[lworld] RFR: 8264414: [lworld] [AArch64] TestBufferTearing.java fails with C1
thartmann at openjdk.java.net
Thu Apr 1 12:55:39 UTC 2021
On Thu, 1 Apr 2021 10:08:45 GMT, Nick Gasson <ngasson at openjdk.org> wrote:
> We see failures like this on AArch64 when MyValue.incrementAndCheck() is
> compiled with C1:
> java.lang.RuntimeException: Inconsistent field values: expected 0 to equal 675128
> at jdk.test.lib.Asserts.fail(Asserts.java:594)
> at jdk.test.lib.Asserts.assertEquals(Asserts.java:205)
> at jdk.test.lib.Asserts.assertEQ(Asserts.java:178)
> at compiler.valhalla.inlinetypes.MyValue.incrementAndCheck(TestBufferTearing.java:81)
> at compiler.valhalla.inlinetypes.TestBufferTearing$Runner.run(TestBufferTearing.java:124)
> The barrier that is usually inserted on return from a method that wrote
> final fields should be sufficient to prevent another thread seeing the
> zero-initialised intermediate state. However this barrier isn't
> inserted at the moment because method()->is_object_constructor() is
> false for primitive class constructors.
> C2 has a similar guard around the memory barrier in Parse::do_exits().
> I'm not sure if that needs amending as well but I've not seen any
> failures due to it.
I think we also need a barrier in `GraphBuilder::access_field` when loading from a flattened field and in `GraphBuilder::load_indexed` (the initialization code is emitted in `LIRGenerator::access_flattened_array`) when loading from a flattened array.
It would be good to have tests that triggers this.
More information about the valhalla-dev