v0.10.0: Fat Teddy 16-bucket SIMD for 33-64 patterns (#68)

Fat Teddy AVX2 implementation with 9+ GB/s throughput for 33-64 patterns.

Key features:
- 16 buckets (vs Slim Teddy's 8) = 2x pattern capacity  
- AVX2 assembly with VPALIGNR half-shift algorithm
- Aho-Corasick fallback for small haystacks (<64 bytes)
- Pure Go scalar fallback for non-AVX2 platforms

Performance:
- 40 patterns: 9.1 GB/s (73x faster than Aho-Corasick)
- Small haystack fallback: 2.4x faster (110ns vs 267ns)

Author: kolkov

CHANGELOG.md

@@ -15,6 +15,49 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ---
 
+## [0.10.0] - 2026-01-07
+
+### Added
+- **Fat Teddy 16-bucket SIMD prefilter for 33-64 patterns**
+  - New strategy tier: Slim Teddy (2-32 patterns) → Fat Teddy (33-64) → Aho-Corasick (>64)
+  - AVX2 assembly implementation with 9+ GB/s throughput
+  - 16 buckets (vs Slim Teddy's 8) = 2x pattern capacity
+  - Pure Go scalar fallback for non-AVX2 platforms
+  - Algorithm from Rust aho-corasick `generic.rs` Fat<V, 2> implementation
+
+- **Aho-Corasick fallback for small haystacks with Fat Teddy**
+  - Fat Teddy's AVX2 SIMD has setup overhead slower than Aho-Corasick on small inputs
+  - Automatic fallback for haystacks < 64 bytes (threshold based on benchmarks)
+  - 2.4x faster on 37-byte haystacks with 50 patterns (267ns → 110ns)
+  - Follows Rust regex's `minimum_len()` approach (`builder.rs:585`)
+
+### Technical Details
+- **fatTeddyMasks struct**: 32-byte SIMD masks (256-bit AVX2 vectors)
+  - Low lane (bytes 0-15): buckets 0-7
+  - High lane (bytes 16-31): buckets 8-15
+- **AVX2 algorithm**:
+  - VBROADCASTI128: Load 16 bytes, duplicate to both lanes
+  - VPSHUFB: Parallel nibble lookup in bucket masks
+  - VPALIGNR $15: Half-shift for 2-byte fingerprint alignment
+  - VPMOVMSKB: Extract 32-bit candidate mask
+
+### Performance
+| Patterns | Engine | Throughput | vs Aho-Corasick |
+|----------|--------|------------|-----------------|
+| 40 | Fat Teddy AVX2 | 9.1 GB/s | **73x faster** |
+| 40 | Fat Teddy scalar | 228 MB/s | **1.5x faster** |
+| 70 | Aho-Corasick | 152 MB/s | baseline |
+
+### Files
+- `prefilter/teddy_fat.go` - Fat Teddy core implementation + MinimumLen()
+- `prefilter/teddy_fat_amd64.go` - AVX2 dispatch
+- `prefilter/teddy_avx2_amd64.s` - AVX2 assembly (~300 lines)
+- `meta/meta.go` - Aho-Corasick fallback for small haystacks
+- `meta/strategy.go` - strategy selection update (32→Fat Teddy, >64→Aho-Corasick)
+- `meta/fat_teddy_fallback_test.go` - tests for fallback logic
+
+---
+
 ## [0.9.5] - 2026-01-06
 
 ### Changed
@@ -1111,8 +1154,9 @@ v0.7.0 → OnePass DFA (DONE ✅)
 v0.8.0 → ReverseInner strategy (DONE ✅)
 v0.8.14-18 → GoAWK integration, Teddy, BoundedBacktracker (DONE ✅)
 v0.8.19 → FindAll ReverseSuffix optimization (DONE ✅)
-v0.8.20 → ReverseSuffixSet for multi-suffix patterns (DONE ✅) ← CURRENT
-v0.9.0 → Beta testing period
+v0.8.20 → ReverseSuffixSet for multi-suffix patterns (DONE ✅)
+v0.9.x → Performance tuning, Teddy 2-byte fingerprint (DONE ✅)
+v0.10.0 → Fat Teddy 33-64 patterns, AVX2 SIMD (DONE ✅) ← CURRENT
 v1.0.0 → Stable release (API frozen)
 ```
 
@@ -1142,7 +1186,13 @@ v1.0.0 → Stable release (API frozen)
 
 ---
 
-[Unreleased]: https://github.com/coregx/coregex/compare/v0.9.0...HEAD
+[Unreleased]: https://github.com/coregx/coregex/compare/v0.10.0...HEAD
+[0.10.0]: https://github.com/coregx/coregex/releases/tag/v0.10.0
+[0.9.5]: https://github.com/coregx/coregex/releases/tag/v0.9.5
+[0.9.4]: https://github.com/coregx/coregex/releases/tag/v0.9.4
+[0.9.3]: https://github.com/coregx/coregex/releases/tag/v0.9.3
+[0.9.2]: https://github.com/coregx/coregex/releases/tag/v0.9.2
+[0.9.1]: https://github.com/coregx/coregex/releases/tag/v0.9.1
 [0.9.0]: https://github.com/coregx/coregex/releases/tag/v0.9.0
 [0.8.24]: https://github.com/coregx/coregex/releases/tag/v0.8.24
 [0.8.23]: https://github.com/coregx/coregex/releases/tag/v0.8.23

README.md

@@ -73,7 +73,7 @@ Cross-language benchmarks on 6MB input ([source](https://github.com/kolkov/regex
 - IP/phone patterns (`\d+\.\d+\.\d+\.\d+`) — optimized DFA strategy
 - Suffix patterns (`.*\.log`, `.*\.txt`) — reverse search optimization
 - Inner literals (`.*error.*`, `.*@example\.com`) — bidirectional DFA
-- Multi-pattern (`foo|bar|baz|...`) — Teddy (≤8) or Aho-Corasick (>8 patterns)
+- Multi-pattern (`foo|bar|baz|...`) — Slim Teddy (≤32), Fat Teddy (33-64), or Aho-Corasick (>64)
 
 ## Features
 
@@ -86,9 +86,10 @@ coregex automatically selects the optimal engine:
 | ReverseInner | `.*keyword.*` | 100-200x |
 | ReverseSuffix | `.*\.txt` | 100-220x |
 | LazyDFA | IP, complex patterns | 10-150x |
-| AhoCorasick | `a\|b\|c\|...\|z` (>8 patterns) | 75-113x |
+| AhoCorasick | `a\|b\|c\|...\|z` (>64 patterns) | 75-113x |
 | CharClassSearcher | `[\w]+`, `\d+` | 4-25x |
-| Teddy | `foo\|bar\|baz` (2-8 patterns) | 15-240x |
+| Slim Teddy | `foo\|bar\|baz` (2-32 patterns) | 15-240x |
+| Fat Teddy | 33-64 patterns | 60-73x |
 | OnePass | Anchored captures | 10x |
 | BoundedBacktracker | Small patterns | 2-5x |
 
@@ -168,7 +169,8 @@ Input → Prefilter (SIMD) → Engine → Match Result
 **SIMD Primitives** (AMD64):
 - `memchr` — single byte search (AVX2)
 - `memmem` — substring search (SSSE3)
-- `teddy` — multi-pattern search (SSSE3)
+- `Slim Teddy` — multi-pattern search, 2-32 patterns (SSSE3)
+- `Fat Teddy` — multi-pattern search, 33-64 patterns (AVX2, 9+ GB/s)
 
 Pure Go fallback on other architectures.
 

ROADMAP.md

@@ -2,7 +2,7 @@
 
 > **Strategic Focus**: Production-grade regex engine with RE2/rust-regex level optimizations
 
-**Last Updated**: 2025-12-13 | **Current Version**: v0.8.22 | **Target**: v1.0.0 stable
+**Last Updated**: 2026-01-07 | **Current Version**: v0.10.0 | **Target**: v1.0.0 stable
 
 ---
 
@@ -12,7 +12,7 @@ Build a **production-ready, high-performance regex engine** for Go that matches
 
 ### Current State vs Target
 
-| Metric | Current (v0.8.22) | Target (v1.0.0) |
+| Metric | Current (v0.10.0) | Target (v1.0.0) |
 |--------|-------------------|-----------------|
 | Inner literal speedup | **87-3154x** | ✅ Achieved |
 | Case-insensitive speedup | **263x** | ✅ Achieved |
@@ -21,7 +21,9 @@ Build a **production-ready, high-performance regex engine** for Go that matches
 | Small string perf | **1.4-20x faster** | ✅ Achieved |
 | Reverse search | **Yes (4 strategies)** | ✅ Achieved |
 | OnePass DFA | **Yes** | ✅ Achieved |
-| Teddy SIMD prefilter | **Yes** | ✅ Achieved |
+| Slim Teddy (2-32 patterns) | **Yes (SSSE3)** | ✅ Achieved |
+| Fat Teddy (33-64 patterns) | **Yes (AVX2, 9GB/s)** | ✅ Achieved |
+| Aho-Corasick (>64 patterns) | **Yes** | ✅ Achieved |
 | BoundedBacktracker | **Yes** | ✅ Achieved |
 | CharClassSearcher | **Yes (23x, 2x vs Rust)** | ✅ Achieved |
 | ARM NEON SIMD | No | Planned |
@@ -32,9 +34,9 @@ Build a **production-ready, high-performance regex engine** for Go that matches
 ## Release Strategy
 
 ```
-v0.8.22 (Current) ✅ → Small string optimization (1.4-20x faster)
+v0.10.0 (Current) ✅ → Fat Teddy 33-64 patterns (AVX2, 9GB/s)
          ↓
-v0.9.x → Beta testing, API stabilization
+v0.11.x → API stabilization, performance tuning
          ↓
 v1.0.0-rc → Feature freeze, API locked
          ↓
@@ -56,6 +58,8 @@ v1.0.0 STABLE → Production release with API stability guarantee
 - ✅ **v0.8.20**: ReverseSuffixSet for multi-suffix patterns (34-385x faster)
 - ✅ **v0.8.21**: CharClassSearcher (23x faster, 2x faster than Rust!)
 - ✅ **v0.8.22**: Small string optimization (1.4-20x faster on ~44B inputs)
+- ✅ **v0.9.x**: DigitPrefilter, Aho-Corasick integration, Teddy 2-byte fingerprint
+- ✅ **v0.10.0**: Fat Teddy 16-bucket SIMD (33-64 patterns, 9+ GB/s)
 
 ---
 
@@ -147,20 +151,21 @@ v1.0.0 STABLE → Production release with API stability guarantee
 
 ## Feature Comparison Matrix
 
-| Feature | RE2 | rust-regex | coregex v0.8.20 | coregex v1.0 |
+| Feature | RE2 | rust-regex | coregex v0.10.0 | coregex v1.0 |
 |---------|-----|------------|-----------------|--------------|
 | Lazy DFA | ✅ | ✅ | ✅ | ✅ |
 | Thompson NFA | ✅ | ✅ | ✅ | ✅ |
 | PikeVM | ✅ | ✅ | ✅ | ✅ |
-| Teddy SIMD | ❌ | ✅ | ✅ | ✅ |
+| Slim Teddy (≤32) | ❌ | ✅ | ✅ | ✅ |
+| Fat Teddy (33-64) | ❌ | ✅ | ✅ | ✅ |
 | Start State Cache | 8 | 6 | 6 | ✅ |
 | Reverse Search | ✅ | ✅ (3) | ✅ (4) | ✅ |
 | ReverseSuffixSet | ❌ | ❌ | ✅ | ✅ |
 | OnePass DFA | ✅ | ✅ | ✅ | ✅ |
 | BoundedBacktracker | ✅ | ✅ | ✅ | ✅ |
 | Named Captures | ✅ | ✅ | ✅ | ✅ |
 | Prefilter Tracking | ✅ | ✅ | ✅ | ✅ |
-| Aho-Corasick | ❌ | ✅ | ❌ | Planned |
+| Aho-Corasick | ❌ | ✅ | ✅ | ✅ |
 | ARM NEON | ❌ | ✅ | ❌ | Planned |
 | Look-around | ✅ | ❌ | ❌ | Planned |
 
@@ -188,7 +193,6 @@ v1.0.0 STABLE → Production release with API stability guarantee
 |---------|--------|----------|
 | ARM NEON SIMD | Planned | Medium |
 | Look-around assertions | Planned | Medium |
-| Aho-Corasick for large sets | Planned | Low |
 | API stability guarantee | Required | High |
 
 ---
@@ -223,7 +227,10 @@ Reference implementations available locally:
 
 | Version | Date | Type | Key Changes |
 |---------|------|------|-------------|
-| **v0.8.20** | 2025-12-12 | Performance | **ReverseSuffixSet (34-385x faster) - NOT in rust-regex!** |
+| **v0.10.0** | 2026-01-07 | Feature | **Fat Teddy 33-64 patterns (AVX2, 9+ GB/s)** |
+| v0.9.5 | 2026-01-06 | Fix | Teddy limit 8→32, literal extraction fix |
+| v0.9.0-v0.9.4 | 2026-01-05 | Performance | DigitPrefilter, Aho-Corasick, 2-byte fingerprint |
+| v0.8.20 | 2025-12-12 | Performance | ReverseSuffixSet (34-385x faster) |
 | v0.8.19 | 2025-12-12 | Performance | FindAll ReverseSuffix (87x faster) |
 | v0.8.18 | 2025-12-12 | Performance | Teddy prefilter for alternations (242x faster) |
 | v0.8.17 | 2025-12-12 | Feature | BoundedBacktracker engine |
@@ -239,4 +246,4 @@ Reference implementations available locally:
 
 ---
 
-*Current: v0.8.20 | Next: v0.9.x (Beta) | Target: v1.0.0*
+*Current: v0.10.0 | Next: v0.11.x (API stabilization) | Target: v1.0.0*

meta/ahocorasick_test.go

@@ -5,27 +5,31 @@ import (
 	"testing"
 )
 
-// TestAhoCorasickStrategySelection verifies that patterns with >32 literals
+// TestAhoCorasickStrategySelection verifies that patterns with >64 literals
 // select UseAhoCorasick strategy.
 func TestAhoCorasickStrategySelection(t *testing.T) {
-	// Pattern with 33 literals (above Teddy's limit of 32)
+	// Pattern with 65 literals (above Teddy's limit of 64)
+	// Teddy supports up to 64 patterns via Slim (2-32) and Fat (33-64) variants.
+	// For >64 patterns, Aho-Corasick is selected.
 	// Each literal >= 3 bytes, all complete (no regex meta-characters)
 	// IMPORTANT: No shared prefixes! Go's regex parser factors common prefixes,
 	// e.g., "two|three" becomes "t(wo|hree)", which extracts only "t" as incomplete.
-	// Using fruits/vegetables/colors with unique first letters.
-	pattern := `apple|banana|cherry|date|elderberry|fig|grape|honeydew|` +
-		`imbe|jackfruit|kiwi|lemon|mango|nectarine|orange|papaya|` +
-		`quince|raspberry|strawberry|tomato|ugli|vanilla|watermelon|` +
-		`ximenia|yuzu|zucchini|apricot|blueberry|coconut|dragonfruit|` +
-		`eggplant|feijoa|guava`
+	// Using unique words with different first characters.
+	pattern := `alpha|bravo|charlie|delta|echo|foxtrot|golf|hotel|india|juliet|` + // 10
+		`kilo|lima|mike|november|oscar|papa|quebec|romeo|sierra|tango|` + // 20
+		`uniform|victor|whiskey|xray|yankee|zulu|anise|basil|cilantro|dill|` + // 30
+		`endive|fennel|ginger|hops|ivory|jasmine|kelp|lavender|mint|nutmeg|` + // 40
+		`oregano|parsley|quassia|rosemary|sage|thyme|urtica|verbena|wasabi|xylose|` + // 50
+		`yarrow|zinnia|acacia|bamboo|cactus|dahlia|ebony|fern|grass|holly|` + // 60
+		`iris|juniper|kudzu|lotus|moss|nettle|oak` // 67
 
 	re, err := Compile(pattern)
 	if err != nil {
 		t.Fatalf("Compile(%q) failed: %v", pattern, err)
 	}
 
 	if re.Strategy() != UseAhoCorasick {
-		t.Errorf("Strategy() = %s, want UseAhoCorasick", re.Strategy())
+		t.Errorf("Strategy() = %s, want UseAhoCorasick for 67 patterns", re.Strategy())
 	}
 }
 
@@ -223,22 +227,25 @@ func TestAhoCorasickCount(t *testing.T) {
 	}
 }
 
-// TestAhoCorasickLargePatternSet tests with many patterns.
+// TestAhoCorasickLargePatternSet tests with many patterns (>64 triggers Aho-Corasick).
 func TestAhoCorasickLargePatternSet(t *testing.T) {
-	// 35 patterns - above Teddy's limit of 32
+	// 70 patterns - above Teddy's limit of 64 (Fat Teddy handles up to 64)
 	// No shared prefixes to avoid Go's regex parser factoring (e.g., "two|three" → "t(wo|hree)")
-	pattern := `alpha|bravo|charlie|delta|echo|foxtrot|golf|hotel|india|juliet|` +
-		`kilo|lima|mike|november|oscar|papa|quebec|romeo|sierra|tango|` +
-		`uniform|victor|whiskey|xray|yankee|zulu|anise|basil|cilantro|dill|` +
-		`endive|fennel|ginger|hops|ivy`
+	pattern := `alpha|bravo|charlie|delta|echo|foxtrot|golf|hotel|india|juliet|` + // 10
+		`kilo|lima|mike|november|oscar|papa|quebec|romeo|sierra|tango|` + // 20
+		`uniform|victor|whiskey|xray|yankee|zulu|anise|basil|cilantro|dill|` + // 30
+		`endive|fennel|ginger|hops|ivory|jasmine|kelp|lavender|mint|nutmeg|` + // 40
+		`oregano|parsley|quassia|rosemary|sage|thyme|urtica|verbena|wasabi|xylose|` + // 50
+		`yarrow|zinnia|acacia|bamboo|cactus|dahlia|ebony|fern|grass|holly|` + // 60
+		`iris|juniper|kudzu|lotus|moss|nettle|oak|plum|reed|sorrel` // 70
 
 	re, err := Compile(pattern)
 	if err != nil {
 		t.Fatalf("Compile(%q) failed: %v", pattern, err)
 	}
 
 	if re.Strategy() != UseAhoCorasick {
-		t.Errorf("Strategy() = %s, want UseAhoCorasick for 35 patterns", re.Strategy())
+		t.Errorf("Strategy() = %s, want UseAhoCorasick for 70 patterns", re.Strategy())
 	}
 
 	haystack := []byte("this is alpha and omega, with bravo and tango at the end")

meta/config.go

@@ -55,7 +55,8 @@ type Config struct {
 	MinLiteralLen int
 
 	// MaxLiterals limits the number of literals to extract for prefiltering.
-	// Default: 64
+	// Must be > 64 to properly detect patterns that exceed Teddy's capacity.
+	// Default: 256 (allows detecting patterns with >64 literals for Aho-Corasick)
 	MaxLiterals int
 
 	// MaxRecursionDepth limits recursion during NFA compilation.
@@ -83,8 +84,8 @@ func DefaultConfig() Config {
 		EnablePrefilter:      true,
 		MaxDFAStates:         10000,
 		DeterminizationLimit: 1000,
-		MinLiteralLen:        1, // Allow single-byte prefilters (memchr) like Rust
-		MaxLiterals:          64,
+		MinLiteralLen:        1,   // Allow single-byte prefilters (memchr) like Rust
+		MaxLiterals:          256, // Allow detecting >64 literals for Aho-Corasick
 		MaxRecursionDepth:    100,
 	}
 }