* [fuzzing] Guard raster/vector create_or_fail results
The raster and vector fuzzers called create_or_fail() and then
immediately dereferenced the returned pointers. Under CIFuzz's
allocation-failure coverage, that turns a valid NULL return into a
harness crash instead of treating it as an out-of-memory case.
Return early after cleaning up when draw or paint context creation
fails. This keeps the fuzzers aligned with the create_or_fail()
contract and avoids null-deref crashes unrelated to library logic.
Tested:
- meson compile -C build hb-raster-fuzzer hb-vector-fuzzer
- build/test/fuzzing/hb-vector-fuzzer test/fuzzing/fonts/TwemojiMozilla.subset.ttf
- build/test/fuzzing/hb-raster-fuzzer test/fuzzing/fonts/TwemojiMozilla.subset.ttf
Assisted-by: OpenAI Codex
* [vector] Avoid OOM leak in paint initialization
hb_vector_paint_ensure_initialized() used group_stack.push() and
then treated the returned pointer as a normal success indicator.
On allocation failure push() returns the Crap() sentinel, so the
subsequent root->alloc(4096) allocated backing storage into the
failure sink and leaked under malloc-failure fuzzing.
Use push_or_fail() instead and only allocate the root body after a
real push succeeds.
Tested:
- build/test/fuzzing/hb-vector-fuzzer test/fuzzing/fonts/TwemojiMozilla.subset.ttf
Assisted-by: OpenAI Codex
* [vector] Avoid double-free on blob creation failure
hb_svg_blob_from_buffer() handed buffer ownership to hb_blob_create()
in the non-recycled path and then ran its own failure cleanup. When
hb_blob_create() failed, it had already invoked the destroy callback,
so the local cleanup freed the same buffer again.
Use hb_blob_create_or_fail() there instead so allocation failure
returns nullptr without triggering the empty-blob fallback and its
conflicting cleanup semantics.
Tested:
- meson compile -C build hb-vector-fuzzer
Assisted-by: OpenAI Codex
* [vector] Avoid double-free in draw blob creation
Mirror the blob-creation failure fix from vector paint in the SVG
draw path. The non-recycled branch handed ownership to
hb_blob_create(), which may already run the destroy callback on
failure, and then fell into local cleanup that freed the same buffer
again.
Switch that branch to hb_blob_create_or_fail() so failure reports
nullptr without conflicting cleanup.
Tested:
- meson compile -C build hb-vector-fuzzer
Assisted-by: OpenAI Codex
* [vector] Fix OOM handling in SVG subset id collection
hb_svg_add_unique_id() used v->push() and tested the returned
pointer for null. hb_vector_t::push() reports allocation failure via
the vector error state and returns the Crap() sentinel, so the old
check did not reliably stop on OOM.
Check v->in_error() after push() before writing into the new slot.
Assisted-by: OpenAI Codex
* [vector] Guard paint render after init failure
hb_vector_paint_render() assumed ensure_initialized() had created
the root group body and indexed group_stack[0] unconditionally.
Under malloc-failure fuzzing, initialization can fail while extents
are already set, leaving group_stack empty and causing a null-deref.
Return nullptr if initialization did not produce the root group.
Assisted-by: OpenAI Codex
* [vector] Propagate paint init failures
hb_vector_paint_ensure_initialized() can fail under malloc-failure
fuzzing, but several paint callbacks assumed the root group existed
immediately afterward and accessed current_body() or group_stack[0].
Make initialization report success and bail out at the callers that
require the root group. Also use push_or_fail() for the extra group
stack entries created by grouping and glyph caching.
Assisted-by: OpenAI Codex
* [raster] Bound SVG path parsing work
hb_raster_svg_parse_path_data() had no complexity cap, unlike the
other new raster SVG subsystems that already bound recursion depth,
attribute count, and gradient stops. Malformed path data could
therefore spend excessive time in parsing and bbox/render work under
fuzzing.
Add a hard limit on parsed path segments and expose it through
hb-limits.hh so the budget is defined alongside the other shared
parser/work limits.
Assisted-by: OpenAI Codex
* [raster] Bound temporary clip mask buffers
hb_raster_paint_push_clip_from_emitter() and the general clip
rectangle path could allocate and clear full-surface temporary A8 clip
masks with no explicit size cap, leading to timeout-scale work on
malformed inputs.
Reuse the raster buffer size policy for these temporary masks as well,
and move that limit into hb-limits.hh so raster images and clip masks
share the same bound.
Assisted-by: OpenAI Codex
* [raster] Reduce buffer size cap
Lower HB_RASTER_MAX_BUFFER_SIZE from 1<<30 to 1<<24 so the shared
raster image / clip-mask limit meaningfully constrains pathological
fuzzer inputs instead of allowing timeout-scale temporary buffers.
Assisted-by: OpenAI Codex
hb_raster_paint_push_clip_from_emitter() and the general clip
rectangle path could allocate and clear full-surface temporary A8 clip
masks with no explicit size cap, leading to timeout-scale work on
malformed inputs.
Reuse the raster buffer size policy for these temporary masks as well,
and move that limit into hb-limits.hh so raster images and clip masks
share the same bound.
Assisted-by: OpenAI Codex
hb_raster_svg_parse_path_data() had no complexity cap, unlike the
other new raster SVG subsystems that already bound recursion depth,
attribute count, and gradient stops. Malformed path data could
therefore spend excessive time in parsing and bbox/render work under
fuzzing.
Add a hard limit on parsed path segments and expose it through
hb-limits.hh so the budget is defined alongside the other shared
parser/work limits.
Assisted-by: OpenAI Codex
hb_vector_paint_ensure_initialized() can fail under malloc-failure
fuzzing, but several paint callbacks assumed the root group existed
immediately afterward and accessed current_body() or group_stack[0].
Make initialization report success and bail out at the callers that
require the root group. Also use push_or_fail() for the extra group
stack entries created by grouping and glyph caching.
Assisted-by: OpenAI Codex
hb_vector_paint_render() assumed ensure_initialized() had created
the root group body and indexed group_stack[0] unconditionally.
Under malloc-failure fuzzing, initialization can fail while extents
are already set, leaving group_stack empty and causing a null-deref.
Return nullptr if initialization did not produce the root group.
Assisted-by: OpenAI Codex
hb_svg_add_unique_id() used v->push() and tested the returned
pointer for null. hb_vector_t::push() reports allocation failure via
the vector error state and returns the Crap() sentinel, so the old
check did not reliably stop on OOM.
Check v->in_error() after push() before writing into the new slot.
Assisted-by: OpenAI Codex
Mirror the blob-creation failure fix from vector paint in the SVG
draw path. The non-recycled branch handed ownership to
hb_blob_create(), which may already run the destroy callback on
failure, and then fell into local cleanup that freed the same buffer
again.
Switch that branch to hb_blob_create_or_fail() so failure reports
nullptr without conflicting cleanup.
Tested:
- meson compile -C build hb-vector-fuzzer
Assisted-by: OpenAI Codex
hb_svg_blob_from_buffer() handed buffer ownership to hb_blob_create()
in the non-recycled path and then ran its own failure cleanup. When
hb_blob_create() failed, it had already invoked the destroy callback,
so the local cleanup freed the same buffer again.
Use hb_blob_create_or_fail() there instead so allocation failure
returns nullptr without triggering the empty-blob fallback and its
conflicting cleanup semantics.
Tested:
- meson compile -C build hb-vector-fuzzer
Assisted-by: OpenAI Codex
Extend hb-cplusplus.hh with shared_ptr/unique_ptr vtables for the
public raster and vector object types when their headers are
included.
Also simplify the local vtable macro so the get_empty/nullptr cases
share one definition instead of duplicating the exact same template
body.
Tested:
- meson compile -C build
Assisted-by: OpenAI Codex
hb_vector_paint_ensure_initialized() used group_stack.push() and
then treated the returned pointer as a normal success indicator.
On allocation failure push() returns the Crap() sentinel, so the
subsequent root->alloc(4096) allocated backing storage into the
failure sink and leaked under malloc-failure fuzzing.
Use push_or_fail() instead and only allocate the root body after a
real push succeeds.
Tested:
- build/test/fuzzing/hb-vector-fuzzer test/fuzzing/fonts/TwemojiMozilla.subset.ttf
Assisted-by: OpenAI Codex
The raster and vector fuzzers called create_or_fail() and then
immediately dereferenced the returned pointers. Under CIFuzz's
allocation-failure coverage, that turns a valid NULL return into a
harness crash instead of treating it as an out-of-memory case.
Return early after cleaning up when draw or paint context creation
fails. This keeps the fuzzers aligned with the create_or_fail()
contract and avoids null-deref crashes unrelated to library logic.
Tested:
- meson compile -C build hb-raster-fuzzer hb-vector-fuzzer
- build/test/fuzzing/hb-vector-fuzzer test/fuzzing/fonts/TwemojiMozilla.subset.ttf
- build/test/fuzzing/hb-raster-fuzzer test/fuzzing/fonts/TwemojiMozilla.subset.ttf
Assisted-by: OpenAI Codex
Skip the SVG table accelerator and do not define the public
SVG color APIs when HB_NO_SVG is set. Guard internal
callers so SVG-table-dependent code compiles out cleanly,
and add HB_NO_SVG to the HB_TINY profile.
Tested: meson setup build-nosvg -Dc_args=-DHB_NO_SVG -Dcpp_args=-DHB_NO_SVG
Tested: meson compile -C build-nosvg hb-shape hb-info hb-vector main
Tested: meson compile -C build-nosvg (libharfbuzz/libharfbuzz-raster/libharfbuzz-vector linked; full tree stopped on unrelated missing chafa.h in util/hb-raster.cc)
Tested: nm -gU build-nosvg/src/libharfbuzz.0.dylib | grep -E "hb_ot_color_.*svg" (no matches)
Assisted-by: OpenAI Codex
Root cause: the guide required descriptive commit messages,
but it did not encode the preferred body width or the
preference to skip irrelevant testing boilerplate.
Update AGENTS.md to recommend wrapping commit message
bodies to about 70 columns and to mention testing only
when it was actually performed or otherwise relevant.
Assisted-by: OpenAI Codex
Root cause: the previous guide had accumulated useful project-specific guidance, but too much detail made the core constraints and workflow harder to scan quickly.
Rewrite AGENTS.md into a shorter structure that keeps the important HarfBuzz-specific rules around API/ABI stability, build and test expectations, feature-guarded builds, generator-driven updates, commit conventions, and project mindset, while removing lower-signal detail. Add back a short Wisdom section to preserve project tone without re-expanding the file.
Testing: not run (documentation-only change).
Assisted-by: OpenAI Codex
Root cause: the repository guide captured broad commit hygiene expectations, but it did not explicitly encode the preferred subject prefix style, issue-linking trailers, or the required Assisted-by trailer.
Update AGENTS.md to require consistent bracketed subsystem prefixes, recommend Fixes: trailers when relevant, and require Assisted-by on commits written through the agent.
Testing: not run (documentation-only change).
Assisted-by: OpenAI Codex
Root cause: the repository guide captured explicit project rules, but it did not yet encode several stable engineering patterns visible across the last 1000 commits.
Update AGENTS.md to reflect recurring HarfBuzz practices: incremental helper extraction over large rewrites, pairing bug fixes with reproducible regressions or fuzz coverage, keeping docs and build/install wiring aligned with feature changes, treating generated data as generator-driven work, and validating simplifications carefully because semantic regressions are often only caught later.
Testing: not run (documentation-only change).
Root cause: the guide had accumulated useful project-specific guidance, but the structure made it harder to scan and apply consistently.
Reorganize AGENTS.md around core rules, build and verification flow, repository layout, change-specific guidance, and working mindset while preserving the existing instructions and recent review-driven additions.
Testing: not run (documentation-only change).
Root cause: the initial repository guide omitted several project-specific expectations raised during review and listed non-installed utilities in the util map.
Update AGENTS.md to reflect HarfBuzz conventions around API/ABI stability, new API documentation with XSince: REPLACEME, warning-free rebuild expectations, OOM and error-handling guidance, root-cause-first debugging, and descriptive commit messages. Also limit the utility list to installed command-line tools.
Testing: not run (documentation-only change).
Use checked arithmetic when calculating the number of extra glyphs
needed during stch processing. Includes a new hb_unsigned_add_overflows
helper in hb-algs.hh.
Co-authored-by: Codex (AI assistant)
Co-authored-by: Gemini (AI assistant)