bug#35119: 26.1; narrow-to-region loses word-start/symbol-start information at end

[Sam Halliday]

Hmm, on further investigation I think this may just be regexp behaviour.

I came up with this as an alternative to `looking-back'

(defun my-looking-back (regexp lower)
  (let ((upper (point))
        (start lower))
    (save-excursion
      (catch 'hit
        (while (< start upper)
          (goto-char start)
          (re-search-forward regexp upper 't)
          (when (= (point) upper)
            (throw 'hit 't))
          (setq start (+ 1 start)))
        nil))))

and it also fails to match the : in the example. So perhaps limit is
also excluding the zero-length implied by the subsequent character.

On Wed, 3 Apr 2019 at 13:30, Sam Halliday <sam.halliday@gmail.com> wrote:
>
> Hi Eli,
>
> Sorry that was a terrible bug report.
>
> This impacts me in `looking-back'. Here's an interactive snippet to
> demonstrate the problem (not minimised to`narrow-to-region'):
>
> (defun look-for-35119 ()
>   (interactive)
>   (if (looking-back
>        (rx (: word-end ":" word-start))
>        ;;(rx (: word-end ":"))
>        (- (point) 1) 't)
>       (message "hit")
>     (message "miss")))
>
> in emacs-lisp-mode, which defines : as non-word, interactively
> evaluate look-for-35119 when the point is just after the colon in this
> example text
>
>   wibble:wobble
>
> I would expect to see "hit", but we get "miss". To demonstrate that
> the word-start is the cause of the problem, try the commented regexp
> and try again, you'll get "hit" but of course this regexp is not what
> is intended. For example, it would also match in between :: in the
> following:
>
>   wibble::wobble
>
> The cause is that the `narrow-to-region' call inside `looking-back' is
> dropping the word-start zero length match at the beginning of wobble.
> This may or may not be a bug in narrow-to-region, but I'm quite sure
> it's a bug in `looking-back'. There is most likely a similar example
> demonstrating that the zero lengths are missing at the start as well
> as the end.
>
> I've tried playing around with multiple alternative implementations of
> `looking-back' but none are working for me. Probably the best
> workaround I can think of is to extend the `narrow-to-region' call by
> one more character at the start and the end. Dealing with the start is
> easy, we just goto-char limit+1, but dealing with the end is difficult
> as we need to put an anychar \\. matcher in the doctored regexp and
> then the match-end is off-by-one from what the user expects, so then
> we have to doctor that, and then all hell breaks loose.
>
> Does that make sense?
>
>
> On Wed, 3 Apr 2019 at 12:25, Eli Zaretskii <eliz@gnu.org> wrote:
> >
> > > From: Sam Halliday <sam.halliday@gmail.com>
> > > Date: Wed, 03 Apr 2019 12:19:08 +0100
> > >
> > > If the function `narrow-to-region' (as it is in `looking-back') is used
> > > to restrict the region prior to an invocation of re-search-forward or
> > > looking-at, then zero length regexp patterns are lost at the boundaries.
> >
> > Could you please provide a recipe to reproduce the issue?  I'm not
> > sure I understand what is the problem you are describing.
> >
> > Thanks.