1//===------ MemoryBuiltins.cpp - Identify calls to memory builtins --------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This family of functions identifies calls to builtin functions that allocate
11// or free memory.
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/Analysis/MemoryBuiltins.h"
16#include "llvm/ADT/STLExtras.h"
17#include "llvm/ADT/Statistic.h"
18#include "llvm/Analysis/ValueTracking.h"
19#include "llvm/IR/DataLayout.h"
20#include "llvm/IR/GlobalVariable.h"
21#include "llvm/IR/Instructions.h"
22#include "llvm/IR/Intrinsics.h"
23#include "llvm/IR/Metadata.h"
24#include "llvm/IR/Module.h"
25#include "llvm/Support/Debug.h"
26#include "llvm/Support/MathExtras.h"
27#include "llvm/Support/raw_ostream.h"
28#include "llvm/Target/TargetLibraryInfo.h"
29#include "llvm/Transforms/Utils/Local.h"
30using namespace llvm;
31
32#define DEBUG_TYPE "memory-builtins"
33
34enum AllocType {
35  OpNewLike          = 1<<0, // allocates; never returns null
36  MallocLike         = 1<<1 | OpNewLike, // allocates; may return null
37  CallocLike         = 1<<2, // allocates + bzero
38  ReallocLike        = 1<<3, // reallocates
39  StrDupLike         = 1<<4,
40  AllocLike          = MallocLike | CallocLike | StrDupLike,
41  AnyAlloc           = AllocLike | ReallocLike
42};
43
44struct AllocFnsTy {
45  LibFunc::Func Func;
46  AllocType AllocTy;
47  unsigned char NumParams;
48  // First and Second size parameters (or -1 if unused)
49  signed char FstParam, SndParam;
50};
51
52// FIXME: certain users need more information. E.g., SimplifyLibCalls needs to
53// know which functions are nounwind, noalias, nocapture parameters, etc.
54static const AllocFnsTy AllocationFnData[] = {
55  {LibFunc::malloc,              MallocLike,  1, 0,  -1},
56  {LibFunc::valloc,              MallocLike,  1, 0,  -1},
57  {LibFunc::Znwj,                OpNewLike,   1, 0,  -1}, // new(unsigned int)
58  {LibFunc::ZnwjRKSt9nothrow_t,  MallocLike,  2, 0,  -1}, // new(unsigned int, nothrow)
59  {LibFunc::Znwm,                OpNewLike,   1, 0,  -1}, // new(unsigned long)
60  {LibFunc::ZnwmRKSt9nothrow_t,  MallocLike,  2, 0,  -1}, // new(unsigned long, nothrow)
61  {LibFunc::Znaj,                OpNewLike,   1, 0,  -1}, // new[](unsigned int)
62  {LibFunc::ZnajRKSt9nothrow_t,  MallocLike,  2, 0,  -1}, // new[](unsigned int, nothrow)
63  {LibFunc::Znam,                OpNewLike,   1, 0,  -1}, // new[](unsigned long)
64  {LibFunc::ZnamRKSt9nothrow_t,  MallocLike,  2, 0,  -1}, // new[](unsigned long, nothrow)
65  {LibFunc::calloc,              CallocLike,  2, 0,   1},
66  {LibFunc::realloc,             ReallocLike, 2, 1,  -1},
67  {LibFunc::reallocf,            ReallocLike, 2, 1,  -1},
68  {LibFunc::strdup,              StrDupLike,  1, -1, -1},
69  {LibFunc::strndup,             StrDupLike,  2, 1,  -1}
70  // TODO: Handle "int posix_memalign(void **, size_t, size_t)"
71};
72
73
74static Function *getCalledFunction(const Value *V, bool LookThroughBitCast) {
75  if (LookThroughBitCast)
76    V = V->stripPointerCasts();
77
78  CallSite CS(const_cast<Value*>(V));
79  if (!CS.getInstruction())
80    return nullptr;
81
82  if (CS.isNoBuiltin())
83    return nullptr;
84
85  Function *Callee = CS.getCalledFunction();
86  if (!Callee || !Callee->isDeclaration())
87    return nullptr;
88  return Callee;
89}
90
91/// \brief Returns the allocation data for the given value if it is a call to a
92/// known allocation function, and NULL otherwise.
93static const AllocFnsTy *getAllocationData(const Value *V, AllocType AllocTy,
94                                           const TargetLibraryInfo *TLI,
95                                           bool LookThroughBitCast = false) {
96  // Skip intrinsics
97  if (isa<IntrinsicInst>(V))
98    return nullptr;
99
100  Function *Callee = getCalledFunction(V, LookThroughBitCast);
101  if (!Callee)
102    return nullptr;
103
104  // Make sure that the function is available.
105  StringRef FnName = Callee->getName();
106  LibFunc::Func TLIFn;
107  if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn))
108    return nullptr;
109
110  unsigned i = 0;
111  bool found = false;
112  for ( ; i < array_lengthof(AllocationFnData); ++i) {
113    if (AllocationFnData[i].Func == TLIFn) {
114      found = true;
115      break;
116    }
117  }
118  if (!found)
119    return nullptr;
120
121  const AllocFnsTy *FnData = &AllocationFnData[i];
122  if ((FnData->AllocTy & AllocTy) != FnData->AllocTy)
123    return nullptr;
124
125  // Check function prototype.
126  int FstParam = FnData->FstParam;
127  int SndParam = FnData->SndParam;
128  FunctionType *FTy = Callee->getFunctionType();
129
130  if (FTy->getReturnType() == Type::getInt8PtrTy(FTy->getContext()) &&
131      FTy->getNumParams() == FnData->NumParams &&
132      (FstParam < 0 ||
133       (FTy->getParamType(FstParam)->isIntegerTy(32) ||
134        FTy->getParamType(FstParam)->isIntegerTy(64))) &&
135      (SndParam < 0 ||
136       FTy->getParamType(SndParam)->isIntegerTy(32) ||
137       FTy->getParamType(SndParam)->isIntegerTy(64)))
138    return FnData;
139  return nullptr;
140}
141
142static bool hasNoAliasAttr(const Value *V, bool LookThroughBitCast) {
143  ImmutableCallSite CS(LookThroughBitCast ? V->stripPointerCasts() : V);
144  return CS && CS.hasFnAttr(Attribute::NoAlias);
145}
146
147
148/// \brief Tests if a value is a call or invoke to a library function that
149/// allocates or reallocates memory (either malloc, calloc, realloc, or strdup
150/// like).
151bool llvm::isAllocationFn(const Value *V, const TargetLibraryInfo *TLI,
152                          bool LookThroughBitCast) {
153  return getAllocationData(V, AnyAlloc, TLI, LookThroughBitCast);
154}
155
156/// \brief Tests if a value is a call or invoke to a function that returns a
157/// NoAlias pointer (including malloc/calloc/realloc/strdup-like functions).
158bool llvm::isNoAliasFn(const Value *V, const TargetLibraryInfo *TLI,
159                       bool LookThroughBitCast) {
160  // it's safe to consider realloc as noalias since accessing the original
161  // pointer is undefined behavior
162  return isAllocationFn(V, TLI, LookThroughBitCast) ||
163         hasNoAliasAttr(V, LookThroughBitCast);
164}
165
166/// \brief Tests if a value is a call or invoke to a library function that
167/// allocates uninitialized memory (such as malloc).
168bool llvm::isMallocLikeFn(const Value *V, const TargetLibraryInfo *TLI,
169                          bool LookThroughBitCast) {
170  return getAllocationData(V, MallocLike, TLI, LookThroughBitCast);
171}
172
173/// \brief Tests if a value is a call or invoke to a library function that
174/// allocates zero-filled memory (such as calloc).
175bool llvm::isCallocLikeFn(const Value *V, const TargetLibraryInfo *TLI,
176                          bool LookThroughBitCast) {
177  return getAllocationData(V, CallocLike, TLI, LookThroughBitCast);
178}
179
180/// \brief Tests if a value is a call or invoke to a library function that
181/// allocates memory (either malloc, calloc, or strdup like).
182bool llvm::isAllocLikeFn(const Value *V, const TargetLibraryInfo *TLI,
183                         bool LookThroughBitCast) {
184  return getAllocationData(V, AllocLike, TLI, LookThroughBitCast);
185}
186
187/// \brief Tests if a value is a call or invoke to a library function that
188/// reallocates memory (such as realloc).
189bool llvm::isReallocLikeFn(const Value *V, const TargetLibraryInfo *TLI,
190                           bool LookThroughBitCast) {
191  return getAllocationData(V, ReallocLike, TLI, LookThroughBitCast);
192}
193
194/// \brief Tests if a value is a call or invoke to a library function that
195/// allocates memory and never returns null (such as operator new).
196bool llvm::isOperatorNewLikeFn(const Value *V, const TargetLibraryInfo *TLI,
197                               bool LookThroughBitCast) {
198  return getAllocationData(V, OpNewLike, TLI, LookThroughBitCast);
199}
200
201/// extractMallocCall - Returns the corresponding CallInst if the instruction
202/// is a malloc call.  Since CallInst::CreateMalloc() only creates calls, we
203/// ignore InvokeInst here.
204const CallInst *llvm::extractMallocCall(const Value *I,
205                                        const TargetLibraryInfo *TLI) {
206  return isMallocLikeFn(I, TLI) ? dyn_cast<CallInst>(I) : nullptr;
207}
208
209static Value *computeArraySize(const CallInst *CI, const DataLayout *DL,
210                               const TargetLibraryInfo *TLI,
211                               bool LookThroughSExt = false) {
212  if (!CI)
213    return nullptr;
214
215  // The size of the malloc's result type must be known to determine array size.
216  Type *T = getMallocAllocatedType(CI, TLI);
217  if (!T || !T->isSized() || !DL)
218    return nullptr;
219
220  unsigned ElementSize = DL->getTypeAllocSize(T);
221  if (StructType *ST = dyn_cast<StructType>(T))
222    ElementSize = DL->getStructLayout(ST)->getSizeInBytes();
223
224  // If malloc call's arg can be determined to be a multiple of ElementSize,
225  // return the multiple.  Otherwise, return NULL.
226  Value *MallocArg = CI->getArgOperand(0);
227  Value *Multiple = nullptr;
228  if (ComputeMultiple(MallocArg, ElementSize, Multiple,
229                      LookThroughSExt))
230    return Multiple;
231
232  return nullptr;
233}
234
235/// isArrayMalloc - Returns the corresponding CallInst if the instruction
236/// is a call to malloc whose array size can be determined and the array size
237/// is not constant 1.  Otherwise, return NULL.
238const CallInst *llvm::isArrayMalloc(const Value *I,
239                                    const DataLayout *DL,
240                                    const TargetLibraryInfo *TLI) {
241  const CallInst *CI = extractMallocCall(I, TLI);
242  Value *ArraySize = computeArraySize(CI, DL, TLI);
243
244  if (ConstantInt *ConstSize = dyn_cast_or_null<ConstantInt>(ArraySize))
245    if (ConstSize->isOne())
246      return CI;
247
248  // CI is a non-array malloc or we can't figure out that it is an array malloc.
249  return nullptr;
250}
251
252/// getMallocType - Returns the PointerType resulting from the malloc call.
253/// The PointerType depends on the number of bitcast uses of the malloc call:
254///   0: PointerType is the calls' return type.
255///   1: PointerType is the bitcast's result type.
256///  >1: Unique PointerType cannot be determined, return NULL.
257PointerType *llvm::getMallocType(const CallInst *CI,
258                                 const TargetLibraryInfo *TLI) {
259  assert(isMallocLikeFn(CI, TLI) && "getMallocType and not malloc call");
260
261  PointerType *MallocType = nullptr;
262  unsigned NumOfBitCastUses = 0;
263
264  // Determine if CallInst has a bitcast use.
265  for (Value::const_user_iterator UI = CI->user_begin(), E = CI->user_end();
266       UI != E;)
267    if (const BitCastInst *BCI = dyn_cast<BitCastInst>(*UI++)) {
268      MallocType = cast<PointerType>(BCI->getDestTy());
269      NumOfBitCastUses++;
270    }
271
272  // Malloc call has 1 bitcast use, so type is the bitcast's destination type.
273  if (NumOfBitCastUses == 1)
274    return MallocType;
275
276  // Malloc call was not bitcast, so type is the malloc function's return type.
277  if (NumOfBitCastUses == 0)
278    return cast<PointerType>(CI->getType());
279
280  // Type could not be determined.
281  return nullptr;
282}
283
284/// getMallocAllocatedType - Returns the Type allocated by malloc call.
285/// The Type depends on the number of bitcast uses of the malloc call:
286///   0: PointerType is the malloc calls' return type.
287///   1: PointerType is the bitcast's result type.
288///  >1: Unique PointerType cannot be determined, return NULL.
289Type *llvm::getMallocAllocatedType(const CallInst *CI,
290                                   const TargetLibraryInfo *TLI) {
291  PointerType *PT = getMallocType(CI, TLI);
292  return PT ? PT->getElementType() : nullptr;
293}
294
295/// getMallocArraySize - Returns the array size of a malloc call.  If the
296/// argument passed to malloc is a multiple of the size of the malloced type,
297/// then return that multiple.  For non-array mallocs, the multiple is
298/// constant 1.  Otherwise, return NULL for mallocs whose array size cannot be
299/// determined.
300Value *llvm::getMallocArraySize(CallInst *CI, const DataLayout *DL,
301                                const TargetLibraryInfo *TLI,
302                                bool LookThroughSExt) {
303  assert(isMallocLikeFn(CI, TLI) && "getMallocArraySize and not malloc call");
304  return computeArraySize(CI, DL, TLI, LookThroughSExt);
305}
306
307
308/// extractCallocCall - Returns the corresponding CallInst if the instruction
309/// is a calloc call.
310const CallInst *llvm::extractCallocCall(const Value *I,
311                                        const TargetLibraryInfo *TLI) {
312  return isCallocLikeFn(I, TLI) ? cast<CallInst>(I) : nullptr;
313}
314
315
316/// isFreeCall - Returns non-null if the value is a call to the builtin free()
317const CallInst *llvm::isFreeCall(const Value *I, const TargetLibraryInfo *TLI) {
318  const CallInst *CI = dyn_cast<CallInst>(I);
319  if (!CI || isa<IntrinsicInst>(CI))
320    return nullptr;
321  Function *Callee = CI->getCalledFunction();
322  if (Callee == nullptr || !Callee->isDeclaration())
323    return nullptr;
324
325  StringRef FnName = Callee->getName();
326  LibFunc::Func TLIFn;
327  if (!TLI || !TLI->getLibFunc(FnName, TLIFn) || !TLI->has(TLIFn))
328    return nullptr;
329
330  unsigned ExpectedNumParams;
331  if (TLIFn == LibFunc::free ||
332      TLIFn == LibFunc::ZdlPv || // operator delete(void*)
333      TLIFn == LibFunc::ZdaPv)   // operator delete[](void*)
334    ExpectedNumParams = 1;
335  else if (TLIFn == LibFunc::ZdlPvRKSt9nothrow_t || // delete(void*, nothrow)
336           TLIFn == LibFunc::ZdaPvRKSt9nothrow_t)   // delete[](void*, nothrow)
337    ExpectedNumParams = 2;
338  else
339    return nullptr;
340
341  // Check free prototype.
342  // FIXME: workaround for PR5130, this will be obsolete when a nobuiltin
343  // attribute will exist.
344  FunctionType *FTy = Callee->getFunctionType();
345  if (!FTy->getReturnType()->isVoidTy())
346    return nullptr;
347  if (FTy->getNumParams() != ExpectedNumParams)
348    return nullptr;
349  if (FTy->getParamType(0) != Type::getInt8PtrTy(Callee->getContext()))
350    return nullptr;
351
352  return CI;
353}
354
355
356
357//===----------------------------------------------------------------------===//
358//  Utility functions to compute size of objects.
359//
360
361
362/// \brief Compute the size of the object pointed by Ptr. Returns true and the
363/// object size in Size if successful, and false otherwise.
364/// If RoundToAlign is true, then Size is rounded up to the aligment of allocas,
365/// byval arguments, and global variables.
366bool llvm::getObjectSize(const Value *Ptr, uint64_t &Size, const DataLayout *DL,
367                         const TargetLibraryInfo *TLI, bool RoundToAlign) {
368  if (!DL)
369    return false;
370
371  ObjectSizeOffsetVisitor Visitor(DL, TLI, Ptr->getContext(), RoundToAlign);
372  SizeOffsetType Data = Visitor.compute(const_cast<Value*>(Ptr));
373  if (!Visitor.bothKnown(Data))
374    return false;
375
376  APInt ObjSize = Data.first, Offset = Data.second;
377  // check for overflow
378  if (Offset.slt(0) || ObjSize.ult(Offset))
379    Size = 0;
380  else
381    Size = (ObjSize - Offset).getZExtValue();
382  return true;
383}
384
385
386STATISTIC(ObjectVisitorArgument,
387          "Number of arguments with unsolved size and offset");
388STATISTIC(ObjectVisitorLoad,
389          "Number of load instructions with unsolved size and offset");
390
391
392APInt ObjectSizeOffsetVisitor::align(APInt Size, uint64_t Align) {
393  if (RoundToAlign && Align)
394    return APInt(IntTyBits, RoundUpToAlignment(Size.getZExtValue(), Align));
395  return Size;
396}
397
398ObjectSizeOffsetVisitor::ObjectSizeOffsetVisitor(const DataLayout *DL,
399                                                 const TargetLibraryInfo *TLI,
400                                                 LLVMContext &Context,
401                                                 bool RoundToAlign)
402: DL(DL), TLI(TLI), RoundToAlign(RoundToAlign) {
403  // Pointer size must be rechecked for each object visited since it could have
404  // a different address space.
405}
406
407SizeOffsetType ObjectSizeOffsetVisitor::compute(Value *V) {
408  IntTyBits = DL->getPointerTypeSizeInBits(V->getType());
409  Zero = APInt::getNullValue(IntTyBits);
410
411  V = V->stripPointerCasts();
412  if (Instruction *I = dyn_cast<Instruction>(V)) {
413    // If we have already seen this instruction, bail out. Cycles can happen in
414    // unreachable code after constant propagation.
415    if (!SeenInsts.insert(I))
416      return unknown();
417
418    if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
419      return visitGEPOperator(*GEP);
420    return visit(*I);
421  }
422  if (Argument *A = dyn_cast<Argument>(V))
423    return visitArgument(*A);
424  if (ConstantPointerNull *P = dyn_cast<ConstantPointerNull>(V))
425    return visitConstantPointerNull(*P);
426  if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
427    return visitGlobalAlias(*GA);
428  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
429    return visitGlobalVariable(*GV);
430  if (UndefValue *UV = dyn_cast<UndefValue>(V))
431    return visitUndefValue(*UV);
432  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
433    if (CE->getOpcode() == Instruction::IntToPtr)
434      return unknown(); // clueless
435    if (CE->getOpcode() == Instruction::GetElementPtr)
436      return visitGEPOperator(cast<GEPOperator>(*CE));
437  }
438
439  DEBUG(dbgs() << "ObjectSizeOffsetVisitor::compute() unhandled value: " << *V
440        << '\n');
441  return unknown();
442}
443
444SizeOffsetType ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) {
445  if (!I.getAllocatedType()->isSized())
446    return unknown();
447
448  APInt Size(IntTyBits, DL->getTypeAllocSize(I.getAllocatedType()));
449  if (!I.isArrayAllocation())
450    return std::make_pair(align(Size, I.getAlignment()), Zero);
451
452  Value *ArraySize = I.getArraySize();
453  if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) {
454    Size *= C->getValue().zextOrSelf(IntTyBits);
455    return std::make_pair(align(Size, I.getAlignment()), Zero);
456  }
457  return unknown();
458}
459
460SizeOffsetType ObjectSizeOffsetVisitor::visitArgument(Argument &A) {
461  // no interprocedural analysis is done at the moment
462  if (!A.hasByValOrInAllocaAttr()) {
463    ++ObjectVisitorArgument;
464    return unknown();
465  }
466  PointerType *PT = cast<PointerType>(A.getType());
467  APInt Size(IntTyBits, DL->getTypeAllocSize(PT->getElementType()));
468  return std::make_pair(align(Size, A.getParamAlignment()), Zero);
469}
470
471SizeOffsetType ObjectSizeOffsetVisitor::visitCallSite(CallSite CS) {
472  const AllocFnsTy *FnData = getAllocationData(CS.getInstruction(), AnyAlloc,
473                                               TLI);
474  if (!FnData)
475    return unknown();
476
477  // handle strdup-like functions separately
478  if (FnData->AllocTy == StrDupLike) {
479    APInt Size(IntTyBits, GetStringLength(CS.getArgument(0)));
480    if (!Size)
481      return unknown();
482
483    // strndup limits strlen
484    if (FnData->FstParam > 0) {
485      ConstantInt *Arg= dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam));
486      if (!Arg)
487        return unknown();
488
489      APInt MaxSize = Arg->getValue().zextOrSelf(IntTyBits);
490      if (Size.ugt(MaxSize))
491        Size = MaxSize + 1;
492    }
493    return std::make_pair(Size, Zero);
494  }
495
496  ConstantInt *Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->FstParam));
497  if (!Arg)
498    return unknown();
499
500  APInt Size = Arg->getValue().zextOrSelf(IntTyBits);
501  // size determined by just 1 parameter
502  if (FnData->SndParam < 0)
503    return std::make_pair(Size, Zero);
504
505  Arg = dyn_cast<ConstantInt>(CS.getArgument(FnData->SndParam));
506  if (!Arg)
507    return unknown();
508
509  Size *= Arg->getValue().zextOrSelf(IntTyBits);
510  return std::make_pair(Size, Zero);
511
512  // TODO: handle more standard functions (+ wchar cousins):
513  // - strdup / strndup
514  // - strcpy / strncpy
515  // - strcat / strncat
516  // - memcpy / memmove
517  // - strcat / strncat
518  // - memset
519}
520
521SizeOffsetType
522ObjectSizeOffsetVisitor::visitConstantPointerNull(ConstantPointerNull&) {
523  return std::make_pair(Zero, Zero);
524}
525
526SizeOffsetType
527ObjectSizeOffsetVisitor::visitExtractElementInst(ExtractElementInst&) {
528  return unknown();
529}
530
531SizeOffsetType
532ObjectSizeOffsetVisitor::visitExtractValueInst(ExtractValueInst&) {
533  // Easy cases were already folded by previous passes.
534  return unknown();
535}
536
537SizeOffsetType ObjectSizeOffsetVisitor::visitGEPOperator(GEPOperator &GEP) {
538  SizeOffsetType PtrData = compute(GEP.getPointerOperand());
539  APInt Offset(IntTyBits, 0);
540  if (!bothKnown(PtrData) || !GEP.accumulateConstantOffset(*DL, Offset))
541    return unknown();
542
543  return std::make_pair(PtrData.first, PtrData.second + Offset);
544}
545
546SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalAlias(GlobalAlias &GA) {
547  if (GA.mayBeOverridden())
548    return unknown();
549  return compute(GA.getAliasee());
550}
551
552SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalVariable(GlobalVariable &GV){
553  if (!GV.hasDefinitiveInitializer())
554    return unknown();
555
556  APInt Size(IntTyBits, DL->getTypeAllocSize(GV.getType()->getElementType()));
557  return std::make_pair(align(Size, GV.getAlignment()), Zero);
558}
559
560SizeOffsetType ObjectSizeOffsetVisitor::visitIntToPtrInst(IntToPtrInst&) {
561  // clueless
562  return unknown();
563}
564
565SizeOffsetType ObjectSizeOffsetVisitor::visitLoadInst(LoadInst&) {
566  ++ObjectVisitorLoad;
567  return unknown();
568}
569
570SizeOffsetType ObjectSizeOffsetVisitor::visitPHINode(PHINode&) {
571  // too complex to analyze statically.
572  return unknown();
573}
574
575SizeOffsetType ObjectSizeOffsetVisitor::visitSelectInst(SelectInst &I) {
576  SizeOffsetType TrueSide  = compute(I.getTrueValue());
577  SizeOffsetType FalseSide = compute(I.getFalseValue());
578  if (bothKnown(TrueSide) && bothKnown(FalseSide) && TrueSide == FalseSide)
579    return TrueSide;
580  return unknown();
581}
582
583SizeOffsetType ObjectSizeOffsetVisitor::visitUndefValue(UndefValue&) {
584  return std::make_pair(Zero, Zero);
585}
586
587SizeOffsetType ObjectSizeOffsetVisitor::visitInstruction(Instruction &I) {
588  DEBUG(dbgs() << "ObjectSizeOffsetVisitor unknown instruction:" << I << '\n');
589  return unknown();
590}
591
592ObjectSizeOffsetEvaluator::ObjectSizeOffsetEvaluator(const DataLayout *DL,
593                                                     const TargetLibraryInfo *TLI,
594                                                     LLVMContext &Context,
595                                                     bool RoundToAlign)
596: DL(DL), TLI(TLI), Context(Context), Builder(Context, TargetFolder(DL)),
597  RoundToAlign(RoundToAlign) {
598  // IntTy and Zero must be set for each compute() since the address space may
599  // be different for later objects.
600}
601
602SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute(Value *V) {
603  // XXX - Are vectors of pointers possible here?
604  IntTy = cast<IntegerType>(DL->getIntPtrType(V->getType()));
605  Zero = ConstantInt::get(IntTy, 0);
606
607  SizeOffsetEvalType Result = compute_(V);
608
609  if (!bothKnown(Result)) {
610    // erase everything that was computed in this iteration from the cache, so
611    // that no dangling references are left behind. We could be a bit smarter if
612    // we kept a dependency graph. It's probably not worth the complexity.
613    for (PtrSetTy::iterator I=SeenVals.begin(), E=SeenVals.end(); I != E; ++I) {
614      CacheMapTy::iterator CacheIt = CacheMap.find(*I);
615      // non-computable results can be safely cached
616      if (CacheIt != CacheMap.end() && anyKnown(CacheIt->second))
617        CacheMap.erase(CacheIt);
618    }
619  }
620
621  SeenVals.clear();
622  return Result;
623}
624
625SizeOffsetEvalType ObjectSizeOffsetEvaluator::compute_(Value *V) {
626  ObjectSizeOffsetVisitor Visitor(DL, TLI, Context, RoundToAlign);
627  SizeOffsetType Const = Visitor.compute(V);
628  if (Visitor.bothKnown(Const))
629    return std::make_pair(ConstantInt::get(Context, Const.first),
630                          ConstantInt::get(Context, Const.second));
631
632  V = V->stripPointerCasts();
633
634  // check cache
635  CacheMapTy::iterator CacheIt = CacheMap.find(V);
636  if (CacheIt != CacheMap.end())
637    return CacheIt->second;
638
639  // always generate code immediately before the instruction being
640  // processed, so that the generated code dominates the same BBs
641  Instruction *PrevInsertPoint = Builder.GetInsertPoint();
642  if (Instruction *I = dyn_cast<Instruction>(V))
643    Builder.SetInsertPoint(I);
644
645  // now compute the size and offset
646  SizeOffsetEvalType Result;
647
648  // Record the pointers that were handled in this run, so that they can be
649  // cleaned later if something fails. We also use this set to break cycles that
650  // can occur in dead code.
651  if (!SeenVals.insert(V)) {
652    Result = unknown();
653  } else if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
654    Result = visitGEPOperator(*GEP);
655  } else if (Instruction *I = dyn_cast<Instruction>(V)) {
656    Result = visit(*I);
657  } else if (isa<Argument>(V) ||
658             (isa<ConstantExpr>(V) &&
659              cast<ConstantExpr>(V)->getOpcode() == Instruction::IntToPtr) ||
660             isa<GlobalAlias>(V) ||
661             isa<GlobalVariable>(V)) {
662    // ignore values where we cannot do more than what ObjectSizeVisitor can
663    Result = unknown();
664  } else {
665    DEBUG(dbgs() << "ObjectSizeOffsetEvaluator::compute() unhandled value: "
666          << *V << '\n');
667    Result = unknown();
668  }
669
670  if (PrevInsertPoint)
671    Builder.SetInsertPoint(PrevInsertPoint);
672
673  // Don't reuse CacheIt since it may be invalid at this point.
674  CacheMap[V] = Result;
675  return Result;
676}
677
678SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitAllocaInst(AllocaInst &I) {
679  if (!I.getAllocatedType()->isSized())
680    return unknown();
681
682  // must be a VLA
683  assert(I.isArrayAllocation());
684  Value *ArraySize = I.getArraySize();
685  Value *Size = ConstantInt::get(ArraySize->getType(),
686                                 DL->getTypeAllocSize(I.getAllocatedType()));
687  Size = Builder.CreateMul(Size, ArraySize);
688  return std::make_pair(Size, Zero);
689}
690
691SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitCallSite(CallSite CS) {
692  const AllocFnsTy *FnData = getAllocationData(CS.getInstruction(), AnyAlloc,
693                                               TLI);
694  if (!FnData)
695    return unknown();
696
697  // handle strdup-like functions separately
698  if (FnData->AllocTy == StrDupLike) {
699    // TODO
700    return unknown();
701  }
702
703  Value *FirstArg = CS.getArgument(FnData->FstParam);
704  FirstArg = Builder.CreateZExt(FirstArg, IntTy);
705  if (FnData->SndParam < 0)
706    return std::make_pair(FirstArg, Zero);
707
708  Value *SecondArg = CS.getArgument(FnData->SndParam);
709  SecondArg = Builder.CreateZExt(SecondArg, IntTy);
710  Value *Size = Builder.CreateMul(FirstArg, SecondArg);
711  return std::make_pair(Size, Zero);
712
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
720}
721
722SizeOffsetEvalType
723ObjectSizeOffsetEvaluator::visitExtractElementInst(ExtractElementInst&) {
724  return unknown();
725}
726
727SizeOffsetEvalType
728ObjectSizeOffsetEvaluator::visitExtractValueInst(ExtractValueInst&) {
729  return unknown();
730}
731
732SizeOffsetEvalType
733ObjectSizeOffsetEvaluator::visitGEPOperator(GEPOperator &GEP) {
734  SizeOffsetEvalType PtrData = compute_(GEP.getPointerOperand());
735  if (!bothKnown(PtrData))
736    return unknown();
737
738  Value *Offset = EmitGEPOffset(&Builder, *DL, &GEP, /*NoAssumptions=*/true);
739  Offset = Builder.CreateAdd(PtrData.second, Offset);
740  return std::make_pair(PtrData.first, Offset);
741}
742
743SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitIntToPtrInst(IntToPtrInst&) {
744  // clueless
745  return unknown();
746}
747
748SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitLoadInst(LoadInst&) {
749  return unknown();
750}
751
752SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitPHINode(PHINode &PHI) {
753  // create 2 PHIs: one for size and another for offset
754  PHINode *SizePHI   = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues());
755  PHINode *OffsetPHI = Builder.CreatePHI(IntTy, PHI.getNumIncomingValues());
756
757  // insert right away in the cache to handle recursive PHIs
758  CacheMap[&PHI] = std::make_pair(SizePHI, OffsetPHI);
759
760  // compute offset/size for each PHI incoming pointer
761  for (unsigned i = 0, e = PHI.getNumIncomingValues(); i != e; ++i) {
762    Builder.SetInsertPoint(PHI.getIncomingBlock(i)->getFirstInsertionPt());
763    SizeOffsetEvalType EdgeData = compute_(PHI.getIncomingValue(i));
764
765    if (!bothKnown(EdgeData)) {
766      OffsetPHI->replaceAllUsesWith(UndefValue::get(IntTy));
767      OffsetPHI->eraseFromParent();
768      SizePHI->replaceAllUsesWith(UndefValue::get(IntTy));
769      SizePHI->eraseFromParent();
770      return unknown();
771    }
772    SizePHI->addIncoming(EdgeData.first, PHI.getIncomingBlock(i));
773    OffsetPHI->addIncoming(EdgeData.second, PHI.getIncomingBlock(i));
774  }
775
776  Value *Size = SizePHI, *Offset = OffsetPHI, *Tmp;
777  if ((Tmp = SizePHI->hasConstantValue())) {
778    Size = Tmp;
779    SizePHI->replaceAllUsesWith(Size);
780    SizePHI->eraseFromParent();
781  }
782  if ((Tmp = OffsetPHI->hasConstantValue())) {
783    Offset = Tmp;
784    OffsetPHI->replaceAllUsesWith(Offset);
785    OffsetPHI->eraseFromParent();
786  }
787  return std::make_pair(Size, Offset);
788}
789
790SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitSelectInst(SelectInst &I) {
791  SizeOffsetEvalType TrueSide  = compute_(I.getTrueValue());
792  SizeOffsetEvalType FalseSide = compute_(I.getFalseValue());
793
794  if (!bothKnown(TrueSide) || !bothKnown(FalseSide))
795    return unknown();
796  if (TrueSide == FalseSide)
797    return TrueSide;
798
799  Value *Size = Builder.CreateSelect(I.getCondition(), TrueSide.first,
800                                     FalseSide.first);
801  Value *Offset = Builder.CreateSelect(I.getCondition(), TrueSide.second,
802                                       FalseSide.second);
803  return std::make_pair(Size, Offset);
804}
805
806SizeOffsetEvalType ObjectSizeOffsetEvaluator::visitInstruction(Instruction &I) {
807  DEBUG(dbgs() << "ObjectSizeOffsetEvaluator unknown instruction:" << I <<'\n');
808  return unknown();
809}
810