Make atomic_ptr optional for active_implementation? about simdutf

[lemire]

I don’t see how you can use simdutf without runtime dispatching. The implementations are system specific.

How would that work? How does the aarch64 implementation/kernel kicks in if you do not detect it?

[TerrorJack]

Ah, I really mean using a regular pointer for runtime dispatching instead of using an atomic, sorry for the confusion.

[lemire]

@TerrorJack We use an atomic to ensure thread safety. We are open to changing this code, but one needs to prove that thread safety is maintained.

Honestly, a C++ platform that does not support atomics does not seem like a serious platform. Are you sure it is not supported?

[TerrorJack]

We are open to changing this code, but one needs to prove that thread safety is maintained.

Platforms that do support atomic will continue to use it by default.

Honestly, a C++ platform that does not support atomics does not seem like a serious platform. Are you sure it is not supported?

I'm sure wasm32-wasi doesn't have threads or atomics at the moment. The bundled libc does not have pthread, and the bundled libc++ does not ship anything threading related. And it is a serious platform! (well, at least serious enough for me to get paid.)

[WojciechMula]

Daniel, since the standard allows std::hardware_concurrency to return 0, I wouldn't be worried about thread safety. ;)

I'd be happy if we end up with some macro definition that explicitly enables the use of bare pointers. However, by default we should assume <atomic> is there.

[WojciechMula]

@TerrorJack It would nicer if we ended up with the following code:

#ifdef SIMDUTF_HAS_ATOMIC
template <typename T>
using simdutf::atomic = std::atomic;
#else
template <typaneme T>
using simdutf::atomic = dummy::atomic;
#endif

Where dummy::atomic is a template providing the methods from std::atomic we actually use.

[TerrorJack]

@WojciechMula IMHO adding dummy::atomic logic would make the PR much more bloated. I'll let it sit for a bit longer to gather more opinions.

[WojciechMula]

@TerrorJack technically your solution is fine. However, there are too many ifdefs --- five. :) There should be exactly one place in the code where we use an ifdef. Please rethink your approach. It'd be fine if we introduce a new header file for atomics.