perf(mir): 模2^n运算优化——AND+符号修正替代 Msub

x % 2^n: (x & (2^n-1)) + cmp + csel 替代 asr+msub
huffman -20, crypto -33, crc -13, h-9 -7

src/mir/Lowering.cpp

@@ -546,56 +546,54 @@ namespace mir
             int val = rhs_const->GetValue();
             if (val > 0 && (val & (val - 1)) == 0)
             {
-              int bias = val - 1;
-              int biased = function.CreateVReg(VRegClass::Int);
-              if (bias <= 4095)
+              // x % 2^n → and + sign fixup: (x & mask) then if x<0 subtract 2^n
+              int mask = val - 1;
+              int masked = function.CreateVReg(VRegClass::Int);
+              if (mask <= 4095)
               {
-                block.Append(Opcode::AddRR,
-                             {Operand::VReg(biased, VRegClass::Int),
+                block.Append(Opcode::AndRR,
+                             {Operand::VReg(masked, VRegClass::Int),
                               Operand::VReg(lhs, VRegClass::Int),
-                              Operand::Imm(bias)});
+                              Operand::Imm(mask)});
               }
               else
               {
-                int bias_reg = function.CreateVReg(VRegClass::Int);
+                int mask_reg = function.CreateVReg(VRegClass::Int);
                 block.Append(Opcode::MovImm,
-                             {Operand::VReg(bias_reg, VRegClass::Int),
-                              Operand::Imm(bias)}).SetRematerializable(true).SetRematImm(bias);
-                block.Append(Opcode::AddRR,
-                             {Operand::VReg(biased, VRegClass::Int),
+                             {Operand::VReg(mask_reg, VRegClass::Int),
+                              Operand::Imm(mask)}).SetRematerializable(true).SetRematImm(mask);
+                block.Append(Opcode::AndRR,
+                             {Operand::VReg(masked, VRegClass::Int),
                               Operand::VReg(lhs, VRegClass::Int),
-                              Operand::VReg(bias_reg, VRegClass::Int)});
-              }
-              int shift = 0;
-              int tmp = val;
-              while (tmp > 1)
-              {
-                tmp >>= 1;
-                ++shift;
+                              Operand::VReg(mask_reg, VRegClass::Int)});
               }
               block.Append(Opcode::CmpImm,
                            {Operand::VReg(lhs, VRegClass::Int),
                             Operand::Imm(0)});
-              int selected = function.CreateVReg(VRegClass::Int);
+              int neg_fixup = function.CreateVReg(VRegClass::Int);
+              if (static_cast<unsigned int>(val) <= 4095)
+              {
+                block.Append(Opcode::SubRR,
+                             {Operand::VReg(neg_fixup, VRegClass::Int),
+                              Operand::VReg(masked, VRegClass::Int),
+                              Operand::Imm(val)});
+              }
+              else
+              {
+                int val_reg = function.CreateVReg(VRegClass::Int);
+                block.Append(Opcode::MovImm,
+                             {Operand::VReg(val_reg, VRegClass::Int),
+                              Operand::Imm(val)}).SetRematerializable(true).SetRematImm(val);
+                block.Append(Opcode::SubRR,
+                             {Operand::VReg(neg_fixup, VRegClass::Int),
+                              Operand::VReg(masked, VRegClass::Int),
+                              Operand::VReg(val_reg, VRegClass::Int)});
+              }
               block.Append(Opcode::Csel,
-                           {Operand::VReg(selected, VRegClass::Int),
-                            Operand::VReg(biased, VRegClass::Int),
-                            Operand::VReg(lhs, VRegClass::Int),
-                            Operand::Imm(static_cast<int>(CondCode::LT))});
-              int q_dst = function.CreateVReg(VRegClass::Int);
-              block.Append(Opcode::AsrRR,
-                           {Operand::VReg(q_dst, VRegClass::Int),
-                            Operand::VReg(selected, VRegClass::Int),
-                            Operand::Imm(shift)});
-              int d_reg = function.CreateVReg(VRegClass::Int);
-              block.Append(Opcode::MovImm,
-                           {Operand::VReg(d_reg, VRegClass::Int),
-                            Operand::Imm(val)}).SetRematerializable(true).SetRematImm(val);
-              block.Append(Opcode::Msub,
                            {Operand::VReg(dst, VRegClass::Int),
-                            Operand::VReg(q_dst, VRegClass::Int),
-                            Operand::VReg(d_reg, VRegClass::Int),
-                            Operand::VReg(lhs, VRegClass::Int)});
+                            Operand::VReg(neg_fixup, VRegClass::Int),
+                            Operand::VReg(masked, VRegClass::Int),
+                            Operand::Imm(static_cast<int>(CondCode::LT))});
               value_vregs[value] = dst;
               return dst;
             }