RFR: 8335358: [premain] Explore alternative ways to trigger the end of training run [v5]

[Mat Carter]

On Tue, 17 Sep 2024 04:22:44 GMT, Ioi Lam <iklam at openjdk.org> wrote:

>> Mat Carter has updated the pull request incrementally with one additional commit since the last revision:
>> 
>>   removed extra space left over from reverted change
>
> src/hotspot/cpu/aarch64/templateInterpreterGenerator_aarch64.cpp line 1280:
> 
>> 1278:   // AOT training run support
>> 1279:   __ end_training_check();
>> 1280: 
> 
> I wonder how much this will slow down the interpreter, as we trap into the VM on every interpreter entry. Could you do something like
> 
> 
> long start = System.currentTimeMillis();
> for (int i = 0; i < 10000000; i++) {
>     emptyMethod();
> }
> System.out.println(System.currentMillis() - start();
> 
> 
> and run it with something like this? 
> 
> 
> rm foo.cds
> java -Xint  -XX:+UnlockDiagnosticVMOptions -XX:+CDSManualFinalImage -XX:CacheDataStore=foo.cds ...

The latest commit changes how we add triggers to the interpreter.  I added variants for the prologs that can have triggers, so only flagged methods will trap in the VM.  Also as the triggers are now only in flagged methods the VM code is faster as it no longer has to wrangle the method handle to test the flags.  While the traps remain even after the end of training has been triggered, the early out tests reduce the performance impact

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

PR Review Comment: https://git.openjdk.org/leyden/pull/21#discussion_r1765205875