cryptocode / fluent_args

Experiment: A macro and pattern to get named and fluent C++17 arguments

This is a C++17 experiment to get named parameters in a fluent way. It seems like an idea worth exploring for certain use-cases (while we wait for meta classes so we can do these things correctly.) This is similar to the Named Parameter Idiom, but applied to functions rather than object construction.

Ingredients

• A tiny, optional macro for convenience
• A function implementation pattern for fluent argument passing
• C++17 copy elision guarantees
• [[nodiscard]] combined with -Werror=unused-result so the call operator won't be forgotten.

Example

#include <fluent.hpp>
#include <iostream>

struct MyThing {
    int simple(int num);
    auto complicatedProcessing() -> auto;
};

int MyThing::simple(int num) 
{ 
    return num * 2; 
}

auto MyThing::complicatedProcessing() -> auto {
    struct action {
        fluent_arg(double, wage);
        fluent_arg(int, id);
        fluent_arg(bool, admin);
    
        // This is where your logic goes. You can add statically required (unnamed) arguments if you wish,
        // and any return type will do.
        int operator ()() {
            // In this implementation, we refer to fluent arguments using an underscore
            std::cout << "Processing: " << _id << std::endl;
    
            // More code...
            return 5;
        }
    };
  
    return action();
}

int main() {
    MyThing thing;
    thing.simple(2);
    int result = thing.complicatedProcessing().id(5).admin(true)();

    // nodiscard ensures it's an error to forget the final call operator (assuming -Werror=unused-result):
    // thing.complicatedProcessing().id(5).admin(true);

    // We get partial application for free:
    auto process = thing.complicatedProcessing().id(6);
    // ...time passes
    process = process.admin(false);
    // ...we're ready to process
    result = process();
}

The pattern

The pattern is simple enough: you create an action type to handle argument passing, and you make sure these are all [[nodiscard]]. The fluent_arg macro helps you get that right. The actual logic sits in the call operator. Since we require -Werror=unused-result, you'll get an error if you forget to call it, because all the argument handlers are [[nodiscard]]!

As already stated, the macro is just for convience. In fact, it's sometimes desirable to write the argument-passing function yourself to add logging, validation and so on.

Benefits

• Arguments are named at the call site, especially important for booleans. Great readability.
• Optional arguments by design (which is also a downside bullet point)
• Can hook into an error propagation system, possibly form monadic chains.
• Offers an obvious place to put argument validation.
• Partial application, i.e. add arguments over time as they become available, call function when ready.
• Common actions (parameter sets) can be shared between related functions. Put action into a namespace, make the call operator virtual, then specialize in various related functions who share the parameter set by deriving and overloading operator ()

Drawbacks

• Arguments not in interface, as they're replaced by auto-return. However, IDEs pick up the argument functions for intellisense purposes. Doxygen may or may not be helpful: doxygen/doxygen#5492
• Adds verbosity to the function implementation. May or may not be worth it to get a nice call site/partial application.
• Cannot statically force certain named parameter methods to be called, though unnamed parameters in the call operator is possible. A maybe type to propagate errors may be helpful, such as Boost Outcome.
• Macro-induced fields may be hard to rename/change type of through IDE refactorings. Expanding a macro temporarily to refactor isn't hard though, but definitely inconvenient.