Remove super cluster stuff from TT and just use a 128 bit multiply.

[mstembera]

Remove super cluster stuff from TT and just use a 128 bit multiply.

STC https://tests.stockfishchess.org/tests/view/5ee719b3aae8aec816ab7548
LLR: 2.94 (-2.94,2.94) {-1.50,0.50}
Total: 12736 W: 2502 L: 2333 D: 7901
Ptnml(0-2): 191, 1452, 2944, 1559, 222

LTC https://tests.stockfishchess.org/tests/view/5ee732d1aae8aec816ab7556
LLR: 2.93 (-2.94,2.94) {-1.50,0.50}
Total: 27584 W: 3431 L: 3350 D: 20803
Ptnml(0-2): 173, 2500, 8400, 2511, 208

Scheme back to being derived from https://lemire.me/blog/2016/06/27/a-fast-alternative-to-the-modulo-reduction/ pointed out by @vondele

Also the default optimized version of the index calculation now uses fewer instructions.
https://godbolt.org/z/Tktxbv

The reason MaxHashMB is currently capped to 1073741824 MB is because it's the largest power of two that can be stored in a 32bit int which is what our Options use.

Related PR #2722 #1349

bench: 4320954

[protonspring]

Could we use (1 << 30) instead of 1073741824 ? Then you wouldn't have to explain the value to less experienced developers.

[mstembera]

@protonspring I think that's reasonable as well. Whatever the maintainers prefer.

[vondele]

actually, as soon as one adds bmi2, the compiler will emit mulx instead of mul:
https://godbolt.org/z/MhDHtH
That's presumably what most of noob's machines do. I would assume there is a performance benefit for this new instruction (but couldn't quite figure out). On AMD Zen, we suggest that people stick to 'modern' which doesn't use bmi2, because pext is slower, so they might have a somewhat suboptimal hash access, unless they enable bmi2 but not pext. I don't know how much the performance difference would be.
Edit: I can't measure a difference between -mbmi2 and not on Zen.

concerning the maximum size, I would leave it at the current 32TB, that's still more than current processors support.

[mstembera]

Ok max hash reverted back to 32TB.

[vondele]

Thanks!

[mstembera]

Just for completeness if we ever want to optimize for MSVC the code is:
uint64_t highProduct;
_umul128(a, b, &highProduct);
return highProduct;