Version 1.5 of ai12s/ai12-0207-1.txt

Unformatted version of ai12s/ai12-0207-1.txt version 1.5
Other versions for file ai12s/ai12-0207-1.txt

!standard 6.3.1(13.1/3)          17-08-03 AI12-0207-1/03
!standard B.1(19)
!standard B.1(21/3)
!class binding interpretation 16-12-15
!status Amendment 1-2012 17-08-03
!status WG9 Approved 17-10-13
!status ARG Approved 9-0-1 17-06-16
!status work item 16-12-15
!status received 16-09-19
!priority Low
!difficulty Medium
!qualifier Omission
!subject Convention of anonymous access types
!summary
The convention of an anonymous access-to-subprogram type in the profile of a protected operation is Ada.
The convention of an anonymous access type of a component of an array or record type is the convention of the array or record type. The convention of an anonymous access type of a stand-alone object declaration is the convention of the stand-alone object.
!question
(1) RM 6.3.1(13.1/3) currently says:
The calling convention for an anonymous access-to-subprogram parameter or anonymous access-to-subprogram result is protected if the reserved word protected appears in its definition; otherwise, it is the convention of the subprogram that contains the parameter.
This seems to have the wrong effect if an access-to-subprogram type appears in the profile of a protected operation. For instance:
protected P is procedure Do_Something (What_To_Do : access procedure(X : in out Integer)); ... end P;
The above wording says that the convention of What_To_Do is protected. But that could be declared if it is desired. Shouldn't the convention be Ada? (Yes.)
(2) 6.3.1(2/1) says that the default convention of any entity is Ada, unless the Standard says otherwise. Therefore, the convention of an anonymous access-to-subprogram type of a component is always Ada.
If such a component is in a record type that is specified to have convention C, that doesn't make any sense. Should anonymous access-to-subprogram types take their convention from the enclosing record type? (Yes.)
!recommendation
(See Summary.)
!wording
Modify 6.3.1(13.1/3):
The calling convention for an anonymous access-to-subprogram parameter or anonymous access-to-subprogram result is /protected/ if the reserved word protected appears in its definition; otherwise, it is the convention of the {entity}[subprogram] that {has}[contains] the parameter {or result, unless that entity has convention /protected/, /entry/, or Intrinsic, in which case the convention is Ada}.
Add after B.1(19):
* T is an anonymous access type, and T is eligible for convention L,
AARM Reason: We say this so that the presence of an anonymous access component does not necessarily prevent a type from being eligible for convention L. We want the anonymous access type to take the convention from the enclosing type, but if we only said that, the definition would be circular (one can only portably apply the convention L to a record type R if the components of R already have convention L; but the anonymous components of R have to take the convention from R). We include the part of about T being eligible for convention L so that we don't force convention L on some type that is incompatible with it.
[Editor's note: specifically, "eligible for convention L" means that T is only L-compatible if B.1(17/3) or B.1(18) are met. We can't say that above since we're not allowed to use paragraph numbers in the RM or AARM notes.]
Add after B.1(21/3):
If convention L is specified for a type T, for each component of T that has an anonymous access type, the convention of the anonymous access type is L. If convention L is specified for an object that has an anonymous access type, the convention of the anonymous access type is L.
AARM Ramification: This applies to both anonymous access-to-object and anonymous access-to-subprogram types.
!discussion
(1) We want an access-to-subprogram type to only be protected if it includes the keyword protected. If it doesn't, it should inherit the convention from the containing operation. But it shouldn't inherit that convention if it is protected, else the usual case of passing an Ada subprogram wouldn't be possible for protected operations and entries.
We change the wording from "subprogram" to "entity" since this rule applies to all entities with a profile, meaning that it applies to entries as well as subprograms.
(2) This issue applies to any anonymous access type used as a component type, as the convention of an anonymous access type cannot be separately specified. It should be the same as the enclosing type.
A similar issue also applies to stand-alone object declarations that have an anonymous access type.
The wording to fix this is complicated, as the rules currently say that conventions can only be applied to an array or record type if the conventions of the components are appropriate (L-compatible). But we want the convention of some components to come from the convention of the array or record.
We accomplish this in two steps. First we say by fiat that anonymous access types are L-compatible (thus taking them out of consideration as to whether a convention can be specified for a type). Then we specify the convention of anonymous access types used as the type of components (and as the type of objects).
---
Both of these rules are formally incompatible. If some code depends on the convention of some anonymous access type being protected (question 1) or Ada (question 2), then that code will break with these rule changes. But it is much more likely to make these constructs useful in interfacing code. In particular, an implementation could have simply disallowed anonymous access types in interfacing code (other than as access parameters), which doesn't seem helpful. These rules will make such uses portable.
!corrigendum 6.3.1(13.1/3)
Replace the paragraph:
by:
!corrigendum B.1(19)
Insert after the paragraph:
the new paragraph:
!corrigendum B.1(21/3)
Insert after the paragraph:
If the Convention aspect is specified for a type, then the type shall either be compatible with or eligible for the specified convention.
the new paragraph:
If convention L is specified for a type T, for each component of T that has an anonymous access type, the convention of the anonymous access type is L. If convention L is specified for an object that has an anonymous access type, the convention of the anonymous access type is L.
!ASIS
No ASIS effect.
!ACATS test
ACATS B-Tests (and possibly C-Tests) should be constructed to check the appropriate conventions are used.
!appendix

From: Tucker Taft
Sent: Monday, September 19, 2016  2:37 PM

It has been suggested that the calling convention for a component of an
anonymous access-to-subprogram type should come from the enclosing composite
type.  Currently they seem to default to Ada, which doesn't make a lot of
sense if the record or array type is of convention C, for example.  This
would also be more consistent with how formal parameters of an anonymous
access-to-subprogram type work.

Any objections to creating such an AI?

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

From: Tucker Taft
Sent: Tuesday, September 27, 2016  3:43 PM

RM 6.3.1(13.1/3) currently says:

    The calling convention for an anonymous access-to-subprogram
    parameter or anonymous access-to-subprogram result is protected if
    the reserved word protected appears in its definition; otherwise, it
    is the convention of the subprogram that contains the parameter.

It should really say:

    The calling convention for an anonymous access-to-subprogram
    parameter or anonymous access-to-subprogram result is protected if
    the reserved word protected appears in its definition; otherwise, it
    is the convention of the subprogram that contains the parameter,
    unless the containing subprogram is itself protected, in which case the
    convention is Ada.

For example, given:

    protected P is
        procedure Do_Something
          (What_To_Do : access procedure(X : in out Integer));
        ...
    end P;

we clearly want the convention of What_To_Do to be Ada, not "protected."

If we did want What_To_Do to be an access-to-protected procedure, there is
syntax to accomplish that:

    protected P is
        procedure Do_Something
          (What_To_Do : access protected procedure(X : in out Integer));
        ...
    end P;

------

While on the topic, we probably want to generalize paragraph 13.1 above to
handle other places where parameter profiles appear, namely in
access-to-subprogram definitions and entries.  Hence:

    The calling convention for an anonymous access-to-subprogram
    parameter or anonymous access-to-subprogram result is /protected/ if
    the reserved word PROTECTED appears in its definition; otherwise, it
    is the convention of the entity that has the parameter or result,
    unless that entity has convention /protected/, /entry/, or Intrinsic,
    in which case the convention is Ada.

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

From: Randy Brukardt
Sent: Thursday, December 15, 2016  10:02 PM

> It has been suggested that the calling convention for a component of 
> an anonymous access-to-subprogram type should come from the enclosing 
> composite type.  Currently they seem to default to Ada, which doesn't 
> make a lot of sense if the record or array type is of convention C, 
> for example.  This would also be more consistent with how formal 
> parameters of an anonymous access-to-subprogram type work.
> 
> Any objections to creating such an AI?

I was writing up an initial version of a convention AI (this and your other
access-to-subprogram convention question).

It strikes me that this issue also occurs for anonymous access-to-object
types used as components. As with anonymous access-to-subprogram types, you
can't explicitly give the convention for an anonymous type. And, as you say,
it doesn't make much sense for a C record to contain a bunch of components
having Ada convention access types.

Indeed, taking B.1(12-20) literally, an implementation would be allowed to
reject allowing convention C on any record or array with anonymous access
components, as there would be no requirement that an Ada convention anonymous
access type to be C-compatible (that is implementation-defined). That seems
like madness to me. (Of course, anonymous access types are madness in general,
so maybe that fits well. ;-)

Unfortunately, exactly how to word this escapes me at the moment (and Tucker
didn't suggest anything). The wording used for access-to-subprograms in 6.3.1
would be backwards for components: B.1 says that one can only apply a
convention L to a record type if all of its components are L-compatible. Thus
saying that the convention of an anonymous access type is L if the enclosing
type has convention L would be begging the question.

We could try to allow anonymous access types to be L-compatible. Adding
additional bullet after B.1(19) would seem to do the trick:

* T is an anonymous access type, and T is eligible for convention L,

(We want B.1(17/3 & 18) to apply to these anonymous types.)

I think this is OK, as it doesn't seem to allow anything additional beyond
components (one can't derive from an anonymous type).

We'd still have to say something about the convention of the anonymous access
type itself (probably after B.1(22)):

If convention L is specified for a type T, for each component of T that
has an anonymous access type, the convention of the anonymous access type is
L.

Does this work? Any better ideas?

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

From: Randy Brukardt
Sent: Thursday, December 15, 2016  10:11 PM

> It strikes me that this issue also occurs for anonymous 
> access-to-object types used as components. As with anonymous 
> access-to-subprogram types, you can't explicitly give the convention 
> for an anonymous type. And, as you say, it doesn't make much sense for 
> a C record to contain a bunch of components having Ada convention 
> access types.

I wonder if this problem also applies to stand-alone object declarations.
Consider:

    C : access Interfaces.C.Int with Convention C;

The convention of the anonymous access type here is Ada, since we never say
that it is anything else, and therefore 6.3.1(2/1) applies.

I'll leave the wording to fix this as an exercise for the reader (or at least
Tucker).

P.S. I hate anonymous access types.

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

From: Randy Brukardt
Sent: Thursday, December 15, 2016  10:36 PM

Following is my draft of this AI. Constructive comments welcome. (Forget the
non-constructive comments. ;-)

[Editor's note: This was followed by version /01 of the AI.]

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

From: Tucker Taft
Sent: Friday, December 16, 2016  4:13 AM

> Following is my draft of this AI. Constructive comments welcome. 
> (Forget the non-constructive comments. ;-)

Looks good!

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

From: Erhard Ploedereder
Sent: Saturday, December 17, 2016  5:48 PM

Makes sense.

Nit: "type used in a component" (in the summary) is very colloquial.
"used in" -> "of"

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

From: Randy Brukardt
Sent: Monday, December 19, 2016  5:40 PM

OK. Fixed this and a couple of other typos in the AI.

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

From: Randy Brukardt
Sent: Friday, August 4, 2017  12:12 PM

The draft of AI12-0207-1 approved in Vienna had the second insertion after
paragraph B.1(22/3). I moved that after B.1(21/3), as paragraph 21 is about
conventions, but paragraph 22 is about Import. It makes more sense to keep
all of the rules about conventions together.

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

Questions? Ask the ACAA Technical Agent