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

> Probably, yes.  It would turn test-completion and try-completion into
> calls to all-completions and then cache one "arg+result" of
> all-completions (this pair would be sufficient to cover all calls to
> test/try/all-completion for any argument string which has `arg' as its
> prefix).

How does this patch look?

(The Octave part is 100% untested).