What my above code does is speed up a part of the primality test when you are in the long long domain for the number under test. so every long long tested will benefit from that.
Testing ..65535 will reject a lot of numbers much faster, but not all that's true.
How much?:
for every prime long long found there will be 21845 faster tests (1/3 of 65535 using the 6k+-1 method)
savings per calculation in this range is 414-126 uSec = 288 usec
21845 * 288 usec = 6.291.360 usec or about 6.2 seconds per prime found. (>20% of the 27 seconds)
As there is also a gain - far more difficult to calculate - for all the non primes.
I expect this to be at least 10%, probably substantial more, as most non primes are filtered by the lower numbers.
but the proof is in the pudding test.