1 //===-- ThreadSanitizer.cpp - race detector -------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file is a part of ThreadSanitizer, a race detector.
10 //
11 // The tool is under development, for the details about previous versions see
12 // http://code.google.com/p/data-race-test
13 //
14 // The instrumentation phase is quite simple:
15 //   - Insert calls to run-time library before every memory access.
16 //      - Optimizations may apply to avoid instrumenting some of the accesses.
17 //   - Insert calls at function entry/exit.
18 // The rest is handled by the run-time library.
19 //===----------------------------------------------------------------------===//
20 
21 #include "llvm/Transforms/Instrumentation/ThreadSanitizer.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallString.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/Analysis/CaptureTracking.h"
28 #include "llvm/Analysis/TargetLibraryInfo.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/InitializePasses.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/Instrumentation.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Transforms/Utils/EscapeEnumerator.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Transforms/Utils/ModuleUtils.h"
50 
51 using namespace llvm;
52 
53 #define DEBUG_TYPE "tsan"
54 
55 static cl::opt<bool>  ClInstrumentMemoryAccesses(
56     "tsan-instrument-memory-accesses", cl::init(true),
57     cl::desc("Instrument memory accesses"), cl::Hidden);
58 static cl::opt<bool>  ClInstrumentFuncEntryExit(
59     "tsan-instrument-func-entry-exit", cl::init(true),
60     cl::desc("Instrument function entry and exit"), cl::Hidden);
61 static cl::opt<bool>  ClHandleCxxExceptions(
62     "tsan-handle-cxx-exceptions", cl::init(true),
63     cl::desc("Handle C++ exceptions (insert cleanup blocks for unwinding)"),
64     cl::Hidden);
65 static cl::opt<bool>  ClInstrumentAtomics(
66     "tsan-instrument-atomics", cl::init(true),
67     cl::desc("Instrument atomics"), cl::Hidden);
68 static cl::opt<bool>  ClInstrumentMemIntrinsics(
69     "tsan-instrument-memintrinsics", cl::init(true),
70     cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
71 
72 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
73 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
74 STATISTIC(NumOmittedReadsBeforeWrite,
75           "Number of reads ignored due to following writes");
76 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
77 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
78 STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
79 STATISTIC(NumOmittedReadsFromConstantGlobals,
80           "Number of reads from constant globals");
81 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
82 STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");
83 
84 static const char *const kTsanModuleCtorName = "tsan.module_ctor";
85 static const char *const kTsanInitName = "__tsan_init";
86 
87 namespace {
88 
89 /// ThreadSanitizer: instrument the code in module to find races.
90 ///
91 /// Instantiating ThreadSanitizer inserts the tsan runtime library API function
92 /// declarations into the module if they don't exist already. Instantiating
93 /// ensures the __tsan_init function is in the list of global constructors for
94 /// the module.
95 struct ThreadSanitizer {
96   bool sanitizeFunction(Function &F, const TargetLibraryInfo &TLI);
97 
98 private:
99   void initialize(Module &M);
100   bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL);
101   bool instrumentAtomic(Instruction *I, const DataLayout &DL);
102   bool instrumentMemIntrinsic(Instruction *I);
103   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
104                                       SmallVectorImpl<Instruction *> &All,
105                                       const DataLayout &DL);
106   bool addrPointsToConstantData(Value *Addr);
107   int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL);
108   void InsertRuntimeIgnores(Function &F);
109 
110   Type *IntptrTy;
111   FunctionCallee TsanFuncEntry;
112   FunctionCallee TsanFuncExit;
113   FunctionCallee TsanIgnoreBegin;
114   FunctionCallee TsanIgnoreEnd;
115   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
116   static const size_t kNumberOfAccessSizes = 5;
117   FunctionCallee TsanRead[kNumberOfAccessSizes];
118   FunctionCallee TsanWrite[kNumberOfAccessSizes];
119   FunctionCallee TsanUnalignedRead[kNumberOfAccessSizes];
120   FunctionCallee TsanUnalignedWrite[kNumberOfAccessSizes];
121   FunctionCallee TsanAtomicLoad[kNumberOfAccessSizes];
122   FunctionCallee TsanAtomicStore[kNumberOfAccessSizes];
123   FunctionCallee TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1]
124                               [kNumberOfAccessSizes];
125   FunctionCallee TsanAtomicCAS[kNumberOfAccessSizes];
126   FunctionCallee TsanAtomicThreadFence;
127   FunctionCallee TsanAtomicSignalFence;
128   FunctionCallee TsanVptrUpdate;
129   FunctionCallee TsanVptrLoad;
130   FunctionCallee MemmoveFn, MemcpyFn, MemsetFn;
131 };
132 
133 struct ThreadSanitizerLegacyPass : FunctionPass {
134   ThreadSanitizerLegacyPass() : FunctionPass(ID) {}
135   StringRef getPassName() const override;
136   void getAnalysisUsage(AnalysisUsage &AU) const override;
137   bool runOnFunction(Function &F) override;
138   bool doInitialization(Module &M) override;
139   static char ID; // Pass identification, replacement for typeid.
140 private:
141   Optional<ThreadSanitizer> TSan;
142 };
143 
144 void insertModuleCtor(Module &M) {
145   getOrCreateSanitizerCtorAndInitFunctions(
146       M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{},
147       /*InitArgs=*/{},
148       // This callback is invoked when the functions are created the first
149       // time. Hook them into the global ctors list in that case:
150       [&](Function *Ctor, FunctionCallee) { appendToGlobalCtors(M, Ctor, 0); });
151 }
152 
153 }  // namespace
154 
155 PreservedAnalyses ThreadSanitizerPass::run(Function &F,
156                                            FunctionAnalysisManager &FAM) {
157   ThreadSanitizer TSan;
158   if (TSan.sanitizeFunction(F, FAM.getResult<TargetLibraryAnalysis>(F)))
159     return PreservedAnalyses::none();
160   return PreservedAnalyses::all();
161 }
162 
163 PreservedAnalyses ThreadSanitizerPass::run(Module &M,
164                                            ModuleAnalysisManager &MAM) {
165   insertModuleCtor(M);
166   return PreservedAnalyses::none();
167 }
168 
169 char ThreadSanitizerLegacyPass::ID = 0;
170 INITIALIZE_PASS_BEGIN(ThreadSanitizerLegacyPass, "tsan",
171                       "ThreadSanitizer: detects data races.", false, false)
172 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
173 INITIALIZE_PASS_END(ThreadSanitizerLegacyPass, "tsan",
174                     "ThreadSanitizer: detects data races.", false, false)
175 
176 StringRef ThreadSanitizerLegacyPass::getPassName() const {
177   return "ThreadSanitizerLegacyPass";
178 }
179 
180 void ThreadSanitizerLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
181   AU.addRequired<TargetLibraryInfoWrapperPass>();
182 }
183 
184 bool ThreadSanitizerLegacyPass::doInitialization(Module &M) {
185   insertModuleCtor(M);
186   TSan.emplace();
187   return true;
188 }
189 
190 bool ThreadSanitizerLegacyPass::runOnFunction(Function &F) {
191   auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
192   TSan->sanitizeFunction(F, TLI);
193   return true;
194 }
195 
196 FunctionPass *llvm::createThreadSanitizerLegacyPassPass() {
197   return new ThreadSanitizerLegacyPass();
198 }
199 
200 void ThreadSanitizer::initialize(Module &M) {
201   const DataLayout &DL = M.getDataLayout();
202   IntptrTy = DL.getIntPtrType(M.getContext());
203 
204   IRBuilder<> IRB(M.getContext());
205   AttributeList Attr;
206   Attr = Attr.addAttribute(M.getContext(), AttributeList::FunctionIndex,
207                            Attribute::NoUnwind);
208   // Initialize the callbacks.
209   TsanFuncEntry = M.getOrInsertFunction("__tsan_func_entry", Attr,
210                                         IRB.getVoidTy(), IRB.getInt8PtrTy());
211   TsanFuncExit =
212       M.getOrInsertFunction("__tsan_func_exit", Attr, IRB.getVoidTy());
213   TsanIgnoreBegin = M.getOrInsertFunction("__tsan_ignore_thread_begin", Attr,
214                                           IRB.getVoidTy());
215   TsanIgnoreEnd =
216       M.getOrInsertFunction("__tsan_ignore_thread_end", Attr, IRB.getVoidTy());
217   IntegerType *OrdTy = IRB.getInt32Ty();
218   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
219     const unsigned ByteSize = 1U << i;
220     const unsigned BitSize = ByteSize * 8;
221     std::string ByteSizeStr = utostr(ByteSize);
222     std::string BitSizeStr = utostr(BitSize);
223     SmallString<32> ReadName("__tsan_read" + ByteSizeStr);
224     TsanRead[i] = M.getOrInsertFunction(ReadName, Attr, IRB.getVoidTy(),
225                                         IRB.getInt8PtrTy());
226 
227     SmallString<32> WriteName("__tsan_write" + ByteSizeStr);
228     TsanWrite[i] = M.getOrInsertFunction(WriteName, Attr, IRB.getVoidTy(),
229                                          IRB.getInt8PtrTy());
230 
231     SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr);
232     TsanUnalignedRead[i] = M.getOrInsertFunction(
233         UnalignedReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
234 
235     SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr);
236     TsanUnalignedWrite[i] = M.getOrInsertFunction(
237         UnalignedWriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
238 
239     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
240     Type *PtrTy = Ty->getPointerTo();
241     SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load");
242     TsanAtomicLoad[i] =
243         M.getOrInsertFunction(AtomicLoadName, Attr, Ty, PtrTy, OrdTy);
244 
245     SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store");
246     TsanAtomicStore[i] = M.getOrInsertFunction(
247         AtomicStoreName, Attr, IRB.getVoidTy(), PtrTy, Ty, OrdTy);
248 
249     for (int op = AtomicRMWInst::FIRST_BINOP;
250         op <= AtomicRMWInst::LAST_BINOP; ++op) {
251       TsanAtomicRMW[op][i] = nullptr;
252       const char *NamePart = nullptr;
253       if (op == AtomicRMWInst::Xchg)
254         NamePart = "_exchange";
255       else if (op == AtomicRMWInst::Add)
256         NamePart = "_fetch_add";
257       else if (op == AtomicRMWInst::Sub)
258         NamePart = "_fetch_sub";
259       else if (op == AtomicRMWInst::And)
260         NamePart = "_fetch_and";
261       else if (op == AtomicRMWInst::Or)
262         NamePart = "_fetch_or";
263       else if (op == AtomicRMWInst::Xor)
264         NamePart = "_fetch_xor";
265       else if (op == AtomicRMWInst::Nand)
266         NamePart = "_fetch_nand";
267       else
268         continue;
269       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
270       TsanAtomicRMW[op][i] =
271           M.getOrInsertFunction(RMWName, Attr, Ty, PtrTy, Ty, OrdTy);
272     }
273 
274     SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr +
275                                   "_compare_exchange_val");
276     TsanAtomicCAS[i] = M.getOrInsertFunction(AtomicCASName, Attr, Ty, PtrTy, Ty,
277                                              Ty, OrdTy, OrdTy);
278   }
279   TsanVptrUpdate =
280       M.getOrInsertFunction("__tsan_vptr_update", Attr, IRB.getVoidTy(),
281                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy());
282   TsanVptrLoad = M.getOrInsertFunction("__tsan_vptr_read", Attr,
283                                        IRB.getVoidTy(), IRB.getInt8PtrTy());
284   TsanAtomicThreadFence = M.getOrInsertFunction("__tsan_atomic_thread_fence",
285                                                 Attr, IRB.getVoidTy(), OrdTy);
286   TsanAtomicSignalFence = M.getOrInsertFunction("__tsan_atomic_signal_fence",
287                                                 Attr, IRB.getVoidTy(), OrdTy);
288 
289   MemmoveFn =
290       M.getOrInsertFunction("memmove", Attr, IRB.getInt8PtrTy(),
291                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy);
292   MemcpyFn =
293       M.getOrInsertFunction("memcpy", Attr, IRB.getInt8PtrTy(),
294                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy);
295   MemsetFn =
296       M.getOrInsertFunction("memset", Attr, IRB.getInt8PtrTy(),
297                             IRB.getInt8PtrTy(), IRB.getInt32Ty(), IntptrTy);
298 }
299 
300 static bool isVtableAccess(Instruction *I) {
301   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
302     return Tag->isTBAAVtableAccess();
303   return false;
304 }
305 
306 // Do not instrument known races/"benign races" that come from compiler
307 // instrumentatin. The user has no way of suppressing them.
308 static bool shouldInstrumentReadWriteFromAddress(const Module *M, Value *Addr) {
309   // Peel off GEPs and BitCasts.
310   Addr = Addr->stripInBoundsOffsets();
311 
312   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
313     if (GV->hasSection()) {
314       StringRef SectionName = GV->getSection();
315       // Check if the global is in the PGO counters section.
316       auto OF = Triple(M->getTargetTriple()).getObjectFormat();
317       if (SectionName.endswith(
318               getInstrProfSectionName(IPSK_cnts, OF, /*AddSegmentInfo=*/false)))
319         return false;
320     }
321 
322     // Check if the global is private gcov data.
323     if (GV->getName().startswith("__llvm_gcov") ||
324         GV->getName().startswith("__llvm_gcda"))
325       return false;
326   }
327 
328   // Do not instrument acesses from different address spaces; we cannot deal
329   // with them.
330   if (Addr) {
331     Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
332     if (PtrTy->getPointerAddressSpace() != 0)
333       return false;
334   }
335 
336   return true;
337 }
338 
339 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
340   // If this is a GEP, just analyze its pointer operand.
341   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
342     Addr = GEP->getPointerOperand();
343 
344   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
345     if (GV->isConstant()) {
346       // Reads from constant globals can not race with any writes.
347       NumOmittedReadsFromConstantGlobals++;
348       return true;
349     }
350   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
351     if (isVtableAccess(L)) {
352       // Reads from a vtable pointer can not race with any writes.
353       NumOmittedReadsFromVtable++;
354       return true;
355     }
356   }
357   return false;
358 }
359 
360 // Instrumenting some of the accesses may be proven redundant.
361 // Currently handled:
362 //  - read-before-write (within same BB, no calls between)
363 //  - not captured variables
364 //
365 // We do not handle some of the patterns that should not survive
366 // after the classic compiler optimizations.
367 // E.g. two reads from the same temp should be eliminated by CSE,
368 // two writes should be eliminated by DSE, etc.
369 //
370 // 'Local' is a vector of insns within the same BB (no calls between).
371 // 'All' is a vector of insns that will be instrumented.
372 void ThreadSanitizer::chooseInstructionsToInstrument(
373     SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
374     const DataLayout &DL) {
375   SmallPtrSet<Value*, 8> WriteTargets;
376   // Iterate from the end.
377   for (Instruction *I : reverse(Local)) {
378     if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
379       Value *Addr = Store->getPointerOperand();
380       if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
381         continue;
382       WriteTargets.insert(Addr);
383     } else {
384       LoadInst *Load = cast<LoadInst>(I);
385       Value *Addr = Load->getPointerOperand();
386       if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
387         continue;
388       if (WriteTargets.count(Addr)) {
389         // We will write to this temp, so no reason to analyze the read.
390         NumOmittedReadsBeforeWrite++;
391         continue;
392       }
393       if (addrPointsToConstantData(Addr)) {
394         // Addr points to some constant data -- it can not race with any writes.
395         continue;
396       }
397     }
398     Value *Addr = isa<StoreInst>(*I)
399         ? cast<StoreInst>(I)->getPointerOperand()
400         : cast<LoadInst>(I)->getPointerOperand();
401     if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) &&
402         !PointerMayBeCaptured(Addr, true, true)) {
403       // The variable is addressable but not captured, so it cannot be
404       // referenced from a different thread and participate in a data race
405       // (see llvm/Analysis/CaptureTracking.h for details).
406       NumOmittedNonCaptured++;
407       continue;
408     }
409     All.push_back(I);
410   }
411   Local.clear();
412 }
413 
414 static bool isAtomic(Instruction *I) {
415   // TODO: Ask TTI whether synchronization scope is between threads.
416   if (LoadInst *LI = dyn_cast<LoadInst>(I))
417     return LI->isAtomic() && LI->getSyncScopeID() != SyncScope::SingleThread;
418   if (StoreInst *SI = dyn_cast<StoreInst>(I))
419     return SI->isAtomic() && SI->getSyncScopeID() != SyncScope::SingleThread;
420   if (isa<AtomicRMWInst>(I))
421     return true;
422   if (isa<AtomicCmpXchgInst>(I))
423     return true;
424   if (isa<FenceInst>(I))
425     return true;
426   return false;
427 }
428 
429 void ThreadSanitizer::InsertRuntimeIgnores(Function &F) {
430   IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
431   IRB.CreateCall(TsanIgnoreBegin);
432   EscapeEnumerator EE(F, "tsan_ignore_cleanup", ClHandleCxxExceptions);
433   while (IRBuilder<> *AtExit = EE.Next()) {
434     AtExit->CreateCall(TsanIgnoreEnd);
435   }
436 }
437 
438 bool ThreadSanitizer::sanitizeFunction(Function &F,
439                                        const TargetLibraryInfo &TLI) {
440   // This is required to prevent instrumenting call to __tsan_init from within
441   // the module constructor.
442   if (F.getName() == kTsanModuleCtorName)
443     return false;
444   // Naked functions can not have prologue/epilogue
445   // (__tsan_func_entry/__tsan_func_exit) generated, so don't instrument them at
446   // all.
447   if (F.hasFnAttribute(Attribute::Naked))
448     return false;
449   initialize(*F.getParent());
450   SmallVector<Instruction*, 8> AllLoadsAndStores;
451   SmallVector<Instruction*, 8> LocalLoadsAndStores;
452   SmallVector<Instruction*, 8> AtomicAccesses;
453   SmallVector<Instruction*, 8> MemIntrinCalls;
454   bool Res = false;
455   bool HasCalls = false;
456   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
457   const DataLayout &DL = F.getParent()->getDataLayout();
458 
459   // Traverse all instructions, collect loads/stores/returns, check for calls.
460   for (auto &BB : F) {
461     for (auto &Inst : BB) {
462       if (isAtomic(&Inst))
463         AtomicAccesses.push_back(&Inst);
464       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
465         LocalLoadsAndStores.push_back(&Inst);
466       else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
467         if (CallInst *CI = dyn_cast<CallInst>(&Inst))
468           maybeMarkSanitizerLibraryCallNoBuiltin(CI, &TLI);
469         if (isa<MemIntrinsic>(Inst))
470           MemIntrinCalls.push_back(&Inst);
471         HasCalls = true;
472         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
473                                        DL);
474       }
475     }
476     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
477   }
478 
479   // We have collected all loads and stores.
480   // FIXME: many of these accesses do not need to be checked for races
481   // (e.g. variables that do not escape, etc).
482 
483   // Instrument memory accesses only if we want to report bugs in the function.
484   if (ClInstrumentMemoryAccesses && SanitizeFunction)
485     for (auto Inst : AllLoadsAndStores) {
486       Res |= instrumentLoadOrStore(Inst, DL);
487     }
488 
489   // Instrument atomic memory accesses in any case (they can be used to
490   // implement synchronization).
491   if (ClInstrumentAtomics)
492     for (auto Inst : AtomicAccesses) {
493       Res |= instrumentAtomic(Inst, DL);
494     }
495 
496   if (ClInstrumentMemIntrinsics && SanitizeFunction)
497     for (auto Inst : MemIntrinCalls) {
498       Res |= instrumentMemIntrinsic(Inst);
499     }
500 
501   if (F.hasFnAttribute("sanitize_thread_no_checking_at_run_time")) {
502     assert(!F.hasFnAttribute(Attribute::SanitizeThread));
503     if (HasCalls)
504       InsertRuntimeIgnores(F);
505   }
506 
507   // Instrument function entry/exit points if there were instrumented accesses.
508   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
509     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
510     Value *ReturnAddress = IRB.CreateCall(
511         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
512         IRB.getInt32(0));
513     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
514 
515     EscapeEnumerator EE(F, "tsan_cleanup", ClHandleCxxExceptions);
516     while (IRBuilder<> *AtExit = EE.Next()) {
517       AtExit->CreateCall(TsanFuncExit, {});
518     }
519     Res = true;
520   }
521   return Res;
522 }
523 
524 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I,
525                                             const DataLayout &DL) {
526   IRBuilder<> IRB(I);
527   bool IsWrite = isa<StoreInst>(*I);
528   Value *Addr = IsWrite
529       ? cast<StoreInst>(I)->getPointerOperand()
530       : cast<LoadInst>(I)->getPointerOperand();
531 
532   // swifterror memory addresses are mem2reg promoted by instruction selection.
533   // As such they cannot have regular uses like an instrumentation function and
534   // it makes no sense to track them as memory.
535   if (Addr->isSwiftError())
536     return false;
537 
538   int Idx = getMemoryAccessFuncIndex(Addr, DL);
539   if (Idx < 0)
540     return false;
541   if (IsWrite && isVtableAccess(I)) {
542     LLVM_DEBUG(dbgs() << "  VPTR : " << *I << "\n");
543     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
544     // StoredValue may be a vector type if we are storing several vptrs at once.
545     // In this case, just take the first element of the vector since this is
546     // enough to find vptr races.
547     if (isa<VectorType>(StoredValue->getType()))
548       StoredValue = IRB.CreateExtractElement(
549           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
550     if (StoredValue->getType()->isIntegerTy())
551       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
552     // Call TsanVptrUpdate.
553     IRB.CreateCall(TsanVptrUpdate,
554                    {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
555                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
556     NumInstrumentedVtableWrites++;
557     return true;
558   }
559   if (!IsWrite && isVtableAccess(I)) {
560     IRB.CreateCall(TsanVptrLoad,
561                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
562     NumInstrumentedVtableReads++;
563     return true;
564   }
565   const unsigned Alignment = IsWrite
566       ? cast<StoreInst>(I)->getAlignment()
567       : cast<LoadInst>(I)->getAlignment();
568   Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
569   const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
570   FunctionCallee OnAccessFunc = nullptr;
571   if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0)
572     OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
573   else
574     OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
575   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
576   if (IsWrite) NumInstrumentedWrites++;
577   else         NumInstrumentedReads++;
578   return true;
579 }
580 
581 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
582   uint32_t v = 0;
583   switch (ord) {
584     case AtomicOrdering::NotAtomic:
585       llvm_unreachable("unexpected atomic ordering!");
586     case AtomicOrdering::Unordered:              LLVM_FALLTHROUGH;
587     case AtomicOrdering::Monotonic:              v = 0; break;
588     // Not specified yet:
589     // case AtomicOrdering::Consume:                v = 1; break;
590     case AtomicOrdering::Acquire:                v = 2; break;
591     case AtomicOrdering::Release:                v = 3; break;
592     case AtomicOrdering::AcquireRelease:         v = 4; break;
593     case AtomicOrdering::SequentiallyConsistent: v = 5; break;
594   }
595   return IRB->getInt32(v);
596 }
597 
598 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
599 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
600 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
601 // instead we simply replace them with regular function calls, which are then
602 // intercepted by the run-time.
603 // Since tsan is running after everyone else, the calls should not be
604 // replaced back with intrinsics. If that becomes wrong at some point,
605 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
606 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
607   IRBuilder<> IRB(I);
608   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
609     IRB.CreateCall(
610         MemsetFn,
611         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
612          IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
613          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
614     I->eraseFromParent();
615   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
616     IRB.CreateCall(
617         isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
618         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
619          IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
620          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
621     I->eraseFromParent();
622   }
623   return false;
624 }
625 
626 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
627 // standards.  For background see C++11 standard.  A slightly older, publicly
628 // available draft of the standard (not entirely up-to-date, but close enough
629 // for casual browsing) is available here:
630 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
631 // The following page contains more background information:
632 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
633 
634 bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
635   IRBuilder<> IRB(I);
636   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
637     Value *Addr = LI->getPointerOperand();
638     int Idx = getMemoryAccessFuncIndex(Addr, DL);
639     if (Idx < 0)
640       return false;
641     const unsigned ByteSize = 1U << Idx;
642     const unsigned BitSize = ByteSize * 8;
643     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
644     Type *PtrTy = Ty->getPointerTo();
645     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
646                      createOrdering(&IRB, LI->getOrdering())};
647     Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
648     Value *C = IRB.CreateCall(TsanAtomicLoad[Idx], Args);
649     Value *Cast = IRB.CreateBitOrPointerCast(C, OrigTy);
650     I->replaceAllUsesWith(Cast);
651   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
652     Value *Addr = SI->getPointerOperand();
653     int Idx = getMemoryAccessFuncIndex(Addr, DL);
654     if (Idx < 0)
655       return false;
656     const unsigned ByteSize = 1U << Idx;
657     const unsigned BitSize = ByteSize * 8;
658     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
659     Type *PtrTy = Ty->getPointerTo();
660     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
661                      IRB.CreateBitOrPointerCast(SI->getValueOperand(), Ty),
662                      createOrdering(&IRB, SI->getOrdering())};
663     CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
664     ReplaceInstWithInst(I, C);
665   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
666     Value *Addr = RMWI->getPointerOperand();
667     int Idx = getMemoryAccessFuncIndex(Addr, DL);
668     if (Idx < 0)
669       return false;
670     FunctionCallee F = TsanAtomicRMW[RMWI->getOperation()][Idx];
671     if (!F)
672       return false;
673     const unsigned ByteSize = 1U << Idx;
674     const unsigned BitSize = ByteSize * 8;
675     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
676     Type *PtrTy = Ty->getPointerTo();
677     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
678                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
679                      createOrdering(&IRB, RMWI->getOrdering())};
680     CallInst *C = CallInst::Create(F, Args);
681     ReplaceInstWithInst(I, C);
682   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
683     Value *Addr = CASI->getPointerOperand();
684     int Idx = getMemoryAccessFuncIndex(Addr, DL);
685     if (Idx < 0)
686       return false;
687     const unsigned ByteSize = 1U << Idx;
688     const unsigned BitSize = ByteSize * 8;
689     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
690     Type *PtrTy = Ty->getPointerTo();
691     Value *CmpOperand =
692       IRB.CreateBitOrPointerCast(CASI->getCompareOperand(), Ty);
693     Value *NewOperand =
694       IRB.CreateBitOrPointerCast(CASI->getNewValOperand(), Ty);
695     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
696                      CmpOperand,
697                      NewOperand,
698                      createOrdering(&IRB, CASI->getSuccessOrdering()),
699                      createOrdering(&IRB, CASI->getFailureOrdering())};
700     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
701     Value *Success = IRB.CreateICmpEQ(C, CmpOperand);
702     Value *OldVal = C;
703     Type *OrigOldValTy = CASI->getNewValOperand()->getType();
704     if (Ty != OrigOldValTy) {
705       // The value is a pointer, so we need to cast the return value.
706       OldVal = IRB.CreateIntToPtr(C, OrigOldValTy);
707     }
708 
709     Value *Res =
710       IRB.CreateInsertValue(UndefValue::get(CASI->getType()), OldVal, 0);
711     Res = IRB.CreateInsertValue(Res, Success, 1);
712 
713     I->replaceAllUsesWith(Res);
714     I->eraseFromParent();
715   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
716     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
717     FunctionCallee F = FI->getSyncScopeID() == SyncScope::SingleThread
718                            ? TsanAtomicSignalFence
719                            : TsanAtomicThreadFence;
720     CallInst *C = CallInst::Create(F, Args);
721     ReplaceInstWithInst(I, C);
722   }
723   return true;
724 }
725 
726 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr,
727                                               const DataLayout &DL) {
728   Type *OrigPtrTy = Addr->getType();
729   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
730   assert(OrigTy->isSized());
731   uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
732   if (TypeSize != 8  && TypeSize != 16 &&
733       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
734     NumAccessesWithBadSize++;
735     // Ignore all unusual sizes.
736     return -1;
737   }
738   size_t Idx = countTrailingZeros(TypeSize / 8);
739   assert(Idx < kNumberOfAccessSizes);
740   return Idx;
741 }
742