830 lines · 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-exception6//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 see12// http://code.google.com/p/data-race-test13//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/SmallString.h"24#include "llvm/ADT/SmallVector.h"25#include "llvm/ADT/Statistic.h"26#include "llvm/ADT/StringExtras.h"27#include "llvm/Analysis/CaptureTracking.h"28#include "llvm/Analysis/TargetLibraryInfo.h"29#include "llvm/Analysis/ValueTracking.h"30#include "llvm/IR/DataLayout.h"31#include "llvm/IR/Function.h"32#include "llvm/IR/IRBuilder.h"33#include "llvm/IR/Instructions.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/raw_ostream.h"44#include "llvm/Transforms/Utils/EscapeEnumerator.h"45#include "llvm/Transforms/Utils/Instrumentation.h"46#include "llvm/Transforms/Utils/Local.h"47#include "llvm/Transforms/Utils/ModuleUtils.h"48 49using namespace llvm;50 51#define DEBUG_TYPE "tsan"52 53static cl::opt<bool> ClInstrumentMemoryAccesses(54 "tsan-instrument-memory-accesses", cl::init(true),55 cl::desc("Instrument memory accesses"), cl::Hidden);56static cl::opt<bool>57 ClInstrumentFuncEntryExit("tsan-instrument-func-entry-exit", cl::init(true),58 cl::desc("Instrument function entry and exit"),59 cl::Hidden);60static 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);64static cl::opt<bool> ClInstrumentAtomics("tsan-instrument-atomics",65 cl::init(true),66 cl::desc("Instrument atomics"),67 cl::Hidden);68static cl::opt<bool> ClInstrumentMemIntrinsics(69 "tsan-instrument-memintrinsics", cl::init(true),70 cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);71static cl::opt<bool> ClDistinguishVolatile(72 "tsan-distinguish-volatile", cl::init(false),73 cl::desc("Emit special instrumentation for accesses to volatiles"),74 cl::Hidden);75static cl::opt<bool> ClInstrumentReadBeforeWrite(76 "tsan-instrument-read-before-write", cl::init(false),77 cl::desc("Do not eliminate read instrumentation for read-before-writes"),78 cl::Hidden);79static cl::opt<bool> ClCompoundReadBeforeWrite(80 "tsan-compound-read-before-write", cl::init(false),81 cl::desc("Emit special compound instrumentation for reads-before-writes"),82 cl::Hidden);83static cl::opt<bool>84 ClOmitNonCaptured("tsan-omit-by-pointer-capturing", cl::init(true),85 cl::desc("Omit accesses due to pointer capturing"),86 cl::Hidden);87 88STATISTIC(NumInstrumentedReads, "Number of instrumented reads");89STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");90STATISTIC(NumOmittedReadsBeforeWrite,91 "Number of reads ignored due to following writes");92STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");93STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");94STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");95STATISTIC(NumOmittedReadsFromConstantGlobals,96 "Number of reads from constant globals");97STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");98STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");99 100const char kTsanModuleCtorName[] = "tsan.module_ctor";101const char kTsanInitName[] = "__tsan_init";102 103namespace {104 105/// ThreadSanitizer: instrument the code in module to find races.106///107/// Instantiating ThreadSanitizer inserts the tsan runtime library API function108/// declarations into the module if they don't exist already. Instantiating109/// ensures the __tsan_init function is in the list of global constructors for110/// the module.111struct ThreadSanitizer {112 ThreadSanitizer() {113 // 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 123private:124 // Internal Instruction wrapper that contains more information about the125 // Instruction from prior analysis.126 struct InstructionInfo {127 // Instrumentation emitted for this instruction is for a compounded set of128 // 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, const TargetLibraryInfo &TLI);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 177void insertModuleCtor(Module &M) {178 getOrCreateSanitizerCtorAndInitFunctions(179 M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{},180 /*InitArgs=*/{},181 // This callback is invoked when the functions are created the first182 // time. Hook them into the global ctors list in that case:183 [&](Function *Ctor, FunctionCallee) { appendToGlobalCtors(M, Ctor, 0); });184}185} // namespace186 187PreservedAnalyses ThreadSanitizerPass::run(Function &F,188 FunctionAnalysisManager &FAM) {189 ThreadSanitizer TSan;190 if (TSan.sanitizeFunction(F, FAM.getResult<TargetLibraryAnalysis>(F)))191 return PreservedAnalyses::none();192 return PreservedAnalyses::all();193}194 195PreservedAnalyses ModuleThreadSanitizerPass::run(Module &M,196 ModuleAnalysisManager &MAM) {197 // Return early if nosanitize_thread module flag is present for the module.198 if (checkIfAlreadyInstrumented(M, "nosanitize_thread"))199 return PreservedAnalyses::all();200 insertModuleCtor(M);201 return PreservedAnalyses::none();202}203void ThreadSanitizer::initialize(Module &M, const TargetLibraryInfo &TLI) {204 const DataLayout &DL = M.getDataLayout();205 LLVMContext &Ctx = M.getContext();206 IntptrTy = DL.getIntPtrType(Ctx);207 208 IRBuilder<> IRB(Ctx);209 AttributeList Attr;210 Attr = Attr.addFnAttribute(Ctx, Attribute::NoUnwind);211 // Initialize the callbacks.212 TsanFuncEntry = M.getOrInsertFunction("__tsan_func_entry", Attr,213 IRB.getVoidTy(), IRB.getPtrTy());214 TsanFuncExit =215 M.getOrInsertFunction("__tsan_func_exit", Attr, IRB.getVoidTy());216 TsanIgnoreBegin = M.getOrInsertFunction("__tsan_ignore_thread_begin", Attr,217 IRB.getVoidTy());218 TsanIgnoreEnd =219 M.getOrInsertFunction("__tsan_ignore_thread_end", Attr, IRB.getVoidTy());220 IntegerType *OrdTy = IRB.getInt32Ty();221 for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {222 const unsigned ByteSize = 1U << i;223 const unsigned BitSize = ByteSize * 8;224 std::string ByteSizeStr = utostr(ByteSize);225 std::string BitSizeStr = utostr(BitSize);226 SmallString<32> ReadName("__tsan_read" + ByteSizeStr);227 TsanRead[i] = M.getOrInsertFunction(ReadName, Attr, IRB.getVoidTy(),228 IRB.getPtrTy());229 230 SmallString<32> WriteName("__tsan_write" + ByteSizeStr);231 TsanWrite[i] = M.getOrInsertFunction(WriteName, Attr, IRB.getVoidTy(),232 IRB.getPtrTy());233 234 SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr);235 TsanUnalignedRead[i] = M.getOrInsertFunction(236 UnalignedReadName, Attr, IRB.getVoidTy(), IRB.getPtrTy());237 238 SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr);239 TsanUnalignedWrite[i] = M.getOrInsertFunction(240 UnalignedWriteName, Attr, IRB.getVoidTy(), IRB.getPtrTy());241 242 SmallString<64> VolatileReadName("__tsan_volatile_read" + ByteSizeStr);243 TsanVolatileRead[i] = M.getOrInsertFunction(244 VolatileReadName, Attr, IRB.getVoidTy(), IRB.getPtrTy());245 246 SmallString<64> VolatileWriteName("__tsan_volatile_write" + ByteSizeStr);247 TsanVolatileWrite[i] = M.getOrInsertFunction(248 VolatileWriteName, Attr, IRB.getVoidTy(), IRB.getPtrTy());249 250 SmallString<64> UnalignedVolatileReadName("__tsan_unaligned_volatile_read" +251 ByteSizeStr);252 TsanUnalignedVolatileRead[i] = M.getOrInsertFunction(253 UnalignedVolatileReadName, Attr, IRB.getVoidTy(), IRB.getPtrTy());254 255 SmallString<64> UnalignedVolatileWriteName(256 "__tsan_unaligned_volatile_write" + ByteSizeStr);257 TsanUnalignedVolatileWrite[i] = M.getOrInsertFunction(258 UnalignedVolatileWriteName, Attr, IRB.getVoidTy(), IRB.getPtrTy());259 260 SmallString<64> CompoundRWName("__tsan_read_write" + ByteSizeStr);261 TsanCompoundRW[i] = M.getOrInsertFunction(262 CompoundRWName, Attr, IRB.getVoidTy(), IRB.getPtrTy());263 264 SmallString<64> UnalignedCompoundRWName("__tsan_unaligned_read_write" +265 ByteSizeStr);266 TsanUnalignedCompoundRW[i] = M.getOrInsertFunction(267 UnalignedCompoundRWName, Attr, IRB.getVoidTy(), IRB.getPtrTy());268 269 Type *Ty = Type::getIntNTy(Ctx, BitSize);270 Type *PtrTy = PointerType::get(Ctx, 0);271 SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load");272 TsanAtomicLoad[i] =273 M.getOrInsertFunction(AtomicLoadName,274 TLI.getAttrList(&Ctx, {1}, /*Signed=*/true,275 /*Ret=*/BitSize <= 32, Attr),276 Ty, PtrTy, OrdTy);277 278 // Args of type Ty need extension only when BitSize is 32 or less.279 using Idxs = std::vector<unsigned>;280 Idxs Idxs2Or12 ((BitSize <= 32) ? Idxs({1, 2}) : Idxs({2}));281 Idxs Idxs34Or1234((BitSize <= 32) ? Idxs({1, 2, 3, 4}) : Idxs({3, 4}));282 SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store");283 TsanAtomicStore[i] = M.getOrInsertFunction(284 AtomicStoreName,285 TLI.getAttrList(&Ctx, Idxs2Or12, /*Signed=*/true, /*Ret=*/false, Attr),286 IRB.getVoidTy(), PtrTy, Ty, OrdTy);287 288 for (unsigned Op = AtomicRMWInst::FIRST_BINOP;289 Op <= AtomicRMWInst::LAST_BINOP; ++Op) {290 TsanAtomicRMW[Op][i] = nullptr;291 const char *NamePart = nullptr;292 if (Op == AtomicRMWInst::Xchg)293 NamePart = "_exchange";294 else if (Op == AtomicRMWInst::Add)295 NamePart = "_fetch_add";296 else if (Op == AtomicRMWInst::Sub)297 NamePart = "_fetch_sub";298 else if (Op == AtomicRMWInst::And)299 NamePart = "_fetch_and";300 else if (Op == AtomicRMWInst::Or)301 NamePart = "_fetch_or";302 else if (Op == AtomicRMWInst::Xor)303 NamePart = "_fetch_xor";304 else if (Op == AtomicRMWInst::Nand)305 NamePart = "_fetch_nand";306 else307 continue;308 SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);309 TsanAtomicRMW[Op][i] = M.getOrInsertFunction(310 RMWName,311 TLI.getAttrList(&Ctx, Idxs2Or12, /*Signed=*/true,312 /*Ret=*/BitSize <= 32, Attr),313 Ty, PtrTy, Ty, OrdTy);314 }315 316 SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr +317 "_compare_exchange_val");318 TsanAtomicCAS[i] = M.getOrInsertFunction(319 AtomicCASName,320 TLI.getAttrList(&Ctx, Idxs34Or1234, /*Signed=*/true,321 /*Ret=*/BitSize <= 32, Attr),322 Ty, PtrTy, Ty, Ty, OrdTy, OrdTy);323 }324 TsanVptrUpdate =325 M.getOrInsertFunction("__tsan_vptr_update", Attr, IRB.getVoidTy(),326 IRB.getPtrTy(), IRB.getPtrTy());327 TsanVptrLoad = M.getOrInsertFunction("__tsan_vptr_read", Attr,328 IRB.getVoidTy(), IRB.getPtrTy());329 TsanAtomicThreadFence = M.getOrInsertFunction(330 "__tsan_atomic_thread_fence",331 TLI.getAttrList(&Ctx, {0}, /*Signed=*/true, /*Ret=*/false, Attr),332 IRB.getVoidTy(), OrdTy);333 334 TsanAtomicSignalFence = M.getOrInsertFunction(335 "__tsan_atomic_signal_fence",336 TLI.getAttrList(&Ctx, {0}, /*Signed=*/true, /*Ret=*/false, Attr),337 IRB.getVoidTy(), OrdTy);338 339 MemmoveFn =340 M.getOrInsertFunction("__tsan_memmove", Attr, IRB.getPtrTy(),341 IRB.getPtrTy(), IRB.getPtrTy(), IntptrTy);342 MemcpyFn =343 M.getOrInsertFunction("__tsan_memcpy", Attr, IRB.getPtrTy(),344 IRB.getPtrTy(), IRB.getPtrTy(), IntptrTy);345 MemsetFn = M.getOrInsertFunction(346 "__tsan_memset",347 TLI.getAttrList(&Ctx, {1}, /*Signed=*/true, /*Ret=*/false, Attr),348 IRB.getPtrTy(), IRB.getPtrTy(), IRB.getInt32Ty(), IntptrTy);349}350 351static bool isVtableAccess(Instruction *I) {352 if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))353 return Tag->isTBAAVtableAccess();354 return false;355}356 357// Do not instrument known races/"benign races" that come from compiler358// instrumentation. The user has no way of suppressing them.359static bool shouldInstrumentReadWriteFromAddress(const Module *M, Value *Addr) {360 // Peel off GEPs and BitCasts.361 Addr = Addr->stripInBoundsOffsets();362 363 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {364 if (GV->hasSection()) {365 StringRef SectionName = GV->getSection();366 // Check if the global is in the PGO counters section.367 auto OF = M->getTargetTriple().getObjectFormat();368 if (SectionName.ends_with(369 getInstrProfSectionName(IPSK_cnts, OF, /*AddSegmentInfo=*/false)))370 return false;371 }372 }373 374 // Do not instrument accesses from different address spaces; we cannot deal375 // with them.376 if (Addr) {377 Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());378 if (PtrTy->getPointerAddressSpace() != 0)379 return false;380 }381 382 return true;383}384 385bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {386 // If this is a GEP, just analyze its pointer operand.387 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr))388 Addr = GEP->getPointerOperand();389 390 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {391 if (GV->isConstant()) {392 // Reads from constant globals can not race with any writes.393 NumOmittedReadsFromConstantGlobals++;394 return true;395 }396 } else if (LoadInst *L = dyn_cast<LoadInst>(Addr)) {397 if (isVtableAccess(L)) {398 // Reads from a vtable pointer can not race with any writes.399 NumOmittedReadsFromVtable++;400 return true;401 }402 }403 return false;404}405 406// Instrumenting some of the accesses may be proven redundant.407// Currently handled:408// - read-before-write (within same BB, no calls between)409// - not captured variables410//411// We do not handle some of the patterns that should not survive412// after the classic compiler optimizations.413// E.g. two reads from the same temp should be eliminated by CSE,414// two writes should be eliminated by DSE, etc.415//416// 'Local' is a vector of insns within the same BB (no calls between).417// 'All' is a vector of insns that will be instrumented.418void ThreadSanitizer::chooseInstructionsToInstrument(419 SmallVectorImpl<Instruction *> &Local,420 SmallVectorImpl<InstructionInfo> &All, const DataLayout &DL) {421 DenseMap<Value *, size_t> WriteTargets; // Map of addresses to index in All422 // Iterate from the end.423 for (Instruction *I : reverse(Local)) {424 const bool IsWrite = isa<StoreInst>(*I);425 Value *Addr = IsWrite ? cast<StoreInst>(I)->getPointerOperand()426 : cast<LoadInst>(I)->getPointerOperand();427 428 if (!shouldInstrumentReadWriteFromAddress(I->getModule(), Addr))429 continue;430 431 if (!IsWrite) {432 const auto WriteEntry = WriteTargets.find(Addr);433 if (!ClInstrumentReadBeforeWrite && WriteEntry != WriteTargets.end()) {434 auto &WI = All[WriteEntry->second];435 // If we distinguish volatile accesses and if either the read or write436 // is volatile, do not omit any instrumentation.437 const bool AnyVolatile =438 ClDistinguishVolatile && (cast<LoadInst>(I)->isVolatile() ||439 cast<StoreInst>(WI.Inst)->isVolatile());440 if (!AnyVolatile) {441 // We will write to this temp, so no reason to analyze the read.442 // Mark the write instruction as compound.443 WI.Flags |= InstructionInfo::kCompoundRW;444 NumOmittedReadsBeforeWrite++;445 continue;446 }447 }448 449 if (addrPointsToConstantData(Addr)) {450 // Addr points to some constant data -- it can not race with any writes.451 continue;452 }453 }454 455 const AllocaInst *AI = findAllocaForValue(Addr);456 // Instead of Addr, we should check whether its base pointer is captured.457 if (AI && !PointerMayBeCaptured(AI, /*ReturnCaptures=*/true) &&458 ClOmitNonCaptured) {459 // The variable is addressable but not captured, so it cannot be460 // referenced from a different thread and participate in a data race461 // (see llvm/Analysis/CaptureTracking.h for details).462 NumOmittedNonCaptured++;463 continue;464 }465 466 // Instrument this instruction.467 All.emplace_back(I);468 if (IsWrite) {469 // For read-before-write and compound instrumentation we only need one470 // write target, and we can override any previous entry if it exists.471 WriteTargets[Addr] = All.size() - 1;472 }473 }474 Local.clear();475}476 477static bool isTsanAtomic(const Instruction *I) {478 // TODO: Ask TTI whether synchronization scope is between threads.479 auto SSID = getAtomicSyncScopeID(I);480 if (!SSID)481 return false;482 if (isa<LoadInst>(I) || isa<StoreInst>(I))483 return *SSID != SyncScope::SingleThread;484 return true;485}486 487void ThreadSanitizer::InsertRuntimeIgnores(Function &F) {488 InstrumentationIRBuilder IRB(&F.getEntryBlock(),489 F.getEntryBlock().getFirstNonPHIIt());490 IRB.CreateCall(TsanIgnoreBegin);491 EscapeEnumerator EE(F, "tsan_ignore_cleanup", ClHandleCxxExceptions);492 while (IRBuilder<> *AtExit = EE.Next()) {493 InstrumentationIRBuilder::ensureDebugInfo(*AtExit, F);494 AtExit->CreateCall(TsanIgnoreEnd);495 }496}497 498bool ThreadSanitizer::sanitizeFunction(Function &F,499 const TargetLibraryInfo &TLI) {500 // This is required to prevent instrumenting call to __tsan_init from within501 // the module constructor.502 if (F.getName() == kTsanModuleCtorName)503 return false;504 // Naked functions can not have prologue/epilogue505 // (__tsan_func_entry/__tsan_func_exit) generated, so don't instrument them at506 // all.507 if (F.hasFnAttribute(Attribute::Naked))508 return false;509 510 // __attribute__(disable_sanitizer_instrumentation) prevents all kinds of511 // instrumentation.512 if (F.hasFnAttribute(Attribute::DisableSanitizerInstrumentation))513 return false;514 515 initialize(*F.getParent(), TLI);516 SmallVector<InstructionInfo, 8> AllLoadsAndStores;517 SmallVector<Instruction*, 8> LocalLoadsAndStores;518 SmallVector<Instruction*, 8> AtomicAccesses;519 SmallVector<Instruction*, 8> MemIntrinCalls;520 bool Res = false;521 bool HasCalls = false;522 bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);523 const DataLayout &DL = F.getDataLayout();524 525 // Traverse all instructions, collect loads/stores/returns, check for calls.526 for (auto &BB : F) {527 for (auto &Inst : BB) {528 // Skip instructions inserted by another instrumentation.529 if (Inst.hasMetadata(LLVMContext::MD_nosanitize))530 continue;531 if (isTsanAtomic(&Inst))532 AtomicAccesses.push_back(&Inst);533 else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))534 LocalLoadsAndStores.push_back(&Inst);535 else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {536 if (CallInst *CI = dyn_cast<CallInst>(&Inst))537 maybeMarkSanitizerLibraryCallNoBuiltin(CI, &TLI);538 if (isa<MemIntrinsic>(Inst))539 MemIntrinCalls.push_back(&Inst);540 HasCalls = true;541 chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,542 DL);543 }544 }545 chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);546 }547 548 // We have collected all loads and stores.549 // FIXME: many of these accesses do not need to be checked for races550 // (e.g. variables that do not escape, etc).551 552 // Instrument memory accesses only if we want to report bugs in the function.553 if (ClInstrumentMemoryAccesses && SanitizeFunction)554 for (const auto &II : AllLoadsAndStores) {555 Res |= instrumentLoadOrStore(II, DL);556 }557 558 // Instrument atomic memory accesses in any case (they can be used to559 // implement synchronization).560 if (ClInstrumentAtomics)561 for (auto *Inst : AtomicAccesses) {562 Res |= instrumentAtomic(Inst, DL);563 }564 565 if (ClInstrumentMemIntrinsics && SanitizeFunction)566 for (auto *Inst : MemIntrinCalls) {567 Res |= instrumentMemIntrinsic(Inst);568 }569 570 if (F.hasFnAttribute("sanitize_thread_no_checking_at_run_time")) {571 assert(!F.hasFnAttribute(Attribute::SanitizeThread));572 if (HasCalls)573 InsertRuntimeIgnores(F);574 }575 576 // Instrument function entry/exit points if there were instrumented accesses.577 if ((Res || HasCalls) && ClInstrumentFuncEntryExit) {578 InstrumentationIRBuilder IRB(&F.getEntryBlock(),579 F.getEntryBlock().getFirstNonPHIIt());580 Value *ReturnAddress =581 IRB.CreateIntrinsic(Intrinsic::returnaddress, IRB.getInt32(0));582 IRB.CreateCall(TsanFuncEntry, ReturnAddress);583 584 EscapeEnumerator EE(F, "tsan_cleanup", ClHandleCxxExceptions);585 while (IRBuilder<> *AtExit = EE.Next()) {586 InstrumentationIRBuilder::ensureDebugInfo(*AtExit, F);587 AtExit->CreateCall(TsanFuncExit, {});588 }589 Res = true;590 }591 return Res;592}593 594bool ThreadSanitizer::instrumentLoadOrStore(const InstructionInfo &II,595 const DataLayout &DL) {596 InstrumentationIRBuilder IRB(II.Inst);597 const bool IsWrite = isa<StoreInst>(*II.Inst);598 Value *Addr = IsWrite ? cast<StoreInst>(II.Inst)->getPointerOperand()599 : cast<LoadInst>(II.Inst)->getPointerOperand();600 Type *OrigTy = getLoadStoreType(II.Inst);601 602 // swifterror memory addresses are mem2reg promoted by instruction selection.603 // As such they cannot have regular uses like an instrumentation function and604 // it makes no sense to track them as memory.605 if (Addr->isSwiftError())606 return false;607 608 int Idx = getMemoryAccessFuncIndex(OrigTy, Addr, DL);609 if (Idx < 0)610 return false;611 if (IsWrite && isVtableAccess(II.Inst)) {612 LLVM_DEBUG(dbgs() << " VPTR : " << *II.Inst << "\n");613 Value *StoredValue = cast<StoreInst>(II.Inst)->getValueOperand();614 // StoredValue may be a vector type if we are storing several vptrs at once.615 // In this case, just take the first element of the vector since this is616 // enough to find vptr races.617 if (isa<VectorType>(StoredValue->getType()))618 StoredValue = IRB.CreateExtractElement(619 StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));620 if (StoredValue->getType()->isIntegerTy())621 StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getPtrTy());622 // Call TsanVptrUpdate.623 IRB.CreateCall(TsanVptrUpdate, {Addr, StoredValue});624 NumInstrumentedVtableWrites++;625 return true;626 }627 if (!IsWrite && isVtableAccess(II.Inst)) {628 IRB.CreateCall(TsanVptrLoad, Addr);629 NumInstrumentedVtableReads++;630 return true;631 }632 633 const Align Alignment = IsWrite ? cast<StoreInst>(II.Inst)->getAlign()634 : cast<LoadInst>(II.Inst)->getAlign();635 const bool IsCompoundRW =636 ClCompoundReadBeforeWrite && (II.Flags & InstructionInfo::kCompoundRW);637 const bool IsVolatile = ClDistinguishVolatile &&638 (IsWrite ? cast<StoreInst>(II.Inst)->isVolatile()639 : cast<LoadInst>(II.Inst)->isVolatile());640 assert((!IsVolatile || !IsCompoundRW) && "Compound volatile invalid!");641 642 const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);643 FunctionCallee OnAccessFunc = nullptr;644 if (Alignment >= Align(8) || (Alignment.value() % (TypeSize / 8)) == 0) {645 if (IsCompoundRW)646 OnAccessFunc = TsanCompoundRW[Idx];647 else if (IsVolatile)648 OnAccessFunc = IsWrite ? TsanVolatileWrite[Idx] : TsanVolatileRead[Idx];649 else650 OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];651 } else {652 if (IsCompoundRW)653 OnAccessFunc = TsanUnalignedCompoundRW[Idx];654 else if (IsVolatile)655 OnAccessFunc = IsWrite ? TsanUnalignedVolatileWrite[Idx]656 : TsanUnalignedVolatileRead[Idx];657 else658 OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];659 }660 IRB.CreateCall(OnAccessFunc, Addr);661 if (IsCompoundRW || IsWrite)662 NumInstrumentedWrites++;663 if (IsCompoundRW || !IsWrite)664 NumInstrumentedReads++;665 return true;666}667 668static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {669 uint32_t v = 0;670 switch (ord) {671 case AtomicOrdering::NotAtomic:672 llvm_unreachable("unexpected atomic ordering!");673 case AtomicOrdering::Unordered: [[fallthrough]];674 case AtomicOrdering::Monotonic: v = 0; break;675 // Not specified yet:676 // case AtomicOrdering::Consume: v = 1; break;677 case AtomicOrdering::Acquire: v = 2; break;678 case AtomicOrdering::Release: v = 3; break;679 case AtomicOrdering::AcquireRelease: v = 4; break;680 case AtomicOrdering::SequentiallyConsistent: v = 5; break;681 }682 return IRB->getInt32(v);683}684 685// If a memset intrinsic gets inlined by the code gen, we will miss races on it.686// So, we either need to ensure the intrinsic is not inlined, or instrument it.687// We do not instrument memset/memmove/memcpy intrinsics (too complicated),688// instead we simply replace them with regular function calls, which are then689// intercepted by the run-time.690// Since tsan is running after everyone else, the calls should not be691// replaced back with intrinsics. If that becomes wrong at some point,692// we will need to call e.g. __tsan_memset to avoid the intrinsics.693bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {694 InstrumentationIRBuilder IRB(I);695 if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {696 Value *Cast1 = IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false);697 Value *Cast2 = IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false);698 IRB.CreateCall(699 MemsetFn,700 {M->getArgOperand(0),701 Cast1,702 Cast2});703 I->eraseFromParent();704 } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {705 IRB.CreateCall(706 isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,707 {M->getArgOperand(0),708 M->getArgOperand(1),709 IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});710 I->eraseFromParent();711 }712 return false;713}714 715// Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x716// standards. For background see C++11 standard. A slightly older, publicly717// available draft of the standard (not entirely up-to-date, but close enough718// for casual browsing) is available here:719// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf720// The following page contains more background information:721// http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/722 723bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {724 InstrumentationIRBuilder IRB(I);725 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {726 Value *Addr = LI->getPointerOperand();727 Type *OrigTy = LI->getType();728 int Idx = getMemoryAccessFuncIndex(OrigTy, Addr, DL);729 if (Idx < 0)730 return false;731 Value *Args[] = {Addr,732 createOrdering(&IRB, LI->getOrdering())};733 Value *C = IRB.CreateCall(TsanAtomicLoad[Idx], Args);734 Value *Cast = IRB.CreateBitOrPointerCast(C, OrigTy);735 I->replaceAllUsesWith(Cast);736 I->eraseFromParent();737 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {738 Value *Addr = SI->getPointerOperand();739 int Idx =740 getMemoryAccessFuncIndex(SI->getValueOperand()->getType(), Addr, DL);741 if (Idx < 0)742 return false;743 const unsigned ByteSize = 1U << Idx;744 const unsigned BitSize = ByteSize * 8;745 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);746 Value *Args[] = {Addr,747 IRB.CreateBitOrPointerCast(SI->getValueOperand(), Ty),748 createOrdering(&IRB, SI->getOrdering())};749 IRB.CreateCall(TsanAtomicStore[Idx], Args);750 SI->eraseFromParent();751 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {752 Value *Addr = RMWI->getPointerOperand();753 int Idx =754 getMemoryAccessFuncIndex(RMWI->getValOperand()->getType(), Addr, DL);755 if (Idx < 0)756 return false;757 FunctionCallee F = TsanAtomicRMW[RMWI->getOperation()][Idx];758 if (!F)759 return false;760 const unsigned ByteSize = 1U << Idx;761 const unsigned BitSize = ByteSize * 8;762 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);763 Value *Val = RMWI->getValOperand();764 Value *Args[] = {Addr, IRB.CreateBitOrPointerCast(Val, Ty),765 createOrdering(&IRB, RMWI->getOrdering())};766 Value *C = IRB.CreateCall(F, Args);767 I->replaceAllUsesWith(IRB.CreateBitOrPointerCast(C, Val->getType()));768 I->eraseFromParent();769 } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {770 Value *Addr = CASI->getPointerOperand();771 Type *OrigOldValTy = CASI->getNewValOperand()->getType();772 int Idx = getMemoryAccessFuncIndex(OrigOldValTy, Addr, DL);773 if (Idx < 0)774 return false;775 const unsigned ByteSize = 1U << Idx;776 const unsigned BitSize = ByteSize * 8;777 Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);778 Value *CmpOperand =779 IRB.CreateBitOrPointerCast(CASI->getCompareOperand(), Ty);780 Value *NewOperand =781 IRB.CreateBitOrPointerCast(CASI->getNewValOperand(), Ty);782 Value *Args[] = {Addr,783 CmpOperand,784 NewOperand,785 createOrdering(&IRB, CASI->getSuccessOrdering()),786 createOrdering(&IRB, CASI->getFailureOrdering())};787 CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);788 Value *Success = IRB.CreateICmpEQ(C, CmpOperand);789 Value *OldVal = C;790 if (Ty != OrigOldValTy) {791 // The value is a pointer, so we need to cast the return value.792 OldVal = IRB.CreateIntToPtr(C, OrigOldValTy);793 }794 795 Value *Res =796 IRB.CreateInsertValue(PoisonValue::get(CASI->getType()), OldVal, 0);797 Res = IRB.CreateInsertValue(Res, Success, 1);798 799 I->replaceAllUsesWith(Res);800 I->eraseFromParent();801 } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {802 Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};803 FunctionCallee F = FI->getSyncScopeID() == SyncScope::SingleThread804 ? TsanAtomicSignalFence805 : TsanAtomicThreadFence;806 IRB.CreateCall(F, Args);807 FI->eraseFromParent();808 }809 return true;810}811 812int ThreadSanitizer::getMemoryAccessFuncIndex(Type *OrigTy, Value *Addr,813 const DataLayout &DL) {814 assert(OrigTy->isSized());815 if (OrigTy->isScalableTy()) {816 // FIXME: support vscale.817 return -1;818 }819 uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);820 if (TypeSize != 8 && TypeSize != 16 &&821 TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {822 NumAccessesWithBadSize++;823 // Ignore all unusual sizes.824 return -1;825 }826 size_t Idx = llvm::countr_zero(TypeSize / 8);827 assert(Idx < kNumberOfAccessSizes);828 return Idx;829}830