Solved a bug in the bilinear interpolated brush transform algorithm that
sometimes caused pixel artifacts to appear on brush edges for brushes
rotated 90, -90, -180, 180 degrees.
Make sure the duplicated group layer actually has a properly set up
tile manager taken from its projection, and not just a dumb copy of
the original group's tiles. Also optimizes away useless calls to
gimp_group_layer_update_size().
(gimp_image_merge_visible_layers): merge the visible layers in the
active layer's group. We can't possibly merge across different groups
anyway because there is no logical place to add the merged layer.
Moreoever, this change makes the group behave more like a sub-image,
which is our metaphor anyway.
There is no reason to disallow this, the merged-down group layer will
simply disappear from the image just as a normal layer, and its
projection composited with the layer below.
(gimp_group_layer_duplicate): change the allowed type of the duplicate
from GIMP_TYPE_GROUP_LAYER to GIMP_TYPE_DRAWABLE. The former was
simply a braino when copying and modifying the GimpLayer code.
* app/core/gimpdrawable.[ch]: add "gboolean push_undo" to
GimpDrawable::convert_type().
* app/core/gimpdrawable-convert.[ch]: same here for the gray and rgb
conversion functions.
* app/core/gimpchannel.c
* app/core/gimplayer.c: pass FALSE when called from GimpItem::convert()
because it can be called on unattached items only.
* app/core/gimpimage-convert.c: pass TRUE.
* app/core/core-enums.[ch]
* app/core/gimpgrouplayerundo.[ch]
* app/core/gimpimage-undo-push.[ch]: add GIMP_UNDO_GROUP_LAYER_CONVERT
which simply calls gimp_drawable_convert_type() with the old type
when undone/redone.
* app/core/gimpgrouplayer.c: push a group layer convert undo so this
can be properly undone/redone.
Needed because a group layer's projection can have a different type
than its image, at least at some pathologic moment during image type
conversion. But even if it didn't, it's cleaner this way anyway.
* app/core/gimpprojectable.[ch]: add the new vfunc plus public API.
* app/core/gimpprojection.c: use it instead of using the type of the
image returned by gimp_projectable_get_image(). The get_image()
function is now unused in the projection and only needs to stay
because the projection is also a GimpPickable.
* app/core/gimpgrouplayer.c: implement GimpProjectable::get_image_type().
Move pixel conversion code from gimp_layer_convert() to the new
gimp_layer_convert_type() implementation and call
gimp_drawable_convert_type() from convert(). In convert_type(), simply
chain up to convert to gray and rgb and only implement indexed
conversion ourselves.
This may look like duplication of GimpItem::convert() but in fact will
fix the longstanding uglyness that GimpItem::convert() both transfers
an item to another image *and* converts the image type of drawables.
When this refactoring is done, GimpItem::convert() will only move an
item to another image, and its implementation in GimpDrawable classes
will call GimpDrawable::convert_type() to convert the pixels to
whatever format.
Takes a "dest_image" parameter anyway because for converting to
indexed we need the destination colormap. The default impl in
GimpDrawable can only convert to gray and rgb however.
Don't blindly fill the dest region's alpha channel even if it has
none. Fixes longstanding bug that made things without alpha dropped
to indexed images to arrive broken.
The new functions reall convert the drawable this time, using the
previously renamed convert_tiles functions. Remove tile manager
fiddling from all callers and leave it there only for converting to
indexed.
Rename them from gimp_drawable_convert_foo() to
gimp_drawable_convert_tiles_foo() because they don't convert the
drawable itself, they convert its tiles into passed-in tiles.
* app/core/gimpgrouplayer.[ch]: add gimp_group_layer_suspend_resize()
and gimp_group_layer_resume_resize() and call them around functions
where all a group's children are transformed (translated, resized
etc). This way we go from the worst case of reallocating the
group's projection tiles once for each child down to exactly one
reallocation.
* app/core/Makefile.am
* app/core/core-enums.[ch]
* app/core/core-types.h
* app/core/gimpimage-undo-push.[ch]
* app/core/gimpgrouplayerundo.[ch]: add new undo class
GimpGroupLayerUndo which implements undos for suspend/resume of
group layers and calls them in reverse order when undoing.
When the projectable's size or image type changes, stop the idle
renderer and remove all queued update area because everything needs to
be re-rendered anyway. Also honor the projectable's offset when
invalidating the entire projection after clearing the pyramid.
The optimization here which would change only the group layer's offset
if the extents of the union of its children has not changed is
completely bogus. That case can only happen if one of the chldren was
moved in a way that does not change the extents of all children's
union, but this doesn't mean that the childrens' positions relative to
each other have not changed.
Instead, invalidate the entire projection to at least avoid constant
reallocation of the tile pyramid.
Found by group layer testing hero Tobias Jakobs.
(will have to optimize real translations of the entire group (or of the
only child in the group) differently)
Having a function that only abstracts whether there is an active
layer or not is pretty useless. This also doesn't make the code in
selection_generate_segs() more complex but rather more obvious.
