Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dr Wenowdis: Specializing dynamic language C extensions using type information

Maxwell Bernstein, CF Bolz-Tereick

TL;DR

The paper addresses the overhead incurred when dynamic languages call into CPython C extensions, which is opaque to static analysis and hinders JIT optimizations. It proposes Dr Wenowdis, a lightweight, backward-compatible mechanism to attach type and effect metadata to C extension functions, enabling fast call paths in PyPy by reducing boxing/unboxing and argument checks. Micro-benchmark results show substantial speedups in Python-to-C calls, including dramatic improvements on hot paths and even gains with the JIT disabled, suggesting the technique can improve performance without invasive rewrites. The work has implications for static analysis and could be generalized to other runtimes and binding tools, potentially enabling automatic generation of such annotations by existing binding generators like Cython or PyO3.

Abstract

C-based interpreters such as CPython make extensive use of C "extension" code, which is opaque to static analysis tools and faster runtimes with JIT compilers, such as PyPy. Not only are the extensions opaque, but the interface between the dynamic language types and the C types can introduce impedance. We hypothesise that frequent calls to C extension code introduce significant overhead that is often unnecessary. We validate this hypothesis by introducing a simple technique, "typed methods", which allow selected C extension functions to have additional metadata attached to them in a backward-compatible way. This additional metadata makes it much easier for a JIT compiler (and as we show, even an interpreter!) to significantly reduce the call and return overhead. Although we have prototyped typed methods in PyPy, we suspect that the same technique is applicable to a wider variety of language runtimes and that the information can also be consumed by static analysis tooling.

Dr Wenowdis: Specializing dynamic language C extensions using type information

TL;DR

The paper addresses the overhead incurred when dynamic languages call into CPython C extensions, which is opaque to static analysis and hinders JIT optimizations. It proposes Dr Wenowdis, a lightweight, backward-compatible mechanism to attach type and effect metadata to C extension functions, enabling fast call paths in PyPy by reducing boxing/unboxing and argument checks. Micro-benchmark results show substantial speedups in Python-to-C calls, including dramatic improvements on hot paths and even gains with the JIT disabled, suggesting the technique can improve performance without invasive rewrites. The work has implications for static analysis and could be generalized to other runtimes and binding tools, potentially enabling automatic generation of such annotations by existing binding generators like Cython or PyO3.

Abstract

C-based interpreters such as CPython make extensive use of C "extension" code, which is opaque to static analysis tools and faster runtimes with JIT compilers, such as PyPy. Not only are the extensions opaque, but the interface between the dynamic language types and the C types can introduce impedance. We hypothesise that frequent calls to C extension code introduce significant overhead that is often unnecessary. We validate this hypothesis by introducing a simple technique, "typed methods", which allow selected C extension functions to have additional metadata attached to them in a backward-compatible way. This additional metadata makes it much easier for a JIT compiler (and as we show, even an interpreter!) to significantly reduce the call and return overhead. Although we have prototyped typed methods in PyPy, we suspect that the same technique is applicable to a wider variety of language runtimes and that the information can also be consumed by static analysis tooling.
Paper Structure (18 sections, 3 figures)

This paper contains 18 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Background
  3. CPython
  4. The CPython C API
  5. PyPy
  6. cpyext and its problems
  7. Adding type information to C extensions
  8. Exposing Type Information
  9. The stable ABI
  10. Sneaking in pointers
  11. Using type information in PyPy
  12. Evaluation
  13. Prior work
  14. Sulong
  15. Argument Clinic
  16. ...and 3 more sections

Figures (3)

  • Figure 1: Example call to from Python code optimized by the JIT to a C function, passing one integer argument. The diagram shows all the needed conversions in the process.
  • Figure 2: Benchmark results
  • Figure 3: Benchmark for PyPy with JIT turned off