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 
44 using namespace llvm;
45 
46 #define DEBUG_TYPE "tsan"
47 
48 static cl::opt<bool>  ClInstrumentMemoryAccesses(
49     "tsan-instrument-memory-accesses", cl::init(true),
50     cl::desc("Instrument memory accesses"), cl::Hidden);
51 static cl::opt<bool>  ClInstrumentFuncEntryExit(
52     "tsan-instrument-func-entry-exit", cl::init(true),
53     cl::desc("Instrument function entry and exit"), cl::Hidden);
54 static cl::opt<bool>  ClInstrumentAtomics(
55     "tsan-instrument-atomics", cl::init(true),
56     cl::desc("Instrument atomics"), cl::Hidden);
57 static cl::opt<bool>  ClInstrumentMemIntrinsics(
58     "tsan-instrument-memintrinsics", cl::init(true),
59     cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
60 
61 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
62 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
63 STATISTIC(NumOmittedReadsBeforeWrite,
64           "Number of reads ignored due to following writes");
65 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
66 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
67 STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
68 STATISTIC(NumOmittedReadsFromConstantGlobals,
69           "Number of reads from constant globals");
70 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
71 
72 namespace {
73 
74 /// ThreadSanitizer: instrument the code in module to find races.
75 struct ThreadSanitizer : public FunctionPass {
76   ThreadSanitizer() : FunctionPass(ID), DL(nullptr) {}
77   const char *getPassName() const override;
78   bool runOnFunction(Function &F) override;
79   bool doInitialization(Module &M) override;
80   static char ID;  // Pass identification, replacement for typeid.
81 
82  private:
83   void initializeCallbacks(Module &M);
84   bool instrumentLoadOrStore(Instruction *I);
85   bool instrumentAtomic(Instruction *I);
86   bool instrumentMemIntrinsic(Instruction *I);
87   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction*> &Local,
88                                       SmallVectorImpl<Instruction*> &All);
89   bool addrPointsToConstantData(Value *Addr);
90   int getMemoryAccessFuncIndex(Value *Addr);
91 
92   const DataLayout *DL;
93   Type *IntptrTy;
94   IntegerType *OrdTy;
95   // Callbacks to run-time library are computed in doInitialization.
96   Function *TsanFuncEntry;
97   Function *TsanFuncExit;
98   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
99   static const size_t kNumberOfAccessSizes = 5;
100   Function *TsanRead[kNumberOfAccessSizes];
101   Function *TsanWrite[kNumberOfAccessSizes];
102   Function *TsanUnalignedRead[kNumberOfAccessSizes];
103   Function *TsanUnalignedWrite[kNumberOfAccessSizes];
104   Function *TsanAtomicLoad[kNumberOfAccessSizes];
105   Function *TsanAtomicStore[kNumberOfAccessSizes];
106   Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes];
107   Function *TsanAtomicCAS[kNumberOfAccessSizes];
108   Function *TsanAtomicThreadFence;
109   Function *TsanAtomicSignalFence;
110   Function *TsanVptrUpdate;
111   Function *TsanVptrLoad;
112   Function *MemmoveFn, *MemcpyFn, *MemsetFn;
113 };
114 }  // namespace
115 
116 char ThreadSanitizer::ID = 0;
117 INITIALIZE_PASS(ThreadSanitizer, "tsan",
118     "ThreadSanitizer: detects data races.",
119     false, false)
120 
121 const char *ThreadSanitizer::getPassName() const {
122   return "ThreadSanitizer";
123 }
124 
125 FunctionPass *llvm::createThreadSanitizerPass() {
126   return new ThreadSanitizer();
127 }
128 
129 static Function *checkInterfaceFunction(Constant *FuncOrBitcast) {
130   if (Function *F = dyn_cast<Function>(FuncOrBitcast))
131      return F;
132   FuncOrBitcast->dump();
133   report_fatal_error("ThreadSanitizer interface function redefined");
134 }
135 
136 void ThreadSanitizer::initializeCallbacks(Module &M) {
137   IRBuilder<> IRB(M.getContext());
138   // Initialize the callbacks.
139   TsanFuncEntry = checkInterfaceFunction(M.getOrInsertFunction(
140       "__tsan_func_entry", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
141   TsanFuncExit = checkInterfaceFunction(M.getOrInsertFunction(
142       "__tsan_func_exit", IRB.getVoidTy(), nullptr));
143   OrdTy = IRB.getInt32Ty();
144   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
145     const size_t ByteSize = 1 << i;
146     const size_t BitSize = ByteSize * 8;
147     SmallString<32> ReadName("__tsan_read" + itostr(ByteSize));
148     TsanRead[i] = checkInterfaceFunction(M.getOrInsertFunction(
149         ReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
150 
151     SmallString<32> WriteName("__tsan_write" + itostr(ByteSize));
152     TsanWrite[i] = checkInterfaceFunction(M.getOrInsertFunction(
153         WriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
154 
155     SmallString<64> UnalignedReadName("__tsan_unaligned_read" +
156         itostr(ByteSize));
157     TsanUnalignedRead[i] = checkInterfaceFunction(M.getOrInsertFunction(
158         UnalignedReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
159 
160     SmallString<64> UnalignedWriteName("__tsan_unaligned_write" +
161         itostr(ByteSize));
162     TsanUnalignedWrite[i] = checkInterfaceFunction(M.getOrInsertFunction(
163         UnalignedWriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
164 
165     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
166     Type *PtrTy = Ty->getPointerTo();
167     SmallString<32> AtomicLoadName("__tsan_atomic" + itostr(BitSize) +
168                                    "_load");
169     TsanAtomicLoad[i] = checkInterfaceFunction(M.getOrInsertFunction(
170         AtomicLoadName, Ty, PtrTy, OrdTy, nullptr));
171 
172     SmallString<32> AtomicStoreName("__tsan_atomic" + itostr(BitSize) +
173                                     "_store");
174     TsanAtomicStore[i] = checkInterfaceFunction(M.getOrInsertFunction(
175         AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy,
176         nullptr));
177 
178     for (int op = AtomicRMWInst::FIRST_BINOP;
179         op <= AtomicRMWInst::LAST_BINOP; ++op) {
180       TsanAtomicRMW[op][i] = nullptr;
181       const char *NamePart = nullptr;
182       if (op == AtomicRMWInst::Xchg)
183         NamePart = "_exchange";
184       else if (op == AtomicRMWInst::Add)
185         NamePart = "_fetch_add";
186       else if (op == AtomicRMWInst::Sub)
187         NamePart = "_fetch_sub";
188       else if (op == AtomicRMWInst::And)
189         NamePart = "_fetch_and";
190       else if (op == AtomicRMWInst::Or)
191         NamePart = "_fetch_or";
192       else if (op == AtomicRMWInst::Xor)
193         NamePart = "_fetch_xor";
194       else if (op == AtomicRMWInst::Nand)
195         NamePart = "_fetch_nand";
196       else
197         continue;
198       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
199       TsanAtomicRMW[op][i] = checkInterfaceFunction(M.getOrInsertFunction(
200           RMWName, Ty, PtrTy, Ty, OrdTy, nullptr));
201     }
202 
203     SmallString<32> AtomicCASName("__tsan_atomic" + itostr(BitSize) +
204                                   "_compare_exchange_val");
205     TsanAtomicCAS[i] = checkInterfaceFunction(M.getOrInsertFunction(
206         AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, nullptr));
207   }
208   TsanVptrUpdate = checkInterfaceFunction(M.getOrInsertFunction(
209       "__tsan_vptr_update", IRB.getVoidTy(), IRB.getInt8PtrTy(),
210       IRB.getInt8PtrTy(), nullptr));
211   TsanVptrLoad = checkInterfaceFunction(M.getOrInsertFunction(
212       "__tsan_vptr_read", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
213   TsanAtomicThreadFence = checkInterfaceFunction(M.getOrInsertFunction(
214       "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, nullptr));
215   TsanAtomicSignalFence = checkInterfaceFunction(M.getOrInsertFunction(
216       "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, nullptr));
217 
218   MemmoveFn = checkInterfaceFunction(M.getOrInsertFunction(
219     "memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
220     IRB.getInt8PtrTy(), IntptrTy, nullptr));
221   MemcpyFn = checkInterfaceFunction(M.getOrInsertFunction(
222     "memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
223     IntptrTy, nullptr));
224   MemsetFn = checkInterfaceFunction(M.getOrInsertFunction(
225     "memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
226     IntptrTy, nullptr));
227 }
228 
229 bool ThreadSanitizer::doInitialization(Module &M) {
230   DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
231   if (!DLP)
232     report_fatal_error("data layout missing");
233   DL = &DLP->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 //
276 // We do not handle some of the patterns that should not survive
277 // after the classic compiler optimizations.
278 // E.g. two reads from the same temp should be eliminated by CSE,
279 // two writes should be eliminated by DSE, etc.
280 //
281 // 'Local' is a vector of insns within the same BB (no calls between).
282 // 'All' is a vector of insns that will be instrumented.
283 void ThreadSanitizer::chooseInstructionsToInstrument(
284     SmallVectorImpl<Instruction*> &Local,
285     SmallVectorImpl<Instruction*> &All) {
286   SmallSet<Value*, 8> WriteTargets;
287   // Iterate from the end.
288   for (SmallVectorImpl<Instruction*>::reverse_iterator It = Local.rbegin(),
289        E = Local.rend(); It != E; ++It) {
290     Instruction *I = *It;
291     if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
292       WriteTargets.insert(Store->getPointerOperand());
293     } else {
294       LoadInst *Load = cast<LoadInst>(I);
295       Value *Addr = Load->getPointerOperand();
296       if (WriteTargets.count(Addr)) {
297         // We will write to this temp, so no reason to analyze the read.
298         NumOmittedReadsBeforeWrite++;
299         continue;
300       }
301       if (addrPointsToConstantData(Addr)) {
302         // Addr points to some constant data -- it can not race with any writes.
303         continue;
304       }
305     }
306     All.push_back(I);
307   }
308   Local.clear();
309 }
310 
311 static bool isAtomic(Instruction *I) {
312   if (LoadInst *LI = dyn_cast<LoadInst>(I))
313     return LI->isAtomic() && LI->getSynchScope() == CrossThread;
314   if (StoreInst *SI = dyn_cast<StoreInst>(I))
315     return SI->isAtomic() && SI->getSynchScope() == CrossThread;
316   if (isa<AtomicRMWInst>(I))
317     return true;
318   if (isa<AtomicCmpXchgInst>(I))
319     return true;
320   if (isa<FenceInst>(I))
321     return true;
322   return false;
323 }
324 
325 bool ThreadSanitizer::runOnFunction(Function &F) {
326   if (!DL) return false;
327   initializeCallbacks(*F.getParent());
328   SmallVector<Instruction*, 8> RetVec;
329   SmallVector<Instruction*, 8> AllLoadsAndStores;
330   SmallVector<Instruction*, 8> LocalLoadsAndStores;
331   SmallVector<Instruction*, 8> AtomicAccesses;
332   SmallVector<Instruction*, 8> MemIntrinCalls;
333   bool Res = false;
334   bool HasCalls = false;
335   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
336 
337   // Traverse all instructions, collect loads/stores/returns, check for calls.
338   for (auto &BB : F) {
339     for (auto &Inst : BB) {
340       if (isAtomic(&Inst))
341         AtomicAccesses.push_back(&Inst);
342       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
343         LocalLoadsAndStores.push_back(&Inst);
344       else if (isa<ReturnInst>(Inst))
345         RetVec.push_back(&Inst);
346       else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
347         if (isa<MemIntrinsic>(Inst))
348           MemIntrinCalls.push_back(&Inst);
349         HasCalls = true;
350         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
351       }
352     }
353     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
354   }
355 
356   // We have collected all loads and stores.
357   // FIXME: many of these accesses do not need to be checked for races
358   // (e.g. variables that do not escape, etc).
359 
360   // Instrument memory accesses only if we want to report bugs in the function.
361   if (ClInstrumentMemoryAccesses && SanitizeFunction)
362     for (auto Inst : AllLoadsAndStores) {
363       Res |= instrumentLoadOrStore(Inst);
364     }
365 
366   // Instrument atomic memory accesses in any case (they can be used to
367   // implement synchronization).
368   if (ClInstrumentAtomics)
369     for (auto Inst : AtomicAccesses) {
370       Res |= instrumentAtomic(Inst);
371     }
372 
373   if (ClInstrumentMemIntrinsics && SanitizeFunction)
374     for (auto Inst : MemIntrinCalls) {
375       Res |= instrumentMemIntrinsic(Inst);
376     }
377 
378   // Instrument function entry/exit points if there were instrumented accesses.
379   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
380     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
381     Value *ReturnAddress = IRB.CreateCall(
382         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
383         IRB.getInt32(0));
384     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
385     for (auto RetInst : RetVec) {
386       IRBuilder<> IRBRet(RetInst);
387       IRBRet.CreateCall(TsanFuncExit);
388     }
389     Res = true;
390   }
391   return Res;
392 }
393 
394 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I) {
395   IRBuilder<> IRB(I);
396   bool IsWrite = isa<StoreInst>(*I);
397   Value *Addr = IsWrite
398       ? cast<StoreInst>(I)->getPointerOperand()
399       : cast<LoadInst>(I)->getPointerOperand();
400   int Idx = getMemoryAccessFuncIndex(Addr);
401   if (Idx < 0)
402     return false;
403   if (IsWrite && isVtableAccess(I)) {
404     DEBUG(dbgs() << "  VPTR : " << *I << "\n");
405     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
406     // StoredValue may be a vector type if we are storing several vptrs at once.
407     // In this case, just take the first element of the vector since this is
408     // enough to find vptr races.
409     if (isa<VectorType>(StoredValue->getType()))
410       StoredValue = IRB.CreateExtractElement(
411           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
412     if (StoredValue->getType()->isIntegerTy())
413       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
414     // Call TsanVptrUpdate.
415     IRB.CreateCall2(TsanVptrUpdate,
416                     IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
417                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy()));
418     NumInstrumentedVtableWrites++;
419     return true;
420   }
421   if (!IsWrite && isVtableAccess(I)) {
422     IRB.CreateCall(TsanVptrLoad,
423                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
424     NumInstrumentedVtableReads++;
425     return true;
426   }
427   const unsigned Alignment = IsWrite
428       ? cast<StoreInst>(I)->getAlignment()
429       : cast<LoadInst>(I)->getAlignment();
430   Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
431   const uint32_t TypeSize = DL->getTypeStoreSizeInBits(OrigTy);
432   Value *OnAccessFunc = nullptr;
433   if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0)
434     OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
435   else
436     OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
437   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
438   if (IsWrite) NumInstrumentedWrites++;
439   else         NumInstrumentedReads++;
440   return true;
441 }
442 
443 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
444   uint32_t v = 0;
445   switch (ord) {
446     case NotAtomic: llvm_unreachable("unexpected atomic ordering!");
447     case Unordered:              // Fall-through.
448     case Monotonic:              v = 0; break;
449     // case Consume:                v = 1; break;  // Not specified yet.
450     case Acquire:                v = 2; break;
451     case Release:                v = 3; break;
452     case AcquireRelease:         v = 4; break;
453     case SequentiallyConsistent: v = 5; break;
454   }
455   return IRB->getInt32(v);
456 }
457 
458 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
459 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
460 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
461 // instead we simply replace them with regular function calls, which are then
462 // intercepted by the run-time.
463 // Since tsan is running after everyone else, the calls should not be
464 // replaced back with intrinsics. If that becomes wrong at some point,
465 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
466 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
467   IRBuilder<> IRB(I);
468   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
469     IRB.CreateCall3(MemsetFn,
470       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
471       IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
472       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
473     I->eraseFromParent();
474   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
475     IRB.CreateCall3(isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
476       IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
477       IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
478       IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false));
479     I->eraseFromParent();
480   }
481   return false;
482 }
483 
484 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
485 // standards.  For background see C++11 standard.  A slightly older, publicly
486 // available draft of the standard (not entirely up-to-date, but close enough
487 // for casual browsing) is available here:
488 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
489 // The following page contains more background information:
490 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
491 
492 bool ThreadSanitizer::instrumentAtomic(Instruction *I) {
493   IRBuilder<> IRB(I);
494   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
495     Value *Addr = LI->getPointerOperand();
496     int Idx = getMemoryAccessFuncIndex(Addr);
497     if (Idx < 0)
498       return false;
499     const size_t ByteSize = 1 << Idx;
500     const size_t BitSize = ByteSize * 8;
501     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
502     Type *PtrTy = Ty->getPointerTo();
503     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
504                      createOrdering(&IRB, LI->getOrdering())};
505     CallInst *C = CallInst::Create(TsanAtomicLoad[Idx], Args);
506     ReplaceInstWithInst(I, C);
507 
508   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
509     Value *Addr = SI->getPointerOperand();
510     int Idx = getMemoryAccessFuncIndex(Addr);
511     if (Idx < 0)
512       return false;
513     const size_t ByteSize = 1 << Idx;
514     const size_t BitSize = ByteSize * 8;
515     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
516     Type *PtrTy = Ty->getPointerTo();
517     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
518                      IRB.CreateIntCast(SI->getValueOperand(), Ty, false),
519                      createOrdering(&IRB, SI->getOrdering())};
520     CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
521     ReplaceInstWithInst(I, C);
522   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
523     Value *Addr = RMWI->getPointerOperand();
524     int Idx = getMemoryAccessFuncIndex(Addr);
525     if (Idx < 0)
526       return false;
527     Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
528     if (!F)
529       return false;
530     const size_t ByteSize = 1 << Idx;
531     const size_t BitSize = ByteSize * 8;
532     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
533     Type *PtrTy = Ty->getPointerTo();
534     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
535                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
536                      createOrdering(&IRB, RMWI->getOrdering())};
537     CallInst *C = CallInst::Create(F, Args);
538     ReplaceInstWithInst(I, C);
539   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
540     Value *Addr = CASI->getPointerOperand();
541     int Idx = getMemoryAccessFuncIndex(Addr);
542     if (Idx < 0)
543       return false;
544     const size_t ByteSize = 1 << Idx;
545     const size_t BitSize = ByteSize * 8;
546     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
547     Type *PtrTy = Ty->getPointerTo();
548     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
549                      IRB.CreateIntCast(CASI->getCompareOperand(), Ty, false),
550                      IRB.CreateIntCast(CASI->getNewValOperand(), Ty, false),
551                      createOrdering(&IRB, CASI->getSuccessOrdering()),
552                      createOrdering(&IRB, CASI->getFailureOrdering())};
553     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
554     Value *Success = IRB.CreateICmpEQ(C, CASI->getCompareOperand());
555 
556     Value *Res = IRB.CreateInsertValue(UndefValue::get(CASI->getType()), C, 0);
557     Res = IRB.CreateInsertValue(Res, Success, 1);
558 
559     I->replaceAllUsesWith(Res);
560     I->eraseFromParent();
561   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
562     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
563     Function *F = FI->getSynchScope() == SingleThread ?
564         TsanAtomicSignalFence : TsanAtomicThreadFence;
565     CallInst *C = CallInst::Create(F, Args);
566     ReplaceInstWithInst(I, C);
567   }
568   return true;
569 }
570 
571 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr) {
572   Type *OrigPtrTy = Addr->getType();
573   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
574   assert(OrigTy->isSized());
575   uint32_t TypeSize = DL->getTypeStoreSizeInBits(OrigTy);
576   if (TypeSize != 8  && TypeSize != 16 &&
577       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
578     NumAccessesWithBadSize++;
579     // Ignore all unusual sizes.
580     return -1;
581   }
582   size_t Idx = countTrailingZeros(TypeSize / 8);
583   assert(Idx < kNumberOfAccessSizes);
584   return Idx;
585 }
586