BasicAliasAnalysis.cpp revision 22c031296ce91cdabb193a203b06b339ef500f4b
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 (TD) { 311 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) { 312 switch (II->getIntrinsicID()) { 313 default: break; 314 case Intrinsic::atomic_cmp_swap: 315 case Intrinsic::atomic_swap: 316 case Intrinsic::atomic_load_add: 317 case Intrinsic::atomic_load_sub: 318 case Intrinsic::atomic_load_and: 319 case Intrinsic::atomic_load_nand: 320 case Intrinsic::atomic_load_or: 321 case Intrinsic::atomic_load_xor: 322 case Intrinsic::atomic_load_max: 323 case Intrinsic::atomic_load_min: 324 case Intrinsic::atomic_load_umax: 325 case Intrinsic::atomic_load_umin: { 326 Value *Op1 = II->getOperand(1); 327 unsigned Op1Size = TD->getTypeStoreSize(Op1->getType()); 328 if (alias(Op1, Op1Size, P, Size) == NoAlias) 329 return NoModRef; 330 break; 331 } 332 } 333 } 334 } 335 } 336 337 // The AliasAnalysis base class has some smarts, lets use them. 338 return AliasAnalysis::getModRefInfo(CS, P, Size); 339} 340 341 342AliasAnalysis::ModRefResult 343BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) { 344 // If CS1 or CS2 are readnone, they don't interact. 345 ModRefBehavior CS1B = AliasAnalysis::getModRefBehavior(CS1); 346 if (CS1B == DoesNotAccessMemory) return NoModRef; 347 348 ModRefBehavior CS2B = AliasAnalysis::getModRefBehavior(CS2); 349 if (CS2B == DoesNotAccessMemory) return NoModRef; 350 351 // If they both only read from memory, just return ref. 352 if (CS1B == OnlyReadsMemory && CS2B == OnlyReadsMemory) 353 return Ref; 354 355 // Otherwise, fall back to NoAA (mod+ref). 356 return NoAA::getModRefInfo(CS1, CS2); 357} 358 359// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction 360// against another. 361// 362AliasAnalysis::AliasResult 363BasicAliasAnalysis::aliasGEP(const Value *V1, unsigned V1Size, 364 const Value *V2, unsigned V2Size) { 365 // If we have two gep instructions with must-alias'ing base pointers, figure 366 // out if the indexes to the GEP tell us anything about the derived pointer. 367 // Note that we also handle chains of getelementptr instructions as well as 368 // constant expression getelementptrs here. 369 // 370 if (isGEP(V1) && isGEP(V2)) { 371 const User *GEP1 = cast<User>(V1); 372 const User *GEP2 = cast<User>(V2); 373 374 // If V1 and V2 are identical GEPs, just recurse down on both of them. 375 // This allows us to analyze things like: 376 // P = gep A, 0, i, 1 377 // Q = gep B, 0, i, 1 378 // by just analyzing A and B. This is even safe for variable indices. 379 if (GEP1->getType() == GEP2->getType() && 380 GEP1->getNumOperands() == GEP2->getNumOperands() && 381 GEP1->getOperand(0)->getType() == GEP2->getOperand(0)->getType() && 382 // All operands are the same, ignoring the base. 383 std::equal(GEP1->op_begin()+1, GEP1->op_end(), GEP2->op_begin()+1)) 384 return aliasCheck(GEP1->getOperand(0), V1Size, 385 GEP2->getOperand(0), V2Size); 386 387 // Drill down into the first non-gep value, to test for must-aliasing of 388 // the base pointers. 389 while (isGEP(GEP1->getOperand(0)) && 390 GEP1->getOperand(1) == 391 Constant::getNullValue(GEP1->getOperand(1)->getType())) 392 GEP1 = cast<User>(GEP1->getOperand(0)); 393 const Value *BasePtr1 = GEP1->getOperand(0); 394 395 while (isGEP(GEP2->getOperand(0)) && 396 GEP2->getOperand(1) == 397 Constant::getNullValue(GEP2->getOperand(1)->getType())) 398 GEP2 = cast<User>(GEP2->getOperand(0)); 399 const Value *BasePtr2 = GEP2->getOperand(0); 400 401 // Do the base pointers alias? 402 AliasResult BaseAlias = aliasCheck(BasePtr1, ~0U, BasePtr2, ~0U); 403 if (BaseAlias == NoAlias) return NoAlias; 404 if (BaseAlias == MustAlias) { 405 // If the base pointers alias each other exactly, check to see if we can 406 // figure out anything about the resultant pointers, to try to prove 407 // non-aliasing. 408 409 // Collect all of the chained GEP operands together into one simple place 410 SmallVector<Value*, 16> GEP1Ops, GEP2Ops; 411 BasePtr1 = GetGEPOperands(V1, GEP1Ops); 412 BasePtr2 = GetGEPOperands(V2, GEP2Ops); 413 414 // If GetGEPOperands were able to fold to the same must-aliased pointer, 415 // do the comparison. 416 if (BasePtr1 == BasePtr2) { 417 AliasResult GAlias = 418 CheckGEPInstructions(BasePtr1->getType(), 419 &GEP1Ops[0], GEP1Ops.size(), V1Size, 420 BasePtr2->getType(), 421 &GEP2Ops[0], GEP2Ops.size(), V2Size); 422 if (GAlias != MayAlias) 423 return GAlias; 424 } 425 } 426 } 427 428 // Check to see if these two pointers are related by a getelementptr 429 // instruction. If one pointer is a GEP with a non-zero index of the other 430 // pointer, we know they cannot alias. 431 // 432 if (V1Size == ~0U || V2Size == ~0U) 433 return MayAlias; 434 435 SmallVector<Value*, 16> GEPOperands; 436 const Value *BasePtr = GetGEPOperands(V1, GEPOperands); 437 438 AliasResult R = aliasCheck(BasePtr, ~0U, V2, V2Size); 439 if (R != MustAlias) 440 // If V2 may alias GEP base pointer, conservatively returns MayAlias. 441 // If V2 is known not to alias GEP base pointer, then the two values 442 // cannot alias per GEP semantics: "A pointer value formed from a 443 // getelementptr instruction is associated with the addresses associated 444 // with the first operand of the getelementptr". 445 return R; 446 447 // If there is at least one non-zero constant index, we know they cannot 448 // alias. 449 bool ConstantFound = false; 450 bool AllZerosFound = true; 451 for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i) 452 if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) { 453 if (!C->isNullValue()) { 454 ConstantFound = true; 455 AllZerosFound = false; 456 break; 457 } 458 } else { 459 AllZerosFound = false; 460 } 461 462 // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases 463 // the ptr, the end result is a must alias also. 464 if (AllZerosFound) 465 return MustAlias; 466 467 if (ConstantFound) { 468 if (V2Size <= 1 && V1Size <= 1) // Just pointer check? 469 return NoAlias; 470 471 // Otherwise we have to check to see that the distance is more than 472 // the size of the argument... build an index vector that is equal to 473 // the arguments provided, except substitute 0's for any variable 474 // indexes we find... 475 if (TD && 476 cast<PointerType>(BasePtr->getType())->getElementType()->isSized()) { 477 for (unsigned i = 0; i != GEPOperands.size(); ++i) 478 if (!isa<ConstantInt>(GEPOperands[i])) 479 GEPOperands[i] = Constant::getNullValue(GEPOperands[i]->getType()); 480 int64_t Offset = TD->getIndexedOffset(BasePtr->getType(), 481 &GEPOperands[0], 482 GEPOperands.size()); 483 484 if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size) 485 return NoAlias; 486 } 487 } 488 489 return MayAlias; 490} 491 492// aliasPHI - Provide a bunch of ad-hoc rules to disambiguate a PHI instruction 493// against another. 494AliasAnalysis::AliasResult 495BasicAliasAnalysis::aliasPHI(const PHINode *PN, unsigned PNSize, 496 const Value *V2, unsigned V2Size) { 497 // The PHI node has already been visited, avoid recursion any further. 498 if (!VisitedPHIs.insert(PN)) 499 return MayAlias; 500 501 SmallSet<Value*, 4> UniqueSrc; 502 SmallVector<Value*, 4> V1Srcs; 503 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 504 Value *PV1 = PN->getIncomingValue(i); 505 if (isa<PHINode>(PV1)) 506 // If any of the source itself is a PHI, return MayAlias conservatively 507 // to avoid compile time explosion. The worst possible case is if both 508 // sides are PHI nodes. In which case, this is O(m x n) time where 'm' 509 // and 'n' are the number of PHI sources. 510 return MayAlias; 511 if (UniqueSrc.insert(PV1)) 512 V1Srcs.push_back(PV1); 513 } 514 515 AliasResult Alias = aliasCheck(V1Srcs[0], PNSize, V2, V2Size); 516 // Early exit if the check of the first PHI source against V2 is MayAlias. 517 // Other results are not possible. 518 if (Alias == MayAlias) 519 return MayAlias; 520 521 // If all sources of the PHI node NoAlias or MustAlias V2, then returns 522 // NoAlias / MustAlias. Otherwise, returns MayAlias. 523 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) { 524 Value *V = V1Srcs[i]; 525 AliasResult ThisAlias = aliasCheck(V, PNSize, V2, V2Size); 526 if (ThisAlias != Alias || ThisAlias == MayAlias) 527 return MayAlias; 528 } 529 530 return Alias; 531} 532 533// aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases, 534// such as array references. 535// 536AliasAnalysis::AliasResult 537BasicAliasAnalysis::aliasCheck(const Value *V1, unsigned V1Size, 538 const Value *V2, unsigned V2Size) { 539 // Strip off any casts if they exist. 540 V1 = V1->stripPointerCasts(); 541 V2 = V2->stripPointerCasts(); 542 543 // Are we checking for alias of the same value? 544 if (V1 == V2) return MustAlias; 545 546 if (!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType())) 547 return NoAlias; // Scalars cannot alias each other 548 549 // Figure out what objects these things are pointing to if we can. 550 const Value *O1 = V1->getUnderlyingObject(); 551 const Value *O2 = V2->getUnderlyingObject(); 552 553 if (O1 != O2) { 554 // If V1/V2 point to two different objects we know that we have no alias. 555 if (isIdentifiedObject(O1) && isIdentifiedObject(O2)) 556 return NoAlias; 557 558 // Arguments can't alias with local allocations or noalias calls. 559 if ((isa<Argument>(O1) && (isa<AllocationInst>(O2) || isNoAliasCall(O2))) || 560 (isa<Argument>(O2) && (isa<AllocationInst>(O1) || isNoAliasCall(O1)))) 561 return NoAlias; 562 563 // Most objects can't alias null. 564 if ((isa<ConstantPointerNull>(V2) && isKnownNonNull(O1)) || 565 (isa<ConstantPointerNull>(V1) && isKnownNonNull(O2))) 566 return NoAlias; 567 } 568 569 // If the size of one access is larger than the entire object on the other 570 // side, then we know such behavior is undefined and can assume no alias. 571 LLVMContext &Context = V1->getContext(); 572 if (TD) 573 if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, Context, *TD)) || 574 (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, Context, *TD))) 575 return NoAlias; 576 577 // If one pointer is the result of a call/invoke and the other is a 578 // non-escaping local object, then we know the object couldn't escape to a 579 // point where the call could return it. 580 if ((isa<CallInst>(O1) || isa<InvokeInst>(O1)) && 581 isNonEscapingLocalObject(O2) && O1 != O2) 582 return NoAlias; 583 if ((isa<CallInst>(O2) || isa<InvokeInst>(O2)) && 584 isNonEscapingLocalObject(O1) && O1 != O2) 585 return NoAlias; 586 587 if (!isGEP(V1) && isGEP(V2)) { 588 std::swap(V1, V2); 589 std::swap(V1Size, V2Size); 590 } 591 if (isGEP(V1)) 592 return aliasGEP(V1, V1Size, V2, V2Size); 593 594 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) { 595 std::swap(V1, V2); 596 std::swap(V1Size, V2Size); 597 } 598 if (const PHINode *PN = dyn_cast<PHINode>(V1)) 599 return aliasPHI(PN, V1Size, V2, V2Size); 600 601 return MayAlias; 602} 603 604// This function is used to determine if the indices of two GEP instructions are 605// equal. V1 and V2 are the indices. 606static bool IndexOperandsEqual(Value *V1, Value *V2, LLVMContext &Context) { 607 if (V1->getType() == V2->getType()) 608 return V1 == V2; 609 if (Constant *C1 = dyn_cast<Constant>(V1)) 610 if (Constant *C2 = dyn_cast<Constant>(V2)) { 611 // Sign extend the constants to long types, if necessary 612 if (C1->getType() != Type::getInt64Ty(Context)) 613 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(Context)); 614 if (C2->getType() != Type::getInt64Ty(Context)) 615 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(Context)); 616 return C1 == C2; 617 } 618 return false; 619} 620 621/// CheckGEPInstructions - Check two GEP instructions with known must-aliasing 622/// base pointers. This checks to see if the index expressions preclude the 623/// pointers from aliasing... 624AliasAnalysis::AliasResult 625BasicAliasAnalysis::CheckGEPInstructions( 626 const Type* BasePtr1Ty, Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1S, 627 const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2S) { 628 // We currently can't handle the case when the base pointers have different 629 // primitive types. Since this is uncommon anyway, we are happy being 630 // extremely conservative. 631 if (BasePtr1Ty != BasePtr2Ty) 632 return MayAlias; 633 634 const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty); 635 636 LLVMContext &Context = GEPPointerTy->getContext(); 637 638 // Find the (possibly empty) initial sequence of equal values... which are not 639 // necessarily constants. 640 unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops; 641 unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands); 642 unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands); 643 unsigned UnequalOper = 0; 644 while (UnequalOper != MinOperands && 645 IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper], 646 Context)) { 647 // Advance through the type as we go... 648 ++UnequalOper; 649 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty)) 650 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]); 651 else { 652 // If all operands equal each other, then the derived pointers must 653 // alias each other... 654 BasePtr1Ty = 0; 655 assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands && 656 "Ran out of type nesting, but not out of operands?"); 657 return MustAlias; 658 } 659 } 660 661 // If we have seen all constant operands, and run out of indexes on one of the 662 // getelementptrs, check to see if the tail of the leftover one is all zeros. 663 // If so, return mustalias. 664 if (UnequalOper == MinOperands) { 665 if (NumGEP1Ops < NumGEP2Ops) { 666 std::swap(GEP1Ops, GEP2Ops); 667 std::swap(NumGEP1Ops, NumGEP2Ops); 668 } 669 670 bool AllAreZeros = true; 671 for (unsigned i = UnequalOper; i != MaxOperands; ++i) 672 if (!isa<Constant>(GEP1Ops[i]) || 673 !cast<Constant>(GEP1Ops[i])->isNullValue()) { 674 AllAreZeros = false; 675 break; 676 } 677 if (AllAreZeros) return MustAlias; 678 } 679 680 681 // So now we know that the indexes derived from the base pointers, 682 // which are known to alias, are different. We can still determine a 683 // no-alias result if there are differing constant pairs in the index 684 // chain. For example: 685 // A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S)) 686 // 687 // We have to be careful here about array accesses. In particular, consider: 688 // A[1][0] vs A[0][i] 689 // In this case, we don't *know* that the array will be accessed in bounds: 690 // the index could even be negative. Because of this, we have to 691 // conservatively *give up* and return may alias. We disregard differing 692 // array subscripts that are followed by a variable index without going 693 // through a struct. 694 // 695 unsigned SizeMax = std::max(G1S, G2S); 696 if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work. 697 698 // Scan for the first operand that is constant and unequal in the 699 // two getelementptrs... 700 unsigned FirstConstantOper = UnequalOper; 701 for (; FirstConstantOper != MinOperands; ++FirstConstantOper) { 702 const Value *G1Oper = GEP1Ops[FirstConstantOper]; 703 const Value *G2Oper = GEP2Ops[FirstConstantOper]; 704 705 if (G1Oper != G2Oper) // Found non-equal constant indexes... 706 if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper))) 707 if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){ 708 if (G1OC->getType() != G2OC->getType()) { 709 // Sign extend both operands to long. 710 if (G1OC->getType() != Type::getInt64Ty(Context)) 711 G1OC = ConstantExpr::getSExt(G1OC, Type::getInt64Ty(Context)); 712 if (G2OC->getType() != Type::getInt64Ty(Context)) 713 G2OC = ConstantExpr::getSExt(G2OC, Type::getInt64Ty(Context)); 714 GEP1Ops[FirstConstantOper] = G1OC; 715 GEP2Ops[FirstConstantOper] = G2OC; 716 } 717 718 if (G1OC != G2OC) { 719 // Handle the "be careful" case above: if this is an array/vector 720 // subscript, scan for a subsequent variable array index. 721 if (const SequentialType *STy = 722 dyn_cast<SequentialType>(BasePtr1Ty)) { 723 const Type *NextTy = STy; 724 bool isBadCase = false; 725 726 for (unsigned Idx = FirstConstantOper; 727 Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) { 728 const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx]; 729 if (!isa<Constant>(V1) || !isa<Constant>(V2)) { 730 isBadCase = true; 731 break; 732 } 733 // If the array is indexed beyond the bounds of the static type 734 // at this level, it will also fall into the "be careful" case. 735 // It would theoretically be possible to analyze these cases, 736 // but for now just be conservatively correct. 737 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 738 if (cast<ConstantInt>(G1OC)->getZExtValue() >= 739 ATy->getNumElements() || 740 cast<ConstantInt>(G2OC)->getZExtValue() >= 741 ATy->getNumElements()) { 742 isBadCase = true; 743 break; 744 } 745 if (const VectorType *VTy = dyn_cast<VectorType>(STy)) 746 if (cast<ConstantInt>(G1OC)->getZExtValue() >= 747 VTy->getNumElements() || 748 cast<ConstantInt>(G2OC)->getZExtValue() >= 749 VTy->getNumElements()) { 750 isBadCase = true; 751 break; 752 } 753 STy = cast<SequentialType>(NextTy); 754 NextTy = cast<SequentialType>(NextTy)->getElementType(); 755 } 756 757 if (isBadCase) G1OC = 0; 758 } 759 760 // Make sure they are comparable (ie, not constant expressions), and 761 // make sure the GEP with the smaller leading constant is GEP1. 762 if (G1OC) { 763 Constant *Compare = ConstantExpr::getICmp(ICmpInst::ICMP_SGT, 764 G1OC, G2OC); 765 if (ConstantInt *CV = dyn_cast<ConstantInt>(Compare)) { 766 if (CV->getZExtValue()) { // If they are comparable and G2 > G1 767 std::swap(GEP1Ops, GEP2Ops); // Make GEP1 < GEP2 768 std::swap(NumGEP1Ops, NumGEP2Ops); 769 } 770 break; 771 } 772 } 773 } 774 } 775 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper); 776 } 777 778 // No shared constant operands, and we ran out of common operands. At this 779 // point, the GEP instructions have run through all of their operands, and we 780 // haven't found evidence that there are any deltas between the GEP's. 781 // However, one GEP may have more operands than the other. If this is the 782 // case, there may still be hope. Check this now. 783 if (FirstConstantOper == MinOperands) { 784 // Without TargetData, we won't know what the offsets are. 785 if (!TD) 786 return MayAlias; 787 788 // Make GEP1Ops be the longer one if there is a longer one. 789 if (NumGEP1Ops < NumGEP2Ops) { 790 std::swap(GEP1Ops, GEP2Ops); 791 std::swap(NumGEP1Ops, NumGEP2Ops); 792 } 793 794 // Is there anything to check? 795 if (NumGEP1Ops > MinOperands) { 796 for (unsigned i = FirstConstantOper; i != MaxOperands; ++i) 797 if (isa<ConstantInt>(GEP1Ops[i]) && 798 !cast<ConstantInt>(GEP1Ops[i])->isZero()) { 799 // Yup, there's a constant in the tail. Set all variables to 800 // constants in the GEP instruction to make it suitable for 801 // TargetData::getIndexedOffset. 802 for (i = 0; i != MaxOperands; ++i) 803 if (!isa<ConstantInt>(GEP1Ops[i])) 804 GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType()); 805 // Okay, now get the offset. This is the relative offset for the full 806 // instruction. 807 int64_t Offset1 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops, 808 NumGEP1Ops); 809 810 // Now check without any constants at the end. 811 int64_t Offset2 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops, 812 MinOperands); 813 814 // Make sure we compare the absolute difference. 815 if (Offset1 > Offset2) 816 std::swap(Offset1, Offset2); 817 818 // If the tail provided a bit enough offset, return noalias! 819 if ((uint64_t)(Offset2-Offset1) >= SizeMax) 820 return NoAlias; 821 // Otherwise break - we don't look for another constant in the tail. 822 break; 823 } 824 } 825 826 // Couldn't find anything useful. 827 return MayAlias; 828 } 829 830 // If there are non-equal constants arguments, then we can figure 831 // out a minimum known delta between the two index expressions... at 832 // this point we know that the first constant index of GEP1 is less 833 // than the first constant index of GEP2. 834 835 // Advance BasePtr[12]Ty over this first differing constant operand. 836 BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)-> 837 getTypeAtIndex(GEP2Ops[FirstConstantOper]); 838 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)-> 839 getTypeAtIndex(GEP1Ops[FirstConstantOper]); 840 841 // We are going to be using TargetData::getIndexedOffset to determine the 842 // offset that each of the GEP's is reaching. To do this, we have to convert 843 // all variable references to constant references. To do this, we convert the 844 // initial sequence of array subscripts into constant zeros to start with. 845 const Type *ZeroIdxTy = GEPPointerTy; 846 for (unsigned i = 0; i != FirstConstantOper; ++i) { 847 if (!isa<StructType>(ZeroIdxTy)) 848 GEP1Ops[i] = GEP2Ops[i] = 849 Constant::getNullValue(Type::getInt32Ty(Context)); 850 851 if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy)) 852 ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]); 853 } 854 855 // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok 856 857 // Loop over the rest of the operands... 858 for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) { 859 const Value *Op1 = i < NumGEP1Ops ? GEP1Ops[i] : 0; 860 const Value *Op2 = i < NumGEP2Ops ? GEP2Ops[i] : 0; 861 // If they are equal, use a zero index... 862 if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) { 863 if (!isa<ConstantInt>(Op1)) 864 GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType()); 865 // Otherwise, just keep the constants we have. 866 } else { 867 if (Op1) { 868 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 869 // If this is an array index, make sure the array element is in range. 870 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) { 871 if (Op1C->getZExtValue() >= AT->getNumElements()) 872 return MayAlias; // Be conservative with out-of-range accesses 873 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) { 874 if (Op1C->getZExtValue() >= VT->getNumElements()) 875 return MayAlias; // Be conservative with out-of-range accesses 876 } 877 878 } else { 879 // GEP1 is known to produce a value less than GEP2. To be 880 // conservatively correct, we must assume the largest possible 881 // constant is used in this position. This cannot be the initial 882 // index to the GEP instructions (because we know we have at least one 883 // element before this one with the different constant arguments), so 884 // we know that the current index must be into either a struct or 885 // array. Because we know it's not constant, this cannot be a 886 // structure index. Because of this, we can calculate the maximum 887 // value possible. 888 // 889 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) 890 GEP1Ops[i] = 891 ConstantInt::get(Type::getInt64Ty(Context), 892 AT->getNumElements()-1); 893 else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) 894 GEP1Ops[i] = 895 ConstantInt::get(Type::getInt64Ty(Context), 896 VT->getNumElements()-1); 897 } 898 } 899 900 if (Op2) { 901 if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) { 902 // If this is an array index, make sure the array element is in range. 903 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr2Ty)) { 904 if (Op2C->getZExtValue() >= AT->getNumElements()) 905 return MayAlias; // Be conservative with out-of-range accesses 906 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr2Ty)) { 907 if (Op2C->getZExtValue() >= VT->getNumElements()) 908 return MayAlias; // Be conservative with out-of-range accesses 909 } 910 } else { // Conservatively assume the minimum value for this index 911 GEP2Ops[i] = Constant::getNullValue(Op2->getType()); 912 } 913 } 914 } 915 916 if (BasePtr1Ty && Op1) { 917 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty)) 918 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]); 919 else 920 BasePtr1Ty = 0; 921 } 922 923 if (BasePtr2Ty && Op2) { 924 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty)) 925 BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]); 926 else 927 BasePtr2Ty = 0; 928 } 929 } 930 931 if (TD && GEPPointerTy->getElementType()->isSized()) { 932 int64_t Offset1 = 933 TD->getIndexedOffset(GEPPointerTy, GEP1Ops, NumGEP1Ops); 934 int64_t Offset2 = 935 TD->getIndexedOffset(GEPPointerTy, GEP2Ops, NumGEP2Ops); 936 assert(Offset1 != Offset2 && 937 "There is at least one different constant here!"); 938 939 // Make sure we compare the absolute difference. 940 if (Offset1 > Offset2) 941 std::swap(Offset1, Offset2); 942 943 if ((uint64_t)(Offset2-Offset1) >= SizeMax) { 944 //cerr << "Determined that these two GEP's don't alias [" 945 // << SizeMax << " bytes]: \n" << *GEP1 << *GEP2; 946 return NoAlias; 947 } 948 } 949 return MayAlias; 950} 951 952// Make sure that anything that uses AliasAnalysis pulls in this file... 953DEFINING_FILE_FOR(BasicAliasAnalysis) 954