Eli Zaretskii <eliz@gnu.org> writes:

> My advice is to have a prototype working, then time it on local
> filesystems.

See the attached patch.

I used the following simplified reproducer that does not involve loading
Org (which skews the relative numbers as Org loading is relatively slow):

(dotimes (i 1000) (with-temp-file (format "/tmp/test/%d.txt" i) (insert "* This is test")))
(dolist (file (directory-files "/tmp/test/" t "txt"))
  (find-file-noselect file))

Without the patch (cpu-profiler-without-patch): 4.3 sec
With the patch    (cpu-profiler-w-patch      ): 2.5 sec

> ... Optimizing for networked filesystems is the next step,
> assuming it is needed.  Please keep in mind that the current code does
> that as well, only from Lisp: the call to file-attributes calls
> 'stat', then conses the 11-member list that is the return value.  So
> the C implementation cannot possibly be worse.

I left the `file-attributes' call in Elisp. Looking at the
cpu-profiler-w-patch, `find-buffer-visiting' is no longer the main
contributor to CPU time. I am not sure if we really need to squeeze the
performance yet further from `find-buffer-visiting' - `file-attributes'
is taking pretty much no time:

(reverse call-tree)
         924  36%   Automatic GC
         173   6% + inhibit-local-variables-p
         172   6% + locate-dominating-file
         139   5% + abbreviate-file-name
         113   4% + dir-locals--all-files
         109   4% + file-truename
          92   3% + find-buffer-visiting 
...
           6   0% + file-attributes 

comapare with cpu-profiler-without-patch:

(reverse call-tree)
        1714  39% + find-buffer-visiting
        1131  26%   Automatic GC
         202   4% + locate-dominating-file
         147   3% + abbreviate-file-name
         140   3% + inhibit-local-variables-p
         104   2% + dir-locals--all-files
          98   2% + uniquify-rationalize-file-buffer-names
          91   2% + file-truename