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