Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Fixes undercompilation on inheritance on same source
### background In sbt 0.13 days, we could ignore the relationship between two classes defined in the same `*.scala` source file, because they will be compiled anyway, and the invalidation was done at the source file level. With class-based namehashing, the invalidation is done at the class level, so we can no longer ignore inheritance relationship coming from the same source, but we still have old assumptions scattered around the xsbt-dependency implementation. ### what we see without the fix ``` [info] Compiling 1 Scala source to ... .... [debug] [inv] internalDependencies: [debug] [inv] DependencyByInheritance Relation [ [debug] [inv] xx.B -> gg.table.A [debug] [inv] xx.Foo -> xx.C [debug] [inv] ] [debug] [inv] DependencyByMemberRef Relation [ [debug] [inv] xx.B -> gg.table.A [debug] [inv] xx.Hello -> gg.table.A [debug] [inv] xx.Foo -> xx.C [debug] [inv] ] .... Caused by: java.lang.AbstractMethodError: xx.Foo.buildNonemptyObjects(II)V ``` First, we see that `xx.C -> xx.B DependencyByInheritance` relationship is missing. Second, the error message seen is `java.lang.AbstractMethodError` happening on `xx.Foo`. ### what this changes This change changes two if expressions that was used to filter out dependency info coming from the same source. One might wonder why it's necessary to keep the local inheritance info, if two classes involved are compiled together anyways. The answer is transitive dependencies. Here's likely what was happening: 1. `gg.table.A` was changed, 2. causing `xx.B` to invalidate. 3. However, because of the missing same-source inheritance, it did not invalidate `xx.C`. 4. This meant that neither `xx.Foo` was invalidated. 5. Calling transform method on a new `xx.Foo` causes runtime error. By tracking same-source inheritance, we will now correctly invalidate `xx.C` and `xx.Foo`. I think the assumption that's broken here is that "we don't need to track inheritance that is happening between two classes in a same source." ### Is this 2.11 only issue? No. The simple trait-trait inheritance reproduction alone will not cause problem in Scala 2.12 because of the [compile-to-interface](http://www.scala-lang.org/news/2.12.0/#traits-compile-to-interfaces) traits. However, not all traits will compile to interface. This means that if we want to take advantage of the compile-to-interface traits, we still should keep track of the same-source inheritance, but introduce some more logic to determine whether recompilation is necessary. Fixes sbt/zinc#417 Rewritten from sbt/zinc@05482d1
- Loading branch information