mp_limb_t
mpn_dcpi1_divappr_q_n (mp_ptr qp, mp_ptr np, mp_srcptr dp, mp_size_t n,
gmp_pi1_t *dinv, mp_ptr tp)
{
mp_size_t lo, hi;
mp_limb_t cy, qh, ql;
lo = n >> 1; /* floor(n/2) */
hi = n - lo; /* ceil(n/2) */
if (BELOW_THRESHOLD (hi, DC_DIV_QR_THRESHOLD))
qh = mpn_sbpi1_div_qr (qp + lo, np + 2 * lo, 2 * hi, dp + lo, hi, dinv->inv32);
else
qh = mpn_dcpi1_div_qr_n (qp + lo, np + 2 * lo, dp + lo, hi, dinv, tp);
mpn_mul (tp, qp + lo, hi, dp, lo);
cy = mpn_sub_n (np + lo, np + lo, tp, n);
if (qh != 0)
cy += mpn_sub_n (np + n, np + n, dp, lo);
while (cy != 0)
{
qh -= mpn_sub_1 (qp + lo, qp + lo, hi, 1);
cy -= mpn_add_n (np + lo, np + lo, dp, n);
}
if (BELOW_THRESHOLD (lo, DC_DIVAPPR_Q_THRESHOLD))
ql = mpn_sbpi1_divappr_q (qp, np + hi, 2 * lo, dp + hi, lo, dinv->inv32);
else
ql = mpn_dcpi1_divappr_q_n (qp, np + hi, dp + hi, lo, dinv, tp);
if (UNLIKELY (ql != 0))
{
mp_size_t i;
for (i = 0; i < lo; i++)
qp[i] = GMP_NUMB_MASK;
}
return qh;
}
mp_limb_t
mpn_dcpi1_divappr_q (mp_ptr qp, mp_ptr np, mp_size_t nn,
mp_srcptr dp, mp_size_t dn, gmp_pi1_t *dinv)
{
mp_size_t qn;
mp_limb_t qh, cy, qsave;
mp_ptr tp;
TMP_DECL;
TMP_MARK;
ASSERT (dn >= 6);
ASSERT (nn > dn);
ASSERT (dp[dn-1] & GMP_NUMB_HIGHBIT);
qn = nn - dn;
qp += qn;
np += nn;
dp += dn;
if (qn >= dn)
{
qn++; /* pretend we'll need an extra limb */
/* Reduce qn mod dn without division, optimizing small operations. */
do
qn -= dn;
while (qn > dn);
qp -= qn; /* point at low limb of next quotient block */
np -= qn; /* point in the middle of partial remainder */
tp = TMP_SALLOC_LIMBS (dn);
/* Perform the typically smaller block first. */
if (qn == 1)
{
mp_limb_t q, n2, n1, n0, d1, d0;
/* Handle qh up front, for simplicity. */
qh = mpn_cmp (np - dn + 1, dp - dn, dn) >= 0;
if (qh)
ASSERT_NOCARRY (mpn_sub_n (np - dn + 1, np - dn + 1, dp - dn, dn));
/* A single iteration of schoolbook: One 3/2 division,
followed by the bignum update and adjustment. */
n2 = np[0];
n1 = np[-1];
n0 = np[-2];
d1 = dp[-1];
d0 = dp[-2];
ASSERT (n2 < d1 || (n2 == d1 && n1 <= d0));
if (UNLIKELY (n2 == d1) && n1 == d0)
{
q = GMP_NUMB_MASK;
cy = mpn_submul_1 (np - dn, dp - dn, dn, q);
ASSERT (cy == n2);
}
else
{
udiv_qr_3by2 (q, n1, n0, n2, n1, n0, d1, d0, dinv->inv32);
if (dn > 2)
{
mp_limb_t cy, cy1;
cy = mpn_submul_1 (np - dn, dp - dn, dn - 2, q);
cy1 = n0 < cy;
n0 = (n0 - cy) & GMP_NUMB_MASK;
cy = n1 < cy1;
n1 = (n1 - cy1) & GMP_NUMB_MASK;
np[-2] = n0;
if (UNLIKELY (cy != 0))
{
n1 += d1 + mpn_add_n (np - dn, np - dn, dp - dn, dn - 1);
qh -= (q == 0);
q = (q - 1) & GMP_NUMB_MASK;
}
}
else
np[-2] = n0;
np[-1] = n1;
}
qp[0] = q;
}
else
{
if (qn == 2)
qh = mpn_divrem_2 (qp, 0L, np - 2, 4, dp - 2);
else if (BELOW_THRESHOLD (qn, DC_DIV_QR_THRESHOLD))
qh = mpn_sbpi1_div_qr (qp, np - qn, 2 * qn, dp - qn, qn, dinv->inv32);
else
qh = mpn_dcpi1_div_qr_n (qp, np - qn, dp - qn, qn, dinv, tp);
if (qn != dn)
{
if (qn > dn - qn)
mpn_mul (tp, qp, qn, dp - dn, dn - qn);
else
mpn_mul (tp, dp - dn, dn - qn, qp, qn);
cy = mpn_sub_n (np - dn, np - dn, tp, dn);
if (qh != 0)
cy += mpn_sub_n (np - dn + qn, np - dn + qn, dp - dn, dn - qn);
while (cy != 0)
{
qh -= mpn_sub_1 (qp, qp, qn, 1);
cy -= mpn_add_n (np - dn, np - dn, dp - dn, dn);
}
}
}
qn = nn - dn - qn + 1;
while (qn > dn)
{
qp -= dn;
np -= dn;
mpn_dcpi1_div_qr_n (qp, np - dn, dp - dn, dn, dinv, tp);
qn -= dn;
}
/* Since we pretended we'd need an extra quotient limb before, we now
have made sure the code above left just dn-1=qn quotient limbs to
develop. Develop that plus a guard limb. */
qn--;
qp -= qn;
np -= dn;
qsave = qp[qn];
mpn_dcpi1_divappr_q_n (qp, np - dn, dp - dn, dn, dinv, tp);
MPN_COPY_INCR (qp, qp + 1, qn);
qp[qn] = qsave;
}
else /* (qn < dn) */
{
mp_ptr q2p;
#if 0 /* not possible since we demand nn > dn */
if (qn == 0)
{
qh = mpn_cmp (np - dn, dp - dn, dn) >= 0;
if (qh)
mpn_sub_n (np - dn, np - dn, dp - dn, dn);
TMP_FREE;
return qh;
}
#endif
qp -= qn; /* point at low limb of next quotient block */
np -= qn; /* point in the middle of partial remainder */
q2p = TMP_SALLOC_LIMBS (qn + 1);
/* Should we at all check DC_DIVAPPR_Q_THRESHOLD here, or reply on
callers not to be silly? */
if (BELOW_THRESHOLD (qn, DC_DIVAPPR_Q_THRESHOLD))
{
qh = mpn_sbpi1_divappr_q (q2p, np - qn - 2, 2 * (qn + 1),
dp - (qn + 1), qn + 1, dinv->inv32);
}
else
{
/* It is tempting to use qp for recursive scratch and put quotient in
tp, but the recursive scratch needs one limb too many. */
tp = TMP_SALLOC_LIMBS (qn + 1);
qh = mpn_dcpi1_divappr_q_n (q2p, np - qn - 2, dp - (qn + 1), qn + 1, dinv, tp);
}
MPN_COPY (qp, q2p + 1, qn);
}
TMP_FREE;
return qh;
}
int
main (int argc, char **argv)
{
gmp_randstate_ptr rands;
unsigned long maxnbits, maxdbits, nbits, dbits;
mpz_t n, d, q, r, tz, junk;
mp_size_t maxnn, maxdn, nn, dn, clearn, i;
mp_ptr np, dup, dnp, qp, rp, junkp;
mp_limb_t t;
gmp_pi1_t dinv;
long count = COUNT;
mp_ptr scratch;
mp_limb_t ran;
mp_size_t alloc, itch;
mp_limb_t rran0, rran1, qran0, qran1;
TMP_DECL;
if (argc > 1)
{
char *end;
count = strtol (argv[1], &end, 0);
if (*end || count <= 0)
{
fprintf (stderr, "Invalid test count: %s.\n", argv[1]);
return 1;
}
}
maxdbits = MAX_DN;
maxnbits = MAX_NN;
tests_start ();
rands = RANDS;
mpz_init (n);
mpz_init (d);
mpz_init (q);
mpz_init (r);
mpz_init (tz);
mpz_init (junk);
maxnn = maxnbits / GMP_NUMB_BITS + 1;
maxdn = maxdbits / GMP_NUMB_BITS + 1;
TMP_MARK;
qp = TMP_ALLOC_LIMBS (maxnn + 2) + 1;
rp = TMP_ALLOC_LIMBS (maxnn + 2) + 1;
dnp = TMP_ALLOC_LIMBS (maxdn);
alloc = 1;
scratch = __GMP_ALLOCATE_FUNC_LIMBS (alloc);
for (test = -300; test < count; test++)
{
nbits = random_word (rands) % (maxnbits - GMP_NUMB_BITS) + 2 * GMP_NUMB_BITS;
if (test < 0)
dbits = (test + 300) % (nbits - 1) + 1;
else
dbits = random_word (rands) % (nbits - 1) % maxdbits + 1;
#if RAND_UNIFORM
#define RANDFUNC mpz_urandomb
#else
#define RANDFUNC mpz_rrandomb
#endif
do
RANDFUNC (d, rands, dbits);
while (mpz_sgn (d) == 0);
dn = SIZ (d);
dup = PTR (d);
MPN_COPY (dnp, dup, dn);
dnp[dn - 1] |= GMP_NUMB_HIGHBIT;
if (test % 2 == 0)
{
RANDFUNC (n, rands, nbits);
nn = SIZ (n);
ASSERT_ALWAYS (nn >= dn);
}
else
{
do
{
RANDFUNC (q, rands, random_word (rands) % (nbits - dbits + 1));
RANDFUNC (r, rands, random_word (rands) % mpz_sizeinbase (d, 2));
mpz_mul (n, q, d);
mpz_add (n, n, r);
nn = SIZ (n);
}
while (nn > maxnn || nn < dn);
}
ASSERT_ALWAYS (nn <= maxnn);
ASSERT_ALWAYS (dn <= maxdn);
mpz_urandomb (junk, rands, nbits);
junkp = PTR (junk);
np = PTR (n);
mpz_urandomb (tz, rands, 32);
t = mpz_get_ui (tz);
if (t % 17 == 0)
{
dnp[dn - 1] = GMP_NUMB_MAX;
dup[dn - 1] = GMP_NUMB_MAX;
}
switch ((int) t % 16)
{
case 0:
clearn = random_word (rands) % nn;
for (i = clearn; i < nn; i++)
np[i] = 0;
break;
case 1:
mpn_sub_1 (np + nn - dn, dnp, dn, random_word (rands));
break;
case 2:
mpn_add_1 (np + nn - dn, dnp, dn, random_word (rands));
break;
}
if (dn >= 2)
invert_pi1 (dinv, dnp[dn - 1], dnp[dn - 2]);
rran0 = random_word (rands);
rran1 = random_word (rands);
qran0 = random_word (rands);
qran1 = random_word (rands);
qp[-1] = qran0;
qp[nn - dn + 1] = qran1;
rp[-1] = rran0;
ran = random_word (rands);
if ((double) (nn - dn) * dn < 1e5)
{
/* Test mpn_sbpi1_div_qr */
if (dn > 2)
{
MPN_COPY (rp, np, nn);
if (nn > dn)
MPN_COPY (qp, junkp, nn - dn);
qp[nn - dn] = mpn_sbpi1_div_qr (qp, rp, nn, dnp, dn, dinv.inv32);
check_one (qp, rp, np, nn, dnp, dn, "mpn_sbpi1_div_qr", 0);
}
/* Test mpn_sbpi1_divappr_q */
if (dn > 2)
{
MPN_COPY (rp, np, nn);
if (nn > dn)
MPN_COPY (qp, junkp, nn - dn);
qp[nn - dn] = mpn_sbpi1_divappr_q (qp, rp, nn, dnp, dn, dinv.inv32);
check_one (qp, NULL, np, nn, dnp, dn, "mpn_sbpi1_divappr_q", 1);
}
/* Test mpn_sbpi1_div_q */
if (dn > 2)
{
MPN_COPY (rp, np, nn);
if (nn > dn)
MPN_COPY (qp, junkp, nn - dn);
qp[nn - dn] = mpn_sbpi1_div_q (qp, rp, nn, dnp, dn, dinv.inv32);
check_one (qp, NULL, np, nn, dnp, dn, "mpn_sbpi1_div_q", 0);
}
/* Test mpn_sb_div_qr_sec */
itch = 3 * nn + 4;
if (itch + 1 > alloc)
{
scratch = __GMP_REALLOCATE_FUNC_LIMBS (scratch, alloc, itch + 1);
alloc = itch + 1;
}
scratch[itch] = ran;
MPN_COPY (rp, np, nn);
if (nn >= dn)
MPN_COPY (qp, junkp, nn - dn + 1);
mpn_sb_div_qr_sec (qp, rp, nn, dup, dn, scratch);
ASSERT_ALWAYS (ran == scratch[itch]);
check_one (qp, rp, np, nn, dup, dn, "mpn_sb_div_qr_sec", 0);
/* Test mpn_sb_div_r_sec */
itch = nn + 2 * dn + 2;
if (itch + 1 > alloc)
{
scratch = __GMP_REALLOCATE_FUNC_LIMBS (scratch, alloc, itch + 1);
alloc = itch + 1;
}
scratch[itch] = ran;
MPN_COPY (rp, np, nn);
mpn_sb_div_r_sec (rp, nn, dup, dn, scratch);
ASSERT_ALWAYS (ran == scratch[itch]);
/* Note: Since check_one cannot cope with random-only functions, we
pass qp[] from the previous function, mpn_sb_div_qr_sec. */
check_one (qp, rp, np, nn, dup, dn, "mpn_sb_div_r_sec", 0);
}
/* Test mpn_dcpi1_div_qr */
if (dn >= 6 && nn - dn >= 3)
{
MPN_COPY (rp, np, nn);
if (nn > dn)
MPN_COPY (qp, junkp, nn - dn);
qp[nn - dn] = mpn_dcpi1_div_qr (qp, rp, nn, dnp, dn, &dinv);
ASSERT_ALWAYS (qp[-1] == qran0); ASSERT_ALWAYS (qp[nn - dn + 1] == qran1);
ASSERT_ALWAYS (rp[-1] == rran0);
check_one (qp, rp, np, nn, dnp, dn, "mpn_dcpi1_div_qr", 0);
}
/* Test mpn_dcpi1_divappr_q */
if (dn >= 6 && nn - dn >= 3)
{
MPN_COPY (rp, np, nn);
if (nn > dn)
MPN_COPY (qp, junkp, nn - dn);
qp[nn - dn] = mpn_dcpi1_divappr_q (qp, rp, nn, dnp, dn, &dinv);
ASSERT_ALWAYS (qp[-1] == qran0); ASSERT_ALWAYS (qp[nn - dn + 1] == qran1);
ASSERT_ALWAYS (rp[-1] == rran0);
check_one (qp, NULL, np, nn, dnp, dn, "mpn_dcpi1_divappr_q", 1);
}
/* Test mpn_dcpi1_div_q */
if (dn >= 6 && nn - dn >= 3)
{
MPN_COPY (rp, np, nn);
if (nn > dn)
MPN_COPY (qp, junkp, nn - dn);
qp[nn - dn] = mpn_dcpi1_div_q (qp, rp, nn, dnp, dn, &dinv);
ASSERT_ALWAYS (qp[-1] == qran0); ASSERT_ALWAYS (qp[nn - dn + 1] == qran1);
ASSERT_ALWAYS (rp[-1] == rran0);
check_one (qp, NULL, np, nn, dnp, dn, "mpn_dcpi1_div_q", 0);
}
/* Test mpn_mu_div_qr */
if (nn - dn > 2 && dn >= 2)
{
itch = mpn_mu_div_qr_itch (nn, dn, 0);
if (itch + 1 > alloc)
{
scratch = __GMP_REALLOCATE_FUNC_LIMBS (scratch, alloc, itch + 1);
alloc = itch + 1;
}
scratch[itch] = ran;
MPN_COPY (qp, junkp, nn - dn);
MPN_ZERO (rp, dn);
rp[dn] = rran1;
qp[nn - dn] = mpn_mu_div_qr (qp, rp, np, nn, dnp, dn, scratch);
ASSERT_ALWAYS (ran == scratch[itch]);
ASSERT_ALWAYS (qp[-1] == qran0); ASSERT_ALWAYS (qp[nn - dn + 1] == qran1);
ASSERT_ALWAYS (rp[-1] == rran0); ASSERT_ALWAYS (rp[dn] == rran1);
check_one (qp, rp, np, nn, dnp, dn, "mpn_mu_div_qr", 0);
}
/* Test mpn_mu_divappr_q */
if (nn - dn > 2 && dn >= 2)
{
itch = mpn_mu_divappr_q_itch (nn, dn, 0);
if (itch + 1 > alloc)
{
scratch = __GMP_REALLOCATE_FUNC_LIMBS (scratch, alloc, itch + 1);
alloc = itch + 1;
}
scratch[itch] = ran;
MPN_COPY (qp, junkp, nn - dn);
qp[nn - dn] = mpn_mu_divappr_q (qp, np, nn, dnp, dn, scratch);
ASSERT_ALWAYS (ran == scratch[itch]);
ASSERT_ALWAYS (qp[-1] == qran0); ASSERT_ALWAYS (qp[nn - dn + 1] == qran1);
check_one (qp, NULL, np, nn, dnp, dn, "mpn_mu_divappr_q", 4);
}
/* Test mpn_mu_div_q */
if (nn - dn > 2 && dn >= 2)
{
itch = mpn_mu_div_q_itch (nn, dn, 0);
if (itch + 1> alloc)
{
scratch = __GMP_REALLOCATE_FUNC_LIMBS (scratch, alloc, itch + 1);
alloc = itch + 1;
}
scratch[itch] = ran;
MPN_COPY (qp, junkp, nn - dn);
qp[nn - dn] = mpn_mu_div_q (qp, np, nn, dnp, dn, scratch);
ASSERT_ALWAYS (ran == scratch[itch]);
ASSERT_ALWAYS (qp[-1] == qran0); ASSERT_ALWAYS (qp[nn - dn + 1] == qran1);
check_one (qp, NULL, np, nn, dnp, dn, "mpn_mu_div_q", 0);
}
if (1)
{
itch = nn + 1;
if (itch + 1> alloc)
{
scratch = __GMP_REALLOCATE_FUNC_LIMBS (scratch, alloc, itch + 1);
alloc = itch + 1;
}
scratch[itch] = ran;
mpn_div_q (qp, np, nn, dup, dn, scratch);
ASSERT_ALWAYS (ran == scratch[itch]);
ASSERT_ALWAYS (qp[-1] == qran0); ASSERT_ALWAYS (qp[nn - dn + 1] == qran1);
check_one (qp, NULL, np, nn, dup, dn, "mpn_div_q", 0);
}
if (dn >= 2 && nn >= 2)
{
mp_limb_t qh;
/* mpn_divrem_2 */
MPN_COPY (rp, np, nn);
qp[nn - 2] = qp[nn-1] = qran1;
qh = mpn_divrem_2 (qp, 0, rp, nn, dnp + dn - 2);
ASSERT_ALWAYS (qp[nn - 2] == qran1);
ASSERT_ALWAYS (qp[-1] == qran0); ASSERT_ALWAYS (qp[nn - 1] == qran1);
qp[nn - 2] = qh;
check_one (qp, rp, np, nn, dnp + dn - 2, 2, "mpn_divrem_2", 0);
/* Missing: divrem_2 with fraction limbs. */
/* mpn_div_qr_2 */
qp[nn - 2] = qran1;
qh = mpn_div_qr_2 (qp, rp, np, nn, dup + dn - 2);
ASSERT_ALWAYS (qp[nn - 2] == qran1);
ASSERT_ALWAYS (qp[-1] == qran0); ASSERT_ALWAYS (qp[nn - 1] == qran1);
qp[nn - 2] = qh;
check_one (qp, rp, np, nn, dup + dn - 2, 2, "mpn_div_qr_2", 0);
}
}
__GMP_FREE_FUNC_LIMBS (scratch, alloc);
TMP_FREE;
mpz_clear (n);
mpz_clear (d);
mpz_clear (q);
mpz_clear (r);
mpz_clear (tz);
mpz_clear (junk);
tests_end ();
return 0;
}
void
mpn_div_q (mp_ptr qp,
mp_srcptr np, mp_size_t nn,
mp_srcptr dp, mp_size_t dn, mp_ptr scratch)
{
mp_ptr new_dp, new_np, tp, rp;
mp_limb_t cy, dh, qh;
mp_size_t new_nn, qn;
gmp_pi1_t dinv;
int cnt;
TMP_DECL;
TMP_MARK;
ASSERT (nn >= dn);
ASSERT (dn > 0);
ASSERT (dp[dn - 1] != 0);
ASSERT (! MPN_OVERLAP_P (qp, nn - dn + 1, np, nn));
ASSERT (! MPN_OVERLAP_P (qp, nn - dn + 1, dp, dn));
ASSERT (MPN_SAME_OR_SEPARATE_P (np, scratch, nn));
ASSERT_ALWAYS (FUDGE >= 2);
if (dn == 1)
{
mpn_divrem_1 (qp, 0L, np, nn, dp[dn - 1]);
return;
}
qn = nn - dn + 1; /* Quotient size, high limb might be zero */
if (qn + FUDGE >= dn)
{
/* |________________________|
|_______| */
new_np = scratch;
dh = dp[dn - 1];
if (LIKELY ((dh & GMP_NUMB_HIGHBIT) == 0))
{
count_leading_zeros (cnt, dh);
cy = mpn_lshift (new_np, np, nn, cnt);
new_np[nn] = cy;
new_nn = nn + (cy != 0);
new_dp = TMP_ALLOC_LIMBS (dn);
mpn_lshift (new_dp, dp, dn, cnt);
if (dn == 2)
{
qh = mpn_divrem_2 (qp, 0L, new_np, new_nn, new_dp);
}
else if (BELOW_THRESHOLD (dn, DC_DIV_Q_THRESHOLD) ||
BELOW_THRESHOLD (new_nn - dn, DC_DIV_Q_THRESHOLD))
{
invert_pi1 (dinv, new_dp[dn - 1], new_dp[dn - 2]);
qh = mpn_sbpi1_div_q (qp, new_np, new_nn, new_dp, dn, dinv.inv32);
}
else if (BELOW_THRESHOLD (dn, MUPI_DIV_Q_THRESHOLD) || /* fast condition */
BELOW_THRESHOLD (nn, 2 * MU_DIV_Q_THRESHOLD) || /* fast condition */
(double) (2 * (MU_DIV_Q_THRESHOLD - MUPI_DIV_Q_THRESHOLD)) * dn /* slow... */
+ (double) MUPI_DIV_Q_THRESHOLD * nn > (double) dn * nn) /* ...condition */
{
invert_pi1 (dinv, new_dp[dn - 1], new_dp[dn - 2]);
qh = mpn_dcpi1_div_q (qp, new_np, new_nn, new_dp, dn, &dinv);
}
else
{
mp_size_t itch = mpn_mu_div_q_itch (new_nn, dn, 0);
mp_ptr scratch = TMP_ALLOC_LIMBS (itch);
qh = mpn_mu_div_q (qp, new_np, new_nn, new_dp, dn, scratch);
}
if (cy == 0)
qp[qn - 1] = qh;
else if (UNLIKELY (qh != 0))
{
/* This happens only when the quotient is close to B^n and
mpn_*_divappr_q returned B^n. */
mp_size_t i, n;
n = new_nn - dn;
for (i = 0; i < n; i++)
qp[i] = GMP_NUMB_MAX;
qh = 0; /* currently ignored */
}
}
else /* divisor is already normalised */
{
if (new_np != np)
MPN_COPY (new_np, np, nn);
if (dn == 2)
{
qh = mpn_divrem_2 (qp, 0L, new_np, nn, dp);
}
else if (BELOW_THRESHOLD (dn, DC_DIV_Q_THRESHOLD) ||
BELOW_THRESHOLD (nn - dn, DC_DIV_Q_THRESHOLD))
{
invert_pi1 (dinv, dh, dp[dn - 2]);
qh = mpn_sbpi1_div_q (qp, new_np, nn, dp, dn, dinv.inv32);
}
else if (BELOW_THRESHOLD (dn, MUPI_DIV_Q_THRESHOLD) || /* fast condition */
BELOW_THRESHOLD (nn, 2 * MU_DIV_Q_THRESHOLD) || /* fast condition */
(double) (2 * (MU_DIV_Q_THRESHOLD - MUPI_DIV_Q_THRESHOLD)) * dn /* slow... */
+ (double) MUPI_DIV_Q_THRESHOLD * nn > (double) dn * nn) /* ...condition */
{
invert_pi1 (dinv, dh, dp[dn - 2]);
qh = mpn_dcpi1_div_q (qp, new_np, nn, dp, dn, &dinv);
}
else
{
mp_size_t itch = mpn_mu_div_q_itch (nn, dn, 0);
mp_ptr scratch = TMP_ALLOC_LIMBS (itch);
qh = mpn_mu_div_q (qp, np, nn, dp, dn, scratch);
}
qp[nn - dn] = qh;
}
}
else
{
/* |________________________|
|_________________| */
tp = TMP_ALLOC_LIMBS (qn + 1);
new_np = scratch;
new_nn = 2 * qn + 1;
if (new_np == np)
/* We need {np,nn} to remain untouched until the final adjustment, so
we need to allocate separate space for new_np. */
new_np = TMP_ALLOC_LIMBS (new_nn + 1);
dh = dp[dn - 1];
if (LIKELY ((dh & GMP_NUMB_HIGHBIT) == 0))
{
count_leading_zeros (cnt, dh);
cy = mpn_lshift (new_np, np + nn - new_nn, new_nn, cnt);
new_np[new_nn] = cy;
new_nn += (cy != 0);
new_dp = TMP_ALLOC_LIMBS (qn + 1);
mpn_lshift (new_dp, dp + dn - (qn + 1), qn + 1, cnt);
new_dp[0] |= dp[dn - (qn + 1) - 1] >> (GMP_NUMB_BITS - cnt);
if (qn + 1 == 2)
{
qh = mpn_divrem_2 (tp, 0L, new_np, new_nn, new_dp);
}
else if (BELOW_THRESHOLD (qn, DC_DIVAPPR_Q_THRESHOLD - 1))
{
invert_pi1 (dinv, new_dp[qn], new_dp[qn - 1]);
qh = mpn_sbpi1_divappr_q (tp, new_np, new_nn, new_dp, qn + 1, dinv.inv32);
}
else if (BELOW_THRESHOLD (qn, MU_DIVAPPR_Q_THRESHOLD - 1))
{
invert_pi1 (dinv, new_dp[qn], new_dp[qn - 1]);
qh = mpn_dcpi1_divappr_q (tp, new_np, new_nn, new_dp, qn + 1, &dinv);
}
else
{
mp_size_t itch = mpn_mu_divappr_q_itch (new_nn, qn + 1, 0);
mp_ptr scratch = TMP_ALLOC_LIMBS (itch);
qh = mpn_mu_divappr_q (tp, new_np, new_nn, new_dp, qn + 1, scratch);
}
if (cy == 0)
tp[qn] = qh;
else if (UNLIKELY (qh != 0))
{
/* This happens only when the quotient is close to B^n and
mpn_*_divappr_q returned B^n. */
mp_size_t i, n;
n = new_nn - (qn + 1);
for (i = 0; i < n; i++)
tp[i] = GMP_NUMB_MAX;
qh = 0; /* currently ignored */
}
}
else /* divisor is already normalised */
{