README.md

scala-fusion

GHC-inspired stream fusion for scala

This repository implements techniques found in "Stream Fusion. From Lists to Streams to Nothing at All", by Duncan Coutts, Roman Leshchinskiy, and Don Stewart (ICFP '07).

Usage

import fusion.syntax._

LazyList(0,1,2,3,4)
  .fuse
  .map { _ + 1 }
  .filter { _ <= 2 }
  .take(1)
  .toList

// output:  List(1)
// runtime: LazyList(<not computed>) -> Stream(<instruction set>, LazyList(<not computed>)) -> List(1)

Comparison

Although the usage is similar to SeqView in Scala 2.13+ and the resulting optimizations are similar, the implementation itself is very different. SeqView is implemented imperatively, while scala-fusion is written functionally but generates an efficient imperative instruction set at compile time.

import fusion.syntax._

// no fusion
List(0,1,2,3,4)
  .map { _ + 1 }
  .filter { _ <= 2 }
  .take(1)

// output:  List(2)
// runtime: List(0,1,2,3,4) -> List(1,2,3,4,5) -> List(1,2) -> List(1)

// SeqView
List(0,1,2,3,4)
  .view
  .map { _ + 1 }
  .filter { _ <= 2 }
  .take(1)
  .toList

// output:  List(1)
// runtime: List(0,1,2,3,4) -> SeqOps(<instruction set>) -> List(1)

// scala-fusion (this library)
LazyList(0,1,2,3,4)
  .fuse
  .map { _ + 1 }
  .filter { _ <= 2 }
  .take(1)
  .toList

// output:  List(1)
// runtime: LazyList(<not computed>) -> Stream(<instruction set>, LazyList(<not computed>)) -> List(1)

Implementation

The original version relies on GHC rewrite rules to reduce stream compose unstream to the identity on streams. Since this is unavailable in scalac, The Fuser type was introduced which wraps the underlying stream and contains an instance of a type-threaded list which stores the sequence of procedures. Composing this type-threaded list results in a single set of fused instructions to run on the original input Stream. This implementation adds a very small amount of overhead proportional to the number of operations called on the stream.