[BUG] `constexpr` without `==`

[JohelEGP]

trafficstars

Title: constexpr without ==.

Description:

In reading P2996R0 Reflection for C++26 2.3 List of Types to List of Sizes, I couldn't help but wonder, how would we declare an uninitialized constexpr variable in Cpp2. This isn't a thing in Cpp1, yet. But I think it's been mentioned as a possible relaxation to constexpr, including modifiable constexpr globals (per TU?).

But there's actually a context that already affects Cpp2. An @interface with a constexpr function (https://compiler-explorer.com/z/aro9K15b1):

struct X {
  constexpr virtual char f() const = 0;
};
struct Y : X {
  constexpr char f() const override { return 'Y'; }
};
constexpr char call_f(const X* x) { return x->f(); }
static_assert([y=Y{}] { return call_f(&y); }() == 'Y');

In Cpp2, the best approximation is (https://cpp2.godbolt.org/z/7Gha4YdK7):

X: @interface type = {
  f: (this) -> char;
}
Y: type = {
  this: X = ();
  operator=: (out this) == { }
  f: (override this) -> char == 'Y';
}
call_f: (x: * const X) -> char == x*.f();
main: () = {
  static_assert(:() call_f(Y()$&);() == 'Y');
}

// Hack for `constexpr`.
namespace cpp2 {
constexpr auto assert_not_null(auto&& p CPP2_SOURCE_LOCATION_PARAM_WITH_DEFAULT) -> decltype(auto) requires true
{
    return CPP2_FORWARD(p);
}
}

But that errors with:

main.cpp2:11:17: error: static assertion expression is not an integral constant expression
   11 |   static_assert([_0 = Y()]() -> auto { return call_f(&_0);  }() == 'Y');
      |                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
main.cpp2:11:17: note: non-literal type '(lambda at main.cpp2:11:17)' cannot be used in a constant expression
1 error generated.

It's not possible for X to declare a constexpr default constructor or f. We could attempt using @polymorphic_base. In Cpp1 (https://compiler-explorer.com/z/Y653WPM7j):

struct X {
  constexpr virtual char f() const { return 'X'; }
};
struct Y : X {
  constexpr char f() const override { return 'Y'; }
};
constexpr char call_f(const X* x) { return x->f(); }
static_assert([x=X{}] { return call_f(&x); }() == 'X');
static_assert([y=Y{}] { return call_f(&y); }() == 'Y');

In Cpp2 (https://cpp2.godbolt.org/z/4MYv5E9vz):

X: @polymorphic_base type = {
  operator=: (out this) == { }
  operator=: (virtual move this) == { }
  f: (virtual this) -> char == 'X';
}
Y: type = {
  this: X = ();
  operator=: (out this) == { }
  f: (override this) -> char == 'Y';
}
call_f: (x: * const X) -> char == x*.f();
main: () = {
  static_assert(:() call_f(X()$&);() == 'X');
  static_assert(:() call_f(Y()$&);() == 'Y');
}

// Hack for `constexpr`.
namespace cpp2 {
constexpr auto assert_not_null(auto&& p CPP2_SOURCE_LOCATION_PARAM_WITH_DEFAULT) -> decltype(auto) requires true
{
    return CPP2_FORWARD(p);
}
}

That's a big jump in the API of X (a variable can be instantiated, calling f on isn't a compile-time error, calling std::terminate() in f helps at compile-time but otherwise delays to runtime). And it happens to work because the return type is a mere char (it could be a more complicated data type, the data-less polymorphic base might need the help of globals).

Minimal reproducer (https://cpp2.godbolt.org/z/7Gha4YdK7):

X: @interface type = {
  f: (this) -> char;
}
Y: type = {
  this: X = ();
  operator=: (out this) == { }
  f: (override this) -> char == 'Y';
}
call_f: (x: * const X) -> char == x*.f();
main: () = {
  static_assert(:() call_f(Y()$&);() == 'Y');
}

// Hack for `constexpr`.
namespace cpp2 {
constexpr auto assert_not_null(auto&& p CPP2_SOURCE_LOCATION_PARAM_WITH_DEFAULT) -> decltype(auto) requires true
{
    return CPP2_FORWARD(p);
}
}

Commands:

cppfront main.cpp2
clang++18 -std=c++23 -stdlib=libc++ -lc++abi -pedantic-errors -Wall -Wextra -Wconversion -Werror=unused-result -I . main.cpp

Expected result:

Consideration to constexpr without ==. It might be relevant to some users today. Cpp1 language evolution might also make it very relevant.

Herb has mentioned limiting global variables to constexpr. That could make x: int; at namespace scope an uninitialized constexpr variable. But that wouldn't help with other entities that can be left uninitialized.

Actual result and error:

Cpp2 lowered to Cpp1:

//=== Cpp2 type declarations ====================================================


#include "cpp2util.h"

class X;


class Y;


//=== Cpp2 type definitions and function declarations ===========================

class X {
  public: [[nodiscard]] virtual auto f() const -> char = 0;

public: virtual ~X() noexcept;

  public: X() = default;
  public: X(X const&) = delete; /* No 'that' constructor, suppress copy */
  public: auto operator=(X const&) -> void = delete;
};
class Y: public X {

  public: constexpr explicit Y();
  public: [[nodiscard]] constexpr auto f() const -> char override;

  public: Y(Y const&) = delete; /* No 'that' constructor, suppress copy */
  public: auto operator=(Y const&) -> void = delete;
};
[[nodiscard]] constexpr auto call_f(X const* x) -> char;
auto main() -> int;


//=== Cpp2 function definitions =================================================



  X::~X() noexcept{}
  constexpr Y::Y()
                             : X{  }
  {}
  [[nodiscard]] constexpr auto Y::f() const -> char { return 'Y';  }

[[nodiscard]] constexpr auto call_f(X const* x) -> char { return CPP2_UFCS_0(f, (*cpp2::assert_not_null(x)));  }
auto main() -> int{
  static_assert([_0 = Y()]() -> auto { return call_f(&_0);  }() == 'Y');
}

Output:

main.cpp2:11:17: error: static assertion expression is not an integral constant expression
   11 |   static_assert([_0 = Y()]() -> auto { return call_f(&_0);  }() == 'Y');
      |                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
main.cpp2:11:17: note: non-literal type '(lambda at main.cpp2:11:17)' cannot be used in a constant expression
1 error generated.

See also:

• Doing away with = and ==: #742.
• f: (x) x; is never constexpr: https://github.com/hsutter/cppfront/discussions/714#discussioncomment-7158721.
• Entities that can be left uninitialized: https://github.com/hsutter/cppfront/issues/273#issuecomment-1538234710.

[JohelEGP]

This could be considered a sufficiently niche issue for the sake of the general simplicity of Cpp2.

There's similar issues with the mapping of Cpp1 specifiers to Cpp2.

• static constexpr non-member objects: https://github.com/hsutter/cppfront/commit/b589f5d25e5acdbfd94791e23c0ba0f6fdcd59c6#commitcomment-129028816.
  • static non-member objects: #212, #322 (https://github.com/hsutter/cppfront/issues?q=is%3Aissue+in%3Atitle+static+is%3Aclosed).
• static call operator for function expressions: P1169R4 3.2.2 Can the static-ness of lambdas be implementation-defined?.

The last one is for performance. Cpp2 has no opt-in and Cppfront doesn't lower to it, so we're backwards-compatible with Cpp1. If static was the default, we'd lose backwards-compatibility. An explicit this parameter (not currently supported) would lower to a generic lambda, and its function pointers are not member function pointers. The only truly backwards-compatible way would be to explicitly opt-into static. So, for performance, maybe we'd need something like a static virtual-specifier.

[JohelEGP]

From https://wg21.link/p2996r0#enum-to-string:

  template for (constexpr auto e : std::meta::members_of(^E)) {

The non-template for equivalent in Cpp2 is:

for range do (e) {

A way to support the proposed Cpp1 template for is with Cpp2 template for. Or, alternatively, if we could specify e as constexpr somehow. Although, arguably, the template keyword upfront is a good indicator of the loop's nature.

Another proposed syntax was for..., but that doesn't work for reflection. std::meta::members_of(^E) is not a pack, but a range to perform heterogeneous splicing.

[JohelEGP]

It also seems that constexpr with == doesn't compose. Only a subset of the built-in metafunctions generate constexpr functions. The built-in metafunctions are

interface, polymorphic_base, ordered, weakly_ordered, partially_ordered, copyable, basic_value, value, weakly_ordered_value, partially_ordered_value, struct, enum, flag_enum, union, print

• struct and print don't generate functions.
• Only the ordering and enum metafunctions will generate constexpr functions.

The ordering metafunctions will be constexpr because they lower to =defaulted functions. In commit b589f5d25e5acdbfd94791e23c0ba0f6fdcd59c6, the enum metafunctions had to stop using basic_value to explicitly opt-into the == syntax for constexpr.

Only interface and polymorphic_base are really sensitive to constexpr. You can only use virtual dispatch from a base pointer if its function is constexpr.

The copyable and value metafunctions never generate constexpr functions. Since P2448R2, those can simply be constexpr functions. From https://en.cppreference.com/w/cpp/compiler_support:

C++23 feature	Paper(s)	GCC	Clang	MSVC	Apple Clang
Relaxing some constexpr restrictions	P2448R2	13	17 (partial)

Or maybe constexpr should still be considered a contract that requires explicit opt-in. Just like how operator= works. I've considered a @constexpr type metafunction to mark all member functions as constexpr. But that might be too broad (e.g., @enum @constexpr would mark streaming operator<< as constexpr). I've also thought of giving an argument to copyable and the value metafunctions to opt-into constexpr.

[JohelEGP]

I've considered a @constexpr type metafunction to mark all member functions as constexpr. But that might be too broad (e.g., @enum @constexpr would mark streaming operator<< as constexpr). I've also thought of giving an argument to copyable and the value metafunctions to opt-into constexpr.

Maybe @constexpr should be a variadic meta-metafunction. It would take the metafunctions to invoke and mark the added members as constexpr. The reflection API would have to be extended to allow declaring never-constexpr member functions. For example:

• The enum metafunctions would switch to using .add_runexpr_functionᵖʳᵒᵛⁱᵗⁱᵒⁿᵃˡ for operator<<. Then @constexpr(enum) wouldn't declare the operator<< as constexpr.
• @constexpr(basic_value) would declare the functions added by basic_value as constexpr. Then, the enum metafunctions can resume using basic_value.

[JohelEGP]

Maybe @constexpr should be a variadic meta-metafunction.

Doesn't seem possible without extending the reflection API. Metafunction template arguments are just strings.

[JohelEGP]

See also #959.