Use propagate_inbounds to restore functionality on 0.7

[vchuravy]

• uses @propagate_inbounds
• drops support for Julia 0.4, and 0.5
• removes deprecation warnings

fixes #10

[KristofferC]

FWIW, it seems like LLVM now uses SIMD even for the regular array.

[vchuravy]

Might be the SLPVectoriser? But in general SoA is still valuable since it saves a bunch of scatter/gather operations

[KristofferC]

Yeah, I didn't notice a big difference now though (using https://github.com/piever/StructArrays.jl/)

using StructArrays
regular = randn(1000000) + im*randn(1000000);
soa = convert(StructArray, regular);

function f(x, a)
    s = zero(eltype(x))
    @simd for i in 1:length(x)
        @inbounds s += x[i] * a
    end
    s
end

julia> @btime f(regular, 0.5+0.5im)
  929.258 μs (2 allocations: 64 bytes)

julia> @btime f(soa, 0.5+0.5im);
  960.669 μs (2 allocations: 64 bytes)

When using @fastmath the soa version goes over to use 256 bit SIMD registries though (while something terrible happens for the regular version):

function f_fast(x, a)
    s = zero(eltype(x))
    @simd for i in 1:length(x)
        @inbounds @fastmath s += x[i] * a
    end
    s
end

julia> @btime f_fast(regular, 0.5+0.5im)
  2.957 ms (2 allocations: 64 bytes)

julia> @btime f_fast(soa, 0.5+0.5im)
  701.956 μs (2 allocations: 64 bytes)

[KristofferC]

I can repeat the timings above with this package as well so it doesn't seem to be an implementation detail of the specific SoA package.