which rounding mode should `convert` use?

[nsajko]

julia> setprecision(BigFloat, 4) do
           (
               setrounding(BigFloat, RoundDown) do
                   convert(BigFloat, 1/3)
               end,
               setrounding(BigFloat, RoundUp) do
                   convert(BigFloat, 1/3)
               end,
           )
       end
(0.312, 0.344)

julia> ==(ans...)
false

I think it'd make more sense for convert(::Type{BigFloat}, ::Any) not to depend on the current rounding mode. It could always use RoundNearest, no? Would that be a breaking change?

[nsajko]

I think it was wrong to ask the question specifically in terms of BigFloat. Broadening this to AbstractFloat, the question is: Which of the following should convert(T, x) behave as:

• T(x)
• T(x, RoundNearest)
• T(x, rounding(T))

The goal is to better specify the behavior of convert, which seems like it's currently underspecified.