View allocation elimination stumped by conditional?

[KristofferC]

Tracked down after performance report on Distances.jl (JuliaStats/Distances.jl#100)

Code to reproduce:

module Distances2

function colwise!(r, a, b)
    @inbounds for j = 1:size(a,2)
        r[j] = evaluate(view(a, :, j), view(b, :, j))
    end
    r
end

@inline function evaluate(a, b)
    length(a) == 0 && return 0.0 # comment out and 0.7 is super fast
    @inbounds begin
        s = 0.0
        @simd for I in eachindex(a, b)
            ai = a[I]
            bi = b[I]
            s += abs2(ai - bi)
        end
        return s
    end
end

end

using BenchmarkTools
z = zeros(41); A = rand(2, 41); B = rand(2, 41);
@btime Distances2.colwise!($z, $A, $B);

On 0.7:

 562.371 ns (82 allocations: 3.84 KiB)

On 0.6:

 255.466 ns (0 allocations: 0 bytes)

Now, if we comment away the length(a) == 0 && return 0.0 line:

0.7:

100.298 ns (0 allocations: 0 bytes)

0.6:

 206.193 ns (0 allocations: 0 bytes)

[KristofferC]

Actually, changing out eachindex(a, b) to 1:size(a, 2) also seems to help.

[Keno]

Smells like the inlining heuristic change.

[KristofferC]

I checked out the commit before #27857 and it didn't change anything.

[KristofferC]

In fact, manually inlining shows the same regression:

using BenchmarkTools

function colwise!(r, A, B)
    @inbounds for j = 1:size(A,2)
        a = view(A, :, j)
        b = view(B, :, j)
        r[j] = begin
            if #= false && =# length(a) == 0  # Remove comment and fast
                0.0
            else
                @inbounds begin
                    s = 0.0
                    for I in eachindex(a, b)
                        ai = a[I]
                        bi = b[I]
                        s += abs2(ai - bi)
                    end
                    s
                end
            end
        end
    end
    r
end


using BenchmarkTools
z = zeros(41); A = rand(2, 41); B = rand(2, 41);
@btime colwise!($z, $A, $B);

With and without false commented out respectively

  115.017 ns (0 allocations: 0 bytes)

  589.823 ns (82 allocations: 3.84 KiB)

So I think inlining can be ruled out.

[KristofferC]

Changing the if to a ifelse makes the allocation go away, I will use that to work around the original issue.

Edit: Not enough on the original issue though. Removing @simd make allocations go away (but then no simd).....

Edit2: The final fix was to change eachindex(a,b) to 1:length(a) which was safe in this specific case (https://github.com/JuliaStats/Distances.jl/pull/101/files#diff-08d86a8ef0a417f2c2269c73f1dbfe5b).

[KristofferC]

Fixed, now zero allocations.