Bug #30355: Incorrect ordering of UDF results

There's currently no way of knowing the determinicity of an UDF.
And the optimizer and the sequence() UDFs were making wrong
assumptions about what the is_const member means.
Plus there was no implementation of update_system_tables()
causing the optimizer to overwrite the information returned by
the <udf>_init function.

Fixed by equating the assumptions about the semantics of 
is_const and providing a implementation of update_used_tables().
Added a TODO item for the UDF API change needed to make a better 
implementation.


include/mysql_com.h:
  Bug #30355: comment added
mysql-test/r/udf.result:
  Bug #30355: test case
mysql-test/t/udf.test:
  Bug #30355: test case
sql/item_func.cc:
  Bug #30355: keep const_item_cache and used_tables_cache in sync
sql/item_func.h:
  Bug #30355: 
   - a better implementation of update_used_tables()
   - keep const_item_cache and used_tables_cache in sync
sql/udf_example.c:
  Bug #30355: Wrong value for const_item fixed.

1 files changed, 50 insertions, 0 deletions

diff --git a/sql/item_func.h b/sql/item_func.h
index a31294c0395..734b215ddc0 100644
--- a/sql/item_func.h
+++ b/sql/item_func.h
@@ -1016,6 +1016,56 @@ public:
     fixed= 1;
     return res;
   }
+  void update_used_tables() 
+  {
+    /*
+      TODO: Make a member in UDF_INIT and return if a UDF is deterministic or
+      not.
+      Currently UDF_INIT has a member (const_item) that is an in/out 
+      parameter to the init() call.
+      The code in udf_handler::fix_fields also duplicates the arguments 
+      handling code in Item_func::fix_fields().
+      
+      The lack of information if a UDF is deterministic makes writing
+      a correct update_used_tables() for UDFs impossible.
+      One solution to this would be :
+       - Add a is_deterministic member of UDF_INIT
+       - (optionally) deprecate the const_item member of UDF_INIT
+       - Take away the duplicate code from udf_handler::fix_fields() and
+         make Item_udf_func call Item_func::fix_fields() to process its 
+         arguments as for any other function.
+       - Store the deterministic flag returned by <udf>_init into the 
+       udf_handler. 
+       - Don't implement Item_udf_func::fix_fields, implement
+       Item_udf_func::fix_length_and_dec() instead (similar to non-UDF
+       functions).
+       - Override Item_func::update_used_tables to call 
+       Item_func::update_used_tables() and add a RAND_TABLE_BIT to the 
+       result of Item_func::update_used_tables() if the UDF is 
+       non-deterministic.
+       - (optionally) rename RAND_TABLE_BIT to NONDETERMINISTIC_BIT to
+       better describe its usage.
+       
+      The above would require a change of the UDF API.
+      Until that change is done here's how the current code works:
+      We call Item_func::update_used_tables() only when we know that
+      the function depends on real non-const tables and is deterministic.
+      This can be done only because we know that the optimizer will
+      call update_used_tables() only when there's possibly a new const
+      table. So update_used_tables() can only make a Item_func more
+      constant than it is currently.
+      That's why we don't need to do anything if a function is guaranteed
+      to return non-constant (it's non-deterministic) or is already a
+      const.
+    */  
+    if ((used_tables_cache & ~PSEUDO_TABLE_BITS) && 
+        !(used_tables_cache & RAND_TABLE_BIT))
+    {
+      Item_func::update_used_tables();
+      if (!const_item_cache && !used_tables_cache)
+        used_tables_cache= RAND_TABLE_BIT;
+    }
+  }
   void cleanup();
   Item_result result_type () const { return udf.result_type(); }
   table_map not_null_tables() const { return 0; }