Re: Interresting, possibly buggy behavior in GNAT generics w/ expression function.

[Shark8 <onewingedshark@gmail.com>]

On Friday, March 29, 2013 10:21:39 AM UTC-6, Dmitry A. Kazakov wrote:
> On Fri, 29 Mar 2013 14:45:33 +0000, Simon Wright wrote:
> 
> > "Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de> writes:
> > 
> >> On Fri, 29 Mar 2013 07:45:04 +0000, Simon Wright wrote:
> >>
> >>> I don't know whether this is a language problem
> >>
> >> It is a general language problem that generic specifications and bodies
> >> cannot be fully checked. Matched formal parameters only is not sufficient.
> >> Much better than with C++ templates, but still same mess.
> > 
> > But, do you think it should have failed at instantiation?
> 
> 
> It should fail to compile, instantiation time is too late. Generic
> contracts are much too weak.

I'm not sure I believe that [they are too weak] -- in any case this isn't a problem of the weakness of generics, it's a case of the compiler choosing a function which it should not dependent upon information it should not "know".

Keeping the instantiation of 
    Package Margins is new Generic_Attribute_Set(
	Attribute_Name   => "margin-left",
	Attribute_Values => Integer, --J,
	Attribute_Units  => Screen,
	Units            => Units
	);
altering the profile of Image [in the generic]
from:	Function Image(Value : Positive) Return String is
to:	Function Image(Value : Natural) Return String is
alters the behavior
from:	margin-left="88%"
to:	margin-left="884"

The issue disappears entirely if an enumeration is supplied for attribute_values so that the correct Image is chosen in either case of Image/Natural or Image/Positive.

> 
> For example:
> 
> generic
>    type S is new T with private;
> package P is
>    type Q is new S with null record;
>    not overriding procedure Foo (X : Q);
> 
> It is uncheckable if S does not have Foo already. Thus, this must be a
> compile error.
> 
> If we wanted properly contracted generics (personally, I don't care), we
> would need some syntax to tell that S may not have operation Foo, e.g.
> 
> generic
>    type S is new T with private;
>    not with procedure  Foo (X : S);
> package P is
>    type Q is new S with null record;
> 

That's a good point, perhaps you should submit an AI?