bug#44273: "total used in directory 19 available 5.2 GiB"

[Eli Zaretskii <eliz@gnu.org>]

> From: Lars Ingebrigtsen <larsi@gnus.org>
> Date: Wed, 28 Oct 2020 12:02:06 +0100
> Cc: 積丹尼 Dan Jacobson <jidanni@jidanni.org>,
>  44273@debbugs.gnu.org
> 
> The argument that Tramp needs to parse "ls" output is valid, and that
> "ls" is faster than `ls-lisp' is, too, but it's still backwards: Dired
> should take a well-defined data structure and then render it according
> to however the user wants.

I think it would be useful to measure the difference in speed between
ls-lisp and 'ls' the program on various platforms.  For example, use
benchmark-run to time (dired "foo") with some large directory 'foo',
with and without ls-lisp loaded.  (This should be done with a warm
disk cache.)  This should tell us how badly we need to speed up
ls-lisp.

Profiling ls-lisp could also be educational.  For example, I see in
the profile on my system that this loop:

	(let ((locale system-time-locale))
	  (if (not locale)
	      (let ((vars '("LC_ALL" "LC_TIME" "LANG")))
		(while (and vars (not (setq locale (getenv (car vars)))))
		  (setq vars (cdr vars)))))

is run for each file, which is definitely a waste of cycles.  My
measurements indicate that running this loop just once could speed up
ls-lisp by 25%.

> The data can come from `ls-lisp', but it could also come from "ls": We
> just need to write a parser that parses the data.

The data which comes from 'ls' can take many different formats,
depending on the switches.  Parsing each and every one of them sounds
daunting.