RE: [PATCH] x86_64: new and improved memset()

[David Laight <David.Laight@ACULAB.COM>]

From: Alexey Dobriyan
> Sent: 14 September 2019 11:34
...
> +ENTRY(memset0_rep_stosq)
> +	xor	eax, eax
> +.globl memsetx_rep_stosq
> +memsetx_rep_stosq:
> +	lea	rsi, [rdi + rcx]
> +	shr	rcx, 3
> +	rep stosq
> +	cmp	rdi, rsi
> +	je	1f
> +2:
> +	mov	[rdi], al
> +	add	rdi, 1
> +	cmp	rdi, rsi
> +	jne	2b
> +1:
> +	ret

You can do the 'trailing bytes' first with a potentially misaligned store.
Something like (modulo asm syntax and argument ordering):
	lea	rsi, [rdi + rdx]
	shr	rcx, 3
	jcxz	1f		# Short buffer
	mov	-8[rsi], rax
	rep stosq
	ret
1:
	mov	[rdi], al
	add	rdi, 1
	cmp	rdi, rsi
	jne	1b
	ret

The final loop can be one instruction shorter by arranging to do:
1:
	mov	[rdi+rxx], al
	add	rdi, 1
	jnz	1b
	ret

Last I looked 'jcxz' was 'ok' on all recent amd and intel cpus.
OTOH 'loop' is horrid on intel ones.

The same applies to the other versions.

I suspect it isn't worth optimising to realign misaligned buffers
they are unlikely to happen often enough.

I also think that gcc's __builtin version does some of the short
buffer optimisations already.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)