string serialization escaping optimisations

[conradludgate]

While serializing strings, we enter the hot loop checking 1 byte at a time. By massaging the code and using a little bit of unsafe, we can optimise this hot loop a bit.

This benchmark shows about a 2-5% speedup when working with string heavy documents

Benchmark was run with target x86_64-unknown-linux-gnu on an Intel Xeon Platinum 8375C CPU @ 2.90GHz, as well as on an Apple M4 Max.

use std::hint::black_box;

use criterion::{criterion_group, criterion_main, Criterion};
use serde_json::Value;

pub fn k8s(c: &mut Criterion) {
    // https://raw.githubusercontent.com/kubernetes/kubernetes/v1.33.3/api/openapi-spec/swagger.json
    let k8s = std::fs::read_to_string("benches/k8s-openapi.json").unwrap();
    let value: Value = serde_json::from_str(&k8s).unwrap();
    drop(k8s);

    let mut v = Vec::new();
    serde_json::to_writer_pretty(&mut v, &value).unwrap();

    c.bench_function("pretty", |b| {
        b.iter(|| {
            v.clear();
            serde_json::to_writer_pretty(&mut v, &value).unwrap();
            black_box(&v[..]);
        })
    });

    c.bench_function("compact", |b| {
        b.iter(|| {
            v.clear();
            serde_json::to_writer(&mut v, &value).unwrap();
            black_box(&v[..]);
        })
    });
}

criterion_group!(benches, k8s);
criterion_main!(benches);

[dtolnay]

Thanks!

Would there be any benefit to coordinating the discriminants of CharEscape with the values that are being converted to/from it in this code, even if as cannot be used?

#[repr(u8)]
pub enum CharEscape {
    Quote = b'"',
    ReverseSolidus = b'\\',
    Solidus = b'/',
    Backspace = b'b',
    FormFeed = b'f',
    LineFeed = b'n',
    CarriageReturn = b'r',
    Tab = b't',
    AsciiControl(u8) = b'u',
}

[conradludgate]

Thanks!

Would there be any benefit to coordinating the discriminants of CharEscape with the values that are being converted to/from it in this code, even if as cannot be used?

#[repr(u8)]
pub enum CharEscape {
    Quote = b'"',
    ReverseSolidus = b'\\',
    Solidus = b'/',
    Backspace = b'b',
    FormFeed = b'f',
    LineFeed = b'n',
    CarriageReturn = b'r',
    Tab = b't',
    AsciiControl(u8) = b'u',
}

I had that in a draft of this PR, but I couldn't get any benefit from it unfortunately. Maybe it still makes sense from a semantic point of view