svn commit: r227812 - head/lib/libc/string

[mdf at FreeBSD.org]

On Mon, Nov 21, 2011 at 6:50 PM, Eitan Adler <eadler at freebsd.org> wrote:
> Author: eadler (ports committer)
> Date: Tue Nov 22 02:50:24 2011
> New Revision: 227812
> URL: http://svn.freebsd.org/changeset/base/227812
>
> Log:
>  - fix some style(9) nits with my last commit
>  - add a comment explaining why I used '|' instead of '||'
>
>  Submitted by: danfe@
>  Approved by:  emaste@
>
> Modified:
>  head/lib/libc/string/strcasecmp.c
>  head/lib/libc/string/strncmp.c
>
> Modified: head/lib/libc/string/strcasecmp.c
> ==============================================================================
> --- head/lib/libc/string/strcasecmp.c   Tue Nov 22 02:27:59 2011        (r227811)
> +++ head/lib/libc/string/strcasecmp.c   Tue Nov 22 02:50:24 2011        (r227812)
> @@ -49,7 +49,7 @@ strcasecmp_l(const char *s1, const char
>                        *us1 = (const u_char *)s1,
>                        *us2 = (const u_char *)s2;
>        if (s1 == s2)
> -           return (0);
> +               return (0);
>
>        FIX_LOCALE(locale);
>
> @@ -73,8 +73,9 @@ strncasecmp_l(const char *s1, const char
>                        *us1 = (const u_char *)s1,
>                        *us2 = (const u_char *)s2;
>
> -       if (( s1 == s2) | (n == 0))
> -           return (0);
> +       /* use a bitwise or to avoid an additional branch instruction */
> +       if ((s1 == s2) | (n == 0))
> +               return (0);

I guess I'm a little confused.  Do we really have profiling
information at this level that suggests the overhead of the branch is
significant?  I thought most hardware had pretty good
branch-prediction, particularly with speculative execution.

Wouldn't something like __predict_false() have more value for
performance, or is all this just guess-work?  I would much rather have
the code say what it means unless there's real, measurable performance
differences from doing otherwise.

Thanks,
matthew