YOMM2

This library implements fast, open, multi-methods for C++17. It is strongly inspired by the papers by Peter Pirkelbauer, Yuriy Solodkyy, and Bjarne Stroustrup.

• YOMM2
  • TL;DR
  • Open Methods in a Nutshell
    • Cross-cutting Concerns and the Expression Problem
    • Multiple Dispatch
  • Performance
  • Installation
  • Going Further
  • Roadmap

TL;DR

If you are familiar with the concept of open multi-methods, or if you prefer to learn by reading code, go directly to the synopsis. The documentation is here

Open Methods in a Nutshell

Cross-cutting Concerns and the Expression Problem

You have a matrix math library. It deals with all sort of matrices: dense, diagonal, tri-diagonal, etc. Each matrix subtype has a corresponding class in a hierarchy rooted in Matrix.

Now you would like to render Matrix objects as JSON strings. The representation will vary depending on the exact type of the object; for example, if a matrix is a DiagonalMatrix, you only need to store the diagonal - the other elements are all zeroes.

This is an example of a "cross-cutting concern". How do you do it?

It turns out that OOP doesn't offer a good solution to this.

You can stick a pure virtual to_json function in the Matrix base class and override it in the subclasses. It is an easy solution but it has severe drawbacks. It requires you to change the Matrix class and its subclasses, and recompile the library. And now all the applications that use it will contain the to_json functions even if they don't need them, because of the way virtual functions are implemented.

Or you may resort on a "type switch": have the application test for each category and generate the JSON accordingly. This is tedious, error prone and, above all, not extensible. Adding a new matrix subclass requires updating all the type switches. The Visitor pattern also suffers from this flaw.

Wouldn't it be nice if you could add behavior to existing types, just as easily and unintrusively as you can extend existing class hierarchies via derivation? What if you could solve the so-called Expression Problem:

existing behaviors += new types
existing types += new behaviors

This is exactly what Open Methods are all about: solving the Expression Problem.

Let's look at an example.

// -----------------------------------------------------------------------------
// library code

struct matrix {
    virtual ~matrix() {
    }
    // ...
};

struct dense_matrix : matrix { /* ... */
};
struct diagonal_matrix : matrix { /* ... */
};

// -----------------------------------------------------------------------------
// application code

#include <memory>
#include <yorel/yomm2/keywords.hpp>

register_classes(matrix, dense_matrix, diagonal_matrix);

declare_method(std::string, to_json, (virtual_<const matrix&>));

define_method(std::string, to_json, (const dense_matrix& m)) {
    return "json for dense matrix...";
}

define_method(std::string, to_json, (const diagonal_matrix& m)) {
    return "json for diagonal matrix...";
}

int main() {
    yorel::yomm2::update();

    const matrix& a = dense_matrix();
    const matrix& b = diagonal_matrix();

    std::cout << to_json(a) << "\n"; // json for dense matrix
    std::cout << to_json(b) << "\n"; // json for diagonal matrix

    return 0;
}

<yorel/yomm2/keywords.hpp> is the library's main entry point. It declares a set of macros, and injects a single name, virtual_, in the global namespace. The purpose of the header is to make it look as if open methods are part of the language.

register_classes informs the library of the existence of the classes, and their inheritance relationships. Any class that can appear in a method call needs to be registered, even if it is not directly referenced by a method.

declare_method declares an open method called to_json, which takes one virtual argument of type const matrix& and returns a std::string. The virtual_<> decorator specifies that the argument must be taken into account to select the appropriate specialization. In essence, this is the same thing as having a virtual std::string to_json() const inside class Matrix - except that the virtual function lives outside of any classes, and you can add as many as you want without changing the classes. NOTE: DO NOT specify argument names, i.e. virtual_<const matrix&> arg is not permitted.

define_method defines two implementations for the to_json method: one for dense matrices, and one for diagonal matrices.

yorel::yomm2::update() creates the dispatch tables; it must be called before any method is called, and after dynamically loading and unloading shared libraries.

The example can be compiled (from the root of the repository) with:

clang++- -I include -std=c++17 tutorials/README.cpp -o example

Multiple Dispatch

Methods can have more than one virtual argument. This is handy in certain situations, for example to implement binary operations on matrices:

// -----------------------------------------------------------------------------
// matrix * matrix

declare_method(
    std::shared_ptr<const matrix>,
    times, (virtual_<const matrix&>, virtual_<const matrix&>));

// catch-all matrix * matrix -> dense_matrix
define_method(
    std::shared_ptr<const matrix>,
    times, (const matrix& a, const matrix& b)) {
    return std::make_shared<dense_matrix>();
}

// diagonal_matrix * diagonal_matrix -> diagonal_matrix
define_method(
    std::shared_ptr<const matrix>,
    times, (const diagonal_matrix& a, const diagonal_matrix& b)) {
    return std::make_shared<diagonal_matrix>();
}

Performance

Open methods are almost as fast as ordinary virtual member functions once you turn on optimization (-O2). With both clang and gcc, dispatching a call to a method with one virtual argument takes 15-30% more time than calling the equivalent virtual member function (unless the call goes through a virtual base, which requires a dynamic cast). It does not involve branching or looping, only a few memory reads (which the CPU can be parallelize), a multiplication, a bit shift, a final memory read, then an indirect call. If the body of the method does any amount of work, the difference is unnoticeable.

virtual_ptr, a fat pointer class, can be used to make method dispatch even faster - three instructions and two memory reads.

Examples are available on Compiler Explorer.

Installation

YOMM2 is available on both major package managers. This is the easiest way of integrating it in your project, along with its dependencies. See the vcpkg example and the Conan2 example.

YOMM2 can also be built and installed from the sources, using straight cmake.

First clone the repository:

git clone https://github.com/jll63/yomm2.git

Run cmake:

cmake -S yomm2 -B build.yomm2
cmake --build build.yomm2

If you want to run the tests, specify it when running cmake:

cmake -S yomm2 -B build.yomm2 -DYOMM2_ENABLE_TESTS=1
cmake --build build.yomm2
ctest --test-dir build.yomm2

YOMM2 uses the following Boost libraries:

• Mp11, Preprocessor, DynamicBitset: included by YOMM2 headers
• Test: only used to run the test suite

If you want to run the benchmarks (and in this case you really want a release build):

cmake -S yomm2 -B build.yomm2 -DYOMM2_ENABLE_TESTS=1 -DYOMM2_ENABLE_BENCHMARKS=1 -DCMAKE_BUILD_TYPE=Release
./build.yomm2/tests/benchmarks

The benchmarks use the Google benchmark library.

If you like YOMM2, and you want to install it, either system-wide:

sudo cmake --install build.yomm2

...or to a specific directory:

DESTDIR=/path/to/my/libs cmake --install build.yomm2

This will install the headers and a CMake package configuration. By default, YOMM2 is installed as a headers only library. The examples can be compiled like this (after installation):

clang++ -std=c++17 -O3 examples/synopsis.cpp -o synopsis

Or directly from the repository (i.e. without installing):

clang++ -std=c++17 -O3 -Iinclude examples/synopsis.cpp -o synopsis

The YOMM2 runtime - responsible for building the dispatch tables - adds ~75K to the image, or ~64K after stripping.

The runtime can also be built and installed as a shared library, by adding -DYOMM2_SHARED=1 to the cmake command line.

A CMake package configuration is also installed. If the install location is in CMAKE_PREFIX_PATH, you can use find_package(YOMM2) to locate YOMM2, then target_link_libraries(<your_target> YOMM2::yomm2) to add the necessary include paths and the library. See this example.

Make sure to add the install location to CMAKE_PREFIX_PATH so that you can use find_package(YOMM2) from your including project. For linking, the use target_link_library(<your_target> YOMM2::yomm2). This will automatically add the necessary include directories, so this should be all you need to do to link to yomm2.

Going Further

The documentation is here. Since version 1.3.0, some of the internals are documented, which make it possible to use the library without using macros - see the API tutorial.

YOMM2 has experimental support for writing templatized methods and definitions

• see the templates tutorial.

The library comes with a series of examples:

• The complete matrix example

• The Asteroids example used in Wikipedia's article on Multiple Dispatch

• Process an AST sans clumsy Visitor

• Adventure: a 3-method example

• friendship: an example with namespaces, method containers and friend declarations

I presented the library at CppCon 2018. Here are the video recording and the slides.

Roadmap

YOMM2 has been stable (in the sense of being backward-compatible) for many years, but it is still evolving. Here are the items on which I intend to work in the future. No promises, no time table.

• Dispatch on std::any.
• Static offsets (i.e. set at compile time).
• Static linking of dispatch data.
• Minimal perfect hash tables as an option.
• Multi-threaded hash search.
• Get closer to Stroustrup et al's papers (version 2.0):
  • use compatible return types for disambiguation
  • move support for std::shared_ptr and unique_ptr to an optional header

If you have ideas, comments, suggestions...get in touch! If you use YOMM2, I would appreciate it if you take the time to send me a description of your use case(s), and links to the project(s), if they are publicly available.