lower bound interferes with typevar value inference

[oxinabox]

Consider the following:

struct XX end
struct YY end

bar(a::A, b::B, c::C) where Tuple{A,B}<:C<:Union{Tuple{XX,YY}, Tuple{YY,XX}} where A where B = C

@show bar(XX(), YY(), (XX(),YY()))

gives:

julia> @show bar(XX(), YY(), (XX(),YY()))
ERROR: UndefVarError: C not defined
Stacktrace:
 [1] bar(::XX, ::YY, ::Tuple{XX,YY}) at ./REPL[3]:1

Expected is to return Tuple{XX,YY}

Similarly:

julia> @show bar(YY(), XX(), (YY(),XX()))
ERROR: UndefVarError: C not defined
Stacktrace:
 [1] bar(::YY, ::XX, ::Tuple{YY,XX}) at ./REPL[3]:1

Expected is to return Tuple{YY,XX}

For interest:
bar(XX(), YY(), (XX(), XX()) and bar(XX(), XX(), (XX(), XX()) , and bar(XX(), XX(), (YY(),YY())) and bar(YY(), YY(), (YY(),YY())) and bar(YY(),XX(), (YY(),YY())) give method errors

[charleskawczynski]

Shouldn't the assignment be to something else (e.g., "xor") and not C

[JeffBezanson]

Yes, I think the lower bound just makes it too hard for us to derive a unique value for C. It works without the lower bound.

[oxinabox]

It looks to me, based on the MethodErrors the dispatch is actually working,
it is just losing C somewhere.

[JeffBezanson]

Yes, dispatch is now (as of 0.7) totally de-coupled from determining the values of static parameters. We know the arguments are a subtype, and so the method matches, we just don't know what value C should have. It's not losing the value; it is unable to determine a value for C in the first place.

There are various cases where we can't determine the value of a static parameter, and we decided it was better to give errors in those cases than to affect method matching --- that way (1) if you don't use the value of the static parameter there's no problem, (2) if you hit the error you can work around it with an @isdefined check, and (3) we can improve static parameter value inference without changing dispatch behavior.