SQRT of negative numbers

[zx70]

In C the result of SQRT of negative numbers is often forced to zero.
It appears to be the preferred behavior, e.g. it fixes this: https://github.com/z88dk/z88dk-ext/blob/master/graphics/hat.c

The default high precision math library works fine, the target specific ones and the alternative ones might require a calibration.

[feilipu]

For math32 the sqrt() of negative numbers is a NaN, which is as designed.

But I'm not sure NaNs are handled consistently by libraries and compilers if provided as input to a function chain. I'll need to think about that one.

[zx70]

I had to do something for the Sinclair ROMs, it simply interrupted the runtime by jumping to BASIC and giving an "Invalid argument" error.
The example program was originally for the MS BASIC which IMO put zero as well.
I took the opportunity to correct a bug in the "TINY" version of the ZX FP libraries.

[feilipu]

Well a bit of a rough check through shows that NaN values are not generated, except for the sqrt(-n) already discussed.

The log(0) or log(-n) should return a NaN, but is currently returning negative infinity. Perhaps that could be fixed to return a NaN?

There is no NaN included in the C header files to return. Perhaps that should be added (below for IEEE754), and then returned as needed?

#define NAN_POS_F32    ((unsigned long)0x7FFFFFFF)
#define NAN_NEG_F32    ((unsigned long)0xFFFFFFFF)

The acos() and asin() functions should return a NaN when the operator is greater than 1. This could be tested and added too.

But otherwise the fsclassify() function is properly implemented, so if that is used from C then things can work as intended.

[feilipu]

Resolved some weaknesses with #2166.