Auto merge of #128204 - GuillaumeGomez:integers-opti, r=workingjubilee

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::imp::<impl core::fmt::Display for i16>::fmt |   1300000 | -    70000 -  5.385%   1230000
core::fmt::num::imp::<impl core::fmt::Display for i32>::fmt |   1910000 | -   100000 -  5.236%   1810000
core::fmt::num::imp::<impl core::fmt::Display for i64>::fmt |   2430000 | -   110000 -  4.527%   2320000
core::fmt::num::imp::<impl core::fmt::Display for i8>::fmt  |   1080000 | -   170000 - 15.741%    910000
core::fmt::num::imp::<impl core::fmt::Display for u16>::fmt |    960000 | +    10000 +  1.042%    970000
core::fmt::num::imp::<impl core::fmt::Display for u32>::fmt |   1300000 | +    30000 +  2.308%   1330000
core::fmt::num::imp::<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`

Author: bors