new register allocator and calling convention

[Stephen Weeks]

> Yes, but that just means you pass in a ref to the C procedure and it fills it
> in for you.  This way the `secret place' comes from the caller (who makes it
> just before the call) so there is no problem with context switches, and it gets
> GC'd automatically.  I.e.,
> 	copyCurrent: ref ??? -> unit
> where ??? is what ever the type of the thing that Thread.save() currently
> returns.

Yes, but this is annoying too, because it means that the runtime
system has to know about the representation of ref ???.  Admittedly
not much knowledge, but the less the better.  It's actually slightly
messier, because it needs to be ??? option ref.  I'll think about it.