Possible improvement to dynamic dispatch (breaking)

[AntoinePrv]

Right now this is the code the dispatcher runs on every call:

xsimd/include/xsimd/config/xsimd_arch.hpp

Lines 223 to 227 in 6b61b1a

	template <class... Tys>
	XSIMD_INLINE auto operator()(Tys&&... args) noexcept
	{
	return walk_archs(ArchList {}, std::forward<Tys>(args)...);
	}

There is some recursive template machinery, coupled with runtime checks on supported_arch::has().
There is potentially quite a few steps for the compiler to prove that it can replace that with a direct jump (to be further investigated).

In Apache Arrow, we use an internal dispatch mechanism where we evaluate available architecture once and store the function pointer (during a dispatcher static initialization).
Here however, we do not have a unique type for all candidate function because the first parameter, the arch, are distinct struct.

xsimd/include/xsimd/config/xsimd_arch.hpp

Lines 210 to 211 in 6b61b1a

	if (availables_archs.has(Arch {}))
	return functor(Arch {}, std::forward<Tys>(args)...);

Example from the doc:

struct sum
{
    template <class Arch, class T>
    T operator()(Arch, T const* data, unsigned size);
};

Now the proposed breaking change would be to take a free function templated by arch, but not using it as a parameter.
I am not immediately sure how we'd handle T but I think we can manage.

  template <class Arch>
  T sum1(float const* data, unsigned size);

Or perhaps with a static method of a functor.

 template <class Arch>
struct sum2
{
    template <class T>
    static T call(T const* data, unsigned size);
};

In both cases there is an underlying function pointer that can be directly stored (sum1<sse4_2>/sum1<avx2> or sum2<sse4_2>::call<float>/sum2<avx2>::call<float>).

@serge-sans-paille do you think it is worth investigating? I can try to find a bit more time to look at the current generated assembly.
CC @JohanMabille

[DiamonDinoia]

In my c++ code I usually dispatch the constructor using xsimd and fill a unique pointer the way you presented.

Maybe this can inspire you https://github.com/DiamonDinoia/random/blob/main/include/random/xoshiro_simd.hpp

That implementation might change when I have time to find a better way so any suggestion is welcome