Recent blog entries for mbrubeck

Let's build a browser engine! Part 4: Style

Welcome back to my series on building your own toy browser engine. If you’re just tuning in, you can find the previous episodes here:

This article will cover what the CSS standard calls assigning property values, or what I call the style module. This module takes DOM nodes and CSS rules as input, and matches them up to determine the value of each CSS property for any given node.

This part doesn’t contain a lot of code, since I’ve left out all the really complicated parts. However, I think what’s left is still quite interesting, and I’ll also explain how some of the missing pieces can be implemented.

The Style Tree

The output of the style module is something I call the style tree. Each node in this tree includes a pointer to a DOM node, plus its CSS property values:

    /// Map from CSS property names to values.
type PropertyMap = HashMap<String, Value>;

/// A node with associated style data.
struct StyledNode<'a> {
    node: &'a Node, // pointer to a DOM node
    specified_values: PropertyMap,
    children: Vec<StyledNode<'a>>,
}

  

What’s with all the 'a stuff? These are lifetime annotations, part of how Rust guarantees that pointers are memory-safe without requiring garbage collection. If you are not working in Rust you can safely ignore them; they aren’t critical to the meaning of this code.

We could add style information directly to the dom::Node struct from Part 1 instead, but I wanted to keep this code out of the earlier “lessons.” This is also a good opportunity to talk about the parallel trees that inhabit most layout engines.

A browser engine module often takes one tree as input, and produces a different but related tree as output. For example, Gecko’s layout code takes a DOM tree and produces a frame tree, which is then used to build a view tree. Blink and WebKit transform the DOM tree into a render tree. Later stages in all these engines produce still more trees, including layer trees and widget trees.

The pipeline for our toy browser engines will look something like this after we complete a few more stages:

In my implementation, each node in the DOM tree produces exactly one node in the style tree. But in a more complicated pipeline stage, several input nodes could collapse into a single output node. Or one input node might expand into several output nodes, or be skipped completely. For example, the style tree could exclude elements whose display property is set to 'none'. (Instead this will happen in the layout stage, because my code turned out a bit simpler that way.)

Selector Matching

The first step in building the style tree is selector matching. This will be very easy, since my CSS parser supports only simple selectors. You can tell whether a simple selector matches an element just by looking at the element itself. Matching compound selectors would require traversing the DOM tree to look at the element’s siblings, parents, etc.

    fn matches(elem: &ElementData, selector: &Selector) -> bool {
    match *selector {
        Simple(ref simple_selector) => matches_simple_selector(elem, simple_selector)
    }
}

  

To help, we’ll add some convenient ID and class accessors to our DOM element type. The class attribute can contain multiple class names separated by spaces, which we return in a hash table. [Note: The Rust types below look a bit hairy because we are passing around pointers rather than copying values. This code should be a lot more concise in languages that are not so concerned with this distinction.]

    impl ElementData {
    fn get_attribute<'a>(&'a self, key: &str) -> Option<&'a String> {
        self.attributes.find_equiv(&key)
    }

    fn id<'a>(&'a self) -> Option<&'a String> {
        self.get_attribute("id")
    }

    fn classes<'a>(&'a self) -> HashSet<&'a str> {
        match self.get_attribute("class") {
            Some(classlist) => classlist.as_slice().split(' ').collect(),
            None => HashSet::new()
        }
    }
}

  

To test whether a simple selector matches an element, just look at each selector component, and return false if the element doesn’t have a matching class, ID, or tag name.

    fn matches_simple_selector(elem: &ElementData, selector: &SimpleSelector) -> bool {
    // Check type selector
    if selector.tag_name.iter().any(|name| elem.tag_name != *name) {
        return false;
    }

    // Check ID selector
    if selector.id.iter().any(|id| elem.id() != Some(id)) {
        return false;
    }

    // Check class selectors
    let elem_classes = elem.classes();
    if selector.class.iter().any(|class| !elem_classes.contains(&class.as_slice())) {
        return false;
    }

    // We didn't find any non-matching selector components.
    return true;
}

  

Rust note: This function uses the any method, which returns true if an iterator contains an element that passes the provided test. This is the same as the any function in Python (or Haskell), or the some method in JavaScript.

When comparing two rules that match the same element, we need to use the highest-specificity selector from each match. Because our CSS parser stores the selectors from most- to least-specific, we can stop as soon as we find a matching one, and return its specificity along with a pointer to the rule.

    /// A single CSS rule and the specificity of its most specific matching selector.
type MatchedRule<'a> = (Specificity, &'a Rule);

/// If `rule` matches `elem`, return a `MatchedRule`. Otherwise return `None`.
fn match_rule<'a>(elem: &ElementData, rule: &'a Rule) -> Option<MatchedRule<'a>> {
    // Find the first (highest-specificity) matching selector.
    rule.selectors.iter().find(|selector| matches(elem, *selector))
        .map(|selector| (selector.specificity(), rule))
}

  

To find all the rules that match an element we call filter_map, which does a linear scan through the style sheet, checking every rule and throwing out ones that don’t match. A real browser engine would speed this up by storing the rules in multiple hash tables based on tag name, id, class, etc.

    /// Find all CSS rules that match the given element.
fn matching_rules<'a>(elem: &ElementData, stylesheet: &'a Stylesheet) -> Vec<MatchedRule<'a>> {
    stylesheet.rules.iter().filter_map(|rule| match_rule(elem, rule)).collect()
}

  

Once we have the matching rules, we can find the specified values for the element. We insert each rule’s property values into a HashMap. We sort the matches by specificity, so the higher specificity rules are processed after the lower ones and can overwrite their values in the HashMap.

    /// Apply styles to a single element, returning the specified styles.
fn specified_values(elem: &ElementData, stylesheet: &Stylesheet) -> PropertyMap {
    let mut values = HashMap::new();
    let mut rules = matching_rules(elem, stylesheet);

    // Go through the rules from lowest to highest specificity.
    rules.sort_by(|&(a, _), &(b, _)| a.cmp(&b));
    for &(_, rule) in rules.iter() {
        for declaration in rule.declarations.iter() {
            values.insert(declaration.name.clone(), declaration.value.clone());
        }
    }
    return values;
}

  

Now we have everything we need to walk through the DOM tree and build the style tree. Note that selector matching works only on elements, so the specified values for a text node are just an empty map.

    /// Apply a stylesheet to an entire DOM tree, returning a StyledNode tree.
pub fn style_tree<'a>(root: &'a Node, stylesheet: &'a Stylesheet) -> StyledNode<'a> {
    StyledNode {
        node: root,
        specified_values: match root.node_type {
            Element(ref elem) => specified_values(elem, stylesheet),
            Text(_) => HashMap::new()
        },
        children: root.children.iter().map(|child| style_tree(child, stylesheet)).collect(),
    }
}

  

That’s all of robinson’s code for building the style tree. Next I’ll talk about some glaring omissions.

The Cascade

Style sheets provided by the author of a web page are called author style sheets. In addition to these, browsers also provide default styles via user agent style sheets. And they may allow users to add custom styles through user style sheets (like Gecko’s userContent.css).

The cascade defines which of these three “origins” takes precedence over another. There are six levels to the cascade: one for each origin’s “normal” declarations, plus one for each origin’s !important declarations.

Robinson’s style code does not implement the cascade; it uses only a single style sheet. The lack of a default style sheet means that HTML elements will not have any of the default styles you might expect. For example, the <head> element’s contents will not be hidden unless you explicitly add this rule to your style sheet:

    head { display: none; }

  

Implementing the cascade should by fairly easy: Just track the origin of each rule, and sort declarations by origin and importance in addition to specificity. A simplified, two-level cascade should be enough to support the most common cases: normal user agent styles and normal author styles.

Computed Values

In addition to the “specified values” mentioned above, CSS defines initial, computed, used, and actual values.

Initial values are defaults for properties that aren’t specified in the cascade. Computed values are based on specified values, but may have some property-specific normalization rules applied.

Implementing these correctly requires separate code for each property, based on its definition in the CSS specs. This work is necessary for a real-world browser engine, but I’m hoping to avoid it in this toy project. In later stages, code that uses these values will (sort of) simulate initial values by using a default when the specified value is missing.

Used values and actual values are calculated during and after layout, which I’ll cover in future articles.

Inheritance

If text nodes can’t match selectors, how do they get colors and fonts and other styles? Through the magic of inheritance.

When a property is inherited, any node without a cascaded value will receive its parent’s value for that property. Some properties, like 'color', are inherited by default; others only if the cascade specifies the special value 'inherit'.

My code does not support inheritance. To implement it, you could pass the parent’s style data into the specified_values function, and use a hard-coded lookup table to decide which properties should be inherited.

Style Attributes

Any HTML element can include a style attribute containing a list of CSS declarations. There are no selectors, because these declarations automatically apply only to the element itself.

    <span style="color: red; background: yellow;">

  

If you want to support the style attribute, make the specified_values function check for the attribute. If the attribute is present, pass it to parse_declarations from the CSS parser. Apply the resulting declarations after the normal author declarations, since the attribute is more specific than any CSS selector.

Exercises

In addition to writing your own selector matching and value assignment code, for further exercise you can implement one or more of the missing pieces discussed above, in your own project or a fork of robinson:

  1. Cascading
  2. Initial and/or computed values
  3. Inheritance
  4. The style attribute

Also, if you extended the CSS parser from Part 3 to include compound selectors, you can now implement matching for those compound selectors.

To be continued…

Part 5 will introduce the layout module. I haven’t finished the code for this yet, so there will be another short before I can start writing the article. I plan to split layout into at least two articles (one for block layout and one for inline layout, probably).

In the meantime, I’d love to see anything you’ve created based on these articles or exercises. If your code is online somewhere, feel free to add a link to the comments below! So far I have seen Martin Tomasi’s Java implementation and Pohl Longsine’s Swift version.

Syndicated 2014-08-25 22:45:00 from Matt Brubeck

Let's build a browser engine! Part 3: CSS

This is the third in a series of articles on building a toy browser rendering engine. Want to build your own? Start at the beginning to learn more:

This article introduces code for reading Cascading Style Sheets (CSS). As usual, I won’t try to cover everything in the spec. Instead, I tried to implement just enough to illustrate some concepts and produce input for later stages in the rendering pipeline.

Anatomy of a Stylesheet

Here’s an example of CSS source code:

    h1, h2, h3 { margin: auto; color: #cc0000; }
div.note { margin-bottom: 20px; padding: 10px; }
#answer { display: none; }

  

Now I’ll walk through some the CSS code from my toy browser engine, robinson.

A CSS stylesheet is a series of rules. (In the example stylesheet above, each line contains one rule.)

    struct Stylesheet {
    rules: Vec<Rule>,
}

  

A rule includes one or more selectors separated by commas, followed by a list of declarations enclosed in braces.

    struct Rule {
    selectors: Vec<Selector>,
    declarations: Vec<Declaration>,
}

  

A selector can be a simple selector, or it can be a chain of selectors joined by combinators. Robinson supports only simple selectors for now.

Note: Confusingly, the newer Selectors Level 3 standard uses the same terms to mean slightly different things. In this article I’ll mostly refer to CSS2.1. Although outdated, it’s a useful starting point because it’s smaller and more self-contained than CSS3 (which is split into myriad specs that reference both each other and CSS2.1).

In robinson, a simple selector can include a tag name, an ID prefixed by '#', any number of class names prefixed by '.', or some combination of the above. If the tag name is empty or '*' then it is a “universal selector” that can match any tag.

There are many other types of selector (especially in CSS3), but this will do for now.

    enum Selector {
    Simple(SimpleSelector),
}

struct SimpleSelector {
    tag_name: Option<String>,
    id: Option<String>,
    class: Vec<String>,
}

  

A declaration is just a name/value pair, separated by a colon and ending with a semicolon. For example, "margin: auto;" is a declaration.

    struct Declaration {
    name: String,
    value: Value,
}

  

My toy engine supports only a handful of CSS’s many value types.

    enum Value {
    Keyword(String),
    Color(u8, u8, u8, u8), // RGBA
    Length(f32, Unit),
    // insert more values here
}

enum Unit { Px, /* insert more units here */ }

  

All other CSS syntax is unsupported, including at-rules, comments, and any selectors/values/units not mentioned above.

Parsing

CSS has a regular grammar, making it easier to parse correctly than its quirky cousin HTML. When a standards-compliant CSS parser encounters a parse error, it discards the unrecognized part of the stylesheet but still processes the remaining portions. This is useful because it allows stylesheets to include new syntax but still produce well-defined output in older browsers.

Robinson uses a very simplistic (and totally not standards-compliant) parser, built the same way as the HTML parser from Part 2. Rather than go through the whole thing line-by-line again, I’ll just paste in a few snippets. For example, here is the code for parsing a single selector:

        /// Parse one simple selector, e.g.: `type#id.class1.class2.class3`
    fn parse_simple_selector(&mut self) -> SimpleSelector {
        let mut result = SimpleSelector { tag_name: None, id: None, class: Vec::new() };
        while !self.eof() {
            match self.next_char() {
                '#' => {
                    self.consume_char();
                    result.id = Some(self.parse_identifier());
                }
                '.' => {
                    self.consume_char();
                    result.class.push(self.parse_identifier());
                }
                '*' => {
                    // universal selector
                    self.consume_char();
                }
                c if valid_identifier_char(c) => {
                    result.tag_name = Some(self.parse_identifier());
                }
                _ => break
            }
        }
        result
    }

  

Note the lack of error checking. Some malformed input like ### or *foo* will parse successfully and produce weird results. A real CSS parser would discard these invalid selectors.

Specificity

Specificity is one of the ways a rendering engine decides which style overrides the other in a conflict. If a stylesheet contains two rules that match an element, the rule with the matching selector of higher specificity can override values from the one with lower specificity.

The specificity of a selector is based on its components. An ID selector is more specific than a class selector, which is more specific than a tag selector. Within each of these “levels,” more selectors beats fewer.

    pub type Specificity = (uint, uint, uint);

impl Selector {
    pub fn specificity(&self) -> Specificity {
        // http://www.w3.org/TR/selectors/#specificity
        let Simple(ref simple) = *self;
        let a = simple.id.iter().len();
        let b = simple.class.len();
        let c = simple.tag_name.iter().len();
        (a, b, c)
    }
}

  

[If we supported chained selectors, we could calculate the specificity of a chain just by adding up the specificities of its parts.]

The selectors for each rule are stored in a sorted vector, most-specific first. This will be important in matching, which I’ll cover in the next article.

        /// Parse a rule set: `<selectors> { <declarations> }`.
    fn parse_rule(&mut self) -> Rule {
        Rule {
            selectors: self.parse_selectors(),
            declarations: self.parse_declarations()
        }
    }

    /// Parse a comma-separated list of selectors.
    fn parse_selectors(&mut self) -> Vec<Selector> {
        let mut selectors = Vec::new();
        loop {
            selectors.push(Simple(self.parse_simple_selector()));
            self.consume_whitespace();
            match self.next_char() {
                ',' => { self.consume_char(); }
                '{' => break, // start of declarations
                c   => fail!("Unexpected character {} in selector list", c)
            }
        }
        // Return selectors with highest specificity first, for use in matching.
        selectors.sort_by(|a,b| b.specificity().cmp(&a.specificity()));
        selectors
    }

  

The rest of the CSS parser is fairly straightforward. You can read the whole thing on GitHub. And if you didn’t already do it for Part 2, this would be a great time to try out a parser generator. My hand-rolled parser gets the job done for simple example files, but it has a lot of hacky bits and will fail badly if you violate its assumptions. Eventually I hope to replace it with a “real” parser built on something like rust-peg.

Exercises

As before, you should decide which of these exercises you want to do, and skip the rest:

  1. Implement your own simplified CSS parser and specificity calculation.

  2. Extend robinson’s CSS parser to support more values, or one or more selector combinators.

  3. Extend the CSS parser to discard any declaration that contains a parse error, and follow the error handling rules to resume parsing after the end of the declaration.

  4. Make the HTML parser pass the contents of any <style> nodes to the CSS parser, and return a Document object that includes a list of Stylesheets in addition to the DOM tree.

Shortcuts

Just like in Part 2, you can skip parsing by hard-coding CSS data structures directly into your program, or by writing them in an alternate format like JSON that you already have a parser for.

To be continued…

The next article will introduce the style module. This is where everything starts to come together, with selector matching to apply CSS styles to DOM nodes.

The pace of this series might slow down soon, since I’ll be busy later this month and I haven’t even written the code for some upcoming articles. I’ll keep them coming as fast as I can!

Syndicated 2014-08-13 19:30:00 from Matt Brubeck

Let's build a browser engine!

This is the second in a series of articles on building a toy browser rendering engine:

This article is about parsing HTML source code to produce a tree of DOM nodes. Parsing is a fascinating topic, but I don’t have the time or expertise to give it the introduction it deserves. You can get a detailed introduction to parsing from any good course or book on compilers. Or get a hands-on start by going through the documentation for a parser generator that works with your chosen programming language.

HTML has its own unique parsing algorithm. Unlike parsers for most programming languages and file formats, the HTML parsing algorithm does not reject invalid input. Instead it includes specific error-handling instructions, so web browsers can agree on how to display every web page, even ones that don’t conform to the syntax rules. Web browsers have to do this to be usable: Since non-conforming HTML has been supported since the early days of the web, it is now used in a huge portion of existing web pages.

A Simple HTML Dialect

I didn’t even try to implement the standard HTML parsing algorithm. Instead I wrote a basic parser for a tiny subset of HTML syntax. My parser can handle simple pages like this:

    <html>
    <body>
        <h1>Title</h1>
        <div id="main" class="test">
            <p>Hello <em>world</em>!</p>
        </div>
    </body>
</html>

  

The following syntax is allowed:

  • Balanced tags: <p>...</p>
  • Attributes with quoted values: id="main"
  • Text nodes: <em>world</em>

Everything else is unsupported, including:

  • Namespaces: <html:body>
  • Self-closing tags: <br/> or <br> with no closing tag
  • Character encoding detection.
  • Escaped characters (like &amp;) and CDATA blocks.
  • Comments, processing instructions, and doctype declarations.
  • Error handling (e.g. unbalanced or improperly nested tags).

At each stage of this project I’m writing more or less the minimum code needed to support the later stages. But if you want to learn more about parsing theory and tools, you can be much more ambitious in your own project!

Example Code

Next, let’s walk through my toy HTML parser, keeping in mind that this is just one way to do it (and probably not the best way). Its structure is based loosely on the tokenizer module from Servo’s cssparser library. It has no real error handling; in most cases, it just aborts when faced with unexpected syntax. The code is in Rust, but I hope it’s fairly readable to anyone who’s used similar-looking languages like Java, C++, or C#. It makes use of the DOM data structures from part 1.

The parser stores its input string and a current position within the string. The position is the index of the next character we haven’t processed yet.

    struct Parser {
    pos: uint,
    input: String,
}

  

We can use this to implement some simple methods for peeking at the next characters in the input:

    impl Parser {
    /// Read the next character without consuming it.
    fn next_char(&self) -> char {
        self.input.as_slice().char_at(self.pos)
    }

    /// Do the next characters start with the given string?
    fn starts_with(&self, s: &str) -> bool {
        self.input.as_slice().slice_from(self.pos).starts_with(s)
    }

    /// Return true if all input is consumed.
    fn eof(&self) -> bool {
        self.pos >= self.input.len()
    }

    // ...
}

  

Rust strings are stored as UTF-8 byte arrays. To go to the next character, we can’t just advance by one byte. Instead we use char_range_at which correctly handles multi-byte characters. (If our string used fixed-width characters, we could just increment pos.)

        /// Return the current character, and advance to the next character.
    fn consume_char(&mut self) -> char {
        let range = self.input.as_slice().char_range_at(self.pos);
        self.pos = range.next;
        range.ch
    }

  

Often we will want to consume a string of consecutive characters. The consume_while method consumes characters that meet a given condition, and returns them as a string:

        /// Consume characters until `test` returns false.
    fn consume_while(&mut self, test: |char| -> bool) -> String {
        let mut result = String::new();
        while !self.eof() && test(self.next_char()) {
            result.push_char(self.consume_char());
        }
        result
    }

  

We can use this to ignore a sequence of space characters, or to consume a string of alphanumeric characters:

        /// Consume and discard zero or more whitespace characters.
    fn consume_whitespace(&mut self) {
        self.consume_while(|c| c.is_whitespace());
    }

    /// Parse a tag or attribute name.
    fn parse_tag_name(&mut self) -> String {
        self.consume_while(|c| match c {
            'a'..'z' | 'A'..'Z' | '0'..'9' => true,
            _ => false
        })
    }

  

Now we’re ready to start parsing HTML. To parse a single node, we look at its first character to see if it is an element or a text node. In our simplified version of HTML, a text node can contain any character except <.

        /// Parse a single node.
    fn parse_node(&mut self) -> dom::Node {
        match self.next_char() {
            '<' => self.parse_element(),
            _   => self.parse_text()
        }
    }

    /// Parse a text node.
    fn parse_text(&mut self) -> dom::Node {
        dom::text(self.consume_while(|c| c != '<'))
    }

  

An element is more complicated. It includes opening and closing tags, and between them any number of child nodes:

        /// Parse a single element, including its open tag, contents, and closing tag.
    fn parse_element(&mut self) -> dom::Node {
        // Opening tag.
        assert!(self.consume_char() == '<');
        let tag_name = self.parse_tag_name();
        let attrs = self.parse_attributes();
        assert!(self.consume_char() == '>');

        // Contents.
        let children = self.parse_nodes();

        // Closing tag.
        assert!(self.consume_char() == '<');
        assert!(self.consume_char() == '/');
        assert!(self.parse_tag_name() == tag_name);
        assert!(self.consume_char() == '>');

        dom::elem(tag_name, attrs, children)
    }

  

Parsing attributes is pretty easy in our simplified syntax. Until we reach the end of the opening tag (>) we repeatedly look for a name followed by = and then a string enclosed in quotes.

        /// Parse a single name="value" pair.
    fn parse_attr(&mut self) -> (String, String) {
        let name = self.parse_tag_name();
        assert!(self.consume_char() == '=');
        let value = self.parse_attr_value();
        (name, value)
    }

    /// Parse a quoted value.
    fn parse_attr_value(&mut self) -> String {
        let open_quote = self.consume_char();
        assert!(open_quote == '"' || open_quote == '\'');
        let value = self.consume_while(|c| c != open_quote);
        assert!(self.consume_char() == open_quote);
        value
    }

    /// Parse a list of name="value" pairs, separated by whitespace.
    fn parse_attributes(&mut self) -> dom::AttrMap {
        let mut attributes = HashMap::new();
        loop {
            self.consume_whitespace();
            if self.next_char() == '>' {
                break;
            }
            let (name, value) = self.parse_attr();
            attributes.insert(name, value);
        }
        attributes
    }

  

To parse the child nodes, we recursively call parse_node in a loop until we reach the closing tag:

        /// Parse a sequence of sibling nodes.
    fn parse_nodes(&mut self) -> Vec<dom::Node> {
        let mut nodes = vec!();
        loop {
            self.consume_whitespace();
            if self.eof() || self.starts_with("</") {
                break;
            }
            nodes.push(self.parse_node());
        }
        nodes
    }

  

Finally, we can put this all together to parse an entire HTML document into a DOM tree. This function will create a root node for the document if it doesn’t include one explicitly; this is similar to what a real HTML parser does.

    /// Parse an HTML document and return the root element.
pub fn parse(source: String) -> dom::Node {
    let mut nodes = Parser { pos: 0u, input: source }.parse_nodes();

    // If the document contains a root element, just return it. Otherwise, create one.
    if nodes.len() == 1 {
        nodes.swap_remove(0).unwrap()
    } else {
        dom::elem("html".to_string(), HashMap::new(), nodes)
    }
}

  

That’s it! The entire code for the robinson HTML parser. The whole thing weighs in at just over 100 lines of code (not counting blank lines and comments). If you use a good library or parser generator, you can probably build a similar toy parser in even less space.

Exercises

Here are a few alternate ways to try this out yourself. As before, you can choose one or more of them and ignore the others.

  1. Build a parser (either “by hand” or with a library or parser generator) that takes a subset of HTML as input and produces a tree of DOM nodes.

  2. Modify robinson’s HTML parser to add some missing features, like comments. Or replace it with a better parser, perhaps built with a library or generator.

  3. Create an invalid HTML file that causes your parser (or mine) to fail. Modify the parser to recover from the error and produce a DOM tree for your test file.

Shortcuts

If you want to skip parsing completely, you can build a DOM tree programmatically instead, by adding some code like this to your program (in pseudo-code; adjust it to match the DOM code you wrote in Part 1):

    // <html><body>Hello, world!</body></html>
let root = element("html");
let body = element("body");
root.children.push(body);
body.children.push(text("Hello, world!"));

  

Or you can find an existing HTML parser and incorporate it into your program.

The next article in this series will cover CSS data structures and parsing.

Syndicated 2014-08-11 15:00:00 from Matt Brubeck

Let's build a browser engine!

I’m building a toy HTML rendering engine, and I think you should too. This is the first in a series of articles describing my project and how you can make your own. But first, let me explain why.

You’re building a what?

Let’s talk terminology. A browser engine is the portion of a web browser that works “under the hood” to fetch a web page from the internet, and translate its contents into forms you can read, watch, hear, etc. Blink, Gecko, WebKit, and Trident are browser engines. In contrast, the the browser’s own UI—tabs, toolbar, menu and such—is called the chrome. Firefox and SeaMonkey are two browsers with different chrome but the same Gecko engine.

A browser engine includes many sub-components: an HTTP client, an HTML parser, a CSS parser, a JavaScript engine (itself composed of parsers, interpreters, and compilers), and much more. The many components involved in parsing web formats like HTML and CSS and translating them into what you see on-screen are sometimes called the layout engine or rendering engine.

Why a “toy” rendering engine?

A full-featured browser engine is hugely complex. Blink, Gecko, WebKit—these are millions of lines of code each. Even younger, simpler rendering engines like Servo and WeasyPrint are each tens of thousands of lines. Not the easiest thing for a newcomer to comprehend!

Speaking of hugely complex software: If you take a class on compilers or operating systems, at some point you will probably create or modify a “toy” compiler or kernel. This is a simple model designed for learning; it may never be run by anyone besides the person who wrote it. But making a toy system is a useful tool for learning how the real thing works. Even if you never build a real-world compiler or kernel, understanding how they work can help you make better use of them when writing your own programs.

So, if you want to become a browser developer, or just to understand what happens inside a browser engine, why not build a toy one? Like a toy compiler that implements a subset of a “real” programming language, a toy rendering engine could implement a small subset of HTML and CSS. It won’t replace the engine in your everyday browser, but should nonetheless illustrate the basic steps needed for rendering a simple HTML document.

Try this at home.

I hope I’ve convinced you to give it a try. This series will be easiest to follow if you already have some solid programming experience and know some high-level HTML and CSS concepts. However, if you’re just getting started with this stuff, or run into things you don’t understand, feel free to ask questions and I’ll try to make it clearer.

Before you start, a few remarks on some choices you can make:

On Programming Languages

You can build a toy layout engine in any programming language. Really! Go ahead and use a language you know and love. Or use this as an excuse to learn a new language if that sounds like fun.

If you want to start contributing to major browser engines like Gecko or WebKit, you might want to work in C++ because it’s the main language used in those engines, and using it will make it easier to compare your code to theirs. My own toy project, robinson, is written in Rust. I’m part of the Servo team at Mozilla, so I’ve become very fond of Rust programming. Plus, one of my goals with this project is to understand more of Servo’s implementation. (I’ve written a lot of browser chrome code, and a few small patches for Gecko, but before joining the Servo project I knew nothing about many areas of the browser engine.) Robinson sometimes uses simplified versions of Servo’s data structures and code. If you too want to start contributing to Servo, try some of the exercises in Rust!

On Libraries and Shortcuts

In a learning exercise like this, you have to decide whether it’s “cheating” to use someone else’s code instead of writing your own from scratch. My advice is to write your own code for the parts that you really want to understand, but don’t be shy about using libraries for everything else. Learning how to use a particular library can be a worthwhile exercise in itself.

I’m writing robinson not just for myself, but also to serve as example code for these articles and exercises. For this and other reasons, I want it to be as tiny and self-contained as possible. So far I’ve used no external code except for the Rust standard library. (This also side-steps the minor hassle of getting multiple dependencies to build with the same version of Rust while the language is still in development.) This rule isn’t set in stone, though. For example, I may decide later to use a graphics library rather than write my own low-level drawing code.

Another way to avoid writing code is to just leave things out. For example, robinson has no networking code yet; it can only read local files. In a toy program, it’s fine to just skip things if you feel like it. I’ll point out potential shortcuts like this as I go along, so you can bypass steps that don’t interest you and jump straight to the good stuff. You can always fill in the gaps later if you change your mind.

First Step: The DOM

Are you ready to write some code? We’ll start with something small: data structures for the DOM. Let’s look at robinson’s dom module.

The DOM is a tree of nodes. A node has zero or more children. (It also has various other attributes and methods, but we can ignore most of those for now.)

    struct Node {
    // data common to all nodes:
    children: Vec<Node>,

    // data specific to each node type:
    node_type: NodeType,
}

  

There are several node types, but for now we will ignore most of them and say that a node is either an Element or a Text node. In a language with inheritance these would be subtypes of Node. In Rust they can be an enum (Rust’s keyword for a “tagged union” or “sum type”):

    enum NodeType {
    Text(String),
    Element(ElementData),
}

  

An element includes a tag name and any number of attributes, which can be stored as a map from names to values. Robinson doesn’t support namespaces, so it just stores tag and attribute names as simple strings.

    struct ElementData {
    tag_name: String,
    attributes: AttrMap,
}

type AttrMap = HashMap<String, String>;

  

Finally, some constructor functions to make it easy to create new nodes:

    impl Node {
    fn new(children: Vec<Node>, node_type: NodeType) -> Node {
        Node { children: children, node_type: node_type }
    }
}

fn text(data: String) -> Node {
    Node::new(vec!(), Text(data))
}

fn elem(name: String, attrs: AttrMap, children: Vec<Node>) -> Node {
    Node::new(children, Element(ElementData {
        tag_name: name,
        attributes: attrs,
    }))
}

  

And that’s it! A full-blown DOM implementation would include a lot more data and dozens of methods, but this is all we need to get started. In the next article, we’ll add a parser that turns HTML source code into a tree of these DOM nodes.

Exercises

These are just a few suggested ways to follow along at home. Do the exercises that interest you and skip any that don’t.

  1. Start a new program in the language of your choice, and write code to represent a tree of DOM text nodes and elements.

  2. Install the latest version of Rust, then download and build robinson. Open up dom.rs and extend NodeType to include additional types like comment nodes.

  3. Write code to pretty-print a tree of DOM nodes.

References

Here’s a short list of “small” open source web rendering engines. Most of them are many times bigger than robinson, but still way smaller than Gecko or WebKit. WebWhirr, at 2000 lines of code, is the only other one I would call a “toy” engine.

You may find these useful for inspiration or reference. If you know of any other similar projects—or if you start your own—please let me know!

Syndicated 2014-08-08 16:40:00 from Matt Brubeck

Better automated detection of Firefox performance regressions

Last spring I spent some of my spare time improving the automated script that detects regressions in Talos and other Firefox performance data. I’m finally writing up some of that work in case it’s useful or interesting to anyone else.

Talos is a system for running performance benchmarks; we use it to run a suite of benchmarks every time a change is pushed to the Firefox source repository. The Talos test harness reports these results to the graph server which stores them and can plot the recorded data to show how it changes over time.

Like most performance measurements, Talos benchmarks can be noisy. We need to use statistics to separate signal from noise. To determine whether a change to the source code caused a change in the benchmark results, an automated script takes multiple datapoints from before and after each push. It computes the average and standard deviation of the “before” datapoints and the “after” datapoints, and uses a Student’s t-test to estimate the likelihood that the datasets are significantly different. If the t-test exceeds a certain threshold, the script sends email to the author(s) of the responsible patches and to the dev-tree-management mailing list.

By nature, these statistical estimates can never be 100% certain. However, we need to make sure that they are correct as often as possible. False negatives mean that real regressions go undetected. But false positives will generate extra work, and may cause developers to ignore future regression alerts. I started inspecting graph server data and regression alerts by hand, recording and investigating any false negatives or false positives I found, and filed bugs to fix the causes of those errors.

Some of these were straightforward implementation bugs, like one where an infinite t-test score (near certain likelihood of regression) was treated as a zero score (no regression at all). Others involved tuning the number of datapoints and the threshold for sending alerts.

Some fixes required more involved changes to the analysis. For example, if one code change actually caused a regression, the pushes right before or after that change will also appear somewhat likely to be responsible for the regression (because they will also have large differences in average score between their “before” and “after” windows). If multiple pushes in a row had t-test scores over the threshold, the script used to send an alert for the first of those pushes, even if it was not the most likely culprit. Now the script blames the push with the highest t-test score, which is almost always the guilty party. This change had the biggest impact in reducing incorrect regression alerts.

After those changes, there was still one common cause of false alarms that I found. The regression analyzer compares the 12 datapoints before each push to the 12 datapoints after it. But these 12-point moving averages could change suddenly not just at the point where a regression happened, but also at an unrelated point that happens to be 12 pushes earlier or later. This caused spooky action at a distance where a regression in one push would cause a false alarm in a completely different push. To fix this, we now compute weighted averages with “triangular” weighting functions that give more weight to the point being analyzed, and fall off gradually with increasing distance from that point. This smooths out changes at the opposite ends of the moving windows.

There are still occasional errors in regression detection, but as far as I can tell most of them are caused by genuinely misleading random noise or bimodal data. If you see any problems with regression emails, please file a bug (and CC :mbrubeck) and we’ll take a look at it.

Syndicated 2013-11-10 19:53:00 from Matt Brubeck

A good time to try Firefox for Metro

“Firefox for Metro” is our project to build a new Firefox user interface designed for touch-screen devices running Windows 8. (“Metro” was Microsoft’s code name for the new, touch-friendly user interface mode in Windows 8.) I’m part of the small team working on this project.

For the past year we’ve been fairly quiet, partly because the browser has been under heavy construction and not really suitable for regular use. It started as a fork of the old Fennec (mobile Firefox) UI, plus a new port of Gecko’s widget layer to Microsoft’s WinRT API. We spent part of that time ripping out and rebuilding old Fennec features to make them work on Windows 8, and finding and fixing bugs in the new widget code. More recently we’ve been focused on reworking the touch input layer. With a ton of help from the graphics team, we replaced Fennec’s old multi-process JavaScript touch support with a new off-main-thread compositing backend for the Windows Direct3D API, and added WinRT support to the async pan/zoom module that implements touch scrolling and zooming on Firefox OS.

All this work is still underway, but in the past week we finally reached a tipping point where I’m able to use Firefox for Metro for most of my everyday browsing. There are still bugs, and we are still actively working on performance and completing the UI work, but I’m now finding very few cases where I need to switch to another browser because of a problem with Firefox for Metro. If you are using Window 8 (especially on a touch-screen PC) and are the type of brave person who uses Firefox nightly builds, this would be a great time to try Metro-style Firefox and let us know what you think!

Looking to the future, here are some of our remaining development priorities for the first release of Firefox for Metro:

  • Improve the installation and first-run experience, to help users figure out how to use the new UI and switch between “Metro” and desktop modes. (Our UX designer has user testing planned to help identify issues here and throughout the product.)

  • Fix any performance and rendering issues with scrolling and zooming, and add support for double-tap to zoom in on a specific page element.

  • Make the Metro and desktop interfaces share a profile, so they can seamlessly use the same bookmarks and other data without connecting to a Firefox Sync account.

And here are some things that I hope we can spend more time on once that work has shipped:

  • Improve the experience on pages with plugins, which currently require the user to switch to the desktop Firefox interface (bug 936907).

  • Implement a “Reader Mode,” like Firefox for Android. (A pair of students have started working on this project, and their work should also be useful for adding Reader Mode to Firefox for desktop.)

  • Add more features, and more ways to customize and tweak the Metro UI.

If you want to contribute to any of this work, please check out our developer documentation and come chat with us in #windev on irc.mozilla.org or on our project mailing list!

Syndicated 2013-11-10 17:20:00 from Matt Brubeck

Congratulating the IE10 team

Back when Firefox 2 was released (six years ago this week!), the Internet Explorer team started a friendly tradition of sending Mozilla a cake as congratulations. This continued for Firefox 3 and Firefox 4. After Firefox switched from major releases once or twice a year to incremental updates every six weeks, they sent us cupcakes for the next few updates instead. :)

Since IE10 for Windows 8 is shipping today, I thought it would be fun to revive the tradition by delivering a cake to congratulate the IE team. Here’s the cake right after I picked it up from Baked Custom Cakes, with the Firefox logo in painted fondant:

Fellow Mozilla developer Eitan Isaacson drove with my wife Sarah and me to Microsoft Building 50 in Redmond, where program manager Jacob Rossi helped us deliver the cake to a group of IE team members:

The IE team posted their thanks through their official Twitter account. (As you can see from their picture, the bottom border of the cake got sligthly restyled in transit, but it still looks quite edible.) Less than 30 minutes later, Microsoftie Michael Bolan tweeted that cake was already gone. (I hear that the sugary Firefox logo was also eaten shortly afterward.)

So congratulations to the Internet Explorer team on your latest release, and we hope you enjoyed the cake!

Syndicated 2012-10-26 21:38:00 from Matt Brubeck

Metro Firefox without Windows 8

A few weeks ago I started working on the Firefox “Metro UI” project, for Windows 8’s Metro (or Modern) touch-screen environment. While we’re still working on getting our first preview builds ready for Windows 8 users to try out, you can already check out the current source code from the elm branch and build it yourself if you want to get involved and help us fix some bugs.

What you might not know is that you can run “Metro” Firefox even if you don’t have Windows 8. It’s been possible for a while to build and run on older versions of Windows using the -metrodesktop flag. Today I landed a patch to make this work on other platforms too. To build the latest elm source code on Linux or Mac OS X, follow these instructions:

  1. Clone the elm repo: hg clone http://hg.mozilla.org/projects/elm/
    (If you have already cloned mozilla-central or some other repo that shares with it, there’s a faster way to do this.)

  2. Create a .mozconfig file with ac_add_options --enable-metro

  3. Build Firefox as you normally would.

  4. From your objdir, run dist/bin/firefox -metrodesktop (Linux)
    or dist/Nightly.app/Contents/MacOS/firefox -metrodesktop (Mac)

  5. You can visit about:config and enable metro.debug.treatmouseastouch (then restart the browser) to simulate touch interaction with the mouse. Right-click to simulate the Windows 8 edge-swipe gesture, which displays the toolbars.

This is still experimental and mostly untested. Elm might accidentally break on non-Windows platforms from time to time (because of course we are doing all our main development and testing on Windows). While it’s not a perfect replacement for running in the real Windows 8 environment, I hope this is a useful option for adventurous Firefox contributors who want to experiment with the Metro code but don’t have convenient access to Windows 8.

Syndicated 2012-09-20 00:42:00 from Matt Brubeck

Mobile web developers: Your users hate it when you do this

Mobile Firefox beta releases include a “Feedback” add-on (like the one in Firefox 4 beta for desktop), which lets users tell us what they think about the new browser. Based on a sample of feedback from mobile beta testers, the most common complaints are about:

  1. Speed
  2. Fitting text to the screen when zoomed in
  3. Mobile vs. desktop versions of web sites

The first two are straightforward, though not necessarily easy. We’re always working on performance, and we have experimental text reflow code (currently available in the Easy Reading add-on). But the last item is more complicated…

Browser detection pitfalls

Web sites can read the User-Agent header sent by your browser to see what browser and OS you are using. Some sites use that information to decide whether to send a “full” version of a web page, or a version formatted for mobile devices.

This can go wrong in several ways. If your browser or device is new, or wasn’t tested when a site was developed, that site has no way of knowing whether it is “mobile.” Users may also change their User-Agent to work around content restrictions or access different media formats. And some sites make incorrect assumptions, like that all browsers with “Android” in their User-Agent string are based on WebKit.

Even when the browser is known, readers and publishers might not agree about whether the mobile or desktop version is better. Based on our feedback, some users want to switch from full sites to mobile sites while others want just the opposite. And some devices, like large touch-screen tablets, combine aspects of handheld and desktop computers.

Solutions

Looking through these complaints, many people are under the mistaken impression that the browser, rather than the web site, decides whether to display mobile-formatted pages. Even the New York Times' David Pogue gets this wrong in his Galaxy Tab review:

When you visit sites like nytimes.com, CNBC.com and Amazon.com, the Galaxy’s browser shows the stripped-down, mobile versions of those sites. According to Samsung, there’s no way to turn that feature off and no way to visit the full-size sites. You can delete the little “m.” in the Web address until you’re blue in the browser, but the Galaxy always puts it right back.

Web developers: your readers are begging us to display your content in their preferred format. We want to help them, but we can’t do it alone.

(I wrote an add-on called Phony that lets mobile Firefox impersonate the User-Agent strings of other browsers. While this improves the experience on some sites, it breaks it on others. Masquerading as another browser can lead sites to serve non-standard markup that do not work in Firefox.)

Because browser detection is never perfect, web sites should let readers choose between mobile and full content. They can try to guess the right version by default, but please let users opt in or out.

Best practices for web developers

Here are some first steps typical mobile web sites can take to make their readers happier:

  • When possible, serve the same content to all browsers. You can use stylesheets and scripts to customize your layout for different display sizes, as in this beautiful site by Jon Hicks.

  • There are valid reasons to use User-Agent sniffing. But if you must use it, test in as many browsers and devices as possible and learn the correct way to detect various browsers. For example, you can detect Gecko-based browsers by looking for Gecko and rv:, and you can detect mobile Firefox by looking for Fennec/.

  • If a “mobile” user requests a page that isn’t available on your mobile site, serve the full version to them anyway, rather than redirecting them to an unrelated mobile landing page.

  • Let users switch from your mobile site to your full site and back. You may remember users' previous choices for convenience, but let them change their minds again.

Further reading

For much more comprehensive development advice, see Yiibu’s thoughtful and practical approach to building sites that work across many different browsers and mobile devices.

Coming from a different perspective, Andrea Trasatti (developer of the device-detection library WURFL) talks about problems in mobile User-Agent strings and how they could be more useful for device detection.

Syndicated 2010-11-19 15:15:00 from Matt Brubeck

125 older entries...

New Advogato Features

New HTML Parser: The long-awaited libxml2 based HTML parser code is live. It needs further work but already handles most markup better than the original parser.

Keep up with the latest Advogato features by reading the Advogato status blog.

If you're a C programmer with some spare time, take a look at the mod_virgule project page and help us with one of the tasks on the ToDo list!