Currently information_schema tables work like:

1. prepare information_schema table
   • this creates a temporary table
2. call the information_schema implementation code
   • it sets values using Field::store() and calls schema_table_store_record() per row
   • schema_table_store_record() uses handler::ha_write_row() to store the row in he temporary table
3. when the temporary table is filled with data, it's used in the query.

For queries like SELECT f1, f2, ... FROM INFORMATION_SCHEMA.tbl the above adds a lot of overhead. The server can recognize that case, not create a temporary table in the step 1. And modify schema_table_store_record() to send results directly to the client.

Another case when a temporary table can be avoided is when a subquery with the information schema table is used in a context implying that the result can be only one row (or an error). Like

SET @a=(SELECT VARIABLE_VALUE FROM INFORMATION_SCHEMA.GLOBAL_STATUS WHERE VARIABLE_NAME='xxx')