[jira] [Commented] (STDCXX-1071) numpunct facet cache initialization is not thread -safe
[ https://issues.apache.org/jira/browse/STDCXX-1071?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13464667#comment-13464667 ] Liviu Nicoara commented on STDCXX-1071: --- I have re-attached the timing measurements, file patch-timings.tgz, after re-verifying the results. They show an improved performance of the facet when not caching the locale data, in multi-threaded builds, with high contention. > numpunct facet cache initialization is not thread -safe > --- > > Key: STDCXX-1071 > URL: https://issues.apache.org/jira/browse/STDCXX-1071 > Project: C++ Standard Library > Issue Type: Bug > Components: 22. Localization, Thread Safety >Affects Versions: 4.2.1, 4.2.x, 4.3.x, 5.0.0 > Environment: $ uname -a; CC -V > Linux behemoth 2.6.37.6 #3 SMP Sat Apr 9 22:49:32 CDT 2011 x86_64 AMD > Opteron(tm) Processor 6134 AuthenticAMD GNU/Linux > CC: Sun C++ 5.12 Linux_i386 2011/11/16 >Reporter: Liviu Nicoara >Priority: Critical > Labels: DCII, localization, numpunct, thread-safety > Fix For: 4.2.1, 4.2.x, 4.3.x, 5.0.0 > > Attachments: library-reduction.cpp, patch-4.2.x.diff, > patch-4.3.x.diff, patch-timings.tgz, punct-mt.cpp > > Original Estimate: 8h > Remaining Estimate: 8h > > The bug affects the initialization of the cached data in numpunct facet. The > failing test is 22.locale.numpunct.cpp in certain builds: Linux with SUNPro > 5.12, optimized, most likely to fail. It has been reported (see linked > incidents) to fail on SPARC as well. It manifests in variate ways: sometimes > as heap corruption, sometimes as access violations, etc. Test cases and > library reduction programs to follow. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (STDCXX-1071) numpunct facet cache initialization is not thread -safe
[ https://issues.apache.org/jira/browse/STDCXX-1071?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13464302#comment-13464302 ] Liviu Nicoara commented on STDCXX-1071: --- I have attached two minimal patches, one for 4.2.x and one for 4.3.x. They both eliminate the facet data caching. The former is binary compatible with 4.2.1, the other not. Application of either eliminates the failures in all the builds I have performed. As recorded in measurements performed for STDCXX-1056, perfect forwarding in the facet interface actually improves performance in multi-threaded environ. However, performance is worse in single-threaded builds. > numpunct facet cache initialization is not thread -safe > --- > > Key: STDCXX-1071 > URL: https://issues.apache.org/jira/browse/STDCXX-1071 > Project: C++ Standard Library > Issue Type: Bug > Components: 22. Localization, Thread Safety >Affects Versions: 4.2.1, 4.2.x, 4.3.x, 5.0.0 > Environment: $ uname -a; CC -V > Linux behemoth 2.6.37.6 #3 SMP Sat Apr 9 22:49:32 CDT 2011 x86_64 AMD > Opteron(tm) Processor 6134 AuthenticAMD GNU/Linux > CC: Sun C++ 5.12 Linux_i386 2011/11/16 >Reporter: Liviu Nicoara >Priority: Critical > Labels: DCII, localization, numpunct, thread-safety > Fix For: 4.2.1, 4.2.x, 4.3.x, 5.0.0 > > Attachments: library-reduction.cpp, patch-4.2.x.diff, > patch-4.3.x.diff, punct-mt.cpp > > Original Estimate: 8h > Remaining Estimate: 8h > > The bug affects the initialization of the cached data in numpunct facet. The > failing test is 22.locale.numpunct.cpp in certain builds: Linux with SUNPro > 5.12, optimized, most likely to fail. It has been reported (see linked > incidents) to fail on SPARC as well. It manifests in variate ways: sometimes > as heap corruption, sometimes as access violations, etc. Test cases and > library reduction programs to follow. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (STDCXX-1071) numpunct facet cache initialization is not thread -safe
[ https://issues.apache.org/jira/browse/STDCXX-1071?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13464298#comment-13464298 ] Liviu Nicoara commented on STDCXX-1071: --- The file punct-mt.cpp is a multi-threaded program that tests either the numpunct or the moneypunct facet accessor interface. If TEST_MONEYPUNCT is defined, the test tests moneypunct. The test fails with 4.2.x and presumably with all others. The moneypunct version does not fail in my local runs. I have used SUNPro 5.12 on Linux, but I noticed failures with gcc 4.7.1 as well. The test case is pointing at the numpunct facet caching code. > numpunct facet cache initialization is not thread -safe > --- > > Key: STDCXX-1071 > URL: https://issues.apache.org/jira/browse/STDCXX-1071 > Project: C++ Standard Library > Issue Type: Bug > Components: 22. Localization, Thread Safety >Affects Versions: 4.2.1, 4.2.x, 4.3.x, 5.0.0 > Environment: $ uname -a; CC -V > Linux behemoth 2.6.37.6 #3 SMP Sat Apr 9 22:49:32 CDT 2011 x86_64 AMD > Opteron(tm) Processor 6134 AuthenticAMD GNU/Linux > CC: Sun C++ 5.12 Linux_i386 2011/11/16 >Reporter: Liviu Nicoara >Priority: Critical > Labels: DCII, localization, numpunct, thread-safety > Fix For: 4.2.1, 4.2.x, 4.3.x, 5.0.0 > > Attachments: library-reduction.cpp, punct-mt.cpp > > Original Estimate: 8h > Remaining Estimate: 8h > > The bug affects the initialization of the cached data in numpunct facet. The > failing test is 22.locale.numpunct.cpp in certain builds: Linux with SUNPro > 5.12, optimized, most likely to fail. It has been reported (see linked > incidents) to fail on SPARC as well. It manifests in variate ways: sometimes > as heap corruption, sometimes as access violations, etc. Test cases and > library reduction programs to follow. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (STDCXX-1071) numpunct facet cache initialization is not thread -safe
[
https://issues.apache.org/jira/browse/STDCXX-1071?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13464295#comment-13464295
]
Liviu Nicoara commented on STDCXX-1071:
---
The file library-reduction.cpp is a reduction of the numpunct library code down
to the essentials that still have the defect. The test case is intentionally
verbose (180LOC) so that it still shows the organization of the library code.
It exhibits the same failures as the original library code it has been
extracted from. It takes two defines on the command line:
1. NO_USE_STDCXX_LOCALES -- if defined it mimics the setting of the facet data
from the system locales
2. NO_USE_NUMPUNCT_CACHE -- if set, it does perfect forwarding, i.e., it does
not cache "locale data" in the facet
The runs, with SUNPro 5.12 on Linux, exhibit failures (renamed to t.cpp):
{noformat}
$ CC -O -mt t.cpp && ./a.out
"t.cpp", line 113: Warning: String literal converted to char* in assignment.
1 Warning(s) detected.
CC: Warning: failed to detect system linker version, falling back to custom
linker usage
16, 1000
*** glibc detected *** ./a.out: double free or corruption (fasttop):
0x00622b40 ***
*** glibc detected *** ./a.out: double free or corruption (fasttop):
0x7f6654000920 ***
=== Backtrace: =
/lib64/libc.so.6(+0x78f85)[0x7f666c031f85]
/lib64/libc.so.6(cfree+0x73)[0x7f666c035d93]
...
{noformat}
{noformat}
$ CC -DNO_USE_STDCXX_LOCALES -O -mt t.cpp && ./a.out
"t.cpp", line 113: Warning: String literal converted to char* in assignment.
1 Warning(s) detected.
CC: Warning: failed to detect system linker version, falling back to custom
linker usage
16, 1000
*** glibc detected *** ./a.out: double free or corruption (fasttop):
0x00622b40 ***
*** glibc detected *** ./a.out: double free or corruption (fasttop):
0x7fda68b0 ***
=== Backtrace: =
/lib64/libc.so.6=== Backtrace: =
/lib64/libc.so.6(cfree+0x73)[0x7fda6c8f2d93]
/lib64/libc.so.6(+0x78f85)[0x7fda6c8eef85]
The reduced program did not fail in my runs with NO_USE_NUMPUNCT_CACHE defined.
> numpunct facet cache initialization is not thread -safe
> ---
>
> Key: STDCXX-1071
> URL: https://issues.apache.org/jira/browse/STDCXX-1071
> Project: C++ Standard Library
> Issue Type: Bug
> Components: 22. Localization, Thread Safety
>Affects Versions: 4.2.1, 4.2.x, 4.3.x, 5.0.0
> Environment: $ uname -a; CC -V
> Linux behemoth 2.6.37.6 #3 SMP Sat Apr 9 22:49:32 CDT 2011 x86_64 AMD
> Opteron(tm) Processor 6134 AuthenticAMD GNU/Linux
> CC: Sun C++ 5.12 Linux_i386 2011/11/16
>Reporter: Liviu Nicoara
>Priority: Critical
> Labels: DCII, localization, numpunct, thread-safety
> Fix For: 4.2.1, 4.2.x, 4.3.x, 5.0.0
>
> Attachments: library-reduction.cpp
>
> Original Estimate: 8h
> Remaining Estimate: 8h
>
> The bug affects the initialization of the cached data in numpunct facet. The
> failing test is 22.locale.numpunct.cpp in certain builds: Linux with SUNPro
> 5.12, optimized, most likely to fail. It has been reported (see linked
> incidents) to fail on SPARC as well. It manifests in variate ways: sometimes
> as heap corruption, sometimes as access violations, etc. Test cases and
> library reduction programs to follow.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
