Hi Egor, On Wed, Apr 15, 2026 at 06:07:54PM +0000, Egor Shestakov wrote: > Hi List! > > I found a code duplication in run_tasks_from_lists(), when a task's > process() callback runs. Initially I wanted just eliminate the duplication, > but discovered there was a branch prediction optimization (see f0c531ab5 > (MEDIUM: tasks: implement a lockless scheduler for single-thread usage, > 2017-11-05): > > /* This is an optimisation to help the processor's branch > * predictor take this most common call. > */ > if (likely(t->process == process_stream)) > t = process_stream(t); > else > t = t->process(t); > > Along the way, this became into this: > > if (process == process_stream) > t = EXEC_CTX_WITH_RET(EXEC_CTX_MAKE(TH_EX_CTX_TASK, process_stream), > process_stream(t, ctx, state)); > else > t = EXEC_CTX_WITH_RET(EXEC_CTX_MAKE(TH_EX_CTX_TASK, process), > process(t, ctx, state)); > > Should this optimization be restored? If not, maybe it's better to just > remove the code duplication?
I remember testing it again around 2 years ago and it was still worth it. It's due to branch prediction units in CPUs which are extremely effective and allow to prefetch instructions from process_stream() before the branch is decided, and since it's by far the most called function when under high HTTP loads, the few tens of cycles saved there are a net overall gain. We could recheck at some point though. Regards, Willy
