The default mapreduce_impl doesn't work for non-Bool ==

[ararslan]

See https://github.com/JuliaLang/julia/blob/v1.1.0/base/reduce.jl#L462. If v == v produces something other than Bool, this fails, producing errors such as that seen in jump-dev/Convex.jl#275. In the case of Convex, == produces an EqConstraint.

Is the right solution here for packages to specialize mapreduce_impl on arrays with a particular eltype so that this logic doesn't get hit, or is there more generic solution that can replace the v == v check?

[mbauman]

We should probably find a better solution — the same thing happens for missing.

julia> maximum([missing;1:15])
ERROR: TypeError: non-boolean (Missing) used in boolean context

Maybe just (v == v) === true?

[ararslan]

Actually, master seems to handle this just fine:

julia> maximum([missing; 1:15])
missing

The v == v check has actually been removed as of #30320, which inadvertently fixed this issue, both for missing and for Convex.

[ararslan]

I guess I'll close this, as there's nothing to be done for master. Whether this is fixed on 1.1 will be a matter of whether the change gets backported to 1.1.1.