lto-cgraph.c (get_alias_symbol): Remove weakref sanity check.
[gcc.git] / gcc / cfgcleanup.c
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2013 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
10
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
19
20 /* This file contains optimizer of the control flow. The main entry point is
21 cleanup_cfg. Following optimizations are performed:
22
23 - Unreachable blocks removal
24 - Edge forwarding (edge to the forwarder block is forwarded to its
25 successor. Simplification of the branch instruction is performed by
26 underlying infrastructure so branch can be converted to simplejump or
27 eliminated).
28 - Cross jumping (tail merging)
29 - Conditional jump-around-simplejump simplification
30 - Basic block merging. */
31
32 #include "config.h"
33 #include "system.h"
34 #include "coretypes.h"
35 #include "tm.h"
36 #include "rtl.h"
37 #include "hard-reg-set.h"
38 #include "regs.h"
39 #include "insn-config.h"
40 #include "flags.h"
41 #include "recog.h"
42 #include "diagnostic-core.h"
43 #include "cselib.h"
44 #include "params.h"
45 #include "tm_p.h"
46 #include "target.h"
47 #include "function.h" /* For inline functions in emit-rtl.h they need crtl. */
48 #include "emit-rtl.h"
49 #include "tree-pass.h"
50 #include "cfgloop.h"
51 #include "expr.h"
52 #include "df.h"
53 #include "dce.h"
54 #include "dbgcnt.h"
55
56 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
57
58 /* Set to true when we are running first pass of try_optimize_cfg loop. */
59 static bool first_pass;
60
61 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
62 static bool crossjumps_occured;
63
64 /* Set to true if we couldn't run an optimization due to stale liveness
65 information; we should run df_analyze to enable more opportunities. */
66 static bool block_was_dirty;
67
68 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
69 static bool try_crossjump_bb (int, basic_block);
70 static bool outgoing_edges_match (int, basic_block, basic_block);
71 static enum replace_direction old_insns_match_p (int, rtx, rtx);
72
73 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
74 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
75 static bool try_optimize_cfg (int);
76 static bool try_simplify_condjump (basic_block);
77 static bool try_forward_edges (int, basic_block);
78 static edge thread_jump (edge, basic_block);
79 static bool mark_effect (rtx, bitmap);
80 static void notice_new_block (basic_block);
81 static void update_forwarder_flag (basic_block);
82 static int mentions_nonequal_regs (rtx *, void *);
83 static void merge_memattrs (rtx, rtx);
84 \f
85 /* Set flags for newly created block. */
86
87 static void
88 notice_new_block (basic_block bb)
89 {
90 if (!bb)
91 return;
92
93 if (forwarder_block_p (bb))
94 bb->flags |= BB_FORWARDER_BLOCK;
95 }
96
97 /* Recompute forwarder flag after block has been modified. */
98
99 static void
100 update_forwarder_flag (basic_block bb)
101 {
102 if (forwarder_block_p (bb))
103 bb->flags |= BB_FORWARDER_BLOCK;
104 else
105 bb->flags &= ~BB_FORWARDER_BLOCK;
106 }
107 \f
108 /* Simplify a conditional jump around an unconditional jump.
109 Return true if something changed. */
110
111 static bool
112 try_simplify_condjump (basic_block cbranch_block)
113 {
114 basic_block jump_block, jump_dest_block, cbranch_dest_block;
115 edge cbranch_jump_edge, cbranch_fallthru_edge;
116 rtx cbranch_insn;
117
118 /* Verify that there are exactly two successors. */
119 if (EDGE_COUNT (cbranch_block->succs) != 2)
120 return false;
121
122 /* Verify that we've got a normal conditional branch at the end
123 of the block. */
124 cbranch_insn = BB_END (cbranch_block);
125 if (!any_condjump_p (cbranch_insn))
126 return false;
127
128 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
129 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
130
131 /* The next block must not have multiple predecessors, must not
132 be the last block in the function, and must contain just the
133 unconditional jump. */
134 jump_block = cbranch_fallthru_edge->dest;
135 if (!single_pred_p (jump_block)
136 || jump_block->next_bb == EXIT_BLOCK_PTR
137 || !FORWARDER_BLOCK_P (jump_block))
138 return false;
139 jump_dest_block = single_succ (jump_block);
140
141 /* If we are partitioning hot/cold basic blocks, we don't want to
142 mess up unconditional or indirect jumps that cross between hot
143 and cold sections.
144
145 Basic block partitioning may result in some jumps that appear to
146 be optimizable (or blocks that appear to be mergeable), but which really
147 must be left untouched (they are required to make it safely across
148 partition boundaries). See the comments at the top of
149 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
150
151 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
152 || (cbranch_jump_edge->flags & EDGE_CROSSING))
153 return false;
154
155 /* The conditional branch must target the block after the
156 unconditional branch. */
157 cbranch_dest_block = cbranch_jump_edge->dest;
158
159 if (cbranch_dest_block == EXIT_BLOCK_PTR
160 || !can_fallthru (jump_block, cbranch_dest_block))
161 return false;
162
163 /* Invert the conditional branch. */
164 if (!invert_jump (cbranch_insn, block_label (jump_dest_block), 0))
165 return false;
166
167 if (dump_file)
168 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
169 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
170
171 /* Success. Update the CFG to match. Note that after this point
172 the edge variable names appear backwards; the redirection is done
173 this way to preserve edge profile data. */
174 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
175 cbranch_dest_block);
176 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
177 jump_dest_block);
178 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
179 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
180 update_br_prob_note (cbranch_block);
181
182 /* Delete the block with the unconditional jump, and clean up the mess. */
183 delete_basic_block (jump_block);
184 tidy_fallthru_edge (cbranch_jump_edge);
185 update_forwarder_flag (cbranch_block);
186
187 return true;
188 }
189 \f
190 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
191 on register. Used by jump threading. */
192
193 static bool
194 mark_effect (rtx exp, regset nonequal)
195 {
196 int regno;
197 rtx dest;
198 switch (GET_CODE (exp))
199 {
200 /* In case we do clobber the register, mark it as equal, as we know the
201 value is dead so it don't have to match. */
202 case CLOBBER:
203 if (REG_P (XEXP (exp, 0)))
204 {
205 dest = XEXP (exp, 0);
206 regno = REGNO (dest);
207 if (HARD_REGISTER_NUM_P (regno))
208 bitmap_clear_range (nonequal, regno,
209 hard_regno_nregs[regno][GET_MODE (dest)]);
210 else
211 bitmap_clear_bit (nonequal, regno);
212 }
213 return false;
214
215 case SET:
216 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
217 return false;
218 dest = SET_DEST (exp);
219 if (dest == pc_rtx)
220 return false;
221 if (!REG_P (dest))
222 return true;
223 regno = REGNO (dest);
224 if (HARD_REGISTER_NUM_P (regno))
225 bitmap_set_range (nonequal, regno,
226 hard_regno_nregs[regno][GET_MODE (dest)]);
227 else
228 bitmap_set_bit (nonequal, regno);
229 return false;
230
231 default:
232 return false;
233 }
234 }
235
236 /* Return nonzero if X is a register set in regset DATA.
237 Called via for_each_rtx. */
238 static int
239 mentions_nonequal_regs (rtx *x, void *data)
240 {
241 regset nonequal = (regset) data;
242 if (REG_P (*x))
243 {
244 int regno;
245
246 regno = REGNO (*x);
247 if (REGNO_REG_SET_P (nonequal, regno))
248 return 1;
249 if (regno < FIRST_PSEUDO_REGISTER)
250 {
251 int n = hard_regno_nregs[regno][GET_MODE (*x)];
252 while (--n > 0)
253 if (REGNO_REG_SET_P (nonequal, regno + n))
254 return 1;
255 }
256 }
257 return 0;
258 }
259 /* Attempt to prove that the basic block B will have no side effects and
260 always continues in the same edge if reached via E. Return the edge
261 if exist, NULL otherwise. */
262
263 static edge
264 thread_jump (edge e, basic_block b)
265 {
266 rtx set1, set2, cond1, cond2, insn;
267 enum rtx_code code1, code2, reversed_code2;
268 bool reverse1 = false;
269 unsigned i;
270 regset nonequal;
271 bool failed = false;
272 reg_set_iterator rsi;
273
274 if (b->flags & BB_NONTHREADABLE_BLOCK)
275 return NULL;
276
277 /* At the moment, we do handle only conditional jumps, but later we may
278 want to extend this code to tablejumps and others. */
279 if (EDGE_COUNT (e->src->succs) != 2)
280 return NULL;
281 if (EDGE_COUNT (b->succs) != 2)
282 {
283 b->flags |= BB_NONTHREADABLE_BLOCK;
284 return NULL;
285 }
286
287 /* Second branch must end with onlyjump, as we will eliminate the jump. */
288 if (!any_condjump_p (BB_END (e->src)))
289 return NULL;
290
291 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
292 {
293 b->flags |= BB_NONTHREADABLE_BLOCK;
294 return NULL;
295 }
296
297 set1 = pc_set (BB_END (e->src));
298 set2 = pc_set (BB_END (b));
299 if (((e->flags & EDGE_FALLTHRU) != 0)
300 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
301 reverse1 = true;
302
303 cond1 = XEXP (SET_SRC (set1), 0);
304 cond2 = XEXP (SET_SRC (set2), 0);
305 if (reverse1)
306 code1 = reversed_comparison_code (cond1, BB_END (e->src));
307 else
308 code1 = GET_CODE (cond1);
309
310 code2 = GET_CODE (cond2);
311 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
312
313 if (!comparison_dominates_p (code1, code2)
314 && !comparison_dominates_p (code1, reversed_code2))
315 return NULL;
316
317 /* Ensure that the comparison operators are equivalent.
318 ??? This is far too pessimistic. We should allow swapped operands,
319 different CCmodes, or for example comparisons for interval, that
320 dominate even when operands are not equivalent. */
321 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
322 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
323 return NULL;
324
325 /* Short circuit cases where block B contains some side effects, as we can't
326 safely bypass it. */
327 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
328 insn = NEXT_INSN (insn))
329 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
330 {
331 b->flags |= BB_NONTHREADABLE_BLOCK;
332 return NULL;
333 }
334
335 cselib_init (0);
336
337 /* First process all values computed in the source basic block. */
338 for (insn = NEXT_INSN (BB_HEAD (e->src));
339 insn != NEXT_INSN (BB_END (e->src));
340 insn = NEXT_INSN (insn))
341 if (INSN_P (insn))
342 cselib_process_insn (insn);
343
344 nonequal = BITMAP_ALLOC (NULL);
345 CLEAR_REG_SET (nonequal);
346
347 /* Now assume that we've continued by the edge E to B and continue
348 processing as if it were same basic block.
349 Our goal is to prove that whole block is an NOOP. */
350
351 for (insn = NEXT_INSN (BB_HEAD (b));
352 insn != NEXT_INSN (BB_END (b)) && !failed;
353 insn = NEXT_INSN (insn))
354 {
355 if (INSN_P (insn))
356 {
357 rtx pat = PATTERN (insn);
358
359 if (GET_CODE (pat) == PARALLEL)
360 {
361 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
362 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
363 }
364 else
365 failed |= mark_effect (pat, nonequal);
366 }
367
368 cselib_process_insn (insn);
369 }
370
371 /* Later we should clear nonequal of dead registers. So far we don't
372 have life information in cfg_cleanup. */
373 if (failed)
374 {
375 b->flags |= BB_NONTHREADABLE_BLOCK;
376 goto failed_exit;
377 }
378
379 /* cond2 must not mention any register that is not equal to the
380 former block. */
381 if (for_each_rtx (&cond2, mentions_nonequal_regs, nonequal))
382 goto failed_exit;
383
384 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
385 goto failed_exit;
386
387 BITMAP_FREE (nonequal);
388 cselib_finish ();
389 if ((comparison_dominates_p (code1, code2) != 0)
390 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
391 return BRANCH_EDGE (b);
392 else
393 return FALLTHRU_EDGE (b);
394
395 failed_exit:
396 BITMAP_FREE (nonequal);
397 cselib_finish ();
398 return NULL;
399 }
400 \f
401 /* Attempt to forward edges leaving basic block B.
402 Return true if successful. */
403
404 static bool
405 try_forward_edges (int mode, basic_block b)
406 {
407 bool changed = false;
408 edge_iterator ei;
409 edge e, *threaded_edges = NULL;
410
411 /* If we are partitioning hot/cold basic blocks, we don't want to
412 mess up unconditional or indirect jumps that cross between hot
413 and cold sections.
414
415 Basic block partitioning may result in some jumps that appear to
416 be optimizable (or blocks that appear to be mergeable), but which really
417 must be left untouched (they are required to make it safely across
418 partition boundaries). See the comments at the top of
419 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
420
421 if (find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX))
422 return false;
423
424 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
425 {
426 basic_block target, first;
427 int counter, goto_locus;
428 bool threaded = false;
429 int nthreaded_edges = 0;
430 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
431
432 /* Skip complex edges because we don't know how to update them.
433
434 Still handle fallthru edges, as we can succeed to forward fallthru
435 edge to the same place as the branch edge of conditional branch
436 and turn conditional branch to an unconditional branch. */
437 if (e->flags & EDGE_COMPLEX)
438 {
439 ei_next (&ei);
440 continue;
441 }
442
443 target = first = e->dest;
444 counter = NUM_FIXED_BLOCKS;
445 goto_locus = e->goto_locus;
446
447 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
448 up jumps that cross between hot/cold sections.
449
450 Basic block partitioning may result in some jumps that appear
451 to be optimizable (or blocks that appear to be mergeable), but which
452 really must be left untouched (they are required to make it safely
453 across partition boundaries). See the comments at the top of
454 bb-reorder.c:partition_hot_cold_basic_blocks for complete
455 details. */
456
457 if (first != EXIT_BLOCK_PTR
458 && find_reg_note (BB_END (first), REG_CROSSING_JUMP, NULL_RTX))
459 return false;
460
461 while (counter < n_basic_blocks)
462 {
463 basic_block new_target = NULL;
464 bool new_target_threaded = false;
465 may_thread |= (target->flags & BB_MODIFIED) != 0;
466
467 if (FORWARDER_BLOCK_P (target)
468 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
469 && single_succ (target) != EXIT_BLOCK_PTR)
470 {
471 /* Bypass trivial infinite loops. */
472 new_target = single_succ (target);
473 if (target == new_target)
474 counter = n_basic_blocks;
475 else if (!optimize)
476 {
477 /* When not optimizing, ensure that edges or forwarder
478 blocks with different locus are not optimized out. */
479 int new_locus = single_succ_edge (target)->goto_locus;
480 int locus = goto_locus;
481
482 if (new_locus != UNKNOWN_LOCATION
483 && locus != UNKNOWN_LOCATION
484 && new_locus != locus)
485 new_target = NULL;
486 else
487 {
488 rtx last;
489
490 if (new_locus != UNKNOWN_LOCATION)
491 locus = new_locus;
492
493 last = BB_END (target);
494 if (DEBUG_INSN_P (last))
495 last = prev_nondebug_insn (last);
496
497 new_locus = last && INSN_P (last)
498 ? INSN_LOCATION (last) : 0;
499
500 if (new_locus != UNKNOWN_LOCATION
501 && locus != UNKNOWN_LOCATION
502 && new_locus != locus)
503 new_target = NULL;
504 else
505 {
506 if (new_locus != UNKNOWN_LOCATION)
507 locus = new_locus;
508
509 goto_locus = locus;
510 }
511 }
512 }
513 }
514
515 /* Allow to thread only over one edge at time to simplify updating
516 of probabilities. */
517 else if ((mode & CLEANUP_THREADING) && may_thread)
518 {
519 edge t = thread_jump (e, target);
520 if (t)
521 {
522 if (!threaded_edges)
523 threaded_edges = XNEWVEC (edge, n_basic_blocks);
524 else
525 {
526 int i;
527
528 /* Detect an infinite loop across blocks not
529 including the start block. */
530 for (i = 0; i < nthreaded_edges; ++i)
531 if (threaded_edges[i] == t)
532 break;
533 if (i < nthreaded_edges)
534 {
535 counter = n_basic_blocks;
536 break;
537 }
538 }
539
540 /* Detect an infinite loop across the start block. */
541 if (t->dest == b)
542 break;
543
544 gcc_assert (nthreaded_edges < n_basic_blocks - NUM_FIXED_BLOCKS);
545 threaded_edges[nthreaded_edges++] = t;
546
547 new_target = t->dest;
548 new_target_threaded = true;
549 }
550 }
551
552 if (!new_target)
553 break;
554
555 counter++;
556 target = new_target;
557 threaded |= new_target_threaded;
558 }
559
560 if (counter >= n_basic_blocks)
561 {
562 if (dump_file)
563 fprintf (dump_file, "Infinite loop in BB %i.\n",
564 target->index);
565 }
566 else if (target == first)
567 ; /* We didn't do anything. */
568 else
569 {
570 /* Save the values now, as the edge may get removed. */
571 gcov_type edge_count = e->count;
572 int edge_probability = e->probability;
573 int edge_frequency;
574 int n = 0;
575
576 e->goto_locus = goto_locus;
577
578 /* Don't force if target is exit block. */
579 if (threaded && target != EXIT_BLOCK_PTR)
580 {
581 notice_new_block (redirect_edge_and_branch_force (e, target));
582 if (dump_file)
583 fprintf (dump_file, "Conditionals threaded.\n");
584 }
585 else if (!redirect_edge_and_branch (e, target))
586 {
587 if (dump_file)
588 fprintf (dump_file,
589 "Forwarding edge %i->%i to %i failed.\n",
590 b->index, e->dest->index, target->index);
591 ei_next (&ei);
592 continue;
593 }
594
595 /* We successfully forwarded the edge. Now update profile
596 data: for each edge we traversed in the chain, remove
597 the original edge's execution count. */
598 edge_frequency = apply_probability (b->frequency, edge_probability);
599
600 do
601 {
602 edge t;
603
604 if (!single_succ_p (first))
605 {
606 gcc_assert (n < nthreaded_edges);
607 t = threaded_edges [n++];
608 gcc_assert (t->src == first);
609 update_bb_profile_for_threading (first, edge_frequency,
610 edge_count, t);
611 update_br_prob_note (first);
612 }
613 else
614 {
615 first->count -= edge_count;
616 if (first->count < 0)
617 first->count = 0;
618 first->frequency -= edge_frequency;
619 if (first->frequency < 0)
620 first->frequency = 0;
621 /* It is possible that as the result of
622 threading we've removed edge as it is
623 threaded to the fallthru edge. Avoid
624 getting out of sync. */
625 if (n < nthreaded_edges
626 && first == threaded_edges [n]->src)
627 n++;
628 t = single_succ_edge (first);
629 }
630
631 t->count -= edge_count;
632 if (t->count < 0)
633 t->count = 0;
634 first = t->dest;
635 }
636 while (first != target);
637
638 changed = true;
639 continue;
640 }
641 ei_next (&ei);
642 }
643
644 free (threaded_edges);
645 return changed;
646 }
647 \f
648
649 /* Blocks A and B are to be merged into a single block. A has no incoming
650 fallthru edge, so it can be moved before B without adding or modifying
651 any jumps (aside from the jump from A to B). */
652
653 static void
654 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
655 {
656 rtx barrier;
657
658 /* If we are partitioning hot/cold basic blocks, we don't want to
659 mess up unconditional or indirect jumps that cross between hot
660 and cold sections.
661
662 Basic block partitioning may result in some jumps that appear to
663 be optimizable (or blocks that appear to be mergeable), but which really
664 must be left untouched (they are required to make it safely across
665 partition boundaries). See the comments at the top of
666 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
667
668 if (BB_PARTITION (a) != BB_PARTITION (b))
669 return;
670
671 barrier = next_nonnote_insn (BB_END (a));
672 gcc_assert (BARRIER_P (barrier));
673 delete_insn (barrier);
674
675 /* Scramble the insn chain. */
676 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
677 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
678 df_set_bb_dirty (a);
679
680 if (dump_file)
681 fprintf (dump_file, "Moved block %d before %d and merged.\n",
682 a->index, b->index);
683
684 /* Swap the records for the two blocks around. */
685
686 unlink_block (a);
687 link_block (a, b->prev_bb);
688
689 /* Now blocks A and B are contiguous. Merge them. */
690 merge_blocks (a, b);
691 }
692
693 /* Blocks A and B are to be merged into a single block. B has no outgoing
694 fallthru edge, so it can be moved after A without adding or modifying
695 any jumps (aside from the jump from A to B). */
696
697 static void
698 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
699 {
700 rtx barrier, real_b_end;
701 rtx label, table;
702
703 /* If we are partitioning hot/cold basic blocks, we don't want to
704 mess up unconditional or indirect jumps that cross between hot
705 and cold sections.
706
707 Basic block partitioning may result in some jumps that appear to
708 be optimizable (or blocks that appear to be mergeable), but which really
709 must be left untouched (they are required to make it safely across
710 partition boundaries). See the comments at the top of
711 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
712
713 if (BB_PARTITION (a) != BB_PARTITION (b))
714 return;
715
716 real_b_end = BB_END (b);
717
718 /* If there is a jump table following block B temporarily add the jump table
719 to block B so that it will also be moved to the correct location. */
720 if (tablejump_p (BB_END (b), &label, &table)
721 && prev_active_insn (label) == BB_END (b))
722 {
723 BB_END (b) = table;
724 }
725
726 /* There had better have been a barrier there. Delete it. */
727 barrier = NEXT_INSN (BB_END (b));
728 if (barrier && BARRIER_P (barrier))
729 delete_insn (barrier);
730
731
732 /* Scramble the insn chain. */
733 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
734
735 /* Restore the real end of b. */
736 BB_END (b) = real_b_end;
737
738 if (dump_file)
739 fprintf (dump_file, "Moved block %d after %d and merged.\n",
740 b->index, a->index);
741
742 /* Now blocks A and B are contiguous. Merge them. */
743 merge_blocks (a, b);
744 }
745
746 /* Attempt to merge basic blocks that are potentially non-adjacent.
747 Return NULL iff the attempt failed, otherwise return basic block
748 where cleanup_cfg should continue. Because the merging commonly
749 moves basic block away or introduces another optimization
750 possibility, return basic block just before B so cleanup_cfg don't
751 need to iterate.
752
753 It may be good idea to return basic block before C in the case
754 C has been moved after B and originally appeared earlier in the
755 insn sequence, but we have no information available about the
756 relative ordering of these two. Hopefully it is not too common. */
757
758 static basic_block
759 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
760 {
761 basic_block next;
762
763 /* If we are partitioning hot/cold basic blocks, we don't want to
764 mess up unconditional or indirect jumps that cross between hot
765 and cold sections.
766
767 Basic block partitioning may result in some jumps that appear to
768 be optimizable (or blocks that appear to be mergeable), but which really
769 must be left untouched (they are required to make it safely across
770 partition boundaries). See the comments at the top of
771 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
772
773 if (BB_PARTITION (b) != BB_PARTITION (c))
774 return NULL;
775
776 /* If B has a fallthru edge to C, no need to move anything. */
777 if (e->flags & EDGE_FALLTHRU)
778 {
779 int b_index = b->index, c_index = c->index;
780
781 /* Protect the loop latches. */
782 if (current_loops && c->loop_father->latch == c)
783 return NULL;
784
785 merge_blocks (b, c);
786 update_forwarder_flag (b);
787
788 if (dump_file)
789 fprintf (dump_file, "Merged %d and %d without moving.\n",
790 b_index, c_index);
791
792 return b->prev_bb == ENTRY_BLOCK_PTR ? b : b->prev_bb;
793 }
794
795 /* Otherwise we will need to move code around. Do that only if expensive
796 transformations are allowed. */
797 else if (mode & CLEANUP_EXPENSIVE)
798 {
799 edge tmp_edge, b_fallthru_edge;
800 bool c_has_outgoing_fallthru;
801 bool b_has_incoming_fallthru;
802
803 /* Avoid overactive code motion, as the forwarder blocks should be
804 eliminated by edge redirection instead. One exception might have
805 been if B is a forwarder block and C has no fallthru edge, but
806 that should be cleaned up by bb-reorder instead. */
807 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
808 return NULL;
809
810 /* We must make sure to not munge nesting of lexical blocks,
811 and loop notes. This is done by squeezing out all the notes
812 and leaving them there to lie. Not ideal, but functional. */
813
814 tmp_edge = find_fallthru_edge (c->succs);
815 c_has_outgoing_fallthru = (tmp_edge != NULL);
816
817 tmp_edge = find_fallthru_edge (b->preds);
818 b_has_incoming_fallthru = (tmp_edge != NULL);
819 b_fallthru_edge = tmp_edge;
820 next = b->prev_bb;
821 if (next == c)
822 next = next->prev_bb;
823
824 /* Otherwise, we're going to try to move C after B. If C does
825 not have an outgoing fallthru, then it can be moved
826 immediately after B without introducing or modifying jumps. */
827 if (! c_has_outgoing_fallthru)
828 {
829 merge_blocks_move_successor_nojumps (b, c);
830 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
831 }
832
833 /* If B does not have an incoming fallthru, then it can be moved
834 immediately before C without introducing or modifying jumps.
835 C cannot be the first block, so we do not have to worry about
836 accessing a non-existent block. */
837
838 if (b_has_incoming_fallthru)
839 {
840 basic_block bb;
841
842 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR)
843 return NULL;
844 bb = force_nonfallthru (b_fallthru_edge);
845 if (bb)
846 notice_new_block (bb);
847 }
848
849 merge_blocks_move_predecessor_nojumps (b, c);
850 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
851 }
852
853 return NULL;
854 }
855 \f
856
857 /* Removes the memory attributes of MEM expression
858 if they are not equal. */
859
860 void
861 merge_memattrs (rtx x, rtx y)
862 {
863 int i;
864 int j;
865 enum rtx_code code;
866 const char *fmt;
867
868 if (x == y)
869 return;
870 if (x == 0 || y == 0)
871 return;
872
873 code = GET_CODE (x);
874
875 if (code != GET_CODE (y))
876 return;
877
878 if (GET_MODE (x) != GET_MODE (y))
879 return;
880
881 if (code == MEM && MEM_ATTRS (x) != MEM_ATTRS (y))
882 {
883 if (! MEM_ATTRS (x))
884 MEM_ATTRS (y) = 0;
885 else if (! MEM_ATTRS (y))
886 MEM_ATTRS (x) = 0;
887 else
888 {
889 HOST_WIDE_INT mem_size;
890
891 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
892 {
893 set_mem_alias_set (x, 0);
894 set_mem_alias_set (y, 0);
895 }
896
897 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
898 {
899 set_mem_expr (x, 0);
900 set_mem_expr (y, 0);
901 clear_mem_offset (x);
902 clear_mem_offset (y);
903 }
904 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
905 || (MEM_OFFSET_KNOWN_P (x)
906 && MEM_OFFSET (x) != MEM_OFFSET (y)))
907 {
908 clear_mem_offset (x);
909 clear_mem_offset (y);
910 }
911
912 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
913 {
914 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
915 set_mem_size (x, mem_size);
916 set_mem_size (y, mem_size);
917 }
918 else
919 {
920 clear_mem_size (x);
921 clear_mem_size (y);
922 }
923
924 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
925 set_mem_align (y, MEM_ALIGN (x));
926 }
927 }
928
929 fmt = GET_RTX_FORMAT (code);
930 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
931 {
932 switch (fmt[i])
933 {
934 case 'E':
935 /* Two vectors must have the same length. */
936 if (XVECLEN (x, i) != XVECLEN (y, i))
937 return;
938
939 for (j = 0; j < XVECLEN (x, i); j++)
940 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
941
942 break;
943
944 case 'e':
945 merge_memattrs (XEXP (x, i), XEXP (y, i));
946 }
947 }
948 return;
949 }
950
951
952 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
953 different single sets S1 and S2. */
954
955 static bool
956 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
957 {
958 int i;
959 rtx e1, e2;
960
961 if (p1 == s1 && p2 == s2)
962 return true;
963
964 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
965 return false;
966
967 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
968 return false;
969
970 for (i = 0; i < XVECLEN (p1, 0); i++)
971 {
972 e1 = XVECEXP (p1, 0, i);
973 e2 = XVECEXP (p2, 0, i);
974 if (e1 == s1 && e2 == s2)
975 continue;
976 if (reload_completed
977 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
978 continue;
979
980 return false;
981 }
982
983 return true;
984 }
985
986 /* Examine register notes on I1 and I2 and return:
987 - dir_forward if I1 can be replaced by I2, or
988 - dir_backward if I2 can be replaced by I1, or
989 - dir_both if both are the case. */
990
991 static enum replace_direction
992 can_replace_by (rtx i1, rtx i2)
993 {
994 rtx s1, s2, d1, d2, src1, src2, note1, note2;
995 bool c1, c2;
996
997 /* Check for 2 sets. */
998 s1 = single_set (i1);
999 s2 = single_set (i2);
1000 if (s1 == NULL_RTX || s2 == NULL_RTX)
1001 return dir_none;
1002
1003 /* Check that the 2 sets set the same dest. */
1004 d1 = SET_DEST (s1);
1005 d2 = SET_DEST (s2);
1006 if (!(reload_completed
1007 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1008 return dir_none;
1009
1010 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1011 set dest to the same value. */
1012 note1 = find_reg_equal_equiv_note (i1);
1013 note2 = find_reg_equal_equiv_note (i2);
1014 if (!note1 || !note2 || !rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0))
1015 || !CONST_INT_P (XEXP (note1, 0)))
1016 return dir_none;
1017
1018 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1019 return dir_none;
1020
1021 /* Although the 2 sets set dest to the same value, we cannot replace
1022 (set (dest) (const_int))
1023 by
1024 (set (dest) (reg))
1025 because we don't know if the reg is live and has the same value at the
1026 location of replacement. */
1027 src1 = SET_SRC (s1);
1028 src2 = SET_SRC (s2);
1029 c1 = CONST_INT_P (src1);
1030 c2 = CONST_INT_P (src2);
1031 if (c1 && c2)
1032 return dir_both;
1033 else if (c2)
1034 return dir_forward;
1035 else if (c1)
1036 return dir_backward;
1037
1038 return dir_none;
1039 }
1040
1041 /* Merges directions A and B. */
1042
1043 static enum replace_direction
1044 merge_dir (enum replace_direction a, enum replace_direction b)
1045 {
1046 /* Implements the following table:
1047 |bo fw bw no
1048 ---+-----------
1049 bo |bo fw bw no
1050 fw |-- fw no no
1051 bw |-- -- bw no
1052 no |-- -- -- no. */
1053
1054 if (a == b)
1055 return a;
1056
1057 if (a == dir_both)
1058 return b;
1059 if (b == dir_both)
1060 return a;
1061
1062 return dir_none;
1063 }
1064
1065 /* Examine I1 and I2 and return:
1066 - dir_forward if I1 can be replaced by I2, or
1067 - dir_backward if I2 can be replaced by I1, or
1068 - dir_both if both are the case. */
1069
1070 static enum replace_direction
1071 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx i1, rtx i2)
1072 {
1073 rtx p1, p2;
1074
1075 /* Verify that I1 and I2 are equivalent. */
1076 if (GET_CODE (i1) != GET_CODE (i2))
1077 return dir_none;
1078
1079 /* __builtin_unreachable() may lead to empty blocks (ending with
1080 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1081 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1082 return dir_both;
1083
1084 /* ??? Do not allow cross-jumping between different stack levels. */
1085 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1086 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1087 if (p1 && p2)
1088 {
1089 p1 = XEXP (p1, 0);
1090 p2 = XEXP (p2, 0);
1091 if (!rtx_equal_p (p1, p2))
1092 return dir_none;
1093
1094 /* ??? Worse, this adjustment had better be constant lest we
1095 have differing incoming stack levels. */
1096 if (!frame_pointer_needed
1097 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1098 return dir_none;
1099 }
1100 else if (p1 || p2)
1101 return dir_none;
1102
1103 p1 = PATTERN (i1);
1104 p2 = PATTERN (i2);
1105
1106 if (GET_CODE (p1) != GET_CODE (p2))
1107 return dir_none;
1108
1109 /* If this is a CALL_INSN, compare register usage information.
1110 If we don't check this on stack register machines, the two
1111 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1112 numbers of stack registers in the same basic block.
1113 If we don't check this on machines with delay slots, a delay slot may
1114 be filled that clobbers a parameter expected by the subroutine.
1115
1116 ??? We take the simple route for now and assume that if they're
1117 equal, they were constructed identically.
1118
1119 Also check for identical exception regions. */
1120
1121 if (CALL_P (i1))
1122 {
1123 /* Ensure the same EH region. */
1124 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1125 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1126
1127 if (!n1 && n2)
1128 return dir_none;
1129
1130 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1131 return dir_none;
1132
1133 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1134 CALL_INSN_FUNCTION_USAGE (i2))
1135 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1136 return dir_none;
1137
1138 /* For address sanitizer, never crossjump __asan_report_* builtins,
1139 otherwise errors might be reported on incorrect lines. */
1140 if (flag_asan)
1141 {
1142 rtx call = get_call_rtx_from (i1);
1143 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1144 {
1145 rtx symbol = XEXP (XEXP (call, 0), 0);
1146 if (SYMBOL_REF_DECL (symbol)
1147 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1148 {
1149 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1150 == BUILT_IN_NORMAL)
1151 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1152 >= BUILT_IN_ASAN_REPORT_LOAD1
1153 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1154 <= BUILT_IN_ASAN_REPORT_STORE16)
1155 return dir_none;
1156 }
1157 }
1158 }
1159 }
1160
1161 #ifdef STACK_REGS
1162 /* If cross_jump_death_matters is not 0, the insn's mode
1163 indicates whether or not the insn contains any stack-like
1164 regs. */
1165
1166 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1167 {
1168 /* If register stack conversion has already been done, then
1169 death notes must also be compared before it is certain that
1170 the two instruction streams match. */
1171
1172 rtx note;
1173 HARD_REG_SET i1_regset, i2_regset;
1174
1175 CLEAR_HARD_REG_SET (i1_regset);
1176 CLEAR_HARD_REG_SET (i2_regset);
1177
1178 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1179 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1180 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1181
1182 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1183 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1184 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1185
1186 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1187 return dir_none;
1188 }
1189 #endif
1190
1191 if (reload_completed
1192 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1193 return dir_both;
1194
1195 return can_replace_by (i1, i2);
1196 }
1197 \f
1198 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1199 flow_find_head_matching_sequence, ensure the notes match. */
1200
1201 static void
1202 merge_notes (rtx i1, rtx i2)
1203 {
1204 /* If the merged insns have different REG_EQUAL notes, then
1205 remove them. */
1206 rtx equiv1 = find_reg_equal_equiv_note (i1);
1207 rtx equiv2 = find_reg_equal_equiv_note (i2);
1208
1209 if (equiv1 && !equiv2)
1210 remove_note (i1, equiv1);
1211 else if (!equiv1 && equiv2)
1212 remove_note (i2, equiv2);
1213 else if (equiv1 && equiv2
1214 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1215 {
1216 remove_note (i1, equiv1);
1217 remove_note (i2, equiv2);
1218 }
1219 }
1220
1221 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1222 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1223 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1224 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1225 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1226
1227 static void
1228 walk_to_nondebug_insn (rtx *i1, basic_block *bb1, bool follow_fallthru,
1229 bool *did_fallthru)
1230 {
1231 edge fallthru;
1232
1233 *did_fallthru = false;
1234
1235 /* Ignore notes. */
1236 while (!NONDEBUG_INSN_P (*i1))
1237 {
1238 if (*i1 != BB_HEAD (*bb1))
1239 {
1240 *i1 = PREV_INSN (*i1);
1241 continue;
1242 }
1243
1244 if (!follow_fallthru)
1245 return;
1246
1247 fallthru = find_fallthru_edge ((*bb1)->preds);
1248 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun)
1249 || !single_succ_p (fallthru->src))
1250 return;
1251
1252 *bb1 = fallthru->src;
1253 *i1 = BB_END (*bb1);
1254 *did_fallthru = true;
1255 }
1256 }
1257
1258 /* Look through the insns at the end of BB1 and BB2 and find the longest
1259 sequence that are either equivalent, or allow forward or backward
1260 replacement. Store the first insns for that sequence in *F1 and *F2 and
1261 return the sequence length.
1262
1263 DIR_P indicates the allowed replacement direction on function entry, and
1264 the actual replacement direction on function exit. If NULL, only equivalent
1265 sequences are allowed.
1266
1267 To simplify callers of this function, if the blocks match exactly,
1268 store the head of the blocks in *F1 and *F2. */
1269
1270 int
1271 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx *f1, rtx *f2,
1272 enum replace_direction *dir_p)
1273 {
1274 rtx i1, i2, last1, last2, afterlast1, afterlast2;
1275 int ninsns = 0;
1276 rtx p1;
1277 enum replace_direction dir, last_dir, afterlast_dir;
1278 bool follow_fallthru, did_fallthru;
1279
1280 if (dir_p)
1281 dir = *dir_p;
1282 else
1283 dir = dir_both;
1284 afterlast_dir = dir;
1285 last_dir = afterlast_dir;
1286
1287 /* Skip simple jumps at the end of the blocks. Complex jumps still
1288 need to be compared for equivalence, which we'll do below. */
1289
1290 i1 = BB_END (bb1);
1291 last1 = afterlast1 = last2 = afterlast2 = NULL_RTX;
1292 if (onlyjump_p (i1)
1293 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1294 {
1295 last1 = i1;
1296 i1 = PREV_INSN (i1);
1297 }
1298
1299 i2 = BB_END (bb2);
1300 if (onlyjump_p (i2)
1301 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1302 {
1303 last2 = i2;
1304 /* Count everything except for unconditional jump as insn. */
1305 if (!simplejump_p (i2) && !returnjump_p (i2) && last1)
1306 ninsns++;
1307 i2 = PREV_INSN (i2);
1308 }
1309
1310 while (true)
1311 {
1312 /* In the following example, we can replace all jumps to C by jumps to A.
1313
1314 This removes 4 duplicate insns.
1315 [bb A] insn1 [bb C] insn1
1316 insn2 insn2
1317 [bb B] insn3 insn3
1318 insn4 insn4
1319 jump_insn jump_insn
1320
1321 We could also replace all jumps to A by jumps to C, but that leaves B
1322 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1323 step, all jumps to B would be replaced with jumps to the middle of C,
1324 achieving the same result with more effort.
1325 So we allow only the first possibility, which means that we don't allow
1326 fallthru in the block that's being replaced. */
1327
1328 follow_fallthru = dir_p && dir != dir_forward;
1329 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1330 if (did_fallthru)
1331 dir = dir_backward;
1332
1333 follow_fallthru = dir_p && dir != dir_backward;
1334 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1335 if (did_fallthru)
1336 dir = dir_forward;
1337
1338 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1339 break;
1340
1341 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1342 if (dir == dir_none || (!dir_p && dir != dir_both))
1343 break;
1344
1345 merge_memattrs (i1, i2);
1346
1347 /* Don't begin a cross-jump with a NOTE insn. */
1348 if (INSN_P (i1))
1349 {
1350 merge_notes (i1, i2);
1351
1352 afterlast1 = last1, afterlast2 = last2;
1353 last1 = i1, last2 = i2;
1354 afterlast_dir = last_dir;
1355 last_dir = dir;
1356 p1 = PATTERN (i1);
1357 if (!(GET_CODE (p1) == USE || GET_CODE (p1) == CLOBBER))
1358 ninsns++;
1359 }
1360
1361 i1 = PREV_INSN (i1);
1362 i2 = PREV_INSN (i2);
1363 }
1364
1365 #ifdef HAVE_cc0
1366 /* Don't allow the insn after a compare to be shared by
1367 cross-jumping unless the compare is also shared. */
1368 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
1369 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1370 #endif
1371
1372 /* Include preceding notes and labels in the cross-jump. One,
1373 this may bring us to the head of the blocks as requested above.
1374 Two, it keeps line number notes as matched as may be. */
1375 if (ninsns)
1376 {
1377 bb1 = BLOCK_FOR_INSN (last1);
1378 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1379 last1 = PREV_INSN (last1);
1380
1381 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1382 last1 = PREV_INSN (last1);
1383
1384 bb2 = BLOCK_FOR_INSN (last2);
1385 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1386 last2 = PREV_INSN (last2);
1387
1388 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1389 last2 = PREV_INSN (last2);
1390
1391 *f1 = last1;
1392 *f2 = last2;
1393 }
1394
1395 if (dir_p)
1396 *dir_p = last_dir;
1397 return ninsns;
1398 }
1399
1400 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1401 the head of the two blocks. Do not include jumps at the end.
1402 If STOP_AFTER is nonzero, stop after finding that many matching
1403 instructions. */
1404
1405 int
1406 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx *f1,
1407 rtx *f2, int stop_after)
1408 {
1409 rtx i1, i2, last1, last2, beforelast1, beforelast2;
1410 int ninsns = 0;
1411 edge e;
1412 edge_iterator ei;
1413 int nehedges1 = 0, nehedges2 = 0;
1414
1415 FOR_EACH_EDGE (e, ei, bb1->succs)
1416 if (e->flags & EDGE_EH)
1417 nehedges1++;
1418 FOR_EACH_EDGE (e, ei, bb2->succs)
1419 if (e->flags & EDGE_EH)
1420 nehedges2++;
1421
1422 i1 = BB_HEAD (bb1);
1423 i2 = BB_HEAD (bb2);
1424 last1 = beforelast1 = last2 = beforelast2 = NULL_RTX;
1425
1426 while (true)
1427 {
1428 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1429 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1430 {
1431 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1432 break;
1433 i1 = NEXT_INSN (i1);
1434 }
1435
1436 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1437 {
1438 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1439 break;
1440 i2 = NEXT_INSN (i2);
1441 }
1442
1443 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1444 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1445 break;
1446
1447 if (NOTE_P (i1) || NOTE_P (i2)
1448 || JUMP_P (i1) || JUMP_P (i2))
1449 break;
1450
1451 /* A sanity check to make sure we're not merging insns with different
1452 effects on EH. If only one of them ends a basic block, it shouldn't
1453 have an EH edge; if both end a basic block, there should be the same
1454 number of EH edges. */
1455 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1456 && nehedges1 > 0)
1457 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1458 && nehedges2 > 0)
1459 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1460 && nehedges1 != nehedges2))
1461 break;
1462
1463 if (old_insns_match_p (0, i1, i2) != dir_both)
1464 break;
1465
1466 merge_memattrs (i1, i2);
1467
1468 /* Don't begin a cross-jump with a NOTE insn. */
1469 if (INSN_P (i1))
1470 {
1471 merge_notes (i1, i2);
1472
1473 beforelast1 = last1, beforelast2 = last2;
1474 last1 = i1, last2 = i2;
1475 ninsns++;
1476 }
1477
1478 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1479 || (stop_after > 0 && ninsns == stop_after))
1480 break;
1481
1482 i1 = NEXT_INSN (i1);
1483 i2 = NEXT_INSN (i2);
1484 }
1485
1486 #ifdef HAVE_cc0
1487 /* Don't allow a compare to be shared by cross-jumping unless the insn
1488 after the compare is also shared. */
1489 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && sets_cc0_p (last1))
1490 last1 = beforelast1, last2 = beforelast2, ninsns--;
1491 #endif
1492
1493 if (ninsns)
1494 {
1495 *f1 = last1;
1496 *f2 = last2;
1497 }
1498
1499 return ninsns;
1500 }
1501
1502 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1503 the branch instruction. This means that if we commonize the control
1504 flow before end of the basic block, the semantic remains unchanged.
1505
1506 We may assume that there exists one edge with a common destination. */
1507
1508 static bool
1509 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1510 {
1511 int nehedges1 = 0, nehedges2 = 0;
1512 edge fallthru1 = 0, fallthru2 = 0;
1513 edge e1, e2;
1514 edge_iterator ei;
1515
1516 /* If we performed shrink-wrapping, edges to the EXIT_BLOCK_PTR can
1517 only be distinguished for JUMP_INSNs. The two paths may differ in
1518 whether they went through the prologue. Sibcalls are fine, we know
1519 that we either didn't need or inserted an epilogue before them. */
1520 if (crtl->shrink_wrapped
1521 && single_succ_p (bb1) && single_succ (bb1) == EXIT_BLOCK_PTR
1522 && !JUMP_P (BB_END (bb1))
1523 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1524 return false;
1525
1526 /* If BB1 has only one successor, we may be looking at either an
1527 unconditional jump, or a fake edge to exit. */
1528 if (single_succ_p (bb1)
1529 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1530 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1531 return (single_succ_p (bb2)
1532 && (single_succ_edge (bb2)->flags
1533 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1534 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1535
1536 /* Match conditional jumps - this may get tricky when fallthru and branch
1537 edges are crossed. */
1538 if (EDGE_COUNT (bb1->succs) == 2
1539 && any_condjump_p (BB_END (bb1))
1540 && onlyjump_p (BB_END (bb1)))
1541 {
1542 edge b1, f1, b2, f2;
1543 bool reverse, match;
1544 rtx set1, set2, cond1, cond2;
1545 enum rtx_code code1, code2;
1546
1547 if (EDGE_COUNT (bb2->succs) != 2
1548 || !any_condjump_p (BB_END (bb2))
1549 || !onlyjump_p (BB_END (bb2)))
1550 return false;
1551
1552 b1 = BRANCH_EDGE (bb1);
1553 b2 = BRANCH_EDGE (bb2);
1554 f1 = FALLTHRU_EDGE (bb1);
1555 f2 = FALLTHRU_EDGE (bb2);
1556
1557 /* Get around possible forwarders on fallthru edges. Other cases
1558 should be optimized out already. */
1559 if (FORWARDER_BLOCK_P (f1->dest))
1560 f1 = single_succ_edge (f1->dest);
1561
1562 if (FORWARDER_BLOCK_P (f2->dest))
1563 f2 = single_succ_edge (f2->dest);
1564
1565 /* To simplify use of this function, return false if there are
1566 unneeded forwarder blocks. These will get eliminated later
1567 during cleanup_cfg. */
1568 if (FORWARDER_BLOCK_P (f1->dest)
1569 || FORWARDER_BLOCK_P (f2->dest)
1570 || FORWARDER_BLOCK_P (b1->dest)
1571 || FORWARDER_BLOCK_P (b2->dest))
1572 return false;
1573
1574 if (f1->dest == f2->dest && b1->dest == b2->dest)
1575 reverse = false;
1576 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1577 reverse = true;
1578 else
1579 return false;
1580
1581 set1 = pc_set (BB_END (bb1));
1582 set2 = pc_set (BB_END (bb2));
1583 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1584 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1585 reverse = !reverse;
1586
1587 cond1 = XEXP (SET_SRC (set1), 0);
1588 cond2 = XEXP (SET_SRC (set2), 0);
1589 code1 = GET_CODE (cond1);
1590 if (reverse)
1591 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1592 else
1593 code2 = GET_CODE (cond2);
1594
1595 if (code2 == UNKNOWN)
1596 return false;
1597
1598 /* Verify codes and operands match. */
1599 match = ((code1 == code2
1600 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1601 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1602 || (code1 == swap_condition (code2)
1603 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1604 XEXP (cond2, 0))
1605 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1606 XEXP (cond2, 1))));
1607
1608 /* If we return true, we will join the blocks. Which means that
1609 we will only have one branch prediction bit to work with. Thus
1610 we require the existing branches to have probabilities that are
1611 roughly similar. */
1612 if (match
1613 && optimize_bb_for_speed_p (bb1)
1614 && optimize_bb_for_speed_p (bb2))
1615 {
1616 int prob2;
1617
1618 if (b1->dest == b2->dest)
1619 prob2 = b2->probability;
1620 else
1621 /* Do not use f2 probability as f2 may be forwarded. */
1622 prob2 = REG_BR_PROB_BASE - b2->probability;
1623
1624 /* Fail if the difference in probabilities is greater than 50%.
1625 This rules out two well-predicted branches with opposite
1626 outcomes. */
1627 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1628 {
1629 if (dump_file)
1630 fprintf (dump_file,
1631 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1632 bb1->index, bb2->index, b1->probability, prob2);
1633
1634 return false;
1635 }
1636 }
1637
1638 if (dump_file && match)
1639 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1640 bb1->index, bb2->index);
1641
1642 return match;
1643 }
1644
1645 /* Generic case - we are seeing a computed jump, table jump or trapping
1646 instruction. */
1647
1648 /* Check whether there are tablejumps in the end of BB1 and BB2.
1649 Return true if they are identical. */
1650 {
1651 rtx label1, label2;
1652 rtx table1, table2;
1653
1654 if (tablejump_p (BB_END (bb1), &label1, &table1)
1655 && tablejump_p (BB_END (bb2), &label2, &table2)
1656 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1657 {
1658 /* The labels should never be the same rtx. If they really are same
1659 the jump tables are same too. So disable crossjumping of blocks BB1
1660 and BB2 because when deleting the common insns in the end of BB1
1661 by delete_basic_block () the jump table would be deleted too. */
1662 /* If LABEL2 is referenced in BB1->END do not do anything
1663 because we would loose information when replacing
1664 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1665 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1666 {
1667 /* Set IDENTICAL to true when the tables are identical. */
1668 bool identical = false;
1669 rtx p1, p2;
1670
1671 p1 = PATTERN (table1);
1672 p2 = PATTERN (table2);
1673 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1674 {
1675 identical = true;
1676 }
1677 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1678 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1679 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1680 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1681 {
1682 int i;
1683
1684 identical = true;
1685 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1686 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1687 identical = false;
1688 }
1689
1690 if (identical)
1691 {
1692 replace_label_data rr;
1693 bool match;
1694
1695 /* Temporarily replace references to LABEL1 with LABEL2
1696 in BB1->END so that we could compare the instructions. */
1697 rr.r1 = label1;
1698 rr.r2 = label2;
1699 rr.update_label_nuses = false;
1700 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1701
1702 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1703 == dir_both);
1704 if (dump_file && match)
1705 fprintf (dump_file,
1706 "Tablejumps in bb %i and %i match.\n",
1707 bb1->index, bb2->index);
1708
1709 /* Set the original label in BB1->END because when deleting
1710 a block whose end is a tablejump, the tablejump referenced
1711 from the instruction is deleted too. */
1712 rr.r1 = label2;
1713 rr.r2 = label1;
1714 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1715
1716 return match;
1717 }
1718 }
1719 return false;
1720 }
1721 }
1722
1723 rtx last1 = BB_END (bb1);
1724 rtx last2 = BB_END (bb2);
1725 if (DEBUG_INSN_P (last1))
1726 last1 = prev_nondebug_insn (last1);
1727 if (DEBUG_INSN_P (last2))
1728 last2 = prev_nondebug_insn (last2);
1729 /* First ensure that the instructions match. There may be many outgoing
1730 edges so this test is generally cheaper. */
1731 if (old_insns_match_p (mode, last1, last2) != dir_both)
1732 return false;
1733
1734 /* Search the outgoing edges, ensure that the counts do match, find possible
1735 fallthru and exception handling edges since these needs more
1736 validation. */
1737 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1738 return false;
1739
1740 bool nonfakeedges = false;
1741 FOR_EACH_EDGE (e1, ei, bb1->succs)
1742 {
1743 e2 = EDGE_SUCC (bb2, ei.index);
1744
1745 if ((e1->flags & EDGE_FAKE) == 0)
1746 nonfakeedges = true;
1747
1748 if (e1->flags & EDGE_EH)
1749 nehedges1++;
1750
1751 if (e2->flags & EDGE_EH)
1752 nehedges2++;
1753
1754 if (e1->flags & EDGE_FALLTHRU)
1755 fallthru1 = e1;
1756 if (e2->flags & EDGE_FALLTHRU)
1757 fallthru2 = e2;
1758 }
1759
1760 /* If number of edges of various types does not match, fail. */
1761 if (nehedges1 != nehedges2
1762 || (fallthru1 != 0) != (fallthru2 != 0))
1763 return false;
1764
1765 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1766 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1767 attempt to optimize, as the two basic blocks might have different
1768 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1769 traps there should be REG_ARG_SIZE notes, they could be missing
1770 for __builtin_unreachable () uses though. */
1771 if (!nonfakeedges
1772 && !ACCUMULATE_OUTGOING_ARGS
1773 && (!INSN_P (last1)
1774 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1775 return false;
1776
1777 /* fallthru edges must be forwarded to the same destination. */
1778 if (fallthru1)
1779 {
1780 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1781 ? single_succ (fallthru1->dest): fallthru1->dest);
1782 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1783 ? single_succ (fallthru2->dest): fallthru2->dest);
1784
1785 if (d1 != d2)
1786 return false;
1787 }
1788
1789 /* Ensure the same EH region. */
1790 {
1791 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1792 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1793
1794 if (!n1 && n2)
1795 return false;
1796
1797 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1798 return false;
1799 }
1800
1801 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1802 version of sequence abstraction. */
1803 FOR_EACH_EDGE (e1, ei, bb2->succs)
1804 {
1805 edge e2;
1806 edge_iterator ei;
1807 basic_block d1 = e1->dest;
1808
1809 if (FORWARDER_BLOCK_P (d1))
1810 d1 = EDGE_SUCC (d1, 0)->dest;
1811
1812 FOR_EACH_EDGE (e2, ei, bb1->succs)
1813 {
1814 basic_block d2 = e2->dest;
1815 if (FORWARDER_BLOCK_P (d2))
1816 d2 = EDGE_SUCC (d2, 0)->dest;
1817 if (d1 == d2)
1818 break;
1819 }
1820
1821 if (!e2)
1822 return false;
1823 }
1824
1825 return true;
1826 }
1827
1828 /* Returns true if BB basic block has a preserve label. */
1829
1830 static bool
1831 block_has_preserve_label (basic_block bb)
1832 {
1833 return (bb
1834 && block_label (bb)
1835 && LABEL_PRESERVE_P (block_label (bb)));
1836 }
1837
1838 /* E1 and E2 are edges with the same destination block. Search their
1839 predecessors for common code. If found, redirect control flow from
1840 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1841 or the other way around (dir_backward). DIR specifies the allowed
1842 replacement direction. */
1843
1844 static bool
1845 try_crossjump_to_edge (int mode, edge e1, edge e2,
1846 enum replace_direction dir)
1847 {
1848 int nmatch;
1849 basic_block src1 = e1->src, src2 = e2->src;
1850 basic_block redirect_to, redirect_from, to_remove;
1851 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1852 rtx newpos1, newpos2;
1853 edge s;
1854 edge_iterator ei;
1855
1856 newpos1 = newpos2 = NULL_RTX;
1857
1858 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1859 to try this optimization.
1860
1861 Basic block partitioning may result in some jumps that appear to
1862 be optimizable (or blocks that appear to be mergeable), but which really
1863 must be left untouched (they are required to make it safely across
1864 partition boundaries). See the comments at the top of
1865 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1866
1867 if (crtl->has_bb_partition && reload_completed)
1868 return false;
1869
1870 /* Search backward through forwarder blocks. We don't need to worry
1871 about multiple entry or chained forwarders, as they will be optimized
1872 away. We do this to look past the unconditional jump following a
1873 conditional jump that is required due to the current CFG shape. */
1874 if (single_pred_p (src1)
1875 && FORWARDER_BLOCK_P (src1))
1876 e1 = single_pred_edge (src1), src1 = e1->src;
1877
1878 if (single_pred_p (src2)
1879 && FORWARDER_BLOCK_P (src2))
1880 e2 = single_pred_edge (src2), src2 = e2->src;
1881
1882 /* Nothing to do if we reach ENTRY, or a common source block. */
1883 if (src1 == ENTRY_BLOCK_PTR || src2 == ENTRY_BLOCK_PTR)
1884 return false;
1885 if (src1 == src2)
1886 return false;
1887
1888 /* Seeing more than 1 forwarder blocks would confuse us later... */
1889 if (FORWARDER_BLOCK_P (e1->dest)
1890 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1891 return false;
1892
1893 if (FORWARDER_BLOCK_P (e2->dest)
1894 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1895 return false;
1896
1897 /* Likewise with dead code (possibly newly created by the other optimizations
1898 of cfg_cleanup). */
1899 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1900 return false;
1901
1902 /* Look for the common insn sequence, part the first ... */
1903 if (!outgoing_edges_match (mode, src1, src2))
1904 return false;
1905
1906 /* ... and part the second. */
1907 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1908
1909 osrc1 = src1;
1910 osrc2 = src2;
1911 if (newpos1 != NULL_RTX)
1912 src1 = BLOCK_FOR_INSN (newpos1);
1913 if (newpos2 != NULL_RTX)
1914 src2 = BLOCK_FOR_INSN (newpos2);
1915
1916 if (dir == dir_backward)
1917 {
1918 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1919 SWAP (basic_block, osrc1, osrc2);
1920 SWAP (basic_block, src1, src2);
1921 SWAP (edge, e1, e2);
1922 SWAP (rtx, newpos1, newpos2);
1923 #undef SWAP
1924 }
1925
1926 /* Don't proceed with the crossjump unless we found a sufficient number
1927 of matching instructions or the 'from' block was totally matched
1928 (such that its predecessors will hopefully be redirected and the
1929 block removed). */
1930 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
1931 && (newpos1 != BB_HEAD (src1)))
1932 return false;
1933
1934 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1935 if (block_has_preserve_label (e1->dest)
1936 && (e1->flags & EDGE_ABNORMAL))
1937 return false;
1938
1939 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1940 will be deleted.
1941 If we have tablejumps in the end of SRC1 and SRC2
1942 they have been already compared for equivalence in outgoing_edges_match ()
1943 so replace the references to TABLE1 by references to TABLE2. */
1944 {
1945 rtx label1, label2;
1946 rtx table1, table2;
1947
1948 if (tablejump_p (BB_END (osrc1), &label1, &table1)
1949 && tablejump_p (BB_END (osrc2), &label2, &table2)
1950 && label1 != label2)
1951 {
1952 replace_label_data rr;
1953 rtx insn;
1954
1955 /* Replace references to LABEL1 with LABEL2. */
1956 rr.r1 = label1;
1957 rr.r2 = label2;
1958 rr.update_label_nuses = true;
1959 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
1960 {
1961 /* Do not replace the label in SRC1->END because when deleting
1962 a block whose end is a tablejump, the tablejump referenced
1963 from the instruction is deleted too. */
1964 if (insn != BB_END (osrc1))
1965 for_each_rtx (&insn, replace_label, &rr);
1966 }
1967 }
1968 }
1969
1970 /* Avoid splitting if possible. We must always split when SRC2 has
1971 EH predecessor edges, or we may end up with basic blocks with both
1972 normal and EH predecessor edges. */
1973 if (newpos2 == BB_HEAD (src2)
1974 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
1975 redirect_to = src2;
1976 else
1977 {
1978 if (newpos2 == BB_HEAD (src2))
1979 {
1980 /* Skip possible basic block header. */
1981 if (LABEL_P (newpos2))
1982 newpos2 = NEXT_INSN (newpos2);
1983 while (DEBUG_INSN_P (newpos2))
1984 newpos2 = NEXT_INSN (newpos2);
1985 if (NOTE_P (newpos2))
1986 newpos2 = NEXT_INSN (newpos2);
1987 while (DEBUG_INSN_P (newpos2))
1988 newpos2 = NEXT_INSN (newpos2);
1989 }
1990
1991 if (dump_file)
1992 fprintf (dump_file, "Splitting bb %i before %i insns\n",
1993 src2->index, nmatch);
1994 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
1995 }
1996
1997 if (dump_file)
1998 fprintf (dump_file,
1999 "Cross jumping from bb %i to bb %i; %i common insns\n",
2000 src1->index, src2->index, nmatch);
2001
2002 /* We may have some registers visible through the block. */
2003 df_set_bb_dirty (redirect_to);
2004
2005 if (osrc2 == src2)
2006 redirect_edges_to = redirect_to;
2007 else
2008 redirect_edges_to = osrc2;
2009
2010 /* Recompute the frequencies and counts of outgoing edges. */
2011 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2012 {
2013 edge s2;
2014 edge_iterator ei;
2015 basic_block d = s->dest;
2016
2017 if (FORWARDER_BLOCK_P (d))
2018 d = single_succ (d);
2019
2020 FOR_EACH_EDGE (s2, ei, src1->succs)
2021 {
2022 basic_block d2 = s2->dest;
2023 if (FORWARDER_BLOCK_P (d2))
2024 d2 = single_succ (d2);
2025 if (d == d2)
2026 break;
2027 }
2028
2029 s->count += s2->count;
2030
2031 /* Take care to update possible forwarder blocks. We verified
2032 that there is no more than one in the chain, so we can't run
2033 into infinite loop. */
2034 if (FORWARDER_BLOCK_P (s->dest))
2035 {
2036 single_succ_edge (s->dest)->count += s2->count;
2037 s->dest->count += s2->count;
2038 s->dest->frequency += EDGE_FREQUENCY (s);
2039 }
2040
2041 if (FORWARDER_BLOCK_P (s2->dest))
2042 {
2043 single_succ_edge (s2->dest)->count -= s2->count;
2044 if (single_succ_edge (s2->dest)->count < 0)
2045 single_succ_edge (s2->dest)->count = 0;
2046 s2->dest->count -= s2->count;
2047 s2->dest->frequency -= EDGE_FREQUENCY (s);
2048 if (s2->dest->frequency < 0)
2049 s2->dest->frequency = 0;
2050 if (s2->dest->count < 0)
2051 s2->dest->count = 0;
2052 }
2053
2054 if (!redirect_edges_to->frequency && !src1->frequency)
2055 s->probability = (s->probability + s2->probability) / 2;
2056 else
2057 s->probability
2058 = ((s->probability * redirect_edges_to->frequency +
2059 s2->probability * src1->frequency)
2060 / (redirect_edges_to->frequency + src1->frequency));
2061 }
2062
2063 /* Adjust count and frequency for the block. An earlier jump
2064 threading pass may have left the profile in an inconsistent
2065 state (see update_bb_profile_for_threading) so we must be
2066 prepared for overflows. */
2067 tmp = redirect_to;
2068 do
2069 {
2070 tmp->count += src1->count;
2071 tmp->frequency += src1->frequency;
2072 if (tmp->frequency > BB_FREQ_MAX)
2073 tmp->frequency = BB_FREQ_MAX;
2074 if (tmp == redirect_edges_to)
2075 break;
2076 tmp = find_fallthru_edge (tmp->succs)->dest;
2077 }
2078 while (true);
2079 update_br_prob_note (redirect_edges_to);
2080
2081 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2082
2083 /* Skip possible basic block header. */
2084 if (LABEL_P (newpos1))
2085 newpos1 = NEXT_INSN (newpos1);
2086
2087 while (DEBUG_INSN_P (newpos1))
2088 newpos1 = NEXT_INSN (newpos1);
2089
2090 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2091 newpos1 = NEXT_INSN (newpos1);
2092
2093 while (DEBUG_INSN_P (newpos1))
2094 newpos1 = NEXT_INSN (newpos1);
2095
2096 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2097 to_remove = single_succ (redirect_from);
2098
2099 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2100 delete_basic_block (to_remove);
2101
2102 update_forwarder_flag (redirect_from);
2103 if (redirect_to != src2)
2104 update_forwarder_flag (src2);
2105
2106 return true;
2107 }
2108
2109 /* Search the predecessors of BB for common insn sequences. When found,
2110 share code between them by redirecting control flow. Return true if
2111 any changes made. */
2112
2113 static bool
2114 try_crossjump_bb (int mode, basic_block bb)
2115 {
2116 edge e, e2, fallthru;
2117 bool changed;
2118 unsigned max, ix, ix2;
2119
2120 /* Nothing to do if there is not at least two incoming edges. */
2121 if (EDGE_COUNT (bb->preds) < 2)
2122 return false;
2123
2124 /* Don't crossjump if this block ends in a computed jump,
2125 unless we are optimizing for size. */
2126 if (optimize_bb_for_size_p (bb)
2127 && bb != EXIT_BLOCK_PTR
2128 && computed_jump_p (BB_END (bb)))
2129 return false;
2130
2131 /* If we are partitioning hot/cold basic blocks, we don't want to
2132 mess up unconditional or indirect jumps that cross between hot
2133 and cold sections.
2134
2135 Basic block partitioning may result in some jumps that appear to
2136 be optimizable (or blocks that appear to be mergeable), but which really
2137 must be left untouched (they are required to make it safely across
2138 partition boundaries). See the comments at the top of
2139 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2140
2141 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2142 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2143 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2144 return false;
2145
2146 /* It is always cheapest to redirect a block that ends in a branch to
2147 a block that falls through into BB, as that adds no branches to the
2148 program. We'll try that combination first. */
2149 fallthru = NULL;
2150 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2151
2152 if (EDGE_COUNT (bb->preds) > max)
2153 return false;
2154
2155 fallthru = find_fallthru_edge (bb->preds);
2156
2157 changed = false;
2158 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2159 {
2160 e = EDGE_PRED (bb, ix);
2161 ix++;
2162
2163 /* As noted above, first try with the fallthru predecessor (or, a
2164 fallthru predecessor if we are in cfglayout mode). */
2165 if (fallthru)
2166 {
2167 /* Don't combine the fallthru edge into anything else.
2168 If there is a match, we'll do it the other way around. */
2169 if (e == fallthru)
2170 continue;
2171 /* If nothing changed since the last attempt, there is nothing
2172 we can do. */
2173 if (!first_pass
2174 && !((e->src->flags & BB_MODIFIED)
2175 || (fallthru->src->flags & BB_MODIFIED)))
2176 continue;
2177
2178 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2179 {
2180 changed = true;
2181 ix = 0;
2182 continue;
2183 }
2184 }
2185
2186 /* Non-obvious work limiting check: Recognize that we're going
2187 to call try_crossjump_bb on every basic block. So if we have
2188 two blocks with lots of outgoing edges (a switch) and they
2189 share lots of common destinations, then we would do the
2190 cross-jump check once for each common destination.
2191
2192 Now, if the blocks actually are cross-jump candidates, then
2193 all of their destinations will be shared. Which means that
2194 we only need check them for cross-jump candidacy once. We
2195 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2196 choosing to do the check from the block for which the edge
2197 in question is the first successor of A. */
2198 if (EDGE_SUCC (e->src, 0) != e)
2199 continue;
2200
2201 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2202 {
2203 e2 = EDGE_PRED (bb, ix2);
2204
2205 if (e2 == e)
2206 continue;
2207
2208 /* We've already checked the fallthru edge above. */
2209 if (e2 == fallthru)
2210 continue;
2211
2212 /* The "first successor" check above only prevents multiple
2213 checks of crossjump(A,B). In order to prevent redundant
2214 checks of crossjump(B,A), require that A be the block
2215 with the lowest index. */
2216 if (e->src->index > e2->src->index)
2217 continue;
2218
2219 /* If nothing changed since the last attempt, there is nothing
2220 we can do. */
2221 if (!first_pass
2222 && !((e->src->flags & BB_MODIFIED)
2223 || (e2->src->flags & BB_MODIFIED)))
2224 continue;
2225
2226 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2227 direction. */
2228 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2229 {
2230 changed = true;
2231 ix = 0;
2232 break;
2233 }
2234 }
2235 }
2236
2237 if (changed)
2238 crossjumps_occured = true;
2239
2240 return changed;
2241 }
2242
2243 /* Search the successors of BB for common insn sequences. When found,
2244 share code between them by moving it across the basic block
2245 boundary. Return true if any changes made. */
2246
2247 static bool
2248 try_head_merge_bb (basic_block bb)
2249 {
2250 basic_block final_dest_bb = NULL;
2251 int max_match = INT_MAX;
2252 edge e0;
2253 rtx *headptr, *currptr, *nextptr;
2254 bool changed, moveall;
2255 unsigned ix;
2256 rtx e0_last_head, cond, move_before;
2257 unsigned nedges = EDGE_COUNT (bb->succs);
2258 rtx jump = BB_END (bb);
2259 regset live, live_union;
2260
2261 /* Nothing to do if there is not at least two outgoing edges. */
2262 if (nedges < 2)
2263 return false;
2264
2265 /* Don't crossjump if this block ends in a computed jump,
2266 unless we are optimizing for size. */
2267 if (optimize_bb_for_size_p (bb)
2268 && bb != EXIT_BLOCK_PTR
2269 && computed_jump_p (BB_END (bb)))
2270 return false;
2271
2272 cond = get_condition (jump, &move_before, true, false);
2273 if (cond == NULL_RTX)
2274 {
2275 #ifdef HAVE_cc0
2276 if (reg_mentioned_p (cc0_rtx, jump))
2277 move_before = prev_nonnote_nondebug_insn (jump);
2278 else
2279 #endif
2280 move_before = jump;
2281 }
2282
2283 for (ix = 0; ix < nedges; ix++)
2284 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR)
2285 return false;
2286
2287 for (ix = 0; ix < nedges; ix++)
2288 {
2289 edge e = EDGE_SUCC (bb, ix);
2290 basic_block other_bb = e->dest;
2291
2292 if (df_get_bb_dirty (other_bb))
2293 {
2294 block_was_dirty = true;
2295 return false;
2296 }
2297
2298 if (e->flags & EDGE_ABNORMAL)
2299 return false;
2300
2301 /* Normally, all destination blocks must only be reachable from this
2302 block, i.e. they must have one incoming edge.
2303
2304 There is one special case we can handle, that of multiple consecutive
2305 jumps where the first jumps to one of the targets of the second jump.
2306 This happens frequently in switch statements for default labels.
2307 The structure is as follows:
2308 FINAL_DEST_BB
2309 ....
2310 if (cond) jump A;
2311 fall through
2312 BB
2313 jump with targets A, B, C, D...
2314 A
2315 has two incoming edges, from FINAL_DEST_BB and BB
2316
2317 In this case, we can try to move the insns through BB and into
2318 FINAL_DEST_BB. */
2319 if (EDGE_COUNT (other_bb->preds) != 1)
2320 {
2321 edge incoming_edge, incoming_bb_other_edge;
2322 edge_iterator ei;
2323
2324 if (final_dest_bb != NULL
2325 || EDGE_COUNT (other_bb->preds) != 2)
2326 return false;
2327
2328 /* We must be able to move the insns across the whole block. */
2329 move_before = BB_HEAD (bb);
2330 while (!NONDEBUG_INSN_P (move_before))
2331 move_before = NEXT_INSN (move_before);
2332
2333 if (EDGE_COUNT (bb->preds) != 1)
2334 return false;
2335 incoming_edge = EDGE_PRED (bb, 0);
2336 final_dest_bb = incoming_edge->src;
2337 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2338 return false;
2339 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2340 if (incoming_bb_other_edge != incoming_edge)
2341 break;
2342 if (incoming_bb_other_edge->dest != other_bb)
2343 return false;
2344 }
2345 }
2346
2347 e0 = EDGE_SUCC (bb, 0);
2348 e0_last_head = NULL_RTX;
2349 changed = false;
2350
2351 for (ix = 1; ix < nedges; ix++)
2352 {
2353 edge e = EDGE_SUCC (bb, ix);
2354 rtx e0_last, e_last;
2355 int nmatch;
2356
2357 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2358 &e0_last, &e_last, 0);
2359 if (nmatch == 0)
2360 return false;
2361
2362 if (nmatch < max_match)
2363 {
2364 max_match = nmatch;
2365 e0_last_head = e0_last;
2366 }
2367 }
2368
2369 /* If we matched an entire block, we probably have to avoid moving the
2370 last insn. */
2371 if (max_match > 0
2372 && e0_last_head == BB_END (e0->dest)
2373 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2374 || control_flow_insn_p (e0_last_head)))
2375 {
2376 max_match--;
2377 if (max_match == 0)
2378 return false;
2379 do
2380 e0_last_head = prev_real_insn (e0_last_head);
2381 while (DEBUG_INSN_P (e0_last_head));
2382 }
2383
2384 if (max_match == 0)
2385 return false;
2386
2387 /* We must find a union of the live registers at each of the end points. */
2388 live = BITMAP_ALLOC (NULL);
2389 live_union = BITMAP_ALLOC (NULL);
2390
2391 currptr = XNEWVEC (rtx, nedges);
2392 headptr = XNEWVEC (rtx, nedges);
2393 nextptr = XNEWVEC (rtx, nedges);
2394
2395 for (ix = 0; ix < nedges; ix++)
2396 {
2397 int j;
2398 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2399 rtx head = BB_HEAD (merge_bb);
2400
2401 while (!NONDEBUG_INSN_P (head))
2402 head = NEXT_INSN (head);
2403 headptr[ix] = head;
2404 currptr[ix] = head;
2405
2406 /* Compute the end point and live information */
2407 for (j = 1; j < max_match; j++)
2408 do
2409 head = NEXT_INSN (head);
2410 while (!NONDEBUG_INSN_P (head));
2411 simulate_backwards_to_point (merge_bb, live, head);
2412 IOR_REG_SET (live_union, live);
2413 }
2414
2415 /* If we're moving across two blocks, verify the validity of the
2416 first move, then adjust the target and let the loop below deal
2417 with the final move. */
2418 if (final_dest_bb != NULL)
2419 {
2420 rtx move_upto;
2421
2422 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2423 jump, e0->dest, live_union,
2424 NULL, &move_upto);
2425 if (!moveall)
2426 {
2427 if (move_upto == NULL_RTX)
2428 goto out;
2429
2430 while (e0_last_head != move_upto)
2431 {
2432 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2433 live_union);
2434 e0_last_head = PREV_INSN (e0_last_head);
2435 }
2436 }
2437 if (e0_last_head == NULL_RTX)
2438 goto out;
2439
2440 jump = BB_END (final_dest_bb);
2441 cond = get_condition (jump, &move_before, true, false);
2442 if (cond == NULL_RTX)
2443 {
2444 #ifdef HAVE_cc0
2445 if (reg_mentioned_p (cc0_rtx, jump))
2446 move_before = prev_nonnote_nondebug_insn (jump);
2447 else
2448 #endif
2449 move_before = jump;
2450 }
2451 }
2452
2453 do
2454 {
2455 rtx move_upto;
2456 moveall = can_move_insns_across (currptr[0], e0_last_head,
2457 move_before, jump, e0->dest, live_union,
2458 NULL, &move_upto);
2459 if (!moveall && move_upto == NULL_RTX)
2460 {
2461 if (jump == move_before)
2462 break;
2463
2464 /* Try again, using a different insertion point. */
2465 move_before = jump;
2466
2467 #ifdef HAVE_cc0
2468 /* Don't try moving before a cc0 user, as that may invalidate
2469 the cc0. */
2470 if (reg_mentioned_p (cc0_rtx, jump))
2471 break;
2472 #endif
2473
2474 continue;
2475 }
2476
2477 if (final_dest_bb && !moveall)
2478 /* We haven't checked whether a partial move would be OK for the first
2479 move, so we have to fail this case. */
2480 break;
2481
2482 changed = true;
2483 for (;;)
2484 {
2485 if (currptr[0] == move_upto)
2486 break;
2487 for (ix = 0; ix < nedges; ix++)
2488 {
2489 rtx curr = currptr[ix];
2490 do
2491 curr = NEXT_INSN (curr);
2492 while (!NONDEBUG_INSN_P (curr));
2493 currptr[ix] = curr;
2494 }
2495 }
2496
2497 /* If we can't currently move all of the identical insns, remember
2498 each insn after the range that we'll merge. */
2499 if (!moveall)
2500 for (ix = 0; ix < nedges; ix++)
2501 {
2502 rtx curr = currptr[ix];
2503 do
2504 curr = NEXT_INSN (curr);
2505 while (!NONDEBUG_INSN_P (curr));
2506 nextptr[ix] = curr;
2507 }
2508
2509 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2510 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2511 if (final_dest_bb != NULL)
2512 df_set_bb_dirty (final_dest_bb);
2513 df_set_bb_dirty (bb);
2514 for (ix = 1; ix < nedges; ix++)
2515 {
2516 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2517 delete_insn_chain (headptr[ix], currptr[ix], false);
2518 }
2519 if (!moveall)
2520 {
2521 if (jump == move_before)
2522 break;
2523
2524 /* For the unmerged insns, try a different insertion point. */
2525 move_before = jump;
2526
2527 #ifdef HAVE_cc0
2528 /* Don't try moving before a cc0 user, as that may invalidate
2529 the cc0. */
2530 if (reg_mentioned_p (cc0_rtx, jump))
2531 break;
2532 #endif
2533
2534 for (ix = 0; ix < nedges; ix++)
2535 currptr[ix] = headptr[ix] = nextptr[ix];
2536 }
2537 }
2538 while (!moveall);
2539
2540 out:
2541 free (currptr);
2542 free (headptr);
2543 free (nextptr);
2544
2545 crossjumps_occured |= changed;
2546
2547 return changed;
2548 }
2549
2550 /* Return true if BB contains just bb note, or bb note followed
2551 by only DEBUG_INSNs. */
2552
2553 static bool
2554 trivially_empty_bb_p (basic_block bb)
2555 {
2556 rtx insn = BB_END (bb);
2557
2558 while (1)
2559 {
2560 if (insn == BB_HEAD (bb))
2561 return true;
2562 if (!DEBUG_INSN_P (insn))
2563 return false;
2564 insn = PREV_INSN (insn);
2565 }
2566 }
2567
2568 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2569 instructions etc. Return nonzero if changes were made. */
2570
2571 static bool
2572 try_optimize_cfg (int mode)
2573 {
2574 bool changed_overall = false;
2575 bool changed;
2576 int iterations = 0;
2577 basic_block bb, b, next;
2578
2579 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2580 clear_bb_flags ();
2581
2582 crossjumps_occured = false;
2583
2584 FOR_EACH_BB (bb)
2585 update_forwarder_flag (bb);
2586
2587 if (! targetm.cannot_modify_jumps_p ())
2588 {
2589 first_pass = true;
2590 /* Attempt to merge blocks as made possible by edge removal. If
2591 a block has only one successor, and the successor has only
2592 one predecessor, they may be combined. */
2593 do
2594 {
2595 block_was_dirty = false;
2596 changed = false;
2597 iterations++;
2598
2599 if (dump_file)
2600 fprintf (dump_file,
2601 "\n\ntry_optimize_cfg iteration %i\n\n",
2602 iterations);
2603
2604 for (b = ENTRY_BLOCK_PTR->next_bb; b != EXIT_BLOCK_PTR;)
2605 {
2606 basic_block c;
2607 edge s;
2608 bool changed_here = false;
2609
2610 /* Delete trivially dead basic blocks. This is either
2611 blocks with no predecessors, or empty blocks with no
2612 successors. However if the empty block with no
2613 successors is the successor of the ENTRY_BLOCK, it is
2614 kept. This ensures that the ENTRY_BLOCK will have a
2615 successor which is a precondition for many RTL
2616 passes. Empty blocks may result from expanding
2617 __builtin_unreachable (). */
2618 if (EDGE_COUNT (b->preds) == 0
2619 || (EDGE_COUNT (b->succs) == 0
2620 && trivially_empty_bb_p (b)
2621 && single_succ_edge (ENTRY_BLOCK_PTR)->dest != b))
2622 {
2623 c = b->prev_bb;
2624 if (EDGE_COUNT (b->preds) > 0)
2625 {
2626 edge e;
2627 edge_iterator ei;
2628
2629 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2630 {
2631 if (BB_FOOTER (b)
2632 && BARRIER_P (BB_FOOTER (b)))
2633 FOR_EACH_EDGE (e, ei, b->preds)
2634 if ((e->flags & EDGE_FALLTHRU)
2635 && BB_FOOTER (e->src) == NULL)
2636 {
2637 if (BB_FOOTER (b))
2638 {
2639 BB_FOOTER (e->src) = BB_FOOTER (b);
2640 BB_FOOTER (b) = NULL;
2641 }
2642 else
2643 {
2644 start_sequence ();
2645 BB_FOOTER (e->src) = emit_barrier ();
2646 end_sequence ();
2647 }
2648 }
2649 }
2650 else
2651 {
2652 rtx last = get_last_bb_insn (b);
2653 if (last && BARRIER_P (last))
2654 FOR_EACH_EDGE (e, ei, b->preds)
2655 if ((e->flags & EDGE_FALLTHRU))
2656 emit_barrier_after (BB_END (e->src));
2657 }
2658 }
2659 delete_basic_block (b);
2660 changed = true;
2661 /* Avoid trying to remove ENTRY_BLOCK_PTR. */
2662 b = (c == ENTRY_BLOCK_PTR ? c->next_bb : c);
2663 continue;
2664 }
2665
2666 /* Remove code labels no longer used. */
2667 if (single_pred_p (b)
2668 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2669 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2670 && LABEL_P (BB_HEAD (b))
2671 /* If the previous block ends with a branch to this
2672 block, we can't delete the label. Normally this
2673 is a condjump that is yet to be simplified, but
2674 if CASE_DROPS_THRU, this can be a tablejump with
2675 some element going to the same place as the
2676 default (fallthru). */
2677 && (single_pred (b) == ENTRY_BLOCK_PTR
2678 || !JUMP_P (BB_END (single_pred (b)))
2679 || ! label_is_jump_target_p (BB_HEAD (b),
2680 BB_END (single_pred (b)))))
2681 {
2682 delete_insn (BB_HEAD (b));
2683 if (dump_file)
2684 fprintf (dump_file, "Deleted label in block %i.\n",
2685 b->index);
2686 }
2687
2688 /* If we fall through an empty block, we can remove it. */
2689 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2690 && single_pred_p (b)
2691 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2692 && !LABEL_P (BB_HEAD (b))
2693 && FORWARDER_BLOCK_P (b)
2694 /* Note that forwarder_block_p true ensures that
2695 there is a successor for this block. */
2696 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2697 && n_basic_blocks > NUM_FIXED_BLOCKS + 1)
2698 {
2699 if (dump_file)
2700 fprintf (dump_file,
2701 "Deleting fallthru block %i.\n",
2702 b->index);
2703
2704 c = b->prev_bb == ENTRY_BLOCK_PTR ? b->next_bb : b->prev_bb;
2705 redirect_edge_succ_nodup (single_pred_edge (b),
2706 single_succ (b));
2707 delete_basic_block (b);
2708 changed = true;
2709 b = c;
2710 continue;
2711 }
2712
2713 /* Merge B with its single successor, if any. */
2714 if (single_succ_p (b)
2715 && (s = single_succ_edge (b))
2716 && !(s->flags & EDGE_COMPLEX)
2717 && (c = s->dest) != EXIT_BLOCK_PTR
2718 && single_pred_p (c)
2719 && b != c)
2720 {
2721 /* When not in cfg_layout mode use code aware of reordering
2722 INSN. This code possibly creates new basic blocks so it
2723 does not fit merge_blocks interface and is kept here in
2724 hope that it will become useless once more of compiler
2725 is transformed to use cfg_layout mode. */
2726
2727 if ((mode & CLEANUP_CFGLAYOUT)
2728 && can_merge_blocks_p (b, c))
2729 {
2730 merge_blocks (b, c);
2731 update_forwarder_flag (b);
2732 changed_here = true;
2733 }
2734 else if (!(mode & CLEANUP_CFGLAYOUT)
2735 /* If the jump insn has side effects,
2736 we can't kill the edge. */
2737 && (!JUMP_P (BB_END (b))
2738 || (reload_completed
2739 ? simplejump_p (BB_END (b))
2740 : (onlyjump_p (BB_END (b))
2741 && !tablejump_p (BB_END (b),
2742 NULL, NULL))))
2743 && (next = merge_blocks_move (s, b, c, mode)))
2744 {
2745 b = next;
2746 changed_here = true;
2747 }
2748 }
2749
2750 /* Simplify branch over branch. */
2751 if ((mode & CLEANUP_EXPENSIVE)
2752 && !(mode & CLEANUP_CFGLAYOUT)
2753 && try_simplify_condjump (b))
2754 changed_here = true;
2755
2756 /* If B has a single outgoing edge, but uses a
2757 non-trivial jump instruction without side-effects, we
2758 can either delete the jump entirely, or replace it
2759 with a simple unconditional jump. */
2760 if (single_succ_p (b)
2761 && single_succ (b) != EXIT_BLOCK_PTR
2762 && onlyjump_p (BB_END (b))
2763 && !find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX)
2764 && try_redirect_by_replacing_jump (single_succ_edge (b),
2765 single_succ (b),
2766 (mode & CLEANUP_CFGLAYOUT) != 0))
2767 {
2768 update_forwarder_flag (b);
2769 changed_here = true;
2770 }
2771
2772 /* Simplify branch to branch. */
2773 if (try_forward_edges (mode, b))
2774 {
2775 update_forwarder_flag (b);
2776 changed_here = true;
2777 }
2778
2779 /* Look for shared code between blocks. */
2780 if ((mode & CLEANUP_CROSSJUMP)
2781 && try_crossjump_bb (mode, b))
2782 changed_here = true;
2783
2784 if ((mode & CLEANUP_CROSSJUMP)
2785 /* This can lengthen register lifetimes. Do it only after
2786 reload. */
2787 && reload_completed
2788 && try_head_merge_bb (b))
2789 changed_here = true;
2790
2791 /* Don't get confused by the index shift caused by
2792 deleting blocks. */
2793 if (!changed_here)
2794 b = b->next_bb;
2795 else
2796 changed = true;
2797 }
2798
2799 if ((mode & CLEANUP_CROSSJUMP)
2800 && try_crossjump_bb (mode, EXIT_BLOCK_PTR))
2801 changed = true;
2802
2803 if (block_was_dirty)
2804 {
2805 /* This should only be set by head-merging. */
2806 gcc_assert (mode & CLEANUP_CROSSJUMP);
2807 df_analyze ();
2808 }
2809
2810 #ifdef ENABLE_CHECKING
2811 if (changed)
2812 verify_flow_info ();
2813 #endif
2814
2815 changed_overall |= changed;
2816 first_pass = false;
2817 }
2818 while (changed);
2819 }
2820
2821 FOR_ALL_BB (b)
2822 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2823
2824 return changed_overall;
2825 }
2826 \f
2827 /* Delete all unreachable basic blocks. */
2828
2829 bool
2830 delete_unreachable_blocks (void)
2831 {
2832 bool changed = false;
2833 basic_block b, prev_bb;
2834
2835 find_unreachable_blocks ();
2836
2837 /* When we're in GIMPLE mode and there may be debug insns, we should
2838 delete blocks in reverse dominator order, so as to get a chance
2839 to substitute all released DEFs into debug stmts. If we don't
2840 have dominators information, walking blocks backward gets us a
2841 better chance of retaining most debug information than
2842 otherwise. */
2843 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2844 && dom_info_available_p (CDI_DOMINATORS))
2845 {
2846 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2847 {
2848 prev_bb = b->prev_bb;
2849
2850 if (!(b->flags & BB_REACHABLE))
2851 {
2852 /* Speed up the removal of blocks that don't dominate
2853 others. Walking backwards, this should be the common
2854 case. */
2855 if (!first_dom_son (CDI_DOMINATORS, b))
2856 delete_basic_block (b);
2857 else
2858 {
2859 vec<basic_block> h
2860 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2861
2862 while (h.length ())
2863 {
2864 b = h.pop ();
2865
2866 prev_bb = b->prev_bb;
2867
2868 gcc_assert (!(b->flags & BB_REACHABLE));
2869
2870 delete_basic_block (b);
2871 }
2872
2873 h.release ();
2874 }
2875
2876 changed = true;
2877 }
2878 }
2879 }
2880 else
2881 {
2882 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2883 {
2884 prev_bb = b->prev_bb;
2885
2886 if (!(b->flags & BB_REACHABLE))
2887 {
2888 delete_basic_block (b);
2889 changed = true;
2890 }
2891 }
2892 }
2893
2894 if (changed)
2895 tidy_fallthru_edges ();
2896 return changed;
2897 }
2898
2899 /* Delete any jump tables never referenced. We can't delete them at the
2900 time of removing tablejump insn as they are referenced by the preceding
2901 insns computing the destination, so we delay deleting and garbagecollect
2902 them once life information is computed. */
2903 void
2904 delete_dead_jumptables (void)
2905 {
2906 basic_block bb;
2907
2908 /* A dead jump table does not belong to any basic block. Scan insns
2909 between two adjacent basic blocks. */
2910 FOR_EACH_BB (bb)
2911 {
2912 rtx insn, next;
2913
2914 for (insn = NEXT_INSN (BB_END (bb));
2915 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
2916 insn = next)
2917 {
2918 next = NEXT_INSN (insn);
2919 if (LABEL_P (insn)
2920 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
2921 && JUMP_TABLE_DATA_P (next))
2922 {
2923 rtx label = insn, jump = next;
2924
2925 if (dump_file)
2926 fprintf (dump_file, "Dead jumptable %i removed\n",
2927 INSN_UID (insn));
2928
2929 next = NEXT_INSN (next);
2930 delete_insn (jump);
2931 delete_insn (label);
2932 }
2933 }
2934 }
2935 }
2936
2937 \f
2938 /* Tidy the CFG by deleting unreachable code and whatnot. */
2939
2940 bool
2941 cleanup_cfg (int mode)
2942 {
2943 bool changed = false;
2944
2945 /* Set the cfglayout mode flag here. We could update all the callers
2946 but that is just inconvenient, especially given that we eventually
2947 want to have cfglayout mode as the default. */
2948 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2949 mode |= CLEANUP_CFGLAYOUT;
2950
2951 timevar_push (TV_CLEANUP_CFG);
2952 if (delete_unreachable_blocks ())
2953 {
2954 changed = true;
2955 /* We've possibly created trivially dead code. Cleanup it right
2956 now to introduce more opportunities for try_optimize_cfg. */
2957 if (!(mode & (CLEANUP_NO_INSN_DEL))
2958 && !reload_completed)
2959 delete_trivially_dead_insns (get_insns (), max_reg_num ());
2960 }
2961
2962 compact_blocks ();
2963
2964 /* To tail-merge blocks ending in the same noreturn function (e.g.
2965 a call to abort) we have to insert fake edges to exit. Do this
2966 here once. The fake edges do not interfere with any other CFG
2967 cleanups. */
2968 if (mode & CLEANUP_CROSSJUMP)
2969 add_noreturn_fake_exit_edges ();
2970
2971 if (!dbg_cnt (cfg_cleanup))
2972 return changed;
2973
2974 while (try_optimize_cfg (mode))
2975 {
2976 delete_unreachable_blocks (), changed = true;
2977 if (!(mode & CLEANUP_NO_INSN_DEL))
2978 {
2979 /* Try to remove some trivially dead insns when doing an expensive
2980 cleanup. But delete_trivially_dead_insns doesn't work after
2981 reload (it only handles pseudos) and run_fast_dce is too costly
2982 to run in every iteration.
2983
2984 For effective cross jumping, we really want to run a fast DCE to
2985 clean up any dead conditions, or they get in the way of performing
2986 useful tail merges.
2987
2988 Other transformations in cleanup_cfg are not so sensitive to dead
2989 code, so delete_trivially_dead_insns or even doing nothing at all
2990 is good enough. */
2991 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
2992 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
2993 break;
2994 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
2995 run_fast_dce ();
2996 }
2997 else
2998 break;
2999 }
3000
3001 if (mode & CLEANUP_CROSSJUMP)
3002 remove_fake_exit_edges ();
3003
3004 /* Don't call delete_dead_jumptables in cfglayout mode, because
3005 that function assumes that jump tables are in the insns stream.
3006 But we also don't _have_ to delete dead jumptables in cfglayout
3007 mode because we shouldn't even be looking at things that are
3008 not in a basic block. Dead jumptables are cleaned up when
3009 going out of cfglayout mode. */
3010 if (!(mode & CLEANUP_CFGLAYOUT))
3011 delete_dead_jumptables ();
3012
3013 /* ??? We probably do this way too often. */
3014 if (current_loops
3015 && (changed
3016 || (mode & CLEANUP_CFG_CHANGED)))
3017 {
3018 timevar_push (TV_REPAIR_LOOPS);
3019 /* The above doesn't preserve dominance info if available. */
3020 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3021 calculate_dominance_info (CDI_DOMINATORS);
3022 fix_loop_structure (NULL);
3023 free_dominance_info (CDI_DOMINATORS);
3024 timevar_pop (TV_REPAIR_LOOPS);
3025 }
3026
3027 timevar_pop (TV_CLEANUP_CFG);
3028
3029 return changed;
3030 }
3031 \f
3032 static unsigned int
3033 execute_jump (void)
3034 {
3035 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3036 if (dump_file)
3037 dump_flow_info (dump_file, dump_flags);
3038 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3039 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3040 return 0;
3041 }
3042
3043 struct rtl_opt_pass pass_jump =
3044 {
3045 {
3046 RTL_PASS,
3047 "jump", /* name */
3048 OPTGROUP_NONE, /* optinfo_flags */
3049 NULL, /* gate */
3050 execute_jump, /* execute */
3051 NULL, /* sub */
3052 NULL, /* next */
3053 0, /* static_pass_number */
3054 TV_JUMP, /* tv_id */
3055 0, /* properties_required */
3056 0, /* properties_provided */
3057 0, /* properties_destroyed */
3058 0, /* todo_flags_start */
3059 TODO_verify_rtl_sharing, /* todo_flags_finish */
3060 }
3061 };
3062 \f
3063 static unsigned int
3064 execute_jump2 (void)
3065 {
3066 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3067 return 0;
3068 }
3069
3070 struct rtl_opt_pass pass_jump2 =
3071 {
3072 {
3073 RTL_PASS,
3074 "jump2", /* name */
3075 OPTGROUP_NONE, /* optinfo_flags */
3076 NULL, /* gate */
3077 execute_jump2, /* execute */
3078 NULL, /* sub */
3079 NULL, /* next */
3080 0, /* static_pass_number */
3081 TV_JUMP, /* tv_id */
3082 0, /* properties_required */
3083 0, /* properties_provided */
3084 0, /* properties_destroyed */
3085 0, /* todo_flags_start */
3086 TODO_verify_rtl_sharing, /* todo_flags_finish */
3087 }
3088 };