A quick overview of the undo system
-----------------------------------

Actions on the image by the user are pushed onto an undo stack.  Each
action object includes all the information needed to undo or redo an
operation, plus an UndoType.  The type can be converted to text to
show to the user.  Actions may be run forwards (UndoState == REDO) or
backwards (UndoState == UNDO).  As the action is run, it swaps the
image's current state and the recorded state.  A run action is moved
from the undo stack to the redo stack (or vice-versa if UndoState ==
REDO).  Pushing something onto the undo stack causes the redo stack to
be cleared, since the actions on the redo stack may depend on the
image being in a particular state (eg consider: layer add, rename,
undo rename, layer delete.  If the redo stack weren't cleared on undo,
then there would still be a "rename" operation on the redo stack which
could be run on a non-existent layer.  Bad news.)

Undo groups
-----------
In order to group many basic operations together into a more useful
whole, code can push group start and end markers.  A group is treated
as a single action for the purposes of the undo and redo user
commands.  It is legal to nest groups, in which case the outermost
group is the only user-visible one.

Groups boundaries used to be implemented by pushing a NULL pointer on
the undo (or redo) stack.  Now they are a special action which has the
"group_boundary" bit set.  This allows the group boundaries to include
the undo type associated with the whole group.  The individual actions
need to preserve their own undo type since the undo_free_* functions
sometimes need to know which action is being freed.

Undo events
-----------
Images emit UNDO_EVENT signals, to say that the user has performed an
undo or redo action on that image.  This allows interested parties to
track image mutation actions.  So far, only the undo history dialog
uses this feature.  The other way to discover the undo status of an
image is to use the iterator functions undo_map_over_undo_stack() and
undo_map_over_redo_stack().  These call your function on each action
(or group) on the stack.  There is also undo_get_undo_name() and
undo_get_redo_name() to peek at the top items on each stack.  This
could be used (eg) to change the undo/redo menu strings to something
more meaningful, but currently lack synchronisation.

Dirtying images
---------------
NOTE about the gimage->dirty counter:
  If 0, then the image is clean (ie, copy on disk is the same as the one 
      in memory).
  If positive, then that's the number of dirtying operations done
      on the image since the last save.
  If negative, then user has hit undo and gone back in time prior
      to the saved copy.  Hitting redo will eventually come back to
      the saved copy.
  The image is dirty (ie, needs saving) if counter is non-zero.
  If the counter is around 10000, this is due to undo-ing back
  before a saved version, then mutating the image (thus destroying
  the redo stack).  Once this has happened, it's impossible to get
  the image back to the state on disk, since the redo info has been
  freed.  See undo.c for the gorey details.

NEVER CALL gimp_image_dirty() directly!

If your code has just dirtied the image, push an undo instead.
Failing that, push the trivial undo which tells the user the
command is not undoable: undo_push_cantundo() (But really, it would
be best to push a proper undo).  If you just dirty the image
without pushing an undo then the dirty count is increased, but
popping that many undo actions won't lead to a clean image.

Austin