// hakmem_tiny_alloc_v3.inc
// Phase 3: Allocation Hot Path Simplification (mimalloc-style)
//
// Goal: Reduce malloc/alloc overhead from 27% to 20%
// Expected improvement: +5-10% (16.53 → 17.5-18.0 M ops/sec)
//
// Key changes:
// - Single-tier magazine hot path (2-3 cycles)
// - Eliminate 6+ stage fallback chain → 2 stages
// - Minimal stack frame
// - Zero TLS overhead on hit

// ============================================================================
#include "mid_tcache.h"
// Phase 3 Helper: Magazine Refill from SuperSlab
// ============================================================================

// Refill magazine from SuperSlab (batch allocation)
// Returns: number of items added to magazine
static int mag_refill_from_superslab_v3(int class_idx, TinyTLSMag* mag, int want) {
    if (!mag || want <= 0) return 0;

    int room = mag->cap - mag->top;
    if (room <= 0) return 0;
    if (want > room) want = room;

    int filled = 0;
    for (int i = 0; i < want; i++) {
        void* ptr = hak_tiny_alloc_superslab(class_idx);
        if (!ptr) break;

        mag->items[mag->top].ptr = ptr;
#if HAKMEM_TINY_MAG_OWNER
        // Owner tracking not critical for SuperSlab allocations
        mag->items[mag->top].owner = NULL;
#endif
        mag->top++;
        filled++;
    }

    return filled;
}

// ============================================================================
// Phase 3: Slow Path (Cold, Noinline)
// ============================================================================

static void* __attribute__((cold, noinline))
hak_tiny_alloc_slow_v3(size_t size, int class_idx) {
    (void)size; // size already validated by caller

    if (class_idx < 0 || class_idx >= TINY_NUM_CLASSES) {
        return NULL;
    }

    TinyTLSMag* mag = &g_tls_mags[class_idx];

    // Step 1: Try refilling magazine from SuperSlab (batch 32-64 items)
    int refill_count = (class_idx <= 3) ? 64 : 32;
    if (mag_refill_from_superslab_v3(class_idx, mag, refill_count) > 0) {
        // Magazine refilled successfully - pop one item
        void* ptr = mag->items[--mag->top].ptr;
        return ptr;
    }

    // Step 2: Direct SuperSlab allocation (magazine full or refill failed)
    void* ptr = hak_tiny_alloc_superslab(class_idx);
    return ptr;
}

// ============================================================================
// Phase 3: Hot Path (Ultra-Simple, Inline Candidate)
// ============================================================================

void* hak_tiny_alloc_v3(size_t size) {
    // Phase 3 hot path: minimal branches, single TLS variable

    // 1. Size → class (branchless LUT)
    int class_idx = hak_tiny_size_to_class(size);
    if (__builtin_expect(class_idx < 0, 0)) {
        return NULL;  // >1KB or size=0
    }

    // 2. MidTC (class>=4) — TLS tcache最優先
    if (__builtin_expect(class_idx > 3, 0)) {
        void* mp = midtc_pop(class_idx);
        if (mp) return mp;
    }

    // 3. Single-tier TLS magazine (HOT PATH - 2-3 cycles!)
    TinyTLSMag* mag = &g_tls_mags[class_idx];
    int top = mag->top;
    if (__builtin_expect(top > 0, 1)) {
        // Fast path: pop from magazine
        void* ptr = mag->items[--top].ptr;
        mag->top = top;
        return ptr;  // ← 最速パス！ 🚀
    }

    // 4. Slow path: refill + fallback
    return hak_tiny_alloc_slow_v3(size, class_idx);
}