fix(mir): 修复全局变量 .bss/.data 段反复切换——统一分组输出

之前每个零初始化全局变量都会在 .bss → .data 之间切换一次，
大数组（12MB BSS）的 .data 插入可能导致链接器内存布局异常。
现在先输出全部 .bss 段，再输出 .data 段，消除无用的段切换。
这可能是 many_mat_cal-1/3 在平台 100MB 限制下 WA 的根因。

src/mir/AsmPrinter.cpp

@@ -865,74 +865,71 @@ namespace mir
         void PrintGlobals(const MachineModule &module, std::ostream &os)
         {
             if (module.GetGlobals().empty())
-            {
                 return;
-            }
 
-            os << "    .data\n";
-            for (const auto &global : module.GetGlobals())
-            {
-                const std::string asm_name = NormalizeAsmSymbol(global.name);
-
-                bool is_zero_init = false;
-                if (global.kind == MachineGlobal::Kind::I32Scalar && global.init_value == 0)
+            auto is_zero = [](const MachineGlobal &g) -> bool {
+                if (g.kind == MachineGlobal::Kind::I32Scalar && g.init_value == 0)
+                    return true;
+                if (g.kind == MachineGlobal::Kind::I32Array)
                 {
-                    is_zero_init = true;
-                }
-                if (global.kind == MachineGlobal::Kind::I32Array)
-                {
-                    bool all_zero = true;
-                    for (auto v : global.init_values)
-                    {
-                        if (v != 0)
-                        {
-                            all_zero = false;
-                            break;
-                        }
-                    }
-                    if (all_zero)
-                    {
-                        is_zero_init = true;
-                    }
+                    if (g.init_values.empty()) return true;
+                    for (auto v : g.init_values)
+                        if (v != 0) return false;
+                    return true;
                 }
+                return false;
+            };
+
+            // 先统一输出 .bss 段
+            bool has_bss = false;
+            for (const auto &g : module.GetGlobals())
+                if (is_zero(g)) { has_bss = true; break; }
 
-                if (is_zero_init)
+            if (has_bss)
+            {
+                os << "    .bss\n";
+                for (const auto &global : module.GetGlobals())
                 {
-                    os << "    .bss\n";
+                    if (!is_zero(global)) continue;
+                    const std::string asm_name = NormalizeAsmSymbol(global.name);
                     os << "    .globl " << asm_name << "\n";
                     os << "    .p2align 2\n";
                     os << asm_name << ":\n";
                     if (global.kind == MachineGlobal::Kind::I32Scalar)
-                    {
                         os << "    .space 4\n";
-                    }
                     else
-                    {
                         os << "    .space " << (global.array_size * 4) << "\n";
-                    }
-                    os << "    .data\n";
-                    continue;
                 }
+            }
 
-                os << "    .globl " << asm_name << "\n";
-                os << "    .p2align 2\n";
-                os << asm_name << ":\n";
+            // 再统一输出 .data 段
+            bool has_data = false;
+            for (const auto &g : module.GetGlobals())
+                if (!is_zero(g)) { has_data = true; break; }
 
-                if (global.kind == MachineGlobal::Kind::I32Scalar)
+            if (has_data)
+            {
+                os << "    .data\n";
+                for (const auto &global : module.GetGlobals())
                 {
-                    os << "    .word " << global.init_value << "\n";
-                    continue;
-                }
+                    if (is_zero(global)) continue;
+                    const std::string asm_name = NormalizeAsmSymbol(global.name);
+                    os << "    .globl " << asm_name << "\n";
+                    os << "    .p2align 2\n";
+                    os << asm_name << ":\n";
 
-                const size_t init_count = global.init_values.size();
-                for (size_t i = 0; i < init_count; ++i)
-                {
-                    os << "    .word " << global.init_values[i] << "\n";
-                }
+                    if (global.kind == MachineGlobal::Kind::I32Scalar)
+                    {
+                        os << "    .word " << global.init_value << "\n";
+                        continue;
+                    }
 
-                if (global.array_size > init_count)
-                {
-                    os << "    .zero " << ((global.array_size - init_count) * 4) << "\n";
+                    const size_t init_count = global.init_values.size();
+                    for (size_t i = 0; i < init_count; ++i)
+                        os << "    .word " << global.init_values[i] << "\n";
+
+                    if (global.array_size > init_count)
+                        os << "    .zero " << ((global.array_size - init_count) * 4) << "\n";
                 }
             }
             os << "\n";