No more signal handler is needed, therefore the code can also get rid of
tracking the treemodel. And we use a faster approach for iterating the
changed cellrenderers: We just iterate all columns instead of over all
cell accessibles, as that number is likely quite a bit smaller.
There was a corner case where the changed signal was not emitted.
If rows were built like this:
1 (not selected)
+ 2 (selected)
+ 3 (not selected)
And row 1 was removed, no signal would be emitted.
I like it when writing tests actually finds bugs that have been around
since 2003 - introduced by 4a03ea2334
actually. :)
We get certain cases, in particular with SELECTION_MULTIPLE, where we
cannot figure out in advance of real_set_cursor() if the selection will
actually change.
Previously, the cursor would just become invalid, which used to
reselect the first row in the treeview later on (without a
cursor-changed signal). This leads to a crash now with the recent
refactorings.
The patch is longer than I'd like it to be, but the situation is kinda
complicated, because we want to make sure to move the cursor to a good
row. It works like this:
1) From the deleted row, search forward for the first row that is not
going to be deleted and is not a separator.
2) If no such row exists, search backwards for a row that is not a
separator.
3) If no such node exists, clear the cursor.
Previously the code used a GtkTreeRowReference, which was (a) less
performant and more importantly (b) hiding errors.
The errors being hidden were the referenced row becoming invalid or
collapsed, because such rows would not be valid cursor rows and it would
be necesary to select a new row and emit a "cursor-changed" signal.
So if a crash bisects down to this commit, it is very likely that the
cursor has not correctly been updated and the cursor row is invalid.
Previously, when the cursor was a descendant of the collapsed row,
the cursor path was set to the collapsed row, but this was not
communicated via cursor-changed events.
_gtk_tree_view_find_path() was not a name that not really described what
the function does. And I kept forgetting it. Also, it took the tree view
as an argument and that was completely unnecessary.
Turned assertion into silent return.
This assertion is only hit when dragging from an empty tree view. In
this case, gtk_tree_view_begin_drag() is triggered from gtkdnd.c and not
from gtk_tree_view_maybe_begin_dragging_row(). We actually want to
cancel the drag at this point, but that is not possible with the GTK+
API as far as I can see.
The alternative is to not allowing the drag to start. This could be
done by simply unsetting the tree view as drag source when it is empty
and setting it as drag source again when rows are added. I didn't
choose to go with this for now, since this will likely break third party
code.
In gtk_tree_view_state_flags_changed() we were setting the background on
the bin_window without the necessary "view" style class, making the
treeview render with the wrong color in some circumstances (such as when
adding an empty treeview in a window, as spotted by Brian Smith).
This way, the a11y code knows if a column was reordered, added or
removed and can do the right things instead of trying to guess and
getting it wrong.
Also, this patch finalizes the changes so that only visible columns
exist to the accessibility interface.
.. as a replacement for _gtk_rbtree_node_find_parity(). Instead of 1 or
0, the function now returns the index of node in the complete tree
(counting from the root). And this is of course identical to the row
number.
Track the RBNode/RBTree instead of keeping a TreeRowReference. This is a
whole lot faster and less error-prone.
Also, notify the accessible of removal of rows before actually removing
them, so we have a chance to clean up.
The default (out) transfer mode is "full", but the passed "model" pointer gets
set to gtk_tree_view_get_model() which is "transfer none". This caused Python
programs to free the model after calling this, causing crashes.
