GNU bug report logs - #60655
30.0.50; tree-sitter: `treesit-transpose-sexps' is broken.

Previous Next

Full log

Message #11 received at 60655 <at> debbugs.gnu.org (full text, mbox):

From: Mickey Petersen <mickey <at> masteringemacs.org>
To: Theodor Thornhill <theo <at> thornhill.no>
Cc: 60655 <at> debbugs.gnu.org
Subject: Re: bug#60655: 30.0.50; tree-sitter: `treesit-transpose-sexps' is
 broken.
Date: Mon, 09 Jan 2023 08:48:04 +0000

Theodor Thornhill <theo <at> thornhill.no> writes:

> Mickey Petersen <mickey <at> masteringemacs.org> > The tree-sitter-enabled
> function, `treesit-transpose-sexps', that is called by
> transpose-sexps, is broken.
>>
>> It uses a naive method of sibling adjacency to determine
>> transpositions. But it is unfortunately not correct.
>>
>> Python:
>>
>>
>>   def -!-foo():
>>       pass
>>
>> Turns into this with `C-M-t':
>>
>>   def ()foo:
>>       pass
>>
>> But it ought to be:
>>
>>   foo def():
>>       pass
>>
>>
>> It's swapping two siblings that are indeed adjacent in the tree, but
>> not on screen, which is confusing and a regression from its previous
>> behaviour.
>>
>
> I can try to make transpose-sexps rely on only swapping "allowed"
> node-types?  That would be able to keep the new, better function, yet
> still disallow these syntax-breaking transposes.  What do you think?
>

This is a hard problem. I'm building the self-same in Combobulate, so
when I saw this implementation I saw a well-trodden path by myself.
There's a lot of subtlety to it, and it is not immediately possible to
accurately gauge the right things to swap with simple (or not so
simple) sibling transpositions.

Using a defined list is better, but with the caveat that it requires manual
intervention per mode. This is a really tricky thing to build well.



>> You could make a cogent argument that both approaches are wrong from a
>> syntactic perspective, but I think that misses the broader point that
>> `C-M-t' now does something errant and unexpected.
>
> I don't really see how "foo def():" is any better at all.  We gain some
> great improvements with this "naive" method - namely:
>
> if 5 + 5 == 10 then 10 else 100 + 100.  If point is on the else the 100
> + 100 wil be swapped by 10, but the old behavior will be broken.
>

The old behaviour was consistent. It had a simple *modus operandi*:
swap two things around point. As someone who has used `C-M-t' for
decades, I know what it'll do in pretty much all situations, because I
know what `C-M-k` and `C-M-f/b` do at all times.

Neither approach is great if you holistically approach this task as
"making it correct at all times", and it is easy to confect scenarios
that result in something that is semantically wrong, but syntactically
correct; something that is plain wrong, both semantically and
syntactically; and something that is occasionally correct.

'Like' siblings are an easy way out of this mess with the caveat, as
you'll see, but now you need to carefully pluck the right nodes from
the tree!

Consider the node type `pair' in a dict in Python. They are easily transposable for
that very reason, notwithstanding the anonymous "," node betwixt them.

That is why Combobulate has a list of stuff that it can safely
transpose, and for everything else it defaults to the "classic"
transpose.

>>
>> Worse, it's not possible to revert to the old behaviour (see
>> bug#60654)
>>
>>

Thanks for fixing that!

Kind regards,

Mickey.

>
> Right.
>
> Thanks,
> Theo

Previous Next

GNU bug tracking system
Copyright (C) 1999 Darren O. Benham, 1997,2003 nCipher Corporation Ltd, 1994-97 Ian Jackson.