projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallivm: Pass texture coords derivates as scalars.
[mesa.git] / src / gallium / auxiliary / gallivm / lp_bld_tgsi_soa.c
1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28
29 /**
30 * @file
31 * TGSI to LLVM IR translation -- SoA.
32 *
33 * @author Jose Fonseca <jfonseca@vmware.com>
34 *
35 * Based on tgsi_sse2.c code written by Michal Krol, Keith Whitwell,
36 * Brian Paul, and others.
37 */
38
39 #include "pipe/p_config.h"
40 #include "pipe/p_shader_tokens.h"
41 #include "util/u_debug.h"
42 #include "util/u_math.h"
43 #include "util/u_memory.h"
44 #include "tgsi/tgsi_dump.h"
45 #include "tgsi/tgsi_info.h"
46 #include "tgsi/tgsi_parse.h"
47 #include "tgsi/tgsi_util.h"
48 #include "tgsi/tgsi_scan.h"
49 #include "lp_bld_type.h"
50 #include "lp_bld_const.h"
51 #include "lp_bld_arit.h"
52 #include "lp_bld_bitarit.h"
53 #include "lp_bld_gather.h"
54 #include "lp_bld_logic.h"
55 #include "lp_bld_swizzle.h"
56 #include "lp_bld_flow.h"
57 #include "lp_bld_quad.h"
58 #include "lp_bld_tgsi.h"
59 #include "lp_bld_limits.h"
60 #include "lp_bld_debug.h"
61
62
63 #define FOR_EACH_CHANNEL( CHAN )\
64 for (CHAN = 0; CHAN < NUM_CHANNELS; CHAN++)
65
66 #define IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
67 ((INST)->Dst[0].Register.WriteMask & (1 << (CHAN)))
68
69 #define IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
70 if (IS_DST0_CHANNEL_ENABLED( INST, CHAN ))
71
72 #define FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\
73 FOR_EACH_CHANNEL( CHAN )\
74 IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )
75
76 #define CHAN_X 0
77 #define CHAN_Y 1
78 #define CHAN_Z 2
79 #define CHAN_W 3
80 #define NUM_CHANNELS 4
81
82 #define LP_MAX_INSTRUCTIONS 256
83
84
85 struct lp_exec_mask {
86 struct lp_build_context *bld;
87
88 boolean has_mask;
89
90 LLVMTypeRef int_vec_type;
91
92 LLVMValueRef cond_stack[LP_MAX_TGSI_NESTING];
93 int cond_stack_size;
94 LLVMValueRef cond_mask;
95
96 LLVMBasicBlockRef loop_block;
97 LLVMValueRef cont_mask;
98 LLVMValueRef break_mask;
99 LLVMValueRef break_var;
100 struct {
101 LLVMBasicBlockRef loop_block;
102 LLVMValueRef cont_mask;
103 LLVMValueRef break_mask;
104 LLVMValueRef break_var;
105 } loop_stack[LP_MAX_TGSI_NESTING];
106 int loop_stack_size;
107
108 LLVMValueRef ret_mask;
109 struct {
110 int pc;
111 LLVMValueRef ret_mask;
112 } call_stack[LP_MAX_TGSI_NESTING];
113 int call_stack_size;
114
115 LLVMValueRef exec_mask;
116 };
117
118 struct lp_build_tgsi_soa_context
119 {
120 struct lp_build_context base;
121
122 /* Builder for integer masks and indices */
123 struct lp_build_context uint_bld;
124
125 LLVMValueRef consts_ptr;
126 const LLVMValueRef *pos;
127 const LLVMValueRef (*inputs)[NUM_CHANNELS];
128 LLVMValueRef (*outputs)[NUM_CHANNELS];
129
130 const struct lp_build_sampler_soa *sampler;
131
132 LLVMValueRef immediates[LP_MAX_TGSI_IMMEDIATES][NUM_CHANNELS];
133 LLVMValueRef temps[LP_MAX_TGSI_TEMPS][NUM_CHANNELS];
134 LLVMValueRef addr[LP_MAX_TGSI_ADDRS][NUM_CHANNELS];
135 LLVMValueRef preds[LP_MAX_TGSI_PREDS][NUM_CHANNELS];
136
137 /* We allocate/use this array of temps if (1 << TGSI_FILE_TEMPORARY) is
138 * set in the indirect_files field.
139 * The temps[] array above is unused then.
140 */
141 LLVMValueRef temps_array;
142
143 const struct tgsi_shader_info *info;
144 /** bitmask indicating which register files are accessed indirectly */
145 unsigned indirect_files;
146
147 struct lp_build_mask_context *mask;
148 struct lp_exec_mask exec_mask;
149
150 struct tgsi_full_instruction *instructions;
151 uint max_instructions;
152 };
153
154 static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
155 {
156 mask->bld = bld;
157 mask->has_mask = FALSE;
158 mask->cond_stack_size = 0;
159 mask->loop_stack_size = 0;
160 mask->call_stack_size = 0;
161
162 mask->int_vec_type = lp_build_int_vec_type(mask->bld->type);
163 mask->exec_mask = mask->ret_mask = mask->break_mask = mask->cont_mask = mask->cond_mask =
164 LLVMConstAllOnes(mask->int_vec_type);
165 }
166
167 static void lp_exec_mask_update(struct lp_exec_mask *mask)
168 {
169 if (mask->loop_stack_size) {
170 /*for loops we need to update the entire mask at runtime */
171 LLVMValueRef tmp;
172 assert(mask->break_mask);
173 tmp = LLVMBuildAnd(mask->bld->builder,
174 mask->cont_mask,
175 mask->break_mask,
176 "maskcb");
177 mask->exec_mask = LLVMBuildAnd(mask->bld->builder,
178 mask->cond_mask,
179 tmp,
180 "maskfull");
181 } else
182 mask->exec_mask = mask->cond_mask;
183
184 if (mask->call_stack_size) {
185 mask->exec_mask = LLVMBuildAnd(mask->bld->builder,
186 mask->exec_mask,
187 mask->ret_mask,
188 "callmask");
189 }
190
191 mask->has_mask = (mask->cond_stack_size > 0 ||
192 mask->loop_stack_size > 0 ||
193 mask->call_stack_size > 0);
194 }
195
196 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
197 LLVMValueRef val)
198 {
199 assert(mask->cond_stack_size < LP_MAX_TGSI_NESTING);
200 if (mask->cond_stack_size == 0) {
201 assert(mask->cond_mask == LLVMConstAllOnes(mask->int_vec_type));
202 }
203 mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
204 assert(LLVMTypeOf(val) == mask->int_vec_type);
205 mask->cond_mask = LLVMBuildAnd(mask->bld->builder,
206 mask->cond_mask,
207 val,
208 "");
209 lp_exec_mask_update(mask);
210 }
211
212 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
213 {
214 LLVMValueRef prev_mask;
215 LLVMValueRef inv_mask;
216
217 assert(mask->cond_stack_size);
218 prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
219 if (mask->cond_stack_size == 1) {
220 assert(prev_mask == LLVMConstAllOnes(mask->int_vec_type));
221 }
222
223 inv_mask = LLVMBuildNot(mask->bld->builder, mask->cond_mask, "");
224
225 mask->cond_mask = LLVMBuildAnd(mask->bld->builder,
226 inv_mask,
227 prev_mask, "");
228 lp_exec_mask_update(mask);
229 }
230
231 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
232 {
233 assert(mask->cond_stack_size);
234 mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
235 lp_exec_mask_update(mask);
236 }
237
238 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
239 {
240 if (mask->loop_stack_size == 0) {
241 assert(mask->loop_block == NULL);
242 assert(mask->cont_mask == LLVMConstAllOnes(mask->int_vec_type));
243 assert(mask->break_mask == LLVMConstAllOnes(mask->int_vec_type));
244 assert(mask->break_var == NULL);
245 }
246
247 assert(mask->loop_stack_size < LP_MAX_TGSI_NESTING);
248
249 mask->loop_stack[mask->loop_stack_size].loop_block = mask->loop_block;
250 mask->loop_stack[mask->loop_stack_size].cont_mask = mask->cont_mask;
251 mask->loop_stack[mask->loop_stack_size].break_mask = mask->break_mask;
252 mask->loop_stack[mask->loop_stack_size].break_var = mask->break_var;
253 ++mask->loop_stack_size;
254
255 mask->break_var = lp_build_alloca(mask->bld->builder, mask->int_vec_type, "");
256 LLVMBuildStore(mask->bld->builder, mask->break_mask, mask->break_var);
257
258 mask->loop_block = lp_build_insert_new_block(mask->bld->builder, "bgnloop");
259 LLVMBuildBr(mask->bld->builder, mask->loop_block);
260 LLVMPositionBuilderAtEnd(mask->bld->builder, mask->loop_block);
261
262 mask->break_mask = LLVMBuildLoad(mask->bld->builder, mask->break_var, "");
263
264 lp_exec_mask_update(mask);
265 }
266
267 static void lp_exec_break(struct lp_exec_mask *mask)
268 {
269 LLVMValueRef exec_mask = LLVMBuildNot(mask->bld->builder,
270 mask->exec_mask,
271 "break");
272
273 mask->break_mask = LLVMBuildAnd(mask->bld->builder,
274 mask->break_mask,
275 exec_mask, "break_full");
276
277 lp_exec_mask_update(mask);
278 }
279
280 static void lp_exec_continue(struct lp_exec_mask *mask)
281 {
282 LLVMValueRef exec_mask = LLVMBuildNot(mask->bld->builder,
283 mask->exec_mask,
284 "");
285
286 mask->cont_mask = LLVMBuildAnd(mask->bld->builder,
287 mask->cont_mask,
288 exec_mask, "");
289
290 lp_exec_mask_update(mask);
291 }
292
293
294 static void lp_exec_endloop(struct lp_exec_mask *mask)
295 {
296 LLVMBasicBlockRef endloop;
297 LLVMTypeRef reg_type = LLVMIntType(mask->bld->type.width*
298 mask->bld->type.length);
299 LLVMValueRef i1cond;
300
301 assert(mask->break_mask);
302
303 /*
304 * Restore the cont_mask, but don't pop
305 */
306 assert(mask->loop_stack_size);
307 mask->cont_mask = mask->loop_stack[mask->loop_stack_size - 1].cont_mask;
308 lp_exec_mask_update(mask);
309
310 /*
311 * Unlike the continue mask, the break_mask must be preserved across loop
312 * iterations
313 */
314 LLVMBuildStore(mask->bld->builder, mask->break_mask, mask->break_var);
315
316 /* i1cond = (mask == 0) */
317 i1cond = LLVMBuildICmp(
318 mask->bld->builder,
319 LLVMIntNE,
320 LLVMBuildBitCast(mask->bld->builder, mask->exec_mask, reg_type, ""),
321 LLVMConstNull(reg_type), "");
322
323 endloop = lp_build_insert_new_block(mask->bld->builder, "endloop");
324
325 LLVMBuildCondBr(mask->bld->builder,
326 i1cond, mask->loop_block, endloop);
327
328 LLVMPositionBuilderAtEnd(mask->bld->builder, endloop);
329
330 assert(mask->loop_stack_size);
331 --mask->loop_stack_size;
332 mask->loop_block = mask->loop_stack[mask->loop_stack_size].loop_block;
333 mask->cont_mask = mask->loop_stack[mask->loop_stack_size].cont_mask;
334 mask->break_mask = mask->loop_stack[mask->loop_stack_size].break_mask;
335 mask->break_var = mask->loop_stack[mask->loop_stack_size].break_var;
336
337 lp_exec_mask_update(mask);
338 }
339
340 /* stores val into an address pointed to by dst.
341 * mask->exec_mask is used to figure out which bits of val
342 * should be stored into the address
343 * (0 means don't store this bit, 1 means do store).
344 */
345 static void lp_exec_mask_store(struct lp_exec_mask *mask,
346 LLVMValueRef pred,
347 LLVMValueRef val,
348 LLVMValueRef dst)
349 {
350 /* Mix the predicate and execution mask */
351 if (mask->has_mask) {
352 if (pred) {
353 pred = LLVMBuildAnd(mask->bld->builder, pred, mask->exec_mask, "");
354 } else {
355 pred = mask->exec_mask;
356 }
357 }
358
359 if (pred) {
360 LLVMValueRef real_val, dst_val;
361
362 dst_val = LLVMBuildLoad(mask->bld->builder, dst, "");
363 real_val = lp_build_select(mask->bld,
364 pred,
365 val, dst_val);
366
367 LLVMBuildStore(mask->bld->builder, real_val, dst);
368 } else
369 LLVMBuildStore(mask->bld->builder, val, dst);
370 }
371
372 static void lp_exec_mask_call(struct lp_exec_mask *mask,
373 int func,
374 int *pc)
375 {
376 assert(mask->call_stack_size < LP_MAX_TGSI_NESTING);
377 mask->call_stack[mask->call_stack_size].pc = *pc;
378 mask->call_stack[mask->call_stack_size].ret_mask = mask->ret_mask;
379 mask->call_stack_size++;
380 *pc = func;
381 }
382
383 static void lp_exec_mask_ret(struct lp_exec_mask *mask, int *pc)
384 {
385 LLVMValueRef exec_mask;
386
387 if (mask->call_stack_size == 0) {
388 /* returning from main() */
389 *pc = -1;
390 return;
391 }
392 exec_mask = LLVMBuildNot(mask->bld->builder,
393 mask->exec_mask,
394 "ret");
395
396 mask->ret_mask = LLVMBuildAnd(mask->bld->builder,
397 mask->ret_mask,
398 exec_mask, "ret_full");
399
400 lp_exec_mask_update(mask);
401 }
402
403 static void lp_exec_mask_bgnsub(struct lp_exec_mask *mask)
404 {
405 }
406
407 static void lp_exec_mask_endsub(struct lp_exec_mask *mask, int *pc)
408 {
409 assert(mask->call_stack_size);
410 mask->call_stack_size--;
411 *pc = mask->call_stack[mask->call_stack_size].pc;
412 mask->ret_mask = mask->call_stack[mask->call_stack_size].ret_mask;
413 lp_exec_mask_update(mask);
414 }
415
416
417 /**
418 * Return pointer to a temporary register channel (src or dest).
419 * Note that indirect addressing cannot be handled here.
420 * \param index which temporary register
421 * \param chan which channel of the temp register.
422 */
423 static LLVMValueRef
424 get_temp_ptr(struct lp_build_tgsi_soa_context *bld,
425 unsigned index,
426 unsigned chan)
427 {
428 assert(chan < 4);
429 if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
430 LLVMValueRef lindex = lp_build_const_int32(index * 4 + chan);
431 return LLVMBuildGEP(bld->base.builder, bld->temps_array, &lindex, 1, "");
432 }
433 else {
434 return bld->temps[index][chan];
435 }
436 }
437
438
439 /**
440 * Gather vector.
441 * XXX the lp_build_gather() function should be capable of doing this
442 * with a little work.
443 */
444 static LLVMValueRef
445 build_gather(struct lp_build_tgsi_soa_context *bld,
446 LLVMValueRef base_ptr,
447 LLVMValueRef indexes)
448 {
449 LLVMValueRef res = bld->base.undef;
450 unsigned i;
451
452 /*
453 * Loop over elements of index_vec, load scalar value, insert it into 'res'.
454 */
455 for (i = 0; i < bld->base.type.length; i++) {
456 LLVMValueRef ii = LLVMConstInt(LLVMInt32Type(), i, 0);
457 LLVMValueRef index = LLVMBuildExtractElement(bld->base.builder,
458 indexes, ii, "");
459 LLVMValueRef scalar_ptr = LLVMBuildGEP(bld->base.builder, base_ptr,
460 &index, 1, "");
461 LLVMValueRef scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
462
463 res = LLVMBuildInsertElement(bld->base.builder, res, scalar, ii, "");
464 }
465
466 return res;
467 }
468
469
470 /**
471 * Read the current value of the ADDR register, convert the floats to
472 * ints, multiply by four and return the vector of offsets.
473 * The offsets will be used to index into the constant buffer or
474 * temporary register file.
475 */
476 static LLVMValueRef
477 get_indirect_index(struct lp_build_tgsi_soa_context *bld,
478 unsigned reg_file, unsigned reg_index,
479 const struct tgsi_src_register *indirect_reg)
480 {
481 struct lp_build_context *uint_bld = &bld->uint_bld;
482 /* always use X component of address register */
483 unsigned swizzle = indirect_reg->SwizzleX;
484 LLVMValueRef base;
485 LLVMValueRef rel;
486 LLVMValueRef max_index;
487 LLVMValueRef index;
488
489 assert(bld->indirect_files & (1 << reg_file));
490
491 base = lp_build_const_int_vec(uint_bld->type, reg_index);
492
493 assert(swizzle < 4);
494 rel = LLVMBuildLoad(bld->base.builder,
495 bld->addr[indirect_reg->Index][swizzle],
496 "load addr reg");
497
498 /* for indexing we want integers */
499 rel = LLVMBuildFPToSI(bld->base.builder,
500 rel,
501 uint_bld->vec_type, "");
502
503 index = lp_build_add(uint_bld, base, rel);
504
505 max_index = lp_build_const_int_vec(uint_bld->type,
506 bld->info->file_max[reg_file]);
507
508 assert(!uint_bld->type.sign);
509 index = lp_build_min(uint_bld, index, max_index);
510
511 return index;
512 }
513
514
515 /**
516 * Register fetch.
517 */
518 static LLVMValueRef
519 emit_fetch(
520 struct lp_build_tgsi_soa_context *bld,
521 const struct tgsi_full_instruction *inst,
522 unsigned src_op,
523 const unsigned chan_index )
524 {
525 struct lp_build_context *uint_bld = &bld->uint_bld;
526 const struct tgsi_full_src_register *reg = &inst->Src[src_op];
527 const unsigned swizzle =
528 tgsi_util_get_full_src_register_swizzle(reg, chan_index);
529 LLVMValueRef res;
530 LLVMValueRef indirect_index = NULL;
531
532 if (swizzle > 3) {
533 assert(0 && "invalid swizzle in emit_fetch()");
534 return bld->base.undef;
535 }
536
537 if (reg->Register.Indirect) {
538 indirect_index = get_indirect_index(bld,
539 reg->Register.File,
540 reg->Register.Index,
541 &reg->Indirect);
542 } else {
543 assert(reg->Register.Index <= bld->info->file_max[reg->Register.File]);
544 }
545
546 switch (reg->Register.File) {
547 case TGSI_FILE_CONSTANT:
548 if (reg->Register.Indirect) {
549 LLVMValueRef swizzle_vec =
550 lp_build_const_int_vec(uint_bld->type, swizzle);
551 LLVMValueRef index_vec; /* index into the const buffer */
552
553 /* index_vec = indirect_index * 4 + swizzle */
554 index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
555 index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
556
557 /* Gather values from the constant buffer */
558 res = build_gather(bld, bld->consts_ptr, index_vec);
559 }
560 else {
561 LLVMValueRef index; /* index into the const buffer */
562 LLVMValueRef scalar, scalar_ptr;
563
564 index = lp_build_const_int32(reg->Register.Index*4 + swizzle);
565
566 scalar_ptr = LLVMBuildGEP(bld->base.builder, bld->consts_ptr,
567 &index, 1, "");
568 scalar = LLVMBuildLoad(bld->base.builder, scalar_ptr, "");
569
570 res = lp_build_broadcast_scalar(&bld->base, scalar);
571 }
572 break;
573
574 case TGSI_FILE_IMMEDIATE:
575 res = bld->immediates[reg->Register.Index][swizzle];
576 assert(res);
577 break;
578
579 case TGSI_FILE_INPUT:
580 res = bld->inputs[reg->Register.Index][swizzle];
581 assert(res);
582 break;
583
584 case TGSI_FILE_TEMPORARY:
585 if (reg->Register.Indirect) {
586 LLVMValueRef swizzle_vec =
587 lp_build_const_int_vec(uint_bld->type, swizzle);
588 LLVMValueRef length_vec =
589 lp_build_const_int_vec(uint_bld->type, bld->base.type.length);
590 LLVMValueRef index_vec; /* index into the const buffer */
591 LLVMValueRef temps_array;
592 LLVMTypeRef float4_ptr_type;
593
594 /* index_vec = (indirect_index * 4 + swizzle) * length */
595 index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
596 index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
597 index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
598
599 /* cast temps_array pointer to float* */
600 float4_ptr_type = LLVMPointerType(LLVMFloatType(), 0);
601 temps_array = LLVMBuildBitCast(uint_bld->builder, bld->temps_array,
602 float4_ptr_type, "");
603
604 /* Gather values from the temporary register array */
605 res = build_gather(bld, temps_array, index_vec);
606 }
607 else {
608 LLVMValueRef temp_ptr;
609 temp_ptr = get_temp_ptr(bld, reg->Register.Index, swizzle);
610 res = LLVMBuildLoad(bld->base.builder, temp_ptr, "");
611 if (!res)
612 return bld->base.undef;
613 }
614 break;
615
616 default:
617 assert(0 && "invalid src register in emit_fetch()");
618 return bld->base.undef;
619 }
620
621 switch( tgsi_util_get_full_src_register_sign_mode( reg, chan_index ) ) {
622 case TGSI_UTIL_SIGN_CLEAR:
623 res = lp_build_abs( &bld->base, res );
624 break;
625
626 case TGSI_UTIL_SIGN_SET:
627 res = lp_build_abs( &bld->base, res );
628 /* fall through */
629 case TGSI_UTIL_SIGN_TOGGLE:
630 res = lp_build_negate( &bld->base, res );
631 break;
632
633 case TGSI_UTIL_SIGN_KEEP:
634 break;
635 }
636
637 return res;
638 }
639
640
641 /**
642 * Register fetch with derivatives.
643 */
644 static void
645 emit_fetch_deriv(
646 struct lp_build_tgsi_soa_context *bld,
647 const struct tgsi_full_instruction *inst,
648 unsigned index,
649 const unsigned chan_index,
650 LLVMValueRef *res,
651 LLVMValueRef *ddx,
652 LLVMValueRef *ddy)
653 {
654 LLVMValueRef src;
655
656 src = emit_fetch(bld, inst, index, chan_index);
657
658 if(res)
659 *res = src;
660
661 /* TODO: use interpolation coeffs for inputs */
662
663 if(ddx)
664 *ddx = lp_build_ddx(&bld->base, src);
665
666 if(ddy)
667 *ddy = lp_build_ddy(&bld->base, src);
668 }
669
670
671 /**
672 * Predicate.
673 */
674 static void
675 emit_fetch_predicate(
676 struct lp_build_tgsi_soa_context *bld,
677 const struct tgsi_full_instruction *inst,
678 LLVMValueRef *pred)
679 {
680 unsigned index;
681 unsigned char swizzles[4];
682 LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
683 LLVMValueRef value;
684 unsigned chan;
685
686 if (!inst->Instruction.Predicate) {
687 FOR_EACH_CHANNEL( chan ) {
688 pred[chan] = NULL;
689 }
690 return;
691 }
692
693 swizzles[0] = inst->Predicate.SwizzleX;
694 swizzles[1] = inst->Predicate.SwizzleY;
695 swizzles[2] = inst->Predicate.SwizzleZ;
696 swizzles[3] = inst->Predicate.SwizzleW;
697
698 index = inst->Predicate.Index;
699 assert(index < LP_MAX_TGSI_PREDS);
700
701 FOR_EACH_CHANNEL( chan ) {
702 unsigned swizzle = swizzles[chan];
703
704 /*
705 * Only fetch the predicate register channels that are actually listed
706 * in the swizzles
707 */
708 if (!unswizzled[swizzle]) {
709 value = LLVMBuildLoad(bld->base.builder,
710 bld->preds[index][swizzle], "");
711
712 /*
713 * Convert the value to an integer mask.
714 *
715 * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
716 * is needlessly causing two comparisons due to storing the intermediate
717 * result as float vector instead of an integer mask vector.
718 */
719 value = lp_build_compare(bld->base.builder,
720 bld->base.type,
721 PIPE_FUNC_NOTEQUAL,
722 value,
723 bld->base.zero);
724 if (inst->Predicate.Negate) {
725 value = LLVMBuildNot(bld->base.builder, value, "");
726 }
727
728 unswizzled[swizzle] = value;
729 } else {
730 value = unswizzled[swizzle];
731 }
732
733 pred[chan] = value;
734 }
735 }
736
737
738 /**
739 * Register store.
740 */
741 static void
742 emit_store(
743 struct lp_build_tgsi_soa_context *bld,
744 const struct tgsi_full_instruction *inst,
745 unsigned index,
746 unsigned chan_index,
747 LLVMValueRef pred,
748 LLVMValueRef value)
749 {
750 const struct tgsi_full_dst_register *reg = &inst->Dst[index];
751 LLVMValueRef indirect_index = NULL;
752
753 switch( inst->Instruction.Saturate ) {
754 case TGSI_SAT_NONE:
755 break;
756
757 case TGSI_SAT_ZERO_ONE:
758 value = lp_build_max(&bld->base, value, bld->base.zero);
759 value = lp_build_min(&bld->base, value, bld->base.one);
760 break;
761
762 case TGSI_SAT_MINUS_PLUS_ONE:
763 value = lp_build_max(&bld->base, value, lp_build_const_vec(bld->base.type, -1.0));
764 value = lp_build_min(&bld->base, value, bld->base.one);
765 break;
766
767 default:
768 assert(0);
769 }
770
771 if (reg->Register.Indirect) {
772 indirect_index = get_indirect_index(bld,
773 reg->Register.File,
774 reg->Register.Index,
775 &reg->Indirect);
776 } else {
777 assert(reg->Register.Index <= bld->info->file_max[reg->Register.File]);
778 }
779
780 switch( reg->Register.File ) {
781 case TGSI_FILE_OUTPUT:
782 lp_exec_mask_store(&bld->exec_mask, pred, value,
783 bld->outputs[reg->Register.Index][chan_index]);
784 break;
785
786 case TGSI_FILE_TEMPORARY:
787 if (reg->Register.Indirect) {
788 /* XXX not done yet */
789 debug_printf("WARNING: LLVM scatter store of temp regs"
790 " not implemented\n");
791 }
792 else {
793 LLVMValueRef temp_ptr = get_temp_ptr(bld, reg->Register.Index,
794 chan_index);
795 lp_exec_mask_store(&bld->exec_mask, pred, value, temp_ptr);
796 }
797 break;
798
799 case TGSI_FILE_ADDRESS:
800 lp_exec_mask_store(&bld->exec_mask, pred, value,
801 bld->addr[reg->Indirect.Index][chan_index]);
802 break;
803
804 case TGSI_FILE_PREDICATE:
805 lp_exec_mask_store(&bld->exec_mask, pred, value,
806 bld->preds[reg->Register.Index][chan_index]);
807 break;
808
809 default:
810 assert( 0 );
811 }
812 }
813
814
815 /**
816 * High-level instruction translators.
817 */
818
819 static void
820 emit_tex( struct lp_build_tgsi_soa_context *bld,
821 const struct tgsi_full_instruction *inst,
822 enum lp_build_tex_modifier modifier,
823 LLVMValueRef *texel)
824 {
825 unsigned unit;
826 LLVMValueRef lod_bias, explicit_lod;
827 LLVMValueRef oow = NULL;
828 LLVMValueRef coords[3];
829 LLVMValueRef ddx[3];
830 LLVMValueRef ddy[3];
831 unsigned num_coords;
832 unsigned i;
833
834 if (!bld->sampler) {
835 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
836 for (i = 0; i < 4; i++) {
837 texel[i] = bld->base.undef;
838 }
839 return;
840 }
841
842 switch (inst->Texture.Texture) {
843 case TGSI_TEXTURE_1D:
844 num_coords = 1;
845 break;
846 case TGSI_TEXTURE_2D:
847 case TGSI_TEXTURE_RECT:
848 num_coords = 2;
849 break;
850 case TGSI_TEXTURE_SHADOW1D:
851 case TGSI_TEXTURE_SHADOW2D:
852 case TGSI_TEXTURE_SHADOWRECT:
853 case TGSI_TEXTURE_3D:
854 case TGSI_TEXTURE_CUBE:
855 num_coords = 3;
856 break;
857 default:
858 assert(0);
859 return;
860 }
861
862 if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
863 lod_bias = emit_fetch( bld, inst, 0, 3 );
864 explicit_lod = NULL;
865 }
866 else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
867 lod_bias = NULL;
868 explicit_lod = emit_fetch( bld, inst, 0, 3 );
869 }
870 else {
871 lod_bias = NULL;
872 explicit_lod = NULL;
873 }
874
875 if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
876 oow = emit_fetch( bld, inst, 0, 3 );
877 oow = lp_build_rcp(&bld->base, oow);
878 }
879
880 for (i = 0; i < num_coords; i++) {
881 coords[i] = emit_fetch( bld, inst, 0, i );
882 if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
883 coords[i] = lp_build_mul(&bld->base, coords[i], oow);
884 }
885 for (i = num_coords; i < 3; i++) {
886 coords[i] = bld->base.undef;
887 }
888
889 if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
890 LLVMTypeRef i32t = LLVMInt32Type();
891 LLVMValueRef index0 = LLVMConstInt(i32t, 0, 0);
892 for (i = 0; i < num_coords; i++) {
893 LLVMValueRef src1 = emit_fetch( bld, inst, 1, i );
894 LLVMValueRef src2 = emit_fetch( bld, inst, 2, i );
895 ddx[i] = LLVMBuildExtractElement(bld->base.builder, src1, index0, "");
896 ddy[i] = LLVMBuildExtractElement(bld->base.builder, src2, index0, "");
897 }
898 unit = inst->Src[3].Register.Index;
899 } else {
900 for (i = 0; i < num_coords; i++) {
901 ddx[i] = lp_build_scalar_ddx( &bld->base, coords[i] );
902 ddy[i] = lp_build_scalar_ddy( &bld->base, coords[i] );
903 }
904 unit = inst->Src[1].Register.Index;
905 }
906 for (i = num_coords; i < 3; i++) {
907 ddx[i] = LLVMGetUndef(bld->base.elem_type);
908 ddy[i] = LLVMGetUndef(bld->base.elem_type);
909 }
910
911 bld->sampler->emit_fetch_texel(bld->sampler,
912 bld->base.builder,
913 bld->base.type,
914 unit, num_coords, coords,
915 ddx, ddy,
916 lod_bias, explicit_lod,
917 texel);
918 }
919
920
921 /**
922 * Kill fragment if any of the src register values are negative.
923 */
924 static void
925 emit_kil(
926 struct lp_build_tgsi_soa_context *bld,
927 const struct tgsi_full_instruction *inst )
928 {
929 const struct tgsi_full_src_register *reg = &inst->Src[0];
930 LLVMValueRef terms[NUM_CHANNELS];
931 LLVMValueRef mask;
932 unsigned chan_index;
933
934 memset(&terms, 0, sizeof terms);
935
936 FOR_EACH_CHANNEL( chan_index ) {
937 unsigned swizzle;
938
939 /* Unswizzle channel */
940 swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
941
942 /* Check if the component has not been already tested. */
943 assert(swizzle < NUM_CHANNELS);
944 if( !terms[swizzle] )
945 /* TODO: change the comparison operator instead of setting the sign */
946 terms[swizzle] = emit_fetch(bld, inst, 0, chan_index );
947 }
948
949 mask = NULL;
950 FOR_EACH_CHANNEL( chan_index ) {
951 if(terms[chan_index]) {
952 LLVMValueRef chan_mask;
953
954 /*
955 * If term < 0 then mask = 0 else mask = ~0.
956 */
957 chan_mask = lp_build_cmp(&bld->base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->base.zero);
958
959 if(mask)
960 mask = LLVMBuildAnd(bld->base.builder, mask, chan_mask, "");
961 else
962 mask = chan_mask;
963 }
964 }
965
966 if(mask) {
967 lp_build_mask_update(bld->mask, mask);
968
969 /* XXX: figure out if we are at the end of the shader and skip this:
970 */
971 lp_build_mask_check(bld->mask);
972 }
973 }
974
975
976 /**
977 * Predicated fragment kill.
978 * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
979 * The only predication is the execution mask which will apply if
980 * we're inside a loop or conditional.
981 */
982 static void
983 emit_kilp(struct lp_build_tgsi_soa_context *bld,
984 const struct tgsi_full_instruction *inst)
985 {
986 LLVMValueRef mask;
987
988 /* For those channels which are "alive", disable fragment shader
989 * execution.
990 */
991 if (bld->exec_mask.has_mask) {
992 mask = LLVMBuildNot(bld->base.builder, bld->exec_mask.exec_mask, "kilp");
993 }
994 else {
995 mask = bld->base.zero;
996 }
997
998 lp_build_mask_update(bld->mask, mask);
999
1000 /* XXX: figure out if we are at the end of the shader and skip this:
1001 */
1002 lp_build_mask_check(bld->mask);
1003 }
1004
1005 static void
1006 emit_declaration(
1007 struct lp_build_tgsi_soa_context *bld,
1008 const struct tgsi_full_declaration *decl)
1009 {
1010 LLVMTypeRef vec_type = bld->base.vec_type;
1011
1012 unsigned first = decl->Range.First;
1013 unsigned last = decl->Range.Last;
1014 unsigned idx, i;
1015
1016 for (idx = first; idx <= last; ++idx) {
1017 assert(last <= bld->info->file_max[decl->Declaration.File]);
1018 switch (decl->Declaration.File) {
1019 case TGSI_FILE_TEMPORARY:
1020 assert(idx < LP_MAX_TGSI_TEMPS);
1021 if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
1022 LLVMValueRef array_size = LLVMConstInt(LLVMInt32Type(),
1023 last*4 + 4, 0);
1024 bld->temps_array = lp_build_array_alloca(bld->base.builder,
1025 vec_type, array_size, "");
1026 } else {
1027 for (i = 0; i < NUM_CHANNELS; i++)
1028 bld->temps[idx][i] = lp_build_alloca(bld->base.builder,
1029 vec_type, "");
1030 }
1031 break;
1032
1033 case TGSI_FILE_OUTPUT:
1034 for (i = 0; i < NUM_CHANNELS; i++)
1035 bld->outputs[idx][i] = lp_build_alloca(bld->base.builder,
1036 vec_type, "");
1037 break;
1038
1039 case TGSI_FILE_ADDRESS:
1040 assert(idx < LP_MAX_TGSI_ADDRS);
1041 for (i = 0; i < NUM_CHANNELS; i++)
1042 bld->addr[idx][i] = lp_build_alloca(bld->base.builder,
1043 vec_type, "");
1044 break;
1045
1046 case TGSI_FILE_PREDICATE:
1047 assert(idx < LP_MAX_TGSI_PREDS);
1048 for (i = 0; i < NUM_CHANNELS; i++)
1049 bld->preds[idx][i] = lp_build_alloca(bld->base.builder,
1050 vec_type, "");
1051 break;
1052
1053 default:
1054 /* don't need to declare other vars */
1055 break;
1056 }
1057 }
1058 }
1059
1060
1061 /**
1062 * Emit LLVM for one TGSI instruction.
1063 * \param return TRUE for success, FALSE otherwise
1064 */
1065 static boolean
1066 emit_instruction(
1067 struct lp_build_tgsi_soa_context *bld,
1068 const struct tgsi_full_instruction *inst,
1069 const struct tgsi_opcode_info *info,
1070 int *pc)
1071 {
1072 unsigned chan_index;
1073 LLVMValueRef src0, src1, src2;
1074 LLVMValueRef tmp0, tmp1, tmp2;
1075 LLVMValueRef tmp3 = NULL;
1076 LLVMValueRef tmp4 = NULL;
1077 LLVMValueRef tmp5 = NULL;
1078 LLVMValueRef tmp6 = NULL;
1079 LLVMValueRef tmp7 = NULL;
1080 LLVMValueRef res;
1081 LLVMValueRef dst0[NUM_CHANNELS];
1082
1083 /*
1084 * Stores and write masks are handled in a general fashion after the long
1085 * instruction opcode switch statement.
1086 *
1087 * Although not stricitly necessary, we avoid generating instructions for
1088 * channels which won't be stored, in cases where's that easy. For some
1089 * complex instructions, like texture sampling, it is more convenient to
1090 * assume a full writemask and then let LLVM optimization passes eliminate
1091 * redundant code.
1092 */
1093
1094 (*pc)++;
1095
1096 assert(info->num_dst <= 1);
1097 if (info->num_dst) {
1098 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1099 dst0[chan_index] = bld->base.undef;
1100 }
1101 }
1102
1103 switch (inst->Instruction.Opcode) {
1104 case TGSI_OPCODE_ARL:
1105 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1106 tmp0 = emit_fetch( bld, inst, 0, chan_index );
1107 tmp0 = lp_build_floor(&bld->base, tmp0);
1108 dst0[chan_index] = tmp0;
1109 }
1110 break;
1111
1112 case TGSI_OPCODE_MOV:
1113 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1114 dst0[chan_index] = emit_fetch( bld, inst, 0, chan_index );
1115 }
1116 break;
1117
1118 case TGSI_OPCODE_LIT:
1119 if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ) {
1120 dst0[CHAN_X] = bld->base.one;
1121 }
1122 if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
1123 src0 = emit_fetch( bld, inst, 0, CHAN_X );
1124 dst0[CHAN_Y] = lp_build_max( &bld->base, src0, bld->base.zero);
1125 }
1126 if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1127 /* XMM[1] = SrcReg[0].yyyy */
1128 tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1129 /* XMM[1] = max(XMM[1], 0) */
1130 tmp1 = lp_build_max( &bld->base, tmp1, bld->base.zero);
1131 /* XMM[2] = SrcReg[0].wwww */
1132 tmp2 = emit_fetch( bld, inst, 0, CHAN_W );
1133 tmp1 = lp_build_pow( &bld->base, tmp1, tmp2);
1134 tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1135 tmp2 = lp_build_cmp(&bld->base, PIPE_FUNC_GREATER, tmp0, bld->base.zero);
1136 dst0[CHAN_Z] = lp_build_select(&bld->base, tmp2, tmp1, bld->base.zero);
1137 }
1138 if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) ) {
1139 dst0[CHAN_W] = bld->base.one;
1140 }
1141 break;
1142
1143 case TGSI_OPCODE_RCP:
1144 /* TGSI_OPCODE_RECIP */
1145 src0 = emit_fetch( bld, inst, 0, CHAN_X );
1146 res = lp_build_rcp(&bld->base, src0);
1147 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1148 dst0[chan_index] = res;
1149 }
1150 break;
1151
1152 case TGSI_OPCODE_RSQ:
1153 /* TGSI_OPCODE_RECIPSQRT */
1154 src0 = emit_fetch( bld, inst, 0, CHAN_X );
1155 src0 = lp_build_abs(&bld->base, src0);
1156 res = lp_build_rsqrt(&bld->base, src0);
1157 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1158 dst0[chan_index] = res;
1159 }
1160 break;
1161
1162 case TGSI_OPCODE_EXP:
1163 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1164 IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
1165 IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
1166 LLVMValueRef *p_exp2_int_part = NULL;
1167 LLVMValueRef *p_frac_part = NULL;
1168 LLVMValueRef *p_exp2 = NULL;
1169
1170 src0 = emit_fetch( bld, inst, 0, CHAN_X );
1171
1172 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
1173 p_exp2_int_part = &tmp0;
1174 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
1175 p_frac_part = &tmp1;
1176 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
1177 p_exp2 = &tmp2;
1178
1179 lp_build_exp2_approx(&bld->base, src0, p_exp2_int_part, p_frac_part, p_exp2);
1180
1181 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
1182 dst0[CHAN_X] = tmp0;
1183 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
1184 dst0[CHAN_Y] = tmp1;
1185 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
1186 dst0[CHAN_Z] = tmp2;
1187 }
1188 /* dst.w = 1.0 */
1189 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
1190 dst0[CHAN_W] = bld->base.one;
1191 }
1192 break;
1193
1194 case TGSI_OPCODE_LOG:
1195 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1196 IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
1197 IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z )) {
1198 LLVMValueRef *p_floor_log2 = NULL;
1199 LLVMValueRef *p_exp = NULL;
1200 LLVMValueRef *p_log2 = NULL;
1201
1202 src0 = emit_fetch( bld, inst, 0, CHAN_X );
1203 src0 = lp_build_abs( &bld->base, src0 );
1204
1205 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
1206 p_floor_log2 = &tmp0;
1207 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ))
1208 p_exp = &tmp1;
1209 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
1210 p_log2 = &tmp2;
1211
1212 lp_build_log2_approx(&bld->base, src0, p_exp, p_floor_log2, p_log2);
1213
1214 /* dst.x = floor(lg2(abs(src.x))) */
1215 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ))
1216 dst0[CHAN_X] = tmp0;
1217 /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */
1218 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y )) {
1219 dst0[CHAN_Y] = lp_build_div( &bld->base, src0, tmp1);
1220 }
1221 /* dst.z = lg2(abs(src.x)) */
1222 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ))
1223 dst0[CHAN_Z] = tmp2;
1224 }
1225 /* dst.w = 1.0 */
1226 if (IS_DST0_CHANNEL_ENABLED( inst, CHAN_W )) {
1227 dst0[CHAN_W] = bld->base.one;
1228 }
1229 break;
1230
1231 case TGSI_OPCODE_MUL:
1232 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1233 src0 = emit_fetch( bld, inst, 0, chan_index );
1234 src1 = emit_fetch( bld, inst, 1, chan_index );
1235 dst0[chan_index] = lp_build_mul(&bld->base, src0, src1);
1236 }
1237 break;
1238
1239 case TGSI_OPCODE_ADD:
1240 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1241 src0 = emit_fetch( bld, inst, 0, chan_index );
1242 src1 = emit_fetch( bld, inst, 1, chan_index );
1243 dst0[chan_index] = lp_build_add(&bld->base, src0, src1);
1244 }
1245 break;
1246
1247 case TGSI_OPCODE_DP3:
1248 /* TGSI_OPCODE_DOT3 */
1249 tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1250 tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1251 tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1252 tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1253 tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1254 tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1255 tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1256 tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1257 tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1258 tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1259 tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1260 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1261 dst0[chan_index] = tmp0;
1262 }
1263 break;
1264
1265 case TGSI_OPCODE_DP4:
1266 /* TGSI_OPCODE_DOT4 */
1267 tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1268 tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1269 tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1270 tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1271 tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1272 tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1273 tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1274 tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1275 tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1276 tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1277 tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1278 tmp1 = emit_fetch( bld, inst, 0, CHAN_W );
1279 tmp2 = emit_fetch( bld, inst, 1, CHAN_W );
1280 tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1281 tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1282 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1283 dst0[chan_index] = tmp0;
1284 }
1285 break;
1286
1287 case TGSI_OPCODE_DST:
1288 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1289 dst0[CHAN_X] = bld->base.one;
1290 }
1291 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1292 tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1293 tmp1 = emit_fetch( bld, inst, 1, CHAN_Y );
1294 dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp0, tmp1);
1295 }
1296 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1297 dst0[CHAN_Z] = emit_fetch( bld, inst, 0, CHAN_Z );
1298 }
1299 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1300 dst0[CHAN_W] = emit_fetch( bld, inst, 1, CHAN_W );
1301 }
1302 break;
1303
1304 case TGSI_OPCODE_MIN:
1305 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1306 src0 = emit_fetch( bld, inst, 0, chan_index );
1307 src1 = emit_fetch( bld, inst, 1, chan_index );
1308 dst0[chan_index] = lp_build_min( &bld->base, src0, src1 );
1309 }
1310 break;
1311
1312 case TGSI_OPCODE_MAX:
1313 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1314 src0 = emit_fetch( bld, inst, 0, chan_index );
1315 src1 = emit_fetch( bld, inst, 1, chan_index );
1316 dst0[chan_index] = lp_build_max( &bld->base, src0, src1 );
1317 }
1318 break;
1319
1320 case TGSI_OPCODE_SLT:
1321 /* TGSI_OPCODE_SETLT */
1322 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1323 src0 = emit_fetch( bld, inst, 0, chan_index );
1324 src1 = emit_fetch( bld, inst, 1, chan_index );
1325 tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, src1 );
1326 dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1327 }
1328 break;
1329
1330 case TGSI_OPCODE_SGE:
1331 /* TGSI_OPCODE_SETGE */
1332 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1333 src0 = emit_fetch( bld, inst, 0, chan_index );
1334 src1 = emit_fetch( bld, inst, 1, chan_index );
1335 tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GEQUAL, src0, src1 );
1336 dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1337 }
1338 break;
1339
1340 case TGSI_OPCODE_MAD:
1341 /* TGSI_OPCODE_MADD */
1342 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1343 tmp0 = emit_fetch( bld, inst, 0, chan_index );
1344 tmp1 = emit_fetch( bld, inst, 1, chan_index );
1345 tmp2 = emit_fetch( bld, inst, 2, chan_index );
1346 tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1347 tmp0 = lp_build_add( &bld->base, tmp0, tmp2);
1348 dst0[chan_index] = tmp0;
1349 }
1350 break;
1351
1352 case TGSI_OPCODE_SUB:
1353 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1354 tmp0 = emit_fetch( bld, inst, 0, chan_index );
1355 tmp1 = emit_fetch( bld, inst, 1, chan_index );
1356 dst0[chan_index] = lp_build_sub( &bld->base, tmp0, tmp1);
1357 }
1358 break;
1359
1360 case TGSI_OPCODE_LRP:
1361 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1362 src0 = emit_fetch( bld, inst, 0, chan_index );
1363 src1 = emit_fetch( bld, inst, 1, chan_index );
1364 src2 = emit_fetch( bld, inst, 2, chan_index );
1365 tmp0 = lp_build_sub( &bld->base, src1, src2 );
1366 tmp0 = lp_build_mul( &bld->base, src0, tmp0 );
1367 dst0[chan_index] = lp_build_add( &bld->base, tmp0, src2 );
1368 }
1369 break;
1370
1371 case TGSI_OPCODE_CND:
1372 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1373 src0 = emit_fetch( bld, inst, 0, chan_index );
1374 src1 = emit_fetch( bld, inst, 1, chan_index );
1375 src2 = emit_fetch( bld, inst, 2, chan_index );
1376 tmp1 = lp_build_const_vec(bld->base.type, 0.5);
1377 tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src2, tmp1);
1378 dst0[chan_index] = lp_build_select( &bld->base, tmp0, src0, src1 );
1379 }
1380 break;
1381
1382 case TGSI_OPCODE_DP2A:
1383 tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); /* xmm0 = src[0].x */
1384 tmp1 = emit_fetch( bld, inst, 1, CHAN_X ); /* xmm1 = src[1].x */
1385 tmp0 = lp_build_mul( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 * xmm1 */
1386 tmp1 = emit_fetch( bld, inst, 0, CHAN_Y ); /* xmm1 = src[0].y */
1387 tmp2 = emit_fetch( bld, inst, 1, CHAN_Y ); /* xmm2 = src[1].y */
1388 tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); /* xmm1 = xmm1 * xmm2 */
1389 tmp0 = lp_build_add( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 + xmm1 */
1390 tmp1 = emit_fetch( bld, inst, 2, CHAN_X ); /* xmm1 = src[2].x */
1391 tmp0 = lp_build_add( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 + xmm1 */
1392 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1393 dst0[chan_index] = tmp0; /* dest[ch] = xmm0 */
1394 }
1395 break;
1396
1397 case TGSI_OPCODE_FRC:
1398 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1399 src0 = emit_fetch( bld, inst, 0, chan_index );
1400 tmp0 = lp_build_floor(&bld->base, src0);
1401 tmp0 = lp_build_sub(&bld->base, src0, tmp0);
1402 dst0[chan_index] = tmp0;
1403 }
1404 break;
1405
1406 case TGSI_OPCODE_CLAMP:
1407 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1408 tmp0 = emit_fetch( bld, inst, 0, chan_index );
1409 src1 = emit_fetch( bld, inst, 1, chan_index );
1410 src2 = emit_fetch( bld, inst, 2, chan_index );
1411 tmp0 = lp_build_max(&bld->base, tmp0, src1);
1412 tmp0 = lp_build_min(&bld->base, tmp0, src2);
1413 dst0[chan_index] = tmp0;
1414 }
1415 break;
1416
1417 case TGSI_OPCODE_FLR:
1418 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1419 tmp0 = emit_fetch( bld, inst, 0, chan_index );
1420 dst0[chan_index] = lp_build_floor(&bld->base, tmp0);
1421 }
1422 break;
1423
1424 case TGSI_OPCODE_ROUND:
1425 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1426 tmp0 = emit_fetch( bld, inst, 0, chan_index );
1427 dst0[chan_index] = lp_build_round(&bld->base, tmp0);
1428 }
1429 break;
1430
1431 case TGSI_OPCODE_EX2: {
1432 tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1433 tmp0 = lp_build_exp2( &bld->base, tmp0);
1434 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1435 dst0[chan_index] = tmp0;
1436 }
1437 break;
1438 }
1439
1440 case TGSI_OPCODE_LG2:
1441 tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1442 tmp0 = lp_build_log2( &bld->base, tmp0);
1443 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1444 dst0[chan_index] = tmp0;
1445 }
1446 break;
1447
1448 case TGSI_OPCODE_POW:
1449 src0 = emit_fetch( bld, inst, 0, CHAN_X );
1450 src1 = emit_fetch( bld, inst, 1, CHAN_X );
1451 res = lp_build_pow( &bld->base, src0, src1 );
1452 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1453 dst0[chan_index] = res;
1454 }
1455 break;
1456
1457 case TGSI_OPCODE_XPD:
1458 if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1459 IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ) {
1460 tmp1 = emit_fetch( bld, inst, 1, CHAN_Z );
1461 tmp3 = emit_fetch( bld, inst, 0, CHAN_Z );
1462 }
1463 if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) ||
1464 IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1465 tmp0 = emit_fetch( bld, inst, 0, CHAN_Y );
1466 tmp4 = emit_fetch( bld, inst, 1, CHAN_Y );
1467 }
1468 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1469 tmp2 = tmp0;
1470 tmp2 = lp_build_mul( &bld->base, tmp2, tmp1);
1471 tmp5 = tmp3;
1472 tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1473 tmp2 = lp_build_sub( &bld->base, tmp2, tmp5);
1474 dst0[CHAN_X] = tmp2;
1475 }
1476 if( IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) ||
1477 IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) ) {
1478 tmp2 = emit_fetch( bld, inst, 1, CHAN_X );
1479 tmp5 = emit_fetch( bld, inst, 0, CHAN_X );
1480 }
1481 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1482 tmp3 = lp_build_mul( &bld->base, tmp3, tmp2);
1483 tmp1 = lp_build_mul( &bld->base, tmp1, tmp5);
1484 tmp3 = lp_build_sub( &bld->base, tmp3, tmp1);
1485 dst0[CHAN_Y] = tmp3;
1486 }
1487 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1488 tmp5 = lp_build_mul( &bld->base, tmp5, tmp4);
1489 tmp0 = lp_build_mul( &bld->base, tmp0, tmp2);
1490 tmp5 = lp_build_sub( &bld->base, tmp5, tmp0);
1491 dst0[CHAN_Z] = tmp5;
1492 }
1493 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1494 dst0[CHAN_W] = bld->base.one;
1495 }
1496 break;
1497
1498 case TGSI_OPCODE_ABS:
1499 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1500 tmp0 = emit_fetch( bld, inst, 0, chan_index );
1501 dst0[chan_index] = lp_build_abs( &bld->base, tmp0 );
1502 }
1503 break;
1504
1505 case TGSI_OPCODE_RCC:
1506 /* deprecated? */
1507 assert(0);
1508 return FALSE;
1509
1510 case TGSI_OPCODE_DPH:
1511 tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1512 tmp1 = emit_fetch( bld, inst, 1, CHAN_X );
1513 tmp0 = lp_build_mul( &bld->base, tmp0, tmp1);
1514 tmp1 = emit_fetch( bld, inst, 0, CHAN_Y );
1515 tmp2 = emit_fetch( bld, inst, 1, CHAN_Y );
1516 tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1517 tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1518 tmp1 = emit_fetch( bld, inst, 0, CHAN_Z );
1519 tmp2 = emit_fetch( bld, inst, 1, CHAN_Z );
1520 tmp1 = lp_build_mul( &bld->base, tmp1, tmp2);
1521 tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1522 tmp1 = emit_fetch( bld, inst, 1, CHAN_W );
1523 tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1524 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1525 dst0[chan_index] = tmp0;
1526 }
1527 break;
1528
1529 case TGSI_OPCODE_COS:
1530 tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1531 tmp0 = lp_build_cos( &bld->base, tmp0 );
1532 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1533 dst0[chan_index] = tmp0;
1534 }
1535 break;
1536
1537 case TGSI_OPCODE_DDX:
1538 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1539 emit_fetch_deriv( bld, inst, 0, chan_index, NULL, &dst0[chan_index], NULL);
1540 }
1541 break;
1542
1543 case TGSI_OPCODE_DDY:
1544 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1545 emit_fetch_deriv( bld, inst, 0, chan_index, NULL, NULL, &dst0[chan_index]);
1546 }
1547 break;
1548
1549 case TGSI_OPCODE_KILP:
1550 /* predicated kill */
1551 emit_kilp( bld, inst );
1552 break;
1553
1554 case TGSI_OPCODE_KIL:
1555 /* conditional kill */
1556 emit_kil( bld, inst );
1557 break;
1558
1559 case TGSI_OPCODE_PK2H:
1560 return FALSE;
1561 break;
1562
1563 case TGSI_OPCODE_PK2US:
1564 return FALSE;
1565 break;
1566
1567 case TGSI_OPCODE_PK4B:
1568 return FALSE;
1569 break;
1570
1571 case TGSI_OPCODE_PK4UB:
1572 return FALSE;
1573 break;
1574
1575 case TGSI_OPCODE_RFL:
1576 return FALSE;
1577 break;
1578
1579 case TGSI_OPCODE_SEQ:
1580 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1581 src0 = emit_fetch( bld, inst, 0, chan_index );
1582 src1 = emit_fetch( bld, inst, 1, chan_index );
1583 tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_EQUAL, src0, src1 );
1584 dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1585 }
1586 break;
1587
1588 case TGSI_OPCODE_SFL:
1589 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1590 dst0[chan_index] = bld->base.zero;
1591 }
1592 break;
1593
1594 case TGSI_OPCODE_SGT:
1595 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1596 src0 = emit_fetch( bld, inst, 0, chan_index );
1597 src1 = emit_fetch( bld, inst, 1, chan_index );
1598 tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_GREATER, src0, src1 );
1599 dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1600 }
1601 break;
1602
1603 case TGSI_OPCODE_SIN:
1604 tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1605 tmp0 = lp_build_sin( &bld->base, tmp0 );
1606 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1607 dst0[chan_index] = tmp0;
1608 }
1609 break;
1610
1611 case TGSI_OPCODE_SLE:
1612 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1613 src0 = emit_fetch( bld, inst, 0, chan_index );
1614 src1 = emit_fetch( bld, inst, 1, chan_index );
1615 tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LEQUAL, src0, src1 );
1616 dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1617 }
1618 break;
1619
1620 case TGSI_OPCODE_SNE:
1621 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1622 src0 = emit_fetch( bld, inst, 0, chan_index );
1623 src1 = emit_fetch( bld, inst, 1, chan_index );
1624 tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_NOTEQUAL, src0, src1 );
1625 dst0[chan_index] = lp_build_select( &bld->base, tmp0, bld->base.one, bld->base.zero );
1626 }
1627 break;
1628
1629 case TGSI_OPCODE_STR:
1630 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1631 dst0[chan_index] = bld->base.one;
1632 }
1633 break;
1634
1635 case TGSI_OPCODE_TEX:
1636 emit_tex( bld, inst, LP_BLD_TEX_MODIFIER_NONE, dst0 );
1637 break;
1638
1639 case TGSI_OPCODE_TXD:
1640 emit_tex( bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV, dst0 );
1641 break;
1642
1643 case TGSI_OPCODE_UP2H:
1644 /* deprecated */
1645 assert (0);
1646 return FALSE;
1647 break;
1648
1649 case TGSI_OPCODE_UP2US:
1650 /* deprecated */
1651 assert(0);
1652 return FALSE;
1653 break;
1654
1655 case TGSI_OPCODE_UP4B:
1656 /* deprecated */
1657 assert(0);
1658 return FALSE;
1659 break;
1660
1661 case TGSI_OPCODE_UP4UB:
1662 /* deprecated */
1663 assert(0);
1664 return FALSE;
1665 break;
1666
1667 case TGSI_OPCODE_X2D:
1668 /* deprecated? */
1669 assert(0);
1670 return FALSE;
1671 break;
1672
1673 case TGSI_OPCODE_ARA:
1674 /* deprecated */
1675 assert(0);
1676 return FALSE;
1677 break;
1678
1679 case TGSI_OPCODE_ARR:
1680 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1681 tmp0 = emit_fetch( bld, inst, 0, chan_index );
1682 tmp0 = lp_build_round(&bld->base, tmp0);
1683 dst0[chan_index] = tmp0;
1684 }
1685 break;
1686
1687 case TGSI_OPCODE_BRA:
1688 /* deprecated */
1689 assert(0);
1690 return FALSE;
1691 break;
1692
1693 case TGSI_OPCODE_CAL:
1694 lp_exec_mask_call(&bld->exec_mask,
1695 inst->Label.Label,
1696 pc);
1697
1698 break;
1699
1700 case TGSI_OPCODE_RET:
1701 lp_exec_mask_ret(&bld->exec_mask, pc);
1702 break;
1703
1704 case TGSI_OPCODE_END:
1705 *pc = -1;
1706 break;
1707
1708 case TGSI_OPCODE_SSG:
1709 /* TGSI_OPCODE_SGN */
1710 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1711 tmp0 = emit_fetch( bld, inst, 0, chan_index );
1712 dst0[chan_index] = lp_build_sgn( &bld->base, tmp0 );
1713 }
1714 break;
1715
1716 case TGSI_OPCODE_CMP:
1717 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1718 src0 = emit_fetch( bld, inst, 0, chan_index );
1719 src1 = emit_fetch( bld, inst, 1, chan_index );
1720 src2 = emit_fetch( bld, inst, 2, chan_index );
1721 tmp0 = lp_build_cmp( &bld->base, PIPE_FUNC_LESS, src0, bld->base.zero );
1722 dst0[chan_index] = lp_build_select( &bld->base, tmp0, src1, src2);
1723 }
1724 break;
1725
1726 case TGSI_OPCODE_SCS:
1727 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_X ) {
1728 tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1729 dst0[CHAN_X] = lp_build_cos( &bld->base, tmp0 );
1730 }
1731 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Y ) {
1732 tmp0 = emit_fetch( bld, inst, 0, CHAN_X );
1733 dst0[CHAN_Y] = lp_build_sin( &bld->base, tmp0 );
1734 }
1735 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_Z ) {
1736 dst0[CHAN_Z] = bld->base.zero;
1737 }
1738 IF_IS_DST0_CHANNEL_ENABLED( inst, CHAN_W ) {
1739 dst0[CHAN_W] = bld->base.one;
1740 }
1741 break;
1742
1743 case TGSI_OPCODE_TXB:
1744 emit_tex( bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS, dst0 );
1745 break;
1746
1747 case TGSI_OPCODE_NRM:
1748 /* fall-through */
1749 case TGSI_OPCODE_NRM4:
1750 /* 3 or 4-component normalization */
1751 {
1752 uint dims = (inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
1753
1754 if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) ||
1755 IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y) ||
1756 IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z) ||
1757 (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 4)) {
1758
1759 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1760
1761 /* xmm4 = src.x */
1762 /* xmm0 = src.x * src.x */
1763 tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1764 if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1765 tmp4 = tmp0;
1766 }
1767 tmp0 = lp_build_mul( &bld->base, tmp0, tmp0);
1768
1769 /* xmm5 = src.y */
1770 /* xmm0 = xmm0 + src.y * src.y */
1771 tmp1 = emit_fetch(bld, inst, 0, CHAN_Y);
1772 if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1773 tmp5 = tmp1;
1774 }
1775 tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1776 tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1777
1778 /* xmm6 = src.z */
1779 /* xmm0 = xmm0 + src.z * src.z */
1780 tmp1 = emit_fetch(bld, inst, 0, CHAN_Z);
1781 if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1782 tmp6 = tmp1;
1783 }
1784 tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1785 tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1786
1787 if (dims == 4) {
1788 /* xmm7 = src.w */
1789 /* xmm0 = xmm0 + src.w * src.w */
1790 tmp1 = emit_fetch(bld, inst, 0, CHAN_W);
1791 if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W)) {
1792 tmp7 = tmp1;
1793 }
1794 tmp1 = lp_build_mul( &bld->base, tmp1, tmp1);
1795 tmp0 = lp_build_add( &bld->base, tmp0, tmp1);
1796 }
1797
1798 /* xmm1 = 1 / sqrt(xmm0) */
1799 tmp1 = lp_build_rsqrt( &bld->base, tmp0);
1800
1801 /* dst.x = xmm1 * src.x */
1802 if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X)) {
1803 dst0[CHAN_X] = lp_build_mul( &bld->base, tmp4, tmp1);
1804 }
1805
1806 /* dst.y = xmm1 * src.y */
1807 if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Y)) {
1808 dst0[CHAN_Y] = lp_build_mul( &bld->base, tmp5, tmp1);
1809 }
1810
1811 /* dst.z = xmm1 * src.z */
1812 if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_Z)) {
1813 dst0[CHAN_Z] = lp_build_mul( &bld->base, tmp6, tmp1);
1814 }
1815
1816 /* dst.w = xmm1 * src.w */
1817 if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_X) && dims == 4) {
1818 dst0[CHAN_W] = lp_build_mul( &bld->base, tmp7, tmp1);
1819 }
1820 }
1821
1822 /* dst.w = 1.0 */
1823 if (IS_DST0_CHANNEL_ENABLED(inst, CHAN_W) && dims == 3) {
1824 dst0[CHAN_W] = bld->base.one;
1825 }
1826 }
1827 break;
1828
1829 case TGSI_OPCODE_DIV:
1830 /* deprecated */
1831 assert( 0 );
1832 return FALSE;
1833 break;
1834
1835 case TGSI_OPCODE_DP2:
1836 tmp0 = emit_fetch( bld, inst, 0, CHAN_X ); /* xmm0 = src[0].x */
1837 tmp1 = emit_fetch( bld, inst, 1, CHAN_X ); /* xmm1 = src[1].x */
1838 tmp0 = lp_build_mul( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 * xmm1 */
1839 tmp1 = emit_fetch( bld, inst, 0, CHAN_Y ); /* xmm1 = src[0].y */
1840 tmp2 = emit_fetch( bld, inst, 1, CHAN_Y ); /* xmm2 = src[1].y */
1841 tmp1 = lp_build_mul( &bld->base, tmp1, tmp2); /* xmm1 = xmm1 * xmm2 */
1842 tmp0 = lp_build_add( &bld->base, tmp0, tmp1); /* xmm0 = xmm0 + xmm1 */
1843 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1844 dst0[chan_index] = tmp0; /* dest[ch] = xmm0 */
1845 }
1846 break;
1847
1848 case TGSI_OPCODE_TXL:
1849 emit_tex( bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD, dst0 );
1850 break;
1851
1852 case TGSI_OPCODE_TXP:
1853 emit_tex( bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED, dst0 );
1854 break;
1855
1856 case TGSI_OPCODE_BRK:
1857 lp_exec_break(&bld->exec_mask);
1858 break;
1859
1860 case TGSI_OPCODE_IF:
1861 tmp0 = emit_fetch(bld, inst, 0, CHAN_X);
1862 tmp0 = lp_build_cmp(&bld->base, PIPE_FUNC_NOTEQUAL,
1863 tmp0, bld->base.zero);
1864 lp_exec_mask_cond_push(&bld->exec_mask, tmp0);
1865 break;
1866
1867 case TGSI_OPCODE_BGNLOOP:
1868 lp_exec_bgnloop(&bld->exec_mask);
1869 break;
1870
1871 case TGSI_OPCODE_BGNSUB:
1872 lp_exec_mask_bgnsub(&bld->exec_mask);
1873 break;
1874
1875 case TGSI_OPCODE_ELSE:
1876 lp_exec_mask_cond_invert(&bld->exec_mask);
1877 break;
1878
1879 case TGSI_OPCODE_ENDIF:
1880 lp_exec_mask_cond_pop(&bld->exec_mask);
1881 break;
1882
1883 case TGSI_OPCODE_ENDLOOP:
1884 lp_exec_endloop(&bld->exec_mask);
1885 break;
1886
1887 case TGSI_OPCODE_ENDSUB:
1888 lp_exec_mask_endsub(&bld->exec_mask, pc);
1889 break;
1890
1891 case TGSI_OPCODE_PUSHA:
1892 /* deprecated? */
1893 assert(0);
1894 return FALSE;
1895 break;
1896
1897 case TGSI_OPCODE_POPA:
1898 /* deprecated? */
1899 assert(0);
1900 return FALSE;
1901 break;
1902
1903 case TGSI_OPCODE_CEIL:
1904 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1905 tmp0 = emit_fetch( bld, inst, 0, chan_index );
1906 dst0[chan_index] = lp_build_ceil(&bld->base, tmp0);
1907 }
1908 break;
1909
1910 case TGSI_OPCODE_I2F:
1911 /* deprecated? */
1912 assert(0);
1913 return FALSE;
1914 break;
1915
1916 case TGSI_OPCODE_NOT:
1917 /* deprecated? */
1918 assert(0);
1919 return FALSE;
1920 break;
1921
1922 case TGSI_OPCODE_TRUNC:
1923 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1924 tmp0 = emit_fetch( bld, inst, 0, chan_index );
1925 dst0[chan_index] = lp_build_trunc(&bld->base, tmp0);
1926 }
1927 break;
1928
1929 case TGSI_OPCODE_SHL:
1930 /* deprecated? */
1931 assert(0);
1932 return FALSE;
1933 break;
1934
1935 case TGSI_OPCODE_ISHR:
1936 /* deprecated? */
1937 assert(0);
1938 return FALSE;
1939 break;
1940
1941 case TGSI_OPCODE_AND:
1942 /* deprecated? */
1943 assert(0);
1944 return FALSE;
1945 break;
1946
1947 case TGSI_OPCODE_OR:
1948 /* deprecated? */
1949 assert(0);
1950 return FALSE;
1951 break;
1952
1953 case TGSI_OPCODE_MOD:
1954 /* deprecated? */
1955 assert(0);
1956 return FALSE;
1957 break;
1958
1959 case TGSI_OPCODE_XOR:
1960 /* deprecated? */
1961 assert(0);
1962 return FALSE;
1963 break;
1964
1965 case TGSI_OPCODE_SAD:
1966 /* deprecated? */
1967 assert(0);
1968 return FALSE;
1969 break;
1970
1971 case TGSI_OPCODE_TXF:
1972 /* deprecated? */
1973 assert(0);
1974 return FALSE;
1975 break;
1976
1977 case TGSI_OPCODE_TXQ:
1978 /* deprecated? */
1979 assert(0);
1980 return FALSE;
1981 break;
1982
1983 case TGSI_OPCODE_CONT:
1984 lp_exec_continue(&bld->exec_mask);
1985 break;
1986
1987 case TGSI_OPCODE_EMIT:
1988 return FALSE;
1989 break;
1990
1991 case TGSI_OPCODE_ENDPRIM:
1992 return FALSE;
1993 break;
1994
1995 case TGSI_OPCODE_NOP:
1996 break;
1997
1998 default:
1999 return FALSE;
2000 }
2001
2002 if(info->num_dst) {
2003 LLVMValueRef pred[NUM_CHANNELS];
2004
2005 emit_fetch_predicate( bld, inst, pred );
2006
2007 FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
2008 emit_store( bld, inst, 0, chan_index, pred[chan_index], dst0[chan_index]);
2009 }
2010 }
2011
2012 return TRUE;
2013 }
2014
2015
2016 void
2017 lp_build_tgsi_soa(LLVMBuilderRef builder,
2018 const struct tgsi_token *tokens,
2019 struct lp_type type,
2020 struct lp_build_mask_context *mask,
2021 LLVMValueRef consts_ptr,
2022 const LLVMValueRef *pos,
2023 const LLVMValueRef (*inputs)[NUM_CHANNELS],
2024 LLVMValueRef (*outputs)[NUM_CHANNELS],
2025 struct lp_build_sampler_soa *sampler,
2026 const struct tgsi_shader_info *info)
2027 {
2028 struct lp_build_tgsi_soa_context bld;
2029 struct tgsi_parse_context parse;
2030 uint num_immediates = 0;
2031 uint num_instructions = 0;
2032 unsigned i;
2033 int pc = 0;
2034
2035 struct lp_type res_type;
2036
2037 assert(type.length <= LP_MAX_VECTOR_LENGTH);
2038 memset(&res_type, 0, sizeof res_type);
2039 res_type.width = type.width;
2040 res_type.length = type.length;
2041 res_type.sign = 1;
2042
2043 /* Setup build context */
2044 memset(&bld, 0, sizeof bld);
2045 lp_build_context_init(&bld.base, builder, type);
2046 lp_build_context_init(&bld.uint_bld, builder, lp_uint_type(type));
2047 bld.mask = mask;
2048 bld.pos = pos;
2049 bld.inputs = inputs;
2050 bld.outputs = outputs;
2051 bld.consts_ptr = consts_ptr;
2052 bld.sampler = sampler;
2053 bld.info = info;
2054 bld.indirect_files = info->indirect_files;
2055 bld.instructions = (struct tgsi_full_instruction *)
2056 MALLOC( LP_MAX_INSTRUCTIONS * sizeof(struct tgsi_full_instruction) );
2057 bld.max_instructions = LP_MAX_INSTRUCTIONS;
2058
2059 if (!bld.instructions) {
2060 return;
2061 }
2062
2063 lp_exec_mask_init(&bld.exec_mask, &bld.base);
2064
2065 tgsi_parse_init( &parse, tokens );
2066
2067 while( !tgsi_parse_end_of_tokens( &parse ) ) {
2068 tgsi_parse_token( &parse );
2069
2070 switch( parse.FullToken.Token.Type ) {
2071 case TGSI_TOKEN_TYPE_DECLARATION:
2072 /* Inputs already interpolated */
2073 emit_declaration( &bld, &parse.FullToken.FullDeclaration );
2074 break;
2075
2076 case TGSI_TOKEN_TYPE_INSTRUCTION:
2077 {
2078 /* save expanded instruction */
2079 if (num_instructions == bld.max_instructions) {
2080 struct tgsi_full_instruction *instructions;
2081 instructions = REALLOC(bld.instructions,
2082 bld.max_instructions
2083 * sizeof(struct tgsi_full_instruction),
2084 (bld.max_instructions + LP_MAX_INSTRUCTIONS)
2085 * sizeof(struct tgsi_full_instruction));
2086 if (!instructions) {
2087 break;
2088 }
2089 bld.instructions = instructions;
2090 bld.max_instructions += LP_MAX_INSTRUCTIONS;
2091 }
2092
2093 memcpy(bld.instructions + num_instructions,
2094 &parse.FullToken.FullInstruction,
2095 sizeof(bld.instructions[0]));
2096
2097 num_instructions++;
2098 }
2099
2100 break;
2101
2102 case TGSI_TOKEN_TYPE_IMMEDIATE:
2103 /* simply copy the immediate values into the next immediates[] slot */
2104 {
2105 const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
2106 assert(size <= 4);
2107 assert(num_immediates < LP_MAX_TGSI_IMMEDIATES);
2108 for( i = 0; i < size; ++i )
2109 bld.immediates[num_immediates][i] =
2110 lp_build_const_vec(type, parse.FullToken.FullImmediate.u[i].Float);
2111 for( i = size; i < 4; ++i )
2112 bld.immediates[num_immediates][i] = bld.base.undef;
2113 num_immediates++;
2114 }
2115 break;
2116
2117 case TGSI_TOKEN_TYPE_PROPERTY:
2118 break;
2119
2120 default:
2121 assert( 0 );
2122 }
2123 }
2124
2125 while (pc != -1) {
2126 struct tgsi_full_instruction *instr = bld.instructions + pc;
2127 const struct tgsi_opcode_info *opcode_info =
2128 tgsi_get_opcode_info(instr->Instruction.Opcode);
2129 if (!emit_instruction( &bld, instr, opcode_info, &pc ))
2130 _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
2131 opcode_info->mnemonic);
2132 }
2133
2134 if (0) {
2135 LLVMBasicBlockRef block = LLVMGetInsertBlock(builder);
2136 LLVMValueRef function = LLVMGetBasicBlockParent(block);
2137 debug_printf("11111111111111111111111111111 \n");
2138 tgsi_dump(tokens, 0);
2139 lp_debug_dump_value(function);
2140 debug_printf("2222222222222222222222222222 \n");
2141 }
2142 tgsi_parse_free( &parse );
2143
2144 if (0) {
2145 LLVMModuleRef module = LLVMGetGlobalParent(
2146 LLVMGetBasicBlockParent(LLVMGetInsertBlock(bld.base.builder)));
2147 LLVMDumpModule(module);
2148
2149 }
2150
2151 FREE( bld.instructions );
2152 }
2153