OpenSSL and panama-foreign

[Rémy Maucherat]

On Thu, Nov 11, 2021 at 3:01 PM Maurizio Cimadamore
<maurizio.cimadamore at oracle.com> wrote:
>
> Hi Remy,
> I've updated the PR:
>
> https://git.openjdk.java.net/jdk/pull/5907
>
> With the fix discussed below.
>
> Your reproducer seems to work fine on my end - but it would be great if
> you could confirm either way.

It works great. I can see it is consistently a bit faster than the
Java 17 branch.

Rémy

>
> Thanks!
> Maurizio
>
> On 11/11/2021 01:52, Maurizio Cimadamore wrote:
> > I think I managed to figure out what introduced the regression.
> >
> > Basically, it's down to the removal of customization for upcall stubs.
> >
> > ```
> >
> > @@ -123,7 +120,7 @@ public static UpcallHandler make(ABIDescriptor
> > abi, MethodHandle target, Calling
> >                  .anyMatch(s -> abi.arch.isStackType(s.type()));
> >          if (USE_INTRINSICS && isSimple && !usesStackArgs &&
> > supportsOptimizedUpcalls()) {
> >              checkPrimitive(doBindings.type());
> > -            JLI.ensureCustomized(doBindings);
> > +            doBindings =
> > insertArguments(exactInvoker(doBindings.type()), 0, doBindings);
> >              VMStorage[] args =
> > Arrays.stream(argMoves).map(Binding.Move::storage).toArray(VMStorage[]::new);
> >              VMStorage[] rets =
> > Arrays.stream(retMoves).map(Binding.Move::storage).toArray(VMStorage[]::new);
> >              CallRegs conv = new CallRegs(args, rets);
> >
> > ```
> >
> > If this patch is reverted, then everything works as expected (I've run
> > your `ab` test 10 consecutive times, no crashes). Which seems like a
> > latent bug in the upcall machinery.
> >
> > That said, I think that, for the time being, it would be best to
> > revert this code to what it was in 17, to avoid these spurious crashes.
> >
> > I wonder if the reason behind the issue we're seeing is caused by the
> > Addressable vs. MemoryAddress mismatch in the method type - e.g. maybe
> > `exactInvoker(doBindings.type())` doesn't do what we think it should.
> > I'll keep investigating tomorrow (well, later today :-) ).
> >
> > Cheers
> > Maurizio
> >
> > On 10/11/2021 22:38, Maurizio Cimadamore wrote:
> >> And, on this same vein, even with the resource scope handshake
> >> disabled, Tomcat still crashes with the following Java option:
> >>
> >> export JAVA_OPTS="-XX:+DeoptimizeALot
> >> --enable-native-access=ALL-UNNAMED --add-modules jdk.incubator.foreign"
> >>
> >> DeoptimizeALot causes frequent deopt events in the VM. The handshake
> >> we use for shared segments also used deopt events. Something got
> >> broken here and the upcall intrinsic doesn't seem to be able to
> >> handle deopt events.
> >>
> >> Maurizio
> >>
> >>
> >> On 10/11/2021 22:27, Maurizio Cimadamore wrote:
> >>> This seems to point at some issues with the shared scope logic
> >>> interacting badly with upcall intrinsics. But neither areas have
> >>> changed significantly since 17, which makes this very odd.