Re: [Qemu-devel] [PATCH 06/10] include: add xxhash.h

[Alex Bennée]

Emilio G. Cota <address@hidden> writes:

> xxhash is a fast, high-quality hashing function. The appended
> brings in the 32-bit version of it, with the small modification that
> it assumes the data to be hashed is made of 32-bit chunks; this increases
> speed slightly for the use-case we care about, i.e. tb-hash.
>
> The original algorithm, as well as a 64-bit implementation, can be found at:
>   https://github.com/Cyan4973/xxHash
>
> Signed-off-by: Emilio G. Cota <address@hidden>
> ---
>  include/qemu/xxhash.h | 106 
> ++++++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 106 insertions(+)
>  create mode 100644 include/qemu/xxhash.h
>
> diff --git a/include/qemu/xxhash.h b/include/qemu/xxhash.h
> new file mode 100644
> index 0000000..a13a665
> --- /dev/null
> +++ b/include/qemu/xxhash.h
> @@ -0,0 +1,106 @@
<snip>
> +
> +/* u32 hash of @n contiguous chunks of u32's */
> +static inline uint32_t qemu_xxh32(const uint32_t *p, size_t n, uint32_t seed)
> +{

What is the point of seed here? I looked on the original site to see if
there was any guidance on tuning seed but couldn't find anything. I
appreciate the compiler can inline the constant away but perhaps we
should #define it and drop the parameter if we are not intending to
modify it?

Also it might be helpful to wrap the call to avoid getting the
boilerplate sizing wrong:

  #define qemu_xxh32(s) qemu_xxh32_impl((const uint32_t *)s, 
sizeof(*s)/sizeof(uint32_t), 1)

Then calls become a little simpler for the user:

  return qemu_xxh32(&k);

Do we need to include a compile time check for structures that don't
neatly divide into uint32_t chunks?

> +    const uint32_t *end = p + n;
> +    uint32_t h32;
> +
> +    if (n >= 4) {
> +        const uint32_t * const limit = end - 4;
> +        uint32_t v1 = seed + PRIME32_1 + PRIME32_2;
> +        uint32_t v2 = seed + PRIME32_2;
> +        uint32_t v3 = seed + 0;
> +        uint32_t v4 = seed - PRIME32_1;
> +
> +        do {
> +            v1 += *p * PRIME32_2;
> +            v1 = XXH_rotl32(v1, 13);
> +            v1 *= PRIME32_1;
> +            p++;
> +            v2 += *p * PRIME32_2;
> +            v2 = XXH_rotl32(v2, 13);
> +            v2 *= PRIME32_1;
> +            p++;
> +            v3 += *p * PRIME32_2;
> +            v3 = XXH_rotl32(v3, 13);
> +            v3 *= PRIME32_1;
> +            p++;
> +            v4 += *p * PRIME32_2;
> +            v4 = XXH_rotl32(v4, 13);
> +            v4 *= PRIME32_1;
> +            p++;
> +        } while (p <= limit);
> +        h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7) + XXH_rotl32(v3, 12) +
> +            XXH_rotl32(v4, 18);
> +    } else {
> +        h32  = seed + PRIME32_5;
> +    }

I don't plead any particular knowledge of hashing codes but I note the
test cases we add only exercise the n == 1 path but in actual usage we
exercise n = 5 (at least for arm32, I guess aarch64 would be more).

> +
> +    h32 += n * sizeof(uint32_t);
> +
> +    while (p < end) {
> +        h32 += *p * PRIME32_3;
> +        h32  = XXH_rotl32(h32, 17) * PRIME32_4 ;
> +        p++;
> +    }
> +
> +    h32 ^= h32 >> 15;
> +    h32 *= PRIME32_2;
> +    h32 ^= h32 >> 13;
> +    h32 *= PRIME32_3;
> +    h32 ^= h32 >> 16;
> +
> +    return h32;
> +}
> +
> +#endif /* QEMU_XXHASH_H */


--
Alex Bennée