Cannot implicitly convert between a subclass of std::vector<T*> and tcb::span<const T* const>

[tcbrindle]

Cannot implicitly convert between a subclass of std::vector and tcb::span

from span.

[tcbrindle]

Subclassing std::vector<int*> and converting to span<const int* const> seems to work correctly in a quick test:

https://gcc.godbolt.org/z/Y68boPvYo

Are you able to give any more information about the error message that might help me recreate the problem, or perhaps try with another compiler?

[matthew-wozniczka]

I was trying to reproduce it w/ something closer to my code, but it was still working, but when I switched to MSVC even the std::vector case broke! (maybe I got confused earlier about it working)

See https://gcc.godbolt.org/z/e3q5e4PW5

It looks like it doesn't like it when it needs to add const?

[kimci86]

I think the problem boils down to this:

#include <type_traits>
static_assert(std::is_convertible<int* (*)[], const int* const (*)[]>::value, "");
int main() {}

On compiler explorer, this compiles fine with:

• x64 MSVC 19.14 with /std:c++17
• GCC >= 10.1 with -std:c++11
• clang >= 7.0.0 with -std:c++11

This does not compile with x64 MSVC 19.14 with /std:c++14 nor with older GCC or clang.

If we remove the first const, it compile everywhere.

[tcbrindle]

Thank you very much for the detective work @kimci86!

After some more investigation, this seems to be caused by CWG issue 330, which was reported in 2002 and finally resolved via N4261 in 2014.

So that's the cause, but unfortunately there are no simple solutions that I can think of.

The condition that's being checked -- can we convert a pointer-to-array-of-U into a pointer-to-array-of-T -- is directly lifted from the std::span specification, and I'm very reluctant to change it as it what prevents, for example, instantiating a span<Base> from an array of Derived.

Unless anyone has any clever ideas, I'm afraid that this may just be something that we have to say is unsupported in older compilers.

(Note that gsl::span uses the exact same check and so I guess it has precisely the same issue.)

[matthew-wozniczka]

Wouldn't it be possible to 'wrap' std::is_convertible in something which handles this case by first removing all the consts from the type? IMO it would make sense to try & make this work for older compilers since this library is meant to support older C++ standards?

[tcbrindle]

The goal is to provide span for older compilers to the extent supported by the language. If the core language is telling us that it doesn't think the pointers are convertible, well, we probably shouldn't do the conversion.