Version 1.10 of ais/ai-00161.txt
!standard 10.02.01 (09) 01-10-17 AI95-00161/06
!standard 10.02.01 (11)
!standard C.04 (04)
!class binding interpretation 99-02-23
!status ARG approved 7-0-1 01-10-07
!status work item 98-04-16
!status received 96-09-10
!priority Medium
!difficulty Medium
!subject Default-initialized objects
!summary
The creation of an object whose type has preelaborable initialization is allowed
in a preelaborable unit. This makes it possible to declare objects of some
private types in preelaborable units.
The notion of a type having "preelaborable initialization" is defined by
this AI.
!question
The paragraphs RM95 10.2.1(5-9) list a number of constructs which are
forbidden in preelaborable units. In particular, RM95 10.2.1(9) states that:
"The creation of a default-initialized object (including a component) of a
descendant of a private type, private extension, controlled type, task type
or protected type with entry declarations."
1. In this rule, what is the meaning of "default-initialized object"?
2. This rule seems extremely restrictive. For instance, if (as recommended)
System.Address is a private type, the declaration:
A : System.Address;
makes the enclosing unit non-preelaborable. As another example, the
implementation of Ada.Strings.Unbounded suggested in the Rationale is
impossible because the initial value for the deferred constant
Null_Unbounded_String would have to include an extension aggregate for
a controlled type, which is illegal in a preelaborable unit.
!recommendation
(See wording.)
!wording
Change the RM95 10.2.1(9) to read: "The creation of an object (including a
component) of a type which does not have preelaborable initialization.
Similarly the evaluation of an extension_aggregate with an ancestor
subtype_mark denoting a subtype of such a type."
Add a new bullet after C.4(4): "No subtype_mark denotes a controlled type, a
private type, a private extension, a generic formal private type, a generic
formal derived type, or a descendant of such a type."
Add the following after RM95 10.2.1(11):
"The following rules specify which entities have preelaborable initialization:
- The partial view of a private type or private extension, a protected type
without entry_declarations, a generic formal private type, or a generic
formal derived type, have preelaborable initialization if and only if the pragma
Preelaborable_Initialization has been applied to them.
- Protected types with entry_declarations and task types do not have
preelaborable initialization.
- A component (including a discriminant) of a record or protected type has
preelaborable initialization if its declaration includes a default_expression
whose execution does not perform any of the actions listed in RM95 10.2.1(6-9),
or if its declaration does not include a default expression and its type has
preelaborable initialization.
- A derived type has preelaborable initialization if its parent type has
preelaborable initialization and (in the case of a derived record or protected
type) if the non-inherited components all have preelaborable initialization.
Moreover, a user-defined controlled type with a local override for the
Initialize procedure does not have preelaborable initialization.
- A view of a type has preelaborable initialization if it is an elementary
type, an array type whose component type has preelaborable initialization,
or a record type whose components all have preelaborable initialization.
A pragma Preelaborable_Initialization specifies that a type has preelaborable
initialization. The form of this pragma is as follows:
pragma Preelaborable_Initialization (direct_name);
This pragma must appear in the visible part of a package or generic package.
If the pragma appears in the first list of declarative_items of a
package_specification, then the direct_name shall denote the first subtype of a
protected type without entry_declarations, private type or private extension
declared within the same package as the pragma. If the pragma is applied to a
protected type, each component of the protected type shall have preelaborable
initialization. If the pragma is applied to a private type or a private
extension, the full view of the type shall have preelaborable initialization.
In addition to the places where Legality Rules normally apply, these rules apply
also in the private part of an instance of a generic unit.
If the pragma appears in a generic_formal_part, then the direct_name shall
denote a generic formal private type or a generic formal derived type declared
in the same generic_formal_part as the pragma. In a generic_instantiation the
corresponding actual type shall have preelaborable initialization."
!discussion
1. A "default-initialized object" is an object for which the implicit
initialization described in RM95 3.3.1(10-14) takes place. Examples of
default-initialized objects include an object_declaration without an
initialization expression, and the (anonymous) object created by the
evaluation of an extension_aggregate whose ancestor_part is a subtype name.
Note that an object of a task type or of a protected type is always default-
initialized. Also note that for a private type or a private extension,
providing an initialization expression always requires a reference to the
name of an object, or a call to a non-static function, which are forbidden by
other rules. Finally note that for a controlled type, an initialization
expression may also be provided by an extension aggregate (if visibility
allows), which must have a controlled subtype name as the ancestor part,
which is forbidden.
The conclusion is that the phrase "default-initialized" can be removed from
RM95 10.2.1(9).
2. RM95 10.2.1(9) is indeed too restrictive. For preelaboration to be usable
(in particular in the context of distributed systems) we must loosen up the
rules dramatically. However, we do not want to impose an additional burden
on implementations which support annex C and have to comply with the
requirements of RM95 C.4. Therefore, each time we remove a restriction from
RM95 10.2.1(9), we must add a corresponding restriction to RM95 C.4.
Clearly we must keep task objects and protected objects with entries in RM95
10.2.1(9). Note that the current wording seems to allow the creation of
objects of entry-less protected types, which is strange since the
component_declarations may include default_expressions which execute non-
trivial code.
To precisely specify which object declarations are acceptable in a
preelaborable unit, we introduce the notion of a type having preelaborable
initialization. In order to take advantage of the fact that some two-part
types (such as private types and private extensions) may have preelaborable
initialization, without breaking the privacy of private types, we introduce
the pragma Preelaborable_Initialization. This pragma is applied to the
partial view, and it promises that the full view will have preelaborable
initialization. This pragma is applied to some of the predefined private
types, to make it easier for users to write preelaborable units.
The following table lists all the private types and private extensions
declared in preelaborated predefined packages. For each such type, the table
specifies if a pragma Preelaborable_Initialization is applied to the type
(yes) or not (no). For those packages that have a Wide_ variant, the types
declared in the Wide_ variant have the same classification as the types
declared in the 'narrow' variant. The same remark applies to the non-generic
equivalents of the various generic packages.
Ada.Calendar.Time yes
Ada.Finalization.Controlled yes
Ada.Finalization.Limited_Controlled yes
Ada.Numerics.Generic_Complex_Types.Imaginary yes
Ada.Streams.Root_Stream_Type yes
Ada.Strings.Bounded.Generic_Bounded_Length.Bounded_String no
Ada.Strings.Maps.Character_Mapping yes
Ada.Strings.Maps.Character_Set yes
Ada.Strings.Unbounded.Unbounded_String yes
Interfaces.C.Strings.Chars_Ptr yes
System.Address yes
System.Storage_Pool.Root_Storage_Pool yes
Note: The type Bounded_String cannot have a pragma
Preelaborable_Initialization, because that would make it impossible to
instantiate Generic_Bounded_Length with a non-static expression for the
parameter Max.
!corrigendum 10.02.01(9)
Replace the paragraph:
- The creation of a default-initialized object (including a component) of
a descendant of a private type, private extension, controlled type, task type,
or protected type with entry_declarations; similarly the evaluation of an
extension_aggregate with an ancestor subtype_mark denoting a subtype
of such a type.
by:
- The creation of an object (including a component) of a type which does
not have preelaborable initialization. Similarly the evaluation of an
extension_aggregate with an ancestor subtype_mark denoting a subtype
of such a type.
!corrigendum 10.02.01(11)
Insert after the paragraph:
If a pragma Preelaborate (or pragma Pure -- see below) applies to a
library unit, then it is preelaborated. If a library unit is preelaborated,
then its declaration, if any, and body, if any, are elaborated prior to all
non-preelaborated library_items of the partition. The declaration and body
of a preelaborated library unit, and all subunits that are elaborated as part
of elaborating the library unit,All compilation units of a preelaborated
library unit shall be preelaborable. In addition to the places where Legality
Rules normally apply (see 12.3), this rule applies also in the private part of
an instance of a generic unit. In addition, all compilation units of a
preelaborated library unit shall depend semantically only on compilation units
of other preelaborated library units.
the new paragraphs:
The following rules specify which entities have preelaborable initialization:
- The partial view of a private type or private extension, a protected
type without entry_declarations, a generic formal private type, or a
generic formal derived type, have preelaborable initialization if and only if
the pragma Preelaborable_Initialization has been applied to them.
- Protected types with entry_declarations and task types do not
have preelaborable initialization.
- A component (including a discriminant) of a record or protected type
has preelaborable initialization if its declaration includes a
default_expression whose execution does not perform any actions prohibited
in preelaborable constructs as described above, or if its declaration does not
include a default expression and its type has preelaborable initialization.
- A derived type has preelaborable initialization if its parent type has
preelaborable initialization and (in the case of a derived record or protected
type) if the non-inherited components all have preelaborable initialization.
Moreover, a user-defined controlled type with a local override for the
Initialize procedure does not have preelaborable initialization.
- A view of a type has preelaborable initialization if it is an
elementary type, an array type whose component type has preelaborable
initialization, or a record type whose components all have preelaborable
initialization.
A pragma Preelaborable_Initialization specifies that a type has
preelaborable initialization. The form of this pragma is as follows:
pragma Preelaborable_Initialization (direct_name);
This pragma must appear in the visible part of a package or generic package.
If the pragma appears in the first list of declarative_items of a
package_specification, then the direct_name shall denote the
first subtype of a protected type without entry_declarations, private
type, or private extension declared within the same package as the pragma.
If the pragma is applied to a protected type, each component of the
protected type shall have preelaborable initialization. If the pragma is
applied to a private type or a private extension, the full view of the type
shall have preelaborable initialization. In addition to the places where
Legality Rules normally apply, these rules apply also in the private part of
an instance of a generic unit.
If the pragma appears in a generic_formal_part, then the
direct_name shall denote a generic formal private type or a generic formal
derived type declared in the same generic_formal_part as the pragma.
In a generic_instantiation the corresponding actual type shall have
preelaborable initialization.
!corrigendum C.4(4)
Insert after the paragraph:
- Any subtype_mark denotes a statically constrained subtype, with
statically constrained subcomponents, if any;
the new paragraph:
- No subtype_mark denotes a controlled type, a
private type, a private extension, a generic formal private type, a generic
formal derived type, or a descendant of such a type;
!ACATS test
A C-Test is needed to test the newly legal cases allowed by this AI. B-Test
BA21A02 checks for the (still) illegal cases.
!appendix
!section 10.2.1(9)
!subject Default-initialized objects
!reference RM95-10.2.1(9)
!from Pascal Leroy 96-08-30
!reference 96-5697.a Pascal Leroy 96-9-10>>
!discussion
The referenced paragraph says that a construct is not preelaborable if it
performs "The creation of a default-initialized object (including a component)
of a descendant of a private type, private extension, controlled type, task
type or protected type with entry declarations;"
In this rule, what is the meaning of "default-initialized object"? I couldn't
find a definition of this phrase.
Bob Duff (private communication) tells me that this phrase is intended to
cover (among other things?) an object declaration which doesn't have an
initialization expression.
Even accepting this interpretation, there remain open questions. For
instance: is an object to which a pragma Import is applied a
default-initialized object? B.1(38) and B.1(24) say that no initialization
occurs in this case, so it would seem that the answer is 'no'.
It is interesting to note that, with this interpretation, 10.2.1(9) becomes
very restrictive: for a private type, a private extension or a controlled
type, I am not sure how one could write an initial value without either
referencing the name of an object, calling a non-static function, or writing
an extension aggregate. But these things are forbidden by 10.2.1(7-9). So
10.2.1 seems to state that (except in the case of imported objects) a
preelaborable units can never declare an object of a private type, a private
extension, or a controlled type. Is it the intent?
(It looks like all objects of task or protected types are
"default-initialized", so the rule is redundant. But that's not important:
it
is clear that we don't want objects of such types in preelaborated units.)
****************************************************************
From: Randy Brukardt [SMTP:Randy@rrsoftware.com]
Sent: Thursday, May 07, 1998 8:12 PM
Subject: Re: AI95-00161/02
This AI has class "Confirmation", yet proposes language which extends the
language (allowing things previously disallowed). I wouldn't expect
"Confirmation" Ais to have wording changes. Shouldn't this AI be
re-classified?
****************************************************************
From: Pascal Leroy[SMTP:phl@Rational.Com]
Sent: Friday, May 08, 1998 6:04 AM
Subject: Re: AI95-00161/02
> This AI has class "Confirmation", yet proposes language which extends the
> language (allowing things previously disallowed). I wouldn't expect
> "Confirmation" Ais to have wording changes. Shouldn't this AI be
> re-classified?
Surely it should be a binding interpretation. Thanks for noticing.
Pascal
****************************************************************
From: Tucker Taft
Sent: Friday, February 26, 1999 9:47 AM
Subject: Re: Updated AIs
In AI 131, you have proposed that entry-less protected
objects are not preelaborable. This is a troublesome
change, as Shared_Passive packages are required to
be preelaborable, and entry-less protected objects are
the only means of synchronization available via a
Shared_Passive partition.
Hence, the AI needs to accommodate entry-less protected objects
somehow, ideally without causing incompatibilities with
existing code which uses Shared_Passive packages, though
that goal may not be achievable...
-Tuck
****************************************************************
From: Randy Brukardt
Sent: Friday, February 26, 1999 3:06 PM
Subject: RE: Updated AIs
Tucker writes:
>In AI 131, you have proposed that entry-less protected
>objects are not preelaborable.
You mean AI 161. 131 is on interfacing to C.
Randy's comments on AI-161:
There is something wrong with the Wording change in this AI, as it depends on the
(undefined) term "preelaboratable initialization". I think the entire definition of this
term, and the pragma, needs to be moved to the "Wording" section. That would
leave just the actual discussion behind, not the technical definition.
Randy.
****************************************************************
From: Pascal Leroy
Sent: Monday, March 01, 1999 4:16 AM
Yeah, I know someone would notice...
The reason why I didn't structure the AI in this way is that I was under the
impression that the stuff in !discussion was insufficiently rigorous for a
!wording section. But then you're right: having the wording depend on an
undefined term is not too good.
Pascal
****************************************************************
From: Pascal Leroy
Sent: Monday, March 01, 1999 4:14 AM
> In AI 161, you have proposed that entry-less protected
> objects are not preelaborable. This is a troublesome
> change, as Shared_Passive packages are required to
> be preelaborable, and entry-less protected objects are
> the only means of synchronization available via a
> Shared_Passive partition.
This was a mistake. My intent was _not_ to make them non-preelaborable.
However, now that you have drawn my attention to this topic, I believe there
is a specific problem with protected objects. The private part of a protected
object may include default expressions for components, and these default
expressions may or may not be preelaborable. We need to decide if we want to
respect the privacy of the private part of protected objects. If we do, it
seems that the pragma should apply in this case too: i.e., in the absence of a
pragma, the declaration of a protected object should be non-preelaborable; if
on the other hand you use the pragma, then the private part must have
preelaborable initialization.
What do you think?
Pascal
****************************************************************
From: Tucker Taft
Sent: Monday, March 01, 1999 8:18 AM
> > In AI 161, you have proposed that entry-less protected
> > objects are not preelaborable. This is a troublesome
> > change, as Shared_Passive packages are required to
> > be preelaborable, and entry-less protected objects are
> > the only means of synchronization available via a
> > Shared_Passive partition.
>
> This was a mistake. My intent was _not_ to make them non-preelaborable.
>
> However, now that you have drawn my attention to this topic, I believe there
> is a specific problem with protected objects. The private part of a protected
> object may include default expressions for components, and these default
> expressions may or may not be preelaborable. We need to decide if we want to
> respect the privacy of the private part of protected objects. If we do, it
> seems that the pragma should apply in this case too: i.e., in the absence of a
> pragma, the declaration of a protected object should be non-preelaborable; if
> on the other hand you use the pragma, then the private part must have
> preelaborable initialization.
>
> What do you think?
Yes, I suppose the pragma ought to apply to entry-less protected
types in the same way it applies to private types. On the other
hand, there seems no need for the pragma for a singleton protected
object, since the only use of the type is for that object, and
if that object is in a preelaborated package, clearly the
default initialization needs to be preelaborable.
-Tuck
****************************************************************
!topic Suggested implementation of Ada.Strings.Unbounded, Preelaborate
!reference RM95-A.4.5(3), 10.2.1(9); AARM95-A.4.5; Rationale A.2.6
!from Adam Beneschan 03-29-02
!discussion
In the AARM, Section A.4.5 (describing Ada.Strings.Unbounded) ends
with the comment, "A sample implementation of the private part of the
package and several of the subprograms appears in the Rationale."
The package defines a constant Null_Unbounded_String. In the
Rationale, section A.2.6 (see http://www.adaic.org/standards/95rat/
RAThtml/rat95-p3-a.html#2), the suggested implementation defines the
constant this way, in the private part:
Null_Unbounded_String : constant Unbounded_String :=
(Controlled with Reference => Null_String'Access);
where Controlled is Ada.Finalization.Controlled.
However, it appears to me that this suggestion is illegal.
RM95-A.4.5(3) says that Ada.Strings.Unbounded has a Preelaborate
pragma, which means it must be preelaborable, which means, according
to 10.2.1(9), that the elaboration of the package must not evaluate an
extension_aggregate with an ancestor subtype_mark denoting a subtype
of a controlled type. The above constant declaration violates this
rule. (3.4.1(10) defines a type to be a descendant of itself, and
Ada.Finalization.Controlled is a controlled type by 7.6(9).)
Is my interpretation incorrect? It seems strange to me that an
implementation which is suggested by an official document (and also
seems to be the natural way to implement the constant) should be
illegal.
****************************************************************
From: Pascal Leroy
Sent: Saturday, March 30, 2002 2:30 AM
This is a problem that has been known for about 8 years and is addressed by
AI 161. I guess the folks who wrote the Rationale didn't have, at the time,
a full-fledged Ada compiler to check the code examples.
****************************************************************
From: Robert Dewar
Sent: Saturday, March 30, 2002 8:04 AM
Indeed this problem is well known :-)
What we do in GNAT is to use the illegal code as given, but then we have a
special kludge in the compiler to allow this for internal implementation
units (there is no rule in the RM that requires library packages to be
written in Ada :-)
****************************************************************
Questions? Ask the ACAA Technical Agent