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 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
22 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
23 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
24 * OTHER DEALINGS IN THE SOFTWARE.
28 * NOTE: Avoid using spaces in between '(' ')' and arguments, especially
29 * with macros like CONST, LLBL that expand to CONCAT(...). Putting spaces
30 * in there will break the build on some platforms.
35 #include "xform_args.h"
39 #define FP_ONE 1065353216
42 #define SRC0 REGOFF(0, ESI)
43 #define SRC1 REGOFF(4, ESI)
44 #define SRC2 REGOFF(8, ESI)
45 #define SRC3 REGOFF(12, ESI)
46 #define DST0 REGOFF(0, EDI)
47 #define DST1 REGOFF(4, EDI)
48 #define DST2 REGOFF(8, EDI)
49 #define DST3 REGOFF(12, EDI)
50 #define MAT0 REGOFF(0, EDX)
51 #define MAT1 REGOFF(4, EDX)
52 #define MAT2 REGOFF(8, EDX)
53 #define MAT3 REGOFF(12, EDX)
54 #define MAT4 REGOFF(16, EDX)
55 #define MAT5 REGOFF(20, EDX)
56 #define MAT6 REGOFF(24, EDX)
57 #define MAT7 REGOFF(28, EDX)
58 #define MAT8 REGOFF(32, EDX)
59 #define MAT9 REGOFF(36, EDX)
60 #define MAT10 REGOFF(40, EDX)
61 #define MAT11 REGOFF(44, EDX)
62 #define MAT12 REGOFF(48, EDX)
63 #define MAT13 REGOFF(52, EDX)
64 #define MAT14 REGOFF(56, EDX)
65 #define MAT15 REGOFF(60, EDX)
69 GLOBL GLNAME( _mesa_x86_transform_points4_general )
70 HIDDEN(_mesa_x86_transform_points4_general)
71 GLNAME( _mesa_x86_transform_points4_general ):
73 #define FRAME_OFFSET 8
77 MOV_L( ARG_SOURCE, ESI )
78 MOV_L( ARG_DEST, EDI )
80 MOV_L( ARG_MATRIX, EDX )
81 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
84 JZ( LLBL(x86_p4_gr_done) )
86 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
87 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
89 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
90 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
92 SHL_L( CONST(4), ECX )
93 MOV_L( REGOFF(V4F_START, ESI), ESI )
95 MOV_L( REGOFF(V4F_START, EDI), EDI )
101 FLD_S( SRC0 ) /* F4 */
103 FLD_S( SRC0 ) /* F5 F4 */
105 FLD_S( SRC0 ) /* F6 F5 F4 */
107 FLD_S( SRC0 ) /* F7 F6 F5 F4 */
110 FLD_S( SRC1 ) /* F0 F7 F6 F5 F4 */
112 FLD_S( SRC1 ) /* F1 F0 F7 F6 F5 F4 */
114 FLD_S( SRC1 ) /* F2 F1 F0 F7 F6 F5 F4 */
116 FLD_S( SRC1 ) /* F3 F2 F1 F0 F7 F6 F5 F4 */
119 FXCH( ST(3) ) /* F0 F2 F1 F3 F7 F6 F5 F4 */
120 FADDP( ST0, ST(7) ) /* F2 F1 F3 F7 F6 F5 F4 */
121 FXCH( ST(1) ) /* F1 F2 F3 F7 F6 F5 F4 */
122 FADDP( ST0, ST(5) ) /* F2 F3 F7 F6 F5 F4 */
123 FADDP( ST0, ST(3) ) /* F3 F7 F6 F5 F4 */
124 FADDP( ST0, ST(1) ) /* F7 F6 F5 F4 */
126 FLD_S( SRC2 ) /* F0 F7 F6 F5 F4 */
128 FLD_S( SRC2 ) /* F1 F0 F7 F6 F5 F4 */
130 FLD_S( SRC2 ) /* F2 F1 F0 F7 F6 F5 F4 */
132 FLD_S( SRC2 ) /* F3 F2 F1 F0 F7 F6 F5 F4 */
135 FXCH( ST(3) ) /* F0 F2 F1 F3 F7 F6 F5 F4 */
136 FADDP( ST0, ST(7) ) /* F2 F1 F3 F7 F6 F5 F4 */
137 FXCH( ST(1) ) /* F1 F2 F3 F7 F6 F5 F4 */
138 FADDP( ST0, ST(5) ) /* F2 F3 F7 F6 F5 F4 */
139 FADDP( ST0, ST(3) ) /* F3 F7 F6 F5 F4 */
140 FADDP( ST0, ST(1) ) /* F7 F6 F5 F4 */
142 FLD_S( SRC3 ) /* F0 F7 F6 F5 F4 */
144 FLD_S( SRC3 ) /* F1 F0 F7 F6 F5 F4 */
146 FLD_S( SRC3 ) /* F2 F1 F0 F7 F6 F5 F4 */
148 FLD_S( SRC3 ) /* F3 F2 F1 F0 F7 F6 F5 F4 */
151 FXCH( ST(3) ) /* F0 F2 F1 F3 F7 F6 F5 F4 */
152 FADDP( ST0, ST(7) ) /* F2 F1 F3 F7 F6 F5 F4 */
153 FXCH( ST(1) ) /* F1 F2 F3 F7 F6 F5 F4 */
154 FADDP( ST0, ST(5) ) /* F2 F3 F7 F6 F5 F4 */
155 FADDP( ST0, ST(3) ) /* F3 F7 F6 F5 F4 */
156 FADDP( ST0, ST(1) ) /* F7 F6 F5 F4 */
158 FXCH( ST(3) ) /* F4 F6 F5 F7 */
159 FSTP_S( DST0 ) /* F6 F5 F7 */
160 FXCH( ST(1) ) /* F5 F6 F7 */
161 FSTP_S( DST1 ) /* F6 F7 */
162 FSTP_S( DST2 ) /* F7 */
165 LLBL(x86_p4_gr_skip):
167 ADD_L( CONST(16), EDI )
170 JNE( LLBL(x86_p4_gr_loop) )
172 LLBL(x86_p4_gr_done):
183 GLOBL GLNAME( _mesa_x86_transform_points4_perspective )
184 HIDDEN(_mesa_x86_transform_points4_perspective)
185 GLNAME( _mesa_x86_transform_points4_perspective ):
187 #define FRAME_OFFSET 12
192 MOV_L( ARG_SOURCE, ESI )
193 MOV_L( ARG_DEST, EDI )
195 MOV_L( ARG_MATRIX, EDX )
196 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
199 JZ( LLBL(x86_p4_pr_done) )
201 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
202 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
204 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
205 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
207 SHL_L( CONST(4), ECX )
208 MOV_L( REGOFF(V4F_START, ESI), ESI )
210 MOV_L( REGOFF(V4F_START, EDI), EDI )
214 LLBL(x86_p4_pr_loop):
216 FLD_S( SRC0 ) /* F4 */
219 FLD_S( SRC1 ) /* F5 F4 */
222 FLD_S( SRC2 ) /* F0 F5 F4 */
224 FLD_S( SRC2 ) /* F1 F0 F5 F4 */
226 FLD_S( SRC2 ) /* F6 F1 F0 F5 F4 */
229 FXCH( ST(2) ) /* F0 F1 F6 F5 F4 */
230 FADDP( ST0, ST(4) ) /* F1 F6 F5 F4 */
231 FADDP( ST0, ST(2) ) /* F6 F5 F4 */
233 FLD_S( SRC3 ) /* F2 F6 F5 F4 */
236 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
239 XOR_L( CONST(-2147483648), EBX )/* change sign */
241 FXCH( ST(2) ) /* F4 F5 F6 */
242 FSTP_S( DST0 ) /* F5 F6 */
243 FSTP_S( DST1 ) /* F6 */
247 LLBL(x86_p4_pr_skip):
249 ADD_L( CONST(16), EDI )
252 JNE( LLBL(x86_p4_pr_loop) )
254 LLBL(x86_p4_pr_done):
266 GLOBL GLNAME( _mesa_x86_transform_points4_3d )
267 HIDDEN(_mesa_x86_transform_points4_3d)
268 GLNAME( _mesa_x86_transform_points4_3d ):
270 #define FRAME_OFFSET 12
275 MOV_L( ARG_SOURCE, ESI )
276 MOV_L( ARG_DEST, EDI )
278 MOV_L( ARG_MATRIX, EDX )
279 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
282 JZ( LLBL(x86_p4_3dr_done) )
284 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
285 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
287 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
288 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
290 SHL_L( CONST(4), ECX )
291 MOV_L( REGOFF(V4F_START, ESI), ESI )
293 MOV_L( REGOFF(V4F_START, EDI), EDI )
297 LLBL(x86_p4_3dr_loop):
299 FLD_S( SRC0 ) /* F4 */
301 FLD_S( SRC0 ) /* F5 F4 */
303 FLD_S( SRC0 ) /* F6 F5 F4 */
306 FLD_S( SRC1 ) /* F0 F6 F5 F4 */
308 FLD_S( SRC1 ) /* F1 F0 F6 F5 F4 */
310 FLD_S( SRC1 ) /* 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 FLD_S( SRC2 ) /* F0 F6 F5 F4 */
320 FLD_S( SRC2 ) /* F1 F0 F6 F5 F4 */
322 FLD_S( SRC2 ) /* F2 F1 F0 F6 F5 F4 */
325 FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */
326 FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */
327 FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */
328 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
330 FLD_S( SRC3 ) /* F0 F6 F5 F4 */
332 FLD_S( SRC3 ) /* F1 F0 F6 F5 F4 */
334 FLD_S( SRC3 ) /* F2 F1 F0 F6 F5 F4 */
337 FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */
338 FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */
339 FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */
340 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
344 FXCH( ST(2) ) /* F4 F5 F6 */
345 FSTP_S( DST0 ) /* F5 F6 */
346 FSTP_S( DST1 ) /* F6 */
350 LLBL(x86_p4_3dr_skip):
352 ADD_L( CONST(16), EDI )
355 JNE( LLBL(x86_p4_3dr_loop) )
357 LLBL(x86_p4_3dr_done):
369 GLOBL GLNAME(_mesa_x86_transform_points4_3d_no_rot)
370 HIDDEN(_mesa_x86_transform_points4_3d_no_rot)
371 GLNAME(_mesa_x86_transform_points4_3d_no_rot):
373 #define FRAME_OFFSET 12
378 MOV_L( ARG_SOURCE, ESI )
379 MOV_L( ARG_DEST, EDI )
381 MOV_L( ARG_MATRIX, EDX )
382 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
385 JZ( LLBL(x86_p4_3dnrr_done) )
387 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
388 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
390 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
391 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
393 SHL_L( CONST(4), ECX )
394 MOV_L( REGOFF(V4F_START, ESI), ESI )
396 MOV_L( REGOFF(V4F_START, EDI), EDI )
400 LLBL(x86_p4_3dnrr_loop):
402 FLD_S( SRC0 ) /* F4 */
405 FLD_S( SRC1 ) /* F5 F4 */
408 FLD_S( SRC2 ) /* F6 F5 F4 */
411 FLD_S( SRC3 ) /* F0 F6 F5 F4 */
413 FLD_S( SRC3 ) /* F1 F0 F6 F5 F4 */
415 FLD_S( SRC3 ) /* F2 F1 F0 F6 F5 F4 */
418 FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */
419 FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */
420 FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */
421 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
425 FXCH( ST(2) ) /* F4 F5 F6 */
426 FSTP_S( DST0 ) /* F5 F6 */
427 FSTP_S( DST1 ) /* F6 */
431 LLBL(x86_p4_3dnrr_skip):
433 ADD_L( CONST(16), EDI )
436 JNE( LLBL(x86_p4_3dnrr_loop) )
438 LLBL(x86_p4_3dnrr_done):
450 GLOBL GLNAME( _mesa_x86_transform_points4_2d )
451 HIDDEN(_mesa_x86_transform_points4_2d)
452 GLNAME( _mesa_x86_transform_points4_2d ):
454 #define FRAME_OFFSET 16
460 MOV_L( ARG_SOURCE, ESI )
461 MOV_L( ARG_DEST, EDI )
463 MOV_L( ARG_MATRIX, EDX )
464 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
467 JZ( LLBL(x86_p4_2dr_done) )
469 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
470 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
472 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
473 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
475 SHL_L( CONST(4), ECX )
476 MOV_L( REGOFF(V4F_START, ESI), ESI )
478 MOV_L( REGOFF(V4F_START, EDI), EDI )
482 LLBL(x86_p4_2dr_loop):
484 FLD_S( SRC0 ) /* F4 */
486 FLD_S( SRC0 ) /* F5 F4 */
489 FLD_S( SRC1 ) /* F0 F5 F4 */
491 FLD_S( SRC1 ) /* F1 F0 F5 F4 */
494 FXCH( ST(1) ) /* F0 F1 F5 F4 */
495 FADDP( ST0, ST(3) ) /* F1 F5 F4 */
496 FADDP( ST0, ST(1) ) /* F5 F4 */
498 FLD_S( SRC3 ) /* F0 F5 F4 */
500 FLD_S( SRC3 ) /* F1 F0 F5 F4 */
503 FXCH( ST(1) ) /* F0 F1 F5 F4 */
504 FADDP( ST0, ST(3) ) /* F1 F5 F4 */
505 FADDP( ST0, ST(1) ) /* F5 F4 */
510 FXCH( ST(1) ) /* F4 F5 */
511 FSTP_S( DST0 ) /* F5 */
516 LLBL(x86_p4_2dr_skip):
518 ADD_L( CONST(16), EDI )
521 JNE( LLBL(x86_p4_2dr_loop) )
523 LLBL(x86_p4_2dr_done):
536 GLOBL GLNAME( _mesa_x86_transform_points4_2d_no_rot )
537 HIDDEN(_mesa_x86_transform_points4_2d_no_rot)
538 GLNAME( _mesa_x86_transform_points4_2d_no_rot ):
540 #define FRAME_OFFSET 16
546 MOV_L( ARG_SOURCE, ESI )
547 MOV_L( ARG_DEST, EDI )
549 MOV_L( ARG_MATRIX, EDX )
550 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
553 JZ( LLBL(x86_p4_2dnrr_done) )
555 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
556 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
558 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
559 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
561 SHL_L( CONST(4), ECX )
562 MOV_L( REGOFF(V4F_START, ESI), ESI )
564 MOV_L( REGOFF(V4F_START, EDI), EDI )
568 LLBL(x86_p4_2dnrr_loop):
570 FLD_S( SRC0 ) /* F4 */
573 FLD_S( SRC1 ) /* F5 F4 */
576 FLD_S( SRC3 ) /* F0 F5 F4 */
578 FLD_S( SRC3 ) /* F1 F0 F5 F4 */
581 FXCH( ST(1) ) /* F0 F1 F5 F4 */
582 FADDP( ST0, ST(3) ) /* F1 F5 F4 */
583 FADDP( ST0, ST(1) ) /* F5 F4 */
588 FXCH( ST(1) ) /* F4 F5 */
589 FSTP_S( DST0 ) /* F5 */
594 LLBL(x86_p4_2dnrr_skip):
596 ADD_L( CONST(16), EDI )
599 JNE( LLBL(x86_p4_2dnrr_loop) )
601 LLBL(x86_p4_2dnrr_done):
614 GLOBL GLNAME( _mesa_x86_transform_points4_identity )
615 HIDDEN(_mesa_x86_transform_points4_identity)
616 GLNAME( _mesa_x86_transform_points4_identity ):
618 #define FRAME_OFFSET 12
623 MOV_L( ARG_SOURCE, ESI )
624 MOV_L( ARG_DEST, EDI )
626 MOV_L( ARG_MATRIX, EDX )
627 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
630 JZ( LLBL(x86_p4_ir_done) )
632 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
633 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
635 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
636 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
638 SHL_L( CONST(4), ECX )
639 MOV_L( REGOFF(V4F_START, ESI), ESI )
641 MOV_L( REGOFF(V4F_START, EDI), EDI )
645 JE( LLBL(x86_p4_ir_done) )
648 LLBL(x86_p4_ir_loop):
662 LLBL(x86_p4_ir_skip):
664 ADD_L( CONST(16), EDI )
667 JNE( LLBL(x86_p4_ir_loop) )
669 LLBL(x86_p4_ir_done):
676 #if defined (__ELF__) && defined (__linux__)
677 .section .note.GNU-stack,"",%progbits