This is a tricky commit. The key point is that the Fiber.complete
method, which commits a rendering to the DOM works like this: it
traverses the component tree, patch the corresponding DOM for each
component, calls the mounted/destroy hooks, and reset the currentfiber
of components to null, all synchronously.
However, this means that it is possible for components to initiate a
rendering (which create a new currentFiber) before the currentFiber is
reset to null, so the internal state of owl is corrupted. This can
occurs in a crash, as in the test that accompanies this commit.
To fix this, we take care of resetting the currentFiber first, while we
walk the component tree. Then, the internal state is always consistent
(i.e. a currentFiber to null means that there is no pending rendering)
closes#904
The initial problem solved by this commit is that it was possible to get
into a situation where a mounting/rendering was started, then the component was
updated, but then another mounting operation begins, and it tries to
reuse the previous rendering operation, which is no longer uptodate.
The underlying issue is that Owl did not track properly the various
internal state change of a component. These issue should be solved by
the introduction of the status enum, which currently tracks 6 possible
states:
- CREATED
- WILLSTARTED
- RENDERED
- MOUNTED
- UNMOUNTED
- DESTROYED
This status number replaces the isMounted and isDestroyed boolean flags.
It has the advantage of making sure that the component is in a
consistent state (it is no longer possible to be destroyed and mounted,
for example)
Another advantage is that it gives us an easy way to track the fact that
a component has been rendered, but is not in the DOM. This is a subtle
situation where some various events can happen, and we need to be able
to react to that case.
Note that there is a change of behaviour: if a component is mounted in a
specific target, then before the mounting is complete, the component is
mounted in another target, we no longer reject the first mounting
operation.
In some cases, a rendering initiated in some component is then remapped
into a larger rendering initiated by some parent.
If we have some components which implement shouldUpdate to return false,
then the following scenario can happen:
- some parent component is mounted (which triggers a rendering with
force: true => bypass the shouldUpdate)
- some sub component is updated and rerendered, AFTER the previous
rendering goes through it
- the sub component notices that there is an ongoing rendering, and
remaps itself in the parent rendering
Before this commit, the new fiber in the subcomponent does not have
force flag set to true, so the new rendering for the subcomponent does
not go through its own children (if they have shouldUpdate=false)
Another more complex kind of scenaria can happen when a remapped
rendering happen with sub components with dynamic shouldUpdate. The
problem is the same at the end: the new rendering should ignore the
shouldUpdate, to make sure we have the last correct information.
With this commit, we make sure that the flag of the new fiber is set to
true.
closes#818
Since ee956a197, the rendering is skipped if the currentFiber is
completed. Unfortunately, cancelled fibers remain set in __owl__,
so when a fiber is cancelled, subsequent calls to render are
skipped.
It would be nice to reset __owl__.currentFiber to null when the
fiber is cancelled, but when trying to do so, a lot of tests fail.
Part of issue #622Closes#665
Asynchronous rendering is subtle. Before this commit, it was possible
(though not easy) to get into a situation where the Owl rendering
pipeline decremented twice the counter of a fiber that indicates that
its work is complete.
The situation occurs in the `render` method, where we create a fiber,
wait for a micro tick before starting the actual rendering (the goal was
to batch all changes coming from a single call stack). However, in this
microtask tick, it was possible that the fiber was cancelled, and we did
not have a check for that.
In most cases, we just want Component<any, Env>. But since it was so
annoying to have always the type Env, we actually used
Component<any,any> everywhere.
With this commit, the generic types have a default (and their order is
swapped), so we can simply use Component in most cases, and
Component<Props> when we want to type the props.
Géry Debongnie