Re: Adding support for xref jumping to headers/interfaces

[Dmitry Gutov <dmitry@gutov.dev>]

On 27/11/2023 17:19, João Távora wrote:

>>> Alright, so since this is contentious (who else but you is pushing
>>> this idea?)
>>
>> Indeed, I wonder if nobody else is interested in having the additional
>> commands have pre-defined bindings, or having the same bindings across
>> languages.
> 
> Of course if you put it like that then EVERYBODY is interested in that.
> Everybody likes consistency, but this consistency you're proposing
> is a mirage.  An naive Xref user would be very proud to use
> "xref-find-declaration" in 3 different languages but then would
> start to doubt xref that the things he and the backend author meant
> by declaration are entirely different.

It's not like things are very different now: a backend might fail to 
implement "references", or the new "kinds" thingy, leaving the most 
recent addition not working anyway.

Or, again, different languages might have different notions for what a 
"definition" is. E.g. C vs. TeX.

> Again, leave this to major modes, which is what we always do (or to backend
> which in 95% of cases would live in the major mode when they are
> language-specific).

I wonder where you're getting this statistic from: advanced language 
analysis features don't usually live in major modes, not in the current 
state of the Emacs ecosystem.

>>> what about we start with 0 in xref.el.  We can always
>>> add to 0, no problem, but taking away  from some other number
>>> isn't so easy.
>>
>> We could indeed start with 0, but then I already see the same set of
>> extra commands supported in Eglot, Elisp, lsp-mode
> 
> They only exist in Eglot and LSP-mode because these are LSP things,
> it makes full sense they would exist in any Emacs/LSP interface.

Before LSP, they would similarly make sense for eclim, or omnisharp, or 
rtags/irony-mode, and etc. And that would be the case after LSP, unless 
the current crop of popular languages is simply washed away. We might 
have taken the names from LSP where they're nicely categorized, but they 
correspond to existing navigation features in IDEs that's been there for 
years.

> Elisp?  What commands are you talking about, only if you force them.
> I see no elisp-find-{declaration,type-definition,implementation}
> or anything of the sort.
> 
> The 7 Elisp cross-reference kinds you defined in your original
> patch are
> 
>    "defalias"
>    "face"
>    "function"
>    "constructor"
>    "generic"
>    "variable"
>    "feature'"

That's a good critique. But the list above is a different ontology than 
the declaration/implementation/type-definition "drawers" we were talking 
about. I wonder how much of it really language-specific.

E.g., a C++/Java/Ruby program also has symbols like 
modules/namespaces/classes/methods/functions... Variables, too. Some 
have global ones, some allow one local variables.

What makes the drawers called 
declaration/definition/implementation/type-definition different? I think 
that these ones, in many cases, apply several at once to the same 
symbol, and not just the same name, but the same entity in the code, 
such as for example a method -- one method can have a declaration and an 
implementation (which is these are the "definition" -- perhaps both -- 
is up to interpretation). Or a variable can have a definition (or 
declaration -- for variables they're usually the same) -- and a type 
definition. Point is, there is no way for the Xref backend to decide 
which of these you want to see when invoking "M-.", that's why there are 
separate actions for them in the protocol.

Whereas for Elisp's 
defalias/face/function/constructor/generic/variable/feature, 95% of the 
time the choice is easy to detect from the usage context. Same for 
modules/classes/methods for Ruby or C++. That's actually the reason why 
I felt my implementation for Elisp's "kinds" was useless -- it's simply 
slower and more awkward to work with than the existing 'M-.' (which has 
the automatic "infer namespace" functionality since 2021 courtesy of 
Mattias).

So we could also say that 
declaration/definition/implementation/type-definition are the names of 
"search actions" whereas namespace/class/face/function are "symbol 
namespaces". And to make interaction with "symbol namespaces" more 
useful, I'm guessing we would need something more - e.g. 
namespace-specific symbol completion. For xref-find-all-definitions (or 
whatever new name it has) to prompt for the namespace first and then 
fetch its specific identifier completion table (all functions or all 
faces, etc). And that, by the way, could also work in different 
languages, with very similar nomenclature (I think only "face" in the 
above list is really Elisp-specific).

I concede that the "search actions" such as 
find-declaration/implementation (and especially type-definition) are 
less useful in Elisp than in OO languages where inheritance (or at least 
traits/mixins) are used more widely. Maybe just for the cl-generic code, 
which is a minority. OTOH, perhaps they would be more relevant in CL?

>>>>    and to allow frictionless extensions for special capabilities.
>>>
>>> As to frictionless extension, the macro I proposed already
>>> xref-define-finder seems the easiest way by far.
>>
>> Sorry, I can't find it.
>>
>> But the definition of xref-find-declarations takes about 3 lines. There
>> is not much potential for making it even shorter.
> 
> It's very similar and analogous to:
> 
> (defmacro eglot--code-action (name kind)
>    "Define NAME to execute KIND code action."
>    `(defun ,name (beg &optional end)
>       ,(format "Execute `%s' code actions between BEG and END." kind)
>       (interactive (eglot--code-action-bounds))
>       (eglot-code-actions beg end ,kind t)))
> 
> And allows us to control exactly the 'interactive' spec of such
> commands to give us consistency among them.

I'm wary of using macros as framework/library interface in Elisp because 
unless they are very stable, they can cause versioning problems when 
upgrading or reverting to an earlier version. You change a macro, but a 
dependent package is already installed and byte-compiled, and there 
won't be a recompilation until its next update. We might not even intend 
to change them much, but then you don't always anticipate a bugfix.