portrait
Fill impl-trait blocks with default, delegation and more.
Motivation
Rust traits support provided methods, which are great for backwards compatibility and implementation coding efficiency. However they come with some limitations:
- There is no reasonable way to implement an associated function if its return type is an associated type.
- If a trait contains many highly similar associated functions, writing the defaults involves a lot of boilerplate. But users can only provide one default implementation for each method through procedural macros.
With portrait
, the default implementations are provided at impl
-level granularity instead of trait-level.
Usage
First of all, make a portrait of the trait to implement with the #[portrait::make]
attribute:
#[portrait::make]
trait FooBar {
// ...
}
Implement the trait partially and leave the rest to the #[portrait::fill]
attribute:
#[portrait::fill(portrait::default)]
impl FooBar {}
The portrait::default
part is the path to the "filler macro", which is the item that actually fills the impl
block. The syntax is similar to #[derive(path::to::Macro)]
.
If there are implementations in a different module, the imports used in trait items need to be manually passed to the make attribute:
#[portrait::make(import(
foo, bar::*,
// same syntax as use statements
))]
trait FooBar {...}
If the fill attribute fails with an error about undefined foo_bar_portrait
, import it manually together with the FooBar trait; the #[portrait::make]
attribute generates this new module in the same module as the FooBar
trait.
Provided fillers
portrait
provides the following filler macros:
default
: Implements each missing method and constant by delegating toDefault::default()
(Default
is const-unstable and requires nightly with#![feature(const_default_impls)]
).delegate
: Proxies each missing method, constant and type to an expression (usuallyself.field
) or another type implementing the same trait.lo
]: Calls aformat!
-like macro with the method arguments.
How this works
Rust macros are invoked at an early stage of the compilation chain. As a result, attribute macros only have access to the literal representation of the item they are applied on, and cross-item derivation is not directly possible. Most macros evade this problem by trying to generate code that works regardless of the inaccessible information, e.g. the Default
derive macro works by invoking Default::default()
without checking whether the actual field type actually implements Default
(since the compiler would do at a later stage anyway).
Unfortunately this approach does not work in the use case of portrait
, where the attribute macro requires compile time (procedural macro runtime) access to the items of the trait referenced in the impl
block; the only available information is the path to the trait (which could even be renamed to a different identifier).
portrait
addresses this challenge by asking the trait to export its information (its "portrait") in the form of a token stream in advance. Through the #[portrait::make]
attribute, a declarative macro with the same identifier is derived, containing the trait items. The (#[portrait::fill]
) attribute on the impl
block then passes its inputs to the declarative macro, which in turn forwards them to the actual attribute implementation (e.g. #[portrait::make]
) along with the trait items.
Now the actual attribute has access to both the trait items and the user impl, but that's not quite yet the end of story. The trait items are written in the scope of the trait definition, but the attribute macro output is in the scope of the impl definition. The most apparent effect is that imports in the trait module do not take effect on the impl output. To avoid updating implementors frequently due to changes in the trait module, the #[portrait::make]
attribute also derives a module that contains the imports used in the trait to be automatically imported in the impl block.
It turns out that, as of current compiler limitations, private imports actually cannot be re-exported publicly even though the imported type is public, so it becomes impractical to scan the trait item automatically for paths to re-export (prelude types also need special treatment since they are not part of the module). The problem of heterogeneous scope thus becomes exposed to users inevitably: either type all required re-exports manually through import, or make the imports visible to a further scope.
Another difficulty is that module identifiers share the same symbol space as trait identifiers (because module::Foo
is indistinguishable from Trait::Foo
). Thus, the module containing the imports cannot be imported together with the trait, and users have to manually import/export both symbols unless the trait is referenced through its enclosing module.
Disclaimer
portrait
is not the first one to use declarative macros in attributes. macro_rules_attribute
also implements a similar idea, although without involving the framework of generating the macro_rules!
part.