IEEE754 non-conformance: is_sign_negative does not apply on NANs

[nagisa]

IEEE754 says

― boolean isSignMinus(source)
isSignMinus(x) is true if and only if x has negative sign. isSignMinus applies to zeros and NaNs
as well.

However our is_sign_negative/is_sign_positive do not properly inspect the sign for NaNs.

[Thiez]

This program:

fn main() {
    let nums = [::std::f64::NEG_INFINITY, -1.0, -0.0, ::std::f64::NAN, 0.0, 1.0, ::std::f64::INFINITY];
    for &num in &nums[..] {
        println!("Number: {}\tSign negative: {}\tSign positive: {}", num, num.is_sign_negative(), num.is_sign_positive());
    }
}

prints the following output when compiled with current master and nightly:

Number: -inf	Sign negative: true	Sign positive: false
Number: -1	Sign negative: true	Sign positive: false
Number: 0	Sign negative: true	Sign positive: false
Number: NaN	Sign negative: false	Sign positive: false
Number: 0	Sign negative: false	Sign positive: true
Number: 1	Sign negative: false	Sign positive: true
Number: inf	Sign negative: false	Sign positive: true

What would be the expected result according to the standard? I would expect something that isn't a number to be neither negative nor positive, so I think the current results are correct?

[nagisa]

My reading of the spec suggests the answer should be equivalent to is_sign_negative = signbit(x) != 0, where signbit is non-zero if the binary representation has sign bit set.

E.g. Section 6.2.1:

When encoded, all NaNs have a sign bit and a pattern of bits necessary to identify the encoding as a NaN and which determines its kind (sNaN vs. qNaN).

alludes to NaNs having a sign bit. The description of isSignMinus seems to quite clearly signal towards the fact that for NaNs the implementation ought to be inspecting this sign bit.

You can set the sign of NaN in rust with -::std::f64::NAN as well.