Re: master 49e06183f5 1/3: Allow REQUIRE-MATCH to be a function

[Stefan Monnier <monnier@iro.umontreal.ca>]

Hi Lars,

>     Allow REQUIRE-MATCH to be a function
>     
>     * doc/lispref/minibuf.texi (Minibuffer Completion): Document it.
>     
>     * lisp/minibuffer.el (completion--complete-and-exit): Allow
>     REQUIRE-MATCH to be a function.
>     (read-file-name): Mention it.

I see the motivation is for things like `read-file-name` where we
can't specify a completion table (otherwise we could also get the same
result with an appropriate completion table whose `test-completion` is
adjusted accordingly).

But besides the `test-completion` part of the completion table, this new
feature also overlaps with the `predicate` argument.  I think it would
be good to write down somewhere how those three compare.

I wonder if that can also be used to solve the problem of completing
file names that "end in .pdf": previously using `test-completion` (via
the completion table or via `predicate`) had the problem that such
a test either ruled out completions to "foo/bar/" (which can be
indispensable on the way to the final "foo/bar/baz.pdf") or spuriously
allow hitting RET on "foo/bar/" even tho it doesn't end in `.pdf`.

> +   ;; The CONFIRM argument is a predicate.
> +   ((and (functionp minibuffer-completion-confirm)
> +         (funcall minibuffer-completion-confirm
> +                  (buffer-substring beg end)))
> +    (funcall exit-function))
> +   ;; See if we have a completion from the table.
> +   ((test-completion (buffer-substring beg end)
> +                     minibuffer-completion-table
> +                     minibuffer-completion-predicate)
> +    ;; FIXME: completion-ignore-case has various slightly
> +    ;; incompatible meanings.  E.g. it can reflect whether the user
> +    ;; wants completion to pay attention to case, or whether the
> +    ;; string will be used in a context where case is significant.
> +    ;; E.g. usually try-completion should obey the first, whereas
> +    ;; test-completion should obey the second.
> +    (when completion-ignore-case
> +      ;; Fixup case of the field, if necessary.
> +      (let* ((string (buffer-substring beg end))
> +             (compl (try-completion
> +                     string
> +                     minibuffer-completion-table
> +                     minibuffer-completion-predicate)))
> +        (when (and (stringp compl) (not (equal string compl))
> +                   ;; If it weren't for this piece of paranoia, I'd replace
> +                   ;; the whole thing with a call to do-completion.
> +                   ;; This is important, e.g. when the current minibuffer's
> +                   ;; content is a directory which only contains a single
> +                   ;; file, so `try-completion' actually completes to
> +                   ;; that file.
> +                   (= (length string) (length compl)))
> +          (completion--replace beg end compl))))
> +    (funcall exit-function))
> +   ;; The user is permitted to exit with an input that's rejected
> +   ;; by test-completion, after confirming her choice.
> +   ((memq minibuffer-completion-confirm '(confirm confirm-after-completion))
> +    (if (or (eq last-command this-command)
> +            ;; For `confirm-after-completion' we only ask for confirmation
> +            ;; if trying to exit immediately after typing TAB (this
> +            ;; catches most minibuffer typos).
> +            (and (eq minibuffer-completion-confirm 'confirm-after-completion)
> +                 (not (memq last-command minibuffer-confirm-exit-commands))))
>          (funcall exit-function)
> -        (minibuffer-message "Confirm")
> -        nil))
> +      (minibuffer-message "Confirm")
> +      nil))

I see that the `minibuffer-completion-confirm` function completely
replaces the usual handling of require-match (i.e. the attempt to fix
the case and the prompting for confirmation).

Maybe this should be better documented, and also AFAICT currently
a `minibuffer-completion-confirm` function cannot reliably reproduce this
behavior by hand because it doesn't have access to `beg` and `end`.
I think we should make this "default behavior" more easily accessible
(e.g. put it into its own function and document it as something that can
be called from `minibuffer-completion-confirm`?).


        Stefan