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