社内勉強会資料_Two Papers Contribute to Faster Python.pdf

社内エンジニア・リサーチャー勉強会の発表資料「Two Papers Contribute to Faster Python」を公開しました！

Pythonの実行速度を向上させるための2つのアプローチ、
Specialized Adaptive InterpreterとCopy-and-patch JIT Compilationについて紹介しています。

1. Two Papers
Contribute to
Faster Python

2. Table of Contents
Faster CPython Project

Specializing Adaptive Interpreter

Copy-and-patch JIT Compilation

3. Faster Python Project

4. x5 Speed
Python
3.10
Python
3.11
Python
3.12
Python
3.13
Summarized in Implementation plan for speeding up CPython
⬆ 50
%
⬆50
%
⬆50
%
⬆50
%
Table of contents

5. Specialized Adaptive Interpreter
Motivated by Multi-level Quickening: The Key to Interpreter Performance

6. Multi-level Quickening: The Key to Interpreter Performance
Background
Interpreters for dynamically-typed programming languages are slow.

CPython
Interpreter
Python
Source
Code
Python
Virtual
Machine
Bytecode

7. Multi-level Quickening: The Key to Interpreter Performance
First-level Quickening: Type Feedback

8. Multi-level Quickening: The Key to Interpreter Performance
Second-level Quickening: Type Propagation

9. Multi-level Quickening: The Key to Interpreter Performance
Optimized Instruction Derivatives
i.e. Make instructions more optimized for machine data types
●Mapping Integer Numbers
○uint64_t ? int64_t ? int32_t
●Mapping Floating Point Number
●Mapping Complex Numbers.

10. Multi-level Quickening: The Key to Interpreter Performance
Super-instructions:
concatenates frequently occurring sequences of interpreter instructions
>>> dis.dis(add_two)
1 0 RESUME 0
2 LOAD_FAST 0 (a)
4 LOAD_FAST 1 (b)
6 BINARY_OP 0 (+)
10 RETURN_VALUE

>>> dis.dis(add_two, adaptive=True)
1 0 RESUME 0
2 LOAD_FAST__LOAD_FAST 0 (a)
4 LOAD_FAST 1 (b)
6 BINARY_OP_ADD_FLOAT 0 (+)
10 RETURN_VALUE
>>> def add_two(a, b): return a + b

11. Multi-level Quickening: The Key to Interpreter Performance
Performance
Results

12. Quickening & Specialization in Reality
Python 3.11
●Specialization clicks in after 8
executions.
●25% faster than Python 3.10.
Python 3.12
●Specialization clicks in after 2
executions.
●Special more new bytecodes.
Table of contents

13. Copy-and-patch JIT Compilation
Motivated by Copy-and-Patch Compilation: A fast compilation
algorithm for high-level languages and bytecode

14. Copy-and-Patch Compilation:
A fast compilation algorithm for high-level languages and bytecode
What’s JIT Compilation
●Just in time Compilation, only compile on demand when running for the 1st time.
●In contrast, compiled languages is using “Ahead of time”, compile before being used.

15. Copy-and-Patch Compilation:
A fast compilation algorithm for high-level languages and bytecode
Interpreting
Parse into
Bytecode
Execution
Copy-and-Patch JIT Compilation
Copy
Bytecode
Execution
Patch
arguments
Running a
function
several
times

16. Copy-and-Patch Compilation:
A fast compilation algorithm for high-level languages and bytecode
Normal Interpreting

Each time we run a function, CPython parse it into bytecodes, in a loop:
>>> def get_one():
... a = 1
... return a

>>> get_one()

>>> get_one()
1 0 RESUME 0
2 2 LOAD_CONST 1 (1)
4 STORE_FAST 0 (a)
3 6 LOAD_FAST 0 (a)
8 RETURN_VALUE
1 0 RESUME 0
2 2 LOAD_CONST 1 (1)
4 STORE_FAST 0 (a)
3 6 LOAD_FAST 0 (a)
8 RETURN_VALUE

17. Copy-and-Patch Compilation:
A fast compilation algorithm for high-level languages and bytecode
Copy-and-Patch JIT
>>> def get_one():
... a = 1
... return a
A function with the same
functionality, but written
in a lower language, e.g. C
or Assembly.

Table of contents
Caveat

Copy Instruction is kind of code generation, may expose security concerns.
iOS prohibits such dynamic code generation.

18. Copy-and-Patch Compilation in Reality
Table of contents
A basic JIT compiler will
be added in Python 3.13
(disabled in beta at the moment)

19. Take Away
●Upgrade to the latest Python version (official releases)
○Save energy
○Save money
○Save environment

20. References
●Implementation plan for speeding up CPython
●Multi-Level Quickening: Ten Years Later
●Copy-and-Patch Compilation: A fast compilation algorithm for high-level languages and bytecode
●Python 3.13 gets a JIT