I instrumented the code and using conditionals resulted in 2x the generation time per loop. These loops are executed only once so the VM doesn't have a chance to optimise them.

I've done similar in the past and can concur that it does have large performance benefits, at least in Chrome where I tested it.

Stating the obvious, but I'm pretty sure that there are no longer any purely interpreting JS engines around, it's all JIT or AOT compiled, so optimizing for a compilation scenario definitely makes sense.

Testing beats imagination.