xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Instrumentation/ThreadSanitizer.cpp (revision a7dea1671b87c07d2d266f836bfa8b58efc7c134)
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   bool sanitizeFunction(Function &F, const TargetLibraryInfo &TLI);
96 
97 private:
98   void initialize(Module &M);
99   bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL);
100   bool instrumentAtomic(Instruction *I, const DataLayout &DL);
101   bool instrumentMemIntrinsic(Instruction *I);
102   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
103                                       SmallVectorImpl<Instruction *> &All,
104                                       const DataLayout &DL);
105   bool addrPointsToConstantData(Value *Addr);
106   int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL);
107   void InsertRuntimeIgnores(Function &F);
108 
109   Type *IntptrTy;
110   FunctionCallee TsanFuncEntry;
111   FunctionCallee TsanFuncExit;
112   FunctionCallee TsanIgnoreBegin;
113   FunctionCallee TsanIgnoreEnd;
114   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
115   static const size_t kNumberOfAccessSizes = 5;
116   FunctionCallee TsanRead[kNumberOfAccessSizes];
117   FunctionCallee TsanWrite[kNumberOfAccessSizes];
118   FunctionCallee TsanUnalignedRead[kNumberOfAccessSizes];
119   FunctionCallee TsanUnalignedWrite[kNumberOfAccessSizes];
120   FunctionCallee TsanAtomicLoad[kNumberOfAccessSizes];
121   FunctionCallee TsanAtomicStore[kNumberOfAccessSizes];
122   FunctionCallee TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1]
123                               [kNumberOfAccessSizes];
124   FunctionCallee TsanAtomicCAS[kNumberOfAccessSizes];
125   FunctionCallee TsanAtomicThreadFence;
126   FunctionCallee TsanAtomicSignalFence;
127   FunctionCallee TsanVptrUpdate;
128   FunctionCallee TsanVptrLoad;
129   FunctionCallee MemmoveFn, MemcpyFn, MemsetFn;
130 };
131 
132 struct ThreadSanitizerLegacyPass : FunctionPass {
133   ThreadSanitizerLegacyPass() : FunctionPass(ID) {}
134   StringRef getPassName() const override;
135   void getAnalysisUsage(AnalysisUsage &AU) const override;
136   bool runOnFunction(Function &F) override;
137   bool doInitialization(Module &M) override;
138   static char ID; // Pass identification, replacement for typeid.
139 private:
140   Optional<ThreadSanitizer> TSan;
141 };
142 
143 void insertModuleCtor(Module &M) {
144   getOrCreateSanitizerCtorAndInitFunctions(
145       M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{},
146       /*InitArgs=*/{},
147       // This callback is invoked when the functions are created the first
148       // time. Hook them into the global ctors list in that case:
149       [&](Function *Ctor, FunctionCallee) { appendToGlobalCtors(M, Ctor, 0); });
150 }
151 
152 }  // namespace
153 
154 PreservedAnalyses ThreadSanitizerPass::run(Function &F,
155                                            FunctionAnalysisManager &FAM) {
156   ThreadSanitizer TSan;
157   if (TSan.sanitizeFunction(F, FAM.getResult<TargetLibraryAnalysis>(F)))
158     return PreservedAnalyses::none();
159   return PreservedAnalyses::all();
160 }
161 
162 PreservedAnalyses ThreadSanitizerPass::run(Module &M,
163                                            ModuleAnalysisManager &MAM) {
164   insertModuleCtor(M);
165   return PreservedAnalyses::none();
166 }
167 
168 char ThreadSanitizerLegacyPass::ID = 0;
169 INITIALIZE_PASS_BEGIN(ThreadSanitizerLegacyPass, "tsan",
170                       "ThreadSanitizer: detects data races.", false, false)
171 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
172 INITIALIZE_PASS_END(ThreadSanitizerLegacyPass, "tsan",
173                     "ThreadSanitizer: detects data races.", false, false)
174 
175 StringRef ThreadSanitizerLegacyPass::getPassName() const {
176   return "ThreadSanitizerLegacyPass";
177 }
178 
179 void ThreadSanitizerLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
180   AU.addRequired<TargetLibraryInfoWrapperPass>();
181 }
182 
183 bool ThreadSanitizerLegacyPass::doInitialization(Module &M) {
184   insertModuleCtor(M);
185   TSan.emplace();
186   return true;
187 }
188 
189 bool ThreadSanitizerLegacyPass::runOnFunction(Function &F) {
190   auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
191   TSan->sanitizeFunction(F, TLI);
192   return true;
193 }
194 
195 FunctionPass *llvm::createThreadSanitizerLegacyPassPass() {
196   return new ThreadSanitizerLegacyPass();
197 }
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.addAttribute(M.getContext(), AttributeList::FunctionIndex,
206                            Attribute::NoUnwind);
207   // Initialize the callbacks.
208   TsanFuncEntry = M.getOrInsertFunction("__tsan_func_entry", Attr,
209                                         IRB.getVoidTy(), IRB.getInt8PtrTy());
210   TsanFuncExit =
211       M.getOrInsertFunction("__tsan_func_exit", Attr, IRB.getVoidTy());
212   TsanIgnoreBegin = M.getOrInsertFunction("__tsan_ignore_thread_begin", Attr,
213                                           IRB.getVoidTy());
214   TsanIgnoreEnd =
215       M.getOrInsertFunction("__tsan_ignore_thread_end", Attr, IRB.getVoidTy());
216   IntegerType *OrdTy = IRB.getInt32Ty();
217   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
218     const unsigned ByteSize = 1U << i;
219     const unsigned BitSize = ByteSize * 8;
220     std::string ByteSizeStr = utostr(ByteSize);
221     std::string BitSizeStr = utostr(BitSize);
222     SmallString<32> ReadName("__tsan_read" + ByteSizeStr);
223     TsanRead[i] = M.getOrInsertFunction(ReadName, Attr, IRB.getVoidTy(),
224                                         IRB.getInt8PtrTy());
225 
226     SmallString<32> WriteName("__tsan_write" + ByteSizeStr);
227     TsanWrite[i] = M.getOrInsertFunction(WriteName, Attr, IRB.getVoidTy(),
228                                          IRB.getInt8PtrTy());
229 
230     SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr);
231     TsanUnalignedRead[i] = M.getOrInsertFunction(
232         UnalignedReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
233 
234     SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr);
235     TsanUnalignedWrite[i] = M.getOrInsertFunction(
236         UnalignedWriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
237 
238     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
239     Type *PtrTy = Ty->getPointerTo();
240     SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load");
241     TsanAtomicLoad[i] =
242         M.getOrInsertFunction(AtomicLoadName, Attr, Ty, PtrTy, OrdTy);
243 
244     SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store");
245     TsanAtomicStore[i] = M.getOrInsertFunction(
246         AtomicStoreName, Attr, IRB.getVoidTy(), PtrTy, Ty, OrdTy);
247 
248     for (int op = AtomicRMWInst::FIRST_BINOP;
249         op <= AtomicRMWInst::LAST_BINOP; ++op) {
250       TsanAtomicRMW[op][i] = nullptr;
251       const char *NamePart = nullptr;
252       if (op == AtomicRMWInst::Xchg)
253         NamePart = "_exchange";
254       else if (op == AtomicRMWInst::Add)
255         NamePart = "_fetch_add";
256       else if (op == AtomicRMWInst::Sub)
257         NamePart = "_fetch_sub";
258       else if (op == AtomicRMWInst::And)
259         NamePart = "_fetch_and";
260       else if (op == AtomicRMWInst::Or)
261         NamePart = "_fetch_or";
262       else if (op == AtomicRMWInst::Xor)
263         NamePart = "_fetch_xor";
264       else if (op == AtomicRMWInst::Nand)
265         NamePart = "_fetch_nand";
266       else
267         continue;
268       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
269       TsanAtomicRMW[op][i] =
270           M.getOrInsertFunction(RMWName, Attr, Ty, PtrTy, Ty, OrdTy);
271     }
272 
273     SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr +
274                                   "_compare_exchange_val");
275     TsanAtomicCAS[i] = M.getOrInsertFunction(AtomicCASName, Attr, Ty, PtrTy, Ty,
276                                              Ty, OrdTy, OrdTy);
277   }
278   TsanVptrUpdate =
279       M.getOrInsertFunction("__tsan_vptr_update", Attr, IRB.getVoidTy(),
280                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy());
281   TsanVptrLoad = M.getOrInsertFunction("__tsan_vptr_read", Attr,
282                                        IRB.getVoidTy(), IRB.getInt8PtrTy());
283   TsanAtomicThreadFence = M.getOrInsertFunction("__tsan_atomic_thread_fence",
284                                                 Attr, IRB.getVoidTy(), OrdTy);
285   TsanAtomicSignalFence = M.getOrInsertFunction("__tsan_atomic_signal_fence",
286                                                 Attr, IRB.getVoidTy(), OrdTy);
287 
288   MemmoveFn =
289       M.getOrInsertFunction("memmove", Attr, IRB.getInt8PtrTy(),
290                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy);
291   MemcpyFn =
292       M.getOrInsertFunction("memcpy", Attr, IRB.getInt8PtrTy(),
293                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy);
294   MemsetFn =
295       M.getOrInsertFunction("memset", Attr, IRB.getInt8PtrTy(),
296                             IRB.getInt8PtrTy(), IRB.getInt32Ty(), IntptrTy);
297 }
298 
299 static bool isVtableAccess(Instruction *I) {
300   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
301     return Tag->isTBAAVtableAccess();
302   return false;
303 }
304 
305 // Do not instrument known races/"benign races" that come from compiler
306 // instrumentatin. The user has no way of suppressing them.
307 static bool shouldInstrumentReadWriteFromAddress(const Module *M, Value *Addr) {
308   // Peel off GEPs and BitCasts.
309   Addr = Addr->stripInBoundsOffsets();
310 
311   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
312     if (GV->hasSection()) {
313       StringRef SectionName = GV->getSection();
314       // Check if the global is in the PGO counters section.
315       auto OF = Triple(M->getTargetTriple()).getObjectFormat();
316       if (SectionName.endswith(
317               getInstrProfSectionName(IPSK_cnts, OF, /*AddSegmentInfo=*/false)))
318         return false;
319     }
320 
321     // Check if the global is private gcov data.
322     if (GV->getName().startswith("__llvm_gcov") ||
323         GV->getName().startswith("__llvm_gcda"))
324       return false;
325   }
326 
327   // Do not instrument acesses from different address spaces; we cannot deal
328   // with them.
329   if (Addr) {
330     Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
331     if (PtrTy->getPointerAddressSpace() != 0)
332       return false;
333   }
334 
335   return true;
336 }
337 
338 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
339   // If this is a GEP, just analyze its pointer operand.
340   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
341     Addr = GEP->getPointerOperand();
342 
343   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
344     if (GV->isConstant()) {
345       // Reads from constant globals can not race with any writes.
346       NumOmittedReadsFromConstantGlobals++;
347       return true;
348     }
349   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
350     if (isVtableAccess(L)) {
351       // Reads from a vtable pointer can not race with any writes.
352       NumOmittedReadsFromVtable++;
353       return true;
354     }
355   }
356   return false;
357 }
358 
359 // Instrumenting some of the accesses may be proven redundant.
360 // Currently handled:
361 //  - read-before-write (within same BB, no calls between)
362 //  - not captured variables
363 //
364 // We do not handle some of the patterns that should not survive
365 // after the classic compiler optimizations.
366 // E.g. two reads from the same temp should be eliminated by CSE,
367 // two writes should be eliminated by DSE, etc.
368 //
369 // 'Local' is a vector of insns within the same BB (no calls between).
370 // 'All' is a vector of insns that will be instrumented.
371 void ThreadSanitizer::chooseInstructionsToInstrument(
372     SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
373     const DataLayout &DL) {
374   SmallPtrSet<Value*, 8> WriteTargets;
375   // Iterate from the end.
376   for (Instruction *I : reverse(Local)) {
377     if (StoreInst *Store = dyn_cast<StoreInst>(I)) {
378       Value *Addr = Store->getPointerOperand();
379       if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
380         continue;
381       WriteTargets.insert(Addr);
382     } else {
383       LoadInst *Load = cast<LoadInst>(I);
384       Value *Addr = Load->getPointerOperand();
385       if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
386         continue;
387       if (WriteTargets.count(Addr)) {
388         // We will write to this temp, so no reason to analyze the read.
389         NumOmittedReadsBeforeWrite++;
390         continue;
391       }
392       if (addrPointsToConstantData(Addr)) {
393         // Addr points to some constant data -- it can not race with any writes.
394         continue;
395       }
396     }
397     Value *Addr = isa<StoreInst>(*I)
398         ? cast<StoreInst>(I)->getPointerOperand()
399         : cast<LoadInst>(I)->getPointerOperand();
400     if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) &&
401         !PointerMayBeCaptured(Addr, true, true)) {
402       // The variable is addressable but not captured, so it cannot be
403       // referenced from a different thread and participate in a data race
404       // (see llvm/Analysis/CaptureTracking.h for details).
405       NumOmittedNonCaptured++;
406       continue;
407     }
408     All.push_back(I);
409   }
410   Local.clear();
411 }
412 
413 static bool isAtomic(Instruction *I) {
414   // TODO: Ask TTI whether synchronization scope is between threads.
415   if (LoadInst *LI = dyn_cast<LoadInst>(I))
416     return LI->isAtomic() && LI->getSyncScopeID() != SyncScope::SingleThread;
417   if (StoreInst *SI = dyn_cast<StoreInst>(I))
418     return SI->isAtomic() && SI->getSyncScopeID() != SyncScope::SingleThread;
419   if (isa<AtomicRMWInst>(I))
420     return true;
421   if (isa<AtomicCmpXchgInst>(I))
422     return true;
423   if (isa<FenceInst>(I))
424     return true;
425   return false;
426 }
427 
428 void ThreadSanitizer::InsertRuntimeIgnores(Function &F) {
429   IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
430   IRB.CreateCall(TsanIgnoreBegin);
431   EscapeEnumerator EE(F, "tsan_ignore_cleanup", ClHandleCxxExceptions);
432   while (IRBuilder<> *AtExit = EE.Next()) {
433     AtExit->CreateCall(TsanIgnoreEnd);
434   }
435 }
436 
437 bool ThreadSanitizer::sanitizeFunction(Function &F,
438                                        const TargetLibraryInfo &TLI) {
439   // This is required to prevent instrumenting call to __tsan_init from within
440   // the module constructor.
441   if (F.getName() == kTsanModuleCtorName)
442     return false;
443   initialize(*F.getParent());
444   SmallVector<Instruction*, 8> AllLoadsAndStores;
445   SmallVector<Instruction*, 8> LocalLoadsAndStores;
446   SmallVector<Instruction*, 8> AtomicAccesses;
447   SmallVector<Instruction*, 8> MemIntrinCalls;
448   bool Res = false;
449   bool HasCalls = false;
450   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
451   const DataLayout &DL = F.getParent()->getDataLayout();
452 
453   // Traverse all instructions, collect loads/stores/returns, check for calls.
454   for (auto &BB : F) {
455     for (auto &Inst : BB) {
456       if (isAtomic(&Inst))
457         AtomicAccesses.push_back(&Inst);
458       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
459         LocalLoadsAndStores.push_back(&Inst);
460       else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
461         if (CallInst *CI = dyn_cast<CallInst>(&Inst))
462           maybeMarkSanitizerLibraryCallNoBuiltin(CI, &TLI);
463         if (isa<MemIntrinsic>(Inst))
464           MemIntrinCalls.push_back(&Inst);
465         HasCalls = true;
466         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
467                                        DL);
468       }
469     }
470     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
471   }
472 
473   // We have collected all loads and stores.
474   // FIXME: many of these accesses do not need to be checked for races
475   // (e.g. variables that do not escape, etc).
476 
477   // Instrument memory accesses only if we want to report bugs in the function.
478   if (ClInstrumentMemoryAccesses && SanitizeFunction)
479     for (auto Inst : AllLoadsAndStores) {
480       Res |= instrumentLoadOrStore(Inst, DL);
481     }
482 
483   // Instrument atomic memory accesses in any case (they can be used to
484   // implement synchronization).
485   if (ClInstrumentAtomics)
486     for (auto Inst : AtomicAccesses) {
487       Res |= instrumentAtomic(Inst, DL);
488     }
489 
490   if (ClInstrumentMemIntrinsics && SanitizeFunction)
491     for (auto Inst : MemIntrinCalls) {
492       Res |= instrumentMemIntrinsic(Inst);
493     }
494 
495   if (F.hasFnAttribute("sanitize_thread_no_checking_at_run_time")) {
496     assert(!F.hasFnAttribute(Attribute::SanitizeThread));
497     if (HasCalls)
498       InsertRuntimeIgnores(F);
499   }
500 
501   // Instrument function entry/exit points if there were instrumented accesses.
502   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
503     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
504     Value *ReturnAddress = IRB.CreateCall(
505         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
506         IRB.getInt32(0));
507     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
508 
509     EscapeEnumerator EE(F, "tsan_cleanup", ClHandleCxxExceptions);
510     while (IRBuilder<> *AtExit = EE.Next()) {
511       AtExit->CreateCall(TsanFuncExit, {});
512     }
513     Res = true;
514   }
515   return Res;
516 }
517 
518 bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I,
519                                             const DataLayout &DL) {
520   IRBuilder<> IRB(I);
521   bool IsWrite = isa<StoreInst>(*I);
522   Value *Addr = IsWrite
523       ? cast<StoreInst>(I)->getPointerOperand()
524       : cast<LoadInst>(I)->getPointerOperand();
525 
526   // swifterror memory addresses are mem2reg promoted by instruction selection.
527   // As such they cannot have regular uses like an instrumentation function and
528   // it makes no sense to track them as memory.
529   if (Addr->isSwiftError())
530     return false;
531 
532   int Idx = getMemoryAccessFuncIndex(Addr, DL);
533   if (Idx < 0)
534     return false;
535   if (IsWrite && isVtableAccess(I)) {
536     LLVM_DEBUG(dbgs() << "  VPTR : " << *I << "\n");
537     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
538     // StoredValue may be a vector type if we are storing several vptrs at once.
539     // In this case, just take the first element of the vector since this is
540     // enough to find vptr races.
541     if (isa<VectorType>(StoredValue->getType()))
542       StoredValue = IRB.CreateExtractElement(
543           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
544     if (StoredValue->getType()->isIntegerTy())
545       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
546     // Call TsanVptrUpdate.
547     IRB.CreateCall(TsanVptrUpdate,
548                    {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
549                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
550     NumInstrumentedVtableWrites++;
551     return true;
552   }
553   if (!IsWrite && isVtableAccess(I)) {
554     IRB.CreateCall(TsanVptrLoad,
555                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
556     NumInstrumentedVtableReads++;
557     return true;
558   }
559   const unsigned Alignment = IsWrite
560       ? cast<StoreInst>(I)->getAlignment()
561       : cast<LoadInst>(I)->getAlignment();
562   Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
563   const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
564   FunctionCallee OnAccessFunc = nullptr;
565   if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0)
566     OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
567   else
568     OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
569   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
570   if (IsWrite) NumInstrumentedWrites++;
571   else         NumInstrumentedReads++;
572   return true;
573 }
574 
575 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
576   uint32_t v = 0;
577   switch (ord) {
578     case AtomicOrdering::NotAtomic:
579       llvm_unreachable("unexpected atomic ordering!");
580     case AtomicOrdering::Unordered:              LLVM_FALLTHROUGH;
581     case AtomicOrdering::Monotonic:              v = 0; break;
582     // Not specified yet:
583     // case AtomicOrdering::Consume:                v = 1; break;
584     case AtomicOrdering::Acquire:                v = 2; break;
585     case AtomicOrdering::Release:                v = 3; break;
586     case AtomicOrdering::AcquireRelease:         v = 4; break;
587     case AtomicOrdering::SequentiallyConsistent: v = 5; break;
588   }
589   return IRB->getInt32(v);
590 }
591 
592 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
593 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
594 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
595 // instead we simply replace them with regular function calls, which are then
596 // intercepted by the run-time.
597 // Since tsan is running after everyone else, the calls should not be
598 // replaced back with intrinsics. If that becomes wrong at some point,
599 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
600 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
601   IRBuilder<> IRB(I);
602   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
603     IRB.CreateCall(
604         MemsetFn,
605         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
606          IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
607          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
608     I->eraseFromParent();
609   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
610     IRB.CreateCall(
611         isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
612         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
613          IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
614          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
615     I->eraseFromParent();
616   }
617   return false;
618 }
619 
620 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
621 // standards.  For background see C++11 standard.  A slightly older, publicly
622 // available draft of the standard (not entirely up-to-date, but close enough
623 // for casual browsing) is available here:
624 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
625 // The following page contains more background information:
626 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
627 
628 bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
629   IRBuilder<> IRB(I);
630   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
631     Value *Addr = LI->getPointerOperand();
632     int Idx = getMemoryAccessFuncIndex(Addr, DL);
633     if (Idx < 0)
634       return false;
635     const unsigned ByteSize = 1U << Idx;
636     const unsigned BitSize = ByteSize * 8;
637     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
638     Type *PtrTy = Ty->getPointerTo();
639     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
640                      createOrdering(&IRB, LI->getOrdering())};
641     Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
642     Value *C = IRB.CreateCall(TsanAtomicLoad[Idx], Args);
643     Value *Cast = IRB.CreateBitOrPointerCast(C, OrigTy);
644     I->replaceAllUsesWith(Cast);
645   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
646     Value *Addr = SI->getPointerOperand();
647     int Idx = getMemoryAccessFuncIndex(Addr, DL);
648     if (Idx < 0)
649       return false;
650     const unsigned ByteSize = 1U << Idx;
651     const unsigned BitSize = ByteSize * 8;
652     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
653     Type *PtrTy = Ty->getPointerTo();
654     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
655                      IRB.CreateBitOrPointerCast(SI->getValueOperand(), Ty),
656                      createOrdering(&IRB, SI->getOrdering())};
657     CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
658     ReplaceInstWithInst(I, C);
659   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
660     Value *Addr = RMWI->getPointerOperand();
661     int Idx = getMemoryAccessFuncIndex(Addr, DL);
662     if (Idx < 0)
663       return false;
664     FunctionCallee F = TsanAtomicRMW[RMWI->getOperation()][Idx];
665     if (!F)
666       return false;
667     const unsigned ByteSize = 1U << Idx;
668     const unsigned BitSize = ByteSize * 8;
669     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
670     Type *PtrTy = Ty->getPointerTo();
671     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
672                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
673                      createOrdering(&IRB, RMWI->getOrdering())};
674     CallInst *C = CallInst::Create(F, Args);
675     ReplaceInstWithInst(I, C);
676   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
677     Value *Addr = CASI->getPointerOperand();
678     int Idx = getMemoryAccessFuncIndex(Addr, DL);
679     if (Idx < 0)
680       return false;
681     const unsigned ByteSize = 1U << Idx;
682     const unsigned BitSize = ByteSize * 8;
683     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
684     Type *PtrTy = Ty->getPointerTo();
685     Value *CmpOperand =
686       IRB.CreateBitOrPointerCast(CASI->getCompareOperand(), Ty);
687     Value *NewOperand =
688       IRB.CreateBitOrPointerCast(CASI->getNewValOperand(), Ty);
689     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
690                      CmpOperand,
691                      NewOperand,
692                      createOrdering(&IRB, CASI->getSuccessOrdering()),
693                      createOrdering(&IRB, CASI->getFailureOrdering())};
694     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
695     Value *Success = IRB.CreateICmpEQ(C, CmpOperand);
696     Value *OldVal = C;
697     Type *OrigOldValTy = CASI->getNewValOperand()->getType();
698     if (Ty != OrigOldValTy) {
699       // The value is a pointer, so we need to cast the return value.
700       OldVal = IRB.CreateIntToPtr(C, OrigOldValTy);
701     }
702 
703     Value *Res =
704       IRB.CreateInsertValue(UndefValue::get(CASI->getType()), OldVal, 0);
705     Res = IRB.CreateInsertValue(Res, Success, 1);
706 
707     I->replaceAllUsesWith(Res);
708     I->eraseFromParent();
709   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
710     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
711     FunctionCallee F = FI->getSyncScopeID() == SyncScope::SingleThread
712                            ? TsanAtomicSignalFence
713                            : TsanAtomicThreadFence;
714     CallInst *C = CallInst::Create(F, Args);
715     ReplaceInstWithInst(I, C);
716   }
717   return true;
718 }
719 
720 int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr,
721                                               const DataLayout &DL) {
722   Type *OrigPtrTy = Addr->getType();
723   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
724   assert(OrigTy->isSized());
725   uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
726   if (TypeSize != 8  && TypeSize != 16 &&
727       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
728     NumAccessesWithBadSize++;
729     // Ignore all unusual sizes.
730     return -1;
731   }
732   size_t Idx = countTrailingZeros(TypeSize / 8);
733   assert(Idx < kNumberOfAccessSizes);
734   return Idx;
735 }
736