Re: xref-query-replace-in-results error message after xref-find-definitions, was: Re: bug#58158: 29.0.50; [overlay] Interval tree iteration considered harmful

[Dmitry Gutov]

On 11.10.2022 19:01, Eli Zaretskii wrote:

> > > So maybe just saying
> > >
> > >     Cannot do global replacement using results of \\[xref-find-definitions]
> > >
> > > should be okay?
> >
> > Isn't it almost the same as I suggested upthread? Except I suggested
> > "this search" instead of naming the specific command.
>
> I'm not trying to argue, I'm trying to find the best message text.  It
> doesn't matter who suggested it first, what matters whether it is
> accurate and self-explanatory.

I wasn't playing "who said it first" either, but pointing at the 
previous suggestion as a little more inherently-accurate.

> > Do you think naming it will be helpful enough to sacrifice the 10%
> > accuracy of the message? I suppose someone might have indeed forgotten
> > that they did the search using xref-find-definitions.
>
> What are the 10% of inaccuracy?

Sorry, losing 100% accuracy, to get a slightly smaller percentage.

> If the message will only ever be
> shown when 'r' is invoked after M-., then what is inaccurate in it?

Almost only ever. With potential for exceptions, unless we try to 
enforce their absence.

> > > > That would mean that one 'r' can work in lsp-mode's
> > > > xref-find-definitions results (they define a bunch of custom commands
> > > > like lsp-find-definition and lsp-find-declaration, but that probably
> > > > doesn't matter). Not sure if we should do something about that.
> > >
> > > If 'r' happens to work in that case, we don't have to worry about the
> > > error message, right?
> >
> > That's correct, but having the command succeed might be a problem by
> > itself, couldn't it? It will rename the definitions (and/or
> > declarations), but not other occurrences.
>
> No error message could possibly prevent that from happening, could it?

Only if it happens somewhere else, or if the check before the errors is 
changed significantly.

> > If we go in from this direction, we can have
> > xref-show-definitions-buffer (the default
> > xref-show-definitions-function) ensure that the binding for 'r' is set
> > to some command that always reports an error (like 'cannot replace in
> > definitions'), or is unbound.
>
> But users can always invoke the command by name as well.

That's, uh, true. But having commands intended for one major mode 
silently failing (or doing that in a weird fashion) when invoked with 
'M-x' in another major mode has not usually been a concern for us.

> > This would do nothing for custom values of
> > xref-show-definitions-function, but should remove most of the confusion
> > with default configuration. And some non-default ones as well (lsp-mode
> > doesn't change the value of xref-show-definitions-function).
> >
> > If the docstring of xref-show-definitions-function looks okay to you, we
> > can use its vocabulary.
> >
> >     Cannot replace in definition search results
> >
> > should cover xref-find-definitions, lsp-find-definition and
> > lsp-find-declaration. Wouldn't help with lsp-find-implementation, though
> > (its results are also questionable WRT renaming because they don't
> > include all references either), but it won't make it worse.
>
> Hmm... maybe
>
>    Cannot perform global replacement in find-definition results
>
> Another idea would be for the error message to be constructed
> dynamically, and include the precise command that produced the Xref
> buffer, if xref.el could record that in some buffer-local variable.

I suppose it could. But then some backend (such as lsp or possibly 
eglot) might return 'definitions' results in format suitable for 
replacements, and that will mean that one can replace in 
xref-find-definitions's results, just with some backends but not others.

And with that, the error message

  Cannot do global replacement using results of \\[xref-find-definitions]

shown to the same user at a different time (perhaps when they're editing 
Elisp) will become a lie.

And then, okay, we could try to add the name of the backend symbol to 
the error message as well, but it's much harder to capture that one, 
especially given that not every command using Xref output will go 
through xref-backend-functions (project-find-regexp is a counter-example).