Skip to content

Latest commit

 

History

History
55 lines (44 loc) · 3.45 KB

README.md

File metadata and controls

55 lines (44 loc) · 3.45 KB

Pypyrust, a Python-implemented Python to Rust compiler

Python is the world's most popular language, in part because it is so easy to write. Rust is arguably the world's best general-purpose language: it is fast, with concise syntax and rigorous safety rules. Moreover, it has features that make it potentially good for replicating Python. It supports functional and object-oriented programming, though without inheritance. Its traits are similar to interfaces in Python, though it does not support duck typing (if it waddles like a duck and quacks like a duck, it is a duck).

Restrictions on the Python side

Python natively uses variants for everything, where a variant can be an integer, float, complex, boolean, string, tuple, list, set or dict. (Objects are semantic sugar around dicts.) It would be possible to compile Python into Rust code that used variants in the same way, and this is how Cython by default compiles Python into C code. However, the resulting Rust code would be horrible both to read and to execute, being only a few times quicker than CPython.

We therefore require the Python code to be supplied with sufficient type hints to allow the Rust compiler to recognise which Rust types to use. Rust is clever enough to deduce types where they are not ambiguous, so really only the function headers need hinting.

Rust does not support execution of arbitrary code fragments, so this is not supported in Python either. In practice, Python tends not to be written like this anyway, and Python is written in modules, with all the code exported in functions or classes.

Restrictions on the Rust side

One of the things that makes Rust so fast is that it allows programmers to manage the lifetime of objects in a safe way. In comparison, C is as flexible, allowing items to be allocated statically, on the stack or the heap, but C is not at all safe, allowing stack items to be accessed after they cease to exist, and allowing heap items to be deleted multiple times or not at all. In contrast, Python is safe but not at all flexible -- all items are represented by the same type of variant, which is initially allocated on the stack but has internal pointers that refer to items, including other variants, which are allocated on a garbage-collected heap.

In compiling Python to Rust, we cannot reproduce the richness of Rust lifetimes. We represent those items that Python would allocate on the heap as Rust heap-items, which must be garbage collected. Rust has no built-in garbage collector. There is a third-party garbage collector implementation, shifgrethor, which we can experiment with. For now, we just use reference counting, and cross our fingers about circular references.

Interoperability

One of the main design features of Cython, and in fact the main thing that slows it down, is that Cython code (Python code compiled with Cython into C code, and then compiled into binary object code) is 100% interoperable with Python. You can use Cython on some of your modules but not others, and you can mix libraries that were compiled with Cython with those that are just interpreted Python.

Pypyrust is not like this. It compiles Python into native Rust code, without interoperability wrappers. If you want interoperability between Pypyrust modules and Python modules or libraries, you have to design and write the iterop layer. (An obvious extension to Pypyrust is to auto-generate this interop layer. We leave this as an exercise.)