Details
Description
http://buildbot.askmonty.org/buildbot/builders/kvm-deb-sid-x86/builds/9229
|
10.4 903f5fea30cb236c5980a07b7fa63450 |
main.func_math w4 [ fail ]
|
Test ended at 2019-11-02 15:51:08
|
|
|
CURRENT_TEST: main.func_math
|
mysqltest: At line 329: query 'SELECT -9223372036854775808 DIV -1' succeeded - should have failed with errno 1690...
|
|
|
The result from queries just before the failure was:
|
< snip >
|
CREATE OR REPLACE VIEW v1 AS SELECT NULL AS a;
|
SELECT RAND(a) FROM v1;
|
RAND(a)
|
0.15522042769493574
|
DROP VIEW v1;
|
SELECT RAND(a) FROM (SELECT NULL AS a) b;
|
RAND(a)
|
0.15522042769493574
|
CREATE TABLE t1 (i INT);
|
INSERT INTO t1 VALUES (NULL);
|
SELECT RAND(i) FROM t1;
|
RAND(i)
|
0.15522042769493574
|
DROP TABLE t1;
|
#
|
# Bug#57477 SIGFPE when dividing a huge number a negative number
|
#
|
SELECT -9999999999999999991 DIV -1;
|
ERROR 22003: BIGINT value is out of range in '-9999999999999999991 DIV -1'
|
SELECT -9223372036854775808 DIV -1;
|
|
|
More results from queries before failure can be found in /dev/shm/var/4/log/func_math.log
|
|
|
Warnings from just before the error:
|
Warning 1916 Got overflow when converting '9999999999999999991' to INT. Value truncated
|
Attachments
Issue Links
- relates to
-
MDBF-536 Remove Fedora 36 builder reporting to GitHub
-
- Closed
-
-
MDEV-21977 main.func_math fails due to undefined behaviour
-
- Closed
-
-
-
- Closed
-
Activity
The test main.func_math fails due to undefined behaviour on 10.2, 10.3, 10.4 when built with gcc-12 -O2 for AMD64, or even some older GCC on non-AMD64 instruction set architectures, such as POWER or IBM Z Series (s390x).
For 10.5 and later, MDEV-26645 fixed this by removing an ‘optimization’. Here is a recent example:
CURRENT_TEST: main.func_math
|
mysqltest: At line 425: query 'SELECT 9223372036854775807 + 9223372036854775807' succeeded - should have failed with errno 1690...
|
The reason for this failure is that the compiler is allowed to assume that no undefined behaviour exists. Specifically, if the compiler was asked to evaluate a+b, and the code subsequently checks for overflow, the compiler is allowed to assume that no overflow occurred in the first place, and to optimize away the overflow check.
This affects the Fedora 36 builder on new buildbot for 10.3 & 10.4.
Attempted to backport the UB fixed and failed https://github.com/MariaDB/server/pull/2458.
As Fedora 36 comes with 10.5 and and with luck and a lot of hard work, Debian 12 will come with 10.11 I think the set distros with new compilers needing this fixed for 10.3/10.4 has been reduce to an empty set..
func_math passing on this recent build http://buildbot.askmonty.org/buildbot/builders/kvm-deb-sid-x86/builds/12362/steps/mtr/logs/stdio ( 10.5)
http://buildbot.askmonty.org/buildbot/builders/kvm-deb-sid-x86/builds/12360/steps/mtr/logs/stdio (10.6)
Can't find recent 10.4 test on deb-sid-x86