We still have SPECIAL_JUMP which is used for all sorts of enumerable messages and so SpecialJump for a simple marker flow control exception instance is overkill so we should consider killing SpecialJumpException altogether for:

Exception SPECIAL_JUMP = new Exception();

There is a tiny risk some native exceptions use Return/Next/Redo but since they are never thrown in 9 I think it is correct for the extensions to error out and know about it sooner than later...

How does IR represent control flow now, if not by exceptions? Especially NextException with blocks - that is handled by the yield-site, not by the block itself, is it? So it could be compiled separately from the yield-site and the control flow is non-local.

IR uses exceptions as well. These older AST-runtime exceptions weren't being used. As for next, a next in a closure becomes a non-local return (which is handled using non-local exceptions anyway). https://github.com/jruby/jruby/blob/master/core/src/main/java/org/jruby/ir/IRBuilder.java#L2781-L2786

Uh .. I am clearly rusty at this point, but, no, that is not a non-local return, just a simple return which transfers control back to the yielder, and so no exceptions involved in next handling.