l4 evict and swiss api improve

.editorconfig

@@ -0,0 +1,5 @@
+root = true
+
+[*.cs]
+dotnet_diagnostic.IDE0032.severity = none
+dotnet_diagnostic.IDE0055.severity = none

src/Nitload.Common/HighPerformance/Memory/L4.CacheLine.cs

@@ -6,6 +6,7 @@
 // -----------------------------------------------------------------------------
 
 using System.Runtime.InteropServices;
+using System.Xml.Serialization;
 
 namespace Nitload.Common.HighPerformance.Memory
 {
@@ -27,10 +28,38 @@ namespace Nitload.Common.HighPerformance.Memory
 
             internal const uint MaxBufferSize = 0xFFFFFFFF - 64U;
 
-            [FieldOffset(16)] internal AllocationState State;
+            [FieldOffset(31)] internal AllocationState State;
 
             public static uint SizeOf { get; } = (uint)Unsafe.SizeOf<CacheLineMetadata>();
 
+            /// <summary>
+            ///   only available in L4.Predict()
+            /// </summary>
+            public void CheckCount()
+            {
+                uint count = Volatile.Read(ref ConsumeCount);
+                if (ushort.MaxValue < count)
+                {
+                    uint delta = unchecked(0u - count);
+                    _ = Interlocked.Add(ref ConsumeCount, delta);
+                    _ = Interlocked.Add(ref PredictCount, delta);
+                }
+            }
+
+            public void IncreaseConsumeCount()
+            {
+                unchecked {
+                    _ = Interlocked.Increment(ref ConsumeCount);
+                }
+            }
+
+            public void IncreasePredictCount()
+            {
+                unchecked {
+                    _ = Interlocked.Increment(ref PredictCount);
+                }
+            }
+
             #region tlsf linked list
 
             [FieldOffset(0)] public short NextLevel;
@@ -42,8 +71,15 @@ namespace Nitload.Common.HighPerformance.Memory
 
             #region buffer
 
-            [FieldOffset(8)] public uint PositionMask;
+            [FieldOffset(8)] public nuint Position;
-            [FieldOffset(12)] public uint SizeMask;
+            [FieldOffset(16)] public uint SizeMask;
+
+            #endregion
+
+            #region count
+
+            [FieldOffset(20)] public uint PredictCount;
+            [FieldOffset(24)] public uint ConsumeCount;
 
         #endregion
         }

src/Nitload.Common/HighPerformance/Memory/L4.Design.md

@@ -83,17 +83,20 @@ published buffer 地址以 `ulong` 保存：
 
 ## 6. 预测器注册、任务状态与在途缓冲
 
-这三者职责不同，不应合并成一个队列：
+预测器调度和预测任务执行是两条相关但不同的生命周期，不能把“预测器已预测完”当成“它产生的工作已完成”。
 
-- **预测器管理数组**：`PredictorMonitor` 的固定槽位数组，保存注册顺序、稳定 ID 和生命周期状态。容量上限（曾讨论 256）是预测器数量，与缓存行计数无关。预测器按依赖顺序依次执行，一个 L4 同一时刻只有一个 predictor worker 处于 prefetch 阶段，因此不需要多 lane 或 thread-local lane 归属。
+- **待调度预测器（多个）**：保存在 PredictorMonitor 的有序管理数组/队列中，等待轮到自己。固定槽位容量（曾讨论 256）表示预测器注册上限，与缓存行计数无关。
-- **Swiss 预测任务状态**：保存/查找待映射的 `AddressSource` 条目；这是预测任务/映射状态，不是预测器注册表。
+- **活动预测器（最多一个）**：当前 L4 正在执行其同步 `PredictMission`。预测器之间按依赖顺序切换，不需要多 lane 或 thread-local lane 归属。
-- **在途预测器缓冲**：预测器已经产出预测，但对应实际工作尚未完成时，保留该预测器及其 mission/写回上下文。它与注册槽位不同：注册槽位描述“有哪些预测器”，在途缓冲描述“哪些预测工作还未完成”。具体容量、完成标记和回收时机尚待设计。
+- **在途预测批次（可以多个）**：某预测器已经产出候选，但候选对应的实际任务尚未全部执行。此时要保留稳定的 predictor ID/代次、mission 或写回上下文；该记录属于任务 bag，不等同于预测器注册槽位。`PredictMission` 按预测器依赖顺序串行调用；前一批次的实际任务可以仍在途，是否与后续批次重叠由任务依赖决定。
+- **已完成批次**：对应实际工作全部完成，且所需软件写回或显式丢弃已完成。此时释放在途记录及其引用资源，不再由 L4 保存该批次。预测器本身是否注销是另一项生命周期操作。
+
+Swiss 负责预测任务/映射条目的查找状态；Monitor 负责预测器注册和调度；在途预测器缓冲负责尚未完成的预测工作及其归属。三者不可混为同一个队列或数组。
 
 `PredictMission(in Span<AddressSource>)` 是同步接口，L4 可传入 stackalloc span。若返回后仍需跨线程或跨 NUMA 处理其内容，必须将内容复制到拥有明确生命周期的缓冲区；不得让 stack span 逃逸。由于预测器顺序有依赖，待处理预测必须按既定顺序交接。若采用异步 worker，队列应有界且不阻塞 Map 快路径。
 
 ## 7. 驱逐与写回策略
 
-L4 不得推断或解码预测器的 packed 格式。若丢弃一个尚未写回、且对应实际工作仍在途的预测条目会丢失必要状态，就不能静默驱逐它。按缓存行的保护/引用字段或 sidecar 正在讨论中，引用/owner metadata 布局尚未定稿。
+L4 不得推断或解码预测器的 packed 格式。若丢弃一个尚未写回、且对应实际工作仍在途的预测条目会丢失必要状态，就不能静默驱逐它；必须由缓存行保护/引用状态阻止替换，或由 Monitor/预测器先完成写回、撤销或显式释放。相关状态可放在 KVP metadata 或 sidecar，布局尚未定稿。
 
 owner ID 可用于路由显式写回；但若多个预测器能独立声明同一条目，单一 owner 字段不够。引用计数可保护共享条目，而写回路径仍需获得预测器 identity/source spec。
 

src/Nitload.Common/HighPerformance/Memory/L4.Map.cs

@@ -18,9 +18,9 @@ namespace Nitload.Common.HighPerformance.Memory
     {
 
         private const int _low_predict = 512;
-        private readonly int _predict_slow_path_count;
 
         private int _predict_count;
+        private readonly int _predict_slow_path_count;
         private int _swiss_reader_counter;
         private int _swiss_swap_lock;
         private int _swiss_writer_lock;
@@ -45,6 +45,7 @@ namespace Nitload.Common.HighPerformance.Memory
 
                             _addressMap.SwapBuffers();
                             Volatile.Write(ref _predict_count, effectivePredictionCount);
+                            _addressMap.TryRehash();
                         }
 
                         if (_addressMap.TryGet(source, out handle)) {
@@ -61,8 +62,9 @@ namespace Nitload.Common.HighPerformance.Memory
                 }
             } else if (predictCount < _low_predict + _predict_slow_path_count)
             {
-                if (!TryGetAddress(source, out handle))
+                if (!TryGetAddress(source, out handle)) {
                     throw new InvalidOperationException($"Unexpected map request: [address source: 0x{source.Address:x16}; layout profile: {source.Description:x16}].");
+                }
             } else if (!_addressMap.TryGet(source, out handle)) {
                 throw new InvalidOperationException($"Unexpected map request: [address source: 0x{source.Address:x16}; layout profile: {source.Description:x16}].");
             }
@@ -70,8 +72,8 @@ namespace Nitload.Common.HighPerformance.Memory
             CacheLineMetadata* metadata = (CacheLineMetadata*)handle;
 
             // process metadata, unfinished
-            uint offset = metadata->PositionMask;
+            ulong offset = metadata->Position;
-            nuint address = /*base ptr + */ offset;
+            nuint address = /*base ptr + */ (nuint)offset;
             _ = Interlocked.Decrement(ref _predict_count);
 
             return new Handle(address, (nuint)handle);

src/Nitload.Common/HighPerformance/Memory/L4.Predict.cs

@@ -16,7 +16,96 @@ namespace Nitload.Common.HighPerformance.Memory
     public unsafe partial class L4
     {
 
-        internal partial int Predict() => 0;
+        private int _active_line_count;
+
+        private int _last_evict_pos;
+
+        internal partial int Predict()
+        {
+            if (!TryGetCurrentPredictor(out IL4Predictor? p)) {
+                return 0;
+            }
+            int total_count = 0, count;
+            Span<L4.AddressSource> as_sp = stackalloc L4.AddressSource[4];
+            SwissEnumerator e = _addressMap.GetWriteBufferEnumerator();
+            Predict:
+            while ((count = p.PredictMission(in as_sp)) > 0)
+            {
+                for (int i = 0; i < count; i++)
+                {
+                    if (_addressMap.TryGetWrite(as_sp[i], out ulong map_handle_adr))
+                    {
+                        L4.CacheLineMetadata* map_handle = (L4.CacheLineMetadata*)map_handle_adr;
+                        map_handle->IncreasePredictCount();
+                        map_handle->CheckCount();
+                    } else
+                    {
+                        uint size = as_sp[i].Size;
+                        int handle_idx = _bufferAllocator.Alloc(size);
+                        L4.CacheLineMetadata* handle = _bufferAllocator.MetadataAt(handle_idx);
+                        p.Prefetch(as_sp[i], handle->Position, handle->SizeMask);
+                        if (!_addressMap.TryAddOrSet(as_sp[i], (ulong)handle)) {
+                            throw new InvalidOperationException();
+                        }
+                        _active_line_count++;
+
+                        if (_active_line_count >= MaxActiveCacheLineCount)
+                        {
+                            int retry = 0;
+                            const int max_retry = 3;
+                            for (; retry < max_retry && EvictOnce(ref e); retry++) {
+                                e = _addressMap.GetWriteBufferEnumerator(_last_evict_pos);
+                            }
+                            if (retry >= max_retry) {
+                                throw new OverflowException("Evict failed, maybe all cache lines are in use.");
+                            }
+                            _active_line_count--;
+                        }
+                    }
+                }
+                total_count += count;
+                if (total_count >= 2048)
+                {
+                    _ = Interlocked.Add(ref _predict_count, total_count);
+                    _last_evict_pos = e.CurrentOffset;
+                    return total_count;
+                }
+            }
+            if (MoveNextPredictor())
+            {
+                _ = TryGetCurrentPredictor(out p);
+                goto Predict;
+            }
+            _ = Interlocked.Add(ref _predict_count, total_count);
+            _last_evict_pos = e.CurrentOffset;
+            return total_count;
+        }
+
+        private bool EvictOnce(
+                     ref SwissEnumerator e
+        )
+        {
+            while (e.MoveNext())
+            {
+                (AddressSource source, ulong cache_head_ptr) = e.Current;
+                L4.CacheLineMetadata* handle = (L4.CacheLineMetadata*)cache_head_ptr;
+                if (Volatile.Read(ref handle->ConsumeCount) == Volatile.Read(ref handle->ConsumeCount))
+                {
+                    // evict
+                    _ = _addressMap.TryRemove(source, out _);
+                    _bufferAllocator.Free(handle);
+                    return true;
+                }
+            }
+            return false;
+        }
 
     }
 }
+
+// Eviction follow-up:
+// 1. Compare ConsumeCount with PredictCount; reset both on TLSF slot reuse and count new predictions.
+// 2. Fix retry/result handling; decrement _active_line_count only after a successful eviction.
+// 3. Correct TLSF reverse-slab pointer validation before freeing metadata pointers.
+// 4. Defer TLSF Free until retired published buffers and outstanding handles can no longer reference the line.
+// 5. Add Swiss TryRemoveAt(offset) to tombstone the enumerated slot without hashing and probing again.

src/Nitload.Common/HighPerformance/Memory/L4.PredictorQueue.cs

@@ -1,4 +1,11 @@
-using System;
+// -----------------------------------------------------------------------------
+// file="L4.PredictorQueue"
+// Project: StgSharp
+// Copyright (c) Nitload.
+// SPDX-License-Identifier: MIT
+// -----------------------------------------------------------------------------
+
+using System;
 using System.Collections.Generic;
 using System.Linq;
 using System.Text;
@@ -6,7 +13,133 @@ using System.Threading.Tasks;
 
 namespace Nitload.Common.HighPerformance.Memory
 {
-    public class L4
+    public partial class L4
     {
+
+        private volatile int _consume;
+        private volatile int _enter;
+        private volatile int _predict;
+
+        private int _queue_sync;
+
+        internal void DropPredictor(
+                      IL4Predictor predictor
+        )
+        {
+            if (predictor is null) {
+                return;
+            }
+            SpinWait sp = new();
+            while (Interlocked.CompareExchange(ref _queue_sync, 1, 0) != 0) {
+                sp.SpinOnce();
+            }
+
+            try
+            {
+                unchecked
+                {
+                    for (int i = _consume; i < _predict; i++)
+                    {
+                        byte idx = (byte)i;
+                        if (ReferenceEquals(predictor, _ring[idx]))
+                        {
+                            int end = _predict;
+                            for (int move = i; move < end - 1; move++) {
+                                _ring[(byte)move] = _ring[(byte)(move + 1)];
+                            }
+                            _ring[(byte)(end - 1)] = default!;
+                            _predict = end - 1;
+                            return;
+                        }
+                    }
+                }
+            }
+            finally
+            {
+                _ = Interlocked.Exchange(ref _queue_sync, 0);
+            }
+        }
+
+        internal bool MoveNextPredictor()
+        {
+            int cur = Interlocked.Increment(ref _predict);
+            return cur < _enter;
+        }
+
+        internal void RegistPredictor(
+                      IL4Predictor predictor
+        )
+        {
+            if (predictor is null) {
+                return;
+            }
+            SpinWait sp = new();
+
+            wait_for_spare:
+            WaitForSpare();
+
+            while (Interlocked.CompareExchange(ref _queue_sync, 1, 0) != 0) {
+                sp.SpinOnce();
+            }
+            if (_enter - _consume >= PredictorRingCapacity)
+            {
+                _ = Interlocked.Exchange(ref _queue_sync, 0);
+                goto wait_for_spare;
+            }
+
+            try
+            {
+                unchecked
+                {
+                    byte idx = (byte)_enter;
+                    _enter++;
+                    _ring[idx] = predictor;
+                }
+            }
+            finally
+            {
+                _ = Interlocked.Exchange(ref _queue_sync, 0);
+            }
+        }
+
+        internal bool TryGetCurrentPredictor(
+                      out IL4Predictor predictor
+        )
+        {
+            int cur = _predict;
+            if (cur >= _enter)
+            {
+                predictor = default!;
+                return false;
+            }
+            unchecked
+            {
+                predictor = _ring[(byte)cur];
+                return true;
+            }
+        }
+
+        private void WaitForSpare()
+        {
+            int spin_count = 0;
+
+            while (_enter - _consume >= PredictorRingCapacity)
+            {
+                switch (spin_count)
+                {
+                    case < 1024:
+                        spin_count++;
+                        break;
+                    case < 65536:
+                        spin_count++;
+                        Thread.SpinWait(1);
+                        break;
+                    default:
+                        Thread.Sleep(1);
+                        break;
+                }
+            }
+        }
+
     }
 }

src/Nitload.Common/HighPerformance/Memory/L4.SwissTable.cs

@@ -6,8 +6,10 @@
 // -----------------------------------------------------------------------------
 
 using System.Buffers;
+using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
+using System.Runtime.Intrinsics;
 using System.Runtime.Intrinsics.X86;
 
 namespace Nitload.Common.HighPerformance.Memory
@@ -37,8 +39,9 @@ namespace Nitload.Common.HighPerformance.Memory
         public struct AddressSource : IEquatable<AddressSource>
         {
 
+            public uint Description;
+            public uint Size;
             public ulong Address;
-            public ulong Description;
 
             public readonly bool Equals(
                                  AddressSource other
@@ -93,11 +96,70 @@ namespace Nitload.Common.HighPerformance.Memory
 
         }
 
+        internal unsafe struct SwissEnumerator(
+                               byte* buffer,
+                               int begins_at = 0
+        )
+        {
+
+            private readonly SwissBucket* _buckets = (SwissBucket*)(buffer + IdealCount);
+            private readonly Vector128<byte>* _control_bytes = (Vector128<byte>*)buffer;
+            private bool _group_loaded = false;
+            private int _visited_group = 0;
+            private int _current_group = (begins_at / BucketCapacity) & (GroupCount - 1);
+            private int _current_mask = 0;
+            private int _current_offset = 0;
+
+            public (AddressSource key, ulong value) Current
+            {
+                get
+                {
+                    SwissBucket* bucket = &_buckets[_current_group];
+                    AddressSource key = bucket->Keys[_current_offset];
+                    ulong value = bucket->Values[_current_offset];
+                    return (key, value);
+                }
+            }
+
+            public readonly int CurrentOffset => (_current_group * BucketCapacity) + _current_offset;
+
+            public bool MoveNext()
+            {
+                while (_visited_group < GroupCount)
+                {
+                    if (!_group_loaded)
+                    {
+                        if (Sse42.IsSupported)
+                        {
+                            _current_mask = SwissReaderX86.CompareOccupiedMask(*(_control_bytes + _current_group));
+                        } else
+                        {
+                            throw new PlatformNotSupportedException();
+                        }
+                        _group_loaded = true;
+                    }
+
+                    if (_current_mask != 0)
+                    {
+                        _current_offset = BitOperations.TrailingZeroCount(_current_mask);
+                        _current_mask &= _current_mask - 1;
+                        return true;
+                    }
+
+                    _current_group = (_current_group + 1) & (GroupCount - 1);
+                    _visited_group++;
+                    _group_loaded = false;
+                }
+                return false;
+            }
+
+        }
+
         internal unsafe struct SwissTable : IDisposable
         {
 
             private byte* _buffer_write;
-
+            private int _delete_count;
             private ulong _published_buffer;
 
             private SwissTable(
@@ -133,13 +195,13 @@ namespace Nitload.Common.HighPerformance.Memory
             /// <summary>
             ///   Writes to the inactive buffer. The caller must serialize all writes.
             /// </summary>
-            public readonly bool TryAddOrSet(
+            public bool TryAddOrSet(
                         AddressSource key,
                         ulong value
             )
             {
                 if (Sse42.IsSupported) {
-                    return SwissWriterX86.TryAddOrSet(_buffer_write, key, value);
+                    return SwissWriterX86.TryAddOrSet(_buffer_write, key, value, ref _delete_count);
                 }
 
                 throw new NotImplementedException("The Swiss backend is not implemented for this ISA.");
@@ -161,6 +223,56 @@ namespace Nitload.Common.HighPerformance.Memory
                 throw new NotImplementedException("The Swiss backend is not implemented for this ISA.");
             }
 
+            /// <summary>
+            ///   Reads the inactive write buffer. The caller must serialize this query with all
+            ///   writes.
+            /// </summary>
+            public bool TryGetWrite(
+                        AddressSource key,
+                        out ulong value
+            )
+            {
+                if (Sse42.IsSupported) {
+                    return SwissReaderX86.TryGet(_buffer_write, key, out value);
+                }
+
+                throw new NotImplementedException("The Swiss backend is not implemented for this ISA.");
+            }
+
+            public void TryRehash()
+            {
+                if (_delete_count < 2048)
+                {
+                    NativeMemory.Copy(_buffer_write, (byte*)_published_buffer, (nuint)TableBufferSize);
+                    return;
+                }
+                try
+                {
+                    SwissEnumerator e = new((byte*)_published_buffer);
+                    InitializeBuffer(_buffer_write);
+                    if (Sse42.IsSupported)
+                    {
+                        while (e.MoveNext())
+                        {
+                            (AddressSource key, ulong value) = e.Current;
+                            if (!SwissWriterX86.TryAddOrSet(_buffer_write, key, value,
+                                                            ref _delete_count)) {
+                                throw new InvalidOperationException("The Swiss table could not fit all active entries during rehash.");
+                            }
+                        }
+                        _delete_count = 0;
+                    } else
+                    {
+                        throw new PlatformNotSupportedException();
+                    }
+                }
+                catch (Exception)
+                {
+                    NativeMemory.Copy(_buffer_write, (byte*)_published_buffer, (nuint)TableBufferSize);
+                    throw;
+                }
+            }
+
             /// <summary>
             ///   Removes from the inactive buffer. The caller must serialize all writes.
             /// </summary>
@@ -169,12 +281,15 @@ namespace Nitload.Common.HighPerformance.Memory
                         out ulong value
             )
             {
-                if (Sse42.IsSupported) {
+                if (Sse42.IsSupported)
-                    return SwissWriterX86.TryRemove(_buffer_write, key, out value);
+                {
-                }
+                    return SwissWriterX86.TryRemove(_buffer_write, key, out value,
-
+                                                    ref _delete_count);
+                } else
+                {
                     throw new NotImplementedException("The Swiss backend is not implemented for this ISA.");
                 }
+            }
 
             internal static SwissTable Create(
                                        byte* buffers
@@ -184,6 +299,14 @@ namespace Nitload.Common.HighPerformance.Memory
                 return new SwissTable(buffers);
             }
 
+            // Enumerates the write buffer; the caller must serialize writers and prevent swaps.
+            internal readonly SwissEnumerator GetWriteBufferEnumerator(
+                                              int begins_at = 0
+            )
+            {
+                return new SwissEnumerator(_buffer_write, begins_at);
+            }
+
             private static void InitializeBuffer(
                                 byte* buffer
             )

src/Nitload.Common/HighPerformance/Memory/L4.Swiss_X86.cs

@@ -39,7 +39,8 @@ namespace Nitload.Common.HighPerformance.Memory
             public static bool TryAddOrSet(
                                byte* buffer,
                                AddressSource key,
-                               ulong value
+                               ulong value,
+                               ref int delete_count
             )
             {
                 Vector128<byte>* control_bytes = (Vector128<byte>*)buffer;
@@ -101,6 +102,7 @@ namespace Nitload.Common.HighPerformance.Memory
                             bucket->Keys[first_tomb_idx] = key;
                             bucket->Values[first_tomb_idx] = value;
                             ((byte*)(control_bytes + first_tomb_group))[first_tomb_idx] = target_hash;
+                            delete_count--;
                         }
                         return true;
                     }
@@ -113,6 +115,7 @@ namespace Nitload.Common.HighPerformance.Memory
                     bucket->Keys[first_tomb_idx] = key;
                     bucket->Values[first_tomb_idx] = value;
                     ((byte*)(control_bytes + first_tomb_group))[first_tomb_idx] = target_hash;
+                    delete_count--;
                     return true;
                 }
                 value = 0;
@@ -122,7 +125,8 @@ namespace Nitload.Common.HighPerformance.Memory
             public static bool TryRemove(
                                byte* buffer,
                                AddressSource key,
-                               out ulong value
+                               out ulong value,
+                               ref int delete_count
             )
             {
                 Vector128<byte>* control_bytes = (Vector128<byte>*)buffer;
@@ -148,6 +152,7 @@ namespace Nitload.Common.HighPerformance.Memory
                         {
                             ((byte*)(control_bytes + begin))[index] = (byte)SwissControl.Deleted;
                             value = bucket->Values[index];
+                            delete_count++;
                             return true;
                         }
                         mask &= ~(1 << index);
@@ -171,6 +176,20 @@ namespace Nitload.Common.HighPerformance.Memory
         internal readonly unsafe struct SwissReaderX86
         {
 
+            private static readonly Vector128<byte> _empty_mask = Vector128.Create((byte)SwissControl.Empty);
+            private static readonly Vector128<byte> _tomb_mask = Vector128.Create((byte)SwissControl.Deleted);
+
+            [MethodImpl(MethodImplOptions.AggressiveInlining)]
+            public static int CompareOccupiedMask(
+                              Vector128<byte> control_bytes_v128
+            )
+            {
+                Vector128<byte> cmp_eql = Sse2.CompareEqual(control_bytes_v128, _empty_mask);
+                Vector128<byte> cmp_del = Sse2.CompareEqual(control_bytes_v128, _tomb_mask);
+                int specialMask = Sse2.MoveMask(Sse2.Or(cmp_eql, cmp_del));
+                return (ushort)~(ushort)specialMask;
+            }
+
             public static bool TryGet(
                                byte* buffer,
                                AddressSource key,

src/Nitload.Common/HighPerformance/Memory/L4.Tlsf.cs

@@ -5,27 +5,29 @@
 // SPDX-License-Identifier: MIT
 // -----------------------------------------------------------------------------
 
-using Nitload.Common.HighPerformance.ProcessorAbstraction;
-using System;
-using System.Collections.Generic;
-using System.Linq;
 using System.Numerics;
-using System.Runtime.InteropServices;
+using System.Reflection.Metadata;
-using System.Text;
-using System.Threading.Tasks;
 
 namespace Nitload.Common.HighPerformance.Memory
 {
     public partial class L4
     {
 
+        /// <summary>
+        ///   L4-specific TLSF allocator whose metadata nodes also serve as cache-line headers.
+        /// </summary>
+        /// <remarks>
+        ///   Payload occupies the first arena region; a fixed metadata slab follows it and is
+        ///   carved backward from its aligned high end. Handle 0 is the sentinel, and positive
+        ///   handles are metadata slot indices. This layout is intentionally coupled to L4.
+        /// </remarks>
         private unsafe struct TLSF
         {
 
-            public const int MaxCacheLineCount = 8192;
+            public const int MaxCacheLineCount = MaxActiveCacheLineCount;
 
             public static readonly nuint RequestedAllocation =
-                    ((nuint)MaxCacheLineCount + 1u) * CacheLineMetadata.SizeOf;
+                    (((nuint)MaxCacheLineCount + 1u) * CacheLineMetadata.SizeOf) + CacheLineMetadata.SizeOf - 1u;
             private readonly CacheLineMetadata* _spareMemory;
             private readonly int _align;
             private readonly int _maxLevel;
@@ -52,12 +54,16 @@ namespace Nitload.Common.HighPerformance.Memory
 
                 // _bucket = new short[GetLevel(_maxAllocSize) + 2];
                 _bucket[63] = (short)(GetLevel(_maxAllocSize) + 2);
-                _spareMemory = (CacheLineMetadata*)(ulong)(_basePtr + _totalSize);
+
+                // Metadata occupies the reservation after payload: handle 0 is the sentinel, and handles 1..MaxCacheLineCount are carved backward from the aligned high end until the slab is full.
+                nuint metadataRegionEnd = (nuint)externalPointer + byteSize + RequestedAllocation;
+                nuint alignedMetadataEnd = metadataRegionEnd & ~((nuint)CacheLineMetadata.SizeOf - 1u);
+                _spareMemory = (CacheLineMetadata*)(alignedMetadataEnd - CacheLineMetadata.SizeOf);
                 _metadataEnd = 0;
 
                 // init spare memory node
 
-                _spareMemory->PositionMask = 0;
+                _spareMemory->Position = (nuint)(ulong)_basePtr;
                 _spareMemory->SizeMask = _totalSize;
                 _spareMemory->NextLevel = 0;
                 _spareMemory->NextNear = 0;
@@ -66,7 +72,15 @@ namespace Nitload.Common.HighPerformance.Memory
             }
 
             [MethodImpl(MethodImplOptions.AggressiveInlining)]
-            private int GetLevel(
+            internal readonly CacheLineMetadata* MetadataAt(
+                                                 int handle
+            )
+            {
+                return _spareMemory - handle;
+            }
+
+            [MethodImpl(MethodImplOptions.AggressiveInlining)]
+            private readonly int GetLevel(
                                  uint size
             )
             {
@@ -93,11 +107,11 @@ namespace Nitload.Common.HighPerformance.Memory
                 // remove from position chain
                 short prevNearIdx = node->PreviousNear;
                 short nextNearIdx = node->NextNear;
-                CacheLineMetadata* prevNear = &_spareMemory[prevNearIdx];
+                CacheLineMetadata* prevNear = MetadataAt(prevNearIdx);
                 prevNear->NextNear = nextNearIdx;
                 if (nextNearIdx != 0)
                 {
-                    CacheLineMetadata* nextNear = &_spareMemory[nextNearIdx];
+                    CacheLineMetadata* nextNear = MetadataAt(nextNearIdx);
                     nextNear->PreviousNear = prevNearIdx;
                 } else
                 {
@@ -106,7 +120,7 @@ namespace Nitload.Common.HighPerformance.Memory
                 node->PreviousNear = 0;
                 node->NextNear = 0;
 
-                node->PositionMask = 0;
+                node->Position = (nuint)(ulong)_basePtr;
                 node->SizeMask = 0;
                 node->State = AllocationState.SpareOrCollect;
                 node->PreviousLevel = 0;
@@ -114,7 +128,7 @@ namespace Nitload.Common.HighPerformance.Memory
                 short nextIdx = _spareMemory->NextLevel;
                 if (nextIdx != 0)
                 {
-                    CacheLineMetadata* next = &_spareMemory[nextIdx];
+                    CacheLineMetadata* next = MetadataAt(nextIdx);
                     next->PreviousLevel = nodeIdx;
                 }
                 node->NextLevel = nextIdx;
@@ -145,12 +159,12 @@ namespace Nitload.Common.HighPerformance.Memory
                     _bucket[currentLevel] = currentNextLevelIdx;
                 } else if (currentPrevLevelIdx != 0)
                 {
-                    CacheLineMetadata* prevLevel = &_spareMemory[currentPrevLevelIdx];
+                    CacheLineMetadata* prevLevel = MetadataAt(currentPrevLevelIdx);
                     prevLevel->NextLevel = currentNextLevelIdx;
                 }
                 if (currentNextLevelIdx != 0)
                 {
-                    CacheLineMetadata* nextLevel = &_spareMemory[currentNextLevelIdx];
+                    CacheLineMetadata* nextLevel = MetadataAt(currentNextLevelIdx);
                     nextLevel->PreviousLevel = currentPrevLevelIdx;
                 }
                 node->PreviousLevel = 0;
@@ -203,11 +217,11 @@ namespace Nitload.Common.HighPerformance.Memory
                         meta = null;
                         return false;
                     }
-                    meta = &_spareMemory[idx];
+                    meta = MetadataAt(idx);
                     short nextSpareIdx = meta->NextLevel;
                     if (nextSpareIdx != 0)
                     {
-                        CacheLineMetadata* nextSpare = &_spareMemory[nextSpareIdx];
+                        CacheLineMetadata* nextSpare = MetadataAt(nextSpareIdx);
                         nextSpare->PreviousLevel = 0;
                     }
                     _spareMemory->NextLevel = nextSpareIdx;
@@ -215,18 +229,19 @@ namespace Nitload.Common.HighPerformance.Memory
                 if (!TryGetNewMetadata(out meta)) {
                     return false;
                 }
+                idx = (short)(_spareMemory - meta);
                 spareSize -= size;
                 meta->SizeMask = size;
-                meta->PositionMask = _spareMemory->PositionMask;
+                meta->Position = _spareMemory->Position;
                 meta->State = AllocationState.Free;
                 meta->PreviousLevel = 0;
                 meta->NextLevel = 0;
                 _spareMemory->SizeMask = spareSize;
-                _spareMemory->PositionMask += size;
+                _spareMemory->Position += size;
 
                 // insert the new chunk to position chain
                 short nearIdx = _spareMemory->PreviousNear;
-                CacheLineMetadata* near = &_spareMemory[nearIdx];
+                CacheLineMetadata* near = MetadataAt(nearIdx);
                 near->NextNear = idx;
                 meta->PreviousNear = nearIdx;
                 _spareMemory->PreviousNear = idx;
@@ -245,7 +260,7 @@ namespace Nitload.Common.HighPerformance.Memory
                     return false;
                 }
                 _metadataEnd++;
-                meta = &_spareMemory[_metadataEnd];
+                meta = MetadataAt(_metadataEnd);
                 return true;
             }
 
@@ -261,7 +276,7 @@ namespace Nitload.Common.HighPerformance.Memory
                     int idx = _bucket[level];
                     while (idx != 0)
                     {
-                        CacheLineMetadata* candidate = &_spareMemory[idx];
+                        CacheLineMetadata* candidate = MetadataAt(idx);
                         if (candidate->SizeMask >= minimumSize)
                         {
                             RemoveFromBucket(level, candidate, idx);
@@ -276,7 +291,7 @@ namespace Nitload.Common.HighPerformance.Memory
                     int idx = _bucket[i];
                     if (idx != 0)
                     {
-                        CacheLineMetadata* candidate = &_spareMemory[idx];
+                        CacheLineMetadata* candidate = MetadataAt(idx);
                         RemoveFromBucketHead(i, candidate, idx);
                         entry = candidate;
                         return true;
@@ -313,7 +328,7 @@ namespace Nitload.Common.HighPerformance.Memory
                     TryGetNewMetadata(out CacheLineMetadata* newEntry))
                 {
                     short newEntryIdx = (short)(_spareMemory - newEntry);
-                    newEntry->PositionMask = entry->PositionMask + allocSize;
+                    newEntry->Position = entry->Position + allocSize;
                     newEntry->SizeMask = remain;
                     newEntry->State = AllocationState.Free;
                     int newLevel = GetLevel(remain);
@@ -321,7 +336,7 @@ namespace Nitload.Common.HighPerformance.Memory
                     if (bucketHeadIdx != 0)
                     {
                         // if there is already an entry in the bucket, link it to the new entry
-                        CacheLineMetadata* bucketHead = &_spareMemory[bucketHeadIdx];
+                        CacheLineMetadata* bucketHead = MetadataAt(bucketHeadIdx);
                         bucketHead->PreviousLevel = newEntryIdx;
                     }
                     newEntry->NextLevel = bucketHeadIdx;
@@ -331,14 +346,14 @@ namespace Nitload.Common.HighPerformance.Memory
 
                     // push to position chain
                     short afterIdx = entry->NextNear;
-                    CacheLineMetadata* after = &_spareMemory[afterIdx];
+                    CacheLineMetadata* after = MetadataAt(afterIdx);
                     newEntry->NextNear = afterIdx;
                     after->PreviousNear = newEntryIdx;
                     entry->NextNear = newEntryIdx;
                     newEntry->PreviousNear = entryIdx;
                 }
                 entry->State = AllocationState.Alloc;
-                if (entry->PositionMask - (ulong)_basePtr > _totalSize) {
+                if (entry->Position - (ulong)_basePtr > _totalSize) {
                     throw new OverflowException("Allocated memory address exceeds the maximum block size.");
                 }
                 return entryIdx;
@@ -353,7 +368,7 @@ namespace Nitload.Common.HighPerformance.Memory
                     {
                         break;
                     }
-                    CacheLineMetadata* current = &_spareMemory[currentIdx];
+                    CacheLineMetadata* current = MetadataAt(currentIdx);
                     if (current->State == AllocationState.Free)
                     {
                         // scan next
@@ -361,7 +376,7 @@ namespace Nitload.Common.HighPerformance.Memory
                         int nextIdx = current->NextNear;
                         while (nextIdx != 0)
                         {
-                            CacheLineMetadata* next = &_spareMemory[nextIdx];
+                            CacheLineMetadata* next = MetadataAt(nextIdx);
                             if (next->State != AllocationState.Free)
                             {
                                 break;
@@ -378,7 +393,7 @@ namespace Nitload.Common.HighPerformance.Memory
                             short bucketHeadIdx = _bucket[newLevel];
                             if (bucketHeadIdx != 0)
                             {
-                                CacheLineMetadata* bucketHead = &_spareMemory[bucketHeadIdx];
+                                CacheLineMetadata* bucketHead = MetadataAt(bucketHeadIdx);
                                 bucketHead->PreviousLevel = currentIdx;
                             }
                             current->NextLevel = bucketHeadIdx;
@@ -391,11 +406,13 @@ namespace Nitload.Common.HighPerformance.Memory
             }
 
             public void Free(
-                        int handle,
+                        CacheLineMetadata* head,
                         int collectScanWindow = 4
             )
             {
-                CacheLineMetadata* head = &_spareMemory[handle];
+                if (head < _spareMemory || head >= _spareMemory + MaxCacheLineCount) {
+                    throw new ArgumentOutOfRangeException(nameof(head), "Handle is out of range.");
+                }
                 short headIdx = (short)(_spareMemory - head);
                 uint size = head->SizeMask;
                 for (int i = 0; i < collectScanWindow; i++)
@@ -406,13 +423,13 @@ namespace Nitload.Common.HighPerformance.Memory
                     {
                         break;
                     }
-                    CacheLineMetadata* prev = &_spareMemory[forwardIdx];
+                    CacheLineMetadata* prev = MetadataAt(forwardIdx);
                     if (prev->State != AllocationState.Free)
                     {
                         break;
                     }
                     size += prev->SizeMask;
-                    head->PositionMask = prev->PositionMask;
+                    head->Position = prev->Position;
                     RecycleNode(prev);
                 }
                 for (int i = 0; i < collectScanWindow; i++)
@@ -423,7 +440,7 @@ namespace Nitload.Common.HighPerformance.Memory
                     {
                         break;
                     }
-                    CacheLineMetadata* prev = &_spareMemory[nextIdx];
+                    CacheLineMetadata* prev = MetadataAt(nextIdx);
                     if (prev->State != AllocationState.Free)
                     {
                         break;
@@ -441,7 +458,7 @@ namespace Nitload.Common.HighPerformance.Memory
                 head->NextLevel = 0;
                 if (bucketHeadIdx != 0)
                 {
-                    CacheLineMetadata* bucketHead = &_spareMemory[bucketHeadIdx];
+                    CacheLineMetadata* bucketHead = MetadataAt(bucketHeadIdx);
                     bucketHead->PreviousLevel = headIdx;
                     head->NextLevel = bucketHeadIdx;
                 }
@@ -449,6 +466,27 @@ namespace Nitload.Common.HighPerformance.Memory
                 return;
             }
 
+            public void Free(
+                        ulong handle,
+                        int collectScanWindow = 4
+            )
+            {
+                if (handle < (ulong)_basePtr || handle >= (ulong)_basePtr + _totalSize) {
+                    throw new ArgumentOutOfRangeException(nameof(handle), "Handle is out of range.");
+                }
+                CacheLineMetadata* head = (CacheLineMetadata*)handle;
+                Free(head, collectScanWindow);
+            }
+
+            public void Free(
+                        int handle,
+                        int collectScanWindow = 4
+            )
+            {
+                CacheLineMetadata* head = MetadataAt(handle);
+                Free(head, collectScanWindow);
+            }
+
             public bool TryAlloc(
                         uint size,
                         out int h
@@ -479,7 +517,7 @@ namespace Nitload.Common.HighPerformance.Memory
                 if (remain > (entrySize >> 7) && TryGetNewMetadata(out CacheLineMetadata* newEntry))
                 {
                     short newEntryIdx = (short)(_spareMemory - newEntry);
-                    newEntry->PositionMask = entry->PositionMask + allocSize;
+                    newEntry->Position = entry->Position + allocSize;
                     newEntry->SizeMask = remain;
                     newEntry->State = AllocationState.Free;
                     int newLevel = GetLevel(remain);
@@ -487,7 +525,7 @@ namespace Nitload.Common.HighPerformance.Memory
                     if (bucketHeadIdx != 0)
                     {
                         // if there is already an entry in the bucket, link it to the new entry
-                        CacheLineMetadata* bucketHead = &_spareMemory[bucketHeadIdx];
+                        CacheLineMetadata* bucketHead = MetadataAt(bucketHeadIdx);
                         newEntry->NextLevel = bucketHeadIdx;
                         bucketHead->PreviousLevel = newEntryIdx;
                     } else
@@ -499,14 +537,14 @@ namespace Nitload.Common.HighPerformance.Memory
 
                     // push to position chain
                     short afterIdx = entry->NextNear;
-                    CacheLineMetadata* after = &_spareMemory[afterIdx];
+                    CacheLineMetadata* after = MetadataAt(afterIdx);
                     newEntry->NextNear = afterIdx;
                     after->PreviousNear = newEntryIdx;
                     entry->NextNear = newEntryIdx;
                     newEntry->PreviousNear = entryIdx;
                 }
                 entry->State = AllocationState.Alloc;
-                if (entry->PositionMask - (ulong)_basePtr > _totalSize)
+                if (entry->Position - (ulong)_basePtr > _totalSize)
                 {
                     h = default;
                     throw new OverflowException("Allocated memory address exceeds the maximum block size.");

src/Nitload.Common/HighPerformance/Memory/L4.cs

@@ -12,7 +12,12 @@ namespace Nitload.Common.HighPerformance.Memory
     public sealed unsafe partial class L4 : IDisposable
     {
 
+        private const int MaxActiveCacheLineCount = 8192;
+
         private const nuint PageSize = 4096;
+        private const int PredictorRingCapacity = 256;
+
+        private readonly IL4Predictor[] _ring;
 
         private bool _disposed;
         private readonly int _bindedNuma;
@@ -29,6 +34,7 @@ namespace Nitload.Common.HighPerformance.Memory
         {
             ArgumentOutOfRangeException.ThrowIfGreaterThan(bufferSize, CacheLineMetadata.MaxBufferSize);
             ArgumentOutOfRangeException.ThrowIfGreaterThan(maxCacheLineSize, CacheLineMetadata.MaxBufferSize);
+            _ring = new IL4Predictor[PredictorRingCapacity];
             _bindedNuma = bindedNuma;
             _predict_slow_path_count = GetPredictSlowPathCount(bindedNuma);
             nuint normalizedBufferSize = (bufferSize + (PageSize - 1)) & (~(PageSize - 1));
@@ -73,6 +79,8 @@ namespace Nitload.Common.HighPerformance.Memory
             _bufferAllocator.Free(handle);
         }
 
+        internal partial int Predict();
+
         internal bool TryAllocate(
                       uint size,
                       out int handle
@@ -82,15 +90,16 @@ namespace Nitload.Common.HighPerformance.Memory
             return _bufferAllocator.TryAlloc(size, out handle);
         }
 
-        internal partial int Predict();
+        private static int GetPredictSlowPathCount(
-
+                           int bindedNuma
-        private static int GetPredictSlowPathCount(int bindedNuma)
+        )
         {
             foreach (NumaNode node in Numa.AllNumaNode)
             {
-                if (node.NumaNodeId == (uint)bindedNuma)
+                if (node.NumaNodeId == (uint)bindedNuma) {
                     return checked(node.LogicalCoreCount * 2);
                 }
+            }
 
             throw new ArgumentOutOfRangeException(nameof(bindedNuma), "No topology information exists for the bound NUMA node.");
         }

src/Nitload.Common/HighPerformance/Memory/L4Predictor.cs

@@ -22,6 +22,12 @@ namespace Nitload.Common.HighPerformance.Memory
             in Span<L4.AddressSource> RequiredBuffer
         );
 
+        void Prefetch(
+             L4.AddressSource RequiredBuffer,
+             nuint allocBuffer,
+             nuint bufferSize
+        );
+
     }
 
     public interface IL4Predictor<TMission> : IL4Predictor
@@ -32,6 +38,4 @@ namespace Nitload.Common.HighPerformance.Memory
         );
 
     }
-
-    internal class PredictorMonitor { }
 }

src/Nitload.Native/CMakeLists.txt

@@ -17,39 +17,16 @@ nfn_resolve_target_identifiers(
   NFN_RESOLVED_ARCH_NAME
   NFN_RESOLVED_ARCH_DEFINE)
 
-set(NFN_NATIVE_SOURCES
+# CONFIGURE_DEPENDS makes CMake reconfigure when files are added or removed,
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/dllmain.c
+# so native sources, headers, and templates need no per-file CMake edits.
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/load_glfw.c
+file(GLOB_RECURSE NFN_NATIVE_SOURCES CONFIGURE_DEPENDS
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/load_intrinsic.x86.c
+  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/*.c)
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/m.bufop.x86.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/m.solve.x86.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/m.transpose.x86.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/mk.fma.pack.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/mk.fma.x86.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/mk.p_scan.x86.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/s.factorial.x86.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/s.hash.x86.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/simdid_compare.x86.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/simdid.x86.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/sk.fma.x86.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/thread.linux.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/thread.windows.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/vector.c
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-src/vector_dot.c)
 
-set(NFN_NATIVE_HEADERS
+file(GLOB_RECURSE NFN_NATIVE_HEADERS CONFIGURE_DEPENDS
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-head/glfw_function.h
+  ${CMAKE_CURRENT_LIST_DIR}/nfn-head/*.h)
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_internal.h
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_intrinsic.context.base.h
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_intrinsic.context.x86.h
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_intrinsic.h
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_intrinsic.matkernel.h
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_intrinsic.std.h
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_target.h
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-head/NitloadFrameWork.h)
 
-set(NFN_NATIVE_TEMPLATES
+file(GLOB_RECURSE NFN_NATIVE_TEMPLATES CONFIGURE_DEPENDS
-  ${CMAKE_CURRENT_LIST_DIR}/nfn-template/nfn_template.fma.x86.h)
+  ${CMAKE_CURRENT_LIST_DIR}/nfn-template/*.h)
 
 # Nitload.Native now lives under <root>/src/Nitload.Native, so walk up two
 # levels (project dir -> src -> solution root) to reach the solution root.

src/Nitload.Native/nfn-head/NitloadFrameWork.h

@@ -27,9 +27,11 @@ NFN_API SIMDID NFN_DECL nfn_get_simd_level_global();
 NFN_API SIMDID NFN_DECL nfn_get_simd_level_local();
 NFN_API SIMDID NFN_DECL nfn_get_unite_simd(SIMDID left, SIMDID right);
 NFN_API int NFN_DECL    nfn_compare_simd(SIMDID left, SIMDID right);
-NFN_API void NFN_DECL nfn_set_thread_affinity(int core_id, int numa_id);
+NFN_API void NFN_DECL   nfn_set_thread_affinity(int cpu_index, int processor_group_id);
-NFN_API void* NFN_DECL nfn_numa_alloc(size_t size, int numa_id);
+NFN_API void *NFN_DECL  nfn_numa_alloc(size_t size, int numa_id);
 NFN_API void NFN_DECL   nfn_numa_free(void *ptr, size_t size, int numa_id);
-NFN_API int NFN_DECL nfn_get_numa_each(int numa_id, uint64_t *internal, uint32_t *global);
+NFN_API int NFN_DECL    nfn_get_numa_each(int numa_id, uint32_t *processor_group_ids,
+                                          uint32_t *socket_ids, uint32_t *cpu_indices,
+                                          uint32_t *numa_node_ids);
 NFN_API int NFN_DECL    nfn_get_numa_overall(int *out_numa_count, int *count_of_each_numa);
 #endif

src/Nitload.Native/nfn-src/thread.linux.c

@@ -26,18 +26,12 @@ typedef struct nfn_linux_max_context {
         int max_value;
 } nfn_linux_max_context;
 
-<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.linux.c
 typedef struct nfn_linux_fill_cpu_context {
-        uint64_t *internal;
-        uint32_t *global;
-=======
-typedef struct nif_linux_fill_cpu_context {
         uint32_t *processor_group_ids;
         uint32_t *socket_ids;
         uint32_t *cpu_indices;
         uint32_t *numa_node_ids;
         int numa_id;
->>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.linux.c
         int index;
 } nfn_linux_fill_cpu_context;
 
@@ -127,12 +121,7 @@ static int nfn_linux_track_max_id(int value, void *context)
         return 1;
 }
 
-<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.linux.c
+static int nfn_linux_read_socket_id(int cpu_id, uint32_t *socket_id)
-static int nfn_linux_fill_cpu_id(int value, void *context)
-{
-        nfn_linux_fill_cpu_context *fill_context = (nfn_linux_fill_cpu_context *)context;
-=======
-static int nif_linux_read_socket_id(int cpu_id, uint32_t *socket_id)
 {
         char path[160];
         char buffer[64];
@@ -140,25 +129,24 @@ static int nif_linux_read_socket_id(int cpu_id, uint32_t *socket_id)
         long parsed_value;
         snprintf(path, sizeof(path),
                  "/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu_id);
-        if (!nif_linux_read_first_line(path, buffer, sizeof(buffer)))
+        if (!nfn_linux_read_first_line(path, buffer, sizeof(buffer)))
                 return 0;
 
         errno = 0;
         parsed_value = strtol(buffer, &end, 10);
-        if (end == buffer || errno == ERANGE || parsed_value < 0 || parsed_value > UINT32_MAX)
+        if (end == buffer || errno == ERANGE || parsed_value < 0 || (uintmax_t)parsed_value > UINT32_MAX)
                 return 0;
 
         *socket_id = (uint32_t)parsed_value;
         return 1;
 }
 
-static int nif_linux_fill_cpu_id(int value, void *context)
+static int nfn_linux_fill_cpu_id(int value, void *context)
 {
-        nif_linux_fill_cpu_context *fill_context = (nif_linux_fill_cpu_context *)context;
+        nfn_linux_fill_cpu_context *fill_context = (nfn_linux_fill_cpu_context *)context;
         uint32_t socket_id = 0;
-        if (!nif_linux_read_socket_id(value, &socket_id))
+        if (!nfn_linux_read_socket_id(value, &socket_id))
                 return 0;
->>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.linux.c
 
         fill_context->processor_group_ids[fill_context->index] = 0u;
         fill_context->socket_ids[fill_context->index] = socket_id;
@@ -249,15 +237,8 @@ static int nfn_linux_populate_writable_mapping(void *address, size_t size)
 #endif
 }
 
-<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.linux.c
-/* Pin the calling thread to a single logical processor.
- * core_id >= 0 : bind to that logical processor (global index, across groups).
- * core_id <  0 : clear affinity constraint within the current group. */
-NFN_API void NFN_DECL nfn_set_thread_affinity(int core_id, int numa_id)
-=======
 /* Linux has no Windows-style processor groups; processor_group_id is ignored. */
-NIF_API void NIF_DECL nif_set_thread_affinity(int cpu_index, int processor_group_id)
+NFN_API void NFN_DECL nfn_set_thread_affinity(int cpu_index, int processor_group_id)
->>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.linux.c
 {
         (void)processor_group_id;
 
@@ -311,15 +292,11 @@ NFN_API int NFN_DECL nfn_get_numa_overall(int *out_numa_count, int *count_of_eac
         return 1;
 }
 
-<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.linux.c
+NFN_API int NFN_DECL nfn_get_numa_each(int numa_id,
-NFN_API int NFN_DECL nfn_get_numa_each(int numa_id, uint64_t *internal, uint32_t *global)
-=======
-NIF_API int NIF_DECL nif_get_numa_each(int numa_id,
                                        uint32_t *processor_group_ids,
                                        uint32_t *socket_ids,
                                        uint32_t *cpu_indices,
                                        uint32_t *numa_node_ids)
->>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.linux.c
 {
         if (numa_id < 0 || processor_group_ids == NULL || socket_ids == NULL ||
             cpu_indices == NULL || numa_node_ids == NULL)
@@ -331,11 +308,7 @@ NIF_API int NIF_DECL nif_get_numa_each(int numa_id,
         if (!nfn_linux_read_first_line(path, buffer, sizeof(buffer)))
                 return 0;
 
-<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.linux.c
+        nfn_linux_fill_cpu_context fill_context = {
-        nfn_linux_fill_cpu_context fill_context = { internal, global, 0 };
-        return nfn_linux_visit_id_list(buffer, nfn_linux_fill_cpu_id, &fill_context);
-=======
-        nif_linux_fill_cpu_context fill_context = {
                 processor_group_ids,
                 socket_ids,
                 cpu_indices,
@@ -343,8 +316,7 @@ NIF_API int NIF_DECL nif_get_numa_each(int numa_id,
                 numa_id,
                 0
         };
-        return nif_linux_visit_id_list(buffer, nif_linux_fill_cpu_id, &fill_context);
+        return nfn_linux_visit_id_list(buffer, nfn_linux_fill_cpu_id, &fill_context);
->>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.linux.c
 }
 
 NFN_API void *NFN_DECL nfn_numa_alloc(size_t size, int numa_id)

src/Nitload.Native/nfn-src/thread.windows.c

@@ -2,15 +2,12 @@
 #include "NitloadFrameWork.h"
 
 #if NFN_IS_OS(NFN_OS_WINDOWS)
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <windows.h>
 
-<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.windows.c
+static int nfn_windows_get_socket_map(uint32_t **out_socket_map, DWORD *out_group_count)
-/* Pin the calling thread to a single logical processor.
- * core_id >= 0 : bind to that logical processor (global index, across groups).
- * core_id <  0 : clear affinity constraint within the current group. */
-NFN_API void NFN_DECL nfn_set_thread_affinity(int core_id, int numa_id)
-=======
-static int nif_windows_get_socket_map(uint32_t **out_socket_map, DWORD *out_group_count)
->>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.windows.c
 {
         DWORD group_count = GetActiveProcessorGroupCount();
         DWORD buffer_size = 0;
@@ -71,7 +68,7 @@ static int nif_windows_get_socket_map(uint32_t **out_socket_map, DWORD *out_grou
         return 1;
 }
 
-static int nif_windows_get_processor_node(WORD group, BYTE cpu, USHORT *node_id)
+static int nfn_windows_get_processor_node(WORD group, BYTE cpu, USHORT *node_id)
 {
         PROCESSOR_NUMBER processor;
         memset(&processor, 0, sizeof(processor));
@@ -81,7 +78,7 @@ static int nif_windows_get_processor_node(WORD group, BYTE cpu, USHORT *node_id)
 }
 
 /* Pin the calling thread using its processor group and group-local CPU index. */
-NIF_API void NIF_DECL nif_set_thread_affinity(int cpu_index, int processor_group_id)
+NFN_API void NFN_DECL nfn_set_thread_affinity(int cpu_index, int processor_group_id)
 {
         if (cpu_index < 0 || processor_group_id < 0) {
                 GROUP_AFFINITY current;
@@ -98,7 +95,7 @@ NIF_API void NIF_DECL nif_set_thread_affinity(int cpu_index, int processor_group
         if ((DWORD)processor_group_id >= GetActiveProcessorGroupCount() || cpu_index >= 64)
                 return;
         USHORT node_id = 0;
-        if (!nif_windows_get_processor_node((WORD)processor_group_id, (BYTE)cpu_index, &node_id))
+        if (!nfn_windows_get_processor_node((WORD)processor_group_id, (BYTE)cpu_index, &node_id))
                 return;
 
         GROUP_AFFINITY affinity;
@@ -124,7 +121,7 @@ NFN_API int NFN_DECL nfn_get_numa_overall(int *out_numa_count, int *count_of_eac
         for (DWORD group = 0; group < group_count; group++) {
                 for (DWORD cpu = 0; cpu < 64; cpu++) {
                         USHORT node_id = 0;
-                        if (nif_windows_get_processor_node((WORD)group, (BYTE)cpu, &node_id) &&
+                        if (nfn_windows_get_processor_node((WORD)group, (BYTE)cpu, &node_id) &&
                             count_of_each_numa != NULL && node_id <= highest)
                                 count_of_each_numa[node_id]++;
                 }
@@ -132,15 +129,11 @@ NFN_API int NFN_DECL nfn_get_numa_overall(int *out_numa_count, int *count_of_eac
         return 1;
 }
 
-<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.windows.c
+NFN_API int NFN_DECL nfn_get_numa_each(int numa_id,
-NFN_API int NFN_DECL nfn_get_numa_each(int numa_id, uint64_t *internal, uint32_t *global)
-=======
-NIF_API int NIF_DECL nif_get_numa_each(int numa_id,
                                        uint32_t *processor_group_ids,
                                        uint32_t *socket_ids,
                                        uint32_t *cpu_indices,
                                        uint32_t *numa_node_ids)
->>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.windows.c
 {
         if (numa_id < 0 || processor_group_ids == NULL || socket_ids == NULL ||
             cpu_indices == NULL || numa_node_ids == NULL)
@@ -148,14 +141,14 @@ NIF_API int NIF_DECL nif_get_numa_each(int numa_id,
 
         uint32_t *socket_map = NULL;
         DWORD group_count = 0;
-        if (!nif_windows_get_socket_map(&socket_map, &group_count))
+        if (!nfn_windows_get_socket_map(&socket_map, &group_count))
                 return 0;
 
         int index = 0;
         for (DWORD group = 0; group < group_count; group++) {
                 for (DWORD cpu = 0; cpu < 64; cpu++) {
                         USHORT node_id = 0;
-                        if (!nif_windows_get_processor_node((WORD)group, (BYTE)cpu, &node_id) ||
+                        if (!nfn_windows_get_processor_node((WORD)group, (BYTE)cpu, &node_id) ||
                             node_id != (USHORT)numa_id)
                                 continue;