1c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/*
2c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru * include/asm-generic/xor.h
3c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru *
4c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru * Generic optimized RAID-5 checksumming functions.
5c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru *
6c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru * This program is free software; you can redistribute it and/or modify
7c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru * it under the terms of the GNU General Public License as published by
8c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru * the Free Software Foundation; either version 2, or (at your option)
9c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru * any later version.
10c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru *
11c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru * You should have received a copy of the GNU General Public License
12c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru * (for example /usr/src/linux/COPYING); if not, write to the Free
13c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
14c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru */
15c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
16c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include <asm/processor.h>
17c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
18c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
19c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_8regs_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
20c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
21c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8;
22c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
23c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
24c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] ^= p2[0];
25c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] ^= p2[1];
26c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] ^= p2[2];
27c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] ^= p2[3];
28c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] ^= p2[4];
29c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] ^= p2[5];
30c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] ^= p2[6];
31c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] ^= p2[7];
32c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
33c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
34c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
35c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
36c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
37c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
38c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_8regs_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
39c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3)
40c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
41c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8;
42c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
43c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
44c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] ^= p2[0] ^ p3[0];
45c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] ^= p2[1] ^ p3[1];
46c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] ^= p2[2] ^ p3[2];
47c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] ^= p2[3] ^ p3[3];
48c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] ^= p2[4] ^ p3[4];
49c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] ^= p2[5] ^ p3[5];
50c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] ^= p2[6] ^ p3[6];
51c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] ^= p2[7] ^ p3[7];
52c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
53c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
54c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
55c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
56c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
57c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
58c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
59c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_8regs_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
60c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3, unsigned long *p4)
61c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
62c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8;
63c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
64c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
65c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] ^= p2[0] ^ p3[0] ^ p4[0];
66c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] ^= p2[1] ^ p3[1] ^ p4[1];
67c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] ^= p2[2] ^ p3[2] ^ p4[2];
68c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] ^= p2[3] ^ p3[3] ^ p4[3];
69c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] ^= p2[4] ^ p3[4] ^ p4[4];
70c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] ^= p2[5] ^ p3[5] ^ p4[5];
71c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] ^= p2[6] ^ p3[6] ^ p4[6];
72c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] ^= p2[7] ^ p3[7] ^ p4[7];
73c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
74c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
75c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
76c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p4 += 8;
77c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
78c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
79c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
80c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
81c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_8regs_5(unsigned long bytes, unsigned long *p1, unsigned long *p2,
82c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3, unsigned long *p4, unsigned long *p5)
83c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
84c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8;
85c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
86c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
87c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] ^= p2[0] ^ p3[0] ^ p4[0] ^ p5[0];
88c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] ^= p2[1] ^ p3[1] ^ p4[1] ^ p5[1];
89c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] ^= p2[2] ^ p3[2] ^ p4[2] ^ p5[2];
90c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] ^= p2[3] ^ p3[3] ^ p4[3] ^ p5[3];
91c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] ^= p2[4] ^ p3[4] ^ p4[4] ^ p5[4];
92c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] ^= p2[5] ^ p3[5] ^ p4[5] ^ p5[5];
93c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] ^= p2[6] ^ p3[6] ^ p4[6] ^ p5[6];
94c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] ^= p2[7] ^ p3[7] ^ p4[7] ^ p5[7];
95c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
96c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
97c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
98c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p4 += 8;
99c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p5 += 8;
100c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
101c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
102c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
103c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
104c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_32regs_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
105c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
106c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8;
107c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
108c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
109c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		register long d0, d1, d2, d3, d4, d5, d6, d7;
110c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 = p1[0];	/* Pull the stuff into registers	*/
111c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 = p1[1];	/*  ... in bursts, if possible.		*/
112c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 = p1[2];
113c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 = p1[3];
114c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 = p1[4];
115c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 = p1[5];
116c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 = p1[6];
117c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 = p1[7];
118c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p2[0];
119c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p2[1];
120c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p2[2];
121c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p2[3];
122c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p2[4];
123c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p2[5];
124c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p2[6];
125c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p2[7];
126c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] = d0;	/* Store the result (in bursts)		*/
127c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] = d1;
128c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] = d2;
129c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] = d3;
130c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] = d4;
131c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] = d5;
132c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] = d6;
133c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] = d7;
134c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
135c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
136c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
137c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
138c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
139c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
140c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_32regs_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
141c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3)
142c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
143c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8;
144c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
145c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
146c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		register long d0, d1, d2, d3, d4, d5, d6, d7;
147c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 = p1[0];	/* Pull the stuff into registers	*/
148c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 = p1[1];	/*  ... in bursts, if possible.		*/
149c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 = p1[2];
150c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 = p1[3];
151c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 = p1[4];
152c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 = p1[5];
153c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 = p1[6];
154c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 = p1[7];
155c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p2[0];
156c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p2[1];
157c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p2[2];
158c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p2[3];
159c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p2[4];
160c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p2[5];
161c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p2[6];
162c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p2[7];
163c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p3[0];
164c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p3[1];
165c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p3[2];
166c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p3[3];
167c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p3[4];
168c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p3[5];
169c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p3[6];
170c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p3[7];
171c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] = d0;	/* Store the result (in bursts)		*/
172c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] = d1;
173c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] = d2;
174c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] = d3;
175c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] = d4;
176c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] = d5;
177c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] = d6;
178c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] = d7;
179c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
180c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
181c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
182c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
183c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
184c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
185c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
186c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_32regs_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
187c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3, unsigned long *p4)
188c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
189c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8;
190c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
191c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
192c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		register long d0, d1, d2, d3, d4, d5, d6, d7;
193c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 = p1[0];	/* Pull the stuff into registers	*/
194c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 = p1[1];	/*  ... in bursts, if possible.		*/
195c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 = p1[2];
196c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 = p1[3];
197c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 = p1[4];
198c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 = p1[5];
199c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 = p1[6];
200c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 = p1[7];
201c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p2[0];
202c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p2[1];
203c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p2[2];
204c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p2[3];
205c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p2[4];
206c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p2[5];
207c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p2[6];
208c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p2[7];
209c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p3[0];
210c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p3[1];
211c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p3[2];
212c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p3[3];
213c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p3[4];
214c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p3[5];
215c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p3[6];
216c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p3[7];
217c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p4[0];
218c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p4[1];
219c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p4[2];
220c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p4[3];
221c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p4[4];
222c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p4[5];
223c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p4[6];
224c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p4[7];
225c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] = d0;	/* Store the result (in bursts)		*/
226c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] = d1;
227c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] = d2;
228c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] = d3;
229c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] = d4;
230c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] = d5;
231c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] = d6;
232c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] = d7;
233c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
234c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
235c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
236c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p4 += 8;
237c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
238c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
239c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
240c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
241c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_32regs_5(unsigned long bytes, unsigned long *p1, unsigned long *p2,
242c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3, unsigned long *p4, unsigned long *p5)
243c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
244c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8;
245c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
246c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
247c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		register long d0, d1, d2, d3, d4, d5, d6, d7;
248c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 = p1[0];	/* Pull the stuff into registers	*/
249c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 = p1[1];	/*  ... in bursts, if possible.		*/
250c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 = p1[2];
251c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 = p1[3];
252c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 = p1[4];
253c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 = p1[5];
254c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 = p1[6];
255c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 = p1[7];
256c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p2[0];
257c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p2[1];
258c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p2[2];
259c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p2[3];
260c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p2[4];
261c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p2[5];
262c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p2[6];
263c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p2[7];
264c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p3[0];
265c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p3[1];
266c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p3[2];
267c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p3[3];
268c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p3[4];
269c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p3[5];
270c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p3[6];
271c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p3[7];
272c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p4[0];
273c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p4[1];
274c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p4[2];
275c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p4[3];
276c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p4[4];
277c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p4[5];
278c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p4[6];
279c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p4[7];
280c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p5[0];
281c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p5[1];
282c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p5[2];
283c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p5[3];
284c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p5[4];
285c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p5[5];
286c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p5[6];
287c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p5[7];
288c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] = d0;	/* Store the result (in bursts)		*/
289c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] = d1;
290c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] = d2;
291c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] = d3;
292c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] = d4;
293c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] = d5;
294c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] = d6;
295c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] = d7;
296c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
297c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
298c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
299c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p4 += 8;
300c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p5 += 8;
301c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
302c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
303c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
304c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
305c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_8regs_p_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
306c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
307c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8 - 1;
308c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetchw(p1);
309c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p2);
310c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
311c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
312c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetchw(p1+8);
313c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p2+8);
314c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru once_more:
315c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] ^= p2[0];
316c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] ^= p2[1];
317c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] ^= p2[2];
318c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] ^= p2[3];
319c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] ^= p2[4];
320c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] ^= p2[5];
321c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] ^= p2[6];
322c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] ^= p2[7];
323c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
324c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
325c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
326c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	if (lines == 0)
327c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		goto once_more;
328c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
329c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
330c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
331c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_8regs_p_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
332c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3)
333c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
334c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8 - 1;
335c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetchw(p1);
336c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p2);
337c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p3);
338c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
339c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
340c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetchw(p1+8);
341c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p2+8);
342c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p3+8);
343c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru once_more:
344c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] ^= p2[0] ^ p3[0];
345c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] ^= p2[1] ^ p3[1];
346c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] ^= p2[2] ^ p3[2];
347c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] ^= p2[3] ^ p3[3];
348c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] ^= p2[4] ^ p3[4];
349c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] ^= p2[5] ^ p3[5];
350c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] ^= p2[6] ^ p3[6];
351c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] ^= p2[7] ^ p3[7];
352c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
353c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
354c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
355c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
356c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	if (lines == 0)
357c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		goto once_more;
358c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
359c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
360c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
361c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_8regs_p_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
362c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3, unsigned long *p4)
363c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
364c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8 - 1;
365c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
366c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetchw(p1);
367c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p2);
368c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p3);
369c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p4);
370c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
371c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
372c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetchw(p1+8);
373c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p2+8);
374c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p3+8);
375c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p4+8);
376c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru once_more:
377c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] ^= p2[0] ^ p3[0] ^ p4[0];
378c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] ^= p2[1] ^ p3[1] ^ p4[1];
379c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] ^= p2[2] ^ p3[2] ^ p4[2];
380c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] ^= p2[3] ^ p3[3] ^ p4[3];
381c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] ^= p2[4] ^ p3[4] ^ p4[4];
382c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] ^= p2[5] ^ p3[5] ^ p4[5];
383c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] ^= p2[6] ^ p3[6] ^ p4[6];
384c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] ^= p2[7] ^ p3[7] ^ p4[7];
385c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
386c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
387c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
388c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p4 += 8;
389c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
390c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	if (lines == 0)
391c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		goto once_more;
392c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
393c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
394c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
395c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_8regs_p_5(unsigned long bytes, unsigned long *p1, unsigned long *p2,
396c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3, unsigned long *p4, unsigned long *p5)
397c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
398c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8 - 1;
399c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
400c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetchw(p1);
401c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p2);
402c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p3);
403c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p4);
404c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p5);
405c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
406c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
407c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetchw(p1+8);
408c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p2+8);
409c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p3+8);
410c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p4+8);
411c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p5+8);
412c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru once_more:
413c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] ^= p2[0] ^ p3[0] ^ p4[0] ^ p5[0];
414c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] ^= p2[1] ^ p3[1] ^ p4[1] ^ p5[1];
415c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] ^= p2[2] ^ p3[2] ^ p4[2] ^ p5[2];
416c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] ^= p2[3] ^ p3[3] ^ p4[3] ^ p5[3];
417c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] ^= p2[4] ^ p3[4] ^ p4[4] ^ p5[4];
418c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] ^= p2[5] ^ p3[5] ^ p4[5] ^ p5[5];
419c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] ^= p2[6] ^ p3[6] ^ p4[6] ^ p5[6];
420c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] ^= p2[7] ^ p3[7] ^ p4[7] ^ p5[7];
421c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
422c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
423c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
424c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p4 += 8;
425c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p5 += 8;
426c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
427c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	if (lines == 0)
428c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		goto once_more;
429c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
430c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
431c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
432c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_32regs_p_2(unsigned long bytes, unsigned long *p1, unsigned long *p2)
433c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
434c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8 - 1;
435c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
436c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetchw(p1);
437c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p2);
438c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
439c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
440c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		register long d0, d1, d2, d3, d4, d5, d6, d7;
441c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
442c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetchw(p1+8);
443c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p2+8);
444c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru once_more:
445c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 = p1[0];	/* Pull the stuff into registers	*/
446c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 = p1[1];	/*  ... in bursts, if possible.		*/
447c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 = p1[2];
448c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 = p1[3];
449c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 = p1[4];
450c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 = p1[5];
451c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 = p1[6];
452c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 = p1[7];
453c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p2[0];
454c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p2[1];
455c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p2[2];
456c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p2[3];
457c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p2[4];
458c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p2[5];
459c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p2[6];
460c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p2[7];
461c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] = d0;	/* Store the result (in bursts)		*/
462c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] = d1;
463c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] = d2;
464c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] = d3;
465c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] = d4;
466c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] = d5;
467c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] = d6;
468c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] = d7;
469c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
470c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
471c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
472c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	if (lines == 0)
473c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		goto once_more;
474c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
475c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
476c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
477c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_32regs_p_3(unsigned long bytes, unsigned long *p1, unsigned long *p2,
478c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3)
479c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
480c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8 - 1;
481c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
482c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetchw(p1);
483c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p2);
484c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p3);
485c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
486c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
487c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		register long d0, d1, d2, d3, d4, d5, d6, d7;
488c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
489c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetchw(p1+8);
490c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p2+8);
491c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p3+8);
492c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru once_more:
493c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 = p1[0];	/* Pull the stuff into registers	*/
494c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 = p1[1];	/*  ... in bursts, if possible.		*/
495c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 = p1[2];
496c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 = p1[3];
497c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 = p1[4];
498c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 = p1[5];
499c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 = p1[6];
500c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 = p1[7];
501c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p2[0];
502c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p2[1];
503c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p2[2];
504c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p2[3];
505c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p2[4];
506c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p2[5];
507c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p2[6];
508c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p2[7];
509c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p3[0];
510c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p3[1];
511c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p3[2];
512c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p3[3];
513c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p3[4];
514c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p3[5];
515c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p3[6];
516c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p3[7];
517c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] = d0;	/* Store the result (in bursts)		*/
518c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] = d1;
519c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] = d2;
520c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] = d3;
521c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] = d4;
522c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] = d5;
523c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] = d6;
524c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] = d7;
525c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
526c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
527c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
528c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
529c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	if (lines == 0)
530c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		goto once_more;
531c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
532c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
533c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
534c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_32regs_p_4(unsigned long bytes, unsigned long *p1, unsigned long *p2,
535c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3, unsigned long *p4)
536c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
537c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8 - 1;
538c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
539c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetchw(p1);
540c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p2);
541c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p3);
542c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p4);
543c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
544c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
545c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		register long d0, d1, d2, d3, d4, d5, d6, d7;
546c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
547c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetchw(p1+8);
548c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p2+8);
549c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p3+8);
550c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p4+8);
551c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru once_more:
552c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 = p1[0];	/* Pull the stuff into registers	*/
553c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 = p1[1];	/*  ... in bursts, if possible.		*/
554c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 = p1[2];
555c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 = p1[3];
556c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 = p1[4];
557c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 = p1[5];
558c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 = p1[6];
559c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 = p1[7];
560c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p2[0];
561c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p2[1];
562c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p2[2];
563c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p2[3];
564c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p2[4];
565c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p2[5];
566c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p2[6];
567c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p2[7];
568c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p3[0];
569c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p3[1];
570c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p3[2];
571c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p3[3];
572c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p3[4];
573c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p3[5];
574c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p3[6];
575c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p3[7];
576c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p4[0];
577c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p4[1];
578c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p4[2];
579c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p4[3];
580c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p4[4];
581c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p4[5];
582c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p4[6];
583c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p4[7];
584c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] = d0;	/* Store the result (in bursts)		*/
585c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] = d1;
586c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] = d2;
587c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] = d3;
588c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] = d4;
589c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] = d5;
590c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] = d6;
591c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] = d7;
592c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
593c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
594c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
595c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p4 += 8;
596c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
597c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	if (lines == 0)
598c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		goto once_more;
599c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
600c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
601c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic void
602c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruxor_32regs_p_5(unsigned long bytes, unsigned long *p1, unsigned long *p2,
603c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	    unsigned long *p3, unsigned long *p4, unsigned long *p5)
604c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru{
605c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	long lines = bytes / (sizeof (long)) / 8 - 1;
606c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
607c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetchw(p1);
608c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p2);
609c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p3);
610c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p4);
611c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	prefetch(p5);
612c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
613c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {
614c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		register long d0, d1, d2, d3, d4, d5, d6, d7;
615c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
616c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetchw(p1+8);
617c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p2+8);
618c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p3+8);
619c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p4+8);
620c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		prefetch(p5+8);
621c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru once_more:
622c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 = p1[0];	/* Pull the stuff into registers	*/
623c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 = p1[1];	/*  ... in bursts, if possible.		*/
624c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 = p1[2];
625c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 = p1[3];
626c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 = p1[4];
627c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 = p1[5];
628c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 = p1[6];
629c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 = p1[7];
630c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p2[0];
631c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p2[1];
632c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p2[2];
633c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p2[3];
634c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p2[4];
635c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p2[5];
636c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p2[6];
637c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p2[7];
638c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p3[0];
639c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p3[1];
640c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p3[2];
641c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p3[3];
642c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p3[4];
643c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p3[5];
644c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p3[6];
645c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p3[7];
646c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p4[0];
647c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p4[1];
648c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p4[2];
649c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p4[3];
650c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p4[4];
651c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p4[5];
652c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p4[6];
653c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p4[7];
654c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d0 ^= p5[0];
655c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d1 ^= p5[1];
656c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d2 ^= p5[2];
657c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d3 ^= p5[3];
658c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d4 ^= p5[4];
659c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d5 ^= p5[5];
660c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d6 ^= p5[6];
661c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		d7 ^= p5[7];
662c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[0] = d0;	/* Store the result (in bursts)		*/
663c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[1] = d1;
664c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[2] = d2;
665c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[3] = d3;
666c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[4] = d4;
667c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[5] = d5;
668c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[6] = d6;
669c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1[7] = d7;
670c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p1 += 8;
671c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p2 += 8;
672c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p3 += 8;
673c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p4 += 8;
674c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		p5 += 8;
675c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (--lines > 0);
676c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	if (lines == 0)
677c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		goto once_more;
678c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru}
679c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
680c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic struct xor_block_template xor_block_8regs = {
681c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.name = "8regs",
682c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_2 = xor_8regs_2,
683c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_3 = xor_8regs_3,
684c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_4 = xor_8regs_4,
685c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_5 = xor_8regs_5,
686c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru};
687c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
688c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic struct xor_block_template xor_block_32regs = {
689c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.name = "32regs",
690c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_2 = xor_32regs_2,
691c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_3 = xor_32regs_3,
692c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_4 = xor_32regs_4,
693c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_5 = xor_32regs_5,
694c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru};
695c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
696c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic struct xor_block_template xor_block_8regs_p = {
697c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.name = "8regs_prefetch",
698c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_2 = xor_8regs_p_2,
699c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_3 = xor_8regs_p_3,
700c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_4 = xor_8regs_p_4,
701c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_5 = xor_8regs_p_5,
702c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru};
703c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
704c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querustatic struct xor_block_template xor_block_32regs_p = {
705c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.name = "32regs_prefetch",
706c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_2 = xor_32regs_p_2,
707c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_3 = xor_32regs_p_3,
708c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_4 = xor_32regs_p_4,
709c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	.do_5 = xor_32regs_p_5,
710c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru};
711c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru
712c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#define XOR_TRY_TEMPLATES			\
713c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	do {					\
714c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		xor_speed(&xor_block_8regs);	\
715c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		xor_speed(&xor_block_8regs_p);	\
716c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		xor_speed(&xor_block_32regs);	\
717c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru		xor_speed(&xor_block_32regs_p);	\
718c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru	} while (0)
719