Implement vertical-align via Parley - #832
Draft
nicoburns wants to merge 15 commits into
Draft
Conversation
Map Stylo's alignment-baseline/baseline-shift longhands to Parley's new VerticalAlign property on text styles and atomic inline boxes, report the inline-block baseline (last line box, unless overflow is not visible) to Parley, and drop the root line-height floor workaround now that Parley adds the strut and ancestor inline boxes itself.
Contributor
🤖 Devin AI EngineerI'll be helping with this pull request! Here's what you should know: ✅ I will automatically:
Note: I can only respond to comments from users who have write access to this repository. ⚙️ Control Options:
|
Contributor
Blitz real-page benchmark: Barack Obama Wikipedia article, 1280 pxHeadless parse + style + layout of Variants
Identical features for all (default features of the root Methodmkdir -p ~/bench/obama && cd ~/bench/obama
curl -L -o obama.html https://en.wikipedia.org/wiki/Barack_Obama
# the two load.php stylesheets referenced by <link rel=stylesheet> were fetched to style0.css / style1.css
# and the <link href> attributes rewritten to those local names. No other network input.
# one git worktree per variant, scratch example copied in (not committed anywhere):
cp obama_bench.rs ~/repos/blitz-{A,B,C,D}/examples/
(cd ~/repos/blitz-$V && cargo build --release --example obama_bench)
# 3 interleaved rounds A B C D, then medians of the per-run medians:
for round in 1 2 3; do for v in A B C D; do
/usr/bin/time -v ~/repos/blitz-$v/target/release/examples/obama_bench ~/bench/obama/obama.html 20 3
done; done
Sanity check that all variants do the same work: an instrumented build counted the items Blitz walks in Results (median of 3 interleaved rounds; run 2, which includes D)
Per-round medians (first / relayout, ms) — run 2: A 282/158, 269/146, 265/147 · B 848/741, 847/735, 845/734 · C 1070/959, 1080/950, 1068/961 · D 1045/925, 1050/931, 1051/941. Run 1 (A B C only) gave the same picture: A 267/161, B 853/742, C 1088/952. Run-to-run noise is ~±3%. Where the time goes (perf,
|
| A | B | C | D | |
|---|---|---|---|---|
| total | 1 670 | 5 906 | 7 206 | 7 188 |
parley::layout::line::GlyphRunIter::next (called from Blitz compute_inline_layout_inner's for item in line.items() loop) |
216 | 3 826 | 4 929 | 4 980 |
of which libc memcpy (self) |
— | 2 093 | 3 286 | 3 427 |
build_inline_layout_into / parley::builder::build_into_layout (tree build + shaping + style metrics) |
448 | 524 | 484 | 517 |
fontique::Query::matches_with (all callers) |
104 | 102 | 21 | 97 |
parley::layout::style_metrics::resolve_style_metrics |
— | 16.5 | 6.5 | — |
harfrust … hb_font_t::shape (sanity, should be constant) |
112 | 111 | 103 | 110 |
Interpretation
-
The 3–4× regression from A to B/C/D is not Add patch for naga to fix macOS support #18 and not the cache; it is Parley
mainvs the0.11.1release, and it is entirely inGlyphRunIter::next. D (Blitzmaincode, only the Parley pin moved to Parleymain) is as slow as C. Blitz'scompute_inline_layout_inneriteratesline.items()for every line to position inline boxes;GlyphRunIter::nextre-walksrun.visual_clusters().flat_map(|c| c.glyphs()).skip(glyph_start)from the start of the run for every glyph run (already O(glyphs × glyph-runs) in 0.11.1), and on Parleymaineach step now copies a largeCluster(Copystruct holding anAtom/ShapedSlice) — therep movsbmemcpy is 35–46% of the whole run. Same item/run/glyph counts in A and B, ~18× more time per item. This is worth fixing upstream (keep a cursor inGlyphRunIterinstead ofskip, and/or slimCluster); it dwarfs everything else on this page. -
Is Add patch for naga to fix macOS support #18's regression visible on a real page? Barely. Comparing B with D (same Parley base, ± Add patch for naga to fix macOS support #18): the parley construction stage (
build_inline_layout_into) is 524 vs 517 ms over 8 layouts (~+1%),resolve_style_metricscosts 16.5 ms per 8 layouts (~2 ms per full page), and heap grows 84.6 → 91.1 MB (+6.5 MB live, plausibly the per-layoutstyle_metrics/ per-runbox_metricsdata kept for every inline context) — that last part is the most tangible cost of Add patch for naga to fix macOS support #18 here. Wall-time-wise B is actually faster than D (847 vs 1050 ms), but see 4. -
Does the cache recover it? Yes for the part it targets. In C,
fontique::Query::matches_withdrops from 102 ms to 21 ms (−80%) andresolve_style_metricsfrom 16.5 to 6.5 ms per 8 layouts; total construction time 524 → 484 ms (−8%, i.e. ~5 ms per full-page layout). Only 6 primary-font entries and 84 metrics entries are needed for the whole article, so the cache's own footprint is negligible: heap after layout is identical to B to within 10 kB (91.08 vs 91.07 MB), peak RSS 199.5 vs 199.4 MB. Relative to A's absolute numbers (~270 ms/page) those 5 ms are ~2%; relative to the Tango micro-benchmarks they line up with the 5–13% Latin wins there. -
Unexplained: B is ~20% faster end-to-end than both C and D, and perf puts all of that difference in
GlyphRunIter::next/memcpy (3 826 vs 4 929/4 980 ms), a code path neither add support for font-size #19 nor Blitz Implementvertical-alignvia Parley #832 touches (line.rs is byte-identical between B and C; Cluster/Run types are unchanged). Shaping and query time in C are lower than B, so this is not the cache being slow. The most likely explanation is codegen (inlining / copy-elision) differences in theflat_map().skip()chain between the three builds; I did not chase it further because the fix in point 1 makes the whole path irrelevant. Treat the B-vs-C wall-clock gap as noise from that path, and the stage-level numbers above as the signal.
Memory summary
| A | B | C | D | |
|---|---|---|---|---|
| peak RSS | 172.8 MB | 199.4 MB | 199.5 MB | 186.9 MB |
| live heap after layout | 77.5 MB | 91.3 MB | 91.3 MB | 84.7 MB |
Parley main vs 0.11.1: +7 MB; #18: +6.5 MB; #19 cache: +~10 kB (6 + 84 entries).
Caveats
- Images / web fonts are not loaded (only the HTML and its two stylesheets), so text is laid out in the system fallback fonts available on the VM; font-face-heavy pages would exercise the metrics cache more (here only 84 (font, size) pairs).
- The relayout path re-shapes all text (Blitz has no cheaper incremental relayout that keeps shaped text), so "steady-state relayout" ≈ "layout without parsing".
- D required 5 one-line API shims in Blitz
mainto compile against Parleymain(InlineBox::baseline,append_inline_box_to_linearity,Cluster::style_index, plus Implementvertical-alignvia Parley #832'sblitz-paint/src/text.rsdiff); none are on a timed path. - Nothing from this benchmark was committed to any PR branch; the worktrees/scratch example live only on the benchmark VM.
obama_bench.rs (scratch, not committed)
//! Scratch benchmark: parse + style + layout of a local HTML file with Blitz's headless path.
//!
//! Usage: obama_bench <path/to/page.html> [iterations] [warmups]
//! Outputs one JSON line with the timings.
use std::alloc::{GlobalAlloc, Layout, System};
use std::sync::Arc;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::time::Instant;
use blitz_dom::DocumentConfig;
use blitz_html::HtmlDocument;
use blitz_traits::net::{Bytes, NetHandler, NetProvider, Request};
use blitz_traits::shell::{ColorScheme, Viewport};
struct Counting;
static ALLOCATED: AtomicUsize = AtomicUsize::new(0);
static PEAK: AtomicUsize = AtomicUsize::new(0);
unsafe impl GlobalAlloc for Counting {
unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
let p = unsafe { System.alloc(layout) };
if !p.is_null() {
let cur = ALLOCATED.fetch_add(layout.size(), Ordering::Relaxed) + layout.size();
PEAK.fetch_max(cur, Ordering::Relaxed);
}
p
}
unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
ALLOCATED.fetch_sub(layout.size(), Ordering::Relaxed);
unsafe { System.dealloc(ptr, layout) }
}
unsafe fn realloc(&self, ptr: *mut u8, layout: Layout, new_size: usize) -> *mut u8 {
let p = unsafe { System.realloc(ptr, layout, new_size) };
if !p.is_null() {
if new_size >= layout.size() {
let cur =
ALLOCATED.fetch_add(new_size - layout.size(), Ordering::Relaxed) + new_size
- layout.size();
PEAK.fetch_max(cur, Ordering::Relaxed);
} else {
ALLOCATED.fetch_sub(layout.size() - new_size, Ordering::Relaxed);
}
}
p
}
}
#[global_allocator]
static GLOBAL: Counting = Counting;
/// Serves `file://` URLs synchronously from disk; ignores everything else (no network).
struct LocalFiles;
impl NetProvider for LocalFiles {
fn fetch(&self, _doc_id: usize, request: Request, handler: Box<dyn NetHandler>) {
if request.url.scheme() != "file" {
return;
}
let Ok(path) = request.url.to_file_path() else {
return;
};
let Ok(bytes) = std::fs::read(&path) else {
eprintln!("missing local file: {}", path.display());
return;
};
handler.bytes(request.url.to_string(), Bytes::from(bytes));
}
}
const WIDTH: u32 = 1280;
const HEIGHT: u32 = 800;
fn viewport(scale: f32) -> Viewport {
Viewport::new(WIDTH, HEIGHT, scale, ColorScheme::Light)
}
fn build(html: &str, base_url: &str, net: &Arc<LocalFiles>) -> HtmlDocument {
let mut doc = HtmlDocument::from_html(
html,
DocumentConfig {
base_url: Some(base_url.to_string()),
net_provider: Some(Arc::clone(net) as _),
viewport: Some(viewport(1.0)),
..Default::default()
},
);
// Stylesheets are delivered synchronously by `LocalFiles`; two resolves make sure
// any nested `@import`s delivered during the first are applied too.
doc.resolve(0.0);
doc.resolve(0.0);
doc
}
fn stats(mut v: Vec<f64>) -> (f64, f64) {
v.sort_by(|a, b| a.partial_cmp(b).unwrap());
let median = v[v.len() / 2];
let p90 = v[((v.len() as f64 * 0.9).ceil() as usize).min(v.len()) - 1];
(median, p90)
}
fn main() {
let path = std::env::args().nth(1).expect("html path");
let iterations: usize = std::env::args()
.nth(2)
.and_then(|s| s.parse().ok())
.unwrap_or(20);
let warmups: usize = std::env::args()
.nth(3)
.and_then(|s| s.parse().ok())
.unwrap_or(3);
let path = std::fs::canonicalize(&path).unwrap();
let html = std::fs::read_to_string(&path).unwrap();
let base_url = format!("file://{}", path.display());
let net = Arc::new(LocalFiles);
let heap_before = ALLOCATED.load(Ordering::Relaxed);
// (i) full parse + style + layout, repeated from scratch
let mut full = Vec::new();
let mut doc = None;
let mut heap_after_first_layout = 0;
let mut root_height = 0.0;
for i in 0..warmups + iterations {
let start = Instant::now();
let d = build(&html, &base_url, &net);
let elapsed = start.elapsed().as_secs_f64() * 1e3;
if i == 0 {
heap_after_first_layout = ALLOCATED.load(Ordering::Relaxed) - heap_before;
root_height = d.as_ref().root_element().final_layout().size.height;
}
if i >= warmups {
full.push(elapsed);
}
doc = Some(d);
}
let mut doc = doc.unwrap();
let heap_after_layout = ALLOCATED.load(Ordering::Relaxed) - heap_before;
// (ii) steady-state relayout: toggling the scale invalidates every inline (text)
// layout, rebuilds the stylist device (restyle) and re-runs layout.
let scales = [1.0_f32, 1.00001_f32];
let mut relayout = Vec::new();
for i in 0..warmups + iterations {
doc.as_mut().set_viewport(viewport(scales[i % 2]));
let start = Instant::now();
doc.resolve(0.0);
let elapsed = start.elapsed().as_secs_f64() * 1e3;
if i >= warmups {
relayout.push(elapsed);
}
}
let (full_med, full_p90) = stats(full);
let (re_med, re_p90) = stats(relayout);
#[allow(unused_mut)]
let mut cache = String::from("null");
#[cfg(font_cache)]
{
let (p, m) = doc.as_ref().font_ctx_cache_len();
cache = format!("[{p}, {m}]");
}
println!(
"{{\"first_layout_ms\": {{\"median\": {full_med:.2}, \"p90\": {full_p90:.2}}}, \
\"relayout_ms\": {{\"median\": {re_med:.2}, \"p90\": {re_p90:.2}}}, \
\"heap_after_first_layout_mb\": {:.2}, \"heap_after_layout_mb\": {:.2}, \
\"heap_peak_mb\": {:.2}, \"root_height\": {root_height:.1}, \"cache_entries\": {cache}}}",
heap_after_first_layout as f64 / 1e6,
heap_after_layout as f64 / 1e6,
PEAK.load(Ordering::Relaxed) as f64 / 1e6,
);
}
Contributor
WPT follow-up for Blitz #832 / Parley #18 (
|
| Repo | PR | Base | Content |
|---|---|---|---|
| parley | DioxusLabs/parley#20 (head 1fd63fba72b38df9ec14892bd6f20e96fe2c1176) |
devin/1788383056-vertical-align |
trailing/consecutive hanging spaces no longer start an extra line; Style::parent() + Line::style_baseline() made public; 2 regression tests + 5 snapshots |
| blitz | #833 | devin/1788386630-vertical-align |
decorations painted at the decorating box's baseline (blitz-paint/src/text.rs::decorating_box_baseline); parley repinned to 1fd63fba72b38df9ec14892bd6f20e96fe2c1176 |
Verification: parley cargo fmt --all / cargo test --workspace (172 parley_tests + others, all ok) / cargo clippy --workspace --all-targets -D warnings clean. Blitz cargo fmt --check, cargo check --workspace --all-targets, cargo clippy -p blitz-dom -D warnings clean. (cargo clippy -p blitz-paint -D warnings fails on a pre-existing needless_return at packages/blitz-dom/src/mutator.rs:1250, untouched by these PRs and also present on the #832 head.)
Per-directory pass counts
| Directory | main | #832 | #832 + fixes (#833) |
|---|---|---|---|
css/CSS2/linebox |
137/201 | 169/201 | 169/201 |
css/CSS2/text |
268/410 | 270/410 | 271/410 |
css/css-text |
982/1872 | 1006/1872 | 1021/1872 |
css/css-inline |
47/234 | 49/234 | 49/234 |
css/CSS2/visudet |
15/39 | 17/39 | 17/39 |
css/CSS2/normal-flow |
648/772 | 648/772 | 648/772 |
css/CSS2/css1 |
94/164 | 116/164 | 116/164 |
css/css-values |
229/496 | 232/496 | 232/496 |
| total | 2420/4188 | 2507/4188 | 2523/4188 |
#832 → #832+fixes: 19 tests FAIL→PASS, 3 PASS→FAIL (all three are white-space: break-spaces tests, see rows below — they passed on #832 only because of the wrong break-after-first-hanging-space behaviour, since Blitz maps break-spaces to pre-wrap; stylo_to_parley.rs:291).
Newly passing: CSS2/text/text-decoration-va-length-001.xht, css-text/white-space/pre-wrap-00{1..7}.html, pre-wrap-leading-spaces-015/017, white-space-pre-wrap-trailing-spaces-005/007/008/011/021, hanging-whitespace-004, trailing-space-align-start.tentative, line-break-anywhere-and-white-space-006, hyphens-auto-002.
Root causes fixed
- (a) Parley – hanging spaces (
parley/src/layout/line_break.rs:1060-1085, theis_space && Wrapoverflow branch). After appending an overflowing space the breaker always calledstart_new_line. Effects:"AAA "at end of layout → a second, empty line (every block ending in a space got a phantom line box;layout.height()doubled);"XXXX XX"in 4ch broke after the first space instead of hanging both. Fix:continuewhen the next atom is another U+0020, or when this is the last atom of the last item. NBSP deliberately not chained (Whitespace::Spaceonly; a broader variant chaining every space-class regressedwhite-space-intrinsic-size-001). Tests:trailing_whitespace_does_not_add_line,vertical_align_style_baseline. - (b) Blitz – decoration baseline (
packages/blitz-paint/src/text.rs,stroke_text). Decorations were positioned at each glyph run's shifted baseline; CSS 2 §16.3.1 positions them from the decorating box. Fix walksStyle::parent()from the run's style to the style whose brush is the decorating node and usesLine::style_baseline. Fixestext-decoration-va-length-001.
Investigated failing tests
Status columns: main → #832 → #832+fixes.
| Test | Class | Diagnosis | Fix / reason not fixed | Status |
|---|---|---|---|---|
css/CSS2/text/text-decoration-va-length-001.xht |
b | underline of <u> followed the vertical-align: <length> child's shifted glyph baseline (blitz-paint/src/text.rs, DecorationRunGeometry.baseline = glyph_run.baseline()) |
blitz#833 | F→F→P |
css/css-text/white-space/pre-wrap-001…007.html, white-space-pre-wrap-trailing-spaces-005/007/008/011/021, pre-wrap-leading-spaces-015/017, hanging-whitespace-004, trailing-space-align-start.tentative, CSS2/text/white-space-processing-002/003/008/009/010 (the last five only pass with line-height: normal → metrics, see question 1) |
a | second and following hanging spaces were pushed to a new line; trailing space at end of content created an empty line box (line_break.rs:1060) |
parley#20 | F→F→P (the white-space-processing-* five: F→F→F) |
css/css-text/line-break/line-break-anywhere-and-white-space-006.html, css/css-text/hyphens/hyphens-auto-002.html |
a | same phantom trailing line | parley#20 | F→F→P |
css/css-text/white-space/break-spaces-001.html, break-spaces-010.html, white-space-intrinsic-size-001.html |
c | white-space: break-spaces is mapped to Preserve/pre-wrap (stylo_to_parley.rs:289-291 TODO) so spaces hang instead of wrapping and don't count toward min-content; passed on #832 only via the bug fixed above |
needs a break-spaces mode in Parley (WhiteSpaceCollapse::BreakSpaces: spaces are break opportunities and never hang) |
P→P→F |
css/CSS2/visudet/line-height-204.html |
d (+ design question) | Test sets line-height on one div to getComputedStyle(other).height and requires equal baselines. Font Revalia.woff has positive hhea.descender = 382 (non-standard sign), which skrifa/parley read as descent −18.65px at 100px (ascent 104.8). Blitz maps line-height: normal to 1.2 × font-size = 120px (stylo_to_parley.rs:354), so the strut baseline is at 105+floor((120−86)/2)=122; the 20px inline-block sitting on the baseline extends the line box to 122px, the script then sets line-height: 122px → half-leading 18 → baseline 123 → 1px shift. Browsers compute normal from font metrics (ascent+descent+lineGap = 86px here), which the test relies on. |
not fixed: changing normal → LineHeight::MetricsRelative(1.0) was tried; net +2 (+8/−6: fixes white-space-processing-002/003/008/009/010, pre-wrap-008/009, block-in-inline-first-line-001; regresses inline-formatting-context-012, block-formatting-contexts-011, inline-block-width-002a/b, overflow-wrap-cluster-001/002) and doesn't fix this test either; it is a design decision for #832 owner (question 1) |
P→F→F |
css/CSS2/linebox/baseline-block-with-overflow-001.html |
b (undiagnosed detail) | inline-block whose only child is an `overflow: hidden | scroll | autoblock. Blitz uses taffy'sbaselines.last.or(first) (layout/inline.rs:~660); taffy 0.14 compute_block_layoutsynthesizes a baseline from a scroll-container child's border-box bottom, so the inline-block gets a baseline where CSS 2 §10.8.1 says "bottom margin edge" *only for the inline-block itself*. Pixel diff: first XX groups aligned, later groups off. Suppressing baselines ofoverflow != visibleblock containers inlayout/mod.rs` compiled but did not make the test pass, so the wrong detail is not yet isolated (could also be the inline-block's own margin handling). |
css/CSS2/linebox/vertical-align-negative-leading-001.html |
a (undiagnosed detail) | negative half-leading (line-height < content area); line-box blue spans matched the ref at sampled columns but glyph/inline-box y differs, so the strut is right and the shift/quantisation of the child is off. Not narrowed to a line. |
not fixed | F→F→F |
css/CSS2/linebox/empty-inline-002/003.xht, inline-box-001/002.xht, border-padding-bleed-001…003.xht, split-inline-borders.html, inline-formatting-context-002…007/022/023.xht |
c | need inline border/padding/margin boxes on non-atomic inlines (border: 25px solid on <span>, etc.); Blitz has no inline fragment boxes |
missing feature | F→F→F |
css/CSS2/linebox/vertical-align-baseline-003.xht, -004a.xht, -005a.xht, -006.xht |
b (not individually diagnosed) | all are inline-block baseline = last line box, with margins (-004a: margin-bottom: 99px, margin: 8px); the inline-block baseline path is inline.rs ibox.baseline = (margin.top + baseline) * scale; source read, screenshots not compared |
not fixed | F→F→F |
css/CSS2/linebox/vertical-align-baseline-009.xht, vertical-align-applies-to-007/015.xht, line-height-applies-to-015.xht |
c | display: table-caption / table-cell vertical-align and table layout |
tables unsupported | F→F→F |
css/CSS2/linebox/vertical-align-117a/118a.xht, -121.xht |
b/a (not individually diagnosed) | text-bottom with margin: auto -1em on inline-blocks (117a/118a); top/bottom inline-blocks with border-bottom (121); source read only |
not fixed | F→F→F |
css/CSS2/linebox/line-height-125.xht |
c | padding-top on an inline <fontsize> element (inline padding box) |
missing feature | F→F→F |
css/CSS2/linebox/inline-negative-margin-001.html |
b (not diagnosed) | inline-block with margin-left: 10ch |
not fixed | F→F→F |
css/CSS2/visudet/inline-block-baseline-001…006/011/014/016 |
b (not individually diagnosed) | inline-block baseline family; also fail on main | not fixed | F→F→F |
css/CSS2/visudet/line-height-205.html |
d | same Revalia / line-height: normal issue as -204 |
see question 1 | F→F→F |
css/CSS2/visudet/content-height-004/005.html, replaced-elements-*.html |
c | replaced-element sizing, not line-box geometry | out of scope | F→F→F |
css/css-inline/alignment-baseline/alignment-baseline-horizontal-002/003.html, alignment-baseline-nested-001/002.html |
c | need central/ideographic/mathematical/alphabetic/hanging baseline tables (BASE table / synthesized) — only `baseline |
middle | text-top |
css/css-inline/alignment-baseline/*-vertical-rl-*, dominant-baseline/dominant-baseline-mixed-writing-modes-002.html |
c | vertical writing modes | out of scope (the latter's main PASS was a false positive: unsupported writing-mode rendering happened to match) | P→F→F / F→F→F |
css/css-inline/baseline-shift/baseline-shift-sub-super.html, baseline-shift-length-percentage.html |
d / c | sub/super use WebKit's font-size/5, /3; the ref is built with explicit lengths matching the reference browser's constants. Percentage test also uses baseline-source. A spec-neutral alternative would be the font's OS/2 subscriptYOffset/superscriptYOffset, which is what css-inline-3 suggests when available |
not changed (constants are a design choice, see question 2) | F→F→F |
css/css-inline/baseline-shift/*-svg.html, baseline-shift-top/bottom/center-svg |
c | SVG text | out of scope | F→F→F |
css/css-inline/empty-span-size-001/002.html |
c | empty <span> with padding/border (inline padding box); 002's main PASS was accidental (no vertical shift applied to the empty box) |
missing feature | F→F→F / P→F→F |
css/css-values/lh-rlh-on-root-001.html, lh-unit-003/004.html |
c | lh/rlh units with line-height: normal need the metrics-based normal (Stylo resolves lh from the computed value; normal has none) |
see question 1 | F→F→F |
css/CSS2/normal-flow/inline-block-zorder-002/004.xht |
b (undiagnosed) | overlapping red/green inline-blocks; looked at paint order, no explanation found | not fixed | F→F→F |
Not investigated: failures whose feature flags/contents are floats, direction/writing-mode, calc, script-driven, text-box-trim, ruby, initial-letter, hanging-punctuation, text-wrap balance/pretty, hyphenation, text-autospace, letter-spacing, text-align/justify.
Questions for the #832 design
line-height: normal= 1.2 × font-size (stylo_to_parley.rs:330,354, kept from main). Browsers use ascent+descent+lineGap of the primary font. Switching toparley::LineHeight::MetricsRelative(1.0)is net +2 on the eight directories (list above) and is whatline-height-204/205,lh-unit-003/004,white-space-processing-*assume, but it changes the height of every block of text in Blitz. Not done here; flagged.sub/superconstants (style_metrics.rs:255-258): WebKit/Blink constants chosen; Firefox uses the font's OS/2 sub/superscript offsets.baseline-shift-sub-super.htmlcannot pass with either unless the reference browser's choice is matched.text-top/text-bottomuse the parent's font ascent/descent (content area) — matches spec; no test was found failing because of this.
Not fixed but in scope (honest gaps)
baseline-block-with-overflow-001, vertical-align-negative-leading-001, the inline-block baseline families (vertical-align-baseline-003…006, visudet/inline-block-baseline-*, vertical-align-117a/118a/121, inline-negative-margin-001), and inline-block-zorder-002/004 were opened and screenshot-compared (or source-read as marked) but the wrong line was not isolated within this session; they were not changed.
This was referenced Sep 3, 2026
Contributor
Blitz #832 (
|
| build | what | purpose |
|---|---|---|
main |
Blitz main, parley 0.11.1 (crates.io, 92b63a2) |
baseline |
main+parley-main |
Blitz main + parley main (83acf8f, the merge-base of parley#18) with API-only shims |
isolates the dependency bump from the vertical-align logic |
#832 |
Blitz #832 head a875dd47 |
reproduce |
All 51 regressions reproduce locally on #832 vs main (statuses in the table below are local; the "fixed" column is the head of blitz#834 + parley#22).
Key finding: 29 of the 51 also fail on main+parley-main, i.e. they are caused by the 70-commit parley bump (0.11.1 → 83acf8f: #734 NBSP preservation, #738 spacing model, #697 CSS line-box sizing, #743 negative half-leading, …) and not by parley#18 / blitz#832's vertical-align code. The remaining 22 are attributable to the vertical-align branches; of those, 9 were genuine bugs and are fixed, 13 are pre-existing gaps that the old always-tall strut / ignored-inline-box-baseline behaviour masked.
PRs opened
- Parley: Fix line-box content detection for negative-height boxes, NBSP width handling and inline box quantization parley#22 (
devin/1788404063-vertical-align-fixes→devin/1788383056-vertical-align, head62c35980d58afa6e13fc1caa148fa346c086f530), 3 regression tests inparley/src/tests/test_style_metrics.rs. - Blitz: Fix atomic-inline baseline export, empty line boxes and root decoration baseline #834 (
devin/1788404083-vertical-align-fixes→devin/1788386630-vertical-align), repins parley to62c35980….
Validation: parley cargo fmt/test/clippy -D warnings clean; blitz cargo check --workspace --all-targets, cargo clippy -p blitz-dom -D warnings, cargo fmt --check clean. Full css/ WPT run of the fixed head vs pristine #832: 18 FAIL→PASS, 1 PASS→FAIL (css/css-contain/contain-size-064.html, see "Exposed gap" below), no subtest regressions. The 8 non-listed improvements (absolute-non-replaced-max-height-002, multi-line-row-flex-fragmentation-081a–d-print, svg-g-no-size-container, flex-direction-row-reverse, flexbox-align-self-vert-003) all come from the negative-margin / quantization fixes to add_inline_box.
Shared root causes (clusters)
| cluster | shared cause | class |
|---|---|---|
| NBSP (zorder-002/004/005, inline-table-zorder-001, t41-html4-keywords-a, contain-animation-001, highlight-001/002, empty-span-size-002) | parley #734 stopped collapsing U+00A0 but is_space_or_nbsp() still classified it as trailing/hangable space in calculate_content_widths (parley/src/layout/data.rs:373), so -only shrink-to-fit boxes had 0 min/max-content width. On main old parley collapsed NBSP to empty text, which sent Blitz down the collapse-through path (blitz-dom/src/layout/inline.rs:192-240) producing 0×0 boxes — the "passing" tests were test==ref both blank. |
a (width) / c (what the width now exposes) |
| Atomic-inline baseline export (baseline-of-scrollable-1a/1b, single-axis, contain-layout-baseline-001/005, grid-baseline-, grid-inline-, list-and-block-in-inline, font-family-applies-to-005, flexbox-baseline-multi-line-vert-002) | #832 started honouring InlineBox::baseline from taffy. Blitz's own mapping (inline.rs:300-360) was wrong for scroll containers / contain:layout (fixed, b). The rest are taffy 0.14 limits: block containers only export a first baseline (taffy/src/compute/block.rs:712 Baselines::from_first) and child scroll-container baselines are clamped, not synthesized (block.rs:1501); grid only measures item baselines when ≥2 items baseline-align (grid/track_sizing.rs:496-530). On main the baseline was ignored (box bottom sat on the baseline), which happened to coincide with the expected result. |
b + c |
| Root strut now always present (contain-size-flexbox-002, intrinsic-percent-replaced-008/dynamic-008, abs-pos-with-replaced-child, ruby-overhang-spaces-022) | Old inline layout dropped the strut on lines containing only inline boxes; #18 applies CSS2 §10.8.1 strut (style-table entry 0) correctly. Tests that relied on the missing strut (vertical writing modes, scrollbars, ruby) now expose the actual unsupported feature. | c |
| Ruby (ruby-003, br-clear-all-000, empty-ruby-, ruby-base-container-) | display: ruby* is not implemented — construct.rs has no DisplayInside::Ruby* arm so <ruby>/<rt> fall through to inline flow (construct.rs:599), with rt { font-size:50%; line-height:1 } from default.css:1009-1015. Bump-attributable: parley #697 CSS line-box sizing made <rt> text contribute its own inline box metrics; on main per-inline metrics were ignored. |
c |
| White-space / hanging (hanging-punctuation-first-002, full-width-leading-spaces-004, hanging-whitespace-002, trailing-ideographic-space-013/014, text-decoration-skip-spaces-001) | U+3000 and other Zs are Whitespace::None (parley_engine/src/shape/data.rs:239-250); hanging glyphs / hanging-punctuation are unimplemented. Bump-attributable (#738 spacing model stops mutating advances; these previously passed through accidental 0-advance handling). |
c |
| Floats (floats-141 fixed; floats-149) | floats-141: add_inline_box treated a finite negative-height box (negative margins) as no-content (a). floats-149: parley #697 line box sizing + float placement of an empty float in an inline with margin-left; not VA. |
a / c |
| lh/rlh, transforms, small-caps, line-height-204 | bump-attributable, see rows. | c |
Per-test table
Status columns: main / #832 (a875dd47) / fixed (blitz#834 head); x/y = passing subtests.
| # | test | class | root cause | fix | main → #832 → fixed |
|---|---|---|---|---|---|
| 1 | css/CSS2/floats-clear/floats-141.xht | a | Two parley#18 bugs in LineBoxMetrics::add_inline_box (parley/src/layout/line_break.rs:296): (1) a box whose ascent+descent <= 0 (negative margin-top atomic inline) was treated as no content → strut dropped, line collapsed; (2) with quantize, ascent and descent were rounded separately, reserving 1px more than the box's own (integral) height when the baseline split it into .5 halves. |
parley#22 | PASS → FAIL → PASS |
| 2 | css/CSS2/floats-clear/floats-149.xht | c | Fails on main+parley-main (not VA). Empty float:left inside an inline with margin-left:40px gets placed on line 1 after parley #697 line-box sizing changed where the float anchor is measured (inline.rs float placement via break_remaining). Old behaviour passed by accident. |
won't fix: parley bump / float placement, unrelated to vertical-align | PASS → FAIL → FAIL |
| 3 | css/CSS2/fonts/font-family-applies-to-005.xht | c | Inline-block with 3 block lines: taffy 0.14 block.rs:1493-1513 only computes Baselines::first, last is None, so inline.rs:337-350 falls back to the first-line baseline (14px) instead of the last (52.4px, CSS2 §10.8.1) → line grows to 59px and box shifts 1–2px. Main ignored the baseline (bottom edge on baseline = same as last-line for this test). Verified with dbg: first=Some(14), last=None. Not a font-fallback issue. |
won't fix: needs taffy last-baseline for block containers | PASS → FAIL → FAIL |
| 4 | css/CSS2/linebox/baseline-block-with-overflow-001.html | c | Confirmed #832's explanation: inline-block with overflow:hidden on an inner block child; taffy block.rs:1501 clamps the child scroll-container baseline instead of synthesizing from its bottom margin edge (css-align-3 §9.1). Main passed because the inline-block's baseline was ignored. |
won't fix: taffy | PASS → FAIL → FAIL |
| 5 | css/CSS2/normal-flow/inline-block-zorder-002.xht | c | NBSP cluster. -only positioned block was 0×0 on main (collapse-through, inline.rs:192-240) so test and ref were both blank. Now it has size and paints in tree order after the inline-block: blitz-paint/src/render.rs:965 draw_children has no CSS2 Appendix E phase separation (in-flow block backgrounds must paint before atomic-inline content). |
won't fix: pre-existing paint-order gap | PASS → FAIL → FAIL |
| 6 | css/CSS2/normal-flow/inline-block-zorder-004.xht | a | NBSP cluster: calculate_content_widths (data.rs:373) trimmed NBSP as trailing whitespace → inline-block collapsed to 0 width. |
parley#22 | PASS → FAIL → PASS |
| 7 | css/CSS2/normal-flow/inline-block-zorder-005.xht | a | same as #6 | parley#22 | PASS → FAIL → PASS |
| 8 | css/CSS2/normal-flow/inline-table-zorder-001.xht | c | same mechanism as #5 (inline-table instead of inline-block); Appendix E ordering. | won't fix: paint order | PASS → FAIL → FAIL |
| 9 | css/CSS2/text/text-decoration-va-length-001.xht | b | blitz-paint/src/text.rs flush_line_decorations (≈:395): a decorating inline root with no run of its own fell back to the first descendant run's geometry, including its vertical-align-shifted baseline, so the underline moved with the shifted child. Now uses the line baseline for the root. (#832 called this pre-existing; it is a fixable painting bug.) |
blitz#834 | PASS → FAIL → PASS |
| 10 | css/CSS2/visudet/line-height-204.html | c | Fails on main+parley-main. Revalia.woff has a positive hhea.descender; parley_engine FontMetrics::from_skrifa negates it → descent −18.65. parley #743 (preserve negative half-leading) makes the strut over=122/under=−2, so used height ≠ line-height: normal; the test's explicit 122px then puts the baseline 1px lower. Not #18. |
won't fix: font-metric sign handling in parley_engine (parley bump) | PASS → FAIL → FAIL |
| 11 | css/css-align/baseline-of-scrollable-1a.html | b | inline.rs:300-350 exported the content baseline of an overflow:hidden inline-block. css-align-3 §9.1: a block-axis scroll container's baseline is its bottom margin edge. Now is_block_axis_scroll_container ⇒ exports_baseline = false (bottom-edge fallback). |
blitz#834 | PASS → FAIL → PASS |
| 12 | css/css-align/baseline-of-scrollable-1b.html | c | Scroll container is a child block of the inline-block; taffy block.rs:1501 clamps its baseline to the border box instead of synthesizing from the bottom margin edge. Main passed because inline-block baseline was ignored. |
won't fix: taffy | PASS → FAIL → FAIL |
| 13 | css/css-align/baseline-of-single-axis-scroll-container.html | c | overflow-y: clip; overflow-x: scroll is fixed up by stylo to overflow-y: hidden (css-overflow-3 §3.1 computed-value rule), so inline.rs:325 sees a two-axis scroll container and synthesizes the bottom-edge baseline; Chrome treats the clip axis as non-scrolling. Needs the pre-fixup value from stylo. |
won't fix: needs stylo-level info | 1/2 → 0/2 → 0/2 |
| 14 | css/css-break/ruby-003.html | c | Ruby cluster; fails on main+parley-main. <rt> laid out as inline text with own metrics (parley #697) makes the ruby line 2 lines tall; multicol + ruby unsupported. |
won't fix: ruby unimplemented | PASS → FAIL → FAIL |
| 15 | css/css-color/t41-html4-keywords-a.xht | a | NBSP cluster: swatches are -only inline-blocks that collapsed to 0 width. Geometry, not colour parsing. |
parley#22 | PASS → FAIL → PASS |
| 16 | css/css-contain/contain-animation-001.html | c | contain: size unimplemented (no Contain::SIZE handling in blitz-dom/taffy). Test passed on main only because the child was 0×0, giving the 100×100 the ref expects by accident. |
won't fix: contain:size unsupported | PASS → FAIL → FAIL |
| 17 | css/css-contain/contain-layout-baseline-001.html | b | contain: layout suppresses baseline export (css-contain-1 §3.2); inline.rs exported the content baseline. Now contain_layout ⇒ exports_baseline = false. |
blitz#834 | PASS → FAIL → PASS |
| 18 | css/css-contain/contain-layout-baseline-005.html | b | same as #17 (fieldset variant). | blitz#834 | PASS → FAIL → PASS |
| 19 | css/css-contain/contain-size-flexbox-002.html | d/c | Chrome/Firefox pass only because overflow:scroll adds a ≈15px scrollbar making both the flex item and the inline-block ≥16px tall. Blitz has 0-width scrollbars, so flex row = 2px vs inline-block row = 16px (strut). Passed on main because the old inline layout omitted the strut on inline-box-only lines (a bug). |
won't fix: depends on scrollbar geometry | PASS → FAIL → FAIL |
| 20 | css/css-flexbox/flexbox-baseline-multi-line-vert-002.html | c | Test side is correct (first baseline 13). The ref is misrendered: .flexContainer > * { display:inline-block; width/height:20px } matches <br>, and construct.rs:981/1192 only special-cases <br> when its display is inline, so the <br> becomes a 20×20 atomic inline and a 2nd line. Masked on main because inline-box baselines were ignored. |
won't fix: <br> with non-inline display (pre-existing) |
PASS → FAIL → FAIL |
| 21 | css/css-flexbox/flexbox_inline.html | b | Atomic inline with negative vertical margins: inline.rs:357 clamped the reserved height to ≥0 and inline.rs:875 positioned with margin.top.max(0), so the box no longer sat on the baseline that parley#18 now honours. Baseline-bearing boxes now use the (possibly negative) margin box. |
blitz#834 | PASS → FAIL → PASS |
| 22 | css/css-fonts/small-caps-letter-spacing-002.html | c | Fails on main+parley-main; unrelated to vertical-align. Most likely parley #738 (spacing model) / #731 (ligature suppression with letter-spacing) changed spacing of synthesized small-caps; not narrowed further. |
won't fix: parley bump, not investigated beyond attribution | PASS → FAIL → FAIL |
| 23 | css/css-grid/alignment/grid-baseline-001.html | c | Baseline cluster: taffy 0.14 grid only measures item baselines when ≥2 items participate in baseline alignment (grid/track_sizing.rs:496-530), else container baseline = first item's bottom (grid/mod.rs:853). Now that the inline-grid's baseline is used by parley, the wrong one shows. |
won't fix: taffy | PASS → FAIL → FAIL |
| 24 | css/css-grid/alignment/grid-baseline-002.html | c | same as #23 | won't fix: taffy | PASS → FAIL → FAIL |
| 25 | css/css-grid/alignment/grid-inline-baseline.html | c | same as #23 | won't fix: taffy | PASS → FAIL → FAIL |
| 26 | css/css-grid/grid-items/grid-inline-items-002.html | c | same as #23 | won't fix: taffy | PASS → FAIL → FAIL |
| 27 | css/css-highlight-api/painting/custom-highlight-painting-inheritance-001.html | c | Highlight API painting unimplemented. Test & ref both contain -only spans which were 0-width on main, so neither showed green → accidental match. |
won't fix: Highlight API unsupported | PASS → FAIL → FAIL |
| 28 | css/css-highlight-api/painting/custom-highlight-painting-inheritance-002.html | c | same as #27 | won't fix | PASS → FAIL → FAIL |
| 29 | css/css-inline/dominant-baseline/dominant-baseline-mixed-writing-modes-002.html | c | Confirmed #832's explanation: ref uses alignment-baseline: text-top (now mapped in stylo_to_parley.rs:301-303) while the test uses dominant-baseline (unsupported) plus vertical writing modes (unsupported). |
won't fix | PASS → FAIL → FAIL |
| 30 | css/css-inline/empty-span-size-002.html | b | Parley (since the bump, incl. #18) lays out empty text as one strut-height line; Blitz only short-circuited empty content when the block had no border/padding (inline.rs:192-240), so a bordered block with only empty spans got a 19px line. Fix: has_inline_content == false ⇒ height 0, no baselines (CSS2 §9.4.2). (#832 called this pre-existing; it is fixable in Blitz.) |
blitz#834 | PASS → FAIL → PASS |
| 31 | css/css-lists/list-and-block-in-inline.html | c | Same as #3: taffy block containers export only the first baseline (block.rs:712); an inline-block's baseline is its last line box. Main matched by accident (bottom edge). |
won't fix: taffy last-baseline | PASS → FAIL → FAIL |
| 32 | css/css-ruby/br-clear-all-000.html | c | Ruby cluster (bump). <rt> as inline text with own line-height contributes extra height; clear + ruby annotation overflow into padding not modelled. |
won't fix: ruby unimplemented | 1/1 → 0/1 → 0/1 |
| 33 | css/css-ruby/empty-ruby-base-container.html | c | Ruby cluster (bump): a<ruby><rt>b</rt></ruby>c with line-height:3 — <rt> (50% font, line-height:1) is an inline with its own metrics, ref has no annotation. |
won't fix | PASS → FAIL → FAIL |
| 34 | css/css-ruby/empty-ruby-text-container-abs.html | c | Ruby cluster (bump), as #33. | won't fix | PASS → FAIL → FAIL |
| 35 | css/css-ruby/empty-ruby-text-container-float.html | c | Ruby cluster (bump), as #33. | won't fix | PASS → FAIL → FAIL |
| 36 | css/css-ruby/ruby-base-container-abs.html | c | Ruby cluster (bump), as #33 (+ NBSP now has width). | won't fix | PASS → FAIL → FAIL |
| 37 | css/css-ruby/ruby-base-container-float.html | c | Ruby cluster (bump), as #33. | won't fix | PASS → FAIL → FAIL |
| 38 | css/css-ruby/ruby-overhang-spaces-022.html | c | Ruby unsupported: <rt> laid out as inline text contributes its own inline-box metrics to the line (25px vs 24px in ref) now that #18 honours per-inline metrics. |
won't fix | PASS → FAIL → FAIL |
| 39 | css/css-sizing/intrinsic-percent-replaced-008.html | c | Vertical writing modes unsupported. Abs-pos block containing only a canvas: with a proper root strut the line is canvas + strut descent (104px) instead of 100. Chrome gets 100 because in vertical-lr the strut extends the inline size. Main passed because the strut was dropped on inline-box-only lines. |
won't fix: vertical writing modes | PASS → FAIL → FAIL |
| 40 | css/css-sizing/intrinsic-percent-replaced-dynamic-008.html | c | same as #39 | won't fix | PASS → FAIL → FAIL |
| 41 | css/css-text-decor/text-decoration-skip-spaces-001.html | c | White-space cluster (bump): U+2000–U+200A/U+205F/U+3000/U+1680 are Whitespace::None (parley_engine/src/shape/data.rs:239-250), so they neither hang nor are skipped by the underline; text-decoration-skip not implemented. |
won't fix | PASS → FAIL → FAIL |
| 42 | css/css-text/hanging-punctuation/hanging-punctuation-first-002.html | c | hanging-punctuation unimplemented; U+3000 is Whitespace::None. Bump-attributable. |
won't fix | PASS → FAIL → FAIL |
| 43 | css/css-text/white-space/full-width-leading-spaces-004.html | c | U+3000 not treated as hangable space for min-content sizing (data.rs:373 only knows Space/NBSP). Bump-attributable. |
won't fix | PASS → FAIL → FAIL |
| 44 | css/css-text/white-space/hanging-whitespace-002.tentative.html | c | pre-wrap spaces followed by an unconditionally-hanging U+3000 must hang; U+3000 isn't a hanging glyph in parley. Bump-attributable. | won't fix | PASS → FAIL → FAIL |
| 45 | css/css-text/white-space/trailing-ideographic-space-013.html | c | Trailing U+3000 must hang; Whitespace::None. Bump-attributable. |
won't fix | PASS → FAIL → FAIL |
| 46 | css/css-text/white-space/trailing-ideographic-space-014.html | c | same as #45 | won't fix | PASS → FAIL → FAIL |
| 47 | css/css-text/white-space/white-space-zero-fontsize-001.html | c | Parley emits a phantom line for a block-final \n (line_break.rs pushes an empty run for the trailing newline; the height correction never fires because item_range is non-empty — same on parley main). Browsers create no line box for a block-final forced break. On main both test (2 lines) and ref (3 lines) were 45px due to the old copy-previous-line hack; #18 (1c93d69) correctly sizes the phantom line by the newline's font-size:0 style so the test is 31px (= browsers) but the ref is still 45px. |
won't fix here: parley block-final-newline line box (pre-existing) | PASS → FAIL → FAIL |
| 48 | css/css-transforms/3d-point-mapping-2-transforminterop.html | c | 5/12 → 2/12 on main+parley-main too: elementFromPoint through 3D transforms; the 3 lost subtests depend on text line geometry changed by parley #697/#743, not on vertical-align (stale-baseline hit-testing ruled out: identical on #832 and fixed head). |
won't fix: transform hit-testing (bump) | 5/12 → 2/12 → 2/12 |
| 49 | css/css-values/lh-rlh-on-root-001.html | c | 4/8 → 2/8 on main+parley-main too. lh/rlh on the root resolve against the root's line-height: normal, whose value changed with parley #743 (negative half-leading preserved). Not #18. |
won't fix: parley bump / root lh resolution | 4/8 → 2/8 → 2/8 |
| 50 | css/css-viewport/zoom/vertical-align.html | c | stylo 0.20: inherit on a reset zoom-dependent property copies the parent's computed value verbatim (properties.mako.rs inherit_*; cascade.rs:1134-1141 only rescales inherited props). Firefox fails this test too (wpt.fyi). Masked on main because vertical-align lengths weren't implemented at all. |
won't fix: stylo | PASS → FAIL → FAIL |
| 51 | css/css-writing-modes/abs-pos-with-replaced-child.html | c | same mechanism as #39 (vertical writing mode; root strut adds descent under the replaced element). | won't fix | PASS → FAIL → FAIL |
Totals: a 4 (floats-141, zorder-004, zorder-005, t41) · b 6 (text-decoration-va-length-001, baseline-of-scrollable-1a, contain-layout-baseline-001/005, flexbox_inline, empty-span-size-002) · c 40 · d 1 (contain-size-flexbox-002, also c). 10 of 51 fixed; 29 of the remaining 41 are the parley dependency bump rather than #832/#18.
Exposed gap introduced by the fixes
css/css-contain/contain-size-064.html PASS → FAIL on the fixed head. contain: size is unimplemented; the test passed only because its -only grid cells had zero min-content width on both test and ref. With the (spec-correct) NBSP width fix the ref shows the green cells the test can't produce. Same mechanism as #16 (contain-animation-001). Flagging rather than reverting: the NBSP fix repairs 3 listed tests + 8 others.
Things deliberately not changed / questions
- Line-break-time NBSP handling (
parley/src/layout/line_break.rs:944 is_space_or_nbsp()) was left as-is. Changing it to== Spacemakes 8 more tests pass (letter-spacing-justify-001,text-justify-and-trailing-spaces-001…004,white-space-nowrap-011,word-break-break-all-018/021) but regressesline-break-anywhere-overrides-uax-behavior-004/006. Worth a separate parley PR with an investigation of theline-break: anywhereinteraction. - No vertical-align semantics were changed (root strut = style entry 0, parent-relative alignment, compound
VerticalAlign { alignment, shift }all intact). Thequantizeparameter ofLineBoxMetrics::add_inline_boxis now unused (kept to avoid changingappend_inline_box_to_line's public signature, whichparley_tests/tests/floats.rs:183calls) — question: should the publicquantizearg be dropped in a follow-up? - Implement
vertical-alignvia Parley #832's description explanations:baseline-block-with-overflow-001anddominant-baseline-mixed-writing-modes-002are confirmed pre-existing gaps;text-decoration-va-length-001andempty-span-size-002were actually fixable in Blitz and are fixed in Fix atomic-inline baseline export, empty line boxes and root decoration baseline #834.
…seline - Inline-blocks that are block-axis scroll containers export no baseline (bottom margin edge is used); other atomic inlines (flex/grid/table) keep their content baseline, clamped to the border box when scrollable. contain: layout boxes export no baseline. - Atomic inlines with a baseline may reserve negative space in the line (negative vertical margins) and are positioned from that baseline. - An inline root with no text and no inline boxes is a zero-height line box and exports no baselines. - Text decorations on the inline root fall back to the line baseline rather than the (vertical-align shifted) first run's baseline. - Repin parley to ac4c61c33c5bbc63303570b336b4e27cfa0cdd83.
… NBSP width fixes)
staging-devin-ai-integration Bot
pushed a commit
that referenced
this pull request
Sep 3, 2026
…faces without U+0020)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Companion to DioxusLabs/parley#18 (Parley now builds the CSS inline alignment tree itself: root strut, ancestor inline boxes, parent-relative baseline shifts,
top/bottomaligned subtrees).Cargo.tomlpins Parley to that branch's head (6268e53); this should be repointed once #18 lands.stylo_to_parley::vertical_align: Stylo storesvertical-alignas the css-inline-3 longhands, and Parley'sVerticalAlign { alignment, shift }is the same compound, so both are mapped losslessly:alignment-baseline(baseline|text-top|text-bottom|middle) →parley::AlignmentBaseline,baseline-shift(sub|super|top|bottom|<length-percentage>) →parley::BaselineShift. Percentages resolve against the element's own line-height.baseline-shift: centerhas no Parley equivalent yet and is approximated asmiddle;baseline-sourceis ignored.TextStyle::vertical_alignon every span andInlineBox::vertical_alignon atomic inlines.InlineBox::baselineis now populated for inline-blocks from taffy'sbaselines.last.or(first)(+ top margin), orNone(bottom margin edge) whenoverflowisn't visible / there are no line boxes (CSS 2 §10.8.1). Inline layout now also reportsbaselines.lastupward.root_line_heightfloor on each span'sline-height— Parley adds the strut and every ancestor span's own box to the line, so the floor was double counting and broke shifted spans.append_inline_box_to_line(.., NEG_INFINITY, NEG_INFINITY, ..)so they contribute no height.Cluster::style_index()replacesGlyph::style_index();Run::font_metrics(),run.font().font,NormalizedCoord::to_bits().WPT (
css/CSS2/linebox css/css-inline css/CSS2/text)454 → 490 tests passing, 1198 → 1299 subtests, 0 crashes. Four previously-passing tests now fail; all are pre-existing gaps that the old always-tall strut happened to hide:
text-decoration-va-length-001.xht— Blitz paints decorations per glyph run at the run's (now shifted) baseline; they should sit at the decorating ancestor's baseline. Needs decoration propagation in blitz-paint.linebox/baseline-block-with-overflow-001.html— taffy block layout propagates a child's baseline through anoverflow: hiddenblock; should synthesize from the bottom margin edge (taffy issue).css-inline/empty-span-size-002.html— Blitz doesn't model inline border/padding boxes, so an empty bordered span doesn't make the line visible.dominant-baseline-mixed-writing-modes-002.html— vertical writing modes /dominant-baselineunsupported.WPT results
Subtests: 321 newly passing, 48 newly failing (net +273).
Full diff (272 changed tests)
Generated by the WPT workflow.
Link to Devin session: https://dioxus.staging.devinenterprise.com/sessions/da7669e341814cc7976e4ebf32b7eacb
Open in Devin Desktop: https://dioxus.staging.devinenterprise.com/desktop/session/da7669e341814cc7976e4ebf32b7eacb?variant=devin-insiders
Requested by: @nicoburns