1 //===-- ThreadSanitizer.cpp - race detector -------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file is a part of ThreadSanitizer, a race detector.
12 // The tool is under development, for the details about previous versions see
13 // http://code.google.com/p/data-race-test
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 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "tsan"
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"
48 static cl::opt<std::string> ClBlackListFile("tsan-blacklist",
49 cl::desc("Blacklist file"), cl::Hidden);
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);
60 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
61 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
62 STATISTIC(NumOmittedReadsBeforeWrite,
63 "Number of reads ignored due to following writes");
64 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
65 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
66 STATISTIC(NumOmittedReadsFromConstantGlobals,
67 "Number of reads from constant globals");
68 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
72 /// ThreadSanitizer: instrument the code in module to find races.
73 struct ThreadSanitizer : public FunctionPass {
75 const char *getPassName() const;
76 bool runOnFunction(Function &F);
77 bool doInitialization(Module &M);
78 static char ID; // Pass identification, replacement for typeid.
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);
89 OwningPtr<BlackList> BL;
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;
108 char ThreadSanitizer::ID = 0;
109 INITIALIZE_PASS(ThreadSanitizer, "tsan",
110 "ThreadSanitizer: detects data races.",
113 const char *ThreadSanitizer::getPassName() const {
114 return "ThreadSanitizer";
117 ThreadSanitizer::ThreadSanitizer()
122 FunctionPass *llvm::createThreadSanitizerPass() {
123 return new ThreadSanitizer();
126 static Function *checkInterfaceFunction(Constant *FuncOrBitcast) {
127 if (Function *F = dyn_cast<Function>(FuncOrBitcast))
129 FuncOrBitcast->dump();
130 report_fatal_error("ThreadSanitizer interface function redefined");
133 bool ThreadSanitizer::doInitialization(Module &M) {
134 TD = getAnalysisIfAvailable<DataLayout>();
137 BL.reset(new BlackList(ClBlackListFile));
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);
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));
158 SmallString<32> WriteName("__tsan_write" + itostr(ByteSize));
159 TsanWrite[i] = checkInterfaceFunction(M.getOrInsertFunction(
160 WriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
162 Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
163 Type *PtrTy = Ty->getPointerTo();
164 SmallString<32> AtomicLoadName("__tsan_atomic" + itostr(BitSize) +
166 TsanAtomicLoad[i] = checkInterfaceFunction(M.getOrInsertFunction(
167 AtomicLoadName, Ty, PtrTy, OrdTy, NULL));
169 SmallString<32> AtomicStoreName("__tsan_atomic" + itostr(BitSize) +
171 TsanAtomicStore[i] = checkInterfaceFunction(M.getOrInsertFunction(
172 AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy,
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 if (op == AtomicRMWInst::Nand)
192 NamePart = "_fetch_nand";
195 SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
196 TsanAtomicRMW[op][i] = checkInterfaceFunction(M.getOrInsertFunction(
197 RMWName, Ty, PtrTy, Ty, OrdTy, NULL));
200 SmallString<32> AtomicCASName("__tsan_atomic" + itostr(BitSize) +
201 "_compare_exchange_val");
202 TsanAtomicCAS[i] = checkInterfaceFunction(M.getOrInsertFunction(
203 AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, NULL));
205 TsanVptrUpdate = checkInterfaceFunction(M.getOrInsertFunction(
206 "__tsan_vptr_update", IRB.getVoidTy(), IRB.getInt8PtrTy(),
207 IRB.getInt8PtrTy(), NULL));
208 TsanAtomicThreadFence = checkInterfaceFunction(M.getOrInsertFunction(
209 "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, NULL));
210 TsanAtomicSignalFence = checkInterfaceFunction(M.getOrInsertFunction(
211 "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, NULL));
215 static bool isVtableAccess(Instruction *I) {
216 if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa)) {
217 if (Tag->getNumOperands() < 1) return false;
218 if (MDString *Tag1 = dyn_cast<MDString>(Tag->getOperand(0))) {
219 if (Tag1->getString() == "vtable pointer") return true;
225 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
226 // If this is a GEP, just analyze its pointer operand.
227 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
228 Addr = GEP->getPointerOperand();
230 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
231 if (GV->isConstant()) {
232 // Reads from constant globals can not race with any writes.
233 NumOmittedReadsFromConstantGlobals++;
236 } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
237 if (isVtableAccess(L)) {
238 // Reads from a vtable pointer can not race with any writes.
239 NumOmittedReadsFromVtable++;
246 // Instrumenting some of the accesses may be proven redundant.
247 // Currently handled:
248 // - read-before-write (within same BB, no calls between)
250 // We do not handle some of the patterns that should not survive
251 // after the classic compiler optimizations.
252 // E.g. two reads from the same temp should be eliminated by CSE,
253 // two writes should be eliminated by DSE, etc.
255 // 'Local' is a vector of insns within the same BB (no calls between).
256 // 'All' is a vector of insns that will be instrumented.
257 void ThreadSanitizer::chooseInstructionsToInstrument(
258 SmallVectorImpl<Instruction*> &Local,
259 SmallVectorImpl<Instruction*> &All) {
260 SmallSet<Value*, 8> WriteTargets;
261 // Iterate from the end.
262 for (SmallVectorImpl<Instruction*>::reverse_iterator It = Local.rbegin(),
263 E = Local.rend(); It != E; ++It) {
264 Instruction *I = *It;
265 if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
266 WriteTargets.insert(Store->getPointerOperand());
268 LoadInst *Load = cast<LoadInst>(I);
269 Value *Addr = Load->getPointerOperand();
270 if (WriteTargets.count(Addr)) {
271 // We will write to this temp, so no reason to analyze the read.
272 NumOmittedReadsBeforeWrite++;
275 if (addrPointsToConstantData(Addr)) {
276 // Addr points to some constant data -- it can not race with any writes.
285 static bool isAtomic(Instruction *I) {
286 if (LoadInst *LI = dyn_cast<LoadInst>(I))
287 return LI->isAtomic() && LI->getSynchScope() == CrossThread;
288 if (StoreInst *SI = dyn_cast<StoreInst>(I))
289 return SI->isAtomic() && SI->getSynchScope() == CrossThread;
290 if (isa<AtomicRMWInst>(I))
292 if (isa<AtomicCmpXchgInst>(I))
294 if (isa<FenceInst>(I))
299 bool ThreadSanitizer::runOnFunction(Function &F) {
300 if (!TD) return false;
301 if (BL->isIn(F)) return false;
302 SmallVector<Instruction*, 8> RetVec;
303 SmallVector<Instruction*, 8> AllLoadsAndStores;
304 SmallVector<Instruction*, 8> LocalLoadsAndStores;
305 SmallVector<Instruction*, 8> AtomicAccesses;
307 bool HasCalls = false;
309 // Traverse all instructions, collect loads/stores/returns, check for calls.
310 for (Function::iterator FI = F.begin(), FE = F.end();
312 BasicBlock &BB = *FI;
313 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end();
316 AtomicAccesses.push_back(BI);
317 else if (isa<LoadInst>(BI) || isa<StoreInst>(BI))
318 LocalLoadsAndStores.push_back(BI);
319 else if (isa<ReturnInst>(BI))
320 RetVec.push_back(BI);
321 else if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
323 chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
326 chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
329 // We have collected all loads and stores.
330 // FIXME: many of these accesses do not need to be checked for races
331 // (e.g. variables that do not escape, etc).
333 // Instrument memory accesses.
334 if (ClInstrumentMemoryAccesses)
335 for (size_t i = 0, n = AllLoadsAndStores.size(); i < n; ++i) {
336 Res |= instrumentLoadOrStore(AllLoadsAndStores[i]);
339 // Instrument atomic memory accesses.
340 if (ClInstrumentAtomics)
341 for (size_t i = 0, n = AtomicAccesses.size(); i < n; ++i) {
342 Res |= instrumentAtomic(AtomicAccesses[i]);
345 // Instrument function entry/exit points if there were instrumented accesses.
346 if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
347 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
348 Value *ReturnAddress = IRB.CreateCall(
349 Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
351 IRB.CreateCall(TsanFuncEntry, ReturnAddress);
352 for (size_t i = 0, n = RetVec.size(); i < n; ++i) {
353 IRBuilder<> IRBRet(RetVec[i]);
354 IRBRet.CreateCall(TsanFuncExit);
361 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I) {
363 bool IsWrite = isa<StoreInst>(*I);
364 Value *Addr = IsWrite
365 ? cast<StoreInst>(I)->getPointerOperand()
366 : cast<LoadInst>(I)->getPointerOperand();
367 int Idx = getMemoryAccessFuncIndex(Addr);
370 if (IsWrite && isVtableAccess(I)) {
371 DEBUG(dbgs() << " VPTR : " << *I << "\n");
372 Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
373 // StoredValue does not necessary have a pointer type.
374 if (isa<IntegerType>(StoredValue->getType()))
375 StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
376 // Call TsanVptrUpdate.
377 IRB.CreateCall2(TsanVptrUpdate,
378 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
379 IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy()));
380 NumInstrumentedVtableWrites++;
383 Value *OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
384 IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
385 if (IsWrite) NumInstrumentedWrites++;
386 else NumInstrumentedReads++;
390 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
393 case NotAtomic: assert(false);
394 case Unordered: // Fall-through.
395 case Monotonic: v = 0; break;
396 // case Consume: v = 1; break; // Not specified yet.
397 case Acquire: v = 2; break;
398 case Release: v = 3; break;
399 case AcquireRelease: v = 4; break;
400 case SequentiallyConsistent: v = 5; break;
402 return IRB->getInt32(v);
405 static ConstantInt *createFailOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
408 case NotAtomic: assert(false);
409 case Unordered: // Fall-through.
410 case Monotonic: v = 0; break;
411 // case Consume: v = 1; break; // Not specified yet.
412 case Acquire: v = 2; break;
413 case Release: v = 0; break;
414 case AcquireRelease: v = 2; break;
415 case SequentiallyConsistent: v = 5; break;
417 return IRB->getInt32(v);
420 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
421 // standards. For background see C++11 standard. A slightly older, publically
422 // available draft of the standard (not entirely up-to-date, but close enough
423 // for casual browsing) is available here:
424 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
425 // The following page contains more background information:
426 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
428 bool ThreadSanitizer::instrumentAtomic(Instruction *I) {
430 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
431 Value *Addr = LI->getPointerOperand();
432 int Idx = getMemoryAccessFuncIndex(Addr);
435 const size_t ByteSize = 1 << Idx;
436 const size_t BitSize = ByteSize * 8;
437 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
438 Type *PtrTy = Ty->getPointerTo();
439 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
440 createOrdering(&IRB, LI->getOrdering())};
441 CallInst *C = CallInst::Create(TsanAtomicLoad[Idx],
442 ArrayRef<Value*>(Args));
443 ReplaceInstWithInst(I, C);
445 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
446 Value *Addr = SI->getPointerOperand();
447 int Idx = getMemoryAccessFuncIndex(Addr);
450 const size_t ByteSize = 1 << Idx;
451 const size_t BitSize = ByteSize * 8;
452 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
453 Type *PtrTy = Ty->getPointerTo();
454 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
455 IRB.CreateIntCast(SI->getValueOperand(), Ty, false),
456 createOrdering(&IRB, SI->getOrdering())};
457 CallInst *C = CallInst::Create(TsanAtomicStore[Idx],
458 ArrayRef<Value*>(Args));
459 ReplaceInstWithInst(I, C);
460 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
461 Value *Addr = RMWI->getPointerOperand();
462 int Idx = getMemoryAccessFuncIndex(Addr);
465 Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
468 const size_t ByteSize = 1 << Idx;
469 const size_t BitSize = ByteSize * 8;
470 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
471 Type *PtrTy = Ty->getPointerTo();
472 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
473 IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
474 createOrdering(&IRB, RMWI->getOrdering())};
475 CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args));
476 ReplaceInstWithInst(I, C);
477 } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
478 Value *Addr = CASI->getPointerOperand();
479 int Idx = getMemoryAccessFuncIndex(Addr);
482 const size_t ByteSize = 1 << Idx;
483 const size_t BitSize = ByteSize * 8;
484 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
485 Type *PtrTy = Ty->getPointerTo();
486 Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
487 IRB.CreateIntCast(CASI->getCompareOperand(), Ty, false),
488 IRB.CreateIntCast(CASI->getNewValOperand(), Ty, false),
489 createOrdering(&IRB, CASI->getOrdering()),
490 createFailOrdering(&IRB, CASI->getOrdering())};
491 CallInst *C = CallInst::Create(TsanAtomicCAS[Idx], ArrayRef<Value*>(Args));
492 ReplaceInstWithInst(I, C);
493 } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
494 Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
495 Function *F = FI->getSynchScope() == SingleThread ?
496 TsanAtomicSignalFence : TsanAtomicThreadFence;
497 CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args));
498 ReplaceInstWithInst(I, C);
503 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr) {
504 Type *OrigPtrTy = Addr->getType();
505 Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
506 assert(OrigTy->isSized());
507 uint32_t TypeSize = TD->getTypeStoreSizeInBits(OrigTy);
508 if (TypeSize != 8 && TypeSize != 16 &&
509 TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
510 NumAccessesWithBadSize++;
511 // Ignore all unusual sizes.
514 size_t Idx = CountTrailingZeros_32(TypeSize / 8);
515 assert(Idx < kNumberOfAccessSizes);