RFR(M): 8154736: enhancement of cmpxchg and copy_to_survivor for ppc64

[Kim Barrett]

> On Apr 25, 2018, at 8:45 AM, David Holmes <david.holmes at oracle.com> wrote:
> 
> Hi Michihiro,
> 
> On 23/04/2018 8:33 PM, Michihiro Horie wrote:
>> Dear all,
>> I would like to ask reviews on 8154736 “enhancement of cmpxchg and
>> copy_to_survivor”. The change adds options to avoid expensive syncs with
>> compare-and-exchange. An experiment by using SPECjbb2015 showed 6%
>> improvement in critical-jOPS. This change focuses on ppc64 but would be
>> potentially beneficial for aarch64.
>> Although discussions stopped at
>> http://mail.openjdk.java.net/pipermail/ppc-aix-port-dev/2016-October/002718.html
>> , I would like to restart the review by taking over Hiroshi's work if the
>> discussion is still open.
> 
> So the very last comment there was about not implicitly assuming memory_order_consume, yet that has not been addressed in the proposal.
> 
> Further the discussion on hotspot-runtime-dev through September and October was far more illuminating. I think my post here:
> 
> http://mail.openjdk.java.net/pipermail/hotspot-runtime-dev/2016-October/021617.html
> 
> and the closely following one from Thomas Schatzl summed up the concerns about the proposed changes.
> 
> AFAICS the restarted proposal addresses none of those concerns but simply takes up where the previous implementation suggestion left off.
> 
> This is a proposal to change the memory ordering semantics of part of the shared GC code _not_ just the PPC64 implementation, but I have seen no analysis to demonstrate the correctness of such a proposal.

I agree with David here. So far we've seen no such analysis. All we
have seen is a series of proposed changes and non-failing test
results, all of which have then been shown to have holes. (Among other
things, this suggests the set of tests being applied is inadequate.)
Part of the author's job is to convince reviewers that the proposed
change is correct. I'm not expecting a formal proof, but I am
expecting a lot more than has been provided so far.

In this latest proposal, the conditional acquire doesn't look right to
me. If the logging is disabled so there is no acquire, the object is
then returned to the callers, who might do what with it? Is that safe?
For all callers? And is it stable through future maintenance? This is
not to say that I think making those acquires unconditional is
sufficient.