Inline cbor_incref and cbor_decref fast paths

[PJK]

Summary

Move the reference-counting hot path (increment; decrement + zero check) into static inline definitions in cbor/common.h. The deallocation branch of cbor_decref stays out of line as _cbor_decref_free, called from the inline once refcount hits zero.

For DOM-building workloads — every construction touches an incref/decref that previously crossed a function-call boundary. Inlining removes that overhead at every call site, including libcbor's own internal recursion in the free branch.

Perf context

Measured on the GLD serializer benchmark (default matrix, N=100). Combined with two wrapper-side fixes (narrowest-width integer builders + cbor_set_allocs arena), the DOM path goes from being the slowest CBOR encoder in the published chart to competitive with tinycbor. This library change contributes roughly the first ~1.3–1.5× of that speedup on its own; the rest comes from the wrapper. Sizes are unaffected.

Caveats

Switching from CBOR_EXPORT extern declarations to static inline in the public header removes cbor_incref and cbor_decref from the shared library's exported symbol table. That is technically an ABI change:

• Existing binaries linked against an older libcbor.so continue to work (they carry their own extern resolution).
• Binaries compiled against the new header cannot be re-linked against an older libcbor.so.
• Consumers using dlsym(\"cbor_incref\") / dlsym(\"cbor_decref\") would break.

If ABI compatibility matters, an alternative is C99 inline + extern inline in common.c, which keeps the exported symbols alive. Happy to reshape into that form.

Test plan

• All 28 existing ctest suites pass
• CI green

🤖 Generated with Claude Code