Re: [PATCH] AArch64: Update A64FX memset not to degrade at 16KB

[Szabolcs Nagy via Libc-alpha <libc-alpha@sourceware.org>]

The 08/27/2021 05:03, Naohiro Tamura via Libc-alpha wrote:
> This patch updates unroll8 code so as not to degrade at the peak
> performance 16KB for both FX1000 and FX700.
> 
> Inserted 2 instructions at the beginning of the unroll8 loop,
> cmp and branch, are a workaround that is found heuristically.
> 
> Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>

thanks, i committed this now.


> ---
>  sysdeps/aarch64/multiarch/memset_a64fx.S | 9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/sysdeps/aarch64/multiarch/memset_a64fx.S b/sysdeps/aarch64/multiarch/memset_a64fx.S
> index 7bf759b6a753..f7dfdaace7cf 100644
> --- a/sysdeps/aarch64/multiarch/memset_a64fx.S
> +++ b/sysdeps/aarch64/multiarch/memset_a64fx.S
> @@ -96,7 +96,14 @@ L(vl_agnostic): // VL Agnostic
>  L(unroll8):
>  	sub	count, count, tmp1
>  	.p2align 4
> -1:	st1b_unroll 0, 7
> +	// The 2 instructions at the beginning of the following loop,
> +	// cmp and branch, are a workaround so as not to degrade at
> +	// the peak performance 16KB.
> +	// It is found heuristically and the branch condition, b.ne,
> +	// is chosen intentionally never to jump.
> +1:	cmp	xzr, xzr
> +	b.ne	1b
> +	st1b_unroll 0, 7
>  	add	dst, dst, tmp1
>  	subs	count, count, tmp1
>  	b.hi	1b
> -- 
> 2.17.1
>