[Piglit] [PATCH v2 10/13] piglit-dispatch: Switch to using piglit-dispatch instead of GLEW.

[Jose Fonseca]

----- Original Message ----- 

> On 16 March 2012 12:04, Jose Fonseca < jfonseca at vmware.com > wrote:

> > ----- Original Message -----
> 
> > > On 03/12/2012 02:41 PM, Paul Berry wrote:
> 
> > > > Previous patches set up the piglit-dispatch infrastructure but
> > > > did
> 
> > > > not
> 
> > > > use it. This patch enables piglit-dispatch and uses it instead
> > > > of
> 
> > > > GLEW.
> 
> > > >
> 
> > > > No piglit regressions on Intel SandyBridge (Linux) or Mac OSX.
> 
> > >
> 
> > > Nice. Only one more platform to conquer.
> 
> > >
> 
> > > > #if defined(USE_OPENGL)
> 
> > > > -# include "glew.h"
> 
> > > > - /* Include the real headers too, in case GLEW misses
> > > > something.
> 
> > > > */
> 
> > > > +# include "piglit-dispatch.h"
> 
> > > > + /* Include the real headers too, in case piglit-dispatch
> > > > misses
> 
> > > > something. */
> 
> > > > # ifdef __APPLE__
> 
> > > > # include <OpenGL/gl.h>
> 
> > > > # include <OpenGL/glu.h>
> 
> > >
> 
> > > Shouldn't Apple's <OpenGL/gl.h> be removed too?
> 
> > > I think we discussed this before, but I don't remember the
> 
> > > conclusion.
> 

> > It's probably pointless if all GL defitions are done before.
> 

> > But note that Apple's glext.h is non standard:
> 

> > https://github.com/apitrace/apitrace/blob/master/thirdparty/khronos/GL/glext.patch
> 

> > It's probably better to replicate this on piglit-dispatch too.
> 

> > Jose
> 

> Interesting: on my mac, GLhandleARB is defined as a void *. Your
> patch defines it as an unsigned long (which is equivalent to void *
> from an ABI perspective).

This was deliberate. As handling both "void *" would make the apitrace code generation unnecessarily complex.

> GL defines it as an unsigned int, which is certainly *not* equivalent
> on 64-bit systems.

> This throws a bit of a monkey wrench into things, since it breaks
> some of the aliases defined in the official gl.spec file. for
> example GetAttachedObjectsARB is marked as an alias for
> GetAttachedShaders, but this can't possibly work on mac if one of
> them takes a GLhandleARB * as its 4th argument, and the other takes
> a GLuint * as its 4th argument (since the two types have different
> sizes).

Never noticed that gl.spec had aliases.  I don't think that information is very reliable.

For example, gl.spec also marks ARB_framebuffer_object as aliases of EXT_framebuffer_object functions, but semantics are different, and I heard reports of segfaults on NVIDIA drivers due to inadvertedly mixing up ARB_framebuffer_object and EXT_framebuffer_object functions.

I think handling aliases under the hood causes more harm than good. It's better that apps are well aware of what extensions they are using.

Jose