xref: /netbsd-src/external/lgpl3/gmp/dist/mpn/arm64/aorsmul_1.asm (revision 3f351f34c6d827cf017cdcff3543f6ec0c88b420)
1dnl  ARM64 mpn_addmul_1 and mpn_submul_1
2
3dnl  Contributed to the GNU project by Torbjörn Granlund.
4
5dnl  Copyright 2013, 2015, 2017 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	9.3-9.8
37C Cortex-A57	 7.0
38C X-Gene	 5.0
39
40C NOTES
41C  * It is possible to keep the carry chain alive between the addition blocks
42C    and thus avoid csinc, but only for addmul_1.  Since that saves no time
43C    on the tested pipelines, we keep addmul_1 and submul_1 similar.
44C  * We could separate feed-in into 4 blocks, one for each residue (mod 4).
45C    That is likely to save a few cycles.
46
47changecom(blah)
48
49define(`rp', `x0')
50define(`up', `x1')
51define(`n',  `x2')
52define(`v0', `x3')
53
54ifdef(`OPERATION_addmul_1', `
55  define(`ADDSUB',	adds)
56  define(`ADDSUBC',	adcs)
57  define(`COND',	`cc')
58  define(`func',	mpn_addmul_1)')
59ifdef(`OPERATION_submul_1', `
60  define(`ADDSUB',	subs)
61  define(`ADDSUBC',	sbcs)
62  define(`COND',	`cs')
63  define(`func',	mpn_submul_1)')
64
65MULFUNC_PROLOGUE(mpn_addmul_1 mpn_submul_1)
66
67PROLOGUE(func)
68	adds	x15, xzr, xzr
69
70	tbz	n, #0, L(1)
71
72	ldr	x4, [up],#8
73	mul	x8, x4, v0
74	umulh	x12, x4, v0
75	ldr	x4, [rp]
76	ADDSUB	x8, x4, x8
77	csinc	x15, x12, x12, COND
78	str	x8, [rp],#8
79
80L(1):	tbz	n, #1, L(2)
81
82	ldp	x4, x5, [up],#16
83	mul	x8, x4, v0
84	umulh	x12, x4, v0
85	mul	x9, x5, v0
86	umulh	x13, x5, v0
87	adds	x8, x8, x15
88	adcs	x9, x9, x12
89	ldp	x4, x5, [rp]
90	adc	x15, x13, xzr
91	ADDSUB	x8, x4, x8
92	ADDSUBC	x9, x5, x9
93	csinc	x15, x15, x15, COND
94	stp	x8, x9, [rp],#16
95
96L(2):	lsr	n, n, #2
97	cbz	n, L(le3)
98	ldp	x4, x5, [up],#32
99	ldp	x6, x7, [up,#-16]
100	b	L(mid)
101L(le3):	mov	x0, x15
102	ret
103
104	ALIGN(16)
105L(top):	ldp	x4, x5, [up],#32
106	ldp	x6, x7, [up,#-16]
107	ADDSUB	x8, x16, x8
108	ADDSUBC	x9, x17, x9
109	stp	x8, x9, [rp],#32
110	ADDSUBC	x10, x12, x10
111	ADDSUBC	x11, x13, x11
112	stp	x10, x11, [rp,#-16]
113	csinc	x15, x15, x15, COND
114L(mid):	sub	n, n, #1
115	mul	x8, x4, v0
116	umulh	x12, x4, v0
117	mul	x9, x5, v0
118	umulh	x13, x5, v0
119	adds	x8, x8, x15
120	mul	x10, x6, v0
121	umulh	x14, x6, v0
122	adcs	x9, x9, x12
123	mul	x11, x7, v0
124	umulh	x15, x7, v0
125	adcs	x10, x10, x13
126	ldp	x16, x17, [rp]
127	adcs	x11, x11, x14
128	ldp	x12, x13, [rp,#16]
129	adc	x15, x15, xzr
130	cbnz	n, L(top)
131
132	ADDSUB	x8, x16, x8
133	ADDSUBC	x9, x17, x9
134	ADDSUBC	x10, x12, x10
135	ADDSUBC	x11, x13, x11
136	stp	x8, x9, [rp]
137	stp	x10, x11, [rp,#16]
138	csinc	x0, x15, x15, COND
139	ret
140EPILOGUE()
141