bug#23928: 25.0.95; Performance regression observable with smartparens

[Eli Zaretskii]

> From: Aaron Jensen <aaronjensen@gmail.com>
> Date: Sun, 10 Jul 2016 02:55:31 +0000
> Cc: 23928@debbugs.gnu.org
> 
> I’ve added profiles of both here:  
> https://gist.github.com/aaronjensen/a1c2b56eac192b0d9cac76b493822070
> 
> In this profile, they’re showing pretty much the exact same performance, so 
> that doesn’t tell us much. I believe the performance degradation is something 
> that happens over time, and these are fresh runs of emacs. I’ll have to try 
> and get from 25 next time it slows down on me.

Thanks.

That rings a bell.  The hottest function on these profiles is
sp--looking-back, which calls in a loop sp--looking-at, which does
this:

  (defun sp--looking-at (regexp)
    "Like `looking-at', but always case sensitive."
    (let ((case-fold-search nil))
      (looking-at regexp)))

As you will see from bug#18522, binding case-fold-search can be very
expensive, especially if you have a lot of buffers.  See
http://debbugs.gnu.org/cgi/bugreport.cgi?bug=18522#201 for a summary
of the findings there.  It clearly shows in your profile:

                 sp--looking-at                171  12%
                  let                           12   0%

As to why you see a slowdown in Emacs 25, I don't know.  Maybe you
have more buffers crop up there as the session goes on?

> One curious thing is that the emacs 24 profile shows all of the builtins 
> (let/if/save-excusion) and the emacs 25 profile does not. Is there something 
> I’m doing wrong that is causing that?

Could it be that in Emacs 24 you loaded the library as a .el file, and
in Emacs 25 as a .elc file?

> I’ll work on this. It seems possible that it will be very hard to track down 
> given the nature—the performance isn’t *always* bad, it just becomes bad at 
> some point in the session.

When it becomes bad, count the number of buffers in each session.
Bug#18522 provides some tools for that.

>     And please don't use elp.el to produce profiles, use profiler.el,
>     because the latter can profile primitives as well.
> 
> I don’t even know how to use elp.el, was there some indication that I was?

Sorry, my bad.  It looked like elp.el output to me.