xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Instrumentation/ThreadSanitizer.cpp (revision 162ae9c834f6d9f9cb443bd62cceb23e0b5fef48)
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/Transforms/Utils/Local.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/Instrumentation.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Transforms/Utils/EscapeEnumerator.h"
48 #include "llvm/Transforms/Utils/ModuleUtils.h"
49 
50 using namespace llvm;
51 
52 #define DEBUG_TYPE "tsan"
53 
54 static cl::opt<bool>  ClInstrumentMemoryAccesses(
55     "tsan-instrument-memory-accesses", cl::init(true),
56     cl::desc("Instrument memory accesses"), cl::Hidden);
57 static cl::opt<bool>  ClInstrumentFuncEntryExit(
58     "tsan-instrument-func-entry-exit", cl::init(true),
59     cl::desc("Instrument function entry and exit"), cl::Hidden);
60 static cl::opt<bool>  ClHandleCxxExceptions(
61     "tsan-handle-cxx-exceptions", cl::init(true),
62     cl::desc("Handle C++ exceptions (insert cleanup blocks for unwinding)"),
63     cl::Hidden);
64 static cl::opt<bool>  ClInstrumentAtomics(
65     "tsan-instrument-atomics", cl::init(true),
66     cl::desc("Instrument atomics"), cl::Hidden);
67 static cl::opt<bool>  ClInstrumentMemIntrinsics(
68     "tsan-instrument-memintrinsics", cl::init(true),
69     cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
70 
71 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
72 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
73 STATISTIC(NumOmittedReadsBeforeWrite,
74           "Number of reads ignored due to following writes");
75 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
76 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
77 STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
78 STATISTIC(NumOmittedReadsFromConstantGlobals,
79           "Number of reads from constant globals");
80 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
81 STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");
82 
83 static const char *const kTsanModuleCtorName = "tsan.module_ctor";
84 static const char *const kTsanInitName = "__tsan_init";
85 
86 namespace {
87 
88 /// ThreadSanitizer: instrument the code in module to find races.
89 ///
90 /// Instantiating ThreadSanitizer inserts the tsan runtime library API function
91 /// declarations into the module if they don't exist already. Instantiating
92 /// ensures the __tsan_init function is in the list of global constructors for
93 /// the module.
94 struct ThreadSanitizer {
95   ThreadSanitizer(Module &M);
96   bool sanitizeFunction(Function &F, const TargetLibraryInfo &TLI);
97 
98 private:
99   void initializeCallbacks(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   IntegerType *OrdTy;
112   // Callbacks to run-time library are computed in doInitialization.
113   FunctionCallee TsanFuncEntry;
114   FunctionCallee TsanFuncExit;
115   FunctionCallee TsanIgnoreBegin;
116   FunctionCallee TsanIgnoreEnd;
117   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
118   static const size_t kNumberOfAccessSizes = 5;
119   FunctionCallee TsanRead[kNumberOfAccessSizes];
120   FunctionCallee TsanWrite[kNumberOfAccessSizes];
121   FunctionCallee TsanUnalignedRead[kNumberOfAccessSizes];
122   FunctionCallee TsanUnalignedWrite[kNumberOfAccessSizes];
123   FunctionCallee TsanAtomicLoad[kNumberOfAccessSizes];
124   FunctionCallee TsanAtomicStore[kNumberOfAccessSizes];
125   FunctionCallee TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1]
126                               [kNumberOfAccessSizes];
127   FunctionCallee TsanAtomicCAS[kNumberOfAccessSizes];
128   FunctionCallee TsanAtomicThreadFence;
129   FunctionCallee TsanAtomicSignalFence;
130   FunctionCallee TsanVptrUpdate;
131   FunctionCallee TsanVptrLoad;
132   FunctionCallee MemmoveFn, MemcpyFn, MemsetFn;
133   Function *TsanCtorFunction;
134 };
135 
136 struct ThreadSanitizerLegacyPass : FunctionPass {
137   ThreadSanitizerLegacyPass() : FunctionPass(ID) {}
138   StringRef getPassName() const override;
139   void getAnalysisUsage(AnalysisUsage &AU) const override;
140   bool runOnFunction(Function &F) override;
141   bool doInitialization(Module &M) override;
142   static char ID; // Pass identification, replacement for typeid.
143 private:
144   Optional<ThreadSanitizer> TSan;
145 };
146 }  // namespace
147 
148 PreservedAnalyses ThreadSanitizerPass::run(Function &F,
149                                            FunctionAnalysisManager &FAM) {
150   ThreadSanitizer TSan(*F.getParent());
151   if (TSan.sanitizeFunction(F, FAM.getResult<TargetLibraryAnalysis>(F)))
152     return PreservedAnalyses::none();
153   return PreservedAnalyses::all();
154 }
155 
156 char ThreadSanitizerLegacyPass::ID = 0;
157 INITIALIZE_PASS_BEGIN(ThreadSanitizerLegacyPass, "tsan",
158                       "ThreadSanitizer: detects data races.", false, false)
159 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
160 INITIALIZE_PASS_END(ThreadSanitizerLegacyPass, "tsan",
161                     "ThreadSanitizer: detects data races.", false, false)
162 
163 StringRef ThreadSanitizerLegacyPass::getPassName() const {
164   return "ThreadSanitizerLegacyPass";
165 }
166 
167 void ThreadSanitizerLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
168   AU.addRequired<TargetLibraryInfoWrapperPass>();
169 }
170 
171 bool ThreadSanitizerLegacyPass::doInitialization(Module &M) {
172   TSan.emplace(M);
173   return true;
174 }
175 
176 bool ThreadSanitizerLegacyPass::runOnFunction(Function &F) {
177   auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
178   TSan->sanitizeFunction(F, TLI);
179   return true;
180 }
181 
182 FunctionPass *llvm::createThreadSanitizerLegacyPassPass() {
183   return new ThreadSanitizerLegacyPass();
184 }
185 
186 void ThreadSanitizer::initializeCallbacks(Module &M) {
187   IRBuilder<> IRB(M.getContext());
188   AttributeList Attr;
189   Attr = Attr.addAttribute(M.getContext(), AttributeList::FunctionIndex,
190                            Attribute::NoUnwind);
191   // Initialize the callbacks.
192   TsanFuncEntry = M.getOrInsertFunction("__tsan_func_entry", Attr,
193                                         IRB.getVoidTy(), IRB.getInt8PtrTy());
194   TsanFuncExit =
195       M.getOrInsertFunction("__tsan_func_exit", Attr, IRB.getVoidTy());
196   TsanIgnoreBegin = M.getOrInsertFunction("__tsan_ignore_thread_begin", Attr,
197                                           IRB.getVoidTy());
198   TsanIgnoreEnd =
199       M.getOrInsertFunction("__tsan_ignore_thread_end", Attr, IRB.getVoidTy());
200   OrdTy = IRB.getInt32Ty();
201   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
202     const unsigned ByteSize = 1U << i;
203     const unsigned BitSize = ByteSize * 8;
204     std::string ByteSizeStr = utostr(ByteSize);
205     std::string BitSizeStr = utostr(BitSize);
206     SmallString<32> ReadName("__tsan_read" + ByteSizeStr);
207     TsanRead[i] = M.getOrInsertFunction(ReadName, Attr, IRB.getVoidTy(),
208                                         IRB.getInt8PtrTy());
209 
210     SmallString<32> WriteName("__tsan_write" + ByteSizeStr);
211     TsanWrite[i] = M.getOrInsertFunction(WriteName, Attr, IRB.getVoidTy(),
212                                          IRB.getInt8PtrTy());
213 
214     SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr);
215     TsanUnalignedRead[i] = M.getOrInsertFunction(
216         UnalignedReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
217 
218     SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr);
219     TsanUnalignedWrite[i] = M.getOrInsertFunction(
220         UnalignedWriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
221 
222     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
223     Type *PtrTy = Ty->getPointerTo();
224     SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load");
225     TsanAtomicLoad[i] =
226         M.getOrInsertFunction(AtomicLoadName, Attr, Ty, PtrTy, OrdTy);
227 
228     SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store");
229     TsanAtomicStore[i] = M.getOrInsertFunction(
230         AtomicStoreName, Attr, IRB.getVoidTy(), PtrTy, Ty, OrdTy);
231 
232     for (int op = AtomicRMWInst::FIRST_BINOP;
233         op <= AtomicRMWInst::LAST_BINOP; ++op) {
234       TsanAtomicRMW[op][i] = nullptr;
235       const char *NamePart = nullptr;
236       if (op == AtomicRMWInst::Xchg)
237         NamePart = "_exchange";
238       else if (op == AtomicRMWInst::Add)
239         NamePart = "_fetch_add";
240       else if (op == AtomicRMWInst::Sub)
241         NamePart = "_fetch_sub";
242       else if (op == AtomicRMWInst::And)
243         NamePart = "_fetch_and";
244       else if (op == AtomicRMWInst::Or)
245         NamePart = "_fetch_or";
246       else if (op == AtomicRMWInst::Xor)
247         NamePart = "_fetch_xor";
248       else if (op == AtomicRMWInst::Nand)
249         NamePart = "_fetch_nand";
250       else
251         continue;
252       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
253       TsanAtomicRMW[op][i] =
254           M.getOrInsertFunction(RMWName, Attr, Ty, PtrTy, Ty, OrdTy);
255     }
256 
257     SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr +
258                                   "_compare_exchange_val");
259     TsanAtomicCAS[i] = M.getOrInsertFunction(AtomicCASName, Attr, Ty, PtrTy, Ty,
260                                              Ty, OrdTy, OrdTy);
261   }
262   TsanVptrUpdate =
263       M.getOrInsertFunction("__tsan_vptr_update", Attr, IRB.getVoidTy(),
264                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy());
265   TsanVptrLoad = M.getOrInsertFunction("__tsan_vptr_read", Attr,
266                                        IRB.getVoidTy(), IRB.getInt8PtrTy());
267   TsanAtomicThreadFence = M.getOrInsertFunction("__tsan_atomic_thread_fence",
268                                                 Attr, IRB.getVoidTy(), OrdTy);
269   TsanAtomicSignalFence = M.getOrInsertFunction("__tsan_atomic_signal_fence",
270                                                 Attr, IRB.getVoidTy(), OrdTy);
271 
272   MemmoveFn =
273       M.getOrInsertFunction("memmove", Attr, IRB.getInt8PtrTy(),
274                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy);
275   MemcpyFn =
276       M.getOrInsertFunction("memcpy", Attr, IRB.getInt8PtrTy(),
277                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy);
278   MemsetFn =
279       M.getOrInsertFunction("memset", Attr, IRB.getInt8PtrTy(),
280                             IRB.getInt8PtrTy(), IRB.getInt32Ty(), IntptrTy);
281 }
282 
283 ThreadSanitizer::ThreadSanitizer(Module &M) {
284   const DataLayout &DL = M.getDataLayout();
285   IntptrTy = DL.getIntPtrType(M.getContext());
286   std::tie(TsanCtorFunction, std::ignore) =
287       getOrCreateSanitizerCtorAndInitFunctions(
288           M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{},
289           /*InitArgs=*/{},
290           // This callback is invoked when the functions are created the first
291           // time. Hook them into the global ctors list in that case:
292           [&](Function *Ctor, FunctionCallee) {
293             appendToGlobalCtors(M, Ctor, 0);
294           });
295 }
296 
297 static bool isVtableAccess(Instruction *I) {
298   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
299     return Tag->isTBAAVtableAccess();
300   return false;
301 }
302 
303 // Do not instrument known races/"benign races" that come from compiler
304 // instrumentatin. The user has no way of suppressing them.
305 static bool shouldInstrumentReadWriteFromAddress(const Module *M, Value *Addr) {
306   // Peel off GEPs and BitCasts.
307   Addr = Addr->stripInBoundsOffsets();
308 
309   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
310     if (GV->hasSection()) {
311       StringRef SectionName = GV->getSection();
312       // Check if the global is in the PGO counters section.
313       auto OF = Triple(M->getTargetTriple()).getObjectFormat();
314       if (SectionName.endswith(
315               getInstrProfSectionName(IPSK_cnts, OF, /*AddSegmentInfo=*/false)))
316         return false;
317     }
318 
319     // Check if the global is private gcov data.
320     if (GV->getName().startswith("__llvm_gcov") ||
321         GV->getName().startswith("__llvm_gcda"))
322       return false;
323   }
324 
325   // Do not instrument acesses from different address spaces; we cannot deal
326   // with them.
327   if (Addr) {
328     Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
329     if (PtrTy->getPointerAddressSpace() != 0)
330       return false;
331   }
332 
333   return true;
334 }
335 
336 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
337   // If this is a GEP, just analyze its pointer operand.
338   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
339     Addr = GEP->getPointerOperand();
340 
341   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
342     if (GV->isConstant()) {
343       // Reads from constant globals can not race with any writes.
344       NumOmittedReadsFromConstantGlobals++;
345       return true;
346     }
347   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
348     if (isVtableAccess(L)) {
349       // Reads from a vtable pointer can not race with any writes.
350       NumOmittedReadsFromVtable++;
351       return true;
352     }
353   }
354   return false;
355 }
356 
357 // Instrumenting some of the accesses may be proven redundant.
358 // Currently handled:
359 //  - read-before-write (within same BB, no calls between)
360 //  - not captured variables
361 //
362 // We do not handle some of the patterns that should not survive
363 // after the classic compiler optimizations.
364 // E.g. two reads from the same temp should be eliminated by CSE,
365 // two writes should be eliminated by DSE, etc.
366 //
367 // 'Local' is a vector of insns within the same BB (no calls between).
368 // 'All' is a vector of insns that will be instrumented.
369 void ThreadSanitizer::chooseInstructionsToInstrument(
370     SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
371     const DataLayout &DL) {
372   SmallPtrSet<Value*, 8> WriteTargets;
373   // Iterate from the end.
374   for (Instruction *I : reverse(Local)) {
375     if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
376       Value *Addr = Store->getPointerOperand();
377       if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
378         continue;
379       WriteTargets.insert(Addr);
380     } else {
381       LoadInst *Load = cast<LoadInst>(I);
382       Value *Addr = Load->getPointerOperand();
383       if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
384         continue;
385       if (WriteTargets.count(Addr)) {
386         // We will write to this temp, so no reason to analyze the read.
387         NumOmittedReadsBeforeWrite++;
388         continue;
389       }
390       if (addrPointsToConstantData(Addr)) {
391         // Addr points to some constant data -- it can not race with any writes.
392         continue;
393       }
394     }
395     Value *Addr = isa<StoreInst>(*I)
396         ? cast<StoreInst>(I)->getPointerOperand()
397         : cast<LoadInst>(I)->getPointerOperand();
398     if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) &&
399         !PointerMayBeCaptured(Addr, true, true)) {
400       // The variable is addressable but not captured, so it cannot be
401       // referenced from a different thread and participate in a data race
402       // (see llvm/Analysis/CaptureTracking.h for details).
403       NumOmittedNonCaptured++;
404       continue;
405     }
406     All.push_back(I);
407   }
408   Local.clear();
409 }
410 
411 static bool isAtomic(Instruction *I) {
412   // TODO: Ask TTI whether synchronization scope is between threads.
413   if (LoadInst *LI = dyn_cast<LoadInst>(I))
414     return LI->isAtomic() && LI->getSyncScopeID() != SyncScope::SingleThread;
415   if (StoreInst *SI = dyn_cast<StoreInst>(I))
416     return SI->isAtomic() && SI->getSyncScopeID() != SyncScope::SingleThread;
417   if (isa<AtomicRMWInst>(I))
418     return true;
419   if (isa<AtomicCmpXchgInst>(I))
420     return true;
421   if (isa<FenceInst>(I))
422     return true;
423   return false;
424 }
425 
426 void ThreadSanitizer::InsertRuntimeIgnores(Function &F) {
427   IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
428   IRB.CreateCall(TsanIgnoreBegin);
429   EscapeEnumerator EE(F, "tsan_ignore_cleanup", ClHandleCxxExceptions);
430   while (IRBuilder<> *AtExit = EE.Next()) {
431     AtExit->CreateCall(TsanIgnoreEnd);
432   }
433 }
434 
435 bool ThreadSanitizer::sanitizeFunction(Function &F,
436                                        const TargetLibraryInfo &TLI) {
437   // This is required to prevent instrumenting call to __tsan_init from within
438   // the module constructor.
439   if (&F == TsanCtorFunction)
440     return false;
441   initializeCallbacks(*F.getParent());
442   SmallVector<Instruction*, 8> AllLoadsAndStores;
443   SmallVector<Instruction*, 8> LocalLoadsAndStores;
444   SmallVector<Instruction*, 8> AtomicAccesses;
445   SmallVector<Instruction*, 8> MemIntrinCalls;
446   bool Res = false;
447   bool HasCalls = false;
448   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
449   const DataLayout &DL = F.getParent()->getDataLayout();
450 
451   // Traverse all instructions, collect loads/stores/returns, check for calls.
452   for (auto &BB : F) {
453     for (auto &Inst : BB) {
454       if (isAtomic(&Inst))
455         AtomicAccesses.push_back(&Inst);
456       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
457         LocalLoadsAndStores.push_back(&Inst);
458       else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
459         if (CallInst *CI = dyn_cast<CallInst>(&Inst))
460           maybeMarkSanitizerLibraryCallNoBuiltin(CI, &TLI);
461         if (isa<MemIntrinsic>(Inst))
462           MemIntrinCalls.push_back(&Inst);
463         HasCalls = true;
464         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
465                                        DL);
466       }
467     }
468     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
469   }
470 
471   // We have collected all loads and stores.
472   // FIXME: many of these accesses do not need to be checked for races
473   // (e.g. variables that do not escape, etc).
474 
475   // Instrument memory accesses only if we want to report bugs in the function.
476   if (ClInstrumentMemoryAccesses && SanitizeFunction)
477     for (auto Inst : AllLoadsAndStores) {
478       Res |= instrumentLoadOrStore(Inst, DL);
479     }
480 
481   // Instrument atomic memory accesses in any case (they can be used to
482   // implement synchronization).
483   if (ClInstrumentAtomics)
484     for (auto Inst : AtomicAccesses) {
485       Res |= instrumentAtomic(Inst, DL);
486     }
487 
488   if (ClInstrumentMemIntrinsics && SanitizeFunction)
489     for (auto Inst : MemIntrinCalls) {
490       Res |= instrumentMemIntrinsic(Inst);
491     }
492 
493   if (F.hasFnAttribute("sanitize_thread_no_checking_at_run_time")) {
494     assert(!F.hasFnAttribute(Attribute::SanitizeThread));
495     if (HasCalls)
496       InsertRuntimeIgnores(F);
497   }
498 
499   // Instrument function entry/exit points if there were instrumented accesses.
500   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
501     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
502     Value *ReturnAddress = IRB.CreateCall(
503         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
504         IRB.getInt32(0));
505     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
506 
507     EscapeEnumerator EE(F, "tsan_cleanup", ClHandleCxxExceptions);
508     while (IRBuilder<> *AtExit = EE.Next()) {
509       AtExit->CreateCall(TsanFuncExit, {});
510     }
511     Res = true;
512   }
513   return Res;
514 }
515 
516 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I,
517                                             const DataLayout &DL) {
518   IRBuilder<> IRB(I);
519   bool IsWrite = isa<StoreInst>(*I);
520   Value *Addr = IsWrite
521       ? cast<StoreInst>(I)->getPointerOperand()
522       : cast<LoadInst>(I)->getPointerOperand();
523 
524   // swifterror memory addresses are mem2reg promoted by instruction selection.
525   // As such they cannot have regular uses like an instrumentation function and
526   // it makes no sense to track them as memory.
527   if (Addr->isSwiftError())
528     return false;
529 
530   int Idx = getMemoryAccessFuncIndex(Addr, DL);
531   if (Idx < 0)
532     return false;
533   if (IsWrite && isVtableAccess(I)) {
534     LLVM_DEBUG(dbgs() << "  VPTR : " << *I << "\n");
535     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
536     // StoredValue may be a vector type if we are storing several vptrs at once.
537     // In this case, just take the first element of the vector since this is
538     // enough to find vptr races.
539     if (isa<VectorType>(StoredValue->getType()))
540       StoredValue = IRB.CreateExtractElement(
541           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
542     if (StoredValue->getType()->isIntegerTy())
543       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
544     // Call TsanVptrUpdate.
545     IRB.CreateCall(TsanVptrUpdate,
546                    {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
547                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
548     NumInstrumentedVtableWrites++;
549     return true;
550   }
551   if (!IsWrite && isVtableAccess(I)) {
552     IRB.CreateCall(TsanVptrLoad,
553                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
554     NumInstrumentedVtableReads++;
555     return true;
556   }
557   const unsigned Alignment = IsWrite
558       ? cast<StoreInst>(I)->getAlignment()
559       : cast<LoadInst>(I)->getAlignment();
560   Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
561   const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
562   FunctionCallee OnAccessFunc = nullptr;
563   if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0)
564     OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
565   else
566     OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
567   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
568   if (IsWrite) NumInstrumentedWrites++;
569   else         NumInstrumentedReads++;
570   return true;
571 }
572 
573 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
574   uint32_t v = 0;
575   switch (ord) {
576     case AtomicOrdering::NotAtomic:
577       llvm_unreachable("unexpected atomic ordering!");
578     case AtomicOrdering::Unordered:              LLVM_FALLTHROUGH;
579     case AtomicOrdering::Monotonic:              v = 0; break;
580     // Not specified yet:
581     // case AtomicOrdering::Consume:                v = 1; break;
582     case AtomicOrdering::Acquire:                v = 2; break;
583     case AtomicOrdering::Release:                v = 3; break;
584     case AtomicOrdering::AcquireRelease:         v = 4; break;
585     case AtomicOrdering::SequentiallyConsistent: v = 5; break;
586   }
587   return IRB->getInt32(v);
588 }
589 
590 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
591 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
592 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
593 // instead we simply replace them with regular function calls, which are then
594 // intercepted by the run-time.
595 // Since tsan is running after everyone else, the calls should not be
596 // replaced back with intrinsics. If that becomes wrong at some point,
597 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
598 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
599   IRBuilder<> IRB(I);
600   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
601     IRB.CreateCall(
602         MemsetFn,
603         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
604          IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
605          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
606     I->eraseFromParent();
607   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
608     IRB.CreateCall(
609         isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
610         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
611          IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
612          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
613     I->eraseFromParent();
614   }
615   return false;
616 }
617 
618 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
619 // standards.  For background see C++11 standard.  A slightly older, publicly
620 // available draft of the standard (not entirely up-to-date, but close enough
621 // for casual browsing) is available here:
622 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
623 // The following page contains more background information:
624 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
625 
626 bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
627   IRBuilder<> IRB(I);
628   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
629     Value *Addr = LI->getPointerOperand();
630     int Idx = getMemoryAccessFuncIndex(Addr, DL);
631     if (Idx < 0)
632       return false;
633     const unsigned ByteSize = 1U << Idx;
634     const unsigned BitSize = ByteSize * 8;
635     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
636     Type *PtrTy = Ty->getPointerTo();
637     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
638                      createOrdering(&IRB, LI->getOrdering())};
639     Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
640     Value *C = IRB.CreateCall(TsanAtomicLoad[Idx], Args);
641     Value *Cast = IRB.CreateBitOrPointerCast(C, OrigTy);
642     I->replaceAllUsesWith(Cast);
643   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
644     Value *Addr = SI->getPointerOperand();
645     int Idx = getMemoryAccessFuncIndex(Addr, DL);
646     if (Idx < 0)
647       return false;
648     const unsigned ByteSize = 1U << Idx;
649     const unsigned BitSize = ByteSize * 8;
650     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
651     Type *PtrTy = Ty->getPointerTo();
652     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
653                      IRB.CreateBitOrPointerCast(SI->getValueOperand(), Ty),
654                      createOrdering(&IRB, SI->getOrdering())};
655     CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
656     ReplaceInstWithInst(I, C);
657   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
658     Value *Addr = RMWI->getPointerOperand();
659     int Idx = getMemoryAccessFuncIndex(Addr, DL);
660     if (Idx < 0)
661       return false;
662     FunctionCallee F = TsanAtomicRMW[RMWI->getOperation()][Idx];
663     if (!F)
664       return false;
665     const unsigned ByteSize = 1U << Idx;
666     const unsigned BitSize = ByteSize * 8;
667     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
668     Type *PtrTy = Ty->getPointerTo();
669     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
670                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
671                      createOrdering(&IRB, RMWI->getOrdering())};
672     CallInst *C = CallInst::Create(F, Args);
673     ReplaceInstWithInst(I, C);
674   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
675     Value *Addr = CASI->getPointerOperand();
676     int Idx = getMemoryAccessFuncIndex(Addr, DL);
677     if (Idx < 0)
678       return false;
679     const unsigned ByteSize = 1U << Idx;
680     const unsigned BitSize = ByteSize * 8;
681     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
682     Type *PtrTy = Ty->getPointerTo();
683     Value *CmpOperand =
684       IRB.CreateBitOrPointerCast(CASI->getCompareOperand(), Ty);
685     Value *NewOperand =
686       IRB.CreateBitOrPointerCast(CASI->getNewValOperand(), Ty);
687     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
688                      CmpOperand,
689                      NewOperand,
690                      createOrdering(&IRB, CASI->getSuccessOrdering()),
691                      createOrdering(&IRB, CASI->getFailureOrdering())};
692     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
693     Value *Success = IRB.CreateICmpEQ(C, CmpOperand);
694     Value *OldVal = C;
695     Type *OrigOldValTy = CASI->getNewValOperand()->getType();
696     if (Ty != OrigOldValTy) {
697       // The value is a pointer, so we need to cast the return value.
698       OldVal = IRB.CreateIntToPtr(C, OrigOldValTy);
699     }
700 
701     Value *Res =
702       IRB.CreateInsertValue(UndefValue::get(CASI->getType()), OldVal, 0);
703     Res = IRB.CreateInsertValue(Res, Success, 1);
704 
705     I->replaceAllUsesWith(Res);
706     I->eraseFromParent();
707   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
708     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
709     FunctionCallee F = FI->getSyncScopeID() == SyncScope::SingleThread
710                            ? TsanAtomicSignalFence
711                            : TsanAtomicThreadFence;
712     CallInst *C = CallInst::Create(F, Args);
713     ReplaceInstWithInst(I, C);
714   }
715   return true;
716 }
717 
718 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr,
719                                               const DataLayout &DL) {
720   Type *OrigPtrTy = Addr->getType();
721   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
722   assert(OrigTy->isSized());
723   uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
724   if (TypeSize != 8  && TypeSize != 16 &&
725       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
726     NumAccessesWithBadSize++;
727     // Ignore all unusual sizes.
728     return -1;
729   }
730   size_t Idx = countTrailingZeros(TypeSize / 8);
731   assert(Idx < kNumberOfAccessSizes);
732   return Idx;
733 }
734