AnyRef specialization seems broken in master

[scabug]

So, while working on #3575 I noticed that AnyRef specialization didn't seem to be working. I created a simple test of AnyRef specialization, noted the failure with master, then tried compiling an example with Scala 2.9.1 and it worked.

Here's the program:

class C[@specialized(Int, AnyRef) A](v:A)                                                                                   
                                                                                                                            
object Test {                                                                                                               
  def main(args:Array[String]) {                                                                                            
    println(new C(123).getClass.getName)                                                                                    
    println(new C("abc").getClass.getName)                                                                                  
  }                                                                                                                         
}

The output on 2.9.1 is:

C$mcI$sp
C$mcT$sp

Whereas on master it is:

C$mcI$sp
C

I know that there was a change recently from using the (totally arbitrary) T to the (Java standard) L for referring to Object/AnyRef, as well as other AnyRef bugfixes. I'm currently trying to do a bisect to figure out when this example stopped working, but haven't figured it out yet.

[scabug]

Imported From: https://issues.scala-lang.org/browse/SI-5488?orig=1
Reporter: @non

[scabug]

@non said (edited on Feb 16, 2012 4:00:40 AM UTC):
I can confirm that this code worked as of commit a2ce52f83, with the following output:

C$mcI$sp
C$mcT$sp

So this bug was introduced after that point.

[scabug]

@paulp said:
The oddsmakers will be taking a huge, huge beating if it wasn't bb23d766bce .

[scabug]

@paulp said:
Wait, what's the point of specializing a single type parameter on AnyRef?

If I make it realistic, it crashes 2.9, but works a little bit in trunk.

class C[@specialized(Int, AnyRef) A, @specialized(Int, AnyRef) B](v:A, w:B)                                                                                  
                                                                                                                            
object Test {                                                                                                               
  def main(args:Array[String]) {                                                                                            
    println(new C(123, 123).getClass.getName)                                                                                    
    println(new C("abc", 123).getClass.getName)                                                                                  
    println(new C(123, "abc").getClass.getName)                                                                                  
    println(new C("abc", "abc").getClass.getName)
  }                                                                                                                         
}

% scalac29 -d /tmp test/files/run/t5488.scala error: scala.tools.nsc.symtab.Types$TypeError: type mismatch;
 found   : java.lang.String("abc")
 required: A$sp
	at scala.tools.nsc.typechecker.Contexts$Context.error(Contexts.scala:298)
	at scala.tools.nsc.typechecker.Infer$Inferencer.error(Infer.scala:207)
	at scala.tools.nsc.typechecker.Infer$Inferencer.typeError(Infer.scala:217)


% scalac3 -d /tmp test/files/run/t5488.scala 
$ scala3 -cp /tmp Test
C$mcII$sp
C
C
C

[scabug]

@paulp said:
Where of course "a little bit" means "at least doesn't crash us out of seeing mcII."

[scabug]

@paulp said:
Every nontrivial thing I try crashes 2.9, no wonder I wasn't setting the bar too high.

Here's a nice realistic one.

class C[@specialized(Int, AnyRef) A, @specialized(Int, AnyRef) B, @specialized(Int, AnyRef) C](f: (A, B) => C)

object Test {
  def main(args:Array[String]) {
    def show(x: Any) = println(x.getClass.getName)
    
    show(new C((x: Int, y: Int) => 1))
    show(new C((x: Int, y: Int) => "abc"))
    show(new C((x: Int, y: AnyRef) => 1))
    show(new C((x: Int, y: AnyRef) => "abc"))
    show(new C((x: AnyRef, y: Int) => 1))
    show(new C((x: AnyRef, y: Int) => "abc"))
    show(new C((x: AnyRef, y: AnyRef) => 1))
    show(new C((x: AnyRef, y: AnyRef) => "abc"))
  }
}

Only works for III.

C$mcIII$sp
C
C
C
C
C
C
C

[scabug]

@non said:
The reason why specializing a single parameter is nice is that in principle it will let us use @final/inline methods from C$mcL$sp instead of from C. I guess that's a little bit perverse since it's working around the weird subtyping/specialization stuff, but I think we're stuck with that design so we might as well make the best of it.

But I agree, your example is a more realistic case.

[scabug]

@paulp said:
I found it. Thanks for catching it so quickly.

[scabug]

@paulp said:
If I create LL/LLL/etc specializations, the compiler crashes during mixin finding duplicate members. I don't think it makes any sense to create L specializations but not LL, LLL, etc. so the upshot is that your test case still behaves the same even though this is fixed. Mine on the other hand prints

C0$mcIII$sp
C0$mcIIL$sp
C0$mcILI$sp
C0$mcILL$sp
C0$mcLII$sp
C0$mcLIL$sp
C0$mcLLI$sp
C0

If there is sufficient motivation we can look into whether it makes sense to generate those all-AnyRef specializations, but I have no more time for this right now.

[scabug]

@non said:
OK, that's fair. I'll put that on my list of things to look at later.

[scabug]

@paulp said:
1e648c3862