bug#17124: 24.3.50; Occasional Extremely Slow Redraws in OSX Emacs

[Jan Djärv]

27 jun 2014 kl. 21:57 skrev Eli Zaretskii <eliz@gnu.org>:

>> From: Jan Djärv <jan.h.d@swipnet.se>
>> Date: Fri, 27 Jun 2014 19:13:00 +0200
>> Cc: ericfroemling@gmail.com,
>> 17124@debbugs.gnu.org
>> 
>>>> With the "shake the divider" recepie (see below), redisplay_internal is 
>>>> called more than 30 times per second.  On an old computer (end of 2008) I 
>>>> get about 37 times per second.
>>>> But each redisplay results in multiple draw_begin/end, so for drawing, it 
>>>> is more than 37 times per second.
>>> 
>>> Does it help to set redisplay-dont-pause to nil?
>> 
>> The "shake the divider" case becomes much worse, lots of flickering and 
>> incomplete text.
> 
> But do you see less drawing requests sent to the backend?

No.

> 
>>> Incidentally, I don't think your suggestion will help in the "shake
>>> the divider" scenario: when window dimensions are changed, we toss the
>>> glyph matrices of the affected windows, and then allocate them anew
>>> (and perform a thorough redisplay of those windows).  So this will
>>> anyhow require to redisplay the entire window completely, and the
>>> backend will not be able to save us any redraws.
>> 
>> Not by itself, but if the backend is responsible for when actual drawing 
>> happens we can make sure we don't draw faster than the screen can update.
> 
> I actually find it hard to believe that we overwhelm the backend,
> except, maybe, when the X client is on a remote machine.  E.g., on
> Windows the "shake the divider" recipe doesn't show any signs of a
> problem, and the CPU load is never more than a single execution unit
> being busy, which means not much is at work except Emacs itself.  With
> today's multi-core fast machines, how come it's impossible to redraw a
> region 37 times a second?

Well, the actual number is more than 37.  37 is the number of times redisplay 
is called.
But within one redisplay, there is a couple of separate update_begin/end blocks.

There was a couple of extra redraws happening in the NS port, but after 
removing them, I can still see problems with shake the divider.

It all goes away if drawing is done the normal way, i.e. from the event handler.
Maybe OSX is optimized for that case, as it covers about every application 
except Emacs.
The Emacs way is really only ment for small updates outside the event loop.

> 
>>>> I think there is room for optimizations in the generic display also, for 
>>>> example moving the mouse redraws the entire mode line, even if the mouse 
>>>> pointer is outside the frame.
>>> 
>>> Please show the recipe to reproduce this.  I just tried a naive way of
>>> doing that, and didn't see any mode-line updates even when the mouse
>>> is inside the frame.  So I must be missing something.
>>> 
>> 
>> I had problems seeing what was drawn and when so I added debug code to see 
>> what the font driver actually draws.  But I see now that it is different for 
>> X11.  So it might be NS specific.  What can make the modeline redraw in one 
>> version of Emacs and not in another?  I thought all that was generic code.
> 
> It _is_ generic code.  Perhaps we are not talking about the same
> things: when you say that the mode line is redrawn, what exactly do
> you see?

I see the mode line redrawn by inspecting what strings are sent to the font 
backend.  Actually the whole buffer is redrawn, I was just looking at an empty 
buffer.  But these are the extra redrawn I mentioned above, they are gone in 
trunk now.

> 
>> Well, shake the divider is not really something normal a user does.  It is 
>> just a way to force the issue.  But slow redraws happens in normal usage 
>> also, i.e. switching buffers and editing.  It solves the second case, but 
>> makes shake the divider worse in terms of smooth redraws.
> 
> We need to compare the performance with this proposed feature with the
> current implementation.  I think it's hard to talk about this without
> some measurements, and probably also some reasonably important use
> cases (which "shake the divider" isn't, IMO).

With the extra redrawns gone, shake the divider performes rather well now if I 
force draws to the event loop.  There is the occational pause in redraw, but I 
guess that is expected.
But I can't do this all in the backend, the downside is that there are double 
redraws.  One for the redisplay call and one from the event loop when the 
redisplay is done.  Also, the event loop redraw redraws the whole frame.

If I get some time I'll try to make a test.

        Jan D.