Hasher.S
val hash_fold_t : Hash.state -> t -> Hash.state
hash_fold_t state x
mixes the content of x
into the state
.
By default, all our hash_fold_t
functions (derived or not) should satisfy the following properties.
1. hash_fold_t state x
should mix all the information present in x
in the state. That is, by default, hash_fold_t
will traverse the full term x
(this is a significant change for Hashtbl.hash which by default stops traversing the term after after considering a small number of "significant values"). hash_fold_t
must not discard the state
.
2. hash_fold_t
must be compatible with the associated compare
function: that is, for all x
y
and s
, compare x y = 0
must imply hash_fold_t s x = hash_fold_t
s y
.
3. To avoid avoid systematic collisions, hash_fold_t
should expand to different sequences of built-in mixing functions for different values of x
. No such sequence is allowed to be a prefix of another.
A common mistake is to implement hash_fold_t
of a collection by just folding all the elements. This makes the folding sequence of a
be a prefix of a @ b
, thereby violating the requirement. This creates large families of collisions: all of the following collections would hash the same:
[[]; [1;2;3]] [[1]; [2;3]] [[1; 2]; [3]] [[1; 2; 3]; []] [[1]; [2]; []; [3];] ...
A good way to avoid this is to mix in the size of the collection to the beginning (fold ~init:(hash_fold_int state length) ~f:hash_fold_elem
). The default in our libraries is to mix the length of the structure before folding. To prevent the aforementioned collisions, one should respect this ordering.