I think this is working-as-intended -- not everything that could be inferred, is inferred.

@griesemer or @mdempsky would know better.

/cc @findleyr @ianlancetaylor

The compiler is correct according to the current type inference rules. There is no type inference from the type of an argument to the instantiation of a generic type used as an ordinary parameter.

should we change the rules? it doesn't make sense to infer it. it only can be int64

My understanding of The Plan is that we want to let the current rules soak for a while, and see how things work as they are. Once we extend an inference rule, we can't retract it, and there might be impacts on other parts of type inference, analysis time, other tools, error message quality, and perhaps readability (which is more important than write-ability). We're still dealing with some of the long-tail problems of the generics addition (build speed, IR changes, export data changes, deciding how we feel about cyclic interfaces, and "comparable").

Anyhow, I'll close with working-as-intended-for-now. We ought to revisit this in a year or so.

Update: This works now on the dev branch (eventual Go 1.21).