`ordNumber` is not compatible with IEEE 754 floats
anttih opened this issue · comments
purs
compiles the following:
(1.0 < nan) == false
to
1.0 < nan === false
but if we instead go through the Ord
API we can observe different behavior that is not compatible with IEEE 754:
compare 1.0 nan == GT
In addition to just being wrong, the more practical issue is that any alternative backend that does not inline the use of comparison operators to operators that implement IEEE 754 correctly won't pass the prelude test suite. For example backend-es
does not pass the prelude test suite.
One simple fix that was floated around in Discord is to just remove the tests using NaN
from the test suite and leave the behavior undefined.
This is one of the situations where Poset
would have been nice to have, since comparing NaN with something should be NC
.
There are issues with Eq
too though, since NaN comparisons aren't reflexive.
I think the only solution for status quo is to consider NaN behavior undefined for the purposes of Eq and Ord, and we just shouldn't test it.