swift-atomics uses non-standard _Atomic(bool) operations

[lorentey]

Swift Atomics currently exposes and/xor/or operations on booleans, and maps them to atomic_fetch_XXX_explicit calls in C. However, C17 does not define these operations on _Atomic(_Bool); it only defines them for other atomic integer types.

This is silently accepted by Clang, but it causes build issues when the _AtomicsShims module is compiled with GCC, such as with CMake.

Review Bool's atomic operations and figure out if we should deprecate loadThenLogicalOr and friends, or if we can continue exposing them (with this implementation or through compareExchange).

[lorentey]

For now, I'm going to #ifdef out the shims when compiled outside Swift's Clang Importer.

[glessard]

Interesting. I had relied on cppreference to be correct, and assumed I did not need to read the standard.

[stephentyrone]

This is silently accepted by Clang, but it causes build issues when the _AtomicsShims module is compiled with GCC, such as with CMake.

CMake builds not using a swift-toolchain-compatible clang seems like a bigger problem, since System and Atomics and Numerics would all benefit considerably from a known C environment.

[compnerd]

@stephentyrone - at the very least, MSVC needs to be supported as a C compiler. There are subtle bugs in clang/LLVM when it comes to some platforms, and so assuming a swift toolchain compatible clang is just as big an issue IMO.

[lorentey]

The current #if __swift__ clauses should nicely step around C compiler problems.

Meanwhile, #74 starts walking towards getting rid of the C module altogether. It'll require compiler and stdlib changes to land; I'm going to do my best to deliver them.

The use of non-standard _Atomic(bool) operations is an issue specific to the current _AtomicsShims module. We could easily fix that, but it does not currently cause problems, so I'm not going to bother.

Closing.

[lorentey]

#74 refactored things so that Bool now goes through Int8 atomics instead, even when we're using C atomics.