bug#15876: 24.3.50; Highly degraded performance between rev 114715 and 115006

[Eli Zaretskii <eliz@gnu.org>]

> Date: Mon, 02 Dec 2013 14:45:25 +0400
> From: Dmitry Antipov <dmantipov@yandex.ru>
> CC: 15876@debbugs.gnu.org
> 
> On 12/01/2013 08:31 PM, Eli Zaretskii wrote:
> 
> > Dmitry, can you please look into this?  I'm not familiar enough with
> > the font stuff, so I don't see how was the font-spec and its
> > font-entities stored in the font cache supposed to be referenced from
> > some other Lisp object, to make sure they are marked and not GC'ed.
> 
> It should be easy - just use this:
> 
> === modified file 'src/alloc.c'
> --- src/alloc.c	2013-12-01 22:33:13 +0000
> +++ src/alloc.c	2013-12-02 09:32:55 +0000
> @@ -5731,7 +5731,18 @@
>     eassert (!VECTOR_MARKED_P (ptr));
>     VECTOR_MARK (ptr);		/* Else mark it.  */
>     if (size & PSEUDOVECTOR_FLAG)
> -    size &= PSEUDOVECTOR_SIZE_MASK;
> +    {
> +      size &= PSEUDOVECTOR_SIZE_MASK;
> +      if (PSEUDOVECTOR_TYPEP (&ptr->header, PVEC_FONT))
> +	{
> +	  Lisp_Object tmp;
> +	  XSETFONT (tmp, ptr);
> +	  if (FONT_SPEC_P (tmp))
> +	    fprintf (stderr, "mark font-spec\n");
> +	  else if (FONT_ENTITY_P (tmp))
> +	    fprintf (stderr, "mark font-entity\n");
> +	}
> +    }
> 
> and set breakpoints to fprintf(). I tried this and see that font-spec
> objects can be referenced from face caches:

Thanks.  But I think I didn't make myself clear: the issue is not just
to see ANY font-spec objects being marked.  The issue is with those
font-spec objects that are recorded in the font caches that are
compacted by compact_font_caches.  I don't see any code that makes
sure some Lisp object references those caches.  Without that, they
cannot be possibly marked, and will be GC'ed, right?

> But the most of font-spec and font-entity objects are referenced via
> staticpro'ed globals Vfontset_table and ft_face_cache.

Those staticpro'ed objects might just be the reason why you don't see
the problem.  Your build uses the ftfont driver, doesn't it?  Because
ftfont.c has this implementation of the get_cache method:

  static Lisp_Object
  ftfont_get_cache (struct frame *f)
  {
    return freetype_font_cache;
  }

and freetype_font_cache is a staticpro'ed variable.  By contrast,
w32font.c does this:

  Lisp_Object
  w32font_get_cache (struct frame *f)
  {
    struct w32_display_info *dpyinfo = FRAME_DISPLAY_INFO (f);

    return (dpyinfo->name_list_element);
  }

and dpyinfo->name_list_element is what compact_font_caches examines.

(Note that xfont.c and xftfont.c also use a cache stored in
dpyinfo->name_list_element, so this issue is not entirely Windows
specific.)

> > In any case, it seems like they are never marked on Windows, because
> > if I set a breakpoint in the line marked below:
> >
> > 	  /* If font-spec is not marked, most likely all font-entities
> > 	     are not marked too.  But we must be sure that nothing is
> > 	     marked within OBJ before we really drop it.  */
> > 	  for (i = 0; i < size; i++)
> > 	    if (VECTOR_MARKED_P (XFONT_ENTITY (AREF (XCDR (obj), i))))
> > 	      break;
> >
> > 	  if (i == size)  <<<<<<<<<<<<<<<<<<<<<
> > 	    drop = 1;
> > 	}
> >
> > with a condition i != size, that breakpoint never breaks.
> 
> The best way to hit this breakpoint is to run something like:
> 
> (defun bloat-font ()
>    (interactive)
>    (let ((fonts (x-list-fonts "*")))
>      (while fonts
>        (condition-case nil (set-frame-font (car fonts)) (error nil))
>        (setq fonts (cdr fonts))
>        (redisplay))))
> 
> (This is X-specific; I believe there should be a similar MS-Windows method).

What is X-specific here?  If you mean x-list-fonts, then it exists on
Windows and does the same.

Anyway, I think we are mis-communicating again: the above Lisp code
will cause the breakpoint to fire with i == size, i.e. we will remove
the font-entities from the cache.  I wanted the opposite: to see at
least 1 font-entity that is NOT removed.  So my breakpoint condition
was "i != size", which means we will NOT drop the font-entity.  What I
saw is that this condition is never true, which means we remove ALL of
the font-entities from the cache.

> Subtle "triangle example" works just fine for me (with default font used by
> GTK3 build and 'emacs -Q' at least).

If the font cache is not cleared, you will not see any problem at all.

> Also I'm curious how to reproduce this issue with X. In particular, how to
> find a font so "heavy" that an attempt to load it creates a lot of Lisp data
> (hundreds Kbytes, enough to reach gc_cons_threshold each time)?

Well, you could lower gc_cons_threshold instead, couldn't you?