Remove NaN tests for Ord and Eq

[anttih]

These tests rely on the purs inliner to provide correct behavior. With this change we leave the Ord and Eq NaN behavior undefined. A backend can choose how comparisons with NaN work.

Fixes #306

[garyb]

I don't know if this is right exactly, these NaN operations are part of IEEE-754 which we probably use to make other assumptions about numbers too.

[anttih]

@garyb Sure, but how do we solve #306? I thought it was agreed that we can consider NaN behavior as undefined for Ord, thus we could remove the tests. Do you have other suggestions?

[anttih]

I don't think we do other assumptions with regards to NaN. If we do, then we at least don't have tests for those.

[natefaubion]

While these are more problematic due to operator inlining, looks like there are some more Ord/Eq tests?

[anttih]

I've now removed all Ord and Eq tests for NaN.

[natefaubion]

@garyb Right now the instances/inlining situation is just incoherent. These tests aren't testing IEEE conformance so much as they are testing that the compiler inlines these operators. For example, use something other than prelude comparison operators, and these won't work!

I think if we have these tests anywhere they would need to be in the compiler with the JS backend because that's the only place this might make sense to validate.

[garyb]

Yeah, that's all fine, I'd completely forgotten about the previous discussion @anttih linked even though I participated. 😄

[anttih]

Can someone merge this?