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