Circumvent type-instability of deepcopy

[timholy]

Due to the use of the ObjectIdDict, deepcopy is type-unstable. Obviously this dict is important for avoiding cycles, so it seems the best solution is to prevent the type-instability from propagating.

Ref JuliaImages/Images.jl#356

[timholy]

Hmmm, more complicated than I thought. I guess the fix for this depends on what deepcopy really means. Ref #13124.

[Keno]

I don't really understand why that issue is related?

[timholy]

The part about equivalence to "serialize" and "deserialize"---that's one way of stating a principle-of-operation for deepcopy, but there might be others. For example, the fact that SharedArrays don't serialize their data would mean you're not actually copying the data.

[vtjnash]

i think there's a related issue (noted in the serialization code), which is that if the exact type isn't returned, then it might get assigned to a field of a different type causing issues.

i think we should either merge this, or perhaps insert the following slow path to deserialize and deepcopy:

if typeof(elt) != fieldtype(T, i)
  elt = convert(fieldtype(T, i), elt)
end

[JeffBezanson]

I also think we should merge this. Unless I'm missing something, I don't think there is a need for deepcopy to literally give the same result as deserialize. That's just one way to understand the sort of thing it does.
Ideally, serialize of a subarray would still store a subarray, but with the backing array trimmed down. It's pretty bad not to preserve type. Of course now that serialize preserves reference topology even this behavior is questionable, unfortunately.

[timholy]

serialize(::SubArray) does return a SubArray, now, indeed with a trimmed backing array. See #14625 and #14708.

Fix for SharedArrays (ref #6362) coming once local tests finish.

[timholy]

Hmm, the AppVeyor failure seems odd because one can see the final "SUCCESS", but then it's judged as failing. Anything to worry about here?

[vtjnash]

try restarting? it looks all of the tests finished, but one of the subprocesses failed to shutdown

[JeffBezanson]

👍