* [raster] Fix background color premultiplication and stride handling
ensure_initialized()'s background fill had two bugs:
1. The background pixel was built with HB_COLOR(blue, green, red,
alpha), which packs channels through HB_TAG's bit layout
(b<<24 | g<<16 | r<<8 | a). Every other pixel in this buffer is
premultiplied BGRA32 (b | g<<8 | r<<16 | a<<24, see hb-raster.hh's
SRC_OVER comment: "premultiplied src over premultiplied dst"), so
the background was both unpremultiplied and in the wrong byte
positions: a half-transparent red background (blue=0, green=0,
red=255, alpha=128) filled the buffer with 0x0000ff80 instead of
the correct premultiplied 0x80800000.
2. The fill then wrote that pixel into a contiguous run of
width * height uint32_t slots starting at the buffer origin,
assuming every row is packed at exactly width * 4 bytes. Once a
caller sets a stride wider than that (row padding, alignment,
embedding the glyph image inside a larger atlas), the fill drifts
across row boundaries after the first row, leaving later rows
partially uninitialized or wrongly colored.
Build bg with color_to_premul_pixel(), the same helper solid fills
already use, and fill row by row using the actual stride.
Added two regression tests in test/api/test-raster.cc:
test_background_premultiplied (default stride, half-alpha background,
checks the stored pixel is premultiplied) and test_background_stride
(stride much wider than width * 4, opaque background, checks every
row).
Assisted-by: Claude
* [raster] Handle unaligned background strides
A caller-supplied stride does not have to be word-aligned, so casting
each row to uint32_t * can perform misaligned stores.
Use hb_packed_t<uint32_t> for the row writes, and exercise an odd
stride in the regression test with byte-wise pixel checks.
Tests: meson test -C build
Assisted-by: Codex <noreply@openai.com>
---------
Co-authored-by: Arthur031221 <levi74108520963@gmail.com>
reference_ft_library() returns an owned FT_Library reference that the
success paths hand to the FT_Face finalizer or the blob user-data
destructor. Four early returns in hb_ft_face_create_from_file_or_fail,
hb_ft_face_create_from_blob_or_fail and hb_ft_font_set_funcs returned
without releasing it, pinning the static library on invalid input and
OOM. Release it via destroy_ft_library() on each path.
Sweep-gradient color-stop angles come from the font, so the
interpolation fraction t can fall outside [0,1] (or be NaN). The
interpolated channel was then cast float -> unsigned directly; a
negative value (e.g. -387.671) is UB (UBSan float-cast-overflow) and
yields meaningless colors on release builds.
Clamp the interpolated channel to [0,255], mapping NaN to 0, before
the cast. In-range behavior is unchanged. Add the reported fuzzer
input as a seed.
Reported by iceray-Li.
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
AxisValueMap::instantiate() renormalizes the from/to coordinates and
stores them with set_float() without clamping. With malformed input
the result can fall outside [-1,1] (e.g. 2.192 -> 35923), overflowing
the int16 F2DOT14 storage: UB under UBSan's float-cast-overflow, and a
silently truncated avar table otherwise.
Clamp both coordinates to [-1,1] before storing. Add the reported
fuzzer input as a seed.
Reported by iceray-Li.
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
* [gpu] Check overflow when summing paint sub-blob sizes
hb_gpu_paint_encode() accumulated the lengths of all sub-blobs into
a 32-bit unsigned without an overflow check, then allocated the
output from the wrapped total. A COLRv1 font whose paint DAG
revisits a max-size (65535-stop) gradient 4096 times makes the sum
exactly 2^32, so the buffer was tiny while the copy loop wrote the
real data, overflowing the heap.
Use hb_unsigned_add_overflows() for both the sub-blob sum and the
total size, and fail the encode on overflow, as hb_gpu_draw_encode()
already does.
Not added a regression test: triggering it needs ~4 GiB of blobs.
Tests: meson test -C build.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* [gpu] Cap total paint sub-blob bytes
Gradient color-stop blobs (up to 1 MiB each) were unmetered, so a
COLRv1 paint DAG revisiting one thousands of times could force
multi-GiB allocations and copies in a single glyph encode.
Route all sub-blob pushes through hb_gpu_paint_t::push_sub_blob(),
which fails (marking the paint unsupported) once the running total
exceeds 64 MiB.
Tests: meson test -C build.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
on a stream-backed FT_Face
hb_ft_font_set_funcs() creates its own internal FT_Face from the
font's face blob, and hooks that blob to the FT_Face's generic
finalizer (_release_blob), with the shared static FT_Library hooked
to the blob's user-data destructor (destroy_ft_library).
On hb_font_destroy(), _hb_ft_face_destroy_static() locks
static_ft_library_mutex and calls FT_Done_Face(), which synchronously
runs the finalizer chain: _release_blob() -> hb_blob_destroy(). For a
face created from a custom FreeType stream (FT_OPEN_STREAM), each
hb_face_reference_blob() call synthesizes a brand-new, unshared blob
(via _hb_ft_reference_table()), so this hb_blob_destroy() call drops
it straight to zero, which triggers destroy_ft_library(), which tries
to lock the same non-recursive static_ft_library_mutex again on the
same thread. Self-deadlock.
This didn't reproduce for a file-path or memory-blob face, because
there the blob handed to the internal FT_Face is the same blob shared
by the outer hb_face_t, so the drop inside the lock only decrements
its refcount without reaching zero; the final unref (and
destroy_ft_library() call) happens later, outside the lock, when the
outer face itself is destroyed.
Fix by holding an extra reference on the blob across the locked
FT_Done_Face() call, so it can only reach a zero refcount (and invoke
destroy_ft_library()) after static_ft_library_mutex has been
released. FT_Done_Face() itself remains fully serialized, as intended
by 46644f885d.
Fixes: https://github.com/harfbuzz/harfbuzz/issues/6276
Tests: meson test -C build (260/260 passed); added
test_native_ft_set_funcs_stream_face_no_deadlock to test/api/test-ft.c,
which hangs on unfixed code and passes with the fix. Also verified
against the reporter's original repro.c.
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* [font] fix signed integer overflow in synthetic_glyph_extents
synthetic_glyph_extents applies synthetic slant and embolden to a
glyph's extents with plain hb_position_t arithmetic: x_bearing +
width and y_bearing + height in the slant branch, plus the += / -=
shifts in the embolden branch. glyf's contour_bounds_t and the sbix
png-dimension path each clamp x_bearing/width and y_bearing/height to
the full hb_position_t range independently via
hb_clamp_to<hb_position_t>(), so one field can reach INT32_MIN/MAX on
its own and the sum overflows, which is undefined behavior (UBSan
signed-integer-overflow at hb-font.hh:221). Reached from
hb_font_get_glyph_extents when synthetic slant/bold is set and the
glyph has extents but no drawable outline.
This is the sibling of 873dbc1 (scale_glyph_extents), which fixed the
same x_bearing + width overflow in the adjacent function. Accumulate
the slant terms in double and the embolden terms in int64, saturating
each result to hb_position_t with hb_clamp_to, matching that fix and
the copy in hb-ft.cc. Valid in-range extents are unchanged.
Add a regression test in test/api/test-font.c that drives the sink
with INT32-extreme extents; it aborts under the asan/ubsan job without
the fix. Full meson test -C build stays green.
* [font] use int64_t for slant accumulation in synthetic_glyph_extents
After instancing, GDEF maps all-zero delta rows to
HB_OT_LAYOUT_NO_VARIATIONS_INDEX. copy() still emitted a device with
the out-of-range varidx (0xFFFFFFFF), even though the delta is
already folded into the base value.
Drop the device in that case. gdef_partial_instance is marked
no_fonttools since fonttools keeps the dead device; the
empty_region_vardata and update_def_wght expected outputs are
regenerated.
Tested with: meson test -C build
Assisted-by: opencode
* [wasm] Fix memory leak inside _hb_wasm_shape()
Correct variable should be freed otherwise memory leaks inside the
container can take place.
* [wasm] extract features ptr to variable for clarity
serialize_rangeoffset_glyid() treated every zero idDelta as "this
range needs a glyphIdArray", which is what to_ranges() emits for
non-run ranges. A real run with delta 0 (glyph id == codepoint) got
the same treatment and cost two bytes per codepoint.
Only zero deltas are ambiguous. Check those against the mapping and
keep the ones that are genuinely identity as delta ranges.
Adds a subset test with a small hand-built glyf font whose cmap maps
U+0001..U+0007 to gids 1..7.
Tested with: meson test -C build
Assisted-by: opencode
hb_font_t::scale_glyph_extents() computed
extents->x_bearing + extents->width (and the y equivalent) as a
plain hb_position_t (int32_t) addition before scaling to float.
glyf's contour_bounds_t::get_extents() clamps x_bearing and width
independently via hb_clamp_to<hb_position_t>(), so each can reach
INT32_MIN/INT32_MAX on its own for a malformed/adversarial glyf
outline. Their sum can then overflow a 32-bit int, which is
undefined behavior and was flagged by UBSan.
Fix by doing the addition in float, after each term is already
scaled, instead of adding the raw hb_position_t values first.
Reproduced with the OSS-Fuzz hb-raster-fuzzer testcase at
https://oss-fuzz.com/testcase?key=5782717883809792, built locally
with -Db_sanitize=undefined:
../src/hb-font.hh:205:28: runtime error: signed integer overflow:
32 + 2147483647 cannot be represented in type 'int'
Confirmed the UBSan report no longer triggers on that testcase
after the fix, and `meson test -C build` passes in full (278/278,
1 pre-existing skip).
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Axis-value variation deltas can push a rounded component coordinate
outside the range of int. Assigning that float to coord_setter then
triggers undefined behavior in the conversion.
Use hb_clamp_to<int>() after rounding, preserving in-range behavior
while saturating extreme values. Add the minimized raster-fuzzer input.
Tested with meson test -C build and the ASan/UBSan raster fuzzer.
Assisted-by: OpenAI Codex
The analytic rasterizer stored every row's area and cover in 32- and
16-bit accumulators. Highly overlapping outlines could exceed those
ranges and trigger signed overflow before coverage was clamped.
Keep compact SIMD accumulators for rows with at most 127 active edges.
Lazily use 64-bit scalar accumulators for denser rows, where the narrow
cover bound no longer holds. Add regressions for area overflow and
cover wrapping.
Tested with meson test -C build, the ASan/UBSan raster API test, all
raster fuzzer corpus chunks, and the supplied reproducer.
Assisted-by: OpenAI Codex <noreply@openai.com>
The current build for v14.5.0 fails to compile on gcc 4.9.4 which is
used on embedded Kobo systems.
Ref: https://github.com/OGKevin/cadmus/pull/1005
Change-Id: 7a4bec5b5418bd0921b273d8d4a1e79b
Change-Id-Short: spvolnuouvyr
Update the generated Unicode and language-tag tables for Unicode 18.0.
Add script values for Jurchen, Proto-Cuneiform, and Seal, route them
through USE, and select ideographic baselines for Jurchen and Seal.
Document the new script values and extend Unicode and baseline tests.
Tested with meson test -C build: 278 passed, 1 skipped, 0 failed.
Assisted-by: Codex <codex@openai.com>
* [subset] Fix empty ClassDefFormat2 not intersecting class zero
An empty ClassDef (rangeCount == 0) assigns class zero to every glyph, but
ClassDefFormat2::intersects_class() and intersected_classes() reported no
intersection when glyphs is not empty.
Adds a subset test that fails before and passes after.
* [subset] Fix empty ClassDefFormat1 not intersecting class zero
Same bug as the previous commit, in the format 1 variant. When glyphCount is
zero and startGlyphID is zero, `startGlyph + classValue.len - 1` underflows
to 0xFFFFFFFF, so neither of the two comparisons can be true and class zero
is never reported. (With a nonzero startGlyphID it happens to work.)
harfbuzz always serializes an empty ClassDef as format 2, so this is only
reachable with fonts produced by other tools.
Extends the previous test with a hand-built format 1 variant of the font.
The glyf outline budget was only charged after a glyph loaded
successfully: Glyph::get_points() spent 16 units per node up front, and
glyf_accelerator_t::get_points() charged one unit per point only after
the whole tree came back. A glyph that loaded ~175K points and then
tripped a structural cap (HB_GLYF_MAX_POINTS, edge count) paid for the
nodes but not the points. Driven from a COLR paint walk through
hb_font_get_glyph_extents(), one such glyph was loaded and thrown away
over 8000 times per session within the budget, ~1.4G point loads.
Charge at the point where the work is attempted instead:
- SimpleGlyph::get_contour_points() takes the budget and charges
num_points before allocating, clearing, and reading, so truncated or
later-rejected glyphs pay for the points they allocate.
- Composite levels charge the component subtree size before
transform_points(), since each nesting level transforms every point
below it and that multiplication was previously invisible to the
budget.
The terminal per-point charge for path emission is unchanged.
Local hb-vector-fuzzer run on the testcase: 4.3s -> 0.12s. Full test
suite passes. The testcase is added as a fuzz seed.
Fixes: https://oss-fuzz.com/testcase-detail/5714817420427264
Claude-Session: https://claude.ai/code/session_013ksLGe7wUVwbuANqR56YaN
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
In the DirectWrite shaper, vis_clusters was allocated from the scratch
buffer with a static size based on Uniscribe scratch math (glyphs_size).
This caused three memory safety issues:
1. vis_clusters was indexed using log_clusters, which stores user
cluster IDs rather than glyph indices, allowing arbitrary
out-of-bounds array access. In DirectWrite, the character-to-glyph
mapping is returned in clusterMap, not log_clusters.
2. If glyphCount exceeded glyphs_size, the initialization loop
overwrote past the allocated scratch buffer.
3. buffer->ensure(glyphCount) could reallocate buffer->pos,
invalidating the scratch-allocated vis_clusters before its values
were read into info->cluster (use-after-free).
Fix this by allocating vis_clusters dynamically as an hb_vector_t
sized to glyphCount, and correctly mapping clusters using clusterMap
with bounds checks.
Assisted-by: Gemini
* Add null checking and fix memory leak in hb-coretext.cc
Assisted by: Gemini
* Add RAII hb_cf_ptr_t to manage coretext object lifetimes.
* Remove HAVE_CORETEXT guard in hb-coretext.hh
hb-coretext.hh will only be included when HAVE_CORETEXT is defined, so no need for the guard.
* [raster] Clip the edge sweep's column walk to the surface
edge_sweep_row() walked every fixed-point pixel column an edge spans,
even though columns right of the surface contribute nothing and
columns left of it only contribute a cover delta that telescopes into
column 0. A near-horizontal edge with extreme coordinates could span
~2^24 columns per row, making sweep cost proportional to coordinate
range instead of surface width: the ClusterFuzz testcase took 114s in
a plain release build.
Fold the left-of-surface portion of the walk into a single column-0
cover update, skip the right-of-surface portion, and enter/exit the
per-column loop at the visible window. Jumping the incremental fy
accumulator by k*delta_fy is exact, since the increment is constant
and the running sums are bounded, so output is bit-identical; verified
by comparing 178 hb-raster renders (draw and paint paths, several
fonts, sizes, and negative-margin crops that clip edges off-surface)
byte-for-byte against the previous code. The testcase now runs in
0.17s. Also add it as a fuzz seed.
Fixes https://oss-fuzz.com/testcase-detail/5267980968329216
Assisted-by: Claude Fable 5 <noreply@anthropic.com>
* [raster] Keep an unclipped fast path in the edge sweep
The window clipping added in the previous commit made every
multi-column edge-row pay for the clip setup, measurable as ~1% more
instructions rasterizing all glyphs of NotoSansCJKsc/SourceHanSans-VF
at 256px with hb-raster-all. Guard it with a two-compare
fully-inside-the-surface check and walk such edges with the original
unclipped code, which is also the everyday case: auto-sized surfaces
always contain their edges.
Interleaved perf-stat A/B (5 rounds, all glyphs at 256px): instruction
overhead drops to +0.24% (CJK) / +0.9% (AwamiNastaliq, tiny workload);
cycles flat to -1.3%, within noise. Output remains bit-identical on
the 178-render comparison matrix, and the oss-fuzz 5267980968329216
testcase stays at 0.19s.
Assisted-by: Claude Fable 5 <noreply@anthropic.com>
* [raster] Charge column spans in the pixel-work estimate
hb_raster_draw_get_pixel_work() charged each edge only its scanline
span, so a nearly-horizontal edge crossing one pixel row but sweeping
the full surface width cost up to `width` cells while being charged 2
units — the paint path's pixel-budget precharge could undercharge the
sweep by a factor of the surface width. This is the same blind spot
that let the previous commits' column-walk timeout through the paint
path's precharge.
Charge each edge its column span too, clipped to the surface width the
same way the row span is clipped to its height. Since an edge is a
straight segment, row span plus clipped column span is a true upper
bound on the post-clipping sweep work. Normal outlines are barely
affected: flattening yields short segments whose column spans are
commensurate with their row spans, far below the 2^26 budget.
No rendering change: the 178-render comparison matrix remains
bit-identical and the full test suite passes.
Assisted-by: Claude Fable 5 <noreply@anthropic.com>
Short-circuit unchanged-index checks and simplify glyph closure to use
one pending set with the output glyph set tracking visited glyphs.
Strengthen retained-glyph-ID coverage by comparing the source and subset
outlines using the original glyph ID.
Tested with: meson test -C build --print-errorlogs
Assisted-by: OpenAI Codex <openai-codex@users.noreply.github.com>
Move table-specific planning and record rewriting phases out of
VARC::subset() into a dedicated helper. This leaves the entry point to
coordinate the phases and serialize their results without changing
subsetting behavior.
Tested with: meson test -C build
Assisted-by: OpenAI Codex <openai-codex@users.noreply.github.com>
Propagate allocation and component-decoding failures from VARC closure
instead of continuing with an incomplete glyph set. Treat malformed VARC
records as dependency-extraction failures as well.
Document VARC dependency edges and exercise subset variation-store
remapping at non-default coordinates.
Tested with:
- meson test -C build
- standalone lean and world syntax checks
- CMake harfbuzz-subset build
Assisted-by: OpenAI Codex <openai-codex@users.noreply.github.com>
When a gvar table has no shared tuples, point sharedTuplesOffset at the
end of the glyph-variation offsets array, where the empty shared-tuples
array resides, instead of emitting a null offset.
This lets strict consumers retain glyph variations in partially
instantiated fonts with no shared tuples.
Assisted-by: OpenAI Codex <codex@openai.com>
Resolve the VARC record-decoding conflict by preserving the shared
decoder while threading main's rendering budget through axis and
variation work.
Tests: meson test -C build (277 passed, 1 skipped)
Assisted-by: OpenAI Codex <codex@openai.com>
The unified glyph budget raised outline work limits for VARC, raster,
vector, and GPU rendering to 1 << 26. Pathological fuzz inputs can
consume that entire allowance, doubling the full test-suite runtime.
Restore the common outline budget to 1 << 24 while retaining the
rasterizer's separate 1 << 26 pixel-work floor. Precharge raster
scanline work against that pixel budget before rendering so complex
COLR masks cannot do unbounded edge sweeps before their mask area is
charged.
Tests: meson test -C build (276 passed, 1 skipped; 2.82-2.83s)
Assisted-by: OpenAI Codex <codex@openai.com>
Use Skrifa's bitmap-strike API to provide glyph extents and paint PNG
and BGRA images from CBDT and sbix tables. Select strikes using the
font's requested ppem and preserve the OT backend's table precedence.
Keep the face blob alive through image blobs retained by paint
callbacks.
Tests: meson test -C build --no-rebuild
Assisted-by: OpenAI Codex <codex@openai.com>
Skrifa exposes resolved COLRv1 clip boxes independently of outline
glyph metrics. Use them when present so glyph extents report the clip
bounds instead of falling back to painted-content bounds.
Keep the existing paint fallback for color glyphs without a clip box.
Assisted-by: OpenAI Codex <codex@openai.com>
The RuleSet fast paths filtered out rules requiring another glyph when
the skipping iterator reached the end of the buffer. Unlike the normal
rule walk, that early return did not record the resulting
unsafe-to-concat range.
Keep the end-of-input optimization, but track filtered rules and mark
their concat hazard before returning. Cover both contextual and chained
contextual rule sets with a five-rule font that forces the fast paths.
Assisted-by: OpenAI Codex <codex@openai.com>
A zero-valued PairPos record can select a different, non-zero record
if the second glyph changes, so keep the pair unsafe to concat even
when no positioning was applied.
Likewise, make the ligature second-glyph digest rejection report the
same hazard as the full ligature walk.
Tests: meson test -C build
Port of https://github.com/harfbuzz/harfrust/pull/465
Assisted-by: OpenAI Codex <codex@openai.com>
Build a compact class digest for every format-2 contextual RuleSet in
the subtable's existing external cache. Reject a RuleSet before walking
its rules when the next input glyph cannot match any rule's first class.
Rules without another input position force the digest open. Digest
collisions only cause the normal rule walk, and allocation failure falls
back to the uncached path.
Assisted-by: Codex <codex@openai.com>