Some threads are created before the GC barrier set, and the barrier set is used to construct and destroy the _gc_data in a thread when the thread is constructed or destroyed.  To cope with this initialization order problem, the initialization of the _gc_data is deferred for such threads. The barrier set is "installed" via the set_barrier_set function, which is responsible for also performing that deferred _gc_data initialization for threads that have been constructed so far.

However, set_barrier_set presently only performs that initialization for the current thread, which is the main thread (a JavaThread).  But there are configurations where some non-JavaThreads are also created before the barrier set.  set_barrier_set should initialize their _gc_data too.

This doesn't seem to be causing any problems yet, probably because we aren't using the GCThreadLocalData from non-JavaThreads.