]> git.ipfire.org Git - thirdparty/gcc.git/blob - gcc/cfgcleanup.c
Merge in trunk.
[thirdparty/gcc.git] / gcc / cfgcleanup.c
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2013 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
10
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
19
20 /* This file contains optimizer of the control flow. The main entry point is
21 cleanup_cfg. Following optimizations are performed:
22
23 - Unreachable blocks removal
24 - Edge forwarding (edge to the forwarder block is forwarded to its
25 successor. Simplification of the branch instruction is performed by
26 underlying infrastructure so branch can be converted to simplejump or
27 eliminated).
28 - Cross jumping (tail merging)
29 - Conditional jump-around-simplejump simplification
30 - Basic block merging. */
31
32 #include "config.h"
33 #include "system.h"
34 #include "coretypes.h"
35 #include "tm.h"
36 #include "rtl.h"
37 #include "hard-reg-set.h"
38 #include "regs.h"
39 #include "insn-config.h"
40 #include "flags.h"
41 #include "recog.h"
42 #include "diagnostic-core.h"
43 #include "cselib.h"
44 #include "params.h"
45 #include "tm_p.h"
46 #include "target.h"
47 #include "function.h" /* For inline functions in emit-rtl.h they need crtl. */
48 #include "emit-rtl.h"
49 #include "tree-pass.h"
50 #include "cfgloop.h"
51 #include "expr.h"
52 #include "df.h"
53 #include "dce.h"
54 #include "dbgcnt.h"
55
56 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
57
58 /* Set to true when we are running first pass of try_optimize_cfg loop. */
59 static bool first_pass;
60
61 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
62 static bool crossjumps_occured;
63
64 /* Set to true if we couldn't run an optimization due to stale liveness
65 information; we should run df_analyze to enable more opportunities. */
66 static bool block_was_dirty;
67
68 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
69 static bool try_crossjump_bb (int, basic_block);
70 static bool outgoing_edges_match (int, basic_block, basic_block);
71 static enum replace_direction old_insns_match_p (int, rtx, rtx);
72
73 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
74 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
75 static bool try_optimize_cfg (int);
76 static bool try_simplify_condjump (basic_block);
77 static bool try_forward_edges (int, basic_block);
78 static edge thread_jump (edge, basic_block);
79 static bool mark_effect (rtx, bitmap);
80 static void notice_new_block (basic_block);
81 static void update_forwarder_flag (basic_block);
82 static int mentions_nonequal_regs (rtx *, void *);
83 static void merge_memattrs (rtx, rtx);
84 \f
85 /* Set flags for newly created block. */
86
87 static void
88 notice_new_block (basic_block bb)
89 {
90 if (!bb)
91 return;
92
93 if (forwarder_block_p (bb))
94 bb->flags |= BB_FORWARDER_BLOCK;
95 }
96
97 /* Recompute forwarder flag after block has been modified. */
98
99 static void
100 update_forwarder_flag (basic_block bb)
101 {
102 if (forwarder_block_p (bb))
103 bb->flags |= BB_FORWARDER_BLOCK;
104 else
105 bb->flags &= ~BB_FORWARDER_BLOCK;
106 }
107 \f
108 /* Simplify a conditional jump around an unconditional jump.
109 Return true if something changed. */
110
111 static bool
112 try_simplify_condjump (basic_block cbranch_block)
113 {
114 basic_block jump_block, jump_dest_block, cbranch_dest_block;
115 edge cbranch_jump_edge, cbranch_fallthru_edge;
116 rtx cbranch_insn;
117
118 /* Verify that there are exactly two successors. */
119 if (EDGE_COUNT (cbranch_block->succs) != 2)
120 return false;
121
122 /* Verify that we've got a normal conditional branch at the end
123 of the block. */
124 cbranch_insn = BB_END (cbranch_block);
125 if (!any_condjump_p (cbranch_insn))
126 return false;
127
128 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
129 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
130
131 /* The next block must not have multiple predecessors, must not
132 be the last block in the function, and must contain just the
133 unconditional jump. */
134 jump_block = cbranch_fallthru_edge->dest;
135 if (!single_pred_p (jump_block)
136 || jump_block->next_bb == EXIT_BLOCK_PTR
137 || !FORWARDER_BLOCK_P (jump_block))
138 return false;
139 jump_dest_block = single_succ (jump_block);
140
141 /* If we are partitioning hot/cold basic blocks, we don't want to
142 mess up unconditional or indirect jumps that cross between hot
143 and cold sections.
144
145 Basic block partitioning may result in some jumps that appear to
146 be optimizable (or blocks that appear to be mergeable), but which really
147 must be left untouched (they are required to make it safely across
148 partition boundaries). See the comments at the top of
149 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
150
151 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
152 || (cbranch_jump_edge->flags & EDGE_CROSSING))
153 return false;
154
155 /* The conditional branch must target the block after the
156 unconditional branch. */
157 cbranch_dest_block = cbranch_jump_edge->dest;
158
159 if (cbranch_dest_block == EXIT_BLOCK_PTR
160 || !can_fallthru (jump_block, cbranch_dest_block))
161 return false;
162
163 /* Invert the conditional branch. */
164 if (!invert_jump (cbranch_insn, block_label (jump_dest_block), 0))
165 return false;
166
167 if (dump_file)
168 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
169 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
170
171 /* Success. Update the CFG to match. Note that after this point
172 the edge variable names appear backwards; the redirection is done
173 this way to preserve edge profile data. */
174 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
175 cbranch_dest_block);
176 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
177 jump_dest_block);
178 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
179 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
180 update_br_prob_note (cbranch_block);
181
182 /* Delete the block with the unconditional jump, and clean up the mess. */
183 delete_basic_block (jump_block);
184 tidy_fallthru_edge (cbranch_jump_edge);
185 update_forwarder_flag (cbranch_block);
186
187 return true;
188 }
189 \f
190 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
191 on register. Used by jump threading. */
192
193 static bool
194 mark_effect (rtx exp, regset nonequal)
195 {
196 int regno;
197 rtx dest;
198 switch (GET_CODE (exp))
199 {
200 /* In case we do clobber the register, mark it as equal, as we know the
201 value is dead so it don't have to match. */
202 case CLOBBER:
203 if (REG_P (XEXP (exp, 0)))
204 {
205 dest = XEXP (exp, 0);
206 regno = REGNO (dest);
207 if (HARD_REGISTER_NUM_P (regno))
208 bitmap_clear_range (nonequal, regno,
209 hard_regno_nregs[regno][GET_MODE (dest)]);
210 else
211 bitmap_clear_bit (nonequal, regno);
212 }
213 return false;
214
215 case SET:
216 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
217 return false;
218 dest = SET_DEST (exp);
219 if (dest == pc_rtx)
220 return false;
221 if (!REG_P (dest))
222 return true;
223 regno = REGNO (dest);
224 if (HARD_REGISTER_NUM_P (regno))
225 bitmap_set_range (nonequal, regno,
226 hard_regno_nregs[regno][GET_MODE (dest)]);
227 else
228 bitmap_set_bit (nonequal, regno);
229 return false;
230
231 default:
232 return false;
233 }
234 }
235
236 /* Return nonzero if X is a register set in regset DATA.
237 Called via for_each_rtx. */
238 static int
239 mentions_nonequal_regs (rtx *x, void *data)
240 {
241 regset nonequal = (regset) data;
242 if (REG_P (*x))
243 {
244 int regno;
245
246 regno = REGNO (*x);
247 if (REGNO_REG_SET_P (nonequal, regno))
248 return 1;
249 if (regno < FIRST_PSEUDO_REGISTER)
250 {
251 int n = hard_regno_nregs[regno][GET_MODE (*x)];
252 while (--n > 0)
253 if (REGNO_REG_SET_P (nonequal, regno + n))
254 return 1;
255 }
256 }
257 return 0;
258 }
259 /* Attempt to prove that the basic block B will have no side effects and
260 always continues in the same edge if reached via E. Return the edge
261 if exist, NULL otherwise. */
262
263 static edge
264 thread_jump (edge e, basic_block b)
265 {
266 rtx set1, set2, cond1, cond2, insn;
267 enum rtx_code code1, code2, reversed_code2;
268 bool reverse1 = false;
269 unsigned i;
270 regset nonequal;
271 bool failed = false;
272 reg_set_iterator rsi;
273
274 if (b->flags & BB_NONTHREADABLE_BLOCK)
275 return NULL;
276
277 /* At the moment, we do handle only conditional jumps, but later we may
278 want to extend this code to tablejumps and others. */
279 if (EDGE_COUNT (e->src->succs) != 2)
280 return NULL;
281 if (EDGE_COUNT (b->succs) != 2)
282 {
283 b->flags |= BB_NONTHREADABLE_BLOCK;
284 return NULL;
285 }
286
287 /* Second branch must end with onlyjump, as we will eliminate the jump. */
288 if (!any_condjump_p (BB_END (e->src)))
289 return NULL;
290
291 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
292 {
293 b->flags |= BB_NONTHREADABLE_BLOCK;
294 return NULL;
295 }
296
297 set1 = pc_set (BB_END (e->src));
298 set2 = pc_set (BB_END (b));
299 if (((e->flags & EDGE_FALLTHRU) != 0)
300 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
301 reverse1 = true;
302
303 cond1 = XEXP (SET_SRC (set1), 0);
304 cond2 = XEXP (SET_SRC (set2), 0);
305 if (reverse1)
306 code1 = reversed_comparison_code (cond1, BB_END (e->src));
307 else
308 code1 = GET_CODE (cond1);
309
310 code2 = GET_CODE (cond2);
311 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
312
313 if (!comparison_dominates_p (code1, code2)
314 && !comparison_dominates_p (code1, reversed_code2))
315 return NULL;
316
317 /* Ensure that the comparison operators are equivalent.
318 ??? This is far too pessimistic. We should allow swapped operands,
319 different CCmodes, or for example comparisons for interval, that
320 dominate even when operands are not equivalent. */
321 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
322 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
323 return NULL;
324
325 /* Short circuit cases where block B contains some side effects, as we can't
326 safely bypass it. */
327 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
328 insn = NEXT_INSN (insn))
329 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
330 {
331 b->flags |= BB_NONTHREADABLE_BLOCK;
332 return NULL;
333 }
334
335 cselib_init (0);
336
337 /* First process all values computed in the source basic block. */
338 for (insn = NEXT_INSN (BB_HEAD (e->src));
339 insn != NEXT_INSN (BB_END (e->src));
340 insn = NEXT_INSN (insn))
341 if (INSN_P (insn))
342 cselib_process_insn (insn);
343
344 nonequal = BITMAP_ALLOC (NULL);
345 CLEAR_REG_SET (nonequal);
346
347 /* Now assume that we've continued by the edge E to B and continue
348 processing as if it were same basic block.
349 Our goal is to prove that whole block is an NOOP. */
350
351 for (insn = NEXT_INSN (BB_HEAD (b));
352 insn != NEXT_INSN (BB_END (b)) && !failed;
353 insn = NEXT_INSN (insn))
354 {
355 if (INSN_P (insn))
356 {
357 rtx pat = PATTERN (insn);
358
359 if (GET_CODE (pat) == PARALLEL)
360 {
361 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
362 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
363 }
364 else
365 failed |= mark_effect (pat, nonequal);
366 }
367
368 cselib_process_insn (insn);
369 }
370
371 /* Later we should clear nonequal of dead registers. So far we don't
372 have life information in cfg_cleanup. */
373 if (failed)
374 {
375 b->flags |= BB_NONTHREADABLE_BLOCK;
376 goto failed_exit;
377 }
378
379 /* cond2 must not mention any register that is not equal to the
380 former block. */
381 if (for_each_rtx (&cond2, mentions_nonequal_regs, nonequal))
382 goto failed_exit;
383
384 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
385 goto failed_exit;
386
387 BITMAP_FREE (nonequal);
388 cselib_finish ();
389 if ((comparison_dominates_p (code1, code2) != 0)
390 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
391 return BRANCH_EDGE (b);
392 else
393 return FALLTHRU_EDGE (b);
394
395 failed_exit:
396 BITMAP_FREE (nonequal);
397 cselib_finish ();
398 return NULL;
399 }
400 \f
401 /* Attempt to forward edges leaving basic block B.
402 Return true if successful. */
403
404 static bool
405 try_forward_edges (int mode, basic_block b)
406 {
407 bool changed = false;
408 edge_iterator ei;
409 edge e, *threaded_edges = NULL;
410
411 /* If we are partitioning hot/cold basic blocks, we don't want to
412 mess up unconditional or indirect jumps that cross between hot
413 and cold sections.
414
415 Basic block partitioning may result in some jumps that appear to
416 be optimizable (or blocks that appear to be mergeable), but which really
417 must be left untouched (they are required to make it safely across
418 partition boundaries). See the comments at the top of
419 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
420
421 if (find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX))
422 return false;
423
424 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
425 {
426 basic_block target, first;
427 int counter, goto_locus;
428 bool threaded = false;
429 int nthreaded_edges = 0;
430 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
431
432 /* Skip complex edges because we don't know how to update them.
433
434 Still handle fallthru edges, as we can succeed to forward fallthru
435 edge to the same place as the branch edge of conditional branch
436 and turn conditional branch to an unconditional branch. */
437 if (e->flags & EDGE_COMPLEX)
438 {
439 ei_next (&ei);
440 continue;
441 }
442
443 target = first = e->dest;
444 counter = NUM_FIXED_BLOCKS;
445 goto_locus = e->goto_locus;
446
447 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
448 up jumps that cross between hot/cold sections.
449
450 Basic block partitioning may result in some jumps that appear
451 to be optimizable (or blocks that appear to be mergeable), but which
452 really must be left untouched (they are required to make it safely
453 across partition boundaries). See the comments at the top of
454 bb-reorder.c:partition_hot_cold_basic_blocks for complete
455 details. */
456
457 if (first != EXIT_BLOCK_PTR
458 && find_reg_note (BB_END (first), REG_CROSSING_JUMP, NULL_RTX))
459 return changed;
460
461 while (counter < n_basic_blocks)
462 {
463 basic_block new_target = NULL;
464 bool new_target_threaded = false;
465 may_thread |= (target->flags & BB_MODIFIED) != 0;
466
467 if (FORWARDER_BLOCK_P (target)
468 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
469 && single_succ (target) != EXIT_BLOCK_PTR)
470 {
471 /* Bypass trivial infinite loops. */
472 new_target = single_succ (target);
473 if (target == new_target)
474 counter = n_basic_blocks;
475 else if (!optimize)
476 {
477 /* When not optimizing, ensure that edges or forwarder
478 blocks with different locus are not optimized out. */
479 int new_locus = single_succ_edge (target)->goto_locus;
480 int locus = goto_locus;
481
482 if (new_locus != UNKNOWN_LOCATION
483 && locus != UNKNOWN_LOCATION
484 && new_locus != locus)
485 new_target = NULL;
486 else
487 {
488 rtx last;
489
490 if (new_locus != UNKNOWN_LOCATION)
491 locus = new_locus;
492
493 last = BB_END (target);
494 if (DEBUG_INSN_P (last))
495 last = prev_nondebug_insn (last);
496
497 new_locus = last && INSN_P (last)
498 ? INSN_LOCATION (last) : 0;
499
500 if (new_locus != UNKNOWN_LOCATION
501 && locus != UNKNOWN_LOCATION
502 && new_locus != locus)
503 new_target = NULL;
504 else
505 {
506 if (new_locus != UNKNOWN_LOCATION)
507 locus = new_locus;
508
509 goto_locus = locus;
510 }
511 }
512 }
513 }
514
515 /* Allow to thread only over one edge at time to simplify updating
516 of probabilities. */
517 else if ((mode & CLEANUP_THREADING) && may_thread)
518 {
519 edge t = thread_jump (e, target);
520 if (t)
521 {
522 if (!threaded_edges)
523 threaded_edges = XNEWVEC (edge, n_basic_blocks);
524 else
525 {
526 int i;
527
528 /* Detect an infinite loop across blocks not
529 including the start block. */
530 for (i = 0; i < nthreaded_edges; ++i)
531 if (threaded_edges[i] == t)
532 break;
533 if (i < nthreaded_edges)
534 {
535 counter = n_basic_blocks;
536 break;
537 }
538 }
539
540 /* Detect an infinite loop across the start block. */
541 if (t->dest == b)
542 break;
543
544 gcc_assert (nthreaded_edges < n_basic_blocks - NUM_FIXED_BLOCKS);
545 threaded_edges[nthreaded_edges++] = t;
546
547 new_target = t->dest;
548 new_target_threaded = true;
549 }
550 }
551
552 if (!new_target)
553 break;
554
555 counter++;
556 target = new_target;
557 threaded |= new_target_threaded;
558 }
559
560 if (counter >= n_basic_blocks)
561 {
562 if (dump_file)
563 fprintf (dump_file, "Infinite loop in BB %i.\n",
564 target->index);
565 }
566 else if (target == first)
567 ; /* We didn't do anything. */
568 else
569 {
570 /* Save the values now, as the edge may get removed. */
571 gcov_type edge_count = e->count;
572 int edge_probability = e->probability;
573 int edge_frequency;
574 int n = 0;
575
576 e->goto_locus = goto_locus;
577
578 /* Don't force if target is exit block. */
579 if (threaded && target != EXIT_BLOCK_PTR)
580 {
581 notice_new_block (redirect_edge_and_branch_force (e, target));
582 if (dump_file)
583 fprintf (dump_file, "Conditionals threaded.\n");
584 }
585 else if (!redirect_edge_and_branch (e, target))
586 {
587 if (dump_file)
588 fprintf (dump_file,
589 "Forwarding edge %i->%i to %i failed.\n",
590 b->index, e->dest->index, target->index);
591 ei_next (&ei);
592 continue;
593 }
594
595 /* We successfully forwarded the edge. Now update profile
596 data: for each edge we traversed in the chain, remove
597 the original edge's execution count. */
598 edge_frequency = apply_probability (b->frequency, edge_probability);
599
600 do
601 {
602 edge t;
603
604 if (!single_succ_p (first))
605 {
606 gcc_assert (n < nthreaded_edges);
607 t = threaded_edges [n++];
608 gcc_assert (t->src == first);
609 update_bb_profile_for_threading (first, edge_frequency,
610 edge_count, t);
611 update_br_prob_note (first);
612 }
613 else
614 {
615 first->count -= edge_count;
616 if (first->count < 0)
617 first->count = 0;
618 first->frequency -= edge_frequency;
619 if (first->frequency < 0)
620 first->frequency = 0;
621 /* It is possible that as the result of
622 threading we've removed edge as it is
623 threaded to the fallthru edge. Avoid
624 getting out of sync. */
625 if (n < nthreaded_edges
626 && first == threaded_edges [n]->src)
627 n++;
628 t = single_succ_edge (first);
629 }
630
631 t->count -= edge_count;
632 if (t->count < 0)
633 t->count = 0;
634 first = t->dest;
635 }
636 while (first != target);
637
638 changed = true;
639 continue;
640 }
641 ei_next (&ei);
642 }
643
644 free (threaded_edges);
645 return changed;
646 }
647 \f
648
649 /* Blocks A and B are to be merged into a single block. A has no incoming
650 fallthru edge, so it can be moved before B without adding or modifying
651 any jumps (aside from the jump from A to B). */
652
653 static void
654 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
655 {
656 rtx barrier;
657
658 /* If we are partitioning hot/cold basic blocks, we don't want to
659 mess up unconditional or indirect jumps that cross between hot
660 and cold sections.
661
662 Basic block partitioning may result in some jumps that appear to
663 be optimizable (or blocks that appear to be mergeable), but which really
664 must be left untouched (they are required to make it safely across
665 partition boundaries). See the comments at the top of
666 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
667
668 if (BB_PARTITION (a) != BB_PARTITION (b))
669 return;
670
671 barrier = next_nonnote_insn (BB_END (a));
672 gcc_assert (BARRIER_P (barrier));
673 delete_insn (barrier);
674
675 /* Scramble the insn chain. */
676 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
677 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
678 df_set_bb_dirty (a);
679
680 if (dump_file)
681 fprintf (dump_file, "Moved block %d before %d and merged.\n",
682 a->index, b->index);
683
684 /* Swap the records for the two blocks around. */
685
686 unlink_block (a);
687 link_block (a, b->prev_bb);
688
689 /* Now blocks A and B are contiguous. Merge them. */
690 merge_blocks (a, b);
691 }
692
693 /* Blocks A and B are to be merged into a single block. B has no outgoing
694 fallthru edge, so it can be moved after A without adding or modifying
695 any jumps (aside from the jump from A to B). */
696
697 static void
698 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
699 {
700 rtx barrier, real_b_end;
701 rtx label, table;
702
703 /* If we are partitioning hot/cold basic blocks, we don't want to
704 mess up unconditional or indirect jumps that cross between hot
705 and cold sections.
706
707 Basic block partitioning may result in some jumps that appear to
708 be optimizable (or blocks that appear to be mergeable), but which really
709 must be left untouched (they are required to make it safely across
710 partition boundaries). See the comments at the top of
711 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
712
713 if (BB_PARTITION (a) != BB_PARTITION (b))
714 return;
715
716 real_b_end = BB_END (b);
717
718 /* If there is a jump table following block B temporarily add the jump table
719 to block B so that it will also be moved to the correct location. */
720 if (tablejump_p (BB_END (b), &label, &table)
721 && prev_active_insn (label) == BB_END (b))
722 {
723 BB_END (b) = table;
724 }
725
726 /* There had better have been a barrier there. Delete it. */
727 barrier = NEXT_INSN (BB_END (b));
728 if (barrier && BARRIER_P (barrier))
729 delete_insn (barrier);
730
731
732 /* Scramble the insn chain. */
733 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
734
735 /* Restore the real end of b. */
736 BB_END (b) = real_b_end;
737
738 if (dump_file)
739 fprintf (dump_file, "Moved block %d after %d and merged.\n",
740 b->index, a->index);
741
742 /* Now blocks A and B are contiguous. Merge them. */
743 merge_blocks (a, b);
744 }
745
746 /* Attempt to merge basic blocks that are potentially non-adjacent.
747 Return NULL iff the attempt failed, otherwise return basic block
748 where cleanup_cfg should continue. Because the merging commonly
749 moves basic block away or introduces another optimization
750 possibility, return basic block just before B so cleanup_cfg don't
751 need to iterate.
752
753 It may be good idea to return basic block before C in the case
754 C has been moved after B and originally appeared earlier in the
755 insn sequence, but we have no information available about the
756 relative ordering of these two. Hopefully it is not too common. */
757
758 static basic_block
759 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
760 {
761 basic_block next;
762
763 /* If we are partitioning hot/cold basic blocks, we don't want to
764 mess up unconditional or indirect jumps that cross between hot
765 and cold sections.
766
767 Basic block partitioning may result in some jumps that appear to
768 be optimizable (or blocks that appear to be mergeable), but which really
769 must be left untouched (they are required to make it safely across
770 partition boundaries). See the comments at the top of
771 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
772
773 if (BB_PARTITION (b) != BB_PARTITION (c))
774 return NULL;
775
776 /* If B has a fallthru edge to C, no need to move anything. */
777 if (e->flags & EDGE_FALLTHRU)
778 {
779 int b_index = b->index, c_index = c->index;
780
781 /* Protect the loop latches. */
782 if (current_loops && c->loop_father->latch == c)
783 return NULL;
784
785 merge_blocks (b, c);
786 update_forwarder_flag (b);
787
788 if (dump_file)
789 fprintf (dump_file, "Merged %d and %d without moving.\n",
790 b_index, c_index);
791
792 return b->prev_bb == ENTRY_BLOCK_PTR ? b : b->prev_bb;
793 }
794
795 /* Otherwise we will need to move code around. Do that only if expensive
796 transformations are allowed. */
797 else if (mode & CLEANUP_EXPENSIVE)
798 {
799 edge tmp_edge, b_fallthru_edge;
800 bool c_has_outgoing_fallthru;
801 bool b_has_incoming_fallthru;
802
803 /* Avoid overactive code motion, as the forwarder blocks should be
804 eliminated by edge redirection instead. One exception might have
805 been if B is a forwarder block and C has no fallthru edge, but
806 that should be cleaned up by bb-reorder instead. */
807 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
808 return NULL;
809
810 /* We must make sure to not munge nesting of lexical blocks,
811 and loop notes. This is done by squeezing out all the notes
812 and leaving them there to lie. Not ideal, but functional. */
813
814 tmp_edge = find_fallthru_edge (c->succs);
815 c_has_outgoing_fallthru = (tmp_edge != NULL);
816
817 tmp_edge = find_fallthru_edge (b->preds);
818 b_has_incoming_fallthru = (tmp_edge != NULL);
819 b_fallthru_edge = tmp_edge;
820 next = b->prev_bb;
821 if (next == c)
822 next = next->prev_bb;
823
824 /* Otherwise, we're going to try to move C after B. If C does
825 not have an outgoing fallthru, then it can be moved
826 immediately after B without introducing or modifying jumps. */
827 if (! c_has_outgoing_fallthru)
828 {
829 merge_blocks_move_successor_nojumps (b, c);
830 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
831 }
832
833 /* If B does not have an incoming fallthru, then it can be moved
834 immediately before C without introducing or modifying jumps.
835 C cannot be the first block, so we do not have to worry about
836 accessing a non-existent block. */
837
838 if (b_has_incoming_fallthru)
839 {
840 basic_block bb;
841
842 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR)
843 return NULL;
844 bb = force_nonfallthru (b_fallthru_edge);
845 if (bb)
846 notice_new_block (bb);
847 }
848
849 merge_blocks_move_predecessor_nojumps (b, c);
850 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
851 }
852
853 return NULL;
854 }
855 \f
856
857 /* Removes the memory attributes of MEM expression
858 if they are not equal. */
859
860 void
861 merge_memattrs (rtx x, rtx y)
862 {
863 int i;
864 int j;
865 enum rtx_code code;
866 const char *fmt;
867
868 if (x == y)
869 return;
870 if (x == 0 || y == 0)
871 return;
872
873 code = GET_CODE (x);
874
875 if (code != GET_CODE (y))
876 return;
877
878 if (GET_MODE (x) != GET_MODE (y))
879 return;
880
881 if (code == MEM && MEM_ATTRS (x) != MEM_ATTRS (y))
882 {
883 if (! MEM_ATTRS (x))
884 MEM_ATTRS (y) = 0;
885 else if (! MEM_ATTRS (y))
886 MEM_ATTRS (x) = 0;
887 else
888 {
889 HOST_WIDE_INT mem_size;
890
891 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
892 {
893 set_mem_alias_set (x, 0);
894 set_mem_alias_set (y, 0);
895 }
896
897 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
898 {
899 set_mem_expr (x, 0);
900 set_mem_expr (y, 0);
901 clear_mem_offset (x);
902 clear_mem_offset (y);
903 }
904 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
905 || (MEM_OFFSET_KNOWN_P (x)
906 && MEM_OFFSET (x) != MEM_OFFSET (y)))
907 {
908 clear_mem_offset (x);
909 clear_mem_offset (y);
910 }
911
912 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
913 {
914 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
915 set_mem_size (x, mem_size);
916 set_mem_size (y, mem_size);
917 }
918 else
919 {
920 clear_mem_size (x);
921 clear_mem_size (y);
922 }
923
924 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
925 set_mem_align (y, MEM_ALIGN (x));
926 }
927 }
928 if (code == MEM)
929 {
930 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
931 {
932 MEM_READONLY_P (x) = 0;
933 MEM_READONLY_P (y) = 0;
934 }
935 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
936 {
937 MEM_NOTRAP_P (x) = 0;
938 MEM_NOTRAP_P (y) = 0;
939 }
940 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
941 {
942 MEM_VOLATILE_P (x) = 1;
943 MEM_VOLATILE_P (y) = 1;
944 }
945 }
946
947 fmt = GET_RTX_FORMAT (code);
948 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
949 {
950 switch (fmt[i])
951 {
952 case 'E':
953 /* Two vectors must have the same length. */
954 if (XVECLEN (x, i) != XVECLEN (y, i))
955 return;
956
957 for (j = 0; j < XVECLEN (x, i); j++)
958 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
959
960 break;
961
962 case 'e':
963 merge_memattrs (XEXP (x, i), XEXP (y, i));
964 }
965 }
966 return;
967 }
968
969
970 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
971 different single sets S1 and S2. */
972
973 static bool
974 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
975 {
976 int i;
977 rtx e1, e2;
978
979 if (p1 == s1 && p2 == s2)
980 return true;
981
982 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
983 return false;
984
985 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
986 return false;
987
988 for (i = 0; i < XVECLEN (p1, 0); i++)
989 {
990 e1 = XVECEXP (p1, 0, i);
991 e2 = XVECEXP (p2, 0, i);
992 if (e1 == s1 && e2 == s2)
993 continue;
994 if (reload_completed
995 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
996 continue;
997
998 return false;
999 }
1000
1001 return true;
1002 }
1003
1004 /* Examine register notes on I1 and I2 and return:
1005 - dir_forward if I1 can be replaced by I2, or
1006 - dir_backward if I2 can be replaced by I1, or
1007 - dir_both if both are the case. */
1008
1009 static enum replace_direction
1010 can_replace_by (rtx i1, rtx i2)
1011 {
1012 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1013 bool c1, c2;
1014
1015 /* Check for 2 sets. */
1016 s1 = single_set (i1);
1017 s2 = single_set (i2);
1018 if (s1 == NULL_RTX || s2 == NULL_RTX)
1019 return dir_none;
1020
1021 /* Check that the 2 sets set the same dest. */
1022 d1 = SET_DEST (s1);
1023 d2 = SET_DEST (s2);
1024 if (!(reload_completed
1025 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1026 return dir_none;
1027
1028 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1029 set dest to the same value. */
1030 note1 = find_reg_equal_equiv_note (i1);
1031 note2 = find_reg_equal_equiv_note (i2);
1032 if (!note1 || !note2 || !rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0))
1033 || !CONST_INT_P (XEXP (note1, 0)))
1034 return dir_none;
1035
1036 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1037 return dir_none;
1038
1039 /* Although the 2 sets set dest to the same value, we cannot replace
1040 (set (dest) (const_int))
1041 by
1042 (set (dest) (reg))
1043 because we don't know if the reg is live and has the same value at the
1044 location of replacement. */
1045 src1 = SET_SRC (s1);
1046 src2 = SET_SRC (s2);
1047 c1 = CONST_INT_P (src1);
1048 c2 = CONST_INT_P (src2);
1049 if (c1 && c2)
1050 return dir_both;
1051 else if (c2)
1052 return dir_forward;
1053 else if (c1)
1054 return dir_backward;
1055
1056 return dir_none;
1057 }
1058
1059 /* Merges directions A and B. */
1060
1061 static enum replace_direction
1062 merge_dir (enum replace_direction a, enum replace_direction b)
1063 {
1064 /* Implements the following table:
1065 |bo fw bw no
1066 ---+-----------
1067 bo |bo fw bw no
1068 fw |-- fw no no
1069 bw |-- -- bw no
1070 no |-- -- -- no. */
1071
1072 if (a == b)
1073 return a;
1074
1075 if (a == dir_both)
1076 return b;
1077 if (b == dir_both)
1078 return a;
1079
1080 return dir_none;
1081 }
1082
1083 /* Examine I1 and I2 and return:
1084 - dir_forward if I1 can be replaced by I2, or
1085 - dir_backward if I2 can be replaced by I1, or
1086 - dir_both if both are the case. */
1087
1088 static enum replace_direction
1089 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx i1, rtx i2)
1090 {
1091 rtx p1, p2;
1092
1093 /* Verify that I1 and I2 are equivalent. */
1094 if (GET_CODE (i1) != GET_CODE (i2))
1095 return dir_none;
1096
1097 /* __builtin_unreachable() may lead to empty blocks (ending with
1098 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1099 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1100 return dir_both;
1101
1102 /* ??? Do not allow cross-jumping between different stack levels. */
1103 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1104 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1105 if (p1 && p2)
1106 {
1107 p1 = XEXP (p1, 0);
1108 p2 = XEXP (p2, 0);
1109 if (!rtx_equal_p (p1, p2))
1110 return dir_none;
1111
1112 /* ??? Worse, this adjustment had better be constant lest we
1113 have differing incoming stack levels. */
1114 if (!frame_pointer_needed
1115 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1116 return dir_none;
1117 }
1118 else if (p1 || p2)
1119 return dir_none;
1120
1121 p1 = PATTERN (i1);
1122 p2 = PATTERN (i2);
1123
1124 if (GET_CODE (p1) != GET_CODE (p2))
1125 return dir_none;
1126
1127 /* If this is a CALL_INSN, compare register usage information.
1128 If we don't check this on stack register machines, the two
1129 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1130 numbers of stack registers in the same basic block.
1131 If we don't check this on machines with delay slots, a delay slot may
1132 be filled that clobbers a parameter expected by the subroutine.
1133
1134 ??? We take the simple route for now and assume that if they're
1135 equal, they were constructed identically.
1136
1137 Also check for identical exception regions. */
1138
1139 if (CALL_P (i1))
1140 {
1141 /* Ensure the same EH region. */
1142 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1143 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1144
1145 if (!n1 && n2)
1146 return dir_none;
1147
1148 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1149 return dir_none;
1150
1151 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1152 CALL_INSN_FUNCTION_USAGE (i2))
1153 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1154 return dir_none;
1155
1156 /* For address sanitizer, never crossjump __asan_report_* builtins,
1157 otherwise errors might be reported on incorrect lines. */
1158 if (flag_sanitize & SANITIZE_ADDRESS)
1159 {
1160 rtx call = get_call_rtx_from (i1);
1161 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1162 {
1163 rtx symbol = XEXP (XEXP (call, 0), 0);
1164 if (SYMBOL_REF_DECL (symbol)
1165 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1166 {
1167 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1168 == BUILT_IN_NORMAL)
1169 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1170 >= BUILT_IN_ASAN_REPORT_LOAD1
1171 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1172 <= BUILT_IN_ASAN_REPORT_STORE16)
1173 return dir_none;
1174 }
1175 }
1176 }
1177 }
1178
1179 #ifdef STACK_REGS
1180 /* If cross_jump_death_matters is not 0, the insn's mode
1181 indicates whether or not the insn contains any stack-like
1182 regs. */
1183
1184 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1185 {
1186 /* If register stack conversion has already been done, then
1187 death notes must also be compared before it is certain that
1188 the two instruction streams match. */
1189
1190 rtx note;
1191 HARD_REG_SET i1_regset, i2_regset;
1192
1193 CLEAR_HARD_REG_SET (i1_regset);
1194 CLEAR_HARD_REG_SET (i2_regset);
1195
1196 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1197 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1198 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1199
1200 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1201 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1202 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1203
1204 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1205 return dir_none;
1206 }
1207 #endif
1208
1209 if (reload_completed
1210 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1211 return dir_both;
1212
1213 return can_replace_by (i1, i2);
1214 }
1215 \f
1216 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1217 flow_find_head_matching_sequence, ensure the notes match. */
1218
1219 static void
1220 merge_notes (rtx i1, rtx i2)
1221 {
1222 /* If the merged insns have different REG_EQUAL notes, then
1223 remove them. */
1224 rtx equiv1 = find_reg_equal_equiv_note (i1);
1225 rtx equiv2 = find_reg_equal_equiv_note (i2);
1226
1227 if (equiv1 && !equiv2)
1228 remove_note (i1, equiv1);
1229 else if (!equiv1 && equiv2)
1230 remove_note (i2, equiv2);
1231 else if (equiv1 && equiv2
1232 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1233 {
1234 remove_note (i1, equiv1);
1235 remove_note (i2, equiv2);
1236 }
1237 }
1238
1239 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1240 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1241 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1242 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1243 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1244
1245 static void
1246 walk_to_nondebug_insn (rtx *i1, basic_block *bb1, bool follow_fallthru,
1247 bool *did_fallthru)
1248 {
1249 edge fallthru;
1250
1251 *did_fallthru = false;
1252
1253 /* Ignore notes. */
1254 while (!NONDEBUG_INSN_P (*i1))
1255 {
1256 if (*i1 != BB_HEAD (*bb1))
1257 {
1258 *i1 = PREV_INSN (*i1);
1259 continue;
1260 }
1261
1262 if (!follow_fallthru)
1263 return;
1264
1265 fallthru = find_fallthru_edge ((*bb1)->preds);
1266 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun)
1267 || !single_succ_p (fallthru->src))
1268 return;
1269
1270 *bb1 = fallthru->src;
1271 *i1 = BB_END (*bb1);
1272 *did_fallthru = true;
1273 }
1274 }
1275
1276 /* Look through the insns at the end of BB1 and BB2 and find the longest
1277 sequence that are either equivalent, or allow forward or backward
1278 replacement. Store the first insns for that sequence in *F1 and *F2 and
1279 return the sequence length.
1280
1281 DIR_P indicates the allowed replacement direction on function entry, and
1282 the actual replacement direction on function exit. If NULL, only equivalent
1283 sequences are allowed.
1284
1285 To simplify callers of this function, if the blocks match exactly,
1286 store the head of the blocks in *F1 and *F2. */
1287
1288 int
1289 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx *f1, rtx *f2,
1290 enum replace_direction *dir_p)
1291 {
1292 rtx i1, i2, last1, last2, afterlast1, afterlast2;
1293 int ninsns = 0;
1294 rtx p1;
1295 enum replace_direction dir, last_dir, afterlast_dir;
1296 bool follow_fallthru, did_fallthru;
1297
1298 if (dir_p)
1299 dir = *dir_p;
1300 else
1301 dir = dir_both;
1302 afterlast_dir = dir;
1303 last_dir = afterlast_dir;
1304
1305 /* Skip simple jumps at the end of the blocks. Complex jumps still
1306 need to be compared for equivalence, which we'll do below. */
1307
1308 i1 = BB_END (bb1);
1309 last1 = afterlast1 = last2 = afterlast2 = NULL_RTX;
1310 if (onlyjump_p (i1)
1311 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1312 {
1313 last1 = i1;
1314 i1 = PREV_INSN (i1);
1315 }
1316
1317 i2 = BB_END (bb2);
1318 if (onlyjump_p (i2)
1319 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1320 {
1321 last2 = i2;
1322 /* Count everything except for unconditional jump as insn. */
1323 if (!simplejump_p (i2) && !returnjump_p (i2) && last1)
1324 ninsns++;
1325 i2 = PREV_INSN (i2);
1326 }
1327
1328 while (true)
1329 {
1330 /* In the following example, we can replace all jumps to C by jumps to A.
1331
1332 This removes 4 duplicate insns.
1333 [bb A] insn1 [bb C] insn1
1334 insn2 insn2
1335 [bb B] insn3 insn3
1336 insn4 insn4
1337 jump_insn jump_insn
1338
1339 We could also replace all jumps to A by jumps to C, but that leaves B
1340 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1341 step, all jumps to B would be replaced with jumps to the middle of C,
1342 achieving the same result with more effort.
1343 So we allow only the first possibility, which means that we don't allow
1344 fallthru in the block that's being replaced. */
1345
1346 follow_fallthru = dir_p && dir != dir_forward;
1347 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1348 if (did_fallthru)
1349 dir = dir_backward;
1350
1351 follow_fallthru = dir_p && dir != dir_backward;
1352 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1353 if (did_fallthru)
1354 dir = dir_forward;
1355
1356 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1357 break;
1358
1359 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1360 if (dir == dir_none || (!dir_p && dir != dir_both))
1361 break;
1362
1363 merge_memattrs (i1, i2);
1364
1365 /* Don't begin a cross-jump with a NOTE insn. */
1366 if (INSN_P (i1))
1367 {
1368 merge_notes (i1, i2);
1369
1370 afterlast1 = last1, afterlast2 = last2;
1371 last1 = i1, last2 = i2;
1372 afterlast_dir = last_dir;
1373 last_dir = dir;
1374 p1 = PATTERN (i1);
1375 if (!(GET_CODE (p1) == USE || GET_CODE (p1) == CLOBBER))
1376 ninsns++;
1377 }
1378
1379 i1 = PREV_INSN (i1);
1380 i2 = PREV_INSN (i2);
1381 }
1382
1383 #ifdef HAVE_cc0
1384 /* Don't allow the insn after a compare to be shared by
1385 cross-jumping unless the compare is also shared. */
1386 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
1387 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1388 #endif
1389
1390 /* Include preceding notes and labels in the cross-jump. One,
1391 this may bring us to the head of the blocks as requested above.
1392 Two, it keeps line number notes as matched as may be. */
1393 if (ninsns)
1394 {
1395 bb1 = BLOCK_FOR_INSN (last1);
1396 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1397 last1 = PREV_INSN (last1);
1398
1399 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1400 last1 = PREV_INSN (last1);
1401
1402 bb2 = BLOCK_FOR_INSN (last2);
1403 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1404 last2 = PREV_INSN (last2);
1405
1406 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1407 last2 = PREV_INSN (last2);
1408
1409 *f1 = last1;
1410 *f2 = last2;
1411 }
1412
1413 if (dir_p)
1414 *dir_p = last_dir;
1415 return ninsns;
1416 }
1417
1418 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1419 the head of the two blocks. Do not include jumps at the end.
1420 If STOP_AFTER is nonzero, stop after finding that many matching
1421 instructions. */
1422
1423 int
1424 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx *f1,
1425 rtx *f2, int stop_after)
1426 {
1427 rtx i1, i2, last1, last2, beforelast1, beforelast2;
1428 int ninsns = 0;
1429 edge e;
1430 edge_iterator ei;
1431 int nehedges1 = 0, nehedges2 = 0;
1432
1433 FOR_EACH_EDGE (e, ei, bb1->succs)
1434 if (e->flags & EDGE_EH)
1435 nehedges1++;
1436 FOR_EACH_EDGE (e, ei, bb2->succs)
1437 if (e->flags & EDGE_EH)
1438 nehedges2++;
1439
1440 i1 = BB_HEAD (bb1);
1441 i2 = BB_HEAD (bb2);
1442 last1 = beforelast1 = last2 = beforelast2 = NULL_RTX;
1443
1444 while (true)
1445 {
1446 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1447 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1448 {
1449 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1450 break;
1451 i1 = NEXT_INSN (i1);
1452 }
1453
1454 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1455 {
1456 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1457 break;
1458 i2 = NEXT_INSN (i2);
1459 }
1460
1461 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1462 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1463 break;
1464
1465 if (NOTE_P (i1) || NOTE_P (i2)
1466 || JUMP_P (i1) || JUMP_P (i2))
1467 break;
1468
1469 /* A sanity check to make sure we're not merging insns with different
1470 effects on EH. If only one of them ends a basic block, it shouldn't
1471 have an EH edge; if both end a basic block, there should be the same
1472 number of EH edges. */
1473 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1474 && nehedges1 > 0)
1475 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1476 && nehedges2 > 0)
1477 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1478 && nehedges1 != nehedges2))
1479 break;
1480
1481 if (old_insns_match_p (0, i1, i2) != dir_both)
1482 break;
1483
1484 merge_memattrs (i1, i2);
1485
1486 /* Don't begin a cross-jump with a NOTE insn. */
1487 if (INSN_P (i1))
1488 {
1489 merge_notes (i1, i2);
1490
1491 beforelast1 = last1, beforelast2 = last2;
1492 last1 = i1, last2 = i2;
1493 ninsns++;
1494 }
1495
1496 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1497 || (stop_after > 0 && ninsns == stop_after))
1498 break;
1499
1500 i1 = NEXT_INSN (i1);
1501 i2 = NEXT_INSN (i2);
1502 }
1503
1504 #ifdef HAVE_cc0
1505 /* Don't allow a compare to be shared by cross-jumping unless the insn
1506 after the compare is also shared. */
1507 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && sets_cc0_p (last1))
1508 last1 = beforelast1, last2 = beforelast2, ninsns--;
1509 #endif
1510
1511 if (ninsns)
1512 {
1513 *f1 = last1;
1514 *f2 = last2;
1515 }
1516
1517 return ninsns;
1518 }
1519
1520 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1521 the branch instruction. This means that if we commonize the control
1522 flow before end of the basic block, the semantic remains unchanged.
1523
1524 We may assume that there exists one edge with a common destination. */
1525
1526 static bool
1527 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1528 {
1529 int nehedges1 = 0, nehedges2 = 0;
1530 edge fallthru1 = 0, fallthru2 = 0;
1531 edge e1, e2;
1532 edge_iterator ei;
1533
1534 /* If we performed shrink-wrapping, edges to the EXIT_BLOCK_PTR can
1535 only be distinguished for JUMP_INSNs. The two paths may differ in
1536 whether they went through the prologue. Sibcalls are fine, we know
1537 that we either didn't need or inserted an epilogue before them. */
1538 if (crtl->shrink_wrapped
1539 && single_succ_p (bb1) && single_succ (bb1) == EXIT_BLOCK_PTR
1540 && !JUMP_P (BB_END (bb1))
1541 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1542 return false;
1543
1544 /* If BB1 has only one successor, we may be looking at either an
1545 unconditional jump, or a fake edge to exit. */
1546 if (single_succ_p (bb1)
1547 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1548 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1549 return (single_succ_p (bb2)
1550 && (single_succ_edge (bb2)->flags
1551 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1552 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1553
1554 /* Match conditional jumps - this may get tricky when fallthru and branch
1555 edges are crossed. */
1556 if (EDGE_COUNT (bb1->succs) == 2
1557 && any_condjump_p (BB_END (bb1))
1558 && onlyjump_p (BB_END (bb1)))
1559 {
1560 edge b1, f1, b2, f2;
1561 bool reverse, match;
1562 rtx set1, set2, cond1, cond2;
1563 enum rtx_code code1, code2;
1564
1565 if (EDGE_COUNT (bb2->succs) != 2
1566 || !any_condjump_p (BB_END (bb2))
1567 || !onlyjump_p (BB_END (bb2)))
1568 return false;
1569
1570 b1 = BRANCH_EDGE (bb1);
1571 b2 = BRANCH_EDGE (bb2);
1572 f1 = FALLTHRU_EDGE (bb1);
1573 f2 = FALLTHRU_EDGE (bb2);
1574
1575 /* Get around possible forwarders on fallthru edges. Other cases
1576 should be optimized out already. */
1577 if (FORWARDER_BLOCK_P (f1->dest))
1578 f1 = single_succ_edge (f1->dest);
1579
1580 if (FORWARDER_BLOCK_P (f2->dest))
1581 f2 = single_succ_edge (f2->dest);
1582
1583 /* To simplify use of this function, return false if there are
1584 unneeded forwarder blocks. These will get eliminated later
1585 during cleanup_cfg. */
1586 if (FORWARDER_BLOCK_P (f1->dest)
1587 || FORWARDER_BLOCK_P (f2->dest)
1588 || FORWARDER_BLOCK_P (b1->dest)
1589 || FORWARDER_BLOCK_P (b2->dest))
1590 return false;
1591
1592 if (f1->dest == f2->dest && b1->dest == b2->dest)
1593 reverse = false;
1594 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1595 reverse = true;
1596 else
1597 return false;
1598
1599 set1 = pc_set (BB_END (bb1));
1600 set2 = pc_set (BB_END (bb2));
1601 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1602 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1603 reverse = !reverse;
1604
1605 cond1 = XEXP (SET_SRC (set1), 0);
1606 cond2 = XEXP (SET_SRC (set2), 0);
1607 code1 = GET_CODE (cond1);
1608 if (reverse)
1609 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1610 else
1611 code2 = GET_CODE (cond2);
1612
1613 if (code2 == UNKNOWN)
1614 return false;
1615
1616 /* Verify codes and operands match. */
1617 match = ((code1 == code2
1618 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1619 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1620 || (code1 == swap_condition (code2)
1621 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1622 XEXP (cond2, 0))
1623 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1624 XEXP (cond2, 1))));
1625
1626 /* If we return true, we will join the blocks. Which means that
1627 we will only have one branch prediction bit to work with. Thus
1628 we require the existing branches to have probabilities that are
1629 roughly similar. */
1630 if (match
1631 && optimize_bb_for_speed_p (bb1)
1632 && optimize_bb_for_speed_p (bb2))
1633 {
1634 int prob2;
1635
1636 if (b1->dest == b2->dest)
1637 prob2 = b2->probability;
1638 else
1639 /* Do not use f2 probability as f2 may be forwarded. */
1640 prob2 = REG_BR_PROB_BASE - b2->probability;
1641
1642 /* Fail if the difference in probabilities is greater than 50%.
1643 This rules out two well-predicted branches with opposite
1644 outcomes. */
1645 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1646 {
1647 if (dump_file)
1648 fprintf (dump_file,
1649 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1650 bb1->index, bb2->index, b1->probability, prob2);
1651
1652 return false;
1653 }
1654 }
1655
1656 if (dump_file && match)
1657 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1658 bb1->index, bb2->index);
1659
1660 return match;
1661 }
1662
1663 /* Generic case - we are seeing a computed jump, table jump or trapping
1664 instruction. */
1665
1666 /* Check whether there are tablejumps in the end of BB1 and BB2.
1667 Return true if they are identical. */
1668 {
1669 rtx label1, label2;
1670 rtx table1, table2;
1671
1672 if (tablejump_p (BB_END (bb1), &label1, &table1)
1673 && tablejump_p (BB_END (bb2), &label2, &table2)
1674 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1675 {
1676 /* The labels should never be the same rtx. If they really are same
1677 the jump tables are same too. So disable crossjumping of blocks BB1
1678 and BB2 because when deleting the common insns in the end of BB1
1679 by delete_basic_block () the jump table would be deleted too. */
1680 /* If LABEL2 is referenced in BB1->END do not do anything
1681 because we would loose information when replacing
1682 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1683 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1684 {
1685 /* Set IDENTICAL to true when the tables are identical. */
1686 bool identical = false;
1687 rtx p1, p2;
1688
1689 p1 = PATTERN (table1);
1690 p2 = PATTERN (table2);
1691 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1692 {
1693 identical = true;
1694 }
1695 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1696 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1697 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1698 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1699 {
1700 int i;
1701
1702 identical = true;
1703 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1704 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1705 identical = false;
1706 }
1707
1708 if (identical)
1709 {
1710 replace_label_data rr;
1711 bool match;
1712
1713 /* Temporarily replace references to LABEL1 with LABEL2
1714 in BB1->END so that we could compare the instructions. */
1715 rr.r1 = label1;
1716 rr.r2 = label2;
1717 rr.update_label_nuses = false;
1718 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1719
1720 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1721 == dir_both);
1722 if (dump_file && match)
1723 fprintf (dump_file,
1724 "Tablejumps in bb %i and %i match.\n",
1725 bb1->index, bb2->index);
1726
1727 /* Set the original label in BB1->END because when deleting
1728 a block whose end is a tablejump, the tablejump referenced
1729 from the instruction is deleted too. */
1730 rr.r1 = label2;
1731 rr.r2 = label1;
1732 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1733
1734 return match;
1735 }
1736 }
1737 return false;
1738 }
1739 }
1740
1741 rtx last1 = BB_END (bb1);
1742 rtx last2 = BB_END (bb2);
1743 if (DEBUG_INSN_P (last1))
1744 last1 = prev_nondebug_insn (last1);
1745 if (DEBUG_INSN_P (last2))
1746 last2 = prev_nondebug_insn (last2);
1747 /* First ensure that the instructions match. There may be many outgoing
1748 edges so this test is generally cheaper. */
1749 if (old_insns_match_p (mode, last1, last2) != dir_both)
1750 return false;
1751
1752 /* Search the outgoing edges, ensure that the counts do match, find possible
1753 fallthru and exception handling edges since these needs more
1754 validation. */
1755 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1756 return false;
1757
1758 bool nonfakeedges = false;
1759 FOR_EACH_EDGE (e1, ei, bb1->succs)
1760 {
1761 e2 = EDGE_SUCC (bb2, ei.index);
1762
1763 if ((e1->flags & EDGE_FAKE) == 0)
1764 nonfakeedges = true;
1765
1766 if (e1->flags & EDGE_EH)
1767 nehedges1++;
1768
1769 if (e2->flags & EDGE_EH)
1770 nehedges2++;
1771
1772 if (e1->flags & EDGE_FALLTHRU)
1773 fallthru1 = e1;
1774 if (e2->flags & EDGE_FALLTHRU)
1775 fallthru2 = e2;
1776 }
1777
1778 /* If number of edges of various types does not match, fail. */
1779 if (nehedges1 != nehedges2
1780 || (fallthru1 != 0) != (fallthru2 != 0))
1781 return false;
1782
1783 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1784 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1785 attempt to optimize, as the two basic blocks might have different
1786 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1787 traps there should be REG_ARG_SIZE notes, they could be missing
1788 for __builtin_unreachable () uses though. */
1789 if (!nonfakeedges
1790 && !ACCUMULATE_OUTGOING_ARGS
1791 && (!INSN_P (last1)
1792 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1793 return false;
1794
1795 /* fallthru edges must be forwarded to the same destination. */
1796 if (fallthru1)
1797 {
1798 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1799 ? single_succ (fallthru1->dest): fallthru1->dest);
1800 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1801 ? single_succ (fallthru2->dest): fallthru2->dest);
1802
1803 if (d1 != d2)
1804 return false;
1805 }
1806
1807 /* Ensure the same EH region. */
1808 {
1809 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1810 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1811
1812 if (!n1 && n2)
1813 return false;
1814
1815 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1816 return false;
1817 }
1818
1819 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1820 version of sequence abstraction. */
1821 FOR_EACH_EDGE (e1, ei, bb2->succs)
1822 {
1823 edge e2;
1824 edge_iterator ei;
1825 basic_block d1 = e1->dest;
1826
1827 if (FORWARDER_BLOCK_P (d1))
1828 d1 = EDGE_SUCC (d1, 0)->dest;
1829
1830 FOR_EACH_EDGE (e2, ei, bb1->succs)
1831 {
1832 basic_block d2 = e2->dest;
1833 if (FORWARDER_BLOCK_P (d2))
1834 d2 = EDGE_SUCC (d2, 0)->dest;
1835 if (d1 == d2)
1836 break;
1837 }
1838
1839 if (!e2)
1840 return false;
1841 }
1842
1843 return true;
1844 }
1845
1846 /* Returns true if BB basic block has a preserve label. */
1847
1848 static bool
1849 block_has_preserve_label (basic_block bb)
1850 {
1851 return (bb
1852 && block_label (bb)
1853 && LABEL_PRESERVE_P (block_label (bb)));
1854 }
1855
1856 /* E1 and E2 are edges with the same destination block. Search their
1857 predecessors for common code. If found, redirect control flow from
1858 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1859 or the other way around (dir_backward). DIR specifies the allowed
1860 replacement direction. */
1861
1862 static bool
1863 try_crossjump_to_edge (int mode, edge e1, edge e2,
1864 enum replace_direction dir)
1865 {
1866 int nmatch;
1867 basic_block src1 = e1->src, src2 = e2->src;
1868 basic_block redirect_to, redirect_from, to_remove;
1869 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1870 rtx newpos1, newpos2;
1871 edge s;
1872 edge_iterator ei;
1873
1874 newpos1 = newpos2 = NULL_RTX;
1875
1876 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1877 to try this optimization.
1878
1879 Basic block partitioning may result in some jumps that appear to
1880 be optimizable (or blocks that appear to be mergeable), but which really
1881 must be left untouched (they are required to make it safely across
1882 partition boundaries). See the comments at the top of
1883 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1884
1885 if (crtl->has_bb_partition && reload_completed)
1886 return false;
1887
1888 /* Search backward through forwarder blocks. We don't need to worry
1889 about multiple entry or chained forwarders, as they will be optimized
1890 away. We do this to look past the unconditional jump following a
1891 conditional jump that is required due to the current CFG shape. */
1892 if (single_pred_p (src1)
1893 && FORWARDER_BLOCK_P (src1))
1894 e1 = single_pred_edge (src1), src1 = e1->src;
1895
1896 if (single_pred_p (src2)
1897 && FORWARDER_BLOCK_P (src2))
1898 e2 = single_pred_edge (src2), src2 = e2->src;
1899
1900 /* Nothing to do if we reach ENTRY, or a common source block. */
1901 if (src1 == ENTRY_BLOCK_PTR || src2 == ENTRY_BLOCK_PTR)
1902 return false;
1903 if (src1 == src2)
1904 return false;
1905
1906 /* Seeing more than 1 forwarder blocks would confuse us later... */
1907 if (FORWARDER_BLOCK_P (e1->dest)
1908 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1909 return false;
1910
1911 if (FORWARDER_BLOCK_P (e2->dest)
1912 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1913 return false;
1914
1915 /* Likewise with dead code (possibly newly created by the other optimizations
1916 of cfg_cleanup). */
1917 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1918 return false;
1919
1920 /* Look for the common insn sequence, part the first ... */
1921 if (!outgoing_edges_match (mode, src1, src2))
1922 return false;
1923
1924 /* ... and part the second. */
1925 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1926
1927 osrc1 = src1;
1928 osrc2 = src2;
1929 if (newpos1 != NULL_RTX)
1930 src1 = BLOCK_FOR_INSN (newpos1);
1931 if (newpos2 != NULL_RTX)
1932 src2 = BLOCK_FOR_INSN (newpos2);
1933
1934 if (dir == dir_backward)
1935 {
1936 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1937 SWAP (basic_block, osrc1, osrc2);
1938 SWAP (basic_block, src1, src2);
1939 SWAP (edge, e1, e2);
1940 SWAP (rtx, newpos1, newpos2);
1941 #undef SWAP
1942 }
1943
1944 /* Don't proceed with the crossjump unless we found a sufficient number
1945 of matching instructions or the 'from' block was totally matched
1946 (such that its predecessors will hopefully be redirected and the
1947 block removed). */
1948 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
1949 && (newpos1 != BB_HEAD (src1)))
1950 return false;
1951
1952 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1953 if (block_has_preserve_label (e1->dest)
1954 && (e1->flags & EDGE_ABNORMAL))
1955 return false;
1956
1957 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1958 will be deleted.
1959 If we have tablejumps in the end of SRC1 and SRC2
1960 they have been already compared for equivalence in outgoing_edges_match ()
1961 so replace the references to TABLE1 by references to TABLE2. */
1962 {
1963 rtx label1, label2;
1964 rtx table1, table2;
1965
1966 if (tablejump_p (BB_END (osrc1), &label1, &table1)
1967 && tablejump_p (BB_END (osrc2), &label2, &table2)
1968 && label1 != label2)
1969 {
1970 replace_label_data rr;
1971 rtx insn;
1972
1973 /* Replace references to LABEL1 with LABEL2. */
1974 rr.r1 = label1;
1975 rr.r2 = label2;
1976 rr.update_label_nuses = true;
1977 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
1978 {
1979 /* Do not replace the label in SRC1->END because when deleting
1980 a block whose end is a tablejump, the tablejump referenced
1981 from the instruction is deleted too. */
1982 if (insn != BB_END (osrc1))
1983 for_each_rtx (&insn, replace_label, &rr);
1984 }
1985 }
1986 }
1987
1988 /* Avoid splitting if possible. We must always split when SRC2 has
1989 EH predecessor edges, or we may end up with basic blocks with both
1990 normal and EH predecessor edges. */
1991 if (newpos2 == BB_HEAD (src2)
1992 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
1993 redirect_to = src2;
1994 else
1995 {
1996 if (newpos2 == BB_HEAD (src2))
1997 {
1998 /* Skip possible basic block header. */
1999 if (LABEL_P (newpos2))
2000 newpos2 = NEXT_INSN (newpos2);
2001 while (DEBUG_INSN_P (newpos2))
2002 newpos2 = NEXT_INSN (newpos2);
2003 if (NOTE_P (newpos2))
2004 newpos2 = NEXT_INSN (newpos2);
2005 while (DEBUG_INSN_P (newpos2))
2006 newpos2 = NEXT_INSN (newpos2);
2007 }
2008
2009 if (dump_file)
2010 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2011 src2->index, nmatch);
2012 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2013 }
2014
2015 if (dump_file)
2016 fprintf (dump_file,
2017 "Cross jumping from bb %i to bb %i; %i common insns\n",
2018 src1->index, src2->index, nmatch);
2019
2020 /* We may have some registers visible through the block. */
2021 df_set_bb_dirty (redirect_to);
2022
2023 if (osrc2 == src2)
2024 redirect_edges_to = redirect_to;
2025 else
2026 redirect_edges_to = osrc2;
2027
2028 /* Recompute the frequencies and counts of outgoing edges. */
2029 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2030 {
2031 edge s2;
2032 edge_iterator ei;
2033 basic_block d = s->dest;
2034
2035 if (FORWARDER_BLOCK_P (d))
2036 d = single_succ (d);
2037
2038 FOR_EACH_EDGE (s2, ei, src1->succs)
2039 {
2040 basic_block d2 = s2->dest;
2041 if (FORWARDER_BLOCK_P (d2))
2042 d2 = single_succ (d2);
2043 if (d == d2)
2044 break;
2045 }
2046
2047 s->count += s2->count;
2048
2049 /* Take care to update possible forwarder blocks. We verified
2050 that there is no more than one in the chain, so we can't run
2051 into infinite loop. */
2052 if (FORWARDER_BLOCK_P (s->dest))
2053 {
2054 single_succ_edge (s->dest)->count += s2->count;
2055 s->dest->count += s2->count;
2056 s->dest->frequency += EDGE_FREQUENCY (s);
2057 }
2058
2059 if (FORWARDER_BLOCK_P (s2->dest))
2060 {
2061 single_succ_edge (s2->dest)->count -= s2->count;
2062 if (single_succ_edge (s2->dest)->count < 0)
2063 single_succ_edge (s2->dest)->count = 0;
2064 s2->dest->count -= s2->count;
2065 s2->dest->frequency -= EDGE_FREQUENCY (s);
2066 if (s2->dest->frequency < 0)
2067 s2->dest->frequency = 0;
2068 if (s2->dest->count < 0)
2069 s2->dest->count = 0;
2070 }
2071
2072 if (!redirect_edges_to->frequency && !src1->frequency)
2073 s->probability = (s->probability + s2->probability) / 2;
2074 else
2075 s->probability
2076 = ((s->probability * redirect_edges_to->frequency +
2077 s2->probability * src1->frequency)
2078 / (redirect_edges_to->frequency + src1->frequency));
2079 }
2080
2081 /* Adjust count and frequency for the block. An earlier jump
2082 threading pass may have left the profile in an inconsistent
2083 state (see update_bb_profile_for_threading) so we must be
2084 prepared for overflows. */
2085 tmp = redirect_to;
2086 do
2087 {
2088 tmp->count += src1->count;
2089 tmp->frequency += src1->frequency;
2090 if (tmp->frequency > BB_FREQ_MAX)
2091 tmp->frequency = BB_FREQ_MAX;
2092 if (tmp == redirect_edges_to)
2093 break;
2094 tmp = find_fallthru_edge (tmp->succs)->dest;
2095 }
2096 while (true);
2097 update_br_prob_note (redirect_edges_to);
2098
2099 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2100
2101 /* Skip possible basic block header. */
2102 if (LABEL_P (newpos1))
2103 newpos1 = NEXT_INSN (newpos1);
2104
2105 while (DEBUG_INSN_P (newpos1))
2106 newpos1 = NEXT_INSN (newpos1);
2107
2108 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2109 newpos1 = NEXT_INSN (newpos1);
2110
2111 while (DEBUG_INSN_P (newpos1))
2112 newpos1 = NEXT_INSN (newpos1);
2113
2114 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2115 to_remove = single_succ (redirect_from);
2116
2117 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2118 delete_basic_block (to_remove);
2119
2120 update_forwarder_flag (redirect_from);
2121 if (redirect_to != src2)
2122 update_forwarder_flag (src2);
2123
2124 return true;
2125 }
2126
2127 /* Search the predecessors of BB for common insn sequences. When found,
2128 share code between them by redirecting control flow. Return true if
2129 any changes made. */
2130
2131 static bool
2132 try_crossjump_bb (int mode, basic_block bb)
2133 {
2134 edge e, e2, fallthru;
2135 bool changed;
2136 unsigned max, ix, ix2;
2137
2138 /* Nothing to do if there is not at least two incoming edges. */
2139 if (EDGE_COUNT (bb->preds) < 2)
2140 return false;
2141
2142 /* Don't crossjump if this block ends in a computed jump,
2143 unless we are optimizing for size. */
2144 if (optimize_bb_for_size_p (bb)
2145 && bb != EXIT_BLOCK_PTR
2146 && computed_jump_p (BB_END (bb)))
2147 return false;
2148
2149 /* If we are partitioning hot/cold basic blocks, we don't want to
2150 mess up unconditional or indirect jumps that cross between hot
2151 and cold sections.
2152
2153 Basic block partitioning may result in some jumps that appear to
2154 be optimizable (or blocks that appear to be mergeable), but which really
2155 must be left untouched (they are required to make it safely across
2156 partition boundaries). See the comments at the top of
2157 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2158
2159 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2160 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2161 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2162 return false;
2163
2164 /* It is always cheapest to redirect a block that ends in a branch to
2165 a block that falls through into BB, as that adds no branches to the
2166 program. We'll try that combination first. */
2167 fallthru = NULL;
2168 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2169
2170 if (EDGE_COUNT (bb->preds) > max)
2171 return false;
2172
2173 fallthru = find_fallthru_edge (bb->preds);
2174
2175 changed = false;
2176 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2177 {
2178 e = EDGE_PRED (bb, ix);
2179 ix++;
2180
2181 /* As noted above, first try with the fallthru predecessor (or, a
2182 fallthru predecessor if we are in cfglayout mode). */
2183 if (fallthru)
2184 {
2185 /* Don't combine the fallthru edge into anything else.
2186 If there is a match, we'll do it the other way around. */
2187 if (e == fallthru)
2188 continue;
2189 /* If nothing changed since the last attempt, there is nothing
2190 we can do. */
2191 if (!first_pass
2192 && !((e->src->flags & BB_MODIFIED)
2193 || (fallthru->src->flags & BB_MODIFIED)))
2194 continue;
2195
2196 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2197 {
2198 changed = true;
2199 ix = 0;
2200 continue;
2201 }
2202 }
2203
2204 /* Non-obvious work limiting check: Recognize that we're going
2205 to call try_crossjump_bb on every basic block. So if we have
2206 two blocks with lots of outgoing edges (a switch) and they
2207 share lots of common destinations, then we would do the
2208 cross-jump check once for each common destination.
2209
2210 Now, if the blocks actually are cross-jump candidates, then
2211 all of their destinations will be shared. Which means that
2212 we only need check them for cross-jump candidacy once. We
2213 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2214 choosing to do the check from the block for which the edge
2215 in question is the first successor of A. */
2216 if (EDGE_SUCC (e->src, 0) != e)
2217 continue;
2218
2219 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2220 {
2221 e2 = EDGE_PRED (bb, ix2);
2222
2223 if (e2 == e)
2224 continue;
2225
2226 /* We've already checked the fallthru edge above. */
2227 if (e2 == fallthru)
2228 continue;
2229
2230 /* The "first successor" check above only prevents multiple
2231 checks of crossjump(A,B). In order to prevent redundant
2232 checks of crossjump(B,A), require that A be the block
2233 with the lowest index. */
2234 if (e->src->index > e2->src->index)
2235 continue;
2236
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 || (e2->src->flags & BB_MODIFIED)))
2242 continue;
2243
2244 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2245 direction. */
2246 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2247 {
2248 changed = true;
2249 ix = 0;
2250 break;
2251 }
2252 }
2253 }
2254
2255 if (changed)
2256 crossjumps_occured = true;
2257
2258 return changed;
2259 }
2260
2261 /* Search the successors of BB for common insn sequences. When found,
2262 share code between them by moving it across the basic block
2263 boundary. Return true if any changes made. */
2264
2265 static bool
2266 try_head_merge_bb (basic_block bb)
2267 {
2268 basic_block final_dest_bb = NULL;
2269 int max_match = INT_MAX;
2270 edge e0;
2271 rtx *headptr, *currptr, *nextptr;
2272 bool changed, moveall;
2273 unsigned ix;
2274 rtx e0_last_head, cond, move_before;
2275 unsigned nedges = EDGE_COUNT (bb->succs);
2276 rtx jump = BB_END (bb);
2277 regset live, live_union;
2278
2279 /* Nothing to do if there is not at least two outgoing edges. */
2280 if (nedges < 2)
2281 return false;
2282
2283 /* Don't crossjump if this block ends in a computed jump,
2284 unless we are optimizing for size. */
2285 if (optimize_bb_for_size_p (bb)
2286 && bb != EXIT_BLOCK_PTR
2287 && computed_jump_p (BB_END (bb)))
2288 return false;
2289
2290 cond = get_condition (jump, &move_before, true, false);
2291 if (cond == NULL_RTX)
2292 {
2293 #ifdef HAVE_cc0
2294 if (reg_mentioned_p (cc0_rtx, jump))
2295 move_before = prev_nonnote_nondebug_insn (jump);
2296 else
2297 #endif
2298 move_before = jump;
2299 }
2300
2301 for (ix = 0; ix < nedges; ix++)
2302 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR)
2303 return false;
2304
2305 for (ix = 0; ix < nedges; ix++)
2306 {
2307 edge e = EDGE_SUCC (bb, ix);
2308 basic_block other_bb = e->dest;
2309
2310 if (df_get_bb_dirty (other_bb))
2311 {
2312 block_was_dirty = true;
2313 return false;
2314 }
2315
2316 if (e->flags & EDGE_ABNORMAL)
2317 return false;
2318
2319 /* Normally, all destination blocks must only be reachable from this
2320 block, i.e. they must have one incoming edge.
2321
2322 There is one special case we can handle, that of multiple consecutive
2323 jumps where the first jumps to one of the targets of the second jump.
2324 This happens frequently in switch statements for default labels.
2325 The structure is as follows:
2326 FINAL_DEST_BB
2327 ....
2328 if (cond) jump A;
2329 fall through
2330 BB
2331 jump with targets A, B, C, D...
2332 A
2333 has two incoming edges, from FINAL_DEST_BB and BB
2334
2335 In this case, we can try to move the insns through BB and into
2336 FINAL_DEST_BB. */
2337 if (EDGE_COUNT (other_bb->preds) != 1)
2338 {
2339 edge incoming_edge, incoming_bb_other_edge;
2340 edge_iterator ei;
2341
2342 if (final_dest_bb != NULL
2343 || EDGE_COUNT (other_bb->preds) != 2)
2344 return false;
2345
2346 /* We must be able to move the insns across the whole block. */
2347 move_before = BB_HEAD (bb);
2348 while (!NONDEBUG_INSN_P (move_before))
2349 move_before = NEXT_INSN (move_before);
2350
2351 if (EDGE_COUNT (bb->preds) != 1)
2352 return false;
2353 incoming_edge = EDGE_PRED (bb, 0);
2354 final_dest_bb = incoming_edge->src;
2355 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2356 return false;
2357 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2358 if (incoming_bb_other_edge != incoming_edge)
2359 break;
2360 if (incoming_bb_other_edge->dest != other_bb)
2361 return false;
2362 }
2363 }
2364
2365 e0 = EDGE_SUCC (bb, 0);
2366 e0_last_head = NULL_RTX;
2367 changed = false;
2368
2369 for (ix = 1; ix < nedges; ix++)
2370 {
2371 edge e = EDGE_SUCC (bb, ix);
2372 rtx e0_last, e_last;
2373 int nmatch;
2374
2375 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2376 &e0_last, &e_last, 0);
2377 if (nmatch == 0)
2378 return false;
2379
2380 if (nmatch < max_match)
2381 {
2382 max_match = nmatch;
2383 e0_last_head = e0_last;
2384 }
2385 }
2386
2387 /* If we matched an entire block, we probably have to avoid moving the
2388 last insn. */
2389 if (max_match > 0
2390 && e0_last_head == BB_END (e0->dest)
2391 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2392 || control_flow_insn_p (e0_last_head)))
2393 {
2394 max_match--;
2395 if (max_match == 0)
2396 return false;
2397 do
2398 e0_last_head = prev_real_insn (e0_last_head);
2399 while (DEBUG_INSN_P (e0_last_head));
2400 }
2401
2402 if (max_match == 0)
2403 return false;
2404
2405 /* We must find a union of the live registers at each of the end points. */
2406 live = BITMAP_ALLOC (NULL);
2407 live_union = BITMAP_ALLOC (NULL);
2408
2409 currptr = XNEWVEC (rtx, nedges);
2410 headptr = XNEWVEC (rtx, nedges);
2411 nextptr = XNEWVEC (rtx, nedges);
2412
2413 for (ix = 0; ix < nedges; ix++)
2414 {
2415 int j;
2416 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2417 rtx head = BB_HEAD (merge_bb);
2418
2419 while (!NONDEBUG_INSN_P (head))
2420 head = NEXT_INSN (head);
2421 headptr[ix] = head;
2422 currptr[ix] = head;
2423
2424 /* Compute the end point and live information */
2425 for (j = 1; j < max_match; j++)
2426 do
2427 head = NEXT_INSN (head);
2428 while (!NONDEBUG_INSN_P (head));
2429 simulate_backwards_to_point (merge_bb, live, head);
2430 IOR_REG_SET (live_union, live);
2431 }
2432
2433 /* If we're moving across two blocks, verify the validity of the
2434 first move, then adjust the target and let the loop below deal
2435 with the final move. */
2436 if (final_dest_bb != NULL)
2437 {
2438 rtx move_upto;
2439
2440 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2441 jump, e0->dest, live_union,
2442 NULL, &move_upto);
2443 if (!moveall)
2444 {
2445 if (move_upto == NULL_RTX)
2446 goto out;
2447
2448 while (e0_last_head != move_upto)
2449 {
2450 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2451 live_union);
2452 e0_last_head = PREV_INSN (e0_last_head);
2453 }
2454 }
2455 if (e0_last_head == NULL_RTX)
2456 goto out;
2457
2458 jump = BB_END (final_dest_bb);
2459 cond = get_condition (jump, &move_before, true, false);
2460 if (cond == NULL_RTX)
2461 {
2462 #ifdef HAVE_cc0
2463 if (reg_mentioned_p (cc0_rtx, jump))
2464 move_before = prev_nonnote_nondebug_insn (jump);
2465 else
2466 #endif
2467 move_before = jump;
2468 }
2469 }
2470
2471 do
2472 {
2473 rtx move_upto;
2474 moveall = can_move_insns_across (currptr[0], e0_last_head,
2475 move_before, jump, e0->dest, live_union,
2476 NULL, &move_upto);
2477 if (!moveall && move_upto == NULL_RTX)
2478 {
2479 if (jump == move_before)
2480 break;
2481
2482 /* Try again, using a different insertion point. */
2483 move_before = jump;
2484
2485 #ifdef HAVE_cc0
2486 /* Don't try moving before a cc0 user, as that may invalidate
2487 the cc0. */
2488 if (reg_mentioned_p (cc0_rtx, jump))
2489 break;
2490 #endif
2491
2492 continue;
2493 }
2494
2495 if (final_dest_bb && !moveall)
2496 /* We haven't checked whether a partial move would be OK for the first
2497 move, so we have to fail this case. */
2498 break;
2499
2500 changed = true;
2501 for (;;)
2502 {
2503 if (currptr[0] == move_upto)
2504 break;
2505 for (ix = 0; ix < nedges; ix++)
2506 {
2507 rtx curr = currptr[ix];
2508 do
2509 curr = NEXT_INSN (curr);
2510 while (!NONDEBUG_INSN_P (curr));
2511 currptr[ix] = curr;
2512 }
2513 }
2514
2515 /* If we can't currently move all of the identical insns, remember
2516 each insn after the range that we'll merge. */
2517 if (!moveall)
2518 for (ix = 0; ix < nedges; ix++)
2519 {
2520 rtx curr = currptr[ix];
2521 do
2522 curr = NEXT_INSN (curr);
2523 while (!NONDEBUG_INSN_P (curr));
2524 nextptr[ix] = curr;
2525 }
2526
2527 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2528 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2529 if (final_dest_bb != NULL)
2530 df_set_bb_dirty (final_dest_bb);
2531 df_set_bb_dirty (bb);
2532 for (ix = 1; ix < nedges; ix++)
2533 {
2534 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2535 delete_insn_chain (headptr[ix], currptr[ix], false);
2536 }
2537 if (!moveall)
2538 {
2539 if (jump == move_before)
2540 break;
2541
2542 /* For the unmerged insns, try a different insertion point. */
2543 move_before = jump;
2544
2545 #ifdef HAVE_cc0
2546 /* Don't try moving before a cc0 user, as that may invalidate
2547 the cc0. */
2548 if (reg_mentioned_p (cc0_rtx, jump))
2549 break;
2550 #endif
2551
2552 for (ix = 0; ix < nedges; ix++)
2553 currptr[ix] = headptr[ix] = nextptr[ix];
2554 }
2555 }
2556 while (!moveall);
2557
2558 out:
2559 free (currptr);
2560 free (headptr);
2561 free (nextptr);
2562
2563 crossjumps_occured |= changed;
2564
2565 return changed;
2566 }
2567
2568 /* Return true if BB contains just bb note, or bb note followed
2569 by only DEBUG_INSNs. */
2570
2571 static bool
2572 trivially_empty_bb_p (basic_block bb)
2573 {
2574 rtx insn = BB_END (bb);
2575
2576 while (1)
2577 {
2578 if (insn == BB_HEAD (bb))
2579 return true;
2580 if (!DEBUG_INSN_P (insn))
2581 return false;
2582 insn = PREV_INSN (insn);
2583 }
2584 }
2585
2586 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2587 instructions etc. Return nonzero if changes were made. */
2588
2589 static bool
2590 try_optimize_cfg (int mode)
2591 {
2592 bool changed_overall = false;
2593 bool changed;
2594 int iterations = 0;
2595 basic_block bb, b, next;
2596
2597 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2598 clear_bb_flags ();
2599
2600 crossjumps_occured = false;
2601
2602 FOR_EACH_BB (bb)
2603 update_forwarder_flag (bb);
2604
2605 if (! targetm.cannot_modify_jumps_p ())
2606 {
2607 first_pass = true;
2608 /* Attempt to merge blocks as made possible by edge removal. If
2609 a block has only one successor, and the successor has only
2610 one predecessor, they may be combined. */
2611 do
2612 {
2613 block_was_dirty = false;
2614 changed = false;
2615 iterations++;
2616
2617 if (dump_file)
2618 fprintf (dump_file,
2619 "\n\ntry_optimize_cfg iteration %i\n\n",
2620 iterations);
2621
2622 for (b = ENTRY_BLOCK_PTR->next_bb; b != EXIT_BLOCK_PTR;)
2623 {
2624 basic_block c;
2625 edge s;
2626 bool changed_here = false;
2627
2628 /* Delete trivially dead basic blocks. This is either
2629 blocks with no predecessors, or empty blocks with no
2630 successors. However if the empty block with no
2631 successors is the successor of the ENTRY_BLOCK, it is
2632 kept. This ensures that the ENTRY_BLOCK will have a
2633 successor which is a precondition for many RTL
2634 passes. Empty blocks may result from expanding
2635 __builtin_unreachable (). */
2636 if (EDGE_COUNT (b->preds) == 0
2637 || (EDGE_COUNT (b->succs) == 0
2638 && trivially_empty_bb_p (b)
2639 && single_succ_edge (ENTRY_BLOCK_PTR)->dest != b))
2640 {
2641 c = b->prev_bb;
2642 if (EDGE_COUNT (b->preds) > 0)
2643 {
2644 edge e;
2645 edge_iterator ei;
2646
2647 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2648 {
2649 if (BB_FOOTER (b)
2650 && BARRIER_P (BB_FOOTER (b)))
2651 FOR_EACH_EDGE (e, ei, b->preds)
2652 if ((e->flags & EDGE_FALLTHRU)
2653 && BB_FOOTER (e->src) == NULL)
2654 {
2655 if (BB_FOOTER (b))
2656 {
2657 BB_FOOTER (e->src) = BB_FOOTER (b);
2658 BB_FOOTER (b) = NULL;
2659 }
2660 else
2661 {
2662 start_sequence ();
2663 BB_FOOTER (e->src) = emit_barrier ();
2664 end_sequence ();
2665 }
2666 }
2667 }
2668 else
2669 {
2670 rtx last = get_last_bb_insn (b);
2671 if (last && BARRIER_P (last))
2672 FOR_EACH_EDGE (e, ei, b->preds)
2673 if ((e->flags & EDGE_FALLTHRU))
2674 emit_barrier_after (BB_END (e->src));
2675 }
2676 }
2677 delete_basic_block (b);
2678 changed = true;
2679 /* Avoid trying to remove ENTRY_BLOCK_PTR. */
2680 b = (c == ENTRY_BLOCK_PTR ? c->next_bb : c);
2681 continue;
2682 }
2683
2684 /* Remove code labels no longer used. */
2685 if (single_pred_p (b)
2686 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2687 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2688 && LABEL_P (BB_HEAD (b))
2689 /* If the previous block ends with a branch to this
2690 block, we can't delete the label. Normally this
2691 is a condjump that is yet to be simplified, but
2692 if CASE_DROPS_THRU, this can be a tablejump with
2693 some element going to the same place as the
2694 default (fallthru). */
2695 && (single_pred (b) == ENTRY_BLOCK_PTR
2696 || !JUMP_P (BB_END (single_pred (b)))
2697 || ! label_is_jump_target_p (BB_HEAD (b),
2698 BB_END (single_pred (b)))))
2699 {
2700 delete_insn (BB_HEAD (b));
2701 if (dump_file)
2702 fprintf (dump_file, "Deleted label in block %i.\n",
2703 b->index);
2704 }
2705
2706 /* If we fall through an empty block, we can remove it. */
2707 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2708 && single_pred_p (b)
2709 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2710 && !LABEL_P (BB_HEAD (b))
2711 && FORWARDER_BLOCK_P (b)
2712 /* Note that forwarder_block_p true ensures that
2713 there is a successor for this block. */
2714 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2715 && n_basic_blocks > NUM_FIXED_BLOCKS + 1)
2716 {
2717 if (dump_file)
2718 fprintf (dump_file,
2719 "Deleting fallthru block %i.\n",
2720 b->index);
2721
2722 c = b->prev_bb == ENTRY_BLOCK_PTR ? b->next_bb : b->prev_bb;
2723 redirect_edge_succ_nodup (single_pred_edge (b),
2724 single_succ (b));
2725 delete_basic_block (b);
2726 changed = true;
2727 b = c;
2728 continue;
2729 }
2730
2731 /* Merge B with its single successor, if any. */
2732 if (single_succ_p (b)
2733 && (s = single_succ_edge (b))
2734 && !(s->flags & EDGE_COMPLEX)
2735 && (c = s->dest) != EXIT_BLOCK_PTR
2736 && single_pred_p (c)
2737 && b != c)
2738 {
2739 /* When not in cfg_layout mode use code aware of reordering
2740 INSN. This code possibly creates new basic blocks so it
2741 does not fit merge_blocks interface and is kept here in
2742 hope that it will become useless once more of compiler
2743 is transformed to use cfg_layout mode. */
2744
2745 if ((mode & CLEANUP_CFGLAYOUT)
2746 && can_merge_blocks_p (b, c))
2747 {
2748 merge_blocks (b, c);
2749 update_forwarder_flag (b);
2750 changed_here = true;
2751 }
2752 else if (!(mode & CLEANUP_CFGLAYOUT)
2753 /* If the jump insn has side effects,
2754 we can't kill the edge. */
2755 && (!JUMP_P (BB_END (b))
2756 || (reload_completed
2757 ? simplejump_p (BB_END (b))
2758 : (onlyjump_p (BB_END (b))
2759 && !tablejump_p (BB_END (b),
2760 NULL, NULL))))
2761 && (next = merge_blocks_move (s, b, c, mode)))
2762 {
2763 b = next;
2764 changed_here = true;
2765 }
2766 }
2767
2768 /* Simplify branch over branch. */
2769 if ((mode & CLEANUP_EXPENSIVE)
2770 && !(mode & CLEANUP_CFGLAYOUT)
2771 && try_simplify_condjump (b))
2772 changed_here = true;
2773
2774 /* If B has a single outgoing edge, but uses a
2775 non-trivial jump instruction without side-effects, we
2776 can either delete the jump entirely, or replace it
2777 with a simple unconditional jump. */
2778 if (single_succ_p (b)
2779 && single_succ (b) != EXIT_BLOCK_PTR
2780 && onlyjump_p (BB_END (b))
2781 && !find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX)
2782 && try_redirect_by_replacing_jump (single_succ_edge (b),
2783 single_succ (b),
2784 (mode & CLEANUP_CFGLAYOUT) != 0))
2785 {
2786 update_forwarder_flag (b);
2787 changed_here = true;
2788 }
2789
2790 /* Simplify branch to branch. */
2791 if (try_forward_edges (mode, b))
2792 {
2793 update_forwarder_flag (b);
2794 changed_here = true;
2795 }
2796
2797 /* Look for shared code between blocks. */
2798 if ((mode & CLEANUP_CROSSJUMP)
2799 && try_crossjump_bb (mode, b))
2800 changed_here = true;
2801
2802 if ((mode & CLEANUP_CROSSJUMP)
2803 /* This can lengthen register lifetimes. Do it only after
2804 reload. */
2805 && reload_completed
2806 && try_head_merge_bb (b))
2807 changed_here = true;
2808
2809 /* Don't get confused by the index shift caused by
2810 deleting blocks. */
2811 if (!changed_here)
2812 b = b->next_bb;
2813 else
2814 changed = true;
2815 }
2816
2817 if ((mode & CLEANUP_CROSSJUMP)
2818 && try_crossjump_bb (mode, EXIT_BLOCK_PTR))
2819 changed = true;
2820
2821 if (block_was_dirty)
2822 {
2823 /* This should only be set by head-merging. */
2824 gcc_assert (mode & CLEANUP_CROSSJUMP);
2825 df_analyze ();
2826 }
2827
2828 if (changed)
2829 {
2830 /* Edge forwarding in particular can cause hot blocks previously
2831 reached by both hot and cold blocks to become dominated only
2832 by cold blocks. This will cause the verification below to fail,
2833 and lead to now cold code in the hot section. This is not easy
2834 to detect and fix during edge forwarding, and in some cases
2835 is only visible after newly unreachable blocks are deleted,
2836 which will be done in fixup_partitions. */
2837 fixup_partitions ();
2838
2839 #ifdef ENABLE_CHECKING
2840 verify_flow_info ();
2841 #endif
2842 }
2843
2844 changed_overall |= changed;
2845 first_pass = false;
2846 }
2847 while (changed);
2848 }
2849
2850 FOR_ALL_BB (b)
2851 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2852
2853 return changed_overall;
2854 }
2855 \f
2856 /* Delete all unreachable basic blocks. */
2857
2858 bool
2859 delete_unreachable_blocks (void)
2860 {
2861 bool changed = false;
2862 basic_block b, prev_bb;
2863
2864 find_unreachable_blocks ();
2865
2866 /* When we're in GIMPLE mode and there may be debug insns, we should
2867 delete blocks in reverse dominator order, so as to get a chance
2868 to substitute all released DEFs into debug stmts. If we don't
2869 have dominators information, walking blocks backward gets us a
2870 better chance of retaining most debug information than
2871 otherwise. */
2872 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2873 && dom_info_available_p (CDI_DOMINATORS))
2874 {
2875 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2876 {
2877 prev_bb = b->prev_bb;
2878
2879 if (!(b->flags & BB_REACHABLE))
2880 {
2881 /* Speed up the removal of blocks that don't dominate
2882 others. Walking backwards, this should be the common
2883 case. */
2884 if (!first_dom_son (CDI_DOMINATORS, b))
2885 delete_basic_block (b);
2886 else
2887 {
2888 vec<basic_block> h
2889 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2890
2891 while (h.length ())
2892 {
2893 b = h.pop ();
2894
2895 prev_bb = b->prev_bb;
2896
2897 gcc_assert (!(b->flags & BB_REACHABLE));
2898
2899 delete_basic_block (b);
2900 }
2901
2902 h.release ();
2903 }
2904
2905 changed = true;
2906 }
2907 }
2908 }
2909 else
2910 {
2911 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2912 {
2913 prev_bb = b->prev_bb;
2914
2915 if (!(b->flags & BB_REACHABLE))
2916 {
2917 delete_basic_block (b);
2918 changed = true;
2919 }
2920 }
2921 }
2922
2923 if (changed)
2924 tidy_fallthru_edges ();
2925 return changed;
2926 }
2927
2928 /* Delete any jump tables never referenced. We can't delete them at the
2929 time of removing tablejump insn as they are referenced by the preceding
2930 insns computing the destination, so we delay deleting and garbagecollect
2931 them once life information is computed. */
2932 void
2933 delete_dead_jumptables (void)
2934 {
2935 basic_block bb;
2936
2937 /* A dead jump table does not belong to any basic block. Scan insns
2938 between two adjacent basic blocks. */
2939 FOR_EACH_BB (bb)
2940 {
2941 rtx insn, next;
2942
2943 for (insn = NEXT_INSN (BB_END (bb));
2944 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
2945 insn = next)
2946 {
2947 next = NEXT_INSN (insn);
2948 if (LABEL_P (insn)
2949 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
2950 && JUMP_TABLE_DATA_P (next))
2951 {
2952 rtx label = insn, jump = next;
2953
2954 if (dump_file)
2955 fprintf (dump_file, "Dead jumptable %i removed\n",
2956 INSN_UID (insn));
2957
2958 next = NEXT_INSN (next);
2959 delete_insn (jump);
2960 delete_insn (label);
2961 }
2962 }
2963 }
2964 }
2965
2966 \f
2967 /* Tidy the CFG by deleting unreachable code and whatnot. */
2968
2969 bool
2970 cleanup_cfg (int mode)
2971 {
2972 bool changed = false;
2973
2974 /* Set the cfglayout mode flag here. We could update all the callers
2975 but that is just inconvenient, especially given that we eventually
2976 want to have cfglayout mode as the default. */
2977 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2978 mode |= CLEANUP_CFGLAYOUT;
2979
2980 timevar_push (TV_CLEANUP_CFG);
2981 if (delete_unreachable_blocks ())
2982 {
2983 changed = true;
2984 /* We've possibly created trivially dead code. Cleanup it right
2985 now to introduce more opportunities for try_optimize_cfg. */
2986 if (!(mode & (CLEANUP_NO_INSN_DEL))
2987 && !reload_completed)
2988 delete_trivially_dead_insns (get_insns (), max_reg_num ());
2989 }
2990
2991 compact_blocks ();
2992
2993 /* To tail-merge blocks ending in the same noreturn function (e.g.
2994 a call to abort) we have to insert fake edges to exit. Do this
2995 here once. The fake edges do not interfere with any other CFG
2996 cleanups. */
2997 if (mode & CLEANUP_CROSSJUMP)
2998 add_noreturn_fake_exit_edges ();
2999
3000 if (!dbg_cnt (cfg_cleanup))
3001 return changed;
3002
3003 while (try_optimize_cfg (mode))
3004 {
3005 delete_unreachable_blocks (), changed = true;
3006 if (!(mode & CLEANUP_NO_INSN_DEL))
3007 {
3008 /* Try to remove some trivially dead insns when doing an expensive
3009 cleanup. But delete_trivially_dead_insns doesn't work after
3010 reload (it only handles pseudos) and run_fast_dce is too costly
3011 to run in every iteration.
3012
3013 For effective cross jumping, we really want to run a fast DCE to
3014 clean up any dead conditions, or they get in the way of performing
3015 useful tail merges.
3016
3017 Other transformations in cleanup_cfg are not so sensitive to dead
3018 code, so delete_trivially_dead_insns or even doing nothing at all
3019 is good enough. */
3020 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3021 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3022 break;
3023 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
3024 run_fast_dce ();
3025 }
3026 else
3027 break;
3028 }
3029
3030 if (mode & CLEANUP_CROSSJUMP)
3031 remove_fake_exit_edges ();
3032
3033 /* Don't call delete_dead_jumptables in cfglayout mode, because
3034 that function assumes that jump tables are in the insns stream.
3035 But we also don't _have_ to delete dead jumptables in cfglayout
3036 mode because we shouldn't even be looking at things that are
3037 not in a basic block. Dead jumptables are cleaned up when
3038 going out of cfglayout mode. */
3039 if (!(mode & CLEANUP_CFGLAYOUT))
3040 delete_dead_jumptables ();
3041
3042 /* ??? We probably do this way too often. */
3043 if (current_loops
3044 && (changed
3045 || (mode & CLEANUP_CFG_CHANGED)))
3046 {
3047 timevar_push (TV_REPAIR_LOOPS);
3048 /* The above doesn't preserve dominance info if available. */
3049 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3050 calculate_dominance_info (CDI_DOMINATORS);
3051 fix_loop_structure (NULL);
3052 free_dominance_info (CDI_DOMINATORS);
3053 timevar_pop (TV_REPAIR_LOOPS);
3054 }
3055
3056 timevar_pop (TV_CLEANUP_CFG);
3057
3058 return changed;
3059 }
3060 \f
3061 static unsigned int
3062 execute_jump (void)
3063 {
3064 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3065 if (dump_file)
3066 dump_flow_info (dump_file, dump_flags);
3067 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3068 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3069 return 0;
3070 }
3071
3072 namespace {
3073
3074 const pass_data pass_data_jump =
3075 {
3076 RTL_PASS, /* type */
3077 "jump", /* name */
3078 OPTGROUP_NONE, /* optinfo_flags */
3079 false, /* has_gate */
3080 true, /* has_execute */
3081 TV_JUMP, /* tv_id */
3082 0, /* properties_required */
3083 0, /* properties_provided */
3084 0, /* properties_destroyed */
3085 0, /* todo_flags_start */
3086 TODO_verify_rtl_sharing, /* todo_flags_finish */
3087 };
3088
3089 class pass_jump : public rtl_opt_pass
3090 {
3091 public:
3092 pass_jump(gcc::context *ctxt)
3093 : rtl_opt_pass(pass_data_jump, ctxt)
3094 {}
3095
3096 /* opt_pass methods: */
3097 unsigned int execute () { return execute_jump (); }
3098
3099 }; // class pass_jump
3100
3101 } // anon namespace
3102
3103 rtl_opt_pass *
3104 make_pass_jump (gcc::context *ctxt)
3105 {
3106 return new pass_jump (ctxt);
3107 }
3108 \f
3109 static unsigned int
3110 execute_jump2 (void)
3111 {
3112 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3113 return 0;
3114 }
3115
3116 namespace {
3117
3118 const pass_data pass_data_jump2 =
3119 {
3120 RTL_PASS, /* type */
3121 "jump2", /* name */
3122 OPTGROUP_NONE, /* optinfo_flags */
3123 false, /* has_gate */
3124 true, /* has_execute */
3125 TV_JUMP, /* tv_id */
3126 0, /* properties_required */
3127 0, /* properties_provided */
3128 0, /* properties_destroyed */
3129 0, /* todo_flags_start */
3130 TODO_verify_rtl_sharing, /* todo_flags_finish */
3131 };
3132
3133 class pass_jump2 : public rtl_opt_pass
3134 {
3135 public:
3136 pass_jump2(gcc::context *ctxt)
3137 : rtl_opt_pass(pass_data_jump2, ctxt)
3138 {}
3139
3140 /* opt_pass methods: */
3141 unsigned int execute () { return execute_jump2 (); }
3142
3143 }; // class pass_jump2
3144
3145 } // anon namespace
3146
3147 rtl_opt_pass *
3148 make_pass_jump2 (gcc::context *ctxt)
3149 {
3150 return new pass_jump2 (ctxt);
3151 }