On Wed, Jan 22, 2014 at 9:17 AM, Toby Cubitt <
toby-dated-1391609828.ede38b@dr-qubit.org> wrote:

> On Tue, Jan 21, 2014 at 07:05:38PM -0500, Barry OReilly wrote:
> > > The "Unrecognised entry" error suggests the undo-tree-canary symbol
> > > has somehow ended up in a `buffer-undo-tree' entry.
> >
> > You mean "buffer-undo-list" not "buffer-undo-tree" right?
>
> No, I mean `buffer-undo-tree'. The canary should never end up there. In
> undo-tree-mode, primitive-undo only ever gets called on an entry copied
> from buffer-undo-tree. Hence my statement, above.
>
> > I checked Emacs 24.3 and as I suspected it's quite easy to make
> > undo-tree-canary appear in the buffer-undo-list.
>
> It's *supposed* to be in buffer-undo-list. It's not supposed to ever be
> in buffer-undo-tree. (And maybe it isn't, I'm just guessing from the
> error message here. I haven't had time to investigate yet.)
>
> > What changed is the error checking in core Emacs. If you expected that
> > undo-tree-canary would never be there between commands, that has not
> > been so for some time.
>
> I didn't expect that.
>
> > Could you tell me more about the purpose of undo-tree-canary?
>
> It lets undo-tree-mode detect when Emacs has discarded undo history from
> buffer-undo-list "behind undo-tree-mode's back". If the canary has
> vanished when undo-tree-mode looks at buffer-undo-list, Emacs must have
> purged some undo history.
>
> The new error checking in primitive-undo shouldn't affect undo-tree-mode
> in any way. I still strongly suspect this is a bug in the very delicate
> and relatively untested undo-in-region code, and the change to
> primitive-undo is a red herring.
>
>
In undo-tree-redo-1:

  (setf (undo-tree-node-undo current)
        (undo-list-pop-changeset 'discard-pos))

In undo-list-pop-changeset:

  (if (eq (car buffer-undo-list) 'undo-tree-canary)
      (push nil buffer-undo-list)
    [...])

The push call returns (nil 'undo-tree-canary). This is how it gets
into the buffer-undo-tree in my reproduction.

I'll close the Emacs bug since we're fairly sure at this point it's an
undo-tree bug.