1 /*
2 * Copyright © 2012 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Eric Anholt <[email protected]>
25 *
26 */
27
28 #include "brw_cfg.h"
29 #include "util/u_dynarray.h"
30 #include "brw_fs.h"
31
32 /** @file
33 *
34 * Walks the shader instructions generated and creates a set of basic
35 * blocks with successor/predecessor edges connecting them.
36 */
37
38 using namespace brw;
39
40 static bblock_t *
pop_stack(exec_list * list)41 pop_stack(exec_list *list)
42 {
43 bblock_link *link = (bblock_link *)list->get_tail();
44 bblock_t *block = link->block;
45 link->link.remove();
46
47 return block;
48 }
49
50 static exec_node *
link(void * mem_ctx,bblock_t * block,enum bblock_link_kind kind)51 link(void *mem_ctx, bblock_t *block, enum bblock_link_kind kind)
52 {
53 bblock_link *l = new(mem_ctx) bblock_link(block, kind);
54 return &l->link;
55 }
56
57 void
push_stack(exec_list * list,void * mem_ctx,bblock_t * block)58 push_stack(exec_list *list, void *mem_ctx, bblock_t *block)
59 {
60 /* The kind of the link is immaterial, but we need to provide one since
61 * this is (ab)using the edge data structure in order to implement a stack.
62 */
63 list->push_tail(link(mem_ctx, block, bblock_link_logical));
64 }
65
bblock_t(cfg_t * cfg)66 bblock_t::bblock_t(cfg_t *cfg) :
67 cfg(cfg), start_ip(0), end_ip(0), end_ip_delta(0), num(0)
68 {
69 instructions.make_empty();
70 parents.make_empty();
71 children.make_empty();
72 }
73
74 void
add_successor(void * mem_ctx,bblock_t * successor,enum bblock_link_kind kind)75 bblock_t::add_successor(void *mem_ctx, bblock_t *successor,
76 enum bblock_link_kind kind)
77 {
78 successor->parents.push_tail(::link(mem_ctx, this, kind));
79 children.push_tail(::link(mem_ctx, successor, kind));
80 }
81
82 bool
is_predecessor_of(const bblock_t * block,enum bblock_link_kind kind) const83 bblock_t::is_predecessor_of(const bblock_t *block,
84 enum bblock_link_kind kind) const
85 {
86 foreach_list_typed_safe (bblock_link, parent, link, &block->parents) {
87 if (parent->block == this && parent->kind <= kind) {
88 return true;
89 }
90 }
91
92 return false;
93 }
94
95 bool
is_successor_of(const bblock_t * block,enum bblock_link_kind kind) const96 bblock_t::is_successor_of(const bblock_t *block,
97 enum bblock_link_kind kind) const
98 {
99 foreach_list_typed_safe (bblock_link, child, link, &block->children) {
100 if (child->block == this && child->kind <= kind) {
101 return true;
102 }
103 }
104
105 return false;
106 }
107
108 static bool
ends_block(const fs_inst * inst)109 ends_block(const fs_inst *inst)
110 {
111 enum opcode op = inst->opcode;
112
113 return op == BRW_OPCODE_IF ||
114 op == BRW_OPCODE_ELSE ||
115 op == BRW_OPCODE_CONTINUE ||
116 op == BRW_OPCODE_BREAK ||
117 op == BRW_OPCODE_DO ||
118 op == BRW_OPCODE_WHILE;
119 }
120
121 static bool
starts_block(const fs_inst * inst)122 starts_block(const fs_inst *inst)
123 {
124 enum opcode op = inst->opcode;
125
126 return op == BRW_OPCODE_DO ||
127 op == BRW_OPCODE_ENDIF;
128 }
129
130 bool
can_combine_with(const bblock_t * that) const131 bblock_t::can_combine_with(const bblock_t *that) const
132 {
133 if ((const bblock_t *)this->link.next != that)
134 return false;
135
136 if (ends_block(this->end()) ||
137 starts_block(that->start()))
138 return false;
139
140 return true;
141 }
142
143 void
combine_with(bblock_t * that)144 bblock_t::combine_with(bblock_t *that)
145 {
146 assert(this->can_combine_with(that));
147 foreach_list_typed (bblock_link, link, link, &that->parents) {
148 assert(link->block == this);
149 }
150
151 this->end_ip = that->end_ip;
152 this->instructions.append_list(&that->instructions);
153
154 this->cfg->remove_block(that);
155 }
156
157 void
dump(FILE * file) const158 bblock_t::dump(FILE *file) const
159 {
160 const fs_visitor *s = this->cfg->s;
161
162 int ip = this->start_ip;
163 foreach_inst_in_block(fs_inst, inst, this) {
164 fprintf(file, "%5d: ", ip);
165 brw_print_instruction(*s, inst, file);
166 ip++;
167 }
168 }
169
170 void
unlink_list(exec_list * list)171 bblock_t::unlink_list(exec_list *list)
172 {
173 assert(list == &parents || list == &children);
174 const bool remove_parent = list == &children;
175
176 foreach_list_typed_safe(bblock_link, link, link, list) {
177 /* Also break the links from the other block back to this block. */
178 exec_list *sub_list = remove_parent ? &link->block->parents : &link->block->children;
179
180 foreach_list_typed_safe(bblock_link, sub_link, link, sub_list) {
181 if (sub_link->block == this) {
182 sub_link->link.remove();
183 ralloc_free(sub_link);
184 }
185 }
186
187 link->link.remove();
188 ralloc_free(link);
189 }
190 }
191
cfg_t(const fs_visitor * s,exec_list * instructions)192 cfg_t::cfg_t(const fs_visitor *s, exec_list *instructions) :
193 s(s)
194 {
195 mem_ctx = ralloc_context(NULL);
196 block_list.make_empty();
197 blocks = NULL;
198 num_blocks = 0;
199
200 bblock_t *cur = NULL;
201 int ip = 0;
202
203 bblock_t *entry = new_block();
204 bblock_t *cur_if = NULL; /**< BB ending with IF. */
205 bblock_t *cur_else = NULL; /**< BB ending with ELSE. */
206 bblock_t *cur_do = NULL; /**< BB starting with DO. */
207 bblock_t *cur_while = NULL; /**< BB immediately following WHILE. */
208 exec_list if_stack, else_stack, do_stack, while_stack;
209 bblock_t *next;
210
211 set_next_block(&cur, entry, ip);
212
213 foreach_in_list_safe(fs_inst, inst, instructions) {
214 /* set_next_block wants the post-incremented ip */
215 ip++;
216
217 inst->exec_node::remove();
218
219 switch (inst->opcode) {
220 case BRW_OPCODE_IF:
221 cur->instructions.push_tail(inst);
222
223 /* Push our information onto a stack so we can recover from
224 * nested ifs.
225 */
226 push_stack(&if_stack, mem_ctx, cur_if);
227 push_stack(&else_stack, mem_ctx, cur_else);
228
229 cur_if = cur;
230 cur_else = NULL;
231
232 /* Set up our immediately following block, full of "then"
233 * instructions.
234 */
235 next = new_block();
236 cur_if->add_successor(mem_ctx, next, bblock_link_logical);
237
238 set_next_block(&cur, next, ip);
239 break;
240
241 case BRW_OPCODE_ELSE:
242 cur->instructions.push_tail(inst);
243
244 cur_else = cur;
245
246 next = new_block();
247 assert(cur_if != NULL);
248 cur_if->add_successor(mem_ctx, next, bblock_link_logical);
249 cur_else->add_successor(mem_ctx, next, bblock_link_physical);
250
251 set_next_block(&cur, next, ip);
252 break;
253
254 case BRW_OPCODE_ENDIF: {
255 bblock_t *cur_endif;
256
257 if (cur->instructions.is_empty()) {
258 /* New block was just created; use it. */
259 cur_endif = cur;
260 } else {
261 cur_endif = new_block();
262
263 cur->add_successor(mem_ctx, cur_endif, bblock_link_logical);
264
265 set_next_block(&cur, cur_endif, ip - 1);
266 }
267
268 cur->instructions.push_tail(inst);
269
270 if (cur_else) {
271 cur_else->add_successor(mem_ctx, cur_endif, bblock_link_logical);
272 } else {
273 assert(cur_if != NULL);
274 cur_if->add_successor(mem_ctx, cur_endif, bblock_link_logical);
275 }
276
277 assert(cur_if->end()->opcode == BRW_OPCODE_IF);
278 assert(!cur_else || cur_else->end()->opcode == BRW_OPCODE_ELSE);
279
280 /* Pop the stack so we're in the previous if/else/endif */
281 cur_if = pop_stack(&if_stack);
282 cur_else = pop_stack(&else_stack);
283 break;
284 }
285 case BRW_OPCODE_DO:
286 /* Push our information onto a stack so we can recover from
287 * nested loops.
288 */
289 push_stack(&do_stack, mem_ctx, cur_do);
290 push_stack(&while_stack, mem_ctx, cur_while);
291
292 /* Set up the block just after the while. Don't know when exactly
293 * it will start, yet.
294 */
295 cur_while = new_block();
296
297 if (cur->instructions.is_empty()) {
298 /* New block was just created; use it. */
299 cur_do = cur;
300 } else {
301 cur_do = new_block();
302
303 cur->add_successor(mem_ctx, cur_do, bblock_link_logical);
304
305 set_next_block(&cur, cur_do, ip - 1);
306 }
307
308 cur->instructions.push_tail(inst);
309
310 /* Represent divergent execution of the loop as a pair of alternative
311 * edges coming out of the DO instruction: For any physical iteration
312 * of the loop a given logical thread can either start off enabled
313 * (which is represented as the "next" successor), or disabled (if it
314 * has reached a non-uniform exit of the loop during a previous
315 * iteration, which is represented as the "cur_while" successor).
316 *
317 * The disabled edge will be taken by the logical thread anytime we
318 * arrive at the DO instruction through a back-edge coming from a
319 * conditional exit of the loop where divergent control flow started.
320 *
321 * This guarantees that there is a control-flow path from any
322 * divergence point of the loop into the convergence point
323 * (immediately past the WHILE instruction) such that it overlaps the
324 * whole IP region of divergent control flow (potentially the whole
325 * loop) *and* doesn't imply the execution of any instructions part
326 * of the loop (since the corresponding execution mask bit will be
327 * disabled for a diverging thread).
328 *
329 * This way we make sure that any variables that are live throughout
330 * the region of divergence for an inactive logical thread are also
331 * considered to interfere with any other variables assigned by
332 * active logical threads within the same physical region of the
333 * program, since otherwise we would risk cross-channel data
334 * corruption.
335 */
336 next = new_block();
337 cur->add_successor(mem_ctx, next, bblock_link_logical);
338 cur->add_successor(mem_ctx, cur_while, bblock_link_physical);
339 set_next_block(&cur, next, ip);
340 break;
341
342 case BRW_OPCODE_CONTINUE:
343 cur->instructions.push_tail(inst);
344
345 /* A conditional CONTINUE may start a region of divergent control
346 * flow until the start of the next loop iteration (*not* until the
347 * end of the loop which is why the successor is not the top-level
348 * divergence point at cur_do). The live interval of any variable
349 * extending through a CONTINUE edge is guaranteed to overlap the
350 * whole region of divergent execution, because any variable live-out
351 * at the CONTINUE instruction will also be live-in at the top of the
352 * loop, and therefore also live-out at the bottom-most point of the
353 * loop which is reachable from the top (since a control flow path
354 * exists from a definition of the variable through this CONTINUE
355 * instruction, the top of the loop, the (reachable) bottom of the
356 * loop, the top of the loop again, into a use of the variable).
357 */
358 assert(cur_do != NULL);
359 cur->add_successor(mem_ctx, cur_do->next(), bblock_link_logical);
360
361 next = new_block();
362 if (inst->predicate)
363 cur->add_successor(mem_ctx, next, bblock_link_logical);
364 else
365 cur->add_successor(mem_ctx, next, bblock_link_physical);
366
367 set_next_block(&cur, next, ip);
368 break;
369
370 case BRW_OPCODE_BREAK:
371 cur->instructions.push_tail(inst);
372
373 /* A conditional BREAK instruction may start a region of divergent
374 * control flow until the end of the loop if the condition is
375 * non-uniform, in which case the loop will execute additional
376 * iterations with the present channel disabled. We model this as a
377 * control flow path from the divergence point to the convergence
378 * point that overlaps the whole IP range of the loop and skips over
379 * the execution of any other instructions part of the loop.
380 *
381 * See the DO case for additional explanation.
382 */
383 assert(cur_do != NULL);
384 cur->add_successor(mem_ctx, cur_do, bblock_link_physical);
385 cur->add_successor(mem_ctx, cur_while, bblock_link_logical);
386
387 next = new_block();
388 if (inst->predicate)
389 cur->add_successor(mem_ctx, next, bblock_link_logical);
390 else
391 cur->add_successor(mem_ctx, next, bblock_link_physical);
392
393 set_next_block(&cur, next, ip);
394 break;
395
396 case BRW_OPCODE_WHILE:
397 cur->instructions.push_tail(inst);
398
399 assert(cur_do != NULL && cur_while != NULL);
400
401 /* A conditional WHILE instruction may start a region of divergent
402 * control flow until the end of the loop, just like the BREAK
403 * instruction. See the BREAK case for more details. OTOH an
404 * unconditional WHILE instruction is non-divergent (just like an
405 * unconditional CONTINUE), and will necessarily lead to the
406 * execution of an additional iteration of the loop for all enabled
407 * channels, so we may skip over the divergence point at the top of
408 * the loop to keep the CFG as unambiguous as possible.
409 */
410 if (inst->predicate) {
411 cur->add_successor(mem_ctx, cur_do, bblock_link_logical);
412 } else {
413 cur->add_successor(mem_ctx, cur_do->next(), bblock_link_logical);
414 }
415
416 set_next_block(&cur, cur_while, ip);
417
418 /* Pop the stack so we're in the previous loop */
419 cur_do = pop_stack(&do_stack);
420 cur_while = pop_stack(&while_stack);
421 break;
422
423 default:
424 cur->instructions.push_tail(inst);
425 break;
426 }
427 }
428
429 cur->end_ip = ip - 1;
430
431 make_block_array();
432 }
433
~cfg_t()434 cfg_t::~cfg_t()
435 {
436 ralloc_free(mem_ctx);
437 }
438
439 void
remove_block(bblock_t * block)440 cfg_t::remove_block(bblock_t *block)
441 {
442 foreach_list_typed_safe (bblock_link, predecessor, link, &block->parents) {
443 /* cfg_t::validate checks that predecessor and successor lists are well
444 * formed, so it is known that the loop here would find exactly one
445 * block. Set old_link_kind to silence "variable used but not set"
446 * warnings.
447 */
448 bblock_link_kind old_link_kind = bblock_link_logical;
449
450 /* Remove block from all of its predecessors' successor lists. */
451 foreach_list_typed_safe (bblock_link, successor, link,
452 &predecessor->block->children) {
453 if (block == successor->block) {
454 old_link_kind = successor->kind;
455 successor->link.remove();
456 ralloc_free(successor);
457 break;
458 }
459 }
460
461 /* Add removed-block's successors to its predecessors' successor lists. */
462 foreach_list_typed (bblock_link, successor, link, &block->children) {
463 bool need_to_link = true;
464 bblock_link_kind new_link_kind = MAX2(old_link_kind, successor->kind);
465
466 foreach_list_typed_safe (bblock_link, child, link, &predecessor->block->children) {
467 /* There is already a link between the two blocks. If the links
468 * are the same kind or the link is logical, do nothing. If the
469 * existing link is physical and the proposed new link is logical,
470 * promote the existing link to logical.
471 *
472 * This is accomplished by taking the minimum of the existing link
473 * kind and the proposed link kind.
474 */
475 if (child->block == successor->block) {
476 child->kind = MIN2(child->kind, new_link_kind);
477 need_to_link = false;
478 break;
479 }
480 }
481
482 if (need_to_link) {
483 predecessor->block->children.push_tail(link(mem_ctx,
484 successor->block,
485 new_link_kind));
486 }
487 }
488 }
489
490 foreach_list_typed_safe (bblock_link, successor, link, &block->children) {
491 /* cfg_t::validate checks that predecessor and successor lists are well
492 * formed, so it is known that the loop here would find exactly one
493 * block. Set old_link_kind to silence "variable used but not set"
494 * warnings.
495 */
496 bblock_link_kind old_link_kind = bblock_link_logical;
497
498 /* Remove block from all of its childrens' parents lists. */
499 foreach_list_typed_safe (bblock_link, predecessor, link,
500 &successor->block->parents) {
501 if (block == predecessor->block) {
502 old_link_kind = predecessor->kind;
503 predecessor->link.remove();
504 ralloc_free(predecessor);
505 }
506 }
507
508 /* Add removed-block's predecessors to its successors' predecessor lists. */
509 foreach_list_typed (bblock_link, predecessor, link, &block->parents) {
510 bool need_to_link = true;
511 bblock_link_kind new_link_kind = MAX2(old_link_kind, predecessor->kind);
512
513 foreach_list_typed_safe (bblock_link, parent, link, &successor->block->parents) {
514 /* There is already a link between the two blocks. If the links
515 * are the same kind or the link is logical, do nothing. If the
516 * existing link is physical and the proposed new link is logical,
517 * promote the existing link to logical.
518 *
519 * This is accomplished by taking the minimum of the existing link
520 * kind and the proposed link kind.
521 */
522 if (parent->block == predecessor->block) {
523 parent->kind = MIN2(parent->kind, new_link_kind);
524 need_to_link = false;
525 break;
526 }
527 }
528
529 if (need_to_link) {
530 successor->block->parents.push_tail(link(mem_ctx,
531 predecessor->block,
532 new_link_kind));
533 }
534 }
535 }
536
537 block->link.remove();
538
539 for (int b = block->num; b < this->num_blocks - 1; b++) {
540 this->blocks[b] = this->blocks[b + 1];
541 this->blocks[b]->num = b;
542 }
543
544 this->blocks[this->num_blocks - 1]->num = this->num_blocks - 2;
545 this->num_blocks--;
546 }
547
548 bblock_t *
new_block()549 cfg_t::new_block()
550 {
551 bblock_t *block = new(mem_ctx) bblock_t(this);
552
553 return block;
554 }
555
556 void
set_next_block(bblock_t ** cur,bblock_t * block,int ip)557 cfg_t::set_next_block(bblock_t **cur, bblock_t *block, int ip)
558 {
559 if (*cur) {
560 (*cur)->end_ip = ip - 1;
561 }
562
563 block->start_ip = ip;
564 block->num = num_blocks++;
565 block_list.push_tail(&block->link);
566 *cur = block;
567 }
568
569 void
make_block_array()570 cfg_t::make_block_array()
571 {
572 blocks = ralloc_array(mem_ctx, bblock_t *, num_blocks);
573
574 int i = 0;
575 foreach_block (block, this) {
576 blocks[i++] = block;
577 }
578 assert(i == num_blocks);
579 }
580
581 namespace {
582
583 struct link_desc {
584 char kind;
585 int num;
586 };
587
588 int
compare_link_desc(const void * a,const void * b)589 compare_link_desc(const void *a, const void *b)
590 {
591 const link_desc *la = (const link_desc *)a;
592 const link_desc *lb = (const link_desc *)b;
593
594 return la->num < lb->num ? -1 :
595 la->num > lb->num ? +1 :
596 la->kind < lb->kind ? -1 :
597 la->kind > lb->kind ? +1 :
598 0;
599 }
600
601 void
sort_links(util_dynarray * scratch,exec_list * list)602 sort_links(util_dynarray *scratch, exec_list *list)
603 {
604 util_dynarray_clear(scratch);
605 foreach_list_typed(bblock_link, link, link, list) {
606 link_desc l;
607 l.kind = link->kind == bblock_link_logical ? '-' : '~';
608 l.num = link->block->num;
609 util_dynarray_append(scratch, link_desc, l);
610 }
611 qsort(scratch->data, util_dynarray_num_elements(scratch, link_desc),
612 sizeof(link_desc), compare_link_desc);
613 }
614
615 } /* namespace */
616
617 void
dump(FILE * file)618 cfg_t::dump(FILE *file)
619 {
620 const idom_tree *idom = (s ? &s->idom_analysis.require() : NULL);
621
622 /* Temporary storage to sort the lists of blocks. This normalizes the
623 * output, making it possible to use it for certain tests.
624 */
625 util_dynarray scratch;
626 util_dynarray_init(&scratch, NULL);
627
628 foreach_block (block, this) {
629 if (idom && idom->parent(block))
630 fprintf(file, "START B%d IDOM(B%d)", block->num,
631 idom->parent(block)->num);
632 else
633 fprintf(file, "START B%d IDOM(none)", block->num);
634
635 sort_links(&scratch, &block->parents);
636 util_dynarray_foreach(&scratch, link_desc, l)
637 fprintf(file, " <%cB%d", l->kind, l->num);
638 fprintf(file, "\n");
639
640 if (s != NULL)
641 block->dump(file);
642 fprintf(file, "END B%d", block->num);
643
644 sort_links(&scratch, &block->children);
645 util_dynarray_foreach(&scratch, link_desc, l)
646 fprintf(file, " %c>B%d", l->kind, l->num);
647 fprintf(file, "\n");
648 }
649
650 util_dynarray_fini(&scratch);
651 }
652
653 /* Calculates the immediate dominator of each block, according to "A Simple,
654 * Fast Dominance Algorithm" by Keith D. Cooper, Timothy J. Harvey, and Ken
655 * Kennedy.
656 *
657 * The authors claim that for control flow graphs of sizes normally encountered
658 * (less than 1000 nodes) that this algorithm is significantly faster than
659 * others like Lengauer-Tarjan.
660 */
idom_tree(const fs_visitor * s)661 idom_tree::idom_tree(const fs_visitor *s) :
662 num_parents(s->cfg->num_blocks),
663 parents(new bblock_t *[num_parents]())
664 {
665 bool changed;
666
667 parents[0] = s->cfg->blocks[0];
668
669 do {
670 changed = false;
671
672 foreach_block(block, s->cfg) {
673 if (block->num == 0)
674 continue;
675
676 bblock_t *new_idom = NULL;
677 foreach_list_typed(bblock_link, parent_link, link, &block->parents) {
678 if (parent(parent_link->block)) {
679 new_idom = (new_idom ? intersect(new_idom, parent_link->block) :
680 parent_link->block);
681 }
682 }
683
684 if (parent(block) != new_idom) {
685 parents[block->num] = new_idom;
686 changed = true;
687 }
688 }
689 } while (changed);
690 }
691
~idom_tree()692 idom_tree::~idom_tree()
693 {
694 delete[] parents;
695 }
696
697 bblock_t *
intersect(bblock_t * b1,bblock_t * b2) const698 idom_tree::intersect(bblock_t *b1, bblock_t *b2) const
699 {
700 /* Note, the comparisons here are the opposite of what the paper says
701 * because we index blocks from beginning -> end (i.e. reverse post-order)
702 * instead of post-order like they assume.
703 */
704 while (b1->num != b2->num) {
705 while (b1->num > b2->num)
706 b1 = parent(b1);
707 while (b2->num > b1->num)
708 b2 = parent(b2);
709 }
710 assert(b1);
711 return b1;
712 }
713
714 void
dump(FILE * file) const715 idom_tree::dump(FILE *file) const
716 {
717 fprintf(file, "digraph DominanceTree {\n");
718 for (unsigned i = 0; i < num_parents; i++)
719 fprintf(file, "\t%d -> %d\n", parents[i]->num, i);
720 fprintf(file, "}\n");
721 }
722
723 void
dump_cfg()724 cfg_t::dump_cfg()
725 {
726 printf("digraph CFG {\n");
727 for (int b = 0; b < num_blocks; b++) {
728 bblock_t *block = this->blocks[b];
729
730 foreach_list_typed_safe (bblock_link, child, link, &block->children) {
731 printf("\t%d -> %d\n", b, child->block->num);
732 }
733 }
734 printf("}\n");
735 }
736
737 void
brw_calculate_cfg(fs_visitor & s)738 brw_calculate_cfg(fs_visitor &s)
739 {
740 if (s.cfg)
741 return;
742 s.cfg = new(s.mem_ctx) cfg_t(&s, &s.instructions);
743 }
744
745 #define cfgv_assert(assertion) \
746 { \
747 if (!(assertion)) { \
748 fprintf(stderr, "ASSERT: CFG validation in %s failed!\n", stage_abbrev); \
749 fprintf(stderr, "%s:%d: '%s' failed\n", __FILE__, __LINE__, #assertion); \
750 abort(); \
751 } \
752 }
753
754 #ifndef NDEBUG
755 void
validate(const char * stage_abbrev)756 cfg_t::validate(const char *stage_abbrev)
757 {
758 foreach_block(block, this) {
759 foreach_list_typed(bblock_link, successor, link, &block->children) {
760 /* Each successor of a block must have one predecessor link back to
761 * the block.
762 */
763 bool successor_links_back_to_predecessor = false;
764 bblock_t *succ_block = successor->block;
765
766 foreach_list_typed(bblock_link, predecessor, link, &succ_block->parents) {
767 if (predecessor->block == block) {
768 cfgv_assert(!successor_links_back_to_predecessor);
769 cfgv_assert(successor->kind == predecessor->kind);
770 successor_links_back_to_predecessor = true;
771 }
772 }
773
774 cfgv_assert(successor_links_back_to_predecessor);
775
776 /* Each successor block must appear only once in the list of
777 * successors.
778 */
779 foreach_list_typed_from(bblock_link, later_successor, link,
780 &block->children, successor->link.next) {
781 cfgv_assert(successor->block != later_successor->block);
782 }
783 }
784
785 foreach_list_typed(bblock_link, predecessor, link, &block->parents) {
786 /* Each predecessor of a block must have one successor link back to
787 * the block.
788 */
789 bool predecessor_links_back_to_successor = false;
790 bblock_t *pred_block = predecessor->block;
791
792 foreach_list_typed(bblock_link, successor, link, &pred_block->children) {
793 if (successor->block == block) {
794 cfgv_assert(!predecessor_links_back_to_successor);
795 cfgv_assert(successor->kind == predecessor->kind);
796 predecessor_links_back_to_successor = true;
797 }
798 }
799
800 cfgv_assert(predecessor_links_back_to_successor);
801
802 /* Each precessor block must appear only once in the list of
803 * precessors.
804 */
805 foreach_list_typed_from(bblock_link, later_precessor, link,
806 &block->parents, predecessor->link.next) {
807 cfgv_assert(predecessor->block != later_precessor->block);
808 }
809 }
810
811 fs_inst *first_inst = block->start();
812 if (first_inst->opcode == BRW_OPCODE_DO) {
813 /* DO instructions both begin and end a block, so the DO instruction
814 * must be the only instruction in the block.
815 */
816 cfgv_assert(exec_list_is_singular(&block->instructions));
817
818 /* A block starting with DO should have exactly two successors. One
819 * is a physical link to the block starting after the WHILE
820 * instruction. The other is a logical link to the block starting the
821 * body of the loop.
822 */
823 bblock_t *physical_block = nullptr;
824 bblock_t *logical_block = nullptr;
825
826 foreach_list_typed(bblock_link, child, link, &block->children) {
827 if (child->kind == bblock_link_physical) {
828 cfgv_assert(physical_block == nullptr);
829 physical_block = child->block;
830 } else {
831 cfgv_assert(logical_block == nullptr);
832 logical_block = child->block;
833 }
834 }
835
836 cfgv_assert(logical_block != nullptr);
837 cfgv_assert(physical_block != nullptr);
838 }
839 }
840 }
841 #endif
842