void
mpn_invert (mp_ptr ip, mp_srcptr dp, mp_size_t n, mp_ptr scratch)
{
ASSERT (n > 0);
ASSERT (dp[n-1] & GMP_NUMB_HIGHBIT);
ASSERT (! MPN_OVERLAP_P (ip, n, dp, n));
ASSERT (! MPN_OVERLAP_P (ip, n, scratch, mpn_invertappr_itch(n)));
ASSERT (! MPN_OVERLAP_P (dp, n, scratch, mpn_invertappr_itch(n)));
if (n == 1)
invert_limb (*ip, *dp);
else if (BELOW_THRESHOLD (n, INV_APPR_THRESHOLD))
{
/* Maximum scratch needed by this branch: 2*n */
mp_size_t i;
mp_ptr xp;
xp = scratch; /* 2 * n limbs */
/* n > 1 here */
i = n;
do
xp[--i] = GMP_NUMB_MAX;
while (i);
mpn_com (xp + n, dp, n);
if (n == 2) {
mpn_divrem_2 (ip, 0, xp, 4, dp);
} else {
gmp_pi1_t inv;
invert_pi1 (inv, dp[n-1], dp[n-2]);
/* FIXME: should we use dcpi1_div_q, for big sizes? */
mpn_sbpi1_div_q (ip, xp, 2 * n, dp, n, inv.inv32);
}
}
else { /* Use approximated inverse; correct the result if needed. */
mp_limb_t e; /* The possible error in the approximate inverse */
ASSERT ( mpn_invert_itch (n) >= mpn_invertappr_itch (n) );
e = mpn_ni_invertappr (ip, dp, n, scratch);
if (UNLIKELY (e)) { /* Assume the error can only be "0" (no error) or "1". */
/* Code to detect and correct the "off by one" approximation. */
mpn_mul_n (scratch, ip, dp, n);
e = mpn_add_n (scratch, scratch, dp, n); /* FIXME: we only need e.*/
if (LIKELY(e)) /* The high part can not give a carry by itself. */
e = mpn_add_nc (scratch + n, scratch + n, dp, n, e); /* FIXME:e */
/* If the value was wrong (no carry), correct it (increment). */
e ^= CNST_LIMB (1);
MPN_INCR_U (ip, n, e);
}
}
}
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
}
/* 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
check (void)
{
mp_limb_t wp[100], xp[100], yp[100];
mp_size_t size = 100;
refmpn_zero (xp, size);
refmpn_zero (yp, size);
refmpn_zero (wp, size);
pre ("mpn_add_n");
mpn_add_n (wp, xp, yp, size);
post ();
#if HAVE_NATIVE_mpn_add_nc
pre ("mpn_add_nc");
mpn_add_nc (wp, xp, yp, size, CNST_LIMB(0));
post ();
#endif
#if HAVE_NATIVE_mpn_addlsh1_n
pre ("mpn_addlsh1_n");
mpn_addlsh1_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_and_n
pre ("mpn_and_n");
mpn_and_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_andn_n
pre ("mpn_andn_n");
mpn_andn_n (wp, xp, yp, size);
post ();
#endif
pre ("mpn_addmul_1");
mpn_addmul_1 (wp, xp, size, yp[0]);
post ();
#if HAVE_NATIVE_mpn_addmul_1c
pre ("mpn_addmul_1c");
mpn_addmul_1c (wp, xp, size, yp[0], CNST_LIMB(0));
post ();
#endif
#if HAVE_NATIVE_mpn_com_n
pre ("mpn_com_n");
mpn_com_n (wp, xp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_copyd
pre ("mpn_copyd");
mpn_copyd (wp, xp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_copyi
pre ("mpn_copyi");
mpn_copyi (wp, xp, size);
post ();
#endif
pre ("mpn_divexact_1");
mpn_divexact_1 (wp, xp, size, CNST_LIMB(123));
post ();
pre ("mpn_divexact_by3c");
mpn_divexact_by3c (wp, xp, size, CNST_LIMB(0));
post ();
pre ("mpn_divrem_1");
mpn_divrem_1 (wp, (mp_size_t) 0, xp, size, CNST_LIMB(123));
post ();
#if HAVE_NATIVE_mpn_divrem_1c
pre ("mpn_divrem_1c");
mpn_divrem_1c (wp, (mp_size_t) 0, xp, size, CNST_LIMB(123), CNST_LIMB(122));
post ();
#endif
pre ("mpn_gcd_1");
xp[0] |= 1;
notdead += (unsigned long) mpn_gcd_1 (xp, size, CNST_LIMB(123));
post ();
#if HAVE_NATIVE_mpn_gcd_finda
pre ("mpn_gcd_finda");
xp[0] |= 1;
xp[1] |= 1;
notdead += mpn_gcd_finda (xp);
post ();
#endif
pre ("mpn_hamdist");
notdead += mpn_hamdist (xp, yp, size);
post ();
#if HAVE_NATIVE_mpn_ior_n
pre ("mpn_ior_n");
mpn_ior_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_iorn_n
pre ("mpn_iorn_n");
mpn_iorn_n (wp, xp, yp, size);
post ();
#endif
pre ("mpn_lshift");
mpn_lshift (wp, xp, size, 1);
post ();
pre ("mpn_mod_1");
notdead += mpn_mod_1 (xp, size, CNST_LIMB(123));
post ();
#if HAVE_NATIVE_mpn_mod_1c
pre ("mpn_mod_1c");
notdead += mpn_mod_1c (xp, size, CNST_LIMB(123), CNST_LIMB(122));
post ();
#endif
#if GMP_NUMB_BITS % 4 == 0
pre ("mpn_mod_34lsub1");
notdead += mpn_mod_34lsub1 (xp, size);
post ();
#endif
pre ("mpn_modexact_1_odd");
notdead += mpn_modexact_1_odd (xp, size, CNST_LIMB(123));
post ();
pre ("mpn_modexact_1c_odd");
notdead += mpn_modexact_1c_odd (xp, size, CNST_LIMB(123), CNST_LIMB(456));
post ();
pre ("mpn_mul_1");
mpn_mul_1 (wp, xp, size, yp[0]);
post ();
#if HAVE_NATIVE_mpn_mul_1c
pre ("mpn_mul_1c");
mpn_mul_1c (wp, xp, size, yp[0], CNST_LIMB(0));
post ();
#endif
#if HAVE_NATIVE_mpn_mul_2
pre ("mpn_mul_2");
mpn_mul_2 (wp, xp, size-1, yp);
post ();
#endif
pre ("mpn_mul_basecase");
mpn_mul_basecase (wp, xp, (mp_size_t) 3, yp, (mp_size_t) 3);
post ();
#if HAVE_NATIVE_mpn_nand_n
pre ("mpn_nand_n");
mpn_nand_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_nior_n
pre ("mpn_nior_n");
mpn_nior_n (wp, xp, yp, size);
post ();
#endif
pre ("mpn_popcount");
notdead += mpn_popcount (xp, size);
post ();
pre ("mpn_preinv_mod_1");
notdead += mpn_preinv_mod_1 (xp, size, GMP_NUMB_MAX,
refmpn_invert_limb (GMP_NUMB_MAX));
post ();
#if USE_PREINV_DIVREM_1 || HAVE_NATIVE_mpn_preinv_divrem_1
pre ("mpn_preinv_divrem_1");
mpn_preinv_divrem_1 (wp, (mp_size_t) 0, xp, size, GMP_NUMB_MAX,
refmpn_invert_limb (GMP_NUMB_MAX), 0);
post ();
#endif
#if HAVE_NATIVE_mpn_rsh1add_n
pre ("mpn_rsh1add_n");
mpn_rsh1add_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_rsh1sub_n
pre ("mpn_rsh1sub_n");
mpn_rsh1sub_n (wp, xp, yp, size);
post ();
#endif
pre ("mpn_rshift");
mpn_rshift (wp, xp, size, 1);
post ();
pre ("mpn_sqr_basecase");
mpn_sqr_basecase (wp, xp, (mp_size_t) 3);
post ();
pre ("mpn_submul_1");
mpn_submul_1 (wp, xp, size, yp[0]);
post ();
#if HAVE_NATIVE_mpn_submul_1c
pre ("mpn_submul_1c");
mpn_submul_1c (wp, xp, size, yp[0], CNST_LIMB(0));
post ();
#endif
pre ("mpn_sub_n");
mpn_sub_n (wp, xp, yp, size);
post ();
#if HAVE_NATIVE_mpn_sub_nc
pre ("mpn_sub_nc");
mpn_sub_nc (wp, xp, yp, size, CNST_LIMB(0));
post ();
#endif
#if HAVE_NATIVE_mpn_sublsh1_n
pre ("mpn_sublsh1_n");
mpn_sublsh1_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_udiv_qrnnd
pre ("mpn_udiv_qrnnd");
mpn_udiv_qrnnd (&wp[0], CNST_LIMB(122), xp[0], CNST_LIMB(123));
post ();
#endif
#if HAVE_NATIVE_mpn_udiv_qrnnd_r
pre ("mpn_udiv_qrnnd_r");
mpn_udiv_qrnnd (CNST_LIMB(122), xp[0], CNST_LIMB(123), &wp[0]);
post ();
#endif
#if HAVE_NATIVE_mpn_umul_ppmm
pre ("mpn_umul_ppmm");
mpn_umul_ppmm (&wp[0], xp[0], yp[0]);
post ();
#endif
#if HAVE_NATIVE_mpn_umul_ppmm_r
pre ("mpn_umul_ppmm_r");
mpn_umul_ppmm_r (&wp[0], xp[0], yp[0]);
post ();
#endif
#if HAVE_NATIVE_mpn_xor_n
pre ("mpn_xor_n");
mpn_xor_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_xnor_n
pre ("mpn_xnor_n");
mpn_xnor_n (wp, xp, yp, size);
post ();
#endif
}
/* mpn_addsub_n.
r1[] = s1[] + s2[]
r2[] = s1[] - s2[]
All operands have n limbs.
In-place operations allowed. */
mp_limb_t
mpn_addsub_n (mp_ptr r1p, mp_ptr r2p, mp_srcptr s1p, mp_srcptr s2p, mp_size_t n)
{
mp_limb_t acyn, acyo; /* carry for add */
mp_limb_t scyn, scyo; /* carry for subtract */
mp_size_t off; /* offset in operands */
mp_size_t this_n; /* size of current chunk */
/* We alternatingly add and subtract in chunks that fit into the (L1)
cache. Since the chunks are several hundred limbs, the function call
overhead is insignificant, but we get much better locality. */
/* We have three variant of the inner loop, the proper loop is chosen
depending on whether r1 or r2 are the same operand as s1 or s2. */
if (r1p != s1p && r1p != s2p)
{
/* r1 is not identical to either input operand. We can therefore write
to r1 directly, without using temporary storage. */
acyo = 0;
scyo = 0;
for (off = 0; off < n; off += PART_SIZE)
{
this_n = MIN (n - off, PART_SIZE);
#if HAVE_NATIVE_mpn_add_nc || !HAVE_NATIVE_mpn_add_n
acyo = mpn_add_nc (r1p + off, s1p + off, s2p + off, this_n, acyo);
#else
acyn = mpn_add_n (r1p + off, s1p + off, s2p + off, this_n);
acyo = acyn + mpn_add_1 (r1p + off, r1p + off, this_n, acyo);
#endif
#if HAVE_NATIVE_mpn_sub_nc || !HAVE_NATIVE_mpn_sub_n
scyo = mpn_sub_nc (r2p + off, s1p + off, s2p + off, this_n, scyo);
#else
scyn = mpn_sub_n (r2p + off, s1p + off, s2p + off, this_n);
scyo = scyn + mpn_sub_1 (r2p + off, r2p + off, this_n, scyo);
#endif
}
}
else if (r2p != s1p && r2p != s2p)
{
/* r2 is not identical to either input operand. We can therefore write
to r2 directly, without using temporary storage. */
acyo = 0;
scyo = 0;
for (off = 0; off < n; off += PART_SIZE)
{
this_n = MIN (n - off, PART_SIZE);
#if HAVE_NATIVE_mpn_sub_nc || !HAVE_NATIVE_mpn_sub_n
scyo = mpn_sub_nc (r2p + off, s1p + off, s2p + off, this_n, scyo);
#else
scyn = mpn_sub_n (r2p + off, s1p + off, s2p + off, this_n);
scyo = scyn + mpn_sub_1 (r2p + off, r2p + off, this_n, scyo);
#endif
#if HAVE_NATIVE_mpn_add_nc || !HAVE_NATIVE_mpn_add_n
acyo = mpn_add_nc (r1p + off, s1p + off, s2p + off, this_n, acyo);
#else
acyn = mpn_add_n (r1p + off, s1p + off, s2p + off, this_n);
acyo = acyn + mpn_add_1 (r1p + off, r1p + off, this_n, acyo);
#endif
}
}
else
{
/* r1 and r2 are identical to s1 and s2 (r1==s1 and r2=s2 or vice versa)
Need temporary storage. */
mp_limb_t tp[PART_SIZE];
acyo = 0;
scyo = 0;
for (off = 0; off < n; off += PART_SIZE)
{
this_n = MIN (n - off, PART_SIZE);
#if HAVE_NATIVE_mpn_add_nc || !HAVE_NATIVE_mpn_add_n
acyo = mpn_add_nc (tp, s1p + off, s2p + off, this_n, acyo);
#else
acyn = mpn_add_n (tp, s1p + off, s2p + off, this_n);
acyo = acyn + mpn_add_1 (tp, tp, this_n, acyo);
#endif
#if HAVE_NATIVE_mpn_sub_nc || !HAVE_NATIVE_mpn_sub_n
scyo = mpn_sub_nc (r2p + off, s1p + off, s2p + off, this_n, scyo);
#else
scyn = mpn_sub_n (r2p + off, s1p + off, s2p + off, this_n);
scyo = scyn + mpn_sub_1 (r2p + off, r2p + off, this_n, scyo);
#endif
MPN_COPY (r1p + off, tp, this_n);
}
}
return 2 * acyo + scyo;
}
