contrib/llvm/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/TargetLibraryInfo.h"
  30 #include "llvm/Analysis/ValueTracking.h"
  31 #include "llvm/IR/DataLayout.h"
  32 #include "llvm/IR/Function.h"
  33 #include "llvm/IR/IRBuilder.h"
  34 #include "llvm/IR/IntrinsicInst.h"
  35 #include "llvm/IR/Intrinsics.h"
  36 #include "llvm/IR/LLVMContext.h"
  37 #include "llvm/IR/Metadata.h"
  38 #include "llvm/IR/Module.h"
  39 #include "llvm/IR/Type.h"
  40 #include "llvm/ProfileData/InstrProf.h"
  41 #include "llvm/Support/CommandLine.h"
  42 #include "llvm/Support/Debug.h"
  43 #include "llvm/Support/MathExtras.h"
  44 #include "llvm/Support/raw_ostream.h"
  45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
  46 #include "llvm/Transforms/Utils/Local.h"
  47 #include "llvm/Transforms/Utils/ModuleUtils.h"
  48
  49 using namespace llvm;
  50
  51 #define DEBUG_TYPE "tsan"
  52
  53 static cl::opt<bool>  ClInstrumentMemoryAccesses(
  54     "tsan-instrument-memory-accesses", cl::init(true),
  55     cl::desc("Instrument memory accesses"), cl::Hidden);
  56 static cl::opt<bool>  ClInstrumentFuncEntryExit(
  57     "tsan-instrument-func-entry-exit", cl::init(true),
  58     cl::desc("Instrument function entry and exit"), cl::Hidden);
  59 static cl::opt<bool>  ClInstrumentAtomics(
  60     "tsan-instrument-atomics", cl::init(true),
  61     cl::desc("Instrument atomics"), cl::Hidden);
  62 static cl::opt<bool>  ClInstrumentMemIntrinsics(
  63     "tsan-instrument-memintrinsics", cl::init(true),
  64     cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
  65
  66 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
  67 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
  68 STATISTIC(NumOmittedReadsBeforeWrite,
  69           "Number of reads ignored due to following writes");
  70 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
  71 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
  72 STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
  73 STATISTIC(NumOmittedReadsFromConstantGlobals,
  74           "Number of reads from constant globals");
  75 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
  76 STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");
  77
  78 static const char *const kTsanModuleCtorName = "tsan.module_ctor";
  79 static const char *const kTsanInitName = "__tsan_init";
  80
  81 namespace {
  82
  83 /// ThreadSanitizer: instrument the code in module to find races.
  84 struct ThreadSanitizer : public FunctionPass {
  85   ThreadSanitizer() : FunctionPass(ID) {}
  86   const char *getPassName() const override;
  87   void getAnalysisUsage(AnalysisUsage &AU) const override;
  88   bool runOnFunction(Function &F) override;
  89   bool doInitialization(Module &M) override;
  90   static char ID;  // Pass identification, replacement for typeid.
  91
  92  private:
  93   void initializeCallbacks(Module &M);
  94   bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL);
  95   bool instrumentAtomic(Instruction *I, const DataLayout &DL);
  96   bool instrumentMemIntrinsic(Instruction *I);
  97   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
  98                                       SmallVectorImpl<Instruction *> &All,
  99                                       const DataLayout &DL);
 100   bool addrPointsToConstantData(Value *Addr);
 101   int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL);
 102
 103   Type *IntptrTy;
 104   IntegerType *OrdTy;
 105   // Callbacks to run-time library are computed in doInitialization.
 106   Function *TsanFuncEntry;
 107   Function *TsanFuncExit;
 108   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
 109   static const size_t kNumberOfAccessSizes = 5;
 110   Function *TsanRead[kNumberOfAccessSizes];
 111   Function *TsanWrite[kNumberOfAccessSizes];
 112   Function *TsanUnalignedRead[kNumberOfAccessSizes];
 113   Function *TsanUnalignedWrite[kNumberOfAccessSizes];
 114   Function *TsanAtomicLoad[kNumberOfAccessSizes];
 115   Function *TsanAtomicStore[kNumberOfAccessSizes];
 116   Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes];
 117   Function *TsanAtomicCAS[kNumberOfAccessSizes];
 118   Function *TsanAtomicThreadFence;
 119   Function *TsanAtomicSignalFence;
 120   Function *TsanVptrUpdate;
 121   Function *TsanVptrLoad;
 122   Function *MemmoveFn, *MemcpyFn, *MemsetFn;
 123   Function *TsanCtorFunction;
 124 };
 125 }  // namespace
 126
 127 char ThreadSanitizer::ID = 0;
 128 INITIALIZE_PASS_BEGIN(
 129     ThreadSanitizer, "tsan",
 130     "ThreadSanitizer: detects data races.",
 131     false, false)
 132 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
 133 INITIALIZE_PASS_END(
 134     ThreadSanitizer, "tsan",
 135     "ThreadSanitizer: detects data races.",
 136     false, false)
 137
 138 const char *ThreadSanitizer::getPassName() const {
 139   return "ThreadSanitizer";
 140 }
 141
 142 void ThreadSanitizer::getAnalysisUsage(AnalysisUsage &AU) const {
 143   AU.addRequired<TargetLibraryInfoWrapperPass>();
 144 }
 145
 146 FunctionPass *llvm::createThreadSanitizerPass() {
 147   return new ThreadSanitizer();
 148 }
 149
 150 void ThreadSanitizer::initializeCallbacks(Module &M) {
 151   IRBuilder<> IRB(M.getContext());
 152   // Initialize the callbacks.
 153   TsanFuncEntry = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 154       "__tsan_func_entry", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 155   TsanFuncExit = checkSanitizerInterfaceFunction(
 156       M.getOrInsertFunction("__tsan_func_exit", IRB.getVoidTy(), nullptr));
 157   OrdTy = IRB.getInt32Ty();
 158   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
 159     const unsigned ByteSize = 1U << i;
 160     const unsigned BitSize = ByteSize * 8;
 161     std::string ByteSizeStr = utostr(ByteSize);
 162     std::string BitSizeStr = utostr(BitSize);
 163     SmallString<32> ReadName("__tsan_read" + ByteSizeStr);
 164     TsanRead[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 165         ReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 166
 167     SmallString<32> WriteName("__tsan_write" + ByteSizeStr);
 168     TsanWrite[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 169         WriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 170
 171     SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr);
 172     TsanUnalignedRead[i] =
 173         checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 174             UnalignedReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 175
 176     SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr);
 177     TsanUnalignedWrite[i] =
 178         checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 179             UnalignedWriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 180
 181     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
 182     Type *PtrTy = Ty->getPointerTo();
 183     SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load");
 184     TsanAtomicLoad[i] = checkSanitizerInterfaceFunction(
 185         M.getOrInsertFunction(AtomicLoadName, Ty, PtrTy, OrdTy, nullptr));
 186
 187     SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store");
 188     TsanAtomicStore[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 189         AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy, nullptr));
 190
 191     for (int op = AtomicRMWInst::FIRST_BINOP;
 192         op <= AtomicRMWInst::LAST_BINOP; ++op) {
 193       TsanAtomicRMW[op][i] = nullptr;
 194       const char *NamePart = nullptr;
 195       if (op == AtomicRMWInst::Xchg)
 196         NamePart = "_exchange";
 197       else if (op == AtomicRMWInst::Add)
 198         NamePart = "_fetch_add";
 199       else if (op == AtomicRMWInst::Sub)
 200         NamePart = "_fetch_sub";
 201       else if (op == AtomicRMWInst::And)
 202         NamePart = "_fetch_and";
 203       else if (op == AtomicRMWInst::Or)
 204         NamePart = "_fetch_or";
 205       else if (op == AtomicRMWInst::Xor)
 206         NamePart = "_fetch_xor";
 207       else if (op == AtomicRMWInst::Nand)
 208         NamePart = "_fetch_nand";
 209       else
 210         continue;
 211       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
 212       TsanAtomicRMW[op][i] = checkSanitizerInterfaceFunction(
 213           M.getOrInsertFunction(RMWName, Ty, PtrTy, Ty, OrdTy, nullptr));
 214     }
 215
 216     SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr +
 217                                   "_compare_exchange_val");
 218     TsanAtomicCAS[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 219         AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, nullptr));
 220   }
 221   TsanVptrUpdate = checkSanitizerInterfaceFunction(
 222       M.getOrInsertFunction("__tsan_vptr_update", IRB.getVoidTy(),
 223                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), nullptr));
 224   TsanVptrLoad = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 225       "__tsan_vptr_read", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 226   TsanAtomicThreadFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 227       "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, nullptr));
 228   TsanAtomicSignalFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 229       "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, nullptr));
 230
 231   MemmoveFn = checkSanitizerInterfaceFunction(
 232       M.getOrInsertFunction("memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
 233                             IRB.getInt8PtrTy(), IntptrTy, nullptr));
 234   MemcpyFn = checkSanitizerInterfaceFunction(
 235       M.getOrInsertFunction("memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
 236                             IRB.getInt8PtrTy(), IntptrTy, nullptr));
 237   MemsetFn = checkSanitizerInterfaceFunction(
 238       M.getOrInsertFunction("memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
 239                             IRB.getInt32Ty(), IntptrTy, nullptr));
 240 }
 241
 242 bool ThreadSanitizer::doInitialization(Module &M) {
 243   const DataLayout &DL = M.getDataLayout();
 244   IntptrTy = DL.getIntPtrType(M.getContext());
 245   std::tie(TsanCtorFunction, std::ignore) = createSanitizerCtorAndInitFunctions(
 246       M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{},
 247       /*InitArgs=*/{});
 248
 249   appendToGlobalCtors(M, TsanCtorFunction, 0);
 250
 251   return true;
 252 }
 253
 254 static bool isVtableAccess(Instruction *I) {
 255   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
 256     return Tag->isTBAAVtableAccess();
 257   return false;
 258 }
 259
 260 // Do not instrument known races/"benign races" that come from compiler
 261 // instrumentatin. The user has no way of suppressing them.
 262 static bool shouldInstrumentReadWriteFromAddress(Value *Addr) {
 263   // Peel off GEPs and BitCasts.
 264   Addr = Addr->stripInBoundsOffsets();
 265
 266   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
 267     if (GV->hasSection()) {
 268       StringRef SectionName = GV->getSection();
 269       // Check if the global is in the PGO counters section.
 270       if (SectionName.endswith(getInstrProfCountersSectionName(
 271             /*AddSegment=*/false)))
 272         return false;
 273     }
 274
 275     // Check if the global is private gcov data.
 276     if (GV->getName().startswith("__llvm_gcov") ||
 277         GV->getName().startswith("__llvm_gcda"))
 278       return false;
 279   }
 280
 281   // Do not instrument acesses from different address spaces; we cannot deal
 282   // with them.
 283   if (Addr) {
 284     Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
 285     if (PtrTy->getPointerAddressSpace() != 0)
 286       return false;
 287   }
 288
 289   return true;
 290 }
 291
 292 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
 293   // If this is a GEP, just analyze its pointer operand.
 294   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
 295     Addr = GEP->getPointerOperand();
 296
 297   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
 298     if (GV->isConstant()) {
 299       // Reads from constant globals can not race with any writes.
 300       NumOmittedReadsFromConstantGlobals++;
 301       return true;
 302     }
 303   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
 304     if (isVtableAccess(L)) {
 305       // Reads from a vtable pointer can not race with any writes.
 306       NumOmittedReadsFromVtable++;
 307       return true;
 308     }
 309   }
 310   return false;
 311 }
 312
 313 // Instrumenting some of the accesses may be proven redundant.
 314 // Currently handled:
 315 //  - read-before-write (within same BB, no calls between)
 316 //  - not captured variables
 317 //
 318 // We do not handle some of the patterns that should not survive
 319 // after the classic compiler optimizations.
 320 // E.g. two reads from the same temp should be eliminated by CSE,
 321 // two writes should be eliminated by DSE, etc.
 322 //
 323 // 'Local' is a vector of insns within the same BB (no calls between).
 324 // 'All' is a vector of insns that will be instrumented.
 325 void ThreadSanitizer::chooseInstructionsToInstrument(
 326     SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
 327     const DataLayout &DL) {
 328   SmallSet<Value*, 8> WriteTargets;
 329   // Iterate from the end.
 330   for (Instruction *I : reverse(Local)) {
 331     if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
 332       Value *Addr = Store->getPointerOperand();
 333       if (!shouldInstrumentReadWriteFromAddress(Addr))
 334         continue;
 335       WriteTargets.insert(Addr);
 336     } else {
 337       LoadInst *Load = cast<LoadInst>(I);
 338       Value *Addr = Load->getPointerOperand();
 339       if (!shouldInstrumentReadWriteFromAddress(Addr))
 340         continue;
 341       if (WriteTargets.count(Addr)) {
 342         // We will write to this temp, so no reason to analyze the read.
 343         NumOmittedReadsBeforeWrite++;
 344         continue;
 345       }
 346       if (addrPointsToConstantData(Addr)) {
 347         // Addr points to some constant data -- it can not race with any writes.
 348         continue;
 349       }
 350     }
 351     Value *Addr = isa<StoreInst>(*I)
 352         ? cast<StoreInst>(I)->getPointerOperand()
 353         : cast<LoadInst>(I)->getPointerOperand();
 354     if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) &&
 355         !PointerMayBeCaptured(Addr, true, true)) {
 356       // The variable is addressable but not captured, so it cannot be
 357       // referenced from a different thread and participate in a data race
 358       // (see llvm/Analysis/CaptureTracking.h for details).
 359       NumOmittedNonCaptured++;
 360       continue;
 361     }
 362     All.push_back(I);
 363   }
 364   Local.clear();
 365 }
 366
 367 static bool isAtomic(Instruction *I) {
 368   if (LoadInst *LI = dyn_cast<LoadInst>(I))
 369     return LI->isAtomic() && LI->getSynchScope() == CrossThread;
 370   if (StoreInst *SI = dyn_cast<StoreInst>(I))
 371     return SI->isAtomic() && SI->getSynchScope() == CrossThread;
 372   if (isa<AtomicRMWInst>(I))
 373     return true;
 374   if (isa<AtomicCmpXchgInst>(I))
 375     return true;
 376   if (isa<FenceInst>(I))
 377     return true;
 378   return false;
 379 }
 380
 381 bool ThreadSanitizer::runOnFunction(Function &F) {
 382   // This is required to prevent instrumenting call to __tsan_init from within
 383   // the module constructor.
 384   if (&F == TsanCtorFunction)
 385     return false;
 386   initializeCallbacks(*F.getParent());
 387   SmallVector<Instruction*, 8> RetVec;
 388   SmallVector<Instruction*, 8> AllLoadsAndStores;
 389   SmallVector<Instruction*, 8> LocalLoadsAndStores;
 390   SmallVector<Instruction*, 8> AtomicAccesses;
 391   SmallVector<Instruction*, 8> MemIntrinCalls;
 392   bool Res = false;
 393   bool HasCalls = false;
 394   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
 395   const DataLayout &DL = F.getParent()->getDataLayout();
 396   const TargetLibraryInfo *TLI =
 397       &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
 398
 399   // Traverse all instructions, collect loads/stores/returns, check for calls.
 400   for (auto &BB : F) {
 401     for (auto &Inst : BB) {
 402       if (isAtomic(&Inst))
 403         AtomicAccesses.push_back(&Inst);
 404       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
 405         LocalLoadsAndStores.push_back(&Inst);
 406       else if (isa<ReturnInst>(Inst))
 407         RetVec.push_back(&Inst);
 408       else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
 409         if (CallInst *CI = dyn_cast<CallInst>(&Inst))
 410           maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
 411         if (isa<MemIntrinsic>(Inst))
 412           MemIntrinCalls.push_back(&Inst);
 413         HasCalls = true;
 414         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
 415                                        DL);
 416       }
 417     }
 418     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
 419   }
 420
 421   // We have collected all loads and stores.
 422   // FIXME: many of these accesses do not need to be checked for races
 423   // (e.g. variables that do not escape, etc).
 424
 425   // Instrument memory accesses only if we want to report bugs in the function.
 426   if (ClInstrumentMemoryAccesses && SanitizeFunction)
 427     for (auto Inst : AllLoadsAndStores) {
 428       Res |= instrumentLoadOrStore(Inst, DL);
 429     }
 430
 431   // Instrument atomic memory accesses in any case (they can be used to
 432   // implement synchronization).
 433   if (ClInstrumentAtomics)
 434     for (auto Inst : AtomicAccesses) {
 435       Res |= instrumentAtomic(Inst, DL);
 436     }
 437
 438   if (ClInstrumentMemIntrinsics && SanitizeFunction)
 439     for (auto Inst : MemIntrinCalls) {
 440       Res |= instrumentMemIntrinsic(Inst);
 441     }
 442
 443   // Instrument function entry/exit points if there were instrumented accesses.
 444   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
 445     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
 446     Value *ReturnAddress = IRB.CreateCall(
 447         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
 448         IRB.getInt32(0));
 449     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
 450     for (auto RetInst : RetVec) {
 451       IRBuilder<> IRBRet(RetInst);
 452       IRBRet.CreateCall(TsanFuncExit, {});
 453     }
 454     Res = true;
 455   }
 456   return Res;
 457 }
 458
 459 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I,
 460                                             const DataLayout &DL) {
 461   IRBuilder<> IRB(I);
 462   bool IsWrite = isa<StoreInst>(*I);
 463   Value *Addr = IsWrite
 464       ? cast<StoreInst>(I)->getPointerOperand()
 465       : cast<LoadInst>(I)->getPointerOperand();
 466   int Idx = getMemoryAccessFuncIndex(Addr, DL);
 467   if (Idx < 0)
 468     return false;
 469   if (IsWrite && isVtableAccess(I)) {
 470     DEBUG(dbgs() << "  VPTR : " << *I << "\n");
 471     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
 472     // StoredValue may be a vector type if we are storing several vptrs at once.
 473     // In this case, just take the first element of the vector since this is
 474     // enough to find vptr races.
 475     if (isa<VectorType>(StoredValue->getType()))
 476       StoredValue = IRB.CreateExtractElement(
 477           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
 478     if (StoredValue->getType()->isIntegerTy())
 479       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
 480     // Call TsanVptrUpdate.
 481     IRB.CreateCall(TsanVptrUpdate,
 482                    {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
 483                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
 484     NumInstrumentedVtableWrites++;
 485     return true;
 486   }
 487   if (!IsWrite && isVtableAccess(I)) {
 488     IRB.CreateCall(TsanVptrLoad,
 489                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
 490     NumInstrumentedVtableReads++;
 491     return true;
 492   }
 493   const unsigned Alignment = IsWrite
 494       ? cast<StoreInst>(I)->getAlignment()
 495       : cast<LoadInst>(I)->getAlignment();
 496   Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
 497   const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
 498   Value *OnAccessFunc = nullptr;
 499   if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0)
 500     OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
 501   else
 502     OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
 503   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
 504   if (IsWrite) NumInstrumentedWrites++;
 505   else         NumInstrumentedReads++;
 506   return true;
 507 }
 508
 509 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
 510   uint32_t v = 0;
 511   switch (ord) {
 512     case AtomicOrdering::NotAtomic:
 513       llvm_unreachable("unexpected atomic ordering!");
 514     case AtomicOrdering::Unordered:              // Fall-through.
 515     case AtomicOrdering::Monotonic:              v = 0; break;
 516     // Not specified yet:
 517     // case AtomicOrdering::Consume:                v = 1; break;
 518     case AtomicOrdering::Acquire:                v = 2; break;
 519     case AtomicOrdering::Release:                v = 3; break;
 520     case AtomicOrdering::AcquireRelease:         v = 4; break;
 521     case AtomicOrdering::SequentiallyConsistent: v = 5; break;
 522   }
 523   return IRB->getInt32(v);
 524 }
 525
 526 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
 527 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
 528 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
 529 // instead we simply replace them with regular function calls, which are then
 530 // intercepted by the run-time.
 531 // Since tsan is running after everyone else, the calls should not be
 532 // replaced back with intrinsics. If that becomes wrong at some point,
 533 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
 534 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
 535   IRBuilder<> IRB(I);
 536   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
 537     IRB.CreateCall(
 538         MemsetFn,
 539         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
 540          IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
 541          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
 542     I->eraseFromParent();
 543   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
 544     IRB.CreateCall(
 545         isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
 546         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
 547          IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
 548          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
 549     I->eraseFromParent();
 550   }
 551   return false;
 552 }
 553
 554 static Value *createIntOrPtrToIntCast(Value *V, Type* Ty, IRBuilder<> &IRB) {
 555   return isa<PointerType>(V->getType()) ?
 556     IRB.CreatePtrToInt(V, Ty) : IRB.CreateIntCast(V, Ty, false);
 557 }
 558
 559 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
 560 // standards.  For background see C++11 standard.  A slightly older, publicly
 561 // available draft of the standard (not entirely up-to-date, but close enough
 562 // for casual browsing) is available here:
 563 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
 564 // The following page contains more background information:
 565 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
 566
 567 bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
 568   IRBuilder<> IRB(I);
 569   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
 570     Value *Addr = LI->getPointerOperand();
 571     int Idx = getMemoryAccessFuncIndex(Addr, DL);
 572     if (Idx < 0)
 573       return false;
 574     const unsigned ByteSize = 1U << Idx;
 575     const unsigned BitSize = ByteSize * 8;
 576     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 577     Type *PtrTy = Ty->getPointerTo();
 578     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 579                      createOrdering(&IRB, LI->getOrdering())};
 580     Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
 581     if (Ty == OrigTy) {
 582       Instruction *C = CallInst::Create(TsanAtomicLoad[Idx], Args);
 583       ReplaceInstWithInst(I, C);
 584     } else {
 585       // We are loading a pointer, so we need to cast the return value.
 586       Value *C = IRB.CreateCall(TsanAtomicLoad[Idx], Args);
 587       Instruction *Cast = CastInst::Create(Instruction::IntToPtr, C, OrigTy);
 588       ReplaceInstWithInst(I, Cast);
 589     }
 590   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
 591     Value *Addr = SI->getPointerOperand();
 592     int Idx = getMemoryAccessFuncIndex(Addr, DL);
 593     if (Idx < 0)
 594       return false;
 595     const unsigned ByteSize = 1U << Idx;
 596     const unsigned BitSize = ByteSize * 8;
 597     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 598     Type *PtrTy = Ty->getPointerTo();
 599     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 600                      createIntOrPtrToIntCast(SI->getValueOperand(), Ty, IRB),
 601                      createOrdering(&IRB, SI->getOrdering())};
 602     CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
 603     ReplaceInstWithInst(I, C);
 604   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
 605     Value *Addr = RMWI->getPointerOperand();
 606     int Idx = getMemoryAccessFuncIndex(Addr, DL);
 607     if (Idx < 0)
 608       return false;
 609     Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
 610     if (!F)
 611       return false;
 612     const unsigned ByteSize = 1U << Idx;
 613     const unsigned BitSize = ByteSize * 8;
 614     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 615     Type *PtrTy = Ty->getPointerTo();
 616     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 617                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
 618                      createOrdering(&IRB, RMWI->getOrdering())};
 619     CallInst *C = CallInst::Create(F, Args);
 620     ReplaceInstWithInst(I, C);
 621   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
 622     Value *Addr = CASI->getPointerOperand();
 623     int Idx = getMemoryAccessFuncIndex(Addr, DL);
 624     if (Idx < 0)
 625       return false;
 626     const unsigned ByteSize = 1U << Idx;
 627     const unsigned BitSize = ByteSize * 8;
 628     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 629     Type *PtrTy = Ty->getPointerTo();
 630     Value *CmpOperand =
 631       createIntOrPtrToIntCast(CASI->getCompareOperand(), Ty, IRB);
 632     Value *NewOperand =
 633       createIntOrPtrToIntCast(CASI->getNewValOperand(), Ty, IRB);
 634     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 635                      CmpOperand,
 636                      NewOperand,
 637                      createOrdering(&IRB, CASI->getSuccessOrdering()),
 638                      createOrdering(&IRB, CASI->getFailureOrdering())};
 639     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
 640     Value *Success = IRB.CreateICmpEQ(C, CmpOperand);
 641     Value *OldVal = C;
 642     Type *OrigOldValTy = CASI->getNewValOperand()->getType();
 643     if (Ty != OrigOldValTy) {
 644       // The value is a pointer, so we need to cast the return value.
 645       OldVal = IRB.CreateIntToPtr(C, OrigOldValTy);
 646     }
 647
 648     Value *Res =
 649       IRB.CreateInsertValue(UndefValue::get(CASI->getType()), OldVal, 0);
 650     Res = IRB.CreateInsertValue(Res, Success, 1);
 651
 652     I->replaceAllUsesWith(Res);
 653     I->eraseFromParent();
 654   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
 655     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
 656     Function *F = FI->getSynchScope() == SingleThread ?
 657         TsanAtomicSignalFence : TsanAtomicThreadFence;
 658     CallInst *C = CallInst::Create(F, Args);
 659     ReplaceInstWithInst(I, C);
 660   }
 661   return true;
 662 }
 663
 664 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr,
 665                                               const DataLayout &DL) {
 666   Type *OrigPtrTy = Addr->getType();
 667   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
 668   assert(OrigTy->isSized());
 669   uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
 670   if (TypeSize != 8  && TypeSize != 16 &&
 671       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
 672     NumAccessesWithBadSize++;
 673     // Ignore all unusual sizes.
 674     return -1;
 675   }
 676   size_t Idx = countTrailingZeros(TypeSize / 8);
 677   assert(Idx < kNumberOfAccessSizes);
 678   return Idx;
 679 }