Hi Dmitry,
Thanks for looking over the patch. Here's the recipe for the purported
bug in xref.el:
1. Please apply my patch to tex-mode.el (and xref.el).
2. I've attached xref-bug.zip, which contains a directory with 4
identical LaTeX files and one LaTeX file with a single additional
character. Please extract it.
3. emacs -Q
4. C-x C-f xref-bug/mwea.ltx, and please don't visit the other 4
files.
5. Put point on \__hook_debug:n in line 6.
6. M-?, RTN, ... RTN, RTN.
The xref buffer should offer 5 hits, one from each file in the
directory.
7. Comment out the the line I added to xref--collect-matches,
byte-compile and load the file.
8. With point in the same place, M-?, RTN, ... RTN, RTN.
The xref buffer should offer 3 hits. The first is from the
file-visiting buffer (where I also set syntax-propertize--done to 0,
because in my testing there could be some issues here, too). The
second hit is from the first file opened in *xref-temp. Here,
syntax-propertize runs to line-end, and all is well. The next two
files are missed, because syntax-propertize--done is set to line-end
and they have exactly the same line length as file two, and therefore
syntax-propertize thinks that's good enough and doesn't actually
change anything. The fifth file has an additional character in line 6,
so syntax-propertize decides it needs to work on this line because
line-end > syntax-propertize--done.
You can put point on, say, \documentclass, and you'll get all 5 hits,
because this string doesn't begin or end with a non-word, non-symbol
character, and syntax-propertize doesn't need to run. You can make the
search string "\documentclass" and you'll get 2 hits, as line 1 has
the same length in all 5 files. (It's worth trying "\usepackage" as
the search string, too.)
That's my diagnosis anyway. Does it make sense?