lang (syntax | attributes)
param_attrs
Allow attributes in formal function parameter position. For example, consider a Jax-Rs-style HTTP API:
#[resource(path = "/foo/bar")]
impl MyResource {
#[get("/person/:name")]
fn get_person(
&self,
#[path_param = "name"] name: String, // <-- formal function parameter.
#[query_param = "limit"] limit: Option<u32>, // <-- here too.
) {
...
}
}
Allowing attributes on formal function parameters enables external tools and compiler internals to take advantage of the additional information that the attributes provide.
Conditional compilation with #[cfg(..)]
is also
facilitated by allowing more ergonomic addition and removal of parameters.
Moreover, procedural macro authors can use annotations on these parameters and thereby richer DSLs may be encoded by users. We already saw an example of such a DSL in the summary. To further illustrate potential usages, let's go through a few examples.
#[rustc_args_required_const]
A number of platform intrinsics are currently provided by rust compilers.
For example, we have core::arch::wasm32::memory_grow
which,
for soundness reasons, requires that when memory_grow
is applied,
mem
must provided a const
expression:
#[rustc_args_required_const(0)]
pub fn memory_grow(mem: u32, delta: usize) -> usize { .. }
This is specified in a positional manner, referring to mem
by 0
.
While this is serviceable, this RFC enables us encode the invariant more directly:
pub fn memory_grow(
#[rustc_args_required_const] mem: u32,
delta: usize
) -> usize {
..
}
Property based testing a la QuickCheck allows users to state properties they expect their programs to adhere to. These properties are then tested by randomly generating input data and running the properties with those. The properties are can then be falsified by finding counter-examples. If no such example are found, the test passes and the property is "verified". In the Rust ecosystem, property based testing is primarily provided by the proptest and quickcheck crates where the former uses integrated shrinking whereas the latter uses type based shrinking.
Consider a case where we want to test a "polymorphic" function on a number of concrete types.
#[proptest] // N.B. Using proptest doesn't look like this today.
fn prop_my_property(#[types(T = u8, u16, u32)] elem: Vec<T>, ..) { .. }
Here, we've overloaded the test for the types u8
, u16
, and u32
.
The test will then act as if you had written:
#[proptest]
fn prop_my_property_u8(elem: Vec<u8>, ..) { .. }
#[proptest]
fn prop_my_property_u16(elem: Vec<u16>, ..) { .. }
#[proptest]
fn prop_my_property_u32(elem: Vec<u32>, ..) { .. }
By allowing attributes on function parameters, the test can be specified more succinctly and without repetition as done in the first example.
There's interest in using attributes on function parameters for
#[wasm_bindgen]
. For example, to interoperate well
with TypeScript's type system, you could write:
#[wasm_bindgen]
impl RustLayoutEngine {
#[wasm_bindgen(constructor)]
pub fn new() -> Self { Default::default() }
#[wasm_bindgen(typescript(return_type = "MapNode[]"))]
pub fn layout(
&self,
#[wasm_bindgen(typescript(type = "MapNode[]"))]
nodes: Vec<JsValue>,
#[wasm_bindgen(typescript(type = "MapEdge[]"))]
edges: Vec<JsValue>
) -> Vec<JsValue> {
..
}
}
Currently, in #[wasm_bindgen]
, the arguments and return type of layout
are all any[]
. By using allowing the annotations above, tighter types can
be used which can help in catching problems at compile time rather than
having UI bugs later.
For raw pointers that are oftentimes used when operating with C code, additional information could be given to the compiler about the set of parameters. You could for example mirror C's restrict keyword or even be more explicit by stating which pointer arguments may overlap:
fn foo(
#[overlaps_with(in_b)] in_a: *const u8,
#[overlaps_with(in_a)] in_b: *const u8,
#[restrict] out: *mut u8
);
This would tell the compiler or some static analysis tool that the pointers
in_a
and in_b
might overlap but out
is non overlapping. Note that neither
overlaps_with
and restrict
are part of this proposal; rather, they are
examples of what this RFC facilities.
In today's Rust it is possible to prefix the name of an identifier to silence
the compiler about it being unused. With attributes on formal parameters,
we could hypothetically have an attribute like #[unused]
that explicitly
states this for a given parameter. Note that #[unused]
is not part of this
proposal but merely a simple use-case. In other words, we could write (1):
fn foo(#[unused] bar: u32) -> bool { .. }
instead of (2):
fn foo(_bar: u32) -> bool { .. }
Especially Rust beginners might find the meaning of (1) to be clearer than (2).
Formal parameters of fn
definitions as well closures parameters may have
attributes attached to them. Thereby, additional information may be provided.
For the purposes of illustration, let's assume we have the attributes
#[orange]
and #[lemon]
available to us.
The syntax for attaching attributes to parameters is shown in the snippet below:
// Free functions:
fn foo(#[orange] bar: u32) { .. }
impl Alpha { // In inherent implementations.
// - `self` can also be attributed:
fn bar(#[lemon] self, #[orange] x: u8) { .. }
fn baz(#[lemon] &self, #[orange] x: u8) { .. }
fn quux(#[lemon] &mut self, #[orange] x: u8) { .. }
..
}
impl Beta for Alpha { // Also works in trait implementations.
fn bar(#[lemon] self, #[orange] x: u8) { .. }
fn baz(#[lemon] &self, #[orange] x: u8) { .. }
fn quux(#[lemon] &mut self, #[orange] x: u8) { .. }
..
}
fn foo() {
// Closures:
let bar = |#[orange] x| { .. };
let baz = |#[lemon] x: u8, #[orange] y| { .. };
}
An fn
definition doesn't need to have a body to permit parameter attributes.
Thus, in trait
definitions, we may write:
trait Beta {
fn bar(#[lemon] self, #[orange] x: u8);
fn baz(#[lemon] &self, #[orange] x: u8);
fn quux(#[lemon] &mut self, #[orange] x: u8);
}
In Rust 2015, since anonymous parameters are allowed, you may also write:
trait Beta {
fn bar(#[lemon] self, #[orange] u8); // <-- Note the absence of `x`!
}
fn
typesYou can also use attributes in function pointer types. For example, you may write:
type Foo = fn(#[orange] x: u8);
type Bar = fn(#[orange] String, #[lemon] y: String);
Attributes attached to formal parameters do not have an inherent meaning in the type system or in the language. Instead, the meaning is what your procedural macros, the tools you use, or what the compiler interprets certain specific attributes as.
As for the built-in attributes and their semantics, we will, for the time being, only permit the following attributes on parameters:
Lint check attributes, that is:
#[allow(C)]
, #[warn(C)]
, #[deny(C)]
, #[forbid(C)]
,
and tool lint attributes such as #[allow(clippy::foobar)]
.
Conditional compilation attributes:
#[cfg_attr(...)]
, e.g.
fn foo(#[cfg_attr(bar, orange)] x: u8) { .. }
If bar
is active, this is equivalent to:
fn foo(#[orange] x: u8) { .. }
And otherwise equivalent to:
fn foo(x: u8) { .. }
#[cfg(...)]
, e.g.
fn foo(#[cfg(bar)] x: u8, y: u16) { .. }
If bar
is active, this is equivalent to:
fn foo(x: u8, y: u16) { .. }
And otherwise equivalent to:
fn foo(y: u16) { .. }
All other built-in attributes will be rejected with a semantic check. For example, you may not write:
fn foo(#[inline] bar: u32) { .. }
Let OuterAttr
denote the production for an attribute #[...]
.
On the formal parameters of an fn
item, including on method receivers,
and irrespective of whether the fn
has a body or not, OuterAttr+
is allowed
but not required. For example, all the following are valid:
fn g1(#[attr1] #[attr2] pat: Type) { .. }
fn g2(#[attr1] x: u8) { .. }
fn g3(#[attr] self) { .. }
fn g4(#[attr] &self) { .. }
fn g5<'a>(#[attr] &mut self) { .. }
fn g6<'a>(#[attr] &'a self) { .. }
fn g7<'a>(#[attr] &'a mut self) { .. }
fn g8(#[attr] self: Self) { .. }
fn g9(#[attr] self: Rc<Self>) { .. }
The attributes here apply to the parameter as a whole,
e.g. in g2
, #[attr]
applies to pat: Type
as opposed to pat
.
More generally, an fn
item contains a list of formal parameters separated or
terminated by ,
and delimited by (
and )
. Each parameter in that list may
optionally be prefixed by OuterAttr+
.
Attributes may also be attached to ...
on variadic functions, e.g.
extern "C" {
fn foo(x: u8, #[attr] ...);
}
That is, for the purposes of this RFC, ...
is considered as a parameter.
In Rust 2015 edition, as fn
s may have anonymous parameters, e.g.
trait Foo { fn bar(u8); }
attributes are allowed on those, e.g.
trait Foo { fn bar(#[attr] u8); }
fn
pointersAssuming roughly the following type grammar for function pointers (in the lykenware/gll notation):
Type =
| ..
| FnPtr:{
binder:ForAllBinder? unsafety:"unsafe"? { "extern" abi:Abi }?
"fn" "(" inputs:FnSigInputs? ","? ")" { "->" ret_ty:Type }?
}
;
FnSigInputs =
| Regular:FnSigInput+ % ","
| Variadic:VaradicTail
| RegularAndVariadic:{ inputs:FnSigInput+ % "," "," "..." }
;
VaradicTail = "...";
FnSigInput = { pat:Pat ":" }? ty:Type;
we change VaradicTail
to:
VaradicTail = OuterAttr* "...";
and change FnSigInput
to:
FnSigInput = OuterAttr* { pat:Pat ":" }? ty:Type;
Similar to parameters in fn
items, the attributes here also apply to the
pattern and the type if both are present, i.e. pat: ty
as opposed to pat
.
Given roughly the following expression grammar for closures:
Expr = attrs:OuterAttr* kind:ExprKind;
ExprKind =
| ..
| Closure:{
by_val:"move"?
"|" args:ClosureArg* % "," ","? "|" { "->" ret_ty:Type }? body:Expr
}
;
ClosureArg = pat:Pat { ":" ty:Type }?;
we change ClosureArg
into:
ClosureArg = OuterAttr* pat:Pat { ":" ty:Type }?;
As before, when the type is specified, OuterAttr*
applies to pat: Type
as opposed to just pat
.
Attributes on formal parameters of functions, closures and function pointers have no inherent meaning in the type system or elsewhere. Semantics, if there are any, are given by the attributes themselves on a case by case basis or by tools external to a Rust compiler.
The built-in attributes that are permitted on the parameters are:
lint check attributes including tool lint attributes.
cfg_attr(..)
unconditionally.
cfg(..)
unconditionally.
When a cfg(..)
is active, the formal parameter will be included
whereas if it is inactive, the formal parameter will be excluded.
All other built-in attributes are for the time being rejected with a semantic check resulting in a compilation error.
Finally, a registered #[proc_macro_attribute]
may not be attached directly
to a formal parameter. For example, if given:
#[proc_macro_attribute]
pub fn attr(args: TokenStream, input: TokenStream) -> TokenStream { .. }
then it is not legal to write:
fn foo(#[attr] x: u8) { .. }
No changes.
All drawbacks for attributes in any location also count for this proposal.
Having attributes in many different places of the language complicates its grammar.
This proposal goes the path of having attributes in more places of the language. It nicely plays together with the advance of procedural macros and macros 2.0 where users can define their own attributes for their special purposes.
An alternative to having attributes for formal parameters might be to just use the current set of available attributable items to store meta information about formal parameters like in the following example:
#[ignore(param = bar)]
fn foo(bar: bool);
An example of this is #[rustc_args_required_const]
as discussed
in the motivation.
Note that this does not work in all situations (for example closures) and might involve even more complexity in user's code than simply allowing attributes on formal function parameters.
The impact will be that users might create custom attributes when designing procedural macros involving formal function parameters.
There should be no breakage of existing code.
fn
pointersIn this proposal it is legal to write #[attr] ...
as well as fn(#[attr] u8)
.
The primary justification for doing so is that conditional compilation with
#[cfg(..)]
is facilitated. Moreover, since the fn
type grammar and
that of fn
items is somewhat shared, and since ...
is the tail of a
list, allowing attributes there makes for a simpler grammar.
Some example languages that allow for attributes on formal function parameter positions are Java, C#, and C++.
Also note that attributes in other parts of the Rust language could be considered prior art to this proposal.
None as of yet.
In the pursuit of allowing more flexible DSLs and more ergonomic conditional compilation, RFC 2602 builds upon this RFC.
In this RFC, we have not allowed documentation comments on parameters. For example, you may not write:
fn foo(
/// Some description about `bar`.
bar: u32
) {
..
}
Neither may you write the desugared form:
fn foo(
#[doc = "Some description about `bar`."]
bar: u32
) {
..
}
In the future, we may want to consider supporting this form of documentation.
This will require support in rustdoc
to actually display the information.
#[proc_macro_attribute]
In this RFC we stated that fn foo(#[attr] x: u8) { .. }
,
where #[attr]
is a #[proc_macro_attribute]
is not allowed.
In the future, if use cases arise to justify a change, we could lift this
restriction such that transformations can be done directly on x: u8
.