Generate subnormal floats

[Zac-HD]

It turns out that we can generate subnormal values; it's just that we do so with negligible probability because they only occur for one-in-2^11 possible exponents. The obvious workaround is to just add a handful of special values to our NASTY_FLOATS table, and so here's a PR to do just that. Fixes #2976.

[honno]

Noting here that extra.array_api is one use case where we don't want to generate subnormal floats as it's not in scope. Interestingly CuPy handles subnormals differently to builtin Python/NumPy/etc..

Could there be a kwarg like allow_subnormal which defaults to True? Otherwise a recipe in the docs might be helpful, although even then maybe it really is just extra.array_api that cares not to generate these.

st.floats().filter(lambda n: abs(n) >= float_info.min)

[Zac-HD]

Noting here that extra.array_api is one use case where we don't want to generate subnormal floats as it's not in scope. Interestingly CuPy handles subnormals differently to builtin Python/NumPy/etc..

I think this is incorrect - the spec currently says that float dtypes follow IEEE 754, which would require support for subnormals.

Evidently CuPy (/Nvidia hardware) doesn't strictly support IEEE 754, and perhaps the array API standard will be altered to make subnormals optional, but as it stands I consider this a bug in CuPy rather than Hypothesis.

Regardless, I'm open to a new allow_subnormal parameter, but in a different PR to this one since they're already supported.

[honno]

Quite familiar with the existing sub-normal tests now, so LGTM.

[honno]

Maybe this is obvious to somebody who didn't just learn about the concept of sub-normals today (me), but I couldn't figure this out. My guess is that you could use an interpreter built in one architecture but run this in the other architecture, so sys.float_info could be incorrect?

[Zac-HD]

Nah, it'll be if the interpreter was built with various unsafe compiler flags which flush denormals to zero (e.g. -funsafe-math-optimizations). Which is terrible, but here we are.

(I'll clarify the test and comment, and make this a skipif)