bug#68246: 30.0.50; Add non-TS mode as extra parent of TS modes

["João Távora" <joaotavora@gmail.com>]

On Tue, Jan 16, 2024 at 2:09 AM Dmitry Gutov <dmitry@gutov.dev> wrote:

> > We could define them all in batch in a macro, that's not too bad.  And
> > then let the existing fleshed out ones overwrite those definitions by
> > making sure to load them later.
>
> A keyword for define-derived-mode like (:base t)? That would work.

I think that would just clash with the existing PARENT one (when would
one of those bases _not_ be the parent).  For base modes, I think
the current approach is mostly fine.  There's something we can do
for the existing ones (and future ones) though, which is to explicitly
mark them abstract somehow.

> > The main advantages of the foo-base-mode approach is that:
> >
> > * it is easily grokkable by everybody, as it is very simply based on
> >    simple inheritance, which everybody knows already.
> >
> > * there's already a fair number of such modes in the tree.
>
> Agree.
>
> I guess the part I don't quite like is adding a lot more new -base-
> modes (we'll have to add one for every prog mode, at least), most of
> which would stay unused, but unlike hook variables, clutter the function
> namespace.

Agree.  And this is why I'm not crazy about the solution either.  But as
to cluttering the function namespace we could say that (:abstract t) modes
do _not_ generate a function (or do not generate them in the public
namespace -- as I think the function still has to exist for any
concrete submodes down the line to call it).

> > But I do like your patch better, it seems pretty useful to introduce the
> > language concept, as it solves this and more problems more cleanly.  So
> > let's see where that goes.
>
> Great.
>
> >> This choice is coupled with the corresponding logic in 'buffer-language'
> >> (whether to keep the replace-regexp-in-string branch).
> >
> > Yes.  I think we should err on the side on convenience.  What exactly are
> > the defects can we get?  I can't see anything else but the tuareg-mode, and we
> > can plug that on our side.  Maybe you can see more.
>
> For example, it would sometimes return ugly non-existent languages like
> :help-fns--edit-value, :org-lint--report or :xref--xref-buffer.

What if we filter by prog-mode?  It would leave the ':ruby-base' and
':python-base' as false positives, I guess.  But then we could reasonably
say that anything ending with '-base' is abstract (or use the
aforementioned  explicit abstract prop).

It would also make ':lisp-data' a language.  But that's not bad.
lisp-data-mode is actually a useful concrete prog-mode derivative,
so I think it's OK to have ':lisp-data' as a language.

We can then have exceptions for some notable cases.  'lisp-mode' is
as we know, for Common Lisp only.

> In most cases that would be harmless, but OTOH the callers would miss
> out on the opportunity to see that the language is nil and apply their
> own fallbacks right away. I don't have a real problem scenario in mind,
> though.

Neither do I, but I agree we should be as accurate as possible.

> Perhaps some commands that would act on the language of the current
> buffer might want to say "no language is associated", but could not with
> the "convenience" approach.

For sure.

> >> Are there specific uses for get-mode-for-language when there is no
> >> existing buffer?
> >
> > Yes, I'd say this markdown-mode use is exactly that.  Markdown inserts
> > some text into a buffer and all it knows is the language it's supposed
> > to fontify it with.  The major mode has that logic, so it must invoke
> > the correct (and preferred) major-mode function.
>
> Sorry, I meant get-language-for-mode (which is the one implemented as
> buffer-language currently).
>
> > Another use is allowing the user to choose major modes for languages,
> > say from a tutorial or wizard or at Emacs startup.  Say, I prefer
> > ruby-ts-mode for Ruby, but c++-mode for C++.  It'd be helpful to summarize
> > those preferences.
>
> This would require capabilities like "get all modes for a language" (not
> one of the set of functions mentioned so far, and it'll need a full scan
> of major-mode-remap-alist) and "get current mode for a language" (this
> one matches markdown-mode's function you posted).

Yes.  I don't see the full scan of m-m-remap-alist as problematic
from a effiency perspective.  If we decide it's the database, it's
the database.   It's unfortunate that the "alist" implementation is
hardcoded in the name (which is why I would prefer a (:language "Foo")
kwarg to define-derived-mode) but we can abstract the alist aspect
away with accessors and do the usual "Do not change this variable
directly, use these accessors instead".

> BTW, get-current-mode-for-language could be implemented in terms of
> set-buffer-language.

What does get-current-mode-for-language do exactly?

> >> We could have both functions: buffer-language and get-language-for-mode
> >> ('get-mode-language'?). Or define one initially and add the other as needed.
> >
> > Yes.  buffer-language isn't bad, it's a useful helper.  But buffer-language
> > should be just
> >
> >     (with-current-buffer buffer (get-language-for-mode major-mode))
> >
> > Right?  Modulo some caching if it turns out to be very inneficient
> > (which I really doubt)
>
> Again: this won't work for files where no suitable major mode is
> available (e.g. not installed yet).

Right. So maybe

(or (with-current-buffer buffer (get-language-for-mode major-mode))
    (let (kw)
       (and buffer-file-name
            (keywordp (setq kw (alist-get buffer-file-name auto-mode-alist)))
            kw))
    (consult-oracles)
    (error "Don't know what language this is, sorry"))


?