bug#49127: Performance degradation in encode_coding_object

[Eli Zaretskii]

> From: Mattias Engdegård <mattiase@acm.org>
> Date: Mon, 16 Aug 2021 19:32:46 +0200
> Cc: Eli Zaretskii <eliz@gnu.org>, 49127@debbugs.gnu.org
> 
> It looks very much like haskell-mode is producing way too many markers, and 
> that should be easily fixed (I wrote a tentative pull request). However, the 
> code in encode_coding_object appears broken. Look at the first loop in that 
> function:
> 
>   if (EQ (src_object, dst_object))
>     {
>       struct Lisp_Marker *tail;
> 
>       for (tail = BUF_MARKERS (current_buffer); tail; tail = tail->next)
>       {
>         tail->need_adjustment
>           = tail->charpos == (tail->insertion_type ? from : to);
>         need_marker_adjustment |= tail->need_adjustment;
>       }
>     }
> 
> I don't know how this could ever work. We loop through the markers in the 
> current buffer?

Yes.  Why do you think this loop is broken?

> It is called from term.c with src_object and dst_object both being Qnil; no 
> buffer is involved at all. At the very least the outer condition should have 
> an "&& BUFFERP (dst_object)" added, and the loop should run through the 
> markers of XBUFFER (dst_object) instead of current_buffer, no?
> 
> `decode_coding_object` seem to get these things right but I only took a 
> superficial look. It's probably copy-paste errors.

It isn't a copy/paste, I think it's a simple omission.

But if you look up-thread, you will see that I asked for xbacktrace
results from the OP, to make sure I completely understand the reasons
for the performance problem.  I never got any responses for that.  So
if you can produce the Lisp backtraces in that case, and later test a
fix, I'm quite sure I know how to fix this problem.  TIA.