I tried a few micro-optimizations, but they did not make a measurable difference, so I kept the code short instead. But maybe some JIT is particularly bad at loop hoisting, so it might make a difference there.