[awild]

We do something very similar at work, but with runtime generated Java code for an in-memory column oriented database. Our data is analyzed before being loaded and we determine the range of numbers necessary to store the data: eg. If we only have bytes incoming, we generate a class that only stores bytes etc. On top of that all our strings are interned in a trie and then represented as ints, also run through above outlined code.

This works because we are mostly interested in equality of strings and other computation with strings (mostly prefix based classific) is done ahead of time.

For debugging we emit our generated code and both eclipse and intellij picked up our generated code without configuration. It's safed us a lot of headaches.