BasicAliasAnalysis.cpp revision fc2a3ed0c9e32cf7edaf5030fa0972b916cc5f0b
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/Passes.h" 19#include "llvm/Constants.h" 20#include "llvm/DerivedTypes.h" 21#include "llvm/Function.h" 22#include "llvm/GlobalVariable.h" 23#include "llvm/Instructions.h" 24#include "llvm/IntrinsicInst.h" 25#include "llvm/LLVMContext.h" 26#include "llvm/Operator.h" 27#include "llvm/Pass.h" 28#include "llvm/Target/TargetData.h" 29#include "llvm/ADT/SmallVector.h" 30#include "llvm/ADT/STLExtras.h" 31#include "llvm/Support/Compiler.h" 32#include "llvm/Support/ErrorHandling.h" 33#include "llvm/Support/GetElementPtrTypeIterator.h" 34#include <algorithm> 35using namespace llvm; 36 37//===----------------------------------------------------------------------===// 38// Useful predicates 39//===----------------------------------------------------------------------===// 40 41static const GEPOperator *isGEP(const Value *V) { 42 return dyn_cast<GEPOperator>(V); 43} 44 45static const Value *GetGEPOperands(const Value *V, 46 SmallVector<Value*, 16> &GEPOps) { 47 assert(GEPOps.empty() && "Expect empty list to populate!"); 48 GEPOps.insert(GEPOps.end(), cast<User>(V)->op_begin()+1, 49 cast<User>(V)->op_end()); 50 51 // Accumulate all of the chained indexes into the operand array 52 V = cast<User>(V)->getOperand(0); 53 54 while (const User *G = isGEP(V)) { 55 if (!isa<Constant>(GEPOps[0]) || isa<GlobalValue>(GEPOps[0]) || 56 !cast<Constant>(GEPOps[0])->isNullValue()) 57 break; // Don't handle folding arbitrary pointer offsets yet... 58 GEPOps.erase(GEPOps.begin()); // Drop the zero index 59 GEPOps.insert(GEPOps.begin(), G->op_begin()+1, G->op_end()); 60 V = G->getOperand(0); 61 } 62 return V; 63} 64 65/// isKnownNonNull - Return true if we know that the specified value is never 66/// null. 67static bool isKnownNonNull(const Value *V) { 68 // Alloca never returns null, malloc might. 69 if (isa<AllocaInst>(V)) return true; 70 71 // A byval argument is never null. 72 if (const Argument *A = dyn_cast<Argument>(V)) 73 return A->hasByValAttr(); 74 75 // Global values are not null unless extern weak. 76 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) 77 return !GV->hasExternalWeakLinkage(); 78 return false; 79} 80 81/// isNonEscapingLocalObject - Return true if the pointer is to a function-local 82/// object that never escapes from the function. 83static bool isNonEscapingLocalObject(const Value *V) { 84 // If this is a local allocation, check to see if it escapes. 85 if (isa<AllocationInst>(V) || isNoAliasCall(V)) 86 return !PointerMayBeCaptured(V, false); 87 88 // If this is an argument that corresponds to a byval or noalias argument, 89 // then it has not escaped before entering the function. Check if it escapes 90 // inside the function. 91 if (const Argument *A = dyn_cast<Argument>(V)) 92 if (A->hasByValAttr() || A->hasNoAliasAttr()) { 93 // Don't bother analyzing arguments already known not to escape. 94 if (A->hasNoCaptureAttr()) 95 return true; 96 return !PointerMayBeCaptured(V, false); 97 } 98 return false; 99} 100 101 102/// isObjectSmallerThan - Return true if we can prove that the object specified 103/// by V is smaller than Size. 104static bool isObjectSmallerThan(const Value *V, unsigned Size, 105 const TargetData &TD) { 106 const Type *AccessTy; 107 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) { 108 AccessTy = GV->getType()->getElementType(); 109 } else if (const AllocationInst *AI = dyn_cast<AllocationInst>(V)) { 110 if (!AI->isArrayAllocation()) 111 AccessTy = AI->getType()->getElementType(); 112 else 113 return false; 114 } else if (const Argument *A = dyn_cast<Argument>(V)) { 115 if (A->hasByValAttr()) 116 AccessTy = cast<PointerType>(A->getType())->getElementType(); 117 else 118 return false; 119 } else { 120 return false; 121 } 122 123 if (AccessTy->isSized()) 124 return TD.getTypeAllocSize(AccessTy) < Size; 125 return false; 126} 127 128//===----------------------------------------------------------------------===// 129// NoAA Pass 130//===----------------------------------------------------------------------===// 131 132namespace { 133 /// NoAA - This class implements the -no-aa pass, which always returns "I 134 /// don't know" for alias queries. NoAA is unlike other alias analysis 135 /// implementations, in that it does not chain to a previous analysis. As 136 /// such it doesn't follow many of the rules that other alias analyses must. 137 /// 138 struct VISIBILITY_HIDDEN NoAA : public ImmutablePass, public AliasAnalysis { 139 static char ID; // Class identification, replacement for typeinfo 140 NoAA() : ImmutablePass(&ID) {} 141 explicit NoAA(void *PID) : ImmutablePass(PID) { } 142 143 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 144 } 145 146 virtual void initializePass() { 147 TD = getAnalysisIfAvailable<TargetData>(); 148 } 149 150 virtual AliasResult alias(const Value *V1, unsigned V1Size, 151 const Value *V2, unsigned V2Size) { 152 return MayAlias; 153 } 154 155 virtual void getArgumentAccesses(Function *F, CallSite CS, 156 std::vector<PointerAccessInfo> &Info) { 157 llvm_unreachable("This method may not be called on this function!"); 158 } 159 160 virtual void getMustAliases(Value *P, std::vector<Value*> &RetVals) { } 161 virtual bool pointsToConstantMemory(const Value *P) { return false; } 162 virtual ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size) { 163 return ModRef; 164 } 165 virtual ModRefResult getModRefInfo(CallSite CS1, CallSite CS2) { 166 return ModRef; 167 } 168 virtual bool hasNoModRefInfoForCalls() const { return true; } 169 170 virtual void deleteValue(Value *V) {} 171 virtual void copyValue(Value *From, Value *To) {} 172 }; 173} // End of anonymous namespace 174 175// Register this pass... 176char NoAA::ID = 0; 177static RegisterPass<NoAA> 178U("no-aa", "No Alias Analysis (always returns 'may' alias)", true, true); 179 180// Declare that we implement the AliasAnalysis interface 181static RegisterAnalysisGroup<AliasAnalysis> V(U); 182 183ImmutablePass *llvm::createNoAAPass() { return new NoAA(); } 184 185//===----------------------------------------------------------------------===// 186// BasicAA Pass 187//===----------------------------------------------------------------------===// 188 189namespace { 190 /// BasicAliasAnalysis - This is the default alias analysis implementation. 191 /// Because it doesn't chain to a previous alias analysis (like -no-aa), it 192 /// derives from the NoAA class. 193 struct VISIBILITY_HIDDEN BasicAliasAnalysis : public NoAA { 194 static char ID; // Class identification, replacement for typeinfo 195 BasicAliasAnalysis() : NoAA(&ID) {} 196 AliasResult alias(const Value *V1, unsigned V1Size, 197 const Value *V2, unsigned V2Size); 198 199 ModRefResult getModRefInfo(CallSite CS, Value *P, unsigned Size); 200 ModRefResult getModRefInfo(CallSite CS1, CallSite CS2); 201 202 /// hasNoModRefInfoForCalls - We can provide mod/ref information against 203 /// non-escaping allocations. 204 virtual bool hasNoModRefInfoForCalls() const { return false; } 205 206 /// pointsToConstantMemory - Chase pointers until we find a (constant 207 /// global) or not. 208 bool pointsToConstantMemory(const Value *P); 209 210 private: 211 // CheckGEPInstructions - Check two GEP instructions with known 212 // must-aliasing base pointers. This checks to see if the index expressions 213 // preclude the pointers from aliasing... 214 AliasResult 215 CheckGEPInstructions(const Type* BasePtr1Ty, 216 Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1Size, 217 const Type *BasePtr2Ty, 218 Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2Size); 219 }; 220} // End of anonymous namespace 221 222// Register this pass... 223char BasicAliasAnalysis::ID = 0; 224static RegisterPass<BasicAliasAnalysis> 225X("basicaa", "Basic Alias Analysis (default AA impl)", false, true); 226 227// Declare that we implement the AliasAnalysis interface 228static RegisterAnalysisGroup<AliasAnalysis, true> Y(X); 229 230ImmutablePass *llvm::createBasicAliasAnalysisPass() { 231 return new BasicAliasAnalysis(); 232} 233 234 235/// pointsToConstantMemory - Chase pointers until we find a (constant 236/// global) or not. 237bool BasicAliasAnalysis::pointsToConstantMemory(const Value *P) { 238 if (const GlobalVariable *GV = 239 dyn_cast<GlobalVariable>(P->getUnderlyingObject())) 240 return GV->isConstant(); 241 return false; 242} 243 244 245// getModRefInfo - Check to see if the specified callsite can clobber the 246// specified memory object. Since we only look at local properties of this 247// function, we really can't say much about this query. We do, however, use 248// simple "address taken" analysis on local objects. 249// 250AliasAnalysis::ModRefResult 251BasicAliasAnalysis::getModRefInfo(CallSite CS, Value *P, unsigned Size) { 252 if (!isa<Constant>(P)) { 253 const Value *Object = P->getUnderlyingObject(); 254 255 // If this is a tail call and P points to a stack location, we know that 256 // the tail call cannot access or modify the local stack. 257 // We cannot exclude byval arguments here; these belong to the caller of 258 // the current function not to the current function, and a tail callee 259 // may reference them. 260 if (isa<AllocaInst>(Object)) 261 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) 262 if (CI->isTailCall()) 263 return NoModRef; 264 265 // If the pointer is to a locally allocated object that does not escape, 266 // then the call can not mod/ref the pointer unless the call takes the 267 // argument without capturing it. 268 if (isNonEscapingLocalObject(Object) && CS.getInstruction() != Object) { 269 bool passedAsArg = false; 270 // TODO: Eventually only check 'nocapture' arguments. 271 for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end(); 272 CI != CE; ++CI) 273 if (isa<PointerType>((*CI)->getType()) && 274 alias(cast<Value>(CI), ~0U, P, ~0U) != NoAlias) 275 passedAsArg = true; 276 277 if (!passedAsArg) 278 return NoModRef; 279 } 280 } 281 282 // The AliasAnalysis base class has some smarts, lets use them. 283 return AliasAnalysis::getModRefInfo(CS, P, Size); 284} 285 286 287AliasAnalysis::ModRefResult 288BasicAliasAnalysis::getModRefInfo(CallSite CS1, CallSite CS2) { 289 // If CS1 or CS2 are readnone, they don't interact. 290 ModRefBehavior CS1B = AliasAnalysis::getModRefBehavior(CS1); 291 if (CS1B == DoesNotAccessMemory) return NoModRef; 292 293 ModRefBehavior CS2B = AliasAnalysis::getModRefBehavior(CS2); 294 if (CS2B == DoesNotAccessMemory) return NoModRef; 295 296 // If they both only read from memory, just return ref. 297 if (CS1B == OnlyReadsMemory && CS2B == OnlyReadsMemory) 298 return Ref; 299 300 // Otherwise, fall back to NoAA (mod+ref). 301 return NoAA::getModRefInfo(CS1, CS2); 302} 303 304 305// alias - Provide a bunch of ad-hoc rules to disambiguate in common cases, such 306// as array references. 307// 308AliasAnalysis::AliasResult 309BasicAliasAnalysis::alias(const Value *V1, unsigned V1Size, 310 const Value *V2, unsigned V2Size) { 311 LLVMContext &Context = V1->getType()->getContext(); 312 313 // Strip off any constant expression casts if they exist 314 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V1)) 315 if (CE->isCast() && isa<PointerType>(CE->getOperand(0)->getType())) 316 V1 = CE->getOperand(0); 317 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V2)) 318 if (CE->isCast() && isa<PointerType>(CE->getOperand(0)->getType())) 319 V2 = CE->getOperand(0); 320 321 // Are we checking for alias of the same value? 322 if (V1 == V2) return MustAlias; 323 324 if (!isa<PointerType>(V1->getType()) || !isa<PointerType>(V2->getType())) 325 return NoAlias; // Scalars cannot alias each other 326 327 // Strip off cast instructions. Since V1 and V2 are pointers, they must be 328 // pointer<->pointer bitcasts. 329 if (const BitCastInst *I = dyn_cast<BitCastInst>(V1)) 330 return alias(I->getOperand(0), V1Size, V2, V2Size); 331 if (const BitCastInst *I = dyn_cast<BitCastInst>(V2)) 332 return alias(V1, V1Size, I->getOperand(0), V2Size); 333 334 // Figure out what objects these things are pointing to if we can. 335 const Value *O1 = V1->getUnderlyingObject(); 336 const Value *O2 = V2->getUnderlyingObject(); 337 338 if (O1 != O2) { 339 // If V1/V2 point to two different objects we know that we have no alias. 340 if (isIdentifiedObject(O1) && isIdentifiedObject(O2)) 341 return NoAlias; 342 343 // Arguments can't alias with local allocations or noalias calls. 344 if ((isa<Argument>(O1) && (isa<AllocationInst>(O2) || isNoAliasCall(O2))) || 345 (isa<Argument>(O2) && (isa<AllocationInst>(O1) || isNoAliasCall(O1)))) 346 return NoAlias; 347 348 // Most objects can't alias null. 349 if ((isa<ConstantPointerNull>(V2) && isKnownNonNull(O1)) || 350 (isa<ConstantPointerNull>(V1) && isKnownNonNull(O2))) 351 return NoAlias; 352 } 353 354 // If the size of one access is larger than the entire object on the other 355 // side, then we know such behavior is undefined and can assume no alias. 356 if (TD) 357 if ((V1Size != ~0U && isObjectSmallerThan(O2, V1Size, *TD)) || 358 (V2Size != ~0U && isObjectSmallerThan(O1, V2Size, *TD))) 359 return NoAlias; 360 361 // If one pointer is the result of a call/invoke and the other is a 362 // non-escaping local object, then we know the object couldn't escape to a 363 // point where the call could return it. 364 if ((isa<CallInst>(O1) || isa<InvokeInst>(O1)) && 365 isNonEscapingLocalObject(O2) && O1 != O2) 366 return NoAlias; 367 if ((isa<CallInst>(O2) || isa<InvokeInst>(O2)) && 368 isNonEscapingLocalObject(O1) && O1 != O2) 369 return NoAlias; 370 371 // If we have two gep instructions with must-alias'ing base pointers, figure 372 // out if the indexes to the GEP tell us anything about the derived pointer. 373 // Note that we also handle chains of getelementptr instructions as well as 374 // constant expression getelementptrs here. 375 // 376 if (isGEP(V1) && isGEP(V2)) { 377 const User *GEP1 = cast<User>(V1); 378 const User *GEP2 = cast<User>(V2); 379 380 // If V1 and V2 are identical GEPs, just recurse down on both of them. 381 // This allows us to analyze things like: 382 // P = gep A, 0, i, 1 383 // Q = gep B, 0, i, 1 384 // by just analyzing A and B. This is even safe for variable indices. 385 if (GEP1->getType() == GEP2->getType() && 386 GEP1->getNumOperands() == GEP2->getNumOperands() && 387 GEP1->getOperand(0)->getType() == GEP2->getOperand(0)->getType() && 388 // All operands are the same, ignoring the base. 389 std::equal(GEP1->op_begin()+1, GEP1->op_end(), GEP2->op_begin()+1)) 390 return alias(GEP1->getOperand(0), V1Size, GEP2->getOperand(0), V2Size); 391 392 393 // Drill down into the first non-gep value, to test for must-aliasing of 394 // the base pointers. 395 while (isGEP(GEP1->getOperand(0)) && 396 GEP1->getOperand(1) == 397 Context.getNullValue(GEP1->getOperand(1)->getType())) 398 GEP1 = cast<User>(GEP1->getOperand(0)); 399 const Value *BasePtr1 = GEP1->getOperand(0); 400 401 while (isGEP(GEP2->getOperand(0)) && 402 GEP2->getOperand(1) == 403 Context.getNullValue(GEP2->getOperand(1)->getType())) 404 GEP2 = cast<User>(GEP2->getOperand(0)); 405 const Value *BasePtr2 = GEP2->getOperand(0); 406 407 // Do the base pointers alias? 408 AliasResult BaseAlias = alias(BasePtr1, ~0U, BasePtr2, ~0U); 409 if (BaseAlias == NoAlias) return NoAlias; 410 if (BaseAlias == MustAlias) { 411 // If the base pointers alias each other exactly, check to see if we can 412 // figure out anything about the resultant pointers, to try to prove 413 // non-aliasing. 414 415 // Collect all of the chained GEP operands together into one simple place 416 SmallVector<Value*, 16> GEP1Ops, GEP2Ops; 417 BasePtr1 = GetGEPOperands(V1, GEP1Ops); 418 BasePtr2 = GetGEPOperands(V2, GEP2Ops); 419 420 // If GetGEPOperands were able to fold to the same must-aliased pointer, 421 // do the comparison. 422 if (BasePtr1 == BasePtr2) { 423 AliasResult GAlias = 424 CheckGEPInstructions(BasePtr1->getType(), 425 &GEP1Ops[0], GEP1Ops.size(), V1Size, 426 BasePtr2->getType(), 427 &GEP2Ops[0], GEP2Ops.size(), V2Size); 428 if (GAlias != MayAlias) 429 return GAlias; 430 } 431 } 432 } 433 434 // Check to see if these two pointers are related by a getelementptr 435 // instruction. If one pointer is a GEP with a non-zero index of the other 436 // pointer, we know they cannot alias. 437 // 438 if (isGEP(V2)) { 439 std::swap(V1, V2); 440 std::swap(V1Size, V2Size); 441 } 442 443 if (V1Size != ~0U && V2Size != ~0U) 444 if (isGEP(V1)) { 445 SmallVector<Value*, 16> GEPOperands; 446 const Value *BasePtr = GetGEPOperands(V1, GEPOperands); 447 448 AliasResult R = alias(BasePtr, V1Size, V2, V2Size); 449 if (R == MustAlias) { 450 // If there is at least one non-zero constant index, we know they cannot 451 // alias. 452 bool ConstantFound = false; 453 bool AllZerosFound = true; 454 for (unsigned i = 0, e = GEPOperands.size(); i != e; ++i) 455 if (const Constant *C = dyn_cast<Constant>(GEPOperands[i])) { 456 if (!C->isNullValue()) { 457 ConstantFound = true; 458 AllZerosFound = false; 459 break; 460 } 461 } else { 462 AllZerosFound = false; 463 } 464 465 // If we have getelementptr <ptr>, 0, 0, 0, 0, ... and V2 must aliases 466 // the ptr, the end result is a must alias also. 467 if (AllZerosFound) 468 return MustAlias; 469 470 if (ConstantFound) { 471 if (V2Size <= 1 && V1Size <= 1) // Just pointer check? 472 return NoAlias; 473 474 // Otherwise we have to check to see that the distance is more than 475 // the size of the argument... build an index vector that is equal to 476 // the arguments provided, except substitute 0's for any variable 477 // indexes we find... 478 if (TD && cast<PointerType>( 479 BasePtr->getType())->getElementType()->isSized()) { 480 for (unsigned i = 0; i != GEPOperands.size(); ++i) 481 if (!isa<ConstantInt>(GEPOperands[i])) 482 GEPOperands[i] = 483 Context.getNullValue(GEPOperands[i]->getType()); 484 int64_t Offset = 485 TD->getIndexedOffset(BasePtr->getType(), 486 &GEPOperands[0], 487 GEPOperands.size()); 488 489 if (Offset >= (int64_t)V2Size || Offset <= -(int64_t)V1Size) 490 return NoAlias; 491 } 492 } 493 } 494 } 495 496 return MayAlias; 497} 498 499// This function is used to determine if the indices of two GEP instructions are 500// equal. V1 and V2 are the indices. 501static bool IndexOperandsEqual(Value *V1, Value *V2, LLVMContext &Context) { 502 if (V1->getType() == V2->getType()) 503 return V1 == V2; 504 if (Constant *C1 = dyn_cast<Constant>(V1)) 505 if (Constant *C2 = dyn_cast<Constant>(V2)) { 506 // Sign extend the constants to long types, if necessary 507 if (C1->getType() != Type::Int64Ty) 508 C1 = Context.getConstantExprSExt(C1, Type::Int64Ty); 509 if (C2->getType() != Type::Int64Ty) 510 C2 = Context.getConstantExprSExt(C2, Type::Int64Ty); 511 return C1 == C2; 512 } 513 return false; 514} 515 516/// CheckGEPInstructions - Check two GEP instructions with known must-aliasing 517/// base pointers. This checks to see if the index expressions preclude the 518/// pointers from aliasing... 519AliasAnalysis::AliasResult 520BasicAliasAnalysis::CheckGEPInstructions( 521 const Type* BasePtr1Ty, Value **GEP1Ops, unsigned NumGEP1Ops, unsigned G1S, 522 const Type *BasePtr2Ty, Value **GEP2Ops, unsigned NumGEP2Ops, unsigned G2S) { 523 // We currently can't handle the case when the base pointers have different 524 // primitive types. Since this is uncommon anyway, we are happy being 525 // extremely conservative. 526 if (BasePtr1Ty != BasePtr2Ty) 527 return MayAlias; 528 529 const PointerType *GEPPointerTy = cast<PointerType>(BasePtr1Ty); 530 531 LLVMContext &Context = GEPPointerTy->getContext(); 532 533 // Find the (possibly empty) initial sequence of equal values... which are not 534 // necessarily constants. 535 unsigned NumGEP1Operands = NumGEP1Ops, NumGEP2Operands = NumGEP2Ops; 536 unsigned MinOperands = std::min(NumGEP1Operands, NumGEP2Operands); 537 unsigned MaxOperands = std::max(NumGEP1Operands, NumGEP2Operands); 538 unsigned UnequalOper = 0; 539 while (UnequalOper != MinOperands && 540 IndexOperandsEqual(GEP1Ops[UnequalOper], GEP2Ops[UnequalOper], 541 Context)) { 542 // Advance through the type as we go... 543 ++UnequalOper; 544 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty)) 545 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[UnequalOper-1]); 546 else { 547 // If all operands equal each other, then the derived pointers must 548 // alias each other... 549 BasePtr1Ty = 0; 550 assert(UnequalOper == NumGEP1Operands && UnequalOper == NumGEP2Operands && 551 "Ran out of type nesting, but not out of operands?"); 552 return MustAlias; 553 } 554 } 555 556 // If we have seen all constant operands, and run out of indexes on one of the 557 // getelementptrs, check to see if the tail of the leftover one is all zeros. 558 // If so, return mustalias. 559 if (UnequalOper == MinOperands) { 560 if (NumGEP1Ops < NumGEP2Ops) { 561 std::swap(GEP1Ops, GEP2Ops); 562 std::swap(NumGEP1Ops, NumGEP2Ops); 563 } 564 565 bool AllAreZeros = true; 566 for (unsigned i = UnequalOper; i != MaxOperands; ++i) 567 if (!isa<Constant>(GEP1Ops[i]) || 568 !cast<Constant>(GEP1Ops[i])->isNullValue()) { 569 AllAreZeros = false; 570 break; 571 } 572 if (AllAreZeros) return MustAlias; 573 } 574 575 576 // So now we know that the indexes derived from the base pointers, 577 // which are known to alias, are different. We can still determine a 578 // no-alias result if there are differing constant pairs in the index 579 // chain. For example: 580 // A[i][0] != A[j][1] iff (&A[0][1]-&A[0][0] >= std::max(G1S, G2S)) 581 // 582 // We have to be careful here about array accesses. In particular, consider: 583 // A[1][0] vs A[0][i] 584 // In this case, we don't *know* that the array will be accessed in bounds: 585 // the index could even be negative. Because of this, we have to 586 // conservatively *give up* and return may alias. We disregard differing 587 // array subscripts that are followed by a variable index without going 588 // through a struct. 589 // 590 unsigned SizeMax = std::max(G1S, G2S); 591 if (SizeMax == ~0U) return MayAlias; // Avoid frivolous work. 592 593 // Scan for the first operand that is constant and unequal in the 594 // two getelementptrs... 595 unsigned FirstConstantOper = UnequalOper; 596 for (; FirstConstantOper != MinOperands; ++FirstConstantOper) { 597 const Value *G1Oper = GEP1Ops[FirstConstantOper]; 598 const Value *G2Oper = GEP2Ops[FirstConstantOper]; 599 600 if (G1Oper != G2Oper) // Found non-equal constant indexes... 601 if (Constant *G1OC = dyn_cast<ConstantInt>(const_cast<Value*>(G1Oper))) 602 if (Constant *G2OC = dyn_cast<ConstantInt>(const_cast<Value*>(G2Oper))){ 603 if (G1OC->getType() != G2OC->getType()) { 604 // Sign extend both operands to long. 605 if (G1OC->getType() != Type::Int64Ty) 606 G1OC = Context.getConstantExprSExt(G1OC, Type::Int64Ty); 607 if (G2OC->getType() != Type::Int64Ty) 608 G2OC = Context.getConstantExprSExt(G2OC, Type::Int64Ty); 609 GEP1Ops[FirstConstantOper] = G1OC; 610 GEP2Ops[FirstConstantOper] = G2OC; 611 } 612 613 if (G1OC != G2OC) { 614 // Handle the "be careful" case above: if this is an array/vector 615 // subscript, scan for a subsequent variable array index. 616 if (const SequentialType *STy = 617 dyn_cast<SequentialType>(BasePtr1Ty)) { 618 const Type *NextTy = STy; 619 bool isBadCase = false; 620 621 for (unsigned Idx = FirstConstantOper; 622 Idx != MinOperands && isa<SequentialType>(NextTy); ++Idx) { 623 const Value *V1 = GEP1Ops[Idx], *V2 = GEP2Ops[Idx]; 624 if (!isa<Constant>(V1) || !isa<Constant>(V2)) { 625 isBadCase = true; 626 break; 627 } 628 // If the array is indexed beyond the bounds of the static type 629 // at this level, it will also fall into the "be careful" case. 630 // It would theoretically be possible to analyze these cases, 631 // but for now just be conservatively correct. 632 if (const ArrayType *ATy = dyn_cast<ArrayType>(STy)) 633 if (cast<ConstantInt>(G1OC)->getZExtValue() >= 634 ATy->getNumElements() || 635 cast<ConstantInt>(G2OC)->getZExtValue() >= 636 ATy->getNumElements()) { 637 isBadCase = true; 638 break; 639 } 640 if (const VectorType *VTy = dyn_cast<VectorType>(STy)) 641 if (cast<ConstantInt>(G1OC)->getZExtValue() >= 642 VTy->getNumElements() || 643 cast<ConstantInt>(G2OC)->getZExtValue() >= 644 VTy->getNumElements()) { 645 isBadCase = true; 646 break; 647 } 648 STy = cast<SequentialType>(NextTy); 649 NextTy = cast<SequentialType>(NextTy)->getElementType(); 650 } 651 652 if (isBadCase) G1OC = 0; 653 } 654 655 // Make sure they are comparable (ie, not constant expressions), and 656 // make sure the GEP with the smaller leading constant is GEP1. 657 if (G1OC) { 658 Constant *Compare = ConstantExpr::getICmp(ICmpInst::ICMP_SGT, 659 G1OC, G2OC); 660 if (ConstantInt *CV = dyn_cast<ConstantInt>(Compare)) { 661 if (CV->getZExtValue()) { // If they are comparable and G2 > G1 662 std::swap(GEP1Ops, GEP2Ops); // Make GEP1 < GEP2 663 std::swap(NumGEP1Ops, NumGEP2Ops); 664 } 665 break; 666 } 667 } 668 } 669 } 670 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)->getTypeAtIndex(G1Oper); 671 } 672 673 // No shared constant operands, and we ran out of common operands. At this 674 // point, the GEP instructions have run through all of their operands, and we 675 // haven't found evidence that there are any deltas between the GEP's. 676 // However, one GEP may have more operands than the other. If this is the 677 // case, there may still be hope. Check this now. 678 if (FirstConstantOper == MinOperands) { 679 // Without TargetData, we won't know what the offsets are. 680 if (!TD) 681 return MayAlias; 682 683 // Make GEP1Ops be the longer one if there is a longer one. 684 if (NumGEP1Ops < NumGEP2Ops) { 685 std::swap(GEP1Ops, GEP2Ops); 686 std::swap(NumGEP1Ops, NumGEP2Ops); 687 } 688 689 // Is there anything to check? 690 if (NumGEP1Ops > MinOperands) { 691 for (unsigned i = FirstConstantOper; i != MaxOperands; ++i) 692 if (isa<ConstantInt>(GEP1Ops[i]) && 693 !cast<ConstantInt>(GEP1Ops[i])->isZero()) { 694 // Yup, there's a constant in the tail. Set all variables to 695 // constants in the GEP instruction to make it suitable for 696 // TargetData::getIndexedOffset. 697 for (i = 0; i != MaxOperands; ++i) 698 if (!isa<ConstantInt>(GEP1Ops[i])) 699 GEP1Ops[i] = Context.getNullValue(GEP1Ops[i]->getType()); 700 // Okay, now get the offset. This is the relative offset for the full 701 // instruction. 702 int64_t Offset1 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops, 703 NumGEP1Ops); 704 705 // Now check without any constants at the end. 706 int64_t Offset2 = TD->getIndexedOffset(GEPPointerTy, GEP1Ops, 707 MinOperands); 708 709 // Make sure we compare the absolute difference. 710 if (Offset1 > Offset2) 711 std::swap(Offset1, Offset2); 712 713 // If the tail provided a bit enough offset, return noalias! 714 if ((uint64_t)(Offset2-Offset1) >= SizeMax) 715 return NoAlias; 716 // Otherwise break - we don't look for another constant in the tail. 717 break; 718 } 719 } 720 721 // Couldn't find anything useful. 722 return MayAlias; 723 } 724 725 // If there are non-equal constants arguments, then we can figure 726 // out a minimum known delta between the two index expressions... at 727 // this point we know that the first constant index of GEP1 is less 728 // than the first constant index of GEP2. 729 730 // Advance BasePtr[12]Ty over this first differing constant operand. 731 BasePtr2Ty = cast<CompositeType>(BasePtr1Ty)-> 732 getTypeAtIndex(GEP2Ops[FirstConstantOper]); 733 BasePtr1Ty = cast<CompositeType>(BasePtr1Ty)-> 734 getTypeAtIndex(GEP1Ops[FirstConstantOper]); 735 736 // We are going to be using TargetData::getIndexedOffset to determine the 737 // offset that each of the GEP's is reaching. To do this, we have to convert 738 // all variable references to constant references. To do this, we convert the 739 // initial sequence of array subscripts into constant zeros to start with. 740 const Type *ZeroIdxTy = GEPPointerTy; 741 for (unsigned i = 0; i != FirstConstantOper; ++i) { 742 if (!isa<StructType>(ZeroIdxTy)) 743 GEP1Ops[i] = GEP2Ops[i] = Context.getNullValue(Type::Int32Ty); 744 745 if (const CompositeType *CT = dyn_cast<CompositeType>(ZeroIdxTy)) 746 ZeroIdxTy = CT->getTypeAtIndex(GEP1Ops[i]); 747 } 748 749 // We know that GEP1Ops[FirstConstantOper] & GEP2Ops[FirstConstantOper] are ok 750 751 // Loop over the rest of the operands... 752 for (unsigned i = FirstConstantOper+1; i != MaxOperands; ++i) { 753 const Value *Op1 = i < NumGEP1Ops ? GEP1Ops[i] : 0; 754 const Value *Op2 = i < NumGEP2Ops ? GEP2Ops[i] : 0; 755 // If they are equal, use a zero index... 756 if (Op1 == Op2 && BasePtr1Ty == BasePtr2Ty) { 757 if (!isa<ConstantInt>(Op1)) 758 GEP1Ops[i] = GEP2Ops[i] = Context.getNullValue(Op1->getType()); 759 // Otherwise, just keep the constants we have. 760 } else { 761 if (Op1) { 762 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 763 // If this is an array index, make sure the array element is in range. 764 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) { 765 if (Op1C->getZExtValue() >= AT->getNumElements()) 766 return MayAlias; // Be conservative with out-of-range accesses 767 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) { 768 if (Op1C->getZExtValue() >= VT->getNumElements()) 769 return MayAlias; // Be conservative with out-of-range accesses 770 } 771 772 } else { 773 // GEP1 is known to produce a value less than GEP2. To be 774 // conservatively correct, we must assume the largest possible 775 // constant is used in this position. This cannot be the initial 776 // index to the GEP instructions (because we know we have at least one 777 // element before this one with the different constant arguments), so 778 // we know that the current index must be into either a struct or 779 // array. Because we know it's not constant, this cannot be a 780 // structure index. Because of this, we can calculate the maximum 781 // value possible. 782 // 783 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr1Ty)) 784 GEP1Ops[i] = 785 ConstantInt::get(Type::Int64Ty,AT->getNumElements()-1); 786 else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr1Ty)) 787 GEP1Ops[i] = 788 ConstantInt::get(Type::Int64Ty,VT->getNumElements()-1); 789 } 790 } 791 792 if (Op2) { 793 if (const ConstantInt *Op2C = dyn_cast<ConstantInt>(Op2)) { 794 // If this is an array index, make sure the array element is in range. 795 if (const ArrayType *AT = dyn_cast<ArrayType>(BasePtr2Ty)) { 796 if (Op2C->getZExtValue() >= AT->getNumElements()) 797 return MayAlias; // Be conservative with out-of-range accesses 798 } else if (const VectorType *VT = dyn_cast<VectorType>(BasePtr2Ty)) { 799 if (Op2C->getZExtValue() >= VT->getNumElements()) 800 return MayAlias; // Be conservative with out-of-range accesses 801 } 802 } else { // Conservatively assume the minimum value for this index 803 GEP2Ops[i] = Context.getNullValue(Op2->getType()); 804 } 805 } 806 } 807 808 if (BasePtr1Ty && Op1) { 809 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr1Ty)) 810 BasePtr1Ty = CT->getTypeAtIndex(GEP1Ops[i]); 811 else 812 BasePtr1Ty = 0; 813 } 814 815 if (BasePtr2Ty && Op2) { 816 if (const CompositeType *CT = dyn_cast<CompositeType>(BasePtr2Ty)) 817 BasePtr2Ty = CT->getTypeAtIndex(GEP2Ops[i]); 818 else 819 BasePtr2Ty = 0; 820 } 821 } 822 823 if (TD && GEPPointerTy->getElementType()->isSized()) { 824 int64_t Offset1 = 825 TD->getIndexedOffset(GEPPointerTy, GEP1Ops, NumGEP1Ops); 826 int64_t Offset2 = 827 TD->getIndexedOffset(GEPPointerTy, GEP2Ops, NumGEP2Ops); 828 assert(Offset1 != Offset2 && 829 "There is at least one different constant here!"); 830 831 // Make sure we compare the absolute difference. 832 if (Offset1 > Offset2) 833 std::swap(Offset1, Offset2); 834 835 if ((uint64_t)(Offset2-Offset1) >= SizeMax) { 836 //cerr << "Determined that these two GEP's don't alias [" 837 // << SizeMax << " bytes]: \n" << *GEP1 << *GEP2; 838 return NoAlias; 839 } 840 } 841 return MayAlias; 842} 843 844// Make sure that anything that uses AliasAnalysis pulls in this file... 845DEFINING_FILE_FOR(BasicAliasAnalysis) 846