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;
  85   bool runOnFunction(Function &F);
  86   bool doInitialization(Module &M);
  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   DL = getAnalysisIfAvailable<DataLayout>();
 228   if (!DL)
 229     return false;
 230   BL.reset(SpecialCaseList::createOrDie(BlacklistFile));
 231
 232   // Always insert a call to __tsan_init into the module's CTORs.
 233   IRBuilder<> IRB(M.getContext());
 234   IntptrTy = IRB.getIntPtrTy(DL);
 235   Value *TsanInit = M.getOrInsertFunction("__tsan_init",
 236                                           IRB.getVoidTy(), NULL);
 237   appendToGlobalCtors(M, cast<Function>(TsanInit), 0);
 238
 239   return true;
 240 }
 241
 242 static bool isVtableAccess(Instruction *I) {
 243   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
 244     return Tag->isTBAAVtableAccess();
 245   return false;
 246 }
 247
 248 bool 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 //
 273 // We do not handle some of the patterns that should not survive
 274 // after the classic compiler optimizations.
 275 // E.g. two reads from the same temp should be eliminated by CSE,
 276 // two writes should be eliminated by DSE, etc.
 277 //
 278 // 'Local' is a vector of insns within the same BB (no calls between).
 279 // 'All' is a vector of insns that will be instrumented.
 280 void ThreadSanitizer::chooseInstructionsToInstrument(
 281     SmallVectorImpl<Instruction*> &Local,
 282     SmallVectorImpl<Instruction*> &All) {
 283   SmallSet<Value*, 8> WriteTargets;
 284   // Iterate from the end.
 285   for (SmallVectorImpl<Instruction*>::reverse_iterator It = Local.rbegin(),
 286        E = Local.rend(); It != E; ++It) {
 287     Instruction *I = *It;
 288     if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
 289       WriteTargets.insert(Store->getPointerOperand());
 290     } else {
 291       LoadInst *Load = cast<LoadInst>(I);
 292       Value *Addr = Load->getPointerOperand();
 293       if (WriteTargets.count(Addr)) {
 294         // We will write to this temp, so no reason to analyze the read.
 295         NumOmittedReadsBeforeWrite++;
 296         continue;
 297       }
 298       if (addrPointsToConstantData(Addr)) {
 299         // Addr points to some constant data -- it can not race with any writes.
 300         continue;
 301       }
 302     }
 303     All.push_back(I);
 304   }
 305   Local.clear();
 306 }
 307
 308 static bool isAtomic(Instruction *I) {
 309   if (LoadInst *LI = dyn_cast<LoadInst>(I))
 310     return LI->isAtomic() && LI->getSynchScope() == CrossThread;
 311   if (StoreInst *SI = dyn_cast<StoreInst>(I))
 312     return SI->isAtomic() && SI->getSynchScope() == CrossThread;
 313   if (isa<AtomicRMWInst>(I))
 314     return true;
 315   if (isa<AtomicCmpXchgInst>(I))
 316     return true;
 317   if (isa<FenceInst>(I))
 318     return true;
 319   return false;
 320 }
 321
 322 bool ThreadSanitizer::runOnFunction(Function &F) {
 323   if (!DL) return false;
 324   if (BL->isIn(F)) return false;
 325   initializeCallbacks(*F.getParent());
 326   SmallVector<Instruction*, 8> RetVec;
 327   SmallVector<Instruction*, 8> AllLoadsAndStores;
 328   SmallVector<Instruction*, 8> LocalLoadsAndStores;
 329   SmallVector<Instruction*, 8> AtomicAccesses;
 330   SmallVector<Instruction*, 8> MemIntrinCalls;
 331   bool Res = false;
 332   bool HasCalls = false;
 333
 334   // Traverse all instructions, collect loads/stores/returns, check for calls.
 335   for (Function::iterator FI = F.begin(), FE = F.end();
 336        FI != FE; ++FI) {
 337     BasicBlock &BB = *FI;
 338     for (BasicBlock::iterator BI = BB.begin(), BE = BB.end();
 339          BI != BE; ++BI) {
 340       if (isAtomic(BI))
 341         AtomicAccesses.push_back(BI);
 342       else if (isa<LoadInst>(BI) || isa<StoreInst>(BI))
 343         LocalLoadsAndStores.push_back(BI);
 344       else if (isa<ReturnInst>(BI))
 345         RetVec.push_back(BI);
 346       else if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
 347         if (isa<MemIntrinsic>(BI))
 348           MemIntrinCalls.push_back(BI);
 349         HasCalls = true;
 350         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
 351       }
 352     }
 353     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
 354   }
 355
 356   // We have collected all loads and stores.
 357   // FIXME: many of these accesses do not need to be checked for races
 358   // (e.g. variables that do not escape, etc).
 359
 360   // Instrument memory accesses.
 361   if (ClInstrumentMemoryAccesses && F.hasFnAttribute(Attribute::SanitizeThread))
 362     for (size_t i = 0, n = AllLoadsAndStores.size(); i < n; ++i) {
 363       Res |= instrumentLoadOrStore(AllLoadsAndStores[i]);
 364     }
 365
 366   // Instrument atomic memory accesses.
 367   if (ClInstrumentAtomics)
 368     for (size_t i = 0, n = AtomicAccesses.size(); i < n; ++i) {
 369       Res |= instrumentAtomic(AtomicAccesses[i]);
 370     }
 371
 372   if (ClInstrumentMemIntrinsics)
 373     for (size_t i = 0, n = MemIntrinCalls.size(); i < n; ++i) {
 374       Res |= instrumentMemIntrinsic(MemIntrinCalls[i]);
 375     }
 376
 377   // Instrument function entry/exit points if there were instrumented accesses.
 378   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
 379     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
 380     Value *ReturnAddress = IRB.CreateCall(
 381         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
 382         IRB.getInt32(0));
 383     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
 384     for (size_t i = 0, n = RetVec.size(); i < n; ++i) {
 385       IRBuilder<> IRBRet(RetVec[i]);
 386       IRBRet.CreateCall(TsanFuncExit);
 387     }
 388     Res = true;
 389   }
 390   return Res;
 391 }
 392
 393 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I) {
 394   IRBuilder<> IRB(I);
 395   bool IsWrite = isa<StoreInst>(*I);
 396   Value *Addr = IsWrite
 397       ? cast<StoreInst>(I)->getPointerOperand()
 398       : cast<LoadInst>(I)->getPointerOperand();
 399   int Idx = getMemoryAccessFuncIndex(Addr);
 400   if (Idx < 0)
 401     return false;
 402   if (IsWrite && isVtableAccess(I)) {
 403     DEBUG(dbgs() << "  VPTR : " << *I << "\n");
 404     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
 405     // StoredValue may be a vector type if we are storing several vptrs at once.
 406     // In this case, just take the first element of the vector since this is
 407     // enough to find vptr races.
 408     if (isa<VectorType>(StoredValue->getType()))
 409       StoredValue = IRB.CreateExtractElement(
 410           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
 411     if (StoredValue->getType()->isIntegerTy())
 412       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
 413     // Call TsanVptrUpdate.
 414     IRB.CreateCall2(TsanVptrUpdate,
 415                     IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
 416                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy()));
 417     NumInstrumentedVtableWrites++;
 418     return true;
 419   }
 420   if (!IsWrite && isVtableAccess(I)) {
 421     IRB.CreateCall(TsanVptrLoad,
 422                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
 423     NumInstrumentedVtableReads++;
 424     return true;
 425   }
 426   Value *OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
 427   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
 428   if (IsWrite) NumInstrumentedWrites++;
 429   else         NumInstrumentedReads++;
 430   return true;
 431 }
 432
 433 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
 434   uint32_t v = 0;
 435   switch (ord) {
 436     case NotAtomic:              assert(false);
 437     case Unordered:              // Fall-through.
 438     case Monotonic:              v = 0; break;
 439     // case Consume:                v = 1; break;  // Not specified yet.
 440     case Acquire:                v = 2; break;
 441     case Release:                v = 3; break;
 442     case AcquireRelease:         v = 4; break;
 443     case SequentiallyConsistent: v = 5; break;
 444   }
 445   return IRB->getInt32(v);
 446 }
 447
 448 static ConstantInt *createFailOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
 449   uint32_t v = 0;
 450   switch (ord) {
 451     case NotAtomic:              assert(false);
 452     case Unordered:              // Fall-through.
 453     case Monotonic:              v = 0; break;
 454     // case Consume:                v = 1; break;  // Not specified yet.
 455     case Acquire:                v = 2; break;
 456     case Release:                v = 0; break;
 457     case AcquireRelease:         v = 2; break;
 458     case SequentiallyConsistent: v = 5; break;
 459   }
 460   return IRB->getInt32(v);
 461 }
 462
 463 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
 464 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
 465 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
 466 // instead we simply replace them with regular function calls, which are then
 467 // intercepted by the run-time.
 468 // Since tsan is running after everyone else, the calls should not be
 469 // replaced back with intrinsics. If that becomes wrong at some point,
 470 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
 471 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
 472   IRBuilder<> IRB(I);
 473   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
 474     IRB.CreateCall3(MemsetFn,
 475       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
 476       IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
 477       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
 478     I->eraseFromParent();
 479   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
 480     IRB.CreateCall3(isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
 481       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
 482       IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
 483       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
 484     I->eraseFromParent();
 485   }
 486   return false;
 487 }
 488
 489 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
 490 // standards.  For background see C++11 standard.  A slightly older, publicly
 491 // available draft of the standard (not entirely up-to-date, but close enough
 492 // for casual browsing) is available here:
 493 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
 494 // The following page contains more background information:
 495 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
 496
 497 bool ThreadSanitizer::instrumentAtomic(Instruction *I) {
 498   IRBuilder<> IRB(I);
 499   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
 500     Value *Addr = LI->getPointerOperand();
 501     int Idx = getMemoryAccessFuncIndex(Addr);
 502     if (Idx < 0)
 503       return false;
 504     const size_t ByteSize = 1 << Idx;
 505     const size_t BitSize = ByteSize * 8;
 506     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 507     Type *PtrTy = Ty->getPointerTo();
 508     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 509                      createOrdering(&IRB, LI->getOrdering())};
 510     CallInst *C = CallInst::Create(TsanAtomicLoad[Idx],
 511                                    ArrayRef<Value*>(Args));
 512     ReplaceInstWithInst(I, C);
 513
 514   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
 515     Value *Addr = SI->getPointerOperand();
 516     int Idx = getMemoryAccessFuncIndex(Addr);
 517     if (Idx < 0)
 518       return false;
 519     const size_t ByteSize = 1 << Idx;
 520     const size_t BitSize = ByteSize * 8;
 521     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 522     Type *PtrTy = Ty->getPointerTo();
 523     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 524                      IRB.CreateIntCast(SI->getValueOperand(), Ty, false),
 525                      createOrdering(&IRB, SI->getOrdering())};
 526     CallInst *C = CallInst::Create(TsanAtomicStore[Idx],
 527                                    ArrayRef<Value*>(Args));
 528     ReplaceInstWithInst(I, C);
 529   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
 530     Value *Addr = RMWI->getPointerOperand();
 531     int Idx = getMemoryAccessFuncIndex(Addr);
 532     if (Idx < 0)
 533       return false;
 534     Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
 535     if (F == NULL)
 536       return false;
 537     const size_t ByteSize = 1 << Idx;
 538     const size_t BitSize = ByteSize * 8;
 539     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 540     Type *PtrTy = Ty->getPointerTo();
 541     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 542                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
 543                      createOrdering(&IRB, RMWI->getOrdering())};
 544     CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args));
 545     ReplaceInstWithInst(I, C);
 546   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
 547     Value *Addr = CASI->getPointerOperand();
 548     int Idx = getMemoryAccessFuncIndex(Addr);
 549     if (Idx < 0)
 550       return false;
 551     const size_t ByteSize = 1 << Idx;
 552     const size_t BitSize = ByteSize * 8;
 553     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 554     Type *PtrTy = Ty->getPointerTo();
 555     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 556                      IRB.CreateIntCast(CASI->getCompareOperand(), Ty, false),
 557                      IRB.CreateIntCast(CASI->getNewValOperand(), Ty, false),
 558                      createOrdering(&IRB, CASI->getOrdering()),
 559                      createFailOrdering(&IRB, CASI->getOrdering())};
 560     CallInst *C = CallInst::Create(TsanAtomicCAS[Idx], ArrayRef<Value*>(Args));
 561     ReplaceInstWithInst(I, C);
 562   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
 563     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
 564     Function *F = FI->getSynchScope() == SingleThread ?
 565         TsanAtomicSignalFence : TsanAtomicThreadFence;
 566     CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args));
 567     ReplaceInstWithInst(I, C);
 568   }
 569   return true;
 570 }
 571
 572 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr) {
 573   Type *OrigPtrTy = Addr->getType();
 574   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
 575   assert(OrigTy->isSized());
 576   uint32_t TypeSize = DL->getTypeStoreSizeInBits(OrigTy);
 577   if (TypeSize != 8  && TypeSize != 16 &&
 578       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
 579     NumAccessesWithBadSize++;
 580     // Ignore all unusual sizes.
 581     return -1;
 582   }
 583   size_t Idx = countTrailingZeros(TypeSize / 8);
 584   assert(Idx < kNumberOfAccessSizes);
 585   return Idx;
 586 }