Re: Subtype conformance... not what I was expecting.

["Adam Beneschan" <adam@irvine.com>]

Peter C. Chapin wrote:
> I'm experimenting with Ada's handling of access to subprogram types. I
> was surprised to discover that the following example does not work (I'm
> using GNAT GPL 2006):
>
> procedure Check is
>     subtype Narrow is Integer range -10..10;
>     type Function_Ptr is access function(X : Narrow) return Integer;
>
>     function F(Para : Integer) return Narrow is
>     begin
>        return Para;
>     end F;
>
>     G      : Function_Ptr := F'Access;
>     Result : Integer;
> begin
>     Result := G(0);
> end Check;
>
> The compiler complains about the initialization of G with F'Access
> saying that it is not "subtype conformant." However, I believe this
> initialization would be type safe. Since F's argument types are super
> types of G's argument types, there is no context where G can be called
> that would violate the constraints on the underlying arguments of F.
> Similarly since F's return type is a subtype of G's return type,
> anything F might return would be acceptable as a return from G. I
> assumed that this was what subtype conformance was about, but apparently
> not.
>
> In fact, GNAT appears to require the argument and return subtypes to
> match exactly. However, this seems overly restrictive. I'm curious about
> the rationale for this restriction.

I don't know what the language designers' rationale was for requiring
the subtypes to statically match, although I do think that trying to
define a one-way "subtype conformance" that requires subtypes for IN
parameters and function results to be "statically a subset" of another
type may just be too complex to be worth the effort if it's not needed
by real programs.

However, I do foresee a possible problem if the rules were loosened.
In your example, if a function has a parameter of subtype "Narrow", a
compiler could work things out so that the parameter could be passed in
an 8- or 16-bit register, or something like that.  This probably
wouldn't make a difference in a simple example like yours.  But on some
processors, and if there were a larger number of parameters, this sort
of optimization *could* make a difference.  But a compiler wouldn't be
able to do this if there were a possibility that, in some other library
package (say), there were an access-to-subprogram type that were
compatible with *both* this function *and* a similar function whose
parameter has type "Integer" (assuming Integer is 32 bits).  Under the
current language rules, I don't think this is a problem.

                             -- Adam