That 8.6 figure comes from the paper linked below [0]. It’s an average. For many benchmarks, the difference is less than one percent, for others it’s 20 or more. In the freebsd article, they see around 1%.
Depends on what you optimize. If sudo's password feedback code was 8.6% slower, it would be insignificant. If openssl's heartbeat code was 8.6% slower, it would be insignificant.
The numbers on a paper itself are much lower, so I would say it is a worthy idea. If not for "whole" kernel, then at least for the calls that they have replaced.
[0] https://www.microsoft.com/en-us/research/uploads/prod/2018/0...