Hi all,

I've spent a bit more time looking into this. The problem is due to the following three lines being conditioned on `!NILP(val)`: (that's around line 2780 of src/font.c)

    Lisp_Object copy = copy_font_spec (scratch_font_spec);
    ASET (copy, FONT_TYPE_INDEX, driver_list->driver->type);
    XSETCDR (cache, Fcons (Fcons (copy, val), XCDR (cache)));

Before af1a69f, these lines were always run. In af1a69f, they only run if `val` is non nil, causing the regression: on master, moving them back out of the if statement fixes the problem. For clarity I have attached the corresponding patch.

Clément.

On 07/18/2015 03:24 AM, Clément Pit--Claudel wrote:
> On 07/15/2015 05:32 AM, Dmitry Antipov wrote:
>> On 07/10/2015 07:55 PM, Clément Pit--Claudel wrote:
>>
>>> The figures are very similar to the tests above: with that patch inserting 50 lines takes 3 seconds,
>>> and without it it's instantaneous. Thus I think it's safe to say that this particular patch is indeed
>>> responsible for the performance regression. But maybe I'm missing something?
>>
>> As of c40ea1328bb33abaec14f1fc92ac2349b5ee2715, I can't reproduce this issue, with your fontset setup
>> from http://debbugs.gnu.org/cgi/bugreport.cgi?bug=21028#17. Cursor motion and keyboard/mouse selection
>> are smooth, and CPU/memory usage looks normal.
> 
> Thanks for trying this out. Thanks for your suggestions, too.
> 
>> My suggestions are:
>>
>> 1) Re-run your timed tests from http://debbugs.gnu.org/cgi/bugreport.cgi?bug=21028#20 under /usr/bin/time,
>> not time (the latter is a simple shell builtin). /usr/bin/time also shows memory usage; if "bad" (current)
>> instance consumes more memory than "good" (with reverted change) one, there may be nasty GC issue.
> 
> The timings are similar:
> 
> $ /usr/bin/time emacs-bad/src/emacs -Q --eval "(progn (dotimes (_ 3) (set-fontset-font t 'unicode (font-spec :name \"Arial\") nil 'append)) (dotimes (_ 500) (insert \"日本国\n\")) (dotimes (_ 500) (previous-line) (redisplay t)) (kill-emacs))"
> 13.31user 8.57system 0:34.54elapsed 63%CPU (0avgtext+0avgdata 50592maxresident)k
> 0inputs+0outputs (0major+4958minor)pagefaults 0swaps
> 
> $ /usr/bin/time emacs-good/src/emacs -Q --eval "(progn (dotimes (_ 3) (set-fontset-font t 'unicode (font-spec :name \"Arial\") nil 'append)) (dotimes (_ 500) (insert \"日本国\n\")) (dotimes (_ 500) (previous-line) (redisplay t)) (kill-emacs))"
> 0.58user 0.03system 0:01.05elapsed 58%CPU (0avgtext+0avgdata 50392maxresident)k
> 29840inputs+0outputs (105major+4911minor)pagefaults 0swaps
> 
>> 2) 3 seconds is large enough to leave the traces in profiled runs. On GNU/Linux, it may be worth trying
>> to run under perf, both "good" and "bad" cases, and comparing reports.
> 
> I have attached the reports in the good and bad cases (200c532 and af1a69f respectively). I'm not sure what to conclude from them; I can provide the full trace data if needed.
> 
> Clément.
>