Collectors like G1 implementing non-contiguous generations previously used an inexact but conservative area for discovery. Concurrent and STW reference processing could discover the same reference multiple times, potentially missing referents during evacuation. So these collectors had to take extra measures while concurrent marking/reference discovery has been running. This change makes discovery exact for G1 (and any collector using non-contiguous generations) so that concurrent discovery and STW discovery discover on strictly disjoint memory areas. This means that the mentioned situation can not occur any more, and extra work is not required any more too.
Reviewed-by: kbarrett, sjohanss
