From 55dc3a2dc252b327cf4dbdf117d38888881d04c8 Mon Sep 17 00:00:00 2001
From: hm-thinkbook16p <gxms0904hm@163.com>
Date: Wed, 29 Jul 2026 08:43:05 +0800
Subject: [PATCH] fix(msvc): add MSVC compatibility for __builtin_prefetch and
 __builtin_ia32_pause

Replace GCC builtins with cross-platform macros: VDE_PREFETCH (MSVC _mm_prefetch), VDE_CPU_PAUSE (MSVC _mm_pause)

Also add missing #include <unordered_map> in license_manager.cpp
---
 include/vde/core/concurrent_data.h    | 25 ++++++++++++++-----------
 include/vde/core/performance_tuning.h | 26 ++++++++++++++++++++------
 src/core/license_manager.cpp          |  1 +
 3 files changed, 35 insertions(+), 17 deletions(-)

diff --git a/include/vde/core/concurrent_data.h b/include/vde/core/concurrent_data.h
index fc15f5c..802a04a 100644
--- a/include/vde/core/concurrent_data.h
+++ b/include/vde/core/concurrent_data.h
@@ -24,6 +24,18 @@
 #include <utility>
 #include <type_traits>
 
+// MSVC 兼容层：GCC/Clang __builtin_ia32_pause → MSVC _mm_pause
+#if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))
+#include <intrin.h>
+#define VDE_CPU_PAUSE() _mm_pause()
+#elif defined(__x86_64__) || defined(_M_X64)
+#define VDE_CPU_PAUSE() __builtin_ia32_pause()
+#elif defined(__aarch64__)
+#define VDE_CPU_PAUSE() __asm__ volatile("yield")
+#else
+#define VDE_CPU_PAUSE() ((void)0)
+#endif
+
 namespace vde::core {
 
 // ═══════════════════════════════════════════════════════════════════════════
@@ -486,12 +498,7 @@ public:
             }
             // 自适应退避
             if (spins > 100) {
-                // PAUSE / YIELD
-                #if defined(__x86_64__) || defined(_M_X64)
-                __builtin_ia32_pause();
-                #elif defined(__aarch64__)
-                __asm__ volatile("yield");
-                #endif
+                VDE_CPU_PAUSE();
             }
         }
     }
@@ -512,11 +519,7 @@ public:
                 return;
             }
             if (spins > 100) {
-                #if defined(__x86_64__) || defined(_M_X64)
-                __builtin_ia32_pause();
-                #elif defined(__aarch64__)
-                __asm__ volatile("yield");
-                #endif
+                VDE_CPU_PAUSE();
             }
         }
     }
diff --git a/include/vde/core/performance_tuning.h b/include/vde/core/performance_tuning.h
index 388895d..ffaea99 100644
--- a/include/vde/core/performance_tuning.h
+++ b/include/vde/core/performance_tuning.h
@@ -27,6 +27,20 @@
 #include <cstddef>
 #include <cstdlib>
 
+// MSVC 兼容层：GCC/Clang __builtin_prefetch → MSVC _mm_prefetch
+#if defined(_MSC_VER)
+#include <intrin.h>
+#define VDE_PREFETCH(addr)        _mm_prefetch((const char*)(addr), _MM_HINT_T0)
+#define VDE_PREFETCH_WRITE(addr)  _mm_prefetch((const char*)(addr), _MM_HINT_T0)
+#define VDE_PREFETCH_DIST(addr, dist) \
+    _mm_prefetch((const char*)(addr), (dist) > 1 ? _MM_HINT_T2 : _MM_HINT_T1)
+#else
+#define VDE_PREFETCH(addr)        __builtin_prefetch((addr), 0, 3)
+#define VDE_PREFETCH_WRITE(addr)  __builtin_prefetch((addr), 1, 3)
+#define VDE_PREFETCH_DIST(addr, dist) \
+    __builtin_prefetch((addr), 0, static_cast<int>(dist) > 1 ? 2 : 3)
+#endif
+
 namespace vde::core {
 
 // ═══════════════════════════════════════════════════════════════════════════
@@ -307,7 +321,7 @@ struct CacheOptimizationHints {
 ///
 /// @ingroup core
 inline void prefetch(const void* addr) {
-    __builtin_prefetch(addr, 0, 3);
+    VDE_PREFETCH(addr);
 }
 
 /// 预取写入（使缓存行进入修改状态）
@@ -316,7 +330,7 @@ inline void prefetch(const void* addr) {
 ///
 /// @ingroup core
 inline void prefetch_write(const void* addr) {
-    __builtin_prefetch(addr, 1, 3);
+    VDE_PREFETCH_WRITE(addr);
 }
 
 /// 防止假共享的填充字段
@@ -566,7 +580,7 @@ template<size_t Stride = 1, PrefetchDistance Dist = PrefetchDistance::NEAR>
 inline void prefetch_array(const void* base, size_t index) {
     const char* ptr = static_cast<const char*>(base)
                     + index * Stride * CACHE_LINE_SIZE;
-    __builtin_prefetch(ptr, 0, static_cast<int>(Dist) > 1 ? 2 : 3);
+    VDE_PREFETCH_DIST(ptr, Dist);
 }
 
 /// 预取下一个元素（循环优化）
@@ -578,13 +592,13 @@ inline void prefetch_array(const void* base, size_t index) {
 /// @endcode
 template<size_t Lookahead = 4>
 inline void prefetch_next(const void* addr) {
-    __builtin_prefetch(addr, 0, 3);
+    VDE_PREFETCH(addr);
 }
 
 /// 预取写入（用于 scatter 操作）
 template<size_t Lookahead = 4>
 inline void prefetch_write_next(void* addr) {
-    __builtin_prefetch(addr, 1, 3);
+    VDE_PREFETCH_WRITE(addr);
 }
 
 /// 顺序预取批量数据
@@ -598,7 +612,7 @@ inline void prefetch_range(const void* base, size_t count,
     const char* ptr = static_cast<const char*>(base);
     size_t limit = std::min(count, count);
     for (size_t i = 0; i < limit; i += distance) {
-        __builtin_prefetch(ptr + i * stride, 0, 3);
+        VDE_PREFETCH(ptr + i * stride);
     }
 }
 
diff --git a/src/core/license_manager.cpp b/src/core/license_manager.cpp
index b45a005..ce463d4 100644
--- a/src/core/license_manager.cpp
+++ b/src/core/license_manager.cpp
@@ -20,6 +20,7 @@
 #include <ctime>
 #include <sstream>
 #include <stdexcept>
+#include <unordered_map>
 
 namespace vde::core {