RFR: 8227745: Enable Escape Analysis for Better Performance in the Presence of JVMTI Agents [v2]

[Richard Reingruber]

On Mon, 14 Sep 2020 05:08:39 GMT, David Holmes <dholmes at openjdk.org> wrote:

>> Richard Reingruber has updated the pull request incrementally with one additional commit since the last revision:
>> 
>>   Changes based on dholmes' feedback.
>>   
>>   EscapeBarrier::sync_and_suspend_all():
>>   
>>     - Set suspend flags before handshake because then the setting will be visible
>>       after leaving the _thread_blocked state in JavaThread::wait_for_object_deoptimization()
>>   
>>   JavaThread::wait_for_object_deoptimization()
>>   
>>     - Reuse SpinYield instead of new custom spinning code.
>>   
>>     - Do safepoint check after the loop. This is possible because the
>>       set_obj_deopt_flag is done before the handshake.
>>   
>>     - Don't set_suspend_equivalent() anymore for more simplicity. It's just an
>>       optimization (that won't pay off here).
>>   
>>   Added/updated source code comments.
>>   
>>   Additional smaller enhancements.
>
> src/hotspot/share/prims/whitebox.cpp line 884:
> 
>> 882:
>> 883: WB_ENTRY(jboolean, WB_IsFrameDeoptimized(JNIEnv* env, jobject o, jint depth))
>> 884:   JavaThread* t = JavaThread::current();
> 
> A WB_ENTRY is a JNI_ENTRY which means the current JavaThread is already available via the `thread` variable.

Done.

> src/hotspot/share/runtime/thread.cpp line 2660:
> 
>> 2658:     // showed 5% better performance when spinning.
>> 2659:     if (should_spin_wait) {
>> 2660:       // Inspired by HandshakeSpinYield
> 
> Can we not reuse any existing spinning code in the VM?

Good point.
I wasn't aware of SpinYield. I use it now for the spinning.
I use 10 * SpinYield::default_spin_limit as spin limit based on microbenchmarking:
http://cr.openjdk.java.net/~rrich/webrevs/8227745/pr116v2.microbenchmark/

-------------

PR: https://git.openjdk.java.net/jdk/pull/119