floats() does not generate subnormal floating-point values

[Zac-HD]

hypothesis/hypothesis-python/tests/nocover/test_floating.py

Lines 107 to 136 in 3226dd7

	# Tests whether we can represent subnormal floating point numbers.
	# This is essentially a function of how the python interpreter
	# was compiled.
	# Everything is terrible
	if math.ldexp(0.25, -1022) > 0:
	REALLY_SMALL_FLOAT = sys.float_info.min
	else:
	REALLY_SMALL_FLOAT = sys.float_info.min * 2
	# These two tests have been failing for an unknown amount of time, but that
	# failure was previously being masked by a bug in our `@fails` decorator.
	@pytest.mark.xfail
	@fails
	@given(floats())
	@TRY_HARDER
	def test_can_generate_really_small_positive_floats(x):
	assume(x > 0)
	assert x >= REALLY_SMALL_FLOAT
	@pytest.mark.xfail
	@fails
	@given(floats())
	@TRY_HARDER
	def test_can_generate_really_small_negative_floats(x):
	assume(x < 0)
	assert x <= -REALLY_SMALL_FLOAT

This was discovered in #2973, but out of scope for that PR. The first step is to determine when, if ever, this did work; what (if anything) changed to break it, and then to get subnormal-float generation working.

[Zac-HD]

OK, on investigation it is strictly speaking possible for us to 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...