that looks promising! how did you test the perf improv?

by running the same example in the old binary and compare it to the new binary

The idea is to instrument the binary using a typical workload and then use that profile to compile an optimized binary

cool, will try with clojure-lsp, we can test with kondo too

where did you find about that? some graal post?

yes: https://www.graalvm.org/latest/reference-manual/native-image/guides/optimize-native-executable-with-pgo/

nice, will take a look, curious to check how much could affect lsp or kondo

@U04V15CAJ how did you get the PGO for bb, did you buy or something graalvm enterprise to get that?

No, use Oracle GraalVM 21 instead of CE

ah so it should work for graal 21 already? nice

yes, but not the CE version

There is no longer an EE version, it's now just called Oracle GraalVM and it's free to use

yeah, just realized that, awesome!

thanks

Still doing some tests but aparrently, 15s to lint clojure-lsp project without cache and 10s with pgo 👀 the binary size went from 89M -> 94M but seems to be worth it

On Nubank biggest project with cache 52s / 38s which is amazing

Weird, bb went from 80 to 65 mb. How did you create the iprof file?

I could not create it using in an editor since the iprof is only generated after the program exit with status 0, but I used clojure-lsp diagnostics which does almost the same things that does in the editor

I could force a exit 0 when using in the editor coding that, but don't think would change that much

@ericdallo I'm reverting PGO in bb for now since the iprof file cannot reliably be re-used over multiple builds it seems

and I don't want to build twice

what you mean build twice?

well, you have to build once to build the instrumented binary

which then produces the iprof file

and then you build the final optimized binary

ah, so you would need to add that to CI or something?

well yeah

I see, so if you change bb's code you need to always generate a new iprof

so far it worked with a shared iprof file

but when I bumped httpkit, it didn't work anymore

I see

that's bad

I did expect it, I was surprised it did work actually initially

at least would be a new build for all platforms, not one for each platform

so, for clojure-lsp we have 5 builds, which is all platforms we support, that would be 6 then

well, I tried that, making an iprof locally and share that over other platforms, but it didn't work anymore

Damm

I'll probably revert as well so

tks for the headsup

how did you do it so far then?

I implemented that this last weekend and only tested on my machine 😅 so when I release a new lsp version we would probably face that issue with other people using

but how did you implement what

anyway it might still work for lsp, but httpkit maybe does something weird

yeah, good point, we already have running for some days on nightly, so I suppose some people would already have complains :thinking_face:

hehe funny bug, I can repro, feel free to open a issue

That should happen on any suppress that has a comment above I suppose

https://github.com/clojure-lsp/clojure-lsp/issues/1678

thanks!