Ted Unangst wrote: > Michael McConville wrote: > > Ted Unangst wrote: > > > Michael McConville wrote: > > > > We have a pretty strong guarantee that it can only happen once > > > > per process... > > > > > > I don't think this really matters. What does it do to the > > > assmembly, and how does that make things faster? > > > > It lets the compiler know that the body is very unlikely to run so > > that it won't unroll loops, and will maybe bump the condition body > > to the end of the procedure, etc. It can also be used to annotate > > the branch with a hint instruction, but I don't know how many > > architectures still use those. > > I meant in this case specifically. What is the *demonstrated* benefit? > > Generally, not many fans of the annotation for the sake of annontation > in these parts. :)
So, I broke out objdump and looked at how these differ on amd64. No profiling yet (I've never profiled syscalls before, advice welcome). The condition evaluation is identical. The differences: * the annotated version jumps on error to the end of the syscall procedure, while the current version has the error body directly after the condition (in the middle of the procedure's hot path) * the annotated version is slightly smaller, assumedly because the compiler didn't bother doing expansion-based optimizations in the error body. The current version prevents speculative execution from being useful and puts unused fluff in the I-cache. Not super important, but these annotations are easy and pretty consistently beneficial.
