Thread: HeapTupleHeader withoud oid
We have been discussing heap tuple header changes for a while now.
Here is my proposal for omitting the oid, when it is not needed:
First let's eliminate t_oid from HeapTupleHeaderData.
Then add the oid to the end of the structure, if and only if it is
needed. The tricky part here is that there is a variable length field
(t_bits) and the oid has to be properly aligned.
This pseudo code snippet illustrates what I plan to do:
len = offsetof(HeapTupleHeaderData, t_bits); /* 23 */if (hasnulls) { len += BITMAPLEN(NumberOfAttributes);}if
(hasoid){ len += sizeof(Oid);}len = MAXALIGN(len);hoff = len;oidoff = hoff - sizeof(Oid);
#define HeapTupleHeaderGetOid(hth) \( *((Oid *)((char *)(hth) + (hth)->t_hoff - sizeof(Oid))) )
And this is how the structure would look like: 1 2 3 0 4 0 0 34 78
2
now oooo<---------fix--------->.x___X___
+oid <---------fix--------->.oooox___ MAXALIGN 4
+oid <---------fix--------->.....ooooX___ MAXALIGN 8
-oid <---------fix--------->.X___
1:
now oooo<---------fix--------->bx___X___
+oid <---------fix--------->boooox___ MAXALIGN 4
+oid <---------fix--------->b....ooooX___ MAXALIGN 8
-oid <---------fix--------->bX___
2:
now oooo<---------fix--------->bb...X___
+oid <---------fix--------->bb...ooooX___ MAXALIGN 4 und 8
-oid <---------fix--------->bb...x___X___ 3 4
6: 2 6 0
now oooo<---------fix--------->bbbbbb...x___X___
+oid <---------fix--------->bbbbbb...oooox___ MAXALIGN 4
+oid <---------fix--------->bbbbbb.......ooooX___ MAXALIGN 8
-oid <---------fix--------->bbbbbb...X___
where
<---------fix---------> fixed sized part without oid, 23 bytes
oooo oid, 4 bytes
b one bitmap byte
. one padding byte
x start of data area (= hoff) with 4-byte-alignment
X start of data area (= hoff) with 8-byte-alignment
Bytes saved on architectures with 4/8 byte alignment: hoff bytes
natts bitmaplen hoff72 oidoff woo saved 0 28/32 24 24/24 4/8
1-8 1 28/32 24 24/24 4/8
9-40 2-5 32/32 28 28/32 4/0
41-72 6-9 36/40 32 32/32 4/8
As a first step I've already posted a patch that eliminates direct
access to t_oid. The final patch will change not much more than the
getter and setter macros.
Problems I have identified so far:
. heap_formtuple needs a parameter bool withoid
. Does heap_addheader *always* create a header with oid?
. Have to check heap_xlog_xxxx routines
. Occasionally a heap tuple header is copied by memmove.
Comments?
ServusManfred
Manfred Koizar <mkoi-pg@aon.at> writes:
> . Does heap_addheader *always* create a header with oid?
No.
regards, tom lane
On Mon, 01 Jul 2002 12:40:35 +0200, I wrote:
>Bytes saved on architectures with 4/8 byte alignment:
> hoff bytes
>natts bitmaplen hoff72 oidoff woo saved
> 0 28/32 24 24/24 4/8
>1-8 1 28/32 24 24/24 4/8
>9-40 2-5 32/32 28 28/32 4/0
>41-72 6-9 36/40 32 32/32 4/8
In this table oidoff contains wrong values, it is from my first
approach, where I tried to put oid at the first INTALIGNed position
after t_bits. The table should be:
bitmap hoff bytes
natts len hoff1 hoff2 oidoff woo saved 0 32 28/32 24/28 24 4/8
1-8 1 32 28/32 24/28 24 4/8
9-40 2-5 36/40 32 28 28/32 4/0
41-72 6-9 40 36/40 32/36 32 4/8
where hoff1 is the MAXALIGNed length of the tuple header with a v7.2
compatible tuple header format (with bitmaplen patch included);
hoff2 is the header size after the Xmin/Cid/Xmax patch, which is still
being discussed on -patches and -hackers;
with this proposal, if a table has oids, oidoff is the offset of the
oid and header size equals hoff2;
hoff woo is the header size without oid;
bytes saved is relative to hoff2.
I apologize for the confusion.
ServusManfred
Manfred Koizar <mkoi-pg@aon.at> writes:
> Here is my proposal for omitting the oid, when it is not needed:
I do not think you can make this work unless "has oids" is added to
TupleDescs. There are too many places where tuples are manipulated
with only a tupdesc for reference.
It might also be necessary to add a "has oid" bit to t_infomask,
so that a tuple's OID can be fetched with *no* outside information,
but I'd prefer to avoid that if possible. I think adding a tupledesc
parameter to heap_getsysattr might be enough to avoid it.
I'd suggest reworking your "Wrap access to Oid" patch, which currently
increases instead of reducing the dependency on access to a Relation
for the tuple. Also, you could be a little more conservative about
adding Asserts --- those are not free, at least not from a development
point of view, so I object to adding multiple redundant Asserts in
hotspot routines.
regards, tom lane