[PATCH] Item_func_neg relies on undefined behavior and can misbehave

Class Item_func_neg has a lot of conversions between signed and unsigned datatypes. I believe at least to some degree these conversions are undefined behavior in C++, and thus Item_func_neg can misbehave when compiled with optimizations. In particular I've experienced misbehavior in the main.func_math test on line 483. This line expects that negation of -9223372036854775808 will throw DATA_OUT_OF_RANGE error, but sometimes it doesn't. The following patch fixes the problem. Please consider including it (or some modification of it) into MariaDB.

--- a/sql/item_func.cc

+++ b/sql/item_func.cc

@@ -1908,6 +1908,8 @@ longlong Item_func_neg::int_op()

   if (args[0]->unsigned_flag &&

       (ulonglong) value > (ulonglong) LONGLONG_MAX + 1)

     return raise_integer_overflow();

+  if (!args[0]->unsigned_flag && value == LONGLONG_MIN)

+    return raise_integer_overflow();

   return check_integer_overflow(-value, !args[0]->unsigned_flag && value < 0);

 }