<complex>: use hardware FMA on Clang too

[Alex Guteniev (AlexGuteniev)]

Yes it does not compile without [[gnu::target("fma")]] magic spell.

[Stephan T. Lavavej (StephanTLavavej)]

Thanks for figuring out the magic incantation here! 😻

I observe that this is inconsistent with the spelling of

STL/stl/inc/atomic

Line 33 in 43e96b2

#pragma clang attribute _STD_ATOMIC_HEADER.push([[gnu::target("cx16")]], apply_to = function)

but I don't care enough to iterate.

[Alex Guteniev (AlexGuteniev)]

I observe that this is inconsistent with the spelling of

If you want consistency, I'd change <atomic> actually.

The spell in <atomic> allows cx16 insn in any function, and this theoretically permits generating one out of thin air. In contrast, the approach in this PR allows fma only in specific function, for which we have runtime check.

<atomic> does not centralize cx16, and there would be few functions, so it takes more changes, but still may be worth doing.

[Stephan T. Lavavej (StephanTLavavej)]

Oh, all I meant was the spelling of __gnu__::__target__.

[Alex Guteniev (AlexGuteniev)]

Aha! It was suggested by Nikolas Klauser (@philnik777). Makes sense as a defense from user's macro, #define gnu 1 is allowed. And in pragma there's no preprocessor, since pragma itself is considered preprocessor.

[Stephan T. Lavavej (StephanTLavavej)]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[Stephan T. Lavavej (StephanTLavavej)]

Thanks for multiplying and adding to the compilers where we use these intrinsics! 😹 ✖️ ➕