[llvm-dev] __c11_atomic builtins' input requirements

[Ben Trapani via llvm-dev]

I have started working on an implementation of atomic_ref. Implementing
this requires performing atomic operations on arbitrary references. The
behavior is undefined if a reference is provided that is not aligned per
the public export required_alignment. What assumptions do the __c11_atomic
builtins make about destination argument alignment, format, size etc.? Is
it required that the destination argument be annotated as _Atomic like the
corresponding standard c11 functions require?

The following code in libcxx/include/config implies that the c11 atomic
builtins should be substituted for gcc atomic builtins if present:

#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

To implement atomic_ref, it might be cleaner to use the gcc builtins. By
doing so, we avoid depending on the internals of the c11 atomic builtins.
The required alignment of referenced variables would always be the first
power of 2 >= max(type_size, type_alignment). Alternatively, we could use
the c11 atomic builtins when available if it is enough to require that the
input reference is aligned as the atomic type would be. What are your
thoughts?

Regards,
Ben
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20191228/d6b7a740/attachment.html>