[kolkov]

Summary

• CompositeSequenceDFA: 25x speedup for overlapping char class patterns like \w+[0-9]+
• FindAllIndexCompact: Zero per-match allocations API (60974 → 9 allocations)

Performance

Engine	Throughput	vs Stdlib
CompositeSearcher (old)	56 MB/s	4.7x
CompositeSequenceDFA (new)	300 MB/s	25x

Implementation

• NFA subset construction for correct overlap handling
• Byte class reduction: 256 bytes → 3-8 equivalence classes
• First-part skip optimization: O(1) check if position can start match
• Loop unrolling: 4 bytes per iteration (Rust-inspired)

Test plan

• go test ./... passes
• golangci-lint run - 0 issues
• Benchmarks show 25x speedup

Closes #83

[github-actions]

Benchmark Comparison

Comparing main → PR #84

Summary: geomean 243.6n 243.4n -0.05%

⚠️ Potential regressions detected:

geomean                               ³                +0.00%               ³
geomean                               ³                +0.00%               ³
geomean                         ³                +0.00%               ³
geomean                         ³                +0.00%               ³
AhoCorasickVsStdlib/coregex_Count-4                     1.470µ ± ∞ ¹    1.486µ ± ∞ ¹     +1.09% (p=0.008 n=5)
AhoCorasickManyPatterns/coregex_25_patterns-4           71.80n ± ∞ ¹    80.36n ± ∞ ¹    +11.92% (p=0.008 n=5)
BranchDispatch_Stdlib/Digits-4                          133.5n ± ∞ ¹    135.8n ± ∞ ¹     +1.72% (p=0.016 n=5)
BranchDispatch_Stdlib/NoMatch-4                         77.26n ± ∞ ¹    77.85n ± ∞ ¹     +0.76% (p=0.032 n=5)
AnchoredAlt_ManyBranches_Stdlib/NoMatch-4               85.56n ± ∞ ¹    86.07n ± ∞ ¹     +0.60% (p=0.016 n=5)
FirstByteReject_Stdlib-4                                40.32n ± ∞ ¹    40.74n ± ∞ ¹     +1.04% (p=0.008 n=5)

Full results available in workflow artifacts. CI runners have ~10-20% variance.
For accurate benchmarks, run locally: ./scripts/bench.sh --compare