1dnl  ARM64 mpn_mul_1
2
3dnl  Contributed to the GNU project by Torbj��rn 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	7.5-8
37C Cortex-A57	 7
38C Cortex-A72
39C X-Gene	 4
40
41C TODO
42C  * Start first multiply earlier.
43
44changecom(blah)
45
46define(`rp', `x0')
47define(`up', `x1')
48define(`n',  `x2')
49define(`v0', `x3')
50
51
52PROLOGUE(mpn_mul_1c)
53	adds	xzr, xzr, xzr		C clear cy flag
54	b	L(com)
55EPILOGUE()
56
57PROLOGUE(mpn_mul_1)
58	adds	x4, xzr, xzr		C clear register and cy flag
59L(com):	lsr	x18, n, #2
60	tbnz	n, #0, L(bx1)
61
62L(bx0):	mov	x11, x4
63	tbz	n, #1, L(b00)
64
65L(b10):	ldp	x4, x5, [up]
66	mul	x8, x4, v0
67	umulh	x10, x4, v0
68	cbz	x18, L(2)
69	ldp	x6, x7, [up,#16]!
70	mul	x9, x5, v0
71	b	L(mid)-8
72
73L(2):	mul	x9, x5, v0
74	b	L(2e)
75
76L(bx1):	ldr	x7, [up],#8
77	mul	x9, x7, v0
78	umulh	x11, x7, v0
79	adds	x9, x9, x4
80	str	x9, [rp],#8
81	tbnz	n, #1, L(b10)
82
83L(b01):	cbz	x18, L(1)
84
85L(b00):	ldp	x6, x7, [up]
86	mul	x8, x6, v0
87	umulh	x10, x6, v0
88	ldp	x4, x5, [up,#16]
89	mul	x9, x7, v0
90	adcs	x12, x8, x11
91	umulh	x11, x7, v0
92	add	rp, rp, #16
93	sub	x18, x18, #1
94	cbz	x18, L(end)
95
96	ALIGN(16)
97L(top):	mul	x8, x4, v0
98	ldp	x6, x7, [up,#32]!
99	adcs	x13, x9, x10
100	umulh	x10, x4, v0
101	mul	x9, x5, v0
102	stp	x12, x13, [rp,#-16]
103	adcs	x12, x8, x11
104	umulh	x11, x5, v0
105L(mid):	mul	x8, x6, v0
106	ldp	x4, x5, [up,#16]
107	adcs	x13, x9, x10
108	umulh	x10, x6, v0
109	mul	x9, x7, v0
110	stp	x12, x13, [rp],#32
111	adcs	x12, x8, x11
112	umulh	x11, x7, v0
113	sub	x18, x18, #1
114	cbnz	x18, L(top)
115
116L(end):	mul	x8, x4, v0
117	adcs	x13, x9, x10
118	umulh	x10, x4, v0
119	mul	x9, x5, v0
120	stp	x12, x13, [rp,#-16]
121L(2e):	adcs	x12, x8, x11
122	umulh	x11, x5, v0
123	adcs	x13, x9, x10
124	stp	x12, x13, [rp]
125L(1):	adc	x0, x11, xzr
126	ret
127EPILOGUE()
128