|
4148
|
1 /*
|
|
|
2 * This code was taken from http://www.mpg123.org
|
|
|
3 * See ChangeLog of mpg123-0.59s-pre.1 for detail
|
|
|
4 * Applied to mplayer by Nick Kurshev <nickols_k@mail.ru>
|
|
|
5 * Partial 3dnowex-DSP! optimization by Nick Kurshev
|
|
|
6 *
|
|
|
7 * TODO: optimize scalar 3dnow! code
|
|
|
8 * Warning: Phases 7 & 8 are not tested
|
|
|
9 */
|
|
|
10 #define real float /* ugly - but only way */
|
|
|
11
|
|
5291
|
12 #include "../mangle.h"
|
|
|
13
|
|
4148
|
14 static unsigned long long int __attribute__((aligned(8))) x_plus_minus_3dnow = 0x8000000000000000ULL;
|
|
|
15 static float plus_1f = 1.0;
|
|
|
16
|
|
|
17 void __attribute__ (( __stdcall__ )) dct64_MMX_3dnowex(real *a,real *b,real *c)
|
|
|
18 {
|
|
|
19 char tmp[256];
|
|
|
20 __asm __volatile(
|
|
|
21 " movl %2,%%eax\n\t"
|
|
|
22
|
|
|
23 " leal 128+%3,%%edx\n\t"
|
|
|
24 " movl %0,%%esi\n\t"
|
|
|
25 " movl %1,%%edi\n\t"
|
|
5291
|
26 " movl $"MANGLE(costab_mmx)",%%ebx\n\t"
|
|
4148
|
27 " leal %3,%%ecx\n\t"
|
|
|
28
|
|
|
29 /* Phase 1*/
|
|
|
30 " movq (%%eax), %%mm0\n\t"
|
|
|
31 " movq 8(%%eax), %%mm4\n\t"
|
|
|
32 " movq %%mm0, %%mm3\n\t"
|
|
|
33 " movq %%mm4, %%mm7\n\t"
|
|
|
34 " movq 120(%%eax), %%mm1\n\t"
|
|
|
35 " movq 112(%%eax), %%mm5\n\t"
|
|
|
36 " pswapd %%mm1, %%mm1\n\t"
|
|
|
37 " pswapd %%mm5, %%mm5\n\t"
|
|
|
38 " pfadd %%mm1, %%mm0\n\t"
|
|
|
39 " pfadd %%mm5, %%mm4\n\t"
|
|
|
40 " movq %%mm0, (%%edx)\n\t"
|
|
|
41 " movq %%mm4, 8(%%edx)\n\t"
|
|
|
42 " pfsub %%mm1, %%mm3\n\t"
|
|
|
43 " pfsub %%mm5, %%mm7\n\t"
|
|
|
44 " pfmul (%%ebx), %%mm3\n\t"
|
|
|
45 " pfmul 8(%%ebx), %%mm7\n\t"
|
|
|
46 " pswapd %%mm3, %%mm3\n\t"
|
|
|
47 " pswapd %%mm7, %%mm7\n\t"
|
|
|
48 " movq %%mm3, 120(%%edx)\n\t"
|
|
|
49 " movq %%mm7, 112(%%edx)\n\t"
|
|
|
50
|
|
|
51 " movq 16(%%eax), %%mm0\n\t"
|
|
|
52 " movq 24(%%eax), %%mm4\n\t"
|
|
|
53 " movq %%mm0, %%mm3\n\t"
|
|
|
54 " movq %%mm4, %%mm7\n\t"
|
|
|
55 " movq 104(%%eax), %%mm1\n\t"
|
|
|
56 " movq 96(%%eax), %%mm5\n\t"
|
|
|
57 " pswapd %%mm1, %%mm1\n\t"
|
|
|
58 " pswapd %%mm5, %%mm5\n\t"
|
|
|
59 " pfadd %%mm1, %%mm0\n\t"
|
|
|
60 " pfadd %%mm5, %%mm4\n\t"
|
|
|
61 " movq %%mm0, 16(%%edx)\n\t"
|
|
|
62 " movq %%mm4, 24(%%edx)\n\t"
|
|
|
63 " pfsub %%mm1, %%mm3\n\t"
|
|
|
64 " pfsub %%mm5, %%mm7\n\t"
|
|
|
65 " pfmul 16(%%ebx), %%mm3\n\t"
|
|
|
66 " pfmul 24(%%ebx), %%mm7\n\t"
|
|
|
67 " pswapd %%mm3, %%mm3\n\t"
|
|
|
68 " pswapd %%mm7, %%mm7\n\t"
|
|
|
69 " movq %%mm3, 104(%%edx)\n\t"
|
|
|
70 " movq %%mm7, 96(%%edx)\n\t"
|
|
|
71
|
|
|
72 " movq 32(%%eax), %%mm0\n\t"
|
|
|
73 " movq 40(%%eax), %%mm4\n\t"
|
|
|
74 " movq %%mm0, %%mm3\n\t"
|
|
|
75 " movq %%mm4, %%mm7\n\t"
|
|
|
76 " movq 88(%%eax), %%mm1\n\t"
|
|
|
77 " movq 80(%%eax), %%mm5\n\t"
|
|
|
78 " pswapd %%mm1, %%mm1\n\t"
|
|
|
79 " pswapd %%mm5, %%mm5\n\t"
|
|
|
80 " pfadd %%mm1, %%mm0\n\t"
|
|
|
81 " pfadd %%mm5, %%mm4\n\t"
|
|
|
82 " movq %%mm0, 32(%%edx)\n\t"
|
|
|
83 " movq %%mm4, 40(%%edx)\n\t"
|
|
|
84 " pfsub %%mm1, %%mm3\n\t"
|
|
|
85 " pfsub %%mm5, %%mm7\n\t"
|
|
|
86 " pfmul 32(%%ebx), %%mm3\n\t"
|
|
|
87 " pfmul 40(%%ebx), %%mm7\n\t"
|
|
|
88 " pswapd %%mm3, %%mm3\n\t"
|
|
|
89 " pswapd %%mm7, %%mm7\n\t"
|
|
|
90 " movq %%mm3, 88(%%edx)\n\t"
|
|
|
91 " movq %%mm7, 80(%%edx)\n\t"
|
|
|
92
|
|
|
93 " movq 48(%%eax), %%mm0\n\t"
|
|
|
94 " movq 56(%%eax), %%mm4\n\t"
|
|
|
95 " movq %%mm0, %%mm3\n\t"
|
|
|
96 " movq %%mm4, %%mm7\n\t"
|
|
|
97 " movq 72(%%eax), %%mm1\n\t"
|
|
|
98 " movq 64(%%eax), %%mm5\n\t"
|
|
|
99 " pswapd %%mm1, %%mm1\n\t"
|
|
|
100 " pswapd %%mm5, %%mm5\n\t"
|
|
|
101 " pfadd %%mm1, %%mm0\n\t"
|
|
|
102 " pfadd %%mm5, %%mm4\n\t"
|
|
|
103 " movq %%mm0, 48(%%edx)\n\t"
|
|
|
104 " movq %%mm4, 56(%%edx)\n\t"
|
|
|
105 " pfsub %%mm1, %%mm3\n\t"
|
|
|
106 " pfsub %%mm5, %%mm7\n\t"
|
|
|
107 " pfmul 48(%%ebx), %%mm3\n\t"
|
|
|
108 " pfmul 56(%%ebx), %%mm7\n\t"
|
|
|
109 " pswapd %%mm3, %%mm3\n\t"
|
|
|
110 " pswapd %%mm7, %%mm7\n\t"
|
|
|
111 " movq %%mm3, 72(%%edx)\n\t"
|
|
|
112 " movq %%mm7, 64(%%edx)\n\t"
|
|
|
113
|
|
|
114 /* Phase 2*/
|
|
|
115
|
|
|
116 " movq (%%edx), %%mm0\n\t"
|
|
|
117 " movq 8(%%edx), %%mm4\n\t"
|
|
|
118 " movq %%mm0, %%mm3\n\t"
|
|
|
119 " movq %%mm4, %%mm7\n\t"
|
|
|
120 " movq 56(%%edx), %%mm1\n\t"
|
|
|
121 " movq 48(%%edx), %%mm5\n\t"
|
|
|
122 " pswapd %%mm1, %%mm1\n\t"
|
|
|
123 " pswapd %%mm5, %%mm5\n\t"
|
|
|
124 " pfadd %%mm1, %%mm0\n\t"
|
|
|
125 " pfadd %%mm5, %%mm4\n\t"
|
|
|
126 " movq %%mm0, (%%ecx)\n\t"
|
|
|
127 " movq %%mm4, 8(%%ecx)\n\t"
|
|
|
128 " pfsub %%mm1, %%mm3\n\t"
|
|
|
129 " pfsub %%mm5, %%mm7\n\t"
|
|
|
130 " pfmul 64(%%ebx), %%mm3\n\t"
|
|
|
131 " pfmul 72(%%ebx), %%mm7\n\t"
|
|
|
132 " pswapd %%mm3, %%mm3\n\t"
|
|
|
133 " pswapd %%mm7, %%mm7\n\t"
|
|
|
134 " movq %%mm3, 56(%%ecx)\n\t"
|
|
|
135 " movq %%mm7, 48(%%ecx)\n\t"
|
|
|
136
|
|
|
137 " movq 16(%%edx), %%mm0\n\t"
|
|
|
138 " movq 24(%%edx), %%mm4\n\t"
|
|
|
139 " movq %%mm0, %%mm3\n\t"
|
|
|
140 " movq %%mm4, %%mm7\n\t"
|
|
|
141 " movq 40(%%edx), %%mm1\n\t"
|
|
|
142 " movq 32(%%edx), %%mm5\n\t"
|
|
|
143 " pswapd %%mm1, %%mm1\n\t"
|
|
|
144 " pswapd %%mm5, %%mm5\n\t"
|
|
|
145 " pfadd %%mm1, %%mm0\n\t"
|
|
|
146 " pfadd %%mm5, %%mm4\n\t"
|
|
|
147 " movq %%mm0, 16(%%ecx)\n\t"
|
|
|
148 " movq %%mm4, 24(%%ecx)\n\t"
|
|
|
149 " pfsub %%mm1, %%mm3\n\t"
|
|
|
150 " pfsub %%mm5, %%mm7\n\t"
|
|
|
151 " pfmul 80(%%ebx), %%mm3\n\t"
|
|
|
152 " pfmul 88(%%ebx), %%mm7\n\t"
|
|
|
153 " pswapd %%mm3, %%mm3\n\t"
|
|
|
154 " pswapd %%mm7, %%mm7\n\t"
|
|
|
155 " movq %%mm3, 40(%%ecx)\n\t"
|
|
|
156 " movq %%mm7, 32(%%ecx)\n\t"
|
|
|
157
|
|
|
158 /* Phase 3*/
|
|
|
159
|
|
|
160 " movq 64(%%edx), %%mm0\n\t"
|
|
|
161 " movq 72(%%edx), %%mm4\n\t"
|
|
|
162 " movq %%mm0, %%mm3\n\t"
|
|
|
163 " movq %%mm4, %%mm7\n\t"
|
|
|
164 " movq 120(%%edx), %%mm1\n\t"
|
|
|
165 " movq 112(%%edx), %%mm5\n\t"
|
|
|
166 " pswapd %%mm1, %%mm1\n\t"
|
|
|
167 " pswapd %%mm5, %%mm5\n\t"
|
|
|
168 " pfadd %%mm1, %%mm0\n\t"
|
|
|
169 " pfadd %%mm5, %%mm4\n\t"
|
|
|
170 " movq %%mm0, 64(%%ecx)\n\t"
|
|
|
171 " movq %%mm4, 72(%%ecx)\n\t"
|
|
|
172 " pfsubr %%mm1, %%mm3\n\t"
|
|
|
173 " pfsubr %%mm5, %%mm7\n\t"
|
|
|
174 " pfmul 64(%%ebx), %%mm3\n\t"
|
|
|
175 " pfmul 72(%%ebx), %%mm7\n\t"
|
|
|
176 " pswapd %%mm3, %%mm3\n\t"
|
|
|
177 " pswapd %%mm7, %%mm7\n\t"
|
|
|
178 " movq %%mm3, 120(%%ecx)\n\t"
|
|
|
179 " movq %%mm7, 112(%%ecx)\n\t"
|
|
|
180
|
|
|
181 " movq 80(%%edx), %%mm0\n\t"
|
|
|
182 " movq 88(%%edx), %%mm4\n\t"
|
|
|
183 " movq %%mm0, %%mm3\n\t"
|
|
|
184 " movq %%mm4, %%mm7\n\t"
|
|
|
185 " movq 104(%%edx), %%mm1\n\t"
|
|
|
186 " movq 96(%%edx), %%mm5\n\t"
|
|
|
187 " pswapd %%mm1, %%mm1\n\t"
|
|
|
188 " pswapd %%mm5, %%mm5\n\t"
|
|
|
189 " pfadd %%mm1, %%mm0\n\t"
|
|
|
190 " pfadd %%mm5, %%mm4\n\t"
|
|
|
191 " movq %%mm0, 80(%%ecx)\n\t"
|
|
|
192 " movq %%mm4, 88(%%ecx)\n\t"
|
|
|
193 " pfsubr %%mm1, %%mm3\n\t"
|
|
|
194 " pfsubr %%mm5, %%mm7\n\t"
|
|
|
195 " pfmul 80(%%ebx), %%mm3\n\t"
|
|
|
196 " pfmul 88(%%ebx), %%mm7\n\t"
|
|
|
197 " pswapd %%mm3, %%mm3\n\t"
|
|
|
198 " pswapd %%mm7, %%mm7\n\t"
|
|
|
199 " movq %%mm3, 104(%%ecx)\n\t"
|
|
|
200 " movq %%mm7, 96(%%ecx)\n\t"
|
|
|
201
|
|
|
202 /* Phase 4*/
|
|
|
203
|
|
|
204 " movq 96(%%ebx), %%mm2\n\t"
|
|
|
205 " movq 104(%%ebx), %%mm6\n\t"
|
|
|
206
|
|
|
207 " movq (%%ecx), %%mm0\n\t"
|
|
|
208 " movq 8(%%ecx), %%mm4\n\t"
|
|
|
209 " movq %%mm0, %%mm3\n\t"
|
|
|
210 " movq %%mm4, %%mm7\n\t"
|
|
|
211 " movq 24(%%ecx), %%mm1\n\t"
|
|
|
212 " movq 16(%%ecx), %%mm5\n\t"
|
|
|
213 " pswapd %%mm1, %%mm1\n\t"
|
|
|
214 " pswapd %%mm5, %%mm5\n\t"
|
|
|
215 " pfadd %%mm1, %%mm0\n\t"
|
|
|
216 " pfadd %%mm5, %%mm4\n\t"
|
|
|
217 " movq %%mm0, (%%edx)\n\t"
|
|
|
218 " movq %%mm4, 8(%%edx)\n\t"
|
|
|
219 " pfsub %%mm1, %%mm3\n\t"
|
|
|
220 " pfsub %%mm5, %%mm7\n\t"
|
|
|
221 " pfmul %%mm2, %%mm3\n\t"
|
|
|
222 " pfmul %%mm6, %%mm7\n\t"
|
|
|
223 " pswapd %%mm3, %%mm3\n\t"
|
|
|
224 " pswapd %%mm7, %%mm7\n\t"
|
|
|
225 " movq %%mm3, 24(%%edx)\n\t"
|
|
|
226 " movq %%mm7, 16(%%edx)\n\t"
|
|
|
227
|
|
|
228 " movq 32(%%ecx), %%mm0\n\t"
|
|
|
229 " movq 40(%%ecx), %%mm4\n\t"
|
|
|
230 " movq %%mm0, %%mm3\n\t"
|
|
|
231 " movq %%mm4, %%mm7\n\t"
|
|
|
232 " movq 56(%%ecx), %%mm1\n\t"
|
|
|
233 " movq 48(%%ecx), %%mm5\n\t"
|
|
|
234 " pswapd %%mm1, %%mm1\n\t"
|
|
|
235 " pswapd %%mm5, %%mm5\n\t"
|
|
|
236 " pfadd %%mm1, %%mm0\n\t"
|
|
|
237 " pfadd %%mm5, %%mm4\n\t"
|
|
|
238 " movq %%mm0, 32(%%edx)\n\t"
|
|
|
239 " movq %%mm4, 40(%%edx)\n\t"
|
|
|
240 " pfsubr %%mm1, %%mm3\n\t"
|
|
|
241 " pfsubr %%mm5, %%mm7\n\t"
|
|
|
242 " pfmul %%mm2, %%mm3\n\t"
|
|
|
243 " pfmul %%mm6, %%mm7\n\t"
|
|
|
244 " pswapd %%mm3, %%mm3\n\t"
|
|
|
245 " pswapd %%mm7, %%mm7\n\t"
|
|
|
246 " movq %%mm3, 56(%%edx)\n\t"
|
|
|
247 " movq %%mm7, 48(%%edx)\n\t"
|
|
|
248
|
|
|
249 " movq 64(%%ecx), %%mm0\n\t"
|
|
|
250 " movq 72(%%ecx), %%mm4\n\t"
|
|
|
251 " movq %%mm0, %%mm3\n\t"
|
|
|
252 " movq %%mm4, %%mm7\n\t"
|
|
|
253 " movq 88(%%ecx), %%mm1\n\t"
|
|
|
254 " movq 80(%%ecx), %%mm5\n\t"
|
|
|
255 " pswapd %%mm1, %%mm1\n\t"
|
|
|
256 " pswapd %%mm5, %%mm5\n\t"
|
|
|
257 " pfadd %%mm1, %%mm0\n\t"
|
|
|
258 " pfadd %%mm5, %%mm4\n\t"
|
|
|
259 " movq %%mm0, 64(%%edx)\n\t"
|
|
|
260 " movq %%mm4, 72(%%edx)\n\t"
|
|
|
261 " pfsub %%mm1, %%mm3\n\t"
|
|
|
262 " pfsub %%mm5, %%mm7\n\t"
|
|
|
263 " pfmul %%mm2, %%mm3\n\t"
|
|
|
264 " pfmul %%mm6, %%mm7\n\t"
|
|
|
265 " pswapd %%mm3, %%mm3\n\t"
|
|
|
266 " pswapd %%mm7, %%mm7\n\t"
|
|
|
267 " movq %%mm3, 88(%%edx)\n\t"
|
|
|
268 " movq %%mm7, 80(%%edx)\n\t"
|
|
|
269
|
|
|
270 " movq 96(%%ecx), %%mm0\n\t"
|
|
|
271 " movq 104(%%ecx), %%mm4\n\t"
|
|
|
272 " movq %%mm0, %%mm3\n\t"
|
|
|
273 " movq %%mm4, %%mm7\n\t"
|
|
|
274 " movq 120(%%ecx), %%mm1\n\t"
|
|
|
275 " movq 112(%%ecx), %%mm5\n\t"
|
|
|
276 " pswapd %%mm1, %%mm1\n\t"
|
|
|
277 " pswapd %%mm5, %%mm5\n\t"
|
|
|
278 " pfadd %%mm1, %%mm0\n\t"
|
|
|
279 " pfadd %%mm5, %%mm4\n\t"
|
|
|
280 " movq %%mm0, 96(%%edx)\n\t"
|
|
|
281 " movq %%mm4, 104(%%edx)\n\t"
|
|
|
282 " pfsubr %%mm1, %%mm3\n\t"
|
|
|
283 " pfsubr %%mm5, %%mm7\n\t"
|
|
|
284 " pfmul %%mm2, %%mm3\n\t"
|
|
|
285 " pfmul %%mm6, %%mm7\n\t"
|
|
|
286 " pswapd %%mm3, %%mm3\n\t"
|
|
|
287 " pswapd %%mm7, %%mm7\n\t"
|
|
|
288 " movq %%mm3, 120(%%edx)\n\t"
|
|
|
289 " movq %%mm7, 112(%%edx)\n\t"
|
|
|
290
|
|
|
291 /* Phase 5 */
|
|
|
292
|
|
|
293 " movq 112(%%ebx), %%mm2\n\t"
|
|
|
294
|
|
|
295 " movq (%%edx), %%mm0\n\t"
|
|
|
296 " movq 16(%%edx), %%mm4\n\t"
|
|
|
297 " movq %%mm0, %%mm3\n\t"
|
|
|
298 " movq %%mm4, %%mm7\n\t"
|
|
|
299 " movq 8(%%edx), %%mm1\n\t"
|
|
|
300 " movq 24(%%edx), %%mm5\n\t"
|
|
|
301 " pswapd %%mm1, %%mm1\n\t"
|
|
|
302 " pswapd %%mm5, %%mm5\n\t"
|
|
|
303 " pfadd %%mm1, %%mm0\n\t"
|
|
|
304 " pfadd %%mm5, %%mm4\n\t"
|
|
|
305 " movq %%mm0, (%%ecx)\n\t"
|
|
|
306 " movq %%mm4, 16(%%ecx)\n\t"
|
|
|
307 " pfsub %%mm1, %%mm3\n\t"
|
|
|
308 " pfsubr %%mm5, %%mm7\n\t"
|
|
|
309 " pfmul %%mm2, %%mm3\n\t"
|
|
|
310 " pfmul %%mm2, %%mm7\n\t"
|
|
|
311 " pswapd %%mm3, %%mm3\n\t"
|
|
|
312 " pswapd %%mm7, %%mm7\n\t"
|
|
|
313 " movq %%mm3, 8(%%ecx)\n\t"
|
|
|
314 " movq %%mm7, 24(%%ecx)\n\t"
|
|
|
315
|
|
|
316 " movq 32(%%edx), %%mm0\n\t"
|
|
|
317 " movq 48(%%edx), %%mm4\n\t"
|
|
|
318 " movq %%mm0, %%mm3\n\t"
|
|
|
319 " movq %%mm4, %%mm7\n\t"
|
|
|
320 " movq 40(%%edx), %%mm1\n\t"
|
|
|
321 " movq 56(%%edx), %%mm5\n\t"
|
|
|
322 " pswapd %%mm1, %%mm1\n\t"
|
|
|
323 " pswapd %%mm5, %%mm5\n\t"
|
|
|
324 " pfadd %%mm1, %%mm0\n\t"
|
|
|
325 " pfadd %%mm5, %%mm4\n\t"
|
|
|
326 " movq %%mm0, 32(%%ecx)\n\t"
|
|
|
327 " movq %%mm4, 48(%%ecx)\n\t"
|
|
|
328 " pfsub %%mm1, %%mm3\n\t"
|
|
|
329 " pfsubr %%mm5, %%mm7\n\t"
|
|
|
330 " pfmul %%mm2, %%mm3\n\t"
|
|
|
331 " pfmul %%mm2, %%mm7\n\t"
|
|
|
332 " pswapd %%mm3, %%mm3\n\t"
|
|
|
333 " pswapd %%mm7, %%mm7\n\t"
|
|
|
334 " movq %%mm3, 40(%%ecx)\n\t"
|
|
|
335 " movq %%mm7, 56(%%ecx)\n\t"
|
|
|
336
|
|
|
337 " movq 64(%%edx), %%mm0\n\t"
|
|
|
338 " movq 80(%%edx), %%mm4\n\t"
|
|
|
339 " movq %%mm0, %%mm3\n\t"
|
|
|
340 " movq %%mm4, %%mm7\n\t"
|
|
|
341 " movq 72(%%edx), %%mm1\n\t"
|
|
|
342 " movq 88(%%edx), %%mm5\n\t"
|
|
|
343 " pswapd %%mm1, %%mm1\n\t"
|
|
|
344 " pswapd %%mm5, %%mm5\n\t"
|
|
|
345 " pfadd %%mm1, %%mm0\n\t"
|
|
|
346 " pfadd %%mm5, %%mm4\n\t"
|
|
|
347 " movq %%mm0, 64(%%ecx)\n\t"
|
|
|
348 " movq %%mm4, 80(%%ecx)\n\t"
|
|
|
349 " pfsub %%mm1, %%mm3\n\t"
|
|
|
350 " pfsubr %%mm5, %%mm7\n\t"
|
|
|
351 " pfmul %%mm2, %%mm3\n\t"
|
|
|
352 " pfmul %%mm2, %%mm7\n\t"
|
|
|
353 " pswapd %%mm3, %%mm3\n\t"
|
|
|
354 " pswapd %%mm7, %%mm7\n\t"
|
|
|
355 " movq %%mm3, 72(%%ecx)\n\t"
|
|
|
356 " movq %%mm7, 88(%%ecx)\n\t"
|
|
|
357
|
|
|
358 " movq 96(%%edx), %%mm0\n\t"
|
|
|
359 " movq 112(%%edx), %%mm4\n\t"
|
|
|
360 " movq %%mm0, %%mm3\n\t"
|
|
|
361 " movq %%mm4, %%mm7\n\t"
|
|
|
362 " movq 104(%%edx), %%mm1\n\t"
|
|
|
363 " movq 120(%%edx), %%mm5\n\t"
|
|
|
364 " pswapd %%mm1, %%mm1\n\t"
|
|
|
365 " pswapd %%mm5, %%mm5\n\t"
|
|
|
366 " pfadd %%mm1, %%mm0\n\t"
|
|
|
367 " pfadd %%mm5, %%mm4\n\t"
|
|
|
368 " movq %%mm0, 96(%%ecx)\n\t"
|
|
|
369 " movq %%mm4, 112(%%ecx)\n\t"
|
|
|
370 " pfsub %%mm1, %%mm3\n\t"
|
|
|
371 " pfsubr %%mm5, %%mm7\n\t"
|
|
|
372 " pfmul %%mm2, %%mm3\n\t"
|
|
|
373 " pfmul %%mm2, %%mm7\n\t"
|
|
|
374 " pswapd %%mm3, %%mm3\n\t"
|
|
|
375 " pswapd %%mm7, %%mm7\n\t"
|
|
|
376 " movq %%mm3, 104(%%ecx)\n\t"
|
|
|
377 " movq %%mm7, 120(%%ecx)\n\t"
|
|
|
378
|
|
|
379
|
|
|
380 /* Phase 6. This is the end of easy road. */
|
|
|
381 /* Code below is coded in scalar mode. Should be optimized */
|
|
|
382
|
|
5291
|
383 " movd "MANGLE(plus_1f)", %%mm6\n\t"
|
|
4148
|
384 " punpckldq 120(%%ebx), %%mm6\n\t" /* mm6 = 1.0 | 120(%%ebx)*/
|
|
|
385 " movq x_plus_minus_3dnow, %%mm7\n\t" /* mm7 = +1 | -1 */
|
|
|
386
|
|
|
387 " movq 32(%%ecx), %%mm0\n\t"
|
|
|
388 " movq 64(%%ecx), %%mm2\n\t"
|
|
|
389 " movq %%mm0, %%mm1\n\t"
|
|
|
390 " movq %%mm2, %%mm3\n\t"
|
|
|
391 " pxor %%mm7, %%mm1\n\t"
|
|
|
392 " pxor %%mm7, %%mm3\n\t"
|
|
|
393 " pfacc %%mm1, %%mm0\n\t"
|
|
|
394 " pfacc %%mm3, %%mm2\n\t"
|
|
|
395 " pfmul %%mm6, %%mm0\n\t"
|
|
|
396 " pfmul %%mm6, %%mm2\n\t"
|
|
|
397 " movq %%mm0, 32(%%edx)\n\t"
|
|
|
398 " movq %%mm2, 64(%%edx)\n\t"
|
|
|
399
|
|
|
400 " movd 44(%%ecx), %%mm0\n\t"
|
|
|
401 " movd 40(%%ecx), %%mm2\n\t"
|
|
|
402 " movd 120(%%ebx), %%mm3\n\t"
|
|
|
403 " punpckldq 76(%%ecx), %%mm0\n\t"
|
|
|
404 " punpckldq 72(%%ecx), %%mm2\n\t"
|
|
|
405 " punpckldq %%mm3, %%mm3\n\t"
|
|
|
406 " movq %%mm0, %%mm4\n\t"
|
|
|
407 " movq %%mm2, %%mm5\n\t"
|
|
|
408 " pfsub %%mm2, %%mm0\n\t"
|
|
|
409 " pfmul %%mm3, %%mm0\n\t"
|
|
|
410 " movq %%mm0, %%mm1\n\t"
|
|
|
411 " pfadd %%mm5, %%mm0\n\t"
|
|
|
412 " pfadd %%mm4, %%mm0\n\t"
|
|
|
413 " movq %%mm0, %%mm2\n\t"
|
|
|
414 " punpckldq %%mm1, %%mm0\n\t"
|
|
|
415 " punpckhdq %%mm1, %%mm2\n\t"
|
|
|
416 " movq %%mm0, 40(%%edx)\n\t"
|
|
|
417 " movq %%mm2, 72(%%edx)\n\t"
|
|
|
418
|
|
|
419 " movd 48(%%ecx), %%mm3\n\t"
|
|
|
420 " movd 60(%%ecx), %%mm2\n\t"
|
|
|
421 " pfsub 52(%%ecx), %%mm3\n\t"
|
|
|
422 " pfsub 56(%%ecx), %%mm2\n\t"
|
|
|
423 " pfmul 120(%%ebx), %%mm3\n\t"
|
|
|
424 " pfmul 120(%%ebx), %%mm2\n\t"
|
|
|
425 " movq %%mm2, %%mm1\n\t"
|
|
|
426
|
|
|
427 " pfadd 56(%%ecx), %%mm1\n\t"
|
|
|
428 " pfadd 60(%%ecx), %%mm1\n\t"
|
|
|
429 " movq %%mm1, %%mm0\n\t"
|
|
|
430
|
|
|
431 " pfadd 48(%%ecx), %%mm0\n\t"
|
|
|
432 " pfadd 52(%%ecx), %%mm0\n\t"
|
|
|
433 " pfadd %%mm3, %%mm1\n\t"
|
|
|
434 " punpckldq %%mm2, %%mm1\n\t"
|
|
|
435 " pfadd %%mm3, %%mm2\n\t"
|
|
|
436 " punpckldq %%mm2, %%mm0\n\t"
|
|
|
437 " movq %%mm1, 56(%%edx)\n\t"
|
|
|
438 " movq %%mm0, 48(%%edx)\n\t"
|
|
|
439
|
|
|
440 /*---*/
|
|
|
441
|
|
|
442 " movd 92(%%ecx), %%mm1\n\t"
|
|
|
443 " pfsub 88(%%ecx), %%mm1\n\t"
|
|
|
444 " pfmul 120(%%ebx), %%mm1\n\t"
|
|
|
445 " movd %%mm1, 92(%%edx)\n\t"
|
|
|
446 " pfadd 92(%%ecx), %%mm1\n\t"
|
|
|
447 " pfadd 88(%%ecx), %%mm1\n\t"
|
|
|
448 " movq %%mm1, %%mm0\n\t"
|
|
|
449
|
|
|
450 " pfadd 80(%%ecx), %%mm0\n\t"
|
|
|
451 " pfadd 84(%%ecx), %%mm0\n\t"
|
|
|
452 " movd %%mm0, 80(%%edx)\n\t"
|
|
|
453
|
|
|
454 " movd 80(%%ecx), %%mm0\n\t"
|
|
|
455 " pfsub 84(%%ecx), %%mm0\n\t"
|
|
|
456 " pfmul 120(%%ebx), %%mm0\n\t"
|
|
|
457 " pfadd %%mm0, %%mm1\n\t"
|
|
|
458 " pfadd 92(%%edx), %%mm0\n\t"
|
|
|
459 " punpckldq %%mm1, %%mm0\n\t"
|
|
|
460 " movq %%mm0, 84(%%edx)\n\t"
|
|
|
461
|
|
|
462 " movq 96(%%ecx), %%mm0\n\t"
|
|
|
463 " movq %%mm0, %%mm1\n\t"
|
|
|
464 " pxor %%mm7, %%mm1\n\t"
|
|
|
465 " pfacc %%mm1, %%mm0\n\t"
|
|
|
466 " pfmul %%mm6, %%mm0\n\t"
|
|
|
467 " movq %%mm0, 96(%%edx)\n\t"
|
|
|
468
|
|
|
469 " movd 108(%%ecx), %%mm0\n\t"
|
|
|
470 " pfsub 104(%%ecx), %%mm0\n\t"
|
|
|
471 " pfmul 120(%%ebx), %%mm0\n\t"
|
|
|
472 " movd %%mm0, 108(%%edx)\n\t"
|
|
|
473 " pfadd 104(%%ecx), %%mm0\n\t"
|
|
|
474 " pfadd 108(%%ecx), %%mm0\n\t"
|
|
|
475 " movd %%mm0, 104(%%edx)\n\t"
|
|
|
476
|
|
|
477 " movd 124(%%ecx), %%mm1\n\t"
|
|
|
478 " pfsub 120(%%ecx), %%mm1\n\t"
|
|
|
479 " pfmul 120(%%ebx), %%mm1\n\t"
|
|
|
480 " movd %%mm1, 124(%%edx)\n\t"
|
|
|
481 " pfadd 120(%%ecx), %%mm1\n\t"
|
|
|
482 " pfadd 124(%%ecx), %%mm1\n\t"
|
|
|
483 " movq %%mm1, %%mm0\n\t"
|
|
|
484
|
|
|
485 " pfadd 112(%%ecx), %%mm0\n\t"
|
|
|
486 " pfadd 116(%%ecx), %%mm0\n\t"
|
|
|
487 " movd %%mm0, 112(%%edx)\n\t"
|
|
|
488
|
|
|
489 " movd 112(%%ecx), %%mm0\n\t"
|
|
|
490 " pfsub 116(%%ecx), %%mm0\n\t"
|
|
|
491 " pfmul 120(%%ebx), %%mm0\n\t"
|
|
|
492 " pfadd %%mm0,%%mm1\n\t"
|
|
|
493 " pfadd 124(%%edx), %%mm0\n\t"
|
|
|
494 " punpckldq %%mm1, %%mm0\n\t"
|
|
|
495 " movq %%mm0, 116(%%edx)\n\t"
|
|
|
496
|
|
|
497 " jnz .L01\n\t"
|
|
|
498
|
|
|
499 /* Phase 7*/
|
|
|
500 /* Code below is coded in scalar mode. Should be optimized */
|
|
|
501
|
|
|
502 " movd (%%ecx), %%mm0\n\t"
|
|
|
503 " pfadd 4(%%ecx), %%mm0\n\t"
|
|
|
504 " movd %%mm0, 1024(%%esi)\n\t"
|
|
|
505
|
|
|
506 " movd (%%ecx), %%mm0\n\t"
|
|
|
507 " pfsub 4(%%ecx), %%mm0\n\t"
|
|
|
508 " pfmul 120(%%ebx), %%mm0\n\t"
|
|
|
509 " movd %%mm0, (%%esi)\n\t"
|
|
|
510 " movd %%mm0, (%%edi)\n\t"
|
|
|
511
|
|
|
512 " movd 12(%%ecx), %%mm0\n\t"
|
|
|
513 " pfsub 8(%%ecx), %%mm0\n\t"
|
|
|
514 " pfmul 120(%%ebx), %%mm0\n\t"
|
|
|
515 " movd %%mm0, 512(%%edi)\n\t"
|
|
|
516 " pfadd 12(%%ecx), %%mm0\n\t"
|
|
|
517 " pfadd 8(%%ecx), %%mm0\n\t"
|
|
|
518 " movd %%mm0, 512(%%esi)\n\t"
|
|
|
519
|
|
|
520 " movd 16(%%ecx), %%mm0\n\t"
|
|
|
521 " pfsub 20(%%ecx), %%mm0\n\t"
|
|
|
522 " pfmul 120(%%ebx), %%mm0\n\t"
|
|
|
523 " movq %%mm0, %%mm3\n\t"
|
|
|
524
|
|
|
525 " movd 28(%%ecx), %%mm0\n\t"
|
|
|
526 " pfsub 24(%%ecx), %%mm0\n\t"
|
|
|
527 " pfmul 120(%%ebx), %%mm0\n\t"
|
|
|
528 " movd %%mm0, 768(%%edi)\n\t"
|
|
|
529 " movq %%mm0, %%mm2\n\t"
|
|
|
530
|
|
|
531 " pfadd 24(%%ecx), %%mm0\n\t"
|
|
|
532 " pfadd 28(%%ecx), %%mm0\n\t"
|
|
|
533 " movq %%mm0, %%mm1\n\t"
|
|
|
534
|
|
|
535 " pfadd 16(%%ecx), %%mm0\n\t"
|
|
|
536 " pfadd 20(%%ecx), %%mm0\n\t"
|
|
|
537 " movd %%mm0, 768(%%esi)\n\t"
|
|
|
538 " pfadd %%mm3, %%mm1\n\t"
|
|
|
539 " movd %%mm1, 256(%%esi)\n\t"
|
|
|
540 " pfadd %%mm3, %%mm2\n\t"
|
|
|
541 " movd %%mm2, 256(%%edi)\n\t"
|
|
|
542
|
|
|
543 /* Phase 8*/
|
|
|
544
|
|
|
545 " movq 32(%%edx), %%mm0\n\t"
|
|
|
546 " movq 48(%%edx), %%mm1\n\t"
|
|
|
547 " pfadd 48(%%edx), %%mm0\n\t"
|
|
|
548 " pfadd 40(%%edx), %%mm1\n\t"
|
|
|
549 " movd %%mm0, 896(%%esi)\n\t"
|
|
|
550 " movd %%mm1, 640(%%esi)\n\t"
|
|
|
551 " psrlq $32, %%mm0\n\t"
|
|
|
552 " psrlq $32, %%mm1\n\t"
|
|
|
553 " movd %%mm0, 128(%%edi)\n\t"
|
|
|
554 " movd %%mm1, 384(%%edi)\n\t"
|
|
|
555
|
|
|
556 " movd 40(%%edx), %%mm0\n\t"
|
|
|
557 " pfadd 56(%%edx), %%mm0\n\t"
|
|
|
558 " movd %%mm0, 384(%%esi)\n\t"
|
|
|
559
|
|
|
560 " movd 56(%%edx), %%mm0\n\t"
|
|
|
561 " pfadd 36(%%edx), %%mm0\n\t"
|
|
|
562 " movd %%mm0, 128(%%esi)\n\t"
|
|
|
563
|
|
|
564 " movd 60(%%edx), %%mm0\n\t"
|
|
|
565 " movd %%mm0, 896(%%edi)\n\t"
|
|
|
566 " pfadd 44(%%edx), %%mm0\n\t"
|
|
|
567 " movd %%mm0, 640(%%edi)\n\t"
|
|
|
568
|
|
|
569 " movq 96(%%edx), %%mm0\n\t"
|
|
|
570 " movq 112(%%edx), %%mm2\n\t"
|
|
|
571 " movq 104(%%edx), %%mm4\n\t"
|
|
|
572 " pfadd 112(%%edx), %%mm0\n\t"
|
|
|
573 " pfadd 104(%%edx), %%mm2\n\t"
|
|
|
574 " pfadd 120(%%edx), %%mm4\n\t"
|
|
|
575 " movq %%mm0, %%mm1\n\t"
|
|
|
576 " movq %%mm2, %%mm3\n\t"
|
|
|
577 " movq %%mm4, %%mm5\n\t"
|
|
|
578 " pfadd 64(%%edx), %%mm0\n\t"
|
|
|
579 " pfadd 80(%%edx), %%mm2\n\t"
|
|
|
580 " pfadd 72(%%edx), %%mm4\n\t"
|
|
|
581 " movd %%mm0, 960(%%esi)\n\t"
|
|
|
582 " movd %%mm2, 704(%%esi)\n\t"
|
|
|
583 " movd %%mm4, 448(%%esi)\n\t"
|
|
|
584 " psrlq $32, %%mm0\n\t"
|
|
|
585 " psrlq $32, %%mm2\n\t"
|
|
|
586 " psrlq $32, %%mm4\n\t"
|
|
|
587 " movd %%mm0, 64(%%edi)\n\t"
|
|
|
588 " movd %%mm2, 320(%%edi)\n\t"
|
|
|
589 " movd %%mm4, 576(%%edi)\n\t"
|
|
|
590 " pfadd 80(%%edx), %%mm1\n\t"
|
|
|
591 " pfadd 72(%%edx), %%mm3\n\t"
|
|
|
592 " pfadd 88(%%edx), %%mm5\n\t"
|
|
|
593 " movd %%mm1, 832(%%esi)\n\t"
|
|
|
594 " movd %%mm3, 576(%%esi)\n\t"
|
|
|
595 " movd %%mm5, 320(%%esi)\n\t"
|
|
|
596 " psrlq $32, %%mm1\n\t"
|
|
|
597 " psrlq $32, %%mm3\n\t"
|
|
|
598 " psrlq $32, %%mm5\n\t"
|
|
|
599 " movd %%mm1, 192(%%edi)\n\t"
|
|
|
600 " movd %%mm3, 448(%%edi)\n\t"
|
|
|
601 " movd %%mm5, 704(%%edi)\n\t"
|
|
|
602
|
|
|
603 " movd 120(%%edx), %%mm0\n\t"
|
|
|
604 " pfadd 100(%%edx), %%mm0\n\t"
|
|
|
605 " movq %%mm0, %%mm1\n\t"
|
|
|
606 " pfadd 88(%%edx), %%mm0\n\t"
|
|
|
607 " movd %%mm0, 192(%%esi)\n\t"
|
|
|
608 " pfadd 68(%%edx), %%mm1\n\t"
|
|
|
609 " movd %%mm1, 64(%%esi)\n\t"
|
|
|
610
|
|
|
611 " movd 124(%%edx), %%mm0\n\t"
|
|
|
612 " movd %%mm0, 960(%%edi)\n\t"
|
|
|
613 " pfadd 92(%%edx), %%mm0\n\t"
|
|
|
614 " movd %%mm0, 832(%%edi)\n\t"
|
|
|
615
|
|
|
616 " jmp .L_bye\n\t"
|
|
|
617 ".L01: \n\t"
|
|
|
618 /* Phase 9*/
|
|
|
619
|
|
|
620 " movq (%%ecx), %%mm0\n\t"
|
|
|
621 " movq %%mm0, %%mm1\n\t"
|
|
|
622 " pxor %%mm7, %%mm1\n\t"
|
|
|
623 " pfacc %%mm1, %%mm0\n\t"
|
|
|
624 " pfmul %%mm6, %%mm0\n\t"
|
|
|
625 " pf2id %%mm0, %%mm0\n\t"
|
|
|
626 " movd %%mm0, %%eax\n\t"
|
|
|
627 " movw %%ax, 512(%%esi)\n\t"
|
|
|
628 " psrlq $32, %%mm0\n\t"
|
|
|
629 " movd %%mm0, %%eax\n\t"
|
|
|
630 " movw %%ax, (%%esi)\n\t"
|
|
|
631
|
|
|
632 " movd 12(%%ecx), %%mm0\n\t"
|
|
|
633 " pfsub 8(%%ecx), %%mm0\n\t"
|
|
|
634 " pfmul 120(%%ebx), %%mm0\n\t"
|
|
|
635 " pf2id %%mm0, %%mm7\n\t"
|
|
|
636 " movd %%mm7, %%eax\n\t"
|
|
|
637 " movw %%ax, 256(%%edi)\n\t"
|
|
|
638 " pfadd 12(%%ecx), %%mm0\n\t"
|
|
|
639 " pfadd 8(%%ecx), %%mm0\n\t"
|
|
|
640 " pf2id %%mm0, %%mm0\n\t"
|
|
|
641 " movd %%mm0, %%eax\n\t"
|
|
|
642 " movw %%ax, 256(%%esi)\n\t"
|
|
|
643
|
|
|
644 " movd 16(%%ecx), %%mm3\n\t"
|
|
|
645 " pfsub 20(%%ecx), %%mm3\n\t"
|
|
|
646 " pfmul 120(%%ebx), %%mm3\n\t"
|
|
|
647 " movq %%mm3, %%mm2\n\t"
|
|
|
648
|
|
|
649 " movd 28(%%ecx), %%mm2\n\t"
|
|
|
650 " pfsub 24(%%ecx), %%mm2\n\t"
|
|
|
651 " pfmul 120(%%ebx), %%mm2\n\t"
|
|
|
652 " movq %%mm2, %%mm1\n\t"
|
|
|
653
|
|
|
654 " pf2id %%mm2, %%mm7\n\t"
|
|
|
655 " movd %%mm7, %%eax\n\t"
|
|
|
656 " movw %%ax, 384(%%edi)\n\t"
|
|
|
657
|
|
|
658 " pfadd 24(%%ecx), %%mm1\n\t"
|
|
|
659 " pfadd 28(%%ecx), %%mm1\n\t"
|
|
|
660 " movq %%mm1, %%mm0\n\t"
|
|
|
661
|
|
|
662 " pfadd 16(%%ecx), %%mm0\n\t"
|
|
|
663 " pfadd 20(%%ecx), %%mm0\n\t"
|
|
|
664 " pf2id %%mm0, %%mm0\n\t"
|
|
|
665 " movd %%mm0, %%eax\n\t"
|
|
|
666 " movw %%ax, 384(%%esi)\n\t"
|
|
|
667 " pfadd %%mm3, %%mm1\n\t"
|
|
|
668 " pf2id %%mm1, %%mm1\n\t"
|
|
|
669 " movd %%mm1, %%eax\n\t"
|
|
|
670 " movw %%ax, 128(%%esi)\n\t"
|
|
|
671 " pfadd %%mm3, %%mm2\n\t"
|
|
|
672 " pf2id %%mm2, %%mm2\n\t"
|
|
|
673 " movd %%mm2, %%eax\n\t"
|
|
|
674 " movw %%ax, 128(%%edi)\n\t"
|
|
|
675
|
|
|
676 /* Phase 10*/
|
|
|
677
|
|
|
678 " movq 32(%%edx), %%mm0\n\t"
|
|
|
679 " movq 48(%%edx), %%mm1\n\t"
|
|
|
680 " pfadd 48(%%edx), %%mm0\n\t"
|
|
|
681 " pfadd 40(%%edx), %%mm1\n\t"
|
|
|
682 " pf2id %%mm0, %%mm0\n\t"
|
|
|
683 " pf2id %%mm1, %%mm1\n\t"
|
|
|
684 " movd %%mm0, %%eax\n\t"
|
|
|
685 " movd %%mm1, %%ecx\n\t"
|
|
|
686 " movw %%ax, 448(%%esi)\n\t"
|
|
|
687 " movw %%cx, 320(%%esi)\n\t"
|
|
|
688 " psrlq $32, %%mm0\n\t"
|
|
|
689 " psrlq $32, %%mm1\n\t"
|
|
|
690 " movd %%mm0, %%eax\n\t"
|
|
|
691 " movd %%mm1, %%ecx\n\t"
|
|
|
692 " movw %%ax, 64(%%edi)\n\t"
|
|
|
693 " movw %%cx, 192(%%edi)\n\t"
|
|
|
694
|
|
|
695 " movd 40(%%edx), %%mm3\n\t"
|
|
|
696 " movd 56(%%edx), %%mm4\n\t"
|
|
|
697 " movd 60(%%edx), %%mm0\n\t"
|
|
|
698 " movd 44(%%edx), %%mm2\n\t"
|
|
|
699 " movd 120(%%edx), %%mm5\n\t"
|
|
|
700 " punpckldq %%mm4, %%mm3\n\t"
|
|
|
701 " punpckldq 124(%%edx), %%mm0\n\t"
|
|
|
702 " pfadd 100(%%edx), %%mm5\n\t"
|
|
|
703 " punpckldq 36(%%edx), %%mm4\n\t"
|
|
|
704 " punpckldq 92(%%edx), %%mm2\n\t"
|
|
|
705 " movq %%mm5, %%mm6\n\t"
|
|
|
706 " pfadd %%mm4, %%mm3\n\t"
|
|
|
707 " pf2id %%mm0, %%mm1\n\t"
|
|
|
708 " pf2id %%mm3, %%mm3\n\t"
|
|
|
709 " pfadd 88(%%edx), %%mm5\n\t"
|
|
|
710 " movd %%mm1, %%eax\n\t"
|
|
|
711 " movd %%mm3, %%ecx\n\t"
|
|
|
712 " movw %%ax, 448(%%edi)\n\t"
|
|
|
713 " movw %%cx, 192(%%esi)\n\t"
|
|
|
714 " pf2id %%mm5, %%mm5\n\t"
|
|
|
715 " psrlq $32, %%mm1\n\t"
|
|
|
716 " psrlq $32, %%mm3\n\t"
|
|
|
717 " movd %%mm5, %%ebx\n\t"
|
|
|
718 " movd %%mm1, %%eax\n\t"
|
|
|
719 " movd %%mm3, %%ecx\n\t"
|
|
|
720 " movw %%bx, 96(%%esi)\n\t"
|
|
|
721 " movw %%ax, 480(%%edi)\n\t"
|
|
|
722 " movw %%cx, 64(%%esi)\n\t"
|
|
|
723 " pfadd %%mm2, %%mm0\n\t"
|
|
|
724 " pf2id %%mm0, %%mm0\n\t"
|
|
|
725 " movd %%mm0, %%eax\n\t"
|
|
|
726 " pfadd 68(%%edx), %%mm6\n\t"
|
|
|
727 " movw %%ax, 320(%%edi)\n\t"
|
|
|
728 " psrlq $32, %%mm0\n\t"
|
|
|
729 " pf2id %%mm6, %%mm6\n\t"
|
|
|
730 " movd %%mm0, %%eax\n\t"
|
|
|
731 " movd %%mm6, %%ebx\n\t"
|
|
|
732 " movw %%ax, 416(%%edi)\n\t"
|
|
|
733 " movw %%bx, 32(%%esi)\n\t"
|
|
|
734
|
|
|
735 " movq 96(%%edx), %%mm0\n\t"
|
|
|
736 " movq 112(%%edx), %%mm2\n\t"
|
|
|
737 " movq 104(%%edx), %%mm4\n\t"
|
|
|
738 " pfadd %%mm2, %%mm0\n\t"
|
|
|
739 " pfadd %%mm4, %%mm2\n\t"
|
|
|
740 " pfadd 120(%%edx), %%mm4\n\t"
|
|
|
741 " movq %%mm0, %%mm1\n\t"
|
|
|
742 " movq %%mm2, %%mm3\n\t"
|
|
|
743 " movq %%mm4, %%mm5\n\t"
|
|
|
744 " pfadd 64(%%edx), %%mm0\n\t"
|
|
|
745 " pfadd 80(%%edx), %%mm2\n\t"
|
|
|
746 " pfadd 72(%%edx), %%mm4\n\t"
|
|
|
747 " pf2id %%mm0, %%mm0\n\t"
|
|
|
748 " pf2id %%mm2, %%mm2\n\t"
|
|
|
749 " pf2id %%mm4, %%mm4\n\t"
|
|
|
750 " movd %%mm0, %%eax\n\t"
|
|
|
751 " movd %%mm2, %%ecx\n\t"
|
|
|
752 " movd %%mm4, %%ebx\n\t"
|
|
|
753 " movw %%ax, 480(%%esi)\n\t"
|
|
|
754 " movw %%cx, 352(%%esi)\n\t"
|
|
|
755 " movw %%bx, 224(%%esi)\n\t"
|
|
|
756 " psrlq $32, %%mm0\n\t"
|
|
|
757 " psrlq $32, %%mm2\n\t"
|
|
|
758 " psrlq $32, %%mm4\n\t"
|
|
|
759 " movd %%mm0, %%eax\n\t"
|
|
|
760 " movd %%mm2, %%ecx\n\t"
|
|
|
761 " movd %%mm4, %%ebx\n\t"
|
|
|
762 " movw %%ax, 32(%%edi)\n\t"
|
|
|
763 " movw %%cx, 160(%%edi)\n\t"
|
|
|
764 " movw %%bx, 288(%%edi)\n\t"
|
|
|
765 " pfadd 80(%%edx), %%mm1\n\t"
|
|
|
766 " pfadd 72(%%edx), %%mm3\n\t"
|
|
|
767 " pfadd 88(%%edx), %%mm5\n\t"
|
|
|
768 " pf2id %%mm1, %%mm1\n\t"
|
|
|
769 " pf2id %%mm3, %%mm3\n\t"
|
|
|
770 " pf2id %%mm5, %%mm5\n\t"
|
|
|
771 " movd %%mm1, %%eax\n\t"
|
|
|
772 " movd %%mm3, %%ecx\n\t"
|
|
|
773 " movd %%mm5, %%ebx\n\t"
|
|
|
774 " movw %%ax, 416(%%esi)\n\t"
|
|
|
775 " movw %%cx, 288(%%esi)\n\t"
|
|
|
776 " movw %%bx, 160(%%esi)\n\t"
|
|
|
777 " psrlq $32, %%mm1\n\t"
|
|
|
778 " psrlq $32, %%mm3\n\t"
|
|
|
779 " psrlq $32, %%mm5\n\t"
|
|
|
780 " movd %%mm1, %%eax\n\t"
|
|
|
781 " movd %%mm3, %%ecx\n\t"
|
|
|
782 " movd %%mm5, %%ebx\n\t"
|
|
|
783 " movw %%ax, 96(%%edi)\n\t"
|
|
|
784 " movw %%cx, 224(%%edi)\n\t"
|
|
|
785 " movw %%bx, 352(%%edi)\n\t"
|
|
|
786
|
|
|
787 " movsw\n\t"
|
|
|
788
|
|
|
789 ".L_bye:\n\t"
|
|
|
790 " femms\n\t"
|
|
|
791 :
|
|
|
792 :"m"(a),"m"(b),"m"(c),"m"(tmp[0])
|
|
|
793 :"memory","%ebx","%esi","%edi");
|
|
|
794 } |
