Fix and extend Array.apply varargs optimization

[retronym]

The existing optimization for Array.apply[T] overreaches by assuming that the ArrayValue carrying the varargs will have the same element type as the implicit class tag. However, this is not the case when the type argument is a generic type:

scala> def foo[T <: AnyRef : reflect.ClassTag](t: T) = Array.apply[T](t); foo[String]("")
java.lang.ClassCastException: [Ljava.lang.Object; cannot be cast to [Ljava.lang.String;
  ... 32 elided

This PR limits the existing optimization to cases where:

• The class tag is an implicit arg
• The class tag is materialized (not found via normal implicit search)

The second criterion excludes the call above.

A new, fallback optimization is provided that avoids a temporary
array and collection wrapper by emitting:

val arr = classTag.newArray(N)
ScalaRunTime.arrayUpdate(arr, N, elemN)
...
arr

Fixes scala/bug#11882

[retronym]

Found a long standing bug in the existing optimization in the process!

[lrytz]

% scalac t4216.scala
% <in process execution of run/t4216.scala> > t4216-run.log
% diff /home/jenkins/workspace/scala-2.12.x-validate-main@3/test/files/run/t4216.check /home/jenkins/workspace/scala-2.12.x-validate-main@3/test/files/run/t4216-run.log
@@ -9,7 +9,9 @@ scala> f(".")
 res0: java.util.List[String] = [.]
 
 scala> f(0)
-res1: java.util.List[Int] = [0]
+java.lang.ClassCastException: [I cannot be cast to [Ljava.lang.Object;
+  at .f(<console>:12)
+  ... 63 elided

[retronym]

scala> def concrete = Array.apply[String]("")
concrete: Array[String]

scala> :javap concrete#concrete
  public java.lang.String[] concrete();
    descriptor: ()[Ljava/lang/String;
    flags: ACC_PUBLIC
    Code:
      stack=4, locals=1, args_size=1
         0: iconst_1
         1: anewarray     #21                 // class java/lang/String
         4: dup
         5: iconst_0
         6: ldc           #23                 // String
         8: aastore
         9: checkcast     #25                 // class "[Ljava/lang/Object;"
        12: checkcast     #27                 // class "[Ljava/lang/String;"
        15: areturn
      LineNumberTable:
        line 11: 0
      LocalVariableTable:
        Start  Length  Slot  Name   Signature
            0      16     0  this   L$line3/$read$$iw$$iw$;

scala> def generic[T <: AnyRef: reflect.ClassTag](t: T) = Array.apply[T](t); generic[String]("")
generic: [T <: AnyRef](t: T)(implicit evidence$1: scala.reflect.ClassTag[T])Array[T]
res3: Array[String] = Array("")

scala> :javap generic#generic
  public <T extends java.lang.Object> T[] generic(T, scala.reflect.ClassTag<T>);
    descriptor: (Ljava/lang/Object;Lscala/reflect/ClassTag;)[Ljava/lang/Object;
    flags: ACC_PUBLIC
    Code:
      stack=4, locals=4, args_size=3
         0: aload_2
         1: iconst_1
         2: invokeinterface #30,  2           // InterfaceMethod scala/reflect/ClassTag.newArray:(I)Ljava/lang/Object;
         7: astore_3
         8: getstatic     #35                 // Field scala/runtime/ScalaRunTime$.MODULE$:Lscala/runtime/ScalaRunTime$;
        11: aload_3
        12: iconst_0
        13: aload_1
        14: invokevirtual #39                 // Method scala/runtime/ScalaRunTime$.array_update:(Ljava/lang/Object;ILjava/lang/Object;)V
        17: aload_3
        18: checkcast     #41                 // class "[Ljava/lang/Object;"
        21: areturn
      LineNumberTable:
        line 11: 0
      LocalVariableTable:
        Start  Length  Slot  Name   Signature
            0      22     0  this   L$line7/$read$$iw$$iw$;
            0      22     1     t   Ljava/lang/Object;
            0      22     2 evidence$1   Lscala/reflect/ClassTag;
    Signature: #22                          // <T:Ljava/lang/Object;>(TT;Lscala/reflect/ClassTag<TT;>;)[TT;
    MethodParameters:
      Name                           Flags
      t                              final
      evidence$1                     final

[retronym]

@lrytz I've reworked this a little.

[lrytz]

LGTM

[sjrd]

I assume I was tagged as reviewer for the interaction with Scala.js. Don't worry: Scala.js runs before cleanup, so this PR won't affect Scala.js.

Incidentally, I would like it if this particular optimization for Array.apply would happen earlier in the pipeline, so that we could benefit from it for free. Currently we replicate it in our back-end. ;)

[retronym]

Earlier in the pipeline actually sounds good as we could detect whether the type argument was concrete or not. At least worth proposing for dotc.

[retronym]

Detect in a more straightforward manner, I should say.

…

On Sun, 6 Dec 2020 at 20:09, Jason Zaugg ***@***.***> wrote: Earlier in the pipeline actually sounds good as we could detect whether the type argument was concrete or not. At least worth proposing for dotc.

[SethTisue]

@retronym is this ready for merge, or do you want to explore the possibility of moving this to earlier in the pipeline here in Scala 2?

[retronym]

I'm happy enough with the solution here, so booking the progress!

[SethTisue]

regression? see scala/scala-dev#760