223184d554afdf3d27a61a26fb62a610ffea1743
[mesa.git] / src / gallium / auxiliary / gallivm / lp_bld_tgsi_aos.c
1 /**************************************************************************
2 *
3 * Copyright 2010 VMware, Inc.
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 VMWARE 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 /**
29 * @file
30 * TGSI to LLVM IR translation -- AoS.
31 *
32 * FIXME:
33 * - No control flow support: the existing control flow code should be factored
34 * out into from the SoA code into a common module and shared.
35 * - No derivatives. Derivate logic should be pluggable, just like the samplers.
36 *
37 * @author Jose Fonseca <jfonseca@vmware.com>
38 */
39
40 #include "pipe/p_config.h"
41 #include "pipe/p_shader_tokens.h"
42 #include "util/u_debug.h"
43 #include "util/u_math.h"
44 #include "util/u_memory.h"
45 #include "tgsi/tgsi_dump.h"
46 #include "tgsi/tgsi_info.h"
47 #include "tgsi/tgsi_parse.h"
48 #include "tgsi/tgsi_util.h"
49 #include "tgsi/tgsi_scan.h"
50 #include "lp_bld_type.h"
51 #include "lp_bld_const.h"
52 #include "lp_bld_arit.h"
53 #include "lp_bld_logic.h"
54 #include "lp_bld_swizzle.h"
55 #include "lp_bld_flow.h"
56 #include "lp_bld_quad.h"
57 #include "lp_bld_tgsi.h"
58 #include "lp_bld_debug.h"
59 #include "lp_bld_sample.h"
60
61
62 /**
63 * Wrapper around lp_build_swizzle_aos which translates swizzles to another
64 * ordering.
65 */
66 static LLVMValueRef
67 swizzle_aos(struct lp_build_tgsi_context *bld_base,
68 LLVMValueRef a,
69 unsigned swizzle_x,
70 unsigned swizzle_y,
71 unsigned swizzle_z,
72 unsigned swizzle_w)
73 {
74 unsigned char swizzles[4];
75 struct lp_build_tgsi_aos_context *bld = lp_aos_context(bld_base);
76
77 assert(swizzle_x < 4);
78 assert(swizzle_y < 4);
79 assert(swizzle_z < 4);
80 assert(swizzle_w < 4);
81
82 swizzles[bld->inv_swizzles[0]] = bld->swizzles[swizzle_x];
83 swizzles[bld->inv_swizzles[1]] = bld->swizzles[swizzle_y];
84 swizzles[bld->inv_swizzles[2]] = bld->swizzles[swizzle_z];
85 swizzles[bld->inv_swizzles[3]] = bld->swizzles[swizzle_w];
86
87 return lp_build_swizzle_aos(&bld->bld_base.base, a, swizzles);
88 }
89
90
91 static LLVMValueRef
92 swizzle_scalar_aos(struct lp_build_tgsi_aos_context *bld,
93 LLVMValueRef a,
94 unsigned chan)
95 {
96 chan = bld->swizzles[chan];
97 return lp_build_swizzle_scalar_aos(&bld->bld_base.base, a, chan, 4);
98 }
99
100
101 static LLVMValueRef
102 emit_fetch_constant(
103 struct lp_build_tgsi_context * bld_base,
104 const struct tgsi_full_src_register * reg,
105 enum tgsi_opcode_type stype,
106 unsigned swizzle)
107 {
108 struct lp_build_tgsi_aos_context * bld = lp_aos_context(bld_base);
109 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
110 struct lp_type type = bld_base->base.type;
111 LLVMValueRef res;
112 unsigned chan;
113
114 assert(!reg->Register.Indirect);
115
116 /*
117 * Get the constants components
118 */
119
120 res = bld->bld_base.base.undef;
121 for (chan = 0; chan < 4; ++chan) {
122 LLVMValueRef index;
123 LLVMValueRef scalar_ptr;
124 LLVMValueRef scalar;
125 LLVMValueRef swizzle;
126
127 index = lp_build_const_int32(bld->bld_base.base.gallivm,
128 reg->Register.Index * 4 + chan);
129
130 scalar_ptr = LLVMBuildGEP(builder, bld->consts_ptr, &index, 1, "");
131
132 scalar = LLVMBuildLoad(builder, scalar_ptr, "");
133
134 lp_build_name(scalar, "const[%u].%c", reg->Register.Index, "xyzw"[chan]);
135
136 /*
137 * NOTE: constants array is always assumed to be RGBA
138 */
139
140 swizzle = lp_build_const_int32(bld->bld_base.base.gallivm,
141 bld->swizzles[chan]);
142
143 res = LLVMBuildInsertElement(builder, res, scalar, swizzle, "");
144 }
145
146 /*
147 * Broadcast the first quaternion to all others.
148 *
149 * XXX: could be factored into a reusable function.
150 */
151
152 if (type.length > 4) {
153 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
154 unsigned i;
155
156 for (chan = 0; chan < 4; ++chan) {
157 shuffles[chan] = lp_build_const_int32(bld->bld_base.base.gallivm, chan);
158 }
159
160 for (i = 4; i < type.length; ++i) {
161 shuffles[i] = shuffles[i % 4];
162 }
163
164 res = LLVMBuildShuffleVector(builder,
165 res, bld->bld_base.base.undef,
166 LLVMConstVector(shuffles, type.length),
167 "");
168 }
169 return res;
170 }
171
172 static LLVMValueRef
173 emit_fetch_immediate(
174 struct lp_build_tgsi_context * bld_base,
175 const struct tgsi_full_src_register * reg,
176 enum tgsi_opcode_type stype,
177 unsigned swizzle)
178 {
179 struct lp_build_tgsi_aos_context * bld = lp_aos_context(bld_base);
180 LLVMValueRef res = bld->immediates[reg->Register.Index];
181 assert(res);
182 return res;
183 }
184
185 static LLVMValueRef
186 emit_fetch_input(
187 struct lp_build_tgsi_context * bld_base,
188 const struct tgsi_full_src_register * reg,
189 enum tgsi_opcode_type stype,
190 unsigned swizzle)
191 {
192 struct lp_build_tgsi_aos_context * bld = lp_aos_context(bld_base);
193 LLVMValueRef res = bld->inputs[reg->Register.Index];
194 assert(!reg->Register.Indirect);
195 assert(res);
196 return res;
197 }
198
199 static LLVMValueRef
200 emit_fetch_temporary(
201 struct lp_build_tgsi_context * bld_base,
202 const struct tgsi_full_src_register * reg,
203 enum tgsi_opcode_type stype,
204 unsigned swizzle)
205 {
206 struct lp_build_tgsi_aos_context * bld = lp_aos_context(bld_base);
207 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
208 LLVMValueRef temp_ptr = bld->temps[reg->Register.Index];
209 LLVMValueRef res = LLVMBuildLoad(builder, temp_ptr, "");
210 assert(!reg->Register.Indirect);
211 if (!res)
212 return bld->bld_base.base.undef;
213
214 return res;
215 }
216
217 /**
218 * Register store.
219 */
220 void
221 lp_emit_store_aos(
222 struct lp_build_tgsi_aos_context *bld,
223 const struct tgsi_full_instruction *inst,
224 unsigned index,
225 LLVMValueRef value)
226 {
227 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
228 const struct tgsi_full_dst_register *reg = &inst->Dst[index];
229 LLVMValueRef mask = NULL;
230 LLVMValueRef ptr;
231
232 /*
233 * Saturate the value
234 */
235
236 switch (inst->Instruction.Saturate) {
237 case TGSI_SAT_NONE:
238 break;
239
240 case TGSI_SAT_ZERO_ONE:
241 value = lp_build_max(&bld->bld_base.base, value, bld->bld_base.base.zero);
242 value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
243 break;
244
245 case TGSI_SAT_MINUS_PLUS_ONE:
246 value = lp_build_max(&bld->bld_base.base, value, lp_build_const_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, -1.0));
247 value = lp_build_min(&bld->bld_base.base, value, bld->bld_base.base.one);
248 break;
249
250 default:
251 assert(0);
252 }
253
254 /*
255 * Translate the register file
256 */
257
258 assert(!reg->Register.Indirect);
259
260 switch (reg->Register.File) {
261 case TGSI_FILE_OUTPUT:
262 ptr = bld->outputs[reg->Register.Index];
263 break;
264
265 case TGSI_FILE_TEMPORARY:
266 ptr = bld->temps[reg->Register.Index];
267 break;
268
269 case TGSI_FILE_ADDRESS:
270 ptr = bld->addr[reg->Indirect.Index];
271 break;
272
273 case TGSI_FILE_PREDICATE:
274 ptr = bld->preds[reg->Register.Index];
275 break;
276
277 default:
278 assert(0);
279 return;
280 }
281
282 if (!ptr)
283 return;
284 /*
285 * Predicate
286 */
287
288 if (inst->Instruction.Predicate) {
289 LLVMValueRef pred;
290
291 assert(inst->Predicate.Index < LP_MAX_TGSI_PREDS);
292
293 pred = LLVMBuildLoad(builder,
294 bld->preds[inst->Predicate.Index], "");
295
296 /*
297 * Convert the value to an integer mask.
298 */
299 pred = lp_build_compare(bld->bld_base.base.gallivm,
300 bld->bld_base.base.type,
301 PIPE_FUNC_NOTEQUAL,
302 pred,
303 bld->bld_base.base.zero);
304
305 if (inst->Predicate.Negate) {
306 pred = LLVMBuildNot(builder, pred, "");
307 }
308
309 pred = bld->bld_base.emit_swizzle(&bld->bld_base, pred,
310 inst->Predicate.SwizzleX,
311 inst->Predicate.SwizzleY,
312 inst->Predicate.SwizzleZ,
313 inst->Predicate.SwizzleW);
314
315 if (mask) {
316 mask = LLVMBuildAnd(builder, mask, pred, "");
317 } else {
318 mask = pred;
319 }
320 }
321
322 /*
323 * Writemask
324 */
325
326 if (reg->Register.WriteMask != TGSI_WRITEMASK_XYZW) {
327 LLVMValueRef writemask;
328
329 writemask = lp_build_const_mask_aos_swizzled(bld->bld_base.base.gallivm,
330 bld->bld_base.base.type,
331 reg->Register.WriteMask,
332 TGSI_NUM_CHANNELS,
333 bld->swizzles);
334
335 if (mask) {
336 mask = LLVMBuildAnd(builder, mask, writemask, "");
337 } else {
338 mask = writemask;
339 }
340 }
341
342 if (mask) {
343 LLVMValueRef orig_value;
344
345 orig_value = LLVMBuildLoad(builder, ptr, "");
346 value = lp_build_select(&bld->bld_base.base,
347 mask, value, orig_value);
348 }
349
350 LLVMBuildStore(builder, value, ptr);
351 }
352
353
354 /**
355 * High-level instruction translators.
356 */
357
358 static LLVMValueRef
359 emit_tex(struct lp_build_tgsi_aos_context *bld,
360 const struct tgsi_full_instruction *inst,
361 enum lp_build_tex_modifier modifier)
362 {
363 unsigned target;
364 unsigned unit;
365 LLVMValueRef coords;
366 struct lp_derivatives derivs = { {NULL}, {NULL} };
367
368 if (!bld->sampler) {
369 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
370 return bld->bld_base.base.undef;
371 }
372
373 target = inst->Texture.Texture;
374
375 coords = lp_build_emit_fetch( &bld->bld_base, inst, 0 , LP_CHAN_ALL);
376
377 if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
378 /* probably not going to work */
379 derivs.ddx[0] = lp_build_emit_fetch( &bld->bld_base, inst, 1 , LP_CHAN_ALL);
380 derivs.ddy[0] = lp_build_emit_fetch( &bld->bld_base, inst, 2 , LP_CHAN_ALL);
381 unit = inst->Src[3].Register.Index;
382 }
383 else {
384 unit = inst->Src[1].Register.Index;
385 }
386 return bld->sampler->emit_fetch_texel(bld->sampler,
387 &bld->bld_base.base,
388 target, unit,
389 coords, derivs,
390 modifier);
391 }
392
393
394 void
395 lp_emit_declaration_aos(
396 struct lp_build_tgsi_aos_context *bld,
397 const struct tgsi_full_declaration *decl)
398 {
399 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
400 LLVMTypeRef vec_type = lp_build_vec_type(bld->bld_base.base.gallivm, bld->bld_base.base.type);
401
402 unsigned first = decl->Range.First;
403 unsigned last = decl->Range.Last;
404 unsigned idx;
405
406 for (idx = first; idx <= last; ++idx) {
407 switch (decl->Declaration.File) {
408 case TGSI_FILE_TEMPORARY:
409 assert(idx < LP_MAX_TGSI_TEMPS);
410 if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
411 LLVMValueRef array_size = lp_build_const_int32(gallivm, last + 1);
412 bld->temps_array = lp_build_array_alloca(bld->bld_base.base.gallivm,
413 vec_type, array_size, "");
414 } else {
415 bld->temps[idx] = lp_build_alloca(gallivm, vec_type, "");
416 }
417 break;
418
419 case TGSI_FILE_OUTPUT:
420 bld->outputs[idx] = lp_build_alloca(gallivm, vec_type, "");
421 break;
422
423 case TGSI_FILE_ADDRESS:
424 assert(idx < LP_MAX_TGSI_ADDRS);
425 bld->addr[idx] = lp_build_alloca(gallivm, vec_type, "");
426 break;
427
428 case TGSI_FILE_PREDICATE:
429 assert(idx < LP_MAX_TGSI_PREDS);
430 bld->preds[idx] = lp_build_alloca(gallivm, vec_type, "");
431 break;
432
433 default:
434 /* don't need to declare other vars */
435 break;
436 }
437 }
438 }
439
440
441 /**
442 * Emit LLVM for one TGSI instruction.
443 * \param return TRUE for success, FALSE otherwise
444 */
445 boolean
446 lp_emit_instruction_aos(
447 struct lp_build_tgsi_aos_context *bld,
448 const struct tgsi_full_instruction *inst,
449 const struct tgsi_opcode_info *info,
450 int *pc)
451 {
452 LLVMValueRef src0, src1, src2;
453 LLVMValueRef tmp0, tmp1;
454 LLVMValueRef dst0 = NULL;
455
456 /*
457 * Stores and write masks are handled in a general fashion after the long
458 * instruction opcode switch statement.
459 *
460 * Although not stricitly necessary, we avoid generating instructions for
461 * channels which won't be stored, in cases where's that easy. For some
462 * complex instructions, like texture sampling, it is more convenient to
463 * assume a full writemask and then let LLVM optimization passes eliminate
464 * redundant code.
465 */
466
467 (*pc)++;
468
469 assert(info->num_dst <= 1);
470 if (info->num_dst) {
471 dst0 = bld->bld_base.base.undef;
472 }
473
474 switch (inst->Instruction.Opcode) {
475 case TGSI_OPCODE_ARL:
476 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
477 dst0 = lp_build_floor(&bld->bld_base.base, src0);
478 break;
479
480 case TGSI_OPCODE_MOV:
481 dst0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
482 break;
483
484 case TGSI_OPCODE_LIT:
485 return FALSE;
486
487 case TGSI_OPCODE_RCP:
488 /* TGSI_OPCODE_RECIP */
489 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
490 dst0 = lp_build_rcp(&bld->bld_base.base, src0);
491 break;
492
493 case TGSI_OPCODE_RSQ:
494 /* TGSI_OPCODE_RECIPSQRT */
495 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
496 tmp0 = lp_build_emit_llvm_unary(&bld->bld_base, TGSI_OPCODE_ABS, src0);
497 dst0 = lp_build_rsqrt(&bld->bld_base.base, tmp0);
498 break;
499
500 case TGSI_OPCODE_EXP:
501 return FALSE;
502
503 case TGSI_OPCODE_LOG:
504 return FALSE;
505
506 case TGSI_OPCODE_MUL:
507 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
508 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
509 dst0 = lp_build_mul(&bld->bld_base.base, src0, src1);
510 break;
511
512 case TGSI_OPCODE_ADD:
513 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
514 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
515 dst0 = lp_build_add(&bld->bld_base.base, src0, src1);
516 break;
517
518 case TGSI_OPCODE_DP3:
519 /* TGSI_OPCODE_DOT3 */
520 return FALSE;
521
522 case TGSI_OPCODE_DP4:
523 /* TGSI_OPCODE_DOT4 */
524 return FALSE;
525
526 case TGSI_OPCODE_DST:
527 return FALSE;
528
529 case TGSI_OPCODE_MIN:
530 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
531 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
532 dst0 = lp_build_min(&bld->bld_base.base, src0, src1);
533 break;
534
535 case TGSI_OPCODE_MAX:
536 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
537 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
538 dst0 = lp_build_max(&bld->bld_base.base, src0, src1);
539 break;
540
541 case TGSI_OPCODE_SLT:
542 /* TGSI_OPCODE_SETLT */
543 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
544 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
545 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, src1);
546 dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
547 break;
548
549 case TGSI_OPCODE_SGE:
550 /* TGSI_OPCODE_SETGE */
551 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
552 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
553 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, src0, src1);
554 dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
555 break;
556
557 case TGSI_OPCODE_MAD:
558 /* TGSI_OPCODE_MADD */
559 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
560 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
561 src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
562 tmp0 = lp_build_mul(&bld->bld_base.base, src0, src1);
563 dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
564 break;
565
566 case TGSI_OPCODE_SUB:
567 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
568 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
569 dst0 = lp_build_sub(&bld->bld_base.base, src0, src1);
570 break;
571
572 case TGSI_OPCODE_LRP:
573 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
574 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
575 src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
576 tmp0 = lp_build_sub(&bld->bld_base.base, src1, src2);
577 tmp0 = lp_build_mul(&bld->bld_base.base, src0, tmp0);
578 dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
579 break;
580
581 case TGSI_OPCODE_CND:
582 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
583 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
584 src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
585 tmp1 = lp_build_const_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, 0.5);
586 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src2, tmp1);
587 dst0 = lp_build_select(&bld->bld_base.base, tmp0, src0, src1);
588 break;
589
590 case TGSI_OPCODE_DP2A:
591 return FALSE;
592
593 case TGSI_OPCODE_FRC:
594 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
595 tmp0 = lp_build_floor(&bld->bld_base.base, src0);
596 dst0 = lp_build_sub(&bld->bld_base.base, src0, tmp0);
597 break;
598
599 case TGSI_OPCODE_CLAMP:
600 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
601 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
602 src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
603 tmp0 = lp_build_max(&bld->bld_base.base, src0, src1);
604 dst0 = lp_build_min(&bld->bld_base.base, tmp0, src2);
605 break;
606
607 case TGSI_OPCODE_FLR:
608 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
609 dst0 = lp_build_floor(&bld->bld_base.base, src0);
610 break;
611
612 case TGSI_OPCODE_ROUND:
613 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
614 dst0 = lp_build_round(&bld->bld_base.base, src0);
615 break;
616
617 case TGSI_OPCODE_EX2:
618 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
619 tmp0 = lp_build_swizzle_scalar_aos(&bld->bld_base.base, src0, TGSI_SWIZZLE_X, TGSI_NUM_CHANNELS);
620 dst0 = lp_build_exp2(&bld->bld_base.base, tmp0);
621 break;
622
623 case TGSI_OPCODE_LG2:
624 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
625 tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
626 dst0 = lp_build_log2(&bld->bld_base.base, tmp0);
627 break;
628
629 case TGSI_OPCODE_POW:
630 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
631 src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
632 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
633 src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X);
634 dst0 = lp_build_pow(&bld->bld_base.base, src0, src1);
635 break;
636
637 case TGSI_OPCODE_XPD:
638 return FALSE;
639
640 case TGSI_OPCODE_RCC:
641 /* deprecated? */
642 assert(0);
643 return FALSE;
644
645 case TGSI_OPCODE_DPH:
646 return FALSE;
647
648 case TGSI_OPCODE_COS:
649 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
650 tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
651 dst0 = lp_build_cos(&bld->bld_base.base, tmp0);
652 break;
653
654 case TGSI_OPCODE_DDX:
655 return FALSE;
656
657 case TGSI_OPCODE_DDY:
658 return FALSE;
659
660 case TGSI_OPCODE_KILP:
661 /* predicated kill */
662 return FALSE;
663
664 case TGSI_OPCODE_KIL:
665 /* conditional kill */
666 return FALSE;
667
668 case TGSI_OPCODE_PK2H:
669 return FALSE;
670 break;
671
672 case TGSI_OPCODE_PK2US:
673 return FALSE;
674 break;
675
676 case TGSI_OPCODE_PK4B:
677 return FALSE;
678 break;
679
680 case TGSI_OPCODE_PK4UB:
681 return FALSE;
682
683 case TGSI_OPCODE_RFL:
684 return FALSE;
685
686 case TGSI_OPCODE_SEQ:
687 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
688 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
689 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_EQUAL, src0, src1);
690 dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
691 break;
692
693 case TGSI_OPCODE_SFL:
694 dst0 = bld->bld_base.base.zero;
695 break;
696
697 case TGSI_OPCODE_SGT:
698 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
699 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
700 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src0, src1);
701 dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
702 break;
703
704 case TGSI_OPCODE_SIN:
705 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
706 tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
707 dst0 = lp_build_sin(&bld->bld_base.base, tmp0);
708 break;
709
710 case TGSI_OPCODE_SLE:
711 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
712 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
713 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LEQUAL, src0, src1);
714 dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
715 break;
716
717 case TGSI_OPCODE_SNE:
718 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
719 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
720 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_NOTEQUAL, src0, src1);
721 dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
722 break;
723
724 case TGSI_OPCODE_STR:
725 dst0 = bld->bld_base.base.one;
726 break;
727
728 case TGSI_OPCODE_TEX:
729 dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
730 break;
731
732 case TGSI_OPCODE_TXD:
733 dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
734 break;
735
736 case TGSI_OPCODE_UP2H:
737 /* deprecated */
738 assert (0);
739 return FALSE;
740 break;
741
742 case TGSI_OPCODE_UP2US:
743 /* deprecated */
744 assert(0);
745 return FALSE;
746 break;
747
748 case TGSI_OPCODE_UP4B:
749 /* deprecated */
750 assert(0);
751 return FALSE;
752 break;
753
754 case TGSI_OPCODE_UP4UB:
755 /* deprecated */
756 assert(0);
757 return FALSE;
758 break;
759
760 case TGSI_OPCODE_X2D:
761 /* deprecated? */
762 assert(0);
763 return FALSE;
764 break;
765
766 case TGSI_OPCODE_ARA:
767 /* deprecated */
768 assert(0);
769 return FALSE;
770 break;
771
772 case TGSI_OPCODE_ARR:
773 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
774 dst0 = lp_build_round(&bld->bld_base.base, src0);
775 break;
776
777 case TGSI_OPCODE_BRA:
778 /* deprecated */
779 assert(0);
780 return FALSE;
781 break;
782
783 case TGSI_OPCODE_CAL:
784 return FALSE;
785
786 case TGSI_OPCODE_RET:
787 return FALSE;
788
789 case TGSI_OPCODE_END:
790 *pc = -1;
791 break;
792
793 case TGSI_OPCODE_SSG:
794 /* TGSI_OPCODE_SGN */
795 tmp0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
796 dst0 = lp_build_sgn(&bld->bld_base.base, tmp0);
797 break;
798
799 case TGSI_OPCODE_CMP:
800 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
801 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
802 src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
803 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, bld->bld_base.base.zero);
804 dst0 = lp_build_select(&bld->bld_base.base, tmp0, src1, src2);
805 break;
806
807 case TGSI_OPCODE_SCS:
808 return FALSE;
809
810 case TGSI_OPCODE_TXB:
811 dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
812 break;
813
814 case TGSI_OPCODE_NRM:
815 /* fall-through */
816 case TGSI_OPCODE_NRM4:
817 return FALSE;
818
819 case TGSI_OPCODE_DIV:
820 /* deprecated */
821 assert(0);
822 return FALSE;
823 break;
824
825 case TGSI_OPCODE_DP2:
826 return FALSE;
827
828 case TGSI_OPCODE_TXL:
829 dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
830 break;
831
832 case TGSI_OPCODE_TXP:
833 dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
834 break;
835
836 case TGSI_OPCODE_BRK:
837 return FALSE;
838
839 case TGSI_OPCODE_IF:
840 case TGSI_OPCODE_UIF:
841 return FALSE;
842
843 case TGSI_OPCODE_BGNLOOP:
844 return FALSE;
845
846 case TGSI_OPCODE_BGNSUB:
847 return FALSE;
848
849 case TGSI_OPCODE_ELSE:
850 return FALSE;
851
852 case TGSI_OPCODE_ENDIF:
853 return FALSE;
854
855 case TGSI_OPCODE_ENDLOOP:
856 return FALSE;
857
858 case TGSI_OPCODE_ENDSUB:
859 return FALSE;
860
861 case TGSI_OPCODE_PUSHA:
862 /* deprecated? */
863 assert(0);
864 return FALSE;
865 break;
866
867 case TGSI_OPCODE_POPA:
868 /* deprecated? */
869 assert(0);
870 return FALSE;
871 break;
872
873 case TGSI_OPCODE_CEIL:
874 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
875 dst0 = lp_build_ceil(&bld->bld_base.base, src0);
876 break;
877
878 case TGSI_OPCODE_I2F:
879 /* deprecated? */
880 assert(0);
881 return FALSE;
882 break;
883
884 case TGSI_OPCODE_NOT:
885 /* deprecated? */
886 assert(0);
887 return FALSE;
888 break;
889
890 case TGSI_OPCODE_TRUNC:
891 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
892 dst0 = lp_build_trunc(&bld->bld_base.base, src0);
893 break;
894
895 case TGSI_OPCODE_SHL:
896 /* deprecated? */
897 assert(0);
898 return FALSE;
899 break;
900
901 case TGSI_OPCODE_ISHR:
902 /* deprecated? */
903 assert(0);
904 return FALSE;
905 break;
906
907 case TGSI_OPCODE_AND:
908 /* deprecated? */
909 assert(0);
910 return FALSE;
911 break;
912
913 case TGSI_OPCODE_OR:
914 /* deprecated? */
915 assert(0);
916 return FALSE;
917 break;
918
919 case TGSI_OPCODE_MOD:
920 /* deprecated? */
921 assert(0);
922 return FALSE;
923 break;
924
925 case TGSI_OPCODE_XOR:
926 /* deprecated? */
927 assert(0);
928 return FALSE;
929 break;
930
931 case TGSI_OPCODE_SAD:
932 /* deprecated? */
933 assert(0);
934 return FALSE;
935 break;
936
937 case TGSI_OPCODE_TXF:
938 /* deprecated? */
939 assert(0);
940 return FALSE;
941 break;
942
943 case TGSI_OPCODE_TXQ:
944 /* deprecated? */
945 assert(0);
946 return FALSE;
947 break;
948
949 case TGSI_OPCODE_CONT:
950 return FALSE;
951
952 case TGSI_OPCODE_EMIT:
953 return FALSE;
954 break;
955
956 case TGSI_OPCODE_ENDPRIM:
957 return FALSE;
958 break;
959
960 case TGSI_OPCODE_NOP:
961 break;
962
963 default:
964 return FALSE;
965 }
966
967 if (info->num_dst) {
968 lp_emit_store_aos(bld, inst, 0, dst0);
969 }
970
971 return TRUE;
972 }
973
974
975 void
976 lp_build_tgsi_aos(struct gallivm_state *gallivm,
977 const struct tgsi_token *tokens,
978 struct lp_type type,
979 const unsigned char swizzles[4],
980 LLVMValueRef consts_ptr,
981 const LLVMValueRef *inputs,
982 LLVMValueRef *outputs,
983 struct lp_build_sampler_aos *sampler,
984 const struct tgsi_shader_info *info)
985 {
986 struct lp_build_tgsi_aos_context bld;
987 struct tgsi_parse_context parse;
988 uint num_immediates = 0;
989 unsigned chan;
990 int pc = 0;
991
992 /* Setup build context */
993 memset(&bld, 0, sizeof bld);
994 lp_build_context_init(&bld.bld_base.base, gallivm, type);
995 lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
996 lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
997 lp_build_context_init(&bld.int_bld, gallivm, lp_int_type(type));
998
999 for (chan = 0; chan < 4; ++chan) {
1000 bld.swizzles[chan] = swizzles[chan];
1001 bld.inv_swizzles[swizzles[chan]] = chan;
1002 }
1003
1004 bld.inputs = inputs;
1005 bld.outputs = outputs;
1006 bld.consts_ptr = consts_ptr;
1007 bld.sampler = sampler;
1008 bld.indirect_files = info->indirect_files;
1009 bld.bld_base.emit_swizzle = swizzle_aos;
1010 bld.bld_base.info = info;
1011
1012 bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
1013 bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
1014 bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
1015 bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
1016
1017 /* Set opcode actions */
1018 lp_set_default_actions_cpu(&bld.bld_base);
1019
1020 if (!lp_bld_tgsi_list_init(&bld.bld_base)) {
1021 return;
1022 }
1023
1024 tgsi_parse_init(&parse, tokens);
1025
1026 while (!tgsi_parse_end_of_tokens(&parse)) {
1027 tgsi_parse_token(&parse);
1028
1029 switch(parse.FullToken.Token.Type) {
1030 case TGSI_TOKEN_TYPE_DECLARATION:
1031 /* Inputs already interpolated */
1032 lp_emit_declaration_aos(&bld, &parse.FullToken.FullDeclaration);
1033 break;
1034
1035 case TGSI_TOKEN_TYPE_INSTRUCTION:
1036 /* save expanded instruction */
1037 lp_bld_tgsi_add_instruction(&bld.bld_base,
1038 &parse.FullToken.FullInstruction);
1039 break;
1040
1041 case TGSI_TOKEN_TYPE_IMMEDIATE:
1042 /* simply copy the immediate values into the next immediates[] slot */
1043 {
1044 const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1045 float imm[4];
1046 assert(size <= 4);
1047 assert(num_immediates < LP_MAX_TGSI_IMMEDIATES);
1048 for (chan = 0; chan < 4; ++chan) {
1049 imm[chan] = 0.0f;
1050 }
1051 for (chan = 0; chan < size; ++chan) {
1052 unsigned swizzle = bld.swizzles[chan];
1053 imm[swizzle] = parse.FullToken.FullImmediate.u[chan].Float;
1054 }
1055 bld.immediates[num_immediates] =
1056 lp_build_const_aos(gallivm, type,
1057 imm[0], imm[1], imm[2], imm[3],
1058 NULL);
1059 num_immediates++;
1060 }
1061 break;
1062
1063 case TGSI_TOKEN_TYPE_PROPERTY:
1064 break;
1065
1066 default:
1067 assert(0);
1068 }
1069 }
1070
1071 while (pc != -1) {
1072 struct tgsi_full_instruction *instr = bld.bld_base.instructions + pc;
1073 const struct tgsi_opcode_info *opcode_info =
1074 tgsi_get_opcode_info(instr->Instruction.Opcode);
1075 if (!lp_emit_instruction_aos(&bld, instr, opcode_info, &pc))
1076 _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1077 opcode_info->mnemonic);
1078 }
1079
1080 if (0) {
1081 LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
1082 LLVMValueRef function = LLVMGetBasicBlockParent(block);
1083 debug_printf("11111111111111111111111111111 \n");
1084 tgsi_dump(tokens, 0);
1085 lp_debug_dump_value(function);
1086 debug_printf("2222222222222222222222222222 \n");
1087 }
1088 tgsi_parse_free(&parse);
1089 FREE(bld.bld_base.instructions);
1090
1091 if (0) {
1092 LLVMModuleRef module = LLVMGetGlobalParent(
1093 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
1094 LLVMDumpModule(module);
1095 }
1096
1097 }
1098