React.DOM is becoming helper factories to generate ReactElements. They're not
classes. It will be ok to call them directly as functions, but not to use them
where a class is expected.
I grepped for uses of React.DOM. This is going to be a set of helper functions
to generate ReactElements without JSX.
They're not classes so they cannot be used where a React class is expected.
Since we always use JSX, we should have no internal use for these helpers.
We can use strings instead.
We're trying to move to a convention where lower-case variables on JSX are always HTML/SVG tags rather than variables in scope. See:
http://fb.me/react-jsx-lower-case
Therefore, this diff renames existing uses to upper case variables.
This throws an error on all these files, so I manually go rename all the variables. I then run lint to ensure that the rename doesn't leave any undeclared variables.
This makes ReactDOM a simple helper for creating ReactElements with the string tag as the type. The actual class is internal and created by instantiateReactComponent. Configurable using injection.
There's not a separate class for each tag. There's just a generic ReactDOMComponent which could take any tag name.
Invididual tags can be wrapped. When wrapping happens you can return the same tag again. If the wrapper returns the same string, then we fall back to the generic component. This avoids recursion in a single level wrapper.
IE9+ supports `window.getSelection()` but `SelectEventPlugin` gives precedence to `document.selection` instead. `document.selection` is still available in later versions of IE, but the modern Selection object is preferable. Swap the two.
We currently automatically render empty components in place of mocks. However,
we were accidentally transferring the props from the mocked descriptor to the
empty component placeholder. Even children.
This change just cleans that up and should only affect unit tests.
Trying to make the event a bit more performant for events.
Feel free to reject this because the API inevitably isn't great. It's good for perf though, and since we're only using `accumulate` in very restrained places, I think we're fine.
`accumulateInto` is `accumulate` that mutates more and allocates less. I kept `accumulate` in case we want that in the future.
The spec changed to allow null & undefined sources without throwing.
Now our code depends on this new behavior. The only browser that
implements Object.assign has also updated its native implementation to
this behavior (Firefox Nightly).
This fixes some log spew since this pattern is deprecated. This doesn't use
the validator in the descriptor creation so there's no prop type checks here.
I guess that's fine because we still have the second prop type checks.
We need to use ReactLegacyDescriptor because the constructor here will be a
legacy factory that we need to unwrap.
constructAndRenderComponent should be deprecated.
Replace plain function calls to legacy factories with createFactory or
createElement. For ReactDOMComponents the type should be replaced with
strings.
Because we don't have easy access to ReactLegacyDescriptor from within
React, we need to use the .type property to extract the real class.
This will go away later and is covered by unit tests.
This moves all logic around legacy descriptors to ReactLegacyDescriptor. This
is responsible for the layer that knows that createClass exports a legacy
factory. When passed one of these classes, it unwraps it to be a real class.
If it is passed a non legacy factory, it is assumed to be a non-react component
that needs to be invoked as a plain function.
The semantic change is that a descriptor is now always returned if passed a
legacy factory. Even if that factory is a mock. A mock would previously return
undefined.
For mocks, I treat the factory as the authoritative function. I call it to extract
the instance or fill it with an empty component placeholder.
Additionally, I make the classes take props as the first argument to the
constructor. This is what the new class system will do.
We currently need to set up some internals by calling the internal construct
method. Instead of doing that automatically in the constructor, I now move that
to a second pass so that mocks can get the plain props.
This means that we can assert that a mock has been called once it's mounted
with it's final props. Instead of the descriptor factory being called.
Depends on #1758.
Fixes#1698.
Previously, controlled components would update too soon when using something like ReactLayeredComponentMixin (i.e., before the layer's updates could propagate from the parent), causing the cursor to jump even when always updating the new model value to match the DOM state. With this change, we defer the update until after all nested updates have had a chance to finish, which prevents the cursor from misbehaving.
Also cleaned up the logic around updating a bit -- the .value and .checked updates in ReactDOMInput weren't being relied on at all so I removed them and opted for a simple forceUpdate instead. I also got rid of _isChanging which hasn't been necessary since the introduction of update batching.
Test Plan: Tested the example in http://jsfiddle.net/Bobris/ZZtXn/2/ and verified that the cursor didn't jump. Changed the code to filter out numbers and verified that the field prevents typing numbers (attempting to do so still causes the cursor to jump to the end). Also verified that controlled and uncontrolled radio buttons, textareas, and select boxes work.
Callbacks passed to this setImmediate function are called at the end of the current update cycle, which is guaranteed to be asynchronous but in the same event loop (with the default batching strategy).
This is useful for new-style refs (#1373, #1554) and for fixing #1698.
Test Plan: jest
This isn't a good final solution but it makes React actually usable on
its own.
This also makes tests runnable, though only via jest
(./node_modules/.bin/jest)
The `mockTagName` parameter was always optional, and so probably was not
used very often. If you tried to use it, it would be shadowed by the
`var mockTagName` declaration in the `render` method, so only
`module.mockTagName` or `"div"` were ever possible values.
Relax the argument type checks. Currently we throw for non-objects and terminals
but Object.assign does a coercion to Object instead. It also allows merging
Arrays as if they are objects.
This also relaxes the check for dependents such as ImmutableObject. This sucks
but it will allow us to use a fast code path to native Object.assign.
We always have the option of adding warnings to Object.assign or static type
checks.
I'm keeping the null check. Object.assign throws for null checks.
We'll also start returning the result of coercions just like Object.assign.
This detail is going to become more important once the idiom
`className={joinClasses(this.props.className, newClass)}` becomes more
common, as it will when we move away from `this.transferPropsTo`.
Breaking changes
- key/ref are no longer accessible on props but they are accessible on the
descriptors. This means that parents/owners can access it but not the
component itself.
- Descriptor factories are now plain functions and you can't rely on the
prototype or constructors of descriptors to identify the component type.
Existing descriptor factories are now wrapped in a legacy factory. Currently it
does nothing but it will give us a hook to track callers to factories that are
not using JSX but just invoking the function directly. It also proxies static
methods/properties to the underlying class. The newer factories don't have this
feature.
ReactTextComponent has it's own little factory because it's props is not an
object. This is a detail and will go away once ReactTextComponent no longer
needs descriptors.
If we don't error here, we end up with a confusing error later on in this.getDOMNode() where ReactMount doesn't have the proper container registered in its object.
See modification to the test-file: Basically we add a small hint at the end
of the error warning for propType errors to help identify which instantiation
of the component at hand is faulty.
We now forbid calling setState or forceUpdate if any component's render function is higher in the stack, not just the component you're trying to update. (We do this already now for React.renderComponent and React.unmountComponentAtNode.)
This also has the advantage that we can now remove CompositeLifeCycle.RECEIVING_STATE, making _compositeLifeCycleState easier to reason about (not to mention that RECEIVING_STATE was a confusing name) and making it possible to remove a try/finally from the render path which might help with perf.
Error should be thrown in the previous condition is "not" meet. And the href propType is described as optional in the comment, but the null check was missing.
This copies the propType and contextType validation to a wrapper around the
descriptor factory. By doing the validation early, we make it easier to track
down bugs. It also prepares for static type checking which should be done at the
usage site.
This validation is not yet active and is just logged using monitorCodeUse. This
will allow us to clean up callsites which would fail this new type of
validation.
I chose to copy the validation of abstracting it out since this is just an
intermediate step to avoid spamming consoles. This makes more a much cleaner
diff review/history. The original validation in the instance will be deleted as
soon as we can turn on the warnings.
Additionally, getDefaultProps are moved to become a static function which is
only executed once. It should be moved to statics but we don't have a
convenient way to merge mixins in statics right now. Deferring to ES6 classes.
This is still a breaking change since you can return an object or array from
getDefaultProps, which later gets mutated and now the shared instance is
mutated. Mutating an object that is passed into you from props is highly
discouraged and likely to lead to subtle bugs anyway. So I'm not too worried.
The defaultProps will later be resolved in the descriptor factory. This will
enable a perf optimizations where we don't create an unnecessary object
allocation when you use default props. It also means that ReactChildren.map
has access to resolved properties which gives them consistent behavior whether
or not the default prop is specified.
The patternUnits attribute defines how a pattern's coordinate system is
defined (x, y, width, height). This is particularly important when using
repeatable patterns in SVG. This commit adds this attribute to the lists
of known properties.
Closes#1548
binds static methods on the descriptor to the component's actual constructor, so that `foo.constructor.bar()` and `Foo.bar()` run with the same `this`.
This one was an actual behavioral bug rather than a bug with the tests; our intention was that the first element from the `defaultValue` array would remain selected but IE seemed to be choosing the last one instead. Now we set the value for uncontrolled components in componentDidUpdate when switching from multiple to non-multiple to ensure that a consistent option gets selected.
Test Plan: Ran the ReactDOMSelect tests in jest, phantomjs, IE10, Chrome, and Firefox. Also tested an uncontrolled select manually to make sure that nothing crazy happened when switching between options.
Previously this was failing because iframeDocument.body wasn't properly initialized. Creating the document this way seems to work in all environments
Test Plan: Ran the test in jest, phantomjs, IE10, Chrome, and Firefox.
Paul O’Shannessy
