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/DenseMap.h"
23 #include "llvm/ADT/Optional.h"
24 #include "llvm/ADT/SmallString.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/StringExtras.h"
28 #include "llvm/Analysis/CaptureTracking.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/Intrinsics.h"
37 #include "llvm/IR/LLVMContext.h"
38 #include "llvm/IR/Metadata.h"
39 #include "llvm/IR/Module.h"
40 #include "llvm/IR/Type.h"
41 #include "llvm/ProfileData/InstrProf.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Transforms/Instrumentation.h"
47 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
48 #include "llvm/Transforms/Utils/EscapeEnumerator.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Transforms/Utils/ModuleUtils.h"
51 
52 using namespace llvm;
53 
54 #define DEBUG_TYPE "tsan"
55 
56 static cl::opt<bool> ClInstrumentMemoryAccesses(
57     "tsan-instrument-memory-accesses", cl::init(true),
58     cl::desc("Instrument memory accesses"), cl::Hidden);
59 static cl::opt<bool>
60     ClInstrumentFuncEntryExit("tsan-instrument-func-entry-exit", cl::init(true),
61                               cl::desc("Instrument function entry and exit"),
62                               cl::Hidden);
63 static cl::opt<bool> ClHandleCxxExceptions(
64     "tsan-handle-cxx-exceptions", cl::init(true),
65     cl::desc("Handle C++ exceptions (insert cleanup blocks for unwinding)"),
66     cl::Hidden);
67 static cl::opt<bool> ClInstrumentAtomics("tsan-instrument-atomics",
68                                          cl::init(true),
69                                          cl::desc("Instrument atomics"),
70                                          cl::Hidden);
71 static cl::opt<bool> ClInstrumentMemIntrinsics(
72     "tsan-instrument-memintrinsics", cl::init(true),
73     cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
74 static cl::opt<bool> ClDistinguishVolatile(
75     "tsan-distinguish-volatile", cl::init(false),
76     cl::desc("Emit special instrumentation for accesses to volatiles"),
77     cl::Hidden);
78 static cl::opt<bool> ClInstrumentReadBeforeWrite(
79     "tsan-instrument-read-before-write", cl::init(false),
80     cl::desc("Do not eliminate read instrumentation for read-before-writes"),
81     cl::Hidden);
82 static cl::opt<bool> ClCompoundReadBeforeWrite(
83     "tsan-compound-read-before-write", cl::init(false),
84     cl::desc("Emit special compound instrumentation for reads-before-writes"),
85     cl::Hidden);
86 
87 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
88 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
89 STATISTIC(NumOmittedReadsBeforeWrite,
90           "Number of reads ignored due to following writes");
91 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
92 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
93 STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
94 STATISTIC(NumOmittedReadsFromConstantGlobals,
95           "Number of reads from constant globals");
96 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
97 STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");
98 
99 const char kTsanModuleCtorName[] = "tsan.module_ctor";
100 const char kTsanInitName[] = "__tsan_init";
101 
102 namespace {
103 
104 /// ThreadSanitizer: instrument the code in module to find races.
105 ///
106 /// Instantiating ThreadSanitizer inserts the tsan runtime library API function
107 /// declarations into the module if they don't exist already. Instantiating
108 /// ensures the __tsan_init function is in the list of global constructors for
109 /// the module.
110 struct ThreadSanitizer {
111   ThreadSanitizer() {
112     // Check options and warn user.
113     if (ClInstrumentReadBeforeWrite && ClCompoundReadBeforeWrite) {
114       errs()
115           << "warning: Option -tsan-compound-read-before-write has no effect "
116              "when -tsan-instrument-read-before-write is set.\n";
117     }
118   }
119 
120   bool sanitizeFunction(Function &F, const TargetLibraryInfo &TLI);
121 
122 private:
123   // Internal Instruction wrapper that contains more information about the
124   // Instruction from prior analysis.
125   struct InstructionInfo {
126     // Instrumentation emitted for this instruction is for a compounded set of
127     // read and write operations in the same basic block.
128     static constexpr unsigned kCompoundRW = (1U << 0);
129 
130     explicit InstructionInfo(Instruction *Inst) : Inst(Inst) {}
131 
132     Instruction *Inst;
133     unsigned Flags = 0;
134   };
135 
136   void initialize(Module &M);
137   bool instrumentLoadOrStore(const InstructionInfo &II, const DataLayout &DL);
138   bool instrumentAtomic(Instruction *I, const DataLayout &DL);
139   bool instrumentMemIntrinsic(Instruction *I);
140   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
141                                       SmallVectorImpl<InstructionInfo> &All,
142                                       const DataLayout &DL);
143   bool addrPointsToConstantData(Value *Addr);
144   int getMemoryAccessFuncIndex(Type *OrigTy, Value *Addr, const DataLayout &DL);
145   void InsertRuntimeIgnores(Function &F);
146 
147   Type *IntptrTy;
148   FunctionCallee TsanFuncEntry;
149   FunctionCallee TsanFuncExit;
150   FunctionCallee TsanIgnoreBegin;
151   FunctionCallee TsanIgnoreEnd;
152   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
153   static const size_t kNumberOfAccessSizes = 5;
154   FunctionCallee TsanRead[kNumberOfAccessSizes];
155   FunctionCallee TsanWrite[kNumberOfAccessSizes];
156   FunctionCallee TsanUnalignedRead[kNumberOfAccessSizes];
157   FunctionCallee TsanUnalignedWrite[kNumberOfAccessSizes];
158   FunctionCallee TsanVolatileRead[kNumberOfAccessSizes];
159   FunctionCallee TsanVolatileWrite[kNumberOfAccessSizes];
160   FunctionCallee TsanUnalignedVolatileRead[kNumberOfAccessSizes];
161   FunctionCallee TsanUnalignedVolatileWrite[kNumberOfAccessSizes];
162   FunctionCallee TsanCompoundRW[kNumberOfAccessSizes];
163   FunctionCallee TsanUnalignedCompoundRW[kNumberOfAccessSizes];
164   FunctionCallee TsanAtomicLoad[kNumberOfAccessSizes];
165   FunctionCallee TsanAtomicStore[kNumberOfAccessSizes];
166   FunctionCallee TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1]
167                               [kNumberOfAccessSizes];
168   FunctionCallee TsanAtomicCAS[kNumberOfAccessSizes];
169   FunctionCallee TsanAtomicThreadFence;
170   FunctionCallee TsanAtomicSignalFence;
171   FunctionCallee TsanVptrUpdate;
172   FunctionCallee TsanVptrLoad;
173   FunctionCallee MemmoveFn, MemcpyFn, MemsetFn;
174 };
175 
176 void insertModuleCtor(Module &M) {
177   getOrCreateSanitizerCtorAndInitFunctions(
178       M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{},
179       /*InitArgs=*/{},
180       // This callback is invoked when the functions are created the first
181       // time. Hook them into the global ctors list in that case:
182       [&](Function *Ctor, FunctionCallee) { appendToGlobalCtors(M, Ctor, 0); });
183 }
184 }  // namespace
185 
186 PreservedAnalyses ThreadSanitizerPass::run(Function &F,
187                                            FunctionAnalysisManager &FAM) {
188   ThreadSanitizer TSan;
189   if (TSan.sanitizeFunction(F, FAM.getResult<TargetLibraryAnalysis>(F)))
190     return PreservedAnalyses::none();
191   return PreservedAnalyses::all();
192 }
193 
194 PreservedAnalyses ModuleThreadSanitizerPass::run(Module &M,
195                                                  ModuleAnalysisManager &MAM) {
196   insertModuleCtor(M);
197   return PreservedAnalyses::none();
198 }
199 void ThreadSanitizer::initialize(Module &M) {
200   const DataLayout &DL = M.getDataLayout();
201   IntptrTy = DL.getIntPtrType(M.getContext());
202 
203   IRBuilder<> IRB(M.getContext());
204   AttributeList Attr;
205   Attr = Attr.addFnAttribute(M.getContext(), Attribute::NoUnwind);
206   // Initialize the callbacks.
207   TsanFuncEntry = M.getOrInsertFunction("__tsan_func_entry", Attr,
208                                         IRB.getVoidTy(), IRB.getInt8PtrTy());
209   TsanFuncExit =
210       M.getOrInsertFunction("__tsan_func_exit", Attr, IRB.getVoidTy());
211   TsanIgnoreBegin = M.getOrInsertFunction("__tsan_ignore_thread_begin", Attr,
212                                           IRB.getVoidTy());
213   TsanIgnoreEnd =
214       M.getOrInsertFunction("__tsan_ignore_thread_end", Attr, IRB.getVoidTy());
215   IntegerType *OrdTy = IRB.getInt32Ty();
216   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
217     const unsigned ByteSize = 1U << i;
218     const unsigned BitSize = ByteSize * 8;
219     std::string ByteSizeStr = utostr(ByteSize);
220     std::string BitSizeStr = utostr(BitSize);
221     SmallString<32> ReadName("__tsan_read" + ByteSizeStr);
222     TsanRead[i] = M.getOrInsertFunction(ReadName, Attr, IRB.getVoidTy(),
223                                         IRB.getInt8PtrTy());
224 
225     SmallString<32> WriteName("__tsan_write" + ByteSizeStr);
226     TsanWrite[i] = M.getOrInsertFunction(WriteName, Attr, IRB.getVoidTy(),
227                                          IRB.getInt8PtrTy());
228 
229     SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr);
230     TsanUnalignedRead[i] = M.getOrInsertFunction(
231         UnalignedReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
232 
233     SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr);
234     TsanUnalignedWrite[i] = M.getOrInsertFunction(
235         UnalignedWriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
236 
237     SmallString<64> VolatileReadName("__tsan_volatile_read" + ByteSizeStr);
238     TsanVolatileRead[i] = M.getOrInsertFunction(
239         VolatileReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
240 
241     SmallString<64> VolatileWriteName("__tsan_volatile_write" + ByteSizeStr);
242     TsanVolatileWrite[i] = M.getOrInsertFunction(
243         VolatileWriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
244 
245     SmallString<64> UnalignedVolatileReadName("__tsan_unaligned_volatile_read" +
246                                               ByteSizeStr);
247     TsanUnalignedVolatileRead[i] = M.getOrInsertFunction(
248         UnalignedVolatileReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
249 
250     SmallString<64> UnalignedVolatileWriteName(
251         "__tsan_unaligned_volatile_write" + ByteSizeStr);
252     TsanUnalignedVolatileWrite[i] = M.getOrInsertFunction(
253         UnalignedVolatileWriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
254 
255     SmallString<64> CompoundRWName("__tsan_read_write" + ByteSizeStr);
256     TsanCompoundRW[i] = M.getOrInsertFunction(
257         CompoundRWName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
258 
259     SmallString<64> UnalignedCompoundRWName("__tsan_unaligned_read_write" +
260                                             ByteSizeStr);
261     TsanUnalignedCompoundRW[i] = M.getOrInsertFunction(
262         UnalignedCompoundRWName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
263 
264     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
265     Type *PtrTy = Ty->getPointerTo();
266     SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load");
267     {
268       AttributeList AL = Attr;
269       AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
270       TsanAtomicLoad[i] =
271           M.getOrInsertFunction(AtomicLoadName, AL, Ty, PtrTy, OrdTy);
272     }
273 
274     SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store");
275     {
276       AttributeList AL = Attr;
277       AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
278       AL = AL.addParamAttribute(M.getContext(), 2, Attribute::ZExt);
279       TsanAtomicStore[i] = M.getOrInsertFunction(
280           AtomicStoreName, AL, IRB.getVoidTy(), PtrTy, Ty, OrdTy);
281     }
282 
283     for (unsigned Op = AtomicRMWInst::FIRST_BINOP;
284          Op <= AtomicRMWInst::LAST_BINOP; ++Op) {
285       TsanAtomicRMW[Op][i] = nullptr;
286       const char *NamePart = nullptr;
287       if (Op == AtomicRMWInst::Xchg)
288         NamePart = "_exchange";
289       else if (Op == AtomicRMWInst::Add)
290         NamePart = "_fetch_add";
291       else if (Op == AtomicRMWInst::Sub)
292         NamePart = "_fetch_sub";
293       else if (Op == AtomicRMWInst::And)
294         NamePart = "_fetch_and";
295       else if (Op == AtomicRMWInst::Or)
296         NamePart = "_fetch_or";
297       else if (Op == AtomicRMWInst::Xor)
298         NamePart = "_fetch_xor";
299       else if (Op == AtomicRMWInst::Nand)
300         NamePart = "_fetch_nand";
301       else
302         continue;
303       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
304       {
305         AttributeList AL = Attr;
306         AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
307         AL = AL.addParamAttribute(M.getContext(), 2, Attribute::ZExt);
308         TsanAtomicRMW[Op][i] =
309             M.getOrInsertFunction(RMWName, AL, Ty, PtrTy, Ty, OrdTy);
310       }
311     }
312 
313     SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr +
314                                   "_compare_exchange_val");
315     {
316       AttributeList AL = Attr;
317       AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
318       AL = AL.addParamAttribute(M.getContext(), 2, Attribute::ZExt);
319       AL = AL.addParamAttribute(M.getContext(), 3, Attribute::ZExt);
320       AL = AL.addParamAttribute(M.getContext(), 4, Attribute::ZExt);
321       TsanAtomicCAS[i] = M.getOrInsertFunction(AtomicCASName, AL, Ty, PtrTy, Ty,
322                                                Ty, OrdTy, OrdTy);
323     }
324   }
325   TsanVptrUpdate =
326       M.getOrInsertFunction("__tsan_vptr_update", Attr, IRB.getVoidTy(),
327                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy());
328   TsanVptrLoad = M.getOrInsertFunction("__tsan_vptr_read", Attr,
329                                        IRB.getVoidTy(), IRB.getInt8PtrTy());
330   {
331     AttributeList AL = Attr;
332     AL = AL.addParamAttribute(M.getContext(), 0, Attribute::ZExt);
333     TsanAtomicThreadFence = M.getOrInsertFunction("__tsan_atomic_thread_fence",
334                                                   AL, IRB.getVoidTy(), OrdTy);
335   }
336   {
337     AttributeList AL = Attr;
338     AL = AL.addParamAttribute(M.getContext(), 0, Attribute::ZExt);
339     TsanAtomicSignalFence = M.getOrInsertFunction("__tsan_atomic_signal_fence",
340                                                   AL, IRB.getVoidTy(), OrdTy);
341   }
342 
343   MemmoveFn =
344       M.getOrInsertFunction("memmove", Attr, IRB.getInt8PtrTy(),
345                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy);
346   MemcpyFn =
347       M.getOrInsertFunction("memcpy", Attr, IRB.getInt8PtrTy(),
348                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy);
349   MemsetFn =
350       M.getOrInsertFunction("memset", Attr, IRB.getInt8PtrTy(),
351                             IRB.getInt8PtrTy(), IRB.getInt32Ty(), IntptrTy);
352 }
353 
354 static bool isVtableAccess(Instruction *I) {
355   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
356     return Tag->isTBAAVtableAccess();
357   return false;
358 }
359 
360 // Do not instrument known races/"benign races" that come from compiler
361 // instrumentatin. The user has no way of suppressing them.
362 static bool shouldInstrumentReadWriteFromAddress(const Module *M, Value *Addr) {
363   // Peel off GEPs and BitCasts.
364   Addr = Addr->stripInBoundsOffsets();
365 
366   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
367     if (GV->hasSection()) {
368       StringRef SectionName = GV->getSection();
369       // Check if the global is in the PGO counters section.
370       auto OF = Triple(M->getTargetTriple()).getObjectFormat();
371       if (SectionName.endswith(
372               getInstrProfSectionName(IPSK_cnts, OF, /*AddSegmentInfo=*/false)))
373         return false;
374     }
375 
376     // Check if the global is private gcov data.
377     if (GV->getName().startswith("__llvm_gcov") ||
378         GV->getName().startswith("__llvm_gcda"))
379       return false;
380   }
381 
382   // Do not instrument acesses from different address spaces; we cannot deal
383   // with them.
384   if (Addr) {
385     Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
386     if (PtrTy->getPointerAddressSpace() != 0)
387       return false;
388   }
389 
390   return true;
391 }
392 
393 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
394   // If this is a GEP, just analyze its pointer operand.
395   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
396     Addr = GEP->getPointerOperand();
397 
398   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
399     if (GV->isConstant()) {
400       // Reads from constant globals can not race with any writes.
401       NumOmittedReadsFromConstantGlobals++;
402       return true;
403     }
404   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
405     if (isVtableAccess(L)) {
406       // Reads from a vtable pointer can not race with any writes.
407       NumOmittedReadsFromVtable++;
408       return true;
409     }
410   }
411   return false;
412 }
413 
414 // Instrumenting some of the accesses may be proven redundant.
415 // Currently handled:
416 //  - read-before-write (within same BB, no calls between)
417 //  - not captured variables
418 //
419 // We do not handle some of the patterns that should not survive
420 // after the classic compiler optimizations.
421 // E.g. two reads from the same temp should be eliminated by CSE,
422 // two writes should be eliminated by DSE, etc.
423 //
424 // 'Local' is a vector of insns within the same BB (no calls between).
425 // 'All' is a vector of insns that will be instrumented.
426 void ThreadSanitizer::chooseInstructionsToInstrument(
427     SmallVectorImpl<Instruction *> &Local,
428     SmallVectorImpl<InstructionInfo> &All, const DataLayout &DL) {
429   DenseMap<Value *, size_t> WriteTargets; // Map of addresses to index in All
430   // Iterate from the end.
431   for (Instruction *I : reverse(Local)) {
432     const bool IsWrite = isa<StoreInst>(*I);
433     Value *Addr = IsWrite ? cast<StoreInst>(I)->getPointerOperand()
434                           : cast<LoadInst>(I)->getPointerOperand();
435 
436     if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
437       continue;
438 
439     if (!IsWrite) {
440       const auto WriteEntry = WriteTargets.find(Addr);
441       if (!ClInstrumentReadBeforeWrite && WriteEntry != WriteTargets.end()) {
442         auto &WI = All[WriteEntry->second];
443         // If we distinguish volatile accesses and if either the read or write
444         // is volatile, do not omit any instrumentation.
445         const bool AnyVolatile =
446             ClDistinguishVolatile && (cast<LoadInst>(I)->isVolatile() ||
447                                       cast<StoreInst>(WI.Inst)->isVolatile());
448         if (!AnyVolatile) {
449           // We will write to this temp, so no reason to analyze the read.
450           // Mark the write instruction as compound.
451           WI.Flags |= InstructionInfo::kCompoundRW;
452           NumOmittedReadsBeforeWrite++;
453           continue;
454         }
455       }
456 
457       if (addrPointsToConstantData(Addr)) {
458         // Addr points to some constant data -- it can not race with any writes.
459         continue;
460       }
461     }
462 
463     if (isa<AllocaInst>(getUnderlyingObject(Addr)) &&
464         !PointerMayBeCaptured(Addr, true, true)) {
465       // The variable is addressable but not captured, so it cannot be
466       // referenced from a different thread and participate in a data race
467       // (see llvm/Analysis/CaptureTracking.h for details).
468       NumOmittedNonCaptured++;
469       continue;
470     }
471 
472     // Instrument this instruction.
473     All.emplace_back(I);
474     if (IsWrite) {
475       // For read-before-write and compound instrumentation we only need one
476       // write target, and we can override any previous entry if it exists.
477       WriteTargets[Addr] = All.size() - 1;
478     }
479   }
480   Local.clear();
481 }
482 
483 static bool isTsanAtomic(const Instruction *I) {
484   // TODO: Ask TTI whether synchronization scope is between threads.
485   auto SSID = getAtomicSyncScopeID(I);
486   if (!SSID)
487     return false;
488   if (isa<LoadInst>(I) || isa<StoreInst>(I))
489     return SSID.value() != SyncScope::SingleThread;
490   return true;
491 }
492 
493 void ThreadSanitizer::InsertRuntimeIgnores(Function &F) {
494   InstrumentationIRBuilder IRB(F.getEntryBlock().getFirstNonPHI());
495   IRB.CreateCall(TsanIgnoreBegin);
496   EscapeEnumerator EE(F, "tsan_ignore_cleanup", ClHandleCxxExceptions);
497   while (IRBuilder<> *AtExit = EE.Next()) {
498     InstrumentationIRBuilder::ensureDebugInfo(*AtExit, F);
499     AtExit->CreateCall(TsanIgnoreEnd);
500   }
501 }
502 
503 bool ThreadSanitizer::sanitizeFunction(Function &F,
504                                        const TargetLibraryInfo &TLI) {
505   // This is required to prevent instrumenting call to __tsan_init from within
506   // the module constructor.
507   if (F.getName() == kTsanModuleCtorName)
508     return false;
509   // Naked functions can not have prologue/epilogue
510   // (__tsan_func_entry/__tsan_func_exit) generated, so don't instrument them at
511   // all.
512   if (F.hasFnAttribute(Attribute::Naked))
513     return false;
514 
515   // __attribute__(disable_sanitizer_instrumentation) prevents all kinds of
516   // instrumentation.
517   if (F.hasFnAttribute(Attribute::DisableSanitizerInstrumentation))
518     return false;
519 
520   initialize(*F.getParent());
521   SmallVector<InstructionInfo, 8> AllLoadsAndStores;
522   SmallVector<Instruction*, 8> LocalLoadsAndStores;
523   SmallVector<Instruction*, 8> AtomicAccesses;
524   SmallVector<Instruction*, 8> MemIntrinCalls;
525   bool Res = false;
526   bool HasCalls = false;
527   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
528   const DataLayout &DL = F.getParent()->getDataLayout();
529 
530   // Traverse all instructions, collect loads/stores/returns, check for calls.
531   for (auto &BB : F) {
532     for (auto &Inst : BB) {
533       if (isTsanAtomic(&Inst))
534         AtomicAccesses.push_back(&Inst);
535       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
536         LocalLoadsAndStores.push_back(&Inst);
537       else if ((isa<CallInst>(Inst) && !isa<DbgInfoIntrinsic>(Inst)) ||
538                isa<InvokeInst>(Inst)) {
539         if (CallInst *CI = dyn_cast<CallInst>(&Inst))
540           maybeMarkSanitizerLibraryCallNoBuiltin(CI, &TLI);
541         if (isa<MemIntrinsic>(Inst))
542           MemIntrinCalls.push_back(&Inst);
543         HasCalls = true;
544         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
545                                        DL);
546       }
547     }
548     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
549   }
550 
551   // We have collected all loads and stores.
552   // FIXME: many of these accesses do not need to be checked for races
553   // (e.g. variables that do not escape, etc).
554 
555   // Instrument memory accesses only if we want to report bugs in the function.
556   if (ClInstrumentMemoryAccesses && SanitizeFunction)
557     for (const auto &II : AllLoadsAndStores) {
558       Res |= instrumentLoadOrStore(II, DL);
559     }
560 
561   // Instrument atomic memory accesses in any case (they can be used to
562   // implement synchronization).
563   if (ClInstrumentAtomics)
564     for (auto Inst : AtomicAccesses) {
565       Res |= instrumentAtomic(Inst, DL);
566     }
567 
568   if (ClInstrumentMemIntrinsics && SanitizeFunction)
569     for (auto Inst : MemIntrinCalls) {
570       Res |= instrumentMemIntrinsic(Inst);
571     }
572 
573   if (F.hasFnAttribute("sanitize_thread_no_checking_at_run_time")) {
574     assert(!F.hasFnAttribute(Attribute::SanitizeThread));
575     if (HasCalls)
576       InsertRuntimeIgnores(F);
577   }
578 
579   // Instrument function entry/exit points if there were instrumented accesses.
580   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
581     InstrumentationIRBuilder IRB(F.getEntryBlock().getFirstNonPHI());
582     Value *ReturnAddress = IRB.CreateCall(
583         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
584         IRB.getInt32(0));
585     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
586 
587     EscapeEnumerator EE(F, "tsan_cleanup", ClHandleCxxExceptions);
588     while (IRBuilder<> *AtExit = EE.Next()) {
589       InstrumentationIRBuilder::ensureDebugInfo(*AtExit, F);
590       AtExit->CreateCall(TsanFuncExit, {});
591     }
592     Res = true;
593   }
594   return Res;
595 }
596 
597 bool ThreadSanitizer::instrumentLoadOrStore(const InstructionInfo &II,
598                                             const DataLayout &DL) {
599   InstrumentationIRBuilder IRB(II.Inst);
600   const bool IsWrite = isa<StoreInst>(*II.Inst);
601   Value *Addr = IsWrite ? cast<StoreInst>(II.Inst)->getPointerOperand()
602                         : cast<LoadInst>(II.Inst)->getPointerOperand();
603   Type *OrigTy = getLoadStoreType(II.Inst);
604 
605   // swifterror memory addresses are mem2reg promoted by instruction selection.
606   // As such they cannot have regular uses like an instrumentation function and
607   // it makes no sense to track them as memory.
608   if (Addr->isSwiftError())
609     return false;
610 
611   int Idx = getMemoryAccessFuncIndex(OrigTy, Addr, DL);
612   if (Idx < 0)
613     return false;
614   if (IsWrite && isVtableAccess(II.Inst)) {
615     LLVM_DEBUG(dbgs() << "  VPTR : " << *II.Inst << "\n");
616     Value *StoredValue = cast<StoreInst>(II.Inst)->getValueOperand();
617     // StoredValue may be a vector type if we are storing several vptrs at once.
618     // In this case, just take the first element of the vector since this is
619     // enough to find vptr races.
620     if (isa<VectorType>(StoredValue->getType()))
621       StoredValue = IRB.CreateExtractElement(
622           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
623     if (StoredValue->getType()->isIntegerTy())
624       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
625     // Call TsanVptrUpdate.
626     IRB.CreateCall(TsanVptrUpdate,
627                    {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
628                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
629     NumInstrumentedVtableWrites++;
630     return true;
631   }
632   if (!IsWrite && isVtableAccess(II.Inst)) {
633     IRB.CreateCall(TsanVptrLoad,
634                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
635     NumInstrumentedVtableReads++;
636     return true;
637   }
638 
639   const Align Alignment = IsWrite ? cast<StoreInst>(II.Inst)->getAlign()
640                                   : cast<LoadInst>(II.Inst)->getAlign();
641   const bool IsCompoundRW =
642       ClCompoundReadBeforeWrite && (II.Flags & InstructionInfo::kCompoundRW);
643   const bool IsVolatile = ClDistinguishVolatile &&
644                           (IsWrite ? cast<StoreInst>(II.Inst)->isVolatile()
645                                    : cast<LoadInst>(II.Inst)->isVolatile());
646   assert((!IsVolatile || !IsCompoundRW) && "Compound volatile invalid!");
647 
648   const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
649   FunctionCallee OnAccessFunc = nullptr;
650   if (Alignment >= Align(8) || (Alignment.value() % (TypeSize / 8)) == 0) {
651     if (IsCompoundRW)
652       OnAccessFunc = TsanCompoundRW[Idx];
653     else if (IsVolatile)
654       OnAccessFunc = IsWrite ? TsanVolatileWrite[Idx] : TsanVolatileRead[Idx];
655     else
656       OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
657   } else {
658     if (IsCompoundRW)
659       OnAccessFunc = TsanUnalignedCompoundRW[Idx];
660     else if (IsVolatile)
661       OnAccessFunc = IsWrite ? TsanUnalignedVolatileWrite[Idx]
662                              : TsanUnalignedVolatileRead[Idx];
663     else
664       OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
665   }
666   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
667   if (IsCompoundRW || IsWrite)
668     NumInstrumentedWrites++;
669   if (IsCompoundRW || !IsWrite)
670     NumInstrumentedReads++;
671   return true;
672 }
673 
674 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
675   uint32_t v = 0;
676   switch (ord) {
677     case AtomicOrdering::NotAtomic:
678       llvm_unreachable("unexpected atomic ordering!");
679     case AtomicOrdering::Unordered:              LLVM_FALLTHROUGH;
680     case AtomicOrdering::Monotonic:              v = 0; break;
681     // Not specified yet:
682     // case AtomicOrdering::Consume:                v = 1; break;
683     case AtomicOrdering::Acquire:                v = 2; break;
684     case AtomicOrdering::Release:                v = 3; break;
685     case AtomicOrdering::AcquireRelease:         v = 4; break;
686     case AtomicOrdering::SequentiallyConsistent: v = 5; break;
687   }
688   return IRB->getInt32(v);
689 }
690 
691 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
692 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
693 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
694 // instead we simply replace them with regular function calls, which are then
695 // intercepted by the run-time.
696 // Since tsan is running after everyone else, the calls should not be
697 // replaced back with intrinsics. If that becomes wrong at some point,
698 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
699 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
700   IRBuilder<> IRB(I);
701   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
702     IRB.CreateCall(
703         MemsetFn,
704         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
705          IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
706          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
707     I->eraseFromParent();
708   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
709     IRB.CreateCall(
710         isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
711         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
712          IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
713          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
714     I->eraseFromParent();
715   }
716   return false;
717 }
718 
719 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
720 // standards.  For background see C++11 standard.  A slightly older, publicly
721 // available draft of the standard (not entirely up-to-date, but close enough
722 // for casual browsing) is available here:
723 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
724 // The following page contains more background information:
725 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
726 
727 bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
728   InstrumentationIRBuilder IRB(I);
729   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
730     Value *Addr = LI->getPointerOperand();
731     Type *OrigTy = LI->getType();
732     int Idx = getMemoryAccessFuncIndex(OrigTy, Addr, DL);
733     if (Idx < 0)
734       return false;
735     const unsigned ByteSize = 1U << Idx;
736     const unsigned BitSize = ByteSize * 8;
737     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
738     Type *PtrTy = Ty->getPointerTo();
739     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
740                      createOrdering(&IRB, LI->getOrdering())};
741     Value *C = IRB.CreateCall(TsanAtomicLoad[Idx], Args);
742     Value *Cast = IRB.CreateBitOrPointerCast(C, OrigTy);
743     I->replaceAllUsesWith(Cast);
744   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
745     Value *Addr = SI->getPointerOperand();
746     int Idx =
747         getMemoryAccessFuncIndex(SI->getValueOperand()->getType(), Addr, DL);
748     if (Idx < 0)
749       return false;
750     const unsigned ByteSize = 1U << Idx;
751     const unsigned BitSize = ByteSize * 8;
752     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
753     Type *PtrTy = Ty->getPointerTo();
754     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
755                      IRB.CreateBitOrPointerCast(SI->getValueOperand(), Ty),
756                      createOrdering(&IRB, SI->getOrdering())};
757     CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
758     ReplaceInstWithInst(I, C);
759   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
760     Value *Addr = RMWI->getPointerOperand();
761     int Idx =
762         getMemoryAccessFuncIndex(RMWI->getValOperand()->getType(), Addr, DL);
763     if (Idx < 0)
764       return false;
765     FunctionCallee F = TsanAtomicRMW[RMWI->getOperation()][Idx];
766     if (!F)
767       return false;
768     const unsigned ByteSize = 1U << Idx;
769     const unsigned BitSize = ByteSize * 8;
770     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
771     Type *PtrTy = Ty->getPointerTo();
772     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
773                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
774                      createOrdering(&IRB, RMWI->getOrdering())};
775     CallInst *C = CallInst::Create(F, Args);
776     ReplaceInstWithInst(I, C);
777   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
778     Value *Addr = CASI->getPointerOperand();
779     Type *OrigOldValTy = CASI->getNewValOperand()->getType();
780     int Idx = getMemoryAccessFuncIndex(OrigOldValTy, Addr, DL);
781     if (Idx < 0)
782       return false;
783     const unsigned ByteSize = 1U << Idx;
784     const unsigned BitSize = ByteSize * 8;
785     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
786     Type *PtrTy = Ty->getPointerTo();
787     Value *CmpOperand =
788       IRB.CreateBitOrPointerCast(CASI->getCompareOperand(), Ty);
789     Value *NewOperand =
790       IRB.CreateBitOrPointerCast(CASI->getNewValOperand(), Ty);
791     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
792                      CmpOperand,
793                      NewOperand,
794                      createOrdering(&IRB, CASI->getSuccessOrdering()),
795                      createOrdering(&IRB, CASI->getFailureOrdering())};
796     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
797     Value *Success = IRB.CreateICmpEQ(C, CmpOperand);
798     Value *OldVal = C;
799     if (Ty != OrigOldValTy) {
800       // The value is a pointer, so we need to cast the return value.
801       OldVal = IRB.CreateIntToPtr(C, OrigOldValTy);
802     }
803 
804     Value *Res =
805       IRB.CreateInsertValue(UndefValue::get(CASI->getType()), OldVal, 0);
806     Res = IRB.CreateInsertValue(Res, Success, 1);
807 
808     I->replaceAllUsesWith(Res);
809     I->eraseFromParent();
810   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
811     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
812     FunctionCallee F = FI->getSyncScopeID() == SyncScope::SingleThread
813                            ? TsanAtomicSignalFence
814                            : TsanAtomicThreadFence;
815     CallInst *C = CallInst::Create(F, Args);
816     ReplaceInstWithInst(I, C);
817   }
818   return true;
819 }
820 
821 int ThreadSanitizer::getMemoryAccessFuncIndex(Type *OrigTy, Value *Addr,
822                                               const DataLayout &DL) {
823   assert(OrigTy->isSized());
824   assert(
825       cast<PointerType>(Addr->getType())->isOpaqueOrPointeeTypeMatches(OrigTy));
826   uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
827   if (TypeSize != 8  && TypeSize != 16 &&
828       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
829     NumAccessesWithBadSize++;
830     // Ignore all unusual sizes.
831     return -1;
832   }
833   size_t Idx = countTrailingZeros(TypeSize / 8);
834   assert(Idx < kNumberOfAccessSizes);
835   return Idx;
836 }
837