On 08/31/2017 09:39 PM, Hans Boehm wrote:
>> I guess you can make VarHandle.fullFence() between getClassDataLayout0() and
>> storing it to the
>> non-volatile field...
>
> ... with the usual warning about this sort of thing:
>
> According to the JMM, it's not guaranteed to work, because the reader-side
> guarantees are not
> there. In practice, you're relying on dependency-based ordering, which the
> compiler is currently
> unlikely to break in this case. But future implementations may.
Right.
> I presume the real concern here is the cost on the reader side? Presumably
> that could be reduced
> with a VarHandle getAcquire(). I believe that eliminates the heavy-weight
> sync, and just keeps
> an lwsync. Imperfect, but better.
Oh right! This is exactly why acq/rel exist. Since OSC is a heavily used class,
the Unsafe
counterparts might be better:
private static final Unsafe U = ...;
private static final long CDS_OFFSET = U.objectFieldOffset(...);
private volatile ClassDataSlot[] dataLayout; // volatile for safety of naked
reads
ClassDataSlot[] getClassDataLayout() throws InvalidClassException {
ClassDataSlot[] slots = U.getObjectAcquire(this, CDS_OFFSET);
if (slots == null) {
slots = getClassDataLayout0();
U.putObjectRelease(this, CDS_OFFSET, slots);
}
return slots;
}
Ogata, please try if that indeed helps on POWER?
Thanks,
-Aleksey
