InvertedIndices broke on 1.11 about julia HOT 2 CLOSED

It looks like there is a fault with how InvertedIndices implements its own computation of picks, which fails its own boundscheck implementation:

julia> axes(Base.to_indices(A, (Not(z),))[1].picks)
(Base.OneTo(24),)

julia> axes(A)
(Base.OneTo(3), Base.OneTo(2), Base.OneTo(4))

The issue looks like it might have been pirating Base.to_indices before, resulting in slightly different calls to its own internal functions, causing this change to uncover a bug in the package's boundscheck implementation.

from julia.

Yeah, effectively what's happening here is that InvertedIndices needs to guess whether Base would generate linear or cartesian indices out of the non-not flavor of to_indices — and then tries to match it so it can iterate (and compare) both the "picks" and the "skips" in lock-step. It is currently trying to workaround around what was Base's inconsistency that #45869 addressed:

https://github.com/JuliaData/InvertedIndices.jl/blob/b7d7ce078888b2966f4e9acee80170bc48612fa0/src/InvertedIndices.jl#L184-L191

That said, its paired iteration strategy is just too clever for its own good and isn't achieving the performance it intended. That was initially my doing, and we should probably just go back to an eager vector.

from julia.