Subtle case of type instability in convert

[lbenet]

A subtle case of type instability:

julia> R = Interval{BigFloat}
Interval{BigFloat}

julia> convert(R, 1.1)
[1.09999, 1.10001]₂₅₆

julia> typeof(ans)  # this is ok
Interval{BigFloat}

julia> convert(R, Inf)
[1.79769e+308, ∞]

julia> typeof(ans)  # this is wrong!
Interval{Float64}

[dpsanders]

Oops...

cc @yashrajgupta

[lbenet]

The following gives the correct type but the result is quite different than the one above

julia> R(Inf)
[∞, ∞]₂₅₆

julia> typeof(ans)
Interval{BigFloat}

[dpsanders]

That is a bug that I think is reported in an issue: [Inf, Inf] is not a valid interval.

[dpsanders]

I guess that's because we removed error checking from the constructor... Maybe that was a bad idea and we should put it back in? We need to benchmark it though.

[lbenet]

I agree, I think we should also have a set of benchmarks to note if such changes help or not, exploiting BenchmarkTools.jl

[vtjnash]

FWIW, this is why the tests don't pass on v1.2+

[dpsanders]

Fixed by #315

[dpsanders]

Thanks for the report @lbenet and for pointing out the failing tests @vtjnash !

[lbenet]

Thanks for fixing it!