double
__ieee754_atanh (double x)
{
double xa = fabs (x);
double t;
if (isless (xa, 0.5))
{
if (__builtin_expect (xa < 0x1.0p-28, 0))
{
math_force_eval (huge + x);
return x;
}
t = xa + xa;
t = 0.5 * __log1p (t + t * xa / (1.0 - xa));
}
else if (__builtin_expect (isless (xa, 1.0), 1))
t = 0.5 * __log1p ((xa + xa) / (1.0 - xa));
else
{
if (isgreater (xa, 1.0))
return (x - x) / (x - x);
return x / 0.0;
}
return __copysign (t, x);
}
double
__ieee754_atanh (double x)
{
double xa = fabs (x);
double t;
if (isless (xa, 0.5))
{
if (__glibc_unlikely (xa < 0x1.0p-28))
{
math_force_eval (huge + x);
math_check_force_underflow (x);
return x;
}
t = xa + xa;
t = 0.5 * __log1p (t + t * xa / (1.0 - xa));
}
else if (__glibc_likely (isless (xa, 1.0)))
t = 0.5 * __log1p ((xa + xa) / (1.0 - xa));
else
{
if (isgreater (xa, 1.0))
return (x - x) / (x - x);
return x / 0.0;
}
return __copysign (t, x);
}
double
__ieee754_acosh (double x)
{
int64_t hx;
EXTRACT_WORDS64 (hx, x);
if (hx > INT64_C (0x4000000000000000))
{
if (__builtin_expect (hx >= INT64_C (0x41b0000000000000), 0))
{
/* x > 2**28 */
if (hx >= INT64_C (0x7ff0000000000000))
/* x is inf of NaN */
return x + x;
else
return __ieee754_log (x) + ln2;/* acosh(huge)=log(2x) */
}
/* 2**28 > x > 2 */
double t = x * x;
return __ieee754_log (2.0 * x - one / (x + __ieee754_sqrt (t - one)));
}
else if (__builtin_expect (hx > INT64_C (0x3ff0000000000000), 1))
{
/* 1<x<2 */
double t = x - one;
return __log1p (t + __ieee754_sqrt (2.0 * t + t * t));
}
else if (__builtin_expect (hx == INT64_C (0x3ff0000000000000), 1))
return 0.0; /* acosh(1) = 0 */
else /* x < 1 */
return (x - x) / (x - x);
}
double
__asinh (double x)
{
double w;
int32_t hx, ix;
GET_HIGH_WORD (hx, x);
ix = hx & 0x7fffffff;
if (__glibc_unlikely (ix < 0x3e300000)) /* |x|<2**-28 */
{
if (huge + x > one)
return x; /* return x inexact except 0 */
}
if (__glibc_unlikely (ix > 0x41b00000)) /* |x| > 2**28 */
{
if (ix >= 0x7ff00000)
return x + x; /* x is inf or NaN */
w = __ieee754_log (fabs (x)) + ln2;
}
else
{
double xa = fabs (x);
if (ix > 0x40000000) /* 2**28 > |x| > 2.0 */
{
w = __ieee754_log (2.0 * xa + one / (__ieee754_sqrt (xa * xa + one) +
xa));
}
else /* 2.0 > |x| > 2**-28 */
{
double t = xa * xa;
w = __log1p (xa + t / (one + __ieee754_sqrt (one + t)));
}
}
return __copysign (w, x);
}
double log10(double x)
{
double f,hfsq,hi,lo,r,val_hi,val_lo,w,y,y2;
int32_t i,k,hx;
uint32_t lx;
EXTRACT_WORDS(hx, lx, x);
k = 0;
if (hx < 0x00100000) { /* x < 2**-1022 */
if (((hx&0x7fffffff)|lx) == 0)
return -two54/0.0; /* log(+-0)=-inf */
if (hx<0)
return (x-x)/0.0; /* log(-#) = NaN */
/* subnormal number, scale up x */
k -= 54;
x *= two54;
GET_HIGH_WORD(hx, x);
}
if (hx >= 0x7ff00000)
return x+x;
if (hx == 0x3ff00000 && lx == 0)
return 0.0; /* log(1) = +0 */
k += (hx>>20) - 1023;
hx &= 0x000fffff;
i = (hx+0x95f64)&0x100000;
SET_HIGH_WORD(x, hx|(i^0x3ff00000)); /* normalize x or x/2 */
k += i>>20;
y = (double)k;
f = x - 1.0;
hfsq = 0.5*f*f;
r = __log1p(f);
/* See log2.c for details. */
hi = f - hfsq;
SET_LOW_WORD(hi, 0);
lo = (f - hi) - hfsq + r;
val_hi = hi*ivln10hi;
y2 = y*log10_2hi;
val_lo = y*log10_2lo + (lo+hi)*ivln10lo + lo*ivln10hi;
/*
* Extra precision in for adding y*log10_2hi is not strictly needed
* since there is no very large cancellation near x = sqrt(2) or
* x = 1/sqrt(2), but we do it anyway since it costs little on CPUs
* with some parallelism and it reduces the error for many args.
*/
w = y2 + val_hi;
val_lo += (y2 - w) + val_hi;
val_hi = w;
return val_lo + val_hi;
}
Err mathlib_log1p(UInt16 refnum, double x, double *result) {
#pragma unused(refnum)
*result = __log1p(x);
return mlErrNone;
}
double log2(double x)
{
double f,hfsq,hi,lo,r,val_hi,val_lo,w,y;
int32_t i,k,hx;
uint32_t lx;
EXTRACT_WORDS(hx, lx, x);
k = 0;
if (hx < 0x00100000) { /* x < 2**-1022 */
if (((hx&0x7fffffff)|lx) == 0)
return -two54/0.0; /* log(+-0)=-inf */
if (hx < 0)
return (x-x)/0.0; /* log(-#) = NaN */
/* subnormal number, scale up x */
k -= 54;
x *= two54;
GET_HIGH_WORD(hx, x);
}
if (hx >= 0x7ff00000)
return x+x;
if (hx == 0x3ff00000 && lx == 0)
return 0.0; /* log(1) = +0 */
k += (hx>>20) - 1023;
hx &= 0x000fffff;
i = (hx+0x95f64) & 0x100000;
SET_HIGH_WORD(x, hx|(i^0x3ff00000)); /* normalize x or x/2 */
k += i>>20;
y = (double)k;
f = x - 1.0;
hfsq = 0.5*f*f;
r = __log1p(f);
/*
* f-hfsq must (for args near 1) be evaluated in extra precision
* to avoid a large cancellation when x is near sqrt(2) or 1/sqrt(2).
* This is fairly efficient since f-hfsq only depends on f, so can
* be evaluated in parallel with R. Not combining hfsq with R also
* keeps R small (though not as small as a true `lo' term would be),
* so that extra precision is not needed for terms involving R.
*
* Compiler bugs involving extra precision used to break Dekker's
* theorem for spitting f-hfsq as hi+lo, unless double_t was used
* or the multi-precision calculations were avoided when double_t
* has extra precision. These problems are now automatically
* avoided as a side effect of the optimization of combining the
* Dekker splitting step with the clear-low-bits step.
*
* y must (for args near sqrt(2) and 1/sqrt(2)) be added in extra
* precision to avoid a very large cancellation when x is very near
* these values. Unlike the above cancellations, this problem is
* specific to base 2. It is strange that adding +-1 is so much
* harder than adding +-ln2 or +-log10_2.
*
* This uses Dekker's theorem to normalize y+val_hi, so the
* compiler bugs are back in some configurations, sigh. And I
* don't want to used double_t to avoid them, since that gives a
* pessimization and the support for avoiding the pessimization
* is not yet available.
*
* The multi-precision calculations for the multiplications are
* routine.
*/
hi = f - hfsq;
SET_LOW_WORD(hi, 0);
lo = (f - hi) - hfsq + r;
val_hi = hi*ivln2hi;
val_lo = (lo+hi)*ivln2lo + lo*ivln2hi;
/* spadd(val_hi, val_lo, y), except for not using double_t: */
w = y + val_hi;
val_lo += (y - w) + val_hi;
val_hi = w;
return val_lo + val_hi;
}
double log1p( double x ) {
return __log1p( x );
}
__complex__ double
__kernel_casinh (__complex__ double x, int adj)
{
__complex__ double res;
double rx, ix;
__complex__ double y;
/* Avoid cancellation by reducing to the first quadrant. */
rx = fabs (__real__ x);
ix = fabs (__imag__ x);
if (rx >= 1.0 / DBL_EPSILON || ix >= 1.0 / DBL_EPSILON)
{
/* For large x in the first quadrant, x + csqrt (1 + x * x)
is sufficiently close to 2 * x to make no significant
difference to the result; avoid possible overflow from
the squaring and addition. */
__real__ y = rx;
__imag__ y = ix;
if (adj)
{
double t = __real__ y;
__real__ y = __copysign (__imag__ y, __imag__ x);
__imag__ y = t;
}
res = __clog (y);
__real__ res += M_LN2;
}
else if (rx >= 0.5 && ix < DBL_EPSILON / 8.0)
{
double s = __ieee754_hypot (1.0, rx);
__real__ res = __ieee754_log (rx + s);
if (adj)
__imag__ res = __ieee754_atan2 (s, __imag__ x);
else
__imag__ res = __ieee754_atan2 (ix, s);
}
else if (rx < DBL_EPSILON / 8.0 && ix >= 1.5)
{
double s = __ieee754_sqrt ((ix + 1.0) * (ix - 1.0));
__real__ res = __ieee754_log (ix + s);
if (adj)
__imag__ res = __ieee754_atan2 (rx, __copysign (s, __imag__ x));
else
__imag__ res = __ieee754_atan2 (s, rx);
}
else if (ix > 1.0 && ix < 1.5 && rx < 0.5)
{
if (rx < DBL_EPSILON * DBL_EPSILON)
{
double ix2m1 = (ix + 1.0) * (ix - 1.0);
double s = __ieee754_sqrt (ix2m1);
__real__ res = __log1p (2.0 * (ix2m1 + ix * s)) / 2.0;
if (adj)
__imag__ res = __ieee754_atan2 (rx, __copysign (s, __imag__ x));
else
__imag__ res = __ieee754_atan2 (s, rx);
}
else
{
double ix2m1 = (ix + 1.0) * (ix - 1.0);
double rx2 = rx * rx;
double f = rx2 * (2.0 + rx2 + 2.0 * ix * ix);
double d = __ieee754_sqrt (ix2m1 * ix2m1 + f);
double dp = d + ix2m1;
double dm = f / dp;
double r1 = __ieee754_sqrt ((dm + rx2) / 2.0);
double r2 = rx * ix / r1;
__real__ res = __log1p (rx2 + dp + 2.0 * (rx * r1 + ix * r2)) / 2.0;
if (adj)
__imag__ res = __ieee754_atan2 (rx + r1, __copysign (ix + r2,
__imag__ x));
else
__imag__ res = __ieee754_atan2 (ix + r2, rx + r1);
}
}
else if (ix == 1.0 && rx < 0.5)
{
if (rx < DBL_EPSILON / 8.0)
{
__real__ res = __log1p (2.0 * (rx + __ieee754_sqrt (rx))) / 2.0;
if (adj)
__imag__ res = __ieee754_atan2 (__ieee754_sqrt (rx),
__copysign (1.0, __imag__ x));
else
__imag__ res = __ieee754_atan2 (1.0, __ieee754_sqrt (rx));
}
else
{
double d = rx * __ieee754_sqrt (4.0 + rx * rx);
double s1 = __ieee754_sqrt ((d + rx * rx) / 2.0);
double s2 = __ieee754_sqrt ((d - rx * rx) / 2.0);
__real__ res = __log1p (rx * rx + d + 2.0 * (rx * s1 + s2)) / 2.0;
if (adj)
__imag__ res = __ieee754_atan2 (rx + s1, __copysign (1.0 + s2,
__imag__ x));
else
__imag__ res = __ieee754_atan2 (1.0 + s2, rx + s1);
}
}
else if (ix < 1.0 && rx < 0.5)
{
if (ix >= DBL_EPSILON)
{
if (rx < DBL_EPSILON * DBL_EPSILON)
{
double onemix2 = (1.0 + ix) * (1.0 - ix);
double s = __ieee754_sqrt (onemix2);
__real__ res = __log1p (2.0 * rx / s) / 2.0;
if (adj)
__imag__ res = __ieee754_atan2 (s, __imag__ x);
else
__imag__ res = __ieee754_atan2 (ix, s);
}
else
{
double onemix2 = (1.0 + ix) * (1.0 - ix);
double rx2 = rx * rx;
double f = rx2 * (2.0 + rx2 + 2.0 * ix * ix);
double d = __ieee754_sqrt (onemix2 * onemix2 + f);
double dp = d + onemix2;
double dm = f / dp;
double r1 = __ieee754_sqrt ((dp + rx2) / 2.0);
double r2 = rx * ix / r1;
__real__ res
= __log1p (rx2 + dm + 2.0 * (rx * r1 + ix * r2)) / 2.0;
if (adj)
__imag__ res = __ieee754_atan2 (rx + r1,
__copysign (ix + r2,
__imag__ x));
else
__imag__ res = __ieee754_atan2 (ix + r2, rx + r1);
}
}
else
{
double s = __ieee754_hypot (1.0, rx);
__real__ res = __log1p (2.0 * rx * (rx + s)) / 2.0;
if (adj)
__imag__ res = __ieee754_atan2 (s, __imag__ x);
else
__imag__ res = __ieee754_atan2 (ix, s);
}
math_check_force_underflow_nonneg (__real__ res);
}
else
{
__real__ y = (rx - ix) * (rx + ix) + 1.0;
__imag__ y = 2.0 * rx * ix;
y = __csqrt (y);
__real__ y += rx;
__imag__ y += ix;
if (adj)
{
double t = __real__ y;
__real__ y = __copysign (__imag__ y, __imag__ x);
__imag__ y = t;
}
res = __clog (y);
}
/* Give results the correct sign for the original argument. */
__real__ res = __copysign (__real__ res, __real__ x);
__imag__ res = __copysign (__imag__ res, (adj ? 1.0 : __imag__ x));
return res;
}
__complex__ double
__catanh (__complex__ double x)
{
__complex__ double res;
int rcls = fpclassify (__real__ x);
int icls = fpclassify (__imag__ x);
if (__glibc_unlikely (rcls <= FP_INFINITE || icls <= FP_INFINITE))
{
if (icls == FP_INFINITE)
{
__real__ res = __copysign (0.0, __real__ x);
__imag__ res = __copysign (M_PI_2, __imag__ x);
}
else if (rcls == FP_INFINITE || rcls == FP_ZERO)
{
__real__ res = __copysign (0.0, __real__ x);
if (icls >= FP_ZERO)
__imag__ res = __copysign (M_PI_2, __imag__ x);
else
__imag__ res = __nan ("");
}
else
{
__real__ res = __nan ("");
__imag__ res = __nan ("");
}
}
else if (__glibc_unlikely (rcls == FP_ZERO && icls == FP_ZERO))
{
res = x;
}
else
{
if (fabs (__real__ x) >= 16.0 / DBL_EPSILON
|| fabs (__imag__ x) >= 16.0 / DBL_EPSILON)
{
__imag__ res = __copysign (M_PI_2, __imag__ x);
if (fabs (__imag__ x) <= 1.0)
__real__ res = 1.0 / __real__ x;
else if (fabs (__real__ x) <= 1.0)
__real__ res = __real__ x / __imag__ x / __imag__ x;
else
{
double h = __ieee754_hypot (__real__ x / 2.0, __imag__ x / 2.0);
__real__ res = __real__ x / h / h / 4.0;
}
}
else
{
if (fabs (__real__ x) == 1.0
&& fabs (__imag__ x) < DBL_EPSILON * DBL_EPSILON)
__real__ res = (__copysign (0.5, __real__ x)
* (M_LN2 - __ieee754_log (fabs (__imag__ x))));
else
{
double i2 = 0.0;
if (fabs (__imag__ x) >= DBL_EPSILON * DBL_EPSILON)
i2 = __imag__ x * __imag__ x;
double num = 1.0 + __real__ x;
num = i2 + num * num;
double den = 1.0 - __real__ x;
den = i2 + den * den;
double f = num / den;
if (f < 0.5)
__real__ res = 0.25 * __ieee754_log (f);
else
{
num = 4.0 * __real__ x;
__real__ res = 0.25 * __log1p (num / den);
}
}
double absx, absy, den;
absx = fabs (__real__ x);
absy = fabs (__imag__ x);
if (absx < absy)
{
double t = absx;
absx = absy;
absy = t;
}
if (absy < DBL_EPSILON / 2.0)
{
den = (1.0 - absx) * (1.0 + absx);
if (den == -0.0)
den = 0.0;
}
else if (absx >= 1.0)
den = (1.0 - absx) * (1.0 + absx) - absy * absy;
else if (absx >= 0.75 || absy >= 0.5)
den = -__x2y2m1 (absx, absy);
else
den = (1.0 - absx) * (1.0 + absx) - absy * absy;
__imag__ res = 0.5 * __ieee754_atan2 (2.0 * __imag__ x, den);
}
math_check_force_underflow_complex (res);
}
return res;
}
__complex__ double
__clog (__complex__ double x)
{
__complex__ double result;
int rcls = fpclassify (__real__ x);
int icls = fpclassify (__imag__ x);
if (__glibc_unlikely (rcls == FP_ZERO && icls == FP_ZERO))
{
/* Real and imaginary part are 0.0. */
__imag__ result = signbit (__real__ x) ? M_PI : 0.0;
__imag__ result = __copysign (__imag__ result, __imag__ x);
/* Yes, the following line raises an exception. */
__real__ result = -1.0 / fabs (__real__ x);
}
else if (__glibc_likely (rcls != FP_NAN && icls != FP_NAN))
{
/* Neither real nor imaginary part is NaN. */
double absx = fabs (__real__ x), absy = fabs (__imag__ x);
int scale = 0;
if (absx < absy)
{
double t = absx;
absx = absy;
absy = t;
}
if (absx > DBL_MAX / 2.0)
{
scale = -1;
absx = __scalbn (absx, scale);
absy = (absy >= DBL_MIN * 2.0 ? __scalbn (absy, scale) : 0.0);
}
else if (absx < DBL_MIN && absy < DBL_MIN)
{
scale = DBL_MANT_DIG;
absx = __scalbn (absx, scale);
absy = __scalbn (absy, scale);
}
if (absx == 1.0 && scale == 0)
{
__real__ result = __log1p (absy * absy) / 2.0;
math_check_force_underflow_nonneg (__real__ result);
}
else if (absx > 1.0 && absx < 2.0 && absy < 1.0 && scale == 0)
{
double d2m1 = (absx - 1.0) * (absx + 1.0);
if (absy >= DBL_EPSILON)
d2m1 += absy * absy;
__real__ result = __log1p (d2m1) / 2.0;
}
else if (absx < 1.0
&& absx >= 0.5
&& absy < DBL_EPSILON / 2.0
&& scale == 0)
{
double d2m1 = (absx - 1.0) * (absx + 1.0);
__real__ result = __log1p (d2m1) / 2.0;
}
else if (absx < 1.0
&& absx >= 0.5
&& scale == 0
&& absx * absx + absy * absy >= 0.5)
{
double d2m1 = __x2y2m1 (absx, absy);
__real__ result = __log1p (d2m1) / 2.0;
}
else
{
double d = __ieee754_hypot (absx, absy);
__real__ result = __ieee754_log (d) - scale * M_LN2;
}
__imag__ result = __ieee754_atan2 (__imag__ x, __real__ x);
}
else
{
__imag__ result = __nan ("");
if (rcls == FP_INFINITE || icls == FP_INFINITE)
/* Real or imaginary part is infinite. */
__real__ result = HUGE_VAL;
else
__real__ result = __nan ("");
}
return result;
}
__complex__ double
__clog10 (__complex__ double x)
{
__complex__ double result;
int rcls = fpclassify (__real__ x);
int icls = fpclassify (__imag__ x);
if (__builtin_expect (rcls == FP_ZERO && icls == FP_ZERO, 0))
{
/* Real and imaginary part are 0.0. */
__imag__ result = signbit (__real__ x) ? M_PI : 0.0;
__imag__ result = __copysign (__imag__ result, __imag__ x);
/* Yes, the following line raises an exception. */
__real__ result = -1.0 / fabs (__real__ x);
}
else if (__builtin_expect (rcls != FP_NAN && icls != FP_NAN, 1))
{
/* Neither real nor imaginary part is NaN. */
double absx = fabs (__real__ x), absy = fabs (__imag__ x);
int scale = 0;
if (absx < absy)
{
double t = absx;
absx = absy;
absy = t;
}
if (absx > DBL_MAX / 2.0)
{
scale = -1;
absx = __scalbn (absx, scale);
absy = (absy >= DBL_MIN * 2.0 ? __scalbn (absy, scale) : 0.0);
}
else if (absx < DBL_MIN && absy < DBL_MIN)
{
scale = DBL_MANT_DIG;
absx = __scalbn (absx, scale);
absy = __scalbn (absy, scale);
}
if (absx == 1.0 && scale == 0)
{
double absy2 = absy * absy;
if (absy2 <= DBL_MIN * 2.0 * M_LN10)
{
#if __FLT_EVAL_METHOD__ == 0
__real__ result = (absy2 / 2.0 - absy2 * absy2 / 4.0) * M_LOG10E;
#else
volatile double force_underflow = absy2 * absy2 / 4.0;
__real__ result = (absy2 / 2.0 - force_underflow) * M_LOG10E;
#endif
}
else
__real__ result = __log1p (absy2) * (M_LOG10E / 2.0);
}
else if (absx > 1.0 && absx < 2.0 && absy < 1.0 && scale == 0)
{
double d2m1 = (absx - 1.0) * (absx + 1.0);
if (absy >= DBL_EPSILON)
d2m1 += absy * absy;
__real__ result = __log1p (d2m1) * (M_LOG10E / 2.0);
}
else if (absx < 1.0
&& absx >= 0.75
&& absy < DBL_EPSILON / 2.0
&& scale == 0)
{
double d2m1 = (absx - 1.0) * (absx + 1.0);
__real__ result = __log1p (d2m1) * (M_LOG10E / 2.0);
}
else if (absx < 1.0 && (absx >= 0.75 || absy >= 0.5) && scale == 0)
{
double d2m1 = __x2y2m1 (absx, absy);
__real__ result = __log1p (d2m1) * (M_LOG10E / 2.0);
}
else
{
double d = __ieee754_hypot (absx, absy);
__real__ result = __ieee754_log10 (d) - scale * M_LOG10_2;
}
__imag__ result = M_LOG10E * __ieee754_atan2 (__imag__ x, __real__ x);
}
else
{
__imag__ result = __nan ("");
if (rcls == FP_INFINITE || icls == FP_INFINITE)
/* Real or imaginary part is infinite. */
__real__ result = HUGE_VAL;
else
__real__ result = __nan ("");
}
return result;
}
