Even for a simple data structure that would ordinarily expand to an impl
that only refers to the public Serde traits without serde::__private, we
still disallow a (renamed) serde_core dependency. This leaves the
liberty to new usage of private helpers in such impls over time. For
example, similar to the optimization of the standard library's
derive(Debug) that introduced debug_struct_field1_finish to improve
compile time of derived impls.
This is intended to be used by other crates which provide their own proc
macros and use serde internally. Today there's no consistent way to put
`#[derive(Deserialize)]` on a struct that consistently works, since
crates may be using either `features = ["derive"]` or relying on
`serde_derive` separately.
Even if we assume that everyone is using `features = ["derive"]`,
without this commit, any crate which generates
`#[derive(serde::Deserialize)]` forces its consumers to put `serde` in
their `Cargo.toml`, even if they aren't otherwise using serde for
anything.
Examples of crates which suffer from this in the real world are
tower-web and swirl.
With this feature, it's expected that these crates would have `pub
extern crate serde;` in some accessible path, and add
`#[serde(serde_path = "that_crate::wherever::serde")]` anywhere they
place serde's derives. Those crates would also have to derive
`that_crate::whatever::serde::Deserialize`, or `use` the macros
explicitly beforehand.
The test for this is a little funky, as it's testing this in a way that
is not the intended use case, or even one we want to support. It has its
own module which re-exports all of serde, but defines its own
`Serialize` and `Deserialize` traits. We then test that we generated
impls for those traits, instead of serde's. The only other way to test
this would be to create a new test crate which does not depend on serde,
but instead depends on `serde_derive` and a third crate which publicly
re-exports serde. This feels like way too much overhead for a single
test case, hence the funky test given.
I didn't see anywhere in this repo to document this attribute, so I
assume the docs will have to be done as a separate PR to a separate
repo.
Fixes#1487
David Tolnay