xref: /linux/lib/raid6/sse2.c (revision 58e16d792a6a8c6b750f637a4649967fcac853dc)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* -*- linux-c -*- ------------------------------------------------------- *
3  *
4  *   Copyright 2002 H. Peter Anvin - All Rights Reserved
5  *
6  * ----------------------------------------------------------------------- */
7 
8 /*
9  * raid6/sse2.c
10  *
11  * SSE-2 implementation of RAID-6 syndrome functions
12  *
13  */
14 
15 #include <linux/raid/pq.h>
16 #include "x86.h"
17 
18 static const struct raid6_sse_constants {
19 	u64 x1d[2];
20 } raid6_sse_constants  __attribute__((aligned(16))) = {
21 	{ 0x1d1d1d1d1d1d1d1dULL, 0x1d1d1d1d1d1d1d1dULL },
22 };
23 
raid6_have_sse2(void)24 static int raid6_have_sse2(void)
25 {
26 	/* Not really boot_cpu but "all_cpus" */
27 	return boot_cpu_has(X86_FEATURE_MMX) &&
28 		boot_cpu_has(X86_FEATURE_FXSR) &&
29 		boot_cpu_has(X86_FEATURE_XMM) &&
30 		boot_cpu_has(X86_FEATURE_XMM2);
31 }
32 
33 /*
34  * Plain SSE2 implementation
35  */
raid6_sse21_gen_syndrome(int disks,size_t bytes,void ** ptrs)36 static void raid6_sse21_gen_syndrome(int disks, size_t bytes, void **ptrs)
37 {
38 	u8 **dptr = (u8 **)ptrs;
39 	u8 *p, *q;
40 	int d, z, z0;
41 
42 	z0 = disks - 3;		/* Highest data disk */
43 	p = dptr[z0+1];		/* XOR parity */
44 	q = dptr[z0+2];		/* RS syndrome */
45 
46 	kernel_fpu_begin();
47 
48 	asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
49 	asm volatile("pxor %xmm5,%xmm5");	/* Zero temp */
50 
51 	for ( d = 0 ; d < bytes ; d += 16 ) {
52 		asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
53 		asm volatile("movdqa %0,%%xmm2" : : "m" (dptr[z0][d])); /* P[0] */
54 		asm volatile("prefetchnta %0" : : "m" (dptr[z0-1][d]));
55 		asm volatile("movdqa %xmm2,%xmm4"); /* Q[0] */
56 		asm volatile("movdqa %0,%%xmm6" : : "m" (dptr[z0-1][d]));
57 		for ( z = z0-2 ; z >= 0 ; z-- ) {
58 			asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
59 			asm volatile("pcmpgtb %xmm4,%xmm5");
60 			asm volatile("paddb %xmm4,%xmm4");
61 			asm volatile("pand %xmm0,%xmm5");
62 			asm volatile("pxor %xmm5,%xmm4");
63 			asm volatile("pxor %xmm5,%xmm5");
64 			asm volatile("pxor %xmm6,%xmm2");
65 			asm volatile("pxor %xmm6,%xmm4");
66 			asm volatile("movdqa %0,%%xmm6" : : "m" (dptr[z][d]));
67 		}
68 		asm volatile("pcmpgtb %xmm4,%xmm5");
69 		asm volatile("paddb %xmm4,%xmm4");
70 		asm volatile("pand %xmm0,%xmm5");
71 		asm volatile("pxor %xmm5,%xmm4");
72 		asm volatile("pxor %xmm5,%xmm5");
73 		asm volatile("pxor %xmm6,%xmm2");
74 		asm volatile("pxor %xmm6,%xmm4");
75 
76 		asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
77 		asm volatile("pxor %xmm2,%xmm2");
78 		asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
79 		asm volatile("pxor %xmm4,%xmm4");
80 	}
81 
82 	asm volatile("sfence" : : : "memory");
83 	kernel_fpu_end();
84 }
85 
86 
raid6_sse21_xor_syndrome(int disks,int start,int stop,size_t bytes,void ** ptrs)87 static void raid6_sse21_xor_syndrome(int disks, int start, int stop,
88 				     size_t bytes, void **ptrs)
89 {
90 	u8 **dptr = (u8 **)ptrs;
91 	u8 *p, *q;
92 	int d, z, z0;
93 
94 	z0 = stop;		/* P/Q right side optimization */
95 	p = dptr[disks-2];	/* XOR parity */
96 	q = dptr[disks-1];	/* RS syndrome */
97 
98 	kernel_fpu_begin();
99 
100 	asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
101 
102 	for ( d = 0 ; d < bytes ; d += 16 ) {
103 		asm volatile("movdqa %0,%%xmm4" :: "m" (dptr[z0][d]));
104 		asm volatile("movdqa %0,%%xmm2" : : "m" (p[d]));
105 		asm volatile("pxor %xmm4,%xmm2");
106 		/* P/Q data pages */
107 		for ( z = z0-1 ; z >= start ; z-- ) {
108 			asm volatile("pxor %xmm5,%xmm5");
109 			asm volatile("pcmpgtb %xmm4,%xmm5");
110 			asm volatile("paddb %xmm4,%xmm4");
111 			asm volatile("pand %xmm0,%xmm5");
112 			asm volatile("pxor %xmm5,%xmm4");
113 			asm volatile("movdqa %0,%%xmm5" :: "m" (dptr[z][d]));
114 			asm volatile("pxor %xmm5,%xmm2");
115 			asm volatile("pxor %xmm5,%xmm4");
116 		}
117 		/* P/Q left side optimization */
118 		for ( z = start-1 ; z >= 0 ; z-- ) {
119 			asm volatile("pxor %xmm5,%xmm5");
120 			asm volatile("pcmpgtb %xmm4,%xmm5");
121 			asm volatile("paddb %xmm4,%xmm4");
122 			asm volatile("pand %xmm0,%xmm5");
123 			asm volatile("pxor %xmm5,%xmm4");
124 		}
125 		asm volatile("pxor %0,%%xmm4" : : "m" (q[d]));
126 		/* Don't use movntdq for r/w memory area < cache line */
127 		asm volatile("movdqa %%xmm4,%0" : "=m" (q[d]));
128 		asm volatile("movdqa %%xmm2,%0" : "=m" (p[d]));
129 	}
130 
131 	asm volatile("sfence" : : : "memory");
132 	kernel_fpu_end();
133 }
134 
135 const struct raid6_calls raid6_sse2x1 = {
136 	raid6_sse21_gen_syndrome,
137 	raid6_sse21_xor_syndrome,
138 	raid6_have_sse2,
139 	"sse2x1",
140 	1			/* Has cache hints */
141 };
142 
143 /*
144  * Unrolled-by-2 SSE2 implementation
145  */
raid6_sse22_gen_syndrome(int disks,size_t bytes,void ** ptrs)146 static void raid6_sse22_gen_syndrome(int disks, size_t bytes, void **ptrs)
147 {
148 	u8 **dptr = (u8 **)ptrs;
149 	u8 *p, *q;
150 	int d, z, z0;
151 
152 	z0 = disks - 3;		/* Highest data disk */
153 	p = dptr[z0+1];		/* XOR parity */
154 	q = dptr[z0+2];		/* RS syndrome */
155 
156 	kernel_fpu_begin();
157 
158 	asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
159 	asm volatile("pxor %xmm5,%xmm5"); /* Zero temp */
160 	asm volatile("pxor %xmm7,%xmm7"); /* Zero temp */
161 
162 	/* We uniformly assume a single prefetch covers at least 32 bytes */
163 	for ( d = 0 ; d < bytes ; d += 32 ) {
164 		asm volatile("prefetchnta %0" : : "m" (dptr[z0][d]));
165 		asm volatile("movdqa %0,%%xmm2" : : "m" (dptr[z0][d]));    /* P[0] */
166 		asm volatile("movdqa %0,%%xmm3" : : "m" (dptr[z0][d+16])); /* P[1] */
167 		asm volatile("movdqa %xmm2,%xmm4"); /* Q[0] */
168 		asm volatile("movdqa %xmm3,%xmm6"); /* Q[1] */
169 		for ( z = z0-1 ; z >= 0 ; z-- ) {
170 			asm volatile("prefetchnta %0" : : "m" (dptr[z][d]));
171 			asm volatile("pcmpgtb %xmm4,%xmm5");
172 			asm volatile("pcmpgtb %xmm6,%xmm7");
173 			asm volatile("paddb %xmm4,%xmm4");
174 			asm volatile("paddb %xmm6,%xmm6");
175 			asm volatile("pand %xmm0,%xmm5");
176 			asm volatile("pand %xmm0,%xmm7");
177 			asm volatile("pxor %xmm5,%xmm4");
178 			asm volatile("pxor %xmm7,%xmm6");
179 			asm volatile("movdqa %0,%%xmm5" : : "m" (dptr[z][d]));
180 			asm volatile("movdqa %0,%%xmm7" : : "m" (dptr[z][d+16]));
181 			asm volatile("pxor %xmm5,%xmm2");
182 			asm volatile("pxor %xmm7,%xmm3");
183 			asm volatile("pxor %xmm5,%xmm4");
184 			asm volatile("pxor %xmm7,%xmm6");
185 			asm volatile("pxor %xmm5,%xmm5");
186 			asm volatile("pxor %xmm7,%xmm7");
187 		}
188 		asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
189 		asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16]));
190 		asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
191 		asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16]));
192 	}
193 
194 	asm volatile("sfence" : : : "memory");
195 	kernel_fpu_end();
196 }
197 
raid6_sse22_xor_syndrome(int disks,int start,int stop,size_t bytes,void ** ptrs)198 static void raid6_sse22_xor_syndrome(int disks, int start, int stop,
199 				     size_t bytes, void **ptrs)
200 {
201 	u8 **dptr = (u8 **)ptrs;
202 	u8 *p, *q;
203 	int d, z, z0;
204 
205 	z0 = stop;		/* P/Q right side optimization */
206 	p = dptr[disks-2];	/* XOR parity */
207 	q = dptr[disks-1];	/* RS syndrome */
208 
209 	kernel_fpu_begin();
210 
211 	asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0]));
212 
213 	for ( d = 0 ; d < bytes ; d += 32 ) {
214 		asm volatile("movdqa %0,%%xmm4" :: "m" (dptr[z0][d]));
215 		asm volatile("movdqa %0,%%xmm6" :: "m" (dptr[z0][d+16]));
216 		asm volatile("movdqa %0,%%xmm2" : : "m" (p[d]));
217 		asm volatile("movdqa %0,%%xmm3" : : "m" (p[d+16]));
218 		asm volatile("pxor %xmm4,%xmm2");
219 		asm volatile("pxor %xmm6,%xmm3");
220 		/* P/Q data pages */
221 		for ( z = z0-1 ; z >= start ; z-- ) {
222 			asm volatile("pxor %xmm5,%xmm5");
223 			asm volatile("pxor %xmm7,%xmm7");
224 			asm volatile("pcmpgtb %xmm4,%xmm5");
225 			asm volatile("pcmpgtb %xmm6,%xmm7");
226 			asm volatile("paddb %xmm4,%xmm4");
227 			asm volatile("paddb %xmm6,%xmm6");
228 			asm volatile("pand %xmm0,%xmm5");
229 			asm volatile("pand %xmm0,%xmm7");
230 			asm volatile("pxor %xmm5,%xmm4");
231 			asm volatile("pxor %xmm7,%xmm6");
232 			asm volatile("movdqa %0,%%xmm5" :: "m" (dptr[z][d]));
233 			asm volatile("movdqa %0,%%xmm7" :: "m" (dptr[z][d+16]));
234 			asm volatile("pxor %xmm5,%xmm2");
235 			asm volatile("pxor %xmm7,%xmm3");
236 			asm volatile("pxor %xmm5,%xmm4");
237 			asm volatile("pxor %xmm7,%xmm6");
238 		}
239 		/* P/Q left side optimization */
240 		for ( z = start-1 ; z >= 0 ; z-- ) {
241 			asm volatile("pxor %xmm5,%xmm5");
242 			asm volatile("pxor %xmm7,%xmm7");
243 			asm volatile("pcmpgtb %xmm4,%xmm5");
244 			asm volatile("pcmpgtb %xmm6,%xmm7");
245 			asm volatile("paddb %xmm4,%xmm4");
246 			asm volatile("paddb %xmm6,%xmm6");
247 			asm volatile("pand %xmm0,%xmm5");
248 			asm volatile("pand %xmm0,%xmm7");
249 			asm volatile("pxor %xmm5,%xmm4");
250 			asm volatile("pxor %xmm7,%xmm6");
251 		}
252 		asm volatile("pxor %0,%%xmm4" : : "m" (q[d]));
253 		asm volatile("pxor %0,%%xmm6" : : "m" (q[d+16]));
254 		/* Don't use movntdq for r/w memory area < cache line */
255 		asm volatile("movdqa %%xmm4,%0" : "=m" (q[d]));
256 		asm volatile("movdqa %%xmm6,%0" : "=m" (q[d+16]));
257 		asm volatile("movdqa %%xmm2,%0" : "=m" (p[d]));
258 		asm volatile("movdqa %%xmm3,%0" : "=m" (p[d+16]));
259 	}
260 
261 	asm volatile("sfence" : : : "memory");
262 	kernel_fpu_end();
263 }
264 
265 const struct raid6_calls raid6_sse2x2 = {
266 	raid6_sse22_gen_syndrome,
267 	raid6_sse22_xor_syndrome,
268 	raid6_have_sse2,
269 	"sse2x2",
270 	1			/* Has cache hints */
271 };
272 
273 #ifdef CONFIG_X86_64
274 
275 /*
276  * Unrolled-by-4 SSE2 implementation
277  */
raid6_sse24_gen_syndrome(int disks,size_t bytes,void ** ptrs)278 static void raid6_sse24_gen_syndrome(int disks, size_t bytes, void **ptrs)
279 {
280 	u8 **dptr = (u8 **)ptrs;
281 	u8 *p, *q;
282 	int d, z, z0;
283 
284 	z0 = disks - 3;		/* Highest data disk */
285 	p = dptr[z0+1];		/* XOR parity */
286 	q = dptr[z0+2];		/* RS syndrome */
287 
288 	kernel_fpu_begin();
289 
290 	asm volatile("movdqa %0,%%xmm0" :: "m" (raid6_sse_constants.x1d[0]));
291 	asm volatile("pxor %xmm2,%xmm2");	/* P[0] */
292 	asm volatile("pxor %xmm3,%xmm3");	/* P[1] */
293 	asm volatile("pxor %xmm4,%xmm4"); 	/* Q[0] */
294 	asm volatile("pxor %xmm5,%xmm5");	/* Zero temp */
295 	asm volatile("pxor %xmm6,%xmm6"); 	/* Q[1] */
296 	asm volatile("pxor %xmm7,%xmm7"); 	/* Zero temp */
297 	asm volatile("pxor %xmm10,%xmm10");	/* P[2] */
298 	asm volatile("pxor %xmm11,%xmm11");	/* P[3] */
299 	asm volatile("pxor %xmm12,%xmm12"); 	/* Q[2] */
300 	asm volatile("pxor %xmm13,%xmm13");	/* Zero temp */
301 	asm volatile("pxor %xmm14,%xmm14"); 	/* Q[3] */
302 	asm volatile("pxor %xmm15,%xmm15"); 	/* Zero temp */
303 
304 	for ( d = 0 ; d < bytes ; d += 64 ) {
305 		for ( z = z0 ; z >= 0 ; z-- ) {
306 			/* The second prefetch seems to improve performance... */
307 			asm volatile("prefetchnta %0" :: "m" (dptr[z][d]));
308 			asm volatile("prefetchnta %0" :: "m" (dptr[z][d+32]));
309 			asm volatile("pcmpgtb %xmm4,%xmm5");
310 			asm volatile("pcmpgtb %xmm6,%xmm7");
311 			asm volatile("pcmpgtb %xmm12,%xmm13");
312 			asm volatile("pcmpgtb %xmm14,%xmm15");
313 			asm volatile("paddb %xmm4,%xmm4");
314 			asm volatile("paddb %xmm6,%xmm6");
315 			asm volatile("paddb %xmm12,%xmm12");
316 			asm volatile("paddb %xmm14,%xmm14");
317 			asm volatile("pand %xmm0,%xmm5");
318 			asm volatile("pand %xmm0,%xmm7");
319 			asm volatile("pand %xmm0,%xmm13");
320 			asm volatile("pand %xmm0,%xmm15");
321 			asm volatile("pxor %xmm5,%xmm4");
322 			asm volatile("pxor %xmm7,%xmm6");
323 			asm volatile("pxor %xmm13,%xmm12");
324 			asm volatile("pxor %xmm15,%xmm14");
325 			asm volatile("movdqa %0,%%xmm5" :: "m" (dptr[z][d]));
326 			asm volatile("movdqa %0,%%xmm7" :: "m" (dptr[z][d+16]));
327 			asm volatile("movdqa %0,%%xmm13" :: "m" (dptr[z][d+32]));
328 			asm volatile("movdqa %0,%%xmm15" :: "m" (dptr[z][d+48]));
329 			asm volatile("pxor %xmm5,%xmm2");
330 			asm volatile("pxor %xmm7,%xmm3");
331 			asm volatile("pxor %xmm13,%xmm10");
332 			asm volatile("pxor %xmm15,%xmm11");
333 			asm volatile("pxor %xmm5,%xmm4");
334 			asm volatile("pxor %xmm7,%xmm6");
335 			asm volatile("pxor %xmm13,%xmm12");
336 			asm volatile("pxor %xmm15,%xmm14");
337 			asm volatile("pxor %xmm5,%xmm5");
338 			asm volatile("pxor %xmm7,%xmm7");
339 			asm volatile("pxor %xmm13,%xmm13");
340 			asm volatile("pxor %xmm15,%xmm15");
341 		}
342 		asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
343 		asm volatile("pxor %xmm2,%xmm2");
344 		asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16]));
345 		asm volatile("pxor %xmm3,%xmm3");
346 		asm volatile("movntdq %%xmm10,%0" : "=m" (p[d+32]));
347 		asm volatile("pxor %xmm10,%xmm10");
348 		asm volatile("movntdq %%xmm11,%0" : "=m" (p[d+48]));
349 		asm volatile("pxor %xmm11,%xmm11");
350 		asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
351 		asm volatile("pxor %xmm4,%xmm4");
352 		asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16]));
353 		asm volatile("pxor %xmm6,%xmm6");
354 		asm volatile("movntdq %%xmm12,%0" : "=m" (q[d+32]));
355 		asm volatile("pxor %xmm12,%xmm12");
356 		asm volatile("movntdq %%xmm14,%0" : "=m" (q[d+48]));
357 		asm volatile("pxor %xmm14,%xmm14");
358 	}
359 
360 	asm volatile("sfence" : : : "memory");
361 	kernel_fpu_end();
362 }
363 
raid6_sse24_xor_syndrome(int disks,int start,int stop,size_t bytes,void ** ptrs)364 static void raid6_sse24_xor_syndrome(int disks, int start, int stop,
365 				     size_t bytes, void **ptrs)
366 {
367 	u8 **dptr = (u8 **)ptrs;
368 	u8 *p, *q;
369 	int d, z, z0;
370 
371 	z0 = stop;		/* P/Q right side optimization */
372 	p = dptr[disks-2];	/* XOR parity */
373 	q = dptr[disks-1];	/* RS syndrome */
374 
375 	kernel_fpu_begin();
376 
377 	asm volatile("movdqa %0,%%xmm0" :: "m" (raid6_sse_constants.x1d[0]));
378 
379 	for ( d = 0 ; d < bytes ; d += 64 ) {
380 		asm volatile("movdqa %0,%%xmm4" :: "m" (dptr[z0][d]));
381 		asm volatile("movdqa %0,%%xmm6" :: "m" (dptr[z0][d+16]));
382 		asm volatile("movdqa %0,%%xmm12" :: "m" (dptr[z0][d+32]));
383 		asm volatile("movdqa %0,%%xmm14" :: "m" (dptr[z0][d+48]));
384 		asm volatile("movdqa %0,%%xmm2" : : "m" (p[d]));
385 		asm volatile("movdqa %0,%%xmm3" : : "m" (p[d+16]));
386 		asm volatile("movdqa %0,%%xmm10" : : "m" (p[d+32]));
387 		asm volatile("movdqa %0,%%xmm11" : : "m" (p[d+48]));
388 		asm volatile("pxor %xmm4,%xmm2");
389 		asm volatile("pxor %xmm6,%xmm3");
390 		asm volatile("pxor %xmm12,%xmm10");
391 		asm volatile("pxor %xmm14,%xmm11");
392 		/* P/Q data pages */
393 		for ( z = z0-1 ; z >= start ; z-- ) {
394 			asm volatile("prefetchnta %0" :: "m" (dptr[z][d]));
395 			asm volatile("prefetchnta %0" :: "m" (dptr[z][d+32]));
396 			asm volatile("pxor %xmm5,%xmm5");
397 			asm volatile("pxor %xmm7,%xmm7");
398 			asm volatile("pxor %xmm13,%xmm13");
399 			asm volatile("pxor %xmm15,%xmm15");
400 			asm volatile("pcmpgtb %xmm4,%xmm5");
401 			asm volatile("pcmpgtb %xmm6,%xmm7");
402 			asm volatile("pcmpgtb %xmm12,%xmm13");
403 			asm volatile("pcmpgtb %xmm14,%xmm15");
404 			asm volatile("paddb %xmm4,%xmm4");
405 			asm volatile("paddb %xmm6,%xmm6");
406 			asm volatile("paddb %xmm12,%xmm12");
407 			asm volatile("paddb %xmm14,%xmm14");
408 			asm volatile("pand %xmm0,%xmm5");
409 			asm volatile("pand %xmm0,%xmm7");
410 			asm volatile("pand %xmm0,%xmm13");
411 			asm volatile("pand %xmm0,%xmm15");
412 			asm volatile("pxor %xmm5,%xmm4");
413 			asm volatile("pxor %xmm7,%xmm6");
414 			asm volatile("pxor %xmm13,%xmm12");
415 			asm volatile("pxor %xmm15,%xmm14");
416 			asm volatile("movdqa %0,%%xmm5" :: "m" (dptr[z][d]));
417 			asm volatile("movdqa %0,%%xmm7" :: "m" (dptr[z][d+16]));
418 			asm volatile("movdqa %0,%%xmm13" :: "m" (dptr[z][d+32]));
419 			asm volatile("movdqa %0,%%xmm15" :: "m" (dptr[z][d+48]));
420 			asm volatile("pxor %xmm5,%xmm2");
421 			asm volatile("pxor %xmm7,%xmm3");
422 			asm volatile("pxor %xmm13,%xmm10");
423 			asm volatile("pxor %xmm15,%xmm11");
424 			asm volatile("pxor %xmm5,%xmm4");
425 			asm volatile("pxor %xmm7,%xmm6");
426 			asm volatile("pxor %xmm13,%xmm12");
427 			asm volatile("pxor %xmm15,%xmm14");
428 		}
429 		asm volatile("prefetchnta %0" :: "m" (q[d]));
430 		asm volatile("prefetchnta %0" :: "m" (q[d+32]));
431 		/* P/Q left side optimization */
432 		for ( z = start-1 ; z >= 0 ; z-- ) {
433 			asm volatile("pxor %xmm5,%xmm5");
434 			asm volatile("pxor %xmm7,%xmm7");
435 			asm volatile("pxor %xmm13,%xmm13");
436 			asm volatile("pxor %xmm15,%xmm15");
437 			asm volatile("pcmpgtb %xmm4,%xmm5");
438 			asm volatile("pcmpgtb %xmm6,%xmm7");
439 			asm volatile("pcmpgtb %xmm12,%xmm13");
440 			asm volatile("pcmpgtb %xmm14,%xmm15");
441 			asm volatile("paddb %xmm4,%xmm4");
442 			asm volatile("paddb %xmm6,%xmm6");
443 			asm volatile("paddb %xmm12,%xmm12");
444 			asm volatile("paddb %xmm14,%xmm14");
445 			asm volatile("pand %xmm0,%xmm5");
446 			asm volatile("pand %xmm0,%xmm7");
447 			asm volatile("pand %xmm0,%xmm13");
448 			asm volatile("pand %xmm0,%xmm15");
449 			asm volatile("pxor %xmm5,%xmm4");
450 			asm volatile("pxor %xmm7,%xmm6");
451 			asm volatile("pxor %xmm13,%xmm12");
452 			asm volatile("pxor %xmm15,%xmm14");
453 		}
454 		asm volatile("movntdq %%xmm2,%0" : "=m" (p[d]));
455 		asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16]));
456 		asm volatile("movntdq %%xmm10,%0" : "=m" (p[d+32]));
457 		asm volatile("movntdq %%xmm11,%0" : "=m" (p[d+48]));
458 		asm volatile("pxor %0,%%xmm4" : : "m" (q[d]));
459 		asm volatile("pxor %0,%%xmm6" : : "m" (q[d+16]));
460 		asm volatile("pxor %0,%%xmm12" : : "m" (q[d+32]));
461 		asm volatile("pxor %0,%%xmm14" : : "m" (q[d+48]));
462 		asm volatile("movntdq %%xmm4,%0" : "=m" (q[d]));
463 		asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16]));
464 		asm volatile("movntdq %%xmm12,%0" : "=m" (q[d+32]));
465 		asm volatile("movntdq %%xmm14,%0" : "=m" (q[d+48]));
466 	}
467 	asm volatile("sfence" : : : "memory");
468 	kernel_fpu_end();
469 }
470 
471 
472 const struct raid6_calls raid6_sse2x4 = {
473 	raid6_sse24_gen_syndrome,
474 	raid6_sse24_xor_syndrome,
475 	raid6_have_sse2,
476 	"sse2x4",
477 	1			/* Has cache hints */
478 };
479 
480 #endif /* CONFIG_X86_64 */
481