[aarch64-port-dev ] JCStress Failure -WCAS_WCAS

[Aleksey Shipilev]

Hi Stuart,

On 09/16/2016 02:52 PM, Stuart Monteith wrote:
>    I've have tests-chapter-1a compiling by generating a jar to use
> "--add-patch" against so that @sun.misc.Contended will apparently be
> in the jdk.unsupported module, but of course, that's not the problem
> once you actually try to run.

Yes, but I would prefer to have all these little Jigsaw pummeling
options out, if possible.

> Your changes look good to me, with the exception of the Padding. There
> is an Aarch64 system that requires 128 bytes of padding to avoid
> contention. If I interpret what you've written correctly, the booleans
> you use for padding will be 64 bytes in length. Am I correct in
> thinking booleans are 4 bytes in Hotspot fields? 

No. Booleans are 1-byte fields in HotSpot. But Padding plus two classes
generates 16x16 = 256 booleans on each side of the protected field,
which yields 256 bytes around it. See e.g.:

$ java --add-exports java.base/jdk.internal.vm.annotation=ALL-UNNAMED
-jar jol-cli.jar internals -cp tests-custom/target/jcstress.jar
org.openjdk.jcstress.infra.results.CharResult1_jcstress

org.openjdk.jcstress.infra.results.CharResult1_jcstress object internals:
 OFFSET  SIZE    TYPE DESCRIPTION                    VALUE
      0     4         (object header)                ...
      4     4         (object header)                ...
      8     4         (object header)                ...
     12     2    char CharResult1.r1                 \00
     14     2         (alignment/padding gap)        N/A
     16     1 boolean CharResult1_jcstress_c2.p000   false
     17     1 boolean CharResult1_jcstress_c2.p001   false
     18     1 boolean CharResult1_jcstress_c2.p002   false
...
    270     1 boolean CharResult1_jcstress_c2.p254   false
    271     1 boolean CharResult1_jcstress_c2.p255   false
    272     4     int CharResult1_jcstress_c1.trap   0
    276     1 boolean CharResult1_jcstress.p000      false
    277     1 boolean CharResult1_jcstress.p001      false
...
    530     1 boolean CharResult1_jcstress.p254      false
    531     1 boolean CharResult1_jcstress.p255      false
    532     4         (loss due to the next object alignment)
Instance size: 536 bytes


> The @Contended flag has the correct behaviour as it can query the
> appropriate granule size.

Yes, I know about the benefits of @Contended. Alas, using it in
compatible manner outside JDK is a nuisance. Granted, this is partially
my own fault for not pressing this hard enough during the original
development, but hey, I *was* suggesting java.util.concurrent.Contended
:) Hierarchy trick is the next best thing.

Thanks,
-Aleksey