Musings on Tracing in PyPy (2025)

a day ago

Tracing JIT compilers generate code by recording execution paths, often focusing on loops, contrasting with method-based JITs that compile functions individually.
Meta-tracing JITs, like PyPy's RPython framework, allow reuse of JIT infrastructure across multiple languages, providing labor-saving benefits for compiler authors.
Advantages of tracing include aggressive partial inlining and path splitting naturally, simplifying optimizations due to linear control flow, which is effective for complex languages like Python.
Disadvantages involve performance cliffs, unpredictable behavior in branchy code (e.g., interpreters), and complexities in corner cases like recursive loops or trace length management.
Method-based JITs, such as Truffle/Graal, offer more predictable performance but require more development effort and funding, making tracing a pragmatic choice for resource-limited projects like PyPy.
PyPy's compatibility with Python's full semantics without language changes highlights the trade-offs in JIT design, with tracing excelling in pure Python code but struggling with C modules.
The evolution of JITs in Python includes alternatives like GraalPy and CPython's new JIT, indicating ongoing interest in performance improvements while navigating language complexity.

Hasty Briefsbeta