function_ref: a non-owning reference to a Callable

Abstract

This paper proposes the addition of function_ref<R(Args...)> to the Standard Library, a “vocabulary type” for non-owning references to Callable objects.

Table of contents

• Overview
• Motivating example
• Impact on the Standard
• Alternatives
• Polls
• Synopsis
• Specification
• Feature test macro
• Example implementation
• Existing practice
• Possible issues
• Open questions
• Bikeshedding
• Acknowledgments
• References

Overview

Since the advent of C++11 writing more functional code has become easier: functional programming patterns and idioms have become powerful additions to the C++ developer’s toolbox. “Higher-order functions” are one of the key ideas of the functional paradigm - in short, they are functions that take functions as arguments and/or return functions as results.

The need of referring to an existing Callable object comes up often when writing functional C++ code, but the Standard Library unfortunately doesn’t provide a flexible facility that allows to do so. Let’s consider the existing utilities:

• Pointers to functions are only useful when the entity they refer to is stateless (i.e. a non-member function or a capture-less lambda), but they are cumbersome to use otherwise. Fully supporting the Callable concept requires also explicitly dealing with pointers to member functions and pointers to data members.

• std::function seamlessly works with Callable objects, but it’s a “general-purpose polymorphic function wrapper” that may introduce unnecessary overhead and that owns the Callable it stores. std::function is a great choice when an owning type-erased wrapper is required, but it’s often abused when its ownership semantics and its flexibility are not required.

  • Note that when std::function is constructed/assigned with a std::reference_wrapper to a Callable, it has reference semantics.

  • Another limitation of std::function is the fact that the stored Callable must be CopyConstructible.

• Templates can be used to avoid unnecessary costs and to uniformly handle any Callable object, but they are hard to constrain to a particular signature and force code to be defined in headers.

This paper proposes the introduction of a new function_ref class template, which is akin to std::string_view. This paper describes function_ref as a non-owning lightweight wrapper over any Callable object.

Motivating example

Here’s one example use case that benefits from higher-order functions: a retry(n, f) function that attempts to synchronously call f up to n times until success. This example might model the real-world scenario of repeatedly querying a flaky web service.

struct payload { /* ... */ };

// Repeatedly invokes `action` up to `times` repetitions.
// Immediately returns if `action` returns a valid `payload`.
// Returns `std::nullopt` otherwise.
std::optional<payload> retry(std::size_t times, /* ????? */ action);

The passed-in action should be a Callable which takes no arguments and returns std::optional<payload>. Let’s see how retry can be implemented with various techniques:

• Using pointers to functions:

  std::optional<payload> retry(std::size_t times,
                               std::optional<payload>(*action)())
  {
      /* ... */
  }

  (on godbolt.org)

  • Advantages:

    • Easy to implement: no need to use a template or any explicit constraint (e.g. std::enable_if_t<...>). The type of the pointer specifies exactly which functions can be passed, no extra constraints are required.

    • Minimal overhead: no allocations, no exceptions, and action is as big as a pointer.

      • Modern compilers are able to completely inline the call to action, producing optimal assembly.

  • Drawbacks:

    • This technique doesn’t support stateful Callable objects.

• Using a template:

  template <typename F>
  auto retry(std::size_t times, F&& action)
      -> std::enable_if_t<std::is_invocable_r_v<std::optional<payload>, F&&>,
                          std::optional<payload>>
  {
      /* ... */
  }

  (on godbolt.org)

  • Advantages:

    • Supports arbitrary Callable objects, such as stateful closures.

    • Zero-overhead: no allocations, no exceptions, no indirections.

  • Drawbacks:

    • Harder to implement and less readable: users must use std::enable_if_t and std::invocable_r_v to ensure that action’s signature is properly constrained.

    • retry must be defined in a header file. This might be undesiderable when trying to minimize compilation times.

• Using std::function:

  std::optional<payload> retry(std::size_t times,
                               std::function<std::optional<payload>()> action)
  {
      /* ... */
  }

  (on godbolt.org)

  • Advantages:

    • Supports arbitrary Callable objects, such as stateful closures.

    • Easy to implement: no need to use a template or any explicit constraint. The type fully constrains what can be passed.

  • Drawbacks:

    • Unclear ownership semantics: action might either own the the stored Callable, or just refer to an existing Callable if initialized with a std::reference_wrapper.

    • Can potentially have significant overhead:

      • Even though the implementation makes use of SBO (small buffer optimization), std::function might allocate if the stored object is large enough. This requires one extra branch on construction/assignment, one potential dynamic allocation, and makes action as big as the size of the internal buffer.

      • If the implementation doesn’t make use of SBO, std::function will always allocate on construction/assignment.

      • Modern compilers are not able to inline std::function, often resulting in very poor assembly compared to the previously mentioned techniques.

    • Mandatory use of exceptions: std::function might throw if an allocation fails, and throws std::bad_function_call if it’s invoked while unset.

• Using the proposed function_ref:

  std::optional<payload> retry(std::size_t times,
                               function_ref<std::optional<payload>()> action)
  {
      /* ... */
  }

  (on godbolt.org)

  • Advantages:

    • Supports arbitrary Callable objects, such as stateful closures.

    • Easy to implement: no need to use a template or any constraint. The type fully constrains what can be passed.

    • Clear ownership semantics: action is a non-owning reference to an existing Callable.

    • Small overhead: no allocations, no exceptions, and action is as big as two pointers.

      • Modern compilers are able to completely inline the call to action, producing optimal assembly.

Impact on the Standard

This proposal is a pure library extension. It does not require changes to any existing part of the Standard.

Alternatives

The only existing viable alternative to function_ref currently is std::function + std::reference_wrapper. The Standard guarantees that when a std::reference_wrapper is used to construct/assign to a std::function no allocations will occur and no exceptions will be thrown.

Using std::function for non-owning references is suboptimal for various reasons.

1. The ownership semantics of a std::function are unclear - they change depending on whether or not the std::function was constructed/assigned with a std::reference_wrapper.

   void foo(std::function<void()> f);
   // `f` could be referring to an existing Callable, or could own one.
   
   void bar(function_ref<void()> f);
   // `f` unambiguously is a non-owning reference to an existing Callable.

2. This technique doesn’t work with temporaries. This is a huge drawback as it prevents stateful temporary lambdas from being passed as callbacks.

   void foo(std::function<void()> f);
   
   int main()
   {
       int x = 0;
       foo(std::ref([&x]{ ++x; }); // does not compile
   }

   (on godbolt.org)

   The code above doesn’t compile, as std::ref only accepts non-const lvalue references (additionally, std::cref is explicitly deleted for rvalue references). Avoiding the use of std::ref breaks the guarantee that f won’t allocate or throw an exception on construction.

3. std::function is harder for compilers to optimize compared to the proposed function_ref. This is true due to various reasons:

   • std::function can allocate and/or throw exceptions on construction and/or assigment.

   • std::function might use SBO, which could require an additional branch during construction/assignment, make inlining more difficult, and unnecessarily increase memory usage.

   Rough benchmarks comparing the generated assembly of a std::function parameter and a function_ref parameter against a template parameter show that:

   • std::function, on average, generates approximately 5x more assembly than a template parameter.

   • function_ref, on average, generates approximately 1.5x more assembly than a template parameter.

   A description of the benchmarking techniques used and the full results can be found on my article “passing functions to functions” ¹.

Polls

P0792R1

Semantics: pointer versus reference

option 1

function_ref, non-nullable, not default constructible

option 2

function_ptr, nullable, default constructible

We want 1 and 2

SF F N A SA

1 2 8 3 6

ref vs ptr

SR R N P SP

6 5 2 5 0

The poll above clearly shows that the desired direction for function_ref is towards a non nullable, non default-constructible reference type. This revision (P0792R2) removes the “empty state” and default constructibility from the proposed function_ref. If those semantics are required by users, they can trivially wrap function_ref into an std::optional<function_ref</* ... */>>.

target and target_type

We want target and target-type (consistent with std::function) if they have no overhead

Unanimous consent

We want target and target-type (consistent with std::function) even though they have overhead

SF F N A SA

0 0 1 9 4

Unfortunately target and target_type cannot be implemented without overhead. Firstly, target is defined in terms of target_type ², so our attention can be focused on target_type.

std::function::target_type is defined as follows:

const std::type_info& target_type() const noexcept;

Returns: typeid(T) if the stored function has type T, otherwise typeid(void).

This polymorphic behavior requires type erasure of the stored function. As the only type erasure occurring in the proposed function_ref is of the referenced callable’s operator(), providing a way of retrieving the type of the erased callable would require additional overhead (e.g. virtual target_type function).

Alternatively, target_type could be not included and target could simply statically check whether the passed type is the same as Signature. E.g.:

template <typename T>
const T* function_ref<Signature>::target() const
{
    if constexpr(std::is_same_v<signature_of_t<T>, Signature>)
    {
        return /* address of referenced function */;
    }
    else
    {
        return nullptr;
    }
}

Synopsis

namespace std
{
    template <typename>
    class function_ref; /* undefined */

    template <typename Signature>
    class function_ref<Signature>
    {
    public:
        constexpr function_ref(const function_ref&) noexcept;

        template <typename F>
        constexpr function_ref(F&&) noexcept;

        constexpr function_ref& operator=(const function_ref&) noexcept;

        template <typename F>
        constexpr function_ref& operator=(F&&) noexcept;

        constexpr void swap(function_ref&) noexcept;

        R operator()(Args...) /* qualifiers */;
        // `R`, `Args...`, and `qualifiers` are the return type, the parameter-type-list,
        // and the sequence "cv-qualifier-seq-opt noexcept-specifier-opt" of the function
        // type `Signature`, respectively.
    };

    template <typename Signature>
    constexpr void swap(function_ref<Signature>&, function_ref<Signature>&) noexcept;

    template <typename R, typename... Args>
    function_ref(R (*)(Args...)) -> function_ref<R(Args...)>;

    template <typename R, typename... Args>
    function_ref(R (*)(Args...) noexcept) -> function_ref<R(Args...) noexcept>;
}

Specification

template <typename Signature>
constexpr function_ref<Signature>::function_ref(const function_ref& rhs) noexcept;

• Effects: constructs a function_ref referring to the same callable rhs refers to.

template <typename Signature>
template <typename F>
constexpr function_ref<Signature>::function_ref(F&& f) noexcept;

• Remarks: This constructor shall not participate in overload resolution unless !std::is_same_v<std::decay_t<F>, function_ref> and:

  • If Signature is marked noexcept: std::is_nothrow_invocable_r_v<R, /* cv-qualifiers */ F&, Args...>;

  • Otherwise: std::is_invocable_r_v<R, /* cv-qualifiers */ F&, Args...>.

  Where R, Args..., and cv-qualifiers are the return type, the parameter-type-list, and the sequence “cv-qualifier-seq-opt” of the function type Signature, respectively.

• Effects: constructs a function_ref referring to f.

template <typename Signature>
constexpr function_ref& function_ref<Signature>::operator=(const function_ref& rhs) noexcept;

• Effects: *this refers to the same callable rhs refers to.

• Returns: *this.

template <typename Signature>
template <typename F>
constexpr function_ref& function_ref<Signature>::operator=(F&&) noexcept;

• Requires: !std::is_same_v<std::decay_t<F>, function_ref> and:

  • If Signature is marked noexcept: std::is_nothrow_invocable_r_v<R, /* cv-qualifiers */ F&, Args...>;

  • Otherwise: std::is_invocable_r_v<R, /* cv-qualifiers */ F&, Args...>.

• Effects: *this refers to f.

• Returns: *this.

template <typename Signature>
constexpr void function_ref<Signature>::swap(function_ref& rhs) noexcept;

• Effects: exchanges the values of *this and rhs.

template <typename Signature>
R function_ref<Signature>::operator()(Args... xs) /* qualifiers */;

• Effects: equivalent to return INVOKE<R>(f, std::forward<Args>(xs)...);, where f is the callable object referred to by *this, qualified with the same cv-qualifiers as the function type Signature.

template <typename Signature>
constexpr void swap(function_ref<R(Args...)>& lhs, function_ref<R(Args...)>& rhs) noexcept;

• Effects: equivalent to lhs.swap(rhs).

Feature test macro

I propose the feature-testing macro name __cpp_lib_function_ref.

Example implementation

An example implementation is available here on GitHub.

Existing practice

Many facilities similar to function_ref exist and are widely used in large codebases. Here are some examples:

• The llvm::function_ref ³ class template is used throughout LLVM. A quick GitHub search on the LLVM organization reports hundreds of usages both in llvm and clang ⁴.

• Facebook’s Folly libraries ⁵ provide a folly::FunctionRef ⁶ class template. A GitHub search shows that it’s used in projects proxygen and fbthrift ⁷.

• GNU’s popular debugger, gdb ⁸, uses gdb::function_view ⁹ throughout its code base. The documentation in the linked header file ¹⁰ is particularly well-written and greatly motivates the need for this facility.

Additionally, combining results from GitHub searches (excluding “llvm” and “folly”) for “function_ref” ¹¹, “function_view” ¹², “FunctionRef” ¹³, and “FunctionView” ¹⁴ roughly shows more than 2800 occurrences.

Possible issues

Accepting temporaries in function_ref’s constructor is extremely useful in the most common use case: using it as a function parameter:

void foo(function_ref<void()>);

int main()
{
    foo([]{ });
}

The usage shown above is completely safe: the temporary closure generated by the lambda expression is guarantee to live for the entirety of the call to foo. Unfortunately, this also means that the following code snippet will result in undefined behavior:

int main()
{
    function_ref<void()> f{[]{ }};
    // ...
    f(); // undefined behavior
}

The above closure is a temporary whose lifetime ends after the function_ref constructor call. The function_ref will store an address to a “dead” closure - invoking it will produce undefined behavior ¹⁵. As an example, AddressSanitizer detects an invalid memory access in this gist ¹⁶. Note that this problem is not unique to function_ref: the recently standardized std::string_view ¹⁷ has the same problem ¹⁸.

I strongly believe that accepting temporaries is a “necessary evil” for both function_ref and std::string_view, as it enables countless valid use cases. The problem of dangling references has been always present in the language - a more general solution like Herb Sutter and Neil Macintosh’s lifetime tracking ¹⁹ would prevent mistakes without limiting the usefulness of view/reference classes.

Open questions

Below are some unanswered questions for which I kindly ask guidance from members of the commitee and readers of this paper.

• function_ref::operator() is not currently marked as constexpr due to implementation issues. I could not figure a way to implement a constexpr-friendly operator(). Is there any possibility it could be marked as constexpr to increase the usefulness of function_ref?

• function_ref<Signature>’s signature currently only accepts any combination of const and noexcept. Should this be extended to include ref-qualifiers? This would mean that function_ref::operator() would first cast the referenced callable to either an lvalue reference or rvalue reference (depending on Signature’s ref qualifiers) before invoking it. See P0045R1 ²⁰ and N4159 ²¹) for additional context.

• Constructing a std::function<Signature> from a function_ref<Signature> is completely different from constructing a std::string from a std::string_view: the latter does actually create a copy while the former remains a reference. It may be reasonable to prevent implicit conversions from function_ref to std::function in order to avoid surprising dangerous behavior.

Bikeshedding

The name function_ref is subject to bikeshedding. Here are some other potential names:

• function_view

• callable_ref

• callable_view

• invocable_ref

• invocable_view

• fn_view

• fn_ref

Acknowledgments

Thanks to Agustín Bergé, Dietmar Kühl, Eric Niebler, Tim van Deurzen, and Alisdair Meredith for providing very valuable feedback on earlier drafts of this proposal.

References