This fixes a plethora of rounding problems on many websites.
In the future, we may want to replace this with fixed-point arithmetic
(bug #18566) for performance (and consistency with other engines),
but in the meantime this makes the web look a bit better. :^)
There's a lot more things that could be converted to doubles, which
would reduce the amount of casting necessary in this patch.
We can do that incrementally, however.
Introduce optimization that determines if one node is preceding,
following, ancestor or descdendant of another node by comparing
ancestors chains of both nodes which is much cheaper than using
Node::is_before() that can walk whole subtree in the worst case.
SVG presentation attributes are parsed as CSS values, so we also need to
handle CSS variable expansion when handling them.
This (roughly) matches the behavior of other engines. It's also used on
the web, for example on https://stripe.com/ :^)
If we finish parsing a calculation tree and it still contains
UnparsedCalculationNodes, then it's not valid, and we shouldn't create
a StyleValue from it.
This avoids a ton of work when painting large documents. Even though it
would eventually get clipped by the painter anyway, by bailing out
earlier, we avoid a lot more work (UTF-8 iteration, OpenType lookups,
etc).
It would be even nicer if we could skip entire line boxes, but we don't
have a fast way to get the bounding rect of a line box at the moment.
Some of the live HTMLCollection only ever contain children of their root
node. When we know that's the case, we can avoid doing a full subtree
traversal of all descendants and only visit children.
This cuts the ECMA262 spec loading time by over 10 seconds. :^)
Previously, calling `.right()` on a `Gfx::Rect` would return the last
column's coordinate still inside the rectangle, or `left + width - 1`.
This is called 'endpoint inclusive' and does not make a lot of sense for
`Gfx::Rect<float>` where a rectangle of width 5 at position (0, 0) would
return 4 as its right side. This same problem exists for `.bottom()`.
This changes `Gfx::Rect` to be endpoint exclusive, which gives us the
nice property that `width = right - left` and `height = bottom - top`.
It enables us to treat `Gfx::Rect<int>` and `Gfx::Rect<float>` exactly
the same.
All users of `Gfx::Rect` have been updated accordingly.
Before this change, LayoutState essentially had a Vector<UsedValues*>
resized to the exact number of layout nodes in the current document.
When a nested layout is performed (to calculate the intrinsic size of
something), we make a new LayoutState with its own Vector. If an entry
is missing in a nested LayoutState, we check the parent chain all the
way up to the root.
Because each nested LayoutState had to allocate a new Vector with space
for all layout nodes, this could get really nasty on very large pages
(such as the ECMA262 specification).
This patch replaces the Vector with a HashMap. There's now a small cost
to lookups, but what we get in return is the ability to handle huge
layout trees without spending eternity in page faults.
By deferring allocation of StaticNodeList objects until we know somebody
actually wants the MutationRecord, we avoid a *lot* of allocation work.
This shaves several seconds off of loading https://tc39.es/ecma262/
At least one other engine (WebKit) skips creating mutation records if
nobody is interested, so even if this is observable somehow, we would
at least match the behavior of a major engine.
Before this, any style change that mutated a property we consider
"layout-affecting" would trigger a complete teardown and rebuild of the
layout tree.
This isn't actually necessary for the vast majority of CSS properties,
so this patch makes the invalidation a bit finer, and we now only
rebuild the layout tree when the CSS display property changes.
For other layout-affecting properties, we keep the old layout tree (if
we have one) and run the layout algorithms over that once again.
This is significantly faster, since we don't have to run all the CSS
selectors all over again.
Rather than queueing microtasks ad nauseam to check if a media element
has a new source candidate, let the media element tell us when it might
have a new child to inspect. This removes endless CPU churn in cases
where there is never a candidate that we can play.
This fixes a crash in box_baseline, due to cells created for
display: table expecting a box child and getting the inline node wrapper
instead.
Fixes#18972.
This now allows websites such as Discord, YouTube and your favourite
"Right Click" xkcd comic to open a custom context menu when you right
click. You can bypass this by holding shift, just like Firefox.
Avoid possible null optional dereference when creating border radius
clipper, and avoid creating clipper if the clip rect is empty (which
prevents some debug spam). Also remove an unnecessary lambda.
The callee that we're passing it to expects a GCPtr anyway, so there's
no need to explicitly dereference this.
Fixes a crash when loading https://spotify.com/
Using HashTable of grid positions to represent OccupationGrid allows to
simplify positioning code. For example maybe_add_row() and
maybe_add_column() calls are not needed anymore because there is no
Vector<Vector<bool>> that need to be resized.
No observable changes in grid layout are expected :)
It's not safe to allocate from the GC heap while in the constructor of a
GC heap cell. (Because if this ends up triggering a collection, we may
end up trying to call through an uninitialized vtable).
This was already done safely in the initialize() virtual in much of
LibJS and LibWeb. This patch moves the logic for prototypes, mixins,
and CSSStyleDeclaration as well.
Fixes a long-standing GC crash that was pretty easy to reproduce by
refreshing https://vercel.com/
Since the specifications indicate that the algorithm for sizing tracks
without any spanning items is a simplified version of the more general
algorithm used for sizing tracks with spanning items, we can reuse the
code to size both cases.
Implements more parts of sizing algorithm for tracks with spanning
items to archive parity with implementation for sizing of tracks
with non-spanning items.
We have to special-case these, otherwise our normal CSS layout algorithm
will see that some SVG roots have width/height assigned, and make those
the used width/height.
When used in an SVG-as-image context, the outermost viewport must be the
authoritative root size.
This finally makes SVG-as-image show up visually! :^)
We should find a way to share this logic with Layout::SVGSVGBox, but
that will require some finesse since they have to work at different
points in the layout/paint timeline.
In order to separate the SVG content from the rest of the engine, it
gets its very own Page, PageClient, top-level browsing context, etc.
Unfortunately, we do have to get the palette and CSS/device pixel ratios
from the host Page for now, maybe that's something we could refactor in
the future.
Note that this doesn't work visually yet, since we don't calculate the
intrinsic sizes & ratio for SVG images. That comes next. :^)
This allows the painting subsystem to request a bitmap with the exact
size needed for painting, instead of being limited to "just give me a
bitmap" (which was perfectly enough for raster images, but not for
vector graphics).
The existing implementation moves down into a new subclass called
AnimatedBitmapDecodedImageData.
The purpose of this change is to create an extension point where we can
plug in an SVG renderer. :^)
This implements the stop-opacity, fill-opacity, and stroke-opacity
properties (in CSS). This replaces the existing more ad-hoc
fill-opacity attribute handling.
There are a couple of things that went into this:
- We now calculate the intrinsic width/height and aspect ratio of <svg>
elements based on the spec algorithm instead of our previous ad-hoc
guesswork solution.
- Replaced elements with automatic size and intrinsic aspect ratio but
no intrinsic dimensions are now sized with the stretch-fit width
formula.
- We take care to assign both used width and used height to <svg>
elements before running their SVG formatting contexts. This ensures
that the inside SVG content is laid out with knowledge of its
viewport geometry.
- We avoid infinite recursion in tentative_height_for_replaced_element()
by using the already-calculated used width instead of calling the
function that calculates the used width (since that may call us right
back again).
In order to fix this, I also had to reorganize the code so that we
create an independent formatting context even for block-level boxes
that don't have any children. This accidentally improves a table
layout test as well (for empty tables).