Stefan Monnier <monnier@iro.umontreal.ca> writes:

>> Here's another attempt. Like the previous patch, I define
>> key-binding-keymap which finds the keymap by mimicking key-binding, and
>> describe-key--binding-locus which matches the keymap to a symbol and
>> makes a description suitable for describe-key.
>
> Looks pretty good.  Feel free to install it into `trunk'.

I can send a patch, including a ChangeLog entry, but I don't have
permission to commit.

>> +      (when (integerp found)
>> +        ;; prefix was found but not the whole sequence
>> +        (setq found nil)))
>
> If key-binding returned a command, then we should never hit this case.
> It's OK to keep the code, but you might like to add a comment mentioning
> that we don't expect this case to happen.

We do hit the case, because we look in active keymaps one by one,
instead of using the keymap (cons 'keymap active-keymaps) like
key-binding does.

>> +         ;; look into these advertised symbols first
>> +         (while (and (not found) advertised-syms)
>> +           (let ((sym (pop advertised-syms)))
>> +             (setq found
>> +                   (when (and
>> +                          (boundp sym)
>> +                          (eq map (symbol-value sym)))
>> +                     sym))))
>> +         ;; only look in other symbols otherwise
>> +         (when (not found)
>> +           (mapatoms
>> +            (lambda (x)
>> +              (when (and (boundp x)
>> +                         ;; avoid let-bound symbols
>> +                         (special-variable-p x)
>> +                         (eq (symbol-value x) map))
>> +                (push x found))))
>> +           (when (and (consp found)
>> +                      (null (cdr found)))
>> +             (setq found (car found))))
>
> I think this code will be simpler if you use catch/throw (you can then
> use dolist over advertised-syms, for example and you don't need
> `found').

Shall I return only the first symbol that matches, also in the mapatoms
case ? I assume yes because of your next suggestion :

> I think it's better to make this function return the symbol rather than
> a string, and let the caller turn it into a string.

Ok.

Here's the patch.