Merge pull request #1428 from SethTisue/signature-polymorphic-methods

_posts/2023-07-17-signature-polymorphic-methods.md

@@ -0,0 +1,235 @@
+---
+category: blog-detail
+post-type: blog
+by: Seth Tisue, Lightbend
+title: "Signature polymorphic methods in Scala"
+---
+
+Java 7 introduced a curious and little-known feature to the Java
+Virtual Machine: "signature polymorphic" methods. These methods have
+strangely malleable types.
+
+This blog post explains the feature and why it exists. We also delve
+into how it is specified and implemented in both Scala 2 and Scala 3.
+
+The Scala 3 implementation is new, and that's the occasion for this
+blog post. Thanks to this recent work, **Scala 3 users can now access
+the entire Java reflection API**, as of Scala 3.3.0.
+
+## Should I keep reading?
+
+Signature polymorphism is admittedly an obscure feature. When you need
+it you need it, but the need doesn't arise in ordinary Scala
+code. Thus, the remaining material may be of interest primarily to JVM
+aficionados, Scala and Java language mavens, and compiler hackers.
+
+## When is signature polymorphism needed?
+
+Compiler support is needed when you use some portions of the Java
+reflection API, namely `MethodHandle` (since Java 7) and `VarHandle`
+(since Java 11).
+
+`MethodHandle` provides reflective access to methods on JVM classes,
+regardless of whether the methods were defined in Java, Scala, or some
+other JVM language. `VarHandle` does the same, but for fields.
+
+The polymorphism of these methods makes them more efficient, by
+avoiding boxing overhead when primitive values are passed, returned,
+or stored.
+
+## Is signature polymorphism supported in Scala?
+
+Yes: since Scala 2.11.5, and more fully since Scala 2.12.16.  Scala 3
+now has the support too, as of Scala 3.3.0.
+
+The initial Scala 2 implementation was done by [Jason Zaugg] in 2014
+and refined later by [Lukas Rytz]. The latest version, with all fixes,
+landed in Scala 2.12.16 (released June 2022).
+
+Recently, [Dale Wijnand] ported the feature to Scala 3, with the
+assistance of [Guillaume Martres] and myself, [Seth Tisue].
+
+Jason, Lukas, Dale, and myself are members of the Scala compiler team
+at [Lightbend]. We maintain Scala 2 and also contribute to Scala 3.
+Guillaume has worked on the Scala 3 compiler for some years, previously
+at [LAMP] and now at the [Scala Center].
+
+[Jason Zaugg]: https://github.com/retronym
+[Lukas Rytz]: https://github.com/lrytz
+[Dale Wijnand]: https://github.com/dwijnand
+[Seth Tisue]: https://github.com/SethTisue
+[Guillaume Martres]: https://github.com/smarter
+[Lightbend]: https://lightbend.com
+[LAMP]: https://www.epfl.ch/labs/lamp/
+[Scala Center]: https://scala.epfl.ch
+
+## What signature polymorphic methods exist?
+
+You may already have run into this feature if you have used the
+`MethodHandle` and `VarHandle` classes from the Java reflection API in
+the Java standard library.
+
+In fact, `MethodHandle` and `VarHandle` are the _only_ places you
+could possibly have run into this feature!
+
+That's because users are not allowed to define their own signature
+polymorphic methods. Only the Java standard library can do that, and
+so far, the creators of Java have only used the feature in these two
+classes.
+
+## What does "signature polymorphism" mean, exactly?
+
+There is a formal description in [JLS 15.12.3], but a more readable
+version is in the [Javadoc for
+`MethodHandle`](https://docs.oracle.com/en/java/javase/11/docs/api/java.base/java/lang/invoke/MethodHandle.html).
+It says:
+
+> A signature polymorphic method is one which can operate with any of
+> a wide range of call signatures and return types.
+
+and:
+
+> In source code, a call to a signature polymorphic method will
+> compile, regardless of the requested symbolic type descriptor. As
+> usual, the Java compiler emits an invokevirtual instruction with the
+> given symbolic type descriptor against the named method. The unusual
+> part is that the symbolic type descriptor is derived from the actual
+> argument and return types, not from the method declaration.
+
+Note that generics are not sufficient to express this level of
+flexibility, for two reasons:
+
+First, Java generics only work on reference types, not primitive
+types.  Scala does not have this limitation, but pays for it by
+incurring boxing at run-time when primitive types are used in generic
+contexts.
+
+Second, methods (in both languages) may only have a fixed number of
+type parameters, but we need one varying type for each parameter
+of the method we want to call reflectively.
+
+The following example should help make all of this clearer.
+
+[JLS 15.12.3]: https://docs.oracle.com/javase/specs/jls/se17/html/jls-15.html#jls-15.12.3
+
+## How do I call a signature polymorphic method from Scala?
+
+Take `MethodHandle` for example. It provides an `invokeExact`
+method. Its signature as seen from Scala is:
+
+    def invokeExact(args: AnyRef*): AnyRef
+
+Signature polymorphism means that the `AnyRef`s here are just
+placeholders for types to be supplied later.
+
+To see this work in practice, let's adapt an example from
+the Javadoc. From Scala, we'll make a reflective call to the `replace`
+method on a `String`:
+
+    import java.lang.invoke._
+    val mt = MethodType.methodType(
+      classOf[String], classOf[Char], classOf[Char])
+    val mh = MethodHandles.lookup.findVirtual(
+      classOf[String], "replace", mt)
+    val s = mh.invokeExact("daddy", 'd', 'n'): String
+
+If we paste this into the Scala REPL (2 or 3), we see:
+
+    val s: String = nanny
+
+Signature polymorphism helped us here in two ways:
+
+* The arguments `d` and `n` will not be passed as `Object` or boxed to
+  `java.lang.Character` at runtime, but will be passed directly as
+  primitive `Char`s.
+* The result comes back as a `String` without needing to be checked
+  or cast at runtime.
+
+## Are these methods good for anything else?
+
+Great question!
+
+Doesn't it seem puzzling that the designers of Java would go to so
+much trouble to make Java reflection faster? If I care so much about
+performance, shouldn't I avoid using reflection entirely?
+
+The real reason these methods need to be fast is to aid efficient
+implementation of dynamic languages on the JVM. `MethodHandle` was
+added to the JVM at the same time as `invokeDynamic`, as part of
+[JSR-292], which aimed to support efficient implementation of JRuby
+and other alternative JVM languages. (`invokeDynamic` is additionally
+used for implementing lambdas, in both Java and Scala; see [this
+writeup on Stack Overflow].)
+
+[JSR-292]: https://www.infoq.com/articles/invokedynamic/
+[this writeup on Stack Overflow]: https://stackoverflow.com/questions/30002380/why-are-java-8-lambdas-invoked-using-invokedynamic
+
+## How is this implemented in Scala 2?
+
+Jason Zaugg describes his initial JDK 7 implementation in [PR 4139]
+and shows how the resulting bytecode looks.
+
+See also these well-documented followups: [PR 5594] for JDK 9,
+[PR 9530] for JDK 11, and [PR 9930] for JDK 17.
+
+[PR 4139]: https://github.com/scala/scala/pull/4139
+[PR 5594]: https://github.com/scala/scala/pull/5594
+[PR 9530]: https://github.com/scala/scala/pull/9530
+[PR 9930]: https://github.com/scala/scala/pull/9930
+
+## What's different in the Scala 3 version?
+
+We had to work harder in Scala 3 because it wasn't enough to have an
+an in-memory representation for signature polymorphic call sites.  The
+call sites must also have a representation in TASTy, so we had to add
+a new TASTy node type. (Scala 2 pickles only represent method
+signatures; in contrast, TASTy represents method bodies too.)
+
+To represent a signature polymorphic call site internally, we
+synthesize a method type based on the types at the call site.  One can
+imagine the original signature-polymorphic method as being infinitely
+overloaded, with each individual overload only being brought into
+existence as needed.
+
+For details, see [the pull
+request](https://github.com/lampepfl/dotty/pull/16225).
+
+### The path not taken
+
+Along the way we explored an alternative approach, suggested by Jason,
+which involved rewriting each call site to include a cast to a
+structural type containing an appropriately typed method.
+
+In that version, the `replace` call-site in the example above was
+rewritten from:
+
+    mh.invokeExact("daddy", 'd', 'n'): String
+
+to:
+
+    mh.asInstanceOf[
+      MethodHandle {
+        def invokeExact(a0: String, a1: Char, a2: Char): String
+      }
+    ].invokeExact("daddy", 'd', 'n')
+
+(The actual rewrite was applied to in-memory ASTs, rather than to
+source code.)
+
+The transformed code could be written and read as TASTy without
+trouble. Later in compilation, we detected which call sites are the
+product of this transform, drop the cast, and emit the correct
+bytecode.
+
+In the end, we didn't go with this approach. As Sébastien Doeraene
+pointed out, although this approach avoided adding a new TASTy tag, it
+also gave new semantics to existing tags that older compilers wouldn't
+understand. Therefore the work still couldn't ship until the next
+minor version of the compiler.  Besides, avoiding the new tag
+complicated the implementation.
+
+## Questions? Discussion?
+
+These are welcome on the Scala Contributors forum thread at:
+
+* (TODO Discourse link, with link back to this post)