I’ve been migrating our internal annotation proces...

# ksp

[philglass]

I’ve been migrating our internal annotation processors to KSP and I’m having all sorts of trouble getting the aggregating one to work correctly incrementally. They’re a layer on top of Dagger that supports broadly similar things to Hilt or Anvil.

It’s multi-round and uses symbols from both the classpath (via getDeclarationsFromPackage) and the current compilation unit, but is otherwise pretty simple. I reproduced a simplified, standlone version as a KSP integration test here - you can run it with ./gradlew integration-tests:test --tests 'com.google.devtools.ksp.test.BrokenIncrementalIT.*'

Any idea where I’m going wrong? Or am I running into bugs in KSP?

[Ting-Yuan Huang]

Thanks for the repro case! Yes, this is a bug in KSP 😢

In short, the dirtiness isn't propagated properly and therefore AppModule1 and AutoIncludeModule_app_AppModule1 are incorrectly considered as clean*1. So they are not included in subsequent processing and therefore getDeclarationsFromPackage didn't return AutoIncludeModule_app_AppModule1.

A workaround is to associate AppModule1 to MergedModule_AppMergedModule manually.

BTW, there is a missing association from AppModule1 to AutoIncludeModule_app_AppModule1. This needs to be fixed in AutoIncludeProcessor.

*1 They are dirty because MergedModule_AppMergedModule is dirty and it needs AutoIncludeModule_app_AppModule1 , which depends on AppModule1

[philglass]

Thanks for taking a look! Are both tests failing for the same reason (the first one is adding a file to the current compilation unit, and the second to the classpath)?

BTW, there is a missing association from AppModule1 to AutoIncludeModule_app_AppModule1. This needs to be fixed in AutoIncludeProcessor.

Is this not enough? Kotlin Poet should be passing an appropriate Dependencies instance to CodeGenerator for me.

[Ting-Yuan Huang]

Ahh I missed this 🤦 Yes this should be good.