Allowed Enum variants to be individually marked as untagged

[dewert99]

This should hopefully resolve #1402.

This can be thought of as applying the following transformation

#[derive(Serialize, Deserialize)
enum Test {
  A(u32, u32),
  B(u32),
  #[serde(untagged)]
  C(u32),
  #[serde(untagged)]
  D(u32, u32),
}

to

#[derive(Serialize, Deserialize)
enum TestTagged {
  A(u32, u32),
  B(u32),
}

#[derive(Serialize, Deserialize)
#[serde(untagged)]
enum TestUntagged {
  Tagged(TestTagged),
  C(u32),
  D(u32, u32)
}

#[derive(Serialize, Deserialize)
#[serde(from = "TestUntagged")]
#[serde(into = "TestUntagged")]
enum Test {
  ...
}

The actual implementation generates the implementation directly allowing it to also work efficiently for recursive enums.

The current version includes a restriction that all variants marked as untagged appear at the end of the enum (after the tagged variants) in order to keep the implementation strategy consistent with an alternate strategy of attempting to deserialize each variant in order.

Deserialize could also be implemented to accept all of the variants with tags, only falling back to untagged if none of them match, I'm not sure if this would be preferred.

[oli-obk]

Please also add some tests for non-recursive examples like in #1402 (comment) or generic examples like #1402 (comment)

[dewert99]

Thanks for all the feedback. I've tried to implement the suggestions.

[oli-obk]

one thing I just realized that we may want to reject is marking all enum variants as untagged, as then you could instead mark the entire enum as untagged.

lgtm otherwise. I'll ping a few folk from the issue so they can try out your PR and see if it works for their use cases.

[dewert99]

I wasn't sure what the policy was on warning-like cases. Other cases to consider might be:

• marking a variant as untagged when the enum itself is already untagged
• marking a variant with rename/alias as well as untagged (I don't think this has a different effect than just using untagged)
• marking a variant with rename/alias when the enum is untagged (This is currently allowed, so changing it would be a breaking change)

[gavrilikhin-d]

Why only allowing #[serde(untagged)] to be at the end?

[dewert99]

This keeps the implementation more consistent with the deserialisation strategy of going though each variant in order and choosing the first variant that matches.

[gavrilikhin-d]

@oli-obk Hi! This feature is really useful, but sadly there is no activity on PR more than a month. The PR itself looks quite finished except warnings. Can it be merged in this state?

[oli-obk]

Unfortunately no one on the issue has so far tested out this branch to see if it works for their use cases.

[gavrilikhin-d]

@oli-obk I tried to check it, but don't know how to replace serde dependency in serde_json

[oli-obk]

Someone was so kind and wrote instructions at #1402 (comment)

[gavrilikhin-d]

@oli-obk My case was covered too. #1402 (comment)

[dtolnay]

Thanks!