Decorator metadata breaking change between TS 2.3.4 -> 2.4.0

[flushentitypacket]

I couldn't find any documentation on this breaking change between TS 2.3.4 and TS 2.4.0.

I'm not sure whether this is a bug or simply an undocumented change. I suspect the latter--in which case my suggestion is to add documentation regarding that change.

TypeScript Version: >= 2.4.0

Ran tsc with flags --experimentalDecorators --emitDecoratorMetadata.

Using reflect-metadata v0.1.10

Code

import 'reflect-metadata'

const dec = (obj: {}, prop: string) => undefined

class Foo {
  @dec public foo: string | null // seems to happen with any complex type
  @dec public bar: string
}

console.log(Reflect.getMetadata('design:type', Foo.prototype, 'foo'))
console.log(Reflect.getMetadata('design:type', Foo.prototype, 'bar'))

Expected behavior:

output in 2.3.4:

[Function: String]
[Function: String]

Actual behavior:

output in 2.4.0:

[Function: Object]
[Function: String]

[awendland]

Related to #9916 but different because #9916 is a feature request while this is a request to change back to previous behavior, or at least document the breaking change. It looks like this commit is what changed the behavior 0eaa8eb.

[awendland]

Running git tag --contains 0eaa8eb3ab9ad1db8cf95c51d8e02c9d65ee40cf reports v2.4-rc, v2.4.0, v2.4.1, v2.4.2, v2.5-rc, v2.5.0, v2.5.1, and v2.5.2.

[ulrichb]

Hmmm. Using the current TS 2.8.3 and still see this issue. Can someone confirm this?

[weswigham]

Do you have strictNullChecks on?

[ulrichb]

Yes.

BTW, this works: @myDec() public PropN?: boolean;

The following doesn't work:

@myDec() public PropN: boolean | undefined;

or

@myDec() public PropN: boolean | null = null;

[weswigham]

Yeah, if you read the linked fix PR - #19089 - this is the expected behavior under strictNullChecks, as a number | undefined isn't just a Number - it's a Number or undefined (which there's no corresponding way to emit metadata for). We don't emit metadata for nonhomogenous unions, as decorator metadata is syntactic (hence why the ? behaves differently than | undefined), and emit the base Object instead.

[ulrichb]

Okay, this was a misunderstanding. I thought this also concerns strict checking. Thanks for clarification.

[ulrichb]

Regarding the general union "issue" of reflecting to Object: I found a few "closed by design" issues.

Is there really no possibility to get this information? Are there no plans to fully support this scenario?

[mhegazy]

We have no plans to implement a full run-time type serialization system akin to reflection C# or Java (more in #3628). The meta-data emit, continues to be an experimental feature tied to decorators, and has a very limited scope (i.e. classes and primitives). we have no plans to expand that scope either.

[ulrichb]

@mhegazy Thanks for clarification.