lib/Transforms/Instrumentation/ThreadSanitizer.cpp

   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 #define DEBUG_TYPE "tsan"
  23
  24 #include "llvm/Transforms/Instrumentation.h"
  25 #include "llvm/ADT/SmallSet.h"
  26 #include "llvm/ADT/SmallString.h"
  27 #include "llvm/ADT/SmallVector.h"
  28 #include "llvm/ADT/Statistic.h"
  29 #include "llvm/ADT/StringExtras.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 #include "llvm/Transforms/Utils/SpecialCaseList.h"
  46
  47 using namespace llvm;
  48
  49 static cl::opt<std::string>  ClBlacklistFile("tsan-blacklist",
  50        cl::desc("Blacklist file"), cl::Hidden);
  51 static cl::opt<bool>  ClInstrumentMemoryAccesses(
  52     "tsan-instrument-memory-accesses", cl::init(true),
  53     cl::desc("Instrument memory accesses"), cl::Hidden);
  54 static cl::opt<bool>  ClInstrumentFuncEntryExit(
  55     "tsan-instrument-func-entry-exit", cl::init(true),
  56     cl::desc("Instrument function entry and exit"), cl::Hidden);
  57 static cl::opt<bool>  ClInstrumentAtomics(
  58     "tsan-instrument-atomics", cl::init(true),
  59     cl::desc("Instrument atomics"), cl::Hidden);
  60 static cl::opt<bool>  ClInstrumentMemIntrinsics(
  61     "tsan-instrument-memintrinsics", cl::init(true),
  62     cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
  63
  64 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
  65 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
  66 STATISTIC(NumOmittedReadsBeforeWrite,
  67           "Number of reads ignored due to following writes");
  68 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
  69 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
  70 STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
  71 STATISTIC(NumOmittedReadsFromConstantGlobals,
  72           "Number of reads from constant globals");
  73 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
  74
  75 namespace {
  76
  77 /// ThreadSanitizer: instrument the code in module to find races.
  78 struct ThreadSanitizer : public FunctionPass {
  79   ThreadSanitizer(StringRef BlacklistFile = StringRef())
  80       : FunctionPass(ID),
  81         DL(0),
  82         BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
  83                                             : BlacklistFile) { }
  84   const char *getPassName() const override;
  85   bool runOnFunction(Function &F) override;
  86   bool doInitialization(Module &M) override;
  87   static char ID;  // Pass identification, replacement for typeid.
  88
  89  private:
  90   void initializeCallbacks(Module &M);
  91   bool instrumentLoadOrStore(Instruction *I);
  92   bool instrumentAtomic(Instruction *I);
  93   bool instrumentMemIntrinsic(Instruction *I);
  94   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction*> &Local,
  95                                       SmallVectorImpl<Instruction*> &All);
  96   bool addrPointsToConstantData(Value *Addr);
  97   int getMemoryAccessFuncIndex(Value *Addr);
  98
  99   const DataLayout *DL;
 100   Type *IntptrTy;
 101   SmallString<64> BlacklistFile;
 102   OwningPtr<SpecialCaseList> BL;
 103   IntegerType *OrdTy;
 104   // Callbacks to run-time library are computed in doInitialization.
 105   Function *TsanFuncEntry;
 106   Function *TsanFuncExit;
 107   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
 108   static const size_t kNumberOfAccessSizes = 5;
 109   Function *TsanRead[kNumberOfAccessSizes];
 110   Function *TsanWrite[kNumberOfAccessSizes];
 111   Function *TsanAtomicLoad[kNumberOfAccessSizes];
 112   Function *TsanAtomicStore[kNumberOfAccessSizes];
 113   Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes];
 114   Function *TsanAtomicCAS[kNumberOfAccessSizes];
 115   Function *TsanAtomicThreadFence;
 116   Function *TsanAtomicSignalFence;
 117   Function *TsanVptrUpdate;
 118   Function *TsanVptrLoad;
 119   Function *MemmoveFn, *MemcpyFn, *MemsetFn;
 120 };
 121 }  // namespace
 122
 123 char ThreadSanitizer::ID = 0;
 124 INITIALIZE_PASS(ThreadSanitizer, "tsan",
 125     "ThreadSanitizer: detects data races.",
 126     false, false)
 127
 128 const char *ThreadSanitizer::getPassName() const {
 129   return "ThreadSanitizer";
 130 }
 131
 132 FunctionPass *llvm::createThreadSanitizerPass(StringRef BlacklistFile) {
 133   return new ThreadSanitizer(BlacklistFile);
 134 }
 135
 136 static Function *checkInterfaceFunction(Constant *FuncOrBitcast) {
 137   if (Function *F = dyn_cast<Function>(FuncOrBitcast))
 138      return F;
 139   FuncOrBitcast->dump();
 140   report_fatal_error("ThreadSanitizer interface function redefined");
 141 }
 142
 143 void ThreadSanitizer::initializeCallbacks(Module &M) {
 144   IRBuilder<> IRB(M.getContext());
 145   // Initialize the callbacks.
 146   TsanFuncEntry = checkInterfaceFunction(M.getOrInsertFunction(
 147       "__tsan_func_entry", IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
 148   TsanFuncExit = checkInterfaceFunction(M.getOrInsertFunction(
 149       "__tsan_func_exit", IRB.getVoidTy(), NULL));
 150   OrdTy = IRB.getInt32Ty();
 151   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
 152     const size_t ByteSize = 1 << i;
 153     const size_t BitSize = ByteSize * 8;
 154     SmallString<32> ReadName("__tsan_read" + itostr(ByteSize));
 155     TsanRead[i] = checkInterfaceFunction(M.getOrInsertFunction(
 156         ReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
 157
 158     SmallString<32> WriteName("__tsan_write" + itostr(ByteSize));
 159     TsanWrite[i] = checkInterfaceFunction(M.getOrInsertFunction(
 160         WriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
 161
 162     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
 163     Type *PtrTy = Ty->getPointerTo();
 164     SmallString<32> AtomicLoadName("__tsan_atomic" + itostr(BitSize) +
 165                                    "_load");
 166     TsanAtomicLoad[i] = checkInterfaceFunction(M.getOrInsertFunction(
 167         AtomicLoadName, Ty, PtrTy, OrdTy, NULL));
 168
 169     SmallString<32> AtomicStoreName("__tsan_atomic" + itostr(BitSize) +
 170                                     "_store");
 171     TsanAtomicStore[i] = checkInterfaceFunction(M.getOrInsertFunction(
 172         AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy,
 173         NULL));
 174
 175     for (int op = AtomicRMWInst::FIRST_BINOP;
 176         op <= AtomicRMWInst::LAST_BINOP; ++op) {
 177       TsanAtomicRMW[op][i] = NULL;
 178       const char *NamePart = NULL;
 179       if (op == AtomicRMWInst::Xchg)
 180         NamePart = "_exchange";
 181       else if (op == AtomicRMWInst::Add)
 182         NamePart = "_fetch_add";
 183       else if (op == AtomicRMWInst::Sub)
 184         NamePart = "_fetch_sub";
 185       else if (op == AtomicRMWInst::And)
 186         NamePart = "_fetch_and";
 187       else if (op == AtomicRMWInst::Or)
 188         NamePart = "_fetch_or";
 189       else if (op == AtomicRMWInst::Xor)
 190         NamePart = "_fetch_xor";
 191       else if (op == AtomicRMWInst::Nand)
 192         NamePart = "_fetch_nand";
 193       else
 194         continue;
 195       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
 196       TsanAtomicRMW[op][i] = checkInterfaceFunction(M.getOrInsertFunction(
 197           RMWName, Ty, PtrTy, Ty, OrdTy, NULL));
 198     }
 199
 200     SmallString<32> AtomicCASName("__tsan_atomic" + itostr(BitSize) +
 201                                   "_compare_exchange_val");
 202     TsanAtomicCAS[i] = checkInterfaceFunction(M.getOrInsertFunction(
 203         AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, NULL));
 204   }
 205   TsanVptrUpdate = checkInterfaceFunction(M.getOrInsertFunction(
 206       "__tsan_vptr_update", IRB.getVoidTy(), IRB.getInt8PtrTy(),
 207       IRB.getInt8PtrTy(), NULL));
 208   TsanVptrLoad = checkInterfaceFunction(M.getOrInsertFunction(
 209       "__tsan_vptr_read", IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
 210   TsanAtomicThreadFence = checkInterfaceFunction(M.getOrInsertFunction(
 211       "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, NULL));
 212   TsanAtomicSignalFence = checkInterfaceFunction(M.getOrInsertFunction(
 213       "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, NULL));
 214
 215   MemmoveFn = checkInterfaceFunction(M.getOrInsertFunction(
 216     "memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
 217     IRB.getInt8PtrTy(), IntptrTy, NULL));
 218   MemcpyFn = checkInterfaceFunction(M.getOrInsertFunction(
 219     "memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
 220     IntptrTy, NULL));
 221   MemsetFn = checkInterfaceFunction(M.getOrInsertFunction(
 222     "memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
 223     IntptrTy, NULL));
 224 }
 225
 226 bool ThreadSanitizer::doInitialization(Module &M) {
 227   DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
 228   if (!DLP)
 229     return false;
 230   DL = &DLP->getDataLayout();
 231   BL.reset(SpecialCaseList::createOrDie(BlacklistFile));
 232
 233   // Always insert a call to __tsan_init into the module's CTORs.
 234   IRBuilder<> IRB(M.getContext());
 235   IntptrTy = IRB.getIntPtrTy(DL);
 236   Value *TsanInit = M.getOrInsertFunction("__tsan_init",
 237                                           IRB.getVoidTy(), NULL);
 238   appendToGlobalCtors(M, cast<Function>(TsanInit), 0);
 239
 240   return true;
 241 }
 242
 243 static bool isVtableAccess(Instruction *I) {
 244   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
 245     return Tag->isTBAAVtableAccess();
 246   return false;
 247 }
 248
 249 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
 250   // If this is a GEP, just analyze its pointer operand.
 251   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
 252     Addr = GEP->getPointerOperand();
 253
 254   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
 255     if (GV->isConstant()) {
 256       // Reads from constant globals can not race with any writes.
 257       NumOmittedReadsFromConstantGlobals++;
 258       return true;
 259     }
 260   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
 261     if (isVtableAccess(L)) {
 262       // Reads from a vtable pointer can not race with any writes.
 263       NumOmittedReadsFromVtable++;
 264       return true;
 265     }
 266   }
 267   return false;
 268 }
 269
 270 // Instrumenting some of the accesses may be proven redundant.
 271 // Currently handled:
 272 //  - read-before-write (within same BB, no calls between)
 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.
 281 void ThreadSanitizer::chooseInstructionsToInstrument(
 282     SmallVectorImpl<Instruction*> &Local,
 283     SmallVectorImpl<Instruction*> &All) {
 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     All.push_back(I);
 305   }
 306   Local.clear();
 307 }
 308
 309 static bool isAtomic(Instruction *I) {
 310   if (LoadInst *LI = dyn_cast<LoadInst>(I))
 311     return LI->isAtomic() && LI->getSynchScope() == CrossThread;
 312   if (StoreInst *SI = dyn_cast<StoreInst>(I))
 313     return SI->isAtomic() && SI->getSynchScope() == CrossThread;
 314   if (isa<AtomicRMWInst>(I))
 315     return true;
 316   if (isa<AtomicCmpXchgInst>(I))
 317     return true;
 318   if (isa<FenceInst>(I))
 319     return true;
 320   return false;
 321 }
 322
 323 bool ThreadSanitizer::runOnFunction(Function &F) {
 324   if (!DL) return false;
 325   if (BL->isIn(F)) return false;
 326   initializeCallbacks(*F.getParent());
 327   SmallVector<Instruction*, 8> RetVec;
 328   SmallVector<Instruction*, 8> AllLoadsAndStores;
 329   SmallVector<Instruction*, 8> LocalLoadsAndStores;
 330   SmallVector<Instruction*, 8> AtomicAccesses;
 331   SmallVector<Instruction*, 8> MemIntrinCalls;
 332   bool Res = false;
 333   bool HasCalls = false;
 334
 335   // Traverse all instructions, collect loads/stores/returns, check for calls.
 336   for (Function::iterator FI = F.begin(), FE = F.end();
 337        FI != FE; ++FI) {
 338     BasicBlock &BB = *FI;
 339     for (BasicBlock::iterator BI = BB.begin(), BE = BB.end();
 340          BI != BE; ++BI) {
 341       if (isAtomic(BI))
 342         AtomicAccesses.push_back(BI);
 343       else if (isa<LoadInst>(BI) || isa<StoreInst>(BI))
 344         LocalLoadsAndStores.push_back(BI);
 345       else if (isa<ReturnInst>(BI))
 346         RetVec.push_back(BI);
 347       else if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
 348         if (isa<MemIntrinsic>(BI))
 349           MemIntrinCalls.push_back(BI);
 350         HasCalls = true;
 351         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
 352       }
 353     }
 354     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
 355   }
 356
 357   // We have collected all loads and stores.
 358   // FIXME: many of these accesses do not need to be checked for races
 359   // (e.g. variables that do not escape, etc).
 360
 361   // Instrument memory accesses.
 362   if (ClInstrumentMemoryAccesses && F.hasFnAttribute(Attribute::SanitizeThread))
 363     for (size_t i = 0, n = AllLoadsAndStores.size(); i < n; ++i) {
 364       Res |= instrumentLoadOrStore(AllLoadsAndStores[i]);
 365     }
 366
 367   // Instrument atomic memory accesses.
 368   if (ClInstrumentAtomics)
 369     for (size_t i = 0, n = AtomicAccesses.size(); i < n; ++i) {
 370       Res |= instrumentAtomic(AtomicAccesses[i]);
 371     }
 372
 373   if (ClInstrumentMemIntrinsics)
 374     for (size_t i = 0, n = MemIntrinCalls.size(); i < n; ++i) {
 375       Res |= instrumentMemIntrinsic(MemIntrinCalls[i]);
 376     }
 377
 378   // Instrument function entry/exit points if there were instrumented accesses.
 379   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
 380     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
 381     Value *ReturnAddress = IRB.CreateCall(
 382         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
 383         IRB.getInt32(0));
 384     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
 385     for (size_t i = 0, n = RetVec.size(); i < n; ++i) {
 386       IRBuilder<> IRBRet(RetVec[i]);
 387       IRBRet.CreateCall(TsanFuncExit);
 388     }
 389     Res = true;
 390   }
 391   return Res;
 392 }
 393
 394 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I) {
 395   IRBuilder<> IRB(I);
 396   bool IsWrite = isa<StoreInst>(*I);
 397   Value *Addr = IsWrite
 398       ? cast<StoreInst>(I)->getPointerOperand()
 399       : cast<LoadInst>(I)->getPointerOperand();
 400   int Idx = getMemoryAccessFuncIndex(Addr);
 401   if (Idx < 0)
 402     return false;
 403   if (IsWrite && isVtableAccess(I)) {
 404     DEBUG(dbgs() << "  VPTR : " << *I << "\n");
 405     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
 406     // StoredValue may be a vector type if we are storing several vptrs at once.
 407     // In this case, just take the first element of the vector since this is
 408     // enough to find vptr races.
 409     if (isa<VectorType>(StoredValue->getType()))
 410       StoredValue = IRB.CreateExtractElement(
 411           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
 412     if (StoredValue->getType()->isIntegerTy())
 413       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
 414     // Call TsanVptrUpdate.
 415     IRB.CreateCall2(TsanVptrUpdate,
 416                     IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
 417                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy()));
 418     NumInstrumentedVtableWrites++;
 419     return true;
 420   }
 421   if (!IsWrite && isVtableAccess(I)) {
 422     IRB.CreateCall(TsanVptrLoad,
 423                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
 424     NumInstrumentedVtableReads++;
 425     return true;
 426   }
 427   Value *OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
 428   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
 429   if (IsWrite) NumInstrumentedWrites++;
 430   else         NumInstrumentedReads++;
 431   return true;
 432 }
 433
 434 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
 435   uint32_t v = 0;
 436   switch (ord) {
 437     case NotAtomic:              assert(false);
 438     case Unordered:              // Fall-through.
 439     case Monotonic:              v = 0; break;
 440     // case Consume:                v = 1; break;  // Not specified yet.
 441     case Acquire:                v = 2; break;
 442     case Release:                v = 3; break;
 443     case AcquireRelease:         v = 4; break;
 444     case SequentiallyConsistent: v = 5; break;
 445   }
 446   return IRB->getInt32(v);
 447 }
 448
 449 static ConstantInt *createFailOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
 450   uint32_t v = 0;
 451   switch (ord) {
 452     case NotAtomic:              assert(false);
 453     case Unordered:              // Fall-through.
 454     case Monotonic:              v = 0; break;
 455     // case Consume:                v = 1; break;  // Not specified yet.
 456     case Acquire:                v = 2; break;
 457     case Release:                v = 0; break;
 458     case AcquireRelease:         v = 2; break;
 459     case SequentiallyConsistent: v = 5; break;
 460   }
 461   return IRB->getInt32(v);
 462 }
 463
 464 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
 465 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
 466 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
 467 // instead we simply replace them with regular function calls, which are then
 468 // intercepted by the run-time.
 469 // Since tsan is running after everyone else, the calls should not be
 470 // replaced back with intrinsics. If that becomes wrong at some point,
 471 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
 472 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
 473   IRBuilder<> IRB(I);
 474   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
 475     IRB.CreateCall3(MemsetFn,
 476       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
 477       IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
 478       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
 479     I->eraseFromParent();
 480   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
 481     IRB.CreateCall3(isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
 482       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
 483       IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
 484       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
 485     I->eraseFromParent();
 486   }
 487   return false;
 488 }
 489
 490 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
 491 // standards.  For background see C++11 standard.  A slightly older, publicly
 492 // available draft of the standard (not entirely up-to-date, but close enough
 493 // for casual browsing) is available here:
 494 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
 495 // The following page contains more background information:
 496 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
 497
 498 bool ThreadSanitizer::instrumentAtomic(Instruction *I) {
 499   IRBuilder<> IRB(I);
 500   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
 501     Value *Addr = LI->getPointerOperand();
 502     int Idx = getMemoryAccessFuncIndex(Addr);
 503     if (Idx < 0)
 504       return false;
 505     const size_t ByteSize = 1 << Idx;
 506     const size_t BitSize = ByteSize * 8;
 507     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 508     Type *PtrTy = Ty->getPointerTo();
 509     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 510                      createOrdering(&IRB, LI->getOrdering())};
 511     CallInst *C = CallInst::Create(TsanAtomicLoad[Idx],
 512                                    ArrayRef<Value*>(Args));
 513     ReplaceInstWithInst(I, C);
 514
 515   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
 516     Value *Addr = SI->getPointerOperand();
 517     int Idx = getMemoryAccessFuncIndex(Addr);
 518     if (Idx < 0)
 519       return false;
 520     const size_t ByteSize = 1 << Idx;
 521     const size_t BitSize = ByteSize * 8;
 522     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 523     Type *PtrTy = Ty->getPointerTo();
 524     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 525                      IRB.CreateIntCast(SI->getValueOperand(), Ty, false),
 526                      createOrdering(&IRB, SI->getOrdering())};
 527     CallInst *C = CallInst::Create(TsanAtomicStore[Idx],
 528                                    ArrayRef<Value*>(Args));
 529     ReplaceInstWithInst(I, C);
 530   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
 531     Value *Addr = RMWI->getPointerOperand();
 532     int Idx = getMemoryAccessFuncIndex(Addr);
 533     if (Idx < 0)
 534       return false;
 535     Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
 536     if (F == NULL)
 537       return false;
 538     const size_t ByteSize = 1 << Idx;
 539     const size_t BitSize = ByteSize * 8;
 540     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 541     Type *PtrTy = Ty->getPointerTo();
 542     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 543                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
 544                      createOrdering(&IRB, RMWI->getOrdering())};
 545     CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args));
 546     ReplaceInstWithInst(I, C);
 547   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
 548     Value *Addr = CASI->getPointerOperand();
 549     int Idx = getMemoryAccessFuncIndex(Addr);
 550     if (Idx < 0)
 551       return false;
 552     const size_t ByteSize = 1 << Idx;
 553     const size_t BitSize = ByteSize * 8;
 554     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 555     Type *PtrTy = Ty->getPointerTo();
 556     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 557                      IRB.CreateIntCast(CASI->getCompareOperand(), Ty, false),
 558                      IRB.CreateIntCast(CASI->getNewValOperand(), Ty, false),
 559                      createOrdering(&IRB, CASI->getOrdering()),
 560                      createFailOrdering(&IRB, CASI->getOrdering())};
 561     CallInst *C = CallInst::Create(TsanAtomicCAS[Idx], ArrayRef<Value*>(Args));
 562     ReplaceInstWithInst(I, C);
 563   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
 564     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
 565     Function *F = FI->getSynchScope() == SingleThread ?
 566         TsanAtomicSignalFence : TsanAtomicThreadFence;
 567     CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args));
 568     ReplaceInstWithInst(I, C);
 569   }
 570   return true;
 571 }
 572
 573 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr) {
 574   Type *OrigPtrTy = Addr->getType();
 575   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
 576   assert(OrigTy->isSized());
 577   uint32_t TypeSize = DL->getTypeStoreSizeInBits(OrigTy);
 578   if (TypeSize != 8  && TypeSize != 16 &&
 579       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
 580     NumAccessesWithBadSize++;
 581     // Ignore all unusual sizes.
 582     return -1;
 583   }
 584   size_t Idx = countTrailingZeros(TypeSize / 8);
 585   assert(Idx < kNumberOfAccessSizes);
 586   return Idx;
 587 }