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 "FunctionBlackList.h"
25#include "llvm/ADT/SmallSet.h"
26#include "llvm/ADT/SmallString.h"
27#include "llvm/ADT/SmallVector.h"
28#include "llvm/ADT/StringExtras.h"
29#include "llvm/Intrinsics.h"
30#include "llvm/Function.h"
31#include "llvm/LLVMContext.h"
32#include "llvm/Metadata.h"
33#include "llvm/Module.h"
34#include "llvm/Support/CommandLine.h"
35#include "llvm/Support/Debug.h"
36#include "llvm/Support/IRBuilder.h"
37#include "llvm/Support/MathExtras.h"
38#include "llvm/Support/raw_ostream.h"
39#include "llvm/Target/TargetData.h"
40#include "llvm/Transforms/Instrumentation.h"
41#include "llvm/Transforms/Utils/ModuleUtils.h"
42#include "llvm/Type.h"
43
44using namespace llvm;
45
46static cl::opt<std::string>  ClBlackListFile("tsan-blacklist",
47       cl::desc("Blacklist file"), cl::Hidden);
48
49static cl::opt<bool> ClPrintStats("tsan-print-stats",
50       cl::desc("Print ThreadSanitizer instrumentation stats"), cl::Hidden);
51
52namespace {
53
54// Stats counters for ThreadSanitizer instrumentation.
55struct ThreadSanitizerStats {
56  size_t NumInstrumentedReads;
57  size_t NumInstrumentedWrites;
58  size_t NumOmittedReadsBeforeWrite;
59  size_t NumAccessesWithBadSize;
60  size_t NumInstrumentedVtableWrites;
61  size_t NumOmittedReadsFromConstantGlobals;
62  size_t NumOmittedReadsFromVtable;
63};
64
65/// ThreadSanitizer: instrument the code in module to find races.
66struct ThreadSanitizer : public FunctionPass {
67  ThreadSanitizer();
68  bool runOnFunction(Function &F);
69  bool doInitialization(Module &M);
70  bool doFinalization(Module &M);
71  bool instrumentLoadOrStore(Instruction *I);
72  static char ID;  // Pass identification, replacement for typeid.
73
74 private:
75  void choseInstructionsToInstrument(SmallVectorImpl<Instruction*> &Local,
76                                     SmallVectorImpl<Instruction*> &All);
77  bool addrPointsToConstantData(Value *Addr);
78
79  TargetData *TD;
80  OwningPtr<FunctionBlackList> BL;
81  // Callbacks to run-time library are computed in doInitialization.
82  Value *TsanFuncEntry;
83  Value *TsanFuncExit;
84  // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
85  static const size_t kNumberOfAccessSizes = 5;
86  Value *TsanRead[kNumberOfAccessSizes];
87  Value *TsanWrite[kNumberOfAccessSizes];
88  Value *TsanVptrUpdate;
89
90  // Stats are modified w/o synchronization.
91  ThreadSanitizerStats stats;
92};
93}  // namespace
94
95char ThreadSanitizer::ID = 0;
96INITIALIZE_PASS(ThreadSanitizer, "tsan",
97    "ThreadSanitizer: detects data races.",
98    false, false)
99
100ThreadSanitizer::ThreadSanitizer()
101  : FunctionPass(ID),
102  TD(NULL) {
103}
104
105FunctionPass *llvm::createThreadSanitizerPass() {
106  return new ThreadSanitizer();
107}
108
109bool ThreadSanitizer::doInitialization(Module &M) {
110  TD = getAnalysisIfAvailable<TargetData>();
111  if (!TD)
112    return false;
113  BL.reset(new FunctionBlackList(ClBlackListFile));
114  memset(&stats, 0, sizeof(stats));
115
116  // Always insert a call to __tsan_init into the module's CTORs.
117  IRBuilder<> IRB(M.getContext());
118  Value *TsanInit = M.getOrInsertFunction("__tsan_init",
119                                          IRB.getVoidTy(), NULL);
120  appendToGlobalCtors(M, cast<Function>(TsanInit), 0);
121
122  // Initialize the callbacks.
123  TsanFuncEntry = M.getOrInsertFunction("__tsan_func_entry", IRB.getVoidTy(),
124                                        IRB.getInt8PtrTy(), NULL);
125  TsanFuncExit = M.getOrInsertFunction("__tsan_func_exit", IRB.getVoidTy(),
126                                       NULL);
127  for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
128    SmallString<32> ReadName("__tsan_read");
129    ReadName += itostr(1 << i);
130    TsanRead[i] = M.getOrInsertFunction(ReadName, IRB.getVoidTy(),
131                                        IRB.getInt8PtrTy(), NULL);
132    SmallString<32> WriteName("__tsan_write");
133    WriteName += itostr(1 << i);
134    TsanWrite[i] = M.getOrInsertFunction(WriteName, IRB.getVoidTy(),
135                                         IRB.getInt8PtrTy(), NULL);
136  }
137  TsanVptrUpdate = M.getOrInsertFunction("__tsan_vptr_update", IRB.getVoidTy(),
138                                         IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
139                                         NULL);
140  return true;
141}
142
143bool ThreadSanitizer::doFinalization(Module &M) {
144  if (ClPrintStats) {
145    errs() << "ThreadSanitizerStats " << M.getModuleIdentifier()
146           << ": wr " << stats.NumInstrumentedWrites
147           << "; rd " << stats.NumInstrumentedReads
148           << "; vt " << stats.NumInstrumentedVtableWrites
149           << "; bs " << stats.NumAccessesWithBadSize
150           << "; rbw " << stats.NumOmittedReadsBeforeWrite
151           << "; rcg " << stats.NumOmittedReadsFromConstantGlobals
152           << "; rvt " << stats.NumOmittedReadsFromVtable
153           << "\n";
154  }
155  return true;
156}
157
158static bool isVtableAccess(Instruction *I) {
159  if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa)) {
160    if (Tag->getNumOperands() < 1) return false;
161    if (MDString *Tag1 = dyn_cast<MDString>(Tag->getOperand(0))) {
162      if (Tag1->getString() == "vtable pointer") return true;
163    }
164  }
165  return false;
166}
167
168bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
169  // If this is a GEP, just analyze its pointer operand.
170  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
171    Addr = GEP->getPointerOperand();
172
173  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
174    if (GV->isConstant()) {
175      // Reads from constant globals can not race with any writes.
176      stats.NumOmittedReadsFromConstantGlobals++;
177      return true;
178    }
179  } else if(LoadInst *L = dyn_cast<LoadInst>(Addr)) {
180    if (isVtableAccess(L)) {
181      // Reads from a vtable pointer can not race with any writes.
182      stats.NumOmittedReadsFromVtable++;
183      return true;
184    }
185  }
186  return false;
187}
188
189// Instrumenting some of the accesses may be proven redundant.
190// Currently handled:
191//  - read-before-write (within same BB, no calls between)
192//
193// We do not handle some of the patterns that should not survive
194// after the classic compiler optimizations.
195// E.g. two reads from the same temp should be eliminated by CSE,
196// two writes should be eliminated by DSE, etc.
197//
198// 'Local' is a vector of insns within the same BB (no calls between).
199// 'All' is a vector of insns that will be instrumented.
200void ThreadSanitizer::choseInstructionsToInstrument(
201    SmallVectorImpl<Instruction*> &Local,
202    SmallVectorImpl<Instruction*> &All) {
203  SmallSet<Value*, 8> WriteTargets;
204  // Iterate from the end.
205  for (SmallVectorImpl<Instruction*>::reverse_iterator It = Local.rbegin(),
206       E = Local.rend(); It != E; ++It) {
207    Instruction *I = *It;
208    if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
209      WriteTargets.insert(Store->getPointerOperand());
210    } else {
211      LoadInst *Load = cast<LoadInst>(I);
212      Value *Addr = Load->getPointerOperand();
213      if (WriteTargets.count(Addr)) {
214        // We will write to this temp, so no reason to analyze the read.
215        stats.NumOmittedReadsBeforeWrite++;
216        continue;
217      }
218      if (addrPointsToConstantData(Addr)) {
219        // Addr points to some constant data -- it can not race with any writes.
220        continue;
221      }
222    }
223    All.push_back(I);
224  }
225  Local.clear();
226}
227
228bool ThreadSanitizer::runOnFunction(Function &F) {
229  if (!TD) return false;
230  if (BL->isIn(F)) return false;
231  SmallVector<Instruction*, 8> RetVec;
232  SmallVector<Instruction*, 8> AllLoadsAndStores;
233  SmallVector<Instruction*, 8> LocalLoadsAndStores;
234  bool Res = false;
235  bool HasCalls = false;
236
237  // Traverse all instructions, collect loads/stores/returns, check for calls.
238  for (Function::iterator FI = F.begin(), FE = F.end();
239       FI != FE; ++FI) {
240    BasicBlock &BB = *FI;
241    for (BasicBlock::iterator BI = BB.begin(), BE = BB.end();
242         BI != BE; ++BI) {
243      if (isa<LoadInst>(BI) || isa<StoreInst>(BI))
244        LocalLoadsAndStores.push_back(BI);
245      else if (isa<ReturnInst>(BI))
246        RetVec.push_back(BI);
247      else if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
248        HasCalls = true;
249        choseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
250      }
251    }
252    choseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
253  }
254
255  // We have collected all loads and stores.
256  // FIXME: many of these accesses do not need to be checked for races
257  // (e.g. variables that do not escape, etc).
258
259  // Instrument memory accesses.
260  for (size_t i = 0, n = AllLoadsAndStores.size(); i < n; ++i) {
261    Res |= instrumentLoadOrStore(AllLoadsAndStores[i]);
262  }
263
264  // Instrument function entry/exit points if there were instrumented accesses.
265  if (Res || HasCalls) {
266    IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
267    Value *ReturnAddress = IRB.CreateCall(
268        Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
269        IRB.getInt32(0));
270    IRB.CreateCall(TsanFuncEntry, ReturnAddress);
271    for (size_t i = 0, n = RetVec.size(); i < n; ++i) {
272      IRBuilder<> IRBRet(RetVec[i]);
273      IRBRet.CreateCall(TsanFuncExit);
274    }
275    Res = true;
276  }
277  return Res;
278}
279
280bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I) {
281  IRBuilder<> IRB(I);
282  bool IsWrite = isa<StoreInst>(*I);
283  Value *Addr = IsWrite
284      ? cast<StoreInst>(I)->getPointerOperand()
285      : cast<LoadInst>(I)->getPointerOperand();
286  Type *OrigPtrTy = Addr->getType();
287  Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
288  assert(OrigTy->isSized());
289  uint32_t TypeSize = TD->getTypeStoreSizeInBits(OrigTy);
290  if (TypeSize != 8  && TypeSize != 16 &&
291      TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
292    stats.NumAccessesWithBadSize++;
293    // Ignore all unusual sizes.
294    return false;
295  }
296  if (IsWrite && isVtableAccess(I)) {
297    Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
298    IRB.CreateCall2(TsanVptrUpdate,
299                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
300                    IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy()));
301    stats.NumInstrumentedVtableWrites++;
302    return true;
303  }
304  size_t Idx = CountTrailingZeros_32(TypeSize / 8);
305  assert(Idx < kNumberOfAccessSizes);
306  Value *OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
307  IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
308  if (IsWrite) stats.NumInstrumentedWrites++;
309  else         stats.NumInstrumentedReads++;
310  return true;
311}
312