The problem is located at org/hibernate/loader/Loader.java:1027, this still exists in 5.0.0-SNAPSHOT.
A second location org/hibernate/loader/plan/exec/process/internal/ResultSetProcessingContextImpl.java:359 where new instances of SubselectFetch are created is much more efficient.
The question is: is the Loader or ResultSetProcessingContextImpl responsible for subselect fetching now?