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

>> Stefan, a related question: do you think it would be reasonable to split
>> completion-table-with-quoting into two functions, separating the
>> substitute-in-file-name and comint-quote-filename cases?  It seems to me
>> that the comint-quote-filename case is way simpler than the
>> substitute-in-file-name case, and it would be a lot easier to understand
>> if the two cases were done separately, without the abstraction of
>> requote and unquote functions.  I can do that if that seems reasonable.
>> It would make fixing this issue a lot easier.
>
> It's worth a try.  The system we have is the best I could come up with
> back then but is hideous, so I'm interested in a simpler solution.

OK, here is a rework of just the substitute-in-file-name side.  This
fixes this bug (by making completion continue to work when point is at
/usr/|/), and fixes another FIXME in a test (partial completion now
preserves environment variables!).  And I don't think it loses anything
important, but I may be missing something.

(completion--sifn-requote can now be deleted and there are probably
other cleanups possible after this patch, but I didn't do them here to
keep the patch smaller.  Also, rfn-eshadow needs to be updated, but
that's just a visual issue.)