projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallium/util: fix util_color_[un]pack[-ub] to be strict aliasing safe
[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 emit_inputf(
1292 func,
1293 xmm,
1294 reg->Register.Index,
1295 swizzle );
1296 break;
1297
1298 case TGSI_FILE_TEMPORARY:
1299 emit_tempf(
1300 func,
1301 xmm,
1302 reg->Register.Index,
1303 swizzle );
1304 break;
1305
1306 default:
1307 assert( 0 );
1308 }
1309 break;
1310
1311 default:
1312 assert( 0 );
1313 }
1314
1315 switch( tgsi_util_get_full_src_register_sign_mode( reg, chan_index ) ) {
1316 case TGSI_UTIL_SIGN_CLEAR:
1317 emit_abs( func, xmm );
1318 break;
1319
1320 case TGSI_UTIL_SIGN_SET:
1321 emit_setsign( func, xmm );
1322 break;
1323
1324 case TGSI_UTIL_SIGN_TOGGLE:
1325 emit_neg( func, xmm );
1326 break;
1327
1328 case TGSI_UTIL_SIGN_KEEP:
1329 break;
1330 }
1331 }
1332
1333 #define FETCH( FUNC, INST, XMM, INDEX, CHAN )\
1334 emit_fetch( FUNC, XMM, &(INST).Src[INDEX], CHAN )
1335
1336 /**
1337 * Register store.
1338 */
1339
1340 static void
1341 emit_store(
1342 struct x86_function *func,
1343 unsigned xmm,
1344 const struct tgsi_full_dst_register *reg,
1345 const struct tgsi_full_instruction *inst,
1346 unsigned chan_index )
1347 {
1348 switch( inst->Instruction.Saturate ) {
1349 case TGSI_SAT_NONE:
1350 break;
1351
1352 case TGSI_SAT_ZERO_ONE:
1353 sse_maxps(
1354 func,
1355 make_xmm( xmm ),
1356 get_temp(
1357 TGSI_EXEC_TEMP_00000000_I,
1358 TGSI_EXEC_TEMP_00000000_C ) );
1359
1360 sse_minps(
1361 func,
1362 make_xmm( xmm ),
1363 get_temp(
1364 TGSI_EXEC_TEMP_ONE_I,
1365 TGSI_EXEC_TEMP_ONE_C ) );
1366 break;
1367
1368 case TGSI_SAT_MINUS_PLUS_ONE:
1369 assert( 0 );
1370 break;
1371 }
1372
1373
1374 switch( reg->Register.File ) {
1375 case TGSI_FILE_OUTPUT:
1376 emit_output(
1377 func,
1378 xmm,
1379 reg->Register.Index,
1380 chan_index );
1381 break;
1382
1383 case TGSI_FILE_TEMPORARY:
1384 emit_temps(
1385 func,
1386 xmm,
1387 reg->Register.Index,
1388 chan_index );
1389 break;
1390
1391 case TGSI_FILE_ADDRESS:
1392 emit_addrs(
1393 func,
1394 xmm,
1395 reg->Register.Index,
1396 chan_index );
1397 break;
1398
1399 default:
1400 assert( 0 );
1401 }
1402 }
1403
1404 #define STORE( FUNC, INST, XMM, INDEX, CHAN )\
1405 emit_store( FUNC, XMM, &(INST).Dst[INDEX], &(INST), CHAN )
1406
1407
1408 static void PIPE_CDECL
1409 fetch_texel( struct tgsi_sampler **sampler,
1410 float *store )
1411 {
1412 #if 0
1413 uint j;
1414
1415 debug_printf("%s sampler: %p (%p) store: %p\n",
1416 __FUNCTION__,
1417 sampler, *sampler,
1418 store );
1419
1420 debug_printf("lodbias %f\n", store[12]);
1421
1422 for (j = 0; j < 4; j++)
1423 debug_printf("sample %d texcoord %f %f\n",
1424 j,
1425 store[0+j],
1426 store[4+j]);
1427 #endif
1428
1429 {
1430 float rgba[NUM_CHANNELS][QUAD_SIZE];
1431 (*sampler)->get_samples(*sampler,
1432 &store[0], /* s */
1433 &store[4], /* t */
1434 &store[8], /* r */
1435 store[12], /* lodbias */
1436 rgba); /* results */
1437
1438 memcpy( store, rgba, 16 * sizeof(float));
1439 }
1440
1441 #if 0
1442 for (j = 0; j < 4; j++)
1443 debug_printf("sample %d result %f %f %f %f\n",
1444 j,
1445 store[0+j],
1446 store[4+j],
1447 store[8+j],
1448 store[12+j]);
1449 #endif
1450 }
1451
1452 /**
1453 * High-level instruction translators.
1454 */
1455
1456 static void
1457 emit_tex( struct x86_function *func,
1458 const struct tgsi_full_instruction *inst,
1459 boolean lodbias,
1460 boolean projected)
1461 {
1462 const uint unit = inst->Src[1].Register.Index;
1463 struct x86_reg args[2];
1464 unsigned count;
1465 unsigned i;
1466
1467 assert(inst->Instruction.Texture);
1468 switch (inst->Texture.Texture) {
1469 case TGSI_TEXTURE_1D:
1470 count = 1;
1471 break;
1472 case TGSI_TEXTURE_2D:
1473 case TGSI_TEXTURE_RECT:
1474 count = 2;
1475 break;
1476 case TGSI_TEXTURE_SHADOW1D:
1477 case TGSI_TEXTURE_SHADOW2D:
1478 case TGSI_TEXTURE_SHADOWRECT:
1479 case TGSI_TEXTURE_3D:
1480 case TGSI_TEXTURE_CUBE:
1481 count = 3;
1482 break;
1483 default:
1484 assert(0);
1485 return;
1486 }
1487
1488 if (lodbias) {
1489 FETCH( func, *inst, 3, 0, 3 );
1490 }
1491 else {
1492 emit_tempf(
1493 func,
1494 3,
1495 TGSI_EXEC_TEMP_00000000_I,
1496 TGSI_EXEC_TEMP_00000000_C );
1497
1498 }
1499
1500 /* store lodbias whether enabled or not -- fetch_texel currently
1501 * respects it always.
1502 */
1503 sse_movaps( func,
1504 get_temp( TEMP_R0, 3 ),
1505 make_xmm( 3 ) );
1506
1507
1508 if (projected) {
1509 FETCH( func, *inst, 3, 0, 3 );
1510
1511 emit_rcp( func, 3, 3 );
1512 }
1513
1514 for (i = 0; i < count; i++) {
1515 FETCH( func, *inst, i, 0, i );
1516
1517 if (projected) {
1518 sse_mulps(
1519 func,
1520 make_xmm( i ),
1521 make_xmm( 3 ) );
1522 }
1523
1524 /* Store in the argument buffer:
1525 */
1526 sse_movaps(
1527 func,
1528 get_temp( TEMP_R0, i ),
1529 make_xmm( i ) );
1530 }
1531
1532 args[0] = get_temp( TEMP_R0, 0 );
1533 args[1] = get_sampler_ptr( unit );
1534
1535
1536 emit_func_call( func,
1537 0,
1538 args,
1539 Elements(args),
1540 fetch_texel );
1541
1542 /* If all four channels are enabled, could use a pointer to
1543 * dst[0].x instead of TEMP_R0 for store?
1544 */
1545 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, i ) {
1546
1547 sse_movaps(
1548 func,
1549 make_xmm( 0 ),
1550 get_temp( TEMP_R0, i ) );
1551
1552 STORE( func, *inst, 0, 0, i );
1553 }
1554 }
1555
1556
1557 static void
1558 emit_kil(
1559 struct x86_function *func,
1560 const struct tgsi_full_src_register *reg )
1561 {
1562 unsigned uniquemask;
1563 unsigned unique_count = 0;
1564 unsigned chan_index;
1565 unsigned i;
1566
1567 /* This mask stores component bits that were already tested. Note that
1568 * we test if the value is less than zero, so 1.0 and 0.0 need not to be
1569 * tested. */
1570 uniquemask = 0;
1571
1572 FOR_EACH_CHANNEL( chan_index ) {
1573 unsigned swizzle;
1574
1575 /* unswizzle channel */
1576 swizzle = tgsi_util_get_full_src_register_swizzle(
1577 reg,
1578 chan_index );
1579
1580 /* check if the component has not been already tested */
1581 if( !(uniquemask & (1 << swizzle)) ) {
1582 uniquemask |= 1 << swizzle;
1583
1584 /* allocate register */
1585 emit_fetch(
1586 func,
1587 unique_count++,
1588 reg,
1589 chan_index );
1590 }
1591 }
1592
1593 x86_push(
1594 func,
1595 x86_make_reg( file_REG32, reg_AX ) );
1596 x86_push(
1597 func,
1598 x86_make_reg( file_REG32, reg_DX ) );
1599
1600 for (i = 0 ; i < unique_count; i++ ) {
1601 struct x86_reg dataXMM = make_xmm(i);
1602
1603 sse_cmpps(
1604 func,
1605 dataXMM,
1606 get_temp(
1607 TGSI_EXEC_TEMP_00000000_I,
1608 TGSI_EXEC_TEMP_00000000_C ),
1609 cc_LessThan );
1610
1611 if( i == 0 ) {
1612 sse_movmskps(
1613 func,
1614 x86_make_reg( file_REG32, reg_AX ),
1615 dataXMM );
1616 }
1617 else {
1618 sse_movmskps(
1619 func,
1620 x86_make_reg( file_REG32, reg_DX ),
1621 dataXMM );
1622 x86_or(
1623 func,
1624 x86_make_reg( file_REG32, reg_AX ),
1625 x86_make_reg( file_REG32, reg_DX ) );
1626 }
1627 }
1628
1629 x86_or(
1630 func,
1631 get_temp(
1632 TGSI_EXEC_TEMP_KILMASK_I,
1633 TGSI_EXEC_TEMP_KILMASK_C ),
1634 x86_make_reg( file_REG32, reg_AX ) );
1635
1636 x86_pop(
1637 func,
1638 x86_make_reg( file_REG32, reg_DX ) );
1639 x86_pop(
1640 func,
1641 x86_make_reg( file_REG32, reg_AX ) );
1642 }
1643
1644
1645 static void
1646 emit_kilp(
1647 struct x86_function *func )
1648 {
1649 /* XXX todo / fix me */
1650 }
1651
1652
1653 static void
1654 emit_setcc(
1655 struct x86_function *func,
1656 struct tgsi_full_instruction *inst,
1657 enum sse_cc cc )
1658 {
1659 unsigned chan_index;
1660
1661 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1662 FETCH( func, *inst, 0, 0, chan_index );
1663 FETCH( func, *inst, 1, 1, chan_index );
1664 sse_cmpps(
1665 func,
1666 make_xmm( 0 ),
1667 make_xmm( 1 ),
1668 cc );
1669 sse_andps(
1670 func,
1671 make_xmm( 0 ),
1672 get_temp(
1673 TEMP_ONE_I,
1674 TEMP_ONE_C ) );
1675 STORE( func, *inst, 0, 0, chan_index );
1676 }
1677 }
1678
1679 static void
1680 emit_cmp(
1681 struct x86_function *func,
1682 struct tgsi_full_instruction *inst )
1683 {
1684 unsigned chan_index;
1685
1686 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1687 FETCH( func, *inst, 0, 0, chan_index );
1688 FETCH( func, *inst, 1, 1, chan_index );
1689 FETCH( func, *inst, 2, 2, chan_index );
1690 sse_cmpps(
1691 func,
1692 make_xmm( 0 ),
1693 get_temp(
1694 TGSI_EXEC_TEMP_00000000_I,
1695 TGSI_EXEC_TEMP_00000000_C ),
1696 cc_LessThan );
1697 sse_andps(
1698 func,
1699 make_xmm( 1 ),
1700 make_xmm( 0 ) );
1701 sse_andnps(
1702 func,
1703 make_xmm( 0 ),
1704 make_xmm( 2 ) );
1705 sse_orps(
1706 func,
1707 make_xmm( 0 ),
1708 make_xmm( 1 ) );
1709 STORE( func, *inst, 0, 0, chan_index );
1710 }
1711 }
1712
1713
1714 /**
1715 * Check if inst src/dest regs use indirect addressing into temporary
1716 * register file.
1717 */
1718 static boolean
1719 indirect_temp_reference(const struct tgsi_full_instruction *inst)
1720 {
1721 uint i;
1722 for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
1723 const struct tgsi_full_src_register *reg = &inst->Src[i];
1724 if (reg->Register.File == TGSI_FILE_TEMPORARY &&
1725 reg->Register.Indirect)
1726 return TRUE;
1727 }
1728 for (i = 0; i < inst->Instruction.NumDstRegs; i++) {
1729 const struct tgsi_full_dst_register *reg = &inst->Dst[i];
1730 if (reg->Register.File == TGSI_FILE_TEMPORARY &&
1731 reg->Register.Indirect)
1732 return TRUE;
1733 }
1734 return FALSE;
1735 }
1736
1737
1738 static int
1739 emit_instruction(
1740 struct x86_function *func,
1741 struct tgsi_full_instruction *inst )
1742 {
1743 unsigned chan_index;
1744
1745 /* we can't handle indirect addressing into temp register file yet */
1746 if (indirect_temp_reference(inst))
1747 return FALSE;
1748
1749 switch (inst->Instruction.Opcode) {
1750 case TGSI_OPCODE_ARL:
1751 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1752 FETCH( func, *inst, 0, 0, chan_index );
1753 emit_flr(func, 0, 0);
1754 emit_f2it( func, 0 );
1755 STORE( func, *inst, 0, 0, chan_index );
1756 }
1757 break;
1758
1759 case TGSI_OPCODE_MOV:
1760 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1761 FETCH( func, *inst, 4 + chan_index, 0, chan_index );
1762 }
1763 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1764 STORE( func, *inst, 4 + chan_index, 0, chan_index );
1765 }
1766 break;
1767
1768 case TGSI_OPCODE_LIT:
1769 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1770 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) ) {
1771 emit_tempf(
1772 func,
1773 0,
1774 TEMP_ONE_I,
1775 TEMP_ONE_C);
1776 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ) {
1777 STORE( func, *inst, 0, 0, CHAN_X );
1778 }
1779 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) ) {
1780 STORE( func, *inst, 0, 0, CHAN_W );
1781 }
1782 }
1783 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
1784 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
1785 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
1786 FETCH( func, *inst, 0, 0, CHAN_X );
1787 sse_maxps(
1788 func,
1789 make_xmm( 0 ),
1790 get_temp(
1791 TGSI_EXEC_TEMP_00000000_I,
1792 TGSI_EXEC_TEMP_00000000_C ) );
1793 STORE( func, *inst, 0, 0, CHAN_Y );
1794 }
1795 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
1796 /* XMM[1] = SrcReg[0].yyyy */
1797 FETCH( func, *inst, 1, 0, CHAN_Y );
1798 /* XMM[1] = max(XMM[1], 0) */
1799 sse_maxps(
1800 func,
1801 make_xmm( 1 ),
1802 get_temp(
1803 TGSI_EXEC_TEMP_00000000_I,
1804 TGSI_EXEC_TEMP_00000000_C ) );
1805 /* XMM[2] = SrcReg[0].wwww */
1806 FETCH( func, *inst, 2, 0, CHAN_W );
1807 /* XMM[2] = min(XMM[2], 128.0) */
1808 sse_minps(
1809 func,
1810 make_xmm( 2 ),
1811 get_temp(
1812 TGSI_EXEC_TEMP_128_I,
1813 TGSI_EXEC_TEMP_128_C ) );
1814 /* XMM[2] = max(XMM[2], -128.0) */
1815 sse_maxps(
1816 func,
1817 make_xmm( 2 ),
1818 get_temp(
1819 TGSI_EXEC_TEMP_MINUS_128_I,
1820 TGSI_EXEC_TEMP_MINUS_128_C ) );
1821 emit_pow( func, 3, 1, 1, 2 );
1822 FETCH( func, *inst, 0, 0, CHAN_X );
1823 sse_xorps(
1824 func,
1825 make_xmm( 2 ),
1826 make_xmm( 2 ) );
1827 sse_cmpps(
1828 func,
1829 make_xmm( 2 ),
1830 make_xmm( 0 ),
1831 cc_LessThan );
1832 sse_andps(
1833 func,
1834 make_xmm( 2 ),
1835 make_xmm( 1 ) );
1836 STORE( func, *inst, 2, 0, CHAN_Z );
1837 }
1838 }
1839 break;
1840
1841 case TGSI_OPCODE_RCP:
1842 FETCH( func, *inst, 0, 0, CHAN_X );
1843 emit_rcp( func, 0, 0 );
1844 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1845 STORE( func, *inst, 0, 0, chan_index );
1846 }
1847 break;
1848
1849 case TGSI_OPCODE_RSQ:
1850 FETCH( func, *inst, 0, 0, CHAN_X );
1851 emit_abs( func, 0 );
1852 emit_rsqrt( func, 1, 0 );
1853 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1854 STORE( func, *inst, 1, 0, chan_index );
1855 }
1856 break;
1857
1858 case TGSI_OPCODE_EXP:
1859 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1860 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
1861 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1862 FETCH( func, *inst, 0, 0, CHAN_X );
1863 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1864 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1865 emit_MOV( func, 1, 0 );
1866 emit_flr( func, 2, 1 );
1867 /* dst.x = ex2(floor(src.x)) */
1868 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X )) {
1869 emit_MOV( func, 2, 1 );
1870 emit_ex2( func, 3, 2 );
1871 STORE( func, *inst, 2, 0, CHAN_X );
1872 }
1873 /* dst.y = src.x - floor(src.x) */
1874 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1875 emit_MOV( func, 2, 0 );
1876 emit_sub( func, 2, 1 );
1877 STORE( func, *inst, 2, 0, CHAN_Y );
1878 }
1879 }
1880 /* dst.z = ex2(src.x) */
1881 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1882 emit_ex2( func, 3, 0 );
1883 STORE( func, *inst, 0, 0, CHAN_Z );
1884 }
1885 }
1886 /* dst.w = 1.0 */
1887 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W )) {
1888 emit_tempf( func, 0, TEMP_ONE_I, TEMP_ONE_C );
1889 STORE( func, *inst, 0, 0, CHAN_W );
1890 }
1891 break;
1892
1893 case TGSI_OPCODE_LOG:
1894 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1895 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
1896 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1897 FETCH( func, *inst, 0, 0, CHAN_X );
1898 emit_abs( func, 0 );
1899 emit_MOV( func, 1, 0 );
1900 emit_lg2( func, 2, 1 );
1901 /* dst.z = lg2(abs(src.x)) */
1902 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z )) {
1903 STORE( func, *inst, 1, 0, CHAN_Z );
1904 }
1905 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
1906 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1907 emit_flr( func, 2, 1 );
1908 /* dst.x = floor(lg2(abs(src.x))) */
1909 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X )) {
1910 STORE( func, *inst, 1, 0, CHAN_X );
1911 }
1912 /* dst.x = abs(src)/ex2(floor(lg2(abs(src.x)))) */
1913 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y )) {
1914 emit_ex2( func, 2, 1 );
1915 emit_rcp( func, 1, 1 );
1916 emit_mul( func, 0, 1 );
1917 STORE( func, *inst, 0, 0, CHAN_Y );
1918 }
1919 }
1920 }
1921 /* dst.w = 1.0 */
1922 if (IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W )) {
1923 emit_tempf( func, 0, TEMP_ONE_I, TEMP_ONE_C );
1924 STORE( func, *inst, 0, 0, CHAN_W );
1925 }
1926 break;
1927
1928 case TGSI_OPCODE_MUL:
1929 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1930 FETCH( func, *inst, 0, 0, chan_index );
1931 FETCH( func, *inst, 1, 1, chan_index );
1932 emit_mul( func, 0, 1 );
1933 STORE( func, *inst, 0, 0, chan_index );
1934 }
1935 break;
1936
1937 case TGSI_OPCODE_ADD:
1938 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1939 FETCH( func, *inst, 0, 0, chan_index );
1940 FETCH( func, *inst, 1, 1, chan_index );
1941 emit_add( func, 0, 1 );
1942 STORE( func, *inst, 0, 0, chan_index );
1943 }
1944 break;
1945
1946 case TGSI_OPCODE_DP3:
1947 FETCH( func, *inst, 0, 0, CHAN_X );
1948 FETCH( func, *inst, 1, 1, CHAN_X );
1949 emit_mul( func, 0, 1 );
1950 FETCH( func, *inst, 1, 0, CHAN_Y );
1951 FETCH( func, *inst, 2, 1, CHAN_Y );
1952 emit_mul( func, 1, 2 );
1953 emit_add( func, 0, 1 );
1954 FETCH( func, *inst, 1, 0, CHAN_Z );
1955 FETCH( func, *inst, 2, 1, CHAN_Z );
1956 emit_mul( func, 1, 2 );
1957 emit_add( func, 0, 1 );
1958 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1959 STORE( func, *inst, 0, 0, chan_index );
1960 }
1961 break;
1962
1963 case TGSI_OPCODE_DP4:
1964 FETCH( func, *inst, 0, 0, CHAN_X );
1965 FETCH( func, *inst, 1, 1, CHAN_X );
1966 emit_mul( func, 0, 1 );
1967 FETCH( func, *inst, 1, 0, CHAN_Y );
1968 FETCH( func, *inst, 2, 1, CHAN_Y );
1969 emit_mul( func, 1, 2 );
1970 emit_add( func, 0, 1 );
1971 FETCH( func, *inst, 1, 0, CHAN_Z );
1972 FETCH( func, *inst, 2, 1, CHAN_Z );
1973 emit_mul(func, 1, 2 );
1974 emit_add(func, 0, 1 );
1975 FETCH( func, *inst, 1, 0, CHAN_W );
1976 FETCH( func, *inst, 2, 1, CHAN_W );
1977 emit_mul( func, 1, 2 );
1978 emit_add( func, 0, 1 );
1979 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
1980 STORE( func, *inst, 0, 0, chan_index );
1981 }
1982 break;
1983
1984 case TGSI_OPCODE_DST:
1985 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
1986 emit_tempf(
1987 func,
1988 0,
1989 TEMP_ONE_I,
1990 TEMP_ONE_C );
1991 STORE( func, *inst, 0, 0, CHAN_X );
1992 }
1993 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
1994 FETCH( func, *inst, 0, 0, CHAN_Y );
1995 FETCH( func, *inst, 1, 1, CHAN_Y );
1996 emit_mul( func, 0, 1 );
1997 STORE( func, *inst, 0, 0, CHAN_Y );
1998 }
1999 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
2000 FETCH( func, *inst, 0, 0, CHAN_Z );
2001 STORE( func, *inst, 0, 0, CHAN_Z );
2002 }
2003 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
2004 FETCH( func, *inst, 0, 1, CHAN_W );
2005 STORE( func, *inst, 0, 0, CHAN_W );
2006 }
2007 break;
2008
2009 case TGSI_OPCODE_MIN:
2010 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2011 FETCH( func, *inst, 0, 0, chan_index );
2012 FETCH( func, *inst, 1, 1, chan_index );
2013 sse_minps(
2014 func,
2015 make_xmm( 0 ),
2016 make_xmm( 1 ) );
2017 STORE( func, *inst, 0, 0, chan_index );
2018 }
2019 break;
2020
2021 case TGSI_OPCODE_MAX:
2022 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2023 FETCH( func, *inst, 0, 0, chan_index );
2024 FETCH( func, *inst, 1, 1, chan_index );
2025 sse_maxps(
2026 func,
2027 make_xmm( 0 ),
2028 make_xmm( 1 ) );
2029 STORE( func, *inst, 0, 0, chan_index );
2030 }
2031 break;
2032
2033 case TGSI_OPCODE_SLT:
2034 emit_setcc( func, inst, cc_LessThan );
2035 break;
2036
2037 case TGSI_OPCODE_SGE:
2038 emit_setcc( func, inst, cc_NotLessThan );
2039 break;
2040
2041 case TGSI_OPCODE_MAD:
2042 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2043 FETCH( func, *inst, 0, 0, chan_index );
2044 FETCH( func, *inst, 1, 1, chan_index );
2045 FETCH( func, *inst, 2, 2, chan_index );
2046 emit_mul( func, 0, 1 );
2047 emit_add( func, 0, 2 );
2048 STORE( func, *inst, 0, 0, chan_index );
2049 }
2050 break;
2051
2052 case TGSI_OPCODE_SUB:
2053 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2054 FETCH( func, *inst, 0, 0, chan_index );
2055 FETCH( func, *inst, 1, 1, chan_index );
2056 emit_sub( func, 0, 1 );
2057 STORE( func, *inst, 0, 0, chan_index );
2058 }
2059 break;
2060
2061 case TGSI_OPCODE_LRP:
2062 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2063 FETCH( func, *inst, 0, 0, chan_index );
2064 FETCH( func, *inst, 1, 1, chan_index );
2065 FETCH( func, *inst, 2, 2, chan_index );
2066 emit_sub( func, 1, 2 );
2067 emit_mul( func, 0, 1 );
2068 emit_add( func, 0, 2 );
2069 STORE( func, *inst, 0, 0, chan_index );
2070 }
2071 break;
2072
2073 case TGSI_OPCODE_CND:
2074 return 0;
2075 break;
2076
2077 case TGSI_OPCODE_DP2A:
2078 FETCH( func, *inst, 0, 0, CHAN_X ); /* xmm0 = src[0].x */
2079 FETCH( func, *inst, 1, 1, CHAN_X ); /* xmm1 = src[1].x */
2080 emit_mul( func, 0, 1 ); /* xmm0 = xmm0 * xmm1 */
2081 FETCH( func, *inst, 1, 0, CHAN_Y ); /* xmm1 = src[0].y */
2082 FETCH( func, *inst, 2, 1, CHAN_Y ); /* xmm2 = src[1].y */
2083 emit_mul( func, 1, 2 ); /* xmm1 = xmm1 * xmm2 */
2084 emit_add( func, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2085 FETCH( func, *inst, 1, 2, CHAN_X ); /* xmm1 = src[2].x */
2086 emit_add( func, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2087 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2088 STORE( func, *inst, 0, 0, chan_index ); /* dest[ch] = xmm0 */
2089 }
2090 break;
2091
2092 case TGSI_OPCODE_FRC:
2093 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2094 FETCH( func, *inst, 0, 0, chan_index );
2095 emit_frc( func, 0, 0 );
2096 STORE( func, *inst, 0, 0, chan_index );
2097 }
2098 break;
2099
2100 case TGSI_OPCODE_CLAMP:
2101 return 0;
2102 break;
2103
2104 case TGSI_OPCODE_FLR:
2105 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2106 FETCH( func, *inst, 0, 0, chan_index );
2107 emit_flr( func, 0, 0 );
2108 STORE( func, *inst, 0, 0, chan_index );
2109 }
2110 break;
2111
2112 case TGSI_OPCODE_ROUND:
2113 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2114 FETCH( func, *inst, 0, 0, chan_index );
2115 emit_rnd( func, 0, 0 );
2116 STORE( func, *inst, 0, 0, chan_index );
2117 }
2118 break;
2119
2120 case TGSI_OPCODE_EX2:
2121 FETCH( func, *inst, 0, 0, CHAN_X );
2122 emit_ex2( func, 0, 0 );
2123 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2124 STORE( func, *inst, 0, 0, chan_index );
2125 }
2126 break;
2127
2128 case TGSI_OPCODE_LG2:
2129 FETCH( func, *inst, 0, 0, CHAN_X );
2130 emit_lg2( func, 0, 0 );
2131 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2132 STORE( func, *inst, 0, 0, chan_index );
2133 }
2134 break;
2135
2136 case TGSI_OPCODE_POW:
2137 FETCH( func, *inst, 0, 0, CHAN_X );
2138 FETCH( func, *inst, 1, 1, CHAN_X );
2139 emit_pow( func, 0, 0, 0, 1 );
2140 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2141 STORE( func, *inst, 0, 0, chan_index );
2142 }
2143 break;
2144
2145 case TGSI_OPCODE_XPD:
2146 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
2147 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
2148 FETCH( func, *inst, 1, 1, CHAN_Z );
2149 FETCH( func, *inst, 3, 0, CHAN_Z );
2150 }
2151 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
2152 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
2153 FETCH( func, *inst, 0, 0, CHAN_Y );
2154 FETCH( func, *inst, 4, 1, CHAN_Y );
2155 }
2156 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
2157 emit_MOV( func, 2, 0 );
2158 emit_mul( func, 2, 1 );
2159 emit_MOV( func, 5, 3 );
2160 emit_mul( func, 5, 4 );
2161 emit_sub( func, 2, 5 );
2162 STORE( func, *inst, 2, 0, CHAN_X );
2163 }
2164 if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
2165 IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
2166 FETCH( func, *inst, 2, 1, CHAN_X );
2167 FETCH( func, *inst, 5, 0, CHAN_X );
2168 }
2169 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
2170 emit_mul( func, 3, 2 );
2171 emit_mul( func, 1, 5 );
2172 emit_sub( func, 3, 1 );
2173 STORE( func, *inst, 3, 0, CHAN_Y );
2174 }
2175 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
2176 emit_mul( func, 5, 4 );
2177 emit_mul( func, 0, 2 );
2178 emit_sub( func, 5, 0 );
2179 STORE( func, *inst, 5, 0, CHAN_Z );
2180 }
2181 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
2182 emit_tempf(
2183 func,
2184 0,
2185 TEMP_ONE_I,
2186 TEMP_ONE_C );
2187 STORE( func, *inst, 0, 0, CHAN_W );
2188 }
2189 break;
2190
2191 case TGSI_OPCODE_ABS:
2192 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2193 FETCH( func, *inst, 0, 0, chan_index );
2194 emit_abs( func, 0) ;
2195
2196 STORE( func, *inst, 0, 0, chan_index );
2197 }
2198 break;
2199
2200 case TGSI_OPCODE_RCC:
2201 return 0;
2202 break;
2203
2204 case TGSI_OPCODE_DPH:
2205 FETCH( func, *inst, 0, 0, CHAN_X );
2206 FETCH( func, *inst, 1, 1, CHAN_X );
2207 emit_mul( func, 0, 1 );
2208 FETCH( func, *inst, 1, 0, CHAN_Y );
2209 FETCH( func, *inst, 2, 1, CHAN_Y );
2210 emit_mul( func, 1, 2 );
2211 emit_add( func, 0, 1 );
2212 FETCH( func, *inst, 1, 0, CHAN_Z );
2213 FETCH( func, *inst, 2, 1, CHAN_Z );
2214 emit_mul( func, 1, 2 );
2215 emit_add( func, 0, 1 );
2216 FETCH( func, *inst, 1, 1, CHAN_W );
2217 emit_add( func, 0, 1 );
2218 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2219 STORE( func, *inst, 0, 0, chan_index );
2220 }
2221 break;
2222
2223 case TGSI_OPCODE_COS:
2224 FETCH( func, *inst, 0, 0, CHAN_X );
2225 emit_cos( func, 0, 0 );
2226 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2227 STORE( func, *inst, 0, 0, chan_index );
2228 }
2229 break;
2230
2231 case TGSI_OPCODE_DDX:
2232 return 0;
2233 break;
2234
2235 case TGSI_OPCODE_DDY:
2236 return 0;
2237 break;
2238
2239 case TGSI_OPCODE_KILP:
2240 /* predicated kill */
2241 emit_kilp( func );
2242 return 0; /* XXX fix me */
2243 break;
2244
2245 case TGSI_OPCODE_KIL:
2246 /* conditional kill */
2247 emit_kil( func, &inst->Src[0] );
2248 break;
2249
2250 case TGSI_OPCODE_PK2H:
2251 return 0;
2252 break;
2253
2254 case TGSI_OPCODE_PK2US:
2255 return 0;
2256 break;
2257
2258 case TGSI_OPCODE_PK4B:
2259 return 0;
2260 break;
2261
2262 case TGSI_OPCODE_PK4UB:
2263 return 0;
2264 break;
2265
2266 case TGSI_OPCODE_RFL:
2267 return 0;
2268 break;
2269
2270 case TGSI_OPCODE_SEQ:
2271 emit_setcc( func, inst, cc_Equal );
2272 break;
2273
2274 case TGSI_OPCODE_SFL:
2275 return 0;
2276 break;
2277
2278 case TGSI_OPCODE_SGT:
2279 emit_setcc( func, inst, cc_NotLessThanEqual );
2280 break;
2281
2282 case TGSI_OPCODE_SIN:
2283 FETCH( func, *inst, 0, 0, CHAN_X );
2284 emit_sin( func, 0, 0 );
2285 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2286 STORE( func, *inst, 0, 0, chan_index );
2287 }
2288 break;
2289
2290 case TGSI_OPCODE_SLE:
2291 emit_setcc( func, inst, cc_LessThanEqual );
2292 break;
2293
2294 case TGSI_OPCODE_SNE:
2295 emit_setcc( func, inst, cc_NotEqual );
2296 break;
2297
2298 case TGSI_OPCODE_STR:
2299 return 0;
2300 break;
2301
2302 case TGSI_OPCODE_TEX:
2303 emit_tex( func, inst, FALSE, FALSE );
2304 break;
2305
2306 case TGSI_OPCODE_TXD:
2307 return 0;
2308 break;
2309
2310 case TGSI_OPCODE_UP2H:
2311 return 0;
2312 break;
2313
2314 case TGSI_OPCODE_UP2US:
2315 return 0;
2316 break;
2317
2318 case TGSI_OPCODE_UP4B:
2319 return 0;
2320 break;
2321
2322 case TGSI_OPCODE_UP4UB:
2323 return 0;
2324 break;
2325
2326 case TGSI_OPCODE_X2D:
2327 return 0;
2328 break;
2329
2330 case TGSI_OPCODE_ARA:
2331 return 0;
2332 break;
2333
2334 case TGSI_OPCODE_ARR:
2335 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2336 FETCH( func, *inst, 0, 0, chan_index );
2337 emit_rnd( func, 0, 0 );
2338 emit_f2it( func, 0 );
2339 STORE( func, *inst, 0, 0, chan_index );
2340 }
2341 break;
2342
2343 case TGSI_OPCODE_BRA:
2344 return 0;
2345 break;
2346
2347 case TGSI_OPCODE_CAL:
2348 return 0;
2349 break;
2350
2351 case TGSI_OPCODE_RET:
2352 emit_ret( func );
2353 break;
2354
2355 case TGSI_OPCODE_END:
2356 break;
2357
2358 case TGSI_OPCODE_SSG:
2359 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2360 FETCH( func, *inst, 0, 0, chan_index );
2361 emit_sgn( func, 0, 0 );
2362 STORE( func, *inst, 0, 0, chan_index );
2363 }
2364 break;
2365
2366 case TGSI_OPCODE_CMP:
2367 emit_cmp (func, inst);
2368 break;
2369
2370 case TGSI_OPCODE_SCS:
2371 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
2372 FETCH( func, *inst, 0, 0, CHAN_X );
2373 emit_cos( func, 0, 0 );
2374 STORE( func, *inst, 0, 0, CHAN_X );
2375 }
2376 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
2377 FETCH( func, *inst, 0, 0, CHAN_X );
2378 emit_sin( func, 0, 0 );
2379 STORE( func, *inst, 0, 0, CHAN_Y );
2380 }
2381 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
2382 emit_tempf(
2383 func,
2384 0,
2385 TGSI_EXEC_TEMP_00000000_I,
2386 TGSI_EXEC_TEMP_00000000_C );
2387 STORE( func, *inst, 0, 0, CHAN_Z );
2388 }
2389 IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
2390 emit_tempf(
2391 func,
2392 0,
2393 TEMP_ONE_I,
2394 TEMP_ONE_C );
2395 STORE( func, *inst, 0, 0, CHAN_W );
2396 }
2397 break;
2398
2399 case TGSI_OPCODE_TXB:
2400 emit_tex( func, inst, TRUE, FALSE );
2401 break;
2402
2403 case TGSI_OPCODE_NRM:
2404 /* fall-through */
2405 case TGSI_OPCODE_NRM4:
2406 /* 3 or 4-component normalization */
2407 {
2408 uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
2409
2410 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) ||
2411 IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y) ||
2412 IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z) ||
2413 (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W) && dims == 4)) {
2414
2415 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
2416
2417 /* xmm4 = src.x */
2418 /* xmm0 = src.x * src.x */
2419 FETCH(func, *inst, 0, 0, CHAN_X);
2420 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X)) {
2421 emit_MOV(func, 4, 0);
2422 }
2423 emit_mul(func, 0, 0);
2424
2425 /* xmm5 = src.y */
2426 /* xmm0 = xmm0 + src.y * src.y */
2427 FETCH(func, *inst, 1, 0, CHAN_Y);
2428 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) {
2429 emit_MOV(func, 5, 1);
2430 }
2431 emit_mul(func, 1, 1);
2432 emit_add(func, 0, 1);
2433
2434 /* xmm6 = src.z */
2435 /* xmm0 = xmm0 + src.z * src.z */
2436 FETCH(func, *inst, 1, 0, CHAN_Z);
2437 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) {
2438 emit_MOV(func, 6, 1);
2439 }
2440 emit_mul(func, 1, 1);
2441 emit_add(func, 0, 1);
2442
2443 if (dims == 4) {
2444 /* xmm7 = src.w */
2445 /* xmm0 = xmm0 + src.w * src.w */
2446 FETCH(func, *inst, 1, 0, CHAN_W);
2447 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W)) {
2448 emit_MOV(func, 7, 1);
2449 }
2450 emit_mul(func, 1, 1);
2451 emit_add(func, 0, 1);
2452 }
2453
2454 /* xmm1 = 1 / sqrt(xmm0) */
2455 emit_rsqrt(func, 1, 0);
2456
2457 /* dst.x = xmm1 * src.x */
2458 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X)) {
2459 emit_mul(func, 4, 1);
2460 STORE(func, *inst, 4, 0, CHAN_X);
2461 }
2462
2463 /* dst.y = xmm1 * src.y */
2464 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y)) {
2465 emit_mul(func, 5, 1);
2466 STORE(func, *inst, 5, 0, CHAN_Y);
2467 }
2468
2469 /* dst.z = xmm1 * src.z */
2470 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) {
2471 emit_mul(func, 6, 1);
2472 STORE(func, *inst, 6, 0, CHAN_Z);
2473 }
2474
2475 /* dst.w = xmm1 * src.w */
2476 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_X) && dims == 4) {
2477 emit_mul(func, 7, 1);
2478 STORE(func, *inst, 7, 0, CHAN_W);
2479 }
2480 }
2481
2482 /* dst0.w = 1.0 */
2483 if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W) && dims == 3) {
2484 emit_tempf(func, 0, TEMP_ONE_I, TEMP_ONE_C);
2485 STORE(func, *inst, 0, 0, CHAN_W);
2486 }
2487 }
2488 break;
2489
2490 case TGSI_OPCODE_DIV:
2491 return 0;
2492 break;
2493
2494 case TGSI_OPCODE_DP2:
2495 FETCH( func, *inst, 0, 0, CHAN_X ); /* xmm0 = src[0].x */
2496 FETCH( func, *inst, 1, 1, CHAN_X ); /* xmm1 = src[1].x */
2497 emit_mul( func, 0, 1 ); /* xmm0 = xmm0 * xmm1 */
2498 FETCH( func, *inst, 1, 0, CHAN_Y ); /* xmm1 = src[0].y */
2499 FETCH( func, *inst, 2, 1, CHAN_Y ); /* xmm2 = src[1].y */
2500 emit_mul( func, 1, 2 ); /* xmm1 = xmm1 * xmm2 */
2501 emit_add( func, 0, 1 ); /* xmm0 = xmm0 + xmm1 */
2502 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2503 STORE( func, *inst, 0, 0, chan_index ); /* dest[ch] = xmm0 */
2504 }
2505 break;
2506
2507 case TGSI_OPCODE_TXL:
2508 emit_tex( func, inst, TRUE, FALSE );
2509 break;
2510
2511 case TGSI_OPCODE_TXP:
2512 emit_tex( func, inst, FALSE, TRUE );
2513 break;
2514
2515 case TGSI_OPCODE_BRK:
2516 return 0;
2517 break;
2518
2519 case TGSI_OPCODE_IF:
2520 return 0;
2521 break;
2522
2523 case TGSI_OPCODE_BGNFOR:
2524 return 0;
2525 break;
2526
2527 case TGSI_OPCODE_REP:
2528 return 0;
2529 break;
2530
2531 case TGSI_OPCODE_ELSE:
2532 return 0;
2533 break;
2534
2535 case TGSI_OPCODE_ENDIF:
2536 return 0;
2537 break;
2538
2539 case TGSI_OPCODE_ENDFOR:
2540 return 0;
2541 break;
2542
2543 case TGSI_OPCODE_ENDREP:
2544 return 0;
2545 break;
2546
2547 case TGSI_OPCODE_PUSHA:
2548 return 0;
2549 break;
2550
2551 case TGSI_OPCODE_POPA:
2552 return 0;
2553 break;
2554
2555 case TGSI_OPCODE_CEIL:
2556 return 0;
2557 break;
2558
2559 case TGSI_OPCODE_I2F:
2560 return 0;
2561 break;
2562
2563 case TGSI_OPCODE_NOT:
2564 return 0;
2565 break;
2566
2567 case TGSI_OPCODE_TRUNC:
2568 FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
2569 FETCH( func, *inst, 0, 0, chan_index );
2570 emit_f2it( func, 0 );
2571 emit_i2f( func, 0 );
2572 STORE( func, *inst, 0, 0, chan_index );
2573 }
2574 break;
2575
2576 case TGSI_OPCODE_SHL:
2577 return 0;
2578 break;
2579
2580 case TGSI_OPCODE_SHR:
2581 return 0;
2582 break;
2583
2584 case TGSI_OPCODE_AND:
2585 return 0;
2586 break;
2587
2588 case TGSI_OPCODE_OR:
2589 return 0;
2590 break;
2591
2592 case TGSI_OPCODE_MOD:
2593 return 0;
2594 break;
2595
2596 case TGSI_OPCODE_XOR:
2597 return 0;
2598 break;
2599
2600 case TGSI_OPCODE_SAD:
2601 return 0;
2602 break;
2603
2604 case TGSI_OPCODE_TXF:
2605 return 0;
2606 break;
2607
2608 case TGSI_OPCODE_TXQ:
2609 return 0;
2610 break;
2611
2612 case TGSI_OPCODE_CONT:
2613 return 0;
2614 break;
2615
2616 case TGSI_OPCODE_EMIT:
2617 return 0;
2618 break;
2619
2620 case TGSI_OPCODE_ENDPRIM:
2621 return 0;
2622 break;
2623
2624 default:
2625 return 0;
2626 }
2627
2628 return 1;
2629 }
2630
2631 static void
2632 emit_declaration(
2633 struct x86_function *func,
2634 struct tgsi_full_declaration *decl )
2635 {
2636 if( decl->Declaration.File == TGSI_FILE_INPUT ) {
2637 unsigned first, last, mask;
2638 unsigned i, j;
2639
2640 first = decl->Range.First;
2641 last = decl->Range.Last;
2642 mask = decl->Declaration.UsageMask;
2643
2644 for( i = first; i <= last; i++ ) {
2645 for( j = 0; j < NUM_CHANNELS; j++ ) {
2646 if( mask & (1 << j) ) {
2647 switch( decl->Declaration.Interpolate ) {
2648 case TGSI_INTERPOLATE_CONSTANT:
2649 emit_coef_a0( func, 0, i, j );
2650 emit_inputs( func, 0, i, j );
2651 break;
2652
2653 case TGSI_INTERPOLATE_LINEAR:
2654 emit_tempf( func, 0, 0, TGSI_SWIZZLE_X );
2655 emit_coef_dadx( func, 1, i, j );
2656 emit_tempf( func, 2, 0, TGSI_SWIZZLE_Y );
2657 emit_coef_dady( func, 3, i, j );
2658 emit_mul( func, 0, 1 ); /* x * dadx */
2659 emit_coef_a0( func, 4, i, j );
2660 emit_mul( func, 2, 3 ); /* y * dady */
2661 emit_add( func, 0, 4 ); /* x * dadx + a0 */
2662 emit_add( func, 0, 2 ); /* x * dadx + y * dady + a0 */
2663 emit_inputs( func, 0, i, j );
2664 break;
2665
2666 case TGSI_INTERPOLATE_PERSPECTIVE:
2667 emit_tempf( func, 0, 0, TGSI_SWIZZLE_X );
2668 emit_coef_dadx( func, 1, i, j );
2669 emit_tempf( func, 2, 0, TGSI_SWIZZLE_Y );
2670 emit_coef_dady( func, 3, i, j );
2671 emit_mul( func, 0, 1 ); /* x * dadx */
2672 emit_tempf( func, 4, 0, TGSI_SWIZZLE_W );
2673 emit_coef_a0( func, 5, i, j );
2674 emit_rcp( func, 4, 4 ); /* 1.0 / w */
2675 emit_mul( func, 2, 3 ); /* y * dady */
2676 emit_add( func, 0, 5 ); /* x * dadx + a0 */
2677 emit_add( func, 0, 2 ); /* x * dadx + y * dady + a0 */
2678 emit_mul( func, 0, 4 ); /* (x * dadx + y * dady + a0) / w */
2679 emit_inputs( func, 0, i, j );
2680 break;
2681
2682 default:
2683 assert( 0 );
2684 break;
2685 }
2686 }
2687 }
2688 }
2689 }
2690 }
2691
2692 static void aos_to_soa( struct x86_function *func,
2693 uint arg_aos,
2694 uint arg_machine,
2695 uint arg_num,
2696 uint arg_stride )
2697 {
2698 struct x86_reg soa_input = x86_make_reg( file_REG32, reg_AX );
2699 struct x86_reg aos_input = x86_make_reg( file_REG32, reg_BX );
2700 struct x86_reg num_inputs = x86_make_reg( file_REG32, reg_CX );
2701 struct x86_reg stride = x86_make_reg( file_REG32, reg_DX );
2702 int inner_loop;
2703
2704
2705 /* Save EBX */
2706 x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
2707
2708 x86_mov( func, aos_input, x86_fn_arg( func, arg_aos ) );
2709 x86_mov( func, soa_input, x86_fn_arg( func, arg_machine ) );
2710 x86_lea( func, soa_input,
2711 x86_make_disp( soa_input,
2712 Offset(struct tgsi_exec_machine, Inputs) ) );
2713 x86_mov( func, num_inputs, x86_fn_arg( func, arg_num ) );
2714 x86_mov( func, stride, x86_fn_arg( func, arg_stride ) );
2715
2716 /* do */
2717 inner_loop = x86_get_label( func );
2718 {
2719 x86_push( func, aos_input );
2720 sse_movlps( func, make_xmm( 0 ), x86_make_disp( aos_input, 0 ) );
2721 sse_movlps( func, make_xmm( 3 ), x86_make_disp( aos_input, 8 ) );
2722 x86_add( func, aos_input, stride );
2723 sse_movhps( func, make_xmm( 0 ), x86_make_disp( aos_input, 0 ) );
2724 sse_movhps( func, make_xmm( 3 ), x86_make_disp( aos_input, 8 ) );
2725 x86_add( func, aos_input, stride );
2726 sse_movlps( func, make_xmm( 1 ), x86_make_disp( aos_input, 0 ) );
2727 sse_movlps( func, make_xmm( 4 ), x86_make_disp( aos_input, 8 ) );
2728 x86_add( func, aos_input, stride );
2729 sse_movhps( func, make_xmm( 1 ), x86_make_disp( aos_input, 0 ) );
2730 sse_movhps( func, make_xmm( 4 ), x86_make_disp( aos_input, 8 ) );
2731 x86_pop( func, aos_input );
2732
2733 sse_movaps( func, make_xmm( 2 ), make_xmm( 0 ) );
2734 sse_movaps( func, make_xmm( 5 ), make_xmm( 3 ) );
2735 sse_shufps( func, make_xmm( 0 ), make_xmm( 1 ), 0x88 );
2736 sse_shufps( func, make_xmm( 2 ), make_xmm( 1 ), 0xdd );
2737 sse_shufps( func, make_xmm( 3 ), make_xmm( 4 ), 0x88 );
2738 sse_shufps( func, make_xmm( 5 ), make_xmm( 4 ), 0xdd );
2739
2740 sse_movups( func, x86_make_disp( soa_input, 0 ), make_xmm( 0 ) );
2741 sse_movups( func, x86_make_disp( soa_input, 16 ), make_xmm( 2 ) );
2742 sse_movups( func, x86_make_disp( soa_input, 32 ), make_xmm( 3 ) );
2743 sse_movups( func, x86_make_disp( soa_input, 48 ), make_xmm( 5 ) );
2744
2745 /* Advance to next input */
2746 x86_lea( func, aos_input, x86_make_disp(aos_input, 16) );
2747 x86_lea( func, soa_input, x86_make_disp(soa_input, 64) );
2748 }
2749 /* while --num_inputs */
2750 x86_dec( func, num_inputs );
2751 x86_jcc( func, cc_NE, inner_loop );
2752
2753 /* Restore EBX */
2754 x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
2755 }
2756
2757 static void soa_to_aos( struct x86_function *func,
2758 uint arg_aos,
2759 uint arg_machine,
2760 uint arg_num,
2761 uint arg_stride )
2762 {
2763 struct x86_reg soa_output = x86_make_reg( file_REG32, reg_AX );
2764 struct x86_reg aos_output = x86_make_reg( file_REG32, reg_BX );
2765 struct x86_reg num_outputs = x86_make_reg( file_REG32, reg_CX );
2766 struct x86_reg temp = x86_make_reg( file_REG32, reg_DX );
2767 int inner_loop;
2768
2769 /* Save EBX */
2770 x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
2771
2772 x86_mov( func, aos_output, x86_fn_arg( func, arg_aos ) );
2773 x86_mov( func, soa_output, x86_fn_arg( func, arg_machine ) );
2774 x86_lea( func, soa_output,
2775 x86_make_disp( soa_output,
2776 Offset(struct tgsi_exec_machine, Outputs) ) );
2777 x86_mov( func, num_outputs, x86_fn_arg( func, arg_num ) );
2778
2779 /* do */
2780 inner_loop = x86_get_label( func );
2781 {
2782 sse_movups( func, make_xmm( 0 ), x86_make_disp( soa_output, 0 ) );
2783 sse_movups( func, make_xmm( 1 ), x86_make_disp( soa_output, 16 ) );
2784 sse_movups( func, make_xmm( 3 ), x86_make_disp( soa_output, 32 ) );
2785 sse_movups( func, make_xmm( 4 ), x86_make_disp( soa_output, 48 ) );
2786
2787 sse_movaps( func, make_xmm( 2 ), make_xmm( 0 ) );
2788 sse_movaps( func, make_xmm( 5 ), make_xmm( 3 ) );
2789 sse_unpcklps( func, make_xmm( 0 ), make_xmm( 1 ) );
2790 sse_unpckhps( func, make_xmm( 2 ), make_xmm( 1 ) );
2791 sse_unpcklps( func, make_xmm( 3 ), make_xmm( 4 ) );
2792 sse_unpckhps( func, make_xmm( 5 ), make_xmm( 4 ) );
2793
2794 x86_mov( func, temp, x86_fn_arg( func, arg_stride ) );
2795 x86_push( func, aos_output );
2796 sse_movlps( func, x86_make_disp( aos_output, 0 ), make_xmm( 0 ) );
2797 sse_movlps( func, x86_make_disp( aos_output, 8 ), make_xmm( 3 ) );
2798 x86_add( func, aos_output, temp );
2799 sse_movhps( func, x86_make_disp( aos_output, 0 ), make_xmm( 0 ) );
2800 sse_movhps( func, x86_make_disp( aos_output, 8 ), make_xmm( 3 ) );
2801 x86_add( func, aos_output, temp );
2802 sse_movlps( func, x86_make_disp( aos_output, 0 ), make_xmm( 2 ) );
2803 sse_movlps( func, x86_make_disp( aos_output, 8 ), make_xmm( 5 ) );
2804 x86_add( func, aos_output, temp );
2805 sse_movhps( func, x86_make_disp( aos_output, 0 ), make_xmm( 2 ) );
2806 sse_movhps( func, x86_make_disp( aos_output, 8 ), make_xmm( 5 ) );
2807 x86_pop( func, aos_output );
2808
2809 /* Advance to next output */
2810 x86_lea( func, aos_output, x86_make_disp(aos_output, 16) );
2811 x86_lea( func, soa_output, x86_make_disp(soa_output, 64) );
2812 }
2813 /* while --num_outputs */
2814 x86_dec( func, num_outputs );
2815 x86_jcc( func, cc_NE, inner_loop );
2816
2817 /* Restore EBX */
2818 x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
2819 }
2820
2821 /**
2822 * Translate a TGSI vertex/fragment shader to SSE2 code.
2823 * Slightly different things are done for vertex vs. fragment shaders.
2824 *
2825 * \param tokens the TGSI input shader
2826 * \param func the output SSE code/function
2827 * \param immediates buffer to place immediates, later passed to SSE func
2828 * \param return 1 for success, 0 if translation failed
2829 */
2830 unsigned
2831 tgsi_emit_sse2(
2832 const struct tgsi_token *tokens,
2833 struct x86_function *func,
2834 float (*immediates)[4],
2835 boolean do_swizzles )
2836 {
2837 struct tgsi_parse_context parse;
2838 unsigned ok = 1;
2839 uint num_immediates = 0;
2840
2841 util_init_math();
2842
2843 func->csr = func->store;
2844
2845 tgsi_parse_init( &parse, tokens );
2846
2847 /* Can't just use EDI, EBX without save/restoring them:
2848 */
2849 x86_push( func, x86_make_reg( file_REG32, reg_BX ) );
2850 x86_push( func, x86_make_reg( file_REG32, reg_DI ) );
2851
2852 /*
2853 * Different function args for vertex/fragment shaders:
2854 */
2855 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX) {
2856 if (do_swizzles)
2857 aos_to_soa( func,
2858 4, /* aos_input */
2859 1, /* machine */
2860 5, /* num_inputs */
2861 6 ); /* input_stride */
2862 }
2863
2864 x86_mov(
2865 func,
2866 get_machine_base(),
2867 x86_fn_arg( func, 1 ) );
2868 x86_mov(
2869 func,
2870 get_const_base(),
2871 x86_fn_arg( func, 2 ) );
2872 x86_mov(
2873 func,
2874 get_immediate_base(),
2875 x86_fn_arg( func, 3 ) );
2876
2877 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_FRAGMENT) {
2878 x86_mov(
2879 func,
2880 get_coef_base(),
2881 x86_fn_arg( func, 4 ) );
2882 }
2883
2884 x86_mov(
2885 func,
2886 get_sampler_base(),
2887 x86_make_disp( get_machine_base(),
2888 Offset( struct tgsi_exec_machine, Samplers ) ) );
2889
2890
2891 while( !tgsi_parse_end_of_tokens( &parse ) && ok ) {
2892 tgsi_parse_token( &parse );
2893
2894 switch( parse.FullToken.Token.Type ) {
2895 case TGSI_TOKEN_TYPE_DECLARATION:
2896 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_FRAGMENT) {
2897 emit_declaration(
2898 func,
2899 &parse.FullToken.FullDeclaration );
2900 }
2901 break;
2902
2903 case TGSI_TOKEN_TYPE_INSTRUCTION:
2904 ok = emit_instruction(
2905 func,
2906 &parse.FullToken.FullInstruction );
2907
2908 if (!ok) {
2909 uint opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
2910 debug_printf("failed to translate tgsi opcode %d (%s) to SSE (%s)\n",
2911 opcode,
2912 tgsi_get_opcode_name(opcode),
2913 parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX ?
2914 "vertex shader" : "fragment shader");
2915 }
2916
2917 if (tgsi_check_soa_dependencies(&parse.FullToken.FullInstruction)) {
2918 uint opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
2919
2920 /* XXX: we only handle src/dst aliasing in a few opcodes
2921 * currently. Need to use an additional temporay to hold
2922 * the result in the cases where the code is too opaque to
2923 * fix.
2924 */
2925 if (opcode != TGSI_OPCODE_MOV) {
2926 debug_printf("Warning: src/dst aliasing in instruction"
2927 " is not handled:\n");
2928 tgsi_dump_instruction(&parse.FullToken.FullInstruction, 1);
2929 }
2930 }
2931 break;
2932
2933 case TGSI_TOKEN_TYPE_IMMEDIATE:
2934 /* simply copy the immediate values into the next immediates[] slot */
2935 {
2936 const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
2937 uint i;
2938 assert(size <= 4);
2939 assert(num_immediates < TGSI_EXEC_NUM_IMMEDIATES);
2940 for( i = 0; i < size; i++ ) {
2941 immediates[num_immediates][i] =
2942 parse.FullToken.FullImmediate.u[i].Float;
2943 }
2944 #if 0
2945 debug_printf("SSE FS immediate[%d] = %f %f %f %f\n",
2946 num_immediates,
2947 immediates[num_immediates][0],
2948 immediates[num_immediates][1],
2949 immediates[num_immediates][2],
2950 immediates[num_immediates][3]);
2951 #endif
2952 num_immediates++;
2953 }
2954 break;
2955
2956 default:
2957 ok = 0;
2958 assert( 0 );
2959 }
2960 }
2961
2962 if (parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX) {
2963 if (do_swizzles)
2964 soa_to_aos( func,
2965 7, /* aos_output */
2966 1, /* machine */
2967 8, /* num_outputs */
2968 9 ); /* output_stride */
2969 }
2970
2971 /* Can't just use EBX, EDI without save/restoring them:
2972 */
2973 x86_pop( func, x86_make_reg( file_REG32, reg_DI ) );
2974 x86_pop( func, x86_make_reg( file_REG32, reg_BX ) );
2975
2976 emit_ret( func );
2977
2978 tgsi_parse_free( &parse );
2979
2980 return ok;
2981 }
2982
2983 #endif /* PIPE_ARCH_X86 */
2984