[PATCH] aarch64: Fix SVE log1p helper mixed-lane shortcut [BZ #34923]
Adhemerval Zanella Netto
adhemerval.zanella@linaro.org
Fri Nov 6 14:55:28 GMT 2026
On 04/11/26 12:06, Dylan Fleming wrote:
> The log1p helper only selects the original input for k == 0 when all
> active lanes have k == 0. If another lane needs exponent reduction, the
> shortcut is skipped even though the correction term is already zeroed
> for k == 0. This loses accuracy in SVE acosh near 1, for example with
> inputs [0x1.0000000000001p0, 2.0, ...].
>
> Select the original input for k == 0 independently in each lane by
> removing the vector-wide condition around svsel.
LGTM, thanks.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
> ---
> SHARED-FILES | 2 +-
> sysdeps/aarch64/fpu/sve/sv_log1p_inline.h | 5 +----
> 2 files changed, 2 insertions(+), 5 deletions(-)
>
> diff --git a/SHARED-FILES b/SHARED-FILES
> index 8f89400177..1f176aa6f8 100644
> --- a/SHARED-FILES
> +++ b/SHARED-FILES
> @@ -591,7 +591,7 @@ arm-optimized-routines:
> sysdeps/aarch64/fpu/sve/sv_expf_special_inline.h
> # math/aarch64/sve/sv_expm1f_inline.h, revision 78cd003
> sysdeps/aarch64/fpu/sve/sv_expm1f_inline.h
> - # math/aarch64/sve/sv_log1p_inline.h, revision 7d08237
> + # math/aarch64/sve/sv_log1p_inline.h, revision 5e51e52
> sysdeps/aarch64/fpu/sve/sv_log1p_inline.h
> # math/aarch64/sve/sv_log1pf_inline.h, revision a553540
> sysdeps/aarch64/fpu/sve/sv_log1pf_inline.h
> diff --git a/sysdeps/aarch64/fpu/sve/sv_log1p_inline.h b/sysdeps/aarch64/fpu/sve/sv_log1p_inline.h
> index 1b2c542851..2e17280cf6 100644
> --- a/sysdeps/aarch64/fpu/sve/sv_log1p_inline.h
> +++ b/sysdeps/aarch64/fpu/sve/sv_log1p_inline.h
> @@ -102,10 +102,7 @@ sv_log1p_inline (svfloat64_t x, const svbool_t pg)
> that the approximation is solely the polynomial. */
> svbool_t knot0 = svcmpne (pg, k, 0);
> cm = svdiv_z (knot0, c, m);
> - if (__glibc_likely (!svptest_any (pg, knot0)))
> - {
> - f = svsel (knot0, f, x);
> - }
> + f = svsel (knot0, f, x);
> #else
> /* No shortcut. */
> cm = svdiv_x (pg, c, m);
More information about the Libc-alpha
mailing list