[clang] Clang 19 reports "arithmetic on a pointer to an incomplete type" where Clang 18 and GCC do not

[vient]

https://godbolt.org/z/sYPWjWqbo

struct _Destroy_aux {
  template <typename _ForwardIterator> static void __destroy(_ForwardIterator) {
    _ForwardIterator __first;
    ++__first;
  }
};

template <typename _ForwardIterator> void _Destroy(_ForwardIterator __last) {
  _Destroy_aux::__destroy(__last);
}

template <typename> struct allocator;

template <typename> struct allocator_traits;

template <typename _Tp> struct allocator_traits<allocator<_Tp>> {
  using pointer = _Tp *;
};

struct FileInfo;

struct __alloc_traits : allocator_traits<allocator<FileInfo>> {};

__alloc_traits::pointer _M_finish;

template <typename = int> struct vector {
  ~vector() { _Destroy(_M_finish); }
};

struct DatasetFactory {
  virtual ~DatasetFactory();
};

struct FileSystemDatasetFactory : DatasetFactory {
  vector<> v;
};

<source>:4:5: error: arithmetic on a pointer to an incomplete type 'FileInfo'
    4 |     ++__first;
      |     ^ ~~~~~~~
<source>:9:17: note: in instantiation of function template specialization '_Destroy_aux::__destroy<FileInfo *>' requested here
    9 |   _Destroy_aux::__destroy(__last);
      |                 ^
<source>:27:15: note: in instantiation of function template specialization '_Destroy<FileInfo *>' requested here
   27 |   ~vector() { _Destroy(_M_finish); }
      |               ^
<source>:34:8: note: in instantiation of member function 'vector<>::~vector' requested here
   34 | struct FileSystemDatasetFactory : DatasetFactory {
      |        ^
<source>:20:8: note: forward declaration of 'FileInfo'
   20 | struct FileInfo;
      |        ^
1 error generated.

Minified from Apache Arrow sources. Removing template from vector makes all compilers report the same error.

Is it a problem with Clang or with Arrow?

[cor3ntin]

All compilers do error if FileSystemDatasetFactory is actually used. Clang is a bit more eager in instantiating the destructor though, hence the variance that you observe.

https://godbolt.org/z/TTdWae7vK

[vient]

They error because a type is incomplete. In Arrow FileSystemDatasetFactory constructor is called only from a complementary .cpp file where a header with full definition of FileInfo is included - external code uses static Make method, which calls a protected constructor. Otherwise it would've meant that nobody uses this class. Again, I don't know what side is wrong here.

FileSystemDatasetFactory definition: https://github.com/apache/arrow/blob/apache-arrow-17.0.0/cpp/src/arrow/dataset/discovery.h#L204 - issue occurs with std::vector<fs::FileInfo> files_ field.
Search for class name: https://github.com/search?q=repo%3Aapache%2Farrow+%2F%28%5E%7C%5CW%29FileSystemDatasetFactory%28%24%7C%5CW%29%2F+language%3AC%2B%2B+path%3A%2F%5Ecpp%5C%2F%2F&type=code - looks like other code exclusively uses static Make method of this class, so FileSystemDatasetFactory instances are only constructed inside its .cpp where FileInfo struct is known.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: vient (vient)

https://godbolt.org/z/sYPWjWqbo

struct _Destroy_aux {
  template <typename _ForwardIterator> static void __destroy(_ForwardIterator) {
    _ForwardIterator __first;
    ++__first;
  }
};

template <typename _ForwardIterator> void _Destroy(_ForwardIterator __last) {
  _Destroy_aux::__destroy(__last);
}

template <typename> struct allocator;

template <typename> struct allocator_traits;

template <typename _Tp> struct allocator_traits<allocator<_Tp>> {
  using pointer = _Tp *;
};

struct FileInfo;

struct __alloc_traits : allocator_traits<allocator<FileInfo>> {};

__alloc_traits::pointer _M_finish;

template <typename = int> struct vector {
  ~vector() { _Destroy(_M_finish); }
};

struct DatasetFactory {
  virtual ~DatasetFactory();
};

struct FileSystemDatasetFactory : DatasetFactory {
  vector<> v;
};

<source>:4:5: error: arithmetic on a pointer to an incomplete type 'FileInfo'
    4 |     ++__first;
      |     ^ ~~~~~~~
<source>:9:17: note: in instantiation of function template specialization '_Destroy_aux::__destroy<FileInfo *>' requested here
    9 |   _Destroy_aux::__destroy(__last);
      |                 ^
<source>:27:15: note: in instantiation of function template specialization '_Destroy<FileInfo *>' requested here
   27 |   ~vector() { _Destroy(_M_finish); }
      |               ^
<source>:34:8: note: in instantiation of member function 'vector<>::~vector' requested here
   34 | struct FileSystemDatasetFactory : DatasetFactory {
      |        ^
<source>:20:8: note: forward declaration of 'FileInfo'
   20 | struct FileInfo;
      |        ^
1 error generated.

Minified from Apache Arrow sources. Removing template from vector makes all compilers report the same error.

Is it a problem with Clang or with Arrow?

[cor3ntin]

@zygoloid I might need your help here

[dcl.fct.def]

A non-user-provided defaulted function (i.e., implicitly declared or explicitly defaulted in the class) that is not defined as deleted is implicitly defined when it is odr-used ([basic.def.odr]) or needed for constant evaluation ([expr.const]).

[basic.def.odr]

A virtual member function is odr-used if it is not pure

I am not sure this is enough to establish when the destructor is odr-used.

I think clang is correct here, ie it odr-uses the implicitly declared destructor (which is virtual) on declaration.
GCC seems to not do that until the class is used. Not sure who is correct here.

[zygoloid]

We're allowed to instantiate the virtual function here, but for QoI purposes generally we try to avoid doing so unless it's actually needed. It'd be useful to know why we're marking the vtable as used or instantiating the destructor. (If the destructor is constexpr I think we might trigger instantiation, but I think it shouldn't be because the base class dtor is not.)

[Endilll]

Bisected to 99500e8 (#77753)
CC @Fznamznon

[Fznamznon]

(If the destructor is constexpr I think we might trigger instantiation, but I think it shouldn't be because the base class dtor is not.)

Seems like a classic case with clang's eager template instantiation? 99500e8 made clang to not care about constexprness of base class destructor.

[zygoloid]

I see. Maybe we can avoid triggering vtable instantiation due to a dtor that is not "effectively constexpr" (if it's either not constexpr or if it's noexcept and a member or base class dtor is not effectively constexpr), with a special-case note for when this causes constant evaluation failure.

[Fznamznon]

Maybe we can avoid triggering vtable instantiation due to a dtor that is not "effectively constexpr" (if it's either not constexpr or if it's noexcept and a member or base class dtor is not effectively constexpr), with a special-case note for when this causes constant evaluation failure.

Trying...