[Accepted] SE-0206: Hashable Enhancements

[lorentey (Karoy Lorentey)]

@lemire

I think that Alexis makes a good point performance-wise. If you need to hash large objects (spanning many bytes), having to eat them word by word is like drinking from a straw.

There is an overload of Hasher.combine(bits:) that takes an UnsafeRawBufferPointer. Do you think we need to provide more ways to hash bulk data?

Hash memoization and custom hash functions are still possible in the hash(into:) world, although they will incur the cost of feeding a single word to Hasher. (This is similar to how Swift 4.1's Set and Dictionary unconditionally postprocess hash values with _mixInt, it's just more explicit.)

class Foo: Hashable {
  // ...
  var _cachedHash: Int? = nil
  func hash(into hasher: inout Hasher) {
    if let hash = _cachedHash {
      hasher.combine(bits: hash)
      return
    }
    var hash = 0
    // Calculate hash however you like
    _cachedHash = hash
    hasher.combine(bits: hash)
  }
}

@lemire

The approach being proposed seems to restrict the hash functions to the “iterated type”. The universality of iterated hashing is limited. But more critically, you can eat new content but you can’t release it, so it prohibits rolling hash functions . So if I want to use this for a Karp-Rabin search, I can’t.

I think we should be careful not to overgeneralize Hashable; I consider it more as a support protocol for Set and Dictionary, rather than something that should cover all possible forms of hashing.

For example, custom seed values makes it possible to use Hashable types in e.g. some forms of Bloom filters. This is a nice bonus, but it is not why we added this feature -- the proposal adds supports for custom seed values because this turned out to be a requirement for Set/Dictionary. These collections have no need for an uncombine operation, so Hasher doesn't provide it -- rolling hashes would needlessly restrict the pool of hash functions Hasher could implement.

For specialized/generalized hashing requirements, I think it is best to define separate protocol(s).