On Thu, Feb 22, 2018 at 04:30:34PM +0100, Peter Zijlstra wrote:
> On Thu, Feb 22, 2018 at 11:12:10PM +0800, Boqun Feng wrote:
> > On Thu, Feb 22, 2018 at 03:26:14PM +0100, Peter Zijlstra wrote:
>
> > > However, I would suggest:
> > >
> > > static inline bool is_xr(u16 dep)
> > > {
> > > return !!(dep & (DEP_NR_MASK | DEP_RR_MASK));
> > > }
> > >
> > > static inline bool is_rx(u16 dep)
> > > {
> > > return !!(dep & (DEP_RN_MASK | DEP_RR_MASK));
> > > }
> > >
> > > /* Skip *R -> R* relations */
> > > if (have_xr && is_rx(entry->dep))
> > > continue;
> >
> > I don't think this works, if we pick a *R for previous entry, and for
> > current entry, we have RR, NN and NR, your approach will skip the
> > current entry, but actually we can pick NN or NR (of course, in __bfs(),
> > we can greedily pick NN, because if NR causes a deadlock, so does NN).
>
> I don't get it, afaict my suggestion is identical.
>
> You skip condition: pick_dep() < 0, evaluates to:
>
> is_rr && (!NN_MASK && !NR_MASK) :=
> is_rr && (RN_MASK | RR_MASK)
>
> Which is exactly what I have.
Ooh, I think I see what I did wrong, would something like:
if (have_xr && !is_nx(entry-dep))
work? That's a lot harder to argue about though, still much better than
that tri-state pick thing.
> If that is satisfied, you set entry->is_rr to pick_dep(), which his
> harder to evaluate, but is something like:
>
> is_rr && NR_MASK || !(NN_MASK | RN_MASK) :=
> is_rr && NR_MASK || (NR_MASK | RR_MASK) :=
> (NR_MASK | RR_MASK)
>
> (because is_rr && RR_MASK will have been skipped)
>
> > >
> > > entry->have_xr = is_xr(entry->dep);
This one I think is still correct though.
