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_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 {
TMP_DECL;
TMP_MARK;
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 */
for (i = n - 1; i >= 0; i--)
xp[i] = GMP_NUMB_MAX;
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);
ASSERT_NOCARRY (mpn_add_n (scratch + n, scratch + n, dp, n));
if (! mpn_add (scratch, scratch, 2*n, dp, n))
MPN_INCR_U (ip, n, 1); /* The value was wrong, correct it. */
}
}
TMP_FREE;
}
}
int
main (int argc, char **argv)
{
mp_ptr ap, bp, rp, refp;
mp_size_t max_n, n;
gmp_randstate_ptr rands;
long test, reps = 1000;
TMP_SDECL;
TMP_SMARK;
tests_start ();
TESTS_REPS (reps, argv, argc);
rands = RANDS;
max_n = 32;
ap = TMP_SALLOC_LIMBS (max_n);
bp = TMP_SALLOC_LIMBS (max_n);
rp = TMP_SALLOC_LIMBS (max_n);
refp = TMP_SALLOC_LIMBS (max_n);
for (test = 0; test < reps; test++)
{
for (n = 1; n <= max_n; n++)
{
mpn_random2 (ap, n);
mpn_random2 (bp, n);
refmpn_and_n (refp, ap, bp, n);
mpn_and_n (rp, ap, bp, n);
check_one (refp, rp, ap, bp, n, "and_n");
refmpn_ior_n (refp, ap, bp, n);
mpn_ior_n (rp, ap, bp, n);
check_one (refp, rp, ap, bp, n, "ior_n");
refmpn_xor_n (refp, ap, bp, n);
mpn_xor_n (rp, ap, bp, n);
check_one (refp, rp, ap, bp, n, "xor_n");
refmpn_andn_n (refp, ap, bp, n);
mpn_andn_n (rp, ap, bp, n);
check_one (refp, rp, ap, bp, n, "andn_n");
refmpn_iorn_n (refp, ap, bp, n);
mpn_iorn_n (rp, ap, bp, n);
check_one (refp, rp, ap, bp, n, "iorn_n");
refmpn_nand_n (refp, ap, bp, n);
mpn_nand_n (rp, ap, bp, n);
check_one (refp, rp, ap, bp, n, "nand_n");
refmpn_nior_n (refp, ap, bp, n);
mpn_nior_n (rp, ap, bp, n);
check_one (refp, rp, ap, bp, n, "nior_n");
refmpn_xnor_n (refp, ap, bp, n);
mpn_xnor_n (rp, ap, bp, n);
check_one (refp, rp, ap, bp, n, "xnor_n");
refmpn_com (refp, ap, n);
mpn_com (rp, ap, n);
check_one (refp, rp, ap, bp, n, "com");
}
}
TMP_SFREE;
tests_end ();
return 0;
}
REGPARM_ATTR (1) static void
cfdiv_r_2exp (mpz_ptr w, mpz_srcptr u, mp_bitcnt_t cnt, int dir)
{
mp_size_t usize, abs_usize, limb_cnt, i;
mp_srcptr up;
mp_ptr wp;
mp_limb_t high;
usize = SIZ(u);
if (usize == 0)
{
SIZ(w) = 0;
return;
}
limb_cnt = cnt / GMP_NUMB_BITS;
cnt %= GMP_NUMB_BITS;
abs_usize = ABS (usize);
/* MPZ_REALLOC(w) below is only when w!=u, so we can fetch PTR(u) here
nice and early */
up = PTR(u);
if ((usize ^ dir) < 0)
{
/* Round towards zero, means just truncate */
if (w == u)
{
/* if already smaller than limb_cnt then do nothing */
if (abs_usize <= limb_cnt)
return;
wp = PTR(w);
}
else
{
i = MIN (abs_usize, limb_cnt+1);
wp = MPZ_REALLOC (w, i);
MPN_COPY (wp, up, i);
/* if smaller than limb_cnt then only the copy is needed */
if (abs_usize <= limb_cnt)
{
SIZ(w) = usize;
return;
}
}
}
else
{
/* Round away from zero, means twos complement if non-zero */
/* if u!=0 and smaller than divisor, then must negate */
if (abs_usize <= limb_cnt)
goto negate;
/* if non-zero low limb, then must negate */
for (i = 0; i < limb_cnt; i++)
if (up[i] != 0)
goto negate;
/* if non-zero partial limb, then must negate */
if ((up[limb_cnt] & LOW_MASK (cnt)) != 0)
goto negate;
/* otherwise low bits of u are zero, so that's the result */
SIZ(w) = 0;
return;
negate:
/* twos complement negation to get 2**cnt-u */
wp = MPZ_REALLOC (w, limb_cnt+1);
up = PTR(u);
/* Ones complement */
i = MIN (abs_usize, limb_cnt+1);
mpn_com (wp, up, i);
for ( ; i <= limb_cnt; i++)
wp[i] = GMP_NUMB_MAX;
/* Twos complement. Since u!=0 in the relevant part, the twos
complement never gives 0 and a carry, so can use MPN_INCR_U. */
MPN_INCR_U (wp, limb_cnt+1, CNST_LIMB(1));
usize = -usize;
}
/* Mask the high limb */
high = wp[limb_cnt];
high &= LOW_MASK (cnt);
wp[limb_cnt] = high;
/* Strip any consequent high zeros */
while (high == 0)
{
limb_cnt--;
if (limb_cnt < 0)
{
SIZ(w) = 0;
return;
}
high = wp[limb_cnt];
}
limb_cnt++;
SIZ(w) = (usize >= 0 ? limb_cnt : -limb_cnt);
}
