JeffBezanson
commented
5 years ago I believe this no longer causes infinite recursion in the compiler, but rather just during execution. If promote_rule(A,B) gives a different answer than promote_rule(B,A), we recursively call promote_type on those to "re-try". In the example here the process never converges. It might be better to give an "ambiguous promotion" error instead.
Here's an example of a case in Base that takes advantage of the current behavior:
julia> Base.promote_rule(Bool, typeof(pi))
Irrational{:π}
julia> Base.promote_rule(typeof(pi), Bool)
Float64
julia> promote_rule(typeof(pi), Float64)
Float64However, that could be resolved by adding a disambiguating definition. After all, the first result doesn't really make sense --- a Bool cannot be converted to an Irrational{:π}.
vtjnash
This is a more complicated case than #13183, and its root cause seems to be different since it persists after 9f72db1a1a94cbb8e365e00dd92b141e956c1b74. But if you're interested in squashing these, here's an interesting case: given two
Numberwrappers that are themselves numbers and promote numbers to their type, we get an infinite recursion in the compiler/inference if they ever come into contact with each other:(via https://github.com/mbauman/AxisArrays.jl/issues/20#issuecomment-140451393)