TODO:

* Ensure that we can have our 'non-threaded' thread be any thread.

* Turn off the global 'fast_calls' variable, to knobble
  old stubs, and go with the new ABI.

* Cross thread Object references
	+ I think the way to fix this is by using a different
	  interface->adaptor_type: MATECORBA_ADAPTOR_POA_THREADED,
	  such that the fast case will always drop through to
	  small_invoke_stub_n, where we can get far more creative.

* Threading 
	+ GIOPMessageQueueEntry needs to hold a cnx ref.
	+ orb->current_invocations needs to be per-thread ...
	  also, we should push a stack-data pointer.
	+ Need to emit 'broken' connection - via the listen/unlisten
	  stuff in the main thread ... [ and garentee the same
	  invariants as previously ? cnx status etc. ? ]

	+ Create a per-object policy 'PROCESS_REENTERANTLY', using our
	  VPVID space; takes a boolean ... set it on the object ...
	  [ make some libmatecomponent wrappers to improve the situation ]
	+ CORBA_Object_non_existant - needs to do a blocking
	  check for CONNECTED / DISCONNETED.
	+ bin redundant forwards in poa.c
	+ giop_recv_thread_fn - more work needed
	+ ** ERROR **: Need to handle zapped cnx' async callbacks


* User friendliness
	+ tolerate NULL CORBA_Environments in stubs
	+ 'sequence from GList' API.
	+ always return initialized memory on exception
	+ add an X like 'event' processing API for incoming bits (?)

* GIOP fragments
/	+ need to handle a cancel request in the middle of a fragment stream

* It seems the local case optimization can go wrong when
  invoking bogus methods from a different interface on an Object ref.
	+ this is to do with
		a) not checking the vepv offset we get
		b) not expanding the lookup of the vepv offset at all.

* Thread safety
	+ For independent clients, we need some way to
	  arbitrate the giop_recv_buffer_get method - so
	  that either: a single thread does the poll and the
	  rest block, or we all poll.

	+ apparently g_object_ref / unref are not thread safe
	  we need linc_object_ref / unref instead that does
	  a LINC_MUTEX_LOCK on the global object lifecycle lock
	  that needs to move into linc ... sigh !

	+ we need to ensure that linc does no signal emissions,
	  since these are horribly slow and again not safe.

* Optimization
	+ we need 2 de-marshalling routines, 1 with (slow)
	  byteswap and 1 without it, save scads and scads of
	  branch cache misses; easy to do too; make it a virtual
	  method on the recv buffer (?) [ for easy chaining ]

	+ we need some concept of align == native align flagged
	  on TypeCodes, so we can blat things quickly with memcpy
	  in a generic way - eg. sequences of octets.

	+ we need to implement 'wait for completion' on
	  shutdown so that we can activate async unrefs in
	  matecomponent.

* Dragons:
	+ giop_recv_buffer.c (giop_connection_handle_input)
	  manages to go through the error condition a whole
	  load; _why_ is this ? - is it a recursive effect on
	  the method ? why are we called if there is no data ?

	+ We need a typecast in an _allocbuf which we know is
	  aliased, eg. PortableServer_POAList_allocbuf, since
	  it gets assigned to a CORBA_Object *, and the cast
	  is bogus.