add t1ha hash (fastest, pass all tests)

[erthink]

ready = https://github.com/leo-yuriev/t1ha/tree/smhasher-rurban.t1ha

[rurban]

Thanks, looks good

[rurban]

Merged, thanks. I just had no room for your README.md docs. Maybe put them into the header.

[erthink]

Sorry, but I just updated t1ha.
This is a fairly important change.

In general, I was switched to little bit faster 'mux' version (with 64x64 mul) and to C from C++.
But also fixes a possibility of crossing page boundary at the end of key, in case it was unaligned.
This is not a flaw in case using t1ha in the 1Hippeus project, but seriously bug for general hash function.

Please merge again.

[rurban]

Can you add the t1ha_mux variant also?
This is now missing from your smhasher-rurban.t1ha branch

[erthink]

It was decided to leave only the "mux" variant, at the same time renaming it into "t1ha".
Maybe it's not the best solution, but in general should be better.

I think, I should clarify:

• The 'mux' version (which now called as main "t1ha") is faster for small keys, but same speed for large keys.
• The old version (originally called 1Hippeus Hash) was dropped, because it should be fixed/refined against flaw that I noted.
• This should not be mislead about "t1ha" name, because it just published as "Positive Hash" and it have not been used before.

https://github.com/leo-yuriev/t1ha/blob/smhasher-rurban.t1ha/main.cpp#L184

[rurban]

But this would not work on 32bit then. No plan for a 32bit version?

[erthink]

It should already work. More over, much faster than mum-hash.

I implemented the required multiplication for the absence of __int128.

https://github.com/leo-yuriev/t1ha/blob/smhasher-rurban.t1ha/t1ha.c#L149

[rurban]

Great, updating and testing it now

[rurban]

This looks like a bug:

add_with_carry(uint64_t *sum, uint64_t addend) {
  *sum += addend;
  return sum < addend;
}

=>
return *sum < addend;

and it is also invalid C++ with -fpermissive

[erthink]

Oops. Sure this is bug. Just lost de-referense while conversion from C++ to C.

It is question for QA. How it passes the tests?

[rurban]

I was QA. I didn't pass :)

in smhasher it just added randomly a 0 or 1 to it, which apparently qualifies as good enough.

on 32bit some constants are also too long and are capped. e.g. 2166136261 would need a 2166136261L there. but this is in a different function.

[erthink]

;)

+1

I am on the road approximately two hours.

[erthink]

Thank!