Portable, Efficient Software

Almost every language provides a means of calling C functions. The key to high performance software is passing pointers to data from the calling language so memory does not need to be copied.

Portability involves the Application Binary Interface. In order for a language to use a C library it must link to either a static library at compile time or a dynamic library at run-time. Run-time linking requires information about the version of the library being loaded and any dependencies it has on other dynamic libraries. The most portable approach is to write C++ code, compile static libraries, and use extern "C" to provide C linkage of public functions that can be used from almost every language to avoid ABI issues.

The C ABI specifies the size and layout of standard types, structs and arrays of standard types, linker visable names, and calling conventions. As long as the host language respects this, it is possible to provide a single library for each platform.

Garbage collected languages usually have a means to temporarily pin data to prevent garbage collection while calling C functions. If the C function requires memory the best practice is to provide a means to communicate that to the calling function. The C function should do no memory allocation, or offload that to the C++ functions being called.

To make a library easily available to a host language requires an understanding of their Foreign Function Interface.

C++

Existing C++ libraries can be wrapped using a C interface if they use a subset of C++ features. For example arguments of type const std::vector<T>& can be called from C using const T*, size_t but memory will be copied to the vector. This can be avoided if the C++ library uses non-owning memory such as std::span<T> or std::mdspan.

Smart pointers involve memory allocation. Use std::unique_ptr for the tightest control on that. Avoid std::shared_ptr when possible. It is trivial to convert a unique pointer to a shared pointer when necessary, but a shared pointer abdicates control to any code using it.

C#

C# uses pin_ptr to pause garbage collection. It must be defined as a local variable before passing it to a C++/CLI function. When it goes out of scope the GC takes over. C++ does not allow 0 length arrays so 0 length C# arrays require special handling.

    void cli(array<T>^ a) {
    {
        pin_ptr<T> pa = (a.length == 0 ? nullptr : &a[0]);
        cpp(pa, a.length);
    }

Here pa gets passed to the C++/CLI function cpp as a T* to provide access to the array of T managed by a. Passing pa when a.length is zero is undefined behaviour.

Portable, Efficient Software

C++

C#

Java

Python