Java annotations with RetentionPolicy.CLASS are retained at runtime (scala 2.8.1)

[scabug]

If I annotate a class in scala, using a annotation defined in java with RetentionPolicy.CLASS, the annotation is retained at runtime.

Here's a typical example using Peter Kriens' annotations for OSGi:

@aQute.bnd.annotation.component.Component class MyComp

If I launch the REPL with my class in the classpath, but without the definition of the annotation, and I simply try to instantiate the class, I get the following error:

scala> val c = new MyComp
error: error while loading MyComp, Missing dependency 'class aQute.bnd.annotation.component.Component', required by ./MyComp.class
<console>:5: error: MyComp does not have a constructor
       val c = new MyComp

And if I use Peter Kriens' class dump utility, it shows the following information:

2: attribute                   RuntimeVisibleAnnotations(#16)
2: Number of annotations       2
2: type                        LaQute/bnd/annotation/component/Component;(#12)2: num_element_value_pairs     0
2: type                        Lscala/reflect/ScalaSignature;(#13)2: num_element_value_pairs     1

This is very problematic in an OSGi environment where the package should not even need to be imported.

[scabug]

Imported From: https://issues.scala-lang.org/browse/SI-4788?orig=1
Reporter: Arjun Panday (joune)
Affected Versions: 2.8.1
See #5420

[scabug]

@paulp said:
This should be easy to fix, assigning to meeting looking for a victim.

[scabug]

@soc said:
This is related to (and could probably be the cause of) #5420.

[scabug]

@adriaanm said:
Un-assigning to foster work stealing, as announced in https://groups.google.com/forum/?fromgroups=#!topic/scala-internals/o8WG4plpNkw

[scabug]

@adriaanm said:
Unassigning and rescheduling to M6 as previous deadline was missed.

[scabug]

@soc said:
Funny coincidence ... I just fixed this and #5420 5 minutes ago. Is it ok if I create a PR on Thursday?

[scabug]

@adriaanm said:
Sure, doesn't seem like it cause regressions. Thanks!

[scabug]

@soc said:
scala/scala#3459

[scabug]

@soc said (edited on Oct 1, 2014 4:35:29 PM UTC):
Just tried to update this patch to finally get rid of this ticket: https://github.com/soc/scala/tree/SI-4788-new ... apart from the known issue of scalac thinking that classOf doesn't exist in Predef, the test suite now fails with over 400 failures + nondeterminism:

test/partest --failed

Selected 402 tests drawn from previously failed tests

# starting 60 tests in pos
ok  1 - pos/NoCyclicReference.scala             
ok  2 - pos/A.scala                             
ok  3 - pos/S1.scala                            
ok  4 - pos/FPTest.scala                        
ok  5 - pos/S3.scala                            
ok  6 - pos/List1.scala                         
ok  7 - pos/MailBox.scala                       
!!  8 - pos/inline-access-levels                  [compilation failed]
...

Many of the failures are related to annotated elements just "disappearing":

@foo val bar = 1
//Somewhere else:
println(bar)
// scalac: Hey dude, bar doesn't exist, amiright?!

But not all of the failures use annotations.

I'm running out of ideas how a fix to make scalac determine correctly whether a Java annotation should be visible at runtime can wreak such a havoc all over the place.

I'd be happy about any help in this regard.

[scabug]

@lrytz said:
scala/scala#4026

[scabug]

@lrytz said:
Need to re-open this.

In scala/scala#5606 we upgraded partest to a version that changes the mixed compilation mode. This caused a change in the retained annotations, which (probably) should not be the case.

I'll take a look.

[scabug]

@lrytz said:
The fix in scala/scala#4026 only works under separate compilation, because annotations are skipped by the Java parser. We have an open issue for this: #8928, so this one can be closed.

@Retention(value=SOURCE) // not parsed by the Java parser
@interface SAnnotation {}

Tests updated accordingly in scala/scala#5606.