Beware of micro-optimizations, as they depend on non-obvious factors. At best, it's a wash. In a situation like this, I suspect it's actually less efficient to create a new array, push two references on it, take a reference to it, create a (single element) list, and return the list. The other alternative is to create a two element list and return that.

Sure, there's one more element to push on the return stack and one more element to copy, if the caller uses the returned list, but you'll have to copy something in both cases, and if you return a reference to an array of references, you'll have to go through more levels of dereferencing.

I don't know where it's more efficient to do it your way, but it probably doesn't converge until you're returning at least five or six references.