[InstCombine] Signed saturation patterns
[llvm-core.git] / lib / Analysis / MemoryBuiltins.cpp
blob172c86eb46466454acb28452f5e031c4c4ac32b4
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.
12 //===----------------------------------------------------------------------===//
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/ADT/StringRef.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 <utility>
48 using namespace llvm;
50 #define DEBUG_TYPE "memory-builtins"
52 enum AllocType : uint8_t {
53 OpNewLike = 1<<0, // allocates; never returns null
54 MallocLike = 1<<1 | OpNewLike, // allocates; may return null
55 CallocLike = 1<<2, // allocates + bzero
56 ReallocLike = 1<<3, // reallocates
57 StrDupLike = 1<<4,
58 MallocOrCallocLike = MallocLike | CallocLike,
59 AllocLike = MallocLike | CallocLike | StrDupLike,
60 AnyAlloc = AllocLike | ReallocLike
63 struct AllocFnsTy {
64 AllocType AllocTy;
65 unsigned NumParams;
66 // First and Second size parameters (or -1 if unused)
67 int FstParam, SndParam;
70 // FIXME: certain users need more information. E.g., SimplifyLibCalls needs to
71 // know which functions are nounwind, noalias, nocapture parameters, etc.
72 static const std::pair<LibFunc, AllocFnsTy> AllocationFnData[] = {
73 {LibFunc_malloc, {MallocLike, 1, 0, -1}},
74 {LibFunc_valloc, {MallocLike, 1, 0, -1}},
75 {LibFunc_Znwj, {OpNewLike, 1, 0, -1}}, // new(unsigned int)
76 {LibFunc_ZnwjRKSt9nothrow_t, {MallocLike, 2, 0, -1}}, // new(unsigned int, nothrow)
77 {LibFunc_ZnwjSt11align_val_t, {OpNewLike, 2, 0, -1}}, // new(unsigned int, align_val_t)
78 {LibFunc_ZnwjSt11align_val_tRKSt9nothrow_t, // new(unsigned int, align_val_t, nothrow)
79 {MallocLike, 3, 0, -1}},
80 {LibFunc_Znwm, {OpNewLike, 1, 0, -1}}, // new(unsigned long)
81 {LibFunc_ZnwmRKSt9nothrow_t, {MallocLike, 2, 0, -1}}, // new(unsigned long, nothrow)
82 {LibFunc_ZnwmSt11align_val_t, {OpNewLike, 2, 0, -1}}, // new(unsigned long, align_val_t)
83 {LibFunc_ZnwmSt11align_val_tRKSt9nothrow_t, // new(unsigned long, align_val_t, nothrow)
84 {MallocLike, 3, 0, -1}},
85 {LibFunc_Znaj, {OpNewLike, 1, 0, -1}}, // new[](unsigned int)
86 {LibFunc_ZnajRKSt9nothrow_t, {MallocLike, 2, 0, -1}}, // new[](unsigned int, nothrow)
87 {LibFunc_ZnajSt11align_val_t, {OpNewLike, 2, 0, -1}}, // new[](unsigned int, align_val_t)
88 {LibFunc_ZnajSt11align_val_tRKSt9nothrow_t, // new[](unsigned int, align_val_t, nothrow)
89 {MallocLike, 3, 0, -1}},
90 {LibFunc_Znam, {OpNewLike, 1, 0, -1}}, // new[](unsigned long)
91 {LibFunc_ZnamRKSt9nothrow_t, {MallocLike, 2, 0, -1}}, // new[](unsigned long, nothrow)
92 {LibFunc_ZnamSt11align_val_t, {OpNewLike, 2, 0, -1}}, // new[](unsigned long, align_val_t)
93 {LibFunc_ZnamSt11align_val_tRKSt9nothrow_t, // new[](unsigned long, align_val_t, nothrow)
94 {MallocLike, 3, 0, -1}},
95 {LibFunc_msvc_new_int, {OpNewLike, 1, 0, -1}}, // new(unsigned int)
96 {LibFunc_msvc_new_int_nothrow, {MallocLike, 2, 0, -1}}, // new(unsigned int, nothrow)
97 {LibFunc_msvc_new_longlong, {OpNewLike, 1, 0, -1}}, // new(unsigned long long)
98 {LibFunc_msvc_new_longlong_nothrow, {MallocLike, 2, 0, -1}}, // new(unsigned long long, nothrow)
99 {LibFunc_msvc_new_array_int, {OpNewLike, 1, 0, -1}}, // new[](unsigned int)
100 {LibFunc_msvc_new_array_int_nothrow, {MallocLike, 2, 0, -1}}, // new[](unsigned int, nothrow)
101 {LibFunc_msvc_new_array_longlong, {OpNewLike, 1, 0, -1}}, // new[](unsigned long long)
102 {LibFunc_msvc_new_array_longlong_nothrow, {MallocLike, 2, 0, -1}}, // new[](unsigned long long, nothrow)
103 {LibFunc_calloc, {CallocLike, 2, 0, 1}},
104 {LibFunc_realloc, {ReallocLike, 2, 1, -1}},
105 {LibFunc_reallocf, {ReallocLike, 2, 1, -1}},
106 {LibFunc_strdup, {StrDupLike, 1, -1, -1}},
107 {LibFunc_strndup, {StrDupLike, 2, 1, -1}}
108 // TODO: Handle "int posix_memalign(void **, size_t, size_t)"
111 static const Function *getCalledFunction(const Value *V, bool LookThroughBitCast,
112 bool &IsNoBuiltin) {
113 // Don't care about intrinsics in this case.
114 if (isa<IntrinsicInst>(V))
115 return nullptr;
117 if (LookThroughBitCast)
118 V = V->stripPointerCasts();
120 ImmutableCallSite CS(V);
121 if (!CS.getInstruction())
122 return nullptr;
124 IsNoBuiltin = CS.isNoBuiltin();
126 if (const Function *Callee = CS.getCalledFunction())
127 return Callee;
128 return nullptr;
131 /// Returns the allocation data for the given value if it's either a call to a
132 /// known allocation function, or a call to a function with the allocsize
133 /// attribute.
134 static Optional<AllocFnsTy>
135 getAllocationDataForFunction(const Function *Callee, AllocType AllocTy,
136 const TargetLibraryInfo *TLI) {
137 // Make sure that the function is available.
138 StringRef FnName = Callee->getName();
139 LibFunc TLIFn;
140 if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn))
141 return None;
143 const auto *Iter = find_if(
144 AllocationFnData, [TLIFn](const std::pair<LibFunc, AllocFnsTy> &P) {
145 return P.first == TLIFn;
148 if (Iter == std::end(AllocationFnData))
149 return None;
151 const AllocFnsTy *FnData = &Iter->second;
152 if ((FnData->AllocTy & AllocTy) != FnData->AllocTy)
153 return None;
155 // Check function prototype.
156 int FstParam = FnData->FstParam;
157 int SndParam = FnData->SndParam;
158 FunctionType *FTy = Callee->getFunctionType();
160 if (FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) &&
161 FTy->getNumParams() == FnData->NumParams &&
162 (FstParam < 0 ||
163 (FTy->getParamType(FstParam)->isIntegerTy(32) ||
164 FTy->getParamType(FstParam)->isIntegerTy(64))) &&
165 (SndParam < 0 ||
166 FTy->getParamType(SndParam)->isIntegerTy(32) ||
167 FTy->getParamType(SndParam)->isIntegerTy(64)))
168 return *FnData;
169 return None;
172 static Optional<AllocFnsTy> getAllocationData(const Value *V, AllocType AllocTy,
173 const TargetLibraryInfo *TLI,
174 bool LookThroughBitCast = false) {
175 bool IsNoBuiltinCall;
176 if (const Function *Callee =
177 getCalledFunction(V, LookThroughBitCast, IsNoBuiltinCall))
178 if (!IsNoBuiltinCall)
179 return getAllocationDataForFunction(Callee, AllocTy, TLI);
180 return None;
183 static Optional<AllocFnsTy>
184 getAllocationData(const Value *V, AllocType AllocTy,
185 function_ref<const TargetLibraryInfo &(Function &)> GetTLI,
186 bool LookThroughBitCast = false) {
187 bool IsNoBuiltinCall;
188 if (const Function *Callee =
189 getCalledFunction(V, LookThroughBitCast, IsNoBuiltinCall))
190 if (!IsNoBuiltinCall)
191 return getAllocationDataForFunction(
192 Callee, AllocTy, &GetTLI(const_cast<Function &>(*Callee)));
193 return None;
196 static Optional<AllocFnsTy> getAllocationSize(const Value *V,
197 const TargetLibraryInfo *TLI) {
198 bool IsNoBuiltinCall;
199 const Function *Callee =
200 getCalledFunction(V, /*LookThroughBitCast=*/false, IsNoBuiltinCall);
201 if (!Callee)
202 return None;
204 // Prefer to use existing information over allocsize. This will give us an
205 // accurate AllocTy.
206 if (!IsNoBuiltinCall)
207 if (Optional<AllocFnsTy> Data =
208 getAllocationDataForFunction(Callee, AnyAlloc, TLI))
209 return Data;
211 Attribute Attr = Callee->getFnAttribute(Attribute::AllocSize);
212 if (Attr == Attribute())
213 return None;
215 std::pair<unsigned, Optional<unsigned>> Args = Attr.getAllocSizeArgs();
217 AllocFnsTy Result;
218 // Because allocsize only tells us how many bytes are allocated, we're not
219 // really allowed to assume anything, so we use MallocLike.
220 Result.AllocTy = MallocLike;
221 Result.NumParams = Callee->getNumOperands();
222 Result.FstParam = Args.first;
223 Result.SndParam = Args.second.getValueOr(-1);
224 return Result;
227 static bool hasNoAliasAttr(const Value *V, bool LookThroughBitCast) {
228 ImmutableCallSite CS(LookThroughBitCast ? V->stripPointerCasts() : V);
229 return CS && CS.hasRetAttr(Attribute::NoAlias);
232 /// Tests if a value is a call or invoke to a library function that
233 /// allocates or reallocates memory (either malloc, calloc, realloc, or strdup
234 /// like).
235 bool llvm::isAllocationFn(const Value *V, const TargetLibraryInfo *TLI,
236 bool LookThroughBitCast) {
237 return getAllocationData(V, AnyAlloc, TLI, LookThroughBitCast).hasValue();
239 bool llvm::isAllocationFn(
240 const Value *V, function_ref<const TargetLibraryInfo &(Function &)> GetTLI,
241 bool LookThroughBitCast) {
242 return getAllocationData(V, AnyAlloc, GetTLI, LookThroughBitCast).hasValue();
245 /// Tests if a value is a call or invoke to a function that returns a
246 /// NoAlias pointer (including malloc/calloc/realloc/strdup-like functions).
247 bool llvm::isNoAliasFn(const Value *V, const TargetLibraryInfo *TLI,
248 bool LookThroughBitCast) {
249 // it's safe to consider realloc as noalias since accessing the original
250 // pointer is undefined behavior
251 return isAllocationFn(V, TLI, LookThroughBitCast) ||
252 hasNoAliasAttr(V, LookThroughBitCast);
255 /// Tests if a value is a call or invoke to a library function that
256 /// allocates uninitialized memory (such as malloc).
257 bool llvm::isMallocLikeFn(const Value *V, const TargetLibraryInfo *TLI,
258 bool LookThroughBitCast) {
259 return getAllocationData(V, MallocLike, TLI, LookThroughBitCast).hasValue();
261 bool llvm::isMallocLikeFn(
262 const Value *V, function_ref<const TargetLibraryInfo &(Function &)> GetTLI,
263 bool LookThroughBitCast) {
264 return getAllocationData(V, MallocLike, GetTLI, LookThroughBitCast)
265 .hasValue();
268 /// Tests if a value is a call or invoke to a library function that
269 /// allocates zero-filled memory (such as calloc).
270 bool llvm::isCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI,
271 bool LookThroughBitCast) {
272 return getAllocationData(V, CallocLike, TLI, LookThroughBitCast).hasValue();
275 /// Tests if a value is a call or invoke to a library function that
276 /// allocates memory similar to malloc or calloc.
277 bool llvm::isMallocOrCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI,
278 bool LookThroughBitCast) {
279 return getAllocationData(V, MallocOrCallocLike, TLI,
280 LookThroughBitCast).hasValue();
283 /// Tests if a value is a call or invoke to a library function that
284 /// allocates memory (either malloc, calloc, or strdup like).
285 bool llvm::isAllocLikeFn(const Value *V, const TargetLibraryInfo *TLI,
286 bool LookThroughBitCast) {
287 return getAllocationData(V, AllocLike, TLI, LookThroughBitCast).hasValue();
290 /// Tests if a value is a call or invoke to a library function that
291 /// reallocates memory (e.g., realloc).
292 bool llvm::isReallocLikeFn(const Value *V, const TargetLibraryInfo *TLI,
293 bool LookThroughBitCast) {
294 return getAllocationData(V, ReallocLike, TLI, LookThroughBitCast).hasValue();
297 /// Tests if a functions is a call or invoke to a library function that
298 /// reallocates memory (e.g., realloc).
299 bool llvm::isReallocLikeFn(const Function *F, const TargetLibraryInfo *TLI) {
300 return getAllocationDataForFunction(F, ReallocLike, TLI).hasValue();
303 /// Tests if a value is a call or invoke to a library function that
304 /// allocates memory and throws if an allocation failed (e.g., new).
305 bool llvm::isOpNewLikeFn(const Value *V, const TargetLibraryInfo *TLI,
306 bool LookThroughBitCast) {
307 return getAllocationData(V, OpNewLike, TLI, LookThroughBitCast).hasValue();
310 /// Tests if a value is a call or invoke to a library function that
311 /// allocates memory (strdup, strndup).
312 bool llvm::isStrdupLikeFn(const Value *V, const TargetLibraryInfo *TLI,
313 bool LookThroughBitCast) {
314 return getAllocationData(V, StrDupLike, TLI, LookThroughBitCast).hasValue();
317 /// extractMallocCall - Returns the corresponding CallInst if the instruction
318 /// is a malloc call. Since CallInst::CreateMalloc() only creates calls, we
319 /// ignore InvokeInst here.
320 const CallInst *llvm::extractMallocCall(
321 const Value *I,
322 function_ref<const TargetLibraryInfo &(Function &)> GetTLI) {
323 return isMallocLikeFn(I, GetTLI) ? dyn_cast<CallInst>(I) : nullptr;
326 static Value *computeArraySize(const CallInst *CI, const DataLayout &DL,
327 const TargetLibraryInfo *TLI,
328 bool LookThroughSExt = false) {
329 if (!CI)
330 return nullptr;
332 // The size of the malloc's result type must be known to determine array size.
333 Type *T = getMallocAllocatedType(CI, TLI);
334 if (!T || !T->isSized())
335 return nullptr;
337 unsigned ElementSize = DL.getTypeAllocSize(T);
338 if (StructType *ST = dyn_cast<StructType>(T))
339 ElementSize = DL.getStructLayout(ST)->getSizeInBytes();
341 // If malloc call's arg can be determined to be a multiple of ElementSize,
342 // return the multiple. Otherwise, return NULL.
343 Value *MallocArg = CI->getArgOperand(0);
344 Value *Multiple = nullptr;
345 if (ComputeMultiple(MallocArg, ElementSize, Multiple, LookThroughSExt))
346 return Multiple;
348 return nullptr;
351 /// getMallocType - Returns the PointerType resulting from the malloc call.
352 /// The PointerType depends on the number of bitcast uses of the malloc call:
353 /// 0: PointerType is the calls' return type.
354 /// 1: PointerType is the bitcast's result type.
355 /// >1: Unique PointerType cannot be determined, return NULL.
356 PointerType *llvm::getMallocType(const CallInst *CI,
357 const TargetLibraryInfo *TLI) {
358 assert(isMallocLikeFn(CI, TLI) && "getMallocType and not malloc call");
360 PointerType *MallocType = nullptr;
361 unsigned NumOfBitCastUses = 0;
363 // Determine if CallInst has a bitcast use.
364 for (Value::const_user_iterator UI = CI->user_begin(), E = CI->user_end();
365 UI != E;)
366 if (const BitCastInst *BCI = dyn_cast<BitCastInst>(*UI++)) {
367 MallocType = cast<PointerType>(BCI->getDestTy());
368 NumOfBitCastUses++;
371 // Malloc call has 1 bitcast use, so type is the bitcast's destination type.
372 if (NumOfBitCastUses == 1)
373 return MallocType;
375 // Malloc call was not bitcast, so type is the malloc function's return type.
376 if (NumOfBitCastUses == 0)
377 return cast<PointerType>(CI->getType());
379 // Type could not be determined.
380 return nullptr;
383 /// getMallocAllocatedType - Returns the Type allocated by malloc call.
384 /// The Type depends on the number of bitcast uses of the malloc call:
385 /// 0: PointerType is the malloc calls' return type.
386 /// 1: PointerType is the bitcast's result type.
387 /// >1: Unique PointerType cannot be determined, return NULL.
388 Type *llvm::getMallocAllocatedType(const CallInst *CI,
389 const TargetLibraryInfo *TLI) {
390 PointerType *PT = getMallocType(CI, TLI);
391 return PT ? PT->getElementType() : nullptr;
394 /// getMallocArraySize - Returns the array size of a malloc call. If the
395 /// argument passed to malloc is a multiple of the size of the malloced type,
396 /// then return that multiple. For non-array mallocs, the multiple is
397 /// constant 1. Otherwise, return NULL for mallocs whose array size cannot be
398 /// determined.
399 Value *llvm::getMallocArraySize(CallInst *CI, const DataLayout &DL,
400 const TargetLibraryInfo *TLI,
401 bool LookThroughSExt) {
402 assert(isMallocLikeFn(CI, TLI) && "getMallocArraySize and not malloc call");
403 return computeArraySize(CI, DL, TLI, LookThroughSExt);
406 /// extractCallocCall - Returns the corresponding CallInst if the instruction
407 /// is a calloc call.
408 const CallInst *llvm::extractCallocCall(const Value *I,
409 const TargetLibraryInfo *TLI) {
410 return isCallocLikeFn(I, TLI) ? cast<CallInst>(I) : nullptr;
413 /// isLibFreeFunction - Returns true if the function is a builtin free()
414 bool llvm::isLibFreeFunction(const Function *F, const LibFunc TLIFn) {
415 unsigned ExpectedNumParams;
416 if (TLIFn == LibFunc_free ||
417 TLIFn == LibFunc_ZdlPv || // operator delete(void*)
418 TLIFn == LibFunc_ZdaPv || // operator delete[](void*)
419 TLIFn == LibFunc_msvc_delete_ptr32 || // operator delete(void*)
420 TLIFn == LibFunc_msvc_delete_ptr64 || // operator delete(void*)
421 TLIFn == LibFunc_msvc_delete_array_ptr32 || // operator delete[](void*)
422 TLIFn == LibFunc_msvc_delete_array_ptr64) // operator delete[](void*)
423 ExpectedNumParams = 1;
424 else if (TLIFn == LibFunc_ZdlPvj || // delete(void*, uint)
425 TLIFn == LibFunc_ZdlPvm || // delete(void*, ulong)
426 TLIFn == LibFunc_ZdlPvRKSt9nothrow_t || // delete(void*, nothrow)
427 TLIFn == LibFunc_ZdlPvSt11align_val_t || // delete(void*, align_val_t)
428 TLIFn == LibFunc_ZdaPvj || // delete[](void*, uint)
429 TLIFn == LibFunc_ZdaPvm || // delete[](void*, ulong)
430 TLIFn == LibFunc_ZdaPvRKSt9nothrow_t || // delete[](void*, nothrow)
431 TLIFn == LibFunc_ZdaPvSt11align_val_t || // delete[](void*, align_val_t)
432 TLIFn == LibFunc_msvc_delete_ptr32_int || // delete(void*, uint)
433 TLIFn == LibFunc_msvc_delete_ptr64_longlong || // delete(void*, ulonglong)
434 TLIFn == LibFunc_msvc_delete_ptr32_nothrow || // delete(void*, nothrow)
435 TLIFn == LibFunc_msvc_delete_ptr64_nothrow || // delete(void*, nothrow)
436 TLIFn == LibFunc_msvc_delete_array_ptr32_int || // delete[](void*, uint)
437 TLIFn == LibFunc_msvc_delete_array_ptr64_longlong || // delete[](void*, ulonglong)
438 TLIFn == LibFunc_msvc_delete_array_ptr32_nothrow || // delete[](void*, nothrow)
439 TLIFn == LibFunc_msvc_delete_array_ptr64_nothrow) // delete[](void*, nothrow)
440 ExpectedNumParams = 2;
441 else if (TLIFn == LibFunc_ZdaPvSt11align_val_tRKSt9nothrow_t || // delete(void*, align_val_t, nothrow)
442 TLIFn == LibFunc_ZdlPvSt11align_val_tRKSt9nothrow_t) // delete[](void*, align_val_t, nothrow)
443 ExpectedNumParams = 3;
444 else
445 return false;
447 // Check free prototype.
448 // FIXME: workaround for PR5130, this will be obsolete when a nobuiltin
449 // attribute will exist.
450 FunctionType *FTy = F->getFunctionType();
451 if (!FTy->getReturnType()->isVoidTy())
452 return false;
453 if (FTy->getNumParams() != ExpectedNumParams)
454 return false;
455 if (FTy->getParamType(0) != Type::getInt8PtrTy(F->getContext()))
456 return false;
458 return true;
461 /// isFreeCall - Returns non-null if the value is a call to the builtin free()
462 const CallInst *llvm::isFreeCall(const Value *I, const TargetLibraryInfo *TLI) {
463 bool IsNoBuiltinCall;
464 const Function *Callee =
465 getCalledFunction(I, /*LookThroughBitCast=*/false, IsNoBuiltinCall);
466 if (Callee == nullptr || IsNoBuiltinCall)
467 return nullptr;
469 StringRef FnName = Callee->getName();
470 LibFunc TLIFn;
471 if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn))
472 return nullptr;
474 return isLibFreeFunction(Callee, TLIFn) ? dyn_cast<CallInst>(I) : nullptr;
478 //===----------------------------------------------------------------------===//
479 // Utility functions to compute size of objects.
481 static APInt getSizeWithOverflow(const SizeOffsetType &Data) {
482 if (Data.second.isNegative() || Data.first.ult(Data.second))
483 return APInt(Data.first.getBitWidth(), 0);
484 return Data.first - Data.second;
487 /// Compute the size of the object pointed by Ptr. Returns true and the
488 /// object size in Size if successful, and false otherwise.
489 /// If RoundToAlign is true, then Size is rounded up to the alignment of
490 /// allocas, byval arguments, and global variables.
491 bool llvm::getObjectSize(const Value *Ptr, uint64_t &Size, const DataLayout &DL,
492 const TargetLibraryInfo *TLI, ObjectSizeOpts Opts) {
493 ObjectSizeOffsetVisitor Visitor(DL, TLI, Ptr->getContext(), Opts);
494 SizeOffsetType Data = Visitor.compute(const_cast<Value*>(Ptr));
495 if (!Visitor.bothKnown(Data))
496 return false;
498 Size = getSizeWithOverflow(Data).getZExtValue();
499 return true;
502 Value *llvm::lowerObjectSizeCall(IntrinsicInst *ObjectSize,
503 const DataLayout &DL,
504 const TargetLibraryInfo *TLI,
505 bool MustSucceed) {
506 assert(ObjectSize->getIntrinsicID() == Intrinsic::objectsize &&
507 "ObjectSize must be a call to llvm.objectsize!");
509 bool MaxVal = cast<ConstantInt>(ObjectSize->getArgOperand(1))->isZero();
510 ObjectSizeOpts EvalOptions;
511 // Unless we have to fold this to something, try to be as accurate as
512 // possible.
513 if (MustSucceed)
514 EvalOptions.EvalMode =
515 MaxVal ? ObjectSizeOpts::Mode::Max : ObjectSizeOpts::Mode::Min;
516 else
517 EvalOptions.EvalMode = ObjectSizeOpts::Mode::Exact;
519 EvalOptions.NullIsUnknownSize =
520 cast<ConstantInt>(ObjectSize->getArgOperand(2))->isOne();
522 auto *ResultType = cast<IntegerType>(ObjectSize->getType());
523 bool StaticOnly = cast<ConstantInt>(ObjectSize->getArgOperand(3))->isZero();
524 if (StaticOnly) {
525 // FIXME: Does it make sense to just return a failure value if the size won't
526 // fit in the output and `!MustSucceed`?
527 uint64_t Size;
528 if (getObjectSize(ObjectSize->getArgOperand(0), Size, DL, TLI, EvalOptions) &&
529 isUIntN(ResultType->getBitWidth(), Size))
530 return ConstantInt::get(ResultType, Size);
531 } else {
532 LLVMContext &Ctx = ObjectSize->getFunction()->getContext();
533 ObjectSizeOffsetEvaluator Eval(DL, TLI, Ctx, EvalOptions);
534 SizeOffsetEvalType SizeOffsetPair =
535 Eval.compute(ObjectSize->getArgOperand(0));
537 if (SizeOffsetPair != ObjectSizeOffsetEvaluator::unknown()) {
538 IRBuilder<TargetFolder> Builder(Ctx, TargetFolder(DL));
539 Builder.SetInsertPoint(ObjectSize);
541 // If we've outside the end of the object, then we can always access
542 // exactly 0 bytes.
543 Value *ResultSize =
544 Builder.CreateSub(SizeOffsetPair.first, SizeOffsetPair.second);
545 Value *UseZero =
546 Builder.CreateICmpULT(SizeOffsetPair.first, SizeOffsetPair.second);
547 return Builder.CreateSelect(UseZero, ConstantInt::get(ResultType, 0),
548 ResultSize);
552 if (!MustSucceed)
553 return nullptr;
555 return ConstantInt::get(ResultType, MaxVal ? -1ULL : 0);
558 STATISTIC(ObjectVisitorArgument,
559 "Number of arguments with unsolved size and offset");
560 STATISTIC(ObjectVisitorLoad,
561 "Number of load instructions with unsolved size and offset");
563 APInt ObjectSizeOffsetVisitor::align(APInt Size, uint64_t Alignment) {
564 if (Options.RoundToAlign && Alignment)
565 return APInt(IntTyBits, alignTo(Size.getZExtValue(), Align(Alignment)));
566 return Size;
569 ObjectSizeOffsetVisitor::ObjectSizeOffsetVisitor(const DataLayout &DL,
570 const TargetLibraryInfo *TLI,
571 LLVMContext &Context,
572 ObjectSizeOpts Options)
573 : DL(DL), TLI(TLI), Options(Options) {
574 // Pointer size must be rechecked for each object visited since it could have
575 // a different address space.
578 SizeOffsetType ObjectSizeOffsetVisitor::compute(Value *V) {
579 IntTyBits = DL.getPointerTypeSizeInBits(V->getType());
580 Zero = APInt::getNullValue(IntTyBits);
582 V = V->stripPointerCasts();
583 if (Instruction *I = dyn_cast<Instruction>(V)) {
584 // If we have already seen this instruction, bail out. Cycles can happen in
585 // unreachable code after constant propagation.
586 if (!SeenInsts.insert(I).second)
587 return unknown();
589 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
590 return visitGEPOperator(*GEP);
591 return visit(*I);
593 if (Argument *A = dyn_cast<Argument>(V))
594 return visitArgument(*A);
595 if (ConstantPointerNull *P = dyn_cast<ConstantPointerNull>(V))
596 return visitConstantPointerNull(*P);
597 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
598 return visitGlobalAlias(*GA);
599 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
600 return visitGlobalVariable(*GV);
601 if (UndefValue *UV = dyn_cast<UndefValue>(V))
602 return visitUndefValue(*UV);
603 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
604 if (CE->getOpcode() == Instruction::IntToPtr)
605 return unknown(); // clueless
606 if (CE->getOpcode() == Instruction::GetElementPtr)
607 return visitGEPOperator(cast<GEPOperator>(*CE));
610 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetVisitor::compute() unhandled value: "
611 << *V << '\n');
612 return unknown();
615 /// When we're compiling N-bit code, and the user uses parameters that are
616 /// greater than N bits (e.g. uint64_t on a 32-bit build), we can run into
617 /// trouble with APInt size issues. This function handles resizing + overflow
618 /// checks for us. Check and zext or trunc \p I depending on IntTyBits and
619 /// I's value.
620 bool ObjectSizeOffsetVisitor::CheckedZextOrTrunc(APInt &I) {
621 // More bits than we can handle. Checking the bit width isn't necessary, but
622 // it's faster than checking active bits, and should give `false` in the
623 // vast majority of cases.
624 if (I.getBitWidth() > IntTyBits && I.getActiveBits() > IntTyBits)
625 return false;
626 if (I.getBitWidth() != IntTyBits)
627 I = I.zextOrTrunc(IntTyBits);
628 return true;
631 SizeOffsetType ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) {
632 if (!I.getAllocatedType()->isSized())
633 return unknown();
635 APInt Size(IntTyBits, DL.getTypeAllocSize(I.getAllocatedType()));
636 if (!I.isArrayAllocation())
637 return std::make_pair(align(Size, I.getAlignment()), Zero);
639 Value *ArraySize = I.getArraySize();
640 if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) {
641 APInt NumElems = C->getValue();
642 if (!CheckedZextOrTrunc(NumElems))
643 return unknown();
645 bool Overflow;
646 Size = Size.umul_ov(NumElems, Overflow);
647 return Overflow ? unknown() : std::make_pair(align(Size, I.getAlignment()),
648 Zero);
650 return unknown();
653 SizeOffsetType ObjectSizeOffsetVisitor::visitArgument(Argument &A) {
654 // No interprocedural analysis is done at the moment.
655 if (!A.hasByValOrInAllocaAttr()) {
656 ++ObjectVisitorArgument;
657 return unknown();
659 PointerType *PT = cast<PointerType>(A.getType());
660 APInt Size(IntTyBits, DL.getTypeAllocSize(PT->getElementType()));
661 return std::make_pair(align(Size, A.getParamAlignment()), Zero);
664 SizeOffsetType ObjectSizeOffsetVisitor::visitCallSite(CallSite CS) {
665 Optional<AllocFnsTy> FnData = getAllocationSize(CS.getInstruction(), TLI);
666 if (!FnData)
667 return unknown();
669 // Handle strdup-like functions separately.
670 if (FnData->AllocTy == StrDupLike) {
671 APInt Size(IntTyBits, GetStringLength(CS.getArgument(0)));
672 if (!Size)
673 return unknown();
675 // Strndup limits strlen.
676 if (FnData->FstParam > 0) {
677 ConstantInt *Arg =
678 dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam));
679 if (!Arg)
680 return unknown();
682 APInt MaxSize = Arg->getValue().zextOrSelf(IntTyBits);
683 if (Size.ugt(MaxSize))
684 Size = MaxSize + 1;
686 return std::make_pair(Size, Zero);
689 ConstantInt *Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam));
690 if (!Arg)
691 return unknown();
693 APInt Size = Arg->getValue();
694 if (!CheckedZextOrTrunc(Size))
695 return unknown();
697 // Size is determined by just 1 parameter.
698 if (FnData->SndParam < 0)
699 return std::make_pair(Size, Zero);
701 Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->SndParam));
702 if (!Arg)
703 return unknown();
705 APInt NumElems = Arg->getValue();
706 if (!CheckedZextOrTrunc(NumElems))
707 return unknown();
709 bool Overflow;
710 Size = Size.umul_ov(NumElems, Overflow);
711 return Overflow ? unknown() : std::make_pair(Size, Zero);
713 // TODO: handle more standard functions (+ wchar cousins):
714 // - strdup / strndup
715 // - strcpy / strncpy
716 // - strcat / strncat
717 // - memcpy / memmove
718 // - strcat / strncat
719 // - memset
722 SizeOffsetType
723 ObjectSizeOffsetVisitor::visitConstantPointerNull(ConstantPointerNull& CPN) {
724 // If null is unknown, there's nothing we can do. Additionally, non-zero
725 // address spaces can make use of null, so we don't presume to know anything
726 // about that.
728 // TODO: How should this work with address space casts? We currently just drop
729 // them on the floor, but it's unclear what we should do when a NULL from
730 // addrspace(1) gets casted to addrspace(0) (or vice-versa).
731 if (Options.NullIsUnknownSize || CPN.getType()->getAddressSpace())
732 return unknown();
733 return std::make_pair(Zero, Zero);
736 SizeOffsetType
737 ObjectSizeOffsetVisitor::visitExtractElementInst(ExtractElementInst&) {
738 return unknown();
741 SizeOffsetType
742 ObjectSizeOffsetVisitor::visitExtractValueInst(ExtractValueInst&) {
743 // Easy cases were already folded by previous passes.
744 return unknown();
747 SizeOffsetType ObjectSizeOffsetVisitor::visitGEPOperator(GEPOperator &GEP) {
748 SizeOffsetType PtrData = compute(GEP.getPointerOperand());
749 APInt Offset(IntTyBits, 0);
750 if (!bothKnown(PtrData) || !GEP.accumulateConstantOffset(DL, Offset))
751 return unknown();
753 return std::make_pair(PtrData.first, PtrData.second + Offset);
756 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalAlias(GlobalAlias &GA) {
757 if (GA.isInterposable())
758 return unknown();
759 return compute(GA.getAliasee());
762 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalVariable(GlobalVariable &GV){
763 if (!GV.hasDefinitiveInitializer())
764 return unknown();
766 APInt Size(IntTyBits, DL.getTypeAllocSize(GV.getValueType()));
767 return std::make_pair(align(Size, GV.getAlignment()), Zero);
770 SizeOffsetType ObjectSizeOffsetVisitor::visitIntToPtrInst(IntToPtrInst&) {
771 // clueless
772 return unknown();
775 SizeOffsetType ObjectSizeOffsetVisitor::visitLoadInst(LoadInst&) {
776 ++ObjectVisitorLoad;
777 return unknown();
780 SizeOffsetType ObjectSizeOffsetVisitor::visitPHINode(PHINode&) {
781 // too complex to analyze statically.
782 return unknown();
785 SizeOffsetType ObjectSizeOffsetVisitor::visitSelectInst(SelectInst &I) {
786 SizeOffsetType TrueSide = compute(I.getTrueValue());
787 SizeOffsetType FalseSide = compute(I.getFalseValue());
788 if (bothKnown(TrueSide) && bothKnown(FalseSide)) {
789 if (TrueSide == FalseSide) {
790 return TrueSide;
793 APInt TrueResult = getSizeWithOverflow(TrueSide);
794 APInt FalseResult = getSizeWithOverflow(FalseSide);
796 if (TrueResult == FalseResult) {
797 return TrueSide;
799 if (Options.EvalMode == ObjectSizeOpts::Mode::Min) {
800 if (TrueResult.slt(FalseResult))
801 return TrueSide;
802 return FalseSide;
804 if (Options.EvalMode == ObjectSizeOpts::Mode::Max) {
805 if (TrueResult.sgt(FalseResult))
806 return TrueSide;
807 return FalseSide;
810 return unknown();
813 SizeOffsetType ObjectSizeOffsetVisitor::visitUndefValue(UndefValue&) {
814 return std::make_pair(Zero, Zero);
817 SizeOffsetType ObjectSizeOffsetVisitor::visitInstruction(Instruction &I) {
818 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetVisitor unknown instruction:" << I
819 << '\n');
820 return unknown();
823 ObjectSizeOffsetEvaluator::ObjectSizeOffsetEvaluator(
824 const DataLayout &DL, const TargetLibraryInfo *TLI, LLVMContext &Context,
825 ObjectSizeOpts EvalOpts)
826 : DL(DL), TLI(TLI), Context(Context),
827 Builder(Context, TargetFolder(DL),
828 IRBuilderCallbackInserter(
829 [&](Instruction *I) { InsertedInstructions.insert(I); })),
830 EvalOpts(EvalOpts) {
831 // IntTy and Zero must be set for each compute() since the address space may
832 // be different for later objects.
835 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute(Value *V) {
836 // XXX - Are vectors of pointers possible here?
837 IntTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
838 Zero = ConstantInt::get(IntTy, 0);
840 SizeOffsetEvalType Result = compute_(V);
842 if (!bothKnown(Result)) {
843 // Erase everything that was computed in this iteration from the cache, so
844 // that no dangling references are left behind. We could be a bit smarter if
845 // we kept a dependency graph. It's probably not worth the complexity.
846 for (const Value *SeenVal : SeenVals) {
847 CacheMapTy::iterator CacheIt = CacheMap.find(SeenVal);
848 // non-computable results can be safely cached
849 if (CacheIt != CacheMap.end() && anyKnown(CacheIt->second))
850 CacheMap.erase(CacheIt);
853 // Erase any instructions we inserted as part of the traversal.
854 for (Instruction *I : InsertedInstructions) {
855 I->replaceAllUsesWith(UndefValue::get(I->getType()));
856 I->eraseFromParent();
860 SeenVals.clear();
861 InsertedInstructions.clear();
862 return Result;
865 SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute_(Value *V) {
866 ObjectSizeOffsetVisitor Visitor(DL, TLI, Context, EvalOpts);
867 SizeOffsetType Const = Visitor.compute(V);
868 if (Visitor.bothKnown(Const))
869 return std::make_pair(ConstantInt::get(Context, Const.first),
870 ConstantInt::get(Context, Const.second));
872 V = V->stripPointerCasts();
874 // Check cache.
875 CacheMapTy::iterator CacheIt = CacheMap.find(V);
876 if (CacheIt != CacheMap.end())
877 return CacheIt->second;
879 // Always generate code immediately before the instruction being
880 // processed, so that the generated code dominates the same BBs.
881 BuilderTy::InsertPointGuard Guard(Builder);
882 if (Instruction *I = dyn_cast<Instruction>(V))
883 Builder.SetInsertPoint(I);
885 // Now compute the size and offset.
886 SizeOffsetEvalType Result;
888 // Record the pointers that were handled in this run, so that they can be
889 // cleaned later if something fails. We also use this set to break cycles that
890 // can occur in dead code.
891 if (!SeenVals.insert(V).second) {
892 Result = unknown();
893 } else if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
894 Result = visitGEPOperator(*GEP);
895 } else if (Instruction *I = dyn_cast<Instruction>(V)) {
896 Result = visit(*I);
897 } else if (isa<Argument>(V) ||
898 (isa<ConstantExpr>(V) &&
899 cast<ConstantExpr>(V)->getOpcode() == Instruction::IntToPtr) ||
900 isa<GlobalAlias>(V) ||
901 isa<GlobalVariable>(V)) {
902 // Ignore values where we cannot do more than ObjectSizeVisitor.
903 Result = unknown();
904 } else {
905 LLVM_DEBUG(
906 dbgs() << "ObjectSizeOffsetEvaluator::compute() unhandled value: " << *V
907 << '\n');
908 Result = unknown();
911 // Don't reuse CacheIt since it may be invalid at this point.
912 CacheMap[V] = Result;
913 return Result;
916 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitAllocaInst(AllocaInst &I) {
917 if (!I.getAllocatedType()->isSized())
918 return unknown();
920 // must be a VLA
921 assert(I.isArrayAllocation());
922 Value *ArraySize = I.getArraySize();
923 Value *Size = ConstantInt::get(ArraySize->getType(),
924 DL.getTypeAllocSize(I.getAllocatedType()));
925 Size = Builder.CreateMul(Size, ArraySize);
926 return std::make_pair(Size, Zero);
929 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitCallSite(CallSite CS) {
930 Optional<AllocFnsTy> FnData = getAllocationSize(CS.getInstruction(), TLI);
931 if (!FnData)
932 return unknown();
934 // Handle strdup-like functions separately.
935 if (FnData->AllocTy == StrDupLike) {
936 // TODO
937 return unknown();
940 Value *FirstArg = CS.getArgument(FnData->FstParam);
941 FirstArg = Builder.CreateZExt(FirstArg, IntTy);
942 if (FnData->SndParam < 0)
943 return std::make_pair(FirstArg, Zero);
945 Value *SecondArg = CS.getArgument(FnData->SndParam);
946 SecondArg = Builder.CreateZExt(SecondArg, IntTy);
947 Value *Size = Builder.CreateMul(FirstArg, SecondArg);
948 return std::make_pair(Size, Zero);
950 // TODO: handle more standard functions (+ wchar cousins):
951 // - strdup / strndup
952 // - strcpy / strncpy
953 // - strcat / strncat
954 // - memcpy / memmove
955 // - strcat / strncat
956 // - memset
959 SizeOffsetEvalType
960 ObjectSizeOffsetEvaluator::visitExtractElementInst(ExtractElementInst&) {
961 return unknown();
964 SizeOffsetEvalType
965 ObjectSizeOffsetEvaluator::visitExtractValueInst(ExtractValueInst&) {
966 return unknown();
969 SizeOffsetEvalType
970 ObjectSizeOffsetEvaluator::visitGEPOperator(GEPOperator &GEP) {
971 SizeOffsetEvalType PtrData = compute_(GEP.getPointerOperand());
972 if (!bothKnown(PtrData))
973 return unknown();
975 Value *Offset = EmitGEPOffset(&Builder, DL, &GEP, /*NoAssumptions=*/true);
976 Offset = Builder.CreateAdd(PtrData.second, Offset);
977 return std::make_pair(PtrData.first, Offset);
980 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitIntToPtrInst(IntToPtrInst&) {
981 // clueless
982 return unknown();
985 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitLoadInst(LoadInst&) {
986 return unknown();
989 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitPHINode(PHINode &PHI) {
990 // Create 2 PHIs: one for size and another for offset.
991 PHINode *SizePHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues());
992 PHINode *OffsetPHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues());
994 // Insert right away in the cache to handle recursive PHIs.
995 CacheMap[&PHI] = std::make_pair(SizePHI, OffsetPHI);
997 // Compute offset/size for each PHI incoming pointer.
998 for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i) {
999 Builder.SetInsertPoint(&*PHI.getIncomingBlock(i)->getFirstInsertionPt());
1000 SizeOffsetEvalType EdgeData = compute_(PHI.getIncomingValue(i));
1002 if (!bothKnown(EdgeData)) {
1003 OffsetPHI->replaceAllUsesWith(UndefValue::get(IntTy));
1004 OffsetPHI->eraseFromParent();
1005 InsertedInstructions.erase(OffsetPHI);
1006 SizePHI->replaceAllUsesWith(UndefValue::get(IntTy));
1007 SizePHI->eraseFromParent();
1008 InsertedInstructions.erase(SizePHI);
1009 return unknown();
1011 SizePHI->addIncoming(EdgeData.first, PHI.getIncomingBlock(i));
1012 OffsetPHI->addIncoming(EdgeData.second, PHI.getIncomingBlock(i));
1015 Value *Size = SizePHI, *Offset = OffsetPHI;
1016 if (Value *Tmp = SizePHI->hasConstantValue()) {
1017 Size = Tmp;
1018 SizePHI->replaceAllUsesWith(Size);
1019 SizePHI->eraseFromParent();
1020 InsertedInstructions.erase(SizePHI);
1022 if (Value *Tmp = OffsetPHI->hasConstantValue()) {
1023 Offset = Tmp;
1024 OffsetPHI->replaceAllUsesWith(Offset);
1025 OffsetPHI->eraseFromParent();
1026 InsertedInstructions.erase(OffsetPHI);
1028 return std::make_pair(Size, Offset);
1031 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitSelectInst(SelectInst &I) {
1032 SizeOffsetEvalType TrueSide = compute_(I.getTrueValue());
1033 SizeOffsetEvalType FalseSide = compute_(I.getFalseValue());
1035 if (!bothKnown(TrueSide) || !bothKnown(FalseSide))
1036 return unknown();
1037 if (TrueSide == FalseSide)
1038 return TrueSide;
1040 Value *Size = Builder.CreateSelect(I.getCondition(), TrueSide.first,
1041 FalseSide.first);
1042 Value *Offset = Builder.CreateSelect(I.getCondition(), TrueSide.second,
1043 FalseSide.second);
1044 return std::make_pair(Size, Offset);
1047 SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitInstruction(Instruction &I) {
1048 LLVM_DEBUG(dbgs() << "ObjectSizeOffsetEvaluator unknown instruction:" << I
1049 << '\n');
1050 return unknown();