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 MallocOrCallocLike = MallocLike | CallocLike, 41 AllocLike = MallocLike | CallocLike | StrDupLike, 42 AnyAlloc = AllocLike | ReallocLike 43 }; 44 45 struct AllocFnsTy { 46 AllocType AllocTy; 47 unsigned NumParams; 48 // First and Second size parameters (or -1 if unused) 49 int 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 std::pair<LibFunc, 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 static const Function *getCalledFunction(const Value *V, bool LookThroughBitCast, 82 bool &IsNoBuiltin) { 83 // Don't care about intrinsics in this case. 84 if (isa<IntrinsicInst>(V)) 85 return nullptr; 86 87 if (LookThroughBitCast) 88 V = V->stripPointerCasts(); 89 90 ImmutableCallSite CS(V); 91 if (!CS.getInstruction()) 92 return nullptr; 93 94 IsNoBuiltin = CS.isNoBuiltin(); 95 96 const Function *Callee = CS.getCalledFunction(); 97 if (!Callee || !Callee->isDeclaration()) 98 return nullptr; 99 return Callee; 100 } 101 102 /// Returns the allocation data for the given value if it's either a call to a 103 /// known allocation function, or a call to a function with the allocsize 104 /// attribute. 105 static Optional<AllocFnsTy> 106 getAllocationDataForFunction(const Function *Callee, AllocType AllocTy, 107 const TargetLibraryInfo *TLI) { 108 // Make sure that the function is available. 109 StringRef FnName = Callee->getName(); 110 LibFunc TLIFn; 111 if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn)) 112 return None; 113 114 const auto *Iter = find_if( 115 AllocationFnData, [TLIFn](const std::pair<LibFunc, AllocFnsTy> &P) { 116 return P.first == TLIFn; 117 }); 118 119 if (Iter == std::end(AllocationFnData)) 120 return None; 121 122 const AllocFnsTy *FnData = &Iter->second; 123 if ((FnData->AllocTy & AllocTy) != FnData->AllocTy) 124 return None; 125 126 // Check function prototype. 127 int FstParam = FnData->FstParam; 128 int SndParam = FnData->SndParam; 129 FunctionType *FTy = Callee->getFunctionType(); 130 131 if (FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) && 132 FTy->getNumParams() == FnData->NumParams && 133 (FstParam < 0 || 134 (FTy->getParamType(FstParam)->isIntegerTy(32) || 135 FTy->getParamType(FstParam)->isIntegerTy(64))) && 136 (SndParam < 0 || 137 FTy->getParamType(SndParam)->isIntegerTy(32) || 138 FTy->getParamType(SndParam)->isIntegerTy(64))) 139 return *FnData; 140 return None; 141 } 142 143 static Optional<AllocFnsTy> getAllocationData(const Value *V, AllocType AllocTy, 144 const TargetLibraryInfo *TLI, 145 bool LookThroughBitCast = false) { 146 bool IsNoBuiltinCall; 147 if (const Function *Callee = 148 getCalledFunction(V, LookThroughBitCast, IsNoBuiltinCall)) 149 if (!IsNoBuiltinCall) 150 return getAllocationDataForFunction(Callee, AllocTy, TLI); 151 return None; 152 } 153 154 static Optional<AllocFnsTy> getAllocationSize(const Value *V, 155 const TargetLibraryInfo *TLI) { 156 bool IsNoBuiltinCall; 157 const Function *Callee = 158 getCalledFunction(V, /*LookThroughBitCast=*/false, IsNoBuiltinCall); 159 if (!Callee) 160 return None; 161 162 // Prefer to use existing information over allocsize. This will give us an 163 // accurate AllocTy. 164 if (!IsNoBuiltinCall) 165 if (Optional<AllocFnsTy> Data = 166 getAllocationDataForFunction(Callee, AnyAlloc, TLI)) 167 return Data; 168 169 Attribute Attr = Callee->getFnAttribute(Attribute::AllocSize); 170 if (Attr == Attribute()) 171 return None; 172 173 std::pair<unsigned, Optional<unsigned>> Args = Attr.getAllocSizeArgs(); 174 175 AllocFnsTy Result; 176 // Because allocsize only tells us how many bytes are allocated, we're not 177 // really allowed to assume anything, so we use MallocLike. 178 Result.AllocTy = MallocLike; 179 Result.NumParams = Callee->getNumOperands(); 180 Result.FstParam = Args.first; 181 Result.SndParam = Args.second.getValueOr(-1); 182 return Result; 183 } 184 185 static bool hasNoAliasAttr(const Value *V, bool LookThroughBitCast) { 186 ImmutableCallSite CS(LookThroughBitCast ? V->stripPointerCasts() : V); 187 return CS && CS.hasRetAttr(Attribute::NoAlias); 188 } 189 190 191 /// \brief Tests if a value is a call or invoke to a library function that 192 /// allocates or reallocates memory (either malloc, calloc, realloc, or strdup 193 /// like). 194 bool llvm::isAllocationFn(const Value *V, const TargetLibraryInfo *TLI, 195 bool LookThroughBitCast) { 196 return getAllocationData(V, AnyAlloc, TLI, LookThroughBitCast).hasValue(); 197 } 198 199 /// \brief Tests if a value is a call or invoke to a function that returns a 200 /// NoAlias pointer (including malloc/calloc/realloc/strdup-like functions). 201 bool llvm::isNoAliasFn(const Value *V, const TargetLibraryInfo *TLI, 202 bool LookThroughBitCast) { 203 // it's safe to consider realloc as noalias since accessing the original 204 // pointer is undefined behavior 205 return isAllocationFn(V, TLI, LookThroughBitCast) || 206 hasNoAliasAttr(V, LookThroughBitCast); 207 } 208 209 /// \brief Tests if a value is a call or invoke to a library function that 210 /// allocates uninitialized memory (such as malloc). 211 bool llvm::isMallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, 212 bool LookThroughBitCast) { 213 return getAllocationData(V, MallocLike, TLI, LookThroughBitCast).hasValue(); 214 } 215 216 /// \brief Tests if a value is a call or invoke to a library function that 217 /// allocates zero-filled memory (such as calloc). 218 bool llvm::isCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, 219 bool LookThroughBitCast) { 220 return getAllocationData(V, CallocLike, TLI, LookThroughBitCast).hasValue(); 221 } 222 223 /// \brief Tests if a value is a call or invoke to a library function that 224 /// allocates memory similiar to malloc or calloc. 225 bool llvm::isMallocOrCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI, 226 bool LookThroughBitCast) { 227 return getAllocationData(V, MallocOrCallocLike, TLI, 228 LookThroughBitCast).hasValue(); 229 } 230 231 /// \brief Tests if a value is a call or invoke to a library function that 232 /// allocates memory (either malloc, calloc, or strdup like). 233 bool llvm::isAllocLikeFn(const Value *V, const TargetLibraryInfo *TLI, 234 bool LookThroughBitCast) { 235 return getAllocationData(V, AllocLike, TLI, LookThroughBitCast).hasValue(); 236 } 237 238 /// extractMallocCall - Returns the corresponding CallInst if the instruction 239 /// is a malloc call. Since CallInst::CreateMalloc() only creates calls, we 240 /// ignore InvokeInst here. 241 const CallInst *llvm::extractMallocCall(const Value *I, 242 const TargetLibraryInfo *TLI) { 243 return isMallocLikeFn(I, TLI) ? dyn_cast<CallInst>(I) : nullptr; 244 } 245 246 static Value *computeArraySize(const CallInst *CI, const DataLayout &DL, 247 const TargetLibraryInfo *TLI, 248 bool LookThroughSExt = false) { 249 if (!CI) 250 return nullptr; 251 252 // The size of the malloc's result type must be known to determine array size. 253 Type *T = getMallocAllocatedType(CI, TLI); 254 if (!T || !T->isSized()) 255 return nullptr; 256 257 unsigned ElementSize = DL.getTypeAllocSize(T); 258 if (StructType *ST = dyn_cast<StructType>(T)) 259 ElementSize = DL.getStructLayout(ST)->getSizeInBytes(); 260 261 // If malloc call's arg can be determined to be a multiple of ElementSize, 262 // return the multiple. Otherwise, return NULL. 263 Value *MallocArg = CI->getArgOperand(0); 264 Value *Multiple = nullptr; 265 if (ComputeMultiple(MallocArg, ElementSize, Multiple, LookThroughSExt)) 266 return Multiple; 267 268 return nullptr; 269 } 270 271 /// getMallocType - Returns the PointerType resulting from the malloc call. 272 /// The PointerType depends on the number of bitcast uses of the malloc call: 273 /// 0: PointerType is the calls' return type. 274 /// 1: PointerType is the bitcast's result type. 275 /// >1: Unique PointerType cannot be determined, return NULL. 276 PointerType *llvm::getMallocType(const CallInst *CI, 277 const TargetLibraryInfo *TLI) { 278 assert(isMallocLikeFn(CI, TLI) && "getMallocType and not malloc call"); 279 280 PointerType *MallocType = nullptr; 281 unsigned NumOfBitCastUses = 0; 282 283 // Determine if CallInst has a bitcast use. 284 for (Value::const_user_iterator UI = CI->user_begin(), E = CI->user_end(); 285 UI != E;) 286 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(*UI++)) { 287 MallocType = cast<PointerType>(BCI->getDestTy()); 288 NumOfBitCastUses++; 289 } 290 291 // Malloc call has 1 bitcast use, so type is the bitcast's destination type. 292 if (NumOfBitCastUses == 1) 293 return MallocType; 294 295 // Malloc call was not bitcast, so type is the malloc function's return type. 296 if (NumOfBitCastUses == 0) 297 return cast<PointerType>(CI->getType()); 298 299 // Type could not be determined. 300 return nullptr; 301 } 302 303 /// getMallocAllocatedType - Returns the Type allocated by malloc call. 304 /// The Type depends on the number of bitcast uses of the malloc call: 305 /// 0: PointerType is the malloc calls' return type. 306 /// 1: PointerType is the bitcast's result type. 307 /// >1: Unique PointerType cannot be determined, return NULL. 308 Type *llvm::getMallocAllocatedType(const CallInst *CI, 309 const TargetLibraryInfo *TLI) { 310 PointerType *PT = getMallocType(CI, TLI); 311 return PT ? PT->getElementType() : nullptr; 312 } 313 314 /// getMallocArraySize - Returns the array size of a malloc call. If the 315 /// argument passed to malloc is a multiple of the size of the malloced type, 316 /// then return that multiple. For non-array mallocs, the multiple is 317 /// constant 1. Otherwise, return NULL for mallocs whose array size cannot be 318 /// determined. 319 Value *llvm::getMallocArraySize(CallInst *CI, const DataLayout &DL, 320 const TargetLibraryInfo *TLI, 321 bool LookThroughSExt) { 322 assert(isMallocLikeFn(CI, TLI) && "getMallocArraySize and not malloc call"); 323 return computeArraySize(CI, DL, TLI, LookThroughSExt); 324 } 325 326 327 /// extractCallocCall - Returns the corresponding CallInst if the instruction 328 /// is a calloc call. 329 const CallInst *llvm::extractCallocCall(const Value *I, 330 const TargetLibraryInfo *TLI) { 331 return isCallocLikeFn(I, TLI) ? cast<CallInst>(I) : nullptr; 332 } 333 334 335 /// isFreeCall - Returns non-null if the value is a call to the builtin free() 336 const CallInst *llvm::isFreeCall(const Value *I, const TargetLibraryInfo *TLI) { 337 const CallInst *CI = dyn_cast<CallInst>(I); 338 if (!CI || isa<IntrinsicInst>(CI)) 339 return nullptr; 340 Function *Callee = CI->getCalledFunction(); 341 if (Callee == nullptr) 342 return nullptr; 343 344 StringRef FnName = Callee->getName(); 345 LibFunc TLIFn; 346 if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn)) 347 return nullptr; 348 349 unsigned ExpectedNumParams; 350 if (TLIFn == LibFunc_free || 351 TLIFn == LibFunc_ZdlPv || // operator delete(void*) 352 TLIFn == LibFunc_ZdaPv || // operator delete[](void*) 353 TLIFn == LibFunc_msvc_delete_ptr32 || // operator delete(void*) 354 TLIFn == LibFunc_msvc_delete_ptr64 || // operator delete(void*) 355 TLIFn == LibFunc_msvc_delete_array_ptr32 || // operator delete[](void*) 356 TLIFn == LibFunc_msvc_delete_array_ptr64) // operator delete[](void*) 357 ExpectedNumParams = 1; 358 else if (TLIFn == LibFunc_ZdlPvj || // delete(void*, uint) 359 TLIFn == LibFunc_ZdlPvm || // delete(void*, ulong) 360 TLIFn == LibFunc_ZdlPvRKSt9nothrow_t || // delete(void*, nothrow) 361 TLIFn == LibFunc_ZdaPvj || // delete[](void*, uint) 362 TLIFn == LibFunc_ZdaPvm || // delete[](void*, ulong) 363 TLIFn == LibFunc_ZdaPvRKSt9nothrow_t || // delete[](void*, nothrow) 364 TLIFn == LibFunc_msvc_delete_ptr32_int || // delete(void*, uint) 365 TLIFn == LibFunc_msvc_delete_ptr64_longlong || // delete(void*, ulonglong) 366 TLIFn == LibFunc_msvc_delete_ptr32_nothrow || // delete(void*, nothrow) 367 TLIFn == LibFunc_msvc_delete_ptr64_nothrow || // delete(void*, nothrow) 368 TLIFn == LibFunc_msvc_delete_array_ptr32_int || // delete[](void*, uint) 369 TLIFn == LibFunc_msvc_delete_array_ptr64_longlong || // delete[](void*, ulonglong) 370 TLIFn == LibFunc_msvc_delete_array_ptr32_nothrow || // delete[](void*, nothrow) 371 TLIFn == LibFunc_msvc_delete_array_ptr64_nothrow) // delete[](void*, nothrow) 372 ExpectedNumParams = 2; 373 else 374 return nullptr; 375 376 // Check free prototype. 377 // FIXME: workaround for PR5130, this will be obsolete when a nobuiltin 378 // attribute will exist. 379 FunctionType *FTy = Callee->getFunctionType(); 380 if (!FTy->getReturnType()->isVoidTy()) 381 return nullptr; 382 if (FTy->getNumParams() != ExpectedNumParams) 383 return nullptr; 384 if (FTy->getParamType(0) != Type::getInt8PtrTy(Callee->getContext())) 385 return nullptr; 386 387 return CI; 388 } 389 390 391 392 //===----------------------------------------------------------------------===// 393 // Utility functions to compute size of objects. 394 // 395 static APInt getSizeWithOverflow(const SizeOffsetType &Data) { 396 if (Data.second.isNegative() || Data.first.ult(Data.second)) 397 return APInt(Data.first.getBitWidth(), 0); 398 return Data.first - Data.second; 399 } 400 401 /// \brief Compute the size of the object pointed by Ptr. Returns true and the 402 /// object size in Size if successful, and false otherwise. 403 /// If RoundToAlign is true, then Size is rounded up to the aligment of allocas, 404 /// byval arguments, and global variables. 405 bool llvm::getObjectSize(const Value *Ptr, uint64_t &Size, const DataLayout &DL, 406 const TargetLibraryInfo *TLI, ObjectSizeOpts Opts) { 407 ObjectSizeOffsetVisitor Visitor(DL, TLI, Ptr->getContext(), Opts); 408 SizeOffsetType Data = Visitor.compute(const_cast<Value*>(Ptr)); 409 if (!Visitor.bothKnown(Data)) 410 return false; 411 412 Size = getSizeWithOverflow(Data).getZExtValue(); 413 return true; 414 } 415 416 ConstantInt *llvm::lowerObjectSizeCall(IntrinsicInst *ObjectSize, 417 const DataLayout &DL, 418 const TargetLibraryInfo *TLI, 419 bool MustSucceed) { 420 assert(ObjectSize->getIntrinsicID() == Intrinsic::objectsize && 421 "ObjectSize must be a call to llvm.objectsize!"); 422 423 bool MaxVal = cast<ConstantInt>(ObjectSize->getArgOperand(1))->isZero(); 424 ObjectSizeOpts EvalOptions; 425 // Unless we have to fold this to something, try to be as accurate as 426 // possible. 427 if (MustSucceed) 428 EvalOptions.EvalMode = 429 MaxVal ? ObjectSizeOpts::Mode::Max : ObjectSizeOpts::Mode::Min; 430 else 431 EvalOptions.EvalMode = ObjectSizeOpts::Mode::Exact; 432 433 EvalOptions.NullIsUnknownSize = 434 cast<ConstantInt>(ObjectSize->getArgOperand(2))->isOne(); 435 436 // FIXME: Does it make sense to just return a failure value if the size won't 437 // fit in the output and `!MustSucceed`? 438 uint64_t Size; 439 auto *ResultType = cast<IntegerType>(ObjectSize->getType()); 440 if (getObjectSize(ObjectSize->getArgOperand(0), Size, DL, TLI, EvalOptions) && 441 isUIntN(ResultType->getBitWidth(), Size)) 442 return ConstantInt::get(ResultType, Size); 443 444 if (!MustSucceed) 445 return nullptr; 446 447 return ConstantInt::get(ResultType, MaxVal ? -1ULL : 0); 448 } 449 450 STATISTIC(ObjectVisitorArgument, 451 "Number of arguments with unsolved size and offset"); 452 STATISTIC(ObjectVisitorLoad, 453 "Number of load instructions with unsolved size and offset"); 454 455 456 APInt ObjectSizeOffsetVisitor::align(APInt Size, uint64_t Align) { 457 if (Options.RoundToAlign && Align) 458 return APInt(IntTyBits, alignTo(Size.getZExtValue(), Align)); 459 return Size; 460 } 461 462 ObjectSizeOffsetVisitor::ObjectSizeOffsetVisitor(const DataLayout &DL, 463 const TargetLibraryInfo *TLI, 464 LLVMContext &Context, 465 ObjectSizeOpts Options) 466 : DL(DL), TLI(TLI), Options(Options) { 467 // Pointer size must be rechecked for each object visited since it could have 468 // a different address space. 469 } 470 471 SizeOffsetType ObjectSizeOffsetVisitor::compute(Value *V) { 472 IntTyBits = DL.getPointerTypeSizeInBits(V->getType()); 473 Zero = APInt::getNullValue(IntTyBits); 474 475 V = V->stripPointerCasts(); 476 if (Instruction *I = dyn_cast<Instruction>(V)) { 477 // If we have already seen this instruction, bail out. Cycles can happen in 478 // unreachable code after constant propagation. 479 if (!SeenInsts.insert(I).second) 480 return unknown(); 481 482 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) 483 return visitGEPOperator(*GEP); 484 return visit(*I); 485 } 486 if (Argument *A = dyn_cast<Argument>(V)) 487 return visitArgument(*A); 488 if (ConstantPointerNull *P = dyn_cast<ConstantPointerNull>(V)) 489 return visitConstantPointerNull(*P); 490 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) 491 return visitGlobalAlias(*GA); 492 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 493 return visitGlobalVariable(*GV); 494 if (UndefValue *UV = dyn_cast<UndefValue>(V)) 495 return visitUndefValue(*UV); 496 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 497 if (CE->getOpcode() == Instruction::IntToPtr) 498 return unknown(); // clueless 499 if (CE->getOpcode() == Instruction::GetElementPtr) 500 return visitGEPOperator(cast<GEPOperator>(*CE)); 501 } 502 503 DEBUG(dbgs() << "ObjectSizeOffsetVisitor::compute() unhandled value: " << *V 504 << '\n'); 505 return unknown(); 506 } 507 508 SizeOffsetType ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) { 509 if (!I.getAllocatedType()->isSized()) 510 return unknown(); 511 512 APInt Size(IntTyBits, DL.getTypeAllocSize(I.getAllocatedType())); 513 if (!I.isArrayAllocation()) 514 return std::make_pair(align(Size, I.getAlignment()), Zero); 515 516 Value *ArraySize = I.getArraySize(); 517 if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) { 518 Size *= C->getValue().zextOrSelf(IntTyBits); 519 return std::make_pair(align(Size, I.getAlignment()), Zero); 520 } 521 return unknown(); 522 } 523 524 SizeOffsetType ObjectSizeOffsetVisitor::visitArgument(Argument &A) { 525 // No interprocedural analysis is done at the moment. 526 if (!A.hasByValOrInAllocaAttr()) { 527 ++ObjectVisitorArgument; 528 return unknown(); 529 } 530 PointerType *PT = cast<PointerType>(A.getType()); 531 APInt Size(IntTyBits, DL.getTypeAllocSize(PT->getElementType())); 532 return std::make_pair(align(Size, A.getParamAlignment()), Zero); 533 } 534 535 SizeOffsetType ObjectSizeOffsetVisitor::visitCallSite(CallSite CS) { 536 Optional<AllocFnsTy> FnData = getAllocationSize(CS.getInstruction(), TLI); 537 if (!FnData) 538 return unknown(); 539 540 // Handle strdup-like functions separately. 541 if (FnData->AllocTy == StrDupLike) { 542 APInt Size(IntTyBits, GetStringLength(CS.getArgument(0))); 543 if (!Size) 544 return unknown(); 545 546 // Strndup limits strlen. 547 if (FnData->FstParam > 0) { 548 ConstantInt *Arg = 549 dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam)); 550 if (!Arg) 551 return unknown(); 552 553 APInt MaxSize = Arg->getValue().zextOrSelf(IntTyBits); 554 if (Size.ugt(MaxSize)) 555 Size = MaxSize + 1; 556 } 557 return std::make_pair(Size, Zero); 558 } 559 560 ConstantInt *Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam)); 561 if (!Arg) 562 return unknown(); 563 564 // When we're compiling N-bit code, and the user uses parameters that are 565 // greater than N bits (e.g. uint64_t on a 32-bit build), we can run into 566 // trouble with APInt size issues. This function handles resizing + overflow 567 // checks for us. 568 auto CheckedZextOrTrunc = [&](APInt &I) { 569 // More bits than we can handle. Checking the bit width isn't necessary, but 570 // it's faster than checking active bits, and should give `false` in the 571 // vast majority of cases. 572 if (I.getBitWidth() > IntTyBits && I.getActiveBits() > IntTyBits) 573 return false; 574 if (I.getBitWidth() != IntTyBits) 575 I = I.zextOrTrunc(IntTyBits); 576 return true; 577 }; 578 579 APInt Size = Arg->getValue(); 580 if (!CheckedZextOrTrunc(Size)) 581 return unknown(); 582 583 // Size is determined by just 1 parameter. 584 if (FnData->SndParam < 0) 585 return std::make_pair(Size, Zero); 586 587 Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->SndParam)); 588 if (!Arg) 589 return unknown(); 590 591 APInt NumElems = Arg->getValue(); 592 if (!CheckedZextOrTrunc(NumElems)) 593 return unknown(); 594 595 bool Overflow; 596 Size = Size.umul_ov(NumElems, Overflow); 597 return Overflow ? unknown() : std::make_pair(Size, Zero); 598 599 // TODO: handle more standard functions (+ wchar cousins): 600 // - strdup / strndup 601 // - strcpy / strncpy 602 // - strcat / strncat 603 // - memcpy / memmove 604 // - strcat / strncat 605 // - memset 606 } 607 608 SizeOffsetType 609 ObjectSizeOffsetVisitor::visitConstantPointerNull(ConstantPointerNull& CPN) { 610 if (Options.NullIsUnknownSize && CPN.getType()->getAddressSpace() == 0) 611 return unknown(); 612 return std::make_pair(Zero, Zero); 613 } 614 615 SizeOffsetType 616 ObjectSizeOffsetVisitor::visitExtractElementInst(ExtractElementInst&) { 617 return unknown(); 618 } 619 620 SizeOffsetType 621 ObjectSizeOffsetVisitor::visitExtractValueInst(ExtractValueInst&) { 622 // Easy cases were already folded by previous passes. 623 return unknown(); 624 } 625 626 SizeOffsetType ObjectSizeOffsetVisitor::visitGEPOperator(GEPOperator &GEP) { 627 SizeOffsetType PtrData = compute(GEP.getPointerOperand()); 628 APInt Offset(IntTyBits, 0); 629 if (!bothKnown(PtrData) || !GEP.accumulateConstantOffset(DL, Offset)) 630 return unknown(); 631 632 return std::make_pair(PtrData.first, PtrData.second + Offset); 633 } 634 635 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalAlias(GlobalAlias &GA) { 636 if (GA.isInterposable()) 637 return unknown(); 638 return compute(GA.getAliasee()); 639 } 640 641 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalVariable(GlobalVariable &GV){ 642 if (!GV.hasDefinitiveInitializer()) 643 return unknown(); 644 645 APInt Size(IntTyBits, DL.getTypeAllocSize(GV.getType()->getElementType())); 646 return std::make_pair(align(Size, GV.getAlignment()), Zero); 647 } 648 649 SizeOffsetType ObjectSizeOffsetVisitor::visitIntToPtrInst(IntToPtrInst&) { 650 // clueless 651 return unknown(); 652 } 653 654 SizeOffsetType ObjectSizeOffsetVisitor::visitLoadInst(LoadInst&) { 655 ++ObjectVisitorLoad; 656 return unknown(); 657 } 658 659 SizeOffsetType ObjectSizeOffsetVisitor::visitPHINode(PHINode&) { 660 // too complex to analyze statically. 661 return unknown(); 662 } 663 664 SizeOffsetType ObjectSizeOffsetVisitor::visitSelectInst(SelectInst &I) { 665 SizeOffsetType TrueSide = compute(I.getTrueValue()); 666 SizeOffsetType FalseSide = compute(I.getFalseValue()); 667 if (bothKnown(TrueSide) && bothKnown(FalseSide)) { 668 if (TrueSide == FalseSide) { 669 return TrueSide; 670 } 671 672 APInt TrueResult = getSizeWithOverflow(TrueSide); 673 APInt FalseResult = getSizeWithOverflow(FalseSide); 674 675 if (TrueResult == FalseResult) { 676 return TrueSide; 677 } 678 if (Options.EvalMode == ObjectSizeOpts::Mode::Min) { 679 if (TrueResult.slt(FalseResult)) 680 return TrueSide; 681 return FalseSide; 682 } 683 if (Options.EvalMode == ObjectSizeOpts::Mode::Max) { 684 if (TrueResult.sgt(FalseResult)) 685 return TrueSide; 686 return FalseSide; 687 } 688 } 689 return unknown(); 690 } 691 692 SizeOffsetType ObjectSizeOffsetVisitor::visitUndefValue(UndefValue&) { 693 return std::make_pair(Zero, Zero); 694 } 695 696 SizeOffsetType ObjectSizeOffsetVisitor::visitInstruction(Instruction &I) { 697 DEBUG(dbgs() << "ObjectSizeOffsetVisitor unknown instruction:" << I << '\n'); 698 return unknown(); 699 } 700 701 ObjectSizeOffsetEvaluator::ObjectSizeOffsetEvaluator( 702 const DataLayout &DL, const TargetLibraryInfo *TLI, LLVMContext &Context, 703 bool RoundToAlign) 704 : DL(DL), TLI(TLI), Context(Context), Builder(Context, TargetFolder(DL)), 705 RoundToAlign(RoundToAlign) { 706 // IntTy and Zero must be set for each compute() since the address space may 707 // be different for later objects. 708 } 709 710 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute(Value *V) { 711 // XXX - Are vectors of pointers possible here? 712 IntTy = cast<IntegerType>(DL.getIntPtrType(V->getType())); 713 Zero = ConstantInt::get(IntTy, 0); 714 715 SizeOffsetEvalType Result = compute_(V); 716 717 if (!bothKnown(Result)) { 718 // Erase everything that was computed in this iteration from the cache, so 719 // that no dangling references are left behind. We could be a bit smarter if 720 // we kept a dependency graph. It's probably not worth the complexity. 721 for (const Value *SeenVal : SeenVals) { 722 CacheMapTy::iterator CacheIt = CacheMap.find(SeenVal); 723 // non-computable results can be safely cached 724 if (CacheIt != CacheMap.end() && anyKnown(CacheIt->second)) 725 CacheMap.erase(CacheIt); 726 } 727 } 728 729 SeenVals.clear(); 730 return Result; 731 } 732 733 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute_(Value *V) { 734 ObjectSizeOpts ObjSizeOptions; 735 ObjSizeOptions.RoundToAlign = RoundToAlign; 736 737 ObjectSizeOffsetVisitor Visitor(DL, TLI, Context, ObjSizeOptions); 738 SizeOffsetType Const = Visitor.compute(V); 739 if (Visitor.bothKnown(Const)) 740 return std::make_pair(ConstantInt::get(Context, Const.first), 741 ConstantInt::get(Context, Const.second)); 742 743 V = V->stripPointerCasts(); 744 745 // Check cache. 746 CacheMapTy::iterator CacheIt = CacheMap.find(V); 747 if (CacheIt != CacheMap.end()) 748 return CacheIt->second; 749 750 // Always generate code immediately before the instruction being 751 // processed, so that the generated code dominates the same BBs. 752 BuilderTy::InsertPointGuard Guard(Builder); 753 if (Instruction *I = dyn_cast<Instruction>(V)) 754 Builder.SetInsertPoint(I); 755 756 // Now compute the size and offset. 757 SizeOffsetEvalType Result; 758 759 // Record the pointers that were handled in this run, so that they can be 760 // cleaned later if something fails. We also use this set to break cycles that 761 // can occur in dead code. 762 if (!SeenVals.insert(V).second) { 763 Result = unknown(); 764 } else if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 765 Result = visitGEPOperator(*GEP); 766 } else if (Instruction *I = dyn_cast<Instruction>(V)) { 767 Result = visit(*I); 768 } else if (isa<Argument>(V) || 769 (isa<ConstantExpr>(V) && 770 cast<ConstantExpr>(V)->getOpcode() == Instruction::IntToPtr) || 771 isa<GlobalAlias>(V) || 772 isa<GlobalVariable>(V)) { 773 // Ignore values where we cannot do more than ObjectSizeVisitor. 774 Result = unknown(); 775 } else { 776 DEBUG(dbgs() << "ObjectSizeOffsetEvaluator::compute() unhandled value: " 777 << *V << '\n'); 778 Result = unknown(); 779 } 780 781 // Don't reuse CacheIt since it may be invalid at this point. 782 CacheMap[V] = Result; 783 return Result; 784 } 785 786 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitAllocaInst(AllocaInst &I) { 787 if (!I.getAllocatedType()->isSized()) 788 return unknown(); 789 790 // must be a VLA 791 assert(I.isArrayAllocation()); 792 Value *ArraySize = I.getArraySize(); 793 Value *Size = ConstantInt::get(ArraySize->getType(), 794 DL.getTypeAllocSize(I.getAllocatedType())); 795 Size = Builder.CreateMul(Size, ArraySize); 796 return std::make_pair(Size, Zero); 797 } 798 799 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitCallSite(CallSite CS) { 800 Optional<AllocFnsTy> FnData = getAllocationSize(CS.getInstruction(), TLI); 801 if (!FnData) 802 return unknown(); 803 804 // Handle strdup-like functions separately. 805 if (FnData->AllocTy == StrDupLike) { 806 // TODO 807 return unknown(); 808 } 809 810 Value *FirstArg = CS.getArgument(FnData->FstParam); 811 FirstArg = Builder.CreateZExt(FirstArg, IntTy); 812 if (FnData->SndParam < 0) 813 return std::make_pair(FirstArg, Zero); 814 815 Value *SecondArg = CS.getArgument(FnData->SndParam); 816 SecondArg = Builder.CreateZExt(SecondArg, IntTy); 817 Value *Size = Builder.CreateMul(FirstArg, SecondArg); 818 return std::make_pair(Size, Zero); 819 820 // TODO: handle more standard functions (+ wchar cousins): 821 // - strdup / strndup 822 // - strcpy / strncpy 823 // - strcat / strncat 824 // - memcpy / memmove 825 // - strcat / strncat 826 // - memset 827 } 828 829 SizeOffsetEvalType 830 ObjectSizeOffsetEvaluator::visitExtractElementInst(ExtractElementInst&) { 831 return unknown(); 832 } 833 834 SizeOffsetEvalType 835 ObjectSizeOffsetEvaluator::visitExtractValueInst(ExtractValueInst&) { 836 return unknown(); 837 } 838 839 SizeOffsetEvalType 840 ObjectSizeOffsetEvaluator::visitGEPOperator(GEPOperator &GEP) { 841 SizeOffsetEvalType PtrData = compute_(GEP.getPointerOperand()); 842 if (!bothKnown(PtrData)) 843 return unknown(); 844 845 Value *Offset = EmitGEPOffset(&Builder, DL, &GEP, /*NoAssumptions=*/true); 846 Offset = Builder.CreateAdd(PtrData.second, Offset); 847 return std::make_pair(PtrData.first, Offset); 848 } 849 850 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitIntToPtrInst(IntToPtrInst&) { 851 // clueless 852 return unknown(); 853 } 854 855 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitLoadInst(LoadInst&) { 856 return unknown(); 857 } 858 859 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitPHINode(PHINode &PHI) { 860 // Create 2 PHIs: one for size and another for offset. 861 PHINode *SizePHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 862 PHINode *OffsetPHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 863 864 // Insert right away in the cache to handle recursive PHIs. 865 CacheMap[&PHI] = std::make_pair(SizePHI, OffsetPHI); 866 867 // Compute offset/size for each PHI incoming pointer. 868 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i) { 869 Builder.SetInsertPoint(&*PHI.getIncomingBlock(i)->getFirstInsertionPt()); 870 SizeOffsetEvalType EdgeData = compute_(PHI.getIncomingValue(i)); 871 872 if (!bothKnown(EdgeData)) { 873 OffsetPHI->replaceAllUsesWith(UndefValue::get(IntTy)); 874 OffsetPHI->eraseFromParent(); 875 SizePHI->replaceAllUsesWith(UndefValue::get(IntTy)); 876 SizePHI->eraseFromParent(); 877 return unknown(); 878 } 879 SizePHI->addIncoming(EdgeData.first, PHI.getIncomingBlock(i)); 880 OffsetPHI->addIncoming(EdgeData.second, PHI.getIncomingBlock(i)); 881 } 882 883 Value *Size = SizePHI, *Offset = OffsetPHI, *Tmp; 884 if ((Tmp = SizePHI->hasConstantValue())) { 885 Size = Tmp; 886 SizePHI->replaceAllUsesWith(Size); 887 SizePHI->eraseFromParent(); 888 } 889 if ((Tmp = OffsetPHI->hasConstantValue())) { 890 Offset = Tmp; 891 OffsetPHI->replaceAllUsesWith(Offset); 892 OffsetPHI->eraseFromParent(); 893 } 894 return std::make_pair(Size, Offset); 895 } 896 897 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitSelectInst(SelectInst &I) { 898 SizeOffsetEvalType TrueSide = compute_(I.getTrueValue()); 899 SizeOffsetEvalType FalseSide = compute_(I.getFalseValue()); 900 901 if (!bothKnown(TrueSide) || !bothKnown(FalseSide)) 902 return unknown(); 903 if (TrueSide == FalseSide) 904 return TrueSide; 905 906 Value *Size = Builder.CreateSelect(I.getCondition(), TrueSide.first, 907 FalseSide.first); 908 Value *Offset = Builder.CreateSelect(I.getCondition(), TrueSide.second, 909 FalseSide.second); 910 return std::make_pair(Size, Offset); 911 } 912 913 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitInstruction(Instruction &I) { 914 DEBUG(dbgs() << "ObjectSizeOffsetEvaluator unknown instruction:" << I <<'\n'); 915 return unknown(); 916 } 917