Type Analysis Error and Segfault for Data structs with Tuples and Arrays in them in Julia 1.11, but not Julia 1.10

[maximilian-gelbrecht]

This came up while working on applying Enzyme to SpeedyWeather.jl

Sorry, for the lengthy title, not sure how to better put this into words.

As a MWE, without our data structures from there:

import Pkg 
Pkg.activate("test")
using Enzyme

@kwdef struct StructA{A<:AbstractArray, B<:AbstractArray} 
    vor::NTuple{2,A} = ntuple(i -> zeros(20,20), 2)
    div::NTuple{2,A} = ntuple(i -> zeros(20,20), 2)
    pres::NTuple{2,B} = ntuple(i -> zeros(20), 2)
    other::B = zeros(20)
    other_int::Int = 0
end  

@kwdef struct StructB{A<:AbstractArray, B<:AbstractArray} 
    vor_tend::A = zeros(20,20)
    div_tend::A = zeros(20,20)
    pres_tend::B = zeros(20)
    other::B = zeros(20)
    other_int::Int = 0
end 

function lf2!(
    progn::StructA,
    tend::StructB)

    for (varname, tendname) in zip((:vor,:div,:pres), (:vor_tend,:div_tend,:pres_tend))
        var_old, var_new = getfield(progn, varname)
        var_tend = getfield(tend, tendname)
        compute_something!(var_old, var_new, var_tend)
    end

    return nothing
end

function compute_something!(A, B, C)
    A .+= C 
    B .-= C 
    return nothing 
end

A = StructA()
dA = StructA(vor=ntuple(i -> ones(20,20), 2), div=ntuple(i -> ones(20,20), 2))
B = StructB()
dB = make_zero(B)

autodiff(Reverse, lf2!, Const, Duplicated(A, dA), Duplicated(B, dB))

With Enzyme in Julia 1.10 this works, with Enzyme in Julia 1.11 this first got me some type analysis errors, and now in this exact version a seg fault.

With simpler data structures, that still contain different array types (like in StructB), but not nested within tuples as above in StructA it also works.

There are ways I could make the code of lf2! a bit more explicit and avoid doing it like above, but it does already work just fine with Enzyme in 1.10 like it is, just not with Julia 1.11.