Subversion Repositories shark

Rev

Details | Last modification | View Log | RSS feed

Rev Author Line No. Line
70 giacomo 1
/* $Id: 3dnow_xform3.s,v 1.1 2003-03-13 12:11:48 giacomo 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
#include "matypes.h"
28
#include "xform_args.h"
29
 
30
    SEG_TEXT
31
 
32
#define FRAME_OFFSET	4
33
 
34
 
35
ALIGNTEXT16
36
GLOBL GLNAME( _mesa_3dnow_transform_points3_general )
37
GLNAME( _mesa_3dnow_transform_points3_general ):
38
 
39
    PUSH_L    ( ESI )
40
 
41
    MOV_L     ( ARG_DEST, ECX )
42
    MOV_L     ( ARG_MATRIX, ESI )
43
    MOV_L     ( ARG_SOURCE, EAX )
44
    MOV_L     ( CONST(4), REGOFF(V4F_SIZE, ECX) )
45
    OR_B      ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
46
    MOV_L     ( REGOFF(V4F_COUNT, EAX), EDX )
47
    MOV_L     ( EDX, REGOFF(V4F_COUNT, ECX) )
48
 
49
    PUSH_L    ( EDI )
50
 
51
    MOV_L     ( REGOFF(V4F_START, ECX), EDX )
52
    MOV_L     ( ESI, ECX )
53
    MOV_L     ( REGOFF(V4F_COUNT, EAX), ESI )
54
    MOV_L     ( REGOFF(V4F_STRIDE, EAX), EDI )
55
    MOV_L     ( REGOFF(V4F_START, EAX), EAX )
56
 
57
    TEST_L    ( ESI, ESI )
58
    JZ        ( LLBL( G3TPGR_2 ) )
59
 
60
    PREFETCHW ( REGIND(EDX) )
61
 
62
ALIGNTEXT16
63
LLBL( G3TPGR_1 ):
64
 
65
    PREFETCHW ( REGOFF(32, EDX) )	/* prefetch 2 vertices ahead         */
66
 
67
    MOVQ      ( REGIND(EAX), MM0 )	/* x1              | x0              */
68
    MOVD      ( REGOFF(8, EAX), MM2 )	/*                 | x2              */
69
 
70
    ADD_L     ( EDI, EAX )		/* next vertex                       */
71
    PREFETCH  ( REGIND(EAX) )
72
 
73
    MOVQ      ( MM0, MM1 )		/* x1              | x0              */
74
    PUNPCKLDQ ( MM2, MM2 )		/* x2              | x2              */
75
 
76
    PUNPCKLDQ ( MM0, MM0 )		/* x0              | x0              */
77
    MOVQ      ( MM2, MM5 )		/* x2              | x2              */
78
 
79
    PUNPCKHDQ ( MM1, MM1 )		/* x1              | x1              */
80
    PFMUL     ( REGOFF(32, ECX), MM2 )	/* x2*m9           | x2*m8           */
81
 
82
    MOVQ      ( MM0, MM3 )		/* x0              | x0              */
83
    PFMUL     ( REGOFF(40, ECX), MM5 )	/* x2*m11          | x2*m10          */
84
 
85
    MOVQ      ( MM1, MM4 )		/* x1              | x1              */
86
    PFMUL     ( REGIND(ECX), MM0 )	/* x0*m1           | x0*m0           */
87
 
88
    PFADD     ( REGOFF(48, ECX), MM2 )	/* x2*m9+m13       | x2*m8+m12       */
89
    PFMUL     ( REGOFF(16, ECX), MM1 )	/* x1*m5           | x1*m4           */
90
 
91
    PFADD     ( REGOFF(56, ECX), MM5 )	/* x2*m11+m15      | x2*m10+m14      */
92
    PFADD     ( MM0, MM1 )		/* x0*m1+x1*m5     | x0*m0+x1*m4     */
93
 
94
    PFMUL     ( REGOFF(8, ECX), MM3 )	/* x0*m3           | x0*m2           */
95
    PFADD     ( MM1, MM2 )		/* r1              | r0              */
96
 
97
    PFMUL     ( REGOFF(24, ECX), MM4 )	/* x1*m7           | x1*m6           */
98
    ADD_L     ( CONST(16), EDX )	/* next output vertex                */
99
 
100
    PFADD     ( MM3, MM4 )		/* x0*m3+x1*m7     | x0*m2+x1*m6     */
101
    MOVQ      ( MM2, REGOFF(-16, EDX) )	/* write r0, r1                      */
102
 
103
    PFADD     ( MM4, MM5 )		/* r3              | r2              */
104
    MOVQ      ( MM5, REGOFF(-8, EDX) )	/* write r2, r3                      */
105
 
106
    DEC_L     ( ESI )			/* decrement vertex counter          */
107
    JNZ       ( LLBL( G3TPGR_1 ) )	/* cnt > 0 ? -> process next vertex  */
108
 
109
LLBL( G3TPGR_2 ):
110
 
111
    FEMMS
112
    POP_L     ( EDI )
113
    POP_L     ( ESI )
114
    RET
115
 
116
 
117
 
118
 
119
ALIGNTEXT16
120
GLOBL GLNAME( _mesa_3dnow_transform_points3_perspective )
121
GLNAME( _mesa_3dnow_transform_points3_perspective ):
122
 
123
    PUSH_L    ( ESI )
124
 
125
    MOV_L     ( ARG_DEST, ECX )
126
    MOV_L     ( ARG_MATRIX, ESI )
127
    MOV_L     ( ARG_SOURCE, EAX )
128
    MOV_L     ( CONST(4), REGOFF(V4F_SIZE, ECX) )
129
    OR_B      ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
130
    MOV_L     ( REGOFF(V4F_COUNT, EAX), EDX )
131
    MOV_L     ( EDX, REGOFF(V4F_COUNT, ECX) )
132
 
133
    PUSH_L    ( EDI )
134
 
135
    MOV_L     ( REGOFF(V4F_START, ECX), EDX )
136
    MOV_L     ( ESI, ECX )
137
    MOV_L     ( REGOFF(V4F_COUNT, EAX), ESI )
138
    MOV_L     ( REGOFF(V4F_STRIDE, EAX), EDI )
139
    MOV_L     ( REGOFF(V4F_START, EAX), EAX )
140
 
141
    TEST_L    ( ESI, ESI )
142
    JZ        ( LLBL( G3TPPR_2 ) )
143
 
144
    PREFETCH  ( REGIND(EAX) )
145
    PREFETCHW ( REGIND(EDX) )
146
 
147
    MOVD      ( REGIND(ECX), MM0 )	/*                 | m00             */
148
    PUNPCKLDQ ( REGOFF(20, ECX), MM0 )	/* m11             | m00             */
149
 
150
    MOVQ      ( REGOFF(32, ECX), MM1 )	/* m21             | m20             */
151
    MOVD      ( REGOFF(40, ECX), MM2 )	/*                 | m22             */
152
 
153
    MOVD      ( REGOFF(56, ECX), MM3 )	/*                 | m32             */
154
 
155
ALIGNTEXT16
156
LLBL( G3TPPR_1 ):
157
 
158
    PREFETCHW ( REGOFF(32, EDX) )	/* prefetch 2 vertices ahead         */
159
 
160
    MOVD      ( REGOFF(8, EAX), MM5 )	/*                 | x2              */
161
    MOVQ      ( REGIND(EAX), MM4 )	/* x1              | x0              */
162
 
163
    ADD_L     ( EDI, EAX )		/* next vertex                       */
164
    PREFETCH  ( REGIND(EAX) )
165
 
166
    PXOR      ( MM7, MM7 )		/* 0               | 0               */
167
    MOVQ      ( MM5, MM6 )		/*                 | x2              */
168
 
169
    PFMUL     ( MM0, MM4 )		/* x1*m11          | x0*m00          */
170
    PFSUB     ( MM5, MM7 )		/*                 | -x2             */
171
 
172
    PFMUL     ( MM2, MM6 )		/*                 | x2*m22          */
173
    PUNPCKLDQ ( MM5, MM5 )		/* x2              | x2              */
174
 
175
    ADD_L     ( CONST(16), EDX )	/* next r                            */
176
    PFMUL     ( MM1, MM5 )		/* x2*m21          | x2*m20          */
177
 
178
    PFADD     ( MM3, MM6 )		/*                 | x2*m22+m32      */
179
    PFADD     ( MM4, MM5 )		/* x1*m11+x2*m21   | x0*m00+x2*m20   */
180
 
181
    MOVQ      ( MM5, REGOFF(-16, EDX) )	/* write r0, r1                      */
182
    MOVD      ( MM6, REGOFF(-8, EDX) )	/* write r2                          */
183
 
184
    MOVD      ( MM7, REGOFF(-4, EDX) )	/* write r3                          */
185
 
186
    DEC_L     ( ESI )			/* decrement vertex counter          */
187
    JNZ       ( LLBL( G3TPPR_1 ) )	/* cnt > 0 ? -> process next vertex  */
188
 
189
LLBL( G3TPPR_2 ):
190
 
191
    FEMMS
192
    POP_L     ( EDI )
193
    POP_L     ( ESI )
194
    RET
195
 
196
 
197
 
198
 
199
ALIGNTEXT16
200
GLOBL GLNAME( _mesa_3dnow_transform_points3_3d )
201
GLNAME( _mesa_3dnow_transform_points3_3d ):
202
 
203
    PUSH_L    ( ESI )
204
 
205
    MOV_L     ( ARG_DEST, ECX )
206
    MOV_L     ( ARG_MATRIX, ESI )
207
    MOV_L     ( ARG_SOURCE, EAX )
208
    MOV_L     ( CONST(3), REGOFF(V4F_SIZE, ECX) )
209
    OR_B      ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
210
    MOV_L     ( REGOFF(V4F_COUNT, EAX), EDX )
211
    MOV_L     ( EDX, REGOFF(V4F_COUNT, ECX) )
212
 
213
    PUSH_L    ( EDI )
214
 
215
    MOV_L     ( REGOFF(V4F_START, ECX), EDX )
216
    MOV_L     ( ESI, ECX )
217
    MOV_L     ( REGOFF(V4F_COUNT, EAX), ESI )
218
    MOV_L     ( REGOFF(V4F_STRIDE, EAX), EDI )
219
    MOV_L     ( REGOFF(V4F_START, EAX), EAX )
220
 
221
    TEST_L    ( ESI, ESI )
222
    JZ        ( LLBL( G3TP3R_2 ) )
223
 
224
    PREFETCH  ( REGIND(EAX) )
225
    PREFETCH  ( REGIND(EDX) )
226
 
227
    MOVD      ( REGOFF(8, ECX), MM7 )	/*                 | m2              */
228
    PUNPCKLDQ ( REGOFF(24, ECX), MM7 )	/* m6              | m2              */
229
 
230
 
231
ALIGNTEXT16
232
LLBL( G3TP3R_1 ):
233
 
234
    PREFETCHW ( REGOFF(32, EDX) )	/* prefetch 2 vertices ahead         */
235
 
236
    MOVQ      ( REGIND(EAX), MM0 )	/* x1              | x0              */
237
    MOVD      ( REGOFF(8, EAX), MM1 )	/*                 | x2              */
238
 
239
    ADD_L     ( EDI, EAX )		/* next vertex                       */
240
    PREFETCH  ( REGIND(EAX) )
241
 
242
    MOVQ      ( MM0, MM2 )		/* x1              | x0              */
243
    ADD_L     ( CONST(16), EDX )	/* next r                            */
244
 
245
    PUNPCKLDQ ( MM2, MM2 )		/* x0              | x0              */
246
    MOVQ      ( MM0, MM3 )		/* x1              | x0              */
247
 
248
    PFMUL     ( REGIND(ECX), MM2 )	/* x0*m1           | x0*m0           */
249
    PUNPCKHDQ ( MM3, MM3 )		/* x1              | x1              */
250
 
251
    MOVQ      ( MM1, MM4 )		/*                 | x2              */
252
    PFMUL     ( REGOFF(16, ECX), MM3 )	/* x1*m5           | x1*m4           */
253
 
254
    PUNPCKLDQ ( MM4, MM4 )		/* x2              | x2              */
255
    PFADD     ( MM2, MM3 )		/* x0*m1+x1*m5     | x0*m0+x1*m4     */
256
 
257
    PFMUL     ( REGOFF(32, ECX), MM4 )	/* x2*m9           | x2*m8           */
258
    PFADD     ( REGOFF(48, ECX), MM3 )	/* x0*m1+...+m11   | x0*m0+x1*m4+m12 */
259
 
260
    PFMUL     ( MM7, MM0 )		/* x1*m6           | x0*m2           */
261
    PFADD     ( MM4, MM3 )		/* r1              | r0              */
262
 
263
    PFMUL     ( REGOFF(40, ECX), MM1 )	/*                 | x2*m10          */
264
    PUNPCKLDQ ( REGOFF(56, ECX), MM1 )	/* m14             | x2*m10          */
265
 
266
    PFACC     ( MM0, MM1 )
267
 
268
    MOVQ      ( MM3, REGOFF(-16, EDX) )	/* write r0, r1                      */
269
    PFACC     ( MM1, MM1 )		/*                 | r2              */
270
 
271
    MOVD      ( MM1, REGOFF(-8, EDX) )	/* write r2                          */
272
 
273
    DEC_L     ( ESI )			/* decrement vertex counter          */
274
    JNZ       ( LLBL( G3TP3R_1 ) )	/* cnt > 0 ? -> process next vertex  */
275
 
276
LLBL( G3TP3R_2 ):
277
 
278
    FEMMS
279
    POP_L     ( EDI )
280
    POP_L     ( ESI )
281
    RET
282
 
283
 
284
 
285
 
286
ALIGNTEXT16
287
GLOBL GLNAME( _mesa_3dnow_transform_points3_3d_no_rot )
288
GLNAME( _mesa_3dnow_transform_points3_3d_no_rot ):
289
 
290
    PUSH_L    ( ESI )
291
 
292
    MOV_L     ( ARG_DEST, ECX )
293
    MOV_L     ( ARG_MATRIX, ESI )
294
    MOV_L     ( ARG_SOURCE, EAX )
295
    MOV_L     ( CONST(3), REGOFF(V4F_SIZE, ECX) )
296
    OR_B      ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
297
    MOV_L     ( REGOFF(V4F_COUNT, EAX), EDX )
298
    MOV_L     ( EDX, REGOFF(V4F_COUNT, ECX) )
299
 
300
    PUSH_L    ( EDI )
301
 
302
    MOV_L     ( REGOFF(V4F_START, ECX), EDX )
303
    MOV_L     ( ESI, ECX )
304
    MOV_L     ( REGOFF(V4F_COUNT, EAX), ESI )
305
    MOV_L     ( REGOFF(V4F_STRIDE, EAX), EDI )
306
    MOV_L     ( REGOFF(V4F_START, EAX), EAX )
307
 
308
    TEST_L    ( ESI, ESI )
309
    JZ        ( LLBL( G3TP3NRR_2 ) )
310
 
311
    PREFETCH  ( REGIND(EAX) )
312
    PREFETCHW ( REGIND(EDX) )
313
 
314
    MOVD      ( REGIND(ECX), MM0 )	/*                 | m00             */
315
    PUNPCKLDQ ( REGOFF(20, ECX), MM0 )	/* m11             | m00             */
316
 
317
    MOVD      ( REGOFF(40, ECX), MM2 )	/*                 | m22             */
318
    PUNPCKLDQ ( MM2, MM2 )		/* m22             | m22             */
319
 
320
    MOVQ      ( REGOFF(48, ECX), MM1 )	/* m31             | m30             */
321
    MOVD      ( REGOFF(56, ECX), MM3 )	/*                 | m32             */
322
 
323
    PUNPCKLDQ ( MM3, MM3 )		/* m32             | m32             */
324
 
325
 
326
ALIGNTEXT16
327
LLBL( G3TP3NRR_1 ):
328
 
329
    PREFETCHW ( REGOFF(32, EDX) )	/* prefetch 2 vertices ahead         */
330
 
331
    MOVQ      ( REGIND(EAX), MM4 )	/* x1              | x0              */
332
    MOVD      ( REGOFF(8, EAX), MM5 )	/*                 | x2              */
333
 
334
    ADD_L     ( EDI, EAX )		/* next vertex                       */
335
    PREFETCHW ( REGIND(EAX) )
336
 
337
    PFMUL     ( MM0, MM4 )		/* x1*m11          | x0*m00          */
338
 
339
    PFADD     ( MM1, MM4 )		/* x1*m11+m31      | x0*m00+m30      */
340
    PFMUL     ( MM2, MM5 )		/*                 | x2*m22          */
341
 
342
    PFADD     ( MM3, MM5 )		/*                 | x2*m22+m32      */
343
    MOVQ      ( MM4, REGIND(EDX) )	/* write r0, r1                      */
344
 
345
    ADD_L     ( CONST(16), EDX )	/* next r                            */
346
    DEC_L     ( ESI )			/* decrement vertex counter          */
347
 
348
    MOVD      ( MM5, REGOFF(-8, EDX) )	/* write r2                          */
349
    JNZ       ( LLBL( G3TP3NRR_1 ) )	/* cnt > 0 ? -> process next vertex  */
350
 
351
LLBL( G3TP3NRR_2 ):
352
 
353
    FEMMS
354
    POP_L     ( EDI )
355
    POP_L     ( ESI )
356
    RET
357
 
358
 
359
 
360
 
361
ALIGNTEXT16
362
GLOBL GLNAME( _mesa_3dnow_transform_points3_2d )
363
GLNAME( _mesa_3dnow_transform_points3_2d ):
364
 
365
    PUSH_L    ( ESI )
366
 
367
    MOV_L     ( ARG_DEST, ECX )
368
    MOV_L     ( ARG_MATRIX, ESI )
369
    MOV_L     ( ARG_SOURCE, EAX )
370
    MOV_L     ( CONST(3), REGOFF(V4F_SIZE, ECX) )
371
    OR_B      ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
372
    MOV_L     ( REGOFF(V4F_COUNT, EAX), EDX )
373
    MOV_L     ( EDX, REGOFF(V4F_COUNT, ECX) )
374
 
375
    PUSH_L    ( EDI )
376
 
377
    MOV_L     ( REGOFF(V4F_START, ECX), EDX )
378
    MOV_L     ( ESI, ECX )
379
    MOV_L     ( REGOFF(V4F_COUNT, EAX), ESI )
380
    MOV_L     ( REGOFF(V4F_STRIDE, EAX), EDI )
381
    MOV_L     ( REGOFF(V4F_START, EAX), EAX )
382
 
383
    TEST_L    ( ESI, ESI )
384
    JZ        ( LLBL( G3TP2R_3) )
385
 
386
    PREFETCH  ( REGIND(EAX) )
387
    PREFETCHW ( REGIND(EDX) )
388
 
389
    MOVD      ( REGIND(ECX), MM0 )	/*                 | m00             */
390
    PUNPCKLDQ ( REGOFF(16, ECX), MM0 )	/* m10             | m00             */
391
 
392
    MOVD      ( REGOFF(4, ECX), MM1 )	/*                 | m01             */
393
    PUNPCKLDQ ( REGOFF(20, ECX), MM1 )	/* m11             | m01             */
394
 
395
    MOVQ      ( REGOFF(48, ECX), MM2 )	/* m31             | m30             */
396
 
397
ALIGNTEXT16
398
LLBL( G3TP2R_2 ):
399
 
400
    PREFETCHW ( REGOFF(32, EDX) )	/* prefetch 2 vertices ahead         */
401
 
402
    MOVQ      ( REGIND(EAX), MM3 )	/* x1              | x0              */
403
    MOVD      ( REGOFF(8, EAX), MM5 )	/*                 | x2              */
404
 
405
    ADD_L     ( EDI, EAX )		/* next vertex                       */
406
    PREFETCH  ( REGIND(EAX) )
407
 
408
    MOVQ      ( MM3, MM4 )		/* x1              | x0              */
409
    PFMUL     ( MM0, MM3 )		/* x1*m10          | x0*m00          */
410
 
411
    ADD_L     ( CONST(16), EDX )	/* next r                            */
412
    PFMUL     ( MM1, MM4 )		/* x1*m11          | x0*m01          */
413
 
414
    PFACC     ( MM4, MM3 )		/* x0*m00+x1*m10   | x0*m01+x1*m11   */
415
    MOVD      ( MM5, REGOFF(-8, EDX) )	/* write r2 (=x2)                    */
416
 
417
    PFADD     ( MM2, MM3 )		/* x0*...*m10+m30  | x0*...*m11+m31  */
418
    MOVQ      ( MM3, REGOFF(-16, EDX) )	/* write r0, r1                      */
419
 
420
    DEC_L     ( ESI )			/* decrement vertex counter          */
421
    JNZ       ( LLBL( G3TP2R_2 ) )	/* cnt > 0 ? -> process next vertex  */
422
 
423
LLBL( G3TP2R_3 ):
424
 
425
    FEMMS
426
    POP_L     ( EDI )
427
    POP_L     ( ESI )
428
    RET
429
 
430
 
431
 
432
 
433
ALIGNTEXT16
434
GLOBL GLNAME( _mesa_3dnow_transform_points3_2d_no_rot )
435
GLNAME( _mesa_3dnow_transform_points3_2d_no_rot ):
436
 
437
    PUSH_L    ( ESI )
438
 
439
    MOV_L     ( ARG_DEST, ECX )
440
    MOV_L     ( ARG_MATRIX, ESI )
441
    MOV_L     ( ARG_SOURCE, EAX )
442
    MOV_L     ( CONST(3), REGOFF(V4F_SIZE, ECX) )
443
    OR_B      ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
444
    MOV_L     ( REGOFF(V4F_COUNT, EAX), EDX )
445
    MOV_L     ( EDX, REGOFF(V4F_COUNT, ECX) )
446
 
447
    PUSH_L    ( EDI )
448
 
449
    MOV_L     ( REGOFF(V4F_START, ECX), EDX )
450
    MOV_L     ( ESI, ECX )
451
    MOV_L     ( REGOFF(V4F_COUNT, EAX), ESI )
452
    MOV_L     ( REGOFF(V4F_STRIDE, EAX), EDI )
453
    MOV_L     ( REGOFF(V4F_START, EAX), EAX )
454
 
455
    TEST_L    ( ESI, ESI )
456
    JZ        ( LLBL( G3TP2NRR_2 ) )
457
 
458
    PREFETCH  ( REGIND(EAX) )
459
    PREFETCHW ( REGIND(EDX) )
460
 
461
    MOVD      ( REGIND(ECX), MM0 )	/*                 | m00             */
462
    PUNPCKLDQ ( REGOFF(20, ECX), MM0 )	/* m11             | m00             */
463
 
464
    MOVQ      ( REGOFF(48, ECX), MM1 )	/* m31             | m30             */
465
 
466
 
467
ALIGNTEXT16
468
LLBL( G3TP2NRR_1 ):
469
 
470
    PREFETCHW ( REGOFF(32, EDX) )	/* prefetch 2 vertices ahead         */
471
 
472
    MOVQ      ( REGIND(EAX), MM4 )	/* x1              | x0              */
473
    MOVD      ( REGOFF(8, EAX), MM5 )	/*                 | x2              */
474
 
475
    ADD_L     ( EDI, EAX )		/* next vertex                       */
476
    PREFETCH  ( REGIND(EAX) )
477
 
478
    PFMUL     ( MM0, MM4 )		/* x1*m11          | x0*m00          */
479
    ADD_L     ( CONST(16), EDX )	/* next r                            */
480
 
481
    PFADD     ( MM1, MM4 )		/* x1*m11+m31      | x0*m00+m30      */
482
 
483
    MOVQ      ( MM4, REGOFF(-16, EDX) )	/* write r0, r1                      */
484
    MOVD      ( MM5, REGOFF(-8, EDX) )	/* write r2 (=x2)                    */
485
 
486
    DEC_L     ( ESI )			/* decrement vertex counter          */
487
    JNZ       ( LLBL( G3TP2NRR_1 ) )	/* cnt > 0 ? -> process next vertex  */
488
 
489
LLBL( G3TP2NRR_2 ):
490
 
491
    FEMMS
492
    POP_L     ( EDI )
493
    POP_L     ( ESI )
494
    RET
495
 
496
 
497
 
498
 
499
ALIGNTEXT16
500
GLOBL GLNAME( _mesa_3dnow_transform_points3_identity )
501
GLNAME( _mesa_3dnow_transform_points3_identity ):
502
 
503
    PUSH_L    ( ESI )
504
 
505
    MOV_L     ( ARG_DEST, ECX )
506
    MOV_L     ( ARG_MATRIX, ESI )
507
    MOV_L     ( ARG_SOURCE, EAX )
508
    MOV_L     ( CONST(3), REGOFF(V4F_SIZE, ECX) )
509
    OR_B      ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
510
    MOV_L     ( REGOFF(V4F_COUNT, EAX), EDX )
511
    MOV_L     ( EDX, REGOFF(V4F_COUNT, ECX) )
512
 
513
    PUSH_L    ( EDI )
514
 
515
    MOV_L     ( REGOFF(V4F_START, ECX), EDX )
516
    MOV_L     ( ESI, ECX )
517
    MOV_L     ( REGOFF(V4F_COUNT, EAX), ESI )
518
    MOV_L     ( REGOFF(V4F_STRIDE, EAX), EDI )
519
    MOV_L     ( REGOFF(V4F_START, EAX), EAX )
520
 
521
    TEST_L    ( ESI, ESI )
522
    JZ        ( LLBL( G3TPIR_2 ) )
523
 
524
    PREFETCHW ( REGIND(EDX) )
525
 
526
ALIGNTEXT16
527
LLBL( G3TPIR_1 ):
528
 
529
    PREFETCHW ( REGOFF(32, EDX) )
530
 
531
    MOVQ      ( REGIND(EAX), MM0 )	/* x1              | x0              */
532
    MOVD      ( REGOFF(8, EAX), MM1 )	/*                 | x2              */
533
 
534
    ADD_L     ( EDI, EAX )		/* next vertex                       */
535
    ADD_L     ( CONST(16), EDX )	/* next r                            */
536
 
537
    DEC_L     ( ESI )			/* decrement vertex counter          */
538
    MOVQ      ( MM0, REGOFF(-16, EDX) )	/* r1              | r0              */
539
 
540
    MOVD      ( MM1, REGOFF(-8, EDX) )	/*                 | r2              */
541
    JNZ       ( LLBL( G3TPIR_1 ) )	/* cnt > 0 ? -> process next vertex  */
542
 
543
LLBL( G3TPIR_2 ):
544
 
545
    FEMMS
546
    POP_L     ( EDI )
547
    POP_L     ( ESI )
548
    RET