Compile-time Code & Metaprogramming
Overview
Metaprogramming in Noir is comprised of three parts:
comptime
code- Quoting and unquoting
- The metaprogramming API in
std::meta
Each of these are explained in more detail in the next sections but the wide picture is that
comptime
allows us to write code which runs at compile-time. In this comptime
code we
can quote and unquote snippets of the program, manipulate them, and insert them in other
parts of the program. Comptime functions which do this are said to be macros. Additionally,
there's a compile-time API of built-in types and functions provided by the compiler which allows
for greater analysis and modification of programs.
Comptime
comptime
is a new keyword in Noir which marks an item as executing or existing at compile-time. It can be used in several ways:
comptime fn
to define functions which execute exclusively during compile-time.comptime global
to define a global variable which is evaluated at compile-time.- Unlike runtime globals,
comptime global
s can be mutable.
- Unlike runtime globals,
comptime { ... }
to execute a block of statements during compile-time.comptime let
to define a variable whose value is evaluated at compile-time.comptime for
to run a for loop at compile-time. Syntax sugar forcomptime { for .. }
.
Scoping
Note that while in a comptime
context, any runtime variables local to the current function are never visible.
Evaluating
Evaluation rules of comptime
follows the normal unconstrained evaluation rules for other Noir code. There are a few things to note though:
- Certain built-in functions may not be available, although more may be added over time.
- Evaluation order of
comptime {}
blocks within global items is currently unspecified. For example, given the following two functions we can't guarantee whichprintln
will execute first. The ordering of the two printouts will be arbitrary, but should be stable across multiple compilations with the samenargo
version as long as the program is also unchanged.
fn one() {
comptime { println("one"); }
}
fn two() {
comptime { println("two"); }
}
- Since evaluation order is unspecified, care should be taken when using mutable globals so that they do not rely on a particular ordering. For example, using globals to generate unique ids should be fine but relying on certain ids always being produced (especially after edits to the program) should be avoided.
- Although the ordering of comptime code is usually unspecified, there are cases where it is:
- Dependencies of a crate will always be evaluated before the dependent crate.
- Any attributes on a function will be run before the function body is resolved. This is to allow the attribute to modify the function if necessary. Note that if the
function itself was called at compile-time previously, it will already be resolved and cannot be modified. To prevent accidentally calling functions you wish to modify
at compile-time, it may be helpful to sort your
comptime
annotation functions into a different submodule crate along with any dependencies they require. - Unlike raw
comptime {}
blocks, attributes on top-level items in the program do have a set evaluation order. Attributes within a module are evaluated top-down, and attributes in different modules are evaluated submodule-first. Sibling modules to the same parent module are evaluated in order of the module declarations (mod foo; mod bar;
) in their parent module.
Lowering
When a comptime
value is used in runtime code it must be lowered into a runtime value. This means replacing the expression with the literal that it evaluated to. For example, the code:
struct Foo { array: [Field; 2], len: u32 }
fn main() {
println(comptime {
let mut foo = std::mem::zeroed::<Foo>();
foo.array[0] = 4;
foo.len = 1;
foo
});
}
will be converted to the following after comptime
expressions are evaluated:
struct Foo { array: [Field; 2], len: u32 }
fn main() {
println(Foo { array: [4, 0], len: 1 });
}
Not all types of values can be lowered. For example, references, Type
s, and TypeDefinition
s (among other types) cannot be lowered at all.
fn main() {
// There's nothing we could inline here to create a Type value at runtime
// let _ = get_type!();
}
comptime fn get_type() -> Type { ... }
Values of certain types may also change type when they are lowered. For example, a comptime format string will already be formatted, and thus lowers into a runtime string instead:
fn main() {
let foo = comptime {
let i = 2;
f"i = {i}"
};
assert_eq(foo, "i = 2");
}
(Quasi) Quote
Macros in Noir are comptime
functions which return code as a value which is inserted into the call site when it is lowered there.
A code value in this case is of type Quoted
and can be created by a quote { ... }
expression.
More specifically, the code value quote
creates is a token stream - a representation of source code as a series of words, numbers, string literals, or operators.
For example, the expression quote { Hi "there reader"! }
would quote three tokens: the word "hi", the string "there reader", and an exclamation mark.
You'll note that snippets that would otherwise be invalid syntax can still be quoted.
When a Quoted
value is used in runtime code, it is lowered into a quote { ... }
expression. Since this expression is only valid
in compile-time code however, we'd get an error if we tried this. Instead, we can use macro insertion to insert each token into the
program at that point, and parse it as an expression. To do this, we have to add a !
after the function name returning the Quoted
value.
If the value was created locally and there is no function returning it, std::meta::unquote!(_)
can be used instead.
Calling such a function at compile-time without !
will just return the Quoted
value to be further manipulated. For example:
comptime fn quote_one() -> Quoted {
quote { 1 }
}
#[test]
fn returning_versus_macro_insertion() {
comptime {
// let _a: Quoted = quote { 1 };
let _a: Quoted = quote_one();
// let _b: Field = 1;
let _b: Field = quote_one!();
// Since integers default to fields, if we
// want a different type we have to explicitly cast
// let _c: i32 = 1 as i32;
let _c: i32 = quote_one!() as i32;
}
}
For those familiar with quoting from other languages (primarily lisps), Noir's quote
is actually a quasiquote.
This means we can escape the quoting by using the unquote operator to splice values in the middle of quoted code.
In addition to curly braces, you can also use square braces for the quote operator:
comptime {
let q1 = quote { 1 };
let q2 = quote [ 2 ];
assert_eq(q1, q2);
// Square braces can be used to quote mismatched curly braces if needed
let _ = quote[}];
}
Unquote
The unquote operator $
is usable within a quote
expression.
It takes a variable as an argument, evaluates the variable, and splices the resulting value into the quoted token stream at that point. For example,
comptime {
let x = 1 + 2;
let y = quote { $x + 4 };
}
The value of y
above will be the token stream containing 3
, +
, and 4
. We can also use this to combine Quoted
values into larger token streams:
comptime {
let x = quote { 1 + 2 };
let y = quote { $x + 4 };
}
The value of y
above is now the token stream containing five tokens: 1 + 2 + 4
.
Note that to unquote something, a variable name must follow the $
operator in a token stream.
If it is an expression (even a parenthesized one), it will do nothing. Most likely a parse error will be given when the macro is later unquoted.
Unquoting can also be avoided by escaping the $
with a backslash:
comptime {
let x = quote { 1 + 2 };
// y contains the four tokens: `$x + 4`
let y = quote { \$x + 4 };
}
Combining Tokens
Note that Quoted
is internally a series of separate tokens, and that all unquoting does is combine these token vectors.
This means that code which appears to append like a string actually appends like a vector internally:
comptime {
let x = 3;
let q = quote { foo$x }; // This is [foo, 3], not [foo3]
// Spaces are ignored in general, they're never part of a token
assert_eq(q, quote { foo 3 });
}
If you do want string semantics, you can use format strings then convert back to a Quoted
value with .quoted_contents()
.
Note that formatting a quoted value with multiple tokens will always insert a space between each token. If this is
undesired, you'll need to only operate on quoted values containing a single token. To do this, you can iterate
over each token of a larger quoted value with .tokens()
:
comptime fn concatenate(q1: Quoted, q2: Quoted) -> Quoted {
assert(q1.tokens().len() <= 1);
assert(q2.tokens().len() <= 1);
f"{q1}{q2}".quoted_contents()
}
// The CtString type is also useful for a compile-time string of unbounded size
// so that you can append to it in a loop.
comptime fn double_spaced(q: Quoted) -> CtString {
let mut result = "".as_ctstring();
for token in q.tokens() {
if result != "".as_ctstring() {
result = result.append_str(" ");
}
result = result.append_fmtstr(f"{token}");
}
result
}
#[test]
fn concatenate_test() {
comptime {
let result = concatenate(quote {foo}, quote {bar});
assert_eq(result, quote {foobar});
let result = double_spaced(quote {foo bar 3}).as_quoted_str!();
assert_eq(result, "foo bar 3");
}
}
Attributes
Attributes provide a way to run a comptime
function on an item in the program.
When you use an attribute, the function with the same name will be called with that item as an argument:
#[my_struct_attribute]
struct Foo {}
comptime fn my_struct_attribute(s: StructDefinition) {
println("Called my_struct_attribute!");
}
#[my_function_attribute]
fn foo() {}
comptime fn my_function_attribute(f: FunctionDefinition) {
println("Called my_function_attribute!");
}
Anything returned from one of these functions will be inserted at top-level along with the original item.
Note that expressions are not valid at top-level so you'll get an error trying to return 3
or similar just as if you tried to write a program containing 3; struct Foo {}
.
You can insert other top-level items such as trait impls, structs, or functions this way though.
For example, this is the mechanism used to insert additional trait implementations into the program when deriving a trait impl from a struct:
trait FieldCount {
fn field_count() -> u32;
}
#[derive_field_count]
struct Bar {
x: Field,
y: [Field; 2],
}
comptime fn derive_field_count(s: StructDefinition) -> Quoted {
let typ = s.as_type();
let field_count = s.fields().len();
quote {
impl FieldCount for $typ {
fn field_count() -> u32 {
$field_count
}
}
}
}
Calling annotations with additional arguments
Arguments may optionally be given to attributes. When this is done, these additional arguments are passed to the attribute function after the item argument.
#[assert_field_is_type(quote { i32 }.as_type())]
struct MyStruct {
my_field: i32,
}
comptime fn assert_field_is_type(s: StructDefinition, typ: Type) {
// Assert the first field in `s` has type `typ`
let fields = s.fields();
assert_eq(fields[0].1, typ);
}
We can also take any number of arguments by adding the varargs
attribute:
#[assert_three_args(1, 2, 3)]
struct MyOtherStruct {
my_other_field: u32,
}
#[varargs]
comptime fn assert_three_args(_s: StructDefinition, args: [Field]) {
assert_eq(args.len(), 3);
}
Attribute Evaluation Order
Unlike the evaluation order of stray comptime {}
blocks within functions, attributes have a well-defined evaluation
order. Within a module, attributes are evaluated top to bottom. Between modules, attributes in child modules are evaluated
first. Attributes in sibling modules are resolved following the mod foo; mod bar;
declaration order within their parent
modules.
mod foo; // attributes in foo are run first
mod bar; // followed by attributes in bar
// followed by any attributes in the parent module
#[derive(Eq)]
struct Baz {}
Note that because of this evaluation order, you may get an error trying to derive a trait for a struct whose fields have not yet had the trait derived already:
// Error! `Bar` field of `Foo` does not (yet) implement Eq!
#[derive(Eq)]
struct Foo {
bar: Bar
}
#[derive(Eq)]
struct Bar {}
In this case, the issue can be resolved by rearranging the structs.
Comptime API
Although comptime
, quote
, and unquoting provide a flexible base for writing macros,
Noir's true metaprogramming ability comes from being able to interact with the compiler through a compile-time API.
This API can be accessed through built-in functions in std::meta
as well as on methods of several comptime
types.
The following is an incomplete list of some comptime
types along with some useful methods on them. You can see more in the standard library Metaprogramming section.
Quoted
: A token streamType
: The type of a Noir typefn implements(self, constraint: TraitConstraint) -> bool
- Returns true if
self
implements the given trait constraint
- Returns true if
Expr
: A syntactically valid expression. Can be used to recur on a program's parse tree to inspect how it is structured.- Methods:
fn as_function_call(self) -> Option<(Expr, [Expr])>
- If this is a function call expression, return
(function, arguments)
- If this is a function call expression, return
fn as_block(self) -> Option<[Expr]>
- If this is a block, return each statement in the block
- Methods:
FunctionDefinition
: A function definition- Methods:
fn parameters(self) -> [(Quoted, Type)]
- Returns a slice of
(name, type)
pairs for each parameter
- Returns a slice of
- Methods:
StructDefinition
: A struct definition- Methods:
fn as_type(self) -> Type
- Returns this
StructDefinition
as aType
. Any generics are kept as-is
- Returns this
fn generics(self) -> [Quoted]
- Return the name of each generic on this struct
fn fields(self) -> [(Quoted, Type)]
- Return the name and type of each field
- Methods:
TraitConstraint
: A trait constraint such asFrom<Field>
TypedExpr
: A type-checked expression.UnresolvedType
: A syntactic notation that refers to a Noir type that hasn't been resolved yet
There are many more functions available by exploring the std::meta
module and its submodules.
Using these methods is the key to writing powerful metaprogramming libraries.
#[use_callers_scope]
Since certain functions such as Quoted::as_type
, Expression::as_type
, or Quoted::as_trait_constraint
will attempt
to resolve their contents in a particular scope - it can be useful to change the scope they resolve in. By default
these functions will resolve in the current function's scope which is usually the attribute function they are called in.
If you're working on a library however, this may be a completely different module or crate to the item you're trying to
use the attribute on. If you want to be able to use Quoted::as_type
to refer to types local to the caller's scope for
example, you can annotate your attribute function with #[use_callers_scope]
. This will ensure your attribute, and any
closures it uses, can refer to anything in the caller's scope. #[use_callers_scope]
also works recursively. So if both
your attribute function and a helper function it calls use it, then they can both refer to the same original caller.
Example: Derive
Using all of the above, we can write a derive
macro that behaves similarly to Rust's but is not built into the language.
From the user's perspective it will look like this:
// Example usage
#[derive(Default, Eq, Ord)]
struct MyStruct { my_field: u32 }
To implement derive
we'll have to create a comptime
function that accepts
a variable amount of traits.
// These are needed for the unconstrained hashmap we're using to store derive functions
use crate::collections::umap::UHashMap;
use crate::hash::BuildHasherDefault;
use crate::hash::poseidon2::Poseidon2Hasher;
// A derive function is one that given a struct definition can
// create us a quoted trait impl from it.
pub type DeriveFunction = fn(StructDefinition) -> Quoted;
// We'll keep a global HANDLERS map to keep track of the derive handler for each trait
comptime mut global HANDLERS: UHashMap<TraitDefinition, DeriveFunction, BuildHasherDefault<Poseidon2Hasher>> =
UHashMap::default();
// Given a struct and a slice of traits to derive, create trait impls for each.
// This function is as simple as iterating over the slice, checking if we have a trait
// handler registered for the given trait, calling it, and appending the result.
#[varargs]
pub comptime fn derive(s: StructDefinition, traits: [TraitDefinition]) -> Quoted {
let mut result = quote {};
for trait_to_derive in traits {
let handler = HANDLERS.get(trait_to_derive);
assert(handler.is_some(), f"No derive function registered for `{trait_to_derive}`");
let trait_impl = handler.unwrap()(s);
result = quote { $result $trait_impl };
}
result
}
Registering a derive function could be done as follows:
// To register a handler for a trait, just add it to our handlers map
pub comptime fn derive_via(t: TraitDefinition, f: DeriveFunction) {
HANDLERS.insert(t, f);
}
// Finally, to register a handler we call the above function as an annotation
// with our handler function.
#[derive_via(derive_do_nothing)]
trait DoNothing {
fn do_nothing(self);
}
comptime fn derive_do_nothing(s: StructDefinition) -> Quoted {
// This is simplified since we don't handle generics or where clauses!
// In a real example we'd likely also need to introduce each of
// `s.generics()` as well as a trait constraint for each generic
// to ensure they also implement the trait.
let typ = s.as_type();
quote {
impl DoNothing for $typ {
fn do_nothing(self) {
// Traits can't tell us what to do
println("something");
}
}
}
}
// Since `DoNothing` is a simple trait which:
// 1. Only has one method
// 2. Does not have any generics on the trait itself
// We can use `std::meta::make_trait_impl` to help us out.
// This helper function will generate our impl for us along with any
// necessary where clauses and still provides a flexible interface
// for us to work on each field on the struct.
comptime fn derive_do_nothing_alt(s: StructDefinition) -> Quoted {
let trait_name = quote { DoNothing };
let method_signature = quote { fn do_nothing(self) };
// Call `do_nothing` recursively on each field in the struct
let for_each_field = |field_name| quote { self.$field_name.do_nothing(); };
// Some traits like Eq want to join each field expression with something like `&`.
// We don't need that here
let join_fields_with = quote {};
// The body function is a spot to insert any extra setup/teardown needed.
// We'll insert our println here. Since we recur on each field, we should see
// one println for the struct itself, followed by a println for every field (recursively).
let body = |body| quote {
println("something");
$body
};
crate::meta::make_trait_impl(
s,
trait_name,
method_signature,
for_each_field,
join_fields_with,
body,
)
}