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