1 /* $Id: x86_xform4.S,v 1.4 2006/04/17 18:58:24 krh Exp $ */
4 * Mesa 3-D graphics library
7 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
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:
16 * The above copyright notice and this permission notice shall be included
17 * in all copies or substantial portions of the Software.
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.
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.
34 #include "xform_args.h"
38 #define FP_ONE 1065353216
41 #define SRC0 REGOFF(0, ESI)
42 #define SRC1 REGOFF(4, ESI)
43 #define SRC2 REGOFF(8, ESI)
44 #define SRC3 REGOFF(12, ESI)
45 #define DST0 REGOFF(0, EDI)
46 #define DST1 REGOFF(4, EDI)
47 #define DST2 REGOFF(8, EDI)
48 #define DST3 REGOFF(12, EDI)
49 #define MAT0 REGOFF(0, EDX)
50 #define MAT1 REGOFF(4, EDX)
51 #define MAT2 REGOFF(8, EDX)
52 #define MAT3 REGOFF(12, EDX)
53 #define MAT4 REGOFF(16, EDX)
54 #define MAT5 REGOFF(20, EDX)
55 #define MAT6 REGOFF(24, EDX)
56 #define MAT7 REGOFF(28, EDX)
57 #define MAT8 REGOFF(32, EDX)
58 #define MAT9 REGOFF(36, EDX)
59 #define MAT10 REGOFF(40, EDX)
60 #define MAT11 REGOFF(44, EDX)
61 #define MAT12 REGOFF(48, EDX)
62 #define MAT13 REGOFF(52, EDX)
63 #define MAT14 REGOFF(56, EDX)
64 #define MAT15 REGOFF(60, EDX)
68 GLOBL GLNAME( _mesa_x86_transform_points4_general )
69 HIDDEN(_mesa_x86_transform_points4_general)
70 GLNAME( _mesa_x86_transform_points4_general ):
72 #define FRAME_OFFSET 8
76 MOV_L( ARG_SOURCE, ESI )
77 MOV_L( ARG_DEST, EDI )
79 MOV_L( ARG_MATRIX, EDX )
80 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
83 JZ( LLBL(x86_p4_gr_done) )
85 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
86 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
88 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
89 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
91 SHL_L( CONST(4), ECX )
92 MOV_L( REGOFF(V4F_START, ESI), ESI )
94 MOV_L( REGOFF(V4F_START, EDI), EDI )
100 FLD_S( SRC0 ) /* F4 */
102 FLD_S( SRC0 ) /* F5 F4 */
104 FLD_S( SRC0 ) /* F6 F5 F4 */
106 FLD_S( SRC0 ) /* F7 F6 F5 F4 */
109 FLD_S( SRC1 ) /* F0 F7 F6 F5 F4 */
111 FLD_S( SRC1 ) /* F1 F0 F7 F6 F5 F4 */
113 FLD_S( SRC1 ) /* F2 F1 F0 F7 F6 F5 F4 */
115 FLD_S( SRC1 ) /* F3 F2 F1 F0 F7 F6 F5 F4 */
118 FXCH( ST(3) ) /* F0 F2 F1 F3 F7 F6 F5 F4 */
119 FADDP( ST0, ST(7) ) /* F2 F1 F3 F7 F6 F5 F4 */
120 FXCH( ST(1) ) /* F1 F2 F3 F7 F6 F5 F4 */
121 FADDP( ST0, ST(5) ) /* F2 F3 F7 F6 F5 F4 */
122 FADDP( ST0, ST(3) ) /* F3 F7 F6 F5 F4 */
123 FADDP( ST0, ST(1) ) /* F7 F6 F5 F4 */
125 FLD_S( SRC2 ) /* F0 F7 F6 F5 F4 */
127 FLD_S( SRC2 ) /* F1 F0 F7 F6 F5 F4 */
129 FLD_S( SRC2 ) /* F2 F1 F0 F7 F6 F5 F4 */
131 FLD_S( SRC2 ) /* F3 F2 F1 F0 F7 F6 F5 F4 */
134 FXCH( ST(3) ) /* F0 F2 F1 F3 F7 F6 F5 F4 */
135 FADDP( ST0, ST(7) ) /* F2 F1 F3 F7 F6 F5 F4 */
136 FXCH( ST(1) ) /* F1 F2 F3 F7 F6 F5 F4 */
137 FADDP( ST0, ST(5) ) /* F2 F3 F7 F6 F5 F4 */
138 FADDP( ST0, ST(3) ) /* F3 F7 F6 F5 F4 */
139 FADDP( ST0, ST(1) ) /* F7 F6 F5 F4 */
141 FLD_S( SRC3 ) /* F0 F7 F6 F5 F4 */
143 FLD_S( SRC3 ) /* F1 F0 F7 F6 F5 F4 */
145 FLD_S( SRC3 ) /* F2 F1 F0 F7 F6 F5 F4 */
147 FLD_S( SRC3 ) /* F3 F2 F1 F0 F7 F6 F5 F4 */
150 FXCH( ST(3) ) /* F0 F2 F1 F3 F7 F6 F5 F4 */
151 FADDP( ST0, ST(7) ) /* F2 F1 F3 F7 F6 F5 F4 */
152 FXCH( ST(1) ) /* F1 F2 F3 F7 F6 F5 F4 */
153 FADDP( ST0, ST(5) ) /* F2 F3 F7 F6 F5 F4 */
154 FADDP( ST0, ST(3) ) /* F3 F7 F6 F5 F4 */
155 FADDP( ST0, ST(1) ) /* F7 F6 F5 F4 */
157 FXCH( ST(3) ) /* F4 F6 F5 F7 */
158 FSTP_S( DST0 ) /* F6 F5 F7 */
159 FXCH( ST(1) ) /* F5 F6 F7 */
160 FSTP_S( DST1 ) /* F6 F7 */
161 FSTP_S( DST2 ) /* F7 */
164 LLBL(x86_p4_gr_skip):
166 ADD_L( CONST(16), EDI )
169 JNE( LLBL(x86_p4_gr_loop) )
171 LLBL(x86_p4_gr_done):
182 GLOBL GLNAME( _mesa_x86_transform_points4_perspective )
183 HIDDEN(_mesa_x86_transform_points4_perspective)
184 GLNAME( _mesa_x86_transform_points4_perspective ):
186 #define FRAME_OFFSET 12
191 MOV_L( ARG_SOURCE, ESI )
192 MOV_L( ARG_DEST, EDI )
194 MOV_L( ARG_MATRIX, EDX )
195 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
198 JZ( LLBL(x86_p4_pr_done) )
200 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
201 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
203 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
204 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
206 SHL_L( CONST(4), ECX )
207 MOV_L( REGOFF(V4F_START, ESI), ESI )
209 MOV_L( REGOFF(V4F_START, EDI), EDI )
213 LLBL(x86_p4_pr_loop):
215 FLD_S( SRC0 ) /* F4 */
218 FLD_S( SRC1 ) /* F5 F4 */
221 FLD_S( SRC2 ) /* F0 F5 F4 */
223 FLD_S( SRC2 ) /* F1 F0 F5 F4 */
225 FLD_S( SRC2 ) /* F6 F1 F0 F5 F4 */
228 FXCH( ST(2) ) /* F0 F1 F6 F5 F4 */
229 FADDP( ST0, ST(4) ) /* F1 F6 F5 F4 */
230 FADDP( ST0, ST(2) ) /* F6 F5 F4 */
232 FLD_S( SRC3 ) /* F2 F6 F5 F4 */
235 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
238 XOR_L( CONST(-2147483648), EBX )/* change sign */
240 FXCH( ST(2) ) /* F4 F5 F6 */
241 FSTP_S( DST0 ) /* F5 F6 */
242 FSTP_S( DST1 ) /* F6 */
246 LLBL(x86_p4_pr_skip):
248 ADD_L( CONST(16), EDI )
251 JNE( LLBL(x86_p4_pr_loop) )
253 LLBL(x86_p4_pr_done):
265 GLOBL GLNAME( _mesa_x86_transform_points4_3d )
266 HIDDEN(_mesa_x86_transform_points4_3d)
267 GLNAME( _mesa_x86_transform_points4_3d ):
269 #define FRAME_OFFSET 12
274 MOV_L( ARG_SOURCE, ESI )
275 MOV_L( ARG_DEST, EDI )
277 MOV_L( ARG_MATRIX, EDX )
278 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
281 JZ( LLBL(x86_p4_3dr_done) )
283 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
284 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
286 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
287 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
289 SHL_L( CONST(4), ECX )
290 MOV_L( REGOFF(V4F_START, ESI), ESI )
292 MOV_L( REGOFF(V4F_START, EDI), EDI )
296 LLBL(x86_p4_3dr_loop):
298 FLD_S( SRC0 ) /* F4 */
300 FLD_S( SRC0 ) /* F5 F4 */
302 FLD_S( SRC0 ) /* F6 F5 F4 */
305 FLD_S( SRC1 ) /* F0 F6 F5 F4 */
307 FLD_S( SRC1 ) /* F1 F0 F6 F5 F4 */
309 FLD_S( SRC1 ) /* F2 F1 F0 F6 F5 F4 */
312 FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */
313 FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */
314 FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */
315 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
317 FLD_S( SRC2 ) /* F0 F6 F5 F4 */
319 FLD_S( SRC2 ) /* F1 F0 F6 F5 F4 */
321 FLD_S( SRC2 ) /* F2 F1 F0 F6 F5 F4 */
324 FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */
325 FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */
326 FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */
327 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
329 FLD_S( SRC3 ) /* F0 F6 F5 F4 */
331 FLD_S( SRC3 ) /* F1 F0 F6 F5 F4 */
333 FLD_S( SRC3 ) /* F2 F1 F0 F6 F5 F4 */
336 FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */
337 FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */
338 FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */
339 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
343 FXCH( ST(2) ) /* F4 F5 F6 */
344 FSTP_S( DST0 ) /* F5 F6 */
345 FSTP_S( DST1 ) /* F6 */
349 LLBL(x86_p4_3dr_skip):
351 ADD_L( CONST(16), EDI )
354 JNE( LLBL(x86_p4_3dr_loop) )
356 LLBL(x86_p4_3dr_done):
368 GLOBL GLNAME(_mesa_x86_transform_points4_3d_no_rot)
369 HIDDEN(_mesa_x86_transform_points4_3d_no_rot)
370 GLNAME(_mesa_x86_transform_points4_3d_no_rot):
372 #define FRAME_OFFSET 12
377 MOV_L( ARG_SOURCE, ESI )
378 MOV_L( ARG_DEST, EDI )
380 MOV_L( ARG_MATRIX, EDX )
381 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
384 JZ( LLBL(x86_p4_3dnrr_done) )
386 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
387 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
389 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
390 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
392 SHL_L( CONST(4), ECX )
393 MOV_L( REGOFF(V4F_START, ESI), ESI )
395 MOV_L( REGOFF(V4F_START, EDI), EDI )
399 LLBL(x86_p4_3dnrr_loop):
401 FLD_S( SRC0 ) /* F4 */
404 FLD_S( SRC1 ) /* F5 F4 */
407 FLD_S( SRC2 ) /* F6 F5 F4 */
410 FLD_S( SRC3 ) /* F0 F6 F5 F4 */
412 FLD_S( SRC3 ) /* F1 F0 F6 F5 F4 */
414 FLD_S( SRC3 ) /* F2 F1 F0 F6 F5 F4 */
417 FXCH( ST(2) ) /* F0 F1 F2 F6 F5 F4 */
418 FADDP( ST0, ST(5) ) /* F1 F2 F6 F5 F4 */
419 FADDP( ST0, ST(3) ) /* F2 F6 F5 F4 */
420 FADDP( ST0, ST(1) ) /* F6 F5 F4 */
424 FXCH( ST(2) ) /* F4 F5 F6 */
425 FSTP_S( DST0 ) /* F5 F6 */
426 FSTP_S( DST1 ) /* F6 */
430 LLBL(x86_p4_3dnrr_skip):
432 ADD_L( CONST(16), EDI )
435 JNE( LLBL(x86_p4_3dnrr_loop) )
437 LLBL(x86_p4_3dnrr_done):
449 GLOBL GLNAME( _mesa_x86_transform_points4_2d )
450 HIDDEN(_mesa_x86_transform_points4_2d)
451 GLNAME( _mesa_x86_transform_points4_2d ):
453 #define FRAME_OFFSET 16
459 MOV_L( ARG_SOURCE, ESI )
460 MOV_L( ARG_DEST, EDI )
462 MOV_L( ARG_MATRIX, EDX )
463 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
466 JZ( LLBL(x86_p4_2dr_done) )
468 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
469 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
471 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
472 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
474 SHL_L( CONST(4), ECX )
475 MOV_L( REGOFF(V4F_START, ESI), ESI )
477 MOV_L( REGOFF(V4F_START, EDI), EDI )
481 LLBL(x86_p4_2dr_loop):
483 FLD_S( SRC0 ) /* F4 */
485 FLD_S( SRC0 ) /* F5 F4 */
488 FLD_S( SRC1 ) /* F0 F5 F4 */
490 FLD_S( SRC1 ) /* F1 F0 F5 F4 */
493 FXCH( ST(1) ) /* F0 F1 F5 F4 */
494 FADDP( ST0, ST(3) ) /* F1 F5 F4 */
495 FADDP( ST0, ST(1) ) /* F5 F4 */
497 FLD_S( SRC3 ) /* F0 F5 F4 */
499 FLD_S( SRC3 ) /* F1 F0 F5 F4 */
502 FXCH( ST(1) ) /* F0 F1 F5 F4 */
503 FADDP( ST0, ST(3) ) /* F1 F5 F4 */
504 FADDP( ST0, ST(1) ) /* F5 F4 */
509 FXCH( ST(1) ) /* F4 F5 */
510 FSTP_S( DST0 ) /* F5 */
515 LLBL(x86_p4_2dr_skip):
517 ADD_L( CONST(16), EDI )
520 JNE( LLBL(x86_p4_2dr_loop) )
522 LLBL(x86_p4_2dr_done):
535 GLOBL GLNAME( _mesa_x86_transform_points4_2d_no_rot )
536 HIDDEN(_mesa_x86_transform_points4_2d_no_rot)
537 GLNAME( _mesa_x86_transform_points4_2d_no_rot ):
539 #define FRAME_OFFSET 16
545 MOV_L( ARG_SOURCE, ESI )
546 MOV_L( ARG_DEST, EDI )
548 MOV_L( ARG_MATRIX, EDX )
549 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
552 JZ( LLBL(x86_p4_2dnrr_done) )
554 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
555 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
557 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
558 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
560 SHL_L( CONST(4), ECX )
561 MOV_L( REGOFF(V4F_START, ESI), ESI )
563 MOV_L( REGOFF(V4F_START, EDI), EDI )
567 LLBL(x86_p4_2dnrr_loop):
569 FLD_S( SRC0 ) /* F4 */
572 FLD_S( SRC1 ) /* F5 F4 */
575 FLD_S( SRC3 ) /* F0 F5 F4 */
577 FLD_S( SRC3 ) /* F1 F0 F5 F4 */
580 FXCH( ST(1) ) /* F0 F1 F5 F4 */
581 FADDP( ST0, ST(3) ) /* F1 F5 F4 */
582 FADDP( ST0, ST(1) ) /* F5 F4 */
587 FXCH( ST(1) ) /* F4 F5 */
588 FSTP_S( DST0 ) /* F5 */
593 LLBL(x86_p4_2dnrr_skip):
595 ADD_L( CONST(16), EDI )
598 JNE( LLBL(x86_p4_2dnrr_loop) )
600 LLBL(x86_p4_2dnrr_done):
613 GLOBL GLNAME( _mesa_x86_transform_points4_identity )
614 HIDDEN(_mesa_x86_transform_points4_identity)
615 GLNAME( _mesa_x86_transform_points4_identity ):
617 #define FRAME_OFFSET 12
622 MOV_L( ARG_SOURCE, ESI )
623 MOV_L( ARG_DEST, EDI )
625 MOV_L( ARG_MATRIX, EDX )
626 MOV_L( REGOFF(V4F_COUNT, ESI), ECX )
629 JZ( LLBL(x86_p4_ir_done) )
631 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX )
632 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) )
634 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) )
635 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) )
637 SHL_L( CONST(4), ECX )
638 MOV_L( REGOFF(V4F_START, ESI), ESI )
640 MOV_L( REGOFF(V4F_START, EDI), EDI )
644 JE( LLBL(x86_p4_ir_done) )
647 LLBL(x86_p4_ir_loop):
661 LLBL(x86_p4_ir_skip):
663 ADD_L( CONST(16), EDI )
666 JNE( LLBL(x86_p4_ir_loop) )
668 LLBL(x86_p4_ir_done):
675 #if defined (__ELF__) && defined (__linux__)
676 .section .note.GNU-stack,"",%progbits