Version 1.10 of ai05s/ai05-0001-1.txt

• it finalizes V; or

• it calls Assign with V as the Target parameter; or

• it finalizes L; or

• it calls Assign with L as the Target parameter; or

• it finalizes M; or

• it calls Assign with M as the Target parameter; or

• it finalizes S; or

• it calls Assign with S as the Target parameter; or

• Pragma Preelaborate is replaced with pragma Pure.

• The type Vector is declared with a discriminant that specifies the capacity:

• The type Vector needs finalization if and only if type Element_Type needs finalization.

• In function Copy, if the Capacity parameter is equal to or greater than Source.Length, the vector capacity exactly equals the value of the Capacity parameter.

• The description of Reserve_Capacity is replaced by:

• Pragma Preelaborate is replaced with pragma Pure.

• The type List is declared with a discriminant that specifies the capacity (maximum number of elements) as follows:

• The type List needs finalization if and only if type Element_Type needs finalization.

• The allocation of internal storage includes a check that the capacity is not exceeded, and Capacity_Error is raised if this check fails.

• In procedure Assign, if Source length is greater than Target capacity, then Capacity_Error is propagated.

• Function Copy is declared as follows:

• In the three-parameter procedure Splice whose Source has type List, if the sum of Target.Length and Source.Length is greater than Target.Capacity, then Splice propagates Capacity_Error.

• In the four-parameter procedure Splice, if Target.Length equals Target.Capacity, then Splice propagates Capacity_Error.

• Pragma Preelaborate is replaced with pragma Pure.

• The type Map is declared with discriminants that specify both the capacity (number of elements) and modulus (number of distinct hash values) of the hash table as follows:

• The type Map needs finalization if and only if type Key_Type or type Element_Type needs finalization.

• The description of Reserve_Capacity is replaced by:

• An additional operation is added immediately following Reserve_Capacity:

• The function Copy is replaced with:

• Pragma Preelaborate is replaced with pragma Pure.

• The type Map is declared with a discriminant that specifies the capacity (maximum number of elements) as follows:

• The type Map needs finalization if and only if type Key_Type or type Element_Type needs finalization.

• The allocation of a new node includes a check that the capacity is not exceeded, and Capacity_Error is raised if this check fails.

• In procedure Assign, if Source length is greater than Target capacity, then Capacity_Error is propagated.

• The function Copy is replaced with:

• Pragma Preelaborate is replaced with pragma Pure.

• The type Set is declared with discriminants that specify both the capacity (number of elements) and modulus (number of distinct hash values) of the hash table as follows:

• The type Set needs finalization if and only if type Key_Type or type Element_Type needs finalization.

• The description of Reserve_Capacity is replaced by:

• An additional operation is added immediately following Reserve_Capacity:

• The function Copy is replaced with:

• Pragma Preelaborate is replaced with pragma Pure.

• The type Set is declared with a discriminant that specifies the capacity (maximum number of elements) as follows:

• The type Set needs finalization if and only if type Element_Type needs finalization.

• If Insert (or Include) adds an element, a check is made that the capacity is not exceeded, and Capacity_Error is propagated if this check fails.

• In procedure Assign, if Source length is greater than Target capacity, then Capacity_Error is propagated.

• The function Copy is replaced with:

From: Martin Dowie
Date: Sunday, Octoher 2, 2005  1:12 PM

I didn't realize that changes were still being proposed/allowed at this
stage until Randy sent in the "extends" email. :-)

I have a suggestion that I hope will be fairly simple and hopefully
useful...

I think that we could extend the Ada.Containers.* hierarchy to include
"Bounded_" variants of each container type by defining that each bounded
package has exactly the same specification as the unbounded version,
except that the container type has a "(Max_Length : Count_Type)" constraint.

Insert/Append routines would raise a Constraint_Error if "Length
(Container) = Max_Length" (or, a "Length_Error : exception;" could be
added to Ada.Containers).

The only 'tricky' bit is what to with:

1) procedure Set_Length
Either check and raise exception or limit to Max_Length even if "Length
 > Container.Max_Length"?

2) each function "&" =>
   a) if Left or Right is the only container then simply check for space;
   b) if Left and Right are elements then return a Max_Length => 2
container;
   c) if Left and Right are containers then append as much of Right to a
new container with Max_Length of Left.Max_Length.

Alternatively for "&" they could return a new container of
Left.Max_Length + Right.Max_Length but that seems against the nature of
bounded things and not so much in keeping with bounded strings. I'm not
too fussed about what the 'actual' rules are/would be, I'm just pointing
out that these are the only things that would need any discussion.

Anyway, I hope I've shown that it would be /simple/ to double the size
of the Ada.Container.* hierarchy with very little effort (and by effort,
I mean 'text'! :-), in much the same way as "Wide_" packages are added
to the RM.

I have a sample "Ada.Containers.Bounded_Vectors" if anyone would like to
see such a beast.

p.s. My $0.02 would choose "check&raise" for what to do with
Set_Length/"&" subprograms, by the way ;-)

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

From: Bob Duff
Date: Sunday, October 2, 2005  8:33 PM

> I didn't realize that changes were still being proposed/allowed at this
> stage until Randy sent in the "extends" email. :-)

I think Randy's suggestion was for Ada 201X, not Ada 2005.  ;-)

> I have a suggestion that I hope will be fairly simple and hopefully
> useful...
>
> I think that we could extend the Ada.Containers.* hierarchy to include
> "Bounded_" variants of each container type by defining that each bounded
> package has exactly the same specification as the unbounded version,
> except that the container type has a "(Max_Length : Count_Type)" constraint.

Hmm... If I say:

    X: Bounded_Whatever(Max_Length => 80);
    ...
    X := Y & Z;

where Length(Y) + Length(Z) <= 80, I would like it to work.
I don't know how to do that in Ada.

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

From: Jeffrey Carter
Date: Sunday, October 2, 2005  10:04 PM

> I think that we could extend the Ada.Containers.* hierarchy to include
> "Bounded_" variants of each container type by defining that each bounded
> package has exactly the same specification as the unbounded version,
> except that the container type has a "(Max_Length : Count_Type)"
> constraint.

This doesn't work very well, because user-defined assignment doesn't handle
this kind of thing. That's why bounded strings have the limit as a generic
formal constant, not a discriminant on Bounded_String.

Somehow, I doubt anything is being considered for Ada 0X anymore. Randy said
his suggestion was for Ada 1X.

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

From: Matthew Heaney
Date: Monday, October 3, 2005  9:05 AM

> I think that we could extend the Ada.Containers.* hierarchy to include
> "Bounded_" variants of each container type by defining that each bounded
> package has exactly the same specification as the unbounded version,
> except that the container type has a "(Max_Length : Count_Type)"
> constraint.

The container type would have to be limited, because of assignment behavior.

What you refer to as Max_Length above is really just capacity.

> 1) procedure Set_Length
> Either check and raise exception or limit to Max_Length even
 > if "Length > Container.Max_Length"?

Just raise CE.

> 2) each function "&" =>
>   a) if Left or Right is the only container then simply check for space;
>   b) if Left and Right are elements then return a Max_Length => 2
> container;
>   c) if Left and Right are containers then append as much of Right to a
> new container with Max_Length of Left.Max_Length.

No.  Operator "&" is gone, because the type has to be limited.

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

From: Martin Dowie
Date: Monday, October 3, 2005  10:17 AM

I could live with either "Max_Length" or "Capacity".

Why would "&" have to go? They are present in "Bounded_String"-s...

Can you give me an example of the problems with assignment behaviour?
Because just now I'm sitting in front of something that seems to work
fine with everything I try and throw at it.

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

From: Matthew Heaney
Date: Monday, October 3, 2005  10:30 AM

> Why would "&" have to go? They are present in "Bounded_String"-s...

Because there's nothing you can to do with the result, if the type is
limited.

Actually, I should take that back.  Operator "&" could become a
constructor-type function (aka "initializer"), since you would only be
able to use it during the declaration of the bounded contained object
(because the type is limited).

> Can you give me an example of the problems with assignment behaviour?
> Because just now I'm sitting in front of something that seems to work
> fine with everything I try and throw at it.

Assuming the type were non-limited:

declare
    C1 : CT (42);
    C2 : CT (1963);
begin
    C2 := C1;  -- raises CE
end;

(But maybe there were some changes in semantics for Ada 2005, such that
the assignment no longer raises CE.)

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

From: Matthew Heaney
Date: Monday, October 3, 2005  11:29 AM

> Hmm... If I say:
>
>     X: Bounded_Whatever(Max_Length => 80);
>     ...
>     X := Y & Z;
>
> where Length(Y) + Length(Z) <= 80, I would like it to work.
> I don't know how to do that in Ada.

Hmmm... Why not?  If the discriminant of the result of Y & Z is the same
as X, then the assignment will succeed.  And the sum of the lengths can
be calculated in the implementation of "&".

Another issue wrt assignment of bounded forms is the (in)efficiency,
since there will be a bitwise copy of both the active and inactive
("free store") elements.

Martin: the Charles library has bounded list forms (having a clever
implementation of the free store, IMHO).  You can play around with those
if you want to get some more data...

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

From: Martin Dowie
Date: Monday, October 3, 2005  11:50 AM

Thanks Matt - I'll take a look - is there an easy 'zipped' version of
the head, or do I need to download each file individually? (I've never
really used CVS)

And while I can't do:
declare
   C1 : CT (42);
   C2 : CT (1963);
begin
   C2 := C1;  -- raises CE
end;

I can do:
declare
   C1 : CT (42);
   C2 : CT (1963);
begin
   C2 := C2 & C1;  -- works fine
end;

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

From: Bob Duff
Date: Monday, October 3, 2005  12:26 PM

> Hmmm... Why not?  If the discriminant of the result of Y & Z is the same
> as X, then the assignment will succeed.  And the sum of the lengths can
> be calculated in the implementation of "&".

I was thinking of this:

    type Bounded_Sequence(Max_Length: Natural) is
        record
            Elements: Element_Array(1..Max_Length);
            Length: Natural := 0;
        end record;

    X: Bounded_Sequence(Max_Length => 80);
    Y: Bounded_Sequence(Max_Length => 40);
    ...

Now, if Y.Length = 20, I would like to be able to say "X := Y",
since the length of Y is shorter than the Max_Length of X.
But that doesn't work.

Same issue for "X := Y & Z" -- the "&" operation can calculate the right
length for the result, but it can't calculate the right Max_Length.

As you pointed out in another message, this is why Bounded_Strings is
generic on the max length.  That means all objects of a given type will
have the same max length, thus trivially ensuring that, as you said
above, the max length "of the result of Y & Z is the same as X".

> Another issue wrt assignment of bounded forms is the (in)efficiency,
> since there will be a bitwise copy of both the active and inactive
> ("free store") elements.

Good point.  Of course that doesn't matter if the thing is limited,
and if it's limited, the issues about assignment and "&" go away.
In Ada 2005, limited is one good way to go.

In any case, it's not _trivial_ to add these bounded versions,
nor to decide on their exact semantics.  So I think it's unlikely that
they will be included in Ada 2005.  Go for 2015!

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

From: Martin Dowie
Date: Monday, October 3, 2005  11:26 PM

>I think Randy's suggestion was for Ada 201X, not Ada 2005.  ;-)
>
>
Apologies for being a bit early with my Ada1Z proposal then! :-)


>Hmm... If I say:
>
>    X: Bounded_Whatever(Max_Length => 80);
>    ...
>    X := Y & Z;
>
>where Length(Y) + Length(Z) <= 80, I would like it to work.
>I don't know how to do that in Ada.
>
Ok, I'm confused by this one as my implementation seems to do the right
thing for this!...

Anyway, I guess all this can wait for the ISO IWA.

-- bounded_integer_vectors.ads
with Ada.Containers.Bounded_Vectors;

package Bounded_Integer_Vectors is
   new Ada.Containers.Bounded_Vectors (Natural, Integer);

-- test_biv.adb
with Ada.Containers; use Ada.Containers;
with Ada.Exceptions; use Ada.Exceptions;
with Ada.Text_IO;    use Ada.Text_IO;

with Bounded_Integer_Vectors; use Bounded_Integer_Vectors;

procedure Test_BIV is
   subtype Vector_3 is Vector (3);
   V1, V2, V3 : Vector_3;
begin
   Put_Line ("Length =>" & Count_Type'Image (V1.Length));
   V1.Append (1);
   Put_Line ("Length =>" & Count_Type'Image (V1.Length));
   V1.Append (3);
   Put_Line ("Length =>" & Count_Type'Image (V1.Length));
   V1.Append (6);
   Put_Line ("Length =>" & Count_Type'Image (V1.Length));
   begin
      V1.Append (10);
   exception
      when Error : others =>
         Put_Line ("!!! " & Exception_Message (Error));
   end;
   Put_Line ("Length =>" & Count_Type'Image (V1.Length));
   Put_Line ("Length =>" & Count_Type'Image (V3.Length));
   V3 := V1 & V2;
   Put_Line ("Length =>" & Count_Type'Image (V3.Length));
end Test_BIV;

-- result.txt
C:\Ada\Bounded_Containers\gps\obj\test_biv.exe
Length => 0
Length => 1
Length => 2
Length => 3
!!! Insert error, already full.
Length => 3
Length => 0
Length => 3

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

From: Bob Duff
Date: Tuesday, October 4, 2005  12:01 PM

> Ok, I'm confused by this one as my implementation seems to do the right
> thing for this!...

>    subtype Vector_3 is Vector (3);
>    V1, V2, V3 : Vector_3;

Try it like this:

    V1, V2: Vector_3;
    V3: Vector(6);

6 is plenty big enough to store 3 characters, but the assignment
will raise Constraint_Error.

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

From: Martin Dowie
Date: Tuesday, October 4, 2005  12:32 PM

Well, that depends on what I make "&" mean! :-)

   declare
      C1, C2 : Bounded_Integer_Vectors.Vector (3);
      C3     : Bounded_Integer_Vectors.Vector (6);
   begin
      C1.Append (1);
      C1.Append (3);
      C1.Append (7);
      C2.Append (11);
      C2.Append (13);
      C2.Append (17);
      C3 := C1 & C2;
      Put_Line ("+++ No exception raised");
      Put_Line ("Length =>" & Count_Type'Image (C3.Length));
   exception
      when Error : others =>
         Put_Line ("!!! " & Exception_Message (Error));
   end;

-- results.txt
+++ No exception raised
Length => 6

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

From: Jeffrey Carter
Date: Tuesday, October 4, 2005  1:00 PM

>   Well, that depends on what I make "&" mean! :-)

Sure, but the point is that there's no way to make

A := B & C;

work for any A, B, & C such that

Length (B) + Length (C) <= A.Max_Length.

Indeed, if you can make it work for your example and for

A : Vector ( 7);
B : Vector (10);
C : Vector (23);

Length (B) = 3
Length (C) = 3

I'll be impressed.

I suspect this discussion should be moved from Ada-Comment.

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

From: Bibb Latting
Sent: Thursday, February 9, 2006  1:31 PM

!topic Control over storage pool underlying Ada.Containers packages
!reference Ada 2005 RMA.18(all)
!from Bibb Latting 05-02-09
!keywords Containers, Storage Pools
!discussion

The capability to use a derivation of root_storage_pool with a container
is desired to provide flexibility in the management of the storage
underlying the container.  This capability provides the opportunity for
the user to optimize memory allocation for a given application, work-around
fragmentation in the hardware mapping of memory, provide non-core memory
allocation of the container, or to comply with safety/security issues.

Please feel free to change the above discussion as desired.

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

From: Tucker Taft
Sent: Thursday, February 9, 2006  1:58 PM

The general consensus on this suggestion was that it
is good, but would involve significant analysis to
get it right at this point.  A recommendation was to
leave it out of the standard, but encourage implementations
to support the capability.  In general the Containers
packages were designed to be a starting point, not
the final story.  Ordered_ and Hashed_ might easily be
augmented with numerous other capabilities, including
"Bounded_...", "Protected_...", "Pooled_...", etc.
So long as all the packages have nearly identical
contents, it isn't too bad to have implementation-specific
code at the point of the instantiation.  The large
bulk of the user's code will depend only on the package
contents, not its instantiation parameters, and reasonable
portability can be preserved.

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

From: Randy Brukardt
Sent: Thursday, February 9, 2006  3:22 PM

The deadline for informal comments on Ada 2005 was last Friday. At this
point, only editorial comments will be accepted. There may be an opportunity
for formal comments at some point later in the future (although the primary
effect of such comments would be to endanger the future of the Amendment and
potentially of Ada itself).

So, at the point we're primarily talking about future work (beyond Ada
2005).

In any case, there are a number of reasons that user-defined storage pools
are inappropriate for the unbounded containers that Ada 2005 has:

1) There is no good way to name the default storage pool. Thus, there is no
good way to specify the default case of using the default storage pool. We
definitely do not want to make instantiating these containers harder.

Note that we tried early in the Ada 2005 process (see AI-300) to find a way
to make the default pool available. However, it ran into opposition from
some ARG members, who noted that some implementations used multiple default
storage pools, tailored to the characteristics of the type. That model
doesn't make it easy to name the default pool. The issue died because no
solution that worked for such a model could be found.

2) Although it's counter-intuitive, allowing user-defined storage pools
makes the containers less safe. The containers are almost certainly going to
be implemented as controlled types. If the storage pool has some method
other than Deallocate for freeing memory (before the pool itself ceases to
exist) -- and many user-defined pools do -- that could free memory belonging
to container objects and subparts. When those objects are later finalized,
the world will end, as the finalization information could very well have
been destroyed. (For instance, Janus/Ada chains together controlled objects;
if the memory is prematurely freed, those chains can be destroyed, and
almost anything could happen, none of it good.)

This problem cannot happen with the standard storage pool.

One could try to mitigate this problem by trying to define when it would be
safe to do premature deallocation, but such wording would also have the
effect of banning useful implementation techniques (such as caching of
nodes).

We'd also have to add wording to take into account this possibility. That
would make the containers *appear* less safe than they currently are, by
adding additional cases of erroneous behavior. Some reviewers were extremely
concerned about erroneous behavior of the containers, and we spent a lot of
effort eliminating as much such behavior as possible.

Aside: We'd also need wording to allow any operation to call Allocate, and
to require the propagation of any exceptions it generates. We don't need
special wording for the standard storage pool, because all it can do is
propagate Storage_Error, and *any* operation in Ada is *always* allowed to
do that. So no extra wording is needed.

3) The unbounded forms can (and often have to) allocate pretty much any size
or shape of memory as part of pretty much any operation. For instance, a
vector container (depending on it's implementation) could allocate
individual elements, small "chunks" of elements, or the entire data array of
elements at one time. Putting limits on that would be inappropriate for the
unbounded containers - we want to encourage innovative implementations.
Thus, any pool used with the unbounded containers would need to allow
arbitrary-sized allocations.

The effect of (2) and (3) is that many useful storage pools couldn't be used
in the context of the unbounded containers. A pool that is used to manage
memory would be inappropriate; it could only manage memory when no container
instantiation exists, and at that point no memory is allocated anyway. A
pool that does not allow any size allocation also cannot be used. For
instance, a pool that only allows specific sizes of allocations to eliminate
fragmentation could not be used in the context of the unbounded containers.

That doesn't leave many useful pools that could be used with the unbounded
containers. A pool that simply records the memory use could be used, and an
alternative heap could be used, but that's about it. That didn't seem like
enough to be worth the additional complication.

OTOH, it makes perfect sense for other forms of container to provide this
sort of functionality. The intent (as Tucker pointed out) is that the
containers in Ada 2005 provide a starting point for Ada containers, not be
the end point. We certainly hope that additional forms of containers are
developed for future use. (Hopefully, that would contain a solution to (1),
as well.)

One of the things that we'll be doing in future months is considering where
we need future work. Certainly, there has been a lot of interest in other
containers and other forms of containers (such as Bounded), so that will be
considered. And, as these are separate packages, there is no need to wait
another 10 years to standardize them. Moreover, these packages are
relatively easy for implementers to provide, so I would expect that most
implementers will do so. But we'll need energetic people (like Matt Heaney
was for the current set of containers) to participate and draft all of the
tedious wording needed. (A thick skin, to survive the constant nit-picking
at their hard work, also is a prerequisite.)

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

From: Robert A. Duff
Sent: Thursday, February 9, 2006  4:54 PM

>...The containers are almost certainly going to
> be implemented as controlled types. If the storage pool has some method
> other than Deallocate for freeing memory (before the pool itself ceases to
> exist) -- and many user-defined pools do -- that could free memory belonging
> to container objects and subparts.

The container type has finalization.  The heap-allocated memory blocks do not
(unless the actual element type passed in does).  So it would be OK
to use a user-def pool, so long as the element type does not have
controlled parts.

Anyway, I agree with what Tuck and Randy said: allowing user-def pools is a
good idea, but for the future (after Ada 2005).

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

From: Randy Brukardt
Sent: Thursday, February 9, 2006  5:10 PM

Huh? If the heap-allocated blocks are linked together (certainly will be
true in the list container, probably in most of the others too), when the
finalization routine walks the blocks whose memory has been reused to free
them, the results won't be pretty.

We certainly don't want to be mandating implementations that don't include
any links (that would be a major change for the unbounded containers, and
would make it very hard to implement the intended semantics). So I don't see
how one could expect such a pool to work at any time than when no container
objects exist. And even that prevents node caching.

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

From: Robert A. Duff
Sent: Thursday, February 9, 2006  5:53 PM

Yes, you're right.  I was confused.

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

From: Bibb Latting

...
> 1) There is no good way to name the default storage pool. Thus, there is
> no good way to specify the default case of using the default storage pool.

It seems to me that the use of multiple default pools implies an
"overloading" of a "root" default storage pool.  How does this
implementation support a user-defined storage pool?  Are all collections
assigned to the pool allocated from the same user-pool or are there
implementation restrictions that require the user to support an "overloaded"
set of user-pools?

From this perspective, an implementation would be free to do what it
currently does with the default pool, while the use of a user-defined pool
would operate in the same manner and with the same restrictions as the
current implementation.  The implementation always has the option of
exposing the specialized allocation mechanism in some way for their
customer's use.

> 2) Although it's counter-intuitive, allowing user-defined storage pools
> makes the containers less safe.

This problem exists for any controlled type allocated within a user-defined
pool; the coupling between user-defined pools and controlled types is really
what needs to be addressed.  It would seem that the use of a user-defined
pool with any controlled type would be unsafe.

> 3) The unbounded forms can (and often have to) allocate pretty much any
size
> or shape of memory as part of pretty much any operation. For instance, a
> vector container (depending on it's implementation) could allocate
> individual elements, small "chunks" of elements, or the entire data array
of
> elements at one time. Putting limits on that would be inappropriate for
> the
> unbounded containers - we want to encourage innovative implementations.
> Thus, any pool used with the unbounded containers would need to allow
> arbitrary-sized allocations.

Agreed, although the user-defined pool may be attempting to correct
underlying memory issues rather than addressing performance.  Of course this
could be overcome by customizing the implementation of the root pool, if an
interface has been provided by the implementation.

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

From: Randy Brukardt

...
> > 1) There is no good way to name the default storage pool. Thus, there is
> > no good way to specify the default case of using the default
> storage pool.
>
> It seems to me that the use of multiple default pools implies an
> "overloading" of a "root" default storage pool.  How does this
> implementation support a user-defined storage pool?  Are all collections
> assigned to the pool allocated from the same user-pool or are there
> implementation restrictions that require the user to support an
> "overloaded" set of user-pools?

I believe that the implementation uses nested default pools to better handle
memory management. Anyway, I don't understand their implementation very
well, so I can't comment too intelligently.

> From this perspective, an implementation would be free to do what it
> currently does with the default pool, while the use of a user-defined pool
> would operate in the same manner and with the same restrictions as the
> current implementation.  The implementation always has the option of
> exposing the specialized allocation mechanism in some way for their
> customer's use.

Well, the problem is how to represent that default pool in a generic
specification. Surely, the "root" containers have to work with the default
pool; we certainly don't want to require every user to define their own
storage pool! For instance, if we had a container that looked like:

    generic
        type Element_Type is private;
        Pool : in out Root_Storage_Pool'Class;
    package Pool_Unbounded_List is

That would work fine for a user-defined pool, but what would the
instantiation be for the default pool? There is no name for such a thing.
You could use the awful kludge:

    type Junk_Ptr is access all Integer;
    package My_List is new Pool_Unbounded_List (Integer, Junk_Ptr'Storage_Pool);

but the default pool here might not support the allocation of the larger
blocks needed by the list.

> > 2) Although it's counter-intuitive, allowing user-defined storage pools
> > makes the containers less safe.
>
> This problem exists for any controlled type allocated within a user-defined
> pool; the coupling between user-defined pools and controlled types is really
> what needs to be addressed.  It would seem that the use of a user-defined
> pool with any controlled type would be unsafe.

Correct, but the problem is worse with the containers, because the whole
point of the containers is to hide the storage allocation behavior. Thus,
there is no clear point where it is safe to do deallocations via the pool,
and it is less obvious that such deallocations are dangerous.

In any case, this whole area needs work (garbage collection appears to not
be allowed if any controlled objects are in the memory), so it seems that
this will be cleaned up. But not this go-round. (We've only discussed this
problem at dinner; it's never been on the agenda.)

> > 3) The unbounded forms can (and often have to) allocate pretty much any size
> > or shape of memory as part of pretty much any operation. For instance, a
> > vector container (depending on it's implementation) could allocate
> > individual elements, small "chunks" of elements, or the entire data array of
> > elements at one time. Putting limits on that would be inappropriate for the
> > unbounded containers - we want to encourage innovative implementations.
> > Thus, any pool used with the unbounded containers would need to allow
> > arbitrary-sized allocations.
>
> Agreed, although the user-defined pool may be attempting to correct
> underlying memory issues rather than addressing performance.  Of course this
> could be overcome by customizing the implementation of the root pool, if an
> interface has been provided by the implementation.

I agree that there are *some* uses of such pools; there just aren't that
many. The common use of such pools (tracking down storage leaks) shouldn't
be necessary with the predefined containers -- vendors will fix such
problems quickly if they ever existed. In any case, users shouldn't need to
debug the containers implementations.

I have a pool that bypasses all of the heap stuff to directly allocate
memory using the Windows memory allocation functions. It was designed to
allocate arrays that could be expanded in place rather than having to copy
them as more memory is needed. Such a pool would be useful for vectors in
some cases: but it isn't general enough to use for the arbitrary-sized
allocations of the containers library. I don't know how many pools are that
general (unless they fall back to the default pool somehow).

Anyway, it's not that allowing user-defined pools is bad idea, but rather
that such use would work better either with a container tailored for that
use and/or with bounded forms. I think we'll get there, but not immediately.

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

From: Bibb Latting

...
>    generic
>        type Element_Type is private;
>        Pool : in out Root_Storage_Pool'Class;
>    package Pool_Unbounded_List is
>
> That would work fine for a user-defined pool, but what would the
> instantiation be for the default pool? There is no name for such a thing.
> You could use the awful kludge:
>
>    type Junk_Ptr is access all Integer;
>    package My_List is new Pool_Unbounded_List (Integer, Junk_Ptr'Storage_Pool);
>

Indeed, we can cajole the address of the default pool out of the system by
using a variety of kludges.  I'm compelled to confess that I've used several
myself.  In AI-300, there are two aspects of the problem that seem
interesting: 1) a way to name a default pool (not necessarily the root pool;
note the special implementation), and 2) a syntax change to support some
sort of auto-switching in a generic.  I don't think the second is worth the
complexity introduced; the first represents a philosophical problem for the
language.  It seems that supporting some level of visibility for the root
pool would be worthwhile.  However, this exposes the root pool to easier
manipulation (corruption) by a programmer.

I think there are language features in place (Ravenscar is built on the features)
that provide the basis for a mechanism to control what language features are
allowed for a given application.  So, in terms of a default storage pool, I guess
my starting position advocates the presentation of a method to obtain a reference
to a default pool.  What each implementation does with the default reference is
up to the implementation; any derivations would work the same way as user-pools
currently do for an implementation.

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

From: Randy Brukardt
Sent: Tuesday, March 14, 2006  4:30 PM

I'm posting this question here because I don't want to post someone's
contact information on comp.lang.ada (that's a certain way to increase the
amount of spam you get), and it's at least tangentially related to the
Amendment. Please cc Dr. Barry if you respond, since I don't believe that he
is joined here.


---------

From: Dr Alwyn Barry

Hi,

Could you forward this question to the appropriate person?

I have been using Ada 2005 to rewrite some Python code.  In translating
the List processing using the new Doubly_Linked_Lists container class
I constantly have to use:

C : Cursor := First(L);
...
while Has_Element(C) loop
   ... do something using Element(C) or Update_Element(..)
   Next(C);
end loop;

Given the inherent insecurity with First()...Next(), is there a better
iterator, such as the 'natural' extension of the For Loop construct
to iterators ... eg:

for C in L loop
   ... do something using Element (C) or Update_Element(..)
end loop;

Also, the GNAT version binds a cursor to the created list, which
means that when the Move() procedure is called, cursors on the moved
list cannot be used!  Since move is a _genuine_ move, this is very
limiting.  It is impossible to overcome this limitation because the
Node_Type is hidden and so there is no other way to reference into
a list.

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

From: Randy Brukardt
Sent: Wednesday, March 15, 2006  10:42 PM

(I'd hoped someone else would answer this, but since no one has, I'll take a
stab at it...)

A.M. Barry writes:

> I have been using Ada 2005 to rewrite some Python code.  In translating
> the List processing using the new Doubly_Linked_Lists container class
> I constantly have to use:
>
> C : Cursor := First(L);
> ...
> while Has_Element(C) loop
>    ... do something using Element(C) or Update_Element(..)
>    Next(C);
> end loop;
>
> Given the inherent insecurity with First()...Next(), is there a better
> iterator, such as the 'natural' extension of the For Loop construct
> to iterators ... eg:
>
> for C in L loop
>    ... do something using Element (C) or Update_Element(..)
> end loop;

We talked about this "natural" extension a little bit, and it's pretty clear
that it's not obvious what is natural. See the e-mail in AC-112:
http://www.ada-auth.org/cgi-bin/cvsweb.cgi/AIs/AC-00112.TXT (warning, this
link is going to change in the next couple of days to:
http://www.ada-auth.org/cgi-bin/cvsweb.cgi/ACs/AC-00112.TXT - the site is
being reorganized to make room for Ada 2005 AIs and ASIS AIs.)

However, there are the Iterate and Reverse_Iterate procedures in the
containers packages. Is there a reason that you can't use them?? They're not
quite as convinient as the loop syntax given above, but they're just as
safe.

> Also, the GNAT version binds a cursor to the created list, which
> means that when the Move() procedure is called, cursors on the moved
> list cannot be used!  Since move is a _genuine_ move, this is very
> limiting.  It is impossible to overcome this limitation because the
> Node_Type is hidden and so there is no other way to reference into
> a list.

Cursors include the container; that's a fundamental property of them in this
model. That's because you can do operations on them without giving the
container. Also, cursors are not pointers per se, although they may be
implemented that way. We wanted to allow alternative implementations that
don't necessarily use explicit pointers.

In any case, Move destroys all of the cursors; the language defines them as
"invalid", and any future use is erroneous. (If the error was detected, you
are lucky.) In this sense, Move is very similar to destroying the container
itself (say, by calling Unchecked_Deallocation on it).

But Move isn't intended for cases where you need to preserve cursors. The
four parameter Splice is designed for that, as it returns the new cursor for
where the element(s) are inserted. If you want to preserve the original
container, just use ":=".

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

From: Matthew Heaney
Sent: Thursday, March 16, 2006  7:56 AM

> (I'd hoped someone else would answer this, but since no one
> has, I'll take a stab at it...)

I sent him some email privately.

> In any case, Move destroys all of the cursors; the language
> defines them as "invalid", and any future use is erroneous.
> (If the error was detected, you are lucky.)

In the GNAT case (that's the compiler he's using), the cursor is implemented
as a record with a pointer to list object and another pointer to the node
containing the item.  When he attempted to use the old cursor with the new
list object, he probabably got Program_Error, since the first thing
cursor-based operations do is compare the list pointer to the address of the
list parameter.

> In this sense,
> Move is very similar to destroying the container itself (say,
> by calling Unchecked_Deallocation on it).

Well, it's not quite as extreme as that, since the nodes haven't been
deallocated; they've just been moved onto another list.  His algorithm would
work fine if the cursor didn't have a list pointer too.

There is often a tradeoff between flexibility and safety, and Move is an
example where some flexibility was lost as a result of adding the checks.

> But Move isn't intended for cases where you need to preserve
> cursors. The four parameter Splice is designed for that, as
> it returns the new cursor for where the element(s) are
> inserted. If you want to preserve the original container,
> just use ":=".

You could use Splice (in fact that's what I told him), but it can get a
little hairy since you must somehow keep track of all the extant cursors.
As you're iterating over the old list, splicing nodes on the new list, you
have to compare the cursor with old cursors, and then call splice with the
old cursor.

I also told him to just write a generic child procedure to rebind the cursor
from the old list to the new, and he seemed to like that idea better.
There's probably a more elegant way to solve his problem without extending
container behavior, but I didn't really understand the small fragments of
code he sent me.

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

From: Nick Roberts
Sent: Saturday, March 18, 2006  5:20 PM

> Given the inherent insecurity with First()...Next(), is there a better
> iterator, such as the 'natural' extension of the For Loop construct
> to iterators ... eg:
>
> for C in L loop
>    ... do something using Element (C) or Update_Element(..)
> end loop;

For what it is worth, I would point out that Ada is really a somewhat
lower level language than Python. I find iterating awkward in Ada (and
in many other languages, to a lesser or greater degree), but I often
find that, in Ada, an alternative such as:

    declare
       procedure Mogrify (C : in Widget_Lists.Cursor) is
       begin
          ...
       end;
    begin
       Widget_Lists.Iterate (My_Widgets, Mogrify'Access);
    end;

alleviates the dangers of missing something vital out.

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

From: Pascal Leroy
Sent: Monday, March 20, 2006  8:00 AM

> in Ada, an alternative such as:
>
>     declare
>        procedure Mogrify (C : in Widget_Lists.Cursor) is
>        begin
>           ...
>        end;
>     begin
>        Widget_Lists.Iterate (My_Widgets, Mogrify'Access);
>     end;
>
> alleviates the dangers of missing something vital out.

I fail to see what dangers it alleviates.  I for one would much prefer to
write something like:

	for each C in My_Widgets loop
	   ...
	end loop;

Remember, Ada is about readability, too.

However, last time this was discussed it appeared that grafting such a
capability on the language was not easy.

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

From: Randy Brukardt
Sent: Monday, March 20, 2006  8:28 PM

I'm reposting this private mail (with permission) on this topic for the
record.

----
From: Alwyn Barry

Hi,

Thanks for your response.

I did, in the end, use iterators.  However I am using a list
(which is backed-up on each cycle of the algorithm to a
back-up list), and three different index lists onto the list
(which are also backed up).  This means, in the worst case,
up to five iterator functions, each of one line, for five
lines of algorithm.  Its ok, but hardly readable.  For cases
where there are large tasks to do, Iterate is great.  For
single statements it is overkill.  What is particularly
frustrating is that though iterate() is provided [effectively
apply()], there is a lack of reduce() and map() functions
which would remove the need for the iterator functions to
access other variables outside the function in order to do
their task.  These functions can easily be provided, but
could have been included.

Anyway, that was one issue which I could work around, although
frustrated that Ada is not adopting features which other
languages have included exactly so that the "dropped next()"
errors do not occur.  In regard to the other problem...

Splice() could be used, but there is no cursor update
with it.  As Matthew Heaney points out in his reply, trying
to update the cursors too gets nasty (the cursor is invalid
once the item has moved, yet the new list does not contain
the item so you cannot update the cursor before the splice).
This can be programmed around, but thats not helpful either.

Anyway, I've got around it by the old trick of having two
records, one the 'current' and one the 'backup' list and
indexes, and simply switching pointers.  I used to do this
in C, which I guess says something about the elegance or
otherwise of the solution.  It does save copying, but where
speed is not really an issue it is a pity there is not a
better way.  Mind you, I do like the safety constraints of
the current implementation! ... it just needs a means of
switching the cursors if the list is moved, or a further
internal level of indirection so that the cursor to list
pointer remains, but the actual list type is a level removed.

Alwyn

---

From: Randy Brukardt

The general answer is that you can't provide the solution to every possible
need, and keep things efficient. Every solution has trade-offs, and with
containers, *everybody* has a better idea. Undoubtedly there is something
better that could have been done [if I had designed them, they would have
worked quite differently!], but the real value here is that they're
standard. There never was any intent that this version would be the final
word in containers for Ada.

The syntax for iterators *is* clunky. We never seriously considered doing
anything about that; I suspect that the greater use of containers will make
that more important as we go forward. But the "next" style is needed for
maximum flexibility; the only way to avoid it is to give up that flexibility
in some way. So, you need both (no matter how nice the syntax of iterators).

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

From: Nick Roberts
Sent: Tuesday, March 21, 2006  8:52 PM

Alwyn Barry wrote [forwarded by Randy Brukardt]:

> ...  What is particularly frustrating is that though iterate() is
> provided [effectively apply()], there is a lack of reduce() and map()
> functions which would remove the need for the iterator functions to
> access other variables outside the function in order to do their
> task.  These functions can easily be provided, but could have been
> included. ...

These functions (reduce and map) could not be so easily provided in Ada
as they are in Python (which has dynamic typing and comprehensions,
which Ada essentially does not). I think both functions would have to be
generic, the 'reduce' function taking the result type as its generic
parameter, and the 'map' function taking another list or vector package
as its generic parameter. In other words, they would be not be the
elegant one-liners that they are in Python.

> ... Anyway, I've got around it by the old trick of having two
> records, one the 'current' and one the 'backup' list and indexes, and
> simply switching pointers.  I used to do this in C, which I guess
> says something about the elegance or otherwise of the solution.  It
> does save copying, but where speed is not really an issue it is a
> pity there is not a better way.  ...

This is the kind of solution that I would have adopted. I'm sure that it
is generally safer using access values in Ada than using pointers in C,
so I'm not sure it need be considered a particularly inelegant solution
(perhaps not in C either).

Randy Brukardt replied:

> The general answer is that you can't provide the solution to every
> possible need, and keep things efficient. ... [if I had designed
> them, they would have worked quite differently!], but the real value
> here is that they're standard. There never was any intent that this
> version would be the final word in containers for Ada. ...

If I had designed the containers, they would have worked quite
differently as well! But I think Matthew Heaney (the principal designer)
and the ARG have done a good job, in the end, of producing a design that
combines efficiency, safety, and intelligibility, especially as they
didn't have a great deal of time to refine it.

I am currently using Matt's reference implementation of the new
containers, grafted into GNAT 3.15p (which is Ada 95), in various 'real
life' programs I am writing, so I'm getting to know them quite well, and
they are very handy!

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

From: Matthew Heaney
Sent: Tuesday, March 21, 2006  9:17 AM

> For cases
> where there are large tasks to do, Iterate is great.  For
> single statements it is overkill.

The library works by providing primitives that you can use to synthesize
higher-level abstractions.  Any "single statement" can be written as a
generic and instantiated as necessary.

 > What is particularly
> frustrating is that though iterate() is provided [effectively
> apply()], there is a lack of reduce() and map() functions
> which would remove the need for the iterator functions to
> access other variables outside the function in order to do
> their task.

Well this is just silly.  Of *course* local functions "access variables
outside the function" since otherwise there'd be little point in making
them local.  Ada has *always* worked this way; the container library
simply conforms to already-established language idioms.

> These functions can easily be provided, but
> could have been included.

Yes of course, and that is by design.  Obviously we didn't include
everything.  If you want a map function, then it's trivial to write it
once and then keep reusing it.  Here's one for arrays (which of course
could be generalized):

generic
    type Result_Type is private;
    type Array_Type is array (Count_Type range <>) of Result_Type;
    type Container_Type (<>) is limited private;
    type Cursor_Type (<>) is limited private;
    with procedure Process
      (Posn : Cursor_Type;
       Result : out Result_Type);
    with function Length (C : Container_Type) return Count_Type;
    with procedure Iterate
      (C : Container_Type;
       Process : not null access procedure (C : Cursor);
function Generic_Map (C : Container_Type) return Array_Type;

function Generic_Map (C : Container_Type) return Array_Type is
    A : Array_Type (1 .. Length (C));
    I : Count_Type := 0;

    procedure Process (Posn : Cursor_Type) is
    begin
       I := I + 1;
       Process (Posn, A (I));
    end;
begin
    Iterate (C, Process'Access);
    return A (1 .. I);
end;

This algorithm works for any container type, including arrays.

> Anyway, that was one issue which I could work around, although
> frustrated that Ada is not adopting features which other
> languages have included exactly so that the "dropped next()"
> errors do not occur.  In regard to the other problem...

What "other languages" are we speaking about?  Ada is a low-level
systems programming language.  It is not a scripting language.  In
particular it is neither Perl nor Python.

Ada's closest analog is C++.  Anything you can do in the STL you can do
in the Ada container library.

We didn't even attempt to include a generic algorithms library too,
since we were trying to keep the scope of the revision small.  The
simplest thing for you to do it write the generic algorithms you need
and reuse them as necessary.

> Splice() could be used, but there is no cursor update
> with it.

Of course there is cursor update with it.  I don't know what you're
talking about here.  The Position parameter of the 4-parameter Splice is
passed using inout mode, the purpose of which is to rebind the cursor
from the old (source) list to the new (target) list.

> As Matthew Heaney points out in his reply, trying
> to update the cursors too gets nasty (the cursor is invalid
> once the item has moved, yet the new list does not contain
> the item so you cannot update the cursor before the splice).

So you update the cursor *during* the call to Splice, by specifying the
cursor you want to rebind as the Position parameter.

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

From: Randy Brukardt
Sent: Thursday, September 25, 2008  5:55 PM

Following is my notes on the two days of containers meetings. Matt has his own
set of notes that probably differ from mine. I made a quick pass thorough these
notes to eliminate/clarify nonsense, but otherwise they are as I took them.

A quick summary is that we agreed on the form and semantics of the bounded
containers (as little as change is possible, but no pointers and no controlled
types are expected to be used in the implementation of them so that that can be
used in safety-critical environments). We looked at a design for shared
(task-safe) queue abstractions, but concluded that there wasn't enough need for
generalized task-safe containers. (Especially as there are as many meanings of
"task-safe" as there are programmers.)

We rejected storage-optimized variants of the containers, as they are
insufficiently different from the basic containers.

We also studied issues involved with defining containers for limited element
types. This probably would need some language support, or intensive headstands
by an implementer. We looked at some options for that language support.

                  Randy Brukardt.

----------------

Containers Meeting - September 22-23, 2008

Present: Randy Brukardt, Bob Duff, Tucker Taft, Ed Schonberg, Matt Heaney.


Who is the customer? "The programmer of the future".

Why would someone not be able to use the existing containers? Restricted
environments (no dynamic allocation, no finalization). Synchronization issues.

Programs can add synchronization, so that is a secondary issue.

Another dimension is limitedness.

Limited,
No dynamic allocation
No finalization
Synchronization

New container types: achieve parity with C++ (multimaps, multisets).

Surely No dynamic allocation is the most important issue.

There is a meta-rule that the containers can be written in Ada.

Tucker suggests that setting an address with the tampering bits could be use add
detection of copying of these bits.

Tucker is trying to order these kinds of containers.

No dynamic allocation
No controlled types

Turning to protected (task-safe) containers. Bob wonders if there is any real
value to wrapping these containers? But the shared queue is commonly wanted.
Perhaps we should just invent a shared queue abstraction and provide that.

Tucker wonders if a generic that protected a set of operations and provided a
protected operation would work. The idea is the underlying storage would be
provided by the formal, so that alternative storage mechanism could be used.

We talk about limited containers. That brings up issues about control of
finalization. The current containers do not have that. For instance, how can we
use a constructor function in a limited bounded form. C++ has placement new for
this. Tucker thinks that you could play games with storage pools to do that.

We talk about bounded hash table. Matt wonders how you set the hash table size
here, because it can't be calculated after the fact. Tucker suggests two
discriminants (or formal parameter). Bob suggests that one could use a
constructor function to calculate the discriminants (which could be private).

Type Hash_Objects (Hash_Size, Element_Size : Integer) is record
      H : Hash_Table (1..Hash_Size);
      E : Element_Table (1..Element_Size); End record;

Should these be discriminants? Per-object boundedness seems valuable. The
discriminants would be visible, and you would need a subtype in order to be
compose these (container of container).

Thus, the bounded forms, will not be based on storage pools, just based on
discriminants, specified on the object.

Matt suggests another model where the memory is allocated in the package body.
That seems too much like a heap implementation. Per-object makes much more
sense.

If we had done it on the instantiation, we would have the choice of for one
object or for all objects. But we don't want to do that.

We don't think assignment is too likely, so we believe that mismatching
discriminants is not an issue.

---

Synchronized operations: multiple vs. single readers? Is a queue good enough, or
should we make all of the containers protected? Tucker is against that, as there
are many other things that you are likely to need to do.

We have an aside about transactional memory.

Tucker would like to try separating storage and protection aspects for a
synchronized queue.

---

We go through Matt's list of issues.

We discuss what happens with uninitialized elements. We have covered that for
vectors; they're called empty elements, and accessing them is a bounded error.
We should just follow this model; and not worry about last wishes.

Bounded forms are non-limited; we'll look at limited forms later and separately.

Ordered multisets: should we have them? Tucker asks this is a "bag"?
Essentially, yes. It contains objects, possibly duplicated.

What new operations are there? There are additional iterators for iterating over
an equivalence set.

Bob wonders if this makes more sense for multi-maps.

Tucker wonders if Generic_Keys could be used for that; it offers a second
equivalence class. Ed doesn't seem to understand what this is used for.

Matt says that you end up having to have a second-level indirection to use the
existing containers this way.

A.18.8(87/2) shows that it is not intended for Generic_Keys to establish some
other equivalence class. So Tucker's idea won't work.

The difference between a set and a map is that the key exists in the data in the
set, while it is separate in map.

So lets look at a multimap. (That's a one-to-many mapping). Even better, let's
have lunch.

----

Matt explains his uses for this Multimap/Multiset package. It still doesn't
sound that fundamental.

Tucker now suggests breaking the key into multiple pieces (the primary key and a
disambiguating key). Then using Floor and Ceiling to iterate over the subrange.
The uniqueness key would not necessarily participate in the ordering relation
(nor is it very interesting).

So it is obvious that we can use some extra memory to get the effect with the
existing packages. The question is it worth it to have this container? Is it a
lot of extra work (definitionally) compared to the existing containers?

Matt prefers a multiset, others think a multimap is more important. We agree
that this is lower priority than bounded, queues, and even limited is more
important (Tucker notes that limited adds a new capability, multiset is just a
bit easier).

Moving on to "case-insensitive string hash and less operations". GNAT has those.

How do this work? Just apply To_Upper or To_Lower. But which is used matters for
">". Bob notes that Unix programmers would use To_Lower, and everyone else would
use To_Upper. Matt says that he uses To_Lower.

Tucker wonders if implementations would optimize this, that's the main reason to
do it.

Robert Dewar joins us and suggests that we simply try to do whatever C does for
in similar functions (StrCaseCpy). But that doesn't seem to be standard.

We originally dropped these because we thought that using a string map was
better than having extra functions for all four kinds of strings.

Robert says that going to lower is better, because there are lower case
characters that have no upper characters. That gives you weird results.

These are common enough that we should have them. Most uses would not care about
the corner cases, and we should not let good enough to be killed off by best.

No promises that Hash(Lower(Key)) = Insens_Hash(Key). Robert disagrees.

Matt offers the question of storage optimized form. (He specifically mentions
singly linked lists). Tucker says that they are just in the way most of the
time. It's not that much of a savings in normal usage.

Robert talks about a pointer that is always the xor of the forward/backward
list. This scheme allows implementing a doubly linked list with only a single
pointer. We think it might be possible to implement the existing package that
way. Which means that potentially there is no savings at all from a singly
linked list package.

We don't think that storage optimized forms are that important. These would only
be necessary on the bleeding edge, and there is no problem switching the
implementation on their own. Do C++ have competing STLs based on performance
characteristics?

Moving to "more sort algorithms". Matt uses Heap Sort because it is stable and
predictable.

Bob wonders if we should have other sorts of sorting. He mentions a radix sort.
That is not worthwhile. But this has little to do with generic that Matt
proposed, that seems to be worthwhile. But it does not have much to do with
anonymous arrays, so drop that part from the name. (That is, it is just
"Generic_Sorting").

Recursive data structures: Tucker again thinks its harder to figure out how to
instantiate or use these packages than to write them yourself. No one disagrees.

Should we be considering lock-free algorithms? That really belongs to a
multicore group; we don't have the expertise for that. We really ought to have a
multicore research group. If we are doing an Amendment, hopefully the ARG will
set up one of those.

We turn to discussion of "smart pointers".

Function F return T_Access is
    X : Auto_Ptr := Init(new T);
Begin
    ...
    Return To_Access(X);
End F;

The idea is that it is a generic wrapper that adds this functionality.

Tucker thinks that this belongs to the storage management realm, and this is
really outside of the scope of this group. He's interested in it, but agrees
with Randy's (old) opinion that this is probably the wrong level of abstraction.
The ARG is already looking at this (lightly).

We talk about lifetime of pointers. Randy talks about his problems in Claw with
pointers that you want to have a defined lifetime and no more (that is, no
copying).

Tucker brings up another dimension: persistence. Passive partitions have this
feature, but that was not the original idea of them. Anyway, pointers into the
persistent area need lifetime information so we can find out when an object has
become unreferenced by the program.

Robert suggests using reference counts for this, but that doesn't work if you
want to swap out a whole page: you'd have to find all of the reference counts
that point in.

Matt mentions iterator syntax "for each obj in collection". Bob mentions "Sather
iterators"; Ed tries to find more about that on the web. It's in the ACM portal.
In any case, this is an area of interest for future work. Randy notes he floated
a proposal in 2005 (AC-0112).

---

We look at a bounded vector specification. Should we still have the Capacity and
Reserve_Capacity? The function Capacity should be available for consistency.
Reserve_Capacity should be gone, it makes no sense.

But there is some sense to it; just ignored if smaller or same, raises
Storage_Error if bigger. That's the same semantics as the Unbounded case, with
less restrictions.

Tucker asks what is the capacity of the vector objects that are returned from
functions. Since these are constants, the smallest possible is fine. That would
be annoying if it is later assigned.

Tucker suggests something, but it doesn't work.

Also, there is no need to constrain the discriminants subtype.

Matt has Assign, which allows assigning objects with different capacity. That
makes sense, some exception will be raised if the source doesn't fit in the
target. Do nothing if Source = Target.

He also has Copy, which changes the capacity of a object on the fly.

What about Move? It is a destruction Copy; implement by copying Source to Target
then zeroing Source.

Buglet: Move should say something about the Capacity of Source afterwards. It
could be zero. Bob suggests that we say that Reserve_Capacity (Source, 0) is
called (that allows rounding up). This should have another AI (it needs to be
done for all of the containers).

Bob thinks he doesn't want Storage_Error raised here, as this is not running out
of all memory. That is a big difference here, especially since the user can
understand when they have run out. He thinks that it is very similar to a range
1..10, and talking about 11. He would want to raise Constraint_Error.

Tucker suggests that this is similar to Storage_Size on an access type. That
raises Storage_Error even though memory is not exhausted.

Tucker also suggests defining Capacity_Error.

Merge is looked at, but it is decided that there is no problem.

Moving to Bounded_Doubly_Linked_Lists.

The list has to be implemented with indices, as pointers don't work in
assignment. That has to be bitwise copy (remember, no controlled types are
desired).

Matt shows us a poor man's initialization and finalization scheme, we decided we
don't want it.

Splice makes sense, but it probably requires some changes.

Merge also makes sense and should be included.

Maps. Matt wants the hash table to be a prime number. If you allow specifying
it, that no longer works. It should be OK for this to be true.

Function Init (Capacity : Count_Type; Hash_Table_Length : Hash_Type := 0)
   return Map;

The problem is that these things are not very static looking (and probably not
static at all; it would be on the secondary stack at best).

So Type Map (Capacity : Count_Type; Hash_Table_Length : Hash_Type);

Function Default_Hash_Table_Length (Capacity : Count_Type) return Hash_Type;

No restrictions on hash table length, because we are using a bucket
implementation.

Do not need a function Hash_Table_Length, using that would lock them into the
bounded form (and they can read the discriminant anyway.

Tucker suggests "Modulus" to "Hash_Type_Length". The current rules don't say how
the hash value is mapped to the hash table, so this seems like
overspecification. Others think it is good enough. The function then is
Default_Modulus. Put the function at the bottom.

Assign and Copy are in all of the packages. Copy works like Reserve_Capacity;
the hash table also gets shrunk, calling Default_Modulus to figure out the
appropriate hash table size. Copy gets an extra parameter Modulus, use
Default_Modulus if it is zero.

Ordered_Maps.

Looks pretty similar.

Should we add Assign and Copy to the original containers? We should plan to do
it "next time", whenever that is.

To_Set would decide its capacity, it does not get a new parameters.

Hashed_Maps.

Add Default_Modulus, etc. Otherwise works like the Hashed_Maps.

Robert leaves us.

----

We talk a bit about indefinite, bounded forms, where the elements come a storage
pool and the rest of the "control" stuff is bounded and allocated statically.
Randy had previously suggested that, because the unbounded forms allocate all
kinds of control junk, and that would complicate a storage pool that is tailored
to a particular element type. But the idea seems pretty complex, though, and the
users that want this probably wouldn't like the complexity.

----

End of day 1.

Day 2. On to the queue abstraction.

----

Generic
   Type Element_Type is private;
Package Ada.Containers.Queues is
   Type Queue (Capacity : Count_Type) is ...
   Procedure Enqueue (Q : in out Queue; New_Item : in Element);
   Pragma Implemented (Enqueue, By_Entry);
   Procedure Dequeue (Q : in out Queue; Item : out Element);
   Pragma Implemented (Dequeue, By_Entry);
   Function Is_Empty (Q : Queue) return Boolean;
   Function Is_Full (Q : Queue);
   Function Count (Q : Queue) return Count_Type is abstract; End Ada.Containers.Queues;



Dequeuing from an empty queue generally waits; Enqueuing to a full (bounded)
queue, then you wait. (You could also raise an exception in both cases.)

Generic
   Type Element_Type is private;
Package Ada.Containers.Queues is
   Type Queue is synchronized interface;
   Procedure Enqueue (Q : in out Queue; New_Item : in Element_Type) is abstract;
   Procedure Dequeue (Q : in out Queue; Item : out Element_Type) is abstract;
   Function Is_Empty (Q : Queue) return Boolean is abstract;
   Function Is_Full (Q : Queue) is abstract; End Ada.Containers.Queues;

Generic
Package Ada.Containers.Queues.Unbounded Is
   Type Queue is synchronized new Queues.Queue with private;
   Procedure Enqueue (Q : in out Queue; New_Item : in Element_Type);
   Procedure Dequeue (Q : in out Queue; Item : out Element_Type);
   Function Is_Empty (Q : Queue) return Boolean;
   Function Is_Full (Q : Queue);
Private
   Protected type Queue is
      Procedure Enqueue (Item : Element_Type); -- Entry for bounded cases.
      Entry Dequeue (Item : out Element_Type);
      Function Is_Entry return Boolean;
   Private
      L : List;
   End Queue;
   Function Is_Full (Q : Queue);
End Ada.Containers.Queues.Unbounded;

Bob doesn't want this interface, because it is different than the current design
of the containers. We could have done that for all of them. Tucker says he
wanted something like that, but he didn't get it.

We could add child signature packages, even to the original (especially for Sets
and Maps to extract the commonality).

But here we have an interface. Bob thinks we should use a signature here.

Matt says that there is penalty here for the use of the interface; two
instantiations. The solution would to be write a wrapper generic, that should
cut it to one instance for simple uses.

Generic
Package Ada.Containers.Queues.Bounded Is
   Type Queue (Capacity : Count_Type) is
       synchronized new Queues.Queue with private;
   Procedure Enqueue (Q : in out Queue; New_Item : in Element_Type);
   Procedure Dequeue (Q : in out Queue; Item : out Element_Type);
   Function Is_Empty (Q : Queue) return Boolean;
   Function Is_Full (Q : Queue);
Private
   Protected type Queue (Capacity : Count) is
      Procedure Enqueue (Item : Element_Type); -- Entry for bounded cases.
      Entry Dequeue (Item : out Element_Type);
      Function Is_Entry return Boolean;
      Function Is_Full return Boolean;
   Private
      A : Element_Array_Type(1..Capacity);
      Front, Back : Natural;
   End Queue;
End Ada.Containers.Queues.Bounded;


Bob asks about a signature package; he says it has less overhead than an
interface. But we decide it is not necessary because the overhead could be
avoided by declaring a generic with a formal derived type:

Generic
   Type T is new Inst.Queues.Queue with private; Package

Bob worries that copying large objects could be expensive. He thinks about
callback to initialize. Tucker worries that the callback would be in a protected
action, that would greatly limit what could be done. Matt notes that the point
here is to pass an object to the container.

The interface package should be Pure, the unbounded packages should be
Preelaborate, the types should NOT have preelaborable initialization. The
bounded packages (including yesterdays) should have Pure. (We don't want any
pointers in the implementation).

Lock-free queues. Bob says when we get that figured out we should next talk
about bug-free queues.

We should be able to implement this interface with a lock-free algorithm
(presuming the use implementation-defined constructs for the implementation of
the package). That (for now at least) would be in an implementation-defined
subpackage.

----

Limited Containers

An old Bob Duff design for Ada 95:

Lookup_Ptr (Map, Key) return Element_Ptr;

Lookup_CPtr (Map, Key) return Element_CPtr --(access-constant)

Lookup_Optional_Ptr (Map, Key) return Element_Ptr;

Insert_Ptr (Map, Key) return Element_Ptr;

Lookup_and_Insert_Ptr (Map, Key) return Element_Ptr;

These are all functions, they use a Rosen trick/access parameter to handle
updates to the Map.

The idea is rename the returned value:

          O : Element renames Lookup_Ptr (Map, Key).all;

---

We turn to Matt's proposal:

Generic
    Type Element_Type is limited private; Package Limited_Hash_Tables is

    Type Map is tagged limited private;

    Function Insert (M : not null access Map; K : Key)
          return not null access Element_Type;

    Procedure Insert (M : in out Map;
                      K : Key;
                      New_Item : not null access function return Element_Type);
       -- The function is for initialization.


Is vector useful for limited elems? Assume Element_Type is such that Bitwise
copy isn't meaningful (e.g. it has internal ptrs) so make it limited. How do you
make a stack of these things?

When you pop the stack, you want the effect of Unchecked_Deallocation on the
object; and the effect of new on pushing the stack.

Tucker suggests that the container itself is a storage pool with additional
operations

Type Stack is new Root_Storage_Pool with record
    Data : array (...) of Element_Type;
    Top : Int := 0;
End record;

Tucker thinks you would need to export a package with a single object of this
type.

Inst (Element_Type => Scope_Descriptor); Obj : Inst.Stack; Type Scope_Ptr is
access Scope_Descriptor; For Scope_Ptr'Storage_Pool use Obj;

Then new would add an object to the top of the stack, and Unchecked_Deallocation
would pop the item from the stack.

Matt wonders:
   Type Pool_Type (EA : access Storage_Element_Array) is new Root_Storage_Pool ...

   Type Stack is record
      EA : Storage_Element_Array (1..
10*Element_Type'Max_Storage_Size_in_elements);

   Procedure Push (S : Stack) is
      Pool : Pool_Type (S.EA);
      Type Acc is access Element_Type;
      For Acc'Storage_Pool use Pool;
   Begin


But this doesn't work because the object will be finalized when the access type
goes away.

Tucker is trying to figure out a fix for that. We decide to look at his subpool
proposal in detail.

Storage Pool  <-->access type 1, access type 2.
  (subpool)   <--> object.

The idea is to allow a bunch of subpools which have a (potentially) shorter
lifetime than the whole pool.

For instance, a hash table defined globally, with a local object.

    Procedure...

       Subpool <-- HT (access Subpool)
       HT1 : HT (new Subpool); or
             HT (Subpool1'access);

Note that you can't use that now, only in a local access type. So we add
Storage_Pool to new (per-object):

Procedure Insert (H : HT; E : Element_Type) is
   New HT1.SP Node'(...E...);

This works, but is horribly unsafe.

So we add some checks. Goal: Ptr must not outlive designated object.

Tucker wants to do the checking at the subpool level, not the individual object.
So the goal becomes Subpool or stack frame (P) containing Ptr must not outlive
subpool containing designated object (D).

Rule: Check when storing ptr into object resident in subpool (T for target) that
T does not outlive D.

Temp := new (HT1.SP) Node'();
Heap_Obj.X := temp;

Access value must be able to determine what subpool that contains the designated
object. This is an additional requirement on subpools (need a way to get from
the address to the pool, implementer must make that work)

Heap_obj is in T, Temp is in D.

Assert (Does_Not_Outlive(T, D));

Tucker is proposing May_Reference (A, B). This creates a partial ordering on the
lifetime of the subpools. This creates a dependence (a counted reference) from
subpool A to subpool B. The finalization of a subpool involves checking that the
reference count of a pool is zero (but he also wants to detect cycles, all
finalizing at once). A subpool handle is a counted reference (it accesses the
subpool); no direct access to subpools. The subpool is finalized when its counts
gets to zero (modulo cycles).

A subpool handle is intended to be used for checks for stack frames. (For
instance, you can't return one of these pointers.) Also need to check on return
that caller has handle on D.

May_Reference and Does_Not_Outlive are provided by the implementation, so that
they can be optimized out in many cases. The data would part of the private part
of subpools.

Tucker hopes that this totally eliminates the possibility of dangling
references; the subpool sticks around until the (inter-pool) pointers go away.

The purpose is to allow cleaning up sets of data where the type is globally
defined.

Each subpool is master, both for tasks and finalization. Whatever task finds the
last subpool is unreferenced is the one that would do the cleanup.

Weak references (subpool+object pointer): can get either null or a strong
reference to a subpool handle. There also needs to be way to destroy the strong
references when they are not needed.

Weak references need to be managed by the storage pool. A strong reference would
a controlled object to decrement the reference count.

Anonymous access parameter issues. How does that work; perhaps you would want to
have master rather than a level for that check. But we had previously said that
we wanted to preserve the model. Tucker suggests that we could give them a very
deep level (as with anonymous access-to-subprogram), so that they could be
passed but not converted to anything else.

One could use Unchecked_Deallocation for finalizing, even though it isn't really
Unchecked nor would it Deallocate. We'd only be using the finalization
side-effect.

----

We turn to looking at control of initialization and finalization.

X : LT := F(...);
For X'Address use Constant_Value;
For X'Master use Master_Obj; -- Object is user defined, master object.

In this case, initialization happens at the point of X, but it is initialized
into memory previously allocated. The new thing is to be able to specify the
master of X (so it doesn't finalize too early).

For a bounded hash table:

Procedure Insert (HT : in out Hash_Type; Key : in Key_Type) is

X : Element_Type := F(...);
For X'Address use HT.Element_Array(I);
For X'Master use HT.Master; -- Or existing HT'Master

Element_Type'Finalize_All (HT.Element_Array(I));

This is unsafe, because we don't know if the original location is already
initialized. In any case, we have to suppress the initialization on the
original object (else it would be initialized twice).

Pragma Suppress_Default_Initialize (HT.Element_Array);

Element_Type'Initialize_All (HT.Element_Array(I), HT'Master, F(...));

(The last parameter is optional.)

Do we need to pass master to T'Finalize_All? Probably not,
Unchecked_Deallocation does not need it.

Bob doesn't want the overhead of using a linked list for the finalization of an
array of items. That's necessary with this model, because you can't know how the
code is going to call T'Initialize_All. He would like the master to be
user-defined.

Do they get extra space for dope? That's hard to do, because we don't know a
size.

Element_Type'Initialize_All (HT.Element_Array(I), [F(...)]);

Function T'Initialize_All (P : access ET; Initial_Value : ET);
  -- We use access ET because we want to worry about by-copy types; P should be by-reference.
  -- Also could use for pointer math.

But you could Unchecked_Convert to access ET, then pass UC.all.

You would have to apply Suppress_Default_Init before using T'Initialize_All,
else the code is erroneous.

Bob suggests pragma Suppress_Default_Init be applied to types. And then the
Initialize_All and Finalize_All could be called for that type (otherwise, we
wouldn't allow it).

Tucker notes that you define a derivation of Element_Type to make Bob's idea work.

It is erroneous to access an uninitialized object (that is, before the call to
Initialize_All. Tucker notes that you could even add a bit to detect this
problem.

Indeed, one could imagine a type with pragma SDI as an Unchecked_Union of a
bag-o-bits and Element_Type. Then actually an assignment would do the right
thing (but of course that doesn't work for build-in-place). And if you stored
the discriminant, you could actually check for double initialization and
finalization.

The model is that T'Initialize_All is an assignment to a record where we're
changing the discriminants from bag-o-bits to Element_Type, T'Finalize_All goes
in the reverse.

Logically:

    Type ET (Initted : Boolean := False) is record
       Case Initted is
          When False => Bag_o_Bits : Storage_Element_Array (1 .. Right_Size);
          When True => Object : Element_Type;
       End case;
    End record;

Goal, if T'Finalize_All is a no-op (because there is no controlled components),
then there should be nothing put on any lists.

Finalize_All for Obj w. components that suppress_default_init. If
Unchecked_Deallocation needs to remove these, it depends on whether it is
on a the list. Randy says it works the same as now, if it is on a list.

Lots more brainstorming goes on, but I gave up making notes for a while.

There is some discussion about only doing this for types that do not have
controlled/task/protected components.

Bob thinks that the initialized bit ought to exist (with the possibility of
suppressing it), and then the routines should check the bit.

Matt suggests possibly using a magical generic for this purpose. He is trying to
make a way to get to C++ placement new.

Package P is new Magic(Element_Type);
Type EAT is array () of P.T;
HT  Arr : EAT;
Ptr := new (HT.Arr(I) Element_Type'(...); Ptr := New_Object (HT.Arr(I));
Delete_Object (HT.Arr(I));

Tucker has to leave to make his train, so the meeting ends at 4:15 p.m.

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

From: Tucker Taft
Sent: Thursday, September 25, 2008  7:16 PM

Thanks, Randy, for taking notes.

You mention these below in passing, but I believe we (tentatively, of course)
agreed on having case-insensitive string hash and comparison, using To_Lower as
the "folding" before comparison (by the way, I finally found a definitive
statement that that is what the C-based standard requires).

I think we also tentatively agreed to provide a Generic_Sort routine that takes
Swap and comparison formal subprograms, without actually taking the array itself
as a parameter.

Thanks again for the notes.

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

From: Matthew Heaney
Sent: Monday, September 29, 2008  5:55 PM

Minutes for containers meeting at AdaCore on 22-3 Sep 2008

Sep 22-3 2008 (Mon, Tue)

Q: What is our goal for this meeting?

A: Greater demand for restricted ("bounded") forms, rather than matching what's
in the C++ STL.  (There has been actual demand for bounded forms, so that should
have priority.)

In bounded forms, there is the issue of initialization and finalization of
elements.  The objects don't actually get created and destroyed, so if an
element is deleted from a container, and then re-inserted, what value should the
newly-inserted object have?  One solution is to use vector semantics.

[Similar to a bounded form, a vector has storage for elements that doesn't get
deallocated (at least immediately).  For that container, there are a few
insertion forms.  The first form accepts a new value as a parameter, so that's
the value that a newly-inserted element has.  The second form omits a new value,
and in that case a local variable having a default value is assigned to the
newly-inserted element, so in that case the element has the default value of the
type.  In the third case, there's a distinct Insert_Space operation that simply
opens up that slot in the internal array, the intent of which is for the caller
to make a separate call to Replace_Element to properly assign a value to the
newly-inserted element. --MJH]


Can you use a storage pool hack to initialize and finalized storage that is
statically allocated?  No, because when the scope of the (local) access type
ends, it will finalize the storage associated with that access type.  [on Tue we
discussed ways the language could be modified to support such a feature.  --MJH]

AI05-107  deallocation issues?

AI05-0069-1  ??

"standards track technical report"  - for this committee.  MJH has the task of
actually writing the TR.

Next ARG meeting: Oct 30-1 (Thu, Fri), Nov 1-2 (Sat, Sun).  WG9 meeting on Thu
afternoon.  [I will be attending the meeting on Sat-Sun.  --MJH]

Incorporate AI-69 into TR?

AI05-0001-1  (describes need for bounded forms?)

TR: will eventually become a standard.

Synchronized forms - no clear meaning of what a "protected form" even is.

Randy said that there was a discussion about this issue on CLA.  Lock for a
single container vs. lock shared among several containers.

Composibility issues. e.g. set of sets.

Locking mechanism: semaphore (Seize, Release) vs. monitor

Canada: our main customer?

What not let user wrap container in a PO himself?

Dimensions of our problem space:
limited-ness
controlled-ness
dynamic allocation
synchronized
new containers
other non-container features
space-optimized forms
persistance

Issue with tampering bits: there is a pathological condition in which assignment
of the container is made in a tampering context.  The problem is that the
tampering bits are set, and these get copied to the object on the LHS of the
assignment.  Without controlled-ness, there's no way to reset these bits.

However, the goal is to not require that the bounded forms be implemented as a
private derviation from controlled.  [Do we require that controlled-ness cannot
be used? --MJH]  So we could decide that this is a bounded error. The
implementation might be able to detect this state and raise an exception.  Or it
could detect this state and automatically reset the bits. (There's a trick you
can use: with an access component that is set to point to the current instance
of the type.  If you notice that the access value doesn't match the current
instance, then assignment has occured and you can reset the bits.  Alternatively
you could set the access value when you enter the tampering context.)

thread-safety vs. performance: if we can't do it better than a user can do
himself, then why do it?

operations are not potentially blocking [does this refer to synchronized
forms???]

Since we can't decide what a protected container is, we can punt and just supply
a (bounded) queue, which is useful all by itself as a stand-alone abstraction.
One possibility is to write a generic queue that imports the underlying
structure (e.g. a bounded vector, or unblunded list) as a generic formal type.
But then do we supply pre-defined instantiations?  [See notes for Tue, where we
tenatively decided on what the spec should look like. --MJH]

Other abstractions for synchronization: barriers.  Not enough complexity here to
standardize?

(Limited) containers with limited elements: how to finalize elements in a
bounded form?

What does "new T" mean?  Does this count as dynamic allocation, even if (say)
the access type is associated with a storage pool that is itself bounded?  [I
think the answer is yes, because (in GNAT at least) the compiler handles a
restrictions pragma by noting where the allocators are used.  So in order to
satisfy the requirement for "no dynamic allocation" we must use a bounded form
(the type is implemented with an array component) instead of importing some
storage pool as a generic formal object.]

Bounded hash table: issue of how to declare number of elements vs. size of
backbone.  Backbone size: specified via discriminant vs. generic formal
parameter.

Secondary stack
discrim vs. generic formal
instantiation provides defaults (but tagged types cannot have defaults for
discrim)

MJH argues that it would be nice if the user didn't have to specify backbone
size (since he might specify a value that doesn't scatter hash values very
well).  However the consensus is that we should allow user to control both
number of elements and backbone size.

Bounded forms: do not use a storage pool.  Use per-object storage rather than a
"pool" shared among instances.  Per-object storage is specified via discriminant
(rather than generic formal parameter).

Note that assignement of bounded forms will cause CE if discrims don't match.

Corrigendum doesn't do much for users; amendment is more useful.

Tuck's storage pool for persistence AI.

John Barnes/American customers: US decision for proper amendment?

What is role of ARG: to make amendment or corrigendum?

Protected forms: multiple readers?  Or only provide queues?  Which operations
are waiting?

Linda language:  has "tuple spaces"

Thread-safe vs. atomicity: atomicity is more useful.

Ed S. mentioned conference at NYU(?) about hardware-level abort/rety

synchronized queue

Vector has "empty" elements: created as a result of Insert_Space.  It's a
bounded error to read from them, so you can't assign them a value via
Update_Element (e.g. in indefinite form you don't even have an element), so you
must use Replace_Element.

Trick for (re)initializing a statically-allocated element: use pragma Import to
fake placement new.

How to grant last wishes: insert nothing (default init), insert item, insert
space.  Can potentially raise exception if touch "empty" element.

non-limited bounded forms: are not controlled

multiset - same as bag

Tucker wants MJH to supply an example that demonstrates how a problem is more
easily solved using a multiset instead of a existing forms (e.g. map of sets).

To guarantee uniqueness of key you could make a two-part key, with one part just
a number that simply gets incremented (to satisfy uniqueness requirements).

Multiset: how to you designant elements in a subrange.  [Using an additional
passive iterator that accepts a key.  You might also be able to use
floor/ceiling to perform active iteration (but maybe not, since they're
semantics don't match C++ lower_bound/upper bounded).]

Having (limited) container of limited elements is more important that a
multiset.

Case-insensitive string hash, string equality, string less: defined in terms of
To_Upper or To_Lower (answer influences outcome of test)?  In GNAT we use
To_Lower.  _stricmp, _wcsicmp in Microsoft C also uses lowercase compare; gives
example of JOHNSON vs. JOHN_HENRY.  Whatever we do should be the same as what C
does (using 8-bit chars)???

We had originally gotten rid them at the Phoenix meeting.  Specify map function
to specify folding?  But this is more than we really need.

We have tenatively decided that they're a good idea: use same names as in
GNAT:

Ada.Strings.Equal_Case_Insensitive
Ada.Strings.Hash_Case_Insensitive
Ada.Strings.Less_Case_Insensitive

and to fold to lower case.

We don't promise that case-insensitive hash returns same result as
case-sensitive hash will all lower-case letters (because implementation could
use knowledge to make an optimal hash function).

It would be possible to make a space-optimized doubly-linked list by xor'ing
prev and next, but in order to navigate you need to refer to a pair of nodes.

Cursors: for vector cursor can be "ambiguous" (see RM p.243) vs. "invalid"
cursor.

Radix sort?  (But heap sort is probably good enough.  "Heap" came from Knuth,
but this is same as Floyd's "tree" sort.)

Should be have a sort algorithm for an "anonymous" (my term) array?  Yes, here's
what the spec should look like:

generic
   type Index_Type is (<>);
   with function Less (Left, Right : Index_Type) return Boolean is <>;
   with procedure Swap (Left, Right : Index_Type) is <>;

procedure Ada.Containers.Generic_Sort
  (First, Last : Index_Type'Base);

pragma Pure (Ada.Containers.Generic_Sort);  --??? gnat-specific?


"lock-free container": multi-core extensions for Ada; to be discussed at ARG
meeting in Portland.

lock-free algorithms
transactions
multi-core


auto_ptr abstraction for Ada.  Outside charter of this group, but it's part of
"storage management issues" to be disussed by ARG (in Portland?).


Persistance: via passive partition or other?
pointer-swizzling


language extensions for iteration:

e.g. vbscript

  Set objCollection = ...
  For Each objItem in objCollection
     'manipulate objItem
  Next

Sather - nice iteration syntax
CLU
Randy's proposal: see Ada-Comment on 12 April 2005 (AC-112)


super-null arrays

Spec of bounded forms: one of our goals is to minimize changes to spec of
unbounded form, so we keep some functions that don't necessarily make sense for
bounded forms.

The bounded non-hashed forms are declared this way (pass capacity as
discrim):

   type Map (Capacity : Count_Type) is tagged private;

The bounded hashed forms are declared this way (pass both capacity and hash
table length):

   type Map (Capacity : Count_Type; Modulus : Hash_Type) is tagged private;

Capacity function - keep this to support non-prefix function calls; if prefix
form is used, language has preference for discriminant so we're OK.

Keep Reserve_Capacity, since we can follow same semantics as for unbounded form.
One issue (to be resolved) is what exception to propagate if user specifies
capacity parameter larger than capacity of bounded container.  My examples in
the jean repository raise Storage_Error, but there was some debate about whether
some other exception (e.g. Constraint_Error) would be better.

Bounded vector: defined in terms of unbounded vector.

RM language for Move operation of unbounded forms: must state that capacity of
source container is allowed to change.  Should be described in terms of calling
Reserve_Capacity with 0 capacity as value.

Don't worry about tamper bits in bounded form (when assigning in tampering
context), since it's a bounded error.

The vector, list, and ordered containers have the following pair of
operations:

   procedure Assign (Target : in out Vector; Source : Vector);

The semantics of Assign are as follows: if Target denotes same object as Source,
then do nothing.

   function Copy (Source : Vector; Capacity : Count_Type := 0) return Vector;

The semantics of Copy are as follows: if capacity is larger than source
container's length, than use that value as the capacity of LHS.  Otherwise, use
length of source as capacity.

Hashed forms look like this:

   function Copy (Container : Map;
                  Capacity : Count_Type := 0;
                  Modulus  : Hash_Type := 0) return Map;

As with other bounded forms, Capacity=0 means use container.length as capacity
of target.  Modulos=0 means use default_modulous(container.length).

There's also a function (declared at end of spec):

   function Default_Modulus (Capacity : Count_Type) return Hash_Type;

This returns a "good" value for the backbone length for a given capacity.

Shared pool idea: either generic formal object (of type Storage_Pool'Class) or
nodes in package body; see adalib list on 2006/12/10 & 12.

Bounded forms have Pure categorization.

Brad Moore - has API with heavy use of synchronized interfaces


Tue Sep 23 2008

Blocking queue - the minimum we provide for a "protected container."

bounded vs. unbounded
signatures vs. interface

dequeue from empty
enqueue to full


We discussed various ways of implementing a queue.  We finally settled on
using an abstract queue interface, with a pair of concrete children,
reasoning that the interface could be used to implement class-wide message
passing algorithms (see Andy Welling et al):

generic
   type Element_Type is private;

package Ada.Containers.Queues is
   pragma Pure;

   type Queue is synchronized interface;

   --??? param names OK?
   procedure Enqueue (Q : in out Queue; New_Item : Element_Type) is
abstract;
   pragma Implemented (Enqueue, By_Entry);  -- ???
   -- see RM 9.5

   procedure Deque (Q : in out Queue; Element : out Element_Type) is
abstract;
   pragma Implemented (Dequeue, By_Entry);  -- ???

   function Is_Empty (Q : Queue) return Boolean is abstract;

   function Is_Full (Q : Queue) return Boolean is abstract;

   function Length (Q : Queue) return Count_Type is abstract;
   --??? do we need this?  Does this over-specify the interface?

   -- ??? other ops

end Ada.Containers.Queues;

generic
package Queues.Bounded is  --??? should this be Generic_Bounded?
   pragma Pure;  -- ???

   type Queue (Capacity : Count_Type) is synchronized new Queues.Queue with
private;

private

   protected type Queue (Capacity : Count_Type) is ...;

end Queues.Bounded;


generic
package Queues.Unbounded is  --??? should this be Generic_Unbounded?
   pragma Preelaborate;  -- ???

   type Queue is synchronized new Queues.Queue with private;

private

   protected type Queue is ...;

end Queues.Unbounded;


We disussed limited containers.  Bob D. and Tucker T. have an limited hash
table that looks something like:

generic
   type Element_Type is limited private;
   type Key_Type is private;

package Limited_Hash_Tables  -- per Bob Duff and Tucker

  type Map is tagged limited private;

  -- Ada95 version:
  function Insert (M : Map; Key : Key_Type) return Elem_Ptr;

  -- Ada05 version:
  function Insert (M : not null access Map; Key : KT)
     return not null access ET;

  -- another idea to initialize element
  procedure Insert (M : in out Map;
                    K : KT;
                    New_Item : not null access function Init return ET);

  Lookup_Ptr (Map; Key) return Elem_Ptr;
  Lookup_CPtr (Map; Key) return Elem_Const_Ptr;
  Lookup_Optimal_Ptr (Map; Key) return Elem_Ptr
  Insert_PTr return Elem_Ptr;
  Insert_Or_Lookup_Ptr (Map; Key) return Elem_Ptr;

  -- vector - not so useful for limited elems?
  -- ET: bitwise copy isn't meaningful (e.g. it has internal ptrs)
  --     so make it limited
  -- how to handle the case when item is inserted in container
  --                                     deleted from container

private

   type Map is ...;

end Limited_Hash_Tables;


Discussion of limited containers led to a discussion about initialization
and finalization of limited elements.  This led to a discussion abount
memory management and initialize/finalization issues.

We have identified two distinct issues.  Tucker often creates long-lived
data structures, that get populated but then the same elements remain for
the life of the container.  He argues that allocation and deallocation of
nodes is at the wrong granularity, and he would like everything in the
container to be finalized and then deallocated together.

The other issue (identified by Bob and Matt) is initialization and
finalization of the "inactive" elements in a container.  This mostly applies
to bounded forms (it also applies to an unbounded vector with more capacity
than length).  In a bounded form, the storage for elements isn't used
immediately -- it's only used when an element is inserted into the container
(made "active").  The issue is that if the element type has default
initialization, or it's controlled, then when the container object is
declared, there is a penalty because all of the (inactive) elements must be
initialized.  It would be nice if there were a way to defer initialization
of the element until it is actually inserted.

We spent the remainder of the day brainstorming changes to the language to
address these issues.  What follows is a rough sketch of what we discussed.

Here are a couple of examples of binding a hash table to a pool:

 type HT (Pool : not null access Subpool_Type) is ...;

 H : HT(new Subpool_Type);
 H2 : HT(Subpool_Object'Access);

The memory management infrastructure would have to be able to distinguish
between the case of allocating a new pool object (is this an example of
"coextension"?) and binding to an existing pool object.

You can get something like placement new by supplying the pool as an
argument:

procedure Insert (H : in out HT; K : KT) is
   X : ... := new (H.Pool.all) Node_Type;
   -- no new access type here
begin

Checks: ptr must not outlive designate object (RM 3.10.2).

Subpool containing ptr must not outlive subpool containing designated object
(D) (really, subpool or stack frame).

Check when storing ptr into subpool (really, obj in subpool).

temp := new (H.Subpool.all) Node_T
             D

heapobj.X := temp

access value -> subpool that contains its designate object

operation: Does_Not_Outlive (T, D)

heapobj.X := temp;
T

T must not outlive D

procedure May_Reference (A, B)
  create dependence between subpools;
  establishes partial order

Finalization of a subpool ensures no references (must detect cycles)

subpool handle -> subpool
   counted reference to subpool


another check on return that caller has handle on D
  procedure May_Ref (A, B)

weak ref = subpool + object_ptr
  return null or counted ref (strong ptr)

with strong ref, object doesn't go away

cache has weak ref - can ask about object
  result is either null (meaning obj has gone away) or strong ref

subpool is master -> finalization of object???

AI-111 Tucker
  per-object storage pools

masters vs. levels
accessibility checks
conversion from anon access type to named type?

static subpools (on stackframe)
  special case of more general mechanism

subpool handle - limited
  can be passed as arg to new

each ht (or stack) is an instance of subpool

[The rest of this stuff is the init/final idea. --MJH]

X : LT := F;
for X'Addr use XYZ;
  but master of X is local
  for X'Master ...;

suppress default init
T'Init_All
T'Finalize_All

stack of fixed length
Finalize_All - to finalize elem

ET'Init_All (H.Array(i), H.Master, F(X, Y)'Access);
ET'Final_All(H.Array(i));  -- to remove from list

pragma Supproess_Default_Init
  creates extra space for links
  always applies to a type
  gives you ET'Init_All and ET'Final_All

wrapper of ET - provides extra bit; to indicate whether init'd.

discriminated record: discrim says whether init'd or not
data is payload

pragma Supress_Discrim_Check - to remove per-elem overhead

model is that ET'Init is assignment to a discrim rec and changes value of
discrim

HT "contains" a master calls Finalize
"gets finalized at same time"

bounded form is sort of like a light-weight controlled object
the HT is controlled because it has controlled elements

what happens to a composite type with a component for which pragma
Suppress_Default_Init is specified?

limited bounded case, yes, but useful for non-limited types too

useful to prevent vector array from init all junk (inactive) elements
e.g. vector instantiated with access type, don't want to null all elements

tucker's discrim type: another option to allow just discrim to change

I had an idea to abstract-away some of support for light-weight
controlled-ness.  Just as we have a pkg like this:

generic
    type Object(<>) is limited private;
package System.Address_To_Access_Conversions is
   pragma Preelaborate(Address_To_Access_Conversions);
   type Object_Pointer is access all Object;
   function To_Pointer(Value : Address) return Object_Pointer;
   function To_Address(Value : Object_Pointer) return Address;
   pragma Convention(Intrinsic, To_Pointer);
   pragma Convention(Intrinsic, To_Address);
end System.Address_To_Access_Conversions;


We could have a pkg something like:

generic
   type Object is limited private;
package System.Initialization is

   type Object_Storage is private;
   -- has size and alignment of type Object
   -- need separate versions for limited Obj vs. non-limited Obj?

   type Object_Pointer is access all Object;
   for Object_Pointer'Storage_Size use 0;

   -- to default-init object
   function New_Object (Storage : not null access Object_Storage)
      return Object_Pointer;

   -- to copy-assign the object
   function New_Object
     (Storage : not null access Object_Storage;
      Value   : Element_Type)
      return Object_Pointer;

   procedure Delete_Object (Storage : in out Object_Storage);

private

   type Object_Storage is ...;

end System.Initialization;

So now you could implement a bounded stack something like:

with System.Initialization;
generic
   type ET is private;
package Generic_Stacks is
   type Stack (Capacity : Count_Type) is tagged private;
   procedure Push (S : in out Stack; E : ET);
   ...
private
   type Element_Array is array (Count_Type range <>)
      of aliased Object_Storage;

   type Stack (Capacity : Count_Type) is tagged record
      Elements : Element_Array (1 .. Capacity);
      Top      : Count_Type := 0;
   end record;
end;

package Generic_Stacks is
   procedure Push (S : in out Stack; E : ET) is
      I : constant Count_Type := S.Top + 1;

      -- this is the part where you might have to pass S'Master, etc.
      P : Object_Pointer := New_Object (S.Elements (I)'Access);
   begin
      P.all := E;
      S.Top := I;
   end;
   ...
end Generic_Stacks;

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

From: Matthew Heaney
Sent: Wednesday, October 22, 2008  12:10 AM

Grein, Christoph (Fa. ESG) wrote:
...

> Is there some place where I can help? (If I find some spare time, of
> course.)

Yes, of course.  The best place is to post to the adalib list.
Otherwise, I can give you developer status in jean repository, which
would allow you to edit (and check-in) sources directly.


> 1.----------------------------------------------------------------------
> [ARG]: Is there anything we want to say about the relationship between
> the range of Index_Type and vector capacity?  It's possible to declare a
> vector object with a capacity smaller than the range of the index
> subtype (and this makes sense for types as Positive, etc), but it's also
> possible to declare a vector object with a capacity larger than the
> range -- but we could never even add those elements, since the maximum
> vector length is determined by the range of the index subtype.  Should
> we allow this? [END ARG]
>
> I think if someone asks for this, he gets what he asked for. A note in
> the AARM (or, if this is only a secondary standard to the RM, in
> something similar) could be added.

OK, that makes sense.  I don't think we made any explicit statement in
the RM to handle this case for the unbounded forms, so maybe we don't
need to worry about it.


> 2.----------------------------------------------------------------------
> Why is there a Reserve_Capacity? It's completely useless.

Actually, I think that it has semantics identical to the unbounded form.
  (But I have to think about it.  Maybe I over-specified, in which case
no discussion would even be necessary.)


> Is it just for
> compatibility with Ada.Containers.Vectors?

Yes.


> There are already operations
> which are not in Ada.Containers.Vectors, so this breaks compatibility,
> so why shouldn't there be some missing?

I suspect that the operations we add to the bounded form will eventually
migrate into the unbounded forms as well.  (I think the only two
operations are Assign and Copy.  I forget -- were there others?)


> 3.----------------------------------------------------------------------
> My lightweight opinion wrt. Constraint_ vs. Storage_Error:
> I think all bounded containers should raise the same exception in such
> cases.
> For vectors: Arrays raise CE; vectors are kind of arrays, so they also
> should raise CE. Bounded strings raise CE.

OK, that's a data point from a HI user.  Reviewing my code I see that I
raise both SE and CE, inconsistently.  Thanks for the tip.

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

From: Randy Brukardt
Sent: Thursday, October 23, 2008  9:55 PM

A quick glance at this did not turn up any wording for Reserve_Capacity for
bounded vectors and lists. I think you need to say somewhere that the capacity
of the container cannot be changed, and an attempt to set it larger than
specified will always fail with Storage_Error (if that what we decided??).
Surely that needs to be mentioned (it's the whole point of bounded). I see you
do mention it for Maps, but then say it is the same as Unbounded. But that's
not quite true; Unbounded is supposed to try to expand the capacity (assuming
it is expanding) and raise an exception only if that is not possible. Bounded
is not supposed to try, just raise an exception for any attempted expansion
(and do nothing for a shrinkage). That is a difference in intent. Surely we
want to explain that somewhere, else these containers will not be obviously
different.

Also note that if Reserve_Capacity is properly documented, changing Insert, etc.
isn't needed, because they all call Reserve_Capacity when they need to expand
(and that will raise an exception).

And I do think you need to mention (but perhaps only in an AARM implementation
note) that the bounded forms are not expected to be controlled types. If there
is any semantic ramifications of that, we need to change the wording as needed.

I didn't read further than that (I have lots of ASIS work still to do).

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

From: Matthew Heaney
Sent: Friday, October 24, 2008  5:51 AM

> A quick glance at this did not turn up any wording for
> Reserve_Capacity for bounded vectors and lists.

For Vector it's on line #105:

"The semantics of Reserve_Capacity is as follows.  If the specified Capacity
is larger than the Container.Capacity, then it raises Storage_Error.
Otherwise, the operation does nothing."

There is none for lists, since the unbounded form doesn't have that.

Bounded hashed map is on line #227:

"The semantics of Reserve_Capacity is as follows.  If the specified Capacity
is larger than the Container.Capacity, then it raises Storage_Error.
Otherwise, the operation does nothing."

For the bounded hashed set it's on line #358:

The semantics of Reserve_Capacity is as follows.  If the specified Capacity
is larger than the Container.Capacity, then it raises Storage_Error.
Otherwise, the operation does nothing.


> I think you need to say somewhere that the capacity of the container
> cannot be changed, and an attempt to set it larger than specified will
> always fail with Storage_Error (if that what we decided??). Surely
> that needs to be mentioned (it's the whole point of bounded).

Is this requirement satisfied by descriptions above?

> I see you do mention it for Maps, but then say it is the same as
> Unbounded.

But that part was non-normative.  It was in the form of a question to be
answered by the ARG during our meeting:

"[ARG]
Actually, I think this has the same behavior as the unbounded form (since
in principle the unbounded form could also raise an exception "if the
specified capacity is larger than the container capacity". Since this API
is just supposed to specify how the bounded form differs from the unbounded
form, then do we need to mention Reserve_Capacity here?
[END ARG]"

> But that's not quite true; Unbounded is supposed to try to expand
> the capacity (assuming it is expanding) and raise an exception only if that
> is not possible. Bounded is not supposed to try, just raise an exception for
> any attempted expansion (and do nothing for a shrinkage). That is a
> difference in intent. Surely we want to explain that somewhere, else these
> containers will not be obviously different.

OK, that's fine by me.  I was trying to minimize the amount of normative
text, but if you think it's more clear to be explicit about behavior for
bounded forms then we can leave it in.

> Also note that if Reserve_Capacity is properly documented, changing Insert,
> etc. isn't needed, because they all call Reserve_Capacity when they need to
> expand (and that will raise an exception).

Good point.  I'll have to review all the spots where Reserve_Capacity is
called indirectly, as a side-effect of insertion, etc.

> And I do think you need to mention (but perhaps only in an AARM
> implementation note) that the bounded forms are not expected to be
> controlled types. If there is any semantic ramifications of that, we need to
> change the wording as needed.

Right, I asked that question on line #14:

"[ARG] We agreed during the meeting on 2008/09/22 that the bounded forms
would not be controlled.  Do we need to say that?  One consequence is
this is that we cannot detect finalization of the container when it's in
a tampering state, ..."

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

From: Randy Brukardt
Sent: Monday, November 10, 2008  9:11 PM

I'm working on the minutes of the Portland meeting, and I have:

"These containers cannot be controlled, because Ada.Finalization is Preelaborated
and these packages are Pure. There probably should be an AARM implementation note
to that effect."

But when looking at this, I remembered that the container types are defined
to "need finalization" (that is, are potentially controlled). We surely don't
want the bounded forms to have this property, so we need explicit wording to
say that the type Vector in Bounded_Vectors does not "need finalization". (That
is, to repeal that part of A.18.2(84/2), as we're planning to inherit everything
except a specific list of changes from the unbounded forms.)

Similar wording is needed for all of the other bounded form container types.
Otherwise, Adam will make us fix it in future. ;-)

I preliminarily put a mention of this into the minutes, but since we didn't talk
about it explicitly in Portland, I wanted to mention it here as well.

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

From: Tucker Taft
Sent: Monday, November 10, 2008  9:22 PM

Good catch.  Yes, we clearly don't want the bounded container types to
"need finalization."

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

From: Randy Brukardt
Sent: Friday, July 3, 2009  8:25 PM

Having finished the minutes draft, I've moved back to inserting the new text
of AI05-0001-1 into the consolidated RM. That process of course forces me
to re-read the AI very carefully, and bugs pop up.

One of the changes is (A.18.3(88/2), procedure Move):

Replace:

If Target denotes the same object as Source, then Move has no effect. Otherwise,
Move first calls Clear (Target). Then, the nodes in Source are moved to Target
(in the original order). The length of Target is set to the length of Source,
and the length of Source is set to 0.

By:

Equivalent to Target.Assign (Source) followed by Source.Clear.


But this replacement is clearly wrong. If Target and Source are the same object,
this new wording would have the effect of clearing the object, while the old wording
does nothing to the object.

We need to say something like:

If Target denotes the same object as Source, then Move has no effect. Otherwise,
equivalent to Target.Assign (Source) followed by Source.Clear.

We need to make similar changes for A.18.4(43/2) and A.18.7(38/2).

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

From: Randy Brukardt
Sent: Monday, July 6, 2009  9:06 PM

There is no wording in Bounded_Ordered_Maps or Bounded_Ordered_Sets to check for
reaching the capacity of the container. (We don't need such wording for the Hashed
forms, as the Reserve_Capacity call does the dirty work.)

For lists, we added the following for this case:

* The allocation of internal storage includes a check that the
  capacity is not exceeded, and Capacity_Error is raised if this
  check fails.

"Allocates a new node" is the wording of A.18.4(45/2) defining what Insert does,
the other insertions are defined by difference from this one. So, for ordered maps,
we should have the following:

* The allocation of a new node includes a check that the
  capacity is not exceeded, and Capacity_Error is raised if this
  check fails.


The wording of ordered sets is different. "Otherwise, Insert adds New_Item to
Container..." I think "adds" is too vague to use in this technical manner, so I
think we need wording like:

* If Insert (or Include) adds an element, a check is made that the
  capacity is not exceeded, and Capacity_Error is raised if this
  check fails.

[I can't find any routines other than Insert or Include that can add an element.
Did I miss any?]

Any comments?

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

From: Ed Schonberg
Sent: Monday, July 6, 2009  11:28 PM

> [I can't find any routines other than Insert or Include that can add an
> element. Did I miss any?]

Same should apply to Union, and the phrasing has to mention "one element or more".

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

From: Randy Brukardt
Sent: Monday, July 6, 2009  11:59 PM

Union doesn't "add" elements, it "inserts" them. One has to presume that means
calling Insert. As such, this wording is correct as written.

It should be noted that the wording of the entire containers library is completely
confused as to whether elements are removed or deleted from containers. That probably
ought to be fixed; I just finished writing a patch-around to deal with that problem
(cursors designating removed elements are not considered invalid, at least if you read
the standard strictly, which only makes "deleted elements" invalid). It probably
would be best to rewrite all of the text that uses the term "deleted" to say "removed"
instead ("removed" is far more common). We'll have to talk about that at a meeting;
based on what it took me to create the AI, we're probably talking about 6-8 hours of
work to do that (creating AI text, Amendment text, and consolidated Standard text);
I have to wonder if this is good use of my time.

If we do that (Intersection and Difference are among those that would need to be changed),
I suppose we should also change Union to say that it "adds" elements, and then this
wording would need to be changed. (It surely would be clearer, because it isn't clear
what lower case "insert" refers to here, and we really do want all of the rules that
apply to adding nodes to apply here.)

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