Infoscience

Conference paper

Bridging Islands of Specialized Code using Macros and Reified Types

Parametric polymorphism in Scala suffers from the usual drawback of erasure on the Java Virtual Machine: primitive values are boxed, leading to indirect access, wasteful use of heap memory and lack of cache locality. For performance-critical parts of the code, the Scala compiler introduces specialization, a transformation that duplicates and adapts the bodies of classes and methods for primitive types. Specializing code can speed up execution by an order of magnitude, but only if the code is called from monomorphic sites or from other specialized code. Still, if these “islands” of specialized code are called from generic code, their performance becomes similar to that of generic code, losing optimality. To address this, our project builds high performance “bridges” between “islands” of specialized code, removing the requirement that full traces need to be specialized: We use macros to delimit performance-critical “gaps” between specialized code, which we also specialize. We then use reified types to dispatch the correct specialized variant, thus recovering performance across the “islands”. Our transformation obtains speedups up to 30x and around 12x in average compared to generic only code, by enabling specialization to completely remove boxing and reach its full potential

Related material