Mercurial > mplayer.hg

annotate mp3lib/dct64_k7.c @ 13639:12028d59e650

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
10l to me, noticed by Torinthiel
author diego
date Thu, 14 Oct 2004 12:23:42 +0000
parents 114f3d149324
children c0fa83fa7cec
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
4148
nick
parents:
diff changeset
1 /*
nick
parents:
diff changeset
2 * This code was taken from http://www.mpg123.org
nick
parents:
diff changeset
3 * See ChangeLog of mpg123-0.59s-pre.1 for detail
nick
parents:
diff changeset
4 * Applied to mplayer by Nick Kurshev <nickols_k@mail.ru>
nick
parents:
diff changeset
5 * Partial 3dnowex-DSP! optimization by Nick Kurshev
nick
parents:
diff changeset
6 *
nick
parents:
diff changeset
7 * TODO: optimize scalar 3dnow! code
nick
parents:
diff changeset
8 * Warning: Phases 7 & 8 are not tested
nick
parents:
diff changeset
9 */
nick
parents:
diff changeset
10 #define real float /* ugly - but only way */
nick
parents:
diff changeset
11
12292
114f3d149324 attribute_used for gcc3.4
alex
parents: 11240
diff changeset
12 #include "../config.h"
5291
421969d55d5f MANGLE() the 3dnow code
arpi
parents: 4148
diff changeset
13 #include "../mangle.h"
421969d55d5f MANGLE() the 3dnow code
arpi
parents: 4148
diff changeset
14
4148
nick
parents:
diff changeset
15 static unsigned long long int __attribute__((aligned(8))) x_plus_minus_3dnow = 0x8000000000000000ULL;
12292
114f3d149324 attribute_used for gcc3.4
alex
parents: 11240
diff changeset
16 static float attribute_used plus_1f = 1.0;
4148
nick
parents:
diff changeset
17
11240
5403367c7032 removed the stdcall convention
alex
parents: 9039
diff changeset
18 void dct64_MMX_3dnowex(real *a,real *b,real *c)
4148
nick
parents:
diff changeset
19 {
nick
parents:
diff changeset
20 char tmp[256];
nick
parents:
diff changeset
21 __asm __volatile(
nick
parents:
diff changeset
22 " movl %2,%%eax\n\t"
nick
parents:
diff changeset
23
nick
parents:
diff changeset
24 " leal 128+%3,%%edx\n\t"
nick
parents:
diff changeset
25 " movl %0,%%esi\n\t"
nick
parents:
diff changeset
26 " movl %1,%%edi\n\t"
5291
421969d55d5f MANGLE() the 3dnow code
arpi
parents: 4148
diff changeset
27 " movl $"MANGLE(costab_mmx)",%%ebx\n\t"
4148
nick
parents:
diff changeset
28 " leal %3,%%ecx\n\t"
nick
parents:
diff changeset
29
nick
parents:
diff changeset
30 /* Phase 1*/
nick
parents:
diff changeset
31 " movq (%%eax), %%mm0\n\t"
nick
parents:
diff changeset
32 " movq 8(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
33 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
34 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
35 " pshufw $78, 120(%%eax), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
36 " pshufw $78, 112(%%eax), %%mm5\n\t"
4148
nick
parents:
diff changeset
37 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
38 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
39 " movq %%mm0, (%%edx)\n\t"
nick
parents:
diff changeset
40 " movq %%mm4, 8(%%edx)\n\t"
nick
parents:
diff changeset
41 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
42 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
43 " pfmul (%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
44 " pfmul 8(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
45 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
46 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
47 " movq %%mm3, 120(%%edx)\n\t"
nick
parents:
diff changeset
48 " movq %%mm7, 112(%%edx)\n\t"
nick
parents:
diff changeset
49
nick
parents:
diff changeset
50 " movq 16(%%eax), %%mm0\n\t"
nick
parents:
diff changeset
51 " movq 24(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
52 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
53 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
54 " pshufw $78, 104(%%eax), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
55 " pshufw $78, 96(%%eax), %%mm5\n\t"
4148
nick
parents:
diff changeset
56 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
57 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
58 " movq %%mm0, 16(%%edx)\n\t"
nick
parents:
diff changeset
59 " movq %%mm4, 24(%%edx)\n\t"
nick
parents:
diff changeset
60 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
61 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
62 " pfmul 16(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
63 " pfmul 24(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
64 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
65 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
66 " movq %%mm3, 104(%%edx)\n\t"
nick
parents:
diff changeset
67 " movq %%mm7, 96(%%edx)\n\t"
nick
parents:
diff changeset
68
nick
parents:
diff changeset
69 " movq 32(%%eax), %%mm0\n\t"
nick
parents:
diff changeset
70 " movq 40(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
71 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
72 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
73 " pshufw $78, 88(%%eax), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
74 " pshufw $78, 80(%%eax), %%mm5\n\t"
4148
nick
parents:
diff changeset
75 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
76 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
77 " movq %%mm0, 32(%%edx)\n\t"
nick
parents:
diff changeset
78 " movq %%mm4, 40(%%edx)\n\t"
nick
parents:
diff changeset
79 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
80 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
81 " pfmul 32(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
82 " pfmul 40(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
83 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
84 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
85 " movq %%mm3, 88(%%edx)\n\t"
nick
parents:
diff changeset
86 " movq %%mm7, 80(%%edx)\n\t"
nick
parents:
diff changeset
87
nick
parents:
diff changeset
88 " movq 48(%%eax), %%mm0\n\t"
nick
parents:
diff changeset
89 " movq 56(%%eax), %%mm4\n\t"
nick
parents:
diff changeset
90 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
91 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
92 " pshufw $78, 72(%%eax), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
93 " pshufw $78, 64(%%eax), %%mm5\n\t"
4148
nick
parents:
diff changeset
94 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
95 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
96 " movq %%mm0, 48(%%edx)\n\t"
nick
parents:
diff changeset
97 " movq %%mm4, 56(%%edx)\n\t"
nick
parents:
diff changeset
98 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
99 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
100 " pfmul 48(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
101 " pfmul 56(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
102 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
103 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
104 " movq %%mm3, 72(%%edx)\n\t"
nick
parents:
diff changeset
105 " movq %%mm7, 64(%%edx)\n\t"
nick
parents:
diff changeset
106
nick
parents:
diff changeset
107 /* Phase 2*/
nick
parents:
diff changeset
108
nick
parents:
diff changeset
109 " movq (%%edx), %%mm0\n\t"
nick
parents:
diff changeset
110 " movq 8(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
111 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
112 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
113 " pshufw $78, 56(%%edx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
114 " pshufw $78, 48(%%edx), %%mm5\n\t"
4148
nick
parents:
diff changeset
115 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
116 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
117 " movq %%mm0, (%%ecx)\n\t"
nick
parents:
diff changeset
118 " movq %%mm4, 8(%%ecx)\n\t"
nick
parents:
diff changeset
119 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
120 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
121 " pfmul 64(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
122 " pfmul 72(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
123 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
124 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
125 " movq %%mm3, 56(%%ecx)\n\t"
nick
parents:
diff changeset
126 " movq %%mm7, 48(%%ecx)\n\t"
nick
parents:
diff changeset
127
nick
parents:
diff changeset
128 " movq 16(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
129 " movq 24(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
130 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
131 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
132 " pshufw $78, 40(%%edx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
133 " pshufw $78, 32(%%edx), %%mm5\n\t"
4148
nick
parents:
diff changeset
134 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
135 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
136 " movq %%mm0, 16(%%ecx)\n\t"
nick
parents:
diff changeset
137 " movq %%mm4, 24(%%ecx)\n\t"
nick
parents:
diff changeset
138 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
139 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
140 " pfmul 80(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
141 " pfmul 88(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
142 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
143 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
144 " movq %%mm3, 40(%%ecx)\n\t"
nick
parents:
diff changeset
145 " movq %%mm7, 32(%%ecx)\n\t"
nick
parents:
diff changeset
146
nick
parents:
diff changeset
147 /* Phase 3*/
nick
parents:
diff changeset
148
nick
parents:
diff changeset
149 " movq 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
150 " movq 72(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
151 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
152 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
153 " pshufw $78, 120(%%edx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
154 " pshufw $78, 112(%%edx), %%mm5\n\t"
4148
nick
parents:
diff changeset
155 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
156 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
157 " movq %%mm0, 64(%%ecx)\n\t"
nick
parents:
diff changeset
158 " movq %%mm4, 72(%%ecx)\n\t"
nick
parents:
diff changeset
159 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
160 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
161 " pfmul 64(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
162 " pfmul 72(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
163 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
164 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
165 " movq %%mm3, 120(%%ecx)\n\t"
nick
parents:
diff changeset
166 " movq %%mm7, 112(%%ecx)\n\t"
nick
parents:
diff changeset
167
nick
parents:
diff changeset
168 " movq 80(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
169 " movq 88(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
170 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
171 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
172 " pshufw $78, 104(%%edx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
173 " pshufw $78, 96(%%edx), %%mm5\n\t"
4148
nick
parents:
diff changeset
174 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
175 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
176 " movq %%mm0, 80(%%ecx)\n\t"
nick
parents:
diff changeset
177 " movq %%mm4, 88(%%ecx)\n\t"
nick
parents:
diff changeset
178 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
179 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
180 " pfmul 80(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
181 " pfmul 88(%%ebx), %%mm7\n\t"
nick
parents:
diff changeset
182 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
183 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
184 " movq %%mm3, 104(%%ecx)\n\t"
nick
parents:
diff changeset
185 " movq %%mm7, 96(%%ecx)\n\t"
nick
parents:
diff changeset
186
nick
parents:
diff changeset
187 /* Phase 4*/
nick
parents:
diff changeset
188
nick
parents:
diff changeset
189 " movq 96(%%ebx), %%mm2\n\t"
nick
parents:
diff changeset
190 " movq 104(%%ebx), %%mm6\n\t"
nick
parents:
diff changeset
191
nick
parents:
diff changeset
192 " movq (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
193 " movq 8(%%ecx), %%mm4\n\t"
nick
parents:
diff changeset
194 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
195 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
196 " pshufw $78, 24(%%ecx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
197 " pshufw $78, 16(%%ecx), %%mm5\n\t"
4148
nick
parents:
diff changeset
198 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
199 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
200 " movq %%mm0, (%%edx)\n\t"
nick
parents:
diff changeset
201 " movq %%mm4, 8(%%edx)\n\t"
nick
parents:
diff changeset
202 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
203 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
204 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
205 " pfmul %%mm6, %%mm7\n\t"
nick
parents:
diff changeset
206 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
207 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
208 " movq %%mm3, 24(%%edx)\n\t"
nick
parents:
diff changeset
209 " movq %%mm7, 16(%%edx)\n\t"
nick
parents:
diff changeset
210
nick
parents:
diff changeset
211 " movq 32(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
212 " movq 40(%%ecx), %%mm4\n\t"
nick
parents:
diff changeset
213 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
214 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
215 " pshufw $78, 56(%%ecx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
216 " pshufw $78, 48(%%ecx), %%mm5\n\t"
4148
nick
parents:
diff changeset
217 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
218 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
219 " movq %%mm0, 32(%%edx)\n\t"
nick
parents:
diff changeset
220 " movq %%mm4, 40(%%edx)\n\t"
nick
parents:
diff changeset
221 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
222 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
223 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
224 " pfmul %%mm6, %%mm7\n\t"
nick
parents:
diff changeset
225 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
226 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
227 " movq %%mm3, 56(%%edx)\n\t"
nick
parents:
diff changeset
228 " movq %%mm7, 48(%%edx)\n\t"
nick
parents:
diff changeset
229
nick
parents:
diff changeset
230 " movq 64(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
231 " movq 72(%%ecx), %%mm4\n\t"
nick
parents:
diff changeset
232 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
233 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
234 " pshufw $78, 88(%%ecx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
235 " pshufw $78, 80(%%ecx), %%mm5\n\t"
4148
nick
parents:
diff changeset
236 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
237 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
238 " movq %%mm0, 64(%%edx)\n\t"
nick
parents:
diff changeset
239 " movq %%mm4, 72(%%edx)\n\t"
nick
parents:
diff changeset
240 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
241 " pfsub %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
242 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
243 " pfmul %%mm6, %%mm7\n\t"
nick
parents:
diff changeset
244 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
245 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
246 " movq %%mm3, 88(%%edx)\n\t"
nick
parents:
diff changeset
247 " movq %%mm7, 80(%%edx)\n\t"
nick
parents:
diff changeset
248
nick
parents:
diff changeset
249 " movq 96(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
250 " movq 104(%%ecx), %%mm4\n\t"
nick
parents:
diff changeset
251 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
252 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
253 " pshufw $78, 120(%%ecx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
254 " pshufw $78, 112(%%ecx), %%mm5\n\t"
4148
nick
parents:
diff changeset
255 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
256 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
257 " movq %%mm0, 96(%%edx)\n\t"
nick
parents:
diff changeset
258 " movq %%mm4, 104(%%edx)\n\t"
nick
parents:
diff changeset
259 " pfsubr %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
260 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
261 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
262 " pfmul %%mm6, %%mm7\n\t"
nick
parents:
diff changeset
263 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
264 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
265 " movq %%mm3, 120(%%edx)\n\t"
nick
parents:
diff changeset
266 " movq %%mm7, 112(%%edx)\n\t"
nick
parents:
diff changeset
267
nick
parents:
diff changeset
268 /* Phase 5 */
nick
parents:
diff changeset
269
nick
parents:
diff changeset
270 " movq 112(%%ebx), %%mm2\n\t"
nick
parents:
diff changeset
271
nick
parents:
diff changeset
272 " movq (%%edx), %%mm0\n\t"
nick
parents:
diff changeset
273 " movq 16(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
274 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
275 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
276 " pshufw $78, 8(%%edx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
277 " pshufw $78, 24(%%edx), %%mm5\n\t"
4148
nick
parents:
diff changeset
278 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
279 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
280 " movq %%mm0, (%%ecx)\n\t"
nick
parents:
diff changeset
281 " movq %%mm4, 16(%%ecx)\n\t"
nick
parents:
diff changeset
282 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
283 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
284 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
285 " pfmul %%mm2, %%mm7\n\t"
nick
parents:
diff changeset
286 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
287 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
288 " movq %%mm3, 8(%%ecx)\n\t"
nick
parents:
diff changeset
289 " movq %%mm7, 24(%%ecx)\n\t"
nick
parents:
diff changeset
290
nick
parents:
diff changeset
291 " movq 32(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
292 " movq 48(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
293 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
294 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
295 " pshufw $78, 40(%%edx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
296 " pshufw $78, 56(%%edx), %%mm5\n\t"
4148
nick
parents:
diff changeset
297 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
298 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
299 " movq %%mm0, 32(%%ecx)\n\t"
nick
parents:
diff changeset
300 " movq %%mm4, 48(%%ecx)\n\t"
nick
parents:
diff changeset
301 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
302 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
303 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
304 " pfmul %%mm2, %%mm7\n\t"
nick
parents:
diff changeset
305 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
306 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
307 " movq %%mm3, 40(%%ecx)\n\t"
nick
parents:
diff changeset
308 " movq %%mm7, 56(%%ecx)\n\t"
nick
parents:
diff changeset
309
nick
parents:
diff changeset
310 " movq 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
311 " movq 80(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
312 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
313 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
314 " pshufw $78, 72(%%edx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
315 " pshufw $78, 88(%%edx), %%mm5\n\t"
4148
nick
parents:
diff changeset
316 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
317 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
318 " movq %%mm0, 64(%%ecx)\n\t"
nick
parents:
diff changeset
319 " movq %%mm4, 80(%%ecx)\n\t"
nick
parents:
diff changeset
320 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
321 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
322 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
323 " pfmul %%mm2, %%mm7\n\t"
nick
parents:
diff changeset
324 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
325 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
326 " movq %%mm3, 72(%%ecx)\n\t"
nick
parents:
diff changeset
327 " movq %%mm7, 88(%%ecx)\n\t"
nick
parents:
diff changeset
328
nick
parents:
diff changeset
329 " movq 96(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
330 " movq 112(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
331 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
332 " movq %%mm4, %%mm7\n\t"
7307
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
333 " pshufw $78, 104(%%edx), %%mm1\n\t"
7d20830716a1 Minor speedup by nick kurshev.
atmos4
parents: 6177
diff changeset
334 " pshufw $78, 120(%%edx), %%mm5\n\t"
4148
nick
parents:
diff changeset
335 " pfadd %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
336 " pfadd %%mm5, %%mm4\n\t"
nick
parents:
diff changeset
337 " movq %%mm0, 96(%%ecx)\n\t"
nick
parents:
diff changeset
338 " movq %%mm4, 112(%%ecx)\n\t"
nick
parents:
diff changeset
339 " pfsub %%mm1, %%mm3\n\t"
nick
parents:
diff changeset
340 " pfsubr %%mm5, %%mm7\n\t"
nick
parents:
diff changeset
341 " pfmul %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
342 " pfmul %%mm2, %%mm7\n\t"
nick
parents:
diff changeset
343 " pswapd %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
344 " pswapd %%mm7, %%mm7\n\t"
nick
parents:
diff changeset
345 " movq %%mm3, 104(%%ecx)\n\t"
nick
parents:
diff changeset
346 " movq %%mm7, 120(%%ecx)\n\t"
nick
parents:
diff changeset
347
nick
parents:
diff changeset
348
nick
parents:
diff changeset
349 /* Phase 6. This is the end of easy road. */
nick
parents:
diff changeset
350 /* Code below is coded in scalar mode. Should be optimized */
nick
parents:
diff changeset
351
5291
421969d55d5f MANGLE() the 3dnow code
arpi
parents: 4148
diff changeset
352 " movd "MANGLE(plus_1f)", %%mm6\n\t"
4148
nick
parents:
diff changeset
353 " punpckldq 120(%%ebx), %%mm6\n\t" /* mm6 = 1.0 | 120(%%ebx)*/
8236
7e2ca93330d0 forgotten mangle patch by (Bj«Órn Sandell <biorn at dce dot chalmers dot se>)
michael
parents: 7307
diff changeset
354 " movq "MANGLE(x_plus_minus_3dnow)", %%mm7\n\t" /* mm7 = +1 | -1 */
4148
nick
parents:
diff changeset
355
nick
parents:
diff changeset
356 " movq 32(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
357 " movq 64(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
358 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
359 " movq %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
360 " pxor %%mm7, %%mm1\n\t"
nick
parents:
diff changeset
361 " pxor %%mm7, %%mm3\n\t"
nick
parents:
diff changeset
362 " pfacc %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
363 " pfacc %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
364 " pfmul %%mm6, %%mm0\n\t"
nick
parents:
diff changeset
365 " pfmul %%mm6, %%mm2\n\t"
nick
parents:
diff changeset
366 " movq %%mm0, 32(%%edx)\n\t"
nick
parents:
diff changeset
367 " movq %%mm2, 64(%%edx)\n\t"
nick
parents:
diff changeset
368
nick
parents:
diff changeset
369 " movd 44(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
370 " movd 40(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
371 " movd 120(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
372 " punpckldq 76(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
373 " punpckldq 72(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
374 " punpckldq %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
375 " movq %%mm0, %%mm4\n\t"
nick
parents:
diff changeset
376 " movq %%mm2, %%mm5\n\t"
nick
parents:
diff changeset
377 " pfsub %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
378 " pfmul %%mm3, %%mm0\n\t"
nick
parents:
diff changeset
379 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
380 " pfadd %%mm5, %%mm0\n\t"
nick
parents:
diff changeset
381 " pfadd %%mm4, %%mm0\n\t"
nick
parents:
diff changeset
382 " movq %%mm0, %%mm2\n\t"
nick
parents:
diff changeset
383 " punpckldq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
384 " punpckhdq %%mm1, %%mm2\n\t"
nick
parents:
diff changeset
385 " movq %%mm0, 40(%%edx)\n\t"
nick
parents:
diff changeset
386 " movq %%mm2, 72(%%edx)\n\t"
nick
parents:
diff changeset
387
nick
parents:
diff changeset
388 " movd 48(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
389 " movd 60(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
390 " pfsub 52(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
391 " pfsub 56(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
392 " pfmul 120(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
393 " pfmul 120(%%ebx), %%mm2\n\t"
nick
parents:
diff changeset
394 " movq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
395
nick
parents:
diff changeset
396 " pfadd 56(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
397 " pfadd 60(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
398 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
399
nick
parents:
diff changeset
400 " pfadd 48(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
401 " pfadd 52(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
402 " pfadd %%mm3, %%mm1\n\t"
nick
parents:
diff changeset
403 " punpckldq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
404 " pfadd %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
405 " punpckldq %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
406 " movq %%mm1, 56(%%edx)\n\t"
nick
parents:
diff changeset
407 " movq %%mm0, 48(%%edx)\n\t"
nick
parents:
diff changeset
408
nick
parents:
diff changeset
409 /*---*/
nick
parents:
diff changeset
410
nick
parents:
diff changeset
411 " movd 92(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
412 " pfsub 88(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
413 " pfmul 120(%%ebx), %%mm1\n\t"
nick
parents:
diff changeset
414 " movd %%mm1, 92(%%edx)\n\t"
nick
parents:
diff changeset
415 " pfadd 92(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
416 " pfadd 88(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
417 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
418
nick
parents:
diff changeset
419 " pfadd 80(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
420 " pfadd 84(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
421 " movd %%mm0, 80(%%edx)\n\t"
nick
parents:
diff changeset
422
nick
parents:
diff changeset
423 " movd 80(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
424 " pfsub 84(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
425 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
426 " pfadd %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
427 " pfadd 92(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
428 " punpckldq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
429 " movq %%mm0, 84(%%edx)\n\t"
nick
parents:
diff changeset
430
nick
parents:
diff changeset
431 " movq 96(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
432 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
433 " pxor %%mm7, %%mm1\n\t"
nick
parents:
diff changeset
434 " pfacc %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
435 " pfmul %%mm6, %%mm0\n\t"
nick
parents:
diff changeset
436 " movq %%mm0, 96(%%edx)\n\t"
nick
parents:
diff changeset
437
nick
parents:
diff changeset
438 " movd 108(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
439 " pfsub 104(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
440 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
441 " movd %%mm0, 108(%%edx)\n\t"
nick
parents:
diff changeset
442 " pfadd 104(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
443 " pfadd 108(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
444 " movd %%mm0, 104(%%edx)\n\t"
nick
parents:
diff changeset
445
nick
parents:
diff changeset
446 " movd 124(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
447 " pfsub 120(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
448 " pfmul 120(%%ebx), %%mm1\n\t"
nick
parents:
diff changeset
449 " movd %%mm1, 124(%%edx)\n\t"
nick
parents:
diff changeset
450 " pfadd 120(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
451 " pfadd 124(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
452 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
453
nick
parents:
diff changeset
454 " pfadd 112(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
455 " pfadd 116(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
456 " movd %%mm0, 112(%%edx)\n\t"
nick
parents:
diff changeset
457
nick
parents:
diff changeset
458 " movd 112(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
459 " pfsub 116(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
460 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
461 " pfadd %%mm0,%%mm1\n\t"
nick
parents:
diff changeset
462 " pfadd 124(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
463 " punpckldq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
464 " movq %%mm0, 116(%%edx)\n\t"
nick
parents:
diff changeset
465
nick
parents:
diff changeset
466 " jnz .L01\n\t"
nick
parents:
diff changeset
467
nick
parents:
diff changeset
468 /* Phase 7*/
nick
parents:
diff changeset
469 /* Code below is coded in scalar mode. Should be optimized */
nick
parents:
diff changeset
470
nick
parents:
diff changeset
471 " movd (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
472 " pfadd 4(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
473 " movd %%mm0, 1024(%%esi)\n\t"
nick
parents:
diff changeset
474
nick
parents:
diff changeset
475 " movd (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
476 " pfsub 4(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
477 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
478 " movd %%mm0, (%%esi)\n\t"
nick
parents:
diff changeset
479 " movd %%mm0, (%%edi)\n\t"
nick
parents:
diff changeset
480
nick
parents:
diff changeset
481 " movd 12(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
482 " pfsub 8(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
483 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
484 " movd %%mm0, 512(%%edi)\n\t"
nick
parents:
diff changeset
485 " pfadd 12(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
486 " pfadd 8(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
487 " movd %%mm0, 512(%%esi)\n\t"
nick
parents:
diff changeset
488
nick
parents:
diff changeset
489 " movd 16(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
490 " pfsub 20(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
491 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
492 " movq %%mm0, %%mm3\n\t"
nick
parents:
diff changeset
493
nick
parents:
diff changeset
494 " movd 28(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
495 " pfsub 24(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
496 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
497 " movd %%mm0, 768(%%edi)\n\t"
nick
parents:
diff changeset
498 " movq %%mm0, %%mm2\n\t"
nick
parents:
diff changeset
499
nick
parents:
diff changeset
500 " pfadd 24(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
501 " pfadd 28(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
502 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
503
nick
parents:
diff changeset
504 " pfadd 16(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
505 " pfadd 20(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
506 " movd %%mm0, 768(%%esi)\n\t"
nick
parents:
diff changeset
507 " pfadd %%mm3, %%mm1\n\t"
nick
parents:
diff changeset
508 " movd %%mm1, 256(%%esi)\n\t"
nick
parents:
diff changeset
509 " pfadd %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
510 " movd %%mm2, 256(%%edi)\n\t"
nick
parents:
diff changeset
511
nick
parents:
diff changeset
512 /* Phase 8*/
nick
parents:
diff changeset
513
nick
parents:
diff changeset
514 " movq 32(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
515 " movq 48(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
516 " pfadd 48(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
517 " pfadd 40(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
518 " movd %%mm0, 896(%%esi)\n\t"
nick
parents:
diff changeset
519 " movd %%mm1, 640(%%esi)\n\t"
nick
parents:
diff changeset
520 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
521 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
522 " movd %%mm0, 128(%%edi)\n\t"
nick
parents:
diff changeset
523 " movd %%mm1, 384(%%edi)\n\t"
nick
parents:
diff changeset
524
nick
parents:
diff changeset
525 " movd 40(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
526 " pfadd 56(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
527 " movd %%mm0, 384(%%esi)\n\t"
nick
parents:
diff changeset
528
nick
parents:
diff changeset
529 " movd 56(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
530 " pfadd 36(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
531 " movd %%mm0, 128(%%esi)\n\t"
nick
parents:
diff changeset
532
nick
parents:
diff changeset
533 " movd 60(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
534 " movd %%mm0, 896(%%edi)\n\t"
nick
parents:
diff changeset
535 " pfadd 44(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
536 " movd %%mm0, 640(%%edi)\n\t"
nick
parents:
diff changeset
537
nick
parents:
diff changeset
538 " movq 96(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
539 " movq 112(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
540 " movq 104(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
541 " pfadd 112(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
542 " pfadd 104(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
543 " pfadd 120(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
544 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
545 " movq %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
546 " movq %%mm4, %%mm5\n\t"
nick
parents:
diff changeset
547 " pfadd 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
548 " pfadd 80(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
549 " pfadd 72(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
550 " movd %%mm0, 960(%%esi)\n\t"
nick
parents:
diff changeset
551 " movd %%mm2, 704(%%esi)\n\t"
nick
parents:
diff changeset
552 " movd %%mm4, 448(%%esi)\n\t"
nick
parents:
diff changeset
553 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
554 " psrlq $32, %%mm2\n\t"
nick
parents:
diff changeset
555 " psrlq $32, %%mm4\n\t"
nick
parents:
diff changeset
556 " movd %%mm0, 64(%%edi)\n\t"
nick
parents:
diff changeset
557 " movd %%mm2, 320(%%edi)\n\t"
nick
parents:
diff changeset
558 " movd %%mm4, 576(%%edi)\n\t"
nick
parents:
diff changeset
559 " pfadd 80(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
560 " pfadd 72(%%edx), %%mm3\n\t"
nick
parents:
diff changeset
561 " pfadd 88(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
562 " movd %%mm1, 832(%%esi)\n\t"
nick
parents:
diff changeset
563 " movd %%mm3, 576(%%esi)\n\t"
nick
parents:
diff changeset
564 " movd %%mm5, 320(%%esi)\n\t"
nick
parents:
diff changeset
565 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
566 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
567 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
568 " movd %%mm1, 192(%%edi)\n\t"
nick
parents:
diff changeset
569 " movd %%mm3, 448(%%edi)\n\t"
nick
parents:
diff changeset
570 " movd %%mm5, 704(%%edi)\n\t"
nick
parents:
diff changeset
571
nick
parents:
diff changeset
572 " movd 120(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
573 " pfadd 100(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
574 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
575 " pfadd 88(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
576 " movd %%mm0, 192(%%esi)\n\t"
nick
parents:
diff changeset
577 " pfadd 68(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
578 " movd %%mm1, 64(%%esi)\n\t"
nick
parents:
diff changeset
579
nick
parents:
diff changeset
580 " movd 124(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
581 " movd %%mm0, 960(%%edi)\n\t"
nick
parents:
diff changeset
582 " pfadd 92(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
583 " movd %%mm0, 832(%%edi)\n\t"
nick
parents:
diff changeset
584
nick
parents:
diff changeset
585 " jmp .L_bye\n\t"
nick
parents:
diff changeset
586 ".L01: \n\t"
nick
parents:
diff changeset
587 /* Phase 9*/
nick
parents:
diff changeset
588
nick
parents:
diff changeset
589 " movq (%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
590 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
591 " pxor %%mm7, %%mm1\n\t"
nick
parents:
diff changeset
592 " pfacc %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
593 " pfmul %%mm6, %%mm0\n\t"
nick
parents:
diff changeset
594 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
595 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
596 " movw %%ax, 512(%%esi)\n\t"
nick
parents:
diff changeset
597 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
598 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
599 " movw %%ax, (%%esi)\n\t"
nick
parents:
diff changeset
600
nick
parents:
diff changeset
601 " movd 12(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
602 " pfsub 8(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
603 " pfmul 120(%%ebx), %%mm0\n\t"
nick
parents:
diff changeset
604 " pf2id %%mm0, %%mm7\n\t"
nick
parents:
diff changeset
605 " movd %%mm7, %%eax\n\t"
nick
parents:
diff changeset
606 " movw %%ax, 256(%%edi)\n\t"
nick
parents:
diff changeset
607 " pfadd 12(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
608 " pfadd 8(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
609 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
610 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
611 " movw %%ax, 256(%%esi)\n\t"
nick
parents:
diff changeset
612
nick
parents:
diff changeset
613 " movd 16(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
614 " pfsub 20(%%ecx), %%mm3\n\t"
nick
parents:
diff changeset
615 " pfmul 120(%%ebx), %%mm3\n\t"
nick
parents:
diff changeset
616 " movq %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
617
nick
parents:
diff changeset
618 " movd 28(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
619 " pfsub 24(%%ecx), %%mm2\n\t"
nick
parents:
diff changeset
620 " pfmul 120(%%ebx), %%mm2\n\t"
nick
parents:
diff changeset
621 " movq %%mm2, %%mm1\n\t"
nick
parents:
diff changeset
622
nick
parents:
diff changeset
623 " pf2id %%mm2, %%mm7\n\t"
nick
parents:
diff changeset
624 " movd %%mm7, %%eax\n\t"
nick
parents:
diff changeset
625 " movw %%ax, 384(%%edi)\n\t"
nick
parents:
diff changeset
626
nick
parents:
diff changeset
627 " pfadd 24(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
628 " pfadd 28(%%ecx), %%mm1\n\t"
nick
parents:
diff changeset
629 " movq %%mm1, %%mm0\n\t"
nick
parents:
diff changeset
630
nick
parents:
diff changeset
631 " pfadd 16(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
632 " pfadd 20(%%ecx), %%mm0\n\t"
nick
parents:
diff changeset
633 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
634 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
635 " movw %%ax, 384(%%esi)\n\t"
nick
parents:
diff changeset
636 " pfadd %%mm3, %%mm1\n\t"
nick
parents:
diff changeset
637 " pf2id %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
638 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
639 " movw %%ax, 128(%%esi)\n\t"
nick
parents:
diff changeset
640 " pfadd %%mm3, %%mm2\n\t"
nick
parents:
diff changeset
641 " pf2id %%mm2, %%mm2\n\t"
nick
parents:
diff changeset
642 " movd %%mm2, %%eax\n\t"
nick
parents:
diff changeset
643 " movw %%ax, 128(%%edi)\n\t"
nick
parents:
diff changeset
644
nick
parents:
diff changeset
645 /* Phase 10*/
nick
parents:
diff changeset
646
nick
parents:
diff changeset
647 " movq 32(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
648 " movq 48(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
649 " pfadd 48(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
650 " pfadd 40(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
651 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
652 " pf2id %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
653 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
654 " movd %%mm1, %%ecx\n\t"
nick
parents:
diff changeset
655 " movw %%ax, 448(%%esi)\n\t"
nick
parents:
diff changeset
656 " movw %%cx, 320(%%esi)\n\t"
nick
parents:
diff changeset
657 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
658 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
659 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
660 " movd %%mm1, %%ecx\n\t"
nick
parents:
diff changeset
661 " movw %%ax, 64(%%edi)\n\t"
nick
parents:
diff changeset
662 " movw %%cx, 192(%%edi)\n\t"
nick
parents:
diff changeset
663
nick
parents:
diff changeset
664 " movd 40(%%edx), %%mm3\n\t"
nick
parents:
diff changeset
665 " movd 56(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
666 " movd 60(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
667 " movd 44(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
668 " movd 120(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
669 " punpckldq %%mm4, %%mm3\n\t"
nick
parents:
diff changeset
670 " punpckldq 124(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
671 " pfadd 100(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
672 " punpckldq 36(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
673 " punpckldq 92(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
674 " movq %%mm5, %%mm6\n\t"
nick
parents:
diff changeset
675 " pfadd %%mm4, %%mm3\n\t"
nick
parents:
diff changeset
676 " pf2id %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
677 " pf2id %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
678 " pfadd 88(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
679 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
680 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
681 " movw %%ax, 448(%%edi)\n\t"
nick
parents:
diff changeset
682 " movw %%cx, 192(%%esi)\n\t"
nick
parents:
diff changeset
683 " pf2id %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
684 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
685 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
686 " movd %%mm5, %%ebx\n\t"
nick
parents:
diff changeset
687 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
688 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
689 " movw %%bx, 96(%%esi)\n\t"
nick
parents:
diff changeset
690 " movw %%ax, 480(%%edi)\n\t"
nick
parents:
diff changeset
691 " movw %%cx, 64(%%esi)\n\t"
nick
parents:
diff changeset
692 " pfadd %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
693 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
694 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
695 " pfadd 68(%%edx), %%mm6\n\t"
nick
parents:
diff changeset
696 " movw %%ax, 320(%%edi)\n\t"
nick
parents:
diff changeset
697 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
698 " pf2id %%mm6, %%mm6\n\t"
nick
parents:
diff changeset
699 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
700 " movd %%mm6, %%ebx\n\t"
nick
parents:
diff changeset
701 " movw %%ax, 416(%%edi)\n\t"
nick
parents:
diff changeset
702 " movw %%bx, 32(%%esi)\n\t"
nick
parents:
diff changeset
703
nick
parents:
diff changeset
704 " movq 96(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
705 " movq 112(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
706 " movq 104(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
707 " pfadd %%mm2, %%mm0\n\t"
nick
parents:
diff changeset
708 " pfadd %%mm4, %%mm2\n\t"
nick
parents:
diff changeset
709 " pfadd 120(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
710 " movq %%mm0, %%mm1\n\t"
nick
parents:
diff changeset
711 " movq %%mm2, %%mm3\n\t"
nick
parents:
diff changeset
712 " movq %%mm4, %%mm5\n\t"
nick
parents:
diff changeset
713 " pfadd 64(%%edx), %%mm0\n\t"
nick
parents:
diff changeset
714 " pfadd 80(%%edx), %%mm2\n\t"
nick
parents:
diff changeset
715 " pfadd 72(%%edx), %%mm4\n\t"
nick
parents:
diff changeset
716 " pf2id %%mm0, %%mm0\n\t"
nick
parents:
diff changeset
717 " pf2id %%mm2, %%mm2\n\t"
nick
parents:
diff changeset
718 " pf2id %%mm4, %%mm4\n\t"
nick
parents:
diff changeset
719 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
720 " movd %%mm2, %%ecx\n\t"
nick
parents:
diff changeset
721 " movd %%mm4, %%ebx\n\t"
nick
parents:
diff changeset
722 " movw %%ax, 480(%%esi)\n\t"
nick
parents:
diff changeset
723 " movw %%cx, 352(%%esi)\n\t"
nick
parents:
diff changeset
724 " movw %%bx, 224(%%esi)\n\t"
nick
parents:
diff changeset
725 " psrlq $32, %%mm0\n\t"
nick
parents:
diff changeset
726 " psrlq $32, %%mm2\n\t"
nick
parents:
diff changeset
727 " psrlq $32, %%mm4\n\t"
nick
parents:
diff changeset
728 " movd %%mm0, %%eax\n\t"
nick
parents:
diff changeset
729 " movd %%mm2, %%ecx\n\t"
nick
parents:
diff changeset
730 " movd %%mm4, %%ebx\n\t"
nick
parents:
diff changeset
731 " movw %%ax, 32(%%edi)\n\t"
nick
parents:
diff changeset
732 " movw %%cx, 160(%%edi)\n\t"
nick
parents:
diff changeset
733 " movw %%bx, 288(%%edi)\n\t"
nick
parents:
diff changeset
734 " pfadd 80(%%edx), %%mm1\n\t"
nick
parents:
diff changeset
735 " pfadd 72(%%edx), %%mm3\n\t"
nick
parents:
diff changeset
736 " pfadd 88(%%edx), %%mm5\n\t"
nick
parents:
diff changeset
737 " pf2id %%mm1, %%mm1\n\t"
nick
parents:
diff changeset
738 " pf2id %%mm3, %%mm3\n\t"
nick
parents:
diff changeset
739 " pf2id %%mm5, %%mm5\n\t"
nick
parents:
diff changeset
740 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
741 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
742 " movd %%mm5, %%ebx\n\t"
nick
parents:
diff changeset
743 " movw %%ax, 416(%%esi)\n\t"
nick
parents:
diff changeset
744 " movw %%cx, 288(%%esi)\n\t"
nick
parents:
diff changeset
745 " movw %%bx, 160(%%esi)\n\t"
nick
parents:
diff changeset
746 " psrlq $32, %%mm1\n\t"
nick
parents:
diff changeset
747 " psrlq $32, %%mm3\n\t"
nick
parents:
diff changeset
748 " psrlq $32, %%mm5\n\t"
nick
parents:
diff changeset
749 " movd %%mm1, %%eax\n\t"
nick
parents:
diff changeset
750 " movd %%mm3, %%ecx\n\t"
nick
parents:
diff changeset
751 " movd %%mm5, %%ebx\n\t"
nick
parents:
diff changeset
752 " movw %%ax, 96(%%edi)\n\t"
nick
parents:
diff changeset
753 " movw %%cx, 224(%%edi)\n\t"
nick
parents:
diff changeset
754 " movw %%bx, 352(%%edi)\n\t"
nick
parents:
diff changeset
755
nick
parents:
diff changeset
756 " movsw\n\t"
nick
parents:
diff changeset
757
nick
parents:
diff changeset
758 ".L_bye:\n\t"
nick
parents:
diff changeset
759 " femms\n\t"
nick
parents:
diff changeset
760 :
nick
parents:
diff changeset
761 :"m"(a),"m"(b),"m"(c),"m"(tmp[0])
9039
b615c698e994 Propolice catches an overflow in dct64_k7.c:
arpi
parents: 8236
diff changeset
762 :"memory","%eax","%ebx","%ecx","%edx","%esi","%edi");
5842
d6eab895c742 Avoid stdcall on cygwin, it causes undefined ref, code needs testing as I have no athlon or k6-2.
atmos4
parents: 5291
diff changeset
763 }