KSP Incremental Processing: Clean source files should have generated output cache accessible via Resolver.

[bcorso]

Dagger recently had two incremental processing issues related to KSP:

google/dagger#4063
google/dagger#4054

Repro Example

As an example, take the following source files:

class B extends A {
  @Inject Foo foo;
}

@SomeOtherAnnotation
class A {
  @Inject Bar bar;
}

Suppose that after a clean build we get an output map like this:

// Clean build
Accumulated source to outputs map
  src/main/java/A.java:
    build/generated/ksp/debug/java/byRounds/1/A_MembersInjector.java
    build/generated/ksp/debug/java/byRounds/1/A_SomeOtherThing.java
  src/main/java/B.java:
    build/generated/ksp/debug/java/byRounds/1/B_MembersInjector.java

Now, suppose that B changes and becomes "dirty". The issue is that Dagger ends up generating A_MembersInjector.java when it inspects the superclass of B because it thinks that A_MembersInjector.java has not been generated yet since it is not found via the Resolver. However, generating A_MembersInjector.java overwrites the output map's cache for A, which is then missing generated sources from other annotations originating from A, e.g. A_SomeOtherThing.java:

// B dirty (results in incorrect outputs for A)
Accumulated source to outputs map
  src/main/java/A.java:
    build/generated/ksp/debug/java/byRounds/1/A_MembersInjector.java
  src/main/java/B.java:
    build/generated/ksp/debug/java/byRounds/1/B_MembersInjector.java

[bcorso]

(Note: @ting-yuan and I have discussed this offline, but putting it here for completeness.)

Short-term fix:

There is a short-term fix that we can implement within Dagger to avoid generating A_MembersInjector while inspecting class B when A is a source file in the same compilation unit (which we can check using Origin).

However, I think this can be fixed on the KSP side to make things easier (and more efficient) for developers.

Long-term fix:

I think a better, long-term fix is to change how KSP handles its output cache. In particular, if A is "clean" then none of its outputs should change (i.e. any output which originates from A). In addition, we should make the cached output of A available via the Resolver so that processors can "see" the generated output exists and avoid trying to re-generate it. This approach should be more similar to how things work in Javac, and I think it will also be more intuitive to developers using KSP.