Eric Biggers wrote:
> Good news: I tried some different constants for multiplicative hashing, and
> the
> results for two constants were as good as the CRC-based hash function, and
> faster to compute (at least on x86). So here's the revised patch that does
> away
> with static data completely.
Good indeed. Slightly better results than the previous
version.
See comments below.
Jean-Pierre
>
> ---------------
>
> diff --git a/libntfs-3g/compress.c b/libntfs-3g/compress.c
> index 73ad283..b356e35 100644
> --- a/libntfs-3g/compress.c
> +++ b/libntfs-3g/compress.c
[...]
>
> /*
> + * Hash the next 3-byte sequence in the input buffer
> + */
> +static inline unsigned int ntfs_hash(const u8 *p)
> +{
> + u32 str;
> +
> +#if defined(__i386__) || defined(__x86_64__)
> + /* Unaligned access okay */
Also warn about only valid on a little endian cpu.
> + str = *(u32 *)p & 0xFFFFFF;
Please cast to "const u32*" to avoid a compiler warning
about dropping a const qualifier.
> +#else
> + str = ((u32)p[0] << 0) | ((u32)p[1] << 8) | ((u32)p[2] << 16);
> +#endif
> +
> + return (str * HASH_MULTIPLIER) >> (32 - HASH_SHIFT);
This is assuming "int" is 32-bit. Please mask the result
with ((1 << HASH_SHIFT) - 1) for safety.
I have checked that gcc recognizes this as unneeded,
so there is no drop in performance with smart compilers.
[...]
>
> +
> + /* Search the appropriate hash chain for matches. */
> +
> + for (; cur_match >= 0 && depth_remaining--; cur_match =
> prev[cur_match]) {
Please split this line (and other ones which are not
within the 80 chars limit.)
[...]
------------------------------------------------------------------------------
_______________________________________________
ntfs-3g-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/ntfs-3g-devel
