abstract class RenderBox

A render object in a 2D cartesian coordinate system.

The size of each box is expressed as a width and a height. Each box has its own coordinate system in which its upper left corner is placed at (0, 0). The lower right corner of the box is therefore at (width, height). The box contains all the points including the upper left corner and extending to, but not including, the lower right corner.

Box layout is performed by passing a BoxConstraints object down the tree. The box constraints establish a min and max value for the child's width and height. In determining its size, the child must respect the constraints given to it by its parent.

This protocol is sufficient for expressing a number of common box layout data flows. For example, to implement a width-in-height-out data flow, call your child's layout function with a set of box constraints with a tight width value (and pass true for parentUsesSize). After the child determines its height, use the child's height to determine your size.

Writing a RenderBox subclass

One would implement a new RenderBox subclass to describe a new layout model, new paint model, new hit-testing model, or new semantics model, while remaining in the cartesian space defined by the RenderBox protocol.

To create a new protocol, consider subclassing RenderObject instead.

Constructors and properties of a new RenderBox subclass

The constructor will typically take a named argument for each property of the class. The value is then passed to a private field of the class and the constructor asserts its correctness (e.g. if it should not be null, it asserts it's not null).

Properties have the form of a getter/setter/field group like the following:

AxisDirection get axis => _axis;
AxisDirection _axis;
set axis(AxisDirection value) {
  assert(value != null); // same check as in the constructor
  if (value == _axis)
    return;
  _axis = value;
  markNeedsLayout();
}

The setter will typically finish with either a call to markNeedsLayout, if the layout uses this property, or markNeedsPaint, if only the painter function does. (No need to call both, markNeedsLayout implies markNeedsPaint.)

Consider layout and paint to be expensive; be conservative about calling markNeedsLayout or markNeedsPaint. They should only be called if the layout (or paint, respectively) has actually changed.

Children

If a render object is a leaf, that is, it cannot have any children, then ignore this section. (Examples of leaf render objects are RenderImage and RenderParagraph.)

For render objects with children, there are four possible scenarios:

• A single RenderBox child. In this scenario, consider inheriting from RenderProxyBox (if the render object sizes itself to match the child) or RenderShiftedBox (if the child will be smaller than the box and the box will align the child inside itself).

• A single child, but it isn't a RenderBox. Use the RenderObjectWithChildMixin mixin.

• A single list of children. Use the ContainerRenderObjectMixin mixin.

• A more complicated child model.

Using RenderProxyBox

By default, a RenderProxyBox render object sizes itself to fit its child, or to be as small as possible if there is no child; it passes all hit testing and painting on to the child, and intrinsic dimensions and baseline measurements similarly are proxied to the child.

A subclass of RenderProxyBox just needs to override the parts of the RenderBox protocol that matter. For example, RenderOpacity just overrides the paint method (and alwaysNeedsCompositing to reflect what the paint method does, and the visitChildrenForSemantics method so that the child is hidden from accessibility tools when it's invisible), and adds an RenderOpacity.opacity field.

RenderProxyBox assumes that the child is the size of the parent and positioned at 0,0. If this is not true, then use RenderShiftedBox instead.

See proxy_box.dart for examples of inheriting from RenderProxyBox.

Using RenderShiftedBox

By default, a RenderShiftedBox acts much like a RenderProxyBox but without assuming that the child is positioned at 0,0 (the actual position recorded in the child's parentData field is used), and without providing a default layout algorithm.

See shifted_box.dart for examples of inheriting from RenderShiftedBox.

Kinds of children and child-specific data

A RenderBox doesn't have to have RenderBox children. One can use another subclass of RenderObject for a RenderBox's children. See the discussion at RenderObject.

Children can have additional data owned by the parent but stored on the child using the parentData field. The class used for that data must inherit from ParentData. The setupParentData method is used to initialise the parentData field of a child when the child is attached.

By convention, RenderBox objects that have RenderBox children use the BoxParentData class, which has a BoxParentData.offset field to store the position of the child relative to the parent. (RenderProxyBox does not need this offset and therefore is an exception to this rule.)

Using RenderObjectWithChildMixin

If a render object has a single child but it isn't a RenderBox, then the RenderObjectWithChildMixin class, which is a mixin that will handle the boilerplate of managing a child, will be useful.

It's a generic class with one type argument, the type of the child. For example, if you are building a RenderFoo class which takes a single RenderBar child, you would use the mixin as follows:

class RenderFoo extends RenderBox
  with RenderObjectWithChildMixin<RenderBar> {
  // ...
}

Since the RenderFoo class itself is still a RenderBox in this case, you still have to implement the RenderBox layout algorithm, as well as features like intrinsics and baselines, painting, and hit testing.

Using ContainerRenderObjectMixin

If a render box can have multiple children, then the ContainerRenderObjectMixin mixin can be used to handle the boilerplate. It uses a linked list to model the children in a manner that is easy to mutate dynamically and that can be walked efficiently. Random access is not efficient in this model; if you need random access to the children consider the next section on more complicated child models.

The ContainerRenderObjectMixin class has two type arguments. The first is the type of the child objects. The second is the type for their parentData. The class used for parentData must itself have the ContainerParentDataMixin class mixed into it; this is where ContainerRenderObjectMixin stores the linked list. A ParentData class can extend ContainerBoxParentDataMixin; this is essentially BoxParentData mixed with ContainerParentDataMixin. For example, if a RenderFoo class wanted to have a linked list of RenderBox children, one might create a FooParentData class as follows:

class FooParentData extends ContainerBoxParentDataMixin<RenderBox> {
  // (any fields you might need for these children)
}

When using ContainerRenderObjectMixin in a RenderBox, consider mixing in RenderBoxContainerDefaultsMixin, which provides a collection of utility methods that implement common parts of the RenderBox protocol (such as painting the children).

The declaration of the RenderFoo class itself would thus look like this:

class RenderFlex extends RenderBox with
  ContainerRenderObjectMixin<RenderBox, FooParentData>,
  RenderBoxContainerDefaultsMixin<RenderBox, FooParentData> {
  // ...
}

When walking the children (e.g. during layout), the following pattern is commonly used (in this case assuming that the children are all RenderBox objects and that this render object uses FooParentData objects for its children's parentData fields):

RenderBox child = firstChild;
while (child != null) {
  final FooParentData childParentData = child.parentData;
  // ...operate on child and childParentData...
  assert(child.parentData == childParentData);
  child = childParentData.nextSibling;
}

More complicated child models

Render objects can have more complicated models, for example a map of children keyed on an enum, or a 2D grid of efficiently randomly-accessible children, or multiple lists of children, etc. If a render object has a model that can't be handled by the mixins above, it must implement the RenderObject child protocol, as follows:

• Any time a child is removed, call dropChild with the child.

• Any time a child is added, call adoptChild with the child.

• Implement the attach method such that it calls attach on each child.

• Implement the detach method such that it calls detach on each child.

• Implement the redepthChildren method such that it calls redepthChild on each child.

• Implement the visitChildren method such that it calls its argument for each child, typically in paint order (back-most to front-most).

• Implement debugDescribeChildren such that it outputs a string that describes all the children. In principle, for each child you want to include the results of that child's toStringDeep function.

Implementing these seven bullet points is essentially all that the two aforementioned mixins do.

Layout

RenderBox classes implement a layout algorithm. They have a set of constraints provided to them, and they size themselves based on those constraints and whatever other inputs they may have (for example, their children or properties).

When implementing a RenderBox subclass, one must make a choice. Does it size itself exclusively based on the constraints, or does it use any other information in sizing itself? An example of sizing purely based on the constraints would be growing to fit the parent.

Sizing purely based on the constraints allows the system to make some significant optimisations. Classes that use this approach should override sizedByParent to return true, and then override performResize to set the size using nothing but the constraints, e.g.:

@override
bool get sizedByParent => true;

@override
void performResize() {
  size = constraints.smallest;
}

Otherwise, the size is set in the performLayout function.

The performLayout function is where render boxes decide, if they are not sizedByParent, what size they should be, and also where they decide where their children should be.

Layout of RenderBox children

The performLayout function should call the layout function of each (box) child, passing it a BoxConstraints object describing the constraints within which the child can render. Passing tight constraints (see BoxConstraints.isTight) to the child will allow the rendering library to apply some optimisations, as it knows that if the constraints are tight, the child's dimensions cannot change even if the layout of the child itself changes.

If the performLayout function will use the child's size to affect other aspects of the layout, for example if the render box sizes itself around the child, or positions several children based on the size of those children, then it must specify the parentUsesSize argument to the child's layout function, setting it to true.

This flag turns off some optimisations; algorithms that do not rely on the children's sizes will be more efficient. (In particular, relying on the child's size means that if the child is marked dirty for layout, the parent will probably also be marked dirty for layout, unless the constraints given by the parent to the child were tight constraints.)

For RenderBox classes that do not inherit from RenderProxyBox, once they have laid out their children, should also position them, by setting the BoxParentData.offset field of each child's parentData object.

Layout of non-RenderBox children

The children of a RenderBox do not have to be RenderBoxes themselves. If they use another protocol (as discussed at RenderObject), then instead of BoxConstraints, the parent would pass in the appropriate Constraints subclass, and instead of reading the child's size, the parent would read whatever the output of layout is for that layout protocol. The parentUsesSize flag is still used to indicate whether the parent is going to read that output, and optimisations still kick in if the child has tight constraints (as defined by Constraints.isTight).

Painting

To describe how a render box paints, implement the paint method. It is given a PaintingContext object and an Offset. The painting context provides methods to affect the layer tree as well as a PaintingContext.canvas which can be used to add drawing commands. The canvas object should not be cached across calls to the PaintingContext's methods; every time a method on PaintingContext is called, there is a chance that the canvas will change identity. The offset specifies the position of the top left corner of the box in the coordinate system of the PaintingContext.canvas.

To draw text on a canvas, use a TextPainter.

To draw an image to a canvas, use the paintImage method.

A RenderBox that uses methods on PaintingContext that introduce new layers should override the alwaysNeedsCompositing getter and set it to true. If the object sometimes does and sometimes does not, it can have that getter return true in some cases and false in others. In that case, whenever the return value would change, call markNeedsCompositingBitsUpdate. (This is done automatically when a child is added or removed, so you don't have to call it explicitly if the alwaysNeedsCompositing getter only changes value based on the presence or absence of children.)

Anytime anything changes on the object that would cause the paint method to paint something different (but would not cause the layout to change), the object should call markNeedsPaint.

Painting children

The paint method's context argument has a PaintingContext.paintChild method, which should be called for each child that is to be painted. It should be given a reference to the child, and an Offset giving the position of the child relative to the parent.

If the paint method applies a transform to the painting context before painting children (or generally applies an additional offset beyond the offset it was itself given as an argument), then the applyPaintTransform method should also be overridden. That method must adjust the matrix that it is given in the same manner as it transformed the painting context and offset before painting the given child. This is used by the globalToLocal and localToGlobal methods.

Hit Tests

Hit testing for render boxes is implemented by the hitTest method. The default implementation of this method defers to hitTestSelf and hitTestChildren. When implementing hit testing, you can either override these latter two methods, or ignore them and just override hitTest.

The hitTest method itself is given a Point, and must return true if the object or one of its children has absorbed the hit (preventing objects below this one from being hit), or false if the hit can continue to other objects below this one.

For each child RenderBox, the hitTest method on the child should be called with the same HitTestResult argument and with the point transformed into the child's coordinate space (in the same manner that the applyPaintTransform method would). The default implementation defers to hitTestChildren to call the children. RenderBoxContainerDefaultsMixin provides a RenderBoxContainerDefaultsMixin.defaultHitTestChildren method that does this assuming that the children are axis-aligned, not transformed, and positioned according to the BoxParentData.offset field of the parentData; more elaborate boxes can override hitTestChildren accordingly.

If the object is hit, then it should also add itself to the HitTestResult object that is given as an argument to the hitTest method, using HitTestResult.add. The default implementation defers to hitTestSelf to determine if the box is hit. If the object adds itself before the children can add themselves, then it will be as if the object was above the children. If it adds itself after the children, then it will be as if it was below the children. Entries added to the HitTestResult object should use the BoxHitTestEntry class. The entries are subsequently walked by the system in the order they were added, and for each entry, the target's handleEvent method is called, passing in the HitTestEntry object.

Hit testing cannot rely on painting having happened.

Semantics

For a render box to be accessible, implement the describeApproximatePaintClip and visitChildrenForSemantics methods, and the semanticAnnotator getter. The default implementations are sufficient for objects that only affect layout, but nodes that represent interactive components or information (diagrams, text, images, etc) should provide more complete implementations. For more information, see the documentation for these members.

Intrinsics and Baselines

The layout, painting, hit testing, and semantics protocols are common to all render objects. RenderBox objects must implement two additional protocols: intrinsic sizing and baseline measurements.

There are four methods to implement for intrinsic sizing, to compute the minimum and maximum intrinsic width and height of the box. The documentation for these methods discusses the protocol in detail: computeMinIntrinsicWidth, computeMaxIntrinsicWidth, computeMinIntrinsicHeight, computeMaxIntrinsicHeight.

In addition, if the box has any children, it must implement computeDistanceToActualBaseline. RenderProxyBox provides a simple implementation that forwards to the child; RenderShiftedBox provides an implementation that offsets the child's baseline information by the position of the child relative to the parent. If you do not inherited from either of these classes, however, you must implement the algorithm yourself.