]> git.proxmox.com Git - mirror_zfs.git/blame - module/zfs/vdev_raidz_math_sse2.c
ddt: modernise assertions
[mirror_zfs.git] / module / zfs / vdev_raidz_math_sse2.c
CommitLineData
ae25d222
GN
1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
1d3ba0bf 9 * or https://opensource.org/licenses/CDDL-1.0.
ae25d222
GN
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (C) 2016 Gvozden Nešković. All rights reserved.
23 */
24
25#include <sys/isa_defs.h>
26
27#if defined(__x86_64) && defined(HAVE_SSE2)
28
29#include <sys/types.h>
006e9a40 30#include <sys/simd.h>
c5858ff9 31#include <sys/debug.h>
ae25d222 32
3df0c2fa 33#ifdef __linux__
ae25d222 34#define __asm __asm__ __volatile__
3df0c2fa 35#endif
ae25d222
GN
36
37#define _REG_CNT(_0, _1, _2, _3, _4, _5, _6, _7, N, ...) N
38#define REG_CNT(r...) _REG_CNT(r, 8, 7, 6, 5, 4, 3, 2, 1)
39
40#define VR0_(REG, ...) "xmm"#REG
41#define VR1_(_1, REG, ...) "xmm"#REG
42#define VR2_(_1, _2, REG, ...) "xmm"#REG
43#define VR3_(_1, _2, _3, REG, ...) "xmm"#REG
44#define VR4_(_1, _2, _3, _4, REG, ...) "xmm"#REG
45#define VR5_(_1, _2, _3, _4, _5, REG, ...) "xmm"#REG
46#define VR6_(_1, _2, _3, _4, _5, _6, REG, ...) "xmm"#REG
47#define VR7_(_1, _2, _3, _4, _5, _6, _7, REG, ...) "xmm"#REG
48
49#define VR0(r...) VR0_(r, 1, 2, 3, 4, 5, 6)
50#define VR1(r...) VR1_(r, 1, 2, 3, 4, 5, 6)
51#define VR2(r...) VR2_(r, 1, 2, 3, 4, 5, 6)
52#define VR3(r...) VR3_(r, 1, 2, 3, 4, 5, 6)
53#define VR4(r...) VR4_(r, 1, 2, 3, 4, 5, 6)
54#define VR5(r...) VR5_(r, 1, 2, 3, 4, 5, 6)
55#define VR6(r...) VR6_(r, 1, 2, 3, 4, 5, 6)
56#define VR7(r...) VR7_(r, 1, 2, 3, 4, 5, 6)
57
58#define ELEM_SIZE 16
59
60typedef struct v {
61 uint8_t b[ELEM_SIZE] __attribute__((aligned(ELEM_SIZE)));
62} v_t;
63
ae25d222
GN
64#define XOR_ACC(src, r...) \
65{ \
66 switch (REG_CNT(r)) { \
67 case 4: \
68 __asm( \
69 "pxor 0x00(%[SRC]), %%" VR0(r) "\n" \
70 "pxor 0x10(%[SRC]), %%" VR1(r) "\n" \
71 "pxor 0x20(%[SRC]), %%" VR2(r) "\n" \
72 "pxor 0x30(%[SRC]), %%" VR3(r) "\n" \
73 : : [SRC] "r" (src)); \
74 break; \
75 case 2: \
76 __asm( \
77 "pxor 0x00(%[SRC]), %%" VR0(r) "\n" \
78 "pxor 0x10(%[SRC]), %%" VR1(r) "\n" \
79 : : [SRC] "r" (src)); \
80 break; \
81 case 1: \
82 __asm("pxor 0x00(%[SRC]), %%" VR0(r) "\n" \
83 : : [SRC] "r" (src)); \
84 break; \
85 } \
86}
87
88#define XOR(r...) \
89{ \
90 switch (REG_CNT(r)) { \
91 case 8: \
92 __asm( \
93 "pxor %" VR0(r) ", %" VR4(r) "\n" \
94 "pxor %" VR1(r) ", %" VR5(r) "\n" \
95 "pxor %" VR2(r) ", %" VR6(r) "\n" \
96 "pxor %" VR3(r) ", %" VR7(r)); \
97 break; \
98 case 4: \
99 __asm( \
100 "pxor %" VR0(r) ", %" VR2(r) "\n" \
101 "pxor %" VR1(r) ", %" VR3(r)); \
102 break; \
103 case 2: \
104 __asm( \
105 "pxor %" VR0(r) ", %" VR1(r)); \
106 break; \
107 } \
108}
65d71d42
GN
109
110#define ZERO(r...) XOR(r, r)
ae25d222
GN
111
112#define COPY(r...) \
113{ \
114 switch (REG_CNT(r)) { \
115 case 8: \
116 __asm( \
117 "movdqa %" VR0(r) ", %" VR4(r) "\n" \
118 "movdqa %" VR1(r) ", %" VR5(r) "\n" \
119 "movdqa %" VR2(r) ", %" VR6(r) "\n" \
120 "movdqa %" VR3(r) ", %" VR7(r)); \
121 break; \
122 case 4: \
123 __asm( \
124 "movdqa %" VR0(r) ", %" VR2(r) "\n" \
125 "movdqa %" VR1(r) ", %" VR3(r)); \
126 break; \
127 case 2: \
128 __asm( \
129 "movdqa %" VR0(r) ", %" VR1(r)); \
130 break; \
c5858ff9
MM
131 default: \
132 VERIFY(0); \
ae25d222
GN
133 } \
134}
135
136#define LOAD(src, r...) \
137{ \
138 switch (REG_CNT(r)) { \
139 case 4: \
140 __asm( \
141 "movdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
142 "movdqa 0x10(%[SRC]), %%" VR1(r) "\n" \
143 "movdqa 0x20(%[SRC]), %%" VR2(r) "\n" \
144 "movdqa 0x30(%[SRC]), %%" VR3(r) "\n" \
145 : : [SRC] "r" (src)); \
146 break; \
147 case 2: \
148 __asm( \
149 "movdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
150 "movdqa 0x10(%[SRC]), %%" VR1(r) "\n" \
151 : : [SRC] "r" (src)); \
152 break; \
153 case 1: \
154 __asm( \
155 "movdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
156 : : [SRC] "r" (src)); \
157 break; \
158 } \
159}
160
161#define STORE(dst, r...) \
162{ \
163 switch (REG_CNT(r)) { \
164 case 4: \
165 __asm( \
166 "movdqa %%" VR0(r)", 0x00(%[DST])\n" \
167 "movdqa %%" VR1(r)", 0x10(%[DST])\n" \
168 "movdqa %%" VR2(r)", 0x20(%[DST])\n" \
169 "movdqa %%" VR3(r)", 0x30(%[DST])\n" \
170 : : [DST] "r" (dst)); \
171 break; \
172 case 2: \
173 __asm( \
174 "movdqa %%" VR0(r)", 0x00(%[DST])\n" \
175 "movdqa %%" VR1(r)", 0x10(%[DST])\n" \
176 : : [DST] "r" (dst)); \
177 break; \
178 case 1: \
179 __asm( \
180 "movdqa %%" VR0(r)", 0x00(%[DST])\n" \
181 : : [DST] "r" (dst)); \
182 break; \
c5858ff9
MM
183 default: \
184 VERIFY(0); \
ae25d222
GN
185 } \
186}
187
188#define MUL2_SETUP() \
189{ \
190 __asm( \
191 "movd %[mask], %%xmm15\n" \
192 "pshufd $0x0, %%xmm15, %%xmm15\n" \
193 : : [mask] "r" (0x1d1d1d1d)); \
194}
195
196#define _MUL2_x1(a0) \
197{ \
198 __asm( \
199 "pxor %xmm14, %xmm14\n" \
200 "pcmpgtb %" a0", %xmm14\n" \
201 "pand %xmm15, %xmm14\n" \
202 "paddb %" a0", %" a0 "\n" \
203 "pxor %xmm14, %" a0); \
204}
205
206#define _MUL2_x2(a0, a1) \
207{ \
208 __asm( \
209 "pxor %xmm14, %xmm14\n" \
210 "pxor %xmm13, %xmm13\n" \
211 "pcmpgtb %" a0", %xmm14\n" \
212 "pcmpgtb %" a1", %xmm13\n" \
213 "pand %xmm15, %xmm14\n" \
214 "pand %xmm15, %xmm13\n" \
215 "paddb %" a0", %" a0 "\n" \
216 "paddb %" a1", %" a1 "\n" \
217 "pxor %xmm14, %" a0 "\n" \
218 "pxor %xmm13, %" a1); \
219}
220
221#define MUL2(r...) \
222{ \
223 switch (REG_CNT(r)) { \
cbf484f8
GN
224 case 4: \
225 _MUL2_x2(VR0(r), VR1(r)); \
226 _MUL2_x2(VR2(r), VR3(r)); \
227 break; \
ae25d222
GN
228 case 2: \
229 _MUL2_x2(VR0(r), VR1(r)); \
230 break; \
231 case 1: \
232 _MUL2_x1(VR0(r)); \
233 break; \
234 } \
235}
236
237#define MUL4(r...) \
238{ \
239 MUL2(r); \
240 MUL2(r); \
241}
242
243/* General multiplication by adding powers of two */
244
245#define _MUL_PARAM(x, in, acc) \
246{ \
65d71d42 247 if (x & 0x01) { COPY(in, acc); } else { ZERO(acc); } \
ae25d222
GN
248 if (x & 0xfe) { MUL2(in); } \
249 if (x & 0x02) { XOR(in, acc); } \
250 if (x & 0xfc) { MUL2(in); } \
251 if (x & 0x04) { XOR(in, acc); } \
252 if (x & 0xf8) { MUL2(in); } \
253 if (x & 0x08) { XOR(in, acc); } \
254 if (x & 0xf0) { MUL2(in); } \
255 if (x & 0x10) { XOR(in, acc); } \
256 if (x & 0xe0) { MUL2(in); } \
257 if (x & 0x20) { XOR(in, acc); } \
258 if (x & 0xc0) { MUL2(in); } \
259 if (x & 0x40) { XOR(in, acc); } \
260 if (x & 0x80) { MUL2(in); XOR(in, acc); } \
261}
262
cbf484f8
GN
263#define _mul_x1_in 11
264#define _mul_x1_acc 12
ae25d222
GN
265
266#define MUL_x1_DEFINE(x) \
267static void \
268mul_x1_ ## x(void) { _MUL_PARAM(x, _mul_x1_in, _mul_x1_acc); }
269
270#define _mul_x2_in 9, 10
271#define _mul_x2_acc 11, 12
272
273#define MUL_x2_DEFINE(x) \
274static void \
275mul_x2_ ## x(void) { _MUL_PARAM(x, _mul_x2_in, _mul_x2_acc); }
276
277MUL_x1_DEFINE(0); MUL_x1_DEFINE(1); MUL_x1_DEFINE(2); MUL_x1_DEFINE(3);
278MUL_x1_DEFINE(4); MUL_x1_DEFINE(5); MUL_x1_DEFINE(6); MUL_x1_DEFINE(7);
279MUL_x1_DEFINE(8); MUL_x1_DEFINE(9); MUL_x1_DEFINE(10); MUL_x1_DEFINE(11);
280MUL_x1_DEFINE(12); MUL_x1_DEFINE(13); MUL_x1_DEFINE(14); MUL_x1_DEFINE(15);
281MUL_x1_DEFINE(16); MUL_x1_DEFINE(17); MUL_x1_DEFINE(18); MUL_x1_DEFINE(19);
282MUL_x1_DEFINE(20); MUL_x1_DEFINE(21); MUL_x1_DEFINE(22); MUL_x1_DEFINE(23);
283MUL_x1_DEFINE(24); MUL_x1_DEFINE(25); MUL_x1_DEFINE(26); MUL_x1_DEFINE(27);
284MUL_x1_DEFINE(28); MUL_x1_DEFINE(29); MUL_x1_DEFINE(30); MUL_x1_DEFINE(31);
285MUL_x1_DEFINE(32); MUL_x1_DEFINE(33); MUL_x1_DEFINE(34); MUL_x1_DEFINE(35);
286MUL_x1_DEFINE(36); MUL_x1_DEFINE(37); MUL_x1_DEFINE(38); MUL_x1_DEFINE(39);
287MUL_x1_DEFINE(40); MUL_x1_DEFINE(41); MUL_x1_DEFINE(42); MUL_x1_DEFINE(43);
288MUL_x1_DEFINE(44); MUL_x1_DEFINE(45); MUL_x1_DEFINE(46); MUL_x1_DEFINE(47);
289MUL_x1_DEFINE(48); MUL_x1_DEFINE(49); MUL_x1_DEFINE(50); MUL_x1_DEFINE(51);
290MUL_x1_DEFINE(52); MUL_x1_DEFINE(53); MUL_x1_DEFINE(54); MUL_x1_DEFINE(55);
291MUL_x1_DEFINE(56); MUL_x1_DEFINE(57); MUL_x1_DEFINE(58); MUL_x1_DEFINE(59);
292MUL_x1_DEFINE(60); MUL_x1_DEFINE(61); MUL_x1_DEFINE(62); MUL_x1_DEFINE(63);
293MUL_x1_DEFINE(64); MUL_x1_DEFINE(65); MUL_x1_DEFINE(66); MUL_x1_DEFINE(67);
294MUL_x1_DEFINE(68); MUL_x1_DEFINE(69); MUL_x1_DEFINE(70); MUL_x1_DEFINE(71);
295MUL_x1_DEFINE(72); MUL_x1_DEFINE(73); MUL_x1_DEFINE(74); MUL_x1_DEFINE(75);
296MUL_x1_DEFINE(76); MUL_x1_DEFINE(77); MUL_x1_DEFINE(78); MUL_x1_DEFINE(79);
297MUL_x1_DEFINE(80); MUL_x1_DEFINE(81); MUL_x1_DEFINE(82); MUL_x1_DEFINE(83);
298MUL_x1_DEFINE(84); MUL_x1_DEFINE(85); MUL_x1_DEFINE(86); MUL_x1_DEFINE(87);
299MUL_x1_DEFINE(88); MUL_x1_DEFINE(89); MUL_x1_DEFINE(90); MUL_x1_DEFINE(91);
300MUL_x1_DEFINE(92); MUL_x1_DEFINE(93); MUL_x1_DEFINE(94); MUL_x1_DEFINE(95);
301MUL_x1_DEFINE(96); MUL_x1_DEFINE(97); MUL_x1_DEFINE(98); MUL_x1_DEFINE(99);
302MUL_x1_DEFINE(100); MUL_x1_DEFINE(101); MUL_x1_DEFINE(102); MUL_x1_DEFINE(103);
303MUL_x1_DEFINE(104); MUL_x1_DEFINE(105); MUL_x1_DEFINE(106); MUL_x1_DEFINE(107);
304MUL_x1_DEFINE(108); MUL_x1_DEFINE(109); MUL_x1_DEFINE(110); MUL_x1_DEFINE(111);
305MUL_x1_DEFINE(112); MUL_x1_DEFINE(113); MUL_x1_DEFINE(114); MUL_x1_DEFINE(115);
306MUL_x1_DEFINE(116); MUL_x1_DEFINE(117); MUL_x1_DEFINE(118); MUL_x1_DEFINE(119);
307MUL_x1_DEFINE(120); MUL_x1_DEFINE(121); MUL_x1_DEFINE(122); MUL_x1_DEFINE(123);
308MUL_x1_DEFINE(124); MUL_x1_DEFINE(125); MUL_x1_DEFINE(126); MUL_x1_DEFINE(127);
309MUL_x1_DEFINE(128); MUL_x1_DEFINE(129); MUL_x1_DEFINE(130); MUL_x1_DEFINE(131);
310MUL_x1_DEFINE(132); MUL_x1_DEFINE(133); MUL_x1_DEFINE(134); MUL_x1_DEFINE(135);
311MUL_x1_DEFINE(136); MUL_x1_DEFINE(137); MUL_x1_DEFINE(138); MUL_x1_DEFINE(139);
312MUL_x1_DEFINE(140); MUL_x1_DEFINE(141); MUL_x1_DEFINE(142); MUL_x1_DEFINE(143);
313MUL_x1_DEFINE(144); MUL_x1_DEFINE(145); MUL_x1_DEFINE(146); MUL_x1_DEFINE(147);
314MUL_x1_DEFINE(148); MUL_x1_DEFINE(149); MUL_x1_DEFINE(150); MUL_x1_DEFINE(151);
315MUL_x1_DEFINE(152); MUL_x1_DEFINE(153); MUL_x1_DEFINE(154); MUL_x1_DEFINE(155);
316MUL_x1_DEFINE(156); MUL_x1_DEFINE(157); MUL_x1_DEFINE(158); MUL_x1_DEFINE(159);
317MUL_x1_DEFINE(160); MUL_x1_DEFINE(161); MUL_x1_DEFINE(162); MUL_x1_DEFINE(163);
318MUL_x1_DEFINE(164); MUL_x1_DEFINE(165); MUL_x1_DEFINE(166); MUL_x1_DEFINE(167);
319MUL_x1_DEFINE(168); MUL_x1_DEFINE(169); MUL_x1_DEFINE(170); MUL_x1_DEFINE(171);
320MUL_x1_DEFINE(172); MUL_x1_DEFINE(173); MUL_x1_DEFINE(174); MUL_x1_DEFINE(175);
321MUL_x1_DEFINE(176); MUL_x1_DEFINE(177); MUL_x1_DEFINE(178); MUL_x1_DEFINE(179);
322MUL_x1_DEFINE(180); MUL_x1_DEFINE(181); MUL_x1_DEFINE(182); MUL_x1_DEFINE(183);
323MUL_x1_DEFINE(184); MUL_x1_DEFINE(185); MUL_x1_DEFINE(186); MUL_x1_DEFINE(187);
324MUL_x1_DEFINE(188); MUL_x1_DEFINE(189); MUL_x1_DEFINE(190); MUL_x1_DEFINE(191);
325MUL_x1_DEFINE(192); MUL_x1_DEFINE(193); MUL_x1_DEFINE(194); MUL_x1_DEFINE(195);
326MUL_x1_DEFINE(196); MUL_x1_DEFINE(197); MUL_x1_DEFINE(198); MUL_x1_DEFINE(199);
327MUL_x1_DEFINE(200); MUL_x1_DEFINE(201); MUL_x1_DEFINE(202); MUL_x1_DEFINE(203);
328MUL_x1_DEFINE(204); MUL_x1_DEFINE(205); MUL_x1_DEFINE(206); MUL_x1_DEFINE(207);
329MUL_x1_DEFINE(208); MUL_x1_DEFINE(209); MUL_x1_DEFINE(210); MUL_x1_DEFINE(211);
330MUL_x1_DEFINE(212); MUL_x1_DEFINE(213); MUL_x1_DEFINE(214); MUL_x1_DEFINE(215);
331MUL_x1_DEFINE(216); MUL_x1_DEFINE(217); MUL_x1_DEFINE(218); MUL_x1_DEFINE(219);
332MUL_x1_DEFINE(220); MUL_x1_DEFINE(221); MUL_x1_DEFINE(222); MUL_x1_DEFINE(223);
333MUL_x1_DEFINE(224); MUL_x1_DEFINE(225); MUL_x1_DEFINE(226); MUL_x1_DEFINE(227);
334MUL_x1_DEFINE(228); MUL_x1_DEFINE(229); MUL_x1_DEFINE(230); MUL_x1_DEFINE(231);
335MUL_x1_DEFINE(232); MUL_x1_DEFINE(233); MUL_x1_DEFINE(234); MUL_x1_DEFINE(235);
336MUL_x1_DEFINE(236); MUL_x1_DEFINE(237); MUL_x1_DEFINE(238); MUL_x1_DEFINE(239);
337MUL_x1_DEFINE(240); MUL_x1_DEFINE(241); MUL_x1_DEFINE(242); MUL_x1_DEFINE(243);
338MUL_x1_DEFINE(244); MUL_x1_DEFINE(245); MUL_x1_DEFINE(246); MUL_x1_DEFINE(247);
339MUL_x1_DEFINE(248); MUL_x1_DEFINE(249); MUL_x1_DEFINE(250); MUL_x1_DEFINE(251);
340MUL_x1_DEFINE(252); MUL_x1_DEFINE(253); MUL_x1_DEFINE(254); MUL_x1_DEFINE(255);
341
342MUL_x2_DEFINE(0); MUL_x2_DEFINE(1); MUL_x2_DEFINE(2); MUL_x2_DEFINE(3);
343MUL_x2_DEFINE(4); MUL_x2_DEFINE(5); MUL_x2_DEFINE(6); MUL_x2_DEFINE(7);
344MUL_x2_DEFINE(8); MUL_x2_DEFINE(9); MUL_x2_DEFINE(10); MUL_x2_DEFINE(11);
345MUL_x2_DEFINE(12); MUL_x2_DEFINE(13); MUL_x2_DEFINE(14); MUL_x2_DEFINE(15);
346MUL_x2_DEFINE(16); MUL_x2_DEFINE(17); MUL_x2_DEFINE(18); MUL_x2_DEFINE(19);
347MUL_x2_DEFINE(20); MUL_x2_DEFINE(21); MUL_x2_DEFINE(22); MUL_x2_DEFINE(23);
348MUL_x2_DEFINE(24); MUL_x2_DEFINE(25); MUL_x2_DEFINE(26); MUL_x2_DEFINE(27);
349MUL_x2_DEFINE(28); MUL_x2_DEFINE(29); MUL_x2_DEFINE(30); MUL_x2_DEFINE(31);
350MUL_x2_DEFINE(32); MUL_x2_DEFINE(33); MUL_x2_DEFINE(34); MUL_x2_DEFINE(35);
351MUL_x2_DEFINE(36); MUL_x2_DEFINE(37); MUL_x2_DEFINE(38); MUL_x2_DEFINE(39);
352MUL_x2_DEFINE(40); MUL_x2_DEFINE(41); MUL_x2_DEFINE(42); MUL_x2_DEFINE(43);
353MUL_x2_DEFINE(44); MUL_x2_DEFINE(45); MUL_x2_DEFINE(46); MUL_x2_DEFINE(47);
354MUL_x2_DEFINE(48); MUL_x2_DEFINE(49); MUL_x2_DEFINE(50); MUL_x2_DEFINE(51);
355MUL_x2_DEFINE(52); MUL_x2_DEFINE(53); MUL_x2_DEFINE(54); MUL_x2_DEFINE(55);
356MUL_x2_DEFINE(56); MUL_x2_DEFINE(57); MUL_x2_DEFINE(58); MUL_x2_DEFINE(59);
357MUL_x2_DEFINE(60); MUL_x2_DEFINE(61); MUL_x2_DEFINE(62); MUL_x2_DEFINE(63);
358MUL_x2_DEFINE(64); MUL_x2_DEFINE(65); MUL_x2_DEFINE(66); MUL_x2_DEFINE(67);
359MUL_x2_DEFINE(68); MUL_x2_DEFINE(69); MUL_x2_DEFINE(70); MUL_x2_DEFINE(71);
360MUL_x2_DEFINE(72); MUL_x2_DEFINE(73); MUL_x2_DEFINE(74); MUL_x2_DEFINE(75);
361MUL_x2_DEFINE(76); MUL_x2_DEFINE(77); MUL_x2_DEFINE(78); MUL_x2_DEFINE(79);
362MUL_x2_DEFINE(80); MUL_x2_DEFINE(81); MUL_x2_DEFINE(82); MUL_x2_DEFINE(83);
363MUL_x2_DEFINE(84); MUL_x2_DEFINE(85); MUL_x2_DEFINE(86); MUL_x2_DEFINE(87);
364MUL_x2_DEFINE(88); MUL_x2_DEFINE(89); MUL_x2_DEFINE(90); MUL_x2_DEFINE(91);
365MUL_x2_DEFINE(92); MUL_x2_DEFINE(93); MUL_x2_DEFINE(94); MUL_x2_DEFINE(95);
366MUL_x2_DEFINE(96); MUL_x2_DEFINE(97); MUL_x2_DEFINE(98); MUL_x2_DEFINE(99);
367MUL_x2_DEFINE(100); MUL_x2_DEFINE(101); MUL_x2_DEFINE(102); MUL_x2_DEFINE(103);
368MUL_x2_DEFINE(104); MUL_x2_DEFINE(105); MUL_x2_DEFINE(106); MUL_x2_DEFINE(107);
369MUL_x2_DEFINE(108); MUL_x2_DEFINE(109); MUL_x2_DEFINE(110); MUL_x2_DEFINE(111);
370MUL_x2_DEFINE(112); MUL_x2_DEFINE(113); MUL_x2_DEFINE(114); MUL_x2_DEFINE(115);
371MUL_x2_DEFINE(116); MUL_x2_DEFINE(117); MUL_x2_DEFINE(118); MUL_x2_DEFINE(119);
372MUL_x2_DEFINE(120); MUL_x2_DEFINE(121); MUL_x2_DEFINE(122); MUL_x2_DEFINE(123);
373MUL_x2_DEFINE(124); MUL_x2_DEFINE(125); MUL_x2_DEFINE(126); MUL_x2_DEFINE(127);
374MUL_x2_DEFINE(128); MUL_x2_DEFINE(129); MUL_x2_DEFINE(130); MUL_x2_DEFINE(131);
375MUL_x2_DEFINE(132); MUL_x2_DEFINE(133); MUL_x2_DEFINE(134); MUL_x2_DEFINE(135);
376MUL_x2_DEFINE(136); MUL_x2_DEFINE(137); MUL_x2_DEFINE(138); MUL_x2_DEFINE(139);
377MUL_x2_DEFINE(140); MUL_x2_DEFINE(141); MUL_x2_DEFINE(142); MUL_x2_DEFINE(143);
378MUL_x2_DEFINE(144); MUL_x2_DEFINE(145); MUL_x2_DEFINE(146); MUL_x2_DEFINE(147);
379MUL_x2_DEFINE(148); MUL_x2_DEFINE(149); MUL_x2_DEFINE(150); MUL_x2_DEFINE(151);
380MUL_x2_DEFINE(152); MUL_x2_DEFINE(153); MUL_x2_DEFINE(154); MUL_x2_DEFINE(155);
381MUL_x2_DEFINE(156); MUL_x2_DEFINE(157); MUL_x2_DEFINE(158); MUL_x2_DEFINE(159);
382MUL_x2_DEFINE(160); MUL_x2_DEFINE(161); MUL_x2_DEFINE(162); MUL_x2_DEFINE(163);
383MUL_x2_DEFINE(164); MUL_x2_DEFINE(165); MUL_x2_DEFINE(166); MUL_x2_DEFINE(167);
384MUL_x2_DEFINE(168); MUL_x2_DEFINE(169); MUL_x2_DEFINE(170); MUL_x2_DEFINE(171);
385MUL_x2_DEFINE(172); MUL_x2_DEFINE(173); MUL_x2_DEFINE(174); MUL_x2_DEFINE(175);
386MUL_x2_DEFINE(176); MUL_x2_DEFINE(177); MUL_x2_DEFINE(178); MUL_x2_DEFINE(179);
387MUL_x2_DEFINE(180); MUL_x2_DEFINE(181); MUL_x2_DEFINE(182); MUL_x2_DEFINE(183);
388MUL_x2_DEFINE(184); MUL_x2_DEFINE(185); MUL_x2_DEFINE(186); MUL_x2_DEFINE(187);
389MUL_x2_DEFINE(188); MUL_x2_DEFINE(189); MUL_x2_DEFINE(190); MUL_x2_DEFINE(191);
390MUL_x2_DEFINE(192); MUL_x2_DEFINE(193); MUL_x2_DEFINE(194); MUL_x2_DEFINE(195);
391MUL_x2_DEFINE(196); MUL_x2_DEFINE(197); MUL_x2_DEFINE(198); MUL_x2_DEFINE(199);
392MUL_x2_DEFINE(200); MUL_x2_DEFINE(201); MUL_x2_DEFINE(202); MUL_x2_DEFINE(203);
393MUL_x2_DEFINE(204); MUL_x2_DEFINE(205); MUL_x2_DEFINE(206); MUL_x2_DEFINE(207);
394MUL_x2_DEFINE(208); MUL_x2_DEFINE(209); MUL_x2_DEFINE(210); MUL_x2_DEFINE(211);
395MUL_x2_DEFINE(212); MUL_x2_DEFINE(213); MUL_x2_DEFINE(214); MUL_x2_DEFINE(215);
396MUL_x2_DEFINE(216); MUL_x2_DEFINE(217); MUL_x2_DEFINE(218); MUL_x2_DEFINE(219);
397MUL_x2_DEFINE(220); MUL_x2_DEFINE(221); MUL_x2_DEFINE(222); MUL_x2_DEFINE(223);
398MUL_x2_DEFINE(224); MUL_x2_DEFINE(225); MUL_x2_DEFINE(226); MUL_x2_DEFINE(227);
399MUL_x2_DEFINE(228); MUL_x2_DEFINE(229); MUL_x2_DEFINE(230); MUL_x2_DEFINE(231);
400MUL_x2_DEFINE(232); MUL_x2_DEFINE(233); MUL_x2_DEFINE(234); MUL_x2_DEFINE(235);
401MUL_x2_DEFINE(236); MUL_x2_DEFINE(237); MUL_x2_DEFINE(238); MUL_x2_DEFINE(239);
402MUL_x2_DEFINE(240); MUL_x2_DEFINE(241); MUL_x2_DEFINE(242); MUL_x2_DEFINE(243);
403MUL_x2_DEFINE(244); MUL_x2_DEFINE(245); MUL_x2_DEFINE(246); MUL_x2_DEFINE(247);
404MUL_x2_DEFINE(248); MUL_x2_DEFINE(249); MUL_x2_DEFINE(250); MUL_x2_DEFINE(251);
405MUL_x2_DEFINE(252); MUL_x2_DEFINE(253); MUL_x2_DEFINE(254); MUL_x2_DEFINE(255);
406
407
408
409typedef void (*mul_fn_ptr_t)(void);
410
411static const mul_fn_ptr_t __attribute__((aligned(256)))
412gf_x1_mul_fns[256] = {
413 mul_x1_0, mul_x1_1, mul_x1_2, mul_x1_3, mul_x1_4, mul_x1_5,
414 mul_x1_6, mul_x1_7, mul_x1_8, mul_x1_9, mul_x1_10, mul_x1_11,
415 mul_x1_12, mul_x1_13, mul_x1_14, mul_x1_15, mul_x1_16, mul_x1_17,
416 mul_x1_18, mul_x1_19, mul_x1_20, mul_x1_21, mul_x1_22, mul_x1_23,
417 mul_x1_24, mul_x1_25, mul_x1_26, mul_x1_27, mul_x1_28, mul_x1_29,
418 mul_x1_30, mul_x1_31, mul_x1_32, mul_x1_33, mul_x1_34, mul_x1_35,
419 mul_x1_36, mul_x1_37, mul_x1_38, mul_x1_39, mul_x1_40, mul_x1_41,
420 mul_x1_42, mul_x1_43, mul_x1_44, mul_x1_45, mul_x1_46, mul_x1_47,
421 mul_x1_48, mul_x1_49, mul_x1_50, mul_x1_51, mul_x1_52, mul_x1_53,
422 mul_x1_54, mul_x1_55, mul_x1_56, mul_x1_57, mul_x1_58, mul_x1_59,
423 mul_x1_60, mul_x1_61, mul_x1_62, mul_x1_63, mul_x1_64, mul_x1_65,
424 mul_x1_66, mul_x1_67, mul_x1_68, mul_x1_69, mul_x1_70, mul_x1_71,
425 mul_x1_72, mul_x1_73, mul_x1_74, mul_x1_75, mul_x1_76, mul_x1_77,
426 mul_x1_78, mul_x1_79, mul_x1_80, mul_x1_81, mul_x1_82, mul_x1_83,
427 mul_x1_84, mul_x1_85, mul_x1_86, mul_x1_87, mul_x1_88, mul_x1_89,
428 mul_x1_90, mul_x1_91, mul_x1_92, mul_x1_93, mul_x1_94, mul_x1_95,
429 mul_x1_96, mul_x1_97, mul_x1_98, mul_x1_99, mul_x1_100, mul_x1_101,
430 mul_x1_102, mul_x1_103, mul_x1_104, mul_x1_105, mul_x1_106, mul_x1_107,
431 mul_x1_108, mul_x1_109, mul_x1_110, mul_x1_111, mul_x1_112, mul_x1_113,
432 mul_x1_114, mul_x1_115, mul_x1_116, mul_x1_117, mul_x1_118, mul_x1_119,
433 mul_x1_120, mul_x1_121, mul_x1_122, mul_x1_123, mul_x1_124, mul_x1_125,
434 mul_x1_126, mul_x1_127, mul_x1_128, mul_x1_129, mul_x1_130, mul_x1_131,
435 mul_x1_132, mul_x1_133, mul_x1_134, mul_x1_135, mul_x1_136, mul_x1_137,
436 mul_x1_138, mul_x1_139, mul_x1_140, mul_x1_141, mul_x1_142, mul_x1_143,
437 mul_x1_144, mul_x1_145, mul_x1_146, mul_x1_147, mul_x1_148, mul_x1_149,
438 mul_x1_150, mul_x1_151, mul_x1_152, mul_x1_153, mul_x1_154, mul_x1_155,
439 mul_x1_156, mul_x1_157, mul_x1_158, mul_x1_159, mul_x1_160, mul_x1_161,
440 mul_x1_162, mul_x1_163, mul_x1_164, mul_x1_165, mul_x1_166, mul_x1_167,
441 mul_x1_168, mul_x1_169, mul_x1_170, mul_x1_171, mul_x1_172, mul_x1_173,
442 mul_x1_174, mul_x1_175, mul_x1_176, mul_x1_177, mul_x1_178, mul_x1_179,
443 mul_x1_180, mul_x1_181, mul_x1_182, mul_x1_183, mul_x1_184, mul_x1_185,
444 mul_x1_186, mul_x1_187, mul_x1_188, mul_x1_189, mul_x1_190, mul_x1_191,
445 mul_x1_192, mul_x1_193, mul_x1_194, mul_x1_195, mul_x1_196, mul_x1_197,
446 mul_x1_198, mul_x1_199, mul_x1_200, mul_x1_201, mul_x1_202, mul_x1_203,
447 mul_x1_204, mul_x1_205, mul_x1_206, mul_x1_207, mul_x1_208, mul_x1_209,
448 mul_x1_210, mul_x1_211, mul_x1_212, mul_x1_213, mul_x1_214, mul_x1_215,
449 mul_x1_216, mul_x1_217, mul_x1_218, mul_x1_219, mul_x1_220, mul_x1_221,
450 mul_x1_222, mul_x1_223, mul_x1_224, mul_x1_225, mul_x1_226, mul_x1_227,
451 mul_x1_228, mul_x1_229, mul_x1_230, mul_x1_231, mul_x1_232, mul_x1_233,
452 mul_x1_234, mul_x1_235, mul_x1_236, mul_x1_237, mul_x1_238, mul_x1_239,
453 mul_x1_240, mul_x1_241, mul_x1_242, mul_x1_243, mul_x1_244, mul_x1_245,
454 mul_x1_246, mul_x1_247, mul_x1_248, mul_x1_249, mul_x1_250, mul_x1_251,
455 mul_x1_252, mul_x1_253, mul_x1_254, mul_x1_255
456};
457
458static const mul_fn_ptr_t __attribute__((aligned(256)))
459gf_x2_mul_fns[256] = {
460 mul_x2_0, mul_x2_1, mul_x2_2, mul_x2_3, mul_x2_4, mul_x2_5,
461 mul_x2_6, mul_x2_7, mul_x2_8, mul_x2_9, mul_x2_10, mul_x2_11,
462 mul_x2_12, mul_x2_13, mul_x2_14, mul_x2_15, mul_x2_16, mul_x2_17,
463 mul_x2_18, mul_x2_19, mul_x2_20, mul_x2_21, mul_x2_22, mul_x2_23,
464 mul_x2_24, mul_x2_25, mul_x2_26, mul_x2_27, mul_x2_28, mul_x2_29,
465 mul_x2_30, mul_x2_31, mul_x2_32, mul_x2_33, mul_x2_34, mul_x2_35,
466 mul_x2_36, mul_x2_37, mul_x2_38, mul_x2_39, mul_x2_40, mul_x2_41,
467 mul_x2_42, mul_x2_43, mul_x2_44, mul_x2_45, mul_x2_46, mul_x2_47,
468 mul_x2_48, mul_x2_49, mul_x2_50, mul_x2_51, mul_x2_52, mul_x2_53,
469 mul_x2_54, mul_x2_55, mul_x2_56, mul_x2_57, mul_x2_58, mul_x2_59,
470 mul_x2_60, mul_x2_61, mul_x2_62, mul_x2_63, mul_x2_64, mul_x2_65,
471 mul_x2_66, mul_x2_67, mul_x2_68, mul_x2_69, mul_x2_70, mul_x2_71,
472 mul_x2_72, mul_x2_73, mul_x2_74, mul_x2_75, mul_x2_76, mul_x2_77,
473 mul_x2_78, mul_x2_79, mul_x2_80, mul_x2_81, mul_x2_82, mul_x2_83,
474 mul_x2_84, mul_x2_85, mul_x2_86, mul_x2_87, mul_x2_88, mul_x2_89,
475 mul_x2_90, mul_x2_91, mul_x2_92, mul_x2_93, mul_x2_94, mul_x2_95,
476 mul_x2_96, mul_x2_97, mul_x2_98, mul_x2_99, mul_x2_100, mul_x2_101,
477 mul_x2_102, mul_x2_103, mul_x2_104, mul_x2_105, mul_x2_106, mul_x2_107,
478 mul_x2_108, mul_x2_109, mul_x2_110, mul_x2_111, mul_x2_112, mul_x2_113,
479 mul_x2_114, mul_x2_115, mul_x2_116, mul_x2_117, mul_x2_118, mul_x2_119,
480 mul_x2_120, mul_x2_121, mul_x2_122, mul_x2_123, mul_x2_124, mul_x2_125,
481 mul_x2_126, mul_x2_127, mul_x2_128, mul_x2_129, mul_x2_130, mul_x2_131,
482 mul_x2_132, mul_x2_133, mul_x2_134, mul_x2_135, mul_x2_136, mul_x2_137,
483 mul_x2_138, mul_x2_139, mul_x2_140, mul_x2_141, mul_x2_142, mul_x2_143,
484 mul_x2_144, mul_x2_145, mul_x2_146, mul_x2_147, mul_x2_148, mul_x2_149,
485 mul_x2_150, mul_x2_151, mul_x2_152, mul_x2_153, mul_x2_154, mul_x2_155,
486 mul_x2_156, mul_x2_157, mul_x2_158, mul_x2_159, mul_x2_160, mul_x2_161,
487 mul_x2_162, mul_x2_163, mul_x2_164, mul_x2_165, mul_x2_166, mul_x2_167,
488 mul_x2_168, mul_x2_169, mul_x2_170, mul_x2_171, mul_x2_172, mul_x2_173,
489 mul_x2_174, mul_x2_175, mul_x2_176, mul_x2_177, mul_x2_178, mul_x2_179,
490 mul_x2_180, mul_x2_181, mul_x2_182, mul_x2_183, mul_x2_184, mul_x2_185,
491 mul_x2_186, mul_x2_187, mul_x2_188, mul_x2_189, mul_x2_190, mul_x2_191,
492 mul_x2_192, mul_x2_193, mul_x2_194, mul_x2_195, mul_x2_196, mul_x2_197,
493 mul_x2_198, mul_x2_199, mul_x2_200, mul_x2_201, mul_x2_202, mul_x2_203,
494 mul_x2_204, mul_x2_205, mul_x2_206, mul_x2_207, mul_x2_208, mul_x2_209,
495 mul_x2_210, mul_x2_211, mul_x2_212, mul_x2_213, mul_x2_214, mul_x2_215,
496 mul_x2_216, mul_x2_217, mul_x2_218, mul_x2_219, mul_x2_220, mul_x2_221,
497 mul_x2_222, mul_x2_223, mul_x2_224, mul_x2_225, mul_x2_226, mul_x2_227,
498 mul_x2_228, mul_x2_229, mul_x2_230, mul_x2_231, mul_x2_232, mul_x2_233,
499 mul_x2_234, mul_x2_235, mul_x2_236, mul_x2_237, mul_x2_238, mul_x2_239,
500 mul_x2_240, mul_x2_241, mul_x2_242, mul_x2_243, mul_x2_244, mul_x2_245,
501 mul_x2_246, mul_x2_247, mul_x2_248, mul_x2_249, mul_x2_250, mul_x2_251,
502 mul_x2_252, mul_x2_253, mul_x2_254, mul_x2_255
503};
504
505#define MUL(c, r...) \
506{ \
507 switch (REG_CNT(r)) { \
508 case 2: \
509 COPY(r, _mul_x2_in); \
510 gf_x2_mul_fns[c](); \
511 COPY(_mul_x2_acc, r); \
512 break; \
513 case 1: \
514 COPY(r, _mul_x1_in); \
515 gf_x1_mul_fns[c](); \
516 COPY(_mul_x1_acc, r); \
517 break; \
c5858ff9
MM
518 default: \
519 VERIFY(0); \
ae25d222
GN
520 } \
521}
522
523
524#define raidz_math_begin() kfpu_begin()
525#define raidz_math_end() kfpu_end()
526
cbf484f8
GN
527#define SYN_STRIDE 4
528
529#define ZERO_STRIDE 4
530#define ZERO_DEFINE() {}
531#define ZERO_D 0, 1, 2, 3
532
533#define COPY_STRIDE 4
534#define COPY_DEFINE() {}
535#define COPY_D 0, 1, 2, 3
536
537#define ADD_STRIDE 4
538#define ADD_DEFINE() {}
539#define ADD_D 0, 1, 2, 3
540
541#define MUL_STRIDE 2
65d71d42 542#define MUL_DEFINE() MUL2_SETUP()
cbf484f8
GN
543#define MUL_D 0, 1
544
ae25d222 545#define GEN_P_STRIDE 4
cbf484f8 546#define GEN_P_DEFINE() {}
ae25d222
GN
547#define GEN_P_P 0, 1, 2, 3
548
cbf484f8 549#define GEN_PQ_STRIDE 4
ae25d222 550#define GEN_PQ_DEFINE() {}
cbf484f8
GN
551#define GEN_PQ_D 0, 1, 2, 3
552#define GEN_PQ_C 4, 5, 6, 7
ae25d222 553
cbf484f8 554#define GEN_PQR_STRIDE 4
ae25d222 555#define GEN_PQR_DEFINE() {}
cbf484f8
GN
556#define GEN_PQR_D 0, 1, 2, 3
557#define GEN_PQR_C 4, 5, 6, 7
ae25d222 558
cbf484f8
GN
559#define SYN_Q_DEFINE() {}
560#define SYN_Q_D 0, 1, 2, 3
561#define SYN_Q_X 4, 5, 6, 7
ae25d222 562
cbf484f8
GN
563#define SYN_R_DEFINE() {}
564#define SYN_R_D 0, 1, 2, 3
565#define SYN_R_X 4, 5, 6, 7
ae25d222 566
cbf484f8
GN
567#define SYN_PQ_DEFINE() {}
568#define SYN_PQ_D 0, 1, 2, 3
569#define SYN_PQ_X 4, 5, 6, 7
ae25d222 570
ae25d222 571#define REC_PQ_STRIDE 2
65d71d42 572#define REC_PQ_DEFINE() MUL2_SETUP()
ae25d222
GN
573#define REC_PQ_X 0, 1
574#define REC_PQ_Y 2, 3
cbf484f8
GN
575#define REC_PQ_T 4, 5
576
577#define SYN_PR_DEFINE() {}
578#define SYN_PR_D 0, 1, 2, 3
579#define SYN_PR_X 4, 5, 6, 7
ae25d222 580
ae25d222 581#define REC_PR_STRIDE 2
65d71d42 582#define REC_PR_DEFINE() MUL2_SETUP()
ae25d222
GN
583#define REC_PR_X 0, 1
584#define REC_PR_Y 2, 3
cbf484f8
GN
585#define REC_PR_T 4, 5
586
587#define SYN_QR_DEFINE() {}
588#define SYN_QR_D 0, 1, 2, 3
589#define SYN_QR_X 4, 5, 6, 7
ae25d222 590
ae25d222 591#define REC_QR_STRIDE 2
65d71d42 592#define REC_QR_DEFINE() MUL2_SETUP()
ae25d222
GN
593#define REC_QR_X 0, 1
594#define REC_QR_Y 2, 3
cbf484f8
GN
595#define REC_QR_T 4, 5
596
597#define SYN_PQR_DEFINE() {}
598#define SYN_PQR_D 0, 1, 2, 3
599#define SYN_PQR_X 4, 5, 6, 7
ae25d222 600
ae25d222 601#define REC_PQR_STRIDE 1
65d71d42 602#define REC_PQR_DEFINE() MUL2_SETUP()
ae25d222
GN
603#define REC_PQR_X 0
604#define REC_PQR_Y 1
605#define REC_PQR_Z 2
cbf484f8
GN
606#define REC_PQR_XS 3
607#define REC_PQR_YS 4
ae25d222
GN
608
609
610#include <sys/vdev_raidz_impl.h>
611#include "vdev_raidz_math_impl.h"
612
613DEFINE_GEN_METHODS(sse2);
614DEFINE_REC_METHODS(sse2);
615
616static boolean_t
617raidz_will_sse2_work(void)
618{
e5db3134 619 return (kfpu_allowed() && zfs_sse_available() && zfs_sse2_available());
ae25d222
GN
620}
621
622const raidz_impl_ops_t vdev_raidz_sse2_impl = {
623 .init = NULL,
624 .fini = NULL,
625 .gen = RAIDZ_GEN_METHODS(sse2),
626 .rec = RAIDZ_REC_METHODS(sse2),
627 .is_supported = &raidz_will_sse2_work,
628 .name = "sse2"
629};
630
631#endif /* defined(__x86_64) && defined(HAVE_SSE2) */