In order to make Items reentrant (see MDEV-6897) we need to get rid of the boolean member Item::null_value.

To remove this member, we'll need to modify the following methods in Item to return the actual value and the SQL NULL flag at the same time:

virtual bool val_bool();

virtual double val_real();

virtual longlong val_int();

Note, the other methods val_str(), val_decimal(), get_date(), get_time() do not necessarily need to be modified, because they already return SQL NULL as their return value:

• val_str() returns NULL to mean SQL NULL
• val_decimal() returns NULL to mean SQL NULL
• get_date() returns true to mean SQL NULL
• get_time() returns true to mean SQL NULL

Possible options

Add a pointer to the SQL NULL flag:

virtual bool val_bool_null(bool *null_value_arg);

virtual longlong val_int_null(bool *null_value_arg);

virtual double val_real_null(bool *null_value_arg);

Return SQL NULL, pass the value target as a pointer:

virtual bool get_bool(bool *to);

virtual bool get_longlong(longlong *to);

virtual bool get_double(double *to);

Return a structure consisting of the value and the SQL NULL flag

virtual Bool_null to_bool_null();

virtual Longlong_null to_longlong_null();

virtual Double_null to_double_null();

where the structures are about the following:

class Bool_null { bool value; bool is_null };

class Longlong_null { longlong value; bool is_null };

class Double_null { double value; bool is_null };

Mixed

A mixed API could be also considered depending on which exact method performs better for a certain data type. For example:

virtual Bool_null to_bool_null();

virtual longlong val_int_null(bool *null_value_arg);

virtual bool get_double(double *to);

But this probably won't be convenient.

Plan

Under terms of this task we'll:

• Implement a bechmark program to test the performance of all proposed new API versions
• Find the way which performs better for every data type
• Discuss the winner methods
• Decide the new API
• Implement the decided API