Not all type parameters are substituted when checking function type parameter bound compatibility

[stereotype441]

The analyzer function FunctionTypeImpl.relateTypeFormals substitutes type bound inside the same loop that generates the substitution. This means that if one type variable refers to a later type variable, the later type variable won't get substituted when checking the bound for the former. E.g.:

relateTypeFormals(<T extends U, U>() => dynamic, <V extends W, W>() => dynamic)

would return false, since during the first loop iteration, it would only substitute fresh type variables for T and V, causing it to call relation(W, U).

What we should do instead is build up a full substitution of fresh type variables for all of the type variables in f1 and f2 first, and then check all the bounds.

This fix needs to be made in both analyzer and front_end (see TypeConstraintGatherer._matchTypeFormals).

[jmesserly]

good catch! yeah that sounds right to me. I hit that in DDC's runtime impl of bounds checking, the bounds must always be full substituted before they're checked (https://github.com/dart-lang/sdk/blob/master/pkg/dev_compiler/tool/input_sdk/private/ddc_runtime/types.dart#L484) ... I forgot to follow up on Analyzer side.

edit: in DDC, type param bounds are always a function, so in your example they'd be
(T, U) => [U, dynamic] and (V, W) => [W, dynamic] ... that forces us always instantiate the bounds before accessing them. In the subtype relation, this is done by instantiating both sets of bounds with the same fresh type variables, so we'd instantiate with fresh F0, F1 get [F1, dynamic] and [F1, dynamic] as the bounds, correctly determining that T and V have the same bound.