G1PeriodicGCTask::should_start_periodic_gc calls g1h->time_since_last_collection(), possibly concurrently with the GC setting the collection time, leading to potential value tearing on 32bit platforms. There might be other similar problems; that's just the first one I noticed.

should_start_periodic_gc should probably put it's entire body within an STS-joiner. This would be consistent with other G1ServiceTasks that use an STS-joiner to protect against GC concurrency.