6

Assuming I wanted to (ab)use cache coherency to do lock free reads like FaRM, would it be enough to have a struct with a single 64 byte array as data to guarantee that on an architecture with 64 byte cache lines each struct would occupy exactly one cache line?

Shepmaster
  • 388,571
  • 95
  • 1,107
  • 1,366
benjumanji
  • 1,395
  • 7
  • 15

1 Answers1

9

No, that wouldn't guarantee that the alignment was a cache line.

RFC 1358 added the concept of #[repr(align)], allowing the programmer to specify alignment requirements. This attribute was stabilized in Rust 1.25.

For your specific case, you'd use it like:

#[repr(align(64))]
struct Foo {
    value: u8,
}
Shepmaster
  • 388,571
  • 95
  • 1,107
  • 1,366
Chris Emerson
  • 13,041
  • 3
  • 44
  • 66
  • I am confusing the size of the array with the size of the type. I guess there aren't any 64 byte simd types. – benjumanji Dec 11 '16 at 19:41
  • There's [AVX512](https://en.m.wikipedia.org/wiki/AVX-512). Not sure if Rust supports it yet. – Chris Emerson Dec 11 '16 at 19:45
  • @ChrisEmerson: Rust doesn't have SIMD types yet. It's a big topic of discussions, lots of people are waiting for them (and the instructions that go with them) but there are many different needs/expectations/wants so it's taking a long time to converge... – Matthieu M. Dec 12 '16 at 09:55
  • @MatthieuM. Yes, I'm following the discussions. :-) But there is [some unstable `#[repr(simd)]`](http://huonw.github.io/blog/2015/08/simd-in-rust/) support on nightlies, as far as I am aware. – Chris Emerson Dec 12 '16 at 09:59
  • @ChrisEmerson: To be honest, though, I'm not quite clear on the exact effect of applying `#[repr(simd)]`. Does it somehow use the size as alignment? – Matthieu M. Dec 12 '16 at 10:15
  • @MatthieuM. I think so, but that's only an assumption; I haven't checked. – Chris Emerson Dec 12 '16 at 10:17