[cairo] Cairo 1.14.2 problems with Solaris 10, libXrender

[suzuki toshiya]

> No. Symbol checking is done at compile time (by configure). The runtime
> check is performed using the function
> 
>    Status XRenderQueryVersion (Display *dpy,
> 			       int     *major_versionp,
> 			       int     *minor_versionp);
> 
> which is declared in Xrender.h and which queries the version of the X
> server (i.e. display) we have connected to (represented by the Display
> pointer).

Oh, I see. The function availability could be checked by
the XRender version...

Regards,
mpsuzuki

Andreas F. Borchert wrote:
> On Wed, Aug 05, 2015 at 11:07:20PM +0900, suzuki toshiya wrote:
>> The issue was found by the inconsistent header/library in
>> Solaris 10 (so, it would be possible for Solaris 10 users
>> to bypass this problem by tweaking the configure script),
>> but your comment is that the hardwired configuration about
>> the availability of XrenderCreateSolidFill etc is not good
>> idea, they should be checked in runtime.
>>
>> Is this correct understanding?
> 
> We have checks by configure (compile time) and during runtime (using
> XRenderQueryVersion, see below). Currently both is checked and we need
> also both. The point is that we should not invoke functions that were
> no-op'd as they are not included in the libXrender client library when
> they appear to be supported by an actual X server (i.e. display) we
> connect to.
> 
>> If this is correct understanding, the runtime checking of
>> the availability would be a symbol checking in libXrender?
> 
> No. Symbol checking is done at compile time (by configure). The runtime
> check is performed using the function
> 
>    Status XRenderQueryVersion (Display *dpy,
> 			       int     *major_versionp,
> 			       int     *minor_versionp);
> 
> which is declared in Xrender.h and which queries the version of the X
> server (i.e. display) we have connected to (represented by the Display
> pointer).
> 
> The actual code checks the runtime version only. Example:
> 
>    if (CAIRO_RENDER_HAS_GRADIENTS(dst->display)) {
>       picture = XRenderCreateSolidFill (dpy, &xcolor);
>    } else {
>       // ...
>    }
> 
> The macro CAIRO_RENDER_HAS_GRADIENTS has been defined in
> cairo-xlib-private.h as
> 
>    #define CAIRO_RENDER_HAS_GRADIENTS(surface) \
>        CAIRO_RENDER_AT_LEAST((surface), 0, 10)
> 
> where CAIRO_RENDER_AT_LEAST accesses render_major and
> render_minor which are initialized by using XRenderQueryVersion
> (see cairo-xlib-display.c):
> 
>    #define CAIRO_RENDER_AT_LEAST(surface, major, minor) \
>         (((surface)->render_major > major) || \
> 	 (((surface)->render_major == major) && \
> 	  ((surface)->render_minor >= minor)))
> 
> The Cairo library should be prepared for the case that
> the XRender version of the actual X Server is more recent
> than that of the XRender client library Cairo was linked
> against. Hence I would suggest to use #ifdefs as follows:
> 
> #if HAVE_XRENDERCREATELINEARGRADIENT
>    if (CAIRO_RENDER_HAS_GRADIENTS(dst->display)) {
>       picture = XRenderCreateSolidFill (dpy, &xcolor);
>    } else {
> #endif
>       // ...
> #if HAVE_XRENDERCREATELINEARGRADIENT
>    }
> #endif
> 
> And to defend against accidental invocations of a no-op
> I would suggest to replace in cairo-xlib-xrender-private.h
> 
>    __attribute__((__unused__)) static int
>    _int_consume (void *p, ...)   { return 0; }
> 
>    // ...
> 
>    #if !HAVE_XRENDERCREATESOLIDFILL
>    #define XRenderCreateSolidFill _int_consume
>    #endif
> 
> by
> 
>    __attribute__((__unused__)) static int
>    _int_consume_and_abort (void *p, ...)   { assert(0); return 0; }
> 
>    // ...
> 
>    #if !HAVE_XRENDERCREATESOLIDFILL
>    #define XRenderCreateSolidFill _int_consume_and_abort
>    #endif
> 
> or something similar, i.e. if against all odds a no-op is called,
> a crash, error or warning message would be most helpful to track
> this down. Rendering should not be silently skipped.
> 
> Regards, Andreas.
>