It's not unavoidable, it just hasn't yet been avoided. It could be avoided the same way as other JITs: by interpreting until a compiled version of the code is available. That's actually being worked on, but it adds considerable complexity. Among other things, it requires being able to jump out of an executing interpreter loop into an optimized, compiled version of the same function. Doable, but annoying to implement and requires compiling two versions of the function instead of one. It can also cause performance instability — see https://tratt.net/laurie/blog/2022/more_evidence_for_problem... and the older paper referenced there.