The way keys were handled in QWeb was mostly ok, but a little naive. It
is fine when we deal with a list of dom nodes, since the reconciliation
algorithm need only to be able to differentiate/reconcile nodes in that
list, but it is an issue with components, which needs to globally be
able to find themselves in their parent's children map.
Because of the way this was handled, there were situations were Owl
internal virtual dom would crash, since components wrongly assumed that
they were already rendered in a different place.
To fix this, we generalize the way keys are generated, by concatenating
all sub keys coming from each iteration loops.
closes#584
Since t-call is now a function call, we need to properly handle the
internal key used to find previous components. If a t-call is inside a
t-foreach, then we need to transfer the key to the sub template.
Otherwise, each component in the subtemplate will be associated to the
same key, which means that it will lead to big issues: components are
destroyed and reused...
closes#581
Before this commit, when template was t-call'ed and defined
t-component directive within it, there was a crash because extra.parent was never defined
After this commit, this use case works
While it is not tested, nor documented, the reference QWeb
implementation display the `false` value as "false". So, we have to
adapt the Owl implementation to match that behaviour.
At the same time, this commit uses 'let' instead of 'var' in various places,
to make the compiled code more consistant.
Before this commit, the handlers were bound to the current context,
which is often the component but not necessarily. In some cases, a sub
scope can be created (with t-foreach, or slots, or t-call), and the
context is actually an object with the actual component instance it its
prototype chain, but not the component.
This commit is a significant refactoring of the internal of QWeb. It
simplifies the way variables/scoping and slots interact together. The
main idea is that we use a simple scope object instead of a
context/var/scope object.
This commit also implement the actual correct QWeb semantic for the
t-call directive with a sub body. Before, we simply extracted the
variables from the body and injected them at the top of the sub
template. We now simply compile the body before the sub template.
This is a joint work with Lucas (lpe).
closes#541closes#544closes#545closes#557closes#556
A component can now be mounted in different positions: either 'before' (at
the start), 'after' (at the end), or 'attach' (take possession of a target
element)
closes#539
This is a really interesting problem: there was a situation where a
component would not have an event handler properly bound. The reason was
that we were in a situation where the scheduler task queue was flushed
at an unfortunate timing:
- parent component template is rendered
- subcomponent is prepared:
- sub component is willStarted
- it is rendered (so, fiber counter is set to 0)
- scheduler task queue is flushed => we patch the DOM
- the code registered in the __prepare.then(handler) is executed, which
add the createHook to the vnode (but after the dom is patched)
Usually, the last two steps happen in a different order, but in an
environment with connected components, we sometimes flush the task queue
at various moments, so some code could be executed between the end of
__prepare and the registered handler.
The fix is interesting: we give a callback to the __prepare function
instead of registering a .then handler to the deferred. This callback
is then guaranteed to be called at exactly the proper timing.
Thanks seb for the hard work of finding the root cause of this issue
closes#520
Before this rev, using t-component on a div (for example) node
would silently ignore the div and replace it by the root node of
the component. A way to improve this was to create an extra div
node and to put the component inside it. However, it raised
questions: what do we do with other attributes set on this tag?
Do we apply them on the div, or on the component? In addition,
some of them only make sense on the component (e.g. props), so we
have to detect them. Whatever we would have decided, it wouldn't
have been obvious from the user point of view, so we chose not to
support it, and we thus now raise an error in this case.
Closes#487.
It is important to reset the isRendered flag to false to ensure
that other (subsequent) simultaneous calls to render will be
skipped, as the currentFiber is actually no yet (re-)rendered.
Closes#483
It could happen that a component would crash. But then, the __render
code tried to copy the class properties into the vnode, which does not
exist, creating a new error.
So, the solution is to process the classObj only in the successful part
of the render call, which then will not crash, so the normal error
handling will occur.
Following 2922cee6ea
Previous commit was not correct: used children with a cancelled
fiber (for instance, with a new currentFiber) could be destroyed.
This commit fixes the issue differently, by directly detecting
children that are not used anymore (and not mounted), and destroy
them directly (no need to wait for willStart promise to be resolved
anymore).
Owl takes care of not rendering a component that will be unmounted
immediately after, and it works. However, it also should clean
up the newly added fiber.
This rev. moves the logic of batching the notifications of changes
occurring in the same microtask tick, from the observer, to the
listeners (the context, and the render function of components).
By doing so in render, we ensure that similar renderings are
batched in a single one, wherever they come from (state change,
store change, direct call...).
Whenever a rendering is completed, we need to reset the currentFiber
to null to make sure all subsequent renderings will not be
confused.
However, the way it was done before this commit was wrong in a
specific situation: we resetted the currentFiber to null in the
patch method. The idea was that this method was called every time
the component completes a rendering. But this is not true: it
can happen that a component is unmounted and remounted without
changes. Then the component will be patched, but if there is no
change, it will not call recursively the patched methods of its
children.
So, we need to choose a better place to reset it to null.
closes#454
Re-render the component all the time before remounting it, which
seems safer anyway.
In the test, we had to add a nextTick to wait for the changes to
be notified (before, we waited thanks to the additional call to
mount). If we don't do the nextTick, mount is called, and render
is called right after (before the promise returned by mount is
resolved). This scenario doesn't work right now (see #441).
Closes#381
This is a partial revert of commit 7d249d6f09.
The reason is that this changes made it much harder to unit test
components. Before, we could simply instantiate a component
like this:
const my|Comp = new MyComponent(null, props);
and then simply test it. This commit reestablish that possibility.
For an unknown reason, it looks like the way ticks, micro task ticks or other subtle scheduling issues are different on a windows machine (it may be because of other difference, such as a specific version of nodejs).
Anyway, I could not find the cause of the issue, but
simply waiting an extra microtask tick seems to work.
We don't see any usecase for it, it makes the code more complex,
and there were still potential unresolved concurrency issues with
it.
Part of task #295
It now takes two arguments: parent (optional, only for non-root
components) and props (optional). In the case of the root component,
the env is taken from the config (config.defaultEnv). If it doesn't
exists, the default env is created on the fly.
Closes#306
Géry Debongnie