perf(mir): 帧槽按大小重排序——小槽优先获得 ldur 范围内偏移

4字节标量优先分配，大数组靠后，使热访问在 [-256,0] 的 ldur 范围。
crypto -176(-10.3%)，消除大量 mov x13,x29 + sub 地址计算链。

src/mir/FrameLowering.cpp

@@ -39,15 +39,35 @@ namespace mir
       }
       callee_saved_bytes = AlignTo(callee_saved_bytes, 16);
 
+      // 将小槽位（频繁访问的标量）排在前面，使偏移在 ldur 范围 [-256,0] 内
+      std::vector<size_t> slot_order(slots.size());
+      for (size_t i = 0; i < slots.size(); ++i)
+        slot_order[i] = i;
+      std::stable_sort(slot_order.begin(), slot_order.end(),
+                       [&](size_t a, size_t b) {
+                         const auto &sa = slots[a];
+                         const auto &sb = slots[b];
+                         if (sa.is_stack_arg != sb.is_stack_arg)
+                           return !sa.is_stack_arg;
+                         if (sa.is_callee_stack_arg != sb.is_callee_stack_arg)
+                           return !sa.is_callee_stack_arg;
+                         return sa.size < sb.size;
+                       });
+
+      std::vector<int> new_offsets(slots.size(), 0);
       int offset = 0;
-      for (auto &slot : slots)
+      for (size_t si : slot_order)
       {
+        auto &slot = slots[si];
         if (slot.is_stack_arg || slot.is_callee_stack_arg)
           continue;
         offset -= slot.size;
         offset = AlignTo(offset, slot.size >= 8 ? 8 : 4);
-        slot.offset = offset;
+        new_offsets[si] = offset;
       }
+      for (size_t i = 0; i < slots.size(); ++i)
+        if (!slots[i].is_stack_arg && !slots[i].is_callee_stack_arg)
+          slots[i].offset = new_offsets[i];
 
       offset -= callee_saved_bytes;
       offset -= 16;