invert_limb.asm revision 1.3 
1dnl  ARM mpn_invert_limb -- Invert a normalized limb.
2
3dnl  Copyright 2001, 2009, 2011, 2012 Free Software Foundation, Inc.
4
5dnl  This file is part of the GNU MP Library.
6
7dnl  The GNU MP Library is free software; you can redistribute it and/or modify
8dnl  it under the terms of the GNU Lesser General Public License as published
9dnl  by the Free Software Foundation; either version 3 of the License, or (at
10dnl  your option) any later version.
11
12dnl  The GNU MP Library is distributed in the hope that it will be useful, but
13dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
14dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
15dnl  License for more details.
16
17dnl  You should have received a copy of the GNU Lesser General Public License
18dnl  along with the GNU MP Library.  If not, see http://www.gnu.org/licenses/.
19
20include(`../config.m4')
21
22ASM_START()
23PROLOGUE(mpn_invert_limb)
24	LEA(	r2, approx_tab-512)
25	mov	r3, r0, lsr #23
26	mov	r3, r3, asl #1
27	ldrh	r3, [r3, r2]
28	mov	r1, r3, asl #17
29	mul	r12, r3, r3
30	umull	r3, r2, r12, r0
31	sub	r1, r1, r2, asl #1
32	umull	r3, r2, r1, r1
33	umull	r12, r3, r0, r3
34	umull	r2, r12, r0, r2
35	adds	r2, r2, r3
36	adc	r12, r12, #0
37	rsb	r1, r12, r1
38	mvn	r2, r2, lsr #30
39	add	r2, r2, r1, asl #2
40	umull	r12, r3, r0, r2
41	adds	r1, r12, r0
42	adc	r3, r3, r0
43	rsb	r0, r3, r2
44ifdef(`ARM_THUMB_MODE',
45`	bx	lr
46',`	mov	pc, lr
47')
48EPILOGUE()
49
50	.section .rodata
51	ALIGN(2)
52approx_tab:
53	.short    0xffc0,0xfec0,0xfdc0,0xfcc0,0xfbc0,0xfac0,0xfa00,0xf900
54	.short    0xf800,0xf700,0xf640,0xf540,0xf440,0xf380,0xf280,0xf180
55	.short    0xf0c0,0xefc0,0xef00,0xee00,0xed40,0xec40,0xeb80,0xeac0
56	.short    0xe9c0,0xe900,0xe840,0xe740,0xe680,0xe5c0,0xe500,0xe400
57	.short    0xe340,0xe280,0xe1c0,0xe100,0xe040,0xdf80,0xdec0,0xde00
58	.short    0xdd40,0xdc80,0xdbc0,0xdb00,0xda40,0xd980,0xd8c0,0xd800
59	.short    0xd740,0xd680,0xd600,0xd540,0xd480,0xd3c0,0xd340,0xd280
60	.short    0xd1c0,0xd140,0xd080,0xcfc0,0xcf40,0xce80,0xcdc0,0xcd40
61	.short    0xcc80,0xcc00,0xcb40,0xcac0,0xca00,0xc980,0xc8c0,0xc840
62	.short    0xc780,0xc700,0xc640,0xc5c0,0xc540,0xc480,0xc400,0xc380
63	.short    0xc2c0,0xc240,0xc1c0,0xc100,0xc080,0xc000,0xbf80,0xbec0
64	.short    0xbe40,0xbdc0,0xbd40,0xbc80,0xbc00,0xbb80,0xbb00,0xba80
65	.short    0xba00,0xb980,0xb900,0xb840,0xb7c0,0xb740,0xb6c0,0xb640
66	.short    0xb5c0,0xb540,0xb4c0,0xb440,0xb3c0,0xb340,0xb2c0,0xb240
67	.short    0xb1c0,0xb140,0xb0c0,0xb080,0xb000,0xaf80,0xaf00,0xae80
68	.short    0xae00,0xad80,0xad40,0xacc0,0xac40,0xabc0,0xab40,0xaac0
69	.short    0xaa80,0xaa00,0xa980,0xa900,0xa8c0,0xa840,0xa7c0,0xa740
70	.short    0xa700,0xa680,0xa600,0xa5c0,0xa540,0xa4c0,0xa480,0xa400
71	.short    0xa380,0xa340,0xa2c0,0xa240,0xa200,0xa180,0xa140,0xa0c0
72	.short    0xa080,0xa000,0x9f80,0x9f40,0x9ec0,0x9e80,0x9e00,0x9dc0
73	.short    0x9d40,0x9d00,0x9c80,0x9c40,0x9bc0,0x9b80,0x9b00,0x9ac0
74	.short    0x9a40,0x9a00,0x9980,0x9940,0x98c0,0x9880,0x9840,0x97c0
75	.short    0x9780,0x9700,0x96c0,0x9680,0x9600,0x95c0,0x9580,0x9500
76	.short    0x94c0,0x9440,0x9400,0x93c0,0x9340,0x9300,0x92c0,0x9240
77	.short    0x9200,0x91c0,0x9180,0x9100,0x90c0,0x9080,0x9000,0x8fc0
78	.short    0x8f80,0x8f40,0x8ec0,0x8e80,0x8e40,0x8e00,0x8d80,0x8d40
79	.short    0x8d00,0x8cc0,0x8c80,0x8c00,0x8bc0,0x8b80,0x8b40,0x8b00
80	.short    0x8a80,0x8a40,0x8a00,0x89c0,0x8980,0x8940,0x88c0,0x8880
81	.short    0x8840,0x8800,0x87c0,0x8780,0x8740,0x8700,0x8680,0x8640
82	.short    0x8600,0x85c0,0x8580,0x8540,0x8500,0x84c0,0x8480,0x8440
83	.short    0x8400,0x8380,0x8340,0x8300,0x82c0,0x8280,0x8240,0x8200
84	.short    0x81c0,0x8180,0x8140,0x8100,0x80c0,0x8080,0x8040,0x8000
85ASM_END()
86