[tsan] First commit of ThreadSanitizer (TSan) run-time library.

Algorithm description: http://code.google.com/p/thread-sanitizer/wiki/ThreadSanitizerAlgorithm

Status:
The tool is known to work on large real-life applications, but still has quite a few rough edges.
Nothing is guaranteed yet.

The tool works on x86_64 Linux.
Support for 64-bit MacOS 10.7+ is planned for late 2012.
Support for 32-bit OSes is doable, but problematic and not yet planed.

Further commits coming:
  - tests
  - makefiles
  - documentation
  - clang driver patch

The code was previously developed at http://code.google.com/p/data-race-test/source/browse/trunk/v2/
by Dmitry Vyukov and Kostya Serebryany with contributions from
Timur Iskhodzhanov, Alexander Potapenko, Alexey Samsonov and Evgeniy Stepanov.




git-svn-id: https://llvm.org/svn/llvm-project/compiler-rt/trunk@156542 91177308-0d34-0410-b5e6-96231b3b80d8

45 files changed, 7978 insertions, 0 deletions

diff --git a/lib/tsan/rtl/tsan_allocator.cc b/lib/tsan/rtl/tsan_allocator.cc
new file mode 100644
index 000000000..4b21d1ee2
--- /dev/null
+++ b/lib/tsan/rtl/tsan_allocator.cc
@@ -0,0 +1,47 @@
+//===-- tsan_allocator-------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#include "tsan_allocator.h"
+
+// Provisional implementation.
+extern "C" void *__libc_malloc(__tsan::uptr size);
+extern "C" void __libc_free(void *ptr);
+
+namespace __tsan {
+
+u64 kBlockMagic = 0x6A6CB03ABCEBC041ull;
+
+void AllocInit() {
+}
+
+void *Alloc(uptr sz) {
+  void *p = __libc_malloc(sz + sizeof(u64));
+  ((u64*)p)[0] = kBlockMagic;
+  return (char*)p + sizeof(u64);
+}
+
+void Free(void *p) {
+  CHECK_NE(p, (char*)0);
+  p = (char*)p - sizeof(u64);
+  CHECK_EQ(((u64*)p)[0], kBlockMagic);
+  ((u64*)p)[0] = 0;
+  __libc_free(p);
+}
+
+void *AllocBlock(void *p) {
+  CHECK_NE(p, (void*)0);
+  u64 *pp = (u64*)((uptr)p & ~0x7);
+  for (; pp[0] != kBlockMagic; pp--) {}
+  return pp + 1;
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_allocator.h b/lib/tsan/rtl/tsan_allocator.h
new file mode 100644
index 000000000..7018bced2
--- /dev/null
+++ b/lib/tsan/rtl/tsan_allocator.h
@@ -0,0 +1,29 @@
+//===-- tsan_allocator.h ----------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_ALLOCATOR_H
+#define TSAN_ALLOCATOR_H
+
+#include "tsan_defs.h"
+
+namespace __tsan {
+
+void AllocInit();
+void *Alloc(uptr sz);
+void Free(void *p);  // Does not accept NULL.
+// Given the pointer p into a valid allocated block,
+// returns a pointer to the beginning of the block.
+void *AllocBlock(void *p);
+
+}  // namespace __tsan
+
+#endif  // TSAN_ALLOCATOR_H
diff --git a/lib/tsan/rtl/tsan_atomic.h b/lib/tsan/rtl/tsan_atomic.h
new file mode 100644
index 000000000..6fcd9f924
--- /dev/null
+++ b/lib/tsan/rtl/tsan_atomic.h
@@ -0,0 +1,140 @@
+//===-- tsan_rtl.h ----------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+// Atomic operations. For now implies IA-32/Intel64.
+//===----------------------------------------------------------------------===//
+
+#ifndef TSAN_ATOMIC_H
+#define TSAN_ATOMIC_H
+
+#include "tsan_defs.h"
+
+namespace __tsan {
+
+const int kCacheLineSize = 64;
+
+enum memory_order {
+  memory_order_relaxed = 1 << 0,
+  memory_order_consume = 1 << 1,
+  memory_order_acquire = 1 << 2,
+  memory_order_release = 1 << 3,
+  memory_order_acq_rel = 1 << 4,
+  memory_order_seq_cst = 1 << 5,
+};
+
+struct atomic_uint32_t {
+  typedef u32 Type;
+  volatile Type val_dont_use;
+};
+
+struct atomic_uint64_t {
+  typedef u64 Type;
+  volatile Type val_dont_use;
+};
+
+struct atomic_uintptr_t {
+  typedef uptr Type;
+  volatile Type val_dont_use;
+};
+
+INLINE void atomic_signal_fence(memory_order) {
+  __asm__ __volatile__("" ::: "memory");
+}
+
+INLINE void atomic_thread_fence(memory_order) {
+  __asm__ __volatile__("mfence" ::: "memory");
+}
+
+INLINE void proc_yield(int cnt) {
+  __asm__ __volatile__("" ::: "memory");
+  for (int i = 0; i < cnt; i++)
+    __asm__ __volatile__("pause");
+  __asm__ __volatile__("" ::: "memory");
+}
+
+template<typename T>
+INLINE typename T::Type atomic_load(
+    const volatile T *a, memory_order mo) {
+  DCHECK(mo & (memory_order_relaxed | memory_order_consume
+      | memory_order_acquire | memory_order_seq_cst));
+  DCHECK(!((uptr)a % sizeof(*a)));
+  typename T::Type v;
+  if (mo == memory_order_relaxed) {
+    v = a->val_dont_use;
+  } else {
+    atomic_signal_fence(memory_order_seq_cst);
+    v = a->val_dont_use;
+    atomic_signal_fence(memory_order_seq_cst);
+  }
+  return v;
+}
+
+template<typename T>
+INLINE void atomic_store(volatile T *a, typename T::Type v, memory_order mo) {
+  DCHECK(mo & (memory_order_relaxed | memory_order_release
+      | memory_order_seq_cst));
+  DCHECK(!((uptr)a % sizeof(*a)));
+  if (mo == memory_order_relaxed) {
+    a->val_dont_use = v;
+  } else {
+    atomic_signal_fence(memory_order_seq_cst);
+    a->val_dont_use = v;
+    atomic_signal_fence(memory_order_seq_cst);
+  }
+  if (mo == memory_order_seq_cst)
+    atomic_thread_fence(memory_order_seq_cst);
+}
+
+template<typename T>
+INLINE typename T::Type atomic_fetch_add(volatile T *a,
+    typename T::Type v, memory_order mo) {
+  (void)mo;
+  DCHECK(!((uptr)a % sizeof(*a)));
+  return __sync_fetch_and_add(&a->val_dont_use, v);
+}
+
+template<typename T>
+INLINE typename T::Type atomic_fetch_sub(volatile T *a,
+    typename T::Type v, memory_order mo) {
+  (void)mo;
+  DCHECK(!((uptr)a % sizeof(*a)));
+  return __sync_fetch_and_add(&a->val_dont_use, -v);
+}
+
+INLINE uptr atomic_exchange(volatile atomic_uintptr_t *a, uptr v,
+                            memory_order mo) {
+  __asm__ __volatile__("xchg %1, %0" : "+r"(v), "+m"(*a) : : "memory", "cc");
+  return v;
+}
+
+template<typename T>
+INLINE bool atomic_compare_exchange_strong(volatile T *a,
+                                           typename T::Type *cmp,
+                                           typename T::Type xchg,
+                                           memory_order mo) {
+  typedef typename T::Type Type;
+  Type cmpv = *cmp;
+  Type prev = __sync_val_compare_and_swap(&a->val_dont_use, cmpv, xchg);
+  if (prev == cmpv)
+    return true;
+  *cmp = prev;
+  return false;
+}
+
+INLINE bool atomic_compare_exchange_weak(volatile atomic_uintptr_t *a,
+                                         uptr *cmp, uptr xchg,
+                                         memory_order mo) {
+  return atomic_compare_exchange_strong(a, cmp, xchg, mo);
+}
+
+}  // namespace __tsan
+
+#endif  // TSAN_ATOMIC_H
diff --git a/lib/tsan/rtl/tsan_clock.cc b/lib/tsan/rtl/tsan_clock.cc
new file mode 100644
index 000000000..2b0163271
--- /dev/null
+++ b/lib/tsan/rtl/tsan_clock.cc
@@ -0,0 +1,99 @@
+//===-- tsan_clock.cc -------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#include "tsan_clock.h"
+#include "tsan_rtl.h"
+
+// It's possible to optimize clock operations for some important cases
+// so that they are O(1). The cases include singletons, once's, local mutexes.
+// First, SyncClock must be re-implemented to allow indexing by tid.
+// It must not necessarily be a full vector clock, though. For example it may
+// be a multi-level table.
+// Then, each slot in SyncClock must contain a dirty bit (it's united with
+// the clock value, so no space increase). The acquire algorithm looks
+// as follows:
+// void acquire(thr, tid, thr_clock, sync_clock) {
+//   if (!sync_clock[tid].dirty)
+//     return;  // No new info to acquire.
+//              // This handles constant reads of singleton pointers and
+//              // stop-flags.
+//   acquire_impl(thr_clock, sync_clock);  // As usual, O(N).
+//   sync_clock[tid].dirty = false;
+//   sync_clock.dirty_count--;
+// }
+// The release operation looks as follows:
+// void release(thr, tid, thr_clock, sync_clock) {
+//   // thr->sync_cache is a simple fixed-size hash-based cache that holds
+//   // several previous sync_clock's.
+//   if (thr->sync_cache[sync_clock] >= thr->last_acquire_epoch) {
+//     // The thread did no acquire operations since last release on this clock.
+//     // So update only the thread's slot (other slots can't possibly change).
+//     sync_clock[tid].clock = thr->epoch;
+//     if (sync_clock.dirty_count == sync_clock.cnt
+//         || (sync_clock.dirty_count == sync_clock.cnt - 1
+//           && sync_clock[tid].dirty == false))
+//       // All dirty flags are set, bail out.
+//       return;
+//     set all dirty bits, but preserve the thread's bit.  // O(N)
+//     update sync_clock.dirty_count;
+//     return;
+//   }
+//   release_impl(thr_clock, sync_clock);  // As usual, O(N).
+//   set all dirty bits, but preserve the thread's bit.
+//   // The previous step is combined with release_impl(), so that
+//   // we scan the arrays only once.
+//   update sync_clock.dirty_count;
+// }
+
+namespace __tsan {
+
+ThreadClock::ThreadClock() {
+  nclk_ = 0;
+  for (uptr i = 0; i < (uptr)kMaxTid; i++)
+    clk_[i] = 0;
+}
+
+void ThreadClock::acquire(const SyncClock *src) {
+  DCHECK(nclk_ <= kMaxTid);
+  DCHECK(src->clk_.Size() <= kMaxTid);
+
+  const uptr nclk = src->clk_.Size();
+  if (nclk == 0)
+    return;
+  nclk_ = max(nclk_, nclk);
+  for (uptr i = 0; i < nclk; i++) {
+    if (clk_[i] < src->clk_[i])
+      clk_[i] = src->clk_[i];
+  }
+}
+
+void ThreadClock::release(SyncClock *dst) const {
+  DCHECK(nclk_ <= kMaxTid);
+  DCHECK(dst->clk_.Size() <= kMaxTid);
+
+  if (dst->clk_.Size() < nclk_)
+    dst->clk_.Resize(nclk_);
+  for (uptr i = 0; i < nclk_; i++) {
+    if (dst->clk_[i] < clk_[i])
+      dst->clk_[i] = clk_[i];
+  }
+}
+
+void ThreadClock::acq_rel(SyncClock *dst) {
+  acquire(dst);
+  release(dst);
+}
+
+SyncClock::SyncClock()
+  : clk_(MBlockClock) {
+}
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_clock.h b/lib/tsan/rtl/tsan_clock.h
new file mode 100644
index 000000000..c4f819447
--- /dev/null
+++ b/lib/tsan/rtl/tsan_clock.h
@@ -0,0 +1,79 @@
+//===-- tsan_clock.h --------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_CLOCK_H
+#define TSAN_CLOCK_H
+
+#include "tsan_defs.h"
+#include "tsan_vector.h"
+
+namespace __tsan {
+
+// The clock that lives in sync variables (mutexes, atomics, etc).
+class SyncClock {
+ public:
+  SyncClock();
+
+  uptr size() const {
+    return clk_.Size();
+  }
+
+  void Reset() {
+    clk_.Reset();
+  }
+
+ private:
+  Vector<u64> clk_;
+  friend struct ThreadClock;
+};
+
+// The clock that lives in threads.
+struct ThreadClock {
+ public:
+  ThreadClock();
+
+  u64 get(int tid) const {
+    DCHECK(tid < kMaxTid);
+    return clk_[tid];
+  }
+
+  void set(int tid, u64 v) {
+    DCHECK(tid < kMaxTid);
+    DCHECK(v >= clk_[tid]);
+    clk_[tid] = v;
+    if ((int)nclk_ <= tid)
+      nclk_ = tid + 1;
+  }
+
+  void tick(int tid) {
+    DCHECK(tid < kMaxTid);
+    clk_[tid]++;
+    if ((int)nclk_ <= tid)
+      nclk_ = tid + 1;
+  }
+
+  uptr size() const {
+    return nclk_;
+  }
+
+  void acquire(const SyncClock *src);
+  void release(SyncClock *dst) const;
+  void acq_rel(SyncClock *dst);
+
+ private:
+  uptr nclk_;
+  u64 clk_[kMaxTid];
+};
+
+}  // namespace __tsan
+
+#endif  // TSAN_CLOCK_H
diff --git a/lib/tsan/rtl/tsan_compiler.h b/lib/tsan/rtl/tsan_compiler.h
new file mode 100644
index 000000000..6aab097d4
--- /dev/null
+++ b/lib/tsan/rtl/tsan_compiler.h
@@ -0,0 +1,30 @@
+//===-- tsan_rtl.h ----------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+// Compiler-specific definitions.
+//===----------------------------------------------------------------------===//
+
+#ifndef TSAN_COMPILER_H
+#define TSAN_COMPILER_H
+
+#define INLINE        static inline
+#define NOINLINE      __attribute__((noinline))
+#define ALWAYS_INLINE __attribute__((always_inline))
+#define NORETURN      __attribute__((noreturn))
+#define WEAK          __attribute__((weak))
+#define ALIGN(n)      __attribute__((aligned(n)))
+#define LIKELY(x)     __builtin_expect(!!(x), 1)
+#define UNLIKELY(x)   __builtin_expect(!!(x), 0)
+#define THREADLOCAL   __thread
+#define FORMAT(f, a)  __attribute__((format(printf, f, a)))
+#define USED          __attribute__((used))
+
+#endif  // TSAN_COMPILER_H
diff --git a/lib/tsan/rtl/tsan_defs.h b/lib/tsan/rtl/tsan_defs.h
new file mode 100644
index 000000000..d2088d5f3
--- /dev/null
+++ b/lib/tsan/rtl/tsan_defs.h
@@ -0,0 +1,194 @@
+//===-- tsan_defs.h ---------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef TSAN_DEFS_H
+#define TSAN_DEFS_H
+
+#include "tsan_compiler.h"
+#include "tsan_stat.h"
+
+#ifndef TSAN_DEBUG
+#define TSAN_DEBUG 0
+#endif  // TSAN_DEBUG
+
+namespace __tsan {
+
+typedef unsigned u32;  // NOLINT
+typedef unsigned long long u64;  // NOLINT
+typedef   signed long long s64;  // NOLINT
+typedef unsigned long uptr;  // NOLINT
+
+const uptr kPageSize = 4096;
+const int kTidBits = 16;
+const int kMaxTid = 1 << kTidBits;
+const int kClkBits = 40;
+
+#ifdef TSAN_SHADOW_COUNT
+# if TSAN_SHADOW_COUNT == 2 \
+  || TSAN_SHADOW_COUNT == 4 || TSAN_SHADOW_COUNT == 8
+const unsigned kShadowCnt = TSAN_SHADOW_COUNT;
+# else
+#   error "TSAN_SHADOW_COUNT must be one of 2,4,8"
+# endif
+#else
+// Count of shadow values in a shadow cell.
+const unsigned kShadowCnt = 8;
+#endif
+
+// That many user bytes are mapped onto a single shadow cell.
+const unsigned kShadowCell = 8;
+
+// Size of a single shadow value (u64).
+const unsigned kShadowSize = 8;
+
+#if defined(TSAN_COLLECT_STATS) && TSAN_COLLECT_STATS
+const bool kCollectStats = true;
+#else
+const bool kCollectStats = false;
+#endif
+
+#define CHECK_IMPL(c1, op, c2) \
+  do { \
+    __tsan::u64 v1 = (u64)(c1); \
+    __tsan::u64 v2 = (u64)(c2); \
+    if (!(v1 op v2)) \
+      __tsan::CheckFailed(__FILE__, __LINE__, \
+        "(" #c1 ") " #op " (" #c2 ")", v1, v2); \
+  } while (false) \
+/**/
+
+#define CHECK(a)       CHECK_IMPL((a), !=, 0)
+#define CHECK_EQ(a, b) CHECK_IMPL((a), ==, (b))
+#define CHECK_NE(a, b) CHECK_IMPL((a), !=, (b))
+#define CHECK_LT(a, b) CHECK_IMPL((a), <,  (b))
+#define CHECK_LE(a, b) CHECK_IMPL((a), <=, (b))
+#define CHECK_GT(a, b) CHECK_IMPL((a), >,  (b))
+#define CHECK_GE(a, b) CHECK_IMPL((a), >=, (b))
+
+#if TSAN_DEBUG
+#define DCHECK(a)       CHECK(a)
+#define DCHECK_EQ(a, b) CHECK_EQ(a, b)
+#define DCHECK_NE(a, b) CHECK_NE(a, b)
+#define DCHECK_LT(a, b) CHECK_LT(a, b)
+#define DCHECK_LE(a, b) CHECK_LE(a, b)
+#define DCHECK_GT(a, b) CHECK_GT(a, b)
+#define DCHECK_GE(a, b) CHECK_GE(a, b)
+#else
+#define DCHECK(a)
+#define DCHECK_EQ(a, b)
+#define DCHECK_NE(a, b)
+#define DCHECK_LT(a, b)
+#define DCHECK_LE(a, b)
+#define DCHECK_GT(a, b)
+#define DCHECK_GE(a, b)
+#endif
+
+void CheckFailed(const char *file, int line, const char *cond, u64 v1, u64 v2);
+
+// The following "build consistency" machinery ensures that all source files
+// are built in the same configuration. Inconsistent builds lead to
+// hard to debug crashes.
+#if TSAN_DEBUG
+void build_consistency_debug();
+#else
+void build_consistency_release();
+#endif
+
+#if TSAN_COLLECT_STATS
+void build_consistency_stats();
+#else
+void build_consistency_nostats();
+#endif
+
+#if TSAN_SHADOW_COUNT == 1
+void build_consistency_shadow1();
+#elif TSAN_SHADOW_COUNT == 2
+void build_consistency_shadow2();
+#elif TSAN_SHADOW_COUNT == 4
+void build_consistency_shadow4();
+#else
+void build_consistency_shadow8();
+#endif
+
+static inline void USED build_consistency() {
+#if TSAN_DEBUG
+  void(*volatile cfg)() = &build_consistency_debug;
+#else
+  void(*volatile cfg)() = &build_consistency_release;
+#endif
+#if TSAN_COLLECT_STATS
+  void(*volatile stats)() = &build_consistency_stats;
+#else
+  void(*volatile stats)() = &build_consistency_nostats;
+#endif
+#if TSAN_SHADOW_COUNT == 1
+  void(*volatile shadow)() = &build_consistency_shadow1;
+#elif TSAN_SHADOW_COUNT == 2
+  void(*volatile shadow)() = &build_consistency_shadow2;
+#elif TSAN_SHADOW_COUNT == 4
+  void(*volatile shadow)() = &build_consistency_shadow4;
+#else
+  void(*volatile shadow)() = &build_consistency_shadow8;
+#endif
+  (void)cfg;
+  (void)stats;
+  (void)shadow;
+}
+
+template<typename T>
+T min(T a, T b) {
+  return a < b ? a : b;
+}
+
+template<typename T>
+T max(T a, T b) {
+  return a > b ? a : b;
+}
+
+template<typename T>
+T RoundUp(T p, int align) {
+  DCHECK_EQ(align & (align - 1), 0);
+  return (T)(((u64)p + align - 1) & ~(align - 1));
+}
+
+void internal_memset(void *ptr, int c, uptr size);
+void internal_memcpy(void *dst, const void *src, uptr size);
+int internal_memcmp(const void *s1, const void *s2, uptr size);
+int internal_strcmp(const char *s1, const char *s2);
+int internal_strncmp(const char *s1, const char *s2, uptr size);
+void internal_strcpy(char *s1, const char *s2);
+uptr internal_strlen(const char *s);
+char* internal_strdup(const char *s);
+const char *internal_strstr(const char *where, const char *what);
+const char *internal_strchr(const char *where, char what);
+
+struct MD5Hash {
+  u64 hash[2];
+  bool operator==(const MD5Hash &other) const {
+    return hash[0] == other.hash[0] && hash[1] == other.hash[1];
+  }
+};
+
+MD5Hash md5_hash(const void *data, uptr size);
+
+struct ThreadState;
+struct ThreadContext;
+struct Context;
+struct ReportStack;
+class ReportDesc;
+class RegionAlloc;
+class StackTrace;
+
+}  // namespace __tsan
+
+#endif  // TSAN_DEFS_H
diff --git a/lib/tsan/rtl/tsan_flags.cc b/lib/tsan/rtl/tsan_flags.cc
new file mode 100644
index 000000000..77abd5b40
--- /dev/null
+++ b/lib/tsan/rtl/tsan_flags.cc
@@ -0,0 +1,143 @@
+//===-- tsan_flags.cc -------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#include "tsan_flags.h"
+#include "tsan_rtl.h"
+#include "tsan_mman.h"
+
+namespace __tsan {
+
+static void Flag(const char *env, bool *flag, const char *name);
+static void Flag(const char *env, int *flag, const char *name);
+static void Flag(const char *env, const char **flag, const char *name);
+
+Flags *flags() {
+  return &CTX()->flags;
+}
+
+// Can be overriden in frontend.
+void WEAK OverrideFlags(Flags *f) {
+  (void)f;
+}
+
+void InitializeFlags(Flags *f, const char *env) {
+  internal_memset(f, 0, sizeof(*f));
+
+  // Default values.
+  f->enable_annotations = true;
+  f->suppress_equal_stacks = true;
+  f->suppress_equal_addresses = true;
+  f->report_thread_leaks = true;
+  f->report_signal_unsafe = true;
+  f->force_seq_cst_atomics = false;
+  f->strip_path_prefix = internal_strdup("");
+  f->suppressions = internal_strdup("");
+  f->exitcode = 66;
+  f->log_fileno = 2;
+  f->atexit_sleep_ms = 1000;
+  f->verbosity = 0;
+
+  // Let a frontend override.
+  OverrideFlags(f);
+
+  // Override from command line.
+  Flag(env, &f->enable_annotations, "enable_annotations");
+  Flag(env, &f->suppress_equal_stacks, "suppress_equal_stacks");
+  Flag(env, &f->suppress_equal_addresses, "suppress_equal_addresses");
+  Flag(env, &f->report_thread_leaks, "report_thread_leaks");
+  Flag(env, &f->report_signal_unsafe, "report_signal_unsafe");
+  Flag(env, &f->force_seq_cst_atomics, "force_seq_cst_atomics");
+  Flag(env, &f->strip_path_prefix, "strip_path_prefix");
+  Flag(env, &f->suppressions, "suppressions");
+  Flag(env, &f->exitcode, "exitcode");
+  Flag(env, &f->log_fileno, "log_fileno");
+  Flag(env, &f->atexit_sleep_ms, "atexit_sleep_ms");
+  Flag(env, &f->verbosity, "verbosity");
+}
+
+void FinalizeFlags(Flags *flags) {
+  internal_free((void*)flags->strip_path_prefix);
+  internal_free((void*)flags->suppressions);
+}
+
+static const char *GetFlagValue(const char *env, const char *name,
+                                const char **end) {
+  if (env == 0)
+    return *end = 0;
+  const char *pos = internal_strstr(env, name);
+  if (pos == 0)
+    return *end = 0;
+  pos += internal_strlen(name);
+  if (pos[0] != '=')
+    return *end = pos;
+  pos += 1;
+  if (pos[0] == '"') {
+    pos += 1;
+    *end = internal_strchr(pos, '"');
+  } else if (pos[0] == '\'') {
+    pos += 1;
+    *end = internal_strchr(pos, '\'');
+  } else {
+    *end = internal_strchr(pos, ' ');
+  }
+  if (*end == 0)
+    *end = pos + internal_strlen(pos);
+  return pos;
+}
+
+static void Flag(const char *env, bool *flag, const char *name) {
+  const char *end = 0;
+  const char *val = GetFlagValue(env, name, &end);
+  if (val == 0)
+    return;
+  int len = end - val;
+  if (len == 1 && val[0] == '0')
+    *flag = false;
+  else if (len == 1 && val[0] == '1')
+    *flag = true;
+}
+
+static void Flag(const char *env, int *flag, const char *name) {
+  const char *end = 0;
+  const char *val = GetFlagValue(env, name, &end);
+  if (val == 0)
+    return;
+  bool minus = false;
+  if (val != end && val[0] == '-') {
+    minus = true;
+    val += 1;
+  }
+  int v = 0;
+  for (; val != end; val++) {
+    if (val[0] < '0' || val[0] > '9')
+      break;
+    v = v * 10 + val[0] - '0';
+  }
+  if (minus)
+    v = -v;
+  *flag = v;
+}
+
+static void Flag(const char *env, const char **flag, const char *name) {
+  const char *end = 0;
+  const char *val = GetFlagValue(env, name, &end);
+  if (val == 0)
+    return;
+  int len = end - val;
+  char *f = (char*)internal_alloc(MBlockFlag, len + 1);
+  internal_memcpy(f, val, len);
+  f[len] = 0;
+  *flag = f;
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_flags.h b/lib/tsan/rtl/tsan_flags.h
new file mode 100644
index 000000000..1bad405a6
--- /dev/null
+++ b/lib/tsan/rtl/tsan_flags.h
@@ -0,0 +1,56 @@
+//===-- tsan_flags.h --------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef TSAN_FLAGS_H
+#define TSAN_FLAGS_H
+
+namespace __tsan {
+
+struct Flags {
+  // Enable dynamic annotations, otherwise they are no-ops.
+  bool enable_annotations;
+  // Supress a race report if we've already output another race report
+  // with the same stack.
+  bool suppress_equal_stacks;
+  // Supress a race report if we've already output another race report
+  // on the same address.
+  bool suppress_equal_addresses;
+  // Report thread leaks at exit?
+  bool report_thread_leaks;
+  // Report violations of async signal-safety
+  // (e.g. malloc() call from a signal handler).
+  bool report_signal_unsafe;
+  // If set, all atomics are effectively sequentially consistent (seq_cst),
+  // regardless of what user actually specified.
+  bool force_seq_cst_atomics;
+  // Strip that prefix from file paths in reports.
+  const char *strip_path_prefix;
+  // Suppressions filename.
+  const char *suppressions;
+  // Override exit status if something was reported.
+  int exitcode;
+  // Log fileno (1 - stdout, 2 - stderr).
+  int log_fileno;
+  // Sleep in main thread before exiting for that many ms
+  // (useful to catch "at exit" races).
+  int atexit_sleep_ms;
+  // Verbosity level (0 - silent, 1 - a bit of output, 2+ - more output).
+  int verbosity;
+};
+
+Flags *flags();
+void InitializeFlags(Flags *flags, const char *env);
+void FinalizeFlags(Flags *flags);
+}
+
+#endif  // TSAN_FLAGS_H
diff --git a/lib/tsan/rtl/tsan_interceptors.cc b/lib/tsan/rtl/tsan_interceptors.cc
new file mode 100644
index 000000000..3d22181fb
--- /dev/null
+++ b/lib/tsan/rtl/tsan_interceptors.cc
@@ -0,0 +1,1402 @@
+//===-- tsan_interceptors_linux.cc ------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#include "interception/interception.h"
+#include "tsan_rtl.h"
+#include "tsan_interface.h"
+#include "tsan_atomic.h"
+#include "tsan_platform.h"
+#include "tsan_mman.h"
+#include "tsan_placement_new.h"
+
+using namespace __tsan;  // NOLINT
+
+struct sigset_t {
+  u64 val[kSigCount / 8 / sizeof(u64)];
+};
+
+struct ucontext_t {
+  u64 opaque[1024];
+};
+
+extern "C" int pthread_attr_init(void *attr);
+extern "C" int pthread_attr_destroy(void *attr);
+extern "C" int pthread_attr_getdetachstate(void *attr, int *v);
+extern "C" int pthread_attr_setstacksize(void *attr, uptr stacksize);
+extern "C" int pthread_attr_getstacksize(void *attr, uptr *stacksize);
+extern "C" int pthread_key_create(unsigned *key, void (*destructor)(void* v));
+extern "C" int pthread_setspecific(unsigned key, const void *v);
+extern "C" int pthread_mutexattr_gettype(void *a, int *type);
+extern "C" int pthread_yield();
+extern "C" int pthread_sigmask(int how, const sigset_t *set, sigset_t *oldset);
+extern "C" int sigfillset(sigset_t *set);
+extern "C" void *pthread_self();
+extern "C" int getcontext(ucontext_t *ucp);
+extern "C" void _exit(int status);
+extern "C" int __cxa_atexit(void (*func)(void *arg), void *arg, void *dso);
+extern "C" int *__errno_location();
+extern "C" int usleep(unsigned usec);
+const int PTHREAD_MUTEX_RECURSIVE = 1;
+const int PTHREAD_MUTEX_RECURSIVE_NP = 1;
+const int kPthreadAttrSize = 56;
+const int EINVAL = 22;
+const int EBUSY = 16;
+const int EPOLL_CTL_ADD = 1;
+void *const MAP_FAILED = (void*)-1;
+const int PTHREAD_BARRIER_SERIAL_THREAD = -1;
+const int MAP_FIXED = 0x10;
+typedef long long_t;  // NOLINT
+
+typedef void (*sighandler_t)(int sig);
+
+union pthread_attr_t {
+  char size[kPthreadAttrSize];
+  void *align;
+};
+
+struct sigaction_t {
+  union {
+    sighandler_t sa_handler;
+    void (*sa_sigaction)(int sig, my_siginfo_t *siginfo, void *uctx);
+  };
+  sigset_t sa_mask;
+  int sa_flags;
+  void (*sa_restorer)();
+};
+
+const sighandler_t SIG_DFL = (sighandler_t)0;
+const sighandler_t SIG_IGN = (sighandler_t)1;
+const int SA_SIGINFO = 4;
+const int SIG_SETMASK = 2;
+
+static sigaction_t sigactions[kSigCount];
+
+static unsigned g_thread_finalize_key;
+
+static void process_pending_signals(ThreadState *thr);
+
+class ScopedInterceptor {
+ public:
+  ScopedInterceptor(ThreadState *thr, const char *fname, uptr pc)
+      : thr_(thr)
+      , in_rtl_(thr->in_rtl) {
+    if (thr_->in_rtl == 0) {
+      Initialize(thr);
+      FuncEntry(thr, pc);
+      thr_->in_rtl++;
+      DPrintf("#%d: intercept %s()\n", thr_->tid, fname);
+    } else {
+      thr_->in_rtl++;
+    }
+  }
+
+  ~ScopedInterceptor() {
+    thr_->in_rtl--;
+    if (thr_->in_rtl == 0) {
+      FuncExit(thr_);
+      process_pending_signals(thr_);
+    }
+    CHECK_EQ(in_rtl_, thr_->in_rtl);
+  }
+
+ private:
+  ThreadState *const thr_;
+  const int in_rtl_;
+};
+
+#define SCOPED_INTERCEPTOR_RAW(func, ...) \
+    ThreadState *thr = cur_thread(); \
+    StatInc(thr, StatInterceptor); \
+    StatInc(thr, StatInt_##func); \
+    ScopedInterceptor si(thr, #func, \
+        (__tsan::uptr)__builtin_return_address(0)); \
+    const uptr pc = (uptr)&func; \
+    (void)pc; \
+/**/
+
+#define SCOPED_TSAN_INTERCEPTOR(func, ...) \
+    SCOPED_INTERCEPTOR_RAW(func, __VA_ARGS__); \
+    if (thr->in_rtl > 1) \
+      return REAL(func)(__VA_ARGS__); \
+/**/
+
+#define TSAN_INTERCEPTOR(ret, func, ...) INTERCEPTOR(ret, func, __VA_ARGS__)
+#define TSAN_INTERCEPT(func) \
+    if (!INTERCEPT_FUNCTION(func)) \
+      Printf("ThreadSanitizer: failed to intercept '" #func "' function\n"); \
+/**/
+
+class AtExitContext {
+ public:
+  AtExitContext()
+    : mtx_(MutexTypeAtExit, StatMtxAtExit)
+    , pos_() {
+  }
+
+  typedef void(*atexit_t)();
+
+  int atexit(ThreadState *thr, uptr pc, atexit_t f) {
+    Lock l(&mtx_);
+    if (pos_ == kMaxAtExit)
+      return 1;
+    Release(thr, pc, (uptr)this);
+    stack_[pos_] = f;
+    pos_++;
+    return 0;
+  }
+
+  void exit(ThreadState *thr, uptr pc) {
+    CHECK_EQ(thr->in_rtl, 0);
+    for (;;) {
+      atexit_t f = 0;
+      {
+        Lock l(&mtx_);
+        if (pos_) {
+          pos_--;
+          f = stack_[pos_];
+          ScopedInRtl in_rtl;
+          Acquire(thr, pc, (uptr)this);
+        }
+      }
+      if (f == 0)
+        break;
+      DPrintf("#%d: executing atexit func %p\n", thr->tid, f);
+      CHECK_EQ(thr->in_rtl, 0);
+      f();
+    }
+  }
+
+ private:
+  static const int kMaxAtExit = 128;
+  Mutex mtx_;
+  atexit_t stack_[kMaxAtExit];
+  int pos_;
+};
+
+static AtExitContext *atexit_ctx;
+
+static void finalize(void *arg) {
+  ThreadState * thr = cur_thread();
+  uptr pc = 0;
+  atexit_ctx->exit(thr, pc);
+  {
+    ScopedInRtl in_rtl;
+    DestroyAndFree(atexit_ctx);
+    usleep(flags()->atexit_sleep_ms * 1000);
+  }
+  int status = Finalize(cur_thread());
+  _exit(status);
+}
+
+TSAN_INTERCEPTOR(int, atexit, void (*f)()) {
+  SCOPED_TSAN_INTERCEPTOR(atexit, f);
+  return atexit_ctx->atexit(thr, pc, f);
+  return 0;
+}
+
+static uptr fd2addr(int fd) {
+  (void)fd;
+  static u64 addr;
+  return (uptr)&addr;
+}
+
+static uptr epollfd2addr(int fd) {
+  (void)fd;
+  static u64 addr;
+  return (uptr)&addr;
+}
+
+static uptr file2addr(char *path) {
+  (void)path;
+  static u64 addr;
+  return (uptr)&addr;
+}
+
+static uptr dir2addr(char *path) {
+  (void)path;
+  static u64 addr;
+  return (uptr)&addr;
+}
+
+TSAN_INTERCEPTOR(void*, malloc, uptr size) {
+  SCOPED_INTERCEPTOR_RAW(malloc, size);
+  return user_alloc(thr, pc, size);
+}
+
+TSAN_INTERCEPTOR(void*, calloc, uptr size, uptr n) {
+  SCOPED_INTERCEPTOR_RAW(calloc, size, n);
+  void *p = user_alloc(thr, pc, n * size);
+  internal_memset(p, 0, n * size);
+  return p;
+}
+
+TSAN_INTERCEPTOR(void*, realloc, void *p, uptr size) {
+  SCOPED_INTERCEPTOR_RAW(realloc, p, size);
+  return user_realloc(thr, pc, p, size);
+}
+
+TSAN_INTERCEPTOR(void, free, void *p) {
+  if (p == 0)
+    return;
+  SCOPED_INTERCEPTOR_RAW(free, p);
+  user_free(thr, pc, p);
+}
+
+TSAN_INTERCEPTOR(void, cfree, void *p) {
+  if (p == 0)
+    return;
+  SCOPED_INTERCEPTOR_RAW(cfree, p);
+  user_free(thr, pc, p);
+}
+
+TSAN_INTERCEPTOR(uptr, strlen, const void *s) {
+  SCOPED_TSAN_INTERCEPTOR(strlen, s);
+  uptr len = REAL(strlen)(s);
+  MemoryAccessRange(thr, pc, (uptr)s, len + 1, false);
+  return len;
+}
+
+TSAN_INTERCEPTOR(void*, memset, void *dst, int v, uptr size) {
+  SCOPED_TSAN_INTERCEPTOR(memset, dst, v, size);
+  MemoryAccessRange(thr, pc, (uptr)dst, size, true);
+  return REAL(memset)(dst, v, size);
+}
+
+TSAN_INTERCEPTOR(void*, memcpy, void *dst, const void *src, uptr size) {
+  SCOPED_TSAN_INTERCEPTOR(memcpy, dst, src, size);
+  MemoryAccessRange(thr, pc, (uptr)dst, size, true);
+  MemoryAccessRange(thr, pc, (uptr)src, size, false);
+  return REAL(memcpy)(dst, src, size);
+}
+
+TSAN_INTERCEPTOR(int, strcmp, const char *s1, const char *s2) {
+  SCOPED_TSAN_INTERCEPTOR(strcmp, s1, s2);
+  uptr len = 0;
+  for (; s1[len] && s2[len]; len++) {
+    if (s1[len] != s2[len])
+      break;
+  }
+  MemoryAccessRange(thr, pc, (uptr)s1, len + 1, false);
+  MemoryAccessRange(thr, pc, (uptr)s2, len + 1, false);
+  return s1[len] - s2[len];
+}
+
+TSAN_INTERCEPTOR(int, strncmp, const char *s1, const char *s2, uptr n) {
+  SCOPED_TSAN_INTERCEPTOR(strncmp, s1, s2, n);
+  uptr len = 0;
+  for (; s1[len] && s2[len] && len < n; len++) {
+    if (s1[len] != s2[len])
+      break;
+  }
+  MemoryAccessRange(thr, pc, (uptr)s1, len < n ? len + 1 : n, false);
+  MemoryAccessRange(thr, pc, (uptr)s2, len < n ? len + 1 : n, false);
+  return len == n ? 0 : s1[len] - s2[len];
+}
+
+TSAN_INTERCEPTOR(void*, memchr, void *s, int c, uptr n) {
+  SCOPED_TSAN_INTERCEPTOR(memchr, s, c, n);
+  void *res = REAL(memchr)(s, c, n);
+  uptr len = res ? (char*)res - (char*)s + 1 : n;
+  MemoryAccessRange(thr, pc, (uptr)s, len, false);
+  return res;
+}
+
+TSAN_INTERCEPTOR(void*, memrchr, char *s, int c, uptr n) {
+  SCOPED_TSAN_INTERCEPTOR(memrchr, s, c, n);
+  MemoryAccessRange(thr, pc, (uptr)s, n, false);
+  return REAL(memrchr)(s, c, n);
+}
+
+TSAN_INTERCEPTOR(void*, memmove, void *dst, void *src, uptr n) {
+  SCOPED_TSAN_INTERCEPTOR(memmove, dst, src, n);
+  MemoryAccessRange(thr, pc, (uptr)dst, n, true);
+  MemoryAccessRange(thr, pc, (uptr)src, n, false);
+  return REAL(memmove)(dst, src, n);
+}
+
+TSAN_INTERCEPTOR(int, memcmp, const void *s1, const void *s2, uptr n) {
+  SCOPED_TSAN_INTERCEPTOR(memcmp, s1, s2, n);
+  int res = 0;
+  uptr len = 0;
+  for (; len < n; len++) {
+    if ((res = ((unsigned char*)s1)[len] - ((unsigned char*)s2)[len]))
+      break;
+  }
+  MemoryAccessRange(thr, pc, (uptr)s1, len < n ? len + 1 : n, false);
+  MemoryAccessRange(thr, pc, (uptr)s2, len < n ? len + 1 : n, false);
+  return res;
+}
+
+TSAN_INTERCEPTOR(void*, strchr, void *s, int c) {
+  SCOPED_TSAN_INTERCEPTOR(strchr, s, c);
+  void *res = REAL(strchr)(s, c);
+  uptr len = res ? (char*)res - (char*)s + 1 : REAL(strlen)(s) + 1;
+  MemoryAccessRange(thr, pc, (uptr)s, len, false);
+  return res;
+}
+
+TSAN_INTERCEPTOR(void*, strchrnul, void *s, int c) {
+  SCOPED_TSAN_INTERCEPTOR(strchrnul, s, c);
+  void *res = REAL(strchrnul)(s, c);
+  uptr len = (char*)res - (char*)s + 1;
+  MemoryAccessRange(thr, pc, (uptr)s, len, false);
+  return res;
+}
+
+TSAN_INTERCEPTOR(void*, strrchr, void *s, int c) {
+  SCOPED_TSAN_INTERCEPTOR(strrchr, s, c);
+  MemoryAccessRange(thr, pc, (uptr)s, REAL(strlen)(s) + 1, false);
+  return REAL(strrchr)(s, c);
+}
+
+TSAN_INTERCEPTOR(void*, strcpy, void *dst, const void *src) {  // NOLINT
+  SCOPED_TSAN_INTERCEPTOR(strcpy, dst, src);  // NOLINT
+  uptr srclen = REAL(strlen)(src);
+  MemoryAccessRange(thr, pc, (uptr)dst, srclen + 1, true);
+  MemoryAccessRange(thr, pc, (uptr)src, srclen + 1, false);
+  return REAL(strcpy)(dst, src);  // NOLINT
+}
+
+TSAN_INTERCEPTOR(void*, strncpy, void *dst, void *src, uptr n) {
+  SCOPED_TSAN_INTERCEPTOR(strncpy, dst, src, n);
+  uptr srclen = REAL(strlen)(src);
+  MemoryAccessRange(thr, pc, (uptr)dst, n, true);
+  MemoryAccessRange(thr, pc, (uptr)src, min(srclen + 1, n), false);
+  return REAL(strncpy)(dst, src, n);
+}
+
+TSAN_INTERCEPTOR(const char*, strstr, const char *s1, const char *s2) {
+  SCOPED_TSAN_INTERCEPTOR(strstr, s1, s2);
+  const char *res = REAL(strstr)(s1, s2);
+  uptr len1 = REAL(strlen)(s1);
+  uptr len2 = REAL(strlen)(s2);
+  MemoryAccessRange(thr, pc, (uptr)s1, len1 + 1, false);
+  MemoryAccessRange(thr, pc, (uptr)s2, len2 + 1, false);
+  return res;
+}
+
+static bool fix_mmap_addr(void **addr, long_t sz, int flags) {
+  if (*addr) {
+    if (!IsAppMem((uptr)*addr) || !IsAppMem((uptr)*addr + sz - 1)) {
+      if (flags & MAP_FIXED) {
+        *__errno_location() = EINVAL;
+        return false;
+      } else {
+        *addr = 0;
+      }
+    }
+  }
+  return true;
+}
+
+TSAN_INTERCEPTOR(void*, mmap, void *addr, long_t sz, int prot,
+                         int flags, int fd, unsigned off) {
+  SCOPED_TSAN_INTERCEPTOR(mmap, addr, sz, prot, flags, fd, off);
+  if (!fix_mmap_addr(&addr, sz, flags))
+    return MAP_FAILED;
+  void *res = REAL(mmap)(addr, sz, prot, flags, fd, off);
+  if (res != MAP_FAILED) {
+    MemoryResetRange(thr, pc, (uptr)res, sz);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(void*, mmap64, void *addr, long_t sz, int prot,
+                           int flags, int fd, u64 off) {
+  SCOPED_TSAN_INTERCEPTOR(mmap64, addr, sz, prot, flags, fd, off);
+  if (!fix_mmap_addr(&addr, sz, flags))
+    return MAP_FAILED;
+  void *res = REAL(mmap64)(addr, sz, prot, flags, fd, off);
+  if (res != MAP_FAILED) {
+    MemoryResetRange(thr, pc, (uptr)res, sz);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, munmap, void *addr, long_t sz) {
+  SCOPED_TSAN_INTERCEPTOR(munmap, addr, sz);
+  int res = REAL(munmap)(addr, sz);
+  return res;
+}
+
+#ifdef __LP64__
+
+// void *operator new(size_t)
+TSAN_INTERCEPTOR(void*, _Znwm, uptr sz) {
+  SCOPED_TSAN_INTERCEPTOR(_Znwm, sz);
+  return user_alloc(thr, pc, sz);
+}
+
+// void *operator new(size_t, nothrow_t)
+TSAN_INTERCEPTOR(void*, _ZnwmRKSt9nothrow_t, uptr sz) {
+  SCOPED_TSAN_INTERCEPTOR(_ZnwmRKSt9nothrow_t, sz);
+  return user_alloc(thr, pc, sz);
+}
+
+// void *operator new[](size_t)
+TSAN_INTERCEPTOR(void*, _Znam, uptr sz) {
+  SCOPED_TSAN_INTERCEPTOR(_Znam, sz);
+  return user_alloc(thr, pc, sz);
+}
+
+// void *operator new[](size_t, nothrow_t)
+TSAN_INTERCEPTOR(void*, _ZnamRKSt9nothrow_t, uptr sz) {
+  SCOPED_TSAN_INTERCEPTOR(_ZnamRKSt9nothrow_t, sz);
+  return user_alloc(thr, pc, sz);
+}
+
+#else
+#error "Not implemented"
+#endif
+
+// void operator delete(void*)
+TSAN_INTERCEPTOR(void, _ZdlPv, void *p) {
+  if (p == 0)
+    return;
+  SCOPED_TSAN_INTERCEPTOR(_ZdlPv, p);
+  user_free(thr, pc, p);
+}
+
+// void operator delete(void*, nothrow_t)
+TSAN_INTERCEPTOR(void, _ZdlPvRKSt9nothrow_t, void *p) {
+  if (p == 0)
+    return;
+  SCOPED_TSAN_INTERCEPTOR(_ZdlPvRKSt9nothrow_t, p);
+  user_free(thr, pc, p);
+}
+
+// void operator delete[](void*)
+TSAN_INTERCEPTOR(void, _ZdaPv, void *p) {
+  if (p == 0)
+    return;
+  SCOPED_TSAN_INTERCEPTOR(_ZdaPv, p);
+  user_free(thr, pc, p);
+}
+
+// void operator delete[](void*, nothrow_t)
+TSAN_INTERCEPTOR(void, _ZdaPvRKSt9nothrow_t, void *p) {
+  if (p == 0)
+    return;
+  SCOPED_TSAN_INTERCEPTOR(_ZdaPvRKSt9nothrow_t, p);
+  user_free(thr, pc, p);
+}
+
+TSAN_INTERCEPTOR(void*, memalign, uptr align, uptr sz) {
+  SCOPED_TSAN_INTERCEPTOR(memalign, align, sz);
+  return user_alloc_aligned(thr, pc, sz, align);
+}
+
+TSAN_INTERCEPTOR(void*, valloc, uptr sz) {
+  SCOPED_TSAN_INTERCEPTOR(valloc, sz);
+  return user_alloc_aligned(thr, pc, sz, kPageSize);
+}
+
+TSAN_INTERCEPTOR(void*, pvalloc, uptr sz) {
+  SCOPED_TSAN_INTERCEPTOR(pvalloc, sz);
+  sz = RoundUp(sz, kPageSize);
+  return user_alloc_aligned(thr, pc, sz, kPageSize);
+}
+
+TSAN_INTERCEPTOR(int, posix_memalign, void **memptr, uptr align, uptr sz) {
+  SCOPED_TSAN_INTERCEPTOR(posix_memalign, memptr, align, sz);
+  *memptr = user_alloc_aligned(thr, pc, sz, align);
+  return 0;
+}
+
+// Used in thread-safe function static initialization.
+TSAN_INTERCEPTOR(int, __cxa_guard_acquire, char *m) {
+  SCOPED_TSAN_INTERCEPTOR(__cxa_guard_acquire, m);
+  int res = REAL(__cxa_guard_acquire)(m);
+  if (res) {
+    // This thread does the init.
+  } else {
+    Acquire(thr, pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(void, __cxa_guard_release, char *m) {
+  SCOPED_TSAN_INTERCEPTOR(__cxa_guard_release, m);
+  Release(thr, pc, (uptr)m);
+  REAL(__cxa_guard_release)(m);
+}
+
+static void thread_finalize(void *v) {
+  uptr iter = (uptr)v;
+  if (iter > 1) {
+    if (pthread_setspecific(g_thread_finalize_key, (void*)(iter - 1))) {
+      Printf("ThreadSanitizer: failed to set thread key\n");
+      Die();
+    }
+    return;
+  }
+  {
+    ScopedInRtl in_rtl;
+    ThreadFinish(cur_thread());
+  }
+}
+
+
+struct ThreadParam {
+  void* (*callback)(void *arg);
+  void *param;
+  atomic_uintptr_t tid;
+};
+
+extern "C" void *__tsan_thread_start_func(void *arg) {
+  ThreadParam *p = (ThreadParam*)arg;
+  void* (*callback)(void *arg) = p->callback;
+  void *param = p->param;
+  int tid = 0;
+  {
+    ThreadState *thr = cur_thread();
+    ScopedInRtl in_rtl;
+    if (pthread_setspecific(g_thread_finalize_key, (void*)4)) {
+      Printf("ThreadSanitizer: failed to set thread key\n");
+      Die();
+    }
+    while ((tid = atomic_load(&p->tid, memory_order_acquire)) == 0)
+      pthread_yield();
+    atomic_store(&p->tid, 0, memory_order_release);
+    ThreadStart(thr, tid);
+    CHECK_EQ(thr->in_rtl, 1);
+  }
+  void *res = callback(param);
+  // Prevent the callback from being tail called,
+  // it mixes up stack traces.
+  volatile int foo = 42;
+  foo++;
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_create,
+    void *th, void *attr, void *(*callback)(void*), void * param) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_create, th, attr, callback, param);
+  pthread_attr_t myattr;
+  if (attr == 0) {
+    pthread_attr_init(&myattr);
+    attr = &myattr;
+  }
+  int detached = 0;
+  pthread_attr_getdetachstate(attr, &detached);
+  uptr stacksize = 0;
+  pthread_attr_getstacksize(attr, &stacksize);
+  // We place the huge ThreadState object into TLS, account for that.
+  const uptr minstacksize = GetTlsSize() + 128*1024;
+  if (stacksize < minstacksize) {
+    DPrintf("ThreadSanitizer: stacksize %lu->%lu\n", stacksize, minstacksize);
+    pthread_attr_setstacksize(attr, minstacksize);
+  }
+  ThreadParam p;
+  p.callback = callback;
+  p.param = param;
+  atomic_store(&p.tid, 0, memory_order_relaxed);
+  int res = REAL(pthread_create)(th, attr, __tsan_thread_start_func, &p);
+  if (res == 0) {
+    int tid = ThreadCreate(cur_thread(), pc, *(uptr*)th, detached);
+    CHECK_NE(tid, 0);
+    atomic_store(&p.tid, tid, memory_order_release);
+    while (atomic_load(&p.tid, memory_order_acquire) != 0)
+      pthread_yield();
+  }
+  if (attr == &myattr)
+    pthread_attr_destroy(&myattr);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_join, void *th, void **ret) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_join, th, ret);
+  int tid = ThreadTid(thr, pc, (uptr)th);
+  int res = REAL(pthread_join)(th, ret);
+  if (res == 0) {
+    ThreadJoin(cur_thread(), pc, tid);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_detach, void *th) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_detach, th);
+  int tid = ThreadTid(thr, pc, (uptr)th);
+  int res = REAL(pthread_detach)(th);
+  if (res == 0) {
+    ThreadDetach(cur_thread(), pc, tid);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_mutex_init, void *m, void *a) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_mutex_init, m, a);
+  int res = REAL(pthread_mutex_init)(m, a);
+  if (res == 0) {
+    bool recursive = false;
+    if (a) {
+      int type = 0;
+      if (pthread_mutexattr_gettype(a, &type) == 0)
+        recursive = (type == PTHREAD_MUTEX_RECURSIVE
+            || type == PTHREAD_MUTEX_RECURSIVE_NP);
+    }
+    MutexCreate(cur_thread(), pc, (uptr)m, false, recursive);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_mutex_destroy, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_mutex_destroy, m);
+  int res = REAL(pthread_mutex_destroy)(m);
+  if (res == 0 || res == EBUSY) {
+    MutexDestroy(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_mutex_lock, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_mutex_lock, m);
+  int res = REAL(pthread_mutex_lock)(m);
+  if (res == 0) {
+    MutexLock(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_mutex_trylock, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_mutex_trylock, m);
+  int res = REAL(pthread_mutex_trylock)(m);
+  if (res == 0) {
+    MutexLock(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_mutex_timedlock, void *m, void *abstime) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_mutex_timedlock, m, abstime);
+  int res = REAL(pthread_mutex_timedlock)(m, abstime);
+  if (res == 0) {
+    MutexLock(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_mutex_unlock, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_mutex_unlock, m);
+  MutexUnlock(cur_thread(), pc, (uptr)m);
+  int res = REAL(pthread_mutex_unlock)(m);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_spin_init, void *m, int pshared) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_spin_init, m, pshared);
+  int res = REAL(pthread_spin_init)(m, pshared);
+  if (res == 0) {
+    MutexCreate(cur_thread(), pc, (uptr)m, false, false);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_spin_destroy, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_spin_destroy, m);
+  int res = REAL(pthread_spin_destroy)(m);
+  if (res == 0) {
+    MutexDestroy(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_spin_lock, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_spin_lock, m);
+  int res = REAL(pthread_spin_lock)(m);
+  if (res == 0) {
+    MutexLock(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_spin_trylock, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_spin_trylock, m);
+  int res = REAL(pthread_spin_trylock)(m);
+  if (res == 0) {
+    MutexLock(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_spin_unlock, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_spin_unlock, m);
+  MutexUnlock(cur_thread(), pc, (uptr)m);
+  int res = REAL(pthread_spin_unlock)(m);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_rwlock_init, void *m, void *a) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_rwlock_init, m, a);
+  int res = REAL(pthread_rwlock_init)(m, a);
+  if (res == 0) {
+    MutexCreate(cur_thread(), pc, (uptr)m, true, false);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_rwlock_destroy, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_rwlock_destroy, m);
+  int res = REAL(pthread_rwlock_destroy)(m);
+  if (res == 0) {
+    MutexDestroy(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_rwlock_rdlock, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_rwlock_rdlock, m);
+  int res = REAL(pthread_rwlock_rdlock)(m);
+  if (res == 0) {
+    MutexReadLock(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_rwlock_tryrdlock, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_rwlock_tryrdlock, m);
+  int res = REAL(pthread_rwlock_tryrdlock)(m);
+  if (res == 0) {
+    MutexReadLock(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_rwlock_timedrdlock, void *m, void *abstime) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_rwlock_timedrdlock, m, abstime);
+  int res = REAL(pthread_rwlock_timedrdlock)(m, abstime);
+  if (res == 0) {
+    MutexReadLock(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_rwlock_wrlock, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_rwlock_wrlock, m);
+  int res = REAL(pthread_rwlock_wrlock)(m);
+  if (res == 0) {
+    MutexLock(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_rwlock_trywrlock, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_rwlock_trywrlock, m);
+  int res = REAL(pthread_rwlock_trywrlock)(m);
+  if (res == 0) {
+    MutexLock(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_rwlock_timedwrlock, void *m, void *abstime) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_rwlock_timedwrlock, m, abstime);
+  int res = REAL(pthread_rwlock_timedwrlock)(m, abstime);
+  if (res == 0) {
+    MutexLock(cur_thread(), pc, (uptr)m);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_rwlock_unlock, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_rwlock_unlock, m);
+  MutexReadOrWriteUnlock(cur_thread(), pc, (uptr)m);
+  int res = REAL(pthread_rwlock_unlock)(m);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_cond_init, void *c, void *a) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_cond_init, c, a);
+  int res = REAL(pthread_cond_init)(c, a);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_cond_destroy, void *c) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_cond_destroy, c);
+  int res = REAL(pthread_cond_destroy)(c);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_cond_signal, void *c) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_cond_signal, c);
+  int res = REAL(pthread_cond_signal)(c);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_cond_broadcast, void *c) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_cond_broadcast, c);
+  int res = REAL(pthread_cond_broadcast)(c);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_cond_wait, void *c, void *m) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_cond_wait, c, m);
+  MutexUnlock(cur_thread(), pc, (uptr)m);
+  int res = REAL(pthread_cond_wait)(c, m);
+  MutexLock(cur_thread(), pc, (uptr)m);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_cond_timedwait, void *c, void *m, void *abstime) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_cond_timedwait, c, m, abstime);
+  MutexUnlock(cur_thread(), pc, (uptr)m);
+  int res = REAL(pthread_cond_timedwait)(c, m, abstime);
+  MutexLock(cur_thread(), pc, (uptr)m);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_barrier_init, void *b, void *a, unsigned count) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_barrier_init, b, a, count);
+  MemoryWrite1Byte(thr, pc, (uptr)b);
+  int res = REAL(pthread_barrier_init)(b, a, count);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_barrier_destroy, void *b) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_barrier_destroy, b);
+  MemoryWrite1Byte(thr, pc, (uptr)b);
+  int res = REAL(pthread_barrier_destroy)(b);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_barrier_wait, void *b) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_barrier_wait, b);
+  Release(cur_thread(), pc, (uptr)b);
+  MemoryRead1Byte(thr, pc, (uptr)b);
+  int res = REAL(pthread_barrier_wait)(b);
+  MemoryRead1Byte(thr, pc, (uptr)b);
+  if (res == 0 || res == PTHREAD_BARRIER_SERIAL_THREAD) {
+    Acquire(cur_thread(), pc, (uptr)b);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, pthread_once, void *o, void (*f)()) {
+  SCOPED_TSAN_INTERCEPTOR(pthread_once, o, f);
+  if (o == 0 || f == 0)
+    return EINVAL;
+  atomic_uint32_t *a = static_cast<atomic_uint32_t*>(o);
+  u32 v = atomic_load(a, memory_order_acquire);
+  if (v == 0 && atomic_compare_exchange_strong(a, &v, 1,
+                                               memory_order_relaxed)) {
+    const int old_in_rtl = thr->in_rtl;
+    thr->in_rtl = 0;
+    (*f)();
+    CHECK_EQ(thr->in_rtl, 0);
+    thr->in_rtl = old_in_rtl;
+    Release(cur_thread(), pc, (uptr)o);
+    atomic_store(a, 2, memory_order_release);
+  } else {
+    while (v != 2) {
+      pthread_yield();
+      v = atomic_load(a, memory_order_acquire);
+    }
+    Acquire(cur_thread(), pc, (uptr)o);
+  }
+  return 0;
+}
+
+TSAN_INTERCEPTOR(int, sem_init, void *s, int pshared, unsigned value) {
+  SCOPED_TSAN_INTERCEPTOR(sem_init, s, pshared, value);
+  int res = REAL(sem_init)(s, pshared, value);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, sem_destroy, void *s) {
+  SCOPED_TSAN_INTERCEPTOR(sem_destroy, s);
+  int res = REAL(sem_destroy)(s);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, sem_wait, void *s) {
+  SCOPED_TSAN_INTERCEPTOR(sem_wait, s);
+  int res = REAL(sem_wait)(s);
+  if (res == 0) {
+    Acquire(cur_thread(), pc, (uptr)s);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, sem_trywait, void *s) {
+  SCOPED_TSAN_INTERCEPTOR(sem_trywait, s);
+  int res = REAL(sem_trywait)(s);
+  if (res == 0) {
+    Acquire(cur_thread(), pc, (uptr)s);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, sem_timedwait, void *s, void *abstime) {
+  SCOPED_TSAN_INTERCEPTOR(sem_timedwait, s, abstime);
+  int res = REAL(sem_timedwait)(s, abstime);
+  if (res == 0) {
+    Acquire(cur_thread(), pc, (uptr)s);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, sem_post, void *s) {
+  SCOPED_TSAN_INTERCEPTOR(sem_post, s);
+  Release(cur_thread(), pc, (uptr)s);
+  int res = REAL(sem_post)(s);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, sem_getvalue, void *s, int *sval) {
+  SCOPED_TSAN_INTERCEPTOR(sem_getvalue, s, sval);
+  int res = REAL(sem_getvalue)(s, sval);
+  if (res == 0) {
+    Acquire(cur_thread(), pc, (uptr)s);
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, read, int fd, void *buf, long_t sz) {
+  SCOPED_TSAN_INTERCEPTOR(read, fd, buf, sz);
+  int res = REAL(read)(fd, buf, sz);
+  if (res >= 0) {
+    Acquire(cur_thread(), pc, fd2addr(fd));
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, pread, int fd, void *buf, long_t sz, unsigned off) {
+  SCOPED_TSAN_INTERCEPTOR(pread, fd, buf, sz, off);
+  int res = REAL(pread)(fd, buf, sz, off);
+  if (res >= 0) {
+    Acquire(cur_thread(), pc, fd2addr(fd));
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, pread64, int fd, void *buf, long_t sz, u64 off) {
+  SCOPED_TSAN_INTERCEPTOR(pread64, fd, buf, sz, off);
+  int res = REAL(pread64)(fd, buf, sz, off);
+  if (res >= 0) {
+    Acquire(cur_thread(), pc, fd2addr(fd));
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, readv, int fd, void *vec, int cnt) {
+  SCOPED_TSAN_INTERCEPTOR(readv, fd, vec, cnt);
+  int res = REAL(readv)(fd, vec, cnt);
+  if (res >= 0) {
+    Acquire(cur_thread(), pc, fd2addr(fd));
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, preadv64, int fd, void *vec, int cnt, u64 off) {
+  SCOPED_TSAN_INTERCEPTOR(preadv64, fd, vec, cnt, off);
+  int res = REAL(preadv64)(fd, vec, cnt, off);
+  if (res >= 0) {
+    Acquire(cur_thread(), pc, fd2addr(fd));
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, write, int fd, void *buf, long_t sz) {
+  SCOPED_TSAN_INTERCEPTOR(write, fd, buf, sz);
+  Release(cur_thread(), pc, fd2addr(fd));
+  int res = REAL(write)(fd, buf, sz);
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, pwrite, int fd, void *buf, long_t sz, unsigned off) {
+  SCOPED_TSAN_INTERCEPTOR(pwrite, fd, buf, sz, off);
+  Release(cur_thread(), pc, fd2addr(fd));
+  int res = REAL(pwrite)(fd, buf, sz, off);
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, pwrite64, int fd, void *buf, long_t sz, unsigned off) {
+  SCOPED_TSAN_INTERCEPTOR(pwrite64, fd, buf, sz, off);
+  Release(cur_thread(), pc, fd2addr(fd));
+  int res = REAL(pwrite64)(fd, buf, sz, off);
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, writev, int fd, void *vec, int cnt) {
+  SCOPED_TSAN_INTERCEPTOR(writev, fd, vec, cnt);
+  Release(cur_thread(), pc, fd2addr(fd));
+  int res = REAL(writev)(fd, vec, cnt);
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, pwritev64, int fd, void *vec, int cnt, u64 off) {
+  SCOPED_TSAN_INTERCEPTOR(pwritev64, fd, vec, cnt, off);
+  Release(cur_thread(), pc, fd2addr(fd));
+  int res = REAL(pwritev64)(fd, vec, cnt, off);
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, send, int fd, void *buf, long_t len, int flags) {
+  SCOPED_TSAN_INTERCEPTOR(send, fd, buf, len, flags);
+  Release(cur_thread(), pc, fd2addr(fd));
+  int res = REAL(send)(fd, buf, len, flags);
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, sendmsg, int fd, void *msg, int flags) {
+  SCOPED_TSAN_INTERCEPTOR(sendmsg, fd, msg, flags);
+  Release(cur_thread(), pc, fd2addr(fd));
+  int res = REAL(sendmsg)(fd, msg, flags);
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, recv, int fd, void *buf, long_t len, int flags) {
+  SCOPED_TSAN_INTERCEPTOR(recv, fd, buf, len, flags);
+  int res = REAL(recv)(fd, buf, len, flags);
+  if (res >= 0) {
+    Acquire(cur_thread(), pc, fd2addr(fd));
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(long_t, recvmsg, int fd, void *msg, int flags) {
+  SCOPED_TSAN_INTERCEPTOR(recvmsg, fd, msg, flags);
+  int res = REAL(recvmsg)(fd, msg, flags);
+  if (res >= 0) {
+    Acquire(cur_thread(), pc, fd2addr(fd));
+  }
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, unlink, char *path) {
+  SCOPED_TSAN_INTERCEPTOR(unlink, path);
+  Release(cur_thread(), pc, file2addr(path));
+  int res = REAL(unlink)(path);
+  return res;
+}
+
+TSAN_INTERCEPTOR(void*, fopen, char *path, char *mode) {
+  SCOPED_TSAN_INTERCEPTOR(fopen, path, mode);
+  void *res = REAL(fopen)(path, mode);
+  Acquire(cur_thread(), pc, file2addr(path));
+  return res;
+}
+
+TSAN_INTERCEPTOR(uptr, fread, void *ptr, uptr size, uptr nmemb, void *f) {
+  SCOPED_TSAN_INTERCEPTOR(fread, ptr, size, nmemb, f);
+  MemoryAccessRange(thr, pc, (uptr)ptr, size * nmemb, true);
+  return REAL(fread)(ptr, size, nmemb, f);
+}
+
+TSAN_INTERCEPTOR(uptr, fwrite, const void *p, uptr size, uptr nmemb, void *f) {
+  SCOPED_TSAN_INTERCEPTOR(fwrite, p, size, nmemb, f);
+  MemoryAccessRange(thr, pc, (uptr)p, size * nmemb, false);
+  return REAL(fwrite)(p, size, nmemb, f);
+}
+
+TSAN_INTERCEPTOR(int, puts, const char *s) {
+  SCOPED_TSAN_INTERCEPTOR(puts, s);
+  MemoryAccessRange(thr, pc, (uptr)s, REAL(strlen)(s), false);
+  return REAL(puts)(s);
+}
+
+TSAN_INTERCEPTOR(int, rmdir, char *path) {
+  SCOPED_TSAN_INTERCEPTOR(rmdir, path);
+  Release(cur_thread(), pc, dir2addr(path));
+  int res = REAL(rmdir)(path);
+  return res;
+}
+
+TSAN_INTERCEPTOR(void*, opendir, char *path) {
+  SCOPED_TSAN_INTERCEPTOR(opendir, path);
+  void *res = REAL(opendir)(path);
+  Acquire(cur_thread(), pc, dir2addr(path));
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, epoll_ctl, int epfd, int op, int fd, void *ev) {
+  SCOPED_TSAN_INTERCEPTOR(epoll_ctl, epfd, op, fd, ev);
+  if (op == EPOLL_CTL_ADD) {
+    Release(cur_thread(), pc, epollfd2addr(epfd));
+  }
+  int res = REAL(epoll_ctl)(epfd, op, fd, ev);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, epoll_wait, int epfd, void *ev, int cnt, int timeout) {
+  SCOPED_TSAN_INTERCEPTOR(epoll_wait, epfd, ev, cnt, timeout);
+  int res = REAL(epoll_wait)(epfd, ev, cnt, timeout);
+  if (res > 0) {
+    Acquire(cur_thread(), pc, epollfd2addr(epfd));
+  }
+  return res;
+}
+
+static void rtl_sighandler(int sig) {
+  ThreadState *thr = cur_thread();
+  SignalDesc *signal = &thr->pending_signals[sig];
+  if (signal->armed == false) {
+    signal->armed = true;
+    signal->sigaction = false;
+    thr->pending_signal_count++;
+  }
+}
+
+static void rtl_sigaction(int sig, my_siginfo_t *info, void *ctx) {
+  ThreadState *thr = cur_thread();
+  SignalDesc *signal = &thr->pending_signals[sig];
+  if (signal->armed == false) {
+    signal->armed = true;
+    signal->sigaction = true;
+    signal->siginfo = *info;
+    thr->pending_signal_count++;
+  }
+}
+
+TSAN_INTERCEPTOR(int, sigaction, int sig, sigaction_t *act, sigaction_t *old) {
+  SCOPED_TSAN_INTERCEPTOR(sigaction, sig, act, old);
+  int res = 0;
+  if (act == 0 || act->sa_handler == SIG_IGN || act->sa_handler == SIG_DFL) {
+    res = REAL(sigaction)(sig, act, old);
+  } else {
+    sigactions[sig] = *act;
+    sigaction_t newact = *act;
+    if (newact.sa_flags & SA_SIGINFO)
+      newact.sa_sigaction = rtl_sigaction;
+    else
+      newact.sa_handler = rtl_sighandler;
+    res = REAL(sigaction)(sig, &newact, old);
+  }
+  return res;
+}
+
+static void process_pending_signals(ThreadState *thr) {
+  CHECK_EQ(thr->in_rtl, 0);
+  if (thr->pending_signal_count == 0 || thr->in_signal_handler)
+    return;
+  thr->in_signal_handler = true;
+  thr->pending_signal_count = 0;
+  // These are too big for stack.
+  static THREADLOCAL ucontext_t uctx;
+  static THREADLOCAL sigset_t emptyset, oldset;
+  getcontext(&uctx);
+  sigfillset(&emptyset);
+  pthread_sigmask(SIG_SETMASK, &emptyset, &oldset);
+  for (int sig = 0; sig < kSigCount; sig++) {
+    SignalDesc *signal = &thr->pending_signals[sig];
+    if (signal->armed) {
+      signal->armed = false;
+      if (signal->sigaction)
+        sigactions[sig].sa_sigaction(sig, &signal->siginfo, &uctx);
+      else
+        sigactions[sig].sa_handler(sig);
+    }
+  }
+  pthread_sigmask(SIG_SETMASK, &oldset, 0);
+  CHECK_EQ(thr->in_signal_handler, true);
+  thr->in_signal_handler = false;
+}
+
+namespace __tsan {
+
+// Used until we obtain real efficient functions.
+static void* poormans_memset(void *dst, int v, uptr size) {
+  for (uptr i = 0; i < size; i++)
+    ((char*)dst)[i] = (char)v;
+  return dst;
+}
+
+static void* poormans_memcpy(void *dst, const void *src, uptr size) {
+  for (uptr i = 0; i < size; i++)
+    ((char*)dst)[i] = ((char*)src)[i];
+  return dst;
+}
+
+void InitializeInterceptors() {
+  CHECK_GT(cur_thread()->in_rtl, 0);
+
+  // We need to setup it early, because functions like dlsym() can call it.
+  REAL(memset) = poormans_memset;
+  REAL(memcpy) = poormans_memcpy;
+
+  TSAN_INTERCEPT(malloc);
+  TSAN_INTERCEPT(calloc);
+  TSAN_INTERCEPT(realloc);
+  TSAN_INTERCEPT(free);
+  TSAN_INTERCEPT(cfree);
+  TSAN_INTERCEPT(mmap);
+  TSAN_INTERCEPT(mmap64);
+  TSAN_INTERCEPT(munmap);
+  TSAN_INTERCEPT(memalign);
+  TSAN_INTERCEPT(valloc);
+  TSAN_INTERCEPT(pvalloc);
+  TSAN_INTERCEPT(posix_memalign);
+
+  TSAN_INTERCEPT(_Znwm);
+  TSAN_INTERCEPT(_ZnwmRKSt9nothrow_t);
+  TSAN_INTERCEPT(_Znam);
+  TSAN_INTERCEPT(_ZnamRKSt9nothrow_t);
+  TSAN_INTERCEPT(_ZdlPv);
+  TSAN_INTERCEPT(_ZdlPvRKSt9nothrow_t);
+  TSAN_INTERCEPT(_ZdaPv);
+  TSAN_INTERCEPT(_ZdaPvRKSt9nothrow_t);
+
+  TSAN_INTERCEPT(strlen);
+  TSAN_INTERCEPT(memset);
+  TSAN_INTERCEPT(memcpy);
+  TSAN_INTERCEPT(strcmp);
+  TSAN_INTERCEPT(memchr);
+  TSAN_INTERCEPT(memrchr);
+  TSAN_INTERCEPT(memmove);
+  TSAN_INTERCEPT(memcmp);
+  TSAN_INTERCEPT(strchr);
+  TSAN_INTERCEPT(strchrnul);
+  TSAN_INTERCEPT(strrchr);
+  TSAN_INTERCEPT(strncmp);
+  TSAN_INTERCEPT(strcpy);  // NOLINT
+  TSAN_INTERCEPT(strncpy);
+  TSAN_INTERCEPT(strstr);
+
+  TSAN_INTERCEPT(__cxa_guard_acquire);
+  TSAN_INTERCEPT(__cxa_guard_release);
+
+  TSAN_INTERCEPT(pthread_create);
+  TSAN_INTERCEPT(pthread_join);
+  TSAN_INTERCEPT(pthread_detach);
+
+  TSAN_INTERCEPT(pthread_mutex_init);
+  TSAN_INTERCEPT(pthread_mutex_destroy);
+  TSAN_INTERCEPT(pthread_mutex_lock);
+  TSAN_INTERCEPT(pthread_mutex_trylock);
+  TSAN_INTERCEPT(pthread_mutex_timedlock);
+  TSAN_INTERCEPT(pthread_mutex_unlock);
+
+  TSAN_INTERCEPT(pthread_spin_init);
+  TSAN_INTERCEPT(pthread_spin_destroy);
+  TSAN_INTERCEPT(pthread_spin_lock);
+  TSAN_INTERCEPT(pthread_spin_trylock);
+  TSAN_INTERCEPT(pthread_spin_unlock);
+
+  TSAN_INTERCEPT(pthread_rwlock_init);
+  TSAN_INTERCEPT(pthread_rwlock_destroy);
+  TSAN_INTERCEPT(pthread_rwlock_rdlock);
+  TSAN_INTERCEPT(pthread_rwlock_tryrdlock);
+  TSAN_INTERCEPT(pthread_rwlock_timedrdlock);
+  TSAN_INTERCEPT(pthread_rwlock_wrlock);
+  TSAN_INTERCEPT(pthread_rwlock_trywrlock);
+  TSAN_INTERCEPT(pthread_rwlock_timedwrlock);
+  TSAN_INTERCEPT(pthread_rwlock_unlock);
+
+  TSAN_INTERCEPT(pthread_cond_init);
+  TSAN_INTERCEPT(pthread_cond_destroy);
+  TSAN_INTERCEPT(pthread_cond_signal);
+  TSAN_INTERCEPT(pthread_cond_broadcast);
+  TSAN_INTERCEPT(pthread_cond_wait);
+  TSAN_INTERCEPT(pthread_cond_timedwait);
+
+  TSAN_INTERCEPT(pthread_barrier_init);
+  TSAN_INTERCEPT(pthread_barrier_destroy);
+  TSAN_INTERCEPT(pthread_barrier_wait);
+
+  TSAN_INTERCEPT(pthread_once);
+
+  TSAN_INTERCEPT(sem_init);
+  TSAN_INTERCEPT(sem_destroy);
+  TSAN_INTERCEPT(sem_wait);
+  TSAN_INTERCEPT(sem_trywait);
+  TSAN_INTERCEPT(sem_timedwait);
+  TSAN_INTERCEPT(sem_post);
+  TSAN_INTERCEPT(sem_getvalue);
+
+  TSAN_INTERCEPT(read);
+  TSAN_INTERCEPT(pread);
+  TSAN_INTERCEPT(pread64);
+  TSAN_INTERCEPT(readv);
+  TSAN_INTERCEPT(preadv64);
+  TSAN_INTERCEPT(write);
+  TSAN_INTERCEPT(pwrite);
+  TSAN_INTERCEPT(pwrite64);
+  TSAN_INTERCEPT(writev);
+  TSAN_INTERCEPT(pwritev64);
+  TSAN_INTERCEPT(send);
+  TSAN_INTERCEPT(sendmsg);
+  TSAN_INTERCEPT(recv);
+  TSAN_INTERCEPT(recvmsg);
+
+  TSAN_INTERCEPT(unlink);
+  TSAN_INTERCEPT(fopen);
+  TSAN_INTERCEPT(fread);
+  TSAN_INTERCEPT(fwrite);
+  TSAN_INTERCEPT(puts);
+  TSAN_INTERCEPT(rmdir);
+  TSAN_INTERCEPT(opendir);
+
+  TSAN_INTERCEPT(epoll_ctl);
+  TSAN_INTERCEPT(epoll_wait);
+
+  TSAN_INTERCEPT(sigaction);
+
+  atexit_ctx = new(internal_alloc(MBlockAtExit, sizeof(AtExitContext)))
+      AtExitContext();
+
+  if (__cxa_atexit(&finalize, 0, 0)) {
+    Printf("ThreadSanitizer: failed to setup atexit callback\n");
+    Die();
+  }
+
+  if (pthread_key_create(&g_thread_finalize_key, &thread_finalize)) {
+    Printf("ThreadSanitizer: failed to create thread key\n");
+    Die();
+  }
+}
+
+void internal_memset(void *ptr, int c, uptr size) {
+  REAL(memset)(ptr, c, size);
+}
+
+void internal_memcpy(void *dst, const void *src, uptr size) {
+  REAL(memcpy)(dst, src, size);
+}
+
+int internal_memcmp(const void *s1, const void *s2, uptr size) {
+  return REAL(memcmp)(s1, s2, size);
+}
+
+int internal_strcmp(const char *s1, const char *s2) {
+  return REAL(strcmp)(s1, s2);
+}
+
+int internal_strncmp(const char *s1, const char *s2, uptr size) {
+  return REAL(strncmp)(s1, s2, size);
+}
+
+void internal_strcpy(char *s1, const char *s2) {
+  REAL(strcpy)(s1, s2);  // NOLINT
+}
+
+uptr internal_strlen(const char *s) {
+  return REAL(strlen)(s);
+}
+
+char* internal_strdup(const char *s) {
+  uptr len = internal_strlen(s);
+  char *s2 = (char*)internal_alloc(MBlockString, len + 1);
+  internal_memcpy(s2, s, len);
+  s2[len] = 0;
+  return s2;
+}
+
+const char *internal_strstr(const char *where, const char *what) {
+  return REAL(strstr)(where, what);
+}
+
+const char *internal_strchr(const char *where, char what) {
+  return (const char*)REAL(strchr)((void*)where, what);
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_interface.cc b/lib/tsan/rtl/tsan_interface.cc
new file mode 100644
index 000000000..b7b008564
--- /dev/null
+++ b/lib/tsan/rtl/tsan_interface.cc
@@ -0,0 +1,42 @@
+//===-- tsan_interface.cc ---------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#include "tsan_interface.h"
+#include "tsan_interface_ann.h"
+#include "tsan_rtl.h"
+
+#define CALLERPC ((uptr)__builtin_return_address(0))
+
+using namespace __tsan;  // NOLINT
+
+void __tsan_init() {
+  Initialize(cur_thread());
+}
+
+void __tsan_read16(void *addr) {
+  MemoryRead8Byte(cur_thread(), CALLERPC, (uptr)addr);
+  MemoryRead8Byte(cur_thread(), CALLERPC, (uptr)addr + 8);
+}
+
+void __tsan_write16(void *addr) {
+  MemoryWrite8Byte(cur_thread(), CALLERPC, (uptr)addr);
+  MemoryWrite8Byte(cur_thread(), CALLERPC, (uptr)addr + 8);
+}
+
+void __tsan_acquire(void *addr) {
+  Acquire(cur_thread(), CALLERPC, (uptr)addr);
+}
+
+void __tsan_release(void *addr) {
+  Release(cur_thread(), CALLERPC, (uptr)addr);
+}
diff --git a/lib/tsan/rtl/tsan_interface.h b/lib/tsan/rtl/tsan_interface.h
new file mode 100644
index 000000000..ed21ec605
--- /dev/null
+++ b/lib/tsan/rtl/tsan_interface.h
@@ -0,0 +1,51 @@
+//===-- tsan_interface.h ----------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+// The functions declared in this header will be inserted by the instrumentation
+// module.
+// This header can be included by the instrumented program or by TSan tests.
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_INTERFACE_H
+#define TSAN_INTERFACE_H
+
+// This header should NOT include any other headers.
+// All functions in this header are extern "C" and start with __tsan_.
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// This function should be called at the very beginning of the process,
+// before any instrumented code is executed and before any call to malloc.
+void __tsan_init();
+
+void __tsan_read1(void *addr);
+void __tsan_read2(void *addr);
+void __tsan_read4(void *addr);
+void __tsan_read8(void *addr);
+void __tsan_read16(void *addr);
+
+void __tsan_write1(void *addr);
+void __tsan_write2(void *addr);
+void __tsan_write4(void *addr);
+void __tsan_write8(void *addr);
+void __tsan_write16(void *addr);
+
+void __tsan_vptr_update(void **vptr_p, void *new_val);
+
+void __tsan_func_entry(void *call_pc);
+void __tsan_func_exit();
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TSAN_INTERFACE_H
diff --git a/lib/tsan/rtl/tsan_interface_ann.cc b/lib/tsan/rtl/tsan_interface_ann.cc
new file mode 100644
index 000000000..96f1f8abc
--- /dev/null
+++ b/lib/tsan/rtl/tsan_interface_ann.cc
@@ -0,0 +1,344 @@
+//===-- tsan_interface_ann.cc -----------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#include "tsan_interface_ann.h"
+#include "tsan_mutex.h"
+#include "tsan_placement_new.h"
+#include "tsan_report.h"
+#include "tsan_rtl.h"
+#include "tsan_mman.h"
+#include "tsan_flags.h"
+
+#define CALLERPC ((uptr)__builtin_return_address(0))
+
+using namespace __tsan;  // NOLINT
+
+namespace __tsan {
+
+class ScopedAnnotation {
+ public:
+  ScopedAnnotation(ThreadState *thr, const char *aname, const char *f, int l,
+                   uptr pc)
+      : thr_(thr)
+      , in_rtl_(thr->in_rtl) {
+    CHECK_EQ(thr_->in_rtl, 0);
+    FuncEntry(thr_, pc);
+    thr_->in_rtl++;
+    DPrintf("#%d: annotation %s() %s:%d\n", thr_->tid, aname, f, l);
+  }
+
+  ~ScopedAnnotation() {
+    thr_->in_rtl--;
+    CHECK_EQ(in_rtl_, thr_->in_rtl);
+    FuncExit(thr_);
+  }
+ private:
+  ThreadState *const thr_;
+  const int in_rtl_;
+};
+
+#define SCOPED_ANNOTATION(typ) \
+    if (!flags()->enable_annotations) \
+      return; \
+    ThreadState *thr = cur_thread(); \
+    StatInc(thr, StatAnnotation); \
+    StatInc(thr, Stat##typ); \
+    ScopedAnnotation sa(thr, __FUNCTION__, f, l, \
+        (uptr)__builtin_return_address(0)); \
+    const uptr pc = (uptr)&__FUNCTION__; \
+    (void)pc; \
+/**/
+
+static const int kMaxDescLen = 128;
+
+struct ExpectRace {
+  ExpectRace *next;
+  ExpectRace *prev;
+  int hitcount;
+  uptr addr;
+  uptr size;
+  char *file;
+  int line;
+  char desc[kMaxDescLen];
+};
+
+struct DynamicAnnContext {
+  Mutex mtx;
+  ExpectRace expect;
+  ExpectRace benign;
+
+  DynamicAnnContext()
+    : mtx(MutexTypeAnnotations, StatMtxAnnotations) {
+  }
+};
+
+static DynamicAnnContext *dyn_ann_ctx;
+static char dyn_ann_ctx_placeholder[sizeof(DynamicAnnContext)] ALIGN(64);
+
+static void AddExpectRace(ExpectRace *list,
+    char *f, int l, uptr addr, uptr size, char *desc) {
+  ExpectRace *race = list->next;
+  for (; race != list; race = race->next) {
+    if (race->addr == addr && race->size == size)
+      return;
+  }
+  race = (ExpectRace*)internal_alloc(MBlockExpectRace, sizeof(ExpectRace));
+  race->hitcount = 0;
+  race->addr = addr;
+  race->size = size;
+  race->file = f;
+  race->line = l;
+  race->desc[0] = 0;
+  if (desc) {
+    int i = 0;
+    for (; i < kMaxDescLen - 1 && desc[i]; i++)
+      race->desc[i] = desc[i];
+    race->desc[i] = 0;
+  }
+  race->prev = list;
+  race->next = list->next;
+  race->next->prev = race;
+  list->next = race;
+}
+
+static ExpectRace *FindRace(ExpectRace *list, uptr addr, uptr size) {
+  for (ExpectRace *race = list->next; race != list; race = race->next) {
+    uptr maxbegin = max(race->addr, addr);
+    uptr minend = min(race->addr + race->size, addr + size);
+    if (maxbegin < minend)
+      return race;
+  }
+  return 0;
+}
+
+static bool CheckContains(ExpectRace *list, uptr addr, uptr size) {
+  ExpectRace *race = FindRace(list, addr, size);
+  if (race == 0)
+    return false;
+  DPrintf("Hit expected/benign race: %s addr=%lx:%d %s:%d\n",
+      race->desc, race->addr, (int)race->size, race->file, race->line);
+  race->hitcount++;
+  return true;
+}
+
+static void InitList(ExpectRace *list) {
+  list->next = list;
+  list->prev = list;
+}
+
+void InitializeDynamicAnnotations() {
+  dyn_ann_ctx = new(dyn_ann_ctx_placeholder) DynamicAnnContext;
+  InitList(&dyn_ann_ctx->expect);
+  InitList(&dyn_ann_ctx->benign);
+}
+
+bool IsExpectedReport(uptr addr, uptr size) {
+  Lock lock(&dyn_ann_ctx->mtx);
+  if (CheckContains(&dyn_ann_ctx->expect, addr, size))
+    return true;
+  if (CheckContains(&dyn_ann_ctx->benign, addr, size))
+    return true;
+  return false;
+}
+
+}  // namespace __tsan
+
+using namespace __tsan;  // NOLINT
+
+extern "C" {
+void AnnotateHappensBefore(char *f, int l, uptr addr) {
+  SCOPED_ANNOTATION(AnnotateHappensBefore);
+  Release(cur_thread(), CALLERPC, addr);
+}
+
+void AnnotateHappensAfter(char *f, int l, uptr addr) {
+  SCOPED_ANNOTATION(AnnotateHappensAfter);
+  Acquire(cur_thread(), CALLERPC, addr);
+}
+
+void AnnotateCondVarSignal(char *f, int l, uptr cv) {
+  SCOPED_ANNOTATION(AnnotateCondVarSignal);
+}
+
+void AnnotateCondVarSignalAll(char *f, int l, uptr cv) {
+  SCOPED_ANNOTATION(AnnotateCondVarSignalAll);
+}
+
+void AnnotateMutexIsNotPHB(char *f, int l, uptr mu) {
+  SCOPED_ANNOTATION(AnnotateMutexIsNotPHB);
+}
+
+void AnnotateCondVarWait(char *f, int l, uptr cv, uptr lock) {
+  SCOPED_ANNOTATION(AnnotateCondVarWait);
+}
+
+void AnnotateRWLockCreate(char *f, int l, uptr lock) {
+  SCOPED_ANNOTATION(AnnotateRWLockCreate);
+}
+
+void AnnotateRWLockDestroy(char *f, int l, uptr lock) {
+  SCOPED_ANNOTATION(AnnotateRWLockDestroy);
+}
+
+void AnnotateRWLockAcquired(char *f, int l, uptr lock, uptr is_w) {
+  SCOPED_ANNOTATION(AnnotateRWLockAcquired);
+}
+
+void AnnotateRWLockReleased(char *f, int l, uptr lock, uptr is_w) {
+  SCOPED_ANNOTATION(AnnotateRWLockReleased);
+}
+
+void AnnotateTraceMemory(char *f, int l, uptr mem) {
+  SCOPED_ANNOTATION(AnnotateTraceMemory);
+}
+
+void AnnotateFlushState(char *f, int l) {
+  SCOPED_ANNOTATION(AnnotateFlushState);
+}
+
+void AnnotateNewMemory(char *f, int l, uptr mem, uptr size) {
+  SCOPED_ANNOTATION(AnnotateNewMemory);
+}
+
+void AnnotateNoOp(char *f, int l, uptr mem) {
+  SCOPED_ANNOTATION(AnnotateNoOp);
+}
+
+static void ReportMissedExpectedRace(ExpectRace *race) {
+  Printf("==================\n");
+  Printf("WARNING: ThreadSanitizer: missed expected data race\n");
+  Printf("  %s addr=%lx %s:%d\n",
+      race->desc, race->addr, race->file, race->line);
+  Printf("==================\n");
+}
+
+void AnnotateFlushExpectedRaces(char *f, int l) {
+  SCOPED_ANNOTATION(AnnotateFlushExpectedRaces);
+  Lock lock(&dyn_ann_ctx->mtx);
+  while (dyn_ann_ctx->expect.next != &dyn_ann_ctx->expect) {
+    ExpectRace *race = dyn_ann_ctx->expect.next;
+    if (race->hitcount == 0) {
+      CTX()->nmissed_expected++;
+      ReportMissedExpectedRace(race);
+    }
+    race->prev->next = race->next;
+    race->next->prev = race->prev;
+    internal_free(race);
+  }
+}
+
+void AnnotateEnableRaceDetection(char *f, int l, int enable) {
+  SCOPED_ANNOTATION(AnnotateEnableRaceDetection);
+  // FIXME: Reconsider this functionality later. It may be irrelevant.
+}
+
+void AnnotateMutexIsUsedAsCondVar(char *f, int l, uptr mu) {
+  SCOPED_ANNOTATION(AnnotateMutexIsUsedAsCondVar);
+}
+
+void AnnotatePCQGet(char *f, int l, uptr pcq) {
+  SCOPED_ANNOTATION(AnnotatePCQGet);
+}
+
+void AnnotatePCQPut(char *f, int l, uptr pcq) {
+  SCOPED_ANNOTATION(AnnotatePCQPut);
+}
+
+void AnnotatePCQDestroy(char *f, int l, uptr pcq) {
+  SCOPED_ANNOTATION(AnnotatePCQDestroy);
+}
+
+void AnnotatePCQCreate(char *f, int l, uptr pcq) {
+  SCOPED_ANNOTATION(AnnotatePCQCreate);
+}
+
+void AnnotateExpectRace(char *f, int l, uptr mem, char *desc) {
+  SCOPED_ANNOTATION(AnnotateExpectRace);
+  Lock lock(&dyn_ann_ctx->mtx);
+  AddExpectRace(&dyn_ann_ctx->expect,
+                f, l, mem, 1, desc);
+  DPrintf("Add expected race: %s addr=%lx %s:%d\n", desc, mem, f, l);
+}
+
+static void BenignRaceImpl(char *f, int l, uptr mem, uptr size, char *desc) {
+  Lock lock(&dyn_ann_ctx->mtx);
+  AddExpectRace(&dyn_ann_ctx->benign,
+                f, l, mem, size, desc);
+  DPrintf("Add benign race: %s addr=%lx %s:%d\n", desc, mem, f, l);
+}
+
+// FIXME: Turn it off later. WTF is benign race?1?? Go talk to Hans Boehm.
+void AnnotateBenignRaceSized(char *f, int l, uptr mem, uptr size, char *desc) {
+  SCOPED_ANNOTATION(AnnotateBenignRaceSized);
+  BenignRaceImpl(f, l, mem, size, desc);
+}
+
+void AnnotateBenignRace(char *f, int l, uptr mem, char *desc) {
+  SCOPED_ANNOTATION(AnnotateBenignRace);
+  BenignRaceImpl(f, l, mem, 1, desc);
+}
+
+void AnnotateIgnoreReadsBegin(char *f, int l) {
+  SCOPED_ANNOTATION(AnnotateIgnoreReadsBegin);
+  IgnoreCtl(cur_thread(), false, true);
+}
+
+void AnnotateIgnoreReadsEnd(char *f, int l) {
+  SCOPED_ANNOTATION(AnnotateIgnoreReadsEnd);
+  IgnoreCtl(cur_thread(), false, false);
+}
+
+void AnnotateIgnoreWritesBegin(char *f, int l) {
+  SCOPED_ANNOTATION(AnnotateIgnoreWritesBegin);
+  IgnoreCtl(cur_thread(), true, true);
+}
+
+void AnnotateIgnoreWritesEnd(char *f, int l) {
+  SCOPED_ANNOTATION(AnnotateIgnoreWritesEnd);
+  IgnoreCtl(cur_thread(), true, false);
+}
+
+void AnnotatePublishMemoryRange(char *f, int l, uptr addr, uptr size) {
+  SCOPED_ANNOTATION(AnnotatePublishMemoryRange);
+}
+
+void AnnotateUnpublishMemoryRange(char *f, int l, uptr addr, uptr size) {
+  SCOPED_ANNOTATION(AnnotateUnpublishMemoryRange);
+}
+
+void AnnotateThreadName(char *f, int l, char *name) {
+  SCOPED_ANNOTATION(AnnotateThreadName);
+}
+
+void WTFAnnotateHappensBefore(char *f, int l, uptr addr) {
+  SCOPED_ANNOTATION(AnnotateHappensBefore);
+}
+
+void WTFAnnotateHappensAfter(char *f, int l, uptr addr) {
+  SCOPED_ANNOTATION(AnnotateHappensAfter);
+}
+
+void WTFAnnotateBenignRaceSized(char *f, int l, uptr mem, uptr sz, char *desc) {
+  SCOPED_ANNOTATION(AnnotateBenignRaceSized);
+}
+
+int RunningOnValgrind() {
+  return 0;
+}
+
+const char *ThreadSanitizerQuery(const char *query) {
+  if (internal_strcmp(query, "pure_happens_before") == 0)
+    return "1";
+  else
+    return "0";
+}
+}  // extern "C"
diff --git a/lib/tsan/rtl/tsan_interface_ann.h b/lib/tsan/rtl/tsan_interface_ann.h
new file mode 100644
index 000000000..09e807a00
--- /dev/null
+++ b/lib/tsan/rtl/tsan_interface_ann.h
@@ -0,0 +1,31 @@
+//===-- tsan_interface_ann.h ------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+// Interface for dynamic annotations.
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_INTERFACE_ANN_H
+#define TSAN_INTERFACE_ANN_H
+
+// This header should NOT include any other headers.
+// All functions in this header are extern "C" and start with __tsan_.
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void __tsan_acquire(void *addr);
+void __tsan_release(void *addr);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TSAN_INTERFACE_ANN_H
diff --git a/lib/tsan/rtl/tsan_interface_atomic.cc b/lib/tsan/rtl/tsan_interface_atomic.cc
new file mode 100644
index 000000000..7e5f191a5
--- /dev/null
+++ b/lib/tsan/rtl/tsan_interface_atomic.cc
@@ -0,0 +1,194 @@
+//===-- tsan_interface_atomic.cc --------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#include "tsan_interface_atomic.h"
+#include "tsan_placement_new.h"
+#include "tsan_flags.h"
+#include "tsan_rtl.h"
+
+using namespace __tsan;  // NOLINT
+
+class ScopedAtomic {
+ public:
+  ScopedAtomic(ThreadState *thr, uptr pc, const char *func)
+      : thr_(thr) {
+    CHECK_EQ(thr_->in_rtl, 1);  // 1 due to our own ScopedInRtl member.
+    DPrintf("#%d: %s\n", thr_->tid, func);
+  }
+  ~ScopedAtomic() {
+    CHECK_EQ(thr_->in_rtl, 1);
+  }
+ private:
+  ThreadState *thr_;
+  ScopedInRtl in_rtl_;
+};
+
+// Some shortcuts.
+typedef __tsan_memory_order morder;
+typedef __tsan_atomic8 a8;
+typedef __tsan_atomic16 a16;
+typedef __tsan_atomic32 a32;
+typedef __tsan_atomic64 a64;
+const int mo_relaxed = __tsan_memory_order_relaxed;
+const int mo_consume = __tsan_memory_order_consume;
+const int mo_acquire = __tsan_memory_order_acquire;
+const int mo_release = __tsan_memory_order_release;
+const int mo_acq_rel = __tsan_memory_order_acq_rel;
+const int mo_seq_cst = __tsan_memory_order_seq_cst;
+
+static void AtomicStatInc(ThreadState *thr, uptr size, morder mo, StatType t) {
+  StatInc(thr, StatAtomic);
+  StatInc(thr, t);
+  StatInc(thr, size == 1 ? StatAtomic1
+             : size == 2 ? StatAtomic2
+             : size == 4 ? StatAtomic4
+             :             StatAtomic8);
+  StatInc(thr, mo == mo_relaxed ? StatAtomicRelaxed
+             : mo == mo_consume ? StatAtomicConsume
+             : mo == mo_acquire ? StatAtomicAcquire
+             : mo == mo_release ? StatAtomicRelease
+             : mo == mo_acq_rel ? StatAtomicAcq_Rel
+             :                    StatAtomicSeq_Cst);
+}
+
+#define SCOPED_ATOMIC(func, ...) \
+    mo = flags()->force_seq_cst_atomics ? (morder)mo_seq_cst : mo; \
+    ThreadState *const thr = cur_thread(); \
+    const uptr pc = (uptr)__builtin_return_address(0); \
+    AtomicStatInc(thr, sizeof(*a), mo, StatAtomic##func); \
+    ScopedAtomic sa(thr, pc, __FUNCTION__); \
+    return Atomic##func(thr, pc, __VA_ARGS__); \
+/**/
+
+template<typename T>
+static T AtomicLoad(ThreadState *thr, uptr pc, const volatile T *a,
+    morder mo) {
+  CHECK(mo & (mo_relaxed | mo_consume | mo_acquire | mo_seq_cst));
+  T v = *a;
+  if (mo & (mo_consume | mo_acquire | mo_seq_cst))
+    Acquire(thr, pc, (uptr)a);
+  return v;
+}
+
+template<typename T>
+static void AtomicStore(ThreadState *thr, uptr pc, volatile T *a, T v,
+    morder mo) {
+  CHECK(mo & (mo_relaxed | mo_release | mo_seq_cst));
+  if (mo & (mo_release | mo_seq_cst))
+    Release(thr, pc, (uptr)a);
+  *a = v;
+}
+
+template<typename T>
+static T AtomicExchange(ThreadState *thr, uptr pc, volatile T *a, T v,
+    morder mo) {
+  if (mo & (mo_release | mo_acq_rel | mo_seq_cst))
+    Release(thr, pc, (uptr)a);
+  v = __sync_lock_test_and_set(a, v);
+  if (mo & (mo_consume | mo_acquire | mo_acq_rel | mo_seq_cst))
+    Acquire(thr, pc, (uptr)a);
+  return v;
+}
+
+template<typename T>
+static T AtomicFetchAdd(ThreadState *thr, uptr pc, volatile T *a, T v,
+    morder mo) {
+  if (mo & (mo_release | mo_acq_rel | mo_seq_cst))
+    Release(thr, pc, (uptr)a);
+  v = __sync_fetch_and_add(a, v);
+  if (mo & (mo_consume | mo_acquire | mo_acq_rel | mo_seq_cst))
+    Acquire(thr, pc, (uptr)a);
+  return v;
+}
+
+template<typename T>
+static bool AtomicCAS(ThreadState *thr, uptr pc,
+    volatile T *a, T *c, T v, morder mo) {
+  if (mo & (mo_release | mo_acq_rel | mo_seq_cst))
+    Release(thr, pc, (uptr)a);
+  T cc = *c;
+  T pr = __sync_val_compare_and_swap(a, cc, v);
+  if (mo & (mo_consume | mo_acquire | mo_acq_rel | mo_seq_cst))
+    Acquire(thr, pc, (uptr)a);
+  if (pr == cc)
+    return true;
+  *c = pr;
+  return false;
+}
+
+static void AtomicFence(ThreadState *thr, uptr pc, morder mo) {
+  __sync_synchronize();
+}
+
+a8 __tsan_atomic8_load(const volatile a8 *a, morder mo) {
+  SCOPED_ATOMIC(Load, a, mo);
+}
+
+a16 __tsan_atomic16_load(const volatile a16 *a, morder mo) {
+  SCOPED_ATOMIC(Load, a, mo);
+}
+
+a32 __tsan_atomic32_load(const volatile a32 *a, morder mo) {
+  SCOPED_ATOMIC(Load, a, mo);
+}
+
+a64 __tsan_atomic64_load(const volatile a64 *a, morder mo) {
+  SCOPED_ATOMIC(Load, a, mo);
+}
+
+void __tsan_atomic8_store(volatile a8 *a, a8 v, morder mo) {
+  SCOPED_ATOMIC(Store, a, v, mo);
+}
+
+void __tsan_atomic16_store(volatile a16 *a, a16 v, morder mo) {
+  SCOPED_ATOMIC(Store, a, v, mo);
+}
+
+void __tsan_atomic32_store(volatile a32 *a, a32 v, morder mo) {
+  SCOPED_ATOMIC(Store, a, v, mo);
+}
+
+void __tsan_atomic64_store(volatile a64 *a, a64 v, morder mo) {
+  SCOPED_ATOMIC(Store, a, v, mo);
+}
+
+a32 __tsan_atomic32_exchange(volatile a32 *a, a32 v, morder mo) {
+  SCOPED_ATOMIC(Exchange, a, v, mo);
+}
+
+a64 __tsan_atomic64_exchange(volatile a64 *a, a64 v, morder mo) {
+  SCOPED_ATOMIC(Exchange, a, v, mo);
+}
+
+a32 __tsan_atomic32_fetch_add(volatile a32 *a, a32 v, morder mo) {
+  SCOPED_ATOMIC(FetchAdd, a, v, mo);
+}
+
+a64 __tsan_atomic64_fetch_add(volatile a64 *a, a64 v, morder mo) {
+  SCOPED_ATOMIC(FetchAdd, a, v, mo);
+}
+
+int __tsan_atomic32_compare_exchange_strong(volatile a32 *a, a32 *c, a32 v,
+    morder mo) {
+  SCOPED_ATOMIC(CAS, a, c, v, mo);
+}
+
+int __tsan_atomic64_compare_exchange_strong(volatile a64 *a, a64 *c, a64 v,
+    morder mo) {
+  SCOPED_ATOMIC(CAS, a, c, v, mo);
+}
+
+void __tsan_atomic_thread_fence(morder mo) {
+  char* a;
+  SCOPED_ATOMIC(Fence, mo);
+}
diff --git a/lib/tsan/rtl/tsan_interface_atomic.h b/lib/tsan/rtl/tsan_interface_atomic.h
new file mode 100644
index 000000000..724407998
--- /dev/null
+++ b/lib/tsan/rtl/tsan_interface_atomic.h
@@ -0,0 +1,73 @@
+//===-- tsan_interface_atomic.h ---------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_INTERFACE_ATOMIC_H
+#define TSAN_INTERFACE_ATOMIC_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef char  __tsan_atomic8;
+typedef short __tsan_atomic16;  // NOLINT
+typedef int   __tsan_atomic32;
+typedef long  __tsan_atomic64;  // NOLINT
+
+typedef enum {
+  __tsan_memory_order_relaxed = 1 << 0,
+  __tsan_memory_order_consume = 1 << 1,
+  __tsan_memory_order_acquire = 1 << 2,
+  __tsan_memory_order_release = 1 << 3,
+  __tsan_memory_order_acq_rel = 1 << 4,
+  __tsan_memory_order_seq_cst = 1 << 5,
+} __tsan_memory_order;
+
+__tsan_atomic8 __tsan_atomic8_load(const volatile __tsan_atomic8 *a,
+    __tsan_memory_order mo);
+__tsan_atomic16 __tsan_atomic16_load(const volatile __tsan_atomic16 *a,
+    __tsan_memory_order mo);
+__tsan_atomic32 __tsan_atomic32_load(const volatile __tsan_atomic32 *a,
+    __tsan_memory_order mo);
+__tsan_atomic64 __tsan_atomic64_load(const volatile __tsan_atomic64 *a,
+    __tsan_memory_order mo);
+
+void __tsan_atomic8_store(volatile __tsan_atomic8 *a, __tsan_atomic8 v,
+    __tsan_memory_order mo);
+void __tsan_atomic16_store(volatile __tsan_atomic16 *a, __tsan_atomic16 v,
+    __tsan_memory_order mo);
+void __tsan_atomic32_store(volatile __tsan_atomic32 *a, __tsan_atomic32 v,
+    __tsan_memory_order mo);
+void __tsan_atomic64_store(volatile __tsan_atomic64 *a, __tsan_atomic64 v,
+    __tsan_memory_order mo);
+
+__tsan_atomic32 __tsan_atomic32_exchange(volatile __tsan_atomic32 *a,
+    __tsan_atomic32 v, __tsan_memory_order mo);
+__tsan_atomic64 __tsan_atomic64_exchange(volatile __tsan_atomic64 *a,
+    __tsan_atomic64 v, __tsan_memory_order mo);
+
+__tsan_atomic32 __tsan_atomic32_fetch_add(volatile __tsan_atomic32 *a,
+    __tsan_atomic32 v, __tsan_memory_order mo);
+__tsan_atomic64 __tsan_atomic64_fetch_add(volatile __tsan_atomic64 *a,
+    __tsan_atomic64 v, __tsan_memory_order mo);
+
+int __tsan_atomic32_compare_exchange_strong(volatile __tsan_atomic32 *a,
+    __tsan_atomic32 *c, __tsan_atomic32 v, __tsan_memory_order mo);
+int __tsan_atomic64_compare_exchange_strong(volatile __tsan_atomic64 *a,
+    __tsan_atomic64 *c, __tsan_atomic64 v, __tsan_memory_order mo);
+
+void __tsan_atomic_thread_fence(__tsan_memory_order mo);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // #ifndef TSAN_INTERFACE_ATOMIC_H
diff --git a/lib/tsan/rtl/tsan_interface_inl.h b/lib/tsan/rtl/tsan_interface_inl.h
new file mode 100644
index 000000000..233f9028a
--- /dev/null
+++ b/lib/tsan/rtl/tsan_interface_inl.h
@@ -0,0 +1,65 @@
+//===-- tsan_interface_inl.h ------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#include "tsan_interface.h"
+#include "tsan_rtl.h"
+
+#define CALLERPC ((uptr)__builtin_return_address(0))
+
+using namespace __tsan;  // NOLINT
+
+void __tsan_read1(void *addr) {
+  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 0, 0);
+}
+
+void __tsan_read2(void *addr) {
+  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 1, 0);
+}
+
+void __tsan_read4(void *addr) {
+  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 2, 0);
+}
+
+void __tsan_read8(void *addr) {
+  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 3, 0);
+}
+
+void __tsan_write1(void *addr) {
+  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 0, 1);
+}
+
+void __tsan_write2(void *addr) {
+  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 1, 1);
+}
+
+void __tsan_write4(void *addr) {
+  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 2, 1);
+}
+
+void __tsan_write8(void *addr) {
+  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 3, 1);
+}
+
+void __tsan_vptr_update(void **vptr_p, void *new_val) {
+  CHECK_EQ(sizeof(vptr_p), 8);
+  if (*vptr_p != new_val)
+    MemoryAccess(cur_thread(), CALLERPC, (uptr)vptr_p, 3, 1);
+}
+
+void __tsan_func_entry(void *pc) {
+  FuncEntry(cur_thread(), (uptr)pc);
+}
+
+void __tsan_func_exit() {
+  FuncExit(cur_thread());
+}
diff --git a/lib/tsan/rtl/tsan_md5.cc b/lib/tsan/rtl/tsan_md5.cc
new file mode 100644
index 000000000..487336025
--- /dev/null
+++ b/lib/tsan/rtl/tsan_md5.cc
@@ -0,0 +1,245 @@
+//===-- tsan_md5.cc ---------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#include "tsan_defs.h"
+
+namespace __tsan {
+
+#define F(x, y, z)      ((z) ^ ((x) & ((y) ^ (z))))
+#define G(x, y, z)      ((y) ^ ((z) & ((x) ^ (y))))
+#define H(x, y, z)      ((x) ^ (y) ^ (z))
+#define I(x, y, z)      ((y) ^ ((x) | ~(z)))
+
+#define STEP(f, a, b, c, d, x, t, s) \
+  (a) += f((b), (c), (d)) + (x) + (t); \
+  (a) = (((a) << (s)) | (((a) & 0xffffffff) >> (32 - (s)))); \
+  (a) += (b);
+
+#define SET(n) \
+  (*(MD5_u32plus *)&ptr[(n) * 4])
+#define GET(n) \
+  SET(n)
+
+typedef unsigned int MD5_u32plus;
+typedef unsigned long ulong_t;  // NOLINT
+
+typedef struct {
+  MD5_u32plus lo, hi;
+  MD5_u32plus a, b, c, d;
+  unsigned char buffer[64];
+  MD5_u32plus block[16];
+} MD5_CTX;
+
+static void *body(MD5_CTX *ctx, void *data, ulong_t size) {
+  unsigned char *ptr;
+  MD5_u32plus a, b, c, d;
+  MD5_u32plus saved_a, saved_b, saved_c, saved_d;
+
+  ptr = (unsigned char*)data;
+
+  a = ctx->a;
+  b = ctx->b;
+  c = ctx->c;
+  d = ctx->d;
+
+  do {
+    saved_a = a;
+    saved_b = b;
+    saved_c = c;
+    saved_d = d;
+
+    STEP(F, a, b, c, d, SET(0), 0xd76aa478, 7)
+    STEP(F, d, a, b, c, SET(1), 0xe8c7b756, 12)
+    STEP(F, c, d, a, b, SET(2), 0x242070db, 17)
+    STEP(F, b, c, d, a, SET(3), 0xc1bdceee, 22)
+    STEP(F, a, b, c, d, SET(4), 0xf57c0faf, 7)
+    STEP(F, d, a, b, c, SET(5), 0x4787c62a, 12)
+    STEP(F, c, d, a, b, SET(6), 0xa8304613, 17)
+    STEP(F, b, c, d, a, SET(7), 0xfd469501, 22)
+    STEP(F, a, b, c, d, SET(8), 0x698098d8, 7)
+    STEP(F, d, a, b, c, SET(9), 0x8b44f7af, 12)
+    STEP(F, c, d, a, b, SET(10), 0xffff5bb1, 17)
+    STEP(F, b, c, d, a, SET(11), 0x895cd7be, 22)
+    STEP(F, a, b, c, d, SET(12), 0x6b901122, 7)
+    STEP(F, d, a, b, c, SET(13), 0xfd987193, 12)
+    STEP(F, c, d, a, b, SET(14), 0xa679438e, 17)
+    STEP(F, b, c, d, a, SET(15), 0x49b40821, 22)
+
+    STEP(G, a, b, c, d, GET(1), 0xf61e2562, 5)
+    STEP(G, d, a, b, c, GET(6), 0xc040b340, 9)
+    STEP(G, c, d, a, b, GET(11), 0x265e5a51, 14)
+    STEP(G, b, c, d, a, GET(0), 0xe9b6c7aa, 20)
+    STEP(G, a, b, c, d, GET(5), 0xd62f105d, 5)
+    STEP(G, d, a, b, c, GET(10), 0x02441453, 9)
+    STEP(G, c, d, a, b, GET(15), 0xd8a1e681, 14)
+    STEP(G, b, c, d, a, GET(4), 0xe7d3fbc8, 20)
+    STEP(G, a, b, c, d, GET(9), 0x21e1cde6, 5)
+    STEP(G, d, a, b, c, GET(14), 0xc33707d6, 9)
+    STEP(G, c, d, a, b, GET(3), 0xf4d50d87, 14)
+    STEP(G, b, c, d, a, GET(8), 0x455a14ed, 20)
+    STEP(G, a, b, c, d, GET(13), 0xa9e3e905, 5)
+    STEP(G, d, a, b, c, GET(2), 0xfcefa3f8, 9)
+    STEP(G, c, d, a, b, GET(7), 0x676f02d9, 14)
+    STEP(G, b, c, d, a, GET(12), 0x8d2a4c8a, 20)
+
+    STEP(H, a, b, c, d, GET(5), 0xfffa3942, 4)
+    STEP(H, d, a, b, c, GET(8), 0x8771f681, 11)
+    STEP(H, c, d, a, b, GET(11), 0x6d9d6122, 16)
+    STEP(H, b, c, d, a, GET(14), 0xfde5380c, 23)
+    STEP(H, a, b, c, d, GET(1), 0xa4beea44, 4)
+    STEP(H, d, a, b, c, GET(4), 0x4bdecfa9, 11)
+    STEP(H, c, d, a, b, GET(7), 0xf6bb4b60, 16)
+    STEP(H, b, c, d, a, GET(10), 0xbebfbc70, 23)
+    STEP(H, a, b, c, d, GET(13), 0x289b7ec6, 4)
+    STEP(H, d, a, b, c, GET(0), 0xeaa127fa, 11)
+    STEP(H, c, d, a, b, GET(3), 0xd4ef3085, 16)
+    STEP(H, b, c, d, a, GET(6), 0x04881d05, 23)
+    STEP(H, a, b, c, d, GET(9), 0xd9d4d039, 4)
+    STEP(H, d, a, b, c, GET(12), 0xe6db99e5, 11)
+    STEP(H, c, d, a, b, GET(15), 0x1fa27cf8, 16)
+    STEP(H, b, c, d, a, GET(2), 0xc4ac5665, 23)
+
+    STEP(I, a, b, c, d, GET(0), 0xf4292244, 6)
+    STEP(I, d, a, b, c, GET(7), 0x432aff97, 10)
+    STEP(I, c, d, a, b, GET(14), 0xab9423a7, 15)
+    STEP(I, b, c, d, a, GET(5), 0xfc93a039, 21)
+    STEP(I, a, b, c, d, GET(12), 0x655b59c3, 6)
+    STEP(I, d, a, b, c, GET(3), 0x8f0ccc92, 10)
+    STEP(I, c, d, a, b, GET(10), 0xffeff47d, 15)
+    STEP(I, b, c, d, a, GET(1), 0x85845dd1, 21)
+    STEP(I, a, b, c, d, GET(8), 0x6fa87e4f, 6)
+    STEP(I, d, a, b, c, GET(15), 0xfe2ce6e0, 10)
+    STEP(I, c, d, a, b, GET(6), 0xa3014314, 15)
+    STEP(I, b, c, d, a, GET(13), 0x4e0811a1, 21)
+    STEP(I, a, b, c, d, GET(4), 0xf7537e82, 6)
+    STEP(I, d, a, b, c, GET(11), 0xbd3af235, 10)
+    STEP(I, c, d, a, b, GET(2), 0x2ad7d2bb, 15)
+    STEP(I, b, c, d, a, GET(9), 0xeb86d391, 21)
+
+    a += saved_a;
+    b += saved_b;
+    c += saved_c;
+    d += saved_d;
+
+    ptr += 64;
+  } while (size -= 64);
+
+  ctx->a = a;
+  ctx->b = b;
+  ctx->c = c;
+  ctx->d = d;
+
+  return ptr;
+}
+
+void MD5_Init(MD5_CTX *ctx) {
+  ctx->a = 0x67452301;
+  ctx->b = 0xefcdab89;
+  ctx->c = 0x98badcfe;
+  ctx->d = 0x10325476;
+
+  ctx->lo = 0;
+  ctx->hi = 0;
+}
+
+void MD5_Update(MD5_CTX *ctx, void *data, ulong_t size) {
+  MD5_u32plus saved_lo;
+  ulong_t used, free;
+
+  saved_lo = ctx->lo;
+  if ((ctx->lo = (saved_lo + size) & 0x1fffffff) < saved_lo)
+    ctx->hi++;
+  ctx->hi += size >> 29;
+
+  used = saved_lo & 0x3f;
+
+  if (used) {
+    free = 64 - used;
+
+    if (size < free) {
+      internal_memcpy(&ctx->buffer[used], data, size);
+      return;
+    }
+
+    internal_memcpy(&ctx->buffer[used], data, free);
+    data = (unsigned char *)data + free;
+    size -= free;
+    body(ctx, ctx->buffer, 64);
+  }
+
+  if (size >= 64) {
+    data = body(ctx, data, size & ~(ulong_t)0x3f);
+    size &= 0x3f;
+  }
+
+  internal_memcpy(ctx->buffer, data, size);
+}
+
+void MD5_Final(unsigned char *result, MD5_CTX *ctx) {
+  ulong_t used, free;
+
+  used = ctx->lo & 0x3f;
+
+  ctx->buffer[used++] = 0x80;
+
+  free = 64 - used;
+
+  if (free < 8) {
+    internal_memset(&ctx->buffer[used], 0, free);
+    body(ctx, ctx->buffer, 64);
+    used = 0;
+    free = 64;
+  }
+
+  internal_memset(&ctx->buffer[used], 0, free - 8);
+
+  ctx->lo <<= 3;
+  ctx->buffer[56] = ctx->lo;
+  ctx->buffer[57] = ctx->lo >> 8;
+  ctx->buffer[58] = ctx->lo >> 16;
+  ctx->buffer[59] = ctx->lo >> 24;
+  ctx->buffer[60] = ctx->hi;
+  ctx->buffer[61] = ctx->hi >> 8;
+  ctx->buffer[62] = ctx->hi >> 16;
+  ctx->buffer[63] = ctx->hi >> 24;
+
+  body(ctx, ctx->buffer, 64);
+
+  result[0] = ctx->a;
+  result[1] = ctx->a >> 8;
+  result[2] = ctx->a >> 16;
+  result[3] = ctx->a >> 24;
+  result[4] = ctx->b;
+  result[5] = ctx->b >> 8;
+  result[6] = ctx->b >> 16;
+  result[7] = ctx->b >> 24;
+  result[8] = ctx->c;
+  result[9] = ctx->c >> 8;
+  result[10] = ctx->c >> 16;
+  result[11] = ctx->c >> 24;
+  result[12] = ctx->d;
+  result[13] = ctx->d >> 8;
+  result[14] = ctx->d >> 16;
+  result[15] = ctx->d >> 24;
+
+  internal_memset(ctx, 0, sizeof(*ctx));
+}
+
+MD5Hash md5_hash(const void *data, uptr size) {
+  MD5Hash res;
+  MD5_CTX ctx;
+  MD5_Init(&ctx);
+  MD5_Update(&ctx, (void*)data, size);
+  MD5_Final((unsigned char*)&res.hash[0], &ctx);
+  return res;
+}
+}
diff --git a/lib/tsan/rtl/tsan_mman.cc b/lib/tsan/rtl/tsan_mman.cc
new file mode 100644
index 000000000..e24074eeb
--- /dev/null
+++ b/lib/tsan/rtl/tsan_mman.cc
@@ -0,0 +1,137 @@
+//===-- tsan_mman.cc --------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#include "tsan_mman.h"
+#include "tsan_allocator.h"
+#include "tsan_rtl.h"
+#include "tsan_report.h"
+#include "tsan_flags.h"
+
+namespace __tsan {
+
+static void SignalUnsafeCall(ThreadState *thr, uptr pc) {
+  if (!thr->in_signal_handler || !flags()->report_signal_unsafe)
+    return;
+  StackTrace stack;
+  stack.ObtainCurrent(thr, pc);
+  ScopedReport rep(ReportTypeSignalUnsafe);
+  rep.AddStack(&stack);
+  OutputReport(rep);
+}
+
+void *user_alloc(ThreadState *thr, uptr pc, uptr sz) {
+  CHECK_GT(thr->in_rtl, 0);
+  MBlock *b = (MBlock*)Alloc(sz + sizeof(MBlock));
+  b->size = sz;
+  void *p = b + 1;
+  if (CTX() && CTX()->initialized) {
+    MemoryResetRange(thr, pc, (uptr)p, sz);
+  }
+  DPrintf("#%d: alloc(%lu) = %p\n", thr->tid, sz, p);
+  SignalUnsafeCall(thr, pc);
+  return p;
+}
+
+void user_free(ThreadState *thr, uptr pc, void *p) {
+  CHECK_GT(thr->in_rtl, 0);
+  CHECK_NE(p, (void*)0);
+  DPrintf("#%d: free(%p)\n", thr->tid, p);
+  MBlock *b = user_mblock(thr, p);
+  p = b + 1;
+  if (CTX() && CTX()->initialized && thr->in_rtl == 1) {
+    MemoryRangeFreed(thr, pc, (uptr)p, b->size);
+  }
+  Free(b);
+  SignalUnsafeCall(thr, pc);
+}
+
+void *user_realloc(ThreadState *thr, uptr pc, void *p, uptr sz) {
+  CHECK_GT(thr->in_rtl, 0);
+  void *p2 = 0;
+  // FIXME: Handle "shrinking" more efficiently,
+  // it seems that some software actually does this.
+  if (sz) {
+    p2 = user_alloc(thr, pc, sz);
+    if (p) {
+      MBlock *b = user_mblock(thr, p);
+      internal_memcpy(p2, p, min(b->size, sz));
+    }
+  }
+  if (p) {
+    user_free(thr, pc, p);
+  }
+  return p2;
+}
+
+void *user_alloc_aligned(ThreadState *thr, uptr pc, uptr sz, uptr align) {
+  CHECK_GT(thr->in_rtl, 0);
+  void *p = user_alloc(thr, pc, sz + align);
+  void *pa = RoundUp(p, align);
+  DCHECK_LE((uptr)pa + sz, (uptr)p + sz + align);
+  return pa;
+}
+
+MBlock *user_mblock(ThreadState *thr, void *p) {
+  CHECK_GT(thr->in_rtl, 0);
+  CHECK_NE(p, (void*)0);
+  MBlock *b = (MBlock*)AllocBlock(p);
+  // FIXME: Output a warning, it's a user error.
+  if (p < (char*)(b + 1) || p > (char*)(b + 1) + b->size) {
+    Printf("user_mblock p=%p b=%p size=%lu beg=%p end=%p\n",
+        p, b, b->size, (char*)(b + 1), (char*)(b + 1) + b->size);
+    CHECK_GE(p, (char*)(b + 1));
+    CHECK_LE(p, (char*)(b + 1) + b->size);
+  }
+  return b;
+}
+
+#if TSAN_DEBUG
+struct InternalMBlock {
+  static u32 const kMagic = 0xBCEBC041;
+  u32 magic;
+  u32 typ;
+  u64 sz;
+};
+#endif
+
+void *internal_alloc(MBlockType typ, uptr sz) {
+  ThreadState *thr = cur_thread();
+  CHECK_GT(thr->in_rtl, 0);
+#if TSAN_DEBUG
+  InternalMBlock *b = (InternalMBlock*)Alloc(sizeof(InternalMBlock) + sz);
+  b->magic = InternalMBlock::kMagic;
+  b->typ = typ;
+  b->sz = sz;
+  thr->int_alloc_cnt[typ] += 1;
+  thr->int_alloc_siz[typ] += sz;
+  void *p = b + 1;
+  return p;
+#else
+  return Alloc(sz);
+#endif
+}
+
+void internal_free(void *p) {
+  ThreadState *thr = cur_thread();
+  CHECK_GT(thr->in_rtl, 0);
+#if TSAN_DEBUG
+  InternalMBlock *b = (InternalMBlock*)p - 1;
+  CHECK_EQ(b->magic, InternalMBlock::kMagic);
+  thr->int_alloc_cnt[b->typ] -= 1;
+  thr->int_alloc_siz[b->typ] -= b->sz;
+  Free(b);
+#else
+  Free(p);
+#endif
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_mman.h b/lib/tsan/rtl/tsan_mman.h
new file mode 100644
index 000000000..8b51de6f3
--- /dev/null
+++ b/lib/tsan/rtl/tsan_mman.h
@@ -0,0 +1,111 @@
+//===-- tsan_mman.h ---------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_MMAN_H
+#define TSAN_MMAN_H
+
+#include "tsan_defs.h"
+
+namespace __tsan {
+
+// Descriptor of user's memory block.
+struct MBlock {
+  uptr size;
+};
+
+// For user allocations.
+void *user_alloc(ThreadState *thr, uptr pc, uptr sz);
+// Does not accept NULL.
+void user_free(ThreadState *thr, uptr pc, void *p);
+void *user_realloc(ThreadState *thr, uptr pc, void *p, uptr sz);
+void *user_alloc_aligned(ThreadState *thr, uptr pc, uptr sz, uptr align);
+// Given the pointer p into a valid allocated block,
+// returns the descriptor of the block.
+MBlock *user_mblock(ThreadState *thr, void *p);
+
+enum MBlockType {
+  MBlockScopedBuf,
+  MBlockString,
+  MBlockStackTrace,
+  MBlockSync,
+  MBlockClock,
+  MBlockThreadContex,
+  MBlockRacyStacks,
+  MBlockRacyAddresses,
+  MBlockAtExit,
+  MBlockFlag,
+  MBlockReport,
+  MBlockReportMop,
+  MBlockReportThread,
+  MBlockReportMutex,
+  MBlockReportLoc,
+  MBlockReportStack,
+  MBlockSuppression,
+  MBlockExpectRace,
+
+  // This must be the last.
+  MBlockTypeCount,
+};
+
+// For internal data structures.
+void *internal_alloc(MBlockType typ, uptr sz);
+void internal_free(void *p);
+
+template<typename T>
+void DestroyAndFree(T *&p) {
+  p->~T();
+  internal_free(p);
+  p = 0;
+}
+
+template<typename T>
+class InternalScopedBuf {
+ public:
+  explicit InternalScopedBuf(uptr cnt) {
+    cnt_ = cnt;
+    ptr_ = (T*)internal_alloc(MBlockScopedBuf, cnt * sizeof(T));
+  }
+
+  ~InternalScopedBuf() {
+    internal_free(ptr_);
+  }
+
+  operator T *() {
+    return ptr_;
+  }
+
+  T &operator[](uptr i) {
+    return ptr_[i];
+  }
+
+  T *Ptr() {
+    return ptr_;
+  }
+
+  uptr Count() {
+    return cnt_;
+  }
+
+  uptr Size() {
+    return cnt_ * sizeof(T);
+  }
+
+ private:
+  T *ptr_;
+  uptr cnt_;
+
+  InternalScopedBuf(const InternalScopedBuf&);
+  void operator = (const InternalScopedBuf&);
+};
+
+}  // namespace __tsan
+#endif  // TSAN_MMAN_H
diff --git a/lib/tsan/rtl/tsan_mutex.cc b/lib/tsan/rtl/tsan_mutex.cc
new file mode 100644
index 000000000..a343a8bc2
--- /dev/null
+++ b/lib/tsan/rtl/tsan_mutex.cc
@@ -0,0 +1,275 @@
+//===-- tsan_mutex.cc -------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#include "tsan_mutex.h"
+#include "tsan_platform.h"
+#include "tsan_rtl.h"
+
+namespace __tsan {
+
+// Simple reader-writer spin-mutex. Optimized for not-so-contended case.
+// Readers have preference, can possibly starvate writers.
+
+// The table fixes what mutexes can be locked under what mutexes.
+// E.g. if the row for MutexTypeThreads contains MutexTypeReport,
+// then Report mutex can be locked while under Threads mutex.
+// The leaf mutexes can be locked under any other mutexes.
+// Recursive locking is not supported.
+const MutexType MutexTypeLeaf = (MutexType)-1;
+static MutexType CanLockTab[MutexTypeCount][MutexTypeCount] = {
+  /*0 MutexTypeInvalid*/     {},
+  /*1 MutexTypeTrace*/       {MutexTypeLeaf},
+  /*2 MutexTypeThreads*/     {MutexTypeReport},
+  /*3 MutexTypeReport*/      {},
+  /*4 MutexTypeSyncVar*/     {},
+  /*5 MutexTypeSyncTab*/     {MutexTypeSyncVar},
+  /*6 MutexTypeSlab*/        {MutexTypeLeaf},
+  /*7 MutexTypeAnnotations*/ {},
+  /*8 MutexTypeAtExit*/      {MutexTypeSyncTab},
+};
+
+static bool CanLockAdj[MutexTypeCount][MutexTypeCount];
+
+void InitializeMutex() {
+  // Build the "can lock" adjacency matrix.
+  // If [i][j]==true, then one can lock mutex j while under mutex i.
+  const int N = MutexTypeCount;
+  int cnt[N] = {};
+  bool leaf[N] = {};
+  for (int i = 1; i < N; i++) {
+    for (int j = 0; j < N; j++) {
+      int z = CanLockTab[i][j];
+      if (z == MutexTypeInvalid)
+        continue;
+      if (z == MutexTypeLeaf) {
+        CHECK(!leaf[i]);
+        leaf[i] = true;
+        continue;
+      }
+      CHECK(!CanLockAdj[i][z]);
+      CanLockAdj[i][z] = true;
+      cnt[i]++;
+    }
+  }
+  for (int i = 0; i < N; i++) {
+    CHECK(!leaf[i] || cnt[i] == 0);
+  }
+  // Add leaf mutexes.
+  for (int i = 0; i < N; i++) {
+    if (!leaf[i])
+      continue;
+    for (int j = 0; j < N; j++) {
+      if (i == j || leaf[j] || j == MutexTypeInvalid)
+        continue;
+      CHECK(!CanLockAdj[j][i]);
+      CanLockAdj[j][i] = true;
+    }
+  }
+  // Build the transitive closure.
+  bool CanLockAdj2[MutexTypeCount][MutexTypeCount];
+  for (int i = 0; i < N; i++) {
+    for (int j = 0; j < N; j++) {
+      CanLockAdj2[i][j] = CanLockAdj[i][j];
+    }
+  }
+  for (int k = 0; k < N; k++) {
+    for (int i = 0; i < N; i++) {
+      for (int j = 0; j < N; j++) {
+        if (CanLockAdj2[i][k] && CanLockAdj2[k][j]) {
+          CanLockAdj2[i][j] = true;
+        }
+      }
+    }
+  }
+#if 0
+  Printf("Can lock graph:\n");
+  for (int i = 0; i < N; i++) {
+    for (int j = 0; j < N; j++) {
+      Printf("%d ", CanLockAdj[i][j]);
+    }
+    Printf("\n");
+  }
+  Printf("Can lock graph closure:\n");
+  for (int i = 0; i < N; i++) {
+    for (int j = 0; j < N; j++) {
+      Printf("%d ", CanLockAdj2[i][j]);
+    }
+    Printf("\n");
+  }
+#endif
+  // Verify that the graph is acyclic.
+  for (int i = 0; i < N; i++) {
+    if (CanLockAdj2[i][i]) {
+      Printf("Mutex %d participates in a cycle\n", i);
+      Die();
+    }
+  }
+}
+
+DeadlockDetector::DeadlockDetector() {
+  // Rely on zero initialization because some mutexes can be locked before ctor.
+}
+
+void DeadlockDetector::Lock(MutexType t) {
+  // Printf("LOCK %d @%llu\n", t, seq_ + 1);
+  u64 max_seq = 0;
+  u64 max_idx = MutexTypeInvalid;
+  for (int i = 0; i != MutexTypeCount; i++) {
+    if (locked_[i] == 0)
+      continue;
+    CHECK_NE(locked_[i], max_seq);
+    if (max_seq < locked_[i]) {
+      max_seq = locked_[i];
+      max_idx = i;
+    }
+  }
+  locked_[t] = ++seq_;
+  if (max_idx == MutexTypeInvalid)
+    return;
+  // Printf("  last %d @%llu\n", max_idx, max_seq);
+  if (!CanLockAdj[max_idx][t]) {
+    Printf("ThreadSanitizer: internal deadlock detected\n");
+    Printf("ThreadSanitizer: can't lock %d while under %llu\n", t, max_idx);
+    Die();
+  }
+}
+
+void DeadlockDetector::Unlock(MutexType t) {
+  // Printf("UNLO %d @%llu #%llu\n", t, seq_, locked_[t]);
+  CHECK(locked_[t]);
+  locked_[t] = 0;
+}
+
+const uptr kUnlocked = 0;
+const uptr kWriteLock = 1;
+const uptr kReadLock = 2;
+
+class Backoff {
+ public:
+  Backoff()
+    : iter_() {
+  }
+
+  bool Do() {
+    if (iter_++ < kActiveSpinIters)
+      proc_yield(kActiveSpinCnt);
+    else
+      sched_yield();
+    return true;
+  }
+
+  u64 Contention() const {
+    u64 active = iter_ % kActiveSpinIters;
+    u64 passive = iter_ - active;
+    return active + 10 * passive;
+  }
+
+ private:
+  int iter_;
+  static const int kActiveSpinIters = 10;
+  static const int kActiveSpinCnt = 20;
+};
+
+Mutex::Mutex(MutexType type, StatType stat_type) {
+  CHECK_GT(type, MutexTypeInvalid);
+  CHECK_LT(type, MutexTypeCount);
+#if TSAN_DEBUG
+  type_ = type;
+#endif
+#if TSAN_COLLECT_STATS
+  stat_type_ = stat_type;
+#endif
+  atomic_store(&state_, kUnlocked, memory_order_relaxed);
+}
+
+Mutex::~Mutex() {
+  CHECK_EQ(atomic_load(&state_, memory_order_relaxed), kUnlocked);
+}
+
+void Mutex::Lock() {
+#if TSAN_DEBUG
+  cur_thread()->deadlock_detector.Lock(type_);
+#endif
+  uptr cmp = kUnlocked;
+  if (atomic_compare_exchange_strong(&state_, &cmp, kWriteLock,
+                                     memory_order_acquire))
+    return;
+  for (Backoff backoff; backoff.Do();) {
+    if (atomic_load(&state_, memory_order_relaxed) == kUnlocked) {
+      cmp = kUnlocked;
+      if (atomic_compare_exchange_weak(&state_, &cmp, kWriteLock,
+                                       memory_order_acquire)) {
+#if TSAN_COLLECT_STATS
+        StatInc(cur_thread(), stat_type_, backoff.Contention());
+#endif
+        return;
+      }
+    }
+  }
+}
+
+void Mutex::Unlock() {
+  uptr prev = atomic_fetch_sub(&state_, kWriteLock, memory_order_release);
+  (void)prev;
+  DCHECK_NE(prev & kWriteLock, 0);
+#if TSAN_DEBUG
+  cur_thread()->deadlock_detector.Unlock(type_);
+#endif
+}
+
+void Mutex::ReadLock() {
+#if TSAN_DEBUG
+  cur_thread()->deadlock_detector.Lock(type_);
+#endif
+  uptr prev = atomic_fetch_add(&state_, kReadLock, memory_order_acquire);
+  if ((prev & kWriteLock) == 0)
+    return;
+  for (Backoff backoff; backoff.Do();) {
+    prev = atomic_load(&state_, memory_order_acquire);
+    if ((prev & kWriteLock) == 0) {
+#if TSAN_COLLECT_STATS
+      StatInc(cur_thread(), stat_type_, backoff.Contention());
+#endif
+      return;
+    }
+  }
+}
+
+void Mutex::ReadUnlock() {
+  uptr prev = atomic_fetch_sub(&state_, kReadLock, memory_order_release);
+  (void)prev;
+  DCHECK_EQ(prev & kWriteLock, 0);
+  DCHECK_GT(prev & ~kWriteLock, 0);
+#if TSAN_DEBUG
+  cur_thread()->deadlock_detector.Unlock(type_);
+#endif
+}
+
+Lock::Lock(Mutex *m)
+  : m_(m) {
+  m_->Lock();
+}
+
+Lock::~Lock() {
+  m_->Unlock();
+}
+
+ReadLock::ReadLock(Mutex *m)
+  : m_(m) {
+  m_->ReadLock();
+}
+
+ReadLock::~ReadLock() {
+  m_->ReadUnlock();
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_mutex.h b/lib/tsan/rtl/tsan_mutex.h
new file mode 100644
index 000000000..2180978f9
--- /dev/null
+++ b/lib/tsan/rtl/tsan_mutex.h
@@ -0,0 +1,98 @@
+//===-- tsan_mutex.h --------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_MUTEX_H
+#define TSAN_MUTEX_H
+
+#include "tsan_atomic.h"
+#include "tsan_defs.h"
+
+namespace __tsan {
+
+enum MutexType {
+  MutexTypeInvalid,
+  MutexTypeTrace,
+  MutexTypeThreads,
+  MutexTypeReport,
+  MutexTypeSyncVar,
+  MutexTypeSyncTab,
+  MutexTypeSlab,
+  MutexTypeAnnotations,
+  MutexTypeAtExit,
+
+  // This must be the last.
+  MutexTypeCount,
+};
+
+class Mutex {
+ public:
+  explicit Mutex(MutexType type, StatType stat_type);
+  ~Mutex();
+
+  void Lock();
+  void Unlock();
+
+  void ReadLock();
+  void ReadUnlock();
+
+ private:
+  atomic_uintptr_t state_;
+#if TSAN_DEBUG
+  MutexType type_;
+#endif
+#if TSAN_COLLECT_STATS
+  StatType stat_type_;
+#endif
+
+  Mutex(const Mutex&);
+  void operator = (const Mutex&);
+};
+
+class Lock {
+ public:
+  explicit Lock(Mutex *m);
+  ~Lock();
+
+ private:
+  Mutex *m_;
+
+  Lock(const Lock&);
+  void operator = (const Lock&);
+};
+
+class ReadLock {
+ public:
+  explicit ReadLock(Mutex *m);
+  ~ReadLock();
+
+ private:
+  Mutex *m_;
+
+  ReadLock(const ReadLock&);
+  void operator = (const ReadLock&);
+};
+
+class DeadlockDetector {
+ public:
+  DeadlockDetector();
+  void Lock(MutexType t);
+  void Unlock(MutexType t);
+ private:
+  u64 seq_;
+  u64 locked_[MutexTypeCount];
+};
+
+void InitializeMutex();
+
+}  // namespace __tsan
+
+#endif  // TSAN_MUTEX_H
diff --git a/lib/tsan/rtl/tsan_placement_new.h b/lib/tsan/rtl/tsan_placement_new.h
new file mode 100644
index 000000000..7b8ba036e
--- /dev/null
+++ b/lib/tsan/rtl/tsan_placement_new.h
@@ -0,0 +1,24 @@
+//===-- tsan_placement_new.h ------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+// The file provides 'placement new'
+// Do not include it into header files, only into source files.
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_PLACEMENT_NEW_H
+#define TSAN_PLACEMENT_NEW_H
+
+#include "tsan_defs.h"
+
+inline void *operator new(__tsan::uptr sz, void *p) {
+  return p;
+}
+
+#endif  // TSAN_PLACEMENT_NEW_H
diff --git a/lib/tsan/rtl/tsan_platform.h b/lib/tsan/rtl/tsan_platform.h
new file mode 100644
index 000000000..9cfe4763f
--- /dev/null
+++ b/lib/tsan/rtl/tsan_platform.h
@@ -0,0 +1,93 @@
+//===-- tsan_platform.h -----------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+// Platform-specific code.
+//===----------------------------------------------------------------------===//
+
+#ifndef TSAN_LINUX_H
+#define TSAN_LINUX_H
+#ifdef __linux__
+
+#include "tsan_rtl.h"
+
+#if __LP64__
+namespace __tsan {
+
+// TSAN_COMPAT_SHADOW is intended for COMPAT virtual memory layout,
+// when memory addresses are of the 0x2axxxxxxxxxx form.
+// The option is enabled with 'setarch x86_64 -L'.
+#if defined(TSAN_COMPAT_SHADOW) && TSAN_COMPAT_SHADOW
+
+static const uptr kLinuxAppMemBeg = 0x2a0000000000ULL;
+static const uptr kLinuxAppMemEnd = 0x7fffffffffffULL;
+
+#else
+
+static const uptr kLinuxAppMemBeg = 0x7ef000000000ULL;
+static const uptr kLinuxAppMemEnd = 0x7fffffffffffULL;
+
+#endif
+
+static const uptr kLinuxAppMemMsk = 0x7c0000000000ULL;
+
+// This has to be a macro to allow constant initialization of constants below.
+#define MemToShadow(addr) \
+    (((addr) & ~(kLinuxAppMemMsk | (kShadowCell - 1))) * kShadowCnt)
+
+static const uptr kLinuxShadowBeg = MemToShadow(kLinuxAppMemBeg);
+static const uptr kLinuxShadowEnd =
+  MemToShadow(kLinuxAppMemEnd) | (kPageSize - 1);
+
+static inline bool IsAppMem(uptr mem) {
+  return mem >= kLinuxAppMemBeg && mem <= kLinuxAppMemEnd;
+}
+
+static inline bool IsShadowMem(uptr mem) {
+  return mem >= kLinuxShadowBeg && mem <= kLinuxShadowEnd;
+}
+
+static inline uptr ShadowToMem(uptr shadow) {
+  CHECK(IsShadowMem(shadow));
+#if defined(TSAN_COMPAT_SHADOW) && TSAN_COMPAT_SHADOW
+  // COMPAT mapping is not quite one-to-one.
+  return (shadow / kShadowCnt) | 0x280000000000ULL;
+#else
+  return (shadow / kShadowCnt) | kLinuxAppMemMsk;
+#endif
+}
+
+const char *InitializePlatform();
+void FinalizePlatform();
+int GetPid();
+
+void sched_yield();
+
+typedef int fd_t;
+const fd_t kInvalidFd = -1;
+fd_t internal_open(const char *name, bool write);
+void internal_close(fd_t fd);
+uptr internal_filesize(fd_t fd);  // -1 on error.
+uptr internal_read(fd_t fd, void *p, uptr size);
+uptr internal_write(fd_t fd, const void *p, uptr size);
+const char *internal_getpwd();
+
+uptr GetTlsSize();
+void GetThreadStackAndTls(uptr *stk_addr, uptr *stk_size,
+                          uptr *tls_addr, uptr *tls_size);
+
+}  // namespace __tsan
+
+#else  // __LP64__
+# error "Only 64-bit is supported"
+#endif
+
+#endif  // __linux__
+#endif  // TSAN_LINUX_H
diff --git a/lib/tsan/rtl/tsan_platform_linux.cc b/lib/tsan/rtl/tsan_platform_linux.cc
new file mode 100644
index 000000000..171c3486a
--- /dev/null
+++ b/lib/tsan/rtl/tsan_platform_linux.cc
@@ -0,0 +1,250 @@
+//===-- tsan_platform_linux.cc ----------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+// Linux-specific code.
+//===----------------------------------------------------------------------===//
+
+#include "tsan_platform.h"
+#include "tsan_rtl.h"
+#include "tsan_flags.h"
+
+#include <asm/prctl.h>
+#include <fcntl.h>
+#include <pthread.h>
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <stdarg.h>
+#include <sys/mman.h>
+#include <sys/prctl.h>
+#include <sys/syscall.h>
+#include <sys/time.h>
+#include <sys/types.h>
+#include <sys/resource.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#include <errno.h>
+#include <sched.h>
+#include <dlfcn.h>
+
+extern "C" int arch_prctl(int code, __tsan::uptr *addr);
+
+namespace __tsan {
+
+static uptr g_tls_size;
+
+ScopedInRtl::ScopedInRtl()
+    : thr_(cur_thread()) {
+  in_rtl_ = thr_->in_rtl;
+  thr_->in_rtl++;
+  errno_ = errno;
+}
+
+ScopedInRtl::~ScopedInRtl() {
+  thr_->in_rtl--;
+  errno = errno_;
+  CHECK_EQ(in_rtl_, thr_->in_rtl);
+}
+
+void Die() {
+  _exit(1);
+}
+
+static void *my_mmap(void *addr, size_t length, int prot, int flags,
+                    int fd, u64 offset) {
+  ScopedInRtl in_rtl;
+# if __WORDSIZE == 64
+  return (void *)syscall(__NR_mmap, addr, length, prot, flags, fd, offset);
+# else
+  return (void *)syscall(__NR_mmap2, addr, length, prot, flags, fd, offset);
+# endif
+}
+
+void sched_yield() {
+  ScopedInRtl in_rtl;
+  syscall(__NR_sched_yield);
+}
+
+fd_t internal_open(const char *name, bool write) {
+  ScopedInRtl in_rtl;
+  return syscall(__NR_open, name,
+      write ? O_WRONLY | O_CREAT | O_CLOEXEC : O_RDONLY, 0660);
+}
+
+void internal_close(fd_t fd) {
+  ScopedInRtl in_rtl;
+  syscall(__NR_close, fd);
+}
+
+uptr internal_filesize(fd_t fd) {
+  struct stat st = {};
+  if (syscall(__NR_fstat, fd, &st))
+    return -1;
+  return (uptr)st.st_size;
+}
+
+uptr internal_read(fd_t fd, void *p, uptr size) {
+  ScopedInRtl in_rtl;
+  return syscall(__NR_read, fd, p, size);
+}
+
+uptr internal_write(fd_t fd, const void *p, uptr size) {
+  ScopedInRtl in_rtl;
+  return syscall(__NR_write, fd, p, size);
+}
+
+const char *internal_getpwd() {
+  return getenv("PWD");
+}
+
+static void ProtectRange(uptr beg, uptr end) {
+  ScopedInRtl in_rtl;
+  CHECK_LE(beg, end);
+  if (beg == end)
+    return;
+  if (beg != (uptr)my_mmap((void*)(beg), end - beg,
+      PROT_NONE,
+      MAP_PRIVATE | MAP_ANON | MAP_FIXED | MAP_NORESERVE,
+      -1, 0)) {
+    Printf("FATAL: ThreadSanitizer can not protect [%lx,%lx]\n", beg, end);
+    Printf("FATAL: Make sure you are not using unlimited stack\n");
+    Die();
+  }
+}
+
+void InitializeShadowMemory() {
+  const uptr kClosedLowBeg  = 0x200000;
+  const uptr kClosedLowEnd  = kLinuxShadowBeg - 1;
+  const uptr kClosedMidBeg = kLinuxShadowEnd + 1;
+  const uptr kClosedMidEnd = kLinuxAppMemBeg - 1;
+  uptr shadow = (uptr)my_mmap((void*)kLinuxShadowBeg,
+      kLinuxShadowEnd - kLinuxShadowBeg,
+      PROT_READ | PROT_WRITE,
+      MAP_PRIVATE | MAP_ANON | MAP_FIXED | MAP_NORESERVE,
+      0, 0);
+  if (shadow != kLinuxShadowBeg) {
+    Printf("FATAL: ThreadSanitizer can not mmap the shadow memory\n");
+    Printf("FATAL: Make sure to compile with -fPIE and to link with -pie.\n");
+    Die();
+  }
+  ProtectRange(kClosedLowBeg, kClosedLowEnd);
+  ProtectRange(kClosedMidBeg, kClosedMidEnd);
+  DPrintf("kClosedLow   %lx-%lx (%luGB)\n",
+      kClosedLowBeg, kClosedLowEnd, (kClosedLowEnd - kClosedLowBeg) >> 30);
+  DPrintf("kLinuxShadow %lx-%lx (%luGB)\n",
+      kLinuxShadowBeg, kLinuxShadowEnd,
+      (kLinuxShadowEnd - kLinuxShadowBeg) >> 30);
+  DPrintf("kClosedMid   %lx-%lx (%luGB)\n",
+      kClosedMidBeg, kClosedMidEnd, (kClosedMidEnd - kClosedMidBeg) >> 30);
+  DPrintf("kLinuxAppMem %lx-%lx (%luGB)\n",
+      kLinuxAppMemBeg, kLinuxAppMemEnd,
+      (kLinuxAppMemEnd - kLinuxAppMemBeg) >> 30);
+  DPrintf("stack        %lx\n", (uptr)&shadow);
+}
+
+static void CheckPIE() {
+  // Ensure that the binary is indeed compiled with -pie.
+  fd_t fmaps = internal_open("/proc/self/maps", false);
+  if (fmaps == kInvalidFd)
+    return;
+  char buf[20];
+  if (internal_read(fmaps, buf, sizeof(buf)) == sizeof(buf)) {
+    buf[sizeof(buf) - 1] = 0;
+    u64 addr = strtoll(buf, 0, 16);
+    if ((u64)addr < kLinuxAppMemBeg) {
+      Printf("FATAL: ThreadSanitizer can not mmap the shadow memory ("
+             "something is mapped at 0x%llx < 0x%lx)\n",
+             addr, kLinuxAppMemBeg);
+      Printf("FATAL: Make sure to compile with -fPIE"
+             " and to link with -pie.\n");
+      Die();
+    }
+  }
+  internal_close(fmaps);
+}
+
+#ifdef __i386__
+# define INTERNAL_FUNCTION __attribute__((regparm(3), stdcall))
+#else
+# define INTERNAL_FUNCTION
+#endif
+extern "C" void _dl_get_tls_static_info(size_t*, size_t*)
+    __attribute__((weak)) INTERNAL_FUNCTION;
+
+static int InitTlsSize() {
+  typedef void (*get_tls_func)(size_t*, size_t*) INTERNAL_FUNCTION;
+  get_tls_func get_tls = &_dl_get_tls_static_info;
+  if (get_tls == 0)
+    get_tls = (get_tls_func)dlsym(RTLD_NEXT, "_dl_get_tls_static_info");
+  CHECK_NE(get_tls, 0);
+  size_t tls_size = 0;
+  size_t tls_align = 0;
+  get_tls(&tls_size, &tls_align);
+  return tls_size;
+}
+
+const char *InitializePlatform() {
+  void *p = 0;
+  if (sizeof(p) == 8) {
+    // Disable core dumps, dumping of 16TB usually takes a bit long.
+    // The following magic is to prevent clang from replacing it with memset.
+    volatile rlimit lim;
+    lim.rlim_cur = 0;
+    lim.rlim_max = 0;
+    setrlimit(RLIMIT_CORE, (rlimit*)&lim);
+  }
+
+  CheckPIE();
+  g_tls_size = (uptr)InitTlsSize();
+  return getenv("TSAN_OPTIONS");
+}
+
+void FinalizePlatform() {
+  fflush(0);
+}
+
+uptr GetTlsSize() {
+  return g_tls_size;
+}
+
+void GetThreadStackAndTls(uptr *stk_addr, uptr *stk_size,
+                          uptr *tls_addr, uptr *tls_size) {
+  *stk_addr = 0;
+  *stk_size = 0;
+  pthread_attr_t attr;
+  if (pthread_getattr_np(pthread_self(), &attr) == 0) {
+    pthread_attr_getstack(&attr, (void**)stk_addr, (size_t*)stk_size);
+    pthread_attr_destroy(&attr);
+  }
+  arch_prctl(ARCH_GET_FS, tls_addr);
+  *tls_addr -= g_tls_size;
+  *tls_size = g_tls_size;
+
+  // If stack and tls intersect, make them non-intersecting.
+  if (*tls_addr > *stk_addr && *tls_addr < *stk_addr + *stk_size) {
+    CHECK_GT(*tls_addr + *tls_size, *stk_addr);
+    CHECK_LE(*tls_addr + *tls_size, *stk_addr + *stk_size);
+    *stk_size = *tls_addr - *stk_addr;
+    *stk_size = RoundUp(*stk_size, kPageSize);
+    uptr stk_end = *stk_addr + *stk_size;
+    if (stk_end > *tls_addr) {
+      *tls_size -= *tls_addr - stk_end;
+      *tls_addr = stk_end;
+    }
+  }
+}
+
+int GetPid() {
+  return getpid();
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_printf.cc b/lib/tsan/rtl/tsan_printf.cc
new file mode 100644
index 000000000..96c4d24eb
--- /dev/null
+++ b/lib/tsan/rtl/tsan_printf.cc
@@ -0,0 +1,147 @@
+//===-- tsan_printf.cc ------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#include "tsan_defs.h"
+#include "tsan_mman.h"
+#include "tsan_platform.h"
+
+#include <stdarg.h>  // va_list
+
+typedef long long i64;  // NOLINT
+typedef long iptr;  // NOLINT
+
+namespace __tsan {
+
+static int AppendChar(char **buff, const char *buff_end, char c) {
+  if (*buff < buff_end) {
+    **buff = c;
+    (*buff)++;
+  }
+  return 1;
+}
+
+static int AppendUnsigned(char **buff, const char *buff_end, u64 num,
+                          int base, uptr minimal_num_length) {
+  uptr const kMaxLen = 30;
+  uptr num_buffer[kMaxLen];
+  uptr pos = 0;
+  do {
+    num_buffer[pos++] = num % base;
+    num /= base;
+  } while (num > 0);
+  while (pos < minimal_num_length) num_buffer[pos++] = 0;
+  int result = 0;
+  while (pos-- > 0) {
+    uptr digit = num_buffer[pos];
+    result += AppendChar(buff, buff_end, (digit < 10) ? '0' + digit
+                                                      : 'a' + digit - 10);
+  }
+  return result;
+}
+
+static int AppendSignedDecimal(char **buff, const char *buff_end, i64 num) {
+  int result = 0;
+  if (num < 0) {
+    result += AppendChar(buff, buff_end, '-');
+    num = -num;
+  }
+  result += AppendUnsigned(buff, buff_end, (u64)num, 10, 0);
+  return result;
+}
+
+static int AppendString(char **buff, const char *buff_end, const char *s) {
+  if (s == 0)
+    s = "<null>";
+  int result = 0;
+  for (; *s; s++) {
+    result += AppendChar(buff, buff_end, *s);
+  }
+  return result;
+}
+
+static int AppendPointer(char **buff, const char *buff_end, u64 ptr_value) {
+  int result = 0;
+  result += AppendString(buff, buff_end, "0x");
+  result += AppendUnsigned(buff, buff_end, ptr_value, 16,
+      (sizeof(void*) == 8) ? 12 : 8);  // NOLINT
+  return result;
+}
+
+static uptr VSNPrintf(char *buff, int buff_length,
+                     const char *format, va_list args) {
+  const char *buff_end = &buff[buff_length - 1];
+  const char *cur = format;
+  int result = 0;
+  for (; *cur; cur++) {
+    if (*cur != '%') {
+      result += AppendChar(&buff, buff_end, *cur);
+      continue;
+    }
+    cur++;
+    bool is_long = (*cur == 'l');
+    cur += is_long;
+    bool is_llong = (*cur == 'l');
+    cur += is_llong;
+    switch (*cur) {
+      case 'd': {
+        i64 v = is_llong ? va_arg(args, i64)
+            : is_long ? va_arg(args, iptr)
+            : va_arg(args, int);
+        result += AppendSignedDecimal(&buff, buff_end, v);
+        break;
+      }
+      case 'u':
+      case 'x': {
+        u64 v = is_llong ? va_arg(args, u64)
+            : is_long ? va_arg(args, uptr)
+            : va_arg(args, unsigned);
+        result += AppendUnsigned(&buff, buff_end, v, *cur == 'u' ? 10: 16, 0);
+        break;
+      }
+      case 'p': {
+        result += AppendPointer(&buff, buff_end, va_arg(args, uptr));
+        break;
+      }
+      case 's': {
+        result += AppendString(&buff, buff_end, va_arg(args, char*));
+        break;
+      }
+      default: {
+        Die();
+      }
+    }
+  }
+  AppendChar(&buff, buff_end + 1, '\0');
+  return result;
+}
+
+void Printf(const char *format, ...) {
+  ScopedInRtl in_rtl;
+  const uptr kMaxLen = 16 * 1024;
+  InternalScopedBuf<char> buffer(kMaxLen);
+  va_list args;
+  va_start(args, format);
+  uptr len = VSNPrintf(buffer, buffer.Size(), format, args);
+  va_end(args);
+  internal_write(CTX() ? flags()->log_fileno : 2,
+      buffer, len < buffer.Size() ? len : buffer.Size() - 1);
+}
+
+uptr Snprintf(char *buffer, uptr length, const char *format, ...) {
+  va_list args;
+  va_start(args, format);
+  uptr len = VSNPrintf(buffer, length, format, args);
+  va_end(args);
+  return len;
+}
+}
diff --git a/lib/tsan/rtl/tsan_report.cc b/lib/tsan/rtl/tsan_report.cc
new file mode 100644
index 000000000..f231ed07d
--- /dev/null
+++ b/lib/tsan/rtl/tsan_report.cc
@@ -0,0 +1,124 @@
+//===-- tsan_report.cc ------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#include "tsan_report.h"
+#include "tsan_platform.h"
+#include "tsan_rtl.h"
+
+namespace __tsan {
+
+ReportDesc::ReportDesc()
+    : stacks(MBlockReportStack)
+    , mops(MBlockReportMop)
+    , locs(MBlockReportLoc)
+    , mutexes(MBlockReportMutex)
+    , threads(MBlockReportThread) {
+}
+
+ReportDesc::~ReportDesc() {
+}
+
+static void PrintHeader(ReportType typ) {
+  Printf("WARNING: ThreadSanitizer: ");
+
+  if (typ == ReportTypeRace)
+    Printf("data race");
+  else if (typ == ReportTypeThreadLeak)
+    Printf("thread leak");
+  else if (typ == ReportTypeMutexDestroyLocked)
+    Printf("destroy of a locked mutex");
+  else if (typ == ReportTypeSignalUnsafe)
+    Printf("signal-unsafe call inside of a signal");
+
+  Printf(" (pid=%d)\n", GetPid());
+}
+
+static void PrintStack(const ReportStack *ent) {
+  for (int i = 0; ent; ent = ent->next, i++) {
+    Printf("    #%d %s %s:%d", i, ent->func, ent->file, ent->line);
+    if (ent->col)
+      Printf(":%d", ent->col);
+    if (ent->module && ent->offset)
+      Printf(" (%s+%p)\n", ent->module, (void*)ent->offset);
+    else
+      Printf(" (%p)\n", (void*)ent->pc);
+  }
+}
+
+static void PrintMop(const ReportMop *mop, bool first) {
+  Printf("  %s of size %d at %p",
+      (first ? (mop->write ? "Write" : "Read")
+             : (mop->write ? "Previous write" : "Previous read")),
+      mop->size, (void*)mop->addr);
+  if (mop->tid == 0)
+    Printf(" by main thread:\n");
+  else
+    Printf(" by thread %d:\n", mop->tid);
+  PrintStack(mop->stack);
+}
+
+static void PrintLocation(const ReportLocation *loc) {
+  if (loc->type == ReportLocationGlobal) {
+    Printf("  Location is global '%s' of size %lu at %lx %s:%d\n",
+           loc->name, loc->size, loc->addr, loc->file, loc->line);
+  } else if (loc->type == ReportLocationHeap) {
+    Printf("  Location is heap of size %lu at %lx allocated by thread %d:\n",
+           loc->size, loc->addr, loc->tid);
+    PrintStack(loc->stack);
+  } else if (loc->type == ReportLocationStack) {
+    Printf("  Location is stack of thread %d:\n", loc->tid);
+  }
+}
+
+static void PrintMutex(const ReportMutex *rm) {
+  if (rm->stack == 0)
+    return;
+  Printf("  Mutex %d created at:\n", rm->id);
+  PrintStack(rm->stack);
+}
+
+static void PrintThread(const ReportThread *rt) {
+  if (rt->id == 0)  // Little sense in describing the main thread.
+    return;
+  Printf("  Thread %d", rt->id);
+  if (rt->name)
+    Printf(" '%s'", rt->name);
+  Printf(" (%s)", rt->running ? "running" : "finished");
+  if (rt->stack)
+    Printf(" created at:");
+  Printf("\n");
+  PrintStack(rt->stack);
+}
+
+void PrintReport(const ReportDesc *rep) {
+  Printf("==================\n");
+  PrintHeader(rep->typ);
+
+  for (uptr i = 0; i < rep->stacks.Size(); i++)
+    PrintStack(rep->stacks[i]);
+
+  for (uptr i = 0; i < rep->mops.Size(); i++)
+    PrintMop(rep->mops[i], i == 0);
+
+  for (uptr i = 0; i < rep->locs.Size(); i++)
+    PrintLocation(rep->locs[i]);
+
+  for (uptr i = 0; i < rep->mutexes.Size(); i++)
+    PrintMutex(rep->mutexes[i]);
+
+  for (uptr i = 0; i < rep->threads.Size(); i++)
+    PrintThread(rep->threads[i]);
+
+  Printf("==================\n");
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_report.h b/lib/tsan/rtl/tsan_report.h
new file mode 100644
index 000000000..70c59324d
--- /dev/null
+++ b/lib/tsan/rtl/tsan_report.h
@@ -0,0 +1,100 @@
+//===-- tsan_report.h -------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_REPORT_H
+#define TSAN_REPORT_H
+
+#include "tsan_defs.h"
+#include "tsan_vector.h"
+
+namespace __tsan {
+
+enum ReportType {
+  ReportTypeRace,
+  ReportTypeThreadLeak,
+  ReportTypeMutexDestroyLocked,
+  ReportTypeSignalUnsafe,
+};
+
+struct ReportStack {
+  ReportStack *next;
+  char *module;
+  uptr offset;
+  uptr pc;
+  char *func;
+  char *file;
+  int line;
+  int col;
+};
+
+struct ReportMop {
+  int tid;
+  uptr addr;
+  int size;
+  bool write;
+  int nmutex;
+  int *mutex;
+  ReportStack *stack;
+};
+
+enum ReportLocationType {
+  ReportLocationGlobal,
+  ReportLocationHeap,
+  ReportLocationStack,
+};
+
+struct ReportLocation {
+  ReportLocationType type;
+  uptr addr;
+  uptr size;
+  int tid;
+  char *name;
+  char *file;
+  int line;
+  ReportStack *stack;
+};
+
+struct ReportThread {
+  int id;
+  bool running;
+  char *name;
+  ReportStack *stack;
+};
+
+struct ReportMutex {
+  int id;
+  ReportStack *stack;
+};
+
+class ReportDesc {
+ public:
+  ReportType typ;
+  Vector<ReportStack*> stacks;
+  Vector<ReportMop*> mops;
+  Vector<ReportLocation*> locs;
+  Vector<ReportMutex*> mutexes;
+  Vector<ReportThread*> threads;
+
+  ReportDesc();
+  ~ReportDesc();
+
+ private:
+  ReportDesc(const ReportDesc&);
+  void operator = (const ReportDesc&);
+};
+
+// Format and output the report to the console/log. No additional logic.
+void PrintReport(const ReportDesc *rep);
+
+}  // namespace __tsan
+
+#endif  // TSAN_REPORT_H
diff --git a/lib/tsan/rtl/tsan_rtl.cc b/lib/tsan/rtl/tsan_rtl.cc
new file mode 100644
index 000000000..0e04c4494
--- /dev/null
+++ b/lib/tsan/rtl/tsan_rtl.cc
@@ -0,0 +1,460 @@
+//===-- tsan_rtl.cc ---------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+// Main file (entry points) for the TSan run-time.
+//===----------------------------------------------------------------------===//
+
+#include "tsan_defs.h"
+#include "tsan_platform.h"
+#include "tsan_rtl.h"
+#include "tsan_interface.h"
+#include "tsan_atomic.h"
+#include "tsan_mman.h"
+#include "tsan_placement_new.h"
+#include "tsan_suppressions.h"
+
+volatile int __tsan_stop = 0;
+
+extern "C" void __tsan_resume() {
+  __tsan_stop = 0;
+}
+
+namespace __tsan {
+
+THREADLOCAL char cur_thread_placeholder[sizeof(ThreadState)] ALIGN(64);
+static char ctx_placeholder[sizeof(Context)] ALIGN(64);
+
+static Context *ctx;
+Context *CTX() {
+  return ctx;
+}
+
+Context::Context()
+  : initialized()
+  , report_mtx(MutexTypeReport, StatMtxReport)
+  , nreported()
+  , nmissed_expected()
+  , thread_mtx(MutexTypeThreads, StatMtxThreads)
+  , racy_stacks(MBlockRacyStacks)
+  , racy_addresses(MBlockRacyAddresses) {
+}
+
+// The objects are allocated in TLS, so one may rely on zero-initialization.
+ThreadState::ThreadState(Context *ctx, int tid, u64 epoch,
+                         uptr stk_addr, uptr stk_size,
+                         uptr tls_addr, uptr tls_size)
+  : fast_state(tid, epoch)
+  // Do not touch these, rely on zero initialization,
+  // they may be accessed before the ctor.
+  // , fast_ignore_reads()
+  // , fast_ignore_writes()
+  // , in_rtl()
+  , shadow_stack_pos(&shadow_stack[0])
+  , tid(tid)
+  , func_call_count()
+  , stk_addr(stk_addr)
+  , stk_size(stk_size)
+  , tls_addr(tls_addr)
+  , tls_size(tls_size) {
+}
+
+ThreadContext::ThreadContext(int tid)
+  : tid(tid)
+  , unique_id()
+  , user_id()
+  , thr()
+  , status(ThreadStatusInvalid)
+  , detached()
+  , reuse_count()
+  , epoch0()
+  , epoch1()
+  , dead_next() {
+}
+
+void Initialize(ThreadState *thr) {
+  // Thread safe because done before all threads exist.
+  static bool is_initialized = false;
+  if (is_initialized)
+    return;
+  is_initialized = true;
+  ScopedInRtl in_rtl;
+  InitializeInterceptors();
+  const char *env = InitializePlatform();
+  InitializeMutex();
+  InitializeDynamicAnnotations();
+  ctx = new(ctx_placeholder) Context;
+  InitializeShadowMemory();
+  ctx->dead_list_size = 0;
+  ctx->dead_list_head = 0;
+  ctx->dead_list_tail = 0;
+  InitializeFlags(&ctx->flags, env);
+  InitializeSuppressions();
+
+  if (ctx->flags.verbosity)
+    Printf("***** Running under ThreadSanitizer v2 (pid=%d) *****\n", GetPid());
+
+  // Initialize thread 0.
+  ctx->thread_seq = 0;
+  int tid = ThreadCreate(thr, 0, 0, true);
+  CHECK_EQ(tid, 0);
+  ThreadStart(thr, tid);
+  CHECK_EQ(thr->in_rtl, 1);
+  ctx->initialized = true;
+
+  if (__tsan_stop) {
+    Printf("ThreadSanitizer is suspended at startup.\n");
+    while (__tsan_stop);
+  }
+}
+
+int Finalize(ThreadState *thr) {
+  ScopedInRtl in_rtl;
+  Context *ctx = __tsan::ctx;
+  bool failed = false;
+
+  // Be very careful beyond that point.
+  // All bets are off. Everything is destroyed.
+  ThreadFinish(thr);
+  ThreadFinalize(thr);
+  FinalizeFlags(&ctx->flags);
+
+  if (ctx->nreported) {
+    failed = true;
+    Printf("ThreadSanitizer: reported %d warnings\n", ctx->nreported);
+  }
+
+  if (ctx->nmissed_expected) {
+    failed = true;
+    Printf("ThreadSanitizer: missed %d expected races\n",
+        ctx->nmissed_expected);
+  }
+
+  StatOutput(ctx->stat);
+  FinalizeSuppressions();
+  FinalizePlatform();
+
+  const int exitcode = failed ? flags()->exitcode : 0;
+  const int log_fileno = flags()->log_fileno;
+  __tsan::ctx->~Context();
+  __tsan::ctx = 0;
+
+  InternalAllocStatAggregate(ctx, thr);
+
+  for (int i = 0; i < (int)MBlockTypeCount; i++) {
+    if (ctx->int_alloc_cnt[i] == 0 && ctx->int_alloc_siz[i] == 0)
+      continue;
+    InternalScopedBuf<char> tmp(1024);
+    Snprintf(tmp, tmp.Size(), "ThreadSanitizer: Internal memory leak: "
+        "type=%d count=%lld size=%lld\n",
+        (int)i, ctx->int_alloc_cnt[i], ctx->int_alloc_siz[i]);
+    internal_write(log_fileno, tmp, internal_strlen(tmp));
+  }
+
+  return exitcode;
+}
+
+static void TraceSwitch(ThreadState *thr) {
+  ScopedInRtl in_rtl;
+  Lock l(&thr->trace.mtx);
+  unsigned trace = (thr->fast_state.epoch() / kTracePartSize) % kTraceParts;
+  TraceHeader *hdr = &thr->trace.headers[trace];
+  hdr->epoch0 = thr->fast_state.epoch();
+  hdr->stack0.ObtainCurrent(thr, 0);
+}
+
+extern "C" void __tsan_trace_switch() {
+  TraceSwitch(cur_thread());
+}
+
+extern "C" void __tsan_report_race() {
+  ReportRace(cur_thread());
+}
+
+ALWAYS_INLINE
+static Shadow LoadShadow(u64 *p) {
+  u64 raw = atomic_load((atomic_uint64_t*)p, memory_order_relaxed);
+  return Shadow(raw);
+}
+
+ALWAYS_INLINE
+static void StoreShadow(u64 *sp, u64 s) {
+  atomic_store((atomic_uint64_t*)sp, s, memory_order_relaxed);
+}
+
+ALWAYS_INLINE
+static void StoreIfNotYetStored(u64 *sp, u64 *s) {
+  StoreShadow(sp, *s);
+  *s = 0;
+}
+
+static inline void HandleRace(ThreadState *thr, u64 *shadow_mem,
+                              Shadow cur, Shadow old) {
+  thr->racy_state[0] = cur.raw();
+  thr->racy_state[1] = old.raw();
+  thr->racy_shadow_addr = shadow_mem;
+  HACKY_CALL(__tsan_report_race);
+}
+
+static inline bool BothReads(Shadow s, int kAccessIsWrite) {
+  return !kAccessIsWrite && !s.is_write();
+}
+
+static inline bool OldIsRWStronger(Shadow old, int kAccessIsWrite) {
+  return old.is_write() || !kAccessIsWrite;
+}
+
+static inline bool OldIsRWWeaker(Shadow old, int kAccessIsWrite) {
+  return !old.is_write() || kAccessIsWrite;
+}
+
+static inline bool OldIsInSameSynchEpoch(Shadow old, ThreadState *thr) {
+  return old.epoch() >= thr->fast_synch_epoch;
+}
+
+static inline bool HappensBefore(Shadow old, ThreadState *thr) {
+  return thr->clock.get(old.tid()) >= old.epoch();
+}
+
+ALWAYS_INLINE
+void MemoryAccessImpl(ThreadState *thr, uptr addr,
+    int kAccessSizeLog, bool kAccessIsWrite, FastState fast_state,
+    u64 *shadow_mem, Shadow cur) {
+  StatInc(thr, StatMop);
+  StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
+  StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
+
+  // This potentially can live in an MMX/SSE scratch register.
+  // The required intrinsics are:
+  // __m128i _mm_move_epi64(__m128i*);
+  // _mm_storel_epi64(u64*, __m128i);
+  u64 store_word = cur.raw();
+
+  // scan all the shadow values and dispatch to 4 categories:
+  // same, replace, candidate and race (see comments below).
+  // we consider only 3 cases regarding access sizes:
+  // equal, intersect and not intersect. initially I considered
+  // larger and smaller as well, it allowed to replace some
+  // 'candidates' with 'same' or 'replace', but I think
+  // it's just not worth it (performance- and complexity-wise).
+
+  Shadow old(0);
+  if (kShadowCnt == 1) {
+    int idx = 0;
+#include "tsan_update_shadow_word_inl.h"
+  } else if (kShadowCnt == 2) {
+    int idx = 0;
+#include "tsan_update_shadow_word_inl.h"
+    idx = 1;
+#include "tsan_update_shadow_word_inl.h"
+  } else if (kShadowCnt == 4) {
+    int idx = 0;
+#include "tsan_update_shadow_word_inl.h"
+    idx = 1;
+#include "tsan_update_shadow_word_inl.h"
+    idx = 2;
+#include "tsan_update_shadow_word_inl.h"
+    idx = 3;
+#include "tsan_update_shadow_word_inl.h"
+  } else if (kShadowCnt == 8) {
+    int idx = 0;
+#include "tsan_update_shadow_word_inl.h"
+    idx = 1;
+#include "tsan_update_shadow_word_inl.h"
+    idx = 2;
+#include "tsan_update_shadow_word_inl.h"
+    idx = 3;
+#include "tsan_update_shadow_word_inl.h"
+    idx = 4;
+#include "tsan_update_shadow_word_inl.h"
+    idx = 5;
+#include "tsan_update_shadow_word_inl.h"
+    idx = 6;
+#include "tsan_update_shadow_word_inl.h"
+    idx = 7;
+#include "tsan_update_shadow_word_inl.h"
+  } else {
+    CHECK(false);
+  }
+
+  // we did not find any races and had already stored
+  // the current access info, so we are done
+  if (LIKELY(store_word == 0))
+    return;
+  // choose a random candidate slot and replace it
+  StoreShadow(shadow_mem + (cur.epoch() % kShadowCnt), store_word);
+  StatInc(thr, StatShadowReplace);
+  return;
+ RACE:
+  HandleRace(thr, shadow_mem, cur, old);
+  return;
+}
+
+ALWAYS_INLINE
+void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
+    int kAccessSizeLog, bool kAccessIsWrite) {
+  u64 *shadow_mem = (u64*)MemToShadow(addr);
+  DPrintf2("#%d: tsan::OnMemoryAccess: @%p %p size=%d"
+      " is_write=%d shadow_mem=%p {%llx, %llx, %llx, %llx}\n",
+      (int)thr->fast_state.tid(), (void*)pc, (void*)addr,
+      (int)(1 << kAccessSizeLog), kAccessIsWrite, shadow_mem,
+      shadow_mem[0], shadow_mem[1], shadow_mem[2], shadow_mem[3]);
+#if TSAN_DEBUG
+  if (!IsAppMem(addr)) {
+    Printf("Access to non app mem %lx\n", addr);
+    DCHECK(IsAppMem(addr));
+  }
+  if (!IsShadowMem((uptr)shadow_mem)) {
+    Printf("Bad shadow addr %p (%lx)\n", shadow_mem, addr);
+    DCHECK(IsShadowMem((uptr)shadow_mem));
+  }
+#endif
+
+  FastState fast_state = thr->fast_state;
+  if (fast_state.GetIgnoreBit())
+    return;
+  fast_state.IncrementEpoch();
+  thr->fast_state = fast_state;
+  Shadow cur(fast_state);
+  cur.SetAddr0AndSizeLog(addr & 7, kAccessSizeLog);
+  cur.SetWrite(kAccessIsWrite);
+
+  // We must not store to the trace if we do not store to the shadow.
+  // That is, this call must be moved somewhere below.
+  TraceAddEvent(thr, fast_state.epoch(), EventTypeMop, pc);
+
+  MemoryAccessImpl(thr, addr, kAccessSizeLog, kAccessIsWrite, fast_state,
+      shadow_mem, cur);
+}
+
+static void MemoryRangeSet(ThreadState *thr, uptr pc, uptr addr, uptr size,
+                           u64 val) {
+  if (size == 0)
+    return;
+  // FIXME: fix me.
+  uptr offset = addr % kShadowCell;
+  if (offset) {
+    offset = kShadowCell - offset;
+    if (size <= offset)
+      return;
+    addr += offset;
+    size -= offset;
+  }
+  CHECK_EQ(addr % 8, 0);
+  CHECK(IsAppMem(addr));
+  CHECK(IsAppMem(addr + size - 1));
+  (void)thr;
+  (void)pc;
+  // Some programs mmap like hundreds of GBs but actually used a small part.
+  // So, it's better to report a false positive on the memory
+  // then to hang here senselessly.
+  const uptr kMaxResetSize = 1024*1024*1024;
+  if (size > kMaxResetSize)
+    size = kMaxResetSize;
+  size = (size + 7) & ~7;
+  u64 *p = (u64*)MemToShadow(addr);
+  CHECK(IsShadowMem((uptr)p));
+  CHECK(IsShadowMem((uptr)(p + size * kShadowCnt / kShadowCell - 1)));
+  // FIXME: may overwrite a part outside the region
+  for (uptr i = 0; i < size * kShadowCnt / kShadowCell; i++)
+    p[i] = val;
+}
+
+void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size) {
+  MemoryRangeSet(thr, pc, addr, size, 0);
+}
+
+void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size) {
+  MemoryAccessRange(thr, pc, addr, size, true);
+  MemoryRangeSet(thr, pc, addr, size, kShadowFreed);
+}
+
+void FuncEntry(ThreadState *thr, uptr pc) {
+  DCHECK_EQ(thr->in_rtl, 0);
+  StatInc(thr, StatFuncEnter);
+  DPrintf2("#%d: tsan::FuncEntry %p\n", (int)thr->fast_state.tid(), (void*)pc);
+  thr->fast_state.IncrementEpoch();
+  TraceAddEvent(thr, thr->fast_state.epoch(), EventTypeFuncEnter, pc);
+
+  // Shadow stack maintenance can be replaced with
+  // stack unwinding during trace switch (which presumably must be faster).
+  DCHECK(thr->shadow_stack_pos >= &thr->shadow_stack[0]);
+  DCHECK(thr->shadow_stack_pos < &thr->shadow_stack[kShadowStackSize]);
+  thr->shadow_stack_pos[0] = pc;
+  thr->shadow_stack_pos++;
+
+#if 1
+  // While we are testing on single-threaded benchmarks,
+  // emulate some synchronization activity.
+  // FIXME: remove me later.
+  if (((++thr->func_call_count) % 1000) == 0) {
+    thr->clock.set(thr->fast_state.tid(), thr->fast_state.epoch());
+    thr->fast_synch_epoch = thr->fast_state.epoch();
+  }
+#endif
+}
+
+void FuncExit(ThreadState *thr) {
+  DCHECK_EQ(thr->in_rtl, 0);
+  StatInc(thr, StatFuncExit);
+  DPrintf2("#%d: tsan::FuncExit\n", (int)thr->fast_state.tid());
+  thr->fast_state.IncrementEpoch();
+  TraceAddEvent(thr, thr->fast_state.epoch(), EventTypeFuncExit, 0);
+
+  DCHECK(thr->shadow_stack_pos > &thr->shadow_stack[0]);
+  DCHECK(thr->shadow_stack_pos < &thr->shadow_stack[kShadowStackSize]);
+  thr->shadow_stack_pos--;
+}
+
+void IgnoreCtl(ThreadState *thr, bool write, bool begin) {
+  DPrintf("#%d: IgnoreCtl(%d, %d)\n", thr->tid, write, begin);
+  thr->ignore_reads_and_writes += begin ? 1 : -1;
+  CHECK_GE(thr->ignore_reads_and_writes, 0);
+  if (thr->ignore_reads_and_writes)
+    thr->fast_state.SetIgnoreBit();
+  else
+    thr->fast_state.ClearIgnoreBit();
+}
+
+void InternalAllocStatAggregate(Context *ctx, ThreadState *thr) {
+  for (int i = 0; i < (int)MBlockTypeCount; i++) {
+    ctx->int_alloc_cnt[i] += thr->int_alloc_cnt[i];
+    ctx->int_alloc_siz[i] += thr->int_alloc_siz[i];
+    thr->int_alloc_cnt[i] = 0;
+    thr->int_alloc_siz[i] = 0;
+  }
+}
+
+#if TSAN_DEBUG
+void build_consistency_debug() {}
+#else
+void build_consistency_release() {}
+#endif
+
+#if TSAN_COLLECT_STATS
+void build_consistency_stats() {}
+#else
+void build_consistency_nostats() {}
+#endif
+
+#if TSAN_SHADOW_COUNT == 1
+void build_consistency_shadow1() {}
+#elif TSAN_SHADOW_COUNT == 2
+void build_consistency_shadow2() {}
+#elif TSAN_SHADOW_COUNT == 4
+void build_consistency_shadow4() {}
+#else
+void build_consistency_shadow8() {}
+#endif
+
+}  // namespace __tsan
+
+// Must be included in this file to make sure everything is inlined.
+#include "tsan_interface_inl.h"
diff --git a/lib/tsan/rtl/tsan_rtl.h b/lib/tsan/rtl/tsan_rtl.h
new file mode 100644
index 000000000..11179045c
--- /dev/null
+++ b/lib/tsan/rtl/tsan_rtl.h
@@ -0,0 +1,459 @@
+//===-- tsan_rtl.h ----------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+// Main internal TSan header file.
+//
+// Ground rules:
+//   - C++ run-time should not be used (static CTORs, RTTI, exceptions, static
+//     function-scope locals)
+//   - All functions/classes/etc reside in namespace __tsan, except for those
+//     declared in tsan_interface.h.
+//   - Platform-specific files should be used instead of ifdefs (*).
+//   - No system headers included in header files (*).
+//   - Platform specific headres included only into platform-specific files (*).
+//
+//  (*) Except when inlining is critical for performance.
+//===----------------------------------------------------------------------===//
+
+#ifndef TSAN_RTL_H
+#define TSAN_RTL_H
+
+#include "tsan_clock.h"
+#include "tsan_defs.h"
+#include "tsan_flags.h"
+#include "tsan_sync.h"
+#include "tsan_trace.h"
+#include "tsan_vector.h"
+#include "tsan_report.h"
+
+namespace __tsan {
+
+void Printf(const char *format, ...) FORMAT(1, 2);
+uptr Snprintf(char *buffer, uptr length, const char *format, ...)  FORMAT(3, 4);
+
+inline void NOINLINE breakhere() {
+  volatile int x = 42;
+  (void)x;
+}
+
+// FastState (from most significant bit):
+//   tid             : kTidBits
+//   epoch           : kClkBits
+//   unused          :
+//   ignore_bit      : 1
+class FastState {
+ public:
+  FastState(u64 tid, u64 epoch) {
+    x_ = tid << (64 - kTidBits);
+    x_ |= epoch << (64 - kTidBits - kClkBits);
+    CHECK(tid == this->tid());
+    CHECK(epoch == this->epoch());
+  }
+
+  explicit FastState(u64 x)
+      : x_(x) {
+  }
+
+  u64 tid() const {
+    u64 res = x_ >> (64 - kTidBits);
+    return res;
+  }
+  u64 epoch() const {
+    u64 res = (x_ << kTidBits) >> (64 - kClkBits);
+    return res;
+  };
+  void IncrementEpoch() {
+    // u64 old_epoch = epoch();
+    x_ += 1 << (64 - kTidBits - kClkBits);
+    // CHECK(old_epoch + 1 == epoch());
+  }
+  void SetIgnoreBit() { x_ |= 1; }
+  void ClearIgnoreBit() { x_ &= ~(u64)1; }
+  bool GetIgnoreBit() { return x_ & 1; }
+
+ private:
+  friend class Shadow;
+  u64 x_;
+};
+
+// Shadow (from most significant bit):
+//   tid             : kTidBits
+//   epoch           : kClkBits
+//   is_write        : 1
+//   size_log        : 2
+//   addr0           : 3
+class Shadow: public FastState {
+ public:
+  explicit Shadow(u64 x) : FastState(x) { }
+
+  explicit Shadow(const FastState &s) : FastState(s.x_) { }
+
+  void SetAddr0AndSizeLog(u64 addr0, unsigned kAccessSizeLog) {
+    DCHECK_EQ(x_ & 31, 0);
+    DCHECK_LE(addr0, 7);
+    DCHECK_LE(kAccessSizeLog, 3);
+    x_ |= (kAccessSizeLog << 3) | addr0;
+    DCHECK_EQ(kAccessSizeLog, size_log());
+    DCHECK_EQ(addr0, this->addr0());
+  }
+
+  void SetWrite(unsigned kAccessIsWrite) {
+    DCHECK_EQ(x_ & 32, 0);
+    if (kAccessIsWrite)
+      x_ |= 32;
+    DCHECK_EQ(kAccessIsWrite, is_write());
+  }
+
+  bool IsZero() const { return x_ == 0; }
+  u64 raw() const { return x_; }
+
+  static inline bool TidsAreEqual(Shadow s1, Shadow s2) {
+    u64 shifted_xor = (s1.x_ ^ s2.x_) >> (64 - kTidBits);
+    DCHECK_EQ(shifted_xor == 0, s1.tid() == s2.tid());
+    return shifted_xor == 0;
+  }
+  static inline bool Addr0AndSizeAreEqual(Shadow s1, Shadow s2) {
+    u64 masked_xor = (s1.x_ ^ s2.x_) & 31;
+    return masked_xor == 0;
+  }
+
+  static bool TwoRangesIntersectSLOW(Shadow s1, Shadow s2) {
+    if (s1.addr0() == s2.addr0()) return true;
+    if (s1.addr0() < s2.addr0() && s1.addr0() + s1.size() > s2.addr0())
+      return true;
+    if (s2.addr0() < s1.addr0() && s2.addr0() + s2.size() > s1.addr0())
+      return true;
+    return false;
+  }
+
+  static inline bool TwoRangesIntersect(Shadow s1, Shadow s2,
+      unsigned kS2AccessSize) {
+    bool res = false;
+    u64 diff = s1.addr0() - s2.addr0();
+    if ((s64)diff < 0) {  // s1.addr0 < s2.addr0  // NOLINT
+      // if (s1.addr0() + size1) > s2.addr0()) return true;
+      if (s1.size() > -diff)  res = true;
+    } else {
+      // if (s2.addr0() + kS2AccessSize > s1.addr0()) return true;
+      if (kS2AccessSize > diff) res = true;
+    }
+    DCHECK_EQ(res, TwoRangesIntersectSLOW(s1, s2));
+    DCHECK_EQ(res, TwoRangesIntersectSLOW(s2, s1));
+    return res;
+  }
+
+  // The idea behind the offset is as follows.
+  // Consider that we have 8 bool's contained within a single 8-byte block
+  // (mapped to a single shadow "cell"). Now consider that we write to the bools
+  // from a single thread (which we consider the common case).
+  // W/o offsetting each access will have to scan 4 shadow values at average
+  // to find the corresponding shadow value for the bool.
+  // With offsetting we start scanning shadow with the offset so that
+  // each access hits necessary shadow straight off (at least in an expected
+  // optimistic case).
+  // This logic works seamlessly for any layout of user data. For example,
+  // if user data is {int, short, char, char}, then accesses to the int are
+  // offsetted to 0, short - 4, 1st char - 6, 2nd char - 7. Hopefully, accesses
+  // from a single thread won't need to scan all 8 shadow values.
+  unsigned ComputeSearchOffset() {
+    return x_ & 7;
+  }
+  u64 addr0() const { return x_ & 7; }
+  u64 size() const { return 1ull << size_log(); }
+  bool is_write() const { return x_ & 32; }
+
+ private:
+  u64 size_log() const { return (x_ >> 3) & 3; }
+};
+
+// Freed memory.
+// As if 8-byte write by thread 0xff..f at epoch 0xff..f, races with everything.
+const u64 kShadowFreed = 0xfffffffffffffff8ull;
+
+const int kSigCount = 1024;
+const int kShadowStackSize = 1024;
+
+struct my_siginfo_t {
+  int opaque[128];
+};
+
+struct SignalDesc {
+  bool armed;
+  bool sigaction;
+  my_siginfo_t siginfo;
+};
+
+// This struct is stored in TLS.
+struct ThreadState {
+  FastState fast_state;
+  // Synch epoch represents the threads's epoch before the last synchronization
+  // action. It allows to reduce number of shadow state updates.
+  // For example, fast_synch_epoch=100, last write to addr X was at epoch=150,
+  // if we are processing write to X from the same thread at epoch=200,
+  // we do nothing, because both writes happen in the same 'synch epoch'.
+  // That is, if another memory access does not race with the former write,
+  // it does not race with the latter as well.
+  // QUESTION: can we can squeeze this into ThreadState::Fast?
+  // E.g. ThreadState::Fast is a 44-bit, 32 are taken by synch_epoch and 12 are
+  // taken by epoch between synchs.
+  // This way we can save one load from tls.
+  u64 fast_synch_epoch;
+  // This is a slow path flag. On fast path, fast_state.GetIgnoreBit() is read.
+  // We do not distinguish beteween ignoring reads and writes
+  // for better performance.
+  int ignore_reads_and_writes;
+  uptr *shadow_stack_pos;
+  u64 *racy_shadow_addr;
+  u64 racy_state[2];
+  Trace trace;
+  uptr shadow_stack[kShadowStackSize];
+  ThreadClock clock;
+  u64 stat[StatCnt];
+  u64 int_alloc_cnt[MBlockTypeCount];
+  u64 int_alloc_siz[MBlockTypeCount];
+  const int tid;
+  int in_rtl;
+  int func_call_count;
+  const uptr stk_addr;
+  const uptr stk_size;
+  const uptr tls_addr;
+  const uptr tls_size;
+
+  DeadlockDetector deadlock_detector;
+
+  bool in_signal_handler;
+  int pending_signal_count;
+  SignalDesc pending_signals[kSigCount];
+
+  explicit ThreadState(Context *ctx, int tid, u64 epoch,
+                       uptr stk_addr, uptr stk_size,
+                       uptr tls_addr, uptr tls_size);
+};
+
+Context *CTX();
+extern THREADLOCAL char cur_thread_placeholder[];
+
+INLINE ThreadState *cur_thread() {
+  return reinterpret_cast<ThreadState *>(&cur_thread_placeholder);
+}
+
+enum ThreadStatus {
+  ThreadStatusInvalid,   // Non-existent thread, data is invalid.
+  ThreadStatusCreated,   // Created but not yet running.
+  ThreadStatusRunning,   // The thread is currently running.
+  ThreadStatusFinished,  // Joinable thread is finished but not yet joined.
+  ThreadStatusDead,      // Joined, but some info (trace) is still alive.
+};
+
+// An info about a thread that is hold for some time after its termination.
+struct ThreadDeadInfo {
+  Trace trace;
+};
+
+struct ThreadContext {
+  const int tid;
+  int unique_id;  // Non-rolling thread id.
+  uptr user_id;  // Some opaque user thread id (e.g. pthread_t).
+  ThreadState *thr;
+  ThreadStatus status;
+  bool detached;
+  int reuse_count;
+  SyncClock sync;
+  // Epoch at which the thread had started.
+  // If we see an event from the thread stamped by an older epoch,
+  // the event is from a dead thread that shared tid with this thread.
+  u64 epoch0;
+  u64 epoch1;
+  StackTrace creation_stack;
+  ThreadDeadInfo dead_info;
+  ThreadContext* dead_next;  // In dead thread list.
+
+  explicit ThreadContext(int tid);
+};
+
+struct RacyStacks {
+  MD5Hash hash[2];
+  bool operator==(const RacyStacks &other) const {
+    if (hash[0] == other.hash[0] && hash[1] == other.hash[1])
+      return true;
+    if (hash[0] == other.hash[1] && hash[1] == other.hash[0])
+      return true;
+    return false;
+  }
+};
+
+struct RacyAddress {
+  uptr addr_min;
+  uptr addr_max;
+};
+
+struct Context {
+  Context();
+
+  bool initialized;
+
+  SyncTab synctab;
+
+  Mutex report_mtx;
+  int nreported;
+  int nmissed_expected;
+
+  Mutex thread_mtx;
+  int thread_seq;
+  int unique_thread_seq;
+  int alive_threads;
+  int max_alive_threads;
+  ThreadContext *threads[kMaxTid];
+  int dead_list_size;
+  ThreadContext* dead_list_head;
+  ThreadContext* dead_list_tail;
+
+  Vector<RacyStacks> racy_stacks;
+  Vector<RacyAddress> racy_addresses;
+
+  Flags flags;
+
+  u64 stat[StatCnt];
+  u64 int_alloc_cnt[MBlockTypeCount];
+  u64 int_alloc_siz[MBlockTypeCount];
+};
+
+class ScopedInRtl {
+ public:
+  ScopedInRtl();
+  ~ScopedInRtl();
+ private:
+  ThreadState*thr_;
+  int in_rtl_;
+  int errno_;
+};
+
+class ScopedReport {
+ public:
+  explicit ScopedReport(ReportType typ);
+  ~ScopedReport();
+
+  void AddStack(const StackTrace *stack);
+  void AddMemoryAccess(uptr addr, Shadow s, const StackTrace *stack);
+  void AddThread(const ThreadContext *tctx);
+  void AddMutex(const SyncVar *s);
+  void AddLocation(uptr addr, uptr size);
+
+  const ReportDesc *GetReport() const;
+
+ private:
+  Context *ctx_;
+  ReportDesc *rep_;
+
+  ScopedReport(const ScopedReport&);
+  void operator = (const ScopedReport&);
+};
+
+void InternalAllocStatAggregate(Context *ctx, ThreadState *thr);
+void StatAggregate(u64 *dst, u64 *src);
+void StatOutput(u64 *stat);
+void ALWAYS_INLINE INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) {
+  if (kCollectStats)
+    thr->stat[typ] += n;
+}
+
+void InitializeShadowMemory();
+void InitializeInterceptors();
+void InitializeDynamicAnnotations();
+void Die() NORETURN;
+
+void ReportRace(ThreadState *thr);
+bool OutputReport(const ScopedReport &srep, ReportStack *suppress_stack = 0);
+bool IsExpectedReport(uptr addr, uptr size);
+
+#if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 1
+# define DPrintf Printf
+#else
+# define DPrintf(...)
+#endif
+
+#if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 2
+# define DPrintf2 Printf
+#else
+# define DPrintf2(...)
+#endif
+
+void Initialize(ThreadState *thr);
+int Finalize(ThreadState *thr);
+
+void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
+    int kAccessSizeLog, bool kAccessIsWrite);
+void MemoryAccessImpl(ThreadState *thr, uptr addr,
+    int kAccessSizeLog, bool kAccessIsWrite, FastState fast_state,
+    u64 *shadow_mem, Shadow cur);
+void MemoryRead1Byte(ThreadState *thr, uptr pc, uptr addr);
+void MemoryWrite1Byte(ThreadState *thr, uptr pc, uptr addr);
+void MemoryRead8Byte(ThreadState *thr, uptr pc, uptr addr);
+void MemoryWrite8Byte(ThreadState *thr, uptr pc, uptr addr);
+void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
+                       uptr size, bool is_write);
+void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size);
+void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size);
+void IgnoreCtl(ThreadState *thr, bool write, bool begin);
+
+void FuncEntry(ThreadState *thr, uptr pc);
+void FuncExit(ThreadState *thr);
+
+int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached);
+void ThreadStart(ThreadState *thr, int tid);
+void ThreadFinish(ThreadState *thr);
+int ThreadTid(ThreadState *thr, uptr pc, uptr uid);
+void ThreadJoin(ThreadState *thr, uptr pc, int tid);
+void ThreadDetach(ThreadState *thr, uptr pc, int tid);
+void ThreadFinalize(ThreadState *thr);
+
+void MutexCreate(ThreadState *thr, uptr pc, uptr addr, bool rw, bool recursive);
+void MutexDestroy(ThreadState *thr, uptr pc, uptr addr);
+void MutexLock(ThreadState *thr, uptr pc, uptr addr);
+void MutexUnlock(ThreadState *thr, uptr pc, uptr addr);
+void MutexReadLock(ThreadState *thr, uptr pc, uptr addr);
+void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr);
+void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr);
+
+void Acquire(ThreadState *thr, uptr pc, uptr addr);
+void Release(ThreadState *thr, uptr pc, uptr addr);
+
+// The hacky call uses custom calling convention and an assembly thunk.
+// It is considerably faster that a normal call for the caller
+// if it is not executed (it is intended for slow paths from hot functions).
+// The trick is that the call preserves all registers and the compiler
+// does not treat it as a call.
+// If it does not work for you, use normal call.
+#if TSAN_DEBUG == 0
+// The caller may not create the stack frame for itself at all,
+// so we create a reserve stack frame for it (1024b must be enough).
+#define HACKY_CALL(f) \
+  __asm__ __volatile__("sub $0x400, %%rsp;" \
+                       "call " #f "_thunk;" \
+                       "add $0x400, %%rsp;" ::: "memory");
+#else
+#define HACKY_CALL(f) f()
+#endif
+
+extern "C" void __tsan_trace_switch();
+void ALWAYS_INLINE INLINE TraceAddEvent(ThreadState *thr, u64 epoch,
+                                        EventType typ, uptr addr) {
+  StatInc(thr, StatEvents);
+  if (UNLIKELY((epoch % kTracePartSize) == 0))
+    HACKY_CALL(__tsan_trace_switch);
+  Event *evp = &thr->trace.events[epoch % kTraceSize];
+  Event ev = (u64)addr | ((u64)typ << 61);
+  *evp = ev;
+}
+
+}  // namespace __tsan
+
+#endif  // TSAN_RTL_H
diff --git a/lib/tsan/rtl/tsan_rtl_amd64.S b/lib/tsan/rtl/tsan_rtl_amd64.S
new file mode 100644
index 000000000..2028ec508
--- /dev/null
+++ b/lib/tsan/rtl/tsan_rtl_amd64.S
@@ -0,0 +1,71 @@
+.section .text
+
+.globl __tsan_trace_switch_thunk
+__tsan_trace_switch_thunk:
+  # Save scratch registers.
+  push %rax
+  push %rcx
+  push %rdx
+  push %rsi
+  push %rdi
+  push %r8
+  push %r9
+  push %r10
+  push %r11
+  # Align stack frame.
+  push %rbx  # non-scratch
+  mov %rsp, %rbx  # save current rsp
+  shr $4, %rsp  # clear 4 lsb, align to 16
+  shl $4, %rsp
+
+  call __tsan_trace_switch
+
+  # Unalign stack frame back.
+  mov %rbx, %rsp  # restore the original rsp
+  pop %rbx
+  # Restore scratch registers.
+  pop %r11
+  pop %r10
+  pop %r9
+  pop %r8
+  pop %rdi
+  pop %rsi
+  pop %rdx
+  pop %rcx
+  pop %rax
+  ret
+
+.globl __tsan_report_race_thunk
+__tsan_report_race_thunk:
+  # Save scratch registers.
+  push %rax
+  push %rcx
+  push %rdx
+  push %rsi
+  push %rdi
+  push %r8
+  push %r9
+  push %r10
+  push %r11
+  # Align stack frame.
+  push %rbx  # non-scratch
+  mov %rsp, %rbx  # save current rsp
+  shr $4, %rsp  # clear 4 lsb, align to 16
+  shl $4, %rsp
+
+  call __tsan_report_race
+
+  # Unalign stack frame back.
+  mov %rbx, %rsp  # restore the original rsp
+  pop %rbx
+  # Restore scratch registers.
+  pop %r11
+  pop %r10
+  pop %r9
+  pop %r8
+  pop %rdi
+  pop %rsi
+  pop %rdx
+  pop %rcx
+  pop %rax
+  ret
diff --git a/lib/tsan/rtl/tsan_rtl_mutex.cc b/lib/tsan/rtl/tsan_rtl_mutex.cc
new file mode 100644
index 000000000..f144209d8
--- /dev/null
+++ b/lib/tsan/rtl/tsan_rtl_mutex.cc
@@ -0,0 +1,209 @@
+//===-- tsan_rtl_mutex.cc ---------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#include "tsan_rtl.h"
+#include "tsan_sync.h"
+#include "tsan_report.h"
+#include "tsan_symbolize.h"
+
+namespace __tsan {
+
+void MutexCreate(ThreadState *thr, uptr pc, uptr addr,
+                 bool rw, bool recursive) {
+  Context *ctx = CTX();
+  CHECK_GT(thr->in_rtl, 0);
+  DPrintf("#%d: MutexCreate %lx\n", thr->tid, addr);
+  StatInc(thr, StatMutexCreate);
+  MemoryWrite1Byte(thr, pc, addr);
+  SyncVar *s = ctx->synctab.GetAndLock(thr, pc, addr, true);
+  s->is_rw = rw;
+  s->is_recursive = recursive;
+  s->mtx.Unlock();
+}
+
+void MutexDestroy(ThreadState *thr, uptr pc, uptr addr) {
+  Context *ctx = CTX();
+  CHECK_GT(thr->in_rtl, 0);
+  DPrintf("#%d: MutexDestroy %lx\n", thr->tid, addr);
+  StatInc(thr, StatMutexDestroy);
+  MemoryWrite1Byte(thr, pc, addr);
+  SyncVar *s = ctx->synctab.GetAndRemove(thr, pc, addr);
+  if (s == 0)
+    return;
+  if (s->owner_tid != SyncVar::kInvalidTid && !s->is_broken) {
+    s->is_broken = true;
+    ScopedReport rep(ReportTypeMutexDestroyLocked);
+    rep.AddMutex(s);
+    rep.AddLocation(s->addr, 1);
+    OutputReport(rep);
+  }
+  DestroyAndFree(s);
+}
+
+void MutexLock(ThreadState *thr, uptr pc, uptr addr) {
+  CHECK_GT(thr->in_rtl, 0);
+  DPrintf("#%d: MutexLock %lx\n", thr->tid, addr);
+  MemoryRead1Byte(thr, pc, addr);
+  thr->fast_state.IncrementEpoch();
+  TraceAddEvent(thr, thr->fast_state.epoch(), EventTypeLock, addr);
+  SyncVar *s = CTX()->synctab.GetAndLock(thr, pc, addr, true);
+  if (s->owner_tid == SyncVar::kInvalidTid) {
+    CHECK_EQ(s->recursion, 0);
+    s->owner_tid = thr->tid;
+  } else if (s->owner_tid == thr->tid) {
+    CHECK_GT(s->recursion, 0);
+  } else {
+    Printf("ThreadSanitizer WARNING: double lock\n");
+  }
+  if (s->recursion == 0) {
+    StatInc(thr, StatMutexLock);
+    thr->clock.set(thr->tid, thr->fast_state.epoch());
+    thr->clock.acquire(&s->clock);
+    StatInc(thr, StatSyncAcquire);
+    thr->clock.acquire(&s->read_clock);
+    StatInc(thr, StatSyncAcquire);
+  } else if (!s->is_recursive) {
+    StatInc(thr, StatMutexRecLock);
+  }
+  s->recursion++;
+  s->mtx.Unlock();
+}
+
+void MutexUnlock(ThreadState *thr, uptr pc, uptr addr) {
+  CHECK_GT(thr->in_rtl, 0);
+  DPrintf("#%d: MutexUnlock %lx\n", thr->tid, addr);
+  MemoryRead1Byte(thr, pc, addr);
+  thr->fast_state.IncrementEpoch();
+  TraceAddEvent(thr, thr->fast_state.epoch(), EventTypeUnlock, addr);
+  SyncVar *s = CTX()->synctab.GetAndLock(thr, pc, addr, true);
+  if (s->recursion == 0) {
+    if (!s->is_broken) {
+      s->is_broken = true;
+      Printf("ThreadSanitizer WARNING: unlock of unlocked mutex\n");
+    }
+  } else if (s->owner_tid != thr->tid) {
+    if (!s->is_broken) {
+      s->is_broken = true;
+      Printf("ThreadSanitizer WARNING: mutex unlock by another thread\n");
+    }
+  } else {
+    s->recursion--;
+    if (s->recursion == 0) {
+      StatInc(thr, StatMutexUnlock);
+      s->owner_tid = SyncVar::kInvalidTid;
+      thr->clock.set(thr->tid, thr->fast_state.epoch());
+      thr->fast_synch_epoch = thr->fast_state.epoch();
+      thr->clock.release(&s->clock);
+      StatInc(thr, StatSyncRelease);
+    } else {
+      StatInc(thr, StatMutexRecUnlock);
+    }
+  }
+  s->mtx.Unlock();
+}
+
+void MutexReadLock(ThreadState *thr, uptr pc, uptr addr) {
+  CHECK_GT(thr->in_rtl, 0);
+  DPrintf("#%d: MutexReadLock %lx\n", thr->tid, addr);
+  StatInc(thr, StatMutexReadLock);
+  MemoryRead1Byte(thr, pc, addr);
+  thr->fast_state.IncrementEpoch();
+  TraceAddEvent(thr, thr->fast_state.epoch(), EventTypeRLock, addr);
+  SyncVar *s = CTX()->synctab.GetAndLock(thr, pc, addr, false);
+  if (s->owner_tid != SyncVar::kInvalidTid)
+    Printf("ThreadSanitizer WARNING: read lock of a write locked mutex\n");
+  thr->clock.set(thr->tid, thr->fast_state.epoch());
+  thr->clock.acquire(&s->clock);
+  StatInc(thr, StatSyncAcquire);
+  s->mtx.ReadUnlock();
+}
+
+void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr) {
+  CHECK_GT(thr->in_rtl, 0);
+  DPrintf("#%d: MutexReadUnlock %lx\n", thr->tid, addr);
+  StatInc(thr, StatMutexReadUnlock);
+  MemoryRead1Byte(thr, pc, addr);
+  thr->fast_state.IncrementEpoch();
+  TraceAddEvent(thr, thr->fast_state.epoch(), EventTypeRUnlock, addr);
+  SyncVar *s = CTX()->synctab.GetAndLock(thr, pc, addr, true);
+  if (s->owner_tid != SyncVar::kInvalidTid)
+    Printf("ThreadSanitizer WARNING: read unlock of a write locked mutex\n");
+  thr->clock.set(thr->tid, thr->fast_state.epoch());
+  thr->fast_synch_epoch = thr->fast_state.epoch();
+  thr->clock.release(&s->read_clock);
+  StatInc(thr, StatSyncRelease);
+  s->mtx.Unlock();
+}
+
+void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr) {
+  CHECK_GT(thr->in_rtl, 0);
+  DPrintf("#%d: MutexReadOrWriteUnlock %lx\n", thr->tid, addr);
+  MemoryRead1Byte(thr, pc, addr);
+  SyncVar *s = CTX()->synctab.GetAndLock(thr, pc, addr, true);
+  if (s->owner_tid == SyncVar::kInvalidTid) {
+    // Seems to be read unlock.
+    StatInc(thr, StatMutexReadUnlock);
+    thr->fast_state.IncrementEpoch();
+    TraceAddEvent(thr, thr->fast_state.epoch(), EventTypeRUnlock, addr);
+    thr->clock.set(thr->tid, thr->fast_state.epoch());
+    thr->fast_synch_epoch = thr->fast_state.epoch();
+    thr->clock.release(&s->read_clock);
+    StatInc(thr, StatSyncRelease);
+  } else if (s->owner_tid == thr->tid) {
+    // Seems to be write unlock.
+    CHECK_GT(s->recursion, 0);
+    s->recursion--;
+    if (s->recursion == 0) {
+      StatInc(thr, StatMutexUnlock);
+      s->owner_tid = SyncVar::kInvalidTid;
+      // FIXME: Refactor me, plz.
+      // The sequence of events is quite tricky and doubled in several places.
+      // First, it's a bug to increment the epoch w/o writing to the trace.
+      // Then, the acquire/release logic can be factored out as well.
+      thr->fast_state.IncrementEpoch();
+      TraceAddEvent(thr, thr->fast_state.epoch(), EventTypeUnlock, addr);
+      thr->clock.set(thr->tid, thr->fast_state.epoch());
+      thr->fast_synch_epoch = thr->fast_state.epoch();
+      thr->clock.release(&s->clock);
+      StatInc(thr, StatSyncRelease);
+    } else {
+      StatInc(thr, StatMutexRecUnlock);
+    }
+  } else if (!s->is_broken) {
+    s->is_broken = true;
+    Printf("ThreadSanitizer WARNING: mutex unlock by another thread\n");
+  }
+  s->mtx.Unlock();
+}
+
+void Acquire(ThreadState *thr, uptr pc, uptr addr) {
+  CHECK_GT(thr->in_rtl, 0);
+  DPrintf("#%d: Acquire %lx\n", thr->tid, addr);
+  SyncVar *s = CTX()->synctab.GetAndLock(thr, pc, addr, false);
+  thr->clock.set(thr->tid, thr->fast_state.epoch());
+  thr->clock.acquire(&s->clock);
+  StatInc(thr, StatSyncAcquire);
+  s->mtx.ReadUnlock();
+}
+
+void Release(ThreadState *thr, uptr pc, uptr addr) {
+  CHECK_GT(thr->in_rtl, 0);
+  DPrintf("#%d: Release %lx\n", thr->tid, addr);
+  SyncVar *s = CTX()->synctab.GetAndLock(thr, pc, addr, true);
+  thr->clock.set(thr->tid, thr->fast_state.epoch());
+  thr->clock.release(&s->clock);
+  StatInc(thr, StatSyncRelease);
+  s->mtx.Unlock();
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_rtl_report.cc b/lib/tsan/rtl/tsan_rtl_report.cc
new file mode 100644
index 000000000..a5c541753
--- /dev/null
+++ b/lib/tsan/rtl/tsan_rtl_report.cc
@@ -0,0 +1,354 @@
+//===-- tsan_rtl.cc ---------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#include "tsan_platform.h"
+#include "tsan_rtl.h"
+#include "tsan_suppressions.h"
+#include "tsan_symbolize.h"
+#include "tsan_report.h"
+#include "tsan_sync.h"
+#include "tsan_mman.h"
+#include "tsan_flags.h"
+#include "tsan_placement_new.h"
+
+namespace __tsan {
+
+// Can be overriden by an application/test to intercept reports.
+bool WEAK OnReport(const ReportDesc *rep, bool suppressed) {
+  (void)rep;
+  return suppressed;
+}
+
+static void StackStripMain(ReportStack *stack) {
+  ReportStack *last_frame = 0;
+  ReportStack *last_frame2 = 0;
+  const char *prefix = "interception_wrap_";
+  uptr prefix_len = internal_strlen(prefix);
+  const char *path_prefix = flags()->strip_path_prefix;
+  uptr path_prefix_len = internal_strlen(path_prefix);
+  for (ReportStack *ent = stack; ent; ent = ent->next) {
+    if (ent->func && 0 == internal_strncmp(ent->func, prefix, prefix_len))
+      ent->func += prefix_len;
+    if (ent->file && 0 == internal_strncmp(ent->file, path_prefix,
+                                           path_prefix_len))
+      ent->file += path_prefix_len;
+    if (ent->file && ent->file[0] == '.' && ent->file[1] == '/')
+      ent->file += 2;
+    last_frame2 = last_frame;
+    last_frame = ent;
+  }
+
+  if (last_frame2 == 0)
+    return;
+  const char *last = last_frame->func;
+  const char *last2 = last_frame2->func;
+  // Strip frame above 'main'
+  if (last2 && 0 == internal_strcmp(last2, "main")) {
+    last_frame2->next = 0;
+  // Strip our internal thread start routine.
+  } else if (last && 0 == internal_strcmp(last, "__tsan_thread_start_func")) {
+    last_frame2->next = 0;
+  // Strip global ctors init.
+  } else if (last && 0 == internal_strcmp(last, "__do_global_ctors_aux")) {
+    last_frame2->next = 0;
+  // If both are 0, then we probably just failed to symbolize.
+  } else if (last || last2) {
+    // Ensure that we recovered stack completely. Trimmed stack
+    // can actually happen if we do not instrument some code,
+    // so it's only a DCHECK. However we must try hard to not miss it
+    // due to our fault.
+    Printf("Bottom stack frame of stack %lx is missed\n", stack->pc);
+  }
+}
+
+static ReportStack *SymbolizeStack(const StackTrace& trace) {
+  if (trace.IsEmpty())
+    return 0;
+  ReportStack *stack = 0;
+  for (uptr si = 0; si < trace.Size(); si++) {
+    // We obtain the return address, that is, address of the next instruction,
+    // so offset it by 1 byte.
+    bool is_last = (si == trace.Size() - 1);
+    ReportStack *ent = SymbolizeCode(trace.Get(si) - !is_last);
+    CHECK_NE(ent, 0);
+    ReportStack *last = ent;
+    while (last->next) {
+      last->pc += !is_last;
+      last = last->next;
+    }
+    last->pc += !is_last;
+    last->next = stack;
+    stack = ent;
+  }
+  StackStripMain(stack);
+  return stack;
+}
+
+ScopedReport::ScopedReport(ReportType typ) {
+  ctx_ = CTX();
+  void *mem = internal_alloc(MBlockReport, sizeof(ReportDesc));
+  rep_ = new(mem) ReportDesc;
+  rep_->typ = typ;
+  ctx_->report_mtx.Lock();
+}
+
+ScopedReport::~ScopedReport() {
+  ctx_->report_mtx.Unlock();
+  rep_->~ReportDesc();
+  internal_free(rep_);
+}
+
+void ScopedReport::AddStack(const StackTrace *stack) {
+  ReportStack **rs = rep_->stacks.PushBack();
+  *rs = SymbolizeStack(*stack);
+}
+
+void ScopedReport::AddMemoryAccess(uptr addr, Shadow s,
+                                   const StackTrace *stack) {
+  void *mem = internal_alloc(MBlockReportMop, sizeof(ReportMop));
+  ReportMop *mop = new(mem) ReportMop;
+  rep_->mops.PushBack(mop);
+  mop->tid = s.tid();
+  mop->addr = addr + s.addr0();
+  mop->size = s.size();
+  mop->write = s.is_write();
+  mop->nmutex = 0;
+  mop->stack = SymbolizeStack(*stack);
+}
+
+void ScopedReport::AddThread(const ThreadContext *tctx) {
+  void *mem = internal_alloc(MBlockReportThread, sizeof(ReportThread));
+  ReportThread *rt = new(mem) ReportThread();
+  rep_->threads.PushBack(rt);
+  rt->id = tctx->tid;
+  rt->running = (tctx->status == ThreadStatusRunning);
+  rt->stack = SymbolizeStack(tctx->creation_stack);
+}
+
+void ScopedReport::AddMutex(const SyncVar *s) {
+  void *mem = internal_alloc(MBlockReportMutex, sizeof(ReportMutex));
+  ReportMutex *rm = new(mem) ReportMutex();
+  rep_->mutexes.PushBack(rm);
+  rm->id = 42;
+  rm->stack = SymbolizeStack(s->creation_stack);
+}
+
+void ScopedReport::AddLocation(uptr addr, uptr size) {
+  ReportStack *symb = SymbolizeData(addr);
+  if (symb) {
+    void *mem = internal_alloc(MBlockReportLoc, sizeof(ReportLocation));
+    ReportLocation *loc = new(mem) ReportLocation();
+    rep_->locs.PushBack(loc);
+    loc->type = ReportLocationGlobal;
+    loc->addr = addr;
+    loc->size = size;
+    loc->tid = 0;
+    loc->name = symb->func;
+    loc->file = symb->file;
+    loc->line = symb->line;
+    loc->stack = 0;
+    internal_free(symb);
+  }
+}
+
+const ReportDesc *ScopedReport::GetReport() const {
+  return rep_;
+}
+
+static void RestoreStack(int tid, const u64 epoch, StackTrace *stk) {
+  ThreadContext *tctx = CTX()->threads[tid];
+  if (tctx == 0)
+    return;
+  Trace* trace = 0;
+  if (tctx->status == ThreadStatusRunning) {
+    CHECK(tctx->thr);
+    trace = &tctx->thr->trace;
+  } else if (tctx->status == ThreadStatusFinished
+      || tctx->status == ThreadStatusDead) {
+    trace = &tctx->dead_info.trace;
+  } else {
+    return;
+  }
+  Lock l(&trace->mtx);
+  const int partidx = (epoch / (kTraceSize / kTraceParts)) % kTraceParts;
+  TraceHeader* hdr = &trace->headers[partidx];
+  if (epoch < hdr->epoch0)
+    return;
+  const u64 eend = epoch % kTraceSize;
+  const u64 ebegin = eend / kTracePartSize * kTracePartSize;
+  DPrintf("#%d: RestoreStack epoch=%llu ebegin=%llu eend=%llu partidx=%d\n",
+      tid, epoch, ebegin, eend, partidx);
+  InternalScopedBuf<uptr> stack(1024);  // FIXME: de-hardcode 1024
+  for (uptr i = 0; i < hdr->stack0.Size(); i++) {
+    stack[i] = hdr->stack0.Get(i);
+    DPrintf2("  #%02lu: pc=%lx\n", i, stack[i]);
+  }
+  uptr pos = hdr->stack0.Size();
+  for (uptr i = ebegin; i <= eend; i++) {
+    Event ev = trace->events[i];
+    EventType typ = (EventType)(ev >> 61);
+    uptr pc = (uptr)(ev & 0xffffffffffffull);
+    DPrintf2("  %lu typ=%d pc=%lx\n", i, typ, pc);
+    if (typ == EventTypeMop) {
+      stack[pos] = pc;
+    } else if (typ == EventTypeFuncEnter) {
+      stack[pos++] = pc;
+    } else if (typ == EventTypeFuncExit) {
+      // Since we have full stacks, this should never happen.
+      DCHECK_GT(pos, 0);
+      if (pos > 0)
+        pos--;
+    }
+    for (uptr j = 0; j <= pos; j++)
+      DPrintf2("      #%lu: %lx\n", j, stack[j]);
+  }
+  if (pos == 0 && stack[0] == 0)
+    return;
+  pos++;
+  stk->Init(stack, pos);
+}
+
+static bool HandleRacyStacks(ThreadState *thr, const StackTrace (&traces)[2],
+    uptr addr_min, uptr addr_max) {
+  Context *ctx = CTX();
+  bool equal_stack = false;
+  RacyStacks hash = {};
+  if (flags()->suppress_equal_stacks) {
+    hash.hash[0] = md5_hash(traces[0].Begin(), traces[0].Size() * sizeof(uptr));
+    hash.hash[1] = md5_hash(traces[1].Begin(), traces[1].Size() * sizeof(uptr));
+    for (uptr i = 0; i < ctx->racy_stacks.Size(); i++) {
+      if (hash == ctx->racy_stacks[i]) {
+        DPrintf("ThreadSanitizer: suppressing report as doubled (stack)\n");
+        equal_stack = true;
+        break;
+      }
+    }
+  }
+  bool equal_address = false;
+  RacyAddress ra0 = {addr_min, addr_max};
+  if (flags()->suppress_equal_addresses) {
+    for (uptr i = 0; i < ctx->racy_addresses.Size(); i++) {
+      RacyAddress ra2 = ctx->racy_addresses[i];
+      uptr maxbeg = max(ra0.addr_min, ra2.addr_min);
+      uptr minend = min(ra0.addr_max, ra2.addr_max);
+      if (maxbeg < minend) {
+        DPrintf("ThreadSanitizer: suppressing report as doubled (addr)\n");
+        equal_address = true;
+        break;
+      }
+    }
+  }
+  if (equal_stack || equal_address) {
+    if (!equal_stack)
+      ctx->racy_stacks.PushBack(hash);
+    if (!equal_address)
+      ctx->racy_addresses.PushBack(ra0);
+    return true;
+  }
+  return false;
+}
+
+static void AddRacyStacks(ThreadState *thr, const StackTrace (&traces)[2],
+    uptr addr_min, uptr addr_max) {
+  Context *ctx = CTX();
+  if (flags()->suppress_equal_stacks) {
+    RacyStacks hash;
+    hash.hash[0] = md5_hash(traces[0].Begin(), traces[0].Size() * sizeof(uptr));
+    hash.hash[1] = md5_hash(traces[1].Begin(), traces[1].Size() * sizeof(uptr));
+    ctx->racy_stacks.PushBack(hash);
+  }
+  if (flags()->suppress_equal_addresses) {
+    RacyAddress ra0 = {addr_min, addr_max};
+    ctx->racy_addresses.PushBack(ra0);
+  }
+}
+
+bool OutputReport(const ScopedReport &srep, ReportStack *suppress_stack) {
+  const ReportDesc *rep = srep.GetReport();
+  bool suppressed = IsSuppressed(rep->typ, suppress_stack);
+  suppressed = OnReport(rep, suppressed);
+  if (suppressed)
+    return false;
+  PrintReport(rep);
+  CTX()->nreported++;
+  return true;
+}
+
+void ReportRace(ThreadState *thr) {
+  ScopedInRtl in_rtl;
+  uptr addr = ShadowToMem((uptr)thr->racy_shadow_addr);
+  uptr addr_min = 0;
+  uptr addr_max = 0;
+  {
+    uptr a0 = addr + Shadow(thr->racy_state[0]).addr0();
+    uptr a1 = addr + Shadow(thr->racy_state[1]).addr0();
+    uptr e0 = a0 + Shadow(thr->racy_state[0]).size();
+    uptr e1 = a1 + Shadow(thr->racy_state[1]).size();
+    addr_min = min(a0, a1);
+    addr_max = max(e0, e1);
+    if (IsExpectedReport(addr_min, addr_max - addr_min))
+      return;
+  }
+
+  Context *ctx = CTX();
+  Lock l0(&ctx->thread_mtx);
+
+  ScopedReport rep(ReportTypeRace);
+  const uptr nmop = thr->racy_state[1] == kShadowFreed ? 1 : 2;
+
+  StackTrace traces[2];
+  for (uptr i = 0; i < nmop; i++) {
+    Shadow s(thr->racy_state[i]);
+    RestoreStack(s.tid(), s.epoch(), &traces[i]);
+  }
+
+  if (HandleRacyStacks(thr, traces, addr_min, addr_max))
+    return;
+
+  for (uptr i = 0; i < nmop; i++) {
+    Shadow s(thr->racy_state[i]);
+    rep.AddMemoryAccess(addr, s, &traces[i]);
+  }
+
+  // Ensure that we have at least something for the current thread.
+  CHECK_EQ(traces[0].IsEmpty(), false);
+
+  for (uptr i = 0; i < nmop; i++) {
+    FastState s(thr->racy_state[i]);
+    ThreadContext *tctx = ctx->threads[s.tid()];
+    if (s.epoch() < tctx->epoch0 || s.epoch() > tctx->epoch1)
+      continue;
+    rep.AddThread(tctx);
+  }
+
+  if (!OutputReport(rep, rep.GetReport()->mops[0]->stack))
+    return;
+
+  AddRacyStacks(thr, traces, addr_min, addr_max);
+
+  // Bump the thread's clock a bit.
+  // This avoids series of similar reports between the same threads
+  // that happen close to each other (e.g. accessing several fields
+  // of the same object).
+  FastState s(thr->racy_state[1]);
+  thr->clock.set(s.tid(), s.epoch() + 100);
+}
+
+void CheckFailed(const char *file, int line, const char *cond, u64 v1, u64 v2) {
+  ScopedInRtl in_rtl;
+  Printf("FATAL: ThreadSanitizer CHECK failed: %s:%d \"%s\" (%llx, %llx)\n",
+         file, line, cond, v1, v2);
+  Die();
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_rtl_thread.cc b/lib/tsan/rtl/tsan_rtl_thread.cc
new file mode 100644
index 000000000..b87482642
--- /dev/null
+++ b/lib/tsan/rtl/tsan_rtl_thread.cc
@@ -0,0 +1,368 @@
+//===-- tsan_rtl_thread.cc --------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#include "tsan_rtl.h"
+#include "tsan_mman.h"
+#include "tsan_placement_new.h"
+#include "tsan_platform.h"
+#include "tsan_report.h"
+#include "tsan_sync.h"
+
+namespace __tsan {
+
+const int kThreadQuarantineSize = 100;
+
+static void MaybeReportThreadLeak(ThreadContext *tctx) {
+  if (tctx->detached)
+    return;
+  if (tctx->status != ThreadStatusCreated
+      && tctx->status != ThreadStatusRunning
+      && tctx->status != ThreadStatusFinished)
+    return;
+  ScopedReport rep(ReportTypeThreadLeak);
+  rep.AddThread(tctx);
+  OutputReport(rep);
+}
+
+void ThreadFinalize(ThreadState *thr) {
+  CHECK_GT(thr->in_rtl, 0);
+  if (!flags()->report_thread_leaks)
+    return;
+  Context *ctx = CTX();
+  Lock l(&ctx->thread_mtx);
+  for (int i = 0; i < kMaxTid; i++) {
+    ThreadContext *tctx = ctx->threads[i];
+    if (tctx == 0)
+      continue;
+    MaybeReportThreadLeak(tctx);
+    DestroyAndFree(tctx);
+    ctx->threads[i] = 0;
+  }
+}
+
+static void ThreadDead(ThreadState *thr, ThreadContext *tctx) {
+  Context *ctx = CTX();
+  CHECK_GT(thr->in_rtl, 0);
+  CHECK(tctx->status == ThreadStatusRunning
+      || tctx->status == ThreadStatusFinished);
+  DPrintf("#%d: ThreadDead uid=%lu\n", thr->tid, tctx->user_id);
+  tctx->status = ThreadStatusDead;
+  tctx->user_id = 0;
+  tctx->sync.Reset();
+
+  // Put to dead list.
+  tctx->dead_next = 0;
+  if (ctx->dead_list_size == 0)
+    ctx->dead_list_head = tctx;
+  else
+    ctx->dead_list_tail->dead_next = tctx;
+  ctx->dead_list_tail = tctx;
+  ctx->dead_list_size++;
+}
+
+int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached) {
+  CHECK_GT(thr->in_rtl, 0);
+  Context *ctx = CTX();
+  Lock l(&ctx->thread_mtx);
+  StatInc(thr, StatThreadCreate);
+  int tid = -1;
+  ThreadContext *tctx = 0;
+  if (ctx->dead_list_size > kThreadQuarantineSize
+      || ctx->thread_seq >= kMaxTid) {
+    if (ctx->dead_list_size == 0) {
+      Printf("ThreadSanitizer: %d thread limit exceeded. Dying.\n", kMaxTid);
+      Die();
+    }
+    StatInc(thr, StatThreadReuse);
+    tctx = ctx->dead_list_head;
+    ctx->dead_list_head = tctx->dead_next;
+    ctx->dead_list_size--;
+    if (ctx->dead_list_size == 0) {
+      CHECK_EQ(tctx->dead_next, 0);
+      ctx->dead_list_head = 0;
+    }
+    CHECK_EQ(tctx->status, ThreadStatusDead);
+    tctx->status = ThreadStatusInvalid;
+    tctx->reuse_count++;
+    tid = tctx->tid;
+    // The point to reclain dead_info.
+    // delete tctx->dead_info;
+  } else {
+    StatInc(thr, StatThreadMaxTid);
+    tid = ctx->thread_seq++;
+    void *mem = internal_alloc(MBlockThreadContex, sizeof(ThreadContext));
+    tctx = new(mem) ThreadContext(tid);
+    ctx->threads[tid] = tctx;
+  }
+  CHECK_NE(tctx, 0);
+  CHECK_GE(tid, 0);
+  CHECK_LT(tid, kMaxTid);
+  DPrintf("#%d: ThreadCreate tid=%d uid=%lu\n", thr->tid, tid, uid);
+  CHECK_EQ(tctx->status, ThreadStatusInvalid);
+  ctx->alive_threads++;
+  if (ctx->max_alive_threads < ctx->alive_threads) {
+    ctx->max_alive_threads++;
+    CHECK_EQ(ctx->max_alive_threads, ctx->alive_threads);
+    StatInc(thr, StatThreadMaxAlive);
+  }
+  tctx->status = ThreadStatusCreated;
+  tctx->thr = 0;
+  tctx->user_id = uid;
+  tctx->unique_id = ctx->unique_thread_seq++;
+  tctx->detached = detached;
+  if (tid) {
+    thr->fast_state.IncrementEpoch();
+    // Can't increment epoch w/o writing to the trace as well.
+    TraceAddEvent(thr, thr->fast_state.epoch(), EventTypeMop, 0);
+    thr->clock.set(thr->tid, thr->fast_state.epoch());
+    thr->fast_synch_epoch = thr->fast_state.epoch();
+    thr->clock.release(&tctx->sync);
+    StatInc(thr, StatSyncRelease);
+
+    tctx->creation_stack.ObtainCurrent(thr, pc);
+  }
+  return tid;
+}
+
+void ThreadStart(ThreadState *thr, int tid) {
+  CHECK_GT(thr->in_rtl, 0);
+  uptr stk_addr = 0;
+  uptr stk_size = 0;
+  uptr tls_addr = 0;
+  uptr tls_size = 0;
+  GetThreadStackAndTls(&stk_addr, &stk_size, &tls_addr, &tls_size);
+
+  MemoryResetRange(thr, /*pc=*/ 1, stk_addr, stk_size);
+
+  // Check that the thr object is in tls;
+  const uptr thr_beg = (uptr)thr;
+  const uptr thr_end = (uptr)thr + sizeof(*thr);
+  CHECK_GE(thr_beg, tls_addr);
+  CHECK_LE(thr_beg, tls_addr + tls_size);
+  CHECK_GE(thr_end, tls_addr);
+  CHECK_LE(thr_end, tls_addr + tls_size);
+  // Since the thr object is huge, skip it.
+  MemoryResetRange(thr, /*pc=*/ 2, tls_addr, thr_beg - tls_addr);
+  MemoryResetRange(thr, /*pc=*/ 2, thr_end, tls_addr + tls_size - thr_end);
+
+  Lock l(&CTX()->thread_mtx);
+  ThreadContext *tctx = CTX()->threads[tid];
+  CHECK_NE(tctx, 0);
+  CHECK_EQ(tctx->status, ThreadStatusCreated);
+  tctx->status = ThreadStatusRunning;
+  tctx->epoch0 = tctx->epoch1 + 1;
+  tctx->epoch1 = (u64)-1;
+  new(thr) ThreadState(CTX(), tid, tctx->epoch0, stk_addr, stk_size,
+                       tls_addr, tls_size);
+  tctx->thr = thr;
+  thr->fast_synch_epoch = tctx->epoch0;
+  thr->clock.set(tid, tctx->epoch0);
+  thr->clock.acquire(&tctx->sync);
+  StatInc(thr, StatSyncAcquire);
+  DPrintf("#%d: ThreadStart epoch=%llu stk_addr=%lx stk_size=%lx "
+      "tls_addr=%lx tls_size=%lx\n",
+      tid, tctx->epoch0, stk_addr, stk_size, tls_addr, tls_size);
+}
+
+void ThreadFinish(ThreadState *thr) {
+  CHECK_GT(thr->in_rtl, 0);
+  StatInc(thr, StatThreadFinish);
+  // FIXME: Treat it as write.
+  if (thr->stk_addr && thr->stk_size)
+    MemoryResetRange(thr, /*pc=*/ 3, thr->stk_addr, thr->stk_size);
+  if (thr->tls_addr && thr->tls_size) {
+    const uptr thr_beg = (uptr)thr;
+    const uptr thr_end = (uptr)thr + sizeof(*thr);
+    // Since the thr object is huge, skip it.
+    MemoryResetRange(thr, /*pc=*/ 4, thr->tls_addr, thr_beg - thr->tls_addr);
+    MemoryResetRange(thr, /*pc=*/ 5,
+        thr_end, thr->tls_addr + thr->tls_size - thr_end);
+  }
+  Context *ctx = CTX();
+  Lock l(&ctx->thread_mtx);
+  ThreadContext *tctx = ctx->threads[thr->tid];
+  CHECK_NE(tctx, 0);
+  CHECK_EQ(tctx->status, ThreadStatusRunning);
+  CHECK_GT(ctx->alive_threads, 0);
+  ctx->alive_threads--;
+  if (tctx->detached) {
+    ThreadDead(thr, tctx);
+  } else {
+    thr->fast_state.IncrementEpoch();
+    // Can't increment epoch w/o writing to the trace as well.
+    TraceAddEvent(thr, thr->fast_state.epoch(), EventTypeMop, 0);
+    thr->clock.set(thr->tid, thr->fast_state.epoch());
+    thr->fast_synch_epoch = thr->fast_state.epoch();
+    thr->clock.release(&tctx->sync);
+    StatInc(thr, StatSyncRelease);
+    tctx->status = ThreadStatusFinished;
+  }
+
+  // Save from info about the thread.
+  // If dead_info will become dynamically allocated again,
+  // it is the point to allocate it.
+  // tctx->dead_info = new ThreadDeadInfo;
+  internal_memcpy(&tctx->dead_info.trace.events[0],
+      &thr->trace.events[0], sizeof(thr->trace.events));
+  for (int i = 0; i < kTraceParts; i++) {
+    tctx->dead_info.trace.headers[i].stack0.CopyFrom(
+        thr->trace.headers[i].stack0);
+  }
+  tctx->epoch1 = thr->clock.get(tctx->tid);
+
+  thr->~ThreadState();
+  StatAggregate(ctx->stat, thr->stat);
+  InternalAllocStatAggregate(ctx, thr);
+  tctx->thr = 0;
+}
+
+int ThreadTid(ThreadState *thr, uptr pc, uptr uid) {
+  CHECK_GT(thr->in_rtl, 0);
+  DPrintf("#%d: ThreadTid uid=%lu\n", thr->tid, uid);
+  Lock l(&CTX()->thread_mtx);
+  for (int tid = 0; tid < kMaxTid; tid++) {
+    if (CTX()->threads[tid] != 0
+        && CTX()->threads[tid]->user_id == uid
+        && CTX()->threads[tid]->status != ThreadStatusInvalid)
+      return tid;
+  }
+  return -1;
+}
+
+void ThreadJoin(ThreadState *thr, uptr pc, int tid) {
+  CHECK_GT(thr->in_rtl, 0);
+  CHECK_GT(tid, 0);
+  CHECK_LT(tid, kMaxTid);
+  DPrintf("#%d: ThreadJoin tid=%d\n", thr->tid, tid);
+  Context *ctx = CTX();
+  Lock l(&ctx->thread_mtx);
+  ThreadContext *tctx = ctx->threads[tid];
+  if (tctx->status == ThreadStatusInvalid) {
+    Printf("ThreadSanitizer: join of non-existent thread\n");
+    return;
+  }
+  CHECK_EQ(tctx->detached, false);
+  CHECK_EQ(tctx->status, ThreadStatusFinished);
+  thr->clock.acquire(&tctx->sync);
+  StatInc(thr, StatSyncAcquire);
+  ThreadDead(thr, tctx);
+}
+
+void ThreadDetach(ThreadState *thr, uptr pc, int tid) {
+  CHECK_GT(thr->in_rtl, 0);
+  CHECK_GT(tid, 0);
+  CHECK_LT(tid, kMaxTid);
+  Context *ctx = CTX();
+  Lock l(&ctx->thread_mtx);
+  ThreadContext *tctx = ctx->threads[tid];
+  if (tctx->status == ThreadStatusInvalid) {
+    Printf("ThreadSanitizer: detach of non-existent thread\n");
+    return;
+  }
+  if (tctx->status == ThreadStatusFinished) {
+    ThreadDead(thr, tctx);
+  } else {
+    tctx->detached = true;
+  }
+}
+
+void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
+                       uptr size, bool is_write) {
+  if (size == 0)
+    return;
+
+  u64 *shadow_mem = (u64*)MemToShadow(addr);
+  DPrintf2("#%d: MemoryAccessRange: @%p %p size=%d is_write=%d\n",
+      thr->tid, (void*)pc, (void*)addr,
+      (int)size, is_write);
+
+#if TSAN_DEBUG
+  if (!IsAppMem(addr)) {
+    Printf("Access to non app mem %lx\n", addr);
+    DCHECK(IsAppMem(addr));
+  }
+  if (!IsAppMem(addr + size - 1)) {
+    Printf("Access to non app mem %lx\n", addr + size - 1);
+    DCHECK(IsAppMem(addr + size - 1));
+  }
+  if (!IsShadowMem((uptr)shadow_mem)) {
+    Printf("Bad shadow addr %p (%lx)\n", shadow_mem, addr);
+    DCHECK(IsShadowMem((uptr)shadow_mem));
+  }
+  if (!IsShadowMem((uptr)(shadow_mem + size * kShadowCnt / 8 - 1))) {
+    Printf("Bad shadow addr %p (%lx)\n",
+        shadow_mem + size * kShadowCnt / 8 - 1, addr + size - 1);
+    DCHECK(IsShadowMem((uptr)(shadow_mem + size * kShadowCnt / 8 - 1)));
+  }
+#endif
+
+  StatInc(thr, StatMopRange);
+
+  FastState fast_state = thr->fast_state;
+  if (fast_state.GetIgnoreBit())
+    return;
+
+  fast_state.IncrementEpoch();
+  thr->fast_state = fast_state;
+  TraceAddEvent(thr, fast_state.epoch(), EventTypeMop, pc);
+
+  bool unaligned = (addr % kShadowCell) != 0;
+
+  // Handle unaligned beginning, if any.
+  for (; addr % kShadowCell && size; addr++, size--) {
+    int const kAccessSizeLog = 0;
+    Shadow cur(fast_state);
+    cur.SetWrite(is_write);
+    cur.SetAddr0AndSizeLog(addr & (kShadowCell - 1), kAccessSizeLog);
+    MemoryAccessImpl(thr, addr, kAccessSizeLog, is_write, fast_state,
+        shadow_mem, cur);
+  }
+  if (unaligned)
+    shadow_mem += kShadowCnt;
+  // Handle middle part, if any.
+  for (; size >= kShadowCell; addr += kShadowCell, size -= kShadowCell) {
+    int const kAccessSizeLog = 3;
+    Shadow cur(fast_state);
+    cur.SetWrite(is_write);
+    cur.SetAddr0AndSizeLog(0, kAccessSizeLog);
+    MemoryAccessImpl(thr, addr, kAccessSizeLog, is_write, fast_state,
+        shadow_mem, cur);
+    shadow_mem += kShadowCnt;
+  }
+  // Handle ending, if any.
+  for (; size; addr++, size--) {
+    int const kAccessSizeLog = 0;
+    Shadow cur(fast_state);
+    cur.SetWrite(is_write);
+    cur.SetAddr0AndSizeLog(addr & (kShadowCell - 1), kAccessSizeLog);
+    MemoryAccessImpl(thr, addr, kAccessSizeLog, is_write, fast_state,
+        shadow_mem, cur);
+  }
+}
+
+void MemoryRead1Byte(ThreadState *thr, uptr pc, uptr addr) {
+  MemoryAccess(thr, pc, addr, 0, 0);
+}
+
+void MemoryWrite1Byte(ThreadState *thr, uptr pc, uptr addr) {
+  MemoryAccess(thr, pc, addr, 0, 1);
+}
+
+void MemoryRead8Byte(ThreadState *thr, uptr pc, uptr addr) {
+  MemoryAccess(thr, pc, addr, 3, 0);
+}
+
+void MemoryWrite8Byte(ThreadState *thr, uptr pc, uptr addr) {
+  MemoryAccess(thr, pc, addr, 3, 1);
+}
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_stat.cc b/lib/tsan/rtl/tsan_stat.cc
new file mode 100644
index 000000000..21e1ee73f
--- /dev/null
+++ b/lib/tsan/rtl/tsan_stat.cc
@@ -0,0 +1,247 @@
+//===-- tsan_stat.cc --------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#include "tsan_stat.h"
+#include "tsan_rtl.h"
+
+namespace __tsan {
+
+void StatAggregate(u64 *dst, u64 *src) {
+  if (!kCollectStats)
+    return;
+  for (int i = 0; i < StatCnt; i++)
+    dst[i] += src[i];
+}
+
+void StatOutput(u64 *stat) {
+  if (!kCollectStats)
+    return;
+
+  stat[StatShadowNonZero] = stat[StatShadowProcessed] - stat[StatShadowZero];
+
+  static const char *name[StatCnt] = {};
+  name[StatMop]                          = "Memory accesses                   ";
+  name[StatMopRead]                      = "  Including reads                 ";
+  name[StatMopWrite]                     = "            writes                ";
+  name[StatMop1]                         = "  Including size 1                ";
+  name[StatMop2]                         = "            size 2                ";
+  name[StatMop4]                         = "            size 4                ";
+  name[StatMop8]                         = "            size 8                ";
+  name[StatMopSame]                      = "  Including same                  ";
+  name[StatMopRange]                     = "  Including range                 ";
+  name[StatShadowProcessed]              = "Shadow processed                  ";
+  name[StatShadowZero]                   = "  Including empty                 ";
+  name[StatShadowNonZero]                = "  Including non empty             ";
+  name[StatShadowSameSize]               = "  Including same size             ";
+  name[StatShadowIntersect]              = "            intersect             ";
+  name[StatShadowNotIntersect]           = "            not intersect         ";
+  name[StatShadowSameThread]             = "  Including same thread           ";
+  name[StatShadowAnotherThread]          = "            another thread        ";
+  name[StatShadowReplace]                = "  Including evicted               ";
+
+  name[StatFuncEnter]                    = "Function entries                  ";
+  name[StatFuncExit]                     = "Function exits                    ";
+  name[StatEvents]                       = "Events collected                  ";
+
+  name[StatThreadCreate]                 = "Total threads created             ";
+  name[StatThreadFinish]                 = "  threads finished                ";
+  name[StatThreadReuse]                  = "  threads reused                  ";
+  name[StatThreadMaxTid]                 = "  max tid                         ";
+  name[StatThreadMaxAlive]               = "  max alive threads               ";
+
+  name[StatMutexCreate]                  = "Mutexes created                   ";
+  name[StatMutexDestroy]                 = "  destroyed                       ";
+  name[StatMutexLock]                    = "  lock                            ";
+  name[StatMutexUnlock]                  = "  unlock                          ";
+  name[StatMutexRecLock]                 = "  recursive lock                  ";
+  name[StatMutexRecUnlock]               = "  recursive unlock                ";
+  name[StatMutexReadLock]                = "  read lock                       ";
+  name[StatMutexReadUnlock]              = "  read unlock                     ";
+
+  name[StatSyncCreated]                  = "Sync objects created              ";
+  name[StatSyncDestroyed]                = "             destroyed            ";
+  name[StatSyncAcquire]                  = "             acquired             ";
+  name[StatSyncRelease]                  = "             released             ";
+
+  name[StatAtomic]                       = "Atomic operations                 ";
+  name[StatAtomicLoad]                   = "  Including load                  ";
+  name[StatAtomicStore]                  = "            store                 ";
+  name[StatAtomicExchange]               = "            exchange              ";
+  name[StatAtomicFetchAdd]               = "            fetch_add             ";
+  name[StatAtomicCAS]                    = "            compare_exchange      ";
+  name[StatAtomicFence]                  = "            fence                 ";
+  name[StatAtomicRelaxed]                = "  Including relaxed               ";
+  name[StatAtomicConsume]                = "            consume               ";
+  name[StatAtomicAcquire]                = "            acquire               ";
+  name[StatAtomicRelease]                = "            release               ";
+  name[StatAtomicAcq_Rel]                = "            acq_rel               ";
+  name[StatAtomicSeq_Cst]                = "            seq_cst               ";
+  name[StatAtomic1]                      = "  Including size 1                ";
+  name[StatAtomic2]                      = "            size 2                ";
+  name[StatAtomic4]                      = "            size 4                ";
+  name[StatAtomic8]                      = "            size 8                ";
+
+  name[StatInterceptor]                  = "Interceptors                      ";
+  name[StatInt_malloc]                   = "  malloc                          ";
+  name[StatInt_calloc]                   = "  calloc                          ";
+  name[StatInt_realloc]                  = "  realloc                         ";
+  name[StatInt_free]                     = "  free                            ";
+  name[StatInt_cfree]                    = "  cfree                           ";
+  name[StatInt_mmap]                     = "  mmap                            ";
+  name[StatInt_mmap64]                   = "  mmap64                          ";
+  name[StatInt_munmap]                   = "  munmap                          ";
+  name[StatInt_memalign]                 = "  memalign                        ";
+  name[StatInt_valloc]                   = "  valloc                          ";
+  name[StatInt_pvalloc]                  = "  pvalloc                         ";
+  name[StatInt_posix_memalign]           = "  posix_memalign                  ";
+  name[StatInt__Znwm]                    = "  _Znwm                           ";
+  name[StatInt__ZnwmRKSt9nothrow_t]      = "  _ZnwmRKSt9nothrow_t             ";
+  name[StatInt__Znam]                    = "  _Znam                           ";
+  name[StatInt__ZnamRKSt9nothrow_t]      = "  _ZnamRKSt9nothrow_t             ";
+  name[StatInt__ZdlPv]                   = "  _ZdlPv                          ";
+  name[StatInt__ZdlPvRKSt9nothrow_t]     = "  _ZdlPvRKSt9nothrow_t            ";
+  name[StatInt__ZdaPv]                   = "  _ZdaPv                          ";
+  name[StatInt__ZdaPvRKSt9nothrow_t]     = "  _ZdaPvRKSt9nothrow_t            ";
+  name[StatInt_strlen]                   = "  strlen                          ";
+  name[StatInt_memset]                   = "  memset                          ";
+  name[StatInt_memcpy]                   = "  memcpy                          ";
+  name[StatInt_strcmp]                   = "  strcmp                          ";
+  name[StatInt_memchr]                   = "  memchr                          ";
+  name[StatInt_memrchr]                  = "  memrchr                         ";
+  name[StatInt_memmove]                  = "  memmove                         ";
+  name[StatInt_memcmp]                   = "  memcmp                          ";
+  name[StatInt_strchr]                   = "  strchr                          ";
+  name[StatInt_strchrnul]                = "  strchrnul                       ";
+  name[StatInt_strrchr]                  = "  strrchr                         ";
+  name[StatInt_strncmp]                  = "  strncmp                         ";
+  name[StatInt_strcpy]                   = "  strcpy                          ";
+  name[StatInt_strncpy]                  = "  strncpy                         ";
+  name[StatInt_strstr]                   = "  strstr                          ";
+  name[StatInt_atexit]                   = "  atexit                          ";
+  name[StatInt___cxa_guard_acquire]      = "  __cxa_guard_acquire             ";
+  name[StatInt___cxa_guard_release]      = "  __cxa_guard_release             ";
+  name[StatInt_pthread_create]           = "  pthread_create                  ";
+  name[StatInt_pthread_join]             = "  pthread_join                    ";
+  name[StatInt_pthread_detach]           = "  pthread_detach                  ";
+  name[StatInt_pthread_mutex_init]       = "  pthread_mutex_init              ";
+  name[StatInt_pthread_mutex_destroy]    = "  pthread_mutex_destroy           ";
+  name[StatInt_pthread_mutex_lock]       = "  pthread_mutex_lock              ";
+  name[StatInt_pthread_mutex_trylock]    = "  pthread_mutex_trylock           ";
+  name[StatInt_pthread_mutex_timedlock]  = "  pthread_mutex_timedlock         ";
+  name[StatInt_pthread_mutex_unlock]     = "  pthread_mutex_unlock            ";
+  name[StatInt_pthread_spin_init]        = "  pthread_spin_init               ";
+  name[StatInt_pthread_spin_destroy]     = "  pthread_spin_destroy            ";
+  name[StatInt_pthread_spin_lock]        = "  pthread_spin_lock               ";
+  name[StatInt_pthread_spin_trylock]     = "  pthread_spin_trylock            ";
+  name[StatInt_pthread_spin_unlock]      = "  pthread_spin_unlock             ";
+  name[StatInt_pthread_rwlock_init]      = "  pthread_rwlock_init             ";
+  name[StatInt_pthread_rwlock_destroy]   = "  pthread_rwlock_destroy          ";
+  name[StatInt_pthread_rwlock_rdlock]    = "  pthread_rwlock_rdlock           ";
+  name[StatInt_pthread_rwlock_tryrdlock] = "  pthread_rwlock_tryrdlock        ";
+  name[StatInt_pthread_rwlock_timedrdlock]
+                                         = "  pthread_rwlock_timedrdlock      ";
+  name[StatInt_pthread_rwlock_wrlock]    = "  pthread_rwlock_wrlock           ";
+  name[StatInt_pthread_rwlock_trywrlock] = "  pthread_rwlock_trywrlock        ";
+  name[StatInt_pthread_rwlock_timedwrlock]
+                                         = "  pthread_rwlock_timedwrlock      ";
+  name[StatInt_pthread_rwlock_unlock]    = "  pthread_rwlock_unlock           ";
+  name[StatInt_pthread_cond_init]        = "  pthread_cond_init               ";
+  name[StatInt_pthread_cond_destroy]     = "  pthread_cond_destroy            ";
+  name[StatInt_pthread_cond_signal]      = "  pthread_cond_signal             ";
+  name[StatInt_pthread_cond_broadcast]   = "  pthread_cond_broadcast          ";
+  name[StatInt_pthread_cond_wait]        = "  pthread_cond_wait               ";
+  name[StatInt_pthread_cond_timedwait]   = "  pthread_cond_timedwait          ";
+  name[StatInt_pthread_barrier_init]     = "  pthread_barrier_init            ";
+  name[StatInt_pthread_barrier_destroy]  = "  pthread_barrier_destroy         ";
+  name[StatInt_pthread_barrier_wait]     = "  pthread_barrier_wait            ";
+  name[StatInt_pthread_once]             = "  pthread_once                    ";
+  name[StatInt_sem_init]                 = "  sem_init                        ";
+  name[StatInt_sem_destroy]              = "  sem_destroy                     ";
+  name[StatInt_sem_wait]                 = "  sem_wait                        ";
+  name[StatInt_sem_trywait]              = "  sem_trywait                     ";
+  name[StatInt_sem_timedwait]            = "  sem_timedwait                   ";
+  name[StatInt_sem_post]                 = "  sem_post                        ";
+  name[StatInt_sem_getvalue]             = "  sem_getvalue                    ";
+  name[StatInt_read]                     = "  read                            ";
+  name[StatInt_pread]                    = "  pread                           ";
+  name[StatInt_pread64]                  = "  pread64                         ";
+  name[StatInt_readv]                    = "  readv                           ";
+  name[StatInt_preadv64]                 = "  preadv64                        ";
+  name[StatInt_write]                    = "  write                           ";
+  name[StatInt_pwrite]                   = "  pwrite                          ";
+  name[StatInt_pwrite64]                 = "  pwrite64                        ";
+  name[StatInt_writev]                   = "  writev                          ";
+  name[StatInt_pwritev64]                = "  pwritev64                       ";
+  name[StatInt_send]                     = "  send                            ";
+  name[StatInt_sendmsg]                  = "  sendmsg                         ";
+  name[StatInt_recv]                     = "  recv                            ";
+  name[StatInt_recvmsg]                  = "  recvmsg                         ";
+  name[StatInt_unlink]                   = "  unlink                          ";
+  name[StatInt_fopen]                    = "  fopen                           ";
+  name[StatInt_fread]                    = "  fread                           ";
+  name[StatInt_fwrite]                   = "  fwrite                          ";
+  name[StatInt_puts]                     = "  puts                            ";
+  name[StatInt_rmdir]                    = "  rmdir                           ";
+  name[StatInt_opendir]                  = "  opendir                         ";
+  name[StatInt_epoll_ctl]                = "  epoll_ctl                       ";
+  name[StatInt_epoll_wait]               = "  epoll_wait                      ";
+  name[StatInt_sigaction]                = "  sigaction                       ";
+
+  name[StatAnnotation]                   = "Dynamic annotations               ";
+  name[StatAnnotateHappensBefore]        = "  HappensBefore                   ";
+  name[StatAnnotateHappensAfter]         = "  HappensAfter                    ";
+  name[StatAnnotateCondVarSignal]        = "  CondVarSignal                   ";
+  name[StatAnnotateCondVarSignalAll]     = "  CondVarSignalAll                ";
+  name[StatAnnotateMutexIsNotPHB]        = "  MutexIsNotPHB                   ";
+  name[StatAnnotateCondVarWait]          = "  CondVarWait                     ";
+  name[StatAnnotateRWLockCreate]         = "  RWLockCreate                    ";
+  name[StatAnnotateRWLockDestroy]        = "  RWLockDestroy                   ";
+  name[StatAnnotateRWLockAcquired]       = "  RWLockAcquired                  ";
+  name[StatAnnotateRWLockReleased]       = "  RWLockReleased                  ";
+  name[StatAnnotateTraceMemory]          = "  TraceMemory                     ";
+  name[StatAnnotateFlushState]           = "  FlushState                      ";
+  name[StatAnnotateNewMemory]            = "  NewMemory                       ";
+  name[StatAnnotateNoOp]                 = "  NoOp                            ";
+  name[StatAnnotateFlushExpectedRaces]   = "  FlushExpectedRaces              ";
+  name[StatAnnotateEnableRaceDetection]  = "  EnableRaceDetection             ";
+  name[StatAnnotateMutexIsUsedAsCondVar] = "  MutexIsUsedAsCondVar            ";
+  name[StatAnnotatePCQGet]               = "  PCQGet                          ";
+  name[StatAnnotatePCQPut]               = "  PCQPut                          ";
+  name[StatAnnotatePCQDestroy]           = "  PCQDestroy                      ";
+  name[StatAnnotatePCQCreate]            = "  PCQCreate                       ";
+  name[StatAnnotateExpectRace]           = "  ExpectRace                      ";
+  name[StatAnnotateBenignRaceSized]      = "  BenignRaceSized                 ";
+  name[StatAnnotateBenignRace]           = "  BenignRace                      ";
+  name[StatAnnotateIgnoreReadsBegin]     = "  IgnoreReadsBegin                ";
+  name[StatAnnotateIgnoreReadsEnd]       = "  IgnoreReadsEnd                  ";
+  name[StatAnnotateIgnoreWritesBegin]    = "  IgnoreWritesBegin               ";
+  name[StatAnnotateIgnoreWritesEnd]      = "  IgnoreWritesEnd                 ";
+  name[StatAnnotatePublishMemoryRange]   = "  PublishMemoryRange              ";
+  name[StatAnnotateUnpublishMemoryRange] = "  UnpublishMemoryRange            ";
+  name[StatAnnotateThreadName]           = "  ThreadName                      ";
+
+  name[StatMtxTotal]                     = "Contentionz                       ";
+  name[StatMtxTrace]                     = "  Trace                           ";
+  name[StatMtxThreads]                   = "  Threads                         ";
+  name[StatMtxReport]                    = "  Report                          ";
+  name[StatMtxSyncVar]                   = "  SyncVar                         ";
+  name[StatMtxSyncTab]                   = "  SyncTab                         ";
+  name[StatMtxSlab]                      = "  Slab                            ";
+  name[StatMtxAtExit]                    = "  Atexit                          ";
+  name[StatMtxAnnotations]               = "  Annotations                     ";
+
+  Printf("Statistics:\n");
+  for (int i = 0; i < StatCnt; i++)
+    Printf("%s: %llu\n", name[i], stat[i]);
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_stat.h b/lib/tsan/rtl/tsan_stat.h
new file mode 100644
index 000000000..b660b48ae
--- /dev/null
+++ b/lib/tsan/rtl/tsan_stat.h
@@ -0,0 +1,245 @@
+//===-- tsan_stat.h ---------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef TSAN_STAT_H
+#define TSAN_STAT_H
+
+namespace __tsan {
+
+enum StatType {
+  // Memory access processing related stuff.
+  StatMop,
+  StatMopRead,
+  StatMopWrite,
+  StatMop1,  // These must be consequtive.
+  StatMop2,
+  StatMop4,
+  StatMop8,
+  StatMopSame,
+  StatMopRange,
+  StatShadowProcessed,
+  StatShadowZero,
+  StatShadowNonZero,  // Derived.
+  StatShadowSameSize,
+  StatShadowIntersect,
+  StatShadowNotIntersect,
+  StatShadowSameThread,
+  StatShadowAnotherThread,
+  StatShadowReplace,
+
+  // Func processing.
+  StatFuncEnter,
+  StatFuncExit,
+
+  // Trace processing.
+  StatEvents,
+
+  // Threads.
+  StatThreadCreate,
+  StatThreadFinish,
+  StatThreadReuse,
+  StatThreadMaxTid,
+  StatThreadMaxAlive,
+
+  // Mutexes.
+  StatMutexCreate,
+  StatMutexDestroy,
+  StatMutexLock,
+  StatMutexUnlock,
+  StatMutexRecLock,
+  StatMutexRecUnlock,
+  StatMutexReadLock,
+  StatMutexReadUnlock,
+
+  // Synchronization.
+  StatSyncCreated,
+  StatSyncDestroyed,
+  StatSyncAcquire,
+  StatSyncRelease,
+
+  // Atomics.
+  StatAtomic,
+  StatAtomicLoad,
+  StatAtomicStore,
+  StatAtomicExchange,
+  StatAtomicFetchAdd,
+  StatAtomicCAS,
+  StatAtomicFence,
+  StatAtomicRelaxed,
+  StatAtomicConsume,
+  StatAtomicAcquire,
+  StatAtomicRelease,
+  StatAtomicAcq_Rel,
+  StatAtomicSeq_Cst,
+  StatAtomic1,
+  StatAtomic2,
+  StatAtomic4,
+  StatAtomic8,
+
+  // Interceptors.
+  StatInterceptor,
+  StatInt_malloc,
+  StatInt_calloc,
+  StatInt_realloc,
+  StatInt_free,
+  StatInt_cfree,
+  StatInt_mmap,
+  StatInt_mmap64,
+  StatInt_munmap,
+  StatInt_memalign,
+  StatInt_valloc,
+  StatInt_pvalloc,
+  StatInt_posix_memalign,
+  StatInt__Znwm,
+  StatInt__ZnwmRKSt9nothrow_t,
+  StatInt__Znam,
+  StatInt__ZnamRKSt9nothrow_t,
+  StatInt__ZdlPv,
+  StatInt__ZdlPvRKSt9nothrow_t,
+  StatInt__ZdaPv,
+  StatInt__ZdaPvRKSt9nothrow_t,
+  StatInt_strlen,
+  StatInt_memset,
+  StatInt_memcpy,
+  StatInt_strcmp,
+  StatInt_memchr,
+  StatInt_memrchr,
+  StatInt_memmove,
+  StatInt_memcmp,
+  StatInt_strchr,
+  StatInt_strchrnul,
+  StatInt_strrchr,
+  StatInt_strncmp,
+  StatInt_strcpy,
+  StatInt_strncpy,
+  StatInt_strstr,
+  StatInt_atexit,
+  StatInt___cxa_guard_acquire,
+  StatInt___cxa_guard_release,
+  StatInt_pthread_create,
+  StatInt_pthread_join,
+  StatInt_pthread_detach,
+  StatInt_pthread_mutex_init,
+  StatInt_pthread_mutex_destroy,
+  StatInt_pthread_mutex_lock,
+  StatInt_pthread_mutex_trylock,
+  StatInt_pthread_mutex_timedlock,
+  StatInt_pthread_mutex_unlock,
+  StatInt_pthread_spin_init,
+  StatInt_pthread_spin_destroy,
+  StatInt_pthread_spin_lock,
+  StatInt_pthread_spin_trylock,
+  StatInt_pthread_spin_unlock,
+  StatInt_pthread_rwlock_init,
+  StatInt_pthread_rwlock_destroy,
+  StatInt_pthread_rwlock_rdlock,
+  StatInt_pthread_rwlock_tryrdlock,
+  StatInt_pthread_rwlock_timedrdlock,
+  StatInt_pthread_rwlock_wrlock,
+  StatInt_pthread_rwlock_trywrlock,
+  StatInt_pthread_rwlock_timedwrlock,
+  StatInt_pthread_rwlock_unlock,
+  StatInt_pthread_cond_init,
+  StatInt_pthread_cond_destroy,
+  StatInt_pthread_cond_signal,
+  StatInt_pthread_cond_broadcast,
+  StatInt_pthread_cond_wait,
+  StatInt_pthread_cond_timedwait,
+  StatInt_pthread_barrier_init,
+  StatInt_pthread_barrier_destroy,
+  StatInt_pthread_barrier_wait,
+  StatInt_pthread_once,
+  StatInt_sem_init,
+  StatInt_sem_destroy,
+  StatInt_sem_wait,
+  StatInt_sem_trywait,
+  StatInt_sem_timedwait,
+  StatInt_sem_post,
+  StatInt_sem_getvalue,
+  StatInt_read,
+  StatInt_pread,
+  StatInt_pread64,
+  StatInt_readv,
+  StatInt_preadv64,
+  StatInt_write,
+  StatInt_pwrite,
+  StatInt_pwrite64,
+  StatInt_writev,
+  StatInt_pwritev64,
+  StatInt_send,
+  StatInt_sendmsg,
+  StatInt_recv,
+  StatInt_recvmsg,
+  StatInt_unlink,
+  StatInt_fopen,
+  StatInt_fread,
+  StatInt_fwrite,
+  StatInt_puts,
+  StatInt_rmdir,
+  StatInt_opendir,
+  StatInt_epoll_ctl,
+  StatInt_epoll_wait,
+  StatInt_sigaction,
+
+  // Dynamic annotations.
+  StatAnnotation,
+  StatAnnotateHappensBefore,
+  StatAnnotateHappensAfter,
+  StatAnnotateCondVarSignal,
+  StatAnnotateCondVarSignalAll,
+  StatAnnotateMutexIsNotPHB,
+  StatAnnotateCondVarWait,
+  StatAnnotateRWLockCreate,
+  StatAnnotateRWLockDestroy,
+  StatAnnotateRWLockAcquired,
+  StatAnnotateRWLockReleased,
+  StatAnnotateTraceMemory,
+  StatAnnotateFlushState,
+  StatAnnotateNewMemory,
+  StatAnnotateNoOp,
+  StatAnnotateFlushExpectedRaces,
+  StatAnnotateEnableRaceDetection,
+  StatAnnotateMutexIsUsedAsCondVar,
+  StatAnnotatePCQGet,
+  StatAnnotatePCQPut,
+  StatAnnotatePCQDestroy,
+  StatAnnotatePCQCreate,
+  StatAnnotateExpectRace,
+  StatAnnotateBenignRaceSized,
+  StatAnnotateBenignRace,
+  StatAnnotateIgnoreReadsBegin,
+  StatAnnotateIgnoreReadsEnd,
+  StatAnnotateIgnoreWritesBegin,
+  StatAnnotateIgnoreWritesEnd,
+  StatAnnotatePublishMemoryRange,
+  StatAnnotateUnpublishMemoryRange,
+  StatAnnotateThreadName,
+
+  // Internal mutex contentionz.
+  StatMtxTotal,
+  StatMtxTrace,
+  StatMtxThreads,
+  StatMtxReport,
+  StatMtxSyncVar,
+  StatMtxSyncTab,
+  StatMtxSlab,
+  StatMtxAnnotations,
+  StatMtxAtExit,
+
+  // This must be the last.
+  StatCnt,
+};
+
+}  // namespace __tsan
+
+#endif  // TSAN_STAT_H
diff --git a/lib/tsan/rtl/tsan_suppressions.cc b/lib/tsan/rtl/tsan_suppressions.cc
new file mode 100644
index 000000000..d74361469
--- /dev/null
+++ b/lib/tsan/rtl/tsan_suppressions.cc
@@ -0,0 +1,173 @@
+//===-- tsan_suppressions.cc ------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#include "tsan_suppressions.h"
+#include "tsan_rtl.h"
+#include "tsan_flags.h"
+#include "tsan_mman.h"
+#include "tsan_platform.h"
+
+namespace __tsan {
+
+static Suppression *g_suppressions;
+
+static char *ReadFile(const char *filename) {
+  if (filename == 0 || filename[0] == 0)
+    return 0;
+  InternalScopedBuf<char> tmp(4*1024);
+  if (filename[0] == '/')
+    Snprintf(tmp, tmp.Size(), "%s", filename);
+  else
+    Snprintf(tmp, tmp.Size(), "%s/%s", internal_getpwd(), filename);
+  fd_t fd = internal_open(tmp, false);
+  if (fd == kInvalidFd) {
+    Printf("ThreadSanitizer: failed to open suppressions file '%s'\n",
+        tmp.Ptr());
+    Die();
+  }
+  const uptr fsize = internal_filesize(fd);
+  if (fsize == (uptr)-1) {
+    Printf("ThreadSanitizer: failed to stat suppressions file '%s'\n",
+        tmp.Ptr());
+    Die();
+  }
+  char *buf = (char*)internal_alloc(MBlockSuppression, fsize + 1);
+  if (fsize != internal_read(fd, buf, fsize)) {
+    Printf("ThreadSanitizer: failed to read suppressions file '%s'\n",
+        tmp.Ptr());
+    Die();
+  }
+  internal_close(fd);
+  buf[fsize] = 0;
+  return buf;
+}
+
+bool SuppressionMatch(char *templ, const char *str) {
+  char *tpos;
+  const char *spos;
+  while (templ && templ[0]) {
+    if (templ[0] == '*') {
+      templ++;
+      continue;
+    }
+    if (str[0] == 0)
+      return false;
+    tpos = (char*)internal_strchr(templ, '*');
+    if (tpos != 0)
+      tpos[0] = 0;
+    spos = internal_strstr(str, templ);
+    str = spos + internal_strlen(templ);
+    templ = tpos;
+    if (tpos)
+      tpos[0] = '*';
+    if (spos == 0)
+      return false;
+  }
+  return true;
+}
+
+Suppression *SuppressionParse(const char* supp) {
+  Suppression *head = 0;
+  const char *line = supp;
+  while (line) {
+    while (line[0] == ' ' || line[0] == '\t')
+      line++;
+    const char *end = internal_strchr(line, '\n');
+    if (end == 0)
+      end = line + internal_strlen(line);
+    if (line != end && line[0] != '#') {
+      const char *end2 = end;
+      while (line != end2 && (end2[-1] == ' ' || end2[-1] == '\t'))
+        end2--;
+      SuppressionType stype;
+      if (0 == internal_strncmp(line, "race:", sizeof("race:") - 1)) {
+        stype = SuppressionRace;
+        line += sizeof("race:") - 1;
+      } else if (0 == internal_strncmp(line, "thread:",
+          sizeof("thread:") - 1)) {
+        stype = SuppressionThread;
+        line += sizeof("thread:") - 1;
+      } else if (0 == internal_strncmp(line, "mutex:",
+          sizeof("mutex:") - 1)) {
+        stype = SuppressionMutex;
+        line += sizeof("mutex:") - 1;
+      } else if (0 == internal_strncmp(line, "signal:",
+          sizeof("signal:") - 1)) {
+        stype = SuppressionSignal;
+        line += sizeof("signal:") - 1;
+      } else {
+        Printf("ThreadSanitizer: failed to parse suppressions file\n");
+        Die();
+      }
+      Suppression *s = (Suppression*)internal_alloc(MBlockSuppression,
+          sizeof(Suppression));
+      s->next = head;
+      head = s;
+      s->type = stype;
+      s->func = (char*)internal_alloc(MBlockSuppression, end2 - line + 1);
+      internal_memcpy(s->func, line, end2 - line);
+      s->func[end2 - line] = 0;
+    }
+    if (end[0] == 0)
+      break;
+    line = end + 1;
+  }
+  return head;
+}
+
+void SuppressionFree(Suppression *supp) {
+  while (supp) {
+    Suppression *tmp = supp;
+    supp = tmp->next;
+    internal_free(tmp->func);
+    internal_free(tmp);
+  }
+}
+
+void InitializeSuppressions() {
+  char *supp = ReadFile(flags()->suppressions);
+  g_suppressions = SuppressionParse(supp);
+}
+
+void FinalizeSuppressions() {
+  SuppressionFree(g_suppressions);
+  g_suppressions = 0;
+}
+
+bool IsSuppressed(ReportType typ, const ReportStack *stack) {
+  if (g_suppressions == 0 || stack == 0)
+    return false;
+  SuppressionType stype;
+  if (typ == ReportTypeRace)
+    stype = SuppressionRace;
+  else if (typ == ReportTypeThreadLeak)
+    stype = SuppressionThread;
+  else if (typ == ReportTypeMutexDestroyLocked)
+    stype = SuppressionMutex;
+  else if (typ == ReportTypeSignalUnsafe)
+    stype = SuppressionSignal;
+  else
+    return false;
+  for (const ReportStack *frame = stack; frame; frame = frame->next) {
+    if (frame->func == 0)
+      continue;
+    for (Suppression *supp = g_suppressions; supp; supp = supp->next) {
+      if (stype == supp->type && SuppressionMatch(supp->func, frame->func)) {
+        DPrintf("ThreadSanitizer: matched suppression '%s'\n", supp->func);
+        return true;
+      }
+    }
+  }
+  return false;
+}
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_suppressions.h b/lib/tsan/rtl/tsan_suppressions.h
new file mode 100644
index 000000000..06d7307a4
--- /dev/null
+++ b/lib/tsan/rtl/tsan_suppressions.h
@@ -0,0 +1,43 @@
+//===-- tsan_suppressions.h -------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_SUPPRESSIONS_H
+#define TSAN_SUPPRESSIONS_H
+
+#include "tsan_report.h"
+
+namespace __tsan {
+
+void InitializeSuppressions();
+void FinalizeSuppressions();
+bool IsSuppressed(ReportType typ, const ReportStack *stack);
+
+// Exposed for testing.
+enum SuppressionType {
+  SuppressionRace,
+  SuppressionMutex,
+  SuppressionThread,
+  SuppressionSignal,
+};
+
+struct Suppression {
+  Suppression *next;
+  SuppressionType type;
+  char *func;
+};
+Suppression *SuppressionParse(const char* supp);
+bool SuppressionMatch(char *templ, const char *str);
+void SuppressionFree(Suppression *supp);
+
+}  // namespace __tsan
+
+#endif  // TSAN_SUPPRESSIONS_H
diff --git a/lib/tsan/rtl/tsan_symbolize.h b/lib/tsan/rtl/tsan_symbolize.h
new file mode 100644
index 000000000..4eaec4a68
--- /dev/null
+++ b/lib/tsan/rtl/tsan_symbolize.h
@@ -0,0 +1,26 @@
+//===-- tsan_symbolize.h ----------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_SYMBOLIZE_H
+#define TSAN_SYMBOLIZE_H
+
+#include "tsan_defs.h"
+#include "tsan_report.h"
+
+namespace __tsan {
+
+ReportStack *SymbolizeCode(uptr addr);
+ReportStack *SymbolizeData(uptr addr);
+
+}  // namespace __tsan
+
+#endif  // TSAN_SYMBOLIZE_H
diff --git a/lib/tsan/rtl/tsan_symbolize_addr2line_linux.cc b/lib/tsan/rtl/tsan_symbolize_addr2line_linux.cc
new file mode 100644
index 000000000..55b7c5d7c
--- /dev/null
+++ b/lib/tsan/rtl/tsan_symbolize_addr2line_linux.cc
@@ -0,0 +1,178 @@
+//===-- tsan_symbolize_addr2line.cc -----------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#include "tsan_symbolize.h"
+#include "tsan_mman.h"
+#include "tsan_rtl.h"
+
+#include <unistd.h>
+#include <dlfcn.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <errno.h>
+#include <link.h>
+#include <linux/limits.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+
+namespace __tsan {
+
+static bool GetSymbolizerFd(int *infdp, int *outfdp) {
+  static int outfd[2];
+  static int infd[2];
+  static int pid = -1;
+  static int inited = 0;
+  if (inited == 0) {
+    inited = -1;
+    if (pipe(outfd)) {
+      Printf("ThreadSanitizer: pipe() failed (%d)\n", errno);
+      Die();
+    }
+    if (pipe(infd)) {
+      Printf("ThreadSanitizer: pipe() failed (%d)\n", errno);
+      Die();
+    }
+    pid = fork();
+    if (pid == 0) {
+      close(STDOUT_FILENO);
+      close(STDIN_FILENO);
+      dup2(outfd[0], STDIN_FILENO);
+      dup2(infd[1], STDOUT_FILENO);
+      close(outfd[0]);
+      close(outfd[1]);
+      close(infd[0]);
+      close(infd[1]);
+      InternalScopedBuf<char> exe(PATH_MAX);
+      ssize_t len = readlink("/proc/self/exe", exe, exe.Size() - 1);
+      exe.Ptr()[len] = 0;
+      execl("/usr/bin/addr2line", "/usr/bin/addr2line", "-Cfe", exe.Ptr(),
+          NULL);
+      _exit(0);
+    } else if (pid < 0) {
+      Printf("ThreadSanitizer: failed to fork symbolizer\n");
+      Die();
+    }
+    close(outfd[0]);
+    close(infd[1]);
+    inited = 1;
+  } else if (inited > 0) {
+    int status = 0;
+    if (pid == waitpid(pid, &status, WNOHANG)) {
+      Printf("ThreadSanitizer: symbolizer died with status %d\n",
+          WEXITSTATUS(status));
+      Die();
+    }
+  }
+  *infdp = infd[0];
+  *outfdp = outfd[1];
+  return inited > 0;
+}
+
+static int dl_iterate_phdr_cb(dl_phdr_info *info, size_t size, void *ctx) {
+  *(uptr*)ctx = (uptr)info->dlpi_addr;
+  return 1;
+}
+
+static uptr GetImageBase() {
+  static uptr base = 0;
+  if (base == 0)
+    dl_iterate_phdr(dl_iterate_phdr_cb, &base);
+  return base;
+}
+
+ReportStack *SymbolizeCode(uptr addr) {
+  uptr base = GetImageBase();
+  uptr offset = addr - base;
+  int infd = -1;
+  int outfd = -1;
+  if (!GetSymbolizerFd(&infd, &outfd))
+    return 0;
+  char addrstr[32];
+  Snprintf(addrstr, sizeof(addrstr), "%p\n", (void*)offset);
+  if (0 >= write(outfd, addrstr, internal_strlen(addrstr))) {
+    Printf("ThreadSanitizer: can't write from symbolizer\n");
+    Die();
+  }
+  InternalScopedBuf<char> func(1024);
+  ssize_t len = read(infd, func, func.Size() - 1);
+  if (len <= 0) {
+    Printf("ThreadSanitizer: can't read from symbolizer\n");
+    Die();
+  }
+  func.Ptr()[len] = 0;
+  ReportStack *res = (ReportStack*)internal_alloc(MBlockReportStack,
+                                                  sizeof(ReportStack));
+  internal_memset(res, 0, sizeof(*res));
+  res->module = (char*)internal_alloc(MBlockReportStack, 4);
+  internal_memcpy(res->module, "exe", 4);
+  res->offset = offset;
+  res->pc = addr;
+
+  char *pos = strchr(func, '\n');
+  if (pos && func[0] != '?') {
+    res->func = (char*)internal_alloc(MBlockReportStack, pos - func + 1);
+    internal_memcpy(res->func, func, pos - func);
+    res->func[pos - func] = 0;
+    char *pos2 = strchr(pos, ':');
+    if (pos2) {
+      res->file = (char*)internal_alloc(MBlockReportStack, pos2 - pos - 1 + 1);
+      internal_memcpy(res->file, pos + 1, pos2 - pos - 1);
+      res->file[pos2 - pos - 1] = 0;
+      res->line = atoi(pos2 + 1);
+     }
+  }
+  return res;
+}
+
+ReportStack *SymbolizeData(uptr addr) {
+  return 0;
+  /*
+  if (base == 0)
+    base = GetImageBase();
+  int res = 0;
+  InternalScopedBuf<char> cmd(1024);
+  Snprintf(cmd, cmd.Size(),
+  "nm -alC %s|grep \"%lx\"|awk '{printf(\"%%s\\n%%s\", $3, $4)}' > tsan.tmp2",
+    exe, (addr - base));
+  if (system(cmd))
+    return 0;
+  FILE* f3 = fopen("tsan.tmp2", "rb");
+  if (f3) {
+    InternalScopedBuf<char> tmp(1024);
+    if (fread(tmp, 1, tmp.Size(), f3) <= 0)
+      return 0;
+    char *pos = strchr(tmp, '\n');
+    if (pos && tmp[0] != '?') {
+      res = 1;
+      symb[0].module = 0;
+      symb[0].offset = addr;
+      symb[0].name = alloc->Alloc<char>(pos - tmp + 1);
+      internal_memcpy(symb[0].name, tmp, pos - tmp);
+      symb[0].name[pos - tmp] = 0;
+      symb[0].file = 0;
+      symb[0].line = 0;
+      char *pos2 = strchr(pos, ':');
+      if (pos2) {
+        symb[0].file = alloc->Alloc<char>(pos2 - pos - 1 + 1);
+        internal_memcpy(symb[0].file, pos + 1, pos2 - pos - 1);
+        symb[0].file[pos2 - pos - 1] = 0;
+        symb[0].line = atoi(pos2 + 1);
+      }
+    }
+    fclose(f3);
+  }
+  return res;
+  */
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_sync.cc b/lib/tsan/rtl/tsan_sync.cc
new file mode 100644
index 000000000..0b31ab9ce
--- /dev/null
+++ b/lib/tsan/rtl/tsan_sync.cc
@@ -0,0 +1,177 @@
+//===-- tsan_sync.cc --------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#include "tsan_sync.h"
+#include "tsan_placement_new.h"
+#include "tsan_rtl.h"
+#include "tsan_mman.h"
+
+namespace __tsan {
+
+SyncVar::SyncVar(uptr addr)
+  : mtx(MutexTypeSyncVar, StatMtxSyncVar)
+  , addr(addr)
+  , owner_tid(kInvalidTid)
+  , recursion()
+  , is_rw()
+  , is_recursive()
+  , is_broken() {
+}
+
+SyncTab::Part::Part()
+  : mtx(MutexTypeSyncTab, StatMtxSyncTab)
+  , val() {
+}
+
+SyncTab::SyncTab() {
+}
+
+SyncTab::~SyncTab() {
+  for (int i = 0; i < kPartCount; i++) {
+    while (tab_[i].val) {
+      SyncVar *tmp = tab_[i].val;
+      tab_[i].val = tmp->next;
+      DestroyAndFree(tmp);
+    }
+  }
+}
+
+SyncVar* SyncTab::GetAndLock(ThreadState *thr, uptr pc,
+                             uptr addr, bool write_lock) {
+  Part *p = &tab_[PartIdx(addr)];
+  {
+    ReadLock l(&p->mtx);
+    for (SyncVar *res = p->val; res; res = res->next) {
+      if (res->addr == addr) {
+        if (write_lock)
+          res->mtx.Lock();
+        else
+          res->mtx.ReadLock();
+        return res;
+      }
+    }
+  }
+  {
+    Lock l(&p->mtx);
+    SyncVar *res = p->val;
+    for (; res; res = res->next) {
+      if (res->addr == addr)
+        break;
+    }
+    if (res == 0) {
+      StatInc(thr, StatSyncCreated);
+      void *mem = internal_alloc(MBlockSync, sizeof(SyncVar));
+      res = new(mem) SyncVar(addr);
+      res->creation_stack.ObtainCurrent(thr, pc);
+      res->next = p->val;
+      p->val = res;
+    }
+    if (write_lock)
+      res->mtx.Lock();
+    else
+      res->mtx.ReadLock();
+    return res;
+  }
+}
+
+SyncVar* SyncTab::GetAndRemove(ThreadState *thr, uptr pc, uptr addr) {
+  Part *p = &tab_[PartIdx(addr)];
+  SyncVar *res = 0;
+  {
+    Lock l(&p->mtx);
+    SyncVar **prev = &p->val;
+    res = *prev;
+    while (res) {
+      if (res->addr == addr) {
+        *prev = res->next;
+        break;
+      }
+      prev = &res->next;
+      res = *prev;
+    }
+  }
+  if (res) {
+    StatInc(thr, StatSyncDestroyed);
+    res->mtx.Lock();
+    res->mtx.Unlock();
+  }
+  return res;
+}
+
+int SyncTab::PartIdx(uptr addr) {
+  return (addr >> 3) % kPartCount;
+}
+
+StackTrace::StackTrace()
+    : n_()
+    , s_() {
+}
+
+StackTrace::~StackTrace() {
+  Reset();
+}
+
+void StackTrace::Reset() {
+  if (s_) {
+    CHECK_NE(n_, 0);
+    internal_free(s_);
+    s_ = 0;
+    n_ = 0;
+  }
+}
+
+void StackTrace::Init(const uptr *pcs, uptr cnt) {
+  Reset();
+  if (cnt == 0)
+    return;
+  n_ = cnt;
+  s_ = (uptr*)internal_alloc(MBlockStackTrace, cnt * sizeof(s_[0]));
+  internal_memcpy(s_, pcs, cnt * sizeof(s_[0]));
+}
+
+void StackTrace::ObtainCurrent(ThreadState *thr, uptr toppc) {
+  Reset();
+  n_ = thr->shadow_stack_pos - &thr->shadow_stack[0];
+  if (n_ + !!toppc == 0)
+    return;
+  s_ = (uptr*)internal_alloc(MBlockStackTrace, (n_ + !!toppc) * sizeof(s_[0]));
+  for (uptr i = 0; i < n_; i++)
+    s_[i] = thr->shadow_stack[i];
+  if (toppc) {
+    s_[n_] = toppc;
+    n_++;
+  }
+}
+
+void StackTrace::CopyFrom(const StackTrace& other) {
+  Reset();
+  Init(other.Begin(), other.Size());
+}
+
+bool StackTrace::IsEmpty() const {
+  return n_ == 0;
+}
+
+uptr StackTrace::Size() const {
+  return n_;
+}
+
+uptr StackTrace::Get(uptr i) const {
+  CHECK_LT(i, n_);
+  return s_[i];
+}
+
+const uptr *StackTrace::Begin() const {
+  return s_;
+}
+
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_sync.h b/lib/tsan/rtl/tsan_sync.h
new file mode 100644
index 000000000..45251f02e
--- /dev/null
+++ b/lib/tsan/rtl/tsan_sync.h
@@ -0,0 +1,97 @@
+//===-- tsan_sync.h ---------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_SYNC_H
+#define TSAN_SYNC_H
+
+#include "tsan_atomic.h"
+#include "tsan_clock.h"
+#include "tsan_defs.h"
+#include "tsan_mutex.h"
+
+namespace __tsan {
+
+class SlabCache;
+
+class StackTrace {
+ public:
+  StackTrace();
+  ~StackTrace();
+  void Reset();
+
+  void Init(const uptr *pcs, uptr cnt);
+  void ObtainCurrent(ThreadState *thr, uptr toppc);
+  bool IsEmpty() const;
+  uptr Size() const;
+  uptr Get(uptr i) const;
+  const uptr *Begin() const;
+  void CopyFrom(const StackTrace& other);
+
+ private:
+  uptr n_;
+  uptr *s_;
+
+  StackTrace(const StackTrace&);
+  void operator = (const StackTrace&);
+};
+
+struct SyncVar {
+  explicit SyncVar(uptr addr);
+
+  static const int kInvalidTid = -1;
+
+  Mutex mtx;
+  const uptr addr;
+  SyncClock clock;
+  StackTrace creation_stack;
+  SyncClock read_clock;  // Used for rw mutexes only.
+  int owner_tid;  // Set only by exclusive owners.
+  int recursion;
+  bool is_rw;
+  bool is_recursive;
+  bool is_broken;
+  SyncVar *next;  // In SyncTab hashtable.
+};
+
+class SyncTab {
+ public:
+  SyncTab();
+  ~SyncTab();
+
+  // If the SyncVar does not exist yet, it is created.
+  SyncVar* GetAndLock(ThreadState *thr, uptr pc,
+                      uptr addr, bool write_lock);
+
+  // If the SyncVar does not exist, returns 0.
+  SyncVar* GetAndRemove(ThreadState *thr, uptr pc, uptr addr);
+
+ private:
+  struct Part {
+    Mutex mtx;
+    SyncVar *val;
+    char pad[kCacheLineSize - sizeof(Mutex) - sizeof(SyncVar*)];  // NOLINT
+    Part();
+  };
+
+  // FIXME: Implement something more sane.
+  static const int kPartCount = 1009;
+  Part tab_[kPartCount];
+
+  int PartIdx(uptr addr);
+
+  SyncTab(const SyncTab&);  // Not implemented.
+  void operator = (const SyncTab&);  // Not implemented.
+};
+
+}  // namespace __tsan
+
+#endif  // TSAN_SYNC_H
diff --git a/lib/tsan/rtl/tsan_trace.h b/lib/tsan/rtl/tsan_trace.h
new file mode 100644
index 000000000..4a1930d20
--- /dev/null
+++ b/lib/tsan/rtl/tsan_trace.h
@@ -0,0 +1,59 @@
+//===-- tsan_trace.h -------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+#ifndef TSAN_TRACE_H
+#define TSAN_TRACE_H
+
+#include "tsan_defs.h"
+#include "tsan_mutex.h"
+#include "tsan_sync.h"
+
+namespace __tsan {
+
+const int kTraceParts = 8;
+const int kTraceSize = 1024*1024;
+const int kTracePartSize = kTraceSize / kTraceParts;
+
+// Must fit into 3 bits.
+enum EventType {
+  EventTypeMop,
+  EventTypeFuncEnter,
+  EventTypeFuncExit,
+  EventTypeLock,
+  EventTypeUnlock,
+  EventTypeRLock,
+  EventTypeRUnlock,
+};
+
+// Represents a thread event (from most significant bit):
+// u64 typ  : 3;   // EventType.
+// u64 addr : 61;  // Associated pc.
+typedef u64 Event;
+
+struct TraceHeader {
+  StackTrace stack0;  // Start stack for the trace.
+  u64   epoch0;       // Start epoch for the trace.
+};
+
+struct Trace {
+  Event events[kTraceSize];
+  TraceHeader headers[kTraceParts];
+  Mutex mtx;
+
+  Trace()
+    : mtx(MutexTypeTrace, StatMtxTrace) {
+  }
+};
+
+}  // namespace __tsan
+
+#endif  // TSAN_TRACE_H
diff --git a/lib/tsan/rtl/tsan_update_shadow_word_inl.h b/lib/tsan/rtl/tsan_update_shadow_word_inl.h
new file mode 100644
index 000000000..c7864ce00
--- /dev/null
+++ b/lib/tsan/rtl/tsan_update_shadow_word_inl.h
@@ -0,0 +1,79 @@
+//===-- tsan_update_shadow_word_inl.h ---------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+// Body of the hottest inner loop.
+// If we wrap this body into a function, compilers (both gcc and clang)
+// produce sligtly less efficient code.
+//===----------------------------------------------------------------------===//
+do {
+  StatInc(thr, StatShadowProcessed);
+  const unsigned kAccessSize = 1 << kAccessSizeLog;
+  unsigned off = cur.ComputeSearchOffset();
+  u64 *sp = &shadow_mem[(idx + off) % kShadowCnt];
+  old = LoadShadow(sp);
+  if (old.IsZero()) {
+    StatInc(thr, StatShadowZero);
+    if (store_word)
+      StoreIfNotYetStored(sp, &store_word);
+    // The above StoreIfNotYetStored could be done unconditionally
+    // and it even shows 4% gain on synthetic benchmarks (r4307).
+    break;
+  }
+  // is the memory access equal to the previous?
+  if (Shadow::Addr0AndSizeAreEqual(cur, old)) {
+    StatInc(thr, StatShadowSameSize);
+    // same thread?
+    if (Shadow::TidsAreEqual(old, cur)) {
+      StatInc(thr, StatShadowSameThread);
+      if (OldIsInSameSynchEpoch(old, thr)) {
+        if (OldIsRWStronger(old, kAccessIsWrite)) {
+          // found a slot that holds effectively the same info
+          // (that is, same tid, same sync epoch and same size)
+          StatInc(thr, StatMopSame);
+          return;
+        }
+        StoreIfNotYetStored(sp, &store_word);
+        break;
+      }
+      if (OldIsRWWeaker(old, kAccessIsWrite))
+        StoreIfNotYetStored(sp, &store_word);
+      break;
+    }
+    StatInc(thr, StatShadowAnotherThread);
+    if (HappensBefore(old, thr)) {
+      StoreIfNotYetStored(sp, &store_word);
+      break;
+    }
+    if (BothReads(old, kAccessIsWrite))
+      break;
+    goto RACE;
+  }
+
+  // Do the memory access intersect?
+  if (Shadow::TwoRangesIntersect(old, cur, kAccessSize)) {
+    StatInc(thr, StatShadowIntersect);
+    if (Shadow::TidsAreEqual(old, cur)) {
+      StatInc(thr, StatShadowSameThread);
+      break;
+    }
+    StatInc(thr, StatShadowAnotherThread);
+    if (HappensBefore(old, thr))
+      break;
+
+    if (BothReads(old, kAccessIsWrite))
+      break;
+
+    goto RACE;
+  }
+  // The accesses do not intersect.
+  StatInc(thr, StatShadowNotIntersect);
+  break;
+} while (0);
diff --git a/lib/tsan/rtl/tsan_vector.h b/lib/tsan/rtl/tsan_vector.h
new file mode 100644
index 000000000..d41063df3
--- /dev/null
+++ b/lib/tsan/rtl/tsan_vector.h
@@ -0,0 +1,110 @@
+//===-- tsan_vector.h -------------------------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+// Low-fat STL-like vector container.
+
+#ifndef TSAN_VECTOR_H
+#define TSAN_VECTOR_H
+
+#include "tsan_defs.h"
+#include "tsan_mman.h"
+
+namespace __tsan {
+
+template<typename T>
+class Vector {
+ public:
+  explicit Vector(MBlockType typ)
+      : typ_(typ)
+      , begin_()
+      , end_()
+      , last_() {
+  }
+
+  ~Vector() {
+    if (begin_)
+      internal_free(begin_);
+  }
+
+  void Reset() {
+    if (begin_)
+      internal_free(begin_);
+    begin_ = 0;
+    end_ = 0;
+    last_ = 0;
+  }
+
+  uptr Size() const {
+    return end_ - begin_;
+  }
+
+  T &operator[](uptr i) {
+    DCHECK_LT(i, end_ - begin_);
+    return begin_[i];
+  }
+
+  const T &operator[](uptr i) const {
+    DCHECK_LT(i, end_ - begin_);
+    return begin_[i];
+  }
+
+  T *PushBack(T v = T()) {
+    EnsureSize(Size() + 1);
+    end_[-1] = v;
+    return &end_[-1];
+  }
+
+  void Resize(uptr size) {
+    uptr old_size = Size();
+    EnsureSize(size);
+    if (old_size < size) {
+      for (uptr i = old_size; i < size; i++)
+        begin_[i] = T();
+    }
+  }
+
+ private:
+  const MBlockType typ_;
+  T *begin_;
+  T *end_;
+  T *last_;
+
+  void EnsureSize(uptr size) {
+    if (size <= Size())
+      return;
+    if (size <= (uptr)(last_ - begin_)) {
+      end_ = begin_ + size;
+      return;
+    }
+    uptr cap0 = last_ - begin_;
+    uptr cap = 2 * cap0;
+    if (cap == 0)
+      cap = 16;
+    if (cap < size)
+      cap = size;
+    T *p = (T*)internal_alloc(typ_, cap * sizeof(T));
+    if (cap0) {
+      internal_memcpy(p, begin_, cap0 * sizeof(T));
+      internal_free(begin_);
+    }
+    begin_ = p;
+    end_ = begin_ + size;
+    last_ = begin_ + cap;
+  }
+
+  Vector(const Vector&);
+  void operator=(const Vector&);
+};
+}
+
+#endif  // #ifndef TSAN_VECTOR_H