[Mono] Consider switching to C11

[lambdageek]

Currently Mono is supposed to use C99. As a result, we're relying on platform or compiler intrinsics for some functionality that has been added to the standard since then such as stdatomic.h.

We should consider upgrading Mono to a more recent language standard.

This came up as part of #88279, for example.

Some concrete things we could take advantage of:

• stdatomic.h for atomic reads/writes, CAS, etc
• _Alignas
• static assertions
• _Noreturn
• possibly thread locals (although the JIT likes to take advantage of understanding TLS layout, so we need to ensure we don't take a perf hit)

Some reasons we might not want to do it:

• Anecdotally, I've heard that MSVC C11 support has some issues - we will need to understand the specific pitfalls to avoid so we don't break mono-aot-cross running on Win32 and also Mono on non-Windows Win32 platforms

[lambdageek]

cc @vargaz @AaronRobinsonMSFT @lateralusX @am11

[AaronRobinsonMSFT]

/cc @jkoritzinsky

[MichalPetryka]

• I've heard that MSVC C11 support has some issues

MSVC doesn't support C11, only C17.

[AaronRobinsonMSFT]

• I've heard that MSVC C11 support has some issues

MSVC doesn't support C11, only C17.

Yes it does - https://devblogs.microsoft.com/cppblog/c11-and-c17-standard-support-arriving-in-msvc/. There is no support for optional features and there are bugs around atomics, but it has support.

[am11]

In libraries, we are using C11 for couple of years:

runtime/eng/native/configurecompiler.cmake

Lines 10 to 11 in 376e235

	set(CMAKE_C_STANDARD 11)
	set(CMAKE_C_STANDARD_REQUIRED ON)

and e.g. its 'Generic Selection' feature.:

runtime/src/native/libs/System.Native/pal_process.c

Line 559 in 376e235

#define LIMIT_MAX(T) _Generic(((T)0), \

VCR handles it just fine. Only these features are not supported: https://devblogs.microsoft.com/cppblog/c11-and-c17-standard-support-arriving-in-msvc/#whats-not which we don't use in libs.

[AaronRobinsonMSFT]

The atomics in C11 is recent and experimental, announced at https://devblogs.microsoft.com/cppblog/c11-atomics-in-visual-studio-2022-version-17-5-preview-2/.

[barcharcraz]

Hi, I'm on the C++ libraries team and I'm the one who wrote the library support for c11 atomics and threads. We actually thought of providing gcc style atomic intrinsics that could work in c99 mode, but because of how our C++ atomics are designed MSVC really, really needs that new _Atomic keyword.

C11 threads will probably work fine in C99 mode, with the possible exception of the new _Thread_local/thread_local keyword. The TLS layout for implicit TLS is the same as __declspec(thread) or c++ thread_local. For the tss_* family of functions the implementation is new, and not shared with the Tls* or Fls* functions, because tss_* needs to support destructors with slightly different semantics than fls. The layout used here is very basic, with an array for global data and a per-thread array for per-thread data that's allocated when you use a relevant function for the first time on a particular thread.

All the C11 threads stuff is new, and distinct from C++ threads. The ABI is also somewhat better protected than for C++ threads, so layouts and implementation are subject to change (in ABI compatible ways).

[barcharcraz]

oh, I should say that C11 threads is just implemented in terms of windows threading functions (CreateThread and friends), and its threads are "normal" threads and don't have any extra state or anything. They do briefly allocate some memory in order to thunk between cdecl and stdcall, but that's it.

Note that this does mean thrd_t thread handles work a little differently for our C11 threads than they tend to in pthreads based implementations. In particular you can't detach yourself by writing thrd_detach(thrd_current()).

[tgani-msft]

The atomics in C11 is recent and experimental, announced at https://devblogs.microsoft.com/cppblog/c11-atomics-in-visual-studio-2022-version-17-5-preview-2/.

It's currently being actively worked on (both front-end and back-end teams) - we can talk about .NET taking a dependency on this support in our sync meetings.

[lambdageek]

Thinking about _Atomic a little bit more, I think once we bump to C11, the guidance in Mono will be not to use _Atomic(T) or the _Atomic qualifier. There's no good compile-time way to know if that will be lockfree (sadly static_assert (atomic_is_lock_free (&someAtomicStruct)) cannot work). In general we cannot use locks inside Mono (in the runtime itself, I mean - pinvokes are fine) without making the coop GC thread state machine aware of them - otherwise we will deadlock if two threads try to take the lock at the same time that a GC is started

runtime/src/mono/mono/utils/checked-build.c

Lines 207 to 221 in cd4a253

	/* Suppose we're using hybrid suspend and T1 is in GC Unsafe and T2 is
	* GC Safe. T1 will be coop suspended, and T2 will be async suspended.
	* Suppose T1 is in RUNNING, and T2 just changed from RUNNING to
	* BLOCKING and it is in trace_state_change to record this fact.
	*
	* suspend initiator: switches T1 to ASYNC_SUSPEND_REQUESTED
	* suspend initiator: switches T2 to BLOCKING_SUSPEND_REQUESTED and sends a suspend signal
	* T1: calls mono_threads_transition_state_poll (),
	* T1: switches to SELF_SUSPENDED and starts trace_state_change ()
	* T2: is still in checked_build_thread_transition for the RUNNING->BLOCKING transition and calls backtrace ()
	* T2: suspend signal lands while T2 is in backtrace() holding a lock; T2 switches to BLOCKING_ASYNC_SUSPENDED () and waits for resume
	* T1: calls backtrace (), waits for the lock ()
	* suspend initiator: waiting for T1 to suspend.
	*
	* At this point we're deadlocked.

So I think at best we will end up redefining mono_atomic_cas_TYPE, mono_atomic_compare_exchange_TYPE, etc in terms of the C11 atomic ops (and adding something along the lines of static_assert ((sizeof(int32_t) == sizeof(atomic_int) && ATOMIC_INT_LOCK_FREE == 2) || (sizeof(int64_t) == sizeof(atomic_llong) && ATOMIC_LLONG_LOCK_FREE == 2)) somewhere in our headers)

[lambdageek]

This initial work has been completed.

Mono now requires a C11 compiler - and we use static_assert in some places.

Some platforms now opt into using C11 atomics for the mono_atomic_XYZ functions.

We are not using c11 atomics for MSVC builds yet, but we could switching. Created #91748 for follow-up work.

We are using "strong" not "weak" atomics. Created #91747 for follow-up work.

[lambdageek]

We're reverting the atomics PR because of some build failures #91812

A proper fix is in progress in #91808 but it's failing due to some weird clang problem on our linux-arm64 crossrootfs

[lewing]

@lambdageek what is the status of this?

[lambdageek]

We're done. #91808 was good.