tgsi: Rework calls to ureg_emit_insn().
[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_KILL:
661 return FALSE;
662
663 case TGSI_OPCODE_KILL_IF:
664 return FALSE;
665
666 case TGSI_OPCODE_PK2H:
667 return FALSE;
668 break;
669
670 case TGSI_OPCODE_PK2US:
671 return FALSE;
672 break;
673
674 case TGSI_OPCODE_PK4B:
675 return FALSE;
676 break;
677
678 case TGSI_OPCODE_PK4UB:
679 return FALSE;
680
681 case TGSI_OPCODE_RFL:
682 return FALSE;
683
684 case TGSI_OPCODE_SEQ:
685 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
686 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
687 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_EQUAL, src0, src1);
688 dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
689 break;
690
691 case TGSI_OPCODE_SFL:
692 dst0 = bld->bld_base.base.zero;
693 break;
694
695 case TGSI_OPCODE_SGT:
696 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
697 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
698 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src0, src1);
699 dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
700 break;
701
702 case TGSI_OPCODE_SIN:
703 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
704 tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X);
705 dst0 = lp_build_sin(&bld->bld_base.base, tmp0);
706 break;
707
708 case TGSI_OPCODE_SLE:
709 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
710 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
711 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LEQUAL, src0, src1);
712 dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
713 break;
714
715 case TGSI_OPCODE_SNE:
716 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
717 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
718 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_NOTEQUAL, src0, src1);
719 dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero);
720 break;
721
722 case TGSI_OPCODE_STR:
723 dst0 = bld->bld_base.base.one;
724 break;
725
726 case TGSI_OPCODE_TEX:
727 dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE);
728 break;
729
730 case TGSI_OPCODE_TXD:
731 dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV);
732 break;
733
734 case TGSI_OPCODE_UP2H:
735 /* deprecated */
736 assert (0);
737 return FALSE;
738 break;
739
740 case TGSI_OPCODE_UP2US:
741 /* deprecated */
742 assert(0);
743 return FALSE;
744 break;
745
746 case TGSI_OPCODE_UP4B:
747 /* deprecated */
748 assert(0);
749 return FALSE;
750 break;
751
752 case TGSI_OPCODE_UP4UB:
753 /* deprecated */
754 assert(0);
755 return FALSE;
756 break;
757
758 case TGSI_OPCODE_X2D:
759 /* deprecated? */
760 assert(0);
761 return FALSE;
762 break;
763
764 case TGSI_OPCODE_ARA:
765 /* deprecated */
766 assert(0);
767 return FALSE;
768 break;
769
770 case TGSI_OPCODE_ARR:
771 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
772 dst0 = lp_build_round(&bld->bld_base.base, src0);
773 break;
774
775 case TGSI_OPCODE_BRA:
776 /* deprecated */
777 assert(0);
778 return FALSE;
779 break;
780
781 case TGSI_OPCODE_CAL:
782 return FALSE;
783
784 case TGSI_OPCODE_RET:
785 return FALSE;
786
787 case TGSI_OPCODE_END:
788 *pc = -1;
789 break;
790
791 case TGSI_OPCODE_SSG:
792 /* TGSI_OPCODE_SGN */
793 tmp0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
794 dst0 = lp_build_sgn(&bld->bld_base.base, tmp0);
795 break;
796
797 case TGSI_OPCODE_CMP:
798 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
799 src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL);
800 src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL);
801 tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, bld->bld_base.base.zero);
802 dst0 = lp_build_select(&bld->bld_base.base, tmp0, src1, src2);
803 break;
804
805 case TGSI_OPCODE_SCS:
806 return FALSE;
807
808 case TGSI_OPCODE_TXB:
809 dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS);
810 break;
811
812 case TGSI_OPCODE_NRM:
813 /* fall-through */
814 case TGSI_OPCODE_NRM4:
815 return FALSE;
816
817 case TGSI_OPCODE_DIV:
818 /* deprecated */
819 assert(0);
820 return FALSE;
821 break;
822
823 case TGSI_OPCODE_DP2:
824 return FALSE;
825
826 case TGSI_OPCODE_TXL:
827 dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD);
828 break;
829
830 case TGSI_OPCODE_TXP:
831 dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED);
832 break;
833
834 case TGSI_OPCODE_BRK:
835 return FALSE;
836
837 case TGSI_OPCODE_IF:
838 case TGSI_OPCODE_UIF:
839 return FALSE;
840
841 case TGSI_OPCODE_BGNLOOP:
842 return FALSE;
843
844 case TGSI_OPCODE_BGNSUB:
845 return FALSE;
846
847 case TGSI_OPCODE_ELSE:
848 return FALSE;
849
850 case TGSI_OPCODE_ENDIF:
851 return FALSE;
852
853 case TGSI_OPCODE_ENDLOOP:
854 return FALSE;
855
856 case TGSI_OPCODE_ENDSUB:
857 return FALSE;
858
859 case TGSI_OPCODE_PUSHA:
860 /* deprecated? */
861 assert(0);
862 return FALSE;
863 break;
864
865 case TGSI_OPCODE_POPA:
866 /* deprecated? */
867 assert(0);
868 return FALSE;
869 break;
870
871 case TGSI_OPCODE_CEIL:
872 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
873 dst0 = lp_build_ceil(&bld->bld_base.base, src0);
874 break;
875
876 case TGSI_OPCODE_I2F:
877 /* deprecated? */
878 assert(0);
879 return FALSE;
880 break;
881
882 case TGSI_OPCODE_NOT:
883 /* deprecated? */
884 assert(0);
885 return FALSE;
886 break;
887
888 case TGSI_OPCODE_TRUNC:
889 src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
890 dst0 = lp_build_trunc(&bld->bld_base.base, src0);
891 break;
892
893 case TGSI_OPCODE_SHL:
894 /* deprecated? */
895 assert(0);
896 return FALSE;
897 break;
898
899 case TGSI_OPCODE_ISHR:
900 /* deprecated? */
901 assert(0);
902 return FALSE;
903 break;
904
905 case TGSI_OPCODE_AND:
906 /* deprecated? */
907 assert(0);
908 return FALSE;
909 break;
910
911 case TGSI_OPCODE_OR:
912 /* deprecated? */
913 assert(0);
914 return FALSE;
915 break;
916
917 case TGSI_OPCODE_MOD:
918 /* deprecated? */
919 assert(0);
920 return FALSE;
921 break;
922
923 case TGSI_OPCODE_XOR:
924 /* deprecated? */
925 assert(0);
926 return FALSE;
927 break;
928
929 case TGSI_OPCODE_SAD:
930 /* deprecated? */
931 assert(0);
932 return FALSE;
933 break;
934
935 case TGSI_OPCODE_TXF:
936 /* deprecated? */
937 assert(0);
938 return FALSE;
939 break;
940
941 case TGSI_OPCODE_TXQ:
942 /* deprecated? */
943 assert(0);
944 return FALSE;
945 break;
946
947 case TGSI_OPCODE_CONT:
948 return FALSE;
949
950 case TGSI_OPCODE_EMIT:
951 return FALSE;
952 break;
953
954 case TGSI_OPCODE_ENDPRIM:
955 return FALSE;
956 break;
957
958 case TGSI_OPCODE_NOP:
959 break;
960
961 default:
962 return FALSE;
963 }
964
965 if (info->num_dst) {
966 lp_emit_store_aos(bld, inst, 0, dst0);
967 }
968
969 return TRUE;
970 }
971
972
973 void
974 lp_build_tgsi_aos(struct gallivm_state *gallivm,
975 const struct tgsi_token *tokens,
976 struct lp_type type,
977 const unsigned char swizzles[4],
978 LLVMValueRef consts_ptr,
979 const LLVMValueRef *inputs,
980 LLVMValueRef *outputs,
981 struct lp_build_sampler_aos *sampler,
982 const struct tgsi_shader_info *info)
983 {
984 struct lp_build_tgsi_aos_context bld;
985 struct tgsi_parse_context parse;
986 uint num_immediates = 0;
987 unsigned chan;
988 int pc = 0;
989
990 /* Setup build context */
991 memset(&bld, 0, sizeof bld);
992 lp_build_context_init(&bld.bld_base.base, gallivm, type);
993 lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
994 lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
995 lp_build_context_init(&bld.int_bld, gallivm, lp_int_type(type));
996
997 for (chan = 0; chan < 4; ++chan) {
998 bld.swizzles[chan] = swizzles[chan];
999 bld.inv_swizzles[swizzles[chan]] = chan;
1000 }
1001
1002 bld.inputs = inputs;
1003 bld.outputs = outputs;
1004 bld.consts_ptr = consts_ptr;
1005 bld.sampler = sampler;
1006 bld.indirect_files = info->indirect_files;
1007 bld.bld_base.emit_swizzle = swizzle_aos;
1008 bld.bld_base.info = info;
1009
1010 bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
1011 bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
1012 bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
1013 bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
1014
1015 /* Set opcode actions */
1016 lp_set_default_actions_cpu(&bld.bld_base);
1017
1018 if (!lp_bld_tgsi_list_init(&bld.bld_base)) {
1019 return;
1020 }
1021
1022 tgsi_parse_init(&parse, tokens);
1023
1024 while (!tgsi_parse_end_of_tokens(&parse)) {
1025 tgsi_parse_token(&parse);
1026
1027 switch(parse.FullToken.Token.Type) {
1028 case TGSI_TOKEN_TYPE_DECLARATION:
1029 /* Inputs already interpolated */
1030 lp_emit_declaration_aos(&bld, &parse.FullToken.FullDeclaration);
1031 break;
1032
1033 case TGSI_TOKEN_TYPE_INSTRUCTION:
1034 /* save expanded instruction */
1035 lp_bld_tgsi_add_instruction(&bld.bld_base,
1036 &parse.FullToken.FullInstruction);
1037 break;
1038
1039 case TGSI_TOKEN_TYPE_IMMEDIATE:
1040 /* simply copy the immediate values into the next immediates[] slot */
1041 {
1042 const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
1043 float imm[4];
1044 assert(size <= 4);
1045 assert(num_immediates < LP_MAX_TGSI_IMMEDIATES);
1046 for (chan = 0; chan < 4; ++chan) {
1047 imm[chan] = 0.0f;
1048 }
1049 for (chan = 0; chan < size; ++chan) {
1050 unsigned swizzle = bld.swizzles[chan];
1051 imm[swizzle] = parse.FullToken.FullImmediate.u[chan].Float;
1052 }
1053 bld.immediates[num_immediates] =
1054 lp_build_const_aos(gallivm, type,
1055 imm[0], imm[1], imm[2], imm[3],
1056 NULL);
1057 num_immediates++;
1058 }
1059 break;
1060
1061 case TGSI_TOKEN_TYPE_PROPERTY:
1062 break;
1063
1064 default:
1065 assert(0);
1066 }
1067 }
1068
1069 while (pc != -1) {
1070 struct tgsi_full_instruction *instr = bld.bld_base.instructions + pc;
1071 const struct tgsi_opcode_info *opcode_info =
1072 tgsi_get_opcode_info(instr->Instruction.Opcode);
1073 if (!lp_emit_instruction_aos(&bld, instr, opcode_info, &pc))
1074 _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1075 opcode_info->mnemonic);
1076 }
1077
1078 if (0) {
1079 LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
1080 LLVMValueRef function = LLVMGetBasicBlockParent(block);
1081 debug_printf("11111111111111111111111111111 \n");
1082 tgsi_dump(tokens, 0);
1083 lp_debug_dump_value(function);
1084 debug_printf("2222222222222222222222222222 \n");
1085 }
1086 tgsi_parse_free(&parse);
1087 FREE(bld.bld_base.instructions);
1088
1089 if (0) {
1090 LLVMModuleRef module = LLVMGetGlobalParent(
1091 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
1092 LLVMDumpModule(module);
1093 }
1094
1095 }
1096