flint_mpn_divexact

Single-limb divisors are handled by mpn_divexact_1. Otherwise the division is done 2-adically: writing $b = 2^v B^k b'$ with $b'$ odd, the quotient is $(a / (2^v B^k))\, b'^{-1} \bmod B^m$, where $m$ is the number of quotient limbs ($\mathit{an} - \mathit{bn} + 1$, or one less when the top limb of $a$ is below that of $b$). With $d = \min(\mathit{bn}, m)$ significant divisor limbs, $d \le 7$ uses register-resident code specialised for each $d$, with the shift done on the fly and the limb inverse $b'^{-1} \bmod B$ computed inline. Larger $d$ use ports of GMP's schoolbook and divide and conquer Hensel division (mpn_sbpi1_bdiv_q, and mpn_dcpi1_bdiv_q from FLINT_MPN_DC_BDIV_Q_CUTOFF limbs) with FLINT's multiplication, and the Newton iteration flint_mpn_bdiv_q from $d \ge$ FLINT_MPN_DIVEXACT_NEWTON_CUTOFF, or already from $\mathit{bn} \ge$ FLINT_MPN_DIVEXACT_UNBALANCED_CUTOFF when $m \ge 4 \mathit{bn}$.