1 //===- MemoryBuiltins.cpp - Identify calls to memory builtins -------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This family of functions identifies calls to builtin functions that allocate 10 // or free memory. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Analysis/MemoryBuiltins.h" 15 #include "llvm/ADT/APInt.h" 16 #include "llvm/ADT/None.h" 17 #include "llvm/ADT/Optional.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/Statistic.h" 20 #include "llvm/Analysis/AliasAnalysis.h" 21 #include "llvm/Analysis/TargetFolder.h" 22 #include "llvm/Analysis/TargetLibraryInfo.h" 23 #include "llvm/Analysis/Utils/Local.h" 24 #include "llvm/Analysis/ValueTracking.h" 25 #include "llvm/IR/Argument.h" 26 #include "llvm/IR/Attributes.h" 27 #include "llvm/IR/Constants.h" 28 #include "llvm/IR/DataLayout.h" 29 #include "llvm/IR/DerivedTypes.h" 30 #include "llvm/IR/Function.h" 31 #include "llvm/IR/GlobalAlias.h" 32 #include "llvm/IR/GlobalVariable.h" 33 #include "llvm/IR/Instruction.h" 34 #include "llvm/IR/Instructions.h" 35 #include "llvm/IR/IntrinsicInst.h" 36 #include "llvm/IR/Operator.h" 37 #include "llvm/IR/Type.h" 38 #include "llvm/IR/Value.h" 39 #include "llvm/Support/Casting.h" 40 #include "llvm/Support/Debug.h" 41 #include "llvm/Support/MathExtras.h" 42 #include "llvm/Support/raw_ostream.h" 43 #include <cassert> 44 #include <cstdint> 45 #include <iterator> 46 #include <numeric> 47 #include <type_traits> 48 #include <utility> 49 50 using namespace llvm; 51 52 #define DEBUG_TYPE "memory-builtins" 53 54 enum AllocType : uint8_t { 55 OpNewLike = 1<<0, // allocates; never returns null 56 MallocLike = 1<<1, // allocates; may return null 57 AlignedAllocLike = 1<<2, // allocates with alignment; may return null 58 CallocLike = 1<<3, // allocates + bzero 59 ReallocLike = 1<<4, // reallocates 60 StrDupLike = 1<<5, 61 MallocOrOpNewLike = MallocLike | OpNewLike, 62 MallocOrCallocLike = MallocLike | OpNewLike | CallocLike | AlignedAllocLike, 63 AllocLike = MallocOrCallocLike | StrDupLike, 64 AnyAlloc = AllocLike | ReallocLike 65 }; 66 67 enum class MallocFamily { 68 Malloc, 69 CPPNew, // new(unsigned int) 70 CPPNewAligned, // new(unsigned int, align_val_t) 71 CPPNewArray, // new[](unsigned int) 72 CPPNewArrayAligned, // new[](unsigned long, align_val_t) 73 MSVCNew, // new(unsigned int) 74 MSVCArrayNew, // new[](unsigned int) 75 VecMalloc, 76 KmpcAllocShared, 77 }; 78 79 StringRef mangledNameForMallocFamily(const MallocFamily &Family) { 80 switch (Family) { 81 case MallocFamily::Malloc: 82 return "malloc"; 83 case MallocFamily::CPPNew: 84 return "_Znwm"; 85 case MallocFamily::CPPNewAligned: 86 return "_ZnwmSt11align_val_t"; 87 case MallocFamily::CPPNewArray: 88 return "_Znam"; 89 case MallocFamily::CPPNewArrayAligned: 90 return "_ZnamSt11align_val_t"; 91 case MallocFamily::MSVCNew: 92 return "??2@YAPAXI@Z"; 93 case MallocFamily::MSVCArrayNew: 94 return "??_U@YAPAXI@Z"; 95 case MallocFamily::VecMalloc: 96 return "vec_malloc"; 97 case MallocFamily::KmpcAllocShared: 98 return "__kmpc_alloc_shared"; 99 } 100 llvm_unreachable("missing an alloc family"); 101 } 102 103 struct AllocFnsTy { 104 AllocType AllocTy; 105 unsigned NumParams; 106 // First and Second size parameters (or -1 if unused) 107 int FstParam, SndParam; 108 // Alignment parameter for aligned_alloc and aligned new 109 int AlignParam; 110 // Name of default allocator function to group malloc/free calls by family 111 MallocFamily Family; 112 }; 113 114 // clang-format off 115 // FIXME: certain users need more information. E.g., SimplifyLibCalls needs to 116 // know which functions are nounwind, noalias, nocapture parameters, etc. 117 static const std::pair<LibFunc, AllocFnsTy> AllocationFnData[] = { 118 {LibFunc_malloc, {MallocLike, 1, 0, -1, -1, MallocFamily::Malloc}}, 119 {LibFunc_vec_malloc, {MallocLike, 1, 0, -1, -1, MallocFamily::VecMalloc}}, 120 {LibFunc_valloc, {MallocLike, 1, 0, -1, -1, MallocFamily::Malloc}}, 121 {LibFunc_Znwj, {OpNewLike, 1, 0, -1, -1, MallocFamily::CPPNew}}, // new(unsigned int) 122 {LibFunc_ZnwjRKSt9nothrow_t, {MallocLike, 2, 0, -1, -1, MallocFamily::CPPNew}}, // new(unsigned int, nothrow) 123 {LibFunc_ZnwjSt11align_val_t, {OpNewLike, 2, 0, -1, 1, MallocFamily::CPPNewAligned}}, // new(unsigned int, align_val_t) 124 {LibFunc_ZnwjSt11align_val_tRKSt9nothrow_t, {MallocLike, 3, 0, -1, 1, MallocFamily::CPPNewAligned}}, // new(unsigned int, align_val_t, nothrow) 125 {LibFunc_Znwm, {OpNewLike, 1, 0, -1, -1, MallocFamily::CPPNew}}, // new(unsigned long) 126 {LibFunc_ZnwmRKSt9nothrow_t, {MallocLike, 2, 0, -1, -1, MallocFamily::CPPNew}}, // new(unsigned long, nothrow) 127 {LibFunc_ZnwmSt11align_val_t, {OpNewLike, 2, 0, -1, 1, MallocFamily::CPPNewAligned}}, // new(unsigned long, align_val_t) 128 {LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t, {MallocLike, 3, 0, -1, 1, MallocFamily::CPPNewAligned}}, // new(unsigned long, align_val_t, nothrow) 129 {LibFunc_Znaj, {OpNewLike, 1, 0, -1, -1, MallocFamily::CPPNewArray}}, // new[](unsigned int) 130 {LibFunc_ZnajRKSt9nothrow_t, {MallocLike, 2, 0, -1, -1, MallocFamily::CPPNewArray}}, // new[](unsigned int, nothrow) 131 {LibFunc_ZnajSt11align_val_t, {OpNewLike, 2, 0, -1, 1, MallocFamily::CPPNewArrayAligned}}, // new[](unsigned int, align_val_t) 132 {LibFunc_ZnajSt11align_val_tRKSt9nothrow_t, {MallocLike, 3, 0, -1, 1, MallocFamily::CPPNewArrayAligned}}, // new[](unsigned int, align_val_t, nothrow) 133 {LibFunc_Znam, {OpNewLike, 1, 0, -1, -1, MallocFamily::CPPNewArray}}, // new[](unsigned long) 134 {LibFunc_ZnamRKSt9nothrow_t, {MallocLike, 2, 0, -1, -1, MallocFamily::CPPNewArray}}, // new[](unsigned long, nothrow) 135 {LibFunc_ZnamSt11align_val_t, {OpNewLike, 2, 0, -1, 1, MallocFamily::CPPNewArrayAligned}}, // new[](unsigned long, align_val_t) 136 {LibFunc_ZnamSt11align_val_tRKSt9nothrow_t, {MallocLike, 3, 0, -1, 1, MallocFamily::CPPNewArrayAligned}}, // new[](unsigned long, align_val_t, nothrow) 137 {LibFunc_msvc_new_int, {OpNewLike, 1, 0, -1, -1, MallocFamily::MSVCNew}}, // new(unsigned int) 138 {LibFunc_msvc_new_int_nothrow, {MallocLike, 2, 0, -1, -1, MallocFamily::MSVCNew}}, // new(unsigned int, nothrow) 139 {LibFunc_msvc_new_longlong, {OpNewLike, 1, 0, -1, -1, MallocFamily::MSVCNew}}, // new(unsigned long long) 140 {LibFunc_msvc_new_longlong_nothrow, {MallocLike, 2, 0, -1, -1, MallocFamily::MSVCNew}}, // new(unsigned long long, nothrow) 141 {LibFunc_msvc_new_array_int, {OpNewLike, 1, 0, -1, -1, MallocFamily::MSVCArrayNew}}, // new[](unsigned int) 142 {LibFunc_msvc_new_array_int_nothrow, {MallocLike, 2, 0, -1, -1, MallocFamily::MSVCArrayNew}}, // new[](unsigned int, nothrow) 143 {LibFunc_msvc_new_array_longlong, {OpNewLike, 1, 0, -1, -1, MallocFamily::MSVCArrayNew}}, // new[](unsigned long long) 144 {LibFunc_msvc_new_array_longlong_nothrow, {MallocLike, 2, 0, -1, -1, MallocFamily::MSVCArrayNew}}, // new[](unsigned long long, nothrow) 145 {LibFunc_aligned_alloc, {AlignedAllocLike, 2, 1, -1, 0, MallocFamily::Malloc}}, 146 {LibFunc_memalign, {AlignedAllocLike, 2, 1, -1, 0, MallocFamily::Malloc}}, 147 {LibFunc_calloc, {CallocLike, 2, 0, 1, -1, MallocFamily::Malloc}}, 148 {LibFunc_vec_calloc, {CallocLike, 2, 0, 1, -1, MallocFamily::VecMalloc}}, 149 {LibFunc_realloc, {ReallocLike, 2, 1, -1, -1, MallocFamily::Malloc}}, 150 {LibFunc_vec_realloc, {ReallocLike, 2, 1, -1, -1, MallocFamily::VecMalloc}}, 151 {LibFunc_reallocf, {ReallocLike, 2, 1, -1, -1, MallocFamily::Malloc}}, 152 {LibFunc_strdup, {StrDupLike, 1, -1, -1, -1, MallocFamily::Malloc}}, 153 {LibFunc_dunder_strdup, {StrDupLike, 1, -1, -1, -1, MallocFamily::Malloc}}, 154 {LibFunc_strndup, {StrDupLike, 2, 1, -1, -1, MallocFamily::Malloc}}, 155 {LibFunc_dunder_strndup, {StrDupLike, 2, 1, -1, -1, MallocFamily::Malloc}}, 156 {LibFunc___kmpc_alloc_shared, {MallocLike, 1, 0, -1, -1, MallocFamily::KmpcAllocShared}}, 157 }; 158 // clang-format on 159 160 static const Function *getCalledFunction(const Value *V, 161 bool &IsNoBuiltin) { 162 // Don't care about intrinsics in this case. 163 if (isa<IntrinsicInst>(V)) 164 return nullptr; 165 166 const auto *CB = dyn_cast<CallBase>(V); 167 if (!CB) 168 return nullptr; 169 170 IsNoBuiltin = CB->isNoBuiltin(); 171 172 if (const Function *Callee = CB->getCalledFunction()) 173 return Callee; 174 return nullptr; 175 } 176 177 /// Returns the allocation data for the given value if it's a call to a known 178 /// allocation function. 179 static Optional<AllocFnsTy> 180 getAllocationDataForFunction(const Function *Callee, AllocType AllocTy, 181 const TargetLibraryInfo *TLI) { 182 // Don't perform a slow TLI lookup, if this function doesn't return a pointer 183 // and thus can't be an allocation function. 184 if (!Callee->getReturnType()->isPointerTy()) 185 return None; 186 187 // Make sure that the function is available. 188 LibFunc TLIFn; 189 if (!TLI || !TLI->getLibFunc(*Callee, TLIFn) || !TLI->has(TLIFn)) 190 return None; 191 192 const auto *Iter = find_if( 193 AllocationFnData, [TLIFn](const std::pair<LibFunc, AllocFnsTy> &P) { 194 return P.first == TLIFn; 195 }); 196 197 if (Iter == std::end(AllocationFnData)) 198 return None; 199 200 const AllocFnsTy *FnData = &Iter->second; 201 if ((FnData->AllocTy & AllocTy) != FnData->AllocTy) 202 return None; 203 204 // Check function prototype. 205 int FstParam = FnData->FstParam; 206 int SndParam = FnData->SndParam; 207 FunctionType *FTy = Callee->getFunctionType(); 208 209 if (FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) && 210 FTy->getNumParams() == FnData->NumParams && 211 (FstParam < 0 || 212 (FTy->getParamType(FstParam)->isIntegerTy(32) || 213 FTy->getParamType(FstParam)->isIntegerTy(64))) && 214 (SndParam < 0 || 215 FTy->getParamType(SndParam)->isIntegerTy(32) || 216 FTy->getParamType(SndParam)->isIntegerTy(64))) 217 return *FnData; 218 return None; 219 } 220 221 static Optional<AllocFnsTy> getAllocationData(const Value *V, AllocType AllocTy, 222 const TargetLibraryInfo *TLI) { 223 bool IsNoBuiltinCall; 224 if (const Function *Callee = getCalledFunction(V, IsNoBuiltinCall)) 225 if (!IsNoBuiltinCall) 226 return getAllocationDataForFunction(Callee, AllocTy, TLI); 227 return None; 228 } 229 230 static Optional<AllocFnsTy> 231 getAllocationData(const Value *V, AllocType AllocTy, 232 function_ref<const TargetLibraryInfo &(Function &)> GetTLI) { 233 bool IsNoBuiltinCall; 234 if (const Function *Callee = getCalledFunction(V, IsNoBuiltinCall)) 235 if (!IsNoBuiltinCall) 236 return getAllocationDataForFunction( 237 Callee, AllocTy, &GetTLI(const_cast<Function &>(*Callee))); 238 return None; 239 } 240 241 static Optional<AllocFnsTy> getAllocationSize(const Value *V, 242 const TargetLibraryInfo *TLI) { 243 bool IsNoBuiltinCall; 244 const Function *Callee = 245 getCalledFunction(V, IsNoBuiltinCall); 246 if (!Callee) 247 return None; 248 249 // Prefer to use existing information over allocsize. This will give us an 250 // accurate AllocTy. 251 if (!IsNoBuiltinCall) 252 if (Optional<AllocFnsTy> Data = 253 getAllocationDataForFunction(Callee, AnyAlloc, TLI)) 254 return Data; 255 256 Attribute Attr = Callee->getFnAttribute(Attribute::AllocSize); 257 if (Attr == Attribute()) 258 return None; 259 260 std::pair<unsigned, Optional<unsigned>> Args = Attr.getAllocSizeArgs(); 261 262 AllocFnsTy Result; 263 // Because allocsize only tells us how many bytes are allocated, we're not 264 // really allowed to assume anything, so we use MallocLike. 265 Result.AllocTy = MallocLike; 266 Result.NumParams = Callee->getNumOperands(); 267 Result.FstParam = Args.first; 268 Result.SndParam = Args.second.value_or(-1); 269 // Allocsize has no way to specify an alignment argument 270 Result.AlignParam = -1; 271 return Result; 272 } 273 274 /// Tests if a value is a call or invoke to a library function that 275 /// allocates or reallocates memory (either malloc, calloc, realloc, or strdup 276 /// like). 277 bool llvm::isAllocationFn(const Value *V, const TargetLibraryInfo *TLI) { 278 return getAllocationData(V, AnyAlloc, TLI).has_value(); 279 } 280 bool llvm::isAllocationFn( 281 const Value *V, function_ref<const TargetLibraryInfo &(Function &)> GetTLI) { 282 return getAllocationData(V, AnyAlloc, GetTLI).has_value(); 283 } 284 285 /// Tests if a value is a call or invoke to a library function that 286 /// allocates uninitialized memory (such as malloc). 287 static bool isMallocLikeFn(const Value *V, const TargetLibraryInfo *TLI) { 288 return getAllocationData(V, MallocOrOpNewLike, TLI).has_value(); 289 } 290 291 /// Tests if a value is a call or invoke to a library function that 292 /// allocates uninitialized memory with alignment (such as aligned_alloc). 293 static bool isAlignedAllocLikeFn(const Value *V, const TargetLibraryInfo *TLI) { 294 return getAllocationData(V, AlignedAllocLike, TLI).has_value(); 295 } 296 297 /// Tests if a value is a call or invoke to a library function that 298 /// allocates zero-filled memory (such as calloc). 299 static bool isCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI) { 300 return getAllocationData(V, CallocLike, TLI).has_value(); 301 } 302 303 /// Tests if a value is a call or invoke to a library function that 304 /// allocates memory similar to malloc or calloc. 305 bool llvm::isMallocOrCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI) { 306 return getAllocationData(V, MallocOrCallocLike, TLI).has_value(); 307 } 308 309 /// Tests if a value is a call or invoke to a library function that 310 /// allocates memory (either malloc, calloc, or strdup like). 311 bool llvm::isAllocLikeFn(const Value *V, const TargetLibraryInfo *TLI) { 312 return getAllocationData(V, AllocLike, TLI).has_value(); 313 } 314 315 /// Tests if a value is a call or invoke to a library function that 316 /// reallocates memory (e.g., realloc). 317 bool llvm::isReallocLikeFn(const Value *V, const TargetLibraryInfo *TLI) { 318 return getAllocationData(V, ReallocLike, TLI).has_value(); 319 } 320 321 /// Tests if a functions is a call or invoke to a library function that 322 /// reallocates memory (e.g., realloc). 323 bool llvm::isReallocLikeFn(const Function *F, const TargetLibraryInfo *TLI) { 324 return getAllocationDataForFunction(F, ReallocLike, TLI).has_value(); 325 } 326 327 bool llvm::isRemovableAlloc(const CallBase *CB, const TargetLibraryInfo *TLI) { 328 // Note: Removability is highly dependent on the source language. For 329 // example, recent C++ requires direct calls to the global allocation 330 // [basic.stc.dynamic.allocation] to be observable unless part of a new 331 // expression [expr.new paragraph 13]. 332 333 // Historically we've treated the C family allocation routines and operator 334 // new as removable 335 return isAllocLikeFn(CB, TLI); 336 } 337 338 Value *llvm::getAllocAlignment(const CallBase *V, 339 const TargetLibraryInfo *TLI) { 340 const Optional<AllocFnsTy> FnData = getAllocationData(V, AnyAlloc, TLI); 341 if (FnData && FnData->AlignParam >= 0) { 342 return V->getOperand(FnData->AlignParam); 343 } 344 return V->getArgOperandWithAttribute(Attribute::AllocAlign); 345 } 346 347 /// When we're compiling N-bit code, and the user uses parameters that are 348 /// greater than N bits (e.g. uint64_t on a 32-bit build), we can run into 349 /// trouble with APInt size issues. This function handles resizing + overflow 350 /// checks for us. Check and zext or trunc \p I depending on IntTyBits and 351 /// I's value. 352 static bool CheckedZextOrTrunc(APInt &I, unsigned IntTyBits) { 353 // More bits than we can handle. Checking the bit width isn't necessary, but 354 // it's faster than checking active bits, and should give `false` in the 355 // vast majority of cases. 356 if (I.getBitWidth() > IntTyBits && I.getActiveBits() > IntTyBits) 357 return false; 358 if (I.getBitWidth() != IntTyBits) 359 I = I.zextOrTrunc(IntTyBits); 360 return true; 361 } 362 363 Optional<APInt> 364 llvm::getAllocSize(const CallBase *CB, const TargetLibraryInfo *TLI, 365 function_ref<const Value *(const Value *)> Mapper) { 366 // Note: This handles both explicitly listed allocation functions and 367 // allocsize. The code structure could stand to be cleaned up a bit. 368 Optional<AllocFnsTy> FnData = getAllocationSize(CB, TLI); 369 if (!FnData) 370 return None; 371 372 // Get the index type for this address space, results and intermediate 373 // computations are performed at that width. 374 auto &DL = CB->getModule()->getDataLayout(); 375 const unsigned IntTyBits = DL.getIndexTypeSizeInBits(CB->getType()); 376 377 // Handle strdup-like functions separately. 378 if (FnData->AllocTy == StrDupLike) { 379 APInt Size(IntTyBits, GetStringLength(Mapper(CB->getArgOperand(0)))); 380 if (!Size) 381 return None; 382 383 // Strndup limits strlen. 384 if (FnData->FstParam > 0) { 385 const ConstantInt *Arg = 386 dyn_cast<ConstantInt>(Mapper(CB->getArgOperand(FnData->FstParam))); 387 if (!Arg) 388 return None; 389 390 APInt MaxSize = Arg->getValue().zext(IntTyBits); 391 if (Size.ugt(MaxSize)) 392 Size = MaxSize + 1; 393 } 394 return Size; 395 } 396 397 const ConstantInt *Arg = 398 dyn_cast<ConstantInt>(Mapper(CB->getArgOperand(FnData->FstParam))); 399 if (!Arg) 400 return None; 401 402 APInt Size = Arg->getValue(); 403 if (!CheckedZextOrTrunc(Size, IntTyBits)) 404 return None; 405 406 // Size is determined by just 1 parameter. 407 if (FnData->SndParam < 0) 408 return Size; 409 410 Arg = dyn_cast<ConstantInt>(Mapper(CB->getArgOperand(FnData->SndParam))); 411 if (!Arg) 412 return None; 413 414 APInt NumElems = Arg->getValue(); 415 if (!CheckedZextOrTrunc(NumElems, IntTyBits)) 416 return None; 417 418 bool Overflow; 419 Size = Size.umul_ov(NumElems, Overflow); 420 if (Overflow) 421 return None; 422 return Size; 423 } 424 425 Constant *llvm::getInitialValueOfAllocation(const Value *V, 426 const TargetLibraryInfo *TLI, 427 Type *Ty) { 428 auto *Alloc = dyn_cast<CallBase>(V); 429 if (!Alloc) 430 return nullptr; 431 432 // malloc and aligned_alloc are uninitialized (undef) 433 if (isMallocLikeFn(Alloc, TLI) || isAlignedAllocLikeFn(Alloc, TLI)) 434 return UndefValue::get(Ty); 435 436 // calloc zero initializes 437 if (isCallocLikeFn(Alloc, TLI)) 438 return Constant::getNullValue(Ty); 439 440 return nullptr; 441 } 442 443 struct FreeFnsTy { 444 unsigned NumParams; 445 // Name of default allocator function to group malloc/free calls by family 446 MallocFamily Family; 447 }; 448 449 // clang-format off 450 static const std::pair<LibFunc, FreeFnsTy> FreeFnData[] = { 451 {LibFunc_free, {1, MallocFamily::Malloc}}, 452 {LibFunc_vec_free, {1, MallocFamily::VecMalloc}}, 453 {LibFunc_ZdlPv, {1, MallocFamily::CPPNew}}, // operator delete(void*) 454 {LibFunc_ZdaPv, {1, MallocFamily::CPPNewArray}}, // operator delete[](void*) 455 {LibFunc_msvc_delete_ptr32, {1, MallocFamily::MSVCNew}}, // operator delete(void*) 456 {LibFunc_msvc_delete_ptr64, {1, MallocFamily::MSVCNew}}, // operator delete(void*) 457 {LibFunc_msvc_delete_array_ptr32, {1, MallocFamily::MSVCArrayNew}}, // operator delete[](void*) 458 {LibFunc_msvc_delete_array_ptr64, {1, MallocFamily::MSVCArrayNew}}, // operator delete[](void*) 459 {LibFunc_ZdlPvj, {2, MallocFamily::CPPNew}}, // delete(void*, uint) 460 {LibFunc_ZdlPvm, {2, MallocFamily::CPPNew}}, // delete(void*, ulong) 461 {LibFunc_ZdlPvRKSt9nothrow_t, {2, MallocFamily::CPPNew}}, // delete(void*, nothrow) 462 {LibFunc_ZdlPvSt11align_val_t, {2, MallocFamily::CPPNewAligned}}, // delete(void*, align_val_t) 463 {LibFunc_ZdaPvj, {2, MallocFamily::CPPNewArray}}, // delete[](void*, uint) 464 {LibFunc_ZdaPvm, {2, MallocFamily::CPPNewArray}}, // delete[](void*, ulong) 465 {LibFunc_ZdaPvRKSt9nothrow_t, {2, MallocFamily::CPPNewArray}}, // delete[](void*, nothrow) 466 {LibFunc_ZdaPvSt11align_val_t, {2, MallocFamily::CPPNewArrayAligned}}, // delete[](void*, align_val_t) 467 {LibFunc_msvc_delete_ptr32_int, {2, MallocFamily::MSVCNew}}, // delete(void*, uint) 468 {LibFunc_msvc_delete_ptr64_longlong, {2, MallocFamily::MSVCNew}}, // delete(void*, ulonglong) 469 {LibFunc_msvc_delete_ptr32_nothrow, {2, MallocFamily::MSVCNew}}, // delete(void*, nothrow) 470 {LibFunc_msvc_delete_ptr64_nothrow, {2, MallocFamily::MSVCNew}}, // delete(void*, nothrow) 471 {LibFunc_msvc_delete_array_ptr32_int, {2, MallocFamily::MSVCArrayNew}}, // delete[](void*, uint) 472 {LibFunc_msvc_delete_array_ptr64_longlong, {2, MallocFamily::MSVCArrayNew}}, // delete[](void*, ulonglong) 473 {LibFunc_msvc_delete_array_ptr32_nothrow, {2, MallocFamily::MSVCArrayNew}}, // delete[](void*, nothrow) 474 {LibFunc_msvc_delete_array_ptr64_nothrow, {2, MallocFamily::MSVCArrayNew}}, // delete[](void*, nothrow) 475 {LibFunc___kmpc_free_shared, {2, MallocFamily::KmpcAllocShared}}, // OpenMP Offloading RTL free 476 {LibFunc_ZdlPvSt11align_val_tRKSt9nothrow_t, {3, MallocFamily::CPPNewAligned}}, // delete(void*, align_val_t, nothrow) 477 {LibFunc_ZdaPvSt11align_val_tRKSt9nothrow_t, {3, MallocFamily::CPPNewArrayAligned}}, // delete[](void*, align_val_t, nothrow) 478 {LibFunc_ZdlPvjSt11align_val_t, {3, MallocFamily::CPPNewAligned}}, // delete(void*, unsigned int, align_val_t) 479 {LibFunc_ZdlPvmSt11align_val_t, {3, MallocFamily::CPPNewAligned}}, // delete(void*, unsigned long, align_val_t) 480 {LibFunc_ZdaPvjSt11align_val_t, {3, MallocFamily::CPPNewArrayAligned}}, // delete[](void*, unsigned int, align_val_t) 481 {LibFunc_ZdaPvmSt11align_val_t, {3, MallocFamily::CPPNewArrayAligned}}, // delete[](void*, unsigned long, align_val_t) 482 }; 483 // clang-format on 484 485 Optional<FreeFnsTy> getFreeFunctionDataForFunction(const Function *Callee, 486 const LibFunc TLIFn) { 487 const auto *Iter = 488 find_if(FreeFnData, [TLIFn](const std::pair<LibFunc, FreeFnsTy> &P) { 489 return P.first == TLIFn; 490 }); 491 if (Iter == std::end(FreeFnData)) 492 return None; 493 return Iter->second; 494 } 495 496 Optional<StringRef> llvm::getAllocationFamily(const Value *I, 497 const TargetLibraryInfo *TLI) { 498 bool IsNoBuiltin; 499 const Function *Callee = getCalledFunction(I, IsNoBuiltin); 500 if (Callee == nullptr || IsNoBuiltin) 501 return None; 502 LibFunc TLIFn; 503 if (!TLI || !TLI->getLibFunc(*Callee, TLIFn) || !TLI->has(TLIFn)) 504 return None; 505 const auto AllocData = getAllocationDataForFunction(Callee, AnyAlloc, TLI); 506 if (AllocData) 507 return mangledNameForMallocFamily(AllocData.value().Family); 508 const auto FreeData = getFreeFunctionDataForFunction(Callee, TLIFn); 509 if (FreeData) 510 return mangledNameForMallocFamily(FreeData.value().Family); 511 return None; 512 } 513 514 /// isLibFreeFunction - Returns true if the function is a builtin free() 515 bool llvm::isLibFreeFunction(const Function *F, const LibFunc TLIFn) { 516 Optional<FreeFnsTy> FnData = getFreeFunctionDataForFunction(F, TLIFn); 517 if (!FnData) 518 return false; 519 520 // Check free prototype. 521 // FIXME: workaround for PR5130, this will be obsolete when a nobuiltin 522 // attribute will exist. 523 FunctionType *FTy = F->getFunctionType(); 524 if (!FTy->getReturnType()->isVoidTy()) 525 return false; 526 if (FTy->getNumParams() != FnData->NumParams) 527 return false; 528 if (FTy->getParamType(0) != Type::getInt8PtrTy(F->getContext())) 529 return false; 530 531 return true; 532 } 533 534 bool llvm::isFreeCall(const Value *I, const TargetLibraryInfo *TLI) { 535 bool IsNoBuiltinCall; 536 const Function *Callee = getCalledFunction(I, IsNoBuiltinCall); 537 if (Callee == nullptr || IsNoBuiltinCall) 538 return false; 539 540 LibFunc TLIFn; 541 if (!TLI || !TLI->getLibFunc(*Callee, TLIFn) || !TLI->has(TLIFn)) 542 return false; 543 544 return isLibFreeFunction(Callee, TLIFn); 545 } 546 547 Value *llvm::getFreedOperand(const CallBase *CB, const TargetLibraryInfo *TLI) { 548 // All currently supported free functions free the first argument. 549 if (isFreeCall(CB, TLI)) 550 return CB->getArgOperand(0); 551 return nullptr; 552 } 553 554 //===----------------------------------------------------------------------===// 555 // Utility functions to compute size of objects. 556 // 557 static APInt getSizeWithOverflow(const SizeOffsetType &Data) { 558 if (Data.second.isNegative() || Data.first.ult(Data.second)) 559 return APInt(Data.first.getBitWidth(), 0); 560 return Data.first - Data.second; 561 } 562 563 /// Compute the size of the object pointed by Ptr. Returns true and the 564 /// object size in Size if successful, and false otherwise. 565 /// If RoundToAlign is true, then Size is rounded up to the alignment of 566 /// allocas, byval arguments, and global variables. 567 bool llvm::getObjectSize(const Value *Ptr, uint64_t &Size, const DataLayout &DL, 568 const TargetLibraryInfo *TLI, ObjectSizeOpts Opts) { 569 ObjectSizeOffsetVisitor Visitor(DL, TLI, Ptr->getContext(), Opts); 570 SizeOffsetType Data = Visitor.compute(const_cast<Value*>(Ptr)); 571 if (!Visitor.bothKnown(Data)) 572 return false; 573 574 Size = getSizeWithOverflow(Data).getZExtValue(); 575 return true; 576 } 577 578 Value *llvm::lowerObjectSizeCall(IntrinsicInst *ObjectSize, 579 const DataLayout &DL, 580 const TargetLibraryInfo *TLI, 581 bool MustSucceed) { 582 return lowerObjectSizeCall(ObjectSize, DL, TLI, /*AAResults=*/nullptr, 583 MustSucceed); 584 } 585 586 Value *llvm::lowerObjectSizeCall(IntrinsicInst *ObjectSize, 587 const DataLayout &DL, 588 const TargetLibraryInfo *TLI, AAResults *AA, 589 bool MustSucceed) { 590 assert(ObjectSize->getIntrinsicID() == Intrinsic::objectsize && 591 "ObjectSize must be a call to llvm.objectsize!"); 592 593 bool MaxVal = cast<ConstantInt>(ObjectSize->getArgOperand(1))->isZero(); 594 ObjectSizeOpts EvalOptions; 595 EvalOptions.AA = AA; 596 597 // Unless we have to fold this to something, try to be as accurate as 598 // possible. 599 if (MustSucceed) 600 EvalOptions.EvalMode = 601 MaxVal ? ObjectSizeOpts::Mode::Max : ObjectSizeOpts::Mode::Min; 602 else 603 EvalOptions.EvalMode = ObjectSizeOpts::Mode::Exact; 604 605 EvalOptions.NullIsUnknownSize = 606 cast<ConstantInt>(ObjectSize->getArgOperand(2))->isOne(); 607 608 auto *ResultType = cast<IntegerType>(ObjectSize->getType()); 609 bool StaticOnly = cast<ConstantInt>(ObjectSize->getArgOperand(3))->isZero(); 610 if (StaticOnly) { 611 // FIXME: Does it make sense to just return a failure value if the size won't 612 // fit in the output and `!MustSucceed`? 613 uint64_t Size; 614 if (getObjectSize(ObjectSize->getArgOperand(0), Size, DL, TLI, EvalOptions) && 615 isUIntN(ResultType->getBitWidth(), Size)) 616 return ConstantInt::get(ResultType, Size); 617 } else { 618 LLVMContext &Ctx = ObjectSize->getFunction()->getContext(); 619 ObjectSizeOffsetEvaluator Eval(DL, TLI, Ctx, EvalOptions); 620 SizeOffsetEvalType SizeOffsetPair = 621 Eval.compute(ObjectSize->getArgOperand(0)); 622 623 if (SizeOffsetPair != ObjectSizeOffsetEvaluator::unknown()) { 624 IRBuilder<TargetFolder> Builder(Ctx, TargetFolder(DL)); 625 Builder.SetInsertPoint(ObjectSize); 626 627 // If we've outside the end of the object, then we can always access 628 // exactly 0 bytes. 629 Value *ResultSize = 630 Builder.CreateSub(SizeOffsetPair.first, SizeOffsetPair.second); 631 Value *UseZero = 632 Builder.CreateICmpULT(SizeOffsetPair.first, SizeOffsetPair.second); 633 ResultSize = Builder.CreateZExtOrTrunc(ResultSize, ResultType); 634 Value *Ret = Builder.CreateSelect( 635 UseZero, ConstantInt::get(ResultType, 0), ResultSize); 636 637 // The non-constant size expression cannot evaluate to -1. 638 if (!isa<Constant>(SizeOffsetPair.first) || 639 !isa<Constant>(SizeOffsetPair.second)) 640 Builder.CreateAssumption( 641 Builder.CreateICmpNE(Ret, ConstantInt::get(ResultType, -1))); 642 643 return Ret; 644 } 645 } 646 647 if (!MustSucceed) 648 return nullptr; 649 650 return ConstantInt::get(ResultType, MaxVal ? -1ULL : 0); 651 } 652 653 STATISTIC(ObjectVisitorArgument, 654 "Number of arguments with unsolved size and offset"); 655 STATISTIC(ObjectVisitorLoad, 656 "Number of load instructions with unsolved size and offset"); 657 658 APInt ObjectSizeOffsetVisitor::align(APInt Size, MaybeAlign Alignment) { 659 if (Options.RoundToAlign && Alignment) 660 return APInt(IntTyBits, alignTo(Size.getZExtValue(), *Alignment)); 661 return Size; 662 } 663 664 ObjectSizeOffsetVisitor::ObjectSizeOffsetVisitor(const DataLayout &DL, 665 const TargetLibraryInfo *TLI, 666 LLVMContext &Context, 667 ObjectSizeOpts Options) 668 : DL(DL), TLI(TLI), Options(Options) { 669 // Pointer size must be rechecked for each object visited since it could have 670 // a different address space. 671 } 672 673 SizeOffsetType ObjectSizeOffsetVisitor::compute(Value *V) { 674 unsigned InitialIntTyBits = DL.getIndexTypeSizeInBits(V->getType()); 675 676 // Stripping pointer casts can strip address space casts which can change the 677 // index type size. The invariant is that we use the value type to determine 678 // the index type size and if we stripped address space casts we have to 679 // readjust the APInt as we pass it upwards in order for the APInt to match 680 // the type the caller passed in. 681 APInt Offset(InitialIntTyBits, 0); 682 V = V->stripAndAccumulateConstantOffsets( 683 DL, Offset, /* AllowNonInbounds */ true, /* AllowInvariantGroup */ true); 684 685 // Later we use the index type size and zero but it will match the type of the 686 // value that is passed to computeImpl. 687 IntTyBits = DL.getIndexTypeSizeInBits(V->getType()); 688 Zero = APInt::getZero(IntTyBits); 689 690 bool IndexTypeSizeChanged = InitialIntTyBits != IntTyBits; 691 if (!IndexTypeSizeChanged && Offset.isZero()) 692 return computeImpl(V); 693 694 // We stripped an address space cast that changed the index type size or we 695 // accumulated some constant offset (or both). Readjust the bit width to match 696 // the argument index type size and apply the offset, as required. 697 SizeOffsetType SOT = computeImpl(V); 698 if (IndexTypeSizeChanged) { 699 if (knownSize(SOT) && !::CheckedZextOrTrunc(SOT.first, InitialIntTyBits)) 700 SOT.first = APInt(); 701 if (knownOffset(SOT) && !::CheckedZextOrTrunc(SOT.second, InitialIntTyBits)) 702 SOT.second = APInt(); 703 } 704 // If the computed offset is "unknown" we cannot add the stripped offset. 705 return {SOT.first, 706 SOT.second.getBitWidth() > 1 ? SOT.second + Offset : SOT.second}; 707 } 708 709 SizeOffsetType ObjectSizeOffsetVisitor::computeImpl(Value *V) { 710 if (Instruction *I = dyn_cast<Instruction>(V)) { 711 // If we have already seen this instruction, bail out. Cycles can happen in 712 // unreachable code after constant propagation. 713 if (!SeenInsts.insert(I).second) 714 return unknown(); 715 716 return visit(*I); 717 } 718 if (Argument *A = dyn_cast<Argument>(V)) 719 return visitArgument(*A); 720 if (ConstantPointerNull *P = dyn_cast<ConstantPointerNull>(V)) 721 return visitConstantPointerNull(*P); 722 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) 723 return visitGlobalAlias(*GA); 724 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 725 return visitGlobalVariable(*GV); 726 if (UndefValue *UV = dyn_cast<UndefValue>(V)) 727 return visitUndefValue(*UV); 728 729 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetVisitor::compute() unhandled value: " 730 << *V << '\n'); 731 return unknown(); 732 } 733 734 bool ObjectSizeOffsetVisitor::CheckedZextOrTrunc(APInt &I) { 735 return ::CheckedZextOrTrunc(I, IntTyBits); 736 } 737 738 SizeOffsetType ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) { 739 if (!I.getAllocatedType()->isSized()) 740 return unknown(); 741 742 TypeSize ElemSize = DL.getTypeAllocSize(I.getAllocatedType()); 743 if (ElemSize.isScalable() && Options.EvalMode != ObjectSizeOpts::Mode::Min) 744 return unknown(); 745 APInt Size(IntTyBits, ElemSize.getKnownMinSize()); 746 if (!I.isArrayAllocation()) 747 return std::make_pair(align(Size, I.getAlign()), Zero); 748 749 Value *ArraySize = I.getArraySize(); 750 if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) { 751 APInt NumElems = C->getValue(); 752 if (!CheckedZextOrTrunc(NumElems)) 753 return unknown(); 754 755 bool Overflow; 756 Size = Size.umul_ov(NumElems, Overflow); 757 return Overflow ? unknown() 758 : std::make_pair(align(Size, I.getAlign()), Zero); 759 } 760 return unknown(); 761 } 762 763 SizeOffsetType ObjectSizeOffsetVisitor::visitArgument(Argument &A) { 764 Type *MemoryTy = A.getPointeeInMemoryValueType(); 765 // No interprocedural analysis is done at the moment. 766 if (!MemoryTy|| !MemoryTy->isSized()) { 767 ++ObjectVisitorArgument; 768 return unknown(); 769 } 770 771 APInt Size(IntTyBits, DL.getTypeAllocSize(MemoryTy)); 772 return std::make_pair(align(Size, A.getParamAlign()), Zero); 773 } 774 775 SizeOffsetType ObjectSizeOffsetVisitor::visitCallBase(CallBase &CB) { 776 if (Optional<APInt> Size = getAllocSize(&CB, TLI)) 777 return std::make_pair(*Size, Zero); 778 return unknown(); 779 } 780 781 SizeOffsetType 782 ObjectSizeOffsetVisitor::visitConstantPointerNull(ConstantPointerNull& CPN) { 783 // If null is unknown, there's nothing we can do. Additionally, non-zero 784 // address spaces can make use of null, so we don't presume to know anything 785 // about that. 786 // 787 // TODO: How should this work with address space casts? We currently just drop 788 // them on the floor, but it's unclear what we should do when a NULL from 789 // addrspace(1) gets casted to addrspace(0) (or vice-versa). 790 if (Options.NullIsUnknownSize || CPN.getType()->getAddressSpace()) 791 return unknown(); 792 return std::make_pair(Zero, Zero); 793 } 794 795 SizeOffsetType 796 ObjectSizeOffsetVisitor::visitExtractElementInst(ExtractElementInst&) { 797 return unknown(); 798 } 799 800 SizeOffsetType 801 ObjectSizeOffsetVisitor::visitExtractValueInst(ExtractValueInst&) { 802 // Easy cases were already folded by previous passes. 803 return unknown(); 804 } 805 806 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalAlias(GlobalAlias &GA) { 807 if (GA.isInterposable()) 808 return unknown(); 809 return compute(GA.getAliasee()); 810 } 811 812 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalVariable(GlobalVariable &GV){ 813 if (!GV.hasDefinitiveInitializer()) 814 return unknown(); 815 816 APInt Size(IntTyBits, DL.getTypeAllocSize(GV.getValueType())); 817 return std::make_pair(align(Size, GV.getAlign()), Zero); 818 } 819 820 SizeOffsetType ObjectSizeOffsetVisitor::visitIntToPtrInst(IntToPtrInst&) { 821 // clueless 822 return unknown(); 823 } 824 825 SizeOffsetType ObjectSizeOffsetVisitor::findLoadSizeOffset( 826 LoadInst &Load, BasicBlock &BB, BasicBlock::iterator From, 827 SmallDenseMap<BasicBlock *, SizeOffsetType, 8> &VisitedBlocks, 828 unsigned &ScannedInstCount) { 829 constexpr unsigned MaxInstsToScan = 128; 830 831 auto Where = VisitedBlocks.find(&BB); 832 if (Where != VisitedBlocks.end()) 833 return Where->second; 834 835 auto Unknown = [this, &BB, &VisitedBlocks]() { 836 return VisitedBlocks[&BB] = unknown(); 837 }; 838 auto Known = [&BB, &VisitedBlocks](SizeOffsetType SO) { 839 return VisitedBlocks[&BB] = SO; 840 }; 841 842 do { 843 Instruction &I = *From; 844 845 if (I.isDebugOrPseudoInst()) 846 continue; 847 848 if (++ScannedInstCount > MaxInstsToScan) 849 return Unknown(); 850 851 if (!I.mayWriteToMemory()) 852 continue; 853 854 if (auto *SI = dyn_cast<StoreInst>(&I)) { 855 AliasResult AR = 856 Options.AA->alias(SI->getPointerOperand(), Load.getPointerOperand()); 857 switch ((AliasResult::Kind)AR) { 858 case AliasResult::NoAlias: 859 continue; 860 case AliasResult::MustAlias: 861 if (SI->getValueOperand()->getType()->isPointerTy()) 862 return Known(compute(SI->getValueOperand())); 863 else 864 return Unknown(); // No handling of non-pointer values by `compute`. 865 default: 866 return Unknown(); 867 } 868 } 869 870 if (auto *CB = dyn_cast<CallBase>(&I)) { 871 Function *Callee = CB->getCalledFunction(); 872 // Bail out on indirect call. 873 if (!Callee) 874 return Unknown(); 875 876 LibFunc TLIFn; 877 if (!TLI || !TLI->getLibFunc(*CB->getCalledFunction(), TLIFn) || 878 !TLI->has(TLIFn)) 879 return Unknown(); 880 881 // TODO: There's probably more interesting case to support here. 882 if (TLIFn != LibFunc_posix_memalign) 883 return Unknown(); 884 885 AliasResult AR = 886 Options.AA->alias(CB->getOperand(0), Load.getPointerOperand()); 887 switch ((AliasResult::Kind)AR) { 888 case AliasResult::NoAlias: 889 continue; 890 case AliasResult::MustAlias: 891 break; 892 default: 893 return Unknown(); 894 } 895 896 // Is the error status of posix_memalign correctly checked? If not it 897 // would be incorrect to assume it succeeds and load doesn't see the 898 // previous value. 899 Optional<bool> Checked = isImpliedByDomCondition( 900 ICmpInst::ICMP_EQ, CB, ConstantInt::get(CB->getType(), 0), &Load, DL); 901 if (!Checked || !*Checked) 902 return Unknown(); 903 904 Value *Size = CB->getOperand(2); 905 auto *C = dyn_cast<ConstantInt>(Size); 906 if (!C) 907 return Unknown(); 908 909 return Known({C->getValue(), APInt(C->getValue().getBitWidth(), 0)}); 910 } 911 912 return Unknown(); 913 } while (From-- != BB.begin()); 914 915 SmallVector<SizeOffsetType> PredecessorSizeOffsets; 916 for (auto *PredBB : predecessors(&BB)) { 917 PredecessorSizeOffsets.push_back(findLoadSizeOffset( 918 Load, *PredBB, BasicBlock::iterator(PredBB->getTerminator()), 919 VisitedBlocks, ScannedInstCount)); 920 if (!bothKnown(PredecessorSizeOffsets.back())) 921 return Unknown(); 922 } 923 924 if (PredecessorSizeOffsets.empty()) 925 return Unknown(); 926 927 return Known(std::accumulate(PredecessorSizeOffsets.begin() + 1, 928 PredecessorSizeOffsets.end(), 929 PredecessorSizeOffsets.front(), 930 [this](SizeOffsetType LHS, SizeOffsetType RHS) { 931 return combineSizeOffset(LHS, RHS); 932 })); 933 } 934 935 SizeOffsetType ObjectSizeOffsetVisitor::visitLoadInst(LoadInst &LI) { 936 if (!Options.AA) { 937 ++ObjectVisitorLoad; 938 return unknown(); 939 } 940 941 SmallDenseMap<BasicBlock *, SizeOffsetType, 8> VisitedBlocks; 942 unsigned ScannedInstCount = 0; 943 SizeOffsetType SO = 944 findLoadSizeOffset(LI, *LI.getParent(), BasicBlock::iterator(LI), 945 VisitedBlocks, ScannedInstCount); 946 if (!bothKnown(SO)) 947 ++ObjectVisitorLoad; 948 return SO; 949 } 950 951 SizeOffsetType ObjectSizeOffsetVisitor::combineSizeOffset(SizeOffsetType LHS, 952 SizeOffsetType RHS) { 953 if (!bothKnown(LHS) || !bothKnown(RHS)) 954 return unknown(); 955 956 switch (Options.EvalMode) { 957 case ObjectSizeOpts::Mode::Min: 958 return (getSizeWithOverflow(LHS).slt(getSizeWithOverflow(RHS))) ? LHS : RHS; 959 case ObjectSizeOpts::Mode::Max: 960 return (getSizeWithOverflow(LHS).sgt(getSizeWithOverflow(RHS))) ? LHS : RHS; 961 case ObjectSizeOpts::Mode::Exact: 962 return (getSizeWithOverflow(LHS).eq(getSizeWithOverflow(RHS))) ? LHS 963 : unknown(); 964 } 965 llvm_unreachable("missing an eval mode"); 966 } 967 968 SizeOffsetType ObjectSizeOffsetVisitor::visitPHINode(PHINode &PN) { 969 auto IncomingValues = PN.incoming_values(); 970 return std::accumulate(IncomingValues.begin() + 1, IncomingValues.end(), 971 compute(*IncomingValues.begin()), 972 [this](SizeOffsetType LHS, Value *VRHS) { 973 return combineSizeOffset(LHS, compute(VRHS)); 974 }); 975 } 976 977 SizeOffsetType ObjectSizeOffsetVisitor::visitSelectInst(SelectInst &I) { 978 return combineSizeOffset(compute(I.getTrueValue()), 979 compute(I.getFalseValue())); 980 } 981 982 SizeOffsetType ObjectSizeOffsetVisitor::visitUndefValue(UndefValue&) { 983 return std::make_pair(Zero, Zero); 984 } 985 986 SizeOffsetType ObjectSizeOffsetVisitor::visitInstruction(Instruction &I) { 987 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetVisitor unknown instruction:" << I 988 << '\n'); 989 return unknown(); 990 } 991 992 ObjectSizeOffsetEvaluator::ObjectSizeOffsetEvaluator( 993 const DataLayout &DL, const TargetLibraryInfo *TLI, LLVMContext &Context, 994 ObjectSizeOpts EvalOpts) 995 : DL(DL), TLI(TLI), Context(Context), 996 Builder(Context, TargetFolder(DL), 997 IRBuilderCallbackInserter( 998 [&](Instruction *I) { InsertedInstructions.insert(I); })), 999 EvalOpts(EvalOpts) { 1000 // IntTy and Zero must be set for each compute() since the address space may 1001 // be different for later objects. 1002 } 1003 1004 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute(Value *V) { 1005 // XXX - Are vectors of pointers possible here? 1006 IntTy = cast<IntegerType>(DL.getIndexType(V->getType())); 1007 Zero = ConstantInt::get(IntTy, 0); 1008 1009 SizeOffsetEvalType Result = compute_(V); 1010 1011 if (!bothKnown(Result)) { 1012 // Erase everything that was computed in this iteration from the cache, so 1013 // that no dangling references are left behind. We could be a bit smarter if 1014 // we kept a dependency graph. It's probably not worth the complexity. 1015 for (const Value *SeenVal : SeenVals) { 1016 CacheMapTy::iterator CacheIt = CacheMap.find(SeenVal); 1017 // non-computable results can be safely cached 1018 if (CacheIt != CacheMap.end() && anyKnown(CacheIt->second)) 1019 CacheMap.erase(CacheIt); 1020 } 1021 1022 // Erase any instructions we inserted as part of the traversal. 1023 for (Instruction *I : InsertedInstructions) { 1024 I->replaceAllUsesWith(PoisonValue::get(I->getType())); 1025 I->eraseFromParent(); 1026 } 1027 } 1028 1029 SeenVals.clear(); 1030 InsertedInstructions.clear(); 1031 return Result; 1032 } 1033 1034 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute_(Value *V) { 1035 ObjectSizeOffsetVisitor Visitor(DL, TLI, Context, EvalOpts); 1036 SizeOffsetType Const = Visitor.compute(V); 1037 if (Visitor.bothKnown(Const)) 1038 return std::make_pair(ConstantInt::get(Context, Const.first), 1039 ConstantInt::get(Context, Const.second)); 1040 1041 V = V->stripPointerCasts(); 1042 1043 // Check cache. 1044 CacheMapTy::iterator CacheIt = CacheMap.find(V); 1045 if (CacheIt != CacheMap.end()) 1046 return CacheIt->second; 1047 1048 // Always generate code immediately before the instruction being 1049 // processed, so that the generated code dominates the same BBs. 1050 BuilderTy::InsertPointGuard Guard(Builder); 1051 if (Instruction *I = dyn_cast<Instruction>(V)) 1052 Builder.SetInsertPoint(I); 1053 1054 // Now compute the size and offset. 1055 SizeOffsetEvalType Result; 1056 1057 // Record the pointers that were handled in this run, so that they can be 1058 // cleaned later if something fails. We also use this set to break cycles that 1059 // can occur in dead code. 1060 if (!SeenVals.insert(V).second) { 1061 Result = unknown(); 1062 } else if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 1063 Result = visitGEPOperator(*GEP); 1064 } else if (Instruction *I = dyn_cast<Instruction>(V)) { 1065 Result = visit(*I); 1066 } else if (isa<Argument>(V) || 1067 (isa<ConstantExpr>(V) && 1068 cast<ConstantExpr>(V)->getOpcode() == Instruction::IntToPtr) || 1069 isa<GlobalAlias>(V) || 1070 isa<GlobalVariable>(V)) { 1071 // Ignore values where we cannot do more than ObjectSizeVisitor. 1072 Result = unknown(); 1073 } else { 1074 LLVM_DEBUG( 1075 dbgs() << "ObjectSizeOffsetEvaluator::compute() unhandled value: " << *V 1076 << '\n'); 1077 Result = unknown(); 1078 } 1079 1080 // Don't reuse CacheIt since it may be invalid at this point. 1081 CacheMap[V] = Result; 1082 return Result; 1083 } 1084 1085 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitAllocaInst(AllocaInst &I) { 1086 if (!I.getAllocatedType()->isSized()) 1087 return unknown(); 1088 1089 // must be a VLA 1090 assert(I.isArrayAllocation()); 1091 1092 // If needed, adjust the alloca's operand size to match the pointer size. 1093 // Subsequent math operations expect the types to match. 1094 Value *ArraySize = Builder.CreateZExtOrTrunc( 1095 I.getArraySize(), DL.getIntPtrType(I.getContext())); 1096 assert(ArraySize->getType() == Zero->getType() && 1097 "Expected zero constant to have pointer type"); 1098 1099 Value *Size = ConstantInt::get(ArraySize->getType(), 1100 DL.getTypeAllocSize(I.getAllocatedType())); 1101 Size = Builder.CreateMul(Size, ArraySize); 1102 return std::make_pair(Size, Zero); 1103 } 1104 1105 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitCallBase(CallBase &CB) { 1106 Optional<AllocFnsTy> FnData = getAllocationSize(&CB, TLI); 1107 if (!FnData) 1108 return unknown(); 1109 1110 // Handle strdup-like functions separately. 1111 if (FnData->AllocTy == StrDupLike) { 1112 // TODO: implement evaluation of strdup/strndup 1113 return unknown(); 1114 } 1115 1116 Value *FirstArg = CB.getArgOperand(FnData->FstParam); 1117 FirstArg = Builder.CreateZExtOrTrunc(FirstArg, IntTy); 1118 if (FnData->SndParam < 0) 1119 return std::make_pair(FirstArg, Zero); 1120 1121 Value *SecondArg = CB.getArgOperand(FnData->SndParam); 1122 SecondArg = Builder.CreateZExtOrTrunc(SecondArg, IntTy); 1123 Value *Size = Builder.CreateMul(FirstArg, SecondArg); 1124 return std::make_pair(Size, Zero); 1125 } 1126 1127 SizeOffsetEvalType 1128 ObjectSizeOffsetEvaluator::visitExtractElementInst(ExtractElementInst&) { 1129 return unknown(); 1130 } 1131 1132 SizeOffsetEvalType 1133 ObjectSizeOffsetEvaluator::visitExtractValueInst(ExtractValueInst&) { 1134 return unknown(); 1135 } 1136 1137 SizeOffsetEvalType 1138 ObjectSizeOffsetEvaluator::visitGEPOperator(GEPOperator &GEP) { 1139 SizeOffsetEvalType PtrData = compute_(GEP.getPointerOperand()); 1140 if (!bothKnown(PtrData)) 1141 return unknown(); 1142 1143 Value *Offset = EmitGEPOffset(&Builder, DL, &GEP, /*NoAssumptions=*/true); 1144 Offset = Builder.CreateAdd(PtrData.second, Offset); 1145 return std::make_pair(PtrData.first, Offset); 1146 } 1147 1148 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitIntToPtrInst(IntToPtrInst&) { 1149 // clueless 1150 return unknown(); 1151 } 1152 1153 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitLoadInst(LoadInst &LI) { 1154 return unknown(); 1155 } 1156 1157 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitPHINode(PHINode &PHI) { 1158 // Create 2 PHIs: one for size and another for offset. 1159 PHINode *SizePHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 1160 PHINode *OffsetPHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 1161 1162 // Insert right away in the cache to handle recursive PHIs. 1163 CacheMap[&PHI] = std::make_pair(SizePHI, OffsetPHI); 1164 1165 // Compute offset/size for each PHI incoming pointer. 1166 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i) { 1167 Builder.SetInsertPoint(&*PHI.getIncomingBlock(i)->getFirstInsertionPt()); 1168 SizeOffsetEvalType EdgeData = compute_(PHI.getIncomingValue(i)); 1169 1170 if (!bothKnown(EdgeData)) { 1171 OffsetPHI->replaceAllUsesWith(PoisonValue::get(IntTy)); 1172 OffsetPHI->eraseFromParent(); 1173 InsertedInstructions.erase(OffsetPHI); 1174 SizePHI->replaceAllUsesWith(PoisonValue::get(IntTy)); 1175 SizePHI->eraseFromParent(); 1176 InsertedInstructions.erase(SizePHI); 1177 return unknown(); 1178 } 1179 SizePHI->addIncoming(EdgeData.first, PHI.getIncomingBlock(i)); 1180 OffsetPHI->addIncoming(EdgeData.second, PHI.getIncomingBlock(i)); 1181 } 1182 1183 Value *Size = SizePHI, *Offset = OffsetPHI; 1184 if (Value *Tmp = SizePHI->hasConstantValue()) { 1185 Size = Tmp; 1186 SizePHI->replaceAllUsesWith(Size); 1187 SizePHI->eraseFromParent(); 1188 InsertedInstructions.erase(SizePHI); 1189 } 1190 if (Value *Tmp = OffsetPHI->hasConstantValue()) { 1191 Offset = Tmp; 1192 OffsetPHI->replaceAllUsesWith(Offset); 1193 OffsetPHI->eraseFromParent(); 1194 InsertedInstructions.erase(OffsetPHI); 1195 } 1196 return std::make_pair(Size, Offset); 1197 } 1198 1199 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitSelectInst(SelectInst &I) { 1200 SizeOffsetEvalType TrueSide = compute_(I.getTrueValue()); 1201 SizeOffsetEvalType FalseSide = compute_(I.getFalseValue()); 1202 1203 if (!bothKnown(TrueSide) || !bothKnown(FalseSide)) 1204 return unknown(); 1205 if (TrueSide == FalseSide) 1206 return TrueSide; 1207 1208 Value *Size = Builder.CreateSelect(I.getCondition(), TrueSide.first, 1209 FalseSide.first); 1210 Value *Offset = Builder.CreateSelect(I.getCondition(), TrueSide.second, 1211 FalseSide.second); 1212 return std::make_pair(Size, Offset); 1213 } 1214 1215 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitInstruction(Instruction &I) { 1216 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetEvaluator unknown instruction:" << I 1217 << '\n'); 1218 return unknown(); 1219 } 1220