Optimization Roadmap: mimalloc Gap Analysis & Phase 1-3 Plan

Add comprehensive mimalloc vs hakmem performance gap analysis (2.5x).

Gap sources (ranked by ROI):
1. Observation tax (stats macros): +2-3% overhead
2. Policy snapshot: +10-15% overhead (per-call TLS read + atomic sync)
3. Header management: +5-10% overhead (1-byte per block)
4. Wrapper layer: +5-10% overhead (LD_PRELOAD interception)
5. Routing switch: +3-5% overhead (5-way switch)

Optimization roadmap:
- Phase 1 (Quick Wins): +4-7% via FREE adoption + compile-out stats + inline
- Phase 2 (Structural): +5-10% via header tax removal + C0-C3 path + jump table
- Phase 3 (Cache): +12-22% via prefetch + cache optimization + static routing

Expected outcome: 52-68M ops/s (vs current 50.7M, gap from 2.5x → 1.9x)

Architectural reality: hakmem's 4-5 layer design adds 50-100x instruction
overhead vs mimalloc's 1-layer design. Gap closure caps at ~1.9x without
fundamental redesign.

Next immediate step: Implement Phase 1A (FREE adoption + compile-out stats)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

CURRENT_TASK.md

@@ -37,15 +37,77 @@
 
 ---
 
-## 次の攻め先: Profile Adoption & Remaining Optimization
+## 次の攻め先: mimalloc Gap Closure Roadmap (2.5x → 1.9x)
 
-**優先度 A** - Free 昇格:
-- FREE-TINY-FAST-DUALHOT-1 を MIXED_TINYV3_C7_SAFE profile に取り込む準備
-- ENV: HAKMEM_TINY_LARSON_FIX=0 as default (DUALHOT ON)
+**Gap Analysis**: hakmem 50.7M ops/s vs mimalloc 127M ops/s
 
-**優先度 B** - Alloc 構造最適化（deferred）:
-- `malloc` / Front Gate の "構造的" オーバーヘッド改善
-- PGO / const propagation / inline optimizations
+根本原因（ROI順）:
+1. **Observation tax** (+2-3%): Stats macros branch even when OFF
+2. **Policy snapshot** (+10-15%): Per-call TLS policy read + atomic sync
+3. **Header management** (+5-10%): 1-byte header per block
+4. **Wrapper layer** (+5-10%): malloc → tiny_alloc_gate_fast + security checks
+5. **Routing switch** (+3-5%): Per-call switch statement
+
+### Phase 1: Quick Wins (Week 1) - Target: +4-7% (52-56M ops/s)
+
+**優先度 A1** - FREE 勝ち箱の本線昇格:
+- HAKMEM_FREE_TINY_FAST_HOTCOLD=1 を MIXED_TINYV3_C7_SAFE default
+- FREE-TINY-FAST-DUALHOT-1 のデフォルト有効化
+- Expected: +2-3% (DUALHOT 効果は既に測定済み +13%)
+
+**優先度 A2** - 観測税ゼロ化 (Compile-out stats):
+- Add HAKMEM_DEBUG_COUNTERS compile-time flag (default 0)
+- When 0: `#define ALLOC_GATE_STAT_INC(x) do {} while(0)` (zero cost)
+- Files: `alloc_gate_stats_box.h`, `free_path_stats_box.h`, `tiny_front_stats_box.h`, `free_tiny_fast_hotcold_stats_box.h`
+- Expected: +2-3% (eliminate branching on all stats)
+
+**優先度 A3** - Inline header write:
+- Add `__attribute__((always_inline))` to `tiny_region_id_write_header()`
+- Eliminate function call overhead in hot path
+- Expected: +1-2%
+
+### Phase 2: Structural Changes (Weeks 2-3) - Target: +5-10% (55-61M ops/s)
+
+**優先度 B1** - C4-C7 header tax削減:
+- Remove 1-byte header for C6 (512B) / C7 (1024B) allocations
+- Use registry-only lookup on free
+- Expected: +3-5% (C6/C7 = 30% of workload, no header = 10% size savings)
+
+**優先度 B2** - C0-C3 専用 fast path:
+- Create `malloc_tiny_fast_c0c3()` entry point (no policy snapshot)
+- Conditional dispatch from wrapper based on size
+- Expected: +1-2%
+
+**優先度 B3** - Routing jump table:
+- Replace switch(route_kind) with function pointer array
+- Reduce branch prediction misses (5-way switch → direct dispatch)
+- Expected: +1-3%
+
+### Phase 3: Cache Locality (Weeks 4-5) - Target: +12-22% (57-68M ops/s)
+
+**優先度 C1** - TLS cache prefetch:
+- `__builtin_prefetch(g_small_policy_v7, 0, 3)` on malloc entry
+- Improve L1 hit rate on cold start
+- Expected: +2-4%
+
+**優先度 C2** - Slab metadata cache optimization:
+- Profile cache-miss hotspots (policy struct, slab metadata)
+- Hot/cold split of metadata
+- Inline first slab descriptor
+- Expected: +5-10%
+
+**優先度 C3** - Static routing (if no learner):
+- Detect static routes at init
+- Bypass policy snapshot entirely
+- Expected: +5-8%
+
+### Architectural Insight (Long-term)
+
+**Reality check**: hakmem 4-5 layer design (wrapper → gate → policy → route → handler) adds 50-100x instruction overhead vs mimalloc's 1-layer TLS buckets.
+
+**Maximum realistic** without redesign: 65-70M ops/s (still ~1.9x gap)
+
+**Future pivot**: Consider static-compiled routing + optional learner (not per-call policy)
 
 ---