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/Analysis/CaptureTracking.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/Intrinsics.h"
36 #include "llvm/IR/LLVMContext.h"
37 #include "llvm/IR/Metadata.h"
38 #include "llvm/IR/Module.h"
39 #include "llvm/IR/Type.h"
40 #include "llvm/ProfileData/InstrProf.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/MathExtras.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Transforms/Utils/ModuleUtils.h"
48 
49 using namespace llvm;
50 
51 #define DEBUG_TYPE "tsan"
52 
53 static cl::opt<bool>  ClInstrumentMemoryAccesses(
54     "tsan-instrument-memory-accesses", cl::init(true),
55     cl::desc("Instrument memory accesses"), cl::Hidden);
56 static cl::opt<bool>  ClInstrumentFuncEntryExit(
57     "tsan-instrument-func-entry-exit", cl::init(true),
58     cl::desc("Instrument function entry and exit"), cl::Hidden);
59 static cl::opt<bool>  ClInstrumentAtomics(
60     "tsan-instrument-atomics", cl::init(true),
61     cl::desc("Instrument atomics"), cl::Hidden);
62 static cl::opt<bool>  ClInstrumentMemIntrinsics(
63     "tsan-instrument-memintrinsics", cl::init(true),
64     cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
65 
66 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
67 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
68 STATISTIC(NumOmittedReadsBeforeWrite,
69           "Number of reads ignored due to following writes");
70 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
71 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
72 STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
73 STATISTIC(NumOmittedReadsFromConstantGlobals,
74           "Number of reads from constant globals");
75 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
76 STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");
77 
78 static const char *const kTsanModuleCtorName = "tsan.module_ctor";
79 static const char *const kTsanInitName = "__tsan_init";
80 
81 namespace {
82 
83 /// ThreadSanitizer: instrument the code in module to find races.
84 struct ThreadSanitizer : public FunctionPass {
85   ThreadSanitizer() : FunctionPass(ID) {}
86   StringRef getPassName() const override;
87   void getAnalysisUsage(AnalysisUsage &AU) const override;
88   bool runOnFunction(Function &F) override;
89   bool doInitialization(Module &M) override;
90   static char ID;  // Pass identification, replacement for typeid.
91 
92  private:
93   void initializeCallbacks(Module &M);
94   bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL);
95   bool instrumentAtomic(Instruction *I, const DataLayout &DL);
96   bool instrumentMemIntrinsic(Instruction *I);
97   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
98                                       SmallVectorImpl<Instruction *> &All,
99                                       const DataLayout &DL);
100   bool addrPointsToConstantData(Value *Addr);
101   int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL);
102 
103   Type *IntptrTy;
104   IntegerType *OrdTy;
105   // Callbacks to run-time library are computed in doInitialization.
106   Function *TsanFuncEntry;
107   Function *TsanFuncExit;
108   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
109   static const size_t kNumberOfAccessSizes = 5;
110   Function *TsanRead[kNumberOfAccessSizes];
111   Function *TsanWrite[kNumberOfAccessSizes];
112   Function *TsanUnalignedRead[kNumberOfAccessSizes];
113   Function *TsanUnalignedWrite[kNumberOfAccessSizes];
114   Function *TsanAtomicLoad[kNumberOfAccessSizes];
115   Function *TsanAtomicStore[kNumberOfAccessSizes];
116   Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes];
117   Function *TsanAtomicCAS[kNumberOfAccessSizes];
118   Function *TsanAtomicThreadFence;
119   Function *TsanAtomicSignalFence;
120   Function *TsanVptrUpdate;
121   Function *TsanVptrLoad;
122   Function *MemmoveFn, *MemcpyFn, *MemsetFn;
123   Function *TsanCtorFunction;
124 };
125 }  // namespace
126 
127 char ThreadSanitizer::ID = 0;
128 INITIALIZE_PASS_BEGIN(
129     ThreadSanitizer, "tsan",
130     "ThreadSanitizer: detects data races.",
131     false, false)
132 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
133 INITIALIZE_PASS_END(
134     ThreadSanitizer, "tsan",
135     "ThreadSanitizer: detects data races.",
136     false, false)
137 
138 StringRef ThreadSanitizer::getPassName() const { return "ThreadSanitizer"; }
139 
140 void ThreadSanitizer::getAnalysisUsage(AnalysisUsage &AU) const {
141   AU.addRequired<TargetLibraryInfoWrapperPass>();
142 }
143 
144 FunctionPass *llvm::createThreadSanitizerPass() {
145   return new ThreadSanitizer();
146 }
147 
148 void ThreadSanitizer::initializeCallbacks(Module &M) {
149   IRBuilder<> IRB(M.getContext());
150   // Initialize the callbacks.
151   TsanFuncEntry = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
152       "__tsan_func_entry", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
153   TsanFuncExit = checkSanitizerInterfaceFunction(
154       M.getOrInsertFunction("__tsan_func_exit", IRB.getVoidTy(), nullptr));
155   OrdTy = IRB.getInt32Ty();
156   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
157     const unsigned ByteSize = 1U << i;
158     const unsigned BitSize = ByteSize * 8;
159     std::string ByteSizeStr = utostr(ByteSize);
160     std::string BitSizeStr = utostr(BitSize);
161     SmallString<32> ReadName("__tsan_read" + ByteSizeStr);
162     TsanRead[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
163         ReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
164 
165     SmallString<32> WriteName("__tsan_write" + ByteSizeStr);
166     TsanWrite[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
167         WriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
168 
169     SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr);
170     TsanUnalignedRead[i] =
171         checkSanitizerInterfaceFunction(M.getOrInsertFunction(
172             UnalignedReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
173 
174     SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr);
175     TsanUnalignedWrite[i] =
176         checkSanitizerInterfaceFunction(M.getOrInsertFunction(
177             UnalignedWriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
178 
179     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
180     Type *PtrTy = Ty->getPointerTo();
181     SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load");
182     TsanAtomicLoad[i] = checkSanitizerInterfaceFunction(
183         M.getOrInsertFunction(AtomicLoadName, Ty, PtrTy, OrdTy, nullptr));
184 
185     SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store");
186     TsanAtomicStore[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
187         AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy, nullptr));
188 
189     for (int op = AtomicRMWInst::FIRST_BINOP;
190         op <= AtomicRMWInst::LAST_BINOP; ++op) {
191       TsanAtomicRMW[op][i] = nullptr;
192       const char *NamePart = nullptr;
193       if (op == AtomicRMWInst::Xchg)
194         NamePart = "_exchange";
195       else if (op == AtomicRMWInst::Add)
196         NamePart = "_fetch_add";
197       else if (op == AtomicRMWInst::Sub)
198         NamePart = "_fetch_sub";
199       else if (op == AtomicRMWInst::And)
200         NamePart = "_fetch_and";
201       else if (op == AtomicRMWInst::Or)
202         NamePart = "_fetch_or";
203       else if (op == AtomicRMWInst::Xor)
204         NamePart = "_fetch_xor";
205       else if (op == AtomicRMWInst::Nand)
206         NamePart = "_fetch_nand";
207       else
208         continue;
209       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
210       TsanAtomicRMW[op][i] = checkSanitizerInterfaceFunction(
211           M.getOrInsertFunction(RMWName, Ty, PtrTy, Ty, OrdTy, nullptr));
212     }
213 
214     SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr +
215                                   "_compare_exchange_val");
216     TsanAtomicCAS[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
217         AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, nullptr));
218   }
219   TsanVptrUpdate = checkSanitizerInterfaceFunction(
220       M.getOrInsertFunction("__tsan_vptr_update", IRB.getVoidTy(),
221                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), nullptr));
222   TsanVptrLoad = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
223       "__tsan_vptr_read", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
224   TsanAtomicThreadFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
225       "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, nullptr));
226   TsanAtomicSignalFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
227       "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, nullptr));
228 
229   MemmoveFn = checkSanitizerInterfaceFunction(
230       M.getOrInsertFunction("memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
231                             IRB.getInt8PtrTy(), IntptrTy, nullptr));
232   MemcpyFn = checkSanitizerInterfaceFunction(
233       M.getOrInsertFunction("memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
234                             IRB.getInt8PtrTy(), IntptrTy, nullptr));
235   MemsetFn = checkSanitizerInterfaceFunction(
236       M.getOrInsertFunction("memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
237                             IRB.getInt32Ty(), IntptrTy, nullptr));
238 }
239 
240 bool ThreadSanitizer::doInitialization(Module &M) {
241   const DataLayout &DL = M.getDataLayout();
242   IntptrTy = DL.getIntPtrType(M.getContext());
243   std::tie(TsanCtorFunction, std::ignore) = createSanitizerCtorAndInitFunctions(
244       M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{},
245       /*InitArgs=*/{});
246 
247   appendToGlobalCtors(M, TsanCtorFunction, 0);
248 
249   return true;
250 }
251 
252 static bool isVtableAccess(Instruction *I) {
253   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
254     return Tag->isTBAAVtableAccess();
255   return false;
256 }
257 
258 // Do not instrument known races/"benign races" that come from compiler
259 // instrumentatin. The user has no way of suppressing them.
260 static bool shouldInstrumentReadWriteFromAddress(Value *Addr) {
261   // Peel off GEPs and BitCasts.
262   Addr = Addr->stripInBoundsOffsets();
263 
264   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
265     if (GV->hasSection()) {
266       StringRef SectionName = GV->getSection();
267       // Check if the global is in the PGO counters section.
268       if (SectionName.endswith(getInstrProfCountersSectionName(
269             /*AddSegment=*/false)))
270         return false;
271     }
272 
273     // Check if the global is private gcov data.
274     if (GV->getName().startswith("__llvm_gcov") ||
275         GV->getName().startswith("__llvm_gcda"))
276       return false;
277   }
278 
279   // Do not instrument acesses from different address spaces; we cannot deal
280   // with them.
281   if (Addr) {
282     Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
283     if (PtrTy->getPointerAddressSpace() != 0)
284       return false;
285   }
286 
287   return true;
288 }
289 
290 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
291   // If this is a GEP, just analyze its pointer operand.
292   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
293     Addr = GEP->getPointerOperand();
294 
295   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
296     if (GV->isConstant()) {
297       // Reads from constant globals can not race with any writes.
298       NumOmittedReadsFromConstantGlobals++;
299       return true;
300     }
301   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
302     if (isVtableAccess(L)) {
303       // Reads from a vtable pointer can not race with any writes.
304       NumOmittedReadsFromVtable++;
305       return true;
306     }
307   }
308   return false;
309 }
310 
311 // Instrumenting some of the accesses may be proven redundant.
312 // Currently handled:
313 //  - read-before-write (within same BB, no calls between)
314 //  - not captured variables
315 //
316 // We do not handle some of the patterns that should not survive
317 // after the classic compiler optimizations.
318 // E.g. two reads from the same temp should be eliminated by CSE,
319 // two writes should be eliminated by DSE, etc.
320 //
321 // 'Local' is a vector of insns within the same BB (no calls between).
322 // 'All' is a vector of insns that will be instrumented.
323 void ThreadSanitizer::chooseInstructionsToInstrument(
324     SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
325     const DataLayout &DL) {
326   SmallSet<Value*, 8> WriteTargets;
327   // Iterate from the end.
328   for (Instruction *I : reverse(Local)) {
329     if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
330       Value *Addr = Store->getPointerOperand();
331       if (!shouldInstrumentReadWriteFromAddress(Addr))
332         continue;
333       WriteTargets.insert(Addr);
334     } else {
335       LoadInst *Load = cast<LoadInst>(I);
336       Value *Addr = Load->getPointerOperand();
337       if (!shouldInstrumentReadWriteFromAddress(Addr))
338         continue;
339       if (WriteTargets.count(Addr)) {
340         // We will write to this temp, so no reason to analyze the read.
341         NumOmittedReadsBeforeWrite++;
342         continue;
343       }
344       if (addrPointsToConstantData(Addr)) {
345         // Addr points to some constant data -- it can not race with any writes.
346         continue;
347       }
348     }
349     Value *Addr = isa<StoreInst>(*I)
350         ? cast<StoreInst>(I)->getPointerOperand()
351         : cast<LoadInst>(I)->getPointerOperand();
352     if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) &&
353         !PointerMayBeCaptured(Addr, true, true)) {
354       // The variable is addressable but not captured, so it cannot be
355       // referenced from a different thread and participate in a data race
356       // (see llvm/Analysis/CaptureTracking.h for details).
357       NumOmittedNonCaptured++;
358       continue;
359     }
360     All.push_back(I);
361   }
362   Local.clear();
363 }
364 
365 static bool isAtomic(Instruction *I) {
366   if (LoadInst *LI = dyn_cast<LoadInst>(I))
367     return LI->isAtomic() && LI->getSynchScope() == CrossThread;
368   if (StoreInst *SI = dyn_cast<StoreInst>(I))
369     return SI->isAtomic() && SI->getSynchScope() == CrossThread;
370   if (isa<AtomicRMWInst>(I))
371     return true;
372   if (isa<AtomicCmpXchgInst>(I))
373     return true;
374   if (isa<FenceInst>(I))
375     return true;
376   return false;
377 }
378 
379 bool ThreadSanitizer::runOnFunction(Function &F) {
380   // This is required to prevent instrumenting call to __tsan_init from within
381   // the module constructor.
382   if (&F == TsanCtorFunction)
383     return false;
384   initializeCallbacks(*F.getParent());
385   SmallVector<Instruction*, 8> RetVec;
386   SmallVector<Instruction*, 8> AllLoadsAndStores;
387   SmallVector<Instruction*, 8> LocalLoadsAndStores;
388   SmallVector<Instruction*, 8> AtomicAccesses;
389   SmallVector<Instruction*, 8> MemIntrinCalls;
390   bool Res = false;
391   bool HasCalls = false;
392   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
393   const DataLayout &DL = F.getParent()->getDataLayout();
394   const TargetLibraryInfo *TLI =
395       &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
396 
397   // Traverse all instructions, collect loads/stores/returns, check for calls.
398   for (auto &BB : F) {
399     for (auto &Inst : BB) {
400       if (isAtomic(&Inst))
401         AtomicAccesses.push_back(&Inst);
402       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
403         LocalLoadsAndStores.push_back(&Inst);
404       else if (isa<ReturnInst>(Inst))
405         RetVec.push_back(&Inst);
406       else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
407         if (CallInst *CI = dyn_cast<CallInst>(&Inst))
408           maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
409         if (isa<MemIntrinsic>(Inst))
410           MemIntrinCalls.push_back(&Inst);
411         HasCalls = true;
412         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
413                                        DL);
414       }
415     }
416     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
417   }
418 
419   // We have collected all loads and stores.
420   // FIXME: many of these accesses do not need to be checked for races
421   // (e.g. variables that do not escape, etc).
422 
423   // Instrument memory accesses only if we want to report bugs in the function.
424   if (ClInstrumentMemoryAccesses && SanitizeFunction)
425     for (auto Inst : AllLoadsAndStores) {
426       Res |= instrumentLoadOrStore(Inst, DL);
427     }
428 
429   // Instrument atomic memory accesses in any case (they can be used to
430   // implement synchronization).
431   if (ClInstrumentAtomics)
432     for (auto Inst : AtomicAccesses) {
433       Res |= instrumentAtomic(Inst, DL);
434     }
435 
436   if (ClInstrumentMemIntrinsics && SanitizeFunction)
437     for (auto Inst : MemIntrinCalls) {
438       Res |= instrumentMemIntrinsic(Inst);
439     }
440 
441   // Instrument function entry/exit points if there were instrumented accesses.
442   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
443     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
444     Value *ReturnAddress = IRB.CreateCall(
445         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
446         IRB.getInt32(0));
447     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
448     for (auto RetInst : RetVec) {
449       IRBuilder<> IRBRet(RetInst);
450       IRBRet.CreateCall(TsanFuncExit, {});
451     }
452     Res = true;
453   }
454   return Res;
455 }
456 
457 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I,
458                                             const DataLayout &DL) {
459   IRBuilder<> IRB(I);
460   bool IsWrite = isa<StoreInst>(*I);
461   Value *Addr = IsWrite
462       ? cast<StoreInst>(I)->getPointerOperand()
463       : cast<LoadInst>(I)->getPointerOperand();
464   int Idx = getMemoryAccessFuncIndex(Addr, DL);
465   if (Idx < 0)
466     return false;
467   if (IsWrite && isVtableAccess(I)) {
468     DEBUG(dbgs() << "  VPTR : " << *I << "\n");
469     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
470     // StoredValue may be a vector type if we are storing several vptrs at once.
471     // In this case, just take the first element of the vector since this is
472     // enough to find vptr races.
473     if (isa<VectorType>(StoredValue->getType()))
474       StoredValue = IRB.CreateExtractElement(
475           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
476     if (StoredValue->getType()->isIntegerTy())
477       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
478     // Call TsanVptrUpdate.
479     IRB.CreateCall(TsanVptrUpdate,
480                    {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
481                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
482     NumInstrumentedVtableWrites++;
483     return true;
484   }
485   if (!IsWrite && isVtableAccess(I)) {
486     IRB.CreateCall(TsanVptrLoad,
487                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
488     NumInstrumentedVtableReads++;
489     return true;
490   }
491   const unsigned Alignment = IsWrite
492       ? cast<StoreInst>(I)->getAlignment()
493       : cast<LoadInst>(I)->getAlignment();
494   Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
495   const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
496   Value *OnAccessFunc = nullptr;
497   if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0)
498     OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
499   else
500     OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
501   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
502   if (IsWrite) NumInstrumentedWrites++;
503   else         NumInstrumentedReads++;
504   return true;
505 }
506 
507 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
508   uint32_t v = 0;
509   switch (ord) {
510     case AtomicOrdering::NotAtomic:
511       llvm_unreachable("unexpected atomic ordering!");
512     case AtomicOrdering::Unordered:              LLVM_FALLTHROUGH;
513     case AtomicOrdering::Monotonic:              v = 0; break;
514     // Not specified yet:
515     // case AtomicOrdering::Consume:                v = 1; break;
516     case AtomicOrdering::Acquire:                v = 2; break;
517     case AtomicOrdering::Release:                v = 3; break;
518     case AtomicOrdering::AcquireRelease:         v = 4; break;
519     case AtomicOrdering::SequentiallyConsistent: v = 5; break;
520   }
521   return IRB->getInt32(v);
522 }
523 
524 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
525 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
526 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
527 // instead we simply replace them with regular function calls, which are then
528 // intercepted by the run-time.
529 // Since tsan is running after everyone else, the calls should not be
530 // replaced back with intrinsics. If that becomes wrong at some point,
531 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
532 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
533   IRBuilder<> IRB(I);
534   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
535     IRB.CreateCall(
536         MemsetFn,
537         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
538          IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
539          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
540     I->eraseFromParent();
541   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
542     IRB.CreateCall(
543         isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
544         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
545          IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
546          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
547     I->eraseFromParent();
548   }
549   return false;
550 }
551 
552 static Value *createIntOrPtrToIntCast(Value *V, Type* Ty, IRBuilder<> &IRB) {
553   return isa<PointerType>(V->getType()) ?
554     IRB.CreatePtrToInt(V, Ty) : IRB.CreateIntCast(V, Ty, false);
555 }
556 
557 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
558 // standards.  For background see C++11 standard.  A slightly older, publicly
559 // available draft of the standard (not entirely up-to-date, but close enough
560 // for casual browsing) is available here:
561 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
562 // The following page contains more background information:
563 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
564 
565 bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
566   IRBuilder<> IRB(I);
567   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
568     Value *Addr = LI->getPointerOperand();
569     int Idx = getMemoryAccessFuncIndex(Addr, DL);
570     if (Idx < 0)
571       return false;
572     const unsigned ByteSize = 1U << Idx;
573     const unsigned BitSize = ByteSize * 8;
574     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
575     Type *PtrTy = Ty->getPointerTo();
576     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
577                      createOrdering(&IRB, LI->getOrdering())};
578     Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
579     if (Ty == OrigTy) {
580       Instruction *C = CallInst::Create(TsanAtomicLoad[Idx], Args);
581       ReplaceInstWithInst(I, C);
582     } else {
583       // We are loading a pointer, so we need to cast the return value.
584       Value *C = IRB.CreateCall(TsanAtomicLoad[Idx], Args);
585       Instruction *Cast = CastInst::Create(Instruction::IntToPtr, C, OrigTy);
586       ReplaceInstWithInst(I, Cast);
587     }
588   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
589     Value *Addr = SI->getPointerOperand();
590     int Idx = getMemoryAccessFuncIndex(Addr, DL);
591     if (Idx < 0)
592       return false;
593     const unsigned ByteSize = 1U << Idx;
594     const unsigned BitSize = ByteSize * 8;
595     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
596     Type *PtrTy = Ty->getPointerTo();
597     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
598                      createIntOrPtrToIntCast(SI->getValueOperand(), Ty, IRB),
599                      createOrdering(&IRB, SI->getOrdering())};
600     CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
601     ReplaceInstWithInst(I, C);
602   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
603     Value *Addr = RMWI->getPointerOperand();
604     int Idx = getMemoryAccessFuncIndex(Addr, DL);
605     if (Idx < 0)
606       return false;
607     Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
608     if (!F)
609       return false;
610     const unsigned ByteSize = 1U << Idx;
611     const unsigned BitSize = ByteSize * 8;
612     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
613     Type *PtrTy = Ty->getPointerTo();
614     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
615                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
616                      createOrdering(&IRB, RMWI->getOrdering())};
617     CallInst *C = CallInst::Create(F, Args);
618     ReplaceInstWithInst(I, C);
619   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
620     Value *Addr = CASI->getPointerOperand();
621     int Idx = getMemoryAccessFuncIndex(Addr, DL);
622     if (Idx < 0)
623       return false;
624     const unsigned ByteSize = 1U << Idx;
625     const unsigned BitSize = ByteSize * 8;
626     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
627     Type *PtrTy = Ty->getPointerTo();
628     Value *CmpOperand =
629       createIntOrPtrToIntCast(CASI->getCompareOperand(), Ty, IRB);
630     Value *NewOperand =
631       createIntOrPtrToIntCast(CASI->getNewValOperand(), Ty, IRB);
632     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
633                      CmpOperand,
634                      NewOperand,
635                      createOrdering(&IRB, CASI->getSuccessOrdering()),
636                      createOrdering(&IRB, CASI->getFailureOrdering())};
637     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
638     Value *Success = IRB.CreateICmpEQ(C, CmpOperand);
639     Value *OldVal = C;
640     Type *OrigOldValTy = CASI->getNewValOperand()->getType();
641     if (Ty != OrigOldValTy) {
642       // The value is a pointer, so we need to cast the return value.
643       OldVal = IRB.CreateIntToPtr(C, OrigOldValTy);
644     }
645 
646     Value *Res =
647       IRB.CreateInsertValue(UndefValue::get(CASI->getType()), OldVal, 0);
648     Res = IRB.CreateInsertValue(Res, Success, 1);
649 
650     I->replaceAllUsesWith(Res);
651     I->eraseFromParent();
652   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
653     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
654     Function *F = FI->getSynchScope() == SingleThread ?
655         TsanAtomicSignalFence : TsanAtomicThreadFence;
656     CallInst *C = CallInst::Create(F, Args);
657     ReplaceInstWithInst(I, C);
658   }
659   return true;
660 }
661 
662 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr,
663                                               const DataLayout &DL) {
664   Type *OrigPtrTy = Addr->getType();
665   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
666   assert(OrigTy->isSized());
667   uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
668   if (TypeSize != 8  && TypeSize != 16 &&
669       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
670     NumAccessesWithBadSize++;
671     // Ignore all unusual sizes.
672     return -1;
673   }
674   size_t Idx = countTrailingZeros(TypeSize / 8);
675   assert(Idx < kNumberOfAccessSizes);
676   return Idx;
677 }
678