I improved the benchmarking code by reducing the number of cube computations. Interestingly, this gives a huge improvement of guile-2.9.1 performance while improving less or even worsening performance for other scheme interpreters I've tested.

This could indicate that

1. the guile-2.9.1 optimizer is able to convert the extra code I introduced to avoid unnecessary cube computation (let:s and extra args) to something with low overhead