infinity() blocks use of fast math options

[Objexx]

The use of std::numeric_limits<T>::infinity() in parse_number.h is incompatible with "fast math" options.

Perhaps if __FAST_MATH__ is defined std::numeric_limits<T>::max() or, if negative, std::numeric_limits<T>::lowest() would suffice for the returned value? Or maybe some other mechanism that elides/replaces parse_infnan?

I think this is messy and non-ideal but some code that wants fast float will also want fast math ;-). This is more of a feature request than a bug report.

Thanks for the great library!

[lemire]

Would you consider issuing a pull request?

I think that supporting fast math is reasonable as an option that is off by default.

[Objexx]

Accommodating "fast math" opens up a can of worms. I don't think it makes sense to detect "Inf" or "Nan" and emit IEEE infinity or NaN values since fast math builds don't detect/handle those values. And turning Inf into max() or lowest() values might work for some applications but not others. Probably the only sensible thing to do is to remove calls to parse_infnan and use the invalid_argument path in from_chars_float_advanced if __FAST_MATH__ is defined.

This would also require a bit of documentation on the limitations when using fast math and the need to define __FAST_MATH__ explicitly in your builds if the C++ compiler doesn't do it for you.

Input/ideas/feedback are welcomed.

[IRainman]

You can try to use my fork, https://github.com/IRainman/fast_float it already fixed this and many more other problems and improved performance significantly.