hakmem/core/hakmem_tiny_alloc_new.inc

// hakmem_tiny_alloc_new.inc
// New 3-layer Tiny Pool allocation (simplified)
//
// Purpose: Reduce from 6-7 layers to 3 layers
// Target: 100+ instructions/op → 20-30 instructions/op
//
// Part of 3-layer architecture simplification (2025-11-01)
// Based on ChatGPT Pro UltraThink recommendations

// === IMPORTANT: Disable old benchmark fastpath ===
// The old HAKMEM_TINY_BENCH_FASTPATH conflicts with new 3-layer architecture
// We must disable it to ensure our new code runs
#ifdef HAKMEM_TINY_BENCH_FASTPATH
#undef HAKMEM_TINY_BENCH_FASTPATH
#endif

// Phase E1-CORRECT: Box API for next pointer operations
#include "box/tiny_next_ptr_box.h"

// Debug counters (thread-local)
static __thread uint64_t g_3layer_bump_hits = 0;
static __thread uint64_t g_3layer_mag_hits = 0;
static __thread uint64_t g_3layer_slow_hits = 0;
static __thread uint64_t g_3layer_refill_count = 0;
static __thread uint64_t g_3layer_refill_items = 0;
static __thread uint64_t g_3layer_fallback_superslab_disabled = 0;
static __thread uint64_t g_3layer_fallback_no_ss = 0;
static __thread uint64_t g_3layer_fallback_no_meta = 0;
static __thread uint64_t g_3layer_batch_carve_count = 0;

// Active accounting helper (env toggle: HAKMEM_TINY_ACTIVE_FIX=0 to disable)
static inline int tiny_active_fix_enabled(void) {
    static int g_active_fix_en = -1;
    if (__builtin_expect(g_active_fix_en == -1, 0)) {
        const char* e = getenv("HAKMEM_TINY_ACTIVE_FIX");
        g_active_fix_en = (e && atoi(e) == 0) ? 0 : 1;
    }
    return g_active_fix_en;
}

static inline void tiny_active_account_alloc(void* ptr) {
    if (!ptr || !g_use_superslab) return;
    if (!tiny_active_fix_enabled()) return;
    SuperSlab* ss = hak_super_lookup(ptr);
    if (ss && ss->magic == SUPERSLAB_MAGIC) {
        ss_active_inc(ss);
    }
}

// Forward declaration for Layer 3
__attribute__((noinline, cold))
static void* tiny_alloc_slow_new(int class_idx);

// ============================================================================
// Main Allocation Function (3-layer architecture)
// ============================================================================

void* hak_tiny_alloc(size_t size) {
    // Initialization check (cold path, once per thread)
#if !HAKMEM_BUILD_RELEASE
    if (!g_tiny_initialized) hak_tiny_init();
#else
    if (__builtin_expect(!g_tiny_initialized, 0)) {
        hak_tiny_init();
    }
#endif

    // Wrapper guard (safety check, rare)
#if !HAKMEM_BUILD_RELEASE
# if HAKMEM_WRAPPER_TLS_GUARD
    if (!g_wrap_tiny_enabled && __builtin_expect(g_tls_in_wrapper != 0, 0)) return NULL;
# else
    extern int hak_in_wrapper(void);
    if (!g_wrap_tiny_enabled && __builtin_expect(hak_in_wrapper() != 0, 0)) return NULL;
# endif
#endif

    // Size to class index
    int class_idx = hak_tiny_size_to_class(size);
    if (class_idx < 0) return NULL;  // > 1KB
    // Route fingerprint begin (debug-only; no-op unless HAKMEM_ROUTE=1)
    ROUTE_BEGIN(class_idx);

    // Initialize small magazine (once per thread)
    if (__builtin_expect(!g_tiny_small_mag_initialized, 0)) {
        tiny_small_mag_init();
    }

    // ========================================================================
    // === LAYER 1: TLS Bump Allocator (hot classes 0-2: 8B/16B/32B) ===
    // === Target: 2-3 instructions/op ===
    // ========================================================================
    if (likely(class_idx <= 2)) {
        void* p = tiny_bump_alloc(class_idx);
        if (likely(p)) {
            tiny_active_account_alloc(p);
            g_3layer_bump_hits++;
            // Mark: bump hit（便宜的にhot_hitのbitを再利用 8）
            ROUTE_MARK(8); ROUTE_COMMIT(class_idx, 0x40);
            HAK_RET_ALLOC(class_idx, p);
        }
    }

    // ========================================================================
    // === LAYER 2: TLS Small Magazine (all classes, 128 items) ===
    // === Target: 5-10 instructions/op ===
    // ========================================================================
    void* p = small_mag_pop(class_idx);
    if (likely(p)) {
        extern unsigned long long g_front_mag_hit[];
        g_front_mag_hit[class_idx]++;
        tiny_active_account_alloc(p);
        g_3layer_mag_hits++;
        // Mark: small mag hit（bench_hitのbitを便宜的に再利用 10）
        ROUTE_MARK(10); ROUTE_COMMIT(class_idx, 0x41);
        HAK_RET_ALLOC(class_idx, p);
    }

    // ========================================================================
    // === LAYER 3: Slow path (refill, slab allocation) ===
    // === Target: 50-100+ instructions/op (rare) ===
    // ========================================================================
    g_3layer_slow_hits++;
    return tiny_alloc_slow_new(class_idx);
}

// ============================================================================
// Layer 3: Slow Path (refill and slab management)
// ============================================================================

__attribute__((noinline, cold))
static void* tiny_alloc_slow_new(int class_idx) {
    // Return‑First Selector: try Ready/Mailbox/Sticky/Hot/Bench/Registry once
    do {
        static int g_return_first = -1; // env: HAKMEM_TINY_RETURN_FIRST (default ON)
        if (__builtin_expect(g_return_first == -1, 0)) {
            const char* e = getenv("HAKMEM_TINY_RETURN_FIRST");
            g_return_first = (e && *e == '0') ? 0 : 1;
        }
        if (__builtin_expect(g_return_first, 1)) {
            extern __thread TinyTLSSlab g_tls_slabs[];
            TinyTLSSlab* tls = &g_tls_slabs[class_idx];
            SuperSlab* rs = tiny_refill_try_fast(class_idx, tls);
            (void)rs; // On success, tls->ss is bound and Step 2 will carve
        }
    } while (0);

    // ========================================================================
    // Layer 3: Refill Small Magazine and/or Bump from existing infrastructure
    // ========================================================================

    // Step 1: Try to refill Small Magazine from existing TLS Magazine
    tiny_mag_init_if_needed(class_idx);
    TinyTLSMag* large_mag = &g_tls_mags[class_idx];

    if (large_mag->top > 0) {
        // Batch transfer from large magazine (2048) to small magazine
        int batch_size = 64;  // Transfer in batches of 64
        if (batch_size > large_mag->top) batch_size = large_mag->top;

        void* items[64];
        for (int i = 0; i < batch_size; i++) {
            items[i] = large_mag->items[large_mag->top - 1 - i].ptr;
        }
        large_mag->top -= batch_size;

        // Push to Small Magazine
        int pushed = small_mag_batch_push(class_idx, items, batch_size);
        g_3layer_refill_count++;
        g_3layer_refill_items += pushed;

        // Try to pop one and return
        void* p = small_mag_pop(class_idx);
        if (p) {
            tiny_active_account_alloc(p);
            return p;
        }
    }

    // Step 2: Large Magazine empty - batch carve from SuperSlab directly
    // ChatGPT Pro P0: Complete batch化 (based on tls_refill_from_tls_slab:115-126)
    if (!g_use_superslab) {
        g_3layer_fallback_superslab_disabled++;
        return hak_tiny_alloc_slow(0, class_idx);
    }

    TinyTLSSlab* tls_slab = &g_tls_slabs[class_idx];
    if (!tls_slab->ss) {
        if (superslab_refill(class_idx) == NULL) {
            g_3layer_fallback_no_ss++;
            // Optional one-shot debug
            static int g_alloc_dbg = -1; if (__builtin_expect(g_alloc_dbg == -1, 0)) { const char* e=getenv("HAKMEM_TINY_ALLOC_DEBUG"); g_alloc_dbg = (e && atoi(e)!=0)?1:0; }
            if (g_alloc_dbg) {
                static _Atomic int printed_ss[8]; int exp=0;
                if (atomic_compare_exchange_strong(&printed_ss[class_idx], &exp, 1)) {
                    fprintf(stderr, "[ALLOC3] refill returned NULL (no SS) class=%d\n", class_idx);
                }
            }
            return hak_tiny_alloc_slow(0, class_idx);  // Fallback
        }
    }

    TinySlabMeta* meta = tls_slab->meta;
    if (!meta) {
        g_3layer_fallback_no_meta++;
        // Optional one-shot debug
        static int g_alloc_dbg2 = -1; if (__builtin_expect(g_alloc_dbg2 == -1, 0)) { const char* e=getenv("HAKMEM_TINY_ALLOC_DEBUG"); g_alloc_dbg2 = (e && atoi(e)!=0)?1:0; }
        if (g_alloc_dbg2) {
            static _Atomic int printed_meta[8]; int exp=0;
            if (atomic_compare_exchange_strong(&printed_meta[class_idx], &exp, 1)) {
                fprintf(stderr, "[ALLOC3] meta is NULL after refill class=%d\n", class_idx);
            }
        }
        return hak_tiny_alloc_slow(0, class_idx);
    }

    // Batch carve from SuperSlab (P0 optimization - no 64x function calls!)
    uint32_t want = 64;  // Refill target
    void* items[64];
    int got = 0;

    // Try freelist first (small amount, usually 0)
    while (got < (int)want && meta->freelist) {
        void* node = meta->freelist;
        meta->freelist = tiny_next_read(node);  // Phase E1-CORRECT: Box API
        items[got++] = node;
        meta->used++;
    }

    // Then linear carve (KEY OPTIMIZATION - direct array fill!)
    if (got < (int)want && meta->used < meta->capacity) {
        uint32_t need = want - got;
        uint32_t available = meta->capacity - meta->used;
        if (need > available) need = available;

        size_t block_size = g_tiny_class_sizes[class_idx];
        uint8_t* slab_base = tls_slab->slab_base ? tls_slab->slab_base
                                                  : tiny_slab_base_for(tls_slab->ss, tls_slab->slab_idx);
        uint8_t* cursor = slab_base + ((size_t)meta->used * block_size);

        // Batch carve: directly fill items array (no linked list, no 64 function calls!)
        for (uint32_t i = 0; i < need; ++i) {
            items[got++] = (void*)cursor;
            cursor += block_size;
        }

        meta->used += need;  // Reserve to TLS; not active until returned to user
    }

    if (got == 0) {
        // Slab exhausted, try refill and retry once
        if (superslab_refill(class_idx) != NULL) {
            return tiny_alloc_slow_new(class_idx);  // Recursive retry
        }
        static int g_alloc_dbg3 = -1; if (__builtin_expect(g_alloc_dbg3 == -1, 0)) { const char* e=getenv("HAKMEM_TINY_ALLOC_DEBUG"); g_alloc_dbg3 = (e && atoi(e)!=0)?1:0; }
        if (g_alloc_dbg3) {
            static _Atomic int printed_final[8]; int exp=0;
            if (atomic_compare_exchange_strong(&printed_final[class_idx], &exp, 1)) {
                fprintf(stderr, "[ALLOC3] no items after retry (final fallback) class=%d\n", class_idx);
            }
        }
        return hak_tiny_alloc_slow(0, class_idx);  // Ultimate fallback
    }

    // Take one for return, push rest to Small Magazine
    g_3layer_batch_carve_count++;
    void* result = items[0];
    if (got > 1) {
        int pushed = small_mag_batch_push(class_idx, &items[1], got - 1);
        g_3layer_refill_count++;
        g_3layer_refill_items += pushed;
    }

    tiny_active_account_alloc(result);
    // Route: slab carve direct（linear相当の採用扱い）
    ROUTE_MARK(11); ROUTE_COMMIT(class_idx, 0x60);
    return result;
}

// Debug function: print layer statistics
__attribute__((destructor))
static void print_3layer_stats(void) {
    uint64_t total = g_3layer_bump_hits + g_3layer_mag_hits + g_3layer_slow_hits;
    if (total > 0) {
        fprintf(stderr, "\n=== 3-Layer Architecture Stats ===\n");
        fprintf(stderr, "Bump hits:     %10lu (%5.2f%%)\n",
                g_3layer_bump_hits, 100.0 * g_3layer_bump_hits / total);
        fprintf(stderr, "Mag hits:      %10lu (%5.2f%%)\n",
                g_3layer_mag_hits, 100.0 * g_3layer_mag_hits / total);
        fprintf(stderr, "Slow hits:     %10lu (%5.2f%%)\n",
                g_3layer_slow_hits, 100.0 * g_3layer_slow_hits / total);
        fprintf(stderr, "Total allocs:  %10lu\n", total);
        fprintf(stderr, "Refill count:  %10lu\n", g_3layer_refill_count);
        fprintf(stderr, "Refill items:  %10lu (avg %.1f/refill)\n",
                g_3layer_refill_items,
                g_3layer_refill_count > 0 ? (double)g_3layer_refill_items / g_3layer_refill_count : 0.0);
        fprintf(stderr, "=== Fallback Paths ===\n");
        fprintf(stderr, "SuperSlab disabled: %lu\n", g_3layer_fallback_superslab_disabled);
        fprintf(stderr, "No SuperSlab:       %lu\n", g_3layer_fallback_no_ss);
        fprintf(stderr, "No meta:            %lu\n", g_3layer_fallback_no_meta);
        fprintf(stderr, "Batch carve count:  %lu\n", g_3layer_batch_carve_count);
    }
}