Try to appease VS 2015 warnings from http://reviews.llvm.org/D11890
[oota-llvm.git] / lib / Transforms / Instrumentation / ThreadSanitizer.cpp
index 463ca66ca1965de571be1b93d464b0c0af4895d8..9331e1d2b3fdd1cb7f1e059ba7db98e7320ede41 100644 (file)
 // The rest is handled by the run-time library.
 //===----------------------------------------------------------------------===//
 
-#define DEBUG_TYPE "tsan"
-
 #include "llvm/Transforms/Instrumentation.h"
-#include "llvm/Transforms/Utils/BlackList.h"
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/SmallString.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/ADT/StringExtras.h"
+#include "llvm/Analysis/CaptureTracking.h"
+#include "llvm/Analysis/ValueTracking.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/Intrinsics.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/IR/Metadata.h"
@@ -45,8 +45,8 @@
 
 using namespace llvm;
 
-static cl::opt<std::string>  ClBlacklistFile("tsan-blacklist",
-       cl::desc("Blacklist file"), cl::Hidden);
+#define DEBUG_TYPE "tsan"
+
 static cl::opt<bool>  ClInstrumentMemoryAccesses(
     "tsan-instrument-memory-accesses", cl::init(true),
     cl::desc("Instrument memory accesses"), cl::Hidden);
@@ -56,6 +56,9 @@ static cl::opt<bool>  ClInstrumentFuncEntryExit(
 static cl::opt<bool>  ClInstrumentAtomics(
     "tsan-instrument-atomics", cl::init(true),
     cl::desc("Instrument atomics"), cl::Hidden);
+static cl::opt<bool>  ClInstrumentMemIntrinsics(
+    "tsan-instrument-memintrinsics", cl::init(true),
+    cl::desc("Instrument memintrinsics (memset/memcpy/memmove)"), cl::Hidden);
 
 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
@@ -63,36 +66,37 @@ STATISTIC(NumOmittedReadsBeforeWrite,
           "Number of reads ignored due to following writes");
 STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size");
 STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes");
+STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads");
 STATISTIC(NumOmittedReadsFromConstantGlobals,
           "Number of reads from constant globals");
 STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads");
+STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing");
+
+static const char *const kTsanModuleCtorName = "tsan.module_ctor";
+static const char *const kTsanInitName = "__tsan_init";
 
 namespace {
 
 /// ThreadSanitizer: instrument the code in module to find races.
 struct ThreadSanitizer : public FunctionPass {
-  ThreadSanitizer(StringRef BlacklistFile = StringRef())
-      : FunctionPass(ID),
-        TD(0),
-        BlacklistFile(BlacklistFile.empty() ? ClBlacklistFile
-                                            : BlacklistFile) { }
-  const char *getPassName() const;
-  bool runOnFunction(Function &F);
-  bool doInitialization(Module &M);
+  ThreadSanitizer() : FunctionPass(ID) {}
+  const char *getPassName() const override;
+  bool runOnFunction(Function &F) override;
+  bool doInitialization(Module &M) override;
   static char ID;  // Pass identification, replacement for typeid.
 
  private:
   void initializeCallbacks(Module &M);
-  bool instrumentLoadOrStore(Instruction *I);
-  bool instrumentAtomic(Instruction *I);
-  void chooseInstructionsToInstrument(SmallVectorImpl<Instruction*> &Local,
-                                      SmallVectorImpl<Instruction*> &All);
+  bool instrumentLoadOrStore(Instruction *I, const DataLayout &DL);
+  bool instrumentAtomic(Instruction *I, const DataLayout &DL);
+  bool instrumentMemIntrinsic(Instruction *I);
+  void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local,
+                                      SmallVectorImpl<Instruction *> &All,
+                                      const DataLayout &DL);
   bool addrPointsToConstantData(Value *Addr);
-  int getMemoryAccessFuncIndex(Value *Addr);
+  int getMemoryAccessFuncIndex(Value *Addr, const DataLayout &DL);
 
-  DataLayout *TD;
-  SmallString<64> BlacklistFile;
-  OwningPtr<BlackList> BL;
+  Type *IntptrTy;
   IntegerType *OrdTy;
   // Callbacks to run-time library are computed in doInitialization.
   Function *TsanFuncEntry;
@@ -101,6 +105,8 @@ struct ThreadSanitizer : public FunctionPass {
   static const size_t kNumberOfAccessSizes = 5;
   Function *TsanRead[kNumberOfAccessSizes];
   Function *TsanWrite[kNumberOfAccessSizes];
+  Function *TsanUnalignedRead[kNumberOfAccessSizes];
+  Function *TsanUnalignedWrite[kNumberOfAccessSizes];
   Function *TsanAtomicLoad[kNumberOfAccessSizes];
   Function *TsanAtomicStore[kNumberOfAccessSizes];
   Function *TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1][kNumberOfAccessSizes];
@@ -108,6 +114,9 @@ struct ThreadSanitizer : public FunctionPass {
   Function *TsanAtomicThreadFence;
   Function *TsanAtomicSignalFence;
   Function *TsanVptrUpdate;
+  Function *TsanVptrLoad;
+  Function *MemmoveFn, *MemcpyFn, *MemsetFn;
+  Function *TsanCtorFunction;
 };
 }  // namespace
 
@@ -120,53 +129,55 @@ const char *ThreadSanitizer::getPassName() const {
   return "ThreadSanitizer";
 }
 
-FunctionPass *llvm::createThreadSanitizerPass(StringRef BlacklistFile) {
-  return new ThreadSanitizer(BlacklistFile);
-}
-
-static Function *checkInterfaceFunction(Constant *FuncOrBitcast) {
-  if (Function *F = dyn_cast<Function>(FuncOrBitcast))
-     return F;
-  FuncOrBitcast->dump();
-  report_fatal_error("ThreadSanitizer interface function redefined");
+FunctionPass *llvm::createThreadSanitizerPass() {
+  return new ThreadSanitizer();
 }
 
 void ThreadSanitizer::initializeCallbacks(Module &M) {
   IRBuilder<> IRB(M.getContext());
   // Initialize the callbacks.
-  TsanFuncEntry = checkInterfaceFunction(M.getOrInsertFunction(
-      "__tsan_func_entry", IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
-  TsanFuncExit = checkInterfaceFunction(M.getOrInsertFunction(
-      "__tsan_func_exit", IRB.getVoidTy(), NULL));
+  TsanFuncEntry = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
+      "__tsan_func_entry", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
+  TsanFuncExit = checkSanitizerInterfaceFunction(
+      M.getOrInsertFunction("__tsan_func_exit", IRB.getVoidTy(), nullptr));
   OrdTy = IRB.getInt32Ty();
   for (size_t i = 0; i < kNumberOfAccessSizes; ++i) {
-    const size_t ByteSize = 1 << i;
-    const size_t BitSize = ByteSize * 8;
-    SmallString<32> ReadName("__tsan_read" + itostr(ByteSize));
-    TsanRead[i] = checkInterfaceFunction(M.getOrInsertFunction(
-        ReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
-
-    SmallString<32> WriteName("__tsan_write" + itostr(ByteSize));
-    TsanWrite[i] = checkInterfaceFunction(M.getOrInsertFunction(
-        WriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), NULL));
+    const unsigned ByteSize = 1U << i;
+    const unsigned BitSize = ByteSize * 8;
+    std::string ByteSizeStr = utostr(ByteSize);
+    std::string BitSizeStr = utostr(BitSize);
+    SmallString<32> ReadName("__tsan_read" + ByteSizeStr);
+    TsanRead[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
+        ReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
+
+    SmallString<32> WriteName("__tsan_write" + ByteSizeStr);
+    TsanWrite[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
+        WriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
+
+    SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr);
+    TsanUnalignedRead[i] =
+        checkSanitizerInterfaceFunction(M.getOrInsertFunction(
+            UnalignedReadName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
+
+    SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr);
+    TsanUnalignedWrite[i] =
+        checkSanitizerInterfaceFunction(M.getOrInsertFunction(
+            UnalignedWriteName, IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
 
     Type *Ty = Type::getIntNTy(M.getContext(), BitSize);
     Type *PtrTy = Ty->getPointerTo();
-    SmallString<32> AtomicLoadName("__tsan_atomic" + itostr(BitSize) +
-                                   "_load");
-    TsanAtomicLoad[i] = checkInterfaceFunction(M.getOrInsertFunction(
-        AtomicLoadName, Ty, PtrTy, OrdTy, NULL));
+    SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load");
+    TsanAtomicLoad[i] = checkSanitizerInterfaceFunction(
+        M.getOrInsertFunction(AtomicLoadName, Ty, PtrTy, OrdTy, nullptr));
 
-    SmallString<32> AtomicStoreName("__tsan_atomic" + itostr(BitSize) +
-                                    "_store");
-    TsanAtomicStore[i] = checkInterfaceFunction(M.getOrInsertFunction(
-        AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy,
-        NULL));
+    SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store");
+    TsanAtomicStore[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
+        AtomicStoreName, IRB.getVoidTy(), PtrTy, Ty, OrdTy, nullptr));
 
     for (int op = AtomicRMWInst::FIRST_BINOP;
         op <= AtomicRMWInst::LAST_BINOP; ++op) {
-      TsanAtomicRMW[op][i] = NULL;
-      const char *NamePart = NULL;
+      TsanAtomicRMW[op][i] = nullptr;
+      const char *NamePart = nullptr;
       if (op == AtomicRMWInst::Xchg)
         NamePart = "_exchange";
       else if (op == AtomicRMWInst::Add)
@@ -184,46 +195,51 @@ void ThreadSanitizer::initializeCallbacks(Module &M) {
       else
         continue;
       SmallString<32> RMWName("__tsan_atomic" + itostr(BitSize) + NamePart);
-      TsanAtomicRMW[op][i] = checkInterfaceFunction(M.getOrInsertFunction(
-          RMWName, Ty, PtrTy, Ty, OrdTy, NULL));
+      TsanAtomicRMW[op][i] = checkSanitizerInterfaceFunction(
+          M.getOrInsertFunction(RMWName, Ty, PtrTy, Ty, OrdTy, nullptr));
     }
 
-    SmallString<32> AtomicCASName("__tsan_atomic" + itostr(BitSize) +
+    SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr +
                                   "_compare_exchange_val");
-    TsanAtomicCAS[i] = checkInterfaceFunction(M.getOrInsertFunction(
-        AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, NULL));
+    TsanAtomicCAS[i] = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
+        AtomicCASName, Ty, PtrTy, Ty, Ty, OrdTy, OrdTy, nullptr));
   }
-  TsanVptrUpdate = checkInterfaceFunction(M.getOrInsertFunction(
-      "__tsan_vptr_update", IRB.getVoidTy(), IRB.getInt8PtrTy(),
-      IRB.getInt8PtrTy(), NULL));
-  TsanAtomicThreadFence = checkInterfaceFunction(M.getOrInsertFunction(
-      "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, NULL));
-  TsanAtomicSignalFence = checkInterfaceFunction(M.getOrInsertFunction(
-      "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, NULL));
+  TsanVptrUpdate = checkSanitizerInterfaceFunction(
+      M.getOrInsertFunction("__tsan_vptr_update", IRB.getVoidTy(),
+                            IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), nullptr));
+  TsanVptrLoad = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
+      "__tsan_vptr_read", IRB.getVoidTy(), IRB.getInt8PtrTy(), nullptr));
+  TsanAtomicThreadFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
+      "__tsan_atomic_thread_fence", IRB.getVoidTy(), OrdTy, nullptr));
+  TsanAtomicSignalFence = checkSanitizerInterfaceFunction(M.getOrInsertFunction(
+      "__tsan_atomic_signal_fence", IRB.getVoidTy(), OrdTy, nullptr));
+
+  MemmoveFn = checkSanitizerInterfaceFunction(
+      M.getOrInsertFunction("memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
+                            IRB.getInt8PtrTy(), IntptrTy, nullptr));
+  MemcpyFn = checkSanitizerInterfaceFunction(
+      M.getOrInsertFunction("memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
+                            IRB.getInt8PtrTy(), IntptrTy, nullptr));
+  MemsetFn = checkSanitizerInterfaceFunction(
+      M.getOrInsertFunction("memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
+                            IRB.getInt32Ty(), IntptrTy, nullptr));
 }
 
 bool ThreadSanitizer::doInitialization(Module &M) {
-  TD = getAnalysisIfAvailable<DataLayout>();
-  if (!TD)
-    return false;
-  BL.reset(new BlackList(BlacklistFile));
+  const DataLayout &DL = M.getDataLayout();
+  IntptrTy = DL.getIntPtrType(M.getContext());
+  std::tie(TsanCtorFunction, std::ignore) = createSanitizerCtorAndInitFunctions(
+      M, kTsanModuleCtorName, kTsanInitName, /*InitArgTypes=*/{},
+      /*InitArgs=*/{});
 
-  // Always insert a call to __tsan_init into the module's CTORs.
-  IRBuilder<> IRB(M.getContext());
-  Value *TsanInit = M.getOrInsertFunction("__tsan_init",
-                                          IRB.getVoidTy(), NULL);
-  appendToGlobalCtors(M, cast<Function>(TsanInit), 0);
+  appendToGlobalCtors(M, TsanCtorFunction, 0);
 
   return true;
 }
 
 static bool isVtableAccess(Instruction *I) {
-  if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa)) {
-    if (Tag->getNumOperands() < 1) return false;
-    if (MDString *Tag1 = dyn_cast<MDString>(Tag->getOperand(0))) {
-      if (Tag1->getString() == "vtable pointer") return true;
-    }
-  }
+  if (MDNode *Tag = I->getMetadata(LLVMContext::MD_tbaa))
+    return Tag->isTBAAVtableAccess();
   return false;
 }
 
@@ -251,6 +267,7 @@ bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
 // Instrumenting some of the accesses may be proven redundant.
 // Currently handled:
 //  - read-before-write (within same BB, no calls between)
+//  - not captured variables
 //
 // We do not handle some of the patterns that should not survive
 // after the classic compiler optimizations.
@@ -260,8 +277,8 @@ bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) {
 // 'Local' is a vector of insns within the same BB (no calls between).
 // 'All' is a vector of insns that will be instrumented.
 void ThreadSanitizer::chooseInstructionsToInstrument(
-    SmallVectorImpl<Instruction*> &Local,
-    SmallVectorImpl<Instruction*> &All) {
+    SmallVectorImpl<Instruction *> &Local, SmallVectorImpl<Instruction *> &All,
+    const DataLayout &DL) {
   SmallSet<Value*, 8> WriteTargets;
   // Iterate from the end.
   for (SmallVectorImpl<Instruction*>::reverse_iterator It = Local.rbegin(),
@@ -282,6 +299,17 @@ void ThreadSanitizer::chooseInstructionsToInstrument(
         continue;
       }
     }
+    Value *Addr = isa<StoreInst>(*I)
+        ? cast<StoreInst>(I)->getPointerOperand()
+        : cast<LoadInst>(I)->getPointerOperand();
+    if (isa<AllocaInst>(GetUnderlyingObject(Addr, DL)) &&
+        !PointerMayBeCaptured(Addr, true, true)) {
+      // The variable is addressable but not captured, so it cannot be
+      // referenced from a different thread and participate in a data race
+      // (see llvm/Analysis/CaptureTracking.h for details).
+      NumOmittedNonCaptured++;
+      continue;
+    }
     All.push_back(I);
   }
   Local.clear();
@@ -302,50 +330,61 @@ static bool isAtomic(Instruction *I) {
 }
 
 bool ThreadSanitizer::runOnFunction(Function &F) {
-  if (!TD) return false;
-  if (BL->isIn(F)) return false;
+  // This is required to prevent instrumenting call to __tsan_init from within
+  // the module constructor.
+  if (&F == TsanCtorFunction)
+    return false;
   initializeCallbacks(*F.getParent());
   SmallVector<Instruction*, 8> RetVec;
   SmallVector<Instruction*, 8> AllLoadsAndStores;
   SmallVector<Instruction*, 8> LocalLoadsAndStores;
   SmallVector<Instruction*, 8> AtomicAccesses;
+  SmallVector<Instruction*, 8> MemIntrinCalls;
   bool Res = false;
   bool HasCalls = false;
+  bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeThread);
+  const DataLayout &DL = F.getParent()->getDataLayout();
 
   // Traverse all instructions, collect loads/stores/returns, check for calls.
-  for (Function::iterator FI = F.begin(), FE = F.end();
-       FI != FE; ++FI) {
-    BasicBlock &BB = *FI;
-    for (BasicBlock::iterator BI = BB.begin(), BE = BB.end();
-         BI != BE; ++BI) {
-      if (isAtomic(BI))
-        AtomicAccesses.push_back(BI);
-      else if (isa<LoadInst>(BI) || isa<StoreInst>(BI))
-        LocalLoadsAndStores.push_back(BI);
-      else if (isa<ReturnInst>(BI))
-        RetVec.push_back(BI);
-      else if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
+  for (auto &BB : F) {
+    for (auto &Inst : BB) {
+      if (isAtomic(&Inst))
+        AtomicAccesses.push_back(&Inst);
+      else if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
+        LocalLoadsAndStores.push_back(&Inst);
+      else if (isa<ReturnInst>(Inst))
+        RetVec.push_back(&Inst);
+      else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
+        if (isa<MemIntrinsic>(Inst))
+          MemIntrinCalls.push_back(&Inst);
         HasCalls = true;
-        chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
+        chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores,
+                                       DL);
       }
     }
-    chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores);
+    chooseInstructionsToInstrument(LocalLoadsAndStores, AllLoadsAndStores, DL);
   }
 
   // We have collected all loads and stores.
   // FIXME: many of these accesses do not need to be checked for races
   // (e.g. variables that do not escape, etc).
 
-  // Instrument memory accesses.
-  if (ClInstrumentMemoryAccesses)
-    for (size_t i = 0, n = AllLoadsAndStores.size(); i < n; ++i) {
-      Res |= instrumentLoadOrStore(AllLoadsAndStores[i]);
+  // Instrument memory accesses only if we want to report bugs in the function.
+  if (ClInstrumentMemoryAccesses && SanitizeFunction)
+    for (auto Inst : AllLoadsAndStores) {
+      Res |= instrumentLoadOrStore(Inst, DL);
     }
 
-  // Instrument atomic memory accesses.
+  // Instrument atomic memory accesses in any case (they can be used to
+  // implement synchronization).
   if (ClInstrumentAtomics)
-    for (size_t i = 0, n = AtomicAccesses.size(); i < n; ++i) {
-      Res |= instrumentAtomic(AtomicAccesses[i]);
+    for (auto Inst : AtomicAccesses) {
+      Res |= instrumentAtomic(Inst, DL);
+    }
+
+  if (ClInstrumentMemIntrinsics && SanitizeFunction)
+    for (auto Inst : MemIntrinCalls) {
+      Res |= instrumentMemIntrinsic(Inst);
     }
 
   // Instrument function entry/exit points if there were instrumented accesses.
@@ -355,38 +394,59 @@ bool ThreadSanitizer::runOnFunction(Function &F) {
         Intrinsic::getDeclaration(F.getParent(), Intrinsic::returnaddress),
         IRB.getInt32(0));
     IRB.CreateCall(TsanFuncEntry, ReturnAddress);
-    for (size_t i = 0, n = RetVec.size(); i < n; ++i) {
-      IRBuilder<> IRBRet(RetVec[i]);
-      IRBRet.CreateCall(TsanFuncExit);
+    for (auto RetInst : RetVec) {
+      IRBuilder<> IRBRet(RetInst);
+      IRBRet.CreateCall(TsanFuncExit, {});
     }
     Res = true;
   }
   return Res;
 }
 
-bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I) {
+bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I,
+                                            const DataLayout &DL) {
   IRBuilder<> IRB(I);
   bool IsWrite = isa<StoreInst>(*I);
   Value *Addr = IsWrite
       ? cast<StoreInst>(I)->getPointerOperand()
       : cast<LoadInst>(I)->getPointerOperand();
-  int Idx = getMemoryAccessFuncIndex(Addr);
+  int Idx = getMemoryAccessFuncIndex(Addr, DL);
   if (Idx < 0)
     return false;
   if (IsWrite && isVtableAccess(I)) {
     DEBUG(dbgs() << "  VPTR : " << *I << "\n");
     Value *StoredValue = cast<StoreInst>(I)->getValueOperand();
-    // StoredValue does not necessary have a pointer type.
-    if (isa<IntegerType>(StoredValue->getType()))
+    // StoredValue may be a vector type if we are storing several vptrs at once.
+    // In this case, just take the first element of the vector since this is
+    // enough to find vptr races.
+    if (isa<VectorType>(StoredValue->getType()))
+      StoredValue = IRB.CreateExtractElement(
+          StoredValue, ConstantInt::get(IRB.getInt32Ty(), 0));
+    if (StoredValue->getType()->isIntegerTy())
       StoredValue = IRB.CreateIntToPtr(StoredValue, IRB.getInt8PtrTy());
     // Call TsanVptrUpdate.
-    IRB.CreateCall2(TsanVptrUpdate,
-                    IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
-                    IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy()));
+    IRB.CreateCall(TsanVptrUpdate,
+                   {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
+                    IRB.CreatePointerCast(StoredValue, IRB.getInt8PtrTy())});
     NumInstrumentedVtableWrites++;
     return true;
   }
-  Value *OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
+  if (!IsWrite && isVtableAccess(I)) {
+    IRB.CreateCall(TsanVptrLoad,
+                   IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
+    NumInstrumentedVtableReads++;
+    return true;
+  }
+  const unsigned Alignment = IsWrite
+      ? cast<StoreInst>(I)->getAlignment()
+      : cast<LoadInst>(I)->getAlignment();
+  Type *OrigTy = cast<PointerType>(Addr->getType())->getElementType();
+  const uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
+  Value *OnAccessFunc = nullptr;
+  if (Alignment == 0 || Alignment >= 8 || (Alignment % (TypeSize / 8)) == 0)
+    OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx];
+  else
+    OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx];
   IRB.CreateCall(OnAccessFunc, IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()));
   if (IsWrite) NumInstrumentedWrites++;
   else         NumInstrumentedReads++;
@@ -396,7 +456,7 @@ bool ThreadSanitizer::instrumentLoadOrStore(Instruction *I) {
 static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
   uint32_t v = 0;
   switch (ord) {
-    case NotAtomic:              assert(false);
+    case NotAtomic: llvm_unreachable("unexpected atomic ordering!");
     case Unordered:              // Fall-through.
     case Monotonic:              v = 0; break;
     // case Consume:                v = 1; break;  // Not specified yet.
@@ -408,116 +468,134 @@ static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
   return IRB->getInt32(v);
 }
 
-static ConstantInt *createFailOrdering(IRBuilder<> *IRB, AtomicOrdering ord) {
-  uint32_t v = 0;
-  switch (ord) {
-    case NotAtomic:              assert(false);
-    case Unordered:              // Fall-through.
-    case Monotonic:              v = 0; break;
-    // case Consume:                v = 1; break;  // Not specified yet.
-    case Acquire:                v = 2; break;
-    case Release:                v = 0; break;
-    case AcquireRelease:         v = 2; break;
-    case SequentiallyConsistent: v = 5; break;
+// If a memset intrinsic gets inlined by the code gen, we will miss races on it.
+// So, we either need to ensure the intrinsic is not inlined, or instrument it.
+// We do not instrument memset/memmove/memcpy intrinsics (too complicated),
+// instead we simply replace them with regular function calls, which are then
+// intercepted by the run-time.
+// Since tsan is running after everyone else, the calls should not be
+// replaced back with intrinsics. If that becomes wrong at some point,
+// we will need to call e.g. __tsan_memset to avoid the intrinsics.
+bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) {
+  IRBuilder<> IRB(I);
+  if (MemSetInst *M = dyn_cast<MemSetInst>(I)) {
+    IRB.CreateCall(
+        MemsetFn,
+        {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
+         IRB.CreateIntCast(M->getArgOperand(1), IRB.getInt32Ty(), false),
+         IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
+    I->eraseFromParent();
+  } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(I)) {
+    IRB.CreateCall(
+        isa<MemCpyInst>(M) ? MemcpyFn : MemmoveFn,
+        {IRB.CreatePointerCast(M->getArgOperand(0), IRB.getInt8PtrTy()),
+         IRB.CreatePointerCast(M->getArgOperand(1), IRB.getInt8PtrTy()),
+         IRB.CreateIntCast(M->getArgOperand(2), IntptrTy, false)});
+    I->eraseFromParent();
   }
-  return IRB->getInt32(v);
+  return false;
 }
 
 // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x
-// standards.  For background see C++11 standard.  A slightly older, publically
+// standards.  For background see C++11 standard.  A slightly older, publicly
 // available draft of the standard (not entirely up-to-date, but close enough
 // for casual browsing) is available here:
 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf
 // The following page contains more background information:
 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
 
-bool ThreadSanitizer::instrumentAtomic(Instruction *I) {
+bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) {
   IRBuilder<> IRB(I);
   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
     Value *Addr = LI->getPointerOperand();
-    int Idx = getMemoryAccessFuncIndex(Addr);
+    int Idx = getMemoryAccessFuncIndex(Addr, DL);
     if (Idx < 0)
       return false;
-    const size_t ByteSize = 1 << Idx;
-    const size_t BitSize = ByteSize * 8;
+    const unsigned ByteSize = 1U << Idx;
+    const unsigned BitSize = ByteSize * 8;
     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
     Type *PtrTy = Ty->getPointerTo();
     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
                      createOrdering(&IRB, LI->getOrdering())};
-    CallInst *C = CallInst::Create(TsanAtomicLoad[Idx],
-                                   ArrayRef<Value*>(Args));
+    CallInst *C = CallInst::Create(TsanAtomicLoad[Idx], Args);
     ReplaceInstWithInst(I, C);
 
   } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
     Value *Addr = SI->getPointerOperand();
-    int Idx = getMemoryAccessFuncIndex(Addr);
+    int Idx = getMemoryAccessFuncIndex(Addr, DL);
     if (Idx < 0)
       return false;
-    const size_t ByteSize = 1 << Idx;
-    const size_t BitSize = ByteSize * 8;
+    const unsigned ByteSize = 1U << Idx;
+    const unsigned BitSize = ByteSize * 8;
     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
     Type *PtrTy = Ty->getPointerTo();
     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
                      IRB.CreateIntCast(SI->getValueOperand(), Ty, false),
                      createOrdering(&IRB, SI->getOrdering())};
-    CallInst *C = CallInst::Create(TsanAtomicStore[Idx],
-                                   ArrayRef<Value*>(Args));
+    CallInst *C = CallInst::Create(TsanAtomicStore[Idx], Args);
     ReplaceInstWithInst(I, C);
   } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I)) {
     Value *Addr = RMWI->getPointerOperand();
-    int Idx = getMemoryAccessFuncIndex(Addr);
+    int Idx = getMemoryAccessFuncIndex(Addr, DL);
     if (Idx < 0)
       return false;
     Function *F = TsanAtomicRMW[RMWI->getOperation()][Idx];
-    if (F == NULL)
+    if (!F)
       return false;
-    const size_t ByteSize = 1 << Idx;
-    const size_t BitSize = ByteSize * 8;
+    const unsigned ByteSize = 1U << Idx;
+    const unsigned BitSize = ByteSize * 8;
     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
     Type *PtrTy = Ty->getPointerTo();
     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
                      IRB.CreateIntCast(RMWI->getValOperand(), Ty, false),
                      createOrdering(&IRB, RMWI->getOrdering())};
-    CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args));
+    CallInst *C = CallInst::Create(F, Args);
     ReplaceInstWithInst(I, C);
   } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(I)) {
     Value *Addr = CASI->getPointerOperand();
-    int Idx = getMemoryAccessFuncIndex(Addr);
+    int Idx = getMemoryAccessFuncIndex(Addr, DL);
     if (Idx < 0)
       return false;
-    const size_t ByteSize = 1 << Idx;
-    const size_t BitSize = ByteSize * 8;
+    const unsigned ByteSize = 1U << Idx;
+    const unsigned BitSize = ByteSize * 8;
     Type *Ty = Type::getIntNTy(IRB.getContext(), BitSize);
     Type *PtrTy = Ty->getPointerTo();
     Value *Args[] = {IRB.CreatePointerCast(Addr, PtrTy),
                      IRB.CreateIntCast(CASI->getCompareOperand(), Ty, false),
                      IRB.CreateIntCast(CASI->getNewValOperand(), Ty, false),
-                     createOrdering(&IRB, CASI->getOrdering()),
-                     createFailOrdering(&IRB, CASI->getOrdering())};
-    CallInst *C = CallInst::Create(TsanAtomicCAS[Idx], ArrayRef<Value*>(Args));
-    ReplaceInstWithInst(I, C);
+                     createOrdering(&IRB, CASI->getSuccessOrdering()),
+                     createOrdering(&IRB, CASI->getFailureOrdering())};
+    CallInst *C = IRB.CreateCall(TsanAtomicCAS[Idx], Args);
+    Value *Success = IRB.CreateICmpEQ(C, CASI->getCompareOperand());
+
+    Value *Res = IRB.CreateInsertValue(UndefValue::get(CASI->getType()), C, 0);
+    Res = IRB.CreateInsertValue(Res, Success, 1);
+
+    I->replaceAllUsesWith(Res);
+    I->eraseFromParent();
   } else if (FenceInst *FI = dyn_cast<FenceInst>(I)) {
     Value *Args[] = {createOrdering(&IRB, FI->getOrdering())};
     Function *F = FI->getSynchScope() == SingleThread ?
         TsanAtomicSignalFence : TsanAtomicThreadFence;
-    CallInst *C = CallInst::Create(F, ArrayRef<Value*>(Args));
+    CallInst *C = CallInst::Create(F, Args);
     ReplaceInstWithInst(I, C);
   }
   return true;
 }
 
-int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr) {
+int ThreadSanitizer::getMemoryAccessFuncIndex(Value *Addr,
+                                              const DataLayout &DL) {
   Type *OrigPtrTy = Addr->getType();
   Type *OrigTy = cast<PointerType>(OrigPtrTy)->getElementType();
   assert(OrigTy->isSized());
-  uint32_t TypeSize = TD->getTypeStoreSizeInBits(OrigTy);
+  uint32_t TypeSize = DL.getTypeStoreSizeInBits(OrigTy);
   if (TypeSize != 8  && TypeSize != 16 &&
       TypeSize != 32 && TypeSize != 64 && TypeSize != 128) {
     NumAccessesWithBadSize++;
     // Ignore all unusual sizes.
     return -1;
   }
-  size_t Idx = CountTrailingZeros_32(TypeSize / 8);
+  size_t Idx = countTrailingZeros(TypeSize / 8);
   assert(Idx < kNumberOfAccessSizes);
   return Idx;
 }