Added few more stubs so that control reaches to DestroyDevice().
[mesa.git] / src / mesa / x86 / sse_xform3.S
1
2 /*
3 * Mesa 3-D graphics library
4 *
5 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
13 *
14 * The above copyright notice and this permission notice shall be included
15 * in all copies or substantial portions of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
21 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
22 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
23 * OTHER DEALINGS IN THE SOFTWARE.
24 */
25
26 /** TODO:
27 * - insert PREFETCH instructions to avoid cache-misses !
28 * - some more optimizations are possible...
29 * - for 40-50% more performance in the SSE-functions, the
30 * data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned !
31 */
32
33 #ifdef USE_SSE_ASM
34 #include "assyntax.h"
35 #define MATH_ASM_PTR_SIZE 4
36 #include "math/m_vector_asm.h"
37 #include "xform_args.h"
38
39 SEG_TEXT
40
41 #define S(i) REGOFF(i * 4, ESI)
42 #define D(i) REGOFF(i * 4, EDI)
43 #define M(i) REGOFF(i * 4, EDX)
44
45
46 ALIGNTEXT4
47 GLOBL GLNAME(_mesa_sse_transform_points3_general)
48 HIDDEN(_mesa_sse_transform_points3_general)
49 GLNAME( _mesa_sse_transform_points3_general ):
50 _CET_ENDBR
51 #define FRAME_OFFSET 8
52 PUSH_L ( ESI )
53 PUSH_L ( EDI )
54
55 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
56 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
57
58 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
59 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
60
61 CMP_L ( CONST(0), ECX ) /* count == 0 ? */
62 JE ( LLBL(K_GTPGR_finish) ) /* yes -> nothing to do. */
63
64 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
65 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
66
67 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
68 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
69
70 SHL_L( CONST(4), ECX ) /* count *= 16 */
71 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
72
73 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
74 ADD_L( EDI, ECX ) /* count += dest ptr */
75
76
77 ALIGNTEXT32
78 MOVAPS ( REGOFF(0, EDX), XMM0 ) /* m0 | m1 | m2 | m3 */
79 MOVAPS ( REGOFF(16, EDX), XMM1 ) /* m4 | m5 | m6 | m7 */
80 MOVAPS ( REGOFF(32, EDX), XMM2 ) /* m8 | m9 | m10 | m11 */
81 MOVAPS ( REGOFF(48, EDX), XMM3 ) /* m12 | m13 | m14 | m15 */
82
83
84 ALIGNTEXT32
85 LLBL(K_GTPGR_top):
86 MOVSS ( REGOFF(0, ESI), XMM4 ) /* | | | ox */
87 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox | ox */
88 MOVSS ( REGOFF(4, ESI), XMM5 ) /* | | | oy */
89 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy | oy */
90 MOVSS ( REGOFF(8, ESI), XMM6 ) /* | | | oz */
91 SHUFPS ( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz | oz */
92
93 MULPS ( XMM0, XMM4 ) /* m3*ox | m2*ox | m1*ox | m0*ox */
94 MULPS ( XMM1, XMM5 ) /* m7*oy | m6*oy | m5*oy | m4*oy */
95 MULPS ( XMM2, XMM6 ) /* m11*oz | m10*oz | m9*oz | m8*oz */
96
97 ADDPS ( XMM5, XMM4 )
98 ADDPS ( XMM6, XMM4 )
99 ADDPS ( XMM3, XMM4 )
100
101 MOVAPS ( XMM4, REGOFF(0, EDI) )
102
103 LLBL(K_GTPGR_skip):
104 ADD_L ( CONST(16), EDI )
105 ADD_L ( EAX, ESI )
106 CMP_L ( ECX, EDI )
107 JNE ( LLBL(K_GTPGR_top) )
108
109 LLBL(K_GTPGR_finish):
110 POP_L ( EDI )
111 POP_L ( ESI )
112 RET
113 #undef FRAME_OFFSET
114
115
116 ALIGNTEXT4
117 GLOBL GLNAME(_mesa_sse_transform_points3_identity)
118 HIDDEN(_mesa_sse_transform_points3_identity)
119 GLNAME( _mesa_sse_transform_points3_identity ):
120 _CET_ENDBR
121 #define FRAME_OFFSET 8
122 PUSH_L ( ESI )
123 PUSH_L ( EDI )
124
125 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
126 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
127
128 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
129
130 TEST_L( ECX, ECX)
131 JZ( LLBL(K_GTPIR_finish) ) /* count was zero; go to finish */
132
133 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
134 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
135
136 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
137 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
138
139 SHL_L( CONST(4), ECX ) /* count *= 16 */
140 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
141
142 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
143 ADD_L( EDI, ECX ) /* count += dest ptr */
144
145 CMP_L( ESI, EDI )
146 JE( LLBL(K_GTPIR_finish) )
147
148
149 ALIGNTEXT32
150 LLBL(K_GTPIR_top):
151 MOVLPS ( S(0), XMM0 )
152 MOVLPS ( XMM0, D(0) )
153 MOVSS ( S(2), XMM0 )
154 MOVSS ( XMM0, D(2) )
155
156 LLBL(K_GTPIR_skip):
157 ADD_L ( CONST(16), EDI )
158 ADD_L ( EAX, ESI )
159 CMP_L ( ECX, EDI )
160 JNE ( LLBL(K_GTPIR_top) )
161
162 LLBL(K_GTPIR_finish):
163 POP_L ( EDI )
164 POP_L ( ESI )
165 RET
166 #undef FRAME_OFFSET
167
168
169
170
171 ALIGNTEXT4
172 GLOBL GLNAME(_mesa_sse_transform_points3_3d_no_rot)
173 HIDDEN(_mesa_sse_transform_points3_3d_no_rot)
174 GLNAME(_mesa_sse_transform_points3_3d_no_rot):
175 _CET_ENDBR
176 #define FRAME_OFFSET 8
177 PUSH_L( ESI )
178 PUSH_L( EDI )
179
180 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
181 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
182
183
184 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
185 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
186
187 TEST_L( ECX, ECX)
188 JZ( LLBL(K_GTP3DNRR_finish) ) /* count was zero; go to finish */
189
190 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
191 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
192
193 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
194 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
195
196 SHL_L( CONST(4), ECX ) /* count *= 16 */
197 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
198
199 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
200 ADD_L( EDI, ECX ) /* count += dest ptr */
201
202 XORPS( XMM0, XMM0 ) /* clean the working register */
203
204 ALIGNTEXT32
205 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
206 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
207 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
208 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */
209 MOVSS ( M(10), XMM3 ) /* - | - | - | m10 */
210 MOVSS ( M(14), XMM4 ) /* - | - | - | m14 */
211
212 ALIGNTEXT32
213 LLBL(K_GTP3DNRR_top):
214
215 MOVLPS ( S(0), XMM0 ) /* - | - | s1 | s0 */
216 MULPS ( XMM1, XMM0 ) /* - | - | s1*m5 | s0*m0 */
217 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */
218 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */
219
220 MOVSS ( S(2), XMM0 ) /* sz */
221 MULSS ( XMM3, XMM0 ) /* sz*m10 */
222 ADDSS ( XMM4, XMM0 ) /* +m14 */
223 MOVSS ( XMM0, D(2) ) /* -> D(2) */
224
225 LLBL(K_GTP3DNRR_skip):
226 ADD_L ( CONST(16), EDI )
227 ADD_L ( EAX, ESI )
228 CMP_L ( ECX, EDI )
229 JNE ( LLBL(K_GTP3DNRR_top) )
230
231 LLBL(K_GTP3DNRR_finish):
232 POP_L ( EDI )
233 POP_L ( ESI )
234 RET
235 #undef FRAME_OFFSET
236
237
238
239 ALIGNTEXT4
240 GLOBL GLNAME(_mesa_sse_transform_points3_perspective)
241 HIDDEN(_mesa_sse_transform_points3_perspective)
242 GLNAME(_mesa_sse_transform_points3_perspective):
243 _CET_ENDBR
244 #define FRAME_OFFSET 8
245 PUSH_L ( ESI )
246 PUSH_L ( EDI )
247
248 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
249 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
250
251 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
252 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
253
254 TEST_L( ECX, ECX)
255 JZ( LLBL(K_GTP3PR_finish) ) /* count was zero; go to finish */
256
257 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
258 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
259
260 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
261 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
262
263 SHL_L( CONST(4), ECX ) /* count *= 16 */
264 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
265
266 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
267 ADD_L( EDI, ECX ) /* count += dest ptr */
268
269 ALIGNTEXT32
270 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
271 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
272 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
273 MOVLPS ( M(8), XMM2 ) /* - | - | m9 | m8 */
274 MOVSS ( M(10), XMM3 ) /* m10 */
275 MOVSS ( M(14), XMM4 ) /* m14 */
276 XORPS ( XMM6, XMM6 ) /* 0 */
277
278 ALIGNTEXT32
279 LLBL(K_GTP3PR_top):
280 MOVLPS ( S(0), XMM0 ) /* oy | ox */
281 MULPS ( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
282 MOVSS ( S(2), XMM5 ) /* oz */
283 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oz | oz */
284 MULPS ( XMM2, XMM5 ) /* oz*m9 | oz*m8 */
285 ADDPS ( XMM5, XMM0 ) /* +oy*m5 | +ox*m0 */
286 MOVLPS ( XMM0, D(0) ) /* ->D(1) | ->D(0) */
287
288 MOVSS ( S(2), XMM0 ) /* oz */
289 MULSS ( XMM3, XMM0 ) /* oz*m10 */
290 ADDSS ( XMM4, XMM0 ) /* +m14 */
291 MOVSS ( XMM0, D(2) ) /* ->D(2) */
292
293 MOVSS ( S(2), XMM0 ) /* oz */
294 MOVSS ( XMM6, XMM5 ) /* 0 */
295 SUBPS ( XMM0, XMM5 ) /* -oz */
296 MOVSS ( XMM5, D(3) ) /* ->D(3) */
297
298 LLBL(K_GTP3PR_skip):
299 ADD_L( CONST(16), EDI )
300 ADD_L( EAX, ESI )
301 CMP_L( ECX, EDI )
302 JNE( LLBL(K_GTP3PR_top) )
303
304 LLBL(K_GTP3PR_finish):
305 POP_L ( EDI )
306 POP_L ( ESI )
307 RET
308 #undef FRAME_OFFSET
309
310
311
312 ALIGNTEXT4
313 GLOBL GLNAME(_mesa_sse_transform_points3_2d)
314 HIDDEN(_mesa_sse_transform_points3_2d)
315 GLNAME(_mesa_sse_transform_points3_2d):
316 _CET_ENDBR
317 #define FRAME_OFFSET 8
318 PUSH_L( ESI )
319 PUSH_L( EDI )
320
321 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
322 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
323
324 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
325 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
326
327 TEST_L( ECX, ECX)
328 JZ( LLBL(K_GTP3P2DR_finish) ) /* count was zero; go to finish */
329
330 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
331 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
332
333 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
334 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
335
336 SHL_L( CONST(4), ECX ) /* count *= 16 */
337 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
338
339 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
340 ADD_L( EDI, ECX ) /* count += dest ptr */
341
342 ALIGNTEXT32
343 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
344 MOVLPS( M(4), XMM1 ) /* m5 | m4 */
345 MOVLPS( M(12), XMM2 ) /* m13 | m12 */
346
347 ALIGNTEXT32
348 LLBL(K_GTP3P2DR_top):
349 MOVSS ( S(0), XMM3 ) /* ox */
350 SHUFPS ( CONST(0x0), XMM3, XMM3 ) /* ox | ox */
351 MULPS ( XMM0, XMM3 ) /* ox*m1 | ox*m0 */
352 MOVSS ( S(1), XMM4 ) /* oy */
353 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* oy | oy */
354 MULPS ( XMM1, XMM4 ) /* oy*m5 | oy*m4 */
355
356 ADDPS ( XMM4, XMM3 )
357 ADDPS ( XMM2, XMM3 )
358 MOVLPS ( XMM3, D(0) )
359
360 MOVSS ( S(2), XMM3 )
361 MOVSS ( XMM3, D(2) )
362
363 LLBL(K_GTP3P2DR_skip):
364 ADD_L ( CONST(16), EDI )
365 ADD_L ( EAX, ESI )
366 CMP_L ( ECX, EDI )
367 JNE ( LLBL(K_GTP3P2DR_top) )
368
369 LLBL(K_GTP3P2DR_finish):
370 POP_L ( EDI )
371 POP_L ( ESI )
372 RET
373 #undef FRAME_OFFSET
374
375
376
377 ALIGNTEXT4
378 GLOBL GLNAME(_mesa_sse_transform_points3_2d_no_rot)
379 HIDDEN(_mesa_sse_transform_points3_2d_no_rot)
380 GLNAME(_mesa_sse_transform_points3_2d_no_rot):
381 _CET_ENDBR
382 #define FRAME_OFFSET 8
383 PUSH_L( ESI )
384 PUSH_L( EDI )
385
386 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
387 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
388
389 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
390 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
391
392 TEST_L( ECX, ECX)
393 JZ( LLBL(K_GTP3P2DNRR_finish) ) /* count was zero; go to finish */
394
395 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
396 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
397
398 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
399 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
400
401 SHL_L( CONST(4), ECX ) /* count *= 16 */
402 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
403
404 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
405 ADD_L( EDI, ECX ) /* count += dest ptr */
406
407 ALIGNTEXT32
408 MOVSS ( M(0), XMM1 ) /* m0 */
409 MOVSS ( M(5), XMM2 ) /* m5 */
410 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */
411 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */
412
413 ALIGNTEXT32
414 LLBL(K_GTP3P2DNRR_top):
415 MOVLPS( S(0), XMM0 ) /* oy | ox */
416 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
417 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */
418 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */
419
420 MOVSS( S(2), XMM0 )
421 MOVSS( XMM0, D(2) )
422
423 LLBL(K_GTP3P2DNRR_skip):
424 ADD_L( CONST(16), EDI )
425 ADD_L( EAX, ESI )
426 CMP_L( ECX, EDI )
427 JNE( LLBL(K_GTP3P2DNRR_top) )
428
429 LLBL(K_GTP3P2DNRR_finish):
430 POP_L( EDI )
431 POP_L( ESI )
432 RET
433 #undef FRAME_OFFSET
434
435
436
437
438 ALIGNTEXT4
439 GLOBL GLNAME(_mesa_sse_transform_points3_3d)
440 HIDDEN(_mesa_sse_transform_points3_3d)
441 GLNAME(_mesa_sse_transform_points3_3d):
442 _CET_ENDBR
443 #define FRAME_OFFSET 8
444 PUSH_L( ESI )
445 PUSH_L( EDI )
446
447 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
448 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
449
450
451 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
452 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
453
454 TEST_L( ECX, ECX)
455 JZ( LLBL(K_GTP3P3DR_finish) ) /* count was zero; go to finish */
456
457 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
458 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
459
460 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
461 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
462
463 SHL_L( CONST(4), ECX ) /* count *= 16 */
464 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
465
466 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
467 ADD_L( EDI, ECX ) /* count += dest ptr */
468
469
470 ALIGNTEXT32
471 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */
472 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */
473 MOVAPS( M(8), XMM2 ) /* m10 | m9 | m8 */
474 MOVAPS( M(12), XMM3 ) /* m14 | m13 | m12 */
475
476 ALIGNTEXT32
477 LLBL(K_GTP3P3DR_top):
478 MOVSS( S(0), XMM4 )
479 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox */
480 MULPS( XMM0, XMM4 ) /* ox*m2 | ox*m1 | ox*m0 */
481
482 MOVSS( S(1), XMM5 )
483 SHUFPS( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy */
484 MULPS( XMM1, XMM5 ) /* oy*m6 | oy*m5 | oy*m4 */
485
486 MOVSS( S(2), XMM6 )
487 SHUFPS( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz */
488 MULPS( XMM2, XMM6 ) /* oz*m10 | oz*m9 | oz*m8 */
489
490 ADDPS( XMM5, XMM4 ) /* + | + | + */
491 ADDPS( XMM6, XMM4 ) /* + | + | + */
492 ADDPS( XMM3, XMM4 ) /* + | + | + */
493
494 MOVLPS( XMM4, D(0) ) /* => D(1) | => D(0) */
495 UNPCKHPS( XMM4, XMM4 )
496 MOVSS( XMM4, D(2) )
497
498 LLBL(K_GTP3P3DR_skip):
499 ADD_L( CONST(16), EDI )
500 ADD_L( EAX, ESI )
501 CMP_L( ECX, EDI )
502 JNE( LLBL(K_GTP3P3DR_top) )
503
504 LLBL(K_GTP3P3DR_finish):
505 POP_L( EDI )
506 POP_L( ESI )
507 RET
508 #undef FRAME_OFFSET
509 #endif
510
511 #if defined (__ELF__) && defined (__linux__)
512 .section .note.GNU-stack,"",%progbits
513 #endif