If you expected to use thestd::tie trick to also implement the hash so that your type can be a key in a std unordered container … you’ll be disappointed. And here is why.

Tuples don’t have a standard hash

In the code below the person is a user defined type that groups several member variables.

1
2
3
4
struct person {
  std::string first_name;
  std::string last_name;
};

If you want to use the class above as a key in an unordered container e.g. unordered_set or unordered_map you need to implement a hash for the class.

First you’ll disappointed that the hash must be a specialization of std::hash in the std namespace not in the namespace of person.

Once you get over that, you might think you can delegate it to the std::tuple hash using something on the lines of the std::tie trick.

That is not the case, currently std::tuple does not have a hash.

You have been warned!

The beautifully written proposal (by Matthew Austern) N1456 to standardise hash tables warned about the specialization in the std namespace:

Trivial as it may seem, hash function packaging may be the most contentious part of this proposal.

It also suggested addressing the issue of combining hashes for fields (note that std::tuple was not in the standard at the time of the N1456 proposal):

Should we define a general hash combiner, that takes two hash codes and gives a hash code for the combination? Should we define a default hash function for std::pair<T,U>?

Why the current state of affairs?

The reality is that the current unordered containers in C++ are basically the hash containers designed in mid 90s in STL, resurrected for C++11. As programmers start to use them they have different (incompatible) expectations.

Some expect it to be a simple to use container, they expect the tuple to have a hash defined by default, and a simple function to combine hashes.

Others use it places where collisions in hashes expose systems to denial of service (DOS) attacks, and would like very good hashes e.g. of cryptographic hash strength and/or with a random seed mechanism. In those scenarios simple methods of combining like hashing the hash of members in a tuple are not good enough. There is the question of who should provide the hash functions: the standard library or it’s users, and how easy would be to update it as issues/vulnerabilities are found.

Yet others are using it in distributed environments where they would like guarantees that the hash for a value is identical on different systems.

What to do then?

The first thing to remember is that the default container should be std::vector. If data is available upfront, for a simple sorted vector the cost of each lookup is O(log(n)), but locality of data means that the constant is small.

The unordered containers were designed for special cases. For example where the data is not all available upfront. The lookup is usually O(1), but because of the memory jumps to reach the nodes the constant is not small, and the lookup could degrade to O(N). If data is not available upfront then one could use the boost functions to combine hashes, assuming it’s not for scenarios where an attacker can exploit it for DOS attacks.

An alternative is to use a key for which a hash is already defined (e.g. built-in type or std::string). My guess is that most of the users of unordered containers do just this.

For serious usages consider creating your own container.

References

There was a lot of discussion in LEWG about how to improve support for hashing. One approach, proposed in N3980 “Types don’t know #”, is to provide a generic framework for hashing arbitrary types using arbitrary hash algorithms, so that users don’t need to decide on a single hash function for their types but instead can define how their types expose their contents to any hash function that asks for them. This allows the choice of hash function for any part of a program to vary independently to the choices of data type and the definition of those types. This had a lot of support from many people as an elegant and extensible solution, but Google in particular are opposed to making it easy for users to swap in and out different hash functions in this way, as they believe it will lead to poor quality hash functions being used in ways that are hard to fix at a later date. Google’s preference is to rely on a single hash function throughout the program (by using the std::hash template). This means the compiler implementation can ensure it chooses a well-known, strong hash function for std::hash which has been studied and vetted by experts, and if weaknesses are found in that function then the standard library can switch to a different implementation of std::hash, requiring changes in only one place and fixing all the code that uses it. Due to this, Google are keen to make it easier to use std::hash for user-defined types, rather than making it easier to use any arbitrary hash function, and that’s what their proposal, N3983 “Hashing tuple-like types” tries to do. Both proposals were well-received by the LEWG, but it’s not clear whether we want to take two overlapping, and somewhat contradictory, approaches to hashing. Until something is done we have the worst of both worlds, in that users have no choice but to create their own poor quality hash values for their types, because we don’t offer them any way to do it better.