static void
set_special (mpfr_ptr x, unsigned int select)
{
MPFR_ASSERTN (select < SPECIAL_MAX);
switch (select)
{
case 0:
MPFR_SET_NAN (x);
break;
case 1:
MPFR_SET_INF (x);
MPFR_SET_POS (x);
break;
case 2:
MPFR_SET_INF (x);
MPFR_SET_NEG (x);
break;
case 3:
MPFR_SET_ZERO (x);
MPFR_SET_POS (x);
break;
case 4:
MPFR_SET_ZERO (x);
MPFR_SET_NEG (x);
break;
case 5:
mpfr_set_str_binary (x, "1");
break;
case 6:
mpfr_set_str_binary (x, "-1");
break;
case 7:
mpfr_set_str_binary (x, "1e-1");
break;
case 8:
mpfr_set_str_binary (x, "1e+1");
break;
case 9:
mpfr_const_pi (x, MPFR_RNDN);
break;
case 10:
mpfr_const_pi (x, MPFR_RNDN);
MPFR_SET_EXP (x, MPFR_GET_EXP (x)-1);
break;
default:
mpfr_urandomb (x, RANDS);
if (randlimb () & 1)
mpfr_neg (x, x, MPFR_RNDN);
break;
}
}
void
mpfr_set_str_binary (mpfr_ptr x, const char *str)
{
int has_sign;
int res;
if (*str == 'N')
{
MPFR_SET_NAN(x);
__gmpfr_flags |= MPFR_FLAGS_NAN;
return;
}
has_sign = *str == '-' || *str == '+';
if (str[has_sign] == 'I')
{
MPFR_SET_INF(x);
if (*str == '-')
MPFR_SET_NEG(x);
else
MPFR_SET_POS(x);
return;
}
res = mpfr_strtofr (x, str, 0, 2, MPFR_RNDZ);
MPFR_ASSERTN (res == 0);
}
int
mpfr_sqrt_ui (mpfr_ptr r, unsigned long u, mpfr_rnd_t rnd_mode)
{
if (u)
{
mpfr_t uu;
mp_limb_t up[1];
unsigned long cnt;
int inex;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_TMP_INIT1 (up, uu, GMP_NUMB_BITS);
MPFR_ASSERTN (u == (mp_limb_t) u);
count_leading_zeros (cnt, (mp_limb_t) u);
*up = (mp_limb_t) u << cnt;
MPFR_SAVE_EXPO_MARK (expo);
MPFR_SET_EXP (uu, GMP_NUMB_BITS - cnt);
inex = mpfr_sqrt(r, uu, rnd_mode);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range(r, inex, rnd_mode);
}
else /* sqrt(0) = 0 */
{
MPFR_SET_ZERO(r);
MPFR_SET_POS(r);
MPFR_RET(0);
}
}
int
mpfr_urandomb (mpfr_ptr rop, gmp_randstate_t rstate)
{
mp_ptr rp;
mp_prec_t nbits;
mp_size_t nlimbs;
mp_size_t k; /* number of high zero limbs */
mp_exp_t exp;
int cnt;
MPFR_CLEAR_FLAGS (rop);
rp = MPFR_MANT (rop);
nbits = MPFR_PREC (rop);
nlimbs = MPFR_LIMB_SIZE (rop);
MPFR_SET_POS (rop);
/* Uniform non-normalized significand */
_gmp_rand (rp, rstate, nlimbs * BITS_PER_MP_LIMB);
/* If nbits isn't a multiple of BITS_PER_MP_LIMB, mask the low bits */
cnt = nlimbs * BITS_PER_MP_LIMB - nbits;
if (MPFR_LIKELY (cnt != 0))
rp[0] &= ~MPFR_LIMB_MASK (cnt);
/* Count the null significant limbs and remaining limbs */
exp = 0;
k = 0;
while (nlimbs != 0 && rp[nlimbs - 1] == 0)
{
k ++;
nlimbs --;
exp -= BITS_PER_MP_LIMB;
}
if (MPFR_LIKELY (nlimbs != 0)) /* otherwise value is zero */
{
count_leading_zeros (cnt, rp[nlimbs - 1]);
/* Normalization */
if (mpfr_set_exp (rop, exp - cnt))
{
/* If the exponent is not in the current exponent range, we
choose to return a NaN as this is probably a user error.
Indeed this can happen only if the exponent range has been
reduced to a very small interval and/or the precision is
huge (very unlikely). */
MPFR_SET_NAN (rop);
__gmpfr_flags |= MPFR_FLAGS_NAN; /* Can't use MPFR_RET_NAN */
return 1;
}
if (cnt != 0)
mpn_lshift (rp + k, rp, nlimbs, cnt);
if (k != 0)
MPN_ZERO (rp, k);
}
else
MPFR_SET_ZERO (rop);
return 0;
}
static void
test_overflow1 (void)
{
mpfr_t x, y, z, r;
int inex;
mpfr_inits2 (8, x, y, z, r, (void *) 0);
MPFR_SET_POS (x);
mpfr_setmax (x, mpfr_get_emax ()); /* x = 2^emax - ulp */
mpfr_set_ui (y, 2, GMP_RNDN); /* y = 2 */
mpfr_neg (z, x, GMP_RNDN); /* z = -x = -(2^emax - ulp) */
mpfr_clear_flags ();
/* The intermediate multiplication x * y overflows, but x * y + z = x
is representable. */
inex = mpfr_fma (r, x, y, z, GMP_RNDN);
if (inex || ! mpfr_equal_p (r, x))
{
printf ("Error in test_overflow1\nexpected ");
mpfr_out_str (stdout, 2, 0, x, GMP_RNDN);
printf (" with inex = 0\n got ");
mpfr_out_str (stdout, 2, 0, r, GMP_RNDN);
printf (" with inex = %d\n", inex);
exit (1);
}
if (mpfr_overflow_p ())
{
printf ("Error in test_overflow1: overflow flag set\n");
exit (1);
}
mpfr_clears (x, y, z, r, (void *) 0);
}
/* set f to the integer z multiplied by 2^e */
int
mpfr_set_z_2exp (mpfr_ptr f, mpz_srcptr z, mpfr_exp_t e, mpfr_rnd_t rnd_mode)
{
mp_size_t fn, zn, dif, en;
int k, sign_z, inex;
mp_limb_t *fp, *zp;
mpfr_exp_t exp;
sign_z = mpz_sgn (z);
if (MPFR_UNLIKELY (sign_z == 0)) /* ignore the exponent for 0 */
{
MPFR_SET_ZERO(f);
MPFR_SET_POS(f);
MPFR_RET(0);
}
MPFR_ASSERTD (sign_z == MPFR_SIGN_POS || sign_z == MPFR_SIGN_NEG);
zn = ABS(SIZ(z)); /* limb size of z */
/* compute en = floor(e/GMP_NUMB_BITS) */
en = (e >= 0) ? e / GMP_NUMB_BITS : (e + 1) / GMP_NUMB_BITS - 1;
MPFR_ASSERTD (zn >= 1);
if (MPFR_UNLIKELY (zn + en > MPFR_EMAX_MAX / GMP_NUMB_BITS + 1))
return mpfr_overflow (f, rnd_mode, sign_z);
/* because zn + en >= MPFR_EMAX_MAX / GMP_NUMB_BITS + 2
implies (zn + en) * GMP_NUMB_BITS >= MPFR_EMAX_MAX + GMP_NUMB_BITS + 1
and exp = zn * GMP_NUMB_BITS + e - k
>= (zn + en) * GMP_NUMB_BITS - k > MPFR_EMAX_MAX */
fp = MPFR_MANT (f);
fn = MPFR_LIMB_SIZE (f);
dif = zn - fn;
zp = PTR(z);
count_leading_zeros (k, zp[zn-1]);
/* now zn + en <= MPFR_EMAX_MAX / GMP_NUMB_BITS + 1
thus (zn + en) * GMP_NUMB_BITS <= MPFR_EMAX_MAX + GMP_NUMB_BITS
and exp = zn * GMP_NUMB_BITS + e - k
<= (zn + en) * GMP_NUMB_BITS - k + GMP_NUMB_BITS - 1
<= MPFR_EMAX_MAX + 2 * GMP_NUMB_BITS - 1 */
exp = (mpfr_prec_t) zn * GMP_NUMB_BITS + e - k;
/* The exponent will be exp or exp + 1 (due to rounding) */
if (MPFR_UNLIKELY (exp > __gmpfr_emax))
return mpfr_overflow (f, rnd_mode, sign_z);
if (MPFR_UNLIKELY (exp + 1 < __gmpfr_emin))
return mpfr_underflow (f, rnd_mode == MPFR_RNDN ? MPFR_RNDZ : rnd_mode,
sign_z);
if (MPFR_LIKELY (dif >= 0))
{
mp_limb_t rb, sb, ulp;
int sh;
/* number has to be truncated */
if (MPFR_LIKELY (k != 0))
{
mpn_lshift (fp, &zp[dif], fn, k);
if (MPFR_LIKELY (dif > 0))
fp[0] |= zp[dif - 1] >> (GMP_NUMB_BITS - k);
}
int
mpfr_urandomb (mpfr_ptr rop, gmp_randstate_t rstate)
{
mpfr_limb_ptr rp;
mpfr_prec_t nbits;
mp_size_t nlimbs;
mp_size_t k; /* number of high zero limbs */
mpfr_exp_t exp;
int cnt;
rp = MPFR_MANT (rop);
nbits = MPFR_PREC (rop);
nlimbs = MPFR_LIMB_SIZE (rop);
MPFR_SET_POS (rop);
cnt = nlimbs * GMP_NUMB_BITS - nbits;
/* Uniform non-normalized significand */
/* generate exactly nbits so that the random generator stays in the same
state, independent of the machine word size GMP_NUMB_BITS */
mpfr_rand_raw (rp, rstate, nbits);
if (MPFR_LIKELY (cnt != 0)) /* this will put the low bits to zero */
mpn_lshift (rp, rp, nlimbs, cnt);
/* Count the null significant limbs and remaining limbs */
exp = 0;
k = 0;
while (nlimbs != 0 && rp[nlimbs - 1] == 0)
{
k ++;
nlimbs --;
exp -= GMP_NUMB_BITS;
}
if (MPFR_LIKELY (nlimbs != 0)) /* otherwise value is zero */
{
count_leading_zeros (cnt, rp[nlimbs - 1]);
/* Normalization */
if (mpfr_set_exp (rop, exp - cnt))
{
/* If the exponent is not in the current exponent range, we
choose to return a NaN as this is probably a user error.
Indeed this can happen only if the exponent range has been
reduced to a very small interval and/or the precision is
huge (very unlikely). */
MPFR_SET_NAN (rop);
__gmpfr_flags |= MPFR_FLAGS_NAN; /* Can't use MPFR_RET_NAN */
return 1;
}
if (cnt != 0)
mpn_lshift (rp + k, rp, nlimbs, cnt);
if (k != 0)
MPN_ZERO (rp, k);
}
else
MPFR_SET_ZERO (rop);
return 0;
}
static void
check_nan (void)
{
mpfr_t d, q;
mpfr_init2 (d, 100L);
mpfr_init2 (q, 100L);
/* 1/+inf == 0 */
MPFR_CLEAR_FLAGS (d);
MPFR_SET_INF (d);
MPFR_SET_POS (d);
MPFR_ASSERTN (mpfr_ui_div (q, 1L, d, GMP_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_number_p (q));
MPFR_ASSERTN (mpfr_sgn (q) == 0);
/* 1/-inf == -0 */
MPFR_CLEAR_FLAGS (d);
MPFR_SET_INF (d);
MPFR_SET_NEG (d);
MPFR_ASSERTN (mpfr_ui_div (q, 1L, d, GMP_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_number_p (q));
MPFR_ASSERTN (mpfr_sgn (q) == 0);
/* 1/nan == nan */
MPFR_SET_NAN (d);
MPFR_ASSERTN (mpfr_ui_div (q, 1L, d, GMP_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_nan_p (q));
/* 0/0 == nan */
mpfr_set_ui (d, 0L, GMP_RNDN);
MPFR_ASSERTN (mpfr_ui_div (q, 0L, d, GMP_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_nan_p (q));
/* 1/+0 = +inf */
mpfr_set_ui (d, 0L, GMP_RNDN);
MPFR_ASSERTN (mpfr_ui_div (q, 1L, d, GMP_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q) && mpfr_sgn (q) > 0);
/* 1/-0 = -inf */
mpfr_set_ui (d, 0L, GMP_RNDN);
mpfr_neg (d, d, GMP_RNDN);
MPFR_ASSERTN (mpfr_ui_div (q, 1L, d, GMP_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q) && mpfr_sgn (q) < 0);
/* 0/1 = +0 */
mpfr_set_ui (d, 1L, GMP_RNDN);
MPFR_ASSERTN (mpfr_ui_div (q, 0L, d, GMP_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_cmp_ui (q, 0) == 0 && MPFR_IS_POS (q));
/* 0/-1 = -0 */
mpfr_set_si (d, -1, GMP_RNDN);
MPFR_ASSERTN (mpfr_ui_div (q, 0L, d, GMP_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_cmp_ui (q, 0) == 0 && MPFR_IS_NEG (q));
mpfr_clear (d);
mpfr_clear (q);
}
void
mpfr_set_inf (mpfr_ptr x, int sign)
{
MPFR_SET_INF(x);
if (sign >= 0)
MPFR_SET_POS(x);
else
MPFR_SET_NEG(x);
}
int
mpfr_set_si (mpfr_ptr x, long i, mp_rnd_t rnd_mode)
{
int inex;
mp_size_t xn;
unsigned int cnt, nbits;
mp_limb_t ai, *xp;
MPFR_CLEAR_FLAGS(x);
if (i == 0)
{
MPFR_SET_ZERO(x);
MPFR_SET_POS(x);
MPFR_RET(0);
}
xn = (MPFR_PREC(x)-1)/BITS_PER_MP_LIMB;
ai = SAFE_ABS(long, i);
count_leading_zeros(cnt, ai);
xp = MPFR_MANT(x);
xp[xn] = ai << cnt;
/* don't forget to put zero in lower limbs */
MPN_ZERO(xp, xn);
/* set sign */
if ((i < 0) ^ (MPFR_SIGN(x) < 0))
MPFR_CHANGE_SIGN(x);
MPFR_EXP(x) = nbits = BITS_PER_MP_LIMB - cnt;
inex = mpfr_check_range(x, rnd_mode);
if (inex)
return inex; /* underflow or overflow */
/* round if MPFR_PREC(x) smaller than length of i */
if (MPFR_PREC(x) < nbits)
{
int carry;
carry = mpfr_round_raw(xp+xn, xp+xn, nbits, (i < 0), MPFR_PREC(x),
rnd_mode, &inex);
if (carry)
{
mp_exp_t exp = MPFR_EXP(x);
if (exp == __mpfr_emax)
return mpfr_set_overflow(x, rnd_mode, (i < 0 ? -1 : 1));
MPFR_EXP(x)++;
xp[xn] = GMP_LIMB_HIGHBIT;
}
}
MPFR_RET(inex);
}
static int
my_setstr (mpfr_ptr t, const char *s)
{
if (strcmp (s, "min") == 0)
{
mpfr_setmin (t, mpfr_get_emin ());
MPFR_SET_POS (t);
return 0;
}
if (strcmp (s, "min+") == 0)
{
mpfr_setmin (t, mpfr_get_emin ());
MPFR_SET_POS (t);
mpfr_nextabove (t);
return 0;
}
if (strcmp (s, "max") == 0)
{
mpfr_setmax (t, mpfr_get_emax ());
MPFR_SET_POS (t);
return 0;
}
return mpfr_set_str (t, s, 10, MPFR_RNDN);
}
/* set f to the integer z */
int
mpfr_set_z (mpfr_ptr f, mpz_srcptr z, mp_rnd_t rnd_mode)
{
mp_size_t fn, zn, dif;
int k, sign_z, inex;
mp_limb_t *fp, *zp;
mp_exp_t exp;
MPFR_CLEAR_FLAGS (f); /* z cannot be NaN nor Inf */
sign_z = mpz_cmp_ui (z, 0);
if (sign_z == 0)
{
MPFR_SET_ZERO(f);
MPFR_SET_POS(f);
MPFR_RET(0);
}
fp = MPFR_MANT(f);
fn = 1 + (MPFR_PREC(f) - 1) / BITS_PER_MP_LIMB;
zn = ABS(SIZ(z));
dif = zn - fn;
zp = PTR(z);
count_leading_zeros(k, zp[zn-1]);
exp = (mp_prec_t) zn * BITS_PER_MP_LIMB - k;
/* The exponent will be exp or exp + 1 (due to rounding) */
if (exp > __mpfr_emax)
return mpfr_set_overflow(f, rnd_mode, sign_z);
if (exp + 1 < __mpfr_emin)
return mpfr_set_underflow(f, rnd_mode, sign_z);
if (MPFR_SIGN(f) * sign_z < 0)
MPFR_CHANGE_SIGN(f);
if (dif >= 0)
{
mp_limb_t cc;
int sh;
/* number has to be truncated */
if (k != 0)
{
mpn_lshift(fp, zp + dif, fn, k);
if (dif != 0)
fp[0] += zp[dif - 1] >> (BITS_PER_MP_LIMB - k);
}
static void
check_singular (void)
{
mpfr_t x, got;
mpfr_init2 (x, 100L);
mpfr_init2 (got, 100L);
/* sqrt(NaN) == NaN */
MPFR_SET_NAN (x);
MPFR_ASSERTN (test_sqrt (got, x, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_nan_p (got));
/* sqrt(-1) == NaN */
mpfr_set_si (x, -1L, MPFR_RNDZ);
MPFR_ASSERTN (test_sqrt (got, x, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_nan_p (got));
/* sqrt(+inf) == +inf */
MPFR_SET_INF (x);
MPFR_SET_POS (x);
MPFR_ASSERTN (test_sqrt (got, x, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (got));
/* sqrt(-inf) == NaN */
MPFR_SET_INF (x);
MPFR_SET_NEG (x);
MPFR_ASSERTN (test_sqrt (got, x, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_nan_p (got));
/* sqrt(+0) == +0 */
mpfr_set_si (x, 0L, MPFR_RNDZ);
MPFR_ASSERTN (test_sqrt (got, x, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_number_p (got));
MPFR_ASSERTN (mpfr_cmp_ui (got, 0L) == 0);
MPFR_ASSERTN (MPFR_IS_POS (got));
/* sqrt(-0) == -0 */
mpfr_set_si (x, 0L, MPFR_RNDZ);
MPFR_SET_NEG (x);
MPFR_ASSERTN (test_sqrt (got, x, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_number_p (got));
MPFR_ASSERTN (mpfr_cmp_ui (got, 0L) == 0);
MPFR_ASSERTN (MPFR_IS_NEG (got));
mpfr_clear (x);
mpfr_clear (got);
}
int
mpfr_set_si_2exp (mpfr_ptr x, long i, mp_exp_t e, mp_rnd_t rnd_mode)
{
if (i == 0)
{
MPFR_SET_ZERO (x);
MPFR_SET_POS (x);
MPFR_RET (0);
}
else
{
mp_size_t xn;
unsigned int cnt, nbits;
mp_limb_t ai, *xp;
int inex = 0;
/* FIXME: support int limbs (e.g. 16-bit limbs on 16-bit proc) */
ai = SAFE_ABS (unsigned long, i);
MPFR_ASSERTN (SAFE_ABS (unsigned long, i) == ai);
/* Position of the highest limb */
xn = (MPFR_PREC (x) - 1) / BITS_PER_MP_LIMB;
count_leading_zeros (cnt, ai);
MPFR_ASSERTD (cnt < BITS_PER_MP_LIMB); /* OK since i != 0 */
xp = MPFR_MANT(x);
xp[xn] = ai << cnt;
/* Zero the xn lower limbs. */
MPN_ZERO(xp, xn);
MPFR_SET_SIGN (x, i < 0 ? MPFR_SIGN_NEG : MPFR_SIGN_POS);
nbits = BITS_PER_MP_LIMB - cnt;
e += nbits; /* exponent _before_ the rounding */
/* round if MPFR_PREC(x) smaller than length of i */
if (MPFR_UNLIKELY (MPFR_PREC (x) < nbits) &&
MPFR_UNLIKELY (mpfr_round_raw (xp + xn, xp + xn, nbits, i < 0,
MPFR_PREC (x), rnd_mode, &inex)))
{
e++;
xp[xn] = MPFR_LIMB_HIGHBIT;
}
MPFR_CLEAR_FLAGS (x);
MPFR_EXP (x) = e;
return mpfr_check_range (x, inex, rnd_mode);
}
}
/* set f to the rational q */
int
mpfr_set_q (mpfr_ptr f, mpq_srcptr q, mp_rnd_t rnd)
{
mpz_srcptr num, den;
mpfr_t n, d;
int inexact;
mp_prec_t prec;
MPFR_CLEAR_FLAGS (f);
num = mpq_numref (q);
if (mpz_cmp_ui (num, 0) == 0)
{
MPFR_SET_ZERO (f);
MPFR_SET_POS (f);
MPFR_RET (0);
}
den = mpq_denref (q);
mpfr_save_emin_emax ();
prec = mpz_sizeinbase (num, 2);
if (prec < MPFR_PREC_MIN)
prec = MPFR_PREC_MIN;
mpfr_init2 (n, prec);
if (mpfr_set_z (n, num, GMP_RNDZ)) /* result is exact unless overflow */
{
mpfr_clear (n);
mpfr_restore_emin_emax ();
MPFR_SET_NAN (f);
MPFR_RET_NAN;
}
prec = mpz_sizeinbase(den, 2);
if (prec < MPFR_PREC_MIN)
prec = MPFR_PREC_MIN;
mpfr_init2 (d, prec);
if (mpfr_set_z (d, den, GMP_RNDZ)) /* result is exact unless overflow */
{
mpfr_clear (d);
mpfr_clear (n);
mpfr_restore_emin_emax ();
MPFR_SET_NAN (f);
MPFR_RET_NAN;
}
inexact = mpfr_div (f, n, d, rnd);
mpfr_clear (n);
mpfr_clear (d);
MPFR_RESTORE_RET (inexact, f, rnd);
}
static void
check_sgn(void)
{
mpfr_t x;
int i, s1, s2;
mpfr_init(x);
for(i = 0 ; i < 100 ; i++)
{
mpfr_urandomb (x, RANDS);
if (i&1)
{
MPFR_SET_POS(x);
s2 = 1;
}
else
{
MPFR_SET_NEG(x);
s2 = -1;
}
s1 = mpfr_sgn(x);
if (s1 < -1 || s1 > 1)
{
printf("Error for sgn: out of range.\n");
goto lexit;
}
else if (MPFR_IS_NAN(x) || MPFR_IS_ZERO(x))
{
if (s1 != 0)
{
printf("Error for sgn: Nan or Zero should return 0.\n");
goto lexit;
}
}
else if (s1 != s2)
{
printf("Error for sgn. Return %d instead of %d.\n", s1, s2);
goto lexit;
}
}
mpfr_clear(x);
return;
lexit:
mpfr_clear(x);
exit(1);
}
int
mpfr_set_ui_2exp (mpfr_ptr x, unsigned long i, mpfr_exp_t e, mpfr_rnd_t rnd_mode)
{
MPFR_SET_POS (x);
if (i == 0)
{
MPFR_SET_ZERO (x);
MPFR_RET (0);
}
else
{
mp_size_t xn;
unsigned int cnt, nbits;
mp_limb_t *xp;
int inex = 0;
/* FIXME: support int limbs (e.g. 16-bit limbs on 16-bit proc) */
MPFR_ASSERTD (i == (mp_limb_t) i);
/* Position of the highest limb */
xn = (MPFR_PREC (x) - 1) / GMP_NUMB_BITS;
count_leading_zeros (cnt, (mp_limb_t) i);
MPFR_ASSERTD (cnt < GMP_NUMB_BITS); /* OK since i != 0 */
xp = MPFR_MANT(x);
xp[xn] = ((mp_limb_t) i) << cnt;
/* Zero the xn lower limbs. */
MPN_ZERO(xp, xn);
nbits = GMP_NUMB_BITS - cnt;
e += nbits; /* exponent _before_ the rounding */
/* round if MPFR_PREC(x) smaller than length of i */
if (MPFR_UNLIKELY (MPFR_PREC (x) < nbits) &&
MPFR_UNLIKELY (mpfr_round_raw (xp + xn, xp + xn, nbits, 0,
MPFR_PREC (x), rnd_mode, &inex)))
{
e++;
xp[xn] = MPFR_LIMB_HIGHBIT;
}
MPFR_EXP (x) = e;
return mpfr_check_range (x, inex, rnd_mode);
}
}
int
mpfr_dim (mpfr_ptr z, mpfr_srcptr x, mpfr_srcptr y, mpfr_rnd_t rnd_mode)
{
if (MPFR_IS_NAN(x) || MPFR_IS_NAN(y))
{
MPFR_SET_NAN(z);
MPFR_RET_NAN;
}
if (mpfr_cmp (x,y) > 0)
return mpfr_sub (z, x, y, rnd_mode);
else
{
MPFR_SET_ZERO(z);
MPFR_SET_POS(z);
MPFR_RET(0);
}
}
/*
Parameters:
s - the input floating-point number
n, p - parameters from the algorithm
tc - an array of p floating-point numbers tc[1]..tc[p]
Output:
b is the result, i.e.
sum(tc[i]*product((s+2j)*(s+2j-1)/n^2,j=1..i-1), i=1..p)*s*n^(-s-1)
*/
static void
mpfr_zeta_part_b (mpfr_t b, mpfr_srcptr s, int n, int p, mpfr_t *tc)
{
mpfr_t s1, d, u;
unsigned long n2;
int l, t;
MPFR_GROUP_DECL (group);
if (p == 0)
{
MPFR_SET_ZERO (b);
MPFR_SET_POS (b);
return;
}
n2 = n * n;
MPFR_GROUP_INIT_3 (group, MPFR_PREC (b), s1, d, u);
/* t equals 2p-2, 2p-3, ... ; s1 equals s+t */
t = 2 * p - 2;
mpfr_set (d, tc[p], GMP_RNDN);
for (l = 1; l < p; l++)
{
mpfr_add_ui (s1, s, t, GMP_RNDN); /* s + (2p-2l) */
mpfr_mul (d, d, s1, GMP_RNDN);
t = t - 1;
mpfr_add_ui (s1, s, t, GMP_RNDN); /* s + (2p-2l-1) */
mpfr_mul (d, d, s1, GMP_RNDN);
t = t - 1;
mpfr_div_ui (d, d, n2, GMP_RNDN);
mpfr_add (d, d, tc[p-l], GMP_RNDN);
/* since s is positive and the tc[i] have alternate signs,
the following is unlikely */
if (MPFR_UNLIKELY (mpfr_cmpabs (d, tc[p-l]) > 0))
mpfr_set (d, tc[p-l], GMP_RNDN);
}
mpfr_mul (d, d, s, GMP_RNDN);
mpfr_add (s1, s, __gmpfr_one, GMP_RNDN);
mpfr_neg (s1, s1, GMP_RNDN);
mpfr_ui_pow (u, n, s1, GMP_RNDN);
mpfr_mul (b, d, u, GMP_RNDN);
MPFR_GROUP_CLEAR (group);
}
static void
check_special (void)
{
mpfr_t x;
int ret = 0;
mpfr_init (x);
MPFR_SET_ZERO (x);
if ((mpfr_sgn) (x) != 0 || mpfr_sgn (x) != 0)
{
printf("Sgn error for 0.\n");
ret = 1;
}
MPFR_SET_INF (x);
MPFR_SET_POS (x);
if ((mpfr_sgn) (x) != 1 || mpfr_sgn (x) != 1)
{
printf("Sgn error for +Inf.\n");
ret = 1;
}
MPFR_SET_INF (x);
MPFR_SET_NEG (x);
if ((mpfr_sgn) (x) != -1 || mpfr_sgn (x) != -1)
{
printf("Sgn error for -Inf.\n");
ret = 1;
}
MPFR_SET_NAN (x);
mpfr_clear_flags ();
if ((mpfr_sgn) (x) != 0 || !mpfr_erangeflag_p ())
{
printf("Sgn error for NaN.\n");
ret = 1;
}
mpfr_clear_flags ();
if (mpfr_sgn (x) != 0 || !mpfr_erangeflag_p ())
{
printf("Sgn error for NaN.\n");
ret = 1;
}
mpfr_clear (x);
if (ret)
exit (ret);
}
int
mpfr_exp (mpfr_ptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_exp_t expx;
mpfr_prec_t precy;
int inexact;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_LOG_FUNC (("x[%#R]=%R rnd=%d", x, x, rnd_mode),
("y[%#R]=%R inexact=%d", y, y, inexact));
if (MPFR_UNLIKELY( MPFR_IS_SINGULAR(x) ))
{
if (MPFR_IS_NAN(x))
{
MPFR_SET_NAN(y);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF(x))
{
if (MPFR_IS_POS(x))
MPFR_SET_INF(y);
else
MPFR_SET_ZERO(y);
MPFR_SET_POS(y);
MPFR_RET(0);
}
else
{
MPFR_ASSERTD(MPFR_IS_ZERO(x));
return mpfr_set_ui (y, 1, rnd_mode);
}
}
/* First, let's detect most overflow and underflow cases. */
{
mpfr_t e, bound;
/* We must extended the exponent range and save the flags now. */
MPFR_SAVE_EXPO_MARK (expo);
mpfr_init2 (e, sizeof (mpfr_exp_t) * CHAR_BIT);
mpfr_init2 (bound, 32);
inexact = mpfr_set_exp_t (e, expo.saved_emax, MPFR_RNDN);
MPFR_ASSERTD (inexact == 0);
mpfr_const_log2 (bound, expo.saved_emax < 0 ? MPFR_RNDD : MPFR_RNDU);
mpfr_mul (bound, bound, e, MPFR_RNDU);
if (MPFR_UNLIKELY (mpfr_cmp (x, bound) >= 0))
{
/* x > log(2^emax), thus exp(x) > 2^emax */
mpfr_clears (e, bound, (mpfr_ptr) 0);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_overflow (y, rnd_mode, 1);
}
inexact = mpfr_set_exp_t (e, expo.saved_emin, MPFR_RNDN);
MPFR_ASSERTD (inexact == 0);
inexact = mpfr_sub_ui (e, e, 2, MPFR_RNDN);
MPFR_ASSERTD (inexact == 0);
mpfr_const_log2 (bound, expo.saved_emin < 0 ? MPFR_RNDU : MPFR_RNDD);
mpfr_mul (bound, bound, e, MPFR_RNDD);
if (MPFR_UNLIKELY (mpfr_cmp (x, bound) <= 0))
{
/* x < log(2^(emin - 2)), thus exp(x) < 2^(emin - 2) */
mpfr_clears (e, bound, (mpfr_ptr) 0);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_underflow (y, rnd_mode == MPFR_RNDN ? MPFR_RNDZ : rnd_mode,
1);
}
/* Other overflow/underflow cases must be detected
by the generic routines. */
mpfr_clears (e, bound, (mpfr_ptr) 0);
MPFR_SAVE_EXPO_FREE (expo);
}
expx = MPFR_GET_EXP (x);
precy = MPFR_PREC (y);
/* if x < 2^(-precy), then exp(x) i.e. gives 1 +/- 1 ulp(1) */
if (MPFR_UNLIKELY (expx < 0 && (mpfr_uexp_t) (-expx) > precy))
{
mpfr_exp_t emin = __gmpfr_emin;
mpfr_exp_t emax = __gmpfr_emax;
int signx = MPFR_SIGN (x);
MPFR_SET_POS (y);
if (MPFR_IS_NEG_SIGN (signx) && (rnd_mode == MPFR_RNDD ||
rnd_mode == MPFR_RNDZ))
{
__gmpfr_emin = 0;
__gmpfr_emax = 0;
mpfr_setmax (y, 0); /* y = 1 - epsilon */
inexact = -1;
}
else
{
__gmpfr_emin = 1;
__gmpfr_emax = 1;
mpfr_setmin (y, 1); /* y = 1 */
if (MPFR_IS_POS_SIGN (signx) && (rnd_mode == MPFR_RNDU ||
rnd_mode == MPFR_RNDA))
{
mp_size_t yn;
int sh;
yn = 1 + (MPFR_PREC(y) - 1) / GMP_NUMB_BITS;
sh = (mpfr_prec_t) yn * GMP_NUMB_BITS - MPFR_PREC(y);
MPFR_MANT(y)[0] += MPFR_LIMB_ONE << sh;
inexact = 1;
}
else
inexact = -MPFR_FROM_SIGN_TO_INT(signx);
}
__gmpfr_emin = emin;
__gmpfr_emax = emax;
}
else /* General case */
{
if (MPFR_UNLIKELY (precy >= MPFR_EXP_THRESHOLD))
/* mpfr_exp_3 saves the exponent range and flags itself, otherwise
the flag changes in mpfr_exp_3 are lost */
inexact = mpfr_exp_3 (y, x, rnd_mode); /* O(M(n) log(n)^2) */
else
{
MPFR_SAVE_EXPO_MARK (expo);
inexact = mpfr_exp_2 (y, x, rnd_mode); /* O(n^(1/3) M(n)) */
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, __gmpfr_flags);
MPFR_SAVE_EXPO_FREE (expo);
}
}
return mpfr_check_range (y, inexact, rnd_mode);
}
int
main (void)
{
mpfr_t x, y, z;
int i, j, k;
tests_start_mpfr ();
mpfr_init (x);
mpfr_init (y);
mpfr_init (z);
for (i = 0; i <= 1; i++)
for (j = 0; j <= 1; j++)
for (k = 0; k <= 5; k++)
{
mpfr_set_nan (x);
i ? MPFR_SET_NEG (x) : MPFR_SET_POS (x);
mpfr_set_nan (y);
j ? MPFR_SET_NEG (y) : MPFR_SET_POS (y);
copysign_variant (z, x, y, MPFR_RNDN, k);
if (MPFR_SIGN (z) != MPFR_SIGN (y) || !mpfr_nanflag_p ())
{
printf ("Error in mpfr_copysign (%cNaN, %cNaN)\n",
i ? '-' : '+', j ? '-' : '+');
exit (1);
}
mpfr_set_si (x, i ? -1250 : 1250, MPFR_RNDN);
mpfr_set_nan (y);
j ? MPFR_SET_NEG (y) : MPFR_SET_POS (y);
copysign_variant (z, x, y, MPFR_RNDN, k);
if (i != j)
mpfr_neg (x, x, MPFR_RNDN);
if (! mpfr_equal_p (z, x) || mpfr_nanflag_p ())
{
printf ("Error in mpfr_copysign (%c1250, %cNaN)\n",
i ? '-' : '+', j ? '-' : '+');
exit (1);
}
mpfr_set_si (x, i ? -1250 : 1250, MPFR_RNDN);
mpfr_set_si (y, j ? -1717 : 1717, MPFR_RNDN);
copysign_variant (z, x, y, MPFR_RNDN, k);
if (i != j)
mpfr_neg (x, x, MPFR_RNDN);
if (! mpfr_equal_p (z, x) || mpfr_nanflag_p ())
{
printf ("Error in mpfr_copysign (%c1250, %c1717)\n",
i ? '-' : '+', j ? '-' : '+');
exit (1);
}
}
mpfr_clear (x);
mpfr_clear (y);
mpfr_clear (z);
tests_end_mpfr ();
return 0;
}
/* agm(x,y) is between x and y, so we don't need to save exponent range */
int
mpfr_agm (mpfr_ptr r, mpfr_srcptr op2, mpfr_srcptr op1, mp_rnd_t rnd_mode)
{
int compare, inexact;
mp_size_t s;
mp_prec_t p, q;
mp_limb_t *up, *vp, *tmpp;
mpfr_t u, v, tmp;
unsigned long n; /* number of iterations */
unsigned long err = 0;
MPFR_ZIV_DECL (loop);
MPFR_TMP_DECL(marker);
MPFR_LOG_FUNC (("op2[%#R]=%R op1[%#R]=%R rnd=%d", op2,op2,op1,op1,rnd_mode),
("r[%#R]=%R inexact=%d", r, r, inexact));
/* Deal with special values */
if (MPFR_ARE_SINGULAR (op1, op2))
{
/* If a or b is NaN, the result is NaN */
if (MPFR_IS_NAN(op1) || MPFR_IS_NAN(op2))
{
MPFR_SET_NAN(r);
MPFR_RET_NAN;
}
/* now one of a or b is Inf or 0 */
/* If a and b is +Inf, the result is +Inf.
Otherwise if a or b is -Inf or 0, the result is NaN */
else if (MPFR_IS_INF(op1) || MPFR_IS_INF(op2))
{
if (MPFR_IS_STRICTPOS(op1) && MPFR_IS_STRICTPOS(op2))
{
MPFR_SET_INF(r);
MPFR_SET_SAME_SIGN(r, op1);
MPFR_RET(0); /* exact */
}
else
{
MPFR_SET_NAN(r);
MPFR_RET_NAN;
}
}
else /* a and b are neither NaN nor Inf, and one is zero */
{ /* If a or b is 0, the result is +0 since a sqrt is positive */
MPFR_ASSERTD (MPFR_IS_ZERO (op1) || MPFR_IS_ZERO (op2));
MPFR_SET_POS (r);
MPFR_SET_ZERO (r);
MPFR_RET (0); /* exact */
}
}
MPFR_CLEAR_FLAGS (r);
/* If a or b is negative (excluding -Infinity), the result is NaN */
if (MPFR_UNLIKELY(MPFR_IS_NEG(op1) || MPFR_IS_NEG(op2)))
{
MPFR_SET_NAN(r);
MPFR_RET_NAN;
}
/* Precision of the following calculus */
q = MPFR_PREC(r);
p = q + MPFR_INT_CEIL_LOG2(q) + 15;
MPFR_ASSERTD (p >= 7); /* see algorithms.tex */
s = (p - 1) / BITS_PER_MP_LIMB + 1;
/* b (op2) and a (op1) are the 2 operands but we want b >= a */
compare = mpfr_cmp (op1, op2);
if (MPFR_UNLIKELY( compare == 0 ))
{
mpfr_set (r, op1, rnd_mode);
MPFR_RET (0); /* exact */
}
else if (compare > 0)
{
mpfr_srcptr t = op1;
op1 = op2;
op2 = t;
}
/* Now b(=op2) >= a (=op1) */
MPFR_TMP_MARK(marker);
/* Main loop */
MPFR_ZIV_INIT (loop, p);
for (;;)
{
mp_prec_t eq;
/* Init temporary vars */
MPFR_TMP_INIT (up, u, p, s);
MPFR_TMP_INIT (vp, v, p, s);
MPFR_TMP_INIT (tmpp, tmp, p, s);
/* Calculus of un and vn */
mpfr_mul (u, op1, op2, GMP_RNDN); /* Faster since PREC(op) < PREC(u) */
mpfr_sqrt (u, u, GMP_RNDN);
mpfr_add (v, op1, op2, GMP_RNDN); /* add with !=prec is still good*/
mpfr_div_2ui (v, v, 1, GMP_RNDN);
n = 1;
while (mpfr_cmp2 (u, v, &eq) != 0 && eq <= p - 2)
{
mpfr_add (tmp, u, v, GMP_RNDN);
mpfr_div_2ui (tmp, tmp, 1, GMP_RNDN);
/* See proof in algorithms.tex */
if (4*eq > p)
{
mpfr_t w;
/* tmp = U(k) */
mpfr_init2 (w, (p + 1) / 2);
mpfr_sub (w, v, u, GMP_RNDN); /* e = V(k-1)-U(k-1) */
mpfr_sqr (w, w, GMP_RNDN); /* e = e^2 */
mpfr_div_2ui (w, w, 4, GMP_RNDN); /* e*= (1/2)^2*1/4 */
mpfr_div (w, w, tmp, GMP_RNDN); /* 1/4*e^2/U(k) */
mpfr_sub (v, tmp, w, GMP_RNDN);
err = MPFR_GET_EXP (tmp) - MPFR_GET_EXP (v); /* 0 or 1 */
mpfr_clear (w);
break;
}
mpfr_mul (u, u, v, GMP_RNDN);
mpfr_sqrt (u, u, GMP_RNDN);
mpfr_swap (v, tmp);
n ++;
}
/* the error on v is bounded by (18n+51) ulps, or twice if there
was an exponent loss in the final subtraction */
err += MPFR_INT_CEIL_LOG2(18 * n + 51); /* 18n+51 should not overflow
since n is about log(p) */
/* we should have n+2 <= 2^(p/4) [see algorithms.tex] */
if (MPFR_LIKELY (MPFR_INT_CEIL_LOG2(n + 2) <= p / 4 &&
MPFR_CAN_ROUND (v, p - err, q, rnd_mode)))
break; /* Stop the loop */
/* Next iteration */
MPFR_ZIV_NEXT (loop, p);
s = (p - 1) / BITS_PER_MP_LIMB + 1;
}
MPFR_ZIV_FREE (loop);
/* Setting of the result */
inexact = mpfr_set (r, v, rnd_mode);
/* Let's clean */
MPFR_TMP_FREE(marker);
return inexact; /* agm(u,v) can be exact for u, v rational only for u=v.
Proof (due to Nicolas Brisebarre): it suffices to consider
u=1 and v<1. Then 1/AGM(1,v) = 2F1(1/2,1/2,1;1-v^2),
and a theorem due to G.V. Chudnovsky states that for x a
non-zero algebraic number with |x|<1, then
2F1(1/2,1/2,1;x) and 2F1(-1/2,1/2,1;x) are algebraically
independent over Q. */
}
static void
check_special (void)
{
mpfr_t a, d, q;
mpfr_exp_t emax, emin;
int i;
mpfr_init2 (a, 100L);
mpfr_init2 (d, 100L);
mpfr_init2 (q, 100L);
/* 1/nan == nan */
mpfr_set_ui (a, 1L, MPFR_RNDN);
MPFR_SET_NAN (d);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (__gmpfr_flags == MPFR_FLAGS_NAN);
/* nan/1 == nan */
MPFR_SET_NAN (a);
mpfr_set_ui (d, 1L, MPFR_RNDN);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (__gmpfr_flags == MPFR_FLAGS_NAN);
/* +inf/1 == +inf */
MPFR_SET_INF (a);
MPFR_SET_POS (a);
mpfr_set_ui (d, 1L, MPFR_RNDN);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q));
MPFR_ASSERTN (mpfr_sgn (q) > 0);
MPFR_ASSERTN (__gmpfr_flags == 0);
/* +inf/-1 == -inf */
MPFR_SET_INF (a);
MPFR_SET_POS (a);
mpfr_set_si (d, -1, MPFR_RNDN);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q));
MPFR_ASSERTN (mpfr_sgn (q) < 0);
MPFR_ASSERTN (__gmpfr_flags == 0);
/* -inf/1 == -inf */
MPFR_SET_INF (a);
MPFR_SET_NEG (a);
mpfr_set_ui (d, 1L, MPFR_RNDN);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q));
MPFR_ASSERTN (mpfr_sgn (q) < 0);
MPFR_ASSERTN (__gmpfr_flags == 0);
/* -inf/-1 == +inf */
MPFR_SET_INF (a);
MPFR_SET_NEG (a);
mpfr_set_si (d, -1, MPFR_RNDN);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q));
MPFR_ASSERTN (mpfr_sgn (q) > 0);
MPFR_ASSERTN (__gmpfr_flags == 0);
/* 1/+inf == +0 */
mpfr_set_ui (a, 1L, MPFR_RNDN);
MPFR_SET_INF (d);
MPFR_SET_POS (d);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_number_p (q));
MPFR_ASSERTN (mpfr_sgn (q) == 0);
MPFR_ASSERTN (MPFR_IS_POS (q));
MPFR_ASSERTN (__gmpfr_flags == 0);
/* 1/-inf == -0 */
mpfr_set_ui (a, 1L, MPFR_RNDN);
MPFR_SET_INF (d);
MPFR_SET_NEG (d);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_number_p (q));
MPFR_ASSERTN (mpfr_sgn (q) == 0);
MPFR_ASSERTN (MPFR_IS_NEG (q));
MPFR_ASSERTN (__gmpfr_flags == 0);
/* -1/+inf == -0 */
mpfr_set_si (a, -1, MPFR_RNDN);
MPFR_SET_INF (d);
MPFR_SET_POS (d);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_number_p (q));
MPFR_ASSERTN (mpfr_sgn (q) == 0);
MPFR_ASSERTN (MPFR_IS_NEG (q));
MPFR_ASSERTN (__gmpfr_flags == 0);
/* -1/-inf == +0 */
mpfr_set_si (a, -1, MPFR_RNDN);
MPFR_SET_INF (d);
MPFR_SET_NEG (d);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_number_p (q));
MPFR_ASSERTN (mpfr_sgn (q) == 0);
MPFR_ASSERTN (MPFR_IS_POS (q));
MPFR_ASSERTN (__gmpfr_flags == 0);
/* 0/0 == nan */
mpfr_set_ui (a, 0L, MPFR_RNDN);
mpfr_set_ui (d, 0L, MPFR_RNDN);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (__gmpfr_flags == MPFR_FLAGS_NAN);
/* +inf/+inf == nan */
MPFR_SET_INF (a);
MPFR_SET_POS (a);
MPFR_SET_INF (d);
MPFR_SET_POS (d);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (__gmpfr_flags == MPFR_FLAGS_NAN);
/* 1/+0 = +inf */
mpfr_set_ui (a, 1, MPFR_RNDZ);
mpfr_set_ui (d, 0, MPFR_RNDZ);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q) && mpfr_sgn (q) > 0);
MPFR_ASSERTN (__gmpfr_flags == MPFR_FLAGS_DIVBY0);
/* 1/-0 = -inf */
mpfr_set_ui (a, 1, MPFR_RNDZ);
mpfr_set_ui (d, 0, MPFR_RNDZ);
mpfr_neg (d, d, MPFR_RNDZ);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q) && mpfr_sgn (q) < 0);
MPFR_ASSERTN (__gmpfr_flags == MPFR_FLAGS_DIVBY0);
/* -1/+0 = -inf */
mpfr_set_si (a, -1, MPFR_RNDZ);
mpfr_set_ui (d, 0, MPFR_RNDZ);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q) && mpfr_sgn (q) < 0);
MPFR_ASSERTN (__gmpfr_flags == MPFR_FLAGS_DIVBY0);
/* -1/-0 = +inf */
mpfr_set_si (a, -1, MPFR_RNDZ);
mpfr_set_ui (d, 0, MPFR_RNDZ);
mpfr_neg (d, d, MPFR_RNDZ);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q) && mpfr_sgn (q) > 0);
MPFR_ASSERTN (__gmpfr_flags == MPFR_FLAGS_DIVBY0);
/* +inf/+0 = +inf */
MPFR_SET_INF (a);
MPFR_SET_POS (a);
mpfr_set_ui (d, 0, MPFR_RNDZ);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q) && mpfr_sgn (q) > 0);
MPFR_ASSERTN (__gmpfr_flags == 0);
/* +inf/-0 = -inf */
MPFR_SET_INF (a);
MPFR_SET_POS (a);
mpfr_set_ui (d, 0, MPFR_RNDZ);
mpfr_neg (d, d, MPFR_RNDZ);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q) && mpfr_sgn (q) < 0);
MPFR_ASSERTN (__gmpfr_flags == 0);
/* -inf/+0 = -inf */
MPFR_SET_INF (a);
MPFR_SET_NEG (a);
mpfr_set_ui (d, 0, MPFR_RNDZ);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q) && mpfr_sgn (q) < 0);
MPFR_ASSERTN (__gmpfr_flags == 0);
/* -inf/-0 = +inf */
MPFR_SET_INF (a);
MPFR_SET_NEG (a);
mpfr_set_ui (d, 0, MPFR_RNDZ);
mpfr_neg (d, d, MPFR_RNDZ);
mpfr_clear_flags ();
MPFR_ASSERTN (test_div (q, a, d, MPFR_RNDZ) == 0); /* exact */
MPFR_ASSERTN (mpfr_inf_p (q) && mpfr_sgn (q) > 0);
MPFR_ASSERTN (__gmpfr_flags == 0);
/* check overflow */
emax = mpfr_get_emax ();
set_emax (1);
mpfr_set_ui (a, 1, MPFR_RNDZ);
mpfr_set_ui (d, 1, MPFR_RNDZ);
mpfr_div_2exp (d, d, 1, MPFR_RNDZ);
mpfr_clear_flags ();
test_div (q, a, d, MPFR_RNDU); /* 1 / 0.5 = 2 -> overflow */
MPFR_ASSERTN (mpfr_inf_p (q) && mpfr_sgn (q) > 0);
MPFR_ASSERTN (__gmpfr_flags == (MPFR_FLAGS_OVERFLOW | MPFR_FLAGS_INEXACT));
set_emax (emax);
/* check underflow */
emin = mpfr_get_emin ();
set_emin (-1);
mpfr_set_ui (a, 1, MPFR_RNDZ);
mpfr_div_2exp (a, a, 2, MPFR_RNDZ);
mpfr_set_prec (d, mpfr_get_prec (q) + 8);
for (i = -1; i <= 1; i++)
{
int sign;
/* Test 2^(-2) / (+/- (2 + eps)), with eps < 0, eps = 0, eps > 0.
-> underflow.
With div.c r5513, this test fails for eps > 0 in MPFR_RNDN. */
mpfr_set_ui (d, 2, MPFR_RNDZ);
if (i < 0)
mpfr_nextbelow (d);
if (i > 0)
mpfr_nextabove (d);
for (sign = 0; sign <= 1; sign++)
{
mpfr_clear_flags ();
test_div (q, a, d, MPFR_RNDZ); /* result = 0 */
MPFR_ASSERTN (__gmpfr_flags ==
(MPFR_FLAGS_UNDERFLOW | MPFR_FLAGS_INEXACT));
MPFR_ASSERTN (sign ? MPFR_IS_NEG (q) : MPFR_IS_POS (q));
MPFR_ASSERTN (MPFR_IS_ZERO (q));
mpfr_clear_flags ();
test_div (q, a, d, MPFR_RNDN); /* result = 0 iff eps >= 0 */
MPFR_ASSERTN (__gmpfr_flags ==
(MPFR_FLAGS_UNDERFLOW | MPFR_FLAGS_INEXACT));
MPFR_ASSERTN (sign ? MPFR_IS_NEG (q) : MPFR_IS_POS (q));
if (i < 0)
mpfr_nexttozero (q);
MPFR_ASSERTN (MPFR_IS_ZERO (q));
mpfr_neg (d, d, MPFR_RNDN);
}
}
set_emin (emin);
mpfr_clear (a);
mpfr_clear (d);
mpfr_clear (q);
}
static void
test_underflow1 (void)
{
mpfr_t x, y, z, r;
int inex, signy, signz, rnd, err = 0;
mpfr_inits2 (8, x, y, z, r, (void *) 0);
MPFR_SET_POS (x);
mpfr_setmin (x, mpfr_get_emin ()); /* x = 0.1@emin */
for (signy = -1; signy <= 1; signy += 2)
{
mpfr_set_si_2exp (y, signy, -1, GMP_RNDN); /* |y| = 1/2 */
for (signz = -3; signz <= 3; signz += 2)
{
RND_LOOP (rnd)
{
mpfr_set_si (z, signz, GMP_RNDN);
if (ABS (signz) != 1)
mpfr_setmax (z, mpfr_get_emax ());
/* |z| = 1 or 2^emax - ulp */
mpfr_clear_flags ();
inex = mpfr_fma (r, x, y, z, rnd);
#define ERRTU1 "Error in test_underflow1 (signy = %d, signz = %d, %s)\n "
if (mpfr_nanflag_p ())
{
printf (ERRTU1 "NaN flag is set\n", signy, signz,
mpfr_print_rnd_mode (rnd));
err = 1;
}
if (signy < 0 && (rnd == GMP_RNDD ||
(rnd == GMP_RNDZ && signz > 0)))
mpfr_nextbelow (z);
if (signy > 0 && (rnd == GMP_RNDU ||
(rnd == GMP_RNDZ && signz < 0)))
mpfr_nextabove (z);
if ((mpfr_overflow_p () != 0) ^ (mpfr_inf_p (z) != 0))
{
printf (ERRTU1 "wrong overflow flag\n", signy, signz,
mpfr_print_rnd_mode (rnd));
err = 1;
}
if (mpfr_underflow_p ())
{
printf (ERRTU1 "underflow flag is set\n", signy, signz,
mpfr_print_rnd_mode (rnd));
err = 1;
}
if (! mpfr_equal_p (r, z))
{
printf (ERRTU1 "got ", signy, signz,
mpfr_print_rnd_mode (rnd));
mpfr_print_binary (r);
printf (" instead of ");
mpfr_print_binary (z);
printf ("\n");
err = 1;
}
if (inex >= 0 && (rnd == GMP_RNDD ||
(rnd == GMP_RNDZ && signz > 0) ||
(rnd == GMP_RNDN && signy > 0)))
{
printf (ERRTU1 "ternary value = %d instead of < 0\n",
signy, signz, mpfr_print_rnd_mode (rnd), inex);
err = 1;
}
if (inex <= 0 && (rnd == GMP_RNDU ||
(rnd == GMP_RNDZ && signz < 0) ||
(rnd == GMP_RNDN && signy < 0)))
{
printf (ERRTU1 "ternary value = %d instead of > 0\n",
signy, signz, mpfr_print_rnd_mode (rnd), inex);
err = 1;
}
}
}
}
if (err)
exit (1);
mpfr_clears (x, y, z, r, (void *) 0);
}
static void
test_overflow2 (void)
{
mpfr_t x, y, z, r;
int i, inex, rnd, err = 0;
mpfr_inits2 (8, x, y, z, r, (void *) 0);
MPFR_SET_POS (x);
mpfr_setmin (x, mpfr_get_emax ()); /* x = 0.1@emax */
mpfr_set_si (y, -2, GMP_RNDN); /* y = -2 */
/* The intermediate multiplication x * y will overflow. */
for (i = -9; i <= 9; i++)
RND_LOOP (rnd)
{
int inf, overflow;
inf = rnd == GMP_RNDN || rnd == GMP_RNDD;
overflow = inf || i <= 0;
inex = mpfr_set_si_2exp (z, i, mpfr_get_emin (), GMP_RNDN);
MPFR_ASSERTN (inex == 0);
mpfr_clear_flags ();
/* One has: x * y = -1@emax exactly (but not representable). */
inex = mpfr_fma (r, x, y, z, rnd);
if (overflow ^ (mpfr_overflow_p () != 0))
{
printf ("Error in test_overflow2 (i = %d, %s): wrong overflow"
" flag (should be %d)\n", i, mpfr_print_rnd_mode (rnd),
overflow);
err = 1;
}
if (mpfr_nanflag_p ())
{
printf ("Error in test_overflow2 (i = %d, %s): NaN flag should"
" not be set\n", i, mpfr_print_rnd_mode (rnd));
err = 1;
}
if (mpfr_nan_p (r))
{
printf ("Error in test_overflow2 (i = %d, %s): got NaN\n",
i, mpfr_print_rnd_mode (rnd));
err = 1;
}
else if (MPFR_SIGN (r) >= 0)
{
printf ("Error in test_overflow2 (i = %d, %s): wrong sign "
"(+ instead of -)\n", i, mpfr_print_rnd_mode (rnd));
err = 1;
}
else if (inf && ! mpfr_inf_p (r))
{
printf ("Error in test_overflow2 (i = %d, %s): expected -Inf,"
" got\n", i, mpfr_print_rnd_mode (rnd));
mpfr_dump (r);
err = 1;
}
else if (!inf && (mpfr_inf_p (r) ||
(mpfr_nextbelow (r), ! mpfr_inf_p (r))))
{
printf ("Error in test_overflow2 (i = %d, %s): expected -MAX,"
" got\n", i, mpfr_print_rnd_mode (rnd));
mpfr_dump (r);
err = 1;
}
if (inf ? inex >= 0 : inex <= 0)
{
printf ("Error in test_overflow2 (i = %d, %s): wrong inexact"
" flag (got %d)\n", i, mpfr_print_rnd_mode (rnd), inex);
err = 1;
}
}
if (err)
exit (1);
mpfr_clears (x, y, z, r, (void *) 0);
}
static void
check_cmp (int argc, char *argv[])
{
mpfr_t x, y;
int n, k;
mpfr_inits2 (53, x, y, (mpfr_ptr) 0);
mpfr_set_ui(x, 1, MPFR_RNDN);
(mpfr_abs) (x, x, MPFR_RNDN);
if (mpfr_cmp_ui (x, 1))
{
printf ("Error in mpfr_abs(1.0)\n");
exit (1);
}
mpfr_set_si(x, -1, MPFR_RNDN);
mpfr_abs(x, x, MPFR_RNDN);
if (mpfr_cmp_ui (x, 1))
{
printf ("Error in mpfr_abs(1.0)\n");
exit (1);
}
mpfr_set_si(x, -1, MPFR_RNDN);
mpfr_abs(x, x, MPFR_RNDN);
if (mpfr_cmp_ui (x, 1))
{
printf ("Error in mpfr_abs(-1.0)\n");
exit (1);
}
mpfr_set_inf (x, 1);
mpfr_abs (x, x, MPFR_RNDN);
if (!mpfr_inf_p(x) || (mpfr_sgn(x) <= 0))
{
printf ("Error in mpfr_abs(Inf).\n");
exit (1);
}
mpfr_set_inf (x, -1);
mpfr_abs (x, x, MPFR_RNDN);
if (!mpfr_inf_p(x) || (mpfr_sgn(x) <= 0))
{
printf ("Error in mpfr_abs(-Inf).\n");
exit (1);
}
MPFR_SET_NAN(x);
mpfr_abs (x, x, MPFR_RNDN);
if (!MPFR_IS_NAN(x))
{
printf ("Error in mpfr_abs(NAN).\n");
exit (1);
}
n = (argc==1) ? 25000 : atoi(argv[1]);
for (k = 1; k <= n; k++)
{
mpfr_rnd_t rnd;
int sign = SIGN_RAND ();
mpfr_urandomb (x, RANDS);
MPFR_SET_SIGN (x, sign);
rnd = RND_RAND ();
mpfr_abs (y, x, rnd);
MPFR_SET_POS (x);
if (mpfr_cmp (x, y))
{
printf ("Mismatch for sign=%d and x=", sign);
mpfr_print_binary (x);
printf ("\nResults=");
mpfr_print_binary (y);
putchar ('\n');
exit (1);
}
}
mpfr_clears (x, y, (mpfr_ptr) 0);
}
/* We use the reflection formula
Gamma(1+t) Gamma(1-t) = - Pi t / sin(Pi (1 + t))
in order to treat the case x <= 1,
i.e. with x = 1-t, then Gamma(x) = -Pi*(1-x)/sin(Pi*(2-x))/GAMMA(2-x)
*/
int
mpfr_gamma (mpfr_ptr gamma, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_t xp, GammaTrial, tmp, tmp2;
mpz_t fact;
mpfr_prec_t realprec;
int compared, is_integer;
int inex = 0; /* 0 means: result gamma not set yet */
MPFR_GROUP_DECL (group);
MPFR_SAVE_EXPO_DECL (expo);
MPFR_ZIV_DECL (loop);
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg rnd=%d", mpfr_get_prec (x), mpfr_log_prec, x, rnd_mode),
("gamma[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (gamma), mpfr_log_prec, gamma, inex));
/* Trivial cases */
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x)))
{
if (MPFR_IS_NAN (x))
{
MPFR_SET_NAN (gamma);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (x))
{
if (MPFR_IS_NEG (x))
{
MPFR_SET_NAN (gamma);
MPFR_RET_NAN;
}
else
{
MPFR_SET_INF (gamma);
MPFR_SET_POS (gamma);
MPFR_RET (0); /* exact */
}
}
else /* x is zero */
{
MPFR_ASSERTD(MPFR_IS_ZERO(x));
MPFR_SET_INF(gamma);
MPFR_SET_SAME_SIGN(gamma, x);
MPFR_SET_DIVBY0 ();
MPFR_RET (0); /* exact */
}
}
/* Check for tiny arguments, where gamma(x) ~ 1/x - euler + ....
We know from "Bound on Runs of Zeros and Ones for Algebraic Functions",
Proceedings of Arith15, T. Lang and J.-M. Muller, 2001, that the maximal
number of consecutive zeroes or ones after the round bit is n-1 for an
input of n bits. But we need a more precise lower bound. Assume x has
n bits, and 1/x is near a floating-point number y of n+1 bits. We can
write x = X*2^e, y = Y/2^f with X, Y integers of n and n+1 bits.
Thus X*Y^2^(e-f) is near from 1, i.e., X*Y is near from 2^(f-e).
Two cases can happen:
(i) either X*Y is exactly 2^(f-e), but this can happen only if X and Y
are themselves powers of two, i.e., x is a power of two;
(ii) or X*Y is at distance at least one from 2^(f-e), thus
|xy-1| >= 2^(e-f), or |y-1/x| >= 2^(e-f)/x = 2^(-f)/X >= 2^(-f-n).
Since ufp(y) = 2^(n-f) [ufp = unit in first place], this means
that the distance |y-1/x| >= 2^(-2n) ufp(y).
Now assuming |gamma(x)-1/x| <= 1, which is true for x <= 1,
if 2^(-2n) ufp(y) >= 2, the error is at most 2^(-2n-1) ufp(y),
and round(1/x) with precision >= 2n+2 gives the correct result.
If x < 2^E, then y > 2^(-E), thus ufp(y) > 2^(-E-1).
A sufficient condition is thus EXP(x) + 2 <= -2 MAX(PREC(x),PREC(Y)).
*/
if (MPFR_GET_EXP (x) + 2
<= -2 * (mpfr_exp_t) MAX(MPFR_PREC(x), MPFR_PREC(gamma)))
{
int sign = MPFR_SIGN (x); /* retrieve sign before possible override */
int special;
MPFR_BLOCK_DECL (flags);
MPFR_SAVE_EXPO_MARK (expo);
/* for overflow cases, see below; this needs to be done
before x possibly gets overridden. */
special =
MPFR_GET_EXP (x) == 1 - MPFR_EMAX_MAX &&
MPFR_IS_POS_SIGN (sign) &&
MPFR_IS_LIKE_RNDD (rnd_mode, sign) &&
mpfr_powerof2_raw (x);
MPFR_BLOCK (flags, inex = mpfr_ui_div (gamma, 1, x, rnd_mode));
if (inex == 0) /* x is a power of two */
{
/* return RND(1/x - euler) = RND(+/- 2^k - eps) with eps > 0 */
if (rnd_mode == MPFR_RNDN || MPFR_IS_LIKE_RNDU (rnd_mode, sign))
inex = 1;
else
{
mpfr_nextbelow (gamma);
inex = -1;
}
}
else if (MPFR_UNLIKELY (MPFR_OVERFLOW (flags)))
{
/* Overflow in the division 1/x. This is a real overflow, except
in RNDZ or RNDD when 1/x = 2^emax, i.e. x = 2^(-emax): due to
the "- euler", the rounded value in unbounded exponent range
is 0.111...11 * 2^emax (not an overflow). */
if (!special)
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, flags);
}
MPFR_SAVE_EXPO_FREE (expo);
/* Note: an overflow is possible with an infinite result;
in this case, the overflow flag will automatically be
restored by mpfr_check_range. */
return mpfr_check_range (gamma, inex, rnd_mode);
}
is_integer = mpfr_integer_p (x);
/* gamma(x) for x a negative integer gives NaN */
if (is_integer && MPFR_IS_NEG(x))
{
MPFR_SET_NAN (gamma);
MPFR_RET_NAN;
}
compared = mpfr_cmp_ui (x, 1);
if (compared == 0)
return mpfr_set_ui (gamma, 1, rnd_mode);
/* if x is an integer that fits into an unsigned long, use mpfr_fac_ui
if argument is not too large.
If precision is p, fac_ui costs O(u*p), whereas gamma costs O(p*M(p)),
so for u <= M(p), fac_ui should be faster.
We approximate here M(p) by p*log(p)^2, which is not a bad guess.
Warning: since the generic code does not handle exact cases,
we want all cases where gamma(x) is exact to be treated here.
*/
if (is_integer && mpfr_fits_ulong_p (x, MPFR_RNDN))
{
unsigned long int u;
mpfr_prec_t p = MPFR_PREC(gamma);
u = mpfr_get_ui (x, MPFR_RNDN);
if (u < 44787929UL && bits_fac (u - 1) <= p + (rnd_mode == MPFR_RNDN))
/* bits_fac: lower bound on the number of bits of m,
where gamma(x) = (u-1)! = m*2^e with m odd. */
return mpfr_fac_ui (gamma, u - 1, rnd_mode);
/* if bits_fac(...) > p (resp. p+1 for rounding to nearest),
then gamma(x) cannot be exact in precision p (resp. p+1).
FIXME: remove the test u < 44787929UL after changing bits_fac
to return a mpz_t or mpfr_t. */
}
MPFR_SAVE_EXPO_MARK (expo);
/* check for overflow: according to (6.1.37) in Abramowitz & Stegun,
gamma(x) >= exp(-x) * x^(x-1/2) * sqrt(2*Pi)
>= 2 * (x/e)^x / x for x >= 1 */
if (compared > 0)
{
mpfr_t yp;
mpfr_exp_t expxp;
MPFR_BLOCK_DECL (flags);
/* quick test for the default exponent range */
if (mpfr_get_emax () >= 1073741823UL && MPFR_GET_EXP(x) <= 25)
{
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_gamma_aux (gamma, x, rnd_mode);
}
/* 1/e rounded down to 53 bits */
#define EXPM1_STR "0.010111100010110101011000110110001011001110111100111"
mpfr_init2 (xp, 53);
mpfr_init2 (yp, 53);
mpfr_set_str_binary (xp, EXPM1_STR);
mpfr_mul (xp, x, xp, MPFR_RNDZ);
mpfr_sub_ui (yp, x, 2, MPFR_RNDZ);
mpfr_pow (xp, xp, yp, MPFR_RNDZ); /* (x/e)^(x-2) */
mpfr_set_str_binary (yp, EXPM1_STR);
mpfr_mul (xp, xp, yp, MPFR_RNDZ); /* x^(x-2) / e^(x-1) */
mpfr_mul (xp, xp, yp, MPFR_RNDZ); /* x^(x-2) / e^x */
mpfr_mul (xp, xp, x, MPFR_RNDZ); /* lower bound on x^(x-1) / e^x */
MPFR_BLOCK (flags, mpfr_mul_2ui (xp, xp, 1, MPFR_RNDZ));
expxp = MPFR_GET_EXP (xp);
mpfr_clear (xp);
mpfr_clear (yp);
MPFR_SAVE_EXPO_FREE (expo);
return MPFR_OVERFLOW (flags) || expxp > __gmpfr_emax ?
mpfr_overflow (gamma, rnd_mode, 1) :
mpfr_gamma_aux (gamma, x, rnd_mode);
}
/* now compared < 0 */
/* check for underflow: for x < 1,
gamma(x) = Pi*(x-1)/sin(Pi*(2-x))/gamma(2-x).
Since gamma(2-x) >= 2 * ((2-x)/e)^(2-x) / (2-x), we have
|gamma(x)| <= Pi*(1-x)*(2-x)/2/((2-x)/e)^(2-x) / |sin(Pi*(2-x))|
<= 12 * ((2-x)/e)^x / |sin(Pi*(2-x))|.
To avoid an underflow in ((2-x)/e)^x, we compute the logarithm.
*/
if (MPFR_IS_NEG(x))
{
int underflow = 0, sgn, ck;
mpfr_prec_t w;
mpfr_init2 (xp, 53);
mpfr_init2 (tmp, 53);
mpfr_init2 (tmp2, 53);
/* we want an upper bound for x * [log(2-x)-1].
since x < 0, we need a lower bound on log(2-x) */
mpfr_ui_sub (xp, 2, x, MPFR_RNDD);
mpfr_log (xp, xp, MPFR_RNDD);
mpfr_sub_ui (xp, xp, 1, MPFR_RNDD);
mpfr_mul (xp, xp, x, MPFR_RNDU);
/* we need an upper bound on 1/|sin(Pi*(2-x))|,
thus a lower bound on |sin(Pi*(2-x))|.
If 2-x is exact, then the error of Pi*(2-x) is (1+u)^2 with u = 2^(-p)
thus the error on sin(Pi*(2-x)) is less than 1/2ulp + 3Pi(2-x)u,
assuming u <= 1, thus <= u + 3Pi(2-x)u */
w = mpfr_gamma_2_minus_x_exact (x); /* 2-x is exact for prec >= w */
w += 17; /* to get tmp2 small enough */
mpfr_set_prec (tmp, w);
mpfr_set_prec (tmp2, w);
MPFR_DBGRES (ck = mpfr_ui_sub (tmp, 2, x, MPFR_RNDN));
MPFR_ASSERTD (ck == 0); /* tmp = 2-x exactly */
mpfr_const_pi (tmp2, MPFR_RNDN);
mpfr_mul (tmp2, tmp2, tmp, MPFR_RNDN); /* Pi*(2-x) */
mpfr_sin (tmp, tmp2, MPFR_RNDN); /* sin(Pi*(2-x)) */
sgn = mpfr_sgn (tmp);
mpfr_abs (tmp, tmp, MPFR_RNDN);
mpfr_mul_ui (tmp2, tmp2, 3, MPFR_RNDU); /* 3Pi(2-x) */
mpfr_add_ui (tmp2, tmp2, 1, MPFR_RNDU); /* 3Pi(2-x)+1 */
mpfr_div_2ui (tmp2, tmp2, mpfr_get_prec (tmp), MPFR_RNDU);
/* if tmp2<|tmp|, we get a lower bound */
if (mpfr_cmp (tmp2, tmp) < 0)
{
mpfr_sub (tmp, tmp, tmp2, MPFR_RNDZ); /* low bnd on |sin(Pi*(2-x))| */
mpfr_ui_div (tmp, 12, tmp, MPFR_RNDU); /* upper bound */
mpfr_log2 (tmp, tmp, MPFR_RNDU);
mpfr_add (xp, tmp, xp, MPFR_RNDU);
/* The assert below checks that expo.saved_emin - 2 always
fits in a long. FIXME if we want to allow mpfr_exp_t to
be a long long, for instance. */
MPFR_ASSERTN (MPFR_EMIN_MIN - 2 >= LONG_MIN);
underflow = mpfr_cmp_si (xp, expo.saved_emin - 2) <= 0;
}
mpfr_clear (xp);
mpfr_clear (tmp);
mpfr_clear (tmp2);
if (underflow) /* the sign is the opposite of that of sin(Pi*(2-x)) */
{
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_underflow (gamma, (rnd_mode == MPFR_RNDN) ? MPFR_RNDZ : rnd_mode, -sgn);
}
}
realprec = MPFR_PREC (gamma);
/* we want both 1-x and 2-x to be exact */
{
mpfr_prec_t w;
w = mpfr_gamma_1_minus_x_exact (x);
if (realprec < w)
realprec = w;
w = mpfr_gamma_2_minus_x_exact (x);
if (realprec < w)
realprec = w;
}
realprec = realprec + MPFR_INT_CEIL_LOG2 (realprec) + 20;
MPFR_ASSERTD(realprec >= 5);
MPFR_GROUP_INIT_4 (group, realprec + MPFR_INT_CEIL_LOG2 (realprec) + 20,
xp, tmp, tmp2, GammaTrial);
mpz_init (fact);
MPFR_ZIV_INIT (loop, realprec);
for (;;)
{
mpfr_exp_t err_g;
int ck;
MPFR_GROUP_REPREC_4 (group, realprec, xp, tmp, tmp2, GammaTrial);
/* reflection formula: gamma(x) = Pi*(x-1)/sin(Pi*(2-x))/gamma(2-x) */
ck = mpfr_ui_sub (xp, 2, x, MPFR_RNDN); /* 2-x, exact */
MPFR_ASSERTD(ck == 0); (void) ck; /* use ck to avoid a warning */
mpfr_gamma (tmp, xp, MPFR_RNDN); /* gamma(2-x), error (1+u) */
mpfr_const_pi (tmp2, MPFR_RNDN); /* Pi, error (1+u) */
mpfr_mul (GammaTrial, tmp2, xp, MPFR_RNDN); /* Pi*(2-x), error (1+u)^2 */
err_g = MPFR_GET_EXP(GammaTrial);
mpfr_sin (GammaTrial, GammaTrial, MPFR_RNDN); /* sin(Pi*(2-x)) */
/* If tmp is +Inf, we compute exp(lngamma(x)). */
if (mpfr_inf_p (tmp))
{
inex = mpfr_explgamma (gamma, x, &expo, tmp, tmp2, rnd_mode);
if (inex)
goto end;
else
goto ziv_next;
}
err_g = err_g + 1 - MPFR_GET_EXP(GammaTrial);
/* let g0 the true value of Pi*(2-x), g the computed value.
We have g = g0 + h with |h| <= |(1+u^2)-1|*g.
Thus sin(g) = sin(g0) + h' with |h'| <= |(1+u^2)-1|*g.
The relative error is thus bounded by |(1+u^2)-1|*g/sin(g)
<= |(1+u^2)-1|*2^err_g. <= 2.25*u*2^err_g for |u|<=1/4.
With the rounding error, this gives (0.5 + 2.25*2^err_g)*u. */
ck = mpfr_sub_ui (xp, x, 1, MPFR_RNDN); /* x-1, exact */
MPFR_ASSERTD(ck == 0); (void) ck; /* use ck to avoid a warning */
mpfr_mul (xp, tmp2, xp, MPFR_RNDN); /* Pi*(x-1), error (1+u)^2 */
mpfr_mul (GammaTrial, GammaTrial, tmp, MPFR_RNDN);
/* [1 + (0.5 + 2.25*2^err_g)*u]*(1+u)^2 = 1 + (2.5 + 2.25*2^err_g)*u
+ (0.5 + 2.25*2^err_g)*u*(2u+u^2) + u^2.
For err_g <= realprec-2, we have (0.5 + 2.25*2^err_g)*u <=
0.5*u + 2.25/4 <= 0.6875 and u^2 <= u/4, thus
(0.5 + 2.25*2^err_g)*u*(2u+u^2) + u^2 <= 0.6875*(2u+u/4) + u/4
<= 1.8*u, thus the rel. error is bounded by (4.5 + 2.25*2^err_g)*u. */
mpfr_div (GammaTrial, xp, GammaTrial, MPFR_RNDN);
/* the error is of the form (1+u)^3/[1 + (4.5 + 2.25*2^err_g)*u].
For realprec >= 5 and err_g <= realprec-2, [(4.5 + 2.25*2^err_g)*u]^2
<= 0.71, and for |y|<=0.71, 1/(1-y) can be written 1+a*y with a<=4.
(1+u)^3 * (1+4*(4.5 + 2.25*2^err_g)*u)
= 1 + (21 + 9*2^err_g)*u + (57+27*2^err_g)*u^2 + (55+27*2^err_g)*u^3
+ (18+9*2^err_g)*u^4
<= 1 + (21 + 9*2^err_g)*u + (57+27*2^err_g)*u^2 + (56+28*2^err_g)*u^3
<= 1 + (21 + 9*2^err_g)*u + (59+28*2^err_g)*u^2
<= 1 + (23 + 10*2^err_g)*u.
The final error is thus bounded by (23 + 10*2^err_g) ulps,
which is <= 2^6 for err_g<=2, and <= 2^(err_g+4) for err_g >= 2. */
err_g = (err_g <= 2) ? 6 : err_g + 4;
if (MPFR_LIKELY (MPFR_CAN_ROUND (GammaTrial, realprec - err_g,
MPFR_PREC(gamma), rnd_mode)))
break;
ziv_next:
MPFR_ZIV_NEXT (loop, realprec);
}
end:
MPFR_ZIV_FREE (loop);
if (inex == 0)
inex = mpfr_set (gamma, GammaTrial, rnd_mode);
MPFR_GROUP_CLEAR (group);
mpz_clear (fact);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (gamma, inex, rnd_mode);
}
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));
}
int
mpfr_exp2 (mpfr_ptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
int inexact;
long xint;
mpfr_t xfrac;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg rnd=%d", mpfr_get_prec(x), mpfr_log_prec, x, rnd_mode),
("y[%Pu]=%.*Rg inexact=%d", mpfr_get_prec(y), mpfr_log_prec, y,
inexact));
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x)))
{
if (MPFR_IS_NAN (x))
{
MPFR_SET_NAN (y);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (x))
{
if (MPFR_IS_POS (x))
MPFR_SET_INF (y);
else
MPFR_SET_ZERO (y);
MPFR_SET_POS (y);
MPFR_RET (0);
}
else /* 2^0 = 1 */
{
MPFR_ASSERTD (MPFR_IS_ZERO(x));
return mpfr_set_ui (y, 1, rnd_mode);
}
}
/* since the smallest representable non-zero float is 1/2*2^__gmpfr_emin,
if x < __gmpfr_emin - 1, the result is either 1/2*2^__gmpfr_emin or 0 */
MPFR_ASSERTN (MPFR_EMIN_MIN >= LONG_MIN + 2);
if (MPFR_UNLIKELY (mpfr_cmp_si (x, __gmpfr_emin - 1) < 0))
{
mpfr_rnd_t rnd2 = rnd_mode;
/* in round to nearest mode, round to zero when x <= __gmpfr_emin-2 */
if (rnd_mode == MPFR_RNDN &&
mpfr_cmp_si_2exp (x, __gmpfr_emin - 2, 0) <= 0)
rnd2 = MPFR_RNDZ;
return mpfr_underflow (y, rnd2, 1);
}
MPFR_ASSERTN (MPFR_EMAX_MAX <= LONG_MAX);
if (MPFR_UNLIKELY (mpfr_cmp_si (x, __gmpfr_emax) >= 0))
return mpfr_overflow (y, rnd_mode, 1);
/* We now know that emin - 1 <= x < emax. */
MPFR_SAVE_EXPO_MARK (expo);
/* 2^x = 1 + x*log(2) + O(x^2) for x near zero, and for |x| <= 1 we have
|2^x - 1| <= x < 2^EXP(x). If x > 0 we must round away from 0 (dir=1);
if x < 0 we must round toward 0 (dir=0). */
MPFR_SMALL_INPUT_AFTER_SAVE_EXPO (y, __gmpfr_one, - MPFR_GET_EXP (x), 0,
MPFR_IS_POS (x), rnd_mode, expo, {});
xint = mpfr_get_si (x, MPFR_RNDZ);
mpfr_init2 (xfrac, MPFR_PREC (x));
mpfr_sub_si (xfrac, x, xint, MPFR_RNDN); /* exact */
if (MPFR_IS_ZERO (xfrac))
{
mpfr_set_ui (y, 1, MPFR_RNDN);
inexact = 0;
}
else
{
/* Declaration of the intermediary variable */
mpfr_t t;
/* Declaration of the size variable */
mpfr_prec_t Ny = MPFR_PREC(y); /* target precision */
mpfr_prec_t Nt; /* working precision */
mpfr_exp_t err; /* error */
MPFR_ZIV_DECL (loop);
/* compute the precision of intermediary variable */
/* the optimal number of bits : see algorithms.tex */
Nt = Ny + 5 + MPFR_INT_CEIL_LOG2 (Ny);
/* initialize of intermediary variable */
mpfr_init2 (t, Nt);
/* First computation */
MPFR_ZIV_INIT (loop, Nt);
for (;;)
{
/* compute exp(x*ln(2))*/
mpfr_const_log2 (t, MPFR_RNDU); /* ln(2) */
mpfr_mul (t, xfrac, t, MPFR_RNDU); /* xfrac * ln(2) */
err = Nt - (MPFR_GET_EXP (t) + 2); /* Estimate of the error */
mpfr_exp (t, t, MPFR_RNDN); /* exp(xfrac * ln(2)) */
if (MPFR_LIKELY (MPFR_CAN_ROUND (t, err, Ny, rnd_mode)))
break;
/* Actualisation of the precision */
MPFR_ZIV_NEXT (loop, Nt);
mpfr_set_prec (t, Nt);
}
MPFR_ZIV_FREE (loop);
inexact = mpfr_set (y, t, rnd_mode);
mpfr_clear (t);
}
mpfr_clear (xfrac);
MPFR_CLEAR_FLAGS ();
mpfr_mul_2si (y, y, xint, MPFR_RNDN); /* exact or overflow */
/* Note: We can have an overflow only when t was rounded up to 2. */
MPFR_ASSERTD (MPFR_IS_PURE_FP (y) || inexact > 0);
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, __gmpfr_flags);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (y, inexact, rnd_mode);
}
