Method dispatch not functioning as expected

[shashi]

In this particular context (I am not entirely sure what is causing this), although @which gives me the correct method, but this method is not what is actually called, instead it calls the more general metadata(::Any) method.

https://github.com/shashi/IJulia.jl/blob/e638f0013ce1107bc2f42a172309403e1123ad39/src/execute_request.jl#L178

In a previous iteration of the code, if I were to print something (e.g. a blank) inside the metadata(::Signal) method , it would work. This no more works though. So this might have something to do with stdIO being captured by IJulia?

cc @stevengj, @dcjones, @JeffBezanson

[shashi]

The other definition of metadata(::Signal) is here: https://github.com/shashi/IJuliaWidgets.jl/blob/master/src/IJuliaWidgets.jl#L46

[simonster]

If execute_request_0x535c5df2 is called (and thus compiled) before IJuliaWidgets is loaded, this may be a variant of #265. I'd guess that metadata(::Any) is getting inlined into the function at compile time, and printing something used to make the function uninlineable but no longer does because of tweaks to the inlining heuristics. You can check this with code_typed, or alternatively, if I'm right, you may be able to prevent inlining by adding another metadata method in that file (e.g. metadata(x::Int) = Dict()) or by putting a dead loop in the body of metadata(x).

[shashi]

I added a dead loop while true break end to metadata(::Any) and it started working. This does seem to be a variant of #265. Closing this.

[shashi]

What would be a better workaround? Picking out the right method with @which or leaving the dead loop in metadata?

[JeffBezanson]

Using invoke might work, and that should be nicer than the other approaches.