mirror of
https://github.com/bevyengine/bevy
synced 2024-09-20 14:32:06 +00:00
6183b56b5d
# Objective The goal with this PR is to allow the use of types that don't implement `Reflect` within the reflection API. Rust's [orphan rule](https://doc.rust-lang.org/book/ch10-02-traits.html#implementing-a-trait-on-a-type) prevents implementing a trait on an external type when neither type nor trait are owned by the implementor. This means that if a crate, `cool_rust_lib`, defines a type, `Foo`, then a user cannot use it with reflection. What this means is that we have to ignore it most of the time: ```rust #[derive(Reflect)] struct SomeStruct { #[reflect(ignore)] data: cool_rust_lib::Foo } ``` Obviously, it's impossible to implement `Reflect` on `Foo`. But does it *have* to be? Most of reflection doesn't deal with concrete types— it's almost all using `dyn Reflect`. And being very metadata-driven, it should theoretically be possible. I mean, [`serde`](https://serde.rs/remote-derive.html) does it. ## Solution > Special thanks to @danielhenrymantilla for their help reviewing this PR and offering wisdom wrt safety. Taking a page out of `serde`'s book, this PR adds the ability to easily use "remote types" with reflection. In this context, a "remote type" is the external type for which we have no ability to implement `Reflect`. This adds the `#[reflect_remote(...)]` attribute macro, which is used to generate "remote type wrappers". All you have to do is define the wrapper exactly the same as the remote type's definition: ```rust // Pretend this is our external crate mod cool_rust_lib { #[derive(Default)] struct Foo { pub value: String } } #[reflect_remote(cool_rust_lib::Foo)] struct FooWrapper { pub value: String } ``` > **Note:** All fields in the external type *must* be public. This could be addressed with a separate getter/setter attribute either in this PR or in another one. The macro takes this user-defined item and transforms it into a newtype wrapper around the external type, marking it as `#[repr(transparent)]`. The fields/variants defined by the user are simply used to build out the reflection impls. Additionally, it generates an implementation of the new trait, `ReflectRemote`, which helps prevent accidental misuses of this API. Therefore, the output generated by the macro would look something like: ```rust #[repr(transparent)] struct FooWrapper(pub cool_rust_lib::Foo); impl ReflectRemote for FooWrapper { type Remote = cool_rust_lib::Foo; // transmutation methods... } // reflection impls... // these will acknowledge and make use of the `value` field ``` Internally, the reflection API will pass around the `FooWrapper` and [transmute](https://doc.rust-lang.org/std/mem/fn.transmute.html) it where necessary. All we have to do is then tell `Reflect` to do that. So rather than ignoring the field, we tell `Reflect` to use our wrapper using the `#[reflect(remote = ...)]` field attribute: ```rust #[derive(Reflect)] struct SomeStruct { #[reflect(remote = FooWrapper)] data: cool_rust_lib::Foo } ``` #### Other Macros & Type Data Because this macro consumes the defined item and generates a new one, we can't just put our macros anywhere. All macros that should be passed to the generated struct need to come *below* this macro. For example, to derive `Default` and register its associated type data: ```rust // ✅ GOOD #[reflect_remote(cool_rust_lib::Foo)] #[derive(Default)] #[reflect(Default)] struct FooWrapper { pub value: String } // ❌ BAD #[derive(Default)] #[reflect_remote(cool_rust_lib::Foo)] #[reflect(Default)] struct FooWrapper { pub value: String } ``` #### Generics Generics are forwarded to the generated struct as well. They should also be defined in the same order: ```rust #[reflect_remote(RemoteGeneric<'a, T1, T2>)] struct GenericWrapper<'a, T1, T2> { pub foo: &'a T1, pub bar: &'a T2, } ``` > Naming does *not* need to match the original definition's. Only order matters here. > Also note that the code above is just a demonstration and doesn't actually compile since we'd need to enforce certain bounds (e.g. `T1: Reflect`, `'a: 'static`, etc.) #### Nesting And, yes, you can nest remote types: ```rust #[reflect_remote(RemoteOuter)] struct OuterWrapper { #[reflect(remote = InnerWrapper)] pub inner: RemoteInner } #[reflect_remote(RemoteInner)] struct InnerWrapper(usize); ``` #### Assertions This macro will also generate some compile-time assertions to ensure that the correct types are used. It's important we catch this early so users don't have to wait for something to panic. And it also helps keep our `unsafe` a little safer. For example, a wrapper definition that does not match its corresponding remote type will result in an error: ```rust mod external_crate { pub struct TheirStruct(pub u32); } #[reflect_remote(external_crate::TheirStruct)] struct MyStruct(pub String); // ERROR: expected type `u32` but found `String` ``` <details> <summary>Generated Assertion</summary> ```rust const _: () = { #[allow(non_snake_case)] #[allow(unused_variables)] #[allow(unused_assignments)] #[allow(unreachable_patterns)] #[allow(clippy::multiple_bound_locations)] fn assert_wrapper_definition_matches_remote_type( mut __remote__: external_crate::TheirStruct, ) { __remote__.0 = (|| -> ::core::option::Option<String> { None })().unwrap(); } }; ``` </details> Additionally, using the incorrect type in a `#[reflect(remote = ...)]` attribute should result in an error: ```rust mod external_crate { pub struct TheirFoo(pub u32); pub struct TheirBar(pub i32); } #[reflect_remote(external_crate::TheirFoo)] struct MyFoo(pub u32); #[reflect_remote(external_crate::TheirBar)] struct MyBar(pub i32); #[derive(Reflect)] struct MyStruct { #[reflect(remote = MyBar)] // ERROR: expected type `TheirFoo` but found struct `TheirBar` foo: external_crate::TheirFoo } ``` <details> <summary>Generated Assertion</summary> ```rust const _: () = { struct RemoteFieldAssertions; impl RemoteFieldAssertions { #[allow(non_snake_case)] #[allow(clippy::multiple_bound_locations)] fn assert__foo__is_valid_remote() { let _: <MyBar as bevy_reflect::ReflectRemote>::Remote = (|| -> ::core::option::Option<external_crate::TheirFoo> { None })().unwrap(); } } }; ``` </details> ### Discussion There are a couple points that I think still need discussion or validation. - [x] 1. `Any` shenanigans ~~If we wanted to downcast our remote type from a `dyn Reflect`, we'd have to first downcast to the wrapper then extract the inner type. This PR has a [commit](b840db9f74cb6d357f951cb11b150d46bac89ee2) that addresses this by making all the `Reflect::*any` methods return the inner type rather than the wrapper type. This allows us to downcast directly to our remote type.~~ ~~However, I'm not sure if this is something we want to do. For unknowing users, it could be confusing and seemingly inconsistent. Is it worth keeping? Or should this behavior be removed?~~ I think this should be fine. The remote wrapper is an implementation detail and users should not need to downcast to the wrapper type. Feel free to let me know if there are other opinions on this though! - [x] 2. Implementing `Deref/DerefMut` and `From` ~~We don't currently do this, but should we implement other traits on the generated transparent struct? We could implement `Deref`/`DerefMut` to easily access the inner type. And we could implement `From` for easier conversion between the two types (e.g. `T: Into<Foo>`).~~ As mentioned in the comments, we probably don't need to do this. Again, the remote wrapper is an implementation detail, and should generally not be used directly. - [x] 3. ~~Should we define a getter/setter field attribute in this PR as well or leave it for a future one?~~ I think this should be saved for a future PR - [ ] 4. Any foreseeable issues with this implementation? #### Alternatives One alternative to defining our own `ReflectRemote` would be to use [bytemuck's `TransparentWrapper`](https://docs.rs/bytemuck/1.13.1/bytemuck/trait.TransparentWrapper.html) (as suggested by @danielhenrymantilla). This is definitely a viable option, as `ReflectRemote` is pretty much the same thing as `TransparentWrapper`. However, the cost would be bringing in a new crate— though, it is already in use in a few other sub-crates like bevy_render. I think we're okay just defining `ReflectRemote` ourselves, but we can go the bytemuck route if we'd prefer offloading that work to another crate. --- ## Changelog * Added the `#[reflect_remote(...)]` attribute macro to allow `Reflect` to be used on remote types * Added `ReflectRemote` trait for ensuring proper remote wrapper usage |
||
|---|---|---|
| .. | ||
| src | ||
| tests | ||
| .gitignore | ||
| Cargo.toml | ||
| README.md | ||
Compile fail tests for bevy_reflect
This crate is separate from bevy_reflect and not part of the Bevy workspace in order to not fail crater tests for
Bevy.
The tests assert on the exact compiler errors and can easily fail for new Rust versions due to updated compiler errors (e.g. changes in spans).
The CI workflow executes these tests on the stable rust toolchain (see tools/ci).
For information on writing tests see compile_fail_utils/README.md.