data corruption in struct

[tpapp]

MWE:

using Enzyme

# make a test problem
struct Foo
    D::Vector{Float64}
end
const N = 5                      # dimension for MWE
(f::Foo)(x) = mapreduce((D, x) -> log(D) + exp(x), +, f.D, x)

ℓ = Foo(fill(2.0, N)) # note the 2.0
ℓ.D                   # before AD: filled with 2.0s
x = zeros(N)
∂ℓ_∂x = zero(x)
Enzyme.autodiff(Reverse, ℓ, Enzyme.Active, Enzyme.Duplicated(x, ∂ℓ_∂x))
ℓ.D                   # after AD: filled with 2.5s

(EnzymeMWE) pkg> st
Status `~/research/spouses-time-allocation/tmp/EnzymeMWE/Project.toml`
  [7da242da] Enzyme v0.10.17

julia> versioninfo()
Julia Version 1.8.5
Commit 17cfb8e65e* (2023-01-08 06:45 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 8 × 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-13.0.1 (ORCJIT, tigerlake)
  Threads: 8 on 8 virtual cores

(Issue I mentioned in https://github.com/EnzymeAD/Enzyme.jl/issues/584#issuecomment-1416153125)

[wsmoses]

Looks like activity analysis cannot get the right answer statically. If you use Enzyme.API.runtimeActivity!(true) before running it fixes it, but let's leave this open so I can try to update activity analysis.

[tpapp]

Is there a way I could rewrite the above to avoid the problem? I am not even sure which part is causing it.

An explicit loop works, eg

function (f::Foo)(x)
    T = typeof(log(one(eltype(f.D))) + exp(zero(eltype(x))))
    ℓ::T = zero(T)
    for (D, x) in zip(f.D, x)
        ℓ += log(D) + exp(x)
    end
    ℓ
end

but calculating types manually is a bit more tedious.

[wsmoses]

As this now successfully throws a runtime activity error, am going to close (and it succeeds with runtime activity)