projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallivm: handle nan's in min/max
[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_ext(float_bld, value, float_bld->zero,
1400 GALLIVM_NAN_RETURN_SECOND);
1401 value = lp_build_min_ext(float_bld, value, float_bld->one,
1402 GALLIVM_NAN_BEHAVIOR_UNDEFINED);
1403 break;
1404
1405 case TGSI_SAT_MINUS_PLUS_ONE:
1406 assert(dtype == TGSI_TYPE_FLOAT ||
1407 dtype == TGSI_TYPE_UNTYPED);
1408 value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1409 value = lp_build_max_ext(float_bld, value,
1410 lp_build_const_vec(gallivm, float_bld->type, -1.0),
1411 GALLIVM_NAN_RETURN_SECOND);
1412 value = lp_build_min_ext(float_bld, value, float_bld->one,
1413 GALLIVM_NAN_BEHAVIOR_UNDEFINED);
1414 break;
1415
1416 default:
1417 assert(0);
1418 }
1419
1420 if (reg->Register.Indirect) {
1421 indirect_index = get_indirect_index(bld,
1422 reg->Register.File,
1423 reg->Register.Index,
1424 &reg->Indirect);
1425 } else {
1426 assert(reg->Register.Index <=
1427 bld_base->info->file_max[reg->Register.File]);
1428 }
1429
1430 switch( reg->Register.File ) {
1431 case TGSI_FILE_OUTPUT:
1432 /* Outputs are always stored as floats */
1433 value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1434
1435 if (reg->Register.Indirect) {
1436 LLVMValueRef chan_vec =
1437 lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1438 LLVMValueRef length_vec =
1439 lp_build_const_int_vec(gallivm, uint_bld->type, float_bld->type.length);
1440 LLVMValueRef index_vec; /* indexes into the temp registers */
1441 LLVMValueRef outputs_array;
1442 LLVMValueRef pixel_offsets;
1443 LLVMTypeRef float_ptr_type;
1444 int i;
1445
1446 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1447 pixel_offsets = uint_bld->undef;
1448 for (i = 0; i < float_bld->type.length; i++) {
1449 LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1450 pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1451 ii, ii, "");
1452 }
1453
1454 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1455 index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1456 index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1457 index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1458 index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1459
1460 float_ptr_type =
1461 LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1462 outputs_array = LLVMBuildBitCast(builder, bld->outputs_array,
1463 float_ptr_type, "");
1464
1465 /* Scatter store values into temp registers */
1466 emit_mask_scatter(bld, outputs_array, index_vec, value,
1467 &bld->exec_mask, pred);
1468 }
1469 else {
1470 LLVMValueRef out_ptr = lp_get_output_ptr(bld, reg->Register.Index,
1471 chan_index);
1472 lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value, out_ptr);
1473 }
1474 break;
1475
1476 case TGSI_FILE_TEMPORARY:
1477 /* Temporaries are always stored as floats */
1478 value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1479
1480 if (reg->Register.Indirect) {
1481 LLVMValueRef chan_vec =
1482 lp_build_const_int_vec(gallivm, uint_bld->type, chan_index);
1483 LLVMValueRef length_vec =
1484 lp_build_const_int_vec(gallivm, uint_bld->type,
1485 float_bld->type.length);
1486 LLVMValueRef index_vec; /* indexes into the temp registers */
1487 LLVMValueRef temps_array;
1488 LLVMValueRef pixel_offsets;
1489 LLVMTypeRef float_ptr_type;
1490 int i;
1491
1492 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1493 pixel_offsets = uint_bld->undef;
1494 for (i = 0; i < float_bld->type.length; i++) {
1495 LLVMValueRef ii = lp_build_const_int32(gallivm, i);
1496 pixel_offsets = LLVMBuildInsertElement(builder, pixel_offsets,
1497 ii, ii, "");
1498 }
1499
1500 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1501 index_vec = lp_build_shl_imm(uint_bld, indirect_index, 2);
1502 index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
1503 index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
1504 index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
1505
1506 float_ptr_type =
1507 LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
1508 temps_array = LLVMBuildBitCast(builder, bld->temps_array,
1509 float_ptr_type, "");
1510
1511 /* Scatter store values into temp registers */
1512 emit_mask_scatter(bld, temps_array, index_vec, value,
1513 &bld->exec_mask, pred);
1514 }
1515 else {
1516 LLVMValueRef temp_ptr;
1517 temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index,
1518 chan_index);
1519 lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value, temp_ptr);
1520 }
1521 break;
1522
1523 case TGSI_FILE_ADDRESS:
1524 assert(dtype == TGSI_TYPE_SIGNED);
1525 assert(LLVMTypeOf(value) == int_bld->vec_type);
1526 value = LLVMBuildBitCast(builder, value, int_bld->vec_type, "");
1527 lp_exec_mask_store(&bld->exec_mask, int_bld, pred, value,
1528 bld->addr[reg->Register.Index][chan_index]);
1529 break;
1530
1531 case TGSI_FILE_PREDICATE:
1532 assert(LLVMTypeOf(value) == float_bld->vec_type);
1533 value = LLVMBuildBitCast(builder, value, float_bld->vec_type, "");
1534 lp_exec_mask_store(&bld->exec_mask, float_bld, pred, value,
1535 bld->preds[reg->Register.Index][chan_index]);
1536 break;
1537
1538 default:
1539 assert( 0 );
1540 }
1541
1542 (void)dtype;
1543 }
1544
1545 static void
1546 emit_store(
1547 struct lp_build_tgsi_context * bld_base,
1548 const struct tgsi_full_instruction * inst,
1549 const struct tgsi_opcode_info * info,
1550 LLVMValueRef dst[4])
1551
1552 {
1553 unsigned chan_index;
1554 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
1555
1556 if(info->num_dst) {
1557 LLVMValueRef pred[TGSI_NUM_CHANNELS];
1558
1559 emit_fetch_predicate( bld, inst, pred );
1560
1561 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
1562 emit_store_chan(bld_base, inst, 0, chan_index, pred[chan_index], dst[chan_index]);
1563 }
1564 }
1565 }
1566
1567 /**
1568 * High-level instruction translators.
1569 */
1570
1571 static void
1572 emit_tex( struct lp_build_tgsi_soa_context *bld,
1573 const struct tgsi_full_instruction *inst,
1574 enum lp_build_tex_modifier modifier,
1575 LLVMValueRef *texel)
1576 {
1577 unsigned unit;
1578 LLVMValueRef lod_bias, explicit_lod;
1579 LLVMValueRef oow = NULL;
1580 LLVMValueRef coords[4];
1581 LLVMValueRef offsets[3] = { NULL };
1582 struct lp_derivatives derivs;
1583 struct lp_derivatives *deriv_ptr = NULL;
1584 boolean scalar_lod;
1585 unsigned num_coords, num_derivs, num_offsets;
1586 unsigned i;
1587
1588 if (!bld->sampler) {
1589 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1590 for (i = 0; i < 4; i++) {
1591 texel[i] = bld->bld_base.base.undef;
1592 }
1593 return;
1594 }
1595
1596 switch (inst->Texture.Texture) {
1597 case TGSI_TEXTURE_1D:
1598 num_coords = 1;
1599 num_offsets = 1;
1600 num_derivs = 1;
1601 break;
1602 case TGSI_TEXTURE_1D_ARRAY:
1603 num_coords = 2;
1604 num_offsets = 1;
1605 num_derivs = 1;
1606 break;
1607 case TGSI_TEXTURE_2D:
1608 case TGSI_TEXTURE_RECT:
1609 num_coords = 2;
1610 num_offsets = 2;
1611 num_derivs = 2;
1612 break;
1613 case TGSI_TEXTURE_SHADOW1D:
1614 case TGSI_TEXTURE_SHADOW1D_ARRAY:
1615 num_coords = 3;
1616 num_offsets = 1;
1617 num_derivs = 1;
1618 break;
1619 case TGSI_TEXTURE_SHADOW2D:
1620 case TGSI_TEXTURE_SHADOWRECT:
1621 case TGSI_TEXTURE_2D_ARRAY:
1622 num_coords = 3;
1623 num_offsets = 2;
1624 num_derivs = 2;
1625 break;
1626 case TGSI_TEXTURE_CUBE:
1627 num_coords = 3;
1628 num_offsets = 2;
1629 num_derivs = 3;
1630 break;
1631 case TGSI_TEXTURE_3D:
1632 num_coords = 3;
1633 num_offsets = 3;
1634 num_derivs = 3;
1635 break;
1636 case TGSI_TEXTURE_SHADOW2D_ARRAY:
1637 num_coords = 4;
1638 num_offsets = 2;
1639 num_derivs = 2;
1640 break;
1641 case TGSI_TEXTURE_SHADOWCUBE:
1642 num_coords = 4;
1643 num_offsets = 2;
1644 num_derivs = 3;
1645 break;
1646 default:
1647 assert(0);
1648 return;
1649 }
1650
1651 /* Note lod and especially projected are illegal in a LOT of cases */
1652 if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1653 assert(num_coords < 4);
1654 lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1655 explicit_lod = NULL;
1656 }
1657 else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1658 assert(num_coords < 4);
1659 lod_bias = NULL;
1660 explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1661 }
1662 else {
1663 lod_bias = NULL;
1664 explicit_lod = NULL;
1665 }
1666
1667 if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED) {
1668 assert(num_coords < 4);
1669 oow = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1670 oow = lp_build_rcp(&bld->bld_base.base, oow);
1671 }
1672
1673 for (i = 0; i < num_coords; i++) {
1674 coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1675 if (modifier == LP_BLD_TEX_MODIFIER_PROJECTED)
1676 coords[i] = lp_build_mul(&bld->bld_base.base, coords[i], oow);
1677 }
1678 for (i = num_coords; i < 4; i++) {
1679 coords[i] = bld->bld_base.base.undef;
1680 }
1681
1682 if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1683 unsigned dim;
1684 for (dim = 0; dim < num_derivs; ++dim) {
1685 derivs.ddx[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 1, dim );
1686 derivs.ddy[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 2, dim );
1687 }
1688 deriv_ptr = &derivs;
1689 unit = inst->Src[3].Register.Index;
1690 } else {
1691 unit = inst->Src[1].Register.Index;
1692 }
1693
1694 /* some advanced gather instructions (txgo) would require 4 offsets */
1695 if (inst->Texture.NumOffsets == 1) {
1696 unsigned dim;
1697 for (dim = 0; dim < num_offsets; dim++) {
1698 offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
1699 }
1700 }
1701
1702 /* TODO: use scalar lod if explicit_lod, lod_bias or derivs are broadcasted scalars */
1703 scalar_lod = bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT;
1704
1705 bld->sampler->emit_fetch_texel(bld->sampler,
1706 bld->bld_base.base.gallivm,
1707 bld->bld_base.base.type,
1708 FALSE,
1709 unit, unit,
1710 coords,
1711 offsets,
1712 deriv_ptr,
1713 lod_bias, explicit_lod, scalar_lod,
1714 texel);
1715 }
1716
1717 static void
1718 emit_sample(struct lp_build_tgsi_soa_context *bld,
1719 const struct tgsi_full_instruction *inst,
1720 enum lp_build_tex_modifier modifier,
1721 boolean compare,
1722 LLVMValueRef *texel)
1723 {
1724 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
1725 unsigned texture_unit, sampler_unit;
1726 LLVMValueRef lod_bias, explicit_lod;
1727 LLVMValueRef coords[4];
1728 LLVMValueRef offsets[3] = { NULL };
1729 struct lp_derivatives derivs;
1730 struct lp_derivatives *deriv_ptr = NULL;
1731 boolean scalar_lod;
1732 unsigned num_coords, num_offsets, num_derivs;
1733 unsigned i;
1734
1735 if (!bld->sampler) {
1736 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1737 for (i = 0; i < 4; i++) {
1738 texel[i] = bld->bld_base.base.undef;
1739 }
1740 return;
1741 }
1742
1743 /*
1744 * unlike old-style tex opcodes the texture/sampler indices
1745 * always come from src1 and src2 respectively.
1746 */
1747 texture_unit = inst->Src[1].Register.Index;
1748 sampler_unit = inst->Src[2].Register.Index;
1749
1750 /*
1751 * Note inst->Texture.Texture will contain the number of offsets,
1752 * however the target information is NOT there and comes from the
1753 * declared sampler views instead.
1754 */
1755 switch (bld->sv[texture_unit].Resource) {
1756 case TGSI_TEXTURE_1D:
1757 num_coords = 1;
1758 num_offsets = 1;
1759 num_derivs = 1;
1760 break;
1761 case TGSI_TEXTURE_1D_ARRAY:
1762 num_coords = 2;
1763 num_offsets = 1;
1764 num_derivs = 1;
1765 break;
1766 case TGSI_TEXTURE_2D:
1767 case TGSI_TEXTURE_RECT:
1768 num_coords = 2;
1769 num_offsets = 2;
1770 num_derivs = 2;
1771 break;
1772 case TGSI_TEXTURE_2D_ARRAY:
1773 num_coords = 3;
1774 num_offsets = 2;
1775 num_derivs = 2;
1776 break;
1777 case TGSI_TEXTURE_CUBE:
1778 num_coords = 3;
1779 num_offsets = 2;
1780 num_derivs = 3;
1781 break;
1782 case TGSI_TEXTURE_3D:
1783 num_coords = 3;
1784 num_offsets = 3;
1785 num_derivs = 3;
1786 break;
1787 case TGSI_TEXTURE_CUBE_ARRAY:
1788 num_coords = 4;
1789 num_offsets = 2;
1790 num_derivs = 3;
1791 break;
1792 default:
1793 assert(0);
1794 return;
1795 }
1796
1797 if (modifier == LP_BLD_TEX_MODIFIER_LOD_BIAS) {
1798 lod_bias = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
1799 explicit_lod = NULL;
1800 }
1801 else if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_LOD) {
1802 lod_bias = NULL;
1803 explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
1804 }
1805 else if (modifier == LP_BLD_TEX_MODIFIER_LOD_ZERO) {
1806 lod_bias = NULL;
1807 /* XXX might be better to explicitly pass the level zero information */
1808 explicit_lod = lp_build_const_vec(gallivm, bld->bld_base.base.type, 0.0F);
1809 }
1810 else {
1811 lod_bias = NULL;
1812 explicit_lod = NULL;
1813 }
1814
1815 for (i = 0; i < num_coords; i++) {
1816 coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1817 }
1818 for (i = num_coords; i < 4; i++) {
1819 coords[i] = bld->bld_base.base.undef;
1820 }
1821 /*
1822 * XXX: whack shadow comparison value into place.
1823 * Should probably fix the interface for separate value
1824 * (it will not work for cube arrays if it is part of coords).
1825 */
1826 if (compare) {
1827 unsigned c_coord = num_coords > 2 ? 3 : 2;
1828 assert(num_coords < 4);
1829 coords[c_coord] = lp_build_emit_fetch( &bld->bld_base, inst, 3, 0 );
1830 }
1831
1832 if (modifier == LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV) {
1833 unsigned dim;
1834 for (dim = 0; dim < num_derivs; ++dim) {
1835 derivs.ddx[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 3, dim );
1836 derivs.ddy[dim] = lp_build_emit_fetch( &bld->bld_base, inst, 4, dim );
1837 }
1838 deriv_ptr = &derivs;
1839 }
1840
1841 /* some advanced gather instructions (txgo) would require 4 offsets */
1842 if (inst->Texture.NumOffsets == 1) {
1843 unsigned dim;
1844 for (dim = 0; dim < num_offsets; dim++) {
1845 offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
1846 }
1847 }
1848
1849 /* TODO: use scalar lod if explicit_lod, lod_bias or derivs are broadcasted scalars */
1850 scalar_lod = bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT;
1851
1852 bld->sampler->emit_fetch_texel(bld->sampler,
1853 bld->bld_base.base.gallivm,
1854 bld->bld_base.base.type,
1855 FALSE,
1856 texture_unit, sampler_unit,
1857 coords,
1858 offsets,
1859 deriv_ptr,
1860 lod_bias, explicit_lod, scalar_lod,
1861 texel);
1862 }
1863
1864 static void
1865 emit_fetch_texels( struct lp_build_tgsi_soa_context *bld,
1866 const struct tgsi_full_instruction *inst,
1867 LLVMValueRef *texel,
1868 boolean is_samplei)
1869 {
1870 unsigned unit, target;
1871 LLVMValueRef coord_undef = LLVMGetUndef(bld->bld_base.base.int_vec_type);
1872 LLVMValueRef explicit_lod = NULL;
1873 LLVMValueRef coords[3];
1874 LLVMValueRef offsets[3] = { NULL };
1875 boolean scalar_lod;
1876 unsigned num_coords;
1877 unsigned dims;
1878 unsigned i;
1879
1880 if (!bld->sampler) {
1881 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1882 for (i = 0; i < 4; i++) {
1883 texel[i] = coord_undef;
1884 }
1885 return;
1886 }
1887
1888 unit = inst->Src[1].Register.Index;
1889
1890 if (is_samplei) {
1891 target = bld->sv[unit].Resource;
1892 }
1893 else {
1894 target = inst->Texture.Texture;
1895 }
1896
1897 switch (target) {
1898 case TGSI_TEXTURE_1D:
1899 case TGSI_TEXTURE_BUFFER:
1900 num_coords = 1;
1901 dims = 1;
1902 break;
1903 case TGSI_TEXTURE_1D_ARRAY:
1904 num_coords = 2;
1905 dims = 1;
1906 break;
1907 case TGSI_TEXTURE_2D:
1908 case TGSI_TEXTURE_RECT:
1909 num_coords = 2;
1910 dims = 2;
1911 break;
1912 case TGSI_TEXTURE_2D_ARRAY:
1913 num_coords = 3;
1914 dims = 2;
1915 break;
1916 case TGSI_TEXTURE_3D:
1917 num_coords = 3;
1918 dims = 3;
1919 break;
1920 default:
1921 assert(0);
1922 return;
1923 }
1924
1925 /* always have lod except for buffers ? */
1926 if (target != TGSI_TEXTURE_BUFFER) {
1927 explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 3 );
1928 }
1929
1930 for (i = 0; i < num_coords; i++) {
1931 coords[i] = lp_build_emit_fetch( &bld->bld_base, inst, 0, i );
1932 }
1933 for (i = num_coords; i < 3; i++) {
1934 coords[i] = coord_undef;
1935 }
1936
1937 if (inst->Texture.NumOffsets == 1) {
1938 unsigned dim;
1939 for (dim = 0; dim < dims; dim++) {
1940 offsets[dim] = lp_build_emit_fetch_texoffset(&bld->bld_base, inst, 0, dim );
1941 }
1942 }
1943
1944 /* TODO: use scalar lod if explicit_lod is broadcasted scalar */
1945 scalar_lod = bld->bld_base.info->processor == TGSI_PROCESSOR_FRAGMENT;
1946
1947 bld->sampler->emit_fetch_texel(bld->sampler,
1948 bld->bld_base.base.gallivm,
1949 bld->bld_base.base.type,
1950 TRUE,
1951 unit, unit,
1952 coords,
1953 offsets,
1954 NULL,
1955 NULL, explicit_lod, scalar_lod,
1956 texel);
1957 }
1958
1959 static void
1960 emit_size_query( struct lp_build_tgsi_soa_context *bld,
1961 const struct tgsi_full_instruction *inst,
1962 LLVMValueRef *sizes_out,
1963 boolean is_sviewinfo)
1964 {
1965 LLVMValueRef explicit_lod;
1966 unsigned has_lod;
1967 unsigned i;
1968 unsigned unit = inst->Src[1].Register.Index;
1969 unsigned target;
1970
1971 if (is_sviewinfo) {
1972 target = bld->sv[unit].Resource;
1973 }
1974 else {
1975 target = inst->Texture.Texture;
1976 }
1977 switch (target) {
1978 case TGSI_TEXTURE_BUFFER:
1979 case TGSI_TEXTURE_RECT:
1980 case TGSI_TEXTURE_SHADOWRECT:
1981 has_lod = 0;
1982 break;
1983 default:
1984 has_lod = 1;
1985 break;
1986 }
1987
1988 if (!bld->sampler) {
1989 _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
1990 for (i = 0; i < 4; i++)
1991 sizes_out[i] = bld->bld_base.int_bld.undef;
1992 return;
1993 }
1994
1995 if (has_lod)
1996 explicit_lod = lp_build_emit_fetch( &bld->bld_base, inst, 0, 0 );
1997 else
1998 explicit_lod = NULL;
1999
2000 bld->sampler->emit_size_query(bld->sampler,
2001 bld->bld_base.base.gallivm,
2002 bld->bld_base.int_bld.type,
2003 unit,
2004 is_sviewinfo,
2005 explicit_lod,
2006 sizes_out);
2007 }
2008
2009 static boolean
2010 near_end_of_shader(struct lp_build_tgsi_soa_context *bld,
2011 int pc)
2012 {
2013 int i;
2014
2015 for (i = 0; i < 5; i++) {
2016 unsigned opcode;
2017
2018 if (pc + i >= bld->bld_base.info->num_instructions)
2019 return TRUE;
2020
2021 opcode = bld->bld_base.instructions[pc + i].Instruction.Opcode;
2022
2023 if (opcode == TGSI_OPCODE_END)
2024 return TRUE;
2025
2026 if (opcode == TGSI_OPCODE_TEX ||
2027 opcode == TGSI_OPCODE_TXP ||
2028 opcode == TGSI_OPCODE_TXD ||
2029 opcode == TGSI_OPCODE_TXB ||
2030 opcode == TGSI_OPCODE_TXL ||
2031 opcode == TGSI_OPCODE_TXF ||
2032 opcode == TGSI_OPCODE_TXQ ||
2033 opcode == TGSI_OPCODE_CAL ||
2034 opcode == TGSI_OPCODE_CALLNZ ||
2035 opcode == TGSI_OPCODE_IF ||
2036 opcode == TGSI_OPCODE_UIF ||
2037 opcode == TGSI_OPCODE_BGNLOOP ||
2038 opcode == TGSI_OPCODE_SWITCH)
2039 return FALSE;
2040 }
2041
2042 return TRUE;
2043 }
2044
2045
2046
2047 /**
2048 * Kill fragment if any of the src register values are negative.
2049 */
2050 static void
2051 emit_kill_if(
2052 struct lp_build_tgsi_soa_context *bld,
2053 const struct tgsi_full_instruction *inst,
2054 int pc)
2055 {
2056 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2057 const struct tgsi_full_src_register *reg = &inst->Src[0];
2058 LLVMValueRef terms[TGSI_NUM_CHANNELS];
2059 LLVMValueRef mask;
2060 unsigned chan_index;
2061
2062 memset(&terms, 0, sizeof terms);
2063
2064 TGSI_FOR_EACH_CHANNEL( chan_index ) {
2065 unsigned swizzle;
2066
2067 /* Unswizzle channel */
2068 swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
2069
2070 /* Check if the component has not been already tested. */
2071 assert(swizzle < TGSI_NUM_CHANNELS);
2072 if( !terms[swizzle] )
2073 /* TODO: change the comparison operator instead of setting the sign */
2074 terms[swizzle] = lp_build_emit_fetch(&bld->bld_base, inst, 0, chan_index );
2075 }
2076
2077 mask = NULL;
2078 TGSI_FOR_EACH_CHANNEL( chan_index ) {
2079 if(terms[chan_index]) {
2080 LLVMValueRef chan_mask;
2081
2082 /*
2083 * If term < 0 then mask = 0 else mask = ~0.
2084 */
2085 chan_mask = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, terms[chan_index], bld->bld_base.base.zero);
2086
2087 if(mask)
2088 mask = LLVMBuildAnd(builder, mask, chan_mask, "");
2089 else
2090 mask = chan_mask;
2091 }
2092 }
2093
2094 if(mask) {
2095 lp_build_mask_update(bld->mask, mask);
2096
2097 if (!near_end_of_shader(bld, pc))
2098 lp_build_mask_check(bld->mask);
2099 }
2100 }
2101
2102
2103 /**
2104 * Unconditional fragment kill.
2105 * The only predication is the execution mask which will apply if
2106 * we're inside a loop or conditional.
2107 */
2108 static void
2109 emit_kill(struct lp_build_tgsi_soa_context *bld,
2110 int pc)
2111 {
2112 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2113 LLVMValueRef mask;
2114
2115 /* For those channels which are "alive", disable fragment shader
2116 * execution.
2117 */
2118 if (bld->exec_mask.has_mask) {
2119 mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
2120 }
2121 else {
2122 LLVMValueRef zero = LLVMConstNull(bld->bld_base.base.int_vec_type);
2123 mask = zero;
2124 }
2125
2126 lp_build_mask_update(bld->mask, mask);
2127
2128 if (!near_end_of_shader(bld, pc))
2129 lp_build_mask_check(bld->mask);
2130 }
2131
2132
2133 /**
2134 * Emit code which will dump the value of all the temporary registers
2135 * to stdout.
2136 */
2137 static void
2138 emit_dump_temps(struct lp_build_tgsi_soa_context *bld)
2139 {
2140 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2141 LLVMBuilderRef builder = gallivm->builder;
2142 LLVMValueRef temp_ptr;
2143 LLVMValueRef i0 = lp_build_const_int32(gallivm, 0);
2144 LLVMValueRef i1 = lp_build_const_int32(gallivm, 1);
2145 LLVMValueRef i2 = lp_build_const_int32(gallivm, 2);
2146 LLVMValueRef i3 = lp_build_const_int32(gallivm, 3);
2147 int index;
2148 int n = bld->bld_base.info->file_max[TGSI_FILE_TEMPORARY];
2149
2150 for (index = 0; index < n; index++) {
2151 LLVMValueRef idx = lp_build_const_int32(gallivm, index);
2152 LLVMValueRef v[4][4], res;
2153 int chan;
2154
2155 lp_build_printf(gallivm, "TEMP[%d]:\n", idx);
2156
2157 for (chan = 0; chan < 4; chan++) {
2158 temp_ptr = lp_get_temp_ptr_soa(bld, index, chan);
2159 res = LLVMBuildLoad(builder, temp_ptr, "");
2160 v[chan][0] = LLVMBuildExtractElement(builder, res, i0, "");
2161 v[chan][1] = LLVMBuildExtractElement(builder, res, i1, "");
2162 v[chan][2] = LLVMBuildExtractElement(builder, res, i2, "");
2163 v[chan][3] = LLVMBuildExtractElement(builder, res, i3, "");
2164 }
2165
2166 lp_build_printf(gallivm, " X: %f %f %f %f\n",
2167 v[0][0], v[0][1], v[0][2], v[0][3]);
2168 lp_build_printf(gallivm, " Y: %f %f %f %f\n",
2169 v[1][0], v[1][1], v[1][2], v[1][3]);
2170 lp_build_printf(gallivm, " Z: %f %f %f %f\n",
2171 v[2][0], v[2][1], v[2][2], v[2][3]);
2172 lp_build_printf(gallivm, " W: %f %f %f %f\n",
2173 v[3][0], v[3][1], v[3][2], v[3][3]);
2174 }
2175 }
2176
2177
2178
2179 void
2180 lp_emit_declaration_soa(
2181 struct lp_build_tgsi_context *bld_base,
2182 const struct tgsi_full_declaration *decl)
2183 {
2184 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
2185 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2186 LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
2187 const unsigned first = decl->Range.First;
2188 const unsigned last = decl->Range.Last;
2189 unsigned idx, i;
2190
2191 for (idx = first; idx <= last; ++idx) {
2192 assert(last <= bld->bld_base.info->file_max[decl->Declaration.File]);
2193 switch (decl->Declaration.File) {
2194 case TGSI_FILE_TEMPORARY:
2195 assert(idx < LP_MAX_TGSI_TEMPS);
2196 if (!(bld->indirect_files & (1 << TGSI_FILE_TEMPORARY))) {
2197 for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2198 bld->temps[idx][i] = lp_build_alloca(gallivm, vec_type, "temp");
2199 }
2200 break;
2201
2202 case TGSI_FILE_OUTPUT:
2203 if (!(bld->indirect_files & (1 << TGSI_FILE_OUTPUT))) {
2204 for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2205 bld->outputs[idx][i] = lp_build_alloca(gallivm,
2206 vec_type, "output");
2207 }
2208 break;
2209
2210 case TGSI_FILE_ADDRESS:
2211 /* ADDR registers are only allocated with an integer LLVM IR type,
2212 * as they are guaranteed to always have integers.
2213 * XXX: Not sure if this exception is worthwhile (or the whole idea of
2214 * an ADDR register for that matter).
2215 */
2216 assert(idx < LP_MAX_TGSI_ADDRS);
2217 for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2218 bld->addr[idx][i] = lp_build_alloca(gallivm, bld_base->base.int_vec_type, "addr");
2219 break;
2220
2221 case TGSI_FILE_PREDICATE:
2222 assert(idx < LP_MAX_TGSI_PREDS);
2223 for (i = 0; i < TGSI_NUM_CHANNELS; i++)
2224 bld->preds[idx][i] = lp_build_alloca(gallivm, vec_type,
2225 "predicate");
2226 break;
2227
2228 case TGSI_FILE_SAMPLER_VIEW:
2229 /*
2230 * The target stored here MUST match whatever there actually
2231 * is in the set sampler views (what about return type?).
2232 */
2233 assert(idx < PIPE_MAX_SHADER_SAMPLER_VIEWS);
2234 bld->sv[idx] = decl->SamplerView;
2235 break;
2236
2237 default:
2238 /* don't need to declare other vars */
2239 break;
2240 }
2241 }
2242 }
2243
2244
2245 void lp_emit_immediate_soa(
2246 struct lp_build_tgsi_context *bld_base,
2247 const struct tgsi_full_immediate *imm)
2248 {
2249 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
2250 struct gallivm_state * gallivm = bld_base->base.gallivm;
2251
2252 /* simply copy the immediate values into the next immediates[] slot */
2253 unsigned i;
2254 const uint size = imm->Immediate.NrTokens - 1;
2255 assert(size <= 4);
2256 assert(bld->num_immediates < LP_MAX_TGSI_IMMEDIATES);
2257 switch (imm->Immediate.DataType) {
2258 case TGSI_IMM_FLOAT32:
2259 for( i = 0; i < size; ++i )
2260 bld->immediates[bld->num_immediates][i] =
2261 lp_build_const_vec(gallivm, bld_base->base.type, imm->u[i].Float);
2262
2263 break;
2264 case TGSI_IMM_UINT32:
2265 for( i = 0; i < size; ++i ) {
2266 LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->uint_bld.type, imm->u[i].Uint);
2267 bld->immediates[bld->num_immediates][i] =
2268 LLVMConstBitCast(tmp, bld_base->base.vec_type);
2269 }
2270
2271 break;
2272 case TGSI_IMM_INT32:
2273 for( i = 0; i < size; ++i ) {
2274 LLVMValueRef tmp = lp_build_const_vec(gallivm, bld_base->int_bld.type, imm->u[i].Int);
2275 bld->immediates[bld->num_immediates][i] =
2276 LLVMConstBitCast(tmp, bld_base->base.vec_type);
2277 }
2278
2279 break;
2280 }
2281 for( i = size; i < 4; ++i )
2282 bld->immediates[bld->num_immediates][i] = bld_base->base.undef;
2283
2284 if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
2285 unsigned index = bld->num_immediates;
2286 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
2287 LLVMBuilderRef builder = gallivm->builder;
2288 for (i = 0; i < 4; ++i ) {
2289 LLVMValueRef lindex = lp_build_const_int32(
2290 bld->bld_base.base.gallivm, index * 4 + i);
2291 LLVMValueRef imm_ptr = LLVMBuildGEP(builder,
2292 bld->imms_array, &lindex, 1, "");
2293 LLVMBuildStore(builder,
2294 bld->immediates[index][i],
2295 imm_ptr);
2296 }
2297 }
2298
2299 bld->num_immediates++;
2300 }
2301
2302 static void
2303 ddx_emit(
2304 const struct lp_build_tgsi_action * action,
2305 struct lp_build_tgsi_context * bld_base,
2306 struct lp_build_emit_data * emit_data)
2307 {
2308 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2309
2310 emit_fetch_deriv(bld, emit_data->args[0], NULL,
2311 &emit_data->output[emit_data->chan], NULL);
2312 }
2313
2314 static void
2315 ddy_emit(
2316 const struct lp_build_tgsi_action * action,
2317 struct lp_build_tgsi_context * bld_base,
2318 struct lp_build_emit_data * emit_data)
2319 {
2320 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2321
2322 emit_fetch_deriv(bld, emit_data->args[0], NULL, NULL,
2323 &emit_data->output[emit_data->chan]);
2324 }
2325
2326 static void
2327 kill_emit(
2328 const struct lp_build_tgsi_action * action,
2329 struct lp_build_tgsi_context * bld_base,
2330 struct lp_build_emit_data * emit_data)
2331 {
2332 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2333
2334 emit_kill(bld, bld_base->pc - 1);
2335 }
2336
2337 static void
2338 kill_if_emit(
2339 const struct lp_build_tgsi_action * action,
2340 struct lp_build_tgsi_context * bld_base,
2341 struct lp_build_emit_data * emit_data)
2342 {
2343 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2344
2345 emit_kill_if(bld, emit_data->inst, bld_base->pc - 1);
2346 }
2347
2348 static void
2349 tex_emit(
2350 const struct lp_build_tgsi_action * action,
2351 struct lp_build_tgsi_context * bld_base,
2352 struct lp_build_emit_data * emit_data)
2353 {
2354 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2355
2356 emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE, emit_data->output);
2357 }
2358
2359 static void
2360 txb_emit(
2361 const struct lp_build_tgsi_action * action,
2362 struct lp_build_tgsi_context * bld_base,
2363 struct lp_build_emit_data * emit_data)
2364 {
2365 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2366
2367 emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
2368 emit_data->output);
2369 }
2370
2371 static void
2372 txd_emit(
2373 const struct lp_build_tgsi_action * action,
2374 struct lp_build_tgsi_context * bld_base,
2375 struct lp_build_emit_data * emit_data)
2376 {
2377 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2378
2379 emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
2380 emit_data->output);
2381 }
2382
2383 static void
2384 txl_emit(
2385 const struct lp_build_tgsi_action * action,
2386 struct lp_build_tgsi_context * bld_base,
2387 struct lp_build_emit_data * emit_data)
2388 {
2389 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2390
2391 emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
2392 emit_data->output);
2393 }
2394
2395 static void
2396 txp_emit(
2397 const struct lp_build_tgsi_action * action,
2398 struct lp_build_tgsi_context * bld_base,
2399 struct lp_build_emit_data * emit_data)
2400 {
2401 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2402
2403 emit_tex(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_PROJECTED,
2404 emit_data->output);
2405 }
2406
2407 static void
2408 txq_emit(
2409 const struct lp_build_tgsi_action * action,
2410 struct lp_build_tgsi_context * bld_base,
2411 struct lp_build_emit_data * emit_data)
2412 {
2413 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2414
2415 emit_size_query(bld, emit_data->inst, emit_data->output, FALSE);
2416 }
2417
2418 static void
2419 txf_emit(
2420 const struct lp_build_tgsi_action * action,
2421 struct lp_build_tgsi_context * bld_base,
2422 struct lp_build_emit_data * emit_data)
2423 {
2424 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2425
2426 emit_fetch_texels(bld, emit_data->inst, emit_data->output, FALSE);
2427 }
2428
2429 static void
2430 sample_i_emit(
2431 const struct lp_build_tgsi_action * action,
2432 struct lp_build_tgsi_context * bld_base,
2433 struct lp_build_emit_data * emit_data)
2434 {
2435 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2436
2437 emit_fetch_texels(bld, emit_data->inst, emit_data->output, TRUE);
2438 }
2439
2440 static void
2441 sample_emit(
2442 const struct lp_build_tgsi_action * action,
2443 struct lp_build_tgsi_context * bld_base,
2444 struct lp_build_emit_data * emit_data)
2445 {
2446 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2447
2448 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2449 FALSE, emit_data->output);
2450 }
2451
2452 static void
2453 sample_b_emit(
2454 const struct lp_build_tgsi_action * action,
2455 struct lp_build_tgsi_context * bld_base,
2456 struct lp_build_emit_data * emit_data)
2457 {
2458 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2459
2460 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_BIAS,
2461 FALSE, emit_data->output);
2462 }
2463
2464 static void
2465 sample_c_emit(
2466 const struct lp_build_tgsi_action * action,
2467 struct lp_build_tgsi_context * bld_base,
2468 struct lp_build_emit_data * emit_data)
2469 {
2470 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2471
2472 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_NONE,
2473 TRUE, emit_data->output);
2474 }
2475
2476 static void
2477 sample_c_lz_emit(
2478 const struct lp_build_tgsi_action * action,
2479 struct lp_build_tgsi_context * bld_base,
2480 struct lp_build_emit_data * emit_data)
2481 {
2482 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2483
2484 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_LOD_ZERO,
2485 TRUE, emit_data->output);
2486 }
2487
2488 static void
2489 sample_d_emit(
2490 const struct lp_build_tgsi_action * action,
2491 struct lp_build_tgsi_context * bld_base,
2492 struct lp_build_emit_data * emit_data)
2493 {
2494 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2495
2496 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
2497 FALSE, emit_data->output);
2498 }
2499
2500 static void
2501 sample_l_emit(
2502 const struct lp_build_tgsi_action * action,
2503 struct lp_build_tgsi_context * bld_base,
2504 struct lp_build_emit_data * emit_data)
2505 {
2506 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2507
2508 emit_sample(bld, emit_data->inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
2509 FALSE, emit_data->output);
2510 }
2511
2512 static void
2513 sviewinfo_emit(
2514 const struct lp_build_tgsi_action * action,
2515 struct lp_build_tgsi_context * bld_base,
2516 struct lp_build_emit_data * emit_data)
2517 {
2518 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2519
2520 emit_size_query(bld, emit_data->inst, emit_data->output, TRUE);
2521 }
2522
2523 static LLVMValueRef
2524 mask_to_one_vec(struct lp_build_tgsi_context *bld_base)
2525 {
2526 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2527 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2528 LLVMValueRef one_vec = bld_base->int_bld.one;
2529 struct lp_exec_mask *exec_mask = &bld->exec_mask;
2530
2531 if (exec_mask->has_mask) {
2532 one_vec = LLVMBuildAnd(builder, one_vec, exec_mask->exec_mask, "");
2533 }
2534 one_vec = LLVMBuildAnd(builder, one_vec,
2535 lp_build_mask_value(bld->mask), "");
2536 return one_vec;
2537 }
2538
2539 static void
2540 increment_vec_ptr_by_mask(struct lp_build_tgsi_context * bld_base,
2541 LLVMValueRef ptr,
2542 LLVMValueRef mask)
2543 {
2544 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2545
2546 LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
2547
2548 current_vec = LLVMBuildAdd(builder, current_vec, mask, "");
2549
2550 LLVMBuildStore(builder, current_vec, ptr);
2551 }
2552
2553 static void
2554 clear_uint_vec_ptr_from_mask(struct lp_build_tgsi_context * bld_base,
2555 LLVMValueRef ptr,
2556 LLVMValueRef mask)
2557 {
2558 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2559
2560 LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
2561 LLVMValueRef full_mask = lp_build_cmp(&bld_base->uint_bld,
2562 PIPE_FUNC_NOTEQUAL,
2563 mask,
2564 bld_base->uint_bld.zero);
2565
2566 current_vec = lp_build_select(&bld_base->uint_bld,
2567 full_mask,
2568 bld_base->uint_bld.zero,
2569 current_vec);
2570
2571 LLVMBuildStore(builder, current_vec, ptr);
2572 }
2573
2574 static LLVMValueRef
2575 clamp_mask_to_max_output_vertices(struct lp_build_tgsi_soa_context * bld,
2576 LLVMValueRef current_mask_vec,
2577 LLVMValueRef total_emitted_vertices_vec)
2578 {
2579 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2580 struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
2581 LLVMValueRef max_mask = lp_build_cmp(uint_bld, PIPE_FUNC_LESS,
2582 total_emitted_vertices_vec,
2583 bld->max_output_vertices_vec);
2584
2585 return LLVMBuildAnd(builder, current_mask_vec, max_mask, "");
2586 }
2587
2588 static void
2589 emit_vertex(
2590 const struct lp_build_tgsi_action * action,
2591 struct lp_build_tgsi_context * bld_base,
2592 struct lp_build_emit_data * emit_data)
2593 {
2594 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2595 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2596
2597 if (bld->gs_iface->emit_vertex) {
2598 LLVMValueRef masked_ones = mask_to_one_vec(bld_base);
2599 LLVMValueRef total_emitted_vertices_vec =
2600 LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
2601 masked_ones = clamp_mask_to_max_output_vertices(bld, masked_ones,
2602 total_emitted_vertices_vec);
2603 gather_outputs(bld);
2604 bld->gs_iface->emit_vertex(bld->gs_iface, &bld->bld_base,
2605 bld->outputs,
2606 total_emitted_vertices_vec);
2607 increment_vec_ptr_by_mask(bld_base, bld->emitted_vertices_vec_ptr,
2608 masked_ones);
2609 increment_vec_ptr_by_mask(bld_base, bld->total_emitted_vertices_vec_ptr,
2610 masked_ones);
2611 #if DUMP_GS_EMITS
2612 lp_build_print_value(bld->bld_base.base.gallivm,
2613 " +++ emit vertex masked ones = ",
2614 masked_ones);
2615 lp_build_print_value(bld->bld_base.base.gallivm,
2616 " +++ emit vertex emitted = ",
2617 total_emitted_vertices_vec);
2618 #endif
2619 }
2620 }
2621
2622
2623 static void
2624 end_primitive_masked(struct lp_build_tgsi_context * bld_base,
2625 LLVMValueRef masked_ones)
2626 {
2627 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2628 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
2629
2630 if (bld->gs_iface->end_primitive) {
2631 LLVMValueRef emitted_vertices_vec =
2632 LLVMBuildLoad(builder, bld->emitted_vertices_vec_ptr, "");
2633 LLVMValueRef emitted_prims_vec =
2634 LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
2635
2636 bld->gs_iface->end_primitive(bld->gs_iface, &bld->bld_base,
2637 emitted_vertices_vec,
2638 emitted_prims_vec);
2639
2640 #if DUMP_GS_EMITS
2641 lp_build_print_value(bld->bld_base.base.gallivm,
2642 " +++ end prim masked ones = ",
2643 masked_ones);
2644 lp_build_print_value(bld->bld_base.base.gallivm,
2645 " +++ end prim emitted verts1 = ",
2646 emitted_vertices_vec);
2647 lp_build_print_value(bld->bld_base.base.gallivm,
2648 " +++ end prim emitted prims1 = ",
2649 LLVMBuildLoad(builder,
2650 bld->emitted_prims_vec_ptr, ""));
2651 #endif
2652 increment_vec_ptr_by_mask(bld_base, bld->emitted_prims_vec_ptr,
2653 masked_ones);
2654 clear_uint_vec_ptr_from_mask(bld_base, bld->emitted_vertices_vec_ptr,
2655 masked_ones);
2656 #if DUMP_GS_EMITS
2657 lp_build_print_value(bld->bld_base.base.gallivm,
2658 " +++ end prim emitted verts2 = ",
2659 LLVMBuildLoad(builder,
2660 bld->emitted_vertices_vec_ptr, ""));
2661 #endif
2662 }
2663
2664 }
2665
2666 static void
2667 end_primitive(
2668 const struct lp_build_tgsi_action * action,
2669 struct lp_build_tgsi_context * bld_base,
2670 struct lp_build_emit_data * emit_data)
2671 {
2672 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2673
2674 if (bld->gs_iface->end_primitive) {
2675 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2676 LLVMValueRef masked_ones = mask_to_one_vec(bld_base);
2677 struct lp_build_context *uint_bld = &bld_base->uint_bld;
2678 LLVMValueRef emitted_verts = LLVMBuildLoad(
2679 builder, bld->emitted_vertices_vec_ptr, "");
2680 LLVMValueRef emitted_mask = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
2681 emitted_verts,
2682 uint_bld->zero);
2683 /* We need to combine the current execution mask with the mask
2684 telling us which, if any, execution slots actually have
2685 unemitted primitives, this way we make sure that end_primitives
2686 executes only on the paths that have unflushed vertices */
2687 masked_ones = LLVMBuildAnd(builder, masked_ones, emitted_mask, "");
2688
2689 end_primitive_masked(bld_base, masked_ones);
2690 }
2691 }
2692
2693 static void
2694 cal_emit(
2695 const struct lp_build_tgsi_action * action,
2696 struct lp_build_tgsi_context * bld_base,
2697 struct lp_build_emit_data * emit_data)
2698 {
2699 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2700
2701 lp_exec_mask_call(&bld->exec_mask, emit_data->inst->Label.Label,
2702 &bld_base->pc);
2703 }
2704
2705 static void
2706 ret_emit(
2707 const struct lp_build_tgsi_action * action,
2708 struct lp_build_tgsi_context * bld_base,
2709 struct lp_build_emit_data * emit_data)
2710 {
2711 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2712
2713 lp_exec_mask_ret(&bld->exec_mask, &bld_base->pc);
2714 }
2715
2716 static void
2717 brk_emit(
2718 const struct lp_build_tgsi_action * action,
2719 struct lp_build_tgsi_context * bld_base,
2720 struct lp_build_emit_data * emit_data)
2721 {
2722 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2723
2724 lp_exec_break(&bld->exec_mask, bld_base);
2725 }
2726
2727 static void
2728 breakc_emit(
2729 const struct lp_build_tgsi_action * action,
2730 struct lp_build_tgsi_context * bld_base,
2731 struct lp_build_emit_data * emit_data)
2732 {
2733 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2734 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
2735 struct lp_build_context *uint_bld = &bld_base->uint_bld;
2736 LLVMValueRef unsigned_cond =
2737 LLVMBuildBitCast(builder, emit_data->args[0], uint_bld->vec_type, "");
2738 LLVMValueRef cond = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
2739 unsigned_cond,
2740 uint_bld->zero);
2741
2742 lp_exec_break_condition(&bld->exec_mask, cond);
2743 }
2744
2745 static void
2746 if_emit(
2747 const struct lp_build_tgsi_action * action,
2748 struct lp_build_tgsi_context * bld_base,
2749 struct lp_build_emit_data * emit_data)
2750 {
2751 LLVMValueRef tmp;
2752 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2753
2754 tmp = lp_build_cmp(&bld_base->base, PIPE_FUNC_NOTEQUAL,
2755 emit_data->args[0], bld->bld_base.base.zero);
2756 lp_exec_mask_cond_push(&bld->exec_mask, tmp);
2757 }
2758
2759 static void
2760 uif_emit(
2761 const struct lp_build_tgsi_action * action,
2762 struct lp_build_tgsi_context * bld_base,
2763 struct lp_build_emit_data * emit_data)
2764 {
2765 LLVMValueRef tmp;
2766 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2767 struct lp_build_context *uint_bld = &bld_base->uint_bld;
2768
2769 tmp = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL,
2770 emit_data->args[0], uint_bld->zero);
2771 lp_exec_mask_cond_push(&bld->exec_mask, tmp);
2772 }
2773
2774 static void
2775 case_emit(
2776 const struct lp_build_tgsi_action * action,
2777 struct lp_build_tgsi_context * bld_base,
2778 struct lp_build_emit_data * emit_data)
2779 {
2780 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2781
2782 lp_exec_case(&bld->exec_mask, emit_data->args[0]);
2783 }
2784
2785 static void
2786 default_emit(
2787 const struct lp_build_tgsi_action * action,
2788 struct lp_build_tgsi_context * bld_base,
2789 struct lp_build_emit_data * emit_data)
2790 {
2791 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2792
2793 lp_exec_default(&bld->exec_mask, bld_base);
2794 }
2795
2796 static void
2797 switch_emit(
2798 const struct lp_build_tgsi_action * action,
2799 struct lp_build_tgsi_context * bld_base,
2800 struct lp_build_emit_data * emit_data)
2801 {
2802 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2803
2804 lp_exec_switch(&bld->exec_mask, emit_data->args[0]);
2805 }
2806
2807 static void
2808 endswitch_emit(
2809 const struct lp_build_tgsi_action * action,
2810 struct lp_build_tgsi_context * bld_base,
2811 struct lp_build_emit_data * emit_data)
2812 {
2813 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2814
2815 lp_exec_endswitch(&bld->exec_mask, bld_base);
2816 }
2817
2818 static void
2819 bgnloop_emit(
2820 const struct lp_build_tgsi_action * action,
2821 struct lp_build_tgsi_context * bld_base,
2822 struct lp_build_emit_data * emit_data)
2823 {
2824 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2825
2826 lp_exec_bgnloop(&bld->exec_mask);
2827 }
2828
2829 static void
2830 bgnsub_emit(
2831 const struct lp_build_tgsi_action * action,
2832 struct lp_build_tgsi_context * bld_base,
2833 struct lp_build_emit_data * emit_data)
2834 {
2835 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2836
2837 lp_exec_mask_bgnsub(&bld->exec_mask);
2838 }
2839
2840 static void
2841 else_emit(
2842 const struct lp_build_tgsi_action * action,
2843 struct lp_build_tgsi_context * bld_base,
2844 struct lp_build_emit_data * emit_data)
2845 {
2846 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2847
2848 lp_exec_mask_cond_invert(&bld->exec_mask);
2849 }
2850
2851 static void
2852 endif_emit(
2853 const struct lp_build_tgsi_action * action,
2854 struct lp_build_tgsi_context * bld_base,
2855 struct lp_build_emit_data * emit_data)
2856 {
2857 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2858
2859 lp_exec_mask_cond_pop(&bld->exec_mask);
2860 }
2861
2862 static void
2863 endloop_emit(
2864 const struct lp_build_tgsi_action * action,
2865 struct lp_build_tgsi_context * bld_base,
2866 struct lp_build_emit_data * emit_data)
2867 {
2868 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2869
2870 lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
2871 }
2872
2873 static void
2874 endsub_emit(
2875 const struct lp_build_tgsi_action * action,
2876 struct lp_build_tgsi_context * bld_base,
2877 struct lp_build_emit_data * emit_data)
2878 {
2879 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2880
2881 lp_exec_mask_endsub(&bld->exec_mask, &bld_base->pc);
2882 }
2883
2884 static void
2885 cont_emit(
2886 const struct lp_build_tgsi_action * action,
2887 struct lp_build_tgsi_context * bld_base,
2888 struct lp_build_emit_data * emit_data)
2889 {
2890 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2891
2892 lp_exec_continue(&bld->exec_mask);
2893 }
2894
2895 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
2896 *
2897 * XXX: What do the comments about xmm registers mean? Maybe they are left over
2898 * from old code, but there is no garauntee that LLVM will use those registers
2899 * for this code.
2900 *
2901 * XXX: There should be no calls to lp_build_emit_fetch in this function. This
2902 * should be handled by the emit_data->fetch_args function. */
2903 static void
2904 nrm_emit(
2905 const struct lp_build_tgsi_action * action,
2906 struct lp_build_tgsi_context * bld_base,
2907 struct lp_build_emit_data * emit_data)
2908 {
2909 LLVMValueRef tmp0, tmp1;
2910 LLVMValueRef tmp4 = NULL;
2911 LLVMValueRef tmp5 = NULL;
2912 LLVMValueRef tmp6 = NULL;
2913 LLVMValueRef tmp7 = NULL;
2914 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2915
2916 uint dims = (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_NRM) ? 3 : 4;
2917
2918 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) ||
2919 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y) ||
2920 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z) ||
2921 (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 4)) {
2922
2923 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
2924
2925 /* xmm4 = src.x */
2926 /* xmm0 = src.x * src.x */
2927 tmp0 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_X);
2928 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
2929 tmp4 = tmp0;
2930 }
2931 tmp0 = lp_build_mul( &bld->bld_base.base, tmp0, tmp0);
2932
2933 /* xmm5 = src.y */
2934 /* xmm0 = xmm0 + src.y * src.y */
2935 tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Y);
2936 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
2937 tmp5 = tmp1;
2938 }
2939 tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2940 tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2941
2942 /* xmm6 = src.z */
2943 /* xmm0 = xmm0 + src.z * src.z */
2944 tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_Z);
2945 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
2946 tmp6 = tmp1;
2947 }
2948 tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2949 tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2950
2951 if (dims == 4) {
2952 /* xmm7 = src.w */
2953 /* xmm0 = xmm0 + src.w * src.w */
2954 tmp1 = lp_build_emit_fetch(&bld->bld_base, emit_data->inst, 0, TGSI_CHAN_W);
2955 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W)) {
2956 tmp7 = tmp1;
2957 }
2958 tmp1 = lp_build_mul( &bld->bld_base.base, tmp1, tmp1);
2959 tmp0 = lp_build_add( &bld->bld_base.base, tmp0, tmp1);
2960 }
2961 /* xmm1 = 1 / sqrt(xmm0) */
2962 tmp1 = lp_build_rsqrt( &bld->bld_base.base, tmp0);
2963 /* dst.x = xmm1 * src.x */
2964 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X)) {
2965 emit_data->output[TGSI_CHAN_X] = lp_build_mul( &bld->bld_base.base, tmp4, tmp1);
2966 }
2967 /* dst.y = xmm1 * src.y */
2968 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Y)) {
2969 emit_data->output[TGSI_CHAN_Y] = lp_build_mul( &bld->bld_base.base, tmp5, tmp1);
2970 }
2971
2972 /* dst.z = xmm1 * src.z */
2973 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_Z)) {
2974 emit_data->output[TGSI_CHAN_Z] = lp_build_mul( &bld->bld_base.base, tmp6, tmp1);
2975 }
2976 /* dst.w = xmm1 * src.w */
2977 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_X) && dims == 4) {
2978 emit_data->output[TGSI_CHAN_W] = lp_build_mul( &bld->bld_base.base, tmp7, tmp1);
2979 }
2980 }
2981
2982 /* dst.w = 1.0 */
2983 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data->inst, TGSI_CHAN_W) && dims == 3) {
2984 emit_data->output[TGSI_CHAN_W] = bld->bld_base.base.one;
2985 }
2986 }
2987
2988 static void emit_prologue(struct lp_build_tgsi_context * bld_base)
2989 {
2990 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
2991 struct gallivm_state * gallivm = bld_base->base.gallivm;
2992
2993 if (bld->indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
2994 LLVMValueRef array_size =
2995 lp_build_const_int32(gallivm,
2996 bld_base->info->file_max[TGSI_FILE_TEMPORARY] * 4 + 4);
2997 bld->temps_array = lp_build_array_alloca(gallivm,
2998 bld_base->base.vec_type, array_size,
2999 "temp_array");
3000 }
3001
3002 if (bld->indirect_files & (1 << TGSI_FILE_OUTPUT)) {
3003 LLVMValueRef array_size =
3004 lp_build_const_int32(gallivm,
3005 bld_base->info->file_max[TGSI_FILE_OUTPUT] * 4 + 4);
3006 bld->outputs_array = lp_build_array_alloca(gallivm,
3007 bld_base->base.vec_type, array_size,
3008 "output_array");
3009 }
3010
3011 if (bld->indirect_files & (1 << TGSI_FILE_IMMEDIATE)) {
3012 LLVMValueRef array_size =
3013 lp_build_const_int32(gallivm,
3014 bld_base->info->file_max[TGSI_FILE_IMMEDIATE] * 4 + 4);
3015 bld->imms_array = lp_build_array_alloca(gallivm,
3016 bld_base->base.vec_type, array_size,
3017 "imms_array");
3018 }
3019
3020 /* If we have indirect addressing in inputs we need to copy them into
3021 * our alloca array to be able to iterate over them */
3022 if (bld->indirect_files & (1 << TGSI_FILE_INPUT) && !bld->gs_iface) {
3023 unsigned index, chan;
3024 LLVMTypeRef vec_type = bld_base->base.vec_type;
3025 LLVMValueRef array_size = lp_build_const_int32(gallivm,
3026 bld_base->info->file_max[TGSI_FILE_INPUT]*4 + 4);
3027 bld->inputs_array = lp_build_array_alloca(gallivm,
3028 vec_type, array_size,
3029 "input_array");
3030
3031 assert(bld_base->info->num_inputs
3032 <= bld_base->info->file_max[TGSI_FILE_INPUT] + 1);
3033
3034 for (index = 0; index < bld_base->info->num_inputs; ++index) {
3035 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
3036 LLVMValueRef lindex =
3037 lp_build_const_int32(gallivm, index * 4 + chan);
3038 LLVMValueRef input_ptr =
3039 LLVMBuildGEP(gallivm->builder, bld->inputs_array,
3040 &lindex, 1, "");
3041 LLVMValueRef value = bld->inputs[index][chan];
3042 if (value)
3043 LLVMBuildStore(gallivm->builder, value, input_ptr);
3044 }
3045 }
3046 }
3047
3048 if (bld->gs_iface) {
3049 struct lp_build_context *uint_bld = &bld->bld_base.uint_bld;
3050 bld->emitted_prims_vec_ptr =
3051 lp_build_alloca(gallivm,
3052 uint_bld->vec_type,
3053 "emitted_prims_ptr");
3054 bld->emitted_vertices_vec_ptr =
3055 lp_build_alloca(gallivm,
3056 uint_bld->vec_type,
3057 "emitted_vertices_ptr");
3058 bld->total_emitted_vertices_vec_ptr =
3059 lp_build_alloca(gallivm,
3060 uint_bld->vec_type,
3061 "total_emitted_vertices_ptr");
3062
3063 LLVMBuildStore(gallivm->builder, uint_bld->zero,
3064 bld->emitted_prims_vec_ptr);
3065 LLVMBuildStore(gallivm->builder, uint_bld->zero,
3066 bld->emitted_vertices_vec_ptr);
3067 LLVMBuildStore(gallivm->builder, uint_bld->zero,
3068 bld->total_emitted_vertices_vec_ptr);
3069 }
3070 }
3071
3072 static void emit_epilogue(struct lp_build_tgsi_context * bld_base)
3073 {
3074 struct lp_build_tgsi_soa_context * bld = lp_soa_context(bld_base);
3075 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
3076
3077 if (0) {
3078 /* for debugging */
3079 emit_dump_temps(bld);
3080 }
3081
3082 /* If we have indirect addressing in outputs we need to copy our alloca array
3083 * to the outputs slots specified by the caller */
3084 if (bld->gs_iface) {
3085 LLVMValueRef total_emitted_vertices_vec;
3086 LLVMValueRef emitted_prims_vec;
3087 /* implicit end_primitives, needed in case there are any unflushed
3088 vertices in the cache */
3089 end_primitive(NULL, bld_base, NULL);
3090
3091 total_emitted_vertices_vec =
3092 LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
3093 emitted_prims_vec =
3094 LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
3095
3096 bld->gs_iface->gs_epilogue(bld->gs_iface,
3097 &bld->bld_base,
3098 total_emitted_vertices_vec,
3099 emitted_prims_vec);
3100 } else {
3101 gather_outputs(bld);
3102 }
3103 }
3104
3105 void
3106 lp_build_tgsi_soa(struct gallivm_state *gallivm,
3107 const struct tgsi_token *tokens,
3108 struct lp_type type,
3109 struct lp_build_mask_context *mask,
3110 LLVMValueRef consts_ptr,
3111 const struct lp_bld_tgsi_system_values *system_values,
3112 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
3113 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
3114 struct lp_build_sampler_soa *sampler,
3115 const struct tgsi_shader_info *info,
3116 const struct lp_build_tgsi_gs_iface *gs_iface)
3117 {
3118 struct lp_build_tgsi_soa_context bld;
3119
3120 struct lp_type res_type;
3121
3122 assert(type.length <= LP_MAX_VECTOR_LENGTH);
3123 memset(&res_type, 0, sizeof res_type);
3124 res_type.width = type.width;
3125 res_type.length = type.length;
3126 res_type.sign = 1;
3127
3128 /* Setup build context */
3129 memset(&bld, 0, sizeof bld);
3130 lp_build_context_init(&bld.bld_base.base, gallivm, type);
3131 lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
3132 lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
3133 lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
3134 bld.mask = mask;
3135 bld.inputs = inputs;
3136 bld.outputs = outputs;
3137 bld.consts_ptr = consts_ptr;
3138 bld.sampler = sampler;
3139 bld.bld_base.info = info;
3140 bld.indirect_files = info->indirect_files;
3141
3142 bld.bld_base.soa = TRUE;
3143 bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
3144 bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
3145 bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
3146 bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
3147 bld.bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = emit_fetch_system_value;
3148 bld.bld_base.emit_store = emit_store;
3149
3150 bld.bld_base.emit_declaration = lp_emit_declaration_soa;
3151 bld.bld_base.emit_immediate = lp_emit_immediate_soa;
3152
3153 bld.bld_base.emit_prologue = emit_prologue;
3154 bld.bld_base.emit_epilogue = emit_epilogue;
3155
3156 /* Set opcode actions */
3157 lp_set_default_actions_cpu(&bld.bld_base);
3158
3159 bld.bld_base.op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
3160 bld.bld_base.op_actions[TGSI_OPCODE_BGNSUB].emit = bgnsub_emit;
3161 bld.bld_base.op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
3162 bld.bld_base.op_actions[TGSI_OPCODE_BREAKC].emit = breakc_emit;
3163 bld.bld_base.op_actions[TGSI_OPCODE_CAL].emit = cal_emit;
3164 bld.bld_base.op_actions[TGSI_OPCODE_CASE].emit = case_emit;
3165 bld.bld_base.op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
3166 bld.bld_base.op_actions[TGSI_OPCODE_DDX].emit = ddx_emit;
3167 bld.bld_base.op_actions[TGSI_OPCODE_DDY].emit = ddy_emit;
3168 bld.bld_base.op_actions[TGSI_OPCODE_DEFAULT].emit = default_emit;
3169 bld.bld_base.op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
3170 bld.bld_base.op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
3171 bld.bld_base.op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
3172 bld.bld_base.op_actions[TGSI_OPCODE_ENDSUB].emit = endsub_emit;
3173 bld.bld_base.op_actions[TGSI_OPCODE_ENDSWITCH].emit = endswitch_emit;
3174 bld.bld_base.op_actions[TGSI_OPCODE_IF].emit = if_emit;
3175 bld.bld_base.op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
3176 bld.bld_base.op_actions[TGSI_OPCODE_KILL_IF].emit = kill_if_emit;
3177 bld.bld_base.op_actions[TGSI_OPCODE_KILL].emit = kill_emit;
3178 bld.bld_base.op_actions[TGSI_OPCODE_NRM].emit = nrm_emit;
3179 bld.bld_base.op_actions[TGSI_OPCODE_NRM4].emit = nrm_emit;
3180 bld.bld_base.op_actions[TGSI_OPCODE_RET].emit = ret_emit;
3181 bld.bld_base.op_actions[TGSI_OPCODE_SWITCH].emit = switch_emit;
3182 bld.bld_base.op_actions[TGSI_OPCODE_TEX].emit = tex_emit;
3183 bld.bld_base.op_actions[TGSI_OPCODE_TXB].emit = txb_emit;
3184 bld.bld_base.op_actions[TGSI_OPCODE_TXD].emit = txd_emit;
3185 bld.bld_base.op_actions[TGSI_OPCODE_TXL].emit = txl_emit;
3186 bld.bld_base.op_actions[TGSI_OPCODE_TXP].emit = txp_emit;
3187 bld.bld_base.op_actions[TGSI_OPCODE_TXQ].emit = txq_emit;
3188 bld.bld_base.op_actions[TGSI_OPCODE_TXF].emit = txf_emit;
3189 /* DX10 sampling ops */
3190 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE].emit = sample_emit;
3191 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_B].emit = sample_b_emit;
3192 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C].emit = sample_c_emit;
3193 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_C_LZ].emit = sample_c_lz_emit;
3194 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_D].emit = sample_d_emit;
3195 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_I].emit = sample_i_emit;
3196 bld.bld_base.op_actions[TGSI_OPCODE_SAMPLE_L].emit = sample_l_emit;
3197 bld.bld_base.op_actions[TGSI_OPCODE_SVIEWINFO].emit = sviewinfo_emit;
3198
3199 if (gs_iface) {
3200 /* There's no specific value for this because it should always
3201 * be set, but apps using ext_geometry_shader4 quite often
3202 * were forgetting so we're using MAX_VERTEX_VARYING from
3203 * that spec even though we could debug_assert if it's not
3204 * set, but that's a lot uglier. */
3205 uint max_output_vertices = 32;
3206 uint i = 0;
3207 /* inputs are always indirect with gs */
3208 bld.indirect_files |= (1 << TGSI_FILE_INPUT);
3209 bld.gs_iface = gs_iface;
3210 bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_gs_input;
3211 bld.bld_base.op_actions[TGSI_OPCODE_EMIT].emit = emit_vertex;
3212 bld.bld_base.op_actions[TGSI_OPCODE_ENDPRIM].emit = end_primitive;
3213
3214 for (i = 0; i < info->num_properties; ++i) {
3215 if (info->properties[i].name ==
3216 TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES) {
3217 max_output_vertices = info->properties[i].data[0];
3218 }
3219 }
3220 bld.max_output_vertices_vec =
3221 lp_build_const_int_vec(gallivm, bld.bld_base.uint_bld.type,
3222 max_output_vertices);
3223 }
3224
3225 lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.int_bld);
3226
3227 bld.system_values = *system_values;
3228
3229 lp_build_tgsi_llvm(&bld.bld_base, tokens);
3230
3231 if (0) {
3232 LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
3233 LLVMValueRef function = LLVMGetBasicBlockParent(block);
3234 debug_printf("11111111111111111111111111111 \n");
3235 tgsi_dump(tokens, 0);
3236 lp_debug_dump_value(function);
3237 debug_printf("2222222222222222222222222222 \n");
3238 }
3239
3240 if (0) {
3241 LLVMModuleRef module = LLVMGetGlobalParent(
3242 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
3243 LLVMDumpModule(module);
3244
3245 }
3246 }