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 namespace {
  76
  77 /// ThreadSanitizer: instrument the code in module to find races.
  78 struct ThreadSanitizer : public FunctionPass {
  79   ThreadSanitizer() : FunctionPass(ID) {}
  80   const char *getPassName() const override;
  81   bool runOnFunction(Function &F) override;
  82   bool doInitialization(Module &M) override;
  83   static char ID;  // Pass identification, replacement for typeid.
  84
  85  private:
  86   void initializeCallbacks(Module &M);
  87   bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL);
  88   bool instrumentAtomic(Instruction *I, const DataLayout &DL);
  89   bool instrumentMemIntrinsic(Instruction *I);
  90   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
  91                                       SmallVectorImpl<Instruction *> &All,
  92                                       const DataLayout &DL);
  93   bool addrPointsToConstantData(Value *Addr);
  94   int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL);
  95
  96   Type *IntptrTy;
  97   IntegerType *OrdTy;
  98   // Callbacks to run-time library are computed in doInitialization.
  99   Function *TsanFuncEntry;
 100   Function *TsanFuncExit;
 101   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
 102   static const size_t kNumberOfAccessSizes = 5;
 103   Function *TsanRead[kNumberOfAccessSizes];
 104   Function *TsanWrite[kNumberOfAccessSizes];
 105   Function *TsanUnalignedRead[kNumberOfAccessSizes];
 106   Function *TsanUnalignedWrite[kNumberOfAccessSizes];
 107   Function *TsanAtomicLoad[kNumberOfAccessSizes];
 108   Function *TsanAtomicStore[kNumberOfAccessSizes];
 109   Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes];
 110   Function *TsanAtomicCAS[kNumberOfAccessSizes];
 111   Function *TsanAtomicThreadFence;
 112   Function *TsanAtomicSignalFence;
 113   Function *TsanVptrUpdate;
 114   Function *TsanVptrLoad;
 115   Function *MemmoveFn, *MemcpyFn, *MemsetFn;
 116 };
 117 }  // namespace
 118
 119 char ThreadSanitizer::ID = 0;
 120 INITIALIZE_PASS(ThreadSanitizer, "tsan",
 121     "ThreadSanitizer: detects data races.",
 122     false, false)
 123
 124 const char *ThreadSanitizer::getPassName() const {
 125   return "ThreadSanitizer";
 126 }
 127
 128 FunctionPass *llvm::createThreadSanitizerPass() {
 129   return new ThreadSanitizer();
 130 }
 131
 132 void ThreadSanitizer::initializeCallbacks(Module &M) {
 133   IRBuilder<> IRB(M.getContext());
 134   // Initialize the callbacks.
 135   TsanFuncEntry = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 136       "__tsan_func_entry", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 137   TsanFuncExit = checkSanitizerInterfaceFunction(
 138       M.getOrInsertFunction("__tsan_func_exit", IRB.getVoidTy(), nullptr));
 139   OrdTy = IRB.getInt32Ty();
 140   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
 141     const size_t ByteSize = 1 << i;
 142     const size_t BitSize = ByteSize * 8;
 143     SmallString<32> ReadName("__tsan_read" + itostr(ByteSize));
 144     TsanRead[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 145         ReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 146
 147     SmallString<32> WriteName("__tsan_write" + itostr(ByteSize));
 148     TsanWrite[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 149         WriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 150
 151     SmallString<64> UnalignedReadName("__tsan_unaligned_read" +
 152         itostr(ByteSize));
 153     TsanUnalignedRead[i] =
 154         checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 155             UnalignedReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 156
 157     SmallString<64> UnalignedWriteName("__tsan_unaligned_write" +
 158         itostr(ByteSize));
 159     TsanUnalignedWrite[i] =
 160         checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 161             UnalignedWriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 162
 163     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
 164     Type *PtrTy = Ty->getPointerTo();
 165     SmallString<32> AtomicLoadName("__tsan_atomic" + itostr(BitSize) +
 166                                    "_load");
 167     TsanAtomicLoad[i] = checkSanitizerInterfaceFunction(
 168         M.getOrInsertFunction(AtomicLoadName, Ty, PtrTy, OrdTy, nullptr));
 169
 170     SmallString<32> AtomicStoreName("__tsan_atomic" + itostr(BitSize) +
 171                                     "_store");
 172     TsanAtomicStore[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 173         AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy, nullptr));
 174
 175     for (int op = AtomicRMWInst::FIRST_BINOP;
 176         op <= AtomicRMWInst::LAST_BINOP; ++op) {
 177       TsanAtomicRMW[op][i] = nullptr;
 178       const char *NamePart = nullptr;
 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] = checkSanitizerInterfaceFunction(
 197           M.getOrInsertFunction(RMWName, Ty, PtrTy, Ty, OrdTy, nullptr));
 198     }
 199
 200     SmallString<32> AtomicCASName("__tsan_atomic" + itostr(BitSize) +
 201                                   "_compare_exchange_val");
 202     TsanAtomicCAS[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 203         AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, nullptr));
 204   }
 205   TsanVptrUpdate = checkSanitizerInterfaceFunction(
 206       M.getOrInsertFunction("__tsan_vptr_update", IRB.getVoidTy(),
 207                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), nullptr));
 208   TsanVptrLoad = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 209       "__tsan_vptr_read", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 210   TsanAtomicThreadFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 211       "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, nullptr));
 212   TsanAtomicSignalFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
 213       "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, nullptr));
 214
 215   MemmoveFn = checkSanitizerInterfaceFunction(
 216       M.getOrInsertFunction("memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
 217                             IRB.getInt8PtrTy(), IntptrTy, nullptr));
 218   MemcpyFn = checkSanitizerInterfaceFunction(
 219       M.getOrInsertFunction("memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
 220                             IRB.getInt8PtrTy(), IntptrTy, nullptr));
 221   MemsetFn = checkSanitizerInterfaceFunction(
 222       M.getOrInsertFunction("memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
 223                             IRB.getInt32Ty(), IntptrTy, nullptr));
 224 }
 225
 226 bool ThreadSanitizer::doInitialization(Module &M) {
 227   const DataLayout &DL = M.getDataLayout();
 228
 229   // Always insert a call to __tsan_init into the module's CTORs.
 230   IRBuilder<> IRB(M.getContext());
 231   IntptrTy = IRB.getIntPtrTy(DL);
 232   Value *TsanInit = M.getOrInsertFunction("__tsan_init",
 233                                           IRB.getVoidTy(), nullptr);
 234   appendToGlobalCtors(M, cast<Function>(TsanInit), 0);
 235
 236   return true;
 237 }
 238
 239 static bool isVtableAccess(Instruction *I) {
 240   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
 241     return Tag->isTBAAVtableAccess();
 242   return false;
 243 }
 244
 245 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
 246   // If this is a GEP, just analyze its pointer operand.
 247   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
 248     Addr = GEP->getPointerOperand();
 249
 250   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
 251     if (GV->isConstant()) {
 252       // Reads from constant globals can not race with any writes.
 253       NumOmittedReadsFromConstantGlobals++;
 254       return true;
 255     }
 256   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
 257     if (isVtableAccess(L)) {
 258       // Reads from a vtable pointer can not race with any writes.
 259       NumOmittedReadsFromVtable++;
 260       return true;
 261     }
 262   }
 263   return false;
 264 }
 265
 266 // Instrumenting some of the accesses may be proven redundant.
 267 // Currently handled:
 268 //  - read-before-write (within same BB, no calls between)
 269 //  - not captured variables
 270 //
 271 // We do not handle some of the patterns that should not survive
 272 // after the classic compiler optimizations.
 273 // E.g. two reads from the same temp should be eliminated by CSE,
 274 // two writes should be eliminated by DSE, etc.
 275 //
 276 // 'Local' is a vector of insns within the same BB (no calls between).
 277 // 'All' is a vector of insns that will be instrumented.
 278 void ThreadSanitizer::chooseInstructionsToInstrument(
 279     SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
 280     const DataLayout &DL) {
 281   SmallSet<Value*, 8> WriteTargets;
 282   // Iterate from the end.
 283   for (SmallVectorImpl<Instruction*>::reverse_iterator It = Local.rbegin(),
 284        E = Local.rend(); It != E; ++It) {
 285     Instruction *I = *It;
 286     if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
 287       WriteTargets.insert(Store->getPointerOperand());
 288     } else {
 289       LoadInst *Load = cast<LoadInst>(I);
 290       Value *Addr = Load->getPointerOperand();
 291       if (WriteTargets.count(Addr)) {
 292         // We will write to this temp, so no reason to analyze the read.
 293         NumOmittedReadsBeforeWrite++;
 294         continue;
 295       }
 296       if (addrPointsToConstantData(Addr)) {
 297         // Addr points to some constant data -- it can not race with any writes.
 298         continue;
 299       }
 300     }
 301     Value *Addr = isa<StoreInst>(*I)
 302         ? cast<StoreInst>(I)->getPointerOperand()
 303         : cast<LoadInst>(I)->getPointerOperand();
 304     if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) &&
 305         !PointerMayBeCaptured(Addr, true, true)) {
 306       // The variable is addressable but not captured, so it cannot be
 307       // referenced from a different thread and participate in a data race
 308       // (see llvm/Analysis/CaptureTracking.h for details).
 309       NumOmittedNonCaptured++;
 310       continue;
 311     }
 312     All.push_back(I);
 313   }
 314   Local.clear();
 315 }
 316
 317 static bool isAtomic(Instruction *I) {
 318   if (LoadInst *LI = dyn_cast<LoadInst>(I))
 319     return LI->isAtomic() && LI->getSynchScope() == CrossThread;
 320   if (StoreInst *SI = dyn_cast<StoreInst>(I))
 321     return SI->isAtomic() && SI->getSynchScope() == CrossThread;
 322   if (isa<AtomicRMWInst>(I))
 323     return true;
 324   if (isa<AtomicCmpXchgInst>(I))
 325     return true;
 326   if (isa<FenceInst>(I))
 327     return true;
 328   return false;
 329 }
 330
 331 bool ThreadSanitizer::runOnFunction(Function &F) {
 332   initializeCallbacks(*F.getParent());
 333   SmallVector<Instruction*, 8> RetVec;
 334   SmallVector<Instruction*, 8> AllLoadsAndStores;
 335   SmallVector<Instruction*, 8> LocalLoadsAndStores;
 336   SmallVector<Instruction*, 8> AtomicAccesses;
 337   SmallVector<Instruction*, 8> MemIntrinCalls;
 338   bool Res = false;
 339   bool HasCalls = false;
 340   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
 341   const DataLayout &DL = F.getParent()->getDataLayout();
 342
 343   // Traverse all instructions, collect loads/stores/returns, check for calls.
 344   for (auto &BB : F) {
 345     for (auto &Inst : BB) {
 346       if (isAtomic(&Inst))
 347         AtomicAccesses.push_back(&Inst);
 348       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
 349         LocalLoadsAndStores.push_back(&Inst);
 350       else if (isa<ReturnInst>(Inst))
 351         RetVec.push_back(&Inst);
 352       else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
 353         if (isa<MemIntrinsic>(Inst))
 354           MemIntrinCalls.push_back(&Inst);
 355         HasCalls = true;
 356         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
 357                                        DL);
 358       }
 359     }
 360     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
 361   }
 362
 363   // We have collected all loads and stores.
 364   // FIXME: many of these accesses do not need to be checked for races
 365   // (e.g. variables that do not escape, etc).
 366
 367   // Instrument memory accesses only if we want to report bugs in the function.
 368   if (ClInstrumentMemoryAccesses && SanitizeFunction)
 369     for (auto Inst : AllLoadsAndStores) {
 370       Res |= instrumentLoadOrStore(Inst, DL);
 371     }
 372
 373   // Instrument atomic memory accesses in any case (they can be used to
 374   // implement synchronization).
 375   if (ClInstrumentAtomics)
 376     for (auto Inst : AtomicAccesses) {
 377       Res |= instrumentAtomic(Inst, DL);
 378     }
 379
 380   if (ClInstrumentMemIntrinsics && SanitizeFunction)
 381     for (auto Inst : MemIntrinCalls) {
 382       Res |= instrumentMemIntrinsic(Inst);
 383     }
 384
 385   // Instrument function entry/exit points if there were instrumented accesses.
 386   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
 387     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
 388     Value *ReturnAddress = IRB.CreateCall(
 389         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
 390         IRB.getInt32(0));
 391     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
 392     for (auto RetInst : RetVec) {
 393       IRBuilder<> IRBRet(RetInst);
 394       IRBRet.CreateCall(TsanFuncExit);
 395     }
 396     Res = true;
 397   }
 398   return Res;
 399 }
 400
 401 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I,
 402                                             const DataLayout &DL) {
 403   IRBuilder<> IRB(I);
 404   bool IsWrite = isa<StoreInst>(*I);
 405   Value *Addr = IsWrite
 406       ? cast<StoreInst>(I)->getPointerOperand()
 407       : cast<LoadInst>(I)->getPointerOperand();
 408   int Idx = getMemoryAccessFuncIndex(Addr, DL);
 409   if (Idx < 0)
 410     return false;
 411   if (IsWrite && isVtableAccess(I)) {
 412     DEBUG(dbgs() << "  VPTR : " << *I << "\n");
 413     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
 414     // StoredValue may be a vector type if we are storing several vptrs at once.
 415     // In this case, just take the first element of the vector since this is
 416     // enough to find vptr races.
 417     if (isa<VectorType>(StoredValue->getType()))
 418       StoredValue = IRB.CreateExtractElement(
 419           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
 420     if (StoredValue->getType()->isIntegerTy())
 421       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
 422     // Call TsanVptrUpdate.
 423     IRB.CreateCall2(TsanVptrUpdate,
 424                     IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
 425                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy()));
 426     NumInstrumentedVtableWrites++;
 427     return true;
 428   }
 429   if (!IsWrite && isVtableAccess(I)) {
 430     IRB.CreateCall(TsanVptrLoad,
 431                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
 432     NumInstrumentedVtableReads++;
 433     return true;
 434   }
 435   const unsigned Alignment = IsWrite
 436       ? cast<StoreInst>(I)->getAlignment()
 437       : cast<LoadInst>(I)->getAlignment();
 438   Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
 439   const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
 440   Value *OnAccessFunc = nullptr;
 441   if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0)
 442     OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
 443   else
 444     OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
 445   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
 446   if (IsWrite) NumInstrumentedWrites++;
 447   else         NumInstrumentedReads++;
 448   return true;
 449 }
 450
 451 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
 452   uint32_t v = 0;
 453   switch (ord) {
 454     case NotAtomic: llvm_unreachable("unexpected atomic ordering!");
 455     case Unordered:              // Fall-through.
 456     case Monotonic:              v = 0; break;
 457     // case Consume:                v = 1; break;  // Not specified yet.
 458     case Acquire:                v = 2; break;
 459     case Release:                v = 3; break;
 460     case AcquireRelease:         v = 4; break;
 461     case SequentiallyConsistent: v = 5; break;
 462   }
 463   return IRB->getInt32(v);
 464 }
 465
 466 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
 467 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
 468 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
 469 // instead we simply replace them with regular function calls, which are then
 470 // intercepted by the run-time.
 471 // Since tsan is running after everyone else, the calls should not be
 472 // replaced back with intrinsics. If that becomes wrong at some point,
 473 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
 474 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
 475   IRBuilder<> IRB(I);
 476   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
 477     IRB.CreateCall3(MemsetFn,
 478       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
 479       IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
 480       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
 481     I->eraseFromParent();
 482   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
 483     IRB.CreateCall3(isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
 484       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
 485       IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
 486       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
 487     I->eraseFromParent();
 488   }
 489   return false;
 490 }
 491
 492 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
 493 // standards.  For background see C++11 standard.  A slightly older, publicly
 494 // available draft of the standard (not entirely up-to-date, but close enough
 495 // for casual browsing) is available here:
 496 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
 497 // The following page contains more background information:
 498 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
 499
 500 bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
 501   IRBuilder<> IRB(I);
 502   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
 503     Value *Addr = LI->getPointerOperand();
 504     int Idx = getMemoryAccessFuncIndex(Addr, DL);
 505     if (Idx < 0)
 506       return false;
 507     const size_t ByteSize = 1 << Idx;
 508     const size_t BitSize = ByteSize * 8;
 509     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 510     Type *PtrTy = Ty->getPointerTo();
 511     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 512                      createOrdering(&IRB, LI->getOrdering())};
 513     CallInst *C = CallInst::Create(TsanAtomicLoad[Idx], Args);
 514     ReplaceInstWithInst(I, C);
 515
 516   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
 517     Value *Addr = SI->getPointerOperand();
 518     int Idx = getMemoryAccessFuncIndex(Addr, DL);
 519     if (Idx < 0)
 520       return false;
 521     const size_t ByteSize = 1 << Idx;
 522     const size_t BitSize = ByteSize * 8;
 523     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 524     Type *PtrTy = Ty->getPointerTo();
 525     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 526                      IRB.CreateIntCast(SI->getValueOperand(), Ty, false),
 527                      createOrdering(&IRB, SI->getOrdering())};
 528     CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
 529     ReplaceInstWithInst(I, C);
 530   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
 531     Value *Addr = RMWI->getPointerOperand();
 532     int Idx = getMemoryAccessFuncIndex(Addr, DL);
 533     if (Idx < 0)
 534       return false;
 535     Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
 536     if (!F)
 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, Args);
 546     ReplaceInstWithInst(I, C);
 547   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
 548     Value *Addr = CASI->getPointerOperand();
 549     int Idx = getMemoryAccessFuncIndex(Addr, DL);
 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->getSuccessOrdering()),
 560                      createOrdering(&IRB, CASI->getFailureOrdering())};
 561     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
 562     Value *Success = IRB.CreateICmpEQ(C, CASI->getCompareOperand());
 563
 564     Value *Res = IRB.CreateInsertValue(UndefValue::get(CASI->getType()), C, 0);
 565     Res = IRB.CreateInsertValue(Res, Success, 1);
 566
 567     I->replaceAllUsesWith(Res);
 568     I->eraseFromParent();
 569   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
 570     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
 571     Function *F = FI->getSynchScope() == SingleThread ?
 572         TsanAtomicSignalFence : TsanAtomicThreadFence;
 573     CallInst *C = CallInst::Create(F, Args);
 574     ReplaceInstWithInst(I, C);
 575   }
 576   return true;
 577 }
 578
 579 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr,
 580                                               const DataLayout &DL) {
 581   Type *OrigPtrTy = Addr->getType();
 582   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
 583   assert(OrigTy->isSized());
 584   uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
 585   if (TypeSize != 8  && TypeSize != 16 &&
 586       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
 587     NumAccessesWithBadSize++;
 588     // Ignore all unusual sizes.
 589     return -1;
 590   }
 591   size_t Idx = countTrailingZeros(TypeSize / 8);
 592   assert(Idx < kNumberOfAccessSizes);
 593   return Idx;
 594 }