Re: hideshow support for treesitter

[Yuan Fu <casouri@gmail.com>]

> On Sep 15, 2024, at 7:28 AM, pranshu sharma <pranshusharma366@gmail.com> wrote:
> 
> Yuan Fu <casouri@gmail.com> writes:
> 
>>> 
>>> On Fri, 6 Sept 2024 at 03:04, Juri Linkov <juri@linkov.net> wrote:
>>>>>>>> Are there any plans to add treesitter support for hideshow, it seems easy
>>>>>>>> enough to do and also very helpful, especially for langs that don't
>>>>>>>> excessively use parens, such as python
>>>>>>> 
>>>>>>> Indeed, it's easy to do this.  You just need to override
>>>>>>> hs-find-block-beginning-func, hs-find-next-block-func, etc.
>>>>>>> with treesit functions like 'treesit-outline-search'.
>>>>>> 
>>>>>> Ah I see, then I think treesit.el should contain code for this, because
>>>>>> this seems extremely useful task that would not make sense if each tree
>>>>>> sitter based major mode was to implement this separately
>>>>> 
>>>>> Or hideshow code for treesit could be in hideshow.el as well.
>>>>> The reason why 'treesit-outline-search' was added to treesit.el
>>>>> was because treesit.el already contained imenu support.  But
>>>>> imenu support could be in imenu.el as well.
>>>> 
>>>> Is there a preferred way to handle this treesit wise. I know some packages
>>>> (eg imenu) require major modes to define their support for it, but others
>>>> like ispell do it themselves. I personally thing the latter would be better
>>>> in this case.
>>> 
>>> I looked at hideshow.el and found such a list:
>>> 
>>> (defvar hs-special-modes-alist
>>>  (mapcar #'purecopy
>>>  '((c-mode "{" "}" "/[*/]" nil nil)
>>>    (c-ts-mode "{" "}" "/[*/]" nil nil)
>>>    (c++-mode "{" "}" "/[*/]" nil nil)
>>>    (c++-ts-mode "{" "}" "/[*/]" nil nil)
>>>    (bibtex-mode ("@\\S(*\\(\\s(\\)" 1))
>>>    (java-mode "{" "}" "/[*/]" nil nil)
>>>    (java-ts-mode "{" "}" "/[*/]" nil nil)
>>>    (js-mode "{" "}" "/[*/]" nil)
>>>    (js-ts-mode "{" "}" "/[*/]" nil)
>>>    (lua-ts-mode "{\\|\\[\\[" "}\\|\\]\\]" "--" nil)
>>>    (mhtml-mode "{\\|<[^/>]*?" "}\\|</[^/>]*[^/]>" "<!--" mhtml-forward nil)
>>> 
>>> But this is useless for tree-sitter.
>>> 
>>> Line-oriented treesit-outline-search relied on treesit-simple-imenu-settings
>>> that defines header lines for imenu.  And every ts-mode sets own value of
>>> treesit-simple-imenu-settings.  But this can't be used for hideshow.
>>> 
>>> So whoever will implement hideshow support for treesitter
>>> will need to design a setting like buffer-local
>>> treesit-simple-imenu-settings for ts-modes.
>> 
>>> On Sep 5, 2024, at 6:18 PM, Pranshu Sharma <pranshusharma198@gmail.com> wrote:
>>> 
>>> I am working to try get the hide show working for tree sitter, however one thing that is missing in treesit is sexp navigation functions, hideshow requires sexp function.
>> 
>> Master branch should have them, you can check out
>> treesit-forward-sexp. Major modes need to define sexp in
>> treesit-thing-settings for it to work though. Also, hideshow can just
>> use forward-sexp, it’ll use treesit-forward-sexp behind the scenes.
> 
> I played around with sexp navigation in haskell-ts-mode, and found that
> is is currently useless for ambiguous scenarios like:
> ---
> f x = 2_
> ---
> Where '_' represents the point, now. In this place if we do backward
> sexp, there are 2 valid places it would go, we can see this by if we
> put parens around the function:
> ---
> (f x = (2)_)_
> ---
> Now either the point could go to front of 2. or in front of f, but it
> will choose to go infront off 'f', which is kind of unpredictable.
> This is not a problem with treesit, rather treesitter itself.

Yeah, there has been discussion around sexp navigation with tree-sitter. I’ll go a step further and say the problem is us trying to retrofit the concept of sexp on non-lisp languages, where not everything is enclosed with parens—the definition of sexp outside of lisp is very fuzzy. As you shown here, where to move to is entirely subjective. Also, most languages doesn’t have delimiters like paren around syntactic constructs, making which sexp to move over ambiguous. 

My personal opinion is to make sexp navigation in non-lisp languages work like list navigation: statements, arguments in argument list, blocks, fields.

The conclusion from previous discussions, OTOH, is to define sexp as “the largest syntactic construct before/after point”. 

> Btw, are there been any plans on a treesit version  of backwards-up-list?

There are plans for generalizing all structural navigation commands and plug tree-inter into them. Right now most of the navigation functions (like backwards-up-list) are not generic. So the first step is to update them and allow other package/major mode to define custom behavior for them. 

There’s already some progress made here, like the addition of forward-sexp-function and transpose-sexp-function. Other navigation commands can follow the same pattern.

Yuan