mkolod / inliner

scala macros to inline idiomatic scala for maximum performance

Inliner

Inliner is a collection of scala macros to inline and optimize idiomatic scala into while loops or nested if/else statements. The purpose is to allow idiomatic scala without having to give up performance.

How to use?

Add val inliner = ProjectRef(uri("git://github.com/johnynek/inliner.git"), "core") to your sbt project, then add .dependsOn(inliner) to any project where you want to use the code. Then do:

import com.github.johnynek.inliner.{InlineTry, InlineOption, InlineCollection}
import InlineTry._
import InlineOption._
import InlineCollection._

Generally you import methods from an object and replace calls like x.method with x.inline.method

Collections

There are .inline versions of the following TraversableOnce methods: find, forall, foldLeft, foreach, reduceOption.

import com.github.johnynek.inliner.InlineCollection._
object Test {
  println(List(0, 1, 2, 4, 8, 16).inline.foldLeft(0)(_ + _))
}

Expands the foldLeft into:

  val it$macro$1 = immutable.this.List.apply[Int](0, 1, 2, 4, 8, 16).toIterator;
  var x$1 = 0;
  while$1(){
    if (it$macro$1.hasNext)
      {
        {
          val x$2 = it$macro$1.next;
          x$1 = x$1.+(x$2)
        };
        while$1()
      }
    else
      ()
  };
  x$1

scala.util.Try

Normally, creating a Try means a call-by-name parameter, which requires an allocation and a method call. With a macro, we can directly inline into a try/catch block:

import com.github.johnynek.inliner.InlineTry._
def halfEven(x: Int): Int = { require(x % 2 == 0, "not even: " + x); x/2 }
inlineTry {
  val x = halfEven(42)
  val y = halfEven(43)
  x * y
}

which, at the REPL, expands to:

try {
  new _root_.scala.util.Success[Int]({
    val x = $line4.$read.$iw.$iw.$iw.$iw.halfEven(42);
    val y = $line4.$read.$iw.$iw.$iw.$iw.halfEven(43);
    x.*(y)
  })
} catch {
  case _root_.scala.util.control.NonFatal((e @ _)) => new scala.util.Failure[Nothing](e)
}

Similarly, for expressions can be expanded into nested if/else:

 for {
   a <- inlineTry { assert(o1 > 0); o1 }.inline
   b <- o2(a).inline
   c <- o3(b).inline
 } yield c

expands to:

{
  val opt$macro$13 = (try {
    new scala.util.Success[Int]({
      scala.this.Predef.assert(o1.>(0));
      o1
    })
  } catch {
    case _root_.scala.util.control.NonFatal((e @ _)) => new scala.util.Failure[Nothing](e)
  }: scala.util.Try[Int]);
  if (opt$macro$13.isSuccess)
    {
      val a = opt$macro$13.get;
      try {
        ({
          val opt$macro$12 = o2.apply(a);
          if (opt$macro$12.isSuccess)
            {
              val b = opt$macro$12.get;
              try {
                (o3.apply(b): scala.util.Try[Long])
              } catch {
                case _root_.scala.util.control.NonFatal((e @ _)) => new scala.util.Failure[Nothing](e)
              }
            }
          else
            opt$macro$12.asInstanceOf[scala.util.Try[Long]]
        }: scala.util.Try[Long])
      } catch {
        case _root_.scala.util.control.NonFatal((e @ _)) => new scala.util.Failure[Nothing](e)
      }
    }
  else
    opt$macro$13.asInstanceOf[_root_.scala.util.Try[Long]]
}

This gives you inlined versions of: filter, flatMap, flatten, fold, foreach, getOrElse, map, orElse.

Option

Similar to Try, you can inline for loops using .inline on methods filter, flatMap, flatten, fold, foreach, getOrElse, map, orElse.

When should I use Inliner?

Macros are not as reliable as they could be. You should probably only use this library for inner loops that have been profiled. Optimizing without profiling is usually not profitable. Once you find a method that needs maximum optimization, Inliner may allow you to keep idiomatic code with minor modifications to get maximum performance.

Future Work

Check the issues, but generally support for more classes (such as Either) or constructs (such as PartialFunction literals) would be useful. Also, optimizing some of the trees would be really interesting. Once we have a full tree we can see that some of the branches will never be taken in large for-expressions. Also, we could port this approach to a whitebox macro such as def inline(x: Any): Any which could do whole expression optimization along the lines we have here without manually calling .inline. This would have the benefit of being able to optimize things like:

myList
  .map { x => (x, 1) }
  .reduceOption { case (la, lb), (ra, rb) => (la + ra, lb + lb) }

to an expression like:

val it = myList.iterator
if (it.hasNext) {
  val head = it.next
  var result1 = head
  var result2 = 1
  while(it.hasNext) {
    val item = it.next
    val item1 = item
    val item2 = 1
    result1 = result1 + item1
    result2 = result2 + item2
  }
  Some((result1, result2))
} else None

Authors

Best to check the commit history, but this was started by Oscar Boykin.