:wave: I've been testing some of the new incremen...

# compiler

[Zac Sweers]

👋 I've been testing some of the new incremental common compilation props and found a bug in 2.4.20 that's fixed (I think incidentally) in 2.5.0-Beta1

I think that's from one or both of these
- github.com/JetBrains/kotlin/commit/… (hash source file paths for JVM metadata fragment names)
- github.com/JetBrains/kotlin/commit/… (drop the disk-based common metadata KLIB writing)

Would it be worth filing a FR to backport those to 2.4.21? Creating the parent directory in writeTo might be enough there, but y'all would know better

[Zac Sweers]

Context: Metro's contribution hints are FIR-generated top-level declarations, so under KT-75864 every contributing file is always dirty. In KMP projects with contributions in commonMain, that makes every JVM incremental build a full rebuild (KT-62686), even for jvmMain-only edits. On 2.5.0-Beta1 the common-sources flag fixes that for us, and a backport would let 2.4.x users opt in a bit earlier

[Zac Sweers]

The crash I ran into trying it

e: java.nio.file.NoSuchFileException: .../build/kotlin/compileKotlinJvm/cacheable/metadata-jvm/commonMain/default/linkdata/package_metro.hints/metro/hints/testCommonValueDev_zacsweers_metro_AppScope.kt.knm
      at org.jetbrains.kotlin.cli.pipeline.jvm.metadata.JvmMetadataComponentWriter.writeTo(JvmMetadataComponentWriter.kt:38)

The synthetic file's name includes its package path (metro/hints/...kt) and JvmMetadataComponentWriter only creates the package fragment directory, so the nested parent directory doesn't exist

[Pavel Kunyavskiy [JB]]

What's your exact setup? As far I remember IC shouldn't be enabled in multiplatform projects at all in 2.4.20, unless you opt-in with some very scary flag.

[Pavel Kunyavskiy [JB]]

Or, that's exactly the problem, as it not enabled, but tries to do something and it crashes?

[Zac Sweers]

I am trying the scary flags yes 🙂

[Ivet]

Hi, thank you for the report!

Could you please share the flags you used for 2.4.20 and 2.5.0-Beta1?

[Zac Sweers]

will circle back with them!

[Zac Sweers]

the flags in 2.4.20

kotlin.internal.incremental.enableUnsafeOptimizationsForMultiplatform=true
kotlin.internal.jvm.enableKmpClasspathMetadataForIncrementalCompilation=true

For 2.5.0-Beta1 I used

kotlin.jvm.enableIncrementalCompilationOfCommonSources=true

The 2.4.20 configuration triggers the metadata-writing crash with Metro contributions in commonMain. With 2.5.0-Beta1, JVM builds stay incremental when editing only jvmMain

[Zac Sweers]

I subsequently also tested kotlin.kmp.separateCompilation=true, including alongside the common-sources IC flag. That testing is captured in here, where I reported a separate single-target producer metadata issue. The JS/Wasm deletion crash also reproduces with separate compilation disabled

[Ivet]

Thanks for providing the flags, that gives me the full picture.

To answer your question, we don’t plan to backport this functionality to 2.4.21. The 2.4.20 implementation was a prototype that used a different mechanism from 2.5.0: in 2.4.20, we generated knm output per file, while 2.5.0 uses only the IC cache. Given that difference, I don’t think it’s worth spending time testing the metadata behavior in versions before 2.5.0.

The change is intended to enable incremental compilation within a module when you edit a file in that module’s commonMain source set. Could you clarify what you meant by “With 2.5.0-Beta1, JVM builds stay incremental when editing only `jvmMain`”? Builds should already be incremental when you edit jvmMain, by default.

Separate compilation should help in multi-module scenarios.