BasicAliasAnalysis.cpp revision 681a33e26dd3222477f13520b94e7417bff59e32
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 VisitedPHIs.clear(); 207 return aliasCheck(V1, V1Size, V2, V2Size); 208 } 209 210 ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size); 211 ModRefResult getModRefInfo(CallSite CS1, CallSite CS2); 212 213 /// hasNoModRefInfoForCalls - We can provide mod/ref information against 214 /// non-escaping allocations. 215 virtual bool hasNoModRefInfoForCalls() const { return false; } 216 217 /// pointsToConstantMemory - Chase pointers until we find a (constant 218 /// global) or not. 219 bool pointsToConstantMemory(const Value *P); 220 221 private: 222 // VisitedPHIs - Track PHI nodes visited by a aliasCheck() call. 223 SmallSet<const PHINode*, 16> VisitedPHIs; 224 225 // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction 226 // against another. 227 AliasResult aliasGEP(const Value *V1, unsigned V1Size, 228 const Value *V2, unsigned V2Size); 229 230 // aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction 231 // against another. 232 AliasResult aliasPHI(const PHINode *PN, unsigned PNSize, 233 const Value *V2, unsigned V2Size); 234 235 AliasResult aliasCheck(const Value *V1, unsigned V1Size, 236 const Value *V2, unsigned V2Size); 237 238 // CheckGEPInstructions - Check two GEP instructions with known 239 // must-aliasing base pointers. This checks to see if the index expressions 240 // preclude the pointers from aliasing... 241 AliasResult 242 CheckGEPInstructions(const Type* BasePtr1Ty, 243 Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1Size, 244 const Type *BasePtr2Ty, 245 Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size); 246 }; 247} // End of anonymous namespace 248 249// Register this pass... 250char BasicAliasAnalysis::ID = 0; 251static RegisterPass<BasicAliasAnalysis> 252X("basicaa", "Basic Alias Analysis (default AA impl)", false, true); 253 254// Declare that we implement the AliasAnalysis interface 255static RegisterAnalysisGroup<AliasAnalysis, true> Y(X); 256 257ImmutablePass *llvm::createBasicAliasAnalysisPass() { 258 return new BasicAliasAnalysis(); 259} 260 261 262/// pointsToConstantMemory - Chase pointers until we find a (constant 263/// global) or not. 264bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) { 265 if (const GlobalVariable *GV = 266 dyn_cast<GlobalVariable>(P->getUnderlyingObject())) 267 return GV->isConstant(); 268 return false; 269} 270 271 272// getModRefInfo - Check to see if the specified callsite can clobber the 273// specified memory object. Since we only look at local properties of this 274// function, we really can't say much about this query. We do, however, use 275// simple "address taken" analysis on local objects. 276// 277AliasAnalysis::ModRefResult 278BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) { 279 if (!isa<Constant>(P)) { 280 const Value *Object = P->getUnderlyingObject(); 281 282 // If this is a tail call and P points to a stack location, we know that 283 // the tail call cannot access or modify the local stack. 284 // We cannot exclude byval arguments here; these belong to the caller of 285 // the current function not to the current function, and a tail callee 286 // may reference them. 287 if (isa<AllocaInst>(Object)) 288 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) 289 if (CI->isTailCall()) 290 return NoModRef; 291 292 // If the pointer is to a locally allocated object that does not escape, 293 // then the call can not mod/ref the pointer unless the call takes the 294 // argument without capturing it. 295 if (isNonEscapingLocalObject(Object) && CS.getInstruction() != Object) { 296 bool passedAsArg = false; 297 // TODO: Eventually only check 'nocapture' arguments. 298 for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end(); 299 CI != CE; ++CI) 300 if (isa<PointerType>((*CI)->getType()) && 301 alias(cast<Value>(CI), ~0U, P, ~0U) != NoAlias) 302 passedAsArg = true; 303 304 if (!passedAsArg) 305 return NoModRef; 306 } 307 308 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) { 309 switch (II->getIntrinsicID()) { 310 default: break; 311 case Intrinsic::atomic_cmp_swap: 312 case Intrinsic::atomic_swap: 313 case Intrinsic::atomic_load_add: 314 case Intrinsic::atomic_load_sub: 315 case Intrinsic::atomic_load_and: 316 case Intrinsic::atomic_load_nand: 317 case Intrinsic::atomic_load_or: 318 case Intrinsic::atomic_load_xor: 319 case Intrinsic::atomic_load_max: 320 case Intrinsic::atomic_load_min: 321 case Intrinsic::atomic_load_umax: 322 case Intrinsic::atomic_load_umin: 323 if (alias(II->getOperand(1), Size, P, Size) == NoAlias) 324 return NoModRef; 325 break; 326 } 327 } 328 } 329 330 // The AliasAnalysis base class has some smarts, lets use them. 331 return AliasAnalysis::getModRefInfo(CS, P, Size); 332} 333 334 335AliasAnalysis::ModRefResult 336BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) { 337 // If CS1 or CS2 are readnone, they don't interact. 338 ModRefBehavior CS1B = AliasAnalysis::getModRefBehavior(CS1); 339 if (CS1B == DoesNotAccessMemory) return NoModRef; 340 341 ModRefBehavior CS2B = AliasAnalysis::getModRefBehavior(CS2); 342 if (CS2B == DoesNotAccessMemory) return NoModRef; 343 344 // If they both only read from memory, just return ref. 345 if (CS1B == OnlyReadsMemory && CS2B == OnlyReadsMemory) 346 return Ref; 347 348 // Otherwise, fall back to NoAA (mod+ref). 349 return NoAA::getModRefInfo(CS1, CS2); 350} 351 352// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction 353// against another. 354// 355AliasAnalysis::AliasResult 356BasicAliasAnalysis::aliasGEP(const Value *V1, unsigned V1Size, 357 const Value *V2, unsigned V2Size) { 358 // If we have two gep instructions with must-alias'ing base pointers, figure 359 // out if the indexes to the GEP tell us anything about the derived pointer. 360 // Note that we also handle chains of getelementptr instructions as well as 361 // constant expression getelementptrs here. 362 // 363 if (isGEP(V1) && isGEP(V2)) { 364 const User *GEP1 = cast<User>(V1); 365 const User *GEP2 = cast<User>(V2); 366 367 // If V1 and V2 are identical GEPs, just recurse down on both of them. 368 // This allows us to analyze things like: 369 // P = gep A, 0, i, 1 370 // Q = gep B, 0, i, 1 371 // by just analyzing A and B. This is even safe for variable indices. 372 if (GEP1->getType() == GEP2->getType() && 373 GEP1->getNumOperands() == GEP2->getNumOperands() && 374 GEP1->getOperand(0)->getType() == GEP2->getOperand(0)->getType() && 375 // All operands are the same, ignoring the base. 376 std::equal(GEP1->op_begin()+1, GEP1->op_end(), GEP2->op_begin()+1)) 377 return aliasCheck(GEP1->getOperand(0), V1Size, 378 GEP2->getOperand(0), V2Size); 379 380 // Drill down into the first non-gep value, to test for must-aliasing of 381 // the base pointers. 382 while (isGEP(GEP1->getOperand(0)) && 383 GEP1->getOperand(1) == 384 Constant::getNullValue(GEP1->getOperand(1)->getType())) 385 GEP1 = cast<User>(GEP1->getOperand(0)); 386 const Value *BasePtr1 = GEP1->getOperand(0); 387 388 while (isGEP(GEP2->getOperand(0)) && 389 GEP2->getOperand(1) == 390 Constant::getNullValue(GEP2->getOperand(1)->getType())) 391 GEP2 = cast<User>(GEP2->getOperand(0)); 392 const Value *BasePtr2 = GEP2->getOperand(0); 393 394 // Do the base pointers alias? 395 AliasResult BaseAlias = aliasCheck(BasePtr1, ~0U, BasePtr2, ~0U); 396 if (BaseAlias == NoAlias) return NoAlias; 397 if (BaseAlias == MustAlias) { 398 // If the base pointers alias each other exactly, check to see if we can 399 // figure out anything about the resultant pointers, to try to prove 400 // non-aliasing. 401 402 // Collect all of the chained GEP operands together into one simple place 403 SmallVector<Value*, 16> GEP1Ops, GEP2Ops; 404 BasePtr1 = GetGEPOperands(V1, GEP1Ops); 405 BasePtr2 = GetGEPOperands(V2, GEP2Ops); 406 407 // If GetGEPOperands were able to fold to the same must-aliased pointer, 408 // do the comparison. 409 if (BasePtr1 == BasePtr2) { 410 AliasResult GAlias = 411 CheckGEPInstructions(BasePtr1->getType(), 412 &GEP1Ops[0], GEP1Ops.size(), V1Size, 413 BasePtr2->getType(), 414 &GEP2Ops[0], GEP2Ops.size(), V2Size); 415 if (GAlias != MayAlias) 416 return GAlias; 417 } 418 } 419 } 420 421 // Check to see if these two pointers are related by a getelementptr 422 // instruction. If one pointer is a GEP with a non-zero index of the other 423 // pointer, we know they cannot alias. 424 // 425 if (V1Size == ~0U || V2Size == ~0U) 426 return MayAlias; 427 428 SmallVector<Value*, 16> GEPOperands; 429 const Value *BasePtr = GetGEPOperands(V1, GEPOperands); 430 431 AliasResult R = aliasCheck(BasePtr, ~0U, V2, V2Size); 432 if (R != MustAlias) 433 // If V2 may alias GEP base pointer, conservatively returns MayAlias. 434 // If V2 is known not to alias GEP base pointer, then the two values 435 // cannot alias per GEP semantics: "A pointer value formed from a 436 // getelementptr instruction is associated with the addresses associated 437 // with the first operand of the getelementptr". 438 return R; 439 440 // If there is at least one non-zero constant index, we know they cannot 441 // alias. 442 bool ConstantFound = false; 443 bool AllZerosFound = true; 444 for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i) 445 if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) { 446 if (!C->isNullValue()) { 447 ConstantFound = true; 448 AllZerosFound = false; 449 break; 450 } 451 } else { 452 AllZerosFound = false; 453 } 454 455 // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases 456 // the ptr, the end result is a must alias also. 457 if (AllZerosFound) 458 return MustAlias; 459 460 if (ConstantFound) { 461 if (V2Size <= 1 && V1Size <= 1) // Just pointer check? 462 return NoAlias; 463 464 // Otherwise we have to check to see that the distance is more than 465 // the size of the argument... build an index vector that is equal to 466 // the arguments provided, except substitute 0's for any variable 467 // indexes we find... 468 if (TD && 469 cast<PointerType>(BasePtr->getType())->getElementType()->isSized()) { 470 for (unsigned i = 0; i != GEPOperands.size(); ++i) 471 if (!isa<ConstantInt>(GEPOperands[i])) 472 GEPOperands[i] = Constant::getNullValue(GEPOperands[i]->getType()); 473 int64_t Offset = TD->getIndexedOffset(BasePtr->getType(), 474 &GEPOperands[0], 475 GEPOperands.size()); 476 477 if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size) 478 return NoAlias; 479 } 480 } 481 482 return MayAlias; 483} 484 485// aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction 486// against another. 487AliasAnalysis::AliasResult 488BasicAliasAnalysis::aliasPHI(const PHINode *PN, unsigned PNSize, 489 const Value *V2, unsigned V2Size) { 490 // The PHI node has already been visited, avoid recursion any further. 491 if (!VisitedPHIs.insert(PN)) 492 return MayAlias; 493 494 SmallSet<Value*, 4> UniqueSrc; 495 SmallVector<Value*, 4> V1Srcs; 496 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 497 Value *PV1 = PN->getIncomingValue(i); 498 if (isa<PHINode>(PV1)) 499 // If any of the source itself is a PHI, return MayAlias conservatively 500 // to avoid compile time explosion. The worst possible case is if both 501 // sides are PHI nodes. In which case, this is O(m x n) time where 'm' 502 // and 'n' are the number of PHI sources. 503 return MayAlias; 504 if (UniqueSrc.insert(PV1)) 505 V1Srcs.push_back(PV1); 506 } 507 508 AliasResult Alias = aliasCheck(V1Srcs[0], PNSize, V2, V2Size); 509 // Early exit if the check of the first PHI source against V2 is MayAlias. 510 // Other results are not possible. 511 if (Alias == MayAlias) 512 return MayAlias; 513 514 // If all sources of the PHI node NoAlias or MustAlias V2, then returns 515 // NoAlias / MustAlias. Otherwise, returns MayAlias. 516 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) { 517 Value *V = V1Srcs[i]; 518 AliasResult ThisAlias = aliasCheck(V, PNSize, V2, V2Size); 519 if (ThisAlias != Alias || ThisAlias == MayAlias) 520 return MayAlias; 521 } 522 523 return Alias; 524} 525 526// aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases, 527// such as array references. 528// 529AliasAnalysis::AliasResult 530BasicAliasAnalysis::aliasCheck(const Value *V1, unsigned V1Size, 531 const Value *V2, unsigned V2Size) { 532 // Strip off any casts if they exist. 533 V1 = V1->stripPointerCasts(); 534 V2 = V2->stripPointerCasts(); 535 536 // Are we checking for alias of the same value? 537 if (V1 == V2) return MustAlias; 538 539 if (!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType())) 540 return NoAlias; // Scalars cannot alias each other 541 542 // Figure out what objects these things are pointing to if we can. 543 const Value *O1 = V1->getUnderlyingObject(); 544 const Value *O2 = V2->getUnderlyingObject(); 545 546 if (O1 != O2) { 547 // If V1/V2 point to two different objects we know that we have no alias. 548 if (isIdentifiedObject(O1) && isIdentifiedObject(O2)) 549 return NoAlias; 550 551 // Arguments can't alias with local allocations or noalias calls. 552 if ((isa<Argument>(O1) && (isa<AllocationInst>(O2) || isNoAliasCall(O2))) || 553 (isa<Argument>(O2) && (isa<AllocationInst>(O1) || isNoAliasCall(O1)))) 554 return NoAlias; 555 556 // Most objects can't alias null. 557 if ((isa<ConstantPointerNull>(V2) && isKnownNonNull(O1)) || 558 (isa<ConstantPointerNull>(V1) && isKnownNonNull(O2))) 559 return NoAlias; 560 } 561 562 // If the size of one access is larger than the entire object on the other 563 // side, then we know such behavior is undefined and can assume no alias. 564 LLVMContext &Context = V1->getContext(); 565 if (TD) 566 if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, Context, *TD)) || 567 (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, Context, *TD))) 568 return NoAlias; 569 570 // If one pointer is the result of a call/invoke and the other is a 571 // non-escaping local object, then we know the object couldn't escape to a 572 // point where the call could return it. 573 if ((isa<CallInst>(O1) || isa<InvokeInst>(O1)) && 574 isNonEscapingLocalObject(O2) && O1 != O2) 575 return NoAlias; 576 if ((isa<CallInst>(O2) || isa<InvokeInst>(O2)) && 577 isNonEscapingLocalObject(O1) && O1 != O2) 578 return NoAlias; 579 580 if (!isGEP(V1) && isGEP(V2)) { 581 std::swap(V1, V2); 582 std::swap(V1Size, V2Size); 583 } 584 if (isGEP(V1)) 585 return aliasGEP(V1, V1Size, V2, V2Size); 586 587 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) { 588 std::swap(V1, V2); 589 std::swap(V1Size, V2Size); 590 } 591 if (const PHINode *PN = dyn_cast<PHINode>(V1)) 592 return aliasPHI(PN, V1Size, V2, V2Size); 593 594 return MayAlias; 595} 596 597// This function is used to determine if the indices of two GEP instructions are 598// equal. V1 and V2 are the indices. 599static bool IndexOperandsEqual(Value *V1, Value *V2, LLVMContext &Context) { 600 if (V1->getType() == V2->getType()) 601 return V1 == V2; 602 if (Constant *C1 = dyn_cast<Constant>(V1)) 603 if (Constant *C2 = dyn_cast<Constant>(V2)) { 604 // Sign extend the constants to long types, if necessary 605 if (C1->getType() != Type::getInt64Ty(Context)) 606 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(Context)); 607 if (C2->getType() != Type::getInt64Ty(Context)) 608 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(Context)); 609 return C1 == C2; 610 } 611 return false; 612} 613 614/// CheckGEPInstructions - Check two GEP instructions with known must-aliasing 615/// base pointers. This checks to see if the index expressions preclude the 616/// pointers from aliasing... 617AliasAnalysis::AliasResult 618BasicAliasAnalysis::CheckGEPInstructions( 619 const Type* BasePtr1Ty, Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1S, 620 const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2S) { 621 // We currently can't handle the case when the base pointers have different 622 // primitive types. Since this is uncommon anyway, we are happy being 623 // extremely conservative. 624 if (BasePtr1Ty != BasePtr2Ty) 625 return MayAlias; 626 627 const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty); 628 629 LLVMContext &Context = GEPPointerTy->getContext(); 630 631 // Find the (possibly empty) initial sequence of equal values... which are not 632 // necessarily constants. 633 unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops; 634 unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands); 635 unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands); 636 unsigned UnequalOper = 0; 637 while (UnequalOper != MinOperands && 638 IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper], 639 Context)) { 640 // Advance through the type as we go... 641 ++UnequalOper; 642 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty)) 643 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]); 644 else { 645 // If all operands equal each other, then the derived pointers must 646 // alias each other... 647 BasePtr1Ty = 0; 648 assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands && 649 "Ran out of type nesting, but not out of operands?"); 650 return MustAlias; 651 } 652 } 653 654 // If we have seen all constant operands, and run out of indexes on one of the 655 // getelementptrs, check to see if the tail of the leftover one is all zeros. 656 // If so, return mustalias. 657 if (UnequalOper == MinOperands) { 658 if (NumGEP1Ops < NumGEP2Ops) { 659 std::swap(GEP1Ops, GEP2Ops); 660 std::swap(NumGEP1Ops, NumGEP2Ops); 661 } 662 663 bool AllAreZeros = true; 664 for (unsigned i = UnequalOper; i != MaxOperands; ++i) 665 if (!isa<Constant>(GEP1Ops[i]) || 666 !cast<Constant>(GEP1Ops[i])->isNullValue()) { 667 AllAreZeros = false; 668 break; 669 } 670 if (AllAreZeros) return MustAlias; 671 } 672 673 674 // So now we know that the indexes derived from the base pointers, 675 // which are known to alias, are different. We can still determine a 676 // no-alias result if there are differing constant pairs in the index 677 // chain. For example: 678 // A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S)) 679 // 680 // We have to be careful here about array accesses. In particular, consider: 681 // A[1][0] vs A[0][i] 682 // In this case, we don't *know* that the array will be accessed in bounds: 683 // the index could even be negative. Because of this, we have to 684 // conservatively *give up* and return may alias. We disregard differing 685 // array subscripts that are followed by a variable index without going 686 // through a struct. 687 // 688 unsigned SizeMax = std::max(G1S, G2S); 689 if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work. 690 691 // Scan for the first operand that is constant and unequal in the 692 // two getelementptrs... 693 unsigned FirstConstantOper = UnequalOper; 694 for (; FirstConstantOper != MinOperands; ++FirstConstantOper) { 695 const Value *G1Oper = GEP1Ops[FirstConstantOper]; 696 const Value *G2Oper = GEP2Ops[FirstConstantOper]; 697 698 if (G1Oper != G2Oper) // Found non-equal constant indexes... 699 if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper))) 700 if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){ 701 if (G1OC->getType() != G2OC->getType()) { 702 // Sign extend both operands to long. 703 if (G1OC->getType() != Type::getInt64Ty(Context)) 704 G1OC = ConstantExpr::getSExt(G1OC, Type::getInt64Ty(Context)); 705 if (G2OC->getType() != Type::getInt64Ty(Context)) 706 G2OC = ConstantExpr::getSExt(G2OC, Type::getInt64Ty(Context)); 707 GEP1Ops[FirstConstantOper] = G1OC; 708 GEP2Ops[FirstConstantOper] = G2OC; 709 } 710 711 if (G1OC != G2OC) { 712 // Handle the "be careful" case above: if this is an array/vector 713 // subscript, scan for a subsequent variable array index. 714 if (const SequentialType *STy = 715 dyn_cast<SequentialType>(BasePtr1Ty)) { 716 const Type *NextTy = STy; 717 bool isBadCase = false; 718 719 for (unsigned Idx = FirstConstantOper; 720 Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) { 721 const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx]; 722 if (!isa<Constant>(V1) || !isa<Constant>(V2)) { 723 isBadCase = true; 724 break; 725 } 726 // If the array is indexed beyond the bounds of the static type 727 // at this level, it will also fall into the "be careful" case. 728 // It would theoretically be possible to analyze these cases, 729 // but for now just be conservatively correct. 730 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 731 if (cast<ConstantInt>(G1OC)->getZExtValue() >= 732 ATy->getNumElements() || 733 cast<ConstantInt>(G2OC)->getZExtValue() >= 734 ATy->getNumElements()) { 735 isBadCase = true; 736 break; 737 } 738 if (const VectorType *VTy = dyn_cast<VectorType>(STy)) 739 if (cast<ConstantInt>(G1OC)->getZExtValue() >= 740 VTy->getNumElements() || 741 cast<ConstantInt>(G2OC)->getZExtValue() >= 742 VTy->getNumElements()) { 743 isBadCase = true; 744 break; 745 } 746 STy = cast<SequentialType>(NextTy); 747 NextTy = cast<SequentialType>(NextTy)->getElementType(); 748 } 749 750 if (isBadCase) G1OC = 0; 751 } 752 753 // Make sure they are comparable (ie, not constant expressions), and 754 // make sure the GEP with the smaller leading constant is GEP1. 755 if (G1OC) { 756 Constant *Compare = ConstantExpr::getICmp(ICmpInst::ICMP_SGT, 757 G1OC, G2OC); 758 if (ConstantInt *CV = dyn_cast<ConstantInt>(Compare)) { 759 if (CV->getZExtValue()) { // If they are comparable and G2 > G1 760 std::swap(GEP1Ops, GEP2Ops); // Make GEP1 < GEP2 761 std::swap(NumGEP1Ops, NumGEP2Ops); 762 } 763 break; 764 } 765 } 766 } 767 } 768 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper); 769 } 770 771 // No shared constant operands, and we ran out of common operands. At this 772 // point, the GEP instructions have run through all of their operands, and we 773 // haven't found evidence that there are any deltas between the GEP's. 774 // However, one GEP may have more operands than the other. If this is the 775 // case, there may still be hope. Check this now. 776 if (FirstConstantOper == MinOperands) { 777 // Without TargetData, we won't know what the offsets are. 778 if (!TD) 779 return MayAlias; 780 781 // Make GEP1Ops be the longer one if there is a longer one. 782 if (NumGEP1Ops < NumGEP2Ops) { 783 std::swap(GEP1Ops, GEP2Ops); 784 std::swap(NumGEP1Ops, NumGEP2Ops); 785 } 786 787 // Is there anything to check? 788 if (NumGEP1Ops > MinOperands) { 789 for (unsigned i = FirstConstantOper; i != MaxOperands; ++i) 790 if (isa<ConstantInt>(GEP1Ops[i]) && 791 !cast<ConstantInt>(GEP1Ops[i])->isZero()) { 792 // Yup, there's a constant in the tail. Set all variables to 793 // constants in the GEP instruction to make it suitable for 794 // TargetData::getIndexedOffset. 795 for (i = 0; i != MaxOperands; ++i) 796 if (!isa<ConstantInt>(GEP1Ops[i])) 797 GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType()); 798 // Okay, now get the offset. This is the relative offset for the full 799 // instruction. 800 int64_t Offset1 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops, 801 NumGEP1Ops); 802 803 // Now check without any constants at the end. 804 int64_t Offset2 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops, 805 MinOperands); 806 807 // Make sure we compare the absolute difference. 808 if (Offset1 > Offset2) 809 std::swap(Offset1, Offset2); 810 811 // If the tail provided a bit enough offset, return noalias! 812 if ((uint64_t)(Offset2-Offset1) >= SizeMax) 813 return NoAlias; 814 // Otherwise break - we don't look for another constant in the tail. 815 break; 816 } 817 } 818 819 // Couldn't find anything useful. 820 return MayAlias; 821 } 822 823 // If there are non-equal constants arguments, then we can figure 824 // out a minimum known delta between the two index expressions... at 825 // this point we know that the first constant index of GEP1 is less 826 // than the first constant index of GEP2. 827 828 // Advance BasePtr[12]Ty over this first differing constant operand. 829 BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)-> 830 getTypeAtIndex(GEP2Ops[FirstConstantOper]); 831 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)-> 832 getTypeAtIndex(GEP1Ops[FirstConstantOper]); 833 834 // We are going to be using TargetData::getIndexedOffset to determine the 835 // offset that each of the GEP's is reaching. To do this, we have to convert 836 // all variable references to constant references. To do this, we convert the 837 // initial sequence of array subscripts into constant zeros to start with. 838 const Type *ZeroIdxTy = GEPPointerTy; 839 for (unsigned i = 0; i != FirstConstantOper; ++i) { 840 if (!isa<StructType>(ZeroIdxTy)) 841 GEP1Ops[i] = GEP2Ops[i] = 842 Constant::getNullValue(Type::getInt32Ty(Context)); 843 844 if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy)) 845 ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]); 846 } 847 848 // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok 849 850 // Loop over the rest of the operands... 851 for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) { 852 const Value *Op1 = i < NumGEP1Ops ? GEP1Ops[i] : 0; 853 const Value *Op2 = i < NumGEP2Ops ? GEP2Ops[i] : 0; 854 // If they are equal, use a zero index... 855 if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) { 856 if (!isa<ConstantInt>(Op1)) 857 GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType()); 858 // Otherwise, just keep the constants we have. 859 } else { 860 if (Op1) { 861 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 862 // If this is an array index, make sure the array element is in range. 863 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) { 864 if (Op1C->getZExtValue() >= AT->getNumElements()) 865 return MayAlias; // Be conservative with out-of-range accesses 866 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) { 867 if (Op1C->getZExtValue() >= VT->getNumElements()) 868 return MayAlias; // Be conservative with out-of-range accesses 869 } 870 871 } else { 872 // GEP1 is known to produce a value less than GEP2. To be 873 // conservatively correct, we must assume the largest possible 874 // constant is used in this position. This cannot be the initial 875 // index to the GEP instructions (because we know we have at least one 876 // element before this one with the different constant arguments), so 877 // we know that the current index must be into either a struct or 878 // array. Because we know it's not constant, this cannot be a 879 // structure index. Because of this, we can calculate the maximum 880 // value possible. 881 // 882 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) 883 GEP1Ops[i] = 884 ConstantInt::get(Type::getInt64Ty(Context), 885 AT->getNumElements()-1); 886 else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) 887 GEP1Ops[i] = 888 ConstantInt::get(Type::getInt64Ty(Context), 889 VT->getNumElements()-1); 890 } 891 } 892 893 if (Op2) { 894 if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) { 895 // If this is an array index, make sure the array element is in range. 896 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr2Ty)) { 897 if (Op2C->getZExtValue() >= AT->getNumElements()) 898 return MayAlias; // Be conservative with out-of-range accesses 899 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr2Ty)) { 900 if (Op2C->getZExtValue() >= VT->getNumElements()) 901 return MayAlias; // Be conservative with out-of-range accesses 902 } 903 } else { // Conservatively assume the minimum value for this index 904 GEP2Ops[i] = Constant::getNullValue(Op2->getType()); 905 } 906 } 907 } 908 909 if (BasePtr1Ty && Op1) { 910 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty)) 911 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]); 912 else 913 BasePtr1Ty = 0; 914 } 915 916 if (BasePtr2Ty && Op2) { 917 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty)) 918 BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]); 919 else 920 BasePtr2Ty = 0; 921 } 922 } 923 924 if (TD && GEPPointerTy->getElementType()->isSized()) { 925 int64_t Offset1 = 926 TD->getIndexedOffset(GEPPointerTy, GEP1Ops, NumGEP1Ops); 927 int64_t Offset2 = 928 TD->getIndexedOffset(GEPPointerTy, GEP2Ops, NumGEP2Ops); 929 assert(Offset1 != Offset2 && 930 "There is at least one different constant here!"); 931 932 // Make sure we compare the absolute difference. 933 if (Offset1 > Offset2) 934 std::swap(Offset1, Offset2); 935 936 if ((uint64_t)(Offset2-Offset1) >= SizeMax) { 937 //cerr << "Determined that these two GEP's don't alias [" 938 // << SizeMax << " bytes]: \n" << *GEP1 << *GEP2; 939 return NoAlias; 940 } 941 } 942 return MayAlias; 943} 944 945// Make sure that anything that uses AliasAnalysis pulls in this file... 946DEFINING_FILE_FOR(BasicAliasAnalysis) 947