Make broadcast eltype widening type-stable

[cstjean]

The broadcast code that copies to the new array was type-unstable, causing allocation and slowdown in v2:

Before:

julia> using BenchmarkTools

julia> v = fill(2.0, 50000);

julia> v2 = vcat(v, missing);

julia> v3 = vcat(missing, v);

julia> @btime $v .* 3.0;
  49.728 μs (2 allocations: 390.70 KiB)

julia> @btime $v2 .* 3.0;
  1.490 ms (99498 allocations: 2.33 MiB)

julia> @btime $v3 .* 3.0;
  103.577 μs (7 allocations: 439.77 KiB)

After:

julia> @btime $v .* 3.0;
  50.147 μs (2 allocations: 390.70 KiB)

julia> @btime $v2 .* 3.0;
  147.158 μs (8 allocations: 830.45 KiB)

julia> @btime $v3 .* 3.0;
  92.611 μs (7 allocations: 439.77 KiB)

This fixes the second half of #30455. First half was fixed by #30480

[mbauman]

Nice. Makes sense. I know it's a small cost, but since we're outlining here, perhaps we should just outline everything after that similar call. E.g.,

newdest = Base.similar(dest, promote_typejoin(T, typeof(val)))
return restart_copyto_nonleaf!(newdest, dest, bc, val, I, iter, state, count)

function restart_copyto_nonleaf!(newdest, dest, bc, val, I, iter, state, count)
    # Function barrier that makes the copying to newdest type stable
    for II in Iterators.take(iter, count)
        newdest[II] = dest[II]
    end
    newdest[I] = val
    return copyto_nonleaf!(newdest, bc, iter, state, count+1)
end

That should just have one dynamic dispatch instead of three. In my quick tests of the cases you listed, it looks like it saves another ~25%.

[cstjean]

Sure, I just took your code and updated the PR. I updated the benchmarks.

[mbauman]

The Travis failures look unrelated — 64 bit linux is an assertion failure in the Profile tests and Mac is a Pkg timeout.

[cstjean]

Awesome, could it be backported to 1.1?

[mbauman]

We're trying to be a bit more conservative in backporting performance fixes, but this is about as straightforward as a patch could be — it's literally just copy-pasting code into an outlined function. I'd support a backport.