False Positive of ODR Violation for Functions with Internal Linkage in Modules

[ckwastra]

Code (CE):

• m.cpp:

  module;
  #include <cstdio>
  export module m;
  static void f() { std::puts(__FILE__); }
  export auto &fr = f;

• main.cpp:

  #include <cstdio>
  import m;
  static void f() { std::puts(__FILE__); }
  constexpr auto &fr1 = f;
  auto &fr2 = fr;
  int main() {
    fr1();
    fr2();
  }

Clang raises the following error:

m.cpp:4:13: error: 'f' has different definitions in different modules; definition in module 'm' first difference is function body
    4 | static void f() { std::puts(__FILE__); }
      | ~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~
main.cpp:3:13: note: but in '' found a different body
    3 | static void f() { std::puts(__FILE__); }
      | ~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~

GCC accepts the code (as expected), and running the produced executable outputs:

/app/main.cpp
/app/m.cpp

[llvmbot]

@llvm/issue-subscribers-clang-modules

Author: Vincent X (ckwastra)

Code ([CE](https://godbolt.org/z/7rqfqKvjq)):

• m.cpp:

  module;
  #include <cstdio>
  export module m;
  static void f() { std::puts(__FILE__); }
  export auto &fr = f;

• main.cpp:

  #include <cstdio>
  import m;
  static void f() { std::puts(__FILE__); }
  constexpr auto &fr1 = f;
  auto &fr2 = fr;
  int main() {
    fr1();
    fr2();
  }

Clang raises the following error:

m.cpp:4:13: error: 'f' has different definitions in different modules; definition in module 'm' first difference is function body
    4 | static void f() { std::puts(__FILE__); }
      | ~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~
main.cpp:3:13: note: but in '' found a different body
    3 | static void f() { std::puts(__FILE__); }
      | ~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~

GCC accepts the code (as expected), and running the produced executable outputs:

/app/main.cpp
/app/m.cpp

[shafik]

CC @ChuanqiXu9 @mpark

[ChuanqiXu9]

The hard part to fix the problem properly is, how do we model the concept of local entities in different TUs. Previously, before modules, we won't see multiple duplicated local entities.

It is easy to suppress the diagnostics. But it is problematic if the contents of the definition differs, then we may fall into real ODRs. Although actually, maybe, I am not sure, it may work due to the static linkage, but it is a bad model. We'd better give it a better model to handle this.