macro_jit_x64: Should not skip zero registers for certain ALU ops

The code generated for these ALU ops assume src_a and src_b are always valid.

src/video_core/macro/macro_jit_x64.cpp

@@ -56,11 +56,13 @@ void MacroJITx64Impl::Compile_ALU(Macro::Opcode opcode) {
     const bool valid_operation = !is_a_zero && !is_b_zero;
     const bool is_move_operation = !is_a_zero && is_b_zero;
     const bool has_zero_register = is_a_zero || is_b_zero;
+    const bool no_zero_reg_skip = opcode.alu_operation == Macro::ALUOperation::AddWithCarry ||
+                                  opcode.alu_operation == Macro::ALUOperation::SubtractWithBorrow;
 
     Xbyak::Reg32 src_a;
     Xbyak::Reg32 src_b;
 
-    if (!optimizer.zero_reg_skip) {
+    if (!optimizer.zero_reg_skip || no_zero_reg_skip) {
         src_a = Compile_GetRegister(opcode.src_a, RESULT);
         src_b = Compile_GetRegister(opcode.src_b, eax);
     } else {