Hi,

I have a wrapper for a "object aware" c library (cairo). Take for example two classes, Surface and a subclass, ImageSurface. Now this code has to be valid:
-----------------------
auto ptr = cairo_image_surface_create(CAIRO_FORMAT_ARGB32, 512, 512);
Surface s = new Surface(ptr);
ImageSurface imgs = cast(ImageSurface)s;
-----------------------

As D cannot know that 's' really should be an ImageSurface, I have implemented opCast to get this example working:
-----------------------
class Surface
{
    static Surface castFrom(Surface other)
    {
        return other;
    }

    T opCast(T)() if(isImplicitlyConvertible!(T, Surface))
    {
        return T.castFrom(this);
    }
}
class ImageSurface : Surface
{
    static ImageSurface castFrom(Surface other)
    {
        auto type = cairo_surface_get_type(other.nativePointer);
        if(type == cairo_surface_type_t.CAIRO_SURFACE_TYPE_IMAGE)
        {
            return new ImageSurface(other.nativePointer);
        }
        else
            return null;
    }
}
-----------------------

This code works quite well. But it performs unnecessary calls to cairo_surface_get_type (and allocates unnecessary objects) for simple cases:
-----------------------
auto surface = new ImageSurface(Format.CAIRO_FORMAT_ARGB32, 400, 400);
Surface tmp = cast(Surface)surface;
ImageSurface test = cast(ImageSurface)as;
-----------------------

In this case, the first D object already is an ImageSurface so the custom opCast code isn't needed for the last line.

So the question is: Is there some way to check in the opCast function if a normal D object cast would succeed and then just return it's result?

-- 
Johannes Pfau

Johannes Pfau wrote:
>Hi,
>
>I have a wrapper for a "object aware" c library (cairo). Take for example two classes, Surface and a subclass, ImageSurface. Now this code has to be valid:
>-----------------------
>auto ptr = cairo_image_surface_create(CAIRO_FORMAT_ARGB32, 512, 512);
>Surface s = new Surface(ptr);
>ImageSurface imgs = cast(ImageSurface)s;
>-----------------------
>
>As D cannot know that 's' really should be an ImageSurface, I have implemented opCast to get this example working:
>-----------------------
>class Surface
>{
>    static Surface castFrom(Surface other)
>    {
>        return other;
>    }
> 
>    T opCast(T)() if(isImplicitlyConvertible!(T, Surface))
>    {
>        return T.castFrom(this);
>    }
>}
>class ImageSurface : Surface
>{
>    static ImageSurface castFrom(Surface other)
>    {
>        auto type = cairo_surface_get_type(other.nativePointer);
>        if(type == cairo_surface_type_t.CAIRO_SURFACE_TYPE_IMAGE)
>        {
>            return new ImageSurface(other.nativePointer);
>        }
>        else
>            return null;
>    }
>}
>-----------------------
>
>This code works quite well. But it performs unnecessary calls to cairo_surface_get_type (and allocates unnecessary objects) for simple cases:
>-----------------------
>auto surface = new ImageSurface(Format.CAIRO_FORMAT_ARGB32, 400, 400);
>Surface tmp = cast(Surface)surface;
>ImageSurface test = cast(ImageSurface)as;
>-----------------------
>
>In this case, the first D object already is an ImageSurface so the custom opCast code isn't needed for the last line.
>
>So the question is: Is there some way to check in the opCast function if a normal D object cast would succeed and then just return it's result?
>

In case anyone's interested:
I think this could be done with _d_dynamic_cast from rt.cast_
(druntime).
I've dropped the opCast/castFrom approach though (not working in some
cases), so I haven't tested this.

-- 
Johannes Pfau

On Sat, 09 Jul 2011 05:47:51 -0400, Johannes Pfau <spam@example.com> wrote:

> Hi,
>
> I have a wrapper for a "object aware" c library (cairo). Take for
> example two classes, Surface and a subclass, ImageSurface. Now this
> code has to be valid:
> -----------------------
> auto ptr = cairo_image_surface_create(CAIRO_FORMAT_ARGB32, 512, 512);
> Surface s = new Surface(ptr);
> ImageSurface imgs = cast(ImageSurface)s;
> -----------------------
>
> As D cannot know that 's' really should be an ImageSurface, I have
> implemented opCast to get this example working:
> -----------------------
> class Surface
> {
>     static Surface castFrom(Surface other)
>     {
>         return other;
>     }
>    T opCast(T)() if(isImplicitlyConvertible!(T, Surface))
>     {
>         return T.castFrom(this);
>     }
> }
> class ImageSurface : Surface
> {
>     static ImageSurface castFrom(Surface other)
>     {
>         auto type = cairo_surface_get_type(other.nativePointer);
>         if(type == cairo_surface_type_t.CAIRO_SURFACE_TYPE_IMAGE)
>         {
>             return new ImageSurface(other.nativePointer);
>         }
>         else
>             return null;
>     }
> }
> -----------------------
>
> This code works quite well. But it performs unnecessary calls to
> cairo_surface_get_type (and allocates unnecessary objects) for simple
> cases:
> -----------------------
> auto surface = new ImageSurface(Format.CAIRO_FORMAT_ARGB32, 400, 400);
> Surface tmp = cast(Surface)surface;
> ImageSurface test = cast(ImageSurface)as;
> -----------------------
>
> In this case, the first D object already is an ImageSurface so the
> custom opCast code isn't needed for the last line.
>
> So the question is: Is there some way to check in the opCast function
> if a normal D object cast would succeed and then just return it's
> result?

I think a factory method would work well here.

If you have a finite set of classes you are creating, a factory method can simply use a switch on the cairo_surfase_type, and you could even put it in Surface:

auto s = Surface.create(ptr); // automatically creates the correct derived class.

Then use dynamic cast to get to the expected derived class.

If you do not have a finite set of classes, you may be able to use runtime type info (a la object.factory).  But from your example, it seems like cairo defines an enum which encapsulates all classes.

Another option, judging from your code, if cairo's functions to create surface objects are specific to the derived type (i.e. cairo_image_surface_create => ImageSurface), then you could simply wrap the cairo functions.  Basically avoid calling the C creation routines outside the D class constructors.

-Steve

Steven Schveighoffer wrote:
>On Sat, 09 Jul 2011 05:47:51 -0400, Johannes Pfau <spam@example.com> wrote:
>
>> Hi,
>>
>> I have a wrapper for a "object aware" c library (cairo). Take for example two classes, Surface and a subclass, ImageSurface. Now this code has to be valid:
>> -----------------------
>> auto ptr = cairo_image_surface_create(CAIRO_FORMAT_ARGB32, 512, 512);
>> Surface s = new Surface(ptr);
>> ImageSurface imgs = cast(ImageSurface)s;
>> -----------------------
>>
>> As D cannot know that 's' really should be an ImageSurface, I have implemented opCast to get this example working:
>> -----------------------
>> class Surface
>> {
>>     static Surface castFrom(Surface other)
>>     {
>>         return other;
>>     }
>>    T opCast(T)() if(isImplicitlyConvertible!(T, Surface))
>>     {
>>         return T.castFrom(this);
>>     }
>> }
>> class ImageSurface : Surface
>> {
>>     static ImageSurface castFrom(Surface other)
>>     {
>>         auto type = cairo_surface_get_type(other.nativePointer);
>>         if(type == cairo_surface_type_t.CAIRO_SURFACE_TYPE_IMAGE)
>>         {
>>             return new ImageSurface(other.nativePointer);
>>         }
>>         else
>>             return null;
>>     }
>> }
>> -----------------------
>>
>> This code works quite well. But it performs unnecessary calls to cairo_surface_get_type (and allocates unnecessary objects) for simple cases:
>> -----------------------
>> auto surface = new ImageSurface(Format.CAIRO_FORMAT_ARGB32, 400,
>> 400); Surface tmp = cast(Surface)surface;
>> ImageSurface test = cast(ImageSurface)as;
>> -----------------------
>>
>> In this case, the first D object already is an ImageSurface so the custom opCast code isn't needed for the last line.
>>
>> So the question is: Is there some way to check in the opCast function if a normal D object cast would succeed and then just return it's result?
>
>I think a factory method would work well here.
>
>If you have a finite set of classes you are creating, a factory method can simply use a switch on the cairo_surfase_type, and you could even put it in Surface:
>
>auto s = Surface.create(ptr); // automatically creates the correct
>derived class.
>
>Then use dynamic cast to get to the expected derived class.
>
>If you do not have a finite set of classes, you may be able to use runtime type info (a la object.factory).  But from your example, it seems like cairo defines an enum which encapsulates all classes.
>
>Another option, judging from your code, if cairo's functions to create surface objects are specific to the derived type (i.e. cairo_image_surface_create => ImageSurface), then you could simply wrap the cairo functions.  Basically avoid calling the C creation routines outside the D class constructors.
>
>-Steve

Thanks for answering.
I already use such a factory method (called createFromNative). I also
create the correct D objects if the Surfaces are created from D
as in your second suggestion (I still need createFromNative, as there
are methods like cairo_pattern_get_surface which receive Surfaces from
cairo).
Thinking about it the whole opCast stuff is indeed obsolete as long as
all code correctly calls createFromNative. I first wanted to make it
possible to add Surface wrappers outside of cairoD, but I dropped that
idea. Without that feature a factory method should be enough.

-- 
Johannes Pfau