3 * Mesa 3-D graphics library
6 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice shall be included
16 * in all copies or substantial portions of the Software.
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
22 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
23 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 * NOTE: Avoid using spaces in between '(' ')' and arguments, especially
28 * with macros like CONST, LLBL that expand to CONCAT(...). Putting spaces
29 * in there will break the build on some platforms.
33 #include "xform_args.h"
37 #define FP_ONE 1065353216
40 #define SRC0 REGOFF(0, ESI)
41 #define SRC1 REGOFF(4, ESI)
42 #define SRC2 REGOFF(8, ESI)
43 #define SRC3 REGOFF(12, ESI)
44 #define DST0 REGOFF(0, EDI)
45 #define DST1 REGOFF(4, EDI)
46 #define DST2 REGOFF(8, EDI)
47 #define DST3 REGOFF(12, EDI)
48 #define MAT0 REGOFF(0, EDX)
49 #define MAT1 REGOFF(4, EDX)
50 #define MAT2 REGOFF(8, EDX)
51 #define MAT3 REGOFF(12, EDX)
52 #define MAT4 REGOFF(16, EDX)
53 #define MAT5 REGOFF(20, EDX)
54 #define MAT6 REGOFF(24, EDX)
55 #define MAT7 REGOFF(28, EDX)
56 #define MAT8 REGOFF(32, EDX)
57 #define MAT9 REGOFF(36, EDX)
58 #define MAT10 REGOFF(40, EDX)
59 #define MAT11 REGOFF(44, EDX)
60 #define MAT12 REGOFF(48, EDX)
61 #define MAT13 REGOFF(52, EDX)
62 #define MAT14 REGOFF(56, EDX)
63 #define MAT15 REGOFF(60, EDX)
67 GLOBL GLNAME( _mesa_x86_transform_points3_general )
68 HIDDEN(_mesa_x86_transform_points3_general)
69 GLNAME( _mesa_x86_transform_points3_general ):
71 #define FRAME_OFFSET 8
75 MOV_L( ARG_SOURCE, ESI )
76 MOV_L( ARG_DEST, EDI )
78 MOV_L( ARG_MATRIX, EDX )
79 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
82 JZ( LLBL(x86_p3_gr_done) )
84 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
85 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
87 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
88 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
90 SHL_L( CONST(4), ECX )
91 MOV_L( REGOFF(V4F_START, ESI), ESI )
93 MOV_L( REGOFF(V4F_START, EDI), EDI )
99 FLD_S( SRC0 ) /* F4 */
101 FLD_S( SRC0 ) /* F5 F4 */
103 FLD_S( SRC0 ) /* F6 F5 F4 */
105 FLD_S( SRC0 ) /* F7 F6 F5 F4 */
108 FLD_S( SRC1 ) /* F0 F7 F6 F5 F4 */
110 FLD_S( SRC1 ) /* F1 F0 F7 F6 F5 F4 */
112 FLD_S( SRC1 ) /* F2 F1 F0 F7 F6 F5 F4 */
114 FLD_S( SRC1 ) /* F3 F2 F1 F0 F7 F6 F5 F4 */
117 FXCH( ST(3) ) /* F0 F2 F1 F3 F7 F6 F5 F4 */
118 FADDP( ST0, ST(7) ) /* F2 F1 F3 F7 F6 F5 F4 */
119 FXCH( ST(1) ) /* F1 F2 F3 F7 F6 F5 F4 */
120 FADDP( ST0, ST(5) ) /* F2 F3 F7 F6 F5 F4 */
121 FADDP( ST0, ST(3) ) /* F3 F7 F6 F5 F4 */
122 FADDP( ST0, ST(1) ) /* F7 F6 F5 F4 */
124 FLD_S( SRC2 ) /* F0 F7 F6 F5 F4 */
126 FLD_S( SRC2 ) /* F1 F0 F7 F6 F5 F4 */
128 FLD_S( SRC2 ) /* F2 F1 F0 F7 F6 F5 F4 */
130 FLD_S( SRC2 ) /* F3 F2 F1 F0 F7 F6 F5 F4 */
133 FXCH( ST(3) ) /* F0 F2 F1 F3 F7 F6 F5 F4 */
134 FADDP( ST0, ST(7) ) /* F2 F1 F3 F7 F6 F5 F4 */
135 FXCH( ST(1) ) /* F1 F2 F3 F7 F6 F5 F4 */
136 FADDP( ST0, ST(5) ) /* F2 F3 F7 F6 F5 F4 */
137 FADDP( ST0, ST(3) ) /* F3 F7 F6 F5 F4 */
138 FADDP( ST0, ST(1) ) /* F7 F6 F5 F4 */
140 FXCH( ST(3) ) /* F4 F6 F5 F7 */
142 FXCH( ST(2) ) /* F5 F6 F4 F7 */
144 FXCH( ST(1) ) /* F6 F5 F4 F7 */
146 FXCH( ST(3) ) /* F7 F5 F4 F6 */
149 FXCH( ST(2) ) /* F4 F5 F7 F6 */
150 FSTP_S( DST0 ) /* F5 F7 F6 */
151 FSTP_S( DST1 ) /* F7 F6 */
152 FXCH( ST(1) ) /* F6 F7 */
153 FSTP_S( DST2 ) /* F7 */
156 LLBL(x86_p3_gr_skip):
158 ADD_L( CONST(16), EDI )
161 JNE( LLBL(x86_p3_gr_loop) )
163 LLBL(x86_p3_gr_done):
174 GLOBL GLNAME( _mesa_x86_transform_points3_perspective )
175 HIDDEN(_mesa_x86_transform_points3_perspective)
176 GLNAME( _mesa_x86_transform_points3_perspective ):
178 #define FRAME_OFFSET 12
183 MOV_L( ARG_SOURCE, ESI )
184 MOV_L( ARG_DEST, EDI )
186 MOV_L( ARG_MATRIX, EDX )
187 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
190 JZ( LLBL(x86_p3_pr_done) )
192 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
193 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
195 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
196 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
198 SHL_L( CONST(4), ECX )
199 MOV_L( REGOFF(V4F_START, ESI), ESI )
201 MOV_L( REGOFF(V4F_START, EDI), EDI )
205 LLBL(x86_p3_pr_loop):
207 FLD_S( SRC0 ) /* F4 */
210 FLD_S( SRC1 ) /* F5 F4 */
213 FLD_S( SRC2 ) /* F0 F5 F4 */
215 FLD_S( SRC2 ) /* F1 F0 F5 F4 */
217 FLD_S( SRC2 ) /* F2 F1 F0 F5 F4 */
220 FXCH( ST(2) ) /* F0 F1 F2 F5 F4 */
221 FADDP( ST0, ST(4) ) /* F1 F2 F5 F4 */
222 FADDP( ST0, ST(2) ) /* F2 F5 F4 */
223 FLD_S( MAT14 ) /* F6 F2 F5 F4 */
224 FXCH( ST(1) ) /* F2 F6 F5 F4 */
225 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
228 XOR_L( CONST(-2147483648), EBX )/* change sign */
230 FXCH( ST(2) ) /* F4 F5 F6 */
231 FSTP_S( DST0 ) /* F5 F6 */
232 FSTP_S( DST1 ) /* F6 */
236 LLBL(x86_p3_pr_skip):
238 ADD_L( CONST(16), EDI )
241 JNE( LLBL(x86_p3_pr_loop) )
243 LLBL(x86_p3_pr_done):
255 GLOBL GLNAME( _mesa_x86_transform_points3_3d )
256 HIDDEN(_mesa_x86_transform_points3_3d)
257 GLNAME( _mesa_x86_transform_points3_3d ):
259 #define FRAME_OFFSET 8
263 MOV_L( ARG_SOURCE, ESI )
264 MOV_L( ARG_DEST, EDI )
266 MOV_L( ARG_MATRIX, EDX )
267 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
270 JZ( LLBL(x86_p3_3dr_done) )
272 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
273 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) )
275 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
276 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) )
278 SHL_L( CONST(4), ECX )
279 MOV_L( REGOFF(V4F_START, ESI), ESI )
281 MOV_L( REGOFF(V4F_START, EDI), EDI )
285 LLBL(x86_p3_3dr_loop):
287 FLD_S( SRC0 ) /* F4 */
289 FLD_S( SRC0 ) /* F5 F4 */
291 FLD_S( SRC0 ) /* F6 F5 F4 */
294 FLD_S( SRC1 ) /* F0 F6 F5 F4 */
296 FLD_S( SRC1 ) /* F1 F0 F6 F5 F4 */
298 FLD_S( SRC1 ) /* F2 F1 F0 F6 F5 F4 */
301 FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */
302 FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */
303 FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */
304 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
306 FLD_S( SRC2 ) /* F0 F6 F5 F4 */
308 FLD_S( SRC2 ) /* F1 F0 F6 F5 F4 */
310 FLD_S( SRC2 ) /* F2 F1 F0 F6 F5 F4 */
313 FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */
314 FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */
315 FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */
316 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
318 FXCH( ST(2) ) /* F4 F5 F6 */
320 FXCH( ST(1) ) /* F5 F4 F6 */
322 FXCH( ST(2) ) /* F6 F4 F5 */
325 FXCH( ST(1) ) /* F4 F6 F5 */
326 FSTP_S( DST0 ) /* F6 F5 */
327 FXCH( ST(1) ) /* F5 F6 */
328 FSTP_S( DST1 ) /* F6 */
331 LLBL(x86_p3_3dr_skip):
333 ADD_L( CONST(16), EDI )
336 JNE( LLBL(x86_p3_3dr_loop) )
338 LLBL(x86_p3_3dr_done):
349 GLOBL GLNAME( _mesa_x86_transform_points3_3d_no_rot )
350 HIDDEN(_mesa_x86_transform_points3_3d_no_rot)
351 GLNAME( _mesa_x86_transform_points3_3d_no_rot ):
353 #define FRAME_OFFSET 8
357 MOV_L( ARG_SOURCE, ESI )
358 MOV_L( ARG_DEST, EDI )
361 MOV_L( ARG_MATRIX, EDX )
362 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
365 JZ( LLBL(x86_p3_3dnrr_done) )
367 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
368 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) )
370 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
371 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) )
373 SHL_L( CONST(4), ECX )
374 MOV_L( REGOFF(V4F_START, ESI), ESI )
376 MOV_L( REGOFF(V4F_START, EDI), EDI )
380 LLBL(x86_p3_3dnrr_loop):
382 FLD_S( SRC0 ) /* F4 */
385 FLD_S( SRC1 ) /* F1 F4 */
388 FLD_S( SRC2 ) /* F2 F1 F4 */
391 FXCH( ST(2) ) /* F4 F1 F2 */
393 FLD_S( MAT13 ) /* F5 F4 F1 F2 */
394 FXCH( ST(2) ) /* F1 F4 F5 F2 */
395 FADDP( ST0, ST(2) ) /* F4 F5 F2 */
396 FLD_S( MAT14 ) /* F6 F4 F5 F2 */
397 FXCH( ST(3) ) /* F2 F4 F5 F6 */
398 FADDP( ST0, ST(3) ) /* F4 F5 F6 */
400 FSTP_S( DST0 ) /* F5 F6 */
401 FSTP_S( DST1 ) /* F6 */
404 LLBL(x86_p3_3dnrr_skip):
406 ADD_L( CONST(16), EDI )
409 JNE( LLBL(x86_p3_3dnrr_loop) )
411 LLBL(x86_p3_3dnrr_done):
422 GLOBL GLNAME( _mesa_x86_transform_points3_2d )
423 HIDDEN(_mesa_x86_transform_points3_2d)
424 GLNAME( _mesa_x86_transform_points3_2d ):
426 #define FRAME_OFFSET 12
431 MOV_L( ARG_SOURCE, ESI )
432 MOV_L( ARG_DEST, EDI )
434 MOV_L( ARG_MATRIX, EDX )
435 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
438 JZ( LLBL(x86_p3_2dr_done) )
440 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
441 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) )
443 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
444 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) )
446 SHL_L( CONST(4), ECX )
447 MOV_L( REGOFF(V4F_START, ESI), ESI )
449 MOV_L( REGOFF(V4F_START, EDI), EDI )
453 LLBL(x86_p3_2dr_loop):
455 FLD_S( SRC0 ) /* F4 */
457 FLD_S( SRC0 ) /* F5 F4 */
460 FLD_S( SRC1 ) /* F0 F5 F4 */
462 FLD_S( SRC1 ) /* F1 F0 F5 F4 */
465 FXCH( ST(1) ) /* F0 F1 F5 F4 */
466 FADDP( ST0, ST(3) ) /* F1 F5 F4 */
467 FADDP( ST0, ST(1) ) /* F5 F4 */
469 FXCH( ST(1) ) /* F4 F5 */
471 FXCH( ST(1) ) /* F5 F4 */
476 FXCH( ST(1) ) /* F4 F5 */
477 FSTP_S( DST0 ) /* F5 */
481 LLBL(x86_p3_2dr_skip):
483 ADD_L( CONST(16), EDI )
486 JNE( LLBL(x86_p3_2dr_loop) )
488 LLBL(x86_p3_2dr_done):
500 GLOBL GLNAME( _mesa_x86_transform_points3_2d_no_rot )
501 HIDDEN(_mesa_x86_transform_points3_2d_no_rot)
502 GLNAME( _mesa_x86_transform_points3_2d_no_rot ):
504 #define FRAME_OFFSET 12
509 MOV_L( ARG_SOURCE, ESI )
510 MOV_L( ARG_DEST, EDI )
512 MOV_L( ARG_MATRIX, EDX )
513 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
516 JZ( LLBL(x86_p3_2dnrr_done) )
518 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
519 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) )
521 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
522 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) )
524 SHL_L( CONST(4), ECX )
525 MOV_L( REGOFF(V4F_START, ESI), ESI )
527 MOV_L( REGOFF(V4F_START, EDI), EDI )
531 LLBL(x86_p3_2dnrr_loop):
533 FLD_S( SRC0 ) /* F4 */
536 FLD_S( SRC1 ) /* F1 F4 */
539 FXCH( ST(1) ) /* F4 F1 */
541 FLD_S( MAT13 ) /* F5 F4 F1 */
543 FXCH( ST(2) ) /* F1 F4 F5 */
544 FADDP( ST0, ST(2) ) /* F4 F5 */
548 FSTP_S( DST0 ) /* F5 */
552 LLBL(x86_p3_2dnrr_skip):
554 ADD_L( CONST(16), EDI )
557 JNE( LLBL(x86_p3_2dnrr_loop) )
559 LLBL(x86_p3_2dnrr_done):
571 GLOBL GLNAME( _mesa_x86_transform_points3_identity )
572 HIDDEN(_mesa_x86_transform_points3_identity)
573 GLNAME(_mesa_x86_transform_points3_identity ):
575 #define FRAME_OFFSET 16
581 MOV_L( ARG_SOURCE, ESI )
582 MOV_L( ARG_DEST, EDI )
584 MOV_L( ARG_MATRIX, EDX )
585 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
588 JZ( LLBL(x86_p3_ir_done) )
590 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
591 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) )
593 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
594 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) )
596 SHL_L( CONST(4), ECX )
597 MOV_L( REGOFF(V4F_START, ESI), ESI )
599 MOV_L( REGOFF(V4F_START, EDI), EDI )
603 JE( LLBL(x86_p3_ir_done) )
606 LLBL(x86_p3_ir_loop):
626 LLBL(x86_p3_ir_skip):
628 ADD_L( CONST(16), EDI )
631 JNE( LLBL(x86_p3_ir_loop) )
633 LLBL(x86_p3_ir_done):
641 #if defined (__ELF__) && defined (__linux__)
642 .section .note.GNU-stack,"",%progbits