perf(decode): hot-loop register-pressure trims + release-plz concurrency guard

[polaz]

Summary

Three independent, low-risk changes from the level-1 decodecorpus decode
investigation:

• ci(release) — serialize release-plz runs with a concurrency group.
  Two PRs merged back-to-back start two overlapping release-plz runs that race on
  the release-PR git refs; one closes the release PR while the other's recreate
  fails on the already-taken ref, leaving it closed with no replacement (this
  happened during the work that led to this branch). cancel-in-progress stays
  false so the publish step is never interrupted mid-run.
• perf(decode) — sequence loop — read OF/ML/LL as three separate pipelined
  reads instead of one serial PEXT-triple (long-pipeline arm), and advance the
  FSE states before executing the current sequence (short arm, mirrors upstream
  ZSTD_decodeSequence; the execute reads no bits so the reorder is
  byte-identical).
• perf(decode) — HUF burst — drop the four carried per-stream sub-byte
  phase registers; recompute the value from trailing_zeros of the
  just-reloaded register at burst writeback.

Both decode changes are byte-identical and perf-neutral — kept as valid
register-pressure / op-count reductions, not as measured wins.

Why perf-neutral, not regressions

The investigation's main finding: the level-1 decode gap vs the C reference is
instruction count (≈1.23×, ≈ the cycle ratio), spread across the decoder
(the sequence decoder is the largest share). That is the cost of a memory-safe
pure-Rust decoder (bounds checks, Result plumbing) plus the absence of a
hand-tuned assembly HUF loop — not a missed optimisation in any single kernel. A
standalone global_asm! HUF burst was prototyped and measured across three
register layouts; it cut instruction count but did not beat the LLVM-scheduled
Rust burst on cycles (lower IPC), so it was dropped. These neutral op-reductions
are the parts of that work that are strictly-better constructions.

Methodology note: per-change deltas were measured with perf stat -e instructions (deterministic) and -e cycles under low load on a dedicated host.
Criterion on these fast small-frame fixtures without a matched control arm
produced false ±2-4% "regressions" (scheduler noise); those are not real.

Testing

• cargo nextest run -p structured-zstd — full suite green (819 tests),
  including the 1000-iteration random round-trip.
• Output byte-identical on the decodecorpus corpus (round-trip + length +
  checksum).
• Per-change instruction count unchanged-or-lower; cycles flat within noise.

The ci(release) change is independent and can be merged first if the release
tooling needs it sooner.

[coderabbitai]

Warning

Review limit reached

@polaz, we couldn't start this review because you've reached your PR review rate limit.

More reviews will be available in 32 minutes and 52 seconds. Learn how PR review limits work.

Your organization has used up its prepaid credits, and credit purchases are no longer available. Enable the review add-on in the billing tab to keep reviews running — you're only billed for reviews past your plan's rate limits ($0.25/file).

⌛ How to resolve this issue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based credits.

🚦 How do rate limits work?

CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability.

For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window.

Please see our Fair Usage Limits Policy for further information.

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 749af4ac-4675-43d3-9307-3ee1389a1fd7

📥 Commits

Reviewing files that changed from the base of the PR and between 562f380 and 57b926d.

📒 Files selected for processing (3)

• .github/workflows/release-plz.yml
• zstd/src/decoding/literals_section_decoder.rs
• zstd/src/decoding/seq_decoder_avx2.rs

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch perf/decode-l1-fast-decodecorpus

---

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR updates release automation and decode hot loops. The main changes are:

• Adds a queued release-plz concurrency group for pushes to main.
• Recomputes HUF burst writeback phase from the sentinel-bearing bit register instead of carrying separate phase registers.
• Reads AVX2 sequence extra bits as separate unchecked reads after one upfront refill.
• Moves short-loop FSE state updates before sequence execution while preserving bit consumption order.

Confidence Score: 5/5

The changes appear merge-safe: the release workflow serialization is isolated, and the decode-loop edits are intended to preserve byte output while reducing register pressure.

The touched code paths are narrow and the described validation covers the full crate tests plus byte-identical corpus checks for the decoder behavior.

T-Rex Logs

What T-Rex did

• Compared the before and after logs for the release-plz concurrency guard to verify the changes.
• Compared base and head decode-perf-neutral outputs to confirm AVX2 presence and consistent test counts, and that the final TREX_CASE checksums matched.
• Observed an environment blocker: cargo is not available, causing decode-perf-neutral runs to stop before execution.

_{Ran code and verified through T-Rex}

_{Reviews (1): Last reviewed commit: "Merge branch 'main' into perf/decode-l1-..." | Re-trigger Greptile}