better instance Hashable IntSet?

[jwaldmann]

cross reference to haskell-unordered-containers/hashable#269

[jwaldmann]

for a given set of elements, the IntSet representation is unique? E.g., "the mask is the largest position ..." https://github.com/haskell/containers/blob/master/containers/src/Data/IntSet/Internal.hs#L261

Then a hash function could just use structure and contents of the internal representation generically?

[treeowl]

@jwaldmann Correct.

[meooow25]

Correct, but it would be wasteful. The Tips contain all the elements, it would be sufficient to just hash the contents of those (ignoring Bins).

[treeowl]

Correct, but it would be wasteful. The Tips contain all the elements, it would be sufficient to just hash the contents of those (ignoring Bins).

Oh yeah, I think there is redundant info there. I'd have to review the representation; it's been a little while.

[jwaldmann]

Ah yes, the Tip includes the prefix.

[treeowl]

We could offer

serialize :: (Word -> r -> r) -> r -> IntSet -> r

This could be used to hash efficiently, with serialize written so it can inline. What would be a good matching deserialize type though?

[jwaldmann]

• serialize: could we write instance Ord IntSet with that? (Improve Ord IntSet instance #787 Improve Ord instances for IntSet and IntMap #470)
• deserialize: don't know. what would be the use case?

[treeowl]

I doubt that's quite what we'd want for Ord, but maybe close. The use case is that I'd be more comfortable with "publicly" exporting serialize/deserialize than a function that's only good for writing the Hashable instance.

[meooow25]

I think we should not mix up serialization and hashing.
Serialization typically requires the size before the elements, which means we must make two traversals for IntMap/IntSet.
Hashing also wants the size for a good hash, but it is fine to add it at the end, which means we can be more efficient and calculate the hash and size in one traversal.