!standard 3.7(10/2)                                    08-02-01    AI05-0063-1/02
!class binding interpretation 07-09-06
!status work item 07-09-06
!status received 07-09-06
!priority Low
!difficulty Easy
!qualifier Error
!subject Access discriminants on derived formal types

!summary

Formal non-tagged limited types are not inherently limited.

A type derived from such a formal type in a generic unit cannot have access

discriminants with defaults.


!question

ACATS 3.0 test B370001 has the following example (note that this test is a
revision of an Ada 95 ACATS test):

   generic
      type FP (DD: access Disc) is limited private;
   package GenLP is
      type DFP (D: access Disc := A_Disc) is new FP(D);  -- Checked in instance.
   end GenLP;

   package Instance_Tsk       is new GenLP (Tsk);                     -- OK.
   package Instance_PT        is new GenLP (PT);                      -- OK.
   package Instance_LimRec    is new GenLP (LimRec);                  -- OK.
   package Instance_NonLimRec is new GenLP (NonLimRec);               -- ERROR:

The test writer thought that this derived type would be legal here and
rechecked in the instance.

However, the actual wording of the rule 3.7(10/2) allows default expressions only
on descendants of task, protected, or explicitly limited record types. FP is
none of these things (an explicitly limited record type needs to be declared
by a record_type_definition).

What is the intent here?

!recommendation

The declaration of a type derived from an untagged limited formal type

cannot have defaulted discriminants.


!wording

[No wording is needed here; AI05-0059-1 includes the needed wording change to
make 3.7(10/2) depend on the definition of "inherently limited".]

AARM note after definition of "inherently limited" after 7.5 (6):

A limited formal type is not inherently limited if it is not explicitly tagged.




!discussion

The current rule works well for generic bodies, where it easily provides
an assume-the-worst rule. But this case seems to need an assume-the-best rule
for generic specifications. Otherwise, it would not be possible to give a
default for an access discriminant given in a generic specification.

[Editor's note: One has to wonder whether this is important enough to bother
with fixing. A default is mainly useful to allow changing of the discriminant (in
a mutable type), and that is of course not possible for limited types anyway.
Surely the easiest fix is to confirm the wording -- thus I didn't propose a
wording fix here.

OTOH, it seems like a bit of wart. It is unusual that we assume-the-worst everywhere
in a generic and provide no way to do this rather than allowing movement to the
specification.]

*****************************

If default discriminants are needed, it is always possible to make the formal

explicitly tagged, or to use a derived type whose ancestor is inherently limited

so this is a very mild restriction that is simpler to describe that the

"legal in spec/assume the worst in body" formulation..

[Editor's note: If the formal is explicitly tagged, then defaults are not allowed
by 3.7(9.1/2) -- tagged types can never have discriminant defaults. This is only
talking about untagged types.]
*****************************



--!corrigendum 7.4(6/2)

!ACATS Test

Replace this case in the ACATS B-Test B370001.

!appendix

****************************************************************