llvmcpy 0.2.0

Python bindings for LLVM auto-generated from the LLVM-C API

llvmcpy

llvmcpy automatically generates Python wrappers for the LLVM-C API.

Goal

The main goal of llvmcpy is to provide Python bindings for the LLVM project that are fast and require the lowest possible maintainance effort. To achive this, we use CFFI to parse the (slightly adapted) LLVM-C API header files and automatically generate a set of classes and functions to interact with them in a Pythonic way.

This project is in an early stage, but allows you to run the following code:

import sys
from llvmcpy import LLVMCPy

llvm = LLVMCPy()
buffer = llvm.create_memory_buffer_with_contents_of_file(sys.argv[1])
context = llvm.get_global_context()
module = context.parse_ir(buffer)
for function in module.iter_functions():
    for bb in function.iter_basic_blocks():
        for instruction in bb.iter_instructions():
            instruction.dump()

It's tested on all LLVM versions from 5 to 19 and on Python from 3.7 to 3.13. Supporting newer versions of the LLVM-C API should be basically effortless.

To try it out, install LLVM and install llvmcpy:

sudo apt-get install llvm
python -m venv venv
source venv/bin/activate
pip install llvmcpy

Naming of the generated classes/functions

The basic idea behind this project is to take the LLVM-C API, create a class for each data type and create a method for that class for each function in the API taking an argument of that data type as first argument.

This means that the following functions:

LLVMModuleRef LLVMCloneModule (LLVMModuleRef M)

Will become:

class Module(object):
    def clone(self):
        # ...

Note the change in the case. Use help(Module.clone) to see which LLVM-C API function a certain method is using.

Each class in llvmcpy is basically a wrapper around a pointer to an LLVM object.

If an API function doesn't take an LLVM object as a first argument, it will be part of the llvm module.

Additionally, we have some generated properties and generators for certain well known patterns in the API.

Properties

For each function starting with LLVMGet or LLVMSet in the LLVM-C API, we generate a property. For example, consider the following two functions:

void LLVMSetValueName (LLVMValueRef Val, const char *Name);
const char* LLVMGetValueName(LLVMValueRef Val);

In llvmcpy the Get/Set prefixes disappear, along with Value (the name of the class) and you can use them as properties of the Value class, e.g.:

my_value.name = "sum"
print(my_value.name)

Generators

The LLVM-C API has a recurrent pattern which allows you to navigate through the hierarchy of its class hierarchy, i.e. the pair of LLVMGetSomething and LLVMGetNextSomething functions. Something can be Function, BasicBlock and so on. llvmcpy identifies these patterns and produces a generator method which allows you to iterate over these objects in a Pythonic way:

for function in module.iter_functions():
    for bb in function.iter_basic_blocks():
        for instruction in bb.iter_instructions():
            # ...

Where are my bindings?

Bindings are automatically generated in a lazy way. Multiple installations of LLVM are supported, just set the LLVM_CONFIG environment variable to the llvm-config program in the bin/ directory of your LLVM installation and everything should work fine.

The bindings are generated in a Python script which is stored in $XDG_CACHE_DIR/llvmcpy/ (typically ~/.cache/llvmcpy) in a directory whose name is obtained by hashing the full path of the output of llvm-config --prefix concatenated with the LLVM version number. For example, for LLVM 19 installed in /usr you'll find the API bindings in ~/.cache/llvmcpy/7fea08f2e9d5108688f692e686c8528b914eda563e7069b25ef18c49ba96d7f2-19.

To generate the bindings, a working C preprocessor must be available in the system. By default, cpp (the C preprocessor part of GCC) is used. If it's not available we check if clang is available in the LLVM installation and use it.

License and credits

This project is developed and maintained by rev.ng Labs as part of the rev.ng project, and it's released under the MIT license.