optimize Divider.mixedness

[niklasf]

Before:

[info] PlayBench.divider  thrpt   45  21.868 ± 1.357  ops/s

After, first attempt:

[info] PlayBench.divider  thrpt   45  33.767 ± 1.497  ops/s

Additionally, precomputing mixedness regions again:

[info] PlayBench.divider  thrpt   45  51.566 ± 4.220  ops/s

Additionally, using mapReduce:

[info] PlayBench.divider  thrpt   45  65.580 ± 3.128  ops/s

Additionally, using Long instead of Bitboard region masks:

[info] PlayBench.divider  thrpt   45  91.231 ± 5.867  ops/s

[lenguyenthanh]

This is really nice, thanks Niklas. I was thinking to do the same but too scared. Just added small suggestion, otherwise LGTM.

[niklasf]

Switching from Bitboard to Long for the region masks is (to me) a surprisingly large win:

[info] PlayBench.divider  thrpt   45  91.231 ± 5.867  ops/s

[lenguyenthanh]

I didn't expect that big of improvement either. It makes me wondering how much we can gain if we don't use Bitboard abstraction.

[lenguyenthanh]

I guess the performance penalty is mostly caused by TotalWrapper because with it, Bitboard is be boxed object , not just a primitive Long anymore.

but maybe the opaque type has some responsibility as well.