From 422f1848fc7360d7bcaf32d6a316f1079328731a Mon Sep 17 00:00:00 2001
From: lzkk <956449176@qq.com>
Date: Thu, 28 May 2026 14:26:21 +0800
Subject: [PATCH] =?UTF-8?q?perf(mir):=20Lowering=20=E7=9B=B4=E6=8E=A5?=
 =?UTF-8?q?=E7=94=9F=E6=88=90=20AddShiftRR=E2=80=94=E2=80=94x+(x*2^n)=20?=
 =?UTF-8?q?=E2=86=92=20add=20x,x,lsl#n?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

在 Lowering 阶段检测 Add 的某个操作数是 Mul by 幂次方常量，
直接发射 AddShiftRR 而非分开发射 ShlRR+AddRR。
配合 Peephole 残余合并，共生成 5 处 add lsl。
---
 src/mir/Lowering.cpp | 46 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 46 insertions(+)

diff --git a/src/mir/Lowering.cpp b/src/mir/Lowering.cpp
index 4bcb494d..a6ea0660 100644
--- a/src/mir/Lowering.cpp
+++ b/src/mir/Lowering.cpp
@@ -410,6 +410,52 @@ namespace mir
 
         int dst = function.CreateVReg(VRegClass::Int);
 
+        // AddShift 折叠：x + (x * 2^n) → add x, x, lsl #n
+        if (opcode == Opcode::AddRR)
+        {
+          auto *mul_rhs2 = dynamic_cast<const ir::BinaryInst *>(bin->GetRhs());
+          if (mul_rhs2 && mul_rhs2->GetOpcode() == ir::Opcode::Mul)
+          {
+            int shift_val = 0;
+            if (TryGetConstantInt(mul_rhs2->GetRhs(), shift_val) &&
+                shift_val > 0 && (shift_val & (shift_val - 1)) == 0)
+            {
+              // rhs is x * 2^n, lhs should match x
+              int mul_lhs = EmitIntValue(mul_rhs2->GetLhs(), function, value_vregs,
+                                         scalar_slots, array_slots, block);
+              int add_lhs = EmitIntValue(bin->GetLhs(), function, value_vregs,
+                                         scalar_slots, array_slots, block);
+              int sh = 0; while (shift_val > 1) { shift_val >>= 1; ++sh; }
+              block.Append(Opcode::AddShiftRR,
+                           {Operand::VReg(dst, VRegClass::Int),
+                            Operand::VReg(add_lhs, VRegClass::Int),
+                            Operand::VReg(mul_lhs, VRegClass::Int),
+                            Operand::Imm(sh)});
+              value_vregs[value] = dst;
+              value_vregs[mul_rhs2] = dst;
+              return dst;
+            }
+            if (TryGetConstantInt(mul_rhs2->GetLhs(), shift_val) &&
+                shift_val > 0 && (shift_val & (shift_val - 1)) == 0)
+            {
+              // rhs is 2^n * x, lhs should match x
+              int mul_rhs = EmitIntValue(mul_rhs2->GetRhs(), function, value_vregs,
+                                         scalar_slots, array_slots, block);
+              int add_lhs = EmitIntValue(bin->GetLhs(), function, value_vregs,
+                                         scalar_slots, array_slots, block);
+              int sh = 0; while (shift_val > 1) { shift_val >>= 1; ++sh; }
+              block.Append(Opcode::AddShiftRR,
+                           {Operand::VReg(dst, VRegClass::Int),
+                            Operand::VReg(add_lhs, VRegClass::Int),
+                            Operand::VReg(mul_rhs, VRegClass::Int),
+                            Operand::Imm(sh)});
+              value_vregs[value] = dst;
+              value_vregs[mul_rhs2] = dst;
+              return dst;
+            }
+          }
+        }
+
         // Madd 折叠：sum + (a * b) → madd sum, a, b, sum（必须在 EmitIntValue 之前）
         if (opcode == Opcode::AddRR)
         {