RFR: Prefer to take AOT method from the SC queue faster

[Vladimir Kozlov]

On Fri, 6 Sep 2024 19:41:12 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

> Probably WIP. Humor me for a second: why could not / should not we take the AOT task for "compilation" (installation, really) right away like this? I suspect when SC queues are overwhelmed with N elements, we keep walking the entire SC queue for no particular reason?
> 
> This improves javac benchmark considerably (look at both wall time and user time):
> 
> 
> # BEFORE
> Benchmark 1: build/linux-x86_64-server-release/images/jdk/bin/java -XX:CacheDataStore=app.cds -Xmx256m -Xms256m -XX:+UnlockExperimentalVMOptions -XX:+UseEpsilonGC -cp JavacBenchApp.jar JavacBenchApp 50
>   Time (mean ± σ):     348.3 ms ±   2.8 ms    [User: 949.0 ms, System: 101.4 ms]
>   Range (min … max):   344.0 ms … 355.3 ms    100 runs
> 
> # AFTER
> Benchmark 1: build/linux-x86_64-server-release/images/jdk/bin/java -XX:CacheDataStore=app.cds -Xmx256m -Xms256m -XX:+UnlockExperimentalVMOptions -XX:+UseEpsilonGC -cp JavacBenchApp.jar JavacBenchApp 50
>   Time (mean ± σ):     312.2 ms ±   2.6 ms    [User: 577.8 ms, System: 95.4 ms]
>   Range (min … max):   307.1 ms … 318.5 ms    100 runs

src/hotspot/share/compiler/compilationPolicy.cpp line 914:

> 912:       max_task = task;
> 913:       max_method = method;
> 914:       break;

I think we can just return `task` here.

We don't cache `CompLevel_full_profile` Level3 C1 compiled code and as result we do nothing in the rest of this method for AOT code.

We put SCC tasks on separate queues which will have only these tasks.

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

PR Review Comment: https://git.openjdk.org/leyden/pull/17#discussion_r1747775671