Timing of input-method output (was: [Emacs-diffs] master 7b31de4: Add hook for all events)

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

>> Hmm... this is a bummer (in some input methods, many/most chars are
>> prefixes of something else).  IIUC this is not a new problem introduced
>> by the use of input-event-functions, tho, right?
> My experience is that completion doesn't work when an input method is
> half way through, although I only use simple input methods.

Right: so this is not a new problem.

> Anyway, I checked (should have done this before I sent the last email,
> sorry). The buffer is not changed by an input method half way through
> (or at least "before-change-function" is not called, which I am assuming
> is the same thing). So, when the buffer appears to say "pizza_[`]", if
> it does offer a completion at all, it's going to offer one based on
> "pizz" because they "a" is not in the buffer.

Actually, quail.el does insert the "a", but it does so only temporarily,
within a with-silent-modification, and it removes it as soon as we know
what will actually be the next result of the input method (to the user,
that's basically when the underscore is (re)moved).

IOW, the underlined text behaves more or less like it's not in the
buffer (technically it is, but it's transient and the *-change-functions
aren't notified).

For latin-based input methods (especially postfix-style ones) where it's
extremely frequent for the first char passed to an input method to end up
passed through unmodified, it would make sense to try and change the
behavior of quail so it really inserts the first char right away, even
if it may later modify it because of some subsequent user input.

E.g. currently, in latin-postfix methods, when I type

    Time0   Time1
    -----   -----
    e       n

the input looks (to the rest of Emacs) as if I had typed

    Time0   Time1
    -----   -----
            e n

this is so that when I type

    Time0   Time1
    -----   -----
    e       '

Emacs only sees

    Time0   Time1
    -----   -----
            é

but we it would make sense to try and make it behave such that an input like:

    Time0   Time1   Time2   Time3
    -----   -----   -----   -----
    e       n       e       '

results in the following (decoded) input events:

    Time0   Time1   Time2   Time3
    -----   -----   -----   -----
    e       n       e       DEL é

[ More or less: the DEL there can't be right since the user can bind it
  to anything else, so we'd have to delete the previous char "manually"
  rather than by emitting a DEL event.  ]

This would be a pretty significant change in quail.el (which is likely
to require experimentation before we can find something that works
well), and is likely to only make sense in some input methods.

I think it'd be nice for someone to play with such a quail variant
(it's probably easiest to do it be adding a config var to quail.el to
control that new behavior) to see if it can fly.


        Stefan