Dynamic languages

The Common Language Infrastructure (CLI) is a virtual machine expressly designed for implementing statically typed languages such as C#, therefore programs written in dynamically typed languages are typically much slower than C# when executed on .NET. Recent developments show that Just In Time (JIT) compilers can exploit runtime type information to generate quite efficient code. Unfortunately, writing a JIT compiler is far from being simple. In this paper we report our positive experience with automatic generation of JIT compilers as supported by the PyPy infrastructure, by focusing on JIT compilation for .NET. Following this approach, we have in fact added a second layer of JIT compilation, by allowing dynamic generation of more efficient .NET bytecode, which in turn can be compiled to machine code by the .NET JIT compiler. The main and novel contribution of this paper is to show that this two-layers JIT technique is effective, since programs written in dynamic languages can run on .NET as fast as (and in some cases even faster than) the equivalent C# programs. The practicality of the approach is demonstrated by showing some promising experiments done with benchmarks written in a simple dynamic language.

Writing an optimizing static compiler for dynamic languages is not an easy task, since quite complex static analysis is required. On the other hand, recent developments show that JIT compilers can exploit runtime type information to generate quite efficient code. Unfortunately, writing a JIT compiler is far from being simple. In this paper we report our positive experience with automatic generation of JIT compilers as supported by the PyPy infrastructure, by focusing on JIT compilation for .NET. The paper presents two main and novel contributions: we show that partial evaluation can be used in practice for generating a JIT compiler, and we experiment with the potentiality offered by the ability to add a further level of JIT compilation on top of .NET. The practicality of the approach is demonstrated by showing some promising experiments done with benchmarks written in a simple dynamic language.

Although the C-based interpreter of Python is reasonably fast, implementations on the CLI or the JVM platforms offers some advantages in terms of robustness and interoperability. Unfortunately, because the CLI and JVM are primarily designed to execute statically typed, object-oriented languages, most dynamic language implementations cannot use the native bytecodes for common operations like method calls and exception handling; as a result, they are not able to take full advantage of the power offered by the CLI and JVM. We describe a different approach that attempts to preserve the flexibility of Python, while still allowing for efficient execution. This is achieved by limiting the use of the more dynamic features of Python to an initial, bootstrapping phase. This phase is used to construct a final RPython (Restricted Python) program that is actually executed. RPython is a proper subset of Python, is statically typed, and does not allow dynamic modification of class or method definitions; however, it can still take advantage of Python features such as mixins and first-class methods and classes. This paper presents an overview of RPython, including its design and its translation to both CLI and JVM bytecode. We show how the bootstrapping phase can be used to implement advanced features, like extensible classes and generative programming. We also discuss what work remains before RPython is truly ready for general use, and compare the performance of RPython with that of other approaches.