perf(backend): Peephole 新增全局变量 store-load 转发和 load CSE

StoreGlobal 后紧跟 LoadGlobal 同一符号时，同寄存器则删除 load，
否则转为 MovReg。LoadGlobal 连续出现时同样处理。

shuffle -6, conv2d -3, crypto -3, h-9 -3。总计 -15 条，零退化。

src/mir/passes/Peephole.cpp

@@ -93,6 +93,36 @@ namespace mir
              s_ops[0].GetReg() != l_ops[0].GetReg();
     }
 
+    // 全局变量：StoreGlobal 后紧跟 LoadGlobal 同一符号 → 用 MovReg 替代 LoadGlobal
+    static bool IsGlobalFwdStoreLoad(const MachineInstr &a, const MachineInstr &b)
+    {
+      if (a.GetOpcode() != Opcode::StoreGlobal || b.GetOpcode() != Opcode::LoadGlobal)
+        return false;
+      const auto &a_ops = a.GetOperands();
+      const auto &b_ops = b.GetOperands();
+      if (a_ops.size() < 2 || b_ops.size() < 2)
+        return false;
+      if (a_ops[1].GetKind() != Operand::Kind::Symbol ||
+          b_ops[1].GetKind() != Operand::Kind::Symbol)
+        return false;
+      return a_ops[1].GetSymbol() == b_ops[1].GetSymbol();
+    }
+
+    // 全局变量：LoadGlobal 后紧跟 LoadGlobal 同一符号 → 用 MovReg 替代第二个
+    static bool IsGlobalRedundantLoad(const MachineInstr &a, const MachineInstr &b)
+    {
+      if (a.GetOpcode() != Opcode::LoadGlobal || b.GetOpcode() != Opcode::LoadGlobal)
+        return false;
+      const auto &a_ops = a.GetOperands();
+      const auto &b_ops = b.GetOperands();
+      if (a_ops.size() < 2 || b_ops.size() < 2)
+        return false;
+      if (a_ops[1].GetKind() != Operand::Kind::Symbol ||
+          b_ops[1].GetKind() != Operand::Kind::Symbol)
+        return false;
+      return a_ops[1].GetSymbol() == b_ops[1].GetSymbol();
+    }
+
     static bool TryMergeZeroStores(MachineInstr &first, MachineInstr &second)
     {
       if (first.GetOpcode() != Opcode::StoreStack ||
@@ -212,6 +242,57 @@ namespace mir
         }
       }
 
+      // 全局变量 StoreGlobal → LoadGlobal 同一符号转发
+      if (!changed)
+      {
+        for (auto it = insts.begin(); it != insts.end(); ++it)
+        {
+          if (it->GetOpcode() == Opcode::StoreGlobal)
+          {
+            auto next = std::next(it);
+            if (next != insts.end() && IsGlobalFwdStoreLoad(*it, *next))
+            {
+              const auto &s_ops = it->GetOperands();
+              const auto &l_ops = next->GetOperands();
+              // 若已是同一寄存器则直接删除 load，否则用 MovReg 替代
+              if (s_ops[0].GetKind() == l_ops[0].GetKind() &&
+                  s_ops[0].GetKind() == Operand::Kind::Reg &&
+                  s_ops[0].GetReg() == l_ops[0].GetReg())
+                next = insts.erase(next);
+              else
+                *next = MachineInstr(Opcode::MovReg, {l_ops[0], s_ops[0]});
+              changed = true;
+              break;
+            }
+          }
+        }
+      }
+
+      // 全局变量 LoadGlobal → LoadGlobal 同一符号消除
+      if (!changed)
+      {
+        for (auto it = insts.begin(); it != insts.end(); ++it)
+        {
+          if (it->GetOpcode() == Opcode::LoadGlobal)
+          {
+            auto next = std::next(it);
+            if (next != insts.end() && IsGlobalRedundantLoad(*it, *next))
+            {
+              const auto &first_ops = it->GetOperands();
+              const auto &second_ops = next->GetOperands();
+              if (first_ops[0].GetKind() == second_ops[0].GetKind() &&
+                  first_ops[0].GetKind() == Operand::Kind::Reg &&
+                  first_ops[0].GetReg() == second_ops[0].GetReg())
+                next = insts.erase(next);
+              else
+                *next = MachineInstr(Opcode::MovReg, {second_ops[0], first_ops[0]});
+              changed = true;
+              break;
+            }
+          }
+        }
+      }
+
       // 分支 fallthrough: 末尾 Br 的目标是紧邻下一个块 → 删除 Br
       if (!insts.empty())
       {

优化记录.md

@@ -96,3 +96,15 @@
 - **退化**：matmul +3（+0.3%），寄存器分配差异，在容忍范围内
 - **功能测试**：87/88 functional 通过（1 个不稳定故障 87_many_params），30/31 h_functional 通过（1 个预存故障 30_many_dimensions）
 - **已知局限**：sdiv 在 Cortex-A53 上延迟较高（4-12 周期），但 QEMU 不精确模拟流水线，且指令数减少足以弥补
+
+---
+
+## 2026-05-25 | 全局变量 Peephole 优化
+
+- **类型**：后端（MIR Peephole）
+- **假设**：同一基本块内，StoreGlobal 后紧跟 LoadGlobal 同一符号时可转发存储值（或相同寄存器则直接消除）；LoadGlobal 后紧跟 LoadGlobal 同一符号时可复用第一次加载的值
+- **实现**：Peephole.cpp 新增 IsGlobalFwdStoreLoad/IsGlobalRedundantLoad 检测函数，RunPeepholeOnBlock 新增两个迭代 pass
+- **指令数效果**：减少 15 条，shuffle -6、conv2d -3、crypto -3、h-9 -3
+- **退化**：无（matmul +3 是之前 sdiv 优化的残留退化）
+- **功能测试**：87/88 functional 通过（1 个不稳定故障 87_many_params）
+- **已知局限**：仅处理同寄存器复用的特例；不同寄存器间的转发/复用转为 MovReg（指令数不减少）

指令数基线.md

@@ -26,15 +26,15 @@
 | performance/03_sort1 | 625 |
 | performance/03_sort2 | 625 |
 | performance/03_sort3 | 625 |
-| performance/conv2d-1 | 627 |
-| performance/conv2d-2 | 627 |
-| performance/conv2d-3 | 627 |
+| performance/conv2d-1 | 626 |
+| performance/conv2d-2 | 626 |
+| performance/conv2d-3 | 626 |
 | performance/crc1 | 242 |
 | performance/crc2 | 242 |
 | performance/crc3 | 242 |
-| performance/crypto-1 | 1809 |
-| performance/crypto-2 | 1809 |
-| performance/crypto-3 | 1809 |
+| performance/crypto-1 | 1808 |
+| performance/crypto-2 | 1808 |
+| performance/crypto-3 | 1808 |
 | performance/fft0 | 564 |
 | performance/fft1 | 564 |
 | performance/fft2 | 564 |
@@ -53,9 +53,9 @@
 | performance/h-8-01 | 410 |
 | performance/h-8-02 | 410 |
 | performance/h-8-03 | 410 |
-| performance/h-9-01 | 198 |
-| performance/h-9-02 | 198 |
-| performance/h-9-03 | 198 |
+| performance/h-9-01 | 197 |
+| performance/h-9-02 | 197 |
+| performance/h-9-03 | 197 |
 | performance/huffman-01 | 694 |
 | performance/huffman-02 | 694 |
 | performance/huffman-03 | 694 |
@@ -71,9 +71,9 @@
 | performance/optimization_scheduling1 | 110 |
 | performance/optimization_scheduling2 | 110 |
 | performance/optimization_scheduling3 | 110 |
-| performance/shuffle0 | 452 |
-| performance/shuffle1 | 452 |
-| performance/shuffle2 | 452 |
+| performance/shuffle0 | 450 |
+| performance/shuffle1 | 450 |
+| performance/shuffle2 | 450 |
 | performance/sl1 | 246 |
 | performance/sl2 | 246 |
 | performance/sl3 | 246 |