1dnl  SPARC v9 mpn_lshift
2
3dnl  Contributed to the GNU project by David Miller.
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 UltraSPARC 1&2:	 2
37C UltraSPARC 3:		 2.5
38C UltraSPARC T1:	17.5
39C UltraSPARC T3:	 8
40C UltraSPARC T4:	 3
41
42C INPUT PARAMETERS
43define(`rp',     `%i0')
44define(`up',     `%i1')
45define(`n',      `%i2')
46define(`cnt',    `%i3')
47
48define(`tcnt',   `%i4')
49define(`retval', `%i5')
50define(`u0',     `%l0')
51define(`u1',     `%l1')
52define(`r0',     `%l6')
53define(`r1',     `%l7')
54define(`u0_off', `%o0')
55define(`u1_off', `%o1')
56define(`r0_off', `%o2')
57define(`r1_off', `%o3')
58
59ASM_START()
60	REGISTER(%g2,#scratch)
61	REGISTER(%g3,#scratch)
62PROLOGUE(mpn_lshift)
63	save	%sp, -176, %sp
64
65	sllx	n, 3, n
66	sub	%g0, cnt, tcnt
67
68	sub	up, 8, u1_off
69	add	rp, (5 * 8), r1_off
70
71	ldx	[n + u1_off], u1	C WAS: up - 8
72	add	u1_off, (3 * 8), u1_off
73
74	sub	r1_off, 8, r0_off
75	sub	u1_off, 8, u0_off
76
77	subcc	n, (3 * 8), n
78	srlx	u1, tcnt, retval
79
80	bl,pn	%xcc, L(end12)
81	 sllx	u1, cnt, %l3
82
83	ldx	[n + u0_off], u0	C WAS: up - 16
84	subcc	n, (2 * 8), n
85
86	ldx	[n + u1_off], u1	C WAS: up - 24
87
88	bl,pn	%xcc, L(end34)
89	 srlx	u0, tcnt, %l4
90
91	b,a	L(top)
92	ALIGN(16)
93L(top):
94	sllx	u0, cnt, %l2
95	or	%l4, %l3, r0
96
97	ldx	[n + u0_off], u0	C WAS: up - 16
98	srlx	u1, tcnt, %l5
99
100	stx	r0, [n + r0_off]	C WAS: rp - 8
101	subcc	n, (2 * 8), n
102
103	sllx	u1, cnt, %l3
104	or	%l2, %l5, r1
105
106	ldx	[n + u1_off], u1	C WAS: up - 24
107	srlx	u0, tcnt, %l4
108
109	bge,pt	%xcc, L(top)
110	 stx	r1, [n + r1_off]	C WAS: rp - 16
111
112L(end34):
113	sllx	u0, cnt, %l2
114	or	%l4, %l3, r0
115
116	srlx	u1, tcnt, %l5
117	stx	r0, [n + r0_off]	C WAS: rp - 8
118
119	or	%l2, %l5, r1
120	sub	n, (2 * 8), %o5
121
122	sllx	u1, cnt, %l3
123	stx	r1, [%o5 + r1_off]	C WAS: rp - 16
124
125L(end12):
126	andcc	n, 8, %g0
127	bz,pn	%xcc, L(done)
128	 nop
129
130	ldx	[n + u0_off], u1
131	srlx	u1, tcnt, %l4
132	or	%l4, %l3, r0
133	stx	r0, [r0_off - 24]
134	sllx	u1, cnt, %l3
135L(done):
136	stx	%l3, [r0_off - 32]
137
138	ret
139	restore retval, 0, %o0
140EPILOGUE()
141