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 #include "llvm/Analysis/MemoryBuiltins.h" 16 #include "llvm/ADT/STLExtras.h" 17 #include "llvm/ADT/Statistic.h" 18 #include "llvm/Analysis/TargetLibraryInfo.h" 19 #include "llvm/Analysis/ValueTracking.h" 20 #include "llvm/IR/DataLayout.h" 21 #include "llvm/IR/GlobalVariable.h" 22 #include "llvm/IR/Instructions.h" 23 #include "llvm/IR/Intrinsics.h" 24 #include "llvm/IR/Metadata.h" 25 #include "llvm/IR/Module.h" 26 #include "llvm/Support/Debug.h" 27 #include "llvm/Support/MathExtras.h" 28 #include "llvm/Support/raw_ostream.h" 29 #include "llvm/Transforms/Utils/Local.h" 30 using namespace llvm; 31 32 #define DEBUG_TYPE "memory-builtins" 33 34 enum AllocType : uint8_t { 35 OpNewLike = 1<<0, // allocates; never returns null 36 MallocLike = 1<<1 | OpNewLike, // allocates; may return null 37 CallocLike = 1<<2, // allocates + bzero 38 ReallocLike = 1<<3, // reallocates 39 StrDupLike = 1<<4, 40 AllocLike = MallocLike | CallocLike | StrDupLike, 41 AnyAlloc = AllocLike | ReallocLike 42 }; 43 44 struct AllocFnsTy { 45 LibFunc::Func Func; 46 AllocType AllocTy; 47 unsigned char NumParams; 48 // First and Second size parameters (or -1 if unused) 49 signed char FstParam, SndParam; 50 }; 51 52 // FIXME: certain users need more information. E.g., SimplifyLibCalls needs to 53 // know which functions are nounwind, noalias, nocapture parameters, etc. 54 static const AllocFnsTy AllocationFnData[] = { 55 {LibFunc::malloc, MallocLike, 1, 0, -1}, 56 {LibFunc::valloc, MallocLike, 1, 0, -1}, 57 {LibFunc::Znwj, OpNewLike, 1, 0, -1}, // new(unsigned int) 58 {LibFunc::ZnwjRKSt9nothrow_t, MallocLike, 2, 0, -1}, // new(unsigned int, nothrow) 59 {LibFunc::Znwm, OpNewLike, 1, 0, -1}, // new(unsigned long) 60 {LibFunc::ZnwmRKSt9nothrow_t, MallocLike, 2, 0, -1}, // new(unsigned long, nothrow) 61 {LibFunc::Znaj, OpNewLike, 1, 0, -1}, // new[](unsigned int) 62 {LibFunc::ZnajRKSt9nothrow_t, MallocLike, 2, 0, -1}, // new[](unsigned int, nothrow) 63 {LibFunc::Znam, OpNewLike, 1, 0, -1}, // new[](unsigned long) 64 {LibFunc::ZnamRKSt9nothrow_t, MallocLike, 2, 0, -1}, // new[](unsigned long, nothrow) 65 {LibFunc::msvc_new_int, OpNewLike, 1, 0, -1}, // new(unsigned int) 66 {LibFunc::msvc_new_int_nothrow, MallocLike, 2, 0, -1}, // new(unsigned int, nothrow) 67 {LibFunc::msvc_new_longlong, OpNewLike, 1, 0, -1}, // new(unsigned long long) 68 {LibFunc::msvc_new_longlong_nothrow, MallocLike, 2, 0, -1}, // new(unsigned long long, nothrow) 69 {LibFunc::msvc_new_array_int, OpNewLike, 1, 0, -1}, // new[](unsigned int) 70 {LibFunc::msvc_new_array_int_nothrow, MallocLike, 2, 0, -1}, // new[](unsigned int, nothrow) 71 {LibFunc::msvc_new_array_longlong, OpNewLike, 1, 0, -1}, // new[](unsigned long long) 72 {LibFunc::msvc_new_array_longlong_nothrow, MallocLike, 2, 0, -1}, // new[](unsigned long long, nothrow) 73 {LibFunc::calloc, CallocLike, 2, 0, 1}, 74 {LibFunc::realloc, ReallocLike, 2, 1, -1}, 75 {LibFunc::reallocf, ReallocLike, 2, 1, -1}, 76 {LibFunc::strdup, StrDupLike, 1, -1, -1}, 77 {LibFunc::strndup, StrDupLike, 2, 1, -1} 78 // TODO: Handle "int posix_memalign(void **, size_t, size_t)" 79 }; 80 81 82 static Function *getCalledFunction(const Value *V, bool LookThroughBitCast) { 83 if (LookThroughBitCast) 84 V = V->stripPointerCasts(); 85 86 CallSite CS(const_cast<Value*>(V)); 87 if (!CS.getInstruction()) 88 return nullptr; 89 90 if (CS.isNoBuiltin()) 91 return nullptr; 92 93 Function *Callee = CS.getCalledFunction(); 94 if (!Callee || !Callee->isDeclaration()) 95 return nullptr; 96 return Callee; 97 } 98 99 /// \brief Returns the allocation data for the given value if it is a call to a 100 /// known allocation function, and NULL otherwise. 101 static const AllocFnsTy *getAllocationData(const Value *V, AllocType AllocTy, 102 const TargetLibraryInfo *TLI, 103 bool LookThroughBitCast = false) { 104 // Skip intrinsics 105 if (isa<IntrinsicInst>(V)) 106 return nullptr; 107 108 Function *Callee = getCalledFunction(V, LookThroughBitCast); 109 if (!Callee) 110 return nullptr; 111 112 // Make sure that the function is available. 113 StringRef FnName = Callee->getName(); 114 LibFunc::Func TLIFn; 115 if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn)) 116 return nullptr; 117 118 const AllocFnsTy *FnData = 119 std::find_if(std::begin(AllocationFnData), std::end(AllocationFnData), 120 [TLIFn](const AllocFnsTy &Fn) { return Fn.Func == TLIFn; }); 121 122 if (FnData == std::end(AllocationFnData)) 123 return nullptr; 124 125 if ((FnData->AllocTy & AllocTy) != FnData->AllocTy) 126 return nullptr; 127 128 // Check function prototype. 129 int FstParam = FnData->FstParam; 130 int SndParam = FnData->SndParam; 131 FunctionType *FTy = Callee->getFunctionType(); 132 133 if (FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) && 134 FTy->getNumParams() == FnData->NumParams && 135 (FstParam < 0 || 136 (FTy->getParamType(FstParam)->isIntegerTy(32) || 137 FTy->getParamType(FstParam)->isIntegerTy(64))) && 138 (SndParam < 0 || 139 FTy->getParamType(SndParam)->isIntegerTy(32) || 140 FTy->getParamType(SndParam)->isIntegerTy(64))) 141 return FnData; 142 return nullptr; 143 } 144 145 static bool hasNoAliasAttr(const Value *V, bool LookThroughBitCast) { 146 ImmutableCallSite CS(LookThroughBitCast ? V->stripPointerCasts() : V); 147 return CS && CS.hasFnAttr(Attribute::NoAlias); 148 } 149 150 151 /// \brief Tests if a value is a call or invoke to a library function that 152 /// allocates or reallocates memory (either malloc, calloc, realloc, or strdup 153 /// like). 154 bool llvm::isAllocationFn(const Value *V, const TargetLibraryInfo *TLI, 155 bool LookThroughBitCast) { 156 return getAllocationData(V, AnyAlloc, TLI, LookThroughBitCast); 157 } 158 159 /// \brief Tests if a value is a call or invoke to a function that returns a 160 /// NoAlias pointer (including malloc/calloc/realloc/strdup-like functions). 161 bool llvm::isNoAliasFn(const Value *V, const TargetLibraryInfo *TLI, 162 bool LookThroughBitCast) { 163 // it's safe to consider realloc as noalias since accessing the original 164 // pointer is undefined behavior 165 return isAllocationFn(V, TLI, LookThroughBitCast) || 166 hasNoAliasAttr(V, LookThroughBitCast); 167 } 168 169 /// \brief Tests if a value is a call or invoke to a library function that 170 /// allocates uninitialized memory (such as malloc). 171 bool llvm::isMallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, 172 bool LookThroughBitCast) { 173 return getAllocationData(V, MallocLike, TLI, LookThroughBitCast); 174 } 175 176 /// \brief Tests if a value is a call or invoke to a library function that 177 /// allocates zero-filled memory (such as calloc). 178 bool llvm::isCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, 179 bool LookThroughBitCast) { 180 return getAllocationData(V, CallocLike, TLI, LookThroughBitCast); 181 } 182 183 /// \brief Tests if a value is a call or invoke to a library function that 184 /// allocates memory (either malloc, calloc, or strdup like). 185 bool llvm::isAllocLikeFn(const Value *V, const TargetLibraryInfo *TLI, 186 bool LookThroughBitCast) { 187 return getAllocationData(V, AllocLike, TLI, LookThroughBitCast); 188 } 189 190 /// \brief Tests if a value is a call or invoke to a library function that 191 /// reallocates memory (such as realloc). 192 bool llvm::isReallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, 193 bool LookThroughBitCast) { 194 return getAllocationData(V, ReallocLike, TLI, LookThroughBitCast); 195 } 196 197 /// \brief Tests if a value is a call or invoke to a library function that 198 /// allocates memory and never returns null (such as operator new). 199 bool llvm::isOperatorNewLikeFn(const Value *V, const TargetLibraryInfo *TLI, 200 bool LookThroughBitCast) { 201 return getAllocationData(V, OpNewLike, TLI, LookThroughBitCast); 202 } 203 204 /// extractMallocCall - Returns the corresponding CallInst if the instruction 205 /// is a malloc call. Since CallInst::CreateMalloc() only creates calls, we 206 /// ignore InvokeInst here. 207 const CallInst *llvm::extractMallocCall(const Value *I, 208 const TargetLibraryInfo *TLI) { 209 return isMallocLikeFn(I, TLI) ? dyn_cast<CallInst>(I) : nullptr; 210 } 211 212 static Value *computeArraySize(const CallInst *CI, const DataLayout &DL, 213 const TargetLibraryInfo *TLI, 214 bool LookThroughSExt = false) { 215 if (!CI) 216 return nullptr; 217 218 // The size of the malloc's result type must be known to determine array size. 219 Type *T = getMallocAllocatedType(CI, TLI); 220 if (!T || !T->isSized()) 221 return nullptr; 222 223 unsigned ElementSize = DL.getTypeAllocSize(T); 224 if (StructType *ST = dyn_cast<StructType>(T)) 225 ElementSize = DL.getStructLayout(ST)->getSizeInBytes(); 226 227 // If malloc call's arg can be determined to be a multiple of ElementSize, 228 // return the multiple. Otherwise, return NULL. 229 Value *MallocArg = CI->getArgOperand(0); 230 Value *Multiple = nullptr; 231 if (ComputeMultiple(MallocArg, ElementSize, Multiple, 232 LookThroughSExt)) 233 return Multiple; 234 235 return nullptr; 236 } 237 238 /// getMallocType - Returns the PointerType resulting from the malloc call. 239 /// The PointerType depends on the number of bitcast uses of the malloc call: 240 /// 0: PointerType is the calls' return type. 241 /// 1: PointerType is the bitcast's result type. 242 /// >1: Unique PointerType cannot be determined, return NULL. 243 PointerType *llvm::getMallocType(const CallInst *CI, 244 const TargetLibraryInfo *TLI) { 245 assert(isMallocLikeFn(CI, TLI) && "getMallocType and not malloc call"); 246 247 PointerType *MallocType = nullptr; 248 unsigned NumOfBitCastUses = 0; 249 250 // Determine if CallInst has a bitcast use. 251 for (Value::const_user_iterator UI = CI->user_begin(), E = CI->user_end(); 252 UI != E;) 253 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(*UI++)) { 254 MallocType = cast<PointerType>(BCI->getDestTy()); 255 NumOfBitCastUses++; 256 } 257 258 // Malloc call has 1 bitcast use, so type is the bitcast's destination type. 259 if (NumOfBitCastUses == 1) 260 return MallocType; 261 262 // Malloc call was not bitcast, so type is the malloc function's return type. 263 if (NumOfBitCastUses == 0) 264 return cast<PointerType>(CI->getType()); 265 266 // Type could not be determined. 267 return nullptr; 268 } 269 270 /// getMallocAllocatedType - Returns the Type allocated by malloc call. 271 /// The Type depends on the number of bitcast uses of the malloc call: 272 /// 0: PointerType is the malloc calls' return type. 273 /// 1: PointerType is the bitcast's result type. 274 /// >1: Unique PointerType cannot be determined, return NULL. 275 Type *llvm::getMallocAllocatedType(const CallInst *CI, 276 const TargetLibraryInfo *TLI) { 277 PointerType *PT = getMallocType(CI, TLI); 278 return PT ? PT->getElementType() : nullptr; 279 } 280 281 /// getMallocArraySize - Returns the array size of a malloc call. If the 282 /// argument passed to malloc is a multiple of the size of the malloced type, 283 /// then return that multiple. For non-array mallocs, the multiple is 284 /// constant 1. Otherwise, return NULL for mallocs whose array size cannot be 285 /// determined. 286 Value *llvm::getMallocArraySize(CallInst *CI, const DataLayout &DL, 287 const TargetLibraryInfo *TLI, 288 bool LookThroughSExt) { 289 assert(isMallocLikeFn(CI, TLI) && "getMallocArraySize and not malloc call"); 290 return computeArraySize(CI, DL, TLI, LookThroughSExt); 291 } 292 293 294 /// extractCallocCall - Returns the corresponding CallInst if the instruction 295 /// is a calloc call. 296 const CallInst *llvm::extractCallocCall(const Value *I, 297 const TargetLibraryInfo *TLI) { 298 return isCallocLikeFn(I, TLI) ? cast<CallInst>(I) : nullptr; 299 } 300 301 302 /// isFreeCall - Returns non-null if the value is a call to the builtin free() 303 const CallInst *llvm::isFreeCall(const Value *I, const TargetLibraryInfo *TLI) { 304 const CallInst *CI = dyn_cast<CallInst>(I); 305 if (!CI || isa<IntrinsicInst>(CI)) 306 return nullptr; 307 Function *Callee = CI->getCalledFunction(); 308 if (Callee == nullptr) 309 return nullptr; 310 311 StringRef FnName = Callee->getName(); 312 LibFunc::Func TLIFn; 313 if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn)) 314 return nullptr; 315 316 unsigned ExpectedNumParams; 317 if (TLIFn == LibFunc::free || 318 TLIFn == LibFunc::ZdlPv || // operator delete(void*) 319 TLIFn == LibFunc::ZdaPv || // operator delete[](void*) 320 TLIFn == LibFunc::msvc_delete_ptr32 || // operator delete(void*) 321 TLIFn == LibFunc::msvc_delete_ptr64 || // operator delete(void*) 322 TLIFn == LibFunc::msvc_delete_array_ptr32 || // operator delete[](void*) 323 TLIFn == LibFunc::msvc_delete_array_ptr64) // operator delete[](void*) 324 ExpectedNumParams = 1; 325 else if (TLIFn == LibFunc::ZdlPvj || // delete(void*, uint) 326 TLIFn == LibFunc::ZdlPvm || // delete(void*, ulong) 327 TLIFn == LibFunc::ZdlPvRKSt9nothrow_t || // delete(void*, nothrow) 328 TLIFn == LibFunc::ZdaPvj || // delete[](void*, uint) 329 TLIFn == LibFunc::ZdaPvm || // delete[](void*, ulong) 330 TLIFn == LibFunc::ZdaPvRKSt9nothrow_t || // delete[](void*, nothrow) 331 TLIFn == LibFunc::msvc_delete_ptr32_int || // delete(void*, uint) 332 TLIFn == LibFunc::msvc_delete_ptr64_longlong || // delete(void*, ulonglong) 333 TLIFn == LibFunc::msvc_delete_ptr32_nothrow || // delete(void*, nothrow) 334 TLIFn == LibFunc::msvc_delete_ptr64_nothrow || // delete(void*, nothrow) 335 TLIFn == LibFunc::msvc_delete_array_ptr32_int || // delete[](void*, uint) 336 TLIFn == LibFunc::msvc_delete_array_ptr64_longlong || // delete[](void*, ulonglong) 337 TLIFn == LibFunc::msvc_delete_array_ptr32_nothrow || // delete[](void*, nothrow) 338 TLIFn == LibFunc::msvc_delete_array_ptr64_nothrow) // delete[](void*, nothrow) 339 ExpectedNumParams = 2; 340 else 341 return nullptr; 342 343 // Check free prototype. 344 // FIXME: workaround for PR5130, this will be obsolete when a nobuiltin 345 // attribute will exist. 346 FunctionType *FTy = Callee->getFunctionType(); 347 if (!FTy->getReturnType()->isVoidTy()) 348 return nullptr; 349 if (FTy->getNumParams() != ExpectedNumParams) 350 return nullptr; 351 if (FTy->getParamType(0) != Type::getInt8PtrTy(Callee->getContext())) 352 return nullptr; 353 354 return CI; 355 } 356 357 358 359 //===----------------------------------------------------------------------===// 360 // Utility functions to compute size of objects. 361 // 362 363 364 /// \brief Compute the size of the object pointed by Ptr. Returns true and the 365 /// object size in Size if successful, and false otherwise. 366 /// If RoundToAlign is true, then Size is rounded up to the aligment of allocas, 367 /// byval arguments, and global variables. 368 bool llvm::getObjectSize(const Value *Ptr, uint64_t &Size, const DataLayout &DL, 369 const TargetLibraryInfo *TLI, bool RoundToAlign) { 370 ObjectSizeOffsetVisitor Visitor(DL, TLI, Ptr->getContext(), RoundToAlign); 371 SizeOffsetType Data = Visitor.compute(const_cast<Value*>(Ptr)); 372 if (!Visitor.bothKnown(Data)) 373 return false; 374 375 APInt ObjSize = Data.first, Offset = Data.second; 376 // check for overflow 377 if (Offset.slt(0) || ObjSize.ult(Offset)) 378 Size = 0; 379 else 380 Size = (ObjSize - Offset).getZExtValue(); 381 return true; 382 } 383 384 385 STATISTIC(ObjectVisitorArgument, 386 "Number of arguments with unsolved size and offset"); 387 STATISTIC(ObjectVisitorLoad, 388 "Number of load instructions with unsolved size and offset"); 389 390 391 APInt ObjectSizeOffsetVisitor::align(APInt Size, uint64_t Align) { 392 if (RoundToAlign && Align) 393 return APInt(IntTyBits, RoundUpToAlignment(Size.getZExtValue(), Align)); 394 return Size; 395 } 396 397 ObjectSizeOffsetVisitor::ObjectSizeOffsetVisitor(const DataLayout &DL, 398 const TargetLibraryInfo *TLI, 399 LLVMContext &Context, 400 bool RoundToAlign) 401 : DL(DL), TLI(TLI), RoundToAlign(RoundToAlign) { 402 // Pointer size must be rechecked for each object visited since it could have 403 // a different address space. 404 } 405 406 SizeOffsetType ObjectSizeOffsetVisitor::compute(Value *V) { 407 IntTyBits = DL.getPointerTypeSizeInBits(V->getType()); 408 Zero = APInt::getNullValue(IntTyBits); 409 410 V = V->stripPointerCasts(); 411 if (Instruction *I = dyn_cast<Instruction>(V)) { 412 // If we have already seen this instruction, bail out. Cycles can happen in 413 // unreachable code after constant propagation. 414 if (!SeenInsts.insert(I).second) 415 return unknown(); 416 417 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) 418 return visitGEPOperator(*GEP); 419 return visit(*I); 420 } 421 if (Argument *A = dyn_cast<Argument>(V)) 422 return visitArgument(*A); 423 if (ConstantPointerNull *P = dyn_cast<ConstantPointerNull>(V)) 424 return visitConstantPointerNull(*P); 425 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) 426 return visitGlobalAlias(*GA); 427 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 428 return visitGlobalVariable(*GV); 429 if (UndefValue *UV = dyn_cast<UndefValue>(V)) 430 return visitUndefValue(*UV); 431 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 432 if (CE->getOpcode() == Instruction::IntToPtr) 433 return unknown(); // clueless 434 if (CE->getOpcode() == Instruction::GetElementPtr) 435 return visitGEPOperator(cast<GEPOperator>(*CE)); 436 } 437 438 DEBUG(dbgs() << "ObjectSizeOffsetVisitor::compute() unhandled value: " << *V 439 << '\n'); 440 return unknown(); 441 } 442 443 SizeOffsetType ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) { 444 if (!I.getAllocatedType()->isSized()) 445 return unknown(); 446 447 APInt Size(IntTyBits, DL.getTypeAllocSize(I.getAllocatedType())); 448 if (!I.isArrayAllocation()) 449 return std::make_pair(align(Size, I.getAlignment()), Zero); 450 451 Value *ArraySize = I.getArraySize(); 452 if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) { 453 Size *= C->getValue().zextOrSelf(IntTyBits); 454 return std::make_pair(align(Size, I.getAlignment()), Zero); 455 } 456 return unknown(); 457 } 458 459 SizeOffsetType ObjectSizeOffsetVisitor::visitArgument(Argument &A) { 460 // no interprocedural analysis is done at the moment 461 if (!A.hasByValOrInAllocaAttr()) { 462 ++ObjectVisitorArgument; 463 return unknown(); 464 } 465 PointerType *PT = cast<PointerType>(A.getType()); 466 APInt Size(IntTyBits, DL.getTypeAllocSize(PT->getElementType())); 467 return std::make_pair(align(Size, A.getParamAlignment()), Zero); 468 } 469 470 SizeOffsetType ObjectSizeOffsetVisitor::visitCallSite(CallSite CS) { 471 const AllocFnsTy *FnData = getAllocationData(CS.getInstruction(), AnyAlloc, 472 TLI); 473 if (!FnData) 474 return unknown(); 475 476 // handle strdup-like functions separately 477 if (FnData->AllocTy == StrDupLike) { 478 APInt Size(IntTyBits, GetStringLength(CS.getArgument(0))); 479 if (!Size) 480 return unknown(); 481 482 // strndup limits strlen 483 if (FnData->FstParam > 0) { 484 ConstantInt *Arg= dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam)); 485 if (!Arg) 486 return unknown(); 487 488 APInt MaxSize = Arg->getValue().zextOrSelf(IntTyBits); 489 if (Size.ugt(MaxSize)) 490 Size = MaxSize + 1; 491 } 492 return std::make_pair(Size, Zero); 493 } 494 495 ConstantInt *Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam)); 496 if (!Arg) 497 return unknown(); 498 499 APInt Size = Arg->getValue().zextOrSelf(IntTyBits); 500 // size determined by just 1 parameter 501 if (FnData->SndParam < 0) 502 return std::make_pair(Size, Zero); 503 504 Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->SndParam)); 505 if (!Arg) 506 return unknown(); 507 508 Size *= Arg->getValue().zextOrSelf(IntTyBits); 509 return std::make_pair(Size, Zero); 510 511 // TODO: handle more standard functions (+ wchar cousins): 512 // - strdup / strndup 513 // - strcpy / strncpy 514 // - strcat / strncat 515 // - memcpy / memmove 516 // - strcat / strncat 517 // - memset 518 } 519 520 SizeOffsetType 521 ObjectSizeOffsetVisitor::visitConstantPointerNull(ConstantPointerNull&) { 522 return std::make_pair(Zero, Zero); 523 } 524 525 SizeOffsetType 526 ObjectSizeOffsetVisitor::visitExtractElementInst(ExtractElementInst&) { 527 return unknown(); 528 } 529 530 SizeOffsetType 531 ObjectSizeOffsetVisitor::visitExtractValueInst(ExtractValueInst&) { 532 // Easy cases were already folded by previous passes. 533 return unknown(); 534 } 535 536 SizeOffsetType ObjectSizeOffsetVisitor::visitGEPOperator(GEPOperator &GEP) { 537 SizeOffsetType PtrData = compute(GEP.getPointerOperand()); 538 APInt Offset(IntTyBits, 0); 539 if (!bothKnown(PtrData) || !GEP.accumulateConstantOffset(DL, Offset)) 540 return unknown(); 541 542 return std::make_pair(PtrData.first, PtrData.second + Offset); 543 } 544 545 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalAlias(GlobalAlias &GA) { 546 if (GA.mayBeOverridden()) 547 return unknown(); 548 return compute(GA.getAliasee()); 549 } 550 551 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalVariable(GlobalVariable &GV){ 552 if (!GV.hasDefinitiveInitializer()) 553 return unknown(); 554 555 APInt Size(IntTyBits, DL.getTypeAllocSize(GV.getType()->getElementType())); 556 return std::make_pair(align(Size, GV.getAlignment()), Zero); 557 } 558 559 SizeOffsetType ObjectSizeOffsetVisitor::visitIntToPtrInst(IntToPtrInst&) { 560 // clueless 561 return unknown(); 562 } 563 564 SizeOffsetType ObjectSizeOffsetVisitor::visitLoadInst(LoadInst&) { 565 ++ObjectVisitorLoad; 566 return unknown(); 567 } 568 569 SizeOffsetType ObjectSizeOffsetVisitor::visitPHINode(PHINode&) { 570 // too complex to analyze statically. 571 return unknown(); 572 } 573 574 SizeOffsetType ObjectSizeOffsetVisitor::visitSelectInst(SelectInst &I) { 575 SizeOffsetType TrueSide = compute(I.getTrueValue()); 576 SizeOffsetType FalseSide = compute(I.getFalseValue()); 577 if (bothKnown(TrueSide) && bothKnown(FalseSide) && TrueSide == FalseSide) 578 return TrueSide; 579 return unknown(); 580 } 581 582 SizeOffsetType ObjectSizeOffsetVisitor::visitUndefValue(UndefValue&) { 583 return std::make_pair(Zero, Zero); 584 } 585 586 SizeOffsetType ObjectSizeOffsetVisitor::visitInstruction(Instruction &I) { 587 DEBUG(dbgs() << "ObjectSizeOffsetVisitor unknown instruction:" << I << '\n'); 588 return unknown(); 589 } 590 591 ObjectSizeOffsetEvaluator::ObjectSizeOffsetEvaluator( 592 const DataLayout &DL, const TargetLibraryInfo *TLI, LLVMContext &Context, 593 bool RoundToAlign) 594 : DL(DL), TLI(TLI), Context(Context), Builder(Context, TargetFolder(DL)), 595 RoundToAlign(RoundToAlign) { 596 // IntTy and Zero must be set for each compute() since the address space may 597 // be different for later objects. 598 } 599 600 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute(Value *V) { 601 // XXX - Are vectors of pointers possible here? 602 IntTy = cast<IntegerType>(DL.getIntPtrType(V->getType())); 603 Zero = ConstantInt::get(IntTy, 0); 604 605 SizeOffsetEvalType Result = compute_(V); 606 607 if (!bothKnown(Result)) { 608 // erase everything that was computed in this iteration from the cache, so 609 // that no dangling references are left behind. We could be a bit smarter if 610 // we kept a dependency graph. It's probably not worth the complexity. 611 for (PtrSetTy::iterator I=SeenVals.begin(), E=SeenVals.end(); I != E; ++I) { 612 CacheMapTy::iterator CacheIt = CacheMap.find(*I); 613 // non-computable results can be safely cached 614 if (CacheIt != CacheMap.end() && anyKnown(CacheIt->second)) 615 CacheMap.erase(CacheIt); 616 } 617 } 618 619 SeenVals.clear(); 620 return Result; 621 } 622 623 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute_(Value *V) { 624 ObjectSizeOffsetVisitor Visitor(DL, TLI, Context, RoundToAlign); 625 SizeOffsetType Const = Visitor.compute(V); 626 if (Visitor.bothKnown(Const)) 627 return std::make_pair(ConstantInt::get(Context, Const.first), 628 ConstantInt::get(Context, Const.second)); 629 630 V = V->stripPointerCasts(); 631 632 // check cache 633 CacheMapTy::iterator CacheIt = CacheMap.find(V); 634 if (CacheIt != CacheMap.end()) 635 return CacheIt->second; 636 637 // always generate code immediately before the instruction being 638 // processed, so that the generated code dominates the same BBs 639 BuilderTy::InsertPointGuard Guard(Builder); 640 if (Instruction *I = dyn_cast<Instruction>(V)) 641 Builder.SetInsertPoint(I); 642 643 // now compute the size and offset 644 SizeOffsetEvalType Result; 645 646 // Record the pointers that were handled in this run, so that they can be 647 // cleaned later if something fails. We also use this set to break cycles that 648 // can occur in dead code. 649 if (!SeenVals.insert(V).second) { 650 Result = unknown(); 651 } else if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 652 Result = visitGEPOperator(*GEP); 653 } else if (Instruction *I = dyn_cast<Instruction>(V)) { 654 Result = visit(*I); 655 } else if (isa<Argument>(V) || 656 (isa<ConstantExpr>(V) && 657 cast<ConstantExpr>(V)->getOpcode() == Instruction::IntToPtr) || 658 isa<GlobalAlias>(V) || 659 isa<GlobalVariable>(V)) { 660 // ignore values where we cannot do more than what ObjectSizeVisitor can 661 Result = unknown(); 662 } else { 663 DEBUG(dbgs() << "ObjectSizeOffsetEvaluator::compute() unhandled value: " 664 << *V << '\n'); 665 Result = unknown(); 666 } 667 668 // Don't reuse CacheIt since it may be invalid at this point. 669 CacheMap[V] = Result; 670 return Result; 671 } 672 673 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitAllocaInst(AllocaInst &I) { 674 if (!I.getAllocatedType()->isSized()) 675 return unknown(); 676 677 // must be a VLA 678 assert(I.isArrayAllocation()); 679 Value *ArraySize = I.getArraySize(); 680 Value *Size = ConstantInt::get(ArraySize->getType(), 681 DL.getTypeAllocSize(I.getAllocatedType())); 682 Size = Builder.CreateMul(Size, ArraySize); 683 return std::make_pair(Size, Zero); 684 } 685 686 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitCallSite(CallSite CS) { 687 const AllocFnsTy *FnData = getAllocationData(CS.getInstruction(), AnyAlloc, 688 TLI); 689 if (!FnData) 690 return unknown(); 691 692 // handle strdup-like functions separately 693 if (FnData->AllocTy == StrDupLike) { 694 // TODO 695 return unknown(); 696 } 697 698 Value *FirstArg = CS.getArgument(FnData->FstParam); 699 FirstArg = Builder.CreateZExt(FirstArg, IntTy); 700 if (FnData->SndParam < 0) 701 return std::make_pair(FirstArg, Zero); 702 703 Value *SecondArg = CS.getArgument(FnData->SndParam); 704 SecondArg = Builder.CreateZExt(SecondArg, IntTy); 705 Value *Size = Builder.CreateMul(FirstArg, SecondArg); 706 return std::make_pair(Size, Zero); 707 708 // TODO: handle more standard functions (+ wchar cousins): 709 // - strdup / strndup 710 // - strcpy / strncpy 711 // - strcat / strncat 712 // - memcpy / memmove 713 // - strcat / strncat 714 // - memset 715 } 716 717 SizeOffsetEvalType 718 ObjectSizeOffsetEvaluator::visitExtractElementInst(ExtractElementInst&) { 719 return unknown(); 720 } 721 722 SizeOffsetEvalType 723 ObjectSizeOffsetEvaluator::visitExtractValueInst(ExtractValueInst&) { 724 return unknown(); 725 } 726 727 SizeOffsetEvalType 728 ObjectSizeOffsetEvaluator::visitGEPOperator(GEPOperator &GEP) { 729 SizeOffsetEvalType PtrData = compute_(GEP.getPointerOperand()); 730 if (!bothKnown(PtrData)) 731 return unknown(); 732 733 Value *Offset = EmitGEPOffset(&Builder, DL, &GEP, /*NoAssumptions=*/true); 734 Offset = Builder.CreateAdd(PtrData.second, Offset); 735 return std::make_pair(PtrData.first, Offset); 736 } 737 738 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitIntToPtrInst(IntToPtrInst&) { 739 // clueless 740 return unknown(); 741 } 742 743 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitLoadInst(LoadInst&) { 744 return unknown(); 745 } 746 747 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitPHINode(PHINode &PHI) { 748 // create 2 PHIs: one for size and another for offset 749 PHINode *SizePHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 750 PHINode *OffsetPHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 751 752 // insert right away in the cache to handle recursive PHIs 753 CacheMap[&PHI] = std::make_pair(SizePHI, OffsetPHI); 754 755 // compute offset/size for each PHI incoming pointer 756 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i) { 757 Builder.SetInsertPoint(&*PHI.getIncomingBlock(i)->getFirstInsertionPt()); 758 SizeOffsetEvalType EdgeData = compute_(PHI.getIncomingValue(i)); 759 760 if (!bothKnown(EdgeData)) { 761 OffsetPHI->replaceAllUsesWith(UndefValue::get(IntTy)); 762 OffsetPHI->eraseFromParent(); 763 SizePHI->replaceAllUsesWith(UndefValue::get(IntTy)); 764 SizePHI->eraseFromParent(); 765 return unknown(); 766 } 767 SizePHI->addIncoming(EdgeData.first, PHI.getIncomingBlock(i)); 768 OffsetPHI->addIncoming(EdgeData.second, PHI.getIncomingBlock(i)); 769 } 770 771 Value *Size = SizePHI, *Offset = OffsetPHI, *Tmp; 772 if ((Tmp = SizePHI->hasConstantValue())) { 773 Size = Tmp; 774 SizePHI->replaceAllUsesWith(Size); 775 SizePHI->eraseFromParent(); 776 } 777 if ((Tmp = OffsetPHI->hasConstantValue())) { 778 Offset = Tmp; 779 OffsetPHI->replaceAllUsesWith(Offset); 780 OffsetPHI->eraseFromParent(); 781 } 782 return std::make_pair(Size, Offset); 783 } 784 785 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitSelectInst(SelectInst &I) { 786 SizeOffsetEvalType TrueSide = compute_(I.getTrueValue()); 787 SizeOffsetEvalType FalseSide = compute_(I.getFalseValue()); 788 789 if (!bothKnown(TrueSide) || !bothKnown(FalseSide)) 790 return unknown(); 791 if (TrueSide == FalseSide) 792 return TrueSide; 793 794 Value *Size = Builder.CreateSelect(I.getCondition(), TrueSide.first, 795 FalseSide.first); 796 Value *Offset = Builder.CreateSelect(I.getCondition(), TrueSide.second, 797 FalseSide.second); 798 return std::make_pair(Size, Offset); 799 } 800 801 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitInstruction(Instruction &I) { 802 DEBUG(dbgs() << "ObjectSizeOffsetEvaluator unknown instruction:" << I <<'\n'); 803 return unknown(); 804 } 805