Fortran has never specified a function like signs_differ(x,y)
, so one generally implements such a thing personally.
x*y<0
(and x*y.lt.0
) is not asking the same thing as "are x and y of different sign?". While the product of x and y being positive means x and y are the same sign in the (mathematical) real numbers, this is not true for (computational) floating point numbers.
Floating point multiplication x*y
may overflow, result in a signed infinite value (raising a IEEE flag) with the comparison returning the expected logical value, but that isn't always true. There were many non-IEEE systems and IEEE systems may see that flag being raised and abort (or have some expensive handling diversion). That's totally not the same thing as "do x and y have the same sign?".
x*(y/dabs(y))
doesn't overflow, is "portable" and is potentially cheaper than (x/dabs(x))*(y/dabs(y))
- ignoring the issues surrounding dabs()
and signed zeros.
Modern Fortran has functions such as sign
, ieee_copy_sign
and ieee_signbit
which didn't exist 40 years ago.