On Tue, Oct 15, 2019 at 7:48 AM Eli Zaretskii <eliz@gnu.org> wrote:

>
> This is encouraging, but please try C-n near the end of xdisp.c, so
> that it includes syms_of_xdisp.  The first 350 lines of xdisp.c is one
> large comment, and there's a very small number of strings after that.

OK. So same setup as before but running on Windows 10 on a i7
processor:

30 seconds from syms_of_xdisp (void) in src/xdisp.c:
  with antiblink: 34969, around 7% load measured by Process Explorer
  without: 34977, also around 7% load

So I'd venture to say this didn't change much.

I think the reason the C-n and C-v tests escape a big performance hit is
that the potentially expensive operation, 'syntax-ppss', doesn't get
called when you you have switched lines.  So the presence or absence of
strings is irrelevant.  The most expensive operation will probably be
`line-beginning-position` which is in C and apparently fast enough to
not have any kind of relative impact.

But now that I notice, the beginning of jit-lock--antiblink-post-command
could be improved slightly to shortcut earlier and avoid one extra
`line-beginning-position' call in these cases (though probably these
rough benchmarks wouldn't show it.)

So to summarize, for the "paging" scenario, the overhead of the
post-command-based antiblink can be as low a 'copy-marker', an argless
'line-beginning-position' and few tests.

Assuming you're OK with the performace aspects of the thing, I'll do the
fixes you suggested earlier and re-submit (or push?)

João