mesa: add support for CET to x86/x86-64 asm files.
[mesa.git] / src / mesa / x86 / sse_xform1.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_points1_general)
48 HIDDEN( _mesa_sse_transform_points1_general )
49 GLNAME( _mesa_sse_transform_points1_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_GTP1GR_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( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
79 MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
80
81 ALIGNTEXT32
82 LLBL(K_GTP1GR_top):
83 MOVSS( S(0), XMM2 ) /* ox */
84 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
85 MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
86 ADDPS( XMM1, XMM2 ) /* + | + | + | + */
87 MOVUPS( XMM2, D(0) )
88
89 LLBL(K_GTP1GR_skip):
90 ADD_L ( CONST(16), EDI )
91 ADD_L ( EAX, ESI )
92 CMP_L ( ECX, EDI )
93 JNE ( LLBL(K_GTP1GR_top) )
94
95 LLBL(K_GTP1GR_finish):
96 POP_L ( EDI )
97 POP_L ( ESI )
98 RET
99 #undef FRAME_OFFSET
100
101
102
103 ALIGNTEXT4
104 GLOBL GLNAME(_mesa_sse_transform_points1_identity)
105 HIDDEN(_mesa_sse_transform_points1_identity)
106 GLNAME( _mesa_sse_transform_points1_identity ):
107 _CET_ENDBR
108 #define FRAME_OFFSET 8
109 PUSH_L ( ESI )
110 PUSH_L ( EDI )
111
112 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
113 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
114
115 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
116
117 TEST_L( ECX, ECX)
118 JZ( LLBL(K_GTP1IR_finish) ) /* count was zero; go to finish */
119
120 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
121 OR_L( CONST(VEC_SIZE_1), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
122
123 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
124 MOV_L( CONST(1), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
125
126 SHL_L( CONST(4), ECX ) /* count *= 16 */
127 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
128
129 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
130 ADD_L( EDI, ECX ) /* count += dest ptr */
131
132 CMP_L( ESI, EDI )
133 JE( LLBL(K_GTP1IR_finish) )
134
135
136 ALIGNTEXT32
137 LLBL(K_GTP1IR_top):
138 MOV_L( S(0), EDX )
139 MOV_L( EDX, D(0) )
140
141 LLBL(K_GTP1IR_skip):
142 ADD_L ( CONST(16), EDI )
143 ADD_L ( EAX, ESI )
144 CMP_L ( ECX, EDI )
145 JNE ( LLBL(K_GTP1IR_top) )
146
147 LLBL(K_GTP1IR_finish):
148 POP_L ( EDI )
149 POP_L ( ESI )
150 RET
151 #undef FRAME_OFFSET
152
153
154
155 ALIGNTEXT4
156 GLOBL GLNAME(_mesa_sse_transform_points1_3d_no_rot)
157 HIDDEN(_mesa_sse_transform_points1_3d_no_rot)
158 GLNAME(_mesa_sse_transform_points1_3d_no_rot):
159 _CET_ENDBR
160 #define FRAME_OFFSET 8
161 PUSH_L( ESI )
162 PUSH_L( EDI )
163
164 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
165 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
166
167 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
168 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
169
170 TEST_L( ECX, ECX)
171 JZ( LLBL(K_GTP13DNRR_finish) ) /* count was zero; go to finish */
172
173 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
174 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
175
176 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
177 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
178
179 SHL_L( CONST(4), ECX ) /* count *= 16 */
180 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
181
182 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
183 ADD_L( EDI, ECX ) /* count += dest ptr */
184
185
186 ALIGNTEXT32
187 MOVSS( M(0), XMM0 ) /* m0 */
188 MOVSS( M(12), XMM1 ) /* m12 */
189 MOVSS( M(13), XMM2 ) /* m13 */
190 MOVSS( M(14), XMM3 ) /* m14 */
191
192 ALIGNTEXT32
193 LLBL(K_GTP13DNRR_top):
194 MOVSS( S(0), XMM4 ) /* ox */
195 MULSS( XMM0, XMM4 ) /* ox*m0 */
196 ADDSS( XMM1, XMM4 ) /* ox*m0+m12 */
197 MOVSS( XMM4, D(0) )
198
199 MOVSS( XMM2, D(1) )
200 MOVSS( XMM3, D(2) )
201
202 LLBL(K_GTP13DNRR_skip):
203 ADD_L ( CONST(16), EDI )
204 ADD_L ( EAX, ESI )
205 CMP_L ( ECX, EDI )
206 JNE ( LLBL(K_GTP13DNRR_top) )
207
208 LLBL(K_GTP13DNRR_finish):
209 POP_L ( EDI )
210 POP_L ( ESI )
211 RET
212 #undef FRAME_OFFSET
213
214
215
216 ALIGNTEXT4
217 GLOBL GLNAME(_mesa_sse_transform_points1_perspective)
218 HIDDEN(_mesa_sse_transform_points1_perspective)
219 GLNAME(_mesa_sse_transform_points1_perspective):
220 _CET_ENDBR
221 #define FRAME_OFFSET 8
222 PUSH_L ( ESI )
223 PUSH_L ( EDI )
224
225 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
226 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
227
228 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
229 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
230
231 TEST_L( ECX, ECX)
232 JZ( LLBL(K_GTP13PR_finish) ) /* count was zero; go to finish */
233
234 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
235 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
236
237 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
238 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
239
240 SHL_L( CONST(4), ECX ) /* count *= 16 */
241 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
242
243 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
244 ADD_L( EDI, ECX ) /* count += dest ptr */
245
246
247 ALIGNTEXT32
248 XORPS( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
249 MOVSS( M(0), XMM1 ) /* m0 */
250 MOVSS( M(14), XMM2 ) /* m14 */
251
252 ALIGNTEXT32
253 LLBL(K_GTP13PR_top):
254 MOVSS( S(0), XMM3 ) /* ox */
255 MULSS( XMM1, XMM3 ) /* ox*m0 */
256 MOVSS( XMM3, D(0) ) /* ox*m0->D(0) */
257 MOVSS( XMM2, D(2) ) /* m14->D(2) */
258
259 MOVSS( XMM0, D(1) )
260 MOVSS( XMM0, D(3) )
261
262 LLBL(K_GTP13PR_skip):
263 ADD_L( CONST(16), EDI )
264 ADD_L( EAX, ESI )
265 CMP_L( ECX, EDI )
266 JNE( LLBL(K_GTP13PR_top) )
267
268 LLBL(K_GTP13PR_finish):
269 POP_L ( EDI )
270 POP_L ( ESI )
271 RET
272 #undef FRAME_OFFSET
273
274
275 ALIGNTEXT4
276 GLOBL GLNAME(_mesa_sse_transform_points1_2d)
277 HIDDEN(_mesa_sse_transform_points1_2d)
278 GLNAME(_mesa_sse_transform_points1_2d):
279 _CET_ENDBR
280 #define FRAME_OFFSET 8
281 PUSH_L( ESI )
282 PUSH_L( EDI )
283
284 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
285 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
286
287 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
288 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
289
290 TEST_L( ECX, ECX)
291 JZ( LLBL(K_GTP13P2DR_finish) ) /* count was zero; go to finish */
292
293 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
294 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
295
296 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
297 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
298
299 SHL_L( CONST(4), ECX ) /* count *= 16 */
300 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
301
302 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
303 ADD_L( EDI, ECX ) /* count += dest ptr */
304
305 ALIGNTEXT32
306 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
307 MOVLPS( M(12), XMM1 ) /* m13 | m12 */
308
309 ALIGNTEXT32
310 LLBL(K_GTP13P2DR_top):
311 MOVSS( S(0), XMM2 ) /* ox */
312 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
313 MULPS( XMM0, XMM2 ) /* - | - | ox*m1 | ox*m0 */
314 ADDPS( XMM1, XMM2 ) /* - | - | ox*m1+m13 | ox*m0+m12 */
315 MOVLPS( XMM2, D(0) )
316
317 LLBL(K_GTP13P2DR_skip):
318 ADD_L ( CONST(16), EDI )
319 ADD_L ( EAX, ESI )
320 CMP_L ( ECX, EDI )
321 JNE ( LLBL(K_GTP13P2DR_top) )
322
323 LLBL(K_GTP13P2DR_finish):
324 POP_L ( EDI )
325 POP_L ( ESI )
326 RET
327 #undef FRAME_OFFSET
328
329
330 ALIGNTEXT4
331 GLOBL GLNAME(_mesa_sse_transform_points1_2d_no_rot)
332 HIDDEN(_mesa_sse_transform_points1_2d_no_rot)
333 GLNAME(_mesa_sse_transform_points1_2d_no_rot):
334 _CET_ENDBR
335 #define FRAME_OFFSET 8
336 PUSH_L( ESI )
337 PUSH_L( EDI )
338
339 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
340 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
341
342 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
343 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
344
345 TEST_L( ECX, ECX)
346 JZ( LLBL(K_GTP13P2DNRR_finish) ) /* count was zero; go to finish */
347
348 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
349 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
350
351 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
352 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
353
354 SHL_L( CONST(4), ECX ) /* count *= 16 */
355 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
356
357 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
358 ADD_L( EDI, ECX ) /* count += dest ptr */
359
360 ALIGNTEXT32
361 MOVSS( M(0), XMM0 ) /* m0 */
362 MOVSS( M(12), XMM1 ) /* m12 */
363 MOVSS( M(13), XMM2 ) /* m13 */
364
365 ALIGNTEXT32
366 LLBL(K_GTP13P2DNRR_top):
367 MOVSS( S(0), XMM3 ) /* ox */
368 MULSS( XMM0, XMM3 ) /* ox*m0 */
369 ADDSS( XMM1, XMM3 ) /* ox*m0+m12 */
370 MOVSS( XMM3, D(0) )
371 MOVSS( XMM2, D(1) )
372
373 LLBL(K_GTP13P2DNRR_skip):
374 ADD_L( CONST(16), EDI )
375 ADD_L( EAX, ESI )
376 CMP_L( ECX, EDI )
377 JNE( LLBL(K_GTP13P2DNRR_top) )
378
379 LLBL(K_GTP13P2DNRR_finish):
380 POP_L( EDI )
381 POP_L( ESI )
382 RET
383 #undef FRAME_OFFSET
384
385
386
387 ALIGNTEXT4
388 GLOBL GLNAME(_mesa_sse_transform_points1_3d)
389 HIDDEN(_mesa_sse_transform_points1_3d)
390 GLNAME(_mesa_sse_transform_points1_3d):
391 _CET_ENDBR
392 #define FRAME_OFFSET 8
393 PUSH_L( ESI )
394 PUSH_L( EDI )
395
396 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
397 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
398
399 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
400 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
401
402 TEST_L( ECX, ECX)
403 JZ( LLBL(K_GTP13P3DR_finish) ) /* count was zero; go to finish */
404
405 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
406 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
407
408 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
409 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
410
411 SHL_L( CONST(4), ECX ) /* count *= 16 */
412 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
413
414 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
415 ADD_L( EDI, ECX ) /* count += dest ptr */
416
417
418 ALIGNTEXT32
419 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
420 MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
421
422 ALIGNTEXT32
423 LLBL(K_GTP13P3DR_top):
424 MOVSS( S(0), XMM2 ) /* ox */
425 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
426 MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
427 ADDPS( XMM1, XMM2 ) /* +m15 | +m14 | +m13 | +m12 */
428 MOVLPS( XMM2, D(0) ) /* - | - | ->D(1)| ->D(0)*/
429 UNPCKHPS( XMM2, XMM2 ) /* ox*m3+m15 | ox*m3+m15 | ox*m2+m14 | ox*m2+m14 */
430 MOVSS( XMM2, D(2) )
431
432 LLBL(K_GTP13P3DR_skip):
433 ADD_L( CONST(16), EDI )
434 ADD_L( EAX, ESI )
435 CMP_L( ECX, EDI )
436 JNE( LLBL(K_GTP13P3DR_top) )
437
438 LLBL(K_GTP13P3DR_finish):
439 POP_L( EDI )
440 POP_L( ESI )
441 RET
442 #undef FRAME_OFFSET
443 #endif
444
445 #if defined (__ELF__) && defined (__linux__)
446 .section .note.GNU-stack,"",%progbits
447 #endif