views:

679

answers:

3

I need to check a double value for infinity in a C++ app on Linux. On most platforms this works by comparing with std::numeric_limits<double>::infinity(). However, on some old platforms (RedHat 9 for example, with gcc 3.2.2) this is not available, and std::numeric_limits<double>::has_infinity is false there.

What workaround would you recommend for those platforms?

A: 

For most cases, std::numeric_limits::max() can be a valid replacement for std::numeric_limits::infinity().

You would have to be more careful about using it, however. With overflows I believe you would have to manually detect the overflow and set the result to the max explicitly if an overflow is detected.

Andrei Krotkov
+3  A: 

If you're using IEEE 754 arithmetic, as you almost certainly are, infinities are well defined values and have defined outcomes for all arithmetic operations. In particular,

infinity - infinity = NaN

Positive and negative infinity and NaN values are the only values for which this is true. NaNs are special "not-a-number" values used to indicate domain errors of functions, e.g. sqrt(-1). Also:

NaN != NaN

NaNs are the only values for which this is true.

Therefore:

bool is_infinite(double x) {
    double y = x - x;
    return x == x && y != y;
}

will return true if and only if x is either positive or negative infinity. Add a test for x > 0 if you only want to check for positive infinity.

j_random_hacker
std::numeric_limits<double>::has_infinity==false suggests you do NOT have IEEE 754
MSalters
That may be true, but in the OP's case it's likely to mean he has an incomplete C++ Standard Library implementation that **has** IEEE 754 but doesn't "realise it". Why? Because IEEE 754 is usually done in hardware, and upgrading to a newer compiler fixed the problem.
j_random_hacker
But I agree that in the general case, if you do not know otherwise, then you should assume that std::numeric_limits<double>::has_infinity==false implies no IEEE 754.
j_random_hacker
A: 

Ok, I have now resorted to using the INFINITY and NAN macros on that particular machine - seems to work fine. They come from math.h.

oliver