User-code BoundsError inside a Base.getindex method silently displays as #undef

[yurivish]

Example:

struct Foo2{T, N} <: AbstractArray{T, N}
end
Base.getindex(x::Foo2, I...) = throw(BoundsError())
Base.size(x::Foo2) = (1,)

x = Foo2{Int, 1}()
println(x) # => [#undef]

This is an MWE that originally arose in a more complicated context, where I had an indexing bug that was silently turning the array into one that displayed as [#undef].

[JeffBezanson]

I think this behavior is somewhat reasonable --- with the proliferation of @inbounds annotations, you're lucky it didn't crash. There's only so much we can do to support array types that misbehave, e.g. reporting their size incorrectly.

I believe this is due to the fallback for isassigned that turns bounds errors and undef ref errors into false answers. The printing code then assumes that if isassigned returns false the element must be undefined, since it knows it only accesses valid indices. One possible (breaking) change here is for isassigned to only check for undefined elements, and throw bounds errors for out-of-bounds queries. The current behavior could be moved to haskey.