RFR: 8358693: [leyden] Bootstrapping circularity leads to never starting preload threads

[Aleksey Shipilev]

On Thu, 5 Jun 2025 12:33:00 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

> This relates to [JDK-8358690](https://bugs.openjdk.org/browse/JDK-8358690), but we want to make sure Leyden preload runs until we have a fuller fix.
> 
> Additional testing:
>  - [x] Ad-hoc perf tests
>  - [ ] Linux x86_64 server fastdebug, `runtime/cds`

You can clearly see it in `-Xlog:jit+thread=debug`:


# Before
[0.016s][debug][jit,thread] Added initial compiler thread C2 CompilerThread0
[0.016s][debug][jit,thread] Added initial compiler thread C1 CompilerThread0
[0.026s][debug][jit,thread] Added compiler thread C2 CompilerThread1 (free memory: 62377MB, available non-profiled code cache: 115MB)
[0.026s][debug][jit,thread] Added compiler thread C2 CompilerThread2 (free memory: 62377MB, available non-profiled code cache: 115MB)
[0.026s][debug][jit,thread] Added compiler thread C2 CompilerThread3 (free memory: 62377MB, available non-profiled code cache: 115MB)
[0.026s][debug][jit,thread] Added compiler thread C2 CompilerThread4 (free memory: 62377MB, available non-profiled code cache: 115MB)
[0.026s][debug][jit,thread] Added compiler thread C2 CompilerThread5 (free memory: 62377MB, available non-profiled code cache: 115MB)
[0.026s][debug][jit,thread] Added compiler thread C2 CompilerThread6 (free memory: 62377MB, available non-profiled code cache: 115MB)
[0.026s][debug][jit,thread] Added compiler thread C2 CompilerThread7 (free memory: 62377MB, available non-profiled code cache: 115MB)
[0.026s][debug][jit,thread] Added compiler thread C2 CompilerThread8 (free memory: 62377MB, available non-profiled code cache: 115MB)
[0.027s][debug][jit,thread] Added compiler thread C2 CompilerThread9 (free memory: 62377MB, available non-profiled code cache: 115MB)
[0.027s][debug][jit,thread] Added compiler thread C1 CompilerThread1 (free memory: 62377MB, available profiled code cache: 116MB)
[0.027s][debug][jit,thread] Added compiler thread C1 CompilerThread2 (free memory: 62377MB, available profiled code cache: 116MB)
[0.027s][debug][jit,thread] Added compiler thread C1 CompilerThread3 (free memory: 62377MB, available profiled code cache: 116MB)
[0.027s][debug][jit,thread] Added compiler thread C1 CompilerThread4 (free memory: 62377MB, available profiled code cache: 116MB)

# After
[0.015s][debug][jit,thread] Added initial compiler thread C2 CompilerThread0
[0.015s][debug][jit,thread] Added initial compiler thread C1 CompilerThread0
[0.016s][debug][jit,thread] Added initial compiler thread C1 AOT code caching CompilerThread
[0.016s][debug][jit,thread] Added initial compiler thread C2 AOT code caching CompilerThread
[0.026s][debug][jit,thread] Added compiler thread C1 CompilerThread1 (free memory: 62495MB, available profiled code cache: 116MB)
[0.027s][debug][jit,thread] Added compiler thread C1 CompilerThread2 (free memory: 62495MB, available profiled code cache: 116MB)
[0.027s][debug][jit,thread] Added compiler thread C1 CompilerThread3 (free memory: 62495MB, available profiled code cache: 116MB)
[0.027s][debug][jit,thread] Added compiler thread C1 CompilerThread4 (free memory: 62495MB, available profiled code cache: 116MB)


The bug regressed premain as we ported up JDK-8354887 from mainline.


* | commit be24c2646c8c9e6009f52b891cf058d915f22072            <------ SLOW
| | Author: Vladimir Kozlov <vladimir.kozlov at oracle.com>
| | Date:   Wed May 28 16:43:31 2025 -0700
| | 
| |     Port JDK-8354887 from mainline
| | 
* | commit 31bf3be94f26f983ade75369d4c257b06dbb2010            <------ FAST
|| Merge: ab4d18e2d15 2595fcc7cc4
| | Author: Vladimir Kozlov <vladimir.kozlov at oracle.com>
| | Date:   Wed May 28 08:12:57 2025 -0700
| | 
| |     Merge 8356192


See:


Benchmark 1: build/linux-x86_64-server-release/images/jdk/bin/java -Xms64m -Xmx1g -XX:+UseSerialGC -cp JavacBenchApp.jar -XX:AOTCache=app.aot JavacBenchApp 50

# at be24c2646c8c9e6009f52b891cf058d915f22072 + this fix
  Time (mean ± σ):     316.2 ms ±   2.1 ms    [User: 655.8 ms, System: 105.6 ms]
  Range (min … max):   312.3 ms … 320.8 ms    30 runs

# at be24c2646c8c9e6009f52b891cf058d915f22072
  Time (mean ± σ):     365.4 ms ±   5.0 ms    [User: 894.8 ms, System: 131.4 ms]
  Range (min … max):   355.9 ms … 374.4 ms    30 runs


# at 31bf3be94f26f983ade75369d4c257b06dbb2010
  Time (mean ± σ):     321.8 ms ±   4.8 ms    [User: 657.1 ms, System: 101.6 ms]
  Range (min … max):   316.0 ms … 335.2 ms    30 runs


The effect on current premain is less pronounced, because I think there are more regressions. I am chasing those down as well.

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

PR Comment: https://git.openjdk.org/leyden/pull/75#issuecomment-2944196111