Re: undo boundaries

[martin rudalics]

> (defun test ()
>   (interactive)
>   (insert "a")
>   (undo-boundary)
>   (insert "b"))
>
> If I do M-x test, I get:
> ab
>   ^
>
> Then, if I undo once in Emacs 22, I get:
> a
> ^
>
> What I was expecting (and what happens in Emacs 21):
> a
>  ^
>
>
> Also note:
>
> bojohan+news@dd.chalmers.se (Johan Bockgård) wrote:
>
>
>>The likely suspect is this change:
>>
>>undo.c
>>revision 1.55
>>date: 2002-04-04 22:42:56 +0200;  author: monnier;
>>(record_point): New fun.
>>(record_delete, record_insert): Use it.

Your example is a bit contrived.  If you do

a M-: (undo-boundary) RET b

subsequent undoing behaves as expected.  Hence, the culprit is probably
`execute-extended-command'.

FWIW, your problem is caused by the (insert "b").  Due to the preceding
(undo-boundary) record_point called by record_insert sets at_boundary to
1, the test last_point_position != pt succeeds, and last_point_position
(the position of `point' before `test' is executed) gets recorded in
undo_list.  However, that position (in your example the position before
the "a") is inherently wrong here since it gets recorded _after_ the
insertion of "a".  Maybe someone wants to debug this to verify my claim.

We could try to convince record_insert and record_delete to set another
variable, say last_point_position_is_valid, to nil, set that variable to
t at the time last_point_position is recorded in the command loop, and
inhibit point recording when last_point_position_is_valid is nil.

BTW, with Emacs 23 you can use

(defun test ()
  (interactive)
  (let ((undo-inhibit-record-point t))
    (insert "a")                      
    (undo-boundary)             
    (insert "b")))