Relaxing some constexpr restrictions

Document #: P2448R1
Date: 2022-01-15
Project: Programming Language C++
Audience: CWG
Reply-to: Barry Revzin
<>

1 Abstract

There are two rules about constexpr programming that make code ill-formed or ill-formed (no diagnostic required) when functions or function templates are marked constexpr that might never evaluate to a constant expression. But… so what if they don’t? The goal of this paper is to stop diagnosing problems that don’t exist.

2 Revision History

Since [P2448R0], CWG telecon pointed out that there were several other rules that could be striken in the same theme. Updated wording.

A draft of the first revision of this paper was discussed in an EWG telecon, where the following poll was taken:

send P2448 to electronic polling, targeting CWG for C++23.

SF
F
N
A
SA
10 8 1 0 0

This first published revision thus targets CWG.

3 Maybe Not Now, But Soon

constexpr functions and function templates in C++ generally speaking mean maybe constexpr. Not all instantiations or evaluations must be invocable at compile time, it’s just that there must be at least one set of function arguments in at least one instantiation that works.

And this isn’t just generally speaking, this is enshrined as a rule: 9.2.6 [dcl.constexpr]/6:

For a constexpr function or constexpr constructor that is neither defaulted nor a template, if no argument values exist such that an invocation of the function or constructor could be an evaluated subexpression of a core constant expression, or, for a constructor, an evaluated subexpression of the initialization full-expression of some constant-initialized object ([basic.start.static]), the program is ill-formed, no diagnostic required.

Here is an example of a program that violates this rule:

void f(int& i) {
    i = 0;
}

constexpr void g(int& i) {
    f(i);
}

g unconditionally calls f, which is not a constexpr function, so there does not exist any invocation that would be a constant expression. Ill-formed, no diagnostic required. gcc and msvc both diagnose this.

Now, one could argue that this diagnosis is a good thing: that constexpr annotation on g makes no sense! It can’t be a constant expression, so having that specifier on the function is misleading. Diagnosing this error helps the programmer realize that they were mistaken and they can take steps to change this. That seems compelling enough.

Consider instead this example:

#include <optional>

constexpr void h(std::optional<int>& o) {
    o.reset();
}

Here, h is the exact same kind of function as g: we have a function unconditionally calling another non-constexpr function. At least, that’s true for C++17. It’s won’t be true in C++23, and the answer for C++20 depends on how vendors choose to implement [P2231R1]. Here the answer depends: some functions can easily be constexpr in one standard but not in earlier ones.

Now, the sanctioned way to fix this code is to write:

#include <optional>

#if __cpp_lib_optional >= 202106
constexpr
#endif
void h(std::optional<int>& o) {
    o.reset();
}

This way, h will be constexpr when the appropriate library changes are made, allowing it to be. But… is this better?

I would argue it’s not. The language rules for constexpr have expanded in every language standard since constexpr was introduced. Standard library support will always lag that. Third-party library support likely even more so. So the answer to the question “are there any arguments for which this function can be a constant expression?” can easily be No in C++N but Yes in C++N+1, for a wide variety of functions. Does forcing conscientious library authors to take painstaking care in conditionally marking function constexpr provide value to the ecosystem? I’m skeptical that it does.

Moreover, while it’s possible to write the above for std::optional, I’m not sure that it’s common for other library to provide macros that can be used to mark functions conditionally constexpr like this. Or, indeed, if there is even another such example. So if I’m a consumer of a library that might have some functionality constexpr in one version but more functionality constexpr in the next, I always have to lag.

Such diagnosis may have made sense in the C++11 days, but now that we’re approaching C++23 where more and more things are constexpr and more and more libraries will mark more of their functions constexpr because they can be, it seems strictly better to just reserve diagnosing constexpr violations to the place where we already have to diagnose them: when you write code that must be evaluated at compile time.

Put differently, the current rule is there must be some tuple (function arguments, template arguments) for which a constexpr function invocation is a constant expression. But there’s really another input here: (function arguments, template arguments, version). The version here might be the language version, it might be a bunch of library versions. Ultimately, it’s a question of time. A function may not be constexpr today, but it may be constexpr soon. Diagnosing it as not being constexpr yet seems harmful to the question of evolving code.

4 Sometimes Maybe, Sometimes Always

constexpr usually means maybe constexpr. But sometimes it actually means always constexpr. That case comes up with explicitly defaulted functions. From 9.5.2 [dcl.fct.def.default]/3:

An explicitly-defaulted function that is not defined as deleted may be declared constexpr or consteval only if it is constexpr-compatible ([special], [class.compare.default]). A function explicitly defaulted on its first declaration is implicitly inline ([dcl.inline]), and is implicitly constexpr ([dcl.constexpr]) if it is constexpr-compatible.

Let’s say I’m writing a wrapper class template that I’m intending to be usable during compile time:

template <typename T>
struct Wrapper {
    constexpr Wrapper() = default;
    constexpr Wrapper(Wrapper const&) = default;
    constexpr Wrapper(T const& t) : t(t) { }

    constexpr T get() const { return t; }
    constexpr bool operator==(Wrapper const&) const = default;
private:
    T t;
};

I might take the strategy of just marking every function constexpr. Both for consistency and also as a strategy to avoid forgetting to mark some functions constexpr.

But then I try to use it:

struct X {
    X();
    bool operator==(X const&) const;
};

Wrapper<X> x;

None of this code is trying to evaluate anything during constant evaluation, yet it is already ill-formed. gcc and clang already diagnose at this point. msvc and icc do not, but this rule is a mandatory diagnostic, so they are mistaken. Although, even here, none of the compilers care that I erroneously marked the default constructor constexpr. And none of them care that I erroneously marked the copy constructor constexpr. gcc and clang are only diagnosing operator== here (even though I’m not even using it).

But why do we even have this rule in the first place? Note that if I wrote my equality operator this way:

constexpr bool operator==(Wrapper const& rhs) const {
    return t == rhs.t;
}

That is, manually writing out what the defaulted version does, even with marking it constexpr, this code is perfectly valid C++ code. It doesn’t matter that I’m marking this function constexpr, because in this context, the only requirement is that some instantiation can be a constant expression. But because I want to = default it, suddenly every instantiation has to be a constant expression?

What this means is that the correct way to write my wrapper type is:

template <typename T>
struct Wrapper {
    Wrapper() = default;
    Wrapper(Wrapper const&) = default;
    constexpr Wrapper(T const& t) : t(t) { }

    constexpr T get() const { return t; }
    bool operator==(Wrapper const&) const = default;
private:
    T t;
};

So I have some functions marked constexpr and some not, but all of them are still usable during constant evaluation time (where appropriate based on T). The lack of consistency here is a bit jarring. constexpr functions should always be maybe constexpr, not must be constexpr.

5 Onwards to constexpr classes

We have a proposal in front of us to allow annotating the entire class as being constexpr, to avoid all these extra annotations [P2350R1]. And that proposal currently runs into both issues:

template <typename T>
struct Wrapper constexpr {
    Wrapper() = default;
    Wrapper(Wrapper const&) = default;
    Wrapper(T const& t) : t(t) { }

    void reset() { t.reset(); }
    bool operator==(Wrapper const&) const = default;
private:
    std::optional<T> t;
};

Are all of these functions okay? I would argue that they should all be okay, but per the wording they’re currently not.

I want Wrapper to be entirely constexpr where feasible. Some of those functions may not be constexpr for all types, and that’s fine. Some of these functions may not be able to be constexpr in C++N but may be later, and I don’t want to have to go back and either annotate against this (which I don’t think P2350 even allows room for) or stop using this feature and go back to manually marking and even more annotations.

6 Getting rid of constexpr-compatible

With the wording suggested in the first revision of this paper [P2448R0], functions (even defaulted special member functions) could be declared constexpr without this leading to either a diagnostic or leading to a program being ill-formed, no diagnostic required.

During the CWG telecon discussing that paper, it was brought up that we can go further. For instance, we have a term constexpr-compatible. Currently used by 9.5.2 [dcl.fct.def.default]/3:

3 An explicitly-defaulted function that is not defined as deleted may be declared constexpr or consteval only if it is constexpr-compatible ([special], [class.compare.default]). A function explicitly defaulted on its first declaration is implicitly inline ([dcl.inline]), and is implicitly constexpr ([dcl.constexpr]) if it is constexpr-compatible.

Where special member functions are considered constexpr-compatible when:

The special member case depends on the kind of special member function:

It would be, within the spirit of this paper, to significantly reduce these restrictions as follows. First, we can remove the restrictions on constexpr constructors and destructors(from 9.2.6 [dcl.constexpr]/4 and /5):

4 The definition of a constexpr constructor whose function-body is not = delete shall additionally satisfy the following requirements:

  • (4.1) for a non-delegating constructor, every constructor selected to initialize non-static data members and base class subobjects shall be a constexpr constructor;
  • (4.2) for a delegating constructor, the target constructor shall be a constexpr constructor.

5 The definition of a constexpr destructor whose function-body is not = delete shall additionally satisfy the following requirement:

  • (5.1) for every subobject of class type or (possibly multi-dimensional) array thereof, that class type shall have a constexpr destructor.

After this, if we allow a constexpr copy/move assignment even for non-literal classes, then we can basically make all defaulted functions constexpr except for constructors and destructors for types that have virtual base classes. This also means that we can remove the term constexpr-compatible since we would no longer need to use it anywhere. That’s a nice chunk of specification improvement, removing rules that nobody really needs.

6.1 Going Deeper

We could go one step further and drop further uses of literal type in 9.2.6 [dcl.constexpr]/3:

3 The definition of a constexpr function shall satisfy the following requirements:

  • (3.1) its return type (if any) shall be a literal type;
  • (3.2) each of its parameter types shall be a literal type;
  • (3.3) it shall not be a coroutine;
  • (3.4) if the function is a constructor or destructor, its class shall not have any virtual base classes.

The first two sub-bullets are also very much in the spirit of this paper. A type could be not literal yet in C++N but could become literal in C++N+1, it would be nice if we could simply mark such functions constexpr regardless (as I’ve already noted the desire to do for copy/move assignment in the previous section).

6.2 Going Deeper

Once we eliminate those two bullets, we only have two rules for the requirements of a constexpr function: not a coroutine, and not a constuctor/destructor of a class that has virtual base classes. I’m not entirely sure why we need the latter rule either, but it could also be moved elsewhere – that is, the problem isn’t declaring a constexpr constructor for a class with a virtual base, the problem is trying to initialize such a type during constant evaluation. Similar to how we removed the restriction on try/catch while still disallowing throwing.

This would actually allow further specification cleanup, since now we could just say that all implicit constructors are constexpr. But it’s a much bigger step and I’m not sure that we should take it at this time.

7 Proposal

Strike two bullets from 9.2.6 [dcl.constexpr]/3, as well as paragraphs /4, /5, /6, their examples, and part of 7:

3 The definition of a constexpr function shall satisfy the following requirements:

  • (3.1) its return type (if any) shall be a literal type;
  • (3.2) each of its parameter types shall be a literal type;
  • (3.3) it shall not be a coroutine;
  • (3.4) if the function is a constructor or destructor, its class shall not have any virtual base classes.

[Example 2:

// ...

-end example]

4 The definition of a constexpr constructor whose function-body is not = delete shall additionally satisfy the following requirements:

  • (4.1) for a non-delegating constructor, every constructor selected to initialize non-static data members and base class subobjects shall be a constexpr constructor;
  • (4.2) for a delegating constructor, the target constructor shall be a constexpr constructor.

[Example 3:

struct Length {
  constexpr explicit Length(int i = 0) : val(i) { }
private:
  int val;
};

— end example]

5 The definition of a constexpr destructor whose function-body is not = delete shall additionally satisfy the following requirement:

  • (5.1) for every subobject of class type or (possibly multi-dimensional) array thereof, that class type shall have a constexpr destructor.

6 For a constexpr function or constexpr constructor that is neither defaulted nor a template, if no argument values exist such that an invocation of the function or constructor could be an evaluated subexpression of a core constant expression, or, for a constructor, an evaluated subexpression of the initialization full-expression of some constant-initialized object ([basic.start.static]), the program is ill-formed, no diagnostic required.

[Example 4:

constexpr int f(bool b)
  { return b ? throw 0 : 0; }           // OK
constexpr int f() { return f(true); }   // ill-formed, no diagnostic required

struct B {
  constexpr B(int x) : i(0) { }         // x is unused
  int i;
};

int global;

struct D : B {
  constexpr D() : B(global) { }         // ill-formed, no diagnostic required
                                        // lvalue-to-rvalue conversion on non-constant global
};
  • end example]

7 If the instantiated template specialization of a constexpr function template or member function of a class template would fail to satisfy the requirements for a constexpr function, that specialization is still a constexpr function, even though a call to such a function cannot appear in a constant expression. If no specialization of the template would satisfy the requirements for a constexpr function when considered as a non-template function, the template is ill-formed, no diagnostic required.

Adjust 9.5.2 [dcl.fct.def.default]/3 and fix the example (which is already wrong at the moment, since default-initializing an int during constant evaluation is ok):

3 An explicitly-defaulted function that is not defined as deleted may be declared constexpr or consteval only if it is constexpr-compatible ([special], [class.compare.default]). A function explicitly defaulted on its first declaration is implicitly inline ([dcl.inline]), and is implicitly constexpr ([dcl.constexpr]) if it is constexpr-compatible satisfies the requirements for a constexpr function.

4 [Example 1:

  struct S {
-   constexpr S() = default;              // error: implicit S() is not constexpr
    S(int a = 0) = default;               // error: default argument
    void operator=(const S&) = default;   // error: non-matching return type
    ~S() noexcept(false) = default;       // OK, despite mismatched exception specification
  private:
    int i;
    S(S&);                                // OK: private copy constructor
  };
  S::S(S&) = default;                     // OK: defines copy constructor

  struct T {
    T();
    T(T &&) noexcept(false);
  };
  struct U {
    T t;
    U();
    U(U &&) noexcept = default;
  };
  U u1;
  U u2 = static_cast<U&&>(u1);            // OK, calls std​::​terminate if T​::​T(T&&) throws

end example]

Strike use of constexpr-compatible in 11.4.4 [special]/8:

8 A defaulted special member function is constexpr-compatible if the corresponding implicitly-declared special member function would be a constexpr function.

Mark assignment as being constexpr in 11.4.6 [class.copy.assign]/10:

10 A copy/move assignment operator for a class X that is defaulted and not defined as deleted is implicitly defined when it is odr-used ([term.odr.use]) (e.g., when it is selected by overload resolution to assign to an object of its class type), when it is needed for constant evaluation ([expr.const]), or when it is explicitly defaulted after its first declaration. The implicitly-defined copy/move assignment operator is constexpr. if

  • (10.1) X is a literal type, and
  • (10.2) the assignment operator selected to copy/move each direct base class subobject is a constexpr function, and
  • (10.3) for each non-static data member of X that is of class type (or array thereof), the assignment operator selected to copy/move that member is a constexpr function.

Remove this no-longer-necessary rule in 11.4.7 [class.dtor]/9 (the requirements for constexpr destructor are now just the requirements for constexpr function, which is now covered by the [dcl.fct.def.default] rule):

9 A defaulted destructor is a constexpr destructor if it satisfies the requirements for a constexpr destructor ([dcl.constexpr]).

Strike use of constexpr-compatible in 11.10.1 [class.compare.default]/4:

4 A defaulted comparison function is constexpr-compatible if it satisfies the requirements for a constexpr function ([dcl.constexpr]) and no overload resolution performed when determining whether to delete the function results in a usable candidate that is a non-constexpr function.

[Note 1: This includes the overload resolutions performed:

  • (4.1) for an operator<=> whose return type is not auto, when determining whether a synthesized three-way comparison is defined,
  • (4.2) for an operator<=> whose return type is auto or for an operator==, for a comparison between an element of the expanded list of subobjects and itself, or
  • (4.3) for a secondary comparison operator @, for the expression x @ y.

end note]

8 References

[P2231R1] Barry Revzin. 2021-02-12. Add further constexpr support for optional/variant.
https://wg21.link/p2231r1

[P2350R1] Andreas Fertig. 2021-07-15. constexpr class.
https://wg21.link/p2350r1

[P2448R0] Barry Revzin. 2021-10-14. Relaxing some constexpr restrictions.
https://wg21.link/p2448r0