A single downcall currently involves a lot of code generation, such as record hashCode when putting into hash maps, indy switches and lambdas/streams. We can reduce some of these overheads.