enum_dispatch 0.1.5

Near drop-in replacement for dynamic-dispatched method calls with up to 10x the speed
enum_dispatch-0.1.5 doesn't have any documentation.

enum_dispatch

Docs License

enum_dispatch transforms your trait objects into concrete compound types, increasing their method call speed up to 10x.

example

If you have the following code...

// We already have defined MyImplementorA and MyImplementorB, and implemented MyBehavior for each.

trait MyBehavior {
    fn my_trait_method(&self);
}

// Any pointer type -- Box, &, etc.
let a: Box<dyn MyBehavior> = Box::new(MyImplementorA::new());

a.my_trait_method();    //dynamic dispatch

...then you can improve its performance using enum_dispatch like this:

#[enum_dispatch]
enum MyBehaviorEnum {
    MyImplementorA,
    MyImplementorB,
}

#[enum_dispatch(MyBehaviorEnum)]
trait MyBehavior {
    fn my_trait_method(&self);
}

let a: MyBehaviorEnum = MyImplementorA::new().into();

a.my_trait_method();    //no dynamic dispatch

Notice the differences:

  1. The new enum, MyBehaviorEnum, whose variants are simply types implementing the trait MyBehavior.
  2. The new enum_dispatch attributes applied to the enum and trait, linking the two together.
  3. The removal of the Box allocation.
  4. Faster trait method calls!

how to use

  1. Add enum_dispatch as a Cargo.toml dependency, and import enum_dispatch::enum_dispatch to your code.
  2. Create a new enum whose variants are any in-scope trait implementors you've defined.
  3. Add an #[enum_dispatch] attribute to either the enum or trait definition. This will "register" it with the enum_dispatch library. Take note of the name of the enum or trait it was applied to -- we'll call it FirstBlockName.
  4. Add an #[enum_dispatch(FirstBlockName)] attribute to the remaining definition. This will "link" it with the previously registered definition.
  5. Update your dynamic types to use the new enum instead. You can use .into() from any trait implementor to automatically turn it into an enum variant.

performance

More information on performance can be found in the docs, and benchmarks are available in the benches directory. The following benchmark results give a taste of what can be achieved using enum_dispatch. They compare the speed of repeatedly accessing method calls on a Vec of 1024 trait objects of randomized concrete types using either Boxed trait objects, & referenced trait objects, or enum_dispatched enum types.

test benches::boxdyn_homogeneous_vec       ... bench:   5,900,191 ns/iter (+/- 95,169)
test benches::refdyn_homogeneous_vec       ... bench:   5,658,461 ns/iter (+/- 137,128)
test benches::enumdispatch_homogeneous_vec ... bench:     479,630 ns/iter (+/- 3,531)

bonus features

While enum_dispatch was built with performance in mind, the transformations it applies make all your data structures much more visible to the compiler. That means you can use serde or other similar tools on your trait objects!

custom variant names

By default, enum_dispatch will expand each enum variants into one with a single unnamed field of the same name as the internal type. If for some reason you'd like to use a custom name for a particular type in an enum_dispatch variant, you can do so as shown below:

#[enum_dispatch]
enum MyTypes {
    TypeA,
    CustomVariantName(TypeB),
}

let mt: MyTypes = TypeB::new().into();
match mt {
    TypeA(a) => { /* `a` is a TypeA */ },
    CustomVariantName(b) => { /* `b` is a TypeB */ },
}

Custom variant names are required for enums and traits with generic type arguments, which can also be optimized by enum_dispatch. Check out this generics example to see how that works.

troubleshooting

no impls created?

Be careful not to forget an attribute or mistype the name in a linking attribute. If parsing is completed before a linking attribute is found, no implementations will be generated. Due to technical limitations of the macro system, it's impossible to properly warn the user in this scenario.

can't parse enum?

Types must be fully in scope to be usable as an enum variant. For example, the following will fail to compile:

#[enum_dispatch]
enum Fails {
    crate::A::TypeA,
    crate::B::TypeB,
}

This is because the enum must be correctly parseable before macro expansion. Instead, import the types first:

use crate::A::TypeA;
use crate::B::TypeB;

#[enum_dispatch]
enum Succeeds {
    TypeA,
    TypeB,
}

technical details

enum_dispatch is a procedural macro that implements a trait for a fixed set of types in the form of an enum. This is faster than using dynamic dispatch because type information will be "built-in" to each enum, avoiding a costly vtable lookup.

Since enum_dispatch is a procedural macro, it works by processing and expanding attributed code at compile time. The folowing sections explain how the example above might be transformed.

enum processing

There's no way to define an enum whose variants are actual concrete types. To get around this, enum_dispatch rewrites its body by generating a name for each variant and using the provided type as its single tuple-style argument. The name for each variant isn't particularly important for most purposes, but enum_dispatch will currently just use the name of the provided type.

enum MyBehaviorEnum {
    MyImplementorA(MyImplementorA),
    MyImplementorB(MyImplementorB),
}

trait processing

enum_dispatch doesn't actually process annotated traits! However, it still requires access to the trait definition so it can take note of the trait's name, as well as the function signatures of any methods inside it.

trait impl creation

Whenever enum_dispatch is able to "link" two definitions together, it will generate an impl block, implementing the trait for the enum. In the above example, the linkage is completed by the MyBehavior trait definition, so impl blocks will be generated directly below that trait. The generated impl block might look something like this:

impl MyBehavior for MyBehaviorEnum {
    fn my_trait_method(&self) {
        match self {
            MyImplementorA(inner) => inner.my_trait_method(),
            MyImplementorB(inner) => inner.my_trait_method(),
        }
    }
}

Additional trait methods would be expanded accordingly, and additional enum variants would correspond to additional match arms in each method definition. It's easy to see how quickly this can become unmanageable in manually written code!

'From' impl creation

Normally, it would be impossible to initialize one of the new enum variants without knowing its name. However, with implementations of From<T> for each variant, that requirement is alleviated. The generated implementations could look like the following:

impl From<MyImplementorA> for MyBehaviorEnum {
    fn from(inner: MyImplementorA) -> MyBehaviorEnum {
        MyBehaviorEnum::MyImplementorA(inner)
    }
}

impl From<MyImplementorB> for MyBehaviorEnum {
    fn from(inner: MyImplementorB) -> MyBehaviorEnum {
        MyBehaviorEnum::MyImplementorB(inner)
    }
}

As with above, having a large number of possible type variants would make this very difficult to maintain by hand.

registry and linkage

Anyone closely familiar with writing macros will know that they must be processed locally, with no context about the surrounding source code. Additionally, parsed syntax items in syn are !Send and !Sync. This is for good reason -- with multithreaded compilation and macro expansion, there are no guarantees on the order or lifetime of a reference to any given block of code. Unfortunately, it also prevents referencing syntax between separate macro invocations.

In the interest of convenience, enum_dispatch circumvents these restrictions by converting syntax into a String and storing it in once_cell lazily initialized Mutex<HashMap<String, String>>s whose keys are either the trait or enum names.

There is also a similar HashMap dedicated to "deferred" links, since definitions in different files could be encountered in arbitrary orders. If a linking attribute (with one argument) occurs before the corresponding registry attribute (with no arguments), the argument will be stored as a deferred link. Once that argument's definition is encountered, impl blocks can be created as normal.

Because of the link deferral mechanism, it's not an error to encounter a linking attribute without being able to implement it. enum_dispatch will simply expect to find the corresponding registry attribute later in parsing. However, there's no way to insert a callback to check that all deferred links have been processed once all the original source code has been parsed, explaining the impossibility of warning the user of unlinked attributes.