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/EscapeEnumerator.h"
47 #include "llvm/Transforms/Utils/Local.h"
48 #include "llvm/Transforms/Utils/ModuleUtils.h"
49 
50 using namespace llvm;
51 
52 #define DEBUG_TYPE "tsan"
53 
54 static cl::opt<bool>  ClInstrumentMemoryAccesses(
55     "tsan-instrument-memory-accesses", cl::init(true),
56     cl::desc("Instrument memory accesses"), cl::Hidden);
57 static cl::opt<bool>  ClInstrumentFuncEntryExit(
58     "tsan-instrument-func-entry-exit", cl::init(true),
59     cl::desc("Instrument function entry and exit"), cl::Hidden);
60 static cl::opt<bool>  ClHandleCxxExceptions(
61     "tsan-handle-cxx-exceptions", cl::init(true),
62     cl::desc("Handle C++ exceptions (insert cleanup blocks for unwinding)"),
63     cl::Hidden);
64 static cl::opt<bool>  ClInstrumentAtomics(
65     "tsan-instrument-atomics", cl::init(true),
66     cl::desc("Instrument atomics"), cl::Hidden);
67 static cl::opt<bool>  ClInstrumentMemIntrinsics(
68     "tsan-instrument-memintrinsics", cl::init(true),
69     cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
70 
71 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
72 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
73 STATISTIC(NumOmittedReadsBeforeWrite,
74           "Number of reads ignored due to following writes");
75 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
76 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
77 STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
78 STATISTIC(NumOmittedReadsFromConstantGlobals,
79           "Number of reads from constant globals");
80 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
81 STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");
82 
83 static const char *const kTsanModuleCtorName = "tsan.module_ctor";
84 static const char *const kTsanInitName = "__tsan_init";
85 
86 namespace {
87 
88 /// ThreadSanitizer: instrument the code in module to find races.
89 struct ThreadSanitizer : public FunctionPass {
90   ThreadSanitizer() : FunctionPass(ID) {}
91   StringRef getPassName() const override;
92   void getAnalysisUsage(AnalysisUsage &AU) const override;
93   bool runOnFunction(Function &F) override;
94   bool doInitialization(Module &M) override;
95   static char ID;  // Pass identification, replacement for typeid.
96 
97  private:
98   void initializeCallbacks(Module &M);
99   bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL);
100   bool instrumentAtomic(Instruction *I, const DataLayout &DL);
101   bool instrumentMemIntrinsic(Instruction *I);
102   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
103                                       SmallVectorImpl<Instruction *> &All,
104                                       const DataLayout &DL);
105   bool addrPointsToConstantData(Value *Addr);
106   int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL);
107   void InsertRuntimeIgnores(Function &F);
108 
109   Type *IntptrTy;
110   IntegerType *OrdTy;
111   // Callbacks to run-time library are computed in doInitialization.
112   Function *TsanFuncEntry;
113   Function *TsanFuncExit;
114   Function *TsanIgnoreBegin;
115   Function *TsanIgnoreEnd;
116   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
117   static const size_t kNumberOfAccessSizes = 5;
118   Function *TsanRead[kNumberOfAccessSizes];
119   Function *TsanWrite[kNumberOfAccessSizes];
120   Function *TsanUnalignedRead[kNumberOfAccessSizes];
121   Function *TsanUnalignedWrite[kNumberOfAccessSizes];
122   Function *TsanAtomicLoad[kNumberOfAccessSizes];
123   Function *TsanAtomicStore[kNumberOfAccessSizes];
124   Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes];
125   Function *TsanAtomicCAS[kNumberOfAccessSizes];
126   Function *TsanAtomicThreadFence;
127   Function *TsanAtomicSignalFence;
128   Function *TsanVptrUpdate;
129   Function *TsanVptrLoad;
130   Function *MemmoveFn, *MemcpyFn, *MemsetFn;
131   Function *TsanCtorFunction;
132 };
133 }  // namespace
134 
135 char ThreadSanitizer::ID = 0;
136 INITIALIZE_PASS_BEGIN(
137     ThreadSanitizer, "tsan",
138     "ThreadSanitizer: detects data races.",
139     false, false)
140 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
141 INITIALIZE_PASS_END(
142     ThreadSanitizer, "tsan",
143     "ThreadSanitizer: detects data races.",
144     false, false)
145 
146 StringRef ThreadSanitizer::getPassName() const { return "ThreadSanitizer"; }
147 
148 void ThreadSanitizer::getAnalysisUsage(AnalysisUsage &AU) const {
149   AU.addRequired<TargetLibraryInfoWrapperPass>();
150 }
151 
152 FunctionPass *llvm::createThreadSanitizerPass() {
153   return new ThreadSanitizer();
154 }
155 
156 void ThreadSanitizer::initializeCallbacks(Module &M) {
157   IRBuilder<> IRB(M.getContext());
158   AttributeSet Attr;
159   Attr = Attr.addAttribute(M.getContext(), AttributeSet::FunctionIndex, Attribute::NoUnwind);
160   // Initialize the callbacks.
161   TsanFuncEntry = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
162       "__tsan_func_entry", Attr, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
163   TsanFuncExit = checkSanitizerInterfaceFunction(
164       M.getOrInsertFunction("__tsan_func_exit", Attr, IRB.getVoidTy(), nullptr));
165   TsanIgnoreBegin = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
166       "__tsan_ignore_thread_begin", Attr, IRB.getVoidTy(), nullptr));
167   TsanIgnoreEnd = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
168       "__tsan_ignore_thread_end", Attr, IRB.getVoidTy(), nullptr));
169   OrdTy = IRB.getInt32Ty();
170   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
171     const unsigned ByteSize = 1U << i;
172     const unsigned BitSize = ByteSize * 8;
173     std::string ByteSizeStr = utostr(ByteSize);
174     std::string BitSizeStr = utostr(BitSize);
175     SmallString<32> ReadName("__tsan_read" + ByteSizeStr);
176     TsanRead[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
177         ReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
178 
179     SmallString<32> WriteName("__tsan_write" + ByteSizeStr);
180     TsanWrite[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
181         WriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
182 
183     SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr);
184     TsanUnalignedRead[i] =
185         checkSanitizerInterfaceFunction(M.getOrInsertFunction(
186             UnalignedReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
187 
188     SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr);
189     TsanUnalignedWrite[i] =
190         checkSanitizerInterfaceFunction(M.getOrInsertFunction(
191             UnalignedWriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
192 
193     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
194     Type *PtrTy = Ty->getPointerTo();
195     SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load");
196     TsanAtomicLoad[i] = checkSanitizerInterfaceFunction(
197         M.getOrInsertFunction(AtomicLoadName, Attr, Ty, PtrTy, OrdTy, nullptr));
198 
199     SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store");
200     TsanAtomicStore[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
201         AtomicStoreName, Attr, IRB.getVoidTy(), PtrTy, Ty, OrdTy, nullptr));
202 
203     for (int op = AtomicRMWInst::FIRST_BINOP;
204         op <= AtomicRMWInst::LAST_BINOP; ++op) {
205       TsanAtomicRMW[op][i] = nullptr;
206       const char *NamePart = nullptr;
207       if (op == AtomicRMWInst::Xchg)
208         NamePart = "_exchange";
209       else if (op == AtomicRMWInst::Add)
210         NamePart = "_fetch_add";
211       else if (op == AtomicRMWInst::Sub)
212         NamePart = "_fetch_sub";
213       else if (op == AtomicRMWInst::And)
214         NamePart = "_fetch_and";
215       else if (op == AtomicRMWInst::Or)
216         NamePart = "_fetch_or";
217       else if (op == AtomicRMWInst::Xor)
218         NamePart = "_fetch_xor";
219       else if (op == AtomicRMWInst::Nand)
220         NamePart = "_fetch_nand";
221       else
222         continue;
223       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
224       TsanAtomicRMW[op][i] = checkSanitizerInterfaceFunction(
225           M.getOrInsertFunction(RMWName, Attr, Ty, PtrTy, Ty, OrdTy, nullptr));
226     }
227 
228     SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr +
229                                   "_compare_exchange_val");
230     TsanAtomicCAS[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
231         AtomicCASName, Attr, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, nullptr));
232   }
233   TsanVptrUpdate = checkSanitizerInterfaceFunction(
234       M.getOrInsertFunction("__tsan_vptr_update", Attr, IRB.getVoidTy(),
235                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), nullptr));
236   TsanVptrLoad = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
237       "__tsan_vptr_read", Attr, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
238   TsanAtomicThreadFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
239       "__tsan_atomic_thread_fence", Attr, IRB.getVoidTy(), OrdTy, nullptr));
240   TsanAtomicSignalFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
241       "__tsan_atomic_signal_fence", Attr, IRB.getVoidTy(), OrdTy, nullptr));
242 
243   MemmoveFn = checkSanitizerInterfaceFunction(
244       M.getOrInsertFunction("memmove", Attr, IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
245                             IRB.getInt8PtrTy(), IntptrTy, nullptr));
246   MemcpyFn = checkSanitizerInterfaceFunction(
247       M.getOrInsertFunction("memcpy", Attr, IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
248                             IRB.getInt8PtrTy(), IntptrTy, nullptr));
249   MemsetFn = checkSanitizerInterfaceFunction(
250       M.getOrInsertFunction("memset", Attr, IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
251                             IRB.getInt32Ty(), IntptrTy, nullptr));
252 }
253 
254 bool ThreadSanitizer::doInitialization(Module &M) {
255   const DataLayout &DL = M.getDataLayout();
256   IntptrTy = DL.getIntPtrType(M.getContext());
257   std::tie(TsanCtorFunction, std::ignore) = createSanitizerCtorAndInitFunctions(
258       M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{},
259       /*InitArgs=*/{});
260 
261   appendToGlobalCtors(M, TsanCtorFunction, 0);
262 
263   return true;
264 }
265 
266 static bool isVtableAccess(Instruction *I) {
267   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
268     return Tag->isTBAAVtableAccess();
269   return false;
270 }
271 
272 // Do not instrument known races/"benign races" that come from compiler
273 // instrumentatin. The user has no way of suppressing them.
274 static bool shouldInstrumentReadWriteFromAddress(Value *Addr) {
275   // Peel off GEPs and BitCasts.
276   Addr = Addr->stripInBoundsOffsets();
277 
278   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
279     if (GV->hasSection()) {
280       StringRef SectionName = GV->getSection();
281       // Check if the global is in the PGO counters section.
282       if (SectionName.endswith(getInstrProfCountersSectionName(
283             /*AddSegment=*/false)))
284         return false;
285     }
286 
287     // Check if the global is private gcov data.
288     if (GV->getName().startswith("__llvm_gcov") ||
289         GV->getName().startswith("__llvm_gcda"))
290       return false;
291   }
292 
293   // Do not instrument acesses from different address spaces; we cannot deal
294   // with them.
295   if (Addr) {
296     Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
297     if (PtrTy->getPointerAddressSpace() != 0)
298       return false;
299   }
300 
301   return true;
302 }
303 
304 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
305   // If this is a GEP, just analyze its pointer operand.
306   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
307     Addr = GEP->getPointerOperand();
308 
309   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
310     if (GV->isConstant()) {
311       // Reads from constant globals can not race with any writes.
312       NumOmittedReadsFromConstantGlobals++;
313       return true;
314     }
315   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
316     if (isVtableAccess(L)) {
317       // Reads from a vtable pointer can not race with any writes.
318       NumOmittedReadsFromVtable++;
319       return true;
320     }
321   }
322   return false;
323 }
324 
325 // Instrumenting some of the accesses may be proven redundant.
326 // Currently handled:
327 //  - read-before-write (within same BB, no calls between)
328 //  - not captured variables
329 //
330 // We do not handle some of the patterns that should not survive
331 // after the classic compiler optimizations.
332 // E.g. two reads from the same temp should be eliminated by CSE,
333 // two writes should be eliminated by DSE, etc.
334 //
335 // 'Local' is a vector of insns within the same BB (no calls between).
336 // 'All' is a vector of insns that will be instrumented.
337 void ThreadSanitizer::chooseInstructionsToInstrument(
338     SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
339     const DataLayout &DL) {
340   SmallSet<Value*, 8> WriteTargets;
341   // Iterate from the end.
342   for (Instruction *I : reverse(Local)) {
343     if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
344       Value *Addr = Store->getPointerOperand();
345       if (!shouldInstrumentReadWriteFromAddress(Addr))
346         continue;
347       WriteTargets.insert(Addr);
348     } else {
349       LoadInst *Load = cast<LoadInst>(I);
350       Value *Addr = Load->getPointerOperand();
351       if (!shouldInstrumentReadWriteFromAddress(Addr))
352         continue;
353       if (WriteTargets.count(Addr)) {
354         // We will write to this temp, so no reason to analyze the read.
355         NumOmittedReadsBeforeWrite++;
356         continue;
357       }
358       if (addrPointsToConstantData(Addr)) {
359         // Addr points to some constant data -- it can not race with any writes.
360         continue;
361       }
362     }
363     Value *Addr = isa<StoreInst>(*I)
364         ? cast<StoreInst>(I)->getPointerOperand()
365         : cast<LoadInst>(I)->getPointerOperand();
366     if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) &&
367         !PointerMayBeCaptured(Addr, true, true)) {
368       // The variable is addressable but not captured, so it cannot be
369       // referenced from a different thread and participate in a data race
370       // (see llvm/Analysis/CaptureTracking.h for details).
371       NumOmittedNonCaptured++;
372       continue;
373     }
374     All.push_back(I);
375   }
376   Local.clear();
377 }
378 
379 static bool isAtomic(Instruction *I) {
380   if (LoadInst *LI = dyn_cast<LoadInst>(I))
381     return LI->isAtomic() && LI->getSynchScope() == CrossThread;
382   if (StoreInst *SI = dyn_cast<StoreInst>(I))
383     return SI->isAtomic() && SI->getSynchScope() == CrossThread;
384   if (isa<AtomicRMWInst>(I))
385     return true;
386   if (isa<AtomicCmpXchgInst>(I))
387     return true;
388   if (isa<FenceInst>(I))
389     return true;
390   return false;
391 }
392 
393 void ThreadSanitizer::InsertRuntimeIgnores(Function &F) {
394   IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
395   IRB.CreateCall(TsanIgnoreBegin);
396   EscapeEnumerator EE(F, "tsan_ignore_cleanup", ClHandleCxxExceptions);
397   while (IRBuilder<> *AtExit = EE.Next()) {
398     AtExit->CreateCall(TsanIgnoreEnd);
399   }
400 }
401 
402 bool ThreadSanitizer::runOnFunction(Function &F) {
403   // This is required to prevent instrumenting call to __tsan_init from within
404   // the module constructor.
405   if (&F == TsanCtorFunction)
406     return false;
407   initializeCallbacks(*F.getParent());
408   SmallVector<Instruction*, 8> AllLoadsAndStores;
409   SmallVector<Instruction*, 8> LocalLoadsAndStores;
410   SmallVector<Instruction*, 8> AtomicAccesses;
411   SmallVector<Instruction*, 8> MemIntrinCalls;
412   bool Res = false;
413   bool HasCalls = false;
414   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
415   const DataLayout &DL = F.getParent()->getDataLayout();
416   const TargetLibraryInfo *TLI =
417       &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
418 
419   // Traverse all instructions, collect loads/stores/returns, check for calls.
420   for (auto &BB : F) {
421     for (auto &Inst : BB) {
422       if (isAtomic(&Inst))
423         AtomicAccesses.push_back(&Inst);
424       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
425         LocalLoadsAndStores.push_back(&Inst);
426       else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
427         if (CallInst *CI = dyn_cast<CallInst>(&Inst))
428           maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
429         if (isa<MemIntrinsic>(Inst))
430           MemIntrinCalls.push_back(&Inst);
431         HasCalls = true;
432         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
433                                        DL);
434       }
435     }
436     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
437   }
438 
439   // We have collected all loads and stores.
440   // FIXME: many of these accesses do not need to be checked for races
441   // (e.g. variables that do not escape, etc).
442 
443   // Instrument memory accesses only if we want to report bugs in the function.
444   if (ClInstrumentMemoryAccesses && SanitizeFunction)
445     for (auto Inst : AllLoadsAndStores) {
446       Res |= instrumentLoadOrStore(Inst, DL);
447     }
448 
449   // Instrument atomic memory accesses in any case (they can be used to
450   // implement synchronization).
451   if (ClInstrumentAtomics)
452     for (auto Inst : AtomicAccesses) {
453       Res |= instrumentAtomic(Inst, DL);
454     }
455 
456   if (ClInstrumentMemIntrinsics && SanitizeFunction)
457     for (auto Inst : MemIntrinCalls) {
458       Res |= instrumentMemIntrinsic(Inst);
459     }
460 
461   if (F.hasFnAttribute("sanitize_thread_no_checking_at_run_time")) {
462     assert(!F.hasFnAttribute(Attribute::SanitizeThread));
463     if (HasCalls)
464       InsertRuntimeIgnores(F);
465   }
466 
467   // Instrument function entry/exit points if there were instrumented accesses.
468   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
469     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
470     Value *ReturnAddress = IRB.CreateCall(
471         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
472         IRB.getInt32(0));
473     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
474 
475     EscapeEnumerator EE(F, "tsan_cleanup", ClHandleCxxExceptions);
476     while (IRBuilder<> *AtExit = EE.Next()) {
477       AtExit->CreateCall(TsanFuncExit, {});
478     }
479     Res = true;
480   }
481   return Res;
482 }
483 
484 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I,
485                                             const DataLayout &DL) {
486   IRBuilder<> IRB(I);
487   bool IsWrite = isa<StoreInst>(*I);
488   Value *Addr = IsWrite
489       ? cast<StoreInst>(I)->getPointerOperand()
490       : cast<LoadInst>(I)->getPointerOperand();
491   int Idx = getMemoryAccessFuncIndex(Addr, DL);
492   if (Idx < 0)
493     return false;
494   if (IsWrite && isVtableAccess(I)) {
495     DEBUG(dbgs() << "  VPTR : " << *I << "\n");
496     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
497     // StoredValue may be a vector type if we are storing several vptrs at once.
498     // In this case, just take the first element of the vector since this is
499     // enough to find vptr races.
500     if (isa<VectorType>(StoredValue->getType()))
501       StoredValue = IRB.CreateExtractElement(
502           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
503     if (StoredValue->getType()->isIntegerTy())
504       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
505     // Call TsanVptrUpdate.
506     IRB.CreateCall(TsanVptrUpdate,
507                    {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
508                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
509     NumInstrumentedVtableWrites++;
510     return true;
511   }
512   if (!IsWrite && isVtableAccess(I)) {
513     IRB.CreateCall(TsanVptrLoad,
514                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
515     NumInstrumentedVtableReads++;
516     return true;
517   }
518   const unsigned Alignment = IsWrite
519       ? cast<StoreInst>(I)->getAlignment()
520       : cast<LoadInst>(I)->getAlignment();
521   Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
522   const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
523   Value *OnAccessFunc = nullptr;
524   if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0)
525     OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
526   else
527     OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
528   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
529   if (IsWrite) NumInstrumentedWrites++;
530   else         NumInstrumentedReads++;
531   return true;
532 }
533 
534 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
535   uint32_t v = 0;
536   switch (ord) {
537     case AtomicOrdering::NotAtomic:
538       llvm_unreachable("unexpected atomic ordering!");
539     case AtomicOrdering::Unordered:              LLVM_FALLTHROUGH;
540     case AtomicOrdering::Monotonic:              v = 0; break;
541     // Not specified yet:
542     // case AtomicOrdering::Consume:                v = 1; break;
543     case AtomicOrdering::Acquire:                v = 2; break;
544     case AtomicOrdering::Release:                v = 3; break;
545     case AtomicOrdering::AcquireRelease:         v = 4; break;
546     case AtomicOrdering::SequentiallyConsistent: v = 5; break;
547   }
548   return IRB->getInt32(v);
549 }
550 
551 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
552 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
553 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
554 // instead we simply replace them with regular function calls, which are then
555 // intercepted by the run-time.
556 // Since tsan is running after everyone else, the calls should not be
557 // replaced back with intrinsics. If that becomes wrong at some point,
558 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
559 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
560   IRBuilder<> IRB(I);
561   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
562     IRB.CreateCall(
563         MemsetFn,
564         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
565          IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
566          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
567     I->eraseFromParent();
568   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
569     IRB.CreateCall(
570         isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
571         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
572          IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
573          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
574     I->eraseFromParent();
575   }
576   return false;
577 }
578 
579 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
580 // standards.  For background see C++11 standard.  A slightly older, publicly
581 // available draft of the standard (not entirely up-to-date, but close enough
582 // for casual browsing) is available here:
583 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
584 // The following page contains more background information:
585 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
586 
587 bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
588   IRBuilder<> IRB(I);
589   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
590     Value *Addr = LI->getPointerOperand();
591     int Idx = getMemoryAccessFuncIndex(Addr, DL);
592     if (Idx < 0)
593       return false;
594     const unsigned ByteSize = 1U << Idx;
595     const unsigned BitSize = ByteSize * 8;
596     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
597     Type *PtrTy = Ty->getPointerTo();
598     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
599                      createOrdering(&IRB, LI->getOrdering())};
600     Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
601     Value *C = IRB.CreateCall(TsanAtomicLoad[Idx], Args);
602     Value *Cast = IRB.CreateBitOrPointerCast(C, OrigTy);
603     I->replaceAllUsesWith(Cast);
604   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
605     Value *Addr = SI->getPointerOperand();
606     int Idx = getMemoryAccessFuncIndex(Addr, DL);
607     if (Idx < 0)
608       return false;
609     const unsigned ByteSize = 1U << Idx;
610     const unsigned BitSize = ByteSize * 8;
611     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
612     Type *PtrTy = Ty->getPointerTo();
613     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
614                      IRB.CreateBitOrPointerCast(SI->getValueOperand(), Ty),
615                      createOrdering(&IRB, SI->getOrdering())};
616     CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
617     ReplaceInstWithInst(I, C);
618   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
619     Value *Addr = RMWI->getPointerOperand();
620     int Idx = getMemoryAccessFuncIndex(Addr, DL);
621     if (Idx < 0)
622       return false;
623     Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
624     if (!F)
625       return false;
626     const unsigned ByteSize = 1U << Idx;
627     const unsigned BitSize = ByteSize * 8;
628     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
629     Type *PtrTy = Ty->getPointerTo();
630     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
631                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
632                      createOrdering(&IRB, RMWI->getOrdering())};
633     CallInst *C = CallInst::Create(F, Args);
634     ReplaceInstWithInst(I, C);
635   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
636     Value *Addr = CASI->getPointerOperand();
637     int Idx = getMemoryAccessFuncIndex(Addr, DL);
638     if (Idx < 0)
639       return false;
640     const unsigned ByteSize = 1U << Idx;
641     const unsigned BitSize = ByteSize * 8;
642     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
643     Type *PtrTy = Ty->getPointerTo();
644     Value *CmpOperand =
645       IRB.CreateBitOrPointerCast(CASI->getCompareOperand(), Ty);
646     Value *NewOperand =
647       IRB.CreateBitOrPointerCast(CASI->getNewValOperand(), Ty);
648     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
649                      CmpOperand,
650                      NewOperand,
651                      createOrdering(&IRB, CASI->getSuccessOrdering()),
652                      createOrdering(&IRB, CASI->getFailureOrdering())};
653     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
654     Value *Success = IRB.CreateICmpEQ(C, CmpOperand);
655     Value *OldVal = C;
656     Type *OrigOldValTy = CASI->getNewValOperand()->getType();
657     if (Ty != OrigOldValTy) {
658       // The value is a pointer, so we need to cast the return value.
659       OldVal = IRB.CreateIntToPtr(C, OrigOldValTy);
660     }
661 
662     Value *Res =
663       IRB.CreateInsertValue(UndefValue::get(CASI->getType()), OldVal, 0);
664     Res = IRB.CreateInsertValue(Res, Success, 1);
665 
666     I->replaceAllUsesWith(Res);
667     I->eraseFromParent();
668   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
669     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
670     Function *F = FI->getSynchScope() == SingleThread ?
671         TsanAtomicSignalFence : TsanAtomicThreadFence;
672     CallInst *C = CallInst::Create(F, Args);
673     ReplaceInstWithInst(I, C);
674   }
675   return true;
676 }
677 
678 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr,
679                                               const DataLayout &DL) {
680   Type *OrigPtrTy = Addr->getType();
681   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
682   assert(OrigTy->isSized());
683   uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
684   if (TypeSize != 8  && TypeSize != 16 &&
685       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
686     NumAccessesWithBadSize++;
687     // Ignore all unusual sizes.
688     return -1;
689   }
690   size_t Idx = countTrailingZeros(TypeSize / 8);
691   assert(Idx < kNumberOfAccessSizes);
692   return Idx;
693 }
694