bug#18923: Alternative scrolling model

[Eli Zaretskii <eliz@gnu.org>]

> From: E Sabof <esabof@gmail.com>
> Date: Sun, 02 Nov 2014 02:31:28 +0000
> 
> > Sounds nice.  Do you imagine it as a replacement for the existing
> > scroll-up/down functions?  Or rather (at least at first) as a separate
> > package?
> > Also, if something's missing for st-height to get more accurate
> > measurements, I suggest you make it a bug report asking for that missing
> > info/feature.
> 
> I was mostly thinking the first.

If this is intended as a replacement for the existing functionality,
then it needs to support all the features that the current code
supports.  The list of those features should include at least the
following:

 . the argument to the commands can be nil, which means "almost the
   full window", where "almost full" depends on the value of
   next-screen-context-lines

 . the auto-window-vscroll variable

 . the scroll-preserve-screen-position option

 . signal an error at beginning and end of buffer, subject to the
   value of scroll-error-top-bottom

 . don't let point enter the scroll margin as result of scrolling

 . the window's old_point marker needs to be set after scrolling

There's also a bug when scrolling near the end of buffer: the result
is that the cursor us shown on a line beyond EOB, which should never
happen.

> The only potentially downside I can think of 
> is that it might be slower -- then again I'm just measuring line-heights, and 
> of these there is (at most) only one line that won't eventually be displayed.

It is indeed much slower.  I timed it on xdisp.c using Dmitry's
scroll-up-benchmark function, and found this code to be 3 times slower
than the current implementation.  Turning off font-lock slashes about
40% of the benchmark time, so CC mode fontifications are not the main
reason for the slowdown.  If I compare the existing implementation
with this one on xdisp.c with font-lock-mode turned off in both cases,
this implementation is 16 times slower than what we have now.

For the record, my timings are from an unoptimized build of a recent
trunk, with your code byte-compiled.

The general algorithm seems to be the same as in the current C
implementation, so I doubt an ELisp implementation could match what we
have in speed, let alone be faster.

Now, I personally don't regard the scrolling command as something that
needs to be lightning-fast (although others obviously do, see the
on-going discussions on emacs-devel about that).  But in this case, a
single PageDown keypress takes close to a second to execute, which is
slow enough to annoy.  By contrast, the current implementation is
almost instantaneous.  (Again, this is in an unoptimized build; an
optimized build should be about twice faster, but I think 0.4 sec for
a single scroll might still annoy.)

Finally, it looks like this code forces Emacs to display every single
screen it scrolls through, even when it cannot keep up.  I guess
that's due to the 'redisplay' calls.  This makes the situation where
someone leans on the PageDown key and then releases it very
unpleasant: Emacs keeps scrolling for a long time, and I didn't find a
way of interrupting that.

> If it were to remain mostly elisp, it would need a reliable way to measure the 
> height of a line (essentially a `st-height' replacement), irrespective of 
> whether it's displayed.

Did you try to use pos-visible-in-window-p?  AFAIU, it gives you what
you want, including for lines that are taller than the window.

> It has also proven rather difficult to set the window 
> start "absolutely". I've documented my findings in `st-move'.

Does this happen only when point is on an image?  (The comments in
st-move seem to talk only about this situation.)  If so, could you
show a simple test case to demonstrate the problem?