Announcing Dotty 0.2.0-RC1, with new optimizations, improved stability and IDE support

Wednesday 12 July 2017

Dmitry Petrashko

Today, we are excited to release Dotty version 0.2.0-RC1. This release serves as a technology preview that demonstrates new language features and the compiler supporting them.

This release is based on the previous milestone. The highlights of this release are:

substantial improvement of quality of generated code for pattern matching
improvements in VS Code IDE stability
support Windows in VS Code IDE
improved compatibility with scalac
initial support for reproducible builds

This is our second scheduled release according to our 6-week release schedule.

What’s in the 0.2.0-RC1 technology preview?

The previous technology preview has shipped new language features planned for Scala 3:
Intersection Types, Union Types, Trait Parameters, Enumerations, Algebraic Data Types, By-Name Implicits.

This technology preview is geared towards improving stability and reliability. It includes:

Local optimizations upstreamed from the Dotty Linker, 2647 by (@OlivierBlanvillain). See more details below.
Optimizing Pattern Matcher by (@odersky)
Idempotency checks are the first step to reproducible builds
Faster Base class sets by (@odersky) and (@darkdimius)
Numerous fixes to IDE and Dotty Language Server covering:
Scalac compatibility improvements:
Java compatibility improvements:
- Fixes to handing of Java generic signatures
- java.lang.System.out is final but that’s a lie
Improved error messages:
Improved command line handling:
- Support params in a file like @file.txt
Type system stability:
- Handle wildcard types in unions and intersections
Fixes to implicit search:
- Fix shadowing of higher order implicits

Better generated code:

As was spotted by @gkossakowski in the previous release Dotty was on par with Scala 2.11 in speed. But why is that? The reason is that Dotty compiled by Dotty had really horrible code generated for pattern matching.

Let’s illustrate on a simple example:

case class CC(a: Int, b: Object)

  def foo(x: Any): Int = {
    val (a, b) = x match {
      case CC(s @ 1, CC(t, _)) =>
        (s , 2)
      case _ => (42, 43)
    }
    a + b
  }
  
  def booleans(a: Object) = {
    val (b1, b2) = (a.isInstanceOf[CC], a.isInstanceOf[List[Int]])
    (b1, b2) match {
      case (true, true) => true
      case (false, false) => true
      case _ => false
    }
  }

The Dotty that was released in the previous milestone didn’t contain any optimizations and generated inefficient code for it. The java-with-goto code below is equivalent to what Dotty generated.

// output of dotc 0.1.2-RC1
    public int foo(Object x) {
        var3_2 = x;
        if (!(var3_2 instanceof CC)) ** GOTO lbl-1000
        var4_3 = (CC)var3_2;
        if (CC$.MODULE$.unapply((CC)var3_2) == null) ** GOTO lbl-1000
        var5_4 = CC$.MODULE$.unapply((CC)var3_2);
        s = var5_4._1();
        var7_6 = var5_4._2();
        if (1 != s) ** GOTO lbl-1000
        var8_7 = s;
        if (!(var7_6 instanceof CC)) ** GOTO lbl-1000
        var9_8 = (CC)var7_6;
        if (CC$.MODULE$.unapply((CC)var7_6) != null) {
            var10_9 = CC$.MODULE$.unapply((CC)var7_6);
            var11_10 = var10_9._2();
            v0 = Tuple2..MODULE$.apply((Object)BoxesRunTime.boxToInteger((int)1), (Object)BoxesRunTime.boxToInteger((int)2));
        } else lbl-1000: // 5 sources:
        {
            v0 = Tuple2..MODULE$.apply((Object)BoxesRunTime.boxToInteger((int)42), (Object)BoxesRunTime.boxToInteger((int)43));
        }
        var2_11 = v0;
        a = BoxesRunTime.unboxToInt((Object)var2_11._1());
        b = BoxesRunTime.unboxToInt((Object)var2_11._2());
        return a + b;
    }

    public boolean booleans(Object a) {
        Tuple2 tuple2 = Tuple2..MODULE$.apply((Object)BoxesRunTime.boxToBoolean((boolean)(a instanceof CC)), (Object)BoxesRunTime.boxToBoolean((boolean)(a instanceof List)));
        boolean b1 = BoxesRunTime.unboxToBoolean((Object)tuple2._1());
        boolean b2 = BoxesRunTime.unboxToBoolean((Object)tuple2._2());
        Tuple2 tuple22 = Tuple2..MODULE$.apply((Object)BoxesRunTime.boxToBoolean((boolean)b1), (Object)BoxesRunTime.boxToBoolean((boolean)b2));
        Option option = Tuple2..MODULE$.unapply(tuple22);
        if (!option.isEmpty()) {
            Tuple2 tuple23 = (Tuple2)option.get();
            boolean bl = BoxesRunTime.unboxToBoolean((Object)tuple23._1());
            boolean bl2 = BoxesRunTime.unboxToBoolean((Object)tuple23._2());
            if (bl) {
                boolean bl3 = bl;
                if (bl2) {
                    boolean bl4 = bl2;
                    return true;
                }
            }
        }
        Option option2 = Tuple2..MODULE$.unapply(tuple22);
        if (option2.isEmpty()) return false;
        Tuple2 tuple24 = (Tuple2)option2.get();
        boolean bl = BoxesRunTime.unboxToBoolean((Object)tuple24._1());
        boolean bl5 = BoxesRunTime.unboxToBoolean((Object)tuple24._2());
        if (bl) return false;
        boolean bl6 = bl;
        if (bl5) return false;
        boolean bl7 = bl5;
        return true;
    }

Due to the new optimizing pattern matcher, Dotty now is able to generate the code below without -optimise

// output of 0.2.0-RC1 without -optimise
    public int foo(Object x) {
        var3_2 = x;
        if (!(var3_2 instanceof CC)) ** GOTO lbl-1000
        var4_3 = CC$.MODULE$.unapply((CC)var3_2);
        s = var5_4 = var4_3._1();
        if (1 == var5_4 && (var7_6 = var4_3._2()) instanceof CC) {
            t = CC$.MODULE$.unapply((CC)var7_6)._1();
            v0 = Tuple2..MODULE$.apply((Object)BoxesRunTime.boxToInteger((int)1), (Object)BoxesRunTime.boxToInteger((int)2));
        } else lbl-1000: // 2 sources:
        {
            v0 = Tuple2..MODULE$.apply((Object)BoxesRunTime.boxToInteger((int)42), (Object)BoxesRunTime.boxToInteger((int)43));
        }
        var2_8 = v0;
        a = BoxesRunTime.unboxToInt((Object)var2_8._1());
        b = BoxesRunTime.unboxToInt((Object)var2_8._2());
        return a + b;
    }

    public boolean booleans(Object a) {
        Tuple2 tuple2 = Tuple2..MODULE$.apply((Object)BoxesRunTime.boxToBoolean((boolean)(a instanceof CC)), (Object)BoxesRunTime.boxToBoolean((boolean)(a instanceof List)));
        boolean b1 = BoxesRunTime.unboxToBoolean((Object)tuple2._1());
        boolean b2 = BoxesRunTime.unboxToBoolean((Object)tuple2._2());
        Tuple2 tuple22 = Tuple2..MODULE$.apply((Object)BoxesRunTime.boxToBoolean((boolean)b1), (Object)BoxesRunTime.boxToBoolean((boolean)b2));
        if (tuple22 != null) {
            boolean bl;
            boolean bl2 = BoxesRunTime.unboxToBoolean((Object)tuple22._1());
            if (!bl2) {
                bl = bl2;
            } else {
                if (BoxesRunTime.unboxToBoolean((Object)tuple22._2())) {
                    return true;
                }
                bl = bl2;
            }
            if (!bl) {
                if (false != BoxesRunTime.unboxToBoolean((Object)tuple22._2())) return false;
                return true;
            }
        }
        return false;
    }

You can clearly see that it’s shorter ;-) and it actually does less work. If you additionally enable local optimizations, you get decent generated code:

// output of 0.2.0-RC1 with -optimise

    public int foo(Object x) {
        int n;
        Tuple2 tuple2;
        CC cC;
        Object object;
        if (x instanceof CC && 1 == (n = (cC = (CC)x)._1()) && (object = cC._2()) instanceof CC) {
            ((CC)object)._1();
            tuple2 = new Tuple2((Object)BoxesRunTime.boxToInteger((int)1), (Object)BoxesRunTime.boxToInteger((int)2));
        } else {
            tuple2 = new Tuple2((Object)BoxesRunTime.boxToInteger((int)42), (Object)BoxesRunTime.boxToInteger((int)43));
        }
        Tuple2 tuple22 = tuple2;
        return BoxesRunTime.unboxToInt((Object)tuple22._1()) + BoxesRunTime.unboxToInt((Object)tuple22._2());
    }

    public boolean booleans(Object a) {
        boolean bl = a instanceof CC;
        boolean bl2 = a instanceof List;
        new Tuple2((Object)BoxesRunTime.boxToBoolean((boolean)bl), (Object)BoxesRunTime.boxToBoolean((boolean)bl2));
        new Tuple2((Object)BoxesRunTime.boxToBoolean((boolean)bl), (Object)BoxesRunTime.boxToBoolean((boolean)bl2));
        if (bl && bl2) {
            return true;
        }
        boolean bl3 = bl;
        if (bl3) return false;
        if (bl2) return false;
        return true;
    }

This code still has a major inefficiency; it allocates tuples. We plan to continue the migration of local optimizations from the Dotty Linker that should allow us to generate code that is as good the code generated by the Dotty Linker with global analysis disabled:

  // output of Dotty linker https://github.com/dotty-linker/dotty/tree/opto
    public int foo(Object x) {
         CC cC;
         int n = 0;
         int n2 = 0;
         if (x instanceof CC && 1 == (cC = (CC)x)._1() && cC._2() instanceof CC) {
             n = 1;
             n2 = 2;
         } else {
             n = 42;
             n2 = 43;
         }
         return n + n2;
     }
 
     public boolean booleans(Object a) {
         boolean bl = a instanceof CC;
         boolean bl2 = a instanceof List;
         if (bl && bl2 || !bl && !bl2) {
             return true;
         }
         return false;
 }

How can you try it out?

We ship with tools that help you try out the Dotty platform:

IDE features for Visual Studio Code
sbt support, including retro-compatibility with Scala 2

You have several alternatives; use the sbt-dotty plugin, get a standalone installation, or try it online on Scastie.

sbt

Using sbt 0.13.13 or newer, do:

sbt new lampepfl/dotty.g8

This will setup a new sbt project with Dotty as compiler. For more details on using Dotty with sbt, see the example project.

Standalone installation

Releases are available for download on the Releases section of the Dotty repository: https://github.com/scala/scala3/releases

We also provide a homebrew package that can be installed by running:

brew install lampepfl/brew/dotty

In case you have already installed Dotty via brew, you should instead update it:

brew upgrade dotty

Scastie

Scastie, the online Scala playground, supports Dotty. You can try it out there without installing anything.

What are the next steps?

Over the coming weeks and months, we plan to work on the following topics:

Add support for using Dotty generated classes with Scala 2.12
Add Language-level support for HMaps and HLists;
Upstream more optimizations from Dotty Linker
Add support for existing in the same classpath with Scala 2.12

If you want to get your hands dirty with any of this, now is a good moment to get involved! Join the team of contributors, including Martin Odersky (@odersky) Dmitry Petrashko (@DarkDimius), Guillaume Martres (@smarter), Felix Mulder (@felixmulder), Nicolas Stucki (@nicolasstucki), Liu Fengyun (@liufengyun), Olivier Blanvillain (@OlivierBlanvillain), and others!

Library authors: Join our community build

Dotty now has a set of libraries that are built against every nightly snapshot. Currently this includes scalatest, squants and algebra. Join our community build to make sure that our regression suite includes your library.

To get started, see https://github.com/scala/scala3.