[CS] Allow binding closure param to pack expansion

[jameesbrown]

In resolveClosureType allow binding a closure parameter (when it's a free type variable) to a contextual type that is a pack expansion and otherwise valid.

Resolves: #78426

[xedin]

I think this should go into ConstraintSystem::resolveClosure because that's the spot where we know that there is a contextual type.

[jameesbrown]

Thanks! I actually started with ConstraintSystem::resolveClosure, but the solution didn’t quite fall into place as cleanly, which led me to try this approach. I definitely missed something though and will try that way again.

[xedin]

This approach would only work for cases were contextual type is known in advance but if closure is passed as an argument to an overloaded function there is no contextual type available upfront, it becomes available once an overload of that function is bound.

[jameesbrown]

That makes sense, thanks for pointing it out, I was having trouble thinking of how we'd not have the contextual type other than the most basic version: let _ = { x in }.

[xedin]

I think the important question here is why doesn't it get bound, is this an inference problem?

[jameesbrown]

If we start with a plain type variable that doesn't have TVO_PackExpansion, it's not allowed to be bound to a PackExpansionType (there's a comment about this in matchTypesBindTypeVar). In some places we do the:

auto *pack = PackType::get(Ctx, tyVar)
// bind this to the pack expansion
auto *expansion = PackExpansionType::get(pack, pack)

But that always fails when the pack expansion we are binding to has a pattern type that's a PackArchetypeType because we'll try to match the reduced shape of Pack{$T1} to a PackArchetypeType.

What we are expecting is to have opened the pack expansion type via

swift/lib/Sema/TypeOfReference.cpp

Line 296 in 52ebd4a

Type ConstraintSystem::openPackExpansionType(PackExpansionType *expansion,

and use that as our type variable, but at this point it seems to late. The only place I can think of is when we initially infer the closure type, but there are problems with that too. Is there a way we can say "forget about this type variable" and replace it?

[xedin]

Okay, so I think what we need to do instead of assigning is to flip the flag on parameter type variable to allow it to be bound to a pack expansion.

[xedin]

This might be okay to do by default in inferClosureType since we already set TVO_CanBindInOut there.

[jameesbrown]

Okay I like that way better too, thank you! Going to try that.

[xedin]

@jameesbrown are you still planning to work on this?