Hmm... On checking down some of the code paths and realizing there were some issues in 3.5 (redundant code, and what looked like two memory leaks), I checked tip (to avoid opening bugs on stale code), and discovered that #22557 rewrote the import code, reducing the cost of top level reimport by ~60%, so my microbenchmarks (run on Python 3.5.0) are already out of date for 3.6's faster re-import. Even so, caching wasn't a wholly unreasonable optimization before now, and undoing it now still has a cost, if a smaller one.