C++20 modules broken in clang++17

[ar-visions]

I cannot use this compiler for experiments. What I'm noticing is you cant implement a simple lambda without having to resort to random hacks to get the compiler to not delete default constructors.

Clang version 17.0.0 (https://github.com/llvm/llvm-project f66b9fe)

Can't use GCC either because it doesnt have AArch64 on Mac. Utterly stuck, and it means write a transpiler or goto Rust (same as many others).

a.cppm:

module;

#include <cstdlib>
#include <cstdio>
#include <cstdarg>
#include <new>
#include <memory>

export module a;

export template<typename T>
class func;

export template<typename R, typename... Args>
class func<R(Args...)> {
public:
    func() { };

    template<typename F>
    func(F&& func) : impl_(new Impl<typename std::decay<F>::type>(std::forward<F>(func))) {}

    R operator()(Args... args) const {
        if (impl_)
            return impl_->call(args...);
        throw std::runtime_error("Empty func object");
    }

    explicit operator bool() const {
        return static_cast<bool>(impl_);
    }

private:
    struct Base {
        virtual ~Base() { };
        virtual R call(Args... args) const = 0;
    };

    template<typename F>
    struct Impl : Base {
        explicit Impl(F&& func) : func_(std::forward<F>(func)) {}

        R call(Args... args) const override {
            return func_(args...);
        }

        F func_;
    };

    std::unique_ptr<Base> impl_ = nullptr;
};

// compiles ONLY when this line is uncommented.  its not even used
// -----------> Broken <------------
// func<void()> abc2 = func<void()>([]() -> void { });

export int foo() {
    return 43;
}

b.cppm:

export module b;
import a;
export int bar() {
    return 43;
}

CMakeLists.txt:

cmake_minimum_required(VERSION 3.26)
project(two_libraries CXX C ASM)

set(CMAKE_EXPERIMENTAL_CXX_MODULE_CMAKE_API "2182bf5c-ef0d-489a-91da-49dbc3090d2a")

include(../cxx_modules_rules_clang.cmake)

set(CMAKE_CXX_STANDARD 20)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CMAKE_CXX_EXTENSIONS OFF)

add_library(a STATIC)
target_sources(a
  PUBLIC
    FILE_SET cxx_modules TYPE CXX_MODULES FILES
    a.cppm
)

add_library(b STATIC)
target_sources(b
  PUBLIC
    FILE_SET cxx_modules TYPE CXX_MODULES FILES
    b.cppm
)

# The compiler won't find module 'a' if we delete the following line.
target_link_libraries(b PUBLIC a)

add_executable(main main.cpp)
target_link_libraries(main PUBLIC a b)

main.cpp:

import a;
import b;

int main() {
    foo();
    bar();

    func<void()> abc3 = func<void()>([]() -> void { });
    abc3();
    
    return 0;
}

build log with abc2 commented out:

kalen@Kalens-Mac-mini build% ninja
[8/10] Building CXX object CMakeFiles/main.dir/main.cpp.o
FAILED: CMakeFiles/main.dir/main.cpp.o 
/Users/kalen/Dropbox/src/llvm-project/build/bin/clang++   -std=c++20 -arch arm64 -isysroot /Library/Developer/CommandLineTools/SDKs/MacOSX13.1.sdk -MD -MT CMakeFiles/main.dir/main.cpp.o -MF CMakeFiles/main.dir/main.cpp.o.d @CMakeFiles/main.dir/main.cpp.o.modmap -o CMakeFiles/main.dir/main.cpp.o -c /Users/kalen/src/CMake-Modules-Examples/two_static_libraries/main.cpp
In file included from /Users/kalen/src/CMake-Modules-Examples/two_static_libraries/main.cpp:2:
/Users/kalen/src/CMake-Modules-Examples/two_static_libraries/a.cppm:20:22: error: no matching constructor for initialization of 'std::unique_ptr<Base>'
    func(F&& func) : impl_(new Impl<typename std::decay<F>::type>(std::forward<F>(func))) {}
                     ^     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/Users/kalen/src/CMake-Modules-Examples/two_static_libraries/main.cpp:9:25: note: in instantiation of function template specialization 'func<void ()>::func<(lambda at /Users/kalen/src/CMake-Modules-Examples/two_static_libraries/main.cpp:9:38)>' requested here
    func<void()> abc3 = func<void()>([]() -> void { });
                        ^
/Library/Developer/CommandLineTools/SDKs/MacOSX13.1.sdk/usr/include/c++/v1/__memory/unique_ptr.h:121:59: note: candidate constructor (the implicit copy constructor) not viable: no known conversion from 'Impl<typename std::decay<(lambda at /Users/kalen/src/CMake-Modules-Examples/two_static_libraries/main.cpp:9:38)>::type> *' (aka 'Impl<(lambda at /Users/kalen/src/CMake-Modules-Examples/two_static_libraries/main.cpp:9:38)> *') to 'const unique_ptr<Base>' for 1st argument
class _LIBCPP_UNIQUE_PTR_TRIVIAL_ABI _LIBCPP_TEMPLATE_VIS unique_ptr {
                                                          ^
/Library/Developer/CommandLineTools/SDKs/MacOSX13.1.sdk/usr/include/c++/v1/__memory/unique_ptr.h:185:21: note: candidate template ignored: requirement 'is_default_constructible<std::default_delete<func<void ()>::Base>>::value' was not satisfied [with _Dummy = true]
  _LIBCPP_CONSTEXPR unique_ptr(nullptr_t) _NOEXCEPT : __ptr_(pointer(), __default_init_tag()) {}
                    ^
/Library/Developer/CommandLineTools/SDKs/MacOSX13.1.sdk/usr/include/c++/v1/__memory/unique_ptr.h:190:12: note: candidate template ignored: requirement 'is_default_constructible<std::default_delete<func<void ()>::Base>>::value' was not satisfied [with _Dummy = true]
  explicit unique_ptr(pointer __p) _NOEXCEPT : __ptr_(__p, __default_init_tag()) {}
           ^
/Library/Developer/CommandLineTools/SDKs/MacOSX13.1.sdk/usr/include/c++/v1/__memory/unique_ptr.h:222:3: note: candidate template ignored: could not match 'unique_ptr<_Up, _Ep>' against 'Impl<typename std::decay<(lambda at /Users/kalen/src/CMake-Modules-Examples/two_static_libraries/main.cpp:9:38)>::type> *' (aka 'Impl<(lambda at /Users/kalen/src/CMake-Modules-Examples/two_static_libraries/main.cpp:9:38)> *')
  unique_ptr(unique_ptr<_Up, _Ep>&& __u) _NOEXCEPT
  ^
/Library/Developer/CommandLineTools/SDKs/MacOSX13.1.sdk/usr/include/c++/v1/__memory/unique_ptr.h:180:21: note: candidate constructor template not viable: requires 0 arguments, but 1 was provided
  _LIBCPP_CONSTEXPR unique_ptr() _NOEXCEPT : __ptr_(pointer(), __default_init_tag()) {}
                    ^
/Library/Developer/CommandLineTools/SDKs/MacOSX13.1.sdk/usr/include/c++/v1/__memory/unique_ptr.h:195:3: note: candidate constructor template not viable: requires 2 arguments, but 1 was provided
  unique_ptr(pointer __p, _LValRefType<_Dummy> __d) _NOEXCEPT
  ^
/Library/Developer/CommandLineTools/SDKs/MacOSX13.1.sdk/usr/include/c++/v1/__memory/unique_ptr.h:201:3: note: candidate constructor template not viable: requires 2 arguments, but 1 was provided
  unique_ptr(pointer __p, _GoodRValRefType<_Dummy> __d) _NOEXCEPT
  ^
/Library/Developer/CommandLineTools/SDKs/MacOSX13.1.sdk/usr/include/c++/v1/__memory/unique_ptr.h:210:3: note: candidate constructor template not viable: requires 2 arguments, but 1 was provided

when its not commented, it works.

I've been waiting years for Clang to do something with C++ modules and it seems unusable. I can't EXPERIMENT with it at all. It cant DO anything. Its like a laboratory that blows up when you step foot in it. Definitely not experimental as-is.

new T(args) is also broken, it tells you to use include new> but that doesnt fix it, as it still complains. Using ::new doesnt do it either. Its broken.

As is, it seems a waste of time to proceed with Clang. Not a single experiment works for me. So I wouldnt classify this code as 'experimental', I cant experiment with it as-is! Unless I want to experiment with errors doing anything modular in nature?

[llvmbot]

@llvm/issue-subscribers-clang-modules

[royjacobson]

Please be mindful of the LLVM community's code of conduct.

[ar-visions]

Please me give an example of this not being followed. I understand that and going to its very letter. My time is valuable and I define my own time as my own, and it has been set back by attempting to experiment with clang. So, clang has wasted my time, as investment into MSVCs working C++23 prototype may be a waste due to a lack of portability. Omitting harm is not some required code of conduct from what I understand? I've used Clang for years but this 2020 draft standard is not even at experimental phase, as you cant experiment with it in Clang++17. Thats 3 years past its draft date, and by compiler history that is unusual and noteworthy.

Using standard include and a templated wrapper case of it would be test #1 in 2020. Its striking that this simple example does not work. I only hope that its a compiler switch I am missing where its shaking away even when using modules: -x c++-module. My small case demonstrates the most important feature of modules and their exports seem broken, and I hadn't even begun experimenting yet. All optimizations done by Clang seem to be with the inclusion of a switch, so unless I am mistaken I don't see how I am doing this wrong.

Experimenting would be including std libraries and checking performance. I can't do step 1, so I can't experiment. At multiple years into a 2020 draft thats worth saying this issue makes it so I can't form generics in any possible way with modules (the point of the language). That means I cannot experiment with C++20 with clang++17 as-is. The laboratory picture makes sense. You can't experiment in a blown up laboratory. At 3 years into expected experimental basis you simply receive these messages describing the error behavior in detail (5k+)

[ar-visions]

I have tried using c++2b as well as c++20. that didnt help.

This is the modmap file which includes both pcms when compiling main.cpp
CMakeFiles/main.dir/main.cpp.o.modmap:

-fmodule-file=a=CMakeFiles/a.dir/a.pcm
-fmodule-file=b=CMakeFiles/b.dir/b.pcm

[ChuanqiXu9]

This should be a duplication for #59513.