[lizthegrey]

Speeds up SpeedDefault encoding on arm64 (+2–4%) with byte-identical output, by giving the default-level encoder the same hash-prime treatment as the better encoder got in #1230. amd64 is unchanged.

Changes (362a8de)

doubleFastEncoder and doubleFastEncoderDict hash with the same two 64-bit primes as betterFastEncoder. On arm64 each prime takes 4 instructions to build, and the compiler rebuilds it at every use.

• betterHashPrimes becomes hashPrimes8and5 in hash.go, shared by both encoders. It still loads the primes from a package var only when hashPrimesInRegs is set (arm64), and returns the constants elsewhere.
• dFastHashL/dFastHashS replace hashLen in the default encoders' loops, taking the primes as arguments. Reset and ResetPrefix keep hashLen.
• The better encoder's generated code is unchanged.

Why output is unchanged

• dFastHashL/dFastHashS compute exactly what hashLen computes for these table sizes. TestBetterHashMatchesHashLen in encoder_test.go becomes TestHashHelpersMatchHashLen and gains the two default-encoder cases.
• sha256 of every frame's output matches master for 8 corpora × {default, better}, on arm64 and amd64.

Benchmarks

Same setup as #1230: 32 MiB decompressed prefixes of your real corpora, each encoded as 4 MiB EncodeAll frames with a 4 MiB window, CRC on and concurrency 1. Each result is 10 master/PR rounds, alternating within each round, on one pinned core, compared with benchstat. Figures are SpeedDefault MB/s vs master; ~ means not significant.

corpus	Graviton2 (c6g)	Graviton3 (c7g)	Graviton4 (c8g)	Sapphire Rapids (c7i)
apache.log	+2.1%	~	~	~
cockroach.node1.log	+2.4%	+0.6%	+1.3%	~
github-june-2days-2019.json	+5.4%	+1.8%	+2.4%	~
github-ranks-backup.bin	+7.3%	+3.2%	+2.5%	~
gob-stream	+4.8%	+1.6%	+1.9%	~
nyc-taxi-data-10M.csv	+2.0%	+1.1%	+1.0%	−0.2% (p=0.050)
rawstudio-mint14.tar	+9.7%	+5.9%	+3.8%	~
sofia-air-quality-dataset.tar	+1.9%	+1.6%	+1.5%	~
geomean	+4.4%	+2.1%	+1.9%	flat

The arm64 cells are all improvements, and all are significant except the ~ ones. On amd64 the helpers return the constants, so the hashing is the same as master's; the one borderline Sapphire Rapids cell sits at the edge of significance, and the other seven are flat.

These runs were taken together with #1230's; that PR does not touch the default encoder, so the SpeedDefault rows measure this change alone.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Performance
  • Improved internal hash calculations used by fast compression paths, including dictionary-based encoding. Compression results and behavior remain unchanged.

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Advanced

Run ID: 0124a38a-80aa-42d6-9701-cf0591c6b5e5

📥 Commits

Reviewing files that changed from the base of the PR and between 0533a67 and 362a8de.

📒 Files selected for processing (4)

• zstd/enc_better.go
• zstd/enc_dfast.go
• zstd/encoder_test.go
• zstd/hash.go

Included review availability: This review used your included allowance. Your plan provides up to 8 included reviews per hour; 4 remain after this review.

---

📝 Walkthrough

Walkthrough

The zstd better and double-fast encoders now use shared hash primes and prime-based hash helpers. The hash-helper test checks the better and double-fast long- and short-hash results against hashLen.

Changes

Zstd hash helper updates

Layer / File(s)	Summary
Shared primes and better encoders zstd/hash.go, zstd/enc_better.go	hashPrimes8and5 provides the long and short primes. Both better encoders use it instead of betterHashPrimes, which was removed.
Double-fast hash indexing zstd/enc_dfast.go, zstd/encoder_test.go	The double-fast encoders use dFastHashL and dFastHashS for match-table lookups and updates. The test compares the better and double-fast helper results with hashLen.

Priority: ➖ Normal

Estimated code review effort: 2 (Simple) | ~12 minutes

Change: Refactor

Merge Risk: ⚪ Minimal · up to 362a8

The identified repeat-offset path does not require a fix, and no actionable merge-blocking issue remains after normal checks.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 71.43% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 7 functions across 4 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly describes the main change: keeping the default encoder hash primes in registers on arm64. It matches the changes to the better and double-fast encoders.
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.

• Fix all pre-merge checks with AI

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

---

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.}