Possibly wrong reachability analysis with global module fragments in module partitions

[rohanlean]

The following C++23 code is rejected by clang version 20.1.6, saying that the declaration of S is unreachable in I.cpp.

// H.hpp
struct S{};

// M.cppm
export module M;

// P.cppm
module;
#include "H.hpp"
module M:P;

using T = S;

// I.cpp
module M;
import :P;

T f() { return {}; }

A variation involving no module partitions is accepted, and I do not find a reason for the rejection in the standard. My reading of 10.4 and 10.7 is that the definition is reachable, with emphasis on 10.4.3.1 and 10.7.3.2.

[llvmbot]

@llvm/issue-subscribers-clang-modules

Author: None (rohanlean)

The following C++23 code is rejected by clang version 20.1.6, saying that the declaration of `S` is unreachable in `I.cpp`. ``` // H.hpp struct S{};

// M.cppm
export module M;

// P.cppm
module;
#include "H.hpp"
module M:P;

using T = S;

// I.cpp
module M;
import :P;

T f() { return {}; }

A variation involving no module partitions is accepted, and I do not find a reason for the rejection in the standard.  My reading of 10.4 and 10.7 is that the definition is reachable, with emphasis on 10.4.3.1 and 10.7.3.2.
</details>

[ChuanqiXu9]

See https://clang.llvm.org/docs/StandardCPlusPlusModules.html#id36

[rohanlean]

Thank you for your answer @ChuanqiXu9. I am not sure if I fully understand how it explains the situation, though. Your link takes me to a section titled “Interactions with Reduced BMI”, but I am not using reduced BMI when compiling the example. The linked section also does not show any examples of standard violations arising from reduced BMIs or experimental non-cascading changes. (And neither does the preceding section on experimental non-cascading changes.)

If you are violating the standard on purpose to achieve better build performance, the error message should be improved in any case.

[ChuanqiXu9]

Thank you for your answer @ChuanqiXu9. I am not sure if I fully understand how it explains the situation, though. Your link takes me to a section titled “Interactions with Reduced BMI”, but I am not using reduced BMI when compiling the example. The linked section also does not show any examples of standard violations arising from reduced BMIs or experimental non-cascading changes. (And neither does the preceding section on experimental non-cascading changes.)

If you are violating the standard on purpose to achieve better build performance, the error message should be improved in any case.

Sorry, there are some problems in the docs. Please see https://clang.llvm.org/docs/StandardCPlusPlusModules.html#reachability-of-internal-partition-units

[rohanlean]

Thanks! I still do not think that explains it, because while in my example M:P is not an interface unit, the highlighted part in this quote should take effect:

A translation unit U is necessarily reachable from a point P if U is a module interface unit on which the translation unit containing P has an interface dependency, or the translation unit containing P imports U, in either case prior to P.

[ChuanqiXu9]

It seems you're correct. Sorry for reading it too quickly.