Training courses

Kernel and Embedded Linux

Bootlin training courses

Embedded Linux, kernel,
Yocto Project, Buildroot, real-time,
graphics, boot time, debugging...

Bootlin logo

Elixir Cross Referencer

dnl  AMD64 mpn_com optimised for CPUs with fast SSE copying and SSSE3.

dnl  Copyright 2012, 2013, 2015 Free Software Foundation, Inc.

dnl  Contributed to the GNU project by Torbjorn Granlund.

dnl  This file is part of the GNU MP Library.
dnl
dnl  The GNU MP Library is free software; you can redistribute it and/or modify
dnl  it under the terms of either:
dnl
dnl    * the GNU Lesser General Public License as published by the Free
dnl      Software Foundation; either version 3 of the License, or (at your
dnl      option) any later version.
dnl
dnl  or
dnl
dnl    * the GNU General Public License as published by the Free Software
dnl      Foundation; either version 2 of the License, or (at your option) any
dnl      later version.
dnl
dnl  or both in parallel, as here.
dnl
dnl  The GNU MP Library is distributed in the hope that it will be useful, but
dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
dnl  for more details.
dnl
dnl  You should have received copies of the GNU General Public License and the
dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
dnl  see https://www.gnu.org/licenses/.

include(`../config.m4')

C	     cycles/limb     cycles/limb     cycles/limb      good
C              aligned	      unaligned	      best seen	     for cpu?
C AMD K8,K9	 2.0		 illop		1.0/1.0		N
C AMD K10	 0.85		 illop				Y/N
C AMD bd1	 1.39		 ? 1.45				Y/N
C AMD bd2     0.8-1.4	       0.7-1.4				Y
C AMD bd3
C AMD bd4
C AMD bobcat	 1.97		 ? 8.17		1.5/1.5		N
C AMD jaguar	 1.02		 1.02		0.91/0.91	N
C Intel P4	 2.26		 illop				Y/N
C Intel core	 0.58		 0.87		opt/0.74	Y
C Intel NHM	 0.64		 1.14		opt/bad		Y
C Intel SBR	 0.51		 0.65		opt/opt		Y
C Intel IBR	 0.50		 0.64		opt/0.57	Y
C Intel HWL	 0.51		 0.58		opt/opt		Y
C Intel BWL	 0.52		 0.64		opt/opt		Y
C Intel SKL	 0.51		 0.63		opt/opt		Y
C Intel atom	 1.16		 1.70		opt/opt		Y
C Intel SLM	 1.02		 1.52				N
C VIA nano	 1.09		 1.10		opt/opt		Y

C We use only 16-byte operations, except for unaligned top-most and bottom-most
C limbs.  We use the SSSE3 palignr instruction when rp - up = 8 (mod 16).  That
C instruction is better adapted to mpn_copyd's needs, we need to contort the
C code to use it here.
C
C For operands of < COM_SSE_THRESHOLD limbs, we use a plain 64-bit loop, taken
C from the x86_64 default code.

C INPUT PARAMETERS
define(`rp', `%rdi')
define(`up', `%rsi')
define(`n',  `%rdx')

C There are three instructions for loading an aligned 128-bit quantity.  We use
C movaps, since it has the shortest coding.
define(`movdqa', ``movaps'')

ifdef(`COM_SSE_THRESHOLD',`',`define(`COM_SSE_THRESHOLD', 7)')

ASM_START()
	TEXT
	ALIGN(64)
PROLOGUE(mpn_com)
	FUNC_ENTRY(3)

	cmp	$COM_SSE_THRESHOLD, n
	jbe	L(bc)

	pcmpeqb	%xmm5, %xmm5		C set to 111...111

	test	$8, R8(rp)		C is rp 16-byte aligned?
	jz	L(rp_aligned)		C jump if rp aligned

	mov	(up), %r8
	lea	8(up), up
	not	%r8
	mov	%r8, (rp)
	lea	8(rp), rp
	dec	n

L(rp_aligned):
	test	$8, R8(up)
	jnz	L(uent)

ifelse(eval(COM_SSE_THRESHOLD >= 8),1,
`	sub	$8, n',
`	jmp	L(am)')

	ALIGN(16)
L(atop):movdqa	0(up), %xmm0
	movdqa	16(up), %xmm1
	movdqa	32(up), %xmm2
	movdqa	48(up), %xmm3
	lea	64(up), up
	pxor	%xmm5, %xmm0
	pxor	%xmm5, %xmm1
	pxor	%xmm5, %xmm2
	pxor	%xmm5, %xmm3
	movdqa	%xmm0, (rp)
	movdqa	%xmm1, 16(rp)
	movdqa	%xmm2, 32(rp)
	movdqa	%xmm3, 48(rp)
	lea	64(rp), rp
L(am):	sub	$8, n
	jnc	L(atop)

	test	$4, R8(n)
	jz	1f
	movdqa	(up), %xmm0
	movdqa	16(up), %xmm1
	lea	32(up), up
	pxor	%xmm5, %xmm0
	pxor	%xmm5, %xmm1
	movdqa	%xmm0, (rp)
	movdqa	%xmm1, 16(rp)
	lea	32(rp), rp

1:	test	$2, R8(n)
	jz	1f
	movdqa	(up), %xmm0
	lea	16(up), up
	pxor	%xmm5, %xmm0
	movdqa	%xmm0, (rp)
	lea	16(rp), rp

1:	test	$1, R8(n)
	jz	1f
	mov	(up), %r8
	not	%r8
	mov	%r8, (rp)

1:	FUNC_EXIT()
	ret

L(uent):
C Code handling up - rp = 8 (mod 16)

C FIXME: The code below only handles overlap if it is close to complete, or
C quite separate: up-rp < 5 or up-up > 15 limbs
	lea	-40(up), %rax		C 40 = 5 * GMP_LIMB_BYTES
	sub	rp, %rax
	cmp	$80, %rax		C 80 = (15-5) * GMP_LIMB_BYTES
	jbe	L(bc)			C deflect to plain loop

	sub	$16, n
	jc	L(uend)

	movdqa	120(up), %xmm3

	sub	$16, n
	jmp	L(um)

	ALIGN(16)
L(utop):movdqa	120(up), %xmm3
	pxor	%xmm5, %xmm0
	movdqa	%xmm0, -128(rp)
	sub	$16, n
L(um):	movdqa	104(up), %xmm2
	palignr($8, %xmm2, %xmm3)
	movdqa	88(up), %xmm1
	pxor	%xmm5, %xmm3
	movdqa	%xmm3, 112(rp)
	palignr($8, %xmm1, %xmm2)
	movdqa	72(up), %xmm0
	pxor	%xmm5, %xmm2
	movdqa	%xmm2, 96(rp)
	palignr($8, %xmm0, %xmm1)
	movdqa	56(up), %xmm3
	pxor	%xmm5, %xmm1
	movdqa	%xmm1, 80(rp)
	palignr($8, %xmm3, %xmm0)
	movdqa	40(up), %xmm2
	pxor	%xmm5, %xmm0
	movdqa	%xmm0, 64(rp)
	palignr($8, %xmm2, %xmm3)
	movdqa	24(up), %xmm1
	pxor	%xmm5, %xmm3
	movdqa	%xmm3, 48(rp)
	palignr($8, %xmm1, %xmm2)
	movdqa	8(up), %xmm0
	pxor	%xmm5, %xmm2
	movdqa	%xmm2, 32(rp)
	palignr($8, %xmm0, %xmm1)
	movdqa	-8(up), %xmm3
	pxor	%xmm5, %xmm1
	movdqa	%xmm1, 16(rp)
	palignr($8, %xmm3, %xmm0)
	lea	128(up), up
	lea	128(rp), rp
	jnc	L(utop)

	pxor	%xmm5, %xmm0
	movdqa	%xmm0, -128(rp)

L(uend):test	$8, R8(n)
	jz	1f
	movdqa	56(up), %xmm3
	movdqa	40(up), %xmm2
	palignr($8, %xmm2, %xmm3)
	movdqa	24(up), %xmm1
	pxor	%xmm5, %xmm3
	movdqa	%xmm3, 48(rp)
	palignr($8, %xmm1, %xmm2)
	movdqa	8(up), %xmm0
	pxor	%xmm5, %xmm2
	movdqa	%xmm2, 32(rp)
	palignr($8, %xmm0, %xmm1)
	movdqa	-8(up), %xmm3
	pxor	%xmm5, %xmm1
	movdqa	%xmm1, 16(rp)
	palignr($8, %xmm3, %xmm0)
	lea	64(up), up
	pxor	%xmm5, %xmm0
	movdqa	%xmm0, (rp)
	lea	64(rp), rp

1:	test	$4, R8(n)
	jz	1f
	movdqa	24(up), %xmm1
	movdqa	8(up), %xmm0
	palignr($8, %xmm0, %xmm1)
	movdqa	-8(up), %xmm3
	pxor	%xmm5, %xmm1
	movdqa	%xmm1, 16(rp)
	palignr($8, %xmm3, %xmm0)
	lea	32(up), up
	pxor	%xmm5, %xmm0
	movdqa	%xmm0, (rp)
	lea	32(rp), rp

1:	test	$2, R8(n)
	jz	1f
	movdqa	8(up), %xmm0
	movdqa	-8(up), %xmm3
	palignr($8, %xmm3, %xmm0)
	lea	16(up), up
	pxor	%xmm5, %xmm0
	movdqa	%xmm0, (rp)
	lea	16(rp), rp

1:	test	$1, R8(n)
	jz	1f
	mov	(up), %r8
	not	%r8
	mov	%r8, (rp)

1:	FUNC_EXIT()
	ret

C Basecase code.  Needed for good small operands speed, not for
C correctness as the above code is currently written.

L(bc):	lea	-8(rp), rp
	sub	$4, R32(n)
	jc	L(end)

ifelse(eval(1 || COM_SSE_THRESHOLD >= 8),1,
`	ALIGN(16)')
L(top):	mov	(up), %r8
	mov	8(up), %r9
	lea	32(rp), rp
	mov	16(up), %r10
	mov	24(up), %r11
	lea	32(up), up
	not	%r8
	not	%r9
	not	%r10
	not	%r11
	mov	%r8, -24(rp)
	mov	%r9, -16(rp)
ifelse(eval(1 || COM_SSE_THRESHOLD >= 8),1,
`	sub	$4, R32(n)')
	mov	%r10, -8(rp)
	mov	%r11, (rp)
ifelse(eval(1 || COM_SSE_THRESHOLD >= 8),1,
`	jnc	L(top)')

L(end):	test	$1, R8(n)
	jz	1f
	mov	(up), %r8
	not	%r8
	mov	%r8, 8(rp)
	lea	8(rp), rp
	lea	8(up), up
1:	test	$2, R8(n)
	jz	1f
	mov	(up), %r8
	mov	8(up), %r9
	not	%r8
	not	%r9
	mov	%r8, 8(rp)
	mov	%r9, 16(rp)
1:	FUNC_EXIT()
	ret
EPILOGUE()