Enforce return types on interface view types

[cocreature]

The design tries to keep the view types equivalent to templates, i.e., records with no type arguments. Daml Engine & the typechecker does not.

We need to change one of the two. I think changing Daml Engine makes more sense.

[S11001001]

I believe this matches the rules we want:

daml/daml-lf/interface/src/main/scala/com/digitalasset/daml/lf/iface/reader/InterfaceReader.scala

Lines 248 to 256 in 531221d

	rawViewType <- toIfaceType(name, astIf.view)
	viewType <- rawViewType match {
	case TypeCon(TypeConName(tcn), Seq()) => \/-(Some(tcn))
	case TypePrim(PrimType.Unit, _) => \/-(None)
	case _ =>
	invalidDataTypeDefinition(
	name,
	s"interface view type ${astIf.view.pretty} must be either a no-argument type reference or unit",
	)

In other words

1. a reference to a record type with no type arguments, or
2. unit

So the LF representation of DefInterface#view can be a lot smaller than Type.

[cocreature]

I'm not sure we want to allow unit. That still means we cannot just expose a Record on the ledger API.

[S11001001]

We can use absence of a Record in the InterfaceView as a representation of (). e: This would mean either it's () or include_interface_views=false. Is that a problem? e2: Or absence of the InterfaceView itself means the latter. I'm not 100% sure of the semantics here, as I've only been concerned with cases where include_interface_views=true.

One thing that was unclear from the design was whether a view would be required. It makes sense to me that it wouldn't be required, but permitting () is functionally identical.

On the ledger API, we can interpret absence of a view as just stated, though it might make more sense to just reject transaction filters that specify interfaces whose viewtype is ().

[cocreature]

Currently we make it mandatory and I’d keep this. It means interface subscriptions always work exactly like template subscriptions and there are no interfaces you cannot subscribe to or get no meaningful return value. Relaxing it later to make them optional is also backwards compatible while the inverse isn’t so I’d start with the more restricted version.

If in our tests this becomes fairly annoying, we can define a shared record type (like Archive) that we use instead of unit where we don’t care about the view.

[S11001001]

we make it mandatory and I’d keep this

@cocreature Do you mean, "it must be present, and must refer to a record type with no type arguments"?

If in our tests this becomes fairly annoying

It already would be, I'd wager. #14456 contains 79 uses of viewtype ().

[cocreature]

Do you mean, "it must be present, and must refer to a record type with no type arguments"?

yes

[dylant-da]

It already would be, I'd wager. #14456 contains 79 uses of viewtype ().

Many of them are typechecked, too, unfortunately.

[cocreature]

Many of them are typechecked, too, unfortunately.

True but

data EmptyInterfaceView = EmptyInterfaceView {}

viewtype EmptyInterfaceView

doesn’t seem that terrible.

[dylant-da]

data EmptyInterfaceView = EmptyInterfaceView {}

viewtype EmptyInterfaceView

doesn’t seem that terrible.

You're right it looks fine - started work on this in https://github.com/digital-asset/daml/tree/non-unit-viewtypes

[akrmn]

Fixed by #14802 and #14932