Oh -- this has a name? And it is not commonly known to be better than modulo? Forgotten? I feel old.

The reason multiply+use of high bits is better than modulo is not the golden ratio. In fact, any number using a lot of bits is good (well, not all are the same, like 0 is extremely bad, and 1 and all powers of two are quite bad, because they just shift. But OTOH, the golden ratio is not outstandingly good either -- many values are good.). This is because in the result of a multiplication, higher bits depend on all lower-bits of the operands, because the carry propagates from low bits to higher bits through the resulting value. In contrast, modulo essentially cuts off the upper bits, and so the upper bits are lost for the distribution entropy. This means that if you take the high bits of a multiplication result to distribute into a small set of integers, then you get a better distribution than using only the lower bits, because more bits of the input value are significant for the output result.

I thought this was common knowledge. The man page on rand() used to tell you that you should always multiply into the target range instead of modulo into the target range. It is surprising to see that all big hash libraries seem to miss this.