complex sqrt bug

[achan001]

b = xim / (a 2) may underflow to 0.0 if xim is tiny, (a 2) big (if we test signbit(b), underflow is not a problem, but we are not)

If xre < 0, it is possible yim have the wrong sign.

Free42 Decimal: -1.e5000-1.e-5000i SQRT      → 0+1.e2500i      // wrong sign !

Above example, b = -1e-5000 / (1e2500 * 2) = -5e-7501, underflow to 0.0 This is more of an issue for Free42 Binary, with much smaller exponent range.

static int mappable_sqrt_c(phloat xre, phloat xim, phloat *yre, phloat *yim) {
    ...
    phloat r = hypot(xre, xim);
    phloat a = sqrt((r + fabs(xre)) / 2);
    phloat b = xim / (a * 2);
    ...

    if (xre >= 0) {
        *yre = a;
        *yim = b;
    } else if (b >= 0) {     // should test xim instead
        *yre = b;
        *yim = a;
    } else {
        *yre = -b;
        *yim = -a;
    }
    return ERR_NONE;
}

[thomasokken]

Confirmed and fixed: 364d7dd17bc1a3e1ead50d6fc6041c64cf889e87

Thanks!