moq-mux: SIMD start-code scanning via memchr

[kixelated]

Summary

Replaces the hand-rolled byte-at-a-time Annex-B start-code scanner in rs/moq-mux/src/codec/annexb.rs with a single SIMD-accelerated substring search.

The key insight: both start-code forms share the 00 00 01 suffix (3-byte is 00 00 01, 4-byte is 00 00 00 01). So you don't need a bespoke vectorized scanner that disambiguates 3-vs-4 inline. You do one substring search for the 00 00 01 core, then peek a single byte backward to see if a leading zero promotes it to 4-byte:

pub fn find_start_code(b: &[u8]) -> Option<(usize, usize)> {
	let core = memchr::memmem::find(b, &[0, 0, 1])?;
	if core > 0 && b[core - 1] == 0 {
		Some((core - 1, 4))
	} else {
		Some((core, 3))
	}
}

This answers the long-standing TODO in the file (Is this the type of thing that SIMD could further improve?).

Why memchr

• It's the maintained, stable-Rust, runtime-CPU-feature-detected (SSE2/AVX2/NEON) implementation. No unsafe, no nightly std::simd, no #[cfg(target_arch)] matrix to maintain. That matches this repo's "prefer a maintained crate over hand-rolling non-core functionality" rule.
• memchr was already in the tree transitively; this just makes it a direct dependency.

A criterion benchmark (four-way: naive / scalar / memmem / a memchr-single-byte variant) confirmed the SIMD win: ~4-5x over the old scalar scanner on a realistic iterator workload, and 25-100x on pathological zero-heavy scans. The memmem and memchr variants traded blows depending on byte statistics, so we kept the simpler memmem one. The benchmark is not committed.

Behavior

Unchanged. All 36 existing annexb tests pass, including:

• 3-byte vs 4-byte disambiguation
• the consecutive-leading-zeros case (ff 00 00 00 00 00 01), where only one zero is promoted so the result stays (3, 4) and the extra zeros remain outside the code.

Public API

find_start_code stays pub with the same signature fn(&[u8]) -> Option<(usize, usize)> (the dropped mut is a binding mode, not part of the signature). No breaking change, so this targets main.

Test plan

• cargo test -p moq-mux --lib codec::annexb — 36 passed

(Written by Claude)

[coderabbitai]

Walkthrough

This pull request optimizes Annex-B start code detection in MPEG-TS codec operations. A new memchr dependency (version 2) is added to Cargo.toml, enabling SIMD-accelerated binary pattern matching. The find_start_code function is rewritten to use memchr::memmem to search for the 0x000001 core pattern, then conditionally expands matches preceded by 0x00 into 4-byte start codes, replacing the previous manual byte-skipping logic. The public function signature remains unchanged.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title 'moq-mux: SIMD start-code scanning via memchr' directly summarizes the main change: replacing hand-rolled start-code scanning with SIMD-accelerated memchr implementation.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description check	✅ Passed	The pull request description clearly explains the changes: replacing a hand-rolled byte-scanner with a SIMD-accelerated memchr substring search, including rationale, implementation details, behavior preservation, and test results.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

✨ Simplify code

• Create PR with simplified code
• Commit simplified code in branch claude/wonderful-rhodes-aad20e

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}