Speed up Slice search, UTF-8 case conversion, and trims

[wendigo]

A batch of performance improvements to Slice and SliceUtf8, plus one bug fix. Each commit is independent and individually benchmarked (JMH, JDK 25, Apple Silicon; multi-fork runs where single-fork variance was significant).

Bug fix

• Slices.ensureSize threw ArrayIndexOutOfBoundsException when growing a view slice whose backing array extends more than the new capacity past the view start — the tail zero-fill used a bound in the old array's coordinates. Now clamped; regression test added.

Search

• indexOf: replaced the 4-byte scan window with an 8-byte SWAR loop using exact zero-byte detection (no false-positive candidates), candidate iteration via bit tricks instead of re-reading memory, and a new tail anchor (last four pattern bytes) alongside the head anchor to reject false candidates before the full comparison. 1.2–4× faster across sparse and repetitive inputs; scan throughput ~4 → ~11 GB/s.
• lastIndexOfByte: switched to the exact zero-byte mask, eliminating the per-candidate re-check loop. 2.9× when the match sits below bytes differing only in the lowest bit (e.g. n vs o), 1.19× on plain backward scans.
• lastIndexOf: same head/tail anchors as indexOf. ~1.7× on repetitive inputs whose candidates fail deep in the pattern; unchanged on random data.

UTF-8

• ASCII case conversion (toUpperCase/toLowerCase): scans and translates eight bytes at a time by flipping the case bit of bytes in range (SWAR range test, branchless tail). 2.7–3.7× at 100 B, 5–9× at 1 KB+; the zero-copy return for already-correct input is preserved and also 8× faster to reach. Non-ASCII inputs unchanged.
• Case-mapping and whitespace tables: the four flat int[0x110000]/boolean[0x110000] arrays (≈13.7 MB of statics) become 256-entry pages where unmapped pages share a single zero/empty page — ≈160 KB total (33–36 distinct pages per case table). Trade-off: translating uniformly random non-ASCII code points measures up to 10% slower from the extra indirection; ASCII paths and trims are unchanged.
• Trims: whitespace scanning uses a SWAR stop mask over the two contiguous ASCII whitespace ranges, and the forward scan skips homogeneous runs (e.g. space padding) with vectorized Arrays.mismatch. 1.8–3.8× on ASCII whitespace at 100 B+. Trade-off: inputs of randomly interleaved multi-byte whitespace measure 4–19% slower (unpredictable run-detection branch); noted in the commit message.

I/O

• DynamicSliceOutput.writeZero: one ensureSize + one Arrays.fill instead of a writeLong(0) loop that re-entered ensureSize every eight bytes — ~2×.
• BasicSliceInput.readByte: single bounds check and unchecked read instead of routing through read()'s −1 sentinel and a second check; semantics (including the exception at end of input) unchanged.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 7299bce8-85ab-4609-8a09-b89a003f128f

📥 Commits

Reviewing files that changed from the base of the PR and between 95a137d and 7b32fa1.

📒 Files selected for processing (9)

• src/main/java/io/airlift/slice/BasicSliceInput.java
• src/main/java/io/airlift/slice/DynamicSliceOutput.java
• src/main/java/io/airlift/slice/Slice.java
• src/main/java/io/airlift/slice/SliceUtf8.java
• src/main/java/io/airlift/slice/Slices.java
• src/test/java/io/airlift/slice/TestSlice.java
• src/test/java/io/airlift/slice/TestSliceOutput.java
• src/test/java/io/airlift/slice/TestSliceUtf8.java
• src/test/java/io/airlift/slice/TestSlices.java

---

📝 Walkthrough

Walkthrough

The changes optimize forward and reverse slice searches with word-based candidate detection and anchor verification. UTF-8 case conversion and whitespace trimming now use paged lookup data and word-at-a-time scanning. Basic slice input uses explicit bounds checks with unchecked reads, while dynamic zero writing validates lengths and clears backing regions directly. Slice growth padding is bounded, and randomized and regression tests cover these behaviors.

🚥 Pre-merge checks | ✅ 2

✅ Passed checks (2 passed)

Check name	Status	Explanation
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.

---

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.}