On Wed, Feb 27, 2019 at 9:34 AM Peter Zijlstra <pet...@infradead.org> wrote: > > But the ONCE thing defeats CSE (on purpose!) so for code-gen that > likes/wants that we then have to use the __this_cpu crud.
Right. But I do think that if you want CSE on a percpu access, you might want to write it as such (ie load the value once and then re-use the value). But I'm not sure exactly which accesses Nadav noticed to be a problem, so.. I guess I should just look at the other patches, but I found them rather nasty and ad-hoc too so I just skimmed them with an "eww" ;) Linus