The difference is that you compile user control flow as-is unless you have overwhelming evidence that you should do otherwise.

And yes you are right. This is promising for run-once code, but all of the cost will be in dynamic things like the envGet. An interpreter can actually do better here because most environment resolution can be done as part of bytecode generation. So it’s possible that this experiment leads to something that is slower than JSC’s interpreter.