void tc4_rshift_inplace(mp_ptr rp, mp_size_t * rn, mp_size_t bits)
{
if (*rn)
{
if ((*rn) > 0)
{
mpn_rshift(rp, rp, *rn, bits);
if (rp[(*rn) - 1] == CNST_LIMB(0)) (*rn)--;
} else
{
mpn_rshift(rp, rp, -(*rn), bits);
if (rp[-(*rn) - 1] == CNST_LIMB(0)) (*rn)++;
}
}
}
void
_gst_mpz_div_2exp (gst_mpz *w, const gst_mpz *u, unsigned cnt)
{
mp_size_t usize = u->size;
mp_size_t wsize;
mp_size_t abs_usize = ABS (usize);
mp_size_t limb_cnt;
limb_cnt = cnt / BITS_PER_MP_LIMB;
wsize = abs_usize - limb_cnt;
if (wsize <= 0)
wsize = 0;
else
{
if (w->alloc < wsize)
gst_mpz_realloc (w, wsize);
if (cnt % BITS_PER_MP_LIMB)
mpn_rshift (w->d, u->d + limb_cnt, abs_usize - limb_cnt,
cnt % BITS_PER_MP_LIMB);
else
MPN_COPY (w->d, u->d + limb_cnt, abs_usize - limb_cnt);
wsize -= w->d[wsize - 1] == 0;
}
w->size = (usize >= 0) ? wsize : -wsize;
}
/* Twos-complement version of 'integer_gmp_mpn_rshift' for performing
* arithmetic right shifts on "negative" MPNs.
*
* Same pre-conditions as 'integer_gmp_mpn_rshift'
*
* This variant is needed to operate on MPNs interpreted as negative
* numbers, which require "rounding" towards minus infinity iff a
* non-zero bit is shifted out.
*/
mp_limb_t
integer_gmp_mpn_rshift_2c (mp_limb_t rp[], const mp_limb_t sp[],
const mp_size_t sn, const mp_bitcnt_t count)
{
const mp_size_t limb_shift = count / GMP_NUMB_BITS;
const unsigned int bit_shift = count % GMP_NUMB_BITS;
const mp_size_t rn = sn - limb_shift;
// whether non-zero bits were shifted out
bool nz_shift_out = false;
if (bit_shift) {
if (mpn_rshift(rp, &sp[limb_shift], rn, bit_shift))
nz_shift_out = true;
} else
memcpy(rp, &sp[limb_shift], rn*sizeof(mp_limb_t));
if (!nz_shift_out)
for (unsigned i = 0; i < limb_shift; i++)
if (sp[i]) {
nz_shift_out = true;
break;
}
// round if non-zero bits were shifted out
if (nz_shift_out)
if (mpn_add_1(rp, rp, rn, 1))
abort(); /* should never happen */
return rp[rn-1];
}
void fmpz_div_2exp(fmpz_t output, fmpz_t x, unsigned long exp)
{
unsigned long limbs = (exp >> FLINT_LG_BITS_PER_LIMB);
unsigned long bits = (exp & (FLINT_BITS - 1));
if ((x[0] == 0) || (limbs >= FLINT_ABS(x[0])))
{
output[0] = 0L;
return;
}
if (bits)
{
fmpz_t temp = fmpz_init(FLINT_ABS(x[0]) - limbs);
mpn_rshift(temp + 1, x + limbs + 1, FLINT_ABS(x[0]) - limbs, bits);
if ((long) x[0] >= 0L) temp[0] = x[0] - limbs;
else temp[0] = limbs + x[0];
NORM(temp);
fmpz_set(output, temp);
fmpz_clear(temp);
} else
{
F_mpn_copy(output + 1, x + limbs + 1, FLINT_ABS(x[0]) - limbs);
if ((long) x[0] >= 0L) output[0] = x[0] - limbs;
else output[0] = limbs + x[0];
}
}
REGPARM_ATTR (1) static void
cfdiv_q_2exp (mpz_ptr w, mpz_srcptr u, mp_bitcnt_t cnt, int dir)
{
mp_size_t wsize, usize, abs_usize, limb_cnt, i;
mp_srcptr up;
mp_ptr wp;
mp_limb_t round, rmask;
usize = SIZ (u);
abs_usize = ABS (usize);
limb_cnt = cnt / GMP_NUMB_BITS;
wsize = abs_usize - limb_cnt;
if (wsize <= 0)
{
/* u < 2**cnt, so result 1, 0 or -1 according to rounding */
PTR(w)[0] = 1;
SIZ(w) = (usize == 0 || (usize ^ dir) < 0 ? 0 : dir);
return;
}
/* +1 limb to allow for mpn_add_1 below */
MPZ_REALLOC (w, wsize+1);
/* Check for rounding if direction matches u sign.
Set round if we're skipping non-zero limbs. */
up = PTR(u);
round = 0;
rmask = ((usize ^ dir) >= 0 ? MP_LIMB_T_MAX : 0);
if (rmask != 0)
for (i = 0; i < limb_cnt && round == 0; i++)
round = up[i];
wp = PTR(w);
cnt %= GMP_NUMB_BITS;
if (cnt != 0)
{
round |= rmask & mpn_rshift (wp, up + limb_cnt, wsize, cnt);
wsize -= (wp[wsize - 1] == 0);
}
else
MPN_COPY_INCR (wp, up + limb_cnt, wsize);
if (round != 0)
{
if (wsize != 0)
{
mp_limb_t cy;
cy = mpn_add_1 (wp, wp, wsize, CNST_LIMB(1));
wp[wsize] = cy;
wsize += cy;
}
else
{
/* We shifted something to zero. */
wp[0] = 1;
wsize = 1;
}
}
SIZ(w) = (usize >= 0 ? wsize : -wsize);
}
void
my__gmpn_tdiv_qr (mp_ptr qp, mp_ptr rp, mp_size_t qxn,
mp_srcptr np, mp_size_t nn, mp_srcptr dp, mp_size_t dn)
{
ASSERT_ALWAYS (qxn == 0);
ASSERT (nn >= 0);
ASSERT (dn >= 0);
ASSERT (dn == 0 || dp[dn - 1] != 0);
ASSERT (! MPN_OVERLAP_P (qp, nn - dn + 1 + qxn, np, nn));
ASSERT (! MPN_OVERLAP_P (qp, nn - dn + 1 + qxn, dp, dn));
int adjust;
gmp_pi1_t dinv;
TMP_DECL;
TMP_MARK;
/* conservative tests for quotient size */
adjust = np[nn - 1] >= dp[dn - 1];
mp_ptr n2p, d2p;
mp_limb_t cy;
int cnt;
qp[nn - dn] = 0; /* zero high quotient limb */
count_leading_zeros (cnt, dp[dn - 1]);
cnt -= GMP_NAIL_BITS;
d2p = TMP_ALLOC_LIMBS (dn);
mpn_lshift (d2p, dp, dn, cnt);
for (int i=0; i<dn; i+=1)
{
printf("d2p %08x\n", *( (int*) (((void*)(d2p))+(i*4))));
}
n2p = TMP_ALLOC_LIMBS (nn + 1);
cy = mpn_lshift (n2p, np, nn, cnt);
for (int i=0; i<nn; i+=1)
{
printf("n2p %08x\n", *( (int*) (((void*)(n2p))+(i*4))));
}
n2p[nn] = cy;
nn += adjust;
printf("d2p[dn-1] = %08lx\nd2p[dn-2] = %08lx\n", d2p[dn-1], d2p[dn-2]);
invert_pi1 (dinv, d2p[dn - 1], d2p[dn - 2]);
printf("dinv %08lx\n", dinv.inv32);
my_mpn_sbpi1_div_qr (qp, n2p, nn, d2p, dn, dinv.inv32);
for (int i=0; i<nn; i+=1)
{
printf("inside qp %08x\n", *( (int*) (((void*)(qp))+(i*4))));
}
n2p[nn] = cy;
mpn_rshift (rp, n2p, dn, cnt);
TMP_FREE;
return;
}
static void fp_halve(element_ptr r, element_ptr a) {
fp_field_data_ptr p = r->field->data;
const size_t t = p->limbs;
int carry = 0;
mp_limb_t *alimb = a->data;
mp_limb_t *rlimb = r->data;
if (alimb[0] & 1) carry = mpn_add_n(rlimb, alimb, p->primelimbs, t);
else fp_set(r, a);
mpn_rshift(rlimb, rlimb, t, 1);
if (carry) rlimb[t - 1] |= ((mp_limb_t) 1) << (sizeof(mp_limb_t) * 8 - 1);
}
/* truncates, returns inexact */
int
_arf_get_integer_mpn(mp_ptr y, mp_srcptr x, mp_size_t xn, slong exp)
{
slong bot_exp = exp - xn * FLINT_BITS;
if (bot_exp >= 0)
{
mp_size_t bot_limbs;
mp_bitcnt_t bot_bits;
bot_limbs = bot_exp / FLINT_BITS;
bot_bits = bot_exp % FLINT_BITS;
flint_mpn_zero(y, bot_limbs);
if (bot_bits == 0)
flint_mpn_copyi(y + bot_limbs, x, xn);
else
y[bot_limbs + xn] = mpn_lshift(y + bot_limbs, x, xn, bot_bits);
/* exact */
return 0;
}
else if (exp <= 0)
{
/* inexact */
return 1;
}
else
{
mp_size_t top_limbs;
mp_bitcnt_t top_bits;
mp_limb_t cy;
top_limbs = exp / FLINT_BITS;
top_bits = exp % FLINT_BITS;
if (top_bits == 0)
{
flint_mpn_copyi(y, x + xn - top_limbs, top_limbs);
/* inexact */
return 1;
}
else
{
/* can be inexact */
cy = mpn_rshift(y, x + xn - top_limbs - 1,
top_limbs + 1, FLINT_BITS - top_bits);
return (cy != 0) || (top_limbs + 1 != xn);
}
}
}
/* Compute r such that r^2 * y = 1 (mod 2^{b+1}).
Return non-zero if such an integer r exists.
Iterates
r' <-- (3r - r^3 y) / 2
using Hensel lifting. Since we divide by two, the Hensel lifting is
somewhat degenerates. Therefore, we lift from 2^b to 2^{b+1}-1.
FIXME:
(1) Simplify to do precision book-keeping in limbs rather than bits.
(2) Rewrite iteration as
r' <-- r - r (r^2 y - 1) / 2
and take advantage of zero low part of r^2 y - 1.
(3) Use wrap-around trick.
(4) Use a small table to get starting value.
*/
int
mpn_bsqrtinv (mp_ptr rp, mp_srcptr yp, mp_bitcnt_t bnb, mp_ptr tp)
{
mp_ptr tp2, tp3;
mp_limb_t k;
mp_size_t bn, order[GMP_LIMB_BITS + 1];
int i, d;
ASSERT (bnb > 0);
bn = 1 + bnb / GMP_LIMB_BITS;
tp2 = tp + bn;
tp3 = tp + 2 * bn;
k = 3;
rp[0] = 1;
if (bnb == 1)
{
if ((yp[0] & 3) != 1)
return 0;
}
else
{
if ((yp[0] & 7) != 1)
return 0;
d = 0;
for (; bnb != 2; bnb = (bnb + 2) >> 1)
order[d++] = bnb;
for (i = d - 1; i >= 0; i--)
{
bnb = order[i];
bn = 1 + bnb / GMP_LIMB_BITS;
mpn_mul_1 (tp, rp, bn, k);
mpn_powlo (tp2, rp, &k, 1, bn, tp3);
mpn_mullo_n (rp, yp, tp2, bn);
#if HAVE_NATIVE_mpn_rsh1sub_n
mpn_rsh1sub_n (rp, tp, rp, bn);
#else
mpn_sub_n (tp2, tp, rp, bn);
mpn_rshift (rp, tp2, bn, 1);
#endif
}
}
return 1;
}
/* Perform arithmetic right shift on MPNs (multi-precision naturals)
*
* pre-conditions:
* - 0 < count < sn*GMP_NUMB_BITS
* - rn = sn - floor(count / GMP_NUMB_BITS)
* - sn > 0
*
* write {sp,sn} right-shifted by count bits into {rp,rn}
*
* return value: most-significant limb stored in {rp,rn} result
*/
mp_limb_t
integer_gmp_mpn_rshift (mp_limb_t rp[], const mp_limb_t sp[], mp_size_t sn,
mp_bitcnt_t count)
{
const mp_size_t limb_shift = count / GMP_NUMB_BITS;
const unsigned int bit_shift = count % GMP_NUMB_BITS;
const mp_size_t rn = sn - limb_shift;
if (bit_shift)
mpn_rshift(rp, &sp[limb_shift], rn, bit_shift);
else
memcpy(rp, &sp[limb_shift], rn*sizeof(mp_limb_t));
return rp[rn-1];
}
static void fp_halve(element_ptr c, element_ptr a) {
eptr ad = (eptr)a->data, cd = (eptr)c->data;
if (!ad->flag) {
cd->flag = 0;
}
else {
fptr p = (fptr)c->field->data;
const size_t t = p->limbs;
int carry = 0;
mp_limb_t *alimb = ad->d;
mp_limb_t *climb = cd->d;
if (alimb[0] & 1) {
carry = mpn_add_n(climb, alimb, p->primelimbs, t);
}
else fp_set(c, a);
mpn_rshift(climb, climb, t, 1);
if (carry) climb[t - 1] |= ((mp_limb_t)1) << (sizeof(mp_limb_t)* 8 - 1);
}
}
mp_exp_t
mpfr_get_z_exp (mpz_ptr z, mpfr_srcptr f)
{
mp_size_t fn;
int sh;
MPFR_ASSERTD (MPFR_IS_FP (f));
if (MPFR_UNLIKELY (MPFR_IS_ZERO (f)))
{
mpz_set_ui (z, 0);
return __gmpfr_emin;
}
fn = MPFR_LIMB_SIZE(f);
/* check whether allocated space for z is enough */
if (MPFR_UNLIKELY (ALLOC (z) < fn))
MPZ_REALLOC (z, fn);
MPFR_UNSIGNED_MINUS_MODULO (sh, MPFR_PREC (f));
if (MPFR_LIKELY (sh))
mpn_rshift (PTR (z), MPFR_MANT (f), fn, sh);
else
MPN_COPY (PTR (z), MPFR_MANT (f), fn);
SIZ(z) = MPFR_IS_NEG (f) ? -fn : fn;
/* Test if the result is representable. Later, we could choose
to return MPFR_EXP_MIN if it isn't, or perhaps MPFR_EXP_MAX
to signal an error. The mantissa would still be meaningful. */
MPFR_ASSERTD ((mp_exp_unsigned_t) MPFR_GET_EXP (f) - MPFR_EXP_MIN
>= (mp_exp_unsigned_t) MPFR_PREC(f));
return MPFR_GET_EXP (f) - MPFR_PREC (f);
}
void
mpn_toom_interpolate_5pts (mp_ptr c, mp_ptr v2, mp_ptr vm1,
mp_size_t k, mp_size_t twor, int sa,
mp_limb_t vinf0, mp_ptr ws)
{
mp_limb_t cy, saved;
mp_size_t twok = k + k;
mp_size_t kk1 = twok + 1;
mp_ptr c1, v1, c3, vinf, c5;
mp_limb_t cout; /* final carry, should be zero at the end */
c1 = c + k;
v1 = c1 + k;
c3 = v1 + k;
vinf = c3 + k;
c5 = vinf + k;
#define v0 (c)
/* (1) v2 <- v2-vm1 < v2+|vm1|, (16 8 4 2 1) - (1 -1 1 -1 1) =
thus 0 <= v2 < 50*B^(2k) < 2^6*B^(2k) (15 9 3 3 0)
*/
if (sa <= 0)
mpn_add_n (v2, v2, vm1, kk1);
else
mpn_sub_n (v2, v2, vm1, kk1);
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 v1 hi(vinf) |vm1| v2-vm1 EMPTY */
ASSERT_NOCARRY (mpn_divexact_by3 (v2, v2, kk1)); /* v2 <- v2 / 3 */
/* (5 3 1 1 0)*/
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 v1 hi(vinf) |vm1| (v2-vm1)/3 EMPTY */
/* (2) vm1 <- tm1 := (v1 - sa*vm1) / 2 [(1 1 1 1 1) - (1 -1 1 -1 1)] / 2 =
tm1 >= 0 (0 1 0 1 0)
No carry comes out from {v1, kk1} +/- {vm1, kk1},
and the division by two is exact */
if (sa <= 0)
{
#ifdef HAVE_NATIVE_mpn_rsh1add_n
mpn_rsh1add_n (vm1, v1, vm1, kk1);
#else
mpn_add_n (vm1, v1, vm1, kk1);
mpn_rshift (vm1, vm1, kk1, 1);
#endif
}
else
{
#ifdef HAVE_NATIVE_mpn_rsh1sub_n
mpn_rsh1sub_n (vm1, v1, vm1, kk1);
#else
mpn_sub_n (vm1, v1, vm1, kk1);
mpn_rshift (vm1, vm1, kk1, 1);
#endif
}
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 v1 hi(vinf) tm1 (v2-vm1)/3 EMPTY */
/* (3) v1 <- t1 := v1 - v0 (1 1 1 1 1) - (0 0 0 0 1) = (1 1 1 1 0)
t1 >= 0
*/
vinf[0] -= mpn_sub_n (v1, v1, c, twok);
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 v1-v0 hi(vinf) tm1 (v2-vm1)/3 EMPTY */
/* (4) v2 <- t2 := ((v2-vm1)/3-t1)/2 = (v2-vm1-3*t1)/6
t2 >= 0 [(5 3 1 1 0) - (1 1 1 1 0)]/2 = (2 1 0 0 0)
*/
#ifdef HAVE_NATIVE_mpn_rsh1sub_n
mpn_rsh1sub_n (v2, v2, v1, kk1);
#else
mpn_sub_n (v2, v2, v1, kk1);
mpn_rshift (v2, v2, kk1, 1);
#endif
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 v1-v0 hi(vinf) tm1 (v2-vm1-3t1)/6 EMPTY */
/* (5) v1 <- t1-tm1 (1 1 1 1 0) - (0 1 0 1 0) = (1 0 1 0 0)
result is v1 >= 0
*/
mpn_sub_n (v1, v1, vm1, kk1);
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 v1-v0-tm1 hi(vinf) tm1 (v2-vm1-3t1)/6 EMPTY */
/* (6) v2 <- v2 - 2*vinf, (2 1 0 0 0) - 2*(1 0 0 0 0) = (0 1 0 0 0)
result is v2 >= 0 */
saved = vinf[0]; /* Remember v1's highest byte (will be overwritten). */
vinf[0] = vinf0; /* Set the right value for vinf0 */
#ifdef HAVE_NATIVE_mpn_sublsh1_n
cy = mpn_sublsh1_n (v2, v2, vinf, twor);
#else
cy = mpn_lshift (ws, vinf, twor, 1);
cy += mpn_sub_n (v2, v2, ws, twor);
#endif
MPN_DECR_U (v2 + twor, kk1 - twor, cy);
/* (7) v1 <- v1 - vinf, (1 0 1 0 0) - (1 0 0 0 0) = (0 0 1 0 0)
result is >= 0 */
cy = mpn_sub_n (v1, v1, vinf, twor); /* vinf is at most twor long. */
vinf[0] = saved;
MPN_DECR_U (v1 + twor, kk1 - twor, cy); /* Treat the last bytes. */
__GMPN_ADD_1 (cout, vinf, vinf, twor, vinf0); /* Add vinf0, propagate carry. */
/* (8) vm1 <- vm1-t2 (0 1 0 1 0) - (0 1 0 0 0) = (0 0 0 1 0)
vm1 >= 0
*/
mpn_sub_n (vm1, vm1, v2, kk1); /* No overlapping here. */
/********************* Beginning the final phase **********************/
/* {c,2k} {c+2k,2k } {c+4k ,2r } {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 t1 hi(t1)+vinf tm1 (v2-vm1-3t1)/6 EMPTY */
/* (9) add t2 in {c+3k, ...} */
cy = mpn_add_n (c3, c3, v2, kk1);
__GMPN_ADD_1 (cout, c5 + 1, c5 + 1, twor - k - 1, cy); /* 2n-(5k+1) = 2r-k-1 */
/* {c,2k} {c+2k,2k } {c+4k ,2r } {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 t1 hi(t1)+vinf tm1 (v2-vm1-3t1)/6 EMPTY */
/* c c+k c+2k c+3k c+4k t t+2k+1 t+4k+2
v0 t1 vinf tm1 t2
+t2 */
/* add vm1 in {c+k, ...} */
cy = mpn_add_n (c1, c1, vm1, kk1);
__GMPN_ADD_1 (cout, c3 + 1, c3 + 1, twor + k - 1, cy); /* 2n-(3k+1) = 2r+k-1 */
/* c c+k c+2k c+3k c+4k t t+2k+1 t+4k+2
v0 t1 vinf tm1 t2
+tm1 +t2 */
#undef v0
#undef t2
}
void
mpn_toom_interpolate_12pts (mp_ptr pp, mp_ptr r1, mp_ptr r3, mp_ptr r5,
mp_size_t n, mp_size_t spt, int half, mp_ptr wsi)
{
mp_limb_t cy;
mp_size_t n3;
mp_size_t n3p1;
n3 = 3 * n;
n3p1 = n3 + 1;
#define r4 (pp + n3) /* 3n+1 */
#define r2 (pp + 7 * n) /* 3n+1 */
#define r0 (pp +11 * n) /* s+t <= 2*n */
/******************************* interpolation *****************************/
if (half != 0) {
cy = mpn_sub_n (r3, r3, r0, spt);
MPN_DECR_U (r3 + spt, n3p1 - spt, cy);
cy = DO_mpn_sublsh_n (r2, r0, spt, 10, wsi);
MPN_DECR_U (r2 + spt, n3p1 - spt, cy);
DO_mpn_subrsh(r5, n3p1, r0, spt, 2, wsi);
cy = DO_mpn_sublsh_n (r1, r0, spt, 20, wsi);
MPN_DECR_U (r1 + spt, n3p1 - spt, cy);
DO_mpn_subrsh(r4, n3p1, r0, spt, 4, wsi);
};
r4[n3] -= DO_mpn_sublsh_n (r4 + n, pp, 2 * n, 20, wsi);
DO_mpn_subrsh(r1 + n, 2 * n + 1, pp, 2 * n, 4, wsi);
#if HAVE_NATIVE_mpn_add_n_sub_n
mpn_add_n_sub_n (r1, r4, r4, r1, n3p1);
#else
ASSERT_NOCARRY(mpn_add_n (wsi, r1, r4, n3p1));
mpn_sub_n (r4, r4, r1, n3p1); /* can be negative */
MP_PTR_SWAP(r1, wsi);
#endif
r5[n3] -= DO_mpn_sublsh_n (r5 + n, pp, 2 * n, 10, wsi);
DO_mpn_subrsh(r2 + n, 2 * n + 1, pp, 2 * n, 2, wsi);
#if HAVE_NATIVE_mpn_add_n_sub_n
mpn_add_n_sub_n (r2, r5, r5, r2, n3p1);
#else
mpn_sub_n (wsi, r5, r2, n3p1); /* can be negative */
ASSERT_NOCARRY(mpn_add_n (r2, r2, r5, n3p1));
MP_PTR_SWAP(r5, wsi);
#endif
r3[n3] -= mpn_sub_n (r3+n, r3+n, pp, 2 * n);
#if AORSMUL_FASTER_AORS_AORSLSH
mpn_submul_1 (r4, r5, n3p1, 257); /* can be negative */
#else
mpn_sub_n (r4, r4, r5, n3p1); /* can be negative */
DO_mpn_sublsh_n (r4, r5, n3p1, 8, wsi); /* can be negative */
#endif
/* A division by 2835x4 follows. Warning: the operand can be negative! */
mpn_divexact_by2835x4(r4, r4, n3p1);
if ((r4[n3] & (GMP_NUMB_MAX << (GMP_NUMB_BITS-3))) != 0)
r4[n3] |= (GMP_NUMB_MAX << (GMP_NUMB_BITS-2));
#if AORSMUL_FASTER_2AORSLSH
mpn_addmul_1 (r5, r4, n3p1, 60); /* can be negative */
#else
DO_mpn_sublsh_n (r5, r4, n3p1, 2, wsi); /* can be negative */
DO_mpn_addlsh_n (r5, r4, n3p1, 6, wsi); /* can give a carry */
#endif
mpn_divexact_by255(r5, r5, n3p1);
ASSERT_NOCARRY(DO_mpn_sublsh_n (r2, r3, n3p1, 5, wsi));
#if AORSMUL_FASTER_3AORSLSH
ASSERT_NOCARRY(mpn_submul_1 (r1, r2, n3p1, 100));
#else
ASSERT_NOCARRY(DO_mpn_sublsh_n (r1, r2, n3p1, 6, wsi));
ASSERT_NOCARRY(DO_mpn_sublsh_n (r1, r2, n3p1, 5, wsi));
ASSERT_NOCARRY(DO_mpn_sublsh_n (r1, r2, n3p1, 2, wsi));
#endif
ASSERT_NOCARRY(DO_mpn_sublsh_n (r1, r3, n3p1, 9, wsi));
mpn_divexact_by42525(r1, r1, n3p1);
#if AORSMUL_FASTER_AORS_2AORSLSH
ASSERT_NOCARRY(mpn_submul_1 (r2, r1, n3p1, 225));
#else
ASSERT_NOCARRY(mpn_sub_n (r2, r2, r1, n3p1));
ASSERT_NOCARRY(DO_mpn_addlsh_n (r2, r1, n3p1, 5, wsi));
ASSERT_NOCARRY(DO_mpn_sublsh_n (r2, r1, n3p1, 8, wsi));
#endif
mpn_divexact_by9x4(r2, r2, n3p1);
ASSERT_NOCARRY(mpn_sub_n (r3, r3, r2, n3p1));
mpn_sub_n (r4, r2, r4, n3p1);
ASSERT_NOCARRY(mpn_rshift(r4, r4, n3p1, 1));
ASSERT_NOCARRY(mpn_sub_n (r2, r2, r4, n3p1));
mpn_add_n (r5, r5, r1, n3p1);
ASSERT_NOCARRY(mpn_rshift(r5, r5, n3p1, 1));
/* last interpolation steps... */
ASSERT_NOCARRY(mpn_sub_n (r3, r3, r1, n3p1));
ASSERT_NOCARRY(mpn_sub_n (r1, r1, r5, n3p1));
/* ... could be mixed with recomposition
||H-r5|M-r5|L-r5| ||H-r1|M-r1|L-r1|
*/
/***************************** recomposition *******************************/
/*
pp[] prior to operations:
|M r0|L r0|___||H r2|M r2|L r2|___||H r4|M r4|L r4|____|H_r6|L r6|pp
summation scheme for remaining operations:
|__12|n_11|n_10|n__9|n__8|n__7|n__6|n__5|n__4|n__3|n__2|n___|n___|pp
|M r0|L r0|___||H r2|M r2|L r2|___||H r4|M r4|L r4|____|H_r6|L r6|pp
||H r1|M r1|L r1| ||H r3|M r3|L r3| ||H_r5|M_r5|L_r5|
*/
cy = mpn_add_n (pp + n, pp + n, r5, n);
cy = mpn_add_1 (pp + 2 * n, r5 + n, n, cy);
#if HAVE_NATIVE_mpn_add_nc
cy = r5[n3] + mpn_add_nc(pp + n3, pp + n3, r5 + 2 * n, n, cy);
#else
MPN_INCR_U (r5 + 2 * n, n + 1, cy);
cy = r5[n3] + mpn_add_n (pp + n3, pp + n3, r5 + 2 * n, n);
#endif
MPN_INCR_U (pp + n3 + n, 2 * n + 1, cy);
pp[2 * n3]+= mpn_add_n (pp + 5 * n, pp + 5 * n, r3, n);
cy = mpn_add_1 (pp + 2 * n3, r3 + n, n, pp[2 * n3]);
#if HAVE_NATIVE_mpn_add_nc
cy = r3[n3] + mpn_add_nc(pp + 7 * n, pp + 7 * n, r3 + 2 * n, n, cy);
#else
MPN_INCR_U (r3 + 2 * n, n + 1, cy);
cy = r3[n3] + mpn_add_n (pp + 7 * n, pp + 7 * n, r3 + 2 * n, n);
#endif
MPN_INCR_U (pp + 8 * n, 2 * n + 1, cy);
pp[10*n]+=mpn_add_n (pp + 9 * n, pp + 9 * n, r1, n);
if (half) {
cy = mpn_add_1 (pp + 10 * n, r1 + n, n, pp[10 * n]);
#if HAVE_NATIVE_mpn_add_nc
if (LIKELY (spt > n)) {
cy = r1[n3] + mpn_add_nc(pp + 11 * n, pp + 11 * n, r1 + 2 * n, n, cy);
MPN_INCR_U (pp + 4 * n3, spt - n, cy);
} else {
ASSERT_NOCARRY(mpn_add_nc(pp + 11 * n, pp + 11 * n, r1 + 2 * n, spt, cy));
}
#else
MPN_INCR_U (r1 + 2 * n, n + 1, cy);
if (LIKELY (spt > n)) {
cy = r1[n3] + mpn_add_n (pp + 11 * n, pp + 11 * n, r1 + 2 * n, n);
MPN_INCR_U (pp + 4 * n3, spt - n, cy);
} else {
ASSERT_NOCARRY(mpn_add_n (pp + 11 * n, pp + 11 * n, r1 + 2 * n, spt));
}
#endif
} else {
ASSERT_NOCARRY(mpn_add_1 (pp + 10 * n, r1 + n, spt, pp[10 * n]));
}
#undef r0
#undef r2
#undef r4
}
/*
Toom 4 interpolation. Interpolates the value at 2^(sn*B) of a
polynomial p(x) with 7 coefficients given the values
p(oo), p(2), p(1), p(-1), 2^6*p(1/2), 2^6*p(-1/2), p(0).
The output is placed in rp and the final number of limbs of the
output is given in rpn.
The 4th and 6th values may be negative, and if so, n4 and n6
should be set to a negative value respectively.
To save space we pass r3, r5, r7 in place in the output rp.
The other r's are stored separately in space tp.
The low limb of r3 is stored in r30, as it will be overwritten
by the high limb of r5.
rp rp1 rp2 rp3 rp4 rp5 rp6 rp7
<----------- r7-----------><------------r5-------------->
<-------------r3------------->
We assume that r1 is stored at tp, r2 at (tp + t4), r4 at (tp + 2*t4)
and r6 (tp + 3*t4). Each of these r's has t4 = s4 + 1 limbs allocated.
*/
void mpn_toom4_interpolate(mp_ptr rp, mp_size_t * rpn, mp_size_t sn,
mp_ptr tp, mp_size_t s4, mp_size_t n4, mp_size_t n6, mp_limb_t r30)
{
mp_size_t n1, n2, n3, n5, n7, t4;
mp_limb_t saved, saved2, cy;
t4 = s4 + 1;
mpn_add_n(r2, r2, r5, s4);
if (n6 < 0)
mpn_add_n(r6, r5, r6, s4);
else
mpn_sub_n(r6, r5, r6, s4);
/* r6 is now in twos complement format */
saved = r3[0];
r3[0] = r30;
if (n4 < 0)
mpn_add_n(r4, r3, r4, s4);
else
mpn_sub_n(r4, r3, r4, s4);
r3[0] = saved;
/* r4 is now in twos complement format */
mpn_sub_n(r5, r5, r1, s4);
#if HAVE_NATIVE_mpn_sublsh_n
r5[s4-1] -= mpn_sublsh_n(r5, r5, r7, s4-1, 6);
#else
r5[s4-1] -= mpn_submul_1(r5, r7, s4-1, 64);
#endif
TC4_RSHIFT1(r4, s4);
saved = r3[0];
r3[0] = r30;
mpn_sub_n(r3, r3, r4, s4);
r30 = r3[0];
r3[0] = saved;
mpn_double(r5, s4);
mpn_sub_n(r5, r5, r6, s4);
saved = r3[0];
r3[0] = r30;
mpn_submul_1(r2, r3, s4, 65);
r3[0] = saved;
saved2 = r7[s4-1];
r7[s4-1] = CNST_LIMB(0); // r7 is always positive so no sign extend needed
saved = r3[0];
r3[0] = r30;
#if HAVE_NATIVE_mpn_subadd_n
mpn_subadd_n(r3, r3, r7, r1, s4);
#else
mpn_sub_n(r3, r3, r7, s4);
mpn_sub_n(r3, r3, r1, s4);
#endif
r7[s4-1] = saved2;
r30 = r3[0];
mpn_addmul_1(r2, r3, s4, 45);
#if HAVE_NATIVE_mpn_sublsh_n
cy = mpn_sublsh_n(r5, r5, r3, s4 - 1, 3);
#else
cy = mpn_submul_1(r5, r3, s4 - 1, 8);
#endif
r3[0] = saved;
r3[0] -= (cy + 8*r3[s4-1]);
mpn_rshift(r5, r5, s4, 3);
mpn_divexact_by3(r5, r5, s4);
mpn_sub_n(r6, r6, r2, s4);
#if HAVE_NATIVE_mpn_sublsh_n
mpn_sublsh_n(r2, r2, r4, s4, 4);
#else
mpn_submul_1(r2, r4, s4, 16);
#endif
mpn_rshift(r2, r2, s4, 1);
mpn_divexact_by3(r2, r2, s4);
mpn_divexact_by3(r2, r2, s4);
saved = r3[0];
r3[0] = r30;
cy = mpn_sub_n(r3, r3, r5, s4 - 1);
r30 = r3[0];
r3[0] = saved;
r3[s4-1] -= (cy + r5[s4-1]);
mpn_sub_n(r4, r4, r2, s4);
mpn_addmul_1(r6, r2, s4, 30);
mpn_divexact_byfobm1(r6, r6, s4, CNST_LIMB(15), CNST_LIMB(~0/15));
mpn_rshift(r6, r6, s4, 2);
mpn_sub_n(r2, r2, r6, s4);
TC4_NORM(r1, n1, s4);
TC4_NORM(r2, n2, s4);
(*rpn) = 6*sn+1;
cy = mpn_add_1(r3, r3, *rpn - 4*sn, r30); /* don't forget to add r3[0] back in */
if (cy)
{
rp[*rpn] = cy;
(*rpn)++;
}
tc4_copy(rp, rpn, 5*sn, r2, n2);
tc4_copy(rp, rpn, 6*sn, r1, n1);
tc4_copy(rp, rpn, sn, r6, s4);
tc4_copy(rp, rpn, 3*sn, r4, s4);
}
int
mpfr_sqrt (mpfr_ptr r, mpfr_srcptr u, mpfr_rnd_t rnd_mode)
{
mp_size_t rsize; /* number of limbs of r (plus 1 if exact limb multiple) */
mp_size_t rrsize;
mp_size_t usize; /* number of limbs of u */
mp_size_t tsize; /* number of limbs of the sqrtrem remainder */
mp_size_t k;
mp_size_t l;
mpfr_limb_ptr rp, rp0;
mpfr_limb_ptr up;
mpfr_limb_ptr sp;
mp_limb_t sticky0; /* truncated part of input */
mp_limb_t sticky1; /* truncated part of rp[0] */
mp_limb_t sticky;
int odd_exp;
int sh; /* number of extra bits in rp[0] */
int inexact; /* return ternary flag */
mpfr_exp_t expr;
MPFR_TMP_DECL(marker);
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg rnd=%d", mpfr_get_prec (u), mpfr_log_prec, u, rnd_mode),
("y[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (r), mpfr_log_prec, r, inexact));
if (MPFR_UNLIKELY(MPFR_IS_SINGULAR(u)))
{
if (MPFR_IS_NAN(u))
{
MPFR_SET_NAN(r);
MPFR_RET_NAN;
}
else if (MPFR_IS_ZERO(u))
{
/* 0+ or 0- */
MPFR_SET_SAME_SIGN(r, u);
MPFR_SET_ZERO(r);
MPFR_RET(0); /* zero is exact */
}
else
{
MPFR_ASSERTD(MPFR_IS_INF(u));
/* sqrt(-Inf) = NAN */
if (MPFR_IS_NEG(u))
{
MPFR_SET_NAN(r);
MPFR_RET_NAN;
}
MPFR_SET_POS(r);
MPFR_SET_INF(r);
MPFR_RET(0);
}
}
if (MPFR_UNLIKELY(MPFR_IS_NEG(u)))
{
MPFR_SET_NAN(r);
MPFR_RET_NAN;
}
MPFR_SET_POS(r);
MPFR_TMP_MARK (marker);
MPFR_UNSIGNED_MINUS_MODULO(sh,MPFR_PREC(r));
if (sh == 0 && rnd_mode == MPFR_RNDN)
sh = GMP_NUMB_BITS; /* ugly case */
rsize = MPFR_LIMB_SIZE(r) + (sh == GMP_NUMB_BITS);
/* rsize is the number of limbs of r + 1 if exact limb multiple and rounding
to nearest, this is the number of wanted limbs for the square root */
rrsize = rsize + rsize;
usize = MPFR_LIMB_SIZE(u); /* number of limbs of u */
rp0 = MPFR_MANT(r);
rp = (sh < GMP_NUMB_BITS) ? rp0 : MPFR_TMP_LIMBS_ALLOC (rsize);
up = MPFR_MANT(u);
sticky0 = MPFR_LIMB_ZERO; /* truncated part of input */
sticky1 = MPFR_LIMB_ZERO; /* truncated part of rp[0] */
odd_exp = (unsigned int) MPFR_GET_EXP (u) & 1;
inexact = -1; /* return ternary flag */
sp = MPFR_TMP_LIMBS_ALLOC (rrsize);
/* copy the most significant limbs of u to {sp, rrsize} */
if (MPFR_LIKELY(usize <= rrsize)) /* in case r and u have the same precision,
we have indeed rrsize = 2 * usize */
{
k = rrsize - usize;
if (MPFR_LIKELY(k))
MPN_ZERO (sp, k);
if (odd_exp)
{
if (MPFR_LIKELY(k))
sp[k - 1] = mpn_rshift (sp + k, up, usize, 1);
else
sticky0 = mpn_rshift (sp, up, usize, 1);
}
else
MPN_COPY (sp + rrsize - usize, up, usize);
}
else /* usize > rrsize: truncate the input */
{
k = usize - rrsize;
if (odd_exp)
sticky0 = mpn_rshift (sp, up + k, rrsize, 1);
else
MPN_COPY (sp, up + k, rrsize);
l = k;
while (sticky0 == MPFR_LIMB_ZERO && l != 0)
sticky0 = up[--l];
}
/* sticky0 is non-zero iff the truncated part of the input is non-zero */
/* mpn_rootrem with NULL 2nd argument is faster than mpn_sqrtrem, thus use
it if available and if the user asked to use GMP internal functions */
#if defined(WANT_GMP_INTERNALS) && defined(HAVE___GMPN_ROOTREM)
tsize = __gmpn_rootrem (rp, NULL, sp, rrsize, 2);
#else
tsize = mpn_sqrtrem (rp, NULL, sp, rrsize);
#endif
/* a return value of zero in mpn_sqrtrem indicates a perfect square */
sticky = sticky0 || tsize != 0;
/* truncate low bits of rp[0] */
sticky1 = rp[0] & ((sh < GMP_NUMB_BITS) ? MPFR_LIMB_MASK(sh)
: ~MPFR_LIMB_ZERO);
rp[0] -= sticky1;
sticky = sticky || sticky1;
expr = (MPFR_GET_EXP(u) + odd_exp) / 2; /* exact */
if (rnd_mode == MPFR_RNDZ || rnd_mode == MPFR_RNDD || sticky == MPFR_LIMB_ZERO)
{
inexact = (sticky == MPFR_LIMB_ZERO) ? 0 : -1;
goto truncate;
}
else if (rnd_mode == MPFR_RNDN)
{
/* if sh < GMP_NUMB_BITS, the round bit is bit (sh-1) of sticky1
and the sticky bit is formed by the low sh-1 bits from
sticky1, together with the sqrtrem remainder and sticky0. */
if (sh < GMP_NUMB_BITS)
{
if (sticky1 & (MPFR_LIMB_ONE << (sh - 1)))
{ /* round bit is set */
if (sticky1 == (MPFR_LIMB_ONE << (sh - 1)) && tsize == 0
&& sticky0 == 0)
goto even_rule;
else
goto add_one_ulp;
}
else /* round bit is zero */
goto truncate; /* with the default inexact=-1 */
}
else /* sh = GMP_NUMB_BITS: the round bit is the most significant bit
of rp[0], and the remaining GMP_NUMB_BITS-1 bits contribute to
the sticky bit */
{
if (sticky1 & MPFR_LIMB_HIGHBIT)
{ /* round bit is set */
if (sticky1 == MPFR_LIMB_HIGHBIT && tsize == 0 && sticky0 == 0)
goto even_rule;
else
goto add_one_ulp;
}
else /* round bit is zero */
goto truncate; /* with the default inexact=-1 */
}
}
else /* rnd_mode=GMP_RDNU, necessarily sticky <> 0, thus add 1 ulp */
goto add_one_ulp;
even_rule: /* has to set inexact */
if (sh < GMP_NUMB_BITS)
inexact = (rp[0] & (MPFR_LIMB_ONE << sh)) ? 1 : -1;
else
inexact = (rp[1] & MPFR_LIMB_ONE) ? 1 : -1;
if (inexact == -1)
goto truncate;
/* else go through add_one_ulp */
add_one_ulp:
inexact = 1; /* always here */
if (sh == GMP_NUMB_BITS)
{
rp ++;
rsize --;
sh = 0;
}
if (mpn_add_1 (rp0, rp, rsize, MPFR_LIMB_ONE << sh))
{
expr ++;
rp[rsize - 1] = MPFR_LIMB_HIGHBIT;
}
goto end;
truncate: /* inexact = 0 or -1 */
if (sh == GMP_NUMB_BITS)
MPN_COPY (rp0, rp + 1, rsize - 1);
end:
MPFR_ASSERTN (expr >= MPFR_EMIN_MIN && expr <= MPFR_EMAX_MAX);
MPFR_EXP (r) = expr;
MPFR_TMP_FREE(marker);
return mpfr_check_range (r, inexact, rnd_mode);
}
void
mpq_set_f (mpq_ptr q, mpf_srcptr f)
{
mp_size_t fexp = EXP(f);
mp_ptr fptr = PTR(f);
mp_size_t fsize = SIZ(f);
mp_size_t abs_fsize = ABS(fsize);
mp_limb_t flow;
if (fsize == 0)
{
/* set q=0 */
SIZ(NUM(q)) = 0;
SIZ(DEN(q)) = 1;
PTR(DEN(q))[0] = 1;
return;
}
/* strip low zero limbs from f */
flow = *fptr;
MPN_STRIP_LOW_ZEROS_NOT_ZERO (fptr, abs_fsize, flow);
if (fexp >= abs_fsize)
{
/* radix point is to the right of the limbs, no denominator */
mp_ptr num_ptr;
num_ptr = MPZ_NEWALLOC (mpq_numref (q), fexp);
MPN_ZERO (num_ptr, fexp - abs_fsize);
MPN_COPY (num_ptr + fexp - abs_fsize, fptr, abs_fsize);
SIZ(NUM(q)) = fsize >= 0 ? fexp : -fexp;
SIZ(DEN(q)) = 1;
PTR(DEN(q))[0] = 1;
}
else
{
/* radix point is within or to the left of the limbs, use denominator */
mp_ptr num_ptr, den_ptr;
mp_size_t den_size;
den_size = abs_fsize - fexp;
num_ptr = MPZ_NEWALLOC (mpq_numref (q), abs_fsize);
den_ptr = MPZ_NEWALLOC (mpq_denref (q), den_size+1);
if (flow & 1)
{
/* no powers of two to strip from numerator */
MPN_COPY (num_ptr, fptr, abs_fsize);
MPN_ZERO (den_ptr, den_size);
den_ptr[den_size] = 1;
}
else
{
/* right shift numerator, adjust denominator accordingly */
int shift;
den_size--;
count_trailing_zeros (shift, flow);
mpn_rshift (num_ptr, fptr, abs_fsize, shift);
abs_fsize -= (num_ptr[abs_fsize-1] == 0);
MPN_ZERO (den_ptr, den_size);
den_ptr[den_size] = GMP_LIMB_HIGHBIT >> (shift-1);
}
SIZ(NUM(q)) = fsize >= 0 ? abs_fsize : -abs_fsize;
SIZ(DEN(q)) = den_size + 1;
}
}
static unsigned long int
lc (mp_ptr rp, gmp_randstate_t rstate)
{
mp_ptr tp, seedp, ap;
mp_size_t ta;
mp_size_t tn, seedn, an;
unsigned long int m2exp;
unsigned long int bits;
int cy;
mp_size_t xn;
gmp_rand_lc_struct *p;
TMP_DECL;
p = (gmp_rand_lc_struct *) RNG_STATE (rstate);
m2exp = p->_mp_m2exp;
seedp = PTR (p->_mp_seed);
seedn = SIZ (p->_mp_seed);
ap = PTR (p->_mp_a);
an = SIZ (p->_mp_a);
/* Allocate temporary storage. Let there be room for calculation of
(A * seed + C) % M, or M if bigger than that. */
TMP_MARK;
ta = an + seedn + 1;
tn = BITS_TO_LIMBS (m2exp);
if (ta <= tn) /* that is, if (ta < tn + 1) */
{
mp_size_t tmp = an + seedn;
ta = tn + 1;
tp = (mp_ptr) TMP_ALLOC (ta * BYTES_PER_MP_LIMB);
MPN_ZERO (&tp[tmp], ta - tmp); /* mpn_mul won't zero it out. */
}
else
tp = (mp_ptr) TMP_ALLOC (ta * BYTES_PER_MP_LIMB);
/* t = a * seed. NOTE: an is always > 0; see initialization. */
ASSERT (seedn >= an && an > 0);
mpn_mul (tp, seedp, seedn, ap, an);
/* t = t + c. NOTE: tn is always >= p->_cn (precondition for __GMPN_ADD);
see initialization. */
ASSERT (tn >= p->_cn);
__GMPN_ADD (cy, tp, tp, tn, p->_cp, p->_cn);
/* t = t % m */
tp[m2exp / GMP_NUMB_BITS] &= (CNST_LIMB (1) << m2exp % GMP_NUMB_BITS) - 1;
/* Save result as next seed. */
MPN_COPY (PTR (p->_mp_seed), tp, tn);
/* Discard the lower m2exp/2 of the result. */
bits = m2exp / 2;
xn = bits / GMP_NUMB_BITS;
tn -= xn;
if (tn > 0)
{
unsigned int cnt = bits % GMP_NUMB_BITS;
if (cnt != 0)
{
mpn_rshift (tp, tp + xn, tn, cnt);
MPN_COPY_INCR (rp, tp, xn + 1);
}
else /* Even limb boundary. */
MPN_COPY_INCR (rp, tp + xn, tn);
}
TMP_FREE;
/* Return number of valid bits in the result. */
return (m2exp + 1) / 2;
}
/* Computes {rp,MIN(rn,an+bn)} <- {ap,an}*{bp,bn} Mod(B^rn-1)
*
* The result is expected to be ZERO if and only if one of the operand
* already is. Otherwise the class [0] Mod(B^rn-1) is represented by
* B^rn-1. This should not be a problem if mulmod_bnm1 is used to
* combine results and obtain a natural number when one knows in
* advance that the final value is less than (B^rn-1).
* Moreover it should not be a problem if mulmod_bnm1 is used to
* compute the full product with an+bn <= rn, because this condition
* implies (B^an-1)(B^bn-1) < (B^rn-1) .
*
* Requires 0 < bn <= an <= rn and an + bn > rn/2
* Scratch need: rn + (need for recursive call OR rn + 4). This gives
*
* S(n) <= rn + MAX (rn + 4, S(n/2)) <= 2rn + 4
*/
void
mpn_mulmod_bnm1 (mp_ptr rp, mp_size_t rn, mp_srcptr ap, mp_size_t an, mp_srcptr bp, mp_size_t bn, mp_ptr tp)
{
ASSERT (0 < bn);
ASSERT (bn <= an);
ASSERT (an <= rn);
if ((rn & 1) != 0 || BELOW_THRESHOLD (rn, MULMOD_BNM1_THRESHOLD))
{
if (UNLIKELY (bn < rn))
{
if (UNLIKELY (an + bn <= rn))
{
mpn_mul (rp, ap, an, bp, bn);
}
else
{
mp_limb_t cy;
mpn_mul (tp, ap, an, bp, bn);
cy = mpn_add (rp, tp, rn, tp + rn, an + bn - rn);
MPN_INCR_U (rp, rn, cy);
}
}
else
mpn_bc_mulmod_bnm1 (rp, ap, bp, rn, tp);
}
else
{
mp_size_t n;
mp_limb_t cy;
mp_limb_t hi;
n = rn >> 1;
/* We need at least an + bn >= n, to be able to fit one of the
recursive products at rp. Requiring strict inequality makes
the coded slightly simpler. If desired, we could avoid this
restriction by initially halving rn as long as rn is even and
an + bn <= rn/2. */
ASSERT (an + bn > n);
/* Compute xm = a*b mod (B^n - 1), xp = a*b mod (B^n + 1)
and crt together as
x = -xp * B^n + (B^n + 1) * [ (xp + xm)/2 mod (B^n-1)]
*/
#define a0 ap
#define a1 (ap + n)
#define b0 bp
#define b1 (bp + n)
#define xp tp /* 2n + 2 */
/* am1 maybe in {xp, n} */
/* bm1 maybe in {xp + n, n} */
#define sp1 (tp + 2*n + 2)
/* ap1 maybe in {sp1, n + 1} */
/* bp1 maybe in {sp1 + n + 1, n + 1} */
{
mp_srcptr am1, bm1;
mp_size_t anm, bnm;
mp_ptr so;
bm1 = b0;
bnm = bn;
if (LIKELY (an > n))
{
am1 = xp;
cy = mpn_add (xp, a0, n, a1, an - n);
MPN_INCR_U (xp, n, cy);
anm = n;
so = xp + n;
if (LIKELY (bn > n))
{
bm1 = so;
cy = mpn_add (so, b0, n, b1, bn - n);
MPN_INCR_U (so, n, cy);
bnm = n;
so += n;
}
}
else
{
so = xp;
am1 = a0;
anm = an;
}
mpn_mulmod_bnm1 (rp, n, am1, anm, bm1, bnm, so);
}
{
int k;
mp_srcptr ap1, bp1;
mp_size_t anp, bnp;
bp1 = b0;
bnp = bn;
if (LIKELY (an > n)) {
ap1 = sp1;
cy = mpn_sub (sp1, a0, n, a1, an - n);
sp1[n] = 0;
MPN_INCR_U (sp1, n + 1, cy);
anp = n + ap1[n];
if (LIKELY (bn > n)) {
bp1 = sp1 + n + 1;
cy = mpn_sub (sp1 + n + 1, b0, n, b1, bn - n);
sp1[2*n+1] = 0;
MPN_INCR_U (sp1 + n + 1, n + 1, cy);
bnp = n + bp1[n];
}
} else {
ap1 = a0;
anp = an;
}
if (BELOW_THRESHOLD (n, MUL_FFT_MODF_THRESHOLD))
k=0;
else
{
int mask;
k = mpn_fft_best_k (n, 0);
mask = (1<<k) - 1;
while (n & mask) {k--; mask >>=1;};
}
if (k >= FFT_FIRST_K)
xp[n] = mpn_mul_fft (xp, n, ap1, anp, bp1, bnp, k);
else if (UNLIKELY (bp1 == b0))
{
ASSERT (anp + bnp <= 2*n+1);
ASSERT (anp + bnp > n);
ASSERT (anp >= bnp);
mpn_mul (xp, ap1, anp, bp1, bnp);
anp = anp + bnp - n;
ASSERT (anp <= n || xp[2*n]==0);
anp-= anp > n;
cy = mpn_sub (xp, xp, n, xp + n, anp);
xp[n] = 0;
MPN_INCR_U (xp, n+1, cy);
}
else
mpn_bc_mulmod_bnp1 (xp, ap1, bp1, n, xp);
}
/* Here the CRT recomposition begins.
xm <- (xp + xm)/2 = (xp + xm)B^n/2 mod (B^n-1)
Division by 2 is a bitwise rotation.
Assumes xp normalised mod (B^n+1).
The residue class [0] is represented by [B^n-1]; except when
both input are ZERO.
*/
#if HAVE_NATIVE_mpn_rsh1add_n || HAVE_NATIVE_mpn_rsh1add_nc
#if HAVE_NATIVE_mpn_rsh1add_nc
cy = mpn_rsh1add_nc(rp, rp, xp, n, xp[n]); /* B^n = 1 */
hi = cy << (GMP_NUMB_BITS - 1);
cy = 0;
/* next update of rp[n-1] will set cy = 1 only if rp[n-1]+=hi
overflows, i.e. a further increment will not overflow again. */
#else /* ! _nc */
cy = xp[n] + mpn_rsh1add_n(rp, rp, xp, n); /* B^n = 1 */
hi = (cy<<(GMP_NUMB_BITS-1))&GMP_NUMB_MASK; /* (cy&1) << ... */
cy >>= 1;
/* cy = 1 only if xp[n] = 1 i.e. {xp,n} = ZERO, this implies that
the rsh1add was a simple rshift: the top bit is 0. cy=1 => hi=0. */
#endif
#if GMP_NAIL_BITS == 0
add_ssaaaa(cy, rp[n-1], cy, rp[n-1], 0, hi);
#else
cy += (hi & rp[n-1]) >> (GMP_NUMB_BITS-1);
rp[n-1] ^= hi;
#endif
#else /* ! HAVE_NATIVE_mpn_rsh1add_n */
#if HAVE_NATIVE_mpn_add_nc
cy = mpn_add_nc(rp, rp, xp, n, xp[n]);
#else /* ! _nc */
cy = xp[n] + mpn_add_n(rp, rp, xp, n); /* xp[n] == 1 implies {xp,n} == ZERO */
#endif
cy += (rp[0]&1);
mpn_rshift(rp, rp, n, 1);
ASSERT (cy <= 2);
hi = (cy<<(GMP_NUMB_BITS-1))&GMP_NUMB_MASK; /* (cy&1) << ... */
cy >>= 1;
/* We can have cy != 0 only if hi = 0... */
ASSERT ((rp[n-1] & GMP_NUMB_HIGHBIT) == 0);
rp[n-1] |= hi;
/* ... rp[n-1] + cy can not overflow, the following INCR is correct. */
#endif
ASSERT (cy <= 1);
/* Next increment can not overflow, read the previous comments about cy. */
ASSERT ((cy == 0) || ((rp[n-1] & GMP_NUMB_HIGHBIT) == 0));
MPN_INCR_U(rp, n, cy);
/* Compute the highest half:
([(xp + xm)/2 mod (B^n-1)] - xp ) * B^n
*/
if (UNLIKELY (an + bn < rn))
{
/* Note that in this case, the only way the result can equal
zero mod B^{rn} - 1 is if one of the inputs is zero, and
then the output of both the recursive calls and this CRT
reconstruction is zero, not B^{rn} - 1. Which is good,
since the latter representation doesn't fit in the output
area.*/
cy = mpn_sub_n (rp + n, rp, xp, an + bn - n);
/* FIXME: This subtraction of the high parts is not really
necessary, we do it to get the carry out, and for sanity
checking. */
cy = xp[n] + mpn_sub_nc (xp + an + bn - n, rp + an + bn - n,
xp + an + bn - n, rn - (an + bn), cy);
ASSERT (an + bn == rn - 1 ||
mpn_zero_p (xp + an + bn - n + 1, rn - 1 - (an + bn)));
cy = mpn_sub_1 (rp, rp, an + bn, cy);
ASSERT (cy == (xp + an + bn - n)[0]);
}
else
{
cy = xp[n] + mpn_sub_n (rp + n, rp, xp, n);
/* cy = 1 only if {xp,n+1} is not ZERO, i.e. {rp,n} is not ZERO.
DECR will affect _at most_ the lowest n limbs. */
MPN_DECR_U (rp, 2*n, cy);
}
#undef a0
#undef a1
#undef b0
#undef b1
#undef xp
#undef sp1
}
}
void _arb_exp_taylor_rs(mp_ptr y, mp_limb_t * error,
mp_srcptr x, mp_size_t xn, ulong N)
{
mp_ptr s, t, xpow;
mp_limb_t new_denom, old_denom, c;
slong power, k, m;
TMP_INIT;
TMP_START;
if (N >= FACTORIAL_TAB_SIZE - 1)
{
flint_printf("_arb_exp_taylor_rs: N too large!\n");
abort();
}
if (N <= 3)
{
if (N <= 1)
{
flint_mpn_zero(y, xn);
y[xn] = N;
error[0] = 0;
}
else if (N == 2)
{
flint_mpn_copyi(y, x, xn);
y[xn] = 1;
error[0] = 0;
}
else
{
/* 1 + x + x^2 / 2 */
t = TMP_ALLOC_LIMBS(2 * xn);
mpn_sqr(t, x, xn);
mpn_rshift(t + xn, t + xn, xn, 1);
y[xn] = mpn_add_n(y, x, t + xn, xn) + 1;
error[0] = 2;
}
}
else
{
/* Choose m ~= sqrt(num_terms) (m must be even, >= 2) */
/* TODO: drop evenness assumption since we don't have sign issues here? */
/* TODO: then just need to fix power construction below... */
m = 2;
while (m * m < N)
m += 2;
/* todo: merge allocations */
xpow = TMP_ALLOC_LIMBS((m + 1) * xn);
s = TMP_ALLOC_LIMBS(xn + 2);
t = TMP_ALLOC_LIMBS(2 * xn + 2); /* todo: 1 limb too much? */
/* higher index ---> */
/* | ---xn--- | */
/* xpow = | <temp> | x^m | x^(m-1) | ... | x^2 | x | */
#define XPOW_WRITE(__k) (xpow + (m - (__k)) * xn)
#define XPOW_READ(__k) (xpow + (m - (__k) + 1) * xn)
flint_mpn_copyi(XPOW_READ(1), x, xn);
mpn_sqr(XPOW_WRITE(2), XPOW_READ(1), xn);
for (k = 4; k <= m; k += 2)
{
mpn_mul_n(XPOW_WRITE(k - 1), XPOW_READ(k / 2), XPOW_READ(k / 2 - 1), xn);
mpn_sqr(XPOW_WRITE(k), XPOW_READ(k / 2), xn);
}
flint_mpn_zero(s, xn + 1);
/* todo: skip one nonscalar multiplication (use x^m)
when starting on x^0 */
power = (N - 1) % m;
for (k = N - 1; k >= 0; k--)
{
c = factorial_tab_numer[k];
new_denom = factorial_tab_denom[k];
old_denom = factorial_tab_denom[k+1];
/* change denominators */
if (new_denom != old_denom && k < N - 1)
{
mpn_divrem_1(s, 0, s, xn + 1, old_denom);
}
if (power == 0)
{
/* add c * x^0 -- only top limb is affected */
s[xn] += c;
/* Outer polynomial evaluation: multiply by x^m */
if (k != 0)
{
mpn_mul(t, s, xn + 1, XPOW_READ(m), xn);
flint_mpn_copyi(s, t + xn, xn + 1);
}
power = m - 1;
}
else
{
s[xn] += mpn_addmul_1(s, XPOW_READ(power), xn, c);
power--;
}
}
/* finally divide by denominator */
mpn_divrem_1(y, 0, s, xn + 1, factorial_tab_denom[0]);
/* error bound (ulp) */
error[0] = 2;
}
TMP_END;
}
/* Since MPFR-3.0, return the usual inexact value.
The erange flag is set if an error occurred in the conversion
(y is NaN, +Inf, or -Inf that have no equivalent in mpf)
*/
int
mpfr_get_f (mpf_ptr x, mpfr_srcptr y, mpfr_rnd_t rnd_mode)
{
int inex;
mp_size_t sx, sy;
mpfr_prec_t precx, precy;
mp_limb_t *xp;
int sh;
if (MPFR_UNLIKELY(MPFR_IS_SINGULAR(y)))
{
if (MPFR_IS_ZERO(y))
{
mpf_set_ui (x, 0);
return 0;
}
else if (MPFR_IS_NAN (y))
{
MPFR_SET_ERANGEFLAG ();
return 0;
}
else /* y is plus infinity (resp. minus infinity), set x to the maximum
value (resp. the minimum value) in precision PREC(x) */
{
int i;
mp_limb_t *xp;
MPFR_SET_ERANGEFLAG ();
/* To this day, [mp_exp_t] and mp_size_t are #defined as the same
type */
EXP (x) = MP_SIZE_T_MAX;
sx = PREC (x);
SIZ (x) = sx;
xp = PTR (x);
for (i = 0; i < sx; i++)
xp[i] = MPFR_LIMB_MAX;
if (MPFR_IS_POS (y))
return -1;
else
{
mpf_neg (x, x);
return +1;
}
}
}
sx = PREC(x); /* number of limbs of the mantissa of x */
precy = MPFR_PREC(y);
precx = (mpfr_prec_t) sx * GMP_NUMB_BITS;
sy = MPFR_LIMB_SIZE (y);
xp = PTR (x);
/* since mpf numbers are represented in base 2^GMP_NUMB_BITS,
we loose -EXP(y) % GMP_NUMB_BITS bits in the most significant limb */
sh = MPFR_GET_EXP(y) % GMP_NUMB_BITS;
sh = sh <= 0 ? - sh : GMP_NUMB_BITS - sh;
MPFR_ASSERTD (sh >= 0);
if (precy + sh <= precx) /* we can copy directly */
{
mp_size_t ds;
MPFR_ASSERTN (sx >= sy);
ds = sx - sy;
if (sh != 0)
{
mp_limb_t out;
out = mpn_rshift (xp + ds, MPFR_MANT(y), sy, sh);
MPFR_ASSERTN (ds > 0 || out == 0);
if (ds > 0)
xp[--ds] = out;
}
else
MPN_COPY (xp + ds, MPFR_MANT (y), sy);
if (ds > 0)
MPN_ZERO (xp, ds);
EXP(x) = (MPFR_GET_EXP(y) + sh) / GMP_NUMB_BITS;
inex = 0;
}
else /* we have to round to precx - sh bits */
{
mpfr_t z;
mp_size_t sz;
/* Recall that precx = (mpfr_prec_t) sx * GMP_NUMB_BITS, thus removing
sh bits (sh < GMP_NUMB_BITSS) won't reduce the number of limbs. */
mpfr_init2 (z, precx - sh);
sz = MPFR_LIMB_SIZE (z);
MPFR_ASSERTN (sx == sz);
inex = mpfr_set (z, y, rnd_mode);
/* warning, sh may change due to rounding, but then z is a power of two,
thus we can safely ignore its last bit which is 0 */
sh = MPFR_GET_EXP(z) % GMP_NUMB_BITS;
sh = sh <= 0 ? - sh : GMP_NUMB_BITS - sh;
MPFR_ASSERTD (sh >= 0);
if (sh != 0)
{
mp_limb_t out;
out = mpn_rshift (xp, MPFR_MANT(z), sz, sh);
/* If sh hasn't changed, it is the number of the non-significant
bits in the lowest limb of z. Therefore out == 0. */
MPFR_ASSERTD (out == 0); (void) out; /* avoid a warning */
}
else
MPN_COPY (xp, MPFR_MANT(z), sz);
EXP(x) = (MPFR_GET_EXP(z) + sh) / GMP_NUMB_BITS;
mpfr_clear (z);
}
/* set size and sign */
SIZ(x) = (MPFR_FROM_SIGN_TO_INT(MPFR_SIGN(y)) < 0) ? -sx : sx;
return inex;
}
/* put in {c, 2n} where n = 2k+r the value of {v0,2k} (already in place)
+ B^k * [{v1, 2k+1} - {t1, 2k+1}]
+ B^(2k) * [{t2, 2k+1} - {v0+vinf, 2k}]
+ B^(3k) * [{t1, 2k+1} - {t2, 2k+1}]
+ B^(4k) * {vinf,2r} (high 2r-1 limbs already in place)
where {t1, 2k+1} = (3*{v0,2k}+2*sa*{vm1,2k+1}+{v2,2k+1})/6-2*{vinf,2r}
{t2, 2k+1} = ({v1, 2k+1} + sa * {vm1, 2k+1})/2
(sa is the sign of {vm1, 2k+1}).
{vinf, 2r} stores the content of {v0, 2r} + {vinf, 2r}, with carry in cinf0.
vinf0 is the low limb of vinf.
ws is temporary space, and should have at least 2r limbs.
Think about:
The evaluated point a-b+c stands a good chance of having a zero carry
limb, a+b+c would have a 1/4 chance, and 4*a+2*b+c a 1/8 chance, roughly.
Perhaps this could be tested and stripped. Doing so before recursing
would be better than stripping at the start of mpn_toom3_mul_n/sqr_n,
since then the recursion could be based on the new size. Although in
truth the kara vs toom3 crossover is never so exact that one limb either
way makes a difference.
A small value like 1 or 2 for the carry could perhaps also be handled
with an add_n or addlsh1_n. Would that be faster than an extra limb on a
(recursed) multiply/square?
*/
static void
toom3_interpolate (mp_ptr c, mp_srcptr v1, mp_ptr v2, mp_ptr vm1,
mp_ptr vinf, mp_size_t k, mp_size_t r, int sa,
mp_limb_t vinf0, mp_limb_t cinf0, mp_ptr ws)
{
mp_limb_t cy, saved;
unsigned long twok = k + k;
unsigned long kk1 = twok + 1;
unsigned long twor = r + r;
mp_ptr c1, c2, c3, c4, c5;
mp_limb_t cout; /* final carry, should be zero at the end */
c1 = c + k;
c2 = c1 + k;
c3 = c2 + k;
c4 = c3 + k;
c5 = c4 + k;
#define v0 (c)
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 |vm1| hi(vinf) v1 v2+2vm1 vinf
+lo(v0) */
ASSERT_NOCARRY (mpn_divexact_by3 (v2, v2, kk1)); /* v2 <- v2 / 3 */
#ifdef HAVE_NATIVE_mpn_rsh1add_n
mpn_rsh1add_n (v2, v2, v0, twok); /* v2 <- (lo(v2)+v0) / 2, exact */
cy = v2[twok] & 1; /* add high limb of v2 divided by 2 */
v2[twok] >>= 1;
MPN_INCR_U (v2 + twok - 1, 2, cy << (GMP_NUMB_BITS - 1));
#else
v2[twok] += mpn_add_n (v2, v2, v0, twok);
mpn_rshift (v2, v2, kk1, 1);
#endif
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 |vm1| hi(vinf) v1 (3v0+2vm1+v2) vinf
/6 +lo(v0) */
/* vm1 <- t2 := (v1 + sa*vm1) / 2
t2 = a0*b0+a0*b2+a1*b1+a2*b0+a2*b2 >= 0
No carry comes out from {v1, kk1} +/- {vm1, kk1},
and the division by two is exact */
if (sa >= 0)
{
#ifdef HAVE_NATIVE_mpn_rsh1add_n
mpn_rsh1add_n (vm1, v1, vm1, kk1);
#else
mpn_add_n (vm1, vm1, v1, kk1);
mpn_rshift (vm1, vm1, kk1, 1);
#endif
}
else
{
#ifdef HAVE_NATIVE_mpn_rsh1sub_n
mpn_rsh1sub_n (vm1, v1, vm1, kk1);
#else
mpn_sub_n (vm1, v1, vm1, kk1);
mpn_rshift (vm1, vm1, kk1, 1);
#endif
}
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 t2 hi(vinf) v1 t1 vinf+lo(v0) */
/* subtract 2*vinf to v2,
result is t1 := a0*b0+a0*b2+a1*b1+a1*b2+a2*b0+a2*b1+a2*b2 >= 0 */
saved = c4[0];
c4[0] = vinf0;
#ifdef HAVE_NATIVE_mpn_sublsh1_n
cy = mpn_sublsh1_n (v2, v2, c4, twor);
#else
cy = mpn_lshift (ws, c4, twor, 1);
cy += mpn_sub_n (v2, v2, ws, twor);
#endif
MPN_DECR_U (v2 + twor, kk1 - twor, cy);
c4[0] = saved;
/* subtract {t2, 2k+1} in {c+3k, 2k+1} i.e. in {t2+k, 2k+1}:
by chunks of k limbs from right to left to avoid overlap */
#define t2 (vm1)
/* a borrow may occur in one of the 2 following __GMPN_SUB_1 calls, but since
the final result is nonnegative, it will be compensated later on */
__GMPN_SUB_1 (cout, c5, c5, twor - k, t2[twok]);
cy = mpn_sub_n (c4, c4, t2 + k, k);
__GMPN_SUB_1 (cout, c5, c5, twor - k, cy);
cy = mpn_sub_n (c3, c3, t2, k);
__GMPN_SUB_1 (cout, c4, c4, twor, cy);
/* don't forget to add vinf0 in {c+4k, ...} */
__GMPN_ADD_1 (cout, c4, c4, twor, vinf0);
/* c c+k c+2k c+3k c+4k+1 t t+2k+1 t+4k+2
v0 t2 hi(vinf) v1 t1 vinf
-t2 +lo(v0)
*/
/* c c+k c+2k c+3k c+4k t t+2k+1 t+4k+2
v0 t2 vinf v1 t1 vinf
-t2 +lo(v0)
*/
/* subtract v0+vinf in {c+2k, ...} */
cy = cinf0 + mpn_sub_n (c2, c2, vinf, twor);
if (twor < twok)
{
__GMPN_SUB_1 (cy, c2 + twor, c2 + twor, twok - twor, cy);
cy += mpn_sub_n (c2 + twor, c2 + twor, v0 + twor, twok - twor);
}
__GMPN_SUB_1 (cout, c4, c4, twor, cy); /* 2n-4k = 2r */
/* c c+k c+2k c+3k c+4k t t+2k+1 t+4k+2
v0 t2 vinf v1 t1 vinf
-v0 -t2 +lo(v0)
-vinf */
/* subtract t1 in {c+k, ...} */
cy = mpn_sub_n (c1, c1, v2, kk1);
__GMPN_SUB_1 (cout, c3 + 1, c3 + 1, twor + k - 1, cy); /* 2n-(3k+1)=k+2r-1 */
/* c c+k c+2k c+3k c+4k t t+2k+1 t+4k+2
v0 t2 vinf v1 t1 vinf
-t1 -v0 -t2
-vinf */
/* add t1 in {c+3k, ...} */
cy = mpn_add_n (c3, c3, v2, kk1);
__GMPN_ADD_1 (cout, c5 + 1, c5 + 1, twor - k - 1, cy); /* 2n-(5k+1) = 2r-k-1 */
/* c c+k c+2k c+3k c+4k t t+2k+1 t+4k+2
v0 t2 t1 vinf v1 t1 vinf
-t1 -v0 -t2
-vinf */
/* add v1 in {c+k, ...} */
cy = mpn_add_n (c1, c1, v1, kk1);
__GMPN_ADD_1 (cout, c3 + 1, c3 + 1, twor + k - 1, cy); /* 2n-(3k+1) = 2r+k-1 */
/* c c+k c+2k c+3k c+4k t t+2k+1 t+4k+2
v0 v1 t2 t1 vinf v1 t1 vinf
-t1 -v0 -t2
-vinf */
#undef v0
#undef t2
}
void
check_functions (void)
{
mp_limb_t wp[2], wp2[2], xp[2], yp[2], r;
int i;
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 123;
yp[0] = 456;
mpn_add_n (wp, xp, yp, (mp_size_t) 1);
ASSERT_ALWAYS (wp[0] == 579);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 123;
wp[0] = 456;
r = mpn_addmul_1 (wp, xp, (mp_size_t) 1, CNST_LIMB(2));
ASSERT_ALWAYS (wp[0] == 702);
ASSERT_ALWAYS (r == 0);
}
#if HAVE_NATIVE_mpn_copyd
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 123;
xp[1] = 456;
mpn_copyd (xp+1, xp, (mp_size_t) 1);
ASSERT_ALWAYS (xp[1] == 123);
}
#endif
#if HAVE_NATIVE_mpn_copyi
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 123;
xp[1] = 456;
mpn_copyi (xp, xp+1, (mp_size_t) 1);
ASSERT_ALWAYS (xp[0] == 456);
}
#endif
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 1605;
mpn_divexact_1 (wp, xp, (mp_size_t) 1, CNST_LIMB(5));
ASSERT_ALWAYS (wp[0] == 321);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 1296;
r = mpn_divexact_by3c (wp, xp, (mp_size_t) 1, CNST_LIMB(0));
ASSERT_ALWAYS (wp[0] == 432);
ASSERT_ALWAYS (r == 0);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 578;
r = mpn_divexact_byfobm1 (wp, xp, (mp_size_t) 1, CNST_LIMB(17),CNST_LIMB(-1)/CNST_LIMB(17));
ASSERT_ALWAYS (wp[0] == 34);
ASSERT_ALWAYS (r == 0);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 287;
r = mpn_divrem_1 (wp, (mp_size_t) 1, xp, (mp_size_t) 1, CNST_LIMB(7));
ASSERT_ALWAYS (wp[1] == 41);
ASSERT_ALWAYS (wp[0] == 0);
ASSERT_ALWAYS (r == 0);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 290;
r = mpn_divrem_euclidean_qr_1 (wp, 0, xp, (mp_size_t) 1, CNST_LIMB(7));
ASSERT_ALWAYS (wp[0] == 41);
ASSERT_ALWAYS (r == 3);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 12;
r = mpn_gcd_1 (xp, (mp_size_t) 1, CNST_LIMB(9));
ASSERT_ALWAYS (r == 3);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 0x1001;
mpn_lshift (wp, xp, (mp_size_t) 1, 1);
ASSERT_ALWAYS (wp[0] == 0x2002);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 14;
r = mpn_mod_1 (xp, (mp_size_t) 1, CNST_LIMB(4));
ASSERT_ALWAYS (r == 2);
}
#if (GMP_NUMB_BITS % 4) == 0
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
int bits = (GMP_NUMB_BITS / 4) * 3;
mp_limb_t mod = (CNST_LIMB(1) << bits) - 1;
mp_limb_t want = GMP_NUMB_MAX % mod;
xp[0] = GMP_NUMB_MAX;
r = mpn_mod_34lsub1 (xp, (mp_size_t) 1);
ASSERT_ALWAYS (r % mod == want);
}
#endif
// DECL_modexact_1c_odd ((*modexact_1c_odd));
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 14;
r = mpn_mul_1 (wp, xp, (mp_size_t) 1, CNST_LIMB(4));
ASSERT_ALWAYS (wp[0] == 56);
ASSERT_ALWAYS (r == 0);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 5;
yp[0] = 7;
mpn_mul_basecase (wp, xp, (mp_size_t) 1, yp, (mp_size_t) 1);
ASSERT_ALWAYS (wp[0] == 35);
ASSERT_ALWAYS (wp[1] == 0);
}
#if HAVE_NATIVE_mpn_preinv_divrem_1 && GMP_NAIL_BITS == 0
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 0x101;
r = mpn_preinv_divrem_1 (wp, (mp_size_t) 1, xp, (mp_size_t) 1,
GMP_LIMB_HIGHBIT,
refmpn_invert_limb (GMP_LIMB_HIGHBIT), 0);
ASSERT_ALWAYS (wp[0] == 0x202);
ASSERT_ALWAYS (wp[1] == 0);
ASSERT_ALWAYS (r == 0);
}
#endif
#if GMP_NAIL_BITS == 0
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = GMP_LIMB_HIGHBIT+123;
r = mpn_preinv_mod_1 (xp, (mp_size_t) 1, GMP_LIMB_HIGHBIT,
refmpn_invert_limb (GMP_LIMB_HIGHBIT));
ASSERT_ALWAYS (r == 123);
}
#endif
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 5;
modlimb_invert(r,xp[0]);
r=-r;
yp[0]=43;
yp[1]=75;
mpn_redc_1 (wp, yp, xp, (mp_size_t) 1,r);
ASSERT_ALWAYS (wp[0] == 78);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0]=5;
yp[0]=3;
mpn_sumdiff_n (wp, wp2,xp, yp,1);
ASSERT_ALWAYS (wp[0] == 8);
ASSERT_ALWAYS (wp2[0] == 2);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 0x8008;
mpn_rshift (wp, xp, (mp_size_t) 1, 1);
ASSERT_ALWAYS (wp[0] == 0x4004);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 5;
mpn_sqr_basecase (wp, xp, (mp_size_t) 1);
ASSERT_ALWAYS (wp[0] == 25);
ASSERT_ALWAYS (wp[1] == 0);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 999;
yp[0] = 666;
mpn_sub_n (wp, xp, yp, (mp_size_t) 1);
ASSERT_ALWAYS (wp[0] == 333);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 123;
wp[0] = 456;
r = mpn_submul_1 (wp, xp, (mp_size_t) 1, CNST_LIMB(2));
ASSERT_ALWAYS (wp[0] == 210);
ASSERT_ALWAYS (r == 0);
}
}
void
mpn_toom_interpolate_5pts (mp_ptr c, mp_ptr v2, mp_ptr vm1,
mp_size_t k, mp_size_t twor, int sa,
mp_limb_t vinf0)
{
mp_limb_t cy, saved;
mp_size_t twok;
mp_size_t kk1;
mp_ptr c1, v1, c3, vinf;
twok = k + k;
kk1 = twok + 1;
c1 = c + k;
v1 = c1 + k;
c3 = v1 + k;
vinf = c3 + k;
#define v0 (c)
/* (1) v2 <- v2-vm1 < v2+|vm1|, (16 8 4 2 1) - (1 -1 1 -1 1) =
thus 0 <= v2 < 50*B^(2k) < 2^6*B^(2k) (15 9 3 3 0)
*/
if (sa)
ASSERT_NOCARRY (mpn_add_n (v2, v2, vm1, kk1));
else
ASSERT_NOCARRY (mpn_sub_n (v2, v2, vm1, kk1));
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 v1 hi(vinf) |vm1| v2-vm1 EMPTY */
ASSERT_NOCARRY (mpn_divexact_by3 (v2, v2, kk1)); /* v2 <- v2 / 3 */
/* (5 3 1 1 0)*/
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 v1 hi(vinf) |vm1| (v2-vm1)/3 EMPTY */
/* (2) vm1 <- tm1 := (v1 - vm1) / 2 [(1 1 1 1 1) - (1 -1 1 -1 1)] / 2 =
tm1 >= 0 (0 1 0 1 0)
No carry comes out from {v1, kk1} +/- {vm1, kk1},
and the division by two is exact.
If (sa!=0) the sign of vm1 is negative */
if (sa)
{
#ifdef HAVE_NATIVE_mpn_rsh1add_n
mpn_rsh1add_n (vm1, v1, vm1, kk1);
#else
ASSERT_NOCARRY (mpn_add_n (vm1, v1, vm1, kk1));
ASSERT_NOCARRY (mpn_rshift (vm1, vm1, kk1, 1));
#endif
}
else
{
#ifdef HAVE_NATIVE_mpn_rsh1sub_n
mpn_rsh1sub_n (vm1, v1, vm1, kk1);
#else
ASSERT_NOCARRY (mpn_sub_n (vm1, v1, vm1, kk1));
ASSERT_NOCARRY (mpn_rshift (vm1, vm1, kk1, 1));
#endif
}
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 v1 hi(vinf) tm1 (v2-vm1)/3 EMPTY */
/* (3) v1 <- t1 := v1 - v0 (1 1 1 1 1) - (0 0 0 0 1) = (1 1 1 1 0)
t1 >= 0
*/
vinf[0] -= mpn_sub_n (v1, v1, c, twok);
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 v1-v0 hi(vinf) tm1 (v2-vm1)/3 EMPTY */
/* (4) v2 <- t2 := ((v2-vm1)/3-t1)/2 = (v2-vm1-3*t1)/6
t2 >= 0 [(5 3 1 1 0) - (1 1 1 1 0)]/2 = (2 1 0 0 0)
*/
#ifdef HAVE_NATIVE_mpn_rsh1sub_n
mpn_rsh1sub_n (v2, v2, v1, kk1);
#else
ASSERT_NOCARRY (mpn_sub_n (v2, v2, v1, kk1));
ASSERT_NOCARRY (mpn_rshift (v2, v2, kk1, 1));
#endif
/* {c,2k} {c+2k,2k+1} {c+4k+1,2r-1} {t,2k+1} {t+2k+1,2k+1} {t+4k+2,2r}
v0 v1-v0 hi(vinf) tm1 (v2-vm1-3t1)/6 EMPTY */
/* (5) v1 <- t1-tm1 (1 1 1 1 0) - (0 1 0 1 0) = (1 0 1 0 0)
result is v1 >= 0
*/
ASSERT_NOCARRY (mpn_sub_n (v1, v1, vm1, kk1));
/* We do not need to read the value in vm1, so we add it in {c+k, ...} */
cy = mpn_add_n (c1, c1, vm1, kk1);
MPN_INCR_U (c3 + 1, twor + k - 1, cy); /* 2n-(3k+1) = 2r+k-1 */
/* Memory allocated for vm1 is now free, it can be recycled ...*/
/* (6) v2 <- v2 - 2*vinf, (2 1 0 0 0) - 2*(1 0 0 0 0) = (0 1 0 0 0)
result is v2 >= 0 */
saved = vinf[0]; /* Remember v1's highest byte (will be overwritten). */
vinf[0] = vinf0; /* Set the right value for vinf0 */
#ifdef HAVE_NATIVE_mpn_sublsh1_n
cy = mpn_sublsh1_n (v2, v2, vinf, twor);
#else
/* Overwrite unused vm1 */
cy = mpn_lshift (vm1, vinf, twor, 1);
cy += mpn_sub_n (v2, v2, vm1, twor);
#endif
MPN_DECR_U (v2 + twor, kk1 - twor, cy);
/* Current matrix is
[1 0 0 0 0; vinf
0 1 0 0 0; v2
1 0 1 0 0; v1
0 1 0 1 0; vm1
0 0 0 0 1] v0
Some vaues already are in-place (we added vm1 in the correct position)
| vinf| v1 | v0 |
| vm1 |
One still is in a separated area
| +v2 |
We have to compute v1-=vinf; vm1 -= v2,
|-vinf|
| -v2 |
Carefully reordering operations we can avoid to compute twice the sum
of the high half of v2 plus the low half of vinf.
*/
/* Add the high half of t2 in {vinf} */
if ( LIKELY(twor > k + 1) ) { /* This is the expected flow */
cy = mpn_add_n (vinf, vinf, v2 + k, k + 1);
MPN_INCR_U (c3 + kk1, twor - k - 1, cy); /* 2n-(5k+1) = 2r-k-1 */
} else { /* triggered only by very unbalanced cases like
(k+k+(k-2))x(k+k+1) , should be handled by toom32 */
ASSERT_NOCARRY (mpn_add_n (vinf, vinf, v2 + k, twor));
}
/* (7) v1 <- v1 - vinf, (1 0 1 0 0) - (1 0 0 0 0) = (0 0 1 0 0)
result is >= 0 */
/* Side effect: we also subtracted (high half) vm1 -= v2 */
cy = mpn_sub_n (v1, v1, vinf, twor); /* vinf is at most twor long. */
vinf0 = vinf[0]; /* Save again the right value for vinf0 */
vinf[0] = saved;
MPN_DECR_U (v1 + twor, kk1 - twor, cy); /* Treat the last bytes. */
/* (8) vm1 <- vm1-v2 (0 1 0 1 0) - (0 1 0 0 0) = (0 0 0 1 0)
Operate only on the low half.
*/
cy = mpn_sub_n (c1, c1, v2, k);
MPN_DECR_U (v1, kk1, cy);
/********************* Beginning the final phase **********************/
/* Most of the recomposition was done */
/* add t2 in {c+3k, ...}, but only the low half */
cy = mpn_add_n (c3, c3, v2, k);
vinf[0] += cy;
ASSERT(vinf[0] >= cy); /* No carry */
MPN_INCR_U (vinf, twor, vinf0); /* Add vinf0, propagate carry. */
#undef v0
}
static
unsigned long int
lc (mp_ptr rp, gmp_randstate_t rstate)
{
mp_ptr tp, seedp, ap;
mp_size_t ta;
mp_size_t tn, seedn, an;
unsigned long int m2exp;
mp_limb_t c;
TMP_DECL (mark);
m2exp = rstate->_mp_algdata._mp_lc->_mp_m2exp;
/* The code below assumes the mod part is a power of two. Make sure
that is the case. */
ASSERT_ALWAYS (m2exp != 0);
c = (mp_limb_t) rstate->_mp_algdata._mp_lc->_mp_c;
seedp = PTR (rstate->_mp_seed);
seedn = SIZ (rstate->_mp_seed);
if (seedn == 0)
{
/* Seed is 0. Result is C % M. Assume table is sensibly stored,
with C smaller than M*/
*rp = c;
*seedp = c;
SIZ (rstate->_mp_seed) = 1;
return m2exp;
}
ap = PTR (rstate->_mp_algdata._mp_lc->_mp_a);
an = SIZ (rstate->_mp_algdata._mp_lc->_mp_a);
/* Allocate temporary storage. Let there be room for calculation of
(A * seed + C) % M, or M if bigger than that. */
TMP_MARK (mark);
ta = an + seedn + 1;
tp = (mp_ptr) TMP_ALLOC (ta * BYTES_PER_MP_LIMB);
/* t = a * seed */
if (seedn >= an)
mpn_mul (tp, seedp, seedn, ap, an);
else
mpn_mul (tp, ap, an, seedp, seedn);
tn = an + seedn;
/* t = t + c */
tp[tn] = 0; /* sentinel, stops MPN_INCR_U */
MPN_INCR_U (tp, tn, c);
ASSERT_ALWAYS (m2exp / GMP_NUMB_BITS < ta);
/* t = t % m */
tp[m2exp / GMP_NUMB_BITS] &= ((mp_limb_t) 1 << m2exp % GMP_NUMB_BITS) - 1;
tn = (m2exp + GMP_NUMB_BITS - 1) / GMP_NUMB_BITS;
/* Save result as next seed. */
MPN_COPY (PTR (rstate->_mp_seed), tp, tn);
SIZ (rstate->_mp_seed) = tn;
{
/* Discard the lower m2exp/2 bits of result. */
unsigned long int bits = m2exp / 2;
mp_size_t xn = bits / GMP_NUMB_BITS;
tn -= xn;
if (tn > 0)
{
unsigned int cnt = bits % GMP_NUMB_BITS;
if (cnt != 0)
{
mpn_rshift (tp, tp + xn, tn, cnt);
MPN_COPY_INCR (rp, tp, xn + 1);
}
else /* Even limb boundary. */
MPN_COPY_INCR (rp, tp + xn, tn);
}
}
TMP_FREE (mark);
/* Return number of valid bits in the result. */
return (m2exp + 1) / 2;
}
void
mpz_powm_ui (mpz_ptr r, mpz_srcptr b, unsigned long int el, mpz_srcptr m)
{
mp_ptr xp, tp, qp, mp, bp;
mp_size_t xn, tn, mn, bn;
int m_zero_cnt;
int c;
mp_limb_t e;
TMP_DECL;
mp = PTR(m);
mn = ABSIZ(m);
if (mn == 0)
DIVIDE_BY_ZERO;
if (el == 0)
{
/* Exponent is zero, result is 1 mod MOD, i.e., 1 or 0
depending on if MOD equals 1. */
SIZ(r) = (mn == 1 && mp[0] == 1) ? 0 : 1;
PTR(r)[0] = 1;
return;
}
TMP_MARK;
/* Normalize m (i.e. make its most significant bit set) as required by
division functions below. */
count_leading_zeros (m_zero_cnt, mp[mn - 1]);
m_zero_cnt -= GMP_NAIL_BITS;
if (m_zero_cnt != 0)
{
mp_ptr new_mp = TMP_ALLOC_LIMBS (mn);
mpn_lshift (new_mp, mp, mn, m_zero_cnt);
mp = new_mp;
}
bn = ABSIZ(b);
bp = PTR(b);
if (bn > mn)
{
/* Reduce possibly huge base. Use a function call to reduce, since we
don't want the quotient allocation to live until function return. */
mp_ptr new_bp = TMP_ALLOC_LIMBS (mn);
reduce (new_bp, bp, bn, mp, mn);
bp = new_bp;
bn = mn;
/* Canonicalize the base, since we are potentially going to multiply with
it quite a few times. */
MPN_NORMALIZE (bp, bn);
}
if (bn == 0)
{
SIZ(r) = 0;
TMP_FREE;
return;
}
tp = TMP_ALLOC_LIMBS (2 * mn + 1);
xp = TMP_ALLOC_LIMBS (mn);
qp = TMP_ALLOC_LIMBS (mn + 1);
MPN_COPY (xp, bp, bn);
xn = bn;
e = el;
count_leading_zeros (c, e);
e = (e << c) << 1; /* shift the exp bits to the left, lose msb */
c = BITS_PER_MP_LIMB - 1 - c;
/* Main loop. */
/* If m is already normalized (high bit of high limb set), and b is the
same size, but a bigger value, and e==1, then there's no modular
reductions done and we can end up with a result out of range at the
end. */
if (c == 0)
{
if (xn == mn && mpn_cmp (xp, mp, mn) >= 0)
mpn_sub_n (xp, xp, mp, mn);
goto finishup;
}
while (c != 0)
{
mpn_sqr_n (tp, xp, xn);
tn = 2 * xn; tn -= tp[tn - 1] == 0;
if (tn < mn)
{
MPN_COPY (xp, tp, tn);
xn = tn;
}
else
{
mpn_tdiv_qr (qp, xp, 0L, tp, tn, mp, mn);
xn = mn;
}
if ((mp_limb_signed_t) e < 0)
{
mpn_mul (tp, xp, xn, bp, bn);
tn = xn + bn; tn -= tp[tn - 1] == 0;
if (tn < mn)
{
MPN_COPY (xp, tp, tn);
xn = tn;
}
else
{
mpn_tdiv_qr (qp, xp, 0L, tp, tn, mp, mn);
xn = mn;
}
}
e <<= 1;
c--;
}
finishup:
/* We shifted m left m_zero_cnt steps. Adjust the result by reducing
it with the original MOD. */
if (m_zero_cnt != 0)
{
mp_limb_t cy;
cy = mpn_lshift (tp, xp, xn, m_zero_cnt);
tp[xn] = cy; xn += cy != 0;
if (xn < mn)
{
MPN_COPY (xp, tp, xn);
}
else
{
mpn_tdiv_qr (qp, xp, 0L, tp, xn, mp, mn);
xn = mn;
}
mpn_rshift (xp, xp, xn, m_zero_cnt);
}
MPN_NORMALIZE (xp, xn);
if ((el & 1) != 0 && SIZ(b) < 0 && xn != 0)
{
mp = PTR(m); /* want original, unnormalized m */
mpn_sub (xp, mp, mn, xp, xn);
xn = mn;
MPN_NORMALIZE (xp, xn);
}
MPZ_REALLOC (r, xn);
SIZ (r) = xn;
MPN_COPY (PTR(r), xp, xn);
TMP_FREE;
}
void
mpn_toom_interpolate_8pts (mp_ptr pp, mp_size_t n,
mp_ptr r3, mp_ptr r7,
mp_size_t spt, mp_ptr ws)
{
mp_limb_signed_t cy;
mp_ptr r5, r1;
r5 = (pp + 3 * n); /* 3n+1 */
r1 = (pp + 7 * n); /* spt */
/******************************* interpolation *****************************/
DO_mpn_subrsh(r3+n, 2 * n + 1, pp, 2 * n, 4, ws);
cy = DO_mpn_sublsh_n (r3, r1, spt, 12, ws);
MPN_DECR_U (r3 + spt, 3 * n + 1 - spt, cy);
DO_mpn_subrsh(r5+n, 2 * n + 1, pp, 2 * n, 2, ws);
cy = DO_mpn_sublsh_n (r5, r1, spt, 6, ws);
MPN_DECR_U (r5 + spt, 3 * n + 1 - spt, cy);
r7[3*n] -= mpn_sub_n (r7+n, r7+n, pp, 2 * n);
cy = mpn_sub_n (r7, r7, r1, spt);
MPN_DECR_U (r7 + spt, 3 * n + 1 - spt, cy);
ASSERT_NOCARRY(mpn_sub_n (r3, r3, r5, 3 * n + 1));
ASSERT_NOCARRY(mpn_rshift(r3, r3, 3 * n + 1, 2));
ASSERT_NOCARRY(mpn_sub_n (r5, r5, r7, 3 * n + 1));
ASSERT_NOCARRY(mpn_sub_n (r3, r3, r5, 3 * n + 1));
mpn_divexact_by45 (r3, r3, 3 * n + 1);
ASSERT_NOCARRY(mpn_divexact_by3 (r5, r5, 3 * n + 1));
ASSERT_NOCARRY(DO_mpn_sublsh2_n (r5, r3, 3 * n + 1, ws));
/* last interpolation steps... */
/* ... are mixed with recomposition */
/***************************** recomposition *******************************/
/*
pp[] prior to operations:
|_H r1|_L r1|____||_H r5|_M_r5|_L r5|_____|_H r8|_L r8|pp
summation scheme for remaining operations:
|____8|n___7|n___6|n___5|n___4|n___3|n___2|n____|n____|pp
|_H r1|_L r1|____||_H*r5|_M r5|_L r5|_____|_H_r8|_L r8|pp
||_H r3|_M r3|_L*r3|
||_H_r7|_M_r7|_L_r7|
||-H r3|-M r3|-L*r3|
||-H*r5|-M_r5|-L_r5|
*/
cy = mpn_add_n (pp + n, pp + n, r7, n); /* Hr8+Lr7-Lr5 */
cy-= mpn_sub_n (pp + n, pp + n, r5, n);
if (0 > cy)
MPN_DECR_U (r7 + n, 2*n + 1, 1);
else
MPN_INCR_U (r7 + n, 2*n + 1, cy);
cy = mpn_sub_n (pp + 2*n, r7 + n, r5 + n, n); /* Mr7-Mr5 */
MPN_DECR_U (r7 + 2*n, n + 1, cy);
cy = mpn_add_n (pp + 3*n, r5, r7+ 2*n, n+1); /* Hr7+Lr5 */
r5[3*n]+= mpn_add_n (r5 + 2*n, r5 + 2*n, r3, n); /* Hr5+Lr3 */
cy-= mpn_sub_n (pp + 3*n, pp + 3*n, r5 + 2*n, n+1); /* Hr7-Hr5+Lr5-Lr3 */
if (UNLIKELY(0 > cy))
MPN_DECR_U (r5 + n + 1, 2*n, 1);
else
MPN_INCR_U (r5 + n + 1, 2*n, cy);
ASSERT_NOCARRY(mpn_sub_n(pp + 4*n, r5 + n, r3 + n, 2*n +1)); /* Mr5-Mr3,Hr5-Hr3 */
cy = mpn_add_1 (pp + 6*n, r3 + n, n, pp[6*n]);
MPN_INCR_U (r3 + 2*n, n + 1, cy);
cy = mpn_add_n (pp + 7*n, pp + 7*n, r3 + 2*n, n);
if (LIKELY(spt != n))
MPN_INCR_U (pp + 8*n, spt - n, cy + r3[3*n]);
else
ASSERT (r3[3*n] | cy == 0);
}
void
_gst_mpz_gcd (gst_mpz *g, const gst_mpz *u, const gst_mpz *v)
{
int g_zero_bits, u_zero_bits, v_zero_bits;
mp_size_t g_zero_limbs, u_zero_limbs, v_zero_limbs;
mp_ptr tp;
mp_ptr up = u->d;
mp_size_t usize = ABS (u->size);
mp_ptr vp = v->d;
mp_size_t vsize = ABS (v->size);
mp_size_t gsize;
/* GCD(0, V) == GCD (U, 1) == V. */
if (usize == 0 || (vsize == 1 && vp[0] == 1))
{
gst_mpz_copy_abs (g, v);
return;
}
/* GCD(U, 0) == GCD (1, V) == U. */
if (vsize == 0 || (usize == 1 && up[0] == 1))
{
gst_mpz_copy_abs (g, u);
return;
}
if (usize == 1)
{
gst_mpz_realloc (g, 1);
g->size = 1;
g->d[0] = mpn_gcd_1 (vp, vsize, up[0]);
return;
}
if (vsize == 1)
{
gst_mpz_realloc (g, 1);
g->size = 1;
g->d[0] = mpn_gcd_1 (up, usize, vp[0]);
return;
}
/* Eliminate low zero bits from U and V and move to temporary storage. */
u_zero_bits = mpn_scan1 (up, 0);
u_zero_limbs = u_zero_bits / BITS_PER_MP_LIMB;
u_zero_bits &= BITS_PER_MP_LIMB - 1;
up += u_zero_limbs;
usize -= u_zero_limbs;
/* Operands could be destroyed for big-endian case, but let's be tidy. */
tp = up;
up = (mp_ptr) alloca (usize * SIZEOF_MP_LIMB_T);
if (u_zero_bits != 0)
{
mpn_rshift (up, tp, usize, u_zero_bits);
usize -= up[usize - 1] == 0;
}
else
MPN_COPY (up, tp, usize);
v_zero_bits = mpn_scan1 (vp, 0);
v_zero_limbs = v_zero_bits / BITS_PER_MP_LIMB;
v_zero_bits &= BITS_PER_MP_LIMB - 1;
vp += v_zero_limbs;
vsize -= v_zero_limbs;
/* Operands could be destroyed for big-endian case, but let's be tidy. */
tp = vp;
vp = (mp_ptr) alloca (vsize * SIZEOF_MP_LIMB_T);
if (v_zero_bits != 0)
{
mpn_rshift (vp, tp, vsize, v_zero_bits);
vsize -= vp[vsize - 1] == 0;
}
else
MPN_COPY (vp, tp, vsize);
if (u_zero_limbs > v_zero_limbs)
{
g_zero_limbs = v_zero_limbs;
g_zero_bits = v_zero_bits;
}
else if (u_zero_limbs < v_zero_limbs)
{
g_zero_limbs = u_zero_limbs;
g_zero_bits = u_zero_bits;
}
else /* Equal. */
{
g_zero_limbs = u_zero_limbs;
g_zero_bits = MIN (u_zero_bits, v_zero_bits);
}
/* Call mpn_gcd. The 2nd argument must not have more bits than the 1st. */
vsize = (usize < vsize || (usize == vsize && up[usize-1] < vp[vsize-1]))
? mpn_gcd (vp, vp, vsize, up, usize)
: mpn_gcd (vp, up, usize, vp, vsize);
/* Here G <-- V << (g_zero_limbs*BITS_PER_MP_LIMB + g_zero_bits). */
gsize = vsize + g_zero_limbs;
if (g_zero_bits != 0)
{
mp_limb_t cy_limb;
gsize += (vp[vsize - 1] >> (BITS_PER_MP_LIMB - g_zero_bits)) != 0;
if (g->alloc < gsize)
gst_mpz_realloc (g, gsize);
MPN_ZERO (g->d, g_zero_limbs);
tp = g->d + g_zero_limbs;
cy_limb = mpn_lshift (tp, vp, vsize, g_zero_bits);
if (cy_limb != 0)
tp[vsize] = cy_limb;
}
int
mpfr_sub1sp (mpfr_ptr a, mpfr_srcptr b, mpfr_srcptr c, mpfr_rnd_t rnd_mode)
{
mpfr_exp_t bx,cx;
mpfr_uexp_t d;
mpfr_prec_t p, sh, cnt;
mp_size_t n;
mp_limb_t *ap, *bp, *cp;
mp_limb_t limb;
int inexact;
mp_limb_t bcp,bcp1; /* Cp and C'p+1 */
mp_limb_t bbcp = (mp_limb_t) -1, bbcp1 = (mp_limb_t) -1; /* Cp+1 and C'p+2,
gcc claims that they might be used uninitialized. We fill them with invalid
values, which should produce a failure if so. See README.dev file. */
MPFR_TMP_DECL(marker);
MPFR_TMP_MARK(marker);
MPFR_ASSERTD(MPFR_PREC(a) == MPFR_PREC(b) && MPFR_PREC(b) == MPFR_PREC(c));
MPFR_ASSERTD(MPFR_IS_PURE_FP(b));
MPFR_ASSERTD(MPFR_IS_PURE_FP(c));
/* Read prec and num of limbs */
p = MPFR_PREC (b);
n = MPFR_PREC2LIMBS (p);
/* Fast cmp of |b| and |c|*/
bx = MPFR_GET_EXP (b);
cx = MPFR_GET_EXP (c);
if (MPFR_UNLIKELY(bx == cx))
{
mp_size_t k = n - 1;
/* Check mantissa since exponent are equals */
bp = MPFR_MANT(b);
cp = MPFR_MANT(c);
while (k>=0 && MPFR_UNLIKELY(bp[k] == cp[k]))
k--;
if (MPFR_UNLIKELY(k < 0))
/* b == c ! */
{
/* Return exact number 0 */
if (rnd_mode == MPFR_RNDD)
MPFR_SET_NEG(a);
else
MPFR_SET_POS(a);
MPFR_SET_ZERO(a);
MPFR_RET(0);
}
else if (bp[k] > cp[k])
goto BGreater;
else
{
MPFR_ASSERTD(bp[k]<cp[k]);
goto CGreater;
}
}
else if (MPFR_UNLIKELY(bx < cx))
{
/* Swap b and c and set sign */
mpfr_srcptr t;
mpfr_exp_t tx;
CGreater:
MPFR_SET_OPPOSITE_SIGN(a,b);
t = b; b = c; c = t;
tx = bx; bx = cx; cx = tx;
}
else
{
/* b > c */
BGreater:
MPFR_SET_SAME_SIGN(a,b);
}
/* Now b > c */
MPFR_ASSERTD(bx >= cx);
d = (mpfr_uexp_t) bx - cx;
DEBUG (printf ("New with diff=%lu\n", (unsigned long) d));
if (MPFR_UNLIKELY(d <= 1))
{
if (MPFR_LIKELY(d < 1))
{
/* <-- b -->
<-- c --> : exact sub */
ap = MPFR_MANT(a);
mpn_sub_n (ap, MPFR_MANT(b), MPFR_MANT(c), n);
/* Normalize */
ExactNormalize:
limb = ap[n-1];
if (MPFR_LIKELY(limb))
{
/* First limb is not zero. */
count_leading_zeros(cnt, limb);
/* cnt could be == 0 <= SubD1Lose */
if (MPFR_LIKELY(cnt))
{
mpn_lshift(ap, ap, n, cnt); /* Normalize number */
bx -= cnt; /* Update final expo */
}
/* Last limb should be ok */
MPFR_ASSERTD(!(ap[0] & MPFR_LIMB_MASK((unsigned int) (-p)
% GMP_NUMB_BITS)));
}
else
{
/* First limb is zero */
mp_size_t k = n-1, len;
/* Find the first limb not equal to zero.
FIXME:It is assume it exists (since |b| > |c| and same prec)*/
do
{
MPFR_ASSERTD( k > 0 );
limb = ap[--k];
}
while (limb == 0);
MPFR_ASSERTD(limb != 0);
count_leading_zeros(cnt, limb);
k++;
len = n - k; /* Number of last limb */
MPFR_ASSERTD(k >= 0);
if (MPFR_LIKELY(cnt))
mpn_lshift(ap+len, ap, k, cnt); /* Normalize the High Limb*/
else
{
/* Must use DECR since src and dest may overlap & dest>=src*/
MPN_COPY_DECR(ap+len, ap, k);
}
MPN_ZERO(ap, len); /* Zeroing the last limbs */
bx -= cnt + len*GMP_NUMB_BITS; /* Update Expo */
/* Last limb should be ok */
MPFR_ASSERTD(!(ap[len]&MPFR_LIMB_MASK((unsigned int) (-p)
% GMP_NUMB_BITS)));
}
/* Check expo underflow */
if (MPFR_UNLIKELY(bx < __gmpfr_emin))
{
MPFR_TMP_FREE(marker);
/* inexact=0 */
DEBUG( printf("(D==0 Underflow)\n") );
if (rnd_mode == MPFR_RNDN &&
(bx < __gmpfr_emin - 1 ||
(/*inexact >= 0 &&*/ mpfr_powerof2_raw (a))))
rnd_mode = MPFR_RNDZ;
return mpfr_underflow (a, rnd_mode, MPFR_SIGN(a));
}
MPFR_SET_EXP (a, bx);
/* No rounding is necessary since the result is exact */
MPFR_ASSERTD(ap[n-1] > ~ap[n-1]);
MPFR_TMP_FREE(marker);
return 0;
}
else /* if (d == 1) */
{
/* | <-- b -->
| <-- c --> */
mp_limb_t c0, mask;
mp_size_t k;
MPFR_UNSIGNED_MINUS_MODULO(sh, p);
/* If we lose at least one bit, compute 2*b-c (Exact)
* else compute b-c/2 */
bp = MPFR_MANT(b);
cp = MPFR_MANT(c);
k = n-1;
limb = bp[k] - cp[k]/2;
if (limb > MPFR_LIMB_HIGHBIT)
{
/* We can't lose precision: compute b-c/2 */
/* Shift c in the allocated temporary block */
SubD1NoLose:
c0 = cp[0] & (MPFR_LIMB_ONE<<sh);
cp = MPFR_TMP_LIMBS_ALLOC (n);
mpn_rshift(cp, MPFR_MANT(c), n, 1);
if (MPFR_LIKELY(c0 == 0))
{
/* Result is exact: no need of rounding! */
ap = MPFR_MANT(a);
mpn_sub_n (ap, bp, cp, n);
MPFR_SET_EXP(a, bx); /* No expo overflow! */
/* No truncate or normalize is needed */
MPFR_ASSERTD(ap[n-1] > ~ap[n-1]);
/* No rounding is necessary since the result is exact */
MPFR_TMP_FREE(marker);
return 0;
}
ap = MPFR_MANT(a);
mask = ~MPFR_LIMB_MASK(sh);
cp[0] &= mask; /* Delete last bit of c */
mpn_sub_n (ap, bp, cp, n);
MPFR_SET_EXP(a, bx); /* No expo overflow! */
MPFR_ASSERTD( !(ap[0] & ~mask) ); /* Check last bits */
/* No normalize is needed */
MPFR_ASSERTD(ap[n-1] > ~ap[n-1]);
/* Rounding is necessary since c0 = 1*/
/* Cp =-1 and C'p+1=0 */
bcp = 1; bcp1 = 0;
if (MPFR_LIKELY(rnd_mode == MPFR_RNDN))
{
/* Even Rule apply: Check Ap-1 */
if (MPFR_LIKELY( (ap[0] & (MPFR_LIMB_ONE<<sh)) == 0) )
goto truncate;
else
goto sub_one_ulp;
}
MPFR_UPDATE_RND_MODE(rnd_mode, MPFR_IS_NEG(a));
if (rnd_mode == MPFR_RNDZ)
goto sub_one_ulp;
else
goto truncate;
}
else if (MPFR_LIKELY(limb < MPFR_LIMB_HIGHBIT))
{
/* We lose at least one bit of prec */
/* Calcul of 2*b-c (Exact) */
/* Shift b in the allocated temporary block */
SubD1Lose:
bp = MPFR_TMP_LIMBS_ALLOC (n);
mpn_lshift (bp, MPFR_MANT(b), n, 1);
ap = MPFR_MANT(a);
mpn_sub_n (ap, bp, cp, n);
bx--;
goto ExactNormalize;
}
else
{
/* Case: limb = 100000000000 */
/* Check while b[k] == c'[k] (C' is C shifted by 1) */
/* If b[k]<c'[k] => We lose at least one bit*/
/* If b[k]>c'[k] => We don't lose any bit */
/* If k==-1 => We don't lose any bit
AND the result is 100000000000 0000000000 00000000000 */
mp_limb_t carry;
do {
carry = cp[k]&MPFR_LIMB_ONE;
k--;
} while (k>=0 &&
bp[k]==(carry=cp[k]/2+(carry<<(GMP_NUMB_BITS-1))));
if (MPFR_UNLIKELY(k<0))
{
/*If carry then (sh==0 and Virtual c'[-1] > Virtual b[-1]) */
if (MPFR_UNLIKELY(carry)) /* carry = cp[0]&MPFR_LIMB_ONE */
{
/* FIXME: Can be faster? */
MPFR_ASSERTD(sh == 0);
goto SubD1Lose;
}
/* Result is a power of 2 */
ap = MPFR_MANT (a);
MPN_ZERO (ap, n);
ap[n-1] = MPFR_LIMB_HIGHBIT;
MPFR_SET_EXP (a, bx); /* No expo overflow! */
/* No Normalize is needed*/
/* No Rounding is needed */
MPFR_TMP_FREE (marker);
return 0;
}
/* carry = cp[k]/2+(cp[k-1]&1)<<(GMP_NUMB_BITS-1) = c'[k]*/
else if (bp[k] > carry)
goto SubD1NoLose;
else
{
MPFR_ASSERTD(bp[k]<carry);
goto SubD1Lose;
}
}
}
}
else if (MPFR_UNLIKELY(d >= p))
{
ap = MPFR_MANT(a);
MPFR_UNSIGNED_MINUS_MODULO(sh, p);
/* We can't set A before since we use cp for rounding... */
/* Perform rounding: check if a=b or a=b-ulp(b) */
if (MPFR_UNLIKELY(d == p))
{
/* cp == -1 and c'p+1 = ? */
bcp = 1;
/* We need Cp+1 later for a very improbable case. */
bbcp = (MPFR_MANT(c)[n-1] & (MPFR_LIMB_ONE<<(GMP_NUMB_BITS-2)));
/* We need also C'p+1 for an even more unprobable case... */
if (MPFR_LIKELY( bbcp ))
bcp1 = 1;
else
{
cp = MPFR_MANT(c);
if (MPFR_UNLIKELY(cp[n-1] == MPFR_LIMB_HIGHBIT))
{
mp_size_t k = n-1;
do {
k--;
} while (k>=0 && cp[k]==0);
bcp1 = (k>=0);
}
else
bcp1 = 1;
}
DEBUG( printf("(D=P) Cp=-1 Cp+1=%d C'p+1=%d \n", bbcp!=0, bcp1!=0) );
bp = MPFR_MANT (b);
/* Even if src and dest overlap, it is ok using MPN_COPY */
if (MPFR_LIKELY(rnd_mode == MPFR_RNDN))
{
if (MPFR_UNLIKELY( bcp && bcp1==0 ))
/* Cp=-1 and C'p+1=0: Even rule Apply! */
/* Check Ap-1 = Bp-1 */
if ((bp[0] & (MPFR_LIMB_ONE<<sh)) == 0)
{
MPN_COPY(ap, bp, n);
goto truncate;
}
MPN_COPY(ap, bp, n);
goto sub_one_ulp;
}
MPFR_UPDATE_RND_MODE(rnd_mode, MPFR_IS_NEG(a));
if (rnd_mode == MPFR_RNDZ)
{
MPN_COPY(ap, bp, n);
goto sub_one_ulp;
}
else
{
MPN_COPY(ap, bp, n);
goto truncate;
}
}
else
{
/* Cp=0, Cp+1=-1 if d==p+1, C'p+1=-1 */
bcp = 0; bbcp = (d==p+1); bcp1 = 1;
DEBUG( printf("(D>P) Cp=%d Cp+1=%d C'p+1=%d\n", bcp!=0,bbcp!=0,bcp1!=0) );
/* Need to compute C'p+2 if d==p+1 and if rnd_mode=NEAREST
(Because of a very improbable case) */
if (MPFR_UNLIKELY(d==p+1 && rnd_mode==MPFR_RNDN))
{
cp = MPFR_MANT(c);
if (MPFR_UNLIKELY(cp[n-1] == MPFR_LIMB_HIGHBIT))
{
mp_size_t k = n-1;
do {
k--;
} while (k>=0 && cp[k]==0);
bbcp1 = (k>=0);
}
else
bbcp1 = 1;
DEBUG( printf("(D>P) C'p+2=%d\n", bbcp1!=0) );
}
/* Copy mantissa B in A */
MPN_COPY(ap, MPFR_MANT(b), n);
/* Round */
if (MPFR_LIKELY(rnd_mode == MPFR_RNDN))
goto truncate;
MPFR_UPDATE_RND_MODE(rnd_mode, MPFR_IS_NEG(a));
if (rnd_mode == MPFR_RNDZ)
goto sub_one_ulp;
else /* rnd_mode = AWAY */
goto truncate;
}
}
else
{
mpfr_uexp_t dm;
mp_size_t m;
mp_limb_t mask;
/* General case: 2 <= d < p */
MPFR_UNSIGNED_MINUS_MODULO(sh, p);
cp = MPFR_TMP_LIMBS_ALLOC (n);
/* Shift c in temporary allocated place */
dm = d % GMP_NUMB_BITS;
m = d / GMP_NUMB_BITS;
if (MPFR_UNLIKELY(dm == 0))
{
/* dm = 0 and m > 0: Just copy */
MPFR_ASSERTD(m!=0);
MPN_COPY(cp, MPFR_MANT(c)+m, n-m);
MPN_ZERO(cp+n-m, m);
}
else if (MPFR_LIKELY(m == 0))
{
/* dm >=2 and m == 0: just shift */
MPFR_ASSERTD(dm >= 2);
mpn_rshift(cp, MPFR_MANT(c), n, dm);
}
else
{
/* dm > 0 and m > 0: shift and zero */
mpn_rshift(cp, MPFR_MANT(c)+m, n-m, dm);
MPN_ZERO(cp+n-m, m);
}
DEBUG( mpfr_print_mant_binary("Before", MPFR_MANT(c), p) );
DEBUG( mpfr_print_mant_binary("B= ", MPFR_MANT(b), p) );
DEBUG( mpfr_print_mant_binary("After ", cp, p) );
/* Compute bcp=Cp and bcp1=C'p+1 */
if (MPFR_LIKELY(sh))
{
/* Try to compute them from C' rather than C (FIXME: Faster?) */
bcp = (cp[0] & (MPFR_LIMB_ONE<<(sh-1))) ;
if (MPFR_LIKELY( cp[0] & MPFR_LIMB_MASK(sh-1) ))
bcp1 = 1;
else
{
/* We can't compute C'p+1 from C'. Compute it from C */
/* Start from bit x=p-d+sh in mantissa C
(+sh since we have already looked sh bits in C'!) */
mpfr_prec_t x = p-d+sh-1;
if (MPFR_LIKELY(x>p))
/* We are already looked at all the bits of c, so C'p+1 = 0*/
bcp1 = 0;
else
{
mp_limb_t *tp = MPFR_MANT(c);
mp_size_t kx = n-1 - (x / GMP_NUMB_BITS);
mpfr_prec_t sx = GMP_NUMB_BITS-1-(x%GMP_NUMB_BITS);
DEBUG (printf ("(First) x=%lu Kx=%ld Sx=%lu\n",
(unsigned long) x, (long) kx,
(unsigned long) sx));
/* Looks at the last bits of limb kx (if sx=0 does nothing)*/
if (tp[kx] & MPFR_LIMB_MASK(sx))
bcp1 = 1;
else
{
/*kx += (sx==0);*/
/*If sx==0, tp[kx] hasn't been checked*/
do {
kx--;
} while (kx>=0 && tp[kx]==0);
bcp1 = (kx >= 0);
}
}
}
}
else
{
/* Compute Cp and C'p+1 from C with sh=0 */
mp_limb_t *tp = MPFR_MANT(c);
/* Start from bit x=p-d in mantissa C */
mpfr_prec_t x = p-d;
mp_size_t kx = n-1 - (x / GMP_NUMB_BITS);
mpfr_prec_t sx = GMP_NUMB_BITS-1-(x%GMP_NUMB_BITS);
MPFR_ASSERTD(p >= d);
bcp = (tp[kx] & (MPFR_LIMB_ONE<<sx));
/* Looks at the last bits of limb kx (If sx=0, does nothing)*/
if (tp[kx] & MPFR_LIMB_MASK(sx))
bcp1 = 1;
else
{
/*kx += (sx==0);*/ /*If sx==0, tp[kx] hasn't been checked*/
do {
kx--;
} while (kx>=0 && tp[kx]==0);
bcp1 = (kx>=0);
}
}
DEBUG( printf("sh=%lu Cp=%d C'p+1=%d\n", sh, bcp!=0, bcp1!=0) );
/* Check if we can lose a bit, and if so compute Cp+1 and C'p+2 */
bp = MPFR_MANT(b);
if (MPFR_UNLIKELY((bp[n-1]-cp[n-1]) <= MPFR_LIMB_HIGHBIT))
{
/* We can lose a bit so we precompute Cp+1 and C'p+2 */
/* Test for trivial case: since C'p+1=0, Cp+1=0 and C'p+2 =0 */
if (MPFR_LIKELY(bcp1 == 0))
{
bbcp = 0;
bbcp1 = 0;
}
else /* bcp1 != 0 */
{
/* We can lose a bit:
compute Cp+1 and C'p+2 from mantissa C */
mp_limb_t *tp = MPFR_MANT(c);
/* Start from bit x=(p+1)-d in mantissa C */
mpfr_prec_t x = p+1-d;
mp_size_t kx = n-1 - (x/GMP_NUMB_BITS);
mpfr_prec_t sx = GMP_NUMB_BITS-1-(x%GMP_NUMB_BITS);
MPFR_ASSERTD(p > d);
DEBUG (printf ("(pre) x=%lu Kx=%ld Sx=%lu\n",
(unsigned long) x, (long) kx,
(unsigned long) sx));
bbcp = (tp[kx] & (MPFR_LIMB_ONE<<sx)) ;
/* Looks at the last bits of limb kx (If sx=0, does nothing)*/
/* If Cp+1=0, since C'p+1!=0, C'p+2=1 ! */
if (MPFR_LIKELY(bbcp==0 || (tp[kx]&MPFR_LIMB_MASK(sx))))
bbcp1 = 1;
else
{
/*kx += (sx==0);*/ /*If sx==0, tp[kx] hasn't been checked*/
do {
kx--;
} while (kx>=0 && tp[kx]==0);
bbcp1 = (kx>=0);
DEBUG (printf ("(Pre) Scan done for %ld\n", (long) kx));
}
} /*End of Bcp1 != 0*/
DEBUG( printf("(Pre) Cp+1=%d C'p+2=%d\n", bbcp!=0, bbcp1!=0) );
} /* End of "can lose a bit" */
/* Clean shifted C' */
mask = ~MPFR_LIMB_MASK (sh);
cp[0] &= mask;
/* Subtract the mantissa c from b in a */
ap = MPFR_MANT(a);
mpn_sub_n (ap, bp, cp, n);
DEBUG( mpfr_print_mant_binary("Sub= ", ap, p) );
/* Normalize: we lose at max one bit*/
if (MPFR_UNLIKELY(MPFR_LIMB_MSB(ap[n-1]) == 0))
{
/* High bit is not set and we have to fix it! */
/* Ap >= 010000xxx001 */
mpn_lshift(ap, ap, n, 1);
/* Ap >= 100000xxx010 */
if (MPFR_UNLIKELY(bcp!=0)) /* Check if Cp = -1 */
/* Since Cp == -1, we have to substract one more */
{
mpn_sub_1(ap, ap, n, MPFR_LIMB_ONE<<sh);
MPFR_ASSERTD(MPFR_LIMB_MSB(ap[n-1]) != 0);
}
/* Ap >= 10000xxx001 */
/* Final exponent -1 since we have shifted the mantissa */
bx--;
/* Update bcp and bcp1 */
MPFR_ASSERTN(bbcp != (mp_limb_t) -1);
MPFR_ASSERTN(bbcp1 != (mp_limb_t) -1);
bcp = bbcp;
bcp1 = bbcp1;
/* We dont't have anymore a valid Cp+1!
But since Ap >= 100000xxx001, the final sub can't unnormalize!*/
}
MPFR_ASSERTD( !(ap[0] & ~mask) );
/* Rounding */
if (MPFR_LIKELY(rnd_mode == MPFR_RNDN))
{
if (MPFR_LIKELY(bcp==0))
goto truncate;
else if ((bcp1) || ((ap[0] & (MPFR_LIMB_ONE<<sh)) != 0))
goto sub_one_ulp;
else
goto truncate;
}
/* Update rounding mode */
MPFR_UPDATE_RND_MODE(rnd_mode, MPFR_IS_NEG(a));
if (rnd_mode == MPFR_RNDZ && (MPFR_LIKELY(bcp || bcp1)))
goto sub_one_ulp;
goto truncate;
}
MPFR_RET_NEVER_GO_HERE ();
/* Sub one ulp to the result */
sub_one_ulp:
mpn_sub_1 (ap, ap, n, MPFR_LIMB_ONE << sh);
/* Result should be smaller than exact value: inexact=-1 */
inexact = -1;
/* Check normalisation */
if (MPFR_UNLIKELY(MPFR_LIMB_MSB(ap[n-1]) == 0))
{
/* ap was a power of 2, and we lose a bit */
/* Now it is 0111111111111111111[00000 */
mpn_lshift(ap, ap, n, 1);
bx--;
/* And the lost bit x depends on Cp+1, and Cp */
/* Compute Cp+1 if it isn't already compute (ie d==1) */
/* FIXME: Is this case possible? */
if (MPFR_UNLIKELY(d == 1))
bbcp = 0;
DEBUG( printf("(SubOneUlp)Cp=%d, Cp+1=%d C'p+1=%d\n", bcp!=0,bbcp!=0,bcp1!=0));
/* Compute the last bit (Since we have shifted the mantissa)
we need one more bit!*/
MPFR_ASSERTN(bbcp != (mp_limb_t) -1);
if ( (rnd_mode == MPFR_RNDZ && bcp==0)
|| (rnd_mode==MPFR_RNDN && bbcp==0)
|| (bcp && bcp1==0) ) /*Exact result*/
{
ap[0] |= MPFR_LIMB_ONE<<sh;
if (rnd_mode == MPFR_RNDN)
inexact = 1;
DEBUG( printf("(SubOneUlp) Last bit set\n") );
}
/* Result could be exact if C'p+1 = 0 and rnd == Zero
since we have had one more bit to the result */
/* Fixme: rnd_mode == MPFR_RNDZ needed ? */
if (bcp1==0 && rnd_mode==MPFR_RNDZ)
{
DEBUG( printf("(SubOneUlp) Exact result\n") );
inexact = 0;
}
}
goto end_of_sub;
truncate:
/* Check if the result is an exact power of 2: 100000000000
in which cases, we could have to do sub_one_ulp due to some nasty reasons:
If Result is a Power of 2:
+ If rnd = AWAY,
| If Cp=-1 and C'p+1 = 0, SubOneUlp and the result is EXACT.
If Cp=-1 and C'p+1 =-1, SubOneUlp and the result is above.
Otherwise truncate
+ If rnd = NEAREST,
If Cp= 0 and Cp+1 =-1 and C'p+2=-1, SubOneUlp and the result is above
If cp=-1 and C'p+1 = 0, SubOneUlp and the result is exact.
Otherwise truncate.
X bit should always be set if SubOneUlp*/
if (MPFR_UNLIKELY(ap[n-1] == MPFR_LIMB_HIGHBIT))
{
mp_size_t k = n-1;
do {
k--;
} while (k>=0 && ap[k]==0);
if (MPFR_UNLIKELY(k<0))
{
/* It is a power of 2! */
/* Compute Cp+1 if it isn't already compute (ie d==1) */
/* FIXME: Is this case possible? */
if (d == 1)
bbcp=0;
DEBUG( printf("(Truncate) Cp=%d, Cp+1=%d C'p+1=%d C'p+2=%d\n", \
bcp!=0, bbcp!=0, bcp1!=0, bbcp1!=0) );
MPFR_ASSERTN(bbcp != (mp_limb_t) -1);
MPFR_ASSERTN((rnd_mode != MPFR_RNDN) || (bcp != 0) || (bbcp == 0) || (bbcp1 != (mp_limb_t) -1));
if (((rnd_mode != MPFR_RNDZ) && bcp)
||
((rnd_mode == MPFR_RNDN) && (bcp == 0) && (bbcp) && (bbcp1)))
{
DEBUG( printf("(Truncate) Do sub\n") );
mpn_sub_1 (ap, ap, n, MPFR_LIMB_ONE << sh);
mpn_lshift(ap, ap, n, 1);
ap[0] |= MPFR_LIMB_ONE<<sh;
bx--;
/* FIXME: Explain why it works (or why not)... */
inexact = (bcp1 == 0) ? 0 : (rnd_mode==MPFR_RNDN) ? -1 : 1;
goto end_of_sub;
}
}
}
/* Calcul of Inexact flag.*/
inexact = MPFR_LIKELY(bcp || bcp1) ? 1 : 0;
end_of_sub:
/* Update Expo */
/* FIXME: Is this test really useful?
If d==0 : Exact case. This is never called.
if 1 < d < p : bx=MPFR_EXP(b) or MPFR_EXP(b)-1 > MPFR_EXP(c) > emin
if d == 1 : bx=MPFR_EXP(b). If we could lose any bits, the exact
normalisation is called.
if d >= p : bx=MPFR_EXP(b) >= MPFR_EXP(c) + p > emin
After SubOneUlp, we could have one bit less.
if 1 < d < p : bx >= MPFR_EXP(b)-2 >= MPFR_EXP(c) > emin
if d == 1 : bx >= MPFR_EXP(b)-1 = MPFR_EXP(c) > emin.
if d >= p : bx >= MPFR_EXP(b)-1 > emin since p>=2.
*/
MPFR_ASSERTD( bx >= __gmpfr_emin);
/*
if (MPFR_UNLIKELY(bx < __gmpfr_emin))
{
DEBUG( printf("(Final Underflow)\n") );
if (rnd_mode == MPFR_RNDN &&
(bx < __gmpfr_emin - 1 ||
(inexact >= 0 && mpfr_powerof2_raw (a))))
rnd_mode = MPFR_RNDZ;
MPFR_TMP_FREE(marker);
return mpfr_underflow (a, rnd_mode, MPFR_SIGN(a));
}
*/
MPFR_SET_EXP (a, bx);
MPFR_TMP_FREE(marker);
MPFR_RET (inexact * MPFR_INT_SIGN (a));
}
void
mpn_toom_interpolate_7pts (mp_ptr rp, mp_size_t n, enum toom7_flags flags,
mp_ptr w1, mp_ptr w3, mp_ptr w4, mp_ptr w5,
mp_size_t w6n, mp_ptr tp)
{
mp_size_t m;
mp_limb_t cy;
m = 2*n + 1;
#define w0 rp
#define w2 (rp + 2*n)
#define w6 (rp + 6*n)
ASSERT (w6n > 0);
ASSERT (w6n <= 2*n);
/* Using formulas similar to Marco Bodrato's
W5 = W5 + W4
W1 =(W4 - W1)/2
W4 = W4 - W0
W4 =(W4 - W1)/4 - W6*16
W3 =(W2 - W3)/2
W2 = W2 - W3
W5 = W5 - W2*65 May be negative.
W2 = W2 - W6 - W0
W5 =(W5 + W2*45)/2 Now >= 0 again.
W4 =(W4 - W2)/3
W2 = W2 - W4
W1 = W5 - W1 May be negative.
W5 =(W5 - W3*8)/9
W3 = W3 - W5
W1 =(W1/15 + W5)/2 Now >= 0 again.
W5 = W5 - W1
where W0 = f(0), W1 = f(-2), W2 = f(1), W3 = f(-1),
W4 = f(2), W5 = f(1/2), W6 = f(oo),
Note that most intermediate results are positive; the ones that
may be negative are represented in two's complement. We must
never shift right a value that may be negative, since that would
invalidate the sign bit. On the other hand, divexact by odd
numbers work fine with two's complement.
*/
mpn_add_n (w5, w5, w4, m);
if (flags & toom7_w1_neg)
{
#ifdef HAVE_NATIVE_mpn_rsh1add_n
mpn_rsh1add_n (w1, w1, w4, m);
#else
mpn_add_n (w1, w1, w4, m); ASSERT (!(w1[0] & 1));
mpn_rshift (w1, w1, m, 1);
#endif
}
else
{
#ifdef HAVE_NATIVE_mpn_rsh1sub_n
mpn_rsh1sub_n (w1, w4, w1, m);
#else
mpn_sub_n (w1, w4, w1, m); ASSERT (!(w1[0] & 1));
mpn_rshift (w1, w1, m, 1);
#endif
}
mpn_sub (w4, w4, m, w0, 2*n);
mpn_sub_n (w4, w4, w1, m); ASSERT (!(w4[0] & 3));
mpn_rshift (w4, w4, m, 2); /* w4>=0 */
tp[w6n] = mpn_lshift (tp, w6, w6n, 4);
mpn_sub (w4, w4, m, tp, w6n+1);
if (flags & toom7_w3_neg)
{
#ifdef HAVE_NATIVE_mpn_rsh1add_n
mpn_rsh1add_n (w3, w3, w2, m);
#else
mpn_add_n (w3, w3, w2, m); ASSERT (!(w3[0] & 1));
mpn_rshift (w3, w3, m, 1);
#endif
}
else
{
#ifdef HAVE_NATIVE_mpn_rsh1sub_n
mpn_rsh1sub_n (w3, w2, w3, m);
#else
mpn_sub_n (w3, w2, w3, m); ASSERT (!(w3[0] & 1));
mpn_rshift (w3, w3, m, 1);
#endif
}
mpn_sub_n (w2, w2, w3, m);
mpn_submul_1 (w5, w2, m, 65);
mpn_sub (w2, w2, m, w6, w6n);
mpn_sub (w2, w2, m, w0, 2*n);
mpn_addmul_1 (w5, w2, m, 45); ASSERT (!(w5[0] & 1));
mpn_rshift (w5, w5, m, 1);
mpn_sub_n (w4, w4, w2, m);
mpn_divexact_by3 (w4, w4, m);
mpn_sub_n (w2, w2, w4, m);
mpn_sub_n (w1, w5, w1, m);
mpn_lshift (tp, w3, m, 3);
mpn_sub_n (w5, w5, tp, m);
mpn_divexact_by9 (w5, w5, m);
mpn_sub_n (w3, w3, w5, m);
mpn_divexact_by15 (w1, w1, m);
mpn_add_n (w1, w1, w5, m); ASSERT (!(w1[0] & 1));
mpn_rshift (w1, w1, m, 1); /* w1>=0 now */
mpn_sub_n (w5, w5, w1, m);
/* These bounds are valid for the 4x4 polynomial product of toom44,
* and they are conservative for toom53 and toom62. */
ASSERT (w1[2*n] < 2);
ASSERT (w2[2*n] < 3);
ASSERT (w3[2*n] < 4);
ASSERT (w4[2*n] < 3);
ASSERT (w5[2*n] < 2);
/* Addition chain. Note carries and the 2n'th limbs that need to be
* added in.
*
* Special care is needed for w2[2n] and the corresponding carry,
* since the "simple" way of adding it all together would overwrite
* the limb at wp[2*n] and rp[4*n] (same location) with the sum of
* the high half of w3 and the low half of w4.
*
* 7 6 5 4 3 2 1 0
* | | | | | | | | |
* ||w3 (2n+1)|
* ||w4 (2n+1)|
* ||w5 (2n+1)| ||w1 (2n+1)|
* + | w6 (w6n)| ||w2 (2n+1)| w0 (2n) | (share storage with r)
* -----------------------------------------------
* r | | | | | | | | |
* c7 c6 c5 c4 c3 Carries to propagate
*/
cy = mpn_add_n (rp + n, rp + n, w1, m);
MPN_INCR_U (w2 + n + 1, n , cy);
cy = mpn_add_n (rp + 3*n, rp + 3*n, w3, n);
MPN_INCR_U (w3 + n, n + 1, w2[2*n] + cy);
cy = mpn_add_n (rp + 4*n, w3 + n, w4, n);
MPN_INCR_U (w4 + n, n + 1, w3[2*n] + cy);
cy = mpn_add_n (rp + 5*n, w4 + n, w5, n);
MPN_INCR_U (w5 + n, n + 1, w4[2*n] + cy);
if (w6n > n + 1)
ASSERT_NOCARRY (mpn_add (rp + 6*n, rp + 6*n, w6n, w5 + n, n + 1));
else
{
ASSERT_NOCARRY (mpn_add_n (rp + 6*n, rp + 6*n, w5 + n, w6n));
#if WANT_ASSERT
{
mp_size_t i;
for (i = w6n; i <= n; i++)
ASSERT (w5[n + i] == 0);
}
#endif
}
}
