Make a better story for split

[treeowl]

I think we should look to Data.Bits for inspiration here. Add a class

class RandomGen g => SplittableRandomGen g where
  safeSplit :: g -> (g, g)

safeSplit is the same as split, except it should only be defined for generators that are actually splittable.

Question: can we improve efficiency or flexibility for any mutable generators by offering something like this?

class StatefulGen g m => SplittableStatefulGen g m where
  splitStatefulGenM :: g -> m g

I imagine this might be able to avoid some unnecessary copying in some cases.

[treeowl]

It looks like this was discussed in idontgetoutmuch#7 but for reasons I don't know was rejected there? Doesn't make sense to me.

[treeowl]

FYI: I want to play with splittable generators specifically for incremental sorting of sequences (for Data.Sequence).

[lehins]

@treeowl Yes we did discuss this approach at lengths when working on random-1.2.0, but we decided not to tackle this problem for that version, since there was already tremendous amount of changes being introduced and we wanted to keep breakage to a minimum.

Here are my comments on your suggestion:

• SplittableStatefulGen is not useful at all, because only pure PRNGs that have instance of RandomGen are capable of splitting. For that reason I prefer to use RandomGenM class instead and use the approach that is in the works in WIP: Improve RandomGenM interface: #93
• class RandomGen g => SplittableRandomGen g where:
  • I really don't like the name SplittableRandomGen, I think SplittableGen would be nicer
  • renaming split -> safeSplit would introduce more work on the users side, it is better to migrate split into SplittableGen right away.
  • Introducing new class would mean that we must go through all PRNG libraries and create that instance, which is problematic, since majority of them are unmaintained.

I thought about this issue quite a bit before, because I also share your dislike of partial functions, and I think I came up with slightly less invasive solution to this problem in #94, please let me know what you think.