Align base addr of multiple io

Author: Grarak

src/jit/emitter/emit_transfer.rs

@@ -375,16 +375,21 @@ impl<const CPU: CpuType> JitAsm<'_, CPU> {
                 block_asm.bfc(base_reg, mmu::MMU_PAGE_SHIFT as u8, 32 - mmu::MMU_PAGE_SHIFT as u8);
                 block_asm.add(base_reg, mmu_offset_reg, base_reg);
                 block_asm.add(base_reg, base_reg, shm_ptr as u32);
+                block_asm.bic(base_reg, base_reg, 3);
 
                 for (guest_reg, fixed_reg) in non_gp_regs_mappings {
                     block_asm.mov(BlockReg::Fixed(fixed_reg), guest_reg);
                 }
 
-                block_asm.guest_transfer_write_multiple(base_reg, base_reg_out, gp_regs, fixed_regs, write_back, pre, !decrement);
+                block_asm.guest_transfer_write_multiple(base_reg, base_reg_out, gp_regs, fixed_regs, false, pre, !decrement);
 
                 if write_back {
-                    block_asm.sub(base_reg, base_reg_out, base_reg);
-                    block_asm.add(op0, op0, base_reg);
+                    let len = gp_regs.len() + fixed_regs.len();
+                    if decrement {
+                        block_asm.sub(op0, op0, len as u32 * 4);
+                    } else {
+                        block_asm.add(op0, op0, len as u32 * 4);
+                    }
                 }
 
                 block_asm.msr_cpsr(cpsr_backup_reg);
@@ -403,17 +408,22 @@ impl<const CPU: CpuType> JitAsm<'_, CPU> {
                 let base_reg_out = block_asm.new_reg();
                 let mmu = get_mmu!(self.emu, CPU);
                 let base_ptr = mmu.get_base_tcm_ptr();
-                block_asm.bic(base_reg, op0, 0xF0000000);
+                block_asm.bic(base_reg, op0, 0xF0000003);
                 block_asm.add(base_reg, base_reg, base_ptr as u32);
-                block_asm.guest_transfer_read_multiple(base_reg, base_reg_out, gp_regs, fixed_regs, write_back, pre, !decrement);
+
+                block_asm.guest_transfer_read_multiple(base_reg, base_reg_out, gp_regs, fixed_regs, false, pre, !decrement);
 
                 for (guest_reg, fixed_reg) in non_gp_regs_mappings {
                     block_asm.mov(guest_reg, BlockReg::Fixed(fixed_reg));
                 }
 
                 if write_back {
-                    block_asm.sub(base_reg, base_reg_out, base_reg);
-                    block_asm.add(op0, base_reg, op0);
+                    let len = gp_regs.len() + fixed_regs.len();
+                    if decrement {
+                        block_asm.sub(op0, op0, len as u32 * 4);
+                    } else {
+                        block_asm.add(op0, op0, len as u32 * 4);
+                    }
                 }
 
                 for guest_reg in rlist {