Hash for Matrix doesn't match Borrow<[[T; R]; C]> semantics

[CAD97]

Borrow<T> for Self implies that trait implementations on Self and T behave identically, in notable particular Eq, Ord, and Hash. The Hash implementation for Matrix is different from the one for arrays, resulting in e.g. the hash of Matrix3<i32> differing from that of the same matrix Borrow::borrowed as [[i32; 3]; 3]. [playground]

use nalgebra; // 0.32.3

use std::borrow::Borrow;
use std::hash::{BuildHasher, Hash};
use std::collections::hash_map::RandomState;

fn main() {
    let s = RandomState::new();
    let v = nalgebra::matrix![3, 1, 4; 1, 5, 9; 2, 6, 5];
    assert_eq!(s.hash_one(&v), s.hash_one(<_ as Borrow<[_; 3]>>::borrow(&v))); // fails, should pass
}

Either the Borrow implementation should be removed or the Hash implementation should be changed to try to match the array Hash implementation. I think the only way to do so properly¹ would be to hash the data field directly; ArrayStorage has the properly equivalent Hash implementation, so it's a matter of requiring all impl RawStorage to provide a Hash impl. I think the way to do so in a non-API-breaking manner would be to add a copy of the Hash::hash method into the RawStorage trait and use that, as a default of the current behavior can be kept which ArrayStorage uses its derived Hash implementation. (This will result in hashing slightly more data, since R gets prefixed to each column individually, but that should be marginal for ArrayStorage-backed Matrix.)

Footnotes

1. The presence of Hash::hash_slice as a customization point means any and all attempts to mirror the Hash behavior will fail in an edge case, short of constructing &[[T; R]] or &[&[T]] (which would require allocation in the general case). ↩