[Clang][Regression:20][Modules] Name lookup failed to find hidden friend in non-exported base class

[MagentaTreehouse]

m.cppm

export module m;
struct Base {
    template <class T>
    friend constexpr auto f(T) { return 0; }
};
export struct A: Base {};

main.cpp

import m;
int main() {
    f(A{});
}

See https://compiler-explorer.com/z/v4xdnT7eW.

Error since Clang 20:

main.cpp:3:5: error: use of undeclared identifier 'f'
    3 |     f(A{});
      |     ^

[llvmbot]

@llvm/issue-subscribers-clang-modules

Author: None (MagentaTreehouse)

m.cppm ```c++ export module m; struct Base { template <class T> friend constexpr auto f(T) { return 0; } }; export struct A: Base {}; ```

main.cpp

import m;
int main() {
    f(A{});
}

See https://compiler-explorer.com/z/v4xdnT7eW.

Error since Clang 20:

main.cpp:3:5: error: use of undeclared identifier 'f'
    3 |     f(A{});
      |     ^

[ayushpareek2003]

What if we update the getExplicitVisibility() function in "clang/lib/AST/Decl.cpp" to ensure that friend functions in modules are considered for visibility, even if they are not explicitly exported? This change would involve modifying the LVComputationKind to not ignore explicit visibility for friend functions, which might address the issue

[ChuanqiXu9]

What if we update the getExplicitVisibility() function in "clang/lib/AST/Decl.cpp" to ensure that friend functions in modules are considered for visibility, even if they are not explicitly exported? This change would involve modifying the LVComputationKind to not ignore explicit visibility for friend functions, which might address the issue

The concern I have is style points. Now we judge visibility in Sema. It seems slightly odd/bad to make it in AST.