Cross-Language Interop with .so Files

Ever wish you could write your performance-critical code in a single performant language, yet still tap into that functionality from other languages like Python, Go etc? Shared Object files (.so files) are your secret weapon for bridging language boundaries. In this post, we'll explore how .so files hold the key to language-agnostic bliss.

What are Shared Object files?

A shared object file (.so file on Linux) is a binary file that contains compiled code that multiple programs or languages can dynamically link to at runtime. Shared libraries are commonly used to encapsulate functionality and promote code reuse across applications.

Core Concepts

C ABI (Application Binary Interface): The C ABI defines how functions and data structures are represented in machine code, ensuring compatibility between compiled languages. By adhering to the C ABI, languages like Python, Go, Rust, and C++ can interact with the shared library.
Exporting Functions: Functions must be explicitly exported using specific language constructs like extern "C" in C++ or #[no_mangle] in Rust to ensure they are accessible.
Dynamic Linking: Programs load .so files at runtime using mechanisms like dlopen on Linux or language-specific bindings (e.g., ctypes in Python).
Data Exchange: Use simple data types like integers, floats, or strings for compatibility. Complex types may require serialization or custom bindings.

Common Challenges

Name Mangling: Compilers often change function names during compilation (e.g., in C++). Use extern "C" or #[no_mangle] to avoid this issue.
Memory Management: Shared libraries may allocate memory that must be freed by the calling application, requiring careful management.
Platform Differences: .so files work on Linux. We need .dll on Windows and .dylib on macOS.
Dependencies: Ensure all dependencies of the shared library are available on the target system.

Does Using a .so File Speed Up Processing?

Yes! Using a .so file can speed up processing, but the extent of the speed improvement depends on several factors:

Algorithm Optimization: Writing performance-critical code in a compiled language (e.g., C, Rust) and then exposing it via a .so file can lead to substantial speedups. The actual speedup depends on how optimized the code is. Poorly written code, even in a compiled language, won’t perform well.
Overhead of Interfacing: When using .so files in a high-level language like Python, there may be some overhead for calls between the language runtime (e.g., Python interpreter) and the shared library. This is negligible if the computational task in the .so file is significant.

Does the Speed Depend on the Source Language?

Yes! The source language can affect the execution speed of the .so file, depending on:

Language Performance Characteristics:
- C and C++: These are low-level, highly efficient languages with direct control over memory and system resources. .so files compiled from C or C++ generally offer the best performance for computation-heavy tasks.
- Python (via Cython): Compiling Python code with Cython can yield significant improvements, but it is generally slower than .so files from C or C++ due to Python’s runtime dependencies.
Garbage Collection and Runtime: Languages with garbage collection (like Python) may experience slight slowdowns due to runtime memory management. C gives developers complete control over memory, eliminating runtime overhead.

Remember, we are discussing about computationally expensive tasks. For I/O-bound tasks (e.g., reading/writing files, making network requests), the speed of .so files from different languages may not matter because:

The bottleneck is not CPU-bound but dependent on external systems.
Overhead from Python interfacing may dominate execution time.

Example

Let's try to implement BFS to find length of shortest path between a source and a destination, in both C++ and Cython, compile them to .so files and we'll use these .so files with Python and Golang.

C++ implementation

// bfs.cpp
#include <bits/stdc++.h>

extern "C" {
    int find_shortest_path(int n, int edgesCount, int* edges, int source, int destination) {
        std::vector<std::vector<int>> g(n);

        for(int i = 0; i < edgesCount; ++i) {
            int u = edges[2*i];
            int v = edges[2*i + 1];
            g[u].emplace_back(v);
            g[v].emplace_back(u);
        }

        int *dist = new int[n];
        std::memset(dist, -1, sizeof(int) * n);
        std::deque<int> q;

        dist[source] = 0;
        q.push_back(source);

        while (!q.empty()) {
            int u = q.front();
            q.pop_front();

            if(u == destination) {
                int ret = dist[u];
                delete[] dist;
                return ret;
            }

            for(int neighbor : g[u]) {
                if(dist[neighbor] == -1) {
                    dist[neighbor] = dist[u] + 1;
                    q.push_back(neighbor);
                }
            }
        }

        delete[] dist;
        return -1;
    }
}

We compile this using g++ -O3 -shared -fPIC bfs.cpp -o libbfs.so

-fPIC: Generate position-independent code (required for shared libraries).
- The -fPIC flag is vital when creating shared libraries because it ensures that the generated code is position-independent. This enables the shared library to be loaded at any memory address, supports efficient memory sharing across processes, facilitates dynamic symbol resolution, and prevents linking and runtime errors. Without -fPIC, creating robust, flexible, and efficient shared libraries would be significantly more challenging.
-shared: Produce a .so.

Cython implementation

We disabled bounds checking and wraparound with the directives at the top (# cython: boundscheck=False, etc.) for performance.

# bfs.pyx
# cython: boundscheck=False
# cython: wraparound=False

def _helper_bfs(n, edges_count, edges, source, destination):
    from collections import deque

    adjacency = [[] for _ in range(n)]
    for i in range(edges_count):
        u = edges[2*i]
        v = edges[2*i + 1]
        adjacency[u].append(v)
        adjacency[v].append(u)

    dist = [-1] * n
    dist[source] = 0

    q = deque()
    q.append(source)

    while q:
        current = q.popleft()
        if current == destination:
            return dist[current]

        for neighbor in adjacency[current]:
            if dist[neighbor] == -1:
                dist[neighbor] = dist[current] + 1
                q.append(neighbor)

    return -1

def find_shortest_path(int n, int edges_count, int[:] edges, int source, int destination) -> int:
    cdef list py_edges = [0]*(2 * edges_count)
    cdef int i
    for i in range(2 * edges_count):
        py_edges[i] = edges[i]

    return _helper_bfs(n, edges_count, py_edges, source, destination)

To compile bfs.pyx file, we also need to create a setup.py.

# setup.py
from setuptools import setup, Extension
from Cython.Build import cythonize

extensions = [
    Extension(
        name="bfs",
        sources=["bfs.pyx"],
    )
]

setup(
    name="bfs",
    ext_modules=cythonize(extensions, language_level="3"),
)

And, now to get .so file from it, we run python setup.py build_ext --inplace. This gives us a .so file named bfs.cpython-38-x86_64-linux-gnu.so.

Inspecting the .so files

Comparing size

$ du -sh libbfs.so 
20K	libbfs.so

$ du -sh bfs.cpython-38-x86_64-linux-gnu.so 
1.3M	bfs.cpython-38-x86_64-linux-gnu.so

Lets use objdump to see the symbols in each .so file.

$ objdump -T libbfs.so | wc -l
31

$ objdump -T bfs.cpython-38-x86_64-linux-gnu.so | wc -l 
162

The one compiled from C++ has fewer symbols because it mostly has just our BFS function and standard library calls, while the one compiled from cython includes Python/C API symbols or references to them as well.

We are going to use following egdes as setup for all experiments and number of nodes as 300.

Edges

edges = [(0, 11), (0, 177), (0, 194), (0, 288), (1, 8), (1, 68), (1, 178), (1, 243), (1, 251), (2, 202), (3, 76), (3, 114), (3, 135), (3, 139), (4, 80), (4, 154), (4, 176), (4, 187), (4, 235), (5, 88), (5, 94), (5, 96), (5, 99), (5, 123), (5, 151), (5, 241), (5, 290), (6, 100), (6, 105), (6, 263), (7, 33), (7, 52), (7, 62), (7, 74), (7, 178), (7, 199), (7, 276), (8, 29), (8, 166), (8, 177), (8, 187), (9, 250), (10, 233), (10, 237), (11, 137), (11, 143), (11, 194), (12, 146), (12, 243), (12, 245), (13, 24), (13, 99), (13, 238), (13, 255), (14, 24), (14, 46), (14, 57), (14, 70), (14, 82), (14, 147), (14, 169), (14, 189), (14, 268), (15, 66), (15, 149), (15, 193), (15, 247), (16, 165), (16, 216), (16, 226), (16, 227), (16, 253), (17, 90), (17, 182), (18, 201), (18, 235), (18, 289), (19, 248), (20, 103), (20, 109), (20, 263), (21, 113), (21, 138), (21, 170), (21, 171), (21, 265), (22, 79), (22, 105), (22, 211), (22, 272), (23, 78), (23, 200), (23, 226), (24, 27), (24, 121), (24, 143), (24, 268), (25, 66), (25, 95), (25, 154), (25, 169), (25, 201), (26, 156), (27, 140), (27, 264), (27, 276), (27, 295), (28, 63), (28, 123), (29, 51), (29, 153), (29, 283), (29, 289), (30, 39), (30, 72), (30, 73), (31, 233), (32, 67), (32, 93), (32, 287), (33, 140), (33, 141), (33, 209), (33, 227), (33, 264), (33, 289), (34, 66), (34, 81), (34, 113), (34, 156), (34, 182), (34, 260), (34, 297), (35, 75), (35, 123), (35, 176), (35, 222), (36, 61), (36, 118), (36, 119), (36, 121), (36, 130), (36, 170), (36, 218), (37, 153), (37, 181), (37, 234), (38, 181), (39, 87), (39, 105), (39, 106), (40, 278), (41, 55), (41, 211), (41, 271), (41, 298), (42, 217), (42, 223), (42, 239), (43, 52), (43, 60), (43, 139), (43, 251), (43, 290), (44, 48), (44, 99), (44, 112), (44, 164), (44, 211), (45, 98), (45, 196), (46, 102), (46, 111), (46, 180), (46, 248), (46, 294), (47, 122), (47, 189), (48, 71), (48, 132), (48, 230), (49, 234), (49, 291), (50, 57), (50, 73), (51, 62), (51, 155), (51, 196), (52, 120), (52, 180), (53, 195), (53, 216), (53, 226), (53, 288), (54, 87), (54, 185), (54, 186), (55, 76), (55, 180), (56, 63), (56, 90), (56, 277), (57, 161), (58, 96), (58, 124), (58, 130), (58, 206), (58, 207), (58, 223), (58, 252), (59, 90), (59, 138), (59, 184), (60, 82), (61, 127), (61, 285), (62, 123), (62, 141), (62, 163), (62, 196), (62, 296), (63, 113), (63, 164), (64, 89), (64, 131), (65, 135), (65, 159), (65, 236), (65, 253), (65, 275), (65, 289), (66, 144), (66, 164), (66, 282), (66, 284), (67, 74), (67, 136), (67, 208), (67, 285), (67, 288), (68, 82), (68, 140), (69, 91), (69, 111), (69, 113), (69, 175), (69, 178), (70, 159), (70, 224), (70, 237), (70, 294), (71, 81), (71, 160), (72, 158), (72, 172), (72, 196), (72, 230), (73, 279), (74, 108), (74, 126), (74, 265), (75, 86), (75, 202), (76, 101), (76, 169), (76, 294), (77, 126), (77, 137), (77, 148), (77, 151), (77, 186), (77, 264), (78, 115), (78, 265), (79, 92), (79, 167), (79, 194), (80, 169), (80, 173), (80, 220), (80, 243), (80, 277), (81, 198), (81, 221), (82, 112), (82, 194), (82, 205), (82, 219), (82, 234), (82, 242), (82, 259), (82, 264), (82, 292), (83, 93), (83, 202), (83, 210), (83, 275), (84, 216), (85, 183), (85, 226), (85, 228), (86, 165), (86, 268), (87, 144), (88, 99), (88, 242), (88, 293), (89, 96), (89, 173), (89, 234), (89, 244), (89, 254), (90, 124), (90, 130), (90, 189), (90, 252), (91, 269), (92, 98), (92, 128), (92, 234), (92, 265), (93, 171), (93, 190), (93, 275), (94, 119), (94, 148), (94, 203), (94, 206), (94, 274), (95, 211), (96, 144), (96, 207), (97, 124), (97, 157), (97, 229), (97, 249), (98, 117), (98, 124), (98, 126), (98, 172), (98, 257), (98, 265), (99, 123), (99, 147), (99, 179), (100, 106), (100, 138), (100, 170), (100, 221), (100, 223), (101, 112), (101, 170), (101, 280), (101, 291), (102, 208), (102, 271), (103, 128), (103, 173), (103, 268), (104, 299), (105, 114), (105, 274), (106, 203), (107, 165), (108, 215), (108, 262), (109, 146), (110, 298), (113, 291), (114, 142), (114, 143), (114, 282), (115, 222), (115, 258), (115, 264), (116, 217), (116, 293), (117, 192), (118, 169), (118, 263), (118, 269), (119, 170), (119, 188), (119, 220), (119, 242), (119, 279), (120, 224), (120, 229), (120, 287), (121, 180), (121, 186), (121, 270), (121, 277), (122, 127), (122, 134), (122, 137), (123, 169), (123, 186), (123, 227), (123, 242), (123, 293), (124, 286), (125, 185), (126, 186), (126, 188), (126, 243), (126, 265), (127, 181), (127, 187), (127, 213), (128, 164), (129, 253), (130, 189), (130, 210), (131, 170), (131, 285), (132, 224), (132, 296), (133, 171), (133, 255), (135, 206), (137, 165), (137, 176), (137, 271), (137, 299), (138, 204), (138, 220), (138, 235), (139, 170), (139, 188), (140, 196), (140, 284), (141, 169), (141, 267), (142, 170), (142, 236), (143, 150), (144, 257), (144, 299), (145, 161), (145, 191), (145, 222), (145, 268), (145, 291), (146, 179), (146, 201), (147, 255), (148, 252), (148, 261), (148, 270), (149, 163), (149, 294), (150, 211), (150, 247), (150, 266), (152, 171), (152, 191), (152, 284), (153, 155), (155, 225), (155, 291), (156, 276), (157, 220), (157, 234), (157, 236), (157, 298), (158, 235), (159, 207), (159, 295), (160, 190), (160, 203), (160, 206), (160, 253), (161, 228), (161, 268), (161, 272), (162, 174), (163, 216), (163, 238), (164, 294), (165, 181), (166, 170), (167, 172), (167, 237), (168, 269), (168, 273), (169, 184), (169, 222), (169, 269), (170, 232), (170, 255), (170, 263), (172, 206), (172, 233), (172, 291), (174, 185), (174, 186), (174, 192), (174, 207), (174, 214), (174, 255), (175, 244), (175, 288), (177, 235), (177, 278), (178, 264), (179, 189), (180, 190), (180, 227), (180, 234), (182, 186), (182, 258), (184, 203), (187, 206), (187, 291), (188, 248), (188, 264), (188, 273), (188, 285), (189, 190), (189, 201), (190, 293), (194, 265), (195, 207), (195, 260), (198, 272), (199, 210), (199, 293), (200, 264), (201, 241), (202, 266), (202, 272), (202, 275), (203, 209), (204, 214), (205, 265), (208, 222), (208, 240), (208, 258), (209, 242), (209, 247), (209, 279), (210, 224), (210, 282), (210, 290), (213, 216), (213, 249), (213, 251), (213, 288), (214, 230), (214, 234), (215, 247), (215, 269), (216, 242), (217, 235), (218, 262), (219, 256), (219, 263), (220, 254), (221, 275), (221, 293), (224, 235), (225, 239), (225, 292), (227, 292), (227, 297), (228, 254), (228, 276), (229, 278), (229, 286), (230, 259), (230, 287), (233, 289), (236, 244), (237, 254), (237, 297), (238, 267), (239, 277), (241, 281), (243, 281), (243, 293), (243, 296), (249, 260), (250, 289), (250, 293), (250, 297), (252, 268), (254, 292), (261, 276), (261, 278), (262, 278), (262, 289), (263, 278), (263, 279), (263, 296), (272, 286), (273, 278), (282, 297), (284, 298), (294, 298)]

Using libbfs.so In Python

# bfs_from_cpp_so.py
import ctypes

libbfs = ctypes.CDLL('./libbfs.so')

libbfs.find_shortest_path.argtypes = [
    ctypes.c_int,
    ctypes.c_int,
    ctypes.POINTER(ctypes.c_int),
    ctypes.c_int,
    ctypes.c_int
]
libbfs.find_shortest_path.restype = ctypes.c_int

def get_shortest_path_length(n, edges, source, destination):
    edges_flat = []
    for (u, v) in edges:
        edges_flat.append(u)
        edges_flat.append(v)

    edges_array = (ctypes.c_int * len(edges_flat))(*edges_flat)
    
    return libbfs.find_shortest_path(n, len(edges), edges_array, source, destination)

if __name__ == "__main__":
    n = 300
    edges = ...

    max_bfs_len = -1
    max_bfs_src, max_bfs_dst = None, None
    for source in range(n):
        for destination in range(n):
            if source == destination:
                continue
            distance = get_shortest_path_length(n, edges, source, destination)
            if distance == -1:
                continue
            if distance > max_bfs_len:
                max_bfs_len = distance
                max_bfs_src = source
                max_bfs_dst = destination

    if max_bfs_len > -1:
        print(f"longest 'shortest distance' is from {max_bfs_src} to {max_bfs_dst}:", max_bfs_len)

$ /bin/time -v python bfs_from_cpp_so.py
longest 'shortest distance' is from 2 to 19: 8
	User time (seconds): 9.84
	...

Using bfs.cpython-38-x86_64-linux-gnu.so In Python

# bfs_from_pyx_so.py
import ctypes
import bfs


def get_shortest_path_length(n, edges, source, destination):
    edges_flat = []
    for (u, v) in edges:
        edges_flat.append(u)
        edges_flat.append(v)

    edges_array = (ctypes.c_int * len(edges_flat))(*edges_flat)
    
    return bfs.find_shortest_path(n, len(edges), edges_array, source, destination)

if __name__ == "__main__":
    n = 300
    edges = ...

    max_bfs_len = -1
    max_bfs_src, max_bfs_dst = None, None
    for source in range(n):
        for destination in range(n):
            if source == destination:
                continue
            distance = get_shortest_path_length(n, edges, source, destination)
            if distance == -1:
                continue
            if distance > max_bfs_len:
                max_bfs_len = distance
                max_bfs_src = source
                max_bfs_dst = destination

    if max_bfs_len > -1:
        print(f"longest 'shortest distance' is from {max_bfs_src} to {max_bfs_dst}:", max_bfs_len)

$ /bin/time -v python bfs_from_pyx_so.py
longest 'shortest distance' is from 2 to 19: 8
	User time (seconds): 13.79
	...

Using libbfs.so In Golang

// main.go
package main

/*
#cgo LDFLAGS: -L. -lbfs

// Forward-declare the BFS function
int find_shortest_path(int n, int edgesCount, int* edges, int source, int destination);
*/
import "C"
import (
	"fmt"
	"unsafe"
)

type edge struct {
	u int
	v int
}

func main() {
	edges := []edge{...}
	flattenedEdges := []int{}
	for _, e := range edges {
		flattenedEdges = append(flattenedEdges, e.u)
		flattenedEdges = append(flattenedEdges, e.v)
	}
	n := 300
	edgesCount := len(edges)

	cEdges := make([]C.int, len(flattenedEdges))
	for i, v := range flattenedEdges {
		cEdges[i] = C.int(v)
	}

	maxBfsLen := C.int(-1)
	maxBfsSrc := -1
	maxBfsDst := -1
	for source := 0; source < n; source++ {
		for destination := 0; destination < n; destination++ {
			if source == destination {
				continue
			}
			dist := C.find_shortest_path(C.int(n), C.int(edgesCount), (*C.int)(unsafe.Pointer(&cEdges[0])), C.int(source), C.int(destination))
			if dist == -1 {
				continue
			}
			if dist > maxBfsLen {
				maxBfsLen = dist
				maxBfsSrc = source
				maxBfsDst = destination
			}
		}
	}
	fmt.Printf("longest 'shortest distance' is from %d to %d: %d\n", maxBfsSrc, maxBfsDst, maxBfsLen)
}

The cgo comments at the top define how the Go toolchain should link your .so file:
- -L. says “look in the current directory for libraries.”
- -lbfs links against libbfs.so.
We declare the function int find_shortest_path(...) so the Go compiler knows the correct signature.

To build, run export LD_LIBRARY_PATH=.:$LD_LIBRARY_PATH && go build -o main main.go

$ ./main
longest 'shortest distance' is from 2 to 19: 8

Keep In Mind While Using .so Files

CPU-Bound + Large Data, Single Call: If you can pass a big chunk of data once, do lots of work in C/C++, and return the result, that’s where you’ll see big speedups over a naive Python implementation.
Minimize Overhead: Repeated small calls across the boundary can kill performance.