Re: [patch suggestion] Mitigating the poor Emacs performance on huge org files: Do not use overlays for PROPERTY and LOGBOOK drawers

[Christian Heinrich]

Hi,

thanks for your (initial) patch! I traced another error down today and found 
your code by chance. I
tested it on an org-drill file that I had (with over 3500 items and hence 3500 
drawers) and this
patch helps *a lot* already. (Performance broke in 
4403d4685e19fb99ba9bfec2bd4ff6781c66981f when
outline-flag-region was replaced with org-flag-region, as drawers are no longer 
opened using
outline-show-all which I had to use anyways to deal with my huge file.)

I am not sure I understand how your follow-up code (below) needs to be 
incorporated. Would you mind
sending a patch file? I hope that this ends up in the master branch at some 
point.

Thanks again!
Christian

On Mon, 2020-04-27 at 00:04 +0800, Ihor Radchenko wrote:
> > You cannot. You may however mimic it with `cursor-sensor-functions' text
> > property. These assume Cursor Sensor minor mode is active, tho.
> > I haven't tested it, but I assume it would slow down text properties
> > a bit, too, but hopefully not as much as overlays.
> 
> Unfortunately, isearch sets inhibit-point-motion-hooks to non-nil
> internally. Anyway, I came up with some workaround, which seems to work
> (see below). Though it would be better if isearch supported hidden text
> in addition to overlays.
> 
> > Missing `isearch-open-invisible' is a deal breaker, IMO. It may be worth
> > experimenting with `cursor-sensor-functions'.
> 
> So far, I came up with the following partial solution searching and
> showing hidden text.
> 
> ;; Unfortunately isearch, sets inhibit-point-motion-hooks and we
> ;; cannot even use cursor-sensor-functions as a workaround
> ;; I used a less ideas approach with advice to isearch-search-string as
> ;; a workaround 
> 
> (defun org-find-text-property-region (pos prop)
>   "Find a region containing PROP text property around point POS."
>   (require 'org-macs) ;; org-with-point-at
>   (org-with-point-at pos
>     (let* ((beg (and (get-text-property pos prop) pos))
>          (end beg))
>       (when beg
>       (setq beg (or (previous-single-property-change pos prop)
>                     beg))
>       (setq end (or (next-single-property-change pos prop)
>                     end))
>         (unless (equal beg end)
>           (cons beg end))))))
> 
> ;; :FIXME: re-hide properties when point moves away
> (define-advice isearch-search-string (:after (&rest _) put-overlay)
>   "Reveal hidden text at point."
>   (when-let ((region (org-find-text-property-region (point) 'invisible)))
>     (with-silent-modifications
>       (put-text-property (car region) (cdr region) 'org-invisible 
> (get-text-property (point)
> 'invisible)))
>       (remove-text-properties (car region) (cdr region) '(invisible nil))))
> 
> ;; this seems to be unstable, but I cannot figure out why
> (defun org-restore-invisibility-specs (&rest _)
>   ""
>    (let ((pos (point-min)))
>      (while (< (setq pos (next-single-property-change pos 'org-invisible nil 
> (point-max))) (point-
> max))
>        (when-let ((region (org-find-text-property-region pos 'org-invisible)))
>          (with-silent-modifications
>            (put-text-property (car region) (cdr region) 'invisible 
> (get-text-property pos 'org-
> invisible))
>            (remove-text-properties (car region) (cdr region) '(org-invisible 
> nil)))))))
> 
> (add-hook 'post-command-hook #'org-restore-invisibility-specs)
> 
> (defun org-flag-region (from to flag spec)
>   "Hide or show lines from FROM to TO, according to FLAG.
> SPEC is the invisibility spec, as a symbol."
>   (pcase spec
>     ('outline
>      (remove-overlays from to 'invisible spec)
>      ;; Use `front-advance' since text right before to the beginning of
>      ;; the overlay belongs to the visible line than to the contents.
>      (when flag
>        (let ((o (make-overlay from to nil 'front-advance)))
>        (overlay-put o 'evaporate t)
>        (overlay-put o 'invisible spec)
>        (overlay-put o 'isearch-open-invisible #'delete-overlay))))
>     (_
>      (with-silent-modifications
>        (remove-text-properties from to '(invisible nil))
>        (when flag
>        (put-text-property from to 'invisible spec)
>        )))))
> 
> ;; This normally deletes invisible text property. We do not want this now.
> (defun org-unfontify-region (beg end &optional _maybe_loudly)
>   "Remove fontification and activation overlays from links."
>   (font-lock-default-unfontify-region beg end)
>   (let* ((buffer-undo-list t)
>        (inhibit-read-only t) (inhibit-point-motion-hooks t)
>        (inhibit-modification-hooks t)
>        deactivate-mark buffer-file-name buffer-file-truename)
>     (decompose-region beg end)
>     (remove-text-properties beg end
>                           '(mouse-face t keymap t org-linked-text t
>                                        ;; Do not remove invisible during 
> fontification                        
>                
>                                        ;; invisible t
>                                          intangible t
>                                        org-emphasis t))
>     (org-remove-font-lock-display-properties beg end)))
> 
> > Anyway, the real fix should come from Emacs itself. There are ways to
> > make overlays faster. These ways have already been discussed on the
> > Emacs devel mailing list, but no one implemented them. It is a bit sad
> > that we have to find workarounds for that.
> 
> I guess that it is a very old story starting from the times when XEmacs
> was a thing [1]. I recently heard about binary tree implementation of
> overlays (there should be a branch in emacs git repo) [2], but there was
> no update on that branch for a while. So, I do not have much hope on
> Emacs implementing efficient overlay access in the near future. (And I
> have problems with huge org files already).
> 
> [1] 
> https://www.reddit.com/r/planetemacs/comments/e9lgwn/history_of_lucid_emacs_fsf_emacs_schism/
> [2] https://lists.gnu.org/archive/html/emacs-devel/2019-12/msg00323.html
> 
> 
> Nicolas Goaziou <address@hidden> writes:
> 
> > Hello,
> > 
> > Ihor Radchenko <address@hidden> writes:
> > 
> > > To my surprise, the patch did not break org to unusable state and
> > > the performance on the sample org file [3] improved drastically. You can
> > > try by yourself!
> > 
> > It is not a surprise, really. Text properties are much faster than
> > overlays, and very close to them features-wise. They are a bit more
> > complex to handle, however.
> > 
> > > However, this did introduce some visual glitches with drawer display.
> > > Though drawers can still be folded/unfolded with <tab>, they are not
> > > folded on org-mode startup for some reason (can be fixed by running
> > > (org-cycle-hide-drawers 'all)). Also, some drawers (or parts of drawers)
> > > are unfolded for no apparent reason sometimes. A blind guess is that it
> > > is something to do with lack of 'isearch-open-invisible, which I am not
> > > sure how to set via text properties.
> > 
> > You cannot. You may however mimic it with `cursor-sensor-functions' text
> > property. These assume Cursor Sensor minor mode is active, tho.
> > I haven't tested it, but I assume it would slow down text properties
> > a bit, too, but hopefully not as much as overlays.
> > 
> > Note there are clear advantages using text properties. For example, when
> > you move contents around, text properties are preserved. So there's no
> > more need for the `org-cycle-hide-drawer' dance, i.e., it is not
> > necessary anymore to re-hide drawers.
> > 
> > > Any thoughts about the use of text properties or about the patch
> > > suggestion are welcome.  
> > 
> > Missing `isearch-open-invisible' is a deal breaker, IMO. It may be worth
> > experimenting with `cursor-sensor-functions'.
> > 
> > We could also use text properties for property drawers, and overlays for
> > regular ones. This might give us a reasonable speed-up with an
> > acceptable feature trade-off.
> > 
> > Anyway, the real fix should come from Emacs itself. There are ways to
> > make overlays faster. These ways have already been discussed on the
> > Emacs devel mailing list, but no one implemented them. It is a bit sad
> > that we have to find workarounds for that.
> > 
> > Regards,
> > 
> > -- 
> > Nicolas Goaziou

signature.asc
Description: This is a digitally signed message part