Обсуждение: FPE on Alpha in Select with PKEY
============================================================================
POSTGRESQL BUG REPORT TEMPLATE
============================================================================
Your name : Adriaan Joubert
Your email address : a.joubert@albourne.com
System Configuration
---------------------
Architecture (example: Intel Pentium) : DEC Alpha
Operating System (example: Linux 2.0.26 ELF) : DU 4.0 D
PostgreSQL version (example: PostgreSQL-6.4.2) : PostgreSQL-6.4.2
Compiler used (example: gcc 2.8.0) : DEC C V5.6-071
Please enter a FULL description of your problem:
------------------------------------------------
When doing a select with multiple OR's or an IN statement where this
involves a primary key, it causes a floating point exception.
Creating two tables A and B as follows:
CREATE TABLE A (
PRIMARY KEY (a),
a int4 NOT NULL,
b int4 NOT NULL
);
INSERT INTO A values (73,2);
INSERT INTO A values (74,2);
CREATE TABLE B (
a int4 NOT NULL,
b int4 NOT NULL
);
INSERT INTO B values (73,2);
INSERT INTO B values (74,2);
we get the following behaviour
tt=> select * from B where a in (73,74);
a|b
--+-
73|2
74|2
(2 rows)
tt=> select * from A where a in (73,74);
ERROR: floating point exception! The last floating point operation
either exceeded legal ranges or was a divide by zero
tt=> select * from A where a=73 or a=74;
ERROR: floating point exception! The last floating point operation
either exceeded legal ranges or was a divide by zero
Please describe a way to repeat the problem. Please try to provide a
concise reproducible example, if at all possible:
----------------------------------------------------------------------
See above.
If you know how this problem might be fixed, list the solution below:
---------------------------------------------------------------------
I know it crashed in create_or_index_paths: following stack-trace
0 0x1200c83b4 in create_or_index_paths(0x0, 0x140201e80, 0x1f5, 0x1,
0x4, 0x0) in ./postgres
#1 0x1200c2924 in UnknownProcedure1FromFile134(0x0, 0x140201e80, 0x1f5,
0x1, 0x4, 0x0) in ./postgres
#2 0x1200c281c in find_paths(0x0, 0x140201e80, 0x1f5, 0x1, 0x4, 0x0) in
./postgres
#3 0x1200cb074 in UnknownProcedure1FromFile147(0x0, 0x140201e80, 0x1f5,
0x1, 0x4, 0x0) in ./postgres
#4 0x1200caedc in query_planner(0x0, 0x140201e80, 0x1f5, 0x1, 0x4, 0x0)
in ./postgres
#5 0x1200cb9bc in union_planner(0x0, 0x140201e80, 0x1f5, 0x1, 0x4, 0x0)
in ./postgres
#6 0x1200cb75c in planner(0x0, 0x140201e80, 0x1f5, 0x1, 0x4, 0x0) in
./postgres
#7 0x12011e3b0 in pg_parse_and_plan(0x0, 0x140201e80, 0x1f5, 0x1, 0x4,
0x0) in ./postgres
#8 0x12011e654 in pg_exec_query_dest(0x0, 0x140201e80, 0x1f5, 0x1, 0x4,
0x0) in ./postgres
#9 0x12011fffc in PostgresMain(0x0, 0x140201e80, 0x1f5, 0x1, 0x4, 0x0)
in ./postgres
#10 0x1200b029c in main(0x0, 0x140201e80, 0x1f5, 0x1, 0x4, 0x0) in
./postgres
#11 0x12003b1a8 in __start(0x0, 0x140201e80, 0x1f5, 0x1, 0x4, 0x0) in
./postgres
I'm going to have to recompile with debugging switched on to have a
better
look at it. Anybody have any idea which source file this is in?
--
--------------------------------------+-------------------------------
Dr Adriaan Joubert | Phone: +357-2-750 652
APL Financial Services (Overseas) Ltd| Fax: +357-2-750 654
3 D. Vikella St | e-mail: a.joubert@albourne.com
1061 Nicosia, CYPRUS |
--------------------------------------+-------------------------------
Further to the above: the error occurs on line 136 of the file
src/backend/optimizer/path/orindxpath.c
which reads
> 136 clausenode->selectivity = (Cost) floatVal(selecs);
floatVal is a macro
#define floatVal(v) (((Value *)v)->val.dval)
where value is
typedef struct Value
{
NodeTag type; /* tag appropriately
(eg. T_String) */
union ValUnion
{
char *str; /* string */
long ival;
double dval;
} val;
} Value;
The variable selecs is of type list, i.e.
typedef struct List
{
NodeTag type;
union
{
void *ptr_value;
int int_value;
} elem;
struct List *next;
} List;
The variable selecs is filled in in the routine best_or_subclause_index
in orindxpath.c, where the value that is read via the floatVal macro is
assigned from a float (called selec as well). So this seems wrong
anyway.
Unfortunately casting ptr_value to float gives garbage as well:
(ladebug) p *(double*)selecs->elem.ptr_value
2.48515019858147e-321
(ladebug) p *(float*)selecs->elem.ptr_value
1.3342e-312
Any ideas?
Adriaan
Further to the above: the error occurs on line 136 of the file
src/backend/optimizer/path/orindxpath.c
> 136 clausenode->selectivity = (Cost) floatVal(selecs);
floatVal is a macro
#define floatVal(v) (((Value *)v)->val.dval)
where value is
typedef struct Value
{
NodeTag type; /* tag appropriately
(eg. T_String) */
union ValUnion
{
char *str; /* string */
long ival;
double dval;
} val;
} Value;
The variable selecs is of type list, i.e.
typedef struct List
{
NodeTag type;
union
{
void *ptr_value;
int int_value;
} elem;
struct List *next;
} List;
The variable selecs is filled in in the routine best_or_subclause_index
in orindxpath.c, where the value that is read via the floatVal macro is
assigned from a float (called selec as well). So this seems wrong
anyway.
Unfortunately casting ptr_value to float gives garbage as well:
(ladebug) p *(double*)selecs->elem.ptr_value
2.48515019858147e-321
(ladebug) p *(float*)selecs->elem.ptr_value
1.3342e-312
Any ideas?
Adriaan
Hi,
I seem to have fixed the floating point problem on the Alphas -- at
least for my test cases. It requires the following patch at the end of
this mail. Could somebody who know postgres a bit better than me please
double-check it, and let me know whether it is ok? The regression test
seemed to run ok'ish (i.e. no more than before failed).
Adriaan
*** ./src/backend/optimizer/path/orindxpath.c.orig Mon Sep 21
18:41:27 1998
--- ./src/backend/optimizer/path/orindxpath.c Tue Mar 2 16:25:11 1999
***************
*** 133,139 ****
xfunc_get_path_cost((Path) pathnode);
}
#endif
! clausenode->selectivity = (Cost)
floatVal(selecs);
t_list = lappend(t_list, pathnode);
}
}
--- 133,139 ----
xfunc_get_path_cost((Path) pathnode);
}
#endif
! clausenode->selectivity = (Cost)
floatVal(selecs->elem.ptr_value);
t_list = lappend(t_list, pathnode);
}
}
Yes. I have an idea. Did you add a column to the table using ALTER
TABLE. There was a bug in 6.4.* and earlier that caused this to happend
for added columns. Do a vacuum analyze on the table and the problem
will go away. Will be fixed in the next release.
> Further to the above: the error occurs on line 136 of the file
> src/backend/optimizer/path/orindxpath.c
>
> > 136 clausenode->selectivity = (Cost) floatVal(selecs);
>
>
>
> floatVal is a macro
>
> #define floatVal(v) (((Value *)v)->val.dval)
>
> where value is
>
> typedef struct Value
> {
> NodeTag type; /* tag appropriately
> (eg. T_String) */
> union ValUnion
> {
> char *str; /* string */
> long ival;
> double dval;
> } val;
> } Value;
>
> The variable selecs is of type list, i.e.
>
> typedef struct List
> {
> NodeTag type;
> union
> {
> void *ptr_value;
> int int_value;
> } elem;
> struct List *next;
> } List;
>
>
> The variable selecs is filled in in the routine best_or_subclause_index
> in orindxpath.c, where the value that is read via the floatVal macro is
> assigned from a float (called selec as well). So this seems wrong
> anyway.
>
> Unfortunately casting ptr_value to float gives garbage as well:
>
> (ladebug) p *(double*)selecs->elem.ptr_value
> 2.48515019858147e-321
> (ladebug) p *(float*)selecs->elem.ptr_value
> 1.3342e-312
>
>
> Any ideas?
>
> Adriaan
>
>
--
Bruce Momjian | http://www.op.net/~candle
maillist@candle.pha.pa.us | (610) 853-3000
+ If your life is a hard drive, | 830 Blythe Avenue
+ Christ can be your backup. | Drexel Hill, Pennsylvania 19026