xref: /OK3568_Linux_fs/kernel/arch/c6x/lib/strasgi.S (revision 4882a59341e53eb6f0b4789bf948001014eff981)
1*4882a593Smuzhiyun;; SPDX-License-Identifier: GPL-2.0-or-later
2*4882a593Smuzhiyun;;  Copyright 2010  Free Software Foundation, Inc.
3*4882a593Smuzhiyun;;  Contributed by Bernd Schmidt <bernds@codesourcery.com>.
4*4882a593Smuzhiyun;;
5*4882a593Smuzhiyun
6*4882a593Smuzhiyun#include <linux/linkage.h>
7*4882a593Smuzhiyun
8*4882a593Smuzhiyun	.text
9*4882a593Smuzhiyun
10*4882a593SmuzhiyunENTRY(__c6xabi_strasgi)
11*4882a593Smuzhiyun	;; This is essentially memcpy, with alignment known to be at least
12*4882a593Smuzhiyun	;; 4, and the size a multiple of 4 greater than or equal to 28.
13*4882a593Smuzhiyun	 ldw	.d2t1	*B4++, A0
14*4882a593Smuzhiyun||	 mvk	.s2	16, B1
15*4882a593Smuzhiyun	 ldw	.d2t1	*B4++, A1
16*4882a593Smuzhiyun||	 mvk	.s2	20, B2
17*4882a593Smuzhiyun||	 sub	.d1	A6, 24, A6
18*4882a593Smuzhiyun	 ldw	.d2t1	*B4++, A5
19*4882a593Smuzhiyun	 ldw	.d2t1	*B4++, A7
20*4882a593Smuzhiyun||	 mv	.l2x	A6, B7
21*4882a593Smuzhiyun	 ldw	.d2t1	*B4++, A8
22*4882a593Smuzhiyun	 ldw	.d2t1	*B4++, A9
23*4882a593Smuzhiyun||	 mv	.s2x	A0, B5
24*4882a593Smuzhiyun||	 cmpltu	.l2	B2, B7, B0
25*4882a593Smuzhiyun
26*4882a593Smuzhiyun_strasgi_loop:
27*4882a593Smuzhiyun	 stw	.d1t2	B5, *A4++
28*4882a593Smuzhiyun|| [B0]	 ldw	.d2t1	*B4++, A0
29*4882a593Smuzhiyun||	 mv	.s2x	A1, B5
30*4882a593Smuzhiyun||	 mv	.l2	B7, B6
31*4882a593Smuzhiyun
32*4882a593Smuzhiyun   [B0]	 sub	.d2	B6, 24, B7
33*4882a593Smuzhiyun|| [B0]	 b	.s2	_strasgi_loop
34*4882a593Smuzhiyun||	 cmpltu	.l2	B1, B6, B0
35*4882a593Smuzhiyun
36*4882a593Smuzhiyun   [B0]	 ldw	.d2t1	*B4++, A1
37*4882a593Smuzhiyun||	 stw	.d1t2	B5, *A4++
38*4882a593Smuzhiyun||	 mv	.s2x	A5, B5
39*4882a593Smuzhiyun||	 cmpltu	.l2	12, B6, B0
40*4882a593Smuzhiyun
41*4882a593Smuzhiyun   [B0]	 ldw	.d2t1	*B4++, A5
42*4882a593Smuzhiyun||	 stw	.d1t2	B5, *A4++
43*4882a593Smuzhiyun||	 mv	.s2x	A7, B5
44*4882a593Smuzhiyun||	 cmpltu	.l2	8, B6, B0
45*4882a593Smuzhiyun
46*4882a593Smuzhiyun   [B0]	 ldw	.d2t1	*B4++, A7
47*4882a593Smuzhiyun||	 stw	.d1t2	B5, *A4++
48*4882a593Smuzhiyun||	 mv	.s2x	A8, B5
49*4882a593Smuzhiyun||	 cmpltu	.l2	4, B6, B0
50*4882a593Smuzhiyun
51*4882a593Smuzhiyun   [B0]	 ldw	.d2t1	*B4++, A8
52*4882a593Smuzhiyun||	 stw	.d1t2	B5, *A4++
53*4882a593Smuzhiyun||	 mv	.s2x	A9, B5
54*4882a593Smuzhiyun||	 cmpltu	.l2	0, B6, B0
55*4882a593Smuzhiyun
56*4882a593Smuzhiyun   [B0]	 ldw	.d2t1	*B4++, A9
57*4882a593Smuzhiyun||	 stw	.d1t2	B5, *A4++
58*4882a593Smuzhiyun||	 mv	.s2x	A0, B5
59*4882a593Smuzhiyun||	 cmpltu	.l2	B2, B7, B0
60*4882a593Smuzhiyun
61*4882a593Smuzhiyun	;; loop back branch happens here
62*4882a593Smuzhiyun
63*4882a593Smuzhiyun	 cmpltu	.l2	B1, B6, B0
64*4882a593Smuzhiyun||	 ret	.s2	b3
65*4882a593Smuzhiyun
66*4882a593Smuzhiyun   [B0]	 stw	.d1t1	A1, *A4++
67*4882a593Smuzhiyun||	 cmpltu	.l2	12, B6, B0
68*4882a593Smuzhiyun   [B0]	 stw	.d1t1	A5, *A4++
69*4882a593Smuzhiyun||	 cmpltu	.l2	8, B6, B0
70*4882a593Smuzhiyun   [B0]	 stw	.d1t1	A7, *A4++
71*4882a593Smuzhiyun||	 cmpltu	.l2	4, B6, B0
72*4882a593Smuzhiyun   [B0]	 stw	.d1t1	A8, *A4++
73*4882a593Smuzhiyun||	 cmpltu	.l2	0, B6, B0
74*4882a593Smuzhiyun   [B0]	 stw	.d1t1	A9, *A4++
75*4882a593Smuzhiyun
76*4882a593Smuzhiyun	;; return happens here
77*4882a593SmuzhiyunENDPROC(__c6xabi_strasgi)
78