Phase MID-V35-HOTPATH-OPT-1 complete: +7.3% on C6-heavy

Step 0: Geometry SSOT
  - New: core/box/smallobject_mid_v35_geom_box.h (L1/L2 consistency)
  - Fix: C6 slots/page 102→128 in L2 (smallobject_cold_iface_mid_v3.c)
  - Applied: smallobject_mid_v35.c, smallobject_segment_mid_v3.c

Step 1-3: ENV gates for hotpath optimizations
  - New: core/box/mid_v35_hotpath_env_box.h
    * HAKMEM_MID_V35_HEADER_PREFILL (default 0)
    * HAKMEM_MID_V35_HOT_COUNTS (default 1)
    * HAKMEM_MID_V35_C6_FASTPATH (default 0)
  - Implementation: smallobject_mid_v35.c
    * Header prefill at refill boundary (Step 1)
    * Gated alloc_count++ in hot path (Step 2)
    * C6 specialized fast path with constant slot_size (Step 3)

A/B Results:
  C6-heavy (257–768B): 8.75M→9.39M ops/s (+7.3%, 5-run mean) ✅
  Mixed (16–1024B): 9.98M→9.96M ops/s (-0.2%, within noise) ✓

Decision: FROZEN - defaults OFF, C6-heavy推奨ON, Mixed現状維持
Documentation: ENV_PROFILE_PRESETS.md updated

🤖 Generated with Claude Code

Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>

CURRENT_TASK.md

@@ -1,6 +1,27 @@
 # 本線タスク（現在）
 
-## 現在地: Phase MID-V35-HOTPATH-OPT-1-DESIGN へ
+## 現在地: Phase MID-V35-HOTPATH-OPT-1 完了 → 次フェーズ選定待ち
+
+---
+
+### Status: Phase MID-V35-HOTPATH-OPT-1 FROZEN ✅
+
+**Summary**:
+- **Design**: Step 0-3（Geometry SSOT + Header prefill + Hot counts + C6 fastpath）
+- **C6-heavy (257–768B)**: **+7.3%** improvement ✅ (8.75M → 9.39M ops/s, 5-run mean)
+- **Mixed (16–1024B)**: **-0.2%** (誤差範囲, ±2%以内) ✓
+- **Decision**: デフォルトOFF/FROZEN（全3ノブ）、C6-heavy推奨ON、Mixed現状維持
+- **Key Finding**:
+  - Step 0: L1/L2 geometry mismatch 修正（C6 102→128 slots）
+  - Step 1-3: refill 境界移動 + 分岐削減 + constant 最適化で +7.3%
+  - Mixed では MID_V3(C6-only) 固定なため効果微小
+
+**Deliverables**:
+- `core/box/smallobject_mid_v35_geom_box.h` (新規)
+- `core/box/mid_v35_hotpath_env_box.h` (新規)
+- `core/smallobject_mid_v35.c` (Step 1-3 統合)
+- `core/smallobject_cold_iface_mid_v3.c` (Step 0 + Step 1)
+- `docs/analysis/ENV_PROFILE_PRESETS.md` (更新)
 
 ---
 

core/box/mid_v35_hotpath_env_box.h

@@ -0,0 +1,52 @@
+// mid_v35_hotpath_env_box.h - Phase MID-V35-HOTPATH-OPT-1 ENV gates
+//
+// Step 1: HAKMEM_MID_V35_HEADER_PREFILL (default 0)
+//   - ON: prefill headers at refill boundary, skip in hot path
+// Step 2: HAKMEM_MID_V35_HOT_COUNTS (default 1)
+//   - OFF: skip alloc_count++ in hot path (free_count/retire kept)
+// Step 3: HAKMEM_MID_V35_C6_FASTPATH (default 0)
+//   - ON: use specialized C6 fast path with constant slot size
+//
+#ifndef HAKMEM_MID_V35_HOTPATH_ENV_BOX_H
+#define HAKMEM_MID_V35_HOTPATH_ENV_BOX_H
+
+#include <stdlib.h>
+#include <stdbool.h>
+
+// Step 1: Header prefill at refill boundary
+// Default OFF (conservative)
+static inline bool mid_v35_header_prefill_enabled(void) {
+    static int g_enabled = -1;
+    if (__builtin_expect(g_enabled >= 0, 1)) {
+        return g_enabled == 1;
+    }
+    const char* e = getenv("HAKMEM_MID_V35_HEADER_PREFILL");
+    g_enabled = (e && *e == '1') ? 1 : 0;
+    return g_enabled == 1;
+}
+
+// Step 2: Hot counts (alloc_count++ in hot path)
+// Default ON (for compatibility/correctness)
+static inline bool mid_v35_hot_counts_enabled(void) {
+    static int g_enabled = -1;
+    if (__builtin_expect(g_enabled >= 0, 1)) {
+        return g_enabled == 1;
+    }
+    const char* e = getenv("HAKMEM_MID_V35_HOT_COUNTS");
+    g_enabled = (e && *e == '0') ? 0 : 1;  // default ON
+    return g_enabled == 1;
+}
+
+// Step 3: C6 specialized fast path
+// Default OFF (conservative)
+static inline bool mid_v35_c6_fastpath_enabled(void) {
+    static int g_enabled = -1;
+    if (__builtin_expect(g_enabled >= 0, 1)) {
+        return g_enabled == 1;
+    }
+    const char* e = getenv("HAKMEM_MID_V35_C6_FASTPATH");
+    g_enabled = (e && *e == '1') ? 1 : 0;
+    return g_enabled == 1;
+}
+
+#endif // HAKMEM_MID_V35_HOTPATH_ENV_BOX_H

core/box/smallobject_mid_v35_geom_box.h

@@ -0,0 +1,45 @@
+// smallobject_mid_v35_geom_box.h - Phase MID-V35-HOTPATH-OPT-1 Step 0
+//
+// Single Source of Truth for MID v3.5 geometry (C5-C7)
+// All layers (L1 HotBox, L2 ColdIface, Segment) must use these helpers.
+//
+// Geometry (64KB page):
+//   C5: 384B slots,  170 slots/page (257-384B)
+//   C6: 512B slots,  128 slots/page (385-512B)
+//   C7: 1024B slots, 64 slots/page  (513-1024B)
+//
+#ifndef HAKMEM_SMALLOBJECT_MID_V35_GEOM_BOX_H
+#define HAKMEM_SMALLOBJECT_MID_V35_GEOM_BOX_H
+
+#include <stdint.h>
+#include <stddef.h>
+
+// Page size for MID v3.5 (64KB)
+#define MID_V35_PAGE_SIZE (64 * 1024)
+
+// Slot size by class_idx (C5-C7 only, others return 0)
+static inline size_t mid_v35_slot_size(uint32_t class_idx) {
+    switch (class_idx) {
+        case 5: return 384;   // C5: 257-384B
+        case 6: return 512;   // C6: 385-512B
+        case 7: return 1024;  // C7: 513-1024B
+        default: return 0;
+    }
+}
+
+// Slots per page by class_idx (C5-C7 only, others return 0)
+static inline uint32_t mid_v35_slots_per_page(uint32_t class_idx) {
+    switch (class_idx) {
+        case 5: return 170;  // 65536 / 384 = 170
+        case 6: return 128;  // 65536 / 512 = 128
+        case 7: return 64;   // 65536 / 1024 = 64
+        default: return 0;
+    }
+}
+
+// Validate class_idx is in MID v3.5 range (C5-C7)
+static inline int mid_v35_class_valid(uint32_t class_idx) {
+    return (class_idx >= 5 && class_idx <= 7);
+}
+
+#endif // HAKMEM_SMALLOBJECT_MID_V35_GEOM_BOX_H

core/smallobject_cold_iface_mid_v3.c

@@ -5,6 +5,9 @@
 #include <stdio.h>
 #include "box/smallobject_cold_iface_mid_v3_box.h"
 #include "box/smallobject_stats_mid_v3_box.h"
+#include "box/smallobject_mid_v35_geom_box.h"  // Phase MID-V35-HOTPATH-OPT-1: geometry SSOT
+#include "box/mid_v35_hotpath_env_box.h"       // Phase MID-V35-HOTPATH-OPT-1: Step 1-3 ENV gates
+#include "tiny_region_id.h"                    // For tiny_region_id_write_header
 
 // SmallPageMeta is defined in smallobject_segment_mid_v3_box.h
 #include "box/smallobject_segment_mid_v3_box.h"
@@ -19,15 +22,9 @@ static __thread SmallSegment_MID_v3 *tls_mid_segment = NULL;
 // ============================================================================
 // Helper: class_idx to slots
 // ============================================================================
-
-static uint32_t class_idx_to_slots(uint32_t class_idx) {
-    switch (class_idx) {
-        case 5: return 170;  // C5: 257-384B
-        case 6: return 102;  // C6: 385-640B
-        case 7: return 64;   // C7: 641-1024B
-        default: return 0;
-    }
-}
+// Phase MID-V35-HOTPATH-OPT-1: Use geom_box as Single Source of Truth
+// See: core/box/smallobject_mid_v35_geom_box.h
+// (Removed local class_idx_to_slots() which had wrong C6 value: 102 instead of 128)
 
 // ============================================================================
 // Cold Interface Implementation
@@ -63,10 +60,21 @@ SmallPageMeta_MID_v3* small_cold_mid_v3_refill_page(uint32_t class_idx) {
 
     // Initialize page for allocation
     page->class_idx = class_idx;
-    page->capacity = class_idx_to_slots(class_idx);
+    page->capacity = mid_v35_slots_per_page(class_idx);
     page->alloc_count = 0;
     page->free_count = 0;
 
+    // Phase MID-V35-HOTPATH-OPT-1 Step 1: Header prefill at refill boundary
+    // When enabled, write all headers now so hot path can skip per-alloc header writes
+    if (mid_v35_header_prefill_enabled()) {
+        size_t slot_size = mid_v35_slot_size(class_idx);
+        uint32_t cap = page->capacity;
+        uint8_t *p = (uint8_t*)page_ptr;
+        for (uint32_t i = 0; i < cap; i++) {
+            tiny_region_id_write_header(p + i * slot_size, class_idx);
+        }
+    }
+
     return page;
 }
 

core/smallobject_mid_v35.c

@@ -13,6 +13,8 @@
 #include "box/smallobject_mid_v35_box.h"
 #include "box/smallobject_segment_mid_v3_box.h"
 #include "box/smallobject_cold_iface_mid_v3_box.h"
+#include "box/smallobject_mid_v35_geom_box.h"  // Phase MID-V35-HOTPATH-OPT-1: geometry SSOT
+#include "box/mid_v35_hotpath_env_box.h"       // Phase MID-V35-HOTPATH-OPT-1: Step 1-3 ENV gates
 #include "tiny_region_id.h"  // For tiny_region_id_write_header
 
 // SmallPageMeta is defined in smallobject_segment_mid_v3_box.h
@@ -34,26 +36,8 @@ static __thread SmallMidV35TlsCtx tls_mid_v35_ctx = {0};
 // ============================================================================
 // Slot Configuration (C5/C6/C7)
 // ============================================================================
-
-// Slot sizes for C5, C6, C7
-static const size_t g_slot_sizes[8] = {
-    0,     // C0: not used
-    0,     // C1: not used
-    0,     // C2: not used
-    0,     // C3: not used
-    0,     // C4: not used (ULTRA handles this)
-    384,   // C5: 257-384 bytes → 384 byte slots
-    512,   // C6: 385-512 bytes → 512 byte slots
-    1024,  // C7: 513-1024 bytes → 1024 byte slots (ULTRA handles this)
-};
-
-// Slots per 64KB page
-static const uint32_t g_slots_per_page[8] = {
-    0, 0, 0, 0, 0,
-    170,  // C5: 65536 / 384 = 170
-    128,  // C6: 65536 / 512 = 128
-    64,   // C7: 65536 / 1024 = 64
-};
+// Phase MID-V35-HOTPATH-OPT-1: Use geom_box as Single Source of Truth
+// See: core/box/smallobject_mid_v35_geom_box.h
 
 // ============================================================================
 // Init
@@ -74,26 +58,57 @@ void* small_mid_v35_alloc(uint32_t class_idx, size_t size) {
 
     SmallMidV35TlsCtx *ctx = &tls_mid_v35_ctx;
 
-    // Fast path: allocate from TLS cached page
+    // ========================================================================
+    // Step 3: C6 specialized fast path (constant slot size = 512)
+    // ========================================================================
+    if (mid_v35_c6_fastpath_enabled() && class_idx == 6) {
+        void *page = ctx->page[6];
+        uint32_t off = ctx->offset[6];
+        if (page && off < ctx->capacity[6]) {
+            // C6: slot_size = 512 (constant, compiler can optimize)
+            void *base = (char*)page + off * 512;
+            ctx->offset[6] = off + 1;
+
+            // Step 2: HOT_COUNTS gate
+            if (mid_v35_hot_counts_enabled() && ctx->meta[6]) {
+                ctx->meta[6]->alloc_count++;
+            }
+
+            // Step 1: HEADER_PREFILL gate
+            if (!mid_v35_header_prefill_enabled()) {
+                tiny_region_id_write_header(base, 6);
+            }
+
+            return (char*)base + 1;
+        }
+        // Fall through to slow path
+    }
+
+    // ========================================================================
+    // Generic fast path: allocate from TLS cached page
+    // ========================================================================
     if (ctx->page[class_idx] && ctx->offset[class_idx] < ctx->capacity[class_idx]) {
-        size_t slot_size = g_slot_sizes[class_idx];
+        size_t slot_size = mid_v35_slot_size(class_idx);
         void *base = (char*)ctx->page[class_idx] + ctx->offset[class_idx] * slot_size;
         ctx->offset[class_idx]++;
 
-        // Update page metadata
-        if (ctx->meta[class_idx]) {
+        // Step 2: HOT_COUNTS gate - Update page metadata
+        if (mid_v35_hot_counts_enabled() && ctx->meta[class_idx]) {
             ctx->meta[class_idx]->alloc_count++;
         }
 
-        // Write header (1-byte Tiny header with class_idx)
-        // Note: Assumes HAKMEM_TINY_HEADER_CLASSIDX is enabled
+        // Step 1: HEADER_PREFILL gate - Write header if not prefilled
+        if (!mid_v35_header_prefill_enabled()) {
             tiny_region_id_write_header(base, class_idx);
+        }
 
         // Return USER pointer (BASE + 1 byte header)
         return (char*)base + 1;
     }
 
+    // ========================================================================
     // Slow path: need new page via ColdIface
+    // ========================================================================
     SmallPageMeta_MID_v3 *page = small_cold_mid_v3_refill_page(class_idx);
     if (!page) {
         // Fallback to legacy or return NULL
@@ -103,14 +118,18 @@ void* small_mid_v35_alloc(uint32_t class_idx, size_t size) {
     // Update TLS cache
     ctx->page[class_idx] = page->ptr;
     ctx->offset[class_idx] = 1;  // First slot already allocated
-    ctx->capacity[class_idx] = g_slots_per_page[class_idx];
+    ctx->capacity[class_idx] = mid_v35_slots_per_page(class_idx);
     ctx->meta[class_idx] = page;
 
-    // Record first allocation in page metadata
+    // Step 2: HOT_COUNTS gate - Record first allocation in page metadata
+    if (mid_v35_hot_counts_enabled()) {
         page->alloc_count = 1;
+    }
 
-    // Write header for first slot
+    // Step 1: HEADER_PREFILL gate - Write header for first slot if not prefilled
+    if (!mid_v35_header_prefill_enabled()) {
         tiny_region_id_write_header(page->ptr, class_idx);
+    }
 
     // Return first slot (USER pointer)
     return (char*)page->ptr + 1;

core/smallobject_segment_mid_v3.c

@@ -7,6 +7,7 @@
 #include <stdio.h>
 #include "box/smallobject_segment_mid_v3_box.h"
 #include "box/region_id_v6_box.h"
+#include "box/smallobject_mid_v35_geom_box.h"  // Phase MID-V35-HOTPATH-OPT-1: geometry SSOT
 
 // SmallPageMeta is now defined in smallobject_segment_mid_v3_box.h
 
@@ -21,20 +22,9 @@
 // ============================================================================
 // Helper: class_idx to slots mapping
 // ============================================================================
-
-static uint32_t class_idx_to_slots(uint32_t class_idx) {
-    // MID v3.5 targets C5-C7 (257-1024B)
-    // C5: 257-384B  → 64KiB / 384B  ≈ 170 slots
-    // C6: 385-640B  → 64KiB / 640B  ≈ 102 slots
-    // C7: 641-1024B → 64KiB / 1024B ≈ 64 slots
-
-    switch (class_idx) {
-        case 5: return 170;  // C5
-        case 6: return 102;  // C6
-        case 7: return 64;   // C7
-        default: return 0;
-    }
-}
+// Phase MID-V35-HOTPATH-OPT-1: Use geom_box as Single Source of Truth
+// See: core/box/smallobject_mid_v35_geom_box.h
+// (Removed local class_idx_to_slots() which had wrong C6 value: 102 instead of 128)
 
 // ============================================================================
 // Segment Lifecycle

docs/analysis/ENV_PROFILE_PRESETS.md

@@ -48,6 +48,19 @@ HAKMEM_TINY_HEAP_STATS=1
 HAKMEM_TINY_HEAP_STATS_DUMP=1
 HAKMEM_SMALL_HEAP_V3_STATS=1
 ```
+- **Phase MID-V35-HOTPATH-OPT-1** (FROZEN - research only):
+```sh
+HAKMEM_MID_V35_HEADER_PREFILL=1     # refill境界でheader先行書き
+HAKMEM_MID_V35_HOT_COUNTS=0         # alloc_count削除
+HAKMEM_MID_V35_C6_FASTPATH=1        # C6特化 fast path
+```
+  - **Status**: Default OFF, FROZEN (all 3 knobs)
+  - **Actual Results** (Phase MID-V35-HOTPATH-OPT-1 Mixed A/B):
+    - Mixed (16–1024B, MID_V35_OFF): **-0.2%** (誤差範囲, ±2%以内) ✓
+    - C6-heavy (257–768B, MID_V35_ON): **+7.3%** improvement ✅
+  - **Finding**: Mixed は MID_V3(C6-only) 固定で効果微小；C6-heavy のみ効果大
+  - **Recommendation**: C6_HEAVY_LEGACY_POOLV1 プリセットで推奨ON
+  - **NOT recommended for**: MIXED_TINYV3_C7_SAFE mainline (keep all defaults OFF)
 - **Phase POLICY-FAST-PATH-V2** (FROZEN - research only):
 ```sh
 HAKMEM_TINY_FREE_POLICY_FAST_V2=1  # Fast-path free optimization
@@ -100,6 +113,12 @@ HAKMEM_MID_V3_ENABLED=1             # Phase MID-V3: 257-768B, C6 only
 HAKMEM_MID_V3_CLASSES=0x40          # C6 only (+11% on C6-heavy)
 HAKMEM_MID_V35_ENABLED=1            # Phase v11a-5: C6-heavy で +8% 改善
 HAKMEM_MID_V35_CLASSES=0x40         # C6 only (53.1M ops/s)
+
+# Phase MID-V35-HOTPATH-OPT-1: C6-heavy 最速セット（推奨ON）
+# 機能: header prefill + hot counts削除 + C6 fast path (組み合わせで +7.3%)
+HAKMEM_MID_V35_HEADER_PREFILL=1     # refill境界でheader先行書き
+HAKMEM_MID_V35_HOT_COUNTS=0         # alloc_count削除（free_count/retire残す）
+HAKMEM_MID_V35_C6_FASTPATH=1        # C6特化 fast path (constant slot_size=512)
 ```
 - mid_desc_lookup TLS キャッシュを試すときだけ: `HAKMEM_MID_DESC_CACHE_ENABLED=1` を上乗せ（デフォルトは OFF）。