including LoopVectorization causes enzyme to segfault or mutate inputs its not supposed to

[ArbitRandomUser]

importing LoopVectorization or importing any package that imports it will cause the following code to segfault or it mutates res2 in this example (it shouldn't)

using LoopVectorization
using Enzyme
mutable struct SS
    dim::Int
    repart::UnitRange
    impart::UnitRange
end

SS(n) = SS(n,1:n,n+1:2n)


function loss(u,res)
    return sum(u[1:10].^2 .+ u[11:20].^2 .- res)
end

function loss2(sys,u,res)
    return sum( u[sys.repart].^2 .+ u[sys.impart].^2 .- res )
end

sys2 = SS(10)
res = rand(sys2.dim)
res2 = deepcopy(res)

u = rand(2*sys2.dim)
out = zero(u)

display(res)
out.=0
autodiff(Reverse,loss,Active,Duplicated(u,out),Const(res))
display("---")
display(res)

out.=0
autodiff(Reverse,loss2,Active,Const(sys2),Duplicated(u,out),Const(res2))
display("---")
display(res2)

[vchuravy]

Does LV change anything about sum?

[wsmoses]

I think broadcasting is changed, slightly smaller MWE:

using LoopVectorization
using Enzyme

mutable struct SS
    repart::UnitRange
end

function loss2(sys,u,res)
  m5 = Base.broadcasted(+, Base.getindex(u, sys.repart), res)
  # m5 = Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{1}, Nothing, typeof(+), Tuple{Vector{Float64}, Vector{Float64}}}(+, (Base.getindex(u, sys.repart), res), nothing)
  return @inbounds iterate( m5 )[1]::Float64
end

sys = SS(1:4)
res = [1., 3., 5., 7.]

u = [2., 4., 6., 8.]

out=zero(res)

autodiff(Reverse,loss2,Active,Const(sys),Duplicated(u,out),Const(res))
@show res
@show u
@show out

[wsmoses]

Problem is not throwing a mismatched activity error on an insert value for a constant inserted pointer.

A simpler, forward mode, for example:

julia> Enzyme.autodiff(Forward, (x,y)->(x,y), Duplicated(u, out), Const(res))

[wsmoses]

Fixed on main by jll bump