The Rust Programming Language

Using Trait Objects That Allow for Values of Different Types

In Chapter 8, we mentioned that one limitation of vectors is that they can store elements of only one type. We created a workaround in Listing 8-9 where we defined a SpreadsheetCell enum that had variants to hold integers, floats, and text. This meant we could store different types of data in each cell and still have a vector that represented a row of cells. This is a perfectly good solution when our interchangeable items are a fixed set of types that we know when our code is compiled.

However, sometimes we want our library user to be able to extend the set of types that are valid in a particular situation. To show how we might achieve this, we’ll create an example graphical user interface (GUI) tool that iterates through a list of items, calling a draw method on each one to draw it to the screen—a common technique for GUI tools. We’ll create a library crate called gui that contains the structure of a GUI library. This crate might include some types for people to use, such as Button or TextField. In addition, gui users will want to create their own types that can be drawn: for instance, one programmer might add an Image and another might add a SelectBox.

We won’t implement a fully fledged GUI library for this example but will show how the pieces would fit together. At the time of writing the library, we can’t know and define all the types other programmers might want to create. But we do know that gui needs to keep track of many values of different types, and it needs to call a draw method on each of these differently typed values. It doesn’t need to know exactly what will happen when we call the draw method, just that the value will have that method available for us to call.

To do this in a language with inheritance, we might define a class named Component that has a method named draw on it. The other classes, such as Button, Image, and SelectBox, would inherit from Component and thus inherit the draw method. They could each override the draw method to define their custom behavior, but the framework could treat all of the types as if they were Component instances and call draw on them. But because Rust doesn’t have inheritance, we need another way to structure the gui library to allow users to extend it with new types.

Defining a Trait for Common Behavior

To implement the behavior we want gui to have, we’ll define a trait named Draw that will have one method named draw. Then we can define a vector that takes a trait object. A trait object points to both an instance of a type implementing our specified trait and a table used to look up trait methods on that type at runtime. We create a trait object by specifying some sort of pointer, such as a & reference or a Box<T> smart pointer, then the dyn keyword, and then specifying the relevant trait. (We’ll talk about the reason trait objects must use a pointer in Chapter 19 in the section “Dynamically Sized Types and the Sized Trait.”) We can use trait objects in place of a generic or concrete type. Wherever we use a trait object, Rust’s type system will ensure at compile time that any value used in that context will implement the trait object’s trait. Consequently, we don’t need to know all the possible types at compile time.

We’ve mentioned that, in Rust, we refrain from calling structs and enums “objects” to distinguish them from other languages’ objects. In a struct or enum, the data in the struct fields and the behavior in impl blocks are separated, whereas in other languages, the data and behavior combined into one concept is often labeled an object. However, trait objects are more like objects in other languages in the sense that they combine data and behavior. But trait objects differ from traditional objects in that we can’t add data to a trait object. Trait objects aren’t as generally useful as objects in other languages: their specific purpose is to allow abstraction across common behavior.

Listing 17-3 shows how to define a trait named Draw with one method named draw:

Filename: src/lib.rs

pub trait Draw {
    fn draw(&self);
}

Listing 17-3: Definition of the Draw trait

This syntax should look familiar from our discussions on how to define traits in Chapter 10. Next comes some new syntax: Listing 17-4 defines a struct named Screen that holds a vector named components. This vector is of type Box<dyn Draw>, which is a trait object; it’s a stand-in for any type inside a Box that implements the Draw trait.

Filename: src/lib.rs

pub trait Draw {
    fn draw(&self);
}

pub struct Screen {
    pub components: Vec<Box<dyn Draw>>,
}

Listing 17-4: Definition of the Screen struct with a components field holding a vector of trait objects that implement the Draw trait

On the Screen struct, we’ll define a method named run that will call the draw method on each of its components, as shown in Listing 17-5:

Filename: src/lib.rs

pub trait Draw {
    fn draw(&self);
}

pub struct Screen {
    pub components: Vec<Box<dyn Draw>>,
}

impl Screen {
    pub fn run(&self) {
        for component in self.components.iter() {
            component.draw();
        }
    }
}

Listing 17-5: A run method on Screen that calls the draw method on each component

This works differently from defining a struct that uses a generic type parameter with trait bounds. A generic type parameter can only be substituted with one concrete type at a time, whereas trait objects allow for multiple concrete types to fill in for the trait object at runtime. For example, we could have defined the Screen struct using a generic type and a trait bound as in Listing 17-6:

Filename: src/lib.rs

pub trait Draw {
    fn draw(&self);
}

pub struct Screen<T: Draw> {
    pub components: Vec<T>,
}

impl<T> Screen<T>
where
    T: Draw,
{
    pub fn run(&self) {
        for component in self.components.iter() {
            component.draw();
        }
    }
}

Listing 17-6: An alternate implementation of the Screen struct and its run method using generics and trait bounds

This restricts us to a Screen instance that has a list of components all of type Button or all of type TextField. If you’ll only ever have homogeneous collections, using generics and trait bounds is preferable because the definitions will be monomorphized at compile time to use the concrete types.

On the other hand, with the method using trait objects, one Screen instance can hold a Vec<T> that contains a Box<Button> as well as a Box<TextField>. Let’s look at how this works, and then we’ll talk about the runtime performance implications.

Implementing the Trait

Now we’ll add some types that implement the Draw trait. We’ll provide the Button type. Again, actually implementing a GUI library is beyond the scope of this book, so the draw method won’t have any useful implementation in its body. To imagine what the implementation might look like, a Button struct might have fields for width, height, and label, as shown in Listing 17-7:

Filename: src/lib.rs

pub trait Draw {
    fn draw(&self);
}

pub struct Screen {
    pub components: Vec<Box<dyn Draw>>,
}

impl Screen {
    pub fn run(&self) {
        for component in self.components.iter() {
            component.draw();
        }
    }
}

pub struct Button {
    pub width: u32,
    pub height: u32,
    pub label: String,
}

impl Draw for Button {
    fn draw(&self) {
        // code to actually draw a button
    }
}

Listing 17-7: A Button struct that implements the Draw trait

The width, height, and label fields on Button will differ from the fields on other components; for example, a TextField type might have those same fields plus a placeholder field. Each of the types we want to draw on the screen will implement the Draw trait but will use different code in the draw method to define how to draw that particular type, as Button has here (without the actual GUI code, as mentioned). The Button type, for instance, might have an additional impl block containing methods related to what happens when a user clicks the button. These kinds of methods won’t apply to types like TextField.

If someone using our library decides to implement a SelectBox struct that has width, height, and options fields, they implement the Draw trait on the SelectBox type as well, as shown in Listing 17-8:

Filename: src/main.rs

use gui::Draw;

struct SelectBox {
    width: u32,
    height: u32,
    options: Vec<String>,
}

impl Draw for SelectBox {
    fn draw(&self) {
        // code to actually draw a select box
    }
}

fn main() {}

Listing 17-8: Another crate using gui and implementing the Draw trait on a SelectBox struct

Using the Trait

Our library’s user can now write their main function to create a Screen instance. To the Screen instance, they can add a SelectBox and a Button by putting each in a Box<T> to become a trait object. They can then call the run method on the Screen instance, which will call draw on each of the components. Listing 17-9 shows this implementation:

Filename: src/main.rs

use gui::Draw;

struct SelectBox {
    width: u32,
    height: u32,
    options: Vec<String>,
}

impl Draw for SelectBox {
    fn draw(&self) {
        // code to actually draw a select box
    }
}

use gui::{Button, Screen};

fn main() {
    let screen = Screen {
        components: vec![
            Box::new(SelectBox {
                width: 75,
                height: 10,
                options: vec![
                    String::from("Yes"),
                    String::from("Maybe"),
                    String::from("No"),
                ],
            }),
            Box::new(Button {
                width: 50,
                height: 10,
                label: String::from("OK"),
            }),
        ],
    };

    screen.run();
}

Listing 17-9: Using trait objects to store values of different types that implement the same trait

When we wrote the library, we didn’t know that someone might add the SelectBox type, but our Screen implementation was able to operate on the new type and draw it because SelectBox implements the Draw trait, which means it implements the draw method.

This concept—of being concerned only with the messages a value responds to rather than the value’s concrete type—is similar to the concept of duck typing in dynamically typed languages: if it walks like a duck and quacks like a duck, then it must be a duck! In the implementation of run on Screen in Listing 17-5, run doesn’t need to know what the concrete type of each component is. It doesn’t check whether a component is an instance of a Button or a SelectBox, it just calls the draw method on the component. By specifying Box<dyn Draw> as the type of the values in the components vector, we’ve defined Screen to need values that we can call the draw method on.

The advantage of using trait objects and Rust’s type system to write code similar to code using duck typing is that we never have to check whether a value implements a particular method at runtime or worry about getting errors if a value doesn’t implement a method but we call it anyway. Rust won’t compile our code if the values don’t implement the traits that the trait objects need.

For example, Listing 17-10 shows what happens if we try to create a Screen with a String as a component:

Filename: src/main.rs

use gui::Screen;

fn main() {
    let screen = Screen {
        components: vec![Box::new(String::from("Hi"))],
    };

    screen.run();
}

Listing 17-10: Attempting to use a type that doesn’t implement the trait object’s trait

We’ll get this error because String doesn’t implement the Draw trait:

$ cargo run
   Compiling gui v0.1.0 (file:///projects/gui)
error[E0277]: the trait bound `String: Draw` is not satisfied
 --> src/main.rs:5:26
  |
5 |         components: vec![Box::new(String::from("Hi"))],
  |                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ the trait `Draw` is not implemented for `String`
  |
  = help: the trait `Draw` is implemented for `Button`
  = note: required for the cast from `Box<String>` to `Box<dyn Draw>`

For more information about this error, try `rustc --explain E0277`.
error: could not compile `gui` (bin "gui") due to 1 previous error

This error lets us know that either we’re passing something to Screen we didn’t mean to pass and so should pass a different type or we should implement Draw on String so that Screen is able to call draw on it.

Trait Objects and Type Inference

One downside to using trait objects is how they interact with type inference. For example, consider type inference for Vec<T>. When T is not a trait object, Rust just needs to know the type of a single element in the vector to infer T. So an empty vector causes a type inference error:

fn main() {
let v = vec![];
// error[E0282]: type annotations needed for `Vec<T>`
}

But adding an element enables Rust to infer the type of the vector:

fn main() {
let v = vec!["Hello world"];
// ok, v : Vec<&str>
}

Type inference is trickier for trait objects. For example, say we tried to factor the components array in Listing 17-9 into a separate variable, like this:

fn main() {
    let components = vec![
        Box::new(SelectBox { /* .. */ }),
        Box::new(Button { /* .. */ }),
    ];
    let screen = Screen { components };
    screen.run();
}

Listing 17-11: Factoring the components array causes a type error

This refactor causes the program to no longer compile! The compiler rejects this program with the following error:

error[E0308]: mismatched types
   --> test.rs:55:14
    |
55  |       Box::new(Button {
    |  _____--------_^
    | |     |
    | |     arguments to this function are incorrect
56  | |       width: 50,
57  | |       height: 10,
58  | |       label: String::from("OK"),
59  | |     }),
    | |_____^ expected `SelectBox`, found `Button`

In Listing 17-09, the compiler understands that the components vector must have the type Vec<Box<dyn Draw>> because that’s specified in the Screen struct definition. But in Listing 17-11, the compiler loses that information at the point where components is defined. To fix the issue, you have to give a hint to the type inference algorithm. That can either be via an explicit cast on any element of the vector, like this:

  let components = vec![
        Box::new(SelectBox { /* .. */ }) as Box<dyn Draw>,
        Box::new(Button { /* .. */ }),
  ];

Or it can be via a type annotation on the let-binding, like this:

  let components: Vec<Box<dyn Draw>> = vec![
        Box::new(SelectBox { /* .. */ }),
        Box::new(Button { /* .. */ }),
  ];

In general, it is good to be aware that using trait objects can cause a worse developer experience for API clients in the case of type inference.

Trait Objects Perform Dynamic Dispatch

Recall in the “Performance of Code Using Generics” section in Chapter 10 our discussion on the monomorphization process performed by the compiler when we use trait bounds on generics: the compiler generates nongeneric implementations of functions and methods for each concrete type that we use in place of a generic type parameter. The code that results from monomorphization is doing static dispatch, which is when the compiler knows what method you’re calling at compile time. This is opposed to dynamic dispatch, which is when the compiler can’t tell at compile time which method you’re calling. In dynamic dispatch cases, the compiler emits code that at runtime will figure out which method to call.

When we use trait objects, Rust must use dynamic dispatch. The compiler doesn’t know all the types that might be used with the code that’s using trait objects, so it doesn’t know which method implemented on which type to call. Instead, at runtime, Rust uses the pointers inside the trait object to know which method to call. This lookup incurs a runtime cost that doesn’t occur with static dispatch. Dynamic dispatch also prevents the compiler from choosing to inline a method’s code, which in turn prevents some optimizations. However, we did get extra flexibility in the code that we wrote in Listing 17-5 and were able to support in Listing 17-9, so it’s a trade-off to consider.