While working on 10.6 and 10.7 I've noticed that this test is permanently failing:
main.func_math w1 [ fail ]
Test ended at 2021-09-18 22:58:28
CURRENT_TEST: main.func_math
mysqltest: At line 425: query 'SELECT 9223372036854775807 + 9223372036854775807' succeeded - should have failed with error ER_DATA_OUT_OF_RANGE (1690)...
The result from queries just before the failure was:
< snip >
ERROR 22003: DOUBLE value is out of range in 'exp(750)'
SELECT POW(10, 309);
ERROR 22003: DOUBLE value is out of range in 'pow(10,309)'
SELECT COT(0);
ERROR 22003: DOUBLE value is out of range in 'cot(0)'
SELECT DEGREES(1e307);
ERROR 22003: DOUBLE value is out of range in 'degrees(1e307)'
SELECT 9223372036854775808 + 9223372036854775808;
ERROR 22003: BIGINT UNSIGNED value is out of range in '9223372036854775808 + 9223372036854775808'
SELECT 18446744073709551615 + 1;
ERROR 22003: BIGINT UNSIGNED value is out of range in '18446744073709551615 + 1'
SELECT 1 + 18446744073709551615;
ERROR 22003: BIGINT UNSIGNED value is out of range in '1 + 18446744073709551615'
SELECT -2 + CAST(1 AS UNSIGNED);
ERROR 22003: BIGINT UNSIGNED value is out of range in '-2 + cast(1 as unsigned)'
SELECT CAST(1 AS UNSIGNED) + -2;
ERROR 22003: BIGINT UNSIGNED value is out of range in 'cast(1 as unsigned) + -2'
The patch func_math_tests_MDEV-26645.diff I just attached "fixes" the test failure - but as you can see in the patched expected results, there are three queries that end up with a mathematically wrong result (probably an undetected overflow).
Therefore: please use that patch to get an idea which queries don't fail as expected, but DO NOT apply it (maybe except as a workaround to un-break the build in distribution packages for s390x and ppc64 until this bug is fixed).
BTW: This bug also seems to depend on the version of toolchain, libraries etc. - I see it when building for openSUSE Tumbleweed (rolling release), but in openSUSE Leap (stable release, which has older versions of everything) the queries fail as expected (= the tests succeed).
Christian Boltz
added a comment - The patch func_math_tests_ MDEV-26645 .diff I just attached "fixes" the test failure - but as you can see in the patched expected results, there are three queries that end up with a mathematically wrong result (probably an undetected overflow).
Therefore: please use that patch to get an idea which queries don't fail as expected, but DO NOT apply it (maybe except as a workaround to un-break the build in distribution packages for s390x and ppc64 until this bug is fixed).
BTW: This bug also seems to depend on the version of toolchain, libraries etc. - I see it when building for openSUSE Tumbleweed (rolling release), but in openSUSE Leap (stable release, which has older versions of everything) the queries fail as expected (= the tests succeed).
Otto Kekäläinen
added a comment - This test failure is still in 10.6, and visible on newly uploaded 10.6 in Debian builders on several non-amd64 archs:
mips64el: https://buildd.debian.org/status/fetch.php?pkg=mariadb-10.6&arch=mips64el&ver=1%3A10.6.5-1&stamp=1638072372&raw=0
mipsel: https://buildd.debian.org/status/fetch.php?pkg=mariadb-10.6&arch=mipsel&ver=1%3A10.6.5-1&stamp=1638124892&raw=0
powerpc: https://buildd.debian.org/status/fetch.php?pkg=mariadb-10.6&arch=powerpc&ver=1%3A10.6.5-1&stamp=1638053797&raw=0
riscv64: https://buildd.debian.org/status/fetch.php?pkg=mariadb-10.6&arch=riscv64&ver=1%3A10.6.5-1&stamp=1638104467&raw=0
Overview on latest build status in Debian at: https://buildd.debian.org/status/package.php?p=mariadb-10.6
This fails on 10.6.5 on x86_64 too, when compiling MariaDB using GCC-12 on openSUSE Tumbleweed. This is the log from the tests: mariadb-gcc12-x86_64.log.
Danilo Spinella
added a comment - This fails on 10.6.5 on x86_64 too, when compiling MariaDB using GCC-12 on openSUSE Tumbleweed. This is the log from the tests: mariadb-gcc12-x86_64.log .
I write the code that test for the overflow and it seams to work on most machines.
To fix this, I would need access to a machine where it fails so that I can test this in a debugger.
This couldeasily be the compiler that is trying to optimize an arithmetic operation that it shouldn't.
Michael Widenius
added a comment - I write the code that test for the overflow and it seams to work on most machines.
To fix this, I would need access to a machine where it fails so that I can test this in a debugger.
This couldeasily be the compiler that is trying to optimize an arithmetic operation that it shouldn't.
This is similar to MDEV-21977, which affected integer division.
An integer overflow in an expression like a+b or a-b is undefined behavior. The compiler is allowed to assume that no such overflow is possible, and optimize away some code accordingly. Apparently, such optimizations were implemented in GCC a little earlier for the POWER and s390x back-ends, and with GCC 12 it starts to affect the AMD64 backend as well.
Marko Mäkelä
added a comment - This is similar to MDEV-21977 , which affected integer division.
An integer overflow in an expression like a+b or a-b is undefined behavior. The compiler is allowed to assume that no such overflow is possible, and optimize away some code accordingly. Apparently, such optimizations were implemented in GCC a little earlier for the POWER and s390x back-ends, and with GCC 12 it starts to affect the AMD64 backend as well.
The fix is simple: Remove some special handling for WITH_UBSAN .
Otto Kekäläinen
added a comment - I backported this to 10.6 in Debian and it worked so far on both Launchpad and official Debian builders. Thanks!
https://salsa.debian.org/mariadb-team/mariadb-server/-/commit/4a96913a88cd868885813f20eb66524cbae85229
Bug
Minor
.
also on 10.5 https://buildbot.mariadb.org/#/builders/309/builds/937/steps/6/logs/stdio