Pads and aligns a value to the length of a cache line.

In concurrent programming, sometimes it is desirable to make sure commonly accessed pieces of data are not placed into the same cache line. Updating an atomic value invalidates the whole cache line it belongs to, which makes the next access to the same cache line slower for other CPU cores. Use CachePadded to ensure updating one piece of data doesn't invalidate other cached data.

Size and alignment

Cache lines are assumed to be N bytes long, depending on the architecture:

• On x86-64, aarch64, and powerpc64, N = 128.
• On arm, mips, mips64, sparc, and hexagon, N = 32.
• On m68k, N = 16.
• On s390x, N = 256.
• On all others, N = 64.

Note that N is just a reasonable guess and is not guaranteed to match the actual cache line length of the machine the program is running on. On modern Intel architectures, spatial prefetcher is pulling pairs of 64-byte cache lines at a time, so we pessimistically assume that cache lines are 128 bytes long.

The size of CachePadded<T> is the smallest multiple of N bytes large enough to accommodate a value of type T.

The alignment of CachePadded<T> is the maximum of N bytes and the alignment of T.

Layout

Since crossbeam-utils 0.8.22, this type is #[repr(C)] and is guaranteed that the pointer to CachePadded<T> has the same address as the pointer to the underlying T.

Examples

Alignment and padding:

use crossbeam_utils::CachePadded;

let array = [CachePadded::new(1i8), CachePadded::new(2i8)];
let addr1 = &*array[0] as *const i8 as usize;
let addr2 = &*array[1] as *const i8 as usize;

assert!(addr2 - addr1 >= 32);
assert_eq!(addr1 % 32, 0);
assert_eq!(addr2 % 32, 0);

When building a concurrent queue with a head and a tail index, it is wise to place them in different cache lines so that concurrent threads pushing and popping elements don't invalidate each other's cache lines:

use crossbeam_utils::CachePadded;
use std::sync::atomic::AtomicUsize;

struct Queue<T> {
    head: CachePadded<AtomicUsize>,
    tail: CachePadded<AtomicUsize>,
    buffer: *mut T,
}