Yep, sophisticated implementations try to infer the patterns in which objects occur and then dynamically generate classes for them. However this is not without

Yep, sophisticated implementations try to infer the patterns in which objects occur and then dynamically generate classes for them. However this is not without problems, because if you assign to a new slot then you need to change the object's class, it needs to gracefully degrade to when the objects are used as hash tables, etc. LLVM IL would be excellent for speed. The problem is that it's not memory safe (i.e. it allows reading and writing to arbitrary memory locations). NativeClient at Google is trying to solve that. http://code.google.com/p/nativeclient/ http://code.google.com/p/nativeclient/