Re: Optimze usage of immutable functions as relation

Hello,

I reworked a patch to make more stable in different cases. I decided to
use simplify_function instead of eval_const_expression to prevent
inlining of the function. The only possible outputs of the
simplify_function are Const node and NULL.

Also, I block pre-evaluation of functions with types other than
TYPTYPE_BASE, cause there is no special logic for compound (and others)
values yet.

There is still a problem with memory leak in case of simplified
arguments. The only way I see is a creation of temporary memory context,
but it cost some performance. Maybe we can store simplified arguments
in the pointed function itself for later use. But eval_const_expression
and friends doesn't change the content of the nodes inside the tree, it
generates new nodes and returns it as a result.

The last point to mention is a fixed plan for the query in the initial
letter of the thread. As I mentioned before, new versions of the patch
replace var not with a function call, but with a function execution
result. After the patch, the following plan is used instead of Nested
Loop with Sequence Scan:

explain select '|'||subject||'|', ts_rank_cd(body_tsvector,q) from messages, to_tsquery('english', 'tuple&header&overhead') q where body_tsvector @@ q limit 10;
QUERY PLAN
----------------------------------------------------------------------------------------------------
Limit (cost=224.16..266.11 rows=3 width=36)
-> Nested Loop (cost=224.16..266.11 rows=3 width=36)
-> Function Scan on q (cost=0.00..0.01 rows=1 width=0)
-> Bitmap Heap Scan on messages (cost=224.16..266.04 rows=3 width=275)
Recheck Cond: (body_tsvector @@ '''tupl'' & ''header'' & ''overhead'''::tsquery)
-> Bitmap Index Scan on message_body_idx (cost=0.00..224.16 rows=3 width=0)
Index Cond: (body_tsvector @@ '''tupl'' & ''header'' & ''overhead'''::tsquery)

--
Aleksandr Parfenov
Postgres Professional: http://www.postgrespro.com
Russian Postgres Company