The gnulib CI failed today, with this error:
depbase=`echo c-vsnprintf.o | sed 's|[^/]*$|.deps/&|;s|\.o$||'`;\
ccache gcc -DHAVE_CONFIG_H -DEXEEXT=\"\" -DEXEEXT=\"\" -DNO_XMALLOC
-DEXEEXT=\"\" -I. -I.. -DGNULIB_STRICT_CHECKING=1 -fvisibility=hidden -g -O2
-MT c-vsnprintf.o -MD -MP -MF $depbase
On 2023-10-02 04:33, Bruno Haible wrote:
But your portable code appears to produce fairly decent x86_64 assembly
code as well.
Thanks for your fixes. Yes, the goal was portability rather than saving
a few instructions. A high-performance implementation of totalorder
would need help from the c
> * Solaris
> x86_64: ABIs 32, 64 on Solaris 10..., Solaris 11.4
> except ABI 64 with cc on Solaris 10: FAIL: test-totalorderl
The cause of this test failure is that the initialization of xu, yu, zu
in
union { unsigned long long i[2]; long double f; } volatile
xu = {0}, yu = {0}
Since this is obviously CPU dependent code, I tested it on various platforms.
There were failures only on two platforms (see 'FAIL' below); the tests
passed on all other platforms.
* Linux/glibc
x86_64: Ubuntu 22.04, Fedora 31, Debian 9.6
x86: Ubuntu 22.04, Debian 9.1
x86_64-x32: Ubuntu 22.
Note: it looks like the libstdc++ has similar functions, with a templated
implementation:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96526
https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=9805965e3551b66b5bd751d6076791d00cdeb137
Here's a discussion how to get nearly optimal code for this functi
