/* This function does not change the flags. */
static void
mpfr_get_zexp (mpz_ptr ez, mpfr_srcptr x)
{
mpz_init (ez);
if (MPFR_IS_UBF (x))
mpz_set (ez, MPFR_ZEXP (x));
else
{
mp_limb_t e_limb[MPFR_EXP_LIMB_SIZE];
mpfr_t e;
int inex;
MPFR_SAVE_EXPO_DECL (expo);
/* TODO: Once this has been tested, optimize based on whether
_MPFR_EXP_FORMAT <= 3. */
MPFR_TMP_INIT1 (e_limb, e, sizeof (mpfr_exp_t) * CHAR_BIT);
MPFR_SAVE_EXPO_MARK (expo);
MPFR_DBGRES (inex = mpfr_set_exp_t (e, MPFR_GET_EXP (x), MPFR_RNDN));
MPFR_ASSERTD (inex == 0);
MPFR_DBGRES (inex = mpfr_get_z (ez, e, MPFR_RNDN));
MPFR_ASSERTD (inex == 0);
MPFR_SAVE_EXPO_FREE (expo);
}
}
static void
underflow_up3 (void)
{
mpfr_t x, y, z, z0;
int inex;
int rnd;
int i;
mpfr_init2 (x, 64);
mpfr_init2 (y, sizeof (mpfr_exp_t) * CHAR_BIT);
mpfr_init2 (z, 32);
mpfr_init2 (z0, 2);
inex = mpfr_set_exp_t (y, mpfr_get_emin () - 2, MPFR_RNDN);
MPFR_ASSERTN (inex == 0);
for (i = -1; i <= 1; i++)
RND_LOOP (rnd)
{
unsigned int ufinex = MPFR_FLAGS_UNDERFLOW | MPFR_FLAGS_INEXACT;
int expected_inex;
mpfr_set_ui (x, 2, MPFR_RNDN);
if (i < 0)
mpfr_nextbelow (x);
if (i > 0)
mpfr_nextabove (x);
/* x = 2 + i * eps, y = emin - 2, x^y ~= 2^(emin - 2) */
expected_inex = rnd == MPFR_RNDU || rnd == MPFR_RNDA
|| (rnd == MPFR_RNDN && i < 0) ? 1 : -1;
mpfr_set_ui (z0, 0, MPFR_RNDN);
if (expected_inex > 0)
mpfr_nextabove (z0);
mpfr_clear_flags ();
inex = mpfr_pow (z, x, y, (mpfr_rnd_t) rnd);
cmpres (0, x, "emin - 2", (mpfr_rnd_t) rnd, z0, expected_inex, z, inex,
ufinex, "underflow_up3", "mpfr_pow");
test_others (NULL, "emin - 2", (mpfr_rnd_t) rnd, x, y, z, inex, ufinex,
"underflow_up3");
}
mpfr_clears (x, y, z, z0, (mpfr_ptr) 0);
}
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);
}
/* 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);
}