Improve performance using nightly features

[kdy1]

Context

I'm working to improve the performance of the deserializer. (See swc-project/swc#3160)

I found that serde::__private::de::Content is created quite frequently from __deserializer and __deserializer is serde::__private::de::ContentDeserailizer in many cases.

Try: specialization

So I copied code from serde, and specialized implementation of deserialize for Content.

impl<'de> Deserialize<'de> for Content<'de> {
    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
    where
        D: Deserializer<'de>,
    {
        <Content as DeserImpl<'de, D>>::deserialize_spec(deserializer)
    }
}

trait DeserImpl<'de, D>: Sized
where
    D: Deserializer<'de>,
{
    fn deserialize_spec(deserializer: D) -> Result<Self, D::Error>;
}

impl<'de, D> DeserImpl<'de, D> for Content<'de>
where
    D: Deserializer<'de>,
{
    default fn deserialize_spec(deserializer: D) -> Result<Self, D::Error> {
        // Untagged and internally tagged enums are only supported in
        // self-describing formats.
        let visitor = ContentVisitor { value: PhantomData };
        deserializer.deserialize_any(visitor)
    }
}

impl<'de, E> DeserImpl<'de, ContentDeserializer<'de, E>> for Content<'de>
where
    ContentDeserializer<'de, E>: Deserializer<'de, Error = E>,
{
    fn deserialize_spec(deserializer: ContentDeserializer<'de, E>) -> Result<Self, E> {
        Ok(deserializer.content)
    }
}

Previous

test json_deserialize                     ... bench:     1,226,131 ns/iter (+/- 58,655)

After

test json_deserialize                     ... bench:     621,480 ns/iter (+/- 69,766)

I spotted few more creation of Content from serde::__private::de::ContentDeserailizer due to #[flatten] and going to implement custom #[derive(Deserialize)] which uses my fork of ContentDeserailizer.

But this is useful for many users ofserde, so I'm filing an issue.

Using fork

To quickly apply specialization for ContentDeserailizer of #[flatten], I created a fork (https://github.com/kdy1/serde/tree/content-specialization). Using it, I could get the result before creating a custom dervie macro.

test json_deserialize                     ... bench:     339,526 ns/iter (+/- 41,530)

[dtolnay]

specialization is not suitable for this because of the unsound interaction with lifetimes, and I would guess min_specialization is not expressive enough to implement this.

The same behavior can be achieved without specialization though. #2148