Parallel gc uses ParMarkBitMap::live_words_in_range() a lot to calculate the amount of live words within a range of memory during full gc.

Tests showed that its return value could be cached and reused to improve full gc performance significantly.

Contributed by T. Lei, see http://mail.openjdk.java.net/pipermail/hotspot-gc-dev/2015-December/015899.html for the initial thread.