/* 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_mul_2si (mpfr_ptr y, mpfr_srcptr x, long int n, mpfr_rnd_t rnd_mode)
{
int inexact;
MPFR_LOG_FUNC (("x[%#R]=%R n=%ld rnd=%d", x, x, n, rnd_mode),
("y[%#R]=%R inexact=%d", y, y, inexact));
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x)))
return mpfr_set (y, x, rnd_mode);
else
{
mpfr_exp_t exp = MPFR_GET_EXP (x);
MPFR_SETRAW (inexact, y, x, exp, rnd_mode);
if (MPFR_UNLIKELY( n > 0 && (__gmpfr_emax < MPFR_EMIN_MIN + n ||
exp > __gmpfr_emax - n)))
return mpfr_overflow (y, rnd_mode, MPFR_SIGN(y));
else if (MPFR_UNLIKELY(n < 0 && (__gmpfr_emin > MPFR_EMAX_MAX + n ||
exp < __gmpfr_emin - n)))
{
if (rnd_mode == MPFR_RNDN &&
(__gmpfr_emin > MPFR_EMAX_MAX + (n + 1) ||
exp < __gmpfr_emin - (n + 1) ||
(inexact >= 0 && mpfr_powerof2_raw (y))))
rnd_mode = MPFR_RNDZ;
return mpfr_underflow (y, rnd_mode, MPFR_SIGN(y));
}
MPFR_SET_EXP (y, exp + n);
}
MPFR_RET (inexact);
}
int
mpfr_div_2si (mpfr_ptr y, mpfr_srcptr x, long int n, mp_rnd_t rnd_mode)
{
int inexact;
MPFR_LOG_FUNC (("x[%#R]=%R n=%ld rnd=%d", x, x, n, rnd_mode),
("y[%#R]=%R inexact=%d", y, y, inexact));
inexact = MPFR_UNLIKELY(y != x) ? mpfr_set (y, x, rnd_mode) : 0;
if (MPFR_LIKELY( MPFR_IS_PURE_FP(y) ))
{
mp_exp_t exp = MPFR_GET_EXP (y);
if (MPFR_UNLIKELY( n > 0 && (__gmpfr_emin > MPFR_EMAX_MAX - n ||
exp < __gmpfr_emin + n)) )
{
if (rnd_mode == GMP_RNDN &&
(__gmpfr_emin > MPFR_EMAX_MAX - (n - 1) ||
exp < __gmpfr_emin + (n - 1) ||
(inexact >= 0 && mpfr_powerof2_raw (y))))
rnd_mode = GMP_RNDZ;
return mpfr_underflow (y, rnd_mode, MPFR_SIGN(y));
}
if (MPFR_UNLIKELY(n < 0 && (__gmpfr_emax < MPFR_EMIN_MIN - n ||
exp > __gmpfr_emax + n)) )
return mpfr_overflow (y, rnd_mode, MPFR_SIGN(y));
MPFR_SET_EXP (y, exp - n);
}
return inexact;
}
int
mpfr_check_range (mpfr_ptr x, int t, mp_rnd_t rnd_mode)
{
if (MPFR_LIKELY( MPFR_IS_PURE_FP(x)) )
{ /* x is a non-zero FP */
mp_exp_t exp = MPFR_EXP (x); /* Do not use MPFR_GET_EXP */
if (MPFR_UNLIKELY( exp < __gmpfr_emin) )
{
/* The following test is necessary because in the rounding to the
* nearest mode, mpfr_underflow always rounds away from 0. In
* this rounding mode, we need to round to 0 if:
* _ |x| < 2^(emin-2), or
* _ |x| = 2^(emin-2) and the absolute value of the exact
* result is <= 2^(emin-2).
*/
if (rnd_mode == GMP_RNDN &&
(exp + 1 < __gmpfr_emin ||
(mpfr_powerof2_raw(x) &&
(MPFR_IS_NEG(x) ? t <= 0 : t >= 0))))
rnd_mode = GMP_RNDZ;
return mpfr_underflow(x, rnd_mode, MPFR_SIGN(x));
}
if (MPFR_UNLIKELY( exp > __gmpfr_emax) )
return mpfr_overflow(x, rnd_mode, MPFR_SIGN(x));
}
MPFR_RET (t); /* propagate inexact ternary value, unlike most functions */
}
int
mpfr_check_range (mpfr_ptr x, int t, mpfr_rnd_t rnd_mode)
{
if (MPFR_LIKELY( MPFR_IS_PURE_FP(x)) )
{ /* x is a non-zero FP */
mpfr_exp_t exp = MPFR_EXP (x); /* Do not use MPFR_GET_EXP */
/* Even if it is unlikely that exp is lower than emin,
this function is called by MPFR functions only if it is
already the case (or if exp is greater than emax) */
if (exp < __gmpfr_emin)
{
/* The following test is necessary because in the rounding to the
* nearest mode, mpfr_underflow always rounds away from 0. In
* this rounding mode, we need to round to 0 if:
* _ |x| < 2^(emin-2), or
* _ |x| = 2^(emin-2) and the absolute value of the exact
* result is <= 2^(emin-2).
*/
if (rnd_mode == MPFR_RNDN &&
(exp + 1 < __gmpfr_emin ||
(mpfr_powerof2_raw(x) &&
(MPFR_IS_NEG(x) ? t <= 0 : t >= 0))))
rnd_mode = MPFR_RNDZ;
return mpfr_underflow(x, rnd_mode, MPFR_SIGN(x));
}
if (exp > __gmpfr_emax)
return mpfr_overflow (x, rnd_mode, MPFR_SIGN(x));
}
else if (MPFR_UNLIKELY (t != 0 && MPFR_IS_INF (x)))
{
/* We need to do the following because most MPFR functions are
* implemented in the following way:
* Ziv's loop:
* | Compute an approximation to the result and an error bound.
* | Possible underflow/overflow detection -> return.
* | If can_round, break (exit the loop).
* | Otherwise, increase the working precision and loop.
* Round the approximation in the target precision. <== See below
* Restore the flags (that could have been set due to underflows
* or overflows during the internal computations).
* Execute: return mpfr_check_range (...).
* The problem is that an overflow could be generated when rounding the
* approximation (in general, such an overflow could not be detected
* earlier), and the overflow flag is lost when the flags are restored.
* This can occur only when the rounding yields an exponent change
* and the new exponent is larger than the maximum exponent, so that
* an infinity is necessarily obtained.
* So, the simplest solution is to detect this overflow case here in
* mpfr_check_range, which is easy to do since the rounded result is
* necessarily an inexact infinity.
*/
__gmpfr_flags |= MPFR_FLAGS_OVERFLOW;
}
MPFR_RET (t); /* propagate inexact ternary value, unlike most functions */
}
int
mpfr_mul_2ui (mpfr_ptr y, mpfr_srcptr x, unsigned long int n, mpfr_rnd_t rnd_mode)
{
int inexact;
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg n=%lu rnd=%d",
mpfr_get_prec (x), mpfr_log_prec, x, n, rnd_mode),
("y[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (y), mpfr_log_prec, y, inexact));
inexact = MPFR_UNLIKELY(y != x) ? mpfr_set (y, x, rnd_mode) : 0;
if (MPFR_LIKELY( MPFR_IS_PURE_FP(y)) )
{
/* n will have to be casted to long to make sure that the addition
and subtraction below (for overflow detection) are signed */
while (MPFR_UNLIKELY(n > LONG_MAX))
{
int inex2;
n -= LONG_MAX;
inex2 = mpfr_mul_2ui(y, y, LONG_MAX, rnd_mode);
if (inex2)
return inex2; /* overflow */
}
/* MPFR_EMIN_MIN + (long) n is signed and doesn't lead to an overflow;
the first test useful so that the real test can't lead to an
overflow. */
{
mpfr_exp_t exp = MPFR_GET_EXP (y);
if (MPFR_UNLIKELY( __gmpfr_emax < MPFR_EMIN_MIN + (long) n ||
exp > __gmpfr_emax - (long) n))
return mpfr_overflow (y, rnd_mode, MPFR_SIGN(y));
MPFR_SET_EXP (y, exp + (long) n);
}
}
return inexact;
}
static void
test_set_overflow (void)
{
mpfr_t x, inf, max;
mpfr_ptr r[MPFR_RND_MAX];
int t[MPFR_RND_MAX] = { 1, -1, 1, -1, 1 }; /* RNDN, RNDZ, RNDU, RNDD, RNDA */
int i;
int s;
mpfr_inits2 (32, x, inf, max, (mpfr_ptr) 0);
mpfr_set_inf (inf, 1);
mpfr_set_inf (max, 1);
mpfr_nextbelow (max);
r[0] = r[2] = r[4] = inf; /* RNDN, RNDU, RNDA */
r[1] = r[3] = max; /* RNDZ, RNDD */
for (s = 1; s > 0; s = -1)
{
for (i = 0; i < MPFR_RND_MAX ; i++)
{
int j;
int inex;
j = s < 0 && i > 1 ? 5 - i : i;
inex = mpfr_overflow (x, (mpfr_rnd_t) i, s);
if (mpfr_cmp (x, r[j]) || inex * t[j] <= 0)
{
printf ("Error in test_set_overflow, sign = %d,"
" rnd_mode = %s\n", s, mpfr_print_rnd_mode ((mpfr_rnd_t) i));
printf ("Got\n");
mpfr_out_str (stdout, 2, 0, x, MPFR_RNDN);
printf (", inex = %d\ninstead of\n", inex);
mpfr_out_str (stdout, 2, 0, r[j], MPFR_RNDN);
printf (", inex = %d\n", t[j]);
exit (1);
}
}
mpfr_neg (inf, inf, MPFR_RNDN);
mpfr_neg (max, max, MPFR_RNDN);
}
mpfr_clears (x, inf, max, (mpfr_ptr) 0);
}
int
mpfr_rint_floor (mpfr_ptr r, mpfr_srcptr u, mpfr_rnd_t rnd_mode)
{
if (MPFR_UNLIKELY( MPFR_IS_SINGULAR(u) ) || mpfr_integer_p (u))
return mpfr_set (r, u, rnd_mode);
else
{
mpfr_t tmp;
int inex;
unsigned int saved_flags = __gmpfr_flags;
MPFR_BLOCK_DECL (flags);
mpfr_init2 (tmp, MPFR_PREC (u));
/* floor(u) is representable in tmp unless an overflow occurs */
MPFR_BLOCK (flags, mpfr_floor (tmp, u));
__gmpfr_flags = saved_flags;
inex = (MPFR_OVERFLOW (flags)
? mpfr_overflow (r, rnd_mode, MPFR_SIGN_NEG)
: mpfr_set (r, tmp, rnd_mode));
mpfr_clear (tmp);
return inex;
}
}
MPFR_HOT_FUNCTION_ATTR int
mpfr_div_2si (mpfr_ptr y, mpfr_srcptr x, long int n, mpfr_rnd_t rnd_mode)
{
int inexact;
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg n=%ld rnd=%d",
mpfr_get_prec(x), mpfr_log_prec, x, n, rnd_mode),
("y[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec(y), mpfr_log_prec, y, inexact));
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x)))
return mpfr_set (y, x, rnd_mode);
else
{
mpfr_exp_t exp = MPFR_GET_EXP (x);
MPFR_SETRAW (inexact, y, x, exp, rnd_mode);
if (MPFR_UNLIKELY( n > 0 && (__gmpfr_emin > MPFR_EMAX_MAX - n ||
exp < __gmpfr_emin + n)) )
{
if (rnd_mode == MPFR_RNDN &&
(__gmpfr_emin > MPFR_EMAX_MAX - (n - 1) ||
exp < __gmpfr_emin + (n - 1) ||
((MPFR_IS_NEG (y) ? inexact <= 0 : inexact >= 0) &&
mpfr_powerof2_raw (y))))
rnd_mode = MPFR_RNDZ;
return mpfr_underflow (y, rnd_mode, MPFR_SIGN(y));
}
else if (MPFR_UNLIKELY(n <= 0 && (__gmpfr_emax < MPFR_EMIN_MIN - n ||
exp > __gmpfr_emax + n)) )
return mpfr_overflow (y, rnd_mode, MPFR_SIGN(y));
MPFR_SET_EXP (y, exp - n);
}
MPFR_RET (inexact);
}
int
mpfr_rint_floor (mpfr_ptr r, mpfr_srcptr u, mpfr_rnd_t rnd_mode)
{
if (MPFR_UNLIKELY( MPFR_IS_SINGULAR(u) ) || mpfr_integer_p (u))
return mpfr_set (r, u, rnd_mode);
else
{
mpfr_t tmp;
int inex;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_BLOCK_DECL (flags);
MPFR_SAVE_EXPO_MARK (expo);
mpfr_init2 (tmp, MPFR_PREC (u));
/* floor(u) is representable in tmp unless an overflow occurs */
MPFR_BLOCK (flags, mpfr_floor (tmp, u));
inex = (MPFR_OVERFLOW (flags)
? mpfr_overflow (r, rnd_mode, MPFR_SIGN_NEG)
: mpfr_set (r, tmp, rnd_mode));
mpfr_clear (tmp);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (r, inex, rnd_mode);
}
}
int
mpfr_exp_3 (mpfr_ptr y, mpfr_srcptr x, mp_rnd_t rnd_mode)
{
mpfr_t t, x_copy, tmp;
mpz_t uk;
mp_exp_t ttt, shift_x;
unsigned long twopoweri;
mpz_t *P;
mp_prec_t *mult;
int i, k, loop;
int prec_x;
mp_prec_t realprec, Prec;
int iter;
int inexact = 0;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_ZIV_DECL (ziv_loop);
MPFR_LOG_FUNC (("x[%#R]=%R rnd=%d", x, x, rnd_mode),
("y[%#R]=%R inexact=%d", y, y, inexact));
MPFR_SAVE_EXPO_MARK (expo);
/* decompose x */
/* we first write x = 1.xxxxxxxxxxxxx
----- k bits -- */
prec_x = MPFR_INT_CEIL_LOG2 (MPFR_PREC (x)) - MPFR_LOG2_BITS_PER_MP_LIMB;
if (prec_x < 0)
prec_x = 0;
ttt = MPFR_GET_EXP (x);
mpfr_init2 (x_copy, MPFR_PREC(x));
mpfr_set (x_copy, x, GMP_RNDD);
/* we shift to get a number less than 1 */
if (ttt > 0)
{
shift_x = ttt;
mpfr_div_2ui (x_copy, x, ttt, GMP_RNDN);
ttt = MPFR_GET_EXP (x_copy);
}
else
shift_x = 0;
MPFR_ASSERTD (ttt <= 0);
/* Init prec and vars */
realprec = MPFR_PREC (y) + MPFR_INT_CEIL_LOG2 (prec_x + MPFR_PREC (y));
Prec = realprec + shift + 2 + shift_x;
mpfr_init2 (t, Prec);
mpfr_init2 (tmp, Prec);
mpz_init (uk);
/* Main loop */
MPFR_ZIV_INIT (ziv_loop, realprec);
for (;;)
{
int scaled = 0;
MPFR_BLOCK_DECL (flags);
k = MPFR_INT_CEIL_LOG2 (Prec) - MPFR_LOG2_BITS_PER_MP_LIMB;
/* now we have to extract */
twopoweri = BITS_PER_MP_LIMB;
/* Allocate tables */
P = (mpz_t*) (*__gmp_allocate_func) (3*(k+2)*sizeof(mpz_t));
for (i = 0; i < 3*(k+2); i++)
mpz_init (P[i]);
mult = (mp_prec_t*) (*__gmp_allocate_func) (2*(k+2)*sizeof(mp_prec_t));
/* Particular case for i==0 */
mpfr_extract (uk, x_copy, 0);
MPFR_ASSERTD (mpz_cmp_ui (uk, 0) != 0);
mpfr_exp_rational (tmp, uk, shift + twopoweri - ttt, k + 1, P, mult);
for (loop = 0; loop < shift; loop++)
mpfr_sqr (tmp, tmp, GMP_RNDD);
twopoweri *= 2;
/* General case */
iter = (k <= prec_x) ? k : prec_x;
for (i = 1; i <= iter; i++)
{
mpfr_extract (uk, x_copy, i);
if (MPFR_LIKELY (mpz_cmp_ui (uk, 0) != 0))
{
mpfr_exp_rational (t, uk, twopoweri - ttt, k - i + 1, P, mult);
mpfr_mul (tmp, tmp, t, GMP_RNDD);
}
MPFR_ASSERTN (twopoweri <= LONG_MAX/2);
twopoweri *=2;
}
/* Clear tables */
for (i = 0; i < 3*(k+2); i++)
mpz_clear (P[i]);
(*__gmp_free_func) (P, 3*(k+2)*sizeof(mpz_t));
(*__gmp_free_func) (mult, 2*(k+2)*sizeof(mp_prec_t));
if (shift_x > 0)
{
MPFR_BLOCK (flags, {
for (loop = 0; loop < shift_x - 1; loop++)
mpfr_sqr (tmp, tmp, GMP_RNDD);
mpfr_sqr (t, tmp, GMP_RNDD);
} );
if (MPFR_UNLIKELY (MPFR_OVERFLOW (flags)))
{
/* tmp <= exact result, so that it is a real overflow. */
inexact = mpfr_overflow (y, rnd_mode, 1);
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_OVERFLOW);
break;
}
if (MPFR_UNLIKELY (MPFR_UNDERFLOW (flags)))
{
/* This may be a spurious underflow. So, let's scale
the result. */
mpfr_mul_2ui (tmp, tmp, 1, GMP_RNDD); /* no overflow, exact */
mpfr_sqr (t, tmp, GMP_RNDD);
if (MPFR_IS_ZERO (t))
{
/* approximate result < 2^(emin - 3), thus
exact result < 2^(emin - 2). */
inexact = mpfr_underflow (y, (rnd_mode == GMP_RNDN) ?
GMP_RNDZ : rnd_mode, 1);
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_UNDERFLOW);
break;
}
scaled = 1;
}
}
/* compute sign(b) * (|b| + |c|)
Returns 0 iff result is exact,
a negative value when the result is less than the exact value,
a positive value otherwise. */
int
mpfr_add1sp (mpfr_ptr a, mpfr_srcptr b, mpfr_srcptr c, mpfr_rnd_t rnd_mode)
{
mpfr_uexp_t d;
mpfr_prec_t p;
unsigned int sh;
mp_size_t n;
mp_limb_t *ap, *cp;
mpfr_exp_t bx;
mp_limb_t limb;
int inexact;
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));
MPFR_ASSERTD(MPFR_GET_EXP(b) >= MPFR_GET_EXP(c));
/* Read prec and num of limbs */
p = MPFR_PREC(b);
n = MPFR_PREC2LIMBS (p);
MPFR_UNSIGNED_MINUS_MODULO(sh, p);
bx = MPFR_GET_EXP(b);
d = (mpfr_uexp_t) (bx - MPFR_GET_EXP(c));
DEBUG (printf ("New add1sp with diff=%lu\n", (unsigned long) d));
if (MPFR_UNLIKELY(d == 0))
{
/* d==0 */
DEBUG( mpfr_print_mant_binary("C= ", MPFR_MANT(c), p) );
DEBUG( mpfr_print_mant_binary("B= ", MPFR_MANT(b), p) );
bx++; /* exp + 1 */
ap = MPFR_MANT(a);
limb = mpn_add_n(ap, MPFR_MANT(b), MPFR_MANT(c), n);
DEBUG( mpfr_print_mant_binary("A= ", ap, p) );
MPFR_ASSERTD(limb != 0); /* There must be a carry */
limb = ap[0]; /* Get LSB (In fact, LSW) */
mpn_rshift(ap, ap, n, 1); /* Shift mantissa A */
ap[n-1] |= MPFR_LIMB_HIGHBIT; /* Set MSB */
ap[0] &= ~MPFR_LIMB_MASK(sh); /* Clear LSB bit */
if (MPFR_LIKELY((limb&(MPFR_LIMB_ONE<<sh)) == 0)) /* Check exact case */
{ inexact = 0; goto set_exponent; }
/* Zero: Truncate
Nearest: Even Rule => truncate or add 1
Away: Add 1 */
if (MPFR_LIKELY(rnd_mode==MPFR_RNDN))
{
if (MPFR_LIKELY((ap[0]&(MPFR_LIMB_ONE<<sh))==0))
{ inexact = -1; goto set_exponent; }
else
goto add_one_ulp;
}
MPFR_UPDATE_RND_MODE(rnd_mode, MPFR_IS_NEG(b));
if (rnd_mode==MPFR_RNDZ)
{ inexact = -1; goto set_exponent; }
else
goto add_one_ulp;
}
else if (MPFR_UNLIKELY (d >= p))
{
if (MPFR_LIKELY (d > p))
{
/* d > p : Copy B in A */
/* Away: Add 1
Nearest: Trunc
Zero: Trunc */
if (MPFR_LIKELY (rnd_mode==MPFR_RNDN
|| MPFR_IS_LIKE_RNDZ (rnd_mode, MPFR_IS_NEG (b))))
{
copy_set_exponent:
ap = MPFR_MANT (a);
MPN_COPY (ap, MPFR_MANT(b), n);
inexact = -1;
goto set_exponent;
}
else
{
copy_add_one_ulp:
ap = MPFR_MANT(a);
MPN_COPY (ap, MPFR_MANT(b), n);
goto add_one_ulp;
}
}
else
{
/* d==p : Copy B in A */
/* Away: Add 1
Nearest: Even Rule if C is a power of 2, else Add 1
Zero: Trunc */
if (MPFR_LIKELY(rnd_mode==MPFR_RNDN))
{
/* Check if C was a power of 2 */
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);
if (MPFR_UNLIKELY(k<0))
/* Power of 2: Even rule */
if ((MPFR_MANT (b)[0]&(MPFR_LIMB_ONE<<sh))==0)
goto copy_set_exponent;
}
/* Not a Power of 2 */
goto copy_add_one_ulp;
}
else if (MPFR_IS_LIKE_RNDZ (rnd_mode, MPFR_IS_NEG (b)))
goto copy_set_exponent;
else
goto copy_add_one_ulp;
}
}
else
{
mp_limb_t mask;
mp_limb_t bcp, bcp1; /* Cp and C'p+1 */
/* General case: 1 <= d < p */
cp = MPFR_TMP_LIMBS_ALLOC (n);
/* Shift c in temporary allocated place */
{
mpfr_uexp_t dm;
mp_size_t m;
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 >=1 and m == 0: just shift */
MPFR_ASSERTD(dm >= 1);
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 > 0))
{
/* Try to compute them from C' rather than C */
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 /* sh == 0 */
{
/* 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
{
do {
kx--;
} while (kx>=0 && tp[kx]==0);
bcp1 = (kx>=0);
}
}
DEBUG (printf("sh=%u Cp=%lu C'p+1=%lu\n", sh,
(unsigned long) bcp, (unsigned long) bcp1));
/* Clean shifted C' */
mask = ~MPFR_LIMB_MASK(sh);
cp[0] &= mask;
/* Add the mantissa c from b in a */
ap = MPFR_MANT(a);
limb = mpn_add_n (ap, MPFR_MANT(b), cp, n);
DEBUG( mpfr_print_mant_binary("Add= ", ap, p) );
/* Check for overflow */
if (MPFR_UNLIKELY (limb))
{
limb = ap[0] & (MPFR_LIMB_ONE<<sh); /* Get LSB */
mpn_rshift (ap, ap, n, 1); /* Shift mantissa*/
bx++; /* Fix exponent */
ap[n-1] |= MPFR_LIMB_HIGHBIT; /* Set MSB */
ap[0] &= mask; /* Clear LSB bit */
bcp1 |= bcp; /* Recompute C'p+1 */
bcp = limb; /* Recompute Cp */
DEBUG (printf ("(Overflow) Cp=%lu C'p+1=%lu\n",
(unsigned long) bcp, (unsigned long) bcp1));
DEBUG (mpfr_print_mant_binary ("Add= ", ap, p));
}
/* Round:
Zero: Truncate but could be exact.
Away: Add 1 if Cp or C'p+1 !=0
Nearest: Truncate but could be exact if Cp==0
Add 1 if C'p+1 !=0,
Even rule else */
if (MPFR_LIKELY(rnd_mode == MPFR_RNDN))
{
if (MPFR_LIKELY(bcp == 0))
{ inexact = MPFR_LIKELY(bcp1) ? -1 : 0; goto set_exponent; }
else if (MPFR_UNLIKELY(bcp1==0) && (ap[0]&(MPFR_LIMB_ONE<<sh))==0)
{ inexact = -1; goto set_exponent; }
else
goto add_one_ulp;
}
MPFR_UPDATE_RND_MODE(rnd_mode, MPFR_IS_NEG(b));
if (rnd_mode == MPFR_RNDZ)
{
inexact = MPFR_LIKELY(bcp || bcp1) ? -1 : 0;
goto set_exponent;
}
else
{
if (MPFR_UNLIKELY(bcp==0 && bcp1==0))
{ inexact = 0; goto set_exponent; }
else
goto add_one_ulp;
}
}
MPFR_ASSERTN(0);
add_one_ulp:
/* add one unit in last place to a */
DEBUG( printf("AddOneUlp\n") );
if (MPFR_UNLIKELY( mpn_add_1(ap, ap, n, MPFR_LIMB_ONE<<sh) ))
{
/* Case 100000x0 = 0x1111x1 + 1*/
DEBUG( printf("Pow of 2\n") );
bx++;
ap[n-1] = MPFR_LIMB_HIGHBIT;
}
inexact = 1;
set_exponent:
if (MPFR_UNLIKELY(bx > __gmpfr_emax)) /* Check for overflow */
{
DEBUG( printf("Overflow\n") );
MPFR_TMP_FREE(marker);
MPFR_SET_SAME_SIGN(a,b);
return mpfr_overflow(a, rnd_mode, MPFR_SIGN(a));
}
MPFR_SET_EXP (a, bx);
MPFR_SET_SAME_SIGN(a,b);
MPFR_TMP_FREE(marker);
MPFR_RET (inexact * MPFR_INT_SIGN (a));
}
int
mpfr_rint (mpfr_ptr r, mpfr_srcptr u, mpfr_rnd_t rnd_mode)
{
int sign;
int rnd_away;
mp_exp_t exp;
if (MPFR_UNLIKELY( MPFR_IS_SINGULAR(u) ))
{
if (MPFR_IS_NAN(u))
{
MPFR_SET_NAN(r);
MPFR_RET_NAN;
}
MPFR_SET_SAME_SIGN(r, u);
if (MPFR_IS_INF(u))
{
MPFR_SET_INF(r);
MPFR_RET(0); /* infinity is exact */
}
else /* now u is zero */
{
MPFR_ASSERTD(MPFR_IS_ZERO(u));
MPFR_SET_ZERO(r);
MPFR_RET(0); /* zero is exact */
}
}
MPFR_SET_SAME_SIGN (r, u); /* Does nothing if r==u */
sign = MPFR_INT_SIGN (u);
exp = MPFR_GET_EXP (u);
rnd_away =
rnd_mode == GMP_RNDD ? sign < 0 :
rnd_mode == GMP_RNDU ? sign > 0 :
rnd_mode == GMP_RNDZ ? 0 : -1;
/* rnd_away:
1 if round away from zero,
0 if round to zero,
-1 if not decided yet.
*/
if (MPFR_UNLIKELY (exp <= 0)) /* 0 < |u| < 1 ==> round |u| to 0 or 1 */
{
/* Note: in the GMP_RNDN mode, 0.5 must be rounded to 0. */
if (rnd_away != 0 &&
(rnd_away > 0 ||
(exp == 0 && (rnd_mode == GMP_RNDNA ||
!mpfr_powerof2_raw (u)))))
{
mp_limb_t *rp;
mp_size_t rm;
rp = MPFR_MANT(r);
rm = (MPFR_PREC(r) - 1) / BITS_PER_MP_LIMB;
rp[rm] = MPFR_LIMB_HIGHBIT;
MPN_ZERO(rp, rm);
MPFR_SET_EXP (r, 1); /* |r| = 1 */
MPFR_RET(sign > 0 ? 2 : -2);
}
else
{
MPFR_SET_ZERO(r); /* r = 0 */
MPFR_RET(sign > 0 ? -2 : 2);
}
}
else /* exp > 0, |u| >= 1 */
{
mp_limb_t *up, *rp;
mp_size_t un, rn, ui;
int sh, idiff;
int uflags;
/*
* uflags will contain:
* _ 0 if u is an integer representable in r,
* _ 1 if u is an integer not representable in r,
* _ 2 if u is not an integer.
*/
up = MPFR_MANT(u);
rp = MPFR_MANT(r);
un = MPFR_LIMB_SIZE(u);
rn = MPFR_LIMB_SIZE(r);
MPFR_UNSIGNED_MINUS_MODULO (sh, MPFR_PREC (r));
MPFR_SET_EXP (r, exp); /* Does nothing if r==u */
if ((exp - 1) / BITS_PER_MP_LIMB >= un)
{
ui = un;
idiff = 0;
uflags = 0; /* u is an integer, representable or not in r */
}
else
{
mp_size_t uj;
ui = (exp - 1) / BITS_PER_MP_LIMB + 1; /* #limbs of the int part */
MPFR_ASSERTD (un >= ui);
uj = un - ui; /* lowest limb of the integer part */
idiff = exp % BITS_PER_MP_LIMB; /* #int-part bits in up[uj] or 0 */
uflags = idiff == 0 || (up[uj] << idiff) == 0 ? 0 : 2;
if (uflags == 0)
while (uj > 0)
if (up[--uj] != 0)
{
uflags = 2;
break;
}
}
if (ui > rn)
{
/* More limbs in the integer part of u than in r.
Just round u with the precision of r. */
MPFR_ASSERTD (rp != up && un > rn);
MPN_COPY (rp, up + (un - rn), rn); /* r != u */
if (rnd_away < 0)
{
/* This is a rounding to nearest mode (GMP_RNDN or GMP_RNDNA).
Decide the rounding direction here. */
if (rnd_mode == GMP_RNDN &&
(rp[0] & (MPFR_LIMB_ONE << sh)) == 0)
{ /* halfway cases rounded toward zero */
mp_limb_t a, b;
/* a: rounding bit and some of the following bits */
/* b: boundary for a (weight of the rounding bit in a) */
if (sh != 0)
{
a = rp[0] & ((MPFR_LIMB_ONE << sh) - 1);
b = MPFR_LIMB_ONE << (sh - 1);
}
else
{
a = up[un - rn - 1];
b = MPFR_LIMB_HIGHBIT;
}
rnd_away = a > b;
if (a == b)
{
mp_size_t i;
for (i = un - rn - 1 - (sh == 0); i >= 0; i--)
if (up[i] != 0)
{
rnd_away = 1;
break;
}
}
}
else /* halfway cases rounded away from zero */
rnd_away = /* rounding bit */
((sh != 0 && (rp[0] & (MPFR_LIMB_ONE << (sh - 1))) != 0) ||
(sh == 0 && (up[un - rn - 1] & MPFR_LIMB_HIGHBIT) != 0));
}
if (uflags == 0)
{ /* u is an integer; determine if it is representable in r */
if (sh != 0 && rp[0] << (BITS_PER_MP_LIMB - sh) != 0)
uflags = 1; /* u is not representable in r */
else
{
mp_size_t i;
for (i = un - rn - 1; i >= 0; i--)
if (up[i] != 0)
{
uflags = 1; /* u is not representable in r */
break;
}
}
}
}
else /* ui <= rn */
{
mp_size_t uj, rj;
int ush;
uj = un - ui; /* lowest limb of the integer part in u */
rj = rn - ui; /* lowest limb of the integer part in r */
if (MPFR_LIKELY (rp != up))
MPN_COPY(rp + rj, up + uj, ui);
/* Ignore the lowest rj limbs, all equal to zero. */
rp += rj;
rn = ui;
/* number of fractional bits in whole rp[0] */
ush = idiff == 0 ? 0 : BITS_PER_MP_LIMB - idiff;
if (rj == 0 && ush < sh)
{
/* If u is an integer (uflags == 0), we need to determine
if it is representable in r, i.e. if its sh - ush bits
in the non-significant part of r are all 0. */
if (uflags == 0 && (rp[0] & ((MPFR_LIMB_ONE << sh) -
(MPFR_LIMB_ONE << ush))) != 0)
uflags = 1; /* u is an integer not representable in r */
}
else /* The integer part of u fits in r, we'll round to it. */
sh = ush;
if (rnd_away < 0)
{
/* This is a rounding to nearest mode.
Decide the rounding direction here. */
if (uj == 0 && sh == 0)
rnd_away = 0; /* rounding bit = 0 (not represented in u) */
else if (rnd_mode == GMP_RNDN &&
(rp[0] & (MPFR_LIMB_ONE << sh)) == 0)
{ /* halfway cases rounded toward zero */
mp_limb_t a, b;
/* a: rounding bit and some of the following bits */
/* b: boundary for a (weight of the rounding bit in a) */
if (sh != 0)
{
a = rp[0] & ((MPFR_LIMB_ONE << sh) - 1);
b = MPFR_LIMB_ONE << (sh - 1);
}
else
{
MPFR_ASSERTD (uj >= 1); /* see above */
a = up[uj - 1];
b = MPFR_LIMB_HIGHBIT;
}
rnd_away = a > b;
if (a == b)
{
mp_size_t i;
for (i = uj - 1 - (sh == 0); i >= 0; i--)
if (up[i] != 0)
{
rnd_away = 1;
break;
}
}
}
else /* halfway cases rounded away from zero */
rnd_away = /* rounding bit */
((sh != 0 && (rp[0] & (MPFR_LIMB_ONE << (sh - 1))) != 0) ||
(sh == 0 && (MPFR_ASSERTD (uj >= 1),
up[uj - 1] & MPFR_LIMB_HIGHBIT) != 0));
}
/* Now we can make the low rj limbs to 0 */
MPN_ZERO (rp-rj, rj);
}
if (sh != 0)
rp[0] &= MP_LIMB_T_MAX << sh;
/* If u is a representable integer, there is no rounding. */
if (uflags == 0)
MPFR_RET(0);
MPFR_ASSERTD (rnd_away >= 0); /* rounding direction is defined */
if (rnd_away && mpn_add_1(rp, rp, rn, MPFR_LIMB_ONE << sh))
{
if (exp == __gmpfr_emax)
return mpfr_overflow(r, rnd_mode, MPFR_SIGN(r)) >= 0 ?
uflags : -uflags;
else
{
MPFR_SET_EXP(r, exp + 1);
rp[rn-1] = MPFR_LIMB_HIGHBIT;
}
}
MPFR_RET (rnd_away ^ (sign < 0) ? uflags : -uflags);
} /* exp > 0, |u| >= 1 */
}
int
mpfr_sinh (mpfr_ptr y, mpfr_srcptr xt, mp_rnd_t rnd_mode)
{
mpfr_t x;
int inexact;
MPFR_LOG_FUNC (("x[%#R]=%R rnd=%d", xt, xt, rnd_mode),
("y[%#R]=%R inexact=%d", y, y, inexact));
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (xt)))
{
if (MPFR_IS_NAN (xt))
{
MPFR_SET_NAN (y);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (xt))
{
MPFR_SET_INF (y);
MPFR_SET_SAME_SIGN (y, xt);
MPFR_RET (0);
}
else /* xt is zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO (xt));
MPFR_SET_ZERO (y); /* sinh(0) = 0 */
MPFR_SET_SAME_SIGN (y, xt);
MPFR_RET (0);
}
}
/* sinh(x) = x + x^3/6 + ... so the error is < 2^(3*EXP(x)-2) */
MPFR_FAST_COMPUTE_IF_SMALL_INPUT (y, xt, -2 * MPFR_GET_EXP(xt), 2, 1,
rnd_mode, {});
MPFR_TMP_INIT_ABS (x, xt);
{
mpfr_t t, ti;
mp_exp_t d;
mp_prec_t Nt; /* Precision of the intermediary variable */
long int err; /* Precision of error */
MPFR_ZIV_DECL (loop);
MPFR_SAVE_EXPO_DECL (expo);
MPFR_GROUP_DECL (group);
MPFR_SAVE_EXPO_MARK (expo);
/* compute the precision of intermediary variable */
Nt = MAX (MPFR_PREC (x), MPFR_PREC (y));
/* the optimal number of bits : see algorithms.ps */
Nt = Nt + MPFR_INT_CEIL_LOG2 (Nt) + 4;
/* If x is near 0, exp(x) - 1/exp(x) = 2*x+x^3/3+O(x^5) */
if (MPFR_GET_EXP (x) < 0)
Nt -= 2*MPFR_GET_EXP (x);
/* initialise of intermediary variables */
MPFR_GROUP_INIT_2 (group, Nt, t, ti);
/* First computation of sinh */
MPFR_ZIV_INIT (loop, Nt);
for (;;) {
/* compute sinh */
mpfr_clear_flags ();
mpfr_exp (t, x, GMP_RNDD); /* exp(x) */
/* exp(x) can overflow! */
/* BUG/TODO/FIXME: exp can overflow but sinh may be representable! */
if (MPFR_UNLIKELY (mpfr_overflow_p ())) {
inexact = mpfr_overflow (y, rnd_mode, MPFR_SIGN (xt));
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_OVERFLOW);
break;
}
d = MPFR_GET_EXP (t);
mpfr_ui_div (ti, 1, t, GMP_RNDU); /* 1/exp(x) */
mpfr_sub (t, t, ti, GMP_RNDN); /* exp(x) - 1/exp(x) */
mpfr_div_2ui (t, t, 1, GMP_RNDN); /* 1/2(exp(x) - 1/exp(x)) */
/* it may be that t is zero (in fact, it can only occur when te=1,
and thus ti=1 too) */
if (MPFR_IS_ZERO (t))
err = Nt; /* double the precision */
else
{
/* calculation of the error */
d = d - MPFR_GET_EXP (t) + 2;
/* error estimate: err = Nt-(__gmpfr_ceil_log2(1+pow(2,d)));*/
err = Nt - (MAX (d, 0) + 1);
if (MPFR_LIKELY (MPFR_CAN_ROUND (t, err, MPFR_PREC (y), rnd_mode)))
{
inexact = mpfr_set4 (y, t, rnd_mode, MPFR_SIGN (xt));
break;
}
}
/* actualisation of the precision */
Nt += err;
MPFR_ZIV_NEXT (loop, Nt);
MPFR_GROUP_REPREC_2 (group, Nt, t, ti);
}
MPFR_ZIV_FREE (loop);
MPFR_GROUP_CLEAR (group);
MPFR_SAVE_EXPO_FREE (expo);
}
return mpfr_check_range (y, inexact, 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;
int cn, cd;
long shift;
mp_size_t sn, sd;
MPFR_SAVE_EXPO_DECL (expo);
num = mpq_numref (q);
den = mpq_denref (q);
/* NAN and INF for mpq are not really documented, but could be found */
if (MPFR_UNLIKELY (mpz_sgn (num) == 0))
{
if (MPFR_UNLIKELY (mpz_sgn (den) == 0))
{
MPFR_SET_NAN (f);
MPFR_RET_NAN;
}
else
{
MPFR_SET_ZERO (f);
MPFR_SET_POS (f);
MPFR_RET (0);
}
}
if (MPFR_UNLIKELY (mpz_sgn (den) == 0))
{
MPFR_SET_INF (f);
MPFR_SET_SIGN (f, mpz_sgn (num));
MPFR_RET (0);
}
MPFR_SAVE_EXPO_MARK (expo);
cn = set_z (n, num, &sn);
cd = set_z (d, den, &sd);
sn -= sd;
if (MPFR_UNLIKELY (sn > MPFR_EMAX_MAX / BITS_PER_MP_LIMB))
{
inexact = mpfr_overflow (f, rnd, MPFR_SIGN (f));
goto end;
}
if (MPFR_UNLIKELY (sn < MPFR_EMIN_MIN / BITS_PER_MP_LIMB -1))
{
if (rnd == GMP_RNDN)
rnd = GMP_RNDZ;
inexact = mpfr_underflow (f, rnd, MPFR_SIGN (f));
goto end;
}
inexact = mpfr_div (f, n, d, rnd);
shift = BITS_PER_MP_LIMB*sn+cn-cd;
MPFR_ASSERTD (shift == BITS_PER_MP_LIMB*sn+cn-cd);
cd = mpfr_mul_2si (f, f, shift, rnd);
MPFR_SAVE_EXPO_FREE (expo);
if (MPFR_UNLIKELY (cd != 0))
inexact = cd;
else
inexact = mpfr_check_range (f, inexact, rnd);
end:
mpfr_clear (d);
mpfr_clear (n);
return inexact;
}
/* The computation of z = pow(x,y) is done by
z = exp(y * log(x)) = x^y
For the special cases, see Section F.9.4.4 of the C standard:
_ pow(±0, y) = ±inf for y an odd integer < 0.
_ pow(±0, y) = +inf for y < 0 and not an odd integer.
_ pow(±0, y) = ±0 for y an odd integer > 0.
_ pow(±0, y) = +0 for y > 0 and not an odd integer.
_ pow(-1, ±inf) = 1.
_ pow(+1, y) = 1 for any y, even a NaN.
_ pow(x, ±0) = 1 for any x, even a NaN.
_ pow(x, y) = NaN for finite x < 0 and finite non-integer y.
_ pow(x, -inf) = +inf for |x| < 1.
_ pow(x, -inf) = +0 for |x| > 1.
_ pow(x, +inf) = +0 for |x| < 1.
_ pow(x, +inf) = +inf for |x| > 1.
_ pow(-inf, y) = -0 for y an odd integer < 0.
_ pow(-inf, y) = +0 for y < 0 and not an odd integer.
_ pow(-inf, y) = -inf for y an odd integer > 0.
_ pow(-inf, y) = +inf for y > 0 and not an odd integer.
_ pow(+inf, y) = +0 for y < 0.
_ pow(+inf, y) = +inf for y > 0. */
int
mpfr_pow (mpfr_ptr z, mpfr_srcptr x, mpfr_srcptr y, mpfr_rnd_t rnd_mode)
{
int inexact;
int cmp_x_1;
int y_is_integer;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg y[%Pu]=%.*Rg rnd=%d",
mpfr_get_prec (x), mpfr_log_prec, x,
mpfr_get_prec (y), mpfr_log_prec, y, rnd_mode),
("z[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (z), mpfr_log_prec, z, inexact));
if (MPFR_ARE_SINGULAR (x, y))
{
/* pow(x, 0) returns 1 for any x, even a NaN. */
if (MPFR_UNLIKELY (MPFR_IS_ZERO (y)))
return mpfr_set_ui (z, 1, rnd_mode);
else if (MPFR_IS_NAN (x))
{
MPFR_SET_NAN (z);
MPFR_RET_NAN;
}
else if (MPFR_IS_NAN (y))
{
/* pow(+1, NaN) returns 1. */
if (mpfr_cmp_ui (x, 1) == 0)
return mpfr_set_ui (z, 1, rnd_mode);
MPFR_SET_NAN (z);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (y))
{
if (MPFR_IS_INF (x))
{
if (MPFR_IS_POS (y))
MPFR_SET_INF (z);
else
MPFR_SET_ZERO (z);
MPFR_SET_POS (z);
MPFR_RET (0);
}
else
{
int cmp;
cmp = mpfr_cmpabs (x, __gmpfr_one) * MPFR_INT_SIGN (y);
MPFR_SET_POS (z);
if (cmp > 0)
{
/* Return +inf. */
MPFR_SET_INF (z);
MPFR_RET (0);
}
else if (cmp < 0)
{
/* Return +0. */
MPFR_SET_ZERO (z);
MPFR_RET (0);
}
else
{
/* Return 1. */
return mpfr_set_ui (z, 1, rnd_mode);
}
}
}
else if (MPFR_IS_INF (x))
{
int negative;
/* Determine the sign now, in case y and z are the same object */
negative = MPFR_IS_NEG (x) && is_odd (y);
if (MPFR_IS_POS (y))
MPFR_SET_INF (z);
else
MPFR_SET_ZERO (z);
if (negative)
MPFR_SET_NEG (z);
else
MPFR_SET_POS (z);
MPFR_RET (0);
}
else
{
int negative;
MPFR_ASSERTD (MPFR_IS_ZERO (x));
/* Determine the sign now, in case y and z are the same object */
negative = MPFR_IS_NEG(x) && is_odd (y);
if (MPFR_IS_NEG (y))
{
MPFR_ASSERTD (! MPFR_IS_INF (y));
MPFR_SET_INF (z);
mpfr_set_divby0 ();
}
else
MPFR_SET_ZERO (z);
if (negative)
MPFR_SET_NEG (z);
else
MPFR_SET_POS (z);
MPFR_RET (0);
}
}
/* x^y for x < 0 and y not an integer is not defined */
y_is_integer = mpfr_integer_p (y);
if (MPFR_IS_NEG (x) && ! y_is_integer)
{
MPFR_SET_NAN (z);
MPFR_RET_NAN;
}
/* now the result cannot be NaN:
(1) either x > 0
(2) or x < 0 and y is an integer */
cmp_x_1 = mpfr_cmpabs (x, __gmpfr_one);
if (cmp_x_1 == 0)
return mpfr_set_si (z, MPFR_IS_NEG (x) && is_odd (y) ? -1 : 1, rnd_mode);
/* now we have:
(1) either x > 0
(2) or x < 0 and y is an integer
and in addition |x| <> 1.
*/
/* detect overflow: an overflow is possible if
(a) |x| > 1 and y > 0
(b) |x| < 1 and y < 0.
FIXME: this assumes 1 is always representable.
FIXME2: maybe we can test overflow and underflow simultaneously.
The idea is the following: first compute an approximation to
y * log2|x|, using rounding to nearest. If |x| is not too near from 1,
this approximation should be accurate enough, and in most cases enable
one to prove that there is no underflow nor overflow.
Otherwise, it should enable one to check only underflow or overflow,
instead of both cases as in the present case.
*/
if (cmp_x_1 * MPFR_SIGN (y) > 0)
{
mpfr_t t;
int negative, overflow;
MPFR_SAVE_EXPO_MARK (expo);
mpfr_init2 (t, 53);
/* we want a lower bound on y*log2|x|:
(i) if x > 0, it suffices to round log2(x) toward zero, and
to round y*o(log2(x)) toward zero too;
(ii) if x < 0, we first compute t = o(-x), with rounding toward 1,
and then follow as in case (1). */
if (MPFR_SIGN (x) > 0)
mpfr_log2 (t, x, MPFR_RNDZ);
else
{
mpfr_neg (t, x, (cmp_x_1 > 0) ? MPFR_RNDZ : MPFR_RNDU);
mpfr_log2 (t, t, MPFR_RNDZ);
}
mpfr_mul (t, t, y, MPFR_RNDZ);
overflow = mpfr_cmp_si (t, __gmpfr_emax) > 0;
mpfr_clear (t);
MPFR_SAVE_EXPO_FREE (expo);
if (overflow)
{
MPFR_LOG_MSG (("early overflow detection\n", 0));
negative = MPFR_SIGN(x) < 0 && is_odd (y);
return mpfr_overflow (z, rnd_mode, negative ? -1 : 1);
}
}
/* Basic underflow checking. One has:
* - if y > 0, |x^y| < 2^(EXP(x) * y);
* - if y < 0, |x^y| <= 2^((EXP(x) - 1) * y);
* so that one can compute a value ebound such that |x^y| < 2^ebound.
* If we have ebound <= emin - 2 (emin - 1 in directed rounding modes),
* then there is an underflow and we can decide the return value.
*/
if (MPFR_IS_NEG (y) ? (MPFR_GET_EXP (x) > 1) : (MPFR_GET_EXP (x) < 0))
{
mpfr_t tmp;
mpfr_eexp_t ebound;
int inex2;
/* We must restore the flags. */
MPFR_SAVE_EXPO_MARK (expo);
mpfr_init2 (tmp, sizeof (mpfr_exp_t) * CHAR_BIT);
inex2 = mpfr_set_exp_t (tmp, MPFR_GET_EXP (x), MPFR_RNDN);
MPFR_ASSERTN (inex2 == 0);
if (MPFR_IS_NEG (y))
{
inex2 = mpfr_sub_ui (tmp, tmp, 1, MPFR_RNDN);
MPFR_ASSERTN (inex2 == 0);
}
mpfr_mul (tmp, tmp, y, MPFR_RNDU);
if (MPFR_IS_NEG (y))
mpfr_nextabove (tmp);
/* tmp doesn't necessarily fit in ebound, but that doesn't matter
since we get the minimum value in such a case. */
ebound = mpfr_get_exp_t (tmp, MPFR_RNDU);
mpfr_clear (tmp);
MPFR_SAVE_EXPO_FREE (expo);
if (MPFR_UNLIKELY (ebound <=
__gmpfr_emin - (rnd_mode == MPFR_RNDN ? 2 : 1)))
{
/* warning: mpfr_underflow rounds away from 0 for MPFR_RNDN */
MPFR_LOG_MSG (("early underflow detection\n", 0));
return mpfr_underflow (z,
rnd_mode == MPFR_RNDN ? MPFR_RNDZ : rnd_mode,
MPFR_SIGN (x) < 0 && is_odd (y) ? -1 : 1);
}
}
/* If y is an integer, we can use mpfr_pow_z (based on multiplications),
but if y is very large (I'm not sure about the best threshold -- VL),
we shouldn't use it, as it can be very slow and take a lot of memory
(and even crash or make other programs crash, as several hundred of
MBs may be necessary). Note that in such a case, either x = +/-2^b
(this case is handled below) or x^y cannot be represented exactly in
any precision supported by MPFR (the general case uses this property).
*/
if (y_is_integer && (MPFR_GET_EXP (y) <= 256))
{
mpz_t zi;
MPFR_LOG_MSG (("special code for y not too large integer\n", 0));
mpz_init (zi);
mpfr_get_z (zi, y, MPFR_RNDN);
inexact = mpfr_pow_z (z, x, zi, rnd_mode);
mpz_clear (zi);
return inexact;
}
/* Special case (+/-2^b)^Y which could be exact. If x is negative, then
necessarily y is a large integer. */
{
mpfr_exp_t b = MPFR_GET_EXP (x) - 1;
MPFR_ASSERTN (b >= LONG_MIN && b <= LONG_MAX); /* FIXME... */
if (mpfr_cmp_si_2exp (x, MPFR_SIGN(x), b) == 0)
{
mpfr_t tmp;
int sgnx = MPFR_SIGN (x);
MPFR_LOG_MSG (("special case (+/-2^b)^Y\n", 0));
/* now x = +/-2^b, so x^y = (+/-1)^y*2^(b*y) is exact whenever b*y is
an integer */
MPFR_SAVE_EXPO_MARK (expo);
mpfr_init2 (tmp, MPFR_PREC (y) + sizeof (long) * CHAR_BIT);
inexact = mpfr_mul_si (tmp, y, b, MPFR_RNDN); /* exact */
MPFR_ASSERTN (inexact == 0);
/* Note: as the exponent range has been extended, an overflow is not
possible (due to basic overflow and underflow checking above, as
the result is ~ 2^tmp), and an underflow is not possible either
because b is an integer (thus either 0 or >= 1). */
MPFR_CLEAR_FLAGS ();
inexact = mpfr_exp2 (z, tmp, rnd_mode);
mpfr_clear (tmp);
if (sgnx < 0 && is_odd (y))
{
mpfr_neg (z, z, rnd_mode);
inexact = -inexact;
}
/* Without the following, the overflows3 test in tpow.c fails. */
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, __gmpfr_flags);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (z, inexact, rnd_mode);
}
}
MPFR_SAVE_EXPO_MARK (expo);
/* Case where |y * log(x)| is very small. Warning: x can be negative, in
that case y is a large integer. */
{
mpfr_t t;
mpfr_exp_t err;
/* We need an upper bound on the exponent of y * log(x). */
mpfr_init2 (t, 16);
if (MPFR_IS_POS(x))
mpfr_log (t, x, cmp_x_1 < 0 ? MPFR_RNDD : MPFR_RNDU); /* away from 0 */
else
{
/* if x < -1, round to +Inf, else round to zero */
mpfr_neg (t, x, (mpfr_cmp_si (x, -1) < 0) ? MPFR_RNDU : MPFR_RNDD);
mpfr_log (t, t, (mpfr_cmp_ui (t, 1) < 0) ? MPFR_RNDD : MPFR_RNDU);
}
MPFR_ASSERTN (MPFR_IS_PURE_FP (t));
err = MPFR_GET_EXP (y) + MPFR_GET_EXP (t);
mpfr_clear (t);
MPFR_CLEAR_FLAGS ();
MPFR_SMALL_INPUT_AFTER_SAVE_EXPO (z, __gmpfr_one, - err, 0,
(MPFR_SIGN (y) > 0) ^ (cmp_x_1 < 0),
rnd_mode, expo, {});
}
/* General case */
inexact = mpfr_pow_general (z, x, y, rnd_mode, y_is_integer, &expo);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (z, inexact, rnd_mode);
}
int
mpfr_pow_si (mpfr_ptr y, mpfr_srcptr x, long int n, mp_rnd_t rnd)
{
if (n >= 0)
return mpfr_pow_ui (y, x, n, rnd);
else
{
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))
{
MPFR_SET_ZERO (y);
if (MPFR_IS_POS (x) || ((unsigned) n & 1) == 0)
MPFR_SET_POS (y);
else
MPFR_SET_NEG (y);
MPFR_RET (0);
}
else /* x is zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO (x));
MPFR_SET_INF(y);
if (MPFR_IS_POS (x) || ((unsigned) n & 1) == 0)
MPFR_SET_POS (y);
else
MPFR_SET_NEG (y);
MPFR_RET(0);
}
}
MPFR_CLEAR_FLAGS (y);
/* detect exact powers: x^(-n) is exact iff x is a power of 2 */
if (mpfr_cmp_si_2exp (x, MPFR_SIGN(x), MPFR_EXP(x) - 1) == 0)
{
mp_exp_t expx = MPFR_EXP (x) - 1, expy;
MPFR_ASSERTD (n < 0);
/* Warning: n * expx may overflow!
* Some systems (apparently alpha-freebsd) abort with
* LONG_MIN / 1, and LONG_MIN / -1 is undefined.
* Proof of the overflow checking. The expressions below are
* assumed to be on the rational numbers, but the word "overflow"
* still has its own meaning in the C context. / still denotes
* the integer (truncated) division, and // denotes the exact
* division.
* - First, (__gmpfr_emin - 1) / n and (__gmpfr_emax - 1) / n
* cannot overflow due to the constraints on the exponents of
* MPFR numbers.
* - If n = -1, then n * expx = - expx, which is representable
* because of the constraints on the exponents of MPFR numbers.
* - If expx = 0, then n * expx = 0, which is representable.
* - If n < -1 and expx > 0:
* + If expx > (__gmpfr_emin - 1) / n, then
* expx >= (__gmpfr_emin - 1) / n + 1
* > (__gmpfr_emin - 1) // n,
* and
* n * expx < __gmpfr_emin - 1,
* i.e.
* n * expx <= __gmpfr_emin - 2.
* This corresponds to an underflow, with a null result in
* the rounding-to-nearest mode.
* + If expx <= (__gmpfr_emin - 1) / n, then n * expx cannot
* overflow since 0 < expx <= (__gmpfr_emin - 1) / n and
* 0 > n * expx >= n * ((__gmpfr_emin - 1) / n)
* >= __gmpfr_emin - 1.
* - If n < -1 and expx < 0:
* + If expx < (__gmpfr_emax - 1) / n, then
* expx <= (__gmpfr_emax - 1) / n - 1
* < (__gmpfr_emax - 1) // n,
* and
* n * expx > __gmpfr_emax - 1,
* i.e.
* n * expx >= __gmpfr_emax.
* This corresponds to an overflow (2^(n * expx) has an
* exponent > __gmpfr_emax).
* + If expx >= (__gmpfr_emax - 1) / n, then n * expx cannot
* overflow since 0 > expx >= (__gmpfr_emax - 1) / n and
* 0 < n * expx <= n * ((__gmpfr_emax - 1) / n)
* <= __gmpfr_emax - 1.
* Note: one could use expx bounds based on MPFR_EXP_MIN and
* MPFR_EXP_MAX instead of __gmpfr_emin and __gmpfr_emax. The
* current bounds do not lead to noticeably slower code and
* allow us to avoid a bug in Sun's compiler for Solaris/x86
* (when optimizations are enabled).
*/
expy =
n != -1 && expx > 0 && expx > (__gmpfr_emin - 1) / n ?
MPFR_EMIN_MIN - 2 /* Underflow */ :
n != -1 && expx < 0 && expx < (__gmpfr_emax - 1) / n ?
MPFR_EMAX_MAX /* Overflow */ : n * expx;
return mpfr_set_si_2exp (y, n % 2 ? MPFR_INT_SIGN (x) : 1,
expy, rnd);
}
/* General case */
{
/* Declaration of the intermediary variable */
mpfr_t t;
/* Declaration of the size variable */
mp_prec_t Ny = MPFR_PREC (y); /* target precision */
mp_prec_t Nt; /* working precision */
mp_exp_t err; /* error */
int inexact;
unsigned long abs_n;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_ZIV_DECL (loop);
abs_n = - (unsigned long) n;
/* compute the precision of intermediary variable */
/* the optimal number of bits : see algorithms.tex */
Nt = Ny + 3 + MPFR_INT_CEIL_LOG2 (Ny);
MPFR_SAVE_EXPO_MARK (expo);
/* initialise of intermediary variable */
mpfr_init2 (t, Nt);
MPFR_ZIV_INIT (loop, Nt);
for (;;)
{
/* compute 1/(x^n), with n > 0 */
mpfr_pow_ui (t, x, abs_n, GMP_RNDN);
mpfr_ui_div (t, 1, t, GMP_RNDN);
/* FIXME: old code improved, but I think this is still incorrect. */
if (MPFR_UNLIKELY (MPFR_IS_ZERO (t)))
{
MPFR_ZIV_FREE (loop);
mpfr_clear (t);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_underflow (y, rnd == GMP_RNDN ? GMP_RNDZ : rnd,
abs_n & 1 ? MPFR_SIGN (x) :
MPFR_SIGN_POS);
}
if (MPFR_UNLIKELY (MPFR_IS_INF (t)))
{
MPFR_ZIV_FREE (loop);
mpfr_clear (t);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_overflow (y, rnd, abs_n & 1 ? MPFR_SIGN (x) :
MPFR_SIGN_POS);
}
/* error estimate -- see pow function in algorithms.ps */
err = Nt - 3;
if (MPFR_LIKELY (MPFR_CAN_ROUND (t, err, Ny, rnd)))
break;
/* actualisation of the precision */
Nt += BITS_PER_MP_LIMB;
mpfr_set_prec (t, Nt);
}
MPFR_ZIV_FREE (loop);
inexact = mpfr_set (y, t, rnd);
mpfr_clear (t);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (y, inexact, rnd);
}
}
}
int
mpfr_pow_si (mpfr_ptr y, mpfr_srcptr x, long int n, mpfr_rnd_t rnd)
{
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg n=%ld rnd=%d",
mpfr_get_prec (x), mpfr_log_prec, x, n, rnd),
("y[%Pu]=%.*Rg", mpfr_get_prec (y), mpfr_log_prec, y));
if (n >= 0)
return mpfr_pow_ui (y, x, n, rnd);
else
{
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x)))
{
if (MPFR_IS_NAN (x))
{
MPFR_SET_NAN (y);
MPFR_RET_NAN;
}
else
{
int positive = MPFR_IS_POS (x) || ((unsigned long) n & 1) == 0;
if (MPFR_IS_INF (x))
MPFR_SET_ZERO (y);
else /* x is zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO (x));
MPFR_SET_INF (y);
mpfr_set_divby0 ();
}
if (positive)
MPFR_SET_POS (y);
else
MPFR_SET_NEG (y);
MPFR_RET (0);
}
}
/* detect exact powers: x^(-n) is exact iff x is a power of 2 */
if (mpfr_cmp_si_2exp (x, MPFR_SIGN(x), MPFR_EXP(x) - 1) == 0)
{
mpfr_exp_t expx = MPFR_EXP (x) - 1, expy;
MPFR_ASSERTD (n < 0);
/* Warning: n * expx may overflow!
*
* Some systems (apparently alpha-freebsd) abort with
* LONG_MIN / 1, and LONG_MIN / -1 is undefined.
* http://www.freebsd.org/cgi/query-pr.cgi?pr=72024
*
* Proof of the overflow checking. The expressions below are
* assumed to be on the rational numbers, but the word "overflow"
* still has its own meaning in the C context. / still denotes
* the integer (truncated) division, and // denotes the exact
* division.
* - First, (__gmpfr_emin - 1) / n and (__gmpfr_emax - 1) / n
* cannot overflow due to the constraints on the exponents of
* MPFR numbers.
* - If n = -1, then n * expx = - expx, which is representable
* because of the constraints on the exponents of MPFR numbers.
* - If expx = 0, then n * expx = 0, which is representable.
* - If n < -1 and expx > 0:
* + If expx > (__gmpfr_emin - 1) / n, then
* expx >= (__gmpfr_emin - 1) / n + 1
* > (__gmpfr_emin - 1) // n,
* and
* n * expx < __gmpfr_emin - 1,
* i.e.
* n * expx <= __gmpfr_emin - 2.
* This corresponds to an underflow, with a null result in
* the rounding-to-nearest mode.
* + If expx <= (__gmpfr_emin - 1) / n, then n * expx cannot
* overflow since 0 < expx <= (__gmpfr_emin - 1) / n and
* 0 > n * expx >= n * ((__gmpfr_emin - 1) / n)
* >= __gmpfr_emin - 1.
* - If n < -1 and expx < 0:
* + If expx < (__gmpfr_emax - 1) / n, then
* expx <= (__gmpfr_emax - 1) / n - 1
* < (__gmpfr_emax - 1) // n,
* and
* n * expx > __gmpfr_emax - 1,
* i.e.
* n * expx >= __gmpfr_emax.
* This corresponds to an overflow (2^(n * expx) has an
* exponent > __gmpfr_emax).
* + If expx >= (__gmpfr_emax - 1) / n, then n * expx cannot
* overflow since 0 > expx >= (__gmpfr_emax - 1) / n and
* 0 < n * expx <= n * ((__gmpfr_emax - 1) / n)
* <= __gmpfr_emax - 1.
* Note: one could use expx bounds based on MPFR_EXP_MIN and
* MPFR_EXP_MAX instead of __gmpfr_emin and __gmpfr_emax. The
* current bounds do not lead to noticeably slower code and
* allow us to avoid a bug in Sun's compiler for Solaris/x86
* (when optimizations are enabled); known affected versions:
* cc: Sun C 5.8 2005/10/13
* cc: Sun C 5.8 Patch 121016-02 2006/03/31
* cc: Sun C 5.8 Patch 121016-04 2006/10/18
*/
expy =
n != -1 && expx > 0 && expx > (__gmpfr_emin - 1) / n ?
MPFR_EMIN_MIN - 2 /* Underflow */ :
n != -1 && expx < 0 && expx < (__gmpfr_emax - 1) / n ?
MPFR_EMAX_MAX /* Overflow */ : n * expx;
return mpfr_set_si_2exp (y, n % 2 ? MPFR_INT_SIGN (x) : 1,
expy, rnd);
}
/* General case */
{
/* Declaration of the intermediary variable */
mpfr_t t;
/* Declaration of the size variable */
mpfr_prec_t Ny; /* target precision */
mpfr_prec_t Nt; /* working precision */
mpfr_rnd_t rnd1;
int size_n;
int inexact;
unsigned long abs_n;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_ZIV_DECL (loop);
abs_n = - (unsigned long) n;
count_leading_zeros (size_n, (mp_limb_t) abs_n);
size_n = GMP_NUMB_BITS - size_n;
/* initial working precision */
Ny = MPFR_PREC (y);
Nt = Ny + size_n + 3 + MPFR_INT_CEIL_LOG2 (Ny);
MPFR_SAVE_EXPO_MARK (expo);
/* initialise of intermediary variable */
mpfr_init2 (t, Nt);
/* We will compute rnd(rnd1(1/x) ^ |n|), where rnd1 is the rounding
toward sign(x), to avoid spurious overflow or underflow, as in
mpfr_pow_z. */
rnd1 = MPFR_EXP (x) < 1 ? MPFR_RNDZ :
(MPFR_SIGN (x) > 0 ? MPFR_RNDU : MPFR_RNDD);
MPFR_ZIV_INIT (loop, Nt);
for (;;)
{
MPFR_BLOCK_DECL (flags);
/* compute (1/x)^|n| */
MPFR_BLOCK (flags, mpfr_ui_div (t, 1, x, rnd1));
MPFR_ASSERTD (! MPFR_UNDERFLOW (flags));
/* t = (1/x)*(1+theta) where |theta| <= 2^(-Nt) */
if (MPFR_UNLIKELY (MPFR_OVERFLOW (flags)))
goto overflow;
MPFR_BLOCK (flags, mpfr_pow_ui (t, t, abs_n, rnd));
/* t = (1/x)^|n|*(1+theta')^(|n|+1) where |theta'| <= 2^(-Nt).
If (|n|+1)*2^(-Nt) <= 1/2, which is satisfied as soon as
Nt >= bits(n)+2, then we can use Lemma \ref{lemma_graillat}
from algorithms.tex, which yields x^n*(1+theta) with
|theta| <= 2(|n|+1)*2^(-Nt), thus the error is bounded by
2(|n|+1) ulps <= 2^(bits(n)+2) ulps. */
if (MPFR_UNLIKELY (MPFR_OVERFLOW (flags)))
{
overflow:
MPFR_ZIV_FREE (loop);
mpfr_clear (t);
MPFR_SAVE_EXPO_FREE (expo);
MPFR_LOG_MSG (("overflow\n", 0));
return mpfr_overflow (y, rnd, abs_n & 1 ?
MPFR_SIGN (x) : MPFR_SIGN_POS);
}
if (MPFR_UNLIKELY (MPFR_UNDERFLOW (flags)))
{
MPFR_ZIV_FREE (loop);
mpfr_clear (t);
MPFR_LOG_MSG (("underflow\n", 0));
if (rnd == MPFR_RNDN)
{
mpfr_t y2, nn;
/* We cannot decide now whether the result should be
rounded toward zero or away from zero. So, like
in mpfr_pow_pos_z, let's use the general case of
mpfr_pow in precision 2. */
MPFR_ASSERTD (mpfr_cmp_si_2exp (x, MPFR_SIGN (x),
MPFR_EXP (x) - 1) != 0);
mpfr_init2 (y2, 2);
mpfr_init2 (nn, sizeof (long) * CHAR_BIT);
inexact = mpfr_set_si (nn, n, MPFR_RNDN);
MPFR_ASSERTN (inexact == 0);
inexact = mpfr_pow_general (y2, x, nn, rnd, 1,
(mpfr_save_expo_t *) NULL);
mpfr_clear (nn);
mpfr_set (y, y2, MPFR_RNDN);
mpfr_clear (y2);
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_UNDERFLOW);
goto end;
}
else
{
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_underflow (y, rnd, abs_n & 1 ?
MPFR_SIGN (x) : MPFR_SIGN_POS);
}
}
/* error estimate -- see pow function in algorithms.ps */
if (MPFR_LIKELY (MPFR_CAN_ROUND (t, Nt - size_n - 2, Ny, rnd)))
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);
mpfr_clear (t);
end:
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (y, inexact, rnd);
}
}
}
/* Assumes that the exponent range has already been extended and if y is
an integer, then the result is not exact in unbounded exponent range. */
int
mpfr_pow_general (mpfr_ptr z, mpfr_srcptr x, mpfr_srcptr y,
mpfr_rnd_t rnd_mode, int y_is_integer, mpfr_save_expo_t *expo)
{
mpfr_t t, u, k, absx;
int neg_result = 0;
int k_non_zero = 0;
int check_exact_case = 0;
int inexact;
/* Declaration of the size variable */
mpfr_prec_t Nz = MPFR_PREC(z); /* target precision */
mpfr_prec_t Nt; /* working precision */
mpfr_exp_t err; /* error */
MPFR_ZIV_DECL (ziv_loop);
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg y[%Pu]=%.*Rg rnd=%d",
mpfr_get_prec (x), mpfr_log_prec, x,
mpfr_get_prec (y), mpfr_log_prec, y, rnd_mode),
("z[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (z), mpfr_log_prec, z, inexact));
/* We put the absolute value of x in absx, pointing to the significand
of x to avoid allocating memory for the significand of absx. */
MPFR_ALIAS(absx, x, /*sign=*/ 1, /*EXP=*/ MPFR_EXP(x));
/* We will compute the absolute value of the result. So, let's
invert the rounding mode if the result is negative. */
if (MPFR_IS_NEG (x) && is_odd (y))
{
neg_result = 1;
rnd_mode = MPFR_INVERT_RND (rnd_mode);
}
/* compute the precision of intermediary variable */
/* the optimal number of bits : see algorithms.tex */
Nt = Nz + 5 + MPFR_INT_CEIL_LOG2 (Nz);
/* initialise of intermediary variable */
mpfr_init2 (t, Nt);
MPFR_ZIV_INIT (ziv_loop, Nt);
for (;;)
{
MPFR_BLOCK_DECL (flags1);
/* compute exp(y*ln|x|), using MPFR_RNDU to get an upper bound, so
that we can detect underflows. */
mpfr_log (t, absx, MPFR_IS_NEG (y) ? MPFR_RNDD : MPFR_RNDU); /* ln|x| */
mpfr_mul (t, y, t, MPFR_RNDU); /* y*ln|x| */
if (k_non_zero)
{
MPFR_LOG_MSG (("subtract k * ln(2)\n", 0));
mpfr_const_log2 (u, MPFR_RNDD);
mpfr_mul (u, u, k, MPFR_RNDD);
/* Error on u = k * log(2): < k * 2^(-Nt) < 1. */
mpfr_sub (t, t, u, MPFR_RNDU);
MPFR_LOG_MSG (("t = y * ln|x| - k * ln(2)\n", 0));
MPFR_LOG_VAR (t);
}
/* estimate of the error -- see pow function in algorithms.tex.
The error on t is at most 1/2 + 3*2^(EXP(t)+1) ulps, which is
<= 2^(EXP(t)+3) for EXP(t) >= -1, and <= 2 ulps for EXP(t) <= -2.
Additional error if k_no_zero: treal = t * errk, with
1 - |k| * 2^(-Nt) <= exp(-|k| * 2^(-Nt)) <= errk <= 1,
i.e., additional absolute error <= 2^(EXP(k)+EXP(t)-Nt).
Total error <= 2^err1 + 2^err2 <= 2^(max(err1,err2)+1). */
err = MPFR_NOTZERO (t) && MPFR_GET_EXP (t) >= -1 ?
MPFR_GET_EXP (t) + 3 : 1;
if (k_non_zero)
{
if (MPFR_GET_EXP (k) > err)
err = MPFR_GET_EXP (k);
err++;
}
MPFR_BLOCK (flags1, mpfr_exp (t, t, MPFR_RNDN)); /* exp(y*ln|x|)*/
/* We need to test */
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (t) || MPFR_UNDERFLOW (flags1)))
{
mpfr_prec_t Ntmin;
MPFR_BLOCK_DECL (flags2);
MPFR_ASSERTN (!k_non_zero);
MPFR_ASSERTN (!MPFR_IS_NAN (t));
/* Real underflow? */
if (MPFR_IS_ZERO (t))
{
/* Underflow. We computed rndn(exp(t)), where t >= y*ln|x|.
Therefore rndn(|x|^y) = 0, and we have a real underflow on
|x|^y. */
inexact = mpfr_underflow (z, rnd_mode == MPFR_RNDN ? MPFR_RNDZ
: rnd_mode, MPFR_SIGN_POS);
if (expo != NULL)
MPFR_SAVE_EXPO_UPDATE_FLAGS (*expo, MPFR_FLAGS_INEXACT
| MPFR_FLAGS_UNDERFLOW);
break;
}
/* Real overflow? */
if (MPFR_IS_INF (t))
{
/* Note: we can probably use a low precision for this test. */
mpfr_log (t, absx, MPFR_IS_NEG (y) ? MPFR_RNDU : MPFR_RNDD);
mpfr_mul (t, y, t, MPFR_RNDD); /* y * ln|x| */
MPFR_BLOCK (flags2, mpfr_exp (t, t, MPFR_RNDD));
/* t = lower bound on exp(y * ln|x|) */
if (MPFR_OVERFLOW (flags2))
{
/* We have computed a lower bound on |x|^y, and it
overflowed. Therefore we have a real overflow
on |x|^y. */
inexact = mpfr_overflow (z, rnd_mode, MPFR_SIGN_POS);
if (expo != NULL)
MPFR_SAVE_EXPO_UPDATE_FLAGS (*expo, MPFR_FLAGS_INEXACT
| MPFR_FLAGS_OVERFLOW);
break;
}
}
k_non_zero = 1;
Ntmin = sizeof(mpfr_exp_t) * CHAR_BIT;
if (Ntmin > Nt)
{
Nt = Ntmin;
mpfr_set_prec (t, Nt);
}
mpfr_init2 (u, Nt);
mpfr_init2 (k, Ntmin);
mpfr_log2 (k, absx, MPFR_RNDN);
mpfr_mul (k, y, k, MPFR_RNDN);
mpfr_round (k, k);
MPFR_LOG_VAR (k);
/* |y| < 2^Ntmin, therefore |k| < 2^Nt. */
continue;
}
if (MPFR_LIKELY (MPFR_CAN_ROUND (t, Nt - err, Nz, rnd_mode)))
{
inexact = mpfr_set (z, t, rnd_mode);
break;
}
/* check exact power, except when y is an integer (since the
exact cases for y integer have already been filtered out) */
if (check_exact_case == 0 && ! y_is_integer)
{
if (mpfr_pow_is_exact (z, absx, y, rnd_mode, &inexact))
break;
check_exact_case = 1;
}
/* reactualisation of the precision */
MPFR_ZIV_NEXT (ziv_loop, Nt);
mpfr_set_prec (t, Nt);
if (k_non_zero)
mpfr_set_prec (u, Nt);
}
MPFR_ZIV_FREE (ziv_loop);
if (k_non_zero)
{
int inex2;
long lk;
/* The rounded result in an unbounded exponent range is z * 2^k. As
* MPFR chooses underflow after rounding, the mpfr_mul_2si below will
* correctly detect underflows and overflows. However, in rounding to
* nearest, if z * 2^k = 2^(emin - 2), then the double rounding may
* affect the result. We need to cope with that before overwriting z.
* This can occur only if k < 0 (this test is necessary to avoid a
* potential integer overflow).
* If inexact >= 0, then the real result is <= 2^(emin - 2), so that
* o(2^(emin - 2)) = +0 is correct. If inexact < 0, then the real
* result is > 2^(emin - 2) and we need to round to 2^(emin - 1).
*/
MPFR_ASSERTN (MPFR_EXP_MAX <= LONG_MAX);
lk = mpfr_get_si (k, MPFR_RNDN);
/* Due to early overflow detection, |k| should not be much larger than
* MPFR_EMAX_MAX, and as MPFR_EMAX_MAX <= MPFR_EXP_MAX/2 <= LONG_MAX/2,
* an overflow should not be possible in mpfr_get_si (and lk is exact).
* And one even has the following assertion. TODO: complete proof.
*/
MPFR_ASSERTD (lk > LONG_MIN && lk < LONG_MAX);
/* Note: even in case of overflow (lk inexact), the code is correct.
* Indeed, for the 3 occurrences of lk:
* - The test lk < 0 is correct as sign(lk) = sign(k).
* - In the test MPFR_GET_EXP (z) == __gmpfr_emin - 1 - lk,
* if lk is inexact, then lk = LONG_MIN <= MPFR_EXP_MIN
* (the minimum value of the mpfr_exp_t type), and
* __gmpfr_emin - 1 - lk >= MPFR_EMIN_MIN - 1 - 2 * MPFR_EMIN_MIN
* >= - MPFR_EMIN_MIN - 1 = MPFR_EMAX_MAX - 1. However, from the
* choice of k, z has been chosen to be around 1, so that the
* result of the test is false, as if lk were exact.
* - In the mpfr_mul_2si (z, z, lk, rnd_mode), if lk is inexact,
* then |lk| >= LONG_MAX >= MPFR_EXP_MAX, and as z is around 1,
* mpfr_mul_2si underflows or overflows in the same way as if
* lk were exact.
* TODO: give a bound on |t|, then on |EXP(z)|.
*/
if (rnd_mode == MPFR_RNDN && inexact < 0 && lk < 0 &&
MPFR_GET_EXP (z) == __gmpfr_emin - 1 - lk && mpfr_powerof2_raw (z))
{
/* Rounding to nearest, real result > z * 2^k = 2^(emin - 2),
* underflow case: as the minimum precision is > 1, we will
* obtain the correct result and exceptions by replacing z by
* nextabove(z).
*/
MPFR_ASSERTN (MPFR_PREC_MIN > 1);
mpfr_nextabove (z);
}
MPFR_CLEAR_FLAGS ();
inex2 = mpfr_mul_2si (z, z, lk, rnd_mode);
if (inex2) /* underflow or overflow */
{
inexact = inex2;
if (expo != NULL)
MPFR_SAVE_EXPO_UPDATE_FLAGS (*expo, __gmpfr_flags);
}
mpfr_clears (u, k, (mpfr_ptr) 0);
}
mpfr_clear (t);
/* update the sign of the result if x was negative */
if (neg_result)
{
MPFR_SET_NEG(z);
inexact = -inexact;
}
return inexact;
}
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);
}
int
mpfr_mul (mpfr_ptr a, mpfr_srcptr b, mpfr_srcptr c, mpfr_rnd_t rnd_mode)
{
int sign, inexact;
mpfr_exp_t ax, ax2;
mp_limb_t *tmp;
mp_limb_t b1;
mpfr_prec_t bq, cq;
mp_size_t bn, cn, tn, k;
MPFR_TMP_DECL (marker);
MPFR_LOG_FUNC (("b[%#R]=%R c[%#R]=%R rnd=%d", b, b, c, c, rnd_mode),
("a[%#R]=%R inexact=%d", a, a, inexact));
/* deal with special cases */
if (MPFR_ARE_SINGULAR (b, c))
{
if (MPFR_IS_NAN (b) || MPFR_IS_NAN (c))
{
MPFR_SET_NAN (a);
MPFR_RET_NAN;
}
sign = MPFR_MULT_SIGN (MPFR_SIGN (b), MPFR_SIGN (c));
if (MPFR_IS_INF (b))
{
if (!MPFR_IS_ZERO (c))
{
MPFR_SET_SIGN (a, sign);
MPFR_SET_INF (a);
MPFR_RET (0);
}
else
{
MPFR_SET_NAN (a);
MPFR_RET_NAN;
}
}
else if (MPFR_IS_INF (c))
{
if (!MPFR_IS_ZERO (b))
{
MPFR_SET_SIGN (a, sign);
MPFR_SET_INF (a);
MPFR_RET(0);
}
else
{
MPFR_SET_NAN (a);
MPFR_RET_NAN;
}
}
else
{
MPFR_ASSERTD (MPFR_IS_ZERO(b) || MPFR_IS_ZERO(c));
MPFR_SET_SIGN (a, sign);
MPFR_SET_ZERO (a);
MPFR_RET (0);
}
}
sign = MPFR_MULT_SIGN (MPFR_SIGN (b), MPFR_SIGN (c));
ax = MPFR_GET_EXP (b) + MPFR_GET_EXP (c);
/* Note: the exponent of the exact result will be e = bx + cx + ec with
ec in {-1,0,1} and the following assumes that e is representable. */
/* FIXME: Useful since we do an exponent check after ?
* It is useful iff the precision is big, there is an overflow
* and we are doing further mults...*/
#ifdef HUGE
if (MPFR_UNLIKELY (ax > __gmpfr_emax + 1))
return mpfr_overflow (a, rnd_mode, sign);
if (MPFR_UNLIKELY (ax < __gmpfr_emin - 2))
return mpfr_underflow (a, rnd_mode == MPFR_RNDN ? MPFR_RNDZ : rnd_mode,
sign);
#endif
bq = MPFR_PREC (b);
cq = MPFR_PREC (c);
MPFR_ASSERTD (bq+cq > bq); /* PREC_MAX is /2 so no integer overflow */
bn = (bq+GMP_NUMB_BITS-1)/GMP_NUMB_BITS; /* number of limbs of b */
cn = (cq+GMP_NUMB_BITS-1)/GMP_NUMB_BITS; /* number of limbs of c */
k = bn + cn; /* effective nb of limbs used by b*c (= tn or tn+1) below */
tn = (bq + cq + GMP_NUMB_BITS - 1) / GMP_NUMB_BITS;
MPFR_ASSERTD (tn <= k); /* tn <= k, thus no int overflow */
/* Check for no size_t overflow*/
MPFR_ASSERTD ((size_t) k <= ((size_t) -1) / BYTES_PER_MP_LIMB);
MPFR_TMP_MARK (marker);
tmp = (mp_limb_t *) MPFR_TMP_ALLOC ((size_t) k * BYTES_PER_MP_LIMB);
/* multiplies two mantissa in temporary allocated space */
if (MPFR_UNLIKELY (bn < cn))
{
mpfr_srcptr z = b;
mp_size_t zn = bn;
b = c;
bn = cn;
c = z;
cn = zn;
}
MPFR_ASSERTD (bn >= cn);
if (MPFR_LIKELY (bn <= 2))
{
if (bn == 1)
{
/* 1 limb * 1 limb */
umul_ppmm (tmp[1], tmp[0], MPFR_MANT (b)[0], MPFR_MANT (c)[0]);
b1 = tmp[1];
}
else if (MPFR_UNLIKELY (cn == 1))
{
/* 2 limbs * 1 limb */
mp_limb_t t;
umul_ppmm (tmp[1], tmp[0], MPFR_MANT (b)[0], MPFR_MANT (c)[0]);
umul_ppmm (tmp[2], t, MPFR_MANT (b)[1], MPFR_MANT (c)[0]);
add_ssaaaa (tmp[2], tmp[1], tmp[2], tmp[1], 0, t);
b1 = tmp[2];
}
else
{
/* 2 limbs * 2 limbs */
mp_limb_t t1, t2, t3;
/* First 2 limbs * 1 limb */
umul_ppmm (tmp[1], tmp[0], MPFR_MANT (b)[0], MPFR_MANT (c)[0]);
umul_ppmm (tmp[2], t1, MPFR_MANT (b)[1], MPFR_MANT (c)[0]);
add_ssaaaa (tmp[2], tmp[1], tmp[2], tmp[1], 0, t1);
/* Second, the other 2 limbs * 1 limb product */
umul_ppmm (t1, t2, MPFR_MANT (b)[0], MPFR_MANT (c)[1]);
umul_ppmm (tmp[3], t3, MPFR_MANT (b)[1], MPFR_MANT (c)[1]);
add_ssaaaa (tmp[3], t1, tmp[3], t1, 0, t3);
/* Sum those two partial products */
add_ssaaaa (tmp[2], tmp[1], tmp[2], tmp[1], t1, t2);
tmp[3] += (tmp[2] < t1);
b1 = tmp[3];
}
b1 >>= (GMP_NUMB_BITS - 1);
tmp += k - tn;
if (MPFR_UNLIKELY (b1 == 0))
mpn_lshift (tmp, tmp, tn, 1); /* tn <= k, so no stack corruption */
}
else
/* Mulders' mulhigh. Disable if squaring, since it is not tuned for
such a case */
if (MPFR_UNLIKELY (bn > MPFR_MUL_THRESHOLD && b != c))
int
mpfr_sqr (mpfr_ptr a, mpfr_srcptr b, mpfr_rnd_t rnd_mode)
{
int cc, inexact;
mpfr_exp_t ax;
mp_limb_t *tmp;
mp_limb_t b1;
mpfr_prec_t bq;
mp_size_t bn, tn;
MPFR_TMP_DECL(marker);
MPFR_LOG_FUNC (("x[%#R]=%R rnd=%d", b, b, rnd_mode),
("y[%#R]=%R inexact=%d", a, a, inexact));
/* deal with special cases */
if (MPFR_UNLIKELY(MPFR_IS_SINGULAR(b)))
{
if (MPFR_IS_NAN(b))
{
MPFR_SET_NAN(a);
MPFR_RET_NAN;
}
MPFR_SET_POS (a);
if (MPFR_IS_INF(b))
MPFR_SET_INF(a);
else
( MPFR_ASSERTD(MPFR_IS_ZERO(b)), MPFR_SET_ZERO(a) );
MPFR_RET(0);
}
ax = 2 * MPFR_GET_EXP (b);
bq = MPFR_PREC(b);
MPFR_ASSERTD (2 * bq > bq); /* PREC_MAX is /2 so no integer overflow */
bn = MPFR_LIMB_SIZE(b); /* number of limbs of b */
tn = 1 + (2 * bq - 1) / GMP_NUMB_BITS; /* number of limbs of square,
2*bn or 2*bn-1 */
MPFR_TMP_MARK(marker);
tmp = (mp_limb_t *) MPFR_TMP_ALLOC((size_t) 2 * bn * BYTES_PER_MP_LIMB);
/* Multiplies the mantissa in temporary allocated space */
mpn_sqr_n (tmp, MPFR_MANT(b), bn);
b1 = tmp[2 * bn - 1];
/* now tmp[0]..tmp[2*bn-1] contains the product of both mantissa,
with tmp[2*bn-1]>=2^(GMP_NUMB_BITS-2) */
b1 >>= GMP_NUMB_BITS - 1; /* msb from the product */
/* if the mantissas of b and c are uniformly distributed in ]1/2, 1],
then their product is in ]1/4, 1/2] with probability 2*ln(2)-1 ~ 0.386
and in [1/2, 1] with probability 2-2*ln(2) ~ 0.614 */
tmp += 2 * bn - tn; /* +0 or +1 */
if (MPFR_UNLIKELY(b1 == 0))
mpn_lshift (tmp, tmp, tn, 1); /* tn <= k, so no stack corruption */
cc = mpfr_round_raw (MPFR_MANT (a), tmp, 2 * bq, 0,
MPFR_PREC (a), rnd_mode, &inexact);
/* cc = 1 ==> result is a power of two */
if (MPFR_UNLIKELY(cc))
MPFR_MANT(a)[MPFR_LIMB_SIZE(a)-1] = MPFR_LIMB_HIGHBIT;
MPFR_TMP_FREE(marker);
{
mpfr_exp_t ax2 = ax + (mpfr_exp_t) (b1 - 1 + cc);
if (MPFR_UNLIKELY( ax2 > __gmpfr_emax))
return mpfr_overflow (a, rnd_mode, MPFR_SIGN_POS);
if (MPFR_UNLIKELY( ax2 < __gmpfr_emin))
{
/* In the rounding to the nearest mode, if the exponent of the exact
result (i.e. before rounding, i.e. without taking cc into account)
is < __gmpfr_emin - 1 or the exact result is a power of 2 (i.e. if
both arguments are powers of 2), then round to zero. */
if (rnd_mode == MPFR_RNDN &&
(ax + (mpfr_exp_t) b1 < __gmpfr_emin || mpfr_powerof2_raw (b)))
rnd_mode = MPFR_RNDZ;
return mpfr_underflow (a, rnd_mode, MPFR_SIGN_POS);
}
MPFR_SET_EXP (a, ax2);
MPFR_SET_POS (a);
}
MPFR_RET (inexact);
}
int
mpfr_sinh_cosh (mpfr_ptr sh, mpfr_ptr ch, mpfr_srcptr xt, mpfr_rnd_t rnd_mode)
{
mpfr_t x;
int inexact_sh, inexact_ch;
MPFR_ASSERTN (sh != ch);
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg rnd=%d",
mpfr_get_prec (xt), mpfr_log_prec, xt, rnd_mode),
("sh[%Pu]=%.*Rg ch[%Pu]=%.*Rg",
mpfr_get_prec (sh), mpfr_log_prec, sh,
mpfr_get_prec (ch), mpfr_log_prec, ch));
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (xt)))
{
if (MPFR_IS_NAN (xt))
{
MPFR_SET_NAN (ch);
MPFR_SET_NAN (sh);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (xt))
{
MPFR_SET_INF (sh);
MPFR_SET_SAME_SIGN (sh, xt);
MPFR_SET_INF (ch);
MPFR_SET_POS (ch);
MPFR_RET (0);
}
else /* xt is zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO (xt));
MPFR_SET_ZERO (sh); /* sinh(0) = 0 */
MPFR_SET_SAME_SIGN (sh, xt);
inexact_sh = 0;
inexact_ch = mpfr_set_ui (ch, 1, rnd_mode); /* cosh(0) = 1 */
return INEX(inexact_sh,inexact_ch);
}
}
/* Warning: if we use MPFR_FAST_COMPUTE_IF_SMALL_INPUT here, make sure
that the code also works in case of overlap (see sin_cos.c) */
MPFR_TMP_INIT_ABS (x, xt);
{
mpfr_t s, c, ti;
mpfr_exp_t d;
mpfr_prec_t N; /* Precision of the intermediary variables */
long int err; /* Precision of error */
MPFR_ZIV_DECL (loop);
MPFR_SAVE_EXPO_DECL (expo);
MPFR_GROUP_DECL (group);
MPFR_SAVE_EXPO_MARK (expo);
/* compute the precision of intermediary variable */
N = MPFR_PREC (ch);
N = MAX (N, MPFR_PREC (sh));
/* the optimal number of bits : see algorithms.ps */
N = N + MPFR_INT_CEIL_LOG2 (N) + 4;
/* initialise of intermediary variables */
MPFR_GROUP_INIT_3 (group, N, s, c, ti);
/* First computation of sinh_cosh */
MPFR_ZIV_INIT (loop, N);
for (;;)
{
MPFR_BLOCK_DECL (flags);
/* compute sinh_cosh */
MPFR_BLOCK (flags, mpfr_exp (s, x, MPFR_RNDD));
if (MPFR_OVERFLOW (flags))
/* exp(x) does overflow */
{
/* since cosh(x) >= exp(x), cosh(x) overflows too */
inexact_ch = mpfr_overflow (ch, rnd_mode, MPFR_SIGN_POS);
/* sinh(x) may be representable */
inexact_sh = mpfr_sinh (sh, xt, rnd_mode);
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_OVERFLOW);
break;
}
d = MPFR_GET_EXP (s);
mpfr_ui_div (ti, 1, s, MPFR_RNDU); /* 1/exp(x) */
mpfr_add (c, s, ti, MPFR_RNDU); /* exp(x) + 1/exp(x) */
mpfr_sub (s, s, ti, MPFR_RNDN); /* exp(x) - 1/exp(x) */
mpfr_div_2ui (c, c, 1, MPFR_RNDN); /* 1/2(exp(x) + 1/exp(x)) */
mpfr_div_2ui (s, s, 1, MPFR_RNDN); /* 1/2(exp(x) - 1/exp(x)) */
/* it may be that s is zero (in fact, it can only occur when exp(x)=1,
and thus ti=1 too) */
if (MPFR_IS_ZERO (s))
err = N; /* double the precision */
else
{
/* calculation of the error */
d = d - MPFR_GET_EXP (s) + 2;
/* error estimate: err = N-(__gmpfr_ceil_log2(1+pow(2,d)));*/
err = N - (MAX (d, 0) + 1);
if (MPFR_LIKELY (MPFR_CAN_ROUND (s, err, MPFR_PREC (sh),
rnd_mode) && \
MPFR_CAN_ROUND (c, err, MPFR_PREC (ch),
rnd_mode)))
{
inexact_sh = mpfr_set4 (sh, s, rnd_mode, MPFR_SIGN (xt));
inexact_ch = mpfr_set (ch, c, rnd_mode);
break;
}
}
/* actualisation of the precision */
N += err;
MPFR_ZIV_NEXT (loop, N);
MPFR_GROUP_REPREC_3 (group, N, s, c, ti);
}
MPFR_ZIV_FREE (loop);
MPFR_GROUP_CLEAR (group);
MPFR_SAVE_EXPO_FREE (expo);
}
/* now, let's raise the flags if needed */
inexact_sh = mpfr_check_range (sh, inexact_sh, rnd_mode);
inexact_ch = mpfr_check_range (ch, inexact_ch, rnd_mode);
return INEX(inexact_sh,inexact_ch);
}
static int
mpfr_mul3 (mpfr_ptr a, mpfr_srcptr b, mpfr_srcptr c, mpfr_rnd_t rnd_mode)
{
/* Old implementation */
int sign_product, cc, inexact;
mpfr_exp_t ax;
mp_limb_t *tmp;
mp_limb_t b1;
mpfr_prec_t bq, cq;
mp_size_t bn, cn, tn, k;
MPFR_TMP_DECL(marker);
/* deal with special cases */
if (MPFR_ARE_SINGULAR(b,c))
{
if (MPFR_IS_NAN(b) || MPFR_IS_NAN(c))
{
MPFR_SET_NAN(a);
MPFR_RET_NAN;
}
sign_product = MPFR_MULT_SIGN( MPFR_SIGN(b) , MPFR_SIGN(c) );
if (MPFR_IS_INF(b))
{
if (MPFR_IS_INF(c) || MPFR_NOTZERO(c))
{
MPFR_SET_SIGN(a,sign_product);
MPFR_SET_INF(a);
MPFR_RET(0); /* exact */
}
else
{
MPFR_SET_NAN(a);
MPFR_RET_NAN;
}
}
else if (MPFR_IS_INF(c))
{
if (MPFR_NOTZERO(b))
{
MPFR_SET_SIGN(a, sign_product);
MPFR_SET_INF(a);
MPFR_RET(0); /* exact */
}
else
{
MPFR_SET_NAN(a);
MPFR_RET_NAN;
}
}
else
{
MPFR_ASSERTD(MPFR_IS_ZERO(b) || MPFR_IS_ZERO(c));
MPFR_SET_SIGN(a, sign_product);
MPFR_SET_ZERO(a);
MPFR_RET(0); /* 0 * 0 is exact */
}
}
sign_product = MPFR_MULT_SIGN( MPFR_SIGN(b) , MPFR_SIGN(c) );
ax = MPFR_GET_EXP (b) + MPFR_GET_EXP (c);
bq = MPFR_PREC(b);
cq = MPFR_PREC(c);
MPFR_ASSERTD(bq+cq > bq); /* PREC_MAX is /2 so no integer overflow */
bn = (bq+GMP_NUMB_BITS-1)/GMP_NUMB_BITS; /* number of limbs of b */
cn = (cq+GMP_NUMB_BITS-1)/GMP_NUMB_BITS; /* number of limbs of c */
k = bn + cn; /* effective nb of limbs used by b*c (= tn or tn+1) below */
tn = (bq + cq + GMP_NUMB_BITS - 1) / GMP_NUMB_BITS;
/* <= k, thus no int overflow */
MPFR_ASSERTD(tn <= k);
/* Check for no size_t overflow*/
MPFR_ASSERTD((size_t) k <= ((size_t) -1) / BYTES_PER_MP_LIMB);
MPFR_TMP_MARK(marker);
tmp = (mp_limb_t *) MPFR_TMP_ALLOC((size_t) k * BYTES_PER_MP_LIMB);
/* multiplies two mantissa in temporary allocated space */
b1 = (MPFR_LIKELY(bn >= cn)) ?
mpn_mul (tmp, MPFR_MANT(b), bn, MPFR_MANT(c), cn)
: mpn_mul (tmp, MPFR_MANT(c), cn, MPFR_MANT(b), bn);
/* now tmp[0]..tmp[k-1] contains the product of both mantissa,
with tmp[k-1]>=2^(GMP_NUMB_BITS-2) */
b1 >>= GMP_NUMB_BITS - 1; /* msb from the product */
/* if the mantissas of b and c are uniformly distributed in ]1/2, 1],
then their product is in ]1/4, 1/2] with probability 2*ln(2)-1 ~ 0.386
and in [1/2, 1] with probability 2-2*ln(2) ~ 0.614 */
tmp += k - tn;
if (MPFR_UNLIKELY(b1 == 0))
mpn_lshift (tmp, tmp, tn, 1); /* tn <= k, so no stack corruption */
cc = mpfr_round_raw (MPFR_MANT (a), tmp, bq + cq,
MPFR_IS_NEG_SIGN(sign_product),
MPFR_PREC (a), rnd_mode, &inexact);
/* cc = 1 ==> result is a power of two */
if (MPFR_UNLIKELY(cc))
MPFR_MANT(a)[MPFR_LIMB_SIZE(a)-1] = MPFR_LIMB_HIGHBIT;
MPFR_TMP_FREE(marker);
{
mpfr_exp_t ax2 = ax + (mpfr_exp_t) (b1 - 1 + cc);
if (MPFR_UNLIKELY( ax2 > __gmpfr_emax))
return mpfr_overflow (a, rnd_mode, sign_product);
if (MPFR_UNLIKELY( ax2 < __gmpfr_emin))
{
/* In the rounding to the nearest mode, if the exponent of the exact
result (i.e. before rounding, i.e. without taking cc into account)
is < __gmpfr_emin - 1 or the exact result is a power of 2 (i.e. if
both arguments are powers of 2), then round to zero. */
if (rnd_mode == MPFR_RNDN &&
(ax + (mpfr_exp_t) b1 < __gmpfr_emin ||
(mpfr_powerof2_raw (b) && mpfr_powerof2_raw (c))))
rnd_mode = MPFR_RNDZ;
return mpfr_underflow (a, rnd_mode, sign_product);
}
MPFR_SET_EXP (a, ax2);
MPFR_SET_SIGN(a, sign_product);
}
MPFR_RET (inexact);
}
int
mpfr_mul_ui (mpfr_ptr y, mpfr_srcptr x, unsigned long int u, mpfr_rnd_t rnd_mode)
{
mp_limb_t *yp;
mp_size_t xn;
int cnt, inexact;
MPFR_TMP_DECL (marker);
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 (u != 0)
{
MPFR_SET_INF (y);
MPFR_SET_SAME_SIGN (y, x);
MPFR_RET (0); /* infinity is exact */
}
else /* 0 * infinity */
{
MPFR_SET_NAN (y);
MPFR_RET_NAN;
}
}
else /* x is zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO (x));
MPFR_SET_ZERO (y);
MPFR_SET_SAME_SIGN (y, x);
MPFR_RET (0); /* zero is exact */
}
}
else if (MPFR_UNLIKELY (u <= 1))
{
if (u < 1)
{
MPFR_SET_ZERO (y);
MPFR_SET_SAME_SIGN (y, x);
MPFR_RET (0); /* zero is exact */
}
else
return mpfr_set (y, x, rnd_mode);
}
else if (MPFR_UNLIKELY (IS_POW2 (u)))
return mpfr_mul_2si (y, x, MPFR_INT_CEIL_LOG2 (u), rnd_mode);
yp = MPFR_MANT (y);
xn = MPFR_LIMB_SIZE (x);
MPFR_ASSERTD (xn < MP_SIZE_T_MAX);
MPFR_TMP_MARK(marker);
yp = MPFR_TMP_LIMBS_ALLOC (xn + 1);
MPFR_ASSERTN (u == (mp_limb_t) u);
yp[xn] = mpn_mul_1 (yp, MPFR_MANT (x), xn, u);
/* x * u is stored in yp[xn], ..., yp[0] */
/* since the case u=1 was treated above, we have u >= 2, thus
yp[xn] >= 1 since x was msb-normalized */
MPFR_ASSERTD (yp[xn] != 0);
if (MPFR_LIKELY (MPFR_LIMB_MSB (yp[xn]) == 0))
{
count_leading_zeros (cnt, yp[xn]);
mpn_lshift (yp, yp, xn + 1, cnt);
}
else
{
cnt = 0;
}
/* now yp[xn], ..., yp[0] is msb-normalized too, and has at most
PREC(x) + (GMP_NUMB_BITS - cnt) non-zero bits */
MPFR_RNDRAW (inexact, y, yp, (mpfr_prec_t) (xn + 1) * GMP_NUMB_BITS,
rnd_mode, MPFR_SIGN (x), cnt -- );
MPFR_TMP_FREE (marker);
cnt = GMP_NUMB_BITS - cnt;
if (MPFR_UNLIKELY (__gmpfr_emax < MPFR_EMAX_MIN + cnt
|| MPFR_GET_EXP (x) > __gmpfr_emax - cnt))
return mpfr_overflow (y, rnd_mode, MPFR_SIGN(x));
MPFR_SET_EXP (y, MPFR_GET_EXP (x) + cnt);
MPFR_SET_SAME_SIGN (y, x);
return inexact;
}
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
mpfr_sinh (mpfr_ptr y, mpfr_srcptr xt, mpfr_rnd_t rnd_mode)
{
mpfr_t x;
int inexact;
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg rnd=%d", mpfr_get_prec (xt), mpfr_log_prec, xt, rnd_mode),
("y[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (y), mpfr_log_prec, y, inexact));
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (xt)))
{
if (MPFR_IS_NAN (xt))
{
MPFR_SET_NAN (y);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (xt))
{
MPFR_SET_INF (y);
MPFR_SET_SAME_SIGN (y, xt);
MPFR_RET (0);
}
else /* xt is zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO (xt));
MPFR_SET_ZERO (y); /* sinh(0) = 0 */
MPFR_SET_SAME_SIGN (y, xt);
MPFR_RET (0);
}
}
/* sinh(x) = x + x^3/6 + ... so the error is < 2^(3*EXP(x)-2) */
MPFR_FAST_COMPUTE_IF_SMALL_INPUT (y, xt, -2 * MPFR_GET_EXP(xt), 2, 1,
rnd_mode, {});
MPFR_TMP_INIT_ABS (x, xt);
{
mpfr_t t, ti;
mpfr_exp_t d;
mpfr_prec_t Nt; /* Precision of the intermediary variable */
long int err; /* Precision of error */
MPFR_ZIV_DECL (loop);
MPFR_SAVE_EXPO_DECL (expo);
MPFR_GROUP_DECL (group);
MPFR_SAVE_EXPO_MARK (expo);
/* compute the precision of intermediary variable */
Nt = MAX (MPFR_PREC (x), MPFR_PREC (y));
/* the optimal number of bits : see algorithms.ps */
Nt = Nt + MPFR_INT_CEIL_LOG2 (Nt) + 4;
/* If x is near 0, exp(x) - 1/exp(x) = 2*x+x^3/3+O(x^5) */
if (MPFR_GET_EXP (x) < 0)
Nt -= 2*MPFR_GET_EXP (x);
/* initialise of intermediary variables */
MPFR_GROUP_INIT_2 (group, Nt, t, ti);
/* First computation of sinh */
MPFR_ZIV_INIT (loop, Nt);
for (;;)
{
MPFR_BLOCK_DECL (flags);
/* compute sinh */
MPFR_BLOCK (flags, mpfr_exp (t, x, MPFR_RNDD));
if (MPFR_OVERFLOW (flags))
/* exp(x) does overflow */
{
/* sinh(x) = 2 * sinh(x/2) * cosh(x/2) */
mpfr_div_2ui (ti, x, 1, MPFR_RNDD); /* exact */
/* t <- cosh(x/2): error(t) <= 1 ulp(t) */
MPFR_BLOCK (flags, mpfr_cosh (t, ti, MPFR_RNDD));
if (MPFR_OVERFLOW (flags))
/* when x>1 we have |sinh(x)| >= cosh(x/2), so sinh(x)
overflows too */
{
inexact = mpfr_overflow (y, rnd_mode, MPFR_SIGN (xt));
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_OVERFLOW);
break;
}
/* ti <- sinh(x/2): , error(ti) <= 1 ulp(ti)
cannot overflow because 0 < sinh(x) < cosh(x) when x > 0 */
mpfr_sinh (ti, ti, MPFR_RNDD);
/* multiplication below, error(t) <= 5 ulp(t) */
MPFR_BLOCK (flags, mpfr_mul (t, t, ti, MPFR_RNDD));
if (MPFR_OVERFLOW (flags))
{
inexact = mpfr_overflow (y, rnd_mode, MPFR_SIGN (xt));
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_OVERFLOW);
break;
}
/* doubling below, exact */
MPFR_BLOCK (flags, mpfr_mul_2ui (t, t, 1, MPFR_RNDN));
if (MPFR_OVERFLOW (flags))
{
inexact = mpfr_overflow (y, rnd_mode, MPFR_SIGN (xt));
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_OVERFLOW);
break;
}
/* we have lost at most 3 bits of precision */
err = Nt - 3;
if (MPFR_LIKELY (MPFR_CAN_ROUND (t, err, MPFR_PREC (y),
rnd_mode)))
{
inexact = mpfr_set4 (y, t, rnd_mode, MPFR_SIGN (xt));
break;
}
err = Nt; /* double the precision */
}
else
{
d = MPFR_GET_EXP (t);
mpfr_ui_div (ti, 1, t, MPFR_RNDU); /* 1/exp(x) */
mpfr_sub (t, t, ti, MPFR_RNDN); /* exp(x) - 1/exp(x) */
mpfr_div_2ui (t, t, 1, MPFR_RNDN); /* 1/2(exp(x) - 1/exp(x)) */
/* it may be that t is zero (in fact, it can only occur when te=1,
and thus ti=1 too) */
if (MPFR_IS_ZERO (t))
err = Nt; /* double the precision */
else
{
/* calculation of the error */
d = d - MPFR_GET_EXP (t) + 2;
/* error estimate: err = Nt-(__gmpfr_ceil_log2(1+pow(2,d)));*/
err = Nt - (MAX (d, 0) + 1);
if (MPFR_LIKELY (MPFR_CAN_ROUND (t, err, MPFR_PREC (y),
rnd_mode)))
{
inexact = mpfr_set4 (y, t, rnd_mode, MPFR_SIGN (xt));
break;
}
}
}
/* actualisation of the precision */
Nt += err;
MPFR_ZIV_NEXT (loop, Nt);
MPFR_GROUP_REPREC_2 (group, Nt, t, ti);
}
MPFR_ZIV_FREE (loop);
MPFR_GROUP_CLEAR (group);
MPFR_SAVE_EXPO_FREE (expo);
}
return mpfr_check_range (y, inexact, rnd_mode);
}
int
mpfr_eint (mpfr_ptr y, mpfr_srcptr x, mp_rnd_t rnd)
{
int inex;
mpfr_t tmp, ump;
mp_exp_t err, te;
mp_prec_t prec;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_ZIV_DECL (loop);
MPFR_LOG_FUNC (("x[%#R]=%R rnd=%d", x, x, rnd),
("y[%#R]=%R inexact=%d", y, y, inex));
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x)))
{
/* exp(NaN) = exp(-Inf) = NaN */
if (MPFR_IS_NAN (x) || (MPFR_IS_INF (x) && MPFR_IS_NEG(x)))
{
MPFR_SET_NAN (y);
MPFR_RET_NAN;
}
/* eint(+inf) = +inf */
else if (MPFR_IS_INF (x))
{
MPFR_SET_INF(y);
MPFR_SET_POS(y);
MPFR_RET(0);
}
else /* eint(+/-0) = -Inf */
{
MPFR_SET_INF(y);
MPFR_SET_NEG(y);
MPFR_RET(0);
}
}
/* eint(x) = NaN for x < 0 */
if (MPFR_IS_NEG(x))
{
MPFR_SET_NAN (y);
MPFR_RET_NAN;
}
MPFR_SAVE_EXPO_MARK (expo);
/* Since eint(x) >= exp(x)/x, we have log2(eint(x)) >= (x-log(x))/log(2).
Let's compute k <= (x-log(x))/log(2) in a low precision. If k >= emax,
then log2(eint(x)) >= emax, and eint(x) >= 2^emax, i.e. it overflows. */
mpfr_init2 (tmp, 64);
mpfr_init2 (ump, 64);
mpfr_log (tmp, x, GMP_RNDU);
mpfr_sub (ump, x, tmp, GMP_RNDD);
mpfr_const_log2 (tmp, GMP_RNDU);
mpfr_div (ump, ump, tmp, GMP_RNDD);
/* FIXME: We really need mpfr_set_exp_t and mpfr_cmp_exp_t functions. */
MPFR_ASSERTN (MPFR_EMAX_MAX <= LONG_MAX);
if (mpfr_cmp_ui (ump, __gmpfr_emax) >= 0)
{
mpfr_clear (tmp);
mpfr_clear (ump);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_overflow (y, rnd, 1);
}
/* Init stuff */
prec = MPFR_PREC (y) + 2 * MPFR_INT_CEIL_LOG2 (MPFR_PREC (y)) + 6;
/* eint() has a root 0.37250741078136663446..., so if x is near,
already take more bits */
if (MPFR_GET_EXP(x) == -1) /* 1/4 <= x < 1/2 */
{
double d;
d = mpfr_get_d (x, GMP_RNDN) - 0.37250741078136663;
d = (d == 0.0) ? -53 : __gmpfr_ceil_log2 (d);
prec += -d;
}
mpfr_set_prec (tmp, prec);
mpfr_set_prec (ump, prec);
MPFR_ZIV_INIT (loop, prec); /* Initialize the ZivLoop controler */
for (;;) /* Infinite loop */
{
/* We need that the smallest value of k!/x^k is smaller than 2^(-p).
The minimum is obtained for x=k, and it is smaller than e*sqrt(x)/e^x
for x>=1. */
if (MPFR_GET_EXP (x) > 0 && mpfr_cmp_d (x, ((double) prec +
0.5 * (double) MPFR_GET_EXP (x)) * LOG2 + 1.0) > 0)
err = mpfr_eint_asympt (tmp, x);
else
{
err = mpfr_eint_aux (tmp, x); /* error <= 2^err ulp(tmp) */
te = MPFR_GET_EXP(tmp);
mpfr_const_euler (ump, GMP_RNDN); /* 0.577 -> EXP(ump)=0 */
mpfr_add (tmp, tmp, ump, GMP_RNDN);
/* error <= 1/2 + 1/2*2^(EXP(ump)-EXP(tmp)) + 2^(te-EXP(tmp)+err)
<= 1/2 + 2^(MAX(EXP(ump), te+err+1) - EXP(tmp))
<= 2^(MAX(0, 1 + MAX(EXP(ump), te+err+1) - EXP(tmp))) */
err = MAX(1, te + err + 2) - MPFR_GET_EXP(tmp);
err = MAX(0, err);
te = MPFR_GET_EXP(tmp);
mpfr_log (ump, x, GMP_RNDN);
mpfr_add (tmp, tmp, ump, GMP_RNDN);
/* same formula as above, except now EXP(ump) is not 0 */
err += te + 1;
if (MPFR_LIKELY (!MPFR_IS_ZERO (ump)))
err = MAX (MPFR_GET_EXP (ump), err);
err = MAX(0, err - MPFR_GET_EXP (tmp));
}
if (MPFR_LIKELY (MPFR_CAN_ROUND (tmp, prec - err, MPFR_PREC (y), rnd)))
break;
MPFR_ZIV_NEXT (loop, prec); /* Increase used precision */
mpfr_set_prec (tmp, prec);
mpfr_set_prec (ump, prec);
}
MPFR_ZIV_FREE (loop); /* Free the ZivLoop Controler */
inex = mpfr_set (y, tmp, rnd); /* Set y to the computed value */
mpfr_clear (tmp);
mpfr_clear (ump);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (y, inex, rnd);
}
/* 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_set_uj_2exp (mpfr_t x, uintmax_t j, intmax_t e, mp_rnd_t rnd)
{
unsigned int cnt, i;
mp_size_t k, len;
mp_limb_t limb;
mp_limb_t yp[sizeof(uintmax_t) / sizeof(mp_limb_t)];
mpfr_t y;
unsigned long uintmax_bit_size = sizeof(uintmax_t) * CHAR_BIT;
unsigned long bpml = BITS_PER_MP_LIMB % uintmax_bit_size;
/* Special case */
if (j == 0)
{
MPFR_SET_POS(x);
MPFR_SET_ZERO(x);
MPFR_RET(0);
}
MPFR_ASSERTN (sizeof(uintmax_t) % sizeof(mp_limb_t) == 0);
/* Create an auxillary var */
MPFR_TMP_INIT1 (yp, y, uintmax_bit_size);
k = numberof (yp);
if (k == 1)
limb = yp[0] = j;
else
{
/* Note: either BITS_PER_MP_LIMB = uintmax_bit_size, then k = 1 the
shift j >>= bpml is never done, or BITS_PER_MP_LIMB < uintmax_bit_size
and bpml = BITS_PER_MP_LIMB. */
for (i = 0; i < k; i++, j >>= bpml)
yp[i] = j; /* Only the low bits are copied */
/* Find the first limb not equal to zero. */
do
{
MPFR_ASSERTD (k > 0);
limb = yp[--k];
}
while (limb == 0);
k++;
}
count_leading_zeros(cnt, limb);
len = numberof (yp) - k;
/* Normalize it: len = number of last 0 limb, k number of non-zero limbs */
if (MPFR_LIKELY(cnt))
mpn_lshift (yp+len, yp, k, cnt); /* Normalize the High Limb*/
else if (len != 0)
MPN_COPY_DECR (yp+len, yp, k); /* Must use DECR */
if (len != 0)
/* Note: when numberof(yp)==1, len is constant and null, so the compiler
can optimize out this code. */
{
if (len == 1)
yp[0] = (mp_limb_t) 0;
else
MPN_ZERO (yp, len); /* Zeroing the last limbs */
}
e += k * BITS_PER_MP_LIMB - cnt; /* Update Expo */
MPFR_ASSERTD (MPFR_LIMB_MSB(yp[numberof (yp) - 1]) != 0);
/* Check expo underflow / overflow (can't use mpfr_check_range) */
if (MPFR_UNLIKELY(e < __gmpfr_emin))
{
/* The following test is necessary because in the rounding to the
* nearest mode, mpfr_underflow always rounds away from 0. In
* this rounding mode, we need to round to 0 if:
* _ |x| < 2^(emin-2), or
* _ |x| = 2^(emin-2) and the absolute value of the exact
* result is <= 2^(emin-2). */
if (rnd == GMP_RNDN && (e+1 < __gmpfr_emin || mpfr_powerof2_raw(y)))
rnd = GMP_RNDZ;
return mpfr_underflow (x, rnd, MPFR_SIGN_POS);
}
if (MPFR_UNLIKELY(e > __gmpfr_emax))
return mpfr_overflow (x, rnd, MPFR_SIGN_POS);
MPFR_SET_EXP (y, e);
/* Final: set x to y (rounding if necessary) */
return mpfr_set (x, y, rnd);
}
