MINOR: tools: improve ptr hash distribution on 64 bits

When testing the pointer hash on 64-bit real pointers (map entries),
it appeared that the shift by 33 bits that hoped to compensate for the
3 nul LSB degrades the hash, and the centering is more optimal on
31-(bits+1)/2. This makes sense since the topmost bit of the
multiplicator is 31, so for an input of 1 bit and 1 bit of output we
would always get zero. With the formula adjusted this way, we can get
up to ~15% more unique entries at 10 bits and ~24% more at 11 bits.
This commit is contained in:
Willy Tarreau 2023-08-11 20:11:30 +02:00
parent ab6cb5dea0
commit 58946d44f8

View File

@ -1085,12 +1085,13 @@ static inline uint statistical_prng_range(uint range)
* to compute statistic buckets, in that it's fast and reasonably distributed * to compute statistic buckets, in that it's fast and reasonably distributed
* thanks to mixing the bits via a multiplication by a prime number and using * thanks to mixing the bits via a multiplication by a prime number and using
* the middle bits on 64-bit platforms or remixing the topmost with lowest ones * the middle bits on 64-bit platforms or remixing the topmost with lowest ones
* on 32-bit. It provides ~2588 unique values (~1510 non-colliding) at 100% * on 32-bit. It provides ~2491 unique values (~1354 non-colliding) for 2^12
* fill ratio for 12 bits, ~1296 (~756 non-colliding) at 100% fill ratio for 11 * valid pointers at 12 bits, ~1454 (~941 non-colliding) for 2^11 valid ptrs
* bits, ~648 (~378 non-colliding) at 100% fill ratio for 10 bits, ~163 (95 non * at 11 bits, ~707 (~434 non-colliding) for 2^10 valid ptrs at 10 bits, ~347
* colliding) at 100% fill ratio for 8 bits, hence 1-1/e and 1/e respectively. * (212 non colliding) for 2^9 valid ptrs at 9 bits, and 165/99 for 2^8 ptrs
* It must be inlined so that <bits> is always a compile-time constant. It * at 8 bits, hence 1-1/e and 1/e respectively. It must be inlined so that
* supports output sizes from 0 to 32 bits. * <bits> is always a compile-time constant. It supports output sizes from 0
* to 32 bits.
*/ */
static forceinline uint ptr_hash(const void *p, const int bits) static forceinline uint ptr_hash(const void *p, const int bits)
{ {
@ -1103,7 +1104,7 @@ static forceinline uint ptr_hash(const void *p, const int bits)
if (sizeof(long) == 4) if (sizeof(long) == 4)
x ^= x >> 32; x ^= x >> 32;
else else
x >>= 33 - bits / 2; x >>= 31 - (bits + 1) / 2;
return x & (~0U >> (-bits & 31)); return x & (~0U >> (-bits & 31));
} }