Re: Package hierarchy. Extensibility.

[Niklas Holsti <niklas.holsti@tidorum.invalid>]

Jarno wrote:
> Niklas Holsti:
> 
>> An interface type that defines the common view of "Points" of 
>> whatever representation? That would allow non-hierarchical and
>> independent implementations of the representations as long as
>> every representation implements this interface.
> 
> Can you please post a simple example? Does interface means
> extendable type (�interface� or �tagged�)?

Yes, I meant types that extend (implement) an interface type. I had 
in fact missed that your original posting did not make these 
"Point" types tagged, that is, you did not ask about a *type* 
hierarchy but only about a *package* hierarchy.

This being the case, I don't see enough information in your 
original posting to answer your question. If these types have 
nothing else in common but their use to represent 2D or 3D points, 
and the existence of conversions between them, there is little 
basis for evaluating your alternatives.

The only reason for using child packages would be to organize your 
package names hierarchically, which is of course nice but 
introduces unnecessary compilation dependencies, especially if you 
declare the "preferable representation" in the common parent 
package. So that much can be answered: to minimize compilation 
dependencies, parent packages should contain only stuff that is 
common (in some sense) to all children.

You also said:

> .. it is desirable for representations to be generic.

By "generic", do you mean that you will have generic units that 
take a formal type parameter to specify which "Point" 
representation they use? Such as:

    generic
       type Point_Type is private;
       <formal operations on Point_Type>
    procedure Rotate_Figure ...

If so, you must ensure that the various kinds of "Point" all 
implement compatible operations, to be associated with the <formal 
operations> in the generic instances.

If you use a tagged type hierarchy or an interface type, operations 
like Rotate_Figure can be class-wide, rather than generic, or can 
avoid the declaration of <formal operations> by using instead the 
primitive operations of the formal type parameter.

> Seems that I need to organize a bunch of generic 
> interconvertible types and still be able to add other
> representations later with minimal cost.

Do you need conversion from any "point" type to any other "point" 
type? That would be a quadratic number of conversions, which is to 
be avoided, of course. If all "point" types implement a common 
interface, you could write a single "constructor" function for each 
"point" type, say PT, that creates an object of type PT from any 
parameter that implements this interface. This single function can 
then be used to convert any other "point" object to an object of 
type PT.

There may be a small decrease in floating-point numerical accuracy 
in this approach, because it is like first converting the source 
type to a "preferable" type (the interface), and then from there to 
the target type, with round-offs in both conversions.

 > I think run-time dispatching would be too much heavy.

Using tagged types or interfaces has two penalties: extra space in 
each object to store the tag, and extra execution time *if* 
run-time dispatching is used. But I sense that you want to have 
different "point" representations because you want to use the most 
suitable representation for each algorithm. If you write your 
algorithms as type-specific rather class-wide, the compiler should 
bind all operations statically, even if the types are tagged.

The need to assign and copy tags implies a (small) time penalty 
even if operations are statically bound. If this is important to 
you, or if the space overhead of tags is important, I would 
recommend a two-layer approach: define the various "Point" types as 
untagged, and then define a "wrapping" tagged type hierarchy (or 
interface-implementing types) that has some such untagged "Point" 
type as a component. You can then write your type-specific 
algorithms using the untagged types, and "wrap" them into objects 
of the tagged type(s) when you want to be class-wide. For example:

    -- The untagged representations:

    type Cartesian_Point is record .. end record;
    type Spherical_Point is record .. end record;

    -- The tagged representations:

    type Tagged_Point is abstract tagged null record;

    type Tagged_Cartesian_Point is new Tagged_Point
    with record Point : Cartesian_Point; end record;

    type Spherical_Cartesian_Point is new Tagged_Point
    with record Point : Spherical_Point; end record;

    -- The algorithms:

    procedure Cartesian_Algo (P : Cartesian_Point) is ...
    procedure Spherical_Algo (P : Spherical_Point) is ...

    procedure Class_Wide_Algo (P : Tagged_Point'Class) is ...

HTH,

-- 
Niklas Holsti
Tidorum Ltd
niklas holsti tidorum fi
       .      @       .