@non_differentiable &

[mcabbott]

Marking & non-diff fixes this:

julia> gradient(x -> eachindex(x)[1], rand(3))
ERROR: Rewrite reached intrinsic function and_int. Missing rule?
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:33
  [2] (::∂⃖{1})(::Core.IntrinsicFunction, ::Bool, ::Vararg{Bool})
    @ Diffractor ~/.julia/packages/Diffractor/HYuxt/src/stage1/generated.jl:205
  [3] &
    @ ./bool.jl:38 [inlined]
  [4] (::Diffractor.∂⃖recurse{1})(::typeof(&), ::Bool, ::Bool)

And div just noticed in passing, always returns an integer, so should not have a derivative.

[oxinabox]

don't forget to increment the version

[mcabbott]

This may have broken things. NNlib and Flux's tests now fail with similar, mysterious, errors, from within Zygote's CR code. Limiting to CR 1-1.17 to exclude this PR fixes this.

FluxML/NNlib.jl#375 (comment)

It could be other PRs but this one seems most likely to gets its fingers deep in the works -- that was the point of the above example.

I haven't yet managed to reproduce the problem locally; as tests segfault for me anyway. @darsnack saw similar failures locally.

[darsnack]

@mcabbott is this still the suspected error source or was it #562? If it is, I can reproduce the issue locally and test if reverting fixes things, then submit a PR.

[mcabbott]

I hope it was that. CR 1.18.1 is released, and it looks like NNlib CI is past the point where it used to fail: https://github.com/FluxML/NNlib.jl/runs/4805913627?check_suite_focus=true . (I couldn't get to the failure locally.)