mp_size_t
mpn_binvert_itch (mp_size_t n)
{
#if WANT_FFT
if (ABOVE_THRESHOLD (n, 2 * MUL_FFT_MODF_THRESHOLD))
return mpn_fft_next_size (n, mpn_fft_best_k (n, 0));
else
#endif
return 3 * (n - (n >> 1));
}
void
mpn_binvert (mp_ptr rp, mp_srcptr up, mp_size_t n, mp_ptr scratch)
{
mp_ptr xp;
mp_size_t rn, newrn;
mp_size_t sizes[NPOWS], *sizp;
mp_limb_t di;
/* Compute the computation precisions from highest to lowest, leaving the
base case size in 'rn'. */
sizp = sizes;
for (rn = n; ABOVE_THRESHOLD (rn, BINV_NEWTON_THRESHOLD); rn = (rn + 1) >> 1)
*sizp++ = rn;
xp = scratch;
/* Compute a base value using a low-overhead O(n^2) algorithm. FIXME: We
should call some divide-and-conquer lsb division function here for an
operand subrange. */
MPN_ZERO (xp, rn);
xp[0] = 1;
binvert_limb (di, up[0]);
if (BELOW_THRESHOLD (rn, DC_BDIV_Q_THRESHOLD))
mpn_sb_bdiv_q (rp, xp, rn, up, rn, -di);
else
mpn_dc_bdiv_q (rp, xp, rn, up, rn, -di);
/* Use Newton iterations to get the desired precision. */
for (; rn < n; rn = newrn)
{
newrn = *--sizp;
#if WANT_FFT
if (ABOVE_THRESHOLD (newrn, 2 * MUL_FFT_MODF_THRESHOLD))
{
int k;
mp_size_t m, i;
k = mpn_fft_best_k (newrn, 0);
m = mpn_fft_next_size (newrn, k);
mpn_mul_fft (xp, m, up, newrn, rp, rn, k);
for (i = rn - 1; i >= 0; i--)
if (xp[i] > (i == 0))
{
mpn_add_1 (xp + rn, xp + rn, newrn - rn, 1);
break;
}
}
else
#endif
mpn_mul (xp, up, newrn, rp, rn);
mpn_mullow_n (rp + rn, rp, xp + rn, newrn - rn);
mpn_neg_n (rp + rn, rp + rn, newrn - rn);
}
}
mp_size_t
mpn_mulmod_bnm1_next_size (mp_size_t n)
{
mp_size_t nh;
if (BELOW_THRESHOLD (n, MULMOD_BNM1_THRESHOLD))
return n;
if (BELOW_THRESHOLD (n, 4 * (MULMOD_BNM1_THRESHOLD - 1) + 1))
return (n + (2-1)) & (-2);
if (BELOW_THRESHOLD (n, 8 * (MULMOD_BNM1_THRESHOLD - 1) + 1))
return (n + (4-1)) & (-4);
nh = (n + 1) >> 1;
if (BELOW_THRESHOLD (nh, MUL_FFT_MODF_THRESHOLD))
return (n + (8-1)) & (-8);
return 2 * mpn_fft_next_size (nh, mpn_fft_best_k (nh, 0));
}
