Reordering of attributes passed to `rustc_macro_derive`

[dtolnay]

(Using rustc from #35957)

Macros 1.1 is reordering attributes in a way that breaks Syntex. If I have the following:

#[derive(Serialize, Deserialize)]
#[serde(bound(deserialize = "T::Owned: Deserialize"))]
struct CowT<'a, T: ?Sized + 'a + ToOwned>(Cow<'a, T>);

It gets passed to my #[rustc_macro_derive(Serialize)] as:

#[serde(bound(deserialize = "T::Owned: Deserialize"))]
#[derive(Deserialize)]
struct CowT<'a, T: ?Sized + 'a + ToOwned>(Cow<'a, T>);

I would expect the #[derive(Deserialize)] to stay above #[serde(...)].

cc @alexcrichton

[alexcrichton]

Hm I'm not actually sure how to fix this offhand. When the #[derive] syntax extension is running the #[derive] attribute has already been removed, and then when we call the first Serialize implementation we manually push back on the #[derive(Deserialize)]. We could push it at the front instead of the back but I have a feeling that's probably also not the right place :(

Might require a more intrusive change to #[derive] to not remove the attribute at the beginning.

[nrc]

We could push it at the front instead of the back but I have a feeling that's probably also not the right place

Why? Putting the remaining derives back where the one we just took out seems like the right thing to do to me.

[alexcrichton]

Oh right yeah we should put the #[derive] attribute back where we found it, but I don't think that information is available via the kind of syntax extension #[derive] is today?

[dtolnay]

This only affected syntex and is no longer required as we move our rustc_macro away from syntex (serde-rs/serde#548).

[dtolnay]

Without syntex this no longer matters.