[BUG] @metaclass prevents type inheriting

[realgdman]

Having this: Base prevents type with some @meta from compiling

Reproduce Code

EmptyBase: type = {}
BaseBV: @basic_value type = {}
BV_BV: @basic_value type = { this: BaseBV; } //error
BV_B: @basic_value type = { this: EmptyBase; } //error
I_B: @interface type = { this: EmptyBase; } //error

https://cpp2.godbolt.org/z/Y5sY4a39G

Version
latest (38aec57)

Command lines
cppfront/cppfront $1.cpp2 -p
clang++-15 -Icppfront/include $1.cpp -std=c++20 -o $1

Expected result
Successful compilation or clearer diagnostic if that's intended

Actual result
@basic_value or @value or @struct:
error: expected 'BaseBV = ...' initialization statement - an operator= body must start with a series of 'member = value;'

@interface:
error: while applying @interface - interfaces may not contain data objects

Additional context
This is specifically about what seems as valid combinations, like @interface::@interface

I'm not talking about impossible combinations, like @polymorphic_base::@copyable, that could be another issue.

Which actually compiles, giving cpp1 warning about

[realgdman]

Regarding @basic_value inheriting @basic_value this might be reasonable error, requiring programmer to write this: BaseBV = ()

Regarding @interface, it seems because m.is_object() is true for this:. This may be right on parse level, but not necessary on semantic/reflection level. Currently it can be workaround with m.require( !m.is_object() || m.name()=="this") inside @interface

[JohelEGP]

This is now #779.

@basic_value or @value or @struct:
error: expected 'BaseBV = ...' initialization statement - an operator= body must start with a series of 'member = value;'

With @basic_value, you opt-into @copyable, which is documented as

cppfront/source/reflect.h2

Line 634 in 11fc88d

// A type with (copy and move) x (construction and assignment)

@copyable also injects operator=: (out this, that) = { }.
But types with this data members only lower a copy constructor from that.
So for types with this data members,
injecting operator=: (out this, that) = { } isn't enough for @copyable to meet its documented behavior.
I think @copyable should reject types with this data members.

By the way,
I've been using @basic_values for polymorphic types,
and this is how I ran into this (https://cpp2.godbolt.org/z/xnoEzc6GE):

sf_vertices: @basic_value type = { }
my_vertices: @basic_value type = {
  this: sf_vertices = ();
  operator=: (out this, size: i32) = _ = size;
}
tiles: type = {
  public vertices: my_vertices;
  operator=: (out this, size: i32) = vertices = size;
  private operator=: (inout this, _: i32) = vertices = my_vertices(0);
}
main: () = { }

main.cpp2:6:9: warning: definition of implicit copy assignment operator for 'my_vertices' is deprecated because it has a user-provided copy constructor [-Wdeprecated-copy-with-user-provided-copy]
    6 | public: my_vertices(my_vertices const& that);
      |         ^
main.cpp2:10:54: note: in implicit copy assignment operator for 'my_vertices' first required here
   10 |                                             vertices = my_vertices(0);
      |                                                      ^
1 warning generated.

[JohelEGP]

I've been using @basic_values for polymorphic types,

Not "polymorphic types", but "implementation inheritance".
My intent was to opt-into the bases' rule of N.

[JohelEGP]

So what I really want is Cpp1's defaults.
Which is equivalent to defaulting all special member functions (default constructor included).
I don't think anything in Cpp2 can give you that for a type that has this data members.

[JohelEGP]

For some more context, I'm writing a shim.
I'm adding a strongly typed interface to a library's types.
I use implementation inheritance because not everything needs wrapping.

I don't think anything in Cpp2 can give you that for a type that has this data members.

Given that @basic_value doesn't work for my use case,
these are my current options AFAIK:

• Manually write the default special member functions to forward to the base one.
• Write a CRTP template to opt-into simulating the Cpp1 defaults.

[JohelEGP]

• Manually write the default special member functions to forward to the base one.

This is actually the only option.
A type with a this data member requires manually writing all operator=s,
because none is generated.
So the only way to opt-into the Cpp1 defaults is by having a metafunction to do just that.

[JohelEGP]

Instead of @basic_value, this is what I have to use, omitting as necessary:

  operator=: (out this) = { }
  operator=: (out this, that)        = { }
  operator=: (out this, move that)   = { }
  operator=: (inout this, that)      = { }
  operator=: (inout this, move that) = { }

The default constructor uses the data members' initializers,
and the others default to the same memberwise operation.

[JohelEGP]

Regarding @interface, it seems because m.is_object() is true for this:. This may be right on parse level, but not necessary on semantic/reflection level. Currently it can be workaround with m.require( !m.is_object() || m.name()=="this") inside @interface

This is the documentation for @interface:

cppfront/source/reflect.h2

Line 475 in 11fc88d

// an abstract base class having only pure virtual functions

There's no way to know, nor assert at compile-time,
that a base type makes (or not) the type being defined meet that documented behavior.

[JohelEGP]

So what I really want is Cpp1's defaults.
Which is equivalent to defaulting all special member functions (default constructor included).

I.e., the rule of zero.

Using @basic_value with a data member of std::unique_ptr
can't be used to mimic the fantastic Cpp1's rule of zero (https://cpp2.godbolt.org/z/zafe6WM16):

t: @basic_value type = {
  p: std::unique_ptr<i32> = ();
  operator=: (out this, x: i32) = p = new<i32>(x);
}
main: () = { }

main.cpp2:6:35: error: call to implicitly-deleted copy constructor of 'std::unique_ptr<cpp2::i32>' (aka 'unique_ptr<int>')
    6 |                                 : p{ that.p }{}
      |                                   ^~~~~~~~~~~
main.cpp2:9:35: error: object of type 'std::unique_ptr<cpp2::i32>' (aka 'unique_ptr<int>') cannot be assigned because its copy assignment operator is implicitly deleted
    9 |                                 p = that.p;
      |                                   ^

Instead of @basic_value, this is what I have to use, omitting as necessary:

  operator=: (out this) = { }
  operator=: (out this, that)        = { }
  operator=: (out this, move that)   = { }
  operator=: (inout this, that)      = { }
  operator=: (inout this, move that) = { }

The default constructor uses the data members' initializers,
and the others default to the same memberwise operation.

This happens to work, but is unfortunately mechanic and verbose (https://cpp2.godbolt.org/z/rWfG5cvTz):

t: type = {
  p: std::unique_ptr<i32> = ();
  operator=: (out this, x: i32) = p = new<i32>(x);
  operator=: (out this) = { }
  // operator=: (out this, that)        = { }
  operator=: (out this, move that)   = { }
  // operator=: (inout this, that)      = { }
  operator=: (inout this, move that) = { }
}
main: () = { }

[JohelEGP]

Fixing #547 as requested would make my uses of @basic_value above work (and my actual use case, too).
But then @copyable's behavior would be different from what it documents.

[JohelEGP]

This happens to work, but is unfortunately mechanic and verbose (https://cpp2.godbolt.org/z/rWfG5cvTz):

And this is how it has to look if it's a template of arity 1 (https://cpp2.godbolt.org/z/rYqonvqYP):

t: <T> type = {
  p: T = ();
  operator=: (out this, x: i32) = p = new<i32>(x);
  operator=: (out this) = { }
  operator=: (out this, that) requires std::copyable<T>   = { }
  operator=: (out this, move that)                        = { }
  operator=: (inout this, that) requires std::copyable<T> = { }
  operator=: (inout this, move that)                      = { }
}
main: () = {
  static_assert(std::movable<t<std::unique_ptr<i32>>>);
  static_assert(!std::copyable<t<std::unique_ptr<i32>>>);
}