float128 EvalPoly(float128 x, float128 *arr, unsigned n)
{
float128 x2 = float128_mul(x, x);
unsigned i;
assert(n > 1);
float128 r1 = arr[--n];
i = n;
while(i >= 2) {
r1 = float128_mul(r1, x2);
i -= 2;
r1 = float128_add(r1, arr[i]);
}
if (i) r1 = float128_mul(r1, x);
float128 r2 = arr[--n];
i = n;
while(i >= 2) {
r2 = float128_mul(r2, x2);
i -= 2;
r2 = float128_add(r2, arr[i]);
}
if (i) r2 = float128_mul(r2, x);
return float128_add(r1, r2);
}
long_double __mulxf3(long_double A, long_double B)
{
#if __SIZEOF_LONG_DOUBLE__ == 12
return floatx80_mul(A, B);
#else
return float128_mul(A, B);
#endif
}
static float128 poly_l2p1(float128 x)
{
/* using float128 for approximation */
float128 x_p2 = float128_add(x, float128_two);
x = float128_div(x, x_p2);
x = poly_ln(x);
x = float128_mul(x, float128_ln2inv2);
return x;
}
/* required sqrt(2)/2 < x < sqrt(2) */
static float128 poly_l2(float128 x)
{
/* using float128 for approximation */
float128 x_p1 = float128_add(x, float128_one);
float128 x_m1 = float128_sub(x, float128_one);
x = float128_div(x_m1, x_p1);
x = poly_ln(x);
x = float128_mul(x, float128_ln2inv2);
return x;
}
/* 128-bit FP multiplication RR */
void HELPER(mxbr)(CPUS390XState *env, uint32_t f1, uint32_t f2)
{
CPU_QuadU v1;
CPU_QuadU v2;
CPU_QuadU res;
v1.ll.upper = env->fregs[f1].ll;
v1.ll.lower = env->fregs[f1 + 2].ll;
v2.ll.upper = env->fregs[f2].ll;
v2.ll.lower = env->fregs[f2 + 2].ll;
res.q = float128_mul(v1.q, v2.q, &env->fpu_status);
env->fregs[f1].ll = res.ll.upper;
env->fregs[f1 + 2].ll = res.ll.lower;
}
float128 OddPoly(float128 x, float128 *arr, unsigned n)
{
return float128_mul(x, EvenPoly(x, arr, n));
}
float128 EvenPoly(float128 x, float128 *arr, unsigned n)
{
return EvalPoly(float128_mul(x, x), arr, n);
}
floatx80 fpatan(floatx80 a, floatx80 b, float_status_t &status)
{
// handle unsupported extended double-precision floating encodings
if (floatx80_is_unsupported(a) || floatx80_is_unsupported(b)) {
float_raise(status, float_flag_invalid);
return floatx80_default_nan;
}
Bit64u aSig = extractFloatx80Frac(a);
Bit32s aExp = extractFloatx80Exp(a);
int aSign = extractFloatx80Sign(a);
Bit64u bSig = extractFloatx80Frac(b);
Bit32s bExp = extractFloatx80Exp(b);
int bSign = extractFloatx80Sign(b);
int zSign = aSign ^ bSign;
if (bExp == 0x7FFF)
{
if ((Bit64u) (bSig<<1))
return propagateFloatx80NaN(a, b, status);
if (aExp == 0x7FFF) {
if ((Bit64u) (aSig<<1))
return propagateFloatx80NaN(a, b, status);
if (aSign) { /* return 3PI/4 */
return roundAndPackFloatx80(80, bSign,
FLOATX80_3PI4_EXP, FLOAT_3PI4_HI, FLOAT_3PI4_LO, status);
}
else { /* return PI/4 */
return roundAndPackFloatx80(80, bSign,
FLOATX80_PI4_EXP, FLOAT_PI_HI, FLOAT_PI_LO, status);
}
}
if (aSig && (aExp == 0))
float_raise(status, float_flag_denormal);
/* return PI/2 */
return roundAndPackFloatx80(80, bSign, FLOATX80_PI2_EXP, FLOAT_PI_HI, FLOAT_PI_LO, status);
}
if (aExp == 0x7FFF)
{
if ((Bit64u) (aSig<<1))
return propagateFloatx80NaN(a, b, status);
if (bSig && (bExp == 0))
float_raise(status, float_flag_denormal);
return_PI_or_ZERO:
if (aSign) { /* return PI */
return roundAndPackFloatx80(80, bSign, FLOATX80_PI_EXP, FLOAT_PI_HI, FLOAT_PI_LO, status);
} else { /* return 0 */
return packFloatx80(bSign, 0, 0);
}
}
if (bExp == 0)
{
if (bSig == 0) {
if (aSig && (aExp == 0)) float_raise(status, float_flag_denormal);
goto return_PI_or_ZERO;
}
float_raise(status, float_flag_denormal);
normalizeFloatx80Subnormal(bSig, &bExp, &bSig);
}
if (aExp == 0)
{
if (aSig == 0) /* return PI/2 */
return roundAndPackFloatx80(80, bSign, FLOATX80_PI2_EXP, FLOAT_PI_HI, FLOAT_PI_LO, status);
float_raise(status, float_flag_denormal);
normalizeFloatx80Subnormal(aSig, &aExp, &aSig);
}
float_raise(status, float_flag_inexact);
/* |a| = |b| ==> return PI/4 */
if (aSig == bSig && aExp == bExp)
return roundAndPackFloatx80(80, bSign, FLOATX80_PI4_EXP, FLOAT_PI_HI, FLOAT_PI_LO, status);
/* ******************************** */
/* using float128 for approximation */
/* ******************************** */
float128 a128 = normalizeRoundAndPackFloat128(0, aExp-0x10, aSig, 0, status);
float128 b128 = normalizeRoundAndPackFloat128(0, bExp-0x10, bSig, 0, status);
float128 x;
int swap = 0, add_pi6 = 0, add_pi4 = 0;
if (aExp > bExp || (aExp == bExp && aSig > bSig))
{
x = float128_div(b128, a128, status);
}
else {
x = float128_div(a128, b128, status);
swap = 1;
}
Bit32s xExp = extractFloat128Exp(x);
if (xExp <= EXP_BIAS-40)
goto approximation_completed;
if (x.hi >= BX_CONST64(0x3ffe800000000000)) // 3/4 < x < 1
{
/*
arctan(x) = arctan((x-1)/(x+1)) + pi/4
*/
float128 t1 = float128_sub(x, float128_one, status);
float128 t2 = float128_add(x, float128_one, status);
x = float128_div(t1, t2, status);
add_pi4 = 1;
}
else
{
/* argument correction */
if (xExp >= 0x3FFD) // 1/4 < x < 3/4
{
/*
arctan(x) = arctan((x*sqrt(3)-1)/(x+sqrt(3))) + pi/6
*/
float128 t1 = float128_mul(x, float128_sqrt3, status);
float128 t2 = float128_add(x, float128_sqrt3, status);
x = float128_sub(t1, float128_one, status);
x = float128_div(x, t2, status);
add_pi6 = 1;
}
}
x = poly_atan(x, status);
if (add_pi6) x = float128_add(x, float128_pi6, status);
if (add_pi4) x = float128_add(x, float128_pi4, status);
approximation_completed:
if (swap) x = float128_sub(float128_pi2, x, status);
floatx80 result = float128_to_floatx80(x, status);
if (zSign) floatx80_chs(result);
int rSign = extractFloatx80Sign(result);
if (!bSign && rSign)
return floatx80_add(result, floatx80_pi, status);
if (bSign && !rSign)
return floatx80_sub(result, floatx80_pi, status);
return result;
}
