Will Tebbutt

@oxinabox what are your thoughts on the `AbstractFFTs` idea above? edit: we should consider the solution discussed [here](https://github.com/JuliaDiff/ChainRules.jl/issues/280)

The point is not that the output is inconsistent (I agree that the output is consistent with what you would get if you provided an `Array`), it's that the optimal...

> But what seems a little tricky is that, if the returned gradient escapes from Zygote, the returning the one-number version seems like it violates the fact that Fill Can...

> Maybe the short version is that I'd like more examples of where Fill gets used. One place that I use `Fill`s extensively is in [TemporalGPs.jl](https://github.com/willtebbutt/TemporalGPs.jl) to represent time-invariant dynamics....

I think the discussion has generally moved on since I opened this, but it's worth pointing out that quite literally no subtype of `AbstractArray` will work with the current definition...

Moreover, it's really hard to fix and let AD run on however `getindex` is implemented. In particular, it's not obvious how to write an `@adjoint` which says "just run AD...

Good catch. @CarloLucibello would you mind moving this over the ChainRules?

FWIW, I have noticed substantially-increased compilation times recently (anecdotally -- I don't have a good MWE)

@pxl-th could you please also running your benchmarks on 0.6.29 and 0.6.28 to see if this is a recent regression, or a problem that has been around for a while?

@theogf I know you've resolved this on your end now, but would you be up for pushing this PR through anyway? Seems like something that we should be doing throughout...