| 1 |
#![feature(test, maybe_uninit_uninit_array_transpose)] |
| 2 |
extern crate test; |
| 3 |
|
| 4 |
use std::{ |
| 5 |
mem::{size_of, MaybeUninit}, |
| 6 |
slice, |
| 7 |
}; |
| 8 |
|
| 9 |
// Call getrandom on a zero-initialized stack buffer |
| 10 |
#[inline(always)] |
| 11 |
fn bench_fill<const N: usize>() { |
| 12 |
let mut buf = [0u8; N]; |
| 13 |
getrandom::fill(&mut buf).unwrap(); |
| 14 |
test::black_box(&buf[..]); |
| 15 |
} |
| 16 |
|
| 17 |
// Call fill_uninit on an uninitialized stack buffer |
| 18 |
#[inline(always)] |
| 19 |
fn bench_fill_uninit<const N: usize>() { |
| 20 |
let mut uninit = [MaybeUninit::uninit(); N]; |
| 21 |
let buf: &[u8] = getrandom::fill_uninit(&mut uninit).unwrap(); |
| 22 |
test::black_box(buf); |
| 23 |
} |
| 24 |
|
| 25 |
#[bench] |
| 26 |
pub fn bench_u32(b: &mut test::Bencher) { |
| 27 |
#[inline(never)] |
| 28 |
fn inner() -> u32 { |
| 29 |
getrandom::u32().unwrap() |
| 30 |
} |
| 31 |
b.bytes = 4; |
| 32 |
b.iter(inner); |
| 33 |
} |
| 34 |
#[bench] |
| 35 |
pub fn bench_u32_via_fill(b: &mut test::Bencher) { |
| 36 |
#[inline(never)] |
| 37 |
fn inner() -> u32 { |
| 38 |
let mut res = MaybeUninit::<u32>::uninit(); |
| 39 |
let dst: &mut [MaybeUninit<u8>] = |
| 40 |
unsafe { slice::from_raw_parts_mut(res.as_mut_ptr().cast(), size_of::<u32>()) }; |
| 41 |
getrandom::fill_uninit(dst).unwrap(); |
| 42 |
unsafe { res.assume_init() } |
| 43 |
} |
| 44 |
b.bytes = 4; |
| 45 |
b.iter(inner); |
| 46 |
} |
| 47 |
|
| 48 |
#[bench] |
| 49 |
pub fn bench_u64(b: &mut test::Bencher) { |
| 50 |
#[inline(never)] |
| 51 |
fn inner() -> u64 { |
| 52 |
getrandom::u64().unwrap() |
| 53 |
} |
| 54 |
b.bytes = 8; |
| 55 |
b.iter(inner); |
| 56 |
} |
| 57 |
|
| 58 |
#[bench] |
| 59 |
pub fn bench_u64_via_fill(b: &mut test::Bencher) { |
| 60 |
#[inline(never)] |
| 61 |
fn inner() -> u64 { |
| 62 |
let mut res = MaybeUninit::<u64>::uninit(); |
| 63 |
let dst: &mut [MaybeUninit<u8>] = |
| 64 |
unsafe { slice::from_raw_parts_mut(res.as_mut_ptr().cast(), size_of::<u64>()) }; |
| 65 |
getrandom::fill_uninit(dst).unwrap(); |
| 66 |
unsafe { res.assume_init() } |
| 67 |
} |
| 68 |
b.bytes = 8; |
| 69 |
b.iter(inner); |
| 70 |
} |
| 71 |
|
| 72 |
// We benchmark using #[inline(never)] "inner" functions for two reasons: |
| 73 |
// - Avoiding inlining reduces a source of variance when running benchmarks. |
| 74 |
// - It is _much_ easier to get the assembly or IR for the inner loop. |
| 75 |
// |
| 76 |
// For example, using cargo-show-asm (https://github.com/pacak/cargo-show-asm), |
| 77 |
// we can get the assembly for a particular benchmark's inner loop by running: |
| 78 |
// cargo asm --bench buffer --release buffer::p384::bench_getrandom::inner |
| 79 |
macro_rules! bench { |
| 80 |
( $name:ident, $size:expr ) => { |
| 81 |
pub mod $name { |
| 82 |
#[bench] |
| 83 |
pub fn bench_fill(b: &mut test::Bencher) { |
| 84 |
#[inline(never)] |
| 85 |
fn inner() { |
| 86 |
super::bench_fill::<{ $size }>() |
| 87 |
} |
| 88 |
|
| 89 |
b.bytes = $size as u64; |
| 90 |
b.iter(inner); |
| 91 |
} |
| 92 |
#[bench] |
| 93 |
pub fn bench_fill_uninit(b: &mut test::Bencher) { |
| 94 |
#[inline(never)] |
| 95 |
fn inner() { |
| 96 |
super::bench_fill_uninit::<{ $size }>() |
| 97 |
} |
| 98 |
|
| 99 |
b.bytes = $size as u64; |
| 100 |
b.iter(inner); |
| 101 |
} |
| 102 |
} |
| 103 |
}; |
| 104 |
} |
| 105 |
|
| 106 |
// 16 bytes (128 bits) is the size of an 128-bit AES key/nonce. |
| 107 |
bench!(aes128, 128 / 8); |
| 108 |
|
| 109 |
// 32 bytes (256 bits) is the seed sized used for rand::thread_rng |
| 110 |
// and the `random` value in a ClientHello/ServerHello for TLS. |
| 111 |
// This is also the size of a 256-bit AES/HMAC/P-256/Curve25519 key |
| 112 |
// and/or nonce. |
| 113 |
bench!(p256, 256 / 8); |
| 114 |
|
| 115 |
// A P-384/HMAC-384 key and/or nonce. |
| 116 |
bench!(p384, 384 / 8); |
| 117 |
|
| 118 |
// Initializing larger buffers is not the primary use case of this library, as |
| 119 |
// this should normally be done by a userspace CSPRNG. However, we have a test |
| 120 |
// here to see the effects of a lower (amortized) syscall overhead. |
| 121 |
bench!(page, 4096); |