Compiling generics through user-directed type specialization

Compilation of polymorphic code through type erasure gives compact code but performance on primitive types is significantly hurt. Full specialization gives good performance, but at the cost of increased code size and compilation time. Instead we propose a mixed approach, which allows the programmer to decide what code to specialize. Our approach supports separate compilation, allows mixing of specialized and generic code, and gives very good results in practice.

Compiling Scala for Performance

Iulian Dragos

Scala is a new programming language bringing together object-oriented and functional programming. Its defining features are uniformity and extensibility. Scala offers great flexibility for programmers, allowing them to grow the language through libraries. Oftentimes what seems like a language feature is in fact implemented in a library, effectively giving programmers the power of language designers. The downside of this flexibility is that familiar looking code may hide unexpected performance costs. It is important for Scala compilers to bring down this cost as much as possible. We identify several areas of impact for Scala performance: higher-order functions and closures, and generic containers used with primitive types. We present two complementary approaches for improving performance in these areas: optimizations and specialization. Compiler optimization can bring down the cost through a combination of aggressive inlining of higher-order functions, an extended version of copy-propagation and dead-code elimination. Both anonymous functions and boxing can be eliminated by this approach. We show on a number of benchmarks that these language features can be up to 5 times faster when properly optimized, on current day JVMs. We propose a new approach to compiling parametric polymorphism for performance at primitive types. We mix a homogeneous translation scheme with user-directed specialization for primitive types. Type parameters may be annotated to require specialization of code depending on them. We propose definition-site specialization for primitive types, achieving separate compilation and no boxing when both the definition and call site are specialized. Specialized classes are compatible with unspecialized code, and specialization agnostic code can work with specialized instances, meaning that specialization is opportunistic. We present a formalism of a small subset of Scala with specialization and prove that specialization preserves types. We implemented this translation in the Scala compiler and report on improvements on a set of benchmarks, showing that specialization can make programs more than two times faster.

EPFL2010

Improving the Interoperation between Generics Translations

Romain Michel Beguet, Martin Odersky, Milos Stojanovic, Nicolas Alexander Stucki, Vlad Ureche

Generics on the Java platform are compiled using the erasure transformation, which only supports by-reference values. This causes slowdowns when generics operate on primitive types, such as integers, as they have to be transformed into reference-based objects. Project Valhalla is an effort to remedy this problem by specializing classes at load-time so they can efficiently handle primitive values. In its current early prototype, the Valhalla compilation scheme limits the interaction between specialized and erased generics, thus preventing certain useful code patterns from being expressed. Scala has been using compile-time specialization for 6 years and has three generics compilation schemes working side by side. In Scala, programmers are allowed to write code that freely exercises the interaction between the different compilation schemes, at the expense of introducing subtle performance issues. Similar performance issues can affect Valhalla-enabled bytecode, whether the code was written in Java or translated from other JVM languages. In this context we explain how we help programmers avoid these performance regressions in the miniboxing transformation: (1) by issuing actionable performance advisories that steer programmers away from performance regressions and (2) by providing alternatives to the standard library constructs that use the miniboxing encoding, thus avoiding the conversion overhead.

ACM2015

Compile-Time Type-Driven Data Representation Transformations in Object-Oriented Languages

Vlad Ureche

High-level languages allow programmers to express data structures and algorithms that abstract over the type of data they handle. This improves code reuse and makes it possible to develop general-purpose libraries. Yet, data abstractions slow down program execution, as they require low-level indirection. In this thesis we explore three compile-time approaches that leverage type systems to reduce the cost of data abstractions, thus improving program performance. In the first part of the thesis we present miniboxing, a compile-time transformation that replaces generic classes by more efficient variants, optimized to handle primitive types. These variants use the miniboxed data encoding, producing speedups of up to 20x compared to generic classes. The miniboxing transformation is the main result of this thesis and motivates the other techniques. Generalizing miniboxing, we show the Late Data Layout (LDL) mechanism, which uses the type system to guide performance-oriented program rewritings. It can be instantiated to perform a host of transformations, such as miniboxing generics, inlining value classes and unboxing primitive types. The LDL mechanism has many desirable properties, such as provable correctness in handling different data representations, reduced number of conversions and built-in support for the object-oriented paradigm. Finally, we show Data-centric Metaprogramming, a technique that allows programmers to go beyond standard compiler optimizations by defining custom representations to be used for their data. These representations are then automatically introduced by the compiler when translating programs. This technique, similar in spirit to metaprogramming, opens new directions in programmer-driven optimizations and shows encouraging results, with speedups of up to 25x. Under the hood, Data-centric Metaprogramming uses the Late Data Layout mechanism.