BasicAliasAnalysis.cpp revision 50a5914e129c348e8878d4654b4306e0349281c2
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 SmallSet<const Value*, 16> VisitedPHIs; 207 return aliasCheck(V1, V1Size, V2, V2Size, VisitedPHIs); 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 // aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction 223 // against another. 224 AliasResult aliasGEP(const Value *V1, unsigned V1Size, 225 const Value *V2, unsigned V2Size, 226 SmallSet<const Value*, 16> &VisitedPHIs); 227 228 AliasResult aliasPHI(const Value *V1, unsigned V1Size, 229 const Value *V2, unsigned V2Size, 230 SmallSet<const Value*, 16> &VisitedPHIs); 231 232 AliasResult aliasCheck(const Value *V1, unsigned V1Size, 233 const Value *V2, unsigned V2Size, 234 SmallSet<const Value*, 16> &VisitedPHIs); 235 236 // CheckGEPInstructions - Check two GEP instructions with known 237 // must-aliasing base pointers. This checks to see if the index expressions 238 // preclude the pointers from aliasing... 239 AliasResult 240 CheckGEPInstructions(const Type* BasePtr1Ty, 241 Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1Size, 242 const Type *BasePtr2Ty, 243 Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size); 244 }; 245} // End of anonymous namespace 246 247// Register this pass... 248char BasicAliasAnalysis::ID = 0; 249static RegisterPass<BasicAliasAnalysis> 250X("basicaa", "Basic Alias Analysis (default AA impl)", false, true); 251 252// Declare that we implement the AliasAnalysis interface 253static RegisterAnalysisGroup<AliasAnalysis, true> Y(X); 254 255ImmutablePass *llvm::createBasicAliasAnalysisPass() { 256 return new BasicAliasAnalysis(); 257} 258 259 260/// pointsToConstantMemory - Chase pointers until we find a (constant 261/// global) or not. 262bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) { 263 if (const GlobalVariable *GV = 264 dyn_cast<GlobalVariable>(P->getUnderlyingObject())) 265 return GV->isConstant(); 266 return false; 267} 268 269 270// getModRefInfo - Check to see if the specified callsite can clobber the 271// specified memory object. Since we only look at local properties of this 272// function, we really can't say much about this query. We do, however, use 273// simple "address taken" analysis on local objects. 274// 275AliasAnalysis::ModRefResult 276BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) { 277 if (!isa<Constant>(P)) { 278 const Value *Object = P->getUnderlyingObject(); 279 280 // If this is a tail call and P points to a stack location, we know that 281 // the tail call cannot access or modify the local stack. 282 // We cannot exclude byval arguments here; these belong to the caller of 283 // the current function not to the current function, and a tail callee 284 // may reference them. 285 if (isa<AllocaInst>(Object)) 286 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) 287 if (CI->isTailCall()) 288 return NoModRef; 289 290 // If the pointer is to a locally allocated object that does not escape, 291 // then the call can not mod/ref the pointer unless the call takes the 292 // argument without capturing it. 293 if (isNonEscapingLocalObject(Object) && CS.getInstruction() != Object) { 294 bool passedAsArg = false; 295 // TODO: Eventually only check 'nocapture' arguments. 296 for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end(); 297 CI != CE; ++CI) 298 if (isa<PointerType>((*CI)->getType()) && 299 alias(cast<Value>(CI), ~0U, P, ~0U) != NoAlias) 300 passedAsArg = true; 301 302 if (!passedAsArg) 303 return NoModRef; 304 } 305 306 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) { 307 switch (II->getIntrinsicID()) { 308 default: break; 309 case Intrinsic::atomic_cmp_swap: 310 case Intrinsic::atomic_swap: 311 case Intrinsic::atomic_load_add: 312 case Intrinsic::atomic_load_sub: 313 case Intrinsic::atomic_load_and: 314 case Intrinsic::atomic_load_nand: 315 case Intrinsic::atomic_load_or: 316 case Intrinsic::atomic_load_xor: 317 case Intrinsic::atomic_load_max: 318 case Intrinsic::atomic_load_min: 319 case Intrinsic::atomic_load_umax: 320 case Intrinsic::atomic_load_umin: 321 if (alias(II->getOperand(1), Size, P, Size) == NoAlias) 322 return NoModRef; 323 break; 324 } 325 } 326 } 327 328 // The AliasAnalysis base class has some smarts, lets use them. 329 return AliasAnalysis::getModRefInfo(CS, P, Size); 330} 331 332 333AliasAnalysis::ModRefResult 334BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) { 335 // If CS1 or CS2 are readnone, they don't interact. 336 ModRefBehavior CS1B = AliasAnalysis::getModRefBehavior(CS1); 337 if (CS1B == DoesNotAccessMemory) return NoModRef; 338 339 ModRefBehavior CS2B = AliasAnalysis::getModRefBehavior(CS2); 340 if (CS2B == DoesNotAccessMemory) return NoModRef; 341 342 // If they both only read from memory, just return ref. 343 if (CS1B == OnlyReadsMemory && CS2B == OnlyReadsMemory) 344 return Ref; 345 346 // Otherwise, fall back to NoAA (mod+ref). 347 return NoAA::getModRefInfo(CS1, CS2); 348} 349 350// aliasGEP - Provide a bunch of ad-hoc rules to disambiguate a GEP instruction 351// against another. 352// 353AliasAnalysis::AliasResult 354BasicAliasAnalysis::aliasGEP(const Value *V1, unsigned V1Size, 355 const Value *V2, unsigned V2Size, 356 SmallSet<const Value*, 16> &VisitedPHIs) { 357 // If we have two gep instructions with must-alias'ing base pointers, figure 358 // out if the indexes to the GEP tell us anything about the derived pointer. 359 // Note that we also handle chains of getelementptr instructions as well as 360 // constant expression getelementptrs here. 361 // 362 if (isGEP(V1) && isGEP(V2)) { 363 const User *GEP1 = cast<User>(V1); 364 const User *GEP2 = cast<User>(V2); 365 366 // If V1 and V2 are identical GEPs, just recurse down on both of them. 367 // This allows us to analyze things like: 368 // P = gep A, 0, i, 1 369 // Q = gep B, 0, i, 1 370 // by just analyzing A and B. This is even safe for variable indices. 371 if (GEP1->getType() == GEP2->getType() && 372 GEP1->getNumOperands() == GEP2->getNumOperands() && 373 GEP1->getOperand(0)->getType() == GEP2->getOperand(0)->getType() && 374 // All operands are the same, ignoring the base. 375 std::equal(GEP1->op_begin()+1, GEP1->op_end(), GEP2->op_begin()+1)) 376 return aliasCheck(GEP1->getOperand(0), V1Size, 377 GEP2->getOperand(0), V2Size, VisitedPHIs); 378 379 // Drill down into the first non-gep value, to test for must-aliasing of 380 // the base pointers. 381 while (isGEP(GEP1->getOperand(0)) && 382 GEP1->getOperand(1) == 383 Constant::getNullValue(GEP1->getOperand(1)->getType())) 384 GEP1 = cast<User>(GEP1->getOperand(0)); 385 const Value *BasePtr1 = GEP1->getOperand(0); 386 387 while (isGEP(GEP2->getOperand(0)) && 388 GEP2->getOperand(1) == 389 Constant::getNullValue(GEP2->getOperand(1)->getType())) 390 GEP2 = cast<User>(GEP2->getOperand(0)); 391 const Value *BasePtr2 = GEP2->getOperand(0); 392 393 // Do the base pointers alias? 394 AliasResult BaseAlias = aliasCheck(BasePtr1, ~0U, BasePtr2, ~0U, 395 VisitedPHIs); 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, V1Size, V2, V2Size, VisitedPHIs); 432 if (R == MustAlias) { 433 // If there is at least one non-zero constant index, we know they cannot 434 // alias. 435 bool ConstantFound = false; 436 bool AllZerosFound = true; 437 for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i) 438 if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) { 439 if (!C->isNullValue()) { 440 ConstantFound = true; 441 AllZerosFound = false; 442 break; 443 } 444 } else { 445 AllZerosFound = false; 446 } 447 448 // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases 449 // the ptr, the end result is a must alias also. 450 if (AllZerosFound) 451 return MustAlias; 452 453 if (ConstantFound) { 454 if (V2Size <= 1 && V1Size <= 1) // Just pointer check? 455 return NoAlias; 456 457 // Otherwise we have to check to see that the distance is more than 458 // the size of the argument... build an index vector that is equal to 459 // the arguments provided, except substitute 0's for any variable 460 // indexes we find... 461 if (TD && 462 cast<PointerType>(BasePtr->getType())->getElementType()->isSized()) { 463 for (unsigned i = 0; i != GEPOperands.size(); ++i) 464 if (!isa<ConstantInt>(GEPOperands[i])) 465 GEPOperands[i] = Constant::getNullValue(GEPOperands[i]->getType()); 466 int64_t Offset = 467 TD->getIndexedOffset(BasePtr->getType(), 468 &GEPOperands[0], 469 GEPOperands.size()); 470 471 if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size) 472 return NoAlias; 473 } 474 } 475 } 476 477 return MayAlias; 478} 479 480AliasAnalysis::AliasResult 481BasicAliasAnalysis::aliasPHI(const Value *V1, unsigned V1Size, 482 const Value *V2, unsigned V2Size, 483 SmallSet<const Value*, 16> &VisitedPHIs) { 484 // The PHI node has already been visited, avoid recursion any further. 485 if (!VisitedPHIs.insert(V1)) 486 return MayAlias; 487 488 SmallSet<Value*, 4> UniqueSrc; 489 SmallVector<Value*, 4> V1Srcs; 490 const PHINode *PN = cast<PHINode>(V1); 491 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 492 Value *PV1 = PN->getIncomingValue(i); 493 if (isa<PHINode>(PV1)) 494 // If any of the source itself is a PHI, return MayAlias conservatively 495 // to avoid compile time explosion. 496 return MayAlias; 497 if (UniqueSrc.insert(PV1)) 498 V1Srcs.push_back(PV1); 499 } 500 501 // If all sources of the PHI node NoAlias or MustAlias V2, then returns 502 // NoAlias / MustAlias. Otherwise, returns MayAlias. 503 AliasResult Alias = aliasCheck(V1Srcs[0], V1Size, V2, V2Size, VisitedPHIs); 504 for (unsigned i = 1, e = V1Srcs.size(); i != e; ++i) { 505 Value *V = V1Srcs[i]; 506 AliasResult ThisAlias = aliasCheck(V, V1Size, V2, V2Size, VisitedPHIs); 507 if (ThisAlias != Alias) 508 return MayAlias; 509 } 510 511 return Alias; 512} 513 514// aliasCheck - Provide a bunch of ad-hoc rules to disambiguate in common cases, 515// such as array references. 516// 517AliasAnalysis::AliasResult 518BasicAliasAnalysis::aliasCheck(const Value *V1, unsigned V1Size, 519 const Value *V2, unsigned V2Size, 520 SmallSet<const Value*, 16> &VisitedPHIs) { 521 // Strip off any casts if they exist. 522 V1 = V1->stripPointerCasts(); 523 V2 = V2->stripPointerCasts(); 524 525 // Are we checking for alias of the same value? 526 if (V1 == V2) return MustAlias; 527 528 if (!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType())) 529 return NoAlias; // Scalars cannot alias each other 530 531 // Figure out what objects these things are pointing to if we can. 532 const Value *O1 = V1->getUnderlyingObject(); 533 const Value *O2 = V2->getUnderlyingObject(); 534 535 if (O1 != O2) { 536 // If V1/V2 point to two different objects we know that we have no alias. 537 if (isIdentifiedObject(O1) && isIdentifiedObject(O2)) 538 return NoAlias; 539 540 // Arguments can't alias with local allocations or noalias calls. 541 if ((isa<Argument>(O1) && (isa<AllocationInst>(O2) || isNoAliasCall(O2))) || 542 (isa<Argument>(O2) && (isa<AllocationInst>(O1) || isNoAliasCall(O1)))) 543 return NoAlias; 544 545 // Most objects can't alias null. 546 if ((isa<ConstantPointerNull>(V2) && isKnownNonNull(O1)) || 547 (isa<ConstantPointerNull>(V1) && isKnownNonNull(O2))) 548 return NoAlias; 549 } 550 551 // If the size of one access is larger than the entire object on the other 552 // side, then we know such behavior is undefined and can assume no alias. 553 LLVMContext &Context = V1->getContext(); 554 if (TD) 555 if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, Context, *TD)) || 556 (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, Context, *TD))) 557 return NoAlias; 558 559 // If one pointer is the result of a call/invoke and the other is a 560 // non-escaping local object, then we know the object couldn't escape to a 561 // point where the call could return it. 562 if ((isa<CallInst>(O1) || isa<InvokeInst>(O1)) && 563 isNonEscapingLocalObject(O2) && O1 != O2) 564 return NoAlias; 565 if ((isa<CallInst>(O2) || isa<InvokeInst>(O2)) && 566 isNonEscapingLocalObject(O1) && O1 != O2) 567 return NoAlias; 568 569 if (!isGEP(V1) && isGEP(V2)) { 570 std::swap(V1, V2); 571 std::swap(V1Size, V2Size); 572 } 573 if (isGEP(V1)) 574 return aliasGEP(V1, V1Size, V2, V2Size, VisitedPHIs); 575 576 if (isa<PHINode>(V2) && !isa<PHINode>(V1)) { 577 std::swap(V1, V2); 578 std::swap(V1Size, V2Size); 579 } 580 if (isa<PHINode>(V1)) 581 return aliasPHI(V1, V1Size, V2, V2Size, VisitedPHIs); 582 583 return MayAlias; 584} 585 586// This function is used to determine if the indices of two GEP instructions are 587// equal. V1 and V2 are the indices. 588static bool IndexOperandsEqual(Value *V1, Value *V2, LLVMContext &Context) { 589 if (V1->getType() == V2->getType()) 590 return V1 == V2; 591 if (Constant *C1 = dyn_cast<Constant>(V1)) 592 if (Constant *C2 = dyn_cast<Constant>(V2)) { 593 // Sign extend the constants to long types, if necessary 594 if (C1->getType() != Type::getInt64Ty(Context)) 595 C1 = ConstantExpr::getSExt(C1, Type::getInt64Ty(Context)); 596 if (C2->getType() != Type::getInt64Ty(Context)) 597 C2 = ConstantExpr::getSExt(C2, Type::getInt64Ty(Context)); 598 return C1 == C2; 599 } 600 return false; 601} 602 603/// CheckGEPInstructions - Check two GEP instructions with known must-aliasing 604/// base pointers. This checks to see if the index expressions preclude the 605/// pointers from aliasing... 606AliasAnalysis::AliasResult 607BasicAliasAnalysis::CheckGEPInstructions( 608 const Type* BasePtr1Ty, Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1S, 609 const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2S) { 610 // We currently can't handle the case when the base pointers have different 611 // primitive types. Since this is uncommon anyway, we are happy being 612 // extremely conservative. 613 if (BasePtr1Ty != BasePtr2Ty) 614 return MayAlias; 615 616 const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty); 617 618 LLVMContext &Context = GEPPointerTy->getContext(); 619 620 // Find the (possibly empty) initial sequence of equal values... which are not 621 // necessarily constants. 622 unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops; 623 unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands); 624 unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands); 625 unsigned UnequalOper = 0; 626 while (UnequalOper != MinOperands && 627 IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper], 628 Context)) { 629 // Advance through the type as we go... 630 ++UnequalOper; 631 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty)) 632 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]); 633 else { 634 // If all operands equal each other, then the derived pointers must 635 // alias each other... 636 BasePtr1Ty = 0; 637 assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands && 638 "Ran out of type nesting, but not out of operands?"); 639 return MustAlias; 640 } 641 } 642 643 // If we have seen all constant operands, and run out of indexes on one of the 644 // getelementptrs, check to see if the tail of the leftover one is all zeros. 645 // If so, return mustalias. 646 if (UnequalOper == MinOperands) { 647 if (NumGEP1Ops < NumGEP2Ops) { 648 std::swap(GEP1Ops, GEP2Ops); 649 std::swap(NumGEP1Ops, NumGEP2Ops); 650 } 651 652 bool AllAreZeros = true; 653 for (unsigned i = UnequalOper; i != MaxOperands; ++i) 654 if (!isa<Constant>(GEP1Ops[i]) || 655 !cast<Constant>(GEP1Ops[i])->isNullValue()) { 656 AllAreZeros = false; 657 break; 658 } 659 if (AllAreZeros) return MustAlias; 660 } 661 662 663 // So now we know that the indexes derived from the base pointers, 664 // which are known to alias, are different. We can still determine a 665 // no-alias result if there are differing constant pairs in the index 666 // chain. For example: 667 // A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S)) 668 // 669 // We have to be careful here about array accesses. In particular, consider: 670 // A[1][0] vs A[0][i] 671 // In this case, we don't *know* that the array will be accessed in bounds: 672 // the index could even be negative. Because of this, we have to 673 // conservatively *give up* and return may alias. We disregard differing 674 // array subscripts that are followed by a variable index without going 675 // through a struct. 676 // 677 unsigned SizeMax = std::max(G1S, G2S); 678 if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work. 679 680 // Scan for the first operand that is constant and unequal in the 681 // two getelementptrs... 682 unsigned FirstConstantOper = UnequalOper; 683 for (; FirstConstantOper != MinOperands; ++FirstConstantOper) { 684 const Value *G1Oper = GEP1Ops[FirstConstantOper]; 685 const Value *G2Oper = GEP2Ops[FirstConstantOper]; 686 687 if (G1Oper != G2Oper) // Found non-equal constant indexes... 688 if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper))) 689 if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){ 690 if (G1OC->getType() != G2OC->getType()) { 691 // Sign extend both operands to long. 692 if (G1OC->getType() != Type::getInt64Ty(Context)) 693 G1OC = ConstantExpr::getSExt(G1OC, Type::getInt64Ty(Context)); 694 if (G2OC->getType() != Type::getInt64Ty(Context)) 695 G2OC = ConstantExpr::getSExt(G2OC, Type::getInt64Ty(Context)); 696 GEP1Ops[FirstConstantOper] = G1OC; 697 GEP2Ops[FirstConstantOper] = G2OC; 698 } 699 700 if (G1OC != G2OC) { 701 // Handle the "be careful" case above: if this is an array/vector 702 // subscript, scan for a subsequent variable array index. 703 if (const SequentialType *STy = 704 dyn_cast<SequentialType>(BasePtr1Ty)) { 705 const Type *NextTy = STy; 706 bool isBadCase = false; 707 708 for (unsigned Idx = FirstConstantOper; 709 Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) { 710 const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx]; 711 if (!isa<Constant>(V1) || !isa<Constant>(V2)) { 712 isBadCase = true; 713 break; 714 } 715 // If the array is indexed beyond the bounds of the static type 716 // at this level, it will also fall into the "be careful" case. 717 // It would theoretically be possible to analyze these cases, 718 // but for now just be conservatively correct. 719 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 720 if (cast<ConstantInt>(G1OC)->getZExtValue() >= 721 ATy->getNumElements() || 722 cast<ConstantInt>(G2OC)->getZExtValue() >= 723 ATy->getNumElements()) { 724 isBadCase = true; 725 break; 726 } 727 if (const VectorType *VTy = dyn_cast<VectorType>(STy)) 728 if (cast<ConstantInt>(G1OC)->getZExtValue() >= 729 VTy->getNumElements() || 730 cast<ConstantInt>(G2OC)->getZExtValue() >= 731 VTy->getNumElements()) { 732 isBadCase = true; 733 break; 734 } 735 STy = cast<SequentialType>(NextTy); 736 NextTy = cast<SequentialType>(NextTy)->getElementType(); 737 } 738 739 if (isBadCase) G1OC = 0; 740 } 741 742 // Make sure they are comparable (ie, not constant expressions), and 743 // make sure the GEP with the smaller leading constant is GEP1. 744 if (G1OC) { 745 Constant *Compare = ConstantExpr::getICmp(ICmpInst::ICMP_SGT, 746 G1OC, G2OC); 747 if (ConstantInt *CV = dyn_cast<ConstantInt>(Compare)) { 748 if (CV->getZExtValue()) { // If they are comparable and G2 > G1 749 std::swap(GEP1Ops, GEP2Ops); // Make GEP1 < GEP2 750 std::swap(NumGEP1Ops, NumGEP2Ops); 751 } 752 break; 753 } 754 } 755 } 756 } 757 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper); 758 } 759 760 // No shared constant operands, and we ran out of common operands. At this 761 // point, the GEP instructions have run through all of their operands, and we 762 // haven't found evidence that there are any deltas between the GEP's. 763 // However, one GEP may have more operands than the other. If this is the 764 // case, there may still be hope. Check this now. 765 if (FirstConstantOper == MinOperands) { 766 // Without TargetData, we won't know what the offsets are. 767 if (!TD) 768 return MayAlias; 769 770 // Make GEP1Ops be the longer one if there is a longer one. 771 if (NumGEP1Ops < NumGEP2Ops) { 772 std::swap(GEP1Ops, GEP2Ops); 773 std::swap(NumGEP1Ops, NumGEP2Ops); 774 } 775 776 // Is there anything to check? 777 if (NumGEP1Ops > MinOperands) { 778 for (unsigned i = FirstConstantOper; i != MaxOperands; ++i) 779 if (isa<ConstantInt>(GEP1Ops[i]) && 780 !cast<ConstantInt>(GEP1Ops[i])->isZero()) { 781 // Yup, there's a constant in the tail. Set all variables to 782 // constants in the GEP instruction to make it suitable for 783 // TargetData::getIndexedOffset. 784 for (i = 0; i != MaxOperands; ++i) 785 if (!isa<ConstantInt>(GEP1Ops[i])) 786 GEP1Ops[i] = Constant::getNullValue(GEP1Ops[i]->getType()); 787 // Okay, now get the offset. This is the relative offset for the full 788 // instruction. 789 int64_t Offset1 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops, 790 NumGEP1Ops); 791 792 // Now check without any constants at the end. 793 int64_t Offset2 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops, 794 MinOperands); 795 796 // Make sure we compare the absolute difference. 797 if (Offset1 > Offset2) 798 std::swap(Offset1, Offset2); 799 800 // If the tail provided a bit enough offset, return noalias! 801 if ((uint64_t)(Offset2-Offset1) >= SizeMax) 802 return NoAlias; 803 // Otherwise break - we don't look for another constant in the tail. 804 break; 805 } 806 } 807 808 // Couldn't find anything useful. 809 return MayAlias; 810 } 811 812 // If there are non-equal constants arguments, then we can figure 813 // out a minimum known delta between the two index expressions... at 814 // this point we know that the first constant index of GEP1 is less 815 // than the first constant index of GEP2. 816 817 // Advance BasePtr[12]Ty over this first differing constant operand. 818 BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)-> 819 getTypeAtIndex(GEP2Ops[FirstConstantOper]); 820 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)-> 821 getTypeAtIndex(GEP1Ops[FirstConstantOper]); 822 823 // We are going to be using TargetData::getIndexedOffset to determine the 824 // offset that each of the GEP's is reaching. To do this, we have to convert 825 // all variable references to constant references. To do this, we convert the 826 // initial sequence of array subscripts into constant zeros to start with. 827 const Type *ZeroIdxTy = GEPPointerTy; 828 for (unsigned i = 0; i != FirstConstantOper; ++i) { 829 if (!isa<StructType>(ZeroIdxTy)) 830 GEP1Ops[i] = GEP2Ops[i] = 831 Constant::getNullValue(Type::getInt32Ty(Context)); 832 833 if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy)) 834 ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]); 835 } 836 837 // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok 838 839 // Loop over the rest of the operands... 840 for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) { 841 const Value *Op1 = i < NumGEP1Ops ? GEP1Ops[i] : 0; 842 const Value *Op2 = i < NumGEP2Ops ? GEP2Ops[i] : 0; 843 // If they are equal, use a zero index... 844 if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) { 845 if (!isa<ConstantInt>(Op1)) 846 GEP1Ops[i] = GEP2Ops[i] = Constant::getNullValue(Op1->getType()); 847 // Otherwise, just keep the constants we have. 848 } else { 849 if (Op1) { 850 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 851 // If this is an array index, make sure the array element is in range. 852 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) { 853 if (Op1C->getZExtValue() >= AT->getNumElements()) 854 return MayAlias; // Be conservative with out-of-range accesses 855 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) { 856 if (Op1C->getZExtValue() >= VT->getNumElements()) 857 return MayAlias; // Be conservative with out-of-range accesses 858 } 859 860 } else { 861 // GEP1 is known to produce a value less than GEP2. To be 862 // conservatively correct, we must assume the largest possible 863 // constant is used in this position. This cannot be the initial 864 // index to the GEP instructions (because we know we have at least one 865 // element before this one with the different constant arguments), so 866 // we know that the current index must be into either a struct or 867 // array. Because we know it's not constant, this cannot be a 868 // structure index. Because of this, we can calculate the maximum 869 // value possible. 870 // 871 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) 872 GEP1Ops[i] = 873 ConstantInt::get(Type::getInt64Ty(Context), 874 AT->getNumElements()-1); 875 else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) 876 GEP1Ops[i] = 877 ConstantInt::get(Type::getInt64Ty(Context), 878 VT->getNumElements()-1); 879 } 880 } 881 882 if (Op2) { 883 if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) { 884 // If this is an array index, make sure the array element is in range. 885 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr2Ty)) { 886 if (Op2C->getZExtValue() >= AT->getNumElements()) 887 return MayAlias; // Be conservative with out-of-range accesses 888 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr2Ty)) { 889 if (Op2C->getZExtValue() >= VT->getNumElements()) 890 return MayAlias; // Be conservative with out-of-range accesses 891 } 892 } else { // Conservatively assume the minimum value for this index 893 GEP2Ops[i] = Constant::getNullValue(Op2->getType()); 894 } 895 } 896 } 897 898 if (BasePtr1Ty && Op1) { 899 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty)) 900 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]); 901 else 902 BasePtr1Ty = 0; 903 } 904 905 if (BasePtr2Ty && Op2) { 906 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty)) 907 BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]); 908 else 909 BasePtr2Ty = 0; 910 } 911 } 912 913 if (TD && GEPPointerTy->getElementType()->isSized()) { 914 int64_t Offset1 = 915 TD->getIndexedOffset(GEPPointerTy, GEP1Ops, NumGEP1Ops); 916 int64_t Offset2 = 917 TD->getIndexedOffset(GEPPointerTy, GEP2Ops, NumGEP2Ops); 918 assert(Offset1 != Offset2 && 919 "There is at least one different constant here!"); 920 921 // Make sure we compare the absolute difference. 922 if (Offset1 > Offset2) 923 std::swap(Offset1, Offset2); 924 925 if ((uint64_t)(Offset2-Offset1) >= SizeMax) { 926 //cerr << "Determined that these two GEP's don't alias [" 927 // << SizeMax << " bytes]: \n" << *GEP1 << *GEP2; 928 return NoAlias; 929 } 930 } 931 return MayAlias; 932} 933 934// Make sure that anything that uses AliasAnalysis pulls in this file... 935DEFINING_FILE_FOR(BasicAliasAnalysis) 936