1dnl  ARM64 mpn_invert_limb -- Invert a normalized limb.
2
3dnl  Contributed to the GNU project by Torbj��rn Granlund.
4
5dnl  Copyright 2013 Free Software Foundation, Inc.
6
7dnl  This file is part of the GNU MP Library.
8dnl
9dnl  The GNU MP Library is free software; you can redistribute it and/or modify
10dnl  it under the terms of either:
11dnl
12dnl    * the GNU Lesser General Public License as published by the Free
13dnl      Software Foundation; either version 3 of the License, or (at your
14dnl      option) any later version.
15dnl
16dnl  or
17dnl
18dnl    * the GNU General Public License as published by the Free Software
19dnl      Foundation; either version 2 of the License, or (at your option) any
20dnl      later version.
21dnl
22dnl  or both in parallel, as here.
23dnl
24dnl  The GNU MP Library is distributed in the hope that it will be useful, but
25dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
26dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
27dnl  for more details.
28dnl
29dnl  You should have received copies of the GNU General Public License and the
30dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
31dnl  see https://www.gnu.org/licenses/.
32
33include(`../config.m4')
34
35C            cycles/limb
36C Cortex-A53     ?
37C Cortex-A57     ?
38
39C Compiler generated, mildly edited.  Could surely be further optimised.
40
41ASM_START()
42PROLOGUE(mpn_invert_limb)
43	lsr	x2, x0, #54
44	LEA_HI(	x1, approx_tab)
45	and	x2, x2, #0x1fe
46	LEA_LO(	x1, approx_tab)
47	ldrh	w3, [x1,x2]
48	lsr	x4, x0, #24
49	add	x4, x4, #1
50	ubfiz	x2, x3, #11, #16
51	umull	x3, w3, w3
52	mul	x3, x3, x4
53	sub	x2, x2, #1
54	sub	x2, x2, x3, lsr #40
55	lsl	x3, x2, #60
56	mul	x1, x2, x2
57	msub	x1, x1, x4, x3
58	lsl	x2, x2, #13
59	add	x1, x2, x1, lsr #47
60	and	x2, x0, #1
61	neg	x3, x2
62	and	x3, x3, x1, lsr #1
63	add	x2, x2, x0, lsr #1
64	msub	x2, x1, x2, x3
65	umulh	x2, x2, x1
66	lsl	x1, x1, #31
67	add	x1, x1, x2, lsr #1
68	mul	x3, x1, x0
69	umulh	x2, x1, x0
70	adds	x4, x3, x0
71	adc	x0, x2, x0
72	sub	x0, x1, x0
73	ret
74EPILOGUE()
75
76	RODATA
77	ALIGN(2)
78	TYPE(   approx_tab, object)
79	SIZE(   approx_tab, 512)
80approx_tab:
81forloop(i,256,512-1,dnl
82`	.hword	eval(0x7fd00/i)
83')dnl
84