Re: [AD] END_OF_MAIN removal patch for msvc

[Peter Wang <novalazy@xxxxxxxxxx>]

On 2008-11-17, Evert Glebbeek <eglebbk@xxxxxxxxxx> wrote:
> On 17 Nov 2008, at 20:36, Peter Wang wrote:
> > It might be best just to expose the boilerplate to the user.
> > For any program which wants to be portable to Mac OS X, you'd write:
> >
> >     int real_main(int argc, const char *argv)
> >     {
> > 	al_init();
> > 	...
> >     }
> >
> >     int main(int argc, const char *argv)
> >     {
> > 	return al_run_in_main_thread(real_main, argc, argv);
> >     }
> >
> > Nothing up the sleeves.
> >
> > We might need three different versions of al_run_in_main_thread,
> > for zero arguments, two arguments, and three arguments.  Of course,  
> > we'd
> > probably get sick of writing that and end up providing a macro:
> >
> >     ALLEGRO_MAIN_WRAPPER(real_main, argc, argv);
> >
> > which looks suspiciously like END_OF_MAIN().  *But* it doesn't require
> > the symbol `_mangled_main_address' to exist, which causes problems  
> > when
> > people try to load the Allegro shared object at run-time, and there is
> > no name mangling.
> 
> My main concern would be the number of people who forget to do that,  
> inadvertently creating platform-dependent code.

I've done that with END_OF_MAIN(), too.  Granted, END_OF_MAIN() is
shorter so easier to remember.

> My secondary concern  
> would be how that would interact with other libraries (probably no  
> worse than END_OF_MAIN(), but the thought occured).

No worse.

> My third concern  
> would be that I actually think that looks uglier than END_OF_MAIN() ;)

We can tweak it until you're happy ;-)

Actually, the argc, argv arguments weren't needed.

    int real_main(int argc, const char *argv)
    {
	...
    }
    ALLEGRO_END_OF_MAIN(real_main);

> I think it makes sense to split the problem in two: requiring a copy  
> of argv[] and having a non-standard main/main requirement.
> The former can be solved by providing a function to pass it (we need  
> some sensible default if a user forgets to do this), assuming we  
> really want it still.
> The latter was never needed on *nix and now isn't needed on Windows  
> anymore either. That leaves OS X.
> 
> If OS X is the only platform that needs a magic main at the moment,  
> we should again look for a way to get rid of it there too. The  
> ucontext stuff Peter [Hull] mentioned would be what I'd look into,  
> assuming there's a way to make it work in Tiger (Leopard and up  
> shouldn't be a problem if I understand correctly). Maybe (optionally)  
> not in debug mode though, but then we'd still need END_OF_MAIN() just  
> for that. Some sort of progress...

The ucontext idea is fun, but I have a feeling it's too clever.

Peter