Base.mightalias claims to be conservative but isn't

[jakobnissen]

Base.mightalias 's docstring says (my emphasis)

Perform a conservative test to check if arrays A and B might share the same memory.

Yet, its default implementation basically always returns false , even for objects that does alias:

julia> struct Foo <: AbstractVector{Int}
           x::Vector{Int}
       end;

julia> x = [1,2];

julia> Base.mightalias(x, Foo(x))
false

julia> d = Dict();

julia> Base.mightalias(d, d.keys)
false

I think this is a problem. Presumably, mightalias is going to be used as a guard to prevent aliasing bugs. Having it claim to be conservative, but actually falsely return false for most arguments is a huge footgun.

Luckily for this issue, Base.mightalias does not appear to be documented in the official Julia docs, so it's fair game to change. I believe the best course of action is to remove the default fallback implementation(s) for this and Base.dataids.

See also: #26865

[mbauman]

Yes, the aliasing detection system is opt-in. Once you opt-in, it is a conservative check. Perhaps that one-sentence summary should be clearer, but the rest of the docstring is trying to describe this:

By default, this simply checks if either of the arrays reference the same memory regions, as identified by their Base.dataids.

I fought for a long time trying to come up with a better default fallback, but it is extraordinarily hard to do so. See #26237. Removing the default fallbacks is highly breaking, even in situations where no aliasing occurs.