Re: What is acceptable for -ffast-math? A numerical viewpoint

[dewar at gnat dot com]

With regard to complex arithmetic, you really have to look at the defining
standard for guidance here. I am not sure what the Fortran standard says,
but Annex G of the Ada standard specifies that overflow (or infinities)
can occur only if the *result* is out of bounds, so something like your
method c is required if you claim to support Annex G. But most certainly
a less strict mode can use a) or b). The scaling of c) can indeed be
expensive, though the test is not necessarily that expensive if you have
good scheduling, complex divisions are often surrounded by other operations
that can be scheduled, and indeed on an EPIC machine, you can simply
speculate and do both computations, choosing the right one at the end.