llvm/lib/Transforms/Instrumentation/ThreadSanitizer.cpp

   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/InitializePasses.h"
  42 #include "llvm/ProfileData/InstrProf.h"
  43 #include "llvm/Support/CommandLine.h"
  44 #include "llvm/Support/Debug.h"
  45 #include "llvm/Support/MathExtras.h"
  46 #include "llvm/Support/raw_ostream.h"
  47 #include "llvm/Transforms/Instrumentation.h"
  48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
  49 #include "llvm/Transforms/Utils/EscapeEnumerator.h"
  50 #include "llvm/Transforms/Utils/Local.h"
  51 #include "llvm/Transforms/Utils/ModuleUtils.h"
  52
  53 using namespace llvm;
  54
  55 #define DEBUG_TYPE "tsan"
  56
  57 static cl::opt<bool> ClInstrumentMemoryAccesses(
  58     "tsan-instrument-memory-accesses", cl::init(true),
  59     cl::desc("Instrument memory accesses"), cl::Hidden);
  60 static cl::opt<bool>
  61     ClInstrumentFuncEntryExit("tsan-instrument-func-entry-exit", cl::init(true),
  62                               cl::desc("Instrument function entry and exit"),
  63                               cl::Hidden);
  64 static cl::opt<bool> ClHandleCxxExceptions(
  65     "tsan-handle-cxx-exceptions", cl::init(true),
  66     cl::desc("Handle C++ exceptions (insert cleanup blocks for unwinding)"),
  67     cl::Hidden);
  68 static cl::opt<bool> ClInstrumentAtomics("tsan-instrument-atomics",
  69                                          cl::init(true),
  70                                          cl::desc("Instrument atomics"),
  71                                          cl::Hidden);
  72 static cl::opt<bool> ClInstrumentMemIntrinsics(
  73     "tsan-instrument-memintrinsics", cl::init(true),
  74     cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
  75 static cl::opt<bool> ClDistinguishVolatile(
  76     "tsan-distinguish-volatile", cl::init(false),
  77     cl::desc("Emit special instrumentation for accesses to volatiles"),
  78     cl::Hidden);
  79 static cl::opt<bool> ClInstrumentReadBeforeWrite(
  80     "tsan-instrument-read-before-write", cl::init(false),
  81     cl::desc("Do not eliminate read instrumentation for read-before-writes"),
  82     cl::Hidden);
  83 static cl::opt<bool> ClCompoundReadBeforeWrite(
  84     "tsan-compound-read-before-write", cl::init(false),
  85     cl::desc("Emit special compound instrumentation for reads-before-writes"),
  86     cl::Hidden);
  87
  88 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
  89 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
  90 STATISTIC(NumOmittedReadsBeforeWrite,
  91           "Number of reads ignored due to following writes");
  92 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
  93 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
  94 STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
  95 STATISTIC(NumOmittedReadsFromConstantGlobals,
  96           "Number of reads from constant globals");
  97 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
  98 STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");
  99
 100 const char kTsanModuleCtorName[] = "tsan.module_ctor";
 101 const char kTsanInitName[] = "__tsan_init";
 102
 103 namespace {
 104
 105 /// ThreadSanitizer: instrument the code in module to find races.
 106 ///
 107 /// Instantiating ThreadSanitizer inserts the tsan runtime library API function
 108 /// declarations into the module if they don't exist already. Instantiating
 109 /// ensures the __tsan_init function is in the list of global constructors for
 110 /// the module.
 111 struct ThreadSanitizer {
 112   ThreadSanitizer() {
 113     // Sanity check options and warn user.
 114     if (ClInstrumentReadBeforeWrite && ClCompoundReadBeforeWrite) {
 115       errs()
 116           << "warning: Option -tsan-compound-read-before-write has no effect "
 117              "when -tsan-instrument-read-before-write is set.\n";
 118     }
 119   }
 120
 121   bool sanitizeFunction(Function &F, const TargetLibraryInfo &TLI);
 122
 123 private:
 124   // Internal Instruction wrapper that contains more information about the
 125   // Instruction from prior analysis.
 126   struct InstructionInfo {
 127     // Instrumentation emitted for this instruction is for a compounded set of
 128     // read and write operations in the same basic block.
 129     static constexpr unsigned kCompoundRW = (1U << 0);
 130
 131     explicit InstructionInfo(Instruction *Inst) : Inst(Inst) {}
 132
 133     Instruction *Inst;
 134     unsigned Flags = 0;
 135   };
 136
 137   void initialize(Module &M);
 138   bool instrumentLoadOrStore(const InstructionInfo &II, const DataLayout &DL);
 139   bool instrumentAtomic(Instruction *I, const DataLayout &DL);
 140   bool instrumentMemIntrinsic(Instruction *I);
 141   void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
 142                                       SmallVectorImpl<InstructionInfo> &All,
 143                                       const DataLayout &DL);
 144   bool addrPointsToConstantData(Value *Addr);
 145   int getMemoryAccessFuncIndex(Type *OrigTy, Value *Addr, const DataLayout &DL);
 146   void InsertRuntimeIgnores(Function &F);
 147
 148   Type *IntptrTy;
 149   FunctionCallee TsanFuncEntry;
 150   FunctionCallee TsanFuncExit;
 151   FunctionCallee TsanIgnoreBegin;
 152   FunctionCallee TsanIgnoreEnd;
 153   // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
 154   static const size_t kNumberOfAccessSizes = 5;
 155   FunctionCallee TsanRead[kNumberOfAccessSizes];
 156   FunctionCallee TsanWrite[kNumberOfAccessSizes];
 157   FunctionCallee TsanUnalignedRead[kNumberOfAccessSizes];
 158   FunctionCallee TsanUnalignedWrite[kNumberOfAccessSizes];
 159   FunctionCallee TsanVolatileRead[kNumberOfAccessSizes];
 160   FunctionCallee TsanVolatileWrite[kNumberOfAccessSizes];
 161   FunctionCallee TsanUnalignedVolatileRead[kNumberOfAccessSizes];
 162   FunctionCallee TsanUnalignedVolatileWrite[kNumberOfAccessSizes];
 163   FunctionCallee TsanCompoundRW[kNumberOfAccessSizes];
 164   FunctionCallee TsanUnalignedCompoundRW[kNumberOfAccessSizes];
 165   FunctionCallee TsanAtomicLoad[kNumberOfAccessSizes];
 166   FunctionCallee TsanAtomicStore[kNumberOfAccessSizes];
 167   FunctionCallee TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1]
 168                               [kNumberOfAccessSizes];
 169   FunctionCallee TsanAtomicCAS[kNumberOfAccessSizes];
 170   FunctionCallee TsanAtomicThreadFence;
 171   FunctionCallee TsanAtomicSignalFence;
 172   FunctionCallee TsanVptrUpdate;
 173   FunctionCallee TsanVptrLoad;
 174   FunctionCallee MemmoveFn, MemcpyFn, MemsetFn;
 175 };
 176
 177 struct ThreadSanitizerLegacyPass : FunctionPass {
 178   ThreadSanitizerLegacyPass() : FunctionPass(ID) {
 179     initializeThreadSanitizerLegacyPassPass(*PassRegistry::getPassRegistry());
 180   }
 181   StringRef getPassName() const override;
 182   void getAnalysisUsage(AnalysisUsage &AU) const override;
 183   bool runOnFunction(Function &F) override;
 184   bool doInitialization(Module &M) override;
 185   static char ID; // Pass identification, replacement for typeid.
 186 private:
 187   Optional<ThreadSanitizer> TSan;
 188 };
 189
 190 void insertModuleCtor(Module &M) {
 191   getOrCreateSanitizerCtorAndInitFunctions(
 192       M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{},
 193       /*InitArgs=*/{},
 194       // This callback is invoked when the functions are created the first
 195       // time. Hook them into the global ctors list in that case:
 196       [&](Function *Ctor, FunctionCallee) { appendToGlobalCtors(M, Ctor, 0); });
 197 }
 198
 199 }  // namespace
 200
 201 PreservedAnalyses ThreadSanitizerPass::run(Function &F,
 202                                            FunctionAnalysisManager &FAM) {
 203   ThreadSanitizer TSan;
 204   if (TSan.sanitizeFunction(F, FAM.getResult<TargetLibraryAnalysis>(F)))
 205     return PreservedAnalyses::none();
 206   return PreservedAnalyses::all();
 207 }
 208
 209 PreservedAnalyses ThreadSanitizerPass::run(Module &M,
 210                                            ModuleAnalysisManager &MAM) {
 211   insertModuleCtor(M);
 212   return PreservedAnalyses::none();
 213 }
 214
 215 char ThreadSanitizerLegacyPass::ID = 0;
 216 INITIALIZE_PASS_BEGIN(ThreadSanitizerLegacyPass, "tsan",
 217                       "ThreadSanitizer: detects data races.", false, false)
 218 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
 219 INITIALIZE_PASS_END(ThreadSanitizerLegacyPass, "tsan",
 220                     "ThreadSanitizer: detects data races.", false, false)
 221
 222 StringRef ThreadSanitizerLegacyPass::getPassName() const {
 223   return "ThreadSanitizerLegacyPass";
 224 }
 225
 226 void ThreadSanitizerLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
 227   AU.addRequired<TargetLibraryInfoWrapperPass>();
 228 }
 229
 230 bool ThreadSanitizerLegacyPass::doInitialization(Module &M) {
 231   insertModuleCtor(M);
 232   TSan.emplace();
 233   return true;
 234 }
 235
 236 bool ThreadSanitizerLegacyPass::runOnFunction(Function &F) {
 237   auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
 238   TSan->sanitizeFunction(F, TLI);
 239   return true;
 240 }
 241
 242 FunctionPass *llvm::createThreadSanitizerLegacyPassPass() {
 243   return new ThreadSanitizerLegacyPass();
 244 }
 245
 246 void ThreadSanitizer::initialize(Module &M) {
 247   const DataLayout &DL = M.getDataLayout();
 248   IntptrTy = DL.getIntPtrType(M.getContext());
 249
 250   IRBuilder<> IRB(M.getContext());
 251   AttributeList Attr;
 252   Attr = Attr.addFnAttribute(M.getContext(), Attribute::NoUnwind);
 253   // Initialize the callbacks.
 254   TsanFuncEntry = M.getOrInsertFunction("__tsan_func_entry", Attr,
 255                                         IRB.getVoidTy(), IRB.getInt8PtrTy());
 256   TsanFuncExit =
 257       M.getOrInsertFunction("__tsan_func_exit", Attr, IRB.getVoidTy());
 258   TsanIgnoreBegin = M.getOrInsertFunction("__tsan_ignore_thread_begin", Attr,
 259                                           IRB.getVoidTy());
 260   TsanIgnoreEnd =
 261       M.getOrInsertFunction("__tsan_ignore_thread_end", Attr, IRB.getVoidTy());
 262   IntegerType *OrdTy = IRB.getInt32Ty();
 263   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
 264     const unsigned ByteSize = 1U << i;
 265     const unsigned BitSize = ByteSize * 8;
 266     std::string ByteSizeStr = utostr(ByteSize);
 267     std::string BitSizeStr = utostr(BitSize);
 268     SmallString<32> ReadName("__tsan_read" + ByteSizeStr);
 269     TsanRead[i] = M.getOrInsertFunction(ReadName, Attr, IRB.getVoidTy(),
 270                                         IRB.getInt8PtrTy());
 271
 272     SmallString<32> WriteName("__tsan_write" + ByteSizeStr);
 273     TsanWrite[i] = M.getOrInsertFunction(WriteName, Attr, IRB.getVoidTy(),
 274                                          IRB.getInt8PtrTy());
 275
 276     SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr);
 277     TsanUnalignedRead[i] = M.getOrInsertFunction(
 278         UnalignedReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
 279
 280     SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr);
 281     TsanUnalignedWrite[i] = M.getOrInsertFunction(
 282         UnalignedWriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
 283
 284     SmallString<64> VolatileReadName("__tsan_volatile_read" + ByteSizeStr);
 285     TsanVolatileRead[i] = M.getOrInsertFunction(
 286         VolatileReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
 287
 288     SmallString<64> VolatileWriteName("__tsan_volatile_write" + ByteSizeStr);
 289     TsanVolatileWrite[i] = M.getOrInsertFunction(
 290         VolatileWriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
 291
 292     SmallString<64> UnalignedVolatileReadName("__tsan_unaligned_volatile_read" +
 293                                               ByteSizeStr);
 294     TsanUnalignedVolatileRead[i] = M.getOrInsertFunction(
 295         UnalignedVolatileReadName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
 296
 297     SmallString<64> UnalignedVolatileWriteName(
 298         "__tsan_unaligned_volatile_write" + ByteSizeStr);
 299     TsanUnalignedVolatileWrite[i] = M.getOrInsertFunction(
 300         UnalignedVolatileWriteName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
 301
 302     SmallString<64> CompoundRWName("__tsan_read_write" + ByteSizeStr);
 303     TsanCompoundRW[i] = M.getOrInsertFunction(
 304         CompoundRWName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
 305
 306     SmallString<64> UnalignedCompoundRWName("__tsan_unaligned_read_write" +
 307                                             ByteSizeStr);
 308     TsanUnalignedCompoundRW[i] = M.getOrInsertFunction(
 309         UnalignedCompoundRWName, Attr, IRB.getVoidTy(), IRB.getInt8PtrTy());
 310
 311     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
 312     Type *PtrTy = Ty->getPointerTo();
 313     SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load");
 314     {
 315       AttributeList AL = Attr;
 316       AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
 317       TsanAtomicLoad[i] =
 318           M.getOrInsertFunction(AtomicLoadName, AL, Ty, PtrTy, OrdTy);
 319     }
 320
 321     SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store");
 322     {
 323       AttributeList AL = Attr;
 324       AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
 325       AL = AL.addParamAttribute(M.getContext(), 2, Attribute::ZExt);
 326       TsanAtomicStore[i] = M.getOrInsertFunction(
 327           AtomicStoreName, AL, IRB.getVoidTy(), PtrTy, Ty, OrdTy);
 328     }
 329
 330     for (unsigned Op = AtomicRMWInst::FIRST_BINOP;
 331          Op <= AtomicRMWInst::LAST_BINOP; ++Op) {
 332       TsanAtomicRMW[Op][i] = nullptr;
 333       const char *NamePart = nullptr;
 334       if (Op == AtomicRMWInst::Xchg)
 335         NamePart = "_exchange";
 336       else if (Op == AtomicRMWInst::Add)
 337         NamePart = "_fetch_add";
 338       else if (Op == AtomicRMWInst::Sub)
 339         NamePart = "_fetch_sub";
 340       else if (Op == AtomicRMWInst::And)
 341         NamePart = "_fetch_and";
 342       else if (Op == AtomicRMWInst::Or)
 343         NamePart = "_fetch_or";
 344       else if (Op == AtomicRMWInst::Xor)
 345         NamePart = "_fetch_xor";
 346       else if (Op == AtomicRMWInst::Nand)
 347         NamePart = "_fetch_nand";
 348       else
 349         continue;
 350       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
 351       {
 352         AttributeList AL = Attr;
 353         AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
 354         AL = AL.addParamAttribute(M.getContext(), 2, Attribute::ZExt);
 355         TsanAtomicRMW[Op][i] =
 356             M.getOrInsertFunction(RMWName, AL, Ty, PtrTy, Ty, OrdTy);
 357       }
 358     }
 359
 360     SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr +
 361                                   "_compare_exchange_val");
 362     {
 363       AttributeList AL = Attr;
 364       AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
 365       AL = AL.addParamAttribute(M.getContext(), 2, Attribute::ZExt);
 366       AL = AL.addParamAttribute(M.getContext(), 3, Attribute::ZExt);
 367       AL = AL.addParamAttribute(M.getContext(), 4, Attribute::ZExt);
 368       TsanAtomicCAS[i] = M.getOrInsertFunction(AtomicCASName, AL, Ty, PtrTy, Ty,
 369                                                Ty, OrdTy, OrdTy);
 370     }
 371   }
 372   TsanVptrUpdate =
 373       M.getOrInsertFunction("__tsan_vptr_update", Attr, IRB.getVoidTy(),
 374                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy());
 375   TsanVptrLoad = M.getOrInsertFunction("__tsan_vptr_read", Attr,
 376                                        IRB.getVoidTy(), IRB.getInt8PtrTy());
 377   {
 378     AttributeList AL = Attr;
 379     AL = AL.addParamAttribute(M.getContext(), 0, Attribute::ZExt);
 380     TsanAtomicThreadFence = M.getOrInsertFunction("__tsan_atomic_thread_fence",
 381                                                   AL, IRB.getVoidTy(), OrdTy);
 382   }
 383   {
 384     AttributeList AL = Attr;
 385     AL = AL.addParamAttribute(M.getContext(), 0, Attribute::ZExt);
 386     TsanAtomicSignalFence = M.getOrInsertFunction("__tsan_atomic_signal_fence",
 387                                                   AL, IRB.getVoidTy(), OrdTy);
 388   }
 389
 390   MemmoveFn =
 391       M.getOrInsertFunction("memmove", Attr, IRB.getInt8PtrTy(),
 392                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy);
 393   MemcpyFn =
 394       M.getOrInsertFunction("memcpy", Attr, IRB.getInt8PtrTy(),
 395                             IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IntptrTy);
 396   MemsetFn =
 397       M.getOrInsertFunction("memset", Attr, IRB.getInt8PtrTy(),
 398                             IRB.getInt8PtrTy(), IRB.getInt32Ty(), IntptrTy);
 399 }
 400
 401 static bool isVtableAccess(Instruction *I) {
 402   if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
 403     return Tag->isTBAAVtableAccess();
 404   return false;
 405 }
 406
 407 // Do not instrument known races/"benign races" that come from compiler
 408 // instrumentatin. The user has no way of suppressing them.
 409 static bool shouldInstrumentReadWriteFromAddress(const Module *M, Value *Addr) {
 410   // Peel off GEPs and BitCasts.
 411   Addr = Addr->stripInBoundsOffsets();
 412
 413   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
 414     if (GV->hasSection()) {
 415       StringRef SectionName = GV->getSection();
 416       // Check if the global is in the PGO counters section.
 417       auto OF = Triple(M->getTargetTriple()).getObjectFormat();
 418       if (SectionName.endswith(
 419               getInstrProfSectionName(IPSK_cnts, OF, /*AddSegmentInfo=*/false)))
 420         return false;
 421     }
 422
 423     // Check if the global is private gcov data.
 424     if (GV->getName().startswith("__llvm_gcov") ||
 425         GV->getName().startswith("__llvm_gcda"))
 426       return false;
 427   }
 428
 429   // Do not instrument acesses from different address spaces; we cannot deal
 430   // with them.
 431   if (Addr) {
 432     Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
 433     if (PtrTy->getPointerAddressSpace() != 0)
 434       return false;
 435   }
 436
 437   return true;
 438 }
 439
 440 bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
 441   // If this is a GEP, just analyze its pointer operand.
 442   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))
 443     Addr = GEP->getPointerOperand();
 444
 445   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
 446     if (GV->isConstant()) {
 447       // Reads from constant globals can not race with any writes.
 448       NumOmittedReadsFromConstantGlobals++;
 449       return true;
 450     }
 451   } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {
 452     if (isVtableAccess(L)) {
 453       // Reads from a vtable pointer can not race with any writes.
 454       NumOmittedReadsFromVtable++;
 455       return true;
 456     }
 457   }
 458   return false;
 459 }
 460
 461 // Instrumenting some of the accesses may be proven redundant.
 462 // Currently handled:
 463 //  - read-before-write (within same BB, no calls between)
 464 //  - not captured variables
 465 //
 466 // We do not handle some of the patterns that should not survive
 467 // after the classic compiler optimizations.
 468 // E.g. two reads from the same temp should be eliminated by CSE,
 469 // two writes should be eliminated by DSE, etc.
 470 //
 471 // 'Local' is a vector of insns within the same BB (no calls between).
 472 // 'All' is a vector of insns that will be instrumented.
 473 void ThreadSanitizer::chooseInstructionsToInstrument(
 474     SmallVectorImpl<Instruction *> &Local,
 475     SmallVectorImpl<InstructionInfo> &All, const DataLayout &DL) {
 476   DenseMap<Value *, size_t> WriteTargets; // Map of addresses to index in All
 477   // Iterate from the end.
 478   for (Instruction *I : reverse(Local)) {
 479     const bool IsWrite = isa<StoreInst>(*I);
 480     Value *Addr = IsWrite ? cast<StoreInst>(I)->getPointerOperand()
 481                           : cast<LoadInst>(I)->getPointerOperand();
 482
 483     if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))
 484       continue;
 485
 486     if (!IsWrite) {
 487       const auto WriteEntry = WriteTargets.find(Addr);
 488       if (!ClInstrumentReadBeforeWrite && WriteEntry != WriteTargets.end()) {
 489         auto &WI = All[WriteEntry->second];
 490         // If we distinguish volatile accesses and if either the read or write
 491         // is volatile, do not omit any instrumentation.
 492         const bool AnyVolatile =
 493             ClDistinguishVolatile && (cast<LoadInst>(I)->isVolatile() ||
 494                                       cast<StoreInst>(WI.Inst)->isVolatile());
 495         if (!AnyVolatile) {
 496           // We will write to this temp, so no reason to analyze the read.
 497           // Mark the write instruction as compound.
 498           WI.Flags |= InstructionInfo::kCompoundRW;
 499           NumOmittedReadsBeforeWrite++;
 500           continue;
 501         }
 502       }
 503
 504       if (addrPointsToConstantData(Addr)) {
 505         // Addr points to some constant data -- it can not race with any writes.
 506         continue;
 507       }
 508     }
 509
 510     if (isa<AllocaInst>(getUnderlyingObject(Addr)) &&
 511         !PointerMayBeCaptured(Addr, true, true)) {
 512       // The variable is addressable but not captured, so it cannot be
 513       // referenced from a different thread and participate in a data race
 514       // (see llvm/Analysis/CaptureTracking.h for details).
 515       NumOmittedNonCaptured++;
 516       continue;
 517     }
 518
 519     // Instrument this instruction.
 520     All.emplace_back(I);
 521     if (IsWrite) {
 522       // For read-before-write and compound instrumentation we only need one
 523       // write target, and we can override any previous entry if it exists.
 524       WriteTargets[Addr] = All.size() - 1;
 525     }
 526   }
 527   Local.clear();
 528 }
 529
 530 static bool isAtomic(Instruction *I) {
 531   // TODO: Ask TTI whether synchronization scope is between threads.
 532   if (LoadInst *LI = dyn_cast<LoadInst>(I))
 533     return LI->isAtomic() && LI->getSyncScopeID() != SyncScope::SingleThread;
 534   if (StoreInst *SI = dyn_cast<StoreInst>(I))
 535     return SI->isAtomic() && SI->getSyncScopeID() != SyncScope::SingleThread;
 536   if (isa<AtomicRMWInst>(I))
 537     return true;
 538   if (isa<AtomicCmpXchgInst>(I))
 539     return true;
 540   if (isa<FenceInst>(I))
 541     return true;
 542   return false;
 543 }
 544
 545 void ThreadSanitizer::InsertRuntimeIgnores(Function &F) {
 546   IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
 547   IRB.CreateCall(TsanIgnoreBegin);
 548   EscapeEnumerator EE(F, "tsan_ignore_cleanup", ClHandleCxxExceptions);
 549   while (IRBuilder<> *AtExit = EE.Next()) {
 550     AtExit->CreateCall(TsanIgnoreEnd);
 551   }
 552 }
 553
 554 bool ThreadSanitizer::sanitizeFunction(Function &F,
 555                                        const TargetLibraryInfo &TLI) {
 556   // This is required to prevent instrumenting call to __tsan_init from within
 557   // the module constructor.
 558   if (F.getName() == kTsanModuleCtorName)
 559     return false;
 560   // Naked functions can not have prologue/epilogue
 561   // (__tsan_func_entry/__tsan_func_exit) generated, so don't instrument them at
 562   // all.
 563   if (F.hasFnAttribute(Attribute::Naked))
 564     return false;
 565   initialize(*F.getParent());
 566   SmallVector<InstructionInfo, 8> AllLoadsAndStores;
 567   SmallVector<Instruction*, 8> LocalLoadsAndStores;
 568   SmallVector<Instruction*, 8> AtomicAccesses;
 569   SmallVector<Instruction*, 8> MemIntrinCalls;
 570   bool Res = false;
 571   bool HasCalls = false;
 572   bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
 573   const DataLayout &DL = F.getParent()->getDataLayout();
 574
 575   // Traverse all instructions, collect loads/stores/returns, check for calls.
 576   for (auto &BB : F) {
 577     for (auto &Inst : BB) {
 578       if (isAtomic(&Inst))
 579         AtomicAccesses.push_back(&Inst);
 580       else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
 581         LocalLoadsAndStores.push_back(&Inst);
 582       else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
 583         if (CallInst *CI = dyn_cast<CallInst>(&Inst))
 584           maybeMarkSanitizerLibraryCallNoBuiltin(CI, &TLI);
 585         if (isa<MemIntrinsic>(Inst))
 586           MemIntrinCalls.push_back(&Inst);
 587         HasCalls = true;
 588         chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
 589                                        DL);
 590       }
 591     }
 592     chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
 593   }
 594
 595   // We have collected all loads and stores.
 596   // FIXME: many of these accesses do not need to be checked for races
 597   // (e.g. variables that do not escape, etc).
 598
 599   // Instrument memory accesses only if we want to report bugs in the function.
 600   if (ClInstrumentMemoryAccesses && SanitizeFunction)
 601     for (const auto &II : AllLoadsAndStores) {
 602       Res |= instrumentLoadOrStore(II, DL);
 603     }
 604
 605   // Instrument atomic memory accesses in any case (they can be used to
 606   // implement synchronization).
 607   if (ClInstrumentAtomics)
 608     for (auto Inst : AtomicAccesses) {
 609       Res |= instrumentAtomic(Inst, DL);
 610     }
 611
 612   if (ClInstrumentMemIntrinsics && SanitizeFunction)
 613     for (auto Inst : MemIntrinCalls) {
 614       Res |= instrumentMemIntrinsic(Inst);
 615     }
 616
 617   if (F.hasFnAttribute("sanitize_thread_no_checking_at_run_time")) {
 618     assert(!F.hasFnAttribute(Attribute::SanitizeThread));
 619     if (HasCalls)
 620       InsertRuntimeIgnores(F);
 621   }
 622
 623   // Instrument function entry/exit points if there were instrumented accesses.
 624   if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {
 625     IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
 626     Value *ReturnAddress = IRB.CreateCall(
 627         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
 628         IRB.getInt32(0));
 629     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
 630
 631     EscapeEnumerator EE(F, "tsan_cleanup", ClHandleCxxExceptions);
 632     while (IRBuilder<> *AtExit = EE.Next()) {
 633       AtExit->CreateCall(TsanFuncExit, {});
 634     }
 635     Res = true;
 636   }
 637   return Res;
 638 }
 639
 640 bool ThreadSanitizer::instrumentLoadOrStore(const InstructionInfo &II,
 641                                             const DataLayout &DL) {
 642   IRBuilder<> IRB(II.Inst);
 643   const bool IsWrite = isa<StoreInst>(*II.Inst);
 644   Value *Addr = IsWrite ? cast<StoreInst>(II.Inst)->getPointerOperand()
 645                         : cast<LoadInst>(II.Inst)->getPointerOperand();
 646   Type *OrigTy = getLoadStoreType(II.Inst);
 647
 648   // swifterror memory addresses are mem2reg promoted by instruction selection.
 649   // As such they cannot have regular uses like an instrumentation function and
 650   // it makes no sense to track them as memory.
 651   if (Addr->isSwiftError())
 652     return false;
 653
 654   int Idx = getMemoryAccessFuncIndex(OrigTy, Addr, DL);
 655   if (Idx < 0)
 656     return false;
 657   if (IsWrite && isVtableAccess(II.Inst)) {
 658     LLVM_DEBUG(dbgs() << "  VPTR : " << *II.Inst << "\n");
 659     Value *StoredValue = cast<StoreInst>(II.Inst)->getValueOperand();
 660     // StoredValue may be a vector type if we are storing several vptrs at once.
 661     // In this case, just take the first element of the vector since this is
 662     // enough to find vptr races.
 663     if (isa<VectorType>(StoredValue->getType()))
 664       StoredValue = IRB.CreateExtractElement(
 665           StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
 666     if (StoredValue->getType()->isIntegerTy())
 667       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
 668     // Call TsanVptrUpdate.
 669     IRB.CreateCall(TsanVptrUpdate,
 670                    {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
 671                     IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
 672     NumInstrumentedVtableWrites++;
 673     return true;
 674   }
 675   if (!IsWrite && isVtableAccess(II.Inst)) {
 676     IRB.CreateCall(TsanVptrLoad,
 677                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
 678     NumInstrumentedVtableReads++;
 679     return true;
 680   }
 681
 682   const unsigned Alignment = IsWrite ? cast<StoreInst>(II.Inst)->getAlignment()
 683                                      : cast<LoadInst>(II.Inst)->getAlignment();
 684   const bool IsCompoundRW =
 685       ClCompoundReadBeforeWrite && (II.Flags & InstructionInfo::kCompoundRW);
 686   const bool IsVolatile = ClDistinguishVolatile &&
 687                           (IsWrite ? cast<StoreInst>(II.Inst)->isVolatile()
 688                                    : cast<LoadInst>(II.Inst)->isVolatile());
 689   assert((!IsVolatile || !IsCompoundRW) && "Compound volatile invalid!");
 690
 691   const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
 692   FunctionCallee OnAccessFunc = nullptr;
 693   if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0) {
 694     if (IsCompoundRW)
 695       OnAccessFunc = TsanCompoundRW[Idx];
 696     else if (IsVolatile)
 697       OnAccessFunc = IsWrite ? TsanVolatileWrite[Idx] : TsanVolatileRead[Idx];
 698     else
 699       OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
 700   } else {
 701     if (IsCompoundRW)
 702       OnAccessFunc = TsanUnalignedCompoundRW[Idx];
 703     else if (IsVolatile)
 704       OnAccessFunc = IsWrite ? TsanUnalignedVolatileWrite[Idx]
 705                              : TsanUnalignedVolatileRead[Idx];
 706     else
 707       OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
 708   }
 709   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
 710   if (IsCompoundRW || IsWrite)
 711     NumInstrumentedWrites++;
 712   if (IsCompoundRW || !IsWrite)
 713     NumInstrumentedReads++;
 714   return true;
 715 }
 716
 717 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
 718   uint32_t v = 0;
 719   switch (ord) {
 720     case AtomicOrdering::NotAtomic:
 721       llvm_unreachable("unexpected atomic ordering!");
 722     case AtomicOrdering::Unordered:              LLVM_FALLTHROUGH;
 723     case AtomicOrdering::Monotonic:              v = 0; break;
 724     // Not specified yet:
 725     // case AtomicOrdering::Consume:                v = 1; break;
 726     case AtomicOrdering::Acquire:                v = 2; break;
 727     case AtomicOrdering::Release:                v = 3; break;
 728     case AtomicOrdering::AcquireRelease:         v = 4; break;
 729     case AtomicOrdering::SequentiallyConsistent: v = 5; break;
 730   }
 731   return IRB->getInt32(v);
 732 }
 733
 734 // If a memset intrinsic gets inlined by the code gen, we will miss races on it.
 735 // So, we either need to ensure the intrinsic is not inlined, or instrument it.
 736 // We do not instrument memset/memmove/memcpy intrinsics (too complicated),
 737 // instead we simply replace them with regular function calls, which are then
 738 // intercepted by the run-time.
 739 // Since tsan is running after everyone else, the calls should not be
 740 // replaced back with intrinsics. If that becomes wrong at some point,
 741 // we will need to call e.g. __tsan_memset to avoid the intrinsics.
 742 bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
 743   IRBuilder<> IRB(I);
 744   if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
 745     IRB.CreateCall(
 746         MemsetFn,
 747         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
 748          IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
 749          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
 750     I->eraseFromParent();
 751   } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
 752     IRB.CreateCall(
 753         isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
 754         {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
 755          IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
 756          IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
 757     I->eraseFromParent();
 758   }
 759   return false;
 760 }
 761
 762 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
 763 // standards.  For background see C++11 standard.  A slightly older, publicly
 764 // available draft of the standard (not entirely up-to-date, but close enough
 765 // for casual browsing) is available here:
 766 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
 767 // The following page contains more background information:
 768 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
 769
 770 bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
 771   IRBuilder<> IRB(I);
 772   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
 773     Value *Addr = LI->getPointerOperand();
 774     Type *OrigTy = LI->getType();
 775     int Idx = getMemoryAccessFuncIndex(OrigTy, Addr, DL);
 776     if (Idx < 0)
 777       return false;
 778     const unsigned ByteSize = 1U << Idx;
 779     const unsigned BitSize = ByteSize * 8;
 780     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 781     Type *PtrTy = Ty->getPointerTo();
 782     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 783                      createOrdering(&IRB, LI->getOrdering())};
 784     Value *C = IRB.CreateCall(TsanAtomicLoad[Idx], Args);
 785     Value *Cast = IRB.CreateBitOrPointerCast(C, OrigTy);
 786     I->replaceAllUsesWith(Cast);
 787   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
 788     Value *Addr = SI->getPointerOperand();
 789     int Idx =
 790         getMemoryAccessFuncIndex(SI->getValueOperand()->getType(), Addr, DL);
 791     if (Idx < 0)
 792       return false;
 793     const unsigned ByteSize = 1U << Idx;
 794     const unsigned BitSize = ByteSize * 8;
 795     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 796     Type *PtrTy = Ty->getPointerTo();
 797     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 798                      IRB.CreateBitOrPointerCast(SI->getValueOperand(), Ty),
 799                      createOrdering(&IRB, SI->getOrdering())};
 800     CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
 801     ReplaceInstWithInst(I, C);
 802   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
 803     Value *Addr = RMWI->getPointerOperand();
 804     int Idx =
 805         getMemoryAccessFuncIndex(RMWI->getValOperand()->getType(), Addr, DL);
 806     if (Idx < 0)
 807       return false;
 808     FunctionCallee F = TsanAtomicRMW[RMWI->getOperation()][Idx];
 809     if (!F)
 810       return false;
 811     const unsigned ByteSize = 1U << Idx;
 812     const unsigned BitSize = ByteSize * 8;
 813     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 814     Type *PtrTy = Ty->getPointerTo();
 815     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 816                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
 817                      createOrdering(&IRB, RMWI->getOrdering())};
 818     CallInst *C = CallInst::Create(F, Args);
 819     ReplaceInstWithInst(I, C);
 820   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
 821     Value *Addr = CASI->getPointerOperand();
 822     Type *OrigOldValTy = CASI->getNewValOperand()->getType();
 823     int Idx = getMemoryAccessFuncIndex(OrigOldValTy, Addr, DL);
 824     if (Idx < 0)
 825       return false;
 826     const unsigned ByteSize = 1U << Idx;
 827     const unsigned BitSize = ByteSize * 8;
 828     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
 829     Type *PtrTy = Ty->getPointerTo();
 830     Value *CmpOperand =
 831       IRB.CreateBitOrPointerCast(CASI->getCompareOperand(), Ty);
 832     Value *NewOperand =
 833       IRB.CreateBitOrPointerCast(CASI->getNewValOperand(), Ty);
 834     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
 835                      CmpOperand,
 836                      NewOperand,
 837                      createOrdering(&IRB, CASI->getSuccessOrdering()),
 838                      createOrdering(&IRB, CASI->getFailureOrdering())};
 839     CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
 840     Value *Success = IRB.CreateICmpEQ(C, CmpOperand);
 841     Value *OldVal = C;
 842     if (Ty != OrigOldValTy) {
 843       // The value is a pointer, so we need to cast the return value.
 844       OldVal = IRB.CreateIntToPtr(C, OrigOldValTy);
 845     }
 846
 847     Value *Res =
 848       IRB.CreateInsertValue(UndefValue::get(CASI->getType()), OldVal, 0);
 849     Res = IRB.CreateInsertValue(Res, Success, 1);
 850
 851     I->replaceAllUsesWith(Res);
 852     I->eraseFromParent();
 853   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
 854     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
 855     FunctionCallee F = FI->getSyncScopeID() == SyncScope::SingleThread
 856                            ? TsanAtomicSignalFence
 857                            : TsanAtomicThreadFence;
 858     CallInst *C = CallInst::Create(F, Args);
 859     ReplaceInstWithInst(I, C);
 860   }
 861   return true;
 862 }
 863
 864 int ThreadSanitizer::getMemoryAccessFuncIndex(Type *OrigTy, Value *Addr,
 865                                               const DataLayout &DL) {
 866   assert(OrigTy->isSized());
 867   assert(
 868       cast<PointerType>(Addr->getType())->isOpaqueOrPointeeTypeMatches(OrigTy));
 869   uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
 870   if (TypeSize != 8  && TypeSize != 16 &&
 871       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
 872     NumAccessesWithBadSize++;
 873     // Ignore all unusual sizes.
 874     return -1;
 875   }
 876   size_t Idx = countTrailingZeros(TypeSize / 8);
 877   assert(Idx < kNumberOfAccessSizes);
 878   return Idx;
 879 }