1 // SPDX-License-Identifier: GPL-2.0-only
22 	const u8 x0f = 0x0f;  in raid6_2data_recov_avx2()
24 	p = (u8 *)ptrs[disks-2];  in raid6_2data_recov_avx2()
25 	q = (u8 *)ptrs[disks-1];  in raid6_2data_recov_avx2()
32 	ptrs[disks-2] = dp;  in raid6_2data_recov_avx2()
35 	ptrs[disks-1] = dq;  in raid6_2data_recov_avx2()
42 	ptrs[disks-2] = p;  in raid6_2data_recov_avx2()
43 	ptrs[disks-1] = q;  in raid6_2data_recov_avx2()
46 	pbmul = raid6_vgfmul[raid6_gfexi[failb-faila]];  in raid6_2data_recov_avx2()
53 	asm volatile("vpbroadcastb %0, %%ymm7" : : "m" (x0f));  in raid6_2data_recov_avx2()
57 		asm volatile("vmovdqa %0, %%ymm1" : : "m" (q[0]));  in raid6_2data_recov_avx2()
58 		asm volatile("vmovdqa %0, %%ymm9" : : "m" (q[32]));  in raid6_2data_recov_avx2()
59 		asm volatile("vmovdqa %0, %%ymm0" : : "m" (p[0]));  in raid6_2data_recov_avx2()
60 		asm volatile("vmovdqa %0, %%ymm8" : : "m" (p[32]));  in raid6_2data_recov_avx2()
61 		asm volatile("vpxor %0, %%ymm1, %%ymm1" : : "m" (dq[0]));  in raid6_2data_recov_avx2()
62 		asm volatile("vpxor %0, %%ymm9, %%ymm9" : : "m" (dq[32]));  in raid6_2data_recov_avx2()
63 		asm volatile("vpxor %0, %%ymm0, %%ymm0" : : "m" (dp[0]));  in raid6_2data_recov_avx2()
64 		asm volatile("vpxor %0, %%ymm8, %%ymm8" : : "m" (dp[32]));  in raid6_2data_recov_avx2()
67 		 * 1 = dq[0]  ^ q[0]  in raid6_2data_recov_avx2()
68 		 * 9 = dq[32] ^ q[32]  in raid6_2data_recov_avx2()
69 		 * 0 = dp[0]  ^ p[0]  in raid6_2data_recov_avx2()
70 		 * 8 = dp[32] ^ p[32]  in raid6_2data_recov_avx2()
73 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (qmul[0]));  in raid6_2data_recov_avx2()
74 		asm volatile("vbroadcasti128 %0, %%ymm5" : : "m" (qmul[16]));  in raid6_2data_recov_avx2()
90 		 * 5 = qx[0]  in raid6_2data_recov_avx2()
91 		 * 15 = qx[32]  in raid6_2data_recov_avx2()
94 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (pbmul[0]));  in raid6_2data_recov_avx2()
95 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (pbmul[16]));  in raid6_2data_recov_avx2()
110 		 * 1  = pbmul[px[0]]  in raid6_2data_recov_avx2()
111 		 * 13 = pbmul[px[32]]  in raid6_2data_recov_avx2()
118 		 * 13 = db[32] = DQ[32]  in raid6_2data_recov_avx2()
120 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));  in raid6_2data_recov_avx2()
121 		asm volatile("vmovdqa %%ymm13,%0" : "=m" (dq[32]));  in raid6_2data_recov_avx2()
125 		asm volatile("vmovdqa %%ymm0, %0" : "=m" (dp[0]));  in raid6_2data_recov_avx2()
126 		asm volatile("vmovdqa %%ymm8, %0" : "=m" (dp[32]));  in raid6_2data_recov_avx2()
128 		bytes -= 64;  in raid6_2data_recov_avx2()
134 		asm volatile("vmovdqa %0, %%ymm1" : : "m" (*q));  in raid6_2data_recov_avx2()
135 		asm volatile("vmovdqa %0, %%ymm0" : : "m" (*p));  in raid6_2data_recov_avx2()
136 		asm volatile("vpxor %0, %%ymm1, %%ymm1" : : "m" (*dq));  in raid6_2data_recov_avx2()
137 		asm volatile("vpxor %0, %%ymm0, %%ymm0" : : "m" (*dp));  in raid6_2data_recov_avx2()
139 		/* 1 = dq ^ q;  0 = dp ^ p */  in raid6_2data_recov_avx2()
141 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (qmul[0]));  in raid6_2data_recov_avx2()
142 		asm volatile("vbroadcasti128 %0, %%ymm5" : : "m" (qmul[16]));  in raid6_2data_recov_avx2()
157 		asm volatile("vbroadcasti128 %0, %%ymm4" : : "m" (pbmul[0]));  in raid6_2data_recov_avx2()
158 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (pbmul[16]));  in raid6_2data_recov_avx2()
170 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));  in raid6_2data_recov_avx2()
173 		asm volatile("vmovdqa %%ymm0, %0" : "=m" (dp[0]));  in raid6_2data_recov_avx2()
175 		bytes -= 32;  in raid6_2data_recov_avx2()
176 		p += 32;  in raid6_2data_recov_avx2()
177 		q += 32;  in raid6_2data_recov_avx2()
178 		dp += 32;  in raid6_2data_recov_avx2()
179 		dq += 32;  in raid6_2data_recov_avx2()
191 	const u8 x0f = 0x0f;  in raid6_datap_recov_avx2()
193 	p = (u8 *)ptrs[disks-2];  in raid6_datap_recov_avx2()
194 	q = (u8 *)ptrs[disks-1];  in raid6_datap_recov_avx2()
200 	ptrs[disks-1] = dq;  in raid6_datap_recov_avx2()
206 	ptrs[disks-1] = q;  in raid6_datap_recov_avx2()
213 	asm volatile("vpbroadcastb %0, %%ymm7" : : "m" (x0f));  in raid6_datap_recov_avx2()
217 		asm volatile("vmovdqa %0, %%ymm3" : : "m" (dq[0]));  in raid6_datap_recov_avx2()
218 		asm volatile("vmovdqa %0, %%ymm8" : : "m" (dq[32]));  in raid6_datap_recov_avx2()
219 		asm volatile("vpxor %0, %%ymm3, %%ymm3" : : "m" (q[0]));  in raid6_datap_recov_avx2()
220 		asm volatile("vpxor %0, %%ymm8, %%ymm8" : : "m" (q[32]));  in raid6_datap_recov_avx2()
223 		 * 3 = q[0] ^ dq[0]  in raid6_datap_recov_avx2()
224 		 * 8 = q[32] ^ dq[32]  in raid6_datap_recov_avx2()
226 		asm volatile("vbroadcasti128 %0, %%ymm0" : : "m" (qmul[0]));  in raid6_datap_recov_avx2()
228 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (qmul[16]));  in raid6_datap_recov_avx2()
245 		 * 1  = qmul[q[0]  ^ dq[0]]  in raid6_datap_recov_avx2()
246 		 * 14 = qmul[q[32] ^ dq[32]]  in raid6_datap_recov_avx2()
248 		asm volatile("vmovdqa %0, %%ymm2" : : "m" (p[0]));  in raid6_datap_recov_avx2()
249 		asm volatile("vmovdqa %0, %%ymm12" : : "m" (p[32]));  in raid6_datap_recov_avx2()
254 		 * 2  = p[0]  ^ qmul[q[0]  ^ dq[0]]  in raid6_datap_recov_avx2()
255 		 * 12 = p[32] ^ qmul[q[32] ^ dq[32]]  in raid6_datap_recov_avx2()
258 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));  in raid6_datap_recov_avx2()
259 		asm volatile("vmovdqa %%ymm14, %0" : "=m" (dq[32]));  in raid6_datap_recov_avx2()
260 		asm volatile("vmovdqa %%ymm2, %0" : "=m" (p[0]));  in raid6_datap_recov_avx2()
261 		asm volatile("vmovdqa %%ymm12,%0" : "=m" (p[32]));  in raid6_datap_recov_avx2()
263 		bytes -= 64;  in raid6_datap_recov_avx2()
268 		asm volatile("vmovdqa %0, %%ymm3" : : "m" (dq[0]));  in raid6_datap_recov_avx2()
269 		asm volatile("vpxor %0, %%ymm3, %%ymm3" : : "m" (q[0]));  in raid6_datap_recov_avx2()
273 		asm volatile("vbroadcasti128 %0, %%ymm0" : : "m" (qmul[0]));  in raid6_datap_recov_avx2()
274 		asm volatile("vbroadcasti128 %0, %%ymm1" : : "m" (qmul[16]));  in raid6_datap_recov_avx2()
285 		asm volatile("vmovdqa %0, %%ymm2" : : "m" (p[0]));  in raid6_datap_recov_avx2()
290 		asm volatile("vmovdqa %%ymm1, %0" : "=m" (dq[0]));  in raid6_datap_recov_avx2()
291 		asm volatile("vmovdqa %%ymm2, %0" : "=m" (p[0]));  in raid6_datap_recov_avx2()
293 		bytes -= 32;  in raid6_datap_recov_avx2()
294 		p += 32;  in raid6_datap_recov_avx2()
295 		q += 32;  in raid6_datap_recov_avx2()
296 		dq += 32;  in raid6_datap_recov_avx2()