old glibc forcing syscall

[brendene]

Comment:

Are there any plans to bump the glibc version this package compiles against to 2.17? Or possibly adding the compile flag: --enable-libstdcxx-time

When libstdc++ is compiled against a glibc older than 2.17 there is a compatibility defines set (_GLIBCXX_USE_CLOCK_GETTIME_SYSCALL) in the header so that the built libstdc++ library forces std::chrono::system_clock::now() to use the syscall version of clock_gettime instead of one that can be accelerated by vdso. (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59177)

The effect of this any code built in an environment using this package has a fairly large performance penalty when running querying the system clock this way.

From https://gcc.gnu.org/onlinedocs/libstdc++/manual/configure.html

For Linux targets, if clock_gettime is not used then the [time.clock] implementation
will use a system call to access the realtime and monotonic clocks, which is
significantly slower than the C library's clock_gettime function.

To see if it is set:

$ grep _GLIBCXX_USE_CLOCK_GETTIME_SYSCALL pkgs/libstdcxx-devel_linux-64-12.2.0-h3b97bd3_19/x86_64-conda-linux-gnu/include/c++/12.2.0/x86_64-conda-linux-gnu/bits/c++config.h
#define _GLIBCXX_USE_CLOCK_GETTIME_SYSCALL 1

Thank you

[apmorton]

+1 on this issue.

I think to work around this we need to add "gcc built with sysroot 2.17 or higher for linux64" to the compiler build matrix.

[brendene]

It looks like once conda forge is switched to alma 8 the default glibc will be 2.28 (conda-forge/conda-forge.github.io#1941) Does this feedstock plan to follow the glibc update as well or will it still be pinned at 2.12?

[h-vetinari]

It looks like once conda forge is switched to alma 8 the default glibc will be 2.28 (conda-forge/conda-forge.github.io#1941)

That issue is for introducing the Alma 8 glibc, but we're very far from making that the default. We've announced to increase the default glibc to 2.17 this July, though there's still ongoing work to enable that transition.

[brendene]

Something changed between the release on 2023-06-03 and 2023-09-16 that fixed the #define, its not clear to me which commit it is, possibly 845b9c6

If the change is in fact due to gcc changing from 13.1 to 13.2, then this this commit (gcc-mirror/gcc@1e3e6c7) from Nov 2020 might change the behavior but I am not sure why it wasn't pulled in earier.

https://anaconda.org/conda-forge/libstdcxx-devel_linux-64/12.3.0/download/linux-64/libstdcxx-devel_linux-64-12.3.0-h8bca6fd_0.conda

$ grep _GLIBCXX_USE_CLOCK_GETTIME_SYSCALL ./x86_64-conda-linux-gnu/include/c++/12.3.0/x86_64-conda-linux-gnu/bits/c++config.h
#define _GLIBCXX_USE_CLOCK_GETTIME_SYSCALL 1

https://anaconda.org/conda-forge/libstdcxx-devel_linux-64/12.3.0/download/linux-64/libstdcxx-devel_linux-64-12.3.0-h8bca6fd_1.conda

$ grep _GLIBCXX_USE_CLOCK_GETTIME_SYSCALL ./x86_64-conda-linux-gnu/include/c++/12.3.0/x86_64-conda-linux-gnu/bits/c++config.h
/* #undef _GLIBCXX_USE_CLOCK_GETTIME_SYSCALL */

This is actually the only difference between the c++config.h