Problem with gmp_randinit_set

[Niels Möller]

"Marco Bodrato" <bodrato at mail.dm.unipi.it> writes:

> The problem is that Niels' code, mine, and the current mpz code... do all
> "reduce" modulo (2^19937-20023) obtaining some non-canonical
> representation.

I was thinking that one should convert to canonical representation at
the end of the powering. But if the current code doesn't do that,
staying compatible would need extra care. And it rules out the simple
way of just using mpn_powm.

(And then I realized that one has to produce a canonical representation
also for the initial reduction, since that uses a different modulo, p-4).

> If we "do not want to generate different sequences than earlier GMP", we
> will have to mimic current behaviour, even in the corner cases...

I think that should be doable, if we want to. Without thinking too
deeply about it, it looks like your variant with shift should be
equivalent to the current mpz code, maybe mine too (it essentially
combines the shift with the addmul_1 call).

> Is it worth doing?

I think it's desirable to eliminate the dependency on mpz. How important
it is not change the seed --> sequence mapping, I don't know.

Regards,
/Niels

-- 
Niels Möller. PGP-encrypted email is preferred. Keyid 368C6677.
Internet email is subject to wholesale government surveillance.