Why couldn't hardware floating point do much of the work?
Because if you want IEEE functionality, you need to use IEEE
representation, and IEEE semantics on operations when/if they involve
Inf or NaN.
I *think* that, for the numbers both formats can represent, F-float and
G-float are identical to IEEE single and double.
[...]
But I remember building Python many years ago. It worked just fine,
but the test suite to verify that it worked correctly always failed.
docs.python.org/3/reference/datamodel.html says that python
numbers.Real "represent machine-level double precision floating-point
numbers", going on to say "You are at the mercy of the underlying
[hardware/software implementation] for the accepted range and handling
of overflow". If the test suite still fails on VAXen, I would say it's
broken to the point that filing a bug report against the test suite is
appropriate.