This is an ad-hoc implementation that resolves the ready() promise once
the document and all fonts collected by the style computer are done
loading. A spec-compliant implementation would include creating a proxy
CSS::FontFace for each @font-face and correctly implementing the
specification steps for font fetching, but we are far from there yet.
This hackish implementation should yield good WPT progress because it
will actually start waiting for the Ahem font to load before capturing
layout measurements. For example, it makes
https://wpt.live/css/css-grid/abspos/positioned-grid-descendants-001.html
go from 0/100 to 36/100 passing subtests.
`revert` is supposed to revert to the previous cascade origin, but we
previously had it reverting to the previous layer. To support both,
track them separately during the cascade.
As part of this, we make `set_property_expanding_shorthands()` fall back
to `initial` if it can't find a previous value to revert to. Previously
we would just shrug and do nothing if that happened, which only works
if the value you want to revert to is whatever is currently in `style`.
That's no longer the case, because `revert` should skip over any layer
styles that have been applied since the previous origin.
As useful as they may be to web developers, :has() selectors complicate
the style invalidation process quite a lot.
Let's have StyleComputer keep track of whether they are present at all
in the current set of active style sheets. This will allow us to
implement fast-path optimizations when there are no :has() selectors.
The web server for WPT has a tendency to just disconnect after sending
us a resource. This makes curl think an error occurred, but it's
actually still recoverable and we have the data.
So instead of just bailing, do what we already do for other kinds of
resources and try to parse the data we got. If it works out, great!
It would be nice to solve this in the networking layer instead, but
I'll leave that as an exercise for our future selves.
Instead of throwing all pseudo-element rules in one bucket, let's have
one bucket per pseudo-element.
This means we only run ::before rules for ::before pseudo-elements,
only ::after rules for ::after, etc.
Average style update time on https://tailwindcss.com/ 250ms -> 215ms.
By filtering first, we end up allocating much less vector space
most of the time.
This is mostly helpful in pathological cases where there's a huge number
of rules present, but most of them get rejected early.
We were saving to source declarations for *every* property, even though
we only ever looked it up for animation-name.
This patch gets rid of the per-property source pointer and we now keep
a single pointer to the animation-name source only.
This shrinks StyleProperties from 6512 bytes to 4368 bytes per instance.
Typeface is a more widely used name for the data represented by
class previously named VectorFont.
Now:
- Typeface represents decoded font that is not ready for rendering
- ScaledFont represents the combination of typeface and size for
rendering
We'll want to explicitly load fonts from FontFace and other Web APIs
in the future. A future refactor should also move this completely away
from StyleComputer and call it something like 'FontCache'.
Given a selector like `.foo .bar #baz`, we know that elements with
the class names `foo` and `bar` must be present in the ancestor chain of
the candidate element, or the selector cannot match.
By keeping track of the current ancestor chain during style computation,
and which strings are used in tag names and attribute names, we can do
a quick check before evaluating the selector itself, to see if all the
required ancestors are present.
The way this works:
1. CSS::Selector now has a cache of up to 8 strings that must be present
in the ancestor chain of a matching element. Note that we actually
store string *hashes*, not the strings themselves.
2. When Document performs a recursive style update, we now push and pop
elements to the ancestor chain stack as they are entered and exited.
3. When entering/exiting an ancestor, StyleComputer collects all the
relevant string hashes from that ancestor element and updates a
counting bloom filter.
4. Before evaluating a selector, we first check if any of the hashes
required by the selector are definitely missing from the ancestor
filter. If so, it cannot be a match, and we reject it immediately.
5. Otherwise, we carry on and evaluate the selector as usual.
I originally tried doing this with a HashMap, but we ended up losing
a huge chunk of the time saved to HashMap instead. As it turns out,
a simple counting bloom filter is way better at handling this.
The cost is a flat 8KB per StyleComputer, and since it's a bloom filter,
false positives are a thing.
This is extremely efficient, and allows us to quickly reject the
majority of selectors on many huge websites.
Some example rejection rates:
- https://amazon.com: 77%
- https://github.com/SerenityOS/serenity: 61%
- https://nytimes.com: 57%
- https://store.steampowered.com: 55%
- https://en.wikipedia.org: 45%
- https://youtube.com: 32%
- https://shopify.com: 25%
This also yields a chunky 37% speedup on StyleBench. :^)
When iterating through a @keyframes rule, it isn't possible to resolve
unresolved style properties since there are no elements. This change
allows those properties to simply pass through this helper function.
If we determine that a selector is simple enough, we now run it using a
special matching loop that traverses up the DOM ancestor chain without
recursion.
The criteria for this fast path are:
- All combinators involved must be either descendant or child.
- Only tag name, class, ID and attribute selectors allowed.
It's definitely possible to increase the coverage of this fast path,
but this first version already provides a substantial reduction in time
spent evaluating selectors.
48% of the selectors evaluated when loading our GitHub repo are now
using this fast path.
18% speed-up on the "Descendant and child combinators" subtest of
StyleBench. :^)
This allows us to skip evaluating selectors like "[foo=bar]" for any
element that doesn't have a "foo" attribute.
Note that the bucket is case-insensitively keyed on the attribute name
since case sensitivity is depending on evaluation context. This ensures
we may get some false positives but no false negatives.
Reduces the number of selectors evaluated by 36% when loading our GitHub
repo at https://github.com/SerenityOS/serenity
Patch up existing style properties instead of using the regular style
invalidation path, which requires rule matching for each element in the
invalidated subtree.
- !important properties: this change introduces a flag used to skip the
update of animated properties overridden by !important.
- inherited animated properties: for now, these are invalidated by
traversing animated element's subtree to propagate the update.
- StyleProperties has a separate array for animated properties that
allows the removal animated properties after animation has ended,
without requiring full style invalidation.
If a selector must match a pseudo element, or must match the root
element, we now cache that information in the MatchingRule struct.
We also introduce separate buckets for these rules, so we can avoid
running them altogether if the current element can't possibly match.
This cuts the number of selectors evaluated by 32% when loading our
GitHub repo page https://github.com/SerenityOS/serenity
Now, if an element belongs to a shadow tree, we use only the style
sheets from the corresponding shadow root during style computation,
instead of using all available style sheets as was the case
previously.
The only exception is the user agent style sheets, which are still
taken into account for all elements.
Tests/LibWeb/Layout/input/input-element-with-display-inline.html
is affected because style of document no longer affects shadow tree
of input element, like it is supposed to be.
Co-authored-by: Simon Wanner <simon+git@skyrising.xyz>
All of this error propogation came from a single call to
HashMap::try_ensure_capacity! As part of the ongoing effort to ignore
small allocation failures, lets just assert this works. This has the
nice side-effect of propogating out to a few other classes.
With this commit, we are finally running animations off of the web
animations spec! A lot of the work StyleComputer is doing is now done
elsewhere. For example, fill-forward animations are handled by
Animation::is_relevant() returning true in the after phase, meaning the
"active_state_if_fill_forward" map is no longer needed.
Implements following rule from CSS Overflow Module Level 3:
"The visible/clip values of overflow compute to auto/hidden
(respectively) if one of overflow-x or overflow-y is neither visible
nor clip."
This patch makes a few changes to the way we calculate line-height:
- `line-height: normal` is now resolved using metrics from the used
font (specifically, round(A + D + lineGap)).
- `line-height: calc(...)` is now resolved at style compute time.
- `line-height` values are now absolutized at style compute time.
As a consequence of the above, we no longer need to walk the DOM
ancestor chain looking for line-heights during style computation.
Instead, values are inherited, resolved and absolutized locally.
This is not only much faster, but also makes our line-height metrics
match those of other engines like Gecko and Blink.
Fetching the viewport rect is currently somewhat expensive, since it
requires finding the navigable the document is active in.
We can avoid the cost of repeated calls by simply allowing StyleComputer
to cache the viewport rect at the start of style computation.
Before this change, we would only cache and reuse Gfx::ScaledFont
instances for downloaded CSS fonts.
By moving it into Gfx::VectorFont, we get caching for all vector fonts,
including local system TTFs etc.
This avoids a *lot* of style invalidations in LibWeb, since we now vend
the same Gfx::Font pointer for the same font when used repeatedly.
No functional impact intended. This is just a more complicated way of
writing what we have now.
The goal of this commit is so that we are able to store the 'name' of a
pseudo element for use in serializing 'unknown -webkit-
pseudo-elements', see:
https://www.w3.org/TR/selectors-4/#compat
This is quite awkward, as in pretty much all cases just the selector
type enum is enough, but we will need to cache the name for serializing
these unknown selectors. I can't figure out any reason why we would need
this name anywhere else in the engine, so pretty much everywhere is
still just passing around this raw enum. But this change will allow us
to easily store the name inside of this new struct for when it is needed
for serialization, once those webkit unknown elements are supported by
our engine.
According to the CSS font matching algorithm specification, it is
supposed to be executed for each glyph instead of each text run, as is
currently done. This change partially implements this by having the
font matching algorithm produce a list of fonts against which each
glyph will be tested to find its suitable font.
Now, it becomes possible to have per-glyph fallback fonts: if the
needed glyph is not present in a font, we can check the subsequent
fonts in the list.
This one is a bit fun because it can be `add(<integer>)` or `auto-add`,
but children have to inherit the computed value not the specified one.
We also have to compute it before computing the font-size, because of
`font-size: math` which will be implemented later.
Resolving typed `attr()` functions is going to involve using more
internal Parser methods, so this is the simplest solution for that.
Also... resolving these is basically parsing them, so it makes more
sense for that process to live here.
This is just moving code, with minimal changes so it still works.
This means StyleComputer::resolve_unresolved_style_value() always
returns a value, so we can change its return type.
However, it does still return an UnresolvedStyleValue sometimes, so we
can't remove those checks from the user code.
User styles are applied after the UserAgent's built-in styles, and
before the Author styles that are part of the web page.
Because they're neither part of the page, but can still be modified
while the page is open, caching is a little tricky. The approach here
is to piggy-back on the StyleComputer's rule caches, which already get
rebuilt whenever the styles change. This is not the smartest approach,
since it means re-parsing the style sheet more often than is necessary,
but it's simple and works. :^)
This effectively makes it per-Document, but we hang it off of
StyleComputer since that's what it's used for.
The purpose of this is to prevent downloaded fonts from escaping the
context that loaded them. There's probably a more elegant solution where
we still share caching of system fonts, but let's start here.
Instead, perform the filtering for each rule as we go. This avoids
creating a separate list of rules, which was ~5% of runtime when
mousing around on the Discord web interface.
