Hi Martin,
I need to think about this some more. A critical property of the fast
field accessors are that they are trivial and completely safe. They are
complicated by the need to check if a GC may have happened while we
directly read the field.
If you try to use fast field accessors when you have to post the field
access event then how can you safely go off into a JVM TI event callback ??
Thanks,
David
On 16/07/2019 11:31 pm, Doerr, Martin wrote:
Hi,
the current implementation of FastJNIAccessors ignores the flag -XX:+UseFastJNIAccessors
when the JVMTI capability "can_post_field_access" is enabled.
This is an unnecessary restriction which makes field accesses (Get<Type>Field)
from native code slower when a JVMTI agent is attached which enables this capability.
A better implementation would check at runtime if an agent actually wants to
receive field access events.
Note that the bytecode interpreter already uses this better implementation by
checking if field access watch events were requested
(JvmtiExport::_field_access_count != 0).
I have implemented such a runtime check on all platforms which currently
support FastJNIAccessors.
My new jtreg test runtime/jni/FastGetField/FastGetField.java contains a micro
benchmark:
test-support/jtreg_test_hotspot_jtreg_runtime_jni_FastGetField/runtime/jni/FastGetField/FastGetField.jtr
shows the duration of 10000 iterations with and without UseFastJNIAccessors
(JVMTI agent gets attached in both runs).
My Intel(R) Xeon(R) CPU E5-2660 v3 @ 2.60GHz needed 4.7ms with FastJNIAccessors
and 11.2ms without it.
Webrev:
http://cr.openjdk.java.net/~mdoerr/8227680_FastJNIAccessors/webrev.00/
We have run the test on 64 bit x86 platforms, SPARC and aarch64.
(FastJNIAccessors are not yet available on PPC64 and s390. I'll contribute them
later.)
My webrev contains 32 bit implementations for x86 and arm, but completely
untested. It'd be great if somebody could volunteer to review and test these
platforms.
Please review.
Best regards,
Martin
