nullbroadcasted example

[haakon-e]

In #4095, we observed that the addition of a new term vtt_bc in advection.jl, as follows:

if !(p.atmos.moisture_model isa DryModel) && energy_q_tot_upwinding != Val(:none)
    ᶜq_tot = @. lazy(specific(Y.c.ρq_tot, Y.c.ρ))
    vtt = vertical_transport(ᶜρ, ᶠu³, ᶜq_tot, FT(dt), energy_q_tot_upwinding)
    vtt_central = vertical_transport(ᶜρ, ᶠu³, ᶜq_tot, FT(dt), Val(:none))
    vtt_bc = ᶜρq_tot_vertical_transport_bc(prescribed_flow, thermo_params, t, ᶠu³)
    @. Yₜ.c.ρq_tot += vtt - vtt_central + vtt_bc
end

where vtt_bc = ᶜρq_tot_vertical_transport_bc(prescribed_flow, thermo_params, t, ᶠu³) returns either a @. lazy field or NullBroadcasted().

In the buildkite test :fire: Flame graph: perf target (Callbacks), where vtt_bc should evaluate to NullBroadcasted(), we observed ~10% more allocations, which exceeded the flame_callbacks threshold.

We hypothesize that the expression was not correctly inferred, which led to the (potential) simulation slowdown