Home | History | Annotate | Line # | Download | only in arm64 
      1 dnl  ARM64 mpn_com.
      2 
      3 dnl  Copyright 2013, 2014 Free Software Foundation, Inc.
      4 
      5 dnl  This file is part of the GNU MP Library.
      6 dnl
      7 dnl  The GNU MP Library is free software; you can redistribute it and/or modify
      8 dnl  it under the terms of either:
      9 dnl
     10 dnl    * the GNU Lesser General Public License as published by the Free
     11 dnl      Software Foundation; either version 3 of the License, or (at your
     12 dnl      option) any later version.
     13 dnl
     14 dnl  or
     15 dnl
     16 dnl    * the GNU General Public License as published by the Free Software
     17 dnl      Foundation; either version 2 of the License, or (at your option) any
     18 dnl      later version.
     19 dnl
     20 dnl  or both in parallel, as here.
     21 dnl
     22 dnl  The GNU MP Library is distributed in the hope that it will be useful, but
     23 dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
     24 dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
     25 dnl  for more details.
     26 dnl
     27 dnl  You should have received copies of the GNU General Public License and the
     28 dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
     29 dnl  see https://www.gnu.org/licenses/.
     30 
     31 include(`../config.m4')
     32 
     33 C	     cycles/limb
     34 C Cortex-A53	 2.25
     35 C Cortex-A57	 1.25
     36 C X-Gene	 1.75
     37 
     38 changecom(blah)
     39 
     40 define(`rp', `x0')
     41 define(`up', `x1')
     42 define(`n',  `x2')
     43 
     44 ASM_START()
     45 PROLOGUE(mpn_com)
     46 	cmp	n, #3
     47 	b.le	L(bc)
     48 
     49 C Copy until rp is 128-bit aligned
     50 	tbz	rp, #3, L(al2)
     51 	ld1	{v22.1d}, [up], #8
     52 	sub	n, n, #1
     53 	mvn	v22.8b, v22.8b
     54 	st1	{v22.1d}, [rp], #8
     55 
     56 L(al2):	ld1	{v26.2d}, [up], #16
     57 	subs	n, n, #6
     58 	b.lt	L(end)
     59 
     60 	ALIGN(16)
     61 L(top):	ld1	{v22.2d}, [up], #16
     62 	mvn	v26.16b, v26.16b
     63 	st1	{v26.2d}, [rp], #16
     64 	ld1	{v26.2d}, [up], #16
     65 	mvn	v22.16b, v22.16b
     66 	st1	{v22.2d}, [rp], #16
     67 	subs	n, n, #4
     68 	b.ge	L(top)
     69 
     70 L(end):	mvn	v26.16b, v26.16b
     71 	st1	{v26.2d}, [rp], #16
     72 
     73 C Copy last 0-3 limbs.  Note that rp is aligned after loop, but not when we
     74 C arrive here via L(bc)
     75 L(bc):	tbz	n, #1, L(tl1)
     76 	ld1	{v22.2d}, [up], #16
     77 	mvn	v22.16b, v22.16b
     78 	st1	{v22.2d}, [rp], #16
     79 L(tl1):	tbz	n, #0, L(tl2)
     80 	ld1	{v22.1d}, [up]
     81 	mvn	v22.8b, v22.8b
     82 	st1	{v22.1d}, [rp]
     83 L(tl2):	ret
     84 EPILOGUE()
     85