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/STLExtras.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/Analysis/AliasAnalysis.h"
19 #include "llvm/Analysis/TargetFolder.h"
20 #include "llvm/Analysis/TargetLibraryInfo.h"
21 #include "llvm/Analysis/Utils/Local.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/Argument.h"
24 #include "llvm/IR/Attributes.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GlobalAlias.h"
30 #include "llvm/IR/GlobalVariable.h"
31 #include "llvm/IR/Instruction.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/Operator.h"
35 #include "llvm/IR/Type.h"
36 #include "llvm/IR/Value.h"
37 #include "llvm/Support/Casting.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/MathExtras.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include <cassert>
42 #include <cstdint>
43 #include <iterator>
44 #include <numeric>
45 #include <optional>
46 #include <type_traits>
47 #include <utility>
48
49 using namespace llvm;
50
51 #define DEBUG_TYPE "memory-builtins"
52
53 enum AllocType : uint8_t {
54 OpNewLike = 1<<0, // allocates; never returns null
55 MallocLike = 1<<1, // allocates; may return null
56 StrDupLike = 1<<2,
57 MallocOrOpNewLike = MallocLike | OpNewLike,
58 AllocLike = MallocOrOpNewLike | StrDupLike,
59 AnyAlloc = AllocLike
60 };
61
62 enum class MallocFamily {
63 Malloc,
64 CPPNew, // new(unsigned int)
65 CPPNewAligned, // new(unsigned int, align_val_t)
66 CPPNewArray, // new[](unsigned int)
67 CPPNewArrayAligned, // new[](unsigned long, align_val_t)
68 MSVCNew, // new(unsigned int)
69 MSVCArrayNew, // new[](unsigned int)
70 VecMalloc,
71 KmpcAllocShared,
72 };
73
mangledNameForMallocFamily(const MallocFamily & Family)74 StringRef mangledNameForMallocFamily(const MallocFamily &Family) {
75 switch (Family) {
76 case MallocFamily::Malloc:
77 return "malloc";
78 case MallocFamily::CPPNew:
79 return "_Znwm";
80 case MallocFamily::CPPNewAligned:
81 return "_ZnwmSt11align_val_t";
82 case MallocFamily::CPPNewArray:
83 return "_Znam";
84 case MallocFamily::CPPNewArrayAligned:
85 return "_ZnamSt11align_val_t";
86 case MallocFamily::MSVCNew:
87 return "??2@YAPAXI@Z";
88 case MallocFamily::MSVCArrayNew:
89 return "??_U@YAPAXI@Z";
90 case MallocFamily::VecMalloc:
91 return "vec_malloc";
92 case MallocFamily::KmpcAllocShared:
93 return "__kmpc_alloc_shared";
94 }
95 llvm_unreachable("missing an alloc family");
96 }
97
98 struct AllocFnsTy {
99 AllocType AllocTy;
100 unsigned NumParams;
101 // First and Second size parameters (or -1 if unused)
102 int FstParam, SndParam;
103 // Alignment parameter for aligned_alloc and aligned new
104 int AlignParam;
105 // Name of default allocator function to group malloc/free calls by family
106 MallocFamily Family;
107 };
108
109 // clang-format off
110 // FIXME: certain users need more information. E.g., SimplifyLibCalls needs to
111 // know which functions are nounwind, noalias, nocapture parameters, etc.
112 static const std::pair<LibFunc, AllocFnsTy> AllocationFnData[] = {
113 {LibFunc_Znwj, {OpNewLike, 1, 0, -1, -1, MallocFamily::CPPNew}}, // new(unsigned int)
114 {LibFunc_ZnwjRKSt9nothrow_t, {MallocLike, 2, 0, -1, -1, MallocFamily::CPPNew}}, // new(unsigned int, nothrow)
115 {LibFunc_ZnwjSt11align_val_t, {OpNewLike, 2, 0, -1, 1, MallocFamily::CPPNewAligned}}, // new(unsigned int, align_val_t)
116 {LibFunc_ZnwjSt11align_val_tRKSt9nothrow_t, {MallocLike, 3, 0, -1, 1, MallocFamily::CPPNewAligned}}, // new(unsigned int, align_val_t, nothrow)
117 {LibFunc_Znwm, {OpNewLike, 1, 0, -1, -1, MallocFamily::CPPNew}}, // new(unsigned long)
118 {LibFunc_ZnwmRKSt9nothrow_t, {MallocLike, 2, 0, -1, -1, MallocFamily::CPPNew}}, // new(unsigned long, nothrow)
119 {LibFunc_ZnwmSt11align_val_t, {OpNewLike, 2, 0, -1, 1, MallocFamily::CPPNewAligned}}, // new(unsigned long, align_val_t)
120 {LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t, {MallocLike, 3, 0, -1, 1, MallocFamily::CPPNewAligned}}, // new(unsigned long, align_val_t, nothrow)
121 {LibFunc_Znaj, {OpNewLike, 1, 0, -1, -1, MallocFamily::CPPNewArray}}, // new[](unsigned int)
122 {LibFunc_ZnajRKSt9nothrow_t, {MallocLike, 2, 0, -1, -1, MallocFamily::CPPNewArray}}, // new[](unsigned int, nothrow)
123 {LibFunc_ZnajSt11align_val_t, {OpNewLike, 2, 0, -1, 1, MallocFamily::CPPNewArrayAligned}}, // new[](unsigned int, align_val_t)
124 {LibFunc_ZnajSt11align_val_tRKSt9nothrow_t, {MallocLike, 3, 0, -1, 1, MallocFamily::CPPNewArrayAligned}}, // new[](unsigned int, align_val_t, nothrow)
125 {LibFunc_Znam, {OpNewLike, 1, 0, -1, -1, MallocFamily::CPPNewArray}}, // new[](unsigned long)
126 {LibFunc_ZnamRKSt9nothrow_t, {MallocLike, 2, 0, -1, -1, MallocFamily::CPPNewArray}}, // new[](unsigned long, nothrow)
127 {LibFunc_ZnamSt11align_val_t, {OpNewLike, 2, 0, -1, 1, MallocFamily::CPPNewArrayAligned}}, // new[](unsigned long, align_val_t)
128 {LibFunc_ZnamSt11align_val_tRKSt9nothrow_t, {MallocLike, 3, 0, -1, 1, MallocFamily::CPPNewArrayAligned}}, // new[](unsigned long, align_val_t, nothrow)
129 {LibFunc_msvc_new_int, {OpNewLike, 1, 0, -1, -1, MallocFamily::MSVCNew}}, // new(unsigned int)
130 {LibFunc_msvc_new_int_nothrow, {MallocLike, 2, 0, -1, -1, MallocFamily::MSVCNew}}, // new(unsigned int, nothrow)
131 {LibFunc_msvc_new_longlong, {OpNewLike, 1, 0, -1, -1, MallocFamily::MSVCNew}}, // new(unsigned long long)
132 {LibFunc_msvc_new_longlong_nothrow, {MallocLike, 2, 0, -1, -1, MallocFamily::MSVCNew}}, // new(unsigned long long, nothrow)
133 {LibFunc_msvc_new_array_int, {OpNewLike, 1, 0, -1, -1, MallocFamily::MSVCArrayNew}}, // new[](unsigned int)
134 {LibFunc_msvc_new_array_int_nothrow, {MallocLike, 2, 0, -1, -1, MallocFamily::MSVCArrayNew}}, // new[](unsigned int, nothrow)
135 {LibFunc_msvc_new_array_longlong, {OpNewLike, 1, 0, -1, -1, MallocFamily::MSVCArrayNew}}, // new[](unsigned long long)
136 {LibFunc_msvc_new_array_longlong_nothrow, {MallocLike, 2, 0, -1, -1, MallocFamily::MSVCArrayNew}}, // new[](unsigned long long, nothrow)
137 {LibFunc_strdup, {StrDupLike, 1, -1, -1, -1, MallocFamily::Malloc}},
138 {LibFunc_dunder_strdup, {StrDupLike, 1, -1, -1, -1, MallocFamily::Malloc}},
139 {LibFunc_strndup, {StrDupLike, 2, 1, -1, -1, MallocFamily::Malloc}},
140 {LibFunc_dunder_strndup, {StrDupLike, 2, 1, -1, -1, MallocFamily::Malloc}},
141 {LibFunc___kmpc_alloc_shared, {MallocLike, 1, 0, -1, -1, MallocFamily::KmpcAllocShared}},
142 };
143 // clang-format on
144
getCalledFunction(const Value * V,bool & IsNoBuiltin)145 static const Function *getCalledFunction(const Value *V,
146 bool &IsNoBuiltin) {
147 // Don't care about intrinsics in this case.
148 if (isa<IntrinsicInst>(V))
149 return nullptr;
150
151 const auto *CB = dyn_cast<CallBase>(V);
152 if (!CB)
153 return nullptr;
154
155 IsNoBuiltin = CB->isNoBuiltin();
156
157 if (const Function *Callee = CB->getCalledFunction())
158 return Callee;
159 return nullptr;
160 }
161
162 /// Returns the allocation data for the given value if it's a call to a known
163 /// allocation function.
164 static std::optional<AllocFnsTy>
getAllocationDataForFunction(const Function * Callee,AllocType AllocTy,const TargetLibraryInfo * TLI)165 getAllocationDataForFunction(const Function *Callee, AllocType AllocTy,
166 const TargetLibraryInfo *TLI) {
167 // Don't perform a slow TLI lookup, if this function doesn't return a pointer
168 // and thus can't be an allocation function.
169 if (!Callee->getReturnType()->isPointerTy())
170 return std::nullopt;
171
172 // Make sure that the function is available.
173 LibFunc TLIFn;
174 if (!TLI || !TLI->getLibFunc(*Callee, TLIFn) || !TLI->has(TLIFn))
175 return std::nullopt;
176
177 const auto *Iter = find_if(
178 AllocationFnData, [TLIFn](const std::pair<LibFunc, AllocFnsTy> &P) {
179 return P.first == TLIFn;
180 });
181
182 if (Iter == std::end(AllocationFnData))
183 return std::nullopt;
184
185 const AllocFnsTy *FnData = &Iter->second;
186 if ((FnData->AllocTy & AllocTy) != FnData->AllocTy)
187 return std::nullopt;
188
189 // Check function prototype.
190 int FstParam = FnData->FstParam;
191 int SndParam = FnData->SndParam;
192 FunctionType *FTy = Callee->getFunctionType();
193
194 if (FTy->getReturnType()->isPointerTy() &&
195 FTy->getNumParams() == FnData->NumParams &&
196 (FstParam < 0 ||
197 (FTy->getParamType(FstParam)->isIntegerTy(32) ||
198 FTy->getParamType(FstParam)->isIntegerTy(64))) &&
199 (SndParam < 0 ||
200 FTy->getParamType(SndParam)->isIntegerTy(32) ||
201 FTy->getParamType(SndParam)->isIntegerTy(64)))
202 return *FnData;
203 return std::nullopt;
204 }
205
206 static std::optional<AllocFnsTy>
getAllocationData(const Value * V,AllocType AllocTy,const TargetLibraryInfo * TLI)207 getAllocationData(const Value *V, AllocType AllocTy,
208 const TargetLibraryInfo *TLI) {
209 bool IsNoBuiltinCall;
210 if (const Function *Callee = getCalledFunction(V, IsNoBuiltinCall))
211 if (!IsNoBuiltinCall)
212 return getAllocationDataForFunction(Callee, AllocTy, TLI);
213 return std::nullopt;
214 }
215
216 static std::optional<AllocFnsTy>
getAllocationData(const Value * V,AllocType AllocTy,function_ref<const TargetLibraryInfo & (Function &)> GetTLI)217 getAllocationData(const Value *V, AllocType AllocTy,
218 function_ref<const TargetLibraryInfo &(Function &)> GetTLI) {
219 bool IsNoBuiltinCall;
220 if (const Function *Callee = getCalledFunction(V, IsNoBuiltinCall))
221 if (!IsNoBuiltinCall)
222 return getAllocationDataForFunction(
223 Callee, AllocTy, &GetTLI(const_cast<Function &>(*Callee)));
224 return std::nullopt;
225 }
226
227 static std::optional<AllocFnsTy>
getAllocationSize(const Value * V,const TargetLibraryInfo * TLI)228 getAllocationSize(const Value *V, const TargetLibraryInfo *TLI) {
229 bool IsNoBuiltinCall;
230 const Function *Callee =
231 getCalledFunction(V, IsNoBuiltinCall);
232 if (!Callee)
233 return std::nullopt;
234
235 // Prefer to use existing information over allocsize. This will give us an
236 // accurate AllocTy.
237 if (!IsNoBuiltinCall)
238 if (std::optional<AllocFnsTy> Data =
239 getAllocationDataForFunction(Callee, AnyAlloc, TLI))
240 return Data;
241
242 Attribute Attr = Callee->getFnAttribute(Attribute::AllocSize);
243 if (Attr == Attribute())
244 return std::nullopt;
245
246 std::pair<unsigned, std::optional<unsigned>> Args = Attr.getAllocSizeArgs();
247
248 AllocFnsTy Result;
249 // Because allocsize only tells us how many bytes are allocated, we're not
250 // really allowed to assume anything, so we use MallocLike.
251 Result.AllocTy = MallocLike;
252 Result.NumParams = Callee->getNumOperands();
253 Result.FstParam = Args.first;
254 Result.SndParam = Args.second.value_or(-1);
255 // Allocsize has no way to specify an alignment argument
256 Result.AlignParam = -1;
257 return Result;
258 }
259
getAllocFnKind(const Value * V)260 static AllocFnKind getAllocFnKind(const Value *V) {
261 if (const auto *CB = dyn_cast<CallBase>(V)) {
262 Attribute Attr = CB->getFnAttr(Attribute::AllocKind);
263 if (Attr.isValid())
264 return AllocFnKind(Attr.getValueAsInt());
265 }
266 return AllocFnKind::Unknown;
267 }
268
getAllocFnKind(const Function * F)269 static AllocFnKind getAllocFnKind(const Function *F) {
270 Attribute Attr = F->getFnAttribute(Attribute::AllocKind);
271 if (Attr.isValid())
272 return AllocFnKind(Attr.getValueAsInt());
273 return AllocFnKind::Unknown;
274 }
275
checkFnAllocKind(const Value * V,AllocFnKind Wanted)276 static bool checkFnAllocKind(const Value *V, AllocFnKind Wanted) {
277 return (getAllocFnKind(V) & Wanted) != AllocFnKind::Unknown;
278 }
279
checkFnAllocKind(const Function * F,AllocFnKind Wanted)280 static bool checkFnAllocKind(const Function *F, AllocFnKind Wanted) {
281 return (getAllocFnKind(F) & Wanted) != AllocFnKind::Unknown;
282 }
283
284 /// Tests if a value is a call or invoke to a library function that
285 /// allocates or reallocates memory (either malloc, calloc, realloc, or strdup
286 /// like).
isAllocationFn(const Value * V,const TargetLibraryInfo * TLI)287 bool llvm::isAllocationFn(const Value *V, const TargetLibraryInfo *TLI) {
288 return getAllocationData(V, AnyAlloc, TLI).has_value() ||
289 checkFnAllocKind(V, AllocFnKind::Alloc | AllocFnKind::Realloc);
290 }
isAllocationFn(const Value * V,function_ref<const TargetLibraryInfo & (Function &)> GetTLI)291 bool llvm::isAllocationFn(
292 const Value *V,
293 function_ref<const TargetLibraryInfo &(Function &)> GetTLI) {
294 return getAllocationData(V, AnyAlloc, GetTLI).has_value() ||
295 checkFnAllocKind(V, AllocFnKind::Alloc | AllocFnKind::Realloc);
296 }
297
298 /// Tests if a value is a call or invoke to a library function that
299 /// allocates memory via new.
isNewLikeFn(const Value * V,const TargetLibraryInfo * TLI)300 bool llvm::isNewLikeFn(const Value *V, const TargetLibraryInfo *TLI) {
301 return getAllocationData(V, OpNewLike, TLI).has_value();
302 }
303
304 /// Tests if a value is a call or invoke to a library function that
305 /// allocates memory similar to malloc or calloc.
isMallocOrCallocLikeFn(const Value * V,const TargetLibraryInfo * TLI)306 bool llvm::isMallocOrCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI) {
307 // TODO: Function behavior does not match name.
308 return getAllocationData(V, MallocOrOpNewLike, TLI).has_value();
309 }
310
311 /// Tests if a value is a call or invoke to a library function that
312 /// allocates memory (either malloc, calloc, or strdup like).
isAllocLikeFn(const Value * V,const TargetLibraryInfo * TLI)313 bool llvm::isAllocLikeFn(const Value *V, const TargetLibraryInfo *TLI) {
314 return getAllocationData(V, AllocLike, TLI).has_value() ||
315 checkFnAllocKind(V, AllocFnKind::Alloc);
316 }
317
318 /// Tests if a functions is a call or invoke to a library function that
319 /// reallocates memory (e.g., realloc).
isReallocLikeFn(const Function * F)320 bool llvm::isReallocLikeFn(const Function *F) {
321 return checkFnAllocKind(F, AllocFnKind::Realloc);
322 }
323
getReallocatedOperand(const CallBase * CB)324 Value *llvm::getReallocatedOperand(const CallBase *CB) {
325 if (checkFnAllocKind(CB, AllocFnKind::Realloc))
326 return CB->getArgOperandWithAttribute(Attribute::AllocatedPointer);
327 return nullptr;
328 }
329
isRemovableAlloc(const CallBase * CB,const TargetLibraryInfo * TLI)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
getAllocAlignment(const CallBase * V,const TargetLibraryInfo * TLI)341 Value *llvm::getAllocAlignment(const CallBase *V,
342 const TargetLibraryInfo *TLI) {
343 const std::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.
CheckedZextOrTrunc(APInt & I,unsigned IntTyBits)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 std::optional<APInt>
getAllocSize(const CallBase * CB,const TargetLibraryInfo * TLI,function_ref<const Value * (const Value *)> Mapper)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 std::optional<AllocFnsTy> FnData = getAllocationSize(CB, TLI);
372 if (!FnData)
373 return std::nullopt;
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 std::nullopt;
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 std::nullopt;
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 std::nullopt;
404
405 APInt Size = Arg->getValue();
406 if (!CheckedZextOrTrunc(Size, IntTyBits))
407 return std::nullopt;
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 std::nullopt;
416
417 APInt NumElems = Arg->getValue();
418 if (!CheckedZextOrTrunc(NumElems, IntTyBits))
419 return std::nullopt;
420
421 bool Overflow;
422 Size = Size.umul_ov(NumElems, Overflow);
423 if (Overflow)
424 return std::nullopt;
425 return Size;
426 }
427
getInitialValueOfAllocation(const Value * V,const TargetLibraryInfo * TLI,Type * Ty)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 are uninitialized (undef)
436 if (getAllocationData(Alloc, MallocOrOpNewLike, TLI).has_value())
437 return UndefValue::get(Ty);
438
439 AllocFnKind AK = getAllocFnKind(Alloc);
440 if ((AK & AllocFnKind::Uninitialized) != AllocFnKind::Unknown)
441 return UndefValue::get(Ty);
442 if ((AK & AllocFnKind::Zeroed) != AllocFnKind::Unknown)
443 return Constant::getNullValue(Ty);
444
445 return nullptr;
446 }
447
448 struct FreeFnsTy {
449 unsigned NumParams;
450 // Name of default allocator function to group malloc/free calls by family
451 MallocFamily Family;
452 };
453
454 // clang-format off
455 static const std::pair<LibFunc, FreeFnsTy> FreeFnData[] = {
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
getFreeFunctionDataForFunction(const Function * Callee,const LibFunc TLIFn)488 std::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 std::nullopt;
496 return Iter->second;
497 }
498
499 std::optional<StringRef>
getAllocationFamily(const Value * I,const TargetLibraryInfo * TLI)500 llvm::getAllocationFamily(const Value *I, const TargetLibraryInfo *TLI) {
501 bool IsNoBuiltin;
502 const Function *Callee = getCalledFunction(I, IsNoBuiltin);
503 if (Callee == nullptr || IsNoBuiltin)
504 return std::nullopt;
505 LibFunc TLIFn;
506
507 if (TLI && TLI->getLibFunc(*Callee, TLIFn) && TLI->has(TLIFn)) {
508 // Callee is some known library function.
509 const auto AllocData = getAllocationDataForFunction(Callee, AnyAlloc, TLI);
510 if (AllocData)
511 return mangledNameForMallocFamily(AllocData->Family);
512 const auto FreeData = getFreeFunctionDataForFunction(Callee, TLIFn);
513 if (FreeData)
514 return mangledNameForMallocFamily(FreeData->Family);
515 }
516 // Callee isn't a known library function, still check attributes.
517 if (checkFnAllocKind(I, AllocFnKind::Free | AllocFnKind::Alloc |
518 AllocFnKind::Realloc)) {
519 Attribute Attr = cast<CallBase>(I)->getFnAttr("alloc-family");
520 if (Attr.isValid())
521 return Attr.getValueAsString();
522 }
523 return std::nullopt;
524 }
525
526 /// isLibFreeFunction - Returns true if the function is a builtin free()
isLibFreeFunction(const Function * F,const LibFunc TLIFn)527 bool llvm::isLibFreeFunction(const Function *F, const LibFunc TLIFn) {
528 std::optional<FreeFnsTy> FnData = getFreeFunctionDataForFunction(F, TLIFn);
529 if (!FnData)
530 return checkFnAllocKind(F, AllocFnKind::Free);
531
532 // Check free prototype.
533 // FIXME: workaround for PR5130, this will be obsolete when a nobuiltin
534 // attribute will exist.
535 FunctionType *FTy = F->getFunctionType();
536 if (!FTy->getReturnType()->isVoidTy())
537 return false;
538 if (FTy->getNumParams() != FnData->NumParams)
539 return false;
540 if (!FTy->getParamType(0)->isPointerTy())
541 return false;
542
543 return true;
544 }
545
getFreedOperand(const CallBase * CB,const TargetLibraryInfo * TLI)546 Value *llvm::getFreedOperand(const CallBase *CB, const TargetLibraryInfo *TLI) {
547 bool IsNoBuiltinCall;
548 const Function *Callee = getCalledFunction(CB, IsNoBuiltinCall);
549 if (Callee == nullptr || IsNoBuiltinCall)
550 return nullptr;
551
552 LibFunc TLIFn;
553 if (TLI && TLI->getLibFunc(*Callee, TLIFn) && TLI->has(TLIFn) &&
554 isLibFreeFunction(Callee, TLIFn)) {
555 // All currently supported free functions free the first argument.
556 return CB->getArgOperand(0);
557 }
558
559 if (checkFnAllocKind(CB, AllocFnKind::Free))
560 return CB->getArgOperandWithAttribute(Attribute::AllocatedPointer);
561
562 return nullptr;
563 }
564
565 //===----------------------------------------------------------------------===//
566 // Utility functions to compute size of objects.
567 //
getSizeWithOverflow(const SizeOffsetType & Data)568 static APInt getSizeWithOverflow(const SizeOffsetType &Data) {
569 if (Data.second.isNegative() || Data.first.ult(Data.second))
570 return APInt(Data.first.getBitWidth(), 0);
571 return Data.first - Data.second;
572 }
573
574 /// Compute the size of the object pointed by Ptr. Returns true and the
575 /// object size in Size if successful, and false otherwise.
576 /// If RoundToAlign is true, then Size is rounded up to the alignment of
577 /// allocas, byval arguments, and global variables.
getObjectSize(const Value * Ptr,uint64_t & Size,const DataLayout & DL,const TargetLibraryInfo * TLI,ObjectSizeOpts Opts)578 bool llvm::getObjectSize(const Value *Ptr, uint64_t &Size, const DataLayout &DL,
579 const TargetLibraryInfo *TLI, ObjectSizeOpts Opts) {
580 ObjectSizeOffsetVisitor Visitor(DL, TLI, Ptr->getContext(), Opts);
581 SizeOffsetType Data = Visitor.compute(const_cast<Value*>(Ptr));
582 if (!Visitor.bothKnown(Data))
583 return false;
584
585 Size = getSizeWithOverflow(Data).getZExtValue();
586 return true;
587 }
588
lowerObjectSizeCall(IntrinsicInst * ObjectSize,const DataLayout & DL,const TargetLibraryInfo * TLI,bool MustSucceed)589 Value *llvm::lowerObjectSizeCall(IntrinsicInst *ObjectSize,
590 const DataLayout &DL,
591 const TargetLibraryInfo *TLI,
592 bool MustSucceed) {
593 return lowerObjectSizeCall(ObjectSize, DL, TLI, /*AAResults=*/nullptr,
594 MustSucceed);
595 }
596
lowerObjectSizeCall(IntrinsicInst * ObjectSize,const DataLayout & DL,const TargetLibraryInfo * TLI,AAResults * AA,bool MustSucceed)597 Value *llvm::lowerObjectSizeCall(IntrinsicInst *ObjectSize,
598 const DataLayout &DL,
599 const TargetLibraryInfo *TLI, AAResults *AA,
600 bool MustSucceed) {
601 assert(ObjectSize->getIntrinsicID() == Intrinsic::objectsize &&
602 "ObjectSize must be a call to llvm.objectsize!");
603
604 bool MaxVal = cast<ConstantInt>(ObjectSize->getArgOperand(1))->isZero();
605 ObjectSizeOpts EvalOptions;
606 EvalOptions.AA = AA;
607
608 // Unless we have to fold this to something, try to be as accurate as
609 // possible.
610 if (MustSucceed)
611 EvalOptions.EvalMode =
612 MaxVal ? ObjectSizeOpts::Mode::Max : ObjectSizeOpts::Mode::Min;
613 else
614 EvalOptions.EvalMode = ObjectSizeOpts::Mode::ExactSizeFromOffset;
615
616 EvalOptions.NullIsUnknownSize =
617 cast<ConstantInt>(ObjectSize->getArgOperand(2))->isOne();
618
619 auto *ResultType = cast<IntegerType>(ObjectSize->getType());
620 bool StaticOnly = cast<ConstantInt>(ObjectSize->getArgOperand(3))->isZero();
621 if (StaticOnly) {
622 // FIXME: Does it make sense to just return a failure value if the size won't
623 // fit in the output and `!MustSucceed`?
624 uint64_t Size;
625 if (getObjectSize(ObjectSize->getArgOperand(0), Size, DL, TLI, EvalOptions) &&
626 isUIntN(ResultType->getBitWidth(), Size))
627 return ConstantInt::get(ResultType, Size);
628 } else {
629 LLVMContext &Ctx = ObjectSize->getFunction()->getContext();
630 ObjectSizeOffsetEvaluator Eval(DL, TLI, Ctx, EvalOptions);
631 SizeOffsetEvalType SizeOffsetPair =
632 Eval.compute(ObjectSize->getArgOperand(0));
633
634 if (SizeOffsetPair != ObjectSizeOffsetEvaluator::unknown()) {
635 IRBuilder<TargetFolder> Builder(Ctx, TargetFolder(DL));
636 Builder.SetInsertPoint(ObjectSize);
637
638 // If we've outside the end of the object, then we can always access
639 // exactly 0 bytes.
640 Value *ResultSize =
641 Builder.CreateSub(SizeOffsetPair.first, SizeOffsetPair.second);
642 Value *UseZero =
643 Builder.CreateICmpULT(SizeOffsetPair.first, SizeOffsetPair.second);
644 ResultSize = Builder.CreateZExtOrTrunc(ResultSize, ResultType);
645 Value *Ret = Builder.CreateSelect(
646 UseZero, ConstantInt::get(ResultType, 0), ResultSize);
647
648 // The non-constant size expression cannot evaluate to -1.
649 if (!isa<Constant>(SizeOffsetPair.first) ||
650 !isa<Constant>(SizeOffsetPair.second))
651 Builder.CreateAssumption(
652 Builder.CreateICmpNE(Ret, ConstantInt::get(ResultType, -1)));
653
654 return Ret;
655 }
656 }
657
658 if (!MustSucceed)
659 return nullptr;
660
661 return ConstantInt::get(ResultType, MaxVal ? -1ULL : 0);
662 }
663
664 STATISTIC(ObjectVisitorArgument,
665 "Number of arguments with unsolved size and offset");
666 STATISTIC(ObjectVisitorLoad,
667 "Number of load instructions with unsolved size and offset");
668
align(APInt Size,MaybeAlign Alignment)669 APInt ObjectSizeOffsetVisitor::align(APInt Size, MaybeAlign Alignment) {
670 if (Options.RoundToAlign && Alignment)
671 return APInt(IntTyBits, alignTo(Size.getZExtValue(), *Alignment));
672 return Size;
673 }
674
ObjectSizeOffsetVisitor(const DataLayout & DL,const TargetLibraryInfo * TLI,LLVMContext & Context,ObjectSizeOpts Options)675 ObjectSizeOffsetVisitor::ObjectSizeOffsetVisitor(const DataLayout &DL,
676 const TargetLibraryInfo *TLI,
677 LLVMContext &Context,
678 ObjectSizeOpts Options)
679 : DL(DL), TLI(TLI), Options(Options) {
680 // Pointer size must be rechecked for each object visited since it could have
681 // a different address space.
682 }
683
compute(Value * V)684 SizeOffsetType ObjectSizeOffsetVisitor::compute(Value *V) {
685 unsigned InitialIntTyBits = DL.getIndexTypeSizeInBits(V->getType());
686
687 // Stripping pointer casts can strip address space casts which can change the
688 // index type size. The invariant is that we use the value type to determine
689 // the index type size and if we stripped address space casts we have to
690 // readjust the APInt as we pass it upwards in order for the APInt to match
691 // the type the caller passed in.
692 APInt Offset(InitialIntTyBits, 0);
693 V = V->stripAndAccumulateConstantOffsets(
694 DL, Offset, /* AllowNonInbounds */ true, /* AllowInvariantGroup */ true);
695
696 // Later we use the index type size and zero but it will match the type of the
697 // value that is passed to computeImpl.
698 IntTyBits = DL.getIndexTypeSizeInBits(V->getType());
699 Zero = APInt::getZero(IntTyBits);
700
701 bool IndexTypeSizeChanged = InitialIntTyBits != IntTyBits;
702 if (!IndexTypeSizeChanged && Offset.isZero())
703 return computeImpl(V);
704
705 // We stripped an address space cast that changed the index type size or we
706 // accumulated some constant offset (or both). Readjust the bit width to match
707 // the argument index type size and apply the offset, as required.
708 SizeOffsetType SOT = computeImpl(V);
709 if (IndexTypeSizeChanged) {
710 if (knownSize(SOT) && !::CheckedZextOrTrunc(SOT.first, InitialIntTyBits))
711 SOT.first = APInt();
712 if (knownOffset(SOT) && !::CheckedZextOrTrunc(SOT.second, InitialIntTyBits))
713 SOT.second = APInt();
714 }
715 // If the computed offset is "unknown" we cannot add the stripped offset.
716 return {SOT.first,
717 SOT.second.getBitWidth() > 1 ? SOT.second + Offset : SOT.second};
718 }
719
computeImpl(Value * V)720 SizeOffsetType ObjectSizeOffsetVisitor::computeImpl(Value *V) {
721 if (Instruction *I = dyn_cast<Instruction>(V)) {
722 // If we have already seen this instruction, bail out. Cycles can happen in
723 // unreachable code after constant propagation.
724 if (!SeenInsts.insert(I).second)
725 return unknown();
726
727 return visit(*I);
728 }
729 if (Argument *A = dyn_cast<Argument>(V))
730 return visitArgument(*A);
731 if (ConstantPointerNull *P = dyn_cast<ConstantPointerNull>(V))
732 return visitConstantPointerNull(*P);
733 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
734 return visitGlobalAlias(*GA);
735 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
736 return visitGlobalVariable(*GV);
737 if (UndefValue *UV = dyn_cast<UndefValue>(V))
738 return visitUndefValue(*UV);
739
740 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetVisitor::compute() unhandled value: "
741 << *V << '\n');
742 return unknown();
743 }
744
CheckedZextOrTrunc(APInt & I)745 bool ObjectSizeOffsetVisitor::CheckedZextOrTrunc(APInt &I) {
746 return ::CheckedZextOrTrunc(I, IntTyBits);
747 }
748
visitAllocaInst(AllocaInst & I)749 SizeOffsetType ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) {
750 TypeSize ElemSize = DL.getTypeAllocSize(I.getAllocatedType());
751 if (ElemSize.isScalable() && Options.EvalMode != ObjectSizeOpts::Mode::Min)
752 return unknown();
753 APInt Size(IntTyBits, ElemSize.getKnownMinValue());
754 if (!I.isArrayAllocation())
755 return std::make_pair(align(Size, I.getAlign()), Zero);
756
757 Value *ArraySize = I.getArraySize();
758 if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) {
759 APInt NumElems = C->getValue();
760 if (!CheckedZextOrTrunc(NumElems))
761 return unknown();
762
763 bool Overflow;
764 Size = Size.umul_ov(NumElems, Overflow);
765 return Overflow ? unknown()
766 : std::make_pair(align(Size, I.getAlign()), Zero);
767 }
768 return unknown();
769 }
770
visitArgument(Argument & A)771 SizeOffsetType ObjectSizeOffsetVisitor::visitArgument(Argument &A) {
772 Type *MemoryTy = A.getPointeeInMemoryValueType();
773 // No interprocedural analysis is done at the moment.
774 if (!MemoryTy|| !MemoryTy->isSized()) {
775 ++ObjectVisitorArgument;
776 return unknown();
777 }
778
779 APInt Size(IntTyBits, DL.getTypeAllocSize(MemoryTy));
780 return std::make_pair(align(Size, A.getParamAlign()), Zero);
781 }
782
visitCallBase(CallBase & CB)783 SizeOffsetType ObjectSizeOffsetVisitor::visitCallBase(CallBase &CB) {
784 if (std::optional<APInt> Size = getAllocSize(&CB, TLI))
785 return std::make_pair(*Size, Zero);
786 return unknown();
787 }
788
789 SizeOffsetType
visitConstantPointerNull(ConstantPointerNull & CPN)790 ObjectSizeOffsetVisitor::visitConstantPointerNull(ConstantPointerNull& CPN) {
791 // If null is unknown, there's nothing we can do. Additionally, non-zero
792 // address spaces can make use of null, so we don't presume to know anything
793 // about that.
794 //
795 // TODO: How should this work with address space casts? We currently just drop
796 // them on the floor, but it's unclear what we should do when a NULL from
797 // addrspace(1) gets casted to addrspace(0) (or vice-versa).
798 if (Options.NullIsUnknownSize || CPN.getType()->getAddressSpace())
799 return unknown();
800 return std::make_pair(Zero, Zero);
801 }
802
803 SizeOffsetType
visitExtractElementInst(ExtractElementInst &)804 ObjectSizeOffsetVisitor::visitExtractElementInst(ExtractElementInst&) {
805 return unknown();
806 }
807
808 SizeOffsetType
visitExtractValueInst(ExtractValueInst &)809 ObjectSizeOffsetVisitor::visitExtractValueInst(ExtractValueInst&) {
810 // Easy cases were already folded by previous passes.
811 return unknown();
812 }
813
visitGlobalAlias(GlobalAlias & GA)814 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalAlias(GlobalAlias &GA) {
815 if (GA.isInterposable())
816 return unknown();
817 return compute(GA.getAliasee());
818 }
819
visitGlobalVariable(GlobalVariable & GV)820 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalVariable(GlobalVariable &GV){
821 if (!GV.hasDefinitiveInitializer())
822 return unknown();
823
824 APInt Size(IntTyBits, DL.getTypeAllocSize(GV.getValueType()));
825 return std::make_pair(align(Size, GV.getAlign()), Zero);
826 }
827
visitIntToPtrInst(IntToPtrInst &)828 SizeOffsetType ObjectSizeOffsetVisitor::visitIntToPtrInst(IntToPtrInst&) {
829 // clueless
830 return unknown();
831 }
832
findLoadSizeOffset(LoadInst & Load,BasicBlock & BB,BasicBlock::iterator From,SmallDenseMap<BasicBlock *,SizeOffsetType,8> & VisitedBlocks,unsigned & ScannedInstCount)833 SizeOffsetType ObjectSizeOffsetVisitor::findLoadSizeOffset(
834 LoadInst &Load, BasicBlock &BB, BasicBlock::iterator From,
835 SmallDenseMap<BasicBlock *, SizeOffsetType, 8> &VisitedBlocks,
836 unsigned &ScannedInstCount) {
837 constexpr unsigned MaxInstsToScan = 128;
838
839 auto Where = VisitedBlocks.find(&BB);
840 if (Where != VisitedBlocks.end())
841 return Where->second;
842
843 auto Unknown = [this, &BB, &VisitedBlocks]() {
844 return VisitedBlocks[&BB] = unknown();
845 };
846 auto Known = [&BB, &VisitedBlocks](SizeOffsetType SO) {
847 return VisitedBlocks[&BB] = SO;
848 };
849
850 do {
851 Instruction &I = *From;
852
853 if (I.isDebugOrPseudoInst())
854 continue;
855
856 if (++ScannedInstCount > MaxInstsToScan)
857 return Unknown();
858
859 if (!I.mayWriteToMemory())
860 continue;
861
862 if (auto *SI = dyn_cast<StoreInst>(&I)) {
863 AliasResult AR =
864 Options.AA->alias(SI->getPointerOperand(), Load.getPointerOperand());
865 switch ((AliasResult::Kind)AR) {
866 case AliasResult::NoAlias:
867 continue;
868 case AliasResult::MustAlias:
869 if (SI->getValueOperand()->getType()->isPointerTy())
870 return Known(compute(SI->getValueOperand()));
871 else
872 return Unknown(); // No handling of non-pointer values by `compute`.
873 default:
874 return Unknown();
875 }
876 }
877
878 if (auto *CB = dyn_cast<CallBase>(&I)) {
879 Function *Callee = CB->getCalledFunction();
880 // Bail out on indirect call.
881 if (!Callee)
882 return Unknown();
883
884 LibFunc TLIFn;
885 if (!TLI || !TLI->getLibFunc(*CB->getCalledFunction(), TLIFn) ||
886 !TLI->has(TLIFn))
887 return Unknown();
888
889 // TODO: There's probably more interesting case to support here.
890 if (TLIFn != LibFunc_posix_memalign)
891 return Unknown();
892
893 AliasResult AR =
894 Options.AA->alias(CB->getOperand(0), Load.getPointerOperand());
895 switch ((AliasResult::Kind)AR) {
896 case AliasResult::NoAlias:
897 continue;
898 case AliasResult::MustAlias:
899 break;
900 default:
901 return Unknown();
902 }
903
904 // Is the error status of posix_memalign correctly checked? If not it
905 // would be incorrect to assume it succeeds and load doesn't see the
906 // previous value.
907 std::optional<bool> Checked = isImpliedByDomCondition(
908 ICmpInst::ICMP_EQ, CB, ConstantInt::get(CB->getType(), 0), &Load, DL);
909 if (!Checked || !*Checked)
910 return Unknown();
911
912 Value *Size = CB->getOperand(2);
913 auto *C = dyn_cast<ConstantInt>(Size);
914 if (!C)
915 return Unknown();
916
917 return Known({C->getValue(), APInt(C->getValue().getBitWidth(), 0)});
918 }
919
920 return Unknown();
921 } while (From-- != BB.begin());
922
923 SmallVector<SizeOffsetType> PredecessorSizeOffsets;
924 for (auto *PredBB : predecessors(&BB)) {
925 PredecessorSizeOffsets.push_back(findLoadSizeOffset(
926 Load, *PredBB, BasicBlock::iterator(PredBB->getTerminator()),
927 VisitedBlocks, ScannedInstCount));
928 if (!bothKnown(PredecessorSizeOffsets.back()))
929 return Unknown();
930 }
931
932 if (PredecessorSizeOffsets.empty())
933 return Unknown();
934
935 return Known(std::accumulate(PredecessorSizeOffsets.begin() + 1,
936 PredecessorSizeOffsets.end(),
937 PredecessorSizeOffsets.front(),
938 [this](SizeOffsetType LHS, SizeOffsetType RHS) {
939 return combineSizeOffset(LHS, RHS);
940 }));
941 }
942
visitLoadInst(LoadInst & LI)943 SizeOffsetType ObjectSizeOffsetVisitor::visitLoadInst(LoadInst &LI) {
944 if (!Options.AA) {
945 ++ObjectVisitorLoad;
946 return unknown();
947 }
948
949 SmallDenseMap<BasicBlock *, SizeOffsetType, 8> VisitedBlocks;
950 unsigned ScannedInstCount = 0;
951 SizeOffsetType SO =
952 findLoadSizeOffset(LI, *LI.getParent(), BasicBlock::iterator(LI),
953 VisitedBlocks, ScannedInstCount);
954 if (!bothKnown(SO))
955 ++ObjectVisitorLoad;
956 return SO;
957 }
958
combineSizeOffset(SizeOffsetType LHS,SizeOffsetType RHS)959 SizeOffsetType ObjectSizeOffsetVisitor::combineSizeOffset(SizeOffsetType LHS,
960 SizeOffsetType RHS) {
961 if (!bothKnown(LHS) || !bothKnown(RHS))
962 return unknown();
963
964 switch (Options.EvalMode) {
965 case ObjectSizeOpts::Mode::Min:
966 return (getSizeWithOverflow(LHS).slt(getSizeWithOverflow(RHS))) ? LHS : RHS;
967 case ObjectSizeOpts::Mode::Max:
968 return (getSizeWithOverflow(LHS).sgt(getSizeWithOverflow(RHS))) ? LHS : RHS;
969 case ObjectSizeOpts::Mode::ExactSizeFromOffset:
970 return (getSizeWithOverflow(LHS).eq(getSizeWithOverflow(RHS))) ? LHS
971 : unknown();
972 case ObjectSizeOpts::Mode::ExactUnderlyingSizeAndOffset:
973 return LHS == RHS && LHS.second.eq(RHS.second) ? LHS : unknown();
974 }
975 llvm_unreachable("missing an eval mode");
976 }
977
visitPHINode(PHINode & PN)978 SizeOffsetType ObjectSizeOffsetVisitor::visitPHINode(PHINode &PN) {
979 auto IncomingValues = PN.incoming_values();
980 return std::accumulate(IncomingValues.begin() + 1, IncomingValues.end(),
981 compute(*IncomingValues.begin()),
982 [this](SizeOffsetType LHS, Value *VRHS) {
983 return combineSizeOffset(LHS, compute(VRHS));
984 });
985 }
986
visitSelectInst(SelectInst & I)987 SizeOffsetType ObjectSizeOffsetVisitor::visitSelectInst(SelectInst &I) {
988 return combineSizeOffset(compute(I.getTrueValue()),
989 compute(I.getFalseValue()));
990 }
991
visitUndefValue(UndefValue &)992 SizeOffsetType ObjectSizeOffsetVisitor::visitUndefValue(UndefValue&) {
993 return std::make_pair(Zero, Zero);
994 }
995
visitInstruction(Instruction & I)996 SizeOffsetType ObjectSizeOffsetVisitor::visitInstruction(Instruction &I) {
997 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetVisitor unknown instruction:" << I
998 << '\n');
999 return unknown();
1000 }
1001
ObjectSizeOffsetEvaluator(const DataLayout & DL,const TargetLibraryInfo * TLI,LLVMContext & Context,ObjectSizeOpts EvalOpts)1002 ObjectSizeOffsetEvaluator::ObjectSizeOffsetEvaluator(
1003 const DataLayout &DL, const TargetLibraryInfo *TLI, LLVMContext &Context,
1004 ObjectSizeOpts EvalOpts)
1005 : DL(DL), TLI(TLI), Context(Context),
1006 Builder(Context, TargetFolder(DL),
1007 IRBuilderCallbackInserter(
1008 [&](Instruction *I) { InsertedInstructions.insert(I); })),
1009 EvalOpts(EvalOpts) {
1010 // IntTy and Zero must be set for each compute() since the address space may
1011 // be different for later objects.
1012 }
1013
compute(Value * V)1014 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute(Value *V) {
1015 // XXX - Are vectors of pointers possible here?
1016 IntTy = cast<IntegerType>(DL.getIndexType(V->getType()));
1017 Zero = ConstantInt::get(IntTy, 0);
1018
1019 SizeOffsetEvalType Result = compute_(V);
1020
1021 if (!bothKnown(Result)) {
1022 // Erase everything that was computed in this iteration from the cache, so
1023 // that no dangling references are left behind. We could be a bit smarter if
1024 // we kept a dependency graph. It's probably not worth the complexity.
1025 for (const Value *SeenVal : SeenVals) {
1026 CacheMapTy::iterator CacheIt = CacheMap.find(SeenVal);
1027 // non-computable results can be safely cached
1028 if (CacheIt != CacheMap.end() && anyKnown(CacheIt->second))
1029 CacheMap.erase(CacheIt);
1030 }
1031
1032 // Erase any instructions we inserted as part of the traversal.
1033 for (Instruction *I : InsertedInstructions) {
1034 I->replaceAllUsesWith(PoisonValue::get(I->getType()));
1035 I->eraseFromParent();
1036 }
1037 }
1038
1039 SeenVals.clear();
1040 InsertedInstructions.clear();
1041 return Result;
1042 }
1043
compute_(Value * V)1044 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute_(Value *V) {
1045 ObjectSizeOffsetVisitor Visitor(DL, TLI, Context, EvalOpts);
1046 SizeOffsetType Const = Visitor.compute(V);
1047 if (Visitor.bothKnown(Const))
1048 return std::make_pair(ConstantInt::get(Context, Const.first),
1049 ConstantInt::get(Context, Const.second));
1050
1051 V = V->stripPointerCasts();
1052
1053 // Check cache.
1054 CacheMapTy::iterator CacheIt = CacheMap.find(V);
1055 if (CacheIt != CacheMap.end())
1056 return CacheIt->second;
1057
1058 // Always generate code immediately before the instruction being
1059 // processed, so that the generated code dominates the same BBs.
1060 BuilderTy::InsertPointGuard Guard(Builder);
1061 if (Instruction *I = dyn_cast<Instruction>(V))
1062 Builder.SetInsertPoint(I);
1063
1064 // Now compute the size and offset.
1065 SizeOffsetEvalType Result;
1066
1067 // Record the pointers that were handled in this run, so that they can be
1068 // cleaned later if something fails. We also use this set to break cycles that
1069 // can occur in dead code.
1070 if (!SeenVals.insert(V).second) {
1071 Result = unknown();
1072 } else if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
1073 Result = visitGEPOperator(*GEP);
1074 } else if (Instruction *I = dyn_cast<Instruction>(V)) {
1075 Result = visit(*I);
1076 } else if (isa<Argument>(V) ||
1077 (isa<ConstantExpr>(V) &&
1078 cast<ConstantExpr>(V)->getOpcode() == Instruction::IntToPtr) ||
1079 isa<GlobalAlias>(V) ||
1080 isa<GlobalVariable>(V)) {
1081 // Ignore values where we cannot do more than ObjectSizeVisitor.
1082 Result = unknown();
1083 } else {
1084 LLVM_DEBUG(
1085 dbgs() << "ObjectSizeOffsetEvaluator::compute() unhandled value: " << *V
1086 << '\n');
1087 Result = unknown();
1088 }
1089
1090 // Don't reuse CacheIt since it may be invalid at this point.
1091 CacheMap[V] = Result;
1092 return Result;
1093 }
1094
visitAllocaInst(AllocaInst & I)1095 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitAllocaInst(AllocaInst &I) {
1096 if (!I.getAllocatedType()->isSized())
1097 return unknown();
1098
1099 // must be a VLA
1100 assert(I.isArrayAllocation());
1101
1102 // If needed, adjust the alloca's operand size to match the pointer size.
1103 // Subsequent math operations expect the types to match.
1104 Value *ArraySize = Builder.CreateZExtOrTrunc(
1105 I.getArraySize(), DL.getIntPtrType(I.getContext()));
1106 assert(ArraySize->getType() == Zero->getType() &&
1107 "Expected zero constant to have pointer type");
1108
1109 Value *Size = ConstantInt::get(ArraySize->getType(),
1110 DL.getTypeAllocSize(I.getAllocatedType()));
1111 Size = Builder.CreateMul(Size, ArraySize);
1112 return std::make_pair(Size, Zero);
1113 }
1114
visitCallBase(CallBase & CB)1115 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitCallBase(CallBase &CB) {
1116 std::optional<AllocFnsTy> FnData = getAllocationSize(&CB, TLI);
1117 if (!FnData)
1118 return unknown();
1119
1120 // Handle strdup-like functions separately.
1121 if (FnData->AllocTy == StrDupLike) {
1122 // TODO: implement evaluation of strdup/strndup
1123 return unknown();
1124 }
1125
1126 Value *FirstArg = CB.getArgOperand(FnData->FstParam);
1127 FirstArg = Builder.CreateZExtOrTrunc(FirstArg, IntTy);
1128 if (FnData->SndParam < 0)
1129 return std::make_pair(FirstArg, Zero);
1130
1131 Value *SecondArg = CB.getArgOperand(FnData->SndParam);
1132 SecondArg = Builder.CreateZExtOrTrunc(SecondArg, IntTy);
1133 Value *Size = Builder.CreateMul(FirstArg, SecondArg);
1134 return std::make_pair(Size, Zero);
1135 }
1136
1137 SizeOffsetEvalType
visitExtractElementInst(ExtractElementInst &)1138 ObjectSizeOffsetEvaluator::visitExtractElementInst(ExtractElementInst&) {
1139 return unknown();
1140 }
1141
1142 SizeOffsetEvalType
visitExtractValueInst(ExtractValueInst &)1143 ObjectSizeOffsetEvaluator::visitExtractValueInst(ExtractValueInst&) {
1144 return unknown();
1145 }
1146
1147 SizeOffsetEvalType
visitGEPOperator(GEPOperator & GEP)1148 ObjectSizeOffsetEvaluator::visitGEPOperator(GEPOperator &GEP) {
1149 SizeOffsetEvalType PtrData = compute_(GEP.getPointerOperand());
1150 if (!bothKnown(PtrData))
1151 return unknown();
1152
1153 Value *Offset = emitGEPOffset(&Builder, DL, &GEP, /*NoAssumptions=*/true);
1154 Offset = Builder.CreateAdd(PtrData.second, Offset);
1155 return std::make_pair(PtrData.first, Offset);
1156 }
1157
visitIntToPtrInst(IntToPtrInst &)1158 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitIntToPtrInst(IntToPtrInst&) {
1159 // clueless
1160 return unknown();
1161 }
1162
visitLoadInst(LoadInst & LI)1163 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitLoadInst(LoadInst &LI) {
1164 return unknown();
1165 }
1166
visitPHINode(PHINode & PHI)1167 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitPHINode(PHINode &PHI) {
1168 // Create 2 PHIs: one for size and another for offset.
1169 PHINode *SizePHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues());
1170 PHINode *OffsetPHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues());
1171
1172 // Insert right away in the cache to handle recursive PHIs.
1173 CacheMap[&PHI] = std::make_pair(SizePHI, OffsetPHI);
1174
1175 // Compute offset/size for each PHI incoming pointer.
1176 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i) {
1177 Builder.SetInsertPoint(&*PHI.getIncomingBlock(i)->getFirstInsertionPt());
1178 SizeOffsetEvalType EdgeData = compute_(PHI.getIncomingValue(i));
1179
1180 if (!bothKnown(EdgeData)) {
1181 OffsetPHI->replaceAllUsesWith(PoisonValue::get(IntTy));
1182 OffsetPHI->eraseFromParent();
1183 InsertedInstructions.erase(OffsetPHI);
1184 SizePHI->replaceAllUsesWith(PoisonValue::get(IntTy));
1185 SizePHI->eraseFromParent();
1186 InsertedInstructions.erase(SizePHI);
1187 return unknown();
1188 }
1189 SizePHI->addIncoming(EdgeData.first, PHI.getIncomingBlock(i));
1190 OffsetPHI->addIncoming(EdgeData.second, PHI.getIncomingBlock(i));
1191 }
1192
1193 Value *Size = SizePHI, *Offset = OffsetPHI;
1194 if (Value *Tmp = SizePHI->hasConstantValue()) {
1195 Size = Tmp;
1196 SizePHI->replaceAllUsesWith(Size);
1197 SizePHI->eraseFromParent();
1198 InsertedInstructions.erase(SizePHI);
1199 }
1200 if (Value *Tmp = OffsetPHI->hasConstantValue()) {
1201 Offset = Tmp;
1202 OffsetPHI->replaceAllUsesWith(Offset);
1203 OffsetPHI->eraseFromParent();
1204 InsertedInstructions.erase(OffsetPHI);
1205 }
1206 return std::make_pair(Size, Offset);
1207 }
1208
visitSelectInst(SelectInst & I)1209 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitSelectInst(SelectInst &I) {
1210 SizeOffsetEvalType TrueSide = compute_(I.getTrueValue());
1211 SizeOffsetEvalType FalseSide = compute_(I.getFalseValue());
1212
1213 if (!bothKnown(TrueSide) || !bothKnown(FalseSide))
1214 return unknown();
1215 if (TrueSide == FalseSide)
1216 return TrueSide;
1217
1218 Value *Size = Builder.CreateSelect(I.getCondition(), TrueSide.first,
1219 FalseSide.first);
1220 Value *Offset = Builder.CreateSelect(I.getCondition(), TrueSide.second,
1221 FalseSide.second);
1222 return std::make_pair(Size, Offset);
1223 }
1224
visitInstruction(Instruction & I)1225 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitInstruction(Instruction &I) {
1226 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetEvaluator unknown instruction:" << I
1227 << '\n');
1228 return unknown();
1229 }
1230