mirror of
https://github.com/rust-lang/rust.git
synced 2024-11-26 08:44:35 +00:00
bfe5e8cef6
Small optimization for integers Display implementation
This is a first pass to try to speed up a bit integers `Display` implementation. The idea behind this is to reduce the stack usage for the buffer storing the output (shouldn't be visible in bench normally) and some small specialization which benefits a lot to smaller integers like `u8` and `i8`.
Here are the results of the benchmarks:
| bench name | current std | with this PR |
|-|-|-|
| bench_std_fmt::bench_i16_0 | 16.45 ns/iter (+/- 0.25) | 16.50 ns/iter (+/- 0.15) |
| bench_std_fmt::bench_i16_max | 17.83 ns/iter (+/- 0.66) | 17.58 ns/iter (+/- 0.10) |
| bench_std_fmt::bench_i16_min | 20.97 ns/iter (+/- 0.49) | 20.50 ns/iter (+/- 0.28) |
| bench_std_fmt::bench_i32_0 | 16.63 ns/iter (+/- 0.06) | 16.62 ns/iter (+/- 0.07) |
| bench_std_fmt::bench_i32_max | 19.79 ns/iter (+/- 0.43) | 19.55 ns/iter (+/- 0.14) |
| bench_std_fmt::bench_i32_min | 22.97 ns/iter (+/- 0.50) | 22.08 ns/iter (+/- 0.08) |
| bench_std_fmt::bench_i64_0 | 16.63 ns/iter (+/- 0.39) | 16.69 ns/iter (+/- 0.44) |
| bench_std_fmt::bench_i64_half | 19.60 ns/iter (+/- 0.05) | 19.10 ns/iter (+/- 0.05) |
| bench_std_fmt::bench_i64_max | 25.22 ns/iter (+/- 0.34) | 24.43 ns/iter (+/- 0.02) |
| bench_std_fmt::bench_i8_0 | 16.27 ns/iter (+/- 0.32) | 15.80 ns/iter (+/- 0.17) |
| bench_std_fmt::bench_i8_max | 16.71 ns/iter (+/- 0.09) | 16.25 ns/iter (+/- 0.01) |
| bench_std_fmt::bench_i8_min | 20.07 ns/iter (+/- 0.22) | 19.80 ns/iter (+/- 0.30) |
| bench_std_fmt::bench_u128_0 | 21.37 ns/iter (+/- 0.24) | 21.35 ns/iter (+/- 0.35) |
| bench_std_fmt::bench_u128_max | 48.13 ns/iter (+/- 0.20) | 48.78 ns/iter (+/- 0.29) |
| bench_std_fmt::bench_u16_0 | 16.48 ns/iter (+/- 0.46) | 16.03 ns/iter (+/- 0.39) |
| bench_std_fmt::bench_u16_max | 17.31 ns/iter (+/- 0.32) | 17.41 ns/iter (+/- 0.32) |
| bench_std_fmt::bench_u16_min | 16.40 ns/iter (+/- 0.45) | 16.02 ns/iter (+/- 0.39) |
| bench_std_fmt::bench_u32_0 | 16.17 ns/iter (+/- 0.04) | 16.29 ns/iter (+/- 0.16) |
| bench_std_fmt::bench_u32_max | 19.00 ns/iter (+/- 0.10) | 19.16 ns/iter (+/- 0.28) |
| bench_std_fmt::bench_u32_min | 16.16 ns/iter (+/- 0.09) | 16.28 ns/iter (+/- 0.11) |
| bench_std_fmt::bench_u64_0 | 16.22 ns/iter (+/- 0.22) | 16.14 ns/iter (+/- 0.18) |
| bench_std_fmt::bench_u64_half | 19.25 ns/iter (+/- 0.07) | 18.95 ns/iter (+/- 0.05) |
| bench_std_fmt::bench_u64_max | 24.31 ns/iter (+/- 0.08) | 24.18 ns/iter (+/- 0.08) |
| bench_std_fmt::bench_u8_0 | 15.76 ns/iter (+/- 0.08) | 15.66 ns/iter (+/- 0.08) |
| bench_std_fmt::bench_u8_max | 16.53 ns/iter (+/- 0.03) | 16.29 ns/iter (+/- 0.02) |
| bench_std_fmt::bench_u8_min | 15.77 ns/iter (+/- 0.06) | 15.67 ns/iter (+/- 0.02) |
The source code is:
<details>
<summary>source code</summary>
```rust
#![feature(test)]
#![allow(non_snake_case)]
#![allow(clippy::cast_lossless)]
extern crate test;
macro_rules! benches {
($($name:ident($value:expr))*) => {
mod bench_std_fmt {
use std::io::Write;
use test::{Bencher, black_box};
$(
#[bench]
fn $name(b: &mut Bencher) {
let mut buf = Vec::with_capacity(40);
b.iter(|| {
buf.clear();
write!(&mut buf, "{}", black_box($value)).unwrap();
black_box(&buf);
});
}
)*
}
}
}
benches! {
bench_u64_0(0u64)
bench_u64_half(u32::max_value() as u64)
bench_u64_max(u64::max_value())
bench_i64_0(0i64)
bench_i64_half(i32::max_value() as i64)
bench_i64_max(i64::max_value())
bench_u16_0(0u16)
bench_u16_min(u16::min_value())
bench_u16_max(u16::max_value())
bench_i16_0(0i16)
bench_i16_min(i16::min_value())
bench_i16_max(i16::max_value())
bench_u128_0(0u128)
bench_u128_max(u128::max_value())
bench_i8_0(0i8)
bench_i8_min(i8::min_value())
bench_i8_max(i8::max_value())
bench_u8_0(0u8)
bench_u8_min(u8::min_value())
bench_u8_max(u8::max_value())
bench_u32_0(0u32)
bench_u32_min(u32::min_value())
bench_u32_max(u32::max_value())
bench_i32_0(0i32)
bench_i32_min(i32::min_value())
bench_i32_max(i32::max_value())
}
```
</details>
And then I ran the equivalent code (source code below) in callgrind with [callgrind_differ](https://github.com/Ethiraric/callgrind_differ) to generate a nice output and here's the result:
```
core::fmt::num:👿:<impl core::fmt::Display for i16>::fmt | 1300000 | - 70000 - 5.385% 1230000
core::fmt::num:👿:<impl core::fmt::Display for i32>::fmt | 1910000 | - 100000 - 5.236% 1810000
core::fmt::num:👿:<impl core::fmt::Display for i64>::fmt | 2430000 | - 110000 - 4.527% 2320000
core::fmt::num:👿:<impl core::fmt::Display for i8>::fmt | 1080000 | - 170000 - 15.741% 910000
core::fmt::num:👿:<impl core::fmt::Display for u16>::fmt | 960000 | + 10000 + 1.042% 970000
core::fmt::num:👿:<impl core::fmt::Display for u32>::fmt | 1300000 | + 30000 + 2.308% 1330000
core::fmt::num:👿:<impl core::fmt::Display for u8>::fmt | 820000 | - 30000 - 3.659% 790000
```
<details>
<summary>Source code</summary>
```rust
#![feature(test)]
extern crate test;
use std::io::{stdout, Write};
use std::io::StdoutLock;
use test::black_box;
macro_rules! benches {
($handle:ident, $buf:ident, $($name:ident($value:expr))*) => {
$(
fn $name(handle: &mut StdoutLock, buf: &mut Vec<u8>) {
for _ in 0..10000 {
buf.clear();
write!(buf, "{}", black_box($value)).unwrap();
handle.write_all(buf);
}
}
$name(&mut $handle, &mut $buf);
)*
}
}
fn main() {
let mut handle = stdout().lock();
let mut buf = Vec::with_capacity(40);
benches! {
handle, buf,
bench_u64_0(0u64)
bench_u64_half(u32::max_value() as u64)
bench_u64_max(u64::max_value())
bench_i64_0(0i64)
bench_i64_half(i32::max_value() as i64)
bench_i64_max(i64::max_value())
bench_u16_0(0u16)
bench_u16_min(u16::min_value())
bench_u16_max(u16::max_value())
bench_i16_0(0i16)
bench_i16_min(i16::min_value())
bench_i16_max(i16::max_value())
bench_u128_0(0u128)
bench_u128_max(u128::max_value())
bench_i8_0(0i8)
bench_i8_min(i8::min_value())
bench_i8_max(i8::max_value())
bench_u8_0(0u8)
bench_u8_min(u8::min_value())
bench_u8_max(u8::max_value())
bench_i32_0(0i32)
bench_i32_min(i32::min_value())
bench_i32_max(i32::max_value())
bench_u32_0(0u32)
bench_u32_min(u32::min_value())
bench_u32_max(u32::max_value())
}
}
```
</details>
The next step would be to specialize the `ToString` implementation so it doesn't go through the `Display` trait. I'm not sure if it will improve anything but I think it's worth a try.
r? `@Amanieu`
|
||
|---|---|---|
| .. | ||
| benches | ||
| src | ||
| tests | ||
| Cargo.toml | ||