Re: [HACKERS] makesign() broken in tsearch2

[Teodor Sigaev]

the bottom half might be messed up too.  So we are failing to exploit
the full intended sign space, which presumably is costing something
in index efficiency.

 


You are absolutly right. My fault.


It looks to me like the values calculated by this routine end up on
disk, and therefore we can't fix it without forcing an initdb, or
at least REINDEX of all affected indexes.  Is that correct?
 

No.  query_gist.c exists only in HEAD branch.  That code is devloped  
not a lot time ago, and itsn't tested well yet.


---(end of broadcast)---
TIP 3: Have you checked our extensive FAQ?

  http://www.postgresql.org/docs/faq

Re: [HACKERS] makesign() broken in tsearch2

[Tom Lane]

Teodor Sigaev [EMAIL PROTECTED] writes:
 It looks to me like the values calculated by this routine end up on
 disk, and therefore we can't fix it without forcing an initdb, or
 at least REINDEX of all affected indexes.  Is that correct?
 
 No.  query_gist.c exists only in HEAD branch.

Oh, good, that makes it easy.  Do you want to apply the fix or shall I?

regards, tom lane

---(end of broadcast)---
TIP 6: explain analyze is your friend

Re: [HACKERS] makesign() broken in tsearch2

[Teodor Sigaev]

Oh, good, that makes it easy.  Do you want to apply the fix or shall I?
 

I will have normal access to Internet only on Monday. It can easy wait 
for me :)


---(end of broadcast)---
TIP 9: In versions below 8.0, the planner will ignore your desire to
  choose an index scan if your joining column's datatypes do not
  match

[HACKERS] makesign() broken in tsearch2

[Tom Lane]

I noticed the following code in tsearch2:

typedef uint64 TPQTGist;

static TPQTGist
makesign(QUERYTYPE * a)
{
int i;
ITEM   *ptr = GETQUERY(a);
TPQTGistsign = 0;

for (i = 0; i  a-size; i++)
{
if (ptr-type == VAL)
sign |= 1  (ptr-val % SIGLEN);
ptr++;
}

return sign;
}

This is wrong because 1 is an int constant, not an int64, and
therefore the shift will be done in int width.  Correct would be

sign |= ((TPQTGist) 1)  (ptr-val % SIGLEN);

The effect of this is at least that the high-order half of sign
remains always zero.  Depending on what the machine does with shifts
exceeding the word width (which is undefined according to ANSI C),
the bottom half might be messed up too.  So we are failing to exploit
the full intended sign space, which presumably is costing something
in index efficiency.

It looks to me like the values calculated by this routine end up on
disk, and therefore we can't fix it without forcing an initdb, or
at least REINDEX of all affected indexes.  Is that correct?

regards, tom lane

---(end of broadcast)---
TIP 1: if posting/reading through Usenet, please send an appropriate
   subscribe-nomail command to [EMAIL PROTECTED] so that your
   message can get through to the mailing list cleanly