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