bug#12619: completion-at-point and changing buffer

[Stefan Monnier <monnier@iro.umontreal.ca>]

> When using cycle completion (using `completion-cycle-threshold'), the
> completion cycle gets aborted if the buffer contents are modified. For
> an IRC buffer, this means that the completion cycle terminates when new
> text arrives. The same problem should be the case for other
> process-related buffers, like shells with regular output or similar.

Indeed, good point.

> First, in `completion-at-point', `completion-in-region-mode-predicate'
> is set to a function that compares the old start value with the new
> using `eq'. This prevents markers to work in this case.  Simply
> replacing `eq' with `=' means markers work, so the positions used by
> completion just move together with text changes.

Sounds OK, tho we should make sure that those values can't be nil
or some other non-numeric thingy.

> Second, `completion--cache-all-sorted-completions' adds a function to
> `after-change-functions' that cleans up the cache of sorted
> completions. I'm not entirely sure why it does so, but my patch adds
> that function only if not both of the returned positions are markers.

Hmm...but if the buffer modification happens right in the middle of the
completion text what should we do?  Should we really ignore
this modification?


        Stefan