wrong derivative with array concatenation

[andrewning]

My code used to work in ReverseDiff 1.2.0 but the current version (1.4.2) returns the wrong derivatives. I've stripped away everything to a MWE:

import ForwardDiff
import ReverseDiff

struct Test{TF}
    a::TF
end

function func(x)
    tvec = Test.(x[1:2])
    a = [t.a for t in tvec]
    return sum([a; x[3]])
end

x = [1.0, 2.0, 3.0]
J1 = ForwardDiff.gradient(func, x)
J2 = ReverseDiff.gradient(func, x)

The answer should be [1, 1, 1], but ReverseDiff gives [0, 0, 1]. If I revert back to 1.2 then it works as expected. Trying different versions it seems that from 1.3 onward this no longer works.

[mohamed82008]

I think I know what's happening here. I will work on a fix soon.