ThreadSanitizer.cpp revision 288943 
1//===-- ThreadSanitizer.cpp - race detector -------------------------------===//
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 ThreadSanitizer, a race detector.
11//
12// The tool is under development, for the details about previous versions see
13// http://code.google.com/p/data-race-test
14//
15// The instrumentation phase is quite simple:
16//   - Insert calls to run-time library before every memory access.
17//      - Optimizations may apply to avoid instrumenting some of the accesses.
18//   - Insert calls at function entry/exit.
19// The rest is handled by the run-time library.
20//===----------------------------------------------------------------------===//
21
22#include "llvm/Transforms/Instrumentation.h"
23#include "llvm/ADT/SmallSet.h"
24#include "llvm/ADT/SmallString.h"
25#include "llvm/ADT/SmallVector.h"
26#include "llvm/ADT/Statistic.h"
27#include "llvm/ADT/StringExtras.h"
28#include "llvm/Analysis/CaptureTracking.h"
29#include "llvm/Analysis/ValueTracking.h"
30#include "llvm/IR/DataLayout.h"
31#include "llvm/IR/Function.h"
32#include "llvm/IR/IRBuilder.h"
33#include "llvm/IR/IntrinsicInst.h"
34#include "llvm/IR/Intrinsics.h"
35#include "llvm/IR/LLVMContext.h"
36#include "llvm/IR/Metadata.h"
37#include "llvm/IR/Module.h"
38#include "llvm/IR/Type.h"
39#include "llvm/Support/CommandLine.h"
40#include "llvm/Support/Debug.h"
41#include "llvm/Support/MathExtras.h"
42#include "llvm/Support/raw_ostream.h"
43#include "llvm/Transforms/Utils/BasicBlockUtils.h"
44#include "llvm/Transforms/Utils/ModuleUtils.h"
45
46using namespace llvm;
47
48#define DEBUG_TYPE "tsan"
49
50static cl::opt<bool>  ClInstrumentMemoryAccesses(
51    "tsan-instrument-memory-accesses", cl::init(true),
52    cl::desc("Instrument memory accesses"), cl::Hidden);
53static cl::opt<bool>  ClInstrumentFuncEntryExit(
54    "tsan-instrument-func-entry-exit", cl::init(true),
55    cl::desc("Instrument function entry and exit"), cl::Hidden);
56static cl::opt<bool>  ClInstrumentAtomics(
57    "tsan-instrument-atomics", cl::init(true),
58    cl::desc("Instrument atomics"), cl::Hidden);
59static cl::opt<bool>  ClInstrumentMemIntrinsics(
60    "tsan-instrument-memintrinsics", cl::init(true),
61    cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
62
63STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
64STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
65STATISTIC(NumOmittedReadsBeforeWrite,
66          "Number of reads ignored due to following writes");
67STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
68STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
69STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
70STATISTIC(NumOmittedReadsFromConstantGlobals,
71          "Number of reads from constant globals");
72STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
73STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");
74
75static const char *const kTsanModuleCtorName = "tsan.module_ctor";
76static const char *const kTsanInitName = "__tsan_init";
77
78namespace {
79
80/// ThreadSanitizer: instrument the code in module to find races.
81struct ThreadSanitizer : public FunctionPass {
82  ThreadSanitizer() : FunctionPass(ID) {}
83  const char *getPassName() const override;
84  bool runOnFunction(Function &F) override;
85  bool doInitialization(Module &M) override;
86  static char ID;  // Pass identification, replacement for typeid.
87
88 private:
89  void initializeCallbacks(Module &M);
90  bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL);
91  bool instrumentAtomic(Instruction *I, const DataLayout &DL);
92  bool instrumentMemIntrinsic(Instruction *I);
93  void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
94                                      SmallVectorImpl<Instruction *> &All,
95                                      const DataLayout &DL);
96  bool addrPointsToConstantData(Value *Addr);
97  int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL);
98
99  Type *IntptrTy;
100  IntegerType *OrdTy;
101  // Callbacks to run-time library are computed in doInitialization.
102  Function *TsanFuncEntry;
103  Function *TsanFuncExit;
104  // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
105  static const size_t kNumberOfAccessSizes = 5;
106  Function *TsanRead[kNumberOfAccessSizes];
107  Function *TsanWrite[kNumberOfAccessSizes];
108  Function *TsanUnalignedRead[kNumberOfAccessSizes];
109  Function *TsanUnalignedWrite[kNumberOfAccessSizes];
110  Function *TsanAtomicLoad[kNumberOfAccessSizes];
111  Function *TsanAtomicStore[kNumberOfAccessSizes];
112  Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes];
113  Function *TsanAtomicCAS[kNumberOfAccessSizes];
114  Function *TsanAtomicThreadFence;
115  Function *TsanAtomicSignalFence;
116  Function *TsanVptrUpdate;
117  Function *TsanVptrLoad;
118  Function *MemmoveFn, *MemcpyFn, *MemsetFn;
119  Function *TsanCtorFunction;
120};
121}  // namespace
122
123char ThreadSanitizer::ID = 0;
124INITIALIZE_PASS(ThreadSanitizer, "tsan",
125    "ThreadSanitizer: detects data races.",
126    false, false)
127
128const char *ThreadSanitizer::getPassName() const {
129  return "ThreadSanitizer";
130}
131
132FunctionPass *llvm::createThreadSanitizerPass() {
133  return new ThreadSanitizer();
134}
135
136void ThreadSanitizer::initializeCallbacks(Module &M) {
137  IRBuilder<> IRB(M.getContext());
138  // Initialize the callbacks.
139  TsanFuncEntry = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
140      "__tsan_func_entry", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
141  TsanFuncExit = checkSanitizerInterfaceFunction(
142      M.getOrInsertFunction("__tsan_func_exit", IRB.getVoidTy(), nullptr));
143  OrdTy = IRB.getInt32Ty();
144  for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
145    const size_t ByteSize = 1 << i;
146    const size_t BitSize = ByteSize * 8;
147    SmallString<32> ReadName("__tsan_read" + itostr(ByteSize));
148    TsanRead[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
149        ReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
150
151    SmallString<32> WriteName("__tsan_write" + itostr(ByteSize));
152    TsanWrite[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
153        WriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
154
155    SmallString<64> UnalignedReadName("__tsan_unaligned_read" +
156        itostr(ByteSize));
157    TsanUnalignedRead[i] =
158        checkSanitizerInterfaceFunction(M.getOrInsertFunction(
159            UnalignedReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
160
161    SmallString<64> UnalignedWriteName("__tsan_unaligned_write" +
162        itostr(ByteSize));
163    TsanUnalignedWrite[i] =
164        checkSanitizerInterfaceFunction(M.getOrInsertFunction(
165            UnalignedWriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
166
167    Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
168    Type *PtrTy = Ty->getPointerTo();
169    SmallString<32> AtomicLoadName("__tsan_atomic" + itostr(BitSize) +
170                                   "_load");
171    TsanAtomicLoad[i] = checkSanitizerInterfaceFunction(
172        M.getOrInsertFunction(AtomicLoadName, Ty, PtrTy, OrdTy, nullptr));
173
174    SmallString<32> AtomicStoreName("__tsan_atomic" + itostr(BitSize) +
175                                    "_store");
176    TsanAtomicStore[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
177        AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy, nullptr));
178
179    for (int op = AtomicRMWInst::FIRST_BINOP;
180        op <= AtomicRMWInst::LAST_BINOP; ++op) {
181      TsanAtomicRMW[op][i] = nullptr;
182      const char *NamePart = nullptr;
183      if (op == AtomicRMWInst::Xchg)
184        NamePart = "_exchange";
185      else if (op == AtomicRMWInst::Add)
186        NamePart = "_fetch_add";
187      else if (op == AtomicRMWInst::Sub)
188        NamePart = "_fetch_sub";
189      else if (op == AtomicRMWInst::And)
190        NamePart = "_fetch_and";
191      else if (op == AtomicRMWInst::Or)
192        NamePart = "_fetch_or";
193      else if (op == AtomicRMWInst::Xor)
194        NamePart = "_fetch_xor";
195      else if (op == AtomicRMWInst::Nand)
196        NamePart = "_fetch_nand";
197      else
198        continue;
199      SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
200      TsanAtomicRMW[op][i] = checkSanitizerInterfaceFunction(
201          M.getOrInsertFunction(RMWName, Ty, PtrTy, Ty, OrdTy, nullptr));
202    }
203
204    SmallString<32> AtomicCASName("__tsan_atomic" + itostr(BitSize) +
205                                  "_compare_exchange_val");
206    TsanAtomicCAS[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
207        AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, nullptr));
208  }
209  TsanVptrUpdate = checkSanitizerInterfaceFunction(
210      M.getOrInsertFunction("__tsan_vptr_update", IRB.getVoidTy(),
211                            IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), nullptr));
212  TsanVptrLoad = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
213      "__tsan_vptr_read", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
214  TsanAtomicThreadFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
215      "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, nullptr));
216  TsanAtomicSignalFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
217      "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, nullptr));
218
219  MemmoveFn = checkSanitizerInterfaceFunction(
220      M.getOrInsertFunction("memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
221                            IRB.getInt8PtrTy(), IntptrTy, nullptr));
222  MemcpyFn = checkSanitizerInterfaceFunction(
223      M.getOrInsertFunction("memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
224                            IRB.getInt8PtrTy(), IntptrTy, nullptr));
225  MemsetFn = checkSanitizerInterfaceFunction(
226      M.getOrInsertFunction("memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
227                            IRB.getInt32Ty(), IntptrTy, nullptr));
228}
229
230bool ThreadSanitizer::doInitialization(Module &M) {
231  const DataLayout &DL = M.getDataLayout();
232  IntptrTy = DL.getIntPtrType(M.getContext());
233  std::tie(TsanCtorFunction, std::ignore) = createSanitizerCtorAndInitFunctions(
234      M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{},
235      /*InitArgs=*/{});
236
237  appendToGlobalCtors(M, TsanCtorFunction, 0);
238
239  return true;
240}
241
242static bool isVtableAccess(Instruction *I) {
243  if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
244    return Tag->isTBAAVtableAccess();
245  return false;
246}
247
248bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
249  // If this is a GEP, just analyze its pointer operand.
250  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
251    Addr = GEP->getPointerOperand();
252
253  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
254    if (GV->isConstant()) {
255      // Reads from constant globals can not race with any writes.
256      NumOmittedReadsFromConstantGlobals++;
257      return true;
258    }
259  } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
260    if (isVtableAccess(L)) {
261      // Reads from a vtable pointer can not race with any writes.
262      NumOmittedReadsFromVtable++;
263      return true;
264    }
265  }
266  return false;
267}
268
269// Instrumenting some of the accesses may be proven redundant.
270// Currently handled:
271//  - read-before-write (within same BB, no calls between)
272//  - not captured variables
273//
274// We do not handle some of the patterns that should not survive
275// after the classic compiler optimizations.
276// E.g. two reads from the same temp should be eliminated by CSE,
277// two writes should be eliminated by DSE, etc.
278//
279// 'Local' is a vector of insns within the same BB (no calls between).
280// 'All' is a vector of insns that will be instrumented.
281void ThreadSanitizer::chooseInstructionsToInstrument(
282    SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
283    const DataLayout &DL) {
284  SmallSet<Value*, 8> WriteTargets;
285  // Iterate from the end.
286  for (SmallVectorImpl<Instruction*>::reverse_iterator It = Local.rbegin(),
287       E = Local.rend(); It != E; ++It) {
288    Instruction *I = *It;
289    if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
290      WriteTargets.insert(Store->getPointerOperand());
291    } else {
292      LoadInst *Load = cast<LoadInst>(I);
293      Value *Addr = Load->getPointerOperand();
294      if (WriteTargets.count(Addr)) {
295        // We will write to this temp, so no reason to analyze the read.
296        NumOmittedReadsBeforeWrite++;
297        continue;
298      }
299      if (addrPointsToConstantData(Addr)) {
300        // Addr points to some constant data -- it can not race with any writes.
301        continue;
302      }
303    }
304    Value *Addr = isa<StoreInst>(*I)
305        ? cast<StoreInst>(I)->getPointerOperand()
306        : cast<LoadInst>(I)->getPointerOperand();
307    if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) &&
308        !PointerMayBeCaptured(Addr, true, true)) {
309      // The variable is addressable but not captured, so it cannot be
310      // referenced from a different thread and participate in a data race
311      // (see llvm/Analysis/CaptureTracking.h for details).
312      NumOmittedNonCaptured++;
313      continue;
314    }
315    All.push_back(I);
316  }
317  Local.clear();
318}
319
320static bool isAtomic(Instruction *I) {
321  if (LoadInst *LI = dyn_cast<LoadInst>(I))
322    return LI->isAtomic() && LI->getSynchScope() == CrossThread;
323  if (StoreInst *SI = dyn_cast<StoreInst>(I))
324    return SI->isAtomic() && SI->getSynchScope() == CrossThread;
325  if (isa<AtomicRMWInst>(I))
326    return true;
327  if (isa<AtomicCmpXchgInst>(I))
328    return true;
329  if (isa<FenceInst>(I))
330    return true;
331  return false;
332}
333
334bool ThreadSanitizer::runOnFunction(Function &F) {
335  // This is required to prevent instrumenting call to __tsan_init from within
336  // the module constructor.
337  if (&F == TsanCtorFunction)
338    return false;
339  initializeCallbacks(*F.getParent());
340  SmallVector<Instruction*, 8> RetVec;
341  SmallVector<Instruction*, 8> AllLoadsAndStores;
342  SmallVector<Instruction*, 8> LocalLoadsAndStores;
343  SmallVector<Instruction*, 8> AtomicAccesses;
344  SmallVector<Instruction*, 8> MemIntrinCalls;
345  bool Res = false;
346  bool HasCalls = false;
347  bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
348  const DataLayout &DL = F.getParent()->getDataLayout();
349
350  // Traverse all instructions, collect loads/stores/returns, check for calls.
351  for (auto &BB : F) {
352    for (auto &Inst : BB) {
353      if (isAtomic(&Inst))
354        AtomicAccesses.push_back(&Inst);
355      else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
356        LocalLoadsAndStores.push_back(&Inst);
357      else if (isa<ReturnInst>(Inst))
358        RetVec.push_back(&Inst);
359      else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
360        if (isa<MemIntrinsic>(Inst))
361          MemIntrinCalls.push_back(&Inst);
362        HasCalls = true;
363        chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
364                                       DL);
365      }
366    }
367    chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
368  }
369
370  // We have collected all loads and stores.
371  // FIXME: many of these accesses do not need to be checked for races
372  // (e.g. variables that do not escape, etc).
373
374  // Instrument memory accesses only if we want to report bugs in the function.
375  if (ClInstrumentMemoryAccesses && SanitizeFunction)
376    for (auto Inst : AllLoadsAndStores) {
377      Res |= instrumentLoadOrStore(Inst, DL);
378    }
379
380  // Instrument atomic memory accesses in any case (they can be used to
381  // implement synchronization).
382  if (ClInstrumentAtomics)
383    for (auto Inst : AtomicAccesses) {
384      Res |= instrumentAtomic(Inst, DL);
385    }
386
387  if (ClInstrumentMemIntrinsics && SanitizeFunction)
388    for (auto Inst : MemIntrinCalls) {
389      Res |= instrumentMemIntrinsic(Inst);
390    }
391
392  // Instrument function entry/exit points if there were instrumented accesses.
393  if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
394    IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
395    Value *ReturnAddress = IRB.CreateCall(
396        Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
397        IRB.getInt32(0));
398    IRB.CreateCall(TsanFuncEntry, ReturnAddress);
399    for (auto RetInst : RetVec) {
400      IRBuilder<> IRBRet(RetInst);
401      IRBRet.CreateCall(TsanFuncExit, {});
402    }
403    Res = true;
404  }
405  return Res;
406}
407
408bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I,
409                                            const DataLayout &DL) {
410  IRBuilder<> IRB(I);
411  bool IsWrite = isa<StoreInst>(*I);
412  Value *Addr = IsWrite
413      ? cast<StoreInst>(I)->getPointerOperand()
414      : cast<LoadInst>(I)->getPointerOperand();
415  int Idx = getMemoryAccessFuncIndex(Addr, DL);
416  if (Idx < 0)
417    return false;
418  if (IsWrite && isVtableAccess(I)) {
419    DEBUG(dbgs() << "  VPTR : " << *I << "\n");
420    Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
421    // StoredValue may be a vector type if we are storing several vptrs at once.
422    // In this case, just take the first element of the vector since this is
423    // enough to find vptr races.
424    if (isa<VectorType>(StoredValue->getType()))
425      StoredValue = IRB.CreateExtractElement(
426          StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
427    if (StoredValue->getType()->isIntegerTy())
428      StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
429    // Call TsanVptrUpdate.
430    IRB.CreateCall(TsanVptrUpdate,
431                   {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
432                    IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
433    NumInstrumentedVtableWrites++;
434    return true;
435  }
436  if (!IsWrite && isVtableAccess(I)) {
437    IRB.CreateCall(TsanVptrLoad,
438                   IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
439    NumInstrumentedVtableReads++;
440    return true;
441  }
442  const unsigned Alignment = IsWrite
443      ? cast<StoreInst>(I)->getAlignment()
444      : cast<LoadInst>(I)->getAlignment();
445  Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
446  const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
447  Value *OnAccessFunc = nullptr;
448  if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0)
449    OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
450  else
451    OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
452  IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
453  if (IsWrite) NumInstrumentedWrites++;
454  else         NumInstrumentedReads++;
455  return true;
456}
457
458static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
459  uint32_t v = 0;
460  switch (ord) {
461    case NotAtomic: llvm_unreachable("unexpected atomic ordering!");
462    case Unordered:              // Fall-through.
463    case Monotonic:              v = 0; break;
464    // case Consume:                v = 1; break;  // Not specified yet.
465    case Acquire:                v = 2; break;
466    case Release:                v = 3; break;
467    case AcquireRelease:         v = 4; break;
468    case SequentiallyConsistent: v = 5; break;
469  }
470  return IRB->getInt32(v);
471}
472
473// If a memset intrinsic gets inlined by the code gen, we will miss races on it.
474// So, we either need to ensure the intrinsic is not inlined, or instrument it.
475// We do not instrument memset/memmove/memcpy intrinsics (too complicated),
476// instead we simply replace them with regular function calls, which are then
477// intercepted by the run-time.
478// Since tsan is running after everyone else, the calls should not be
479// replaced back with intrinsics. If that becomes wrong at some point,
480// we will need to call e.g. __tsan_memset to avoid the intrinsics.
481bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
482  IRBuilder<> IRB(I);
483  if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
484    IRB.CreateCall(
485        MemsetFn,
486        {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
487         IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
488         IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
489    I->eraseFromParent();
490  } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
491    IRB.CreateCall(
492        isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
493        {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
494         IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
495         IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
496    I->eraseFromParent();
497  }
498  return false;
499}
500
501// Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
502// standards.  For background see C++11 standard.  A slightly older, publicly
503// available draft of the standard (not entirely up-to-date, but close enough
504// for casual browsing) is available here:
505// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
506// The following page contains more background information:
507// http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
508
509bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
510  IRBuilder<> IRB(I);
511  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
512    Value *Addr = LI->getPointerOperand();
513    int Idx = getMemoryAccessFuncIndex(Addr, DL);
514    if (Idx < 0)
515      return false;
516    const size_t ByteSize = 1 << Idx;
517    const size_t BitSize = ByteSize * 8;
518    Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
519    Type *PtrTy = Ty->getPointerTo();
520    Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
521                     createOrdering(&IRB, LI->getOrdering())};
522    CallInst *C = CallInst::Create(TsanAtomicLoad[Idx], Args);
523    ReplaceInstWithInst(I, C);
524
525  } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
526    Value *Addr = SI->getPointerOperand();
527    int Idx = getMemoryAccessFuncIndex(Addr, DL);
528    if (Idx < 0)
529      return false;
530    const size_t ByteSize = 1 << Idx;
531    const size_t BitSize = ByteSize * 8;
532    Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
533    Type *PtrTy = Ty->getPointerTo();
534    Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
535                     IRB.CreateIntCast(SI->getValueOperand(), Ty, false),
536                     createOrdering(&IRB, SI->getOrdering())};
537    CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
538    ReplaceInstWithInst(I, C);
539  } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
540    Value *Addr = RMWI->getPointerOperand();
541    int Idx = getMemoryAccessFuncIndex(Addr, DL);
542    if (Idx < 0)
543      return false;
544    Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
545    if (!F)
546      return false;
547    const size_t ByteSize = 1 << Idx;
548    const size_t BitSize = ByteSize * 8;
549    Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
550    Type *PtrTy = Ty->getPointerTo();
551    Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
552                     IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
553                     createOrdering(&IRB, RMWI->getOrdering())};
554    CallInst *C = CallInst::Create(F, Args);
555    ReplaceInstWithInst(I, C);
556  } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
557    Value *Addr = CASI->getPointerOperand();
558    int Idx = getMemoryAccessFuncIndex(Addr, DL);
559    if (Idx < 0)
560      return false;
561    const size_t ByteSize = 1 << Idx;
562    const size_t BitSize = ByteSize * 8;
563    Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
564    Type *PtrTy = Ty->getPointerTo();
565    Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
566                     IRB.CreateIntCast(CASI->getCompareOperand(), Ty, false),
567                     IRB.CreateIntCast(CASI->getNewValOperand(), Ty, false),
568                     createOrdering(&IRB, CASI->getSuccessOrdering()),
569                     createOrdering(&IRB, CASI->getFailureOrdering())};
570    CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
571    Value *Success = IRB.CreateICmpEQ(C, CASI->getCompareOperand());
572
573    Value *Res = IRB.CreateInsertValue(UndefValue::get(CASI->getType()), C, 0);
574    Res = IRB.CreateInsertValue(Res, Success, 1);
575
576    I->replaceAllUsesWith(Res);
577    I->eraseFromParent();
578  } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
579    Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
580    Function *F = FI->getSynchScope() == SingleThread ?
581        TsanAtomicSignalFence : TsanAtomicThreadFence;
582    CallInst *C = CallInst::Create(F, Args);
583    ReplaceInstWithInst(I, C);
584  }
585  return true;
586}
587
588int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr,
589                                              const DataLayout &DL) {
590  Type *OrigPtrTy = Addr->getType();
591  Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
592  assert(OrigTy->isSized());
593  uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
594  if (TypeSize != 8  && TypeSize != 16 &&
595      TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
596    NumAccessesWithBadSize++;
597    // Ignore all unusual sizes.
598    return -1;
599  }
600  size_t Idx = countTrailingZeros(TypeSize / 8);
601  assert(Idx < kNumberOfAccessSizes);
602  return Idx;
603}
604