Pavel Stehule wrote:
> 2017-03-02 19:32 GMT+01:00 Alvaro Herrera <>:
> > So in the old (non-executor-node) implementation, you could attach WITH
> > ORDINALITY to the xmltable expression and it would count the output
> > rows, regardless of which XML document it comes from.  With the new
> > implementation, the grammar no longer accepts it.  To count output rows,
> > you still need to use row_number().  Maybe this is okay.  This is the
> > example from the docs, and I add another XML document with two more rows
> > for xmltable.  Look at the three numbering columns ...
> >
> It is expected - now tablefunc are not special case of SRF, so it lost all
> SRF functionality. It is not critical lost - it supports internally FOR
> ORDINALITY column, and classic ROW_NUMBER can be used. It can be enhanced
> to support WITH ORDINALITY in future, but I have not any use case for it.


After looking at the new executor code a bit, I noticed that we don't
need the resultSlot anymore; we can use the ss_ScanTupleSlot instead.
Because resultSlot was being used in the xml.c code (which already
appeared a bit dubious to me), I changed the interface so that instead
the things that it read from it are passed as parameters -- namely, in
InitBuilder we pass natts, and in GetValue we pass typid and typmod.

Secondly, I noticed we have the FetchRow routine produce a minimal
tuple, put it in a slot; then its caller takes the slot and put the
tuple in the tuplestore.  This is pointless; we can just have FetchRow
put the tuple in the tuplestore directly and not bother with any slot
manipulations there.  This simplifies the code a bit.

Here's v47 with those changes.

Álvaro Herrera      
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

Attachment: xmltable-47.patch.gz
Description: application/gunzip

Sent via pgsql-hackers mailing list (
To make changes to your subscription:

Reply via email to