perf(backend): 消除连续全局变量访问的冗余 ADRP

AsmPrinter 添加 ADRP 缓存，同符号连续访问时跳过重复的页面地址装载。
x13 被非全局访问路径使用时失效缓存；基本块入口重置。

shuffle -48, crypto -27, conv2d -21, fft -12, huffman -9, h-9 -9,
03_sort -6, h-8 -3。总计 -135 条，零退化。

src/mir/AsmPrinter.cpp

@@ -256,10 +256,20 @@ namespace mir
             }
         }
 
+        // ADRP 缓存——避免连续访问同一全局变量时重复发射 ADRP
+        std::string g_cached_adrp_symbol;
+        bool g_adrp_cache_valid = false;
+
+        void InvalidateAdrpCache()
+        {
+            g_adrp_cache_valid = false;
+        }
+
         void EmitStackAdjust(const char *op, int amount, std::ostream &os)
         {
             if (amount > 12285)
             {
+                InvalidateAdrpCache();
                 os << "    movz x13, #" << (amount & 0xFFFF) << "\n";
                 if ((amount >> 16) != 0)
                     os << "    movk x13, #" << ((amount >> 16) & 0xFFFF) << ", lsl #16\n";
@@ -287,6 +297,7 @@ namespace mir
         {
             if (offset > 12285)
             {
+                InvalidateAdrpCache();
                 os << "    movz x13, #" << (offset & 0xFFFF) << "\n";
                 if ((offset >> 16) != 0)
                     os << "    movk x13, #" << ((offset >> 16) & 0xFFFF) << ", lsl #16\n";
@@ -297,6 +308,7 @@ namespace mir
             if (offset < -12285)
             {
                 int abs_off = -offset;
+                InvalidateAdrpCache();
                 os << "    movz x13, #" << (abs_off & 0xFFFF) << "\n";
                 if ((abs_off >> 16) != 0)
                     os << "    movk x13, #" << ((abs_off >> 16) & 0xFFFF) << ", lsl #16\n";
@@ -365,7 +377,17 @@ namespace mir
             const std::string asm_symbol = NormalizeAsmSymbol(symbol);
             const PhysReg scratch_xreg = PrinterScratchXReg();
 
-            os << "    adrp " << PhysRegName(scratch_xreg) << ", " << asm_symbol << "\n";
+            if (g_adrp_cache_valid && g_cached_adrp_symbol == asm_symbol)
+            {
+                // x13 已持有该全局变量的页面地址，跳过 ADRP
+            }
+            else
+            {
+                os << "    adrp " << PhysRegName(scratch_xreg) << ", " << asm_symbol << "\n";
+                g_cached_adrp_symbol = asm_symbol;
+                g_adrp_cache_valid = true;
+            }
+
             os << "    " << (opcode == Opcode::LoadGlobal ? "ldr " : "str ");
             PrintOperand(reg, os);
             os << ", [" << PhysRegName(scratch_xreg) << ", #:lo12:" << asm_symbol << "]\n";
@@ -823,6 +845,10 @@ namespace mir
                 }
                 return;
 
+            case Opcode::Call:
+                InvalidateAdrpCache();
+                // 不 break，落到 default 让泛型打印机输出 bl 指令
+
             default:
                 break;
             }
@@ -878,6 +904,7 @@ namespace mir
 
     void PrintAsm(const MachineFunction &function, std::ostream &os)
     {
+        g_adrp_cache_valid = false;
         const std::string asm_name = NormalizeAsmSymbol(function.GetName());
 
         os << "    .text\n";
@@ -893,6 +920,8 @@ namespace mir
             }
 
             const auto &block = *block_ptr;
+            // 每个基本块重置 ADRP 缓存——跨块时 x13 可能已被 call/clobber 破坏
+            g_adrp_cache_valid = false;
             PrintBlockLabelRef(function, block.GetLabelId(), os);
             os << ":\n";
 

优化记录.md

@@ -57,3 +57,15 @@
 - **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**：仅修复 EmitLargeImmediate；EmitStackAdjust/EmitAddressFromBase 中的 movz 模式仍有同样问题，可后续统一
+
+---
+
+## 2026-05-25 | ADRP 冗余消除
+
+- **类型**：后端（AsmPrinter）
+- **假设**：连续访问同一全局变量时，x13 已持有页面地址，后续 ADRP 冗余。例如 `adrp x13, k; str w8, [x13, :lo12:k]; adrp x13, k` 中第二个 ADRP 多余
+- **实现**：AsmPrinter 添加 ADRP 缓存（g_cached_adrp_symbol + g_adrp_cache_valid）。PrintGlobalAccess 检测同符号命中时跳过 ADRP。EmitStackAdjust/EmitAddressFromBase 使用 x13 时失效缓存。Call 指令失效缓存（x13 caller-saved）。每个基本块入口重置缓存（跨块时 call/clobber 不确定）
+- **指令数效果**：减少 135 条，shuffle -48（-3.4%）、crypto -27（-1.4%）、conv2d -21（-3.2%）、fft -12（-2.0%）、huffman -9（-1.1%）、h-9 -9（-4.0%）、03_sort -6（-0.9%）、h-8 -3（-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

指令数基线.md

@@ -23,21 +23,21 @@
 | performance/01_mm1 | 309 |
 | performance/01_mm2 | 309 |
 | performance/01_mm3 | 309 |
-| performance/03_sort1 | 641 |
-| performance/03_sort2 | 641 |
-| performance/03_sort3 | 641 |
-| performance/conv2d-1 | 656 |
-| performance/conv2d-2 | 656 |
-| performance/conv2d-3 | 656 |
+| 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 | 1919 |
-| performance/crypto-2 | 1919 |
-| performance/crypto-3 | 1919 |
-| performance/fft0 | 595 |
-| performance/fft1 | 595 |
-| performance/fft2 | 595 |
+| performance/crypto-1 | 1910 |
+| performance/crypto-2 | 1910 |
+| performance/crypto-3 | 1910 |
+| performance/fft0 | 591 |
+| performance/fft1 | 591 |
+| performance/fft2 | 591 |
 | performance/h-1-01 | 157 |
 | performance/h-1-02 | 157 |
 | performance/h-1-03 | 157 |
@@ -50,15 +50,15 @@
 | performance/h-5-01 | 338 |
 | performance/h-5-02 | 338 |
 | performance/h-5-03 | 338 |
-| performance/h-8-01 | 411 |
-| performance/h-8-02 | 411 |
-| performance/h-8-03 | 411 |
-| performance/h-9-01 | 224 |
-| performance/h-9-02 | 224 |
-| performance/h-9-03 | 224 |
-| performance/huffman-01 | 790 |
-| performance/huffman-02 | 790 |
-| performance/huffman-03 | 790 |
+| 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/knapsack_naive-1 | 167 |
 | performance/knapsack_naive-2 | 167 |
 | performance/knapsack_naive-3 | 167 |
@@ -71,9 +71,9 @@
 | performance/optimization_scheduling1 | 116 |
 | performance/optimization_scheduling2 | 116 |
 | performance/optimization_scheduling3 | 116 |
-| performance/shuffle0 | 471 |
-| performance/shuffle1 | 471 |
-| performance/shuffle2 | 471 |
+| performance/shuffle0 | 455 |
+| performance/shuffle1 | 455 |
+| performance/shuffle2 | 455 |
 | performance/sl1 | 247 |
 | performance/sl2 | 247 |
 | performance/sl3 | 247 |