/**
Cast to 64 bit int (assuming double & long long have same endianness),
convert from sign-magnitude to 2s-complement,
return (magnitude of) difference in ULPs.
The difference between 2 64-bit numbers is a 65-bit number, so only get the
magnitude of the difference. The usual floating point comparisons (<code>a > b</code>, etc)
can be used to find which is bigger.
This currently uses isnan and isinf, if they are not defined they should be
implemented using bitwise comparisons. isnan must not use <code>x != x</code>, or eqivalent.
If no 64bit unsigned int is provided then it becomes very complicated to implement this function.
\param a,b Any 64-bit non-denormalised floating point numbers.
\return The unsigned distance between <code>a</code> and <code>b</code> in 'floating point space'.
*/
uint64_t cbf_ULP64(const double a, const double b)
{
/*
Check for NAN or INF, ignore subnormals completely.
Don't use 'x!=x' to test for NANs, it might be optimised out,
test for a bit pattern manually if 'isnan' or 'isinf' can't be used.
*/
if (isnan64(a) || isinf64(a) || isnan64(b) || isinf64(b)) {
if (isinf64(a) && isinf64(b) && a==b) return 0;
return 0xffffffffffffffffl;
} else {
const uint64_t signmask = 0x8000000000000000l;
/* cast numbers so that ia >= ib */
uint64_t ia = *(int64_t*)(a>b ? &a : &b);
uint64_t ib = *(int64_t*)(a>b ? &b : &a);
/* convert to 2's complement form */
ia = signmask&ia ? signmask^(~ia+1) : ia;
ib = signmask&ib ? signmask^(~ib+1) : ib;
/* subtract, knowing that ia >= ib */
return ia-ib;
}
}
_f_int2
_IEEE_EXPONENT_I2_H(_f_real8 x)
{
int i;
REGISTER_8 s1;
/* if x is a NaN, return HUGE */
if (isnan64(x))
return(HUGE_INT2_F90);
s1.f = x;
/* clear sign bit. */
s1.ui &= ~IEEE_64_SIGN_BIT;
/* if x is + or minus infinity, return a HUGE */
if (s1.ui == IEEE_64_INFINITY)
return(HUGE_INT2_F90);
/* if x is zero, return -HUGE */
if (x == 0.0e0)
return(-HUGE_INT2_F90);
i = __ieee_real4i(x);
return ((_f_int2) i);
}
/* _IEEE_EXPONENT_I6_R - IEEE EXPONENT returns the exponent part of the
* 64-bit argument in 46-bit integer.
*/
_f_int6
_IEEE_EXPONENT_I6_R(_f_real8 x)
{
int i;
REGISTER_8 s1, s2;
/* if x is a NaN, return HUGE */
if (isnan64(x))
return HUGE_INT6_F90;
s1.f = x;
/* clear sign bit */
s1.ui &= ~IEEE_64_SIGN_BIT;
/* if x is + or - infinity, return HUGE */
if (s1.ui == IEEE_64_INFINITY)
return HUGE_INT6_F90;
/* if x is zero, return -HUGE */
if (x == 0.0e0)
return -HUGE_INT6_F90;
/* shift exponent bits to right. */
s1.ui >>= IEEE_64_MANT_BITS;
if (s1.ui == 0) {
/* x is a subnormal number (implicit leading bit is zero
* and the exponent is zero). Calculate the exponent
* based on normalized x.
*
* get mantissa
*/
s2.f = x;
s2.ui = IEEE_64_MANTISSA & s2.ui;
/* get leading zeros in mantissa part. */
i = _leadz8(s2.ui) - IEEE_64_EXPO_BITS;
/* calculate exponent. */
s1.i -= (IEEE_64_EXPO_BIAS + i);
} else {
/* subtract exponent bias. */
s1.i -= (IEEE_64_EXPO_BIAS);
}
return (_f_int6) s1.i;
}
inline static /* for the inline version of this function. */
#endif
_f_real8
_FRACTION(_f_real8 x)
{
int lz;
REGISTER_8 s1, sign_bit, mantissa, exponent;
if (x == 0.0)
return 0.0;
/* if x is either infinity or a NaN, return a Nan */
if (x == IEEE_64_INFINITY)
return _SGL_NaN;
if (isnan64(x))
return x;
s1.f = x;
/* get sign bit. */
sign_bit.ui = IEEE_64_SIGN_BIT & s1.ui;
/* get mantissa. */
mantissa.ui = IEEE_64_MANTISSA & s1.ui;
/* get exponent. */
exponent.ui = IEEE_64_EXPONENT & s1.ui;
if (exponent.ui == LL_CONST(0x0)) {
/* 1. number is subnormal. normalize mantissa.
* Get leading zeros of mantissa.
*/
lz = _leadz8(mantissa.ui) - IEEE_64_EXPO_BITS;
/* 2. normalize by shifting out all leading zeros
* and first 1 bit.
*/
mantissa.ui = (mantissa.ui << lz) & IEEE_64_MANTISSA;
}
/* position exponent bias less the implicit bit. */
exponent.ui = IEEE_64_EXPO_BIAS - 1;
exponent.ui = exponent.ui << IEEE_64_MANT_BITS;
/* extract fraction. */
s1.ui = exponent.ui | mantissa.ui | sign_bit.ui;
return s1.f;
}
_f_real8
_IEEE_BINARY_SCALE_I8(_f_real8 orig_number, _f_int8 orig_scale_factor)
{
int lz, lzm, shift;
REGISTER_8 number, scale_factor, expon, exponent,
orig_mantissa, mantissa, sign_bit;
/* if x is a NaN, return a Nan */
if (isnan64(orig_number))
return orig_number;
number.f = orig_number;
/* extract sign bit. */
sign_bit.ui = IEEE_64_SIGN_BIT & number.ui;
/* extract exponent. */
exponent.ui = IEEE_64_EXPONENT & number.ui;
/* extract mantissa. */
orig_mantissa.ui = IEEE_64_MANTISSA & number.ui;
/* if x is + or -infinity, return same infinity */
number.ui &= ~IEEE_64_SIGN_BIT;
if (number.ui == IEEE_64_INFINITY)
return orig_number;
/* if x is zero, return zero. */
if (orig_number == 0)
return orig_number;
scale_factor.i = orig_scale_factor;
/* check for unnormalized number */
if (exponent.ui == 0) {
/* Denormmal. Get leading zeros in original mantissa */
lz = _leadz8(orig_mantissa.ui);
if (scale_factor.i > 0) {
/* Scale number by a positive power of 2. The
* result may need to be normalized. Get
* leading zeros in mantissa = lzm.
*/
lzm = lz - IEEE_64_EXPO_BITS - 1;
/* Any leading zeros in mantissa? */
if (lzm > 0) {
/* check if number of leading zeros in
* mantissa allows scaling by shifting.
*/
if (scale_factor.i <= lzm) {
/* Enough lead zeros to shift.
* Exponent is unaffected.
*/
shift = scale_factor.i;
expon.i = 0;
} else {
/* Scale by shifting what we can
* and adjust exponent for rest.
* The result is a normalized
* number.
*/
shift = lzm + 1;
expon.i = scale_factor.i - lzm;
}
/* No leading zeros in mantissa. */
} else {
/* Shift by 1 to normalize mantissa. Do
* rest of scaling through exponent.
*/
shift = 1;
expon.i = scale_factor.i;
}
/* position the exponent. */
exponent.ui = expon.ui << IEEE_64_MANT_BITS;
/* scale the mantissa. */
mantissa.ui = orig_mantissa.ui << shift;
/* Scale_factor LE 0. */
} else {
/* scale mantissa. */
mantissa.ui = orig_mantissa.ui >> (-scale_factor.i);
if ((-scale_factor.ui != 0) &&
(orig_mantissa.ui & (0x1 <<
(-scale_factor.i - 1)))) {
mantissa.ui++;
}
}
/* mask out any bits that may have shifted to the left of
* the mantissa area.
*/
mantissa.ui &= IEEE_64_MANTISSA;
/* OR the new mantissa, the new exponent, and the original
* sign bit together to create the result.
*/
number.ui = mantissa.ui | exponent.ui | sign_bit.ui;
} else {
_f_real8
_IEEE_REMAINDER_R_R(_f_real8 argx, _f_real8 argy)
{
/* Union defined to work with IEEE 64-bit floating point. */
union _ieee_double {
struct {
#if __BYTE_ORDER == __LITTLE_ENDIAN
unsigned int mantissa_lo : IEEE_64_MANT_BITS_H2;
unsigned int mantissa_up : IEEE_64_MANT_BITS_H1;
unsigned int exponent : IEEE_64_EXPO_BITS;
unsigned int sign : 1;
#else
unsigned int sign : 1;
unsigned int exponent : IEEE_64_EXPO_BITS;
unsigned int mantissa_up : IEEE_64_MANT_BITS_H1;
unsigned int mantissa_lo : IEEE_64_MANT_BITS_H2;
#endif
} parts1;
_f_real8 dword;
unsigned int lword[2];
unsigned long long llword;
long long int64;
};
#if __BYTE_ORDER == __LITTLE_ENDIAN
const int lword_hi = 1;
const int lword_lo = 0;
#else
const int lword_hi = 0;
const int lword_lo = 1;
#endif
_f_real8 scalet, scaley;
unsigned int sign_x = 0X80000000;
unsigned int even_x = 0X00000001;
union _ieee_double x_val, y_val, tdiv, two_52, evenchk, div_52;
union _ieee_double nearint, y_up, y_lo, tdiv_up, tdiv_lo, res;
x_val.dword = argx;
y_val.dword = argy;
two_52.lword[lword_hi] = 0x43300000; /* 2**52 */
two_52.lword[lword_lo] = 0x00000000;
div_52.lword[lword_hi] = 0x3CB00000; /* 2**-52 */
div_52.lword[lword_lo] = 0x00000000;
/* check input values: x for infinity and y for zero. */
if (((x_val.llword & ~IEEE_64_SIGN_BIT) == IEEE_64_INFINITY) ||
((!(isnan64(y_val.dword))) && (y_val.dword == 0.0))) {
union _ieee_double x_val;
_f_real8 result8;
_f_real8 arg8 = 0.0;
x_val.llword = _SGL_NaN;
/* need to emit invalid exception */
result8 = _raisinvld(arg8, arg8);
return(x_val.dword);
}
tdiv.dword = argx / argy;
/* check for 2**52 or greater = already integer */
if ((int) (tdiv.lword[lword_hi] & (~sign_x)) < (int) two_52.lword[lword_hi]) {
/* calculate fraction */
evenchk.dword =
tdiv.dword - (_f_real8)((_f_int8)tdiv.dword);
if (tdiv.dword < 0.0) {
nearint.int64 = (_f_int8) (tdiv.dword - 0.5);
if ((evenchk.dword == -0.5) &&
((nearint.lword[lword_lo] & even_x) != 0))
nearint.int64 += 1;
} else {
nearint.int64 = (_f_int8) (tdiv.dword + 0.5);
if ((evenchk.dword == 0.5) &&
((nearint.lword[lword_lo] & even_x) != 0))
nearint.int64 -= 1;
}
tdiv.dword = (_f_real8) nearint.int64;
}
/* Calculate upper and lower for y and tdiv. */
y_up.dword = y_val.dword;
tdiv_up.dword = tdiv.dword;
y_up.parts1.mantissa_lo = 0x0;
tdiv_up.parts1.mantissa_lo = 0x0;
scalet = 1.0;
scaley = 1.0;
/* If tdiv_up exponent < 27, scale up to prevent underflow. */
if ((int) tdiv.parts1.exponent < 27) {
tdiv_lo.dword = (tdiv.dword * two_52.dword) -
(tdiv_up.dword * two_52.dword);
scalet = div_52.dword;
} else
tdiv_lo.dword = tdiv.dword - tdiv_up.dword;
/* If y_up exponent < 27, scale up to prevent underflow. */
if ((int) y_val.parts1.exponent < 27) {
y_lo.dword = (y_val.dword * two_52.dword) -
(y_val.dword * two_52.dword);
scaley = div_52.dword;
} else
y_lo.dword = y_val.dword - y_up.dword;
/* algorithm for ieee for x - (x/y)*y. */
res.dword = ((((x_val.dword - (tdiv_up.dword * y_up.dword)) -
(scaley * (tdiv_up.dword * y_lo.dword))) -
(scalet * (tdiv_lo.dword * y_up.dword))) -
(scalet * scaley * (tdiv_lo.dword * y_lo.dword)));
if (res.dword == 0.0)
res.parts1.sign = x_val.parts1.sign;
return(res.dword);
}
_f_log8 _IEEE_IS_NAN_L8( _f_real8 x)
{
/* if x is NaN, return TRUE */
return ((_f_log8) _btol(isnan64(x)));
}
