1Size, 415 BasePtr2->getType(), 416 &GEP2Ops[0], GEP2Ops.size(), V2Size); 417 if (GAlias != MayAlias) 418 return GAlias; 419 } 420 } 421 } 422 423 // Check to see if these two pointers are related by a getelementptr 424 // instruction. If one pointer is a GEP with a non-zero index of the other 425 // pointer, we know they cannot alias. 426 // 427 if (V1Size == ~0U || V2Size == ~0U) 428 return MayAlias; 429 430 SmallVector<Value*, 16> GEPOperands; 431 const Value *BasePtr = GetGEPOperands(V1, GEPOperands); 432 433 AliasResult R = aliasCheck(BasePtr, ~0U, V2, V2Size); 434 if (R != MustAlias) 435 // If V2 may alias GEP base pointer, conservatively returns MayAlias. 436 // If V2 is known not to alias GEP base pointer, then the two values 437 // cannot alias per GEP semantics: "A pointer value formed from a 438 // getelementptr instruction is associated with the addresses associated 439 // with the first operand of the getelementptr". 440 return R; 441 442 // If there is at least one non-zero constant index, we know they cannot 443 // alias. 444 bool ConstantFound = false; 445 bool AllZerosFound = true; 446 for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i) 447 if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) { 448 if (!C->isNullValue()) { 449 ConstantFound = true; 450 AllZerosFound = false; 451 break; 452 } 453 } else { 454 AllZerosFound = false; 455 } 456 457 // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases 458 // the ptr, the end result is a must alias also. 459 if (AllZerosFound) 460 return MustAlias; 461 462 if (ConstantFound) { 463 if (V2Size <= 1 && V1Size <= 1) // Just pointer check? 464 return NoAlias; 465 466 // Otherwise we have to check to see that the distance is more than 467 // the size of the argument... build an index vector that is equal to 468 // the arguments provided, except substitute 0's for any variable 469 // indexes we find... 470 if (TD && 471 cast<PointerType>(BasePtr->getType())->getElementType()->isSized()) { 472 for (unsigned i = 0; i != GEPOperands.size(); ++i) 473 if (!isa<ConstantInt>(GEPOperands[i])) 474 GEPOperands[i] = Constant::getNullValue(GEPOperands[i]->getType()); 475 int64_t Offset = TD->getIndexedOffset(BasePtr->getType(), 476 &GEPOperands[0], 477 GEPOperands.size()); 478 479 if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size) 480 return NoAlias; 481 } 482 } 483 484 return MayAlias; 485} 486 487// aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction 488// against another. 489AliasAnalysis::AliasResult 490BasicAliasAnalysis::aliasPHI(const PHINode *PN, unsigned PNSize, 491 const Value *V2, unsigned V2Size) { 492 // The PHI node has already been visited, avoid recursion any further. 493 if (!VisitedPHIs.insert(PN)) 494 return MayAlias; 495 496 SmallSet<Value*, 4> UniqueSrc; 497 SmallVector<Value*, 4> V1Srcs; 498 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 499 Value *PV1 = PN->getIncomingValue(i); 500 if (isa<PHINode>(PV1)) 501 // If any of the source itself is a PHI, return MayAlias conservatively 502 // to avoid compile time explosion. The worst possible case is if both 503 // sides are PHI nodes. In which case, this is O(m x n) time where 'm' 504 // and 'n' are the number of PHI sources. 505 return MayAlias; 506 if (UniqueSrc.insert(PV1)) 507 V1Srcs.push_back(PV1); 508 } 509 510 AliasResult Alias = aliasCheck(V1Srcs[0], PNSize, V2, V2Size); 511 // Early exit if the check of the first PHI source against V2 is MayAlias. 512 // Other results are not possible. 513 if (Alias == MayAlias) 514 return MayAlias; 515 516 // If all sources of the PHI node NoAlias or MustAlias V2, then returns 517 // NoAlias / MustAlias. Otherwise, returns MayAlias. 518 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) { 519 Value *V = V1Srcs[i]; 520 AliasResult ThisAlias = aliasCheck(V, PNSize, V2, V2Size); 521 if (ThisAlias != Alias || ThisAlias == MayAlias) 522 return MayAlias; 523 } 524 525 return Alias; 526} 527 528// aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases, 529// such as array references. 530// 531AliasAnalysis::AliasResult 532BasicAliasAnalysis::aliasCheck(const Value *V1, unsigned V1Size, 533 const Value *V2, unsigned V2Size) { 534 // Strip off any casts if they exist. 535 V1 = V1->stripPointerCasts(); 536 V2 = V2->stripPointerCasts(); 537 538 // Are we checking for alias of the same value? 539 if (V1 == V2) return MustAlias; 540 541 if (!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType())) 542 return NoAlias; // Scalars cannot alias each other 543 544 // Figure out what objects these things are pointing to if we can. 545 const Value *O1 = V1->getUnderlyingObject(); 546 const Value *O2 = V2->getUnderlyingObject(); 547 548 if (O1 != O2) { 549 // If V1/V2 point to two different objects we know that we have no alias. 550 if (isIdentifiedObject(O1) && isIdentifiedObject(O2)) 551 return NoAlias; 552 553 // Arguments can't alias with local allocations or noalias calls. 554 if ((isa<Argument>(O1) && (isa<AllocationInst>(O2) || isNoAliasCall(O2))) || 555 (isa<Argument>(O2) && (isa<AllocationInst>(O1) || isNoAliasCall(O1)))) 556 return NoAlias; 557 558 // Most objects can't alias null. 559 if ((isa<ConstantPointerNull>(V2) && isKnownNonNull(O1)) || 560 (isa<ConstantPointerNull>(V1) && isKnownNonNull(O2))) 561 return NoAlias; 562 } 563 564 // If the size of one access is larger than the entire object on the other 565 // side, then we know such behavior is undefined and can assume no alias. 566 LLVMContext &Context = V1->getContext(); 567 if (TD) 568 if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, Context, *TD)) || 569 (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, Context, *TD))) 570 return NoAlias; 571 572 // If one pointer is the result of a call/invoke and the other is a 573 // non-escaping local object, then we know the object couldn't escape to a 574 // point where the call could return it. 575 if ((isa<CallInst>(O1) || isa<InvokeInst>(O1)) && 576 isNonEscapingLocalObject(O2) && O1 != O2) 577 return NoAlias; 578 if ((isa<CallInst>(O2) || isa<InvokeInst>(O2)) && 579 isNonEscapingLocalObject(O1) && O1 != O2) 580 return NoAlias; 581 582 if (!isGEP(V1) && isGEP(V2)) { 583 std::swap(V1, V2); 584 std::swap(V1Size, V2Size); 585 } 586 if (isGEP(V1)) 587 return aliasGEP(V1, V1Size, V2, V2Size); 588 589 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) { 590 std::swap(V1, V2); 591 std::swap(V1Size, V2Size); 592 } 593 if (const PHINode *PN = dyn_cast<PHINode>(V1)) 594 return aliasPHI(PN, V1Size, V2, V2Size); 595 596 return MayAlias; 597} 598 599// This function is used to determine if the indices of two GEP instructions are 600// equal. V1 and V2 are the indices. 601static bool IndexOperandsEqual(Value *V1, Value *V2, LLVMContext &Context) { 602 if (V1->getType() == V2->getType()) 603 return V1 == V2; 604 if (Constant *C1 = dyn_cast<Constant>(V1)) 605 if (Constant *C2 = dyn_cast<Constant>(V2)) { 606 // Sign extend the constants to long types, if necessary 607 if (C1->getType() != Type::getInt64Ty(Context)) 608 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(Context)); 609 if (C2->getType() != Type::getInt64Ty(Context)) 610 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(Context)); 611 return C1 == C2; 612 } 613 return false; 614} 615 616/// CheckGEPInstructions - Check two GEP instructions with known must-aliasing 617/// base pointers. This checks to see if the index expressions preclude the 618/// pointers from aliasing... 619AliasAnalysis::AliasResult 620BasicAliasAnalysis::CheckGEPInstructions( 621 const Type* BasePtr1Ty, Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1S, 622 const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2S) { 623 // We currently can't handle the case when the base pointers have different 624 // primitive types. Since this is uncommon anyway, we are happy being 625 // extremely conservative. 626 if (BasePtr1Ty != BasePtr2Ty) 627 return MayAlias; 628 629 const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty); 630 631 LLVMContext &Context = GEPPointerTy->getContext(); 632 633 // Find the (possibly empty) initial sequence of equal values... which are not 634 // necessarily constants. 635 unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops; 636 unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands); 637 unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands); 638 unsigned UnequalOper = 0; 639 while (UnequalOper != MinOperands && 640 IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper], 641 Context)) { 642 // Advance through the type as we go... 643 ++UnequalOper; 644 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty)) 645 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]); 646 else { 647 // If all operands equal each other, then the derived pointers must 648 // alias each other... 649 BasePtr1Ty = 0; 650 assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands && 651 "Ran out of type nesting, but not out of operands?"); 652 return MustAlias; 653 } 654 } 655 656 // If we have seen all constant operands, and run out of indexes on one of the 657 // getelementptrs, check to see if the tail of the leftover one is all zeros. 658 // If so, return mustalias. 659 if (UnequalOper == MinOperands) { 660 if (NumGEP1Ops < NumGEP2Ops) { 661 std::swap(GEP1Ops, GEP2Ops); 662 std::swap(NumGEP1Ops, NumGEP2Ops); 663 } 664 665 bool AllAreZeros = true; 666 for (unsigned i = UnequalOper; i != MaxOperands; ++i) 667 if (!isa<Constant>(GEP1Ops[i]) || 668 !cast<Constant>(GEP1Ops[i])->isNullValue()) { 669 AllAreZeros = false; 670 break; 671 } 672 if (AllAreZeros) return MustAlias; 673 } 674 675 676 // So now we know that the indexes derived from the base pointers, 677 // which are known to alias, are different. We can still determine a 678 // no-alias result if there are differing constant pairs in the index 679 // chain. For example: 680 // A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S)) 681 // 682 // We have to be careful here about array accesses. In particular, consider: 683 // A[1][0] vs A[0][i] 684 // In this case, we don't *know* that the array will be accessed in bounds: 685 // the index could even be negative. Because of this, we have to 686 // conservatively *give up* and return may alias. We disregard differing 687 // array subscripts that are followed by a variable index without going 688 // through a struct. 689 // 690 unsigned SizeMax = std::max(G1S, G2S); 691 if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work. 692 693 // Scan for the first operand that is constant and unequal in the 694 // two getelementptrs... 695 unsigned FirstConstantOper = UnequalOper; 696 for (; FirstConstantOper != MinOperands; ++FirstConstantOper) { 697 const Value *G1Oper = GEP1Ops[FirstConstantOper]; 698 const Value *G2Oper = GEP2Ops[FirstConstantOper]; 699 700 if (G1Oper != G2Oper) // Found non-equal constant indexes... 701 if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper))) 702 if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){ 703 if (G1OC->getType() != G2OC->getType()) { 704 // Sign extend both operands to long. 705 if (G1OC->getType() != Type::getInt64Ty(Context)) 706 G1OC = ConstantExpr::getSExt(G1OC, Type::getInt64Ty(Context)); 707 if (G2OC->getType() != Type::getInt64Ty(Context)) 708 G2OC = ConstantExpr::getSExt(G2OC, Type::getInt64Ty(Context)); 709 GEP1Ops[FirstConstantOper] = G1OC; 710 GEP2Ops[FirstConstantOper] = G2OC; 711 } 712 713 if (G1OC != G2OC) { 714 // Handle the "be careful" case above: if this is an array/vector 715 // subscript, scan for a subsequent variable array index. 716 if (const SequentialType *STy = 717 dyn_cast<SequentialType>(BasePtr1Ty)) { 718 const Type *NextTy = STy; 719 bool isBadCase = false; 720 721 for (unsigned Idx = FirstConstantOper; 722 Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) { 723 const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx]; 724 if (!isa<Constant>(V1) || !isa<Constant>(V2)) { 725 isBadCase = true; 726 break; 727 } 728 // If the array is indexed beyond the bounds of the static type 729 // at this level, it will also fall into the "be careful" case. 730 // It would theoretically be possible to analyze these cases, 731 // but for now just be conservatively correct. 732 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 733 if (cast<ConstantInt>(G1OC)->getZExtValue() >= 734 ATy->getNumElements() || 735 cast<ConstantInt>(G2OC)->getZExtValue() >= 736 ATy->getNumElements()) { 737 isBadCase = true; 738 break; 739 } 740 if (const VectorType *VTy = dyn_cast<VectorType>(STy)) 741 if (cast<ConstantInt>(G1OC)->getZExtValue() >= 742 VTy->getNumElements() || 743 cast<ConstantInt>(G2OC)->getZExtValue() >= 744 VTy->getNumElements()) { 745 isBadCase = true; 746 break; 747 } 748 STy = cast<SequentialType>(NextTy); 749 NextTy = cast<SequentialType>(NextTy)->getElementType(); 750 } 751 752 if (isBadCase) G1OC = 0; 753 } 754 755 // Make sure they are comparable (ie, not constant expressions), and 756 // make sure the GEP with the smaller leading constant is GEP1. 757 if (G1OC) { 758 Constant *Compare = ConstantExpr::getICmp(ICmpInst::ICMP_SGT, 759 G1OC, G2OC); 760 if (ConstantInt *CV = dyn_cast<ConstantInt>(Compare)) { 761 if (CV->getZExtValue()) { // If they are comparable and G2 > G1 762 std::swap(GEP1Ops, GEP2Ops); // Make GEP1 < GEP2 763 std::swap(NumGEP1Ops, NumGEP2Ops); 764 } 765 break; 766 } 767 } 768 } 769 } 770 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper); 771 } 772 773 // No shared constant operands, and we ran out of common operands. At this 774 // point, the GEP instructions have run through all of their operands, and we 775 // haven't found evidence that there are any deltas between the GEP's. 776 // However, one GEP may have more operands than the other. If this is the 777 // case, there may still be hope. Check this now. 778 if (FirstConstantOper == MinOperands) { 779 // Without TargetData, we won't know what the offsets are. 780 if (!TD) 781 return MayAlias; 782 783 // Make GEP1Ops be the longer one if there is a longer one. 784 if (NumGEP1Ops < NumGEP2Ops) { 785 std::swap(GEP1Ops, GEP2Ops); 786 std::swap(NumGEP1Ops, NumGEP2Ops); 787 } 788 789 // Is there anything to check? 790 if (NumGEP1Ops > MinOperands) { 791 for (unsigned i = FirstConstantOper; i != MaxOperands; ++i) 792 if (isa<ConstantInt>(GEP1Ops[i]) && 793 !cast<ConstantInt>(GEP1Ops[i])->isZero()) { 794 // Yup, there's a constant in the tail. Set all variables to 795 // constants in the GEP instruction to make it suitable for 796 // TargetData::getIndexedOffset. 797 for (i = 0; i != MaxOperands; ++i) 798 if (!isa<ConstantInt>(GEP1Ops[i])) 799 GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType()); 800 // Okay, now get the offset. This is the relative offset for the full 801 // instruction. 802 int64_t Offset1 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops, 803 NumGEP1Ops); 804 805 // Now check without any constants at the end. 806 int64_t Offset2 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops, 807 MinOperands); 808 809 // Make sure we compare the absolute difference. 810 if (Offset1 > Offset2) 811 std::swap(Offset1, Offset2); 812 813 // If the tail provided a bit enough offset, return noalias! 814 if ((uint64_t)(Offset2-Offset1) >= SizeMax) 815 return NoAlias; 816 // Otherwise break - we don't look for another constant in the tail. 817 break; 818 } 819 } 820 821 // Couldn't find anything useful. 822 return MayAlias; 823 } 824 825 // If there are non-equal constants arguments, then we can figure 826 // out a minimum known delta between the two index expressions... at 827 // this point we know that the first constant index of GEP1 is less 828 // than the first constant index of GEP2. 829 830 // Advance BasePtr[12]Ty over this first differing constant operand. 831 BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)-> 832 getTypeAtIndex(GEP2Ops[FirstConstantOper]); 833 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)-> 834 getTypeAtIndex(GEP1Ops[FirstConstantOper]); 835 836 // We are going to be using TargetData::getIndexedOffset to determine the 837 // offset that each of the GEP's is reaching. To do this, we have to convert 838 // all variable references to constant references. To do this, we convert the 839 // initial sequence of array subscripts into constant zeros to start with. 840 const Type *ZeroIdxTy = GEPPointerTy; 841 for (unsigned i = 0; i != FirstConstantOper; ++i) { 842 if (!isa<StructType>(ZeroIdxTy)) 843 GEP1Ops[i] = GEP2Ops[i] = 844 Constant::getNullValue(Type::getInt32Ty(Context)); 845 846 if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy)) 847 ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]); 848 } 849 850 // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok 851 852 // Loop over the rest of the operands... 853 for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) { 854 const Value *Op1 = i < NumGEP1Ops ? GEP1Ops[i] : 0; 855 const Value *Op2 = i < NumGEP2Ops ? GEP2Ops[i] : 0; 856 // If they are equal, use a zero index... 857 if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) { 858 if (!isa<ConstantInt>(Op1)) 859 GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType()); 860 // Otherwise, just keep the constants we have. 861 } else { 862 if (Op1) { 863 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 864 // If this is an array index, make sure the array element is in range. 865 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) { 866 if (Op1C->getZExtValue() >= AT->getNumElements()) 867 return MayAlias; // Be conservative with out-of-range accesses 868 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) { 869 if (Op1C->getZExtValue() >= VT->getNumElements()) 870 return MayAlias; // Be conservative with out-of-range accesses 871 } 872 873 } else { 874 // GEP1 is known to produce a value less than GEP2. To be 875 // conservatively correct, we must assume the largest possible 876 // constant is used in this position. This cannot be the initial 877 // index to the GEP instructions (because we know we have at least one 878 // element before this one with the different constant arguments), so 879 // we know that the current index must be into either a struct or 880 // array. Because we know it's not constant, this cannot be a 881 // structure index. Because of this, we can calculate the maximum 882 // value possible. 883 // 884 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) 885 GEP1Ops[i] = 886 ConstantInt::get(Type::getInt64Ty(Context), 887 AT->getNumElements()-1); 888 else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) 889 GEP1Ops[i] = 890 ConstantInt::get(Type::getInt64Ty(Context), 891 VT->getNumElements()-1); 892 } 893 } 894 895 if (Op2) { 896 if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) { 897 // If this is an array index, make sure the array element is in range. 898 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr2Ty)) { 899 if (Op2C->getZExtValue() >= AT->getNumElements()) 900 return MayAlias; // Be conservative with out-of-range accesses 901 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr2Ty)) { 902 if (Op2C->getZExtValue() >= VT->getNumElements()) 903 return MayAlias; // Be conservative with out-of-range accesses 904 } 905 } else { // Conservatively assume the minimum value for this index 906 GEP2Ops[i] = Constant::getNullValue(Op2->getType()); 907 } 908 } 909 } 910 911 if (BasePtr1Ty && Op1) { 912 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty)) 913 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]); 914 else 915 BasePtr1Ty = 0; 916 } 917 918 if (BasePtr2Ty && Op2) { 919 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty)) 920 BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]); 921 else 922 BasePtr2Ty = 0; 923 } 924 } 925 926 if (TD && GEPPointerTy->getElementType()->isSized()) { 927 int64_t Offset1 = 928 TD->getIndexedOffset(GEPPointerTy, GEP1Ops, NumGEP1Ops); 929 int64_t Offset2 = 930 TD->getIndexedOffset(GEPPointerTy, GEP2Ops, NumGEP2Ops); 931 assert(Offset1 != Offset2 && 932 "There is at least one different constant here!"); 933 934 // Make sure we compare the absolute difference. 935 if (Offset1 > Offset2) 936 std::swap(Offset1, Offset2); 937 938 if ((uint64_t)(Offset2-Offset1) >= SizeMax) { 939 //cerr << "Determined that these two GEP's don't alias [" 940 // << SizeMax << " bytes]: \n" << *GEP1 << *GEP2; 941 return NoAlias; 942 } 943 } 944 return MayAlias; 945} 946 947// Make sure that anything that uses AliasAnalysis pulls in this file... 948DEFINING_FILE_FOR(BasicAliasAnalysis) 949