int
mpfr_si_sub (mpfr_ptr y, long int u, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
if (u >= 0)
return mpfr_ui_sub (y, u, x, rnd_mode);
else
{
int res = -mpfr_add_ui (y, x, -u, MPFR_INVERT_RND (rnd_mode));
MPFR_CHANGE_SIGN (y);
return res;
}
}
int mpfr_mul_si (mpfr_ptr y, mpfr_srcptr x, long int u, mpfr_rnd_t rnd_mode)
{
int res;
if (u >= 0)
res = mpfr_mul_ui (y, x, u, rnd_mode);
else
{
res = -mpfr_mul_ui (y, x, -u, MPFR_INVERT_RND (rnd_mode));
MPFR_CHANGE_SIGN (y);
}
return res;
}
int
mpfr_set_sj_2exp (mpfr_t x, intmax_t j, intmax_t e, mp_rnd_t rnd)
{
if (j>=0)
return mpfr_set_uj_2exp (x, j, e, rnd);
else
{
int inex;
inex = mpfr_set_uj_2exp (x, - (uintmax_t) j, e, MPFR_INVERT_RND (rnd));
MPFR_CHANGE_SIGN (x);
return -inex;
}
}
int mpfr_si_div (mpfr_ptr y, long int u, mpfr_srcptr x,mp_rnd_t rnd_mode)
{
int res;
if (u >= 0)
res = mpfr_ui_div (y, u, x, rnd_mode);
else
{
res = -mpfr_ui_div (y, -u, x, MPFR_INVERT_RND(rnd_mode));
MPFR_CHANGE_SIGN (y);
}
return res;
}
static int
pi_div_2ui (mpfr_ptr dest, int i, int neg, mpfr_rnd_t rnd_mode)
{
int inexact;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_SAVE_EXPO_MARK (expo);
if (neg) /* -PI/2^i */
{
inexact = - mpfr_const_pi (dest, MPFR_INVERT_RND (rnd_mode));
MPFR_CHANGE_SIGN (dest);
}
else /* PI/2^i */
{
inexact = mpfr_const_pi (dest, rnd_mode);
}
mpfr_div_2ui (dest, dest, i, rnd_mode); /* exact */
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (dest, inexact, rnd_mode);
}
int
mpfr_cbrt (mpfr_ptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpz_t m;
mpfr_exp_t e, r, sh;
mpfr_prec_t n, size_m, tmp;
int inexact, negative;
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));
/* special values */
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_INF (y);
MPFR_SET_SAME_SIGN (y, x);
MPFR_RET (0);
}
/* case 0: cbrt(+/- 0) = +/- 0 */
else /* x is necessarily 0 */
{
MPFR_ASSERTD (MPFR_IS_ZERO (x));
MPFR_SET_ZERO (y);
MPFR_SET_SAME_SIGN (y, x);
MPFR_RET (0);
}
}
/* General case */
MPFR_SAVE_EXPO_MARK (expo);
mpz_init (m);
e = mpfr_get_z_2exp (m, x); /* x = m * 2^e */
if ((negative = MPFR_IS_NEG(x)))
mpz_neg (m, m);
r = e % 3;
if (r < 0)
r += 3;
/* x = (m*2^r) * 2^(e-r) = (m*2^r) * 2^(3*q) */
MPFR_MPZ_SIZEINBASE2 (size_m, m);
n = MPFR_PREC (y) + (rnd_mode == MPFR_RNDN);
/* we want 3*n-2 <= size_m + 3*sh + r <= 3*n
i.e. 3*sh + size_m + r <= 3*n */
sh = (3 * (mpfr_exp_t) n - (mpfr_exp_t) size_m - r) / 3;
sh = 3 * sh + r;
if (sh >= 0)
{
mpz_mul_2exp (m, m, sh);
e = e - sh;
}
else if (r > 0)
{
mpz_mul_2exp (m, m, r);
e = e - r;
}
/* invariant: x = m*2^e, with e divisible by 3 */
/* we reuse the variable m to store the cube root, since it is not needed
any more: we just need to know if the root is exact */
inexact = mpz_root (m, m, 3) == 0;
MPFR_MPZ_SIZEINBASE2 (tmp, m);
sh = tmp - n;
if (sh > 0) /* we have to flush to 0 the last sh bits from m */
{
inexact = inexact || ((mpfr_exp_t) mpz_scan1 (m, 0) < sh);
mpz_fdiv_q_2exp (m, m, sh);
e += 3 * sh;
}
if (inexact)
{
if (negative)
rnd_mode = MPFR_INVERT_RND (rnd_mode);
if (rnd_mode == MPFR_RNDU || rnd_mode == MPFR_RNDA
|| (rnd_mode == MPFR_RNDN && mpz_tstbit (m, 0)))
inexact = 1, mpz_add_ui (m, m, 1);
else
inexact = -1;
}
/* either inexact is not zero, and the conversion is exact, i.e. inexact
is not changed; or inexact=0, and inexact is set only when
rnd_mode=MPFR_RNDN and bit (n+1) from m is 1 */
inexact += mpfr_set_z (y, m, MPFR_RNDN);
MPFR_SET_EXP (y, MPFR_GET_EXP (y) + e / 3);
if (negative)
{
MPFR_CHANGE_SIGN (y);
inexact = -inexact;
}
mpz_clear (m);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (y, inexact, rnd_mode);
}
static void
special (void)
{
mpfr_t x, y, z;
int r;
int i;
mpfr_init2 (x, 53);
mpfr_init2 (y, 53);
mpfr_init2 (z, 53);
mpfr_set_str_binary (x, "1.0000100110000001100111100011001110101110100111011101");
mpfr_set_str_binary (y, "1.1001101101110100101100110011011101101000011010111110e-1");
mpfr_atan (z, x, MPFR_RNDN);
if (mpfr_cmp (y, z))
{
printf ("Error in mpfr_atan for prec=53, rnd=MPFR_RNDN\n");
printf ("x=");
mpfr_out_str (stdout, 2, 0, x, MPFR_RNDN);
printf ("\nexpected ");
mpfr_out_str (stdout, 2, 0, y, MPFR_RNDN);
printf ("\ngot ");
mpfr_out_str (stdout, 2, 0, z, MPFR_RNDN);
printf ("\n");
exit (1);
}
/* atan(+Inf) = Pi/2 */
for (r = 0; r < MPFR_RND_MAX ; r++)
{
mpfr_set_inf (x, 1);
mpfr_atan (y, x, (mpfr_rnd_t) r);
mpfr_const_pi (x, (mpfr_rnd_t) r);
mpfr_div_2exp (x, x, 1, (mpfr_rnd_t) r);
if (mpfr_cmp (x, y))
{
printf ("Error: mpfr_atan(+Inf), rnd=%s\n",
mpfr_print_rnd_mode ((mpfr_rnd_t) r));
exit (1);
}
}
/* atan(-Inf) = - Pi/2 */
for (r = 0; r < MPFR_RND_MAX ; r++)
{
mpfr_set_inf (x, -1);
mpfr_atan (y, x, (mpfr_rnd_t) r);
mpfr_const_pi (x, MPFR_INVERT_RND((mpfr_rnd_t) r));
mpfr_neg (x, x, (mpfr_rnd_t) r);
mpfr_div_2exp (x, x, 1, (mpfr_rnd_t) r);
if (mpfr_cmp (x, y))
{
printf ("Error: mpfr_atan(-Inf), rnd=%s\n",
mpfr_print_rnd_mode ((mpfr_rnd_t) r));
exit (1);
}
}
/* atan(NaN) = NaN */
mpfr_set_nan (x);
mpfr_atan (y, x, MPFR_RNDN);
if (!mpfr_nan_p (y))
{
printf ("Error: mpfr_atan(NaN) <> NaN\n");
exit (1);
}
/* atan(+/-0) = +/-0 */
mpfr_set_ui (x, 0, MPFR_RNDN);
MPFR_SET_NEG (y);
mpfr_atan (y, x, MPFR_RNDN);
if (mpfr_cmp_ui (y, 0) || MPFR_IS_NEG (y))
{
printf ("Error: mpfr_atan (+0) <> +0\n");
exit (1);
}
mpfr_atan (x, x, MPFR_RNDN);
if (mpfr_cmp_ui (x, 0) || MPFR_IS_NEG (x))
{
printf ("Error: mpfr_atan (+0) <> +0 (in place)\n");
exit (1);
}
mpfr_neg (x, x, MPFR_RNDN);
MPFR_SET_POS (y);
mpfr_atan (y, x, MPFR_RNDN);
if (mpfr_cmp_ui (y, 0) || MPFR_IS_POS (y))
{
printf ("Error: mpfr_atan (-0) <> -0\n");
exit (1);
}
mpfr_atan (x, x, MPFR_RNDN);
if (mpfr_cmp_ui (x, 0) || MPFR_IS_POS (x))
{
printf ("Error: mpfr_atan (-0) <> -0 (in place)\n");
exit (1);
}
mpfr_set_prec (x, 32);
mpfr_set_prec (y, 32);
/* test one random positive argument */
mpfr_set_str_binary (x, "0.10000100001100101001001001011001");
mpfr_atan (x, x, MPFR_RNDN);
mpfr_set_str_binary (y, "0.1111010000001111001111000000011E-1");
if (mpfr_cmp (x, y))
{
printf ("Error in mpfr_atan (1)\n");
exit (1);
}
/* test one random negative argument */
mpfr_set_str_binary (x, "-0.1100001110110000010101011001011");
mpfr_atan (x, x, MPFR_RNDN);
mpfr_set_str_binary (y, "-0.101001110001010010110001110001");
if (mpfr_cmp (x, y))
{
printf ("Error in mpfr_atan (2)\n");
mpfr_print_binary (x);
printf ("\n");
mpfr_print_binary (y);
printf ("\n");
exit (1);
}
mpfr_set_prec (x, 3);
mpfr_set_prec (y, 192);
mpfr_set_prec (z, 192);
mpfr_set_str_binary (x, "-0.100e1");
mpfr_atan (z, x, MPFR_RNDD);
mpfr_set_str_binary (y, "-0.110010010000111111011010101000100010000101101000110000100011010011000100110001100110001010001011100000001101110000011100110100010010100100000010010011100000100010001010011001111100110001110101");
if (mpfr_cmp (z, y))
{
printf ("Error in mpfr_atan (3)\n");
printf ("Expected ");
mpfr_print_binary (y);
printf ("\n");
printf ("Got ");
mpfr_print_binary (z);
printf ("\n");
exit (1);
}
/* Test regression */
mpfr_set_prec (x, 51);
mpfr_set_prec (y, 51);
mpfr_set_str_binary (x,
"0.101100100000101111111010001111111000001000000000000E-11");
i = mpfr_atan (y, x, MPFR_RNDN);
if (mpfr_cmp_str (y,
"1.01100100000101111111001110011001010110100100000000e-12", 2, MPFR_RNDN)
|| i >= 0)
{
printf ("Wrong Regression test (%d)\n", i);
mpfr_dump (y);
exit (1);
}
mpfr_set_si (x, -1, MPFR_RNDN);
mpfr_atan (x, x, MPFR_RNDN);
MPFR_ASSERTN (MPFR_IS_NEG (x));
/* Test regression */
mpfr_set_prec (x, 48);
mpfr_set_prec (y, 48);
mpfr_set_str_binary (x, "1.11001110010000011111100000010000000000000000000e-19");
mpfr_atan (y, x, MPFR_RNDD);
if (mpfr_cmp_str (y, "0.111001110010000011111100000001111111110000010011E-18", 2, MPFR_RNDN))
{
printf ("Error in mpfr_atan (4)\n");
printf ("Input 1.11001110010000011111100000010000000000000000000e-19 [prec=48]\n");
printf ("Expected 0.111001110010000011111100000001111111110000010011E-18\n");
printf ("Got ");
mpfr_dump (y);
exit (1);
}
mpfr_clear (x);
mpfr_clear (y);
mpfr_clear (z);
}
int
mpfr_sub (mpfr_ptr a, mpfr_srcptr b, mpfr_srcptr c, mpfr_rnd_t rnd_mode)
{
MPFR_LOG_FUNC (("b[%#R]=%R c[%#R]=%R rnd=%d", b, b, c, c, rnd_mode),
("a[%#R]=%R", a, a));
if (MPFR_ARE_SINGULAR (b,c))
{
if (MPFR_IS_NAN (b) || MPFR_IS_NAN (c))
{
MPFR_SET_NAN (a);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (b))
{
if (!MPFR_IS_INF (c) || MPFR_SIGN (b) != MPFR_SIGN(c))
{
MPFR_SET_INF (a);
MPFR_SET_SAME_SIGN (a, b);
MPFR_RET (0); /* exact */
}
else
{
MPFR_SET_NAN (a); /* Inf - Inf */
MPFR_RET_NAN;
}
}
else if (MPFR_IS_INF (c))
{
MPFR_SET_INF (a);
MPFR_SET_OPPOSITE_SIGN (a, c);
MPFR_RET (0); /* exact */
}
else if (MPFR_IS_ZERO (b))
{
if (MPFR_IS_ZERO (c))
{
int sign = rnd_mode != MPFR_RNDD
? ((MPFR_IS_NEG(b) && MPFR_IS_POS(c)) ? -1 : 1)
: ((MPFR_IS_POS(b) && MPFR_IS_NEG(c)) ? 1 : -1);
MPFR_SET_SIGN (a, sign);
MPFR_SET_ZERO (a);
MPFR_RET(0); /* 0 - 0 is exact */
}
else
return mpfr_neg (a, c, rnd_mode);
}
else
{
MPFR_ASSERTD (MPFR_IS_ZERO (c));
return mpfr_set (a, b, rnd_mode);
}
}
MPFR_ASSERTD (MPFR_IS_PURE_FP (b) && MPFR_IS_PURE_FP (c));
if (MPFR_LIKELY (MPFR_SIGN (b) == MPFR_SIGN (c)))
{ /* signs are equal, it's a real subtraction */
if (MPFR_LIKELY (MPFR_PREC (a) == MPFR_PREC (b)
&& MPFR_PREC (b) == MPFR_PREC (c)))
return mpfr_sub1sp (a, b, c, rnd_mode);
else
return mpfr_sub1 (a, b, c, rnd_mode);
}
else
{ /* signs differ, it's an addition */
if (MPFR_GET_EXP (b) < MPFR_GET_EXP (c))
{ /* exchange rounding modes toward +/- infinity */
int inexact;
rnd_mode = MPFR_INVERT_RND (rnd_mode);
if (MPFR_LIKELY (MPFR_PREC (a) == MPFR_PREC (b)
&& MPFR_PREC (b) == MPFR_PREC (c)))
inexact = mpfr_add1sp (a, c, b, rnd_mode);
else
inexact = mpfr_add1 (a, c, b, rnd_mode);
MPFR_CHANGE_SIGN (a);
return -inexact;
}
else
{
if (MPFR_LIKELY (MPFR_PREC (a) == MPFR_PREC (b)
&& MPFR_PREC (b) == MPFR_PREC (c)))
return mpfr_add1sp (a, b, c, rnd_mode);
else
return mpfr_add1 (a, b, c, rnd_mode);
}
}
}
/* 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_asin (mpfr_ptr asin, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_t xp;
int compared, inexact;
mpfr_prec_t prec;
mpfr_exp_t xp_exp;
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),
("asin[%Pu]=%.*Rg inexact=%d", mpfr_get_prec (asin), mpfr_log_prec, asin,
inexact));
/* Special cases */
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x)))
{
if (MPFR_IS_NAN (x) || MPFR_IS_INF (x))
{
MPFR_SET_NAN (asin);
MPFR_RET_NAN;
}
else /* x = 0 */
{
MPFR_ASSERTD (MPFR_IS_ZERO (x));
MPFR_SET_ZERO (asin);
MPFR_SET_SAME_SIGN (asin, x);
MPFR_RET (0); /* exact result */
}
}
/* asin(x) = x + x^3/6 + ... so the error is < 2^(3*EXP(x)-2) */
MPFR_FAST_COMPUTE_IF_SMALL_INPUT (asin, x, -2 * MPFR_GET_EXP (x), 2, 1,
rnd_mode, {});
/* Set x_p=|x| (x is a normal number) */
mpfr_init2 (xp, MPFR_PREC (x));
inexact = mpfr_abs (xp, x, MPFR_RNDN);
MPFR_ASSERTD (inexact == 0);
compared = mpfr_cmp_ui (xp, 1);
MPFR_SAVE_EXPO_MARK (expo);
if (MPFR_UNLIKELY (compared >= 0))
{
mpfr_clear (xp);
if (compared > 0) /* asin(x) = NaN for |x| > 1 */
{
MPFR_SAVE_EXPO_FREE (expo);
MPFR_SET_NAN (asin);
MPFR_RET_NAN;
}
else /* x = 1 or x = -1 */
{
if (MPFR_IS_POS (x)) /* asin(+1) = Pi/2 */
inexact = mpfr_const_pi (asin, rnd_mode);
else /* asin(-1) = -Pi/2 */
{
inexact = -mpfr_const_pi (asin, MPFR_INVERT_RND(rnd_mode));
MPFR_CHANGE_SIGN (asin);
}
mpfr_div_2ui (asin, asin, 1, rnd_mode);
}
}
else
{
/* Compute exponent of 1 - ABS(x) */
mpfr_ui_sub (xp, 1, xp, MPFR_RNDD);
MPFR_ASSERTD (MPFR_GET_EXP (xp) <= 0);
MPFR_ASSERTD (MPFR_GET_EXP (x) <= 0);
xp_exp = 2 - MPFR_GET_EXP (xp);
/* Set up initial prec */
prec = MPFR_PREC (asin) + 10 + xp_exp;
/* use asin(x) = atan(x/sqrt(1-x^2)) */
MPFR_ZIV_INIT (loop, prec);
for (;;)
{
mpfr_set_prec (xp, prec);
mpfr_sqr (xp, x, MPFR_RNDN);
mpfr_ui_sub (xp, 1, xp, MPFR_RNDN);
mpfr_sqrt (xp, xp, MPFR_RNDN);
mpfr_div (xp, x, xp, MPFR_RNDN);
mpfr_atan (xp, xp, MPFR_RNDN);
if (MPFR_LIKELY (MPFR_CAN_ROUND (xp, prec - xp_exp,
MPFR_PREC (asin), rnd_mode)))
break;
MPFR_ZIV_NEXT (loop, prec);
}
MPFR_ZIV_FREE (loop);
inexact = mpfr_set (asin, xp, rnd_mode);
mpfr_clear (xp);
}
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (asin, inexact, rnd_mode);
}
int
mpfr_atan2 (mpfr_ptr dest, mpfr_srcptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_t tmp, pi;
int inexact;
mpfr_prec_t prec;
mpfr_exp_t e;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_ZIV_DECL (loop);
MPFR_LOG_FUNC
(("y[%Pu]=%.*Rg x[%Pu]=%.*Rg rnd=%d",
mpfr_get_prec (y), mpfr_log_prec, y,
mpfr_get_prec (x), mpfr_log_prec, x, rnd_mode),
("atan[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (dest), mpfr_log_prec, dest, inexact));
/* Special cases */
if (MPFR_ARE_SINGULAR (x, y))
{
/* atan2(0, 0) does not raise the "invalid" floating-point
exception, nor does atan2(y, 0) raise the "divide-by-zero"
floating-point exception.
-- atan2(±0, -0) returns ±pi.313)
-- atan2(±0, +0) returns ±0.
-- atan2(±0, x) returns ±pi, for x < 0.
-- atan2(±0, x) returns ±0, for x > 0.
-- atan2(y, ±0) returns -pi/2 for y < 0.
-- atan2(y, ±0) returns pi/2 for y > 0.
-- atan2(±oo, -oo) returns ±3pi/4.
-- atan2(±oo, +oo) returns ±pi/4.
-- atan2(±oo, x) returns ±pi/2, for finite x.
-- atan2(±y, -oo) returns ±pi, for finite y > 0.
-- atan2(±y, +oo) returns ±0, for finite y > 0.
*/
if (MPFR_IS_NAN (x) || MPFR_IS_NAN (y))
{
MPFR_SET_NAN (dest);
MPFR_RET_NAN;
}
if (MPFR_IS_ZERO (y))
{
if (MPFR_IS_NEG (x)) /* +/- PI */
{
set_pi:
if (MPFR_IS_NEG (y))
{
inexact = mpfr_const_pi (dest, MPFR_INVERT_RND (rnd_mode));
MPFR_CHANGE_SIGN (dest);
return -inexact;
}
else
return mpfr_const_pi (dest, rnd_mode);
}
else /* +/- 0 */
{
set_zero:
MPFR_SET_ZERO (dest);
MPFR_SET_SAME_SIGN (dest, y);
return 0;
}
}
if (MPFR_IS_ZERO (x))
{
return pi_div_2ui (dest, 1, MPFR_IS_NEG (y), rnd_mode);
}
if (MPFR_IS_INF (y))
{
if (!MPFR_IS_INF (x)) /* +/- PI/2 */
return pi_div_2ui (dest, 1, MPFR_IS_NEG (y), rnd_mode);
else if (MPFR_IS_POS (x)) /* +/- PI/4 */
return pi_div_2ui (dest, 2, MPFR_IS_NEG (y), rnd_mode);
else /* +/- 3*PI/4: Ugly since we have to round properly */
{
mpfr_t tmp2;
MPFR_ZIV_DECL (loop2);
mpfr_prec_t prec2 = MPFR_PREC (dest) + 10;
MPFR_SAVE_EXPO_MARK (expo);
mpfr_init2 (tmp2, prec2);
MPFR_ZIV_INIT (loop2, prec2);
for (;;)
{
mpfr_const_pi (tmp2, MPFR_RNDN);
mpfr_mul_ui (tmp2, tmp2, 3, MPFR_RNDN); /* Error <= 2 */
mpfr_div_2ui (tmp2, tmp2, 2, MPFR_RNDN);
if (mpfr_round_p (MPFR_MANT (tmp2), MPFR_LIMB_SIZE (tmp2),
MPFR_PREC (tmp2) - 2,
MPFR_PREC (dest) + (rnd_mode == MPFR_RNDN)))
break;
MPFR_ZIV_NEXT (loop2, prec2);
mpfr_set_prec (tmp2, prec2);
}
MPFR_ZIV_FREE (loop2);
if (MPFR_IS_NEG (y))
MPFR_CHANGE_SIGN (tmp2);
inexact = mpfr_set (dest, tmp2, rnd_mode);
mpfr_clear (tmp2);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (dest, inexact, rnd_mode);
}
}
MPFR_ASSERTD (MPFR_IS_INF (x));
if (MPFR_IS_NEG (x))
goto set_pi;
else
goto set_zero;
}
/* When x is a power of two, we call directly atan(y/x) since y/x is
exact. */
if (MPFR_UNLIKELY (MPFR_IS_POWER_OF_2 (x)))
{
int r;
mpfr_t yoverx;
unsigned int saved_flags = __gmpfr_flags;
mpfr_init2 (yoverx, MPFR_PREC (y));
if (MPFR_LIKELY (mpfr_div_2si (yoverx, y, MPFR_GET_EXP (x) - 1,
MPFR_RNDN) == 0))
{
/* Here the flags have not changed due to mpfr_div_2si. */
r = mpfr_atan (dest, yoverx, rnd_mode);
mpfr_clear (yoverx);
return r;
}
else
{
/* Division is inexact because of a small exponent range */
mpfr_clear (yoverx);
__gmpfr_flags = saved_flags;
}
}
MPFR_SAVE_EXPO_MARK (expo);
/* Set up initial prec */
prec = MPFR_PREC (dest) + 3 + MPFR_INT_CEIL_LOG2 (MPFR_PREC (dest));
mpfr_init2 (tmp, prec);
MPFR_ZIV_INIT (loop, prec);
if (MPFR_IS_POS (x))
/* use atan2(y,x) = atan(y/x) */
for (;;)
{
int div_inex;
MPFR_BLOCK_DECL (flags);
MPFR_BLOCK (flags, div_inex = mpfr_div (tmp, y, x, MPFR_RNDN));
if (div_inex == 0)
{
/* Result is exact. */
inexact = mpfr_atan (dest, tmp, rnd_mode);
goto end;
}
/* Error <= ulp (tmp) except in case of underflow or overflow. */
/* If the division underflowed, since |atan(z)/z| < 1, we have
an underflow. */
if (MPFR_UNDERFLOW (flags))
{
int sign;
/* In the case MPFR_RNDN with 2^(emin-2) < |y/x| < 2^(emin-1):
The smallest significand value S > 1 of |y/x| is:
* 1 / (1 - 2^(-px)) if py <= px,
* (1 - 2^(-px) + 2^(-py)) / (1 - 2^(-px)) if py >= px.
Therefore S - 1 > 2^(-pz), where pz = max(px,py). We have:
atan(|y/x|) > atan(z), where z = 2^(emin-2) * (1 + 2^(-pz)).
> z - z^3 / 3.
> 2^(emin-2) * (1 + 2^(-pz) - 2^(2 emin - 5))
Assuming pz <= -2 emin + 5, we can round away from zero
(this is what mpfr_underflow always does on MPFR_RNDN).
In the case MPFR_RNDN with |y/x| <= 2^(emin-2), we round
toward zero, as |atan(z)/z| < 1. */
MPFR_ASSERTN (MPFR_PREC_MAX <=
2 * (mpfr_uexp_t) - MPFR_EMIN_MIN + 5);
if (rnd_mode == MPFR_RNDN && MPFR_IS_ZERO (tmp))
rnd_mode = MPFR_RNDZ;
sign = MPFR_SIGN (tmp);
mpfr_clear (tmp);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_underflow (dest, rnd_mode, sign);
}
mpfr_atan (tmp, tmp, MPFR_RNDN); /* Error <= 2*ulp (tmp) since
abs(D(arctan)) <= 1 */
/* TODO: check that the error bound is correct in case of overflow. */
/* FIXME: Error <= ulp(tmp) ? */
if (MPFR_LIKELY (MPFR_CAN_ROUND (tmp, prec - 2, MPFR_PREC (dest),
rnd_mode)))
break;
MPFR_ZIV_NEXT (loop, prec);
mpfr_set_prec (tmp, prec);
}
else /* x < 0 */
/* Use sign(y)*(PI - atan (|y/x|)) */
{
mpfr_init2 (pi, prec);
for (;;)
{
mpfr_div (tmp, y, x, MPFR_RNDN); /* Error <= ulp (tmp) */
/* If tmp is 0, we have |y/x| <= 2^(-emin-2), thus
atan|y/x| < 2^(-emin-2). */
MPFR_SET_POS (tmp); /* no error */
mpfr_atan (tmp, tmp, MPFR_RNDN); /* Error <= 2*ulp (tmp) since
abs(D(arctan)) <= 1 */
mpfr_const_pi (pi, MPFR_RNDN); /* Error <= ulp(pi) /2 */
e = MPFR_NOTZERO(tmp) ? MPFR_GET_EXP (tmp) : __gmpfr_emin - 1;
mpfr_sub (tmp, pi, tmp, MPFR_RNDN); /* see above */
if (MPFR_IS_NEG (y))
MPFR_CHANGE_SIGN (tmp);
/* Error(tmp) <= (1/2+2^(EXP(pi)-EXP(tmp)-1)+2^(e-EXP(tmp)+1))*ulp
<= 2^(MAX (MAX (EXP(PI)-EXP(tmp)-1, e-EXP(tmp)+1),
-1)+2)*ulp(tmp) */
e = MAX (MAX (MPFR_GET_EXP (pi)-MPFR_GET_EXP (tmp) - 1,
e - MPFR_GET_EXP (tmp) + 1), -1) + 2;
if (MPFR_LIKELY (MPFR_CAN_ROUND (tmp, prec - e, MPFR_PREC (dest),
rnd_mode)))
break;
MPFR_ZIV_NEXT (loop, prec);
mpfr_set_prec (tmp, prec);
mpfr_set_prec (pi, prec);
}
mpfr_clear (pi);
}
inexact = mpfr_set (dest, tmp, rnd_mode);
end:
MPFR_ZIV_FREE (loop);
mpfr_clear (tmp);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (dest, inexact, rnd_mode);
}
static void
special (void)
{
mpfr_t x, y;
int r;
mpfr_init (x);
mpfr_init (y);
/* asin(NaN) = NaN */
mpfr_set_nan (x);
mpfr_asin (y, x, MPFR_RNDN);
if (!mpfr_nan_p (y))
{
printf ("Error: mpfr_asin (NaN) <> NaN\n");
exit (1);
}
/* asin(+/-Inf) = NaN */
mpfr_set_inf (x, 1);
mpfr_asin (y, x, MPFR_RNDN);
if (!mpfr_nan_p (y))
{
printf ("Error: mpfr_asin (+Inf) <> NaN\n");
exit (1);
}
mpfr_set_inf (x, -1);
mpfr_asin (y, x, MPFR_RNDN);
if (!mpfr_nan_p (y))
{
printf ("Error: mpfr_asin (-Inf) <> NaN\n");
exit (1);
}
/* asin(+/-2) = NaN */
mpfr_set_ui (x, 2, MPFR_RNDN);
mpfr_asin (y, x, MPFR_RNDN);
if (!mpfr_nan_p (y))
{
printf ("Error: mpfr_asin (+2) <> NaN\n");
exit (1);
}
mpfr_set_si (x, -2, MPFR_RNDN);
mpfr_asin (y, x, MPFR_RNDN);
if (!mpfr_nan_p (y))
{
printf ("Error: mpfr_asin (-2) <> NaN\n");
exit (1);
}
/* asin(+/-0) = +/-0 */
mpfr_set_ui (x, 0, MPFR_RNDN);
mpfr_asin (y, x, MPFR_RNDN);
if (mpfr_cmp_ui (y, 0) || mpfr_sgn (y) < 0)
{
printf ("Error: mpfr_asin (+0) <> +0\n");
exit (1);
}
mpfr_neg (x, x, MPFR_RNDN);
mpfr_asin (y, x, MPFR_RNDN);
if (mpfr_cmp_ui (y, 0) || mpfr_sgn (y) > 0)
{
printf ("Error: mpfr_asin (-0) <> -0\n");
exit (1);
}
/* asin(1) = Pi/2 */
for (r = 0; r < MPFR_RND_MAX; r++)
{
mpfr_set_ui (x, 1, MPFR_RNDN); /* exact */
mpfr_asin (y, x, (mpfr_rnd_t) r);
mpfr_const_pi (x, (mpfr_rnd_t) r);
mpfr_div_2exp (x, x, 1, MPFR_RNDN); /* exact */
if (mpfr_cmp (x, y))
{
printf ("Error: asin(1) != Pi/2 for rnd=%s\n",
mpfr_print_rnd_mode ((mpfr_rnd_t) r));
exit (1);
}
}
/* asin(-1) = -Pi/2 */
for (r = 0; r < MPFR_RND_MAX; r++)
{
mpfr_set_si (x, -1, MPFR_RNDN); /* exact */
mpfr_asin (y, x, (mpfr_rnd_t) r);
mpfr_const_pi (x, MPFR_INVERT_RND((mpfr_rnd_t) r));
mpfr_neg (x, x, MPFR_RNDN); /* exact */
mpfr_div_2exp (x, x, 1, MPFR_RNDN); /* exact */
if (mpfr_cmp (x, y))
{
printf ("Error: asin(-1) != -Pi/2 for rnd=%s\n",
mpfr_print_rnd_mode ((mpfr_rnd_t) r));
exit (1);
}
}
mpfr_set_prec (x, 32);
mpfr_set_prec (y, 32);
mpfr_set_str_binary (x, "0.1101110111111111001011101000101");
mpfr_asin (x, x, MPFR_RNDN);
mpfr_set_str_binary (y, "1.00001100101011000001111100111");
if (mpfr_cmp (x, y))
{
printf ("Error: mpfr_asin (1)\n");
exit (1);
}
mpfr_set_str_binary (x, "-0.01110111000011101010111100000101");
mpfr_asin (x, x, MPFR_RNDN);
mpfr_set_str_binary (y, "-0.0111101111010100011111110101");
if (mpfr_cmp (x, y))
{
printf ("Error: mpfr_asin (2)\n");
mpfr_print_binary (x); printf ("\n");
mpfr_print_binary (y); printf ("\n");
exit (1);
}
mpfr_set_prec (x, 9);
mpfr_set_prec (y, 19);
mpfr_set_str_binary (x, "0.110000000E-6");
mpfr_asin (y, x, MPFR_RNDD);
mpfr_set_prec (x, 19);
mpfr_set_str_binary (x, "0.1100000000000001001E-6");
if (mpfr_cmp (x, y))
{
printf ("Error: mpfr_asin (3)\n");
mpfr_dump (x);
mpfr_dump (y);
exit (1);
}
mpfr_clear (x);
mpfr_clear (y);
}
int
mpfr_erf (mpfr_ptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_t xf;
mp_limb_t xf_limb[(53 - 1) / GMP_NUMB_BITS + 1];
int inex, large;
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, inex));
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)) /* erf(+inf) = +1, erf(-inf) = -1 */
return mpfr_set_si (y, MPFR_INT_SIGN (x), MPFR_RNDN);
else /* erf(+0) = +0, erf(-0) = -0 */
{
MPFR_ASSERTD (MPFR_IS_ZERO (x));
return mpfr_set (y, x, MPFR_RNDN); /* should keep the sign of x */
}
}
/* now x is neither NaN, Inf nor 0 */
/* first try expansion at x=0 when x is small, or asymptotic expansion
where x is large */
MPFR_SAVE_EXPO_MARK (expo);
/* around x=0, we have erf(x) = 2x/sqrt(Pi) (1 - x^2/3 + ...),
with 1 - x^2/3 <= sqrt(Pi)*erf(x)/2/x <= 1 for x >= 0. This means that
if x^2/3 < 2^(-PREC(y)-1) we can decide of the correct rounding,
unless we have a worst-case for 2x/sqrt(Pi). */
if (MPFR_EXP(x) < - (mpfr_exp_t) (MPFR_PREC(y) / 2))
{
/* we use 2x/sqrt(Pi) (1 - x^2/3) <= erf(x) <= 2x/sqrt(Pi) for x > 0
and 2x/sqrt(Pi) <= erf(x) <= 2x/sqrt(Pi) (1 - x^2/3) for x < 0.
In both cases |2x/sqrt(Pi) (1 - x^2/3)| <= |erf(x)| <= |2x/sqrt(Pi)|.
We will compute l and h such that l <= |2x/sqrt(Pi) (1 - x^2/3)|
and |2x/sqrt(Pi)| <= h. If l and h round to the same value to
precision PREC(y) and rounding rnd_mode, then we are done. */
mpfr_t l, h; /* lower and upper bounds for erf(x) */
int ok, inex2;
mpfr_init2 (l, MPFR_PREC(y) + 17);
mpfr_init2 (h, MPFR_PREC(y) + 17);
/* first compute l */
mpfr_mul (l, x, x, MPFR_RNDU);
mpfr_div_ui (l, l, 3, MPFR_RNDU); /* upper bound on x^2/3 */
mpfr_ui_sub (l, 1, l, MPFR_RNDZ); /* lower bound on 1 - x^2/3 */
mpfr_const_pi (h, MPFR_RNDU); /* upper bound of Pi */
mpfr_sqrt (h, h, MPFR_RNDU); /* upper bound on sqrt(Pi) */
mpfr_div (l, l, h, MPFR_RNDZ); /* lower bound on 1/sqrt(Pi) (1 - x^2/3) */
mpfr_mul_2ui (l, l, 1, MPFR_RNDZ); /* 2/sqrt(Pi) (1 - x^2/3) */
mpfr_mul (l, l, x, MPFR_RNDZ); /* |l| is a lower bound on
|2x/sqrt(Pi) (1 - x^2/3)| */
/* now compute h */
mpfr_const_pi (h, MPFR_RNDD); /* lower bound on Pi */
mpfr_sqrt (h, h, MPFR_RNDD); /* lower bound on sqrt(Pi) */
mpfr_div_2ui (h, h, 1, MPFR_RNDD); /* lower bound on sqrt(Pi)/2 */
/* since sqrt(Pi)/2 < 1, the following should not underflow */
mpfr_div (h, x, h, MPFR_IS_POS(x) ? MPFR_RNDU : MPFR_RNDD);
/* round l and h to precision PREC(y) */
inex = mpfr_prec_round (l, MPFR_PREC(y), rnd_mode);
inex2 = mpfr_prec_round (h, MPFR_PREC(y), rnd_mode);
/* Caution: we also need inex=inex2 (inex might be 0). */
ok = SAME_SIGN (inex, inex2) && mpfr_cmp (l, h) == 0;
if (ok)
mpfr_set (y, h, rnd_mode);
mpfr_clear (l);
mpfr_clear (h);
if (ok)
goto end;
/* this test can still fail for small precision, for example
for x=-0.100E-2 with a target precision of 3 bits, since
the error term x^2/3 is not that small. */
}
MPFR_TMP_INIT1(xf_limb, xf, 53);
mpfr_div (xf, x, __gmpfr_const_log2_RNDU, MPFR_RNDZ); /* round to zero
ensures we get a lower bound of |x/log(2)| */
mpfr_mul (xf, xf, x, MPFR_RNDZ);
large = mpfr_cmp_ui (xf, MPFR_PREC (y) + 1) > 0;
/* when x goes to infinity, we have erf(x) = 1 - 1/sqrt(Pi)/exp(x^2)/x + ...
and |erf(x) - 1| <= exp(-x^2) is true for any x >= 0, thus if
exp(-x^2) < 2^(-PREC(y)-1) the result is 1 or 1-epsilon.
This rewrites as x^2/log(2) > p+1. */
if (MPFR_UNLIKELY (large))
/* |erf x| = 1 or 1- */
{
mpfr_rnd_t rnd2 = MPFR_IS_POS (x) ? rnd_mode : MPFR_INVERT_RND(rnd_mode);
if (rnd2 == MPFR_RNDN || rnd2 == MPFR_RNDU || rnd2 == MPFR_RNDA)
{
inex = MPFR_INT_SIGN (x);
mpfr_set_si (y, inex, rnd2);
}
else /* round to zero */
{
inex = -MPFR_INT_SIGN (x);
mpfr_setmax (y, 0); /* warning: setmax keeps the old sign of y */
MPFR_SET_SAME_SIGN (y, x);
}
}
else /* use Taylor */
{
double xf2;
/* FIXME: get rid of doubles/mpfr_get_d here */
xf2 = mpfr_get_d (x, MPFR_RNDN);
xf2 = xf2 * xf2; /* xf2 ~ x^2 */
inex = mpfr_erf_0 (y, x, xf2, rnd_mode);
}
end:
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (y, inex, rnd_mode);
}
/* return in z a lower bound (for rnd = RNDD) or upper bound (for rnd = RNDU)
of |zeta(s)|/2, using:
log(|zeta(s)|/2) = (s-1)*log(2*Pi) + lngamma(1-s)
+ log(|sin(Pi*s/2)| * zeta(1-s)).
Assumes s < 1/2 and s1 = 1-s exactly, thus s1 > 1/2.
y and p are temporary variables.
At input, p is Pi rounded down.
The comments in the code are for rnd = RNDD. */
static void
mpfr_reflection_overflow (mpfr_t z, mpfr_t s1, const mpfr_t s, mpfr_t y,
mpfr_t p, mpfr_rnd_t rnd)
{
mpz_t sint;
MPFR_ASSERTD (rnd == MPFR_RNDD || rnd == MPFR_RNDU);
/* Since log is increasing, we want lower bounds on |sin(Pi*s/2)| and
zeta(1-s). */
mpz_init (sint);
mpfr_get_z (sint, s, MPFR_RNDD); /* sint = floor(s) */
/* We first compute a lower bound of |sin(Pi*s/2)|, which is a periodic
function of period 2. Thus:
if 2k < s < 2k+1, then |sin(Pi*s/2)| is increasing;
if 2k-1 < s < 2k, then |sin(Pi*s/2)| is decreasing.
These cases are distinguished by testing bit 0 of floor(s) as if
represented in two's complement (or equivalently, as an unsigned
integer mod 2):
0: sint = 0 mod 2, thus 2k < s < 2k+1 and |sin(Pi*s/2)| is increasing;
1: sint = 1 mod 2, thus 2k-1 < s < 2k and |sin(Pi*s/2)| is decreasing.
Let's recall that the comments are for rnd = RNDD. */
if (mpz_tstbit (sint, 0) == 0) /* |sin(Pi*s/2)| is increasing: round down
Pi*s to get a lower bound. */
{
mpfr_mul (y, p, s, rnd);
if (rnd == MPFR_RNDD)
mpfr_nextabove (p); /* we will need p rounded above afterwards */
}
else /* |sin(Pi*s/2)| is decreasing: round up Pi*s to get a lower bound. */
{
if (rnd == MPFR_RNDD)
mpfr_nextabove (p);
mpfr_mul (y, p, s, MPFR_INVERT_RND(rnd));
}
mpfr_div_2ui (y, y, 1, MPFR_RNDN); /* exact, rounding mode doesn't matter */
/* The rounding direction of sin depends on its sign. We have:
if -4k-2 < s < -4k, then -2k-1 < s/2 < -2k, thus sin(Pi*s/2) < 0;
if -4k < s < -4k+2, then -2k < s/2 < -2k+1, thus sin(Pi*s/2) > 0.
These cases are distinguished by testing bit 1 of floor(s) as if
represented in two's complement (or equivalently, as an unsigned
integer mod 4):
0: sint = {0,1} mod 4, thus -2k < s/2 < -2k+1 and sin(Pi*s/2) > 0;
1: sint = {2,3} mod 4, thus -2k-1 < s/2 < -2k and sin(Pi*s/2) < 0.
Let's recall that the comments are for rnd = RNDD. */
if (mpz_tstbit (sint, 1) == 0) /* -2k < s/2 < -2k+1; sin(Pi*s/2) > 0 */
{
/* Round sin down to get a lower bound of |sin(Pi*s/2)|. */
mpfr_sin (y, y, rnd);
}
else /* -2k-1 < s/2 < -2k; sin(Pi*s/2) < 0 */
{
/* Round sin up to get a lower bound of |sin(Pi*s/2)|. */
mpfr_sin (y, y, MPFR_INVERT_RND(rnd));
mpfr_abs (y, y, MPFR_RNDN); /* exact, rounding mode doesn't matter */
}
mpz_clear (sint);
/* now y <= |sin(Pi*s/2)| when rnd=RNDD, y >= |sin(Pi*s/2)| when rnd=RNDU */
mpfr_zeta_pos (z, s1, rnd); /* zeta(1-s) */
mpfr_mul (z, z, y, rnd);
/* now z <= |sin(Pi*s/2)|*zeta(1-s) */
mpfr_log (z, z, rnd);
/* now z <= log(|sin(Pi*s/2)|*zeta(1-s)) */
mpfr_lngamma (y, s1, rnd);
mpfr_add (z, z, y, rnd);
/* z <= lngamma(1-s) + log(|sin(Pi*s/2)|*zeta(1-s)) */
/* since s-1 < 0, we want to round log(2*pi) upwards */
mpfr_mul_2ui (y, p, 1, MPFR_INVERT_RND(rnd));
mpfr_log (y, y, MPFR_INVERT_RND(rnd));
mpfr_mul (y, y, s1, MPFR_INVERT_RND(rnd));
mpfr_sub (z, z, y, rnd);
mpfr_exp (z, z, rnd);
if (rnd == MPFR_RNDD)
mpfr_nextbelow (p); /* restore original p */
}
