Avoid OverflowError in mpz.__rshift__()

[skirpichev]

Closes #524

[skirpichev]

@casevh, second patch removes scripts/mp_bitcnt_t.diff, thus I'm marking this as a draft.

It's possible to implement this pr in other way, but I think it triggers a bug in the patch. I think that proper support for 64-bit mp_bitcnt_t on win64 1) requires a more work, 2) we should submit patch to the upstream first.

[casevh]

What bug is triggered in the mp_bitcnt_t patch? A previous bug that I reported upstream was fixed within a few days.

[skirpichev]

What bug is triggered in the mp_bitcnt_t patch?

See failed job: https://github.com/aleaxit/gmpy/actions/runs/11379818079/job/31657941881

It's on Windows, I can't reproduce this. I guess, mpz_fdiv_q_2exp() make some assumptions on max value of the last argument.

[skirpichev]

@casevh, I think failure with mp_bitcnt_t.diff happens, because mpz_fdiv_q_2exp() uses mpn_rshift() to do divisions. This will not work with fourth argument of mp_bitcnt_t type, because it's declared as

mp_limb_t mpn_rshift (mp_ptr, mp_srcptr, mp_size_t, unsigned int);

I don't see an easy fix.

What do you think?

[casevh]

Per the GMP documentation, (the fourth argument) must be in the range 1 to mp_bits_per_limb{}-1. So I think there is a bug somewhere else. I'll try to step through the logic in cfdiv_q_2exp.

[skirpichev]

FYI, failed test: assert (mpz(123) >> 111111111111111111111) == mpz(0)