Primitives specialization for Hasher about gxhash HOT 4 CLOSED

It seems the default write (non-specialized) wasn't that bad even on the smallest primitive types. On MacBook M1 pro, using write_u32 yields a +13% performance (which is still substantial).

Another interesting thing is that the hashset benchmark was biased in some cases. black_boxing the keys prevents compiler optimizations that made this bench biased.

from gxhash.

Current progress involves hashes that are stable in the context of the Hasher, however hashes for an u32 hashed via Hasher::write_u32 are not stable with hashes using the gxhash(&[u8], ...) method. I think this is acceptable because those are two very different contexts. SMHasher should still pass for both contexes.

from gxhash.

Fixed a SIGSEGV when passed [u8] is a null slice (not just an empty slice)

from gxhash.

Merging and releasing 2.3.0

On both my ARM and X86 platforms, I get about -13% of hashing time for small inputs (u8, u16, u32, u64, u128 and signed counterparts). On my ARM PC, gxhash Hasher is now faster than ahash for such inputs. My on X86 PC, gxhash remain a bit slower for these inputs (about 10% slower). I have a doubt in ahash Hasher passing SMHasher quality test for such inputs.

from gxhash.