Wrong gradient returned in case of getindex/ifelse

[rcalxrc08]

This is a weird issue and I have no idea about the cause.
MWE:

using Zygote, Test

function generate_i(i, a, d)
    ifelse(i == 1, a, d)
end

function f_std(a, d)
    N = 3
    x = ones(N)
    v = @. generate_i(1:N, a, d)
    av = v .* x
    return sum(av)
end
a = 0.2
d = 0.5
#Following should be (1.0, 2.0), while it is (0.0, 2.0)
@show res_std = Zygote.gradient(f_std, a, d)
dx = 1e-4
df_da = (f_std(a + dx, d) - f_std(a - dx, d)) / (2 * dx)
df_dd = (f_std(a, d + dx) - f_std(a, d - dx)) / (2 * dx)
@show res_fd = (df_da, df_dd)
@show analytical_res = (1.0, 2.0)
@test sum(abs.(res_std .- res_fd)) < 1e-2
@test all(res_std .≈ analytical_res)

Zygote version: v0.7.10

[rcalxrc08]

The issue is (again) in broadcasting, the following works because of explicit iteration:

using Zygote, Test

function generate_i(i, a, d)
    ifelse(i == 1, a, d)
end

function f_std(a, d)
    N = 3
    x = ones(N)
    v = [generate_i(i, a, d) for i in 1:N] # The only difference
    av = v .* x
    return sum(av)
end
a = 0.2
d = 0.5
@show res_std = Zygote.gradient(f_std, a, d)
dx = 1e-4
df_da = (f_std(a + dx, d) - f_std(a - dx, d)) / (2 * dx)
df_dd = (f_std(a, d + dx) - f_std(a, d - dx)) / (2 * dx)
@show res_fd = (df_da, df_dd)
@show analytical_res = (1.0, 2.0)
@test sum(abs.(res_std .- res_fd)) < 1e-2
@test all(res_std .≈ analytical_res)