With the work to unify the layout logic for replaced and non-replaced
boxes (#37942), I think the `IndependentNonReplacedContents` enum does
no longer make much sense.
Therefore, this removes `IndependentNonReplacedContents`, merging its
values into `IndependentFormattingContextContents`.
The methods defined on `IndependentFormattingContextContents` can now be
on `IndependentFormattingContext`, in particular this implies that the
layout results of a replaced box will now be cached.
Testing: Unneeded (no behavior change)
This part of #37942
Signed-off-by: Oriol Brufau <obrufau@igalia.com>
When a style change does not chang the structure of the box tree, it is
possible to skip box tree rebuilding for an element. This change adds
support for reusing old box trees when no element has that type of
damage. In order to make this happen, there needs to be a type of
"empty" `LayoutDamage` that just indicates that a fragment tree layout
is necessary.
This is the first step toward incremental fragment tree layout.
Testing: This should not change observable behavior and thus is covered
by
existing WPT tests. Performance numbers to follow.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
Previously, anonymous boxes, such for anonymous table parts were not
associated with their non-pseudo ancestor DOM nodes. This presents a
problem when it comes time to clear layout data during incremental
layouts. This change reworks the way that pseudo-elements in general are
stored in their non-pseudo ancestor DOM nodes, allowing for any number
to be placed there.
This trades a bit of performance for space, as just adding a vector to
the node would add something like 24 bytes of storage to every node.
This change should have a neutral runtime memory usage.
Testing: This shouldn't change observable behavior and is thus covered
by
existing WPT tests. It will allow tests to pass in a subsequent PR.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
Clearance only applies to block-level boxes, so it was unnecessary to
require it as a parameter. Instead, in block layout we can set it using
the new `.with_clearance()` method.
Testing: Unnecessary (no behavior change)
Signed-off-by: Oriol Brufau <obrufau@igalia.com>
It was very easy to forget about using `.with_specific_layout_info()` to
set the specific layout info, so it's better to make it a parameter.
In fact this already happened in the past: #36993 fixed the missing
specific layout info for flex items.
This patch fixes it for floats and atomic inlines. It also propagates it
in other cases where not doing so was not a big deal because the
specific layout info was None, but that was a fragile assumption.
Testing: Various WPT improvements
Fixes: #37898
Signed-off-by: Oriol Brufau <obrufau@igalia.com>
This change extends incremental box tree updates to table columns.
Testing: This should not change observable behavior and is thus covered
by existing WPT tests.
Signed-off-by: Oriol Brufau <obrufau@igalia.com>
Co-authored-by: Martin Robinson <mrobinson@igalia.com>
This change extends incremental box tree updates to table cells. In
addition, for simplicity this refactors `BoxSlot::take_layout_box()`
into `BoxSlot::take_layout_box_if_undamaged()`.
Testing: This should not change observable behavior and is thus covered
by existing WPT tests.
Signed-off-by: Oriol Brufau <obrufau@igalia.com>
Co-authored-by: Martin Robinson <mrobinson@igalia.com>
This change extends incremental box tree updates to table captions. In
addition, calls to `LayoutBox::invalidate_cached_fragment()` are moved
to the damage calculation traversal.
Testing: This should not change observable behavior and is thus covered
by existing WPT tests.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
In layout, some parts of the code were still using parallel iterators
from Rayon even when single-thread layout was activated. This change
modifies those parts to use non-parallel iterators when
`LayoutContext::use_rayon` is not active.
Testing: It's very hard to make an automated test for this, but I've
manually
verified this by building with tracing and observing that layout runs
only on
a single thread now when loading https://servo.org.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
This implementation is quite confusing as it makes it harder to tell
that we are just looking for the case that `Contents` contains
`NonReplacedContents`.
Testing: This shouldn't have any functional change, so is covered by
existing tests.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Now that we are standardizing on the `_traits` crates becoming `_api`
and exposing the API of the crate that they get their name from [^1],
`script_layout_interface` becomes `layout_api` as it exposes the API for
`layout` that is used by `script` This brings the crate in line with the
naming of the other ones in `shared`.
[^1]:
https://servo.zulipchat.com/#narrow/channel/263398-general/topic/Organizing.20*_traits.20crates/with/396893711
Testing: This should not change any behavior and thus is covered by
existing tests.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Text decorations have a special kind of propagation. Instead of
propating these during box tree construction, move propagation to
stacking context tree construction. This will allow for using a very
easy type of incremental layout when text decorations change. For
instance, when a link changes color during hovering over it, we can skip
all of box and fragment tree construction.
In addition, propagation works a bit better now and color and style
properly move down from their originating `Fragment`s.
This introduces three new failures, because now we are drawing the
text-decoration with the correct color in more places, which exposes an
issue we have with text-decorations not being drawn in relation to the
baseline (taking into account `vertical-align`).
Testing: There are tests for these changes.
Fixes#31736.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
When making last-minute changes to the repaint-only layout pass, damage
propagation was broken, meaning that full layout was always done. This
change fixes that, meaning that times in the `blaster.html` test case
now reflect those described in the original commit message from #36978.
In addition, some style repair is now fixed:
- `InlineFormattingContext`s now keep a `SharedInlineStyles` for the
root of the IFC
which is updated during style repair.
- `BlockFormattingContext`s now properly update their style.
These changes are verified by turning on repaint only layout for more
properties
in Stylo via servo/stylo#183.
Testing: Manual performance testing via `blaster.html`.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
This change adds the simplest kind of incremental layout. When Servo
detects that all style changes only require a repaint, only run stacking
context tree and WebRender display list generation. This means that
these kind of restyles do not need a re-layout. Instead, the existing
box and fragment trees will be used and the styles of damaged nodes will
be updated in their box and fragment tree nodes.
This requires a new style repair DOM traversal for nodes that have had
their style damaged. In addition, careful accounting of all the places
where we store style must happen in order ot update those styles.
Testing: This is covered by existing WPT tests as it should not change
observable behavior.
We have created a test case which shows a 50% speedup when run
in Servo, even though there still a long way to go to match the speed
of other browsers:
https://gist.github.com/mrobinson/44ec87d028c0198917a7715a06dd98a0
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
`TextRun`s use their parent style to render. Previously, these styles
were cloned and stored directly in the box tree `TextRun` and resulting
`TextFragment`s. This presents a problem for incremental layout.
Wrapping the style in another layer of shared ownership and mutability
will allow updating all `TextFragment`s during repaint-only incremental
layout by simply updating the box tree styles of the original text
parents.
This adds a new set of borrows when accessing text styles, but also
makes it so that during box tree block construction
`InlineFormattingContext`s are created lazily and now
`InlineFormattingContextBuilder::finish` consumes the builder, making
the API make a bit more sense. This should also improve performance of
box tree block construction slightly.
Testing: This should not change observable behavior and thus is covered
by existing WPT tests.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
Table cells share background styles with their track and track group
boxes. When a track and track group style is repaired, this new data
structure will allow reparing the style of the cell `Fragment`s without
having to lay the table out again or walk through `Fragment`s and
individually repair their background styles.
Testing: This doesn't change behavior and is thus tested by existing
WPT tests.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
This makes it so that layout is no longer generic on the node type,
depending directly on `script`'s `ServoLayoutNode`. In addition to
greatly simplifying layout, this is necessary because incremental layout
needs to be able to create pseudo-element styles without having a handle
on the original `impl LayoutNode`. We feel this is a reasonable
tradeoff.
Testing: No functional changes, so covered by existing WPT tests.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
`PositioningContext` held two vectors, one inside an `Option`, to
differentiate between the version used for a containing block for all
descendants (including `position: absolute` and `position: fixed`) or
only for `position: absolute` descendants. This distinction was really
hard to reason about and required a lot of bookkeeping about what kind
of `PositioningContext` a layout box's parent expected. In addition, it
led to a lot of mistakes.
This change simplifies things so that `PositioningContext` only holds a
single vector. When it comes time to lay out hoisted absolutely
positioned
fragments, the code then:
- lays out all of them (in the case of a `PositioningContext` for all
descendants), or
- only lays out the `position: absolute` descendants and preserves the
`position: fixed` descendants (in the case the `PositioningContext`
is only for `position: absolute`.), or
- lays out none of them if the `PositioningContext` was created for
box that did not establish a containing block for absolutes.
It's possible that this way of dealing with hoisted absolutes is a bit
less efficient, but, the number of these descendants is typically quite
small, so it should not be significant. In addition, this decreases the
size in memory of all `PositioningContexts` which are created in more
situations as time goes on.
Testing: There is a new WPT test with this change.
Fixes: #36696.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Previously, spans were partially clamped during layout, but this means
that accessing and setting these properties via script wouldn't behave
according to the HTML specification. In addition, the value wasn't
floored in layout, so could lead to panics. This change improves
clamping and moves it to script.
Testing: This change includes a new WPT test.
Fixes#36699.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
As per
[w3.org/TR/filter-effects-1#FilterProperty](https://www.w3.org/TR/filter-effects-1/#FilterProperty),
`filter` shouldn't make the root element establish a containing block
for absolute and fixed positioned descendants. `will-change: filter` has
matching behavior.
This PR adds a check for if we are the root element before establishing
such a block.
To know if we are the root element, we look at the `FragmentFlags`
passed in. Previously for our function, these were dummy flags, always
constructed as empty. Thus, this PR also makes sure the correct
FragmentFlags are passed down the chain to the function
`establishes_containing_block_for_all_descendants`.
Testing:
- `/css/filter-effects/filtered-html-is-not-container.html` now passes
- `/css/css-will-change/will-change-fixedpos-cb-003.html` now passes
- Manual tests are working
Fixes: #35391
---------
Signed-off-by: haval0 <56519858+haval0@users.noreply.github.com>
Signed-off-by: Oriol Brufau <obrufau@igalia.com>
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
This change switches `offsetParent`, `offsetLeft`, etc queries to use
the BoxTree fragments instead of walking the entire fragment tree. In
addition, fragments are stored for columns and colgroups. In general,
this greatly simplifies the flow of the query and prevents having to do
expensive tree walks.
Testing: This change is covered by newly passing WPT tests and three new
failures:
- /css/filter-effects/backdrop-filter-edge-clipping-2.html
- /css/filter-effects/backdrop-filter-edge-mirror.html
- /css/filter-effects/backdrop-filter-edge-pixels-2.html
These failures are actually progressions, because now the references
start
to render properly whereas before they did not.
Fixes: This is part of #36525 and #36665.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Co-authored-by: Oriol Brufau <obrufau@igalia.com>
We previously ignored the opaque layout data field inside each node when
measuring a DOM node's memory usage. While some of the reachable memory
was accounted for by measuring the layout's box tree, measuring it via
the node ensures that we don't miss anything. Since there are often Arc
values involved, this means that the layout-thread box tree measurements
now look quite small, while reported JS heap usage has increased.
Testing: Manually compared about:memory for servo.org.
---------
Signed-off-by: Josh Matthews <josh@joshmatthews.net>
Now that legacy layout has been removed, the name `layout_2020` doesn't
make much sense any longer, also it's 2025 now for better or worse. The
split between the "layout thread" and "layout" also doesn't make as much
sense since layout doesn't run on it's own thread. There's a possibility
that it will in the future, but that should be something that the user
of the crate controls rather than layout iself.
This is part of the larger layout interface cleanup and optimization
that
@Looriool and I are doing.
Testing: Covered by existing tests as this is just code movement.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
