projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
Merge branch 'gallium-edgeflags'
[mesa.git] / src / gallium / auxiliary / tgsi / tgsi_sse2.c
1 /**************************************************************************
2 *
3 * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 #include "pipe/p_config.h"
29
30 #if defined(PIPE_ARCH_X86)
31
32 #include "util/u_debug.h"
33 #include "pipe/p_shader_tokens.h"
34 #include "util/u_math.h"
35 #include "util/u_memory.h"
36 #if defined(PIPE_ARCH_SSE)
37 #include "util/u_sse.h"
38 #endif
39 #include "tgsi/tgsi_info.h"
40 #include "tgsi/tgsi_parse.h"
41 #include "tgsi/tgsi_util.h"
42 #include "tgsi/tgsi_dump.h"
43 #include "tgsi/tgsi_exec.h"
44 #include "tgsi/tgsi_sse2.h"
45
46 #include "rtasm/rtasm_x86sse.h"
47
48 /* for 1/sqrt()
49 *
50 * This costs about 100fps (close to 10%) in gears:
51 */
52 #define HIGH_PRECISION 1
53
54 #define FAST_MATH 1
55
56
57 #define FOR_EACH_CHANNEL( CHAN )\
58 for (CHAN = 0; CHAN < NUM_CHANNELS; CHAN++)
59
60 #define IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
61 ((INST).Dst[0].Register.WriteMask & (1 << (CHAN)))
62
63 #define IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
64 if (IS_DST0_CHANNEL_ENABLED( INST, CHAN ))
65
66 #define FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\
67 FOR_EACH_CHANNEL( CHAN )\
68 IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )
69
70 #define CHAN_X 0
71 #define CHAN_Y 1
72 #define CHAN_Z 2
73 #define CHAN_W 3
74
75 #define TEMP_ONE_I TGSI_EXEC_TEMP_ONE_I
76 #define TEMP_ONE_C TGSI_EXEC_TEMP_ONE_C
77
78 #define TEMP_R0 TGSI_EXEC_TEMP_R0
79 #define TEMP_ADDR TGSI_EXEC_TEMP_ADDR
80 #define TEMP_EXEC_MASK_I TGSI_EXEC_MASK_I
81 #define TEMP_EXEC_MASK_C TGSI_EXEC_MASK_C
82
83
84 /**
85 * X86 utility functions.
86 */
87
88 static struct x86_reg
89 make_xmm(
90 unsigned xmm )
91 {
92 return x86_make_reg(
93 file_XMM,
94 (enum x86_reg_name) xmm );
95 }
96
97 /**
98 * X86 register mapping helpers.
99 */
100
101 static struct x86_reg
102 get_const_base( void )
103 {
104 return x86_make_reg(
105 file_REG32,
106 reg_AX );
107 }
108
109 static struct x86_reg
110 get_machine_base( void )
111 {
112 return x86_make_reg(
113 file_REG32,
114 reg_CX );
115 }
116
117 static struct x86_reg
118 get_input_base( void )
119 {
120 return x86_make_disp(
121 get_machine_base(),
122 Offset(struct tgsi_exec_machine, Inputs) );
123 }
124
125 static struct x86_reg
126 get_output_base( void )
127 {
128 return x86_make_disp(
129 get_machine_base(),
130 Offset(struct tgsi_exec_machine, Outputs) );
131 }
132
133 static struct x86_reg
134 get_temp_base( void )
135 {
136 return x86_make_disp(
137 get_machine_base(),
138 Offset(struct tgsi_exec_machine, Temps) );
139 }
140
141 static struct x86_reg
142 get_coef_base( void )
143 {
144 return x86_make_reg(
145 file_REG32,
146 reg_BX );
147 }
148
149 static struct x86_reg
150 get_sampler_base( void )
151 {
152 return x86_make_reg(
153 file_REG32,
154 reg_DI );
155 }
156
157 static struct x86_reg
158 get_immediate_base( void )
159 {
160 return x86_make_reg(
161 file_REG32,
162 reg_DX );
163 }
164
165
166 /**
167 * Data access helpers.
168 */
169
170
171 static struct x86_reg
172 get_immediate(
173 unsigned vec,
174 unsigned chan )
175 {
176 return x86_make_disp(
177 get_immediate_base(),
178 (vec * 4 + chan) * 4 );
179 }
180
181 static struct x86_reg
182 get_const(
183 unsigned vec,
184 unsigned chan )
185 {
186 return x86_make_disp(
187 get_const_base(),
188 (vec * 4 + chan) * 4 );
189 }
190
191 static struct x86_reg
192 get_sampler_ptr(
193 unsigned unit )
194 {
195 return x86_make_disp(
196 get_sampler_base(),
197 unit * sizeof( struct tgsi_sampler * ) );
198 }
199
200 static struct x86_reg
201 get_input(
202 unsigned vec,
203 unsigned chan )
204 {
205 return x86_make_disp(
206 get_input_base(),
207 (vec * 4 + chan) * 16 );
208 }
209
210 static struct x86_reg
211 get_output(
212 unsigned vec,
213 unsigned chan )
214 {
215 return x86_make_disp(
216 get_output_base(),
217 (vec * 4 + chan) * 16 );
218 }
219
220 static struct x86_reg
221 get_temp(
222 unsigned vec,
223 unsigned chan )
224 {
225 return x86_make_disp(
226 get_temp_base(),
227 (vec * 4 + chan) * 16 );
228 }
229
230 static struct x86_reg
231 get_coef(
232 unsigned vec,
233 unsigned chan,
234 unsigned member )
235 {
236 return x86_make_disp(
237 get_coef_base(),
238 ((vec * 3 + member) * 4 + chan) * 4 );
239 }
240
241
242 static void
243 emit_ret(
244 struct x86_function *func )
245 {
246 x86_ret( func );
247 }
248
249
250 /**
251 * Data fetch helpers.
252 */
253
254 /**
255 * Copy a shader constant to xmm register
256 * \param xmm the destination xmm register
257 * \param vec the src const buffer index
258 * \param chan src channel to fetch (X, Y, Z or W)
259 */
260 static void
261 emit_const(
262 struct x86_function *func,
263 uint xmm,
264 int vec,
265 uint chan,
266 uint indirect,
267 uint indirectFile,
268 int indirectIndex )
269 {
270 if (indirect) {
271 /* 'vec' is the offset from the address register's value.
272 * We're loading CONST[ADDR+vec] into an xmm register.
273 */
274 struct x86_reg r0 = get_immediate_base();
275 struct x86_reg r1 = get_coef_base();
276 uint i;
277
278 assert( indirectFile == TGSI_FILE_ADDRESS );
279 assert( indirectIndex == 0 );
280 assert( r0.mod == mod_REG );
281 assert( r1.mod == mod_REG );
282
283 x86_push( func, r0 );
284 x86_push( func, r1 );
285
286 /*
287 * Loop over the four pixels or vertices in the quad.
288 * Get the value of the address (offset) register for pixel/vertex[i],
289 * add it to the src offset and index into the constant buffer.
290 * Note that we're working on SOA data.
291 * If any of the pixel/vertex execution channels are unused their
292 * values will be garbage. It's very important that we don't use
293 * those garbage values as indexes into the constant buffer since
294 * that'll cause segfaults.
295 * The solution is to bitwise-AND the offset with the execution mask
296 * register whose values are either 0 or ~0.
297 * The caller must setup the execution mask register to indicate
298 * which channels are valid/alive before running the shader.
299 * The execution mask will also figure into loops and conditionals
300 * someday.
301 */
302 for (i = 0; i < QUAD_SIZE; i++) {
303 /* r1 = address register[i] */
304 x86_mov( func, r1, x86_make_disp( get_temp( TEMP_ADDR, CHAN_X ), i * 4 ) );
305 /* r0 = execution mask[i] */
306 x86_mov( func, r0, x86_make_disp( get_temp( TEMP_EXEC_MASK_I, TEMP_EXEC_MASK_C ), i * 4 ) );
307 /* r1 = r1 & r0 */
308 x86_and( func, r1, r0 );
309 /* r0 = 'vec', the offset */
310 x86_lea( func, r0, get_const( vec, chan ) );
311
312 /* Quick hack to multiply r1 by 16 -- need to add SHL to rtasm.
313 */
314 x86_add( func, r1, r1 );
315 x86_add( func, r1, r1 );
316 x86_add( func, r1, r1 );
317 x86_add( func, r1, r1 );
318
319 x86_add( func, r0, r1 ); /* r0 = r0 + r1 */
320 x86_mov( func, r1, x86_deref( r0 ) );
321 x86_mov( func, x86_make_disp( get_temp( TEMP_R0, CHAN_X ), i * 4 ), r1 );
322 }
323
324 x86_pop( func, r1 );
325 x86_pop( func, r0 );
326
327 sse_movaps(
328 func,
329 make_xmm( xmm ),
330 get_temp( TEMP_R0, CHAN_X ) );
331 }
332 else {
333 /* 'vec' is the index into the src register file, such as TEMP[vec] */
334 assert( vec >= 0 );
335
336 sse_movss(
337 func,
338 make_xmm( xmm ),
339 get_const( vec, chan ) );
340 sse_shufps(
341 func,
342 make_xmm( xmm ),
343 make_xmm( xmm ),
344 SHUF( 0, 0, 0, 0 ) );
345 }
346 }
347
348 static void
349 emit_immediate(
350 struct x86_function *func,
351 unsigned xmm,
352 unsigned vec,
353 unsigned chan )
354 {
355 sse_movss(
356 func,
357 make_xmm( xmm ),
358 get_immediate( vec, chan ) );
359 sse_shufps(
360 func,
361 make_xmm( xmm ),
362 make_xmm( xmm ),
363 SHUF( 0, 0, 0, 0 ) );
364 }
365
366
367 /**
368 * Copy a shader input to xmm register
369 * \param xmm the destination xmm register
370 * \param vec the src input attrib
371 * \param chan src channel to fetch (X, Y, Z or W)
372 */
373 static void
374 emit_inputf(
375 struct x86_function *func,
376 unsigned xmm,
377 unsigned vec,
378 unsigned chan )
379 {
380 sse_movups(
381 func,
382 make_xmm( xmm ),
383 get_input( vec, chan ) );
384 }
385
386 /**
387 * Store an xmm register to a shader output
388 * \param xmm the source xmm register
389 * \param vec the dest output attrib
390 * \param chan src dest channel to store (X, Y, Z or W)
391 */
392 static void
393 emit_output(
394 struct x86_function *func,
395 unsigned xmm,
396 unsigned vec,
397 unsigned chan )
398 {
399 sse_movups(
400 func,
401 get_output( vec, chan ),
402 make_xmm( xmm ) );
403 }
404
405 /**
406 * Copy a shader temporary to xmm register
407 * \param xmm the destination xmm register
408 * \param vec the src temp register
409 * \param chan src channel to fetch (X, Y, Z or W)
410 */
411 static void
412 emit_tempf(
413 struct x86_function *func,
414 unsigned xmm,
415 unsigned vec,
416 unsigned chan )
417 {
418 sse_movaps(
419 func,
420 make_xmm( xmm ),
421 get_temp( vec, chan ) );
422 }
423
424 /**
425 * Load an xmm register with an input attrib coefficient (a0, dadx or dady)
426 * \param xmm the destination xmm register
427 * \param vec the src input/attribute coefficient index
428 * \param chan src channel to fetch (X, Y, Z or W)
429 * \param member 0=a0, 1=dadx, 2=dady
430 */
431 static void
432 emit_coef(
433 struct x86_function *func,
434 unsigned xmm,
435 unsigned vec,
436 unsigned chan,
437 unsigned member )
438 {
439 sse_movss(
440 func,
441 make_xmm( xmm ),
442 get_coef( vec, chan, member ) );
443 sse_shufps(
444 func,
445 make_xmm( xmm ),
446 make_xmm( xmm ),
447 SHUF( 0, 0, 0, 0 ) );
448 }
449
450 /**
451 * Data store helpers.
452 */
453
454 static void
455 emit_inputs(
456 struct x86_function *func,
457 unsigned xmm,
458 unsigned vec,
459 unsigned chan )
460 {
461 sse_movups(
462 func,
463 get_input( vec, chan ),
464 make_xmm( xmm ) );
465 }
466
467 static void
468 emit_temps(
469 struct x86_function *func,
470 unsigned xmm,
471 unsigned vec,
472 unsigned chan )
473 {
474 sse_movaps(
475 func,
476 get_temp( vec, chan ),
477 make_xmm( xmm ) );
478 }
479
480 static void
481 emit_addrs(
482 struct x86_function *func,
483 unsigned xmm,
484 unsigned vec,
485 unsigned chan )
486 {
487 assert( vec == 0 );
488
489 emit_temps(
490 func,
491 xmm,
492 vec + TGSI_EXEC_TEMP_ADDR,
493 chan );
494 }
495
496 /**
497 * Coefficent fetch helpers.
498 */
499
500 static void
501 emit_coef_a0(
502 struct x86_function *func,
503 unsigned xmm,
504 unsigned vec,
505 unsigned chan )
506 {
507 emit_coef(
508 func,
509 xmm,
510 vec,
511 chan,
512 0 );
513 }
514
515 static void
516 emit_coef_dadx(
517 struct x86_function *func,
518 unsigned xmm,
519 unsigned vec,
520 unsigned chan )
521 {
522 emit_coef(
523 func,
524 xmm,
525 vec,
526 chan,
527 1 );
528 }
529
530 static void
531 emit_coef_dady(
532 struct x86_function *func,
533 unsigned xmm,
534 unsigned vec,
535 unsigned chan )
536 {
537 emit_coef(
538 func,
539 xmm,
540 vec,
541 chan,
542 2 );
543 }
544
545 /**
546 * Function call helpers.
547 */
548
549 /**
550 * NOTE: In gcc, if the destination uses the SSE intrinsics, then it must be
551 * defined with __attribute__((force_align_arg_pointer)), as we do not guarantee
552 * that the stack pointer is 16 byte aligned, as expected.
553 */
554 static void
555 emit_func_call(
556 struct x86_function *func,
557 unsigned xmm_save_mask,
558 const struct x86_reg *arg,
559 unsigned nr_args,
560 void (PIPE_CDECL *code)() )
561 {
562 struct x86_reg ecx = x86_make_reg( file_REG32, reg_CX );
563 unsigned i, n;
564
565 x86_push(
566 func,
567 x86_make_reg( file_REG32, reg_AX) );
568 x86_push(
569 func,
570 x86_make_reg( file_REG32, reg_CX) );
571 x86_push(
572 func,
573 x86_make_reg( file_REG32, reg_DX) );
574
575 /* Store XMM regs to the stack
576 */
577 for(i = 0, n = 0; i < 8; ++i)
578 if(xmm_save_mask & (1 << i))
579 ++n;
580
581 x86_sub_imm(
582 func,
583 x86_make_reg( file_REG32, reg_SP ),
584 n*16);
585
586 for(i = 0, n = 0; i < 8; ++i)
587 if(xmm_save_mask & (1 << i)) {
588 sse_movups(
589 func,
590 x86_make_disp( x86_make_reg( file_REG32, reg_SP ), n*16 ),
591 make_xmm( i ) );
592 ++n;
593 }
594
595 for (i = 0; i < nr_args; i++) {
596 /* Load the address of the buffer we use for passing arguments and
597 * receiving results:
598 */
599 x86_lea(
600 func,
601 ecx,
602 arg[i] );
603
604 /* Push actual function arguments (currently just the pointer to
605 * the buffer above), and call the function:
606 */
607 x86_push( func, ecx );
608 }
609
610 x86_mov_reg_imm( func, ecx, (unsigned long) code );
611 x86_call( func, ecx );
612
613 /* Pop the arguments (or just add an immediate to esp)
614 */
615 for (i = 0; i < nr_args; i++) {
616 x86_pop(func, ecx );
617 }
618
619 /* Pop the saved XMM regs:
620 */
621 for(i = 0, n = 0; i < 8; ++i)
622 if(xmm_save_mask & (1 << i)) {
623 sse_movups(
624 func,
625 make_xmm( i ),
626 x86_make_disp( x86_make_reg( file_REG32, reg_SP ), n*16 ) );
627 ++n;
628 }
629
630 x86_add_imm(
631 func,
632 x86_make_reg( file_REG32, reg_SP ),
633 n*16);
634
635 /* Restore GP registers in a reverse order.
636 */
637 x86_pop(
638 func,
639 x86_make_reg( file_REG32, reg_DX) );
640 x86_pop(
641 func,
642 x86_make_reg( file_REG32, reg_CX) );
643 x86_pop(
644 func,
645 x86_make_reg( file_REG32, reg_AX) );
646 }
647
648 static void
649 emit_func_call_dst_src1(
650 struct x86_function *func,
651 unsigned xmm_save,
652 unsigned xmm_dst,
653 unsigned xmm_src0,
654 void (PIPE_CDECL *code)() )
655 {
656 struct x86_reg store = get_temp( TEMP_R0, 0 );
657 unsigned xmm_mask = ((1 << xmm_save) - 1) & ~(1 << xmm_dst);
658
659 /* Store our input parameters (in xmm regs) to the buffer we use
660 * for passing arguments. We will pass a pointer to this buffer as
661 * the actual function argument.
662 */
663 sse_movaps(
664 func,
665 store,
666 make_xmm( xmm_src0 ) );
667
668 emit_func_call( func,
669 xmm_mask,
670 &store,
671 1,
672 code );
673
674 sse_movaps(
675 func,
676 make_xmm( xmm_dst ),
677 store );
678 }
679
680
681 static void
682 emit_func_call_dst_src2(
683 struct x86_function *func,
684 unsigned xmm_save,
685 unsigned xmm_dst,
686 unsigned xmm_src0,
687 unsigned xmm_src1,
688 void (PIPE_CDECL *code)() )
689 {
690 struct x86_reg store = get_temp( TEMP_R0, 0 );
691 unsigned xmm_mask = ((1 << xmm_save) - 1) & ~(1 << xmm_dst);
692
693 /* Store two inputs to parameter buffer.
694 */
695 sse_movaps(
696 func,
697 store,
698 make_xmm( xmm_src0 ) );
699
700 sse_movaps(
701 func,
702 x86_make_disp( store, 4 * sizeof(float) ),
703 make_xmm( xmm_src1 ) );
704
705
706 /* Emit the call
707 */
708 emit_func_call( func,
709 xmm_mask,
710 &store,
711 1,
712 code );
713
714 /* Retrieve the results:
715 */
716 sse_movaps(
717 func,
718 make_xmm( xmm_dst ),
719 store );
720 }
721
722
723
724
725
726 #if defined(PIPE_ARCH_SSE)
727
728 /*
729 * Fast SSE2 implementation of special math functions.
730 */
731
732 #define POLY0(x, c0) _mm_set1_ps(c0)
733 #define POLY1(x, c0, c1) _mm_add_ps(_mm_mul_ps(POLY0(x, c1), x), _mm_set1_ps(c0))
734 #define POLY2(x, c0, c1, c2) _mm_add_ps(_mm_mul_ps(POLY1(x, c1, c2), x), _mm_set1_ps(c0))
735 #define POLY3(x, c0, c1, c2, c3) _mm_add_ps(_mm_mul_ps(POLY2(x, c1, c2, c3), x), _mm_set1_ps(c0))
736 #define POLY4(x, c0, c1, c2, c3, c4) _mm_add_ps(_mm_mul_ps(POLY3(x, c1, c2, c3, c4), x), _mm_set1_ps(c0))
737 #define POLY5(x, c0, c1, c2, c3, c4, c5) _mm_add_ps(_mm_mul_ps(POLY4(x, c1, c2, c3, c4, c5), x), _mm_set1_ps(c0))
738
739 #define EXP_POLY_DEGREE 3
740 #define LOG_POLY_DEGREE 5
741
742 /**
743 * See http://www.devmaster.net/forums/showthread.php?p=43580
744 */
745 static INLINE __m128
746 exp2f4(__m128 x)
747 {
748 __m128i ipart;
749 __m128 fpart, expipart, expfpart;
750
751 x = _mm_min_ps(x, _mm_set1_ps( 129.00000f));
752 x = _mm_max_ps(x, _mm_set1_ps(-126.99999f));
753
754 /* ipart = int(x - 0.5) */
755 ipart = _mm_cvtps_epi32(_mm_sub_ps(x, _mm_set1_ps(0.5f)));
756
757 /* fpart = x - ipart */
758 fpart = _mm_sub_ps(x, _mm_cvtepi32_ps(ipart));
759
760 /* expipart = (float) (1 << ipart) */
761 expipart = _mm_castsi128_ps(_mm_slli_epi32(_mm_add_epi32(ipart, _mm_set1_epi32(127)), 23));
762
763 /* minimax polynomial fit of 2**x, in range [-0.5, 0.5[ */
764 #if EXP_POLY_DEGREE == 5
765 expfpart = POLY5(fpart, 9.9999994e-1f, 6.9315308e-1f, 2.4015361e-1f, 5.5826318e-2f, 8.9893397e-3f, 1.8775767e-3f);
766 #elif EXP_POLY_DEGREE == 4
767 expfpart = POLY4(fpart, 1.0000026f, 6.9300383e-1f, 2.4144275e-1f, 5.2011464e-2f, 1.3534167e-2f);
768 #elif EXP_POLY_DEGREE == 3
769 expfpart = POLY3(fpart, 9.9992520e-1f, 6.9583356e-1f, 2.2606716e-1f, 7.8024521e-2f);
770 #elif EXP_POLY_DEGREE == 2
771 expfpart = POLY2(fpart, 1.0017247f, 6.5763628e-1f, 3.3718944e-1f);
772 #else
773 #error
774 #endif
775
776 return _mm_mul_ps(expipart, expfpart);
777 }
778
779
780 /**
781 * See http://www.devmaster.net/forums/showthread.php?p=43580
782 */
783 static INLINE __m128
784 log2f4(__m128 x)
785 {
786 __m128i expmask = _mm_set1_epi32(0x7f800000);
787 __m128i mantmask = _mm_set1_epi32(0x007fffff);
788 __m128 one = _mm_set1_ps(1.0f);
789
790 __m128i i = _mm_castps_si128(x);
791
792 /* exp = (float) exponent(x) */
793 __m128 exp = _mm_cvtepi32_ps(_mm_sub_epi32(_mm_srli_epi32(_mm_and_si128(i, expmask), 23), _mm_set1_epi32(127)));
794
795 /* mant = (float) mantissa(x) */
796 __m128 mant = _mm_or_ps(_mm_castsi128_ps(_mm_and_si128(i, mantmask)), one);
797
798 __m128 logmant;
799
800 /* Minimax polynomial fit of log2(x)/(x - 1), for x in range [1, 2[
801 * These coefficients can be generate with
802 * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
803 */
804 #if LOG_POLY_DEGREE == 6
805 logmant = POLY5(mant, 3.11578814719469302614f, -3.32419399085241980044f, 2.59883907202499966007f, -1.23152682416275988241f, 0.318212422185251071475f, -0.0344359067839062357313f);
806 #elif LOG_POLY_DEGREE == 5
807 logmant = POLY4(mant, 2.8882704548164776201f, -2.52074962577807006663f, 1.48116647521213171641f, -0.465725644288844778798f, 0.0596515482674574969533f);
808 #elif LOG_POLY_DEGREE == 4
809 logmant = POLY3(mant, 2.61761038894603480148f, -1.75647175389045657003f, 0.688243882994381274313f, -0.107254423828329604454f);
810 #elif LOG_POLY_DEGREE == 3
811 logmant = POLY2(mant, 2.28330284476918490682f, -1.04913055217340124191f, 0.204446009836232697516f);
812 #else
813 #error
814 #endif
815
816 /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
817 logmant = _mm_mul_ps(logmant, _mm_sub_ps(mant, one));
818
819 return _mm_add_ps(logmant, exp);
820 }
821
822
823 static INLINE __m128
824 powf4(__m128 x, __m128 y)
825 {
826 return exp2f4(_mm_mul_ps(log2f4(x), y));
827 }
828
829 #endif /* PIPE_ARCH_SSE */
830
831
832
833 /**
834 * Low-level instruction translators.
835 */
836
837 static void
838 emit_abs(
839 struct x86_function *func,
840 unsigned xmm )
841 {
842 sse_andps(
843 func,
844 make_xmm( xmm ),
845 get_temp(
846 TGSI_EXEC_TEMP_7FFFFFFF_I,
847 TGSI_EXEC_TEMP_7FFFFFFF_C ) );
848 }
849
850 static void
851 emit_add(
852 struct x86_function *func,
853 unsigned xmm_dst,
854 unsigned xmm_src )
855 {
856 sse_addps(
857 func,
858 make_xmm( xmm_dst ),
859 make_xmm( xmm_src ) );
860 }
861
862 static void PIPE_CDECL
863 cos4f(
864 float *store )
865 {
866 store[0] = cosf( store[0] );
867 store[1] = cosf( store[1] );
868 store[2] = cosf( store[2] );
869 store[3] = cosf( store[3] );
870 }
871
872 static void
873 emit_cos(
874 struct x86_function *func,
875 unsigned xmm_save,
876 unsigned xmm_dst )
877 {
878 emit_func_call_dst_src1(
879 func,
880 xmm_save,
881 xmm_dst,
882 xmm_dst,
883 cos4f );
884 }
885
886 static void PIPE_CDECL
887 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_SSE)
888 __attribute__((force_align_arg_pointer))
889 #endif
890 ex24f(
891 float *store )
892 {
893 #if defined(PIPE_ARCH_SSE)
894 _mm_store_ps(&store[0], exp2f4( _mm_load_ps(&store[0]) ));
895 #else
896 store[0] = util_fast_exp2( store[0] );
897 store[1] = util_fast_exp2( store[1] );
898 store[2] = util_fast_exp2( store[2] );
899 store[3] = util_fast_exp2( store[3] );
900 #endif
901 }
902
903 static void
904 emit_ex2(
905 struct x86_function *func,
906 unsigned xmm_save,
907 unsigned xmm_dst )
908 {
909 emit_func_call_dst_src1(
910 func,
911 xmm_save,
912 xmm_dst,
913 xmm_dst,
914 ex24f );
915 }
916
917 static void
918 emit_f2it(
919 struct x86_function *func,
920 unsigned xmm )
921 {
922 sse2_cvttps2dq(
923 func,
924 make_xmm( xmm ),
925 make_xmm( xmm ) );
926 }
927
928 static void
929 emit_i2f(
930 struct x86_function *func,
931 unsigned xmm )
932 {
933 sse2_cvtdq2ps(
934 func,
935 make_xmm( xmm ),
936 make_xmm( xmm ) );
937 }
938
939 static void PIPE_CDECL
940 flr4f(
941 float *store )
942 {
943 store[0] = floorf( store[0] );
944 store[1] = floorf( store[1] );
945 store[2] = floorf( store[2] );
946 store[3] = floorf( store[3] );
947 }
948
949 static void
950 emit_flr(
951 struct x86_function *func,
952 unsigned xmm_save,
953 unsigned xmm_dst )
954 {
955 emit_func_call_dst_src1(
956 func,
957 xmm_save,
958 xmm_dst,
959 xmm_dst,
960 flr4f );
961 }
962
963 static void PIPE_CDECL
964 frc4f(
965 float *store )
966 {
967 store[0] -= floorf( store[0] );
968 store[1] -= floorf( store[1] );
969 store[2] -= floorf( store[2] );
970 store[3] -= floorf( store[3] );
971 }
972
973 static void
974 emit_frc(
975 struct x86_function *func,
976 unsigned xmm_save,
977 unsigned xmm_dst )
978 {
979 emit_func_call_dst_src1(
980 func,
981 xmm_save,
982 xmm_dst,
983 xmm_dst,
984 frc4f );
985 }
986
987 static void PIPE_CDECL
988 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_SSE)
989 __attribute__((force_align_arg_pointer))
990 #endif
991 lg24f(
992 float *store )
993 {
994 #if defined(PIPE_ARCH_SSE)
995 _mm_store_ps(&store[0], log2f4( _mm_load_ps(&store[0]) ));
996 #else
997 store[0] = util_fast_log2( store[0] );
998 store[1] = util_fast_log2( store[1] );
999 store[2] = util_fast_log2( store[2] );
1000 store[3] = util_fast_log2( store[3] );
1001 #endif
1002 }
1003
1004 static void
1005 emit_lg2(
1006 struct x86_function *func,
1007 unsigned xmm_save,
1008 unsigned xmm_dst )
1009 {
1010 emit_func_call_dst_src1(
1011 func,
1012 xmm_save,
1013 xmm_dst,
1014 xmm_dst,
1015 lg24f );
1016 }
1017
1018 static void
1019 emit_MOV(
1020 struct x86_function *func,
1021 unsigned xmm_dst,
1022 unsigned xmm_src )
1023 {
1024 sse_movups(
1025 func,
1026 make_xmm( xmm_dst ),
1027 make_xmm( xmm_src ) );
1028 }
1029
1030 static void
1031 emit_mul (struct x86_function *func,
1032 unsigned xmm_dst,
1033 unsigned xmm_src)
1034 {
1035 sse_mulps(
1036 func,
1037 make_xmm( xmm_dst ),
1038 make_xmm( xmm_src ) );
1039 }
1040
1041 static void
1042 emit_neg(
1043 struct x86_function *func,
1044 unsigned xmm )
1045 {
1046 sse_xorps(
1047 func,
1048 make_xmm( xmm ),
1049 get_temp(
1050 TGSI_EXEC_TEMP_80000000_I,
1051 TGSI_EXEC_TEMP_80000000_C ) );
1052 }
1053
1054 static void PIPE_CDECL
1055 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_SSE)
1056 __attribute__((force_align_arg_pointer))
1057 #endif
1058 pow4f(
1059 float *store )
1060 {
1061 #if defined(PIPE_ARCH_SSE)
1062 _mm_store_ps(&store[0], powf4( _mm_load_ps(&store[0]), _mm_load_ps(&store[4]) ));
1063 #else
1064 store[0] = util_fast_pow( store[0], store[4] );
1065 store[1] = util_fast_pow( store[1], store[5] );
1066 store[2] = util_fast_pow( store[2], store[6] );
1067 store[3] = util_fast_pow( store[3], store[7] );
1068 #endif
1069 }
1070
1071 static void
1072 emit_pow(
1073 struct x86_function *func,
1074 unsigned xmm_save,
1075 unsigned xmm_dst,
1076 unsigned xmm_src0,
1077 unsigned xmm_src1 )
1078 {
1079 emit_func_call_dst_src2(
1080 func,
1081 xmm_save,
1082 xmm_dst,
1083 xmm_src0,
1084 xmm_src1,
1085 pow4f );
1086 }
1087
1088 static void
1089 emit_rcp (
1090 struct x86_function *func,
1091 unsigned xmm_dst,
1092 unsigned xmm_src )
1093 {
1094 /* On Intel CPUs at least, this is only accurate to 12 bits -- not
1095 * good enough. Need to either emit a proper divide or use the
1096 * iterative technique described below in emit_rsqrt().
1097 */
1098 sse2_rcpps(
1099 func,
1100 make_xmm( xmm_dst ),
1101 make_xmm( xmm_src ) );
1102 }
1103
1104 static void PIPE_CDECL
1105 rnd4f(
1106 float *store )
1107 {
1108 store[0] = floorf( store[0] + 0.5f );
1109 store[1] = floorf( store[1] + 0.5f );
1110 store[2] = floorf( store[2] + 0.5f );
1111 store[3] = floorf( store[3] + 0.5f );
1112 }
1113
1114 static void
1115 emit_rnd(
1116 struct x86_function *func,
1117 unsigned xmm_save,
1118 unsigned xmm_dst )
1119 {
1120 emit_func_call_dst_src1(
1121 func,
1122 xmm_save,
1123 xmm_dst,
1124 xmm_dst,
1125 rnd4f );
1126 }
1127
1128 static void
1129 emit_rsqrt(
1130 struct x86_function *func,
1131 unsigned xmm_dst,
1132 unsigned xmm_src )
1133 {
1134 #if HIGH_PRECISION
1135 /* Although rsqrtps() and rcpps() are low precision on some/all SSE
1136 * implementations, it is possible to improve its precision at
1137 * fairly low cost, using a newton/raphson step, as below:
1138 *
1139 * x1 = 2 * rcpps(a) - a * rcpps(a) * rcpps(a)
1140 * x1 = 0.5 * rsqrtps(a) * [3.0 - (a * rsqrtps(a))* rsqrtps(a)]
1141 *
1142 * See: http://softwarecommunity.intel.com/articles/eng/1818.htm
1143 */
1144 {
1145 struct x86_reg dst = make_xmm( xmm_dst );
1146 struct x86_reg src = make_xmm( xmm_src );
1147 struct x86_reg tmp0 = make_xmm( 2 );
1148 struct x86_reg tmp1 = make_xmm( 3 );
1149
1150 assert( xmm_dst != xmm_src );
1151 assert( xmm_dst != 2 && xmm_dst != 3 );
1152 assert( xmm_src != 2 && xmm_src != 3 );
1153
1154 sse_movaps( func, dst, get_temp( TGSI_EXEC_TEMP_HALF_I, TGSI_EXEC_TEMP_HALF_C ) );
1155 sse_movaps( func, tmp0, get_temp( TGSI_EXEC_TEMP_THREE_I, TGSI_EXEC_TEMP_THREE_C ) );
1156 sse_rsqrtps( func, tmp1, src );
1157 sse_mulps( func, src, tmp1 );
1158 sse_mulps( func, dst, tmp1 );
1159 sse_mulps( func, src, tmp1 );
1160 sse_subps( func, tmp0, src );
1161 sse_mulps( func, dst, tmp0 );
1162 }
1163 #else
1164 /* On Intel CPUs at least, this is only accurate to 12 bits -- not
1165 * good enough.
1166 */
1167 sse_rsqrtps(
1168 func,
1169 make_xmm( xmm_dst ),
1170 make_xmm( xmm_src ) );
1171 #endif
1172 }
1173
1174 static void
1175 emit_setsign(
1176 struct x86_function *func,
1177 unsigned xmm )
1178 {
1179 sse_orps(
1180 func,
1181 make_xmm( xmm ),
1182 get_temp(
1183 TGSI_EXEC_TEMP_80000000_I,
1184 TGSI_EXEC_TEMP_80000000_C ) );
1185 }
1186
1187 static void PIPE_CDECL
1188 sgn4f(
1189 float *store )
1190 {
1191 store[0] = store[0] < 0.0f ? -1.0f : store[0] > 0.0f ? 1.0f : 0.0f;
1192 store[1] = store[1] < 0.0f ? -1.0f : store[1] > 0.0f ? 1.0f : 0.0f;
1193 store[2] = store[2] < 0.0f ? -1.0f : store[2] > 0.0f ? 1.0f : 0.0f;
1194 store[3] = store[3] < 0.0f ? -1.0f : store[3] > 0.0f ? 1.0f : 0.0f;
1195 }
1196
1197 static void
1198 emit_sgn(
1199 struct x86_function *func,
1200 unsigned xmm_save,
1201 unsigned xmm_dst )
1202 {
1203 emit_func_call_dst_src1(
1204 func,
1205 xmm_save,
1206 xmm_dst,
1207 xmm_dst,
1208 sgn4f );
1209 }
1210
1211 static void PIPE_CDECL
1212 sin4f(
1213 float *store )
1214 {
1215 store[0] = sinf( store[0] );
1216 store[1] = sinf( store[1] );
1217 store[2] = sinf( store[2] );
1218 store[3] = sinf( store[3] );
1219 }
1220
1221 static void
1222 emit_sin (struct x86_function *func,
1223 unsigned xmm_save,
1224 unsigned xmm_dst)
1225 {
1226 emit_func_call_dst_src1(
1227 func,
1228 xmm_save,
1229 xmm_dst,
1230 xmm_dst,
1231 sin4f );
1232 }
1233
1234 static void
1235 emit_sub(
1236 struct x86_function *func,
1237 unsigned xmm_dst,
1238 unsigned xmm_src )
1239 {
1240 sse_subps(
1241 func,
1242 make_xmm( xmm_dst ),
1243 make_xmm( xmm_src ) );
1244 }
1245
1246
1247
1248
1249
1250
1251
1252 /**
1253 * Register fetch.
1254 */
1255
1256 static void
1257 emit_fetch(
1258 struct x86_function *func,
1259 unsigned xmm,
1260 const struct tgsi_full_src_register *reg,
1261 const unsigned chan_index )
1262 {
1263 unsigned swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
1264
1265 switch (swizzle) {
1266 case TGSI_SWIZZLE_X:
1267 case TGSI_SWIZZLE_Y:
1268 case TGSI_SWIZZLE_Z:
1269 case TGSI_SWIZZLE_W:
1270 switch (reg->Register.File) {
1271 case TGSI_FILE_CONSTANT:
1272 emit_const(
1273 func,
1274 xmm,
1275 reg->Register.Index,
1276 swizzle,
1277 reg->Register.Indirect,
1278 reg->Indirect.File,
1279 reg->Indirect.Index );
1280 break;
1281
1282 case TGSI_FILE_IMMEDIATE:
1283 emit_immediate(
1284 func,
1285 xmm,
1286 reg->Register.Index,
1287 swizzle );
1288 break;
1289
1290 case TGSI_FILE_INPUT:
1291 case TGSI_FILE_SYSTEM_VALUE:
1292 emit_inputf(
1293 func,
1294 xmm,
1295 reg->Register.Index,
1296 swizzle );
1297 break;
1298
1299 case TGSI_FILE_TEMPORARY:
1300 emit_tempf(
1301 func,
1302 xmm,
1303 reg->Register.Index,
1304 swizzle );
1305 break;
1306
1307 default:
1308 assert( 0 );
1309 }
1310 break;
1311
1312 default:
1313 assert( 0 );
1314 }
1315
1316 switch( tgsi_util_get_full_src_register_sign_mode( reg, chan_index ) ) {
1317 case TGSI_UTIL_SIGN_CLEAR:
1318 emit_abs( func, xmm );
1319 break;
1320
1321 case TGSI_UTIL_SIGN_SET:
1322 emit_setsign( func, xmm );
1323 break;
1324
1325 case TGSI_UTIL_SIGN_TOGGLE:
1326 emit_neg( func, xmm );
1327 break;
1328
1329 case TGSI_UTIL_SIGN_KEEP:
1330 break;
1331 }
1332 }
1333
1334 #define FETCH( FUNC, INST, XMM, INDEX, CHAN )\
1335 emit_fetch( FUNC, XMM, &(INST).Src[INDEX], CHAN )
1336
1337 /**
1338 * Register store.
1339 */
1340
1341 static void
1342 emit_store(
1343 struct x86_function *func,
1344 unsigned xmm,
1345 const struct tgsi_full_dst_register *reg,
1346 const struct tgsi_full_instruction *inst,
1347 unsigned chan_index )
1348 {
1349 switch( inst->Instruction.Saturate ) {
1350 case TGSI_SAT_NONE:
1351 break;
1352
1353 case TGSI_SAT_ZERO_ONE:
1354 sse_maxps(
1355 func,
1356 make_xmm( xmm ),
1357 get_temp(
1358 TGSI_EXEC_TEMP_00000000_I,
1359 TGSI_EXEC_TEMP_00000000_C ) );
1360
1361 sse_minps(
1362 func,
1363 make_xmm( xmm ),
1364 get_temp(
1365 TGSI_EXEC_TEMP_ONE_I,
1366 TGSI_EXEC_TEMP_ONE_C ) );
1367 break;
1368
1369 case TGSI_SAT_MINUS_PLUS_ONE:
1370 assert( 0 );
1371 break;
1372 }
1373
1374
1375 switch( reg->Register.File ) {
1376 case TGSI_FILE_OUTPUT:
1377 emit_output(
1378 func,
1379 xmm,
1380 reg->Register.Index,
1381 chan_index );
1382 break;
1383
1384 case TGSI_FILE_TEMPORARY:
1385 emit_temps(
1386 func,
1387 xmm,
1388 reg->Register.Index,
1389 chan_index );
1390 break;
1391
1392 case TGSI_FILE_ADDRESS:
1393 emit_addrs(
1394 func,
1395 xmm,
1396 reg->Register.Index,
1397 chan_index );
1398 break;
1399
1400 default:
1401 assert( 0 );
1402 }
1403 }
1404
1405 #define STORE( FUNC, INST, XMM, INDEX, CHAN )\
1406 emit_store( FUNC, XMM, &(INST).Dst[INDEX], &(INST), CHAN )
1407
1408
1409 static void PIPE_CDECL
1410 fetch_texel( struct tgsi_sampler **sampler,
1411 float *store )
1412 {
1413 #if 0
1414 uint j;
1415
1416 debug_printf("%s sampler: %p (%p) store: %p\n",
1417 __FUNCTION__,
1418 sampler, *sampler,
1419 store );
1420
1421 debug_printf("lodbias %f\n", store[12]);
1422
1423 for (j = 0; j < 4; j++)
1424 debug_printf("sample %d texcoord %f %f\n",
1425 j,
1426 store[0+j],
1427 store[4+j]);
1428 #endif
1429
1430 {
1431 float rgba[NUM_CHANNELS][QUAD_SIZE];
1432 (*sampler)->get_samples(*sampler,
1433 &store[0], /* s */
1434 &store[4], /* t */
1435 &store[8], /* r */
1436 store[12], /* lodbias */
1437 rgba); /* results */
1438
1439 memcpy( store, rgba, 16 * sizeof(float));
1440 }
1441
1442 #if 0
1443 for (j = 0; j < 4; j++)
1444 debug_printf("sample %d result %f %f %f %f\n",
1445 j,
1446 store[0+j],
1447 store[4+j],
1448 store[8+j],
1449 store[12+j]);
1450 #endif
1451 }
1452
1453 /**
1454 * High-level instruction translators.
1455 */
1456
1457 static void
1458 emit_tex( struct x86_function *func,
1459 const struct tgsi_full_instruction *inst,
1460 boolean lodbias,
1461 boolean projected)
1462 {
1463 const uint unit = inst->Src[1].Register.Index;
1464 struct x86_reg args[2];
1465 unsigned count;
1466 unsigned i;
1467
1468 assert(inst->Instruction.Texture);
1469 switch (inst->Texture.Texture) {
1470 case TGSI_TEXTURE_1D:
1471 count = 1;
1472 break;
1473 case TGSI_TEXTURE_2D:
1474 case TGSI_TEXTURE_RECT:
1475 count = 2;
1476 break;
1477 case TGSI_TEXTURE_SHADOW1D:
1478 case TGSI_TEXTURE_SHADOW2D:
1479 case TGSI_TEXTURE_SHADOWRECT:
1480 case TGSI_TEXTURE_3D:
1481 case TGSI_TEXTURE_CUBE:
1482 count = 3;
1483 break;
1484 default:
1485 assert(0);
1486 return;
1487 }
1488
1489 if (lodbias) {
1490 FETCH( func, *inst, 3, 0, 3 );
1491 }
1492 else {
1493 emit_tempf(
1494 func,
1495 3,
1496 TGSI_EXEC_TEMP_00000000_I,
1497 TGSI_EXEC_TEMP_00000000_C );
1498
1499 }
1500
1501 /* store lodbias whether enabled or not -- fetch_texel currently
1502 * respects it always.
1503 */
1504 sse_movaps( func,
1505 get_temp( TEMP_R0, 3 ),
1506 make_xmm( 3 ) );
1507
1508
1509 if (projected) {
1510 FETCH( func, *inst, 3, 0, 3 );
1511
1512 emit_rcp( func, 3, 3 );
1513 }
1514
1515 for (i = 0; i < count; i++) {
1516 FETCH( func, *inst, i, 0, i );
1517
1518 if (projected) {
1519 sse_mulps(
1520 func,
1521 make_xmm( i ),
1522 make_xmm( 3 ) );
1523 }
1524
1525 /* Store in the argument buffer:
1526 */
1527 sse_movaps(
1528 func,
1529 get_temp( TEMP_R0, i ),
1530 make_xmm( i ) );
1531 }
1532
1533 args[0] = get_temp( TEMP_R0, 0 );
1534 args[1] = get_sampler_ptr( unit );
1535
1536
1537 emit_func_call( func,
1538 0,
1539 args,
1540 Elements(args),
1541 fetch_texel );
1542
1543 /* If all four channels are enabled, could use a pointer to
1544 * dst[0].x instead of TEMP_R0 for store?
1545 */
1546 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, i ) {
1547
1548 sse_movaps(
1549 func,
1550 make_xmm( 0 ),
1551 get_temp( TEMP_R0, i ) );
1552
1553 STORE( func, *inst, 0, 0, i );
1554 }
1555 }
1556
1557
1558 static void
1559 emit_kil(
1560 struct x86_function *func,
1561 const struct tgsi_full_src_register *reg )
1562 {
1563 unsigned uniquemask;
1564 unsigned unique_count = 0;
1565 unsigned chan_index;
1566 unsigned i;
1567
1568 /* This mask stores component bits that were already tested. Note that
1569 * we test if the value is less than zero, so 1.0 and 0.0 need not to be
1570 * tested. */
1571 uniquemask = 0;
1572
1573 FOR_EACH_CHANNEL( chan_index ) {
1574 unsigned swizzle;
1575
1576 /* unswizzle channel */
1577 swizzle = tgsi_util_get_full_src_register_swizzle(
1578 reg,
1579 chan_index );
1580
1581 /* check if the component has not been already tested */
1582 if( !(uniquemask & (1 << swizzle)) ) {
1583 uniquemask |= 1 << swizzle;
1584
1585 /* allocate register */
1586 emit_fetch(
1587 func,
1588 unique_count++,
1589 reg,
1590 chan_index );
1591 }
1592 }
1593
1594 x86_push(
1595 func,
1596 x86_make_reg( file_REG32, reg_AX ) );
1597 x86_push(
1598 func,
1599 x86_make_reg( file_REG32, reg_DX ) );
1600
1601 for (i = 0 ; i < unique_count; i++ ) {
1602 struct x86_reg dataXMM = make_xmm(i);
1603
1604 sse_cmpps(
1605 func,
1606 dataXMM,
1607 get_temp(
1608 TGSI_EXEC_TEMP_00000000_I,
1609 TGSI_EXEC_TEMP_00000000_C ),
1610 cc_LessThan );
1611
1612 if( i == 0 ) {
1613 sse_movmskps(
1614 func,
1615 x86_make_reg( file_REG32, reg_AX ),
1616 dataXMM );
1617 }
1618 else {
1619 sse_movmskps(
1620 func,
1621 x86_make_reg( file_REG32, reg_DX ),
1622 dataXMM );
1623 x86_or(
1624 func,
1625 x86_make_reg( file_REG32, reg_AX ),
1626 x86_make_reg( file_REG32, reg_DX ) );
1627 }
1628 }
1629
1630 x86_or(
1631 func,
1632 get_temp(
1633 TGSI_EXEC_TEMP_KILMASK_I,
1634 TGSI_EXEC_TEMP_KILMASK_C ),
1635 x86_make_reg( file_REG32, reg_AX ) );
1636
1637 x86_pop(
1638 func,
1639 x86_make_reg( file_REG32, reg_DX ) );
1640 x86_pop(
1641 func,
1642 x86_make_reg( file_REG32, reg_AX ) );
1643 }
1644
1645
1646 static void
1647 emit_kilp(
1648 struct x86_function *func )
1649 {
1650 /* XXX todo / fix me */
1651 }
1652
1653
1654 static void
1655 emit_setcc(
1656 struct x86_function *func,
1657 struct tgsi_full_instruction *inst,
1658 enum sse_cc cc )
1659 {
1660 unsigned chan_index;
1661
1662 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1663 FETCH( func, *inst, 0, 0, chan_index );
1664 FETCH( func, *inst, 1, 1, chan_index );
1665 sse_cmpps(
1666 func,
1667 make_xmm( 0 ),
1668 make_xmm( 1 ),
1669 cc );
1670 sse_andps(
1671 func,
1672 make_xmm( 0 ),
1673 get_temp(
1674 TEMP_ONE_I,
1675 TEMP_ONE_C ) );
1676 STORE( func, *inst, 0, 0, chan_index );
1677 }
1678 }
1679
1680 static void
1681 emit_cmp(
1682 struct x86_function *func,
1683 struct tgsi_full_instruction *inst )
1684 {
1685 unsigned chan_index;
1686
1687 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1688 FETCH( func, *inst, 0, 0, chan_index );
1689 FETCH( func, *inst, 1, 1, chan_index );
1690 FETCH( func, *inst, 2, 2, chan_index );
1691 sse_cmpps(
1692 func,
1693 make_xmm( 0 ),
1694 get_temp(
1695 TGSI_EXEC_TEMP_00000000_I,
1696 TGSI_EXEC_TEMP_00000000_C ),
1697 cc_LessThan );
1698 sse_andps(
1699 func,
1700 make_xmm( 1 ),
1701 make_xmm( 0 ) );
1702 sse_andnps(
1703 func,
1704 make_xmm( 0 ),
1705 make_xmm( 2 ) );
1706 sse_orps(
1707 func,
1708 make_xmm( 0 ),
1709 make_xmm( 1 ) );
1710 STORE( func, *inst, 0, 0, chan_index );
1711 }
1712 }
1713
1714
1715 /**
1716 * Check if inst src/dest regs use indirect addressing into temporary
1717 * register file.
1718 */
1719 static boolean
1720 indirect_temp_reference(const struct tgsi_full_instruction *inst)
1721 {
1722 uint i;
1723 for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
1724 const struct tgsi_full_src_register *reg = &inst->Src[i];
1725 if (reg->Register.File == TGSI_FILE_TEMPORARY &&
1726 reg->Register.Indirect)
1727 return TRUE;
1728 }
1729 for (i = 0; i < inst->Instruction.NumDstRegs; i++) {
1730 const struct tgsi_full_dst_register *reg = &inst->Dst[i];
1731 if (reg->Register.File == TGSI_FILE_TEMPORARY &&
1732 reg->Register.Indirect)
1733 return TRUE;
1734 }
1735 return FALSE;
1736 }
1737
1738
1739 static int
1740 emit_instruction(
1741 struct x86_function *func,
1742 struct tgsi_full_instruction *inst )
1743 {
1744 unsigned chan_index;
1745
1746 /* we can't handle indirect addressing into temp register file yet */
1747 if (indirect_temp_reference(inst))
1748 return FALSE;
1749
1750 switch (inst->Instruction.Opcode) {
1751 case TGSI_OPCODE_ARL:
1752 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1753 FETCH( func, *inst, 0, 0, chan_index );
1754 emit_flr(func, 0, 0);
1755 emit_f2it( func, 0 );
1756 STORE( func, *inst, 0, 0, chan_index );
1757 }
1758 break;
1759
1760 case TGSI_OPCODE_MOV:
1761 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1762 FETCH( func, *inst, 4 + chan_index, 0, chan_index );
1763 }
1764 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1765 STORE( func, *inst, 4 + chan_index, 0, chan_index );
1766 }
1767 break;
1768
1769 case TGSI_OPCODE_LIT:
1770 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1771 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) ) {
1772 emit_tempf(
1773 func,
1774 0,
1775 TEMP_ONE_I,
1776 TEMP_ONE_C);
1777 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ) {
1778 STORE( func, *inst, 0, 0, CHAN_X );
1779 }
1780 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) ) {
1781 STORE( func, *inst, 0, 0, CHAN_W );
1782 }
1783 }
1784 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
1785 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
1786 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
1787 FETCH( func, *inst, 0, 0, CHAN_X );
1788 sse_maxps(
1789 func,
1790 make_xmm( 0 ),
1791 get_temp(
1792 TGSI_EXEC_TEMP_00000000_I,
1793 TGSI_EXEC_TEMP_00000000_C ) );
1794 STORE( func, *inst, 0, 0, CHAN_Y );
1795 }
1796 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
1797 /* XMM[1] = SrcReg[0].yyyy */
1798 FETCH( func, *inst, 1, 0, CHAN_Y );
1799 /* XMM[1] = max(XMM[1], 0) */
1800 sse_maxps(
1801 func,
1802 make_xmm( 1 ),
1803 get_temp(
1804 TGSI_EXEC_TEMP_00000000_I,
1805 TGSI_EXEC_TEMP_00000000_C ) );
1806 /* XMM[2] = SrcReg[0].wwww */
1807 FETCH( func, *inst, 2, 0, CHAN_W );
1808 /* XMM[2] = min(XMM[2], 128.0) */
1809 sse_minps(
1810 func,
1811 make_xmm( 2 ),
1812 get_temp(
1813 TGSI_EXEC_TEMP_128_I,
1814 TGSI_EXEC_TEMP_128_C ) );
1815 /* XMM[2] = max(XMM[2], -128.0) */
1816 sse_maxps(
1817 func,
1818 make_xmm( 2 ),
1819 get_temp(
1820 TGSI_EXEC_TEMP_MINUS_128_I,
1821 TGSI_EXEC_TEMP_MINUS_128_C ) );
1822 emit_pow( func, 3, 1, 1, 2 );
1823 FETCH( func, *inst, 0, 0, CHAN_X );
1824 sse_xorps(
1825 func,
1826 make_xmm( 2 ),
1827 make_xmm( 2 ) );
1828 sse_cmpps(
1829 func,
1830 make_xmm( 2 ),
1831 make_xmm( 0 ),
1832 cc_LessThan );
1833 sse_andps(
1834 func,
1835 make_xmm( 2 ),
1836 make_xmm( 1 ) );
1837 STORE( func, *inst, 2, 0, CHAN_Z );
1838 }
1839 }
1840 break;
1841
1842 case TGSI_OPCODE_RCP:
1843 FETCH( func, *inst, 0, 0, CHAN_X );
1844 emit_rcp( func, 0, 0 );
1845 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1846 STORE( func, *inst, 0, 0, chan_index );
1847 }
1848 break;
1849
1850 case TGSI_OPCODE_RSQ:
1851 FETCH( func, *inst, 0, 0, CHAN_X );
1852 emit_abs( func, 0 );
1853 emit_rsqrt( func, 1, 0 );
1854 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1855 STORE( func, *inst, 1, 0, chan_index );
1856 }
1857 break;
1858
1859 case TGSI_OPCODE_EXP:
1860 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1861 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
1862 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1863 FETCH( func, *inst, 0, 0, CHAN_X );
1864 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1865 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1866 emit_MOV( func, 1, 0 );
1867 emit_flr( func, 2, 1 );
1868 /* dst.x = ex2(floor(src.x)) */
1869 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X )) {
1870 emit_MOV( func, 2, 1 );
1871 emit_ex2( func, 3, 2 );
1872 STORE( func, *inst, 2, 0, CHAN_X );
1873 }
1874 /* dst.y = src.x - floor(src.x) */
1875 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1876 emit_MOV( func, 2, 0 );
1877 emit_sub( func, 2, 1 );
1878 STORE( func, *inst, 2, 0, CHAN_Y );
1879 }
1880 }
1881 /* dst.z = ex2(src.x) */
1882 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1883 emit_ex2( func, 3, 0 );
1884 STORE( func, *inst, 0, 0, CHAN_Z );
1885 }
1886 }
1887 /* dst.w = 1.0 */
1888 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W )) {
1889 emit_tempf( func, 0, TEMP_ONE_I, TEMP_ONE_C );
1890 STORE( func, *inst, 0, 0, CHAN_W );
1891 }
1892 break;
1893
1894 case TGSI_OPCODE_LOG:
1895 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1896 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
1897 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1898 FETCH( func, *inst, 0, 0, CHAN_X );
1899 emit_abs( func, 0 );
1900 emit_MOV( func, 1, 0 );
1901 emit_lg2( func, 2, 1 );
1902 /* dst.z = lg2(abs(src.x)) */
1903 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1904 STORE( func, *inst, 1, 0, CHAN_Z );
1905 }
1906 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1907 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1908 emit_flr( func, 2, 1 );
1909 /* dst.x = floor(lg2(abs(src.x))) */
1910 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X )) {
1911 STORE( func, *inst, 1, 0, CHAN_X );
1912 }
1913 /* dst.x = abs(src)/ex2(floor(lg2(abs(src.x)))) */
1914 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1915 emit_ex2( func, 2, 1 );
1916 emit_rcp( func, 1, 1 );
1917 emit_mul( func, 0, 1 );
1918 STORE( func, *inst, 0, 0, CHAN_Y );
1919 }
1920 }
1921 }
1922 /* dst.w = 1.0 */
1923 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W )) {
1924 emit_tempf( func, 0, TEMP_ONE_I, TEMP_ONE_C );
1925 STORE( func, *inst, 0, 0, CHAN_W );
1926 }
1927 break;
1928
1929 case TGSI_OPCODE_MUL:
1930 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1931 FETCH( func, *inst, 0, 0, chan_index );
1932 FETCH( func, *inst, 1, 1, chan_index );
1933 emit_mul( func, 0, 1 );
1934 STORE( func, *inst, 0, 0, chan_index );
1935 }
1936 break;
1937
1938 case TGSI_OPCODE_ADD:
1939 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1940 FETCH( func, *inst, 0, 0, chan_index );
1941 FETCH( func, *inst, 1, 1, chan_index );
1942 emit_add( func, 0, 1 );
1943 STORE( func, *inst, 0, 0, chan_index );
1944 }
1945 break;
1946
1947 case TGSI_OPCODE_DP3:
1948 FETCH( func, *inst, 0, 0, CHAN_X );
1949 FETCH( func, *inst, 1, 1, CHAN_X );
1950 emit_mul( func, 0, 1 );
1951 FETCH( func, *inst, 1, 0, CHAN_Y );
1952 FETCH( func, *inst, 2, 1, CHAN_Y );
1953 emit_mul( func, 1, 2 );
1954 emit_add( func, 0, 1 );
1955 FETCH( func, *inst, 1, 0, CHAN_Z );
1956 FETCH( func, *inst, 2, 1, CHAN_Z );
1957 emit_mul( func, 1, 2 );
1958 emit_add( func, 0, 1 );
1959 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1960 STORE( func, *inst, 0, 0, chan_index );
1961 }
1962 break;
1963
1964 case TGSI_OPCODE_DP4:
1965 FETCH( func, *inst, 0, 0, CHAN_X );
1966 FETCH( func, *inst, 1, 1, CHAN_X );
1967 emit_mul( func, 0, 1 );
1968 FETCH( func, *inst, 1, 0, CHAN_Y );
1969 FETCH( func, *inst, 2, 1, CHAN_Y );
1970 emit_mul( func, 1, 2 );
1971 emit_add( func, 0, 1 );
1972 FETCH( func, *inst, 1, 0, CHAN_Z );
1973 FETCH( func, *inst, 2, 1, CHAN_Z );
1974 emit_mul(func, 1, 2 );
1975 emit_add(func, 0, 1 );
1976 FETCH( func, *inst, 1, 0, CHAN_W );
1977 FETCH( func, *inst, 2, 1, CHAN_W );
1978 emit_mul( func, 1, 2 );
1979 emit_add( func, 0, 1 );
1980 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1981 STORE( func, *inst, 0, 0, chan_index );
1982 }
1983 break;
1984
1985 case TGSI_OPCODE_DST:
1986 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
1987 emit_tempf(
1988 func,
1989 0,
1990 TEMP_ONE_I,
1991 TEMP_ONE_C );
1992 STORE( func, *inst, 0, 0, CHAN_X );
1993 }
1994 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
1995 FETCH( func, *inst, 0, 0, CHAN_Y );
1996 FETCH( func, *inst, 1, 1, CHAN_Y );
1997 emit_mul( func, 0, 1 );
1998 STORE( func, *inst, 0, 0, CHAN_Y );
1999 }
2000 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
2001 FETCH( func, *inst, 0, 0, CHAN_Z );
2002 STORE( func, *inst, 0, 0, CHAN_Z );
2003 }
2004 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
2005 FETCH( func, *inst, 0, 1, CHAN_W );
2006 STORE( func, *inst, 0, 0, CHAN_W );
2007 }
2008 break;
2009
2010 case TGSI_OPCODE_MIN:
2011 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2012 FETCH( func, *inst, 0, 0, chan_index );
2013 FETCH( func, *inst, 1, 1, chan_index );
2014 sse_minps(
2015 func,
2016 make_xmm( 0 ),
2017 make_xmm( 1 ) );
2018 STORE( func, *inst, 0, 0, chan_index );
2019 }
2020 break;
2021
2022 case TGSI_OPCODE_MAX:
2023 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2024 FETCH( func, *inst, 0, 0, chan_index );
2025 FETCH( func, *inst, 1, 1, chan_index );
2026 sse_maxps(
2027 func,
2028 make_xmm( 0 ),
2029 make_xmm( 1 ) );
2030 STORE( func, *inst, 0, 0, chan_index );
2031 }
2032 break;
2033
2034 case TGSI_OPCODE_SLT:
2035 emit_setcc( func, inst, cc_LessThan );
2036 break;
2037
2038 case TGSI_OPCODE_SGE:
2039 emit_setcc( func, inst, cc_NotLessThan );
2040 break;
2041
2042 case TGSI_OPCODE_MAD:
2043 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2044 FETCH( func, *inst, 0, 0, chan_index );
2045 FETCH( func, *inst, 1, 1, chan_index );
2046 FETCH( func, *inst, 2, 2, chan_index );
2047 emit_mul( func, 0, 1 );
2048 emit_add( func, 0, 2 );
2049 STORE( func, *inst, 0, 0, chan_index );
2050 }
2051 break;
2052
2053 case TGSI_OPCODE_SUB:
2054 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2055 FETCH( func, *inst, 0, 0, chan_index );
2056 FETCH( func, *inst, 1, 1, chan_index );
2057 emit_sub( func, 0, 1 );
2058 STORE( func, *inst, 0, 0, chan_index );
2059 }
2060 break;
2061
2062 case TGSI_OPCODE_LRP:
2063 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2064 FETCH( func, *inst, 0, 0, chan_index );
2065 FETCH( func, *inst, 1, 1, chan_index );
2066 FETCH( func, *inst, 2, 2, chan_index );
2067 emit_sub( func, 1, 2 );
2068 emit_mul( func, 0, 1 );
2069 emit_add( func, 0, 2 );
2070 STORE( func, *inst, 0, 0, chan_index );
2071 }
2072 break;
2073
2074 case TGSI_OPCODE_CND:
2075 return 0;
2076 break;
2077
2078 case TGSI_OPCODE_DP2A:
2079 FETCH( func, *inst, 0, 0, CHAN_X ); /* xmm0 = src[0].x */
2080 FETCH( func, *inst, 1, 1, CHAN_X ); /* xmm1 = src[1].x */
2081 emit_mul( func, 0, 1 ); /* xmm0 = xmm0 * xmm1 */
2082 FETCH( func, *inst, 1, 0, CHAN_Y ); /* xmm1 = src[0].y */
2083 FETCH( func, *inst, 2, 1, CHAN_Y ); /* xmm2 = src[1].y */
2084 emit_mul( func, 1, 2 ); /* xmm1 = xmm1 * xmm2 */
2085 emit_add( func, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2086 FETCH( func, *inst, 1, 2, CHAN_X ); /* xmm1 = src[2].x */
2087 emit_add( func, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2088 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2089 STORE( func, *inst, 0, 0, chan_index ); /* dest[ch] = xmm0 */
2090 }
2091 break;
2092
2093 case TGSI_OPCODE_FRC:
2094 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2095 FETCH( func, *inst, 0, 0, chan_index );
2096 emit_frc( func, 0, 0 );
2097 STORE( func, *inst, 0, 0, chan_index );
2098 }
2099 break;
2100
2101 case TGSI_OPCODE_CLAMP:
2102 return 0;
2103 break;
2104
2105 case TGSI_OPCODE_FLR:
2106 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2107 FETCH( func, *inst, 0, 0, chan_index );
2108 emit_flr( func, 0, 0 );
2109 STORE( func, *inst, 0, 0, chan_index );
2110 }
2111 break;
2112
2113 case TGSI_OPCODE_ROUND:
2114 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2115 FETCH( func, *inst, 0, 0, chan_index );
2116 emit_rnd( func, 0, 0 );
2117 STORE( func, *inst, 0, 0, chan_index );
2118 }
2119 break;
2120
2121 case TGSI_OPCODE_EX2:
2122 FETCH( func, *inst, 0, 0, CHAN_X );
2123 emit_ex2( func, 0, 0 );
2124 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2125 STORE( func, *inst, 0, 0, chan_index );
2126 }
2127 break;
2128
2129 case TGSI_OPCODE_LG2:
2130 FETCH( func, *inst, 0, 0, CHAN_X );
2131 emit_lg2( func, 0, 0 );
2132 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2133 STORE( func, *inst, 0, 0, chan_index );
2134 }
2135 break;
2136
2137 case TGSI_OPCODE_POW:
2138 FETCH( func, *inst, 0, 0, CHAN_X );
2139 FETCH( func, *inst, 1, 1, CHAN_X );
2140 emit_pow( func, 0, 0, 0, 1 );
2141 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2142 STORE( func, *inst, 0, 0, chan_index );
2143 }
2144 break;
2145
2146 case TGSI_OPCODE_XPD:
2147 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
2148 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
2149 FETCH( func, *inst, 1, 1, CHAN_Z );
2150 FETCH( func, *inst, 3, 0, CHAN_Z );
2151 }
2152 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
2153 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
2154 FETCH( func, *inst, 0, 0, CHAN_Y );
2155 FETCH( func, *inst, 4, 1, CHAN_Y );
2156 }
2157 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
2158 emit_MOV( func, 2, 0 );
2159 emit_mul( func, 2, 1 );
2160 emit_MOV( func, 5, 3 );
2161 emit_mul( func, 5, 4 );
2162 emit_sub( func, 2, 5 );
2163 STORE( func, *inst, 2, 0, CHAN_X );
2164 }
2165 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
2166 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
2167 FETCH( func, *inst, 2, 1, CHAN_X );
2168 FETCH( func, *inst, 5, 0, CHAN_X );
2169 }
2170 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
2171 emit_mul( func, 3, 2 );
2172 emit_mul( func, 1, 5 );
2173 emit_sub( func, 3, 1 );
2174 STORE( func, *inst, 3, 0, CHAN_Y );
2175 }
2176 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
2177 emit_mul( func, 5, 4 );
2178 emit_mul( func, 0, 2 );
2179 emit_sub( func, 5, 0 );
2180 STORE( func, *inst, 5, 0, CHAN_Z );
2181 }
2182 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
2183 emit_tempf(
2184 func,
2185 0,
2186 TEMP_ONE_I,
2187 TEMP_ONE_C );
2188 STORE( func, *inst, 0, 0, CHAN_W );
2189 }
2190 break;
2191
2192 case TGSI_OPCODE_ABS:
2193 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2194 FETCH( func, *inst, 0, 0, chan_index );
2195 emit_abs( func, 0) ;
2196
2197 STORE( func, *inst, 0, 0, chan_index );
2198 }
2199 break;
2200
2201 case TGSI_OPCODE_RCC:
2202 return 0;
2203 break;
2204
2205 case TGSI_OPCODE_DPH:
2206 FETCH( func, *inst, 0, 0, CHAN_X );
2207 FETCH( func, *inst, 1, 1, CHAN_X );
2208 emit_mul( func, 0, 1 );
2209 FETCH( func, *inst, 1, 0, CHAN_Y );
2210 FETCH( func, *inst, 2, 1, CHAN_Y );
2211 emit_mul( func, 1, 2 );
2212 emit_add( func, 0, 1 );
2213 FETCH( func, *inst, 1, 0, CHAN_Z );
2214 FETCH( func, *inst, 2, 1, CHAN_Z );
2215 emit_mul( func, 1, 2 );
2216 emit_add( func, 0, 1 );
2217 FETCH( func, *inst, 1, 1, CHAN_W );
2218 emit_add( func, 0, 1 );
2219 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2220 STORE( func, *inst, 0, 0, chan_index );
2221 }
2222 break;
2223
2224 case TGSI_OPCODE_COS:
2225 FETCH( func, *inst, 0, 0, CHAN_X );
2226 emit_cos( func, 0, 0 );
2227 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2228 STORE( func, *inst, 0, 0, chan_index );
2229 }
2230 break;
2231
2232 case TGSI_OPCODE_DDX:
2233 return 0;
2234 break;
2235
2236 case TGSI_OPCODE_DDY:
2237 return 0;
2238 break;
2239
2240 case TGSI_OPCODE_KILP:
2241 /* predicated kill */
2242 emit_kilp( func );
2243 return 0; /* XXX fix me */
2244 break;
2245
2246 case TGSI_OPCODE_KIL:
2247 /* conditional kill */
2248 emit_kil( func, &inst->Src[0] );
2249 break;
2250
2251 case TGSI_OPCODE_PK2H:
2252 return 0;
2253 break;
2254
2255 case TGSI_OPCODE_PK2US:
2256 return 0;
2257 break;
2258
2259 case TGSI_OPCODE_PK4B:
2260 return 0;
2261 break;
2262
2263 case TGSI_OPCODE_PK4UB:
2264 return 0;
2265 break;
2266
2267 case TGSI_OPCODE_RFL:
2268 return 0;
2269 break;
2270
2271 case TGSI_OPCODE_SEQ:
2272 emit_setcc( func, inst, cc_Equal );
2273 break;
2274
2275 case TGSI_OPCODE_SFL:
2276 return 0;
2277 break;
2278
2279 case TGSI_OPCODE_SGT:
2280 emit_setcc( func, inst, cc_NotLessThanEqual );
2281 break;
2282
2283 case TGSI_OPCODE_SIN:
2284 FETCH( func, *inst, 0, 0, CHAN_X );
2285 emit_sin( func, 0, 0 );
2286 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2287 STORE( func, *inst, 0, 0, chan_index );
2288 }
2289 break;
2290
2291 case TGSI_OPCODE_SLE:
2292 emit_setcc( func, inst, cc_LessThanEqual );
2293 break;
2294
2295 case TGSI_OPCODE_SNE:
2296 emit_setcc( func, inst, cc_NotEqual );
2297 break;
2298
2299 case TGSI_OPCODE_STR:
2300 return 0;
2301 break;
2302
2303 case TGSI_OPCODE_TEX:
2304 emit_tex( func, inst, FALSE, FALSE );
2305 break;
2306
2307 case TGSI_OPCODE_TXD:
2308 return 0;
2309 break;
2310
2311 case TGSI_OPCODE_UP2H:
2312 return 0;
2313 break;
2314
2315 case TGSI_OPCODE_UP2US:
2316 return 0;
2317 break;
2318
2319 case TGSI_OPCODE_UP4B:
2320 return 0;
2321 break;
2322
2323 case TGSI_OPCODE_UP4UB:
2324 return 0;
2325 break;
2326
2327 case TGSI_OPCODE_X2D:
2328 return 0;
2329 break;
2330
2331 case TGSI_OPCODE_ARA:
2332 return 0;
2333 break;
2334
2335 case TGSI_OPCODE_ARR:
2336 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2337 FETCH( func, *inst, 0, 0, chan_index );
2338 emit_rnd( func, 0, 0 );
2339 emit_f2it( func, 0 );
2340 STORE( func, *inst, 0, 0, chan_index );
2341 }
2342 break;
2343
2344 case TGSI_OPCODE_BRA:
2345 return 0;
2346 break;
2347
2348 case TGSI_OPCODE_CAL:
2349 return 0;
2350 break;
2351
2352 case TGSI_OPCODE_RET:
2353 emit_ret( func );
2354 break;
2355
2356 case TGSI_OPCODE_END:
2357 break;
2358
2359 case TGSI_OPCODE_SSG:
2360 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2361 FETCH( func, *inst, 0, 0, chan_index );
2362 emit_sgn( func, 0, 0 );
2363 STORE( func, *inst, 0, 0, chan_index );
2364 }
2365 break;
2366
2367 case TGSI_OPCODE_CMP:
2368 emit_cmp (func, inst);
2369 break;
2370
2371 case TGSI_OPCODE_SCS:
2372 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
2373 FETCH( func, *inst, 0, 0, CHAN_X );
2374 emit_cos( func, 0, 0 );
2375 STORE( func, *inst, 0, 0, CHAN_X );
2376 }
2377 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
2378 FETCH( func, *inst, 0, 0, CHAN_X );
2379 emit_sin( func, 0, 0 );
2380 STORE( func, *inst, 0, 0, CHAN_Y );
2381 }
2382 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
2383 emit_tempf(
2384 func,
2385 0,
2386 TGSI_EXEC_TEMP_00000000_I,
2387 TGSI_EXEC_TEMP_00000000_C );
2388 STORE( func, *inst, 0, 0, CHAN_Z );
2389 }
2390 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
2391 emit_tempf(
2392 func,
2393 0,
2394 TEMP_ONE_I,
2395 TEMP_ONE_C );
2396 STORE( func, *inst, 0, 0, CHAN_W );
2397 }
2398 break;
2399
2400 case TGSI_OPCODE_TXB:
2401 emit_tex( func, inst, TRUE, FALSE );
2402 break;
2403
2404 case TGSI_OPCODE_NRM:
2405 /* fall-through */
2406 case TGSI_OPCODE_NRM4:
2407 /* 3 or 4-component normalization */
2408 {
2409 uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
2410
2411 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) ||
2412 IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y) ||
2413 IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z) ||
2414 (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W) && dims == 4)) {
2415
2416 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
2417
2418 /* xmm4 = src.x */
2419 /* xmm0 = src.x * src.x */
2420 FETCH(func, *inst, 0, 0, CHAN_X);
2421 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X)) {
2422 emit_MOV(func, 4, 0);
2423 }
2424 emit_mul(func, 0, 0);
2425
2426 /* xmm5 = src.y */
2427 /* xmm0 = xmm0 + src.y * src.y */
2428 FETCH(func, *inst, 1, 0, CHAN_Y);
2429 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) {
2430 emit_MOV(func, 5, 1);
2431 }
2432 emit_mul(func, 1, 1);
2433 emit_add(func, 0, 1);
2434
2435 /* xmm6 = src.z */
2436 /* xmm0 = xmm0 + src.z * src.z */
2437 FETCH(func, *inst, 1, 0, CHAN_Z);
2438 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) {
2439 emit_MOV(func, 6, 1);
2440 }
2441 emit_mul(func, 1, 1);
2442 emit_add(func, 0, 1);
2443
2444 if (dims == 4) {
2445 /* xmm7 = src.w */
2446 /* xmm0 = xmm0 + src.w * src.w */
2447 FETCH(func, *inst, 1, 0, CHAN_W);
2448 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W)) {
2449 emit_MOV(func, 7, 1);
2450 }
2451 emit_mul(func, 1, 1);
2452 emit_add(func, 0, 1);
2453 }
2454
2455 /* xmm1 = 1 / sqrt(xmm0) */
2456 emit_rsqrt(func, 1, 0);
2457
2458 /* dst.x = xmm1 * src.x */
2459 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X)) {
2460 emit_mul(func, 4, 1);
2461 STORE(func, *inst, 4, 0, CHAN_X);
2462 }
2463
2464 /* dst.y = xmm1 * src.y */
2465 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) {
2466 emit_mul(func, 5, 1);
2467 STORE(func, *inst, 5, 0, CHAN_Y);
2468 }
2469
2470 /* dst.z = xmm1 * src.z */
2471 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) {
2472 emit_mul(func, 6, 1);
2473 STORE(func, *inst, 6, 0, CHAN_Z);
2474 }
2475
2476 /* dst.w = xmm1 * src.w */
2477 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) && dims == 4) {
2478 emit_mul(func, 7, 1);
2479 STORE(func, *inst, 7, 0, CHAN_W);
2480 }
2481 }
2482
2483 /* dst0.w = 1.0 */
2484 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W) && dims == 3) {
2485 emit_tempf(func, 0, TEMP_ONE_I, TEMP_ONE_C);
2486 STORE(func, *inst, 0, 0, CHAN_W);
2487 }
2488 }
2489 break;
2490
2491 case TGSI_OPCODE_DIV:
2492 return 0;
2493 break;
2494
2495 case TGSI_OPCODE_DP2:
2496 FETCH( func, *inst, 0, 0, CHAN_X ); /* xmm0 = src[0].x */
2497 FETCH( func, *inst, 1, 1, CHAN_X ); /* xmm1 = src[1].x */
2498 emit_mul( func, 0, 1 ); /* xmm0 = xmm0 * xmm1 */
2499 FETCH( func, *inst, 1, 0, CHAN_Y ); /* xmm1 = src[0].y */
2500 FETCH( func, *inst, 2, 1, CHAN_Y ); /* xmm2 = src[1].y */
2501 emit_mul( func, 1, 2 ); /* xmm1 = xmm1 * xmm2 */
2502 emit_add( func, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2503 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2504 STORE( func, *inst, 0, 0, chan_index ); /* dest[ch] = xmm0 */
2505 }
2506 break;
2507
2508 case TGSI_OPCODE_TXL:
2509 emit_tex( func, inst, TRUE, FALSE );
2510 break;
2511
2512 case TGSI_OPCODE_TXP:
2513 emit_tex( func, inst, FALSE, TRUE );
2514 break;
2515
2516 case TGSI_OPCODE_BRK:
2517 return 0;
2518 break;
2519
2520 case TGSI_OPCODE_IF:
2521 return 0;
2522 break;
2523
2524 case TGSI_OPCODE_BGNFOR:
2525 return 0;
2526 break;
2527
2528 case TGSI_OPCODE_REP:
2529 return 0;
2530 break;
2531
2532 case TGSI_OPCODE_ELSE:
2533 return 0;
2534 break;
2535
2536 case TGSI_OPCODE_ENDIF:
2537 return 0;
2538 break;
2539
2540 case TGSI_OPCODE_ENDFOR:
2541 return 0;
2542 break;
2543
2544 case TGSI_OPCODE_ENDREP:
2545 return 0;
2546 break;
2547
2548 case TGSI_OPCODE_PUSHA:
2549 return 0;
2550 break;
2551
2552 case TGSI_OPCODE_POPA:
2553 return 0;
2554 break;
2555
2556 case TGSI_OPCODE_CEIL:
2557 return 0;
2558 break;
2559
2560 case TGSI_OPCODE_I2F:
2561 return 0;
2562 break;
2563
2564 case TGSI_OPCODE_NOT:
2565 return 0;
2566 break;
2567
2568 case TGSI_OPCODE_TRUNC:
2569 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2570 FETCH( func, *inst, 0, 0, chan_index );
2571 emit_f2it( func, 0 );
2572 emit_i2f( func, 0 );
2573 STORE( func, *inst, 0, 0, chan_index );
2574 }
2575 break;
2576
2577 case TGSI_OPCODE_SHL:
2578 return 0;
2579 break;
2580
2581 case TGSI_OPCODE_SHR:
2582 return 0;
2583 break;
2584
2585 case TGSI_OPCODE_AND:
2586 return 0;
2587 break;
2588
2589 case TGSI_OPCODE_OR:
2590 return 0;
2591 break;
2592
2593 case TGSI_OPCODE_MOD:
2594 return 0;
2595 break;
2596
2597 case TGSI_OPCODE_XOR:
2598 return 0;
2599 break;
2600
2601 case TGSI_OPCODE_SAD:
2602 return 0;
2603 break;
2604
2605 case TGSI_OPCODE_TXF:
2606 return 0;
2607 break;
2608
2609 case TGSI_OPCODE_TXQ:
2610 return 0;
2611 break;
2612
2613 case TGSI_OPCODE_CONT:
2614 return 0;
2615 break;
2616
2617 case TGSI_OPCODE_EMIT:
2618 return 0;
2619 break;
2620
2621 case TGSI_OPCODE_ENDPRIM:
2622 return 0;
2623 break;
2624
2625 default:
2626 return 0;
2627 }
2628
2629 return 1;
2630 }
2631
2632 static void
2633 emit_declaration(
2634 struct x86_function *func,
2635 struct tgsi_full_declaration *decl )
2636 {
2637 if( decl->Declaration.File == TGSI_FILE_INPUT ||
2638 decl->Declaration.File == TGSI_FILE_SYSTEM_VALUE ) {
2639 unsigned first, last, mask;
2640 unsigned i, j;
2641
2642 first = decl->Range.First;
2643 last = decl->Range.Last;
2644 mask = decl->Declaration.UsageMask;
2645
2646 for( i = first; i <= last; i++ ) {
2647 for( j = 0; j < NUM_CHANNELS; j++ ) {
2648 if( mask & (1 << j) ) {
2649 switch( decl->Declaration.Interpolate ) {
2650 case TGSI_INTERPOLATE_CONSTANT:
2651 emit_coef_a0( func, 0, i, j );
2652 emit_inputs( func, 0, i, j );
2653 break;
2654
2655 case TGSI_INTERPOLATE_LINEAR:
2656 emit_tempf( func, 0, 0, TGSI_SWIZZLE_X );
2657 emit_coef_dadx( func, 1, i, j );
2658 emit_tempf( func, 2, 0, TGSI_SWIZZLE_Y );
2659 emit_coef_dady( func, 3, i, j );
2660 emit_mul( func, 0, 1 ); /* x * dadx */
2661 emit_coef_a0( func, 4, i, j );
2662 emit_mul( func, 2, 3 ); /* y * dady */
2663 emit_add( func, 0, 4 ); /* x * dadx + a0 */
2664 emit_add( func, 0, 2 ); /* x * dadx + y * dady + a0 */
2665 emit_inputs( func, 0, i, j );
2666 break;
2667
2668 case TGSI_INTERPOLATE_PERSPECTIVE:
2669 emit_tempf( func, 0, 0, TGSI_SWIZZLE_X );
2670 emit_coef_dadx( func, 1, i, j );
2671 emit_tempf( func, 2, 0, TGSI_SWIZZLE_Y );
2672 emit_coef_dady( func, 3, i, j );
2673 emit_mul( func, 0, 1 ); /* x * dadx */
2674 emit_tempf( func, 4, 0, TGSI_SWIZZLE_W );
2675 emit_coef_a0( func, 5, i, j );
2676 emit_rcp( func, 4, 4 ); /* 1.0 / w */
2677 emit_mul( func, 2, 3 ); /* y * dady */
2678 emit_add( func, 0, 5 ); /* x * dadx + a0 */
2679 emit_add( func, 0, 2 ); /* x * dadx + y * dady + a0 */
2680 emit_mul( func, 0, 4 ); /* (x * dadx + y * dady + a0) / w */
2681 emit_inputs( func, 0, i, j );
2682 break;
2683
2684 default:
2685 assert( 0 );
2686 break;
2687 }
2688 }
2689 }
2690 }
2691 }
2692 }
2693
2694 static void aos_to_soa( struct x86_function *func,
2695 uint arg_aos,
2696 uint arg_machine,
2697 uint arg_num,
2698 uint arg_stride )
2699 {
2700 struct x86_reg soa_input = x86_make_reg( file_REG32, reg_AX );
2701 struct x86_reg aos_input = x86_make_reg( file_REG32, reg_BX );
2702 struct x86_reg num_inputs = x86_make_reg( file_REG32, reg_CX );
2703 struct x86_reg stride = x86_make_reg( file_REG32, reg_DX );
2704 int inner_loop;
2705
2706
2707 /* Save EBX */
2708 x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
2709
2710 x86_mov( func, aos_input, x86_fn_arg( func, arg_aos ) );
2711 x86_mov( func, soa_input, x86_fn_arg( func, arg_machine ) );
2712 x86_lea( func, soa_input,
2713 x86_make_disp( soa_input,
2714 Offset(struct tgsi_exec_machine, Inputs) ) );
2715 x86_mov( func, num_inputs, x86_fn_arg( func, arg_num ) );
2716 x86_mov( func, stride, x86_fn_arg( func, arg_stride ) );
2717
2718 /* do */
2719 inner_loop = x86_get_label( func );
2720 {
2721 x86_push( func, aos_input );
2722 sse_movlps( func, make_xmm( 0 ), x86_make_disp( aos_input, 0 ) );
2723 sse_movlps( func, make_xmm( 3 ), x86_make_disp( aos_input, 8 ) );
2724 x86_add( func, aos_input, stride );
2725 sse_movhps( func, make_xmm( 0 ), x86_make_disp( aos_input, 0 ) );
2726 sse_movhps( func, make_xmm( 3 ), x86_make_disp( aos_input, 8 ) );
2727 x86_add( func, aos_input, stride );
2728 sse_movlps( func, make_xmm( 1 ), x86_make_disp( aos_input, 0 ) );
2729 sse_movlps( func, make_xmm( 4 ), x86_make_disp( aos_input, 8 ) );
2730 x86_add( func, aos_input, stride );
2731 sse_movhps( func, make_xmm( 1 ), x86_make_disp( aos_input, 0 ) );
2732 sse_movhps( func, make_xmm( 4 ), x86_make_disp( aos_input, 8 ) );
2733 x86_pop( func, aos_input );
2734
2735 sse_movaps( func, make_xmm( 2 ), make_xmm( 0 ) );
2736 sse_movaps( func, make_xmm( 5 ), make_xmm( 3 ) );
2737 sse_shufps( func, make_xmm( 0 ), make_xmm( 1 ), 0x88 );
2738 sse_shufps( func, make_xmm( 2 ), make_xmm( 1 ), 0xdd );
2739 sse_shufps( func, make_xmm( 3 ), make_xmm( 4 ), 0x88 );
2740 sse_shufps( func, make_xmm( 5 ), make_xmm( 4 ), 0xdd );
2741
2742 sse_movups( func, x86_make_disp( soa_input, 0 ), make_xmm( 0 ) );
2743 sse_movups( func, x86_make_disp( soa_input, 16 ), make_xmm( 2 ) );
2744 sse_movups( func, x86_make_disp( soa_input, 32 ), make_xmm( 3 ) );
2745 sse_movups( func, x86_make_disp( soa_input, 48 ), make_xmm( 5 ) );
2746
2747 /* Advance to next input */
2748 x86_lea( func, aos_input, x86_make_disp(aos_input, 16) );
2749 x86_lea( func, soa_input, x86_make_disp(soa_input, 64) );
2750 }
2751 /* while --num_inputs */
2752 x86_dec( func, num_inputs );
2753 x86_jcc( func, cc_NE, inner_loop );
2754
2755 /* Restore EBX */
2756 x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
2757 }
2758
2759 static void soa_to_aos( struct x86_function *func,
2760 uint arg_aos,
2761 uint arg_machine,
2762 uint arg_num,
2763 uint arg_stride )
2764 {
2765 struct x86_reg soa_output = x86_make_reg( file_REG32, reg_AX );
2766 struct x86_reg aos_output = x86_make_reg( file_REG32, reg_BX );
2767 struct x86_reg num_outputs = x86_make_reg( file_REG32, reg_CX );
2768 struct x86_reg temp = x86_make_reg( file_REG32, reg_DX );
2769 int inner_loop;
2770
2771 /* Save EBX */
2772 x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
2773
2774 x86_mov( func, aos_output, x86_fn_arg( func, arg_aos ) );
2775 x86_mov( func, soa_output, x86_fn_arg( func, arg_machine ) );
2776 x86_lea( func, soa_output,
2777 x86_make_disp( soa_output,
2778 Offset(struct tgsi_exec_machine, Outputs) ) );
2779 x86_mov( func, num_outputs, x86_fn_arg( func, arg_num ) );
2780
2781 /* do */
2782 inner_loop = x86_get_label( func );
2783 {
2784 sse_movups( func, make_xmm( 0 ), x86_make_disp( soa_output, 0 ) );
2785 sse_movups( func, make_xmm( 1 ), x86_make_disp( soa_output, 16 ) );
2786 sse_movups( func, make_xmm( 3 ), x86_make_disp( soa_output, 32 ) );
2787 sse_movups( func, make_xmm( 4 ), x86_make_disp( soa_output, 48 ) );
2788
2789 sse_movaps( func, make_xmm( 2 ), make_xmm( 0 ) );
2790 sse_movaps( func, make_xmm( 5 ), make_xmm( 3 ) );
2791 sse_unpcklps( func, make_xmm( 0 ), make_xmm( 1 ) );
2792 sse_unpckhps( func, make_xmm( 2 ), make_xmm( 1 ) );
2793 sse_unpcklps( func, make_xmm( 3 ), make_xmm( 4 ) );
2794 sse_unpckhps( func, make_xmm( 5 ), make_xmm( 4 ) );
2795
2796 x86_mov( func, temp, x86_fn_arg( func, arg_stride ) );
2797 x86_push( func, aos_output );
2798 sse_movlps( func, x86_make_disp( aos_output, 0 ), make_xmm( 0 ) );
2799 sse_movlps( func, x86_make_disp( aos_output, 8 ), make_xmm( 3 ) );
2800 x86_add( func, aos_output, temp );
2801 sse_movhps( func, x86_make_disp( aos_output, 0 ), make_xmm( 0 ) );
2802 sse_movhps( func, x86_make_disp( aos_output, 8 ), make_xmm( 3 ) );
2803 x86_add( func, aos_output, temp );
2804 sse_movlps( func, x86_make_disp( aos_output, 0 ), make_xmm( 2 ) );
2805 sse_movlps( func, x86_make_disp( aos_output, 8 ), make_xmm( 5 ) );
2806 x86_add( func, aos_output, temp );
2807 sse_movhps( func, x86_make_disp( aos_output, 0 ), make_xmm( 2 ) );
2808 sse_movhps( func, x86_make_disp( aos_output, 8 ), make_xmm( 5 ) );
2809 x86_pop( func, aos_output );
2810
2811 /* Advance to next output */
2812 x86_lea( func, aos_output, x86_make_disp(aos_output, 16) );
2813 x86_lea( func, soa_output, x86_make_disp(soa_output, 64) );
2814 }
2815 /* while --num_outputs */
2816 x86_dec( func, num_outputs );
2817 x86_jcc( func, cc_NE, inner_loop );
2818
2819 /* Restore EBX */
2820 x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
2821 }
2822
2823 /**
2824 * Translate a TGSI vertex/fragment shader to SSE2 code.
2825 * Slightly different things are done for vertex vs. fragment shaders.
2826 *
2827 * \param tokens the TGSI input shader
2828 * \param func the output SSE code/function
2829 * \param immediates buffer to place immediates, later passed to SSE func
2830 * \param return 1 for success, 0 if translation failed
2831 */
2832 unsigned
2833 tgsi_emit_sse2(
2834 const struct tgsi_token *tokens,
2835 struct x86_function *func,
2836 float (*immediates)[4],
2837 boolean do_swizzles )
2838 {
2839 struct tgsi_parse_context parse;
2840 unsigned ok = 1;
2841 uint num_immediates = 0;
2842
2843 util_init_math();
2844
2845 func->csr = func->store;
2846
2847 tgsi_parse_init( &parse, tokens );
2848
2849 /* Can't just use EDI, EBX without save/restoring them:
2850 */
2851 x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
2852 x86_push( func, x86_make_reg( file_REG32, reg_DI ) );
2853
2854 /*
2855 * Different function args for vertex/fragment shaders:
2856 */
2857 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX) {
2858 if (do_swizzles)
2859 aos_to_soa( func,
2860 4, /* aos_input */
2861 1, /* machine */
2862 5, /* num_inputs */
2863 6 ); /* input_stride */
2864 }
2865
2866 x86_mov(
2867 func,
2868 get_machine_base(),
2869 x86_fn_arg( func, 1 ) );
2870 x86_mov(
2871 func,
2872 get_const_base(),
2873 x86_fn_arg( func, 2 ) );
2874 x86_mov(
2875 func,
2876 get_immediate_base(),
2877 x86_fn_arg( func, 3 ) );
2878
2879 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_FRAGMENT) {
2880 x86_mov(
2881 func,
2882 get_coef_base(),
2883 x86_fn_arg( func, 4 ) );
2884 }
2885
2886 x86_mov(
2887 func,
2888 get_sampler_base(),
2889 x86_make_disp( get_machine_base(),
2890 Offset( struct tgsi_exec_machine, Samplers ) ) );
2891
2892
2893 while( !tgsi_parse_end_of_tokens( &parse ) && ok ) {
2894 tgsi_parse_token( &parse );
2895
2896 switch( parse.FullToken.Token.Type ) {
2897 case TGSI_TOKEN_TYPE_DECLARATION:
2898 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_FRAGMENT) {
2899 emit_declaration(
2900 func,
2901 &parse.FullToken.FullDeclaration );
2902 }
2903 break;
2904
2905 case TGSI_TOKEN_TYPE_INSTRUCTION:
2906 ok = emit_instruction(
2907 func,
2908 &parse.FullToken.FullInstruction );
2909
2910 if (!ok) {
2911 uint opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
2912 debug_printf("failed to translate tgsi opcode %d (%s) to SSE (%s)\n",
2913 opcode,
2914 tgsi_get_opcode_name(opcode),
2915 parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX ?
2916 "vertex shader" : "fragment shader");
2917 }
2918
2919 if (tgsi_check_soa_dependencies(&parse.FullToken.FullInstruction)) {
2920 uint opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
2921
2922 /* XXX: we only handle src/dst aliasing in a few opcodes
2923 * currently. Need to use an additional temporay to hold
2924 * the result in the cases where the code is too opaque to
2925 * fix.
2926 */
2927 if (opcode != TGSI_OPCODE_MOV) {
2928 debug_printf("Warning: src/dst aliasing in instruction"
2929 " is not handled:\n");
2930 tgsi_dump_instruction(&parse.FullToken.FullInstruction, 1);
2931 }
2932 }
2933 break;
2934
2935 case TGSI_TOKEN_TYPE_IMMEDIATE:
2936 /* simply copy the immediate values into the next immediates[] slot */
2937 {
2938 const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
2939 uint i;
2940 assert(size <= 4);
2941 assert(num_immediates < TGSI_EXEC_NUM_IMMEDIATES);
2942 for( i = 0; i < size; i++ ) {
2943 immediates[num_immediates][i] =
2944 parse.FullToken.FullImmediate.u[i].Float;
2945 }
2946 #if 0
2947 debug_printf("SSE FS immediate[%d] = %f %f %f %f\n",
2948 num_immediates,
2949 immediates[num_immediates][0],
2950 immediates[num_immediates][1],
2951 immediates[num_immediates][2],
2952 immediates[num_immediates][3]);
2953 #endif
2954 num_immediates++;
2955 }
2956 break;
2957 case TGSI_TOKEN_TYPE_PROPERTY:
2958 /* we just ignore them for now */
2959 break;
2960
2961 default:
2962 ok = 0;
2963 assert( 0 );
2964 }
2965 }
2966
2967 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX) {
2968 if (do_swizzles)
2969 soa_to_aos( func,
2970 7, /* aos_output */
2971 1, /* machine */
2972 8, /* num_outputs */
2973 9 ); /* output_stride */
2974 }
2975
2976 /* Can't just use EBX, EDI without save/restoring them:
2977 */
2978 x86_pop( func, x86_make_reg( file_REG32, reg_DI ) );
2979 x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
2980
2981 emit_ret( func );
2982
2983 tgsi_parse_free( &parse );
2984
2985 return ok;
2986 }
2987
2988 #endif /* PIPE_ARCH_X86 */
2989