1 //===------ MemoryBuiltins.cpp - Identify calls to memory builtins --------===// 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 family of functions identifies calls to builtin functions that allocate 11 // or free memory. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #define DEBUG_TYPE "memory-builtins" 16 #include "llvm/ADT/Statistic.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/Analysis/MemoryBuiltins.h" 19 #include "llvm/GlobalVariable.h" 20 #include "llvm/Instructions.h" 21 #include "llvm/Intrinsics.h" 22 #include "llvm/Metadata.h" 23 #include "llvm/Module.h" 24 #include "llvm/Analysis/ValueTracking.h" 25 #include "llvm/Support/Debug.h" 26 #include "llvm/Support/MathExtras.h" 27 #include "llvm/Support/raw_ostream.h" 28 #include "llvm/Target/TargetData.h" 29 #include "llvm/Transforms/Utils/Local.h" 30 using namespace llvm; 31 32 enum AllocType { 33 MallocLike = 1<<0, // allocates 34 CallocLike = 1<<1, // allocates + bzero 35 ReallocLike = 1<<2, // reallocates 36 StrDupLike = 1<<3, 37 AllocLike = MallocLike | CallocLike | StrDupLike, 38 AnyAlloc = MallocLike | CallocLike | ReallocLike | StrDupLike 39 }; 40 41 struct AllocFnsTy { 42 const char *Name; 43 AllocType AllocTy; 44 unsigned char NumParams; 45 // First and Second size parameters (or -1 if unused) 46 signed char FstParam, SndParam; 47 }; 48 49 // FIXME: certain users need more information. E.g., SimplifyLibCalls needs to 50 // know which functions are nounwind, noalias, nocapture parameters, etc. 51 static const AllocFnsTy AllocationFnData[] = { 52 {"malloc", MallocLike, 1, 0, -1}, 53 {"valloc", MallocLike, 1, 0, -1}, 54 {"_Znwj", MallocLike, 1, 0, -1}, // new(unsigned int) 55 {"_ZnwjRKSt9nothrow_t", MallocLike, 2, 0, -1}, // new(unsigned int, nothrow) 56 {"_Znwm", MallocLike, 1, 0, -1}, // new(unsigned long) 57 {"_ZnwmRKSt9nothrow_t", MallocLike, 2, 0, -1}, // new(unsigned long, nothrow) 58 {"_Znaj", MallocLike, 1, 0, -1}, // new[](unsigned int) 59 {"_ZnajRKSt9nothrow_t", MallocLike, 2, 0, -1}, // new[](unsigned int, nothrow) 60 {"_Znam", MallocLike, 1, 0, -1}, // new[](unsigned long) 61 {"_ZnamRKSt9nothrow_t", MallocLike, 2, 0, -1}, // new[](unsigned long, nothrow) 62 {"posix_memalign", MallocLike, 3, 2, -1}, 63 {"calloc", CallocLike, 2, 0, 1}, 64 {"realloc", ReallocLike, 2, 1, -1}, 65 {"reallocf", ReallocLike, 2, 1, -1}, 66 {"strdup", StrDupLike, 1, -1, -1}, 67 {"strndup", StrDupLike, 2, 1, -1} 68 }; 69 70 71 static Function *getCalledFunction(const Value *V, bool LookThroughBitCast) { 72 if (LookThroughBitCast) 73 V = V->stripPointerCasts(); 74 75 CallSite CS(const_cast<Value*>(V)); 76 if (!CS.getInstruction()) 77 return 0; 78 79 Function *Callee = CS.getCalledFunction(); 80 if (!Callee || !Callee->isDeclaration()) 81 return 0; 82 return Callee; 83 } 84 85 /// \brief Returns the allocation data for the given value if it is a call to a 86 /// known allocation function, and NULL otherwise. 87 static const AllocFnsTy *getAllocationData(const Value *V, AllocType AllocTy, 88 bool LookThroughBitCast = false) { 89 Function *Callee = getCalledFunction(V, LookThroughBitCast); 90 if (!Callee) 91 return 0; 92 93 unsigned i = 0; 94 bool found = false; 95 for ( ; i < array_lengthof(AllocationFnData); ++i) { 96 if (Callee->getName() == AllocationFnData[i].Name) { 97 found = true; 98 break; 99 } 100 } 101 if (!found) 102 return 0; 103 104 const AllocFnsTy *FnData = &AllocationFnData[i]; 105 if ((FnData->AllocTy & AllocTy) == 0) 106 return 0; 107 108 // Check function prototype. 109 // FIXME: Check the nobuiltin metadata?? (PR5130) 110 int FstParam = FnData->FstParam; 111 int SndParam = FnData->SndParam; 112 FunctionType *FTy = Callee->getFunctionType(); 113 114 if (FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) && 115 FTy->getNumParams() == FnData->NumParams && 116 (FstParam < 0 || 117 (FTy->getParamType(FstParam)->isIntegerTy(32) || 118 FTy->getParamType(FstParam)->isIntegerTy(64))) && 119 (SndParam < 0 || 120 FTy->getParamType(SndParam)->isIntegerTy(32) || 121 FTy->getParamType(SndParam)->isIntegerTy(64))) 122 return FnData; 123 return 0; 124 } 125 126 static bool hasNoAliasAttr(const Value *V, bool LookThroughBitCast) { 127 ImmutableCallSite CS(LookThroughBitCast ? V->stripPointerCasts() : V); 128 return CS && CS.hasFnAttr(Attribute::NoAlias); 129 } 130 131 132 /// \brief Tests if a value is a call or invoke to a library function that 133 /// allocates or reallocates memory (either malloc, calloc, realloc, or strdup 134 /// like). 135 bool llvm::isAllocationFn(const Value *V, bool LookThroughBitCast) { 136 return getAllocationData(V, AnyAlloc, LookThroughBitCast); 137 } 138 139 /// \brief Tests if a value is a call or invoke to a function that returns a 140 /// NoAlias pointer (including malloc/calloc/realloc/strdup-like functions). 141 bool llvm::isNoAliasFn(const Value *V, bool LookThroughBitCast) { 142 // it's safe to consider realloc as noalias since accessing the original 143 // pointer is undefined behavior 144 return isAllocationFn(V, LookThroughBitCast) || 145 hasNoAliasAttr(V, LookThroughBitCast); 146 } 147 148 /// \brief Tests if a value is a call or invoke to a library function that 149 /// allocates uninitialized memory (such as malloc). 150 bool llvm::isMallocLikeFn(const Value *V, bool LookThroughBitCast) { 151 return getAllocationData(V, MallocLike, LookThroughBitCast); 152 } 153 154 /// \brief Tests if a value is a call or invoke to a library function that 155 /// allocates zero-filled memory (such as calloc). 156 bool llvm::isCallocLikeFn(const Value *V, bool LookThroughBitCast) { 157 return getAllocationData(V, CallocLike, LookThroughBitCast); 158 } 159 160 /// \brief Tests if a value is a call or invoke to a library function that 161 /// allocates memory (either malloc, calloc, or strdup like). 162 bool llvm::isAllocLikeFn(const Value *V, bool LookThroughBitCast) { 163 return getAllocationData(V, AllocLike, LookThroughBitCast); 164 } 165 166 /// \brief Tests if a value is a call or invoke to a library function that 167 /// reallocates memory (such as realloc). 168 bool llvm::isReallocLikeFn(const Value *V, bool LookThroughBitCast) { 169 return getAllocationData(V, ReallocLike, LookThroughBitCast); 170 } 171 172 /// extractMallocCall - Returns the corresponding CallInst if the instruction 173 /// is a malloc call. Since CallInst::CreateMalloc() only creates calls, we 174 /// ignore InvokeInst here. 175 const CallInst *llvm::extractMallocCall(const Value *I) { 176 return isMallocLikeFn(I) ? dyn_cast<CallInst>(I) : 0; 177 } 178 179 static Value *computeArraySize(const CallInst *CI, const TargetData *TD, 180 bool LookThroughSExt = false) { 181 if (!CI) 182 return NULL; 183 184 // The size of the malloc's result type must be known to determine array size. 185 Type *T = getMallocAllocatedType(CI); 186 if (!T || !T->isSized() || !TD) 187 return NULL; 188 189 unsigned ElementSize = TD->getTypeAllocSize(T); 190 if (StructType *ST = dyn_cast<StructType>(T)) 191 ElementSize = TD->getStructLayout(ST)->getSizeInBytes(); 192 193 // If malloc call's arg can be determined to be a multiple of ElementSize, 194 // return the multiple. Otherwise, return NULL. 195 Value *MallocArg = CI->getArgOperand(0); 196 Value *Multiple = NULL; 197 if (ComputeMultiple(MallocArg, ElementSize, Multiple, 198 LookThroughSExt)) 199 return Multiple; 200 201 return NULL; 202 } 203 204 /// isArrayMalloc - Returns the corresponding CallInst if the instruction 205 /// is a call to malloc whose array size can be determined and the array size 206 /// is not constant 1. Otherwise, return NULL. 207 const CallInst *llvm::isArrayMalloc(const Value *I, const TargetData *TD) { 208 const CallInst *CI = extractMallocCall(I); 209 Value *ArraySize = computeArraySize(CI, TD); 210 211 if (ArraySize && 212 ArraySize != ConstantInt::get(CI->getArgOperand(0)->getType(), 1)) 213 return CI; 214 215 // CI is a non-array malloc or we can't figure out that it is an array malloc. 216 return NULL; 217 } 218 219 /// getMallocType - Returns the PointerType resulting from the malloc call. 220 /// The PointerType depends on the number of bitcast uses of the malloc call: 221 /// 0: PointerType is the calls' return type. 222 /// 1: PointerType is the bitcast's result type. 223 /// >1: Unique PointerType cannot be determined, return NULL. 224 PointerType *llvm::getMallocType(const CallInst *CI) { 225 assert(isMallocLikeFn(CI) && "getMallocType and not malloc call"); 226 227 PointerType *MallocType = NULL; 228 unsigned NumOfBitCastUses = 0; 229 230 // Determine if CallInst has a bitcast use. 231 for (Value::const_use_iterator UI = CI->use_begin(), E = CI->use_end(); 232 UI != E; ) 233 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(*UI++)) { 234 MallocType = cast<PointerType>(BCI->getDestTy()); 235 NumOfBitCastUses++; 236 } 237 238 // Malloc call has 1 bitcast use, so type is the bitcast's destination type. 239 if (NumOfBitCastUses == 1) 240 return MallocType; 241 242 // Malloc call was not bitcast, so type is the malloc function's return type. 243 if (NumOfBitCastUses == 0) 244 return cast<PointerType>(CI->getType()); 245 246 // Type could not be determined. 247 return NULL; 248 } 249 250 /// getMallocAllocatedType - Returns the Type allocated by malloc call. 251 /// The Type depends on the number of bitcast uses of the malloc call: 252 /// 0: PointerType is the malloc calls' return type. 253 /// 1: PointerType is the bitcast's result type. 254 /// >1: Unique PointerType cannot be determined, return NULL. 255 Type *llvm::getMallocAllocatedType(const CallInst *CI) { 256 PointerType *PT = getMallocType(CI); 257 return PT ? PT->getElementType() : NULL; 258 } 259 260 /// getMallocArraySize - Returns the array size of a malloc call. If the 261 /// argument passed to malloc is a multiple of the size of the malloced type, 262 /// then return that multiple. For non-array mallocs, the multiple is 263 /// constant 1. Otherwise, return NULL for mallocs whose array size cannot be 264 /// determined. 265 Value *llvm::getMallocArraySize(CallInst *CI, const TargetData *TD, 266 bool LookThroughSExt) { 267 assert(isMallocLikeFn(CI) && "getMallocArraySize and not malloc call"); 268 return computeArraySize(CI, TD, LookThroughSExt); 269 } 270 271 272 /// extractCallocCall - Returns the corresponding CallInst if the instruction 273 /// is a calloc call. 274 const CallInst *llvm::extractCallocCall(const Value *I) { 275 return isCallocLikeFn(I) ? cast<CallInst>(I) : 0; 276 } 277 278 279 /// isFreeCall - Returns non-null if the value is a call to the builtin free() 280 const CallInst *llvm::isFreeCall(const Value *I) { 281 const CallInst *CI = dyn_cast<CallInst>(I); 282 if (!CI) 283 return 0; 284 Function *Callee = CI->getCalledFunction(); 285 if (Callee == 0 || !Callee->isDeclaration()) 286 return 0; 287 288 if (Callee->getName() != "free" && 289 Callee->getName() != "_ZdlPv" && // operator delete(void*) 290 Callee->getName() != "_ZdaPv") // operator delete[](void*) 291 return 0; 292 293 // Check free prototype. 294 // FIXME: workaround for PR5130, this will be obsolete when a nobuiltin 295 // attribute will exist. 296 FunctionType *FTy = Callee->getFunctionType(); 297 if (!FTy->getReturnType()->isVoidTy()) 298 return 0; 299 if (FTy->getNumParams() != 1) 300 return 0; 301 if (FTy->getParamType(0) != Type::getInt8PtrTy(Callee->getContext())) 302 return 0; 303 304 return CI; 305 } 306 307 308 309 //===----------------------------------------------------------------------===// 310 // Utility functions to compute size of objects. 311 // 312 313 314 /// \brief Compute the size of the object pointed by Ptr. Returns true and the 315 /// object size in Size if successful, and false otherwise. 316 /// If RoundToAlign is true, then Size is rounded up to the aligment of allocas, 317 /// byval arguments, and global variables. 318 bool llvm::getObjectSize(const Value *Ptr, uint64_t &Size, const TargetData *TD, 319 bool RoundToAlign) { 320 if (!TD) 321 return false; 322 323 ObjectSizeOffsetVisitor Visitor(TD, Ptr->getContext(), RoundToAlign); 324 SizeOffsetType Data = Visitor.compute(const_cast<Value*>(Ptr)); 325 if (!Visitor.bothKnown(Data)) 326 return false; 327 328 APInt ObjSize = Data.first, Offset = Data.second; 329 // check for overflow 330 if (Offset.slt(0) || ObjSize.ult(Offset)) 331 Size = 0; 332 else 333 Size = (ObjSize - Offset).getZExtValue(); 334 return true; 335 } 336 337 338 STATISTIC(ObjectVisitorArgument, 339 "Number of arguments with unsolved size and offset"); 340 STATISTIC(ObjectVisitorLoad, 341 "Number of load instructions with unsolved size and offset"); 342 343 344 APInt ObjectSizeOffsetVisitor::align(APInt Size, uint64_t Align) { 345 if (RoundToAlign && Align) 346 return APInt(IntTyBits, RoundUpToAlignment(Size.getZExtValue(), Align)); 347 return Size; 348 } 349 350 ObjectSizeOffsetVisitor::ObjectSizeOffsetVisitor(const TargetData *TD, 351 LLVMContext &Context, 352 bool RoundToAlign) 353 : TD(TD), RoundToAlign(RoundToAlign) { 354 IntegerType *IntTy = TD->getIntPtrType(Context); 355 IntTyBits = IntTy->getBitWidth(); 356 Zero = APInt::getNullValue(IntTyBits); 357 } 358 359 SizeOffsetType ObjectSizeOffsetVisitor::compute(Value *V) { 360 V = V->stripPointerCasts(); 361 362 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) 363 return visitGEPOperator(*GEP); 364 if (Instruction *I = dyn_cast<Instruction>(V)) 365 return visit(*I); 366 if (Argument *A = dyn_cast<Argument>(V)) 367 return visitArgument(*A); 368 if (ConstantPointerNull *P = dyn_cast<ConstantPointerNull>(V)) 369 return visitConstantPointerNull(*P); 370 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 371 return visitGlobalVariable(*GV); 372 if (UndefValue *UV = dyn_cast<UndefValue>(V)) 373 return visitUndefValue(*UV); 374 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 375 if (CE->getOpcode() == Instruction::IntToPtr) 376 return unknown(); // clueless 377 378 DEBUG(dbgs() << "ObjectSizeOffsetVisitor::compute() unhandled value: " << *V 379 << '\n'); 380 return unknown(); 381 } 382 383 SizeOffsetType ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) { 384 if (!I.getAllocatedType()->isSized()) 385 return unknown(); 386 387 APInt Size(IntTyBits, TD->getTypeAllocSize(I.getAllocatedType())); 388 if (!I.isArrayAllocation()) 389 return std::make_pair(align(Size, I.getAlignment()), Zero); 390 391 Value *ArraySize = I.getArraySize(); 392 if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) { 393 Size *= C->getValue().zextOrSelf(IntTyBits); 394 return std::make_pair(align(Size, I.getAlignment()), Zero); 395 } 396 return unknown(); 397 } 398 399 SizeOffsetType ObjectSizeOffsetVisitor::visitArgument(Argument &A) { 400 // no interprocedural analysis is done at the moment 401 if (!A.hasByValAttr()) { 402 ++ObjectVisitorArgument; 403 return unknown(); 404 } 405 PointerType *PT = cast<PointerType>(A.getType()); 406 APInt Size(IntTyBits, TD->getTypeAllocSize(PT->getElementType())); 407 return std::make_pair(align(Size, A.getParamAlignment()), Zero); 408 } 409 410 SizeOffsetType ObjectSizeOffsetVisitor::visitCallSite(CallSite CS) { 411 const AllocFnsTy *FnData = getAllocationData(CS.getInstruction(), AnyAlloc); 412 if (!FnData) 413 return unknown(); 414 415 // handle strdup-like functions separately 416 if (FnData->AllocTy == StrDupLike) { 417 APInt Size(IntTyBits, GetStringLength(CS.getArgument(0))); 418 if (!Size) 419 return unknown(); 420 421 // strndup limits strlen 422 if (FnData->FstParam > 0) { 423 ConstantInt *Arg= dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam)); 424 if (!Arg) 425 return unknown(); 426 427 APInt MaxSize = Arg->getValue().zextOrSelf(IntTyBits); 428 if (Size.ugt(MaxSize)) 429 Size = MaxSize + 1; 430 } 431 return std::make_pair(Size, Zero); 432 } 433 434 ConstantInt *Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam)); 435 if (!Arg) 436 return unknown(); 437 438 APInt Size = Arg->getValue().zextOrSelf(IntTyBits); 439 // size determined by just 1 parameter 440 if (FnData->SndParam < 0) 441 return std::make_pair(Size, Zero); 442 443 Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->SndParam)); 444 if (!Arg) 445 return unknown(); 446 447 Size *= Arg->getValue().zextOrSelf(IntTyBits); 448 return std::make_pair(Size, Zero); 449 450 // TODO: handle more standard functions (+ wchar cousins): 451 // - strdup / strndup 452 // - strcpy / strncpy 453 // - strcat / strncat 454 // - memcpy / memmove 455 // - strcat / strncat 456 // - memset 457 } 458 459 SizeOffsetType 460 ObjectSizeOffsetVisitor::visitConstantPointerNull(ConstantPointerNull&) { 461 return std::make_pair(Zero, Zero); 462 } 463 464 SizeOffsetType 465 ObjectSizeOffsetVisitor::visitExtractElementInst(ExtractElementInst&) { 466 return unknown(); 467 } 468 469 SizeOffsetType 470 ObjectSizeOffsetVisitor::visitExtractValueInst(ExtractValueInst&) { 471 // Easy cases were already folded by previous passes. 472 return unknown(); 473 } 474 475 SizeOffsetType ObjectSizeOffsetVisitor::visitGEPOperator(GEPOperator &GEP) { 476 // Ignore self-referencing GEPs, they can occur in unreachable code. 477 if (&GEP == GEP.getPointerOperand()) 478 return unknown(); 479 480 SizeOffsetType PtrData = compute(GEP.getPointerOperand()); 481 if (!bothKnown(PtrData) || !GEP.hasAllConstantIndices()) 482 return unknown(); 483 484 SmallVector<Value*, 8> Ops(GEP.idx_begin(), GEP.idx_end()); 485 APInt Offset(IntTyBits,TD->getIndexedOffset(GEP.getPointerOperandType(),Ops)); 486 return std::make_pair(PtrData.first, PtrData.second + Offset); 487 } 488 489 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalVariable(GlobalVariable &GV){ 490 if (!GV.hasDefinitiveInitializer()) 491 return unknown(); 492 493 APInt Size(IntTyBits, TD->getTypeAllocSize(GV.getType()->getElementType())); 494 return std::make_pair(align(Size, GV.getAlignment()), Zero); 495 } 496 497 SizeOffsetType ObjectSizeOffsetVisitor::visitIntToPtrInst(IntToPtrInst&) { 498 // clueless 499 return unknown(); 500 } 501 502 SizeOffsetType ObjectSizeOffsetVisitor::visitLoadInst(LoadInst&) { 503 ++ObjectVisitorLoad; 504 return unknown(); 505 } 506 507 SizeOffsetType ObjectSizeOffsetVisitor::visitPHINode(PHINode&) { 508 // too complex to analyze statically. 509 return unknown(); 510 } 511 512 SizeOffsetType ObjectSizeOffsetVisitor::visitSelectInst(SelectInst &I) { 513 // ignore malformed self-looping selects 514 if (I.getTrueValue() == &I || I.getFalseValue() == &I) 515 return unknown(); 516 517 SizeOffsetType TrueSide = compute(I.getTrueValue()); 518 SizeOffsetType FalseSide = compute(I.getFalseValue()); 519 if (bothKnown(TrueSide) && bothKnown(FalseSide) && TrueSide == FalseSide) 520 return TrueSide; 521 return unknown(); 522 } 523 524 SizeOffsetType ObjectSizeOffsetVisitor::visitUndefValue(UndefValue&) { 525 return std::make_pair(Zero, Zero); 526 } 527 528 SizeOffsetType ObjectSizeOffsetVisitor::visitInstruction(Instruction &I) { 529 DEBUG(dbgs() << "ObjectSizeOffsetVisitor unknown instruction:" << I << '\n'); 530 return unknown(); 531 } 532 533 534 ObjectSizeOffsetEvaluator::ObjectSizeOffsetEvaluator(const TargetData *TD, 535 LLVMContext &Context) 536 : TD(TD), Context(Context), Builder(Context, TargetFolder(TD)), 537 Visitor(TD, Context) { 538 IntTy = TD->getIntPtrType(Context); 539 Zero = ConstantInt::get(IntTy, 0); 540 } 541 542 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute(Value *V) { 543 SizeOffsetEvalType Result = compute_(V); 544 545 if (!bothKnown(Result)) { 546 // erase everything that was computed in this iteration from the cache, so 547 // that no dangling references are left behind. We could be a bit smarter if 548 // we kept a dependency graph. It's probably not worth the complexity. 549 for (PtrSetTy::iterator I=SeenVals.begin(), E=SeenVals.end(); I != E; ++I) { 550 CacheMapTy::iterator CacheIt = CacheMap.find(*I); 551 // non-computable results can be safely cached 552 if (CacheIt != CacheMap.end() && anyKnown(CacheIt->second)) 553 CacheMap.erase(CacheIt); 554 } 555 } 556 557 SeenVals.clear(); 558 return Result; 559 } 560 561 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute_(Value *V) { 562 SizeOffsetType Const = Visitor.compute(V); 563 if (Visitor.bothKnown(Const)) 564 return std::make_pair(ConstantInt::get(Context, Const.first), 565 ConstantInt::get(Context, Const.second)); 566 567 V = V->stripPointerCasts(); 568 569 // check cache 570 CacheMapTy::iterator CacheIt = CacheMap.find(V); 571 if (CacheIt != CacheMap.end()) 572 return CacheIt->second; 573 574 // always generate code immediately before the instruction being 575 // processed, so that the generated code dominates the same BBs 576 Instruction *PrevInsertPoint = Builder.GetInsertPoint(); 577 if (Instruction *I = dyn_cast<Instruction>(V)) 578 Builder.SetInsertPoint(I); 579 580 // record the pointers that were handled in this run, so that they can be 581 // cleaned later if something fails 582 SeenVals.insert(V); 583 584 // now compute the size and offset 585 SizeOffsetEvalType Result; 586 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 587 Result = visitGEPOperator(*GEP); 588 } else if (Instruction *I = dyn_cast<Instruction>(V)) { 589 Result = visit(*I); 590 } else if (isa<Argument>(V) || 591 (isa<ConstantExpr>(V) && 592 cast<ConstantExpr>(V)->getOpcode() == Instruction::IntToPtr) || 593 isa<GlobalVariable>(V)) { 594 // ignore values where we cannot do more than what ObjectSizeVisitor can 595 Result = unknown(); 596 } else { 597 DEBUG(dbgs() << "ObjectSizeOffsetEvaluator::compute() unhandled value: " 598 << *V << '\n'); 599 Result = unknown(); 600 } 601 602 if (PrevInsertPoint) 603 Builder.SetInsertPoint(PrevInsertPoint); 604 605 // Don't reuse CacheIt since it may be invalid at this point. 606 CacheMap[V] = Result; 607 return Result; 608 } 609 610 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitAllocaInst(AllocaInst &I) { 611 if (!I.getAllocatedType()->isSized()) 612 return unknown(); 613 614 // must be a VLA 615 assert(I.isArrayAllocation()); 616 Value *ArraySize = I.getArraySize(); 617 Value *Size = ConstantInt::get(ArraySize->getType(), 618 TD->getTypeAllocSize(I.getAllocatedType())); 619 Size = Builder.CreateMul(Size, ArraySize); 620 return std::make_pair(Size, Zero); 621 } 622 623 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitCallSite(CallSite CS) { 624 const AllocFnsTy *FnData = getAllocationData(CS.getInstruction(), AnyAlloc); 625 if (!FnData) 626 return unknown(); 627 628 // handle strdup-like functions separately 629 if (FnData->AllocTy == StrDupLike) { 630 // TODO 631 return unknown(); 632 } 633 634 Value *FirstArg = CS.getArgument(FnData->FstParam); 635 FirstArg = Builder.CreateZExt(FirstArg, IntTy); 636 if (FnData->SndParam < 0) 637 return std::make_pair(FirstArg, Zero); 638 639 Value *SecondArg = CS.getArgument(FnData->SndParam); 640 SecondArg = Builder.CreateZExt(SecondArg, IntTy); 641 Value *Size = Builder.CreateMul(FirstArg, SecondArg); 642 return std::make_pair(Size, Zero); 643 644 // TODO: handle more standard functions (+ wchar cousins): 645 // - strdup / strndup 646 // - strcpy / strncpy 647 // - strcat / strncat 648 // - memcpy / memmove 649 // - strcat / strncat 650 // - memset 651 } 652 653 SizeOffsetEvalType 654 ObjectSizeOffsetEvaluator::visitExtractElementInst(ExtractElementInst&) { 655 return unknown(); 656 } 657 658 SizeOffsetEvalType 659 ObjectSizeOffsetEvaluator::visitExtractValueInst(ExtractValueInst&) { 660 return unknown(); 661 } 662 663 SizeOffsetEvalType 664 ObjectSizeOffsetEvaluator::visitGEPOperator(GEPOperator &GEP) { 665 SizeOffsetEvalType PtrData = compute_(GEP.getPointerOperand()); 666 if (!bothKnown(PtrData)) 667 return unknown(); 668 669 Value *Offset = EmitGEPOffset(&Builder, *TD, &GEP, /*NoAssumptions=*/true); 670 Offset = Builder.CreateAdd(PtrData.second, Offset); 671 return std::make_pair(PtrData.first, Offset); 672 } 673 674 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitIntToPtrInst(IntToPtrInst&) { 675 // clueless 676 return unknown(); 677 } 678 679 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitLoadInst(LoadInst&) { 680 return unknown(); 681 } 682 683 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitPHINode(PHINode &PHI) { 684 // create 2 PHIs: one for size and another for offset 685 PHINode *SizePHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 686 PHINode *OffsetPHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 687 688 // insert right away in the cache to handle recursive PHIs 689 CacheMap[&PHI] = std::make_pair(SizePHI, OffsetPHI); 690 691 // compute offset/size for each PHI incoming pointer 692 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i) { 693 Builder.SetInsertPoint(PHI.getIncomingBlock(i)->getFirstInsertionPt()); 694 SizeOffsetEvalType EdgeData = compute_(PHI.getIncomingValue(i)); 695 696 if (!bothKnown(EdgeData)) { 697 OffsetPHI->replaceAllUsesWith(UndefValue::get(IntTy)); 698 OffsetPHI->eraseFromParent(); 699 SizePHI->replaceAllUsesWith(UndefValue::get(IntTy)); 700 SizePHI->eraseFromParent(); 701 return unknown(); 702 } 703 SizePHI->addIncoming(EdgeData.first, PHI.getIncomingBlock(i)); 704 OffsetPHI->addIncoming(EdgeData.second, PHI.getIncomingBlock(i)); 705 } 706 707 Value *Size = SizePHI, *Offset = OffsetPHI, *Tmp; 708 if ((Tmp = SizePHI->hasConstantValue())) { 709 Size = Tmp; 710 SizePHI->replaceAllUsesWith(Size); 711 SizePHI->eraseFromParent(); 712 } 713 if ((Tmp = OffsetPHI->hasConstantValue())) { 714 Offset = Tmp; 715 OffsetPHI->replaceAllUsesWith(Offset); 716 OffsetPHI->eraseFromParent(); 717 } 718 return std::make_pair(Size, Offset); 719 } 720 721 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitSelectInst(SelectInst &I) { 722 // ignore malformed self-looping selects 723 if (I.getTrueValue() == &I || I.getFalseValue() == &I) 724 return unknown(); 725 726 SizeOffsetEvalType TrueSide = compute_(I.getTrueValue()); 727 SizeOffsetEvalType FalseSide = compute_(I.getFalseValue()); 728 729 if (!bothKnown(TrueSide) || !bothKnown(FalseSide)) 730 return unknown(); 731 if (TrueSide == FalseSide) 732 return TrueSide; 733 734 Value *Size = Builder.CreateSelect(I.getCondition(), TrueSide.first, 735 FalseSide.first); 736 Value *Offset = Builder.CreateSelect(I.getCondition(), TrueSide.second, 737 FalseSide.second); 738 return std::make_pair(Size, Offset); 739 } 740 741 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitInstruction(Instruction &I) { 742 DEBUG(dbgs() << "ObjectSizeOffsetEvaluator unknown instruction:" << I <<'\n'); 743 return unknown(); 744 } 745