[FFmpeg-devel] [PATCH 3/7] avutil/common: Implement av_sat_add64_c() with fewer branches

[Andreas Rheinhardt]

Michael Niedermayer:
> No benchmark because this is not used in any speed relevant pathes nor is it
> used where __builtin_add_overflow is available.
> So I do not know how to realistically benchmark it.
> 
> Signed-off-by: Michael Niedermayer <michael at niedermayer.cc>
> ---
>  libavutil/common.h | 9 ++++-----
>  1 file changed, 4 insertions(+), 5 deletions(-)
> 
> diff --git a/libavutil/common.h b/libavutil/common.h
> index 92b721a59c..a48c0648f8 100644
> --- a/libavutil/common.h
> +++ b/libavutil/common.h
> @@ -303,11 +303,10 @@ static av_always_inline int64_t av_sat_add64_c(int64_t a, int64_t b) {
>      int64_t tmp;
>      return !__builtin_add_overflow(a, b, &tmp) ? tmp : (tmp < 0 ? INT64_MAX : INT64_MIN);
>  #else
> -    if (b >= 0 && a >= INT64_MAX - b)
> -        return INT64_MAX;
> -    if (b <= 0 && a <= INT64_MIN - b)
> -        return INT64_MIN;
> -    return a + b;
> +    int64_t s = a+(uint64_t)b;
> +    if ((int64_t)(a^b | ~s^b) >= 0)
> +        return b < 0 ? INT64_MIN : INT64_MAX;
> +    return s;
>  #endif
>  }
>  
> 
Returning INT64_MAX ^ (b >> 63) would have even less branches
(untested). See https://godbolt.org/z/oYadG3

- Andreas