Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Support Type Hierarchy #1231

Closed
wants to merge 8 commits into from
Closed
Changes from 2 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
220 changes: 220 additions & 0 deletions _specifications/specification-3-17.md
Original file line number Diff line number Diff line change
Expand Up @@ -1927,6 +1927,13 @@ export interface TextDocumentClientCapabilities {
* @since 3.16.0
*/
moniker?: MonikerClientCapabilities;

/**
* Capabilities specific to the various type hierarchy requests.
*
* @since 3.17.0
*/
typeHierarchy?: TypeHierarchyClientCapabilities;
}
```

Expand Down Expand Up @@ -2419,6 +2426,14 @@ interface ServerCapabilities {
}
}

/**
* The server provides type hierarchy support.
*
* @since 3.17.0
*/
typeHierarchyProvider?: boolean | TypeHierarchyOptions
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is a simple boolean necessary/sufficient here since each server capability can be defined under TypeHierarchyOptions?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, we don't need boolean here in this version. We are considering whether to remove the capability inheritanceTreeSuppport to keep the type hierarchy general. I'll update this PR then.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I find it was referring call hierarchy.

callHierarchyProvider?: boolean | CallHierarchyOptions | CallHierarchyRegistrationOptions;

which is equivalent to:

boolean // simplest 
OR
{
  workDoneProgress?: boolean;  // from WorkDoneProgressOptions
} 
OR
{
  workDoneProgress?: boolean;  // from WorkDoneProgressOptions
  id?: string;  // from StaticRegistrationOptions 
  documentSelector: DocumentSelector | null;  // from TextDocumentRegistrationOptions 
}

specifying whether server has capabitility to support work done progress / unregistration / customized scope of the feature.

Is a simple boolean necessary/sufficient here since each server capability can be defined under TypeHierarchyOptions?

@KamasamaK I think boolean is insufficient for advanced features mentioned above. For me, it's somehow necessary . Because on one hand, when I implement it in server side I might not care about any advanced feature, then I can simply return a true; on the other hand it's consistent with other capabilities like call hierarchy.

We are considering whether to remove the capability inheritanceTreeSuppport to keep the type hierarchy general.

@CsCherrYY I'm voting +1 for removing inheritanceTreeSuppport, otherwise for languages supporting multiple inheritance (e.g. C++), you may also need something like inheritanceGraphSuppport?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If you don't care about implementing or can't implement advanced features, you can return false for their capabilities. For example, codeLensProvider only has a type of CodeLensOptions with one optional native capability, so that's already established for LSP. That is the reason I say it is probably unnecessary. But if the extra capability is going to be removed, that might change things.

Perhaps @dbaeumer can weigh in on the necessity of a simple boolean for new operations when *Options exists, and whether them having no or only optional native properties makes a difference in that determination.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you for pointing out CodeLensOptions. It looks:
{ resolveProvider: false } means server doesn't have a resolve provider.
{ resolveProvider: true } means server also support to resolve code lens.
So now I have a question, if server doesn't support code lens at all, what should it return? (if boolean is not allowed)

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Excluding codeLensProvider would indicate that the server does not support it.

A missing property should be interpreted as an absence of the capability. If a missing property normally defines sub properties, all missing sub properties should be interpreted as an absence of the corresponding capability.

| TypeHierarchyRegistrationOptions;

/**
* Experimental server capabilities.
*/
Expand Down Expand Up @@ -8173,6 +8188,211 @@ export interface Moniker {
}
```

#### <a href="#textDocument_prepareTypeHierarchy" name="textDocument_prepareTypeHierarchy" class="anchor">Prepare Type Hierarchy Request (:leftwards_arrow_with_hook:)</a>

> *Since version 3.17.0*

The type hierarchy request is sent from the client to the server to return a type hierarchy for the language element of given text document positions. The type hierarchy requests are executed in two steps:

1. first a type hierarchy item is prepared for the given text document position
1. the client request the server to resolve the given item with its supertypes or/and subtypes

_Client Capability_:

* property name (optional): `textDocument.typeHierarchy`
* property type: `TypeHierarchyClientCapabilities` defined as follows:

```typescript
interface TypeHierarchyClientCapabilities {
/**
* Whether implementation supports dynamic registration. If this is set to
* `true` the client supports the new `(TextDocumentRegistrationOptions &
* StaticRegistrationOptions)` return value for the corresponding server
* capability as well.
*/
dynamicRegistration?: boolean;
}
```

_Server Capability_:

* property name (optional): `typeHierarchyProvider`
* property type: `boolean | TypeHierarchyOptions | TypeHierarchyRegistrationOptions` where `TypeHierarchyOptions` is defined as follows:

```typescript
export interface TypeHierarchyOptions extends WorkDoneProgressOptions {
/**
* The server has support for supporting inheritance tree.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

"support for supporting" -> "support for providing an"

*/
inheritanceTreeSuppport?: boolean;
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

typo in support

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am not convinced that we should add inheritanceTreeSuppport in the first version since I don't see clients to support this rendering. IMO it makes things simply harder to understand. Or do I miss something?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

My apologize for not mentioning the client behaviors about inheritanceTreeSupport in the protocol. I'd like to describe it in detail here:

When using inheritanceTreeSupport, the client has the same mechanism of handling response data as the other two requests. The only difference is how to set classOnly in TypeHierarchySupertypesParams.

  • If the server doesn't set inheritanceTreeSupport or it's false, the classOnly will always be undefined since the server doesn't support inheritance tree. Besides, Inheritance Tree button will hide in the client.
  • When the server set inheritanceTreeSupport to true, the client could show Inheritance Tree button.
    • If the active view is inheritance tree view, after prepare request, to get the most ancestor class of the given type, the client will send several typeHierarchy/supertypes requests with classOnly in their params set to true. The responses include all classes in the inheritance tree of the given item, so that the client can render this tree reversally. The remaining requests work normally.
    • If the active view is not inheritance tree view, classOnly will always be false and everything works normally.
  • In the server, if there is a typeHierarchy/supertypes request with its classOnly is true, the response should contain class only, otherwise it contains all types.

If there is no inheritanceTreeSupport in the protocol, since we'd like to support inheritance tree view for it's the most important type hierarchy view for Java users (might also for single inheritance language users), we have to independently implement the view in java extension instead. We might make it as a command and use workspace.executeCommand to trigger this view. In server, we'll have the similar logics to handle the supertype requests which return only class. Another concern is that if we can put the Inheritance Tree button properly. See the following screenshot:
type2
Currently we could put them together since we contribute ms-vscode.references-view, if the right two buttons are supported in LSP and implemented in client already, can we still find a way to put them together?

What do you think of this?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I was incorrect about what I said. I see the rendering but I am still not convinced for the modes in the API. Explaining this is simply too complicated. Why can a client not render that tree using the existing API by filtering interfaces. And what do we expect that happens if a client calls the API with classOnly for a item that represents a interface.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why can a client not render that tree using the existing API by filtering interfaces.

A client can render that tree by filtering interfaces indeed. It's a good tradeoff to remove classOnly and put all the logics in client implementation to keep the protocol clear. If we remove classOnly and keep inheritanceTreeSupport in the first version, should we define the behaviors of client? If we remove inheritanceTreeSupport from the first version, my concern is whether we can still keep the inheritance tree view (from java extension) and the other two views button (from VSCode) together.

And what do we expect that happens if a client calls the API with classOnly for a item that represents a interface.

Currently we expect calling the API with classOnly for an interface will return an empty array since never a class can be the supertype of an interface.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am not understanding why inheritanceTreeSupport can influence how a client can renderer a class hierarchy. Isn't this something that is under the sole control of the client and a server should never influence this?

Another problem I see with classOnly is that it is specific for a certain use case. It might cause problems with other languages. I could for example imagine a special view for language that have multiple class inheritance. Another view could be special for interface to see the implemented protocols.

If we think that that performance and payload size is a problem then we should add a filter property to the params which would allows to filter hierarchy items on the server side. IMO classOnly is simply to specific for a specific use case.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

About inheritanceTreeSupport

After reading through above comments, I think current naming is causing confusion. Let me help to clarify according to my understanding:

ServerCapabilities.inheritanceTreeSupport proposed by @CsCherrYY indicates whether "the language only supports single inheritance through class". E.g. it's true for Java because multiple inheritance is supported through interfaces. And it's false for C++. (@CsCherrYY correct me if I'm wrong)

"how a client can renderer a class hierarchy" mentioned by @dbaeumer should be related to a Client capability if needed. For example inheritanceTreeSupport, inheritanceGraphSupport (for multiple-inheritance languages?) etc. (@dbaeumer correct me if I'm wrong)

About classOnly

If we think that that performance and payload size is a problem then we should add a filter property to the params

Agree that classOnly is merely a specific filter property, is there any other filtering requirement (e.g. interfaceOnly), how would a generic filter property look like?
I'm testing this proposal on Java language server, so far I don't see any performance/payload size problems as we expand the hierarchy level by level, and I'm fine to filter items on either server/client side.

Here I only have a stupid question as I know little about client implementation details: No matter what the filtering property eventually is, if I want to support showing class view (e.g. using tree items in vscode's reference view) for Java, how does client know when to filter the items or not? Does it provide multiple entries (e.g."show supertypes", "show subtypes", "show class view", "show interface only") and send requests with different params corresponding to the entries?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Filter properties would be properties that are available on a TypeHierarchyItem. I would say typically the kind property.

Regarding the UI: it could either be a different command or some sort of filter buttons in the UI.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If it's possible to keep the Class Hierarchy view available for the Java users, I'm very glad to remove inheritanceTreeSupport in the first version since it's indeed a little complicated. Here is another approach to support Class Hierarchy: The Java extension can contribute another button in reference-view and use a separated command to show class hierarchy. But there is still a concern about the UI implementation.

We want to make the separated class hierarchy button align with the two existing type hierarchy buttons (supertype hierarchy and subtype hierarchy). In the current reference-view implementation, the call hierarchy uses context value to control whether to show the button. Since the call hierarchy has only two views, it works well. But for type hierarchy, can we make the context value extendable to support possible other views? Then we can use other conditions like enablement to show if this button is disabled, like the current Java implementation.

@dbaeumer Could you give any suggestions?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Update: Another concern is about the default view. Currently the default type hierarchy view for java users is Class Hierarchy view, which is most frequently used by java developers. However, the implementation of type hierarchy in vscode-references-view would be based on the protocol, supporting subtype hierarchy view and supertype hierarchy view only. Could it be possible to change the default view of type hierarchy in another extension? e.g. Java extension.

}
```

_Registration Options_: `TypeHierarchyRegistrationOptions` defined as follows:

```typescript
export interface TypeHierarchyRegistrationOptions extends
TextDocumentRegistrationOptions, TypeHierarchyOptions,
StaticRegistrationOptions {
}
```

_Request_:

* method: 'textDocument/prepareTypeHierarchy'
* params: `TypeHierarchyPrepareParams` defined as follows:

```typescript
export interface TypeHierarchyPrepareParams extends TextDocumentPositionParams,
WorkDoneProgressParams {
}
```

_Response_:

* result: `TypeHierarchyItem[] | null` defined as follows:

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why should we allow null here ? What's the client expected to do with a null? It's better to define this to either return an empty array, or an error. Alternatively explicitly define what the meaning of null is (as opposed to an empty array or an error return).

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In the textDocument/prepareTypeHierarchy request, if the element in the location of the params is not a valid type (that depends on the server implementation, whether to infer a valid type), it will return null to indicate that there is no result. For those types have no subtypes in typeHierarchy/subtypes, the server will return an empty array.

Thanks for the comment, I'll add a description in the next commit.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

null is for consistency since lots of LSP requests that return and array can return null to indicate no elements


```typescript
export interface TypeHierarchyItem {
/**
* The name of this item.
*/
name: string;

/**
* The kind of this item.
*/
kind: SymbolKind;

/**
* Tags for this item.
*/
tags?: SymbolTag[];

/**
* More detail for this item, e.g. the signature of a function.
*/
detail?: string;

/**
* The resource identifier of this item.
*/
uri: DocumentUri;

/**
* The range enclosing this symbol not including leading/trailing whitespace
* but everything else, e.g. comments and code.
*/
range?: Range;

/**
* The range that should be selected and revealed when this symbol is being
* picked, e.g. the name of a function. Must be contained by the
* [`range`](#TypeHierarchyItem.range).
*/
selectionRange: Range;

/**
* Indicates if this item has supertypes. When `undefined` the supertypes
* of this item have not been resolved yet.
*/
hasSupertypes?: boolean;

/**
* Indicates if this item has subtypes. When `undefined` the subtypes
* of this item have not been resolved yet.
*/
hasSubtypes?: boolean;

/**
* A data entry field that is preserved between a call hierarchy prepare and
* incoming calls or outgoing calls requests.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

"and incoming calls or outgoing calls requests" - I'm not sure I understand what that's saying.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's copy/pasted from CallHierarchyItem.data and mistakenly unchanged.

*/
data?: unknown;
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Would it make more sense for the prepare response to be an object (versus an array), and have a single data field? I guess I'm not sure to what extent the full call needs data versus each individual item (where if you really just wanted the same data, you'd have to dupe it on each of them).

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

FWIW, in clangd, what we put in the data field is a hash of the internal universal identifier of the symbol represented by the TypeHierarchyItem, so it being per-item makes sense.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think if we need a data field in each individual item would depend on server. If server could hold or cache the whole type hierarchy and use it to find a type in coming requests, one data field is OK, but if server wants to use data field to indicate the specific type, we'd better to keep a data field in each item. In protocol, I suggest keeping this field in each item to support these two usages both.

}
```

* error: code and message set in case an exception happens during the 'textDocument/prepareTypeHierarchy' request

#### <a href="#typeHierarchy_supertypes" name="typeHierarchy_supertypes" class="anchor">Type Hierarchy Supertypes(:leftwards_arrow_with_hook:)</a>

> *Since version 3.17.0*

The request is sent from the client to the server to resolve the supertypes for a given type hierarchy item. The request doesn't define its own client and server capabilities. It is only issued if a server registers for the [`textDocument/prepareTypeHierarchy` request](#textDocument_prepareTypeHierarchy).

_Request_:

* method: 'typeHierarchy/supertypes'
* params: `TypeHierarchySupertypesParams` defined as follows:

```typescript
export interface TypeHierarchySupertypesParams extends
WorkDoneProgressParams, PartialResultParams {
item: TypeHierarchyItem;
}
```
_Response_:

* result: `TypeHierarchyItem[] | null`

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

what does null return mean?

* partial result: `TypeHierarchyItem[]`
* error: code and message set in case an exception happens during the 'typeHierarchy/supertypes' request

#### <a href="#typeHierarchy_subtypes" name="typeHierarchy_subtypes" class="anchor">Type Hierarchy Subtypes(:leftwards_arrow_with_hook:)</a>

> *Since version 3.17.0*

The request is sent from the client to the server to resolve the subtypes for a given type hierarchy item. The request doesn't define its own client and server capabilities. It is only issued if a server registers for the [`textDocument/prepareTypeHierarchy` request](#textDocument_prepareTypeHierarchy).

_Request_:

* method: 'typeHierarchy/subtypes'

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's not obvious why there are so many messages here, nor why the prepare and resolve steps are required.

For clients (and presumably many servers), it would be simpler to return the whole tree in a single message. With this current API proposal there will be a lot of chatter for an ostensibly simple dataset. Could we perhaps incorporate the ability for the server to return the full type hierarchy in response to just a simple /typeHierarchy message as the first instance? Do we know that this is extremely expensive to calculate in many servers (enough to justify the complexity of this back-and-forth prepare/request/request/request API?

It's also unclear to me when a client should use subtypes and super types requests vs the inheritanceTree request - unless I'm missing something, they are implementing the same thing in a different way? What's the rationale for this?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just as you mentioned, calculating the full type hierarchy is extremely expensive for a given type with many subtypes. We currently have an implementation in Java language server, it may cost about 25 seconds to calculate the direct subtypes of java.io.Serializable (1877 subtypes in the example project in total). We want the user to know what server is doing - after prepare request, the client can get the base type itself so that it can be shown. So the user can see the based type with some progress icons at least, with partial result mechanism (if implemented), the client can show the results gradually.

For the second question, we have two common views for all languages, they are subtypes and supertypes. For single inheritance languages, since a given class could only have one superclass, usually they have another view to show all the classes in the inheritance tree. That can be performed in many supertypes requests and a subtypes request as well, so we're still in discussing whether to isolate this request.

* params: `TypeHierarchySubtypesParams` defined as follows:

```typescript
export interface TypeHierarchySubtypesParams extends
WorkDoneProgressParams, PartialResultParams {
item: TypeHierarchyItem;
}
```
_Response_:

* result: `TypeHierarchyItem[] | null`

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

again, what's null for? we need to be explicit about what the valid responses are and what they mean, so that client and servers work together in harmony.

* partial result: `TypeHierarchyItem[]`
* error: code and message set in case an exception happens during the 'typeHierarchy/subtypes' request

#### <a href="#typeHierarchy_inheritanceTree" name="typeHierarchy_inheritanceTree" class="anchor">Type Hierarchy Inheritance Tree(:leftwards_arrow_with_hook:)</a>

> *Since version 3.17.0*

The request is sent from the client to the server to resolve the inheritance tree for a given type hierarchy item. The response result of this request is `TreeItem<T>`.
* `TreeItem<T>` defined as follows:
```typescript
export interface TreeItem<T> {
data: T;
/**
* The children of this TreeItem. When `undefined` the children have not been resolved yet.
*/
children?: TreeItem<T>[];
}
```
The result represents the inheritance tree related to the specified type hierarchy item. Only single inheritance can be represented this way. The root type is expected as the return value. The request doesn't define its own client and server capabilities. It is only issued if a server has the capability for `TypeHierarchyOptions/inheritanceTreeSuppport`.

_Request_:

* method: 'typeHierarchy/inheritanceTree'
* params: `TypeHierarchyInheritanceTreeParams` defined as follows:

```typescript
export interface TypeHierarchyInheritanceTreeParams extends
WorkDoneProgressParams {
item: TypeHierarchyItem;
}
```
_Response_:

* result: `TreeItem<TypeHierarchyItem> | null`

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What does null mean?

* error: code and message set in case an exception happens during the 'typeHierarchy/resolve' request
Copy link
Contributor

@KamasamaK KamasamaK Mar 30, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this supposed to be 'typeHierarchy/inheritanceTree'? There is no 'typeHierarchy/resolve' request defined.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes. I'll fix it in the next commit.


##### Notes

Server implementations of this method should ensure that the moniker calculation matches to those used in the corresponding LSIF implementation to ensure symbols can be associated correctly across IDE sessions and LSIF indexes.
Expand Down