Differentiating through `Threads.@threads` with boxed variables leads to race conditions

[penelopeysm]

The following bug requires a Julia session launched with multiple threads. The primal is deterministic, but the gradient with reverse-mode Enzyme does not always give the correct result:

using Enzyme

struct Container
    vals::Vector{Float64}
end

function f(x)
    a = x[1]
    c = Container(zeros(Threads.maxthreadid()))
    Threads.@threads for i in 1:100
        tid = Threads.threadid()
        c.vals[tid] += a * i
        c = c # forces `c` to be boxed
    end
    return sum(c.vals)
end

julia> unique([f([2.0]) for _ in 1:20])
1-element Vector{Float64}:
 10100.0

julia> unique([Enzyme.gradient(Enzyme.Reverse, f, [2.0]) for _ in 1:20])
4-element Vector{Tuple{Vector{Float64}}}:
 ([5050.0],)
 ([4950.0],)
 ([4995.0],)
 ([4658.0],)

julia> versioninfo()
Julia Version 1.11.9
Commit 53a02c0720c (2026-02-06 00:27 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: macOS (arm64-apple-darwin24.0.0)
  CPU: 10 × Apple M1 Pro
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, apple-m1)
Threads: 8 default, 0 interactive, 4 GC (on 8 virtual cores)

(ppl) pkg> st
Status `~/ppl/Project.toml`
  [7da242da] Enzyme v0.13.132