Re: Undo stack for GTK+ (was: Re: undo in textview)

[Tristan Van Berkom <tvb gnome org>]

On Wed, Dec 30, 2009 at 12:48 AM, Paul Davis <paul linuxaudiosystems com> wrote:
> On Mon, Dec 28, 2009 at 12:02 PM, Holger Berndt <berndth gmx de> wrote:
>> Cross-posting to move the discussion to gtk-devel-list. Anybody interested
>> in the topic, please follow up there.
>>
>> On Do, 24.09.2009 19:23, A. Walton wrote:
>>
>>>It's definitely something many developers would love to see in Gtk+,
>>>but only a few have stepped up to the bat with patches and actually
>>>discussed the problem,
>>
>> Why don't we take the opportunity to discuss the problem now, then? I
>> can start by offering my view on how an undo stack should look like,
>> and provide a reference implementation as a basis of discussion.
>
> We've been using a C++ type-safe Undo/Redo stack in Ardour for many
> years. It was designed after looking at undo models in various other
> applications, and finding them all to have various problems for what
> we really needed (this included the GIMP). We've learned a number of
> things over the years.
>
> 1) you need to be able to group individual actions into transactions
> 2) serialization/deserialization is good, and hard
> 3) its very hard to use only a Memento pattern without running into
> efficiency problems
> 4) it can be quite hard to correctly implement object lifetime
> mamagement for objects involved in callbacks on the undo/redo stack
> 5) deciding whether or not to support nested transactions can be important

Hi Paul,
   Interesting discussion, I dont have loads of time to share either but
I did have a hell of a time making undo/redo work properly in Glade, have
to agree with all of the points above, cheers to point 4 and 5, have some
things to add...

Particularly point 5, nested transactions; I banged my head around that one
for a while and finally did not implement this; my interpretation of
nested transactions
was something like this:
  o User initiates action
  o action goes and modifies data on the model
  o model data change callback initiates "nested action"

  a.) it was complex as hell to implement without re-creating commands
       when the data model changes due to undo/redo or something other
than an action
      (and however nice the api, the complexity spills into the
individual action initiating code).
  b.) I found that if I simply add callbacks to the editors modifying
the data model instead of
       the model itself - I could achieve the same goals using command
groups (or "nested transactions").
       (i.e. connect to the editor before_commit/after_commit signal
and push/pop a command group
       around the editor's commit phase).

Another thing; one interesting feature not present in all undo/redo
implementations I've come across
(as a user of applications in general) that is present in Glade; is
the ->unifies()/->collapse() methods
of the GladeCommand; basically it allows you to make abundant changes
to the same data set
without making a huge unusable command stack (i.e. Glade uses this for
instance; to make sure
that 15 edits to a string property will only show up as one command).

Also finally, it would be nice to see this targeted somewhere below
GTK+ in the stack, would be
a shame to do all this work and not be able to use it from; say;
Clutter apps or such...

Cheers,
         -Tristan


> To browse the current implementation:
>
> http://subversion.ardour.org/svn/ardour2/branches/3.0/libs/pbd/undo.cc
> http://subversion.ardour.org/svn/ardour2/branches/3.0/libs/pbd/pbd/undo.h
>
> I am not suggesting this as an alternative to your proposal, its just
> a real life, battle-tested implementation that avoids pre-registration
> of actions that can be undone, and allows arbitrary functors to be
> used in the undo/redo stack (note that it uses boost::bind for this
> purpose).
>
> --p
> _______________________________________________
> gtk-list mailing list
> gtk-list gnome org
> http://mail.gnome.org/mailman/listinfo/gtk-list
>