[clang] builtin `__cxx_atomic_fetch_add` should support all floating point types

[huixie90]

This issues came up in implementing P0020R6 Floating Point Atomic for libc++
#67799

Basically we are trying to use __cxx_atomic_fetch_add for implementing atomic RMW operations. However, this function would error on some platforms for long double if the representation is x87 fp80 format. On the library (libc++) side, there is no way for us to tell if the function __cxx_atomic_fetch_add is going to error on the current platform or not.

The builtin should handle this properly: generates FADD IR for platforms that support it, and fallback to the compare-exchange loop on the platforms that don't

[llvmbot]

@llvm/issue-subscribers-bug

This issues came up in implementing P0020R6 Floating Point Atomic for libc++ https://github.com//pull/67799

Basically we are trying to use __cxx_atomic_fetch_add for implementing atomic RMW operations. However, this function would error on some platforms for long double if the representation is x87 fp80 format. On the library (libc++) side, there is no way for us to tell if the function __cxx_atomic_fetch_add is going to error on the current platform or not.

The builtin should handle this properly: generates FADD IR for platforms that support it, and fallback to the compare-exchange loop on the platforms that don't