Add support for intrinsics for NTuple{VecElement}

[oscardssmith]

This is the start towards proper SIMD support in Base. Currently the main things missing are support for masked/predicated instructions, and an intrinsic to load an arbitrarily chosen sized NTuple{n, VecElement} from a Memory. Many thanks to Alexandre Prieur for pair programming this with me, and @vtjnash, @gbaraldi and @vchuravy for answering the 50 million questions about C++/LLVM/debugging.

julia> f(a,b) = Core.Intrinsics.mul_float(a, b)
julia> a = ntuple(i->VecElement(sqrt(i)), 8);

julia> @code_llvm f(a,a)
; Function Signature: f(NTuple{8, VecElement{Float64}}, NTuple{8, VecElement{Float64}})
;  @ REPL[1]:1 within `f`
define <8 x double> @julia_f_949(<8 x double> %"a::Tuple", <8 x double> %"b::Tuple") #0 {
top:
  %0 = fmul <8 x double> %"a::Tuple", %"b::Tuple"
  ret <8 x double> %0
}

[tecosaur]

As one of the people rather excited about this, I'm wondering whether I might hope to see this progress to a non-draft PR in the near future, or whether this has some blocker/major further work needed? 🙂

[oscardssmith]

I'm unlikely to have time to do all the things that likely will be necessary to finish this off. there's no blockers, just a bunch more work.

[giordano]

Would this include vscale? 👀

[KristofferC]

It is not obvious to me that these should be defined. When you are doing low level SIMD stuff you probably don't want to accidentally promote things and in case where you really want to work with different types, an explicit convert might be better for clarity?