sbool parallelMap2_debug;

static <A, B> L<B> parallelMap2(Collection<A> l, F1<A, B> f) {
  ret parallelMap2(iterator(l), l(l), f);
}

static <A, B> L<B> parallelMap2(Iterator<A> it, final int count, final F1<A, B> f) ctex {
  final new L<Pair<B, Int>> out;
    
  int poolSize = coresToUse_fixed(), queueSize = 500;
  NotifyingBlockingThreadPoolExecutor e = new(poolSize, queueSize, 15, TimeUnit.SECONDS);
  if (parallelMap2_debug)
    print("parallelMap2: " + count + " elements, " + poolSize + " threads");
    
  // TODO: if poolSize == 1, do it in this thread

  long time = sysNow();
  try {
    int i = 0;
    for (final A o : iterable(it)) {
      ++i;
      final int _i = i;
      e.execute(r { pcall {
        B x = callF(f, o);
        synchronized(out) {
          out.add(pair(x, _i));
        }
      }});
    }
    /*if (parallelMap2_debug)
      print("parallelMap2: scheduling done after " + (sysNow()-time));*/
    e.shutdown();
    e.awaitTermination(1, TimeUnit.DAYS);
  } finally {
    e.shutdown();
  }
  if (parallelMap2_debug)
    print("parallelMap2: main done after " + (sysNow()-time));
    
  ret firstOfPairs(sortBySecondOfPairs_inPlace(out));
}