Recent posts

Why is swc fast?

DongYoon Kang

swc is fast. Very fast. It's 18x faster than babel on a single-core benchmark, and on a parallel benchmark, it's 68x faster than babel on a 4 core (8 HT) machine. Why? Just because it's written in rust? No. Its fundamental design differs from any other tool.

The language#

Just rewriting something in another language does not make it faster.

This is partially correct. Rewriting something in a language with an optimizing compiler can make difference. Some compilers including rustc, which uses llvm as a backend, are good at optimizing codes.

Edit: makes -> can make, created an optimization hints section,

Optimization hints#

Sometimes, rustc cannot deduce that it's dead code even if it is dead. In swc, we give enough hint if it's the case.

For example, not all passes touch all of the ast. Many passes return the given ast node as-is. Especially, most passes have nothing to do with types. So those passes return the typings as-is. So, it would be cool if those handlers are optimized out. rustc failed to optimize it out by default because it changes semantic, but we can help it.

macro_rules! noop_fold_type {
($F:ty, $N:tt) => {
impl Fold<swc_ecma_ast::$N> for $F {
#[inline]
fn fold(&mut self, node: swc_ecma_ast::$N) -> swc_ecma_ast::$N {
node
}
}
};
($F:ty) => {
noop_fold_type!($F, Accessibility);
// ...
// Simillar code for each ast types
};
}

See it on github.

Now, it's optimized out. However, what matters is allocations. Fold<Vec<T>> is implemented as below.

impl<T, F> FoldWith<F> for Vec<T>
where
F: Fold<T>,
{
fn fold_children(self, f: &mut F) -> Self {
self.move_map(|it| f.fold(it))
}
}

move_map() is a hack to prevent reallocation. When Fold<T> is an identity function, Fold<Vec<T>> is cleary no-op. Let's see the generated assembly for the code below. The link contains the actual implementation of move_map.

pub fn ret(v: Vec<i32>) -> Vec<i32> {
v
}
pub fn move_map(v:Vec<i32>) -> Vec<i32> {
v.move_map(|v| v)
}
example::ret:
mov rax, rdi
mov rcx, qword ptr [rsi + 16]
mov qword ptr [rdi + 16], rcx
movups xmm0, xmmword ptr [rsi]
movups xmmword ptr [rdi], xmm0
ret
example::move_map:
sub rsp, 24
mov rax, rdi
movups xmm0, xmmword ptr [rsi]
movaps xmmword ptr [rsp], xmm0
mov rcx, qword ptr [rsi + 16]
movups xmmword ptr [rdi], xmm0
mov qword ptr [rdi + 16], rcx
add rsp, 24
ret

Great. Although it's a bit longer than no-op, no allocation occurs and thus it's fast enough. The same goes for the Box<T>. See https://godbolt.org/z/5fDXQK if you want.

Design#

String caching#

Identifiers are inherently used multiple times. swc utilizes string_cache from the servo project to cache strings. Common identifiers like Object is stored as a constant.

Less scope analysis#

Scope analysis is done only 2 ~ 4 times per file. One at the start of processing, one at the end of processing, one to strip type-only imports (typescript only), and the last one to transcompile it into the other type of modules. (cjs, amd, umd)

I'll explain how it works, and why it's fast. First, before applying any other transformations, resolver pass colors the identifiers.

Note that the symbol is not changed. ___ denotes the context number and it's not part of the symbol.

const foo = 1;
use(foo);
{
const foo = 2;
use(foo);
}

becomes

const foo___1 = 1;
use(foo___1);
{
const foo___2 = 2;
use(foo___2);
}

Then, other passes insert identifiers if required. There's a helper macro to create a private ident. Let's suppose that a pass appended a private identifier named foo to the top level.

It's now

const foo = 1;
use(foo);
{
const foo = 2;
use(foo);
}
const foo = 3;

and with expanded context number, it's

const foo___1 = 1;
use(foo___1);
{
const foo___2 = 2;
use(foo__2);
}
const foo___3 = 3;

The last pass, named hygiene, removes context numbers and changes symbol to the appropriate one. It becomes

const foo = 1;
use(foo);
{
const foo1 = 2;
use(foo1);
}
const foo2 = 3;

It means that passes between resolver and hygiene can inject identifiers as they want without any heavy operation.

For comparison, babel maintains the scope while transcompiling and uses it like scope.rename('require') to ensure that the name does not conflict. However, given the fact that babel does renaming directly instead of queueing, api like this means that all identifiers in the file are visited on each call. For common js module, babel does like

path.scope.rename("exports");
path.scope.rename("module");
path.scope.rename("require");
path.scope.rename("__filename");
path.scope.rename("__dirname");

and it results in 5 scope analysis.

The basic idea of hygiene is taken from the macro system of rustc. Note that no compiler works in this way and I call this approach identifier hygiene. (Please correct me if I'm wrong.)

Note that this will also help writing a fast bundler. While bundling modules, only a single invocation of the resolver pass is enough to distinguish identifiers from multiple modules. Hygiene pass is invoked only once at the end.

Edit: Added how babel works

No use of graph data structures#

swc does not use graph structures for some reason. The most important one is the fact that rust is not a language with a garbage collector. After parsing, we have ast node, which consists of some vectors.

To use a graph data structures like petgraph::DiGraph, we need to destruct a vector and create a graph of nodes. Also, as js code generator takes ast node, we have to reconstruct the nodes from graph structures. This is an obvious overhead, and I wanted to avoid it.

Some graph traversal are possible with vectors, so I decided to go without graph data structures.

Edit: swc uses dfs and bfs, so we can't say it does not do graph traversal

--