This PR introduces a new feature where we will be handling the CSS
nesting ourselves.
We currently still rely on Lightning CSS in most places. But there are
situations where we don't use Lightning CSS out of the box:
1. During development, typically optimization/minification isn't setup
2. In places where it isn't as easy to run Lightning CSS such as in
`@tailwindcss/browser` or in Tailwind Play.
We handle CSS nesting in a single pass over the AST by tracking some
information as we go. It's not the most complex code, but there are some
tricky parts to make this happen in an efficient way, especially for the
few additional optimizations we handle.
While going over the AST, we will only emit CSS the moment we see
declarations or comments. This also means that this has a fun side
effect of removing CSS that ends up with empty nodes automatically.
(Caveat: there are exceptions for body-less rules such as `@layer foo;`
or `@charset "UTF-8";)
This also allowed us to do some cleanup in `optimizeAst` that tried to
do this as well, but now this will be handled by the code that handles
nesting automatically. Which is preferred because the version in
`optimizeAst` mutated the AST.
This also contains some optimizations where we merge adjacent at-rules
(with the same name / params), and adjacent rules with the same
selector, and get rid of declarations that are duplicated in a node.
(Caveat: there are exceptions, in case of `@font-family { … }` where we
don't want to merge them)
~~To ensure that this implementation is correct, I also added an oracle
implementation in the tests. This implementation does multiple passes
over the AST, because it does each step one by one, with minimal code.
Each step contains comments with examples to see what's happening in
that step. We then test the optimized version against this.~~ Once the
implementation was in place, and all the tests were passing, then I
deleted the oracle implementation. That way we don't have to keep it in
sync all the time.
While handling the nesting, we have to make sure that `&` exists and if
we replace it with a parent selector that we do use `:is(…)` semantics.
This means that:
```css
.foo {
&:hover {
color: red;
}
}
```
Becomes:
```css
:is(.foo):hover {
color: red;
}
```
We then also make sure that we optimize the selector by removing the
unnecessary `:is(…)` wrappers, but only if they were introduced by the
nesting logic. If _you_ wrote `:is(…)` in your CSS, we won't touch it.
If you look at the commits, the first thing we did is remove the
optimization step from Lightning CSS in the tests. Then we enabled our
CSS nesting handling code. This allows us to see the effect of the
changes we are making. At the end, we re-enabled Lightning CSS.
For now, this PR will be a step that happens before Lightning CSS is
executed, while still using Lightning CSS. But now this step will also
always happen in places where we don't use Lightning CSS at all.
This should not result in any breaking changes. It could result in
changed CSS output in environments where Lightning CSS isn't used. In
environments where it is being used, then there could be some
differences related to some selectors but they should result in the same
behavior with the same specificity.
While testing things, I noticed that there are some missed opportunities
for performance related to how we extract variables from declaration
values. I want to tackle `optimizeAst` in future PRs to make it simpler,
more performant, and maybe even merge it with the CSS nesting handling.
As part of testing this, I tested it against the tailwindcss.com
codebase which contains a lot of CSS (807.67 KB, 18 174 AST nodes)
because almost every utility is being used in examples.
The oracle implementation is rather slow:
```
[131.59ms] ↳ oracle (step by step)
[129.23ms] ↳ handleNesting(…)
[ 2.32ms] ↳ toCss(…)
```
But the final code is much faster (`<15ms`):
```
[ 10.11ms] ↳ hand written (single pass)
[ 8.40ms] ↳ handleNesting(…)
[ 1.67ms] ↳ toCss(…)
```
In contrast, Lightning CSS takes: `[ 37.48ms] Optimized by Lightning
CSS`
One interesting thing to notice is that in big projects, this could add
`10ms` to the build, but Lightning CSS would then take less time to
process, which results in a no-op with better output.
One thing to keep in mind here is that Lightning CSS does more things,
such as normalizing values, handling vendor prefixes, CSS nesting, etc.
<details>
<summary>Some notes on how the algorithm works:</summary>
### The basic idea
When you have CSS that looks this:
```css
.foo {
.bar {
color: red;
}
}
```
Then the AST looks like this:
```
[
{
kind: 'rule',
selector: '.foo',
nodes: [
{
kind: 'rule',
selector: '.bar',
nodes: [
{
kind: 'declaration',
property: 'color',
value: 'red',
important: false
}
]
}
]
}
]
```
When we walk this tree, and we encounter a `rule`, then we will track
the selector on a stack. When we are done walking over the rule, then we
will pop the selector from the stack. This means that the top-most
selector on the stack will always be the parent selector.
```ts
let selectorStack = []
walk(ast, {
enter(node) {
selectorStack.push(node.selector)
},
exit(node) {
selectorStack.pop()
},
})
```
The moment we encounter a `rule`, and if a previous rule was seen, then
we push the `selector` of the rule onto the stack, but in a way that the
`&` is already replaced by the selector. This way, a sibling rule will
also get the already-prepared parent selector.
The simple version looks like this:
```ts
walk(ast, {
enter(node) {
// In the real code we properly handle `&` replacement, and make sure that
// parent selector is prepended if there is no `&` used in the selector of the
// node.
let selector =
selectorStack.length > 0
// At this point, we don't optimize anything related to the selector yet
? node.selector.replaceAll('&', `:is(${selectorStack.at(-1)})`)
: node.selector
selectorStack.push(selector)
},
exit(node) {
selectorStack.pop()
},
})
```
So far we aren't doing much yet, but the interesting part is when we
encounter a `declaration` (or a `comment`). The moment we see any of
those, then will we emit a node with the information from the
`selectorStack`.
We then also track the last node's `nodes` we created such that we can
push more declarations into it as a shortcut.
```ts
let result: AstNode[] = []
let nodes: AstNode[] | null = null
walk(ast, {
enter(node) {
if (node.kind === 'declaration') {
// `nodes` is available, nothing special to do
if (nodes) {
nodes.push(node)
return
}
// Track new nodes
let nodes = [node]
// Create a new node with a reference to `nodes` for future declarations
let newNode = rule(selectorStack.at(-1), nodes)
result.push(newNode)
}
},
})
```
The last important part is that whenever we see a new `rule`, then we
have to reset that `nodes` tracking variable such that we can create a
fresh node the next time we see a declaration.
For the `at-rules`, something similar happens but they are tracked in a
similar but separate stack. The idea there is that we can then wrap
those `at-rules` around the `newNode` we create. That way the at-rules
naturally float to the top.
I can keep going here, but I think if you're interested in this, then
you could go over the commits in this PR, or you can look at the
`ast.ts` implementation directly to see what's going on.
</details>
## Test plan
1. Existing tests should pass
2. New tests have been added to test the flattening of nested CSS
5835691d21