Clarify behviour when writing to a union field that implements Drop

[joshtriplett]

Discussion in the tracking issue for unions turned up a corner case not addressed by RFC1444: whether assigning to a union field that implements Drop should drop the previous value of the field. Based on that discussion, document that such an assignment does drop the previous value of the field. Also document how to work around that if needed, and add a note that the lint warning for declaring a union field with a type implementing Drop should explicitly document this issue in its explanation.

[nrc]

I think this should be fully specified one way or the other. And, nit, we don't usually use the terminology 'pragma'.

[strega-nil]

What do you mean by "fully specified"?

[joshtriplett]

I assume this means s/may optionally/should/ here.

[nrc]

It would be helpful if you could summarise the discussion from the tracking issue which led to the decision in this RFC, thanks!

[joshtriplett]

@nrc Done.

[nrc]

This RFC PR was discussed (briefly) in the lang team meeting and is now 🔔 entering the final comment period 🔔.

[strega-nil]

union Union {
  string: String,
  vec: Vec<u32>,
}
let mut x: Union;
x = Union { string: "".to_owned() };
x = Union { vec: vec![] };
// both of these are assigning to mutable lvalues of type Union
// in the second case, this is equivalent to 
// `drop_in_place(&mut x); write(&mut x, Union { vec: vec![] });`
// this doesn't run any drop glue, because Union doesn't inherit 
// drop glue from it's internal members (Union only has drop glue
// if it literally impls Drop)

// however
x.vec = vec![];
// in this case you're assigning to an lvalue of type Vec<u32>
// which is equivalent to `drop_in_place(&mut x.vec); write(&mut x.vec, vec![])`
// which obviously drops the vector

Basically, what I'm saying is, 👍 to this change :)

[aturon]

I'm in favor of this specification. It's a subtle issue, but I agree with the conclusion of the earlier discussion that on the whole, this RFC's behavior provides the greatest degree of consistency.

The use of a lint to alert users of the potential footgun here is very good; I expect we won't use many unions with Drop data, except as a means of explicitly controlling when drops happen.

[nikomatsakis]

Huzzah! The @rust-lang/lang team has decided to accept this RFC.