Miniboxing is a Scala compiler transformation that improves the performance of generics for primitive numeric types. Unlike specialization, miniboxing reduces the amount of generated bytecode without sacrificing execution speed, therefore being capable of transforming the Scala collections hierachy.
##Why use Miniboxing?##
Short answer: because it matches the performance of specialization, without the bytecode blowup. For the
case class Tuple3[@specialized +T1, @specialized +T2, @specialized +T3](_1: T1, _2: T2, _3: T3)
Specialization generates 1000 classes. Just change
@miniboxed and you get only 8 classes.
Long answer: Aside from reducing the bytecode size, the miniboxing technique improves several other aspects of specialization:
To see the benchmarks we performed, have a look at the OOPSLA 2013 paper we just prepared. They include performance evaluations, bytecode size comparisons and many more. Also, the docs directory contains a series of papers and presentations which explain many aspects of the miniboxing transformation.
At this point, the miniboxing plugin is not production-ready, although it can compile spire. However, we do publish a nightly maven artifact on Sonatype, so anyone can try the current transformation. To get started, have a look at the example project we prepared.
Mind the gap: there are still many bugs and known limitations, so you're in for a thrill! Also, don't hesitate to add bugs to the tracker, good reductions that can be easily reproduced are highly appreciated!
The development branches are:
components/plugin- the actual Scala compiler plugin
components/runtime- the runtime support for the transformed code
components/classloader- the classloader used for runtime class specialization
tests/benchmarks- the benchmarks for the project
tests/correctness- the tests for the plugin transformation
docs- documents released as the development goes on
##Questions?## If you have any question, you can contact me at vlad dot ureche at epfl dot ch.