There's a bug in the microbenchmark pushed for JDK-8246338 where the trailing slash test is not doing what it intends. This also hid the fact that we could profitably special case normalization of a path that only has a trailing slash.