In the past they've talked about the speed and memory improvements they've gotten from using MultiArrayList in the compiler, storing tags separately from the unions themselves. If you have a union with a size of 16 bytes and you add a tag to that which is 1 byte, a lot of space is wasted due to padding. If you keep the tags in a separate array, both arrays are individually densely packed. Less memory wasted due to padding == less memory use overall and better utilization of cache.

But in terms of the implementation, this means working with untagged unions, because the tags are maintained externally.