On Thu, Nov 05, 2020 at 01:18:02PM -0800, Linus Torvalds wrote: > On Thu, Nov 5, 2020 at 1:13 PM Luc Van Oostenryck > <luc.vanoostenryck@xxxxxxxxx> wrote: > > > > Annotations can't fix anything here. The patch [1] changes htab_lock_bucket(), > > into conditionally taking a lock while before it took it unconditionally. > > So, each point where its value is tested now becomes the confluence of 2 > > distinct paths: one with the lock taken and one with the lock not taken. > > This is what is meant by 'context imbalance'. > > I think the point Song makes is that if sparse did a few more > optimizations, it would see the context imbalance go away, because the > value it tests would become constant. It's not how I understood it, but yes, I agree. It's what I tried to briefly explain the in the last paragraph. >From what I know from these situations, in most cases, simple 'expression' optimizations are not enough but simple forms of code hoisting would often anything that's needed. -- Luc
