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.getValueOr(-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.hasValue() && 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().zextOrSelf(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 CallBase *Alloc, 424 const TargetLibraryInfo *TLI, 425 Type *Ty) { 426 assert(isAllocationFn(Alloc, TLI)); 427 428 // malloc and aligned_alloc are uninitialized (undef) 429 if (isMallocLikeFn(Alloc, TLI) || isAlignedAllocLikeFn(Alloc, TLI)) 430 return UndefValue::get(Ty); 431 432 // calloc zero initializes 433 if (isCallocLikeFn(Alloc, TLI)) 434 return Constant::getNullValue(Ty); 435 436 return nullptr; 437 } 438 439 struct FreeFnsTy { 440 unsigned NumParams; 441 // Name of default allocator function to group malloc/free calls by family 442 MallocFamily Family; 443 }; 444 445 // clang-format off 446 static const std::pair<LibFunc, FreeFnsTy> FreeFnData[] = { 447 {LibFunc_free, {1, MallocFamily::Malloc}}, 448 {LibFunc_ZdlPv, {1, MallocFamily::CPPNew}}, // operator delete(void*) 449 {LibFunc_ZdaPv, {1, MallocFamily::CPPNewArray}}, // operator delete[](void*) 450 {LibFunc_msvc_delete_ptr32, {1, MallocFamily::MSVCNew}}, // operator delete(void*) 451 {LibFunc_msvc_delete_ptr64, {1, MallocFamily::MSVCNew}}, // operator delete(void*) 452 {LibFunc_msvc_delete_array_ptr32, {1, MallocFamily::MSVCArrayNew}}, // operator delete[](void*) 453 {LibFunc_msvc_delete_array_ptr64, {1, MallocFamily::MSVCArrayNew}}, // operator delete[](void*) 454 {LibFunc_ZdlPvj, {2, MallocFamily::CPPNew}}, // delete(void*, uint) 455 {LibFunc_ZdlPvm, {2, MallocFamily::CPPNew}}, // delete(void*, ulong) 456 {LibFunc_ZdlPvRKSt9nothrow_t, {2, MallocFamily::CPPNew}}, // delete(void*, nothrow) 457 {LibFunc_ZdlPvSt11align_val_t, {2, MallocFamily::CPPNewAligned}}, // delete(void*, align_val_t) 458 {LibFunc_ZdaPvj, {2, MallocFamily::CPPNewArray}}, // delete[](void*, uint) 459 {LibFunc_ZdaPvm, {2, MallocFamily::CPPNewArray}}, // delete[](void*, ulong) 460 {LibFunc_ZdaPvRKSt9nothrow_t, {2, MallocFamily::CPPNewArray}}, // delete[](void*, nothrow) 461 {LibFunc_ZdaPvSt11align_val_t, {2, MallocFamily::CPPNewArrayAligned}}, // delete[](void*, align_val_t) 462 {LibFunc_msvc_delete_ptr32_int, {2, MallocFamily::MSVCNew}}, // delete(void*, uint) 463 {LibFunc_msvc_delete_ptr64_longlong, {2, MallocFamily::MSVCNew}}, // delete(void*, ulonglong) 464 {LibFunc_msvc_delete_ptr32_nothrow, {2, MallocFamily::MSVCNew}}, // delete(void*, nothrow) 465 {LibFunc_msvc_delete_ptr64_nothrow, {2, MallocFamily::MSVCNew}}, // delete(void*, nothrow) 466 {LibFunc_msvc_delete_array_ptr32_int, {2, MallocFamily::MSVCArrayNew}}, // delete[](void*, uint) 467 {LibFunc_msvc_delete_array_ptr64_longlong, {2, MallocFamily::MSVCArrayNew}}, // delete[](void*, ulonglong) 468 {LibFunc_msvc_delete_array_ptr32_nothrow, {2, MallocFamily::MSVCArrayNew}}, // delete[](void*, nothrow) 469 {LibFunc_msvc_delete_array_ptr64_nothrow, {2, MallocFamily::MSVCArrayNew}}, // delete[](void*, nothrow) 470 {LibFunc___kmpc_free_shared, {2, MallocFamily::KmpcAllocShared}}, // OpenMP Offloading RTL free 471 {LibFunc_ZdlPvSt11align_val_tRKSt9nothrow_t, {3, MallocFamily::CPPNewAligned}}, // delete(void*, align_val_t, nothrow) 472 {LibFunc_ZdaPvSt11align_val_tRKSt9nothrow_t, {3, MallocFamily::CPPNewArrayAligned}}, // delete[](void*, align_val_t, nothrow) 473 {LibFunc_ZdlPvjSt11align_val_t, {3, MallocFamily::CPPNewAligned}}, // delete(void*, unsigned int, align_val_t) 474 {LibFunc_ZdlPvmSt11align_val_t, {3, MallocFamily::CPPNewAligned}}, // delete(void*, unsigned long, align_val_t) 475 {LibFunc_ZdaPvjSt11align_val_t, {3, MallocFamily::CPPNewArrayAligned}}, // delete[](void*, unsigned int, align_val_t) 476 {LibFunc_ZdaPvmSt11align_val_t, {3, MallocFamily::CPPNewArrayAligned}}, // delete[](void*, unsigned long, align_val_t) 477 }; 478 // clang-format on 479 480 Optional<FreeFnsTy> getFreeFunctionDataForFunction(const Function *Callee, 481 const LibFunc TLIFn) { 482 const auto *Iter = 483 find_if(FreeFnData, [TLIFn](const std::pair<LibFunc, FreeFnsTy> &P) { 484 return P.first == TLIFn; 485 }); 486 if (Iter == std::end(FreeFnData)) 487 return None; 488 return Iter->second; 489 } 490 491 Optional<StringRef> llvm::getAllocationFamily(const Value *I, 492 const TargetLibraryInfo *TLI) { 493 bool IsNoBuiltin; 494 const Function *Callee = getCalledFunction(I, IsNoBuiltin); 495 if (Callee == nullptr || IsNoBuiltin) 496 return None; 497 LibFunc TLIFn; 498 if (!TLI || !TLI->getLibFunc(*Callee, TLIFn) || !TLI->has(TLIFn)) 499 return None; 500 const auto AllocData = getAllocationDataForFunction(Callee, AnyAlloc, TLI); 501 if (AllocData.hasValue()) 502 return mangledNameForMallocFamily(AllocData.getValue().Family); 503 const auto FreeData = getFreeFunctionDataForFunction(Callee, TLIFn); 504 if (FreeData.hasValue()) 505 return mangledNameForMallocFamily(FreeData.getValue().Family); 506 return None; 507 } 508 509 /// isLibFreeFunction - Returns true if the function is a builtin free() 510 bool llvm::isLibFreeFunction(const Function *F, const LibFunc TLIFn) { 511 Optional<FreeFnsTy> FnData = getFreeFunctionDataForFunction(F, TLIFn); 512 if (!FnData.hasValue()) 513 return false; 514 515 // Check free prototype. 516 // FIXME: workaround for PR5130, this will be obsolete when a nobuiltin 517 // attribute will exist. 518 FunctionType *FTy = F->getFunctionType(); 519 if (!FTy->getReturnType()->isVoidTy()) 520 return false; 521 if (FTy->getNumParams() != FnData->NumParams) 522 return false; 523 if (FTy->getParamType(0) != Type::getInt8PtrTy(F->getContext())) 524 return false; 525 526 return true; 527 } 528 529 /// isFreeCall - Returns non-null if the value is a call to the builtin free() 530 const CallInst *llvm::isFreeCall(const Value *I, const TargetLibraryInfo *TLI) { 531 bool IsNoBuiltinCall; 532 const Function *Callee = getCalledFunction(I, IsNoBuiltinCall); 533 if (Callee == nullptr || IsNoBuiltinCall) 534 return nullptr; 535 536 LibFunc TLIFn; 537 if (!TLI || !TLI->getLibFunc(*Callee, TLIFn) || !TLI->has(TLIFn)) 538 return nullptr; 539 540 return isLibFreeFunction(Callee, TLIFn) ? dyn_cast<CallInst>(I) : nullptr; 541 } 542 543 544 //===----------------------------------------------------------------------===// 545 // Utility functions to compute size of objects. 546 // 547 static APInt getSizeWithOverflow(const SizeOffsetType &Data) { 548 if (Data.second.isNegative() || Data.first.ult(Data.second)) 549 return APInt(Data.first.getBitWidth(), 0); 550 return Data.first - Data.second; 551 } 552 553 /// Compute the size of the object pointed by Ptr. Returns true and the 554 /// object size in Size if successful, and false otherwise. 555 /// If RoundToAlign is true, then Size is rounded up to the alignment of 556 /// allocas, byval arguments, and global variables. 557 bool llvm::getObjectSize(const Value *Ptr, uint64_t &Size, const DataLayout &DL, 558 const TargetLibraryInfo *TLI, ObjectSizeOpts Opts) { 559 ObjectSizeOffsetVisitor Visitor(DL, TLI, Ptr->getContext(), Opts); 560 SizeOffsetType Data = Visitor.compute(const_cast<Value*>(Ptr)); 561 if (!Visitor.bothKnown(Data)) 562 return false; 563 564 Size = getSizeWithOverflow(Data).getZExtValue(); 565 return true; 566 } 567 568 Value *llvm::lowerObjectSizeCall(IntrinsicInst *ObjectSize, 569 const DataLayout &DL, 570 const TargetLibraryInfo *TLI, 571 bool MustSucceed) { 572 return lowerObjectSizeCall(ObjectSize, DL, TLI, /*AAResults=*/nullptr, 573 MustSucceed); 574 } 575 576 Value *llvm::lowerObjectSizeCall(IntrinsicInst *ObjectSize, 577 const DataLayout &DL, 578 const TargetLibraryInfo *TLI, AAResults *AA, 579 bool MustSucceed) { 580 assert(ObjectSize->getIntrinsicID() == Intrinsic::objectsize && 581 "ObjectSize must be a call to llvm.objectsize!"); 582 583 bool MaxVal = cast<ConstantInt>(ObjectSize->getArgOperand(1))->isZero(); 584 ObjectSizeOpts EvalOptions; 585 EvalOptions.AA = AA; 586 587 // Unless we have to fold this to something, try to be as accurate as 588 // possible. 589 if (MustSucceed) 590 EvalOptions.EvalMode = 591 MaxVal ? ObjectSizeOpts::Mode::Max : ObjectSizeOpts::Mode::Min; 592 else 593 EvalOptions.EvalMode = ObjectSizeOpts::Mode::Exact; 594 595 EvalOptions.NullIsUnknownSize = 596 cast<ConstantInt>(ObjectSize->getArgOperand(2))->isOne(); 597 598 auto *ResultType = cast<IntegerType>(ObjectSize->getType()); 599 bool StaticOnly = cast<ConstantInt>(ObjectSize->getArgOperand(3))->isZero(); 600 if (StaticOnly) { 601 // FIXME: Does it make sense to just return a failure value if the size won't 602 // fit in the output and `!MustSucceed`? 603 uint64_t Size; 604 if (getObjectSize(ObjectSize->getArgOperand(0), Size, DL, TLI, EvalOptions) && 605 isUIntN(ResultType->getBitWidth(), Size)) 606 return ConstantInt::get(ResultType, Size); 607 } else { 608 LLVMContext &Ctx = ObjectSize->getFunction()->getContext(); 609 ObjectSizeOffsetEvaluator Eval(DL, TLI, Ctx, EvalOptions); 610 SizeOffsetEvalType SizeOffsetPair = 611 Eval.compute(ObjectSize->getArgOperand(0)); 612 613 if (SizeOffsetPair != ObjectSizeOffsetEvaluator::unknown()) { 614 IRBuilder<TargetFolder> Builder(Ctx, TargetFolder(DL)); 615 Builder.SetInsertPoint(ObjectSize); 616 617 // If we've outside the end of the object, then we can always access 618 // exactly 0 bytes. 619 Value *ResultSize = 620 Builder.CreateSub(SizeOffsetPair.first, SizeOffsetPair.second); 621 Value *UseZero = 622 Builder.CreateICmpULT(SizeOffsetPair.first, SizeOffsetPair.second); 623 ResultSize = Builder.CreateZExtOrTrunc(ResultSize, ResultType); 624 Value *Ret = Builder.CreateSelect( 625 UseZero, ConstantInt::get(ResultType, 0), ResultSize); 626 627 // The non-constant size expression cannot evaluate to -1. 628 if (!isa<Constant>(SizeOffsetPair.first) || 629 !isa<Constant>(SizeOffsetPair.second)) 630 Builder.CreateAssumption( 631 Builder.CreateICmpNE(Ret, ConstantInt::get(ResultType, -1))); 632 633 return Ret; 634 } 635 } 636 637 if (!MustSucceed) 638 return nullptr; 639 640 return ConstantInt::get(ResultType, MaxVal ? -1ULL : 0); 641 } 642 643 STATISTIC(ObjectVisitorArgument, 644 "Number of arguments with unsolved size and offset"); 645 STATISTIC(ObjectVisitorLoad, 646 "Number of load instructions with unsolved size and offset"); 647 648 APInt ObjectSizeOffsetVisitor::align(APInt Size, MaybeAlign Alignment) { 649 if (Options.RoundToAlign && Alignment) 650 return APInt(IntTyBits, alignTo(Size.getZExtValue(), Alignment)); 651 return Size; 652 } 653 654 ObjectSizeOffsetVisitor::ObjectSizeOffsetVisitor(const DataLayout &DL, 655 const TargetLibraryInfo *TLI, 656 LLVMContext &Context, 657 ObjectSizeOpts Options) 658 : DL(DL), TLI(TLI), Options(Options) { 659 // Pointer size must be rechecked for each object visited since it could have 660 // a different address space. 661 } 662 663 SizeOffsetType ObjectSizeOffsetVisitor::compute(Value *V) { 664 unsigned InitialIntTyBits = DL.getIndexTypeSizeInBits(V->getType()); 665 666 // Stripping pointer casts can strip address space casts which can change the 667 // index type size. The invariant is that we use the value type to determine 668 // the index type size and if we stripped address space casts we have to 669 // readjust the APInt as we pass it upwards in order for the APInt to match 670 // the type the caller passed in. 671 APInt Offset(InitialIntTyBits, 0); 672 V = V->stripAndAccumulateConstantOffsets( 673 DL, Offset, /* AllowNonInbounds */ true, /* AllowInvariantGroup */ true); 674 675 // Later we use the index type size and zero but it will match the type of the 676 // value that is passed to computeImpl. 677 IntTyBits = DL.getIndexTypeSizeInBits(V->getType()); 678 Zero = APInt::getZero(IntTyBits); 679 680 bool IndexTypeSizeChanged = InitialIntTyBits != IntTyBits; 681 if (!IndexTypeSizeChanged && Offset.isZero()) 682 return computeImpl(V); 683 684 // We stripped an address space cast that changed the index type size or we 685 // accumulated some constant offset (or both). Readjust the bit width to match 686 // the argument index type size and apply the offset, as required. 687 SizeOffsetType SOT = computeImpl(V); 688 if (IndexTypeSizeChanged) { 689 if (knownSize(SOT) && !::CheckedZextOrTrunc(SOT.first, InitialIntTyBits)) 690 SOT.first = APInt(); 691 if (knownOffset(SOT) && !::CheckedZextOrTrunc(SOT.second, InitialIntTyBits)) 692 SOT.second = APInt(); 693 } 694 // If the computed offset is "unknown" we cannot add the stripped offset. 695 return {SOT.first, 696 SOT.second.getBitWidth() > 1 ? SOT.second + Offset : SOT.second}; 697 } 698 699 SizeOffsetType ObjectSizeOffsetVisitor::computeImpl(Value *V) { 700 if (Instruction *I = dyn_cast<Instruction>(V)) { 701 // If we have already seen this instruction, bail out. Cycles can happen in 702 // unreachable code after constant propagation. 703 if (!SeenInsts.insert(I).second) 704 return unknown(); 705 706 return visit(*I); 707 } 708 if (Argument *A = dyn_cast<Argument>(V)) 709 return visitArgument(*A); 710 if (ConstantPointerNull *P = dyn_cast<ConstantPointerNull>(V)) 711 return visitConstantPointerNull(*P); 712 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) 713 return visitGlobalAlias(*GA); 714 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 715 return visitGlobalVariable(*GV); 716 if (UndefValue *UV = dyn_cast<UndefValue>(V)) 717 return visitUndefValue(*UV); 718 719 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetVisitor::compute() unhandled value: " 720 << *V << '\n'); 721 return unknown(); 722 } 723 724 bool ObjectSizeOffsetVisitor::CheckedZextOrTrunc(APInt &I) { 725 return ::CheckedZextOrTrunc(I, IntTyBits); 726 } 727 728 SizeOffsetType ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) { 729 if (!I.getAllocatedType()->isSized()) 730 return unknown(); 731 732 TypeSize ElemSize = DL.getTypeAllocSize(I.getAllocatedType()); 733 if (ElemSize.isScalable() && Options.EvalMode != ObjectSizeOpts::Mode::Min) 734 return unknown(); 735 APInt Size(IntTyBits, ElemSize.getKnownMinSize()); 736 if (!I.isArrayAllocation()) 737 return std::make_pair(align(Size, I.getAlign()), Zero); 738 739 Value *ArraySize = I.getArraySize(); 740 if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) { 741 APInt NumElems = C->getValue(); 742 if (!CheckedZextOrTrunc(NumElems)) 743 return unknown(); 744 745 bool Overflow; 746 Size = Size.umul_ov(NumElems, Overflow); 747 return Overflow ? unknown() 748 : std::make_pair(align(Size, I.getAlign()), Zero); 749 } 750 return unknown(); 751 } 752 753 SizeOffsetType ObjectSizeOffsetVisitor::visitArgument(Argument &A) { 754 Type *MemoryTy = A.getPointeeInMemoryValueType(); 755 // No interprocedural analysis is done at the moment. 756 if (!MemoryTy|| !MemoryTy->isSized()) { 757 ++ObjectVisitorArgument; 758 return unknown(); 759 } 760 761 APInt Size(IntTyBits, DL.getTypeAllocSize(MemoryTy)); 762 return std::make_pair(align(Size, A.getParamAlign()), Zero); 763 } 764 765 SizeOffsetType ObjectSizeOffsetVisitor::visitCallBase(CallBase &CB) { 766 auto Mapper = [](const Value *V) { return V; }; 767 if (Optional<APInt> Size = getAllocSize(&CB, TLI, Mapper)) 768 return std::make_pair(*Size, Zero); 769 return unknown(); 770 } 771 772 SizeOffsetType 773 ObjectSizeOffsetVisitor::visitConstantPointerNull(ConstantPointerNull& CPN) { 774 // If null is unknown, there's nothing we can do. Additionally, non-zero 775 // address spaces can make use of null, so we don't presume to know anything 776 // about that. 777 // 778 // TODO: How should this work with address space casts? We currently just drop 779 // them on the floor, but it's unclear what we should do when a NULL from 780 // addrspace(1) gets casted to addrspace(0) (or vice-versa). 781 if (Options.NullIsUnknownSize || CPN.getType()->getAddressSpace()) 782 return unknown(); 783 return std::make_pair(Zero, Zero); 784 } 785 786 SizeOffsetType 787 ObjectSizeOffsetVisitor::visitExtractElementInst(ExtractElementInst&) { 788 return unknown(); 789 } 790 791 SizeOffsetType 792 ObjectSizeOffsetVisitor::visitExtractValueInst(ExtractValueInst&) { 793 // Easy cases were already folded by previous passes. 794 return unknown(); 795 } 796 797 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalAlias(GlobalAlias &GA) { 798 if (GA.isInterposable()) 799 return unknown(); 800 return compute(GA.getAliasee()); 801 } 802 803 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalVariable(GlobalVariable &GV){ 804 if (!GV.hasDefinitiveInitializer()) 805 return unknown(); 806 807 APInt Size(IntTyBits, DL.getTypeAllocSize(GV.getValueType())); 808 return std::make_pair(align(Size, GV.getAlign()), Zero); 809 } 810 811 SizeOffsetType ObjectSizeOffsetVisitor::visitIntToPtrInst(IntToPtrInst&) { 812 // clueless 813 return unknown(); 814 } 815 816 SizeOffsetType ObjectSizeOffsetVisitor::findLoadSizeOffset( 817 LoadInst &Load, BasicBlock &BB, BasicBlock::iterator From, 818 SmallDenseMap<BasicBlock *, SizeOffsetType, 8> &VisitedBlocks, 819 unsigned &ScannedInstCount) { 820 constexpr unsigned MaxInstsToScan = 128; 821 822 auto Where = VisitedBlocks.find(&BB); 823 if (Where != VisitedBlocks.end()) 824 return Where->second; 825 826 auto Unknown = [this, &BB, &VisitedBlocks]() { 827 return VisitedBlocks[&BB] = unknown(); 828 }; 829 auto Known = [&BB, &VisitedBlocks](SizeOffsetType SO) { 830 return VisitedBlocks[&BB] = SO; 831 }; 832 833 do { 834 Instruction &I = *From; 835 836 if (I.isDebugOrPseudoInst()) 837 continue; 838 839 if (++ScannedInstCount > MaxInstsToScan) 840 return Unknown(); 841 842 if (!I.mayWriteToMemory()) 843 continue; 844 845 if (auto *SI = dyn_cast<StoreInst>(&I)) { 846 AliasResult AR = 847 Options.AA->alias(SI->getPointerOperand(), Load.getPointerOperand()); 848 switch ((AliasResult::Kind)AR) { 849 case AliasResult::NoAlias: 850 continue; 851 case AliasResult::MustAlias: 852 if (SI->getValueOperand()->getType()->isPointerTy()) 853 return Known(compute(SI->getValueOperand())); 854 else 855 return Unknown(); // No handling of non-pointer values by `compute`. 856 default: 857 return Unknown(); 858 } 859 } 860 861 if (auto *CB = dyn_cast<CallBase>(&I)) { 862 Function *Callee = CB->getCalledFunction(); 863 // Bail out on indirect call. 864 if (!Callee) 865 return Unknown(); 866 867 LibFunc TLIFn; 868 if (!TLI || !TLI->getLibFunc(*CB->getCalledFunction(), TLIFn) || 869 !TLI->has(TLIFn)) 870 return Unknown(); 871 872 // TODO: There's probably more interesting case to support here. 873 if (TLIFn != LibFunc_posix_memalign) 874 return Unknown(); 875 876 AliasResult AR = 877 Options.AA->alias(CB->getOperand(0), Load.getPointerOperand()); 878 switch ((AliasResult::Kind)AR) { 879 case AliasResult::NoAlias: 880 continue; 881 case AliasResult::MustAlias: 882 break; 883 default: 884 return Unknown(); 885 } 886 887 // Is the error status of posix_memalign correctly checked? If not it 888 // would be incorrect to assume it succeeds and load doesn't see the 889 // previous value. 890 Optional<bool> Checked = isImpliedByDomCondition( 891 ICmpInst::ICMP_EQ, CB, ConstantInt::get(CB->getType(), 0), &Load, DL); 892 if (!Checked || !*Checked) 893 return Unknown(); 894 895 Value *Size = CB->getOperand(2); 896 auto *C = dyn_cast<ConstantInt>(Size); 897 if (!C) 898 return Unknown(); 899 900 return Known({C->getValue(), APInt(C->getValue().getBitWidth(), 0)}); 901 } 902 903 return Unknown(); 904 } while (From-- != BB.begin()); 905 906 SmallVector<SizeOffsetType> PredecessorSizeOffsets; 907 for (auto *PredBB : predecessors(&BB)) { 908 PredecessorSizeOffsets.push_back(findLoadSizeOffset( 909 Load, *PredBB, BasicBlock::iterator(PredBB->getTerminator()), 910 VisitedBlocks, ScannedInstCount)); 911 if (!bothKnown(PredecessorSizeOffsets.back())) 912 return Unknown(); 913 } 914 915 if (PredecessorSizeOffsets.empty()) 916 return Unknown(); 917 918 return Known(std::accumulate(PredecessorSizeOffsets.begin() + 1, 919 PredecessorSizeOffsets.end(), 920 PredecessorSizeOffsets.front(), 921 [this](SizeOffsetType LHS, SizeOffsetType RHS) { 922 return combineSizeOffset(LHS, RHS); 923 })); 924 } 925 926 SizeOffsetType ObjectSizeOffsetVisitor::visitLoadInst(LoadInst &LI) { 927 if (!Options.AA) { 928 ++ObjectVisitorLoad; 929 return unknown(); 930 } 931 932 SmallDenseMap<BasicBlock *, SizeOffsetType, 8> VisitedBlocks; 933 unsigned ScannedInstCount = 0; 934 SizeOffsetType SO = 935 findLoadSizeOffset(LI, *LI.getParent(), BasicBlock::iterator(LI), 936 VisitedBlocks, ScannedInstCount); 937 if (!bothKnown(SO)) 938 ++ObjectVisitorLoad; 939 return SO; 940 } 941 942 SizeOffsetType ObjectSizeOffsetVisitor::combineSizeOffset(SizeOffsetType LHS, 943 SizeOffsetType RHS) { 944 if (!bothKnown(LHS) || !bothKnown(RHS)) 945 return unknown(); 946 947 switch (Options.EvalMode) { 948 case ObjectSizeOpts::Mode::Min: 949 return (getSizeWithOverflow(LHS).slt(getSizeWithOverflow(RHS))) ? LHS : RHS; 950 case ObjectSizeOpts::Mode::Max: 951 return (getSizeWithOverflow(LHS).sgt(getSizeWithOverflow(RHS))) ? LHS : RHS; 952 case ObjectSizeOpts::Mode::Exact: 953 return (getSizeWithOverflow(LHS).eq(getSizeWithOverflow(RHS))) ? LHS 954 : unknown(); 955 } 956 llvm_unreachable("missing an eval mode"); 957 } 958 959 SizeOffsetType ObjectSizeOffsetVisitor::visitPHINode(PHINode &PN) { 960 auto IncomingValues = PN.incoming_values(); 961 return std::accumulate(IncomingValues.begin() + 1, IncomingValues.end(), 962 compute(*IncomingValues.begin()), 963 [this](SizeOffsetType LHS, Value *VRHS) { 964 return combineSizeOffset(LHS, compute(VRHS)); 965 }); 966 } 967 968 SizeOffsetType ObjectSizeOffsetVisitor::visitSelectInst(SelectInst &I) { 969 return combineSizeOffset(compute(I.getTrueValue()), 970 compute(I.getFalseValue())); 971 } 972 973 SizeOffsetType ObjectSizeOffsetVisitor::visitUndefValue(UndefValue&) { 974 return std::make_pair(Zero, Zero); 975 } 976 977 SizeOffsetType ObjectSizeOffsetVisitor::visitInstruction(Instruction &I) { 978 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetVisitor unknown instruction:" << I 979 << '\n'); 980 return unknown(); 981 } 982 983 ObjectSizeOffsetEvaluator::ObjectSizeOffsetEvaluator( 984 const DataLayout &DL, const TargetLibraryInfo *TLI, LLVMContext &Context, 985 ObjectSizeOpts EvalOpts) 986 : DL(DL), TLI(TLI), Context(Context), 987 Builder(Context, TargetFolder(DL), 988 IRBuilderCallbackInserter( 989 [&](Instruction *I) { InsertedInstructions.insert(I); })), 990 EvalOpts(EvalOpts) { 991 // IntTy and Zero must be set for each compute() since the address space may 992 // be different for later objects. 993 } 994 995 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute(Value *V) { 996 // XXX - Are vectors of pointers possible here? 997 IntTy = cast<IntegerType>(DL.getIndexType(V->getType())); 998 Zero = ConstantInt::get(IntTy, 0); 999 1000 SizeOffsetEvalType Result = compute_(V); 1001 1002 if (!bothKnown(Result)) { 1003 // Erase everything that was computed in this iteration from the cache, so 1004 // that no dangling references are left behind. We could be a bit smarter if 1005 // we kept a dependency graph. It's probably not worth the complexity. 1006 for (const Value *SeenVal : SeenVals) { 1007 CacheMapTy::iterator CacheIt = CacheMap.find(SeenVal); 1008 // non-computable results can be safely cached 1009 if (CacheIt != CacheMap.end() && anyKnown(CacheIt->second)) 1010 CacheMap.erase(CacheIt); 1011 } 1012 1013 // Erase any instructions we inserted as part of the traversal. 1014 for (Instruction *I : InsertedInstructions) { 1015 I->replaceAllUsesWith(UndefValue::get(I->getType())); 1016 I->eraseFromParent(); 1017 } 1018 } 1019 1020 SeenVals.clear(); 1021 InsertedInstructions.clear(); 1022 return Result; 1023 } 1024 1025 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute_(Value *V) { 1026 ObjectSizeOffsetVisitor Visitor(DL, TLI, Context, EvalOpts); 1027 SizeOffsetType Const = Visitor.compute(V); 1028 if (Visitor.bothKnown(Const)) 1029 return std::make_pair(ConstantInt::get(Context, Const.first), 1030 ConstantInt::get(Context, Const.second)); 1031 1032 V = V->stripPointerCasts(); 1033 1034 // Check cache. 1035 CacheMapTy::iterator CacheIt = CacheMap.find(V); 1036 if (CacheIt != CacheMap.end()) 1037 return CacheIt->second; 1038 1039 // Always generate code immediately before the instruction being 1040 // processed, so that the generated code dominates the same BBs. 1041 BuilderTy::InsertPointGuard Guard(Builder); 1042 if (Instruction *I = dyn_cast<Instruction>(V)) 1043 Builder.SetInsertPoint(I); 1044 1045 // Now compute the size and offset. 1046 SizeOffsetEvalType Result; 1047 1048 // Record the pointers that were handled in this run, so that they can be 1049 // cleaned later if something fails. We also use this set to break cycles that 1050 // can occur in dead code. 1051 if (!SeenVals.insert(V).second) { 1052 Result = unknown(); 1053 } else if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 1054 Result = visitGEPOperator(*GEP); 1055 } else if (Instruction *I = dyn_cast<Instruction>(V)) { 1056 Result = visit(*I); 1057 } else if (isa<Argument>(V) || 1058 (isa<ConstantExpr>(V) && 1059 cast<ConstantExpr>(V)->getOpcode() == Instruction::IntToPtr) || 1060 isa<GlobalAlias>(V) || 1061 isa<GlobalVariable>(V)) { 1062 // Ignore values where we cannot do more than ObjectSizeVisitor. 1063 Result = unknown(); 1064 } else { 1065 LLVM_DEBUG( 1066 dbgs() << "ObjectSizeOffsetEvaluator::compute() unhandled value: " << *V 1067 << '\n'); 1068 Result = unknown(); 1069 } 1070 1071 // Don't reuse CacheIt since it may be invalid at this point. 1072 CacheMap[V] = Result; 1073 return Result; 1074 } 1075 1076 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitAllocaInst(AllocaInst &I) { 1077 if (!I.getAllocatedType()->isSized()) 1078 return unknown(); 1079 1080 // must be a VLA 1081 assert(I.isArrayAllocation()); 1082 1083 // If needed, adjust the alloca's operand size to match the pointer size. 1084 // Subsequent math operations expect the types to match. 1085 Value *ArraySize = Builder.CreateZExtOrTrunc( 1086 I.getArraySize(), DL.getIntPtrType(I.getContext())); 1087 assert(ArraySize->getType() == Zero->getType() && 1088 "Expected zero constant to have pointer type"); 1089 1090 Value *Size = ConstantInt::get(ArraySize->getType(), 1091 DL.getTypeAllocSize(I.getAllocatedType())); 1092 Size = Builder.CreateMul(Size, ArraySize); 1093 return std::make_pair(Size, Zero); 1094 } 1095 1096 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitCallBase(CallBase &CB) { 1097 Optional<AllocFnsTy> FnData = getAllocationSize(&CB, TLI); 1098 if (!FnData) 1099 return unknown(); 1100 1101 // Handle strdup-like functions separately. 1102 if (FnData->AllocTy == StrDupLike) { 1103 // TODO: implement evaluation of strdup/strndup 1104 return unknown(); 1105 } 1106 1107 Value *FirstArg = CB.getArgOperand(FnData->FstParam); 1108 FirstArg = Builder.CreateZExtOrTrunc(FirstArg, IntTy); 1109 if (FnData->SndParam < 0) 1110 return std::make_pair(FirstArg, Zero); 1111 1112 Value *SecondArg = CB.getArgOperand(FnData->SndParam); 1113 SecondArg = Builder.CreateZExtOrTrunc(SecondArg, IntTy); 1114 Value *Size = Builder.CreateMul(FirstArg, SecondArg); 1115 return std::make_pair(Size, Zero); 1116 } 1117 1118 SizeOffsetEvalType 1119 ObjectSizeOffsetEvaluator::visitExtractElementInst(ExtractElementInst&) { 1120 return unknown(); 1121 } 1122 1123 SizeOffsetEvalType 1124 ObjectSizeOffsetEvaluator::visitExtractValueInst(ExtractValueInst&) { 1125 return unknown(); 1126 } 1127 1128 SizeOffsetEvalType 1129 ObjectSizeOffsetEvaluator::visitGEPOperator(GEPOperator &GEP) { 1130 SizeOffsetEvalType PtrData = compute_(GEP.getPointerOperand()); 1131 if (!bothKnown(PtrData)) 1132 return unknown(); 1133 1134 Value *Offset = EmitGEPOffset(&Builder, DL, &GEP, /*NoAssumptions=*/true); 1135 Offset = Builder.CreateAdd(PtrData.second, Offset); 1136 return std::make_pair(PtrData.first, Offset); 1137 } 1138 1139 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitIntToPtrInst(IntToPtrInst&) { 1140 // clueless 1141 return unknown(); 1142 } 1143 1144 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitLoadInst(LoadInst &LI) { 1145 return unknown(); 1146 } 1147 1148 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitPHINode(PHINode &PHI) { 1149 // Create 2 PHIs: one for size and another for offset. 1150 PHINode *SizePHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 1151 PHINode *OffsetPHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues()); 1152 1153 // Insert right away in the cache to handle recursive PHIs. 1154 CacheMap[&PHI] = std::make_pair(SizePHI, OffsetPHI); 1155 1156 // Compute offset/size for each PHI incoming pointer. 1157 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i) { 1158 Builder.SetInsertPoint(&*PHI.getIncomingBlock(i)->getFirstInsertionPt()); 1159 SizeOffsetEvalType EdgeData = compute_(PHI.getIncomingValue(i)); 1160 1161 if (!bothKnown(EdgeData)) { 1162 OffsetPHI->replaceAllUsesWith(UndefValue::get(IntTy)); 1163 OffsetPHI->eraseFromParent(); 1164 InsertedInstructions.erase(OffsetPHI); 1165 SizePHI->replaceAllUsesWith(UndefValue::get(IntTy)); 1166 SizePHI->eraseFromParent(); 1167 InsertedInstructions.erase(SizePHI); 1168 return unknown(); 1169 } 1170 SizePHI->addIncoming(EdgeData.first, PHI.getIncomingBlock(i)); 1171 OffsetPHI->addIncoming(EdgeData.second, PHI.getIncomingBlock(i)); 1172 } 1173 1174 Value *Size = SizePHI, *Offset = OffsetPHI; 1175 if (Value *Tmp = SizePHI->hasConstantValue()) { 1176 Size = Tmp; 1177 SizePHI->replaceAllUsesWith(Size); 1178 SizePHI->eraseFromParent(); 1179 InsertedInstructions.erase(SizePHI); 1180 } 1181 if (Value *Tmp = OffsetPHI->hasConstantValue()) { 1182 Offset = Tmp; 1183 OffsetPHI->replaceAllUsesWith(Offset); 1184 OffsetPHI->eraseFromParent(); 1185 InsertedInstructions.erase(OffsetPHI); 1186 } 1187 return std::make_pair(Size, Offset); 1188 } 1189 1190 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitSelectInst(SelectInst &I) { 1191 SizeOffsetEvalType TrueSide = compute_(I.getTrueValue()); 1192 SizeOffsetEvalType FalseSide = compute_(I.getFalseValue()); 1193 1194 if (!bothKnown(TrueSide) || !bothKnown(FalseSide)) 1195 return unknown(); 1196 if (TrueSide == FalseSide) 1197 return TrueSide; 1198 1199 Value *Size = Builder.CreateSelect(I.getCondition(), TrueSide.first, 1200 FalseSide.first); 1201 Value *Offset = Builder.CreateSelect(I.getCondition(), TrueSide.second, 1202 FalseSide.second); 1203 return std::make_pair(Size, Offset); 1204 } 1205 1206 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitInstruction(Instruction &I) { 1207 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetEvaluator unknown instruction:" << I 1208 << '\n'); 1209 return unknown(); 1210 } 1211