GNU bug report logs - #79200
31.0.50; Duplicated elements for '#<marker at' in buffer-undo-list

Previous Next

Full log

View this message in rfc822 format

From: Gerd Möllmann <gerd.moellmann <at> gmail.com>
To: Pip Cet <pipcet <at> protonmail.com>
Cc: oscarfv <at> eclipso.eu, 79200 <at> debbugs.gnu.org, monnier <at> iro.umontreal.ca
Subject: bug#79200: 31.0.50; Duplicated elements for '#<marker at' in buffer-undo-list
Date: Sat, 09 Aug 2025 13:15:18 +0200

Pip Cet <pipcet <at> protonmail.com> writes:

>>> The "obvious" fix would be to use a weak hash table to associate undo
>>> information with markers rather than keeping them in the undo
>>> list. However, that sounds quite hard to do, and it might make undo a
>>> lot more expensive.
>>
>> Hm, yes that sounds hard. Hm.
>>
>> Do you think we could, in compact, locally keep track of markers already
>> seen, and coalesce their entries somehow? My memory of the undo stuff is
>> pretty hazy, alas, and probably outdated, too.
>
> So far, I've seen no markers showing up twice in the same section of the
> undo list; that would be a bug in BUF_MARKERS, I think. So I'm not sure
> how coalescing would work.

Okay. The primitive-undo code of course only indicates that it can
happen, not how often.

> Over here, most of the markers in the undo list appear to be constructed
> in the format_mode_line_unwind_data function. So maybe something like
> this is needed:
>
> From d5cf17f88fbba4cdfc61ebea725e12a2c8b9aa42 Mon Sep 17 00:00:00 2001
> From: Pip Cet <pipcet <at> protonmail.com>
> Date: Sat, 9 Aug 2025 10:32:19 +0000
> Subject: [PATCH] Detach markers explicitly so they don't wind up on the undo
>  list
>
> * src/xdisp.c (unwind_format_mode_line): Detach markers when we're
> done with them.
> ---
>  src/xdisp.c | 1 +
>  1 file changed, 1 insertion(+)
>
> diff --git a/src/xdisp.c b/src/xdisp.c
> index 432dd5dceca..2406bb6f3d9 100644
> --- a/src/xdisp.c
> +++ b/src/xdisp.c
> @@ -13828,6 +13828,7 @@ unwind_format_mode_line (Lisp_Object vector)
>  
>  	      current_buffer = XBUFFER (buffer);
>  	      set_point_from_marker (AREF (vector, 11));
> +	      detach_marker (AREF (vector, 11));
>  	      ASET (vector, 11, Qnil);
>  	      current_buffer = cb;
>  	    }

I think that should be done. The 11 is always made with build_marker, so
it's guaranteed to be correct, and a win.

(And Gnus seems to not cite anything after an inline patch Also
interesting :-/).

> Famous last words: I think this is safe and AREF (vector, 11) is never
> leaked to Lisp. And I'm not sure whether the only reason it helps here
> might be a local peculiarity...

I think it should go in. I've also checked it's safe.
And I'll run with it in my Emacs.

> Of course, none of this is very satisfying: we can perform manual memory
> management for markers in C code, but it's very hard to do so from Lisp
> code.

Yes, the more general problem is that old GC is "eagerly weak" so to
speak, while MPS Is "lazily weak". I have no idea what to do about that. 

Previous Next

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