drd/tests/tsan_unittest.cpp

/*
  This file is part of Valgrind, a dynamic binary instrumentation
  framework.

  Copyright (C) 2008-2008 Google Inc
     opensource@google.com

  This program is free software; you can redistribute it and/or
  modify it under the terms of the GNU General Public License as
  published by the Free Software Foundation; either version 2 of the
  License, or (at your option) any later version.

  This program is distributed in the hope that it will be useful, but
  WITHOUT ANY WARRANTY; without even the implied warranty of
  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  General Public License for more details.

  You should have received a copy of the GNU General Public License
  along with this program; if not, write to the Free Software
  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
  02111-1307, USA.

  The GNU General Public License is contained in the file COPYING.
*/

// Author: Konstantin Serebryany <opensource@google.com>
//
// This file contains a set of unit tests for a data race detection tool.
//
//
//
// This test can be compiled with pthreads (default) or
// with any other library that supports threads, locks, cond vars, etc.
//
// To compile with pthreads:
//   g++  racecheck_unittest.cc dynamic_annotations.cc
//        -lpthread -g -DDYNAMIC_ANNOTATIONS=1
//
// To compile with different library:
//   1. cp thread_wrappers_pthread.h thread_wrappers_yourlib.h
//   2. edit thread_wrappers_yourlib.h
//   3. add '-DTHREAD_WRAPPERS="thread_wrappers_yourlib.h"' to your compilation.
//
//

// This test must not include any other file specific to threading library,
// everything should be inside THREAD_WRAPPERS.
#ifndef THREAD_WRAPPERS
# define THREAD_WRAPPERS "thread_wrappers_pthread.h"
#endif
#include THREAD_WRAPPERS

#ifndef NEEDS_SEPERATE_RW_LOCK
#define RWLock Mutex // Mutex does work as an rw-lock.
#define WriterLockScoped MutexLock
#define ReaderLockScoped ReaderMutexLock
#endif // !NEEDS_SEPERATE_RW_LOCK


// Helgrind memory usage testing stuff
// If not present in dynamic_annotations.h/.cc - ignore
#ifndef ANNOTATE_RESET_STATS
#define ANNOTATE_RESET_STATS() do { } while(0)
#endif
#ifndef ANNOTATE_PRINT_STATS
#define ANNOTATE_PRINT_STATS() do { } while(0)
#endif
#ifndef ANNOTATE_PRINT_MEMORY_USAGE
#define ANNOTATE_PRINT_MEMORY_USAGE(a) do { } while(0)
#endif
//

#include <vector>
#include <string>
#include <map>
#include <queue>
#include <algorithm>
#include <cstring>      // strlen(), index(), rindex()
#include <ctime>
#include <sys/time.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <sys/mman.h>  // mmap
#include <errno.h>
#include <stdlib.h>
#include <dirent.h>

#ifndef _APPLE_
#include <malloc.h>
#endif

// The tests are
// - Stability tests (marked STAB)
// - Performance tests (marked PERF)
// - Feature tests
//   - TN (true negative) : no race exists and the tool is silent.
//   - TP (true positive) : a race exists and reported.
//   - FN (false negative): a race exists but not reported.
//   - FP (false positive): no race exists but the tool reports it.
//
// The feature tests are marked according to the behavior of helgrind 3.3.0.
//
// TP and FP tests are annotated with ANNOTATE_EXPECT_RACE,
// so, no error reports should be seen when running under helgrind.
//
// When some of the FP cases are fixed in helgrind we'll need
// to update this test.
//
// Each test resides in its own namespace.
// Namespaces are named test01, test02, ...
// Please, *DO NOT* change the logic of existing tests nor rename them.
// Create a new test instead.
//
// Some tests use sleep()/usleep().
// This is not a synchronization, but a simple way to trigger
// some specific behaviour of the race detector's scheduler.

// Globals and utilities used by several tests. {{{1
CondVar CV;
int     COND = 0;


typedef void (*void_func_void_t)(void);
enum TEST_FLAG {
  FEATURE           = 1 << 0,
  STABILITY         = 1 << 1,
  PERFORMANCE       = 1 << 2,
  EXCLUDE_FROM_ALL  = 1 << 3,
  NEEDS_ANNOTATIONS = 1 << 4,
  RACE_DEMO         = 1 << 5,
  MEMORY_USAGE      = 1 << 6,
  PRINT_STATS       = 1 << 7
};

// Put everything into stderr.
Mutex printf_mu;
#define printf(args...) \
    do{ \
      printf_mu.Lock();\
      fprintf(stderr, args);\
      printf_mu.Unlock(); \
    }while(0)

long GetTimeInMs() {
   struct timeval tv;
   gettimeofday(&tv, NULL);
   return (tv.tv_sec * 1000L) + (tv.tv_usec / 1000L);
}

struct Test{
  void_func_void_t f_;
  int flags_;
  Test(void_func_void_t f, int flags)
    : f_(f)
    , flags_(flags)
  {}
  Test() : f_(0), flags_(0) {}
  void Run() {
     ANNOTATE_RESET_STATS();
     if (flags_ & PERFORMANCE) {
        long start = GetTimeInMs();
        f_();
        long end = GetTimeInMs();
        printf ("Time: %4ldms\n", end-start);
     } else
        f_();
     if (flags_ & PRINT_STATS)
        ANNOTATE_PRINT_STATS();
     if (flags_ & MEMORY_USAGE)
        ANNOTATE_PRINT_MEMORY_USAGE(0);
  }
};
std::map<int, Test> TheMapOfTests;

#define NOINLINE __attribute__ ((noinline))
extern "C" void NOINLINE AnnotateSetVerbosity(const char *, int, int) {};


struct TestAdder {
  TestAdder(void_func_void_t f, int id, int flags = FEATURE) {
    // AnnotateSetVerbosity(__FILE__, __LINE__, 0);
    CHECK(TheMapOfTests.count(id) == 0);
    TheMapOfTests[id] = Test(f, flags);
  }
};

#define REGISTER_TEST(f, id)         TestAdder add_test_##id (f, id);
#define REGISTER_TEST2(f, id, flags) TestAdder add_test_##id (f, id, flags);

static bool ArgIsOne(int *arg) { return *arg == 1; };
static bool ArgIsZero(int *arg) { return *arg == 0; };
static bool ArgIsTrue(bool *arg) { return *arg == true; };

// Call ANNOTATE_EXPECT_RACE only if 'machine' env variable is defined.
// Useful to test against several different machines.
// Supported machines so far:
//   MSM_HYBRID1             -- aka MSMProp1
//   MSM_HYBRID1_INIT_STATE  -- aka MSMProp1 with --initialization-state=yes
//   MSM_THREAD_SANITIZER    -- ThreadSanitizer's state machine
#define ANNOTATE_EXPECT_RACE_FOR_MACHINE(mem, descr, machine) \
    while(getenv(machine)) {\
      ANNOTATE_EXPECT_RACE(mem, descr); \
      break;\
    }\

#define ANNOTATE_EXPECT_RACE_FOR_TSAN(mem, descr) \
    ANNOTATE_EXPECT_RACE_FOR_MACHINE(mem, descr, "MSM_THREAD_SANITIZER")

inline bool Tsan_PureHappensBefore() {
  return true;
}

inline bool Tsan_FastMode()           {
  return getenv("TSAN_FAST_MODE") != NULL;
}

// Initialize *(mem) to 0 if Tsan_FastMode.
#define FAST_MODE_INIT(mem) do { if (Tsan_FastMode()) { *(mem) = 0; } } while(0)

#ifndef MAIN_INIT_ACTION
#define MAIN_INIT_ACTION
#endif


int main(int argc, char** argv) { // {{{1
  MAIN_INIT_ACTION;
  printf("FLAGS [phb=%i, fm=%i]\n", Tsan_PureHappensBefore(), Tsan_FastMode());
  if (argc == 2 && !strcmp(argv[1], "benchmark")) {
     for (std::map<int,Test>::iterator it = TheMapOfTests.begin();
         it != TheMapOfTests.end(); ++it) {
       if(!(it->second.flags_ & PERFORMANCE)) continue;
       it->second.Run();
     }
  } else if (argc == 2 && !strcmp(argv[1], "demo")) {
     for (std::map<int,Test>::iterator it = TheMapOfTests.begin();
         it != TheMapOfTests.end();  ++it) {
       if(!(it->second.flags_ & RACE_DEMO)) continue;
       it->second.Run();
     }
  } else if (argc > 1) {
    // the tests are listed in command line flags
    for (int i = 1; i < argc; i++) {
      int f_num = atoi(argv[i]);
      CHECK(TheMapOfTests.count(f_num));
      TheMapOfTests[f_num].Run();
    }
  } else {
    bool run_tests_with_annotations = false;
    if (getenv("DRT_ALLOW_ANNOTATIONS")) {
      run_tests_with_annotations = true;
    }
    for (std::map<int,Test>::iterator it = TheMapOfTests.begin();
        it != TheMapOfTests.end();
        ++it) {
      if(it->second.flags_ & EXCLUDE_FROM_ALL) continue;
      if(it->second.flags_ & RACE_DEMO) continue;
      if((it->second.flags_ & NEEDS_ANNOTATIONS)
         && run_tests_with_annotations == false) continue;
      it->second.Run();
    }
  }
}

#ifdef THREAD_WRAPPERS_PTHREAD_H
#endif


// An array of threads. Create/start/join all elements at once. {{{1
class MyThreadArray {
 public:
  static const int kSize = 5;
  typedef void (*F) (void);
  MyThreadArray(F f1, F f2 = NULL, F f3 = NULL, F f4 = NULL, F f5 = NULL) {
    ar_[0] = new MyThread(f1);
    ar_[1] = f2 ? new MyThread(f2) : NULL;
    ar_[2] = f3 ? new MyThread(f3) : NULL;
    ar_[3] = f4 ? new MyThread(f4) : NULL;
    ar_[4] = f5 ? new MyThread(f5) : NULL;
  }
  void Start() {
    for(int i = 0; i < kSize; i++) {
      if(ar_[i]) {
        ar_[i]->Start();
        usleep(10);
      }
    }
  }

  void Join() {
    for(int i = 0; i < kSize; i++) {
      if(ar_[i]) {
        ar_[i]->Join();
      }
    }
  }

  ~MyThreadArray() {
    for(int i = 0; i < kSize; i++) {
      delete ar_[i];
    }
  }
 private:
  MyThread *ar_[kSize];
};


// test00: {{{1
namespace test00 {
int     GLOB = 0;
void Run() {
  printf("test00: negative\n");
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 00)
}  // namespace test00


// test01: TP. Simple race (write vs write). {{{1
namespace test01 {
int     GLOB = 0;
void Worker() {
  GLOB = 1;
}

void Parent() {
  MyThread t(Worker);
  t.Start();
  GLOB = 2;
  t.Join();
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test01. TP.");
  ANNOTATE_TRACE_MEMORY(&GLOB);
  printf("test01: positive\n");
  Parent();
  const int tmp = GLOB;
  printf("\tGLOB=%d\n", tmp);
}
REGISTER_TEST(Run, 1);
}  // namespace test01


// test02: TN. Synchronization via CondVar. {{{1
namespace test02 {
int     GLOB = 0;
// Two write accesses to GLOB are synchronized because
// the pair of CV.Signal() and CV.Wait() establish happens-before relation.
//
// Waiter:                      Waker:
// 1. COND = 0
// 2. Start(Waker)
// 3. MU.Lock()                 a. write(GLOB)
//                              b. MU.Lock()
//                              c. COND = 1
//                         /--- d. CV.Signal()
//  4. while(COND)        /     e. MU.Unlock()
//       CV.Wait(MU) <---/
//  5. MU.Unlock()
//  6. write(GLOB)
Mutex   MU;

void Waker() {
  usleep(100000);  // Make sure the waiter blocks.
  GLOB = 1;

  MU.Lock();
  COND = 1;
  CV.Signal();
  MU.Unlock();
}

void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));
  MU.Lock();
  while(COND != 1)
    CV.Wait(&MU);
  MU.Unlock();
  GLOB = 2;
}
void Run() {
  printf("test02: negative\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 2);
}  // namespace test02


// test03: TN. Synchronization via LockWhen, signaller gets there first. {{{1
namespace test03 {
int     GLOB = 0;
// Two write accesses to GLOB are synchronized via conditional critical section.
// Note that LockWhen() happens first (we use sleep(1) to make sure)!
//
// Waiter:                           Waker:
// 1. COND = 0
// 2. Start(Waker)
//                                   a. write(GLOB)
//                                   b. MU.Lock()
//                                   c. COND = 1
//                              /--- d. MU.Unlock()
// 3. MU.LockWhen(COND==1) <---/
// 4. MU.Unlock()
// 5. write(GLOB)
Mutex   MU;

void Waker() {
  usleep(100000);  // Make sure the waiter blocks.
  GLOB = 1;

  MU.Lock();
  COND = 1; // We are done! Tell the Waiter.
  MU.Unlock(); // calls ANNOTATE_CONDVAR_SIGNAL;
}
void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));
  MU.LockWhen(Condition(&ArgIsOne, &COND));  // calls ANNOTATE_CONDVAR_WAIT
  MU.Unlock();  // Waker is done!

  GLOB = 2;
}
void Run() {
  printf("test03: negative\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 3, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test03

// test04: TN. Synchronization via PCQ. {{{1
namespace test04 {
int     GLOB = 0;
ProducerConsumerQueue Q(INT_MAX);
// Two write accesses to GLOB are separated by PCQ Put/Get.
//
// Putter:                        Getter:
// 1. write(GLOB)
// 2. Q.Put() ---------\          .
//                      \-------> a. Q.Get()
//                                b. write(GLOB)


void Putter() {
  GLOB = 1;
  Q.Put(NULL);
}

void Getter() {
  Q.Get();
  GLOB = 2;
}

void Run() {
  printf("test04: negative\n");
  MyThreadArray t(Putter, Getter);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 4);
}  // namespace test04


// test05: FP. Synchronization via CondVar, but waiter does not block. {{{1
// Since CondVar::Wait() is not called, we get a false positive.
namespace test05 {
int     GLOB = 0;
// Two write accesses to GLOB are synchronized via CondVar.
// But race detector can not see it.
// See this for details:
// http://www.valgrind.org/docs/manual/hg-manual.html#hg-manual.effective-use.
//
// Waiter:                                  Waker:
// 1. COND = 0
// 2. Start(Waker)
// 3. MU.Lock()                             a. write(GLOB)
//                                          b. MU.Lock()
//                                          c. COND = 1
//                                          d. CV.Signal()
//  4. while(COND)                          e. MU.Unlock()
//       CV.Wait(MU) <<< not called
//  5. MU.Unlock()
//  6. write(GLOB)
Mutex   MU;

void Waker() {
  GLOB = 1;
  MU.Lock();
  COND = 1;
  CV.Signal();
  MU.Unlock();
}

void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));
  usleep(100000);  // Make sure the signaller gets first.
  MU.Lock();
  while(COND != 1)
    CV.Wait(&MU);
  MU.Unlock();
  GLOB = 2;
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  if (!Tsan_PureHappensBefore())
    ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test05. FP. Unavoidable in hybrid scheme.");
  printf("test05: unavoidable false positive\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 5);
}  // namespace test05


// test06: TN. Synchronization via CondVar, but Waker gets there first.  {{{1
namespace test06 {
int     GLOB = 0;
// Same as test05 but we annotated the Wait() loop.
//
// Waiter:                                            Waker:
// 1. COND = 0
// 2. Start(Waker)
// 3. MU.Lock()                                       a. write(GLOB)
//                                                    b. MU.Lock()
//                                                    c. COND = 1
//                                           /------- d. CV.Signal()
//  4. while(COND)                          /         e. MU.Unlock()
//       CV.Wait(MU) <<< not called        /
//  6. ANNOTATE_CONDVAR_WAIT(CV, MU) <----/
//  5. MU.Unlock()
//  6. write(GLOB)

Mutex   MU;

void Waker() {
  GLOB = 1;
  MU.Lock();
  COND = 1;
  CV.Signal();
  MU.Unlock();
}

void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));
  usleep(100000);  // Make sure the signaller gets first.
  MU.Lock();
  while(COND != 1)
    CV.Wait(&MU);
  ANNOTATE_CONDVAR_LOCK_WAIT(&CV, &MU);

  MU.Unlock();
  GLOB = 2;
}
void Run() {
  printf("test06: negative\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 6, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test06


// test07: TN. Synchronization via LockWhen(), Signaller is observed first. {{{1
namespace test07 {
int     GLOB = 0;
bool    COND = 0;
// Two write accesses to GLOB are synchronized via conditional critical section.
// LockWhen() is observed after COND has been set (due to sleep).
// Unlock() calls ANNOTATE_CONDVAR_SIGNAL().
//
// Waiter:                           Signaller:
// 1. COND = 0
// 2. Start(Signaller)
//                                   a. write(GLOB)
//                                   b. MU.Lock()
//                                   c. COND = 1
//                              /--- d. MU.Unlock calls ANNOTATE_CONDVAR_SIGNAL
// 3. MU.LockWhen(COND==1) <---/
// 4. MU.Unlock()
// 5. write(GLOB)

Mutex   MU;
void Signaller() {
  GLOB = 1;
  MU.Lock();
  COND = true; // We are done! Tell the Waiter.
  MU.Unlock(); // calls ANNOTATE_CONDVAR_SIGNAL;
}
void Waiter() {
  COND = false;
  MyThread t(Signaller);
  t.Start();
  usleep(100000);  // Make sure the signaller gets there first.

  MU.LockWhen(Condition(&ArgIsTrue, &COND));  // calls ANNOTATE_CONDVAR_WAIT
  MU.Unlock();  // Signaller is done!

  GLOB = 2; // If LockWhen didn't catch the signal, a race may be reported here.
  t.Join();
}
void Run() {
  printf("test07: negative\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 7, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test07

// test08: TN. Synchronization via thread start/join. {{{1
namespace test08 {
int     GLOB = 0;
// Three accesses to GLOB are separated by thread start/join.
//
// Parent:                        Worker:
// 1. write(GLOB)
// 2. Start(Worker) ------------>
//                                a. write(GLOB)
// 3. Join(Worker) <------------
// 4. write(GLOB)
void Worker() {
  GLOB = 2;
}

void Parent() {
  MyThread t(Worker);
  GLOB = 1;
  t.Start();
  t.Join();
  GLOB = 3;
}
void Run() {
  printf("test08: negative\n");
  Parent();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 8);
}  // namespace test08


// test09: TP. Simple race (read vs write). {{{1
namespace test09 {
int     GLOB = 0;
// A simple data race between writer and reader.
// Write happens after read (enforced by sleep).
// Usually, easily detectable by a race detector.
void Writer() {
  usleep(100000);
  GLOB = 3;
}
void Reader() {
  CHECK(GLOB != -777);
}

void Run() {
  ANNOTATE_TRACE_MEMORY(&GLOB);
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test09. TP.");
  printf("test09: positive\n");
  MyThreadArray t(Writer, Reader);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 9);
}  // namespace test09


// test10: FN. Simple race (write vs read). {{{1
namespace test10 {
int     GLOB = 0;
// A simple data race between writer and reader.
// Write happens before Read (enforced by sleep),
// otherwise this test is the same as test09.
//
// Writer:                    Reader:
// 1. write(GLOB)             a. sleep(long enough so that GLOB
//                                is most likely initialized by Writer)
//                            b. read(GLOB)
//
//
// Eraser algorithm does not detect the race here,
// see Section 2.2 of http://citeseer.ist.psu.edu/savage97eraser.html.
//
void Writer() {
  GLOB = 3;
}
void Reader() {
  usleep(100000);
  CHECK(GLOB != -777);
}

void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test10. TP. FN in MSMHelgrind.");
  printf("test10: positive\n");
  MyThreadArray t(Writer, Reader);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 10);
}  // namespace test10


// test11: FP. Synchronization via CondVar, 2 workers. {{{1
// This test is properly synchronized, but currently (Dec 2007)
// helgrind reports a false positive.
//
// Parent:                              Worker1, Worker2:
// 1. Start(workers)                    a. read(GLOB)
// 2. MU.Lock()                         b. MU.Lock()
// 3. while(COND != 2)        /-------- c. CV.Signal()
//      CV.Wait(&MU) <-------/          d. MU.Unlock()
// 4. MU.Unlock()
// 5. write(GLOB)
//
namespace test11 {
int     GLOB = 0;
Mutex   MU;
void Worker() {
  usleep(200000);
  CHECK(GLOB != 777);

  MU.Lock();
  COND++;
  CV.Signal();
  MU.Unlock();
}

void Parent() {
  COND = 0;

  MyThreadArray t(Worker, Worker);
  t.Start();

  MU.Lock();
  while(COND != 2) {
    CV.Wait(&MU);
  }
  MU.Unlock();

  GLOB = 2;

  t.Join();
}

void Run() {
//  ANNOTATE_EXPECT_RACE(&GLOB, "test11. FP. Fixed by MSMProp1.");
  printf("test11: negative\n");
  Parent();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 11);
}  // namespace test11


// test12: FP. Synchronization via Mutex, then via PCQ. {{{1
namespace test12 {
int     GLOB = 0;
// This test is properly synchronized, but currently (Dec 2007)
// helgrind reports a false positive.
//
// First, we write to GLOB under MU, then we synchronize via PCQ,
// which is essentially a semaphore.
//
// Putter:                       Getter:
// 1. MU.Lock()                  a. MU.Lock()
// 2. write(GLOB) <---- MU ----> b. write(GLOB)
// 3. MU.Unlock()                c. MU.Unlock()
// 4. Q.Put()   ---------------> d. Q.Get()
//                               e. write(GLOB)

ProducerConsumerQueue Q(INT_MAX);
Mutex   MU;

void Putter() {
  MU.Lock();
  GLOB++;
  MU.Unlock();

  Q.Put(NULL);
}

void Getter() {
  MU.Lock();
  GLOB++;
  MU.Unlock();

  Q.Get();
  GLOB++;
}

void Run() {
//  ANNOTATE_EXPECT_RACE(&GLOB, "test12. FP. Fixed by MSMProp1.");
  printf("test12: negative\n");
  MyThreadArray t(Putter, Getter);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 12);
}  // namespace test12


// test13: FP. Synchronization via Mutex, then via LockWhen. {{{1
namespace test13 {
int     GLOB = 0;
// This test is essentially the same as test12, but uses LockWhen
// instead of PCQ.
//
// Waker:                                     Waiter:
// 1. MU.Lock()                               a. MU.Lock()
// 2. write(GLOB) <---------- MU ---------->  b. write(GLOB)
// 3. MU.Unlock()                             c. MU.Unlock()
// 4. MU.Lock()                               .
// 5. COND = 1                                .
// 6. ANNOTATE_CONDVAR_SIGNAL -------\        .
// 7. MU.Unlock()                     \       .
//                                     \----> d. MU.LockWhen(COND == 1)
//                                            e. MU.Unlock()
//                                            f. write(GLOB)
Mutex   MU;

void Waker() {
  MU.Lock();
  GLOB++;
  MU.Unlock();

  MU.Lock();
  COND = 1;
  ANNOTATE_CONDVAR_SIGNAL(&MU);
  MU.Unlock();
}

void Waiter() {
  MU.Lock();
  GLOB++;
  MU.Unlock();

  MU.LockWhen(Condition(&ArgIsOne, &COND));
  MU.Unlock();
  GLOB++;
}

void Run() {
//  ANNOTATE_EXPECT_RACE(&GLOB, "test13. FP. Fixed by MSMProp1.");
  printf("test13: negative\n");
  COND = 0;

  MyThreadArray t(Waker, Waiter);
  t.Start();
  t.Join();

  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 13, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test13


// test14: FP. Synchronization via PCQ, reads, 2 workers. {{{1
namespace test14 {
int     GLOB = 0;
// This test is properly synchronized, but currently (Dec 2007)
// helgrind reports a false positive.
//
// This test is similar to test11, but uses PCQ (semaphore).
//
// Putter2:                  Putter1:                     Getter:
// 1. read(GLOB)             a. read(GLOB)
// 2. Q2.Put() ----\         b. Q1.Put() -----\           .
//                  \                          \--------> A. Q1.Get()
//                   \----------------------------------> B. Q2.Get()
//                                                        C. write(GLOB)
ProducerConsumerQueue Q1(INT_MAX), Q2(INT_MAX);

void Putter1() {
  CHECK(GLOB != 777);
  Q1.Put(NULL);
}
void Putter2() {
  CHECK(GLOB != 777);
  Q2.Put(NULL);
}
void Getter() {
  Q1.Get();
  Q2.Get();
  GLOB++;
}
void Run() {
//  ANNOTATE_EXPECT_RACE(&GLOB, "test14. FP. Fixed by MSMProp1.");
  printf("test14: negative\n");
  MyThreadArray t(Getter, Putter1, Putter2);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 14);
}  // namespace test14


// test15: TN. Synchronization via LockWhen. One waker and 2 waiters. {{{1
namespace test15 {
// Waker:                                   Waiter1, Waiter2:
// 1. write(GLOB)
// 2. MU.Lock()
// 3. COND = 1
// 4. ANNOTATE_CONDVAR_SIGNAL ------------> a. MU.LockWhen(COND == 1)
// 5. MU.Unlock()                           b. MU.Unlock()
//                                          c. read(GLOB)

int     GLOB = 0;
Mutex   MU;

void Waker() {
  GLOB = 2;

  MU.Lock();
  COND = 1;
  ANNOTATE_CONDVAR_SIGNAL(&MU);
  MU.Unlock();
};

void Waiter() {
  MU.LockWhen(Condition(&ArgIsOne, &COND));
  MU.Unlock();
  CHECK(GLOB != 777);
}


void Run() {
  COND = 0;
  printf("test15: negative\n");
  MyThreadArray t(Waker, Waiter, Waiter);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 15);
}  // namespace test15


// test16: FP. Barrier (emulated by CV), 2 threads. {{{1
namespace test16 {
// Worker1:                                     Worker2:
// 1. MU.Lock()                                 a. MU.Lock()
// 2. write(GLOB) <------------ MU ---------->  b. write(GLOB)
// 3. MU.Unlock()                               c. MU.Unlock()
// 4. MU2.Lock()                                d. MU2.Lock()
// 5. COND--                                    e. COND--
// 6. ANNOTATE_CONDVAR_SIGNAL(MU2) ---->V       .
// 7. MU2.Await(COND == 0) <------------+------ f. ANNOTATE_CONDVAR_SIGNAL(MU2)
// 8. MU2.Unlock()                      V-----> g. MU2.Await(COND == 0)
// 9. read(GLOB)                                h. MU2.Unlock()
//                                              i. read(GLOB)
//
//
// TODO: This way we may create too many edges in happens-before graph.
// Arndt Mühlenfeld in his PhD (TODO: link) suggests creating special nodes in
// happens-before graph to reduce the total number of edges.
// See figure 3.14.
//
//
int     GLOB = 0;
Mutex   MU;
Mutex MU2;

void Worker() {
  MU.Lock();
  GLOB++;
  MU.Unlock();

  MU2.Lock();
  COND--;
  ANNOTATE_CONDVAR_SIGNAL(&MU2);
  MU2.Await(Condition(&ArgIsZero, &COND));
  MU2.Unlock();

  CHECK(GLOB == 2);
}

void Run() {
//  ANNOTATE_EXPECT_RACE(&GLOB, "test16. FP. Fixed by MSMProp1 + Barrier support.");
  COND = 2;
  printf("test16: negative\n");
  MyThreadArray t(Worker, Worker);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 16, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test16


// test17: FP. Barrier (emulated by CV), 3 threads. {{{1
namespace test17 {
// Same as test16, but with 3 threads.
int     GLOB = 0;
Mutex   MU;
Mutex MU2;

void Worker() {
  MU.Lock();
  GLOB++;
  MU.Unlock();

  MU2.Lock();
  COND--;
  ANNOTATE_CONDVAR_SIGNAL(&MU2);
  MU2.Await(Condition(&ArgIsZero, &COND));
  MU2.Unlock();

  CHECK(GLOB == 3);
}

void Run() {
//  ANNOTATE_EXPECT_RACE(&GLOB, "test17. FP. Fixed by MSMProp1 + Barrier support.");
  COND = 3;
  printf("test17: negative\n");
  MyThreadArray t(Worker, Worker, Worker);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 17, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test17


// test18: TN. Synchronization via Await(), signaller gets there first. {{{1
namespace test18 {
int     GLOB = 0;
Mutex   MU;
// Same as test03, but uses Mutex::Await() instead of Mutex::LockWhen().

void Waker() {
  usleep(100000);  // Make sure the waiter blocks.
  GLOB = 1;

  MU.Lock();
  COND = 1; // We are done! Tell the Waiter.
  MU.Unlock(); // calls ANNOTATE_CONDVAR_SIGNAL;
}
void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));

  MU.Lock();
  MU.Await(Condition(&ArgIsOne, &COND));  // calls ANNOTATE_CONDVAR_WAIT
  MU.Unlock();  // Waker is done!

  GLOB = 2;
}
void Run() {
  printf("test18: negative\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 18, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test18

// test19: TN. Synchronization via AwaitWithTimeout(). {{{1
namespace test19 {
int     GLOB = 0;
// Same as test18, but with AwaitWithTimeout. Do not timeout.
Mutex   MU;
void Waker() {
  usleep(100000);  // Make sure the waiter blocks.
  GLOB = 1;

  MU.Lock();
  COND = 1; // We are done! Tell the Waiter.
  MU.Unlock(); // calls ANNOTATE_CONDVAR_SIGNAL;
}
void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));

  MU.Lock();
  CHECK(MU.AwaitWithTimeout(Condition(&ArgIsOne, &COND), INT_MAX));
  MU.Unlock();

  GLOB = 2;
}
void Run() {
  printf("test19: negative\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 19, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test19

// test20: TP. Incorrect synchronization via AwaitWhen(), timeout. {{{1
namespace test20 {
int     GLOB = 0;
Mutex   MU;
// True race. We timeout in AwaitWhen.
void Waker() {
  GLOB = 1;
  usleep(100 * 1000);
}
void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));

  MU.Lock();
  CHECK(!MU.AwaitWithTimeout(Condition(&ArgIsOne, &COND), 100));
  MU.Unlock();

  GLOB = 2;
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test20. TP.");
  printf("test20: positive\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 20, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test20

// test21: TP. Incorrect synchronization via LockWhenWithTimeout(). {{{1
namespace test21 {
int     GLOB = 0;
// True race. We timeout in LockWhenWithTimeout().
Mutex   MU;
void Waker() {
  GLOB = 1;
  usleep(100 * 1000);
}
void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));

  CHECK(!MU.LockWhenWithTimeout(Condition(&ArgIsOne, &COND), 100));
  MU.Unlock();

  GLOB = 2;
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test21. TP.");
  printf("test21: positive\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 21, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test21

// test22: TP. Incorrect synchronization via CondVar::WaitWithTimeout(). {{{1
namespace test22 {
int     GLOB = 0;
Mutex   MU;
// True race. We timeout in CondVar::WaitWithTimeout().
void Waker() {
  GLOB = 1;
  usleep(100 * 1000);
}
void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));

  int64_t ms_left_to_wait = 100;
  int64_t deadline_ms = GetCurrentTimeMillis() + ms_left_to_wait;
  MU.Lock();
  while(COND != 1 && ms_left_to_wait > 0) {
    CV.WaitWithTimeout(&MU, ms_left_to_wait);
    ms_left_to_wait = deadline_ms - GetCurrentTimeMillis();
  }
  MU.Unlock();

  GLOB = 2;
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test22. TP.");
  printf("test22: positive\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 22);
}  // namespace test22

// test23: TN. TryLock, ReaderLock, ReaderTryLock. {{{1
namespace test23 {
// Correct synchronization with TryLock, Lock, ReaderTryLock, ReaderLock.
int     GLOB = 0;
Mutex   MU;
void Worker_TryLock() {
  for (int i = 0; i < 20; i++) {
    while (true) {
      if (MU.TryLock()) {
        GLOB++;
        MU.Unlock();
        break;
      }
      usleep(1000);
    }
  }
}

void Worker_ReaderTryLock() {
  for (int i = 0; i < 20; i++) {
    while (true) {
      if (MU.ReaderTryLock()) {
        CHECK(GLOB != 777);
        MU.ReaderUnlock();
        break;
      }
      usleep(1000);
    }
  }
}

void Worker_ReaderLock() {
  for (int i = 0; i < 20; i++) {
    MU.ReaderLock();
    CHECK(GLOB != 777);
    MU.ReaderUnlock();
    usleep(1000);
  }
}

void Worker_Lock() {
  for (int i = 0; i < 20; i++) {
    MU.Lock();
    GLOB++;
    MU.Unlock();
    usleep(1000);
  }
}

void Run() {
  printf("test23: negative\n");
  MyThreadArray t(Worker_TryLock,
                  Worker_ReaderTryLock,
                  Worker_ReaderLock,
                  Worker_Lock
                  );
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 23);
}  // namespace test23

// test24: TN. Synchronization via ReaderLockWhen(). {{{1
namespace test24 {
int     GLOB = 0;
Mutex   MU;
// Same as test03, but uses ReaderLockWhen().

void Waker() {
  usleep(100000);  // Make sure the waiter blocks.
  GLOB = 1;

  MU.Lock();
  COND = 1; // We are done! Tell the Waiter.
  MU.Unlock(); // calls ANNOTATE_CONDVAR_SIGNAL;
}
void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));
  MU.ReaderLockWhen(Condition(&ArgIsOne, &COND));
  MU.ReaderUnlock();

  GLOB = 2;
}
void Run() {
  printf("test24: negative\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 24, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test24

// test25: TN. Synchronization via ReaderLockWhenWithTimeout(). {{{1
namespace test25 {
int     GLOB = 0;
Mutex   MU;
// Same as test24, but uses ReaderLockWhenWithTimeout().
// We do not timeout.

void Waker() {
  usleep(100000);  // Make sure the waiter blocks.
  GLOB = 1;

  MU.Lock();
  COND = 1; // We are done! Tell the Waiter.
  MU.Unlock(); // calls ANNOTATE_CONDVAR_SIGNAL;
}
void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));
  CHECK(MU.ReaderLockWhenWithTimeout(Condition(&ArgIsOne, &COND), INT_MAX));
  MU.ReaderUnlock();

  GLOB = 2;
}
void Run() {
  printf("test25: negative\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 25, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test25

// test26: TP. Incorrect synchronization via ReaderLockWhenWithTimeout(). {{{1
namespace test26 {
int     GLOB = 0;
Mutex   MU;
// Same as test25, but we timeout and incorrectly assume happens-before.

void Waker() {
  GLOB = 1;
  usleep(10000);
}
void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));
  CHECK(!MU.ReaderLockWhenWithTimeout(Condition(&ArgIsOne, &COND), 100));
  MU.ReaderUnlock();

  GLOB = 2;
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test26. TP");
  printf("test26: positive\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 26, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test26


// test27: TN. Simple synchronization via SpinLock. {{{1
namespace test27 {
#ifndef NO_SPINLOCK
int     GLOB = 0;
SpinLock MU;
void Worker() {
  MU.Lock();
  GLOB++;
  MU.Unlock();
  usleep(10000);
}

void Run() {
  printf("test27: negative\n");
  MyThreadArray t(Worker, Worker, Worker, Worker);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 27, FEATURE|NEEDS_ANNOTATIONS);
#endif // NO_SPINLOCK
}  // namespace test27


// test28: TN. Synchronization via Mutex, then PCQ. 3 threads {{{1
namespace test28 {
// Putter1:                       Getter:                         Putter2:
// 1. MU.Lock()                                                   A. MU.Lock()
// 2. write(GLOB)                                                 B. write(GLOB)
// 3. MU.Unlock()                                                 C. MU.Unlock()
// 4. Q.Put() ---------\                                 /------- D. Q.Put()
// 5. MU.Lock()         \-------> a. Q.Get()            /         E. MU.Lock()
// 6. read(GLOB)                  b. Q.Get() <---------/          F. read(GLOB)
// 7. MU.Unlock()                   (sleep)                       G. MU.Unlock()
//                                c. read(GLOB)
ProducerConsumerQueue Q(INT_MAX);
int     GLOB = 0;
Mutex   MU;

void Putter() {
  MU.Lock();
  GLOB++;
  MU.Unlock();

  Q.Put(NULL);

  MU.Lock();
  CHECK(GLOB != 777);
  MU.Unlock();
}

void Getter() {
  Q.Get();
  Q.Get();
  usleep(100000);
  CHECK(GLOB == 2);
}

void Run() {
  printf("test28: negative\n");
  MyThreadArray t(Getter, Putter, Putter);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 28);
}  // namespace test28


// test29: TN. Synchronization via Mutex, then PCQ. 4 threads. {{{1
namespace test29 {
// Similar to test28, but has two Getters and two PCQs.
ProducerConsumerQueue *Q1, *Q2;
Mutex   MU;
int     GLOB = 0;

void Putter(ProducerConsumerQueue *q) {
  MU.Lock();
  GLOB++;
  MU.Unlock();

  q->Put(NULL);
  q->Put(NULL);

  MU.Lock();
  CHECK(GLOB != 777);
  MU.Unlock();

}

void Putter1() { Putter(Q1); }
void Putter2() { Putter(Q2); }

void Getter() {
  Q1->Get();
  Q2->Get();
  usleep(100000);
  CHECK(GLOB == 2);
  usleep(48000); //  TODO: remove this when FP in test32 is fixed.
}

void Run() {
  printf("test29: negative\n");
  Q1 = new ProducerConsumerQueue(INT_MAX);
  Q2 = new ProducerConsumerQueue(INT_MAX);
  MyThreadArray t(Getter, Getter, Putter1, Putter2);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
  delete Q1;
  delete Q2;
}
REGISTER_TEST(Run, 29);
}  // namespace test29


// test30: TN. Synchronization via 'safe' race. Writer vs multiple Readers. {{{1
namespace test30 {
// This test shows a very risky kind of synchronization which is very easy
// to get wrong. Actually, I am not sure I've got it right.
//
// Writer:                                 Reader1, Reader2, ..., ReaderN:
// 1. write(GLOB[i]: i >= BOUNDARY)        a. n = BOUNDARY
// 2. HAPPENS_BEFORE(BOUNDARY+1)  -------> b. HAPPENS_AFTER(n)
// 3. BOUNDARY++;                          c. read(GLOB[i]: i < n)
//
// Here we have a 'safe' race on accesses to BOUNDARY and
// no actual races on accesses to GLOB[]:
// Writer writes to GLOB[i] where i>=BOUNDARY and then increments BOUNDARY.
// Readers read BOUNDARY and read GLOB[i] where i<BOUNDARY.
//
// I am not completely sure that this scheme guaranties no race between
// accesses to GLOB since compilers and CPUs
// are free to rearrange memory operations.
// I am actually sure that this scheme is wrong unless we use
// some smart memory fencing...


const int N = 48;
static int GLOB[N];
volatile int BOUNDARY = 0;

void Writer() {
  for (int i = 0; i < N; i++) {
    CHECK(BOUNDARY == i);
    for (int j = i; j < N; j++) {
      GLOB[j] = j;
    }
    ANNOTATE_HAPPENS_BEFORE(reinterpret_cast<void*>(BOUNDARY+1));
    BOUNDARY++;
    usleep(1000);
  }
}

void Reader() {
  int n;
  do {
    n = BOUNDARY;
    if (n == 0) continue;
    ANNOTATE_HAPPENS_AFTER(reinterpret_cast<void*>(n));
    for (int i = 0; i < n; i++) {
      CHECK(GLOB[i] == i);
    }
    usleep(100);
  } while(n < N);
}

void Run() {
  FAST_MODE_INIT(&BOUNDARY);
  ANNOTATE_EXPECT_RACE((void*)(&BOUNDARY), "test30. Sync via 'safe' race.");
  printf("test30: negative\n");
  MyThreadArray t(Writer, Reader, Reader, Reader);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB[N-1]);
}
REGISTER_TEST2(Run, 30, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test30


// test31: TN. Synchronization via 'safe' race. Writer vs Writer. {{{1
namespace test31 {
// This test is similar to test30, but
// it has one Writer instead of mulitple Readers.
//
// Writer1:                                Writer2
// 1. write(GLOB[i]: i >= BOUNDARY)        a. n = BOUNDARY
// 2. HAPPENS_BEFORE(BOUNDARY+1)  -------> b. HAPPENS_AFTER(n)
// 3. BOUNDARY++;                          c. write(GLOB[i]: i < n)
//

const int N = 48;
static int GLOB[N];
volatile int BOUNDARY = 0;

void Writer1() {
  for (int i = 0; i < N; i++) {
    CHECK(BOUNDARY == i);
    for (int j = i; j < N; j++) {
      GLOB[j] = j;
    }
    ANNOTATE_HAPPENS_BEFORE(reinterpret_cast<void*>(BOUNDARY+1));
    BOUNDARY++;
    usleep(1000);
  }
}

void Writer2() {
  int n;
  do {
    n = BOUNDARY;
    if (n == 0) continue;
    ANNOTATE_HAPPENS_AFTER(reinterpret_cast<void*>(n));
    for (int i = 0; i < n; i++) {
      if(GLOB[i] == i) {
        GLOB[i]++;
      }
    }
    usleep(100);
  } while(n < N);
}

void Run() {
  FAST_MODE_INIT(&BOUNDARY);
  ANNOTATE_EXPECT_RACE((void*)(&BOUNDARY), "test31. Sync via 'safe' race.");
  printf("test31: negative\n");
  MyThreadArray t(Writer1, Writer2);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB[N-1]);
}
REGISTER_TEST2(Run, 31, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test31


// test32: FP. Synchronization via thread create/join. W/R. {{{1
namespace test32 {
// This test is well synchronized but helgrind 3.3.0 reports a race.
//
// Parent:                   Writer:               Reader:
// 1. Start(Reader) -----------------------\       .
//                                          \      .
// 2. Start(Writer) ---\                     \     .
//                      \---> a. MU.Lock()    \--> A. sleep(long enough)
//                            b. write(GLOB)
//                      /---- c. MU.Unlock()
// 3. Join(Writer) <---/
//                                                 B. MU.Lock()
//                                                 C. read(GLOB)
//                                   /------------ D. MU.Unlock()
// 4. Join(Reader) <----------------/
// 5. write(GLOB)
//
//
// The call to sleep() in Reader is not part of synchronization,
// it is required to trigger the false positive in helgrind 3.3.0.
//
int     GLOB = 0;
Mutex   MU;

void Writer() {
  MU.Lock();
  GLOB = 1;
  MU.Unlock();
}

void Reader() {
  usleep(480000);
  MU.Lock();
  CHECK(GLOB != 777);
  MU.Unlock();
}

void Parent() {
  MyThread r(Reader);
  MyThread w(Writer);
  r.Start();
  w.Start();

  w.Join();  // 'w' joins first.
  r.Join();

  GLOB = 2;
}

void Run() {
//  ANNOTATE_EXPECT_RACE(&GLOB, "test32. FP. Fixed by MSMProp1.");
  printf("test32: negative\n");
  Parent();
  printf("\tGLOB=%d\n", GLOB);
}

REGISTER_TEST(Run, 32);
}  // namespace test32


// test33: STAB. Stress test for the number of thread sets (TSETs). {{{1
namespace test33 {
int     GLOB = 0;
// Here we access N memory locations from within log(N) threads.
// We do it in such a way that helgrind creates nearly all possible TSETs.
// Then we join all threads and start again (N_iter times).
const int N_iter = 48;
const int Nlog  = 15;
const int N     = 1 << Nlog;
static int ARR[N];
Mutex   MU;

void Worker() {
  MU.Lock();
  int n = ++GLOB;
  MU.Unlock();

  n %= Nlog;
  for (int i = 0; i < N; i++) {
    // ARR[i] is accessed by threads from i-th subset
    if (i & (1 << n)) {
        CHECK(ARR[i] == 0);
    }
  }
}

void Run() {
  printf("test33:\n");

  std::vector<MyThread*> vec(Nlog);

  for (int i = 0; i < N_iter; i++) {
    // Create and start Nlog threads
    for (int i = 0; i < Nlog; i++) {
      vec[i] = new MyThread(Worker);
    }
    for (int i = 0; i < Nlog; i++) {
      vec[i]->Start();
    }
    // Join all threads.
    for (int i = 0; i < Nlog; i++) {
      vec[i]->Join();
      delete vec[i];
    }
    printf("------------------\n");
  }

  printf("\tGLOB=%d; ARR[1]=%d; ARR[7]=%d; ARR[N-1]=%d\n",
         GLOB, ARR[1], ARR[7], ARR[N-1]);
}
REGISTER_TEST2(Run, 33, STABILITY|EXCLUDE_FROM_ALL);
}  // namespace test33


// test34: STAB. Stress test for the number of locks sets (LSETs). {{{1
namespace test34 {
// Similar to test33, but for lock sets.
int     GLOB = 0;
const int N_iter = 48;
const int Nlog = 10;
const int N    = 1 << Nlog;
static int ARR[N];
static Mutex *MUs[Nlog];

void Worker() {
    for (int i = 0; i < N; i++) {
      // ARR[i] is protected by MUs from i-th subset of all MUs
      for (int j = 0; j < Nlog; j++)  if (i & (1 << j)) MUs[j]->Lock();
      CHECK(ARR[i] == 0);
      for (int j = 0; j < Nlog; j++)  if (i & (1 << j)) MUs[j]->Unlock();
    }
}

void Run() {
  printf("test34:\n");
  for (int iter = 0; iter < N_iter; iter++) {
    for (int i = 0; i < Nlog; i++) {
      MUs[i] = new Mutex;
    }
    MyThreadArray t(Worker, Worker);
    t.Start();
    t.Join();
    for (int i = 0; i < Nlog; i++) {
      delete MUs[i];
    }
    printf("------------------\n");
  }
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 34, STABILITY|EXCLUDE_FROM_ALL);
}  // namespace test34


// test35: PERF. Lots of mutexes and lots of call to free().  {{{1
namespace test35 {
// Helgrind 3.3.0 has very slow in shadow_mem_make_NoAccess(). Fixed locally.
// With the fix helgrind runs this test about a minute.
// Without the fix -- about 5 minutes. (on c2d 2.4GHz).
//
// TODO: need to figure out the best way for performance testing.
int **ARR;
const int N_mu   = 25000;
const int N_free = 48000;

void Worker() {
  for (int i = 0; i < N_free; i++)
    CHECK(777 == *ARR[i]);
}

void Run() {
  printf("test35:\n");
  std::vector<Mutex*> mus;

  ARR = new int *[N_free];
  for (int i = 0; i < N_free; i++) {
    const int c = N_free / N_mu;
    if ((i % c) == 0) {
      mus.push_back(new Mutex);
      mus.back()->Lock();
      mus.back()->Unlock();
    }
    ARR[i] = new int(777);
  }

  // Need to put all ARR[i] into shared state in order
  // to trigger the performance bug.
  MyThreadArray t(Worker, Worker);
  t.Start();
  t.Join();

  for (int i = 0; i < N_free; i++) delete ARR[i];
  delete [] ARR;

  for (size_t i = 0; i < mus.size(); i++) {
    delete mus[i];
  }
}
REGISTER_TEST2(Run, 35, PERFORMANCE|EXCLUDE_FROM_ALL);
}  // namespace test35


// test36: TN. Synchronization via Mutex, then PCQ. 3 threads. W/W {{{1
namespace test36 {
// variation of test28 (W/W instead of W/R)

// Putter1:                       Getter:                         Putter2:
// 1. MU.Lock();                                                  A. MU.Lock()
// 2. write(GLOB)                                                 B. write(GLOB)
// 3. MU.Unlock()                                                 C. MU.Unlock()
// 4. Q.Put() ---------\                                 /------- D. Q.Put()
// 5. MU1.Lock()        \-------> a. Q.Get()            /         E. MU1.Lock()
// 6. MU.Lock()                   b. Q.Get() <---------/          F. MU.Lock()
// 7. write(GLOB)                                                 G. write(GLOB)
// 8. MU.Unlock()                                                 H. MU.Unlock()
// 9. MU1.Unlock()                  (sleep)                       I. MU1.Unlock()
//                                c. MU1.Lock()
//                                d. write(GLOB)
//                                e. MU1.Unlock()
ProducerConsumerQueue Q(INT_MAX);
int     GLOB = 0;
Mutex   MU, MU1;

void Putter() {
  MU.Lock();
  GLOB++;
  MU.Unlock();

  Q.Put(NULL);

  MU1.Lock();
  MU.Lock();
  GLOB++;
  MU.Unlock();
  MU1.Unlock();
}

void Getter() {
  Q.Get();
  Q.Get();
  usleep(100000);
  MU1.Lock();
  GLOB++;
  MU1.Unlock();
}

void Run() {
  printf("test36: negative \n");
  MyThreadArray t(Getter, Putter, Putter);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 36);
}  // namespace test36


// test37: TN. Simple synchronization (write vs read). {{{1
namespace test37 {
int     GLOB = 0;
Mutex   MU;
// Similar to test10, but properly locked.
// Writer:             Reader:
// 1. MU.Lock()
// 2. write
// 3. MU.Unlock()
//                    a. MU.Lock()
//                    b. read
//                    c. MU.Unlock();

void Writer() {
  MU.Lock();
  GLOB = 3;
  MU.Unlock();
}
void Reader() {
  usleep(100000);
  MU.Lock();
  CHECK(GLOB != -777);
  MU.Unlock();
}

void Run() {
  printf("test37: negative\n");
  MyThreadArray t(Writer, Reader);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 37);
}  // namespace test37


// test38: TN. Synchronization via Mutexes and PCQ. 4 threads. W/W {{{1
namespace test38 {
// Fusion of test29 and test36.

// Putter1:            Putter2:           Getter1:       Getter2:
//    MU1.Lock()          MU1.Lock()
//    write(GLOB)         write(GLOB)
//    MU1.Unlock()        MU1.Unlock()
//    Q1.Put()            Q2.Put()
//    Q1.Put()            Q2.Put()
//    MU1.Lock()          MU1.Lock()
//    MU2.Lock()          MU2.Lock()
//    write(GLOB)         write(GLOB)
//    MU2.Unlock()        MU2.Unlock()
//    MU1.Unlock()        MU1.Unlock()     sleep          sleep
//                                         Q1.Get()       Q1.Get()
//                                         Q2.Get()       Q2.Get()
//                                         MU2.Lock()     MU2.Lock()
//                                         write(GLOB)    write(GLOB)
//                                         MU2.Unlock()   MU2.Unlock()
//


ProducerConsumerQueue *Q1, *Q2;
int     GLOB = 0;
Mutex   MU, MU1, MU2;

void Putter(ProducerConsumerQueue *q) {
  MU1.Lock();
  GLOB++;
  MU1.Unlock();

  q->Put(NULL);
  q->Put(NULL);

  MU1.Lock();
  MU2.Lock();
  GLOB++;
  MU2.Unlock();
  MU1.Unlock();

}

void Putter1() { Putter(Q1); }
void Putter2() { Putter(Q2); }

void Getter() {
  usleep(100000);
  Q1->Get();
  Q2->Get();

  MU2.Lock();
  GLOB++;
  MU2.Unlock();

  usleep(48000); //  TODO: remove this when FP in test32 is fixed.
}

void Run() {
  printf("test38: negative\n");
  Q1 = new ProducerConsumerQueue(INT_MAX);
  Q2 = new ProducerConsumerQueue(INT_MAX);
  MyThreadArray t(Getter, Getter, Putter1, Putter2);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
  delete Q1;
  delete Q2;
}
REGISTER_TEST(Run, 38);
}  // namespace test38

// test39: FP. Barrier. {{{1
namespace test39 {
#ifndef NO_BARRIER
// Same as test17 but uses Barrier class (pthread_barrier_t).
int     GLOB = 0;
const int N_threads = 3;
Barrier barrier(N_threads);
Mutex   MU;

void Worker() {
  MU.Lock();
  GLOB++;
  MU.Unlock();
  barrier.Block();
  CHECK(GLOB == N_threads);
}
void Run() {
  ANNOTATE_TRACE_MEMORY(&GLOB);
//  ANNOTATE_EXPECT_RACE(&GLOB, "test39. FP. Fixed by MSMProp1. Barrier.");
  printf("test39: negative\n");
  {
    ThreadPool pool(N_threads);
    pool.StartWorkers();
    for (int i = 0; i < N_threads; i++) {
      pool.Add(NewCallback(Worker));
    }
  } // all folks are joined here.
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 39);
#endif // NO_BARRIER
}  // namespace test39


// test40: FP. Synchronization via Mutexes and PCQ. 4 threads. W/W {{{1
namespace test40 {
// Similar to test38 but with different order of events (due to sleep).

// Putter1:            Putter2:           Getter1:       Getter2:
//    MU1.Lock()          MU1.Lock()
//    write(GLOB)         write(GLOB)
//    MU1.Unlock()        MU1.Unlock()
//    Q1.Put()            Q2.Put()
//    Q1.Put()            Q2.Put()
//                                        Q1.Get()       Q1.Get()
//                                        Q2.Get()       Q2.Get()
//                                        MU2.Lock()     MU2.Lock()
//                                        write(GLOB)    write(GLOB)
//                                        MU2.Unlock()   MU2.Unlock()
//
//    MU1.Lock()          MU1.Lock()
//    MU2.Lock()          MU2.Lock()
//    write(GLOB)         write(GLOB)
//    MU2.Unlock()        MU2.Unlock()
//    MU1.Unlock()        MU1.Unlock()


ProducerConsumerQueue *Q1, *Q2;
int     GLOB = 0;
Mutex   MU, MU1, MU2;

void Putter(ProducerConsumerQueue *q) {
  MU1.Lock();
  GLOB++;
  MU1.Unlock();

  q->Put(NULL);
  q->Put(NULL);
  usleep(100000);

  MU1.Lock();
  MU2.Lock();
  GLOB++;
  MU2.Unlock();
  MU1.Unlock();

}

void Putter1() { Putter(Q1); }
void Putter2() { Putter(Q2); }

void Getter() {
  Q1->Get();
  Q2->Get();

  MU2.Lock();
  GLOB++;
  MU2.Unlock();

  usleep(48000); //  TODO: remove this when FP in test32 is fixed.
}

void Run() {
//  ANNOTATE_EXPECT_RACE(&GLOB, "test40. FP. Fixed by MSMProp1. Complex Stuff.");
  printf("test40: negative\n");
  Q1 = new ProducerConsumerQueue(INT_MAX);
  Q2 = new ProducerConsumerQueue(INT_MAX);
  MyThreadArray t(Getter, Getter, Putter1, Putter2);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
  delete Q1;
  delete Q2;
}
REGISTER_TEST(Run, 40);
}  // namespace test40

// test41: TN. Test for race that appears when loading a dynamic symbol. {{{1
namespace test41 {
void Worker() {
  ANNOTATE_NO_OP(NULL); // An empty function, loaded from dll.
}
void Run() {
  printf("test41: negative\n");
  MyThreadArray t(Worker, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 41, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test41


// test42: TN. Using the same cond var several times. {{{1
namespace test42 {
int GLOB = 0;
int COND = 0;
int N_threads = 3;
Mutex   MU;

void Worker1() {
  GLOB=1;

  MU.Lock();
  COND = 1;
  CV.Signal();
  MU.Unlock();

  MU.Lock();
  while (COND != 0)
    CV.Wait(&MU);
  ANNOTATE_CONDVAR_LOCK_WAIT(&CV, &MU);
  MU.Unlock();

  GLOB=3;

}

void Worker2() {

  MU.Lock();
  while (COND != 1)
    CV.Wait(&MU);
  ANNOTATE_CONDVAR_LOCK_WAIT(&CV, &MU);
  MU.Unlock();

  GLOB=2;

  MU.Lock();
  COND = 0;
  CV.Signal();
  MU.Unlock();

}

void Run() {
//  ANNOTATE_EXPECT_RACE(&GLOB, "test42. TN. debugging.");
  printf("test42: negative\n");
  MyThreadArray t(Worker1, Worker2);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 42, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test42


// test43: TN. {{{1
namespace test43 {
//
// Putter:            Getter:
// 1. write
// 2. Q.Put() --\     .
// 3. read       \--> a. Q.Get()
//                    b. read
int     GLOB = 0;
ProducerConsumerQueue Q(INT_MAX);
void Putter() {
  GLOB = 1;
  Q.Put(NULL);
  CHECK(GLOB == 1);
}
void Getter() {
  Q.Get();
  usleep(100000);
  CHECK(GLOB == 1);
}
void Run() {
  printf("test43: negative\n");
  MyThreadArray t(Putter, Getter);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 43)
}  // namespace test43


// test44: FP. {{{1
namespace test44 {
//
// Putter:            Getter:
// 1. read
// 2. Q.Put() --\     .
// 3. MU.Lock()  \--> a. Q.Get()
// 4. write
// 5. MU.Unlock()
//                    b. MU.Lock()
//                    c. write
//                    d. MU.Unlock();
int     GLOB = 0;
Mutex   MU;
ProducerConsumerQueue Q(INT_MAX);
void Putter() {
  CHECK(GLOB == 0);
  Q.Put(NULL);
  MU.Lock();
  GLOB = 1;
  MU.Unlock();
}
void Getter() {
  Q.Get();
  usleep(100000);
  MU.Lock();
  GLOB = 1;
  MU.Unlock();
}
void Run() {
//  ANNOTATE_EXPECT_RACE(&GLOB, "test44. FP. Fixed by MSMProp1.");
  printf("test44: negative\n");
  MyThreadArray t(Putter, Getter);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 44)
}  // namespace test44


// test45: TN. {{{1
namespace test45 {
//
// Putter:            Getter:
// 1. read
// 2. Q.Put() --\     .
// 3. MU.Lock()  \--> a. Q.Get()
// 4. write
// 5. MU.Unlock()
//                    b. MU.Lock()
//                    c. read
//                    d. MU.Unlock();
int     GLOB = 0;
Mutex   MU;
ProducerConsumerQueue Q(INT_MAX);
void Putter() {
  CHECK(GLOB == 0);
  Q.Put(NULL);
  MU.Lock();
  GLOB++;
  MU.Unlock();
}
void Getter() {
  Q.Get();
  usleep(100000);
  MU.Lock();
  CHECK(GLOB <= 1);
  MU.Unlock();
}
void Run() {
  printf("test45: negative\n");
  MyThreadArray t(Putter, Getter);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 45)
}  // namespace test45


// test46: FN. {{{1
namespace test46 {
//
// First:                             Second:
// 1. write
// 2. MU.Lock()
// 3. write
// 4. MU.Unlock()                      (sleep)
//                                    a. MU.Lock()
//                                    b. write
//                                    c. MU.Unlock();
int     GLOB = 0;
Mutex   MU;
void First() {
  GLOB++;
  MU.Lock();
  GLOB++;
  MU.Unlock();
}
void Second() {
  usleep(480000);
  MU.Lock();
  GLOB++;
  MU.Unlock();

  // just a print.
  // If we move it to Run()  we will get report in MSMHelgrind
  // due to its false positive (test32).
  MU.Lock();
  printf("\tGLOB=%d\n", GLOB);
  MU.Unlock();
}
void Run() {
  ANNOTATE_TRACE_MEMORY(&GLOB);
  MyThreadArray t(First, Second);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 46)
}  // namespace test46


// test47: TP. Not detected by pure happens-before detectors. {{{1
namespace test47 {
// A true race that can not be detected by a pure happens-before
// race detector.
//
// First:                             Second:
// 1. write
// 2. MU.Lock()
// 3. MU.Unlock()                      (sleep)
//                                    a. MU.Lock()
//                                    b. MU.Unlock();
//                                    c. write
int     GLOB = 0;
Mutex   MU;
void First() {
  GLOB=1;
  MU.Lock();
  MU.Unlock();
}
void Second() {
  usleep(480000);
  MU.Lock();
  MU.Unlock();
  GLOB++;
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  if (!Tsan_PureHappensBefore())
    ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test47. TP. Not detected by pure HB.");
  printf("test47: positive\n");
  MyThreadArray t(First, Second);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 47)
}  // namespace test47


// test48: FN. Simple race (single write vs multiple reads). {{{1
namespace test48 {
int     GLOB = 0;
// same as test10 but with single writer and  multiple readers
// A simple data race between single writer and  multiple readers.
// Write happens before Reads (enforced by sleep(1)),

//
// Writer:                    Readers:
// 1. write(GLOB)             a. sleep(long enough so that GLOB
//                                is most likely initialized by Writer)
//                            b. read(GLOB)
//
//
// Eraser algorithm does not detect the race here,
// see Section 2.2 of http://citeseer.ist.psu.edu/savage97eraser.html.
//
void Writer() {
  GLOB = 3;
}
void Reader() {
  usleep(100000);
  CHECK(GLOB != -777);
}

void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test48. TP. FN in MSMHelgrind.");
  printf("test48: positive\n");
  MyThreadArray t(Writer, Reader,Reader,Reader);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 48)
}  // namespace test48


// test49: FN. Simple race (single write vs multiple reads). {{{1
namespace test49 {
int     GLOB = 0;
// same as test10 but with multiple read operations done by a single reader
// A simple data race between writer and readers.
// Write happens before Read (enforced by sleep(1)),
//
// Writer:                    Reader:
// 1. write(GLOB)             a. sleep(long enough so that GLOB
//                                is most likely initialized by Writer)
//                            b. read(GLOB)
//                            c. read(GLOB)
//                            d. read(GLOB)
//                            e. read(GLOB)
//
//
// Eraser algorithm does not detect the race here,
// see Section 2.2 of http://citeseer.ist.psu.edu/savage97eraser.html.
//
void Writer() {
  GLOB = 3;
}
void Reader() {
  usleep(100000);
  CHECK(GLOB != -777);
  CHECK(GLOB != -777);
  CHECK(GLOB != -777);
  CHECK(GLOB != -777);
}

void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test49. TP. FN in MSMHelgrind.");
  printf("test49: positive\n");
  MyThreadArray t(Writer, Reader);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 49);
}  // namespace test49


// test50: TP. Synchronization via CondVar. {{{1
namespace test50 {
int     GLOB = 0;
Mutex   MU;
// Two last write accesses to GLOB are not synchronized
//
// Waiter:                      Waker:
// 1. COND = 0
// 2. Start(Waker)
// 3. MU.Lock()                 a. write(GLOB)
//                              b. MU.Lock()
//                              c. COND = 1
//                         /--- d. CV.Signal()
//  4. while(COND != 1)   /     e. MU.Unlock()
//       CV.Wait(MU) <---/
//  5. MU.Unlock()
//  6. write(GLOB)              f. MU.Lock()
//                              g. write(GLOB)
//                              h. MU.Unlock()


void Waker() {
  usleep(100000);  // Make sure the waiter blocks.

  GLOB = 1;

  MU.Lock();
  COND = 1;
  CV.Signal();
  MU.Unlock();

  usleep(100000);
  MU.Lock();
  GLOB = 3;
  MU.Unlock();
}

void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  COND = 0;
  pool.Add(NewCallback(Waker));

  MU.Lock();
  while(COND != 1)
    CV.Wait(&MU);
  ANNOTATE_CONDVAR_LOCK_WAIT(&CV, &MU);
  MU.Unlock();

  GLOB = 2;
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test50. TP.");
  printf("test50: positive\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 50, FEATURE|NEEDS_ANNOTATIONS);
}  // namespace test50


// test51: TP. Synchronization via CondVar: problem with several signals. {{{1
namespace test51 {
int     GLOB = 0;
int     COND = 0;
Mutex   MU;


// scheduler dependent results because of several signals
// second signal will be lost
//
// Waiter:                      Waker:
// 1. Start(Waker)
// 2. MU.Lock()
// 3. while(COND)
//       CV.Wait(MU)<-\         .
// 4. MU.Unlock()      \        .
// 5. write(GLOB)       \       a. write(GLOB)
//                       \      b. MU.Lock()
//                        \     c. COND = 1
//                         \--- d. CV.Signal()
//                              e. MU.Unlock()
//
//                              f. write(GLOB)
//
//                              g. MU.Lock()
//                              h. COND = 1
//                    LOST<---- i. CV.Signal()
//                              j. MU.Unlock()

void Waker() {

  usleep(10000);  // Make sure the waiter blocks.

  GLOB = 1;

  MU.Lock();
  COND = 1;
  CV.Signal();
  MU.Unlock();

  usleep(10000);  // Make sure the waiter is signalled.

  GLOB = 2;

  MU.Lock();
  COND = 1;
  CV.Signal();   //Lost Signal
  MU.Unlock();
}

void Waiter() {

  ThreadPool pool(1);
  pool.StartWorkers();
  pool.Add(NewCallback(Waker));

  MU.Lock();
  while(COND != 1)
    CV.Wait(&MU);
  MU.Unlock();


  GLOB = 3;
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE(&GLOB, "test51. TP.");
  printf("test51: positive\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 51);
}  // namespace test51


// test52: TP. Synchronization via CondVar: problem with several signals. {{{1
namespace test52 {
int     GLOB = 0;
int     COND = 0;
Mutex   MU;

// same as test51 but the first signal will be lost
// scheduler dependent results because of several signals
//
// Waiter:                      Waker:
// 1. Start(Waker)
//                              a. write(GLOB)
//                              b. MU.Lock()
//                              c. COND = 1
//                    LOST<---- d. CV.Signal()
//                              e. MU.Unlock()
//
// 2. MU.Lock()
// 3. while(COND)
//       CV.Wait(MU)<-\         .
// 4. MU.Unlock()      \        f. write(GLOB)
// 5. write(GLOB)       \       .
//                       \      g. MU.Lock()
//                        \     h. COND = 1
//                         \--- i. CV.Signal()
//                              j. MU.Unlock()

void Waker() {

  GLOB = 1;

  MU.Lock();
  COND = 1;
  CV.Signal();    //lost signal
  MU.Unlock();

  usleep(20000);  // Make sure the waiter blocks

  GLOB = 2;

  MU.Lock();
  COND = 1;
  CV.Signal();
  MU.Unlock();
}

void Waiter() {
  ThreadPool pool(1);
  pool.StartWorkers();
  pool.Add(NewCallback(Waker));

  usleep(10000);  // Make sure the first signal will be lost

  MU.Lock();
  while(COND != 1)
    CV.Wait(&MU);
  MU.Unlock();

  GLOB = 3;
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE(&GLOB, "test52. TP.");
  printf("test52: positive\n");
  Waiter();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 52);
}  // namespace test52


// test53: FP. Synchronization via implicit semaphore. {{{1
namespace test53 {
// Correctly synchronized test, but the common lockset is empty.
// The variable FLAG works as an implicit semaphore.
// MSMHelgrind still does not complain since it does not maintain the lockset
// at the exclusive state. But MSMProp1 does complain.
// See also test54.
//
//
// Initializer:                  Users
// 1. MU1.Lock()
// 2. write(GLOB)
// 3. FLAG = true
// 4. MU1.Unlock()
//                               a. MU1.Lock()
//                               b. f = FLAG;
//                               c. MU1.Unlock()
//                               d. if (!f) goto a.
//                               e. MU2.Lock()
//                               f. write(GLOB)
//                               g. MU2.Unlock()
//

int     GLOB = 0;
bool    FLAG = false;
Mutex   MU1, MU2;

void Initializer() {
  MU1.Lock();
  GLOB = 1000;
  FLAG = true;
  MU1.Unlock();
  usleep(100000); // just in case
}

void User() {
  bool f = false;
  while(!f) {
    MU1.Lock();
    f = FLAG;
    MU1.Unlock();
    usleep(10000);
  }
  // at this point Initializer will not access GLOB again
  MU2.Lock();
  CHECK(GLOB >= 1000);
  GLOB++;
  MU2.Unlock();
}

void Run() {
  FAST_MODE_INIT(&GLOB);
  if (!Tsan_PureHappensBefore())
    ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test53. FP. Implicit semaphore");
  printf("test53: FP. false positive, Implicit semaphore\n");
  MyThreadArray t(Initializer, User, User);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 53)
}  // namespace test53


// test54: TN. Synchronization via implicit semaphore. Annotated {{{1
namespace test54 {
// Same as test53, but annotated.
int     GLOB = 0;
bool    FLAG = false;
Mutex   MU1, MU2;

void Initializer() {
  MU1.Lock();
  GLOB = 1000;
  FLAG = true;
  ANNOTATE_CONDVAR_SIGNAL(&GLOB);
  MU1.Unlock();
  usleep(100000); // just in case
}

void User() {
  bool f = false;
  while(!f) {
    MU1.Lock();
    f = FLAG;
    MU1.Unlock();
    usleep(10000);
  }
  // at this point Initializer will not access GLOB again
  ANNOTATE_CONDVAR_WAIT(&GLOB);
  MU2.Lock();
  CHECK(GLOB >= 1000);
  GLOB++;
  MU2.Unlock();
}

void Run() {
  printf("test54: negative\n");
  MyThreadArray t(Initializer, User, User);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 54, FEATURE|NEEDS_ANNOTATIONS)
}  // namespace test54


// test55: FP. Synchronization with TryLock. Not easy for race detectors {{{1
namespace test55 {
// "Correct" synchronization with TryLock and Lock.
//
// This scheme is actually very risky.
// It is covered in detail in this video:
// http://youtube.com/watch?v=mrvAqvtWYb4 (slide 36, near 50-th minute).
int     GLOB = 0;
Mutex   MU;

void Worker_Lock() {
  GLOB = 1;
  MU.Lock();
}

void Worker_TryLock() {
  while (true) {
    if (!MU.TryLock()) {
      MU.Unlock();
      break;
    }
    else
      MU.Unlock();
    usleep(100);
  }
  GLOB = 2;
}

void Run() {
  printf("test55:\n");
  MyThreadArray t(Worker_Lock, Worker_TryLock);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 55, FEATURE|EXCLUDE_FROM_ALL);
}  // namespace test55


// test56: TP. Use of ANNOTATE_BENIGN_RACE. {{{1
namespace test56 {
// For whatever reason the user wants to treat
// a race on GLOB as a benign race.
int     GLOB = 0;
int     GLOB2 = 0;

void Worker() {
  GLOB++;
}

void Run() {
  ANNOTATE_BENIGN_RACE(&GLOB, "test56. Use of ANNOTATE_BENIGN_RACE.");
  ANNOTATE_BENIGN_RACE(&GLOB2, "No race. The tool should be silent");
  printf("test56: positive\n");
  MyThreadArray t(Worker, Worker, Worker, Worker);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 56, FEATURE|NEEDS_ANNOTATIONS)
}  // namespace test56


// test57: TN: Correct use of atomics. {{{1
namespace test57 {
int     GLOB = 0;
void Writer() {
  for (int i = 0; i < 10; i++) {
    AtomicIncrement(&GLOB, 1);
    usleep(1000);
  }
}
void Reader() {
  while (GLOB < 20) usleep(1000);
}
void Run() {
  printf("test57: negative\n");
  MyThreadArray t(Writer, Writer, Reader, Reader);
  t.Start();
  t.Join();
  CHECK(GLOB == 20);
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 57)
}  // namespace test57


// test58: TN. User defined synchronization. {{{1
namespace test58 {
int     GLOB1 = 1;
int     GLOB2 = 2;
int     FLAG1 = 0;
int     FLAG2 = 0;

// Correctly synchronized test, but the common lockset is empty.
// The variables FLAG1 and FLAG2 used for synchronization and as
// temporary variables for swapping two global values.
// Such kind of synchronization is rarely used (Excluded from all tests??).

void Worker2() {
  FLAG1=GLOB2;

  while(!FLAG2)
    ;
  GLOB2=FLAG2;
}

void Worker1() {
  FLAG2=GLOB1;

  while(!FLAG1)
    ;
  GLOB1=FLAG1;
}

void Run() {
  printf("test58:\n");
  MyThreadArray t(Worker1, Worker2);
  t.Start();
  t.Join();
  printf("\tGLOB1=%d\n", GLOB1);
  printf("\tGLOB2=%d\n", GLOB2);
}
REGISTER_TEST2(Run, 58, FEATURE|EXCLUDE_FROM_ALL)
}  // namespace test58


// test59: TN. User defined synchronization. Annotated {{{1
namespace test59 {
int     COND1 = 0;
int     COND2 = 0;
int     GLOB1 = 1;
int     GLOB2 = 2;
int     FLAG1 = 0;
int     FLAG2 = 0;
// same as test 58 but annotated

void Worker2() {
  FLAG1=GLOB2;
  ANNOTATE_CONDVAR_SIGNAL(&COND2);
  while(!FLAG2) usleep(1);
  ANNOTATE_CONDVAR_WAIT(&COND1);
  GLOB2=FLAG2;
}

void Worker1() {
  FLAG2=GLOB1;
  ANNOTATE_CONDVAR_SIGNAL(&COND1);
  while(!FLAG1) usleep(1);
  ANNOTATE_CONDVAR_WAIT(&COND2);
  GLOB1=FLAG1;
}

void Run() {
  printf("test59: negative\n");
  ANNOTATE_BENIGN_RACE(&FLAG1, "synchronization via 'safe' race");
  ANNOTATE_BENIGN_RACE(&FLAG2, "synchronization via 'safe' race");
  MyThreadArray t(Worker1, Worker2);
  t.Start();
  t.Join();
  printf("\tGLOB1=%d\n", GLOB1);
  printf("\tGLOB2=%d\n", GLOB2);
}
REGISTER_TEST2(Run, 59, FEATURE|NEEDS_ANNOTATIONS)
}  // namespace test59


// test60: TN. Correct synchronization using signal-wait {{{1
namespace test60 {
int     COND1 = 0;
int     COND2 = 0;
int     GLOB1 = 1;
int     GLOB2 = 2;
int     FLAG2 = 0;
int     FLAG1 = 0;
Mutex   MU;
// same as test 59 but synchronized with signal-wait.

void Worker2() {
  FLAG1=GLOB2;

  MU.Lock();
  COND1 = 1;
  CV.Signal();
  MU.Unlock();

  MU.Lock();
  while(COND2 != 1)
    CV.Wait(&MU);
  ANNOTATE_CONDVAR_LOCK_WAIT(&CV, &MU);
  MU.Unlock();

  GLOB2=FLAG2;
}

void Worker1() {
  FLAG2=GLOB1;

  MU.Lock();
  COND2 = 1;
  CV.Signal();
  MU.Unlock();

  MU.Lock();
  while(COND1 != 1)
    CV.Wait(&MU);
  ANNOTATE_CONDVAR_LOCK_WAIT(&CV, &MU);
  MU.Unlock();

  GLOB1=FLAG1;
}

void Run() {
  printf("test60: negative\n");
  MyThreadArray t(Worker1, Worker2);
  t.Start();
  t.Join();
  printf("\tGLOB1=%d\n", GLOB1);
  printf("\tGLOB2=%d\n", GLOB2);
}
REGISTER_TEST2(Run, 60, FEATURE|NEEDS_ANNOTATIONS)
}  // namespace test60


// test61: TN. Synchronization via Mutex as in happens-before, annotated. {{{1
namespace test61 {
Mutex MU;
int     GLOB = 0;
int     *P1 = NULL, *P2 = NULL;

// In this test Mutex lock/unlock operations introduce happens-before relation.
// We annotate the code so that MU is treated as in pure happens-before detector.


void Putter() {
  ANNOTATE_MUTEX_IS_USED_AS_CONDVAR(&MU);
  MU.Lock();
  if (P1 == NULL) {
    P1 = &GLOB;
    *P1 = 1;
  }
  MU.Unlock();
}

void Getter() {
  bool done  = false;
  while (!done) {
    MU.Lock();
    if (P1) {
      done = true;
      P2 = P1;
      P1 = NULL;
    }
    MU.Unlock();
  }
  *P2 = 2;
}


void Run() {
  printf("test61: negative\n");
  MyThreadArray t(Putter, Getter);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 61, FEATURE|NEEDS_ANNOTATIONS)
}  // namespace test61


// test62: STAB. Create as many segments as possible. {{{1
namespace test62 {
// Helgrind 3.3.0 will fail as it has a hard limit of < 2^24 segments.
// A better scheme is to implement garbage collection for segments.
ProducerConsumerQueue Q(INT_MAX);
const int N = 1 << 22;

void Putter() {
  for (int i = 0; i < N; i++){
    if ((i % (N / 8)) == 0) {
      printf("i=%d\n", i);
    }
    Q.Put(NULL);
  }
}

void Getter() {
  for (int i = 0; i < N; i++)
    Q.Get();
}

void Run() {
  printf("test62:\n");
  MyThreadArray t(Putter, Getter);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 62, STABILITY|EXCLUDE_FROM_ALL)
}  // namespace test62


// test63: STAB. Create as many segments as possible and do it fast. {{{1
namespace test63 {
// Helgrind 3.3.0 will fail as it has a hard limit of < 2^24 segments.
// A better scheme is to implement garbage collection for segments.
const int N = 1 << 24;
int C = 0;

void Putter() {
  for (int i = 0; i < N; i++){
    if ((i % (N / 8)) == 0) {
      printf("i=%d\n", i);
    }
    ANNOTATE_CONDVAR_SIGNAL(&C);
  }
}

void Getter() {
}

void Run() {
  printf("test63:\n");
  MyThreadArray t(Putter, Getter);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 63, STABILITY|EXCLUDE_FROM_ALL)
}  // namespace test63


// test64: TP. T2 happens-before T3, but T1 is independent. Reads in T1/T2. {{{1
namespace test64 {
// True race between T1 and T3:
//
// T1:                   T2:                   T3:
// 1. read(GLOB)         (sleep)
//                       a. read(GLOB)
//                       b. Q.Put() ----->    A. Q.Get()
//                                            B. write(GLOB)
//
//

int     GLOB = 0;
ProducerConsumerQueue Q(INT_MAX);

void T1() {
  CHECK(GLOB == 0);
}

void T2() {
  usleep(100000);
  CHECK(GLOB == 0);
  Q.Put(NULL);
}

void T3() {
  Q.Get();
  GLOB = 1;
}


void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test64: TP.");
  printf("test64: positive\n");
  MyThreadArray t(T1, T2, T3);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 64)
}  // namespace test64


// test65: TP. T2 happens-before T3, but T1 is independent. Writes in T1/T2. {{{1
namespace test65 {
// Similar to test64.
// True race between T1 and T3:
//
// T1:                   T2:                   T3:
// 1. MU.Lock()
// 2. write(GLOB)
// 3. MU.Unlock()         (sleep)
//                       a. MU.Lock()
//                       b. write(GLOB)
//                       c. MU.Unlock()
//                       d. Q.Put() ----->    A. Q.Get()
//                                            B. write(GLOB)
//
//

int     GLOB = 0;
Mutex   MU;
ProducerConsumerQueue Q(INT_MAX);

void T1() {
  MU.Lock();
  GLOB++;
  MU.Unlock();
}

void T2() {
  usleep(100000);
  MU.Lock();
  GLOB++;
  MU.Unlock();
  Q.Put(NULL);
}

void T3() {
  Q.Get();
  GLOB = 1;
}


void Run() {
  FAST_MODE_INIT(&GLOB);
  if (!Tsan_PureHappensBefore())
    ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test65. TP.");
  printf("test65: positive\n");
  MyThreadArray t(T1, T2, T3);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 65)
}  // namespace test65


// test66: TN. Two separate pairs of signaller/waiter using the same CV. {{{1
namespace test66 {
int     GLOB1 = 0;
int     GLOB2 = 0;
int     C1 = 0;
int     C2 = 0;
Mutex   MU;

void Signaller1() {
  GLOB1 = 1;
  MU.Lock();
  C1 = 1;
  CV.Signal();
  MU.Unlock();
}

void Signaller2() {
  GLOB2 = 1;
  usleep(100000);
  MU.Lock();
  C2 = 1;
  CV.Signal();
  MU.Unlock();
}

void Waiter1() {
  MU.Lock();
  while (C1 != 1) CV.Wait(&MU);
  ANNOTATE_CONDVAR_WAIT(&CV);
  MU.Unlock();
  GLOB1 = 2;
}

void Waiter2() {
  MU.Lock();
  while (C2 != 1) CV.Wait(&MU);
  ANNOTATE_CONDVAR_WAIT(&CV);
  MU.Unlock();
  GLOB2 = 2;
}

void Run() {
  printf("test66: negative\n");
  MyThreadArray t(Signaller1, Signaller2, Waiter1, Waiter2);
  t.Start();
  t.Join();
  printf("\tGLOB=%d/%d\n", GLOB1, GLOB2);
}
REGISTER_TEST2(Run, 66, FEATURE|NEEDS_ANNOTATIONS)
}  // namespace test66


// test67: FN. Race between Signaller1 and Waiter2 {{{1
namespace test67 {
// Similar to test66, but there is a real race here.
//
// Here we create a happens-before arc between Signaller1 and Waiter2
// even though there should be no such arc.
// However, it's probably improssible (or just very hard) to avoid it.
int     GLOB = 0;
int     C1 = 0;
int     C2 = 0;
Mutex   MU;

void Signaller1() {
  GLOB = 1;
  MU.Lock();
  C1 = 1;
  CV.Signal();
  MU.Unlock();
}

void Signaller2() {
  usleep(100000);
  MU.Lock();
  C2 = 1;
  CV.Signal();
  MU.Unlock();
}

void Waiter1() {
  MU.Lock();
  while (C1 != 1) CV.Wait(&MU);
  ANNOTATE_CONDVAR_WAIT(&CV);
  MU.Unlock();
}

void Waiter2() {
  MU.Lock();
  while (C2 != 1) CV.Wait(&MU);
  ANNOTATE_CONDVAR_WAIT(&CV);
  MU.Unlock();
  GLOB = 2;
}

void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE(&GLOB, "test67. FN. Race between Signaller1 and Waiter2");
  printf("test67: positive\n");
  MyThreadArray t(Signaller1, Signaller2, Waiter1, Waiter2);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 67, FEATURE|NEEDS_ANNOTATIONS|EXCLUDE_FROM_ALL)
}  // namespace test67


// test68: TP. Writes are protected by MU, reads are not. {{{1
namespace test68 {
// In this test, all writes to GLOB are protected by a mutex
// but some reads go unprotected.
// This is certainly a race, but in some cases such code could occur in
// a correct program. For example, the unprotected reads may be used
// for showing statistics and are not required to be precise.
int     GLOB = 0;
int     COND = 0;
const int N_writers = 3;
Mutex MU, MU1;

void Writer() {
  for (int i = 0; i < 100; i++) {
    MU.Lock();
    GLOB++;
    MU.Unlock();
  }

  // we are done
  MU1.Lock();
  COND++;
  MU1.Unlock();
}

void Reader() {
  bool cont = true;
  while (cont) {
    CHECK(GLOB >= 0);

    // are we done?
    MU1.Lock();
    if (COND == N_writers)
      cont = false;
    MU1.Unlock();
    usleep(100);
  }
}

void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE(&GLOB, "TP. Writes are protected, reads are not.");
  printf("test68: positive\n");
  MyThreadArray t(Reader, Writer, Writer, Writer);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 68)
}  // namespace test68


// test69:  {{{1
namespace test69 {
// This is the same as test68, but annotated.
// We do not want to annotate GLOB as a benign race
// because we want to allow racy reads only in certain places.
//
// TODO:
int     GLOB = 0;
int     COND = 0;
const int N_writers = 3;
int     FAKE_MU = 0;
Mutex MU, MU1;

void Writer() {
  for (int i = 0; i < 10; i++) {
    MU.Lock();
    GLOB++;
    MU.Unlock();
  }

  // we are done
  MU1.Lock();
  COND++;
  MU1.Unlock();
}

void Reader() {
  bool cont = true;
  while (cont) {
    ANNOTATE_IGNORE_READS_BEGIN();
    CHECK(GLOB >= 0);
    ANNOTATE_IGNORE_READS_END();

    // are we done?
    MU1.Lock();
    if (COND == N_writers)
      cont = false;
    MU1.Unlock();
    usleep(100);
  }
}

void Run() {
  printf("test69: negative\n");
  MyThreadArray t(Reader, Writer, Writer, Writer);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 69)
}  // namespace test69

// test70: STAB. Check that TRACE_MEMORY works. {{{1
namespace test70 {
int     GLOB = 0;
void Run() {
  printf("test70: negative\n");
  ANNOTATE_TRACE_MEMORY(&GLOB);
  GLOB = 1;
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 70)
}  // namespace test70


// test71: TN. strlen, index. {{{1
namespace test71 {
// This test is a reproducer for a benign race in strlen (as well as index, etc).
// Some implementations of strlen may read up to 7 bytes past the end of the string
// thus touching memory which may not belong to this string.
// Such race is benign because the data read past the end of the string is not used.
//
// Here, we allocate a 8-byte aligned string str and initialize first 5 bytes.
// Then one thread calls strlen(str) (as well as index & rindex)
// and another thread initializes str[5]..str[7].
//
// This can be fixed in Helgrind by intercepting strlen and replacing it
// with a simpler implementation.

char    *str;
void WorkerX() {
  usleep(100000);
  CHECK(strlen(str) == 4);
  CHECK(index(str, 'X') == str);
  CHECK(index(str, 'x') == str+1);
  CHECK(index(str, 'Y') == NULL);
  CHECK(rindex(str, 'X') == str+2);
  CHECK(rindex(str, 'x') == str+3);
  CHECK(rindex(str, 'Y') == NULL);
}
void WorkerY() {
  str[5] = 'Y';
  str[6] = 'Y';
  str[7] = '\0';
}

void Run() {
  str = new char[8];
  str[0] = 'X';
  str[1] = 'x';
  str[2] = 'X';
  str[3] = 'x';
  str[4] = '\0';

  printf("test71: negative (strlen & index)\n");
  MyThread t1(WorkerY);
  MyThread t2(WorkerX);
  t1.Start();
  t2.Start();
  t1.Join();
  t2.Join();
  printf("\tstrX=%s; strY=%s\n", str, str+5);
}
REGISTER_TEST(Run, 71)
}  // namespace test71


// test72: STAB. Stress test for the number of segment sets (SSETs). {{{1
namespace test72 {
#ifndef NO_BARRIER
// Variation of test33.
// Instead of creating Nlog*N_iter threads,
// we create Nlog threads and do N_iter barriers.
int     GLOB = 0;
const int N_iter = 30;
const int Nlog  = 16;
const int N     = 1 << Nlog;
static int64_t ARR1[N];
static int64_t ARR2[N];
Barrier *barriers[N_iter];
Mutex   MU;

void Worker() {
  MU.Lock();
  int n = ++GLOB;
  MU.Unlock();

  n %= Nlog;

  long t0 = clock();
  long t = t0;

  for (int it = 0; it < N_iter; it++) {
    if(n == 0) {
      //printf("Iter: %d; %ld %ld\n", it, clock() - t, clock() - t0);
      t = clock();
    }
    // Iterate N_iter times, block on barrier after each iteration.
    // This way Helgrind will create new segments after each barrier.

    for (int x = 0; x < 2; x++) {
      // run the inner loop twice.
      // When a memory location is accessed second time it is likely
      // that the state (SVal) will be unchanged.
      // The memory machine may optimize this case.
      for (int i = 0; i < N; i++) {
        // ARR1[i] and ARR2[N-1-i] are accessed by threads from i-th subset
        if (i & (1 << n)) {
          CHECK(ARR1[i] == 0);
          CHECK(ARR2[N-1-i] == 0);
        }
      }
    }
    barriers[it]->Block();
  }
}


void Run() {
  printf("test72:\n");

  std::vector<MyThread*> vec(Nlog);

  for (int i = 0; i < N_iter; i++)
    barriers[i] = new Barrier(Nlog);

  // Create and start Nlog threads
  for (int i = 0; i < Nlog; i++) {
    vec[i] = new MyThread(Worker);
    vec[i]->Start();
  }

  // Join all threads.
  for (int i = 0; i < Nlog; i++) {
    vec[i]->Join();
    delete vec[i];
  }
  for (int i = 0; i < N_iter; i++)
    delete barriers[i];

  /*printf("\tGLOB=%d; ARR[1]=%d; ARR[7]=%d; ARR[N-1]=%d\n",
         GLOB, (int)ARR1[1], (int)ARR1[7], (int)ARR1[N-1]);*/
}
REGISTER_TEST2(Run, 72, STABILITY|PERFORMANCE|EXCLUDE_FROM_ALL);
#endif // NO_BARRIER
}  // namespace test72


// test73: STAB. Stress test for the number of (SSETs), different access sizes. {{{1
namespace test73 {
#ifndef NO_BARRIER
// Variation of test72.
// We perform accesses of different sizes to the same location.
int     GLOB = 0;
const int N_iter = 2;
const int Nlog  = 16;
const int N     = 1 << Nlog;
static int64_t ARR1[N];
static int ARR2[N];
Barrier *barriers[N_iter];
Mutex   MU;

void Worker() {
  MU.Lock();
  int n = ++GLOB;
  MU.Unlock();

  n %= Nlog;

  for (int it = 0; it < N_iter; it++) {
    // Iterate N_iter times, block on barrier after each iteration.
    // This way Helgrind will create new segments after each barrier.

    for (int x = 0; x < 4; x++) {
      for (int i = 0; i < N; i++) {
        // ARR1[i] are accessed by threads from i-th subset
        if (i & (1 << n)) {
          for (int off = 0; off < (1 << x); off++) {
            switch(x) {
              case 0: CHECK(          ARR1  [i * (1<<x) + off] == 0); break;
              case 1: CHECK(((int*)  (ARR1))[i * (1<<x) + off] == 0); break;
              case 2: CHECK(((short*)(ARR1))[i * (1<<x) + off] == 0); break;
              case 3: CHECK(((char*) (ARR1))[i * (1<<x) + off] == 0); break;
            }
            switch(x) {
              case 1: CHECK(((int*)  (ARR2))[i * (1<<x) + off] == 0); break;
              case 2: CHECK(((short*)(ARR2))[i * (1<<x) + off] == 0); break;
              case 3: CHECK(((char*) (ARR2))[i * (1<<x) + off] == 0); break;
            }
          }
        }
      }
    }
    barriers[it]->Block();
  }
}


void Run() {
  printf("test73:\n");

  std::vector<MyThread*> vec(Nlog);

  for (int i = 0; i < N_iter; i++)
    barriers[i] = new Barrier(Nlog);

  // Create and start Nlog threads
  for (int i = 0; i < Nlog; i++) {
    vec[i] = new MyThread(Worker);
    vec[i]->Start();
  }

  // Join all threads.
  for (int i = 0; i < Nlog; i++) {
    vec[i]->Join();
    delete vec[i];
  }
  for (int i = 0; i < N_iter; i++)
    delete barriers[i];

  /*printf("\tGLOB=%d; ARR[1]=%d; ARR[7]=%d; ARR[N-1]=%d\n",
         GLOB, (int)ARR1[1], (int)ARR1[7], (int)ARR1[N-1]);*/
}
REGISTER_TEST2(Run, 73, STABILITY|PERFORMANCE|EXCLUDE_FROM_ALL);
#endif // NO_BARRIER
}  // namespace test73


// test74: PERF. A lot of lock/unlock calls. {{{1
namespace    test74 {
const int N = 100000;
Mutex   MU;
void Run() {
  printf("test74: perf\n");
  for (int i = 0; i < N; i++ ) {
    MU.Lock();
    MU.Unlock();
  }
}
REGISTER_TEST(Run, 74)
}  // namespace test74


// test75: TN. Test for sem_post, sem_wait, sem_trywait. {{{1
namespace test75 {
int     GLOB = 0;
sem_t   sem[2];

void Poster() {
  GLOB = 1;
  sem_post(&sem[0]);
  sem_post(&sem[1]);
}

void Waiter() {
  sem_wait(&sem[0]);
  CHECK(GLOB==1);
}
void TryWaiter() {
  usleep(500000);
  sem_trywait(&sem[1]);
  CHECK(GLOB==1);
}

void Run() {
#ifndef DRT_NO_SEM
  sem_init(&sem[0], 0, 0);
  sem_init(&sem[1], 0, 0);

  printf("test75: negative\n");
  {
    MyThreadArray t(Poster, Waiter);
    t.Start();
    t.Join();
  }
  GLOB = 2;
  {
    MyThreadArray t(Poster, TryWaiter);
    t.Start();
    t.Join();
  }
  printf("\tGLOB=%d\n", GLOB);

  sem_destroy(&sem[0]);
  sem_destroy(&sem[1]);
#endif
}
REGISTER_TEST(Run, 75)
}  // namespace test75

// RefCountedClass {{{1
struct RefCountedClass {
 public:
  RefCountedClass() {
    annotate_unref_ = false;
    ref_ = 0;
    data_ = 0;
  }

  ~RefCountedClass() {
    CHECK(ref_ == 0);     // race may be reported here
    int data_val = data_; // and here
                          // if MU is not annotated
    data_ = 0;
    ref_ = -1;
    printf("\tRefCountedClass::data_ = %d\n", data_val);
  }

  void AccessData() {
    this->mu_.Lock();
    this->data_++;
    this->mu_.Unlock();
  }

  void Ref() {
    MU.Lock();
    CHECK(ref_ >= 0);
    ref_++;
    MU.Unlock();
  }

  void Unref() {
    MU.Lock();
    CHECK(ref_ > 0);
    ref_--;
    bool do_delete = ref_ == 0;
    if (annotate_unref_) {
      ANNOTATE_CONDVAR_SIGNAL(this);
    }
    MU.Unlock();
    if (do_delete) {
      if (annotate_unref_) {
        ANNOTATE_CONDVAR_WAIT(this);
      }
      delete this;
    }
  }

  static void Annotate_MU() {
    ANNOTATE_MUTEX_IS_USED_AS_CONDVAR(&MU);
  }
  void AnnotateUnref() {
    annotate_unref_ = true;
  }
  void Annotate_Race() {
    ANNOTATE_BENIGN_RACE(&this->data_, "needs annotation");
    ANNOTATE_BENIGN_RACE(&this->ref_, "needs annotation");
  }
 private:
  bool annotate_unref_;

  int data_;
  Mutex mu_; // protects data_

  int ref_;
  static Mutex MU; // protects ref_
};

Mutex RefCountedClass::MU;

// test76: FP. Ref counting, no annotations. {{{1
namespace test76 {
#ifndef NO_BARRIER
int     GLOB = 0;
Barrier barrier(4);
RefCountedClass *object = NULL;
void Worker() {
  object->Ref();
  barrier.Block();
  object->AccessData();
  object->Unref();
}
void Run() {
  printf("test76: false positive (ref counting)\n");
  object = new RefCountedClass;
  object->Annotate_Race();
  MyThreadArray t(Worker, Worker, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 76, FEATURE)
#endif // NO_BARRIER
}  // namespace test76


// test77: TN. Ref counting, MU is annotated. {{{1
namespace test77 {
#ifndef NO_BARRIER
// same as test76, but RefCountedClass::MU is annotated.
int     GLOB = 0;
Barrier barrier(4);
RefCountedClass *object = NULL;
void Worker() {
  object->Ref();
  barrier.Block();
  object->AccessData();
  object->Unref();
}
void Run() {
  printf("test77: true negative (ref counting), mutex is annotated\n");
  RefCountedClass::Annotate_MU();
  object = new RefCountedClass;
  MyThreadArray t(Worker, Worker, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 77)
#endif // NO_BARRIER
}  // namespace test77


// test78: TN. Ref counting, Unref is annotated. {{{1
namespace test78 {
#ifndef NO_BARRIER
// same as test76, but RefCountedClass::Unref is annotated.
int     GLOB = 0;
Barrier barrier(4);
RefCountedClass *object = NULL;
void Worker() {
  object->Ref();
  barrier.Block();
  object->AccessData();
  object->Unref();
}
void Run() {
  printf("test78: true negative (ref counting), Unref is annotated\n");
  RefCountedClass::Annotate_MU();
  object = new RefCountedClass;
  MyThreadArray t(Worker, Worker, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 78)
#endif // NO_BARRIER
}  // namespace test78


// test79 TN. Swap. {{{1
namespace test79 {
#if 0
typedef __gnu_cxx::hash_map<int, int> map_t;
#else
typedef std::map<int, int> map_t;
#endif
map_t   MAP;
Mutex   MU;

// Here we use swap to pass MAP between threads.
// The synchronization is correct, but w/o ANNOTATE_MUTEX_IS_USED_AS_CONDVAR
// Helgrind will complain.

void Worker1() {
  map_t tmp;
  MU.Lock();
  // We swap the new empty map 'tmp' with 'MAP'.
  MAP.swap(tmp);
  MU.Unlock();
  // tmp (which is the old version of MAP) is destroyed here.
}

void Worker2() {
  MU.Lock();
  MAP[1]++;  // Just update MAP under MU.
  MU.Unlock();
}

void Worker3() { Worker1(); }
void Worker4() { Worker2(); }

void Run() {
  ANNOTATE_MUTEX_IS_USED_AS_CONDVAR(&MU);
  printf("test79: negative\n");
  MyThreadArray t(Worker1, Worker2, Worker3, Worker4);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 79)
}  // namespace test79


// AtomicRefCountedClass. {{{1
// Same as RefCountedClass, but using atomic ops instead of mutex.
struct AtomicRefCountedClass {
 public:
  AtomicRefCountedClass() {
    annotate_unref_ = false;
    ref_ = 0;
    data_ = 0;
  }

  ~AtomicRefCountedClass() {
    CHECK(ref_ == 0);     // race may be reported here
    int data_val = data_; // and here
    data_ = 0;
    ref_ = -1;
    printf("\tRefCountedClass::data_ = %d\n", data_val);
  }

  void AccessData() {
    this->mu_.Lock();
    this->data_++;
    this->mu_.Unlock();
  }

  void Ref() {
    AtomicIncrement(&ref_, 1);
  }

  void Unref() {
    // DISCLAIMER: I am not sure I've implemented this correctly
    // (might require some memory barrier, etc).
    // But this implementation of reference counting is enough for
    // the purpose of Helgrind demonstration.
    AtomicIncrement(&ref_, -1);
    if (annotate_unref_) { ANNOTATE_CONDVAR_SIGNAL(this); }
    if (ref_ == 0) {
      if (annotate_unref_) { ANNOTATE_CONDVAR_WAIT(this); }
      delete this;
    }
  }

  void AnnotateUnref() {
    annotate_unref_ = true;
  }
  void Annotate_Race() {
    ANNOTATE_BENIGN_RACE(&this->data_, "needs annotation");
  }
 private:
  bool annotate_unref_;

  Mutex mu_;
  int data_; // under mu_

  int ref_;  // used in atomic ops.
};

// test80: FP. Ref counting with atomics, no annotations. {{{1
namespace test80 {
#ifndef NO_BARRIER
int     GLOB = 0;
Barrier barrier(4);
AtomicRefCountedClass *object = NULL;
void Worker() {
  object->Ref();
  barrier.Block();
  object->AccessData();
  object->Unref(); // All the tricky stuff is here.
}
void Run() {
  printf("test80: false positive (ref counting)\n");
  object = new AtomicRefCountedClass;
  object->Annotate_Race();
  MyThreadArray t(Worker, Worker, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 80, FEATURE|EXCLUDE_FROM_ALL)
#endif // NO_BARRIER
}  // namespace test80


// test81: TN. Ref counting with atomics, Unref is annotated. {{{1
namespace test81 {
#ifndef NO_BARRIER
// same as test80, but Unref is annotated.
int     GLOB = 0;
Barrier barrier(4);
AtomicRefCountedClass *object = NULL;
void Worker() {
  object->Ref();
  barrier.Block();
  object->AccessData();
  object->Unref(); // All the tricky stuff is here.
}
void Run() {
  printf("test81: negative (annotated ref counting)\n");
  object = new AtomicRefCountedClass;
  object->AnnotateUnref();
  MyThreadArray t(Worker, Worker, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 81, FEATURE|EXCLUDE_FROM_ALL)
#endif // NO_BARRIER
}  // namespace test81


// test82: Object published w/o synchronization. {{{1
namespace test82 {

// Writer creates a new object and makes the pointer visible to the Reader.
// Reader waits until the object pointer is non-null and reads the object.
//
// On Core 2 Duo this test will sometimes (quite rarely) fail in
// the CHECK below, at least if compiled with -O2.
//
// The sequence of events::
// Thread1:                  Thread2:
//   a. arr_[...] = ...
//   b. foo[i]    = ...
//                           A. ... = foo[i]; // non NULL
//                           B. ... = arr_[...];
//
//  Since there is no proper synchronization, during the even (B)
//  Thread2 may not see the result of the event (a).
//  On x86 and x86_64 this happens due to compiler reordering instructions.
//  On other arcitectures it may also happen due to cashe inconsistency.

class FOO {
 public:
  FOO() {
    idx_ = rand() % 1024;
    arr_[idx_] = 77777;
  //   __asm__ __volatile__("" : : : "memory"); // this fixes!
  }
  static void check(volatile FOO *foo) {
    CHECK(foo->arr_[foo->idx_] == 77777);
  }
 private:
  int idx_;
  int arr_[1024];
};

const int N = 100000;
static volatile FOO *foo[N];
Mutex   MU;

void Writer() {
  for (int i = 0; i < N; i++) {
    foo[i] = new FOO;
    usleep(100);
  }
}

void Reader() {
  for (int i = 0; i < N; i++) {
    while (!foo[i]) {
      MU.Lock();   // this is NOT a synchronization,
      MU.Unlock(); // it just helps foo[i] to become visible in Reader.
    }
    if ((i % 100) == 0) {
      printf("rd %d\n", i);
    }
    // At this point Reader() sees the new value of foo[i]
    // but in very rare cases will not see the new value of foo[i]->arr_.
    // Thus this CHECK will sometimes fail.
    FOO::check(foo[i]);
  }
}

void Run() {
  printf("test82: positive\n");
  MyThreadArray t(Writer, Reader);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 82, FEATURE|EXCLUDE_FROM_ALL)
}  // namespace test82


// test83: Object published w/o synchronization (simple version){{{1
namespace test83 {
// A simplified version of test83 (example of a wrong code).
// This test, though incorrect, will almost never fail.
volatile static int *ptr = NULL;
Mutex   MU;

void Writer() {
  usleep(100);
  ptr = new int(777);
}

void Reader() {
  while(!ptr) {
    MU.Lock(); // Not a synchronization!
    MU.Unlock();
  }
  CHECK(*ptr == 777);
}

void Run() {
//  printf("test83: positive\n");
  MyThreadArray t(Writer, Reader);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 83, FEATURE|EXCLUDE_FROM_ALL)
}  // namespace test83


// test84: TP. True race (regression test for a bug related to atomics){{{1
namespace test84 {
// Helgrind should not create HB arcs for the bus lock even when
// --pure-happens-before=yes is used.
// Bug found in by Bart Van Assche, the test is taken from
// valgrind file drd/tests/atomic_var.c.
static int s_x = 0;
/* s_dummy[] ensures that s_x and s_y are not in the same cache line. */
static char s_dummy[512] = {0};
static int s_y;

void thread_func_1()
{
  s_y = 1;
  AtomicIncrement(&s_x, 1);
}

void thread_func_2()
{
  while (AtomicIncrement(&s_x, 0) == 0)
    ;
  printf("y = %d\n", s_y);
}


void Run() {
  CHECK(s_dummy[0] == 0);  // Avoid compiler warning about 's_dummy unused'.
  printf("test84: positive\n");
  FAST_MODE_INIT(&s_y);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&s_y, "test84: TP. true race.");
  MyThreadArray t(thread_func_1, thread_func_2);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 84)
}  // namespace test84


// test85: Test for RunningOnValgrind(). {{{1
namespace  test85 {
int     GLOB = 0;
void Run() {
  printf("test85: RunningOnValgrind() = %d\n", RunningOnValgrind());
}
REGISTER_TEST(Run, 85)
}  // namespace test85


// test86: Test for race inside DTOR: racey write to vptr. Benign. {{{1
namespace test86 {
// This test shows a racey access to vptr (the pointer to vtbl).
// We have class A and class B derived from A.
// Both classes have a virtual function f() and a virtual DTOR.
// We create an object 'A *a = new B'
// and pass this object from Thread1 to Thread2.
// Thread2 calls a->f(). This call reads a->vtpr.
// Thread1 deletes the object. B::~B waits untill the object can be destroyed
// (flag_stopped == true) but at the very beginning of B::~B
// a->vptr is written to.
// So, we have a race on a->vptr.
// On this particular test this race is benign, but test87 shows
// how such race could harm.
//
//
//
// Threa1:                                            Thread2:
// 1. A a* = new B;
// 2. Q.Put(a); ------------\                         .
//                           \-------------------->   a. a = Q.Get();
//                                                    b. a->f();
//                                       /---------   c. flag_stopped = true;
// 3. delete a;                         /
//    waits untill flag_stopped <------/
//    inside the dtor
//

bool flag_stopped = false;
Mutex mu;

ProducerConsumerQueue Q(INT_MAX);  // Used to pass A* between threads.

struct A {
  A()  { printf("A::A()\n"); }
  virtual ~A() { printf("A::~A()\n"); }
  virtual void f() { }

  uintptr_t padding[15];
} __attribute__ ((aligned (64)));

struct B: A {
  B()  { printf("B::B()\n"); }
  virtual ~B() {
    // The race is here.    <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
    printf("B::~B()\n");
    // wait until flag_stopped is true.
    mu.LockWhen(Condition(&ArgIsTrue, &flag_stopped));
    mu.Unlock();
    printf("B::~B() done\n");
  }
  virtual void f() { }
};

void Waiter() {
  A *a = new B;
  if (!Tsan_FastMode())
    ANNOTATE_EXPECT_RACE(a, "test86: expected race on a->vptr");
  printf("Waiter: B created\n");
  Q.Put(a);
  usleep(100000); // so that Worker calls a->f() first.
  printf("Waiter: deleting B\n");
  delete a;
  printf("Waiter: B deleted\n");
  usleep(100000);
  printf("Waiter: done\n");
}

void Worker() {
  A *a = reinterpret_cast<A*>(Q.Get());
  printf("Worker: got A\n");
  a->f();

  mu.Lock();
  flag_stopped = true;
  mu.Unlock();
  usleep(200000);
  printf("Worker: done\n");
}

void Run() {
  printf("test86: positive, race inside DTOR\n");
  MyThreadArray t(Waiter, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 86)
}  // namespace test86


// test87: Test for race inside DTOR: racey write to vptr. Harmful.{{{1
namespace test87 {
// A variation of test86 where the race is harmful.
// Here we have class C derived from B.
// We create an object 'A *a = new C' in Thread1 and pass it to Thread2.
// Thread2 calls a->f().
// Thread1 calls 'delete a'.
// It first calls C::~C, then B::~B where it rewrites the vptr to point
// to B::vtbl. This is a problem because Thread2 might not have called a->f()
// and now it will call B::f instead of C::f.
//
bool flag_stopped = false;
Mutex mu;

ProducerConsumerQueue Q(INT_MAX);  // Used to pass A* between threads.

struct A {
  A()  { printf("A::A()\n"); }
  virtual ~A() { printf("A::~A()\n"); }
  virtual void f() = 0; // pure virtual.
};

struct B: A {
  B()  { printf("B::B()\n"); }
  virtual ~B() {
    // The race is here.    <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
    printf("B::~B()\n");
    // wait until flag_stopped is true.
    mu.LockWhen(Condition(&ArgIsTrue, &flag_stopped));
    mu.Unlock();
    printf("B::~B() done\n");
  }
  virtual void f() = 0; // pure virtual.
};

struct C: B {
  C()  { printf("C::C()\n"); }
  virtual ~C() { printf("C::~C()\n"); }
  virtual void f() { }
};

void Waiter() {
  A *a = new C;
  Q.Put(a);
  delete a;
}

void Worker() {
  A *a = reinterpret_cast<A*>(Q.Get());
  a->f();

  mu.Lock();
  flag_stopped = true;
  ANNOTATE_CONDVAR_SIGNAL(&mu);
  mu.Unlock();
}

void Run() {
  printf("test87: positive, race inside DTOR\n");
  MyThreadArray t(Waiter, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 87, FEATURE|EXCLUDE_FROM_ALL)
}  // namespace test87


// test88: Test for ANNOTATE_IGNORE_WRITES_*{{{1
namespace test88 {
// a recey write annotated with ANNOTATE_IGNORE_WRITES_BEGIN/END.
int     GLOB = 0;
void Worker() {
  ANNOTATE_IGNORE_WRITES_BEGIN();
  GLOB = 1;
  ANNOTATE_IGNORE_WRITES_END();
}
void Run() {
  printf("test88: negative, test for ANNOTATE_IGNORE_WRITES_*\n");
  MyThread t(Worker);
  t.Start();
  GLOB = 1;
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 88)
}  // namespace test88


// test89: Test for debug info. {{{1
namespace test89 {
// Simlpe races with different objects (stack, heap globals; scalars, structs).
// Also, if run with --trace-level=2 this test will show a sequence of
// CTOR and DTOR calls.
struct STRUCT {
  int a, b, c;
};

struct A {
  int a;
  A() {
    ANNOTATE_TRACE_MEMORY(&a);
    a = 1;
  }
  virtual ~A() {
    a = 4;
  }
};

struct B : A {
  B()  { CHECK(a == 1); }
  virtual ~B() { CHECK(a == 3); }
};
struct C : B {
  C()  { a = 2; }
  virtual ~C() { a = 3; }
};

int            GLOBAL = 0;
int           *STACK  = 0;
STRUCT         GLOB_STRUCT;
STRUCT        *STACK_STRUCT;
STRUCT        *HEAP_STRUCT;

void Worker() {
  GLOBAL = 1;
  *STACK = 1;
  GLOB_STRUCT.b   = 1;
  STACK_STRUCT->b = 1;
  HEAP_STRUCT->b  = 1;
}

void Run() {
  int stack_var = 0;
  STACK = &stack_var;

  STRUCT stack_struct;
  STACK_STRUCT = &stack_struct;

  HEAP_STRUCT = new STRUCT;

  printf("test89: negative\n");
  MyThreadArray t(Worker, Worker);
  t.Start();
  t.Join();

  delete HEAP_STRUCT;

  A *a = new C;
  printf("Using 'a->a':  %d\n", a->a);
  delete a;
}
REGISTER_TEST2(Run, 89, FEATURE|EXCLUDE_FROM_ALL)
}  // namespace test89


// test90: FP. Test for a safely-published pointer (read-only). {{{1
namespace test90 {
// The Publisher creates an object and safely publishes it under a mutex.
// Readers access the object read-only.
// See also test91.
//
// Without annotations Helgrind will issue a false positive in Reader().
//
// Choices for annotations:
//   -- ANNOTATE_CONDVAR_SIGNAL/ANNOTATE_CONDVAR_WAIT
//   -- ANNOTATE_MUTEX_IS_USED_AS_CONDVAR
//   -- ANNOTATE_PUBLISH_MEMORY_RANGE.

int     *GLOB = 0;
Mutex   MU;

void Publisher() {
  MU.Lock();
  GLOB = (int*)memalign(64, sizeof(int));
  *GLOB = 777;
  if (!Tsan_PureHappensBefore() && !Tsan_FastMode())
    ANNOTATE_EXPECT_RACE_FOR_TSAN(GLOB, "test90. FP. This is a false positve");
  MU.Unlock();
  usleep(200000);
}

void Reader() {
  usleep(10000);
  while (true) {
    MU.Lock();
    int *p = GLOB;
    MU.Unlock();
    if (p) {
      CHECK(*p == 777);  // Race is reported here.
      break;
    }
  }
}

void Run() {
  printf("test90: false positive (safely published pointer).\n");
  MyThreadArray t(Publisher, Reader, Reader, Reader);
  t.Start();
  t.Join();
  printf("\t*GLOB=%d\n", *GLOB);
  free(GLOB);
}
REGISTER_TEST(Run, 90)
}  // namespace test90


// test91: FP. Test for a safely-published pointer (read-write). {{{1
namespace test91 {
// Similar to test90.
// The Publisher creates an object and safely publishes it under a mutex MU1.
// Accessors get the object under MU1 and access it (read/write) under MU2.
//
// Without annotations Helgrind will issue a false positive in Accessor().
//

int     *GLOB = 0;
Mutex   MU, MU1, MU2;

void Publisher() {
  MU1.Lock();
  GLOB = (int*)memalign(64, sizeof(int));
  *GLOB = 777;
  if (!Tsan_PureHappensBefore() && !Tsan_FastMode())
    ANNOTATE_EXPECT_RACE_FOR_TSAN(GLOB, "test91. FP. This is a false positve");
  MU1.Unlock();
}

void Accessor() {
  usleep(10000);
  while (true) {
    MU1.Lock();
    int *p = GLOB;
    MU1.Unlock();
    if (p) {
      MU2.Lock();
      (*p)++;  // Race is reported here.
      CHECK(*p >  777);
      MU2.Unlock();
      break;
    }
  }
}

void Run() {
  printf("test91: false positive (safely published pointer, read/write).\n");
  MyThreadArray t(Publisher, Accessor, Accessor, Accessor);
  t.Start();
  t.Join();
  printf("\t*GLOB=%d\n", *GLOB);
  free(GLOB);
}
REGISTER_TEST(Run, 91)
}  // namespace test91


// test92: TN. Test for a safely-published pointer (read-write), annotated. {{{1
namespace test92 {
// Similar to test91, but annotated with ANNOTATE_PUBLISH_MEMORY_RANGE.
//
//
// Publisher:                                       Accessors:
//
// 1. MU1.Lock()
// 2. Create GLOB.
// 3. ANNOTATE_PUBLISH_...(GLOB) -------\            .
// 4. MU1.Unlock()                       \           .
//                                        \          a. MU1.Lock()
//                                         \         b. Get GLOB
//                                          \        c. MU1.Unlock()
//                                           \-->    d. Access GLOB
//
//  A happens-before arc is created between ANNOTATE_PUBLISH_MEMORY_RANGE and
//  accesses to GLOB.

struct ObjType {
  int arr[10];
};

ObjType *GLOB = 0;
Mutex   MU, MU1, MU2;

void Publisher() {
  MU1.Lock();
  GLOB = new ObjType;
  for (int i = 0; i < 10; i++) {
    GLOB->arr[i] = 777;
  }
  // This annotation should go right before the object is published.
  ANNOTATE_PUBLISH_MEMORY_RANGE(GLOB, sizeof(*GLOB));
  MU1.Unlock();
}

void Accessor(int index) {
  while (true) {
    MU1.Lock();
    ObjType *p = GLOB;
    MU1.Unlock();
    if (p) {
      MU2.Lock();
      p->arr[index]++;  // W/o the annotations the race will be reported here.
      CHECK(p->arr[index] ==  778);
      MU2.Unlock();
      break;
    }
  }
}

void Accessor0() { Accessor(0); }
void Accessor5() { Accessor(5); }
void Accessor9() { Accessor(9); }

void Run() {
  printf("test92: safely published pointer, read/write, annotated.\n");
  MyThreadArray t(Publisher, Accessor0, Accessor5, Accessor9);
  t.Start();
  t.Join();
  printf("\t*GLOB=%d\n", GLOB->arr[0]);
}
REGISTER_TEST(Run, 92)
}  // namespace test92


// test93: TP. Test for incorrect usage of ANNOTATE_PUBLISH_MEMORY_RANGE. {{{1
namespace test93 {
int     GLOB = 0;

void Reader() {
  CHECK(GLOB == 0);
}

void Publisher() {
  usleep(10000);
  // Incorrect, used after the memory has been accessed in another thread.
  ANNOTATE_PUBLISH_MEMORY_RANGE(&GLOB, sizeof(GLOB));
}

void Run() {
  printf("test93: positive, misuse of ANNOTATE_PUBLISH_MEMORY_RANGE\n");
  MyThreadArray t(Reader, Publisher);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 93, FEATURE|EXCLUDE_FROM_ALL)
}  // namespace test93


// test94: TP. Check do_cv_signal/fake segment logic {{{1
namespace test94 {
int     GLOB;

int COND  = 0;
int COND2 = 0;
Mutex MU, MU2;
CondVar CV, CV2;

void Thr1() {
  usleep(10000);  // Make sure the waiter blocks.

  GLOB = 1; // WRITE

  MU.Lock();
  COND = 1;
  CV.Signal();
  MU.Unlock();
}
void Thr2() {
  usleep(1000*1000); // Make sure CV2.Signal() "happens after" CV.Signal()
  usleep(10000);  // Make sure the waiter blocks.

  MU2.Lock();
  COND2 = 1;
  CV2.Signal();
  MU2.Unlock();
}
void Thr3() {
  MU.Lock();
  while(COND != 1)
    CV.Wait(&MU);
  MU.Unlock();
}
void Thr4() {
  MU2.Lock();
  while(COND2 != 1)
    CV2.Wait(&MU2);
  MU2.Unlock();
  GLOB = 2; // READ: no HB-relation between CV.Signal and CV2.Wait !
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test94: TP.");
  printf("test94: TP. Check do_cv_signal/fake segment logic\n");
  MyThreadArray mta(Thr1, Thr2, Thr3, Thr4);
  mta.Start();
  mta.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 94);
}  // namespace test94

// test95: TP. Check do_cv_signal/fake segment logic {{{1
namespace test95 {
int     GLOB = 0;

int COND  = 0;
int COND2 = 0;
Mutex MU, MU2;
CondVar CV, CV2;

void Thr1() {
  usleep(1000*1000); // Make sure CV2.Signal() "happens before" CV.Signal()
  usleep(10000);  // Make sure the waiter blocks.

  GLOB = 1; // WRITE

  MU.Lock();
  COND = 1;
  CV.Signal();
  MU.Unlock();
}
void Thr2() {
  usleep(10000);  // Make sure the waiter blocks.

  MU2.Lock();
  COND2 = 1;
  CV2.Signal();
  MU2.Unlock();
}
void Thr3() {
  MU.Lock();
  while(COND != 1)
    CV.Wait(&MU);
  MU.Unlock();
}
void Thr4() {
  MU2.Lock();
  while(COND2 != 1)
    CV2.Wait(&MU2);
  MU2.Unlock();
  GLOB = 2; // READ: no HB-relation between CV.Signal and CV2.Wait !
}
void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test95: TP.");
  printf("test95: TP. Check do_cv_signal/fake segment logic\n");
  MyThreadArray mta(Thr1, Thr2, Thr3, Thr4);
  mta.Start();
  mta.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 95);
}  // namespace test95

// test96: TN. tricky LockSet behaviour {{{1
// 3 threads access the same memory with three different
// locksets: {A, B}, {B, C}, {C, A}.
// These locksets have empty intersection
namespace test96 {
int     GLOB = 0;

Mutex A, B, C;

void Thread1() {
  MutexLock a(&A);
  MutexLock b(&B);
  GLOB++;
}

void Thread2() {
  MutexLock b(&B);
  MutexLock c(&C);
  GLOB++;
}

void Thread3() {
  MutexLock a(&A);
  MutexLock c(&C);
  GLOB++;
}

void Run() {
  printf("test96: FP. tricky LockSet behaviour\n");
  ANNOTATE_TRACE_MEMORY(&GLOB);
  MyThreadArray mta(Thread1, Thread2, Thread3);
  mta.Start();
  mta.Join();
  CHECK(GLOB == 3);
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 96);
}  // namespace test96

// test97: This test shows false negative with --fast-mode=yes {{{1
namespace test97 {
const int HG_CACHELINE_SIZE = 64;

Mutex MU;

const int ARRAY_SIZE = HG_CACHELINE_SIZE * 4 / sizeof(int);
int array[ARRAY_SIZE];
int * GLOB = &array[ARRAY_SIZE/2];
/*
  We use sizeof(array) == 4 * HG_CACHELINE_SIZE to be sure that GLOB points
  to a memory inside a CacheLineZ which is inside array's memory range
 */

void Reader() {
  usleep(500000);
  CHECK(777 == *GLOB);
}

void Run() {
  MyThreadArray t(Reader);
  if (!Tsan_FastMode())
    ANNOTATE_EXPECT_RACE_FOR_TSAN(GLOB, "test97: TP. FN with --fast-mode=yes");
  printf("test97: This test shows false negative with --fast-mode=yes\n");

  t.Start();
  *GLOB = 777;
  t.Join();
}

REGISTER_TEST2(Run, 97, FEATURE)
}  // namespace test97

// test98: Synchronization via read/write (or send/recv). {{{1
namespace test98 {
// The synchronization here is done by a pair of read/write calls
// that create a happens-before arc. Same may be done with send/recv.
// Such synchronization is quite unusual in real programs
// (why would one synchronizae via a file or socket?), but
// quite possible in unittests where one threads runs for producer
// and one for consumer.
//
// A race detector has to create a happens-before arcs for
// {read,send}->{write,recv} even if the file descriptors are different.
//
int     GLOB = 0;
int fd_out = -1;
int fd_in  = -1;

void Writer() {
  usleep(1000);
  GLOB = 1;
  const char *str = "Hey there!\n";
  write(fd_out, str, strlen(str) + 1);
}

void Reader() {
  char buff[100];
  while (read(fd_in, buff, 100) == 0)
    sleep(1);
  printf("read: %s\n", buff);
  GLOB = 2;
}

void Run() {
  printf("test98: negative, synchronization via I/O\n");
  char in_name[100];
  char out_name[100];
  // we open two files, on for reading and one for writing,
  // but the files are actually the same (symlinked).
  sprintf(in_name,  "/tmp/racecheck_unittest_in.%d", getpid());
  sprintf(out_name, "/tmp/racecheck_unittest_out.%d", getpid());
  fd_out = creat(out_name, O_WRONLY | S_IRWXU);
  symlink(out_name, in_name);
  fd_in  = open(in_name, 0, O_RDONLY);
  CHECK(fd_out >= 0);
  CHECK(fd_in  >= 0);
  MyThreadArray t(Writer, Reader);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
  // cleanup
  close(fd_in);
  close(fd_out);
  unlink(in_name);
  unlink(out_name);
}
REGISTER_TEST(Run, 98)
}  // namespace test98


// test99: TP. Unit test for a bug in LockWhen*. {{{1
namespace test99 {


bool GLOB = false;
Mutex mu;

static void Thread1() {
  for (int i = 0; i < 100; i++) {
    mu.LockWhenWithTimeout(Condition(&ArgIsTrue, &GLOB), 5);
    GLOB = false;
    mu.Unlock();
    usleep(10000);
  }
}

static void Thread2() {
  for (int i = 0; i < 100; i++) {
    mu.Lock();
    mu.Unlock();
    usleep(10000);
  }
}

void Run() {
  printf("test99: regression test for LockWhen*\n");
  MyThreadArray t(Thread1, Thread2);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 99);
}  // namespace test99


// test100: Test for initialization bit. {{{1
namespace test100 {
int     G1 = 0;
int     G2 = 0;
int     G3 = 0;
int     G4 = 0;

void Creator() {
  G1 = 1; CHECK(G1);
  G2 = 1;
  G3 = 1; CHECK(G3);
  G4 = 1;
}

void Worker1() {
  usleep(100000);
  CHECK(G1);
  CHECK(G2);
  G3 = 3;
  G4 = 3;
}

void Worker2() {

}


void Run() {
  printf("test100: test for initialization bit. \n");
  MyThreadArray t(Creator, Worker1, Worker2);
  ANNOTATE_TRACE_MEMORY(&G1);
  ANNOTATE_TRACE_MEMORY(&G2);
  ANNOTATE_TRACE_MEMORY(&G3);
  ANNOTATE_TRACE_MEMORY(&G4);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 100, FEATURE|EXCLUDE_FROM_ALL)
}  // namespace test100


// test101: TN. Two signals and two waits. {{{1
namespace test101 {
Mutex MU;
CondVar CV;
int     GLOB = 0;

int C1 = 0, C2 = 0;

void Signaller() {
  usleep(100000);
  MU.Lock();
  C1 = 1;
  CV.Signal();
  printf("signal\n");
  MU.Unlock();

  GLOB = 1;

  usleep(500000);
  MU.Lock();
  C2 = 1;
  CV.Signal();
  printf("signal\n");
  MU.Unlock();
}

void Waiter() {
  MU.Lock();
  while(!C1)
    CV.Wait(&MU);
  printf("wait\n");
  MU.Unlock();

  MU.Lock();
  while(!C2)
    CV.Wait(&MU);
  printf("wait\n");
  MU.Unlock();

  GLOB = 2;

}

void Run() {
  printf("test101: negative\n");
  MyThreadArray t(Waiter, Signaller);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 101)
}  // namespace test101

// test102: --fast-mode=yes vs. --initialization-bit=yes {{{1
namespace test102 {
const int HG_CACHELINE_SIZE = 64;

Mutex MU;

const int ARRAY_SIZE = HG_CACHELINE_SIZE * 4 / sizeof(int);
int array[ARRAY_SIZE + 1];
int * GLOB = &array[ARRAY_SIZE/2];
/*
  We use sizeof(array) == 4 * HG_CACHELINE_SIZE to be sure that GLOB points
  to a memory inside a CacheLineZ which is inside array's memory range
*/

void Reader() {
  usleep(200000);
  CHECK(777 == GLOB[0]);
  usleep(400000);
  CHECK(777 == GLOB[1]);
}

void Run() {
  MyThreadArray t(Reader);
  if (!Tsan_FastMode())
    ANNOTATE_EXPECT_RACE_FOR_TSAN(GLOB+0, "test102: TP. FN with --fast-mode=yes");
  ANNOTATE_EXPECT_RACE_FOR_TSAN(GLOB+1, "test102: TP");
  printf("test102: --fast-mode=yes vs. --initialization-bit=yes\n");

  t.Start();
  GLOB[0] = 777;
  usleep(400000);
  GLOB[1] = 777;
  t.Join();
}

REGISTER_TEST2(Run, 102, FEATURE)
}  // namespace test102

// test103: Access different memory locations with different LockSets {{{1
namespace test103 {
const int N_MUTEXES = 6;
const int LOCKSET_INTERSECTION_SIZE = 3;

int data[1 << LOCKSET_INTERSECTION_SIZE] = {0};
Mutex MU[N_MUTEXES];

inline int LS_to_idx (int ls) {
  return (ls >> (N_MUTEXES - LOCKSET_INTERSECTION_SIZE))
      & ((1 << LOCKSET_INTERSECTION_SIZE) - 1);
}

void Worker() {
  for (int ls = 0; ls < (1 << N_MUTEXES); ls++) {
    if (LS_to_idx(ls) == 0)
      continue;
    for (int m = 0; m < N_MUTEXES; m++)
      if (ls & (1 << m))
        MU[m].Lock();

    data[LS_to_idx(ls)]++;

    for (int m = N_MUTEXES - 1; m >= 0; m--)
      if (ls & (1 << m))
        MU[m].Unlock();
  }
}

void Run() {
  printf("test103: Access different memory locations with different LockSets\n");
  MyThreadArray t(Worker, Worker, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 103, FEATURE)
}  // namespace test103

// test104: TP. Simple race (write vs write). Heap mem. {{{1
namespace test104 {
int     *GLOB = NULL;
void Worker() {
  *GLOB = 1;
}

void Parent() {
  MyThread t(Worker);
  t.Start();
  usleep(100000);
  *GLOB = 2;
  t.Join();
}
void Run() {
  GLOB = (int*)memalign(64, sizeof(int));
  *GLOB = 0;
  ANNOTATE_EXPECT_RACE(GLOB, "test104. TP.");
  ANNOTATE_TRACE_MEMORY(GLOB);
  printf("test104: positive\n");
  Parent();
  printf("\tGLOB=%d\n", *GLOB);
  free(GLOB);
}
REGISTER_TEST(Run, 104);
}  // namespace test104


// test105: Checks how stack grows. {{{1
namespace test105 {
int     GLOB = 0;

void F1() {
  int ar[32];
//  ANNOTATE_TRACE_MEMORY(&ar[0]);
//  ANNOTATE_TRACE_MEMORY(&ar[31]);
  ar[0] = 1;
  ar[31] = 1;
}

void Worker() {
  int ar[32];
//  ANNOTATE_TRACE_MEMORY(&ar[0]);
//  ANNOTATE_TRACE_MEMORY(&ar[31]);
  ar[0] = 1;
  ar[31] = 1;
  F1();
}

void Run() {
  printf("test105: negative\n");
  Worker();
  MyThread t(Worker);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 105)
}  // namespace test105


// test106: TN. pthread_once. {{{1
namespace test106 {
int     *GLOB = NULL;
static pthread_once_t once = PTHREAD_ONCE_INIT;
void Init() {
  GLOB = new int;
  ANNOTATE_TRACE_MEMORY(GLOB);
  *GLOB = 777;
}

void Worker0() {
  pthread_once(&once, Init);
}
void Worker1() {
  usleep(100000);
  pthread_once(&once, Init);
  CHECK(*GLOB == 777);
}


void Run() {
  printf("test106: negative\n");
  MyThreadArray t(Worker0, Worker1, Worker1, Worker1);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", *GLOB);
}
REGISTER_TEST2(Run, 106, FEATURE)
}  // namespace test106


// test107: Test for ANNOTATE_EXPECT_RACE {{{1
namespace test107 {
int     GLOB = 0;
void Run() {
  printf("test107: negative\n");
  ANNOTATE_EXPECT_RACE(&GLOB, "No race in fact. Just checking the tool.");
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 107, FEATURE|EXCLUDE_FROM_ALL)
}  // namespace test107


// test108: TN. initialization of static object. {{{1
namespace test108 {
// Here we have a function-level static object.
// Starting from gcc 4 this is therad safe,
// but is is not thread safe with many other compilers.
//
// Helgrind supports this kind of initialization by
// intercepting __cxa_guard_acquire/__cxa_guard_release
// and ignoring all accesses between them.
// Helgrind also intercepts pthread_once in the same manner.
class Foo {
 public:
  Foo() {
    ANNOTATE_TRACE_MEMORY(&a_);
    a_ = 42;
  }
  void Check() const { CHECK(a_ == 42); }
 private:
  int a_;
};

const Foo *GetFoo() {
  static const Foo *foo = new Foo();
  return foo;
}
void Worker0() {
  GetFoo();
}

void Worker() {
  usleep(200000);
  const Foo *foo = GetFoo();
  foo->Check();
}


void Run() {
  printf("test108: negative, initialization of static object\n");
  MyThreadArray t(Worker0, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 108, FEATURE)
}  // namespace test108


// test109: TN. Checking happens before between parent and child threads. {{{1
namespace test109 {
// Check that the detector correctly connects
//   pthread_create with the new thread
// and
//   thread exit with pthread_join
const int N = 32;
static int GLOB[N];

void Worker(void *a) {
  usleep(10000);
//  printf("--Worker : %ld %p\n", (int*)a - GLOB, (void*)pthread_self());
  int *arg = (int*)a;
  (*arg)++;
}

void Run() {
  printf("test109: negative\n");
  MyThread *t[N];
  for (int i  = 0; i < N; i++) {
    t[i] = new MyThread(Worker, &GLOB[i]);
  }
  for (int i  = 0; i < N; i++) {
    ANNOTATE_TRACE_MEMORY(&GLOB[i]);
    GLOB[i] = 1;
    t[i]->Start();
//    printf("--Started: %p\n", (void*)t[i]->tid());
  }
  for (int i  = 0; i < N; i++) {
//    printf("--Joining: %p\n", (void*)t[i]->tid());
    t[i]->Join();
//    printf("--Joined : %p\n", (void*)t[i]->tid());
    GLOB[i]++;
  }
  for (int i  = 0; i < N; i++) delete t[i];

  printf("\tGLOB=%d\n", GLOB[13]);
}
REGISTER_TEST(Run, 109)
}  // namespace test109


// test110: TP. Simple races with stack, global and heap objects. {{{1
namespace test110 {
int        GLOB = 0;
static int STATIC;

int       *STACK = 0;

int       *MALLOC;
int       *CALLOC;
int       *REALLOC;
int       *VALLOC;
int       *PVALLOC;
int       *MEMALIGN;
int       *POSIX_MEMALIGN;
int       *MMAP;

int       *NEW;
int       *NEW_ARR;

void Worker() {
  GLOB++;
  STATIC++;

  (*STACK)++;

  (*MALLOC)++;
  (*CALLOC)++;
  (*REALLOC)++;
  (*VALLOC)++;
  (*PVALLOC)++;
  (*MEMALIGN)++;
  (*POSIX_MEMALIGN)++;
  (*MMAP)++;

  (*NEW)++;
  (*NEW_ARR)++;
}
void Run() {
  int x = 0;
  STACK = &x;

  MALLOC = (int*)malloc(sizeof(int));
  CALLOC = (int*)calloc(1, sizeof(int));
  REALLOC = (int*)realloc(NULL, sizeof(int));
  VALLOC = (int*)valloc(sizeof(int));
  PVALLOC = (int*)valloc(sizeof(int));  // TODO: pvalloc breaks helgrind.
  MEMALIGN = (int*)memalign(64, sizeof(int));
  CHECK(0 == posix_memalign((void**)&POSIX_MEMALIGN, 64, sizeof(int)));
  MMAP = (int*)mmap(NULL, sizeof(int), PROT_READ | PROT_WRITE,
                    MAP_PRIVATE | MAP_ANON, -1, 0);

  NEW     = new int;
  NEW_ARR = new int[10];


  FAST_MODE_INIT(STACK);
  ANNOTATE_EXPECT_RACE(STACK, "real race on stack object");
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE(&GLOB, "real race on global object");
  FAST_MODE_INIT(&STATIC);
  ANNOTATE_EXPECT_RACE(&STATIC, "real race on a static global object");
  FAST_MODE_INIT(MALLOC);
  ANNOTATE_EXPECT_RACE(MALLOC, "real race on a malloc-ed object");
  FAST_MODE_INIT(CALLOC);
  ANNOTATE_EXPECT_RACE(CALLOC, "real race on a calloc-ed object");
  FAST_MODE_INIT(REALLOC);
  ANNOTATE_EXPECT_RACE(REALLOC, "real race on a realloc-ed object");
  FAST_MODE_INIT(VALLOC);
  ANNOTATE_EXPECT_RACE(VALLOC, "real race on a valloc-ed object");
  FAST_MODE_INIT(PVALLOC);
  ANNOTATE_EXPECT_RACE(PVALLOC, "real race on a pvalloc-ed object");
  FAST_MODE_INIT(MEMALIGN);
  ANNOTATE_EXPECT_RACE(MEMALIGN, "real race on a memalign-ed object");
  FAST_MODE_INIT(POSIX_MEMALIGN);
  ANNOTATE_EXPECT_RACE(POSIX_MEMALIGN, "real race on a posix_memalign-ed object");
  FAST_MODE_INIT(MMAP);
  ANNOTATE_EXPECT_RACE(MMAP, "real race on a mmap-ed object");

  FAST_MODE_INIT(NEW);
  ANNOTATE_EXPECT_RACE(NEW, "real race on a new-ed object");
  FAST_MODE_INIT(NEW_ARR);
  ANNOTATE_EXPECT_RACE(NEW_ARR, "real race on a new[]-ed object");

  MyThreadArray t(Worker, Worker, Worker);
  t.Start();
  t.Join();
  printf("test110: positive (race on a stack object)\n");
  printf("\tSTACK=%d\n", *STACK);
  CHECK(GLOB <= 3);
  CHECK(STATIC <= 3);

  free(MALLOC);
  free(CALLOC);
  free(REALLOC);
  free(VALLOC);
  free(PVALLOC);
  free(MEMALIGN);
  free(POSIX_MEMALIGN);
  munmap(MMAP, sizeof(int));
  delete NEW;
  delete [] NEW_ARR;
}
REGISTER_TEST(Run, 110)
}  // namespace test110


// test111: TN. Unit test for a bug related to stack handling. {{{1
namespace test111 {
char     *GLOB = 0;
bool COND = false;
Mutex mu;
const int N = 3000;

void write_to_p(char *p, int val) {
  for (int i = 0; i < N; i++)
    p[i] = val;
}

static bool ArgIsTrue(bool *arg) {
//  printf("ArgIsTrue: %d tid=%d\n", *arg, (int)pthread_self());
  return *arg == true;
}

void f1() {
  char some_stack[N];
  write_to_p(some_stack, 1);
  mu.LockWhen(Condition(&ArgIsTrue, &COND));
  mu.Unlock();
}

void f2() {
  char some_stack[N];
  char some_more_stack[N];
  write_to_p(some_stack, 2);
  write_to_p(some_more_stack, 2);
}

void f0() { f2(); }

void Worker1() {
  f0();
  f1();
  f2();
}

void Worker2() {
  usleep(100000);
  mu.Lock();
  COND = true;
  mu.Unlock();
}

void Run() {
  printf("test111: regression test\n");
  MyThreadArray t(Worker1, Worker1, Worker2);
//  AnnotateSetVerbosity(__FILE__, __LINE__, 3);
  t.Start();
  t.Join();
//  AnnotateSetVerbosity(__FILE__, __LINE__, 1);
}
REGISTER_TEST2(Run, 111, FEATURE)
}  // namespace test111

// test112: STAB. Test for ANNOTATE_PUBLISH_MEMORY_RANGE{{{1
namespace test112 {
char     *GLOB = 0;
const int N = 64 * 5;
Mutex mu;
bool ready = false; // under mu
int beg, end; // under mu

Mutex mu1;

void Worker() {

  bool is_ready = false;
  int b, e;
  while (!is_ready) {
    mu.Lock();
    is_ready = ready;
    b = beg;
    e = end;
    mu.Unlock();
    usleep(1000);
  }

  mu1.Lock();
  for (int i = b; i < e; i++) {
    GLOB[i]++;
  }
  mu1.Unlock();
}

void PublishRange(int b, int e) {
  MyThreadArray t(Worker, Worker);
  ready = false; // runs before other threads
  t.Start();

  ANNOTATE_NEW_MEMORY(GLOB + b, e - b);
  ANNOTATE_TRACE_MEMORY(GLOB + b);
  for (int j = b; j < e; j++) {
    GLOB[j] = 0;
  }
  ANNOTATE_PUBLISH_MEMORY_RANGE(GLOB + b, e - b);

  // hand off
  mu.Lock();
  ready = true;
  beg = b;
  end = e;
  mu.Unlock();

  t.Join();
}

void Run() {
  printf("test112: stability (ANNOTATE_PUBLISH_MEMORY_RANGE)\n");
  GLOB = new char [N];

  PublishRange(0, 10);
  PublishRange(3, 5);

  PublishRange(12, 13);
  PublishRange(10, 14);

  PublishRange(15, 17);
  PublishRange(16, 18);

  // do few more random publishes.
  for (int i = 0; i < 20; i++) {
    int beg = rand() % N;
    int size = (rand() % (N - beg)) + 1;
    CHECK(size > 0);
    CHECK(beg + size <= N);
    PublishRange(beg, beg + size);
  }

  printf("GLOB = %d\n", (int)GLOB[0]);
}
REGISTER_TEST2(Run, 112, STABILITY)
}  // namespace test112


// test113: PERF. A lot of lock/unlock calls. Many locks {{{1
namespace    test113 {
const int kNumIter = 100000;
const int kNumLocks = 7;
Mutex   MU[kNumLocks];
void Run() {
  printf("test113: perf\n");
  for (int i = 0; i < kNumIter; i++ ) {
    for (int j = 0; j < kNumLocks; j++) {
      if (i & (1 << j)) MU[j].Lock();
    }
    for (int j = kNumLocks - 1; j >= 0; j--) {
      if (i & (1 << j)) MU[j].Unlock();
    }
  }
}
REGISTER_TEST(Run, 113)
}  // namespace test113


// test114: STAB. Recursive lock. {{{1
namespace    test114 {
int Bar() {
  static int bar = 1;
  return bar;
}
int Foo() {
  static int foo = Bar();
  return foo;
}
void Worker() {
  static int x = Foo();
  CHECK(x == 1);
}
void Run() {
  printf("test114: stab\n");
  MyThreadArray t(Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 114)
}  // namespace test114


// test115: TN. sem_open. {{{1
namespace    test115 {
int tid = 0;
Mutex mu;
const char *kSemName = "drt-test-sem";

int GLOB = 0;

sem_t *DoSemOpen() {
  // TODO: there is some race report inside sem_open
  // for which suppressions do not work... (???)
  ANNOTATE_IGNORE_WRITES_BEGIN();
  sem_t *sem = sem_open(kSemName, O_CREAT, 0600, 3);
  ANNOTATE_IGNORE_WRITES_END();
  return sem;
}

void Worker() {
  mu.Lock();
  int my_tid = tid++;
  mu.Unlock();

  if (my_tid == 0) {
    GLOB = 1;
  }

  // if the detector observes a happens-before arc between
  // sem_open and sem_wait, it will be silent.
  sem_t *sem = DoSemOpen();
  usleep(100000);
  CHECK(sem != SEM_FAILED);
  CHECK(sem_wait(sem) == 0);

  if (my_tid > 0) {
    CHECK(GLOB == 1);
  }
}

void Run() {
  printf("test115: stab (sem_open())\n");

  // just check that sem_open is not completely broken
  sem_unlink(kSemName);
  sem_t* sem = DoSemOpen();
  CHECK(sem != SEM_FAILED);
  CHECK(sem_wait(sem) == 0);
  sem_unlink(kSemName);

  // check that sem_open and sem_wait create a happens-before arc.
  MyThreadArray t(Worker, Worker, Worker);
  t.Start();
  t.Join();
  // clean up
  sem_unlink(kSemName);
}
REGISTER_TEST(Run, 115)
}  // namespace test115


// test116: TN. some operations with string<> objects. {{{1
namespace test116 {

void Worker() {
  string A[10], B[10], C[10];
  for (int i = 0; i < 1000; i++) {
    for (int j = 0; j < 10; j++) {
      string &a = A[j];
      string &b = B[j];
      string &c = C[j];
      a = "sdl;fkjhasdflksj df";
      b = "sdf sdf;ljsd ";
      c = "'sfdf df";
      c = b;
      a = c;
      b = a;
      swap(a,b);
      swap(b,c);
    }
    for (int j = 0; j < 10; j++) {
      string &a = A[j];
      string &b = B[j];
      string &c = C[j];
      a.clear();
      b.clear();
      c.clear();
    }
  }
}

void Run() {
  printf("test116: negative (strings)\n");
  MyThreadArray t(Worker, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 116, FEATURE|EXCLUDE_FROM_ALL)
}  // namespace test116

// test117: TN. Many calls to function-scope static init. {{{1
namespace test117 {
const int N = 50;

int Foo() {
  usleep(20000);
  return 1;
}

void Worker(void *a) {
  static int foo = Foo();
  CHECK(foo == 1);
}

void Run() {
  printf("test117: negative\n");
  MyThread *t[N];
  for (int i  = 0; i < N; i++) {
    t[i] = new MyThread(Worker);
  }
  for (int i  = 0; i < N; i++) {
    t[i]->Start();
  }
  for (int i  = 0; i < N; i++) {
    t[i]->Join();
  }
  for (int i  = 0; i < N; i++) delete t[i];
}
REGISTER_TEST(Run, 117)
}  // namespace test117


// test118 PERF: One signal, multiple waits. {{{1
namespace   test118 {
int     GLOB = 0;
const int kNumIter = 2000000;
void Signaller() {
  usleep(50000);
  ANNOTATE_CONDVAR_SIGNAL(&GLOB);
}
void Waiter() {
  for (int i = 0; i < kNumIter; i++) {
    ANNOTATE_CONDVAR_WAIT(&GLOB);
    if (i == kNumIter / 2)
      usleep(100000);
  }
}
void Run() {
  printf("test118: perf\n");
  MyThreadArray t(Signaller, Waiter, Signaller, Waiter);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 118)
}  // namespace test118


// test119: TP. Testing that malloc does not introduce any HB arc. {{{1
namespace test119 {
int     GLOB = 0;
void Worker1() {
  GLOB = 1;
  free(malloc(123));
}
void Worker2() {
  usleep(100000);
  free(malloc(345));
  GLOB = 2;
}
void Run() {
  printf("test119: positive (checking if malloc creates HB arcs)\n");
  FAST_MODE_INIT(&GLOB);
  if (!(Tsan_PureHappensBefore() && kMallocUsesMutex))
    ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "true race");
  MyThreadArray t(Worker1, Worker2);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 119)
}  // namespace test119


// test120: TP. Thread1: write then read. Thread2: read. {{{1
namespace test120 {
int     GLOB = 0;

void Thread1() {
  GLOB = 1;           // write
  CHECK(GLOB);        // read
}

void Thread2() {
  usleep(100000);
  CHECK(GLOB >= 0);   // read
}

void Run() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "TP (T1: write then read, T2: read)");
  printf("test120: positive\n");
  MyThreadArray t(Thread1, Thread2);
  GLOB = 1;
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 120)
}  // namespace test120


// test121: TP. Example of double-checked-locking  {{{1
namespace test121 {
struct Foo {
  uintptr_t a, b[15];
} __attribute__ ((aligned (64)));

static Mutex mu;
static Foo  *foo;

void InitMe() {
  if (!foo) {
    MutexLock lock(&mu);
    if (!foo) {
      ANNOTATE_EXPECT_RACE_FOR_TSAN(&foo, "test121. Double-checked locking (ptr)");
      foo = new Foo;
      if (!Tsan_FastMode())
        ANNOTATE_EXPECT_RACE_FOR_TSAN(&foo->a, "test121. Double-checked locking (obj)");
      foo->a = 42;
    }
  }
}

void UseMe() {
  InitMe();
  CHECK(foo && foo->a == 42);
}

void Worker1() { UseMe(); }
void Worker2() { UseMe(); }
void Worker3() { UseMe(); }


void Run() {
  FAST_MODE_INIT(&foo);
  printf("test121: TP. Example of double-checked-locking\n");
  MyThreadArray t1(Worker1, Worker2, Worker3);
  t1.Start();
  t1.Join();
  delete foo;
}
REGISTER_TEST(Run, 121)
}  // namespace test121

// test122 TP: Simple test with RWLock {{{1
namespace  test122 {
int     VAR1 = 0;
int     VAR2 = 0;
RWLock mu;

void WriteWhileHoldingReaderLock(int *p) {
  usleep(100000);
  ReaderLockScoped lock(&mu);  // Reader lock for writing. -- bug.
  (*p)++;
}

void CorrectWrite(int *p) {
  WriterLockScoped lock(&mu);
  (*p)++;
}

void Thread1() { WriteWhileHoldingReaderLock(&VAR1); }
void Thread2() { CorrectWrite(&VAR1); }
void Thread3() { CorrectWrite(&VAR2); }
void Thread4() { WriteWhileHoldingReaderLock(&VAR2); }


void Run() {
  printf("test122: positive (rw-lock)\n");
  VAR1 = 0;
  VAR2 = 0;
  ANNOTATE_TRACE_MEMORY(&VAR1);
  ANNOTATE_TRACE_MEMORY(&VAR2);
  if (!Tsan_PureHappensBefore()) {
    ANNOTATE_EXPECT_RACE_FOR_TSAN(&VAR1, "test122. TP. ReaderLock-ed while writing");
    ANNOTATE_EXPECT_RACE_FOR_TSAN(&VAR2, "test122. TP. ReaderLock-ed while writing");
  }
  MyThreadArray t(Thread1, Thread2, Thread3, Thread4);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 122)
}  // namespace test122


// test123 TP: accesses of different sizes. {{{1
namespace test123 {

uint64_t MEM[8];

#define GenericWrite(p,size,off) { \
  if (size == 64) {\
    CHECK(off == 0);\
    (p)[off] = 1;\
  } else if (size == 32) {\
    CHECK(off <= 2);\
    uint32_t *x = (uint32_t*)(p);\
    x[off] = 1;\
  } else if (size == 16) {\
    CHECK(off <= 4);\
    uint16_t *x = (uint16_t*)(p);\
    x[off] = 1;\
  } else if (size == 8) {\
    CHECK(off <= 8);\
    uint8_t *x = (uint8_t*)(p);\
    x[off] = 1;\
  } else {\
    CHECK(0);\
  }\
}\

// Q. Hey dude, why so many functions?
// A. I need different stack traces for different accesses.

void Wr64_0() { GenericWrite(&MEM[0], 64, 0); }
void Wr64_1() { GenericWrite(&MEM[1], 64, 0); }
void Wr64_2() { GenericWrite(&MEM[2], 64, 0); }
void Wr64_3() { GenericWrite(&MEM[3], 64, 0); }
void Wr64_4() { GenericWrite(&MEM[4], 64, 0); }
void Wr64_5() { GenericWrite(&MEM[5], 64, 0); }
void Wr64_6() { GenericWrite(&MEM[6], 64, 0); }
void Wr64_7() { GenericWrite(&MEM[7], 64, 0); }

void Wr32_0() { GenericWrite(&MEM[0], 32, 0); }
void Wr32_1() { GenericWrite(&MEM[1], 32, 1); }
void Wr32_2() { GenericWrite(&MEM[2], 32, 0); }
void Wr32_3() { GenericWrite(&MEM[3], 32, 1); }
void Wr32_4() { GenericWrite(&MEM[4], 32, 0); }
void Wr32_5() { GenericWrite(&MEM[5], 32, 1); }
void Wr32_6() { GenericWrite(&MEM[6], 32, 0); }
void Wr32_7() { GenericWrite(&MEM[7], 32, 1); }

void Wr16_0() { GenericWrite(&MEM[0], 16, 0); }
void Wr16_1() { GenericWrite(&MEM[1], 16, 1); }
void Wr16_2() { GenericWrite(&MEM[2], 16, 2); }
void Wr16_3() { GenericWrite(&MEM[3], 16, 3); }
void Wr16_4() { GenericWrite(&MEM[4], 16, 0); }
void Wr16_5() { GenericWrite(&MEM[5], 16, 1); }
void Wr16_6() { GenericWrite(&MEM[6], 16, 2); }
void Wr16_7() { GenericWrite(&MEM[7], 16, 3); }

void Wr8_0() { GenericWrite(&MEM[0], 8, 0); }
void Wr8_1() { GenericWrite(&MEM[1], 8, 1); }
void Wr8_2() { GenericWrite(&MEM[2], 8, 2); }
void Wr8_3() { GenericWrite(&MEM[3], 8, 3); }
void Wr8_4() { GenericWrite(&MEM[4], 8, 4); }
void Wr8_5() { GenericWrite(&MEM[5], 8, 5); }
void Wr8_6() { GenericWrite(&MEM[6], 8, 6); }
void Wr8_7() { GenericWrite(&MEM[7], 8, 7); }

void WriteAll64() {
  Wr64_0();
  Wr64_1();
  Wr64_2();
  Wr64_3();
  Wr64_4();
  Wr64_5();
  Wr64_6();
  Wr64_7();
}

void WriteAll32() {
  Wr32_0();
  Wr32_1();
  Wr32_2();
  Wr32_3();
  Wr32_4();
  Wr32_5();
  Wr32_6();
  Wr32_7();
}

void WriteAll16() {
  Wr16_0();
  Wr16_1();
  Wr16_2();
  Wr16_3();
  Wr16_4();
  Wr16_5();
  Wr16_6();
  Wr16_7();
}

void WriteAll8() {
  Wr8_0();
  Wr8_1();
  Wr8_2();
  Wr8_3();
  Wr8_4();
  Wr8_5();
  Wr8_6();
  Wr8_7();
}

void W00() { WriteAll64(); }
void W01() { WriteAll64(); }
void W02() { WriteAll64(); }

void W10() { WriteAll32(); }
void W11() { WriteAll32(); }
void W12() { WriteAll32(); }

void W20() { WriteAll16(); }
void W21() { WriteAll16(); }
void W22() { WriteAll16(); }

void W30() { WriteAll8(); }
void W31() { WriteAll8(); }
void W32() { WriteAll8(); }

typedef void (*F)(void);

void TestTwoSizes(F f1, F f2) {
  // first f1, then f2
  ANNOTATE_NEW_MEMORY(MEM, sizeof(MEM));
  memset(MEM, 0, sizeof(MEM));
  MyThreadArray t1(f1, f2);
  t1.Start();
  t1.Join();
  // reverse order
  ANNOTATE_NEW_MEMORY(MEM, sizeof(MEM));
  memset(MEM, 0, sizeof(MEM));
  MyThreadArray t2(f2, f1);
  t2.Start();
  t2.Join();
}

void Run() {
  printf("test123: positive (different sizes)\n");
  TestTwoSizes(W00, W10);
//  TestTwoSizes(W01, W20);
//  TestTwoSizes(W02, W30);
//  TestTwoSizes(W11, W21);
//  TestTwoSizes(W12, W31);
//  TestTwoSizes(W22, W32);

}
REGISTER_TEST2(Run, 123, FEATURE|EXCLUDE_FROM_ALL)
}  // namespace test123


// test124: What happens if we delete an unlocked lock? {{{1
namespace test124 {
// This test does not worg with pthreads (you can't call
// pthread_mutex_destroy on a locked lock).
int     GLOB = 0;
const int N = 1000;
void Worker() {
  Mutex *a_large_local_array_of_mutexes;
  a_large_local_array_of_mutexes = new Mutex[N];
  for (int i = 0; i < N; i++) {
    a_large_local_array_of_mutexes[i].Lock();
  }
  delete []a_large_local_array_of_mutexes;
  GLOB = 1;
}

void Run() {
  printf("test124: negative\n");
  MyThreadArray t(Worker, Worker, Worker);
  t.Start();
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 124, FEATURE|EXCLUDE_FROM_ALL)
}  // namespace test124


// test125 TN: Backwards lock (annotated). {{{1
namespace test125 {
// This test uses "Backwards mutex" locking protocol.
// We take a *reader* lock when writing to a per-thread data
// (GLOB[thread_num])  and we take a *writer* lock when we
// are reading from the entire array at once.
//
// Such locking protocol is not understood by ThreadSanitizer's
// hybrid state machine. So, you either have to use a pure-happens-before
// detector ("tsan --pure-happens-before") or apply pure happens-before mode
// to this particular lock by using ANNOTATE_MUTEX_IS_USED_AS_CONDVAR(&mu).

const int n_threads = 3;
RWLock   mu;
int     GLOB[n_threads];

int adder_num; // updated atomically.

void Adder() {
  int my_num = AtomicIncrement(&adder_num, 1);

  ReaderLockScoped lock(&mu);
  GLOB[my_num]++;
}

void Aggregator() {
  int sum = 0;
  {
    WriterLockScoped lock(&mu);
    for (int i = 0; i < n_threads; i++) {
      sum += GLOB[i];
    }
  }
  printf("sum=%d\n", sum);
}

void Run() {
  printf("test125: negative\n");

  ANNOTATE_MUTEX_IS_USED_AS_CONDVAR(&mu);

  // run Adders, then Aggregator
  {
    MyThreadArray t(Adder, Adder, Adder, Aggregator);
    t.Start();
    t.Join();
  }

  // Run Aggregator first.
  adder_num = 0;
  {
    MyThreadArray t(Aggregator, Adder, Adder, Adder);
    t.Start();
    t.Join();
  }

}
REGISTER_TEST(Run, 125)
}  // namespace test125

// test126 TN: test for BlockingCounter {{{1
namespace  test126 {
BlockingCounter *blocking_counter;
int     GLOB = 0;
void Worker() {
  CHECK(blocking_counter);
  CHECK(GLOB == 0);
  blocking_counter->DecrementCount();
}
void Run() {
  printf("test126: negative\n");
  MyThreadArray t(Worker, Worker, Worker);
  blocking_counter = new BlockingCounter(3);
  t.Start();
  blocking_counter->Wait();
  GLOB = 1;
  t.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST(Run, 126)
}  // namespace test126


// test127. Bad code: unlocking a mutex locked by another thread. {{{1
namespace test127 {
Mutex mu;
void Thread1() {
  mu.Lock();
}
void Thread2() {
  usleep(100000);
  mu.Unlock();
}
void Run() {
  printf("test127: unlocking a mutex locked by another thread.\n");
  MyThreadArray t(Thread1, Thread2);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 127)
}  // namespace test127

// test128. Suppressed code in concurrent accesses {{{1
// Please use --suppressions=unittest.supp flag when running this test.
namespace test128 {
Mutex mu;
int GLOB = 0;
void Worker() {
  usleep(100000);
  mu.Lock();
  GLOB++;
  mu.Unlock();
}
void ThisFunctionShouldBeSuppressed() {
  GLOB++;
}
void Run() {
  printf("test128: Suppressed code in concurrent accesses.\n");
  MyThreadArray t(Worker, ThisFunctionShouldBeSuppressed);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 128, FEATURE | EXCLUDE_FROM_ALL)
}  // namespace test128

// test129: TN. Synchronization via ReaderLockWhen(). {{{1
namespace test129 {
int     GLOB = 0;
Mutex   MU;
bool WeirdCondition(int* param) {
  *param = GLOB;  // a write into Waiter's memory
  return GLOB > 0;
}
void Waiter() {
  int param = 0;
  MU.ReaderLockWhen(Condition(WeirdCondition, &param));
  MU.ReaderUnlock();
  CHECK(GLOB > 0);
  CHECK(param > 0);
}
void Waker() {
  usleep(100000);  // Make sure the waiter blocks.
  MU.Lock();
  GLOB++;
  MU.Unlock();     // calls ANNOTATE_CONDVAR_SIGNAL;
}
void Run() {
  printf("test129: Synchronization via ReaderLockWhen()\n");
  MyThread mt(Waiter, NULL, "Waiter Thread");
  mt.Start();
  Waker();
  mt.Join();
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 129, FEATURE);
}  // namespace test129

// test130: TN. Per-thread. {{{1
namespace test130 {
#ifndef NO_TLS
// This test verifies that the race detector handles
// thread-local storage (TLS) correctly.
// As of 09-03-30 ThreadSanitizer has a bug:
//   - Thread1 starts
//   - Thread1 touches per_thread_global
//   - Thread1 ends
//   - Thread2 starts (and there is no happens-before relation between it and
//   Thread1)
//   - Thread2 touches per_thread_global
// It may happen so that Thread2 will have per_thread_global in the same address
// as Thread1. Since there is no happens-before relation between threads,
// ThreadSanitizer reports a race.
//
// test131 does the same for stack.

static __thread int per_thread_global[10] = {0};

void RealWorker() {  // Touch per_thread_global.
  per_thread_global[1]++;
  errno++;
}

void Worker() {  // Spawn few threads that touch per_thread_global.
  MyThreadArray t(RealWorker, RealWorker);
  t.Start();
  t.Join();
}
void Worker0() { sleep(0); Worker(); }
void Worker1() { sleep(1); Worker(); }
void Worker2() { sleep(2); Worker(); }
void Worker3() { sleep(3); Worker(); }

void Run() {
  printf("test130: Per-thread\n");
  MyThreadArray t1(Worker0, Worker1, Worker2, Worker3);
  t1.Start();
  t1.Join();
  printf("\tper_thread_global=%d\n", per_thread_global[1]);
}
REGISTER_TEST(Run, 130)
#endif // NO_TLS
}  // namespace test130


// test131: TN. Stack. {{{1
namespace test131 {
// Same as test130, but for stack.

void RealWorker() {  // Touch stack.
  int stack_var = 0;
  stack_var++;
}

void Worker() {  // Spawn few threads that touch stack.
  MyThreadArray t(RealWorker, RealWorker);
  t.Start();
  t.Join();
}
void Worker0() { sleep(0); Worker(); }
void Worker1() { sleep(1); Worker(); }
void Worker2() { sleep(2); Worker(); }
void Worker3() { sleep(3); Worker(); }

void Run() {
  printf("test131: stack\n");
  MyThreadArray t(Worker0, Worker1, Worker2, Worker3);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 131)
}  // namespace test131


// test132: TP. Simple race (write vs write). Works in fast-mode. {{{1
namespace test132 {
int     GLOB = 0;
void Worker() { GLOB = 1; }

void Run1() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test132");
  printf("test132: positive; &GLOB=%p\n", &GLOB);
  ANNOTATE_TRACE_MEMORY(&GLOB);
  GLOB = 7;
  MyThreadArray t(Worker, Worker);
  t.Start();
  t.Join();
}

void Run() {
  Run1();
}
REGISTER_TEST(Run, 132);
}  // namespace test132


// test133: TP. Simple race (write vs write). Works in fast mode. {{{1
namespace test133 {
// Same as test132, but everything is run from a separate thread spawned from
// the main thread.
int     GLOB = 0;
void Worker() { GLOB = 1; }

void Run1() {
  FAST_MODE_INIT(&GLOB);
  ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "test133");
  printf("test133: positive; &GLOB=%p\n", &GLOB);
  ANNOTATE_TRACE_MEMORY(&GLOB);
  GLOB = 7;
  MyThreadArray t(Worker, Worker);
  t.Start();
  t.Join();
}
void Run() {
  MyThread t(Run1);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 133);
}  // namespace test133


// test134 TN. Swap. Variant of test79. {{{1
namespace test134 {
#if 0
typedef __gnu_cxx::hash_map<int, int> map_t;
#else
typedef std::map<int, int> map_t;
#endif
map_t   map;
Mutex   mu;
// Here we use swap to pass map between threads.
// The synchronization is correct, but w/o the annotation
// any hybrid detector will complain.

// Swap is very unfriendly to the lock-set (and hybrid) race detectors.
// Since tmp is destructed outside the mutex, we need to have a happens-before
// arc between any prior access to map and here.
// Since the internals of tmp are created ouside the mutex and are passed to
// other thread, we need to have a h-b arc between here and any future access.
// These arcs can be created by HAPPENS_{BEFORE,AFTER} annotations, but it is
// much simpler to apply pure-happens-before mode to the mutex mu.
void Swapper() {
  map_t tmp;
  MutexLock lock(&mu);
  ANNOTATE_HAPPENS_AFTER(&map);
  // We swap the new empty map 'tmp' with 'map'.
  map.swap(tmp);
  ANNOTATE_HAPPENS_BEFORE(&map);
  // tmp (which is the old version of map) is destroyed here.
}

void Worker() {
  MutexLock lock(&mu);
  ANNOTATE_HAPPENS_AFTER(&map);
  map[1]++;
  ANNOTATE_HAPPENS_BEFORE(&map);
}

void Run() {
  printf("test134: negative (swap)\n");
  // ********************** Shorter way: ***********************
  // ANNOTATE_MUTEX_IS_USED_AS_CONDVAR(&mu);
  MyThreadArray t(Worker, Worker, Swapper, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 134)
}  // namespace test134

// test135 TN. Swap. Variant of test79. {{{1
namespace test135 {

void SubWorker() {
  const long SIZE = 65536;
  for (int i = 0; i < 32; i++) {
    int *ptr = (int*)mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
                          MAP_PRIVATE | MAP_ANON, -1, 0);
    *ptr = 42;
    munmap(ptr, SIZE);
  }
}

void Worker() {
  MyThreadArray t(SubWorker, SubWorker, SubWorker, SubWorker);
  t.Start();
  t.Join();
}

void Run() {
  printf("test135: negative (mmap)\n");
  MyThreadArray t(Worker, Worker, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 135)
}  // namespace test135

// test136. Unlock twice. {{{1
namespace test136 {
void Run() {
  printf("test136: unlock twice\n");
  pthread_mutexattr_t attr;
  CHECK(0 == pthread_mutexattr_init(&attr));
  CHECK(0 == pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_ERRORCHECK));

  pthread_mutex_t mu;
  CHECK(0 == pthread_mutex_init(&mu, &attr));
  CHECK(0 == pthread_mutex_lock(&mu));
  CHECK(0 == pthread_mutex_unlock(&mu));
  int ret_unlock = pthread_mutex_unlock(&mu);  // unlocking twice.
  int ret_destroy = pthread_mutex_destroy(&mu);
  printf("  pthread_mutex_unlock returned %d\n", ret_unlock);
  printf("  pthread_mutex_destroy returned %d\n", ret_destroy);

}

REGISTER_TEST(Run, 136)
}  // namespace test136

// test137 TP. Races on stack variables. {{{1
namespace test137 {
int GLOB = 0;
ProducerConsumerQueue q(10);

void Worker() {
  int stack;
  int *tmp = (int*)q.Get();
  (*tmp)++;
  int *racey = &stack;
  q.Put(racey);
  (*racey)++;
  usleep(150000);
  // We may miss the races if we sleep less due to die_memory events...
}

void Run() {
  int tmp = 0;
  printf("test137: TP. Races on stack variables.\n");
  q.Put(&tmp);
  MyThreadArray t(Worker, Worker, Worker, Worker);
  t.Start();
  t.Join();
  q.Get();
}

REGISTER_TEST2(Run, 137, FEATURE | EXCLUDE_FROM_ALL)
}  // namespace test137

// test138 FN. Two closures hit the same thread in ThreadPool. {{{1
namespace test138 {
int GLOB = 0;

void Worker() {
  usleep(100000);
  GLOB++;
}

void Run() {
  FAST_MODE_INIT(&GLOB);
  printf("test138: FN. Two closures hit the same thread in ThreadPool.\n");

  // When using thread pools, two concurrent callbacks might be scheduled
  // onto the same executor thread. As a result, unnecessary happens-before
  // relation may be introduced between callbacks.
  // If we set the number of executor threads to 1, any known data
  // race detector will be silent. However, the same situation may happen
  // with any number of executor threads (with some probability).
  ThreadPool tp(1);
  tp.StartWorkers();
  tp.Add(NewCallback(Worker));
  tp.Add(NewCallback(Worker));
}

REGISTER_TEST2(Run, 138, FEATURE)
}  // namespace test138

// test139: FN. A true race hidden by reference counting annotation. {{{1
namespace test139 {
int GLOB = 0;
RefCountedClass *obj;

void Worker1() {
  GLOB++;  // First access.
  obj->Unref();
}

void Worker2() {
  usleep(100000);
  obj->Unref();
  GLOB++;  // Second access.
}

void Run() {
  FAST_MODE_INIT(&GLOB);
  printf("test139: FN. A true race hidden by reference counting annotation.\n");

  obj = new RefCountedClass;
  obj->AnnotateUnref();
  obj->Ref();
  obj->Ref();
  MyThreadArray mt(Worker1, Worker2);
  mt.Start();
  mt.Join();
}

REGISTER_TEST2(Run, 139, FEATURE)
}  // namespace test139

// test140 TN. Swap. Variant of test79 and test134. {{{1
namespace test140 {
#if 0
typedef __gnu_cxx::hash_map<int, int> Container;
#else
typedef std::map<int,int>             Container;
#endif
Mutex mu;
static Container container;

// Here we use swap to pass a Container between threads.
// The synchronization is correct, but w/o the annotation
// any hybrid detector will complain.
//
// Unlike the test134, we try to have a minimal set of annotations
// so that extra h-b arcs do not hide other races.

// Swap is very unfriendly to the lock-set (and hybrid) race detectors.
// Since tmp is destructed outside the mutex, we need to have a happens-before
// arc between any prior access to map and here.
// Since the internals of tmp are created ouside the mutex and are passed to
// other thread, we need to have a h-b arc between here and any future access.
//
// We want to be able to annotate swapper so that we don't need to annotate
// anything else.
void Swapper() {
  Container tmp;
  tmp[1] = tmp[2] = tmp[3] = 0;
  {
    MutexLock lock(&mu);
    container.swap(tmp);
    // we are unpublishing the old container.
    ANNOTATE_UNPUBLISH_MEMORY_RANGE(&container, sizeof(container));
    // we are publishing the new container.
    ANNOTATE_PUBLISH_MEMORY_RANGE(&container, sizeof(container));
  }
  tmp[1]++;
  tmp[2]++;
  // tmp (which is the old version of container) is destroyed here.
}

void Worker() {
  MutexLock lock(&mu);
  container[1]++;
  int *v = &container[2];
  for (int i = 0; i < 10; i++) {
    // if uncommented, this will break ANNOTATE_UNPUBLISH_MEMORY_RANGE():
    // ANNOTATE_HAPPENS_BEFORE(v);
    if (i % 3) {
      (*v)++;
    }
  }
}

void Run() {
  printf("test140: negative (swap) %p\n", &container);
  MyThreadArray t(Worker, Worker, Swapper, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 140)
}  // namespace test140

// test141 FP. unlink/fopen, rmdir/opendir. {{{1
namespace test141 {
int GLOB1 = 0,
    GLOB2 = 0;
char *dir_name = NULL,
     *filename = NULL;

void Waker1() {
  usleep(100000);
  GLOB1 = 1;  // Write
  // unlink deletes a file 'filename'
  // which exits spin-loop in Waiter1().
  printf("  Deleting file...\n");
  CHECK(unlink(filename) == 0);
}

void Waiter1() {
  FILE *tmp;
  while ((tmp = fopen(filename, "r")) != NULL) {
    fclose(tmp);
    usleep(10000);
  }
  printf("  ...file has been deleted\n");
  GLOB1 = 2;  // Write
}

void Waker2() {
  usleep(100000);
  GLOB2 = 1;  // Write
  // rmdir deletes a directory 'dir_name'
  // which exit spin-loop in Waker().
  printf("  Deleting directory...\n");
  CHECK(rmdir(dir_name) == 0);
}

void Waiter2() {
  DIR *tmp;
  while ((tmp = opendir(dir_name)) != NULL) {
    closedir(tmp);
    usleep(10000);
  }
  printf("  ...directory has been deleted\n");
  GLOB2 = 2;
}

void Run() {
  FAST_MODE_INIT(&GLOB1);
  FAST_MODE_INIT(&GLOB2);
  printf("test141: FP. unlink/fopen, rmdir/opendir.\n");

  dir_name = strdup("/tmp/tsan-XXXXXX");
  mkdtemp(dir_name);

  filename = strdup((std::string() + dir_name + "/XXXXXX").c_str());
  const int fd = mkstemp(filename);
  CHECK(fd >= 0);
  close(fd);

  MyThreadArray mta1(Waker1, Waiter1);
  mta1.Start();
  mta1.Join();

  MyThreadArray mta2(Waker2, Waiter2);
  mta2.Start();
  mta2.Join();
  free(filename);
  filename = 0;
  free(dir_name);
  dir_name = 0;
}
REGISTER_TEST(Run, 141)
}  // namespace test141


// Simple FIFO queue annotated with PCQ annotations. {{{1
class FifoMessageQueue {
 public:
  FifoMessageQueue() { ANNOTATE_PCQ_CREATE(this); }
  ~FifoMessageQueue() { ANNOTATE_PCQ_DESTROY(this); }
  // Send a message. 'message' should be positive.
  void Put(int message) {
    CHECK(message);
    MutexLock lock(&mu_);
    ANNOTATE_PCQ_PUT(this);
    q_.push(message);
  }
  // Return the message from the queue and pop it
  // or return 0 if there are no messages.
  int Get() {
    MutexLock lock(&mu_);
    if (q_.empty()) return 0;
    int res = q_.front();
    q_.pop();
    ANNOTATE_PCQ_GET(this);
    return res;
  }
 private:
  Mutex mu_;
  queue<int> q_;
};


// test142: TN. Check PCQ_* annotations. {{{1
namespace test142 {
// Putter writes to array[i] and sends a message 'i'.
// Getters receive messages and read array[message].
// PCQ_* annotations calm down the hybrid detectors.

const int N = 1000;
int array[N+1];

FifoMessageQueue q;

void Putter() {
  for (int i = 1; i <= N; i++) {
    array[i] = i*i;
    q.Put(i);
    usleep(1000);
  }
}

void Getter() {
  int non_zero_received  = 0;
  for (int i = 1; i <= N; i++) {
    int res = q.Get();
    if (res > 0) {
      CHECK(array[res] = res * res);
      non_zero_received++;
    }
    usleep(1000);
  }
  printf("T=%d: non_zero_received=%d\n",
         (int)pthread_self(), non_zero_received);
}

void Run() {
  printf("test142: tests PCQ annotations\n");
  MyThreadArray t(Putter, Getter, Getter);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 142)
}  // namespace test142


// test143: TP. Check PCQ_* annotations. {{{1
namespace test143 {
// True positive.
// We have a race on GLOB between Putter and one of the Getters.
// Pure h-b will not see it.
// If FifoMessageQueue was annotated using HAPPENS_BEFORE/AFTER, the race would
// be missed too.
// PCQ_* annotations do not hide this race.
int     GLOB = 0;

FifoMessageQueue q;

void Putter() {
  GLOB = 1;
  q.Put(1);
}

void Getter() {
  usleep(10000);
  q.Get();
  CHECK(GLOB == 1);  // Race here
}

void Run() {
  q.Put(1);
  if (!Tsan_PureHappensBefore()) {
    ANNOTATE_EXPECT_RACE_FOR_TSAN(&GLOB, "true races");
  }
  printf("test143: tests PCQ annotations (true positive)\n");
  MyThreadArray t(Putter, Getter, Getter);
  t.Start();
  t.Join();
}
REGISTER_TEST(Run, 143);
}  // namespace test143


// test300: {{{1
namespace test300 {
int     GLOB = 0;
void Run() {
}
REGISTER_TEST2(Run, 300, RACE_DEMO)
}  // namespace test300

// test301: Simple race.  {{{1
namespace test301 {
Mutex mu1;  // This Mutex guards var.
Mutex mu2;  // This Mutex is not related to var.
int   var;  // GUARDED_BY(mu1)

void Thread1() {  // Runs in thread named 'test-thread-1'.
  MutexLock lock(&mu1);  // Correct Mutex.
  var = 1;
}

void Thread2() {  // Runs in thread named 'test-thread-2'.
  MutexLock lock(&mu2);  // Wrong Mutex.
  var = 2;
}

void Run() {
  var = 0;
  printf("test301: simple race.\n");
  MyThread t1(Thread1, NULL, "test-thread-1");
  MyThread t2(Thread2, NULL, "test-thread-2");
  t1.Start();
  t2.Start();
  t1.Join();
  t2.Join();
}
REGISTER_TEST2(Run, 301, RACE_DEMO)
}  // namespace test301

// test302: Complex race which happens at least twice.  {{{1
namespace test302 {
// In this test we have many different accesses to GLOB and only one access
// is not synchronized properly.
int     GLOB = 0;

Mutex MU1;
Mutex MU2;
void Worker() {
  for(int i = 0; i < 100; i++) {
    switch(i % 4) {
      case 0:
        // This read is protected correctly.
        MU1.Lock(); CHECK(GLOB >= 0); MU1.Unlock();
        break;
      case 1:
        // Here we used the wrong lock! The reason of the race is here.
        MU2.Lock(); CHECK(GLOB >= 0); MU2.Unlock();
        break;
      case 2:
        // This read is protected correctly.
        MU1.Lock(); CHECK(GLOB >= 0); MU1.Unlock();
        break;
      case 3:
        // This write is protected correctly.
        MU1.Lock(); GLOB++; MU1.Unlock();
        break;
    }
    // sleep a bit so that the threads interleave
    // and the race happens at least twice.
    usleep(100);
  }
}

void Run() {
  printf("test302: Complex race that happens twice.\n");
  MyThread t1(Worker), t2(Worker);
  t1.Start();
  t2.Start();
  t1.Join();   t2.Join();
}
REGISTER_TEST2(Run, 302, RACE_DEMO)
}  // namespace test302


// test303: Need to trace the memory to understand the report. {{{1
namespace test303 {
int     GLOB = 0;

Mutex MU;
void Worker1() { CHECK(GLOB >= 0); }
void Worker2() { MU.Lock(); GLOB=1;  MU.Unlock();}

void Run() {
  printf("test303: a race that needs annotations.\n");
  ANNOTATE_TRACE_MEMORY(&GLOB);
  MyThreadArray t(Worker1, Worker2);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 303, RACE_DEMO)
}  // namespace test303


// test304: Can not trace the memory, since it is a library object. {{{1
namespace test304 {
string *STR;
Mutex   MU;

void Worker1() {
  sleep(0);
  ANNOTATE_CONDVAR_SIGNAL((void*)0xDEADBEAF);
  MU.Lock(); CHECK(STR->length() >= 4); MU.Unlock();
}
void Worker2() {
  sleep(1);
  ANNOTATE_CONDVAR_SIGNAL((void*)0xDEADBEAF);
  CHECK(STR->length() >= 4); // Unprotected!
}
void Worker3() {
  sleep(2);
  ANNOTATE_CONDVAR_SIGNAL((void*)0xDEADBEAF);
  MU.Lock(); CHECK(STR->length() >= 4); MU.Unlock();
}
void Worker4() {
  sleep(3);
  ANNOTATE_CONDVAR_SIGNAL((void*)0xDEADBEAF);
  MU.Lock(); *STR += " + a very very long string"; MU.Unlock();
}

void Run() {
  STR = new string ("The String");
  printf("test304: a race where memory tracing does not work.\n");
  MyThreadArray t(Worker1, Worker2, Worker3, Worker4);
  t.Start();
  t.Join();

  printf("%s\n", STR->c_str());
  delete STR;
}
REGISTER_TEST2(Run, 304, RACE_DEMO)
}  // namespace test304


// test305: A bit more tricky: two locks used inconsistenly. {{{1
namespace test305 {
int     GLOB = 0;

// In this test GLOB is protected by MU1 and MU2, but inconsistently.
// The TRACES observed by helgrind are:
// TRACE[1]: Access{T2/S2 wr} -> new State{Mod; #LS=2; #SS=1; T2/S2}
// TRACE[2]: Access{T4/S9 wr} -> new State{Mod; #LS=1; #SS=2; T2/S2, T4/S9}
// TRACE[3]: Access{T5/S13 wr} -> new State{Mod; #LS=1; #SS=3; T2/S2, T4/S9, T5/S13}
// TRACE[4]: Access{T6/S19 wr} -> new State{Mod; #LS=0; #SS=4; T2/S2, T4/S9, T5/S13, T6/S19}
//
// The guilty access is either Worker2() or Worker4(), depending on
// which mutex is supposed to protect GLOB.
Mutex MU1;
Mutex MU2;
void Worker1() { MU1.Lock(); MU2.Lock(); GLOB=1; MU2.Unlock(); MU1.Unlock(); }
void Worker2() { MU1.Lock();             GLOB=2;               MU1.Unlock(); }
void Worker3() { MU1.Lock(); MU2.Lock(); GLOB=3; MU2.Unlock(); MU1.Unlock(); }
void Worker4() {             MU2.Lock(); GLOB=4; MU2.Unlock();               }

void Run() {
  ANNOTATE_TRACE_MEMORY(&GLOB);
  printf("test305: simple race.\n");
  MyThread t1(Worker1), t2(Worker2), t3(Worker3), t4(Worker4);
  t1.Start(); usleep(100);
  t2.Start(); usleep(100);
  t3.Start(); usleep(100);
  t4.Start(); usleep(100);
  t1.Join(); t2.Join(); t3.Join(); t4.Join();
}
REGISTER_TEST2(Run, 305, RACE_DEMO)
}  // namespace test305

// test306: Two locks are used to protect a var.  {{{1
namespace test306 {
int     GLOB = 0;
// Thread1 and Thread2 access the var under two locks.
// Thread3 uses no locks.

Mutex MU1;
Mutex MU2;
void Worker1() { MU1.Lock(); MU2.Lock(); GLOB=1; MU2.Unlock(); MU1.Unlock(); }
void Worker2() { MU1.Lock(); MU2.Lock(); GLOB=3; MU2.Unlock(); MU1.Unlock(); }
void Worker3() {                         GLOB=4;               }

void Run() {
  ANNOTATE_TRACE_MEMORY(&GLOB);
  printf("test306: simple race.\n");
  MyThread t1(Worker1), t2(Worker2), t3(Worker3);
  t1.Start(); usleep(100);
  t2.Start(); usleep(100);
  t3.Start(); usleep(100);
  t1.Join(); t2.Join(); t3.Join();
}
REGISTER_TEST2(Run, 306, RACE_DEMO)
}  // namespace test306

// test307: Simple race, code with control flow  {{{1
namespace test307 {
int     *GLOB = 0;
volatile /*to fake the compiler*/ bool some_condition = true;


void SomeFunc() { }

int FunctionWithControlFlow() {
  int unrelated_stuff = 0;
  unrelated_stuff++;
  SomeFunc();                // "--keep-history=1" will point somewhere here.
  if (some_condition) {      // Or here
    if (some_condition) {
      unrelated_stuff++;     // Or here.
      unrelated_stuff++;
      (*GLOB)++;             // "--keep-history=2" will point here (experimental).
    }
  }
  usleep(100000);
  return unrelated_stuff;
}

void Worker1() { FunctionWithControlFlow(); }
void Worker2() { Worker1(); }
void Worker3() { Worker2(); }
void Worker4() { Worker3(); }

void Run() {
  GLOB = new int;
  *GLOB = 1;
  printf("test307: simple race, code with control flow\n");
  MyThreadArray t1(Worker1, Worker2, Worker3, Worker4);
  t1.Start();
  t1.Join();
}
REGISTER_TEST2(Run, 307, RACE_DEMO)
}  // namespace test307

// test308: Example of double-checked-locking  {{{1
namespace test308 {
struct Foo {
  int a;
};

static int   is_inited = 0;
static Mutex lock;
static Foo  *foo;

void InitMe() {
  if (!is_inited) {
    lock.Lock();
      if (!is_inited) {
        foo = new Foo;
        foo->a = 42;
        is_inited = 1;
      }
    lock.Unlock();
  }
}

void UseMe() {
  InitMe();
  CHECK(foo && foo->a == 42);
}

void Worker1() { UseMe(); }
void Worker2() { UseMe(); }
void Worker3() { UseMe(); }


void Run() {
  ANNOTATE_TRACE_MEMORY(&is_inited);
  printf("test308: Example of double-checked-locking\n");
  MyThreadArray t1(Worker1, Worker2, Worker3);
  t1.Start();
  t1.Join();
}
REGISTER_TEST2(Run, 308, RACE_DEMO)
}  // namespace test308

// test309: Simple race on an STL object.  {{{1
namespace test309 {
string  GLOB;

void Worker1() {
  GLOB="Thread1";
}
void Worker2() {
  usleep(100000);
  GLOB="Booooooooooo";
}

void Run() {
  printf("test309: simple race on an STL object.\n");
  MyThread t1(Worker1), t2(Worker2);
  t1.Start();
  t2.Start();
  t1.Join();   t2.Join();
}
REGISTER_TEST2(Run, 309, RACE_DEMO)
}  // namespace test309

// test310: One more simple race.  {{{1
namespace test310 {
int     *PTR = NULL;  // GUARDED_BY(mu1)

Mutex mu1;  // Protects PTR.
Mutex mu2;  // Unrelated to PTR.
Mutex mu3;  // Unrelated to PTR.

void Writer1() {
  MutexLock lock3(&mu3);  // This lock is unrelated to PTR.
  MutexLock lock1(&mu1);  // Protect PTR.
  *PTR = 1;
}

void Writer2() {
  MutexLock lock2(&mu2);  // This lock is unrelated to PTR.
  MutexLock lock1(&mu1);  // Protect PTR.
  int some_unrelated_stuff = 0;
  if (some_unrelated_stuff == 0)
    some_unrelated_stuff++;
  *PTR = 2;
}


void Reader() {
  MutexLock lock2(&mu2);  // Oh, gosh, this is a wrong mutex!
  CHECK(*PTR <= 2);
}

// Some functions to make the stack trace non-trivial.
void DoWrite1() { Writer1();  }
void Thread1()  { DoWrite1(); }

void DoWrite2() { Writer2();  }
void Thread2()  { DoWrite2(); }

void DoRead()  { Reader();  }
void Thread3() { DoRead();  }

void Run() {
  printf("test310: simple race.\n");
  PTR = new int;
  ANNOTATE_TRACE_MEMORY(PTR);
  *PTR = 0;
  MyThread t1(Thread1, NULL, "writer1"),
           t2(Thread2, NULL, "writer2"),
           t3(Thread3, NULL, "buggy reader");
  t1.Start();
  t2.Start();
  usleep(100000);  // Let the writers go first.
  t3.Start();

  t1.Join();
  t2.Join();
  t3.Join();
}
REGISTER_TEST2(Run, 310, RACE_DEMO)
}  // namespace test310

// test311: Yet another simple race.  {{{1
namespace test311 {
int     *PTR = NULL;  // GUARDED_BY(mu1)

Mutex mu1;  // Protects PTR.
Mutex mu2;  // Unrelated to PTR.
Mutex mu3;  // Unrelated to PTR.

void GoodWriter1() {
  MutexLock lock3(&mu3);  // This lock is unrelated to PTR.
  MutexLock lock1(&mu1);  // Protect PTR.
  *PTR = 1;
}

void GoodWriter2() {
  MutexLock lock2(&mu2);  // This lock is unrelated to PTR.
  MutexLock lock1(&mu1);  // Protect PTR.
  *PTR = 2;
}

void GoodReader() {
  MutexLock lock1(&mu1);  // Protect PTR.
  CHECK(*PTR >= 0);
}

void BuggyWriter() {
  MutexLock lock2(&mu2);  // Wrong mutex!
  *PTR = 3;
}

// Some functions to make the stack trace non-trivial.
void DoWrite1() { GoodWriter1();  }
void Thread1()  { DoWrite1(); }

void DoWrite2() { GoodWriter2();  }
void Thread2()  { DoWrite2(); }

void DoGoodRead()  { GoodReader();  }
void Thread3()     { DoGoodRead();  }

void DoBadWrite()  { BuggyWriter(); }
void Thread4()     { DoBadWrite(); }

void Run() {
  printf("test311: simple race.\n");
  PTR = new int;
  ANNOTATE_TRACE_MEMORY(PTR);
  *PTR = 0;
  MyThread t1(Thread1, NULL, "good writer1"),
           t2(Thread2, NULL, "good writer2"),
           t3(Thread3, NULL, "good reader"),
           t4(Thread4, NULL, "buggy writer");
  t1.Start();
  t3.Start();
  // t2 goes after t3. This way a pure happens-before detector has no chance.
  usleep(10000);
  t2.Start();
  usleep(100000);  // Let the good folks go first.
  t4.Start();

  t1.Join();
  t2.Join();
  t3.Join();
  t4.Join();
}
REGISTER_TEST2(Run, 311, RACE_DEMO)
}  // namespace test311

// test312: A test with a very deep stack. {{{1
namespace test312 {
int     GLOB = 0;
void RaceyWrite() { GLOB++; }
void Func1() { RaceyWrite(); }
void Func2() { Func1(); }
void Func3() { Func2(); }
void Func4() { Func3(); }
void Func5() { Func4(); }
void Func6() { Func5(); }
void Func7() { Func6(); }
void Func8() { Func7(); }
void Func9() { Func8(); }
void Func10() { Func9(); }
void Func11() { Func10(); }
void Func12() { Func11(); }
void Func13() { Func12(); }
void Func14() { Func13(); }
void Func15() { Func14(); }
void Func16() { Func15(); }
void Func17() { Func16(); }
void Func18() { Func17(); }
void Func19() { Func18(); }
void Worker() { Func19(); }
void Run() {
  printf("test312: simple race with deep stack.\n");
  MyThreadArray t(Worker, Worker, Worker);
  t.Start();
  t.Join();
}
REGISTER_TEST2(Run, 312, RACE_DEMO)
}  // namespace test312

// test313 TP: test for thread graph output {{{1
namespace  test313 {
BlockingCounter *blocking_counter;
int     GLOB = 0;

// Worker(N) will do 2^N increments of GLOB, each increment in a separate thread
void Worker(int depth) {
  CHECK(depth >= 0);
  if (depth > 0) {
    ThreadPool pool(2);
    pool.StartWorkers();
    pool.Add(NewCallback(Worker, depth-1));
    pool.Add(NewCallback(Worker, depth-1));
  } else {
    GLOB++; // Race here
  }
}
void Run() {
  printf("test313: positive\n");
  Worker(4);
  printf("\tGLOB=%d\n", GLOB);
}
REGISTER_TEST2(Run, 313, RACE_DEMO)
}  // namespace test313


// test400: Demo of a simple false positive. {{{1
namespace test400 {
static Mutex mu;
static vector<int> *vec; // GUARDED_BY(mu);

void InitAllBeforeStartingThreads() {
  vec = new vector<int>;
  vec->push_back(1);
  vec->push_back(2);
}

void Thread1() {
  MutexLock lock(&mu);
  vec->pop_back();
}

void Thread2() {
  MutexLock lock(&mu);
  vec->pop_back();
}

//---- Sub-optimal code ---------
size_t NumberOfElementsLeft() {
  MutexLock lock(&mu);
  return vec->size();
}

void WaitForAllThreadsToFinish_InefficientAndTsanUnfriendly() {
  while(NumberOfElementsLeft()) {
    ; // sleep or print or do nothing.
  }
  // It is now safe to access vec w/o lock.
  // But a hybrid detector (like ThreadSanitizer) can't see it.
  // Solutions:
  //   1. Use pure happens-before detector (e.g. "tsan --pure-happens-before")
  //   2. Call ANNOTATE_MUTEX_IS_USED_AS_CONDVAR(&mu)
  //      in InitAllBeforeStartingThreads()
  //   3. (preferred) Use WaitForAllThreadsToFinish_Good() (see below).
  CHECK(vec->empty());
  delete vec;
}

//----- Better code -----------

bool NoElementsLeft(vector<int> *v) {
  return v->empty();
}

void WaitForAllThreadsToFinish_Good() {
  mu.LockWhen(Condition(NoElementsLeft, vec));
  mu.Unlock();

  // It is now safe to access vec w/o lock.
  CHECK(vec->empty());
  delete vec;
}


void Run() {
  MyThreadArray t(Thread1, Thread2);
  InitAllBeforeStartingThreads();
  t.Start();
  WaitForAllThreadsToFinish_InefficientAndTsanUnfriendly();
//  WaitForAllThreadsToFinish_Good();
  t.Join();
}
REGISTER_TEST2(Run, 400, RACE_DEMO)
}  // namespace test400

// test401: Demo of false positive caused by reference counting. {{{1
namespace test401 {
// A simplified example of reference counting.
// DecRef() does ref count increment in a way unfriendly to race detectors.
// DecRefAnnotated() does the same in a friendly way.

static vector<int> *vec;
static int ref_count;

void InitAllBeforeStartingThreads(int number_of_threads) {
  vec = new vector<int>;
  vec->push_back(1);
  ref_count = number_of_threads;
}

// Correct, but unfriendly to race detectors.
int DecRef() {
  return AtomicIncrement(&ref_count, -1);
}

// Correct and friendly to race detectors.
int DecRefAnnotated() {
  ANNOTATE_CONDVAR_SIGNAL(&ref_count);
  int res = AtomicIncrement(&ref_count, -1);
  if (res == 0) {
    ANNOTATE_CONDVAR_WAIT(&ref_count);
  }
  return res;
}

void ThreadWorker() {
  CHECK(ref_count > 0);
  CHECK(vec->size() == 1);
  if (DecRef() == 0) {  // Use DecRefAnnotated() instead!
    // No one uses vec now ==> delete it.
    delete vec;  // A false race may be reported here.
    vec = NULL;
  }
}

void Run() {
  MyThreadArray t(ThreadWorker, ThreadWorker, ThreadWorker);
  InitAllBeforeStartingThreads(3 /*number of threads*/);
  t.Start();
  t.Join();
  CHECK(vec == 0);
}
REGISTER_TEST2(Run, 401, RACE_DEMO)
}  // namespace test401

// test501: Manually call PRINT_* annotations {{{1
namespace test501 {
int  COUNTER = 0;
int     GLOB = 0;
Mutex muCounter, muGlob[65];

void Worker() {
   muCounter.Lock();
   int myId = ++COUNTER;
   muCounter.Unlock();

   usleep(100);

   muGlob[myId].Lock();
   muGlob[0].Lock();
   GLOB++;
   muGlob[0].Unlock();
   muGlob[myId].Unlock();
}

void Worker_1() {
   MyThreadArray ta (Worker, Worker, Worker, Worker);
   ta.Start();
   usleep(500000);
   ta.Join ();
}

void Worker_2() {
   MyThreadArray ta (Worker_1, Worker_1, Worker_1, Worker_1);
   ta.Start();
   usleep(300000);
   ta.Join ();
}

void Run() {
   ANNOTATE_RESET_STATS();
   printf("test501: Manually call PRINT_* annotations.\n");
   MyThreadArray ta (Worker_2, Worker_2, Worker_2, Worker_2);
   ta.Start();
   usleep(100000);
   ta.Join ();
   ANNOTATE_PRINT_MEMORY_USAGE(0);
   ANNOTATE_PRINT_STATS();
}

REGISTER_TEST2(Run, 501, FEATURE | EXCLUDE_FROM_ALL)
}  // namespace test501

// test502: produce lots of segments without cross-thread relations {{{1
namespace test502 {

/*
 * This test produces ~1Gb of memory usage when run with the following options:
 *
 * --tool=helgrind
 * --trace-after-race=0
 * --num-callers=2
 * --more-context=no
 */

Mutex MU;
int     GLOB = 0;

void TP() {
   for (int i = 0; i < 750000; i++) {
      MU.Lock();
      GLOB++;
      MU.Unlock();
   }
}

void Run() {
   MyThreadArray t(TP, TP);
   printf("test502: produce lots of segments without cross-thread relations\n");

   t.Start();
   t.Join();
}

REGISTER_TEST2(Run, 502, MEMORY_USAGE | PRINT_STATS | EXCLUDE_FROM_ALL
                              | PERFORMANCE)
}  // namespace test502

// test503: produce lots of segments with simple HB-relations {{{1
// HB cache-miss rate is ~55%
namespace test503 {

//  |- |  |  |  |  |
//  | \|  |  |  |  |
//  |  |- |  |  |  |
//  |  | \|  |  |  |
//  |  |  |- |  |  |
//  |  |  | \|  |  |
//  |  |  |  |- |  |
//  |  |  |  | \|  |
//  |  |  |  |  |- |
//  |  |  |  |  | \|
//  |  |  |  |  |  |----
//->|  |  |  |  |  |
//  |- |  |  |  |  |
//  | \|  |  |  |  |
//     ...

const int N_threads = 32;
const int ARRAY_SIZE = 128;
int       GLOB[ARRAY_SIZE];
ProducerConsumerQueue *Q[N_threads];
int GLOB_limit = 100000;
int count = -1;

void Worker(){
   int myId = AtomicIncrement(&count, 1);

   ProducerConsumerQueue &myQ = *Q[myId], &nextQ = *Q[(myId+1) % N_threads];

   // this code produces a new SS with each new segment
   while (myQ.Get() != NULL) {
      for (int i = 0; i < ARRAY_SIZE; i++)
         GLOB[i]++;

      if (myId == 0 && GLOB[0] > GLOB_limit) {
         // Stop all threads
         for (int i = 0; i < N_threads; i++)
            Q[i]->Put(NULL);
      } else
         nextQ.Put(GLOB);
   }
}

void Run() {
   printf("test503: produce lots of segments with simple HB-relations\n");
   for (int i = 0; i < N_threads; i++)
      Q[i] = new ProducerConsumerQueue(1);
   Q[0]->Put(GLOB);

   {
      ThreadPool pool(N_threads);
      pool.StartWorkers();
      for (int i = 0; i < N_threads; i++) {
         pool.Add(NewCallback(Worker));
      }
   } // all folks are joined here.

   for (int i = 0; i < N_threads; i++)
      delete Q[i];
}

REGISTER_TEST2(Run, 503, MEMORY_USAGE | PRINT_STATS
                  | PERFORMANCE | EXCLUDE_FROM_ALL)
}  // namespace test503

// test504: force massive cache fetch-wback (50% misses, mostly CacheLineZ) {{{1
namespace test504 {

const int N_THREADS = 2,
          HG_CACHELINE_COUNT = 1 << 16,
          HG_CACHELINE_SIZE  = 1 << 6,
          HG_CACHE_SIZE = HG_CACHELINE_COUNT * HG_CACHELINE_SIZE;

// int gives us ~4x speed of the byte test
// 4x array size gives us
// total multiplier of 16x over the cachesize
// so we can neglect the cached-at-the-end memory
const int ARRAY_SIZE = 4 * HG_CACHE_SIZE,
          ITERATIONS = 30;
int array[ARRAY_SIZE];

int count = 0;
Mutex count_mu;

void Worker() {
   count_mu.Lock();
   int myId = ++count;
   count_mu.Unlock();

   // all threads write to different memory locations,
   // so no synchronization mechanisms are needed
   int lower_bound = ARRAY_SIZE * (myId-1) / N_THREADS,
       upper_bound = ARRAY_SIZE * ( myId ) / N_THREADS;
   for (int j = 0; j < ITERATIONS; j++)
   for (int i = lower_bound; i < upper_bound;
            i += HG_CACHELINE_SIZE / sizeof(array[0])) {
      array[i] = i; // each array-write generates a cache miss
   }
}

void Run() {
   printf("test504: force massive CacheLineZ fetch-wback\n");
   MyThreadArray t(Worker, Worker);
   t.Start();
   t.Join();
}

REGISTER_TEST2(Run, 504, PERFORMANCE | PRINT_STATS | EXCLUDE_FROM_ALL)
}  // namespace test504

// test505: force massive cache fetch-wback (60% misses) {{{1
// modification of test504 - more threads, byte accesses and lots of mutexes
// so it produces lots of CacheLineF misses (30-50% of CacheLineZ misses)
namespace test505 {

const int N_THREADS = 2,
          HG_CACHELINE_COUNT = 1 << 16,
          HG_CACHELINE_SIZE  = 1 << 6,
          HG_CACHE_SIZE = HG_CACHELINE_COUNT * HG_CACHELINE_SIZE;

const int ARRAY_SIZE = 4 * HG_CACHE_SIZE,
          ITERATIONS = 3;
int64_t array[ARRAY_SIZE];

int count = 0;
Mutex count_mu;

void Worker() {
   const int N_MUTEXES = 5;
   Mutex mu[N_MUTEXES];
   count_mu.Lock();
   int myId = ++count;
   count_mu.Unlock();

   // all threads write to different memory locations,
   // so no synchronization mechanisms are needed
   int lower_bound = ARRAY_SIZE * (myId-1) / N_THREADS,
       upper_bound = ARRAY_SIZE * ( myId ) / N_THREADS;
   for (int j = 0; j < ITERATIONS; j++)
   for (int mutex_id = 0; mutex_id < N_MUTEXES; mutex_id++) {
      Mutex *m = & mu[mutex_id];
      m->Lock();
      for (int i = lower_bound + mutex_id, cnt = 0;
               i < upper_bound;
               i += HG_CACHELINE_SIZE / sizeof(array[0]), cnt++) {
         array[i] = i; // each array-write generates a cache miss
      }
      m->Unlock();
   }
}

void Run() {
   printf("test505: force massive CacheLineF fetch-wback\n");
   MyThreadArray t(Worker, Worker);
   t.Start();
   t.Join();
}

REGISTER_TEST2(Run, 505, PERFORMANCE | PRINT_STATS | EXCLUDE_FROM_ALL)
}  // namespace test505

// test506: massive HB's using Barriers {{{1
// HB cache miss is ~40%
// segments consume 10x more memory than SSs
// modification of test39
namespace test506 {
#ifndef NO_BARRIER
// Same as test17 but uses Barrier class (pthread_barrier_t).
int     GLOB = 0;
const int N_threads = 64,
          ITERATIONS = 1000;
Barrier *barrier[ITERATIONS];
Mutex   MU;

void Worker() {
  for (int i = 0; i < ITERATIONS; i++) {
     MU.Lock();
     GLOB++;
     MU.Unlock();
     barrier[i]->Block();
  }
}
void Run() {
  printf("test506: massive HB's using Barriers\n");
  for (int i = 0; i < ITERATIONS; i++) {
     barrier[i] = new Barrier(N_threads);
  }
  {
    ThreadPool pool(N_threads);
    pool.StartWorkers();
    for (int i = 0; i < N_threads; i++) {
      pool.Add(NewCallback(Worker));
    }
  } // all folks are joined here.
  CHECK(GLOB == N_threads * ITERATIONS);
  for (int i = 0; i < ITERATIONS; i++) {
     delete barrier[i];
  }
}
REGISTER_TEST2(Run, 506, PERFORMANCE | PRINT_STATS | EXCLUDE_FROM_ALL);
#endif // NO_BARRIER
}  // namespace test506

// test507: vgHelgrind_initIterAtFM/stackClear benchmark {{{1
// vgHelgrind_initIterAtFM/stackClear consume ~8.5%/5.5% CPU
namespace test507 {
const int N_THREADS    = 1,
          BUFFER_SIZE  = 1,
          ITERATIONS   = 1 << 20;

void Foo() {
  struct T {
    char temp;
    T() {
      ANNOTATE_RWLOCK_CREATE(&temp);
    }
    ~T() {
      ANNOTATE_RWLOCK_DESTROY(&temp);
    }
  } s[BUFFER_SIZE];
  s->temp = '\0';
}

void Worker() {
  for (int j = 0; j < ITERATIONS; j++) {
    Foo();
  }
}

void Run() {
  printf("test507: vgHelgrind_initIterAtFM/stackClear benchmark\n");
  {
    ThreadPool pool(N_THREADS);
    pool.StartWorkers();
    for (int i = 0; i < N_THREADS; i++) {
      pool.Add(NewCallback(Worker));
    }
  } // all folks are joined here.
}
REGISTER_TEST2(Run, 507, EXCLUDE_FROM_ALL);
}  // namespace test507

// test508: cmp_WordVecs_for_FM benchmark {{{1
// 50+% of CPU consumption by cmp_WordVecs_for_FM
namespace test508 {
const int N_THREADS    = 1,
          BUFFER_SIZE  = 1 << 10,
          ITERATIONS   = 1 << 9;

void Foo() {
  struct T {
    char temp;
    T() {
      ANNOTATE_RWLOCK_CREATE(&temp);
    }
    ~T() {
      ANNOTATE_RWLOCK_DESTROY(&temp);
    }
  } s[BUFFER_SIZE];
  s->temp = '\0';
}

void Worker() {
  for (int j = 0; j < ITERATIONS; j++) {
    Foo();
  }
}

void Run() {
  printf("test508: cmp_WordVecs_for_FM benchmark\n");
  {
    ThreadPool pool(N_THREADS);
    pool.StartWorkers();
    for (int i = 0; i < N_THREADS; i++) {
      pool.Add(NewCallback(Worker));
    }
  } // all folks are joined here.
}
REGISTER_TEST2(Run, 508, EXCLUDE_FROM_ALL);
}  // namespace test508

// test509: avl_find_node benchmark {{{1
// 10+% of CPU consumption by avl_find_node
namespace test509 {
const int N_THREADS    = 16,
          ITERATIONS   = 1 << 8;

void Worker() {
  std::vector<Mutex*> mu_list;
  for (int i = 0; i < ITERATIONS; i++) {
    Mutex * mu = new Mutex();
    mu_list.push_back(mu);
    mu->Lock();
  }
  for (int i = ITERATIONS - 1; i >= 0; i--) {
    Mutex * mu = mu_list[i];
    mu->Unlock();
    delete mu;
  }
}

void Run() {
  printf("test509: avl_find_node benchmark\n");
  {
    ThreadPool pool(N_THREADS);
    pool.StartWorkers();
    for (int i = 0; i < N_THREADS; i++) {
      pool.Add(NewCallback(Worker));
    }
  } // all folks are joined here.
}
REGISTER_TEST2(Run, 509, EXCLUDE_FROM_ALL);
}  // namespace test509

// test510: SS-recycle test {{{1
// this tests shows the case where only ~1% of SS are recycled
namespace test510 {
const int N_THREADS    = 16,
          ITERATIONS   = 1 << 10;
int GLOB = 0;

void Worker() {
  usleep(100000);
  for (int i = 0; i < ITERATIONS; i++) {
    ANNOTATE_CONDVAR_SIGNAL((void*)0xDeadBeef);
    GLOB++;
    usleep(10);
  }
}

void Run() {
  //ANNOTATE_BENIGN_RACE(&GLOB, "Test");
  printf("test510: SS-recycle test\n");
  {
    ThreadPool pool(N_THREADS);
    pool.StartWorkers();
    for (int i = 0; i < N_THREADS; i++) {
      pool.Add(NewCallback(Worker));
    }
  } // all folks are joined here.
}
REGISTER_TEST2(Run, 510, MEMORY_USAGE | PRINT_STATS | EXCLUDE_FROM_ALL);
}  // namespace test510

// test511: Segment refcounting test ('1' refcounting) {{{1
namespace test511 {
int GLOB = 0;

void Run () {
   for (int i = 0; i < 300; i++) {
      ANNOTATE_CONDVAR_SIGNAL(&GLOB);
      usleep(1000);
      GLOB++;
      ANNOTATE_CONDVAR_WAIT(&GLOB);
      if (i % 100 == 0)
         ANNOTATE_PRINT_MEMORY_USAGE(0);
   }
}
REGISTER_TEST2(Run, 511, MEMORY_USAGE | PRINT_STATS | EXCLUDE_FROM_ALL);
}  // namespace test511

// test512: Segment refcounting test ('S' refcounting) {{{1
namespace test512 {
int GLOB = 0;
sem_t SEM;

void Run () {
   sem_init(&SEM, 0, 0);
   for (int i = 0; i < 300; i++) {
      sem_post(&SEM);
      usleep(1000);
      GLOB++;
      sem_wait(&SEM);
      /*if (i % 100 == 0)
         ANNOTATE_PRINT_MEMORY_USAGE(0);*/
   }
   sem_destroy(&SEM);
}
REGISTER_TEST2(Run, 512, MEMORY_USAGE | PRINT_STATS | EXCLUDE_FROM_ALL);
}  // namespace test512

// test513: --fast-mode benchmark {{{1
namespace test513 {

const int N_THREADS = 2,
          HG_CACHELINE_SIZE  = 1 << 6,
          ARRAY_SIZE = HG_CACHELINE_SIZE * 512,
          MUTEX_ID_BITS = 8,
          MUTEX_ID_MASK = (1 << MUTEX_ID_BITS) - 1;

// Each thread has its own cacheline and tackles with it intensively
const int ITERATIONS = 1024;
int array[N_THREADS][ARRAY_SIZE];

int count = 0;
Mutex count_mu;
Mutex mutex_arr[N_THREADS][MUTEX_ID_BITS];

void Worker() {
   count_mu.Lock();
   int myId = count++;
   count_mu.Unlock();

   // all threads write to different memory locations
   for (int j = 0; j < ITERATIONS; j++) {
      int mutex_mask = j & MUTEX_ID_BITS;
      for (int m = 0; m < MUTEX_ID_BITS; m++)
         if (mutex_mask & (1 << m))
            mutex_arr[myId][m].Lock();

      for (int i = 0; i < ARRAY_SIZE; i++) {
         array[myId][i] = i;
      }

      for (int m = 0; m < MUTEX_ID_BITS; m++)
         if (mutex_mask & (1 << m))
            mutex_arr[myId][m].Unlock();
   }
}

void Run() {
   printf("test513: --fast-mode benchmark\n");
   {
      ThreadPool pool(N_THREADS);
      pool.StartWorkers();
      for (int i = 0; i < N_THREADS; i++) {
         pool.Add(NewCallback(Worker));
      }
   } // all folks are joined here.
}

REGISTER_TEST2(Run, 513, PERFORMANCE | PRINT_STATS | EXCLUDE_FROM_ALL)
}  // namespace test513

// End {{{1
// vim:shiftwidth=2:softtabstop=2:expandtab:foldmethod=marker