perf(mir): MIR预分配 MovImm 转发——消除常数二次传递

mov v1,#N; mov v2,v1（v1 无其他使用）→ mov v2,#N
matmul -3.3%, optim_sched -10.2%, 01_mm -4.4%。33基线更新。

src/include/mir/MIR.h

@@ -421,6 +421,9 @@ namespace mir
   void RunBlockLayout(MachineFunction &function);
   void RunBlockLayout(MachineModule &module);
 
+  void RunMIRCleanup(MachineFunction &function);
+  void RunMIRCleanup(MachineModule &module);
+
   void PrintAsm(const MachineFunction &function, std::ostream &os);
   void PrintAsm(const MachineModule &module, std::ostream &os);
 

src/main.cpp

@@ -51,6 +51,7 @@ int main(int argc, char** argv) {
     // 汇编输出到文件或标准输出
     if (opts.emit_asm) {
       auto machine_module = mir::LowerModuleToMIR(*module);
+      mir::RunMIRCleanup(*machine_module);
       mir::RunRegAlloc(*machine_module);
       mir::RunFrameLowering(*machine_module);
       mir::RunBlockLayout(*machine_module);

src/mir/passes/CMakeLists.txt

@@ -2,6 +2,7 @@ add_library(mir_passes STATIC
   PassManager.cpp
   Peephole.cpp
   BlockLayoutOpt.cpp
+  MIRCleanup.cpp
 )
 
 target_link_libraries(mir_passes PUBLIC

src/mir/passes/MIRCleanup.cpp

@@ -0,0 +1,100 @@
+#include "mir/MIR.h"
+
+#include "utils/Log.h"
+
+namespace mir
+{
+namespace
+{
+
+// MovImm 转发：mov v1, #N; mov v2, v1（v1 无其他使用）→ mov v2, #N
+static void ForwardMovImm(MachineFunction &function)
+{
+  for (auto &block : function.GetBlocks())
+  {
+    if (!block)
+      continue;
+    auto &insts = block->GetInstructions();
+
+    for (auto it = insts.begin(); it != insts.end(); ++it)
+    {
+      if (it->GetOpcode() != Opcode::MovImm)
+        continue;
+      const auto &ops = it->GetOperands();
+      if (ops.size() < 2 || ops[0].GetKind() != Operand::Kind::VReg)
+        continue;
+
+      int src_vreg = ops[0].GetVRegId();
+      int imm_val = ops[1].GetImm();
+
+      auto next = std::next(it);
+      if (next == insts.end())
+        continue;
+      if (next->GetOpcode() != Opcode::MovReg)
+        continue;
+
+      const auto &n_ops = next->GetOperands();
+      if (n_ops.size() < 2)
+        continue;
+      if (n_ops[1].GetKind() != Operand::Kind::VReg ||
+          n_ops[1].GetVRegId() != src_vreg)
+        continue;
+      if (n_ops[0].GetKind() != Operand::Kind::VReg)
+        continue;
+
+      int dst_vreg = n_ops[0].GetVRegId();
+
+      // 检查 src_vreg 是否还有其他使用
+      bool other_use = false;
+      for (auto &b2 : function.GetBlocks())
+      {
+        if (!b2)
+          continue;
+        for (auto &inst2 : b2->GetInstructions())
+        {
+          for (const auto &op : inst2.GetOperands())
+          {
+            if (op.GetKind() == Operand::Kind::VReg &&
+                op.GetVRegId() == src_vreg)
+            {
+              // 排除 MovImm 自身（def）和 MovReg（use）
+              MachineInstr *mi_ptr = const_cast<MachineInstr *>(&inst2);
+              if (mi_ptr != &(*it) && mi_ptr != &(*next))
+              {
+                other_use = true;
+                goto done_check;
+              }
+            }
+          }
+        }
+      }
+    done_check:
+
+      if (!other_use)
+      {
+        *next = MachineInstr(Opcode::MovImm,
+                             {Operand::VReg(dst_vreg, function.GetVRegClass(dst_vreg)),
+                              Operand::Imm(imm_val)});
+        it = insts.erase(it);
+      }
+    }
+  }
+}
+
+} // namespace
+
+void RunMIRCleanup(MachineFunction &function)
+{
+  ForwardMovImm(function);
+}
+
+void RunMIRCleanup(MachineModule &module)
+{
+  for (auto &function : module.GetFunctions())
+  {
+    if (function)
+      RunMIRCleanup(*function);
+  }
+}
+
+} // namespace mir