Slow push! and unshift! for large number of arguments

[bkamins]

The following code is slow (even if the push! and unshift! functions are run several times):

x = [1, 2, 3]
push!(x, (1:10000)...) # this is very slow
unshift!(x, (1:10000)...) # this is also very slow

compared to:

x = [1, 2, 3]
for i in 1:10000
    push!(x, i)
end
for i in 1:10000
    unshift!(x, i)
end

For smaller number of values (less than 10000) the speed of current implementation is not prohibitively low, but slower than loop.

The problem was encountered in NullableArrays (PR JuliaStats/NullableArrays.jl#110) and the solution comment by @nalimilan requires fixing the cases mentioned above.

[andreasnoack]

The issue is the splatting because you effectively call a function with 10001 arguments. Wouldn't prepend! work for the problem in NullableArrays?

[bkamins]

I will have to check. But see below that the loop is reasonably fast even with splatting so why not fix it?

x = [1, 2, 3]
@time push!(x, (1:10000)...)
# over 2s
x = [1, 2, 3]
@time push!(x, (1:10000)...)
# over 2s

function npush!(x, vs...)
    for i in vs
        push!(x, i)
    end
    return x
end

x = [1, 2, 3]
@time npush!(x, (1:10000)...)
# less than 0.01s
x = [1, 2, 3]
@time npush!(x, (1:10000)...)
# less than 0.001s

[yuyichao]

Calling a function with splatting of an container is a bad idea. The multi-args push and unshift are optimized for a small number of arguments. append exists for a reason.

[nalimilan]

Sorry if my comment on the other issue wasn't explicit enough, but I meant that there was no performance issue here.