entrait

A proc macro to ease development using Inversion of Control patterns in Rust.

entrait is used to generate a trait from the definition of a regular function. The main use case for this is that other functions may depend upon the trait instead of the concrete implementation, enabling better test isolation.

The macro looks like this:

```rust

[entrait(MyFunction)]

fn my_function(deps: &D) { } ```

which generates the trait MyFunction:

rust trait MyFunction { fn my_function(&self); }

my_function's first and only parameter is deps which is generic over some unknown type D. This would correspond to the self parameter in the trait. But what is this type supposed to be? We can generate an implementation in the same go, using for Type:

```rust struct App;

[entrait::entrait(MyFunction for App)]

fn my_function(deps: &D) { }

// Generated: // trait MyFunction { // fn myfunction(&self); // } // // impl MyFunction for App { // fn myfunction(&self) { // my_function(self) // } // }

fn main() { let app = App; app.my_function(); } ```

The advantage of this pattern comes into play when a function declares its dependencies, as trait bounds:

```rust

[entrait(Foo for App)]

fn foo(deps: &(impl Bar)) { deps.bar(); }

[entrait(Bar for App)]

fn bar(deps: &D) { } ```

The functions may take any number of parameters, but the first one is always considered specially as the "dependency parameter".

Functions may also be non-generic, depending directly on the App:

```rust

[entrait(ExtractSomething for App)]

fn extractsomething(app: &App) -> SomeType { app.something } ```

These kinds of functions may be considered "leaves" of a dependency tree.

"Philosophy"

The idea behind entrait is to explore a specific architectural pattern: * Interfaces with one runtime implementation * named traits as the interface of single functions

entrait does not implement Dependency Injection (DI). DI is a strictly object-oriented concept that will often look awkward in Rust. The author thinks of DI as the "reification of code modules": In a DI-enabled programming environment, code modules are grouped together as objects and other modules may depend upon the interface of such an object by receiving some instance that implements it. When this pattern is applied successively, one ends up with an in-memory dependency graph of high-level modules.

entrait tries to turn this around by saying that the primary abstraction that is depended upon is a set of functions, not a set of code modules.

An architectural consequence is that one ends up with one ubiquitous type that represents a running application that implements all these function abstraction traits. But the point is that this is all loosely coupled: Most function definitions themselves do not refer to this god-like type, they only depend upon traits.

async support

Since Rust at the time of writing does not natively support async methods in traits, you may opt in to having #[async_trait] generated for your trait:

```rust

[entrait(Foo, async_trait=true)]

async fn foo(deps: &D) { } `` This is designed to be forwards compatible with real async fn in traits. When that day comes, you should be able to just remove theasync_trait=true` to get a proper zero-cost future.

Mock support

The macro supports autogenerating [mockall] mock structs:

```rust

[entrait(Foo, mockall=true)]

fn foo(_: &D) -> u32 { unimplemented!() }

fn my_func(deps: &(impl Foo)) -> u32 { deps.foo() }

fn main() { let mut deps = MockFoo::new(); deps.expectfoo().returning(|| 42); asserteq!(42, my_func(&deps)); } `` Usingmockallis easy enough when there is only one trait bound, because the generated trait need only be attributed withmockall::automock`.

multiple trait bounds with unimock

With multiple trait bounds, this becomes a little harder: We need some concrete struct that implement all the given traits. This is easily solved by the crate [unimock], and using unimock = true:

```rust use unimock::Unimock;

[entrait(Foo, unimock=true)]

fn foo(_: &D) -> u32 { unimplemented!() }

[entrait(Bar, unimock=true)]

fn bar(_: &D) -> u32 { unimplemented!() }

fn my_func(deps: &(impl Foo + Bar)) -> u32 { deps.foo() + deps.bar() }

fn main() { let deps = Unimock::new() .mock(|foo: &mut MockFoo| { foo.expectfoo().returning(|| 40); }) .mock(|bar: &mut MockBar| { bar.expectbar().returning(|| 2); });

assert_eq!(42, my_func(&deps));

} ```

unimock = true implies mockall = true.

conditional mock implementations

Most often, you will only need to generate mock implementations in test code, and skip this for production code. For this, there are the test variants:

which puts the corresponding attributes in #[cfg_attr(test, ...)]:

```rust

[entrait(Foo, unimock=test)]

fn foo(_: &D) -> u32 { unimplemented!() }

fn takes_foo(foo: impl Foo) {}

fn main() { // we can still instantiate Unimock, but it's not useful, // because now it doesn't implement Foo: let mock = unimock::Unimock::new(); //takes_foo(mock); //--------- ^^^^ the trait Foo is not implemented for Unimock }

[test]

fn test() { // this compiles! let mock = unimock::Unimock::new(); takes_foo(mock); } ```

This is opt-in because there could be scenarios where this behaviour is not desired, e.g. when you write a library and want mocks exported for those.

License: MIT