Implements the corresponding encoders, selects the appropriate one when
encoding URL search params. If an encoder for the given encoding could
not be found, fallback to utf-8.
`m_needs_repaint = true` is not enough because it doesn't schedule
repaint of a parent navigable.
Fixes the bug when an iframe is not repainted after scrolling.
When converting a `Gfx::Bitmap` to a Skia bitmap, we cannot assume the
color data is unpremultiplied. For example, everything canvas-related
uses premultiplied color data:
https://html.spec.whatwg.org/multipage/canvas.html#premultiplied-alpha-and-the-2d-rendering-context
We were probably assuming unpremultiplied since that is what the PNG
decoder gives us. Since we now make `Gfx::Bitmap` identify what alpha
type is being used, we can instruct Skia a bit better :^)
Update our `EdgeFlagPathRasterizer` to use premultiplied alpha instead
of unpremultiplied so we can apply alpha correctly for path masks.
This fixes the dark borders sometimes visible when SVGs are blended
with a colored background.
This also exposed an issue with our `CanvasRenderingContext2D`, which is
supposed to hold a bitmap with premultiplied alpha internally but expose
a bitmap with unpremultiplied alpha in `CanvasImageData`. Expand our C2D
test to include the alpha channel as well.
Finally, this also exposed an off-by-one issue in
`EdgeFlagPathRasterizer` which caused the last scanlines for edges to
render incorrectly. We had some reference images which included these
corruptions (they were almost unnoticeable), so update them as well.
We use instances of `Gfx::Bitmap` to move pixel data all the way from
raw image bytes up to the Skia renderer. A vital piece of information
for correct blending of bitmaps is the alpha type, i.e. are we dealing
with premultiplied or unpremultiplied color values?
Premultiplied means that the RGB colors have been multiplied with the
associated alpha value, i.e. RGB(255, 255, 255) with an alpha of 2% is
stored as RGBA(5, 5, 5, 2%).
Unpremultiplied means that the original RGB colors are stored,
regardless of the alpha value. I.e. RGB(255, 255, 255) with an alpha of
2% is stored as RGBA(255, 255, 255, 2%).
It is important to know how the color data is stored in a
`Gfx::Bitmap`, because correct blending depends on knowing the alpha
type: premultiplied blending uses `S + (1 - A) * D`, while
unpremultiplied blending uses `A * S + (1 - A) * D`.
This adds the alpha type information to `Gfx::Bitmap` across the board.
It isn't used anywhere yet.
Knowing whether pixel color data is premultiplied or unpremultiplied is
useful in more places; factor it out so we can use the enum elsewhere.
Implement support for premultiplied alpha in `Color::with_alpha()` and
`::set_alpha()`.
If a DOM node is an element with pseudo-elements, but it has no child
DOM nodes and is not a shadow host, then the code that serializes its
pseudo-elements would get skipped, making them not show up in the
inspector.
Right now, we deviate from the CSSOM spec regarding our
CSSStyleDeclaration classes, so this is not as close to the spec as I'd
like. But it works, which means we'll be able to test pseudo-element
styling a lot more easily. :^)
As noted, this is hacky because the parser wasn't written to allow
parsing an individual component of a selector. (Fox example, the
convenient-sounding `parse_pseudo_simple_selector()` assumes the first
colon has already been consumed...) So until that changes, this parses
the input as an entire selector-list, and then throws it away if it's
not a single pseudo-element selector.
It's only temporary though, I promise. 😅
The script pulls in a dependency on the `yaml` python package. Instead
of updating all the docs and CI jobs to account for this, let's guard
calling the script behind our experimental flag instead.
Parsing last as an IPV4 number was not returning true in "ends with a
number" as the parsing of that part was overflowing. This means that the
URL is not considered to be an IPv4 address, and is treated as a valid
domain.
Helpfully, the spec also points out in a note that this step is
equivalent to simply checking that the last part ends with 0x followed
by only hex digits - which doesn't suffer from any overflow problem!
Arguably this is an editorial issue in the spec where this should be
clarified a little bit. But for now, fixing this fixes 3 sub tests in
WPT for:
https://wpt.live/url/url-constructor.any.html
Our heuristic was a bit too simplistic and would not run through the
ToASCII unicode algorithm which performs some extra validation. This
would cause invalid URLs that should fail to be parsed be mistakenly
accepted.
This fixes 8 tests in: https://wpt.live/url/url-constructor.any.html
We can't simply use the base URL as it may need to be modified in some
form. For example - for the included test, the fragment was previously
being included in the resulting URL.
This fixes 1 test on https://wpt.live/url/url-constructor.any.html
Instead of carrying the display list for a mask in each command that
might potentially be affected by "background-clip: text", this change
introduces a new AddMask command that is applied once for all
background layers within one box.
The new AddMask command includes a rectangle for the mask destination
that is translated by the corresponding scroll offset.
Fixes https://github.com/LadybirdBrowser/ladybird/issues/857
The spec didn't match how other browsers behave, and we dutifully did
what the spec said. A spec bug has been filed, so let's fix this locally
for now with a FIXME.
add_clip_rect() accepts a rectangle in viewport-relative coordinates,
so it must be translated by the enclosing scroll offset to be displayed
correctly inside a scrollable box.
This change fixes layering violation by moving to_px() calls to happen
before display list recording. Also it should make display list
recording a bit faster by resolving background properties beforehand.
This was a goof in cc55732332 which
resulted in the URL path getting double percent encoded. Since the path
already comes out percent encoded following the rules in the URL spec -
we don't need to percent encode again.
Fixes: #978
There were some extra steps in there which produced wrong results for
relative file URLs.
Fixes 7 test cases in: https://wpt.live/url/url-constructor.any.html
We also need to adjust the test results in TestURL. The behaviour tested
does not match how URL is specified to work as an absolute relative is
given.
The check for:
```
if (start_index >= end_index)
return {};
```
To prevent an out of bounds when trimming the start and end of the input
of whitespace was preventing valid URLs (only having whitespace in the
input) from being parsed.
Instead, prevent start_index from ever getting above end_index in the
first place, and don't treat empty inputs as an error.
Fixes one WPT test on:
https://wpt.live/url/url-constructor.any.html
This doesn't seem like something we will neccessarily do in the URL
class anyway due to performance reasons - unless strictly needed (like
for the DOMURL implementation).
I previously believed there was no way a detached document should have
layout information, but it turns out there is a way: getComputedStyle().
So we need to account for cases where we have a layout node, but no
navigable, since that is a state we can get into at this moment.
Fixes#354
For the SVG <use> element, we want to support loading HTML documents
that have a SVG element inside of it pointed to by the URL fragment.
In this situation we would need to fetch and parse the entire document
in SharedImageRequest (so that we can still cache the SVGs). Rename
SharedImageRequest to SharedResourceRequest to make the class a little
more generic for future usecases.
This closer mirrors handle_failed_fetch, making the handling slightly
more clear to understand. It also means that we don't need to take a
strong reference to this on a successful SVG resource load.
...When loading images through SharedImageRequest.
There is a small behavioural difference here - handle_failed_fetch
clears the pending callbacks whereas handled_failed_decode was
previously not. This does not seem intentional, and appears like a bug.
Implementing it this way is _slightly_ simpler - and also means we
don't need to take a strong handle to this in the case of loading an
SVG image.
Web specs do not return through javascript percent decoded URL path
components - but we were doing this in a number of places due to the
default behaviour of URL::serialize_path.
Since percent encoded URL paths may not contain valid UTF-8 - this was
resulting in us crashing in these places.
For example - on an HTMLAnchorElement when retrieving the pathname for
the URL of:
http://ladybird.org/foo%C2%91%91
To fix this make the URL class only return the percent encoded
serialized path, matching the URL spec. When the decoded path is
required instead explicitly call URL::percent_decode.
This fixes a crash running WPT URL tests for the anchor element on:
https://wpt.live/url/a-element.html
Previously, calling `setProperty` or `removeProperty` from JS on a
CSSStyleDeclaration returned from `getComputedStyle()` would return
null. We now return a NoModificationAllowedError instead, which aligns
our implementation with the specification.
This also fixes a bug where task IDs were being deallocated from the
wrong IDAllocator. I don't know if it was actually possible to cause any
real trouble with that mistake, nor do I know how to write a test for
it, but this makes the bug go away.
Not every value in a StyleProperties will be non-null by the time we
perform `revert`, so let's make a specialized function for reverting a
property instead of using the path that requires the value to be
non-null.
The "text-overflow" property affects text that may get clipped if it is
larger than its container and does not do any line breaks.
The ellipsis character gets added to the end and the rest of the text
gets trunctated if the property is set to "ellipsis".
This patch implements this behavior in the InlineFormattingContext. :^)
The "text-overflow" property is also added to the
getComputedStyle-print-all test.
To avoid expensive lookups, we now cache a weak pointer from document to
the last known node navigable. Before using the cache, we validate that
the document is still the navigable's active document.
This patch moves the data members of URL to an internal URL::Data struct
that is also reference-counted. URL then uses a CopyOnWrite<T> template
to give itself copy-on-write behavior.
This means that URL itself is now 8 bytes per instance, and copying is
cheap as long as you don't mutate.
This shrinks many data structures over in LibWeb land. As an example,
CSS::ComputedValues goes from 3024 bytes to 2288 bytes per instance.
This shrinks each instance of PercentageOr by 8 bytes and avoids virtual
dispatch when resolving calc() values. It's a win-win!
Many data structures shrink as a result. An example is ComputedValues
which goes from 3376 bytes to 3024 bytes per instance.
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.
Navigables are re-used for navigations within the same tab. Its current
ownership of the cursor position is a bit ad-hoc, so nothing in the spec
indicates when to reset the cursor, nor do we manually do so. So when a
cursor update happens on one page, that cursor is retained on the next
page.
Instead, let's have the document own the cursor. Each navigation results
in a new document, thus we don't need to worry about resetting cursors.
This also makes many of the callsites feel nicer. We were previously
often going from the node, to the document, to the navigable, to the
cursor. This patch removes the navigable hop.
Adds additional padding to the end-side of the scrollable overflow
rectangle as necessary to enable a scroll position that satisfies
the requirements of `place-content: end` alignment.
The spec does not define activation behavior of ctrl/cmd clicks, so we
have to go a bit ad-hoc here. When an anchor element is clicked with the
cmd (on macOS) or ctrl (on other platforms) modifier pressed, we will
skip activation of that element and pass the event on to the chrome. We
still dispatch the event to allow scripts to cancel the event.
ICU 72 began using non-ASCII spaces in some formatted date-time strings.
Every major browser has found that this introduced major breakage in web
compatibility, as many sites and tools expect ASCII spaces. This patch
removes these non-ASCII spaces in the same manner as the major engines.
Such behavior is also tested by WPT.
This removes the --enable-callgrind-profiling flag, and replaces it with
a --profile-process=<process-name> flag. For example:
ladybird --profile-process=WebContent
ladybird --profile-process=RequestServer
This allows profiling any helper process with callgrind.
This removes the --debug-web-content flag, and replaces it with a
--debug-process=<process-name> flag. For example:
ladybird --debug-process=WebContent
ladybird --debug-process=RequestServer
This allows attaching gdb to any helper process.
Overflow clipping is currently implemented as:
1. Create clip frame for each box with hidden overflow
2. Calculate clip rect for each clip frame by intersecting padding boxes
of all boxes with hidden overflow in containing block chain
3. Assign enclosing clip frame (closest clip frame in containing block
chain) to each PaintableBox
4. Apply clip rect of enclosing clip frame in Paintable::before_paint()
It breaks when any CSS transform other than simple translation is lying
between box with hidden overflow and a clipped box, because clip
rectangle will be applied when transform has already changed.
The fix is implemented by relying on the following rule:
"For elements whose layout is governed by the CSS box model, any value
other than none for the transform also causes the element to establish
a containing block for all descendants."
It means everything nested into a stacking context with CSS transform
can't escape its clip, so it's safe to apply its clip for all children.
Adds missing clip of border-radius for clip frame. While we have already
accounted for the border-radius of containing blocks, the box itself was
not being clipped.
Remove `for_each_section_of_type` in favor of making the module's
sections defined as distinct fields. This means it is no longer possible
to have two of the same section (which is invalid in WebAssembly, for
anything other than custom sections).
Now that the implementation is in FormAssociatedElement, the
implementation in HTMLInputElement is effectively just a passthrough,
with some minor differences to handle small behavioural quirks between
the two (such as the difference in nullability of types).
The user is not required to keep the object alive, this commit makes it
so the lifetime of these functions is extended to match the Wasm module
it is imported into.
Fixes the crash in #907.
Currently, if we want to add a new e.g. WebContent command line option,
we have to add it to all of Qt, AppKit, and headless-browser. (Or worse,
we only add it to one of these, and we have feature disparity).
To prevent this, this moves command line flags to WebView::Application.
The flags are assigned to ChromeOptions and WebContentOptions structs.
Each chrome can still add its platform-specific options; for example,
the Qt chrome has a flag to enable Qt networking.
There should be no behavior change here, other than that AppKit will now
support command line flags that were previously only supported by Qt.
We don't want to set the intrinsic Console object's client to non-top-
level clients, created for e.g. subframes. We also want to make sure the
Console client is updated if the top-level document has changed.
We explicitly stopped visting the map of documents to console clients in
commit 44659f2f2a to avoid keeping the
document alive. However, if nothing else visits the console clients, we
may set the top-level console client to a client that has been garbage
collected.
So instead of storing this map, just store the console client on the
document itself. This will allow the document to visit its client.
Moves paint_table_borders() call into PaintableBox::paint() to make
scroll offset and clip rectangle of enclosing scrollable be applied
in ::before_paint().
Previously, pseudo-elements had their style computed while the layout
tree was being built. Instead, do so inside Element::recompute_style(),
using the same invalidation mechanism that the element itself uses.
This also has the effect of invalidating the layout much less often.
Pseudo-elements' style is only computed while building the layout tree.
This meant that previously, they would not have their style recomputed
in some cases. (Such as when :hover is applied to an ancestor.)
Now, when recomputing an element's style, we also return a full
invalidation if one or more pseudo-elements would exist either before or
after style recomputation.
This heuristic produces some false positives, but no false negatives.
Because pseudo-elements' style is computed during layout building, any
computation done here is then thrown away. So this approach minimises
the amount of wasted style computation. Plus it's simple, until we have
data on what approach would be faster.
This fixes the Acid2 nose becoming blue when the .nose div is hovered.
Contrary to LibGfx, where corner clipping was implemented by sampling
and blitting pixels under corners into a temporary bitmap, Skia allows
us to simply apply a mask. As a result, we no longer need the
BlitCornerClipping command, which has become a no-op.
- SampleUnderCorners is renamed to AddRoundedRectClip
- The optimization that skipped unnecessary blit and sample commands has
been removed. However, this should not result in a performance
regression because Skia seems to perform mask rasterization lazily.
In the HTML parser spec, there are 2 instances of the following text:
add the attribute and its corresponding value to that element
The "add the attribute" text does not have a corresponding spec link to
actually specify what to do. We currently use `set_attribute`, which can
throw an exception if the attribute name contains an invalid character
(such as '<'). Instead, switch to `append_attribute`, which allows such
attribute names. This behavior matches Firefox.
Note we cannot yet make the unclosed-html-element.html test match the
expectations of the unclosed-body-element.html due to another bug that
would prevent checking if the expected element has the right attribute.
That will be fixed in an upcoming commit.
That usually happens in "exceptional" states when the client exits
unexpectedly (crash, force quit mid-load, etc), leading to the job flush
timer firing and attempting to write to a nonexistent object (the
client).
This commit makes RS simply cancel such jobs; cancelled jobs in this
state simply go away instead of sending notifications around.
`Module::functions` created clones of all of the functions in the
module. It provided a _slightly_ better API, but ended up costing around
40ms when instantiating spidermonkey.
This code previously only allocated enough space in
m_col_elements_by_index for 1 slot per column, meaning that columns
with a span > 1 would write off the end of it.
These have a few rules that we didn't follow in most cases:
- CSS-wide keywords are not allowed. (inherit, initial, etc)
- `default` is not allowed.
- The above and any other disallowed identifiers must be tested
case-insensitively.
This introduces a `parse_custom_ident_value()` method, which takes a
list of disallowed identifier names, and handles the above rules.
The main incentive is much better performance. We could have gone a bit
further in optimizing the Skia painter to blit glyphs produced by LibGfx
more efficiently from the glyph atlas, but eventually, we also want Skia
to improve correctness.
This change does not completely replace LibGfx in text handling. It's
still used at all stages, including layout, up until display list
replaying.
While conceptually is_supported_property_index is a cheap index lookup,
it is not currently cheap for an container such as HTMLAllCollection
that is operating on an uncached collection. Instead - just do the
lookup once. It also happens to look a little nicer to not blindly
dereference an optional.
This uses a faster hashtable lookup in the case of HTMLCollection.
Also port invoke_named_property_setter to FlyString to avoid a
FlyString->String->FlyString conversion that surfaces from this change.
This removes some ambiguity about what the return value should be if
the index is out of range.
Previously, we would sometimes return a JS null, and other times a JS
undefined.
It will also let us fold together the checks for whether an index is a
supported property index, followed by getting the value just afterwards.
This is `counter(name, style?)` or `counters(name, link, style?)`. The
difference being, `counter()` matches only the nearest level (eg, "1"),
and `counters()` combines all the levels in the tree (eg, "3.4.1").
These control the state of CSS counters.
Parsing code for `reversed(counter-name)` is implemented, but disabled
for now until we are able to resolve values for those.
The new method is Parser::parse_all_as_single_none_value(), which has a
few advantages:
1. There's no need for user code to manually create a StyleValue.
2. It consumes tokens so that doesn't have to be done manually.
3. Whitespace before or after the `none` is consumed correctly.
It does mean we create and then discard a `none` StyleValue in a couple
of places, namely parsing for `grid-*` properties. We may or may not
want to migrate those to returning the IdentifierStyleValue instead.
This seems to have been required when pseudo-elements were first
implemented, but has since become unused. It's also awkward because we
don't have access to the DOM Element and its CountersSet at this point.
So, let's remove it.
For reference, Chrome&Firefox both return the computed value for
`content: counter(foo)` as `counter(foo)`, not as the computed string.
So not computing it here seems like the intended behaviour.
This represents each element's set of CSS counters.
https://drafts.csswg.org/css-lists-3/#css-counters-set
Counters are resolved while building the tree. Most elements will not
have any counters to keep track of, so as an optimization, we don't
create a CountersSet object until the element actually needs one.
In order to properly support counters on pseudo-elements, the
CountersSet needs to go somewhere else. However, my experiments with
placing it on the Layout::Node kept hitting a wall. For now, this is
fairly simple at least.
We now follow the rules from the spec more closely, along with an
unspecified quirk for when the offsetParent is a non-positioned body
element. (Spec bug linked in a comment.)
This fixes a whole bunch of css-flexbox tests on WPT, which already had
correct layout, but the reported metrics from JS API were wrong.
We now ensure that `Node::is_character_data()` returns true for all
nodes of type character data.
Previously, calling `Node::length()` on `CDataSection` or
`ProcessingInstruction` nodes would return an incorrect value.
We were mistakenly executing the current node's script instead of the
document's pending parsing-blocking script.
This caused ~1000 WPT tests to time out, since we never ended up firing
a load event for XHTML pages that load multiple external scripts.
Before this change, "background-clip: text" was implemented by saving a
Vector<Gfx::Path> of all glyphs needed to paint a mask for the
background. The issue with this approach was that once glyphs were
extracted into vector paths, the glyph rasterization cache could no
longer be utilized.
With this change, all text required for mask painting is saved in a
nested display list and rasterized as a regular text.
This change is a preparation for the upcoming changes where display
list will be nested and the same display could be owned by multiple
display list items.
With this change display list player will be able to recurse into
executing another display list, without having to construct new display
list player. It is going to be useful in the upcoming changes to paint
a mask from a display list owned by a command.
Previously, when creating a HTML element with
`document.createElementNS()` we would convert the given local name to
lowercase before deciding which element type to return. We now no
longer perform this lower case conversion, so if an uppercase local
name is provided, an element of type `HTMLUnknownElement` will be
returned. This aligns our implementation with the specification.
Instead of multiple loops and multiple vectors, parse Wasm expressions
in a simple loop. This gets us from ~450ms to instantiate spidermonkey
to ~280ms.
`swizzle` had the wrong operands, and the vector masking boolean logic
was incorrect in the internal `shuffle_or_0` implementation. `shuffle`
was previously implemented as a dynamic swizzle, when it uses an
immediate operand for lane indices in the spec.
CommandResult was needed by LibGfx display list player that could have
failed to allocate a temporary bitmap for painting a stacking context
with CSS transforms. This is no longer an issue with Skia painter, so
we can delete CommandResult::SkipStackingContext handling path.
This is called by https://athenacrisis.com/ and passed through to
AudioNode.connect, which expects an AudioNode.
Implement this function enough so that we return an AudioNode so that
AudioNode.connect does not throw a TypeError.
This is an AudioNode representing the final audio destination and is
what the user will ultimately hear.
This node is used as one of the connecting nodes in athenacrisis.com
Add a placeholder for the interface without anything backing it for now.
With this change, instead of recording a display list item for each
instance of a repeated background, a new DrawRepeatedImmutableBitmap
type is used. This allows the painter to use optimized repeated image
painting and, when the GPU backend is used, avoid re-uploading the image
texture for each repetition.
Some screenshot tests are affected, but there are no visible
regressions.
https://null.com/games/chainstaff works a lof faster with this change.
The :host family of pseudo class selectors select the shadow host
element when matching against a rule from within the element's shadow
tree.
This is a bit convoluted due to the fact that the document-level
StyleComputer keeps track of *all* style rules, and not just the
document-level ones.
In the future, we should refactor style storage so that shadow roots
have their own style scope, and we can simplify a lot of this.
This fixes an issue where :host(foo) would parse as if "foo" was the
on the right side of a descendant combinator.
Not testable yet, but will be in the next commit.
We don't want to set the intrinsic Console object's client to non-top-
level clients, created for e.g. SVG elements or subframes. We also want
to make sure the Console client is updated if the top-level document has
changed.
Checking that the string parsed for the `font` property is not enough,
the spec also wants to rule out CSS-wide keywords like `inherit`. The
simplest way to do so is to check if it's a ShorthandStyleValue, which
also rules out use of `var()`; this matches other browsers' behaviour.
The newly-added test would previously crash, and now doesn't. :^)
The typeof operator has a very small set of possible resulting strings,
so let's make it much faster by caching those strings on the VM.
~8x speed-up on this microbenchmark:
for (let i = 0; i < 10_000_000; ++i) {
typeof i;
}
Before this change, removing a style element from inside a shadow tree
would cause it to be unregistered with the document-level list of sheets
instead of the shadow-root-level list.
This would eventually lead to a verification failure if someone tried to
update the text contents of that style element, since it was still in
the shadow-root-level list, but now with a null owner element.
Fixes a crash on https://www.swedbank.se/
The default limit (at least on Linux) causes us to run out of file
descriptors at around 15 tabs. Increase this limit to 8k. This is a
rather arbitrary number, but matches the limit set by Chrome.
Having resolution of all properties for all paintable types in a single
function was hard to iterate on, so this change separates it into
smaller functions per paintable type.
Previously, if a document had any element with a name attribute that
was set to the empty string, then `document.getElementsByName("")` and
`element.getElementsByName("")` would return a collection including
those elements.
Previously, if a document had an element whose id was the empty string,
then `document.getElementById("")` and `element.getElementById("")`
would return that element.
This change removes wrappers inherited from Gfx::Typeface for WOFF and
WOFF2 fonts. The only purpose they served is owning of ttf ByteBuffer
produced by decoding a WOFF/WOFF2 font. Now new FontData class is
responsible for holding ByteBuffer when a font is constructed from
non-externally owned memory.
It currently doesn't support animated image.
Note that Gfx::Bitmap has no support for get_pixel when the format is
RGBA8888. This is why it has been removed from the tests.
Previously, `SVGSVGBox` would have a natural aspect ratio of 0 if it
had a viewbox with zero width. This led to a division by zero, causing
a crash.
Found by Domato.
Previously calling `PaintableBox::set_scroll_offset()` with a
PaintableBox whose content size was larger than its scrollble overflow
rect would cause a crash.
Found by Domato.
We had a const and non-const version of this function, with slightly
different behavior (oops!)
This patch consolidates the implementations and keeps only the correct
behavior in there.
Fixes an issue where comments were not collapsible on Hacker News.
Skia painter is visibly faster than LibGfx painter and has more complete
CSS transforms support. With this change:
- On Linux, it will try to use Vulkan-backend with fallback to
CPU-backend
- On macOS it will try to use Metal-backend with fallback to
CPU-backend
- headless-browser always runs with CPU-backend in layout mode
LibGfx's output is consistent across different platforms, which allows
us to have one set of expectations for screenshot tests. This
consistency will not hold for Skia, where features like antialiasing and
gradient color interpolation vary slightly depending on the platform. In
upcoming changes, we are going to switch to using Skia as the default
painter, which leaves us with the following options:
- Have per-platform screenshot test expectations.
- Limit screenshot tests to run only on one platform and maintain a
single set of expectation files.
For now, I have decided to choose the latter option, using Linux as it
seems to be the most popular platform among developers.
These were being immediately stored in JS::GCPtrs (and dutifully visited
by HTMLParser), so creating temporary handles for them was a complete
waste of time.
When loading a canned version of reddit.com, we end up parsing many many
shadow tree style sheets of roughly ~170 KiB text each.
None of them have '\r' or '\f', yet we spend 2-3 ms for each sheet just
looping over and reconstructing the text to see if we need to normalize
any newlines.
This patch makes the common case faster in two ways:
- We use TextCodec::Decoder::to_utf8() instead of process()
This way, we do a one-shot fast validation and conversion to UTF-8,
instead of using the generic code-point-at-a-time callback API.
- We scan for '\r' and '\f' before filtering, and if neither is present,
we simply use the unfiltered string.
With these changes, we now spend 0 ms in the filtering function for the
vast majority of style sheets I've seen so far.
This way, we still perform UTF-8 validation, but don't go through the
slow generic code path that rebuilds the decoded string one code point
at a time.
This was a bottleneck when loading a canned copy of reddit.com, which
ended up being ~120 MiB large.
- Time spent decoding UTF-8 before this change: 1192 ms
- Time spent decoding UTF-8 after this change: 154 ms
That's still a long time, but 7.7x faster is nothing to sneeze at! :^)
Note that if the input fails UTF-8 validation, we still fall back to
the slow path and insert replacement characters per the WHATWG Encoding
spec: https://encoding.spec.whatwg.org/#utf-8-decode
This change makes OpenType::Name::string_for_id handle fonts whose names
are UTF-16-encoded (along with handling UTF-8-encoded names).
Otherwise, without this change, the existing code assumes the names are
UTF-8-encoded, fails gracelessly if they’re not, and crashes.
Fixes https://github.com/LadybirdBrowser/ladybird/issues/75
This is the expected behavior per the HTML spec. Fixes an issue where
styling these elements wouldn't have the expected effect unless you also
set the display property.
For performance, rather than slowly incrementing the capacity of the
rope string's buffer, compute an approximate length for that buffer to
be reserved up front.
Instead of allowing arbitrarily large values (which could eventually
overflow an i32), let's just cap them at the same limit as Firefox does.
Found by Domato.
This change will make it easier to disable screenshot comparison tests
on a specific platform or have per-platform expectations.
Additionally, it's nice to be able to tell if a ref-test uses a
screenshot as an expectation by looking at the test path.
The EntryType has three possible values: Fetching, Failed or
ModuleScript. It is possible that we transition from Fetching to Failed
as in #13.1. Change the assertion to include the failed scenario.
Fixes: https://github.com/LadybirdBrowser/ladybird/issues/661
GCC 14 emits a warning when an always succeeding `dynamic_cast`'s return
value is compared to NULL inside the `AK::is<T>(U)` template when `T` ==
`U`.
While warning on tautological `is` calls seems useful, it's a bit
awkward when it comes from a function template where the cast may fail
in some instantiation. There is a GCC bug open for it:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115664
Work around the warning by performing the algorithm on the base type
(`EventTarget`), with a wrapper that casts it to the more specialized
input type.
