1 //===-- ThreadSanitizer.cpp - race detector -------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file is a part of ThreadSanitizer, a race detector. 10 // 11 // The tool is under development, for the details about previous versions see 12 // http://code.google.com/p/data-race-test 13 // 14 // The instrumentation phase is quite simple: 15 // - Insert calls to run-time library before every memory access. 16 // - Optimizations may apply to avoid instrumenting some of the accesses. 17 // - Insert calls at function entry/exit. 18 // The rest is handled by the run-time library. 19 //===----------------------------------------------------------------------===// 20 21 #include "llvm/Transforms/Instrumentation/ThreadSanitizer.h" 22 #include "llvm/ADT/SmallPtrSet.h" 23 #include "llvm/ADT/SmallString.h" 24 #include "llvm/ADT/SmallVector.h" 25 #include "llvm/ADT/Statistic.h" 26 #include "llvm/ADT/StringExtras.h" 27 #include "llvm/Analysis/CaptureTracking.h" 28 #include "llvm/Analysis/TargetLibraryInfo.h" 29 #include "llvm/Transforms/Utils/Local.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/Instrumentation.h" 46 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 47 #include "llvm/Transforms/Utils/EscapeEnumerator.h" 48 #include "llvm/Transforms/Utils/ModuleUtils.h" 49 50 using namespace llvm; 51 52 #define DEBUG_TYPE "tsan" 53 54 static cl::opt<bool> ClInstrumentMemoryAccesses( 55 "tsan-instrument-memory-accesses", cl::init(true), 56 cl::desc("Instrument memory accesses"), cl::Hidden); 57 static cl::opt<bool> ClInstrumentFuncEntryExit( 58 "tsan-instrument-func-entry-exit", cl::init(true), 59 cl::desc("Instrument function entry and exit"), cl::Hidden); 60 static cl::opt<bool> ClHandleCxxExceptions( 61 "tsan-handle-cxx-exceptions", cl::init(true), 62 cl::desc("Handle C++ exceptions (insert cleanup blocks for unwinding)"), 63 cl::Hidden); 64 static cl::opt<bool> ClInstrumentAtomics( 65 "tsan-instrument-atomics", cl::init(true), 66 cl::desc("Instrument atomics"), cl::Hidden); 67 static cl::opt<bool> ClInstrumentMemIntrinsics( 68 "tsan-instrument-memintrinsics", cl::init(true), 69 cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden); 70 71 STATISTIC(NumInstrumentedReads, "Number of instrumented reads"); 72 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes"); 73 STATISTIC(NumOmittedReadsBeforeWrite, 74 "Number of reads ignored due to following writes"); 75 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size"); 76 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes"); 77 STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads"); 78 STATISTIC(NumOmittedReadsFromConstantGlobals, 79 "Number of reads from constant globals"); 80 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads"); 81 STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing"); 82 83 static const char *const kTsanModuleCtorName = "tsan.module_ctor"; 84 static const char *const kTsanInitName = "__tsan_init"; 85 86 namespace { 87 88 /// ThreadSanitizer: instrument the code in module to find races. 89 /// 90 /// Instantiating ThreadSanitizer inserts the tsan runtime library API function 91 /// declarations into the module if they don't exist already. Instantiating 92 /// ensures the __tsan_init function is in the list of global constructors for 93 /// the module. 94 struct ThreadSanitizer { 95 ThreadSanitizer(Module &M); 96 bool sanitizeFunction(Function &F, const TargetLibraryInfo &TLI); 97 98 private: 99 void initializeCallbacks(Module &M); 100 bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL); 101 bool instrumentAtomic(Instruction *I, const DataLayout &DL); 102 bool instrumentMemIntrinsic(Instruction *I); 103 void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local, 104 SmallVectorImpl<Instruction *> &All, 105 const DataLayout &DL); 106 bool addrPointsToConstantData(Value *Addr); 107 int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL); 108 void InsertRuntimeIgnores(Function &F); 109 110 Type *IntptrTy; 111 IntegerType *OrdTy; 112 // Callbacks to run-time library are computed in doInitialization. 113 Function *TsanFuncEntry; 114 Function *TsanFuncExit; 115 Function *TsanIgnoreBegin; 116 Function *TsanIgnoreEnd; 117 // Accesses sizes are powers of two: 1, 2, 4, 8, 16. 118 static const size_t kNumberOfAccessSizes = 5; 119 Function *TsanRead[kNumberOfAccessSizes]; 120 Function *TsanWrite[kNumberOfAccessSizes]; 121 Function *TsanUnalignedRead[kNumberOfAccessSizes]; 122 Function *TsanUnalignedWrite[kNumberOfAccessSizes]; 123 Function *TsanAtomicLoad[kNumberOfAccessSizes]; 124 Function *TsanAtomicStore[kNumberOfAccessSizes]; 125 Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes]; 126 Function *TsanAtomicCAS[kNumberOfAccessSizes]; 127 Function *TsanAtomicThreadFence; 128 Function *TsanAtomicSignalFence; 129 Function *TsanVptrUpdate; 130 Function *TsanVptrLoad; 131 Function *MemmoveFn, *MemcpyFn, *MemsetFn; 132 Function *TsanCtorFunction; 133 }; 134 135 struct ThreadSanitizerLegacyPass : FunctionPass { 136 ThreadSanitizerLegacyPass() : FunctionPass(ID) {} 137 StringRef getPassName() const override; 138 void getAnalysisUsage(AnalysisUsage &AU) const override; 139 bool runOnFunction(Function &F) override; 140 bool doInitialization(Module &M) override; 141 static char ID; // Pass identification, replacement for typeid. 142 private: 143 Optional<ThreadSanitizer> TSan; 144 }; 145 } // namespace 146 147 PreservedAnalyses ThreadSanitizerPass::run(Function &F, 148 FunctionAnalysisManager &FAM) { 149 ThreadSanitizer TSan(*F.getParent()); 150 if (TSan.sanitizeFunction(F, FAM.getResult<TargetLibraryAnalysis>(F))) 151 return PreservedAnalyses::none(); 152 return PreservedAnalyses::all(); 153 } 154 155 char ThreadSanitizerLegacyPass::ID = 0; 156 INITIALIZE_PASS_BEGIN(ThreadSanitizerLegacyPass, "tsan", 157 "ThreadSanitizer: detects data races.", false, false) 158 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 159 INITIALIZE_PASS_END(ThreadSanitizerLegacyPass, "tsan", 160 "ThreadSanitizer: detects data races.", false, false) 161 162 StringRef ThreadSanitizerLegacyPass::getPassName() const { 163 return "ThreadSanitizerLegacyPass"; 164 } 165 166 void ThreadSanitizerLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const { 167 AU.addRequired<TargetLibraryInfoWrapperPass>(); 168 } 169 170 bool ThreadSanitizerLegacyPass::doInitialization(Module &M) { 171 TSan.emplace(M); 172 return true; 173 } 174 175 bool ThreadSanitizerLegacyPass::runOnFunction(Function &F) { 176 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 177 TSan->sanitizeFunction(F, TLI); 178 return true; 179 } 180 181 FunctionPass *llvm::createThreadSanitizerLegacyPassPass() { 182 return new ThreadSanitizerLegacyPass(); 183 } 184 185 void ThreadSanitizer::initializeCallbacks(Module &M) { 186 IRBuilder<> IRB(M.getContext()); 187 AttributeList Attr; 188 Attr = Attr.addAttribute(M.getContext(), AttributeList::FunctionIndex, 189 Attribute::NoUnwind); 190 // Initialize the callbacks. 191 TsanFuncEntry = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 192 "__tsan_func_entry", Attr, IRB.getVoidTy(), IRB.getInt8PtrTy())); 193 TsanFuncExit = checkSanitizerInterfaceFunction( 194 M.getOrInsertFunction("__tsan_func_exit", Attr, IRB.getVoidTy())); 195 TsanIgnoreBegin = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 196 "__tsan_ignore_thread_begin", Attr, IRB.getVoidTy())); 197 TsanIgnoreEnd = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 198 "__tsan_ignore_thread_end", Attr, IRB.getVoidTy())); 199 OrdTy = IRB.getInt32Ty(); 200 for (size_t i = 0; i < kNumberOfAccessSizes; ++i) { 201 const unsigned ByteSize = 1U << i; 202 const unsigned BitSize = ByteSize * 8; 203 std::string ByteSizeStr = utostr(ByteSize); 204 std::string BitSizeStr = utostr(BitSize); 205 SmallString<32> ReadName("__tsan_read" + ByteSizeStr); 206 TsanRead[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 207 ReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy())); 208 209 SmallString<32> WriteName("__tsan_write" + ByteSizeStr); 210 TsanWrite[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 211 WriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy())); 212 213 SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr); 214 TsanUnalignedRead[i] = 215 checkSanitizerInterfaceFunction(M.getOrInsertFunction( 216 UnalignedReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy())); 217 218 SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr); 219 TsanUnalignedWrite[i] = 220 checkSanitizerInterfaceFunction(M.getOrInsertFunction( 221 UnalignedWriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy())); 222 223 Type *Ty = Type::getIntNTy(M.getContext(), BitSize); 224 Type *PtrTy = Ty->getPointerTo(); 225 SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load"); 226 TsanAtomicLoad[i] = checkSanitizerInterfaceFunction( 227 M.getOrInsertFunction(AtomicLoadName, Attr, Ty, PtrTy, OrdTy)); 228 229 SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store"); 230 TsanAtomicStore[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 231 AtomicStoreName, Attr, IRB.getVoidTy(), PtrTy, Ty, OrdTy)); 232 233 for (int op = AtomicRMWInst::FIRST_BINOP; 234 op <= AtomicRMWInst::LAST_BINOP; ++op) { 235 TsanAtomicRMW[op][i] = nullptr; 236 const char *NamePart = nullptr; 237 if (op == AtomicRMWInst::Xchg) 238 NamePart = "_exchange"; 239 else if (op == AtomicRMWInst::Add) 240 NamePart = "_fetch_add"; 241 else if (op == AtomicRMWInst::Sub) 242 NamePart = "_fetch_sub"; 243 else if (op == AtomicRMWInst::And) 244 NamePart = "_fetch_and"; 245 else if (op == AtomicRMWInst::Or) 246 NamePart = "_fetch_or"; 247 else if (op == AtomicRMWInst::Xor) 248 NamePart = "_fetch_xor"; 249 else if (op == AtomicRMWInst::Nand) 250 NamePart = "_fetch_nand"; 251 else 252 continue; 253 SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart); 254 TsanAtomicRMW[op][i] = checkSanitizerInterfaceFunction( 255 M.getOrInsertFunction(RMWName, Attr, Ty, PtrTy, Ty, OrdTy)); 256 } 257 258 SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr + 259 "_compare_exchange_val"); 260 TsanAtomicCAS[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 261 AtomicCASName, Attr, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy)); 262 } 263 TsanVptrUpdate = checkSanitizerInterfaceFunction( 264 M.getOrInsertFunction("__tsan_vptr_update", Attr, IRB.getVoidTy(), 265 IRB.getInt8PtrTy(), IRB.getInt8PtrTy())); 266 TsanVptrLoad = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 267 "__tsan_vptr_read", Attr, IRB.getVoidTy(), IRB.getInt8PtrTy())); 268 TsanAtomicThreadFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 269 "__tsan_atomic_thread_fence", Attr, IRB.getVoidTy(), OrdTy)); 270 TsanAtomicSignalFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction( 271 "__tsan_atomic_signal_fence", Attr, IRB.getVoidTy(), OrdTy)); 272 273 MemmoveFn = checkSanitizerInterfaceFunction( 274 M.getOrInsertFunction("memmove", Attr, IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 275 IRB.getInt8PtrTy(), IntptrTy)); 276 MemcpyFn = checkSanitizerInterfaceFunction( 277 M.getOrInsertFunction("memcpy", Attr, IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 278 IRB.getInt8PtrTy(), IntptrTy)); 279 MemsetFn = checkSanitizerInterfaceFunction( 280 M.getOrInsertFunction("memset", Attr, IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 281 IRB.getInt32Ty(), IntptrTy)); 282 } 283 284 ThreadSanitizer::ThreadSanitizer(Module &M) { 285 const DataLayout &DL = M.getDataLayout(); 286 IntptrTy = DL.getIntPtrType(M.getContext()); 287 std::tie(TsanCtorFunction, std::ignore) = 288 getOrCreateSanitizerCtorAndInitFunctions( 289 M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{}, 290 /*InitArgs=*/{}, 291 // This callback is invoked when the functions are created the first 292 // time. Hook them into the global ctors list in that case: 293 [&](Function *Ctor, Function *) { appendToGlobalCtors(M, Ctor, 0); }); 294 } 295 296 static bool isVtableAccess(Instruction *I) { 297 if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa)) 298 return Tag->isTBAAVtableAccess(); 299 return false; 300 } 301 302 // Do not instrument known races/"benign races" that come from compiler 303 // instrumentatin. The user has no way of suppressing them. 304 static bool shouldInstrumentReadWriteFromAddress(const Module *M, Value *Addr) { 305 // Peel off GEPs and BitCasts. 306 Addr = Addr->stripInBoundsOffsets(); 307 308 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 309 if (GV->hasSection()) { 310 StringRef SectionName = GV->getSection(); 311 // Check if the global is in the PGO counters section. 312 auto OF = Triple(M->getTargetTriple()).getObjectFormat(); 313 if (SectionName.endswith( 314 getInstrProfSectionName(IPSK_cnts, OF, /*AddSegmentInfo=*/false))) 315 return false; 316 } 317 318 // Check if the global is private gcov data. 319 if (GV->getName().startswith("__llvm_gcov") || 320 GV->getName().startswith("__llvm_gcda")) 321 return false; 322 } 323 324 // Do not instrument acesses from different address spaces; we cannot deal 325 // with them. 326 if (Addr) { 327 Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType()); 328 if (PtrTy->getPointerAddressSpace() != 0) 329 return false; 330 } 331 332 return true; 333 } 334 335 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) { 336 // If this is a GEP, just analyze its pointer operand. 337 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr)) 338 Addr = GEP->getPointerOperand(); 339 340 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) { 341 if (GV->isConstant()) { 342 // Reads from constant globals can not race with any writes. 343 NumOmittedReadsFromConstantGlobals++; 344 return true; 345 } 346 } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) { 347 if (isVtableAccess(L)) { 348 // Reads from a vtable pointer can not race with any writes. 349 NumOmittedReadsFromVtable++; 350 return true; 351 } 352 } 353 return false; 354 } 355 356 // Instrumenting some of the accesses may be proven redundant. 357 // Currently handled: 358 // - read-before-write (within same BB, no calls between) 359 // - not captured variables 360 // 361 // We do not handle some of the patterns that should not survive 362 // after the classic compiler optimizations. 363 // E.g. two reads from the same temp should be eliminated by CSE, 364 // two writes should be eliminated by DSE, etc. 365 // 366 // 'Local' is a vector of insns within the same BB (no calls between). 367 // 'All' is a vector of insns that will be instrumented. 368 void ThreadSanitizer::chooseInstructionsToInstrument( 369 SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All, 370 const DataLayout &DL) { 371 SmallPtrSet<Value*, 8> WriteTargets; 372 // Iterate from the end. 373 for (Instruction *I : reverse(Local)) { 374 if (StoreInst *Store = dyn_cast<StoreInst>(I)) { 375 Value *Addr = Store->getPointerOperand(); 376 if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr)) 377 continue; 378 WriteTargets.insert(Addr); 379 } else { 380 LoadInst *Load = cast<LoadInst>(I); 381 Value *Addr = Load->getPointerOperand(); 382 if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr)) 383 continue; 384 if (WriteTargets.count(Addr)) { 385 // We will write to this temp, so no reason to analyze the read. 386 NumOmittedReadsBeforeWrite++; 387 continue; 388 } 389 if (addrPointsToConstantData(Addr)) { 390 // Addr points to some constant data -- it can not race with any writes. 391 continue; 392 } 393 } 394 Value *Addr = isa<StoreInst>(*I) 395 ? cast<StoreInst>(I)->getPointerOperand() 396 : cast<LoadInst>(I)->getPointerOperand(); 397 if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) && 398 !PointerMayBeCaptured(Addr, true, true)) { 399 // The variable is addressable but not captured, so it cannot be 400 // referenced from a different thread and participate in a data race 401 // (see llvm/Analysis/CaptureTracking.h for details). 402 NumOmittedNonCaptured++; 403 continue; 404 } 405 All.push_back(I); 406 } 407 Local.clear(); 408 } 409 410 static bool isAtomic(Instruction *I) { 411 // TODO: Ask TTI whether synchronization scope is between threads. 412 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 413 return LI->isAtomic() && LI->getSyncScopeID() != SyncScope::SingleThread; 414 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 415 return SI->isAtomic() && SI->getSyncScopeID() != SyncScope::SingleThread; 416 if (isa<AtomicRMWInst>(I)) 417 return true; 418 if (isa<AtomicCmpXchgInst>(I)) 419 return true; 420 if (isa<FenceInst>(I)) 421 return true; 422 return false; 423 } 424 425 void ThreadSanitizer::InsertRuntimeIgnores(Function &F) { 426 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI()); 427 IRB.CreateCall(TsanIgnoreBegin); 428 EscapeEnumerator EE(F, "tsan_ignore_cleanup", ClHandleCxxExceptions); 429 while (IRBuilder<> *AtExit = EE.Next()) { 430 AtExit->CreateCall(TsanIgnoreEnd); 431 } 432 } 433 434 bool ThreadSanitizer::sanitizeFunction(Function &F, 435 const TargetLibraryInfo &TLI) { 436 // This is required to prevent instrumenting call to __tsan_init from within 437 // the module constructor. 438 if (&F == TsanCtorFunction) 439 return false; 440 initializeCallbacks(*F.getParent()); 441 SmallVector<Instruction*, 8> AllLoadsAndStores; 442 SmallVector<Instruction*, 8> LocalLoadsAndStores; 443 SmallVector<Instruction*, 8> AtomicAccesses; 444 SmallVector<Instruction*, 8> MemIntrinCalls; 445 bool Res = false; 446 bool HasCalls = false; 447 bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread); 448 const DataLayout &DL = F.getParent()->getDataLayout(); 449 450 // Traverse all instructions, collect loads/stores/returns, check for calls. 451 for (auto &BB : F) { 452 for (auto &Inst : BB) { 453 if (isAtomic(&Inst)) 454 AtomicAccesses.push_back(&Inst); 455 else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst)) 456 LocalLoadsAndStores.push_back(&Inst); 457 else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) { 458 if (CallInst *CI = dyn_cast<CallInst>(&Inst)) 459 maybeMarkSanitizerLibraryCallNoBuiltin(CI, &TLI); 460 if (isa<MemIntrinsic>(Inst)) 461 MemIntrinCalls.push_back(&Inst); 462 HasCalls = true; 463 chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, 464 DL); 465 } 466 } 467 chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL); 468 } 469 470 // We have collected all loads and stores. 471 // FIXME: many of these accesses do not need to be checked for races 472 // (e.g. variables that do not escape, etc). 473 474 // Instrument memory accesses only if we want to report bugs in the function. 475 if (ClInstrumentMemoryAccesses && SanitizeFunction) 476 for (auto Inst : AllLoadsAndStores) { 477 Res |= instrumentLoadOrStore(Inst, DL); 478 } 479 480 // Instrument atomic memory accesses in any case (they can be used to 481 // implement synchronization). 482 if (ClInstrumentAtomics) 483 for (auto Inst : AtomicAccesses) { 484 Res |= instrumentAtomic(Inst, DL); 485 } 486 487 if (ClInstrumentMemIntrinsics && SanitizeFunction) 488 for (auto Inst : MemIntrinCalls) { 489 Res |= instrumentMemIntrinsic(Inst); 490 } 491 492 if (F.hasFnAttribute("sanitize_thread_no_checking_at_run_time")) { 493 assert(!F.hasFnAttribute(Attribute::SanitizeThread)); 494 if (HasCalls) 495 InsertRuntimeIgnores(F); 496 } 497 498 // Instrument function entry/exit points if there were instrumented accesses. 499 if ((Res || HasCalls) && ClInstrumentFuncEntryExit) { 500 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI()); 501 Value *ReturnAddress = IRB.CreateCall( 502 Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress), 503 IRB.getInt32(0)); 504 IRB.CreateCall(TsanFuncEntry, ReturnAddress); 505 506 EscapeEnumerator EE(F, "tsan_cleanup", ClHandleCxxExceptions); 507 while (IRBuilder<> *AtExit = EE.Next()) { 508 AtExit->CreateCall(TsanFuncExit, {}); 509 } 510 Res = true; 511 } 512 return Res; 513 } 514 515 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I, 516 const DataLayout &DL) { 517 IRBuilder<> IRB(I); 518 bool IsWrite = isa<StoreInst>(*I); 519 Value *Addr = IsWrite 520 ? cast<StoreInst>(I)->getPointerOperand() 521 : cast<LoadInst>(I)->getPointerOperand(); 522 523 // swifterror memory addresses are mem2reg promoted by instruction selection. 524 // As such they cannot have regular uses like an instrumentation function and 525 // it makes no sense to track them as memory. 526 if (Addr->isSwiftError()) 527 return false; 528 529 int Idx = getMemoryAccessFuncIndex(Addr, DL); 530 if (Idx < 0) 531 return false; 532 if (IsWrite && isVtableAccess(I)) { 533 LLVM_DEBUG(dbgs() << " VPTR : " << *I << "\n"); 534 Value *StoredValue = cast<StoreInst>(I)->getValueOperand(); 535 // StoredValue may be a vector type if we are storing several vptrs at once. 536 // In this case, just take the first element of the vector since this is 537 // enough to find vptr races. 538 if (isa<VectorType>(StoredValue->getType())) 539 StoredValue = IRB.CreateExtractElement( 540 StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0)); 541 if (StoredValue->getType()->isIntegerTy()) 542 StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy()); 543 // Call TsanVptrUpdate. 544 IRB.CreateCall(TsanVptrUpdate, 545 {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()), 546 IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())}); 547 NumInstrumentedVtableWrites++; 548 return true; 549 } 550 if (!IsWrite && isVtableAccess(I)) { 551 IRB.CreateCall(TsanVptrLoad, 552 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy())); 553 NumInstrumentedVtableReads++; 554 return true; 555 } 556 const unsigned Alignment = IsWrite 557 ? cast<StoreInst>(I)->getAlignment() 558 : cast<LoadInst>(I)->getAlignment(); 559 Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType(); 560 const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy); 561 Value *OnAccessFunc = nullptr; 562 if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0) 563 OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx]; 564 else 565 OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx]; 566 IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy())); 567 if (IsWrite) NumInstrumentedWrites++; 568 else NumInstrumentedReads++; 569 return true; 570 } 571 572 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) { 573 uint32_t v = 0; 574 switch (ord) { 575 case AtomicOrdering::NotAtomic: 576 llvm_unreachable("unexpected atomic ordering!"); 577 case AtomicOrdering::Unordered: LLVM_FALLTHROUGH; 578 case AtomicOrdering::Monotonic: v = 0; break; 579 // Not specified yet: 580 // case AtomicOrdering::Consume: v = 1; break; 581 case AtomicOrdering::Acquire: v = 2; break; 582 case AtomicOrdering::Release: v = 3; break; 583 case AtomicOrdering::AcquireRelease: v = 4; break; 584 case AtomicOrdering::SequentiallyConsistent: v = 5; break; 585 } 586 return IRB->getInt32(v); 587 } 588 589 // If a memset intrinsic gets inlined by the code gen, we will miss races on it. 590 // So, we either need to ensure the intrinsic is not inlined, or instrument it. 591 // We do not instrument memset/memmove/memcpy intrinsics (too complicated), 592 // instead we simply replace them with regular function calls, which are then 593 // intercepted by the run-time. 594 // Since tsan is running after everyone else, the calls should not be 595 // replaced back with intrinsics. If that becomes wrong at some point, 596 // we will need to call e.g. __tsan_memset to avoid the intrinsics. 597 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) { 598 IRBuilder<> IRB(I); 599 if (MemSetInst *M = dyn_cast<MemSetInst>(I)) { 600 IRB.CreateCall( 601 MemsetFn, 602 {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()), 603 IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false), 604 IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)}); 605 I->eraseFromParent(); 606 } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) { 607 IRB.CreateCall( 608 isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn, 609 {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()), 610 IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()), 611 IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)}); 612 I->eraseFromParent(); 613 } 614 return false; 615 } 616 617 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x 618 // standards. For background see C++11 standard. A slightly older, publicly 619 // available draft of the standard (not entirely up-to-date, but close enough 620 // for casual browsing) is available here: 621 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf 622 // The following page contains more background information: 623 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/ 624 625 bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) { 626 IRBuilder<> IRB(I); 627 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 628 Value *Addr = LI->getPointerOperand(); 629 int Idx = getMemoryAccessFuncIndex(Addr, DL); 630 if (Idx < 0) 631 return false; 632 const unsigned ByteSize = 1U << Idx; 633 const unsigned BitSize = ByteSize * 8; 634 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize); 635 Type *PtrTy = Ty->getPointerTo(); 636 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy), 637 createOrdering(&IRB, LI->getOrdering())}; 638 Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType(); 639 Value *C = IRB.CreateCall(TsanAtomicLoad[Idx], Args); 640 Value *Cast = IRB.CreateBitOrPointerCast(C, OrigTy); 641 I->replaceAllUsesWith(Cast); 642 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 643 Value *Addr = SI->getPointerOperand(); 644 int Idx = getMemoryAccessFuncIndex(Addr, DL); 645 if (Idx < 0) 646 return false; 647 const unsigned ByteSize = 1U << Idx; 648 const unsigned BitSize = ByteSize * 8; 649 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize); 650 Type *PtrTy = Ty->getPointerTo(); 651 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy), 652 IRB.CreateBitOrPointerCast(SI->getValueOperand(), Ty), 653 createOrdering(&IRB, SI->getOrdering())}; 654 CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args); 655 ReplaceInstWithInst(I, C); 656 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) { 657 Value *Addr = RMWI->getPointerOperand(); 658 int Idx = getMemoryAccessFuncIndex(Addr, DL); 659 if (Idx < 0) 660 return false; 661 Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx]; 662 if (!F) 663 return false; 664 const unsigned ByteSize = 1U << Idx; 665 const unsigned BitSize = ByteSize * 8; 666 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize); 667 Type *PtrTy = Ty->getPointerTo(); 668 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy), 669 IRB.CreateIntCast(RMWI->getValOperand(), Ty, false), 670 createOrdering(&IRB, RMWI->getOrdering())}; 671 CallInst *C = CallInst::Create(F, Args); 672 ReplaceInstWithInst(I, C); 673 } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) { 674 Value *Addr = CASI->getPointerOperand(); 675 int Idx = getMemoryAccessFuncIndex(Addr, DL); 676 if (Idx < 0) 677 return false; 678 const unsigned ByteSize = 1U << Idx; 679 const unsigned BitSize = ByteSize * 8; 680 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize); 681 Type *PtrTy = Ty->getPointerTo(); 682 Value *CmpOperand = 683 IRB.CreateBitOrPointerCast(CASI->getCompareOperand(), Ty); 684 Value *NewOperand = 685 IRB.CreateBitOrPointerCast(CASI->getNewValOperand(), Ty); 686 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy), 687 CmpOperand, 688 NewOperand, 689 createOrdering(&IRB, CASI->getSuccessOrdering()), 690 createOrdering(&IRB, CASI->getFailureOrdering())}; 691 CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args); 692 Value *Success = IRB.CreateICmpEQ(C, CmpOperand); 693 Value *OldVal = C; 694 Type *OrigOldValTy = CASI->getNewValOperand()->getType(); 695 if (Ty != OrigOldValTy) { 696 // The value is a pointer, so we need to cast the return value. 697 OldVal = IRB.CreateIntToPtr(C, OrigOldValTy); 698 } 699 700 Value *Res = 701 IRB.CreateInsertValue(UndefValue::get(CASI->getType()), OldVal, 0); 702 Res = IRB.CreateInsertValue(Res, Success, 1); 703 704 I->replaceAllUsesWith(Res); 705 I->eraseFromParent(); 706 } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) { 707 Value *Args[] = {createOrdering(&IRB, FI->getOrdering())}; 708 Function *F = FI->getSyncScopeID() == SyncScope::SingleThread ? 709 TsanAtomicSignalFence : TsanAtomicThreadFence; 710 CallInst *C = CallInst::Create(F, Args); 711 ReplaceInstWithInst(I, C); 712 } 713 return true; 714 } 715 716 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr, 717 const DataLayout &DL) { 718 Type *OrigPtrTy = Addr->getType(); 719 Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType(); 720 assert(OrigTy->isSized()); 721 uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy); 722 if (TypeSize != 8 && TypeSize != 16 && 723 TypeSize != 32 && TypeSize != 64 && TypeSize != 128) { 724 NumAccessesWithBadSize++; 725 // Ignore all unusual sizes. 726 return -1; 727 } 728 size_t Idx = countTrailingZeros(TypeSize / 8); 729 assert(Idx < kNumberOfAccessSizes); 730 return Idx; 731 } 732