Revive productElementName to extract case class field names #6951

olafurpg · 2018-07-18T11:30:39Z

This commit adds two methods to the scala.Product trait:

trait Product {
  /** Returns the field name of element at index n */
  def productElementName(n: Int): String
  /** Returns field names of this product. Must have same length as productIterator */
  def productElementNames: Iterator[String]
}

Both methods have a default implementation which returns the empty
string for all field names.

This commit then changes the code-generation for case classes to
synthesize a productElementName method with actual class field names.

The benefit of this change is that it becomes possible to pretty-print
case classes with field names, for example

case class User(name: String, age: Int)

def toPrettyString(p: Product): String =
  p.productElementNames.zip(p.productIterator)
   .map { case (name, value) => s"$name=$value" }
   .mkString(p.productPrefix + "(", ", ", ")")

toPrettyString(User("Susan", 42))
// res0: String = User(name=Susan, age=42)

The downside of this change is that it produces more bytecode for each
case-class definition. Running :javacp -c for a case class with three
fields yields the following results

> case class A(a: Int, b: Int, c: Int)
> :javap -c A
  public java.lang.String productElementName(int);
    Code:
       0: iload_1
       1: istore_2
       2: iload_2
       3: tableswitch   { // 0 to 2
                     0: 28
                     1: 33
                     2: 38
               default: 43
          }
      28: ldc           78                 // String a
      30: goto          58
      33: ldc           79                 // String b
      35: goto          58
      38: ldc           80                 // String c
      40: goto          58
      43: new           67                 // class java/lang/IndexOutOfBoundsException
      46: dup
      47: iload_1
      48: invokestatic  65                 // Method scala/runtime/BoxesRunTime.boxToInteger:(I)Ljava/lang/Integer;
      51: invokevirtual 70                 // Method java/lang/Object.toString:()Ljava/lang/String;
      54: invokespecial 73                 // Method java/lang/IndexOutOfBoundsException."<init>":(Ljava/lang/String;)V
      57: athrow
      58: areturn

Thanks to Adriaan's help, the estimated cost per productElementName
appears to be fixed 56 bytes and then 10 bytes for each field with
the following breakdown:

3 bytes for the
string info
(the actual characters are already in the constant pool)
4 bytes for the tableswitch entry
2 bytes for the ldc to load the string
1 byte for areturn

In my opinion, the bytecode cost is acceptably low thanks to the fact
that field name literals are already available in the constant pool.

This commit adds two methods to the `scala.Product` trait: ```scala trait Product { /** Returns the field name of element at index n */ def productElementName(n: Int): String /** Returns field names of this product. Must have same length as productIterator */ def productElementNames: Iterator[String] } ``` Both methods have a default implementation which returns the empty string for all field names. This commit then changes the code-generation for case classes to synthesize a `productElementName` method with actual class field names. The benefit of this change is that it becomes possible to pretty-print case classes with field names, for example ```scala case class User(name: String, age: Int) def toPrettyString(p: Product): String = p.productElementNames.zip(p.productIterator) .map { case (name, value) => s"$name=$value" } .mkString(p.productPrefix + "(", ", ", ")") toPrettyString(User("Susan", 42)) // res0: String = User(name=Susan, age=42) ``` The downside of this change is that it produces more bytecode for each case-class definition. Running `:javacp -c` for a case class with three fields yields the following results ```scala > case class A(a: Int, b: Int, c: Int) > :javap -c A public java.lang.String productElementName(int); Code: 0: iload_1 1: istore_2 2: iload_2 3: tableswitch { // 0 to 2 0: 28 1: 33 2: 38 default: 43 } 28: ldc 78 // String a 30: goto 58 33: ldc 79 // String b 35: goto 58 38: ldc 80 // String c 40: goto 58 43: new 67 // class java/lang/IndexOutOfBoundsException 46: dup 47: iload_1 48: invokestatic 65 // Method scala/runtime/BoxesRunTime.boxToInteger:(I)Ljava/lang/Integer; 51: invokevirtual 70 // Method java/lang/Object.toString:()Ljava/lang/String; 54: invokespecial 73 // Method java/lang/IndexOutOfBoundsException."<init>":(Ljava/lang/String;)V 57: athrow 58: areturn ``` Thanks to Adriaan's help, the estimated cost per `productElementName` appears to be fixed 56 bytes and then 10 bytes for each field with the following breakdown: * 3 bytes for the [string info](https://docs.oracle.com/javase/specs/jvms/se8/html/jvms-4.html#jvms-4.4.3) (the actual characters are already in the constant pool) * 4 bytes for the tableswitch entry * 2 bytes for the ldc to load the string * 1 byte for areturn In my opinion, the bytecode cost is acceptably low thanks to the fact that field name literals are already available in the constant pool.

olafurpg · 2018-07-18T11:32:23Z

This is an alternative solution for #6936 The difference is that this PR

does not change the behavior of toString to minimize the breaking change
makes it possible to implement custom pretty-printing formats for case classes by exposing productElementName to library developers

olafurpg · 2018-07-18T11:37:55Z

I haven't validated yet the exact bytecode overhead from this change. A good solution would be to compile larger projects and see the total added bytes to the resulting packaged jar.

I opened this PR too fast as I just realized I unfortunately won't have time to complete this PR as I'm already quite occupied with getting a stable Scalafix release out. Feel free to pick this up if you are interested. No attribution needed, the diff is mostly just uncommenting existing code anyways.

julienrf · 2018-07-18T13:39:28Z

What was the reason for removing productElementName?

olafurpg · 2018-07-18T14:57:38Z

I don't know if productElementName ever made it into an official Scala release, but there was a comment saying it should be possible to accomplish the same functionality by inspecting the classfiles instead (I assume via runtime reflection, and it was removed to minimize generated bytecode).

SethTisue · 2018-07-18T16:01:06Z

In my opinion, the bytecode cost is acceptably low

agree

retronym · 2018-07-18T22:58:40Z

We could ratchet the fixed cost down a bit by adding a static factory method to scala.runtime for the IIOBE. Java 9 introduced a constructor overload that accepts an int, sadly we can't use that yet.

We should test the interaction with case classes that define or inherit a non-default implementation of productElementName.

retronym · 2018-07-18T23:01:24Z

src/library/scala/Product.scala

@@ -48,4 +48,15 @@ trait Product extends Any with Equals {
   *  @return   in the default implementation, the empty string
   */
  def productPrefix = ""
+
+  def productElementName(n: Int): String =
+    if (n >= 0 && n  < n) ""


n < productArity?

Yes, good catch.

retronym · 2018-07-18T23:05:54Z

test/files/run/productElementName.scala

+     .map { case (name, value) => s"$name=$value" }
+     .mkString(p.productPrefix + "(", ", ", ")")
+  println(pretty(User("Susan", 42)))
+}


We should also test with:

non alphanumeric names (is the name.decoded done before code gen?)

case classes with non-public elements (just in case there is any problems in the current implementation related to these comments)

case classes with secondary parameter lists or with secondary constructors (again, to make sure we don't have latent bugs in product* code gen)

olafurpg

Thank you for the review @retronym. I won't be able to address the comments (at least not until around September) but maybe someone from the contributors thread https://contributors.scala-lang.org/t/case-class-tostring-new-behavior-proposal-with-implementation/2056/40?u=olafurpg is motivated to pick this up.

olafurpg · 2018-07-19T05:33:50Z

src/library/scala/Product.scala

@@ -48,4 +48,15 @@ trait Product extends Any with Equals {
   *  @return   in the default implementation, the empty string
   */
  def productPrefix = ""
+
+  def productElementName(n: Int): String =
+    if (n >= 0 && n  < n) ""


Yes, good catch.

olafurpg · 2018-07-19T05:34:30Z

test/files/run/productElementName.scala

+     .map { case (name, value) => s"$name=$value" }
+     .mkString(p.productPrefix + "(", ", ", ")")
+  println(pretty(User("Susan", 42)))
+}


cb372 · 2018-07-25T07:44:54Z

I’d like to pick this up if nobody else has already. I’ll look at adding some of the tests suggested above.

fommil · 2018-07-29T09:11:41Z

FYI scalaz-deriving's instance of Show is able to do this without touching the case class, supporting older versions of scala.

import scalaz.{deriving,Show}

@deriving(Show)
case class Foo(s: String)

etc, for arbitrary arity.

scala-jenkins added this to the 2.13.0-M5 milestone Jul 18, 2018

olafurpg closed this Jul 18, 2018

retronym reviewed Jul 18, 2018

View reviewed changes

olafurpg commented Jul 19, 2018

View reviewed changes

cb372 mentioned this pull request Jul 25, 2018

Add productElementName and productElementNames methods to Product #6972

Merged

SethTisue removed this from the 2.13.0-M5 milestone Aug 22, 2018

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Revive productElementName to extract case class field names #6951

Revive productElementName to extract case class field names #6951

olafurpg commented Jul 18, 2018

olafurpg commented Jul 18, 2018

olafurpg commented Jul 18, 2018

julienrf commented Jul 18, 2018

olafurpg commented Jul 18, 2018

SethTisue commented Jul 18, 2018

retronym commented Jul 18, 2018 •

edited

Loading

retronym Jul 18, 2018

olafurpg Jul 19, 2018

retronym Jul 18, 2018

olafurpg Jul 19, 2018

olafurpg left a comment

olafurpg Jul 19, 2018

olafurpg Jul 19, 2018

cb372 commented Jul 25, 2018

fommil commented Jul 29, 2018

Revive productElementName to extract case class field names #6951

Revive productElementName to extract case class field names #6951

Conversation

olafurpg commented Jul 18, 2018

olafurpg commented Jul 18, 2018

olafurpg commented Jul 18, 2018

julienrf commented Jul 18, 2018

olafurpg commented Jul 18, 2018

SethTisue commented Jul 18, 2018

retronym commented Jul 18, 2018 • edited Loading

retronym Jul 18, 2018

Choose a reason for hiding this comment

olafurpg Jul 19, 2018

Choose a reason for hiding this comment

retronym Jul 18, 2018

Choose a reason for hiding this comment

olafurpg Jul 19, 2018

Choose a reason for hiding this comment

olafurpg left a comment

Choose a reason for hiding this comment

olafurpg Jul 19, 2018

Choose a reason for hiding this comment

olafurpg Jul 19, 2018

Choose a reason for hiding this comment

cb372 commented Jul 25, 2018

fommil commented Jul 29, 2018

retronym commented Jul 18, 2018 •

edited

Loading