Okay, here's one I just encountered: In Hopcroft's DFA minimization algorithm, one must collect all states that have edges arriving at a chosen state, grouping them by the character on the edge.  A natural representation for each state, in this algorithm, is as a hash table mapping characters to sets of states.  Merging, then, is computing the union of all sets of states for each character.  In this case, it's probably most natural to use the functional version, but it's good to expose the imperative version for efficiency in cases where accumulation doesn't happen all at once.