C++20 modules available but `__cpp_modules` not defined

[boris-kolpackov]

While Clang now enables C++20 modules by default with -std=c++20 and later, it does not appear to define the __cpp_modules feature test macros (which should have the value of 201907L; see https://en.cppreference.com/w/cpp/feature_test). I've tested this with up to Clang 18 pre-release (20231102103655+18839aec4ed1).

Is there a reason for omitting this macro for now or is it an oversight that could be fixed? In build2 we currently define this macro on the command line which adds quite a bit of noise.

@iains @ChuanqiXu9 @dwblaikie

[llvmbot]

@llvm/issue-subscribers-c-20

Author: Boris Kolpackov (boris-kolpackov)

While Clang now enables C++20 modules by default with `-std=c++20` and later, it does not appear to define the `__cpp_modules` feature test macros (which should have the value of `201907L`; see https://en.cppreference.com/w/cpp/feature_test). I've tested this with up to Clang 18 pre-release (`20231102103655+18839aec4ed1`).

Is there a reason for omitting this macro for now or is it an oversight that could be fixed? In build2 we currently define this macro on the command line which adds quite a bit of noise.

@iains @ChuanqiXu9 @dwblaikie

[llvmbot]

@llvm/issue-subscribers-clang-modules

Author: Boris Kolpackov (boris-kolpackov)

While Clang now enables C++20 modules by default with `-std=c++20` and later, it does not appear to define the `__cpp_modules` feature test macros (which should have the value of `201907L`; see https://en.cppreference.com/w/cpp/feature_test). I've tested this with up to Clang 18 pre-release (`20231102103655+18839aec4ed1`).

Is there a reason for omitting this macro for now or is it an oversight that could be fixed? In build2 we currently define this macro on the command line which adds quite a bit of noise.

@iains @ChuanqiXu9 @dwblaikie

[ChuanqiXu9]

The major reason is that we haven't marked the status of Modules as complete now. See https://clang.llvm.org/cxx_status.html

The reason behind this may be a different strategy to claim a feature as stable. For example, the support of coroutines in clang is marked as partial before Mar this year: https://discourse.llvm.org/t/rfc-could-we-mark-coroutines-as-unreleased-now/69220, while it is already used in production for years.

[boris-kolpackov]

The major reason is that we haven't marked the status of Modules as complete now.

This feels inconsistent to me: you enable modules by default in -std=c++20 (and have dropped -fmodules-ts) but you don't signal in any way that they are available.

Perhaps a compromise would be to define this macro but with a value that is an earlier date. I believe GCC does it this way.

[ChuanqiXu9]

The major reason is that we haven't marked the status of Modules as complete now.

This feels inconsistent to me: you enable modules by default in -std=c++20 (and have dropped -fmodules-ts) but you don't signal in any way that they are available.

Sounds good. I'd like to define the macro later if no objection comes in.

[AaronBallman]

I think there's evidence that defining the macro at all when modules is not complete will cause problems in practice: https://sourcegraph.com/search?q=context:global+%23if+defined%28__cpp_modules%29+-file:.*test.*+-file:.*clang.*+-file:.*gcc.*&patternType=standard&case=yes&sm=1&groupBy=repo

In general, I think we don't want to define feature test macros until we're comfortable claiming we fully support the functionality. The one big exception to this is when system headers (mis)use the feature testing macros in a way that's critical for us to support.

[zowers]

https://sourcegraph.com/search?q=context:global+%23if+defined%28__cpp_modules%29+-file:.*test.*+-file:.*clang.*+-file:.*gcc.*&patternType=standard&case=yes&sm=1&groupBy=repo

it's only 39 results

[ChuanqiXu9]

BTW, the list of missing features in my mind is collected in #112295

And I am open for whether to not to define the macro now. Since although it is not complete in fact, modules are already being used. For example, this is an incomplete list I searched few days ago for the use of boost in modules:

https://github.com/infiniflow/infinity/blob/fa922afd58c973a9a317d2c09ad8cea8984af91e/src/common/boost.cppm#L17
https://github.com/ossia/score/blob/0ce6f62a97651a1248a92b3e13ce6ede10341d1f/src/plugins/score-plugin-avnd/halp.cppm#L74
https://github.com/RichardLuo0/make-dot-cpp/blob/e2ef10fd43060a2b56fc8f87f1407e4e3b1c2b8e/src/utils/module.cppm#L4
https://github.com/slaakko/cmajor-mod/blob/32e6eb8dc099b395fa751340de96f47817f92977/cmajor/system-x/machine/clock.cppm#L7
https://github.com/df-com/dragonfly/blob/9dfcbc1153f292eb8581f1df7459600cc331c690/dragonfly/Chart.cppm#L9
https://github.com/nofe1248/Helium/blob/92ec4452b8c05e6d86c972f843196a4392a7a6da/Events/EventBus/Helium.Events.EventBus.cppm#L20
https://github.com/LagrangeDev/liblagrange/blob/10e3b48f8d47fd65de00b14cfed64c0be18c083d/src/base/transport/service.cppm#L7
https://github.com/yudaichen/feature-db/blob/e94ade969af682ed5dd25d70bf282ba39253cbf0/src/common/small_utils.cppm#L7

So my logic is, if it is already used, it looks fine to me to define the feature macro. And similarly, if we don't want to define the macro now, I don't feel bad too. Since it looks it doesn't prevent people using it : )

[boris-kolpackov]

Still not time to define __cpp_modules by default?

It looks like __cpp_lib_modules is now defined by default in Clang 20 (#89579) and I feel many people will find this combination (modules as a feature not available while standard library modules are) puzzling.

FWIW, GCC 15 defines both.

[ChuanqiXu9]

Still not time to define __cpp_modules by default?

It looks like __cpp_lib_modules is now defined by default in Clang 20 (#89579) and I feel many people will find this combination (modules as a feature not available while standard library modules are) puzzling.

FWIW, GCC 15 defines both.

We're not technically implemented all features in modules. See #112294 (comment). So according to Aaron's words, it is better to not define it now.

According to my observation, I feel this may not prevent people using modules with clang.