nir/cf: Remove phi sources if needed in nir_handle_add_jump
[mesa.git] / src / compiler / nir / nir_control_flow.c
1 /*
2 * Copyright © 2014 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 * Connor Abbott (cwabbott0@gmail.com)
25 *
26 */
27
28 #include "nir_control_flow_private.h"
29
30 /**
31 * \name Control flow modification
32 *
33 * These functions modify the control flow tree while keeping the control flow
34 * graph up-to-date. The invariants respected are:
35 * 1. Each then statement, else statement, or loop body must have at least one
36 * control flow node.
37 * 2. Each if-statement and loop must have one basic block before it and one
38 * after.
39 * 3. Two basic blocks cannot be directly next to each other.
40 * 4. If a basic block has a jump instruction, there must be only one and it
41 * must be at the end of the block.
42 *
43 * The purpose of the second one is so that we have places to insert code during
44 * GCM, as well as eliminating the possibility of critical edges.
45 */
46 /*@{*/
47
48 static inline void
49 block_add_pred(nir_block *block, nir_block *pred)
50 {
51 _mesa_set_add(block->predecessors, pred);
52 }
53
54 static inline void
55 block_remove_pred(nir_block *block, nir_block *pred)
56 {
57 struct set_entry *entry = _mesa_set_search(block->predecessors, pred);
58
59 assert(entry);
60
61 _mesa_set_remove(block->predecessors, entry);
62 }
63
64 static void
65 link_blocks(nir_block *pred, nir_block *succ1, nir_block *succ2)
66 {
67 pred->successors[0] = succ1;
68 if (succ1 != NULL)
69 block_add_pred(succ1, pred);
70
71 pred->successors[1] = succ2;
72 if (succ2 != NULL)
73 block_add_pred(succ2, pred);
74 }
75
76 static void
77 unlink_blocks(nir_block *pred, nir_block *succ)
78 {
79 if (pred->successors[0] == succ) {
80 pred->successors[0] = pred->successors[1];
81 pred->successors[1] = NULL;
82 } else {
83 assert(pred->successors[1] == succ);
84 pred->successors[1] = NULL;
85 }
86
87 block_remove_pred(succ, pred);
88 }
89
90 static void
91 unlink_block_successors(nir_block *block)
92 {
93 if (block->successors[1] != NULL)
94 unlink_blocks(block, block->successors[1]);
95 if (block->successors[0] != NULL)
96 unlink_blocks(block, block->successors[0]);
97 }
98
99 static void
100 link_non_block_to_block(nir_cf_node *node, nir_block *block)
101 {
102 if (node->type == nir_cf_node_if) {
103 /*
104 * We're trying to link an if to a block after it; this just means linking
105 * the last block of the then and else branches.
106 */
107
108 nir_if *if_stmt = nir_cf_node_as_if(node);
109
110 nir_block *last_then_block = nir_if_last_then_block(if_stmt);
111 nir_block *last_else_block = nir_if_last_else_block(if_stmt);
112
113 if (!nir_block_ends_in_jump(last_then_block)) {
114 unlink_block_successors(last_then_block);
115 link_blocks(last_then_block, block, NULL);
116 }
117
118 if (!nir_block_ends_in_jump(last_else_block)) {
119 unlink_block_successors(last_else_block);
120 link_blocks(last_else_block, block, NULL);
121 }
122 } else {
123 assert(node->type == nir_cf_node_loop);
124 }
125 }
126
127 static void
128 link_block_to_non_block(nir_block *block, nir_cf_node *node)
129 {
130 if (node->type == nir_cf_node_if) {
131 /*
132 * We're trying to link a block to an if after it; this just means linking
133 * the block to the first block of the then and else branches.
134 */
135
136 nir_if *if_stmt = nir_cf_node_as_if(node);
137
138 nir_block *first_then_block = nir_if_first_then_block(if_stmt);
139 nir_block *first_else_block = nir_if_first_else_block(if_stmt);
140
141 unlink_block_successors(block);
142 link_blocks(block, first_then_block, first_else_block);
143 } else {
144 /*
145 * For similar reasons as the corresponding case in
146 * link_non_block_to_block(), don't worry about if the loop header has
147 * any predecessors that need to be unlinked.
148 */
149
150 nir_loop *loop = nir_cf_node_as_loop(node);
151
152 nir_block *loop_header_block = nir_loop_first_block(loop);
153
154 unlink_block_successors(block);
155 link_blocks(block, loop_header_block, NULL);
156 }
157
158 }
159
160 /**
161 * Replace a block's successor with a different one.
162 */
163 static void
164 replace_successor(nir_block *block, nir_block *old_succ, nir_block *new_succ)
165 {
166 if (block->successors[0] == old_succ) {
167 block->successors[0] = new_succ;
168 } else {
169 assert(block->successors[1] == old_succ);
170 block->successors[1] = new_succ;
171 }
172
173 block_remove_pred(old_succ, block);
174 block_add_pred(new_succ, block);
175 }
176
177 /**
178 * Takes a basic block and inserts a new empty basic block before it, making its
179 * predecessors point to the new block. This essentially splits the block into
180 * an empty header and a body so that another non-block CF node can be inserted
181 * between the two. Note that this does *not* link the two basic blocks, so
182 * some kind of cleanup *must* be performed after this call.
183 */
184
185 static nir_block *
186 split_block_beginning(nir_block *block)
187 {
188 nir_block *new_block = nir_block_create(ralloc_parent(block));
189 new_block->cf_node.parent = block->cf_node.parent;
190 exec_node_insert_node_before(&block->cf_node.node, &new_block->cf_node.node);
191
192 struct set_entry *entry;
193 set_foreach(block->predecessors, entry) {
194 nir_block *pred = (nir_block *) entry->key;
195 replace_successor(pred, block, new_block);
196 }
197
198 /* Any phi nodes must stay part of the new block, or else their
199 * sourcse will be messed up. This will reverse the order of the phis, but
200 * order shouldn't matter.
201 */
202 nir_foreach_instr_safe(instr, block) {
203 if (instr->type != nir_instr_type_phi)
204 break;
205
206 exec_node_remove(&instr->node);
207 instr->block = new_block;
208 exec_list_push_head(&new_block->instr_list, &instr->node);
209 }
210
211 return new_block;
212 }
213
214 static void
215 rewrite_phi_preds(nir_block *block, nir_block *old_pred, nir_block *new_pred)
216 {
217 nir_foreach_instr_safe(instr, block) {
218 if (instr->type != nir_instr_type_phi)
219 break;
220
221 nir_phi_instr *phi = nir_instr_as_phi(instr);
222 nir_foreach_phi_src(src, phi) {
223 if (src->pred == old_pred) {
224 src->pred = new_pred;
225 break;
226 }
227 }
228 }
229 }
230
231 static void
232 insert_phi_undef(nir_block *block, nir_block *pred)
233 {
234 nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node);
235 nir_foreach_instr(instr, block) {
236 if (instr->type != nir_instr_type_phi)
237 break;
238
239 nir_phi_instr *phi = nir_instr_as_phi(instr);
240 nir_ssa_undef_instr *undef =
241 nir_ssa_undef_instr_create(ralloc_parent(phi),
242 phi->dest.ssa.num_components,
243 phi->dest.ssa.bit_size);
244 nir_instr_insert_before_cf_list(&impl->body, &undef->instr);
245 nir_phi_src *src = ralloc(phi, nir_phi_src);
246 src->pred = pred;
247 src->src.parent_instr = &phi->instr;
248 src->src.is_ssa = true;
249 src->src.ssa = &undef->def;
250
251 list_addtail(&src->src.use_link, &undef->def.uses);
252
253 exec_list_push_tail(&phi->srcs, &src->node);
254 }
255 }
256
257 /**
258 * Moves the successors of source to the successors of dest, leaving both
259 * successors of source NULL.
260 */
261
262 static void
263 move_successors(nir_block *source, nir_block *dest)
264 {
265 nir_block *succ1 = source->successors[0];
266 nir_block *succ2 = source->successors[1];
267
268 if (succ1) {
269 unlink_blocks(source, succ1);
270 rewrite_phi_preds(succ1, source, dest);
271 }
272
273 if (succ2) {
274 unlink_blocks(source, succ2);
275 rewrite_phi_preds(succ2, source, dest);
276 }
277
278 unlink_block_successors(dest);
279 link_blocks(dest, succ1, succ2);
280 }
281
282 /* Given a basic block with no successors that has been inserted into the
283 * control flow tree, gives it the successors it would normally have assuming
284 * it doesn't end in a jump instruction. Also inserts phi sources with undefs
285 * if necessary.
286 */
287 static void
288 block_add_normal_succs(nir_block *block)
289 {
290 if (exec_node_is_tail_sentinel(block->cf_node.node.next)) {
291 nir_cf_node *parent = block->cf_node.parent;
292 if (parent->type == nir_cf_node_if) {
293 nir_cf_node *next = nir_cf_node_next(parent);
294 nir_block *next_block = nir_cf_node_as_block(next);
295
296 link_blocks(block, next_block, NULL);
297 } else if (parent->type == nir_cf_node_loop) {
298 nir_loop *loop = nir_cf_node_as_loop(parent);
299
300 nir_block *head_block = nir_loop_first_block(loop);
301
302 link_blocks(block, head_block, NULL);
303 insert_phi_undef(head_block, block);
304 } else {
305 nir_function_impl *impl = nir_cf_node_as_function(parent);
306 link_blocks(block, impl->end_block, NULL);
307 }
308 } else {
309 nir_cf_node *next = nir_cf_node_next(&block->cf_node);
310 if (next->type == nir_cf_node_if) {
311 nir_if *next_if = nir_cf_node_as_if(next);
312
313 nir_block *first_then_block = nir_if_first_then_block(next_if);
314 nir_block *first_else_block = nir_if_first_else_block(next_if);
315
316 link_blocks(block, first_then_block, first_else_block);
317 } else {
318 nir_loop *next_loop = nir_cf_node_as_loop(next);
319
320 nir_block *first_block = nir_loop_first_block(next_loop);
321
322 link_blocks(block, first_block, NULL);
323 insert_phi_undef(first_block, block);
324 }
325 }
326 }
327
328 static nir_block *
329 split_block_end(nir_block *block)
330 {
331 nir_block *new_block = nir_block_create(ralloc_parent(block));
332 new_block->cf_node.parent = block->cf_node.parent;
333 exec_node_insert_after(&block->cf_node.node, &new_block->cf_node.node);
334
335 if (nir_block_ends_in_jump(block)) {
336 /* Figure out what successor block would've had if it didn't have a jump
337 * instruction, and make new_block have that successor.
338 */
339 block_add_normal_succs(new_block);
340 } else {
341 move_successors(block, new_block);
342 }
343
344 return new_block;
345 }
346
347 static nir_block *
348 split_block_before_instr(nir_instr *instr)
349 {
350 assert(instr->type != nir_instr_type_phi);
351 nir_block *new_block = split_block_beginning(instr->block);
352
353 nir_foreach_instr_safe(cur_instr, instr->block) {
354 if (cur_instr == instr)
355 break;
356
357 exec_node_remove(&cur_instr->node);
358 cur_instr->block = new_block;
359 exec_list_push_tail(&new_block->instr_list, &cur_instr->node);
360 }
361
362 return new_block;
363 }
364
365 /* Splits a basic block at the point specified by the cursor. The "before" and
366 * "after" arguments are filled out with the blocks resulting from the split
367 * if non-NULL. Note that the "beginning" of the block is actually interpreted
368 * as before the first non-phi instruction, and it's illegal to split a block
369 * before a phi instruction.
370 */
371
372 static void
373 split_block_cursor(nir_cursor cursor,
374 nir_block **_before, nir_block **_after)
375 {
376 nir_block *before, *after;
377 switch (cursor.option) {
378 case nir_cursor_before_block:
379 after = cursor.block;
380 before = split_block_beginning(cursor.block);
381 break;
382
383 case nir_cursor_after_block:
384 before = cursor.block;
385 after = split_block_end(cursor.block);
386 break;
387
388 case nir_cursor_before_instr:
389 after = cursor.instr->block;
390 before = split_block_before_instr(cursor.instr);
391 break;
392
393 case nir_cursor_after_instr:
394 /* We lower this to split_block_before_instr() so that we can keep the
395 * after-a-jump-instr case contained to split_block_end().
396 */
397 if (nir_instr_is_last(cursor.instr)) {
398 before = cursor.instr->block;
399 after = split_block_end(cursor.instr->block);
400 } else {
401 after = cursor.instr->block;
402 before = split_block_before_instr(nir_instr_next(cursor.instr));
403 }
404 break;
405
406 default:
407 unreachable("not reached");
408 }
409
410 if (_before)
411 *_before = before;
412 if (_after)
413 *_after = after;
414 }
415
416 /**
417 * Inserts a non-basic block between two basic blocks and links them together.
418 */
419
420 static void
421 insert_non_block(nir_block *before, nir_cf_node *node, nir_block *after)
422 {
423 node->parent = before->cf_node.parent;
424 exec_node_insert_after(&before->cf_node.node, &node->node);
425 link_block_to_non_block(before, node);
426 link_non_block_to_block(node, after);
427 }
428
429 /* walk up the control flow tree to find the innermost enclosed loop */
430 static nir_loop *
431 nearest_loop(nir_cf_node *node)
432 {
433 while (node->type != nir_cf_node_loop) {
434 node = node->parent;
435 }
436
437 return nir_cf_node_as_loop(node);
438 }
439
440 static void
441 remove_phi_src(nir_block *block, nir_block *pred)
442 {
443 nir_foreach_instr(instr, block) {
444 if (instr->type != nir_instr_type_phi)
445 break;
446
447 nir_phi_instr *phi = nir_instr_as_phi(instr);
448 nir_foreach_phi_src_safe(src, phi) {
449 if (src->pred == pred) {
450 list_del(&src->src.use_link);
451 exec_node_remove(&src->node);
452 }
453 }
454 }
455 }
456
457 /*
458 * update the CFG after a jump instruction has been added to the end of a block
459 */
460
461 void
462 nir_handle_add_jump(nir_block *block)
463 {
464 nir_instr *instr = nir_block_last_instr(block);
465 nir_jump_instr *jump_instr = nir_instr_as_jump(instr);
466
467 if (block->successors[0])
468 remove_phi_src(block->successors[0], block);
469 if (block->successors[1])
470 remove_phi_src(block->successors[1], block);
471 unlink_block_successors(block);
472
473 nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node);
474 nir_metadata_preserve(impl, nir_metadata_none);
475
476 if (jump_instr->type == nir_jump_break ||
477 jump_instr->type == nir_jump_continue) {
478 nir_loop *loop = nearest_loop(&block->cf_node);
479
480 if (jump_instr->type == nir_jump_continue) {
481 nir_block *first_block = nir_loop_first_block(loop);
482 link_blocks(block, first_block, NULL);
483 } else {
484 nir_cf_node *after = nir_cf_node_next(&loop->cf_node);
485 nir_block *after_block = nir_cf_node_as_block(after);
486 link_blocks(block, after_block, NULL);
487 }
488 } else {
489 assert(jump_instr->type == nir_jump_return);
490 link_blocks(block, impl->end_block, NULL);
491 }
492 }
493
494 /* Removes the successor of a block with a jump. Note that the jump to be
495 * eliminated may be free-floating.
496 */
497
498 static void
499 unlink_jump(nir_block *block, nir_jump_type type, bool add_normal_successors)
500 {
501 if (block->successors[0])
502 remove_phi_src(block->successors[0], block);
503 if (block->successors[1])
504 remove_phi_src(block->successors[1], block);
505
506 unlink_block_successors(block);
507 if (add_normal_successors)
508 block_add_normal_succs(block);
509 }
510
511 void
512 nir_handle_remove_jump(nir_block *block, nir_jump_type type)
513 {
514 unlink_jump(block, type, true);
515
516 nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node);
517 nir_metadata_preserve(impl, nir_metadata_none);
518 }
519
520 static void
521 update_if_uses(nir_cf_node *node)
522 {
523 if (node->type != nir_cf_node_if)
524 return;
525
526 nir_if *if_stmt = nir_cf_node_as_if(node);
527
528 if_stmt->condition.parent_if = if_stmt;
529 if (if_stmt->condition.is_ssa) {
530 list_addtail(&if_stmt->condition.use_link,
531 &if_stmt->condition.ssa->if_uses);
532 } else {
533 list_addtail(&if_stmt->condition.use_link,
534 &if_stmt->condition.reg.reg->if_uses);
535 }
536 }
537
538 /**
539 * Stitch two basic blocks together into one. The aggregate must have the same
540 * predecessors as the first and the same successors as the second.
541 */
542
543 static void
544 stitch_blocks(nir_block *before, nir_block *after)
545 {
546 /*
547 * We move after into before, so we have to deal with up to 2 successors vs.
548 * possibly a large number of predecessors.
549 *
550 * TODO: special case when before is empty and after isn't?
551 */
552
553 if (nir_block_ends_in_jump(before)) {
554 assert(exec_list_is_empty(&after->instr_list));
555 if (after->successors[0])
556 remove_phi_src(after->successors[0], after);
557 if (after->successors[1])
558 remove_phi_src(after->successors[1], after);
559 unlink_block_successors(after);
560 exec_node_remove(&after->cf_node.node);
561 } else {
562 move_successors(after, before);
563
564 foreach_list_typed(nir_instr, instr, node, &after->instr_list) {
565 instr->block = before;
566 }
567
568 exec_list_append(&before->instr_list, &after->instr_list);
569 exec_node_remove(&after->cf_node.node);
570 }
571 }
572
573 void
574 nir_cf_node_insert(nir_cursor cursor, nir_cf_node *node)
575 {
576 nir_block *before, *after;
577
578 split_block_cursor(cursor, &before, &after);
579
580 if (node->type == nir_cf_node_block) {
581 nir_block *block = nir_cf_node_as_block(node);
582 exec_node_insert_after(&before->cf_node.node, &block->cf_node.node);
583 block->cf_node.parent = before->cf_node.parent;
584 /* stitch_blocks() assumes that any block that ends with a jump has
585 * already been setup with the correct successors, so we need to set
586 * up jumps here as the block is being inserted.
587 */
588 if (nir_block_ends_in_jump(block))
589 nir_handle_add_jump(block);
590
591 stitch_blocks(block, after);
592 stitch_blocks(before, block);
593 } else {
594 update_if_uses(node);
595 insert_non_block(before, node, after);
596 }
597 }
598
599 static bool
600 replace_ssa_def_uses(nir_ssa_def *def, void *void_impl)
601 {
602 nir_function_impl *impl = void_impl;
603 void *mem_ctx = ralloc_parent(impl);
604
605 nir_ssa_undef_instr *undef =
606 nir_ssa_undef_instr_create(mem_ctx, def->num_components,
607 def->bit_size);
608 nir_instr_insert_before_cf_list(&impl->body, &undef->instr);
609 nir_ssa_def_rewrite_uses(def, nir_src_for_ssa(&undef->def));
610 return true;
611 }
612
613 static void
614 cleanup_cf_node(nir_cf_node *node, nir_function_impl *impl)
615 {
616 switch (node->type) {
617 case nir_cf_node_block: {
618 nir_block *block = nir_cf_node_as_block(node);
619 /* We need to walk the instructions and clean up defs/uses */
620 nir_foreach_instr_safe(instr, block) {
621 if (instr->type == nir_instr_type_jump) {
622 nir_jump_type jump_type = nir_instr_as_jump(instr)->type;
623 unlink_jump(block, jump_type, false);
624 } else {
625 nir_foreach_ssa_def(instr, replace_ssa_def_uses, impl);
626 nir_instr_remove(instr);
627 }
628 }
629 break;
630 }
631
632 case nir_cf_node_if: {
633 nir_if *if_stmt = nir_cf_node_as_if(node);
634 foreach_list_typed(nir_cf_node, child, node, &if_stmt->then_list)
635 cleanup_cf_node(child, impl);
636 foreach_list_typed(nir_cf_node, child, node, &if_stmt->else_list)
637 cleanup_cf_node(child, impl);
638
639 list_del(&if_stmt->condition.use_link);
640 break;
641 }
642
643 case nir_cf_node_loop: {
644 nir_loop *loop = nir_cf_node_as_loop(node);
645 foreach_list_typed(nir_cf_node, child, node, &loop->body)
646 cleanup_cf_node(child, impl);
647 break;
648 }
649 case nir_cf_node_function: {
650 nir_function_impl *impl = nir_cf_node_as_function(node);
651 foreach_list_typed(nir_cf_node, child, node, &impl->body)
652 cleanup_cf_node(child, impl);
653 break;
654 }
655 default:
656 unreachable("Invalid CF node type");
657 }
658 }
659
660 void
661 nir_cf_extract(nir_cf_list *extracted, nir_cursor begin, nir_cursor end)
662 {
663 nir_block *block_begin, *block_end, *block_before, *block_after;
664
665 if (nir_cursors_equal(begin, end)) {
666 exec_list_make_empty(&extracted->list);
667 extracted->impl = NULL; /* we shouldn't need this */
668 return;
669 }
670
671 /* In the case where begin points to an instruction in some basic block and
672 * end points to the end of the same basic block, we rely on the fact that
673 * splitting on an instruction moves earlier instructions into a new basic
674 * block. If the later instructions were moved instead, then the end cursor
675 * would be pointing to the same place that begin used to point to, which
676 * is obviously not what we want.
677 */
678 split_block_cursor(begin, &block_before, &block_begin);
679 split_block_cursor(end, &block_end, &block_after);
680
681 extracted->impl = nir_cf_node_get_function(&block_begin->cf_node);
682 exec_list_make_empty(&extracted->list);
683
684 /* Dominance and other block-related information is toast. */
685 nir_metadata_preserve(extracted->impl, nir_metadata_none);
686
687 nir_cf_node *cf_node = &block_begin->cf_node;
688 nir_cf_node *cf_node_end = &block_end->cf_node;
689 while (true) {
690 nir_cf_node *next = nir_cf_node_next(cf_node);
691
692 exec_node_remove(&cf_node->node);
693 cf_node->parent = NULL;
694 exec_list_push_tail(&extracted->list, &cf_node->node);
695
696 if (cf_node == cf_node_end)
697 break;
698
699 cf_node = next;
700 }
701
702 stitch_blocks(block_before, block_after);
703 }
704
705 void
706 nir_cf_reinsert(nir_cf_list *cf_list, nir_cursor cursor)
707 {
708 nir_block *before, *after;
709
710 if (exec_list_is_empty(&cf_list->list))
711 return;
712
713 split_block_cursor(cursor, &before, &after);
714
715 foreach_list_typed_safe(nir_cf_node, node, node, &cf_list->list) {
716 exec_node_remove(&node->node);
717 node->parent = before->cf_node.parent;
718 exec_node_insert_node_before(&after->cf_node.node, &node->node);
719 }
720
721 stitch_blocks(before,
722 nir_cf_node_as_block(nir_cf_node_next(&before->cf_node)));
723 stitch_blocks(nir_cf_node_as_block(nir_cf_node_prev(&after->cf_node)),
724 after);
725 }
726
727 void
728 nir_cf_delete(nir_cf_list *cf_list)
729 {
730 foreach_list_typed(nir_cf_node, node, node, &cf_list->list) {
731 cleanup_cf_node(node, cf_list->impl);
732 }
733 }