The Universe of Discourse
           
Thu, 24 Jan 2008

Emacs and alists
[ This article is a few weeks old now. I wrote it and forgot to publish it at the time. ]

Yesterday I upgraded Emacs, and since it was an upgrade, something that had been working for me for fifteen years stopped working, because that's what "upgrade" means. My .emacs file contains:

        (aput 'auto-mode-alist "\\.pl\\'" (function cperl-mode))
        (aput 'auto-mode-alist "\\.t\\'" (function cperl-mode))
        (aput 'auto-mode-alist "\\.cgi\\'" (function cperl-mode))
        (aput 'auto-mode-alist "\\.pm\\'" (function cperl-mode))
        (aput 'auto-mode-alist "\\.blog\\'" (function text-mode))
        (aput 'auto-mode-alist "\\.sml\\'" (function sml-mode))
I should explain this, since I imagine that most readers of this blog are like me in that they touch Emacs Lisp only once a year on Saint Vibrissa's Day. An alist ("association list") is a common data structure in Lisp programs. It is a list of pairs; the first element of each pair is a key, and the second element is an associated value. The pairs in the special auto-mode-alist variable have regexes as their keys and functions as their values. Whenever Emacs opens a new file, it scans this alist, until it finds a regex that matches the name of the file. It then executes the associated function. Thus the effect of the first line above is to have Emacs enable the cperl-mode function on any file whose name ends in ".pl".

The aput function is for maintaining alists. It takes an alist, a key, and a value, scans the alist looking for a matching key, and then if it finds it, it amends the corresponding value. Otherwise, it appends a new association onto the front of the alist.

When I upgraded emacs, this broke. The aput function was moved into a separate package, which I now had to load with (require 'assoc).

I asked about this on IRC, and was told that the correct way to do this, if I did not want to (require 'assoc), was to use the following abomination:

        (mapc (lambda (x) (when (eq 'perl-mode (cdr x)) (setcdr x 'cperl-mode)))
                 (append auto-mode-alist interpreter-mode-alist))
The effect of this is to scan over auto-mode-alist (and also interpreter-mode-alist, a related variable) looking for any association whose value was the perl-mode function, and using setcdr to replace perl-mode with cperl-mode.

(This does not address the issue of what to do with .t files or .blog files, for which no association exists yet, presumably, but I did not ask about those specifically on IRC.)

I was totally boggled. Choosing the right editing mode for a file is a basic function of emacs. I could not believe that the best and simplest way to add or change associations was to use mapc lambda gobhorn oleo potatopudding quote potrzebie. I was assured that this was indeed the only correct method. Struck almost speechless, I managed to come up with "Bullshit."

Apparently the issue was that if auto-mode-alist already contains an association for ".pl", there is no guarantee that my new association will be found and preferred to the old one, unless I somehow remove the old one, or edit it to be the way I want.

This seemed very unlikely to me. You see, an alist is a list. This means that it is searched from head to tail, because this is the only way a list can be searched. So in particular, if you cons a second association to the front of the list, which has the same key as a later (older) association, the search will find the new one first, and the older one becomes inoperative. I asked if there was not a guarantee that the alist would be searched from front to back. I was told that there is not.

I looked in the manual, and reported that the assoc function, which is the getter that corresponds to aput, taking an alist and a key, and returning the corresponding value, is expressly guaranteed to return the first matching item. I was told that there was no guarantee that assoc would be used.

I pondered the manual some more and found this passage:

However, association lists have their own advantages. Depending on your application, it may be faster to add an association to the front of an association list than to update a property.
That is, it is expressly endorsing the technique of adding a new item to the front of an alist in order to override any later item that might have the same key.

After finding that the add-to-the-front technique really did work, I reasoned that if someday Emacs stopped searching alists sequentially, I would not be in any more trouble than I had been today when they removed the aput function.

So I did not take the advice I was given. Instead, I left it pretty much the way it was. I did take the opportunity to clean up the code a bit:

        (push '("\\.pl\\'" . cperl-mode) auto-mode-alist)
        (push '("\\.t\\'" .  cperl-mode) auto-mode-alist)
        (push '("\\.cgi\\'" . cperl-mode) auto-mode-alist)
        (push '("\\.pm\\'" . cperl-mode) auto-mode-alist)
        (push '("\\.blog\\'" . text-mode) auto-mode-alist)
        (push '("\\.sml\\'" . sml-mode) auto-mode-alist)
The push function simply appends an element to the front of a list, modifying the list in-place.

But wow, the advice I got was phenomenally bad. It was bad in a really interesting way, too. It reminded me of the advice people get on the #math channel, where some guy comes in with some question about triangles and gets the category-theoretic viewpoint on triangles as natural transformations of something or other. The advice was bad because although it was correct, it was completely devoid of common sense.

[ Addendum 20080124: It has been brought to my attention that the Emacs FAQ endorses my solution, which makes the category-theoretic advice proposed by the #emacs blockheads even less defensible. ]

[ Addendum 20080201: Steve Vinoski suggests replacing the aput function. ]


[Other articles in category /prog] permanent link