<PR for AppCDS/AOT in Compose Multiplatform for De...

# compose-desktop

[Alexander Maryanovsky]

PR for AppCDS/AOT in Compose Multiplatform for Desktop including speedup times on a test project.
TL;DR:
On macOS AOT speeds up startup by 2-5x
On Windows AOT speeds up startup by 1.5x

[wwalkingg]

that's great! you always bring so many good news!

[אליהו הדס]

openjdk.org/jeps/8335368
For now, Leyden isn't real AoT, but it's coming!

[Alexander Maryanovsky]

They first need to fix https://youtrack.jetbrains.com/issue/JBR-9098/AppCDS-writes-local-paths-into-.jsa-archive

[אליהו הדס]

It seems to me that the JDK 25 AOT cache (not AppCDS) tolerates the relocation of JARs as long as the relative structure of the classpath remains identical.

[Alexander Maryanovsky]

It doesn't; I checked

[Alexander Maryanovsky]

I mean it does tolerate it, but so does 21 with AppCDS. But it still tries to read the files at the classpath, and when it doesn't find them, it takes forever

[Alexander Maryanovsky]

https://github.com/JetBrains/JetBrainsRuntime/pull/618

[אליהו הדס]

I understand, the problem only exists on Windows, your PR only activates it on Mac and Linux ?

[Alexander Maryanovsky]

No, the PR is for all platforms

[אליהו הדס]

I don't understand anymore, but according to this, it still improves performance on Windows ? it doesn't degrade it then?

[Alexander Maryanovsky]

It depends on the number of AOT-d classes. For a HelloWorld Compose app it's still beneficial.

[אליהו הדס]

So on Windows you recommend doing the training on the client side?

[Alexander Maryanovsky]

If you can tolerate the first run being slow, I think so. Until it's fixed, anyway.
Although actually I’m not sure how this bug presents with the auto-cds option.

[אליהו הדס]

Personally, I activated it 7 or 8 months ago on Windows users who had been using it for a while on a small application I have, and I received very positive feedback despite everything; the cache was created in the CI.

[אליהו הדס]

But I noticed increased RAM usage at the beginning, probably because the cache needs to be loaded in addition to the JVM.

[Alexander Maryanovsky]

Did you measure startup time?

[אליהו הדס]

I'm going to measure; I still have the old builds. These days I only use the native GraalVM compilation; I don't need reflection or class loader.

[אליהו הדס]

I've measured it now; initially, I activated it without class linking, and I got a gain of 1.5x. Then I released another build with class linking, and the gain was 2x compared to the first release.

[אליהו הדס]

That's in hot start mode.

[אליהו הדס]

For the cold start, I missed the measurement for the first release without anything, but the version without class linking takes about 8 seconds and with class linking, 6 seconds.

[Alexander Maryanovsky]

Compare a run on the machine it’s built on (right from the build/ directory) to a run on another machine. Or just move the app to a different directory and delete build/

[אליהו הדס]

I installed old installers that were built from a ci

[Alexander Maryanovsky]

Ah

[Alexander Maryanovsky]

You can still compare it with some effort. Check the classpath and make sure the files listed there exist on your machine.

[אליהו הדס]

I measured a penalty of 60ms

[אליהו הדס]

Now I'm almost certain that out of those 6 seconds, 5 seconds are due to a bug with AWT. I have the same problem with IntelliJ; on the same machine, on Linux, I barely have time to see the splash screen, and on Windows, it takes between 10 and 15 seconds to start.