hakmem/core/tiny_alloc_fast_inline.h

// tiny_alloc_fast_inline.h - Phase 7 Task 2: Aggressive inline TLS cache access
// Purpose: Eliminate function call overhead (5-10 cycles) in hot path
// Design: Macro-based inline expansion of TLS freelist operations
// Performance: Expected +10-15% (22M → 24-25M ops/s)

#ifndef TINY_ALLOC_FAST_INLINE_H
#define TINY_ALLOC_FAST_INLINE_H

#include <stddef.h>
#include <stdint.h>
#include "hakmem_build_flags.h"
#include "tiny_remote.h"  // for TINY_REMOTE_SENTINEL (defense-in-depth)
#include "box/tiny_next_ptr_box.h"  // Phase E1-CORRECT: unified next pointer API
#include "tiny_region_id.h"  // For HEADER_MAGIC, HEADER_CLASS_MASK (Fix #7)

// 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];

#ifndef TINY_NUM_CLASSES
#define TINY_NUM_CLASSES 8
#endif

// ========== Inline Macro: TLS Freelist Pop ==========
//
// Aggressive inline expansion of tiny_alloc_fast_pop()
// Saves: 5-10 cycles (function call overhead + register spilling)
//
// Assembly comparison (x86-64):
//   Function call:
//     push   %rbx                  ; Save registers
//     mov    %edi, %ebx             ; class_idx to %ebx
//     call   tiny_alloc_fast_pop    ; Call (5-10 cycles overhead)
//     pop    %rbx                   ; Restore registers
//     test   %rax, %rax             ; Check result
//
//   Inline macro:
//     mov    g_tls_sll_head(%rdi), %rax  ; Direct access (3-4 cycles)
//     test   %rax, %rax
//     je     .miss
//     mov    (%rax), %rdx
//     mov    %rdx, g_tls_sll_head(%rdi)
//
// Result: 5-10 fewer instructions, better register allocation
//
#define TINY_ALLOC_FAST_POP_INLINE(class_idx, ptr_out) do { \
    void* _head = g_tls_sll_head[(class_idx)]; \
    if (__builtin_expect(_head != NULL, 1)) { \
        if (__builtin_expect((uintptr_t)_head == TINY_REMOTE_SENTINEL, 0)) { \
            /* Break the chain defensively if sentinel leaked into TLS SLL */ \
            g_tls_sll_head[(class_idx)] = NULL; \
            if (g_tls_sll_count[(class_idx)] > 0) g_tls_sll_count[(class_idx)]--; \
            (ptr_out) = NULL; \
        } else { \
            /* Phase E1-CORRECT: Use Box API for next pointer read */ \
            void* _next = tiny_next_read(class_idx, _head); \
            g_tls_sll_head[(class_idx)] = _next; \
            if (g_tls_sll_count[(class_idx)] > 0) { \
                g_tls_sll_count[(class_idx)]--; \
            } \
            /* Phase E1-CORRECT: All classes return user pointer (base+1) */ \
            (ptr_out) = (void*)((uint8_t*)_head + 1); \
        } \
    } else { \
        (ptr_out) = NULL; \
    } \
} while(0)

// ========== Inline Macro: TLS Freelist Push ==========
//
// Aggressive inline expansion of tiny_alloc_fast_push()
// Saves: 5-10 cycles (function call overhead)
//
// Assembly comparison:
//   Function call:
//     mov    %rdi, %rsi             ; ptr to %rsi
//     mov    %ebx, %edi             ; class_idx to %edi
//     call   tiny_alloc_fast_push   ; Call (5-10 cycles)
//
//   Inline macro:
//     mov    g_tls_sll_head(%rdi), %rax  ; Direct inline (2-3 cycles)
//     mov    %rax, (%rsi)
//     mov    %rsi, g_tls_sll_head(%rdi)
//
#if HAKMEM_TINY_HEADER_CLASSIDX
// Phase E1-CORRECT: Restore header on FREE for ALL classes (including C7)
// ROOT CAUSE: User may have overwritten byte 0 (header). tls_sll_splice() checks
// byte 0 for HEADER_MAGIC. Without restoration, it finds 0x00 → uses wrong offset → SEGV.
// COST: 1 byte write (~1-2 cycles per free, negligible).
#define TINY_ALLOC_FAST_PUSH_INLINE(class_idx, ptr) do { \
    *(uint8_t*)(ptr) = HEADER_MAGIC | ((class_idx) & HEADER_CLASS_MASK); \
    tiny_next_write(class_idx, (ptr), g_tls_sll_head[(class_idx)]); \
    g_tls_sll_head[(class_idx)] = (ptr); \
    g_tls_sll_count[(class_idx)]++; \
} while(0)
#else
#define TINY_ALLOC_FAST_PUSH_INLINE(class_idx, ptr) do { \
    tiny_next_write(class_idx, (ptr), g_tls_sll_head[(class_idx)]); \
    g_tls_sll_head[(class_idx)] = (ptr); \
    g_tls_sll_count[(class_idx)]++; \
} while(0)
#endif

// ========== Performance Notes ==========
//
// Benchmark results (expected):
// - Random Mixed 128B: 21M → 23M ops/s (+10%)
// - Random Mixed 256B: 19M → 22M ops/s (+15%)
// - Larson 1T: 2.7M → 3.0M ops/s (+11%)
//
// Key optimizations:
// 1. No function call overhead (save 5-10 cycles)
// 2. Better register allocation (inline knows full context)
// 3. No stack frame setup/teardown
// 4. Compiler can optimize across macro boundaries
//
// Trade-offs:
// 1. Code size: +100-200 bytes (each call site expanded)
// 2. Debug visibility: Macros harder to step through
// 3. Maintenance: Changes must be kept in sync with function version
//
// Recommendation: Use inline macros for CRITICAL hot paths only
// (alloc/free fast path), keep functions for diagnostics/debugging

#endif // TINY_ALLOC_FAST_INLINE_H