[PATCH RFC rebase 3/9] powerpc/64: Use barrier_nospec in syscall entry
npiggin at gmail.com
Fri Mar 16 21:46:47 AEDT 2018
On Fri, 16 Mar 2018 10:15:49 +0100
Michal Suchánek <msuchanek at suse.de> wrote:
> On Fri, 16 Mar 2018 15:18:23 +1000
> Nicholas Piggin <npiggin at gmail.com> wrote:
> > On Thu, 15 Mar 2018 20:15:52 +0100
> > Michal Suchanek <msuchanek at suse.de> wrote:
> > > On powerpc syscall entry is done in assembly so patch in an explicit
> > > barrier_nospec.
> > Same comment as Linus for this -- the barriers are before the branch
> > here, so is it possible the branch instruction can be speculative
> > while the index is used to load the syscall table?
> As far as I understand barriers they separate code before the barrier
> and code after the barrier.
> So inserting barrier_nospec after cmpldi means that the result of the
> cmpldi has to be known before any instruction following barrier_nospec
> that depends on the result can be executed.
> In many cases it is useful to put the barrier after a branch. It allows
> the compiler to speculate on the computed value at compile time and if
> it is constrained optimize out the branch. It may also result in the
> need to include many barriers and less readable code.
> However, you have probably knowledge of the powerpc implementation of
> the barrier so if the semantic is actually different then please
> enlighten me.
I actually don't. I'm assuming we should be able to say that no previous
instruction is speculative when a subsequent one is executed.
But the branch instruction itself that is speculated, not the compare.
Usually even if all sources are ready, the pipeline may take in some
cycles after a branch, before that branch can finish executing and
squash speculation if it was wrong. Perhaps there is only a couple of
cycles of instructions that get a chance to reach execution units and
disturb any caches, but still there could be some window and I don't
think we would have architectural gurantees on that.
I'll try to ask around and see if there's any documentation we can
give you yet.
More information about the Linuxppc-dev