In the case that the timewindows are event time windows, the events they emit will be timestamped as having occurred at the end of the window. In the case of processing time windows, the events won't have timestamps and the CPU time-of-day clock will be used as the source of timing information.
Update:
Time windows in Flink are aligned to the epoch -- they are not relative to the first event, or anything like that. You are guaranteed that two event time windows that have the same duration and offset, e.g., two one-second-long tumbling windows, will collect events for exactly the same interval of time.
The event stream emitted by an event time window is itself a stream with event time timestamps, and can be further windowed just like any other timestamped event stream. Just keep in mind that all of the events produced by one window instance (i.e., for the same second) will have the same timestamp. So if you follow a 1 sec window with a shorter window, e.g., 100 msec, then 9 times out of 10 the shorter window won't see any events.