double
__ieee754_exp10 (double arg)
{
int32_t lx;
double arg_high, arg_low;
double exp_high, exp_low;
if (!isfinite (arg))
return __ieee754_exp (arg);
if (arg < DBL_MIN_10_EXP - DBL_DIG - 10)
return DBL_MIN * DBL_MIN;
else if (arg > DBL_MAX_10_EXP + 1)
return DBL_MAX * DBL_MAX;
else if (fabs (arg) < 0x1p-56)
return 1.0;
GET_LOW_WORD (lx, arg);
lx &= 0xf8000000;
arg_high = arg;
SET_LOW_WORD (arg_high, lx);
arg_low = arg - arg_high;
exp_high = arg_high * log10_high;
exp_low = arg_high * log10_low + arg_low * M_LN10;
return __ieee754_exp (exp_high) * __ieee754_exp (exp_low);
}
/*
* Fused multiply-add: Compute x * y + z with a single rounding error.
*
* A double has more than twice as much precision than a float, so
* direct double-precision arithmetic suffices, except where double
* rounding occurs.
*/
DLLEXPORT float
fmaf(float x, float y, float z)
{
double xy, result;
u_int32_t hr, lr;
xy = (double)x * y;
result = xy + z;
EXTRACT_WORDS(hr, lr, result);
/* Common case: The double precision result is fine. */
if ((lr & 0x1fffffff) != 0x10000000 || /* not a halfway case */
(hr & 0x7ff00000) == 0x7ff00000 || /* NaN */
result - xy == z || /* exact */
fegetround() != FE_TONEAREST) /* not round-to-nearest */
return (result);
/*
* If result is inexact, and exactly halfway between two float values,
* we need to adjust the low-order bit in the direction of the error.
*/
fesetround(FE_TOWARDZERO);
volatile double vxy = xy; /* XXX work around gcc CSE bug */
double adjusted_result = vxy + z;
fesetround(FE_TONEAREST);
if (result == adjusted_result)
SET_LOW_WORD(adjusted_result, lr + 1);
return (adjusted_result);
}
double __tan(double x, double y, int odd)
{
double_t z, r, v, w, s, a;
double w0, a0;
uint32_t hx;
int big, sign;
GET_HIGH_WORD(hx,x);
big = (hx&0x7fffffff) >= 0x3FE59428; /* |x| >= 0.6744 */
if (big) {
sign = hx>>31;
if (sign) {
x = -x;
y = -y;
}
x = (pio4 - x) + (pio4lo - y);
y = 0.0;
}
z = x * x;
w = z * z;
/*
* Break x^5*(T[1]+x^2*T[2]+...) into
* x^5(T[1]+x^4*T[3]+...+x^20*T[11]) +
* x^5(x^2*(T[2]+x^4*T[4]+...+x^22*[T12]))
*/
r = T[1] + w*(T[3] + w*(T[5] + w*(T[7] + w*(T[9] + w*T[11]))));
v = z*(T[2] + w*(T[4] + w*(T[6] + w*(T[8] + w*(T[10] + w*T[12])))));
s = z * x;
r = y + z*(s*(r + v) + y) + s*T[0];
w = x + r;
if (big) {
s = 1 - 2*odd;
v = s - 2.0 * (x + (r - w*w/(w + s)));
return sign ? -v : v;
}
if (!odd)
return w;
/* -1.0/(x+r) has up to 2ulp error, so compute it accurately */
w0 = w;
SET_LOW_WORD(w0, 0);
v = r - (w0 - x); /* w0+v = r+x */
a0 = a = -1.0 / w;
SET_LOW_WORD(a0, 0);
return a0 + a*(1.0 + a0*w0 + a0*v);
}
double acos(double x)
{
double z, p, q, r, w, s, c, df;
s32_t hx, ix;
GET_HIGH_WORD(hx,x);
ix = hx & 0x7fffffff;
if (ix >= 0x3ff00000)
{
u32_t lx;
GET_LOW_WORD(lx,x);
if (((ix - 0x3ff00000) | lx) == 0)
{
if (hx > 0)
return 0.0;
else
return pi + 2.0 * pio2_lo;
}
return (x - x) / (x - x);
}
if (ix < 0x3fe00000)
{
if (ix <= 0x3c600000)
return pio2_hi + pio2_lo;
z = x * x;
p = z * (pS0 + z * (pS1 + z * (pS2 + z * (pS3 + z * (pS4 + z * pS5)))));
q = one + z * (qS1 + z * (qS2 + z * (qS3 + z * qS4)));
r = p / q;
return pio2_hi - (x - (pio2_lo - x * r));
}
else if (hx < 0)
{
z = (one + x) * 0.5;
p = z * (pS0 + z * (pS1 + z * (pS2 + z * (pS3 + z * (pS4 + z * pS5)))));
q = one + z * (qS1 + z * (qS2 + z * (qS3 + z * qS4)));
s = sqrt(z);
r = p / q;
w = r * s - pio2_lo;
return pi - 2.0 * (s + w);
}
else
{
z = (one - x) * 0.5;
s = sqrt(z);
df = s;
SET_LOW_WORD(df,0);
c = (z - df * df) / (s + df);
p = z * (pS0 + z * (pS1 + z * (pS2 + z * (pS3 + z * (pS4 + z * pS5)))));
q = one + z * (qS1 + z * (qS2 + z * (qS3 + z * qS4)));
r = p / q;
w = r * s + c;
return 2.0 * (df + w);
}
}
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;
}
double asin(double x)
{
double t=0.0,w,p,q,c,r,s;
int32_t hx,ix;
GET_HIGH_WORD(hx, x);
ix = hx & 0x7fffffff;
if (ix >= 0x3ff00000) { /* |x|>= 1 */
uint32_t lx;
GET_LOW_WORD(lx, x);
if ((ix-0x3ff00000 | lx) == 0)
/* asin(1) = +-pi/2 with inexact */
return x*pio2_hi + x*pio2_lo;
return (x-x)/(x-x); /* asin(|x|>1) is NaN */
} else if (ix < 0x3fe00000) { /* |x|<0.5 */
if (ix < 0x3e500000) { /* if |x| < 2**-26 */
if (huge+x > 1.0)
return x; /* return x with inexact if x!=0*/
}
t = x*x;
p = t*(pS0+t*(pS1+t*(pS2+t*(pS3+t*(pS4+t*pS5)))));
q = 1.0+t*(qS1+t*(qS2+t*(qS3+t*qS4)));
w = p/q;
return x + x*w;
}
/* 1 > |x| >= 0.5 */
w = 1.0 - fabs(x);
t = w*0.5;
p = t*(pS0+t*(pS1+t*(pS2+t*(pS3+t*(pS4+t*pS5)))));
q = 1.0+t*(qS1+t*(qS2+t*(qS3+t*qS4)));
s = sqrt(t);
if (ix >= 0x3FEF3333) { /* if |x| > 0.975 */
w = p/q;
t = pio2_hi-(2.0*(s+s*w)-pio2_lo);
} else {
w = s;
SET_LOW_WORD(w,0);
c = (t-w*w)/(s+w);
r = p/q;
p = 2.0*s*r-(pio2_lo-2.0*c);
q = pio4_hi - 2.0*w;
t = pio4_hi - (p-q);
}
if (hx > 0)
return t;
return -t;
}
double
__ieee754_acos(double x)
{
double z, p, q, r, w, s, c, df;
int32_t hx, ix;
GET_HIGH_WORD(hx, x);
ix = hx & 0x7fffffff;
if (ix >= 0x3ff00000) { /* |x| >= 1 */
uint32_t lx;
GET_LOW_WORD(lx, x);
if (((ix - 0x3ff00000) | lx) == 0) { /* |x|==1 */
if (hx>0) return 0.0; /* acos(1) = 0 */
else return pi + 2.0*pio2_lo; /* acos(-1)= pi */
}
return (x - x) / (x - x); /* acos(|x|>1) is NaN */
}
if (ix<0x3fe00000) { /* |x| < 0.5 */
if (ix <= 0x3c600000) return pio2_hi + pio2_lo;/*if|x|<2**-57*/
z = x*x;
p = z*(pS0 + z*(pS1 + z*(pS2 + z*(pS3 + z*(pS4 + z*pS5)))));
q = one + z*(qS1 + z*(qS2 + z*(qS3 + z*qS4)));
r = p / q;
return pio2_hi - (x - (pio2_lo - x*r));
}
else if (hx<0) { /* x < -0.5 */
z = (one + x)*0.5;
p = z*(pS0 + z*(pS1 + z*(pS2 + z*(pS3 + z*(pS4 + z*pS5)))));
q = one + z*(qS1 + z*(qS2 + z*(qS3 + z*qS4)));
s = sqrt(z);
r = p / q;
w = r*s - pio2_lo;
return pi - 2.0*(s + w);
}
else { /* x > 0.5 */
z = (one - x)*0.5;
s = sqrt(z);
df = s;
SET_LOW_WORD(df, 0);
c = (z - df*df) / (s + df);
p = z*(pS0 + z*(pS1 + z*(pS2 + z*(pS3 + z*(pS4 + z*pS5)))));
q = one + z*(qS1 + z*(qS2 + z*(qS3 + z*qS4)));
r = p / q;
w = r*s + c;
return 2.0*(df + w);
}
}
double __builtin_cbrt(double x)
{
int hx;
double r,s,t=0.0,w;
unsigned sign;
hx = GET_HI(x); /* high word of x */
sign=hx&0x80000000; /* sign= sign(x) */
hx ^=sign;
if(hx>=0x7ff00000) return(x); /* cbrt(NaN,INF) is itself */
if((hx|GET_LO(x))==0)
return(x); /* cbrt(0) is itself */
SET_HIGH_WORD(x, hx); /* x <- |x| */
/* rough cbrt to 5 bits */
if(hx<0x00100000) /* subnormal number */
{SET_HIGH_WORD(t, 0x43500000); /* set t= 2**54 */
t*=x; SET_HIGH_WORD(t,GET_HI(t)/3+B2);
}
else
SET_HIGH_WORD(t,hx/3+B1);
/* new cbrt to 23 bits, may be implemented in single precision */
r=t*t/x;
s=C+r*t;
t*=G+F/(s+E+D/s);
/* chopped to 20 bits and make it larger than cbrt(x) */
SET_LOW_WORD(t,0); SET_HIGH_WORD(t, GET_HI(t) + 0x00000001);
/* one step newton iteration to 53 bits with error less than 0.667 ulps */
s=t*t; /* t*t is exact */
r=x/s;
w=t+t;
r=(r-t)/(w+r); /* r-s is exact */
t=t+t*r;
/* retore the sign bit */
SET_HIGH_WORD(t, GET_HI(t) | sign);
return(t);
}
double __hide_ieee754_asin(double x)
{
double t=0.0,w,p,q,c,r,s;
int hx,ix;
hx = GET_HI(x);
ix = hx&0x7fffffff;
if(ix>= 0x3ff00000) { /* |x|>= 1 */
if(((ix-0x3ff00000)|GET_LO(x))==0)
/* asin(1)=+-pi/2 with inexact */
return x*pio2_hi+x*pio2_lo;
return __builtin_nan(""); /* asin(|x|>1) is NaN */
} else if (ix<0x3fe00000) { /* |x|<0.5 */
if(ix<0x3e400000) { /* if |x| < 2**-27 */
if(huge+x>one) return x;/* return x with inexact if x!=0*/
} else
t = x*x;
p = t*(pS0+t*(pS1+t*(pS2+t*(pS3+t*(pS4+t*pS5)))));
q = one+t*(qS1+t*(qS2+t*(qS3+t*qS4)));
w = p/q;
return x+x*w;
}
/* 1> |x|>= 0.5 */
w = one-__builtin_fabs(x);
t = w*0.5;
p = t*(pS0+t*(pS1+t*(pS2+t*(pS3+t*(pS4+t*pS5)))));
q = one+t*(qS1+t*(qS2+t*(qS3+t*qS4)));
s = __builtin_sqrt(t);
if(ix>=0x3FEF3333) { /* if |x| > 0.975 */
w = p/q;
t = pio2_hi-(2.0*(s+s*w)-pio2_lo);
} else {
w = s;
SET_LOW_WORD(w, 0);
c = (t-w*w)/(s+w);
r = p/q;
p = 2.0*s*r-(pio2_lo-2.0*c);
q = pio4_hi-2.0*w;
t = pio4_hi-(p-q);
}
if(hx>0) return t; else return -t;
}
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;
}
