Version 1.3 of ai05s/ai05-0207-1.txt

Unformatted version of ai05s/ai05-0207-1.txt version 1.3
Other versions for file ai05s/ai05-0207-1.txt

!standard 6.3.1(16/2) 10-04-02 AI05-0207-1/02
!class binding interpretation 10-02-15
!status Amendment 2012 10-04-02
!status ARG Approved 9-0-1 10-02-28
!status work item 10-02-15
!status received 10-01-25
!priority Low
!difficulty Easy
!qualifier Omission
!subject Access constant is considered for mode conformance

!summary

Mode conformance includes whether access parameters are access-to-constant or access-to-variable.

!question

12.6(8) says that mode conformance is required for an instance. 6.3.1(16/2) says that the designated types of access parameters need to staticaly match, but there are no other requirements.

Consider:

generic type T is private; CONST_VAL : T; with procedure PROC (X : access constant T); package Gen_Pack1 is procedure Try_This; end Gen_Pack1;

package body Gen_Pack1 is

C : aliased constant T := CONST_VAL;

procedure Try_This is begin PROC (C'access); end Try_This;

end Gen_Pack1;

package Pack2 is procedure Modify_The_Constant; end Pack2;

with Gen_Pack1; package body Pack2 is

type Rec is record F1 : Integer; end record;

procedure Bump (X : access Rec) is begin X.F1 := X.F1 + 1; end Bump;

package Inst is new Gen_Pack1 (Rec, (F1 => 10), Bump);

procedure Modify_The_Constant is begin Inst.Try_This; end Modify_The_Constant;

end Pack2;

One hopes that something makes this illegal!

!recommendation

(See Summary.)

!wording

Modify 6.3.1(16/2):

Two profiles are mode conformant if they are type{ }[-]conformant, [and] corresponding parameters have identical modes, and, for access parameters or access result types, the designated subtypes statically match{ and either both or neither are access-to-constant}, or the designated profiles are subtype conformant.

!discussion

We fix this problem by adjusting mode conformance to include access-to-constant matching for access parameters and results.

The basic idea of mode conformance is that it implies that the legality checks for a call will be the same for any conformant subprogram. (Similarly, type conformance implies that resolution will be the same.) As such, it needs to include whether or not a particular parameter is access-to-constant or access-to-variable; this is the access parameter equivalent to mode "in" versus mode "in out".

Note that this change has a small incompatibility compared to the Ada 2005 rules; the following is now illegal:

procedure P (A : access constant T);

procedure R (A : access T) renames P; -- Now illegal.

This seems fairly consistent with the handling of null exclusions in renaming, so it is unlikely to be a surprise.

We considered, but rejected, adjusting the rules for generic matching only. Ada has a design principle that generic formal subprograms work the same way as subprogram renaming, and adopting a generic-only rule would break this principle.

!corrigendum 6.3.1(16/2)

Replace the paragraph:

Two profiles are mode conformant if they are type-conformant, and corresponding parameters have identical modes, and, for access parameters or access result types, the designated subtypes statically match, or the designated profiles are subtype conformant.

by:

Two profiles are mode conformant if they are type conformant, corresponding parameters have identical modes, and, for access parameters or access result types, the designated subtypes statically match and either both or neither are access-to-constant, or the designated profiles are subtype conformant.

!ACATS Test

Create a B-Test to check that the examples in the question and in the discussion are rejected.

!appendix


!topic Conformance rules and "access constant"
!reference 4.9.1(2/2), 6.3.1(16-17), 12.6(8)
!from Adam Beneschan 10-01-25
!discussion

4.9.1(2) says that if one anonymous access type is access-constant, and the
other isn't, the subtypes don't statically match.  The rules in 6.3.1(16-17) say
that static subtype matching is required for two profiles to be
subtype-conformant, but not for mode conformance. 12.6(8) says that for an
instance of a generic formal subprogram, only mode conformance is required, not
subtype conformance.  (The last half of 6.3.1(16) requires that *designated*
subtypes statically match for mode conformance, but not that the access subtypes
themselves statically match.)

So what makes this illegal?

generic
    type T is private;
    CONST_VAL : T;
    with procedure PROC (X : access constant T); package Gen_Pack1 is
    procedure Try_This;
end Gen_Pack1;

package body Gen_Pack1 is

    C : aliased constant T := CONST_VAL;

    procedure Try_This is
    begin
        PROC (C'Access);
    end Try_This;

end Gen_Pack1;

package Pack2 is
    procedure Modify_The_Constant;
end Pack2;

with Gen_Pack1;
package body Pack2 is

    type Rec is record
        F1 : Integer;
    end record;

    procedure Bump (X : access Rec) is
    begin
        X.F1 := X.F1 + 1;
    end Bump;

    package Inst is new Gen_Pack1 (Rec, (F1 => 10), Bump);

    procedure Modify_The_Constant is
    begin
        Inst.Try_This;
    end Modify_The_Constant;

end Pack2;

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

Questions? Ask the ACAA Technical Agent