Making Bitmapsets be valid Nodes
Per the discussion at [1]/messages/by-id/94353655-c177-1f55-7afb-b2090de33341@enterprisedb.com, it seems like it'd be a good idea to make
Bitmapsets into full-fledged, tagged Nodes, so that we could do things
like print or copy lists of them without special-case logic. The
extra space for the NodeTag is basically free due to alignment
considerations, at least on 64-bit hardware.
Attached is a cleaned-up version of Amit's patch v24-0003 at [2]/messages/by-id/CA+HiwqEYCLRZ2Boq_uK0pjLn_9b8XL-LmwKj7HN5kJOivUkYLg@mail.gmail.com.
I fixed the problems with not always tagging Bitmapsets, and changed
the outfuncs/readfuncs logic so that Bitmapsets still print exactly
as they did before (thus, this doesn't require a catversion bump).
As proof of concept, I removed the read_write_ignore labels from
RelOptInfo's unique_for_rels and non_unique_for_rels fields, and
got nice-looking debug printout:
:unique_for_rels ((b 1))
:non_unique_for_rels <>
I also removed some special-case code from indxpath.c because
list_member() can do the same thing now. (There might be other
places that can be simplified; I didn't look very hard.)
It'd be possible to make Bitmapset fields be (mostly) not special cases
in the copy/equal/out/read support. But I chose to leave that alone,
because it'd add a little runtime overhead for indirecting through the
generic support functions while not really saving any code space.
Barring objections, I'd like to go ahead and push this.
regards, tom lane
[1]: /messages/by-id/94353655-c177-1f55-7afb-b2090de33341@enterprisedb.com
[2]: /messages/by-id/CA+HiwqEYCLRZ2Boq_uK0pjLn_9b8XL-LmwKj7HN5kJOivUkYLg@mail.gmail.com
Attachments:
v1-0001-make-Bitmapsets-be-valid-Nodes.patchtext/x-diff; charset=us-ascii; name=v1-0001-make-Bitmapsets-be-valid-Nodes.patchDownload+68-30
On 11.11.22 21:05, Tom Lane wrote:
Per the discussion at [1], it seems like it'd be a good idea to make
Bitmapsets into full-fledged, tagged Nodes, so that we could do things
like print or copy lists of them without special-case logic. The
extra space for the NodeTag is basically free due to alignment
considerations, at least on 64-bit hardware.Attached is a cleaned-up version of Amit's patch v24-0003 at [2].
I fixed the problems with not always tagging Bitmapsets, and changed
the outfuncs/readfuncs logic so that Bitmapsets still print exactly
as they did before (thus, this doesn't require a catversion bump).
This looks good to me.
Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:
On 11.11.22 21:05, Tom Lane wrote:
Attached is a cleaned-up version of Amit's patch v24-0003 at [2].
I fixed the problems with not always tagging Bitmapsets, and changed
the outfuncs/readfuncs logic so that Bitmapsets still print exactly
as they did before (thus, this doesn't require a catversion bump).
This looks good to me.
Pushed, thanks for looking.
regards, tom lane
On Mon, Nov 14, 2022 at 12:23 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
Peter Eisentraut <peter.eisentraut@enterprisedb.com> writes:
On 11.11.22 21:05, Tom Lane wrote:
Attached is a cleaned-up version of Amit's patch v24-0003 at [2].
I fixed the problems with not always tagging Bitmapsets, and changed
the outfuncs/readfuncs logic so that Bitmapsets still print exactly
as they did before (thus, this doesn't require a catversion bump).This looks good to me.
Pushed, thanks for looking.
Thanks a lot for this.
I agree that it may not be worthwhile to add an extra function call by
changing COPY_BITMAPSET_FIELD, etc. that is currently emitted by
gen_node_support.pl for any Bitmapset * / Relid struct members to
COPY_NODE_FIELD, etc.
--
Thanks, Amit Langote
EDB: http://www.enterprisedb.com