[Numeric-interest] rsqrt(-0) and rsqrt(-inf)

[Douglas Priest]

A number of math libraries now include a reciprocal square root
function, rsqrt, such that rsqrt(x) delivers an approximation to
x^(-1/2).  Some implementations, including HP's and IBM's, have
chosen to define rsqrt so that 

	rsqrt(-0) delivers -inf and raises division-by-zero

and

	rsqrt(-inf) delivers NaN and raises invalid operation.

These choices are consistent with the interpretation that rsqrt(x)
is a substitute for the expression "1.0/sqrt(x)", but since the
actual rsqrt functions generally do not deliver the exact same
values as 1.0/sqrt(x) computed by a correctly rounded square root
followed by a correctly rounded divide, there is no reason a priori
why we must make this interpretation.  We could, for example, say
that rsqrt(x) is a substitute for the expression "sqrt(1.0/x)",
which would lead us to choose

	rsqrt(-0) delivers NaN and raises invalid operation

and

	rsqrt(-inf) delivers -0 and raises no exception

instead.

Does anyone know of any examples of software that could benefit from
a particular choice of result for either rsqrt(-0) or rsqrt(-inf)?
I am not asking for consistency arguments or educated guesses, just
real programs in which the "right" definition would make things simpler
or the "wrong" one make them more complicated.

Thank you,
Doug Priest