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/Transforms/Instrumentation.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallString.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/StringExtras.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/IR/Metadata.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Type.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/Transforms/Utils/BasicBlockUtils.h"
42 #include "llvm/Transforms/Utils/ModuleUtils.h"
43 #include "llvm/Transforms/Utils/SpecialCaseList.h"
44 
45 using namespace llvm;
46 
47 #define DEBUG_TYPE "tsan"
48 
49 static cl::opt<bool>  ClInstrumentMemoryAccesses(
50     "tsan-instrument-memory-accesses", cl::init(true),
51     cl::desc("Instrument memory accesses"), cl::Hidden);
52 static cl::opt<bool>  ClInstrumentFuncEntryExit(
53     "tsan-instrument-func-entry-exit", cl::init(true),
54     cl::desc("Instrument function entry and exit"), cl::Hidden);
55 static cl::opt<bool>  ClInstrumentAtomics(
56     "tsan-instrument-atomics", cl::init(true),
57     cl::desc("Instrument atomics"), cl::Hidden);
58 static cl::opt<bool>  ClInstrumentMemIntrinsics(
59     "tsan-instrument-memintrinsics", cl::init(true),
60     cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
61 
62 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
63 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
64 STATISTIC(NumOmittedReadsBeforeWrite,
65           "Number of reads ignored due to following writes");
66 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
67 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
68 STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
69 STATISTIC(NumOmittedReadsFromConstantGlobals,
70           "Number of reads from constant globals");
71 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
72 
73 namespace {
74 
75 /// ThreadSanitizer: instrument the code in module to find races.
76 struct ThreadSanitizer : public FunctionPass {
77   ThreadSanitizer() : FunctionPass(ID), DL(nullptr) {}
78   const char *getPassName() const override;
79   bool runOnFunction(Function &F) override;
80   bool doInitialization(Module &M) override;
81   static char ID;  // Pass identification, replacement for typeid.
82 
83  private:
84   void initializeCallbacks(Module &M);
85   bool instrumentLoadOrStore(Instruction *I);
86   bool instrumentAtomic(Instruction *I);
87   bool instrumentMemIntrinsic(Instruction *I);
88   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction*> &Local,
89                                       SmallVectorImpl<Instruction*> &All);
90   bool addrPointsToConstantData(Value *Addr);
91   int getMemoryAccessFuncIndex(Value *Addr);
92 
93   const DataLayout *DL;
94   Type *IntptrTy;
95   IntegerType *OrdTy;
96   // Callbacks to run-time library are computed in doInitialization.
97   Function *TsanFuncEntry;
98   Function *TsanFuncExit;
99   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
100   static const size_t kNumberOfAccessSizes = 5;
101   Function *TsanRead[kNumberOfAccessSizes];
102   Function *TsanWrite[kNumberOfAccessSizes];
103   Function *TsanAtomicLoad[kNumberOfAccessSizes];
104   Function *TsanAtomicStore[kNumberOfAccessSizes];
105   Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes];
106   Function *TsanAtomicCAS[kNumberOfAccessSizes];
107   Function *TsanAtomicThreadFence;
108   Function *TsanAtomicSignalFence;
109   Function *TsanVptrUpdate;
110   Function *TsanVptrLoad;
111   Function *MemmoveFn, *MemcpyFn, *MemsetFn;
112 };
113 }  // namespace
114 
115 char ThreadSanitizer::ID = 0;
116 INITIALIZE_PASS(ThreadSanitizer, "tsan",
117     "ThreadSanitizer: detects data races.",
118     false, false)
119 
120 const char *ThreadSanitizer::getPassName() const {
121   return "ThreadSanitizer";
122 }
123 
124 FunctionPass *llvm::createThreadSanitizerPass() {
125   return new ThreadSanitizer();
126 }
127 
128 static Function *checkInterfaceFunction(Constant *FuncOrBitcast) {
129   if (Function *F = dyn_cast<Function>(FuncOrBitcast))
130      return F;
131   FuncOrBitcast->dump();
132   report_fatal_error("ThreadSanitizer interface function redefined");
133 }
134 
135 void ThreadSanitizer::initializeCallbacks(Module &M) {
136   IRBuilder<> IRB(M.getContext());
137   // Initialize the callbacks.
138   TsanFuncEntry = checkInterfaceFunction(M.getOrInsertFunction(
139       "__tsan_func_entry", IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
140   TsanFuncExit = checkInterfaceFunction(M.getOrInsertFunction(
141       "__tsan_func_exit", IRB.getVoidTy(), NULL));
142   OrdTy = IRB.getInt32Ty();
143   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
144     const size_t ByteSize = 1 << i;
145     const size_t BitSize = ByteSize * 8;
146     SmallString<32> ReadName("__tsan_read" + itostr(ByteSize));
147     TsanRead[i] = checkInterfaceFunction(M.getOrInsertFunction(
148         ReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
149 
150     SmallString<32> WriteName("__tsan_write" + itostr(ByteSize));
151     TsanWrite[i] = checkInterfaceFunction(M.getOrInsertFunction(
152         WriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
153 
154     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
155     Type *PtrTy = Ty->getPointerTo();
156     SmallString<32> AtomicLoadName("__tsan_atomic" + itostr(BitSize) +
157                                    "_load");
158     TsanAtomicLoad[i] = checkInterfaceFunction(M.getOrInsertFunction(
159         AtomicLoadName, Ty, PtrTy, OrdTy, NULL));
160 
161     SmallString<32> AtomicStoreName("__tsan_atomic" + itostr(BitSize) +
162                                     "_store");
163     TsanAtomicStore[i] = checkInterfaceFunction(M.getOrInsertFunction(
164         AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy,
165         NULL));
166 
167     for (int op = AtomicRMWInst::FIRST_BINOP;
168         op <= AtomicRMWInst::LAST_BINOP; ++op) {
169       TsanAtomicRMW[op][i] = nullptr;
170       const char *NamePart = nullptr;
171       if (op == AtomicRMWInst::Xchg)
172         NamePart = "_exchange";
173       else if (op == AtomicRMWInst::Add)
174         NamePart = "_fetch_add";
175       else if (op == AtomicRMWInst::Sub)
176         NamePart = "_fetch_sub";
177       else if (op == AtomicRMWInst::And)
178         NamePart = "_fetch_and";
179       else if (op == AtomicRMWInst::Or)
180         NamePart = "_fetch_or";
181       else if (op == AtomicRMWInst::Xor)
182         NamePart = "_fetch_xor";
183       else if (op == AtomicRMWInst::Nand)
184         NamePart = "_fetch_nand";
185       else
186         continue;
187       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
188       TsanAtomicRMW[op][i] = checkInterfaceFunction(M.getOrInsertFunction(
189           RMWName, Ty, PtrTy, Ty, OrdTy, NULL));
190     }
191 
192     SmallString<32> AtomicCASName("__tsan_atomic" + itostr(BitSize) +
193                                   "_compare_exchange_val");
194     TsanAtomicCAS[i] = checkInterfaceFunction(M.getOrInsertFunction(
195         AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, NULL));
196   }
197   TsanVptrUpdate = checkInterfaceFunction(M.getOrInsertFunction(
198       "__tsan_vptr_update", IRB.getVoidTy(), IRB.getInt8PtrTy(),
199       IRB.getInt8PtrTy(), NULL));
200   TsanVptrLoad = checkInterfaceFunction(M.getOrInsertFunction(
201       "__tsan_vptr_read", IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
202   TsanAtomicThreadFence = checkInterfaceFunction(M.getOrInsertFunction(
203       "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, NULL));
204   TsanAtomicSignalFence = checkInterfaceFunction(M.getOrInsertFunction(
205       "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, NULL));
206 
207   MemmoveFn = checkInterfaceFunction(M.getOrInsertFunction(
208     "memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
209     IRB.getInt8PtrTy(), IntptrTy, NULL));
210   MemcpyFn = checkInterfaceFunction(M.getOrInsertFunction(
211     "memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
212     IntptrTy, NULL));
213   MemsetFn = checkInterfaceFunction(M.getOrInsertFunction(
214     "memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
215     IntptrTy, NULL));
216 }
217 
218 bool ThreadSanitizer::doInitialization(Module &M) {
219   DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
220   if (!DLP)
221     report_fatal_error("data layout missing");
222   DL = &DLP->getDataLayout();
223 
224   // Always insert a call to __tsan_init into the module's CTORs.
225   IRBuilder<> IRB(M.getContext());
226   IntptrTy = IRB.getIntPtrTy(DL);
227   Value *TsanInit = M.getOrInsertFunction("__tsan_init",
228                                           IRB.getVoidTy(), NULL);
229   appendToGlobalCtors(M, cast<Function>(TsanInit), 0);
230 
231   return true;
232 }
233 
234 static bool isVtableAccess(Instruction *I) {
235   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
236     return Tag->isTBAAVtableAccess();
237   return false;
238 }
239 
240 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
241   // If this is a GEP, just analyze its pointer operand.
242   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
243     Addr = GEP->getPointerOperand();
244 
245   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
246     if (GV->isConstant()) {
247       // Reads from constant globals can not race with any writes.
248       NumOmittedReadsFromConstantGlobals++;
249       return true;
250     }
251   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
252     if (isVtableAccess(L)) {
253       // Reads from a vtable pointer can not race with any writes.
254       NumOmittedReadsFromVtable++;
255       return true;
256     }
257   }
258   return false;
259 }
260 
261 // Instrumenting some of the accesses may be proven redundant.
262 // Currently handled:
263 //  - read-before-write (within same BB, no calls between)
264 //
265 // We do not handle some of the patterns that should not survive
266 // after the classic compiler optimizations.
267 // E.g. two reads from the same temp should be eliminated by CSE,
268 // two writes should be eliminated by DSE, etc.
269 //
270 // 'Local' is a vector of insns within the same BB (no calls between).
271 // 'All' is a vector of insns that will be instrumented.
272 void ThreadSanitizer::chooseInstructionsToInstrument(
273     SmallVectorImpl<Instruction*> &Local,
274     SmallVectorImpl<Instruction*> &All) {
275   SmallSet<Value*, 8> WriteTargets;
276   // Iterate from the end.
277   for (SmallVectorImpl<Instruction*>::reverse_iterator It = Local.rbegin(),
278        E = Local.rend(); It != E; ++It) {
279     Instruction *I = *It;
280     if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
281       WriteTargets.insert(Store->getPointerOperand());
282     } else {
283       LoadInst *Load = cast<LoadInst>(I);
284       Value *Addr = Load->getPointerOperand();
285       if (WriteTargets.count(Addr)) {
286         // We will write to this temp, so no reason to analyze the read.
287         NumOmittedReadsBeforeWrite++;
288         continue;
289       }
290       if (addrPointsToConstantData(Addr)) {
291         // Addr points to some constant data -- it can not race with any writes.
292         continue;
293       }
294     }
295     All.push_back(I);
296   }
297   Local.clear();
298 }
299 
300 static bool isAtomic(Instruction *I) {
301   if (LoadInst *LI = dyn_cast<LoadInst>(I))
302     return LI->isAtomic() && LI->getSynchScope() == CrossThread;
303   if (StoreInst *SI = dyn_cast<StoreInst>(I))
304     return SI->isAtomic() && SI->getSynchScope() == CrossThread;
305   if (isa<AtomicRMWInst>(I))
306     return true;
307   if (isa<AtomicCmpXchgInst>(I))
308     return true;
309   if (isa<FenceInst>(I))
310     return true;
311   return false;
312 }
313 
314 bool ThreadSanitizer::runOnFunction(Function &F) {
315   if (!DL) return false;
316   initializeCallbacks(*F.getParent());
317   SmallVector<Instruction*, 8> RetVec;
318   SmallVector<Instruction*, 8> AllLoadsAndStores;
319   SmallVector<Instruction*, 8> LocalLoadsAndStores;
320   SmallVector<Instruction*, 8> AtomicAccesses;
321   SmallVector<Instruction*, 8> MemIntrinCalls;
322   bool Res = false;
323   bool HasCalls = false;
324   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
325 
326   // Traverse all instructions, collect loads/stores/returns, check for calls.
327   for (auto &BB : F) {
328     for (auto &Inst : BB) {
329       if (isAtomic(&Inst))
330         AtomicAccesses.push_back(&Inst);
331       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
332         LocalLoadsAndStores.push_back(&Inst);
333       else if (isa<ReturnInst>(Inst))
334         RetVec.push_back(&Inst);
335       else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
336         if (isa<MemIntrinsic>(Inst))
337           MemIntrinCalls.push_back(&Inst);
338         HasCalls = true;
339         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
340       }
341     }
342     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
343   }
344 
345   // We have collected all loads and stores.
346   // FIXME: many of these accesses do not need to be checked for races
347   // (e.g. variables that do not escape, etc).
348 
349   // Instrument memory accesses only if we want to report bugs in the function.
350   if (ClInstrumentMemoryAccesses && SanitizeFunction)
351     for (auto Inst : AllLoadsAndStores) {
352       Res |= instrumentLoadOrStore(Inst);
353     }
354 
355   // Instrument atomic memory accesses in any case (they can be used to
356   // implement synchronization).
357   if (ClInstrumentAtomics)
358     for (auto Inst : AtomicAccesses) {
359       Res |= instrumentAtomic(Inst);
360     }
361 
362   if (ClInstrumentMemIntrinsics && SanitizeFunction)
363     for (auto Inst : MemIntrinCalls) {
364       Res |= instrumentMemIntrinsic(Inst);
365     }
366 
367   // Instrument function entry/exit points if there were instrumented accesses.
368   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
369     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
370     Value *ReturnAddress = IRB.CreateCall(
371         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
372         IRB.getInt32(0));
373     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
374     for (auto RetInst : RetVec) {
375       IRBuilder<> IRBRet(RetInst);
376       IRBRet.CreateCall(TsanFuncExit);
377     }
378     Res = true;
379   }
380   return Res;
381 }
382 
383 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I) {
384   IRBuilder<> IRB(I);
385   bool IsWrite = isa<StoreInst>(*I);
386   Value *Addr = IsWrite
387       ? cast<StoreInst>(I)->getPointerOperand()
388       : cast<LoadInst>(I)->getPointerOperand();
389   int Idx = getMemoryAccessFuncIndex(Addr);
390   if (Idx < 0)
391     return false;
392   if (IsWrite && isVtableAccess(I)) {
393     DEBUG(dbgs() << "  VPTR : " << *I << "\n");
394     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
395     // StoredValue may be a vector type if we are storing several vptrs at once.
396     // In this case, just take the first element of the vector since this is
397     // enough to find vptr races.
398     if (isa<VectorType>(StoredValue->getType()))
399       StoredValue = IRB.CreateExtractElement(
400           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
401     if (StoredValue->getType()->isIntegerTy())
402       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
403     // Call TsanVptrUpdate.
404     IRB.CreateCall2(TsanVptrUpdate,
405                     IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
406                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy()));
407     NumInstrumentedVtableWrites++;
408     return true;
409   }
410   if (!IsWrite && isVtableAccess(I)) {
411     IRB.CreateCall(TsanVptrLoad,
412                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
413     NumInstrumentedVtableReads++;
414     return true;
415   }
416   Value *OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
417   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
418   if (IsWrite) NumInstrumentedWrites++;
419   else         NumInstrumentedReads++;
420   return true;
421 }
422 
423 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
424   uint32_t v = 0;
425   switch (ord) {
426     case NotAtomic:              assert(false);
427     case Unordered:              // Fall-through.
428     case Monotonic:              v = 0; break;
429     // case Consume:                v = 1; break;  // Not specified yet.
430     case Acquire:                v = 2; break;
431     case Release:                v = 3; break;
432     case AcquireRelease:         v = 4; break;
433     case SequentiallyConsistent: v = 5; break;
434   }
435   return IRB->getInt32(v);
436 }
437 
438 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
439 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
440 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
441 // instead we simply replace them with regular function calls, which are then
442 // intercepted by the run-time.
443 // Since tsan is running after everyone else, the calls should not be
444 // replaced back with intrinsics. If that becomes wrong at some point,
445 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
446 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
447   IRBuilder<> IRB(I);
448   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
449     IRB.CreateCall3(MemsetFn,
450       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
451       IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
452       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
453     I->eraseFromParent();
454   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
455     IRB.CreateCall3(isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
456       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
457       IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
458       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
459     I->eraseFromParent();
460   }
461   return false;
462 }
463 
464 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
465 // standards.  For background see C++11 standard.  A slightly older, publicly
466 // available draft of the standard (not entirely up-to-date, but close enough
467 // for casual browsing) is available here:
468 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
469 // The following page contains more background information:
470 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
471 
472 bool ThreadSanitizer::instrumentAtomic(Instruction *I) {
473   IRBuilder<> IRB(I);
474   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
475     Value *Addr = LI->getPointerOperand();
476     int Idx = getMemoryAccessFuncIndex(Addr);
477     if (Idx < 0)
478       return false;
479     const size_t ByteSize = 1 << Idx;
480     const size_t BitSize = ByteSize * 8;
481     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
482     Type *PtrTy = Ty->getPointerTo();
483     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
484                      createOrdering(&IRB, LI->getOrdering())};
485     CallInst *C = CallInst::Create(TsanAtomicLoad[Idx],
486                                    ArrayRef<Value*>(Args));
487     ReplaceInstWithInst(I, C);
488 
489   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
490     Value *Addr = SI->getPointerOperand();
491     int Idx = getMemoryAccessFuncIndex(Addr);
492     if (Idx < 0)
493       return false;
494     const size_t ByteSize = 1 << Idx;
495     const size_t BitSize = ByteSize * 8;
496     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
497     Type *PtrTy = Ty->getPointerTo();
498     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
499                      IRB.CreateIntCast(SI->getValueOperand(), Ty, false),
500                      createOrdering(&IRB, SI->getOrdering())};
501     CallInst *C = CallInst::Create(TsanAtomicStore[Idx],
502                                    ArrayRef<Value*>(Args));
503     ReplaceInstWithInst(I, C);
504   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
505     Value *Addr = RMWI->getPointerOperand();
506     int Idx = getMemoryAccessFuncIndex(Addr);
507     if (Idx < 0)
508       return false;
509     Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
510     if (!F)
511       return false;
512     const size_t ByteSize = 1 << Idx;
513     const size_t BitSize = ByteSize * 8;
514     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
515     Type *PtrTy = Ty->getPointerTo();
516     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
517                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
518                      createOrdering(&IRB, RMWI->getOrdering())};
519     CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args));
520     ReplaceInstWithInst(I, C);
521   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
522     Value *Addr = CASI->getPointerOperand();
523     int Idx = getMemoryAccessFuncIndex(Addr);
524     if (Idx < 0)
525       return false;
526     const size_t ByteSize = 1 << Idx;
527     const size_t BitSize = ByteSize * 8;
528     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
529     Type *PtrTy = Ty->getPointerTo();
530     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
531                      IRB.CreateIntCast(CASI->getCompareOperand(), Ty, false),
532                      IRB.CreateIntCast(CASI->getNewValOperand(), Ty, false),
533                      createOrdering(&IRB, CASI->getSuccessOrdering()),
534                      createOrdering(&IRB, CASI->getFailureOrdering())};
535     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
536     Value *Success = IRB.CreateICmpEQ(C, CASI->getCompareOperand());
537 
538     Value *Res = IRB.CreateInsertValue(UndefValue::get(CASI->getType()), C, 0);
539     Res = IRB.CreateInsertValue(Res, Success, 1);
540 
541     I->replaceAllUsesWith(Res);
542     I->eraseFromParent();
543   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
544     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
545     Function *F = FI->getSynchScope() == SingleThread ?
546         TsanAtomicSignalFence : TsanAtomicThreadFence;
547     CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args));
548     ReplaceInstWithInst(I, C);
549   }
550   return true;
551 }
552 
553 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr) {
554   Type *OrigPtrTy = Addr->getType();
555   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
556   assert(OrigTy->isSized());
557   uint32_t TypeSize = DL->getTypeStoreSizeInBits(OrigTy);
558   if (TypeSize != 8  && TypeSize != 16 &&
559       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
560     NumAccessesWithBadSize++;
561     // Ignore all unusual sizes.
562     return -1;
563   }
564   size_t Idx = countTrailingZeros(TypeSize / 8);
565   assert(Idx < kNumberOfAccessSizes);
566   return Idx;
567 }
568