Potential issue with strstr on x86 with sse4.2 in glibc-2.18

[Carlos O'Donell]

On 08/16/2013 08:16 AM, Allan McRae wrote:
> This is more of a heads up as on first glance I suspect this is a gcc
> issue rather than a glibc one.
> 
> I have had a number of reports for i686-pc-linux-gnu builds of glibc
> (with gcc 4.8.1 20130725) giving a segfault in __strstr_sse42.  Here is
> our bugtracker link: https://bugs.archlinux.org/task/36556 .  A bisect
> puts indicates this commit:
> 
> http://sourceware.org/git/?p=glibc.git;a=commit;h=f1d70dad
> 
> and in particular, reverting this change in
> sysdeps/x86_64/multiarch/strstr.c stops the issue:
> 
> -static inline __m128i
> +static __m128i
>  __m128i_strloadu (const unsigned char * p, __m128i zero)
> 
> 
> That's right...  adding an inline solves the segfaults, hence why I
> suspect a gcc issue.  Also, compiling with -O3 instead of -O2 avoids the
> issue.
> 
> Note that this code runs fine on x86_64.  We only notice it on our i686
> builds.

Thanks for the heads up.

We also just saw an i686 and gcc 4.8 inlining issue recently where
adding noinline fixed it. It might not be the same issue, but I just
wanted to add our data point. Our example is a little more complicated
though and harder to explain.

Cheers,
Carlos.