[aarch64-port-dev ] RFR(XS): 8248219: aarch64: missing memory barrier in fast_storefield and fast_accessfield

[Yangfei (Felix)]

Hi Andrew,

  Sorry for the late reply.  It's Dragon Boat Festival here in China.

> -----Original Message-----
> From: Andrew Haley [mailto:aph at redhat.com]
> Sent: Wednesday, June 24, 2020 6:45 PM
> To: Yangfei (Felix) <felix.yang at huawei.com>; hotspot-runtime-
> dev at openjdk.java.net
> Cc: aarch64-port-dev at openjdk.java.net
> Subject: Re: RFR(XS): 8248219: aarch64: missing memory barrier in
> fast_storefield and fast_accessfield
> 
> On 24/06/2020 10:38, Yangfei (Felix) wrote:
> >   Suggestions?
> 
> Great catch!

Thanks for the quick review :-)

> Thanks for that, I completely agree. Please benchmark the two and unless
> there is an advantage for the address dependency we'll go with LoadLoad.  It

I use a simple test [1] to exercise the read & write object field operation and run it in -Xint mode.
Test result on our aarch64 platform show no advantage for the address dependency way.
So I think I am OK to go with the LoadLoad way. 

Webrev: http://cr.openjdk.java.net/~fyang/8248219/webrev.00/ 

> looks like all of the ConstantPoolCacheEntry::set methods use
> Atomic::release_store, so everything should be fine there.
>
> Did you also look to see if we need similar memory barriers elsewhere?

Yes, I checked and only saw these two places.

> We're going to need backports for all extant OpenJDK versions. Please let us
> know if you can handle that too.

Well, I think at least I can handle jdk8u-shenandoah, jdk11u and jdk/jdk.  

Felix

[1]
class TestCP {

  public int num;

  void run(int reps) {
    for (int i = 0; i < reps; i++) {
      num += i;
    }
  }

  public static void main(String[] args) throws Exception {
    int reps = Integer.parseInt(System.getProperty("repos"));
    TestCP t = new TestCP();

    Long startTime = System.nanoTime();
    t.run(reps);
    Long endTime = System.nanoTime();
    System.out.println(endTime - startTime);
  }
}