Stefan Monnier <monnier@iro.umontreal.ca> writes:

>> Actually... I just realized this misses some cases, namely when we have
>> "star point" or "point star".
>
> FWIW, my local patches have included for years optimizations like the
> ones you suggested *and* they replaced `star point` and `point star`
> with just `star`.
>
> Do you think it's important to preserve `point` in those cases?

I think we can decide that question later.

>>> BTW, maybe we should go with something like
>>>
>>>        (`(,(and (pred symbolp) s1) ,(and (pred symbolp) s2) . ,rest)
>>>         (cond ((completion-pcm--<something>-p s1 s2) (setq p (cons s1 rest)))
>>>               ((completion-pcm--<something>-p s2 s1) (setq p (cons s2 rest)))
>>>               (t (push (pop p) n))))
>>>
>>> Where `completion-pcm--<something>-p` is some kind of partial ordering?
>>
>> Interesting thought.  Maybe it would make sense to have something like
>> - completion-pcm--wildcard-grows-on-left-p
>>   (non-nil for star, point, any, any-delim)
>> - completion-pcm--wildcard-grows-on-right-p
>>   (non-nil for star, point, prefix)
>> - completion-pcm--wildcard-in-text-p
>>   (non-nil for star and point)
>> Then this case would be something like "if grows-left/right-p is the
>> same, and at most one of the symbols is in-text-p, delete the
>> non-in-text-p one".
>
> But I guess your `completion-pcm--merge-completions` is still needed as
> long as we can't optimize all sequences of symbols to a single symbol.
> 🙁

Indeed.  So (my own tangents aside) I think this patch is a good fix for
this issue.  Updated it to fix a bug (fixed by using append instead of
nconc) and to add a few more tests including one covering that case.