Contributing to Svelte - Implement {#key}

September 27, 2020
Series: Contributing to Svelte SvelteJavaScriptOpen Source

Background

Unlike the other contributing to Svelte posts [1] [2], which I wrote it while implementing the fix, describing as detailed as possible, today I am going to share the process of how I implemented the {#key} block retrospectively.

The implementation of the {#key} block is much simpler, relative to {#if}, {#await} or {#each}. And I believe the process of implementing the {#key} block helps paint the pratical side of "The Svelte Compiler Handbook" or my "Looking into the Svelte compiler" talk.

The motivation

The idea of {#key} block starts with the feature request 2 years ago (yea, it's that long) for the ability to key a non-each component, GitHub issue #1469.

To key a component, is to force recreation of the component when the key changes.

And you see this ability of destroying and creating new components when using {#each} with key:

<script>
let data = [{ id: 1, name: 'alice' }];
function update() {
data = [{ id: 2, name: 'bob' }];
}
</script>
{#each data as item (item.id)}
<div>{ item.name }</div>
{/each}

REPL

When we call the function update, we removed alice from the data and we added bob. The net effect is still having a list of 1 item. However, instead of reusing the 1 <div /> by updating { item.name } to "bob", Svelte removes and destroys the <div /> and create a new <div /> for bob. This is because of the key we specified to the {#each} block. Svelte will not reuse the <div /> because it was created with a different key.

One of the benefits of having a key for {#each} item is to be able to add transition to the item correctly. Without a key to identify which item is added / removed, the transiion on a {#each} list will always applied to the last item, when the list grows or shrinks in length.

Try with and without the key in this REPL to see the importance of having a key.

This is similar to the key attribute of React, if you are familiar with React. Check this out on how to remount a component with the key attribute in React.

However, the ability of having to key an element / component only exist for the {#each} block. To workaround the constraint, it's common to use the "1-item keyed-each hack":

{#each key as k (k)}
<div />
{/each}

The <div /> will be recreated if the key has changed.

Transitions for reactive data change

Another commonly brought up request, to be able to apply transition: to an element when a reactive data changes (GitHub issue #5119):

<script>
import { fade } from 'svelte/transition'
let count = 0;
const handleClick = () => count +=1
</script>
<button on:click={handleClick}>Click me</button>
<p>You clicked <strong transition:fade>{count}</strong> times</p>

This is another facet of the same issue.

We need an ability to transition the old element out, and transition a new element in when a data, or a key changes.

A workaround, again, is to use the "1-item keyed-each hack":

<script>
import { fade } from 'svelte/transition'
let count = 0;
const handleClick = () => count +=1
</script>
<button on:click={handleClick}>Click me</button>
<p>You clicked
{#each [count] as count (count)}
<strong transition:fade>{count}</strong>
{/each}
times</p>

So the proposal of the feature request was to have a {#key} block:

<p>You clicked
  {#key count}
    <strong transition:fade>{count}</strong>
  {/key}
 times</p>

I've seen this issue months ago, and I passed the issue. I didn't think I know good enough to implement a new logic block. However, the issue recently resurfaced as someone commented on it recently. And this time, I felt I am ready, so here's my journey of implementing the {#key} block.

The implementation

As explained in "The Svelte Compiler Handbook", the Svelte compilation process can be broken into steps:

  • Parsing
  • Tracking references and dependencies
  • Creating code blocks & fragments
  • Generate code

Of course, that's the steps that we are going to work on as well.

Parsing

The actual parsing starts here in src/compiler/parse/index.ts:

let state: ParserState = fragment;

while (this.index < this.template.length) {
  state = state(this) || fragment;
}

There are 4 states in the parser:

  • fragment - in this state, we check the current character and determine which state we should proceed to
  • tag - we enter this state when we encounter < character. In this state, we are going to parse HTML tags (eg: <p>), attributes (eg: class) and directives (eg: on:).
  • mustache - we enter this state when we encounter { character. In this state, we are going to parse expression, { value } and logic blocks {#if}
  • text - In this state, we are going to parse texts that are neither < nor {, which includes whitespace, newlines, and texts!

To be able to parse the {#key} block, we are going to take a look at the mustache state function.

The {#key} block syntax is similar to {#if} without else, we take in an expression in the opening block and that's all:

{#key expression}
<div />
{/key}
<!-- similar to -->
{#if expression}
<div />
{/if}

So over here, when we encounter a {#, we add a case to check if we are starting a {#key} block:

// ...
} else if (parser.eat(#)) {
  // if {#if foo}, {#each foo} or {#await foo}
  let type;

  if (parser.eat('if')) {
    type = 'IfBlock';
  } else if (parser.eat('each')) {
    type = 'EachBlock';
  } else if (parser.eat('await')) {
    type = 'AwaitBlock';
   } else if (parser.eat('key')) {
     type = 'KeyBlock';
  } else {
    parser.error({
      code: `expected-block-type`,
       message: `Expected if, each or await`
       message: `Expected if, each, await or key`
    });
  }

Similarly, for closing block {/, we are going to make sure that {#key} closes with {/key}:

if (parser.eat('/')) {
  let block = parser.current();
  let expected;
  // ...
  if (block.type === 'IfBlock') {
    expected = 'if';
  } else if (block.type === 'EachBlock') {
    expected = 'each';
  } else if (block.type === 'AwaitBlock') {
    expected = 'await';
   } else if (block.type === 'KeyBlock') {
     expected = 'key';
  } else {
    parser.error({
      code: `unexpected-block-close`,
      message: `Unexpected block closing tag`
    });
  }

The next step is to read the JS expression. Since all logic blocks, {#if}, {#each} and {#await} will read the JS expression next, it is no different for {#key} and it is already taken care of:

parser.require_whitespace();

// read the JS expression
const expression = read_expression(parser);

// create the AST node
const block: TemplateNode = {...};

parser.allow_whitespace();

// other logic blocks specific syntax
if (type === 'EachBlock') {
  // {#each} block specific syntax for {#each list as item}
  // ...
}

So, let's move on to the next step!

Tracking references and dependencies

If you noticed in the previous step, the type name we created for {#key} block is called KeyBlock.

So, to keep the name consistent, we are going to create a KeyBlock class in src/compiler/compile/nodes/KeyBlock.ts:

import Expression from './shared/Expression';
import map_children from './shared/map_children';
import AbstractBlock from './shared/AbstractBlock';

export default class KeyBlock extends AbstractBlock {
  // for discriminant property for TypeScript to differentiate types
  type: 'KeyBlock';

  expression: Expression;

  constructor(component, parent, scope, info) {
    super(component, parent, scope, info);

    // create an Expression instance for the expression
    this.expression = new Expression(component, this, scope, info.expression);

    // loop through children and create respective node instance
    this.children = map_children(component, this, scope, info.children);

    // simple validation: make sure the block is not empty
    this.warn_if_empty_block();
  }
}

I've added comments annotating the code above, hopefully it's self-explanatory.

A few more points:

  • info is the AST node we got from the parsing.
  • the class Expression is constructed with the JavaScript AST of the expression and it is where we traverse the AST and marked the variables within the expression as referenced: true.
  • map_children is used to map the children of the KeyBlock AST node to the compile node.

Pardon for my lack of "appropriate" naming to differentiate the nodes in the Svelte codebase.

Throughout the Svelte compilation process, the node is transformed one to another, which in every step of the transformation, new analysis is performed, and new information are added.

Here, I am going to call:

If you managed to keep up so far, you may be sensing where we are heading next.

We need to add KeyBlock into map_children:

// src/compiler/compile/nodes/shared/map_children.ts
function get_constructor(type) {
  switch (type) {
    case 'AwaitBlock':
      return AwaitBlock;
    case 'Body':
      return Body;
    // ...
    case 'KeyBlock':
      return KeyBlock;
    // ...
  }
}

Also, we need to add KeyBlock as one of the INode type for TypeScript:

// src/compiler/compile/nodes/interfaces.ts
export type INode =
  | Action
  | Animation
  // ...
  | KeyBlock;
// ...

And now, let's move on to implementing a render-dom Wrapper for KeyBlock.

Creating code blocks & fragments

At this point, we need to decide how the compiled JS should look like, it's time for us to reverse-compile Svelte in your head!

If you've read my Compile Svelte in your head (Part 4), you've seen how we create a different create_fragment function for each of the logic branches, so we can control the content within a logic branch as a whole.

Similarly, we can create a create_fragment function for the content of the {#key}, then we can control when to create / mount / update / destroy the content.

function create_key_block(ctx) {
  // instructions to create / mount / update / destroy inner content of {#key}
  return {
    c() {},
    m() {},
    p() {},
    d() {},
  };
}

To use the create_key_block:

const key_block = create_key_block(ctx);
// create the elements for the {#key}
key_block.c();

// mount the elements in the {#key}
key_block.m(target, anchor);

// update the elements in the {#key}
key_block.p(ctx, dirty);

// destroy the elements in the {#key}
key_block.d(detaching);

// intro & outro the elements in the {#key}
transition_in(key_block);
transition_out(key_block);

The next thing to do, is to place these statements in the right position:

function create_fragment(ctx) {
  // init
  let key_block = create_key_block(ctx);

  return {
    c() {
      // create
      key_block.c();
    },
    m(target, anchor) {
      // mount
      key_block.m(target, anchor);
    },
    p(ctx, dirty) {
      // update
      key_block.p(ctx, dirty);
    },
    i(local) {
      // intro
      transition_in(key_block);
    },
    o(local) {
      // outro
      transition_out(key_block);
    },
    d(detaching) {
      // destroy
      key_block.d(detaching);
    },
  };
}

Now, the most important piece of the {#key} block, the logic to

  • check if the expression has changed
  • if so, recreate the elements inside the {#key} block
function create_fragment(ctx) {
  // we store the previous key expression value
  let previous_key = value_of_the_key_expression;
  // ...
  return {
    // ...
    p(ctx, dirty) {
      if (
        // if the any variables within the key has changed, and
        dirty & dynamic_variables_in_key_expression &&
        // if the value of the key expression has changed
        previous_key !== (previous_key = value_of_the_key_expression)
      ) {
        // destroy the elements
        // detaching = 1 (true) to remove the elements immediately
        key_block.d(1);
        // create a new key_block
        key_block = create_key_block(ctx);
        key_block.c();
        // mount the new key_block
        key_block.m(...);
      } else {
        // if the key has not changed, make sure the content of {#key} is up to date
        key_block.p();
      }
    }
    // ...
  }
}

If there is transition in the content of the key_block, we need extra code for the transition:

// instead of key_block.d(1);
group_outros();
transition_out(key_block, 1, 1, noop);
check_outros();

// before key_block.m(...)
transition_in(key_block);

I am going to gloss over the details of how outros / intros work, we will cover them in the later parts of "Compile Svelte in your head", so let's assume these code are up for the job.

Now we have done the reverse-compile Svelte in your head, let's reverse the reverse, and write the render code for Svelte {#key} block.

Here are some setup code for the render-dom Wrapper for {#key}:

export default class KeyBlockWrapper extends Wrapper {
  // ...
  // the `key_block` variable
  var: Identifier = { type: 'Identifier', name: 'key_block' };

  constructor(renderer: Renderer, block: Block, parent: Wrapper, node: EachBlock, strip_whitespace: boolean, next_sibling: Wrapper) {
    super(renderer, block, parent, node);

    // deoptimisation, set flag indicate the content is not static
    this.cannot_use_innerhtml();
    this.not_static_content();

    // get all the dynamic variables within the expression
    // useful for later
    this.dependencies = node.expression.dynamic_dependencies();

    // create a new `create_fragment` function
    this.block = block.child({
      comment: create_debugging_comment(node, renderer.component),
      name: renderer.component.get_unique_name('create_key_block'),
      type: 'key',
    });
    renderer.blocks.push(block);

    // create render-dom Wrappers for the children
    this.fragment = new FragmentWrapper(renderer, this.block, node.children, parent, strip_whitespace, next_sibling);
  }
  render(block: Block, parent_node: Identifier, parent_nodes: Identifier) {
    // NOTE: here is where we write the render code
  }
}

A few more points:

  • the block in the render method is the current create_fragment function that the {#key} block is in; this.block is the new create_fragment function that we created to put the content of the {#key} block
    • we named the new create_fragment function "create_key_block"
    • to make sure there's no conflicting names, we use renderer.component.get_unique_name()
  • All render-dom wrappers has a property named var, which is the variable name referencing the element / block to be created by the render-dom wrapper.

Now, let's implement the render method.

Firstly, render the children into this.block:

render(block: Block, parent_node: Identifier, parent_nodes: Identifier) {
  this.fragment.render(
    this.block,
    null,
    (x`#nodes` as unknown) as Identifier
  );
}

We pass in null as parent_node and x`#nodes` as parent_nodes to indicate that the children will be rendered at the root of the this.block.


If I am implementing the render method of an Element render-dom Wrapper, and currently rendering the <div> in the following code snippet:

<div>
  <span />
</div>

then I will render the <span /> with:

spanWrapper.render(
  block,
  this.var, // div's var
  x`${this.var.name}.childNodes`, // div.childNodes
);

so the <span /> will be inserted into the current <div /> and hydrate from the <div />'s childNodes.


Next, I am going to insert code into each of the fragment methods:

// let key_block = create_key_block(ctx);
block.chunks.init.push(
  b`let ${this.var} = ${this.block.name}(#ctx)`
);

// key_block.c();
block.chunks.create.push(b`${this.var}.c();`);

// key_block.m(...);
block.chunks.mount.push(
  b`${this.var}.m(${parent_node || "#target"}, ${parent_node ? "null" : "#anchor"});`
);

// key_block.p(...);
block.chunks.update.push(
  b`${this.var}.p(#ctx, #dirty);`
);

// key_block.d(...);
block.chunks.destroy.push(b`${this.var}.d(detaching)`);

A few more points:

  • we push the code into respective methods of the block, eg: init, create, mount, ...
  • we use tagged templates, b`...` to create a JavaScript AST node. The b tag function allow us to pass in JavaScript AST node as placeholder, so that is very convenient.
    • You can check out more about the b tag function from code-red

Now, to implement the dirty checking, we use this.dependencies

const is_dirty = this.renderer.dirty(this.dependencies);

To determine whether our expression value has changed, we are going to compute the expression and compare it with previous_key and determine whether it has changed.

Here's a recap of the compiled code that we've come up previously:

// we store the previous key expression value
let previous_key = value_of_the_key_expression;
// ...
// if the value of the key expression has changed
previous_key !== (previous_key = value_of_the_key_expression)

We start with declaring the variable, previous_key:

const previous_key = block.get_unique_name('previous_key');
const snippet = this.node.expression.manipulate(block);
block.add_variable(previous_key, snippet);

expression.manipulate(block) will convert the expression to refer to the ctx variable, for example:

human.age + limit
// into something like
ctx[0].age + ctx[2]

Next we are going to compare the new value and assign it to previous_key after that.

const has_change = x`${previous_key} !== (${previous_key} = ${snippet})`

And to combine all of these, we have:

block.chunks.update.push(b`
  if (${is_dirty} && ${has_change}) {
    ${this.var}.d(1);
    ${this.var} = ${this.block.name}(#ctx);
    ${this.var}.c();
    ${this.var}.m(${this.get_update_mount_node(anchor)}, ${anchor});
  } else {
    ${this.var}.p(#ctx, #dirty);
  }
`);

We are using the anchor when we are mounting the new key_block, you can check out Compile Svelte in your head Part 4: the extra text node, explaining why we need the anchor node, and here is how the anchor node being computed:

const anchor = this.get_or_create_anchor(block, parent_node, parent_nodes);

It could be the next sibling, or it could be a new empty() text node created.

Finally, if the content has transition, we need to add code for the transition as well:

const has_transitions = !!(this.block.has_intro_method || this.block.has_outro_method);
const transition_out = b`
  @group_outros();
  @transition_out(${this.var}, 1, 1, @noop);
  @check_outros();
`;
const transition_in = b`
  @transition_in(${this.var});
`;

Where to place them? Well, I'll leave that as your exercise to figure that out. 😉

Creating code for SSR

For SSR, it is much simpler than for the dom. {#key} block has no special meaning in SSR, because, you will only render once in SSR:

import KeyBlock from '../../nodes/KeyBlock';
import Renderer, { RenderOptions } from '../Renderer';

export default function(node: KeyBlock, renderer: Renderer, options: RenderOptions) {
	renderer.render(node.children, options);
}

☝️ That's all the code we need for SSR. We are rendering the children, passing down the options, and add no extra code for the {#key} block.

Generate code

Well, everything in this step is set up generic enough to handle most use case.

So, nothing to change here. 🤷‍♂️

A few other implementation consideration

  • What if the expression in the {#key} block is not dynamic, do we give warnings? or optimise the output?
  • How will <svelte:options immutable={true}> affect the code output?

The testing

You've seen me implementing test cases in the previous "Contributing to Svelte" articles [1] [2], here I am going to skip showing the implementation of the test cases, and probably point out some thoughts I had when coming up with tests:

  1. Happy path: changing the key expression should recreate the content

  2. Happy path: Transition when recreating the content should work ✨

  3. Possible edge case: Changing variables other than the key expression should not recreate the content in {#key}

    <script>
    let reactive1;
    let reactive2;
    let key;
    </script>
    {#key key}
    {key} {reactive1}
    {/key}
    {reactive2}
  4. Possible edge case: Changing the variables within the key expression but the result value of the key expression stay the same

    <script>
    let a = 1;
    let b = 2;
    function update() {
    a = 2;
    b = 1;
    }
    </script>
    {#key a + b}
    <div />
    {/key}

Closing Notes

You can read the Pull Request #5397 to read the final implementation.


If you wish to learn more about Svelte, follow me on Twitter.

If you have anything unclear about this article, find me on Twitter too!