// unnecessary allocations each Execute call, we keep one heap-allocated
  // version for the thread.
  StatusPtr status_ TF_GUARDED_BY(execution_mutex_);

  const std::string device_;
  ExecutorPtr executor_ TF_GUARDED_BY(execution_mutex_);
  mutable OpPtr op_ TF_GUARDED_BY(execution_mutex_);
  std::unique_ptr<Thread> thread_;
};

DeviceThread::~DeviceThread() {
  {
    tensorflow::mutex_lock l(execution_mutex_);

     * there's another race: If the first thread fails with an exception and a second thread
     * immediately fails with the same exception:
     *
     * Thread1: calls setException(), which returns true, context switch before it can CAS
     * seenExceptions to its exception
     *
     * Thread2: calls setException(), which returns false, CASes seenExceptions to its exception,

     * there's another race: If the first thread fails with an exception and a second thread
     * immediately fails with the same exception:
     *
     * Thread1: calls setException(), which returns true, context switch before it can CAS
     * seenExceptions to its exception
     *
     * Thread2: calls setException(), which returns false, CASes seenExceptions to its exception,

          @Override
          public void run() {
            list.execute();
          }
        };
    Thread thread1 = new Thread(execute);
    Thread thread2 = new Thread(execute);
    thread1.start();
    thread2.start();
    assertEquals(0, runCalled.get());
    okayToRun.countDown();
    thread1.join();
    thread2.join();
    assertEquals(1, runCalled.get());
  }

  public void testAddAfterRun() throws Exception {

          @Override
          public void run() {
            list.execute();
          }
        };
    Thread thread1 = new Thread(execute);
    Thread thread2 = new Thread(execute);
    thread1.start();
    thread2.start();
    assertEquals(0, runCalled.get());
    okayToRun.countDown();
    thread1.join();
    thread2.join();
    assertEquals(1, runCalled.get());
  }

  public void testAddAfterRun() throws Exception {

import java.util.concurrent.atomic.AtomicInteger;
import junit.framework.TestCase;

/**
 * Basher test for {@link ConcurrentHashMultiset}: start a bunch of threads, have each of them do
 * operations at random. Each thread keeps track of the per-key deltas that it's directly
 * responsible for; after all threads have completed, we sum the per-key deltas and compare to the
 * existing multiset values.
 *
 * @author mike nonemacher
 */

    // Thread#setPriority() already checks for validity. These error messages
    // are nicer though and will fail-fast.
    checkArgument(
        priority >= Thread.MIN_PRIORITY,
        "Thread priority (%s) must be >= %s",
        priority,
        Thread.MIN_PRIORITY);
    checkArgument(
        priority <= Thread.MAX_PRIORITY,
        "Thread priority (%s) must be <= %s",
        priority,
        Thread.MAX_PRIORITY);

                }
              });
      thread.start();
      // we want to wait until each thread is WAITING - one thread waiting inside CacheLoader.load
      // (in startSignal.await()), and the others waiting for that thread's result.
      while (thread.isAlive() && thread.getState() != Thread.State.WAITING) {
        Thread.yield();
      }
    }
    gettersStartedSignal.countDown();

    for (int i = 0; i < numThreads; i++) {
      threads[i] =
          new Thread() {
            @Override
            public void run() {
              assertSame(Boolean.TRUE, memoizedSupplier.get());
            }
          };
    }
    for (Thread t : threads) {
      t.start();
    }
    for (Thread t : threads) {
      t.join();
    }

    if (thrown.get() != null) {

    for (int i = 0; i < numThreads; i++) {
      threads[i] =
          new Thread() {
            @Override
            public void run() {
              assertSame(Boolean.TRUE, memoizedSupplier.get());
            }
          };
    }
    for (Thread t : threads) {
      t.start();
    }
    for (Thread t : threads) {
      t.join();
    }

    if (thrown.get() != null) {