On Mon, Sep 26, 2011 at 6:53 AM, Roman Leshchinskiy
IMO, anybody who stores Doubles in a Set either really knows what he's doing (in which case NaNs shouldn't be a problem) or is just doing it wrong.
In general, these examples just look broken to me regardless of how Eq and Ord are defined. If I saw something like this happen in real code I'd assume it's a bug.
At the very least there should be more documentation about this problem (I don't know where). Even for those who know about the IEEE standard, it's not obvious at all that these functions won't work. The more complex the example (max → sort → Data.Set), the less straightforward the problem becomes. Instead of this proposal, it may also be interesting to propose something like newtype Total a = Total a type TotalDouble = Total Double type TotalFloat = Total Float which provides meaningful Eq and Ord instances for algorithms that don't care about what a NaN is. Cheers, -- Felipe.