Home

Awesome

lang

The lang module/system/library can be loaded from multiple programming languages to seamlessly integrate all the other supported ones. The currently supported languages are Python and Lisp.

Installation

Each language requires one or more auxiliary libraries to be installed on your system:

Lisp

Requires libsbcl in your library path. Either install libsbcl with your favorite package manager or build and install it yourself.

Python

Requires libpython3.11 or later in your library path. To select a particular python environment, put the following into your .sbclrc:

(progn
  (ql:quickload :cffi)
  (pushnew "~/miniforge3/envs/MYENV/lib/"
           (symbol-value (intern "*FOREIGN-LIBRARY-DIRECTORIES*" "CFFI"))
           :test #'string=))

Showcases

Python from Lisp

(in-package #:cl-user)
(asdf:load-system :lang.python)
(sb-ext:add-package-local-nickname '#:python '#:lang.python)

(python:list #(1 2.0 #c(3 4) "5 6"))
 => [1, 2.0, #C(3 4) "5 6"]

(python:tuple (python:map #'class-of *))
 => (<class 'lang.lisp.common-lisp.fixnum'>,
     <class 'lang.lisp.common-lisp.single-float'>,
     <class 'lang.lisp.common-lisp.complex'>,
     <class 'lang.lisp.sb-kernel.simple-character-string'>)

(asdf:load-system :lang.python.numpy)
(sb-ext:add-package-local-nickname '#:np '#:lang.python.numpy)
(np:array #(1 2 3) np:float32)
 => array([1., 2., 3.], dtype=float32)

(python:* * 3)
 => array([3., 6., 9.], dtype=float32)

Lisp from Python

import lang.lisp

Next Steps

Things I (or some contributor?) plan to do next, in no particular order. Some of these ideas can probably be implemented in a few hours and others will be near impossible, but it is hard to tell which are which in advance.

Convert Lisp Packages to Python Modules

We can already convert Python modules into Lisp packages, and all the necessary machinery is there. The next step is to turn Lisp packages into Python modules. The plan is to fill each resulting Python module according to some fuzzy heuristics, and give each module foo two submodules foo.__functions__ and foo.__values__ that accurately reflect the two Lisp namespaces. Unless there is a collision, these submodules have the aliases foo.functions and foo.values. When a Lisp package name contains dots, we turn it into several nested modules.

A possible next step would be to not just convert packages to modules, but to turn them into modules directly, i.e., to have Lisp packages inherit from the Python module type.

Documentation, Exports

Most of the functionality in this library is not properly exported and documented. The interface also needs some curating. Address this.

Design Document

Write a document that codifies the rules we have used so far to embed one language into another. This document can then be consulted when adding a new language to keep things consistent.

One example of such a rule is that each object of language A has at exactly one mirror object in language B so that reference semantics are preserved. However, there are exceptions to this rule such as that the python class object maps directly to the Lisp class T to preserve the lattice property of each object hierarchy.

Another topic worth writing about in such a document is the mapping of identifiers in one language to those in another.

The document should also describe the object conversion protocol, and the decision of why we convert to Common Lisp by default (most flexible (meta-)object system, no restriction on symbol names, performance, macros).

Sequence Protocol

Use SBCL's extensible sequence protocol to provide the standard CL sequence functions for other languages' sequences.

Number Protocol

Convince the SBCL developers to provide an extensible protocol for numbers, so that we can finally use the standard functions on matrices, quaternions, and numbers of other programming languages.

Zero-copy Python strings.

Right now, Python strings and Lisp strings have to be carefully converted into each other, typically using UTF8-encoded buffers. However, I expect that they both have a very similar internal representation. If so, we could somehow pass them by reference. Might be a stupid idea, though.

Lambda Lists

Right now, the lambda list of each Python function is (&rest rest). Use introspection to make each lambda list resemble that of the corresponding Python function. Do the same for the other direction, so that Lisp functions in Python also display the correct structure of arguments. Bonus points for keeping this in sync when functions are redefined. Extra bonus points for automatically adding type annotations to each function, e.g., symbol_name(symbol: lang.lisp.common-lisp.symbol).

Auto-install

When loading a module that is not yet installed, the default should be to download it from the internet using a configurable approach for each language. Make this mechanism extensible, and also provide an option to not download anything from the internet. Write proper error handling to report when any part of this doesn't work.

Vectorcall Everywhere

We already use Python's fast vectorcall calling convention in some places, but not yet everywhere. Address this.

Bonus point for not using vectorcall for functions that actually expect a Python dictionary of keywords.

Write benchmarks to show whether everything works as intended.

Test Suite

Write a comprehensive test suite that ensures that everything works as expected.

One idea would be to use Python's AST module to turn Python code into Lisp code in the PYTHON package and to use that on the test suites of various Python libraries.

Write tests that check for the presence of memory leaks.

Write tests that use multiple Lisp and Python threads at the same time.

Write tests that perform various consistency checks.

Write tests that detect performance regressions.

Metaclasses in Python 3.12+

Python 3.12 adds the C API function PyType_FromMetaclass, so we can finally specify the correct metaclasses of each Lisp class. Do this. Whatever solution we come up with should still be backward compatible to earlier versions of Python.

Support older versions of Python

Right now we require at least Python 3.11, due to the C API functions we use. It would be nice to see whether those requirements could be lowered with some kludges and hacks.

Documentation Mirroring

Automatically convert Lisp docstrings to Python docstrings and vice versa. Bonus points for figuring out a way that to keep them in sync after a change. Possible solutions for keeping things in sync are the dependent maintenance protocol, or a generic writer function with an :after method.

Displace foreign arrays into Lisp arrays

Make it possible to convert Lisp arrays to Numpy arrays (or others) with by-reference semantics. The Lisp array either has to be displaced into a static vector, or pinned to its memory location.

Lisp Exceptions as Conditions

Make each Python condition a subclass of serious-condition, and arrange that calls between Python and Lisp propagate these conditions correctly. Bonus points for writing a Swank plugin that displays the correct stack trace containing both Python and Lisp stack frames.

Performance Improvements

The goal is to make the PYTHON package of Common Lisp faster than cpython :) The main challenge is probably the overhead of cross-language calls, which can be discussed with the SBCL developers. On the plus side, we have native compilation, and could even compile the bytecode of existing Python functions to Lisp on the fly.

Cross-language Garbage Collection

Right now there is no mechanism to collect circular structures that cross language boundaries more than once. One algorithm that could address this is the following:

Maintain a work list whose entries are pairs of programming languages and sets of reachable mirror objects.

  1. For each programming language, push the pair of that language and the empty set onto the work list.

  2. Pop one language and its set of reachable mirror objects from the work list.

  3. Starting from those mirror objects that are reachable, as well as all the usual GC roots, mark all reachable objects of that language.

  4. Determine the sets mirror objects that haven't been reachable before, but that are reachable now. Update the reachable object sets of all languages accordingly.

  5. Push every language whose set of reachable mirror objects has changed back onto the work list. Go to 2.

Once this algorithm terminates, destroy all mirror objects that are globally unreachable.

The main challenge here is to hack each language's memory management to allow for this kind of reachability analysis with dynamically changing roots. Ideally, the analysis should also be incremental, such that marking stops whenever it reaches an object that has already been marked in a previous step of the same cross-language GC macro cycle.

New Language: APL

Use the April APL compiler to implement lang.apl.

New Language: Forth

New Language: Lua

New Language: R

New Language: Ruby

New Languages: Rust, C, C++, Fortran

Use LLVM to support all these languages.

New Language: Scheme

Could we turn SBCL into a Scheme compiler? The main challenge is probably to implement first-class continuations. Sounds like a task for Kartik :)

New Languages: Java, Scala, Kotlin, Clojure

Make it possible to run JVM languages. One direction could be solved by implementing the JVM in Common Lisp. The other direction could be solved by porting lang to ABCL. This task is a lot of work, but also extremely rewarding.

New Languages: C#, F#

Make it possible to run .NET languages. Approach would be similar to Java.