[PATCH] powerpc: P9 vector load instruction change in memcpy and memmove

[Florian Weimer]

On 10/19/2017 05:25 PM, Tulio Magno Quites Machado Filho wrote:
> diff --git a/sysdeps/powerpc/powerpc64/multiarch/bcopy.c b/sysdeps/powerpc/powerpc64/multiarch/bcopy.c
> index 05d46e2..4a4ee6e 100644
> --- a/sysdeps/powerpc/powerpc64/multiarch/bcopy.c
> +++ b/sysdeps/powerpc/powerpc64/multiarch/bcopy.c
> @@ -22,8 +22,12 @@
>   extern __typeof (bcopy) __bcopy_ppc attribute_hidden;
>   /* __bcopy_power7 symbol is implemented at memmove-power7.S  */
>   extern __typeof (bcopy) __bcopy_power7 attribute_hidden;
> +/* __bcopy_power9 symbol is implemented at memmove-power9.S.  */
> +extern __typeof (bcopy) __bcopy_power9 attribute_hidden;
>   
>   libc_ifunc (bcopy,
> -            (hwcap & PPC_FEATURE_HAS_VSX)
> +	    (hwcap2 & PPC_FEATURE2_ARCH_3_00)
> +	    ? __bcopy_power9
> +	    : (hwcap & PPC_FEATURE_HAS_VSX)
>               ? __bcopy_power7
>               : __bcopy_ppc);
> diff --git a/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c b/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
> index 6a88536..9040bbc 100644
> --- a/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
> +++ b/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
> @@ -51,6 +51,8 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
>   #ifdef SHARED
>     /* Support sysdeps/powerpc/powerpc64/multiarch/memcpy.c.  */
>     IFUNC_IMPL (i, name, memcpy,
> +	      IFUNC_IMPL_ADD (array, i, memcpy, hwcap2 & PPC_FEATURE2_ARCH_3_00,
> +			      __memcpy_power9)
>   	      IFUNC_IMPL_ADD (array, i, memcpy, hwcap & PPC_FEATURE_HAS_VSX,
>   			      __memcpy_power7)
>   	      IFUNC_IMPL_ADD (array, i, memcpy, hwcap & PPC_FEATURE_ARCH_2_06,
> @@ -65,6 +67,8 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
>   
>     /* Support sysdeps/powerpc/powerpc64/multiarch/memmove.c.  */
>     IFUNC_IMPL (i, name, memmove,
> +	      IFUNC_IMPL_ADD (array, i, memmove, hwcap2 & PPC_FEATURE2_ARCH_3_00,
> +			      __memmove_power9)
>   	      IFUNC_IMPL_ADD (array, i, memmove, hwcap & PPC_FEATURE_HAS_VSX,
>   			      __memmove_power7)
>   	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_ppc))
> @@ -168,6 +172,8 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
>   
>     /* Support sysdeps/powerpc/powerpc64/multiarch/bcopy.c.  */
>     IFUNC_IMPL (i, name, bcopy,
> +	      IFUNC_IMPL_ADD (array, i, bcopy, hwcap2 & PPC_FEATURE2_ARCH_3_00,
> +			      __bcopy_power9)
>   	      IFUNC_IMPL_ADD (array, i, bcopy, hwcap & PPC_FEATURE_HAS_VSX,
>   			      __bcopy_power7)
>   	      IFUNC_IMPL_ADD (array, i, bcopy, 1, __bcopy_ppc))

I'm concerned that this needs *another* change to recognize post-DD2.1 
POWER 9 hardware which has the fix (or perhaps OpenPOWER implementations 
which never had the bug).

Is there any other way you could select the workaround that is more 
specific?  Or is the performance hit from avoiding the affected vector 
instructions not so severe that it would be an immediate concern for 
future silicon?

Thanks,
Florian