syntax fixes for Solaris (David Dawes)
[mesa.git] / src / mesa / x86 / sse_xform3.S
1 /* $Id: sse_xform3.S,v 1.1 2001/03/29 06:46:27 gareth 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 /** TODO:
28 * - insert PREFETCH instructions to avoid cache-misses !
29 * - some more optimizations are possible...
30 * - for 40-50% more performance in the SSE-functions, the
31 * data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned !
32 */
33
34 #include "matypes.h"
35 #include "xform_args.h"
36
37 SEG_TEXT
38
39 #define S(i) REGOFF(i * 4, ESI)
40 #define D(i) REGOFF(i * 4, EDI)
41 #define M(i) REGOFF(i * 4, EDX)
42
43
44 ALIGNTEXT4
45 GLOBL GLNAME(_mesa_sse_transform_points3_general)
46 GLNAME( _mesa_sse_transform_points3_general ):
47
48 #define FRAME_OFFSET 8
49 PUSH_L ( ESI )
50 PUSH_L ( EDI )
51
52 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
53 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
54
55 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
56 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
57
58 CMP_L ( CONST(0), ECX ) /* count == 0 ? */
59 JE ( LLBL(K_GTPGR_finish) ) /* yes -> nothing to do. */
60
61 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
62 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
63
64 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
65 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
66
67 SHL_L( CONST(4), ECX ) /* count *= 16 */
68 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
69
70 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
71 ADD_L( EDI, ECX ) /* count += dest ptr */
72
73
74 ALIGNTEXT32
75 MOVAPS ( REGOFF(0, EDX), XMM0 ) /* m0 | m1 | m2 | m3 */
76 MOVAPS ( REGOFF(16, EDX), XMM1 ) /* m4 | m5 | m6 | m7 */
77 MOVAPS ( REGOFF(32, EDX), XMM2 ) /* m8 | m9 | m10 | m11 */
78 MOVAPS ( REGOFF(48, EDX), XMM3 ) /* m12 | m13 | m14 | m15 */
79
80
81 ALIGNTEXT32
82 LLBL(K_GTPGR_top):
83 MOVSS ( REGOFF(0, ESI), XMM4 ) /* | | | ox */
84 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox | ox */
85 MOVSS ( REGOFF(4, ESI), XMM5 ) /* | | | oy */
86 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy | oy */
87 MOVSS ( REGOFF(8, ESI), XMM6 ) /* | | | oz */
88 SHUFPS ( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz | oz */
89
90 MULPS ( XMM0, XMM4 ) /* m3*ox | m2*ox | m1*ox | m0*ox */
91 MULPS ( XMM1, XMM5 ) /* m7*oy | m6*oy | m5*oy | m4*oy */
92 MULPS ( XMM2, XMM6 ) /* m11*oz | m10*oz | m9*oz | m8*oz */
93
94 ADDPS ( XMM5, XMM4 )
95 ADDPS ( XMM6, XMM4 )
96 ADDPS ( XMM3, XMM4 )
97
98 MOVAPS ( XMM4, REGOFF(0, EDI) )
99
100 LLBL(K_GTPGR_skip):
101 ADD_L ( CONST(16), EDI )
102 ADD_L ( EAX, ESI )
103 CMP_L ( ECX, EDI )
104 JNE ( LLBL(K_GTPGR_top) )
105
106 LLBL(K_GTPGR_finish):
107 POP_L ( EDI )
108 POP_L ( ESI )
109 RET
110 #undef FRAME_OFFSET
111
112
113 ALIGNTEXT4
114 GLOBL GLNAME(_mesa_sse_transform_points3_identity)
115 GLNAME( _mesa_sse_transform_points3_identity ):
116
117 #define FRAME_OFFSET 8
118 PUSH_L ( ESI )
119 PUSH_L ( EDI )
120
121 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
122 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
123
124 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
125
126 TEST_L( ECX, ECX)
127 JZ( LLBL(K_GTPIR_finish) ) /* count was zero; go to finish */
128
129 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
130 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
131
132 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
133 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
134
135 SHL_L( CONST(4), ECX ) /* count *= 16 */
136 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
137
138 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
139 ADD_L( EDI, ECX ) /* count += dest ptr */
140
141 CMP_L( ESI, EDI )
142 JE( LLBL(K_GTPIR_finish) )
143
144
145 ALIGNTEXT32
146 LLBL(K_GTPIR_top):
147 MOVLPS ( S(0), XMM0 )
148 MOVLPS ( XMM0, D(0) )
149 MOVSS ( S(2), XMM0 )
150 MOVSS ( XMM0, D(2) )
151
152 LLBL(K_GTPIR_skip):
153 ADD_L ( CONST(16), EDI )
154 ADD_L ( EAX, ESI )
155 CMP_L ( ECX, EDI )
156 JNE ( LLBL(K_GTPIR_top) )
157
158 LLBL(K_GTPIR_finish):
159 POP_L ( EDI )
160 POP_L ( ESI )
161 RET
162 #undef FRAME_OFFSET
163
164
165
166
167 ALIGNTEXT4
168 GLOBL GLNAME(_mesa_sse_transform_points3_3d_no_rot)
169 GLNAME(_mesa_sse_transform_points3_3d_no_rot):
170
171 #define FRAME_OFFSET 8
172 PUSH_L( ESI )
173 PUSH_L( EDI )
174
175 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
176 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
177
178
179 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
180 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
181
182 TEST_L( ECX, ECX)
183 JZ( LLBL(K_GTP3DNRR_finish) ) /* count was zero; go to finish */
184
185 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
186 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
187
188 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
189 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
190
191 SHL_L( CONST(4), ECX ) /* count *= 16 */
192 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
193
194 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
195 ADD_L( EDI, ECX ) /* count += dest ptr */
196
197
198 ALIGNTEXT32
199 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
200 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
201 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
202 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */
203 MOVSS ( M(10), XMM3 ) /* - | - | - | m10 */
204 MOVSS ( M(14), XMM4 ) /* - | - | - | m14 */
205
206 ALIGNTEXT32
207 LLBL(K_GTP3DNRR_top):
208
209 MOVLPS ( S(0), XMM0 ) /* - | - | s1 | s0 */
210 MULPS ( XMM1, XMM0 ) /* - | - | s1*m5 | s0*m0 */
211 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */
212 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */
213
214 MOVSS ( S(2), XMM0 ) /* sz */
215 MULSS ( XMM3, XMM0 ) /* sz*m10 */
216 ADDSS ( XMM4, XMM0 ) /* +m14 */
217 MOVSS ( XMM0, D(2) ) /* -> D(2) */
218
219 LLBL(K_GTP3DNRR_skip):
220 ADD_L ( CONST(16), EDI )
221 ADD_L ( EAX, ESI )
222 CMP_L ( ECX, EDI )
223 JNE ( LLBL(K_GTP3DNRR_top) )
224
225 LLBL(K_GTP3DNRR_finish):
226 POP_L ( EDI )
227 POP_L ( ESI )
228 RET
229 #undef FRAME_OFFSET
230
231
232
233 ALIGNTEXT4
234 GLOBL GLNAME(_mesa_sse_transform_points3_perspective)
235 GLNAME(_mesa_sse_transform_points3_perspective):
236
237 #define FRAME_OFFSET 8
238 PUSH_L ( ESI )
239 PUSH_L ( EDI )
240
241 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
242 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
243
244 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
245 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
246
247 TEST_L( ECX, ECX)
248 JZ( LLBL(K_GTP3PR_finish) ) /* count was zero; go to finish */
249
250 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
251 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
252
253 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
254 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
255
256 SHL_L( CONST(4), ECX ) /* count *= 16 */
257 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
258
259 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
260 ADD_L( EDI, ECX ) /* count += dest ptr */
261
262 ALIGNTEXT32
263 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
264 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
265 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
266 MOVLPS ( M(8), XMM2 ) /* - | - | m9 | m8 */
267 MOVSS ( M(10), XMM3 ) /* m10 */
268 MOVSS ( M(14), XMM4 ) /* m14 */
269 XORPS ( XMM6, XMM6 ) /* 0 */
270
271 ALIGNTEXT32
272 LLBL(K_GTP3PR_top):
273 MOVLPS ( S(0), XMM0 ) /* oy | ox */
274 MULPS ( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
275 MOVSS ( S(2), XMM5 ) /* oz */
276 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oz | oz */
277 MULPS ( XMM2, XMM5 ) /* oz*m9 | oz*m8 */
278 ADDPS ( XMM5, XMM0 ) /* +oy*m5 | +ox*m0 */
279 MOVLPS ( XMM0, D(0) ) /* ->D(1) | ->D(0) */
280
281 MOVSS ( S(2), XMM0 ) /* oz */
282 MULSS ( XMM3, XMM0 ) /* oz*m10 */
283 ADDSS ( XMM4, XMM0 ) /* +m14 */
284 MOVSS ( XMM0, D(2) ) /* ->D(2) */
285
286 MOVSS ( S(2), XMM0 ) /* oz */
287 MOVSS ( XMM6, XMM5 ) /* 0 */
288 SUBPS ( XMM0, XMM5 ) /* -oz */
289 MOVSS ( XMM5, D(3) ) /* ->D(3) */
290
291 LLBL(K_GTP3PR_skip):
292 ADD_L( CONST(16), EDI )
293 ADD_L( EAX, ESI )
294 CMP_L( ECX, EDI )
295 JNE( LLBL(K_GTP3PR_top) )
296
297 LLBL(K_GTP3PR_finish):
298 POP_L ( EDI )
299 POP_L ( ESI )
300 RET
301 #undef FRAME_OFFSET
302
303
304
305 ALIGNTEXT4
306 GLOBL GLNAME(_mesa_sse_transform_points3_2d)
307 GLNAME(_mesa_sse_transform_points3_2d):
308
309 #define FRAME_OFFSET 8
310 PUSH_L( ESI )
311 PUSH_L( EDI )
312
313 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
314 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
315
316 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
317 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
318
319 TEST_L( ECX, ECX)
320 JZ( LLBL(K_GTP3P2DR_finish) ) /* count was zero; go to finish */
321
322 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
323 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
324
325 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
326 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
327
328 SHL_L( CONST(4), ECX ) /* count *= 16 */
329 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
330
331 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
332 ADD_L( EDI, ECX ) /* count += dest ptr */
333
334 ALIGNTEXT32
335 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
336 MOVLPS( M(4), XMM1 ) /* m5 | m4 */
337 MOVLPS( M(12), XMM2 ) /* m13 | m12 */
338
339 ALIGNTEXT32
340 LLBL(K_GTP3P2DR_top):
341 MOVSS ( S(0), XMM3 ) /* ox */
342 SHUFPS ( CONST(0x0), XMM3, XMM3 ) /* ox | ox */
343 MULPS ( XMM0, XMM3 ) /* ox*m1 | ox*m0 */
344 MOVSS ( S(1), XMM4 ) /* oy */
345 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* oy | oy */
346 MULPS ( XMM1, XMM4 ) /* oy*m5 | oy*m4 */
347
348 ADDPS ( XMM4, XMM3 )
349 ADDPS ( XMM2, XMM3 )
350 MOVLPS ( XMM3, D(0) )
351
352 MOVSS ( S(2), XMM3 )
353 MOVSS ( XMM3, D(2) )
354
355 LLBL(K_GTP3P2DR_skip):
356 ADD_L ( CONST(16), EDI )
357 ADD_L ( EAX, ESI )
358 CMP_L ( ECX, EDI )
359 JNE ( LLBL(K_GTP3P2DR_top) )
360
361 LLBL(K_GTP3P2DR_finish):
362 POP_L ( EDI )
363 POP_L ( ESI )
364 RET
365 #undef FRAME_OFFSET
366
367
368
369 ALIGNTEXT4
370 GLOBL GLNAME(_mesa_sse_transform_points3_2d_no_rot)
371 GLNAME(_mesa_sse_transform_points3_2d_no_rot):
372
373 #define FRAME_OFFSET 8
374 PUSH_L( ESI )
375 PUSH_L( EDI )
376
377 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
378 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
379
380 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
381 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
382
383 TEST_L( ECX, ECX)
384 JZ( LLBL(K_GTP3P2DNRR_finish) ) /* count was zero; go to finish */
385
386 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
387 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
388
389 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
390 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
391
392 SHL_L( CONST(4), ECX ) /* count *= 16 */
393 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
394
395 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
396 ADD_L( EDI, ECX ) /* count += dest ptr */
397
398 ALIGNTEXT32
399 MOVSS ( M(0), XMM1 ) /* m0 */
400 MOVSS ( M(5), XMM2 ) /* m5 */
401 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */
402 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */
403
404 ALIGNTEXT32
405 LLBL(K_GTP3P2DNRR_top):
406 MOVLPS( S(0), XMM0 ) /* oy | ox */
407 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
408 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */
409 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */
410
411 MOVSS( S(2), XMM0 )
412 MOVSS( XMM0, D(2) )
413
414 LLBL(K_GTP3P2DNRR_skip):
415 ADD_L( CONST(16), EDI )
416 ADD_L( EAX, ESI )
417 CMP_L( ECX, EDI )
418 JNE( LLBL(K_GTP3P2DNRR_top) )
419
420 LLBL(K_GTP3P2DNRR_finish):
421 POP_L( EDI )
422 POP_L( ESI )
423 RET
424 #undef FRAME_OFFSET
425
426
427
428
429 ALIGNTEXT4
430 GLOBL GLNAME(_mesa_sse_transform_points3_3d)
431 GLNAME(_mesa_sse_transform_points3_3d):
432
433 #define FRAME_OFFSET 8
434 PUSH_L( ESI )
435 PUSH_L( EDI )
436
437 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
438 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
439
440
441 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
442 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
443
444 TEST_L( ECX, ECX)
445 JZ( LLBL(K_GTP3P3DR_finish) ) /* count was zero; go to finish */
446
447 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
448 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
449
450 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
451 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
452
453 SHL_L( CONST(4), ECX ) /* count *= 16 */
454 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
455
456 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
457 ADD_L( EDI, ECX ) /* count += dest ptr */
458
459
460 ALIGNTEXT32
461 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */
462 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */
463 MOVAPS( M(8), XMM2 ) /* m10 | m9 | m8 */
464 MOVAPS( M(12), XMM3 ) /* m14 | m13 | m12 */
465
466 ALIGNTEXT32
467 LLBL(K_GTP3P3DR_top):
468 MOVSS( S(0), XMM4 )
469 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox */
470 MULPS( XMM0, XMM4 ) /* ox*m2 | ox*m1 | ox*m0 */
471
472 MOVSS( S(1), XMM5 )
473 SHUFPS( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy */
474 MULPS( XMM1, XMM5 ) /* oy*m6 | oy*m5 | oy*m4 */
475
476 MOVSS( S(2), XMM6 )
477 SHUFPS( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz */
478 MULPS( XMM2, XMM6 ) /* oz*m10 | oz*m9 | oz*m8 */
479
480 ADDPS( XMM5, XMM4 ) /* + | + | + */
481 ADDPS( XMM6, XMM4 ) /* + | + | + */
482 ADDPS( XMM3, XMM4 ) /* + | + | + */
483
484 MOVLPS( XMM4, D(0) ) /* => D(1) | => D(0) */
485 UNPCKHPS( XMM4, XMM4 )
486 MOVSS( XMM4, D(2) )
487
488 LLBL(K_GTP3P3DR_skip):
489 ADD_L( CONST(16), EDI )
490 ADD_L( EAX, ESI )
491 CMP_L( ECX, EDI )
492 JNE( LLBL(K_GTP3P3DR_top) )
493
494 LLBL(K_GTP3P3DR_finish):
495 POP_L( EDI )
496 POP_L( ESI )
497 RET
498 #undef FRAME_OFFSET