[.]Type completion for (any P). shows it will produce an existential metatype

[AnthonyLatsis]

Test case:

// RUN: %empty-directory(%t)
// RUN: %target-swift-ide-test -batch-code-completion -source-filename %s -filecheck %raw-FileCheck -completion-output-dir %t
protocol P {}

(any P).#^COMPLETE?check=META^# 
// META: Keyword/CurrNominal:  Type[#(any P).Type#]; name=Type

This test will fail, because the current completion is Type[#any P.Type#]; name=Type. Notice how it thinks the completion will produce an existential metatype any P.Type, whereas the actual result is the singleton metatype (any P).Type.

Note
any syntax was proposed in SE-0335.

Environment

• Swift 5.9-dev (c1d5118)

[AnthonyLatsis]

@Rajveer100 This one is somewhat less involved if you’re interested.

[Rajveer100]

So do you suggest working on this, before #65699?

[AnthonyLatsis]

Yeah, I think that’s a good idea if this is your first time working on code completion.

[Rajveer100]

error: 'clang/Driver/Options.inc' file not found

[Rajveer100]

@AnthonyLatsis What could be the cause for this?

[AnthonyLatsis]

Stale build artifacts, for example. Rule no. 1 (assuming your checkout revisions are in sync): try a clean build.

[Rajveer100]

Nope, still doesn't work, my local repo is in-fact synced and I also tried doing a clean rebuild, same error.

[AnthonyLatsis]

I can't tell much about what went wrong without more context. Can you share the steps you took originally and after my suggestion?

[Rajveer100]

I mean the only thing I did is change my target scheme to swift-ide-test, as you said to test for code completion.

[AnthonyLatsis]

Ah, so you're getting this error in Xcode, not as a result of invoking build-script. Does it build fine if you switch back to swift-frontend? Could you have selected a generated swift-ide-test target/scheme instead of manually creating your own, as you did for swift-frontend per the guide?

[Rajveer100]

Oh yeah, I used the existing one!

[Rajveer100]

So I need to create a new scheme with the target as swift-ide-test? Any other configurations that's needed like it was in swift-frontend?

[AnthonyLatsis]

You need to follow the exact same steps, but for the swift-ide-test executable target. Keep in mind that these steps are opt-in and you can always build/run on the command line.

[Rajveer100]

Now it builds successfully and does run with the following output:

action required
OVERVIEW: Swift IDE Test

USAGE: swift-ide-test [options] [input files...]

...

[Rajveer100]

@AnthonyLatsis I added the following as args... in the scheme for swift-ide-test, what would be the source file that's required as per the log message:

ARGS: swift-ide-test --code-completion --code-completion-token="#^COMPLETE?check=META^#"

LOG: source file required

[AnthonyLatsis]

Good job puzzling out how to output completions for a single token! As in the Swift frontend, this tool requires a path to an input source file to read from: -source-filename <path to file>. Also, -code-completion-token expects a token name/id like COMPLETE, not the token itself.

[Rajveer100]

Got this:

found code completion token COMPLETE at offset 951
Begin completions, 3 items
Keyword[self]/CurrNominal:          self[#(any P).Type#]; name=self
Keyword/CurrNominal:                Protocol[#(any P).Type#]; name=Protocol
Keyword/CurrNominal:                Type[#any P.Type#]; name=Type
End completions
LookedupTypeNames: ['swift_ide_test.P']

[Rajveer100]

I am aware of existential meta types, but could you differentiate it from singleton meta types and also brief me out the log message which I obtained above. Regarding the files, I suppose CompletionLookup::getTypeCompletions and CompletionLookup::getPostfixKeywordCompletions seems to be the place to debug.

[AnthonyLatsis]

The singleton metatype is simply the metatype of a given type T written as T.Type (e.g. Int.Type, (any P).Type). The only possible instance of a Swift metatype is the type object T.self, hence singleton. An existential metatype is not really a metatype, but is called as such because it can be assigned type objects. If any P is a box that wraps a value of dynamic type that conforms to P, then any P.Type, aka an existential metatype, is a box that wraps a value of dynamic metatype whose instance type conforms to P. So if Int conforms to P, then any P can wrap 0, and any P.Type can wrap Int.self.

---

The places to debug should be correct.

[Rajveer100]

Yeah that gave me a good idea! Let me try to debug!

[Rajveer100]

@AnthonyLatsis
Under CompletionLookup::getPostfixKeywordCompletions, we convert the ParsedExpr to a type of AnyMetatypeType,
which is where the check for isExistentialType() succeeds and the following snippet probably causes it:

// Just below "Protocol" keyword addition
addKeyword("Type", ExistentialMetatypeType::get(instanceTy),
                   SemanticContextKind::CurrentNominal);

Is my intuition right?

[Rajveer100]

Do we want an extra bracket () surrounded for the BaseType before calling lookupVisibleMemberDecls ?

[Rajveer100]

@AnthonyLatsis
Am I on the right track here?

[AnthonyLatsis]

which is where the check for isExistentialType() succeeds and the following snippet probably causes it

Yes, this very much looks like the cause. You have to understand that any P.Type and (any P).Type are totally different types semantically. It is not that one type simply has an additional layer of parentheses as in Int vs. (Int):

• base.Type → MetatypeType(base) — a regular, singleton, metatype type.
• any base.Type → ExistentialMetatypeType(base) — an existential metatype type.

Hence (any P).Type → MetatypeType(any P) and any P.Type → ExistentialMetatypeType(P).

[Rajveer100]

In this function:

inline bool CanType::isExistentialTypeImpl(CanType type) {
  return isa<ProtocolType>(type) ||
         isa<ProtocolCompositionType>(type) ||
         isa<ExistentialType>(type) ||
         isa<ParameterizedProtocolType>(type);
}

Why are Protocols considered as isExistentialTypeImpl?

[AnthonyLatsis]

For no good reason. That this has not been known to cause problems is a slightly worrying fluke. Before a separate representation for existential types was adopted in the model, this function worked with what is now known as a constraint type — protocol and protocol composition types. Then (apparently during the transition) someone added the isa<ExistentialType>(type) term . It might have been an intentional technical debt.

[Rajveer100]

At the moment I feel there could be the following possibilities:

• Either the problem lies in the check for existential type.
• Or when we get the expression by using getAs.

[AnthonyLatsis]

The current logic assumes that a postfix Type completion will necessarily produce an existential metatype if the instance type of the parsed type expression isAnyExistentialType():

https://github.com/apple/swift/blob/532a7de23e361b7367b168fa76c1082f69b681d0/lib/IDE/CompletionLookup.cpp#L2437-L2445

This is true for any P but false for (any P): the parentheses do not affect the canonical type of the expression, but they do change the outcome of appending Type.

[Rajveer100]

If the canonical type remains the same shouldn't this check succeed:

!instanceTy->getCanonicalType()->is<MetatypeType>()

[AnthonyLatsis]

I’m not catching on to the causal relationship, but yes, the instance type is not a metatype in either of the above cases.

[AnthonyLatsis]

Let me give the explanation another shot.

• (any P).Type (the singleton metatype of existential type any P) ≠ any P.Type = P.Type (an existential metatype constrained to protocol P).
• For P., the correct Type completion is Type[#any P.Type#] because any P.Type is the modern way of spelling P.Type = P. + Type.
• Ditto any P.. In this case, the result is exactly any P.Type.
• For (any P)., the correct Type completion is Type[#(any P).Type#] because (any P). + Type = (any P).Type.

If you feel unconfident around any/metatypes, you may want to put a pause on this and do some research on them.

[Jamezzzb]

Also from the SE-0335 .md: Spell the existential metatype as any P.Type, and the protocol metatype as (any P).Type.

Reiterating part of the above: any P and (any P) are the same but any P.Type and (any P).Type are different. Which is what the current completion logic seems to be missing out on.