> I have optimized XOR and now it's even faster than OpenSSL's code for
> reasonable long passwords. :(

Uh... and it gives the right results? Perhaps OpenSSL should be reimplemented in Python :-)

If your code is resistant to timing attacks, I suppose we don't need the C implementation anymore? Or does it have a larger startup cost?
(what is a "reasonably long password" for you? 10 chars or 64? :-))