Document #: | P0847R5 |
Date: | 2020-10-14 |
Project: | Programming Language C++ |
Audience: |
EWG => CWG |
Reply-to: |
Gašper Ažman <gasper.azman@gmail.com> Sy Brand <simon.brand@microsoft.com> Ben Deane, ben at elbeno dot com <ben@elbeno.com> Barry Revzin <barry.revzin@gmail.com> |
We propose a new mechanism for specifying or deducing the value category of the expression that a member-function is invoked on. In other words, a way to tell from within a member function whether the expression it’s invoked on is an lvalue or an rvalue; whether it is const or volatile; and the expression’s type.
Wording and Implementation. Discussion about implicit vs explicit invocation and interaction with static functions.
The feedback from Belfast in EWG was “This looks good, come back with wording and implementation”. This version adds wording, the implementation is in the works.
[P0847R2] was presented in Kona in Jaunary 2019 to EWGI, with generally enthusiastic support.
This version adds:
[P0847R1] was presented in San Diego in November 2018 with a wide array of syntaxes and name lookup options. Discussion there revealed some potential issues with regards to lambdas that needed to be ironed out. This revision zeroes in on one specific syntax and name lookup semantic which solves all the use-cases.
[P0847R0] was presented in Rapperswil in June 2018 using a syntax adjusted from the one used in that paper, using this Self&& self
to indicate the explicit object parameter rather than the Self&& this self
that appeared in r0 of our paper.
EWG strongly encouraged us to look in two new directions:
This revision carefully explores both of these directions, presents different syntaxes and lookup schemes, and discusses in depth multiple use cases and how each syntax can or cannot address them.
In C++03, member functions could have cv-qualifications, so it was possible to have scenarios where a particular class would want both a const
and non-const
overload of a particular member. (Note that it was also possible to want volatile
overloads, but those are less common and thus are not examined here.) In these cases, both overloads do the same thing — the only difference is in the types being accessed and used. This was handled by either duplicating the function while adjusting types and qualifications as necessary, or having one overload delegate to the other. An example of the latter can be found in Scott Meyers’s “Effective C++” [EffCpp], Item 3:
class TextBlock {
public:
char const& operator[](size_t position) const {
// ...
return text[position];
}
char& operator[](size_t position) {
return const_cast<char&>(
static_cast<TextBlock const&>(*this)[position]
);
}
// ...
};
Arguably, neither duplication nor delegation via const_cast
are great solutions, but they work.
In C++11, member functions acquired a new axis to specialize on: ref-qualifiers. Now, instead of potentially needing two overloads of a single member function, we might need four: &
, const&
, &&
, or const&&
. We have three approaches to deal with this:
One example of the latter might be the overload set for optional<T>::value()
, implemented as:
This is far from a complicated function, but essentially repeating the same code four times — or using artificial delegation to avoid doing so — begs a rewrite. Unfortunately, it’s impossible to improve; we must implement it this way. It seems we should be able to abstract away the qualifiers as we can for non-member functions, where we simply don’t have this problem:
template <typename T>
class optional {
// ...
template <typename Opt>
friend decltype(auto) value(Opt&& o) {
if (o.has_value()) {
return forward<Opt>(o).m_value;
}
throw bad_optional_access();
}
// ...
};
All four cases are now handled with just one function… except it’s a non-member function, not a member function. Different semantics, different syntax, doesn’t help.
There are many cases where we need two or four overloads of the same member function for different const
- or ref-qualifiers. More than that, there are likely additional cases where a class should have four overloads of a particular member function but, due to developer laziness, doesn’t. We think that there are enough such cases to merit a better solution than simply “write it, write it again, then write it two more times.”
We propose a new way of declaring non-static member functions that will allow for deducing the type and value category of the class instance parameter while still being invocable with regular member function syntax. This is a strict extension to the language.
We believe that the ability to write cv-ref qualifier-aware member function templates without duplication will improve code maintainability, decrease the likelihood of bugs, and make fast, correct code easier to write.
The proposal is sufficiently general and orthogonal to allow for several new exciting features and design patterns for C++:
These are explored in detail in the examples section.
This proposal assumes the existence of two library additions, though it does not propose them:
like_t
, a metafunction that applies the cv- and ref-qualifiers of the first type onto the second (e.g. like_t<int&, double>
is double&
, like_t<X const&&, Y>
is Y const&&
, etc.)forward_like
, a version of forward
that is intended to forward a variable not based on its own type but instead based on some other type. forward_like<T>(u)
is short-hand for forward<like_t<T,decltype(u)>>(u)
.The proposed syntax in this paper is to use an explicit this
-annotated parameter.
A non-static member function can be declared to take as its first parameter an explicit object parameter, denoted with the prefixed keyword this
. Once we elevate the object parameter to a proper function parameter, it can be deduced following normal function template deduction rules:
struct X {
void foo(this X const& self, int i);
template <typename Self>
void bar(this Self&& self);
};
struct D : X { };
void ex(X& x, D const& d) {
x.foo(42); // 'self' is bound to 'x', 'i' is 42
x.bar(); // deduces Self as X&, calls X::bar<X&>
move(x).bar(); // deduces Self as X, calls X::bar<X>
d.foo(17); // 'self' is bound to 'd'
d.bar(); // deduces Self as D const&, calls X::bar<D const&>
}
Member functions with an explicit object parameter cannot be static
or have cv- or ref-qualifiers.
A call to a member function will interpret the object argument as the first (this
-annotated) parameter to it; the first argument in the parenthesized expression list is then interpreted as the second parameter, and so forth.
Following normal deduction rules, the template parameter corresponding to the explicit object parameter can deduce to a type derived from the class in which the member function is declared, as in the example above for d.bar()
).
We can use this syntax to implement optional::value()
and optional::operator->()
in just two functions instead of the current six:
template <typename T>
struct optional {
template <typename Self>
constexpr auto&& value(this Self&& self) {
if (!self.has_value()) {
throw bad_optional_access();
}
return forward<Self>(self).m_value;
}
template <typename Self>
constexpr auto operator->(this Self&& self) {
return addressof(self.m_value);
}
};
This syntax can be used in lambdas as well, with the this
-annotated parameter exposing a way to refer to the lambda itself in its body:
vector captured = {1, 2, 3, 4};
[captured](this auto&& self) -> decltype(auto) {
return forward_like<decltype(self)>(captured);
}
[captured]<class Self>(this Self&& self) -> decltype(auto) {
return forward_like<Self>(captured);
}
The lambdas can either move or copy from the capture, depending on whether the lambda is an lvalue or an rvalue.
What follows is a description of how deducing this
affects all important language constructs — name lookup, type deduction, overload resolution, and so forth.
In C++17, name lookup includes both static and non-static member functions found by regular class lookup when invoking a named function or an operator, including the call operator, on an object of class type. Non-static member functions are treated as if there were an implicit object parameter whose type is an lvalue or rvalue reference to cv X
(where the reference and cv qualifiers are determined based on the function’s own qualifiers) which binds to the object on which the function was invoked.
For non-static member functions using an explicit object parameter, lookup will work the same way as other member functions in C++17, with one exception: rather than implicitly determining the type of the object parameter based on the cv- and ref-qualifiers of the member function, these are now explicitly determined by the provided type of the explicit object parameter. The following examples illustrate this concept.
C++17 | Proposed |
---|---|
Name lookup on an expression like obj.foo()
in C++17 would find both overloads of foo
in the first column, with the non-const overload discarded should obj
be const.
With the proposed syntax, obj.foo()
would continue to find both overloads of foo
, with identical behaviour to C++17.
The only change in how we look up candidate functions is in the case of an explicit object parameter, where the argument list is shifted by one. The first listed parameter is bound to the object argument, and the second listed parameter corresponds to the first argument of the call expression.
This paper does not propose any changes to overload resolution but merely suggests extending the candidate set to include non-static member functions and member function templates written in a new syntax. Therefore, given a call to x.foo()
, overload resolution would still select the first foo()
overload if x
is not const
and the second if it is.
The behaviors of the two columns are exactly equivalent as proposed.
The only change as far as candidates are concerned is that the proposal allows for deduction of the object parameter, which is new for the language.
Since in some cases there are multiple ways to declare the same function, it would be ill-formed to declare two functions with the same parameters and the same qualifiers for the object parameter. This is:
struct X {
void bar() &&;
void bar(this X&&); // error: same this parameter type
static void f();
void f(this X const&); // error: two functions taking no parameters
};
But as long as any of the qualifiers are different, it is fine:
The rule in question is 12.2 [over.load]/2.2, and is extended in the wording below.
One of the main motivations of this proposal is to deduce the cv-qualifiers and value category of the class object, which requires that the explicit member object or type be deducible from the object on which the member function is invoked.
If the type of the object parameter is a template parameter, all of the usual template deduction rules apply as expected:
struct X {
template <typename Self>
void foo(this Self&&, int);
};
struct D : X { };
void ex(X& x, D& d) {
x.foo(1); // Self=X&
move(x).foo(2); // Self=X
d.foo(3); // Self=D&
}
It’s important to stress that deduction is able to deduce a derived type, which is extremely powerful. In the last line, regardless of syntax, Self
deduces as D&
. This has implications for name lookup within member functions, and leads to a potential template argument deduction extension.
this
But what if the explicit type does not have reference type? What should this mean?
struct less_than {
template <typename T, typename U>
bool operator()(this less_than, T const& lhs, U const& rhs) {
return lhs < rhs;
}
};
less_than{}(4, 5);
Clearly, the parameter specification should not lie, and the first parameter (less_than{}
) is passed by value.
Following the proposed rules for candidate lookup, the call operator here would be a candidate, with the object parameter binding to the (empty) object and the other two parameters binding to the arguments. Having a value parameter is nothing new in the language at all — it has a clear and obvious meaning, but we’ve never been able to take an object parameter by value before. For cases in which this might be desirable, see by-value member functions.
So far, we’ve only considered how member functions with explicit object parameters are found with name lookup and how they deduce that parameter. Now we move on to how the bodies of these functions actually behave.
Since the explicit object parameter is deduced from the object on which the function is called, this has the possible effect of deducing derived types. We must carefully consider how name lookup works in this context.
struct B {
int i = 0;
template <typename Self> auto&& f1(this Self&&) { return i; }
template <typename Self> auto&& f2(this Self&&) { return this->i; }
template <typename Self> auto&& f3(this Self&&) { return forward_like<Self>(*this).i; }
template <typename Self> auto&& f4(this Self&&) { return forward<Self>(*this).i; }
template <typename Self> auto&& f5(this Self&& self) { return forward<Self>(self).i; }
};
struct D : B {
// shadows B::i
double i = 3.14;
};
The question is, what do each of these five functions do? Should any of them be ill-formed? What is the safest option?
We believe that there are three approaches to choose from:
If there is an explicit object parameter, this
is inaccessible, and each access must be through self
. There is no implicit lookup of members through this
. This makes f1
through f4
ill-formed and only f5
well-formed. However, while B().f5()
returns a reference to B::i
, D().f5()
returns a reference to D::i
, since self
is a reference to D
.
If there is an explicit object parameter, this
is accessible and points to the base subobject. There is no implicit lookup of members; all access must be through this
or self
explicitly. This makes f1
ill-formed. f2
would be well-formed and always return a reference to B::i
. Most importantly, this
would be dependent if the explicit object parameter was deduced. this->i
is always going to be an int
but it could be either an int
or an int const
depending on whether the B
object is const. f3
would always be well-formed and would be the correct way to return a forwarding reference to B::i
. f4
would be well-formed when invoked on B
but ill-formed if invoked on D
because of the requested implicit downcast. As before, f5
would be well-formed.
this
is always accessible and points to the base subobject; we allow implicit lookup as in C++17. This is mostly the same as the previous choice, except that now f1
is well-formed and exactly equivalent to f2
.
Following discussion in San Diego, the option we are proposing is #1. This allows for the clearest model of what a this
-annotated function is: it is a static
member function that offers a more convenient function call syntax. There is no implicit this
in such functions, the only mention of this
would be the annotation on the object parameter. All member access must be done directly through the object parameter.
The consequence of such a choice is that we will need to defend against the object parameter being deduced to a derived type. To ensure that f5()
above is always returning a reference to B::i
, we would need to write one of the following:
template <typename Self>
auto&& f5(this Self&& self) {
// explicitly cast self to the appropriately qualified B
// note that we have to cast self, not self.i
return static_cast<like_t<Self, B>&&>(self).i;
// use the explicit subobject syntax. Note that this is always
// an lvalue reference - not a forwarding reference
return self.B::i;
// use the explicit subobject syntax to get a forwarding reference
return forward<Self>(self).B::i;
}
The worst case for this proposal is the case where we do not intend on deducing a derived object - we only mean to deduce the qualifiers - but that derived type inherits from us privately and shadows one of our members:
class B {
int i;
public:
template <typename Self>
auto&& get(this Self&& self) {
// see above: we need to mitigate against shadowing
return forward<Self>(self).B::i;
}
};
class D : private B {
double i;
public:
using B::get;
};
D().get(); // error
In this example, Self
deduces as D
(not B
), but our choice of shadowing mitigation will not work - we cannot actually access B::i
from a D
because that inheritance is private!
However, we don’t have to rely on D
to friend B
to get this to work. There actually is a way to get this to work correctly and safely. C-style casts get a bad rap, but they are actually the solution here:
class B {
int i;
public:
template <typename Self>
auto&& get(this Self&& self) {
return ((like_t<Self, B>&&)self).i;
}
};
class D : private B {
double i;
public:
using B::get;
};
D().get(); // error
No access checking for the win.
As described in the previous section, the model for a member function with an explicit object parameter is a static
member function.
In other words, given:
While the type of &Y::f
is int(Y::*)(int, int) const&
, the type of &Y::g
is int(*)(Y const&, int, int)
. As these are just function pointers, the usage of these two member functions differs once we drop them to pointers:
Y y;
y.f(1, 2); // ok as usual
y.g(3, 4); // ok, this paper
auto pf = &Y::f;
pf(y, 1, 2); // error: pointers to member functions are not callable
(y.*pf)(1, 2); // okay, same as above
std::invoke(pf, y, 1, 2); // ok
auto pg = &Y::g;
pg(y, 3, 4); // okay, same as above
(y.*pg)(3, 4); // error: pg is not a pointer to member function
std::invoke(pg, y, 3, 4); // ok
The rules are the same when deduction kicks in:
Types are as follows:
&B::foo<B>
is void(*)(B&&)
&B::foo<B const&>
is void(*)(B const&)
&D::foo<B>
is void(*)(B&&)
&B::foo<D>
is void(*)(D&&)
This is exactly what happens if foo
is a normal function.
By-value object parameters give you pointers to function in just the same way, the only difference being that the first parameter being a value parameter instead of a reference parameter:
The type of &less_than<int>::operator()
is bool(*)(less_than<int>, int const&, int const&)
and follows the usual rules of invocation:
less_than<int> lt;
auto p = &less_than<int>::operator();
lt(1, 2); // ok
p(lt, 1, 2); // ok
(lt.*p)(1, 2); // error: p is not a pointer to member function
invoke(p, lt, 1, 2); // ok
It is important to mention the pathological cases. First, what happens if D
is incomplete but becomes valid later?
Following the precedent of [P0929R2], we think this should be fine, albeit strange. If D
is incomplete, we simply postpone checking until the point where we actually need a complete type, as usual. At that point D().foo()
would be a valid expression. We see no reason to reject.
For unrelated complete classes or non-classes:
These are even more unlikely to be actually useful code. In this example, B
is neither convertible to A
nor int
, so neither of these functions is even invocable using normal member syntax. However, they’re still static member functions, so B::bar(42)
is a valid call.
We think these declarations can best be left for compilers to warn about if they so choose, rather than coming up with a language rule to reject them.
Another interesting case, courtesy of Jens Maurer:
struct D;
struct B {
int f1(this D);
};
struct D1 : B { };
struct D2 : B { };
struct D : D1, D2 { };
int x = D().f1(); // error: ambiguous lookup
int y = B().f1(); // error: B is not implicitly convertible to D
auto z = &B::f1; // ok
z(D()); // ok
B::f1(D{}); // ok
Even though both D().f1()
and B().f1()
are ill-formed, for entirely different reasons, taking a pointer to &B::f1
is acceptable — its type is int(*)(D)
— and that function pointer can be invoked with a D
. Actually invoking this function does not require any further name lookup or conversion because by-value member functions do not have an implicit object parameter in this syntax (see by-value this
). The same reasoning holds for the direct function invocation.
Again, we’re not sure if these formulations are actually useful. More so that they don’t seem harmful and attempting to reject these cases may accidentally reject useful ones.
Explicitly naming the object as the this
-designated first parameter fits within many programmers’ mental models of the this
pointer being the first parameter to member functions “under the hood” and is comparable to its usage in other languages, e.g. Python and Rust. It also works as a more obvious way to teach how std::bind
, std::thread
, std::function
, and others work with a member function pointer by making the pointer explicit.
As such, we do not believe there to be any teachability problems.
static
member functions have an explicit object type?No. Static member functions currently do not have an implicit object parameter, and therefore have no reason to provide an explicit one.
[this]
and [*this]
in lambdasInteroperability is perfect, since they do not impact the meaning of this
in a function body. The introduced identifier self
can then be used to refer to the lambda instance from the body.
The proposed syntax has no parsings issue that we are aware of.
There are two programmatic issues with this proposal that we are aware of:
Inadvertently referencing a shadowing member of a derived object in a base class this
-annotated member function. There are some use cases where we would want to do this on purposes (see crtp), but for other use-cases the programmer will have to be aware of potential issues and defend against them in a somewhat verobse way.
Because there is no way to just deduce const
vs non-const
, the only way to deduce the value category would be to take a forwarding reference. This means that potentially we create four instantiations when only two would be minimally necessary to solve the problem. But deferring to a templated implementation is an acceptable option and has been improved by no longer requiring casts. We believe that the problem is minimal.
What follows are several examples of the kinds of problems that can be solved using this proposal.
This proposal can de-duplicate and de-quadruplicate a large amount of code. In each case, the single function is only slightly more complex than the initial two or four, which makes for a huge win. What follows are a few examples of ways to reduce repeated code.
This particular implementation of optional
is Simon’s, and can be viewed on GitHub. It includes some functions proposed in [P0798R0], with minor changes to better suit this format:
There are a few more functions in [P0798R0] responsible for this explosion of overloads, so the difference in both code and clarity is dramatic.
For those that dislike returning auto in these cases, it is easy to write a metafunction matching the appropriate qualifiers from a type. It is certainly a better option than blindly copying and pasting code, hoping that the minor changes were made correctly in each case.
Today, a common design pattern is the Curiously Recurring Template Pattern. This implies passing the derived type as a template parameter to a base class template as a way of achieving static polymorphism. If we wanted to simply outsource implementing postfix incrementation to a base, we could use CRTP for that. But with explicit objects that already deduce to the derived objects, we don’t need any curious recurrence — we can use standard inheritance and let deduction do its thing. The base class doesn’t even need to be a template:
C++17 | Proposed |
---|---|
The proposed examples aren’t much shorter, but they are certainly simpler by comparison.
Once we start to do any more with CRTP, complexity quickly increases, whereas with this proposal, it stays remarkably low.
Let’s say we have a builder that does multiple things. We might start with:
struct Builder {
Builder& a() { /* ... */; return *this; }
Builder& b() { /* ... */; return *this; }
Builder& c() { /* ... */; return *this; }
};
Builder().a().b().a().b().c();
But now we want to create a specialized builder with new operations d()
and e()
. This specialized builder needs new member functions, and we don’t want to burden existing users with them. We also want Special().a().d()
to work, so we need to use CRTP to conditionally return either a Builder&
or a Special&
:
The code on the right is dramatically easier to understand and therefore more accessible to more programmers than the code on the left.
But wait! There’s more!
What if we added a super-specialized builder, a more special form of Special
? Now we need Special
to opt-in to CRTP so that it knows which type to pass to Builder
, ensuring that everything in the hierarchy returns the correct type. It’s about this point that most programmers would give up. But with this proposal, there’s no problem!
The code on the right is much easier in all contexts. There are so many situations where this idiom, if available, would give programmers a better solution for problems that they cannot easily solve today.
Note that the Super
implementations with this proposal opt-in to further derivation, since it’s a no-brainer at this point.
The explicit object parameter syntax offers an alternative solution to implementing a recursive lambda as compared to [P0839R0], since now we’ve opened up the possibility of allowing a lambda to reference itself. To do this, we need a way to name the lambda.
// as proposed in P0839
auto fib = [] self (int n) {
if (n < 2) return n;
return self(n-1) + self(n-2);
};
// this proposal
auto fib = [](this auto self, int n) {
if (n < 2) return n;
return self(n-1) + self(n-2);
};
This works by following the established rules. The call operator of the closure object can also have an explicit object parameter, so in this example, self
is the closure object.
In San Diego, issues of implementability were raised. The proposal ends up being implementable. See the lambda FAQ entry for details.
Combine this with the new style of mixins allowing us to automatically deduce the most derived object, and you get the following example — a simple recursive lambda that counts the number of leaves in a tree.
struct Leaf { };
struct Node;
using Tree = variant<Leaf, Node*>;
struct Node {
Tree left;
Tree right;
};
int num_leaves(Tree const& tree) {
return visit(overload( // <-----------------------------------+
[](Leaf const&) { return 1; }, // |
[](this auto const& self, Node* n) -> int { // |
return visit(self, n->left) + visit(self, n->right); // <----+
}
), tree);
}
In the calls to visit
, self
isn’t the lambda; self
is the overload
wrapper. This works straight out of the box.
This section presents some of the cases for by-value member functions.
Say you wanted to provide a .sorted()
method on a data structure. Such a method naturally wants to operate on a copy. Taking the parameter by value will cleanly and correctly move into the parameter if the original object is an rvalue without requiring templates.
struct my_vector : vector<int> {
auto sorted(this my_vector self) -> my_vector {
sort(self.begin(), self.end());
return self;
}
};
It’s been established that if you want the best performance, you should pass small types by value to avoid an indirection penalty. One such small type is std::string_view
. Abseil Tip #1 for instance, states:
Unlike other string types, you should pass
string_view
by value just like you would anint
or adouble
becausestring_view
is a small value.
There is, however, one place today where you simply cannot pass types like string_view
by value: to their own member functions. The implicit object parameter is always a reference, so any such member functions that do not get inlined incur a double indirection.
As an easy performance optimization, any member function of small types that does not perform any modifications can take the object parameter by value. Here is an example of some member functions of basic_string_view
assuming that we are just using charT const*
as iterator
:
template <class charT, class traits = char_traits<charT>>
class basic_string_view {
private:
const_pointer data_;
size_type size_;
public:
constexpr const_iterator begin(this basic_string_view self) {
return self.data_;
}
constexpr const_iterator end(this basic_string_view self) {
return self.data_ + self.size_;
}
constexpr size_t size(this basic_string_view self) {
return self.size_;
}
constexpr const_reference operator[](this basic_string_view self, size_type pos) {
return self.data_[pos];
}
};
Most of the member functions can be rewritten this way for a free performance boost.
The same can be said for types that aren’t only cheap to copy, but have no state at all. Compare these two implementations of less_than
:
C++17 | Proposed |
---|---|
In C++17, invoking less_than()(x, y)
still requires an implicit reference to the less_than
object — completely unnecessary work when copying it is free. The compiler knows it doesn’t have to do anything. We want to pass less_than
by value here. Indeed, this specific situation is the main motivation for [P1169R0].
A seemingly unrelated problem to the question of code quadruplication is that of writing numerous overloads for function wrappers, as demonstrated in [P0826R0]. Consider what happens if we implement std::not_fn()
as currently specified:
template <typename F>
class call_wrapper {
F f;
public:
// ...
template <typename... Args>
auto operator()(Args&&... ) &
-> decltype(!declval<invoke_result_t<F&, Args...>>());
template <typename... Args>
auto operator()(Args&&... ) const&
-> decltype(!declval<invoke_result_t<F const&, Args...>>());
// ... same for && and const && ...
};
template <typename F>
auto not_fn(F&& f) {
return call_wrapper<decay_t<F>>{forward<F>(f)};
}
As described in the paper, this implementation has two pathological cases: one in which the callable is SFINAE-unfriendly, causing the call to be ill-formed where it would otherwise work; and one in which overload is deleted, causing the call to fall back to a different overload when it should fail instead:
struct unfriendly {
template <typename T>
auto operator()(T v) {
static_assert(is_same_v<T, int>);
return v;
}
template <typename T>
auto operator()(T v) const {
static_assert(is_same_v<T, double>);
return v;
}
};
struct fun {
template <typename... Args>
void operator()(Args&&...) = delete;
template <typename... Args>
bool operator()(Args&&...) const { return true; }
};
std::not_fn(unfriendly{})(1); // static assert!
// even though the non-const overload is viable and would be the
// best match, during overload resolution, both overloads of
// unfriendly have to be instantiated - and the second one is a
// hard compile error.
std::not_fn(fun{})(); // ok!? Returns false
// even though we want the non-const overload to be deleted, the
// const overload of the call_wrapper ends up being viable - and
// the only viable candidate.
Gracefully handling SFINAE-unfriendly callables is not solvable in C++ today. Preventing fallback can be solved by the addition of another four overloads, so that each of the four cv/ref-qualifiers leads to a pair of overloads: one enabled and one deleted
.
This proposal solves both problems by allowing this
to be deduced. The following is a complete implementation of std::not_fn
. For simplicity, it makes use of BOOST_HOF_RETURNS
from Boost.HOF to avoid duplicating expressions:
template <typename F>
struct call_wrapper {
F f;
template <typename Self, typename... Args>
auto operator()(this Self&& self, Args&&... args)
BOOST_HOF_RETURNS(
!invoke(
forward<Self>(self).f,
forward<Args>(args)...))
};
template <typename F>
auto not_fn(F&& f) {
return call_wrapper<decay_t<F>>{forward<F>(f)};
}
Which leads to:
Here, there is only one overload with everything deduced together. The first example now works correctly. Self
gets deduced as call_wrapper<unfriendly>
, and the one operator()
will only consider unfriendly
’s non-const
call operator. The const
one is never even considered, so it does not have an opportunity to cause problems.
The second example now also fails correctly. Previously, we had four candidates. The two non-const
options were removed from the overload set due to fun
’s non-const
call operator being delete
d, and the two const
ones which were viable. But now, we only have one candidate. Self
is deduced as call_wrapper<fun>
, which requires fun
’s non-const
call operator to be well-formed. Since it is not, the call results in an error. There is no opportunity for fallback since only one overload is ever considered.
This singular overload has precisely the desired behavior: working for unfriendly
, and not working for fun
.
This could also be implemented as a lambda completely within the body of not_fn
:
template <typename F>
auto not_fn(F&& f) {
return [f=forward<F>(f)](this auto&& self, auto&&.. args)
BOOST_HOF_RETURNS(
!invoke(
forward_like<decltype(self)>(f),
forward<decltype(args)>(args)...))
;
}
In San Diego, 2018, there was a question of whether recursive lambdas are implementable. They are, details follow.
The specific issue is the way lambdas are parsed. When parsing a non-generic lambda function body with a default capture, the type of this_lambda
would not be dependent, because the body is not a template. This leads to sizeof(this_lambda)
not being dependent either, and must therefore have an answer - and yet, it cannot, as the lambda capture is not complete, and therefore, the type of this_lambda
is not complete.
This is a huge issue for any proposal of recursive lambdas that includes non-generic lambdas.
Notice, however, that the syntax this paper proposes is the following:
There is, quite obviously, no way to spell a non-generic lambda, because the lambda type is unutterable. self
’s type is always dependent.
This makes expressions depending on self
to be parsed using the regular rules of the language. Expressions involving self
become dependent, and the existing language rules apply, which means both nothing new to implement, and nothing new to teach.
This proposal is therefore implementable, unlike any other we’ve seen to date. We would really like to thank Daveed Vandevoorde for thinking through this one with us in Aspen 2019.
In Kona, EWGI asked us to see whether library implementors would use this. The answer seems to be a resounding yes.
We have heard from Casey Carter and Jonathan Wakely that they are interested in this feature. Also, on the ewg/lewg mailing lists, this paper comes up as a solution to a surprising number of questions, and gets referenced in many papers-in-flight. A sampling of papers:
In Herb Sutter’s “Name 5 most important papers for C++”, 10 out of 289 respondents chose it. Given that the cutoff was 5, and that modules, throwing values, contracts, reflection, coroutines, linear algebra, and pattern matching were all in that list, I find the result a strong indication that it is wanted.
We can also report that Gašper is dearly missing this feature in libciabatta, a mixin support library, as well as his regular work writing libraries.
On the question of whether this would get used in the standard library interfaces, the answer was “not without the ability to constrain the deduced type”, which is a feature C++ needs even without this paper, and is an orthogonal feature. The same authors were generally very enthusiastic about using this feature in their implementations.
A valid question to ask is what should be the type of this-annotated functions that have a member function equivalent? There are only two options, each with a trade-off. Please assume the existence of these three functions:
struct Y {
int f(int, int) const&; // exists
int g(this Y const&, int, int); // this paper
int h(this Y, int, int); // this paper, by value
};
g
has a current equivalent (f
), while h
does not. &Y::h
’s type must be a regular function pointer.
If we allow g
’s type to be a pointer-to-member-function, we get non-uniformity between the types of h
and g
. We also get implementation issues because the types a template can result in are non-uniform (is this a template for a member function or a free function? Surprise, it’s both!).
We also get forward compatibility with any concievable proposal for extension methods - those will also have to be free functions by necessity, for roughly the same reasons.
The paper originally proposed it the other way, but this was changed to the current wording through EWG input in Cologne, 2018.
One of the pitfalls of having a deduced object parameter is when the intent is solely to deduce the cv-qualifiers and value category of the object parameter, but a derived type is deduced as well — any access through an object that might have a derived type could inadvertently refer to a shadowed member in the derived class. While this is desirable and very powerful in the case of mixins, it is not always desirable in other situations. Superfluous template instantiations are also unwelcome side effects.
One family of possible solutions could be summarized as make it easy to get the base class pointer. However, all of these solutions still require extra instantiations. For optional::value()
, we really only want four instantiations: &
, const&
, &&
, and const&&
. If something inherits from optional
, we don’t want additional instantiations of those functions for the derived types, which won’t do anything new, anyway. This is code bloat.
This is already a problem for free-function templates: The authors have heard many a complaint about it from library vendors, even before this paper was introduced, as it is desirable to only deduce the ref-qualifier in many contexts. Therefore, it might make sense to tackle this issue in a more general way. A complementary feature could be proposed to constrain type deduction.
The authors strongly believe this feature is orthogonal. However, hoping that mentioning that solutions are in the pipeline helps gain consensus for this paper, we mention one solution here. The proposal is in early stages, and is not in the pre-belfast mailing. It will be present in the post-belfast mailing: computed deduction
This has been implemented in the EDG front end, with gracious help and encouragement from Daveed Vandevoorde. Implementation didn’t turn up any notable issues.
Where previously, member functions were divided into static member functions and non-static member functions, this gets a little more complex because some static member functions still use the implied object parameter (those that have an explicit this parameter) and some do not. This wording introduces the term “object member function” for the union of non-static member functions and static member functions with an explicit this parameter. Many functions were previously restricted to be non-static member functions are now restricted to be object member functions.
The biggest issue in working through the wording for this paper is: just how static is a function with an explicit this parameter? Member functions with an explicit this parameter are sort of half-static, half-non-static. They’re static in the sense that a pointer to a function with an explicit this parameter has pointer-to-function type, not pointer-to-member type, and there is no implicit this
in the body of such functions. They’re non-static in the sense that class access to such a function must be a full call expression, and you can use such functions to declare operators:
struct C {
void nonstatic_fun();
int explicit_fun(this C c) {
nonstatic_fun(); // error
c.nonstatic_fun(); // ok
static_fun(); // ok
auto x = this; // error
return 0;
}
static void static_fun() {
explicit_fun(); // error
explicit_fun(C{}); // error
C{}.explicit_fun(); // ok
}
static void operator()(int); // error
void operator()(this C, char); // ok
};
C c;
int (*a)(C) = &C::explicit_fun; // ok
auto x = c.static_fun; // ok
auto y = c.explicit_fun; // error
auto z = c.explicit_fun(); // ok
One question came up over the course of implementing this feature which was whether or not there should be implicit this syntax for invoking an “explicit this” member function from an “implicit this” one. That is:
struct D {
void explicit_fun(this D const&);
void implicit_fun() const {
this->explicit_fun(); // obviously ok
explicit_fun(); // but what about this?
}
};
It’s tempting to say that given an explicit this
parameter, we should always require an explicit this
argument. But one of the major advantages of having the “implicit this” call support in this context is that it allows derived types to not have to know about the implementation choice. As frequently used as a motivating example, we would like std::optional
to be able to implement its member functions using this language feature if it wants to, without us having to know about it:
struct F : std::optional<int>
{
bool is_big() const {
// if this language feature required explicit call syntax, then this
// call would be valid if and only if the particular implementation of
// optional did _not_ use this feature. This is undesirable
return value() > 42;
}
};
Or, more generally, the implementation strategy of a particular type’s member function should not be relevant for users deriving from that type. So “implicit this” stays.
The way we talk about what it means to invoke a member function, there’s always an object argument — whether explicit or implied. As mentioned above, it’s important that this works:
Where the call in #1
is equivalent to (*this).f();
due to the implicit this
, that’s the implied object argument.
But we say that f
is a static
member function, so what about this:
There is no this
in a static
member function, but there is always an object argument anyway - in this case, in general, we create a “contrived” argument of type B
and then say the call fails if we end up picking a non-static
member function that actually needs a real object argument. But that’s the status quo, what should we do in this case? There are two options to consider:
f
were a “normal” static
member function). But in this case, because f
has an explicit this
parameter, our contrived object argument would bind to the this
parameter and this call would fail. That is, we’re passing two arguments (a contrived object of type B
and a B
) to a function that takes a single parameter (B const&
).f
has an explicit this
parameter. As such, we’re just invoking a single-parameter function with a single argument and this works fine.It’s tempting to say, sure, let’s just choose option 2, since we would be allowing more functionality. But that runs into a wrinkle:
struct C {
static void f(C&); // #3
void f(this C const&); // #4
static void j() {
C c;
f(c); // #5
}
};
What does the call to f(c)
do? #3
is our legacy C++20 (or even C++98) candidate. But the way this candidate works is that f
behaves like a two-parameter function (the first parameter is ignored and matches everything equivalently and the second is C&
) and we are providing two arguments to it (a contrived object of type C
and an lvalue of type C
). If we say that #2
above is a valid call, then #4
here would likewise also be a valid candidate - but here we only have a single-parameter function with a single argument (there is no contrived object argument or fake parameter). In other words, we’re performing overload resolution now with candidates of different arity? This would be a first in C++.
It’s probably doable to come up with wording to properly handle this scenario, ending up with #3
being invoked, but it seems fairly complicated and we’re not even sure that we want to support the f(B{})
or f(c)
calls directly to begin with.
Instead, we propose not to support those calls. That is, while member functions with an explicit this
parameter are considered static member functions, they must still be invoked with an object argument (which can be implied) just like non-static member functions.
This does come with its own quirk. Since a function with an explicit this
parameter is still a static
member function, a pointer to it still has function pointer type, which means you can invoke through the function pointer. Just not directly. That is:
struct B {
void f(this B const&);
static void h() {
B{}.f(); // okay
f(B{}); // proposed: error. The B{} has to be the object
// argument, not just any kind of argument
}
};
void external() {
auto p = &B::f; // p is a void(*)(B const&)
p(B{}); // okay
}
This is, admittedly, a weird place to arrive at.
The status quo here is that a member function has an implicit object parameter, always of reference type, which the implied object argument is bound to. The obvious name for what this paper is proposing is, then, the explicit object parameter. The problem with these names is: well, what is an object parameter? A parameter that takes an object? Isn’t that most parameters?
Instead, the wording introduces the term this parameter, renaming implicit object parameter to implicit this parameter, and introducing the notion of an explicit this parameter. Alternate terms considered were “selector parameter” or “instance parameter”.
[ Editor's note: This paper introduces many new terms that are defined in [dcl.dcl] - so even though the wording here is presented in standard layout order (we obviously want to ensure that is_standard_layout<P0847>
is true
), it may be helpful to refer to those definitions when reviewing the wording ]
Move 11.4.3 [class.mfct.non-static]/3 in front of 7.5.4 [expr.prim.id]/2 (the highlighted diff is relative to the original paragraph):
2* When an id-expression that is not part of a class member access syntax and not used to form a pointer to member (7.6.2.2 [expr.unary.op]) is used in a member of class
X
in a context wherethis
can be used, if name lookup resolves the name in the id-expression to either a non-static non-type member or a static member function with an explicit this parameter ([dcl.fct]) of some classC
, and if either the id-expression is potentially evaluated orC
isX
or a base class ofX
, the id-expression is transformed into a class member access expression using(*this)
as the postfix-expression to the left of the.
operator. [ Note: IfC
is notX
or a base class ofX
, the class member access expression is ill-formed. — end note ] This transformation does not apply in the template definition context ([temp.dep.type]). [ Example: […] - end example ]
Strike the footnote in 7.5.4 [expr.prim.id]/2 as a drive-by-fix:
- (3.1) as part of a class member access in which the object expression refers to the member’s class
[Footnote: This also applies when the object expression is an implicitor a class derived from that class, or(*this)
([class.mfct.non-static]). ]
Change 7.5.5 [expr.prim.lambda]/3:
3 In the decl-specifier-seq of the lambda-declarator, each decl-specifier shall be one of
mutable
,constexpr
, orconsteval
. If the lambda-declarator contains an explicit this parameter ([dcl.fct]), then no decl-specifier in the decl-specifier-seq shall bemutable
. [ Note: The trailing requires-clause is described in [dcl.decl]. — end note ]
Extend the example in 7.5.5.2 [expr.prim.lambda.closure]/3:
auto glambda = [](auto a, auto&& b) { return a < b; }; bool b = glambda(3, 3.14); // OK auto vglambda = [](auto printer) { return [=](auto&& ... ts) { // OK: ts is a function parameter pack printer(std::forward<decltype(ts)>(ts)...); return [=]() { printer(ts ...); }; }; }; auto p = vglambda( [](auto v1, auto v2, auto v3) { std::cout << v1 << v2 << v3; } ); auto q = p(1, 'a', 3.14); // OK: outputs 1a3.14 q(); // OK: outputs 1a3.14 + + auto fact = [](this auto self, int n) -> int { // OK: explicit this parameter + return (n <= 1) ? 1 : n * self(n-1); + }; + std::cout << fact(5); // OK: outputs 120
Add a new paragraph after 7.5.5.2 [expr.prim.lambda.closure]/3:
3* Given a lambda with a lambda-capture, the type of the explicit this parameter, if any, of the lambda’s function call operator (possibly instantiated from a function call operator template) shall be either:
- (3*.1) the closure type,
- (3*.2) a class type derived from the closure type, or
- (3*.3) a reference to a possibly cv-qualified such type.
[ Example:
struct C { template <typename T> C(T); }; void func(int i) { int x = [=](this auto&&) { return i; }(); // ok int y = [=](this C) { return i; }(); // ill-formed int z = [](this C) { return 42; }(); // ok }
- end example ]
Change 7.5.5.2 [expr.prim.lambda.closure]/4:
4 The function call operator or operator template is declared
const
([class.mfct.non-static]) if and only if the lambda-expression’s parameter-declaration-clause is not followed bymutable
and the lambda-declarator does not contain an explicit this parameter.
Change 7.6.1.3 [expr.call]/1-2. The intent is to reduce the restriction on static member functions to just the ones without an explicit this parameter (since otherwise it would prevent the typical x.f()
use-cases from working, since x.f
is later defined as a prvalue). We do not need to change p2 to refer to object member functions since explicit this functions are also regular static member functions, and this restriction would prevent such uses.
1 […] The postfix expression shall have function type or function pointer type. For a call to a non-member function or to a static member function that does not have an explicit this parameter ([dcl.fct]), the postfix expression shall either be an lvalue that refers to a function (in which case the function-to-pointer standard conversion ([conv.func]) is suppressed on the postfix expression), or have function pointer type.
2 For a call to a non-static member function, the postfix expression shall be an implicit ([class.mfct.non-static], [class.static]) or explicit class member access whose id-expression is a function member name, or a pointer-to-member expression selecting a function member; the call is as a member of the class object referred to by the object expression. In the case of an implicit class member access, the implied object is the one pointed to by
this
. [ Note: A member function call of the formf()
is interpreted as(*this).f()
(see[class.mfct.non-static][expr.prim.id]). — end note ]
Change 7.6.1.3 [expr.call]/7:
7 When a function is called, each parameter ([dcl.fct]) is initialized ([dcl.init], [class.copy.ctor]) with its corresponding argument. If the function is a static member function with an explicit this parameter and there is an implied object argument ([over.call.func]), the list of provided arguments is preceded by the implied object argument for the purposes of this correspondence. If there is no corresponding argument, the default argument for the parameter is used. […] If the function is a non-static member function, the
this
parameter of the function is initialized with a pointer to the object of the call, converted as if by an explicit type conversion. [ Note: There is no access or ambiguity checking on this conversion; the access checking and disambiguation are done as part of the (possibly implicit) class member access operator. See [class.member.lookup], [class.access.base], and [expr.ref]. — end note ] When a function is called, the type of any parameter shall not be a class type that is either incomplete or abstract.
Change 7.6.1.5 [expr.ref]/6.3 - flipping the two bullets.
(6.3) If
E2
is a (possibly overloaded) member function, function overload resolution ([over.match]) is used to select the function to whichE2
refers. The type ofE1.E2
is the type ofE2
andE1.E2
refers to the function referred to byE2
.
- (6.3.1)
Otherwise (whenIfE2
refers to a non-static member function)E2
refers to an object member function ([dcl.fct]),E1.E2
is a prvalue. The expression can be used only as the left-hand operand of a member function call ([class.mfct]). [ Note: Any redundant set of parentheses surrounding the expression is ignored ([expr.prim.paren]). — end note ]- (6.3.2)
IfOtherwise,E2
refers to a static member functionE1.E2
is an lvalue.
In 9.3.4.6 [dcl.fct]/3, introduce explicit-this-parameter-declaration and non-this-parameter-declaration as the two kinds of parameter-declaration:
parameter-declaration-list:
parameter-declaration
parameter-declaration-list,
parameter-declaration
parameter-declarationnon-this-parameter-declaration:
attribute-specifier-seqoptdecl-specifier-seq declarator
attribute-specifier-seqoptdecl-specifier-seq declarator=
initializer-clause
attribute-specifier-seqoptdecl-specifier-seq abstract-declaratoropt
attribute-specifier-seqoptdecl-specifier-seq abstract-declaratoropt=
initializer-clauseexplicit-this-parameter-declaration:
this
non-this-parameter-declaration
parameter-declaration:
attribute-specifier-seqopt explicit-this-parameter-declaration
attribute-specifier-seqopt non-this-parameter-declaration
After 9.3.4.6 [dcl.fct]/5, insert a paragraph describing where a function declaration with an explicit this parameter may appear, and renumber section.
5a An explicit-this-parameter-declaration shall appear only as the first parameter-declaration of a parameter-declaration-list of either:
- (5a.1) a member-declarator that declares a member function ([class.mem]), or
- (5a.2) a lambda-declarator ([expr.prim.lambda]).
5b A member-declarator with an explicit-this-parameter-declaration shall not include a ref-qualifier or a cv-qualifier-seq and shall not be declared
static
orvirtual
. [ Note: Such a function is implicitly static ([class.mem]) - end note ][ Example:
struct C { void f(this C& self); template <typename Self> void g(this Self&& self, int); void h(this C) const; // error: const not allowed here }; void test(C c) { c.f(); // ok: calls C::f c.g(42); // ok: calls C::g<C&> std::move(c).g(42); // ok: calls C::g<C> }
- end example ]
5c A function parameter declared with an explicit-this-parameter-declaration is an explicit this parameter. An explicit this parameter shall not be a function parameter pack ([temp.variadic]). An object member function is either a static member function with an explicit this parameter or a non-static member function.
5d An non-this parameter is a function parameter that is not the explicit this parameter. The non-this-parameter-type-list of a member function is the parameter-type-list of that function with the explicit this parameter, if any, omitted. [ Note: The non-this-parameter-type-list consists of the adjusted types of all the non-this parameters. -end note ]
Change 9.5.4 [dcl.fct.def.coroutine]/3-4:
3 The promise type of a coroutine is
std::coroutine_traits<R, P1, …, Pn>::promise_type
, whereR
is the return type of the function, andP1…Pn
are the sequence of types of the non-this function parameters, preceded by the type of the implicit or explicit thisobjectparameter ([over.match.funcs]) if the coroutine isa non-statican object member function. The promise type shall be a class type.4 In the following,
pi
is an lvalue of typePi
, wherep1
denotesthe implicit or explicit this parameter and*this
pi+1
denotes the ith non-this function parameter fora non-statican object member function, andpi
denotes the ith function parameter otherwise.
Change 11.4 [class.mem]/4:
4 A data member or member function may be declared
static
in its member-declaration, in which case it is a static member (see [class.static]) (a static data member ([class.static.data]) or static member function ([class.static.mfct]), respectively) of the class. A member function declared with an explicit this parameter ([dcl.fct]) is a static member function. Any other data member or member function is a non-static member (a non-static data member or non-static member function ([class.mfct.non-static]), respectively). [ Note: A non-static data member of non-reference type is a member subobject of a class object. — end note ]
Remove 11.4.3 [class.mfct.non-static]/3 (was moved into [expr.prim.id] earlier).
Change 11.4.8.3 [class.conv.fct]/1:
1
AAn object member function of a classX
having no non-this parameters ([dcl.fct]) with a name of the form […] specifies a conversion fromX
to the type specified by the conversion-type-id. Such functions are called conversion functions. A decl-specifier in the decl-specifier-seq of a conversion function (if any) shall be neither a defining-type-specifier norstatic
. The type of the conversion function ([dcl.fct]) is “function taking no parameter returning conversion-type-id”.
Change 11.4.9.2 [class.static.mfct]/2:
2 [ Note: A static member function does not have a this pointer ([class.this]). — end note ] A static member function shall not be
virtual
. There shall not be a static and a non-static member function with the same name and the sameparameter typesnon-this-parameter-type-list ([dcl.fct], [over.load]). A static member function shall not be declaredconst
,volatile
, orconst volatile
.
Change 12.2 [over.load]/2.2:
(2.2) Member function declarations with the same name, the same
parameter-type-listnon-this-parameter-type-list ([dcl.fct]), and the same trailing requires-clause (if any) cannot be overloaded if any of them isanot an object member function declarationstatic
([class.static])([dcl.fct]). Likewise, member function template declarations with the same name, the sameparameter-type-listnon-this-parameter-type-list, the same trailing requires-clause (if any), and the same template-head cannot be overloaded if any of them isanot an object member function template declaration. The types of the implicit or explicitstatic
objectthis parameters constructed for the member functions for the purpose of overload resolution ([over.match.funcs]) are not considered when comparingparameter-type-listsnon-this-parameter-type-lists for enforcement of this rule. In contrast, ifthere is noevery member function declaration among a set of member function declarations with the same name, the samestatic
parameter-type-listnon-this-parameter-type-list, and the same trailing requires-clause (if any) is an object member function, then these member function declarations can be overloaded if they differ in the type of their implicit or explicitobjectthis parameter. [ Example: The following illustrates this distinction:class X { static void f(); void f(); // error void f() const; // error void f() const volatile; // error void g(); void g() const; // OK: no static g void g() const volatile; // OK: no static g + void h(this X&, int); + void h(int) &&; // OK: different this parameter type + void j(this const X&); + void j() const&; // error: same this parameter type + void k(this X&); // OK + void k(this X&&); // OK };
— end example ]
Change 12.4 [over.match]/1:
1 The selection criteria for the best function are the number of arguments, how well the arguments match the parameter-type-list of the candidate function, how well (for non-static member functions) the object matches the implicit
objectthis parameter ([dcl.fct]), and certain other properties of the candidate function.
Change 12.4.2 [over.match.funcs]/2-5:
2 So that argument and parameter lists are comparable within this heterogeneous set, a member function that does not have an explicit this parameter is considered to have an extra first parameter, called the implicit
objectthis parameter, which represents the object for which the member function has been called. For the purposes of overload resolution, both static and non-static member functions have an implicit or explicitobjectthis parameter, but constructors do not.3 Similarly, when appropriate, the context can construct an argument list that contains an implied object argument as the first argument in the list to denote the object to be operated on.
4 For non-static member functions, the type of the implicit
objectthis parameter is
- (4.1) “lvalue reference to cv
X
” for functions declared without a ref-qualifier or with the&
ref-qualifier- (4.2) “rvalue reference to cv
X
” for functions declared with the&&
ref-qualifierwhere
X
is the class of which the function is a member and cv is the cv-qualification on the member function declaration. [ Example: For aconst
member function of classX
, the extra parameter is assumed to have type “lvalue reference toconst X
”. — end example ] For conversion functions that are non-static member functions, the function is considered to be a member of the class of the implied object argument for the purpose of defining the type of the implicitobjectthis parameter. For non-conversion functions introduced by a using-declaration into a derived class, the function is considered to be a member of the derived class for the purpose of defining the type of the implicitobjectthis parameter. For static member functions that do not have an explicit this parameter, the implicitobjectthis parameter is considered to match any object (since if the function is selected, the object is discarded). [ Note: No actual type is established for the implicitobjectthis parameter of such a static member function, and no attempt will be made to determine a conversion sequence for that parameter ([over.match.best]). — end note ]5 During overload resolution, the implied object argument is indistinguishable from other arguments. The implicit
objectthis parameter, however, retains its identity since no user-defined conversions can be applied to achieve a type match with it. For non-static member functions declared without a ref-qualifier, even if the implicitobjectthis parameter is not const-qualified, an rvalue can be bound to the parameter as long as in all other respects the argument can be converted to the type of the implicitobjectthis parameter. [ Note: The fact that such an argument is an rvalue does not affect the ranking of implicit conversion sequences. — end note ]
Change 12.4.2.2.2 [over.call.func]/3 and adjust the corresponding footnote, and add an example:
3 Because of the rules for name lookup, the set of candidate functions consists (1) entirely of non-member functions or (2) entirely of member functions of some class T. In case (1), the argument list is the same as the expression-list in the call. In case (2), the argument list is the expression-list in the call augmented by the addition of an implied object argument as in a qualified function call. If the keyword
this
is in scope and refers to classT
, or a derived class ofT
, then the implied object argument is(*this)
. If the keywordthis
is not in scope or refers to another class, then a contrived object of typeT
becomes the implied object argument. 123 If the argument list is augmented by a contrived object and overload resolution selects one of thenon-staticobject member functions ofT
, the call is ill-formed.[ Example:
struct C { void a(); void b() { a(); // ok, (*this).a() } void f(this const C&); void g() const { f(); // ok: (*this).f() f(*this); // error: no viable candidate for (*this).f(*this) this->f(); // ok } static void h() { f(); // error: contrived object argument, but overload resolution picked an object member function f(C{}); // error: no viable candidate C{}.f(); // ok } };
- end example ]
119 An implied object argument must be contrived to correspond to the implicit
objectthis parameter attributed to member functions during overload resolution. It is not used in the call to the selected function. Since the member functions all have the same implicit object parameter, the contrived object will not be the cause to select or reject a function.
Add to 12.4.2.2.3 [over.call.object]/3:
3 The argument list submitted to overload resolution consists of the argument expressions present in the function call syntax preceded by the implied object argument
(E)
. [ Note: When comparing the call against the function call operators, the implied object argument is compared against either the implicit or explicitobjectthis parameter of the function call operator. When comparing the call against a surrogate call function, the implied object argument is compared against the first parameter of the surrogate call function. The conversion function from which the surrogate call function was derived will be used in the conversion sequence for that parameter since it converts the implied object argument to the appropriate function pointer or reference required by that first parameter. — end note ]
Change the note in 12.4.2.3 [over.match.oper]/3.4:
(3.4.5) [ Note: A candidate synthesized from a member candidate has its implicit or explicit
objectthis parameter as the second parameter, thus implicit conversions are considered for the first, but not for the second, parameter. — end note ]
Change the note in 12.4.2.5 [over.match.copy]/2:
2 In both cases, the argument list has one argument, which is the initializer expression. [ Note: This argument will be compared against the first parameter of the constructors and against the implicit or explicit
objectthis parameter of the conversion functions. — end note ]
Change the note in 12.4.2.6 [over.match.conv]/2:
2 The argument list has one argument, which is the initializer expression. [ Note: This argument will be compared against the implicit or explicit
objectthis parameter of the conversion functions. — end note ]
Change the note in 12.4.2.7 [over.match.ref]/2:
2 The argument list has one argument, which is the initializer expression. [ Note: This argument will be compared against the implicit or explicit
objectthis parameter of the conversion functions. — end note ]
Change 12.4.4.2 [over.best.ics]/4:
4 However, if the target is
- (4.1) the first parameter of a constructor or
- (4.2) the implicit or explicit
objectthis parameter of a user-defined conversion functionand the constructor or user-defined conversion function is a candidate by […]
Change 12.4.4.2 [over.best.ics]/7:
7 In all contexts, when converting to the implicit
objectthis parameter or when converting to the left operand of an assignment operation only standard conversion sequences are allowed. [Note: When converting to the explicit this parameter, if any, user-defined conversion sequences are allowed. - end note ]
Change 12.4.4.2.3 [over.ics.user]/1:
1 If the user-defined conversion is specified by a conversion function, the initial standard conversion sequence converts the source type to the implicit or explicit
objectthis parameter of the conversion function.
Change 12.4.4.2.5 [over.ics.ref]/3:
3 Except for an implicit
objectthis parameter, for which see [over.match.funcs], an implicit conversion sequence cannot be formed if it requires binding an lvalue reference other than a reference to a non-volatile const type to an rvalue or binding an rvalue reference to an lvalue other than a function lvalue. [ Note: This means, for example, that a candidate function cannot be a viable function if it has a non-const lvalue reference parameter (other than the implicitobjectthis parameter) and the corresponding argument would require a temporary to be created to initialize the lvalue reference (see [dcl.init.ref]). — end note ]
Change 12.4.4.3 [over.ics.rank]/3.2.3:
(3.2) Standard conversion sequence
S1
is a better conversion sequence than standard conversion sequenceS2
if
- (3.2.3)
S1
andS2
include reference bindings ([dcl.init.ref]) and neither refers to an implicitobjectthis parameter of a non-static member function declared without a ref-qualifier, andS1
binds an rvalue reference to an rvalue andS2
binds an lvalue reference
Change 12.6 [over.oper]/7:
7 An operator function shall either be
a non-statican object member function or be a non-member function that has at least one parameter whose type is a class, a reference to a class, an enumeration, or a reference to an enumeration.
Change 12.6.2 [over.unary]/1:
1 A prefix unary operator function is a function named
operator@
for a prefix unary-operator@
([expr.unary.op]) that is eithera non-statican object member function ([class.mfct]) with no non-this parameters or a non-member function with one parameter.
Change 12.6.3 [over.binary]/1:
1 A binary operator function is a function named
operator@
for a binary operator@
that is eithera non-statican object member function ([class.mfct]) with one non-this parameter or a non-member function with two parameters.
Change 12.6.3.2 [over.ass]/1:
A simple assignment operator function is a binary operator function named
operator=
. A simple assignment operator function shall bea non-statican object member function.
Change 12.6.4 [over.call]/1:
1 A function call operator function is a function named
operator()
that isa non-statican object member function with an arbitrary number of parameters.
Change 12.6.5 [over.sub]/1:
1 A subscripting operator function is a function named
operator[]
that isa non-statican object member function with exactly one non-this parameter.
Change 12.6.6 [over.ref]/1:
1 A class member access operator function is a function named
operator->
that isa non-statican object member function taking no non-this parameters.
Change 12.6.7 [over.inc]/1:
1 An increment operator function is a function named
operator++
. If this function isa non-statican object member function with no non-this parameters, or a non-member function with one parameter, it defines the prefix increment operator++
for objects of that type. If the function isa non-statican object member function with one non-this parameter (which shall be of typeint
) or a non-member function with two parameters (the second of which shall be of typeint
), it defines the postfix increment operator++
for objects of that type.
Add to 15.11 [cpp.predefined]/table 17 ([tab:cpp.predefined.ft]):
__cpp_explicit_this_parameter
with the appropriate value.
The authors would like to thank:
this
inside function bodies[EffCpp] Scott Meyers. Effective C++, Third Edition.
https://www.aristeia.com/books.html
[P0798R0] Simon Brand. 2017. Monadic operations for std::optional.
https://wg21.link/p0798r0
[P0798R3] Simon Brand. 2019. Monadic operations for std::optional.
https://wg21.link/p0798r3
[P0826R0] Agustín Bergé. 2017. SFINAE-friendly std::bind.
https://wg21.link/p0826r0
[P0839R0] Richard Smith. 2017. Recursive Lambdas.
https://wg21.link/p0839r0
[P0847R0] Gašper Ažman, Simon Brand, Ben Deane, Barry Revzin. 2018. Deducing this.
https://wg21.link/p0847r0
[P0847R1] Gašper Ažman, Simon Brand, Ben Deane, Barry Revzin. 2018. Deducing this.
https://wg21.link/p0847r1
[P0847R2] Gašper Ažman, Simon Brand, Ben Deane, Barry Revzin. 2019. Deducing this.
https://wg21.link/p0847r2
[P0929R2] Jens Maurer. 2018. Checking for abstract class types.
https://wg21.link/p0929r2
[P1169R0] Barry Revzin, Casey Carter. 2018. static operator().
https://wg21.link/p1169r0
[P1221R1] Jason Rice. 2018. Parametric Expressions.
https://wg21.link/p1221r1