Ya know, when you're building a Merkle Trie that's exactly the primitive you need.

In my own construction the assumption is that the values are already hashed down to 256 bits when they're passed in, and the tags (which are currently done by sacrificing bits instead of using tags, that needs to be fixed) include three states for either side: empty, unary, or middle. Three of those possibilities are unreachable (empty/empty, empty/unary, unary/empty) so there are 6 possible tags needed. This approach essentially skips doing the unary hashes, a further performance improvement. There doesn't appear to be any downside in leveraging this trick as long as tags are cheap.