Good evening,
The FrankenKotlin experiment is now over, and it failed.
Le 2025-02-28 20:20, Julien Plissonneau Duquène a écrit :
incorrectly generated code in Kotlin plugins
This wasn't as bad as it looked like actually: it was caused by the
Kotlin compiler using the default JDK (21) as its target for API
compatibility, so while the bytecode was compatible with Java 8 it used
some API calls that were introduced in later JDKs instead of
reimplementing them as inline code. Using the JDK 11 again solved that
for the purpose of using that compiler for building Gradle.
After fixing a few more little things the next roadblock was a missing
implicit class declaration for the script that caused the
pluginManagement block (a function call actually) to fail to compile
with an internal assertion error:
* Exception is:
java.lang.AssertionError: Undefined parameter referenced: class
Settings::this
owner=class Settings_gradle; [class Settings_gradle::this]
at
org.jetbrains.kotlin.ir.util.SymbolTable.referenceValueParameter(SymbolTable.kt:420)
at
org.jetbrains.kotlin.psi2ir.generators.ArgumentsGenerationUtilsKt$generateReceiver$$inlined$generateDelegatedValue$1$1.invoke(OnceExpressionValue.kt:63)
at
org.jetbrains.kotlin.psi2ir.generators.ArgumentsGenerationUtilsKt$generateReceiver$$inlined$generateDelegatedValue$1$1.invoke(OnceExpressionValue.kt:32)
at kotlin.SynchronizedLazyImpl.getValue(LazyJVM.kt:74)
It seemed like not much would be needed to fix this, but this was
actually a rabbit hole. My cherry-picks started to touch IR code, then
backends code, then plugins code, and I ended up facing walls of build
errors after applying each of them, and fixing these errors required new
picks, which meant going back to new walls of errors on modules I just
barely got to build. After picking 19 commits and adding 9k lines of
patches I estimated that I wasn't even half done fixing that single
issue and decided that it was definitely not worth continuing.
I computed a few stats to illustrate how this Kotlin project was a
moving target (and probably still is) by then. Between v1.3.31
(2019-04-25) and v1.3.70 (2020-03-03), that's an interval of a bit more
than 10 months, there were 1114 commits that affected some source files
under compiler/ir alone (more than 3 commits/calendar day on average),
for 571k lines of git log -p and 204k lines of git diff. Overall there
were 8236 commits for 4.4M lines of git log -p between the same
releases. For comparison over the same period that's a bit more than 10%
of the activity of the Linux kernel tree in number of commits, and about
44% in number of git log -p lines.
This outcome is a bit frustrating but it was the most likely one, and I
learned a few valuable things in the process.
My plan is now to update the Kotlin package to 2.0.21 (renaming it
kotlin2.0) and use it to build Gradle 8. I'm not sure that both package
(and dependencies) will be in a releasable shape in time for Trixie, but
if they miss the deadline they can still be made available later through
backports.
Cheers,
--
Julien Plissonneau Duquène
