benchmarking Poly/ML & floating point

[Henry Cejtin]

That is certainly what I have seen in assember generated by some C compilers.
I remember all kinds of problems with Mathematica where we tried to test
for NANs by using
	if (x != x)
		NAN
or
	if (! (x == x))
		NAN
Some compilers would optimize out the test entirely.  This was in the early
days of IEEE machines, so it could be that later ANSI specs nailed it down
some what, but I doubt it.  It also could be that gcc does something more
predictable, but I don't see anything in its documentation saying that this
is the case.