Performance regression in commit e98883f50f5794

[hpdeifel]

Ivy is currently so slow for me that it is barely usable. Typing letters has a noticeable lag. git bisect revealed commit e98883f as the culprit.

This lag happened in all ivy commands that I tested (counsel-M-x, ivy-find-file and a few others).

I can't reproduce it with plain emacs -Q, but calling (package-initialize) with an otherwise empty config suffices to show the bug.

Replacing (require 'flx nil 'noerror) with nil in ivy--format-minibuffer-line also makes ivy fast again.

[abo-abo]

Works fine for me. Here's the profiler result:

ivy--exhibit                              33          0.271637048   0.0082314256

Less than 0.01 second is spent on filtering and redisplay after each key press.

[hpdeifel]

I get

 ivy--exhibit   59          8.5247626979  0.1444875033

that doesn't sound much, but feels incredibly laggy.

profiler.el says that 86% of the time is spend in the and in ivy--format-minibuffer-line where the (require 'flx nil 'noerror) is that I mentioned. Strangely it doesn't show any children of this and.

Keep in mind that this only happens when (package-initialize) is called. I have quite a number of packages installed, so maybe the require has to go through all these package include paths to determine that flx is not there. Profiling require itself gives me:

require        1276        8.8469242039  0.0069333261

But that must be wrong, since replacing that require with nil fixes the problem.

[abo-abo]

You mean removing only the require or the whole highlighting statement? Removing the require is, of course, and easy fix.

[hpdeifel]

I mean this diff:

diff --git a/ivy.el b/ivy.el
index e74c0c5..7468e2e 100644
--- a/ivy.el
+++ b/ivy.el
@@ -2695,7 +2695,7 @@ SEPARATOR is used to join the candidates."
                       str))
                    (cl-incf i)))))
             ((and
-              (require 'flx nil 'noerror)
+              nil
               (or (eq ivy--regex-function 'ivy--regex-fuzzy)
                   (and (eq ivy--regex-function 'swiper--re-builder)
                        (let ((caller (ivy-state-caller ivy-last)))

Since I don't have flx installed, this is equivalent in my case and much faster

[jferguson-gnubio]

I also got a huge performance regression, that I assume is the same thing, moving from counsel-20160826.2343 to counsel-20160906.750 from melpa. Sorry I haven't done a proper bisect, but it sounds the same. counsel-M-x and counsel-find-file are very slow to start up.

[jferguson-gnubio]

Apologies I'm not particularly proficient at Elisp, profiling etc, but with some experimentation on a variation of the patch hpdeifel suggested, for counsel-M-x

(require 'flx nil 'noerror) => 5627 cpu samples
(featurep 'flx)             => 942 cpu samples

I think the culprit might be that I have the auto-compile package, with auto-compile-on-load, which seems to result in a lot of the require overhead.

[abo-abo]

Please check if it's better now.

[hpdeifel]

Yes, that's much better. Thanks!

[jferguson-gnubio]

Works nicely - excellent, thank you.

[abo-abo]

Thanks for the confirmation.