Checking foo() === NaN is brittle

[KristofferC]

Tests for Distributions.jl fails on Julia master likely due to the type of NaN that is generated. === on NaN will do a bitwise check. It is likely the tests should use isnan.

[matbesancon]

I see why we should fix this in the package, but on the Julia side, isn't it a breaking change to now return false on NaN === NaN?

[andreasnoack]

So NaN === NaN is still true. I guess we might sometimes receive different kinds of NaNs from our dependencies so we end up with something like

julia> nan1 = reinterpret(Float64, 0x7ff8000000000000)
NaN

julia> nan2 = reinterpret(Float64, 0x7ff8000000000001)
NaN

julia> nan1 === nan2
false

but what have changed in Julia 1.4 that could cause this? To clarify, the example here also returns false on older Julia releases so what might have changed is the NaN produced by some of our functions.

[KristofferC]

JuliaLang/julia#34405 might make this not happen anymore but haven't checked.

[matbesancon]

interesting, I thought NaN was a unique bit value, well TIL.