Hello,
Original goal i'd like to achieve is to completely abstract user of my
`library` from concrete code implementations which would be set in one
centralized place (such as config file). While it's common to do that in C++ at
runtime, i'd like to let compiler optimize more by making this in compile-
time.
So, the first solution i came up is to make a 'cfg' or 'impl' template, which
accepts interface as an argument. It would be something like configurable
std.container : make. Inside this template, i could read and parse simple
configuration file to find proper implementation, and return it for user.
Unfortunately, i'm not aware of any method of instantiating any class in
template like this.
That's why i started to implementing other, simpler method. I assumed, that
CMake which I'm using, will control include paths correctly and swap
implementations on every build. What a dissapointment i met, when after
refactoring my code in this way, user can see only interface because this is
the first matching file module. Well, it's quite my fault because i believe it
was mentioned in TDPL, that module = 1 file.

So, to sum up:
My project is divided in two sections:
* interfaces
* implementations
How to make in my D library possibility of configurable instantiation (lookup
at compile time) by having only interface at client side?

Thanks,
Mariusz Gliwiński

shd Wrote:

> Hello,
> Original goal i'd like to achieve is to completely abstract user of my
> `library` from concrete code implementations which would be set in one
> centralized place (such as config file). While it's common to do that in C++ at
> runtime, i'd like to let compiler optimize more by making this in compile-
> time.
> So, the first solution i came up is to make a 'cfg' or 'impl' template, which
> accepts interface as an argument. It would be something like configurable
> std.container : make. Inside this template, i could read and parse simple
> configuration file to find proper implementation, and return it for user.
> Unfortunately, i'm not aware of any method of instantiating any class in
> template like this.
> That's why i started to implementing other, simpler method. I assumed, that
> CMake which I'm using, will control include paths correctly and swap
> implementations on every build. What a dissapointment i met, when after
> refactoring my code in this way, user can see only interface because this is
> the first matching file module. Well, it's quite my fault because i believe it
> was mentioned in TDPL, that module = 1 file.
> 
> So, to sum up:
> My project is divided in two sections:
> * interfaces
> * implementations
> How to make in my D library possibility of configurable instantiation (lookup
> at compile time) by having only interface at client side?
> 
> Thanks,
> Mariusz Gliwiñski

I'm sorry to say that I don't exactly know what you are trying to achieve or problems you are having. I'm going to trying  and provide details on what I have gotten out of your message.

You can do compile-time configuration, this however is not a replacement to run-time configuration files.

What you are trying to do, you probably don't want to do. It sounds like you should be providing a library that can be expanded on rather than a configurable implementation. Maybe this is what you are trying to do, and it might help to think of it in this form instead.

D provides "Anti-Highjacking" measures to prevent a module from calling into code it did not know existed upon creation. This does prevent some interesting (and useful) patterns from being used, but generally it just results in unexpected, hard to track bugs.

Firstly, thanks for response.
2011/7/11 Jesse Phillips <jessekphillips+D@gmail.com>:
> I'm sorry to say that I don't exactly know what you are trying to achieve or problems you are having. I'm going to trying  and provide details on what I have gotten out of your message.

Hopefully i'll clarify it more.

> You can do compile-time configuration, this however is not a replacement to run-time configuration files.

I'm aware, and that's not my goal - let me clarify difference between
my run-time and compile-time configs.
Run-time configuration for my project might be delegated to optional
`util.cfg` package which is able to read let's say XML file that will
store things like resolution or something.

While i'm trying to make my potentially very big project possible to make in reasonable time, i'm pretty focused on code-reusing and making use of third-party libraries. Because external code quality might change, my preferences might change, and APIs are not what i'd love to work with i'd like to try pay price of composing D object interfaces i like, and implement them with wrappers that are connecting with external code.

So, back on to my previous example of runtime `util.cfg`, it might use filesystem/http reader and xml/cvs formatter from package `io` (this is gonna be run-time configuration) BUT besides of that i might like to change xml phobos wrapper (actually this example is not the best) with rapidxml or whatever and *this* is gonna be compile-time config. So, config has to be configured and it is going to be configured statically.

Ps. Don't get me wrong, i'm not in love with wrapping standard library, but there are other classes which i'm going to use which have no interface in phobos. This was just an example.

> What you are trying to do, you probably don't want to do. It sounds like you should be providing a library that can be expanded on rather than a configurable implementation. Maybe this is what you are trying to do, and it might help to think of it in this form instead.

Well, my end-goal is to provide complete application. Because i embrace code reusability, i love specialized libraries and when i'll need to implement something specialized, then share it - i'd love to do that. I'm afraid it's more like framework than library, but i don't like frameworks because they force you to use everything. What i'd like to achieve is to let someone use only parts of my code for plugging it into his application, lessen dependencies as much as i can, while remain type-safe.

Actually, project structure looks like that:
* project/
  * bin/
  * work/
  * docs/
  * iface/
    * ae
    * ge
    * pe
    (...)
  * impl/
    * glfw
      * ge.window
      * io.hid.keyboard
      * io.hid.mouse
      (you can add other window toolkits, or just separate OpenGL
context creation code and user-input handling from different
libraries)
      * h3d
        * ge.renderer
        (...)
      (you can add any other renderers or parts of game engines if
they let you to do that)
    * project_name
      * <package.interfaces implemented by me>
  * pe
    * bullet
    * ode
    * newton
(code is opensource and publicly accessible but excuse me i don't
quote anything because it's currently ugly and unprofessional so i'm
not proud of it yet, although i figured it out it's best way to do my
way, it just needs time and work to be proud)

So, i could easily change libraries, someone might use parts of my code by implementing bounding interfaces and as side effect i could easily benchmark how different implementations act in my project (like with C++ PAL).

So, if i'll handle task of abstracting all this out, and get reasonable performance that's the other topic, but i believe it's worth a try. All i need is to have a tool that lets me choose implementation at compile time without interface file modification. Any clues?

> D provides "Anti-Highjacking" measures to prevent a module from calling into code it did not know existed upon creation. This does prevent some interesting (and useful) patterns from being used, but generally it just results in unexpected, hard to track bugs.

Any anti-anti-hijacking ideas? :)

shd Wrote:

> Actually, project structure looks like that:
> * project/
>   * bin/
>   * work/
>   * docs/
>   * iface/
>     * ae
>     * ge
>     * pe
>     (...)
>   * impl/
>     * glfw
>       * ge.window
>       * io.hid.keyboard
>       * io.hid.mouse
>       (you can add other window toolkits, or just separate OpenGL
> context creation code and user-input handling from different
> libraries)
>       * h3d
>         * ge.renderer
>         (...)
>       (you can add any other renderers or parts of game engines if
> they let you to do that)
>     * project_name
>       * <package.interfaces implemented by me>
>   * pe
>     * bullet
>     * ode
>     * newton
> (code is opensource and publicly accessible but excuse me i don't
> quote anything because it's currently ugly and unprofessional so i'm
> not proud of it yet, although i figured it out it's best way to do my
> way, it just needs time and work to be proud)
> 
> So, i could easily change libraries, someone might use parts of my code by implementing bounding interfaces and as side effect i could easily benchmark how different implementations act in my project (like with C++ PAL).
> 
> So, if i'll handle task of abstracting all this out, and get reasonable performance that's the other topic, but i believe it's worth a try. All i need is to have a tool that lets me choose implementation at compile time without interface file modification. Any clues?

You want an IOC container like Unity?
And you want the interface-to-class mapping to be configurable externally rather than version out these classes directly in source?
like
version(UseGL)
{
  static import wrappers.gl;
  alias wrappers.gl.WindowImpl Window;
}
else version(UseWhat)
{
  static import wrappers.what;
  alias wrappers.what.WindowImpl Window;
}
container.register!(IWindow,Window)();

client:

IWindow myWindow=container.resolve!(IWindow)();

2011/7/12 Kagamin <spam@here.lot>:
> You want an IOC container like Unity?
exactly
> And you want the interface-to-class mapping to be configurable externally rather than version out these classes directly in source?
> like
> version(UseGL)
> {
>  static import wrappers.gl;
>  alias wrappers.gl.WindowImpl Window;
> }
> else version(UseWhat)
> {
>  static import wrappers.what;
>  alias wrappers.what.WindowImpl Window;
> }
> container.register!(IWindow,Window)();
>
> client:
>
> IWindow myWindow=container.resolve!(IWindow)();
>
that's right :)

> Any anti-anti-hijacking ideas? :)

Hopefully I remember to reply more in-depth when I get home.

But this was a question on SO that sounds kind of like your request. I tried explaining why anti-hijacking was created.

http://stackoverflow.com/questions/6362280/d-template-specialization-in-different-source-file/6366177#6366177

shd Wrote:

> Firstly, thanks for response.

After reading Kagamin's response, I'm not sure I understand you still. You want to have code reuse, you want to have run-time and compile-time configuration, but you don't want to make use of the technologies/techniques/patterns which facilitate these.

You say no to libraries, you say no to frameworks, you say no to versioning and desire the benefits of these without what they entail. If you come up with an alternative, then please do bring it forward. I don't see how anyone will be able to get you what you want.

shd Wrote:

> 2011/7/12 Kagamin <spam@here.lot>:
> > You want an IOC container like Unity?
> exactly
> > And you want the interface-to-class mapping to be configurable externally rather than version out these classes directly in source?
> > like
> > version(UseGL)
> > {
> > �static import wrappers.gl;
> > �alias wrappers.gl.WindowImpl Window;
> > }
> > else version(UseWhat)
> > {
> > �static import wrappers.what;
> > �alias wrappers.what.WindowImpl Window;
> > }
> > container.register!(IWindow,Window)();
> >
> > client:
> >
> > IWindow myWindow=container.resolve!(IWindow)();
> >
> that's right :)

Hmm... If you have
wrappers/gl/guiImpl/...
and
wrappers/what/guiImpl/...
and conditionally add them to includes
for example
with -I wrappers/gl you should be able to import guiImpl; and use guiImpl.Window. If it's -I wrappers/what, guiImpl.Window is still guiImpl.Window.

2011/7/13 Kagamin <spam@here.lot>:
> Hmm... If you have
> wrappers/gl/guiImpl/...
> and
> wrappers/what/guiImpl/...
> and conditionally add them to includes
> for example
> with -I wrappers/gl you should be able to import guiImpl; and use guiImpl.Window. If it's -I wrappers/what, guiImpl.Window is still guiImpl.Window.

Yeah, that might work.

I tend to see everywhere very strange problems (for other people). Like I think of including interface file from user-app, and then give him only this interface compliant instance. And if he wants to use some implementation-depended functionalities he should clearly instantiate IObjectSpecialized or import specialized.path.to.package;

By doing it in way you are talking about (which is kind of less perfectionist version of method mentioned by me in first post, although *it works*), user is able to import path.to.package; then instantiate object by standard instantiation, and accidentally use implementation-dependant method. Then, he commits and someone else wants to compile end-app with different set of implementation libraries. Program won't compile because someone used method which isn't specified in portable interface. In the same time, program is assumed to work so, because it's not clear that person who wrote this call is acting unportable (so we're relying on documentation rather than no-can-do method). So it's a fail of my 'framework', which claims to let use x, y and z libraries interchangeably by only changing cmake definition, while the build breaks. Author of framework didn't gave programmer, clear feedback that he is acting unportable.

I hope i didn't bored you too much with my imagination because there is something more i don't like. Directory hierarchy :) It's not D-related though.

Soo, the only solutions we came up is doing it out of language (by -I impl/library/path), or messing with import/aliases in interfaces which isn't nice. Although first solution isn't as good as i hoped for, it works... which is a good start, so i'll probably do it this way :)

2011/7/13 Jesse Phillips <jessekphillips+D@gmail.com>:
> After reading Kagamin's response, I'm not sure I understand you still. You want to have code reuse, you want to have run-time and compile-time configuration, but you don't want to make use of the technologies/techniques/patterns which facilitate these.

I'm not sure are you referring 'technologies/techniques/patterns' to code reusability or configuration. If the first one, read one. If second, i don't know which techniques/patterns you are talking about. version() doesn't match for me here i think.

> You say no to libraries

That's not true! I'm not sure how did you came up to this conclusions, but i love libraries for applications same as command-line programs (like ffmpeg) for GUI apps.


>, you say no to frameworks,

That's true i don't really like frameworks, but only because i felt like they're trying to be irreplaceable. If you're using one part, it's hard to make other parts they implemented in your own way (so they're too tightly integrated). Maybe i just tried wrong ones, i don't know. Libraries on the other hand, let you compose your program from solutions i like most, and that's why i'm trying to do this clear division between interface and implementation. So i can use different specialized libraries, and later make my own, so if someone likes my entity system he doesn't have to use my poor renderer. So hopefully i won't end-up with some big chunk of code as ogre3d.

> you say no to versioning and desire the benefits of these without what they entail. If you come up with an alternative, then please do bring it forward. I don't see how anyone will be able to get you what you want.

I like version control, and that's why i'm using one (git), even if i don't really need to operate with someone because of current project phase. Versions we're talking about are rather like compilation profiles, so you can use different sets of libraries to try different implementations without code modification. So, in one moment there would be code for many alternative implementations, and versions would affect only resulting binary which isn't commited at all.