[PATCH] aarch64: Fix SVE tanf mixed-lane results [BZ #34922]

Adhemerval Zanella Netto adhemerval.zanella@linaro.org
Wed Nov 4 15:38:19 GMT 2026



On 04/11/26 12:06, Dylan Fleming wrote:
> The special-case path reciprocates every lane before calling scalar tanf
> for large inputs. This also changes ordinary lanes which do not need the
> reciprocal, so inputs [0.0f, 0x1p15f, ...] return infinity in the first
> lane instead of zero.

This requires a regression testcase, as we do for all other math function.

> 
> Fixed by passing the alternate-lane predicate to special_case and selecting
> the reciprocal only where needed before handling the special lanes.
> ---
>  SHARED-FILES                   | 2 +-
>  sysdeps/aarch64/fpu/sve/tanf.c | 6 ++++--
>  2 files changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/SHARED-FILES b/SHARED-FILES
> index 8f89400177..2c6e4546dd 100644
> --- a/SHARED-FILES
> +++ b/SHARED-FILES
> @@ -573,7 +573,7 @@ arm-optimized-routines:
>    sysdeps/aarch64/fpu/sve/sinpif.c
>    # math/aarch64/sve/tan.c, revision 1aad38d
>    sysdeps/aarch64/fpu/sve/tan.c
> -  # math/aarch64/sve/tanf.c, revision 1aad38d
> +  # math/aarch64/sve/tanf.c, revision 5e51e52
>    sysdeps/aarch64/fpu/sve/tanf.c
>    # math/aarch64/sve/tanh.c, revision 1aad38d
>    sysdeps/aarch64/fpu/sve/tanh.c
> diff --git a/sysdeps/aarch64/fpu/sve/tanf.c b/sysdeps/aarch64/fpu/sve/tanf.c
> index 60d93fd7e2..4a627d31e0 100644
> --- a/sysdeps/aarch64/fpu/sve/tanf.c
> +++ b/sysdeps/aarch64/fpu/sve/tanf.c
> @@ -47,8 +47,10 @@ static const struct data
>  };
>  
>  static svfloat32_t NOINLINE
> -special_case (svfloat32_t x, svfloat32_t y, svbool_t cmp)
> +special_case (svfloat32_t x, svfloat32_t y, svbool_t cmp, svbool_t pred_alt)
>  {
> +  svfloat32_t inv_y = svdivr_x (svptrue_b32 (), y, 1.0f);
> +  y = svsel (pred_alt, inv_y, y);
>    return sv_call_f32 (tanf, x, y, cmp);
>  }
>  
> @@ -107,7 +109,7 @@ svfloat32_t SV_NAME_F1 (tan) (svfloat32_t x, const svbool_t pg)
>    /* Determine whether input is too large to perform fast regression.  */
>    svbool_t cmp = svacge (pg, x, d->range_val);
>    if (__glibc_unlikely (svptest_any (pg, cmp)))
> -    return special_case (x, svdivr_x (pg, y, 1.0f), cmp);
> +    return special_case (x, y, cmp, pred_alt);
>  
>    svfloat32_t inv_y = svdivr_x (pg, y, 1.0f);
>    return svsel (pred_alt, inv_y, y);



More information about the Libc-alpha mailing list