Inside Flutter

This document describes the inner workings of the Flutter toolkit that make Flutter's API possible. Because Flutter widgets are built using aggressive composition, user interfaces built with Flutter have a large number of widgets. To support this workload, Flutter uses sublinear algorithms for layout and building widgets as well as data structures that make tree surgery efficient and that have a number of constant-factor optimizations. With some additional details, this design also makes it easy for developers to create infinite scrolling lists using callbacks that build exactly those widgets that are visible to the user.

Aggressive composability

#

One of the most distinctive aspects of Flutter is its aggressive composability. Widgets are built by composing other widgets, which are themselves built out of progressively more basic widgets. For example, Padding is a widget rather than a property of other widgets. As a result, user interfaces built with Flutter consist of many, many widgets.

The widget building recursion bottoms out in RenderObjectWidgets, which are widgets that create nodes in the underlying render tree. The render tree is a data structure that stores the geometry of the user interface, which is computed during layout and used during painting and hit testing. Most Flutter developers do not author render objects directly but instead manipulate the render tree using widgets.

In order to support aggressive composability at the widget layer, Flutter uses a number of efficient algorithms and optimizations at both the widget and render tree layers, which are described in the following subsections.

Sublinear layout

#

With a large number of widgets and render objects, the key to good performance is efficient algorithms. Of paramount importance is the performance of layout, which is the algorithm that determines the geometry (for example, the size and position) of the render objects. Some other toolkits use layout algorithms that are O(N²) or worse (for example, fixed-point iteration in some constraint domain). Flutter aims for linear performance for initial layout, and sublinear layout performance in the common case of subsequently updating an existing layout. Typically, the amount of time spent in layout should scale more slowly than the number of render objects.

Flutter performs one layout per frame, and the layout algorithm works in a single pass. Constraints are passed down the tree by parent objects calling the layout method on each of their children. The children recursively perform their own layout and then return geometry up the tree by returning from their layout method. Importantly, once a render object has returned from its layout method, that render object will not be visited again[1] until the layout for the next frame. This approach combines what might otherwise be separate measure and layout passes into a single pass and, as a result, each render object is visited at most twice[2] during layout: once on the way down the tree, and once on the way up the tree.

Flutter has several specializations of this general protocol. The most common specialization is RenderBox, which operates in two-dimensional, cartesian coordinates. In box layout, the constraints are a min and max width and a min and max height. During layout, the child determines its geometry by choosing a size within these bounds. After the child returns from layout, the parent decides the child's position in the parent's coordinate system[3]. Note that the child's layout cannot depend on its position, as the position is not determined until after the child returns from the layout. As a result, the parent is free to reposition the child without needing to recompute its layout.

More generally, during layout, the only information that flows from parent to child are the constraints and the only information that flows from child to parent is the geometry. These invariants can reduce the amount of work required during layout:

  • If the child has not marked its own layout as dirty, the child can return immediately from layout, cutting off the walk, as long as the parent gives the child the same constraints as the child received during the previous layout.

  • Whenever a parent calls a child's layout method, the parent indicates whether it uses the size information returned from the child. If, as often happens, the parent does not use the size information, then the parent need not recompute its layout if the child selects a new size because the parent is guaranteed that the new size will conform to the existing constraints.

  • Tight constraints are those that can be satisfied by exactly one valid geometry. For example, if the min and max widths are equal to each other and the min and max heights are equal to each other, the only size that satisfies those constraints is one with that width and height. If the parent provides tight constraints, then the parent need not recompute its layout whenever the child recomputes its layout, even if the parent uses the child's size in its layout, because the child cannot change size without new constraints from its parent.

  • A render object can declare that it uses the constraints provided by the parent only to determine its geometry. Such a declaration informs the framework that the parent of that render object does not need to recompute its layout when the child recomputes its layout even if the constraints are not tight and even if the parent's layout depends on the child's size, because the child cannot change size without new constraints from its parent.

As a result of these optimizations, when the render object tree contains dirty nodes, only those nodes and a limited part of the subtree around them are visited during layout.

Sublinear widget building

#

Similar to the layout algorithm, Flutter's widget building algorithm is sublinear. After being built, the widgets are held by the element tree, which retains the logical structure of the user interface. The element tree is necessary because the widgets themselves are immutable, which means (among other things), they cannot remember their parent or child relationships with other widgets. The element tree also holds the state objects associated with stateful widgets.

In response to user input (or other stimuli), an element can become dirty, for example if the developer calls setState() on the associated state object. The framework keeps a list of dirty elements and jumps directly to them during the build phase, skipping over clean elements. During the build phase, information flows unidirectionally down the element tree, which means each element is visited at most once during the build phase. Once cleaned, an element cannot become dirty again because, by induction, all its ancestor elements are also clean[4].

Because widgets are immutable, if an element has not marked itself as dirty, the element can return immediately from build, cutting off the walk, if the parent rebuilds the element with an identical widget. Moreover, the element need only compare the object identity of the two widget references in order to establish that the new widget is the same as the old widget. Developers exploit this optimization to implement the reprojection pattern, in which a widget includes a prebuilt child widget stored as a member variable in its build.

During build, Flutter also avoids walking the parent chain using InheritedWidgets. If widgets commonly walked their parent chain, for example to determine the current theme color, the build phase would become O(N²) in the depth of the tree, which can be quite large due to aggressive composition. To avoid these parent walks, the framework pushes information down the element tree by maintaining a hash table of InheritedWidgets at each element. Typically, many elements will reference the same hash table, which changes only at elements that introduce a new InheritedWidget.

Linear reconciliation

#

Contrary to popular belief, Flutter does not employ a tree-diffing algorithm. Instead, the framework decides whether to reuse elements by examining the child list for each element independently using an O(N) algorithm. The child list reconciliation algorithm optimizes for the following cases:

  • The old child list is empty.
  • The two lists are identical.
  • There is an insertion or removal of one or more widgets in exactly one place in the list.
  • If each list contains a widget with the same key[5], the two widgets are matched.

The general approach is to match up the beginning and end of both child lists by comparing the runtime type and key of each widget, potentially finding a non-empty range in the middle of each list that contains all the unmatched children. The framework then places the children in the range in the old child list into a hash table based on their keys. Next, the framework walks the range in the new child list and queries the hash table by key for matches. Unmatched children are discarded and rebuilt from scratch whereas matched children are rebuilt with their new widgets.

Tree surgery

#

Reusing elements is important for performance because elements own two critical pieces of data: the state for stateful widgets and the underlying render objects. When the framework is able to reuse an element, the state for that logical part of the user interface is preserved and the layout information computed previously can be reused, often avoiding entire subtree walks. In fact, reusing elements is so valuable that Flutter supports non-local tree mutations that preserve state and layout information.

Developers can perform a non-local tree mutation by associating a GlobalKey with one of their widgets. Each global key is unique throughout the entire application and is registered with a thread-specific hash table. During the build phase, the developer can move a widget with a global key to an arbitrary location in the element tree. Rather than building a fresh element at that location, the framework will check the hash table and reparent the existing element from its previous location to its new location, preserving the entire subtree.

The render objects in the reparented subtree are able to preserve their layout information because the layout constraints are the only information that flows from parent to child in the render tree. The new parent is marked dirty for layout because its child list has changed, but if the new parent passes the child the same layout constraints the child received from its old parent, the child can return immediately from layout, cutting off the walk.

Global keys and non-local tree mutations are used extensively by developers to achieve effects such as hero transitions and navigation.

Constant-factor optimizations

#

In addition to these algorithmic optimizations, achieving aggressive composability also relies on several important constant-factor optimizations. These optimizations are most important at the leaves of the major algorithms discussed above.

  • Child-model agnostic. Unlike most toolkits, which use child lists, Flutter's render tree does not commit to a specific child model. For example, the RenderBox class has an abstract visitChildren() method rather than a concrete firstChild and nextSibling interface. Many subclasses support only a single child, held directly as a member variable, rather than a list of children. For example, RenderPadding supports only a single child and, as a result, has a simpler layout method that takes less time to execute.

  • Visual render tree, logical widget tree. In Flutter, the render tree operates in a device-independent, visual coordinate system, which means smaller values in the x coordinate are always towards the left, even if the current reading direction is right-to-left. The widget tree typically operates in logical coordinates, meaning with start and end values whose visual interpretation depends on the reading direction. The transformation from logical to visual coordinates is done in the handoff between the widget tree and the render tree. This approach is more efficient because layout and painting calculations in the render tree happen more often than the widget-to-render tree handoff and can avoid repeated coordinate conversions.

  • Text handled by a specialized render object. The vast majority of render objects are ignorant of the complexities of text. Instead, text is handled by a specialized render object, RenderParagraph, which is a leaf in the render tree. Rather than subclassing a text-aware render object, developers incorporate text into their user interface using composition. This pattern means RenderParagraph can avoid recomputing its text layout as long as its parent supplies the same layout constraints, which is common, even during tree surgery.

  • Observable objects. Flutter uses both the model-observation and the reactive paradigms. Obviously, the reactive paradigm is dominant, but Flutter uses observable model objects for some leaf data structures. For example, Animations notify an observer list when their value changes. Flutter hands off these observable objects from the widget tree to the render tree, which observes them directly and invalidates only the appropriate stage of the pipeline when they change. For example, a change to an Animation<Color> might trigger only the paint phase rather than both the build and paint phases.

Taken together and summed over the large trees created by aggressive composition, these optimizations have a substantial effect on performance.

Separation of the Element and RenderObject trees

#

The RenderObject and Element (Widget) trees in Flutter are isomorphic (strictly speaking, the RenderObject tree is a subset of the Element tree). An obvious simplification would be to combine these trees into one tree. However, in practice there are a number of benefits to having these trees be separate:

  • Performance. When the layout changes, only the relevant parts of the layout tree need to be walked. Due to composition, the element tree frequently has many additional nodes that would have to be skipped.

  • Clarity. The clearer separation of concerns allows the widget protocol and the render object protocol to each be specialized to their specific needs, simplifying the API surface and thus lowering the risk of bugs and the testing burden.

  • Type safety. The render object tree can be more type safe since it can guarantee at runtime that children will be of the appropriate type (each coordinate system, e.g. has its own type of render object). Composition widgets can be agnostic about the coordinate system used during layout (for example, the same widget exposing a part of the app model could be used in both a box layout and a sliver layout), and thus in the element tree, verifying the type of render objects would require a tree walk.

Infinite scrolling

#

Infinite scrolling lists are notoriously difficult for toolkits. Flutter supports infinite scrolling lists with a simple interface based on the builder pattern, in which a ListView uses a callback to build widgets on demand as they become visible to the user during scrolling. Supporting this feature requires viewport-aware layout and building widgets on demand.

Viewport-aware layout

#

Like most things in Flutter, scrollable widgets are built using composition. The outside of a scrollable widget is a Viewport, which is a box that is "bigger on the inside," meaning its children can extend beyond the bounds of the viewport and can be scrolled into view. However, rather than having RenderBox children, a viewport has RenderSliver children, known as slivers, which have a viewport-aware layout protocol.

The sliver layout protocol matches the structure of the box layout protocol in that parents pass constraints down to their children and receive geometry in return. However, the constraint and geometry data differs between the two protocols. In the sliver protocol, children are given information about the viewport, including the amount of visible space remaining. The geometry data they return enables a variety of scroll-linked effects, including collapsible headers and parallax.

Different slivers fill the space available in the viewport in different ways. For example, a sliver that produces a linear list of children lays out each child in order until the sliver either runs out of children or runs out of space. Similarly, a sliver that produces a two-dimensional grid of children fills only the portion of its grid that is visible. Because they are aware of how much space is visible, slivers can produce a finite number of children even if they have the potential to produce an unbounded number of children.

Slivers can be composed to create bespoke scrollable layouts and effects. For example, a single viewport can have a collapsible header followed by a linear list and then a grid. All three slivers will cooperate through the sliver layout protocol to produce only those children that are actually visible through the viewport, regardless of whether those children belong to the header, the list, or the grid[6].

Building widgets on demand

#

If Flutter had a strict build-then-layout-then-paint pipeline, the foregoing would be insufficient to implement an infinite scrolling list because the information about how much space is visible through the viewport is available only during the layout phase. Without additional machinery, the layout phase is too late to build the widgets necessary to fill the space. Flutter solves this problem by interleaving the build and layout phases of the pipeline. At any point in the layout phase, the framework can start building new widgets on demand as long as those widgets are descendants of the render object currently performing layout.

Interleaving build and layout is possible only because of the strict controls on information propagation in the build and layout algorithms. Specifically, during the build phase, information can propagate only down the tree. When a render object is performing layout, the layout walk has not visited the subtree below that render object, which means writes generated by building in that subtree cannot invalidate any information that has entered the layout calculation thus far. Similarly, once layout has returned from a render object, that render object will never be visited again during this layout, which means any writes generated by subsequent layout calculations cannot invalidate the information used to build the render object's subtree.

Additionally, linear reconciliation and tree surgery are essential for efficiently updating elements during scrolling and for modifying the render tree when elements are scrolled into and out of view at the edge of the viewport.

API Ergonomics

#

Being fast only matters if the framework can actually be used effectively. To guide Flutter's API design towards greater usability, Flutter has been repeatedly tested in extensive UX studies with developers. These studies sometimes confirmed pre-existing design decisions, sometimes helped guide the prioritization of features, and sometimes changed the direction of the API design. For instance, Flutter's APIs are heavily documented; UX studies confirmed the value of such documentation, but also highlighted the need specifically for sample code and illustrative diagrams.

This section discusses some of the decisions made in Flutter's API design in aid of usability.

Specializing APIs to match the developer's mindset

#

The base class for nodes in Flutter's Widget, Element, and RenderObject trees does not define a child model. This allows each node to be specialized for the child model that is applicable to that node.

Most Widget objects have a single child Widget, and therefore only expose a single child parameter. Some widgets support an arbitrary number of children, and expose a children parameter that takes a list. Some widgets don't have any children at all and reserve no memory, and have no parameters for them. Similarly, RenderObjects expose APIs specific to their child model. RenderImage is a leaf node, and has no concept of children. RenderPadding takes a single child, so it has storage for a single pointer to a single child. RenderFlex takes an arbitrary number of children and manages it as a linked list.

In some rare cases, more complicated child models are used. The RenderTable render object's constructor takes an array of arrays of children, the class exposes getters and setters that control the number of rows and columns, and there are specific methods to replace individual children by x,y coordinate, to add a row, to provide a new array of arrays of children, and to replace the entire child list with a single array and a column count. In the implementation, the object does not use a linked list like most render objects but instead uses an indexable array.

The Chip widgets and InputDecoration objects have fields that match the slots that exist on the relevant controls. Where a one-size-fits-all child model would force semantics to be layered on top of a list of children, for example, defining the first child to be the prefix value and the second to be the suffix, the dedicated child model allows for dedicated named properties to be used instead.

This flexibility allows each node in these trees to be manipulated in the way most idiomatic for its role. It's rare to want to insert a cell in a table, causing all the other cells to wrap around; similarly, it's rare to want to remove a child from a flex row by index instead of by reference.

The RenderParagraph object is the most extreme case: it has a child of an entirely different type, TextSpan. At the RenderParagraph boundary, the RenderObject tree transitions into being a TextSpan tree.

The overall approach of specializing APIs to meet the developer's expectations is applied to more than just child models.

Some rather trivial widgets exist specifically so that developers will find them when looking for a solution to a problem. Adding a space to a row or column is easily done once one knows how, using the Expanded widget and a zero-sized SizedBox child, but discovering that pattern is unnecessary because searching for space uncovers the Spacer widget, which uses Expanded and SizedBox directly to achieve the effect.

Similarly, hiding a widget subtree is easily done by not including the widget subtree in the build at all. However, developers typically expect there to be a widget to do this, and so the Visibility widget exists to wrap this pattern in a trivial reusable widget.

Explicit arguments

#

UI frameworks tend to have many properties, such that a developer is rarely able to remember the semantic meaning of each constructor argument of each class. As Flutter uses the reactive paradigm, it is common for build methods in Flutter to have many calls to constructors. By leveraging Dart's support for named arguments, Flutter's API is able to keep such build methods clear and understandable.

This pattern is extended to any method with multiple arguments, and in particular is extended to any boolean argument, so that isolated true or false literals in method calls are always self-documenting. Furthermore, to avoid confusion commonly caused by double negatives in APIs, boolean arguments and properties are always named in the positive form (for example, enabled: true rather than disabled: false).

Paving over pitfalls

#

A technique used in a number of places in the Flutter framework is to define the API such that error conditions don't exist. This removes entire classes of errors from consideration.

For example, interpolation functions allow one or both ends of the interpolation to be null, instead of defining that as an error case: interpolating between two null values is always null, and interpolating from a null value or to a null value is the equivalent of interpolating to the zero analog for the given type. This means that developers who accidentally pass null to an interpolation function will not hit an error case, but will instead get a reasonable result.

A more subtle example is in the Flex layout algorithm. The concept of this layout is that the space given to the flex render object is divided among its children, so the size of the flex should be the entirety of the available space. In the original design, providing infinite space would fail: it would imply that the flex should be infinitely sized, a useless layout configuration. Instead, the API was adjusted so that when infinite space is allocated to the flex render object, the render object sizes itself to fit the desired size of the children, reducing the possible number of error cases.

The approach is also used to avoid having constructors that allow inconsistent data to be created. For instance, the PointerDownEvent constructor does not allow the down property of PointerEvent to be set to false (a situation that would be self-contradictory); instead, the constructor does not have a parameter for the down field and always sets it to true.

In general, the approach is to define valid interpretations for all values in the input domain. The simplest example is the Color constructor. Instead of taking four integers, one for red, one for green, one for blue, and one for alpha, each of which could be out of range, the default constructor takes a single integer value, and defines the meaning of each bit (for example, the bottom eight bits define the red component), so that any input value is a valid color value.

A more elaborate example is the paintImage() function. This function takes eleven arguments, some with quite wide input domains, but they have been carefully designed to be mostly orthogonal to each other, such that there are very few invalid combinations.

Reporting error cases aggressively

#

Not all error conditions can be designed out. For those that remain, in debug builds, Flutter generally attempts to catch the errors very early and immediately reports them. Asserts are widely used. Constructor arguments are sanity checked in detail. Lifecycles are monitored and when inconsistencies are detected they immediately cause an exception to be thrown.

In some cases, this is taken to extremes: for example, when running unit tests, regardless of what else the test is doing, every RenderBox subclass that is laid out aggressively inspects whether its intrinsic sizing methods fulfill the intrinsic sizing contract. This helps catch errors in APIs that might otherwise not be exercised.

When exceptions are thrown, they include as much information as is available. Some of Flutter's error messages proactively probe the associated stack trace to determine the most likely location of the actual bug. Others walk the relevant trees to determine the source of bad data. The most common errors include detailed instructions including in some cases sample code for avoiding the error, or links to further documentation.

Reactive paradigm

#

Mutable tree-based APIs suffer from a dichotomous access pattern: creating the tree's original state typically uses a very different set of operations than subsequent updates. Flutter's rendering layer uses this paradigm, as it is an effective way to maintain a persistent tree, which is key for efficient layout and painting. However, it means that direct interaction with the rendering layer is awkward at best and bug-prone at worst.

Flutter's widget layer introduces a composition mechanism using the reactive paradigm[7] to manipulate the underlying rendering tree. This API abstracts out the tree manipulation by combining the tree creation and tree mutation steps into a single tree description (build) step, where, after each change to the system state, the new configuration of the user interface is described by the developer and the framework computes the series of tree mutations necessary to reflect this new configuration.

Interpolation

#

Since Flutter's framework encourages developers to describe the interface configuration matching the current application state, a mechanism exists to implicitly animate between these configurations.

For example, suppose that in state S1 the interface consists of a circle, but in state S2 it consists of a square. Without an animation mechanism, the state change would have a jarring interface change. An implicit animation allows the circle to be smoothly squared over several frames.

Each feature that can be implicitly animated has a stateful widget that keeps a record of the current value of the input, and begins an animation sequence whenever the input value changes, transitioning from the current value to the new value over a specified duration.

This is implemented using lerp (linear interpolation) functions using immutable objects. Each state (circle and square, in this case) is represented as an immutable object that is configured with appropriate settings (color, stroke width, etc) and knows how to paint itself. When it is time to draw the intermediate steps during the animation, the start and end values are passed to the appropriate lerp function along with a t value representing the point along the animation, where 0.0 represents the start and 1.0 represents the end[8], and the function returns a third immutable object representing the intermediate stage.

For the circle-to-square transition, the lerp function would return an object representing a "rounded square" with a radius described as a fraction derived from the t value, a color interpolated using the lerp function for colors, and a stroke width interpolated using the lerp function for doubles. That object, which implements the same interface as circles and squares, would then be able to paint itself when requested to.

This technique allows the state machinery, the mapping of states to configurations, the animation machinery, the interpolation machinery, and the specific logic relating to how to paint each frame to be entirely separated from each other.

This approach is broadly applicable. In Flutter, basic types like Color and Shape can be interpolated, but so can much more elaborate types such as Decoration, TextStyle, or Theme. These are typically constructed from components that can themselves be interpolated, and interpolating the more complicated objects is often as simple as recursively interpolating all the values that describe the complicated objects.

Some interpolatable objects are defined by class hierarchies. For example, shapes are represented by the ShapeBorder interface, and there exists a variety of shapes, including BeveledRectangleBorder, BoxBorder, CircleBorder, RoundedRectangleBorder, and StadiumBorder. A single lerp function can't anticipate all possible types, and therefore the interface instead defines lerpFrom and lerpTo methods, which the static lerp method defers to. When told to interpolate from a shape A to a shape B, first B is asked if it can lerpFrom A, then, if it cannot, A is instead asked if it can lerpTo B. (If neither is possible, then the function returns A from values of t less than 0.5, and returns B otherwise.)

This allows the class hierarchy to be arbitrarily extended, with later additions being able to interpolate between previously-known values and themselves.

In some cases, the interpolation itself cannot be described by any of the available classes, and a private class is defined to describe the intermediate stage. This is the case, for instance, when interpolating between a CircleBorder and a RoundedRectangleBorder.

This mechanism has one further added advantage: it can handle interpolation from intermediate stages to new values. For example, half-way through a circle-to-square transition, the shape could be changed once more, causing the animation to need to interpolate to a triangle. So long as the triangle class can lerpFrom the rounded-square intermediate class, the transition can be seamlessly performed.

Conclusion

#

Flutter's slogan, "everything is a widget," revolves around building user interfaces by composing widgets that are, in turn, composed of progressively more basic widgets. The result of this aggressive composition is a large number of widgets that require carefully designed algorithms and data structures to process efficiently. With some additional design, these data structures also make it easy for developers to create infinite scrolling lists that build widgets on demand when they become visible.


Footnotes:


  1. For layout, at least. It might be revisited for painting, for building the accessibility tree if necessary, and for hit testing if necessary. ↩︎

  2. Reality, of course, is a bit more complicated. Some layouts involve intrinsic dimensions or baseline measurements, which do involve an additional walk of the relevant subtree (aggressive caching is used to mitigate the potential for quadratic performance in the worst case). These cases, however, are surprisingly rare. In particular, intrinsic dimensions are not required for the common case of shrink-wrapping. ↩︎

  3. Technically, the child's position is not part of its RenderBox geometry and therefore need not actually be calculated during layout. Many render objects implicitly position their single child at 0,0 relative to their own origin, which requires no computation or storage at all. Some render objects avoid computing the position of their children until the last possible moment (for example, during the paint phase), to avoid the computation entirely if they are not subsequently painted. ↩︎

  4. There exists one exception to this rule. As discussed in the Building widgets on demand section, some widgets can be rebuilt as a result of a change in layout constraints. If a widget marked itself dirty for unrelated reasons in the same frame that it also is affected by a change in layout constraints, it will be updated twice. This redundant build is limited to the widget itself and does not impact its descendants. ↩︎

  5. A key is an opaque object optionally associated with a widget whose equality operator is used to influence the reconciliation algorithm. ↩︎

  6. For accessibility, and to give applications a few extra milliseconds between when a widget is built and when it appears on the screen, the viewport creates (but does not paint) widgets for a few hundred pixels before and after the visible widgets. ↩︎

  7. This approach was first made popular by Facebook's React library. ↩︎

  8. In practice, the t value is allowed to extend past the 0.0-1.0 range, and does so for some curves. For example, the "elastic" curves overshoot briefly in order to represent a bouncing effect. The interpolation logic typically can extrapolate past the start or end as appropriate. For some types, for example, when interpolating colors, the t value is effectively clamped to the 0.0-1.0 range. ↩︎