5 releases
Uses old Rust 2015
0.1.1474613452 | Sep 23, 2016 |
---|---|
0.1.5 | Sep 23, 2016 |
0.1.3 | Sep 18, 2016 |
0.1.2 | Sep 16, 2016 |
0.1.1 | Sep 16, 2016 |
#395 in Procedural macros
2KB
Named arguments for Rust
An experimental (probably broken) procedural macro. Instead of code like this:
parser.parse_item(vec![], true, false)
this macro lets you write code like this!
parser.parse_item(attrs: vec![], macros_allowed: true, attributes_allowed: false)
Written partly as a proof of concept for a potential future Rust feature, but also to be practical and usable today. Sort of. In theory.
Syntax
Strongly Swift-inspired.
Like Swift, argument names form a mandatory part of the function signature, and are not reorderable; also, an argument may have separate 'external' and 'internal' names, where the former is used when calling the function, and the latter is the name of the resulting variable.
In fact, the declaration syntax is quite similar to Swift, but with two forms swapped around:
Variant | Rust | Swift | Called as | Variable name |
---|---|---|---|---|
Unnamed | fn foo(bar: String) |
func foo(_ bar: String) |
foo(string) |
bar |
Basic named | fn foo(_ bar: String) |
func foo(bar: String) |
foo(bar: string) |
bar |
Different names | fn foo(bar baz: String) |
func foo(bar baz: String) |
foo(bar: string) |
baz |
Swift defaults to named arguments and uses _
for unnamed arguments, while the
extended Rust syntax defaults to unnamed (for compatibility with existing
Rust code), and uses _
to indicate a named argument, with equivalent
external and internal names. For different external and internal names, write
them separated by a space.
Named arguments are implemented by appending the argument names to the function name, at both definitions and call sites. As a natural consequence:
-
You can overload functions with the same name but different argument labels. This doesn't have the same pitfalls as purely type-based overloading as seen in C++/Java/etc., so there's no reason not to use it. For example, consider constructors. For zero- and one-argument constructors, Rust's existing convention of
::new()
and::with_foo(foo)
works pretty well:HashMap::new() HashMap::with_capacity(100)
But with two or more arguments, it starts to get ugly:
HashMap::with_capacity_and_hasher(100, Default::default())
In an API designed with named arguments, the progression can be more natural:
HashMap::new() HashMap::new(capacity: 100) HashMap::new(capacity: 100, hasher: Default::default())
(Note: different numbers of unnamed arguments can also be distinguished, but only after the first named argument.)
-
Named argument syntax works both for bare fns and trait methods (in both trait definitions and impls). But it doesn't work in either lambdas or function types (either bare
fn(...)
or theFn/FnMut/FnOnce(...)
traits); the type drops argument names.To name a function with named arguments without calling it, use this syntax (loosely inspired by Objective-C):
Called as Named as foo()
/foo(123)
foo
foo(bar: 123)
foo{bar:}
foo(123, baz: 456, 789)
foo{:baz::}
It can then be called without labels or coerced to the appropriate
fn
type. -
...Error messages can get a bit confusing. I meant "appending to the function name" literally. Like any other Rust procedural macro, this one performs a purely syntactic transformation and then hands some output code to the compiler, without much ability to control compilation after that. The following input:
fn foo(_ bar: u32) {} foo(bar: 5);
gets transformed into:
// In "nightly" mode: fn foo{bar:}(bar: u32) {} // ^^^^^^^^^ this whole thing is stuffed into an Ident token so it's treated as // a name, even though it's not a valid ident in vanilla Rust) // ditto for the call: foo{bar:}(bar: 5); // In "macros 1.1" mode, we have to use valid code: fn foo_namedarg_bar(bar: u32) {} foo_namedarg_bar(5);
The nightly version makes for somewhat better-looking errors, but either way, if you accidentally call
foo(5)
(without argument names), the error message will just sayunresolved name `foo`
. It doesn't know that you just messed up the labels. (So don't forget!)Note: It would be nice if I could improve this with a lint somehow, but early lints are too early and late lints aren't run if compilation fails.
Patterns
In Rust, unlike Swift, arguments don't actually have 'names'; they have
'patterns' just like let
or match
. So this is valid, standard Rust:
fn foo(Config { a, b }: Config) {}
fn foo(TupleStruct(a, b): TupleStruct) {}
fn foo((a, b): (i32, i32)) {}
With this macro, patterns can be used with named arguments too... sometimes:
fn foo(config Config { a, b }: Config) {} // OK
fn foo(ts TupleStruct(a, b): TupleStruct) {} // OK
fn foo(pair (a, b): (i32, i32)) {} // No good...
Unfortunately, the last one creates a syntactic ambiguity: it's intended to be
a parameter named pair
that takes a tuple, but it's already valid Rust syntax
for an unnamed parameter accepting a tuple-like struct called pair
:
struct pair(i32, i32); // Lowercasing a struct name is bad style, but possible!
fn foo(pair(a, b): pair) {} // Only the space distinguishes this from the other example.
(Well, and the type, but using the type to differentiate wouldn't work because of type aliases.)
Currently, the only workaround is to move destructuring into the function body:
fn foo(pair pair: (i32, i32)) { let (a, b) = pair; ... }
This works well enough, but is inelegant. A future version of this plugin may change the syntax to remove the ambiguity. (Any official named argument feature for Rust should probably do the same...)
Interaction with type ascription
In theory, named argument call syntax can conflict with type ascription (which
is not stable). Does the expression foo(a: b)
contain a named argument a
with value b
, or an unnamed argument whose value is the variable a
coerced
to the type b
?
In this plugin, named argument syntax is only triggered when the start of a
function argument consists of an identifier followed by a colon. So type
ascription is unaffected when its expression is anything other than a plain
variable: foo(5: usize)
or foo(x.collect(): Vec<bool>
works fine. And if
the expression is a variable, you can add parentheses: foo((a: usize))
.
But if you think about it, why would you want to type-ascribe a variable in the first place? In almost all cases, it would be better to give the variable definition a type annotation rather than using type ascription. Exceptions might include variables defined in complex patterns or match statements, which cannot be given types, but even then there's usually a better way to write it.
Usage
Add this to your Cargo.toml:
[dependencies]
namedarg = "0.1"
namedarg_rustc_macro = "0.1"
And wrap your source file in a macro invocation:
#![feature(rustc_macro, custom_derive)]
#[macro_use] extern crate namedarg;
namedarg! {
// your code here
}
Yeah, it's a little ugly. Once macros 1.1 is stabilized, the first line can be omitted.
"Macros 1.1" mode is automatically enabled if a non-nightly rustc is
detected, although at time of writing this is useless because it is not
actually stable yet. It has the sliight downside of breaking line
numbering for everything inside namedarg!
, which is inherent in the design
of macros 1.1. Like, all errors will show up pointing to the line containing
namedarg!
. Not exactly usable for development, but it should be possible
(post stabilization) to develop a crate on nightly and have it still compile on
stable this way.
You can force macros 1.1 mode on using a Cargo feature:
[dependencies.namedarg]
version = "0.1"
features = ["force_macros11_mode"]
[dependencies.namedarg_rustc_macro]
version = "0.1"
features = ["force_macros11_mode"]
Default arguments
There's also a very barebones and even more experimental syntax for default arguments, i.e. arguments that can be omitted when calling for a default value. I, comex, like Swift's syntax for named arguments, but I'm not satisfied with how any languages I know of treat default arguments. The problem comes when writing functions that forward their arguments to other functions, typically the kind of small wrapper you might call a 'convenience function'. For example, imagine we have this Swift code:
struct Player {
func doSomething(useFlubber: Bool = false) {}
}
func doSomethingToPlayer(username: String, useFlubber: Bool = false) {
// Convenience function to lookup a Player by name and call doSomething
Player(username: username)!.doSomething(useFlubber: useFlubber)
}
It works, but we had to write the default value twice, a DRY violation. If we
later decided to change the default for doSomething
, we'd probably want to
make the same change to doSomethingToPlayer
- but if we forgot, the compiler
wouldn't complain! Normally, refactoring functions in static languages pretty
reliably produces type errors on spots we haven't fixed up yet, but having a
different default is never a type error.
One way to avoid this issue might be to use a wrapper "config" struct whose constructor is responsible for dealing with defaults. (This is also often proposed as an alternative to named arguments in general.) This is a good solution if there's a lot of arguments - after all, repeating the argument names is already a DRY violation by itself - but it feels pretty verbose in the common case that there's only one or two arguments to forward.
Instead, the version of default arguments I've implemented in this macro is based on Option. It's pretty simple: if you put #[default] before the argument name, it's passed by default as None. It looks like this:
impl Player {
fn do_something(&self, #[default] _ use_flubber: Option<bool>) {
let use_flubber = use_flubber.unwrap_or(false);
}
}
fn do_something_to_player(username: &str, #[default] _ use_flubber: Option<bool>) {
Player::new(username: username).unwrap().do_something(use_flubber: use_flubber)
}
do_something_to_player("someone"); // implicitly passes None
do_something_to_player("someone", Some(true)); // kinda verbose but...
Note: Unnamed arguments can be defaulted too.
There are two significant limitations to defaults as currently implemented.
The first is apparent in the above snippet: if you want to pass the value
rather than leaving it at the default, you have to explicitly write Some
,
which is more verbose.
The second is that you can't skip over default arguments and specify later keyword arguments:
fn foo(#[default] _ a: Option<bool>, #[default] _ b: Option<bool>) -> u32 { /* ... */ }
foo(b: Some(true)) // doesn't work!
foo(a: None, b: Some(true)) // you have to do this.
Before I explain the source of these limitations, I suppose I should show how the transformation works. The above definition gets expanded to a series of overloaded functions:
fn foo() -> u32 { foo(a: None, b: None) }
fn foo(_ a: Option<bool>) -> u32 { foo(a: a, b: None) }
fn foo(_ a: Option<bool>, _ b: Option<bool>) -> u32 { /* ... */ }
A series of stub functions gets generated, one per prefix of the argument list, which forward to the real function. There's no better way to do it under the function-renaming scheme (and for that matter I don't think there are any particularly good alternatives to function renaming, though there are some theoretical possibilities involving types that are far more complex).
Thus the second limitation: If skipping over default arguments were supported, a stub would have to be generated for every subset of the argument list rather than every prefix, so adding an argument would double the number of stubs; this can get out of hand rather quickly. While I don't recommend having functions with high numbers of default arguments in the first place (as I said at some point it's better to use a struct), I don't like the idea of an implementation that outright fails on such functions.
The first limitation, having to write Some
, would be easier to overcome.
First of all, we could imagine having the generated overloads not take
Option
, something like:
fn foo() -> u32 { foo(a: None, b: None) }
fn foo(_ a: bool) -> u32 { foo(a: Some(a), b: None) }
fn foo(_ a: bool, _ b: bool) -> u32 { foo_IMPL(a: Some(a), b: Some(b)) }
fn foo_IMPL(_ a: bool, _ b: bool) -> u32 { /* ... */ }
But then how would we pass an Option when forwarding?
fn x_foo(#[default] _ a: Option<bool>) {
foo(a: a) // type error: expected bool, found Option<bool>
}
For that matter, how would we pass None when (explicitly) skipping over an argument?
There would have to be some special syntax to mean "pass through this Option argument untouched", though it would be tricky even then (requiring either generating even more stubs or changing the named-argument transformation).
Another possibility would be relying on the type system to disambiguate Option and non-Option arguments:
trait OptionOrJust<T> { ... }
impl<T> OptionOrJust<T> for T { ... }
impl<T> OptionOrJust<T> for Option<T> { ... }
fn foo<A: OptionOrJust<bool>>(_ a: A)
// both work!
foo(a: true);
foo(a: Some(true));
It would even work correctly if the argument was itself Option-typed... usually. But sometimes it would just get confusing:
// original:
fn foo<T>(#[default] _ x: T) { ... }
// transformed by the plugin:
fn foo<T, X: OptionOrJust<T>>(_ x: X) { ... }
// but wait...
foo(x: Some(123)) // what is T??
foo::<u32>(x: Some(123)) // error: too few type parameters provided
foo::<u32, _>(x: Some(123)) // OK, but what is this phantom type parameter?
...And even in nicer cases, it might make type inference a bit worse.
Other limitations
-
macro_rules!
insidenamedarg!
doesn't work properly because of a rustc bug. Trying to work around this by, say, puttingnamedarg!
inside the macro expansion doesn't work very well - for one thing, since the implementation is based on custom derives, the inside ofnamedarg!
must be a full set of items (things likefn
s, struct declarations, etc.), not, say, an expression. Also, becausenamedarg!
's expansion containsextern crate
, it only works if specified once, at the crate root. You can work around that, at least, by usingnamedarg_inner!
, which is the same asnamedarg!
but without theextern crate
: so there must be anamedarg!
in the same file for it to work. But the full-item limitation remains. I should add some kind of escaping mechanism to better work around the rustc bug, allowing any macros to just be defined insidenamedarg!
. -
The nightly implementation is a ludicrous hack. It used to just be a regular nightly compiler plugin, but that presents a worse user experience: it's not possible for the expansion of a
macro_rules!
macro to import a compiler plugin, so the user code would have to contain the logic for deciding whether to use the nightly or macros-1.1 implementation, which is quite a lot of clutter on top of what is already required. A sane person would submit a PR to rustc to improve the interaction ofmacro_rules!
and plugins. I instead noted that expansions can load macros-1.1 custom derives, and wrote some highly unsafe code to make an unstable plugin masquerade as a macros-1.1 plugin.