double
mpfr_get_d (mpfr_srcptr src, mpfr_rnd_t rnd_mode)
{
double d;
int negative;
mpfr_exp_t e;
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (src)))
{
if (MPFR_IS_NAN (src))
return MPFR_DBL_NAN;
negative = MPFR_IS_NEG (src);
if (MPFR_IS_INF (src))
return negative ? MPFR_DBL_INFM : MPFR_DBL_INFP;
MPFR_ASSERTD (MPFR_IS_ZERO(src));
return negative ? DBL_NEG_ZERO : 0.0;
}
e = MPFR_GET_EXP (src);
negative = MPFR_IS_NEG (src);
if (MPFR_UNLIKELY(rnd_mode == MPFR_RNDA))
rnd_mode = negative ? MPFR_RNDD : MPFR_RNDU;
/* the smallest normalized number is 2^(-1022)=0.1e-1021, and the smallest
subnormal is 2^(-1074)=0.1e-1073 */
if (MPFR_UNLIKELY (e < -1073))
{
/* Note: Avoid using a constant expression DBL_MIN * DBL_EPSILON
as this gives 0 instead of the correct result with gcc on some
Alpha machines. */
d = negative ?
(rnd_mode == MPFR_RNDD ||
(rnd_mode == MPFR_RNDN && mpfr_cmp_si_2exp(src, -1, -1075) < 0)
? -DBL_MIN : DBL_NEG_ZERO) :
(rnd_mode == MPFR_RNDU ||
(rnd_mode == MPFR_RNDN && mpfr_cmp_si_2exp(src, 1, -1075) > 0)
? DBL_MIN : 0.0);
if (d != 0.0) /* we multiply DBL_MIN = 2^(-1022) by DBL_EPSILON = 2^(-52)
to get +-2^(-1074) */
d *= DBL_EPSILON;
}
/* the largest normalized number is 2^1024*(1-2^(-53))=0.111...111e1024 */
else if (MPFR_UNLIKELY (e > 1024))
{
d = negative ?
(rnd_mode == MPFR_RNDZ || rnd_mode == MPFR_RNDU ?
-DBL_MAX : MPFR_DBL_INFM) :
(rnd_mode == MPFR_RNDZ || rnd_mode == MPFR_RNDD ?
DBL_MAX : MPFR_DBL_INFP);
}
else
{
int nbits;
mp_size_t np, i;
mp_limb_t tp[ MPFR_LIMBS_PER_DOUBLE ];
int carry;
nbits = IEEE_DBL_MANT_DIG; /* 53 */
if (MPFR_UNLIKELY (e < -1021))
/*In the subnormal case, compute the exact number of significant bits*/
{
nbits += (1021 + e);
MPFR_ASSERTD (nbits >= 1);
}
np = MPFR_PREC2LIMBS (nbits);
MPFR_ASSERTD ( np <= MPFR_LIMBS_PER_DOUBLE );
carry = mpfr_round_raw_4 (tp, MPFR_MANT(src), MPFR_PREC(src), negative,
nbits, rnd_mode);
if (MPFR_UNLIKELY(carry))
d = 1.0;
else
{
/* The following computations are exact thanks to the previous
mpfr_round_raw. */
d = (double) tp[0] / MP_BASE_AS_DOUBLE;
for (i = 1 ; i < np ; i++)
d = (d + tp[i]) / MP_BASE_AS_DOUBLE;
/* d is the mantissa (between 1/2 and 1) of the argument rounded
to 53 bits */
}
d = mpfr_scale2 (d, e);
if (negative)
d = -d;
}
return d;
}
float
mpfr_get_flt (mpfr_srcptr src, mpfr_rnd_t rnd_mode)
{
int negative;
mpfr_exp_t e;
float d;
/* in case of NaN, +Inf, -Inf, +0, -0, the conversion from double to float
is exact */
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (src)))
return (float) mpfr_get_d (src, rnd_mode);
e = MPFR_GET_EXP (src);
negative = MPFR_IS_NEG (src);
if (MPFR_UNLIKELY(rnd_mode == MPFR_RNDA))
rnd_mode = negative ? MPFR_RNDD : MPFR_RNDU;
/* the smallest positive normal float number is 2^(-126) = 0.5*2^(-125),
and the smallest positive subnormal number is 2^(-149) = 0.5*2^(-148) */
if (MPFR_UNLIKELY (e < -148))
{
/* |src| < 2^(-149), i.e., |src| is smaller than the smallest positive
subnormal number.
In round-to-nearest mode, 2^(-150) is rounded to zero.
*/
d = negative ?
(rnd_mode == MPFR_RNDD ||
(rnd_mode == MPFR_RNDN && mpfr_cmp_si_2exp (src, -1, -150) < 0)
? -FLT_MIN : FLT_NEG_ZERO) :
(rnd_mode == MPFR_RNDU ||
(rnd_mode == MPFR_RNDN && mpfr_cmp_si_2exp (src, 1, -150) > 0)
? FLT_MIN : 0.0);
if (d != 0.0) /* we multiply FLT_MIN = 2^(-126) by FLT_EPSILON = 2^(-23)
to get +-2^(-149) */
d *= FLT_EPSILON;
}
/* the largest normal number is 2^128*(1-2^(-24)) = 0.111...111e128 */
else if (MPFR_UNLIKELY (e > 128))
{
d = negative ?
(rnd_mode == MPFR_RNDZ || rnd_mode == MPFR_RNDU ?
-FLT_MAX : MPFR_FLT_INFM) :
(rnd_mode == MPFR_RNDZ || rnd_mode == MPFR_RNDD ?
FLT_MAX : MPFR_FLT_INFP);
}
else /* -148 <= e <= 127 */
{
int nbits;
mp_size_t np, i;
mp_limb_t tp[MPFR_LIMBS_PER_FLT];
int carry;
double dd;
nbits = IEEE_FLT_MANT_DIG; /* 24 */
if (MPFR_UNLIKELY (e < -125))
/*In the subnormal case, compute the exact number of significant bits*/
{
nbits += (125 + e);
MPFR_ASSERTD (nbits >= 1);
}
np = MPFR_PREC2LIMBS (nbits);
MPFR_ASSERTD(np <= MPFR_LIMBS_PER_FLT);
carry = mpfr_round_raw_4 (tp, MPFR_MANT(src), MPFR_PREC(src), negative,
nbits, rnd_mode);
/* we perform the reconstruction using the 'double' type here,
knowing the result is exactly representable as 'float' */
if (MPFR_UNLIKELY(carry))
dd = 1.0;
else
{
/* The following computations are exact thanks to the previous
mpfr_round_raw. */
dd = (double) tp[0] / MP_BASE_AS_DOUBLE;
for (i = 1 ; i < np ; i++)
dd = (dd + tp[i]) / MP_BASE_AS_DOUBLE;
/* dd is the mantissa (between 1/2 and 1) of the argument rounded
to 24 bits */
}
dd = mpfr_scale2 (dd, e);
if (negative)
dd = -dd;
/* convert (exacly) to float */
d = (float) dd;
}
return d;
}
