Quoting from Jagerman's reason: With the std::hash specialization
already here, an unordered_map here seems an obviously better choice:
hash collisions won't be destructive (they are with std::map + hash as
keys), the code is slightly simpler (you can eliminate the hash call),
and performance is better (map is an insertion sorted container, thus
O(NlogN) on N insertions).

This isn't just an optimal data structure thing: using a map here has
a potential (if rare) data omission on collision.