projects / mesa.git / blob
? search:


f64185910369c0de500341dd84ae6d4a22ed0482
[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_exec.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_tgsi_action.h"
51 #include "lp_bld_type.h"
52 #include "lp_bld_const.h"
53 #include "lp_bld_arit.h"
54 #include "lp_bld_bitarit.h"
55 #include "lp_bld_gather.h"
56 #include "lp_bld_init.h"
57 #include "lp_bld_logic.h"
58 #include "lp_bld_swizzle.h"
59 #include "lp_bld_flow.h"
60 #include "lp_bld_quad.h"
61 #include "lp_bld_tgsi.h"
62 #include "lp_bld_limits.h"
63 #include "lp_bld_debug.h"
64 #include "lp_bld_printf.h"
65 #include "lp_bld_sample.h"
66 #include "lp_bld_struct.h"
67
68 #define DUMP_GS_EMITS 0
69
70 static void lp_exec_mask_init(struct lp_exec_mask *mask, struct lp_build_context *bld)
71 {
72 LLVMTypeRef int_type = LLVMInt32TypeInContext(bld->gallivm->context);
73 LLVMBuilderRef builder = bld->gallivm->builder;
74
75 mask->bld = bld;
76 mask->has_mask = FALSE;
77 mask->ret_in_main = FALSE;
78 mask->cond_stack_size = 0;
79 mask->loop_stack_size = 0;
80 mask->call_stack_size = 0;
81 mask->switch_stack_size = 0;
82
83 mask->int_vec_type = lp_build_int_vec_type(bld->gallivm, mask->bld->type);
84 mask->exec_mask = mask->ret_mask = mask->break_mask = mask->cont_mask =
85 mask->cond_mask = mask->switch_mask =
86 LLVMConstAllOnes(mask->int_vec_type);
87
88 mask->loop_limiter = lp_build_alloca(bld->gallivm, int_type, "looplimiter");
89
90 LLVMBuildStore(
91 builder,
92 LLVMConstInt(int_type, LP_MAX_TGSI_LOOP_ITERATIONS, false),
93 mask->loop_limiter);
94 }
95
96 static void lp_exec_mask_update(struct lp_exec_mask *mask)
97 {
98 LLVMBuilderRef builder = mask->bld->gallivm->builder;
99
100 if (mask->loop_stack_size) {
101 /*for loops we need to update the entire mask at runtime */
102 LLVMValueRef tmp;
103 assert(mask->break_mask);
104 tmp = LLVMBuildAnd(builder,
105 mask->cont_mask,
106 mask->break_mask,
107 "maskcb");
108 mask->exec_mask = LLVMBuildAnd(builder,
109 mask->cond_mask,
110 tmp,
111 "maskfull");
112 } else
113 mask->exec_mask = mask->cond_mask;
114
115 if (mask->switch_stack_size) {
116 mask->exec_mask = LLVMBuildAnd(builder,
117 mask->exec_mask,
118 mask->switch_mask,
119 "switchmask");
120 }
121
122 if (mask->call_stack_size || mask->ret_in_main) {
123 mask->exec_mask = LLVMBuildAnd(builder,
124 mask->exec_mask,
125 mask->ret_mask,
126 "callmask");
127 }
128
129 mask->has_mask = (mask->cond_stack_size > 0 ||
130 mask->loop_stack_size > 0 ||
131 mask->call_stack_size > 0 ||
132 mask->switch_stack_size > 0 ||
133 mask->ret_in_main);
134 }
135
136 static void lp_exec_mask_cond_push(struct lp_exec_mask *mask,
137 LLVMValueRef val)
138 {
139 LLVMBuilderRef builder = mask->bld->gallivm->builder;
140
141 assert(mask->cond_stack_size < LP_MAX_TGSI_NESTING);
142 if (mask->cond_stack_size == 0) {
143 assert(mask->cond_mask == LLVMConstAllOnes(mask->int_vec_type));
144 }
145 mask->cond_stack[mask->cond_stack_size++] = mask->cond_mask;
146 assert(LLVMTypeOf(val) == mask->int_vec_type);
147 mask->cond_mask = LLVMBuildAnd(builder,
148 mask->cond_mask,
149 val,
150 "");
151 lp_exec_mask_update(mask);
152 }
153
154 static void lp_exec_mask_cond_invert(struct lp_exec_mask *mask)
155 {
156 LLVMBuilderRef builder = mask->bld->gallivm->builder;
157 LLVMValueRef prev_mask;
158 LLVMValueRef inv_mask;
159
160 assert(mask->cond_stack_size);
161 prev_mask = mask->cond_stack[mask->cond_stack_size - 1];
162 if (mask->cond_stack_size == 1) {
163 assert(prev_mask == LLVMConstAllOnes(mask->int_vec_type));
164 }
165
166 inv_mask = LLVMBuildNot(builder, mask->cond_mask, "");
167
168 mask->cond_mask = LLVMBuildAnd(builder,
169 inv_mask,
170 prev_mask, "");
171 lp_exec_mask_update(mask);
172 }
173
174 static void lp_exec_mask_cond_pop(struct lp_exec_mask *mask)
175 {
176 assert(mask->cond_stack_size);
177 mask->cond_mask = mask->cond_stack[--mask->cond_stack_size];
178 lp_exec_mask_update(mask);
179 }
180
181 static void lp_exec_bgnloop(struct lp_exec_mask *mask)
182 {
183 LLVMBuilderRef builder = mask->bld->gallivm->builder;
184
185 if (mask->loop_stack_size == 0) {
186 assert(mask->loop_block == NULL);
187 assert(mask->cont_mask == LLVMConstAllOnes(mask->int_vec_type));
188 assert(mask->break_mask == LLVMConstAllOnes(mask->int_vec_type));
189 assert(mask->break_var == NULL);
190 }
191
192 assert(mask->loop_stack_size < LP_MAX_TGSI_NESTING);
193
194 mask->break_type_stack[mask->loop_stack_size + mask->switch_stack_size] =
195 mask->break_type;
196 mask->break_type = LP_EXEC_MASK_BREAK_TYPE_LOOP;
197
198 mask->loop_stack[mask->loop_stack_size].loop_block = mask->loop_block;
199 mask->loop_stack[mask->loop_stack_size].cont_mask = mask->cont_mask;
200 mask->loop_stack[mask->loop_stack_size].break_mask = mask->break_mask;
201 mask->loop_stack[mask->loop_stack_size].break_var = mask->break_var;
202 ++mask->loop_stack_size;
203
204 mask->break_var = lp_build_alloca(mask->bld->gallivm, mask->int_vec_type, "");
205 LLVMBuildStore(builder, mask->break_mask, mask->break_var);
206
207 mask->loop_block = lp_build_insert_new_block(mask->bld->gallivm, "bgnloop");
208
209 LLVMBuildBr(builder, mask->loop_block);
210 LLVMPositionBuilderAtEnd(builder, mask->loop_block);
211
212 mask->break_mask = LLVMBuildLoad(builder, mask->break_var, "");
213
214 lp_exec_mask_update(mask);
215 }
216
217 static void lp_exec_break(struct lp_exec_mask *mask,
218 struct lp_build_tgsi_context * bld_base)
219 {
220 LLVMBuilderRef builder = mask->bld->gallivm->builder;
221
222 if (mask->break_type == LP_EXEC_MASK_BREAK_TYPE_LOOP) {
223 LLVMValueRef exec_mask = LLVMBuildNot(builder,
224 mask->exec_mask,
225 "break");
226
227 mask->break_mask = LLVMBuildAnd(builder,
228 mask->break_mask,
229 exec_mask, "break_full");
230 }
231 else {
232 unsigned opcode = bld_base->instructions[bld_base->pc + 1].Instruction.Opcode;
233 boolean break_always = (opcode == TGSI_OPCODE_ENDSWITCH ||
234 opcode == TGSI_OPCODE_CASE);
235
236
237 if (mask->switch_in_default) {
238 /*
239 * stop default execution but only if this is an unconditional switch.
240 * (The condition here is not perfect since dead code after break is
241 * allowed but should be sufficient since false negatives are just
242 * unoptimized - so we don't have to pre-evaluate that).
243 */
244 if(break_always && mask->switch_pc) {
245 bld_base->pc = mask->switch_pc;
246 return;
247 }
248 }
249
250 if (break_always) {
251 mask->switch_mask = LLVMConstNull(mask->bld->int_vec_type);
252 }
253 else {
254 LLVMValueRef exec_mask = LLVMBuildNot(builder,
255 mask->exec_mask,
256 "break");
257 mask->switch_mask = LLVMBuildAnd(builder,
258 mask->switch_mask,
259 exec_mask, "break_switch");
260 }
261 }
262
263 lp_exec_mask_update(mask);
264 }
265
266 static void lp_exec_break_condition(struct lp_exec_mask *mask,
267 LLVMValueRef cond)
268 {
269 LLVMBuilderRef builder = mask->bld->gallivm->builder;
270 LLVMValueRef cond_mask = LLVMBuildAnd(builder,
271 mask->exec_mask,
272 cond, "cond_mask");
273 cond_mask = LLVMBuildNot(builder, cond_mask, "break_cond");
274
275 if (mask->break_type == LP_EXEC_MASK_BREAK_TYPE_LOOP) {
276 mask->break_mask = LLVMBuildAnd(builder,
277 mask->break_mask,
278 cond_mask, "breakc_full");
279 }
280 else {
281 mask->switch_mask = LLVMBuildAnd(builder,
282 mask->switch_mask,
283 cond_mask, "breakc_switch");
284 }
285
286 lp_exec_mask_update(mask);
287 }
288
289 static void lp_exec_continue(struct lp_exec_mask *mask)
290 {
291 LLVMBuilderRef builder = mask->bld->gallivm->builder;
292 LLVMValueRef exec_mask = LLVMBuildNot(builder,
293 mask->exec_mask,
294 "");
295
296 mask->cont_mask = LLVMBuildAnd(builder,
297 mask->cont_mask,
298 exec_mask, "");
299
300 lp_exec_mask_update(mask);
301 }
302
303
304 static void lp_exec_endloop(struct gallivm_state *gallivm,
305 struct lp_exec_mask *mask)
306 {
307 LLVMBuilderRef builder = mask->bld->gallivm->builder;
308 LLVMBasicBlockRef endloop;
309 LLVMTypeRef int_type = LLVMInt32TypeInContext(mask->bld->gallivm->context);
310 LLVMTypeRef reg_type = LLVMIntTypeInContext(gallivm->context,
311 mask->bld->type.width *
312 mask->bld->type.length);
313 LLVMValueRef i1cond, i2cond, icond, limiter;
314
315 assert(mask->break_mask);
316
317 /*
318 * Restore the cont_mask, but don't pop
319 */
320 assert(mask->loop_stack_size);
321 mask->cont_mask = mask->loop_stack[mask->loop_stack_size - 1].cont_mask;
322 lp_exec_mask_update(mask);
323
324 /*
325 * Unlike the continue mask, the break_mask must be preserved across loop
326 * iterations
327 */
328 LLVMBuildStore(builder, mask->break_mask, mask->break_var);
329
330 /* Decrement the loop limiter */
331 limiter = LLVMBuildLoad(builder, mask->loop_limiter, "");
332
333 limiter = LLVMBuildSub(
334 builder,
335 limiter,
336 LLVMConstInt(int_type, 1, false),
337 "");
338
339 LLVMBuildStore(builder, limiter, mask->loop_limiter);
340
341 /* i1cond = (mask != 0) */
342 i1cond = LLVMBuildICmp(
343 builder,
344 LLVMIntNE,
345 LLVMBuildBitCast(builder, mask->exec_mask, reg_type, ""),
346 LLVMConstNull(reg_type), "i1cond");
347
348 /* i2cond = (looplimiter > 0) */
349 i2cond = LLVMBuildICmp(
350 builder,
351 LLVMIntSGT,
352 limiter,
353 LLVMConstNull(int_type), "i2cond");
354
355 /* if( i1cond && i2cond ) */
356 icond = LLVMBuildAnd(builder, i1cond, i2cond, "");
357
358 endloop = lp_build_insert_new_block(mask->bld->gallivm, "endloop");
359
360 LLVMBuildCondBr(builder,
361 icond, mask->loop_block, endloop);
362
363 LLVMPositionBuilderAtEnd(builder, endloop);
364
365 assert(mask->loop_stack_size);
366 --mask->loop_stack_size;
367 mask->loop_block = mask->loop_stack[mask->loop_stack_size].loop_block;
368 mask->cont_mask = mask->loop_stack[mask->loop_stack_size].cont_mask;
369 mask->break_mask = mask->loop_stack[mask->loop_stack_size].break_mask;
370 mask->break_var = mask->loop_stack[mask->loop_stack_size].break_var;
371 mask->break_type = mask->break_type_stack[mask->loop_stack_size + mask->switch_stack_size];
372
373 lp_exec_mask_update(mask);
374 }
375
376 static void lp_exec_switch(struct lp_exec_mask *mask,
377 LLVMValueRef switchval)
378 {
379 mask->break_type_stack[mask->loop_stack_size + mask->switch_stack_size] =
380 mask->break_type;
381 mask->break_type = LP_EXEC_MASK_BREAK_TYPE_SWITCH;
382
383 mask->switch_stack[mask->switch_stack_size].switch_val = mask->switch_val;
384 mask->switch_stack[mask->switch_stack_size].switch_mask = mask->switch_mask;
385 mask->switch_stack[mask->switch_stack_size].switch_mask_default = mask->switch_mask_default;
386 mask->switch_stack[mask->switch_stack_size].switch_in_default = mask->switch_in_default;
387 mask->switch_stack[mask->switch_stack_size].switch_pc = mask->switch_pc;
388 mask->switch_stack_size++;
389
390 mask->switch_val = switchval;
391 mask->switch_mask = LLVMConstNull(mask->int_vec_type);
392 mask->switch_mask_default = LLVMConstNull(mask->int_vec_type);
393 mask->switch_in_default = false;
394 mask->switch_pc = 0;
395
396 lp_exec_mask_update(mask);
397 }
398
399 static void lp_exec_endswitch(struct lp_exec_mask *mask,
400 struct lp_build_tgsi_context * bld_base)
401 {
402 LLVMBuilderRef builder = mask->bld->gallivm->builder;
403
404 /* check if there's deferred default if so do it now */
405 if (mask->switch_pc && !mask->switch_in_default) {
406 LLVMValueRef prevmask, defaultmask;
407 unsigned tmp_pc;
408 prevmask = mask->switch_stack[mask->switch_stack_size - 1].switch_mask;
409 defaultmask = LLVMBuildNot(builder, mask->switch_mask_default, "sw_default_mask");
410 mask->switch_mask = LLVMBuildAnd(builder, prevmask, defaultmask, "sw_mask");
411 mask->switch_in_default = true;
412
413 lp_exec_mask_update(mask);
414
415 assert(bld_base->instructions[mask->switch_pc - 1].Instruction.Opcode ==
416 TGSI_OPCODE_DEFAULT);
417
418 tmp_pc = bld_base->pc;
419 bld_base->pc = mask->switch_pc;
420 /*
421 * re-purpose switch_pc to point to here again, since we stop execution of
422 * the deferred default after next break.
423 */
424 mask->switch_pc = tmp_pc - 1;
425
426 return;
427 }
428
429 else if (mask->switch_pc && mask->switch_in_default) {
430 assert(bld_base->pc == mask->switch_pc + 1);
431 }
432
433 mask->switch_stack_size--;
434 mask->switch_val = mask->switch_stack[mask->switch_stack_size].switch_val;
435 mask->switch_mask = mask->switch_stack[mask->switch_stack_size].switch_mask;
436 mask->switch_mask_default = mask->switch_stack[mask->switch_stack_size].switch_mask_default;
437 mask->switch_in_default = mask->switch_stack[mask->switch_stack_size].switch_in_default;
438 mask->switch_pc = mask->switch_stack[mask->switch_stack_size].switch_pc;
439
440 mask->break_type = mask->break_type_stack[mask->loop_stack_size + mask->switch_stack_size];
441
442 lp_exec_mask_update(mask);
443 }
444
445 static void lp_exec_case(struct lp_exec_mask *mask,
446 LLVMValueRef caseval)
447 {
448 LLVMBuilderRef builder = mask->bld->gallivm->builder;
449
450 LLVMValueRef casemask, prevmask;
451
452 /* skipping case mask evaluation here is NOT optional (not in all cases anyway). */
453 if (!mask->switch_in_default) {
454 prevmask = mask->switch_stack[mask->switch_stack_size - 1].switch_mask;
455 casemask = lp_build_cmp(mask->bld, PIPE_FUNC_EQUAL, caseval, mask->switch_val);
456 mask->switch_mask_default = LLVMBuildOr(builder, casemask,
457 mask->switch_mask_default, "sw_default_mask");
458 casemask = LLVMBuildOr(builder, casemask, mask->switch_mask, "");
459 mask->switch_mask = LLVMBuildAnd(builder, casemask, prevmask, "sw_mask");
460
461 lp_exec_mask_update(mask);
462 }
463 }
464
465 /*
466 * Analyse default statement in a switch.
467 * \return true if default is last statement, false otherwise
468 * \param default_pc_start contains pc of instruction to jump to
469 * if default wasn't last but there's no
470 * fallthrough into default.
471 */
472 static boolean default_analyse_is_last(struct lp_exec_mask *mask,
473 struct lp_build_tgsi_context * bld_base,
474 int *default_pc_start)
475 {
476 unsigned pc = bld_base->pc;
477 unsigned curr_switch_stack = mask->switch_stack_size;
478
479 /* skip over case statements which are together with default */
480 while (bld_base->instructions[pc].Instruction.Opcode == TGSI_OPCODE_CASE) {
481 pc++;
482 }
483
484 while (pc != -1 && pc < bld_base->num_instructions) {
485 unsigned opcode = bld_base->instructions[pc].Instruction.Opcode;
486 switch (opcode) {
487 case TGSI_OPCODE_CASE:
488 if (curr_switch_stack == mask->switch_stack_size) {
489 *default_pc_start = pc - 1;
490 return false;
491 }
492 break;
493 case TGSI_OPCODE_SWITCH:
494 curr_switch_stack++;
495 break;
496 case TGSI_OPCODE_ENDSWITCH:
497 if (curr_switch_stack == mask->switch_stack_size) {
498 *default_pc_start = pc - 1;
499 return true;
500 }
501 curr_switch_stack--;
502 break;
503 }
504 pc++;
505 }
506 /* should never arrive here */
507 assert(0);
508 return true;
509 }
510
511 static void lp_exec_default(struct lp_exec_mask *mask,
512 struct lp_build_tgsi_context * bld_base)
513 {
514 LLVMBuilderRef builder = mask->bld->gallivm->builder;
515
516 int default_exec_pc;
517 boolean default_is_last;
518
519 /*
520 * This is a messy opcode, because it may not be always at the end and
521 * there can be fallthrough in and out of it.
522 */
523
524 default_is_last = default_analyse_is_last(mask, bld_base, &default_exec_pc);
525 /*
526 * If it is last statement in switch (note that case statements appearing
527 * "at the same time" as default don't change that) everything is just fine,
528 * update switch mask and go on. This means we can handle default with
529 * fallthrough INTO it without overhead, if it is last.
530 */
531 if (default_is_last) {
532 LLVMValueRef prevmask, defaultmask;
533 prevmask = mask->switch_stack[mask->switch_stack_size - 1].switch_mask;
534 defaultmask = LLVMBuildNot(builder, mask->switch_mask_default, "sw_default_mask");
535 defaultmask = LLVMBuildOr(builder, defaultmask, mask->switch_mask, "");
536 mask->switch_mask = LLVMBuildAnd(builder, prevmask, defaultmask, "sw_mask");
537 mask->switch_in_default = true;
538
539 lp_exec_mask_update(mask);
540 }
541 else {
542 /*
543 * Technically, "case" immediately before default isn't really a
544 * fallthrough, however we still have to count them as such as we
545 * already have updated the masks.
546 * If that happens in practice could add a switch optimizer pass
547 * which just gets rid of all case statements appearing together with
548 * default (or could do switch analysis at switch start time instead).
549 */
550 unsigned opcode = bld_base->instructions[bld_base->pc - 1].Instruction.Opcode;
551 boolean ft_into = (opcode != TGSI_OPCODE_BRK ||
552 opcode != TGSI_OPCODE_SWITCH);
553 /*
554 * If it is not last statement and there was no fallthrough into it,
555 * we record the PC and continue execution at next case (again, those
556 * case encountered at the same time don't count). At endswitch
557 * time, we update switchmask, and go back executing the code we skipped
558 * until the next break (possibly re-executing some code with changed mask
559 * if there was a fallthrough out of default).
560 * Finally, if it is not last statement and there was a fallthrough into it,
561 * do the same as with the former case, except instead of skipping the code
562 * just execute it without updating the mask, then go back and re-execute.
563 */
564 mask->switch_pc = bld_base->pc;
565 if (!ft_into) {
566 bld_base->pc = default_exec_pc;
567 }
568 }
569 }
570
571
572 /* stores val into an address pointed to by dst_ptr.
573 * mask->exec_mask is used to figure out which bits of val
574 * should be stored into the address
575 * (0 means don't store this bit, 1 means do store).
576 */
577 static void lp_exec_mask_store(struct lp_exec_mask *mask,
578 struct lp_build_context *bld_store,
579 LLVMValueRef pred,
580 LLVMValueRef val,
581 LLVMValueRef dst_ptr)
582 {
583 LLVMBuilderRef builder = mask->bld->gallivm->builder;
584
585 assert(lp_check_value(bld_store->type, val));
586 assert(LLVMGetTypeKind(LLVMTypeOf(dst_ptr)) == LLVMPointerTypeKind);
587 assert(LLVMGetElementType(LLVMTypeOf(dst_ptr)) == LLVMTypeOf(val));
588
589 /* Mix the predicate and execution mask */
590 if (mask->has_mask) {
591 if (pred) {
592 pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
593 } else {
594 pred = mask->exec_mask;
595 }
596 }
597
598 if (pred) {
599 LLVMValueRef res, dst;
600
601 dst = LLVMBuildLoad(builder, dst_ptr, "");
602 res = lp_build_select(bld_store, pred, val, dst);
603 LLVMBuildStore(builder, res, dst_ptr);
604 } else
605 LLVMBuildStore(builder, val, dst_ptr);
606 }
607
608 static void lp_exec_mask_call(struct lp_exec_mask *mask,
609 int func,
610 int *pc)
611 {
612 assert(mask->call_stack_size < LP_MAX_TGSI_NESTING);
613 mask->call_stack[mask->call_stack_size].pc = *pc;
614 mask->call_stack[mask->call_stack_size].ret_mask = mask->ret_mask;
615 mask->call_stack_size++;
616 *pc = func;
617 }
618
619 static void lp_exec_mask_ret(struct lp_exec_mask *mask, int *pc)
620 {
621 LLVMBuilderRef builder = mask->bld->gallivm->builder;
622 LLVMValueRef exec_mask;
623
624 if (mask->cond_stack_size == 0 &&
625 mask->loop_stack_size == 0 &&
626 mask->switch_stack_size == 0 &&
627 mask->call_stack_size == 0) {
628 /* returning from main() */
629 *pc = -1;
630 return;
631 }
632
633 if (mask->call_stack_size == 0) {
634 /*
635 * This requires special handling since we need to ensure
636 * we don't drop the mask even if we have no call stack
637 * (e.g. after a ret in a if clause after the endif)
638 */
639 mask->ret_in_main = TRUE;
640 }
641
642 exec_mask = LLVMBuildNot(builder,
643 mask->exec_mask,
644 "ret");
645
646 mask->ret_mask = LLVMBuildAnd(builder,
647 mask->ret_mask,
648 exec_mask, "ret_full");
649
650 lp_exec_mask_update(mask);
651 }
652
653 static void lp_exec_mask_bgnsub(struct lp_exec_mask *mask)
654 {
655 }
656
657 static void lp_exec_mask_endsub(struct lp_exec_mask *mask, int *pc)
658 {
659 assert(mask->call_stack_size);
660 mask->call_stack_size--;
661 *pc = mask->call_stack[mask->call_stack_size].pc;
662 mask->ret_mask = mask->call_stack[mask->call_stack_size].ret_mask;
663 lp_exec_mask_update(mask);
664 }
665
666
667 /**
668 * Return pointer to a temporary register channel (src or dest).
669 * Note that indirect addressing cannot be handled here.
670 * \param index which temporary register
671 * \param chan which channel of the temp register.
672 */
673 LLVMValueRef
674 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context *bld,
675 unsigned index,
676 unsigned chan)
677 {
678 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
679 assert(chan < 4);
680 if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
681 LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm, index * 4 + chan);
682 return LLVMBuildGEP(builder, bld->temps_array, &lindex, 1, "");
683 }
684 else {
685 return bld->temps[index][chan];
686 }
687 }
688
689 /**
690 * Return pointer to a output register channel (src or dest).
691 * Note that indirect addressing cannot be handled here.
692 * \param index which output register
693 * \param chan which channel of the output register.
694 */
695 LLVMValueRef
696 lp_get_output_ptr(struct lp_build_tgsi_soa_context *bld,
697 unsigned index,
698 unsigned chan)
699 {
700 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
701 assert(chan < 4);
702 if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
703 LLVMValueRef lindex = lp_build_const_int32(bld->bld_base.base.gallivm,
704 index * 4 + chan);
705 return LLVMBuildGEP(builder, bld->outputs_array, &lindex, 1, "");
706 }
707 else {
708 return bld->outputs[index][chan];
709 }
710 }
711
712 /*
713 * If we have indirect addressing in outputs copy our alloca array
714 * to the outputs slots specified by the caller to make sure
715 * our outputs are delivered consistently via the same interface.
716 */
717 static void
718 gather_outputs(struct lp_build_tgsi_soa_context * bld)
719 {
720 if ((bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
721 unsigned index, chan;
722 assert(bld->bld_base.info->num_outputs <=
723 bld->bld_base.info->file_max[TGSI_FILE_OUTPUT] + 1);
724 for (index = 0; index < bld->bld_base.info->num_outputs; ++index) {
725 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
726 bld->outputs[index][chan] = lp_get_output_ptr(bld, index, chan);
727 }
728 }
729 }
730 }
731
732 /**
733 * Gather vector.
734 * XXX the lp_build_gather() function should be capable of doing this
735 * with a little work.
736 */
737 static LLVMValueRef
738 build_gather(struct lp_build_context *bld,
739 LLVMValueRef base_ptr,
740 LLVMValueRef indexes)
741 {
742 LLVMBuilderRef builder = bld->gallivm->builder;
743 LLVMValueRef res = bld->undef;
744 unsigned i;
745
746 /*
747 * Loop over elements of index_vec, load scalar value, insert it into 'res'.
748 */
749 for (i = 0; i < bld->type.length; i++) {
750 LLVMValueRef ii = lp_build_const_int32(bld->gallivm, i);
751 LLVMValueRef index = LLVMBuildExtractElement(builder,
752 indexes, ii, "");
753 LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr,
754 &index, 1, "gather_ptr");
755 LLVMValueRef scalar = LLVMBuildLoad(builder, scalar_ptr, "");
756
757 res = LLVMBuildInsertElement(builder, res, scalar, ii, "");
758 }
759
760 return res;
761 }
762
763
764 /**
765 * Scatter/store vector.
766 */
767 static void
768 emit_mask_scatter(struct lp_build_tgsi_soa_context *bld,
769 LLVMValueRef base_ptr,
770 LLVMValueRef indexes,
771 LLVMValueRef values,
772 struct lp_exec_mask *mask,
773 LLVMValueRef pred)
774 {
775 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
776 LLVMBuilderRef builder = gallivm->builder;
777 unsigned i;
778
779 /* Mix the predicate and execution mask */
780 if (mask->has_mask) {
781 if (pred) {
782 pred = LLVMBuildAnd(builder, pred, mask->exec_mask, "");
783 }
784 else {
785 pred = mask->exec_mask;
786 }
787 }
788
789 /*
790 * Loop over elements of index_vec, store scalar value.
791 */
792 for (i = 0; i < bld->bld_base.base.type.length; i++) {
793 LLVMValueRef ii = lp_build_const_int32(gallivm, i);
794 LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
795 LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
796 LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
797 LLVMValueRef scalar_pred = pred ?
798 LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
799
800 if (0)
801 lp_build_printf(gallivm, "scatter %d: val %f at %d %p\n",
802 ii, val, index, scalar_ptr);
803
804 if (scalar_pred) {
805 LLVMValueRef real_val, dst_val;
806 dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
807 real_val = lp_build_select(&bld->elem_bld, scalar_pred, val, dst_val);
808 LLVMBuildStore(builder, real_val, scalar_ptr);
809 }
810 else {
811 LLVMBuildStore(builder, val, scalar_ptr);
812 }
813 }
814 }
815
816
817 /**
818 * Read the current value of the ADDR register, convert the floats to
819 * ints, add the base index and return the vector of offsets.
820 * The offsets will be used to index into the constant buffer or
821 * temporary register file.
822 */
823 static LLVMValueRef
824 get_indirect_index(struct lp_build_tgsi_soa_context *bld,
825 unsigned reg_file, unsigned reg_index,
826 const struct tgsi_ind_register *indirect_reg)
827 {
828 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
829 struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
830 /* always use X component of address register */
831 unsigned swizzle = indirect_reg->Swizzle;
832 LLVMValueRef base;
833 LLVMValueRef rel;
834 LLVMValueRef max_index;
835 LLVMValueRef index;
836
837 assert(bld->indirect_files & (1 << reg_file));
838
839 base = lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, reg_index);
840
841 assert(swizzle < 4);
842 switch (indirect_reg->File) {
843 case TGSI_FILE_ADDRESS:
844 rel = LLVMBuildLoad(builder,
845 bld->addr[indirect_reg->Index][swizzle],
846 "load addr reg");
847 /* ADDR LLVM values already have LLVM integer type. */
848 break;
849 case TGSI_FILE_TEMPORARY:
850 rel = lp_get_temp_ptr_soa(bld, indirect_reg->Index, swizzle);
851 rel = LLVMBuildLoad(builder, rel, "load temp reg");
852 /* TEMP LLVM values always have LLVM float type, but for indirection, the
853 * value actually stored is expected to be an integer */
854 rel = LLVMBuildBitCast(builder, rel, uint_bld->vec_type, "");
855 break;
856 default:
857 assert(0);
858 rel = uint_bld->zero;
859 }
860
861 index = lp_build_add(uint_bld, base, rel);
862
863 max_index = lp_build_const_int_vec(bld->bld_base.base.gallivm,
864 uint_bld->type,
865 bld->bld_base.info->file_max[reg_file]);
866
867 assert(!uint_bld->type.sign);
868 index = lp_build_min(uint_bld, index, max_index);
869
870 return index;
871 }
872
873 static struct lp_build_context *
874 stype_to_fetch(struct lp_build_tgsi_context * bld_base,
875 enum tgsi_opcode_type stype)
876 {
877 struct lp_build_context *bld_fetch;
878
879 switch (stype) {
880 case TGSI_TYPE_FLOAT:
881 case TGSI_TYPE_UNTYPED:
882 bld_fetch = &bld_base->base;
883 break;
884 case TGSI_TYPE_UNSIGNED:
885 bld_fetch = &bld_base->uint_bld;
886 break;
887 case TGSI_TYPE_SIGNED:
888 bld_fetch = &bld_base->int_bld;
889 break;
890 case TGSI_TYPE_VOID:
891 case TGSI_TYPE_DOUBLE:
892 default:
893 assert(0);
894 bld_fetch = NULL;
895 break;
896 }
897 return bld_fetch;
898 }
899
900 static LLVMValueRef
901 emit_fetch_constant(
902 struct lp_build_tgsi_context * bld_base,
903 const struct tgsi_full_src_register * reg,
904 enum tgsi_opcode_type stype,
905 unsigned swizzle)
906 {
907 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
908 struct gallivm_state *gallivm = bld_base->base.gallivm;
909 LLVMBuilderRef builder = gallivm->builder;
910 struct lp_build_context *uint_bld = &bld_base->uint_bld;
911 LLVMValueRef indirect_index = NULL;
912 unsigned dimension = 0;
913 LLVMValueRef dimension_index;
914 LLVMValueRef consts_ptr;
915 LLVMValueRef res;
916
917 /* XXX: Handle fetching xyzw components as a vector */
918 assert(swizzle != ~0);
919
920 if (reg->Register.Dimension) {
921 assert(!reg->Dimension.Indirect);
922 dimension = reg->Dimension.Index;
923 assert(dimension < LP_MAX_TGSI_CONST_BUFFERS);
924 }
925
926 dimension_index = lp_build_const_int32(gallivm, dimension);
927 consts_ptr = lp_build_array_get(gallivm, bld->consts_ptr, dimension_index);
928
929 if (reg->Register.Indirect) {
930 indirect_index = get_indirect_index(bld,
931 reg->Register.File,
932 reg->Register.Index,
933 &reg->Indirect);
934 }
935
936 if (reg->Register.Indirect) {
937 LLVMValueRef swizzle_vec =
938 lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
939 LLVMValueRef index_vec; /* index into the const buffer */
940
941 /* index_vec = indirect_index * 4 + swizzle */
942 index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
943 index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
944
945 /* Gather values from the constant buffer */
946 res = build_gather(&bld_base->base, consts_ptr, index_vec);
947 }
948 else {
949 LLVMValueRef index; /* index into the const buffer */
950 LLVMValueRef scalar, scalar_ptr;
951
952 index = lp_build_const_int32(gallivm, reg->Register.Index*4 + swizzle);
953
954 scalar_ptr = LLVMBuildGEP(builder, consts_ptr,
955 &index, 1, "");
956 scalar = LLVMBuildLoad(builder, scalar_ptr, "");
957 res = lp_build_broadcast_scalar(&bld_base->base, scalar);
958 }
959
960 if (stype == TGSI_TYPE_SIGNED || stype == TGSI_TYPE_UNSIGNED) {
961 struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
962 res = LLVMBuildBitCast(builder, res, bld_fetch->vec_type, "");
963 }
964 return res;
965 }
966
967 static LLVMValueRef
968 emit_fetch_immediate(
969 struct lp_build_tgsi_context * bld_base,
970 const struct tgsi_full_src_register * reg,
971 enum tgsi_opcode_type stype,
972 unsigned swizzle)
973 {
974 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
975 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
976 LLVMBuilderRef builder = gallivm->builder;
977 struct lp_build_context *uint_bld = &bld_base->uint_bld;
978 struct lp_build_context *float_bld = &bld_base->base;
979 LLVMValueRef res = NULL;
980 LLVMValueRef indirect_index = NULL;
981
982 if (reg->Register.Indirect) {
983 indirect_index = get_indirect_index(bld,
984 reg->Register.File,
985 reg->Register.Index,
986 &reg->Indirect);
987 }
988
989 if (reg->Register.Indirect) {
990 LLVMValueRef swizzle_vec =
991 lp_build_const_int_vec(bld->bld_base.base.gallivm,
992 uint_bld->type, swizzle);
993 LLVMValueRef length_vec =
994 lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type,
995 bld->bld_base.base.type.length);
996 LLVMValueRef index_vec; /* index into the const buffer */
997 LLVMValueRef imms_array;
998 LLVMValueRef pixel_offsets;
999 LLVMValueRef offsets[LP_MAX_VECTOR_LENGTH];
1000 LLVMTypeRef float4_ptr_type;
1001 int i;
1002
1003 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1004 for (i = 0; i < float_bld->type.length; i++) {
1005 offsets[i] = lp_build_const_int32(gallivm, i);
1006 }
1007 pixel_offsets = LLVMConstVector(offsets, float_bld->type.length);
1008
1009 /* index_vec = (indirect_index * 4 + swizzle) * length */
1010 index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1011 index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
1012 index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1013 index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1014
1015 /* cast imms_array pointer to float* */
1016 float4_ptr_type = LLVMPointerType(
1017 LLVMFloatTypeInContext(bld->bld_base.base.gallivm->context), 0);
1018 imms_array = LLVMBuildBitCast(builder, bld->imms_array,
1019 float4_ptr_type, "");
1020
1021 /* Gather values from the temporary register array */
1022 res = build_gather(&bld_base->base, imms_array, index_vec);
1023 }
1024 else {
1025 res = bld->immediates[reg->Register.Index][swizzle];
1026 }
1027
1028 if (stype == TGSI_TYPE_UNSIGNED) {
1029 res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1030 } else if (stype == TGSI_TYPE_SIGNED) {
1031 res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1032 }
1033 return res;
1034 }
1035
1036 static LLVMValueRef
1037 emit_fetch_input(
1038 struct lp_build_tgsi_context * bld_base,
1039 const struct tgsi_full_src_register * reg,
1040 enum tgsi_opcode_type stype,
1041 unsigned swizzle)
1042 {
1043 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1044 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1045 LLVMBuilderRef builder = gallivm->builder;
1046 struct lp_build_context *uint_bld = &bld_base->uint_bld;
1047 LLVMValueRef indirect_index = NULL;
1048 LLVMValueRef res;
1049
1050 if (reg->Register.Indirect) {
1051 indirect_index = get_indirect_index(bld,
1052 reg->Register.File,
1053 reg->Register.Index,
1054 &reg->Indirect);
1055 }
1056
1057 if (reg->Register.Indirect) {
1058 LLVMValueRef swizzle_vec =
1059 lp_build_const_int_vec(gallivm, uint_bld->type, swizzle);
1060 LLVMValueRef length_vec =
1061 lp_build_const_int_vec(gallivm, uint_bld->type, bld->bld_base.base.type.length);
1062 LLVMValueRef index_vec; /* index into the const buffer */
1063 LLVMValueRef inputs_array;
1064 LLVMTypeRef float4_ptr_type;
1065
1066 /* index_vec = (indirect_index * 4 + swizzle) * length */
1067 index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1068 index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
1069 index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1070
1071 /* cast inputs_array pointer to float* */
1072 float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1073 inputs_array = LLVMBuildBitCast(builder, bld->inputs_array,
1074 float4_ptr_type, "");
1075
1076 /* Gather values from the temporary register array */
1077 res = build_gather(&bld_base->base, inputs_array, index_vec);
1078 } else {
1079 if (bld->indirect_files & (1 << TGSI_FILE_INPUT)) {
1080 LLVMValueRef lindex = lp_build_const_int32(gallivm,
1081 reg->Register.Index * 4 + swizzle);
1082 LLVMValueRef input_ptr = LLVMBuildGEP(builder,
1083 bld->inputs_array, &lindex, 1, "");
1084 res = LLVMBuildLoad(builder, input_ptr, "");
1085 }
1086 else {
1087 res = bld->inputs[reg->Register.Index][swizzle];
1088 }
1089 }
1090
1091 assert(res);
1092
1093 if (stype == TGSI_TYPE_UNSIGNED) {
1094 res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1095 } else if (stype == TGSI_TYPE_SIGNED) {
1096 res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1097 }
1098
1099 return res;
1100 }
1101
1102
1103 static LLVMValueRef
1104 emit_fetch_gs_input(
1105 struct lp_build_tgsi_context * bld_base,
1106 const struct tgsi_full_src_register * reg,
1107 enum tgsi_opcode_type stype,
1108 unsigned swizzle)
1109 {
1110 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1111 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1112 LLVMBuilderRef builder = gallivm->builder;
1113 LLVMValueRef attrib_index = NULL;
1114 LLVMValueRef vertex_index = NULL;
1115 LLVMValueRef swizzle_index = lp_build_const_int32(gallivm, swizzle);
1116 LLVMValueRef res;
1117
1118 if (reg->Register.Indirect) {
1119 attrib_index = get_indirect_index(bld,
1120 reg->Register.File,
1121 reg->Register.Index,
1122 &reg->Indirect);
1123 } else {
1124 attrib_index = lp_build_const_int32(gallivm, reg->Register.Index);
1125 }
1126
1127 if (reg->Dimension.Indirect) {
1128 vertex_index = get_indirect_index(bld,
1129 reg->Register.File,
1130 reg->Dimension.Index,
1131 &reg->DimIndirect);
1132 } else {
1133 vertex_index = lp_build_const_int32(gallivm, reg->Dimension.Index);
1134 }
1135
1136 res = bld->gs_iface->fetch_input(bld->gs_iface, bld_base,
1137 reg->Dimension.Indirect,
1138 vertex_index, attrib_index,
1139 swizzle_index);
1140
1141 assert(res);
1142
1143 if (stype == TGSI_TYPE_UNSIGNED) {
1144 res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1145 } else if (stype == TGSI_TYPE_SIGNED) {
1146 res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1147 }
1148
1149 return res;
1150 }
1151
1152 static LLVMValueRef
1153 emit_fetch_temporary(
1154 struct lp_build_tgsi_context * bld_base,
1155 const struct tgsi_full_src_register * reg,
1156 enum tgsi_opcode_type stype,
1157 unsigned swizzle)
1158 {
1159 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1160 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1161 LLVMBuilderRef builder = gallivm->builder;
1162 struct lp_build_context *uint_bld = &bld_base->uint_bld;
1163 struct lp_build_context *float_bld = &bld_base->base;
1164 LLVMValueRef indirect_index = NULL;
1165 LLVMValueRef res;
1166
1167 if (reg->Register.Indirect) {
1168 indirect_index = get_indirect_index(bld,
1169 reg->Register.File,
1170 reg->Register.Index,
1171 &reg->Indirect);
1172 }
1173
1174 if (reg->Register.Indirect) {
1175 LLVMValueRef swizzle_vec =
1176 lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type, swizzle);
1177 LLVMValueRef length_vec =
1178 lp_build_const_int_vec(bld->bld_base.base.gallivm, uint_bld->type,
1179 bld->bld_base.base.type.length);
1180 LLVMValueRef index_vec; /* index into the const buffer */
1181 LLVMValueRef temps_array;
1182 LLVMValueRef pixel_offsets;
1183 LLVMValueRef offsets[LP_MAX_VECTOR_LENGTH];
1184 LLVMTypeRef float4_ptr_type;
1185 int i;
1186
1187 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1188 for (i = 0; i < float_bld->type.length; i++) {
1189 offsets[i] = lp_build_const_int32(gallivm, i);
1190 }
1191 pixel_offsets = LLVMConstVector(offsets, float_bld->type.length);
1192
1193 /* index_vec = (indirect_index * 4 + swizzle) * length */
1194 index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1195 index_vec = lp_build_add(uint_bld, index_vec, swizzle_vec);
1196 index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1197 index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1198
1199 /* cast temps_array pointer to float* */
1200 float4_ptr_type = LLVMPointerType(LLVMFloatTypeInContext(bld->bld_base.base.gallivm->context), 0);
1201 temps_array = LLVMBuildBitCast(builder, bld->temps_array,
1202 float4_ptr_type, "");
1203
1204 /* Gather values from the temporary register array */
1205 res = build_gather(&bld_base->base, temps_array, index_vec);
1206 }
1207 else {
1208 LLVMValueRef temp_ptr;
1209 temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
1210 res = LLVMBuildLoad(builder, temp_ptr, "");
1211 }
1212
1213 if (stype == TGSI_TYPE_SIGNED || stype == TGSI_TYPE_UNSIGNED) {
1214 struct lp_build_context *bld_fetch = stype_to_fetch(bld_base, stype);
1215 res = LLVMBuildBitCast(builder, res, bld_fetch->vec_type, "");
1216 }
1217
1218 return res;
1219 }
1220
1221 static LLVMValueRef
1222 emit_fetch_system_value(
1223 struct lp_build_tgsi_context * bld_base,
1224 const struct tgsi_full_src_register * reg,
1225 enum tgsi_opcode_type stype,
1226 unsigned swizzle)
1227 {
1228 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1229 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1230 const struct tgsi_shader_info *info = bld->bld_base.info;
1231 LLVMBuilderRef builder = gallivm->builder;
1232 LLVMValueRef res;
1233 enum tgsi_opcode_type atype; // Actual type of the value
1234
1235 assert(!reg->Register.Indirect);
1236
1237 switch (info->system_value_semantic_name[reg->Register.Index]) {
1238 case TGSI_SEMANTIC_INSTANCEID:
1239 res = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.instance_id);
1240 atype = TGSI_TYPE_UNSIGNED;
1241 break;
1242
1243 case TGSI_SEMANTIC_VERTEXID:
1244 res = bld->system_values.vertex_id;
1245 atype = TGSI_TYPE_UNSIGNED;
1246 break;
1247
1248 case TGSI_SEMANTIC_PRIMID:
1249 res = bld->system_values.prim_id;
1250 atype = TGSI_TYPE_UNSIGNED;
1251 break;
1252
1253 default:
1254 assert(!"unexpected semantic in emit_fetch_system_value");
1255 res = bld_base->base.zero;
1256 atype = TGSI_TYPE_FLOAT;
1257 break;
1258 }
1259
1260 if (atype != stype) {
1261 if (stype == TGSI_TYPE_FLOAT) {
1262 res = LLVMBuildBitCast(builder, res, bld_base->base.vec_type, "");
1263 } else if (stype == TGSI_TYPE_UNSIGNED) {
1264 res = LLVMBuildBitCast(builder, res, bld_base->uint_bld.vec_type, "");
1265 } else if (stype == TGSI_TYPE_SIGNED) {
1266 res = LLVMBuildBitCast(builder, res, bld_base->int_bld.vec_type, "");
1267 }
1268 }
1269
1270 return res;
1271 }
1272
1273 /**
1274 * Register fetch with derivatives.
1275 */
1276 static void
1277 emit_fetch_deriv(
1278 struct lp_build_tgsi_soa_context *bld,
1279 LLVMValueRef src,
1280 LLVMValueRef *res,
1281 LLVMValueRef *ddx,
1282 LLVMValueRef *ddy)
1283 {
1284 if(res)
1285 *res = src;
1286
1287 /* TODO: use interpolation coeffs for inputs */
1288
1289 if(ddx)
1290 *ddx = lp_build_ddx(&bld->bld_base.base, src);
1291
1292 if(ddy)
1293 *ddy = lp_build_ddy(&bld->bld_base.base, src);
1294 }
1295
1296
1297 /**
1298 * Predicate.
1299 */
1300 static void
1301 emit_fetch_predicate(
1302 struct lp_build_tgsi_soa_context *bld,
1303 const struct tgsi_full_instruction *inst,
1304 LLVMValueRef *pred)
1305 {
1306 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1307 unsigned index;
1308 unsigned char swizzles[4];
1309 LLVMValueRef unswizzled[4] = {NULL, NULL, NULL, NULL};
1310 LLVMValueRef value;
1311 unsigned chan;
1312
1313 if (!inst->Instruction.Predicate) {
1314 TGSI_FOR_EACH_CHANNEL( chan ) {
1315 pred[chan] = NULL;
1316 }
1317 return;
1318 }
1319
1320 swizzles[0] = inst->Predicate.SwizzleX;
1321 swizzles[1] = inst->Predicate.SwizzleY;
1322 swizzles[2] = inst->Predicate.SwizzleZ;
1323 swizzles[3] = inst->Predicate.SwizzleW;
1324
1325 index = inst->Predicate.Index;
1326 assert(index < LP_MAX_TGSI_PREDS);
1327
1328 TGSI_FOR_EACH_CHANNEL( chan ) {
1329 unsigned swizzle = swizzles[chan];
1330
1331 /*
1332 * Only fetch the predicate register channels that are actually listed
1333 * in the swizzles
1334 */
1335 if (!unswizzled[swizzle]) {
1336 value = LLVMBuildLoad(builder,
1337 bld->preds[index][swizzle], "");
1338
1339 /*
1340 * Convert the value to an integer mask.
1341 *
1342 * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
1343 * is needlessly causing two comparisons due to storing the intermediate
1344 * result as float vector instead of an integer mask vector.
1345 */
1346 value = lp_build_compare(bld->bld_base.base.gallivm,
1347 bld->bld_base.base.type,
1348 PIPE_FUNC_NOTEQUAL,
1349 value,
1350 bld->bld_base.base.zero);
1351 if (inst->Predicate.Negate) {
1352 value = LLVMBuildNot(builder, value, "");
1353 }
1354
1355 unswizzled[swizzle] = value;
1356 } else {
1357 value = unswizzled[swizzle];
1358 }
1359
1360 pred[chan] = value;
1361 }
1362 }
1363
1364 /**
1365 * Register store.
1366 */
1367 static void
1368 emit_store_chan(
1369 struct lp_build_tgsi_context *bld_base,
1370 const struct tgsi_full_instruction *inst,
1371 unsigned index,
1372 unsigned chan_index,
1373 LLVMValueRef pred,
1374 LLVMValueRef value)
1375 {
1376 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1377 struct gallivm_state *gallivm = bld_base->base.gallivm;
1378 LLVMBuilderRef builder = gallivm->builder;
1379 const struct tgsi_full_dst_register *reg = &inst->Dst[index];
1380 struct lp_build_context *float_bld = &bld_base->base;
1381 struct lp_build_context *int_bld = &bld_base->int_bld;
1382 struct lp_build_context *uint_bld = &bld_base->uint_bld;
1383 LLVMValueRef indirect_index = NULL;
1384 enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
1385
1386 /*
1387 * Apply saturation.
1388 *
1389 * It is always assumed to be float.
1390 */
1391 switch( inst->Instruction.Saturate ) {
1392 case TGSI_SAT_NONE:
1393 break;
1394
1395 case TGSI_SAT_ZERO_ONE:
1396 assert(dtype == TGSI_TYPE_FLOAT ||
1397 dtype == TGSI_TYPE_UNTYPED);
1398 value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1399 value = lp_build_max(float_bld, value, float_bld->zero);
1400 value = lp_build_min(float_bld, value, float_bld->one);
1401 break;
1402
1403 case TGSI_SAT_MINUS_PLUS_ONE:
1404 assert(dtype == TGSI_TYPE_FLOAT ||
1405 dtype == TGSI_TYPE_UNTYPED);
1406 value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1407 value = lp_build_max(float_bld, value, lp_build_const_vec(gallivm, float_bld->type, -1.0));
1408 value = lp_build_min(float_bld, value, float_bld->one);
1409 break;
1410
1411 default:
1412 assert(0);
1413 }
1414
1415 if (reg->Register.Indirect) {
1416 indirect_index = get_indirect_index(bld,
1417 reg->Register.File,
1418 reg->Register.Index,
1419 &reg->Indirect);
1420 } else {
1421 assert(reg->Register.Index <=
1422 bld_base->info->file_max[reg->Register.File]);
1423 }
1424
1425 switch( reg->Register.File ) {
1426 case TGSI_FILE_OUTPUT:
1427 /* Outputs are always stored as floats */
1428 value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1429
1430 if (reg->Register.Indirect) {
1431 LLVMValueRef chan_vec =
1432 lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1433 LLVMValueRef length_vec =
1434 lp_build_const_int_vec(gallivm, uint_bld->type, float_bld->type.length);
1435 LLVMValueRef index_vec; /* indexes into the temp registers */
1436 LLVMValueRef outputs_array;
1437 LLVMValueRef pixel_offsets;
1438 LLVMTypeRef float_ptr_type;
1439 int i;
1440
1441 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1442 pixel_offsets = uint_bld->undef;
1443 for (i = 0; i < float_bld->type.length; i++) {
1444 LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1445 pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1446 ii, ii, "");
1447 }
1448
1449 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1450 index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1451 index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1452 index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1453 index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1454
1455 float_ptr_type =
1456 LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1457 outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
1458 float_ptr_type, "");
1459
1460 /* Scatter store values into temp registers */
1461 emit_mask_scatter(bld, outputs_array, index_vec, value,
1462 &bld->exec_mask, pred);
1463 }
1464 else {
1465 LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
1466 chan_index);
1467 lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value, out_ptr);
1468 }
1469 break;
1470
1471 case TGSI_FILE_TEMPORARY:
1472 /* Temporaries are always stored as floats */
1473 value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1474
1475 if (reg->Register.Indirect) {
1476 LLVMValueRef chan_vec =
1477 lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1478 LLVMValueRef length_vec =
1479 lp_build_const_int_vec(gallivm, uint_bld->type,
1480 float_bld->type.length);
1481 LLVMValueRef index_vec; /* indexes into the temp registers */
1482 LLVMValueRef temps_array;
1483 LLVMValueRef pixel_offsets;
1484 LLVMTypeRef float_ptr_type;
1485 int i;
1486
1487 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1488 pixel_offsets = uint_bld->undef;
1489 for (i = 0; i < float_bld->type.length; i++) {
1490 LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1491 pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1492 ii, ii, "");
1493 }
1494
1495 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1496 index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1497 index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1498 index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1499 index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1500
1501 float_ptr_type =
1502 LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1503 temps_array = LLVMBuildBitCast(builder, bld->temps_array,
1504 float_ptr_type, "");
1505
1506 /* Scatter store values into temp registers */
1507 emit_mask_scatter(bld, temps_array, index_vec, value,
1508 &bld->exec_mask, pred);
1509 }
1510 else {
1511 LLVMValueRef temp_ptr;
1512 temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1513 chan_index);
1514 lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value, temp_ptr);
1515 }
1516 break;
1517
1518 case TGSI_FILE_ADDRESS:
1519 assert(dtype == TGSI_TYPE_SIGNED);
1520 assert(LLVMTypeOf(value) == int_bld->vec_type);
1521 value = LLVMBuildBitCast(builder, value, int_bld->vec_type, "");
1522 lp_exec_mask_store(&bld->exec_mask, int_bld, pred, value,
1523 bld->addr[reg->Register.Index][chan_index]);
1524 break;
1525
1526 case TGSI_FILE_PREDICATE:
1527 assert(LLVMTypeOf(value) == float_bld->vec_type);
1528 value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1529 lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value,
1530 bld->preds[reg->Register.Index][chan_index]);
1531 break;
1532
1533 default:
1534 assert( 0 );
1535 }
1536
1537 (void)dtype;
1538 }
1539
1540 static void
1541 emit_store(
1542 struct lp_build_tgsi_context * bld_base,
1543 const struct tgsi_full_instruction * inst,
1544 const struct tgsi_opcode_info * info,
1545 LLVMValueRef dst[4])
1546
1547 {
1548 unsigned chan_index;
1549 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1550
1551 if(info->num_dst) {
1552 LLVMValueRef pred[TGSI_NUM_CHANNELS];
1553
1554 emit_fetch_predicate( bld, inst, pred );
1555
1556 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1557 emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1558 }
1559 }
1560 }
1561
1562 /**
1563 * High-level instruction translators.
1564 */
1565
1566 static void
1567 emit_tex( struct lp_build_tgsi_soa_context *bld,
1568 const struct tgsi_full_instruction *inst,
1569 enum lp_build_tex_modifier modifier,
1570 LLVMValueRef *texel)
1571 {
1572 unsigned unit;
1573 LLVMValueRef lod_bias, explicit_lod;
1574 LLVMValueRef oow = NULL;
1575 LLVMValueRef coords[4];
1576 LLVMValueRef offsets[3] = { NULL };
1577 struct lp_derivatives derivs;
1578 struct lp_derivatives *deriv_ptr = NULL;
1579 unsigned num_coords, num_derivs, num_offsets;
1580 unsigned i;
1581
1582 if (!bld->sampler) {
1583 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1584 for (i = 0; i < 4; i++) {
1585 texel[i] = bld->bld_base.base.undef;
1586 }
1587 return;
1588 }
1589
1590 switch (inst->Texture.Texture) {
1591 case TGSI_TEXTURE_1D:
1592 num_coords = 1;
1593 num_offsets = 1;
1594 num_derivs = 1;
1595 break;
1596 case TGSI_TEXTURE_1D_ARRAY:
1597 num_coords = 2;
1598 num_offsets = 1;
1599 num_derivs = 1;
1600 break;
1601 case TGSI_TEXTURE_2D:
1602 case TGSI_TEXTURE_RECT:
1603 num_coords = 2;
1604 num_offsets = 2;
1605 num_derivs = 2;
1606 break;
1607 case TGSI_TEXTURE_SHADOW1D:
1608 case TGSI_TEXTURE_SHADOW1D_ARRAY:
1609 num_coords = 3;
1610 num_offsets = 1;
1611 num_derivs = 1;
1612 break;
1613 case TGSI_TEXTURE_SHADOW2D:
1614 case TGSI_TEXTURE_SHADOWRECT:
1615 case TGSI_TEXTURE_2D_ARRAY:
1616 num_coords = 3;
1617 num_offsets = 2;
1618 num_derivs = 2;
1619 break;
1620 case TGSI_TEXTURE_CUBE:
1621 num_coords = 3;
1622 num_offsets = 2;
1623 num_derivs = 3;
1624 break;
1625 case TGSI_TEXTURE_3D:
1626 num_coords = 3;
1627 num_offsets = 3;
1628 num_derivs = 3;
1629 break;
1630 case TGSI_TEXTURE_SHADOW2D_ARRAY:
1631 num_coords = 4;
1632 num_offsets = 2;
1633 num_derivs = 2;
1634 break;
1635 case TGSI_TEXTURE_SHADOWCUBE:
1636 num_coords = 4;
1637 num_offsets = 2;
1638 num_derivs = 3;
1639 break;
1640 default:
1641 assert(0);
1642 return;
1643 }
1644
1645 /* Note lod and especially projected are illegal in a LOT of cases */
1646 if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1647 assert(num_coords < 4);
1648 lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1649 explicit_lod = NULL;
1650 }
1651 else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1652 assert(num_coords < 4);
1653 lod_bias = NULL;
1654 explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1655 }
1656 else {
1657 lod_bias = NULL;
1658 explicit_lod = NULL;
1659 }
1660
1661 if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
1662 assert(num_coords < 4);
1663 oow = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1664 oow = lp_build_rcp(&bld->bld_base.base, oow);
1665 }
1666
1667 for (i = 0; i < num_coords; i++) {
1668 coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1669 if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
1670 coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
1671 }
1672 for (i = num_coords; i < 4; i++) {
1673 coords[i] = bld->bld_base.base.undef;
1674 }
1675
1676 if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1677 unsigned dim;
1678 for (dim = 0; dim < num_derivs; ++dim) {
1679 derivs.ddx[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 1, dim );
1680 derivs.ddy[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 2, dim );
1681 }
1682 deriv_ptr = &derivs;
1683 unit = inst->Src[3].Register.Index;
1684 } else {
1685 unit = inst->Src[1].Register.Index;
1686 }
1687
1688 /* some advanced gather instructions (txgo) would require 4 offsets */
1689 if (inst->Texture.NumOffsets == 1) {
1690 unsigned dim;
1691 for (dim = 0; dim < num_offsets; dim++) {
1692 offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
1693 }
1694 }
1695
1696 bld->sampler->emit_fetch_texel(bld->sampler,
1697 bld->bld_base.base.gallivm,
1698 bld->bld_base.base.type,
1699 FALSE,
1700 unit, unit,
1701 coords,
1702 offsets,
1703 deriv_ptr,
1704 lod_bias, explicit_lod,
1705 texel);
1706 }
1707
1708 static void
1709 emit_sample(struct lp_build_tgsi_soa_context *bld,
1710 const struct tgsi_full_instruction *inst,
1711 enum lp_build_tex_modifier modifier,
1712 boolean compare,
1713 LLVMValueRef *texel)
1714 {
1715 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1716 unsigned texture_unit, sampler_unit;
1717 LLVMValueRef lod_bias, explicit_lod;
1718 LLVMValueRef coords[4];
1719 LLVMValueRef offsets[3] = { NULL };
1720 struct lp_derivatives derivs;
1721 struct lp_derivatives *deriv_ptr = NULL;
1722 unsigned num_coords, num_offsets, num_derivs;
1723 unsigned i;
1724
1725 if (!bld->sampler) {
1726 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1727 for (i = 0; i < 4; i++) {
1728 texel[i] = bld->bld_base.base.undef;
1729 }
1730 return;
1731 }
1732
1733 /*
1734 * unlike old-style tex opcodes the texture/sampler indices
1735 * always come from src1 and src2 respectively.
1736 */
1737 texture_unit = inst->Src[1].Register.Index;
1738 sampler_unit = inst->Src[2].Register.Index;
1739
1740 /*
1741 * Note inst->Texture.Texture will contain the number of offsets,
1742 * however the target information is NOT there and comes from the
1743 * declared sampler views instead.
1744 */
1745 switch (bld->sv[texture_unit].Resource) {
1746 case TGSI_TEXTURE_1D:
1747 num_coords = 1;
1748 num_offsets = 1;
1749 num_derivs = 1;
1750 break;
1751 case TGSI_TEXTURE_1D_ARRAY:
1752 num_coords = 2;
1753 num_offsets = 1;
1754 num_derivs = 1;
1755 break;
1756 case TGSI_TEXTURE_2D:
1757 case TGSI_TEXTURE_RECT:
1758 num_coords = 2;
1759 num_offsets = 2;
1760 num_derivs = 2;
1761 break;
1762 case TGSI_TEXTURE_2D_ARRAY:
1763 num_coords = 3;
1764 num_offsets = 2;
1765 num_derivs = 2;
1766 break;
1767 case TGSI_TEXTURE_CUBE:
1768 num_coords = 3;
1769 num_offsets = 2;
1770 num_derivs = 3;
1771 break;
1772 case TGSI_TEXTURE_3D:
1773 num_coords = 3;
1774 num_offsets = 3;
1775 num_derivs = 3;
1776 break;
1777 case TGSI_TEXTURE_CUBE_ARRAY:
1778 num_coords = 4;
1779 num_offsets = 2;
1780 num_derivs = 3;
1781 break;
1782 default:
1783 assert(0);
1784 return;
1785 }
1786
1787 if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1788 lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
1789 explicit_lod = NULL;
1790 }
1791 else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1792 lod_bias = NULL;
1793 explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
1794 }
1795 else if (modifier == LP_BLD_TEX_MODIFIER_LOD_ZERO) {
1796 lod_bias = NULL;
1797 /* XXX might be better to explicitly pass the level zero information */
1798 explicit_lod = lp_build_const_vec(gallivm, bld->bld_base.base.type, 0.0F);
1799 }
1800 else {
1801 lod_bias = NULL;
1802 explicit_lod = NULL;
1803 }
1804
1805 for (i = 0; i < num_coords; i++) {
1806 coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1807 }
1808 for (i = num_coords; i < 4; i++) {
1809 coords[i] = bld->bld_base.base.undef;
1810 }
1811 /*
1812 * XXX: whack shadow comparison value into place.
1813 * Should probably fix the interface for separate value
1814 * (it will not work for cube arrays if it is part of coords).
1815 */
1816 if (compare) {
1817 unsigned c_coord = num_coords > 2 ? 3 : 2;
1818 assert(num_coords < 4);
1819 coords[c_coord] = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
1820 }
1821
1822 if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1823 unsigned dim;
1824 for (dim = 0; dim < num_derivs; ++dim) {
1825 derivs.ddx[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 3, dim );
1826 derivs.ddy[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 4, dim );
1827 }
1828 deriv_ptr = &derivs;
1829 }
1830
1831 /* some advanced gather instructions (txgo) would require 4 offsets */
1832 if (inst->Texture.NumOffsets == 1) {
1833 unsigned dim;
1834 for (dim = 0; dim < num_offsets; dim++) {
1835 offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
1836 }
1837 }
1838
1839 bld->sampler->emit_fetch_texel(bld->sampler,
1840 bld->bld_base.base.gallivm,
1841 bld->bld_base.base.type,
1842 FALSE,
1843 texture_unit, sampler_unit,
1844 coords,
1845 offsets,
1846 deriv_ptr,
1847 lod_bias, explicit_lod,
1848 texel);
1849 }
1850
1851 static void
1852 emit_fetch_texels( struct lp_build_tgsi_soa_context *bld,
1853 const struct tgsi_full_instruction *inst,
1854 LLVMValueRef *texel,
1855 boolean is_samplei)
1856 {
1857 unsigned unit, target;
1858 LLVMValueRef coord_undef = LLVMGetUndef(bld->bld_base.base.int_vec_type);
1859 LLVMValueRef explicit_lod = NULL;
1860 LLVMValueRef coords[3];
1861 LLVMValueRef offsets[3] = { NULL };
1862 unsigned num_coords;
1863 unsigned dims;
1864 unsigned i;
1865
1866 if (!bld->sampler) {
1867 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1868 for (i = 0; i < 4; i++) {
1869 texel[i] = coord_undef;
1870 }
1871 return;
1872 }
1873
1874 unit = inst->Src[1].Register.Index;
1875
1876 if (is_samplei) {
1877 target = bld->sv[unit].Resource;
1878 }
1879 else {
1880 target = inst->Texture.Texture;
1881 }
1882
1883 switch (target) {
1884 case TGSI_TEXTURE_1D:
1885 case TGSI_TEXTURE_BUFFER:
1886 num_coords = 1;
1887 dims = 1;
1888 break;
1889 case TGSI_TEXTURE_1D_ARRAY:
1890 num_coords = 2;
1891 dims = 1;
1892 break;
1893 case TGSI_TEXTURE_2D:
1894 case TGSI_TEXTURE_RECT:
1895 num_coords = 2;
1896 dims = 2;
1897 break;
1898 case TGSI_TEXTURE_2D_ARRAY:
1899 num_coords = 3;
1900 dims = 2;
1901 break;
1902 case TGSI_TEXTURE_3D:
1903 num_coords = 3;
1904 dims = 3;
1905 break;
1906 default:
1907 assert(0);
1908 return;
1909 }
1910
1911 /* always have lod except for buffers ? */
1912 if (target != TGSI_TEXTURE_BUFFER) {
1913 explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1914 }
1915
1916 for (i = 0; i < num_coords; i++) {
1917 coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1918 }
1919 for (i = num_coords; i < 3; i++) {
1920 coords[i] = coord_undef;
1921 }
1922
1923 if (inst->Texture.NumOffsets == 1) {
1924 unsigned dim;
1925 for (dim = 0; dim < dims; dim++) {
1926 offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
1927 }
1928 }
1929
1930 bld->sampler->emit_fetch_texel(bld->sampler,
1931 bld->bld_base.base.gallivm,
1932 bld->bld_base.base.type,
1933 TRUE,
1934 unit, unit,
1935 coords,
1936 offsets,
1937 NULL,
1938 NULL, explicit_lod,
1939 texel);
1940 }
1941
1942 static void
1943 emit_size_query( struct lp_build_tgsi_soa_context *bld,
1944 const struct tgsi_full_instruction *inst,
1945 LLVMValueRef *sizes_out,
1946 boolean is_sviewinfo)
1947 {
1948 LLVMValueRef explicit_lod;
1949 unsigned has_lod;
1950 unsigned i;
1951 unsigned unit = inst->Src[1].Register.Index;
1952 unsigned target;
1953
1954 if (is_sviewinfo) {
1955 target = bld->sv[unit].Resource;
1956 }
1957 else {
1958 target = inst->Texture.Texture;
1959 }
1960 switch (target) {
1961 case TGSI_TEXTURE_BUFFER:
1962 case TGSI_TEXTURE_RECT:
1963 case TGSI_TEXTURE_SHADOWRECT:
1964 has_lod = 0;
1965 break;
1966 default:
1967 has_lod = 1;
1968 break;
1969 }
1970
1971 if (!bld->sampler) {
1972 _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
1973 for (i = 0; i < 4; i++)
1974 sizes_out[i] = bld->bld_base.int_bld.undef;
1975 return;
1976 }
1977
1978 if (has_lod)
1979 explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 0 );
1980 else
1981 explicit_lod = NULL;
1982
1983 bld->sampler->emit_size_query(bld->sampler,
1984 bld->bld_base.base.gallivm,
1985 bld->bld_base.int_bld.type,
1986 unit,
1987 is_sviewinfo,
1988 explicit_lod,
1989 sizes_out);
1990 }
1991
1992 static boolean
1993 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
1994 int pc)
1995 {
1996 int i;
1997
1998 for (i = 0; i < 5; i++) {
1999 unsigned opcode;
2000
2001 if (pc + i >= bld->bld_base.info->num_instructions)
2002 return TRUE;
2003
2004 opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
2005
2006 if (opcode == TGSI_OPCODE_END)
2007 return TRUE;
2008
2009 if (opcode == TGSI_OPCODE_TEX ||
2010 opcode == TGSI_OPCODE_TXP ||
2011 opcode == TGSI_OPCODE_TXD ||
2012 opcode == TGSI_OPCODE_TXB ||
2013 opcode == TGSI_OPCODE_TXL ||
2014 opcode == TGSI_OPCODE_TXF ||
2015 opcode == TGSI_OPCODE_TXQ ||
2016 opcode == TGSI_OPCODE_CAL ||
2017 opcode == TGSI_OPCODE_CALLNZ ||
2018 opcode == TGSI_OPCODE_IF ||
2019 opcode == TGSI_OPCODE_UIF ||
2020 opcode == TGSI_OPCODE_BGNLOOP ||
2021 opcode == TGSI_OPCODE_SWITCH)
2022 return FALSE;
2023 }
2024
2025 return TRUE;
2026 }
2027
2028
2029
2030 /**
2031 * Kill fragment if any of the src register values are negative.
2032 */
2033 static void
2034 emit_kil(
2035 struct lp_build_tgsi_soa_context *bld,
2036 const struct tgsi_full_instruction *inst,
2037 int pc)
2038 {
2039 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2040 const struct tgsi_full_src_register *reg = &inst->Src[0];
2041 LLVMValueRef terms[TGSI_NUM_CHANNELS];
2042 LLVMValueRef mask;
2043 unsigned chan_index;
2044
2045 memset(&terms, 0, sizeof terms);
2046
2047 TGSI_FOR_EACH_CHANNEL( chan_index ) {
2048 unsigned swizzle;
2049
2050 /* Unswizzle channel */
2051 swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
2052
2053 /* Check if the component has not been already tested. */
2054 assert(swizzle < TGSI_NUM_CHANNELS);
2055 if( !terms[swizzle] )
2056 /* TODO: change the comparison operator instead of setting the sign */
2057 terms[swizzle] = lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
2058 }
2059
2060 mask = NULL;
2061 TGSI_FOR_EACH_CHANNEL( chan_index ) {
2062 if(terms[chan_index]) {
2063 LLVMValueRef chan_mask;
2064
2065 /*
2066 * If term < 0 then mask = 0 else mask = ~0.
2067 */
2068 chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
2069
2070 if(mask)
2071 mask = LLVMBuildAnd(builder, mask, chan_mask, "");
2072 else
2073 mask = chan_mask;
2074 }
2075 }
2076
2077 if(mask) {
2078 lp_build_mask_update(bld->mask, mask);
2079
2080 if (!near_end_of_shader(bld, pc))
2081 lp_build_mask_check(bld->mask);
2082 }
2083 }
2084
2085
2086 /**
2087 * Predicated fragment kill.
2088 * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
2089 * The only predication is the execution mask which will apply if
2090 * we're inside a loop or conditional.
2091 */
2092 static void
2093 emit_kilp(struct lp_build_tgsi_soa_context *bld,
2094 int pc)
2095 {
2096 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2097 LLVMValueRef mask;
2098
2099 /* For those channels which are "alive", disable fragment shader
2100 * execution.
2101 */
2102 if (bld->exec_mask.has_mask) {
2103 mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
2104 }
2105 else {
2106 LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
2107 mask = zero;
2108 }
2109
2110 lp_build_mask_update(bld->mask, mask);
2111
2112 if (!near_end_of_shader(bld, pc))
2113 lp_build_mask_check(bld->mask);
2114 }
2115
2116
2117 /**
2118 * Emit code which will dump the value of all the temporary registers
2119 * to stdout.
2120 */
2121 static void
2122 emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
2123 {
2124 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2125 LLVMBuilderRef builder = gallivm->builder;
2126 LLVMValueRef temp_ptr;
2127 LLVMValueRef i0 = lp_build_const_int32(gallivm, 0);
2128 LLVMValueRef i1 = lp_build_const_int32(gallivm, 1);
2129 LLVMValueRef i2 = lp_build_const_int32(gallivm, 2);
2130 LLVMValueRef i3 = lp_build_const_int32(gallivm, 3);
2131 int index;
2132 int n = bld->bld_base.info->file_max[TGSI_FILE_TEMPORARY];
2133
2134 for (index = 0; index < n; index++) {
2135 LLVMValueRef idx = lp_build_const_int32(gallivm, index);
2136 LLVMValueRef v[4][4], res;
2137 int chan;
2138
2139 lp_build_printf(gallivm, "TEMP[%d]:\n", idx);
2140
2141 for (chan = 0; chan < 4; chan++) {
2142 temp_ptr = lp_get_temp_ptr_soa(bld, index, chan);
2143 res = LLVMBuildLoad(builder, temp_ptr, "");
2144 v[chan][0] = LLVMBuildExtractElement(builder, res, i0, "");
2145 v[chan][1] = LLVMBuildExtractElement(builder, res, i1, "");
2146 v[chan][2] = LLVMBuildExtractElement(builder, res, i2, "");
2147 v[chan][3] = LLVMBuildExtractElement(builder, res, i3, "");
2148 }
2149
2150 lp_build_printf(gallivm, " X: %f %f %f %f\n",
2151 v[0][0], v[0][1], v[0][2], v[0][3]);
2152 lp_build_printf(gallivm, " Y: %f %f %f %f\n",
2153 v[1][0], v[1][1], v[1][2], v[1][3]);
2154 lp_build_printf(gallivm, " Z: %f %f %f %f\n",
2155 v[2][0], v[2][1], v[2][2], v[2][3]);
2156 lp_build_printf(gallivm, " W: %f %f %f %f\n",
2157 v[3][0], v[3][1], v[3][2], v[3][3]);
2158 }
2159 }
2160
2161
2162
2163 void
2164 lp_emit_declaration_soa(
2165 struct lp_build_tgsi_context *bld_base,
2166 const struct tgsi_full_declaration *decl)
2167 {
2168 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
2169 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2170 LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
2171 const unsigned first = decl->Range.First;
2172 const unsigned last = decl->Range.Last;
2173 unsigned idx, i;
2174
2175 for (idx = first; idx <= last; ++idx) {
2176 assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
2177 switch (decl->Declaration.File) {
2178 case TGSI_FILE_TEMPORARY:
2179 assert(idx < LP_MAX_TGSI_TEMPS);
2180 if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
2181 for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2182 bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
2183 }
2184 break;
2185
2186 case TGSI_FILE_OUTPUT:
2187 if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
2188 for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2189 bld->outputs[idx][i] = lp_build_alloca(gallivm,
2190 vec_type, "output");
2191 }
2192 break;
2193
2194 case TGSI_FILE_ADDRESS:
2195 /* ADDR registers are only allocated with an integer LLVM IR type,
2196 * as they are guaranteed to always have integers.
2197 * XXX: Not sure if this exception is worthwhile (or the whole idea of
2198 * an ADDR register for that matter).
2199 */
2200 assert(idx < LP_MAX_TGSI_ADDRS);
2201 for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2202 bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
2203 break;
2204
2205 case TGSI_FILE_PREDICATE:
2206 assert(idx < LP_MAX_TGSI_PREDS);
2207 for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2208 bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
2209 "predicate");
2210 break;
2211
2212 case TGSI_FILE_SAMPLER_VIEW:
2213 /*
2214 * The target stored here MUST match whatever there actually
2215 * is in the set sampler views (what about return type?).
2216 */
2217 assert(idx < PIPE_MAX_SHADER_SAMPLER_VIEWS);
2218 bld->sv[idx] = decl->SamplerView;
2219 break;
2220
2221 default:
2222 /* don't need to declare other vars */
2223 break;
2224 }
2225 }
2226 }
2227
2228
2229 void lp_emit_immediate_soa(
2230 struct lp_build_tgsi_context *bld_base,
2231 const struct tgsi_full_immediate *imm)
2232 {
2233 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
2234 struct gallivm_state * gallivm = bld_base->base.gallivm;
2235
2236 /* simply copy the immediate values into the next immediates[] slot */
2237 unsigned i;
2238 const uint size = imm->Immediate.NrTokens - 1;
2239 assert(size <= 4);
2240 assert(bld->num_immediates < LP_MAX_TGSI_IMMEDIATES);
2241 switch (imm->Immediate.DataType) {
2242 case TGSI_IMM_FLOAT32:
2243 for( i = 0; i < size; ++i )
2244 bld->immediates[bld->num_immediates][i] =
2245 lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
2246
2247 break;
2248 case TGSI_IMM_UINT32:
2249 for( i = 0; i < size; ++i ) {
2250 LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
2251 bld->immediates[bld->num_immediates][i] =
2252 LLVMConstBitCast(tmp, bld_base->base.vec_type);
2253 }
2254
2255 break;
2256 case TGSI_IMM_INT32:
2257 for( i = 0; i < size; ++i ) {
2258 LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
2259 bld->immediates[bld->num_immediates][i] =
2260 LLVMConstBitCast(tmp, bld_base->base.vec_type);
2261 }
2262
2263 break;
2264 }
2265 for( i = size; i < 4; ++i )
2266 bld->immediates[bld->num_immediates][i] = bld_base->base.undef;
2267
2268 if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
2269 unsigned index = bld->num_immediates;
2270 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2271 LLVMBuilderRef builder = gallivm->builder;
2272 for (i = 0; i < 4; ++i ) {
2273 LLVMValueRef lindex = lp_build_const_int32(
2274 bld->bld_base.base.gallivm, index * 4 + i);
2275 LLVMValueRef imm_ptr = LLVMBuildGEP(builder,
2276 bld->imms_array, &lindex, 1, "");
2277 LLVMBuildStore(builder,
2278 bld->immediates[index][i],
2279 imm_ptr);
2280 }
2281 }
2282
2283 bld->num_immediates++;
2284 }
2285
2286 static void
2287 ddx_emit(
2288 const struct lp_build_tgsi_action * action,
2289 struct lp_build_tgsi_context * bld_base,
2290 struct lp_build_emit_data * emit_data)
2291 {
2292 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2293
2294 emit_fetch_deriv(bld, emit_data->args[0], NULL,
2295 &emit_data->output[emit_data->chan], NULL);
2296 }
2297
2298 static void
2299 ddy_emit(
2300 const struct lp_build_tgsi_action * action,
2301 struct lp_build_tgsi_context * bld_base,
2302 struct lp_build_emit_data * emit_data)
2303 {
2304 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2305
2306 emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
2307 &emit_data->output[emit_data->chan]);
2308 }
2309
2310 static void
2311 kilp_emit(
2312 const struct lp_build_tgsi_action * action,
2313 struct lp_build_tgsi_context * bld_base,
2314 struct lp_build_emit_data * emit_data)
2315 {
2316 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2317
2318 emit_kilp(bld, bld_base->pc - 1);
2319 }
2320
2321 static void
2322 kil_emit(
2323 const struct lp_build_tgsi_action * action,
2324 struct lp_build_tgsi_context * bld_base,
2325 struct lp_build_emit_data * emit_data)
2326 {
2327 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2328
2329 emit_kil(bld, emit_data->inst, bld_base->pc - 1);
2330 }
2331
2332 static void
2333 tex_emit(
2334 const struct lp_build_tgsi_action * action,
2335 struct lp_build_tgsi_context * bld_base,
2336 struct lp_build_emit_data * emit_data)
2337 {
2338 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2339
2340 emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE, emit_data->output);
2341 }
2342
2343 static void
2344 txb_emit(
2345 const struct lp_build_tgsi_action * action,
2346 struct lp_build_tgsi_context * bld_base,
2347 struct lp_build_emit_data * emit_data)
2348 {
2349 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2350
2351 emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
2352 emit_data->output);
2353 }
2354
2355 static void
2356 txd_emit(
2357 const struct lp_build_tgsi_action * action,
2358 struct lp_build_tgsi_context * bld_base,
2359 struct lp_build_emit_data * emit_data)
2360 {
2361 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2362
2363 emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
2364 emit_data->output);
2365 }
2366
2367 static void
2368 txl_emit(
2369 const struct lp_build_tgsi_action * action,
2370 struct lp_build_tgsi_context * bld_base,
2371 struct lp_build_emit_data * emit_data)
2372 {
2373 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2374
2375 emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
2376 emit_data->output);
2377 }
2378
2379 static void
2380 txp_emit(
2381 const struct lp_build_tgsi_action * action,
2382 struct lp_build_tgsi_context * bld_base,
2383 struct lp_build_emit_data * emit_data)
2384 {
2385 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2386
2387 emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
2388 emit_data->output);
2389 }
2390
2391 static void
2392 txq_emit(
2393 const struct lp_build_tgsi_action * action,
2394 struct lp_build_tgsi_context * bld_base,
2395 struct lp_build_emit_data * emit_data)
2396 {
2397 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2398
2399 emit_size_query(bld, emit_data->inst, emit_data->output, FALSE);
2400 }
2401
2402 static void
2403 txf_emit(
2404 const struct lp_build_tgsi_action * action,
2405 struct lp_build_tgsi_context * bld_base,
2406 struct lp_build_emit_data * emit_data)
2407 {
2408 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2409
2410 emit_fetch_texels(bld, emit_data->inst, emit_data->output, FALSE);
2411 }
2412
2413 static void
2414 sample_i_emit(
2415 const struct lp_build_tgsi_action * action,
2416 struct lp_build_tgsi_context * bld_base,
2417 struct lp_build_emit_data * emit_data)
2418 {
2419 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2420
2421 emit_fetch_texels(bld, emit_data->inst, emit_data->output, TRUE);
2422 }
2423
2424 static void
2425 sample_emit(
2426 const struct lp_build_tgsi_action * action,
2427 struct lp_build_tgsi_context * bld_base,
2428 struct lp_build_emit_data * emit_data)
2429 {
2430 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2431
2432 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2433 FALSE, emit_data->output);
2434 }
2435
2436 static void
2437 sample_b_emit(
2438 const struct lp_build_tgsi_action * action,
2439 struct lp_build_tgsi_context * bld_base,
2440 struct lp_build_emit_data * emit_data)
2441 {
2442 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2443
2444 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
2445 FALSE, emit_data->output);
2446 }
2447
2448 static void
2449 sample_c_emit(
2450 const struct lp_build_tgsi_action * action,
2451 struct lp_build_tgsi_context * bld_base,
2452 struct lp_build_emit_data * emit_data)
2453 {
2454 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2455
2456 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2457 TRUE, emit_data->output);
2458 }
2459
2460 static void
2461 sample_c_lz_emit(
2462 const struct lp_build_tgsi_action * action,
2463 struct lp_build_tgsi_context * bld_base,
2464 struct lp_build_emit_data * emit_data)
2465 {
2466 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2467
2468 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_ZERO,
2469 TRUE, emit_data->output);
2470 }
2471
2472 static void
2473 sample_d_emit(
2474 const struct lp_build_tgsi_action * action,
2475 struct lp_build_tgsi_context * bld_base,
2476 struct lp_build_emit_data * emit_data)
2477 {
2478 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2479
2480 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
2481 FALSE, emit_data->output);
2482 }
2483
2484 static void
2485 sample_l_emit(
2486 const struct lp_build_tgsi_action * action,
2487 struct lp_build_tgsi_context * bld_base,
2488 struct lp_build_emit_data * emit_data)
2489 {
2490 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2491
2492 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
2493 FALSE, emit_data->output);
2494 }
2495
2496 static void
2497 sviewinfo_emit(
2498 const struct lp_build_tgsi_action * action,
2499 struct lp_build_tgsi_context * bld_base,
2500 struct lp_build_emit_data * emit_data)
2501 {
2502 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2503
2504 emit_size_query(bld, emit_data->inst, emit_data->output, TRUE);
2505 }
2506
2507 static LLVMValueRef
2508 mask_to_one_vec(struct lp_build_tgsi_context *bld_base)
2509 {
2510 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2511 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2512 LLVMValueRef one_vec = bld_base->int_bld.one;
2513 struct lp_exec_mask *exec_mask = &bld->exec_mask;
2514
2515 if (exec_mask->has_mask) {
2516 one_vec = LLVMBuildAnd(builder, one_vec, exec_mask->exec_mask, "");
2517 }
2518 one_vec = LLVMBuildAnd(builder, one_vec,
2519 lp_build_mask_value(bld->mask), "");
2520 return one_vec;
2521 }
2522
2523 static void
2524 increment_vec_ptr_by_mask(struct lp_build_tgsi_context * bld_base,
2525 LLVMValueRef ptr,
2526 LLVMValueRef mask)
2527 {
2528 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2529
2530 LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
2531
2532 current_vec = LLVMBuildAdd(builder, current_vec, mask, "");
2533
2534 LLVMBuildStore(builder, current_vec, ptr);
2535 }
2536
2537 static void
2538 clear_uint_vec_ptr_from_mask(struct lp_build_tgsi_context * bld_base,
2539 LLVMValueRef ptr,
2540 LLVMValueRef mask)
2541 {
2542 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2543
2544 LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
2545 LLVMValueRef full_mask = lp_build_cmp(&bld_base->uint_bld,
2546 PIPE_FUNC_NOTEQUAL,
2547 mask,
2548 bld_base->uint_bld.zero);
2549
2550 current_vec = lp_build_select(&bld_base->uint_bld,
2551 full_mask,
2552 bld_base->uint_bld.zero,
2553 current_vec);
2554
2555 LLVMBuildStore(builder, current_vec, ptr);
2556 }
2557
2558 static LLVMValueRef
2559 clamp_mask_to_max_output_vertices(struct lp_build_tgsi_soa_context * bld,
2560 LLVMValueRef current_mask_vec,
2561 LLVMValueRef total_emitted_vertices_vec)
2562 {
2563 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2564 struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
2565 LLVMValueRef max_mask = lp_build_cmp(uint_bld, PIPE_FUNC_LESS,
2566 total_emitted_vertices_vec,
2567 bld->max_output_vertices_vec);
2568
2569 return LLVMBuildAnd(builder, current_mask_vec, max_mask, "");
2570 }
2571
2572 static void
2573 emit_vertex(
2574 const struct lp_build_tgsi_action * action,
2575 struct lp_build_tgsi_context * bld_base,
2576 struct lp_build_emit_data * emit_data)
2577 {
2578 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2579 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2580
2581 if (bld->gs_iface->emit_vertex) {
2582 LLVMValueRef masked_ones = mask_to_one_vec(bld_base);
2583 LLVMValueRef total_emitted_vertices_vec =
2584 LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
2585 masked_ones = clamp_mask_to_max_output_vertices(bld, masked_ones,
2586 total_emitted_vertices_vec);
2587 gather_outputs(bld);
2588 bld->gs_iface->emit_vertex(bld->gs_iface, &bld->bld_base,
2589 bld->outputs,
2590 total_emitted_vertices_vec);
2591 increment_vec_ptr_by_mask(bld_base, bld->emitted_vertices_vec_ptr,
2592 masked_ones);
2593 increment_vec_ptr_by_mask(bld_base, bld->total_emitted_vertices_vec_ptr,
2594 masked_ones);
2595 #if DUMP_GS_EMITS
2596 lp_build_print_value(bld->bld_base.base.gallivm,
2597 " +++ emit vertex masked ones = ",
2598 masked_ones);
2599 lp_build_print_value(bld->bld_base.base.gallivm,
2600 " +++ emit vertex emitted = ",
2601 total_emitted_vertices_vec);
2602 #endif
2603 }
2604 }
2605
2606
2607 static void
2608 end_primitive_masked(struct lp_build_tgsi_context * bld_base,
2609 LLVMValueRef masked_ones)
2610 {
2611 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2612 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2613
2614 if (bld->gs_iface->end_primitive) {
2615 LLVMValueRef emitted_vertices_vec =
2616 LLVMBuildLoad(builder, bld->emitted_vertices_vec_ptr, "");
2617 LLVMValueRef emitted_prims_vec =
2618 LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
2619
2620 bld->gs_iface->end_primitive(bld->gs_iface, &bld->bld_base,
2621 emitted_vertices_vec,
2622 emitted_prims_vec);
2623
2624 #if DUMP_GS_EMITS
2625 lp_build_print_value(bld->bld_base.base.gallivm,
2626 " +++ end prim masked ones = ",
2627 masked_ones);
2628 lp_build_print_value(bld->bld_base.base.gallivm,
2629 " +++ end prim emitted verts1 = ",
2630 emitted_vertices_vec);
2631 lp_build_print_value(bld->bld_base.base.gallivm,
2632 " +++ end prim emitted prims1 = ",
2633 LLVMBuildLoad(builder,
2634 bld->emitted_prims_vec_ptr, ""));
2635 #endif
2636 increment_vec_ptr_by_mask(bld_base, bld->emitted_prims_vec_ptr,
2637 masked_ones);
2638 clear_uint_vec_ptr_from_mask(bld_base, bld->emitted_vertices_vec_ptr,
2639 masked_ones);
2640 #if DUMP_GS_EMITS
2641 lp_build_print_value(bld->bld_base.base.gallivm,
2642 " +++ end prim emitted verts2 = ",
2643 LLVMBuildLoad(builder,
2644 bld->emitted_vertices_vec_ptr, ""));
2645 #endif
2646 }
2647
2648 }
2649
2650 static void
2651 end_primitive(
2652 const struct lp_build_tgsi_action * action,
2653 struct lp_build_tgsi_context * bld_base,
2654 struct lp_build_emit_data * emit_data)
2655 {
2656 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2657
2658 if (bld->gs_iface->end_primitive) {
2659 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2660 LLVMValueRef masked_ones = mask_to_one_vec(bld_base);
2661 struct lp_build_context *uint_bld = &bld_base->uint_bld;
2662 LLVMValueRef emitted_verts = LLVMBuildLoad(
2663 builder, bld->emitted_vertices_vec_ptr, "");
2664 LLVMValueRef emitted_mask = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
2665 emitted_verts,
2666 uint_bld->zero);
2667 /* We need to combine the current execution mask with the mask
2668 telling us which, if any, execution slots actually have
2669 unemitted primitives, this way we make sure that end_primitives
2670 executes only on the paths that have unflushed vertices */
2671 masked_ones = LLVMBuildAnd(builder, masked_ones, emitted_mask, "");
2672
2673 end_primitive_masked(bld_base, masked_ones);
2674 }
2675 }
2676
2677 static void
2678 cal_emit(
2679 const struct lp_build_tgsi_action * action,
2680 struct lp_build_tgsi_context * bld_base,
2681 struct lp_build_emit_data * emit_data)
2682 {
2683 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2684
2685 lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
2686 &bld_base->pc);
2687 }
2688
2689 static void
2690 ret_emit(
2691 const struct lp_build_tgsi_action * action,
2692 struct lp_build_tgsi_context * bld_base,
2693 struct lp_build_emit_data * emit_data)
2694 {
2695 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2696
2697 lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
2698 }
2699
2700 static void
2701 brk_emit(
2702 const struct lp_build_tgsi_action * action,
2703 struct lp_build_tgsi_context * bld_base,
2704 struct lp_build_emit_data * emit_data)
2705 {
2706 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2707
2708 lp_exec_break(&bld->exec_mask, bld_base);
2709 }
2710
2711 static void
2712 breakc_emit(
2713 const struct lp_build_tgsi_action * action,
2714 struct lp_build_tgsi_context * bld_base,
2715 struct lp_build_emit_data * emit_data)
2716 {
2717 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2718 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2719 struct lp_build_context *uint_bld = &bld_base->uint_bld;
2720 LLVMValueRef unsigned_cond =
2721 LLVMBuildBitCast(builder, emit_data->args[0], uint_bld->vec_type, "");
2722 LLVMValueRef cond = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
2723 unsigned_cond,
2724 uint_bld->zero);
2725
2726 lp_exec_break_condition(&bld->exec_mask, cond);
2727 }
2728
2729 static void
2730 if_emit(
2731 const struct lp_build_tgsi_action * action,
2732 struct lp_build_tgsi_context * bld_base,
2733 struct lp_build_emit_data * emit_data)
2734 {
2735 LLVMValueRef tmp;
2736 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2737
2738 tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
2739 emit_data->args[0], bld->bld_base.base.zero);
2740 lp_exec_mask_cond_push(&bld->exec_mask, tmp);
2741 }
2742
2743 static void
2744 uif_emit(
2745 const struct lp_build_tgsi_action * action,
2746 struct lp_build_tgsi_context * bld_base,
2747 struct lp_build_emit_data * emit_data)
2748 {
2749 LLVMValueRef tmp;
2750 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2751 struct lp_build_context *uint_bld = &bld_base->uint_bld;
2752
2753 tmp = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
2754 emit_data->args[0], uint_bld->zero);
2755 lp_exec_mask_cond_push(&bld->exec_mask, tmp);
2756 }
2757
2758 static void
2759 case_emit(
2760 const struct lp_build_tgsi_action * action,
2761 struct lp_build_tgsi_context * bld_base,
2762 struct lp_build_emit_data * emit_data)
2763 {
2764 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2765
2766 lp_exec_case(&bld->exec_mask, emit_data->args[0]);
2767 }
2768
2769 static void
2770 default_emit(
2771 const struct lp_build_tgsi_action * action,
2772 struct lp_build_tgsi_context * bld_base,
2773 struct lp_build_emit_data * emit_data)
2774 {
2775 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2776
2777 lp_exec_default(&bld->exec_mask, bld_base);
2778 }
2779
2780 static void
2781 switch_emit(
2782 const struct lp_build_tgsi_action * action,
2783 struct lp_build_tgsi_context * bld_base,
2784 struct lp_build_emit_data * emit_data)
2785 {
2786 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2787
2788 lp_exec_switch(&bld->exec_mask, emit_data->args[0]);
2789 }
2790
2791 static void
2792 endswitch_emit(
2793 const struct lp_build_tgsi_action * action,
2794 struct lp_build_tgsi_context * bld_base,
2795 struct lp_build_emit_data * emit_data)
2796 {
2797 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2798
2799 lp_exec_endswitch(&bld->exec_mask, bld_base);
2800 }
2801
2802 static void
2803 bgnloop_emit(
2804 const struct lp_build_tgsi_action * action,
2805 struct lp_build_tgsi_context * bld_base,
2806 struct lp_build_emit_data * emit_data)
2807 {
2808 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2809
2810 lp_exec_bgnloop(&bld->exec_mask);
2811 }
2812
2813 static void
2814 bgnsub_emit(
2815 const struct lp_build_tgsi_action * action,
2816 struct lp_build_tgsi_context * bld_base,
2817 struct lp_build_emit_data * emit_data)
2818 {
2819 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2820
2821 lp_exec_mask_bgnsub(&bld->exec_mask);
2822 }
2823
2824 static void
2825 else_emit(
2826 const struct lp_build_tgsi_action * action,
2827 struct lp_build_tgsi_context * bld_base,
2828 struct lp_build_emit_data * emit_data)
2829 {
2830 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2831
2832 lp_exec_mask_cond_invert(&bld->exec_mask);
2833 }
2834
2835 static void
2836 endif_emit(
2837 const struct lp_build_tgsi_action * action,
2838 struct lp_build_tgsi_context * bld_base,
2839 struct lp_build_emit_data * emit_data)
2840 {
2841 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2842
2843 lp_exec_mask_cond_pop(&bld->exec_mask);
2844 }
2845
2846 static void
2847 endloop_emit(
2848 const struct lp_build_tgsi_action * action,
2849 struct lp_build_tgsi_context * bld_base,
2850 struct lp_build_emit_data * emit_data)
2851 {
2852 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2853
2854 lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
2855 }
2856
2857 static void
2858 endsub_emit(
2859 const struct lp_build_tgsi_action * action,
2860 struct lp_build_tgsi_context * bld_base,
2861 struct lp_build_emit_data * emit_data)
2862 {
2863 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2864
2865 lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
2866 }
2867
2868 static void
2869 cont_emit(
2870 const struct lp_build_tgsi_action * action,
2871 struct lp_build_tgsi_context * bld_base,
2872 struct lp_build_emit_data * emit_data)
2873 {
2874 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2875
2876 lp_exec_continue(&bld->exec_mask);
2877 }
2878
2879 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
2880 *
2881 * XXX: What do the comments about xmm registers mean? Maybe they are left over
2882 * from old code, but there is no garauntee that LLVM will use those registers
2883 * for this code.
2884 *
2885 * XXX: There should be no calls to lp_build_emit_fetch in this function. This
2886 * should be handled by the emit_data->fetch_args function. */
2887 static void
2888 nrm_emit(
2889 const struct lp_build_tgsi_action * action,
2890 struct lp_build_tgsi_context * bld_base,
2891 struct lp_build_emit_data * emit_data)
2892 {
2893 LLVMValueRef tmp0, tmp1;
2894 LLVMValueRef tmp4 = NULL;
2895 LLVMValueRef tmp5 = NULL;
2896 LLVMValueRef tmp6 = NULL;
2897 LLVMValueRef tmp7 = NULL;
2898 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2899
2900 uint dims = (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
2901
2902 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) ||
2903 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y) ||
2904 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z) ||
2905 (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 4)) {
2906
2907 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
2908
2909 /* xmm4 = src.x */
2910 /* xmm0 = src.x * src.x */
2911 tmp0 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_X);
2912 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
2913 tmp4 = tmp0;
2914 }
2915 tmp0 = lp_build_mul( &bld->bld_base.base, tmp0, tmp0);
2916
2917 /* xmm5 = src.y */
2918 /* xmm0 = xmm0 + src.y * src.y */
2919 tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Y);
2920 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
2921 tmp5 = tmp1;
2922 }
2923 tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2924 tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2925
2926 /* xmm6 = src.z */
2927 /* xmm0 = xmm0 + src.z * src.z */
2928 tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Z);
2929 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
2930 tmp6 = tmp1;
2931 }
2932 tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2933 tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2934
2935 if (dims == 4) {
2936 /* xmm7 = src.w */
2937 /* xmm0 = xmm0 + src.w * src.w */
2938 tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_W);
2939 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W)) {
2940 tmp7 = tmp1;
2941 }
2942 tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2943 tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2944 }
2945 /* xmm1 = 1 / sqrt(xmm0) */
2946 tmp1 = lp_build_rsqrt( &bld->bld_base.base, tmp0);
2947 /* dst.x = xmm1 * src.x */
2948 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
2949 emit_data->output[TGSI_CHAN_X] = lp_build_mul( &bld->bld_base.base, tmp4, tmp1);
2950 }
2951 /* dst.y = xmm1 * src.y */
2952 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
2953 emit_data->output[TGSI_CHAN_Y] = lp_build_mul( &bld->bld_base.base, tmp5, tmp1);
2954 }
2955
2956 /* dst.z = xmm1 * src.z */
2957 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
2958 emit_data->output[TGSI_CHAN_Z] = lp_build_mul( &bld->bld_base.base, tmp6, tmp1);
2959 }
2960 /* dst.w = xmm1 * src.w */
2961 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) && dims == 4) {
2962 emit_data->output[TGSI_CHAN_W] = lp_build_mul( &bld->bld_base.base, tmp7, tmp1);
2963 }
2964 }
2965
2966 /* dst.w = 1.0 */
2967 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 3) {
2968 emit_data->output[TGSI_CHAN_W] = bld->bld_base.base.one;
2969 }
2970 }
2971
2972 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
2973 {
2974 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2975 struct gallivm_state * gallivm = bld_base->base.gallivm;
2976
2977 if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
2978 LLVMValueRef array_size =
2979 lp_build_const_int32(gallivm,
2980 bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
2981 bld->temps_array = lp_build_array_alloca(gallivm,
2982 bld_base->base.vec_type, array_size,
2983 "temp_array");
2984 }
2985
2986 if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
2987 LLVMValueRef array_size =
2988 lp_build_const_int32(gallivm,
2989 bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
2990 bld->outputs_array = lp_build_array_alloca(gallivm,
2991 bld_base->base.vec_type, array_size,
2992 "output_array");
2993 }
2994
2995 if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
2996 LLVMValueRef array_size =
2997 lp_build_const_int32(gallivm,
2998 bld_base->info->file_max[TGSI_FILE_IMMEDIATE] * 4 + 4);
2999 bld->imms_array = lp_build_array_alloca(gallivm,
3000 bld_base->base.vec_type, array_size,
3001 "imms_array");
3002 }
3003
3004 /* If we have indirect addressing in inputs we need to copy them into
3005 * our alloca array to be able to iterate over them */
3006 if (bld->indirect_files & (1 << TGSI_FILE_INPUT) && !bld->gs_iface) {
3007 unsigned index, chan;
3008 LLVMTypeRef vec_type = bld_base->base.vec_type;
3009 LLVMValueRef array_size = lp_build_const_int32(gallivm,
3010 bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
3011 bld->inputs_array = lp_build_array_alloca(gallivm,
3012 vec_type, array_size,
3013 "input_array");
3014
3015 assert(bld_base->info->num_inputs
3016 <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
3017
3018 for (index = 0; index < bld_base->info->num_inputs; ++index) {
3019 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
3020 LLVMValueRef lindex =
3021 lp_build_const_int32(gallivm, index * 4 + chan);
3022 LLVMValueRef input_ptr =
3023 LLVMBuildGEP(gallivm->builder, bld->inputs_array,
3024 &lindex, 1, "");
3025 LLVMValueRef value = bld->inputs[index][chan];
3026 if (value)
3027 LLVMBuildStore(gallivm->builder, value, input_ptr);
3028 }
3029 }
3030 }
3031
3032 if (bld->gs_iface) {
3033 struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
3034 bld->emitted_prims_vec_ptr =
3035 lp_build_alloca(gallivm,
3036 uint_bld->vec_type,
3037 "emitted_prims_ptr");
3038 bld->emitted_vertices_vec_ptr =
3039 lp_build_alloca(gallivm,
3040 uint_bld->vec_type,
3041 "emitted_vertices_ptr");
3042 bld->total_emitted_vertices_vec_ptr =
3043 lp_build_alloca(gallivm,
3044 uint_bld->vec_type,
3045 "total_emitted_vertices_ptr");
3046
3047 LLVMBuildStore(gallivm->builder, uint_bld->zero,
3048 bld->emitted_prims_vec_ptr);
3049 LLVMBuildStore(gallivm->builder, uint_bld->zero,
3050 bld->emitted_vertices_vec_ptr);
3051 LLVMBuildStore(gallivm->builder, uint_bld->zero,
3052 bld->total_emitted_vertices_vec_ptr);
3053 }
3054 }
3055
3056 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
3057 {
3058 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3059 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3060
3061 if (0) {
3062 /* for debugging */
3063 emit_dump_temps(bld);
3064 }
3065
3066 /* If we have indirect addressing in outputs we need to copy our alloca array
3067 * to the outputs slots specified by the caller */
3068 if (bld->gs_iface) {
3069 LLVMValueRef total_emitted_vertices_vec;
3070 LLVMValueRef emitted_prims_vec;
3071 /* implicit end_primitives, needed in case there are any unflushed
3072 vertices in the cache */
3073 end_primitive(NULL, bld_base, NULL);
3074
3075 total_emitted_vertices_vec =
3076 LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
3077 emitted_prims_vec =
3078 LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
3079
3080 bld->gs_iface->gs_epilogue(bld->gs_iface,
3081 &bld->bld_base,
3082 total_emitted_vertices_vec,
3083 emitted_prims_vec);
3084 } else {
3085 gather_outputs(bld);
3086 }
3087 }
3088
3089 void
3090 lp_build_tgsi_soa(struct gallivm_state *gallivm,
3091 const struct tgsi_token *tokens,
3092 struct lp_type type,
3093 struct lp_build_mask_context *mask,
3094 LLVMValueRef consts_ptr,
3095 const struct lp_bld_tgsi_system_values *system_values,
3096 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
3097 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
3098 struct lp_build_sampler_soa *sampler,
3099 const struct tgsi_shader_info *info,
3100 const struct lp_build_tgsi_gs_iface *gs_iface)
3101 {
3102 struct lp_build_tgsi_soa_context bld;
3103
3104 struct lp_type res_type;
3105
3106 assert(type.length <= LP_MAX_VECTOR_LENGTH);
3107 memset(&res_type, 0, sizeof res_type);
3108 res_type.width = type.width;
3109 res_type.length = type.length;
3110 res_type.sign = 1;
3111
3112 /* Setup build context */
3113 memset(&bld, 0, sizeof bld);
3114 lp_build_context_init(&bld.bld_base.base, gallivm, type);
3115 lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
3116 lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
3117 lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
3118 bld.mask = mask;
3119 bld.inputs = inputs;
3120 bld.outputs = outputs;
3121 bld.consts_ptr = consts_ptr;
3122 bld.sampler = sampler;
3123 bld.bld_base.info = info;
3124 bld.indirect_files = info->indirect_files;
3125
3126 bld.bld_base.soa = TRUE;
3127 bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
3128 bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
3129 bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
3130 bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
3131 bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
3132 bld.bld_base.emit_store = emit_store;
3133
3134 bld.bld_base.emit_declaration = lp_emit_declaration_soa;
3135 bld.bld_base.emit_immediate = lp_emit_immediate_soa;
3136
3137 bld.bld_base.emit_prologue = emit_prologue;
3138 bld.bld_base.emit_epilogue = emit_epilogue;
3139
3140 /* Set opcode actions */
3141 lp_set_default_actions_cpu(&bld.bld_base);
3142
3143 bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
3144 bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
3145 bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
3146 bld.bld_base.op_actions[TGSI_OPCODE_BREAKC].emit = breakc_emit;
3147 bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
3148 bld.bld_base.op_actions[TGSI_OPCODE_CASE].emit = case_emit;
3149 bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
3150 bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
3151 bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
3152 bld.bld_base.op_actions[TGSI_OPCODE_DEFAULT].emit = default_emit;
3153 bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
3154 bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
3155 bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
3156 bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
3157 bld.bld_base.op_actions[TGSI_OPCODE_ENDSWITCH].emit = endswitch_emit;
3158 bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
3159 bld.bld_base.op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
3160 bld.bld_base.op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
3161 bld.bld_base.op_actions[TGSI_OPCODE_KILP].emit = kilp_emit;
3162 bld.bld_base.op_actions[TGSI_OPCODE_NRM].emit = nrm_emit;
3163 bld.bld_base.op_actions[TGSI_OPCODE_NRM4].emit = nrm_emit;
3164 bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
3165 bld.bld_base.op_actions[TGSI_OPCODE_SWITCH].emit = switch_emit;
3166 bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
3167 bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
3168 bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
3169 bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
3170 bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
3171 bld.bld_base.op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
3172 bld.bld_base.op_actions[TGSI_OPCODE_TXF].emit = txf_emit;
3173 /* DX10 sampling ops */
3174 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE].emit = sample_emit;
3175 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_B].emit = sample_b_emit;
3176 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C].emit = sample_c_emit;
3177 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C_LZ].emit = sample_c_lz_emit;
3178 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_D].emit = sample_d_emit;
3179 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_I].emit = sample_i_emit;
3180 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_L].emit = sample_l_emit;
3181 bld.bld_base.op_actions[TGSI_OPCODE_SVIEWINFO].emit = sviewinfo_emit;
3182
3183 if (gs_iface) {
3184 /* There's no specific value for this because it should always
3185 * be set, but apps using ext_geometry_shader4 quite often
3186 * were forgetting so we're using MAX_VERTEX_VARYING from
3187 * that spec even though we could debug_assert if it's not
3188 * set, but that's a lot uglier. */
3189 uint max_output_vertices = 32;
3190 uint i = 0;
3191 /* inputs are always indirect with gs */
3192 bld.indirect_files |= (1 << TGSI_FILE_INPUT);
3193 bld.gs_iface = gs_iface;
3194 bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_gs_input;
3195 bld.bld_base.op_actions[TGSI_OPCODE_EMIT].emit = emit_vertex;
3196 bld.bld_base.op_actions[TGSI_OPCODE_ENDPRIM].emit = end_primitive;
3197
3198 for (i = 0; i < info->num_properties; ++i) {
3199 if (info->properties[i].name ==
3200 TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES) {
3201 max_output_vertices = info->properties[i].data[0];
3202 }
3203 }
3204 bld.max_output_vertices_vec =
3205 lp_build_const_int_vec(gallivm, bld.bld_base.uint_bld.type,
3206 max_output_vertices);
3207 }
3208
3209 lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.int_bld);
3210
3211 bld.system_values = *system_values;
3212
3213 lp_build_tgsi_llvm(&bld.bld_base, tokens);
3214
3215 if (0) {
3216 LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
3217 LLVMValueRef function = LLVMGetBasicBlockParent(block);
3218 debug_printf("11111111111111111111111111111 \n");
3219 tgsi_dump(tokens, 0);
3220 lp_debug_dump_value(function);
3221 debug_printf("2222222222222222222222222222 \n");
3222 }
3223
3224 if (0) {
3225 LLVMModuleRef module = LLVMGetGlobalParent(
3226 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
3227 LLVMDumpModule(module);
3228
3229 }
3230 }