[m-dev.] For review: Add implementation of reference types (global heap)

[Warwick Harvey]

Peter Schachte wrote:

> Maybe this zone should only be created on demand.  Perhaps set
> global_heap_zone = NULL, and put something like
> 
>     if (global_heap_zone == NULL) {
> 	global_heap_zone = create_zone("global_heap", 1, global_heap_size,
> 		next_offset(), global_heap_zone_size, default_handler);
> 	global_hp = global_heap_zone->min;
>     }
> 
> in the code for making permanent copies.  I guess it depends on the
> cost of allocating a big chunk of memory you'll never need vs. the
> cost of an extra test for every permanent copy.

Indeed: which is the better trade-off?  I have no idea.  Or is there some 
better way to handle this?

> > :- pragma inline(new_nb_reference/2).
> > :- pragma c_code(new_nb_reference(X::in, Ref::out), will_not_call_mercury, "
> > 	incr_hp(Ref, 1);
> > 	save_transient_registers();
> > 	*(Word *)Ref = MR_make_partially_permanent(X, (Word *) TypeInfo_for_T,
> > 			(Word *) Ref);
> > 	restore_transient_registers();
> > ").
> 
> Are the save_ and restore_transient_registers(); calls really
> necessary here?

YES!

It took me a while to find this bug in your code.  ;-)  I'm sure it worked 
fine on whatever architecture you were using to develop it, but on the SPARC 
I'm using, it's very definitely necessary: the heap pointer gets clobbered 
whenever a C function is called.

Warwick