Fix getindex inlining for Array{Union{T,Missing}}

[nalimilan]

Fixes #27403.

With this PR, a naive sum over non-missing values in a Vector{Union{Int,Missing}} is (probably) as fast as it could theoretically be. Masked SIMD instructions are used, which is almost too good to be true.

function mysum(X)
    s = zero(eltype(X))
    @inbounds @simd for x in X
        if x !== missing
            s += x
        end
    end
    s
end

X1 = rand(Int, 10_000_000);
X2 = Vector{Union{Missing,Int}}(X1);
X3 = ifelse.(rand(length(X2)) .< 0.9, X2, missing);

julia> @btime mysum(X1);
  5.787 ms (1 allocation: 16 bytes)

julia> @btime mysum(X2);
  5.726 ms (1 allocation: 16 bytes)

julia> @btime mysum(X3);
  5.844 ms (1 allocation: 16 bytes)

I've added a test to check that getindex is inlined, but I'm really not sure that's the best way to do it. There are also benchmarks for this in BaseBenchmarks, and I'll add another one (EDIT: see JuliaCI/BaseBenchmarks.jl#207).

[davidanthoff]

I get slightly different numbers, in particular the X1 case is faster than the other two (which kind of makes sense?):

julia> @btime mysum(X1);
  4.812 ms (1 allocation: 16 bytes)

julia> @btime mysum(X2);
  5.757 ms (1 allocation: 16 bytes)

julia> @btime mysum(X3);
  5.753 ms (1 allocation: 16 bytes)

[nalimilan]

Yes, the particular run I posted was a bit extreme. The one I added to the blog post is more similar to yours. I have no idea why it changed.

[vtjnash]

@nanosoldier runbenchmarks(ALL, vs = "@b331c70642e1f792577f2b591a4e5016e3c2d53f")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan

[nalimilan]

Some probably spurious regressions, lots of Union-related improvements (about 2×-3×), and some very dramatic speedups. This one is particularly nice:

ID	time ratio	memory ratio
["union", "array", "("perf_sum", Int8, true)"]	0.04 (15%)	1.00 (1%)