bug#6256: 24.0.50; read-event in `repeat' command

[Drew Adams]

> > That tells me that it exited during `read-key', not
> > because the event/key tested did not match (`while' test = nil).
> 
> That's one possibility, indeed, tho maybe the code never gets to
> read-key at all.

This is what I used:
(unwind-protect
     (while (let ((evt (read-key)))
              (message "EVT: %S, R-R-CHAR: %S"
                       evt repeat-repeat-char)
              (eq (or (car-safe evt) evt)
                  (or (car-safe repeat-repeat-char)
                      repeat-repeat-char)))
       (repeat repeat-arg))
  (setq repeat-undo-count nil))

The only time that code is not entered is if `repeat-repeat-char' is nil (since
the code is inside the (when repeat-repeat-char...)).

And if that code is entered then the only time `message' is not invoked is if
`read-key' barfs.

> > Apparently `read-key' itself exits out to the `unwind-protect'
> > protection code in this scenario.
> 
> That's only if it signals an error, which you should then see 
> somewhere.

I saw no error messages in *Messages*.

Perhaps we can conclude that `repeat-repeat-char' was nil.  Since I have
`repeat-on-final-keystroke'=t (the default), that in turn means that
`last-command-event' was nil.  Does that help you see what went wrong?

`last-command-event' takes us into C code (and perhaps into your code for
`read-key'?).  I can't help much with that.  As I suggested, this might have to
do with other things such as my using a standalone minibuffer.  Dunno.

> > In emacs -Q, however, with just the Bookmark+ files loaded, I do not
> > see the problem.  That is presumably what you are seeing.
> 
> Indeed.  So it does work for you in Emacs-23, tho only for the case of
> "emacs -Q"?  If so, you may want to try and do the binary-search dance
> on your .emacs to see what's interfering.

I really don't want to try that for this, if I can avoid it.

> > Perhaps you can suggest something I can do to determine what is
> > happening that causes it (presumably) to exit the `while' during the
> > `read-key', jumping out to the `unwind-protect'.  Is there 
> > some debug message I can put at the beginning of the the
> > `unwind-protect' protection code to see what happened?  Can I
> > put some debug stuff into `read-key'?
> 
> You can start by adding various `message' calls around the 
> while, inside the while, around the read-key call, etc...
> Using edebug in this code is sadly problematic, so we're left with
> print-debugging.
> 
> > You still have not said anything about what's wrong with 
> > that solution.
> 
> It does not work for me (i.e. for X11 mouse wheel events).

Ah.  I see.  That's the first I heard of it.

I was thinking that that code would work because it remains abstract (in
principle).  It just decomposes the event into its components and compares
those.  There is (in principle) nothing platform-specific about it.

But I see from what you say below that the problem is that on X (and maybe on
some other platforms?) wheel events cannot be treated abstractly without some
massaging first.  They have a different form depending on whether they are the
first of a series.

That form difference might be useful for X (?), and maybe it could sometimes be
useful for Emacs (?), but it is just an obstacle in this context.  In general
(and in particular here) I don't think that Emacs has any need to distinguish
the first wheel event in a given direction from subsequent ones in the same
direction.

IOW, the problem is not that X uses different event names from Windows for its
wheel events.  The problem (here) is that X uses different event names for the
_same_ wheel action (rotation in a given direction).

> You could also try
>  (while (let ((evt  (read-event)))
>           (message "EVT: %S, R-R-CHAR: %S" evt repeat-repeat-char)
>           (and (equal (event-basic-type evt)
>                       (event-basic-type repeat-repeat-char))
>                (equal (event-modifiers evt)
>                       (event-modifiers repeat-repeat-char))))
>    (repeat repeat-arg))
> 
> > It's pretty simple.  It also seems logical, and it says just what we
> > want to be done: if the event's components are all the same as before
> > then repeat.  Dunno why you have a problem with this.  What problems
> > do you see with this approach?
> 
> That on X11, the second event is *not* the same as the first because
> mouse wheel send first a down and then an up event.

I see.  So it sounds like we need an abstraction to deal with that - I'd think
that would be useful anyway.  After all, there is no good reason to distinguish
the first wheel rotation (in a given direction).  At least there is no good
reason to _always_ do that, even if someone might find a reason why that might
be useful sometimes.

The modifiers that I wanted to compare in the `while' test are only the
`control', `meta', `shift', `double', etc. modifiers.  In the case of a wheel
event we could safely abstract from any `down' or `up' modifiers (here, at
least), AFAIK - they don't mean anything here.

What about defining a function that maps X wheel events to "Emacs" wheel events
that strip the `down' and `up' modifiers?  And then using that function here?

But that means being able to recognize a wheel event as such - any wheel event.
We don't want to strip `down' or `up' from any non-wheel mouse events.  That
recognition should be possible using the vars `mouse-wheel-(up|down)-event'.

Can you use something like this to make the decomposing-read-event approach work
for X?

(defun wheel-event (event)
  "Return EVENT, with modifiers `down' and `up' removed if a wheel event."
  (if (memq (event-basic-type event)
            (list mouse-wheel-up-event
                  mouse-wheel-down-event))
      event
    (event-convert-list
     (delq 'down (delq 'up (event-modifiers event))))))