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);
}
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_tanh (mpfr_ptr y, mpfr_srcptr xt , mpfr_rnd_t rnd_mode)
{
/****** Declaration ******/
mpfr_t x;
int inexact;
MPFR_SAVE_EXPO_DECL (expo);
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));
/* Special value checking */
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))
{
/* tanh(inf) = 1 && tanh(-inf) = -1 */
return mpfr_set_si (y, MPFR_INT_SIGN (xt), rnd_mode);
}
else /* tanh (0) = 0 and xt is zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO(xt));
MPFR_SET_ZERO (y);
MPFR_SET_SAME_SIGN (y, xt);
MPFR_RET (0);
}
}
/* tanh(x) = x - x^3/3 + ... so the error is < 2^(3*EXP(x)-1) */
MPFR_FAST_COMPUTE_IF_SMALL_INPUT (y, xt, -2 * MPFR_GET_EXP (xt), 1, 0,
rnd_mode, {});
MPFR_TMP_INIT_ABS (x, xt);
MPFR_SAVE_EXPO_MARK (expo);
/* General case */
{
/* Declaration of the intermediary variable */
mpfr_t t, te;
mpfr_exp_t d;
/* Declaration of the size variable */
mpfr_prec_t Ny = MPFR_PREC(y); /* target precision */
mpfr_prec_t Nt; /* working precision */
long int err; /* error */
int sign = MPFR_SIGN (xt);
MPFR_ZIV_DECL (loop);
MPFR_GROUP_DECL (group);
/* First check for BIG overflow of exp(2*x):
For x > 0, exp(2*x) > 2^(2*x)
If 2 ^(2*x) > 2^emax or x>emax/2, there is an overflow */
if (MPFR_UNLIKELY (mpfr_cmp_si (x, __gmpfr_emax/2) >= 0)) {
/* initialise of intermediary variables
since 'set_one' label assumes the variables have been
initialize */
MPFR_GROUP_INIT_2 (group, MPFR_PREC_MIN, t, te);
goto set_one;
}
/* Compute the precision of intermediary variable */
/* The optimal number of bits: see algorithms.tex */
Nt = Ny + MPFR_INT_CEIL_LOG2 (Ny) + 4;
/* if x is small, there will be a cancellation in exp(2x)-1 */
if (MPFR_GET_EXP (x) < 0)
Nt += -MPFR_GET_EXP (x);
/* initialise of intermediary variable */
MPFR_GROUP_INIT_2 (group, Nt, t, te);
MPFR_ZIV_INIT (loop, Nt);
for (;;) {
/* tanh = (exp(2x)-1)/(exp(2x)+1) */
mpfr_mul_2ui (te, x, 1, MPFR_RNDN); /* 2x */
/* since x > 0, we can only have an overflow */
mpfr_exp (te, te, MPFR_RNDN); /* exp(2x) */
if (MPFR_UNLIKELY (MPFR_IS_INF (te))) {
set_one:
inexact = MPFR_FROM_SIGN_TO_INT (sign);
mpfr_set4 (y, __gmpfr_one, MPFR_RNDN, sign);
if (MPFR_IS_LIKE_RNDZ (rnd_mode, MPFR_IS_NEG_SIGN (sign)))
{
inexact = -inexact;
mpfr_nexttozero (y);
}
break;
}
d = MPFR_GET_EXP (te); /* For Error calculation */
mpfr_add_ui (t, te, 1, MPFR_RNDD); /* exp(2x) + 1*/
mpfr_sub_ui (te, te, 1, MPFR_RNDU); /* exp(2x) - 1*/
d = d - MPFR_GET_EXP (te);
mpfr_div (t, te, t, MPFR_RNDN); /* (exp(2x)-1)/(exp(2x)+1)*/
/* Calculation of the error */
d = MAX(3, d + 1);
err = Nt - (d + 1);
if (MPFR_LIKELY ((d <= Nt / 2) && MPFR_CAN_ROUND (t, err, Ny, rnd_mode)))
{
inexact = mpfr_set4 (y, t, rnd_mode, sign);
break;
}
/* if t=1, we still can round since |sinh(x)| < 1 */
if (MPFR_GET_EXP (t) == 1)
goto set_one;
/* Actualisation of the precision */
MPFR_ZIV_NEXT (loop, Nt);
MPFR_GROUP_REPREC_2 (group, Nt, t, te);
}
MPFR_ZIV_FREE (loop);
MPFR_GROUP_CLEAR (group);
}
MPFR_SAVE_EXPO_FREE (expo);
inexact = mpfr_check_range (y, inexact, rnd_mode);
return inexact;
}
int
mpfr_sin (mpfr_ptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_t c, xr;
mpfr_srcptr xx;
mpfr_exp_t expx, err;
mpfr_prec_t precy, m;
int inexact, sign, reduce;
MPFR_ZIV_DECL (loop);
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_IS_INF (x))
{
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);
}
}
/* sin(x) = x - x^3/6 + ... so the error is < 2^(3*EXP(x)-2) */
MPFR_FAST_COMPUTE_IF_SMALL_INPUT (y, x, -2 * MPFR_GET_EXP (x), 2, 0,
rnd_mode, {});
MPFR_SAVE_EXPO_MARK (expo);
/* Compute initial precision */
precy = MPFR_PREC (y);
if (precy >= MPFR_SINCOS_THRESHOLD)
return mpfr_sin_fast (y, x, rnd_mode);
m = precy + MPFR_INT_CEIL_LOG2 (precy) + 13;
expx = MPFR_GET_EXP (x);
mpfr_init (c);
mpfr_init (xr);
MPFR_ZIV_INIT (loop, m);
for (;;)
{
/* first perform argument reduction modulo 2*Pi (if needed),
also helps to determine the sign of sin(x) */
if (expx >= 2) /* If Pi < x < 4, we need to reduce too, to determine
the sign of sin(x). For 2 <= |x| < Pi, we could avoid
the reduction. */
{
reduce = 1;
/* As expx + m - 1 will silently be converted into mpfr_prec_t
in the mpfr_set_prec call, the assert below may be useful to
avoid undefined behavior. */
MPFR_ASSERTN (expx + m - 1 <= MPFR_PREC_MAX);
mpfr_set_prec (c, expx + m - 1);
mpfr_set_prec (xr, m);
mpfr_const_pi (c, MPFR_RNDN);
mpfr_mul_2ui (c, c, 1, MPFR_RNDN);
mpfr_remainder (xr, x, c, MPFR_RNDN);
/* The analysis is similar to that of cos.c:
|xr - x - 2kPi| <= 2^(2-m). Thus we can decide the sign
of sin(x) if xr is at distance at least 2^(2-m) of both
0 and +/-Pi. */
mpfr_div_2ui (c, c, 1, MPFR_RNDN);
/* Since c approximates Pi with an error <= 2^(2-expx-m) <= 2^(-m),
it suffices to check that c - |xr| >= 2^(2-m). */
if (MPFR_SIGN (xr) > 0)
mpfr_sub (c, c, xr, MPFR_RNDZ);
else
mpfr_add (c, c, xr, MPFR_RNDZ);
if (MPFR_IS_ZERO(xr)
|| MPFR_EXP(xr) < (mpfr_exp_t) 3 - (mpfr_exp_t) m
|| MPFR_EXP(c) < (mpfr_exp_t) 3 - (mpfr_exp_t) m)
goto ziv_next;
/* |xr - x - 2kPi| <= 2^(2-m), thus |sin(xr) - sin(x)| <= 2^(2-m) */
xx = xr;
}
else /* the input argument is already reduced */
{
reduce = 0;
xx = x;
}
sign = MPFR_SIGN(xx);
/* now that the argument is reduced, precision m is enough */
mpfr_set_prec (c, m);
mpfr_cos (c, xx, MPFR_RNDZ); /* can't be exact */
mpfr_nexttoinf (c); /* now c = cos(x) rounded away */
mpfr_mul (c, c, c, MPFR_RNDU); /* away */
mpfr_ui_sub (c, 1, c, MPFR_RNDZ);
mpfr_sqrt (c, c, MPFR_RNDZ);
if (MPFR_IS_NEG_SIGN(sign))
MPFR_CHANGE_SIGN(c);
/* Warning: c may be 0! */
if (MPFR_UNLIKELY (MPFR_IS_ZERO (c)))
{
/* Huge cancellation: increase prec a lot! */
m = MAX (m, MPFR_PREC (x));
m = 2 * m;
}
else
{
/* the absolute error on c is at most 2^(3-m-EXP(c)),
plus 2^(2-m) if there was an argument reduction.
Since EXP(c) <= 1, 3-m-EXP(c) >= 2-m, thus the error
is at most 2^(3-m-EXP(c)) in case of argument reduction. */
err = 2 * MPFR_GET_EXP (c) + (mpfr_exp_t) m - 3 - (reduce != 0);
if (MPFR_CAN_ROUND (c, err, precy, rnd_mode))
break;
/* check for huge cancellation (Near 0) */
if (err < (mpfr_exp_t) MPFR_PREC (y))
m += MPFR_PREC (y) - err;
/* Check if near 1 */
if (MPFR_GET_EXP (c) == 1)
m += m;
}
ziv_next:
/* Else generic increase */
MPFR_ZIV_NEXT (loop, m);
}
MPFR_ZIV_FREE (loop);
inexact = mpfr_set (y, c, rnd_mode);
/* inexact cannot be 0, since this would mean that c was representable
within the target precision, but in that case mpfr_can_round will fail */
mpfr_clear (c);
mpfr_clear (xr);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (y, inexact, rnd_mode);
}
