Just discovered empirically that `is SomeInterface` is slower than `is SomeClass`, maybe about 6x sl...

# random

[Youssef Shoaib [MOD]]

Just discovered empirically that is SomeInterface is slower than is SomeClass, maybe about 6x slower. That pushed me to switch a sealed interface to a sealed class

[Frank Bouwens]

Interessting find!

[Fudge]

Can you share some benchmarks? It seems Platform dependant and something that would be optimized by the JIT

[ephemient]

as classes are only single inheritance, class instanceof can walk up the parent tree only

[ephemient]

whereas interfaces are multiple inheritance, so interface instanceof is more complex

[ephemient]

still I would have hoped that Java's PermittedSubclasses attribute would be able to optimize that

[ephemient]

(which Kotlin sealed compiles to as long as the JVM target 17+ iirc)

[Youssef Shoaib [MOD]]

Oh yeah my JVM target is like 11, so maybe that's why

[ephemient]

https://mail.openjdk.org/pipermail/hotspot-runtime-dev/2020-August/041051.html

[ephemient]

in principle the JVM could optimize it better even without PermittedSubclasses (there's already CHA during classloading) but it's not really significant so it seems the current implementation is considered good enough

[eygraber]

@Youssef Shoaib [MOD] are you planning on filing an issue with your findings? 6x is pretty significant for a common pattern.

[ephemient]

file an issue where? its performance is specific to different JVM implementations (eg hotspot will do differently than j9)

[ephemient]

the mailing list link I sent above is talking about hotspot in particular

[eygraber]

Ah I was assuming this was a Kotlin specific problem. Thanks for the link.

[efemoney]

6x of 0 is 0 AND 6x of 2 is still less than 100 🙂

Point is, does using interfaces really reduce performance of your entire application?

I would wager that for majority of apps, answer would be a no.

The distinction btw sealed interface and sealed class has real world impact either in mental model of a domain or in its maintainability, I would say to be careful in making such a change in your own applications without doing some benchmarking

[Youssef Shoaib [MOD]]

It was making a significant difference in a benchmark I was running. I'm writing a library, not an end-user app, so my goal is to reduce overhead as much as possible. For an app, I would never care about this. For my library, I was running millions of if these checks, so it genuinely made a difference to the overall performance