On Fri, Nov 5, 2021 at 3:20 AM liuhongt <hongtao....@intel.com> wrote:
>
> > Note that this is not safe with -fsignaling-nans, so needs to be disabled
> > for that option (if there isn't already logic somewhere with that effect),
> > because the extend will convert a signaling NaN to quiet (raising
> > "invalid"), but copysign won't, so this transformation could result in a
> > signaling NaN being wrongly returned when the original code would never
> > have returned a signaling NaN.
> >
> > --
> > Joseph S. Myers
> > jos...@codesourcery.com
>
> Bootstrapped and regtested on x86_64-linux-gnu{-m32,}.
> Ok for trunk?
>
> gcc/ChangeLog
>
> * match.pd
> (Simplifcation (trunc)copysign((extend)a, (extend)b) to
> .COPYSIGN (a, b)): Add !flag_signaling_nans.
> ---
> gcc/match.pd | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index fb1065dc0e6..d6a8dd0dd20 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -6176,6 +6176,7 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (simplify
> (convert (copysigns (convert@2 @0) (convert @1)))
> (if (optimize
> + && !flag_signaling_nans
Please use !HONOR_SNANS (@2)
OK with that change.
> && types_match (type, TREE_TYPE (@0))
> && types_match (type, TREE_TYPE (@1))
> && TYPE_PRECISION (type) < TYPE_PRECISION (TREE_TYPE (@2))
> --
> 2.18.1
>
