Scenario: Given two files:
Compile first core, then dispatch:
This scenario more commonly occurs in a leiningen project with :aot :all. Files are compiled in alphabetical order with :all. In this case, dispatch.core will be compiled first, then dispatch.dispatch.
- transitively compiles dispatch.dispatch
- writes .class files to compile-path (which is on the classpath)
- assertion passes
- due to prior compile, load dispatch.dispatch__init is loaded via the appclassloader
- ->T constructor will use new bytecode to instantiate a T instance - this uses appclassloader, loaded from compiled T on disk
- however, T class literals are resolved with RT.classForName, which checks the dynamic classloader cache, so uses old runtime version of T, instead of on-disk version
In 1.6, RT.classForName() did not check dynamic classloader cache, so loaded T from disk as with instances. This was changed in
CLJ-979 to support other redefinition and AOT mixing usages.
1) Compile in reverse dependency order to avoid compiling twice.
Either swap the order of compilation in the first example or specify the order in project.clj:
This is a short-term workaround.
2) Move the deftype into a separate namespace from where it is used so it is not redefined on the second compile. This is another short-term workaround.
3) Do not put compile-path on the classpath (this violates current expectations, but avoids loading dispatch__init)
This is not easy to set up via Leiningen currently.
4) Compile each file with an independent Clojure runtime - avoids using cached classes in DCL for class literals.
Probably too annoying to actually do right now in Leiningen or otherwise.
5) Make compilation non-transitive. This is in the ballpark of CLJ-322, which is another can of worms. Also possibly where we should be headed though.
Screening: I do not believe the proposed patch is a good idea - it papers over the symptom without addressing the root cause. I think we need to re-evaluate how compilation works with regard to compile-path (#3) and transitivity (CLJ-322) (#5), but I think we should do this after 1.7. - Alex