> I think that you have a misunderstanding of how memory context works in
> postgres.
No, I understand how memory contexts work, but I skipped over the fact that the
cleaned up query text is copied into CachedPlanSource memory context during
CreateCachedPlan:
```
plansource->query_string = pstrdup(query_string);
```
Which is long-lived and will remain until DEALLOCATE.
while pstate->p_sourcetext and new_query are in the current execution's
memory context.
The patch is indeed better from a memory perspective. My apologies.
However, the error reporting does break with the patch. Notice with the patch
the cursor for the error reporting shifts incorrectly. This is due to the fact
rawstmt->stmt_location/length are no longer representative of the original
qurey text.
## unpatched
```
postgres=# SELECT 1 \; PREPARE stmt AS SELECT nonexistent_column
FROM users\; SELECT 2;
?column?
----------
1
(1 row)
ERROR: relation "users" does not exist
LINE 1: ... ; PREPARE stmt AS SELECT nonexistent_column FROM users; SEL...
^
```
## patched
```
postgres=# SELECT 1 \; PREPARE stmt AS SELECT nonexistent_column
FROM users\; SELECT 2;
?column?
----------
1
(1 row)
ERROR: relation "users" does not exist
LINE 1: ...LECT 1 ; PREPARE stmt AS SELECT nonexistent_column FROM u...
^
```
--
Sami Imseih
Amazon Web Services (AWS)
