Had mangled the order of if()s in a previous changeset (1.1596) (not pushed),

correcting it now. Thanks Dmitri for spotting this.

mysql-test/r/user_var.result

@@ -5,6 +5,10 @@ set @a := connection_id() + 3;
 select @a - connection_id();
 @a - connection_id()
 3
+set @b := 1;
+select @b;
+@b
+1
 CREATE TABLE t1 ( i int not null, v int not null,index (i));
 insert into t1 values (1,1),(1,3),(2,1);
 create table t2 (i int not null, unique (i));

mysql-test/t/user_var.test

@@ -8,6 +8,9 @@ set @a := foo;
 set @a := connection_id() + 3;
 select @a - connection_id();
 
+set @b := 1;
+select @b;
+
 # Check using and setting variables with SELECT DISTINCT
 
 CREATE TABLE t1 ( i int not null, v int not null,index (i));

sql/item_func.cc

@@ -2343,25 +2343,23 @@ longlong Item_func_get_user_var::val_int()
 void Item_func_get_user_var::fix_length_and_dec()
 {
   THD *thd=current_thd;
-
-  if (!(opt_bin_log && is_update_query(thd->lex.sql_command)))
-       return;
-
   BINLOG_USER_VAR_EVENT *user_var_event;
   maybe_null=1;
   decimals=NOT_FIXED_DEC;
   max_length=MAX_BLOB_WIDTH;
 
+  var_entry= get_variable(&thd->user_vars, name, 0);
+
+  if (!(opt_bin_log && is_update_query(thd->lex.sql_command)))
+    return;
+
+  if (!var_entry)
+  {
     /*
       If the variable does not exist, it's NULL, but we want to create it so
       that it gets into the binlog (if it didn't, the slave could be
       influenced by a variable of the same name previously set by another
       thread).
-  */
-  
-  if (!(var_entry= get_variable(&thd->user_vars, name, 0)))
-  {
-    /*
       We create it like if it had been explicitely set with SET before.
       The 'new' mimicks what sql_yacc.yy does when 'SET @a=10;'.
       sql_set_variables() is what is called from 'case SQLCOM_SET_OPTION'