Properly release GL resource when unmapping a window and detect stale
cache indices using a generation count. This fixes the disappearing
images (GL textures would be invalid and need to be re-created) as well
as the panics when an item's cache index ended up being re-used because
some other item was drawn first and allocated that indices. The latter
is fixed by using a generational counter on the cache that's bumped when
clearing, instead of a single "cache_ok" bool.
The follow basic example reproduces the issue that the items remain around
but the item cache has been destroyed and re-created, which causes get_mut() to fail:
```
let app = App::new();
loop {
app.run(); // now close the window via the close button in the window decoration
}
```
Remove the pos parameter to the render functions and instead let
the item renderer apply the transformation on the rendere (femtovg
canvas or QPainter).
So `draw_*` functions in the backend now always operate in item local
coordinates.
There are two problems that this patch fixes:
* It may happen that the image is not loaded yet, which means
we need to wait with the colorization effect instead of colorizing
the dummy (1x1) texture.
* It may happen that we transition from a regular image to one
that has a valid colorization brush, in which case we need to
invalidate the graphics item cache.
This intends to provide a configurable rectangular "drop shadow". The
API is modeled after CSS/HTML5 Canvas where the element can be "bound"
to an existing rectangular shape (geometry and radius), the offset can
be used to place the shadow and color and blur configure the shadow.
The shadow's color fades into transparent.
TODO (in subsequent changes):
* Documentation
* Qt implementation
Use a save/restore state pattern for the entire renderer instead of passing around clip rectangles as data structure.
That's less data copying and maps naturally to existing mechanisms,
which in turn can use their own dirty handling to make
save_state/restore_state as cheap as possible.
This means going back to an immutable renderer reference and interior
mutability for the canvas and the gpu cache. This is because while
traversing the item tree for rendering we may end up destroying
other items due to the lazyness of the models.
... to a more general gpu_cache and store an enum. That way we can
encapsulate the resource releasing - in the future - entirely in the
backend through run-time polymorphism in the backend (gpu cache data
enum) instead of a vtable method in the item to destruct the correct
type of data.
Instead of determining the focus item through item tree traversal and
pointer comparison and storing the intermediate indices in the
components in the tree, remember the focus item by a pair of
VWeak<ComponentVTable, Dyn> and item_index: usize.
This speeds up determining the focus item as well as delivering events,
which can now be done directly after retrieving an ItemRef with
get_item_ref.
This also fixes the duplicate line edit focus in the 7gui cells
test case.
This is in preparation for allowing the run-time / items to clone VRc's
of the component.
ComponentVTable functions like visit_children_item contine to take a
ComponentRefPin as "self" parameter type, as a VRc would not be
supported by rust right now. That means the implementation then uses
self_weak to obtain a strong self-reference.
Destroying a component releases the item's graphics resources, which requires
a mutable reference to the rendering cache. That reference is acquired fine grained,
but when rendering items itself we also acquired a mutable reference
at the start of the rendering phase. That is too early and too long, and not necessary.
Borrow the rendering and cache individually with different mutability.
We might later, for destruction handling, need to borrow the rendering
cache mutably. That in turn *may* happen even during rendering itself,
as the models are updated lazily during visitor traversal and that may
result in item destruction.
Move fields out of the HighLevelRenderingPrimitive that are suitable to
represent as uniform variables. That reduces fields like Image or
Rectangle to their image or dimension, and only if *they* change, then
we re-do the work of building geometry and uploading it to the GPU.
Instead, the typically animated properties such as the position or the
color are passed separately with each draw call. This avoids re-decoding
PNG files and uploading them again when moving an image element around
on the screen.
