ccall should pass vectorized tuples by value instead of reference

[ArchRobison]

While writing the docs for #15244, I discovered that ccall is passing tuples of VecElement by reference instead of value. Passing by value would be more consistent with the C ABI for SIMD types, e.g. pass a NTuple{8,VecElement{Float32}} as if it were a C AVX __m256.

I'll poke around ccall to see what I need to teach it. Below is the example I was playing with.

typealias m256 NTuple{8,VecElement{Float32}}

convert(::Type{VecElement{Float32}},x::Float32) = VecElement(x)

a = m256(ntuple(i->VecElement(sin(Float32(i))),8))
b = m256(ntuple(i->VecElement(cos(Float32(i))),8))
println(typeof(a))
println(typeof(b))

function call_dist(a::m256, b::m256)
    ccall((:dist, "libdist"), m256, (m256, m256), a, b)  # C prototype is __m256 dist( __m256 a, __m256 b ) 
end

@code_llvm call_dist(a,b)

[ArchRobison]

Looks like the fix involves changing the Classification logic to understand the special tuples. Can someone explain the purpose of the two classes fields here in abi_x86_64.cpp?

struct Classification {
    bool isMemory;
    ArgClass classes[2];

[yuyichao]

Right. All the abi needs to be updated.

[ArchRobison]

What's the difference between classes[0] and classes[1]?

[ArchRobison]

Closed by #16258.