feat(linter): implement class sorting rule (first pass)

[DaniGuardiola]

Related: #1274

Summary

A class sorting rule for Tailwind CSS classes and other tools involving utility classes / atomic classes (initial version).

Test Plan

Unit tests and integration tests.

How it works

The pipeline is this:

• from options (preset + base values + extended values, currently hardcoded to tailwind preset), create the "resolved" sort config. ref: https://github.com/DaniGuardiola/biome/blob/e58fdb7855eba0a716261db3b05ac84e8320860d/crates/biome_js_analyze/src/semantic_analyzers/nursery/use_sorted_classes/sort_config.rs#L1-L6
• a classname string get split and each individual class lexed to get the "tailwind DSL info" (variants, utility, whether they are arbitrary). ref: https://github.com/DaniGuardiola/biome/blob/e58fdb7855eba0a716261db3b05ac84e8320860d/crates/biome_js_analyze/src/semantic_analyzers/nursery/use_sorted_classes/class_lexer.rs#L1-L9
• then, each of these classes are processed to compute some things ("class info" entities), which directly depend on the sort config (e.g. need to know the index, layer, and compute the "variant weight" of each. ref: https://github.com/DaniGuardiola/biome/blob/e58fdb7855eba0a716261db3b05ac84e8320860d/crates/biome_js_analyze/src/semantic_analyzers/nursery/use_sorted_classes/class_info.rs#L1-L11
• finally, these "class info" entities are sorted. ref: https://github.com/DaniGuardiola/biome/blob/e58fdb7855eba0a716261db3b05ac84e8320860d/crates/biome_js_analyze/src/semantic_analyzers/nursery/use_sorted_classes/sort.rs#L55

Some examples for clarity:

"sort config"

ordered layers + index in each layer, e.g:

1. components layer
   index 1 - container
2. utilities layer
   index 1 - p-
   index 2 - px-
   etc

variants are just a flat list e.g.

index 1 - hover
index 2 - focus
etc

"classname"

hover:p-4 px-2 [mask:none]

"classes"

hover:p-4 and px-2 and [mask:none]

"lexed/tokenized classes"

hover:p-4: variant hover and utility p-4
px-2: utility px-2
[mask:none]: utility [mask:none] (arbitrary!)

"class info"

hover:p-4: variant weight 1, utility layer "utilities", utility index "1", text "hover:p-4"
etc

---

To briefly explain this variant weight thing, each variant has a "bit" in a big bit vector, e.g. in my example sort config:

hover: 01
focus: 10

When computed the "total weight", it's just a bitwise "or" between all weights, e.g.

hover:px-2: 01
hover:focus:px-2: 11
focus:px-2: 10
px-2: 00

Which means that the order (given that the utility is the same) will be:

px-2 hover:px-2 focus:px-2 hover:focus:px-2

Why

Many people use utility classes / atomic CSS. In a nutshell, it means using classes that do one thing (e.g. .m-4 sets a margin). There are a few ways in which these are used:

• Through Tailwind CSS. By far the most popular option.
• Through other tools like Uno CSS or Nativewind. Less popular than Tailwind CSS.
• Custom-made: just some manually maintained CSS. I suspect this is very uncommon these days vs. tools.

When authoring in this way, the HTML/JSX code contains elements/components with potentially many classes, e.g.:

<div class="mb-4 top-2 pt-5 p-2 container some-custom-class flex relative" />;

This can get a bit messy, and that's why https://github.com/heybourn/headwind, a tool to sort them in a deterministic way, was created. Then, the Tailwind CSS folks came out with their version in the form of a Prettier plugin, which is widely used these days: https://github.com/tailwindlabs/prettier-plugin-tailwindcss

Many developers rely on it, and many have manifested that they are waiting for Biome to add similar functionality before switching from Prettier. This PR aims to do just that.

Utility classes / atomic CSS

For context: I am a very experienced/advanced Tailwind CSS user, having written a full version of it for CLIs (https://github.com/DaniGuardiola/classy-ink), a rewrite of tailwind-merge (https://twitter.com/daniguardio_la/status/1740184901775970703), and many (unreleased/proprietary) custom plugins/tooling based on it.

I won't go in-depth here as I think utility classes are best explained in the Tailwind CSS docs or any article out there. In this section, I'm going to focus on everything relevant to class sorting.

There are two main concepts: utilities and variants.

• Utilities: what you use to specify CSS properties, e.g. .relative for position: relative.
• Variants: "conditionals", e.g. hover:bg-red-500 results in a red background when the element is hovered. Variants can be stacked, e.g. md:focused:w-full results in width: 100% when the element is focused and when the screen width exceeds the value for the "medium" breakpoint.

Arbitrary values (top-[13px]), arbitrary CSS ([mask-type:luminance]) and arbitrary variants ([&:nth-child(3)]:underline) are also possible.

The three projects mentioned follow this structure. Nativewind works almost exactly like Tailwind CSS (in fact, it uses it under the hood). Uno CSS is more of a "make your own Tailwind" project, with a lot of presets, including a tailwind compat one.

How the Talwind sorting algorithm works

Utilities

The order is pretty straightforward and depends on a specific index per utility. For example, utility a might have an index of 1, and utility b might have an index of 2. So, something like a-24 will go before b-something.

Utilities can have one or both of these traits:

• Values: a utility that supports -suffix. E.g. a-12, b-something, etc.
• Default: a utility that supports not specifying any value, e.g. c.

Real-world examples:

• Only supports values: p-2, mt-6.
• Only supports default: block, sticky, underline.
• Supports values and default: grow, grow-2, border, border-4.

If the utility has both traits then the "default" goes first when sorted, e.g. a a-12.

These three pieces of information (index, has default, has values) are easy to obtain from the tailwind config file by using some tools exported by the tailwind package (already used by other packages like the prettier plugin itself). I already have a working script that does this, and turn it into a more compact format that I will discuss below.

Variants

These are a bit trickier, but after a long night, I was able to work it out and replicate it from scratch 😅

Variants are never alone, as they are "modifiers" for utilities, so they always come with a utility, e.g. hover:bg-red-500. Furthermore, they can be stacked, e.g. md:hover:bg-green-500.

Tailwind assigns a big number (JS bigint type) to each variant (it calls this number variants, I'm gonna call it "weight" because it's more descriptive for this explanation). This "weight" number is crucial for sorting.

This is how the comparison function of the sorting algorithm works in respect to classes:

1. For each of both classes, collect the "weights" of all variants. Derive a "combined" weight by executing a bitwise XOR operation between all of them (e.g. a | b | c where a, b and c are the weights that correspond to each variant).
2. Compare the resulting weights. If they are different, smaller goes first.
3. If weights are the same, compare indexes of the attached utilities. If they are different, the order is the same as described in the previous section.
4. If the indexes are the same, compare lexicographically (the whole class string, including all variants and the utility).

For this algorithm, the necessary information is this variant "weight", which can also be obtained from the tailwind config file, in the same way as described above.

Advanced details

Note that I'm omitting some information to make things easier to understand at first glance. Some details might not be worth talking about, but here's a list with other considerations:

• Utilities are split into "utilities" and "components". You can think of these as "categories". Components go first (the default tailwind config only ships with one: .container - all other classes are utilities).
• Variant sorting has more nuance. For example, some variants specify a custom sorting function (this is used, for example, to sort the screen variants, like sm, md, lg, etc). There are other details of the sorting algorithm that might or might not be worth porting over: layers, parent layers, "parallel indexes" (whatever that might be)...
• Arbitrary values, CSS and variants are not taken into account here.

WIP

You've reached the end, for now. I will expand on this soon, with details about this implementation, ideas, and more.

Check again soon :)

[netlify]

✅ Deploy Preview for biomejs canceled.

Name	Link
🔨 Latest commit	a1b156d
🔍 Latest deploy log	https://app.netlify.com/sites/biomejs/deploys/65b0f38817a9ea0008198a17

[ematipico]

Thank you @DaniGuardiola for kicking off this!

Here's some preliminary feedback, please take it with a grain of salt as I have little knowledge of the tool tailwind, and it's still unclear how you want to implement this feature gradually:

• I'd remove all the options for now, it make the code unclear and we can do it later
• I would start moving functions and logic in a separate file or multiple files. While doing that, I would appreciate if we start adding tests. Not just integration tests, but unit tests too. In Rust is very easy to create unit tests.

   // some_file.rs
   fn class_parser(&str) -> Result<SomeData, String> {
   
   }
   
   #[cfg(test)]
   mod tests {
   	use super::class_parser;
   	#t[test]
   	fn should_parse_something() {
   		let result = class_parser("");
   		expect(result.is_ok());
   	}
   }

  So please do that so we can also see how the logic should work
• Document stuff. This is really important in Biome. We want to make sure that the inner logic of our code is documented and understandable by any contributor. We believe this helps maintainability and PR reviews
• There's a lot of string allocations, but it's fine for now. We can deal with that later once we progress with the developments
• I understood that there's a lot of work to do, and I believe we can gradually add features in separate PRs. So, I think it's best to make a plan for this PR, review it, merge and continue with more work. You can also create new draft PRs based on this one so you feel stuck in case our reviews are slow to come

What do you think? Again, great work and thank you!

[DaniGuardiola]

Thanks for the review @ematipico!

I agree that it might be a good idea to groom and consolidate the current code before adding more features. I'm gonna focus on that so we can merge this and keep working in separate PRs. I would only highlight that the sorting algorithm is very complex, and I have iterated a lot, so the code is a bit frankensteiny at the moment with a mixture of obsolete POC stuff and "big-picture" decisions. My goal is going to be to get this in a good state and in line with the "big-picture" approach consistently (without adding features or unnecessary complexity). Sound good?

Responding to your preliminary feedback:

I'd remove all the options for now, it make the code unclear and we can do it later

I think removing most options is fine, and I can just check this branch's history to re-introduce them later (a decent amount of work went into them, so I'd like to re-use them in the future). I would like to preserve two options though, because they are super-simple and central to the rule: attributes and functions.

For the rest, we can just use the tailwind preset directly for now.

I would start moving functions and logic in a separate file or multiple files. While doing that, I would appreciate if we start adding tests.

Yep, I'd love to do that. Still new to Rust and the module system has me a bit confused, so I'd appreciate any guidance on this. Re:tests, of course, I was just focusing on getting things right first but I think some things can be tested now, like the class parser.

Document stuff.

100%. As with tests, I was waiting for things to stabilize, and for the big picture to emerge.

There's a lot of string allocations

Yes, this was jarring to me too. I even asked about it in the Discord. As a Rust (and Biome) beginner, I was a bit helpless here, partly because the presets are supposed to be in the same format as options, but string slices can't be deserialized apparently. Any help would be super appreciated, I wanna make this rule as fast as possible!

I think it's best to make a plan for this PR, review it, merge, and continue with more work.

Agreed! Let's do that. I'm gonna focus on that: address your comments, groom it, get it in a good state, and merge. Then I can keep working on features. In parallel, I'm thinking about writing a deep dive in my post about how Tailwind sorting works, so that there's something I can reference. Thanks again!

PD: after I go offline today, I'll be afk for a couple of days. Be back when the new years hangover is over :P

[DaniGuardiola]

@ematipico I did the following:

• Replaced the outdated "PoC" bits with code that is in line with the whole thing, leaving room for the remaining features to be implemented.
• Cleaned up a whole lot.
• Documented some things.
• Taken action on your review comments, responded to the ones I couldn't figure out, or where I added some necessary context.
• Removed the chunky part of the options (sort config), leaving only the very simple visitor options.

Before merging, I still wanna do a full documentation pass and address some of the TODOs.

When you have a moment, please re-review :) happy new year!

[ematipico]

Hey @DaniGuardiola

Thank you! I pushed some code in the branch. Let's:

• do one more pass of documentation if this is still needed
• add tests. We can't merge the rule without integration tests

[Conaclos]

I made a quick pass and left several comments :)

[DaniGuardiola]

Thank you very much, @Conaclos! I responded to all of your comments and will take action on them soon.

[Conaclos]

@DaniGuardiola Fell free to resolve the reviews once addressed.

[stabildev]

How do I get this to run on save?

I tried these settings without any luck

"linter": {
    "rules": {
      "nursery": {
        "useSortedClasses": "error"

  "editor.codeActionsOnSave": {
    "quickfix.biome": "explicit",
    "source.fixAll": "explicit"
  },

[ematipico]

You can't. The code action is unsafe. You can only opt-in via editor. Or via --apply-unsafe

[stabildev]

For anyone interested, I found this workaround (automatic class sorting is a must!):

Install the extension Run on Save by Emeraldwalk

Add this to your config:

"emeraldwalk.runonsave": {
    "commands": [
      {
        "match": "\\.(ts|tsx|js|jsx|html)$",
        "cmd": "bunx @biomejs/biome lint ${file} --apply-unsafe"
      }
    ]
  }

Thank you for building this!

[Duckinm]

For anyone interested, I found this workaround (automatic class sorting is a must!):
Install the extension Run on Save by Emeraldwalk
Add this to your config:

"emeraldwalk.runonsave": {
    "commands": [
      {
        "match": "\\.(ts|tsx|js|jsx|html)$",
        "cmd": "bunx @biomejs/biome lint ${file} --apply-unsafe"
      }
    ]
  }

Thank you for building this!

I don't think I have this working, In my case I'm using pnpm

"emeraldwalk.runonsave": {
		"commands": [
			{
				"match": "\\.(ts|tsx|js|jsx|html)$",
				"cmd": "pnpm exec @biomejs/biome lint ${file} --apply-unsafe"
			}
		]
	}

Ah just using pnpm biome lint directly, Thank you!

"cmd": "pnpm biome lint ${file} --apply-unsafe"

[vasyaqwe]

neither "cmd": "pnpm biome lint ${file} --apply-unsafe" nor "cmd": "pnpm biome lint ${file} --write --unsafe" was working for me, I got it to work by enclosing the file in quotes, like so:

"emeraldwalk.runonsave": {
      "commands": [
         {
            "match": "\\.(ts|tsx|js|jsx|html)$",
            "cmd": "pnpm biome lint \"${file}\" --write --unsafe"
         }
      ]
   }

hope this helps someone :)

[fzkhan19]

If you want it to work on explicit save you can go to your vscode's keybindings.json and add this:

 {
     "key": "ctrl+s",
     "command": "workbench.action.terminal.sendSequence",
     "args": { "text": "bunx @biomejs/biome lint ${file} --apply-unsafe\u000D" }
 }

[devcaeg]

When do you think this option will cease to be “unsafe”?

[luckydye]

@devcaeg you can also configure it now:

"useSortedClasses": {
  ...
  "fix": "safe"
}

[devcaeg]

@luckydye

The problem with this is that in the autosave the “useSortedClasses” fix is not applied.

[devcaeg]

I take back what was said before, with the following configuration in Visual Studio Code, the class is automatically ordered when saving the file:

"editor.codeActionsOnSave": {
	"quickfix.biome": "explicit"
}