Free42 possible accuracy flaw

[Werner]

(03-23-2022 08:05 AM)Thomas Okken Wrote:  Estimate for exactness: take the number of significant digits, that is, the mantissa length minus the number of trailing zeroes, and multiply that by the exponent; if <= 34, use squaring?

That doesn't work for single-digits eg 8^35. And even 11^32 < 1e34

Quote:The estimate doesn't have to be sharp: multiply the mantissa by such a power of 10 that you get the smallest possible integer, then perform the repeated-squaring algorithm, and if the result is < 10^35, fix the exponent and return it, and otherwise, let bid128_pow() do its thing.

Yes, even if you then do some superfluous work. But better that than accuracy loss ;-)
Doesn't the lib have an exact flag, like in the 71B? As long as the result is exact, keep on squaring. Once it isn't, go back and do the normal e^(x*LN(y))
Werner