ThreadSanitizer.cpp revision 243830
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 "BlackList.h" 25#include "llvm/Function.h" 26#include "llvm/IRBuilder.h" 27#include "llvm/Intrinsics.h" 28#include "llvm/LLVMContext.h" 29#include "llvm/Metadata.h" 30#include "llvm/Module.h" 31#include "llvm/Type.h" 32#include "llvm/ADT/SmallSet.h" 33#include "llvm/ADT/SmallString.h" 34#include "llvm/ADT/SmallVector.h" 35#include "llvm/ADT/Statistic.h" 36#include "llvm/ADT/StringExtras.h" 37#include "llvm/Support/CommandLine.h" 38#include "llvm/Support/Debug.h" 39#include "llvm/Support/MathExtras.h" 40#include "llvm/Support/raw_ostream.h" 41#include "llvm/DataLayout.h" 42#include "llvm/Transforms/Instrumentation.h" 43#include "llvm/Transforms/Utils/BasicBlockUtils.h" 44#include "llvm/Transforms/Utils/ModuleUtils.h" 45 46using namespace llvm; 47 48static cl::opt<std::string> ClBlackListFile("tsan-blacklist", 49 cl::desc("Blacklist file"), cl::Hidden); 50static cl::opt<bool> ClInstrumentMemoryAccesses( 51 "tsan-instrument-memory-accesses", cl::init(true), 52 cl::desc("Instrument memory accesses"), cl::Hidden); 53static cl::opt<bool> ClInstrumentFuncEntryExit( 54 "tsan-instrument-func-entry-exit", cl::init(true), 55 cl::desc("Instrument function entry and exit"), cl::Hidden); 56static cl::opt<bool> ClInstrumentAtomics( 57 "tsan-instrument-atomics", cl::init(true), 58 cl::desc("Instrument atomics"), cl::Hidden); 59 60STATISTIC(NumInstrumentedReads, "Number of instrumented reads"); 61STATISTIC(NumInstrumentedWrites, "Number of instrumented writes"); 62STATISTIC(NumOmittedReadsBeforeWrite, 63 "Number of reads ignored due to following writes"); 64STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size"); 65STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes"); 66STATISTIC(NumOmittedReadsFromConstantGlobals, 67 "Number of reads from constant globals"); 68STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads"); 69 70namespace { 71 72/// ThreadSanitizer: instrument the code in module to find races. 73struct ThreadSanitizer : public FunctionPass { 74 ThreadSanitizer(); 75 const char *getPassName() const; 76 bool runOnFunction(Function &F); 77 bool doInitialization(Module &M); 78 static char ID; // Pass identification, replacement for typeid. 79 80 private: 81 bool instrumentLoadOrStore(Instruction *I); 82 bool instrumentAtomic(Instruction *I); 83 void chooseInstructionsToInstrument(SmallVectorImpl<Instruction*> &Local, 84 SmallVectorImpl<Instruction*> &All); 85 bool addrPointsToConstantData(Value *Addr); 86 int getMemoryAccessFuncIndex(Value *Addr); 87 88 DataLayout *TD; 89 OwningPtr<BlackList> BL; 90 IntegerType *OrdTy; 91 // Callbacks to run-time library are computed in doInitialization. 92 Function *TsanFuncEntry; 93 Function *TsanFuncExit; 94 // Accesses sizes are powers of two: 1, 2, 4, 8, 16. 95 static const size_t kNumberOfAccessSizes = 5; 96 Function *TsanRead[kNumberOfAccessSizes]; 97 Function *TsanWrite[kNumberOfAccessSizes]; 98 Function *TsanAtomicLoad[kNumberOfAccessSizes]; 99 Function *TsanAtomicStore[kNumberOfAccessSizes]; 100 Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes]; 101 Function *TsanAtomicCAS[kNumberOfAccessSizes]; 102 Function *TsanAtomicThreadFence; 103 Function *TsanAtomicSignalFence; 104 Function *TsanVptrUpdate; 105}; 106} // namespace 107 108char ThreadSanitizer::ID = 0; 109INITIALIZE_PASS(ThreadSanitizer, "tsan", 110 "ThreadSanitizer: detects data races.", 111 false, false) 112 113const char *ThreadSanitizer::getPassName() const { 114 return "ThreadSanitizer"; 115} 116 117ThreadSanitizer::ThreadSanitizer() 118 : FunctionPass(ID), 119 TD(NULL) { 120} 121 122FunctionPass *llvm::createThreadSanitizerPass() { 123 return new ThreadSanitizer(); 124} 125 126static Function *checkInterfaceFunction(Constant *FuncOrBitcast) { 127 if (Function *F = dyn_cast<Function>(FuncOrBitcast)) 128 return F; 129 FuncOrBitcast->dump(); 130 report_fatal_error("ThreadSanitizer interface function redefined"); 131} 132 133bool ThreadSanitizer::doInitialization(Module &M) { 134 TD = getAnalysisIfAvailable<DataLayout>(); 135 if (!TD) 136 return false; 137 BL.reset(new BlackList(ClBlackListFile)); 138 139 // Always insert a call to __tsan_init into the module's CTORs. 140 IRBuilder<> IRB(M.getContext()); 141 Value *TsanInit = M.getOrInsertFunction("__tsan_init", 142 IRB.getVoidTy(), NULL); 143 appendToGlobalCtors(M, cast<Function>(TsanInit), 0); 144 145 // 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 192 continue; 193 SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart); 194 TsanAtomicRMW[op][i] = checkInterfaceFunction(M.getOrInsertFunction( 195 RMWName, Ty, PtrTy, Ty, OrdTy, NULL)); 196 } 197 198 SmallString<32> AtomicCASName("__tsan_atomic" + itostr(BitSize) + 199 "_compare_exchange_val"); 200 TsanAtomicCAS[i] = checkInterfaceFunction(M.getOrInsertFunction( 201 AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, NULL)); 202 } 203 TsanVptrUpdate = checkInterfaceFunction(M.getOrInsertFunction( 204 "__tsan_vptr_update", IRB.getVoidTy(), IRB.getInt8PtrTy(), 205 IRB.getInt8PtrTy(), NULL)); 206 TsanAtomicThreadFence = checkInterfaceFunction(M.getOrInsertFunction( 207 "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, NULL)); 208 TsanAtomicSignalFence = checkInterfaceFunction(M.getOrInsertFunction( 209 "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, NULL)); 210 return true; 211} 212 213static bool isVtableAccess(Instruction *I) { 214 if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa)) { 215 if (Tag->getNumOperands() < 1) return false; 216 if (MDString *Tag1 = dyn_cast<MDString>(Tag->getOperand(0))) { 217 if (Tag1->getString() == "vtable pointer") return true; 218 } 219 } 220 return false; 221} 222 223bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) { 224 // If this is a GEP, just analyze its pointer operand. 225 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr)) 226 Addr = GEP->getPointerOperand(); 227 228 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 229 if (GV->isConstant()) { 230 // Reads from constant globals can not race with any writes. 231 NumOmittedReadsFromConstantGlobals++; 232 return true; 233 } 234 } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) { 235 if (isVtableAccess(L)) { 236 // Reads from a vtable pointer can not race with any writes. 237 NumOmittedReadsFromVtable++; 238 return true; 239 } 240 } 241 return false; 242} 243 244// Instrumenting some of the accesses may be proven redundant. 245// Currently handled: 246// - read-before-write (within same BB, no calls between) 247// 248// We do not handle some of the patterns that should not survive 249// after the classic compiler optimizations. 250// E.g. two reads from the same temp should be eliminated by CSE, 251// two writes should be eliminated by DSE, etc. 252// 253// 'Local' is a vector of insns within the same BB (no calls between). 254// 'All' is a vector of insns that will be instrumented. 255void ThreadSanitizer::chooseInstructionsToInstrument( 256 SmallVectorImpl<Instruction*> &Local, 257 SmallVectorImpl<Instruction*> &All) { 258 SmallSet<Value*, 8> WriteTargets; 259 // Iterate from the end. 260 for (SmallVectorImpl<Instruction*>::reverse_iterator It = Local.rbegin(), 261 E = Local.rend(); It != E; ++It) { 262 Instruction *I = *It; 263 if (StoreInst *Store = dyn_cast<StoreInst>(I)) { 264 WriteTargets.insert(Store->getPointerOperand()); 265 } else { 266 LoadInst *Load = cast<LoadInst>(I); 267 Value *Addr = Load->getPointerOperand(); 268 if (WriteTargets.count(Addr)) { 269 // We will write to this temp, so no reason to analyze the read. 270 NumOmittedReadsBeforeWrite++; 271 continue; 272 } 273 if (addrPointsToConstantData(Addr)) { 274 // Addr points to some constant data -- it can not race with any writes. 275 continue; 276 } 277 } 278 All.push_back(I); 279 } 280 Local.clear(); 281} 282 283static bool isAtomic(Instruction *I) { 284 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 285 return LI->isAtomic() && LI->getSynchScope() == CrossThread; 286 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 287 return SI->isAtomic() && SI->getSynchScope() == CrossThread; 288 if (isa<AtomicRMWInst>(I)) 289 return true; 290 if (isa<AtomicCmpXchgInst>(I)) 291 return true; 292 if (isa<FenceInst>(I)) 293 return true; 294 return false; 295} 296 297bool ThreadSanitizer::runOnFunction(Function &F) { 298 if (!TD) return false; 299 if (BL->isIn(F)) return false; 300 SmallVector<Instruction*, 8> RetVec; 301 SmallVector<Instruction*, 8> AllLoadsAndStores; 302 SmallVector<Instruction*, 8> LocalLoadsAndStores; 303 SmallVector<Instruction*, 8> AtomicAccesses; 304 bool Res = false; 305 bool HasCalls = false; 306 307 // Traverse all instructions, collect loads/stores/returns, check for calls. 308 for (Function::iterator FI = F.begin(), FE = F.end(); 309 FI != FE; ++FI) { 310 BasicBlock &BB = *FI; 311 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); 312 BI != BE; ++BI) { 313 if (isAtomic(BI)) 314 AtomicAccesses.push_back(BI); 315 else if (isa<LoadInst>(BI) || isa<StoreInst>(BI)) 316 LocalLoadsAndStores.push_back(BI); 317 else if (isa<ReturnInst>(BI)) 318 RetVec.push_back(BI); 319 else if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) { 320 HasCalls = true; 321 chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores); 322 } 323 } 324 chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores); 325 } 326 327 // We have collected all loads and stores. 328 // FIXME: many of these accesses do not need to be checked for races 329 // (e.g. variables that do not escape, etc). 330 331 // Instrument memory accesses. 332 if (ClInstrumentMemoryAccesses) 333 for (size_t i = 0, n = AllLoadsAndStores.size(); i < n; ++i) { 334 Res |= instrumentLoadOrStore(AllLoadsAndStores[i]); 335 } 336 337 // Instrument atomic memory accesses. 338 if (ClInstrumentAtomics) 339 for (size_t i = 0, n = AtomicAccesses.size(); i < n; ++i) { 340 Res |= instrumentAtomic(AtomicAccesses[i]); 341 } 342 343 // Instrument function entry/exit points if there were instrumented accesses. 344 if ((Res || HasCalls) && ClInstrumentFuncEntryExit) { 345 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI()); 346 Value *ReturnAddress = IRB.CreateCall( 347 Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress), 348 IRB.getInt32(0)); 349 IRB.CreateCall(TsanFuncEntry, ReturnAddress); 350 for (size_t i = 0, n = RetVec.size(); i < n; ++i) { 351 IRBuilder<> IRBRet(RetVec[i]); 352 IRBRet.CreateCall(TsanFuncExit); 353 } 354 Res = true; 355 } 356 return Res; 357} 358 359bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I) { 360 IRBuilder<> IRB(I); 361 bool IsWrite = isa<StoreInst>(*I); 362 Value *Addr = IsWrite 363 ? cast<StoreInst>(I)->getPointerOperand() 364 : cast<LoadInst>(I)->getPointerOperand(); 365 int Idx = getMemoryAccessFuncIndex(Addr); 366 if (Idx < 0) 367 return false; 368 if (IsWrite && isVtableAccess(I)) { 369 DEBUG(dbgs() << " VPTR : " << *I << "\n"); 370 Value *StoredValue = cast<StoreInst>(I)->getValueOperand(); 371 // StoredValue does not necessary have a pointer type. 372 if (isa<IntegerType>(StoredValue->getType())) 373 StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy()); 374 // Call TsanVptrUpdate. 375 IRB.CreateCall2(TsanVptrUpdate, 376 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()), 377 IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())); 378 NumInstrumentedVtableWrites++; 379 return true; 380 } 381 Value *OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx]; 382 IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy())); 383 if (IsWrite) NumInstrumentedWrites++; 384 else NumInstrumentedReads++; 385 return true; 386} 387 388static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) { 389 uint32_t v = 0; 390 switch (ord) { 391 case NotAtomic: assert(false); 392 case Unordered: // Fall-through. 393 case Monotonic: v = 0; break; 394 // case Consume: v = 1; break; // Not specified yet. 395 case Acquire: v = 2; break; 396 case Release: v = 3; break; 397 case AcquireRelease: v = 4; break; 398 case SequentiallyConsistent: v = 5; break; 399 } 400 return IRB->getInt32(v); 401} 402 403bool ThreadSanitizer::instrumentAtomic(Instruction *I) { 404 IRBuilder<> IRB(I); 405 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 406 Value *Addr = LI->getPointerOperand(); 407 int Idx = getMemoryAccessFuncIndex(Addr); 408 if (Idx < 0) 409 return false; 410 const size_t ByteSize = 1 << Idx; 411 const size_t BitSize = ByteSize * 8; 412 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize); 413 Type *PtrTy = Ty->getPointerTo(); 414 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy), 415 createOrdering(&IRB, LI->getOrdering())}; 416 CallInst *C = CallInst::Create(TsanAtomicLoad[Idx], 417 ArrayRef<Value*>(Args)); 418 ReplaceInstWithInst(I, C); 419 420 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 421 Value *Addr = SI->getPointerOperand(); 422 int Idx = getMemoryAccessFuncIndex(Addr); 423 if (Idx < 0) 424 return false; 425 const size_t ByteSize = 1 << Idx; 426 const size_t BitSize = ByteSize * 8; 427 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize); 428 Type *PtrTy = Ty->getPointerTo(); 429 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy), 430 IRB.CreateIntCast(SI->getValueOperand(), Ty, false), 431 createOrdering(&IRB, SI->getOrdering())}; 432 CallInst *C = CallInst::Create(TsanAtomicStore[Idx], 433 ArrayRef<Value*>(Args)); 434 ReplaceInstWithInst(I, C); 435 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) { 436 Value *Addr = RMWI->getPointerOperand(); 437 int Idx = getMemoryAccessFuncIndex(Addr); 438 if (Idx < 0) 439 return false; 440 Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx]; 441 if (F == NULL) 442 return false; 443 const size_t ByteSize = 1 << Idx; 444 const size_t BitSize = ByteSize * 8; 445 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize); 446 Type *PtrTy = Ty->getPointerTo(); 447 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy), 448 IRB.CreateIntCast(RMWI->getValOperand(), Ty, false), 449 createOrdering(&IRB, RMWI->getOrdering())}; 450 CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args)); 451 ReplaceInstWithInst(I, C); 452 } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) { 453 Value *Addr = CASI->getPointerOperand(); 454 int Idx = getMemoryAccessFuncIndex(Addr); 455 if (Idx < 0) 456 return false; 457 const size_t ByteSize = 1 << Idx; 458 const size_t BitSize = ByteSize * 8; 459 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize); 460 Type *PtrTy = Ty->getPointerTo(); 461 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy), 462 IRB.CreateIntCast(CASI->getCompareOperand(), Ty, false), 463 IRB.CreateIntCast(CASI->getNewValOperand(), Ty, false), 464 createOrdering(&IRB, CASI->getOrdering())}; 465 CallInst *C = CallInst::Create(TsanAtomicCAS[Idx], ArrayRef<Value*>(Args)); 466 ReplaceInstWithInst(I, C); 467 } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) { 468 Value *Args[] = {createOrdering(&IRB, FI->getOrdering())}; 469 Function *F = FI->getSynchScope() == SingleThread ? 470 TsanAtomicSignalFence : TsanAtomicThreadFence; 471 CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args)); 472 ReplaceInstWithInst(I, C); 473 } 474 return true; 475} 476 477int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr) { 478 Type *OrigPtrTy = Addr->getType(); 479 Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType(); 480 assert(OrigTy->isSized()); 481 uint32_t TypeSize = TD->getTypeStoreSizeInBits(OrigTy); 482 if (TypeSize != 8 && TypeSize != 16 && 483 TypeSize != 32 && TypeSize != 64 && TypeSize != 128) { 484 NumAccessesWithBadSize++; 485 // Ignore all unusual sizes. 486 return -1; 487 } 488 size_t Idx = CountTrailingZeros_32(TypeSize / 8); 489 assert(Idx < kNumberOfAccessSizes); 490 return Idx; 491} 492