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/Analysis/CaptureTracking.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/Transforms/Instrumentation.h"
25 #include "llvm/ADT/SmallSet.h"
26 #include "llvm/ADT/SmallString.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/ADT/StringExtras.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), DL(nullptr) {}
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);
88   bool instrumentAtomic(Instruction *I);
89   bool instrumentMemIntrinsic(Instruction *I);
90   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction*> &Local,
91                                       SmallVectorImpl<Instruction*> &All);
92   bool addrPointsToConstantData(Value *Addr);
93   int getMemoryAccessFuncIndex(Value *Addr);
94 
95   const DataLayout *DL;
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 static Function *checkInterfaceFunction(Constant *FuncOrBitcast) {
133   if (Function *F = dyn_cast<Function>(FuncOrBitcast))
134      return F;
135   FuncOrBitcast->dump();
136   report_fatal_error("ThreadSanitizer interface function redefined");
137 }
138 
139 void ThreadSanitizer::initializeCallbacks(Module &M) {
140   IRBuilder<> IRB(M.getContext());
141   // Initialize the callbacks.
142   TsanFuncEntry = checkInterfaceFunction(M.getOrInsertFunction(
143       "__tsan_func_entry", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
144   TsanFuncExit = checkInterfaceFunction(M.getOrInsertFunction(
145       "__tsan_func_exit", IRB.getVoidTy(), nullptr));
146   OrdTy = IRB.getInt32Ty();
147   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
148     const size_t ByteSize = 1 << i;
149     const size_t BitSize = ByteSize * 8;
150     SmallString<32> ReadName("__tsan_read" + itostr(ByteSize));
151     TsanRead[i] = checkInterfaceFunction(M.getOrInsertFunction(
152         ReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
153 
154     SmallString<32> WriteName("__tsan_write" + itostr(ByteSize));
155     TsanWrite[i] = checkInterfaceFunction(M.getOrInsertFunction(
156         WriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
157 
158     SmallString<64> UnalignedReadName("__tsan_unaligned_read" +
159         itostr(ByteSize));
160     TsanUnalignedRead[i] = checkInterfaceFunction(M.getOrInsertFunction(
161         UnalignedReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
162 
163     SmallString<64> UnalignedWriteName("__tsan_unaligned_write" +
164         itostr(ByteSize));
165     TsanUnalignedWrite[i] = checkInterfaceFunction(M.getOrInsertFunction(
166         UnalignedWriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
167 
168     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
169     Type *PtrTy = Ty->getPointerTo();
170     SmallString<32> AtomicLoadName("__tsan_atomic" + itostr(BitSize) +
171                                    "_load");
172     TsanAtomicLoad[i] = checkInterfaceFunction(M.getOrInsertFunction(
173         AtomicLoadName, Ty, PtrTy, OrdTy, nullptr));
174 
175     SmallString<32> AtomicStoreName("__tsan_atomic" + itostr(BitSize) +
176                                     "_store");
177     TsanAtomicStore[i] = checkInterfaceFunction(M.getOrInsertFunction(
178         AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy,
179         nullptr));
180 
181     for (int op = AtomicRMWInst::FIRST_BINOP;
182         op <= AtomicRMWInst::LAST_BINOP; ++op) {
183       TsanAtomicRMW[op][i] = nullptr;
184       const char *NamePart = nullptr;
185       if (op == AtomicRMWInst::Xchg)
186         NamePart = "_exchange";
187       else if (op == AtomicRMWInst::Add)
188         NamePart = "_fetch_add";
189       else if (op == AtomicRMWInst::Sub)
190         NamePart = "_fetch_sub";
191       else if (op == AtomicRMWInst::And)
192         NamePart = "_fetch_and";
193       else if (op == AtomicRMWInst::Or)
194         NamePart = "_fetch_or";
195       else if (op == AtomicRMWInst::Xor)
196         NamePart = "_fetch_xor";
197       else if (op == AtomicRMWInst::Nand)
198         NamePart = "_fetch_nand";
199       else
200         continue;
201       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
202       TsanAtomicRMW[op][i] = checkInterfaceFunction(M.getOrInsertFunction(
203           RMWName, Ty, PtrTy, Ty, OrdTy, nullptr));
204     }
205 
206     SmallString<32> AtomicCASName("__tsan_atomic" + itostr(BitSize) +
207                                   "_compare_exchange_val");
208     TsanAtomicCAS[i] = checkInterfaceFunction(M.getOrInsertFunction(
209         AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, nullptr));
210   }
211   TsanVptrUpdate = checkInterfaceFunction(M.getOrInsertFunction(
212       "__tsan_vptr_update", IRB.getVoidTy(), IRB.getInt8PtrTy(),
213       IRB.getInt8PtrTy(), nullptr));
214   TsanVptrLoad = checkInterfaceFunction(M.getOrInsertFunction(
215       "__tsan_vptr_read", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
216   TsanAtomicThreadFence = checkInterfaceFunction(M.getOrInsertFunction(
217       "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, nullptr));
218   TsanAtomicSignalFence = checkInterfaceFunction(M.getOrInsertFunction(
219       "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, nullptr));
220 
221   MemmoveFn = checkInterfaceFunction(M.getOrInsertFunction(
222     "memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
223     IRB.getInt8PtrTy(), IntptrTy, nullptr));
224   MemcpyFn = checkInterfaceFunction(M.getOrInsertFunction(
225     "memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
226     IntptrTy, nullptr));
227   MemsetFn = checkInterfaceFunction(M.getOrInsertFunction(
228     "memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
229     IntptrTy, nullptr));
230 }
231 
232 bool ThreadSanitizer::doInitialization(Module &M) {
233   DL = &M.getDataLayout();
234 
235   // Always insert a call to __tsan_init into the module's CTORs.
236   IRBuilder<> IRB(M.getContext());
237   IntptrTy = IRB.getIntPtrTy(DL);
238   Value *TsanInit = M.getOrInsertFunction("__tsan_init",
239                                           IRB.getVoidTy(), nullptr);
240   appendToGlobalCtors(M, cast<Function>(TsanInit), 0);
241 
242   return true;
243 }
244 
245 static bool isVtableAccess(Instruction *I) {
246   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
247     return Tag->isTBAAVtableAccess();
248   return false;
249 }
250 
251 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
252   // If this is a GEP, just analyze its pointer operand.
253   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
254     Addr = GEP->getPointerOperand();
255 
256   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
257     if (GV->isConstant()) {
258       // Reads from constant globals can not race with any writes.
259       NumOmittedReadsFromConstantGlobals++;
260       return true;
261     }
262   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
263     if (isVtableAccess(L)) {
264       // Reads from a vtable pointer can not race with any writes.
265       NumOmittedReadsFromVtable++;
266       return true;
267     }
268   }
269   return false;
270 }
271 
272 // Instrumenting some of the accesses may be proven redundant.
273 // Currently handled:
274 //  - read-before-write (within same BB, no calls between)
275 //  - not captured variables
276 //
277 // We do not handle some of the patterns that should not survive
278 // after the classic compiler optimizations.
279 // E.g. two reads from the same temp should be eliminated by CSE,
280 // two writes should be eliminated by DSE, etc.
281 //
282 // 'Local' is a vector of insns within the same BB (no calls between).
283 // 'All' is a vector of insns that will be instrumented.
284 void ThreadSanitizer::chooseInstructionsToInstrument(
285     SmallVectorImpl<Instruction*> &Local,
286     SmallVectorImpl<Instruction*> &All) {
287   SmallSet<Value*, 8> WriteTargets;
288   // Iterate from the end.
289   for (SmallVectorImpl<Instruction*>::reverse_iterator It = Local.rbegin(),
290        E = Local.rend(); It != E; ++It) {
291     Instruction *I = *It;
292     if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
293       WriteTargets.insert(Store->getPointerOperand());
294     } else {
295       LoadInst *Load = cast<LoadInst>(I);
296       Value *Addr = Load->getPointerOperand();
297       if (WriteTargets.count(Addr)) {
298         // We will write to this temp, so no reason to analyze the read.
299         NumOmittedReadsBeforeWrite++;
300         continue;
301       }
302       if (addrPointsToConstantData(Addr)) {
303         // Addr points to some constant data -- it can not race with any writes.
304         continue;
305       }
306     }
307     Value *Addr = isa<StoreInst>(*I)
308         ? cast<StoreInst>(I)->getPointerOperand()
309         : cast<LoadInst>(I)->getPointerOperand();
310     if (isa<AllocaInst>(GetUnderlyingObject(Addr, nullptr)) &&
311         !PointerMayBeCaptured(Addr, true, true)) {
312       // The variable is addressable but not captured, so it cannot be
313       // referenced from a different thread and participate in a data race
314       // (see llvm/Analysis/CaptureTracking.h for details).
315       NumOmittedNonCaptured++;
316       continue;
317     }
318     All.push_back(I);
319   }
320   Local.clear();
321 }
322 
323 static bool isAtomic(Instruction *I) {
324   if (LoadInst *LI = dyn_cast<LoadInst>(I))
325     return LI->isAtomic() && LI->getSynchScope() == CrossThread;
326   if (StoreInst *SI = dyn_cast<StoreInst>(I))
327     return SI->isAtomic() && SI->getSynchScope() == CrossThread;
328   if (isa<AtomicRMWInst>(I))
329     return true;
330   if (isa<AtomicCmpXchgInst>(I))
331     return true;
332   if (isa<FenceInst>(I))
333     return true;
334   return false;
335 }
336 
337 bool ThreadSanitizer::runOnFunction(Function &F) {
338   if (!DL) return false;
339   initializeCallbacks(*F.getParent());
340   SmallVector<Instruction*, 8> RetVec;
341   SmallVector<Instruction*, 8> AllLoadsAndStores;
342   SmallVector<Instruction*, 8> LocalLoadsAndStores;
343   SmallVector<Instruction*, 8> AtomicAccesses;
344   SmallVector<Instruction*, 8> MemIntrinCalls;
345   bool Res = false;
346   bool HasCalls = false;
347   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
348 
349   // Traverse all instructions, collect loads/stores/returns, check for calls.
350   for (auto &BB : F) {
351     for (auto &Inst : BB) {
352       if (isAtomic(&Inst))
353         AtomicAccesses.push_back(&Inst);
354       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
355         LocalLoadsAndStores.push_back(&Inst);
356       else if (isa<ReturnInst>(Inst))
357         RetVec.push_back(&Inst);
358       else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
359         if (isa<MemIntrinsic>(Inst))
360           MemIntrinCalls.push_back(&Inst);
361         HasCalls = true;
362         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
363       }
364     }
365     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
366   }
367 
368   // We have collected all loads and stores.
369   // FIXME: many of these accesses do not need to be checked for races
370   // (e.g. variables that do not escape, etc).
371 
372   // Instrument memory accesses only if we want to report bugs in the function.
373   if (ClInstrumentMemoryAccesses && SanitizeFunction)
374     for (auto Inst : AllLoadsAndStores) {
375       Res |= instrumentLoadOrStore(Inst);
376     }
377 
378   // Instrument atomic memory accesses in any case (they can be used to
379   // implement synchronization).
380   if (ClInstrumentAtomics)
381     for (auto Inst : AtomicAccesses) {
382       Res |= instrumentAtomic(Inst);
383     }
384 
385   if (ClInstrumentMemIntrinsics && SanitizeFunction)
386     for (auto Inst : MemIntrinCalls) {
387       Res |= instrumentMemIntrinsic(Inst);
388     }
389 
390   // Instrument function entry/exit points if there were instrumented accesses.
391   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
392     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
393     Value *ReturnAddress = IRB.CreateCall(
394         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
395         IRB.getInt32(0));
396     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
397     for (auto RetInst : RetVec) {
398       IRBuilder<> IRBRet(RetInst);
399       IRBRet.CreateCall(TsanFuncExit);
400     }
401     Res = true;
402   }
403   return Res;
404 }
405 
406 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I) {
407   IRBuilder<> IRB(I);
408   bool IsWrite = isa<StoreInst>(*I);
409   Value *Addr = IsWrite
410       ? cast<StoreInst>(I)->getPointerOperand()
411       : cast<LoadInst>(I)->getPointerOperand();
412   int Idx = getMemoryAccessFuncIndex(Addr);
413   if (Idx < 0)
414     return false;
415   if (IsWrite && isVtableAccess(I)) {
416     DEBUG(dbgs() << "  VPTR : " << *I << "\n");
417     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
418     // StoredValue may be a vector type if we are storing several vptrs at once.
419     // In this case, just take the first element of the vector since this is
420     // enough to find vptr races.
421     if (isa<VectorType>(StoredValue->getType()))
422       StoredValue = IRB.CreateExtractElement(
423           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
424     if (StoredValue->getType()->isIntegerTy())
425       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
426     // Call TsanVptrUpdate.
427     IRB.CreateCall2(TsanVptrUpdate,
428                     IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
429                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy()));
430     NumInstrumentedVtableWrites++;
431     return true;
432   }
433   if (!IsWrite && isVtableAccess(I)) {
434     IRB.CreateCall(TsanVptrLoad,
435                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
436     NumInstrumentedVtableReads++;
437     return true;
438   }
439   const unsigned Alignment = IsWrite
440       ? cast<StoreInst>(I)->getAlignment()
441       : cast<LoadInst>(I)->getAlignment();
442   Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
443   const uint32_t TypeSize = DL->getTypeStoreSizeInBits(OrigTy);
444   Value *OnAccessFunc = nullptr;
445   if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0)
446     OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
447   else
448     OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
449   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
450   if (IsWrite) NumInstrumentedWrites++;
451   else         NumInstrumentedReads++;
452   return true;
453 }
454 
455 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
456   uint32_t v = 0;
457   switch (ord) {
458     case NotAtomic: llvm_unreachable("unexpected atomic ordering!");
459     case Unordered:              // Fall-through.
460     case Monotonic:              v = 0; break;
461     // case Consume:                v = 1; break;  // Not specified yet.
462     case Acquire:                v = 2; break;
463     case Release:                v = 3; break;
464     case AcquireRelease:         v = 4; break;
465     case SequentiallyConsistent: v = 5; break;
466   }
467   return IRB->getInt32(v);
468 }
469 
470 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
471 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
472 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
473 // instead we simply replace them with regular function calls, which are then
474 // intercepted by the run-time.
475 // Since tsan is running after everyone else, the calls should not be
476 // replaced back with intrinsics. If that becomes wrong at some point,
477 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
478 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
479   IRBuilder<> IRB(I);
480   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
481     IRB.CreateCall3(MemsetFn,
482       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
483       IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
484       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
485     I->eraseFromParent();
486   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
487     IRB.CreateCall3(isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
488       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
489       IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
490       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
491     I->eraseFromParent();
492   }
493   return false;
494 }
495 
496 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
497 // standards.  For background see C++11 standard.  A slightly older, publicly
498 // available draft of the standard (not entirely up-to-date, but close enough
499 // for casual browsing) is available here:
500 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
501 // The following page contains more background information:
502 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
503 
504 bool ThreadSanitizer::instrumentAtomic(Instruction *I) {
505   IRBuilder<> IRB(I);
506   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
507     Value *Addr = LI->getPointerOperand();
508     int Idx = getMemoryAccessFuncIndex(Addr);
509     if (Idx < 0)
510       return false;
511     const size_t ByteSize = 1 << Idx;
512     const size_t BitSize = ByteSize * 8;
513     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
514     Type *PtrTy = Ty->getPointerTo();
515     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
516                      createOrdering(&IRB, LI->getOrdering())};
517     CallInst *C = CallInst::Create(TsanAtomicLoad[Idx], Args);
518     ReplaceInstWithInst(I, C);
519 
520   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
521     Value *Addr = SI->getPointerOperand();
522     int Idx = getMemoryAccessFuncIndex(Addr);
523     if (Idx < 0)
524       return false;
525     const size_t ByteSize = 1 << Idx;
526     const size_t BitSize = ByteSize * 8;
527     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
528     Type *PtrTy = Ty->getPointerTo();
529     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
530                      IRB.CreateIntCast(SI->getValueOperand(), Ty, false),
531                      createOrdering(&IRB, SI->getOrdering())};
532     CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
533     ReplaceInstWithInst(I, C);
534   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
535     Value *Addr = RMWI->getPointerOperand();
536     int Idx = getMemoryAccessFuncIndex(Addr);
537     if (Idx < 0)
538       return false;
539     Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
540     if (!F)
541       return false;
542     const size_t ByteSize = 1 << Idx;
543     const size_t BitSize = ByteSize * 8;
544     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
545     Type *PtrTy = Ty->getPointerTo();
546     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
547                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
548                      createOrdering(&IRB, RMWI->getOrdering())};
549     CallInst *C = CallInst::Create(F, Args);
550     ReplaceInstWithInst(I, C);
551   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
552     Value *Addr = CASI->getPointerOperand();
553     int Idx = getMemoryAccessFuncIndex(Addr);
554     if (Idx < 0)
555       return false;
556     const size_t ByteSize = 1 << Idx;
557     const size_t BitSize = ByteSize * 8;
558     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
559     Type *PtrTy = Ty->getPointerTo();
560     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
561                      IRB.CreateIntCast(CASI->getCompareOperand(), Ty, false),
562                      IRB.CreateIntCast(CASI->getNewValOperand(), Ty, false),
563                      createOrdering(&IRB, CASI->getSuccessOrdering()),
564                      createOrdering(&IRB, CASI->getFailureOrdering())};
565     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
566     Value *Success = IRB.CreateICmpEQ(C, CASI->getCompareOperand());
567 
568     Value *Res = IRB.CreateInsertValue(UndefValue::get(CASI->getType()), C, 0);
569     Res = IRB.CreateInsertValue(Res, Success, 1);
570 
571     I->replaceAllUsesWith(Res);
572     I->eraseFromParent();
573   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
574     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
575     Function *F = FI->getSynchScope() == SingleThread ?
576         TsanAtomicSignalFence : TsanAtomicThreadFence;
577     CallInst *C = CallInst::Create(F, Args);
578     ReplaceInstWithInst(I, C);
579   }
580   return true;
581 }
582 
583 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr) {
584   Type *OrigPtrTy = Addr->getType();
585   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
586   assert(OrigTy->isSized());
587   uint32_t TypeSize = DL->getTypeStoreSizeInBits(OrigTy);
588   if (TypeSize != 8  && TypeSize != 16 &&
589       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
590     NumAccessesWithBadSize++;
591     // Ignore all unusual sizes.
592     return -1;
593   }
594   size_t Idx = countTrailingZeros(TypeSize / 8);
595   assert(Idx < kNumberOfAccessSizes);
596   return Idx;
597 }
598