Pyjion - a JIT extension system for CPython
- 1. Pyjion
- 2. What are we going to discuss ļµ When you develop a Python app, what happens when you execute it ļµ Why does that happen and how does it work? ļµ How can it be improved? ļµ What is a JIT ļµ What is Pyjion
- 3. Abstract Syntax Trees ļµ Clone https://github.com/quantifiedcode/python-ast-visualizer ļµ pip install ār requirements.txt ļµ cat example.py | python astvisualizer.py ļµ Uses built in module ast and function compile
- 4. What is ByteCode ļµ Machine code is the lowest level but not portable between processors ļµ Assembly is short-hand for machine code ļµ C is compiled and linked into machine-code which is efficient but not portable once compiled ļµ Contrast Jython to Cpython, Jython compiles the AST to JVM bytecode, Cpython to Cpython bytecode (pyc files)
- 5. cPython bytecode ļµ The cPython compiler reads the text (command, file etc) ļµ Prepares Abstract Syntax Tree ļµ The cPython compiler then converts the AST into Bytecode ļµ Bytecode operations are low level operations that translate to likely OS/machine level instructions.
- 6. Using DIS to inspect byte-code def helloworld(): print("Hello world!") dis.dis(helloworld) 2 0 LOAD_GLOBAL 0 (print) 3 LOAD_CONST 1 ('Hello world!') 6 CALL_FUNCTION 1 (1 positional, 0 keyword pair) 9 POP_TOP 10 LOAD_CONST 0 (None) 13 RETURN_VALUE Using the standard library ādisā module, you can disassemble a Python object into byte-code in the Python CLI. https://github.com/python/cpython/blob/ab8b3ff250a1 3d506f572abd66c37bce0759ab7d/Lib/dis.py#L31-L65
- 7. Python threads and the Global Interpreter Lock ļµ Python has native support for threading ļµ Python threads are real system threads (either POSIX or Windows) ļµ Those threads are managed by the host operating system ļµ Python threads are simple callable objects ļµ The GIL ensures that sure each thread gets exclusive access to the interpreter internals when it's running (and that call-outs to C extensions play nice) ļµ See: http://www.dabeaz.com/python/GIL.pdf
- 8. How a Python thread executes Thread created Python creates data structure (PyThreadState) for thread The pthread is launched The thread calls PyEval_CallObject (run the callable)
- 9. Just-in-time compilation ļµ Just-in-time compilation requires an intermediate language (IL) to allow the code to be split into chunks (or frames) ļµ Ahead of time (AOT) compilers are designed to ensure that, the CPU can understand every line in the code before any interaction takes place. ļµ In the case of Python, the execution is converted to byte-code when called. Those byte-code statements are cached in .pyc files. ļµ This diagram shows the .NET MSIL to JIT compiler flow
- 10. .NET core ļµ Microsoft.NET is a JIT framework, with a series of languages supported by Microsoft and interopability with the common MSIL intermediate language. The .NET runtime is the MSIL JIT compilation module, but the executable and installer includes 100ās of modules specific for Windows and Windows application development. ļµ Microsoft have redeveloped .NET into a new lightweight framework focusing on the core elements (JIT and compilation) as well as the core data structures and common classes. This is developed in C++ and is cross-platform. ļµ The JIT module itself in .NET core can be used without the MSIL purely for executing byte-code operators on the local CPU. ļµ This is what Pyjion does.
- 11. Python CodeObjects ļµ The goal of the JIT is to convert Python Code Objects (either loaded via imports, on the CLI or in a thread) to machine instructions. ļµ The low level data structure is called PyCodeObject, declared in Cpython. ļµ Pyjion introduces a new abstract type, IPythonCompiler, which takes a pointer to a PyCodeObject and executes (compiles) the code against a target ILGenerator ļµ The ILGenerator generates DotNetCore IL instructions for each Python code
- 12. End to end workflow Python creates PyCodeObject from code file/input PythonCompiler takes PyCodeObject and converts Python OpCodes into intermediate instructions (Pyjion IL?) ILGenerator converts abstract IL operations to send to the ICorJitCompiler ICorJitCompiler compiles stacks to machine code ICorJitInfo is the main interface that the JIT uses to call back to the execution engine (EE) and get information. DotNetCorePyjionPython print(āhello world!ā) 010101110011
- 13. Proposed extensions to CPython ļµ In order to support Pyjion (or another JIT engine) some changes are proposed to Cpython ļµ Load pyjit.dll and run ājitInitā (pylifecycle.cpp) ļµ JitInitFunction jitinit = (JitInitFunction)GetProcAddress(pyjit, "JitInit"); jitinit(); ļµ Ceval (the code that evaluates the .pyc opcode cache) can now ask the JIT module to evaluate frames
- 14. Does this help me run .NET from Python or vice versa? ļµ No. ļµ .NET is not only an intermediate-language and execution engine, but also a huge collection of standard libraries (similar to Python). ļµ Pyjion bridges the gap between the low-level function that CPython performs executing byte-code operations on the local CPU and replaces the last step with a 3rd party execution engine.
- 15. Building the project ļµ You will need ļµ Git, SVN ļµ Cmake ļµ Visual Studio (well MSBuild) 1. Download Pyjion (github.com/Microsoft/Pyjion) 2. Install submodules 3. Apply patches to CoreCLR and Python 4. Compile Pyjion 5. Copy build output to the CPython distribution output (pyjit.dll)
- 16. About me @anthonypjshaw @tonybaloney anthonyshaw@apache.org Head of Innovation at Dimension Data