1//===-- AddressSanitizer.cpp - memory error detector ------------*- C++ -*-===//
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 file is a part of AddressSanitizer, an address sanity checker.
11// Details of the algorithm:
12//  http://code.google.com/p/address-sanitizer/wiki/AddressSanitizerAlgorithm
13//
14//===----------------------------------------------------------------------===//
15
16#define DEBUG_TYPE "asan"
17
18#include "FunctionBlackList.h"
19#include "llvm/ADT/ArrayRef.h"
20#include "llvm/ADT/OwningPtr.h"
21#include "llvm/ADT/SmallSet.h"
22#include "llvm/ADT/SmallString.h"
23#include "llvm/ADT/SmallVector.h"
24#include "llvm/ADT/StringExtras.h"
25#include "llvm/ADT/Triple.h"
26#include "llvm/Function.h"
27#include "llvm/IntrinsicInst.h"
28#include "llvm/LLVMContext.h"
29#include "llvm/Module.h"
30#include "llvm/Support/CommandLine.h"
31#include "llvm/Support/DataTypes.h"
32#include "llvm/Support/Debug.h"
33#include "llvm/Support/IRBuilder.h"
34#include "llvm/Support/raw_ostream.h"
35#include "llvm/Support/system_error.h"
36#include "llvm/Target/TargetData.h"
37#include "llvm/Target/TargetMachine.h"
38#include "llvm/Transforms/Instrumentation.h"
39#include "llvm/Transforms/Utils/BasicBlockUtils.h"
40#include "llvm/Transforms/Utils/ModuleUtils.h"
41#include "llvm/Type.h"
42
43#include <string>
44#include <algorithm>
45
46using namespace llvm;
47
48static const uint64_t kDefaultShadowScale = 3;
49static const uint64_t kDefaultShadowOffset32 = 1ULL << 29;
50static const uint64_t kDefaultShadowOffset64 = 1ULL << 44;
51static const uint64_t kDefaultShadowOffsetAndroid = 0;
52
53static const size_t kMaxStackMallocSize = 1 << 16;  // 64K
54static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3;
55static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E;
56
57static const char *kAsanModuleCtorName = "asan.module_ctor";
58static const char *kAsanModuleDtorName = "asan.module_dtor";
59static const int   kAsanCtorAndCtorPriority = 1;
60static const char *kAsanReportErrorTemplate = "__asan_report_";
61static const char *kAsanRegisterGlobalsName = "__asan_register_globals";
62static const char *kAsanUnregisterGlobalsName = "__asan_unregister_globals";
63static const char *kAsanInitName = "__asan_init";
64static const char *kAsanHandleNoReturnName = "__asan_handle_no_return";
65static const char *kAsanMappingOffsetName = "__asan_mapping_offset";
66static const char *kAsanMappingScaleName = "__asan_mapping_scale";
67static const char *kAsanStackMallocName = "__asan_stack_malloc";
68static const char *kAsanStackFreeName = "__asan_stack_free";
69
70static const int kAsanStackLeftRedzoneMagic = 0xf1;
71static const int kAsanStackMidRedzoneMagic = 0xf2;
72static const int kAsanStackRightRedzoneMagic = 0xf3;
73static const int kAsanStackPartialRedzoneMagic = 0xf4;
74
75// Command-line flags.
76
77// This flag may need to be replaced with -f[no-]asan-reads.
78static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
79       cl::desc("instrument read instructions"), cl::Hidden, cl::init(true));
80static cl::opt<bool> ClInstrumentWrites("asan-instrument-writes",
81       cl::desc("instrument write instructions"), cl::Hidden, cl::init(true));
82// This flag may need to be replaced with -f[no]asan-stack.
83static cl::opt<bool> ClStack("asan-stack",
84       cl::desc("Handle stack memory"), cl::Hidden, cl::init(true));
85// This flag may need to be replaced with -f[no]asan-use-after-return.
86static cl::opt<bool> ClUseAfterReturn("asan-use-after-return",
87       cl::desc("Check return-after-free"), cl::Hidden, cl::init(false));
88// This flag may need to be replaced with -f[no]asan-globals.
89static cl::opt<bool> ClGlobals("asan-globals",
90       cl::desc("Handle global objects"), cl::Hidden, cl::init(true));
91static cl::opt<bool> ClMemIntrin("asan-memintrin",
92       cl::desc("Handle memset/memcpy/memmove"), cl::Hidden, cl::init(true));
93// This flag may need to be replaced with -fasan-blacklist.
94static cl::opt<std::string>  ClBlackListFile("asan-blacklist",
95       cl::desc("File containing the list of functions to ignore "
96                "during instrumentation"), cl::Hidden);
97
98// These flags allow to change the shadow mapping.
99// The shadow mapping looks like
100//    Shadow = (Mem >> scale) + (1 << offset_log)
101static cl::opt<int> ClMappingScale("asan-mapping-scale",
102       cl::desc("scale of asan shadow mapping"), cl::Hidden, cl::init(0));
103static cl::opt<int> ClMappingOffsetLog("asan-mapping-offset-log",
104       cl::desc("offset of asan shadow mapping"), cl::Hidden, cl::init(-1));
105
106// Optimization flags. Not user visible, used mostly for testing
107// and benchmarking the tool.
108static cl::opt<bool> ClOpt("asan-opt",
109       cl::desc("Optimize instrumentation"), cl::Hidden, cl::init(true));
110static cl::opt<bool> ClOptSameTemp("asan-opt-same-temp",
111       cl::desc("Instrument the same temp just once"), cl::Hidden,
112       cl::init(true));
113static cl::opt<bool> ClOptGlobals("asan-opt-globals",
114       cl::desc("Don't instrument scalar globals"), cl::Hidden, cl::init(true));
115
116// Debug flags.
117static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
118                            cl::init(0));
119static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
120                                 cl::Hidden, cl::init(0));
121static cl::opt<std::string> ClDebugFunc("asan-debug-func",
122                                        cl::Hidden, cl::desc("Debug func"));
123static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
124                               cl::Hidden, cl::init(-1));
125static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug man inst"),
126                               cl::Hidden, cl::init(-1));
127
128namespace {
129
130/// AddressSanitizer: instrument the code in module to find memory bugs.
131struct AddressSanitizer : public ModulePass {
132  AddressSanitizer();
133  virtual const char *getPassName() const;
134  void instrumentMop(Instruction *I);
135  void instrumentAddress(Instruction *OrigIns, IRBuilder<> &IRB,
136                         Value *Addr, uint32_t TypeSize, bool IsWrite);
137  Instruction *generateCrashCode(IRBuilder<> &IRB, Value *Addr,
138                                 bool IsWrite, uint32_t TypeSize);
139  bool instrumentMemIntrinsic(MemIntrinsic *MI);
140  void instrumentMemIntrinsicParam(Instruction *OrigIns, Value *Addr,
141                                  Value *Size,
142                                   Instruction *InsertBefore, bool IsWrite);
143  Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
144  bool handleFunction(Module &M, Function &F);
145  bool maybeInsertAsanInitAtFunctionEntry(Function &F);
146  bool poisonStackInFunction(Module &M, Function &F);
147  virtual bool runOnModule(Module &M);
148  bool insertGlobalRedzones(Module &M);
149  BranchInst *splitBlockAndInsertIfThen(Instruction *SplitBefore, Value *Cmp);
150  static char ID;  // Pass identification, replacement for typeid
151
152 private:
153
154  uint64_t getAllocaSizeInBytes(AllocaInst *AI) {
155    Type *Ty = AI->getAllocatedType();
156    uint64_t SizeInBytes = TD->getTypeAllocSize(Ty);
157    return SizeInBytes;
158  }
159  uint64_t getAlignedSize(uint64_t SizeInBytes) {
160    return ((SizeInBytes + RedzoneSize - 1)
161            / RedzoneSize) * RedzoneSize;
162  }
163  uint64_t getAlignedAllocaSize(AllocaInst *AI) {
164    uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
165    return getAlignedSize(SizeInBytes);
166  }
167
168  void PoisonStack(const ArrayRef<AllocaInst*> &AllocaVec, IRBuilder<> IRB,
169                   Value *ShadowBase, bool DoPoison);
170  bool LooksLikeCodeInBug11395(Instruction *I);
171
172  Module      *CurrentModule;
173  LLVMContext *C;
174  TargetData *TD;
175  uint64_t MappingOffset;
176  int MappingScale;
177  size_t RedzoneSize;
178  int LongSize;
179  Type *IntptrTy;
180  Type *IntptrPtrTy;
181  Function *AsanCtorFunction;
182  Function *AsanInitFunction;
183  Instruction *CtorInsertBefore;
184  OwningPtr<FunctionBlackList> BL;
185};
186}  // namespace
187
188char AddressSanitizer::ID = 0;
189INITIALIZE_PASS(AddressSanitizer, "asan",
190    "AddressSanitizer: detects use-after-free and out-of-bounds bugs.",
191    false, false)
192AddressSanitizer::AddressSanitizer() : ModulePass(ID) { }
193ModulePass *llvm::createAddressSanitizerPass() {
194  return new AddressSanitizer();
195}
196
197const char *AddressSanitizer::getPassName() const {
198  return "AddressSanitizer";
199}
200
201// Create a constant for Str so that we can pass it to the run-time lib.
202static GlobalVariable *createPrivateGlobalForString(Module &M, StringRef Str) {
203  Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
204  return new GlobalVariable(M, StrConst->getType(), true,
205                            GlobalValue::PrivateLinkage, StrConst, "");
206}
207
208// Split the basic block and insert an if-then code.
209// Before:
210//   Head
211//   SplitBefore
212//   Tail
213// After:
214//   Head
215//   if (Cmp)
216//     NewBasicBlock
217//   SplitBefore
218//   Tail
219//
220// Returns the NewBasicBlock's terminator.
221BranchInst *AddressSanitizer::splitBlockAndInsertIfThen(
222    Instruction *SplitBefore, Value *Cmp) {
223  BasicBlock *Head = SplitBefore->getParent();
224  BasicBlock *Tail = Head->splitBasicBlock(SplitBefore);
225  TerminatorInst *HeadOldTerm = Head->getTerminator();
226  BasicBlock *NewBasicBlock =
227      BasicBlock::Create(*C, "", Head->getParent());
228  BranchInst *HeadNewTerm = BranchInst::Create(/*ifTrue*/NewBasicBlock,
229                                               /*ifFalse*/Tail,
230                                               Cmp);
231  ReplaceInstWithInst(HeadOldTerm, HeadNewTerm);
232
233  BranchInst *CheckTerm = BranchInst::Create(Tail, NewBasicBlock);
234  return CheckTerm;
235}
236
237Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
238  // Shadow >> scale
239  Shadow = IRB.CreateLShr(Shadow, MappingScale);
240  if (MappingOffset == 0)
241    return Shadow;
242  // (Shadow >> scale) | offset
243  return IRB.CreateOr(Shadow, ConstantInt::get(IntptrTy,
244                                               MappingOffset));
245}
246
247void AddressSanitizer::instrumentMemIntrinsicParam(Instruction *OrigIns,
248    Value *Addr, Value *Size, Instruction *InsertBefore, bool IsWrite) {
249  // Check the first byte.
250  {
251    IRBuilder<> IRB(InsertBefore);
252    instrumentAddress(OrigIns, IRB, Addr, 8, IsWrite);
253  }
254  // Check the last byte.
255  {
256    IRBuilder<> IRB(InsertBefore);
257    Value *SizeMinusOne = IRB.CreateSub(
258        Size, ConstantInt::get(Size->getType(), 1));
259    SizeMinusOne = IRB.CreateIntCast(SizeMinusOne, IntptrTy, false);
260    Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
261    Value *AddrPlusSizeMinisOne = IRB.CreateAdd(AddrLong, SizeMinusOne);
262    instrumentAddress(OrigIns, IRB, AddrPlusSizeMinisOne, 8, IsWrite);
263  }
264}
265
266// Instrument memset/memmove/memcpy
267bool AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) {
268  Value *Dst = MI->getDest();
269  MemTransferInst *MemTran = dyn_cast<MemTransferInst>(MI);
270  Value *Src = MemTran ? MemTran->getSource() : NULL;
271  Value *Length = MI->getLength();
272
273  Constant *ConstLength = dyn_cast<Constant>(Length);
274  Instruction *InsertBefore = MI;
275  if (ConstLength) {
276    if (ConstLength->isNullValue()) return false;
277  } else {
278    // The size is not a constant so it could be zero -- check at run-time.
279    IRBuilder<> IRB(InsertBefore);
280
281    Value *Cmp = IRB.CreateICmpNE(Length,
282                                   Constant::getNullValue(Length->getType()));
283    InsertBefore = splitBlockAndInsertIfThen(InsertBefore, Cmp);
284  }
285
286  instrumentMemIntrinsicParam(MI, Dst, Length, InsertBefore, true);
287  if (Src)
288    instrumentMemIntrinsicParam(MI, Src, Length, InsertBefore, false);
289  return true;
290}
291
292static Value *getLDSTOperand(Instruction *I) {
293  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
294    return LI->getPointerOperand();
295  }
296  return cast<StoreInst>(*I).getPointerOperand();
297}
298
299void AddressSanitizer::instrumentMop(Instruction *I) {
300  int IsWrite = isa<StoreInst>(*I);
301  Value *Addr = getLDSTOperand(I);
302  if (ClOpt && ClOptGlobals && isa<GlobalVariable>(Addr)) {
303    // We are accessing a global scalar variable. Nothing to catch here.
304    return;
305  }
306  Type *OrigPtrTy = Addr->getType();
307  Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
308
309  assert(OrigTy->isSized());
310  uint32_t TypeSize = TD->getTypeStoreSizeInBits(OrigTy);
311
312  if (TypeSize != 8  && TypeSize != 16 &&
313      TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
314    // Ignore all unusual sizes.
315    return;
316  }
317
318  IRBuilder<> IRB(I);
319  instrumentAddress(I, IRB, Addr, TypeSize, IsWrite);
320}
321
322Instruction *AddressSanitizer::generateCrashCode(
323    IRBuilder<> &IRB, Value *Addr, bool IsWrite, uint32_t TypeSize) {
324  // IsWrite and TypeSize are encoded in the function name.
325  std::string FunctionName = std::string(kAsanReportErrorTemplate) +
326      (IsWrite ? "store" : "load") + itostr(TypeSize / 8);
327  Value *ReportWarningFunc = CurrentModule->getOrInsertFunction(
328      FunctionName, IRB.getVoidTy(), IntptrTy, NULL);
329  CallInst *Call = IRB.CreateCall(ReportWarningFunc, Addr);
330  Call->setDoesNotReturn();
331  return Call;
332}
333
334void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
335                                         IRBuilder<> &IRB, Value *Addr,
336                                         uint32_t TypeSize, bool IsWrite) {
337  Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
338
339  Type *ShadowTy  = IntegerType::get(
340      *C, std::max(8U, TypeSize >> MappingScale));
341  Type *ShadowPtrTy = PointerType::get(ShadowTy, 0);
342  Value *ShadowPtr = memToShadow(AddrLong, IRB);
343  Value *CmpVal = Constant::getNullValue(ShadowTy);
344  Value *ShadowValue = IRB.CreateLoad(
345      IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy));
346
347  Value *Cmp = IRB.CreateICmpNE(ShadowValue, CmpVal);
348
349  Instruction *CheckTerm = splitBlockAndInsertIfThen(
350      cast<Instruction>(Cmp)->getNextNode(), Cmp);
351  IRBuilder<> IRB2(CheckTerm);
352
353  size_t Granularity = 1 << MappingScale;
354  if (TypeSize < 8 * Granularity) {
355    // Addr & (Granularity - 1)
356    Value *Lower3Bits = IRB2.CreateAnd(
357        AddrLong, ConstantInt::get(IntptrTy, Granularity - 1));
358    // (Addr & (Granularity - 1)) + size - 1
359    Value *LastAccessedByte = IRB2.CreateAdd(
360        Lower3Bits, ConstantInt::get(IntptrTy, TypeSize / 8 - 1));
361    // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
362    LastAccessedByte = IRB2.CreateIntCast(
363        LastAccessedByte, IRB.getInt8Ty(), false);
364    // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
365    Value *Cmp2 = IRB2.CreateICmpSGE(LastAccessedByte, ShadowValue);
366
367    CheckTerm = splitBlockAndInsertIfThen(CheckTerm, Cmp2);
368  }
369
370  IRBuilder<> IRB1(CheckTerm);
371  Instruction *Crash = generateCrashCode(IRB1, AddrLong, IsWrite, TypeSize);
372  Crash->setDebugLoc(OrigIns->getDebugLoc());
373  ReplaceInstWithInst(CheckTerm, new UnreachableInst(*C));
374}
375
376// This function replaces all global variables with new variables that have
377// trailing redzones. It also creates a function that poisons
378// redzones and inserts this function into llvm.global_ctors.
379bool AddressSanitizer::insertGlobalRedzones(Module &M) {
380  SmallVector<GlobalVariable *, 16> GlobalsToChange;
381
382  for (Module::GlobalListType::iterator G = M.getGlobalList().begin(),
383       E = M.getGlobalList().end(); G != E; ++G) {
384    Type *Ty = cast<PointerType>(G->getType())->getElementType();
385    DEBUG(dbgs() << "GLOBAL: " << *G);
386
387    if (!Ty->isSized()) continue;
388    if (!G->hasInitializer()) continue;
389    // Touch only those globals that will not be defined in other modules.
390    // Don't handle ODR type linkages since other modules may be built w/o asan.
391    if (G->getLinkage() != GlobalVariable::ExternalLinkage &&
392        G->getLinkage() != GlobalVariable::PrivateLinkage &&
393        G->getLinkage() != GlobalVariable::InternalLinkage)
394      continue;
395    // Two problems with thread-locals:
396    //   - The address of the main thread's copy can't be computed at link-time.
397    //   - Need to poison all copies, not just the main thread's one.
398    if (G->isThreadLocal())
399      continue;
400    // For now, just ignore this Alloca if the alignment is large.
401    if (G->getAlignment() > RedzoneSize) continue;
402
403    // Ignore all the globals with the names starting with "\01L_OBJC_".
404    // Many of those are put into the .cstring section. The linker compresses
405    // that section by removing the spare \0s after the string terminator, so
406    // our redzones get broken.
407    if ((G->getName().find("\01L_OBJC_") == 0) ||
408        (G->getName().find("\01l_OBJC_") == 0)) {
409      DEBUG(dbgs() << "Ignoring \\01L_OBJC_* global: " << *G);
410      continue;
411    }
412
413    if (G->hasSection()) {
414      StringRef Section(G->getSection());
415      // Ignore the globals from the __OBJC section. The ObjC runtime assumes
416      // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
417      // them.
418      if ((Section.find("__OBJC,") == 0) ||
419          (Section.find("__DATA, __objc_") == 0)) {
420        DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G);
421        continue;
422      }
423      // See http://code.google.com/p/address-sanitizer/issues/detail?id=32
424      // Constant CFString instances are compiled in the following way:
425      //  -- the string buffer is emitted into
426      //     __TEXT,__cstring,cstring_literals
427      //  -- the constant NSConstantString structure referencing that buffer
428      //     is placed into __DATA,__cfstring
429      // Therefore there's no point in placing redzones into __DATA,__cfstring.
430      // Moreover, it causes the linker to crash on OS X 10.7
431      if (Section.find("__DATA,__cfstring") == 0) {
432        DEBUG(dbgs() << "Ignoring CFString: " << *G);
433        continue;
434      }
435    }
436
437    GlobalsToChange.push_back(G);
438  }
439
440  size_t n = GlobalsToChange.size();
441  if (n == 0) return false;
442
443  // A global is described by a structure
444  //   size_t beg;
445  //   size_t size;
446  //   size_t size_with_redzone;
447  //   const char *name;
448  // We initialize an array of such structures and pass it to a run-time call.
449  StructType *GlobalStructTy = StructType::get(IntptrTy, IntptrTy,
450                                               IntptrTy, IntptrTy, NULL);
451  SmallVector<Constant *, 16> Initializers(n);
452
453  IRBuilder<> IRB(CtorInsertBefore);
454
455  for (size_t i = 0; i < n; i++) {
456    GlobalVariable *G = GlobalsToChange[i];
457    PointerType *PtrTy = cast<PointerType>(G->getType());
458    Type *Ty = PtrTy->getElementType();
459    uint64_t SizeInBytes = TD->getTypeAllocSize(Ty);
460    uint64_t RightRedzoneSize = RedzoneSize +
461        (RedzoneSize - (SizeInBytes % RedzoneSize));
462    Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize);
463
464    StructType *NewTy = StructType::get(Ty, RightRedZoneTy, NULL);
465    Constant *NewInitializer = ConstantStruct::get(
466        NewTy, G->getInitializer(),
467        Constant::getNullValue(RightRedZoneTy), NULL);
468
469    SmallString<2048> DescriptionOfGlobal = G->getName();
470    DescriptionOfGlobal += " (";
471    DescriptionOfGlobal += M.getModuleIdentifier();
472    DescriptionOfGlobal += ")";
473    GlobalVariable *Name = createPrivateGlobalForString(M, DescriptionOfGlobal);
474
475    // Create a new global variable with enough space for a redzone.
476    GlobalVariable *NewGlobal = new GlobalVariable(
477        M, NewTy, G->isConstant(), G->getLinkage(),
478        NewInitializer, "", G, G->isThreadLocal());
479    NewGlobal->copyAttributesFrom(G);
480    NewGlobal->setAlignment(RedzoneSize);
481
482    Value *Indices2[2];
483    Indices2[0] = IRB.getInt32(0);
484    Indices2[1] = IRB.getInt32(0);
485
486    G->replaceAllUsesWith(
487        ConstantExpr::getGetElementPtr(NewGlobal, Indices2, true));
488    NewGlobal->takeName(G);
489    G->eraseFromParent();
490
491    Initializers[i] = ConstantStruct::get(
492        GlobalStructTy,
493        ConstantExpr::getPointerCast(NewGlobal, IntptrTy),
494        ConstantInt::get(IntptrTy, SizeInBytes),
495        ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize),
496        ConstantExpr::getPointerCast(Name, IntptrTy),
497        NULL);
498    DEBUG(dbgs() << "NEW GLOBAL:\n" << *NewGlobal);
499  }
500
501  ArrayType *ArrayOfGlobalStructTy = ArrayType::get(GlobalStructTy, n);
502  GlobalVariable *AllGlobals = new GlobalVariable(
503      M, ArrayOfGlobalStructTy, false, GlobalVariable::PrivateLinkage,
504      ConstantArray::get(ArrayOfGlobalStructTy, Initializers), "");
505
506  Function *AsanRegisterGlobals = cast<Function>(M.getOrInsertFunction(
507      kAsanRegisterGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
508  AsanRegisterGlobals->setLinkage(Function::ExternalLinkage);
509
510  IRB.CreateCall2(AsanRegisterGlobals,
511                  IRB.CreatePointerCast(AllGlobals, IntptrTy),
512                  ConstantInt::get(IntptrTy, n));
513
514  // We also need to unregister globals at the end, e.g. when a shared library
515  // gets closed.
516  Function *AsanDtorFunction = Function::Create(
517      FunctionType::get(Type::getVoidTy(*C), false),
518      GlobalValue::InternalLinkage, kAsanModuleDtorName, &M);
519  BasicBlock *AsanDtorBB = BasicBlock::Create(*C, "", AsanDtorFunction);
520  IRBuilder<> IRB_Dtor(ReturnInst::Create(*C, AsanDtorBB));
521  Function *AsanUnregisterGlobals = cast<Function>(M.getOrInsertFunction(
522      kAsanUnregisterGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy, NULL));
523  AsanUnregisterGlobals->setLinkage(Function::ExternalLinkage);
524
525  IRB_Dtor.CreateCall2(AsanUnregisterGlobals,
526                       IRB.CreatePointerCast(AllGlobals, IntptrTy),
527                       ConstantInt::get(IntptrTy, n));
528  appendToGlobalDtors(M, AsanDtorFunction, kAsanCtorAndCtorPriority);
529
530  DEBUG(dbgs() << M);
531  return true;
532}
533
534// virtual
535bool AddressSanitizer::runOnModule(Module &M) {
536  // Initialize the private fields. No one has accessed them before.
537  TD = getAnalysisIfAvailable<TargetData>();
538  if (!TD)
539    return false;
540  BL.reset(new FunctionBlackList(ClBlackListFile));
541
542  CurrentModule = &M;
543  C = &(M.getContext());
544  LongSize = TD->getPointerSizeInBits();
545  IntptrTy = Type::getIntNTy(*C, LongSize);
546  IntptrPtrTy = PointerType::get(IntptrTy, 0);
547
548  AsanCtorFunction = Function::Create(
549      FunctionType::get(Type::getVoidTy(*C), false),
550      GlobalValue::InternalLinkage, kAsanModuleCtorName, &M);
551  BasicBlock *AsanCtorBB = BasicBlock::Create(*C, "", AsanCtorFunction);
552  CtorInsertBefore = ReturnInst::Create(*C, AsanCtorBB);
553
554  // call __asan_init in the module ctor.
555  IRBuilder<> IRB(CtorInsertBefore);
556  AsanInitFunction = cast<Function>(
557      M.getOrInsertFunction(kAsanInitName, IRB.getVoidTy(), NULL));
558  AsanInitFunction->setLinkage(Function::ExternalLinkage);
559  IRB.CreateCall(AsanInitFunction);
560
561  llvm::Triple targetTriple(M.getTargetTriple());
562  bool isAndroid = targetTriple.getEnvironment() == llvm::Triple::ANDROIDEABI;
563
564  MappingOffset = isAndroid ? kDefaultShadowOffsetAndroid :
565    (LongSize == 32 ? kDefaultShadowOffset32 : kDefaultShadowOffset64);
566  if (ClMappingOffsetLog >= 0) {
567    if (ClMappingOffsetLog == 0) {
568      // special case
569      MappingOffset = 0;
570    } else {
571      MappingOffset = 1ULL << ClMappingOffsetLog;
572    }
573  }
574  MappingScale = kDefaultShadowScale;
575  if (ClMappingScale) {
576    MappingScale = ClMappingScale;
577  }
578  // Redzone used for stack and globals is at least 32 bytes.
579  // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
580  RedzoneSize = std::max(32, (int)(1 << MappingScale));
581
582  bool Res = false;
583
584  if (ClGlobals)
585    Res |= insertGlobalRedzones(M);
586
587  if (ClMappingOffsetLog >= 0) {
588    // Tell the run-time the current values of mapping offset and scale.
589    GlobalValue *asan_mapping_offset =
590        new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
591                       ConstantInt::get(IntptrTy, MappingOffset),
592                       kAsanMappingOffsetName);
593    // Read the global, otherwise it may be optimized away.
594    IRB.CreateLoad(asan_mapping_offset, true);
595  }
596  if (ClMappingScale) {
597    GlobalValue *asan_mapping_scale =
598        new GlobalVariable(M, IntptrTy, true, GlobalValue::LinkOnceODRLinkage,
599                           ConstantInt::get(IntptrTy, MappingScale),
600                           kAsanMappingScaleName);
601    // Read the global, otherwise it may be optimized away.
602    IRB.CreateLoad(asan_mapping_scale, true);
603  }
604
605
606  for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
607    if (F->isDeclaration()) continue;
608    Res |= handleFunction(M, *F);
609  }
610
611  appendToGlobalCtors(M, AsanCtorFunction, kAsanCtorAndCtorPriority);
612
613  return Res;
614}
615
616bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
617  // For each NSObject descendant having a +load method, this method is invoked
618  // by the ObjC runtime before any of the static constructors is called.
619  // Therefore we need to instrument such methods with a call to __asan_init
620  // at the beginning in order to initialize our runtime before any access to
621  // the shadow memory.
622  // We cannot just ignore these methods, because they may call other
623  // instrumented functions.
624  if (F.getName().find(" load]") != std::string::npos) {
625    IRBuilder<> IRB(F.begin()->begin());
626    IRB.CreateCall(AsanInitFunction);
627    return true;
628  }
629  return false;
630}
631
632bool AddressSanitizer::handleFunction(Module &M, Function &F) {
633  if (BL->isIn(F)) return false;
634  if (&F == AsanCtorFunction) return false;
635
636  // If needed, insert __asan_init before checking for AddressSafety attr.
637  maybeInsertAsanInitAtFunctionEntry(F);
638
639  if (!F.hasFnAttr(Attribute::AddressSafety)) return false;
640
641  if (!ClDebugFunc.empty() && ClDebugFunc != F.getName())
642    return false;
643  // We want to instrument every address only once per basic block
644  // (unless there are calls between uses).
645  SmallSet<Value*, 16> TempsToInstrument;
646  SmallVector<Instruction*, 16> ToInstrument;
647  SmallVector<Instruction*, 8> NoReturnCalls;
648
649  // Fill the set of memory operations to instrument.
650  for (Function::iterator FI = F.begin(), FE = F.end();
651       FI != FE; ++FI) {
652    TempsToInstrument.clear();
653    for (BasicBlock::iterator BI = FI->begin(), BE = FI->end();
654         BI != BE; ++BI) {
655      if (LooksLikeCodeInBug11395(BI)) return false;
656      if ((isa<LoadInst>(BI) && ClInstrumentReads) ||
657          (isa<StoreInst>(BI) && ClInstrumentWrites)) {
658        Value *Addr = getLDSTOperand(BI);
659        if (ClOpt && ClOptSameTemp) {
660          if (!TempsToInstrument.insert(Addr))
661            continue;  // We've seen this temp in the current BB.
662        }
663      } else if (isa<MemIntrinsic>(BI) && ClMemIntrin) {
664        // ok, take it.
665      } else {
666        if (CallInst *CI = dyn_cast<CallInst>(BI)) {
667          // A call inside BB.
668          TempsToInstrument.clear();
669          if (CI->doesNotReturn()) {
670            NoReturnCalls.push_back(CI);
671          }
672        }
673        continue;
674      }
675      ToInstrument.push_back(BI);
676    }
677  }
678
679  // Instrument.
680  int NumInstrumented = 0;
681  for (size_t i = 0, n = ToInstrument.size(); i != n; i++) {
682    Instruction *Inst = ToInstrument[i];
683    if (ClDebugMin < 0 || ClDebugMax < 0 ||
684        (NumInstrumented >= ClDebugMin && NumInstrumented <= ClDebugMax)) {
685      if (isa<StoreInst>(Inst) || isa<LoadInst>(Inst))
686        instrumentMop(Inst);
687      else
688        instrumentMemIntrinsic(cast<MemIntrinsic>(Inst));
689    }
690    NumInstrumented++;
691  }
692
693  DEBUG(dbgs() << F);
694
695  bool ChangedStack = poisonStackInFunction(M, F);
696
697  // We must unpoison the stack before every NoReturn call (throw, _exit, etc).
698  // See e.g. http://code.google.com/p/address-sanitizer/issues/detail?id=37
699  for (size_t i = 0, n = NoReturnCalls.size(); i != n; i++) {
700    Instruction *CI = NoReturnCalls[i];
701    IRBuilder<> IRB(CI);
702    IRB.CreateCall(M.getOrInsertFunction(kAsanHandleNoReturnName,
703                                         IRB.getVoidTy(), NULL));
704  }
705
706  return NumInstrumented > 0 || ChangedStack || !NoReturnCalls.empty();
707}
708
709static uint64_t ValueForPoison(uint64_t PoisonByte, size_t ShadowRedzoneSize) {
710  if (ShadowRedzoneSize == 1) return PoisonByte;
711  if (ShadowRedzoneSize == 2) return (PoisonByte << 8) + PoisonByte;
712  if (ShadowRedzoneSize == 4)
713    return (PoisonByte << 24) + (PoisonByte << 16) +
714        (PoisonByte << 8) + (PoisonByte);
715  llvm_unreachable("ShadowRedzoneSize is either 1, 2 or 4");
716}
717
718static void PoisonShadowPartialRightRedzone(uint8_t *Shadow,
719                                            size_t Size,
720                                            size_t RedzoneSize,
721                                            size_t ShadowGranularity,
722                                            uint8_t Magic) {
723  for (size_t i = 0; i < RedzoneSize;
724       i+= ShadowGranularity, Shadow++) {
725    if (i + ShadowGranularity <= Size) {
726      *Shadow = 0;  // fully addressable
727    } else if (i >= Size) {
728      *Shadow = Magic;  // unaddressable
729    } else {
730      *Shadow = Size - i;  // first Size-i bytes are addressable
731    }
732  }
733}
734
735void AddressSanitizer::PoisonStack(const ArrayRef<AllocaInst*> &AllocaVec,
736                                   IRBuilder<> IRB,
737                                   Value *ShadowBase, bool DoPoison) {
738  size_t ShadowRZSize = RedzoneSize >> MappingScale;
739  assert(ShadowRZSize >= 1 && ShadowRZSize <= 4);
740  Type *RZTy = Type::getIntNTy(*C, ShadowRZSize * 8);
741  Type *RZPtrTy = PointerType::get(RZTy, 0);
742
743  Value *PoisonLeft  = ConstantInt::get(RZTy,
744    ValueForPoison(DoPoison ? kAsanStackLeftRedzoneMagic : 0LL, ShadowRZSize));
745  Value *PoisonMid   = ConstantInt::get(RZTy,
746    ValueForPoison(DoPoison ? kAsanStackMidRedzoneMagic : 0LL, ShadowRZSize));
747  Value *PoisonRight = ConstantInt::get(RZTy,
748    ValueForPoison(DoPoison ? kAsanStackRightRedzoneMagic : 0LL, ShadowRZSize));
749
750  // poison the first red zone.
751  IRB.CreateStore(PoisonLeft, IRB.CreateIntToPtr(ShadowBase, RZPtrTy));
752
753  // poison all other red zones.
754  uint64_t Pos = RedzoneSize;
755  for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
756    AllocaInst *AI = AllocaVec[i];
757    uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
758    uint64_t AlignedSize = getAlignedAllocaSize(AI);
759    assert(AlignedSize - SizeInBytes < RedzoneSize);
760    Value *Ptr = NULL;
761
762    Pos += AlignedSize;
763
764    assert(ShadowBase->getType() == IntptrTy);
765    if (SizeInBytes < AlignedSize) {
766      // Poison the partial redzone at right
767      Ptr = IRB.CreateAdd(
768          ShadowBase, ConstantInt::get(IntptrTy,
769                                       (Pos >> MappingScale) - ShadowRZSize));
770      size_t AddressableBytes = RedzoneSize - (AlignedSize - SizeInBytes);
771      uint32_t Poison = 0;
772      if (DoPoison) {
773        PoisonShadowPartialRightRedzone((uint8_t*)&Poison, AddressableBytes,
774                                        RedzoneSize,
775                                        1ULL << MappingScale,
776                                        kAsanStackPartialRedzoneMagic);
777      }
778      Value *PartialPoison = ConstantInt::get(RZTy, Poison);
779      IRB.CreateStore(PartialPoison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
780    }
781
782    // Poison the full redzone at right.
783    Ptr = IRB.CreateAdd(ShadowBase,
784                        ConstantInt::get(IntptrTy, Pos >> MappingScale));
785    Value *Poison = i == AllocaVec.size() - 1 ? PoisonRight : PoisonMid;
786    IRB.CreateStore(Poison, IRB.CreateIntToPtr(Ptr, RZPtrTy));
787
788    Pos += RedzoneSize;
789  }
790}
791
792// Workaround for bug 11395: we don't want to instrument stack in functions
793// with large assembly blobs (32-bit only), otherwise reg alloc may crash.
794// FIXME: remove once the bug 11395 is fixed.
795bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
796  if (LongSize != 32) return false;
797  CallInst *CI = dyn_cast<CallInst>(I);
798  if (!CI || !CI->isInlineAsm()) return false;
799  if (CI->getNumArgOperands() <= 5) return false;
800  // We have inline assembly with quite a few arguments.
801  return true;
802}
803
804// Find all static Alloca instructions and put
805// poisoned red zones around all of them.
806// Then unpoison everything back before the function returns.
807//
808// Stack poisoning does not play well with exception handling.
809// When an exception is thrown, we essentially bypass the code
810// that unpoisones the stack. This is why the run-time library has
811// to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
812// stack in the interceptor. This however does not work inside the
813// actual function which catches the exception. Most likely because the
814// compiler hoists the load of the shadow value somewhere too high.
815// This causes asan to report a non-existing bug on 453.povray.
816// It sounds like an LLVM bug.
817bool AddressSanitizer::poisonStackInFunction(Module &M, Function &F) {
818  if (!ClStack) return false;
819  SmallVector<AllocaInst*, 16> AllocaVec;
820  SmallVector<Instruction*, 8> RetVec;
821  uint64_t TotalSize = 0;
822
823  // Filter out Alloca instructions we want (and can) handle.
824  // Collect Ret instructions.
825  for (Function::iterator FI = F.begin(), FE = F.end();
826       FI != FE; ++FI) {
827    BasicBlock &BB = *FI;
828    for (BasicBlock::iterator BI = BB.begin(), BE = BB.end();
829         BI != BE; ++BI) {
830      if (isa<ReturnInst>(BI)) {
831          RetVec.push_back(BI);
832          continue;
833      }
834
835      AllocaInst *AI = dyn_cast<AllocaInst>(BI);
836      if (!AI) continue;
837      if (AI->isArrayAllocation()) continue;
838      if (!AI->isStaticAlloca()) continue;
839      if (!AI->getAllocatedType()->isSized()) continue;
840      if (AI->getAlignment() > RedzoneSize) continue;
841      AllocaVec.push_back(AI);
842      uint64_t AlignedSize =  getAlignedAllocaSize(AI);
843      TotalSize += AlignedSize;
844    }
845  }
846
847  if (AllocaVec.empty()) return false;
848
849  uint64_t LocalStackSize = TotalSize + (AllocaVec.size() + 1) * RedzoneSize;
850
851  bool DoStackMalloc = ClUseAfterReturn
852      && LocalStackSize <= kMaxStackMallocSize;
853
854  Instruction *InsBefore = AllocaVec[0];
855  IRBuilder<> IRB(InsBefore);
856
857
858  Type *ByteArrayTy = ArrayType::get(IRB.getInt8Ty(), LocalStackSize);
859  AllocaInst *MyAlloca =
860      new AllocaInst(ByteArrayTy, "MyAlloca", InsBefore);
861  MyAlloca->setAlignment(RedzoneSize);
862  assert(MyAlloca->isStaticAlloca());
863  Value *OrigStackBase = IRB.CreatePointerCast(MyAlloca, IntptrTy);
864  Value *LocalStackBase = OrigStackBase;
865
866  if (DoStackMalloc) {
867    Value *AsanStackMallocFunc = M.getOrInsertFunction(
868        kAsanStackMallocName, IntptrTy, IntptrTy, IntptrTy, NULL);
869    LocalStackBase = IRB.CreateCall2(AsanStackMallocFunc,
870        ConstantInt::get(IntptrTy, LocalStackSize), OrigStackBase);
871  }
872
873  // This string will be parsed by the run-time (DescribeStackAddress).
874  SmallString<2048> StackDescriptionStorage;
875  raw_svector_ostream StackDescription(StackDescriptionStorage);
876  StackDescription << F.getName() << " " << AllocaVec.size() << " ";
877
878  uint64_t Pos = RedzoneSize;
879  // Replace Alloca instructions with base+offset.
880  for (size_t i = 0, n = AllocaVec.size(); i < n; i++) {
881    AllocaInst *AI = AllocaVec[i];
882    uint64_t SizeInBytes = getAllocaSizeInBytes(AI);
883    StringRef Name = AI->getName();
884    StackDescription << Pos << " " << SizeInBytes << " "
885                     << Name.size() << " " << Name << " ";
886    uint64_t AlignedSize = getAlignedAllocaSize(AI);
887    assert((AlignedSize % RedzoneSize) == 0);
888    AI->replaceAllUsesWith(
889        IRB.CreateIntToPtr(
890            IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, Pos)),
891            AI->getType()));
892    Pos += AlignedSize + RedzoneSize;
893  }
894  assert(Pos == LocalStackSize);
895
896  // Write the Magic value and the frame description constant to the redzone.
897  Value *BasePlus0 = IRB.CreateIntToPtr(LocalStackBase, IntptrPtrTy);
898  IRB.CreateStore(ConstantInt::get(IntptrTy, kCurrentStackFrameMagic),
899                  BasePlus0);
900  Value *BasePlus1 = IRB.CreateAdd(LocalStackBase,
901                                   ConstantInt::get(IntptrTy, LongSize/8));
902  BasePlus1 = IRB.CreateIntToPtr(BasePlus1, IntptrPtrTy);
903  Value *Description = IRB.CreatePointerCast(
904      createPrivateGlobalForString(M, StackDescription.str()),
905      IntptrTy);
906  IRB.CreateStore(Description, BasePlus1);
907
908  // Poison the stack redzones at the entry.
909  Value *ShadowBase = memToShadow(LocalStackBase, IRB);
910  PoisonStack(ArrayRef<AllocaInst*>(AllocaVec), IRB, ShadowBase, true);
911
912  Value *AsanStackFreeFunc = NULL;
913  if (DoStackMalloc) {
914    AsanStackFreeFunc = M.getOrInsertFunction(
915        kAsanStackFreeName, IRB.getVoidTy(),
916        IntptrTy, IntptrTy, IntptrTy, NULL);
917  }
918
919  // Unpoison the stack before all ret instructions.
920  for (size_t i = 0, n = RetVec.size(); i < n; i++) {
921    Instruction *Ret = RetVec[i];
922    IRBuilder<> IRBRet(Ret);
923
924    // Mark the current frame as retired.
925    IRBRet.CreateStore(ConstantInt::get(IntptrTy, kRetiredStackFrameMagic),
926                       BasePlus0);
927    // Unpoison the stack.
928    PoisonStack(ArrayRef<AllocaInst*>(AllocaVec), IRBRet, ShadowBase, false);
929
930    if (DoStackMalloc) {
931      IRBRet.CreateCall3(AsanStackFreeFunc, LocalStackBase,
932                         ConstantInt::get(IntptrTy, LocalStackSize),
933                         OrigStackBase);
934    }
935  }
936
937  if (ClDebugStack) {
938    DEBUG(dbgs() << F);
939  }
940
941  return true;
942}
943