[concurrency-interest] RFR: 8065804: JEP171:Clarifications/corrections for fence intrinsics

[David Holmes]

Hi Hans,

Given IRIW is a thorn in everyone's side and has no known useful benefit, and can hopefully be killed off in the future, lets not get bogged down in IRIW. But none of what you say below relates to multi-copy-atomicity.

Cheers,
David
  -----Original Message-----
  From: hjkhboehm at gmail.com [mailto:hjkhboehm at gmail.com]On Behalf Of Hans Boehm
  Sent: Wednesday, 26 November 2014 12:04 PM
  To: dholmes at ieee.org
  Cc: Stephan Diestelhorst; concurrency-interest at cs.oswego.edu; core-libs-dev
  Subject: Re: [concurrency-interest] RFR: 8065804: JEP171:Clarifications/corrections for fence intrinsics


  To be concrete here, on Power, loads can normally be ordered by an address dependency or light-weight fence (lwsync).  However, neither is enough to prevent the questionable outcome for IRIW, since it doesn't ensure that the stores in T1 and T2 will be made visible to other threads in a consistent order.  That outcome can be prevented by using heavyweight fences (sync) instructions between the loads instead.  Peter Sewell's group concluded that to enforce correct volatile behavior on Power, you essentially need a a heavyweight fence between every pair of volatile operations on Power.  That cannot be understood based on simple ordering constraints.


  As Stephan pointed out, there are similar issues on ARM, but they're less commonly encountered in a Java implementation.  If you're lucky, you can get to the right implementation recipe by looking at only reordering, I think.




  On Tue, Nov 25, 2014 at 4:36 PM, David Holmes <davidcholmes at aapt.net.au> wrote:

    Stephan Diestelhorst writes:
    >
    > David Holmes wrote:
    > > Stephan Diestelhorst writes:
    > > > Am Dienstag, 25. November 2014, 11:15:36 schrieb Hans Boehm:
    > > > > I'm no hardware architect, but fundamentally it seems to me that
    > > > >
    > > > > load x
    > > > > acquire_fence
    > > > >
    > > > > imposes a much more stringent constraint than
    > > > >
    > > > > load_acquire x
    > > > >
    > > > > Consider the case in which the load from x is an L1 hit, but a
    > > > > preceding load (from say y) is a long-latency miss.  If we enforce
    > > > > ordering by just waiting for completion of prior operation, the
    > > > > former has to wait for the load from y to complete; while the
    > > > > latter doesn't.  I find it hard to believe that this doesn't leave
    > > > > an appreciable amount of performance on the table, at least for
    > > > > some interesting microarchitectures.
    > > >
    > > > I agree, Hans, that this is a reasonable assumption.  Load_acquire x
    > > > does allow roach motel, whereas the acquire fence does not.
    > > >
    > > > >  In addition, for better or worse, fencing requirements on at least
    > > > >  Power are actually driven as much by store atomicity issues, as by
    > > > >  the ordering issues discussed in the cookbook.  This was not
    > > > >  understood in 2005, and unfortunately doesn't seem to be
    > amenable to
    > > > >  the kind of straightforward explanation as in Doug's cookbook.
    > > >
    > > > Coming from a strongly ordered architecture to a weakly ordered one
    > > > myself, I also needed some mental adjustment about store (multi-copy)
    > > > atomicity.  I can imagine others will be unaware of this difference,
    > > > too, even in 2014.
    > >
    > > Sorry I'm missing the connection between fences and multi-copy
    > atomicity.
    >
    > One example is the classic IRIW.  With non-multi copy atomic stores, but
    > ordered (say through a dependency) loads in the following example:
    >
    > Memory: foo = bar = 0
    > _T1_         _T2_         _T3_                              _T4_
    > st (foo),1   st (bar),1   ld r1, (bar)                      ld r3,(foo)
    >                           <addr dep / local "fence" here>   <addr dep>
    >                           ld r2, (foo)                      ld r4, (bar)
    >
    > You may observe r1 = 1, r2 = 0, r3 = 1, r4 = 0 on non-multi-copy atomic
    > machines.  On TSO boxes, this is not possible.  That means that the
    > memory fence that will prevent such a behaviour (DMB on ARM) needs to
    > carry some additional oomph in ensuring multi-copy atomicity, or rather
    > prevent you from seeing it (which is the same thing).


    I take it as given that any code for which you may have ordering
    constraints, must first have basic atomicity properties for loads and
    stores. I would not expect any kind of fence to add multi-copy-atomicity
    where there was none.

    David


    > Stephan
    >
    > _______________________________________________
    > Concurrency-interest mailing list
    > Concurrency-interest at cs.oswego.edu
    > http://cs.oswego.edu/mailman/listinfo/concurrency-interest

    _______________________________________________
    Concurrency-interest mailing list
    Concurrency-interest at cs.oswego.edu
    http://cs.oswego.edu/mailman/listinfo/concurrency-interest