bring over build fixes from stable branch
[mesa.git] / src / mesa / x86 / 3dnow_xform1.S
1 /* $Id: 3dnow_xform1.S,v 1.2 2004/04/26 10:10:25 alanh Exp $ */
2
3 /*
4 * Mesa 3-D graphics library
5 * Version: 3.5
6 *
7 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
8 *
9 * Permission is hereby granted, free of charge, to any person obtaining a
10 * copy of this software and associated documentation files (the "Software"),
11 * to deal in the Software without restriction, including without limitation
12 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
13 * and/or sell copies of the Software, and to permit persons to whom the
14 * Software is furnished to do so, subject to the following conditions:
15 *
16 * The above copyright notice and this permission notice shall be included
17 * in all copies or substantial portions of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
23 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
24 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 */
26
27 #ifdef USE_3DNOW_ASM
28 #include "matypes.h"
29 #include "xform_args.h"
30
31 SEG_TEXT
32
33 #define FRAME_OFFSET 4
34
35
36 ALIGNTEXT16
37 GLOBL GLNAME( _mesa_3dnow_transform_points1_general )
38 GLNAME( _mesa_3dnow_transform_points1_general ):
39
40 PUSH_L ( ESI )
41
42 MOV_L ( ARG_DEST, ECX )
43 MOV_L ( ARG_MATRIX, ESI )
44 MOV_L ( ARG_SOURCE, EAX )
45 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
46 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
47 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
48 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
49
50 PUSH_L ( EDI )
51
52 MOV_L ( REGOFF(4, ECX), EDX )
53 MOV_L ( ESI, ECX )
54 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
55 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
56 MOV_L ( REGOFF(V4F_START, EAX), EAX )
57
58 TEST_L ( ESI, ESI )
59 JZ ( LLBL( G3TPGR_3 ) )
60
61 MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */
62 MOVQ ( REGOFF(8, ECX), MM1 ) /* m03 | m02 */
63
64 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
65 MOVQ ( REGOFF(56, ECX), MM3 ) /* m33 | m32 */
66
67 ALIGNTEXT16
68 LLBL( G3TPGR_2 ):
69
70 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
71 PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */
72
73 MOVQ ( MM4, MM5 ) /* x0 | x0 */
74 PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */
75
76 PFMUL ( MM1, MM5 ) /* x0*m03 | x0*m02 */
77 PFADD ( MM2, MM4 ) /* x0*m01+m31 | x0*m00+m30 */
78
79 PFADD ( MM3, MM5 ) /* x0*m03+m33 | x0*m02+m32 */
80 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
81
82 MOVQ ( MM5, REGOFF(8, EDX) ) /* write r3, r2 */
83 ADD_L ( EDI, EAX ) /* next vertex */
84
85 ADD_L ( CONST(16), EDX ) /* next r */
86 DEC_L ( ESI ) /* decrement vertex counter */
87
88 JNZ ( LLBL( G3TPGR_2 ) ) /* cnt > 0 ? -> process next vertex */
89
90 LLBL( G3TPGR_3 ):
91
92 FEMMS
93 POP_L ( EDI )
94 POP_L ( ESI )
95 RET
96
97
98
99
100 ALIGNTEXT16
101 GLOBL GLNAME( _mesa_3dnow_transform_points1_identity )
102 GLNAME( _mesa_3dnow_transform_points1_identity ):
103
104 PUSH_L ( ESI )
105
106 MOV_L ( ARG_DEST, ECX )
107 MOV_L ( ARG_MATRIX, ESI )
108 MOV_L ( ARG_SOURCE, EAX )
109 MOV_L ( CONST(1), REGOFF(V4F_SIZE, ECX) )
110 OR_B ( CONST(VEC_SIZE_1), REGOFF(V4F_FLAGS, ECX) )
111 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
112 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
113
114 PUSH_L ( EDI )
115
116 MOV_L ( REGOFF(4, ECX), EDX )
117 MOV_L ( ESI, ECX )
118 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
119 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
120 MOV_L ( REGOFF(V4F_START, EAX), EAX )
121
122 TEST_L ( ESI, ESI )
123 JZ ( LLBL( G3TPIR_4) )
124
125 ALIGNTEXT16
126 LLBL( G3TPIR_3 ):
127
128 MOVD ( REGIND(EAX), MM0 ) /* | x0 */
129 ADD_L ( EDI, EAX ) /* next vertex */
130
131 MOVD ( MM0, REGIND(EDX) ) /* | r0 */
132 ADD_L ( CONST(16), EDX ) /* next r */
133
134 DEC_L ( ESI ) /* decrement vertex counter */
135 JNZ ( LLBL( G3TPIR_3 ) ) /* cnt > 0 ? -> process next vertex */
136
137 LLBL( G3TPIR_4 ):
138
139 FEMMS
140 POP_L ( EDI )
141 POP_L ( ESI )
142 RET
143
144
145
146
147 ALIGNTEXT16
148 GLOBL GLNAME( _mesa_3dnow_transform_points1_3d_no_rot )
149 GLNAME( _mesa_3dnow_transform_points1_3d_no_rot ):
150
151 PUSH_L ( ESI )
152
153 MOV_L ( ARG_DEST, ECX )
154 MOV_L ( ARG_MATRIX, ESI )
155 MOV_L ( ARG_SOURCE, EAX )
156 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
157 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
158 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
159 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
160
161 PUSH_L ( EDI )
162
163 MOV_L ( REGOFF(4, ECX), EDX )
164 MOV_L ( ESI, ECX )
165 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
166 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
167 MOV_L ( REGOFF(V4F_START, EAX), EAX )
168
169 TEST_L ( ESI, ESI )
170 JZ ( LLBL( G3TP3NRR_3 ) )
171
172 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
173 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
174
175 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
176
177 ALIGNTEXT16
178 LLBL( G3TP3NRR_2 ):
179
180 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
181 PFMUL ( MM0, MM4 ) /* | x0*m00 */
182
183 PFADD ( MM2, MM4 ) /* m31 | x0*m00+m30 */
184 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
185
186 MOVD ( MM3, REGOFF(8, EDX) ) /* write r2 */
187 ADD_L ( EDI, EAX ) /* next vertex */
188
189 ADD_L ( CONST(16), EDX ) /* next r */
190 DEC_L ( ESI ) /* decrement vertex counter */
191
192 JNZ ( LLBL( G3TP3NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
193
194 LLBL( G3TP3NRR_3 ):
195
196 FEMMS
197 POP_L ( EDI )
198 POP_L ( ESI )
199 RET
200
201
202
203
204 ALIGNTEXT16
205 GLOBL GLNAME( _mesa_3dnow_transform_points1_perspective )
206 GLNAME( _mesa_3dnow_transform_points1_perspective ):
207
208 PUSH_L ( ESI )
209
210 MOV_L ( ARG_DEST, ECX )
211 MOV_L ( ARG_MATRIX, ESI )
212 MOV_L ( ARG_SOURCE, EAX )
213 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
214 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
215 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
216 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
217
218 PUSH_L ( EDI )
219
220 MOV_L ( REGOFF(4, ECX), EDX )
221 MOV_L ( ESI, ECX )
222 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
223 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
224 MOV_L ( REGOFF(V4F_START, EAX), EAX )
225
226 TEST_L ( ESI, ESI )
227 JZ ( LLBL( G3TPPR_3 ) )
228
229 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
230 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
231
232 ALIGNTEXT16
233 LLBL( G3TPPR_2 ):
234
235 MOVD ( REGIND(EAX), MM4 ) /* 0 | x0 */
236 PFMUL ( MM0, MM4 ) /* 0 | x0*m00 */
237
238 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
239 MOVQ ( MM3, REGOFF(8, EDX) ) /* write r2 (=m32), r3 (=0) */
240
241 ADD_L ( EDI, EAX ) /* next vertex */
242 ADD_L ( CONST(16), EDX ) /* next r */
243
244 DEC_L ( ESI ) /* decrement vertex counter */
245 JNZ ( LLBL( G3TPPR_2 ) ) /* cnt > 0 ? -> process next vertex */
246
247 LLBL( G3TPPR_3 ):
248
249 FEMMS
250 POP_L ( EDI )
251 POP_L ( ESI )
252 RET
253
254
255
256
257 ALIGNTEXT16
258 GLOBL GLNAME( _mesa_3dnow_transform_points1_2d )
259 GLNAME( _mesa_3dnow_transform_points1_2d ):
260
261 PUSH_L ( ESI )
262
263 MOV_L ( ARG_DEST, ECX )
264 MOV_L ( ARG_MATRIX, ESI )
265 MOV_L ( ARG_SOURCE, EAX )
266 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
267 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
268 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
269 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
270
271 PUSH_L ( EDI )
272
273 MOV_L ( REGOFF(4, ECX), EDX )
274 MOV_L ( ESI, ECX )
275 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
276 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
277 MOV_L ( REGOFF(V4F_START, EAX), EAX )
278
279 TEST_L ( ESI, ESI )
280 JZ ( LLBL( G3TP2R_3 ) )
281
282 MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */
283 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
284
285 ALIGNTEXT16
286 LLBL( G3TP2R_2 ):
287
288 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
289 PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */
290
291 PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */
292 PFADD ( MM2, MM4 ) /* x0*m01+m31 | x0*m00+m30 */
293
294 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
295 ADD_L ( EDI, EAX ) /* next vertex */
296
297 ADD_L ( CONST(16), EDX ) /* next r */
298 DEC_L ( ESI ) /* decrement vertex counter */
299
300 JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */
301
302 LLBL( G3TP2R_3 ):
303
304 FEMMS
305 POP_L ( EDI )
306 POP_L ( ESI )
307 RET
308
309
310
311
312 ALIGNTEXT16
313 GLOBL GLNAME( _mesa_3dnow_transform_points1_2d_no_rot )
314 GLNAME( _mesa_3dnow_transform_points1_2d_no_rot ):
315
316 PUSH_L ( ESI )
317
318 MOV_L ( ARG_DEST, ECX )
319 MOV_L ( ARG_MATRIX, ESI )
320 MOV_L ( ARG_SOURCE, EAX )
321 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
322 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
323 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
324 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
325
326 PUSH_L ( EDI )
327
328 MOV_L ( REGOFF(4, ECX), EDX )
329 MOV_L ( ESI, ECX )
330 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
331 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
332 MOV_L ( REGOFF(V4F_START, EAX), EAX )
333
334 TEST_L ( ESI, ESI )
335 JZ ( LLBL( G3TP2NRR_3 ) )
336
337 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
338 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
339
340 ALIGNTEXT16
341 LLBL( G3TP2NRR_2 ):
342
343 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
344 ADD_L ( EDI, EAX ) /* next vertex */
345
346 PFMUL ( MM0, MM4 ) /* | x0*m00 */
347 PFADD ( MM2, MM4 ) /* m31 | x0*m00+m30 */
348
349 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
350 ADD_L ( CONST(16), EDX ) /* next r */
351
352 DEC_L ( ESI ) /* decrement vertex counter */
353 JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
354
355 LLBL( G3TP2NRR_3 ):
356
357 FEMMS
358 POP_L ( EDI )
359 POP_L ( ESI )
360 RET
361
362
363
364
365 ALIGNTEXT16
366 GLOBL GLNAME( _mesa_3dnow_transform_points1_3d )
367 GLNAME( _mesa_3dnow_transform_points1_3d ):
368
369 PUSH_L ( ESI )
370
371 MOV_L ( ARG_DEST, ECX )
372 MOV_L ( ARG_MATRIX, ESI )
373 MOV_L ( ARG_SOURCE, EAX )
374 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
375 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
376 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
377 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
378
379 PUSH_L ( EDI )
380
381 MOV_L ( REGOFF(4, ECX), EDX )
382 MOV_L ( ESI, ECX )
383 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
384 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
385 MOV_L ( REGOFF(V4F_START, EAX), EAX )
386
387 TEST_L ( ESI, ESI )
388 JZ ( LLBL( G3TP3R_3 ) )
389
390 MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */
391 MOVD ( REGOFF(8, ECX), MM1 ) /* | m02 */
392
393 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
394 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
395
396 ALIGNTEXT16
397 LLBL( G3TP3R_2 ):
398
399 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
400 PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */
401
402 MOVQ ( MM4, MM5 ) /* | x0 */
403 PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */
404
405 PFMUL ( MM1, MM5 ) /* | x0*m02 */
406 PFADD ( MM2, MM4 ) /* x0*m01+m31 | x0*m00+m30 */
407
408 PFADD ( MM3, MM5 ) /* | x0*m02+m32 */
409 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
410
411 MOVD ( MM5, REGOFF(8, EDX) ) /* write r2 */
412 ADD_L ( EDI, EAX ) /* next vertex */
413
414 ADD_L ( CONST(16), EDX ) /* next r */
415 DEC_L ( ESI ) /* decrement vertex counter */
416
417 JNZ ( LLBL( G3TP3R_2 ) ) /* cnt > 0 ? -> process next vertex */
418
419 LLBL( G3TP3R_3 ):
420
421 FEMMS
422 POP_L ( EDI )
423 POP_L ( ESI )
424 RET
425
426 #endif