Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
3960 | Serge | 1 | /* |
2 | dct64_3dnow.s: Replacement of dct36() with AMD's 3DNow! SIMD operations support |
||
3 | |||
4 | copyright ?-2006 by the mpg123 project - free software under the terms of the LGPL 2.1 |
||
5 | see COPYING and AUTHORS files in distribution or http://mpg123.org |
||
6 | initially written by Syuuhei Kashiyama |
||
7 | |||
8 | This code based 'dct36_3dnow.s' by Syuuhei Kashiyama |
||
9 |
|
||
10 | |||
11 | - remove PREFETCH instruction for speedup |
||
12 | - change function name for support 3DNow! automatic detect |
||
13 | |||
14 | You can find Kashiyama's original 3dnow! support patch |
||
15 | (for mpg123-0.59o) at |
||
16 | http://user.ecc.u-tokyo.ac.jp/~g810370/linux-simd/ (Japanese). |
||
17 | |||
18 | by KIMURA Takuhiro |
||
19 |
|
||
20 | |||
21 | Replacement of dct36() with AMD's 3DNow! SIMD operations support |
||
22 | |||
23 | Syuuhei Kashiyama |
||
24 | |||
25 | The author of this program disclaim whole expressed or implied |
||
26 | warranties with regard to this program, and in no event shall the |
||
27 | author of this program liable to whatever resulted from the use of |
||
28 | this program. Use it at your own risk. |
||
29 | */ |
||
30 | |||
31 | #include "mangle.h" |
||
32 | |||
33 | .globl ASM_NAME(dct36_3dnow) |
||
34 | /* .type ASM_NAME(dct36_3dnow),@function */ |
||
35 | ASM_NAME(dct36_3dnow): |
||
36 | pushl %ebp |
||
37 | movl %esp,%ebp |
||
38 | subl $120,%esp |
||
39 | pushl %esi |
||
40 | pushl %ebx |
||
41 | movl 8(%ebp),%eax |
||
42 | movl 12(%ebp),%esi |
||
43 | movl 16(%ebp),%ecx |
||
44 | movl 20(%ebp),%edx |
||
45 | movl 24(%ebp),%ebx |
||
46 | leal -128(%ebp),%esp |
||
47 | |||
48 | femms |
||
49 | movq (%eax),%mm0 |
||
50 | movq 4(%eax),%mm1 |
||
51 | pfadd %mm1,%mm0 |
||
52 | movq %mm0,4(%eax) |
||
53 | psrlq $32,%mm1 |
||
54 | movq 12(%eax),%mm2 |
||
55 | punpckldq %mm2,%mm1 |
||
56 | pfadd %mm2,%mm1 |
||
57 | movq %mm1,12(%eax) |
||
58 | psrlq $32,%mm2 |
||
59 | movq 20(%eax),%mm3 |
||
60 | punpckldq %mm3,%mm2 |
||
61 | pfadd %mm3,%mm2 |
||
62 | movq %mm2,20(%eax) |
||
63 | psrlq $32,%mm3 |
||
64 | movq 28(%eax),%mm4 |
||
65 | punpckldq %mm4,%mm3 |
||
66 | pfadd %mm4,%mm3 |
||
67 | movq %mm3,28(%eax) |
||
68 | psrlq $32,%mm4 |
||
69 | movq 36(%eax),%mm5 |
||
70 | punpckldq %mm5,%mm4 |
||
71 | pfadd %mm5,%mm4 |
||
72 | movq %mm4,36(%eax) |
||
73 | psrlq $32,%mm5 |
||
74 | movq 44(%eax),%mm6 |
||
75 | punpckldq %mm6,%mm5 |
||
76 | pfadd %mm6,%mm5 |
||
77 | movq %mm5,44(%eax) |
||
78 | psrlq $32,%mm6 |
||
79 | movq 52(%eax),%mm7 |
||
80 | punpckldq %mm7,%mm6 |
||
81 | pfadd %mm7,%mm6 |
||
82 | movq %mm6,52(%eax) |
||
83 | psrlq $32,%mm7 |
||
84 | movq 60(%eax),%mm0 |
||
85 | punpckldq %mm0,%mm7 |
||
86 | pfadd %mm0,%mm7 |
||
87 | movq %mm7,60(%eax) |
||
88 | psrlq $32,%mm0 |
||
89 | movd 68(%eax),%mm1 |
||
90 | pfadd %mm1,%mm0 |
||
91 | movd %mm0,68(%eax) |
||
92 | movd 4(%eax),%mm0 |
||
93 | movd 12(%eax),%mm1 |
||
94 | punpckldq %mm1,%mm0 |
||
95 | punpckldq 20(%eax),%mm1 |
||
96 | pfadd %mm1,%mm0 |
||
97 | movd %mm0,12(%eax) |
||
98 | psrlq $32,%mm0 |
||
99 | movd %mm0,20(%eax) |
||
100 | psrlq $32,%mm1 |
||
101 | movd 28(%eax),%mm2 |
||
102 | punpckldq %mm2,%mm1 |
||
103 | punpckldq 36(%eax),%mm2 |
||
104 | pfadd %mm2,%mm1 |
||
105 | movd %mm1,28(%eax) |
||
106 | psrlq $32,%mm1 |
||
107 | movd %mm1,36(%eax) |
||
108 | psrlq $32,%mm2 |
||
109 | movd 44(%eax),%mm3 |
||
110 | punpckldq %mm3,%mm2 |
||
111 | punpckldq 52(%eax),%mm3 |
||
112 | pfadd %mm3,%mm2 |
||
113 | movd %mm2,44(%eax) |
||
114 | psrlq $32,%mm2 |
||
115 | movd %mm2,52(%eax) |
||
116 | psrlq $32,%mm3 |
||
117 | movd 60(%eax),%mm4 |
||
118 | punpckldq %mm4,%mm3 |
||
119 | punpckldq 68(%eax),%mm4 |
||
120 | pfadd %mm4,%mm3 |
||
121 | movd %mm3,60(%eax) |
||
122 | psrlq $32,%mm3 |
||
123 | movd %mm3,68(%eax) |
||
124 | |||
125 | movq 24(%eax),%mm0 |
||
126 | movq 48(%eax),%mm1 |
||
127 | movd ASM_NAME(COS9)+12,%mm2 |
||
128 | punpckldq %mm2,%mm2 |
||
129 | movd ASM_NAME(COS9)+24,%mm3 |
||
130 | punpckldq %mm3,%mm3 |
||
131 | pfmul %mm2,%mm0 |
||
132 | pfmul %mm3,%mm1 |
||
133 | pushl %eax |
||
134 | movl $1,%eax |
||
135 | movd %eax,%mm7 |
||
136 | pi2fd %mm7,%mm7 |
||
137 | popl %eax |
||
138 | movq 8(%eax),%mm2 |
||
139 | movd ASM_NAME(COS9)+4,%mm3 |
||
140 | punpckldq %mm3,%mm3 |
||
141 | pfmul %mm3,%mm2 |
||
142 | pfadd %mm0,%mm2 |
||
143 | movq 40(%eax),%mm3 |
||
144 | movd ASM_NAME(COS9)+20,%mm4 |
||
145 | punpckldq %mm4,%mm4 |
||
146 | pfmul %mm4,%mm3 |
||
147 | pfadd %mm3,%mm2 |
||
148 | movq 56(%eax),%mm3 |
||
149 | movd ASM_NAME(COS9)+28,%mm4 |
||
150 | punpckldq %mm4,%mm4 |
||
151 | pfmul %mm4,%mm3 |
||
152 | pfadd %mm3,%mm2 |
||
153 | movq (%eax),%mm3 |
||
154 | movq 16(%eax),%mm4 |
||
155 | movd ASM_NAME(COS9)+8,%mm5 |
||
156 | punpckldq %mm5,%mm5 |
||
157 | pfmul %mm5,%mm4 |
||
158 | pfadd %mm4,%mm3 |
||
159 | movq 32(%eax),%mm4 |
||
160 | movd ASM_NAME(COS9)+16,%mm5 |
||
161 | punpckldq %mm5,%mm5 |
||
162 | pfmul %mm5,%mm4 |
||
163 | pfadd %mm4,%mm3 |
||
164 | pfadd %mm1,%mm3 |
||
165 | movq 64(%eax),%mm4 |
||
166 | movd ASM_NAME(COS9)+32,%mm5 |
||
167 | punpckldq %mm5,%mm5 |
||
168 | pfmul %mm5,%mm4 |
||
169 | pfadd %mm4,%mm3 |
||
170 | movq %mm2,%mm4 |
||
171 | pfadd %mm3,%mm4 |
||
172 | movq %mm7,%mm5 |
||
173 | punpckldq ASM_NAME(tfcos36)+0,%mm5 |
||
174 | pfmul %mm5,%mm4 |
||
175 | movq %mm4,%mm5 |
||
176 | pfacc %mm5,%mm5 |
||
177 | movd 108(%edx),%mm6 |
||
178 | punpckldq 104(%edx),%mm6 |
||
179 | pfmul %mm6,%mm5 |
||
180 | movd %mm5,36(%ecx) |
||
181 | psrlq $32,%mm5 |
||
182 | movd %mm5,32(%ecx) |
||
183 | movq %mm4,%mm6 |
||
184 | punpckldq %mm6,%mm5 |
||
185 | pfsub %mm6,%mm5 |
||
186 | punpckhdq %mm5,%mm5 |
||
187 | movd 32(%edx),%mm6 |
||
188 | punpckldq 36(%edx),%mm6 |
||
189 | pfmul %mm6,%mm5 |
||
190 | movd 32(%esi),%mm6 |
||
191 | punpckldq 36(%esi),%mm6 |
||
192 | pfadd %mm6,%mm5 |
||
193 | movd %mm5,1024(%ebx) |
||
194 | psrlq $32,%mm5 |
||
195 | movd %mm5,1152(%ebx) |
||
196 | movq %mm3,%mm4 |
||
197 | pfsub %mm2,%mm4 |
||
198 | movq %mm7,%mm5 |
||
199 | punpckldq ASM_NAME(tfcos36)+32,%mm5 |
||
200 | pfmul %mm5,%mm4 |
||
201 | movq %mm4,%mm5 |
||
202 | pfacc %mm5,%mm5 |
||
203 | movd 140(%edx),%mm6 |
||
204 | punpckldq 72(%edx),%mm6 |
||
205 | pfmul %mm6,%mm5 |
||
206 | movd %mm5,68(%ecx) |
||
207 | psrlq $32,%mm5 |
||
208 | movd %mm5,0(%ecx) |
||
209 | movq %mm4,%mm6 |
||
210 | punpckldq %mm6,%mm5 |
||
211 | pfsub %mm6,%mm5 |
||
212 | punpckhdq %mm5,%mm5 |
||
213 | movd 0(%edx),%mm6 |
||
214 | punpckldq 68(%edx),%mm6 |
||
215 | pfmul %mm6,%mm5 |
||
216 | movd 0(%esi),%mm6 |
||
217 | punpckldq 68(%esi),%mm6 |
||
218 | pfadd %mm6,%mm5 |
||
219 | movd %mm5,0(%ebx) |
||
220 | psrlq $32,%mm5 |
||
221 | movd %mm5,2176(%ebx) |
||
222 | movq 8(%eax),%mm2 |
||
223 | movq 40(%eax),%mm3 |
||
224 | pfsub %mm3,%mm2 |
||
225 | movq 56(%eax),%mm3 |
||
226 | pfsub %mm3,%mm2 |
||
227 | movd ASM_NAME(COS9)+12,%mm3 |
||
228 | punpckldq %mm3,%mm3 |
||
229 | pfmul %mm3,%mm2 |
||
230 | movq 16(%eax),%mm3 |
||
231 | movq 32(%eax),%mm4 |
||
232 | pfsub %mm4,%mm3 |
||
233 | movq 64(%eax),%mm4 |
||
234 | pfsub %mm4,%mm3 |
||
235 | movd ASM_NAME(COS9)+24,%mm4 |
||
236 | punpckldq %mm4,%mm4 |
||
237 | pfmul %mm4,%mm3 |
||
238 | movq 48(%eax),%mm4 |
||
239 | pfsub %mm4,%mm3 |
||
240 | movq (%eax),%mm4 |
||
241 | pfadd %mm4,%mm3 |
||
242 | movq %mm2,%mm4 |
||
243 | pfadd %mm3,%mm4 |
||
244 | movq %mm7,%mm5 |
||
245 | punpckldq ASM_NAME(tfcos36)+4,%mm5 |
||
246 | pfmul %mm5,%mm4 |
||
247 | movq %mm4,%mm5 |
||
248 | pfacc %mm5,%mm5 |
||
249 | movd 112(%edx),%mm6 |
||
250 | punpckldq 100(%edx),%mm6 |
||
251 | pfmul %mm6,%mm5 |
||
252 | movd %mm5,40(%ecx) |
||
253 | psrlq $32,%mm5 |
||
254 | movd %mm5,28(%ecx) |
||
255 | movq %mm4,%mm6 |
||
256 | punpckldq %mm6,%mm5 |
||
257 | pfsub %mm6,%mm5 |
||
258 | punpckhdq %mm5,%mm5 |
||
259 | movd 28(%edx),%mm6 |
||
260 | punpckldq 40(%edx),%mm6 |
||
261 | pfmul %mm6,%mm5 |
||
262 | movd 28(%esi),%mm6 |
||
263 | punpckldq 40(%esi),%mm6 |
||
264 | pfadd %mm6,%mm5 |
||
265 | movd %mm5,896(%ebx) |
||
266 | psrlq $32,%mm5 |
||
267 | movd %mm5,1280(%ebx) |
||
268 | movq %mm3,%mm4 |
||
269 | pfsub %mm2,%mm4 |
||
270 | movq %mm7,%mm5 |
||
271 | punpckldq ASM_NAME(tfcos36)+28,%mm5 |
||
272 | pfmul %mm5,%mm4 |
||
273 | movq %mm4,%mm5 |
||
274 | pfacc %mm5,%mm5 |
||
275 | movd 136(%edx),%mm6 |
||
276 | punpckldq 76(%edx),%mm6 |
||
277 | pfmul %mm6,%mm5 |
||
278 | movd %mm5,64(%ecx) |
||
279 | psrlq $32,%mm5 |
||
280 | movd %mm5,4(%ecx) |
||
281 | movq %mm4,%mm6 |
||
282 | punpckldq %mm6,%mm5 |
||
283 | pfsub %mm6,%mm5 |
||
284 | punpckhdq %mm5,%mm5 |
||
285 | movd 4(%edx),%mm6 |
||
286 | punpckldq 64(%edx),%mm6 |
||
287 | pfmul %mm6,%mm5 |
||
288 | movd 4(%esi),%mm6 |
||
289 | punpckldq 64(%esi),%mm6 |
||
290 | pfadd %mm6,%mm5 |
||
291 | movd %mm5,128(%ebx) |
||
292 | psrlq $32,%mm5 |
||
293 | movd %mm5,2048(%ebx) |
||
294 | |||
295 | movq 8(%eax),%mm2 |
||
296 | movd ASM_NAME(COS9)+20,%mm3 |
||
297 | punpckldq %mm3,%mm3 |
||
298 | pfmul %mm3,%mm2 |
||
299 | pfsub %mm0,%mm2 |
||
300 | movq 40(%eax),%mm3 |
||
301 | movd ASM_NAME(COS9)+28,%mm4 |
||
302 | punpckldq %mm4,%mm4 |
||
303 | pfmul %mm4,%mm3 |
||
304 | pfsub %mm3,%mm2 |
||
305 | movq 56(%eax),%mm3 |
||
306 | movd ASM_NAME(COS9)+4,%mm4 |
||
307 | punpckldq %mm4,%mm4 |
||
308 | pfmul %mm4,%mm3 |
||
309 | pfadd %mm3,%mm2 |
||
310 | movq (%eax),%mm3 |
||
311 | movq 16(%eax),%mm4 |
||
312 | movd ASM_NAME(COS9)+32,%mm5 |
||
313 | punpckldq %mm5,%mm5 |
||
314 | pfmul %mm5,%mm4 |
||
315 | pfsub %mm4,%mm3 |
||
316 | movq 32(%eax),%mm4 |
||
317 | movd ASM_NAME(COS9)+8,%mm5 |
||
318 | punpckldq %mm5,%mm5 |
||
319 | pfmul %mm5,%mm4 |
||
320 | pfsub %mm4,%mm3 |
||
321 | pfadd %mm1,%mm3 |
||
322 | movq 64(%eax),%mm4 |
||
323 | movd ASM_NAME(COS9)+16,%mm5 |
||
324 | punpckldq %mm5,%mm5 |
||
325 | pfmul %mm5,%mm4 |
||
326 | pfadd %mm4,%mm3 |
||
327 | movq %mm2,%mm4 |
||
328 | pfadd %mm3,%mm4 |
||
329 | movq %mm7,%mm5 |
||
330 | punpckldq ASM_NAME(tfcos36)+8,%mm5 |
||
331 | pfmul %mm5,%mm4 |
||
332 | movq %mm4,%mm5 |
||
333 | pfacc %mm5,%mm5 |
||
334 | movd 116(%edx),%mm6 |
||
335 | punpckldq 96(%edx),%mm6 |
||
336 | pfmul %mm6,%mm5 |
||
337 | movd %mm5,44(%ecx) |
||
338 | psrlq $32,%mm5 |
||
339 | movd %mm5,24(%ecx) |
||
340 | movq %mm4,%mm6 |
||
341 | punpckldq %mm6,%mm5 |
||
342 | pfsub %mm6,%mm5 |
||
343 | punpckhdq %mm5,%mm5 |
||
344 | movd 24(%edx),%mm6 |
||
345 | punpckldq 44(%edx),%mm6 |
||
346 | pfmul %mm6,%mm5 |
||
347 | movd 24(%esi),%mm6 |
||
348 | punpckldq 44(%esi),%mm6 |
||
349 | pfadd %mm6,%mm5 |
||
350 | movd %mm5,768(%ebx) |
||
351 | psrlq $32,%mm5 |
||
352 | movd %mm5,1408(%ebx) |
||
353 | movq %mm3,%mm4 |
||
354 | pfsub %mm2,%mm4 |
||
355 | movq %mm7,%mm5 |
||
356 | punpckldq ASM_NAME(tfcos36)+24,%mm5 |
||
357 | pfmul %mm5,%mm4 |
||
358 | movq %mm4,%mm5 |
||
359 | pfacc %mm5,%mm5 |
||
360 | movd 132(%edx),%mm6 |
||
361 | punpckldq 80(%edx),%mm6 |
||
362 | pfmul %mm6,%mm5 |
||
363 | movd %mm5,60(%ecx) |
||
364 | psrlq $32,%mm5 |
||
365 | movd %mm5,8(%ecx) |
||
366 | movq %mm4,%mm6 |
||
367 | punpckldq %mm6,%mm5 |
||
368 | pfsub %mm6,%mm5 |
||
369 | punpckhdq %mm5,%mm5 |
||
370 | movd 8(%edx),%mm6 |
||
371 | punpckldq 60(%edx),%mm6 |
||
372 | pfmul %mm6,%mm5 |
||
373 | movd 8(%esi),%mm6 |
||
374 | punpckldq 60(%esi),%mm6 |
||
375 | pfadd %mm6,%mm5 |
||
376 | movd %mm5,256(%ebx) |
||
377 | psrlq $32,%mm5 |
||
378 | movd %mm5,1920(%ebx) |
||
379 | movq 8(%eax),%mm2 |
||
380 | movd ASM_NAME(COS9)+28,%mm3 |
||
381 | punpckldq %mm3,%mm3 |
||
382 | pfmul %mm3,%mm2 |
||
383 | pfsub %mm0,%mm2 |
||
384 | movq 40(%eax),%mm3 |
||
385 | movd ASM_NAME(COS9)+4,%mm4 |
||
386 | punpckldq %mm4,%mm4 |
||
387 | pfmul %mm4,%mm3 |
||
388 | pfadd %mm3,%mm2 |
||
389 | movq 56(%eax),%mm3 |
||
390 | movd ASM_NAME(COS9)+20,%mm4 |
||
391 | punpckldq %mm4,%mm4 |
||
392 | pfmul %mm4,%mm3 |
||
393 | pfsub %mm3,%mm2 |
||
394 | movq (%eax),%mm3 |
||
395 | movq 16(%eax),%mm4 |
||
396 | movd ASM_NAME(COS9)+16,%mm5 |
||
397 | punpckldq %mm5,%mm5 |
||
398 | pfmul %mm5,%mm4 |
||
399 | pfsub %mm4,%mm3 |
||
400 | movq 32(%eax),%mm4 |
||
401 | movd ASM_NAME(COS9)+32,%mm5 |
||
402 | punpckldq %mm5,%mm5 |
||
403 | pfmul %mm5,%mm4 |
||
404 | pfadd %mm4,%mm3 |
||
405 | pfadd %mm1,%mm3 |
||
406 | movq 64(%eax),%mm4 |
||
407 | movd ASM_NAME(COS9)+8,%mm5 |
||
408 | punpckldq %mm5,%mm5 |
||
409 | pfmul %mm5,%mm4 |
||
410 | pfsub %mm4,%mm3 |
||
411 | movq %mm2,%mm4 |
||
412 | pfadd %mm3,%mm4 |
||
413 | movq %mm7,%mm5 |
||
414 | punpckldq ASM_NAME(tfcos36)+12,%mm5 |
||
415 | pfmul %mm5,%mm4 |
||
416 | movq %mm4,%mm5 |
||
417 | pfacc %mm5,%mm5 |
||
418 | movd 120(%edx),%mm6 |
||
419 | punpckldq 92(%edx),%mm6 |
||
420 | pfmul %mm6,%mm5 |
||
421 | movd %mm5,48(%ecx) |
||
422 | psrlq $32,%mm5 |
||
423 | movd %mm5,20(%ecx) |
||
424 | movq %mm4,%mm6 |
||
425 | punpckldq %mm6,%mm5 |
||
426 | pfsub %mm6,%mm5 |
||
427 | punpckhdq %mm5,%mm5 |
||
428 | movd 20(%edx),%mm6 |
||
429 | punpckldq 48(%edx),%mm6 |
||
430 | pfmul %mm6,%mm5 |
||
431 | movd 20(%esi),%mm6 |
||
432 | punpckldq 48(%esi),%mm6 |
||
433 | pfadd %mm6,%mm5 |
||
434 | movd %mm5,640(%ebx) |
||
435 | psrlq $32,%mm5 |
||
436 | movd %mm5,1536(%ebx) |
||
437 | movq %mm3,%mm4 |
||
438 | pfsub %mm2,%mm4 |
||
439 | movq %mm7,%mm5 |
||
440 | punpckldq ASM_NAME(tfcos36)+20,%mm5 |
||
441 | pfmul %mm5,%mm4 |
||
442 | movq %mm4,%mm5 |
||
443 | pfacc %mm5,%mm5 |
||
444 | movd 128(%edx),%mm6 |
||
445 | punpckldq 84(%edx),%mm6 |
||
446 | pfmul %mm6,%mm5 |
||
447 | movd %mm5,56(%ecx) |
||
448 | psrlq $32,%mm5 |
||
449 | movd %mm5,12(%ecx) |
||
450 | movq %mm4,%mm6 |
||
451 | punpckldq %mm6,%mm5 |
||
452 | pfsub %mm6,%mm5 |
||
453 | punpckhdq %mm5,%mm5 |
||
454 | movd 12(%edx),%mm6 |
||
455 | punpckldq 56(%edx),%mm6 |
||
456 | pfmul %mm6,%mm5 |
||
457 | movd 12(%esi),%mm6 |
||
458 | punpckldq 56(%esi),%mm6 |
||
459 | pfadd %mm6,%mm5 |
||
460 | movd %mm5,384(%ebx) |
||
461 | psrlq $32,%mm5 |
||
462 | movd %mm5,1792(%ebx) |
||
463 | |||
464 | movq (%eax),%mm4 |
||
465 | movq 16(%eax),%mm3 |
||
466 | pfsub %mm3,%mm4 |
||
467 | movq 32(%eax),%mm3 |
||
468 | pfadd %mm3,%mm4 |
||
469 | movq 48(%eax),%mm3 |
||
470 | pfsub %mm3,%mm4 |
||
471 | movq 64(%eax),%mm3 |
||
472 | pfadd %mm3,%mm4 |
||
473 | movq %mm7,%mm5 |
||
474 | punpckldq ASM_NAME(tfcos36)+16,%mm5 |
||
475 | pfmul %mm5,%mm4 |
||
476 | movq %mm4,%mm5 |
||
477 | pfacc %mm5,%mm5 |
||
478 | movd 124(%edx),%mm6 |
||
479 | punpckldq 88(%edx),%mm6 |
||
480 | pfmul %mm6,%mm5 |
||
481 | movd %mm5,52(%ecx) |
||
482 | psrlq $32,%mm5 |
||
483 | movd %mm5,16(%ecx) |
||
484 | movq %mm4,%mm6 |
||
485 | punpckldq %mm6,%mm5 |
||
486 | pfsub %mm6,%mm5 |
||
487 | punpckhdq %mm5,%mm5 |
||
488 | movd 16(%edx),%mm6 |
||
489 | punpckldq 52(%edx),%mm6 |
||
490 | pfmul %mm6,%mm5 |
||
491 | movd 16(%esi),%mm6 |
||
492 | punpckldq 52(%esi),%mm6 |
||
493 | pfadd %mm6,%mm5 |
||
494 | movd %mm5,512(%ebx) |
||
495 | psrlq $32,%mm5 |
||
496 | movd %mm5,1664(%ebx) |
||
497 | |||
498 | femms |
||
499 | popl %ebx |
||
500 | popl %esi |
||
501 | movl %ebp,%esp |
||
502 | popl %ebp |
||
503 | ret |
||
504 | |||
505 | NONEXEC_STACK |