Module std::prim_never

🔬This is a nightly-only experimental API. (never_type #35121)

The ! type, also called “never”.

! represents the type of computations which never resolve to any value at all. For example, the exit function fn exit(code: i32) -> ! exits the process without ever returning, and so returns !.

break, continue and return expressions also have type !. For example we are allowed to write:

#![feature(never_type)]
let x: ! = {
    return 123
};

Although the let is pointless here, it illustrates the meaning of !. Since x is never assigned a value (because return returns from the entire function), x can be given type !. We could also replace return 123 with a panic! or a never-ending loop and this code would still be valid.

A more realistic usage of ! is in this code:

let num: u32 = match get_a_number() {
    Some(num) => num,
    None => break,
};

Both match arms must produce values of type u32, but since break never produces a value at all we know it can never produce a value which isn’t a u32. This illustrates another behaviour of the ! type - expressions with type ! will coerce into any other type.

! and generics

Infallible errors

The main place you’ll see ! used explicitly is in generic code. Consider the FromStr trait:

trait FromStr: Sized {
    type Err;
    fn from_str(s: &str) -> Result<Self, Self::Err>;
}

When implementing this trait for String we need to pick a type for Err. And since converting a string into a string will never result in an error, the appropriate type is !. (Currently the type actually used is an enum with no variants, though this is only because ! was added to Rust at a later date and it may change in the future.) With an Err type of !, if we have to call String::from_str for some reason the result will be a Result<String, !> which we can unpack like this:

#![feature(exhaustive_patterns)]
use std::str::FromStr;
let Ok(s) = String::from_str("hello");

Since the Err variant contains a !, it can never occur. If the exhaustive_patterns feature is present this means we can exhaustively match on Result<T, !> by just taking the Ok variant. This illustrates another behaviour of ! - it can be used to “delete” certain enum variants from generic types like Result.

Infinite loops

While Result<T, !> is very useful for removing errors, ! can also be used to remove successes as well. If we think of Result<T, !> as “if this function returns, it has not errored,” we get a very intuitive idea of Result<!, E> as well: if the function returns, it has errored.

For example, consider the case of a simple web server, which can be simplified to:

loop {
    let (client, request) = get_request().expect("disconnected");
    let response = request.process();
    response.send(client);
}

Currently, this isn’t ideal, because we simply panic whenever we fail to get a new connection. Instead, we’d like to keep track of this error, like this:

loop {
    match get_request() {
        Err(err) => break err,
        Ok((client, request)) => {
            let response = request.process();
            response.send(client);
        },
    }
}

Now, when the server disconnects, we exit the loop with an error instead of panicking. While it might be intuitive to simply return the error, we might want to wrap it in a Result<!, E> instead:

fn server_loop() -> Result<!, ConnectionError> {
    loop {
        let (client, request) = get_request()?;
        let response = request.process();
        response.send(client);
    }
}

Now, we can use ? instead of match, and the return type makes a lot more sense: if the loop ever stops, it means that an error occurred. We don’t even have to wrap the loop in an Ok because ! coerces to Result<!, ConnectionError> automatically.

! and traits

When writing your own traits, ! should have an impl whenever there is an obvious impl which doesn’t panic!. The reason is that functions returning an impl Trait where ! does not have an impl of Trait cannot diverge as their only possible code path. In other words, they can’t return ! from every code path. As an example, this code doesn’t compile:

use std::ops::Add;

fn foo() -> impl Add<u32> {
    unimplemented!()
}

But this code does:

use std::ops::Add;

fn foo() -> impl Add<u32> {
    if true {
        unimplemented!()
    } else {
        0
    }
}

The reason is that, in the first example, there are many possible types that ! could coerce to, because many types implement Add<u32>. However, in the second example, the else branch returns a 0, which the compiler infers from the return type to be of type u32. Since u32 is a concrete type, ! can and will be coerced to it. See issue #36375 for more information on this quirk of !.

As it turns out, though, most traits can have an impl for !. Take Debug for example:

#![feature(never_type)]
impl Debug for ! {
    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
        *self
    }
}

Once again we’re using !’s ability to coerce into any other type, in this case fmt::Result. Since this method takes a &! as an argument we know that it can never be called (because there is no value of type ! for it to be called with). Writing *self essentially tells the compiler “We know that this code can never be run, so just treat the entire function body as having type fmt::Result”. This pattern can be used a lot when implementing traits for !. Generally, any trait which only has methods which take a self parameter should have such an impl.

On the other hand, one trait which would not be appropriate to implement is Default:

trait Default {
    fn default() -> Self;
}

Since ! has no values, it has no default value either. It’s true that we could write an impl for this which simply panics, but the same is true for any type (we could impl Default for (eg.) File by just making default() panic.)