During testing some other changes, I came across a case where humongous objects were being allocated on young regions and copied during collection pauses. This breaks the fundamental G1 assumption that humongous objects are always allocated separately. I've only seen this when running dacapo pmd.
|