Also, I did a bunch of similar performance testing a couple years ago, with similar weird results. Regarding Swift type inference compile-time performance
Reading that thread, it looks like these types of benchmarks have been regressing for years now.