views:

192

answers:

3

The documentation for java.lang.Double.NaN says that it is

A constant holding a Not-a-Number (NaN) value of type double. It is equivalent to the value returned by Double.longBitsToDouble(0x7ff8000000000000L).

This seems to imply there are others. If so, how do I get hold of them, and can this be done portably?

To be clear, I would like to find the double values x such that

Double.doubleToRawLongBits(x) != Double.doubleToRawLongBits(Double.NaN)

and

Double.isNaN(x)

are both true.

+1  A: 

IEEE 754 defines a NaN as a number with all exponent bits which are 1 and a non zero number in the mantissa.

So for a single-precision number you are looking for:

S     E            M
x  11111111   xxxxxx....xxx (with M != 0)

Java handles this like so:

Double n = Double.longBitsToDouble(0x7ff8000000000000L); // default NaN
Double n2 = Double.longBitsToDouble(0x7ff8000000000100L); // also a NaN, but M != 0

System.out.println(n.isNaN()); // true
System.out.println(n2.isNaN()); // true
System.out.println(n2 != Double.doubleToLongBits(Double.NaN)); // true

To sum, you can use any NaN you want which conforms to the rules aforementioned (all bits 1 in exponent and mantissa != 0).

Yuval A
The last statement makes no sense. `Double.doubleToLongBits(...)` is a `long`, which is implicitly converted to `double`, autoboxed and compared by *reference* to `n2`.
finnw
Possibly in the last line should be `Double.doubleToLongBits(n2) != Double.doubleToLongBits(Double.NaN)` ?
Simon Nickerson
That prints `false`, because `doubleToLongBits` converts all `NaN` s to the same value. But it's true if you use `doubleToRawLongBits`.
finnw
I was wrong about the reference comparison (I just looked at the bytecode.) The `long` is converted to `double`, but `n2` is auto-unboxed and normal `double` comparison is applied.
finnw
+2  A: 

Java uses IEEE 754 for its floating point numbers and therefore follows their rules.

According to the Wikipedia page on NaN it is defined like this:

A bit-wise example of a IEEE floating-point standard single precision NaN: x111 1111 1axx xxxx xxxx xxxx xxxx xxxx where x means don't care.

So there are quite a few bit-patterns all of which are NaN values.

Joachim Sauer
Thanks. Is this guaranteed to be portable across VMs?
Simon Nickerson
@simonn: that question would make me stop for a moment: How exactly do you want to make use of that fact? Do you try to submit in-band information inside the NaN value of a `double`?
Joachim Sauer
Sort of. I am trying to emulate the semantics of a system where there are several different values which all act like NaN, but they are distinguishable.
Simon Nickerson
+1  A: 

You need doubleToRawLongBits rather than doubleToLongBits.

doubleToRawLongBits extracts the actual binary representation. doubleToLongBits doesn't, it converts all NaNs to the default NaN first.

double n = Double.longBitsToDouble(0x7ff8000000000000L); // default NaN
double n2 = Double.longBitsToDouble(0x7ff8000000000100L); // also a NaN, but M != 0

System.out.printf("%X\n", Double.doubleToLongBits(n));
System.out.printf("%X\n", Double.doubleToRawLongBits(n));
System.out.printf("%X\n", Double.doubleToLongBits(n2));
System.out.printf("%X\n", Double.doubleToRawLongBits(n2));

output:

7FF8000000000000
7FF8000000000000
7FF8000000000000
7FF8000000000100
finnw
Thanks for the tip about `doubleToRawLongBits`!
Simon Nickerson