Probably fifteen years ago I discussed runtime code generation with some computer architects and they did indeed raise concerns about caching at the time. But now I am not as convinced this is as much of a worry as people seem to think. Sure if you were generating new code ever time you opened a file and did a write perhaps you would blow out the instruction cache with excessive amounts of freshly generated code. But if you waited to generate an optimized version of write until you see it is "hot spot" it seems like you would probably be okay.

I can only speculate, so maybe someone who knows better can correct me, but I would be surprised if many kernel syscall implementations would actually stay in cache after a dwq context switches. Especially with recent timing channel mitigations. But I'm just guessing.

And even if there are architectural issues, with processors seeing less benefit from Moore's Law I would expect computer architects to be looking for ways to eek out more performance such that they'd find solutions.

The synthesis paper directly talks about the cache effects.