re PR debug/66691 (ICE on valid code at -O3 with -g enabled in simplify_subreg, at...
[gcc.git] / gcc / cfgcleanup.c
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2015 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 "alias.h"
38 #include "symtab.h"
39 #include "tree.h"
40 #include "hard-reg-set.h"
41 #include "regs.h"
42 #include "insn-config.h"
43 #include "flags.h"
44 #include "recog.h"
45 #include "diagnostic-core.h"
46 #include "alloc-pool.h"
47 #include "cselib.h"
48 #include "params.h"
49 #include "tm_p.h"
50 #include "target.h"
51 #include "function.h" /* For inline functions in emit-rtl.h they need crtl. */
52 #include "emit-rtl.h"
53 #include "tree-pass.h"
54 #include "cfgloop.h"
55 #include "function.h"
56 #include "expmed.h"
57 #include "dojump.h"
58 #include "explow.h"
59 #include "calls.h"
60 #include "varasm.h"
61 #include "stmt.h"
62 #include "expr.h"
63 #include "dominance.h"
64 #include "cfg.h"
65 #include "cfgrtl.h"
66 #include "cfganal.h"
67 #include "cfgbuild.h"
68 #include "cfgcleanup.h"
69 #include "predict.h"
70 #include "basic-block.h"
71 #include "df.h"
72 #include "dce.h"
73 #include "dbgcnt.h"
74 #include "rtl-iter.h"
75
76 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
77
78 /* Set to true when we are running first pass of try_optimize_cfg loop. */
79 static bool first_pass;
80
81 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
82 static bool crossjumps_occured;
83
84 /* Set to true if we couldn't run an optimization due to stale liveness
85 information; we should run df_analyze to enable more opportunities. */
86 static bool block_was_dirty;
87
88 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
89 static bool try_crossjump_bb (int, basic_block);
90 static bool outgoing_edges_match (int, basic_block, basic_block);
91 static enum replace_direction old_insns_match_p (int, rtx_insn *, rtx_insn *);
92
93 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
94 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
95 static bool try_optimize_cfg (int);
96 static bool try_simplify_condjump (basic_block);
97 static bool try_forward_edges (int, basic_block);
98 static edge thread_jump (edge, basic_block);
99 static bool mark_effect (rtx, bitmap);
100 static void notice_new_block (basic_block);
101 static void update_forwarder_flag (basic_block);
102 static void merge_memattrs (rtx, rtx);
103 \f
104 /* Set flags for newly created block. */
105
106 static void
107 notice_new_block (basic_block bb)
108 {
109 if (!bb)
110 return;
111
112 if (forwarder_block_p (bb))
113 bb->flags |= BB_FORWARDER_BLOCK;
114 }
115
116 /* Recompute forwarder flag after block has been modified. */
117
118 static void
119 update_forwarder_flag (basic_block bb)
120 {
121 if (forwarder_block_p (bb))
122 bb->flags |= BB_FORWARDER_BLOCK;
123 else
124 bb->flags &= ~BB_FORWARDER_BLOCK;
125 }
126 \f
127 /* Simplify a conditional jump around an unconditional jump.
128 Return true if something changed. */
129
130 static bool
131 try_simplify_condjump (basic_block cbranch_block)
132 {
133 basic_block jump_block, jump_dest_block, cbranch_dest_block;
134 edge cbranch_jump_edge, cbranch_fallthru_edge;
135 rtx_insn *cbranch_insn;
136
137 /* Verify that there are exactly two successors. */
138 if (EDGE_COUNT (cbranch_block->succs) != 2)
139 return false;
140
141 /* Verify that we've got a normal conditional branch at the end
142 of the block. */
143 cbranch_insn = BB_END (cbranch_block);
144 if (!any_condjump_p (cbranch_insn))
145 return false;
146
147 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
148 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
149
150 /* The next block must not have multiple predecessors, must not
151 be the last block in the function, and must contain just the
152 unconditional jump. */
153 jump_block = cbranch_fallthru_edge->dest;
154 if (!single_pred_p (jump_block)
155 || jump_block->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
156 || !FORWARDER_BLOCK_P (jump_block))
157 return false;
158 jump_dest_block = single_succ (jump_block);
159
160 /* If we are partitioning hot/cold basic blocks, we don't want to
161 mess up unconditional or indirect jumps that cross between hot
162 and cold sections.
163
164 Basic block partitioning may result in some jumps that appear to
165 be optimizable (or blocks that appear to be mergeable), but which really
166 must be left untouched (they are required to make it safely across
167 partition boundaries). See the comments at the top of
168 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
169
170 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
171 || (cbranch_jump_edge->flags & EDGE_CROSSING))
172 return false;
173
174 /* The conditional branch must target the block after the
175 unconditional branch. */
176 cbranch_dest_block = cbranch_jump_edge->dest;
177
178 if (cbranch_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
179 || !can_fallthru (jump_block, cbranch_dest_block))
180 return false;
181
182 /* Invert the conditional branch. */
183 if (!invert_jump (as_a <rtx_jump_insn *> (cbranch_insn),
184 block_label (jump_dest_block), 0))
185 return false;
186
187 if (dump_file)
188 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
189 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
190
191 /* Success. Update the CFG to match. Note that after this point
192 the edge variable names appear backwards; the redirection is done
193 this way to preserve edge profile data. */
194 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
195 cbranch_dest_block);
196 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
197 jump_dest_block);
198 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
199 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
200 update_br_prob_note (cbranch_block);
201
202 /* Delete the block with the unconditional jump, and clean up the mess. */
203 delete_basic_block (jump_block);
204 tidy_fallthru_edge (cbranch_jump_edge);
205 update_forwarder_flag (cbranch_block);
206
207 return true;
208 }
209 \f
210 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
211 on register. Used by jump threading. */
212
213 static bool
214 mark_effect (rtx exp, regset nonequal)
215 {
216 rtx dest;
217 switch (GET_CODE (exp))
218 {
219 /* In case we do clobber the register, mark it as equal, as we know the
220 value is dead so it don't have to match. */
221 case CLOBBER:
222 dest = XEXP (exp, 0);
223 if (REG_P (dest))
224 bitmap_clear_range (nonequal, REGNO (dest), REG_NREGS (dest));
225 return false;
226
227 case SET:
228 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
229 return false;
230 dest = SET_DEST (exp);
231 if (dest == pc_rtx)
232 return false;
233 if (!REG_P (dest))
234 return true;
235 bitmap_set_range (nonequal, REGNO (dest), REG_NREGS (dest));
236 return false;
237
238 default:
239 return false;
240 }
241 }
242
243 /* Return true if X contains a register in NONEQUAL. */
244 static bool
245 mentions_nonequal_regs (const_rtx x, regset nonequal)
246 {
247 subrtx_iterator::array_type array;
248 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
249 {
250 const_rtx x = *iter;
251 if (REG_P (x))
252 {
253 unsigned int end_regno = END_REGNO (x);
254 for (unsigned int regno = REGNO (x); regno < end_regno; ++regno)
255 if (REGNO_REG_SET_P (nonequal, regno))
256 return true;
257 }
258 }
259 return false;
260 }
261
262 /* Attempt to prove that the basic block B will have no side effects and
263 always continues in the same edge if reached via E. Return the edge
264 if exist, NULL otherwise. */
265
266 static edge
267 thread_jump (edge e, basic_block b)
268 {
269 rtx set1, set2, cond1, cond2;
270 rtx_insn *insn;
271 enum rtx_code code1, code2, reversed_code2;
272 bool reverse1 = false;
273 unsigned i;
274 regset nonequal;
275 bool failed = false;
276 reg_set_iterator rsi;
277
278 if (b->flags & BB_NONTHREADABLE_BLOCK)
279 return NULL;
280
281 /* At the moment, we do handle only conditional jumps, but later we may
282 want to extend this code to tablejumps and others. */
283 if (EDGE_COUNT (e->src->succs) != 2)
284 return NULL;
285 if (EDGE_COUNT (b->succs) != 2)
286 {
287 b->flags |= BB_NONTHREADABLE_BLOCK;
288 return NULL;
289 }
290
291 /* Second branch must end with onlyjump, as we will eliminate the jump. */
292 if (!any_condjump_p (BB_END (e->src)))
293 return NULL;
294
295 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
296 {
297 b->flags |= BB_NONTHREADABLE_BLOCK;
298 return NULL;
299 }
300
301 set1 = pc_set (BB_END (e->src));
302 set2 = pc_set (BB_END (b));
303 if (((e->flags & EDGE_FALLTHRU) != 0)
304 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
305 reverse1 = true;
306
307 cond1 = XEXP (SET_SRC (set1), 0);
308 cond2 = XEXP (SET_SRC (set2), 0);
309 if (reverse1)
310 code1 = reversed_comparison_code (cond1, BB_END (e->src));
311 else
312 code1 = GET_CODE (cond1);
313
314 code2 = GET_CODE (cond2);
315 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
316
317 if (!comparison_dominates_p (code1, code2)
318 && !comparison_dominates_p (code1, reversed_code2))
319 return NULL;
320
321 /* Ensure that the comparison operators are equivalent.
322 ??? This is far too pessimistic. We should allow swapped operands,
323 different CCmodes, or for example comparisons for interval, that
324 dominate even when operands are not equivalent. */
325 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
326 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
327 return NULL;
328
329 /* Short circuit cases where block B contains some side effects, as we can't
330 safely bypass it. */
331 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
332 insn = NEXT_INSN (insn))
333 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
334 {
335 b->flags |= BB_NONTHREADABLE_BLOCK;
336 return NULL;
337 }
338
339 cselib_init (0);
340
341 /* First process all values computed in the source basic block. */
342 for (insn = NEXT_INSN (BB_HEAD (e->src));
343 insn != NEXT_INSN (BB_END (e->src));
344 insn = NEXT_INSN (insn))
345 if (INSN_P (insn))
346 cselib_process_insn (insn);
347
348 nonequal = BITMAP_ALLOC (NULL);
349 CLEAR_REG_SET (nonequal);
350
351 /* Now assume that we've continued by the edge E to B and continue
352 processing as if it were same basic block.
353 Our goal is to prove that whole block is an NOOP. */
354
355 for (insn = NEXT_INSN (BB_HEAD (b));
356 insn != NEXT_INSN (BB_END (b)) && !failed;
357 insn = NEXT_INSN (insn))
358 {
359 if (INSN_P (insn))
360 {
361 rtx pat = PATTERN (insn);
362
363 if (GET_CODE (pat) == PARALLEL)
364 {
365 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
366 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
367 }
368 else
369 failed |= mark_effect (pat, nonequal);
370 }
371
372 cselib_process_insn (insn);
373 }
374
375 /* Later we should clear nonequal of dead registers. So far we don't
376 have life information in cfg_cleanup. */
377 if (failed)
378 {
379 b->flags |= BB_NONTHREADABLE_BLOCK;
380 goto failed_exit;
381 }
382
383 /* cond2 must not mention any register that is not equal to the
384 former block. */
385 if (mentions_nonequal_regs (cond2, nonequal))
386 goto failed_exit;
387
388 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
389 goto failed_exit;
390
391 BITMAP_FREE (nonequal);
392 cselib_finish ();
393 if ((comparison_dominates_p (code1, code2) != 0)
394 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
395 return BRANCH_EDGE (b);
396 else
397 return FALLTHRU_EDGE (b);
398
399 failed_exit:
400 BITMAP_FREE (nonequal);
401 cselib_finish ();
402 return NULL;
403 }
404 \f
405 /* Attempt to forward edges leaving basic block B.
406 Return true if successful. */
407
408 static bool
409 try_forward_edges (int mode, basic_block b)
410 {
411 bool changed = false;
412 edge_iterator ei;
413 edge e, *threaded_edges = NULL;
414
415 /* If we are partitioning hot/cold basic blocks, we don't want to
416 mess up unconditional or indirect jumps that cross between hot
417 and cold sections.
418
419 Basic block partitioning may result in some jumps that appear to
420 be optimizable (or blocks that appear to be mergeable), but which really
421 must be left untouched (they are required to make it safely across
422 partition boundaries). See the comments at the top of
423 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
424
425 if (JUMP_P (BB_END (b)) && CROSSING_JUMP_P (BB_END (b)))
426 return false;
427
428 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
429 {
430 basic_block target, first;
431 location_t goto_locus;
432 int counter;
433 bool threaded = false;
434 int nthreaded_edges = 0;
435 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
436
437 /* Skip complex edges because we don't know how to update them.
438
439 Still handle fallthru edges, as we can succeed to forward fallthru
440 edge to the same place as the branch edge of conditional branch
441 and turn conditional branch to an unconditional branch. */
442 if (e->flags & EDGE_COMPLEX)
443 {
444 ei_next (&ei);
445 continue;
446 }
447
448 target = first = e->dest;
449 counter = NUM_FIXED_BLOCKS;
450 goto_locus = e->goto_locus;
451
452 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
453 up jumps that cross between hot/cold sections.
454
455 Basic block partitioning may result in some jumps that appear
456 to be optimizable (or blocks that appear to be mergeable), but which
457 really must be left untouched (they are required to make it safely
458 across partition boundaries). See the comments at the top of
459 bb-reorder.c:partition_hot_cold_basic_blocks for complete
460 details. */
461
462 if (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
463 && JUMP_P (BB_END (first))
464 && CROSSING_JUMP_P (BB_END (first)))
465 return changed;
466
467 while (counter < n_basic_blocks_for_fn (cfun))
468 {
469 basic_block new_target = NULL;
470 bool new_target_threaded = false;
471 may_thread |= (target->flags & BB_MODIFIED) != 0;
472
473 if (FORWARDER_BLOCK_P (target)
474 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
475 && single_succ (target) != EXIT_BLOCK_PTR_FOR_FN (cfun))
476 {
477 /* Bypass trivial infinite loops. */
478 new_target = single_succ (target);
479 if (target == new_target)
480 counter = n_basic_blocks_for_fn (cfun);
481 else if (!optimize)
482 {
483 /* When not optimizing, ensure that edges or forwarder
484 blocks with different locus are not optimized out. */
485 location_t new_locus = single_succ_edge (target)->goto_locus;
486 location_t locus = goto_locus;
487
488 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
489 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
490 && new_locus != locus)
491 new_target = NULL;
492 else
493 {
494 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
495 locus = new_locus;
496
497 rtx_insn *last = BB_END (target);
498 if (DEBUG_INSN_P (last))
499 last = prev_nondebug_insn (last);
500 if (last && INSN_P (last))
501 new_locus = INSN_LOCATION (last);
502 else
503 new_locus = UNKNOWN_LOCATION;
504
505 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
506 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
507 && new_locus != locus)
508 new_target = NULL;
509 else
510 {
511 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
512 locus = new_locus;
513
514 goto_locus = locus;
515 }
516 }
517 }
518 }
519
520 /* Allow to thread only over one edge at time to simplify updating
521 of probabilities. */
522 else if ((mode & CLEANUP_THREADING) && may_thread)
523 {
524 edge t = thread_jump (e, target);
525 if (t)
526 {
527 if (!threaded_edges)
528 threaded_edges = XNEWVEC (edge,
529 n_basic_blocks_for_fn (cfun));
530 else
531 {
532 int i;
533
534 /* Detect an infinite loop across blocks not
535 including the start block. */
536 for (i = 0; i < nthreaded_edges; ++i)
537 if (threaded_edges[i] == t)
538 break;
539 if (i < nthreaded_edges)
540 {
541 counter = n_basic_blocks_for_fn (cfun);
542 break;
543 }
544 }
545
546 /* Detect an infinite loop across the start block. */
547 if (t->dest == b)
548 break;
549
550 gcc_assert (nthreaded_edges
551 < (n_basic_blocks_for_fn (cfun)
552 - NUM_FIXED_BLOCKS));
553 threaded_edges[nthreaded_edges++] = t;
554
555 new_target = t->dest;
556 new_target_threaded = true;
557 }
558 }
559
560 if (!new_target)
561 break;
562
563 counter++;
564 target = new_target;
565 threaded |= new_target_threaded;
566 }
567
568 if (counter >= n_basic_blocks_for_fn (cfun))
569 {
570 if (dump_file)
571 fprintf (dump_file, "Infinite loop in BB %i.\n",
572 target->index);
573 }
574 else if (target == first)
575 ; /* We didn't do anything. */
576 else
577 {
578 /* Save the values now, as the edge may get removed. */
579 gcov_type edge_count = e->count;
580 int edge_probability = e->probability;
581 int edge_frequency;
582 int n = 0;
583
584 e->goto_locus = goto_locus;
585
586 /* Don't force if target is exit block. */
587 if (threaded && target != EXIT_BLOCK_PTR_FOR_FN (cfun))
588 {
589 notice_new_block (redirect_edge_and_branch_force (e, target));
590 if (dump_file)
591 fprintf (dump_file, "Conditionals threaded.\n");
592 }
593 else if (!redirect_edge_and_branch (e, target))
594 {
595 if (dump_file)
596 fprintf (dump_file,
597 "Forwarding edge %i->%i to %i failed.\n",
598 b->index, e->dest->index, target->index);
599 ei_next (&ei);
600 continue;
601 }
602
603 /* We successfully forwarded the edge. Now update profile
604 data: for each edge we traversed in the chain, remove
605 the original edge's execution count. */
606 edge_frequency = apply_probability (b->frequency, edge_probability);
607
608 do
609 {
610 edge t;
611
612 if (!single_succ_p (first))
613 {
614 gcc_assert (n < nthreaded_edges);
615 t = threaded_edges [n++];
616 gcc_assert (t->src == first);
617 update_bb_profile_for_threading (first, edge_frequency,
618 edge_count, t);
619 update_br_prob_note (first);
620 }
621 else
622 {
623 first->count -= edge_count;
624 if (first->count < 0)
625 first->count = 0;
626 first->frequency -= edge_frequency;
627 if (first->frequency < 0)
628 first->frequency = 0;
629 /* It is possible that as the result of
630 threading we've removed edge as it is
631 threaded to the fallthru edge. Avoid
632 getting out of sync. */
633 if (n < nthreaded_edges
634 && first == threaded_edges [n]->src)
635 n++;
636 t = single_succ_edge (first);
637 }
638
639 t->count -= edge_count;
640 if (t->count < 0)
641 t->count = 0;
642 first = t->dest;
643 }
644 while (first != target);
645
646 changed = true;
647 continue;
648 }
649 ei_next (&ei);
650 }
651
652 free (threaded_edges);
653 return changed;
654 }
655 \f
656
657 /* Blocks A and B are to be merged into a single block. A has no incoming
658 fallthru edge, so it can be moved before B without adding or modifying
659 any jumps (aside from the jump from A to B). */
660
661 static void
662 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
663 {
664 rtx_insn *barrier;
665
666 /* If we are partitioning hot/cold basic blocks, we don't want to
667 mess up unconditional or indirect jumps that cross between hot
668 and cold sections.
669
670 Basic block partitioning may result in some jumps that appear to
671 be optimizable (or blocks that appear to be mergeable), but which really
672 must be left untouched (they are required to make it safely across
673 partition boundaries). See the comments at the top of
674 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
675
676 if (BB_PARTITION (a) != BB_PARTITION (b))
677 return;
678
679 barrier = next_nonnote_insn (BB_END (a));
680 gcc_assert (BARRIER_P (barrier));
681 delete_insn (barrier);
682
683 /* Scramble the insn chain. */
684 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
685 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
686 df_set_bb_dirty (a);
687
688 if (dump_file)
689 fprintf (dump_file, "Moved block %d before %d and merged.\n",
690 a->index, b->index);
691
692 /* Swap the records for the two blocks around. */
693
694 unlink_block (a);
695 link_block (a, b->prev_bb);
696
697 /* Now blocks A and B are contiguous. Merge them. */
698 merge_blocks (a, b);
699 }
700
701 /* Blocks A and B are to be merged into a single block. B has no outgoing
702 fallthru edge, so it can be moved after A without adding or modifying
703 any jumps (aside from the jump from A to B). */
704
705 static void
706 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
707 {
708 rtx_insn *barrier, *real_b_end;
709 rtx label;
710 rtx_jump_table_data *table;
711
712 /* If we are partitioning hot/cold basic blocks, we don't want to
713 mess up unconditional or indirect jumps that cross between hot
714 and cold sections.
715
716 Basic block partitioning may result in some jumps that appear to
717 be optimizable (or blocks that appear to be mergeable), but which really
718 must be left untouched (they are required to make it safely across
719 partition boundaries). See the comments at the top of
720 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
721
722 if (BB_PARTITION (a) != BB_PARTITION (b))
723 return;
724
725 real_b_end = BB_END (b);
726
727 /* If there is a jump table following block B temporarily add the jump table
728 to block B so that it will also be moved to the correct location. */
729 if (tablejump_p (BB_END (b), &label, &table)
730 && prev_active_insn (label) == BB_END (b))
731 {
732 BB_END (b) = table;
733 }
734
735 /* There had better have been a barrier there. Delete it. */
736 barrier = NEXT_INSN (BB_END (b));
737 if (barrier && BARRIER_P (barrier))
738 delete_insn (barrier);
739
740
741 /* Scramble the insn chain. */
742 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
743
744 /* Restore the real end of b. */
745 BB_END (b) = real_b_end;
746
747 if (dump_file)
748 fprintf (dump_file, "Moved block %d after %d and merged.\n",
749 b->index, a->index);
750
751 /* Now blocks A and B are contiguous. Merge them. */
752 merge_blocks (a, b);
753 }
754
755 /* Attempt to merge basic blocks that are potentially non-adjacent.
756 Return NULL iff the attempt failed, otherwise return basic block
757 where cleanup_cfg should continue. Because the merging commonly
758 moves basic block away or introduces another optimization
759 possibility, return basic block just before B so cleanup_cfg don't
760 need to iterate.
761
762 It may be good idea to return basic block before C in the case
763 C has been moved after B and originally appeared earlier in the
764 insn sequence, but we have no information available about the
765 relative ordering of these two. Hopefully it is not too common. */
766
767 static basic_block
768 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
769 {
770 basic_block next;
771
772 /* If we are partitioning hot/cold basic blocks, we don't want to
773 mess up unconditional or indirect jumps that cross between hot
774 and cold sections.
775
776 Basic block partitioning may result in some jumps that appear to
777 be optimizable (or blocks that appear to be mergeable), but which really
778 must be left untouched (they are required to make it safely across
779 partition boundaries). See the comments at the top of
780 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
781
782 if (BB_PARTITION (b) != BB_PARTITION (c))
783 return NULL;
784
785 /* If B has a fallthru edge to C, no need to move anything. */
786 if (e->flags & EDGE_FALLTHRU)
787 {
788 int b_index = b->index, c_index = c->index;
789
790 /* Protect the loop latches. */
791 if (current_loops && c->loop_father->latch == c)
792 return NULL;
793
794 merge_blocks (b, c);
795 update_forwarder_flag (b);
796
797 if (dump_file)
798 fprintf (dump_file, "Merged %d and %d without moving.\n",
799 b_index, c_index);
800
801 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
802 }
803
804 /* Otherwise we will need to move code around. Do that only if expensive
805 transformations are allowed. */
806 else if (mode & CLEANUP_EXPENSIVE)
807 {
808 edge tmp_edge, b_fallthru_edge;
809 bool c_has_outgoing_fallthru;
810 bool b_has_incoming_fallthru;
811
812 /* Avoid overactive code motion, as the forwarder blocks should be
813 eliminated by edge redirection instead. One exception might have
814 been if B is a forwarder block and C has no fallthru edge, but
815 that should be cleaned up by bb-reorder instead. */
816 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
817 return NULL;
818
819 /* We must make sure to not munge nesting of lexical blocks,
820 and loop notes. This is done by squeezing out all the notes
821 and leaving them there to lie. Not ideal, but functional. */
822
823 tmp_edge = find_fallthru_edge (c->succs);
824 c_has_outgoing_fallthru = (tmp_edge != NULL);
825
826 tmp_edge = find_fallthru_edge (b->preds);
827 b_has_incoming_fallthru = (tmp_edge != NULL);
828 b_fallthru_edge = tmp_edge;
829 next = b->prev_bb;
830 if (next == c)
831 next = next->prev_bb;
832
833 /* Otherwise, we're going to try to move C after B. If C does
834 not have an outgoing fallthru, then it can be moved
835 immediately after B without introducing or modifying jumps. */
836 if (! c_has_outgoing_fallthru)
837 {
838 merge_blocks_move_successor_nojumps (b, c);
839 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
840 }
841
842 /* If B does not have an incoming fallthru, then it can be moved
843 immediately before C without introducing or modifying jumps.
844 C cannot be the first block, so we do not have to worry about
845 accessing a non-existent block. */
846
847 if (b_has_incoming_fallthru)
848 {
849 basic_block bb;
850
851 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
852 return NULL;
853 bb = force_nonfallthru (b_fallthru_edge);
854 if (bb)
855 notice_new_block (bb);
856 }
857
858 merge_blocks_move_predecessor_nojumps (b, c);
859 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
860 }
861
862 return NULL;
863 }
864 \f
865
866 /* Removes the memory attributes of MEM expression
867 if they are not equal. */
868
869 static void
870 merge_memattrs (rtx x, rtx y)
871 {
872 int i;
873 int j;
874 enum rtx_code code;
875 const char *fmt;
876
877 if (x == y)
878 return;
879 if (x == 0 || y == 0)
880 return;
881
882 code = GET_CODE (x);
883
884 if (code != GET_CODE (y))
885 return;
886
887 if (GET_MODE (x) != GET_MODE (y))
888 return;
889
890 if (code == MEM && !mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
891 {
892 if (! MEM_ATTRS (x))
893 MEM_ATTRS (y) = 0;
894 else if (! MEM_ATTRS (y))
895 MEM_ATTRS (x) = 0;
896 else
897 {
898 HOST_WIDE_INT mem_size;
899
900 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
901 {
902 set_mem_alias_set (x, 0);
903 set_mem_alias_set (y, 0);
904 }
905
906 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
907 {
908 set_mem_expr (x, 0);
909 set_mem_expr (y, 0);
910 clear_mem_offset (x);
911 clear_mem_offset (y);
912 }
913 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
914 || (MEM_OFFSET_KNOWN_P (x)
915 && MEM_OFFSET (x) != MEM_OFFSET (y)))
916 {
917 clear_mem_offset (x);
918 clear_mem_offset (y);
919 }
920
921 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
922 {
923 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
924 set_mem_size (x, mem_size);
925 set_mem_size (y, mem_size);
926 }
927 else
928 {
929 clear_mem_size (x);
930 clear_mem_size (y);
931 }
932
933 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
934 set_mem_align (y, MEM_ALIGN (x));
935 }
936 }
937 if (code == MEM)
938 {
939 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
940 {
941 MEM_READONLY_P (x) = 0;
942 MEM_READONLY_P (y) = 0;
943 }
944 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
945 {
946 MEM_NOTRAP_P (x) = 0;
947 MEM_NOTRAP_P (y) = 0;
948 }
949 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
950 {
951 MEM_VOLATILE_P (x) = 1;
952 MEM_VOLATILE_P (y) = 1;
953 }
954 }
955
956 fmt = GET_RTX_FORMAT (code);
957 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
958 {
959 switch (fmt[i])
960 {
961 case 'E':
962 /* Two vectors must have the same length. */
963 if (XVECLEN (x, i) != XVECLEN (y, i))
964 return;
965
966 for (j = 0; j < XVECLEN (x, i); j++)
967 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
968
969 break;
970
971 case 'e':
972 merge_memattrs (XEXP (x, i), XEXP (y, i));
973 }
974 }
975 return;
976 }
977
978
979 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
980 different single sets S1 and S2. */
981
982 static bool
983 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
984 {
985 int i;
986 rtx e1, e2;
987
988 if (p1 == s1 && p2 == s2)
989 return true;
990
991 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
992 return false;
993
994 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
995 return false;
996
997 for (i = 0; i < XVECLEN (p1, 0); i++)
998 {
999 e1 = XVECEXP (p1, 0, i);
1000 e2 = XVECEXP (p2, 0, i);
1001 if (e1 == s1 && e2 == s2)
1002 continue;
1003 if (reload_completed
1004 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
1005 continue;
1006
1007 return false;
1008 }
1009
1010 return true;
1011 }
1012
1013
1014 /* NOTE1 is the REG_EQUAL note, if any, attached to an insn
1015 that is a single_set with a SET_SRC of SRC1. Similarly
1016 for NOTE2/SRC2.
1017
1018 So effectively NOTE1/NOTE2 are an alternate form of
1019 SRC1/SRC2 respectively.
1020
1021 Return nonzero if SRC1 or NOTE1 has the same constant
1022 integer value as SRC2 or NOTE2. Else return zero. */
1023 static int
1024 values_equal_p (rtx note1, rtx note2, rtx src1, rtx src2)
1025 {
1026 if (note1
1027 && note2
1028 && CONST_INT_P (XEXP (note1, 0))
1029 && rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0)))
1030 return 1;
1031
1032 if (!note1
1033 && !note2
1034 && CONST_INT_P (src1)
1035 && CONST_INT_P (src2)
1036 && rtx_equal_p (src1, src2))
1037 return 1;
1038
1039 if (note1
1040 && CONST_INT_P (src2)
1041 && rtx_equal_p (XEXP (note1, 0), src2))
1042 return 1;
1043
1044 if (note2
1045 && CONST_INT_P (src1)
1046 && rtx_equal_p (XEXP (note2, 0), src1))
1047 return 1;
1048
1049 return 0;
1050 }
1051
1052 /* Examine register notes on I1 and I2 and return:
1053 - dir_forward if I1 can be replaced by I2, or
1054 - dir_backward if I2 can be replaced by I1, or
1055 - dir_both if both are the case. */
1056
1057 static enum replace_direction
1058 can_replace_by (rtx_insn *i1, rtx_insn *i2)
1059 {
1060 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1061 bool c1, c2;
1062
1063 /* Check for 2 sets. */
1064 s1 = single_set (i1);
1065 s2 = single_set (i2);
1066 if (s1 == NULL_RTX || s2 == NULL_RTX)
1067 return dir_none;
1068
1069 /* Check that the 2 sets set the same dest. */
1070 d1 = SET_DEST (s1);
1071 d2 = SET_DEST (s2);
1072 if (!(reload_completed
1073 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1074 return dir_none;
1075
1076 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1077 set dest to the same value. */
1078 note1 = find_reg_equal_equiv_note (i1);
1079 note2 = find_reg_equal_equiv_note (i2);
1080
1081 src1 = SET_SRC (s1);
1082 src2 = SET_SRC (s2);
1083
1084 if (!values_equal_p (note1, note2, src1, src2))
1085 return dir_none;
1086
1087 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1088 return dir_none;
1089
1090 /* Although the 2 sets set dest to the same value, we cannot replace
1091 (set (dest) (const_int))
1092 by
1093 (set (dest) (reg))
1094 because we don't know if the reg is live and has the same value at the
1095 location of replacement. */
1096 c1 = CONST_INT_P (src1);
1097 c2 = CONST_INT_P (src2);
1098 if (c1 && c2)
1099 return dir_both;
1100 else if (c2)
1101 return dir_forward;
1102 else if (c1)
1103 return dir_backward;
1104
1105 return dir_none;
1106 }
1107
1108 /* Merges directions A and B. */
1109
1110 static enum replace_direction
1111 merge_dir (enum replace_direction a, enum replace_direction b)
1112 {
1113 /* Implements the following table:
1114 |bo fw bw no
1115 ---+-----------
1116 bo |bo fw bw no
1117 fw |-- fw no no
1118 bw |-- -- bw no
1119 no |-- -- -- no. */
1120
1121 if (a == b)
1122 return a;
1123
1124 if (a == dir_both)
1125 return b;
1126 if (b == dir_both)
1127 return a;
1128
1129 return dir_none;
1130 }
1131
1132 /* Examine I1 and I2 and return:
1133 - dir_forward if I1 can be replaced by I2, or
1134 - dir_backward if I2 can be replaced by I1, or
1135 - dir_both if both are the case. */
1136
1137 static enum replace_direction
1138 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1139 {
1140 rtx p1, p2;
1141
1142 /* Verify that I1 and I2 are equivalent. */
1143 if (GET_CODE (i1) != GET_CODE (i2))
1144 return dir_none;
1145
1146 /* __builtin_unreachable() may lead to empty blocks (ending with
1147 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1148 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1149 return dir_both;
1150
1151 /* ??? Do not allow cross-jumping between different stack levels. */
1152 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1153 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1154 if (p1 && p2)
1155 {
1156 p1 = XEXP (p1, 0);
1157 p2 = XEXP (p2, 0);
1158 if (!rtx_equal_p (p1, p2))
1159 return dir_none;
1160
1161 /* ??? Worse, this adjustment had better be constant lest we
1162 have differing incoming stack levels. */
1163 if (!frame_pointer_needed
1164 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1165 return dir_none;
1166 }
1167 else if (p1 || p2)
1168 return dir_none;
1169
1170 p1 = PATTERN (i1);
1171 p2 = PATTERN (i2);
1172
1173 if (GET_CODE (p1) != GET_CODE (p2))
1174 return dir_none;
1175
1176 /* If this is a CALL_INSN, compare register usage information.
1177 If we don't check this on stack register machines, the two
1178 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1179 numbers of stack registers in the same basic block.
1180 If we don't check this on machines with delay slots, a delay slot may
1181 be filled that clobbers a parameter expected by the subroutine.
1182
1183 ??? We take the simple route for now and assume that if they're
1184 equal, they were constructed identically.
1185
1186 Also check for identical exception regions. */
1187
1188 if (CALL_P (i1))
1189 {
1190 /* Ensure the same EH region. */
1191 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1192 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1193
1194 if (!n1 && n2)
1195 return dir_none;
1196
1197 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1198 return dir_none;
1199
1200 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1201 CALL_INSN_FUNCTION_USAGE (i2))
1202 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1203 return dir_none;
1204
1205 /* For address sanitizer, never crossjump __asan_report_* builtins,
1206 otherwise errors might be reported on incorrect lines. */
1207 if (flag_sanitize & SANITIZE_ADDRESS)
1208 {
1209 rtx call = get_call_rtx_from (i1);
1210 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1211 {
1212 rtx symbol = XEXP (XEXP (call, 0), 0);
1213 if (SYMBOL_REF_DECL (symbol)
1214 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1215 {
1216 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1217 == BUILT_IN_NORMAL)
1218 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1219 >= BUILT_IN_ASAN_REPORT_LOAD1
1220 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1221 <= BUILT_IN_ASAN_STOREN)
1222 return dir_none;
1223 }
1224 }
1225 }
1226 }
1227
1228 #ifdef STACK_REGS
1229 /* If cross_jump_death_matters is not 0, the insn's mode
1230 indicates whether or not the insn contains any stack-like
1231 regs. */
1232
1233 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1234 {
1235 /* If register stack conversion has already been done, then
1236 death notes must also be compared before it is certain that
1237 the two instruction streams match. */
1238
1239 rtx note;
1240 HARD_REG_SET i1_regset, i2_regset;
1241
1242 CLEAR_HARD_REG_SET (i1_regset);
1243 CLEAR_HARD_REG_SET (i2_regset);
1244
1245 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1246 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1247 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1248
1249 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1250 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1251 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1252
1253 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1254 return dir_none;
1255 }
1256 #endif
1257
1258 if (reload_completed
1259 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1260 return dir_both;
1261
1262 return can_replace_by (i1, i2);
1263 }
1264 \f
1265 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1266 flow_find_head_matching_sequence, ensure the notes match. */
1267
1268 static void
1269 merge_notes (rtx_insn *i1, rtx_insn *i2)
1270 {
1271 /* If the merged insns have different REG_EQUAL notes, then
1272 remove them. */
1273 rtx equiv1 = find_reg_equal_equiv_note (i1);
1274 rtx equiv2 = find_reg_equal_equiv_note (i2);
1275
1276 if (equiv1 && !equiv2)
1277 remove_note (i1, equiv1);
1278 else if (!equiv1 && equiv2)
1279 remove_note (i2, equiv2);
1280 else if (equiv1 && equiv2
1281 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1282 {
1283 remove_note (i1, equiv1);
1284 remove_note (i2, equiv2);
1285 }
1286 }
1287
1288 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1289 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1290 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1291 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1292 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1293
1294 static void
1295 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1296 bool *did_fallthru)
1297 {
1298 edge fallthru;
1299
1300 *did_fallthru = false;
1301
1302 /* Ignore notes. */
1303 while (!NONDEBUG_INSN_P (*i1))
1304 {
1305 if (*i1 != BB_HEAD (*bb1))
1306 {
1307 *i1 = PREV_INSN (*i1);
1308 continue;
1309 }
1310
1311 if (!follow_fallthru)
1312 return;
1313
1314 fallthru = find_fallthru_edge ((*bb1)->preds);
1315 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1316 || !single_succ_p (fallthru->src))
1317 return;
1318
1319 *bb1 = fallthru->src;
1320 *i1 = BB_END (*bb1);
1321 *did_fallthru = true;
1322 }
1323 }
1324
1325 /* Look through the insns at the end of BB1 and BB2 and find the longest
1326 sequence that are either equivalent, or allow forward or backward
1327 replacement. Store the first insns for that sequence in *F1 and *F2 and
1328 return the sequence length.
1329
1330 DIR_P indicates the allowed replacement direction on function entry, and
1331 the actual replacement direction on function exit. If NULL, only equivalent
1332 sequences are allowed.
1333
1334 To simplify callers of this function, if the blocks match exactly,
1335 store the head of the blocks in *F1 and *F2. */
1336
1337 int
1338 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1339 rtx_insn **f2, enum replace_direction *dir_p)
1340 {
1341 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1342 int ninsns = 0;
1343 enum replace_direction dir, last_dir, afterlast_dir;
1344 bool follow_fallthru, did_fallthru;
1345
1346 if (dir_p)
1347 dir = *dir_p;
1348 else
1349 dir = dir_both;
1350 afterlast_dir = dir;
1351 last_dir = afterlast_dir;
1352
1353 /* Skip simple jumps at the end of the blocks. Complex jumps still
1354 need to be compared for equivalence, which we'll do below. */
1355
1356 i1 = BB_END (bb1);
1357 last1 = afterlast1 = last2 = afterlast2 = NULL;
1358 if (onlyjump_p (i1)
1359 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1360 {
1361 last1 = i1;
1362 i1 = PREV_INSN (i1);
1363 }
1364
1365 i2 = BB_END (bb2);
1366 if (onlyjump_p (i2)
1367 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1368 {
1369 last2 = i2;
1370 /* Count everything except for unconditional jump as insn.
1371 Don't count any jumps if dir_p is NULL. */
1372 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1373 ninsns++;
1374 i2 = PREV_INSN (i2);
1375 }
1376
1377 while (true)
1378 {
1379 /* In the following example, we can replace all jumps to C by jumps to A.
1380
1381 This removes 4 duplicate insns.
1382 [bb A] insn1 [bb C] insn1
1383 insn2 insn2
1384 [bb B] insn3 insn3
1385 insn4 insn4
1386 jump_insn jump_insn
1387
1388 We could also replace all jumps to A by jumps to C, but that leaves B
1389 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1390 step, all jumps to B would be replaced with jumps to the middle of C,
1391 achieving the same result with more effort.
1392 So we allow only the first possibility, which means that we don't allow
1393 fallthru in the block that's being replaced. */
1394
1395 follow_fallthru = dir_p && dir != dir_forward;
1396 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1397 if (did_fallthru)
1398 dir = dir_backward;
1399
1400 follow_fallthru = dir_p && dir != dir_backward;
1401 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1402 if (did_fallthru)
1403 dir = dir_forward;
1404
1405 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1406 break;
1407
1408 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1409 if (dir == dir_none || (!dir_p && dir != dir_both))
1410 break;
1411
1412 merge_memattrs (i1, i2);
1413
1414 /* Don't begin a cross-jump with a NOTE insn. */
1415 if (INSN_P (i1))
1416 {
1417 merge_notes (i1, i2);
1418
1419 afterlast1 = last1, afterlast2 = last2;
1420 last1 = i1, last2 = i2;
1421 afterlast_dir = last_dir;
1422 last_dir = dir;
1423 if (active_insn_p (i1))
1424 ninsns++;
1425 }
1426
1427 i1 = PREV_INSN (i1);
1428 i2 = PREV_INSN (i2);
1429 }
1430
1431 /* Don't allow the insn after a compare to be shared by
1432 cross-jumping unless the compare is also shared. */
1433 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1434 && ! sets_cc0_p (last1))
1435 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1436
1437 /* Include preceding notes and labels in the cross-jump. One,
1438 this may bring us to the head of the blocks as requested above.
1439 Two, it keeps line number notes as matched as may be. */
1440 if (ninsns)
1441 {
1442 bb1 = BLOCK_FOR_INSN (last1);
1443 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1444 last1 = PREV_INSN (last1);
1445
1446 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1447 last1 = PREV_INSN (last1);
1448
1449 bb2 = BLOCK_FOR_INSN (last2);
1450 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1451 last2 = PREV_INSN (last2);
1452
1453 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1454 last2 = PREV_INSN (last2);
1455
1456 *f1 = last1;
1457 *f2 = last2;
1458 }
1459
1460 if (dir_p)
1461 *dir_p = last_dir;
1462 return ninsns;
1463 }
1464
1465 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1466 the head of the two blocks. Do not include jumps at the end.
1467 If STOP_AFTER is nonzero, stop after finding that many matching
1468 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1469 non-zero, only count active insns. */
1470
1471 int
1472 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1473 rtx_insn **f2, int stop_after)
1474 {
1475 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1476 int ninsns = 0;
1477 edge e;
1478 edge_iterator ei;
1479 int nehedges1 = 0, nehedges2 = 0;
1480
1481 FOR_EACH_EDGE (e, ei, bb1->succs)
1482 if (e->flags & EDGE_EH)
1483 nehedges1++;
1484 FOR_EACH_EDGE (e, ei, bb2->succs)
1485 if (e->flags & EDGE_EH)
1486 nehedges2++;
1487
1488 i1 = BB_HEAD (bb1);
1489 i2 = BB_HEAD (bb2);
1490 last1 = beforelast1 = last2 = beforelast2 = NULL;
1491
1492 while (true)
1493 {
1494 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1495 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1496 {
1497 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1498 break;
1499 i1 = NEXT_INSN (i1);
1500 }
1501
1502 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1503 {
1504 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1505 break;
1506 i2 = NEXT_INSN (i2);
1507 }
1508
1509 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1510 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1511 break;
1512
1513 if (NOTE_P (i1) || NOTE_P (i2)
1514 || JUMP_P (i1) || JUMP_P (i2))
1515 break;
1516
1517 /* A sanity check to make sure we're not merging insns with different
1518 effects on EH. If only one of them ends a basic block, it shouldn't
1519 have an EH edge; if both end a basic block, there should be the same
1520 number of EH edges. */
1521 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1522 && nehedges1 > 0)
1523 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1524 && nehedges2 > 0)
1525 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1526 && nehedges1 != nehedges2))
1527 break;
1528
1529 if (old_insns_match_p (0, i1, i2) != dir_both)
1530 break;
1531
1532 merge_memattrs (i1, i2);
1533
1534 /* Don't begin a cross-jump with a NOTE insn. */
1535 if (INSN_P (i1))
1536 {
1537 merge_notes (i1, i2);
1538
1539 beforelast1 = last1, beforelast2 = last2;
1540 last1 = i1, last2 = i2;
1541 if (!stop_after || active_insn_p (i1))
1542 ninsns++;
1543 }
1544
1545 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1546 || (stop_after > 0 && ninsns == stop_after))
1547 break;
1548
1549 i1 = NEXT_INSN (i1);
1550 i2 = NEXT_INSN (i2);
1551 }
1552
1553 /* Don't allow a compare to be shared by cross-jumping unless the insn
1554 after the compare is also shared. */
1555 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1556 && sets_cc0_p (last1))
1557 last1 = beforelast1, last2 = beforelast2, ninsns--;
1558
1559 if (ninsns)
1560 {
1561 *f1 = last1;
1562 *f2 = last2;
1563 }
1564
1565 return ninsns;
1566 }
1567
1568 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1569 the branch instruction. This means that if we commonize the control
1570 flow before end of the basic block, the semantic remains unchanged.
1571
1572 We may assume that there exists one edge with a common destination. */
1573
1574 static bool
1575 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1576 {
1577 int nehedges1 = 0, nehedges2 = 0;
1578 edge fallthru1 = 0, fallthru2 = 0;
1579 edge e1, e2;
1580 edge_iterator ei;
1581
1582 /* If we performed shrink-wrapping, edges to the exit block can
1583 only be distinguished for JUMP_INSNs. The two paths may differ in
1584 whether they went through the prologue. Sibcalls are fine, we know
1585 that we either didn't need or inserted an epilogue before them. */
1586 if (crtl->shrink_wrapped
1587 && single_succ_p (bb1)
1588 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1589 && !JUMP_P (BB_END (bb1))
1590 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1591 return false;
1592
1593 /* If BB1 has only one successor, we may be looking at either an
1594 unconditional jump, or a fake edge to exit. */
1595 if (single_succ_p (bb1)
1596 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1597 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1598 return (single_succ_p (bb2)
1599 && (single_succ_edge (bb2)->flags
1600 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1601 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1602
1603 /* Match conditional jumps - this may get tricky when fallthru and branch
1604 edges are crossed. */
1605 if (EDGE_COUNT (bb1->succs) == 2
1606 && any_condjump_p (BB_END (bb1))
1607 && onlyjump_p (BB_END (bb1)))
1608 {
1609 edge b1, f1, b2, f2;
1610 bool reverse, match;
1611 rtx set1, set2, cond1, cond2;
1612 enum rtx_code code1, code2;
1613
1614 if (EDGE_COUNT (bb2->succs) != 2
1615 || !any_condjump_p (BB_END (bb2))
1616 || !onlyjump_p (BB_END (bb2)))
1617 return false;
1618
1619 b1 = BRANCH_EDGE (bb1);
1620 b2 = BRANCH_EDGE (bb2);
1621 f1 = FALLTHRU_EDGE (bb1);
1622 f2 = FALLTHRU_EDGE (bb2);
1623
1624 /* Get around possible forwarders on fallthru edges. Other cases
1625 should be optimized out already. */
1626 if (FORWARDER_BLOCK_P (f1->dest))
1627 f1 = single_succ_edge (f1->dest);
1628
1629 if (FORWARDER_BLOCK_P (f2->dest))
1630 f2 = single_succ_edge (f2->dest);
1631
1632 /* To simplify use of this function, return false if there are
1633 unneeded forwarder blocks. These will get eliminated later
1634 during cleanup_cfg. */
1635 if (FORWARDER_BLOCK_P (f1->dest)
1636 || FORWARDER_BLOCK_P (f2->dest)
1637 || FORWARDER_BLOCK_P (b1->dest)
1638 || FORWARDER_BLOCK_P (b2->dest))
1639 return false;
1640
1641 if (f1->dest == f2->dest && b1->dest == b2->dest)
1642 reverse = false;
1643 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1644 reverse = true;
1645 else
1646 return false;
1647
1648 set1 = pc_set (BB_END (bb1));
1649 set2 = pc_set (BB_END (bb2));
1650 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1651 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1652 reverse = !reverse;
1653
1654 cond1 = XEXP (SET_SRC (set1), 0);
1655 cond2 = XEXP (SET_SRC (set2), 0);
1656 code1 = GET_CODE (cond1);
1657 if (reverse)
1658 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1659 else
1660 code2 = GET_CODE (cond2);
1661
1662 if (code2 == UNKNOWN)
1663 return false;
1664
1665 /* Verify codes and operands match. */
1666 match = ((code1 == code2
1667 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1668 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1669 || (code1 == swap_condition (code2)
1670 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1671 XEXP (cond2, 0))
1672 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1673 XEXP (cond2, 1))));
1674
1675 /* If we return true, we will join the blocks. Which means that
1676 we will only have one branch prediction bit to work with. Thus
1677 we require the existing branches to have probabilities that are
1678 roughly similar. */
1679 if (match
1680 && optimize_bb_for_speed_p (bb1)
1681 && optimize_bb_for_speed_p (bb2))
1682 {
1683 int prob2;
1684
1685 if (b1->dest == b2->dest)
1686 prob2 = b2->probability;
1687 else
1688 /* Do not use f2 probability as f2 may be forwarded. */
1689 prob2 = REG_BR_PROB_BASE - b2->probability;
1690
1691 /* Fail if the difference in probabilities is greater than 50%.
1692 This rules out two well-predicted branches with opposite
1693 outcomes. */
1694 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1695 {
1696 if (dump_file)
1697 fprintf (dump_file,
1698 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1699 bb1->index, bb2->index, b1->probability, prob2);
1700
1701 return false;
1702 }
1703 }
1704
1705 if (dump_file && match)
1706 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1707 bb1->index, bb2->index);
1708
1709 return match;
1710 }
1711
1712 /* Generic case - we are seeing a computed jump, table jump or trapping
1713 instruction. */
1714
1715 /* Check whether there are tablejumps in the end of BB1 and BB2.
1716 Return true if they are identical. */
1717 {
1718 rtx label1, label2;
1719 rtx_jump_table_data *table1, *table2;
1720
1721 if (tablejump_p (BB_END (bb1), &label1, &table1)
1722 && tablejump_p (BB_END (bb2), &label2, &table2)
1723 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1724 {
1725 /* The labels should never be the same rtx. If they really are same
1726 the jump tables are same too. So disable crossjumping of blocks BB1
1727 and BB2 because when deleting the common insns in the end of BB1
1728 by delete_basic_block () the jump table would be deleted too. */
1729 /* If LABEL2 is referenced in BB1->END do not do anything
1730 because we would loose information when replacing
1731 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1732 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1733 {
1734 /* Set IDENTICAL to true when the tables are identical. */
1735 bool identical = false;
1736 rtx p1, p2;
1737
1738 p1 = PATTERN (table1);
1739 p2 = PATTERN (table2);
1740 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1741 {
1742 identical = true;
1743 }
1744 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1745 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1746 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1747 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1748 {
1749 int i;
1750
1751 identical = true;
1752 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1753 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1754 identical = false;
1755 }
1756
1757 if (identical)
1758 {
1759 bool match;
1760
1761 /* Temporarily replace references to LABEL1 with LABEL2
1762 in BB1->END so that we could compare the instructions. */
1763 replace_label_in_insn (BB_END (bb1), label1, label2, false);
1764
1765 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1766 == dir_both);
1767 if (dump_file && match)
1768 fprintf (dump_file,
1769 "Tablejumps in bb %i and %i match.\n",
1770 bb1->index, bb2->index);
1771
1772 /* Set the original label in BB1->END because when deleting
1773 a block whose end is a tablejump, the tablejump referenced
1774 from the instruction is deleted too. */
1775 replace_label_in_insn (BB_END (bb1), label2, label1, false);
1776
1777 return match;
1778 }
1779 }
1780 return false;
1781 }
1782 }
1783
1784 /* Find the last non-debug non-note instruction in each bb, except
1785 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1786 handles that case specially. old_insns_match_p does not handle
1787 other types of instruction notes. */
1788 rtx_insn *last1 = BB_END (bb1);
1789 rtx_insn *last2 = BB_END (bb2);
1790 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1791 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1792 last1 = PREV_INSN (last1);
1793 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1794 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1795 last2 = PREV_INSN (last2);
1796 gcc_assert (last1 && last2);
1797
1798 /* First ensure that the instructions match. There may be many outgoing
1799 edges so this test is generally cheaper. */
1800 if (old_insns_match_p (mode, last1, last2) != dir_both)
1801 return false;
1802
1803 /* Search the outgoing edges, ensure that the counts do match, find possible
1804 fallthru and exception handling edges since these needs more
1805 validation. */
1806 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1807 return false;
1808
1809 bool nonfakeedges = false;
1810 FOR_EACH_EDGE (e1, ei, bb1->succs)
1811 {
1812 e2 = EDGE_SUCC (bb2, ei.index);
1813
1814 if ((e1->flags & EDGE_FAKE) == 0)
1815 nonfakeedges = true;
1816
1817 if (e1->flags & EDGE_EH)
1818 nehedges1++;
1819
1820 if (e2->flags & EDGE_EH)
1821 nehedges2++;
1822
1823 if (e1->flags & EDGE_FALLTHRU)
1824 fallthru1 = e1;
1825 if (e2->flags & EDGE_FALLTHRU)
1826 fallthru2 = e2;
1827 }
1828
1829 /* If number of edges of various types does not match, fail. */
1830 if (nehedges1 != nehedges2
1831 || (fallthru1 != 0) != (fallthru2 != 0))
1832 return false;
1833
1834 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1835 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1836 attempt to optimize, as the two basic blocks might have different
1837 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1838 traps there should be REG_ARG_SIZE notes, they could be missing
1839 for __builtin_unreachable () uses though. */
1840 if (!nonfakeedges
1841 && !ACCUMULATE_OUTGOING_ARGS
1842 && (!INSN_P (last1)
1843 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1844 return false;
1845
1846 /* fallthru edges must be forwarded to the same destination. */
1847 if (fallthru1)
1848 {
1849 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1850 ? single_succ (fallthru1->dest): fallthru1->dest);
1851 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1852 ? single_succ (fallthru2->dest): fallthru2->dest);
1853
1854 if (d1 != d2)
1855 return false;
1856 }
1857
1858 /* Ensure the same EH region. */
1859 {
1860 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1861 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1862
1863 if (!n1 && n2)
1864 return false;
1865
1866 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1867 return false;
1868 }
1869
1870 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1871 version of sequence abstraction. */
1872 FOR_EACH_EDGE (e1, ei, bb2->succs)
1873 {
1874 edge e2;
1875 edge_iterator ei;
1876 basic_block d1 = e1->dest;
1877
1878 if (FORWARDER_BLOCK_P (d1))
1879 d1 = EDGE_SUCC (d1, 0)->dest;
1880
1881 FOR_EACH_EDGE (e2, ei, bb1->succs)
1882 {
1883 basic_block d2 = e2->dest;
1884 if (FORWARDER_BLOCK_P (d2))
1885 d2 = EDGE_SUCC (d2, 0)->dest;
1886 if (d1 == d2)
1887 break;
1888 }
1889
1890 if (!e2)
1891 return false;
1892 }
1893
1894 return true;
1895 }
1896
1897 /* Returns true if BB basic block has a preserve label. */
1898
1899 static bool
1900 block_has_preserve_label (basic_block bb)
1901 {
1902 return (bb
1903 && block_label (bb)
1904 && LABEL_PRESERVE_P (block_label (bb)));
1905 }
1906
1907 /* E1 and E2 are edges with the same destination block. Search their
1908 predecessors for common code. If found, redirect control flow from
1909 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1910 or the other way around (dir_backward). DIR specifies the allowed
1911 replacement direction. */
1912
1913 static bool
1914 try_crossjump_to_edge (int mode, edge e1, edge e2,
1915 enum replace_direction dir)
1916 {
1917 int nmatch;
1918 basic_block src1 = e1->src, src2 = e2->src;
1919 basic_block redirect_to, redirect_from, to_remove;
1920 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1921 rtx_insn *newpos1, *newpos2;
1922 edge s;
1923 edge_iterator ei;
1924
1925 newpos1 = newpos2 = NULL;
1926
1927 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1928 to try this optimization.
1929
1930 Basic block partitioning may result in some jumps that appear to
1931 be optimizable (or blocks that appear to be mergeable), but which really
1932 must be left untouched (they are required to make it safely across
1933 partition boundaries). See the comments at the top of
1934 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1935
1936 if (crtl->has_bb_partition && reload_completed)
1937 return false;
1938
1939 /* Search backward through forwarder blocks. We don't need to worry
1940 about multiple entry or chained forwarders, as they will be optimized
1941 away. We do this to look past the unconditional jump following a
1942 conditional jump that is required due to the current CFG shape. */
1943 if (single_pred_p (src1)
1944 && FORWARDER_BLOCK_P (src1))
1945 e1 = single_pred_edge (src1), src1 = e1->src;
1946
1947 if (single_pred_p (src2)
1948 && FORWARDER_BLOCK_P (src2))
1949 e2 = single_pred_edge (src2), src2 = e2->src;
1950
1951 /* Nothing to do if we reach ENTRY, or a common source block. */
1952 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1953 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1954 return false;
1955 if (src1 == src2)
1956 return false;
1957
1958 /* Seeing more than 1 forwarder blocks would confuse us later... */
1959 if (FORWARDER_BLOCK_P (e1->dest)
1960 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1961 return false;
1962
1963 if (FORWARDER_BLOCK_P (e2->dest)
1964 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1965 return false;
1966
1967 /* Likewise with dead code (possibly newly created by the other optimizations
1968 of cfg_cleanup). */
1969 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1970 return false;
1971
1972 /* Look for the common insn sequence, part the first ... */
1973 if (!outgoing_edges_match (mode, src1, src2))
1974 return false;
1975
1976 /* ... and part the second. */
1977 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1978
1979 osrc1 = src1;
1980 osrc2 = src2;
1981 if (newpos1 != NULL_RTX)
1982 src1 = BLOCK_FOR_INSN (newpos1);
1983 if (newpos2 != NULL_RTX)
1984 src2 = BLOCK_FOR_INSN (newpos2);
1985
1986 if (dir == dir_backward)
1987 {
1988 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1989 SWAP (basic_block, osrc1, osrc2);
1990 SWAP (basic_block, src1, src2);
1991 SWAP (edge, e1, e2);
1992 SWAP (rtx_insn *, newpos1, newpos2);
1993 #undef SWAP
1994 }
1995
1996 /* Don't proceed with the crossjump unless we found a sufficient number
1997 of matching instructions or the 'from' block was totally matched
1998 (such that its predecessors will hopefully be redirected and the
1999 block removed). */
2000 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
2001 && (newpos1 != BB_HEAD (src1)))
2002 return false;
2003
2004 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
2005 if (block_has_preserve_label (e1->dest)
2006 && (e1->flags & EDGE_ABNORMAL))
2007 return false;
2008
2009 /* Here we know that the insns in the end of SRC1 which are common with SRC2
2010 will be deleted.
2011 If we have tablejumps in the end of SRC1 and SRC2
2012 they have been already compared for equivalence in outgoing_edges_match ()
2013 so replace the references to TABLE1 by references to TABLE2. */
2014 {
2015 rtx label1, label2;
2016 rtx_jump_table_data *table1, *table2;
2017
2018 if (tablejump_p (BB_END (osrc1), &label1, &table1)
2019 && tablejump_p (BB_END (osrc2), &label2, &table2)
2020 && label1 != label2)
2021 {
2022 rtx_insn *insn;
2023
2024 /* Replace references to LABEL1 with LABEL2. */
2025 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2026 {
2027 /* Do not replace the label in SRC1->END because when deleting
2028 a block whose end is a tablejump, the tablejump referenced
2029 from the instruction is deleted too. */
2030 if (insn != BB_END (osrc1))
2031 replace_label_in_insn (insn, label1, label2, true);
2032 }
2033 }
2034 }
2035
2036 /* Avoid splitting if possible. We must always split when SRC2 has
2037 EH predecessor edges, or we may end up with basic blocks with both
2038 normal and EH predecessor edges. */
2039 if (newpos2 == BB_HEAD (src2)
2040 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2041 redirect_to = src2;
2042 else
2043 {
2044 if (newpos2 == BB_HEAD (src2))
2045 {
2046 /* Skip possible basic block header. */
2047 if (LABEL_P (newpos2))
2048 newpos2 = NEXT_INSN (newpos2);
2049 while (DEBUG_INSN_P (newpos2))
2050 newpos2 = NEXT_INSN (newpos2);
2051 if (NOTE_P (newpos2))
2052 newpos2 = NEXT_INSN (newpos2);
2053 while (DEBUG_INSN_P (newpos2))
2054 newpos2 = NEXT_INSN (newpos2);
2055 }
2056
2057 if (dump_file)
2058 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2059 src2->index, nmatch);
2060 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2061 }
2062
2063 if (dump_file)
2064 fprintf (dump_file,
2065 "Cross jumping from bb %i to bb %i; %i common insns\n",
2066 src1->index, src2->index, nmatch);
2067
2068 /* We may have some registers visible through the block. */
2069 df_set_bb_dirty (redirect_to);
2070
2071 if (osrc2 == src2)
2072 redirect_edges_to = redirect_to;
2073 else
2074 redirect_edges_to = osrc2;
2075
2076 /* Recompute the frequencies and counts of outgoing edges. */
2077 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2078 {
2079 edge s2;
2080 edge_iterator ei;
2081 basic_block d = s->dest;
2082
2083 if (FORWARDER_BLOCK_P (d))
2084 d = single_succ (d);
2085
2086 FOR_EACH_EDGE (s2, ei, src1->succs)
2087 {
2088 basic_block d2 = s2->dest;
2089 if (FORWARDER_BLOCK_P (d2))
2090 d2 = single_succ (d2);
2091 if (d == d2)
2092 break;
2093 }
2094
2095 s->count += s2->count;
2096
2097 /* Take care to update possible forwarder blocks. We verified
2098 that there is no more than one in the chain, so we can't run
2099 into infinite loop. */
2100 if (FORWARDER_BLOCK_P (s->dest))
2101 {
2102 single_succ_edge (s->dest)->count += s2->count;
2103 s->dest->count += s2->count;
2104 s->dest->frequency += EDGE_FREQUENCY (s);
2105 }
2106
2107 if (FORWARDER_BLOCK_P (s2->dest))
2108 {
2109 single_succ_edge (s2->dest)->count -= s2->count;
2110 if (single_succ_edge (s2->dest)->count < 0)
2111 single_succ_edge (s2->dest)->count = 0;
2112 s2->dest->count -= s2->count;
2113 s2->dest->frequency -= EDGE_FREQUENCY (s);
2114 if (s2->dest->frequency < 0)
2115 s2->dest->frequency = 0;
2116 if (s2->dest->count < 0)
2117 s2->dest->count = 0;
2118 }
2119
2120 if (!redirect_edges_to->frequency && !src1->frequency)
2121 s->probability = (s->probability + s2->probability) / 2;
2122 else
2123 s->probability
2124 = ((s->probability * redirect_edges_to->frequency +
2125 s2->probability * src1->frequency)
2126 / (redirect_edges_to->frequency + src1->frequency));
2127 }
2128
2129 /* Adjust count and frequency for the block. An earlier jump
2130 threading pass may have left the profile in an inconsistent
2131 state (see update_bb_profile_for_threading) so we must be
2132 prepared for overflows. */
2133 tmp = redirect_to;
2134 do
2135 {
2136 tmp->count += src1->count;
2137 tmp->frequency += src1->frequency;
2138 if (tmp->frequency > BB_FREQ_MAX)
2139 tmp->frequency = BB_FREQ_MAX;
2140 if (tmp == redirect_edges_to)
2141 break;
2142 tmp = find_fallthru_edge (tmp->succs)->dest;
2143 }
2144 while (true);
2145 update_br_prob_note (redirect_edges_to);
2146
2147 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2148
2149 /* Skip possible basic block header. */
2150 if (LABEL_P (newpos1))
2151 newpos1 = NEXT_INSN (newpos1);
2152
2153 while (DEBUG_INSN_P (newpos1))
2154 newpos1 = NEXT_INSN (newpos1);
2155
2156 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2157 newpos1 = NEXT_INSN (newpos1);
2158
2159 while (DEBUG_INSN_P (newpos1))
2160 newpos1 = NEXT_INSN (newpos1);
2161
2162 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2163 to_remove = single_succ (redirect_from);
2164
2165 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2166 delete_basic_block (to_remove);
2167
2168 update_forwarder_flag (redirect_from);
2169 if (redirect_to != src2)
2170 update_forwarder_flag (src2);
2171
2172 return true;
2173 }
2174
2175 /* Search the predecessors of BB for common insn sequences. When found,
2176 share code between them by redirecting control flow. Return true if
2177 any changes made. */
2178
2179 static bool
2180 try_crossjump_bb (int mode, basic_block bb)
2181 {
2182 edge e, e2, fallthru;
2183 bool changed;
2184 unsigned max, ix, ix2;
2185
2186 /* Nothing to do if there is not at least two incoming edges. */
2187 if (EDGE_COUNT (bb->preds) < 2)
2188 return false;
2189
2190 /* Don't crossjump if this block ends in a computed jump,
2191 unless we are optimizing for size. */
2192 if (optimize_bb_for_size_p (bb)
2193 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2194 && computed_jump_p (BB_END (bb)))
2195 return false;
2196
2197 /* If we are partitioning hot/cold basic blocks, we don't want to
2198 mess up unconditional or indirect jumps that cross between hot
2199 and cold sections.
2200
2201 Basic block partitioning may result in some jumps that appear to
2202 be optimizable (or blocks that appear to be mergeable), but which really
2203 must be left untouched (they are required to make it safely across
2204 partition boundaries). See the comments at the top of
2205 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2206
2207 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2208 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2209 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2210 return false;
2211
2212 /* It is always cheapest to redirect a block that ends in a branch to
2213 a block that falls through into BB, as that adds no branches to the
2214 program. We'll try that combination first. */
2215 fallthru = NULL;
2216 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2217
2218 if (EDGE_COUNT (bb->preds) > max)
2219 return false;
2220
2221 fallthru = find_fallthru_edge (bb->preds);
2222
2223 changed = false;
2224 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2225 {
2226 e = EDGE_PRED (bb, ix);
2227 ix++;
2228
2229 /* As noted above, first try with the fallthru predecessor (or, a
2230 fallthru predecessor if we are in cfglayout mode). */
2231 if (fallthru)
2232 {
2233 /* Don't combine the fallthru edge into anything else.
2234 If there is a match, we'll do it the other way around. */
2235 if (e == fallthru)
2236 continue;
2237 /* If nothing changed since the last attempt, there is nothing
2238 we can do. */
2239 if (!first_pass
2240 && !((e->src->flags & BB_MODIFIED)
2241 || (fallthru->src->flags & BB_MODIFIED)))
2242 continue;
2243
2244 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2245 {
2246 changed = true;
2247 ix = 0;
2248 continue;
2249 }
2250 }
2251
2252 /* Non-obvious work limiting check: Recognize that we're going
2253 to call try_crossjump_bb on every basic block. So if we have
2254 two blocks with lots of outgoing edges (a switch) and they
2255 share lots of common destinations, then we would do the
2256 cross-jump check once for each common destination.
2257
2258 Now, if the blocks actually are cross-jump candidates, then
2259 all of their destinations will be shared. Which means that
2260 we only need check them for cross-jump candidacy once. We
2261 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2262 choosing to do the check from the block for which the edge
2263 in question is the first successor of A. */
2264 if (EDGE_SUCC (e->src, 0) != e)
2265 continue;
2266
2267 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2268 {
2269 e2 = EDGE_PRED (bb, ix2);
2270
2271 if (e2 == e)
2272 continue;
2273
2274 /* We've already checked the fallthru edge above. */
2275 if (e2 == fallthru)
2276 continue;
2277
2278 /* The "first successor" check above only prevents multiple
2279 checks of crossjump(A,B). In order to prevent redundant
2280 checks of crossjump(B,A), require that A be the block
2281 with the lowest index. */
2282 if (e->src->index > e2->src->index)
2283 continue;
2284
2285 /* If nothing changed since the last attempt, there is nothing
2286 we can do. */
2287 if (!first_pass
2288 && !((e->src->flags & BB_MODIFIED)
2289 || (e2->src->flags & BB_MODIFIED)))
2290 continue;
2291
2292 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2293 direction. */
2294 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2295 {
2296 changed = true;
2297 ix = 0;
2298 break;
2299 }
2300 }
2301 }
2302
2303 if (changed)
2304 crossjumps_occured = true;
2305
2306 return changed;
2307 }
2308
2309 /* Search the successors of BB for common insn sequences. When found,
2310 share code between them by moving it across the basic block
2311 boundary. Return true if any changes made. */
2312
2313 static bool
2314 try_head_merge_bb (basic_block bb)
2315 {
2316 basic_block final_dest_bb = NULL;
2317 int max_match = INT_MAX;
2318 edge e0;
2319 rtx_insn **headptr, **currptr, **nextptr;
2320 bool changed, moveall;
2321 unsigned ix;
2322 rtx_insn *e0_last_head;
2323 rtx cond;
2324 rtx_insn *move_before;
2325 unsigned nedges = EDGE_COUNT (bb->succs);
2326 rtx_insn *jump = BB_END (bb);
2327 regset live, live_union;
2328
2329 /* Nothing to do if there is not at least two outgoing edges. */
2330 if (nedges < 2)
2331 return false;
2332
2333 /* Don't crossjump if this block ends in a computed jump,
2334 unless we are optimizing for size. */
2335 if (optimize_bb_for_size_p (bb)
2336 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2337 && computed_jump_p (BB_END (bb)))
2338 return false;
2339
2340 cond = get_condition (jump, &move_before, true, false);
2341 if (cond == NULL_RTX)
2342 {
2343 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2344 move_before = prev_nonnote_nondebug_insn (jump);
2345 else
2346 move_before = jump;
2347 }
2348
2349 for (ix = 0; ix < nedges; ix++)
2350 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2351 return false;
2352
2353 for (ix = 0; ix < nedges; ix++)
2354 {
2355 edge e = EDGE_SUCC (bb, ix);
2356 basic_block other_bb = e->dest;
2357
2358 if (df_get_bb_dirty (other_bb))
2359 {
2360 block_was_dirty = true;
2361 return false;
2362 }
2363
2364 if (e->flags & EDGE_ABNORMAL)
2365 return false;
2366
2367 /* Normally, all destination blocks must only be reachable from this
2368 block, i.e. they must have one incoming edge.
2369
2370 There is one special case we can handle, that of multiple consecutive
2371 jumps where the first jumps to one of the targets of the second jump.
2372 This happens frequently in switch statements for default labels.
2373 The structure is as follows:
2374 FINAL_DEST_BB
2375 ....
2376 if (cond) jump A;
2377 fall through
2378 BB
2379 jump with targets A, B, C, D...
2380 A
2381 has two incoming edges, from FINAL_DEST_BB and BB
2382
2383 In this case, we can try to move the insns through BB and into
2384 FINAL_DEST_BB. */
2385 if (EDGE_COUNT (other_bb->preds) != 1)
2386 {
2387 edge incoming_edge, incoming_bb_other_edge;
2388 edge_iterator ei;
2389
2390 if (final_dest_bb != NULL
2391 || EDGE_COUNT (other_bb->preds) != 2)
2392 return false;
2393
2394 /* We must be able to move the insns across the whole block. */
2395 move_before = BB_HEAD (bb);
2396 while (!NONDEBUG_INSN_P (move_before))
2397 move_before = NEXT_INSN (move_before);
2398
2399 if (EDGE_COUNT (bb->preds) != 1)
2400 return false;
2401 incoming_edge = EDGE_PRED (bb, 0);
2402 final_dest_bb = incoming_edge->src;
2403 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2404 return false;
2405 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2406 if (incoming_bb_other_edge != incoming_edge)
2407 break;
2408 if (incoming_bb_other_edge->dest != other_bb)
2409 return false;
2410 }
2411 }
2412
2413 e0 = EDGE_SUCC (bb, 0);
2414 e0_last_head = NULL;
2415 changed = false;
2416
2417 for (ix = 1; ix < nedges; ix++)
2418 {
2419 edge e = EDGE_SUCC (bb, ix);
2420 rtx_insn *e0_last, *e_last;
2421 int nmatch;
2422
2423 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2424 &e0_last, &e_last, 0);
2425 if (nmatch == 0)
2426 return false;
2427
2428 if (nmatch < max_match)
2429 {
2430 max_match = nmatch;
2431 e0_last_head = e0_last;
2432 }
2433 }
2434
2435 /* If we matched an entire block, we probably have to avoid moving the
2436 last insn. */
2437 if (max_match > 0
2438 && e0_last_head == BB_END (e0->dest)
2439 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2440 || control_flow_insn_p (e0_last_head)))
2441 {
2442 max_match--;
2443 if (max_match == 0)
2444 return false;
2445 do
2446 e0_last_head = prev_real_insn (e0_last_head);
2447 while (DEBUG_INSN_P (e0_last_head));
2448 }
2449
2450 if (max_match == 0)
2451 return false;
2452
2453 /* We must find a union of the live registers at each of the end points. */
2454 live = BITMAP_ALLOC (NULL);
2455 live_union = BITMAP_ALLOC (NULL);
2456
2457 currptr = XNEWVEC (rtx_insn *, nedges);
2458 headptr = XNEWVEC (rtx_insn *, nedges);
2459 nextptr = XNEWVEC (rtx_insn *, nedges);
2460
2461 for (ix = 0; ix < nedges; ix++)
2462 {
2463 int j;
2464 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2465 rtx_insn *head = BB_HEAD (merge_bb);
2466
2467 while (!NONDEBUG_INSN_P (head))
2468 head = NEXT_INSN (head);
2469 headptr[ix] = head;
2470 currptr[ix] = head;
2471
2472 /* Compute the end point and live information */
2473 for (j = 1; j < max_match; j++)
2474 do
2475 head = NEXT_INSN (head);
2476 while (!NONDEBUG_INSN_P (head));
2477 simulate_backwards_to_point (merge_bb, live, head);
2478 IOR_REG_SET (live_union, live);
2479 }
2480
2481 /* If we're moving across two blocks, verify the validity of the
2482 first move, then adjust the target and let the loop below deal
2483 with the final move. */
2484 if (final_dest_bb != NULL)
2485 {
2486 rtx_insn *move_upto;
2487
2488 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2489 jump, e0->dest, live_union,
2490 NULL, &move_upto);
2491 if (!moveall)
2492 {
2493 if (move_upto == NULL_RTX)
2494 goto out;
2495
2496 while (e0_last_head != move_upto)
2497 {
2498 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2499 live_union);
2500 e0_last_head = PREV_INSN (e0_last_head);
2501 }
2502 }
2503 if (e0_last_head == NULL_RTX)
2504 goto out;
2505
2506 jump = BB_END (final_dest_bb);
2507 cond = get_condition (jump, &move_before, true, false);
2508 if (cond == NULL_RTX)
2509 {
2510 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2511 move_before = prev_nonnote_nondebug_insn (jump);
2512 else
2513 move_before = jump;
2514 }
2515 }
2516
2517 do
2518 {
2519 rtx_insn *move_upto;
2520 moveall = can_move_insns_across (currptr[0], e0_last_head,
2521 move_before, jump, e0->dest, live_union,
2522 NULL, &move_upto);
2523 if (!moveall && move_upto == NULL_RTX)
2524 {
2525 if (jump == move_before)
2526 break;
2527
2528 /* Try again, using a different insertion point. */
2529 move_before = jump;
2530
2531 /* Don't try moving before a cc0 user, as that may invalidate
2532 the cc0. */
2533 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2534 break;
2535
2536 continue;
2537 }
2538
2539 if (final_dest_bb && !moveall)
2540 /* We haven't checked whether a partial move would be OK for the first
2541 move, so we have to fail this case. */
2542 break;
2543
2544 changed = true;
2545 for (;;)
2546 {
2547 if (currptr[0] == move_upto)
2548 break;
2549 for (ix = 0; ix < nedges; ix++)
2550 {
2551 rtx_insn *curr = currptr[ix];
2552 do
2553 curr = NEXT_INSN (curr);
2554 while (!NONDEBUG_INSN_P (curr));
2555 currptr[ix] = curr;
2556 }
2557 }
2558
2559 /* If we can't currently move all of the identical insns, remember
2560 each insn after the range that we'll merge. */
2561 if (!moveall)
2562 for (ix = 0; ix < nedges; ix++)
2563 {
2564 rtx_insn *curr = currptr[ix];
2565 do
2566 curr = NEXT_INSN (curr);
2567 while (!NONDEBUG_INSN_P (curr));
2568 nextptr[ix] = curr;
2569 }
2570
2571 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2572 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2573 if (final_dest_bb != NULL)
2574 df_set_bb_dirty (final_dest_bb);
2575 df_set_bb_dirty (bb);
2576 for (ix = 1; ix < nedges; ix++)
2577 {
2578 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2579 delete_insn_chain (headptr[ix], currptr[ix], false);
2580 }
2581 if (!moveall)
2582 {
2583 if (jump == move_before)
2584 break;
2585
2586 /* For the unmerged insns, try a different insertion point. */
2587 move_before = jump;
2588
2589 /* Don't try moving before a cc0 user, as that may invalidate
2590 the cc0. */
2591 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2592 break;
2593
2594 for (ix = 0; ix < nedges; ix++)
2595 currptr[ix] = headptr[ix] = nextptr[ix];
2596 }
2597 }
2598 while (!moveall);
2599
2600 out:
2601 free (currptr);
2602 free (headptr);
2603 free (nextptr);
2604
2605 crossjumps_occured |= changed;
2606
2607 return changed;
2608 }
2609
2610 /* Return true if BB contains just bb note, or bb note followed
2611 by only DEBUG_INSNs. */
2612
2613 static bool
2614 trivially_empty_bb_p (basic_block bb)
2615 {
2616 rtx_insn *insn = BB_END (bb);
2617
2618 while (1)
2619 {
2620 if (insn == BB_HEAD (bb))
2621 return true;
2622 if (!DEBUG_INSN_P (insn))
2623 return false;
2624 insn = PREV_INSN (insn);
2625 }
2626 }
2627
2628 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2629 instructions etc. Return nonzero if changes were made. */
2630
2631 static bool
2632 try_optimize_cfg (int mode)
2633 {
2634 bool changed_overall = false;
2635 bool changed;
2636 int iterations = 0;
2637 basic_block bb, b, next;
2638
2639 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2640 clear_bb_flags ();
2641
2642 crossjumps_occured = false;
2643
2644 FOR_EACH_BB_FN (bb, cfun)
2645 update_forwarder_flag (bb);
2646
2647 if (! targetm.cannot_modify_jumps_p ())
2648 {
2649 first_pass = true;
2650 /* Attempt to merge blocks as made possible by edge removal. If
2651 a block has only one successor, and the successor has only
2652 one predecessor, they may be combined. */
2653 do
2654 {
2655 block_was_dirty = false;
2656 changed = false;
2657 iterations++;
2658
2659 if (dump_file)
2660 fprintf (dump_file,
2661 "\n\ntry_optimize_cfg iteration %i\n\n",
2662 iterations);
2663
2664 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2665 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2666 {
2667 basic_block c;
2668 edge s;
2669 bool changed_here = false;
2670
2671 /* Delete trivially dead basic blocks. This is either
2672 blocks with no predecessors, or empty blocks with no
2673 successors. However if the empty block with no
2674 successors is the successor of the ENTRY_BLOCK, it is
2675 kept. This ensures that the ENTRY_BLOCK will have a
2676 successor which is a precondition for many RTL
2677 passes. Empty blocks may result from expanding
2678 __builtin_unreachable (). */
2679 if (EDGE_COUNT (b->preds) == 0
2680 || (EDGE_COUNT (b->succs) == 0
2681 && trivially_empty_bb_p (b)
2682 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2683 != b))
2684 {
2685 c = b->prev_bb;
2686 if (EDGE_COUNT (b->preds) > 0)
2687 {
2688 edge e;
2689 edge_iterator ei;
2690
2691 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2692 {
2693 if (BB_FOOTER (b)
2694 && BARRIER_P (BB_FOOTER (b)))
2695 FOR_EACH_EDGE (e, ei, b->preds)
2696 if ((e->flags & EDGE_FALLTHRU)
2697 && BB_FOOTER (e->src) == NULL)
2698 {
2699 if (BB_FOOTER (b))
2700 {
2701 BB_FOOTER (e->src) = BB_FOOTER (b);
2702 BB_FOOTER (b) = NULL;
2703 }
2704 else
2705 {
2706 start_sequence ();
2707 BB_FOOTER (e->src) = emit_barrier ();
2708 end_sequence ();
2709 }
2710 }
2711 }
2712 else
2713 {
2714 rtx_insn *last = get_last_bb_insn (b);
2715 if (last && BARRIER_P (last))
2716 FOR_EACH_EDGE (e, ei, b->preds)
2717 if ((e->flags & EDGE_FALLTHRU))
2718 emit_barrier_after (BB_END (e->src));
2719 }
2720 }
2721 delete_basic_block (b);
2722 changed = true;
2723 /* Avoid trying to remove the exit block. */
2724 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2725 continue;
2726 }
2727
2728 /* Remove code labels no longer used. */
2729 if (single_pred_p (b)
2730 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2731 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2732 && LABEL_P (BB_HEAD (b))
2733 && !LABEL_PRESERVE_P (BB_HEAD (b))
2734 /* If the previous block ends with a branch to this
2735 block, we can't delete the label. Normally this
2736 is a condjump that is yet to be simplified, but
2737 if CASE_DROPS_THRU, this can be a tablejump with
2738 some element going to the same place as the
2739 default (fallthru). */
2740 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2741 || !JUMP_P (BB_END (single_pred (b)))
2742 || ! label_is_jump_target_p (BB_HEAD (b),
2743 BB_END (single_pred (b)))))
2744 {
2745 delete_insn (BB_HEAD (b));
2746 if (dump_file)
2747 fprintf (dump_file, "Deleted label in block %i.\n",
2748 b->index);
2749 }
2750
2751 /* If we fall through an empty block, we can remove it. */
2752 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2753 && single_pred_p (b)
2754 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2755 && !LABEL_P (BB_HEAD (b))
2756 && FORWARDER_BLOCK_P (b)
2757 /* Note that forwarder_block_p true ensures that
2758 there is a successor for this block. */
2759 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2760 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2761 {
2762 if (dump_file)
2763 fprintf (dump_file,
2764 "Deleting fallthru block %i.\n",
2765 b->index);
2766
2767 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2768 ? b->next_bb : b->prev_bb);
2769 redirect_edge_succ_nodup (single_pred_edge (b),
2770 single_succ (b));
2771 delete_basic_block (b);
2772 changed = true;
2773 b = c;
2774 continue;
2775 }
2776
2777 /* Merge B with its single successor, if any. */
2778 if (single_succ_p (b)
2779 && (s = single_succ_edge (b))
2780 && !(s->flags & EDGE_COMPLEX)
2781 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2782 && single_pred_p (c)
2783 && b != c)
2784 {
2785 /* When not in cfg_layout mode use code aware of reordering
2786 INSN. This code possibly creates new basic blocks so it
2787 does not fit merge_blocks interface and is kept here in
2788 hope that it will become useless once more of compiler
2789 is transformed to use cfg_layout mode. */
2790
2791 if ((mode & CLEANUP_CFGLAYOUT)
2792 && can_merge_blocks_p (b, c))
2793 {
2794 merge_blocks (b, c);
2795 update_forwarder_flag (b);
2796 changed_here = true;
2797 }
2798 else if (!(mode & CLEANUP_CFGLAYOUT)
2799 /* If the jump insn has side effects,
2800 we can't kill the edge. */
2801 && (!JUMP_P (BB_END (b))
2802 || (reload_completed
2803 ? simplejump_p (BB_END (b))
2804 : (onlyjump_p (BB_END (b))
2805 && !tablejump_p (BB_END (b),
2806 NULL, NULL))))
2807 && (next = merge_blocks_move (s, b, c, mode)))
2808 {
2809 b = next;
2810 changed_here = true;
2811 }
2812 }
2813
2814 /* Simplify branch over branch. */
2815 if ((mode & CLEANUP_EXPENSIVE)
2816 && !(mode & CLEANUP_CFGLAYOUT)
2817 && try_simplify_condjump (b))
2818 changed_here = true;
2819
2820 /* If B has a single outgoing edge, but uses a
2821 non-trivial jump instruction without side-effects, we
2822 can either delete the jump entirely, or replace it
2823 with a simple unconditional jump. */
2824 if (single_succ_p (b)
2825 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2826 && onlyjump_p (BB_END (b))
2827 && !CROSSING_JUMP_P (BB_END (b))
2828 && try_redirect_by_replacing_jump (single_succ_edge (b),
2829 single_succ (b),
2830 (mode & CLEANUP_CFGLAYOUT) != 0))
2831 {
2832 update_forwarder_flag (b);
2833 changed_here = true;
2834 }
2835
2836 /* Simplify branch to branch. */
2837 if (try_forward_edges (mode, b))
2838 {
2839 update_forwarder_flag (b);
2840 changed_here = true;
2841 }
2842
2843 /* Look for shared code between blocks. */
2844 if ((mode & CLEANUP_CROSSJUMP)
2845 && try_crossjump_bb (mode, b))
2846 changed_here = true;
2847
2848 if ((mode & CLEANUP_CROSSJUMP)
2849 /* This can lengthen register lifetimes. Do it only after
2850 reload. */
2851 && reload_completed
2852 && try_head_merge_bb (b))
2853 changed_here = true;
2854
2855 /* Don't get confused by the index shift caused by
2856 deleting blocks. */
2857 if (!changed_here)
2858 b = b->next_bb;
2859 else
2860 changed = true;
2861 }
2862
2863 if ((mode & CLEANUP_CROSSJUMP)
2864 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
2865 changed = true;
2866
2867 if (block_was_dirty)
2868 {
2869 /* This should only be set by head-merging. */
2870 gcc_assert (mode & CLEANUP_CROSSJUMP);
2871 df_analyze ();
2872 }
2873
2874 if (changed)
2875 {
2876 /* Edge forwarding in particular can cause hot blocks previously
2877 reached by both hot and cold blocks to become dominated only
2878 by cold blocks. This will cause the verification below to fail,
2879 and lead to now cold code in the hot section. This is not easy
2880 to detect and fix during edge forwarding, and in some cases
2881 is only visible after newly unreachable blocks are deleted,
2882 which will be done in fixup_partitions. */
2883 fixup_partitions ();
2884
2885 #ifdef ENABLE_CHECKING
2886 verify_flow_info ();
2887 #endif
2888 }
2889
2890 changed_overall |= changed;
2891 first_pass = false;
2892 }
2893 while (changed);
2894 }
2895
2896 FOR_ALL_BB_FN (b, cfun)
2897 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2898
2899 return changed_overall;
2900 }
2901 \f
2902 /* Delete all unreachable basic blocks. */
2903
2904 bool
2905 delete_unreachable_blocks (void)
2906 {
2907 bool changed = false;
2908 basic_block b, prev_bb;
2909
2910 find_unreachable_blocks ();
2911
2912 /* When we're in GIMPLE mode and there may be debug insns, we should
2913 delete blocks in reverse dominator order, so as to get a chance
2914 to substitute all released DEFs into debug stmts. If we don't
2915 have dominators information, walking blocks backward gets us a
2916 better chance of retaining most debug information than
2917 otherwise. */
2918 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2919 && dom_info_available_p (CDI_DOMINATORS))
2920 {
2921 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2922 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2923 {
2924 prev_bb = b->prev_bb;
2925
2926 if (!(b->flags & BB_REACHABLE))
2927 {
2928 /* Speed up the removal of blocks that don't dominate
2929 others. Walking backwards, this should be the common
2930 case. */
2931 if (!first_dom_son (CDI_DOMINATORS, b))
2932 delete_basic_block (b);
2933 else
2934 {
2935 vec<basic_block> h
2936 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2937
2938 while (h.length ())
2939 {
2940 b = h.pop ();
2941
2942 prev_bb = b->prev_bb;
2943
2944 gcc_assert (!(b->flags & BB_REACHABLE));
2945
2946 delete_basic_block (b);
2947 }
2948
2949 h.release ();
2950 }
2951
2952 changed = true;
2953 }
2954 }
2955 }
2956 else
2957 {
2958 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2959 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2960 {
2961 prev_bb = b->prev_bb;
2962
2963 if (!(b->flags & BB_REACHABLE))
2964 {
2965 delete_basic_block (b);
2966 changed = true;
2967 }
2968 }
2969 }
2970
2971 if (changed)
2972 tidy_fallthru_edges ();
2973 return changed;
2974 }
2975
2976 /* Delete any jump tables never referenced. We can't delete them at the
2977 time of removing tablejump insn as they are referenced by the preceding
2978 insns computing the destination, so we delay deleting and garbagecollect
2979 them once life information is computed. */
2980 void
2981 delete_dead_jumptables (void)
2982 {
2983 basic_block bb;
2984
2985 /* A dead jump table does not belong to any basic block. Scan insns
2986 between two adjacent basic blocks. */
2987 FOR_EACH_BB_FN (bb, cfun)
2988 {
2989 rtx_insn *insn, *next;
2990
2991 for (insn = NEXT_INSN (BB_END (bb));
2992 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
2993 insn = next)
2994 {
2995 next = NEXT_INSN (insn);
2996 if (LABEL_P (insn)
2997 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
2998 && JUMP_TABLE_DATA_P (next))
2999 {
3000 rtx_insn *label = insn, *jump = next;
3001
3002 if (dump_file)
3003 fprintf (dump_file, "Dead jumptable %i removed\n",
3004 INSN_UID (insn));
3005
3006 next = NEXT_INSN (next);
3007 delete_insn (jump);
3008 delete_insn (label);
3009 }
3010 }
3011 }
3012 }
3013
3014 \f
3015 /* Tidy the CFG by deleting unreachable code and whatnot. */
3016
3017 bool
3018 cleanup_cfg (int mode)
3019 {
3020 bool changed = false;
3021
3022 /* Set the cfglayout mode flag here. We could update all the callers
3023 but that is just inconvenient, especially given that we eventually
3024 want to have cfglayout mode as the default. */
3025 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3026 mode |= CLEANUP_CFGLAYOUT;
3027
3028 timevar_push (TV_CLEANUP_CFG);
3029 if (delete_unreachable_blocks ())
3030 {
3031 changed = true;
3032 /* We've possibly created trivially dead code. Cleanup it right
3033 now to introduce more opportunities for try_optimize_cfg. */
3034 if (!(mode & (CLEANUP_NO_INSN_DEL))
3035 && !reload_completed)
3036 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3037 }
3038
3039 compact_blocks ();
3040
3041 /* To tail-merge blocks ending in the same noreturn function (e.g.
3042 a call to abort) we have to insert fake edges to exit. Do this
3043 here once. The fake edges do not interfere with any other CFG
3044 cleanups. */
3045 if (mode & CLEANUP_CROSSJUMP)
3046 add_noreturn_fake_exit_edges ();
3047
3048 if (!dbg_cnt (cfg_cleanup))
3049 return changed;
3050
3051 while (try_optimize_cfg (mode))
3052 {
3053 delete_unreachable_blocks (), changed = true;
3054 if (!(mode & CLEANUP_NO_INSN_DEL))
3055 {
3056 /* Try to remove some trivially dead insns when doing an expensive
3057 cleanup. But delete_trivially_dead_insns doesn't work after
3058 reload (it only handles pseudos) and run_fast_dce is too costly
3059 to run in every iteration.
3060
3061 For effective cross jumping, we really want to run a fast DCE to
3062 clean up any dead conditions, or they get in the way of performing
3063 useful tail merges.
3064
3065 Other transformations in cleanup_cfg are not so sensitive to dead
3066 code, so delete_trivially_dead_insns or even doing nothing at all
3067 is good enough. */
3068 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3069 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3070 break;
3071 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
3072 run_fast_dce ();
3073 }
3074 else
3075 break;
3076 }
3077
3078 if (mode & CLEANUP_CROSSJUMP)
3079 remove_fake_exit_edges ();
3080
3081 /* Don't call delete_dead_jumptables in cfglayout mode, because
3082 that function assumes that jump tables are in the insns stream.
3083 But we also don't _have_ to delete dead jumptables in cfglayout
3084 mode because we shouldn't even be looking at things that are
3085 not in a basic block. Dead jumptables are cleaned up when
3086 going out of cfglayout mode. */
3087 if (!(mode & CLEANUP_CFGLAYOUT))
3088 delete_dead_jumptables ();
3089
3090 /* ??? We probably do this way too often. */
3091 if (current_loops
3092 && (changed
3093 || (mode & CLEANUP_CFG_CHANGED)))
3094 {
3095 timevar_push (TV_REPAIR_LOOPS);
3096 /* The above doesn't preserve dominance info if available. */
3097 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3098 calculate_dominance_info (CDI_DOMINATORS);
3099 fix_loop_structure (NULL);
3100 free_dominance_info (CDI_DOMINATORS);
3101 timevar_pop (TV_REPAIR_LOOPS);
3102 }
3103
3104 timevar_pop (TV_CLEANUP_CFG);
3105
3106 return changed;
3107 }
3108 \f
3109 namespace {
3110
3111 const pass_data pass_data_jump =
3112 {
3113 RTL_PASS, /* type */
3114 "jump", /* name */
3115 OPTGROUP_NONE, /* optinfo_flags */
3116 TV_JUMP, /* tv_id */
3117 0, /* properties_required */
3118 0, /* properties_provided */
3119 0, /* properties_destroyed */
3120 0, /* todo_flags_start */
3121 0, /* todo_flags_finish */
3122 };
3123
3124 class pass_jump : public rtl_opt_pass
3125 {
3126 public:
3127 pass_jump (gcc::context *ctxt)
3128 : rtl_opt_pass (pass_data_jump, ctxt)
3129 {}
3130
3131 /* opt_pass methods: */
3132 virtual unsigned int execute (function *);
3133
3134 }; // class pass_jump
3135
3136 unsigned int
3137 pass_jump::execute (function *)
3138 {
3139 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3140 if (dump_file)
3141 dump_flow_info (dump_file, dump_flags);
3142 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3143 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3144 return 0;
3145 }
3146
3147 } // anon namespace
3148
3149 rtl_opt_pass *
3150 make_pass_jump (gcc::context *ctxt)
3151 {
3152 return new pass_jump (ctxt);
3153 }
3154 \f
3155 namespace {
3156
3157 const pass_data pass_data_jump2 =
3158 {
3159 RTL_PASS, /* type */
3160 "jump2", /* name */
3161 OPTGROUP_NONE, /* optinfo_flags */
3162 TV_JUMP, /* tv_id */
3163 0, /* properties_required */
3164 0, /* properties_provided */
3165 0, /* properties_destroyed */
3166 0, /* todo_flags_start */
3167 0, /* todo_flags_finish */
3168 };
3169
3170 class pass_jump2 : public rtl_opt_pass
3171 {
3172 public:
3173 pass_jump2 (gcc::context *ctxt)
3174 : rtl_opt_pass (pass_data_jump2, ctxt)
3175 {}
3176
3177 /* opt_pass methods: */
3178 virtual unsigned int execute (function *)
3179 {
3180 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3181 return 0;
3182 }
3183
3184 }; // class pass_jump2
3185
3186 } // anon namespace
3187
3188 rtl_opt_pass *
3189 make_pass_jump2 (gcc::context *ctxt)
3190 {
3191 return new pass_jump2 (ctxt);
3192 }