RFR: 8355296: [leyden] Some methods are stuck at level=0 with -XX:-TieredCompilation

[Aleksey Shipilev]

On Tue, 22 Apr 2025 18:09:38 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

> Discovered this originally when testing Persistent Profiles PR in Leyden. In short, the transition from level=0 would not go to level=4 ever for some methods, if we disable intermediate levels. I believe this is one the causes for at least some performance artifacts I seen in Leyden testing last week. 
> 
> There are two bugs:
>  1. In `CompilationPolicy::common`: Trivial/native method shortcut was lost, so trivial methods enter through the normal profiled policy (without a good reason?)
>  2. In `CompilationPolicy::should_create_mdo`: When training data is used we _only_ create MDOs for methods that are referenced in TD. This has a major knock-on effect on _normal_ JIT compilations that want to see MDO invocation/backedge counters. Instead, MDO is missing, so interpreter only increments MCS counters, and then compilation policy misses the transitions.
> 
> (1) is really subsumed by (2), so we can really only do (2).
> 
> See reproducers and logs in the issue. More performance data in comments.
> 
> Additional testing:
>  - [x] Linux x86_64 server fastdebug, `runtime/cds`

On persistent profiles PR alone:

Before: ![before-fix](https://github.com/user-attachments/assets/efffd541-2d1d-4b8f-ba5b-d8d3c8e13527)
After: ![after-fix](https://github.com/user-attachments/assets/488c3ac5-ed0e-48c2-a9f2-4d552b14a0d6)

For Leyden `premain`, `JavacBenchApp 50`:


# Before
  Time (mean ± σ):     545.4 ms ±   3.8 ms    [User: 777.4 ms, System: 64.6 ms]
  Range (min … max):   538.1 ms … 552.3 ms    30 runs

# After
  Time (mean ± σ):     536.5 ms ±   6.0 ms    [User: 785.0 ms, System: 64.8 ms]
  Range (min … max):   524.6 ms … 547.8 ms    30 runs

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

PR Comment: https://git.openjdk.org/leyden/pull/57#issuecomment-2822107505