Summary

A from-scratch pure C11 reimplementation of tinyobjloader: secure, portable, freestanding-capable, dependency-free, fast, and robust. Not header-only — tiny_obj_c.c + tiny_obj_c.h, plus a separate reusable tessellation library tobj_tess.{c,h}. Carries the C++ feature set over to C11 and adds line/point/free-form-patch primitives.

Highlights

• Dual precision — float (_f) and double (_d) families coexist in one build, generated from a single template via token-pasting.
• Freestanding / no-dependency — the core links zero libc symbols (verified with nm under -ffreestanding -DTOBJ_NO_LIBC): no malloc/strtod/fopen/memcpy/sqrt/pthread. Injectable allocator, bundled ASCII→float parser, no math.h in the core or default tessellator.
• Robust tessellation (tobj_tess) — Newell projection with bbox fallback, magnitude-scaled epsilons (fixes the C++ absolute-epsilon area threshold that is meaningless at the ~8000–14000 coordinate magnitudes in sandbox/zh2-ill.obj, sandbox/ill-tri.obj, models/issue-295-trianguation-failure.obj), a convex fan fast path, and a force-clip backstop that always emits n−2 triangles and never crashes or loops on degenerate, concave, collinear, coincident, non-planar, or self-intersecting input.
• Primitives — faces (triangulated), lines (l), points (p), and free-form patches (cstype/curv/curv2/surf/vp/parm/trim/hole/end) parsed and retained losslessly (no NURBS evaluation).
• Full material support — Phong + PBR + 13 texture channels with options, unknown-parameter map, map_Kd-without-Kd default, standalone .mtl API, material-resolver callback (replaces the virtual MaterialReader), and a streaming callback API.
• Optional perf behind macros (scalar/single-threaded/libc default) — multithreaded two-pass parse, SIMD newline scan (AVX2/SSE2/NEON), mmap + file I/O, arena allocator, fast float.
• Security — bounds-checked indices, overflow-safe size arithmetic, configurable caps, defined teardown.

Validation (all under ASan + UBSan)

• 84-file corpus (models/ + sandbox/) loads with 0 failures, 1.45M triangles, consistent index counts.
• acutest drivers: tests/tess_tester.c (15 unit/property tests incl. pathological classes, area preservation, determinism) and tests/tester_c.c (12 loader tests: corpus, both precisions, triangulation, BOM/CRLF, vertex colors, lines/points, MTL, garbage) — all pass.
• Freestanding driver runs with a custom bump allocator and no libc.
• Multithreaded output is byte-identical to single-threaded on a 20k-vertex varied OBJ.
• libFuzzer harness finds no crashes/leaks.
• Strict -Wall -Wextra -Wpedantic -Wshadow clean (default and all-features).

Build / test

cd tests && make check_c          # tester_c, tester_c_features (SIMD+MT), tess_tester, freestanding
# or via CMake:
cmake -S . -B build -DTINYOBJLOADER_C_BUILD_TESTS=ON && cmake --build build && ctest --test-dir build

Added options in CMakeLists.txt: TINYOBJLOADER_BUILD_C_LIBRARY (ON), TINYOBJLOADER_C_ENABLE_FILE_IO/MMAP/SIMD/MULTITHREADING, TINYOBJLOADER_C_BUILD_TESTS, plus a fuzz_tobj_c target under LIB_FUZZING_ENGINE.

Known limitation

In the multithreaded path, curv2→vp index resolution can differ from single-threaded only when vp lines have inconsistent component counts (a pathological free-form case); it does not affect the corpus or normal use. A reserved tobj_mathops hook is currently unused (the default tessellator needs no sqrt).

🤖 Generated with Claude Code