hakmem/core/tiny_alloc_fast.inc.h

// tiny_alloc_fast.inc.h - Box 5: Allocation Fast Path (3-4 instructions)
// Purpose: Ultra-fast TLS freelist pop (inspired by System tcache & Mid-Large HAKX +171%)
// Invariant: Hit rate > 95% → 3-4 instructions, Miss → refill from backend
// Design: "Simple Front + Smart Back" - Front is dumb & fast, Back is smart
//
// Box 5-NEW: SFC (Super Front Cache) Integration
// Architecture: SFC (Layer 0, 128-256 slots) → SLL (Layer 1, unlimited) → SuperSlab (Layer 2+)
// Cascade Refill: SFC ← SLL (one-way, safe)
// Goal: +200% performance (4.19M → 12M+ ops/s)
//
// Phase 2b: Adaptive TLS Cache Sizing
// Hot classes grow to 2048 slots, cold classes shrink to 16 slots
// Expected: +3-10% performance, -30-50% TLS cache memory overhead
#pragma once
#include "tiny_atomic.h"
#include "hakmem_tiny.h"
#include "tiny_route.h"
#include "tiny_alloc_fast_sfc.inc.h"  // Box 5-NEW: SFC Layer
#include "tiny_region_id.h"            // Phase 7: Header-based class_idx lookup
#include "tiny_adaptive_sizing.h"      // Phase 2b: Adaptive sizing
#include "box/tls_sll_box.h"           // Box TLS-SLL: C7-safe push/pop/splice
#ifdef HAKMEM_TINY_FRONT_GATE_BOX
#include "box/front_gate_box.h"
#endif
#include <stdio.h>

// Phase 7 Task 2: Aggressive inline TLS cache access
// Enable with: make HEADER_CLASSIDX=1 AGGRESSIVE_INLINE=1
#ifndef HAKMEM_TINY_AGGRESSIVE_INLINE
#define HAKMEM_TINY_AGGRESSIVE_INLINE 0
#endif

#if HAKMEM_TINY_AGGRESSIVE_INLINE
#include "tiny_alloc_fast_inline.h"
#endif

// ========== Debug Counters (compile-time gated) ==========
#if HAKMEM_DEBUG_COUNTERS
// Refill-stage counters (defined in hakmem_tiny.c)
extern unsigned long long g_rf_total_calls[];
extern unsigned long long g_rf_hit_bench[];
extern unsigned long long g_rf_hit_hot[];
extern unsigned long long g_rf_hit_mail[];
extern unsigned long long g_rf_hit_slab[];
extern unsigned long long g_rf_hit_ss[];
extern unsigned long long g_rf_hit_reg[];
extern unsigned long long g_rf_mmap_calls[];

// Publish hits (defined in hakmem_tiny.c)
extern unsigned long long g_pub_mail_hits[];
extern unsigned long long g_pub_bench_hits[];
extern unsigned long long g_pub_hot_hits[];

// Free pipeline (defined in hakmem_tiny.c)
extern unsigned long long g_free_via_tls_sll[];
#endif

// ========== Box 5: Allocation Fast Path ==========
// 箱理論の Fast Allocation 層。TLS freelist から直接 pop（3-4命令）。
// 不変条件:
// - TLS freelist が非空なら即座に return (no lock, no sync)
// - Miss なら Backend (Box 3: SuperSlab) に委譲
// - Cross-thread allocation は考慮しない（Backend が処理）

// External TLS variables (defined in hakmem_tiny.c)
extern __thread void* g_tls_sll_head[TINY_NUM_CLASSES];
extern __thread uint32_t g_tls_sll_count[TINY_NUM_CLASSES];

// External backend functions
// P0 Fix: Use appropriate refill function based on P0 status
#if HAKMEM_TINY_P0_BATCH_REFILL
extern int sll_refill_batch_from_ss(int class_idx, int max_take);
#else
extern int sll_refill_small_from_ss(int class_idx, int max_take);
#endif
extern void* hak_tiny_alloc_slow(size_t size, int class_idx);
extern int hak_tiny_size_to_class(size_t size);
extern int tiny_refill_failfast_level(void);
extern const size_t g_tiny_class_sizes[];

// Global Front refill config (parsed at init; defined in hakmem_tiny.c)
extern int g_refill_count_global;
extern int g_refill_count_hot;
extern int g_refill_count_mid;
extern int g_refill_count_class[TINY_NUM_CLASSES];

// HAK_RET_ALLOC macro is now defined in core/hakmem_tiny.c
// See lines 116-152 for single definition point based on HAKMEM_TINY_HEADER_CLASSIDX

// ========== RDTSC Profiling (lightweight) ==========
#ifdef __x86_64__
static inline uint64_t tiny_fast_rdtsc(void) {
    unsigned int lo, hi;
    __asm__ __volatile__ ("rdtsc" : "=a" (lo), "=d" (hi));
    return ((uint64_t)hi << 32) | lo;
}
#else
static inline uint64_t tiny_fast_rdtsc(void) { return 0; }
#endif

// Per-thread profiling counters (enable with HAKMEM_TINY_PROFILE=1)
static __thread uint64_t g_tiny_alloc_hits = 0;
static __thread uint64_t g_tiny_alloc_cycles = 0;
static __thread uint64_t g_tiny_refill_calls = 0;
static __thread uint64_t g_tiny_refill_cycles = 0;
static int g_tiny_profile_enabled = -1;  // -1: uninitialized

static inline int tiny_profile_enabled(void) {
    if (__builtin_expect(g_tiny_profile_enabled == -1, 0)) {
        const char* env = getenv("HAKMEM_TINY_PROFILE");
        g_tiny_profile_enabled = (env && *env && *env != '0') ? 1 : 0;
    }
    return g_tiny_profile_enabled;
}

// Print profiling results at exit
static void tiny_fast_print_profile(void) __attribute__((destructor));
static void tiny_fast_print_profile(void) {
    if (!tiny_profile_enabled()) return;
    if (g_tiny_alloc_hits == 0 && g_tiny_refill_calls == 0) return;

    fprintf(stderr, "\n========== Box Theory Fast Path Profile ==========\n");
    if (g_tiny_alloc_hits > 0) {
        fprintf(stderr, "[ALLOC HIT] count=%lu, avg_cycles=%lu\n",
                (unsigned long)g_tiny_alloc_hits,
                (unsigned long)(g_tiny_alloc_cycles / g_tiny_alloc_hits));
    }
    if (g_tiny_refill_calls > 0) {
        fprintf(stderr, "[REFILL]    count=%lu, avg_cycles=%lu\n",
                (unsigned long)g_tiny_refill_calls,
                (unsigned long)(g_tiny_refill_cycles / g_tiny_refill_calls));
    }
    fprintf(stderr, "===================================================\n\n");
}

// ========== Fast Path: TLS Freelist Pop (3-4 instructions) ==========

// External SFC control (defined in hakmem_tiny_sfc.c)
extern int g_sfc_enabled;

// Allocation fast path (inline for zero-cost)
// Returns: pointer on success, NULL on miss (caller should try refill/slow)
//
// Box 5-NEW Architecture:
//   Layer 0: SFC (128-256 slots, high hit rate) [if enabled]
//   Layer 1: SLL (unlimited, existing)
//   Cascade: SFC miss → try SLL → refill
//
// Assembly (x86-64, optimized):
//   mov    rax, QWORD PTR g_sfc_head[class_idx]      ; SFC: Load head
//   test   rax, rax                                   ; Check NULL
//   jne    .sfc_hit                                   ; If not empty, SFC hit!
//   mov    rax, QWORD PTR g_tls_sll_head[class_idx]  ; SLL: Load head
//   test   rax, rax                                   ; Check NULL
//   je     .miss                                      ; If empty, miss
//   mov    rdx, QWORD PTR [rax]                       ; Load next
//   mov    QWORD PTR g_tls_sll_head[class_idx], rdx  ; Update head
//   ret                                               ; Return ptr
// .sfc_hit:
//   mov    rdx, QWORD PTR [rax]                       ; Load next
//   mov    QWORD PTR g_sfc_head[class_idx], rdx      ; Update head
//   ret
// .miss:
//   ; Fall through to refill
//
// Expected: 3-4 instructions on SFC hit, 6-8 on SLL hit
static inline void* tiny_alloc_fast_pop(int class_idx) {
    // CRITICAL: C7 (1KB) is headerless - delegate to slow path completely
    // Reason: Fast path uses SLL which stores next pointer in user data area
    //         C7's headerless design is incompatible with fast path assumptions
    // Solution: Force C7 to use slow path for both alloc and free
    if (__builtin_expect(class_idx == 7, 0)) {
        return NULL;  // Force slow path
    }

#ifdef HAKMEM_TINY_FRONT_GATE_BOX
    void* out = NULL;
    if (front_gate_try_pop(class_idx, &out)) {
        return out;
    }
    return NULL;
#else
    // Phase 7 Task 3: Profiling overhead removed in release builds
    // In release mode, compiler can completely eliminate profiling code
#if !HAKMEM_BUILD_RELEASE
    uint64_t start = tiny_profile_enabled() ? tiny_fast_rdtsc() : 0;
#endif

    // Box 5-NEW: Layer 0 - Try SFC first (if enabled)
    // Cache g_sfc_enabled in TLS to avoid global load on every allocation
    static __thread int sfc_check_done = 0;
    static __thread int sfc_is_enabled = 0;
    if (__builtin_expect(!sfc_check_done, 0)) {
        sfc_is_enabled = g_sfc_enabled;
        sfc_check_done = 1;
    }

    if (__builtin_expect(sfc_is_enabled, 1)) {
        void* ptr = sfc_alloc(class_idx);
        if (__builtin_expect(ptr != NULL, 1)) {
            // Front Gate: SFC hit
            extern unsigned long long g_front_sfc_hit[];
            g_front_sfc_hit[class_idx]++;
            // 🚀 SFC HIT! (Layer 0)
#if !HAKMEM_BUILD_RELEASE
            if (start) {
                g_tiny_alloc_cycles += (tiny_fast_rdtsc() - start);
                g_tiny_alloc_hits++;
            }
#endif
            return ptr;
        }
        // SFC miss → try SLL (Layer 1)
    }

    // Box Boundary: Layer 1 - TLS SLL freelist の先頭を pop（envで無効化可）
    extern int g_tls_sll_enable;  // set at init via HAKMEM_TINY_TLS_SLL
    if (__builtin_expect(g_tls_sll_enable, 1)) {
        // Use Box TLS-SLL API (C7-safe pop)
        // CRITICAL: Pop FIRST, do NOT read g_tls_sll_head directly (race condition!)
        // Reading head before pop causes stale read → rbp=0xa0 SEGV
        void* head = NULL;
        if (tls_sll_pop(class_idx, &head)) {
            // Front Gate: SLL hit (fast path 3 instructions)
            extern unsigned long long g_front_sll_hit[];
            g_front_sll_hit[class_idx]++;

#if HAKMEM_DEBUG_COUNTERS
            // Track TLS freelist hits (compile-time gated, zero runtime cost when disabled)
            g_free_via_tls_sll[class_idx]++;
#endif

#if !HAKMEM_BUILD_RELEASE
            // Debug: Track profiling (release builds skip this overhead)
            if (start) {
                g_tiny_alloc_cycles += (tiny_fast_rdtsc() - start);
                g_tiny_alloc_hits++;
            }
#endif
            return head;
        }
    }

    // Fast path miss → NULL (caller should refill)
    return NULL;
#endif
}

// ========== Cascade Refill: SFC ← SLL (Box Theory boundary) ==========

// Cascade refill: Transfer blocks from SLL to SFC (one-way, safe)
// Returns: number of blocks transferred
//
// Contract:
// - Transfer ownership: SLL → SFC
// - No circular dependency: one-way only
// - Boundary clear: SLL pop → SFC push
// - Fallback safe: if SFC full, stop (no overflow)
static inline int sfc_refill_from_sll(int class_idx, int target_count) {
    int transferred = 0;
    uint32_t cap = g_sfc_capacity[class_idx];

    while (transferred < target_count && g_tls_sll_count[class_idx] > 0) {
        // Check SFC capacity before transfer
        if (g_sfc_count[class_idx] >= cap) {
            break;  // SFC full, stop
        }

        // Pop from SLL (Layer 1) using Box TLS-SLL API (C7-safe)
        void* ptr = NULL;
        if (!tls_sll_pop(class_idx, &ptr)) {
            break;  // SLL empty
        }

        // Push to SFC (Layer 0) — header-aware
#if HAKMEM_TINY_HEADER_CLASSIDX
        const size_t sfc_next_off = (class_idx == 7) ? 0 : 1;
#else
        const size_t sfc_next_off = 0;
#endif
        *(void**)((uint8_t*)ptr + sfc_next_off) = g_sfc_head[class_idx];
        g_sfc_head[class_idx] = ptr;
        g_sfc_count[class_idx]++;

        transferred++;
    }

    return transferred;
}

// ========== Refill Path: Backend Integration ==========

// Refill TLS freelist from backend (SuperSlab/ACE/Learning layer)
// Returns: number of blocks refilled
//
// Box 5-NEW Architecture:
//   SFC enabled:  SuperSlab → SLL → SFC (cascade)
//   SFC disabled: SuperSlab → SLL (direct, old path)
//
// This integrates with existing HAKMEM infrastructure:
// - SuperSlab provides memory chunks
// - ACE provides adaptive capacity learning
// - L25 provides mid-large integration
//
// Refill count is tunable via HAKMEM_TINY_REFILL_COUNT (default: 16)
// - Smaller count (8-16): better for diverse workloads, faster warmup
// - Larger count (64-128): better for homogeneous workloads, fewer refills
static inline int tiny_alloc_fast_refill(int class_idx) {
    // CRITICAL: C7 (1KB) is headerless - skip refill completely, force slow path
    // Reason: Refill pushes blocks to TLS SLL which stores next pointer in user data
    //         C7's headerless design is incompatible with this mechanism
    if (__builtin_expect(class_idx == 7, 0)) {
        return 0;  // Skip refill, force slow path allocation
    }

    // Phase 7 Task 3: Profiling overhead removed in release builds
    // In release mode, compiler can completely eliminate profiling code
#if !HAKMEM_BUILD_RELEASE
    uint64_t start = tiny_profile_enabled() ? tiny_fast_rdtsc() : 0;
#endif

    // Phase 2b: Check available capacity before refill
    int available_capacity = get_available_capacity(class_idx);
    if (available_capacity <= 0) {
        // Cache is full, don't refill
        return 0;
    }

    // Phase 7 Task 3: Simplified refill count (cached per-class in TLS)
    // Previous: Complex precedence logic on every miss (5-10 cycles overhead)
    // Now: Simple TLS cache lookup (1-2 cycles)
    static __thread int s_refill_count[TINY_NUM_CLASSES] = {0};
    // Simple adaptive booster: bump per-class refill size when refills are frequent.
    static __thread uint8_t s_refill_calls[TINY_NUM_CLASSES] = {0};
    int cnt = s_refill_count[class_idx];
    if (__builtin_expect(cnt == 0, 0)) {
        // First miss: Initialize from globals (parsed at init time)
        int v = HAKMEM_TINY_REFILL_DEFAULT;  // Default from hakmem_build_flags.h

        // Precedence: per-class > hot/mid > global
        if (g_refill_count_class[class_idx] > 0) {
            v = g_refill_count_class[class_idx];
        } else if (class_idx <= 3 && g_refill_count_hot > 0) {
            v = g_refill_count_hot;
        } else if (class_idx >= 4 && g_refill_count_mid > 0) {
            v = g_refill_count_mid;
        } else if (g_refill_count_global > 0) {
            v = g_refill_count_global;
        }

        // Clamp to sane range (min: 8, max: 256)
        if (v < 8) v = 8;        // Minimum: avoid thrashing
        if (v > 256) v = 256;    // Maximum: avoid excessive TLS memory

        s_refill_count[class_idx] = v;
        cnt = v;
    }

    // Phase 2b: Clamp refill count to available capacity
    if (cnt > available_capacity) {
        cnt = available_capacity;
    }

#if HAKMEM_DEBUG_COUNTERS
    // Track refill calls (compile-time gated)
    g_rf_total_calls[class_idx]++;
#endif

    // Box Boundary: Delegate to Backend (Box 3: SuperSlab)
    // This gives us ACE, Learning layer, L25 integration for free!
    // P0 Fix: Use appropriate refill function based on P0 status
#if HAKMEM_TINY_P0_BATCH_REFILL
    int refilled = sll_refill_batch_from_ss(class_idx, cnt);
#else
    int refilled = sll_refill_small_from_ss(class_idx, cnt);
#endif

    // Lightweight adaptation: if refills keep happening, increase per-class refill.
    // Focus on class 7 (1024B) to reduce mmap/refill frequency under Tiny-heavy loads.
    if (refilled > 0) {
        uint8_t c = ++s_refill_calls[class_idx];
        if (class_idx == 7) {
            // Every 4 refills, increase target by +16 up to 128 (unless overridden).
            if ((c & 0x03u) == 0) {
                int target = s_refill_count[class_idx];
                if (target < 128) {
                    target += 16;
                    if (target > 128) target = 128;
                    s_refill_count[class_idx] = target;
                }
            }
        }
    } else {
        // No refill performed (capacity full): slowly decay the counter.
        if (s_refill_calls[class_idx] > 0) s_refill_calls[class_idx]--;
    }

    // Phase 2b: Track refill and adapt cache size
    if (refilled > 0) {
        track_refill_for_adaptation(class_idx);
    }

    // Box 5-NEW: Cascade refill SFC ← SLL (if SFC enabled)
    // This happens AFTER SuperSlab → SLL refill, so SLL has blocks
    static __thread int sfc_check_done_refill = 0;
    static __thread int sfc_is_enabled_refill = 0;
    if (__builtin_expect(!sfc_check_done_refill, 0)) {
        sfc_is_enabled_refill = g_sfc_enabled;
        sfc_check_done_refill = 1;
    }

    if (sfc_is_enabled_refill && refilled > 0) {
        // Transfer half of refilled blocks to SFC (keep half in SLL for future)
        int sfc_target = refilled / 2;
        if (sfc_target > 0) {
#ifdef HAKMEM_TINY_FRONT_GATE_BOX
            front_gate_after_refill(class_idx, refilled);
#else
            int transferred = sfc_refill_from_sll(class_idx, sfc_target);
            (void)transferred;  // Unused, but could track stats
#endif
        }
    }

#if !HAKMEM_BUILD_RELEASE
    // Debug: Track profiling (release builds skip this overhead)
    if (start) {
        g_tiny_refill_cycles += (tiny_fast_rdtsc() - start);
        g_tiny_refill_calls++;
    }
#endif

    return refilled;
}

// ========== Combined Fast Path (Alloc + Refill) ==========

// Complete fast path allocation (inline for zero-cost)
// Returns: pointer on success, NULL on failure (OOM or size too large)
//
// Flow:
// 1. TLS freelist pop (3-4 instructions) - Hit rate ~95%
// 2. Miss → Refill from backend (~5% cases)
// 3. Refill success → Retry pop
// 4. Refill failure → Slow path (OOM or new SuperSlab allocation)
//
// Example usage:
//   void* ptr = tiny_alloc_fast(64);
//   if (!ptr) {
//       // OOM handling
//   }
static inline void* tiny_alloc_fast(size_t size) {
    // 1. Size → class index (inline, fast)
    int class_idx = hak_tiny_size_to_class(size);
    if (__builtin_expect(class_idx < 0, 0)) {
        return NULL;  // Size > 1KB, not Tiny
    }
    ROUTE_BEGIN(class_idx);

    // 2. Fast path: TLS freelist pop (3-4 instructions, 95% hit rate)
    // CRITICAL: Use Box TLS-SLL API (static inline, same performance as macro but SAFE!)
    // The old macro had race condition: read head before pop → rbp=0xa0 SEGV
    void* ptr = NULL;
    tls_sll_pop(class_idx, &ptr);
    if (__builtin_expect(ptr != NULL, 1)) {
        // C7 (1024B, headerless): clear embedded next pointer before returning to user
        if (__builtin_expect(class_idx == 7, 0)) {
            *(void**)ptr = NULL;
        }
        HAK_RET_ALLOC(class_idx, ptr);
    }

    // 3. Miss: Refill from backend (Box 3: SuperSlab)
    int refilled = tiny_alloc_fast_refill(class_idx);
    if (__builtin_expect(refilled > 0, 1)) {
        // Refill success → retry pop using safe Box TLS-SLL API
        ptr = NULL;
        tls_sll_pop(class_idx, &ptr);
        if (ptr) {
            if (__builtin_expect(class_idx == 7, 0)) {
                *(void**)ptr = NULL;
            }
            HAK_RET_ALLOC(class_idx, ptr);
        }
    }

    // 4. Refill failure or still empty → slow path (OOM or new SuperSlab)
    // Box Boundary: Delegate to Slow Path (Box 3 backend)
    ptr = hak_tiny_alloc_slow(size, class_idx);
    if (ptr) {
        HAK_RET_ALLOC(class_idx, ptr);
    }

    return ptr;  // NULL if OOM
}

// ========== Push to TLS Freelist (for free path) ==========

// Push block to TLS freelist (used by free fast path)
// This is a "helper" for Box 6 (Free Fast Path)
//
// Invariant: ptr must belong to current thread (no ownership check here)
// Caller (Box 6) is responsible for ownership verification
static inline void tiny_alloc_fast_push(int class_idx, void* ptr) {
#ifdef HAKMEM_TINY_FRONT_GATE_BOX
    front_gate_push_tls(class_idx, ptr);
#else
    // Box Boundary: Push to TLS freelist using Box TLS-SLL API (C7-safe)
    uint32_t capacity = UINT32_MAX;  // Unlimited for helper function
    if (!tls_sll_push(class_idx, ptr, capacity)) {
        // C7 rejected or SLL somehow full (should not happen)
        // In release builds, this is a no-op (caller expects success)
#if !HAKMEM_BUILD_RELEASE
        fprintf(stderr, "[WARN] tls_sll_push failed in tiny_alloc_fast_push cls=%d ptr=%p\n",
                class_idx, ptr);
#endif
    }
#endif
}

// ========== Statistics & Diagnostics ==========

// Get TLS freelist stats (for debugging/profiling)
typedef struct {
    int class_idx;
    void* head;
    uint32_t count;
} TinyAllocFastStats;

static inline TinyAllocFastStats tiny_alloc_fast_stats(int class_idx) {
    TinyAllocFastStats stats = {
        .class_idx = class_idx,
        .head = g_tls_sll_head[class_idx],
        .count = g_tls_sll_count[class_idx]
    };
    return stats;
}

// Reset TLS freelist (for testing/benchmarking)
// WARNING: This leaks memory! Only use in controlled test environments.
static inline void tiny_alloc_fast_reset(int class_idx) {
    g_tls_sll_head[class_idx] = NULL;
    g_tls_sll_count[class_idx] = 0;
}

// ========== Performance Notes ==========
//
// Expected metrics (based on System tcache & HAKX +171% results):
// - Fast path hit rate: 95%+ (workload dependent)
// - Fast path latency: 3-4 instructions (1-2 cycles on modern CPUs)
// - Miss penalty: ~20-50 instructions (refill from SuperSlab)
// - Throughput improvement: +10-25% vs current multi-layer design
//
// Key optimizations:
// 1. `__builtin_expect` for branch prediction (hot path first)
// 2. `static inline` for zero-cost abstraction
// 3. TLS variables (no atomic ops, no locks)
// 4. Minimal work in fast path (defer stats/accounting to backend)
//
// Comparison with current design:
// - Current: 20+ instructions (Magazine → SuperSlab → ACE → ...)
// - New: 3-4 instructions (TLS freelist pop only)
// - Reduction: -80% instructions in hot path
//
// Inspired by:
// - System tcache (glibc malloc) - 3-4 instruction fast path
// - HAKX Mid-Large (+171%) - "Simple Front + Smart Back"
// - Box Theory - Clear boundaries, minimal coupling