From the post: «In the context of resource-constraint embedded devices, it may make sense to tailor an interpreter to a specific applications to gain best perfo

From the post: «In the context of resource-constraint embedded devices, it may make sense to tailor an interpreter to a specific applications to gain best performance.» The «key problems» are also discussed in the introduction of the paper at the top of page 2 - including why it’s practical in their context. Also in the paper, they say speedups based on optimizing a subset of benchmarks generalize to other benchmarks as well, but not across environments. Last I checked, starting with an «optimizing your environment» task, is far from being considered «not very practical»