bug#53749: 29.0.50; [PATCH] Xref backend for TeX buffers

[Dmitry Gutov]

On 13/09/2023 20:01, David Fussner via Bug reports for GNU Emacs, the 
Swiss army knife of text editors wrote:

> > These won't be affected either way, right? Because project-find-regexp
> > defaults its input to (thing-at-point 'symbol t), and isearch...
> > probably also uses "symbol" if you ask it to.
> >
> > So... why not just make tex-thingatpt-include-escape a boolean? What
> > commands need to be distinguished that way? I think 'find-tag' (it's
> > obsolete but still used sometimes) would need to obey this var as well.
>
> xref-find-apropos and xref-find-references don't work well (or at all)
> with the escape char included in the search string, so I was keeping
> that char away from them. (The buffer-local variables I manipulate for
> project-find-regexp and isearch-forward-thing-at-point have to do with
> ensuring they use the texsymbol thing in the first place -- see
> tex--symbol-or-texsymbol.) Does that make sense?

Hmm, I suppose I skipped over that part of the patch too quickly.

Here's a potential problem with replacing the notion of "symbol": some 
other existing code (also working with TeX/LaTeX) might disagree, as it 
might have some existing notion of what a "symbol" in those modes is (as 
defined by the syntax table).

In general, we change the notion of a symbol by either changing the 
mode's syntax table, or by augmenting its effect using 
syntax-propertize-function (which, for example, could propertize the 
backslashes inside the buffer as "symbol constituent"). The latter might 
actually be a change that would affect how 'M-x xref-find-references' 
works (it will likely start to consider those \tags as symbol 
occurrences together with the backslash). But like other changes of what 
is considered to be a "symbol" in a major mode, it could conflict with 
existing code.

Anyway, I'm not saying you have to change the approach, but that's 
something to be aware of.

And to look at it from another direction: if the default implementation 
of xref-find-references (and etags uses the very generic one) doesn't 
work for you, perhaps it would be worth it to define a TeX-specific Xref 
backend. That would perhaps take 20-30 lines of code total, most of them 
delegating to the etags backend, or the default impl. But while 
delegating, you can modify the passed argument - e.g. if it included a 
backslash, you could forward it to the default impl for "find 
references" without a backslash. Or - alternatively - call 
(project-find-regexp "...") with a more complex regexp of your choice. 
The first alternative could look like this:

  (cl-defmethod xref-backend-references ((_backend (eql 'tex-etags)) 
identifier)
    (xref-backend-references 'etags (string-remove-prefix "\\" 
identifier)))

> I'll look at find-tag, too; thanks for pointing that out.

Doing the above choice on the level of Xref backend's methods 
would/should automatically make it work for all commands appropriately.

> > Why not set the variable find-tag-default-function instead? That seems
> > easier and more appropriate to do inside a major mode function.
>
> I settled on putting the symbol on the modes because I thought it was
> simpler than setting the variable buffer-locally in all the in-tree
> and AUCTeX modes, but I'll revisit this and see whether I can come up
> with something better.

Do AUCTeX modes inherit from tex-mode? Or all call 
tex-common-initialization? Then you could set that variable locally 
inside that function once.

All in all, it might not be wise to modify the behavior of third-party 
packages from inside Emacs this way (they might have other expectations, 
or there's going to appear a new major mode that needs the same 
treatment anyway).

Setting a variable to be used through mode inheritance or delegation is 
fine, but if that doesn't help, I would probably stop at defining a 
helper function or two and documenting how it should be used. And then 
maybe work with AUCTeX people to get the remaining necessary changes in 
from their side (or just leaving that up to the user, depending on how 
functional the default config ends up being).