I don't like this PR as-is. The reason I wanted to split up lookup and invoke instructions is because we can simplify our code: we can have Lookup, LookupSuper, ... and then just Send, SendWithoutBlock. We don't need to fully explode the powerset of combinations (and their optimized variants) into the IR. Ideally we only have ~1 kind of Ruby send that we can analyze, simplify, inline, etc. I don't want to treat invokesuper differently from e.g. invokeblock.

That sounds right. YJIT lets those instructions share the same gen_send_iseq/gen_send_cfunc optimizations and we shouldn't give it up in ZJIT.

At the same time, however, all these instructions need separate fallback implementations. We might not need two InvokeSuper* instructions, but we need one for the invokesuper fallback (and another for invokeblock), just like Insn::Send and (non-direct) Insn::SendWithoutBlock.