Sometimes people make tables with a specific width. In those cases,
we can't just use the auto-sizing algorithm, but instead have to
respect whatever width the content specifies.
This is a bit rickety right now, since we don't implement generation
of anonymous table boxes.
The basic mechanism here is that block layout (which table-cell uses)
now has a virtual way of asking for the width of the logical containing
block. This is necessary as table-row does not produce a block-level
element and so was previously unable to provide a containing block
width for table-cell layout.
If the table has a non-auto specified width, we now interpret that as
a request to use fixed table layout. This will evolve over time. :^)
Margin collapsing is a bit confusing, but if I understand correctly,
when collapsing a box's top margin with the vertically adjacent
sibling box above, we should "skip over" empty (0-height) boxes and
collapse their margin with *their* vertically adjacent sibling box
above, etc. Iterating until we hit the first non-empty in-flow block.
This pulls up the bottom part of the face on ACID2. :^)
We could previously place a box next to a preceding sibling with
position:fixed, which is wrong since fixed-position elements are taken
out of the normal flow.
When highlighting a node in the inspector, we now paint three overlays:
- The content box (magenta)
- The padding box (cyan)
- The margin box (yellow)
This makes it a lot easier to debug layout issues.
Sometimes we end up with an empty line box at the bottom of a block.
Instead of worrying about this in all the places we split into lines,
just remove the trailing box (if any) after splitting is finshed.
To make this possible, I also had to give each LayoutNode a Document&
so it can resolve document-specific colors correctly. There's probably
ways to avoid having this extra member by resolving colors later, but
this works for now.
This patch also adds the ability for Length to contain percentage
values. This is a little off-spec, but will make storing and dealing
with lengths a lot easier.
To resolve a Length to a px-or-auto Length, there are now helpers
for that. After calling them, you no longer have to think about
em, rem, %, and such things.
StyleProperties is really only the specified "input" to what eventually
becomes the used/computed style we use for layout and painting.
Unlike StyleProperties, LayoutStyle will have strongly typed values for
everything it contains (i.e no CSS::ValueID or strings, etc.)
This first patch moves z-index into LayoutStyle.
This avoids having to query the StyleProperties hash map whenever we
need to know if an element is absolutely positioned. This was extremely
hot in interactive window resize profiles.
In this case, we need to undo the right-side offsetting, since the
width computation algorithm will already have stretched the width to
accomodate both the side constraints.
Previously we would always just use the combined content width as the
shrunken width in shrink-to-fit width calculations, but if the element
has a non-auto specified width, we should just let that take over.
This is far from perfect and doesn't take stuff like min/max-width
into account. Will need more work, this just covers the basic case.
We now use the ImageDecoder service in LibWeb for everything except
GIF images (we'll have to deal with them later, ofc.)
This has a little bit of overhead but we should be able to optimize
it until it becomes negligible.
The stacking context tree doesn't affect layout at all, so let's move
it into the Painting/ directory. I'm not sure yet if it's worth going
for a fullly separate painting tree. So far I'm thinking a stacking
context tree with pointers into the layout tree might be enough.
"Paint" matches what we call this in the rest of the system. Let's not
confuse things by mixing paint/render/draw all the time. I'm guilty of
this in more places..
Also rename RenderingContext => PaintContext.
The shrink-to-fit width algorithm actually works a little bit different
in the absolute positioning context, so it can't share all of its code
with non-absolute positioning.
Also, inline-block elements were always inserting unnecessary line
breaks when splitting, which caused the preferred width to be smaller
than it should be. This patch fixes that as well, by just not breaking
after inline-block elements in LayoutMode::OnlyRequiredLineBreaks.
CSS defines a very specific paint order. This patch starts steering us
towards respecting that by introducing the PaintPhase enum with values:
- Background
- Border
- Foreground
- Overlay (internal overlays used by inspector)
Basically, to get the right visual result, we have to render the page
multiple times, going one phase at a time.
The more generic virtual variant is renamed to node_name() and now only
Element has tag_name(). This removes a huge amount of String ctor/dtor
churn in selector matching.
Hey, why not. We did all the hard work for display:inline-block already
and now we can just allow this.
This makes <a><h1>Hello friends!</h1></a> work :^)
To support z-ordering when painting, the layout tree now has a parallel
sparse tree of stacking contexts. The rules for which layout boxes
establish a stacking context are a bit complex, but the intent is to
encapsulate the decision making into establishes_stacking_context().
When we paint, we start from the ICB (LayoutDocument) who always has a
StackingContext and then paint the tree of StackingContexts where each
node has its children sorted by z-index.
This is pretty crude, but gets the basic job done. Note that this does
not yet support hit testing; hit testing is still done using a naive
treewalk from the root.
The only layout nodes that don't have their own style are LayoutText
(they inherit the style from their parent element since text cannot
be styled by CSS.)
However, it never makes sense for text nodes to have absolute position
so don't claim it.
Previously, layout recursively performed these steps (roughly):
1. Compute own width
2. Compute own position
3. Layout in-flow children
4. Compute own height
5. Layout absolutely positioned descendants
However, step (2) was pretty inconsistent. Some things computed their
own position, others had their parent do it for them, etc.
To get closer to CSS spec language, and make things easier in general,
this patch reorganizes the algorithm into:
1. Compute own width & height
2. Compute width & height of in-flow managed descendants
3. Move in-flow managed descendants to their final position
4. Layout absolutely positioned descendants
Block layout is now driven by the containing block, which will iterate
the descendants it's responsible for. There are a lot of inefficient
patterns in this logic right now, but they can easily be replaced with
better iteration functions once we settle on a long-term architecture.
Since the ICB (LayoutDocument) is at (0, 0), it doesn't rely on a
containing block to move it into place.
This code is still evolving along with my understanding of CSS layout,
so it's likely that we'll reorganize this again sooner or later. :^)
