[M3devel] Resummary of tagged reference proposals

[hendrik at topoi.pooq.com]

On Tue, Jun 02, 2009 at 10:04:37AM +1000, Tony Hosking wrote:
> The compiler and run-time support is already in the CVS head.
> 
> On 1 Jun 2009, at 22:21, Darko wrote:
> 
> >This is a great little hack. Is it available now or are you still  
> >working on it? I don't know what other people want to use it for but  
> >this is a great way to link memory structures with external data. It  
> >overcomes one of the downsides of GC compared to say C, where you  
> >can use pointer values for other purposes, but it does it safely.
> >
> >Also I'd like to add my voice for the minimalist approach where no  
> >language changes are made. I don't think there is a semantically  
> >clean way of incorporating this into the language  and I'm not sure  
> >there is any payoff in doing so.

I think there is a semantically clean way of incorporating it into the 
language, but it's not a trivial change.  It becomes an optimisable 
special case of a more general feature -- for any bunch of types A, B, 
C, ... you introduce a type UNION(A, B, C, ...).  There's nothing you 
can do with a union except build it from an A, B, or C ..., copy it, and 
test its type and extract the contents -- more or less what you can do 
to one of our type-tagged values.  These unions have a fixed size (more 
or less the max of the sizes of A, B, C, ... and an extra type-tag.
(I don't know if there has to be a restriction on using open arrays 
here; I hole not.)

In the special case of UNION(A, B), where B is a suitable subrange if 
INTEGER and A is a pointer-to-object, the compiler could choose to 
implement this with our low-bit-and-shift hack.

-- hendrik