bug#39015: `isearch-complete1' should use `completion-at-point' or `completion-in-region'

bug#39015: `isearch-complete1' should use `completion-at-point' or `completion-in-region'

[Alexander Shukaev]

Not sure why in the first place `isearch-complete1' uses some custom 
auto-completion implementation that is different from a conventional 
solution, perhaps historical reasons.  Thus, I believe that reusing 
either `completion-at-point' or `completion-in-region' to implement that 
functionality is much better in the long run.  This would also allow 
packages, which customize completion behavior, to hook into `isearch' 
auto-completion as well.  See also [1] for example.

[1] https://github.com/abo-abo/swiper/issues/1882

Re: `isearch-complete1' should use `completion-at-point' or `completion-in-region' (was: bug#39015)

[Juri Linkov]

> Not sure why in the first place `isearch-complete1' uses some custom
> auto-completion implementation that is different from a conventional 
> solution, perhaps historical reasons.  Thus, I believe that reusing either
> `completion-at-point' or `completion-in-region' to implement that
> functionality is much better in the long run.  This would also allow 
> packages, which customize completion behavior, to hook into `isearch'
> auto-completion as well.  See also [1] for example.
>
> [1] https://github.com/abo-abo/swiper/issues/1882

Indeed, Isearch uses custom completion for historical reasons,
and there is no reason for this anymore.

But there is one thing I don't understand: how packages prefer to hook
into auto-completion?  Is it enough to convert `isearch-complete1' to
just `completing-read' in the minibuffer activated by `isearch-edit-string'?

Re: `isearch-complete1' should use `completion-at-point' or `completion-in-region'

[Alexander Shukaev]

On 13/01/2020 00:36, Juri Linkov wrote:
>> Not sure why in the first place `isearch-complete1' uses some custom
>> auto-completion implementation that is different from a conventional
>> solution, perhaps historical reasons.  Thus, I believe that reusing either
>> `completion-at-point' or `completion-in-region' to implement that
>> functionality is much better in the long run.  This would also allow
>> packages, which customize completion behavior, to hook into `isearch'
>> auto-completion as well.  See also [1] for example.
>>
>> [1] https://github.com/abo-abo/swiper/issues/1882
> 
> Indeed, Isearch uses custom completion for historical reasons,
> and there is no reason for this anymore.
> 
> But there is one thing I don't understand: how packages prefer to hook
> into auto-completion?  Is it enough to convert `isearch-complete1' to
> just `completing-read' in the minibuffer activated by `isearch-edit-string'?
> 

Exactly, e.g. `ivy' [1] would pick it up out of the box.

[1] https://github.com/abo-abo/swiper/blob/master/ivy.el

Re: `isearch-complete1' should use `completion-at-point' or `completion-in-region'

[Stefan Monnier]

> But there is one thing I don't understand: how packages prefer to hook
> into auto-completion?  Is it enough to convert `isearch-complete1' to
> just `completing-read' in the minibuffer activated by `isearch-edit-string'?

First, there are 2 different cases: `isearch-complete-edit` and `isearch-complete`.

For `isearch-complete-edit` we could indeed use `completing-read`, but
I don't think I like this idea (e.g. I think it'd be odd for Icomplete
to kick in when in `isearch-edit`).  So I think the better solution is
to set `completion-at-point-functions` and then turn
`isearch-complete-edit` into an alias for `completion-at-point`.

For `isearch-complete`, it's a bit more tricky because it's a form of
completion that is performed on a text that's kept inside a string
rather than inside a buffer, so I think we'll still need some ad-hoc
code, but it could better use our completion framework.


        Stefan

Re: `isearch-complete1' should use `completion-at-point' or `completion-in-region'

[Juri Linkov]

> First, there are 2 different cases: `isearch-complete-edit` and `isearch-complete`.
>
> For `isearch-complete-edit` we could indeed use `completing-read`, but
> I don't think I like this idea (e.g. I think it'd be odd for Icomplete
> to kick in when in `isearch-edit`).

But the original request was just about allowing completion packages
to kick in when in `isearch-edit`, and Icomplete is one kind of such packages.

> So I think the better solution is to set `completion-at-point-functions`
> and then turn `isearch-complete-edit` into an alias for `completion-at-point`.

Maybe something like this:

(defun isearch-completion-at-point-function ()
  (list (minibuffer-prompt-end) (point-max)
        (if isearch-regexp regexp-search-ring search-ring)))

(defun isearch-completion-at-point ()
  (interactive)
  (let* ((completion-ignore-case case-fold-search)
         (completion-at-point-functions '(isearch-completion-at-point-function)))
    (completion-at-point)))

(define-key minibuffer-local-isearch-map "\M-\t" 'isearch-completion-at-point)

But I'm not sure whether this could help completion packages like swiper
to hook into isearch completion.

> For `isearch-complete`, it's a bit more tricky because it's a form of
> completion that is performed on a text that's kept inside a string
> rather than inside a buffer, so I think we'll still need some ad-hoc
> code, but it could better use our completion framework.

`isearch-complete` already activates the minibuffer conditionally
when there are completions.  It could activate the minibuffer always,
then immediately exit the minibuffer when completion-at-point
finds no completions.

Re: `isearch-complete1' should use `completion-at-point' or `completion-in-region'

[Stefan Monnier]

> Maybe something like this:
>
> (defun isearch-completion-at-point-function ()
>   (list (minibuffer-prompt-end) (point-max)
>         (if isearch-regexp regexp-search-ring search-ring)))

Looks fine, yes.

> (defun isearch-completion-at-point ()
>   (interactive)
>   (let* ((completion-ignore-case case-fold-search)
>          (completion-at-point-functions '(isearch-completion-at-point-function)))
>     (completion-at-point)))

Ah, no I was thinking rather of something like

    (minibuffer-with-setup-hook
        (lambda ()
          (setq-local completion-ignore-case case-fold-search)
          (add-hook 'completion-at-point-functions
                    #'isearch-completion-at-point-function nil t))
      ...)

This way, things like `company-mode` can use it (tho IIUC company-mode
currently doesn't work in the minibuffer) and `isearch-complete-edit`
can be an obsolete alias of `completion-at-point`.

>> For `isearch-complete`, it's a bit more tricky because it's a form of
>> completion that is performed on a text that's kept inside a string
>> rather than inside a buffer, so I think we'll still need some ad-hoc
>> code, but it could better use our completion framework.
> `isearch-complete` already activates the minibuffer conditionally
> when there are completions. It could activate the minibuffer always,
> then immediately exit the minibuffer when completion-at-point
> finds no completions.

Yes.  What I meant is that `isearch-complete` can't be turned into
a mere obsolete alias of `completion-at-point`.

FWIW, I think it might be worthwhile trying to rewrite isearch such that
it always uses the minibuffer (instead of using a transient keymap and
displaying the search string in the echo area).

Then `isearch-edit` would simply change the minibuffer's keymap such
that we only exit the minibuffer in the "usual" way.

In such a setting `isearch-complete` could be a simple alias of
`completion-at-point`.


        Stefan

Re: `isearch-complete1' should use `completion-at-point' or `completion-in-region'

[Juri Linkov]

> FWIW, I think it might be worthwhile trying to rewrite isearch such that
> it always uses the minibuffer (instead of using a transient keymap and
> displaying the search string in the echo area).

For the sake of backward-compatibility it would be easier
to write a new package implementing just a mapping from keys typed
in the isearch minibuffer to the search functions from isearch.el
invoked on the original buffer.

bug#39015: `isearch-complete1' should use `completion-at-point' or `completion-in-region'

[Juri Linkov]

> Not sure why in the first place `isearch-complete1' uses some custom
> auto-completion implementation that is different from a conventional
> solution, perhaps historical reasons.  Thus, I believe that reusing either
> `completion-at-point' or `completion-in-region' to implement that
> functionality is much better in the long run.  This would also allow
> packages, which customize completion behavior, to hook into `isearch'
> auto-completion as well.  See also [1] for example.
>
> [1] https://github.com/abo-abo/swiper/issues/1882

Based on this solution
https://lists.gnu.org/archive/html/emacs-devel/2020-01/msg00447.html
it should be possible to write a patch that would hook into `isearch'
auto-completion.  I seems I can't do this since I don't understand
how to test it.