G1: set_in_progress() and clear_started() needs a barrier on non-TSO platforms

ConcurrentMarkThread::during_cycle() is implemented as:

bool during_cycle() { return started() || in_progress(); }

So, it checks both ConcurrentMarkThread::_started and ConcurrentMarkThread::_in_progress and they are meant to overlap. That is, we should not set _started to false until after we have set _in_progress to true. This is done in sleepBeforeNextCycle():

void ConcurrentMarkThread::sleepBeforeNextCycle() {
  // We join here because we don't want to do the "shouldConcurrentMark()"
  // below while the world is otherwise stopped.
  assert(!in_progress(), "should have been cleared");

  MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
  while (!started() && !_should_terminate) {
    CGC_lock->wait(Mutex::_no_safepoint_check_flag);
  }

  if (started()) {
    set_in_progress();
    clear_started();
  }
}

On non-TSO platforms there is a risk that the write to _in_progress (from set_in_progress()) is seen by other threads after the write to _started (in clear_started()). In that case there is a window when during_cycle() may return false even though we are in a concurrent cycle.