ftoa fails (and unsafe) for floats > INT32_MAX

[jeromekelleher]

The decimal value for FLT_MAX is 340282346638528859811704183484516925440, which is much larger than we currently support. We need some way to either support this, or to protect the underlying C code.

[jeromekelleher]

For reference, htslib uses kputd which is quite a bit more complex.

I think the right approach here is to use the existing code for the common cases, and fallback onto sprintf("%f") when the magnitude of the value is outside the range we want to deal with. Detecting this will be fiddly, though.

@tomwhite - what was the rationale for limiting the int32 + three digits in the numba implementation?

[tomwhite]

Could we eventually use/follow kputd?

what was the rationale for limiting the int32 + three digits in the numba implementation?

It was basically covering the common cases - leaving open the option of a wider range of magnitudes (scientific notation) for later, see sgkit-dev/sgkit#966 (and sgkit-dev/sgkit#953).

[jeromekelleher]

Adapting kputd will have it's own issues probably, but yeah, that may be the simplest route.