perf(backend): 叶函数跳过帧设置，节省 x29/x30 保存/恢复

MachineFunction 添加 HasCall 标记，Lowering 在发射 Call 时设置。
叶函数无帧且无 callee-saved 寄存器的函数完全跳过 prologue/epilogue；
有帧叶函数改用 str/ldr x29 替代 stp/ldp x29,x30。

huffman -93, crypto -54, conv2d -45, crc -27, h-9 -27,
03_sort -18, opt_scheduling -18, h-4 -12, fft -9, shuffle -9。
总计 -312 条，零退化。

src/include/mir/MIR.h

@@ -333,6 +333,9 @@ namespace mir
     int GetFrameSize() const { return frame_size_; }
     void SetFrameSize(int size) { frame_size_ = size; }
 
+    bool HasCall() const { return has_call_; }
+    void SetHasCall(bool v = true) { has_call_ = v; }
+
     int CreateVReg(VRegClass vreg_class);
     VRegClass GetVRegClass(int vreg_id) const;
     int GetNumVRegs() const { return static_cast<int>(vreg_classes_.size()); }
@@ -347,6 +350,7 @@ namespace mir
 
     std::vector<FrameSlot> frame_slots_;
     int frame_size_ = 0;
+    bool has_call_ = false;
     int next_label_id_ = 0;
 
     std::vector<VRegClass> vreg_classes_;

src/mir/AsmPrinter.cpp

@@ -456,12 +456,31 @@ namespace mir
             case Opcode::Prologue:
             {
                 const auto &cs_regs = function.GetCalleeSavedRegs();
-                os << "    stp x29, x30, [sp, #-16]!\n";
+                const bool is_leaf = !function.HasCall();
+                const bool no_frame = (function.GetFrameSize() == 0 && cs_regs.empty());
+
+                // 叶函数无帧且无 callee-saved 寄存器：完全跳过帧设置
+                if (is_leaf && no_frame)
+                {
+                    return;
+                }
+
+                // 叶函数仅保存 x29（LR 不会被修改），非叶函数保存 x29+x30
+                if (is_leaf)
+                {
+                    os << "    str x29, [sp, #-8]!\n";
+                }
+                else
+                {
+                    os << "    stp x29, x30, [sp, #-16]!\n";
+                }
                 os << "    mov x29, sp\n";
+
                 if (function.GetFrameSize() > 0)
                 {
                     EmitStackAdjust("sub", function.GetFrameSize(), os);
                 }
+
                 // X(64-bit) 和 S(32-bit) 分两组配对 stp
                 std::vector<PhysReg> x_regs, s_regs;
                 for (auto r : cs_regs)
@@ -505,6 +524,17 @@ namespace mir
             case Opcode::Epilogue:
             {
                 const auto &cs_regs = function.GetCalleeSavedRegs();
+                const bool is_leaf = !function.HasCall();
+                const bool no_frame = (function.GetFrameSize() == 0 && cs_regs.empty());
+
+                // 叶函数无帧且无 callee-saved 寄存器——直接返回
+                if (is_leaf && no_frame)
+                {
+                    os << "    ret\n";
+                    return;
+                }
+
+                // 恢复 callee-saved 寄存器（叶函数也需要——它们属于调用者）
                 std::vector<PhysReg> x_regs, s_regs;
                 for (auto r : cs_regs)
                 {
@@ -513,7 +543,7 @@ namespace mir
                     else if (r >= PhysReg::S0 && r <= PhysReg::S31)
                         s_regs.push_back(r);
                     else
-                        x_regs.push_back(r); // 兜底：非 X 非 S 按 X 处理
+                        x_regs.push_back(r);
                 }
                 int cs_offset = 0;
                 for (size_t i = 0; i + 1 < x_regs.size(); i += 2)
@@ -541,11 +571,20 @@ namespace mir
                     os << "    ldr " << PhysRegName(s_regs.back())
                        << ", [sp, #" << cs_offset << "]\n";
                 }
+
                 if (function.GetFrameSize() > 0)
                 {
                     EmitStackAdjust("add", function.GetFrameSize(), os);
                 }
-                os << "    ldp x29, x30, [sp], #16\n";
+
+                if (is_leaf)
+                {
+                    os << "    ldr x29, [sp], #8\n";
+                }
+                else
+                {
+                    os << "    ldp x29, x30, [sp], #16\n";
+                }
                 os << "    ret\n";
                 return;
             }

src/mir/Lowering.cpp

@@ -1764,6 +1764,7 @@ namespace mir
         }
 
         block.Append(Opcode::Call, {Operand::Symbol(callee->GetName())});
+        function.SetHasCall();
 
         if (aligned_stack_arg_bytes > 0)
         {

优化记录.md

@@ -69,3 +69,18 @@
 - **退化**：无
 - **功能测试**：100/100 functional 通过，30/31 h_functional 通过（1 个预存故障 30_many_dimensions）
 - **已知局限**：仅缓存 x13 上的 ADRP；LoadGlobalAddr 使用其他寄存器时不参与缓存；同一基本块内优化最有效
+
+---
+
+## 2026-05-25 | 叶函数帧设置优化
+
+- **类型**：后端（AsmPrinter + Lowering）
+- **假设**：叶函数（无 Call 指令）不需要保存/恢复 x30（LR 不会被修改）。无帧且无 callee-saved 寄存器的叶函数可完全跳过帧设置（stp/ldp x29,x30 + mov x29,sp），节省 3 条指令。有帧叶函数改用 str/ldr x29 替代 stp/ldp x29,x30，节省栈空间
+- **实现**：
+  - MIR.h：MachineFunction 新增 has_call_ 字段 + HasCall()/SetHasCall()
+  - Lowering.cpp：每次发射 Call 指令时标记 function.SetHasCall()
+  - AsmPrinter.cpp：Prologue/Epilogue 根据 is_leaf 和 no_frame 条件跳过或简化帧设置
+- **指令数效果**：减少 312 条，huffman -93（-3.9%）、crypto -54（-2.8%）、conv2d -45（-2.3%）、crc -27（-3.2%）、h-9 -27（-4.1%）、03_sort -18（-0.9%）、opt_scheduling -18（-5.2%）、h-4 -12（-2.5%）、fft -9（-0.5%）、shuffle -9（-0.7%）
+- **退化**：无
+- **功能测试**：100/100 functional 通过，30/31 h_functional 通过（1 个预存故障 30_many_dimensions）
+- **已知局限**：仅对无帧且无 callee-saved 寄存器的叶函数完全跳过帧设置；有 callee-saved 的叶函数仍需保存它们（属于调用者）

指令数基线.md

@@ -23,42 +23,42 @@
 | performance/01_mm1 | 309 |
 | performance/01_mm2 | 309 |
 | performance/01_mm3 | 309 |
-| performance/03_sort1 | 639 |
-| performance/03_sort2 | 639 |
-| performance/03_sort3 | 639 |
-| performance/conv2d-1 | 649 |
-| performance/conv2d-2 | 649 |
-| performance/conv2d-3 | 649 |
-| performance/crc1 | 279 |
-| performance/crc2 | 279 |
-| performance/crc3 | 279 |
-| performance/crypto-1 | 1910 |
-| performance/crypto-2 | 1910 |
-| performance/crypto-3 | 1910 |
-| performance/fft0 | 591 |
-| performance/fft1 | 591 |
-| performance/fft2 | 591 |
+| performance/03_sort1 | 633 |
+| performance/03_sort2 | 633 |
+| performance/03_sort3 | 633 |
+| performance/conv2d-1 | 634 |
+| performance/conv2d-2 | 634 |
+| performance/conv2d-3 | 634 |
+| performance/crc1 | 270 |
+| performance/crc2 | 270 |
+| performance/crc3 | 270 |
+| performance/crypto-1 | 1892 |
+| performance/crypto-2 | 1892 |
+| performance/crypto-3 | 1892 |
+| performance/fft0 | 588 |
+| performance/fft1 | 588 |
+| performance/fft2 | 588 |
 | performance/h-1-01 | 157 |
 | performance/h-1-02 | 157 |
 | performance/h-1-03 | 157 |
 | performance/h-10-01 | 327 |
 | performance/h-10-02 | 327 |
 | performance/h-10-03 | 327 |
-| performance/h-4-01 | 162 |
-| performance/h-4-02 | 162 |
-| performance/h-4-03 | 162 |
+| performance/h-4-01 | 158 |
+| performance/h-4-02 | 158 |
+| performance/h-4-03 | 158 |
 | performance/h-5-01 | 338 |
 | performance/h-5-02 | 338 |
 | performance/h-5-03 | 338 |
 | performance/h-8-01 | 410 |
 | performance/h-8-02 | 410 |
 | performance/h-8-03 | 410 |
-| performance/h-9-01 | 221 |
-| performance/h-9-02 | 221 |
-| performance/h-9-03 | 221 |
-| performance/huffman-01 | 787 |
-| performance/huffman-02 | 787 |
-| performance/huffman-03 | 787 |
+| performance/h-9-01 | 212 |
+| performance/h-9-02 | 212 |
+| performance/h-9-03 | 212 |
+| performance/huffman-01 | 756 |
+| performance/huffman-02 | 756 |
+| performance/huffman-03 | 756 |
 | performance/knapsack_naive-1 | 167 |
 | performance/knapsack_naive-2 | 167 |
 | performance/knapsack_naive-3 | 167 |
@@ -68,12 +68,12 @@
 | performance/matmul1 | 379 |
 | performance/matmul2 | 379 |
 | performance/matmul3 | 379 |
-| performance/optimization_scheduling1 | 116 |
-| performance/optimization_scheduling2 | 116 |
-| performance/optimization_scheduling3 | 116 |
-| performance/shuffle0 | 455 |
-| performance/shuffle1 | 455 |
-| performance/shuffle2 | 455 |
+| performance/optimization_scheduling1 | 110 |
+| performance/optimization_scheduling2 | 110 |
+| performance/optimization_scheduling3 | 110 |
+| performance/shuffle0 | 452 |
+| performance/shuffle1 | 452 |
+| performance/shuffle2 | 452 |
 | performance/sl1 | 247 |
 | performance/sl2 | 247 |
 | performance/sl3 | 247 |