"David Allsopp" <dra-news@metastack.com> writes:

> Goswin von Brederlow wrote: 
>> I would actually like to propose that to work on a type
>> level. Currently we have:
>> 
>> I propose that records can be subtyped if their prefix matches and
>> ".." would be used to denote the subtype of any record with this
>> prefix:
>> 
> <snip>
>> One could also declare get_x to accept supertypes:
>> 
>> # let get_x (r : { x : int; .. }) = r.x;
>> val get_x : { x : int; .. } -> int = <fun>
>> 
>> # get_x { x=1; y=2; };;
>> - : int = 1
>
> Bear in mind that for this work the label must be at the same ordinal
> position within both record types for this to work (which I think could
> potentially be irritating). For example, given:
>
> # type base = { x : int }
>   let get_x r = r.x
>   type more = { y : int; x : int };;
>
> # get_x ({ y=2; x=1; } :> base);;
>
> You would have to get a type coercion error message.

Yes. They don't share a common prefix except the empty one.

>> I think there is only one thing that would need to be changed for the
>> compiled code with record subtypes like the above. Copying and
>> modifying a record would need allocate a block matching the size of
>> the input record and not the size of the subtype. Think about this
>> code:
>> 
>> # let change_x (r : { x : int; .. }) x = { r with x = x }
>> val change_x : { x : int; .. } as 'a -> int -> 'a = <fun>
>> # change_x { x=1; y=2; } 2;;
>> - : more = {x = 2; y = 2}
>
> I imagine that it's a reasonably rare thing to do, but any C bindings which
> manipulate record types in this way would break (and probably segfault the
> runtime) if this behaviour were required. The other worry in the back of my

Why? If you declare the binding as { x : int } as 'a -> 'a then it
works just like now. If you declare it as { x : int; .. } as 'a -> 'a
then it has to cope with super types. That is totaly up to you (as
writer of the bindings) to decide and to ensure. Ocaml should probably
provide a copy_record or clone_record helper function to simplify { r
with x = 1 } type alterations in C.

> mind is that despite having considerably more flexible records in SML (you
> don't have to declare the types in advance and label sharing is possible), a
> function in SML still has the restriction of only being over a fixed record
> type ... and when SML has odd restrictions, they're usually to do with the
> more "obvious" type system feature being undecideable. For example, you
> can't say in SML:
>
> fun get_x r = #x r;
>
> as the equivalent of the OCaml get_x you propose.
>
>
> David

MfG
        Goswin