Fix seeding with undefined elements

[devmotion]

Fixes #436. Fixes #740.

[gdalle]

I didn't know about isassigned, this seems appropriate. Could this cause a slowdown in the vast majority of cases where similar actually assigns stuff?

[devmotion]

Could this cause a slowdown in the vast majority of cases where similar actually assigns stuff?

I put it behind an isbitstype check, so it should never be hit for bit types (such as Float64, Float32, Duals of Float64, Complex of FloatXY, etc.)

[gdalle]

Does this test cover all the new branches? It seems CodeCov is sleeping at the wheel.
I think we may also want to add cases where only some of the values are uninitialized?

[devmotion]

Codecov results are available on the codecov webpage: https://app.codecov.io/github/JuliaDiff/ForwardDiff.jl/commit/6f8f6a977d4d69fda7c492f6e0a47cb8a6853045 (I assume the codecov app is not installed in the repo and hence there's no nice summary in the PR - AFAICT I don't have sufficient permissions to install it)

Indeed, you were right, the tests did not cover all new branches. I extended the tests, and locally it seems that they cover all branches as commenting out any of the branches leads to test failures.

[devmotion]

Codecov confirms that all modified lines are covered by the tests: https://app.codecov.io/github/JuliaDiff/ForwardDiff.jl/commit/ec3e483747159b46e0db5328c60c3543c13478a6/indirect-changes

[KristofferC]

:/, this is quite unfortunate, is this really the only way to fix this?

[gdalle]

I think the idea is that if x[idx] is #undef, we want duals[idx] to be the same

[devmotion]

Yeah, I wasn't 100% percent happy but it was the best I could think of and what apparently is used in base. IMO we really want to ensure here that we get the same behaviour as copyto! on the primal part, to avoid that yduals contains any surprising values, possibly from previous chunks or differentations.

[devmotion]

julia> x = BigFloat[1, 2];

julia> y = similar(x); y[2] = 1;

julia> x
2-element Vector{BigFloat}:
 1.0
 2.0

julia> y
2-element Vector{BigFloat}:
 #undef
   1.0

julia> copyto!(x, y);

julia> x
2-element Vector{BigFloat}:
 #undef
   1.0

[gdalle]

What's the word on this one? Do we want it?

[pkofod]

gentle bump from me as well

[pkofod]

Bump again

[KristofferC]

I guess let's go with _unsetindex! for now then. 👍

[KristofferC]

I opened JuliaLang/julia#59394 regarding unsetindex!.