RFR: 8377597: [Leyden] Improve peak performance when AOT code is used [v2]

[Aleksey Shipilev]

On Thu, 12 Feb 2026 01:43:27 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:

>> Currently some AOT code could be used for long time after startup. It could case peak performance regression because AOT code is conservative and have several restrictions on optimizations it can do.
>> 
>> Introduce AOT code entry counter to request JIT compilation and replace AOT code after some threshold is reached. Use invocation count of C2 code during training run as threshold for AOT code replacement during production run.
>> 
>> The counts collected during training run are scaled based on hyperbolic saturation curve formula:
>> 
>> 
>>    int scaled_limit = (AOTCodeInvokeBase + limit / (1.0 + limit / (100000.0 * AOTCodeInvokeScale)));
>> 
>> where `AOTCodeInvokeBase` (default 100.) and `AOTCodeInvokeScale` (default 1.) are diagnostic flags.
>> This scaling limits threshold to 100K for higher counts.
>> 
>> Here some results running JavacBanch JMH benchmark on linux-x64 (numactl -C 0-3 -m 0`)
>> 
>> 
>> java -jar javac.jar -f 1 -bm ss -wi 0 -i 100 JavacBench.helloWorld1k
>> 
>> 
>> <img width="781" height="466" alt="Screenshot 2026-02-10 at 1 25 30 PM" src="https://github.com/user-attachments/assets/58d973bf-9881-45d9-acb8-40b18ca02a06" />
>> 
>> 
>> <img width="486" height="178" alt="Screenshot 2026-02-10 at 1 22 24 PM" src="https://github.com/user-attachments/assets/19fff702-2302-4e43-a093-5c6981a069ba" />
>> 
>> ...
>> <img width="479" height="153" alt="Screenshot 2026-02-10 at 1 24 09 PM" src="https://github.com/user-attachments/assets/72e9d81d-bce2-482c-aaea-a32a192a8899" />
>
> Vladimir Kozlov has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Address comments

> This _does_ improve peak performance for my tests _for sure_, but still not fully there compared to non-AOT config, alas:

...which is a bit weird, because I do see this work well as recompilation policy solution: the bulk of the code is replaced by T4 by the end of the run:

<img width="1800" height="1350" alt="plot-8c-aot-after" src="https://github.com/user-attachments/assets/3c9d78e5-2b6f-482a-b4f8-605bc346b833" />

...compare with the same AOT mode before:

<img width="1800" height="1350" alt="plot-8c-aot-before" src="https://github.com/user-attachments/assets/50a41f91-9a25-4f8e-9e8d-5be9d5e39f17" />

This is 8 cores available for the run. The same configuration, but in JIT-only mode, for reference:
<img width="1800" height="1350" alt="plot-8c-jit" src="https://github.com/user-attachments/assets/d8c684e8-4507-4288-95ef-f877ade251f4" />

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

PR Comment: https://git.openjdk.org/leyden/pull/110#issuecomment-3892228446