<mutex> : std::mutex::mutex is not constexpr

[tiagomacarios]

This is mentioned in other issues, but I wanted a bug report tracking this explicitly. For the curious see #946.

std::mutex::mutex is constexpr as of c++11, but even in c++17 MSVC will not support it.
Minimal repro:
https://godbolt.org/z/5Yc56hEG7

#include <mutex>

struct S {
    constexpr S() noexcept = default;
    std::mutex m_mutex;
};

static constexpr S s{};

[AlexGuteniev]

It is not even documented. I added a PR form std::mutex documentation: MicrosoftDocs/cpp-docs#3467
But probably it also should be mentioned here: https://docs.microsoft.com/en-us/cpp/overview/visual-cpp-language-conformance?view=msvc-160

[AdamBucior]

Could we make shared_mutex constructor constexpr so that people who need it can have an alternative?

[AlexGuteniev]

Could we make shared_mutex constructor constexpr so that people who need it can have an alternative?

See #845:

We are explicitly forbidden from strengthening constexpr by WG21-N4861 16.5.5.7 [constexpr.functions]/1: "An implementation shall not declare any standard library function signature as constexpr except for those where it is explicitly required."

[StephanTLavavej]

Perhaps an LWG issue should be submitted to allow shared_mutex's constructor to be constexpr, as static initialization is preferable to dynamic initialization. As we just store nullptr, we would easily be able to implement such a resolution.

[CaseyCarter]

Perhaps an LWG issue should be submitted to allow shared_mutex's constructor to be constexpr, as static initialization is preferable to dynamic initialization. As we just store nullptr, we would easily be able to implement such a resolution.

Looks like it would be feasible for POSIX platforms as well (https://godbolt.org/z/nqxs1n9xn) if they implement shared_mutex naively.

[CaseyCarter]

Per the discussion in #2286, we believe mutex's default constructor can feasibly be made constexpr absent the need to support Vista. We'll need to figure out precisely how to wiggle the class definition so the in-memory representation of a default-constructed mutex looks exactly like the current result of dynamic-initialization on a Windows-7-or-later box.

[AlexGuteniev]

I've tried to look into it. So we need to implement Concurrency::details::stl_critical_section_win7 and _Mtx_internal_imp_t which contains it in a header. These are not _Ugly. Should we:

• Implement _Ugly version that does not match existing names?
• Make non-_Ugly names visible in headers?

[BillyONeal]

Per the discussion in #2286, we believe mutex's default constructor can feasibly be made constexpr absent the need to support Vista. We'll need to figure out precisely how to wiggle the class definition so the in-memory representation of a default-constructed mutex looks exactly like the current result of dynamic-initialization on a Windows-7-or-later box.

I don't believe this is true, at least for std::mutex's current guarantees, because the vtbl must live in the STL's DLL. Otherwise, mutexes created by a user DLL become radioactive if that user DLL is unloaded. (There's a similar problem in system_error for generic_category and system_category, but error_codes tend not to be super long-lived objects likely to trigger the problem vs. mutexes)

[AlexGuteniev]

Billy O'Neal (@BillyONeal) suggested a possible workaround for his concern: use __builtin_isconstant_evaluated() and do constant initialization to a new implementation only in constexpr context, while keeping old initialization for runtime.

Still there are problems in getting vtable class into mutex:

• mutex subobjects can't contain a vtable directly or via an union member due to is_standard_layout requirement
• vtable cannot be copied via __builtin_bit_cast due to requirement of trivially_copyable
• vtable cannot be set with placement new, as only allocating new is valid in constexpr context
• construct_at does not help either as there should be a pointer to vtable class, and reinterpreting casts won't work in constexpr
• fake vtable could have worked. That is a structure with pointers to functions. __thiscall methods can be emulated with __fastcall function. But this all is way too scary

[Fedr]

There is a question on Stackoverflow about this issue.

[pdimov]

FWIW, I just got a bug report against Boost.System caused by std::mutex's default constructor not using static initialization: boostorg/system#102.