Translating Keys For Fun And Profit

11/09/2015

I’ve got to admit that Vim is pretty clever when it comes to its key bindings. Never did I notice that TAB and C-i are the same thing in the terminal, simply because I’ve always used C-i in normal mode for traversing the jump list and TAB in insert mode for inserting indentation. With Emacs it’s quite different. TAB doesn’t just insert a preset amount of indentation, no, it’s usually bound to a command that adjusts the indentation of the current line or region to be just about right. Therefore this key makes sense outside of Evil’s insert state, simply because it is a convenient way of fixing the indentation of a piece of code. That’s why one of the first tweaks to my Evil setup was unbinding TAB (and SPC and RET) in its motion and visual state map. Too bad this gets rid of C-i for jumping forward. Surely it’s possible to have my cake and eat it?

Turns out it is if I use the GUI version as it distinguishes between the key symbol <tab> which is exclusive to the tab key and TAB, a key code both C-i and TAB can resolve to unless <tab> has been bound. My first attempt did look like this:

(with-eval-after-load 'evil-maps
  (define-key evil-motion-state-map (kbd "TAB") 'evil-jump-forward))

This undoes my earlier unmapping, simply because in modes where <tab> hasn’t been bound yet, both TAB and C-i will now jump forward. Back to square one.

After studying the Emacs Lisp manual a good bit, I’ve learned Emacs comes with not one, not two, but three special keymaps specially designed for translating keypresses, sorted in order of lookup:

input-decode-map
Mainly used for working around terminal-specific oddities. Terminal-local.
local-function-key-map
Turns unbound key symbols into more preferable keys. Terminal-local. function-key-map is the global variant.
key-translation-map
Generic map for translating one key into another. The manual recommends using it for self-inserting keys…

Since I’m neither interested in working around the peculiarities of terminal-local keymaps nor translating unbound symbols, key-translation-map it is! Any of its bindings can either take a key or a function receiving a PROMPT argument you’re unlikely to use:

(defun evil (prompt)
  (if prompt (kbd "C-c") (kbd "C-x")))

(define-key key-translation-map (kbd "C-c") 'evil)

Good luck figuring out why C-x C-c doesn’t work although F1 k C-x C-c shows a correctly looked up command!

Back to the original problem. First of all, I need something less ambiguous than TAB and morally equivalent to <tab>. Evaluating (kbd "<tab>") confirms that this key binding is a vector holding tab as its only element. Therefore it should be fine to use (kbd "<C-i>") for translation purposes.

Next, the key shouldn’t be translated unconditionally. A check whether it’s not part of a longer key sequence and used in Evil’s normal state should suffice:

(defun my-translate-C-i (_prompt)
  (if (and (= (length (this-command-keys-vector)) 1)
           (= (aref (this-command-keys-vector) 0) ?\C-i)
           (bound-and-true-p evil-mode)
           (eq evil-state 'normal))
      (kbd "<C-i>")
    (kbd "TAB")))

With the translation function returning <C-i> for the desired special case, it can be bound to the correct action in one of Evil’s many keymaps:

(define-key key-translation-map (kbd "TAB") 'my-translate-C-i)

(with-eval-after-load 'evil-maps
  (define-key evil-motion-state-map (kbd "<C-i>") 'evil-jump-forward))

Evil is using a similiar trick for a different purpose: Distinguishing the Escape key from key bindings involving Meta. Vim avoids this problem again by not using the Meta key in any of its stock key bindings. If you hit, say M-O in insert mode, this will exit insert mode, open a line above the current one and enter insert mode again[1]. In Emacs hitting ESC will either trigger keyboard-escape-quit or allow for entering a command containing a meta key in an alternative manner.

Their implementation differs from mine in a few important aspects: First of all, input-decode-map is used to allow for doing the translation in terminals only. This results in a minor inconvenience: The binding needs to be done for every terminal with the terminal selected. Second, evil-esc checks for more than just the state and length of the binding: It waits for a customizable amount of time to ensure the key event wasn’t part of a key chord. Finally, they cannot make use of a function like I did because if the translation cannot be applied, translating with the original key sequence will screw up more complex chords.

So, how did they solve this? Well, define-key accepts more definitions than just keyboard macros, symbols and commands. It holds part of the secret of how one defines new menu bindings, simply because in Emacs a menu item is just another key binding with its own set of rules. One of them is a :filter property that takes a function for translating a MAP argument to something different, including another key binding. In case the translation doesn’t end up being chosen, MAP can be simply returned. I’m not sure whether this couldn’t be emulated with one of the many functions returning the keys associated to the current command, but anyway, it works pretty well as is.

[1]Some people even abuse this feature on a regular basis which led to a bug report on Evil’s issue tracker by someone seeking to have it in Emacs as well!