Free42 possible accuracy flaw

[J-F Garnier]

History may help here.
In classic HP BCD machines, there are these special cases to ensure "nice" results for powers of ten:
- LOG10: check for exact power of 10,
- 10^X: check for integer,
- Y^X: check for y-mantissa=1 and x=integer.

(03-22-2022 10:08 PM)Thomas Okken Wrote:  10^2 = 99.99999999999999999999999999999999

Strangely, the Intel lib power function seems to have more problems with small numbers than with large numbers.
100 0.5 y^x ---> 9.999999999999999999999999999999999
0.01 0.5 y^x --> 9.999999999999999999999999999999998e-2
1E6000 0.5 y^x -> 1.e3000 (exact)
It seems counter-intuitive for me.

[added:]
This is the opposite of HP machines, we start to get slightly inaccurate y^x for large numbers
e.g. 1E480 0.5 y^x --> 9.99999999999E239

J-F