Prototype code for golang/go #34381

1. Perfect hash function

Using FNV (variant 1a for better avalanche properties).

1. Use random seed
2. Hash string length (as one byte);
3. If there are multiple same-length strings: find minimal unique prefix of the case strings.
4. If there are collisions, try another seed.

Use 32-bit hash, to fit the hash sum in a word on 32-bit architectures. Could be extended with 64-bit implementation if needed.

2. Minimal perfect hash function (for jump table)

The jump table index is calculated in the following manner, inspired by [0], [1]. The hash function used is the FNV-based perfect hash function mentioned above.

For N pre-defined keys (strings):

1. Define jump table size m: the smallest power of 2 greater than N

2. Assign the keys to buckets: the number of buckets k is the smallest power of 2 greater than N/3. Each key is assigned to bucket hash(key) mod k

3. Each bucket gets a shift value so that all keys in that bucket get a unique jump table index that doesn't collide with any other key:

    sum = hash(key)
    shift = bucketShifts[sum mod k]
    sum' = sum >> shift
    jump table index = (sum' xor sum) mod m
   

4. If we cannot find a suitable shift for some bucket, try a different seed for the hash function.

The number of buckets and jump table size are powers of two so we can use bitmasks instead of the modulo operator. The number of buckets is based on the results presented in [0].

References:

[0] F. C. Botelho, D. Belazzougui and M. Dietzfelbinger. Compress, hash and displace. In Proceedings of the 17th European Symposium on Algorithms (ESA 2009). Springer LNCS, 2009. http://cmph.sourceforge.net/papers/esa09.pdf

[1] Bob Jenkins: Minimal Perfect Hashing, http://www.burtleburtle.net/bob/hash/perfect.html#algo