Extract the file-local is_supported_image_type() helper from
HTMLImageElement into a small standalone translation unit so other
parts of the engine can ask the same question. The next commit reuses
it for the image-set() type() filter.
The list is still hard-coded; deriving it from the registered image
decoders remains a FIXME.
When resolving cross-axis auto margins on a flex item, the outer cross
size calculation omitted all cross-axis margins. We now include
non-auto margins as part of the outer cross size treating auto margins
as zero.
We don't match the C++ output byte-for-byte here, because the key
iteration order is different. So I've taken the opportunity to tidy up
some whitespace too.
Move owned ArrayBuffer storage directly when transferring stream
buffers instead of copying the bytes before detaching the source.
WebAssembly memory continues to copy because its ArrayBuffer wraps
externally-owned storage.
Preserve the abrupt completion from DetachArrayBuffer before moving
storage so non-transferable buffers, such as WebAssembly.Memory-backed
views, still surface TypeError through stream operations instead of
aborting.
This saves ~130ms of main thread time when loading a YouTube video
on my Linux computer. :^)
This matches the behavior of other engines. Some CDNs that do content
negotiation will fall back to non alpha-preserving formats if these
values are not present.
HashMap<_, GC::Ref<_>>::ensure() crashed under UBSan whenever the
initialization callback triggered a GC: lookup_for_writing() stamped
the target bucket as used and added it to the ordered list before the
callback ran, so the marking visitor walked the map, read the
uninitialized slot, and failed the returns_nonnull check in GC::Ref.
Split bucket reservation into two phases. lookup_for_writing() now
hands back the target in the Free state (not in the ordered list,
m_size unchanged); callers placement-new the value and then commit via
commit_inserted_bucket(). The Robin Hood displacement loop still
stamps the slot internally and un-stamps before returning, so probing
is unchanged and the whole operation remains a single hash and a
single probe.
WebAssembly.Memory-backed ArrayBuffers wrap external
ByteBuffer storage. When that memory grows,
ByteBuffer::try_resize() may realloc the backing storage while
old fixed-length buffer objects remain reachable from JS.
TypedArrayBase cached m_data for all fixed-length buffers, and
the asm interpreter fast path dereferenced that cached pointer
directly. For wasm memory views this could leave a stale
pointer behind across grow().
Restrict cached typed-array data pointers to fixed-length
ArrayBuffers that own stable ByteBuffer storage.
External/unowned buffers, including WebAssembly.Memory
buffers, now keep m_data == nullptr and fall back to code that
re-derives buffer().data() on each access.
Add regressions for both the original shared-memory grow case
and the second-grow stale-view case.
Implement the width filtering step of the font matching algorithm.
Without it, system font providers that group all widths under one
family could return a condensed variant for font-width: normal,
producing visibly narrower text.
FontFace.load() set the face's status to "loading" but never switched
the containing FontFaceSets to "loading" or appended to their
`[[LoadingFonts]]` lists. The load-completion handler then found
`[[LoadingFonts]]` already empty and fired switch-to-loaded after the
first face finished, resolving `document.fonts.ready` while faces in
the same set were potentially still loading.
Previously, if search was disabled, entering non-URL text would just
silently drop the search query (and on Qt, we would reload the current
URL). We now detect that the query did not result in a navigation and
load an error page instead, which directs the user to enable search.
Move partial interface, partial namespace, mixin, typedef, and overload
finalization into a context-wide post-parse resolve step.
This lets BindingsGenerator parse all declared IDL files first and then
finalize the shared IDL context before generating bindings.
Teach the bindings build to pass support IDL files alongside the regular
binding IDLs so BindingsGenerator parses the full declared IDL set into
a shared context.
Keep idl_files.cmake as the source of truth for parsed support IDLs, and
let Interface::will_generate_code() decide which parsed interfaces emit
generated bindings.
Semantic IDL lookups no longer rely on imported modules, so stop storing
that state on Interface and Module.
import directives are still resolved during parsing for now, but the
parsed imported-module lists are no longer needed.
Stop using imported modules to decide which binding headers to include.
Instead, collect dependencies from referenced interfaces, dictionaries,
enumerations, callbacks, and nested IDL types in the parsed context.
Derive C++ namespaces from each IDL module's location and use those
qualified names when generating binding code.
Also Teach dictionaries their owning IDL module path so dictionary C++
types can be qualified the same way as interfaces. This removes the need
for the generated `using namespace Web::*` hack and the hard-coded
namespace list.
Also fix DOMURL.idl to refer to the IDL interface name `URL`, not the
C++ implementation name `DOMURL`.
No other third layer folder in LibWeb has its own namespace which
makes this a special case for the IDLGenerator when determining
namespaces. Instead of adding a special case, simply remove the
namespace.
Teach BindingsGenerator to parse and generate bindings for the full
LibWeb IDL set in one invocation, and collapse the CMake bindings
rules from one custom command per IDL file to a single batched codegen
step.
This has the downsides of:
* Any single IDL change now reruns the whole bindings generator
* Per-IDL parallelism at the custom-command level is lost
However, I still feel that this is a worthy trade off as:
* Generated files are written with write_if_changed(), so rebuilds
of generated files should not be significantly impacted.
* It is not a common task to be modifying IDL files
* Most importantly, giving the IDL generator full knowledge of _all_
IDL will allow for some simplifications of the bindings generator as
it has knowledge of all types.
HTMLImageElement's update-the-image-data step 16 queues its state
transition and load event dispatch via a 1 ms BatchingDispatcher, so
the current request does not become CompletelyAvailable synchronously
when the fetch finishes. decode()'s on_finish callback, however, was
queuing its resolve task directly on the event loop, bypassing the
batch. That race meant decode() could resolve while the image request
was still in Unavailable state, so any .then() handler inspecting
img.width / img.height (or anything derived from the bitmap) would see
zeros.
Google Maps hits this on its .9.png road shield icons: after awaiting
img.decode() it reads a.width / a.height and calls
ctx.getImageData(0, 0, 0, 0), which throws IndexSizeError and aborts
the tile rendering pipeline.
Route decode()'s on_finish through the same BatchingDispatcher so both
are processed in the same batch, with the decode resolution queued
after step 16's element task.
During intrinsic sizing, compute_width() ran on block descendants with
an intrinsic-sizing available space. For a non-FC-establishing block
with auto width, used_width stayed auto, and the min-width clamp then
compared AvailableSize::min-content against min-width via operator<,
which always returns true when the left side is min-content. The clamp
fired with min-width: 0 and set content_width to 0 permanently.
Skip the min-width clamp when used_width is still auto, mirroring the
max-width clamp a few lines above which already no-ops via
to_px_or_zero. The real width is then set by the IntrinsicSizing branch
in layout_block_level_children.
invalidate_all_prototype_chains_leading_to_this used to scan every
prototype shape in the realm and walk each one's chain looking for
the mutated shape. That was O(N_prototype_shapes x chain_depth) per
mutation and showed up hot in real profiles when a page churned a
lot of prototype state during startup.
Each prototype shape now keeps a weak list of the prototype shapes
whose immediate [[Prototype]] points at the object that owns this
shape. The list is registered on prototype-shape creation
(clone_for_prototype, set_prototype_shape) and migrated to the new
prototype shape when the owning prototype object transitions to a
new shape. Invalidation is then a recursive walk over this direct-
child registry, costing O(transitive descendants).
Saves ~300 ms of main thread time when loading https://youtube.com/
on my Linux machine. :^)
notify_about_rejected_promises() is called for every related environment
settings object at the end of every microtask checkpoint. It was
unconditionally reading the realm up front, which showed up at 3.0% self
time in a YouTube playback profile.
This patch moves the realm lookup into the queued task callback, which
happens way less often.
Document::is_decoded_svg() was reached through two pointer hops and a
virtual call into PageClient on every invocation. It showed up at 1.9%
self time in a YouTube playback profile, and it's also called for every
document in the hot documents_in_this_event_loop_matching() loop that
runs on every rendering update.
The page's client is fixed for the lifetime of a document, so we can
cache the answer at construction time and serve future calls from a
plain member load.
No other engine defines this function, so it is an observable difference
of our engine. This traces back to the earliest days of LibJS.
We now define `gc` in just the test-js and test262 runners.
Previously, we were accidentally creating temporary copies of custom
property maps on both sides of a ternary in `compute_style_impl()`. We
now bind to a static empty sentinel instead so the reference binds
directly to `own_values()` without copying.
font_for_code_point() was the heaviest function in layout profiles
of a YouTube page (216ms CPU out of 2900ms total). Every call walked
the full cascade and ran a virtual contains_glyph() against each
entry, even though the result is the same for most ASCII code points
across a document.
Add a 128-entry direct-mapped cache keyed by code point that stores
the resolved Font pointer on first lookup. Subsequent ASCII lookups
become a null check plus a load.
No invalidation is needed: m_fonts is append-only, and the cascade
returns the first matching font, so once an entry claims a code
point, later appends cannot change the answer.
