denisrosset / metal   0.16.0.0

GitHub

Metal - fast unboxed data structures for Scala

Scala versions: 2.12

Metal - fast unboxed data structures for Scala

Gitter

See the companion website and the associated tutorials Metal provides fast mutable collections whose performance should be close to hand-written data structures using raw primitive arrays.

In particular:

  • Metal collections are themselves not specialized, but will avoid any boxing/allocation when accessing, storing and updating elements, thanks to macros and specialized methods;
  • Metal provides higher-order methods such as foreach, count, exists, ... that are translated into while loops during compilation. The loop body is inlined, avoiding allocation of closures;
  • Scala iterators are replaced by pointers, represented by value classes that need no allocation;
  • mutable containers can be used as builders for immutable containers.

The library is heavily inspired by Debox. Parts of the implementation are similar (for example the strategy for hash sets and maps).

For performance reasons, Metal's types are neither compatible with Scala's collections framework, nor the Debox implementations; but the methods in Metal's interfaces are either prefixed by ptr, or follow currently used conventions (e.g. def isEmpty: Boolean).

Higher-order methods such as foreach are provided by enrichment methods.

The set of methods available on Metal instances is limited, but guarantees that no allocations occur except when creating or growing containers.

We expose most of the collections internal state and do not try to have proper encapsulation. Metal is intended to be used in performance critical hot spots; it is not a replacement for the standard library collections.

Structure of Metal

Metal is composed of two packages:

  • core contains the definition of pointer types (see below), traits that are inherited by containers to provide capabilities, and macro implementations of the standard collection methods;
  • library provides implementations of standard data structures, with generic providing base types such as Set, Map, Map2, Buffer, mutable variants in mutable with corresponding immutable variants in immutable.

We currently have a dependency on Spire for two reasons:

  • availability of a non-boxing option type spire.util.Opt,
  • compatibility macro shims for Scala 2.10 and 2.11.

Example

scala> import metal._; import syntax._
import metal._
import syntax._

scala> val set = mutable.HashSet(1,2,3,4)
set: metal.mutable.HashSet.S[Int] = Set(1, 2, 3, 4)

scala> set.contains(3)
res0: Boolean = true

scala> set.foreach { k => println(k) }
1
2
3
4


scala> set.result()
res2: metal.immutable.HashSet[Int] = Set(1, 2, 3, 4)

scala> set
res3: metal.mutable.HashSet.S[Int] = Set()

scala> val map = mutable.HashMap(1 -> "test")
map: metal.mutable.HashMap.M[Int,String] = Map(1 -> test)

scala> map.getOrElse(1, sys.error(""))
res4: String = test