enable <atomic> header on systems without thread-support

[ldionne]

Summary

This patch enables the <atomic> header on systems without thread support, such as for example ARMv7em baremetal targets.
It consists essentially of three changes:

1. Do not blindly refuse <atomic> when _LIBCPP_HAS_NO_THREADS is defined. I removed corresponding #error at the top of <atomic>. Furthermore, I changed the logic in <__config> where _LIBCPP_HAS_NO_ATOMIC_HEADER is determined to ignore _LIBCPP_HAS_NO_THREADS.
2. In <__config>, define _LIBCPP_ATOMIC_ONLY_USE_BUILTINS when _LIBCPP_HAS_NO_THREADS is defined (instead of just when _LIBCPP_FREESTANDING is defined), since there is no other way to implement <atomic> without thread support. Arguably, when building without thread support, one is effectively doing a freestanding compilation, so one could require the user to supply “-ffreestanding”. However, it seems that this is not usually done on the systems I work with. Furthermore, it is conceivable that one wants to build libc++ for a hosted platform that does not supports threads but still would want to have support for atomics.
3. In memory.cpp, change #ifdef _LIBCPP_HAS_NO_ATOMIC_HEADER to #ifdef _LIBCPP_HAS_NO_THREADS. Since this ifdef protects against the use of threading primitives, the latter seems more correct, while the former leads to compilation errors with the new logic for _LIBCPP_HAS_NO_ATOMIC_HEADER. It seems to indicate that the implementation of <memory> uses threading primitives instead of atomics. Therefore, in standalone builds, where concurrency comes in the form of interrupts rather than threads, one has to be careful with the use of <memory>. If at some point libc++ on baremetal becomes officially supported, that issue should probably be documented somewhere.

Of course, so far my patch fails to provide tests for the issue, which I’m not sure how to proceed with. I have never run any of the LLVM tests myself, as this seems to be extra challenging for cross-builds for baremetal. I did manually test the built libc++ in a simple (“blinky”) project that uses hardware timer interrupts to trigger state changes, relying on <atomic> for synchronisation. It appears to work, contrary to the non-atomic version of it.

[__simt__]

I think we want people to use -ffreestanding more for things like this, so the part where this is circumvented isn't clearly the right choice. The rest seems fine, and is exactly the reason why I created _LIBCPP_ATOMIC_ONLY_USE_BUILTINS.

[burnpanck]

My reasoning is the following: If _LIBCPP_ATOMIC_ONLY_USE_BUILTINS were only implied by _LIBCPP_FREESTANDING but not by _LIBCPP_HAS_NO_THREADS, the following paradoxical situation would appear for platforms without threads:

• libc++ *with* -ffreestanding would be a conforming freestanding implementation, but not a conforming hosted implementation (no threads)
• libc++ *without* -ffreestanding would still not be a conforming hosted implementation, but now it wouldn't even be a conforming freestanding implementation (because atomics are now missing).

Thus, the flag -ffreestanding, whose purpose is to *lower* the conformance level actually leads to a *higher* conformance level than without! The standard requires atomics in all conforming cases, whereas it would allow to disable threads under -ffreestanding (but it doesn't require it).

To me, -ffreestanding currently is ill-specified and has no good use. As far as I understand, the requirements for freestanding implementations are strictly smaller than for hosted implementations, but allowing implementation-defined extension. A conforming freestanding implementation is allowed to be a conforming hosted implementation, but it is not required to be. Thus, it is not clear what -ffreestanding even should do. It could be understood that *lack of -ffreestanding* requests a fully conforming hosted implementation; in that case, libc++ should choke on any platform that does not support threads unless -ffreestanding is given.

[ldionne]

This change makes sense to me as-is. I agree with @__simt__ that we want people to use -ffreestanding in the future, however I'm also with @burnpanck on the fact that Freestanding is currently ill-defined (both in the Standard and in existing practice -- nobody seems to have agreement on what exactly is in there).

Since we currently allow disabling the use of threads without using -ffreestanding, I think it makes sense to _not_ require -ffreestanding to enable the builtin-only atomics. Once we figure out what exactly is in freestanding, we can change that and guard all those features by #if defined(_LIBCPP_FREESTANDING) && <existing feature> (or something along those lines), with the goal of requiring the use of -ffreestanding to enable those features. @__simt__ Would you still rather requiring the use of -ffreestanding to enable builtin-only atomics?

[ldionne]

Adding @EricWF and @mclow.lists for awareness.

[burnpanck]

Friendly ping. Is there anything I can help with? In the meantime, I have been able to get some code to run, utilising this on a Cortex M4 in a bare-metal environment. To get it to work I have to patch LLVM 9 clang and libc++ in a few other places anyway, so there is no immediate need to get this upstreamed from my side - but I'd definitely would love to see LLVM as a viable alternative to GCC for such tasks at some point.

[ldionne]

@burnpanck I'm willing to look at this again if you're willing to rebase on top of master, FWIW. There might be changes required since we landed the C++20 Synchronization Library.

[ldionne]

[Github PRs transition cleanup]

Abandoning since this needs a heavy rebase.

libcxx/include/__config

 #if __has_feature(cxx_atomic) || __has_extension(c_atomic) || __has_keyword(_Atomic)
 #  define _LIBCPP_HAS_C_ATOMIC_IMP
 #elif defined(_LIBCPP_COMPILER_GCC)
 #  define _LIBCPP_HAS_GCC_ATOMIC_IMP
 #endif
 
 #if (!defined(_LIBCPP_HAS_C_ATOMIC_IMP) && \
      !defined(_LIBCPP_HAS_GCC_ATOMIC_IMP) && \
-     !defined(_LIBCPP_HAS_EXTERNAL_ATOMIC_IMP)) \
-     || defined(_LIBCPP_HAS_NO_THREADS)
+     !defined(_LIBCPP_HAS_EXTERNAL_ATOMIC_IMP))
 #  define _LIBCPP_HAS_NO_ATOMIC_HEADER
 #else
 #  ifndef _LIBCPP_ATOMIC_FLAG_TYPE
 #    define _LIBCPP_ATOMIC_FLAG_TYPE bool
 #  endif
-#  ifdef _LIBCPP_FREESTANDING
+#  if defined(_LIBCPP_FREESTANDING) || defined(_LIBCPP_HAS_NO_THREADS)
 #    define _LIBCPP_ATOMIC_ONLY_USE_BUILTINS
 #  endif
 #endif
 
 #ifndef _LIBCPP_DISABLE_UBSAN_UNSIGNED_INTEGER_CHECK
 #define _LIBCPP_DISABLE_UBSAN_UNSIGNED_INTEGER_CHECK
 #endif
 

libcxx/include/atomic

 #include <cstdint>
 #include <type_traits>
 #include <version>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #pragma GCC system_header
 #endif
 
-#ifdef _LIBCPP_HAS_NO_THREADS
-# error <atomic> is not supported on this single threaded system
-#endif
 #ifdef _LIBCPP_HAS_NO_ATOMIC_HEADER
 # error <atomic> is not implemented
 #endif
 #ifdef kill_dependency
 # error C++ standard library is incompatible with <stdatomic.h>
 #endif
 
 #define _LIBCPP_CHECK_STORE_MEMORY_ORDER(__m) \

libcxx/src/memory.cpp

 const void*
 __shared_weak_count::__get_deleter(const type_info&) const _NOEXCEPT
 {
     return nullptr;
 }
 
 #endif  // _LIBCPP_NO_RTTI
 
-#if !defined(_LIBCPP_HAS_NO_ATOMIC_HEADER)
+#if !defined(_LIBCPP_HAS_NO_THREADS)
 
 _LIBCPP_SAFE_STATIC static const std::size_t __sp_mut_count = 16;
 _LIBCPP_SAFE_STATIC static __libcpp_mutex_t mut_back[__sp_mut_count] =
 {
     _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER,
     _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER,
     _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER,
     _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER, _LIBCPP_MUTEX_INITIALIZER
static __sp_mut muts[__sp_mut_count]
         &mut_back[ 0], &mut_back[ 1], &mut_back[ 2], &mut_back[ 3],
         &mut_back[ 4], &mut_back[ 5], &mut_back[ 6], &mut_back[ 7],
         &mut_back[ 8], &mut_back[ 9], &mut_back[10], &mut_back[11],
         &mut_back[12], &mut_back[13], &mut_back[14], &mut_back[15]
     };
     return muts[hash<const void*>()(p) & (__sp_mut_count-1)];
 }
 
-#endif // !defined(_LIBCPP_HAS_NO_ATOMIC_HEADER)
+#endif // !defined(_LIBCPP_HAS_NO_THREADS)
 
 void
 declare_reachable(void*)
 {
 }
 
 void
 declare_no_pointers(char*, size_t)