re PR fortran/58771 (ICE in transfer_expr, at fortran/trans-io.c:2164)
[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 "tree.h"
38 #include "hard-reg-set.h"
39 #include "regs.h"
40 #include "insn-config.h"
41 #include "flags.h"
42 #include "recog.h"
43 #include "diagnostic-core.h"
44 #include "cselib.h"
45 #include "params.h"
46 #include "tm_p.h"
47 #include "target.h"
48 #include "function.h" /* For inline functions in emit-rtl.h they need crtl. */
49 #include "emit-rtl.h"
50 #include "tree-pass.h"
51 #include "cfgloop.h"
52 #include "expr.h"
53 #include "df.h"
54 #include "dce.h"
55 #include "dbgcnt.h"
56
57 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
58
59 /* Set to true when we are running first pass of try_optimize_cfg loop. */
60 static bool first_pass;
61
62 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
63 static bool crossjumps_occured;
64
65 /* Set to true if we couldn't run an optimization due to stale liveness
66 information; we should run df_analyze to enable more opportunities. */
67 static bool block_was_dirty;
68
69 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
70 static bool try_crossjump_bb (int, basic_block);
71 static bool outgoing_edges_match (int, basic_block, basic_block);
72 static enum replace_direction old_insns_match_p (int, rtx, rtx);
73
74 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
75 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
76 static bool try_optimize_cfg (int);
77 static bool try_simplify_condjump (basic_block);
78 static bool try_forward_edges (int, basic_block);
79 static edge thread_jump (edge, basic_block);
80 static bool mark_effect (rtx, bitmap);
81 static void notice_new_block (basic_block);
82 static void update_forwarder_flag (basic_block);
83 static int mentions_nonequal_regs (rtx *, void *);
84 static void merge_memattrs (rtx, rtx);
85 \f
86 /* Set flags for newly created block. */
87
88 static void
89 notice_new_block (basic_block bb)
90 {
91 if (!bb)
92 return;
93
94 if (forwarder_block_p (bb))
95 bb->flags |= BB_FORWARDER_BLOCK;
96 }
97
98 /* Recompute forwarder flag after block has been modified. */
99
100 static void
101 update_forwarder_flag (basic_block bb)
102 {
103 if (forwarder_block_p (bb))
104 bb->flags |= BB_FORWARDER_BLOCK;
105 else
106 bb->flags &= ~BB_FORWARDER_BLOCK;
107 }
108 \f
109 /* Simplify a conditional jump around an unconditional jump.
110 Return true if something changed. */
111
112 static bool
113 try_simplify_condjump (basic_block cbranch_block)
114 {
115 basic_block jump_block, jump_dest_block, cbranch_dest_block;
116 edge cbranch_jump_edge, cbranch_fallthru_edge;
117 rtx cbranch_insn;
118
119 /* Verify that there are exactly two successors. */
120 if (EDGE_COUNT (cbranch_block->succs) != 2)
121 return false;
122
123 /* Verify that we've got a normal conditional branch at the end
124 of the block. */
125 cbranch_insn = BB_END (cbranch_block);
126 if (!any_condjump_p (cbranch_insn))
127 return false;
128
129 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
130 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
131
132 /* The next block must not have multiple predecessors, must not
133 be the last block in the function, and must contain just the
134 unconditional jump. */
135 jump_block = cbranch_fallthru_edge->dest;
136 if (!single_pred_p (jump_block)
137 || jump_block->next_bb == EXIT_BLOCK_PTR
138 || !FORWARDER_BLOCK_P (jump_block))
139 return false;
140 jump_dest_block = single_succ (jump_block);
141
142 /* If we are partitioning hot/cold basic blocks, we don't want to
143 mess up unconditional or indirect jumps that cross between hot
144 and cold sections.
145
146 Basic block partitioning may result in some jumps that appear to
147 be optimizable (or blocks that appear to be mergeable), but which really
148 must be left untouched (they are required to make it safely across
149 partition boundaries). See the comments at the top of
150 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
151
152 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
153 || (cbranch_jump_edge->flags & EDGE_CROSSING))
154 return false;
155
156 /* The conditional branch must target the block after the
157 unconditional branch. */
158 cbranch_dest_block = cbranch_jump_edge->dest;
159
160 if (cbranch_dest_block == EXIT_BLOCK_PTR
161 || !can_fallthru (jump_block, cbranch_dest_block))
162 return false;
163
164 /* Invert the conditional branch. */
165 if (!invert_jump (cbranch_insn, block_label (jump_dest_block), 0))
166 return false;
167
168 if (dump_file)
169 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
170 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
171
172 /* Success. Update the CFG to match. Note that after this point
173 the edge variable names appear backwards; the redirection is done
174 this way to preserve edge profile data. */
175 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
176 cbranch_dest_block);
177 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
178 jump_dest_block);
179 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
180 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
181 update_br_prob_note (cbranch_block);
182
183 /* Delete the block with the unconditional jump, and clean up the mess. */
184 delete_basic_block (jump_block);
185 tidy_fallthru_edge (cbranch_jump_edge);
186 update_forwarder_flag (cbranch_block);
187
188 return true;
189 }
190 \f
191 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
192 on register. Used by jump threading. */
193
194 static bool
195 mark_effect (rtx exp, regset nonequal)
196 {
197 int regno;
198 rtx dest;
199 switch (GET_CODE (exp))
200 {
201 /* In case we do clobber the register, mark it as equal, as we know the
202 value is dead so it don't have to match. */
203 case CLOBBER:
204 if (REG_P (XEXP (exp, 0)))
205 {
206 dest = XEXP (exp, 0);
207 regno = REGNO (dest);
208 if (HARD_REGISTER_NUM_P (regno))
209 bitmap_clear_range (nonequal, regno,
210 hard_regno_nregs[regno][GET_MODE (dest)]);
211 else
212 bitmap_clear_bit (nonequal, regno);
213 }
214 return false;
215
216 case SET:
217 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
218 return false;
219 dest = SET_DEST (exp);
220 if (dest == pc_rtx)
221 return false;
222 if (!REG_P (dest))
223 return true;
224 regno = REGNO (dest);
225 if (HARD_REGISTER_NUM_P (regno))
226 bitmap_set_range (nonequal, regno,
227 hard_regno_nregs[regno][GET_MODE (dest)]);
228 else
229 bitmap_set_bit (nonequal, regno);
230 return false;
231
232 default:
233 return false;
234 }
235 }
236
237 /* Return nonzero if X is a register set in regset DATA.
238 Called via for_each_rtx. */
239 static int
240 mentions_nonequal_regs (rtx *x, void *data)
241 {
242 regset nonequal = (regset) data;
243 if (REG_P (*x))
244 {
245 int regno;
246
247 regno = REGNO (*x);
248 if (REGNO_REG_SET_P (nonequal, regno))
249 return 1;
250 if (regno < FIRST_PSEUDO_REGISTER)
251 {
252 int n = hard_regno_nregs[regno][GET_MODE (*x)];
253 while (--n > 0)
254 if (REGNO_REG_SET_P (nonequal, regno + n))
255 return 1;
256 }
257 }
258 return 0;
259 }
260 /* Attempt to prove that the basic block B will have no side effects and
261 always continues in the same edge if reached via E. Return the edge
262 if exist, NULL otherwise. */
263
264 static edge
265 thread_jump (edge e, basic_block b)
266 {
267 rtx set1, set2, cond1, cond2, insn;
268 enum rtx_code code1, code2, reversed_code2;
269 bool reverse1 = false;
270 unsigned i;
271 regset nonequal;
272 bool failed = false;
273 reg_set_iterator rsi;
274
275 if (b->flags & BB_NONTHREADABLE_BLOCK)
276 return NULL;
277
278 /* At the moment, we do handle only conditional jumps, but later we may
279 want to extend this code to tablejumps and others. */
280 if (EDGE_COUNT (e->src->succs) != 2)
281 return NULL;
282 if (EDGE_COUNT (b->succs) != 2)
283 {
284 b->flags |= BB_NONTHREADABLE_BLOCK;
285 return NULL;
286 }
287
288 /* Second branch must end with onlyjump, as we will eliminate the jump. */
289 if (!any_condjump_p (BB_END (e->src)))
290 return NULL;
291
292 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
293 {
294 b->flags |= BB_NONTHREADABLE_BLOCK;
295 return NULL;
296 }
297
298 set1 = pc_set (BB_END (e->src));
299 set2 = pc_set (BB_END (b));
300 if (((e->flags & EDGE_FALLTHRU) != 0)
301 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
302 reverse1 = true;
303
304 cond1 = XEXP (SET_SRC (set1), 0);
305 cond2 = XEXP (SET_SRC (set2), 0);
306 if (reverse1)
307 code1 = reversed_comparison_code (cond1, BB_END (e->src));
308 else
309 code1 = GET_CODE (cond1);
310
311 code2 = GET_CODE (cond2);
312 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
313
314 if (!comparison_dominates_p (code1, code2)
315 && !comparison_dominates_p (code1, reversed_code2))
316 return NULL;
317
318 /* Ensure that the comparison operators are equivalent.
319 ??? This is far too pessimistic. We should allow swapped operands,
320 different CCmodes, or for example comparisons for interval, that
321 dominate even when operands are not equivalent. */
322 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
323 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
324 return NULL;
325
326 /* Short circuit cases where block B contains some side effects, as we can't
327 safely bypass it. */
328 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
329 insn = NEXT_INSN (insn))
330 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
331 {
332 b->flags |= BB_NONTHREADABLE_BLOCK;
333 return NULL;
334 }
335
336 cselib_init (0);
337
338 /* First process all values computed in the source basic block. */
339 for (insn = NEXT_INSN (BB_HEAD (e->src));
340 insn != NEXT_INSN (BB_END (e->src));
341 insn = NEXT_INSN (insn))
342 if (INSN_P (insn))
343 cselib_process_insn (insn);
344
345 nonequal = BITMAP_ALLOC (NULL);
346 CLEAR_REG_SET (nonequal);
347
348 /* Now assume that we've continued by the edge E to B and continue
349 processing as if it were same basic block.
350 Our goal is to prove that whole block is an NOOP. */
351
352 for (insn = NEXT_INSN (BB_HEAD (b));
353 insn != NEXT_INSN (BB_END (b)) && !failed;
354 insn = NEXT_INSN (insn))
355 {
356 if (INSN_P (insn))
357 {
358 rtx pat = PATTERN (insn);
359
360 if (GET_CODE (pat) == PARALLEL)
361 {
362 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
363 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
364 }
365 else
366 failed |= mark_effect (pat, nonequal);
367 }
368
369 cselib_process_insn (insn);
370 }
371
372 /* Later we should clear nonequal of dead registers. So far we don't
373 have life information in cfg_cleanup. */
374 if (failed)
375 {
376 b->flags |= BB_NONTHREADABLE_BLOCK;
377 goto failed_exit;
378 }
379
380 /* cond2 must not mention any register that is not equal to the
381 former block. */
382 if (for_each_rtx (&cond2, mentions_nonequal_regs, nonequal))
383 goto failed_exit;
384
385 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
386 goto failed_exit;
387
388 BITMAP_FREE (nonequal);
389 cselib_finish ();
390 if ((comparison_dominates_p (code1, code2) != 0)
391 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
392 return BRANCH_EDGE (b);
393 else
394 return FALLTHRU_EDGE (b);
395
396 failed_exit:
397 BITMAP_FREE (nonequal);
398 cselib_finish ();
399 return NULL;
400 }
401 \f
402 /* Attempt to forward edges leaving basic block B.
403 Return true if successful. */
404
405 static bool
406 try_forward_edges (int mode, basic_block b)
407 {
408 bool changed = false;
409 edge_iterator ei;
410 edge e, *threaded_edges = NULL;
411
412 /* If we are partitioning hot/cold basic blocks, we don't want to
413 mess up unconditional or indirect jumps that cross between hot
414 and cold sections.
415
416 Basic block partitioning may result in some jumps that appear to
417 be optimizable (or blocks that appear to be mergeable), but which really
418 must be left untouched (they are required to make it safely across
419 partition boundaries). See the comments at the top of
420 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
421
422 if (find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX))
423 return false;
424
425 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
426 {
427 basic_block target, first;
428 int counter, goto_locus;
429 bool threaded = false;
430 int nthreaded_edges = 0;
431 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
432
433 /* Skip complex edges because we don't know how to update them.
434
435 Still handle fallthru edges, as we can succeed to forward fallthru
436 edge to the same place as the branch edge of conditional branch
437 and turn conditional branch to an unconditional branch. */
438 if (e->flags & EDGE_COMPLEX)
439 {
440 ei_next (&ei);
441 continue;
442 }
443
444 target = first = e->dest;
445 counter = NUM_FIXED_BLOCKS;
446 goto_locus = e->goto_locus;
447
448 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
449 up jumps that cross between hot/cold sections.
450
451 Basic block partitioning may result in some jumps that appear
452 to be optimizable (or blocks that appear to be mergeable), but which
453 really must be left untouched (they are required to make it safely
454 across partition boundaries). See the comments at the top of
455 bb-reorder.c:partition_hot_cold_basic_blocks for complete
456 details. */
457
458 if (first != EXIT_BLOCK_PTR
459 && find_reg_note (BB_END (first), REG_CROSSING_JUMP, NULL_RTX))
460 return changed;
461
462 while (counter < n_basic_blocks)
463 {
464 basic_block new_target = NULL;
465 bool new_target_threaded = false;
466 may_thread |= (target->flags & BB_MODIFIED) != 0;
467
468 if (FORWARDER_BLOCK_P (target)
469 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
470 && single_succ (target) != EXIT_BLOCK_PTR)
471 {
472 /* Bypass trivial infinite loops. */
473 new_target = single_succ (target);
474 if (target == new_target)
475 counter = n_basic_blocks;
476 else if (!optimize)
477 {
478 /* When not optimizing, ensure that edges or forwarder
479 blocks with different locus are not optimized out. */
480 int new_locus = single_succ_edge (target)->goto_locus;
481 int locus = goto_locus;
482
483 if (new_locus != UNKNOWN_LOCATION
484 && locus != UNKNOWN_LOCATION
485 && new_locus != locus)
486 new_target = NULL;
487 else
488 {
489 rtx last;
490
491 if (new_locus != UNKNOWN_LOCATION)
492 locus = new_locus;
493
494 last = BB_END (target);
495 if (DEBUG_INSN_P (last))
496 last = prev_nondebug_insn (last);
497
498 new_locus = last && INSN_P (last)
499 ? INSN_LOCATION (last) : 0;
500
501 if (new_locus != UNKNOWN_LOCATION
502 && locus != UNKNOWN_LOCATION
503 && new_locus != locus)
504 new_target = NULL;
505 else
506 {
507 if (new_locus != UNKNOWN_LOCATION)
508 locus = new_locus;
509
510 goto_locus = locus;
511 }
512 }
513 }
514 }
515
516 /* Allow to thread only over one edge at time to simplify updating
517 of probabilities. */
518 else if ((mode & CLEANUP_THREADING) && may_thread)
519 {
520 edge t = thread_jump (e, target);
521 if (t)
522 {
523 if (!threaded_edges)
524 threaded_edges = XNEWVEC (edge, n_basic_blocks);
525 else
526 {
527 int i;
528
529 /* Detect an infinite loop across blocks not
530 including the start block. */
531 for (i = 0; i < nthreaded_edges; ++i)
532 if (threaded_edges[i] == t)
533 break;
534 if (i < nthreaded_edges)
535 {
536 counter = n_basic_blocks;
537 break;
538 }
539 }
540
541 /* Detect an infinite loop across the start block. */
542 if (t->dest == b)
543 break;
544
545 gcc_assert (nthreaded_edges < n_basic_blocks - NUM_FIXED_BLOCKS);
546 threaded_edges[nthreaded_edges++] = t;
547
548 new_target = t->dest;
549 new_target_threaded = true;
550 }
551 }
552
553 if (!new_target)
554 break;
555
556 counter++;
557 target = new_target;
558 threaded |= new_target_threaded;
559 }
560
561 if (counter >= n_basic_blocks)
562 {
563 if (dump_file)
564 fprintf (dump_file, "Infinite loop in BB %i.\n",
565 target->index);
566 }
567 else if (target == first)
568 ; /* We didn't do anything. */
569 else
570 {
571 /* Save the values now, as the edge may get removed. */
572 gcov_type edge_count = e->count;
573 int edge_probability = e->probability;
574 int edge_frequency;
575 int n = 0;
576
577 e->goto_locus = goto_locus;
578
579 /* Don't force if target is exit block. */
580 if (threaded && target != EXIT_BLOCK_PTR)
581 {
582 notice_new_block (redirect_edge_and_branch_force (e, target));
583 if (dump_file)
584 fprintf (dump_file, "Conditionals threaded.\n");
585 }
586 else if (!redirect_edge_and_branch (e, target))
587 {
588 if (dump_file)
589 fprintf (dump_file,
590 "Forwarding edge %i->%i to %i failed.\n",
591 b->index, e->dest->index, target->index);
592 ei_next (&ei);
593 continue;
594 }
595
596 /* We successfully forwarded the edge. Now update profile
597 data: for each edge we traversed in the chain, remove
598 the original edge's execution count. */
599 edge_frequency = apply_probability (b->frequency, edge_probability);
600
601 do
602 {
603 edge t;
604
605 if (!single_succ_p (first))
606 {
607 gcc_assert (n < nthreaded_edges);
608 t = threaded_edges [n++];
609 gcc_assert (t->src == first);
610 update_bb_profile_for_threading (first, edge_frequency,
611 edge_count, t);
612 update_br_prob_note (first);
613 }
614 else
615 {
616 first->count -= edge_count;
617 if (first->count < 0)
618 first->count = 0;
619 first->frequency -= edge_frequency;
620 if (first->frequency < 0)
621 first->frequency = 0;
622 /* It is possible that as the result of
623 threading we've removed edge as it is
624 threaded to the fallthru edge. Avoid
625 getting out of sync. */
626 if (n < nthreaded_edges
627 && first == threaded_edges [n]->src)
628 n++;
629 t = single_succ_edge (first);
630 }
631
632 t->count -= edge_count;
633 if (t->count < 0)
634 t->count = 0;
635 first = t->dest;
636 }
637 while (first != target);
638
639 changed = true;
640 continue;
641 }
642 ei_next (&ei);
643 }
644
645 free (threaded_edges);
646 return changed;
647 }
648 \f
649
650 /* Blocks A and B are to be merged into a single block. A has no incoming
651 fallthru edge, so it can be moved before B without adding or modifying
652 any jumps (aside from the jump from A to B). */
653
654 static void
655 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
656 {
657 rtx barrier;
658
659 /* If we are partitioning hot/cold basic blocks, we don't want to
660 mess up unconditional or indirect jumps that cross between hot
661 and cold sections.
662
663 Basic block partitioning may result in some jumps that appear to
664 be optimizable (or blocks that appear to be mergeable), but which really
665 must be left untouched (they are required to make it safely across
666 partition boundaries). See the comments at the top of
667 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
668
669 if (BB_PARTITION (a) != BB_PARTITION (b))
670 return;
671
672 barrier = next_nonnote_insn (BB_END (a));
673 gcc_assert (BARRIER_P (barrier));
674 delete_insn (barrier);
675
676 /* Scramble the insn chain. */
677 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
678 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
679 df_set_bb_dirty (a);
680
681 if (dump_file)
682 fprintf (dump_file, "Moved block %d before %d and merged.\n",
683 a->index, b->index);
684
685 /* Swap the records for the two blocks around. */
686
687 unlink_block (a);
688 link_block (a, b->prev_bb);
689
690 /* Now blocks A and B are contiguous. Merge them. */
691 merge_blocks (a, b);
692 }
693
694 /* Blocks A and B are to be merged into a single block. B has no outgoing
695 fallthru edge, so it can be moved after A without adding or modifying
696 any jumps (aside from the jump from A to B). */
697
698 static void
699 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
700 {
701 rtx barrier, real_b_end;
702 rtx label, table;
703
704 /* If we are partitioning hot/cold basic blocks, we don't want to
705 mess up unconditional or indirect jumps that cross between hot
706 and cold sections.
707
708 Basic block partitioning may result in some jumps that appear to
709 be optimizable (or blocks that appear to be mergeable), but which really
710 must be left untouched (they are required to make it safely across
711 partition boundaries). See the comments at the top of
712 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
713
714 if (BB_PARTITION (a) != BB_PARTITION (b))
715 return;
716
717 real_b_end = BB_END (b);
718
719 /* If there is a jump table following block B temporarily add the jump table
720 to block B so that it will also be moved to the correct location. */
721 if (tablejump_p (BB_END (b), &label, &table)
722 && prev_active_insn (label) == BB_END (b))
723 {
724 BB_END (b) = table;
725 }
726
727 /* There had better have been a barrier there. Delete it. */
728 barrier = NEXT_INSN (BB_END (b));
729 if (barrier && BARRIER_P (barrier))
730 delete_insn (barrier);
731
732
733 /* Scramble the insn chain. */
734 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
735
736 /* Restore the real end of b. */
737 BB_END (b) = real_b_end;
738
739 if (dump_file)
740 fprintf (dump_file, "Moved block %d after %d and merged.\n",
741 b->index, a->index);
742
743 /* Now blocks A and B are contiguous. Merge them. */
744 merge_blocks (a, b);
745 }
746
747 /* Attempt to merge basic blocks that are potentially non-adjacent.
748 Return NULL iff the attempt failed, otherwise return basic block
749 where cleanup_cfg should continue. Because the merging commonly
750 moves basic block away or introduces another optimization
751 possibility, return basic block just before B so cleanup_cfg don't
752 need to iterate.
753
754 It may be good idea to return basic block before C in the case
755 C has been moved after B and originally appeared earlier in the
756 insn sequence, but we have no information available about the
757 relative ordering of these two. Hopefully it is not too common. */
758
759 static basic_block
760 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
761 {
762 basic_block next;
763
764 /* If we are partitioning hot/cold basic blocks, we don't want to
765 mess up unconditional or indirect jumps that cross between hot
766 and cold sections.
767
768 Basic block partitioning may result in some jumps that appear to
769 be optimizable (or blocks that appear to be mergeable), but which really
770 must be left untouched (they are required to make it safely across
771 partition boundaries). See the comments at the top of
772 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
773
774 if (BB_PARTITION (b) != BB_PARTITION (c))
775 return NULL;
776
777 /* If B has a fallthru edge to C, no need to move anything. */
778 if (e->flags & EDGE_FALLTHRU)
779 {
780 int b_index = b->index, c_index = c->index;
781
782 /* Protect the loop latches. */
783 if (current_loops && c->loop_father->latch == c)
784 return NULL;
785
786 merge_blocks (b, c);
787 update_forwarder_flag (b);
788
789 if (dump_file)
790 fprintf (dump_file, "Merged %d and %d without moving.\n",
791 b_index, c_index);
792
793 return b->prev_bb == ENTRY_BLOCK_PTR ? b : b->prev_bb;
794 }
795
796 /* Otherwise we will need to move code around. Do that only if expensive
797 transformations are allowed. */
798 else if (mode & CLEANUP_EXPENSIVE)
799 {
800 edge tmp_edge, b_fallthru_edge;
801 bool c_has_outgoing_fallthru;
802 bool b_has_incoming_fallthru;
803
804 /* Avoid overactive code motion, as the forwarder blocks should be
805 eliminated by edge redirection instead. One exception might have
806 been if B is a forwarder block and C has no fallthru edge, but
807 that should be cleaned up by bb-reorder instead. */
808 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
809 return NULL;
810
811 /* We must make sure to not munge nesting of lexical blocks,
812 and loop notes. This is done by squeezing out all the notes
813 and leaving them there to lie. Not ideal, but functional. */
814
815 tmp_edge = find_fallthru_edge (c->succs);
816 c_has_outgoing_fallthru = (tmp_edge != NULL);
817
818 tmp_edge = find_fallthru_edge (b->preds);
819 b_has_incoming_fallthru = (tmp_edge != NULL);
820 b_fallthru_edge = tmp_edge;
821 next = b->prev_bb;
822 if (next == c)
823 next = next->prev_bb;
824
825 /* Otherwise, we're going to try to move C after B. If C does
826 not have an outgoing fallthru, then it can be moved
827 immediately after B without introducing or modifying jumps. */
828 if (! c_has_outgoing_fallthru)
829 {
830 merge_blocks_move_successor_nojumps (b, c);
831 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
832 }
833
834 /* If B does not have an incoming fallthru, then it can be moved
835 immediately before C without introducing or modifying jumps.
836 C cannot be the first block, so we do not have to worry about
837 accessing a non-existent block. */
838
839 if (b_has_incoming_fallthru)
840 {
841 basic_block bb;
842
843 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR)
844 return NULL;
845 bb = force_nonfallthru (b_fallthru_edge);
846 if (bb)
847 notice_new_block (bb);
848 }
849
850 merge_blocks_move_predecessor_nojumps (b, c);
851 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
852 }
853
854 return NULL;
855 }
856 \f
857
858 /* Removes the memory attributes of MEM expression
859 if they are not equal. */
860
861 void
862 merge_memattrs (rtx x, rtx y)
863 {
864 int i;
865 int j;
866 enum rtx_code code;
867 const char *fmt;
868
869 if (x == y)
870 return;
871 if (x == 0 || y == 0)
872 return;
873
874 code = GET_CODE (x);
875
876 if (code != GET_CODE (y))
877 return;
878
879 if (GET_MODE (x) != GET_MODE (y))
880 return;
881
882 if (code == MEM && MEM_ATTRS (x) != MEM_ATTRS (y))
883 {
884 if (! MEM_ATTRS (x))
885 MEM_ATTRS (y) = 0;
886 else if (! MEM_ATTRS (y))
887 MEM_ATTRS (x) = 0;
888 else
889 {
890 HOST_WIDE_INT mem_size;
891
892 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
893 {
894 set_mem_alias_set (x, 0);
895 set_mem_alias_set (y, 0);
896 }
897
898 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
899 {
900 set_mem_expr (x, 0);
901 set_mem_expr (y, 0);
902 clear_mem_offset (x);
903 clear_mem_offset (y);
904 }
905 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
906 || (MEM_OFFSET_KNOWN_P (x)
907 && MEM_OFFSET (x) != MEM_OFFSET (y)))
908 {
909 clear_mem_offset (x);
910 clear_mem_offset (y);
911 }
912
913 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
914 {
915 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
916 set_mem_size (x, mem_size);
917 set_mem_size (y, mem_size);
918 }
919 else
920 {
921 clear_mem_size (x);
922 clear_mem_size (y);
923 }
924
925 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
926 set_mem_align (y, MEM_ALIGN (x));
927 }
928 }
929 if (code == MEM)
930 {
931 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
932 {
933 MEM_READONLY_P (x) = 0;
934 MEM_READONLY_P (y) = 0;
935 }
936 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
937 {
938 MEM_NOTRAP_P (x) = 0;
939 MEM_NOTRAP_P (y) = 0;
940 }
941 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
942 {
943 MEM_VOLATILE_P (x) = 1;
944 MEM_VOLATILE_P (y) = 1;
945 }
946 }
947
948 fmt = GET_RTX_FORMAT (code);
949 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
950 {
951 switch (fmt[i])
952 {
953 case 'E':
954 /* Two vectors must have the same length. */
955 if (XVECLEN (x, i) != XVECLEN (y, i))
956 return;
957
958 for (j = 0; j < XVECLEN (x, i); j++)
959 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
960
961 break;
962
963 case 'e':
964 merge_memattrs (XEXP (x, i), XEXP (y, i));
965 }
966 }
967 return;
968 }
969
970
971 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
972 different single sets S1 and S2. */
973
974 static bool
975 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
976 {
977 int i;
978 rtx e1, e2;
979
980 if (p1 == s1 && p2 == s2)
981 return true;
982
983 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
984 return false;
985
986 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
987 return false;
988
989 for (i = 0; i < XVECLEN (p1, 0); i++)
990 {
991 e1 = XVECEXP (p1, 0, i);
992 e2 = XVECEXP (p2, 0, i);
993 if (e1 == s1 && e2 == s2)
994 continue;
995 if (reload_completed
996 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
997 continue;
998
999 return false;
1000 }
1001
1002 return true;
1003 }
1004
1005 /* Examine register notes on I1 and I2 and return:
1006 - dir_forward if I1 can be replaced by I2, or
1007 - dir_backward if I2 can be replaced by I1, or
1008 - dir_both if both are the case. */
1009
1010 static enum replace_direction
1011 can_replace_by (rtx i1, rtx i2)
1012 {
1013 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1014 bool c1, c2;
1015
1016 /* Check for 2 sets. */
1017 s1 = single_set (i1);
1018 s2 = single_set (i2);
1019 if (s1 == NULL_RTX || s2 == NULL_RTX)
1020 return dir_none;
1021
1022 /* Check that the 2 sets set the same dest. */
1023 d1 = SET_DEST (s1);
1024 d2 = SET_DEST (s2);
1025 if (!(reload_completed
1026 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1027 return dir_none;
1028
1029 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1030 set dest to the same value. */
1031 note1 = find_reg_equal_equiv_note (i1);
1032 note2 = find_reg_equal_equiv_note (i2);
1033 if (!note1 || !note2 || !rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0))
1034 || !CONST_INT_P (XEXP (note1, 0)))
1035 return dir_none;
1036
1037 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1038 return dir_none;
1039
1040 /* Although the 2 sets set dest to the same value, we cannot replace
1041 (set (dest) (const_int))
1042 by
1043 (set (dest) (reg))
1044 because we don't know if the reg is live and has the same value at the
1045 location of replacement. */
1046 src1 = SET_SRC (s1);
1047 src2 = SET_SRC (s2);
1048 c1 = CONST_INT_P (src1);
1049 c2 = CONST_INT_P (src2);
1050 if (c1 && c2)
1051 return dir_both;
1052 else if (c2)
1053 return dir_forward;
1054 else if (c1)
1055 return dir_backward;
1056
1057 return dir_none;
1058 }
1059
1060 /* Merges directions A and B. */
1061
1062 static enum replace_direction
1063 merge_dir (enum replace_direction a, enum replace_direction b)
1064 {
1065 /* Implements the following table:
1066 |bo fw bw no
1067 ---+-----------
1068 bo |bo fw bw no
1069 fw |-- fw no no
1070 bw |-- -- bw no
1071 no |-- -- -- no. */
1072
1073 if (a == b)
1074 return a;
1075
1076 if (a == dir_both)
1077 return b;
1078 if (b == dir_both)
1079 return a;
1080
1081 return dir_none;
1082 }
1083
1084 /* Examine I1 and I2 and return:
1085 - dir_forward if I1 can be replaced by I2, or
1086 - dir_backward if I2 can be replaced by I1, or
1087 - dir_both if both are the case. */
1088
1089 static enum replace_direction
1090 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx i1, rtx i2)
1091 {
1092 rtx p1, p2;
1093
1094 /* Verify that I1 and I2 are equivalent. */
1095 if (GET_CODE (i1) != GET_CODE (i2))
1096 return dir_none;
1097
1098 /* __builtin_unreachable() may lead to empty blocks (ending with
1099 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1100 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1101 return dir_both;
1102
1103 /* ??? Do not allow cross-jumping between different stack levels. */
1104 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1105 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1106 if (p1 && p2)
1107 {
1108 p1 = XEXP (p1, 0);
1109 p2 = XEXP (p2, 0);
1110 if (!rtx_equal_p (p1, p2))
1111 return dir_none;
1112
1113 /* ??? Worse, this adjustment had better be constant lest we
1114 have differing incoming stack levels. */
1115 if (!frame_pointer_needed
1116 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1117 return dir_none;
1118 }
1119 else if (p1 || p2)
1120 return dir_none;
1121
1122 p1 = PATTERN (i1);
1123 p2 = PATTERN (i2);
1124
1125 if (GET_CODE (p1) != GET_CODE (p2))
1126 return dir_none;
1127
1128 /* If this is a CALL_INSN, compare register usage information.
1129 If we don't check this on stack register machines, the two
1130 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1131 numbers of stack registers in the same basic block.
1132 If we don't check this on machines with delay slots, a delay slot may
1133 be filled that clobbers a parameter expected by the subroutine.
1134
1135 ??? We take the simple route for now and assume that if they're
1136 equal, they were constructed identically.
1137
1138 Also check for identical exception regions. */
1139
1140 if (CALL_P (i1))
1141 {
1142 /* Ensure the same EH region. */
1143 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1144 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1145
1146 if (!n1 && n2)
1147 return dir_none;
1148
1149 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1150 return dir_none;
1151
1152 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1153 CALL_INSN_FUNCTION_USAGE (i2))
1154 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1155 return dir_none;
1156
1157 /* For address sanitizer, never crossjump __asan_report_* builtins,
1158 otherwise errors might be reported on incorrect lines. */
1159 if (flag_sanitize & SANITIZE_ADDRESS)
1160 {
1161 rtx call = get_call_rtx_from (i1);
1162 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1163 {
1164 rtx symbol = XEXP (XEXP (call, 0), 0);
1165 if (SYMBOL_REF_DECL (symbol)
1166 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1167 {
1168 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1169 == BUILT_IN_NORMAL)
1170 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1171 >= BUILT_IN_ASAN_REPORT_LOAD1
1172 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1173 <= BUILT_IN_ASAN_REPORT_STORE16)
1174 return dir_none;
1175 }
1176 }
1177 }
1178 }
1179
1180 #ifdef STACK_REGS
1181 /* If cross_jump_death_matters is not 0, the insn's mode
1182 indicates whether or not the insn contains any stack-like
1183 regs. */
1184
1185 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1186 {
1187 /* If register stack conversion has already been done, then
1188 death notes must also be compared before it is certain that
1189 the two instruction streams match. */
1190
1191 rtx note;
1192 HARD_REG_SET i1_regset, i2_regset;
1193
1194 CLEAR_HARD_REG_SET (i1_regset);
1195 CLEAR_HARD_REG_SET (i2_regset);
1196
1197 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1198 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1199 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1200
1201 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1202 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1203 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1204
1205 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1206 return dir_none;
1207 }
1208 #endif
1209
1210 if (reload_completed
1211 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1212 return dir_both;
1213
1214 return can_replace_by (i1, i2);
1215 }
1216 \f
1217 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1218 flow_find_head_matching_sequence, ensure the notes match. */
1219
1220 static void
1221 merge_notes (rtx i1, rtx i2)
1222 {
1223 /* If the merged insns have different REG_EQUAL notes, then
1224 remove them. */
1225 rtx equiv1 = find_reg_equal_equiv_note (i1);
1226 rtx equiv2 = find_reg_equal_equiv_note (i2);
1227
1228 if (equiv1 && !equiv2)
1229 remove_note (i1, equiv1);
1230 else if (!equiv1 && equiv2)
1231 remove_note (i2, equiv2);
1232 else if (equiv1 && equiv2
1233 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1234 {
1235 remove_note (i1, equiv1);
1236 remove_note (i2, equiv2);
1237 }
1238 }
1239
1240 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1241 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1242 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1243 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1244 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1245
1246 static void
1247 walk_to_nondebug_insn (rtx *i1, basic_block *bb1, bool follow_fallthru,
1248 bool *did_fallthru)
1249 {
1250 edge fallthru;
1251
1252 *did_fallthru = false;
1253
1254 /* Ignore notes. */
1255 while (!NONDEBUG_INSN_P (*i1))
1256 {
1257 if (*i1 != BB_HEAD (*bb1))
1258 {
1259 *i1 = PREV_INSN (*i1);
1260 continue;
1261 }
1262
1263 if (!follow_fallthru)
1264 return;
1265
1266 fallthru = find_fallthru_edge ((*bb1)->preds);
1267 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun)
1268 || !single_succ_p (fallthru->src))
1269 return;
1270
1271 *bb1 = fallthru->src;
1272 *i1 = BB_END (*bb1);
1273 *did_fallthru = true;
1274 }
1275 }
1276
1277 /* Look through the insns at the end of BB1 and BB2 and find the longest
1278 sequence that are either equivalent, or allow forward or backward
1279 replacement. Store the first insns for that sequence in *F1 and *F2 and
1280 return the sequence length.
1281
1282 DIR_P indicates the allowed replacement direction on function entry, and
1283 the actual replacement direction on function exit. If NULL, only equivalent
1284 sequences are allowed.
1285
1286 To simplify callers of this function, if the blocks match exactly,
1287 store the head of the blocks in *F1 and *F2. */
1288
1289 int
1290 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx *f1, rtx *f2,
1291 enum replace_direction *dir_p)
1292 {
1293 rtx i1, i2, last1, last2, afterlast1, afterlast2;
1294 int ninsns = 0;
1295 rtx p1;
1296 enum replace_direction dir, last_dir, afterlast_dir;
1297 bool follow_fallthru, did_fallthru;
1298
1299 if (dir_p)
1300 dir = *dir_p;
1301 else
1302 dir = dir_both;
1303 afterlast_dir = dir;
1304 last_dir = afterlast_dir;
1305
1306 /* Skip simple jumps at the end of the blocks. Complex jumps still
1307 need to be compared for equivalence, which we'll do below. */
1308
1309 i1 = BB_END (bb1);
1310 last1 = afterlast1 = last2 = afterlast2 = NULL_RTX;
1311 if (onlyjump_p (i1)
1312 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1313 {
1314 last1 = i1;
1315 i1 = PREV_INSN (i1);
1316 }
1317
1318 i2 = BB_END (bb2);
1319 if (onlyjump_p (i2)
1320 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1321 {
1322 last2 = i2;
1323 /* Count everything except for unconditional jump as insn. */
1324 if (!simplejump_p (i2) && !returnjump_p (i2) && last1)
1325 ninsns++;
1326 i2 = PREV_INSN (i2);
1327 }
1328
1329 while (true)
1330 {
1331 /* In the following example, we can replace all jumps to C by jumps to A.
1332
1333 This removes 4 duplicate insns.
1334 [bb A] insn1 [bb C] insn1
1335 insn2 insn2
1336 [bb B] insn3 insn3
1337 insn4 insn4
1338 jump_insn jump_insn
1339
1340 We could also replace all jumps to A by jumps to C, but that leaves B
1341 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1342 step, all jumps to B would be replaced with jumps to the middle of C,
1343 achieving the same result with more effort.
1344 So we allow only the first possibility, which means that we don't allow
1345 fallthru in the block that's being replaced. */
1346
1347 follow_fallthru = dir_p && dir != dir_forward;
1348 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1349 if (did_fallthru)
1350 dir = dir_backward;
1351
1352 follow_fallthru = dir_p && dir != dir_backward;
1353 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1354 if (did_fallthru)
1355 dir = dir_forward;
1356
1357 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1358 break;
1359
1360 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1361 if (dir == dir_none || (!dir_p && dir != dir_both))
1362 break;
1363
1364 merge_memattrs (i1, i2);
1365
1366 /* Don't begin a cross-jump with a NOTE insn. */
1367 if (INSN_P (i1))
1368 {
1369 merge_notes (i1, i2);
1370
1371 afterlast1 = last1, afterlast2 = last2;
1372 last1 = i1, last2 = i2;
1373 afterlast_dir = last_dir;
1374 last_dir = dir;
1375 p1 = PATTERN (i1);
1376 if (!(GET_CODE (p1) == USE || GET_CODE (p1) == CLOBBER))
1377 ninsns++;
1378 }
1379
1380 i1 = PREV_INSN (i1);
1381 i2 = PREV_INSN (i2);
1382 }
1383
1384 #ifdef HAVE_cc0
1385 /* Don't allow the insn after a compare to be shared by
1386 cross-jumping unless the compare is also shared. */
1387 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
1388 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1389 #endif
1390
1391 /* Include preceding notes and labels in the cross-jump. One,
1392 this may bring us to the head of the blocks as requested above.
1393 Two, it keeps line number notes as matched as may be. */
1394 if (ninsns)
1395 {
1396 bb1 = BLOCK_FOR_INSN (last1);
1397 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1398 last1 = PREV_INSN (last1);
1399
1400 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1401 last1 = PREV_INSN (last1);
1402
1403 bb2 = BLOCK_FOR_INSN (last2);
1404 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1405 last2 = PREV_INSN (last2);
1406
1407 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1408 last2 = PREV_INSN (last2);
1409
1410 *f1 = last1;
1411 *f2 = last2;
1412 }
1413
1414 if (dir_p)
1415 *dir_p = last_dir;
1416 return ninsns;
1417 }
1418
1419 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1420 the head of the two blocks. Do not include jumps at the end.
1421 If STOP_AFTER is nonzero, stop after finding that many matching
1422 instructions. */
1423
1424 int
1425 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx *f1,
1426 rtx *f2, int stop_after)
1427 {
1428 rtx i1, i2, last1, last2, beforelast1, beforelast2;
1429 int ninsns = 0;
1430 edge e;
1431 edge_iterator ei;
1432 int nehedges1 = 0, nehedges2 = 0;
1433
1434 FOR_EACH_EDGE (e, ei, bb1->succs)
1435 if (e->flags & EDGE_EH)
1436 nehedges1++;
1437 FOR_EACH_EDGE (e, ei, bb2->succs)
1438 if (e->flags & EDGE_EH)
1439 nehedges2++;
1440
1441 i1 = BB_HEAD (bb1);
1442 i2 = BB_HEAD (bb2);
1443 last1 = beforelast1 = last2 = beforelast2 = NULL_RTX;
1444
1445 while (true)
1446 {
1447 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1448 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1449 {
1450 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1451 break;
1452 i1 = NEXT_INSN (i1);
1453 }
1454
1455 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1456 {
1457 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1458 break;
1459 i2 = NEXT_INSN (i2);
1460 }
1461
1462 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1463 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1464 break;
1465
1466 if (NOTE_P (i1) || NOTE_P (i2)
1467 || JUMP_P (i1) || JUMP_P (i2))
1468 break;
1469
1470 /* A sanity check to make sure we're not merging insns with different
1471 effects on EH. If only one of them ends a basic block, it shouldn't
1472 have an EH edge; if both end a basic block, there should be the same
1473 number of EH edges. */
1474 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1475 && nehedges1 > 0)
1476 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1477 && nehedges2 > 0)
1478 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1479 && nehedges1 != nehedges2))
1480 break;
1481
1482 if (old_insns_match_p (0, i1, i2) != dir_both)
1483 break;
1484
1485 merge_memattrs (i1, i2);
1486
1487 /* Don't begin a cross-jump with a NOTE insn. */
1488 if (INSN_P (i1))
1489 {
1490 merge_notes (i1, i2);
1491
1492 beforelast1 = last1, beforelast2 = last2;
1493 last1 = i1, last2 = i2;
1494 ninsns++;
1495 }
1496
1497 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1498 || (stop_after > 0 && ninsns == stop_after))
1499 break;
1500
1501 i1 = NEXT_INSN (i1);
1502 i2 = NEXT_INSN (i2);
1503 }
1504
1505 #ifdef HAVE_cc0
1506 /* Don't allow a compare to be shared by cross-jumping unless the insn
1507 after the compare is also shared. */
1508 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && sets_cc0_p (last1))
1509 last1 = beforelast1, last2 = beforelast2, ninsns--;
1510 #endif
1511
1512 if (ninsns)
1513 {
1514 *f1 = last1;
1515 *f2 = last2;
1516 }
1517
1518 return ninsns;
1519 }
1520
1521 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1522 the branch instruction. This means that if we commonize the control
1523 flow before end of the basic block, the semantic remains unchanged.
1524
1525 We may assume that there exists one edge with a common destination. */
1526
1527 static bool
1528 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1529 {
1530 int nehedges1 = 0, nehedges2 = 0;
1531 edge fallthru1 = 0, fallthru2 = 0;
1532 edge e1, e2;
1533 edge_iterator ei;
1534
1535 /* If we performed shrink-wrapping, edges to the EXIT_BLOCK_PTR can
1536 only be distinguished for JUMP_INSNs. The two paths may differ in
1537 whether they went through the prologue. Sibcalls are fine, we know
1538 that we either didn't need or inserted an epilogue before them. */
1539 if (crtl->shrink_wrapped
1540 && single_succ_p (bb1) && single_succ (bb1) == EXIT_BLOCK_PTR
1541 && !JUMP_P (BB_END (bb1))
1542 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1543 return false;
1544
1545 /* If BB1 has only one successor, we may be looking at either an
1546 unconditional jump, or a fake edge to exit. */
1547 if (single_succ_p (bb1)
1548 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1549 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1550 return (single_succ_p (bb2)
1551 && (single_succ_edge (bb2)->flags
1552 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1553 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1554
1555 /* Match conditional jumps - this may get tricky when fallthru and branch
1556 edges are crossed. */
1557 if (EDGE_COUNT (bb1->succs) == 2
1558 && any_condjump_p (BB_END (bb1))
1559 && onlyjump_p (BB_END (bb1)))
1560 {
1561 edge b1, f1, b2, f2;
1562 bool reverse, match;
1563 rtx set1, set2, cond1, cond2;
1564 enum rtx_code code1, code2;
1565
1566 if (EDGE_COUNT (bb2->succs) != 2
1567 || !any_condjump_p (BB_END (bb2))
1568 || !onlyjump_p (BB_END (bb2)))
1569 return false;
1570
1571 b1 = BRANCH_EDGE (bb1);
1572 b2 = BRANCH_EDGE (bb2);
1573 f1 = FALLTHRU_EDGE (bb1);
1574 f2 = FALLTHRU_EDGE (bb2);
1575
1576 /* Get around possible forwarders on fallthru edges. Other cases
1577 should be optimized out already. */
1578 if (FORWARDER_BLOCK_P (f1->dest))
1579 f1 = single_succ_edge (f1->dest);
1580
1581 if (FORWARDER_BLOCK_P (f2->dest))
1582 f2 = single_succ_edge (f2->dest);
1583
1584 /* To simplify use of this function, return false if there are
1585 unneeded forwarder blocks. These will get eliminated later
1586 during cleanup_cfg. */
1587 if (FORWARDER_BLOCK_P (f1->dest)
1588 || FORWARDER_BLOCK_P (f2->dest)
1589 || FORWARDER_BLOCK_P (b1->dest)
1590 || FORWARDER_BLOCK_P (b2->dest))
1591 return false;
1592
1593 if (f1->dest == f2->dest && b1->dest == b2->dest)
1594 reverse = false;
1595 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1596 reverse = true;
1597 else
1598 return false;
1599
1600 set1 = pc_set (BB_END (bb1));
1601 set2 = pc_set (BB_END (bb2));
1602 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1603 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1604 reverse = !reverse;
1605
1606 cond1 = XEXP (SET_SRC (set1), 0);
1607 cond2 = XEXP (SET_SRC (set2), 0);
1608 code1 = GET_CODE (cond1);
1609 if (reverse)
1610 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1611 else
1612 code2 = GET_CODE (cond2);
1613
1614 if (code2 == UNKNOWN)
1615 return false;
1616
1617 /* Verify codes and operands match. */
1618 match = ((code1 == code2
1619 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1620 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1621 || (code1 == swap_condition (code2)
1622 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1623 XEXP (cond2, 0))
1624 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1625 XEXP (cond2, 1))));
1626
1627 /* If we return true, we will join the blocks. Which means that
1628 we will only have one branch prediction bit to work with. Thus
1629 we require the existing branches to have probabilities that are
1630 roughly similar. */
1631 if (match
1632 && optimize_bb_for_speed_p (bb1)
1633 && optimize_bb_for_speed_p (bb2))
1634 {
1635 int prob2;
1636
1637 if (b1->dest == b2->dest)
1638 prob2 = b2->probability;
1639 else
1640 /* Do not use f2 probability as f2 may be forwarded. */
1641 prob2 = REG_BR_PROB_BASE - b2->probability;
1642
1643 /* Fail if the difference in probabilities is greater than 50%.
1644 This rules out two well-predicted branches with opposite
1645 outcomes. */
1646 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1647 {
1648 if (dump_file)
1649 fprintf (dump_file,
1650 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1651 bb1->index, bb2->index, b1->probability, prob2);
1652
1653 return false;
1654 }
1655 }
1656
1657 if (dump_file && match)
1658 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1659 bb1->index, bb2->index);
1660
1661 return match;
1662 }
1663
1664 /* Generic case - we are seeing a computed jump, table jump or trapping
1665 instruction. */
1666
1667 /* Check whether there are tablejumps in the end of BB1 and BB2.
1668 Return true if they are identical. */
1669 {
1670 rtx label1, label2;
1671 rtx table1, table2;
1672
1673 if (tablejump_p (BB_END (bb1), &label1, &table1)
1674 && tablejump_p (BB_END (bb2), &label2, &table2)
1675 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1676 {
1677 /* The labels should never be the same rtx. If they really are same
1678 the jump tables are same too. So disable crossjumping of blocks BB1
1679 and BB2 because when deleting the common insns in the end of BB1
1680 by delete_basic_block () the jump table would be deleted too. */
1681 /* If LABEL2 is referenced in BB1->END do not do anything
1682 because we would loose information when replacing
1683 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1684 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1685 {
1686 /* Set IDENTICAL to true when the tables are identical. */
1687 bool identical = false;
1688 rtx p1, p2;
1689
1690 p1 = PATTERN (table1);
1691 p2 = PATTERN (table2);
1692 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1693 {
1694 identical = true;
1695 }
1696 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1697 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1698 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1699 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1700 {
1701 int i;
1702
1703 identical = true;
1704 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1705 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1706 identical = false;
1707 }
1708
1709 if (identical)
1710 {
1711 replace_label_data rr;
1712 bool match;
1713
1714 /* Temporarily replace references to LABEL1 with LABEL2
1715 in BB1->END so that we could compare the instructions. */
1716 rr.r1 = label1;
1717 rr.r2 = label2;
1718 rr.update_label_nuses = false;
1719 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1720
1721 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1722 == dir_both);
1723 if (dump_file && match)
1724 fprintf (dump_file,
1725 "Tablejumps in bb %i and %i match.\n",
1726 bb1->index, bb2->index);
1727
1728 /* Set the original label in BB1->END because when deleting
1729 a block whose end is a tablejump, the tablejump referenced
1730 from the instruction is deleted too. */
1731 rr.r1 = label2;
1732 rr.r2 = label1;
1733 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1734
1735 return match;
1736 }
1737 }
1738 return false;
1739 }
1740 }
1741
1742 rtx last1 = BB_END (bb1);
1743 rtx last2 = BB_END (bb2);
1744 if (DEBUG_INSN_P (last1))
1745 last1 = prev_nondebug_insn (last1);
1746 if (DEBUG_INSN_P (last2))
1747 last2 = prev_nondebug_insn (last2);
1748 /* First ensure that the instructions match. There may be many outgoing
1749 edges so this test is generally cheaper. */
1750 if (old_insns_match_p (mode, last1, last2) != dir_both)
1751 return false;
1752
1753 /* Search the outgoing edges, ensure that the counts do match, find possible
1754 fallthru and exception handling edges since these needs more
1755 validation. */
1756 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1757 return false;
1758
1759 bool nonfakeedges = false;
1760 FOR_EACH_EDGE (e1, ei, bb1->succs)
1761 {
1762 e2 = EDGE_SUCC (bb2, ei.index);
1763
1764 if ((e1->flags & EDGE_FAKE) == 0)
1765 nonfakeedges = true;
1766
1767 if (e1->flags & EDGE_EH)
1768 nehedges1++;
1769
1770 if (e2->flags & EDGE_EH)
1771 nehedges2++;
1772
1773 if (e1->flags & EDGE_FALLTHRU)
1774 fallthru1 = e1;
1775 if (e2->flags & EDGE_FALLTHRU)
1776 fallthru2 = e2;
1777 }
1778
1779 /* If number of edges of various types does not match, fail. */
1780 if (nehedges1 != nehedges2
1781 || (fallthru1 != 0) != (fallthru2 != 0))
1782 return false;
1783
1784 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1785 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1786 attempt to optimize, as the two basic blocks might have different
1787 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1788 traps there should be REG_ARG_SIZE notes, they could be missing
1789 for __builtin_unreachable () uses though. */
1790 if (!nonfakeedges
1791 && !ACCUMULATE_OUTGOING_ARGS
1792 && (!INSN_P (last1)
1793 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1794 return false;
1795
1796 /* fallthru edges must be forwarded to the same destination. */
1797 if (fallthru1)
1798 {
1799 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1800 ? single_succ (fallthru1->dest): fallthru1->dest);
1801 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1802 ? single_succ (fallthru2->dest): fallthru2->dest);
1803
1804 if (d1 != d2)
1805 return false;
1806 }
1807
1808 /* Ensure the same EH region. */
1809 {
1810 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1811 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1812
1813 if (!n1 && n2)
1814 return false;
1815
1816 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1817 return false;
1818 }
1819
1820 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1821 version of sequence abstraction. */
1822 FOR_EACH_EDGE (e1, ei, bb2->succs)
1823 {
1824 edge e2;
1825 edge_iterator ei;
1826 basic_block d1 = e1->dest;
1827
1828 if (FORWARDER_BLOCK_P (d1))
1829 d1 = EDGE_SUCC (d1, 0)->dest;
1830
1831 FOR_EACH_EDGE (e2, ei, bb1->succs)
1832 {
1833 basic_block d2 = e2->dest;
1834 if (FORWARDER_BLOCK_P (d2))
1835 d2 = EDGE_SUCC (d2, 0)->dest;
1836 if (d1 == d2)
1837 break;
1838 }
1839
1840 if (!e2)
1841 return false;
1842 }
1843
1844 return true;
1845 }
1846
1847 /* Returns true if BB basic block has a preserve label. */
1848
1849 static bool
1850 block_has_preserve_label (basic_block bb)
1851 {
1852 return (bb
1853 && block_label (bb)
1854 && LABEL_PRESERVE_P (block_label (bb)));
1855 }
1856
1857 /* E1 and E2 are edges with the same destination block. Search their
1858 predecessors for common code. If found, redirect control flow from
1859 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1860 or the other way around (dir_backward). DIR specifies the allowed
1861 replacement direction. */
1862
1863 static bool
1864 try_crossjump_to_edge (int mode, edge e1, edge e2,
1865 enum replace_direction dir)
1866 {
1867 int nmatch;
1868 basic_block src1 = e1->src, src2 = e2->src;
1869 basic_block redirect_to, redirect_from, to_remove;
1870 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1871 rtx newpos1, newpos2;
1872 edge s;
1873 edge_iterator ei;
1874
1875 newpos1 = newpos2 = NULL_RTX;
1876
1877 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1878 to try this optimization.
1879
1880 Basic block partitioning may result in some jumps that appear to
1881 be optimizable (or blocks that appear to be mergeable), but which really
1882 must be left untouched (they are required to make it safely across
1883 partition boundaries). See the comments at the top of
1884 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1885
1886 if (crtl->has_bb_partition && reload_completed)
1887 return false;
1888
1889 /* Search backward through forwarder blocks. We don't need to worry
1890 about multiple entry or chained forwarders, as they will be optimized
1891 away. We do this to look past the unconditional jump following a
1892 conditional jump that is required due to the current CFG shape. */
1893 if (single_pred_p (src1)
1894 && FORWARDER_BLOCK_P (src1))
1895 e1 = single_pred_edge (src1), src1 = e1->src;
1896
1897 if (single_pred_p (src2)
1898 && FORWARDER_BLOCK_P (src2))
1899 e2 = single_pred_edge (src2), src2 = e2->src;
1900
1901 /* Nothing to do if we reach ENTRY, or a common source block. */
1902 if (src1 == ENTRY_BLOCK_PTR || src2 == ENTRY_BLOCK_PTR)
1903 return false;
1904 if (src1 == src2)
1905 return false;
1906
1907 /* Seeing more than 1 forwarder blocks would confuse us later... */
1908 if (FORWARDER_BLOCK_P (e1->dest)
1909 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1910 return false;
1911
1912 if (FORWARDER_BLOCK_P (e2->dest)
1913 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1914 return false;
1915
1916 /* Likewise with dead code (possibly newly created by the other optimizations
1917 of cfg_cleanup). */
1918 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1919 return false;
1920
1921 /* Look for the common insn sequence, part the first ... */
1922 if (!outgoing_edges_match (mode, src1, src2))
1923 return false;
1924
1925 /* ... and part the second. */
1926 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1927
1928 osrc1 = src1;
1929 osrc2 = src2;
1930 if (newpos1 != NULL_RTX)
1931 src1 = BLOCK_FOR_INSN (newpos1);
1932 if (newpos2 != NULL_RTX)
1933 src2 = BLOCK_FOR_INSN (newpos2);
1934
1935 if (dir == dir_backward)
1936 {
1937 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1938 SWAP (basic_block, osrc1, osrc2);
1939 SWAP (basic_block, src1, src2);
1940 SWAP (edge, e1, e2);
1941 SWAP (rtx, newpos1, newpos2);
1942 #undef SWAP
1943 }
1944
1945 /* Don't proceed with the crossjump unless we found a sufficient number
1946 of matching instructions or the 'from' block was totally matched
1947 (such that its predecessors will hopefully be redirected and the
1948 block removed). */
1949 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
1950 && (newpos1 != BB_HEAD (src1)))
1951 return false;
1952
1953 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1954 if (block_has_preserve_label (e1->dest)
1955 && (e1->flags & EDGE_ABNORMAL))
1956 return false;
1957
1958 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1959 will be deleted.
1960 If we have tablejumps in the end of SRC1 and SRC2
1961 they have been already compared for equivalence in outgoing_edges_match ()
1962 so replace the references to TABLE1 by references to TABLE2. */
1963 {
1964 rtx label1, label2;
1965 rtx table1, table2;
1966
1967 if (tablejump_p (BB_END (osrc1), &label1, &table1)
1968 && tablejump_p (BB_END (osrc2), &label2, &table2)
1969 && label1 != label2)
1970 {
1971 replace_label_data rr;
1972 rtx insn;
1973
1974 /* Replace references to LABEL1 with LABEL2. */
1975 rr.r1 = label1;
1976 rr.r2 = label2;
1977 rr.update_label_nuses = true;
1978 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
1979 {
1980 /* Do not replace the label in SRC1->END because when deleting
1981 a block whose end is a tablejump, the tablejump referenced
1982 from the instruction is deleted too. */
1983 if (insn != BB_END (osrc1))
1984 for_each_rtx (&insn, replace_label, &rr);
1985 }
1986 }
1987 }
1988
1989 /* Avoid splitting if possible. We must always split when SRC2 has
1990 EH predecessor edges, or we may end up with basic blocks with both
1991 normal and EH predecessor edges. */
1992 if (newpos2 == BB_HEAD (src2)
1993 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
1994 redirect_to = src2;
1995 else
1996 {
1997 if (newpos2 == BB_HEAD (src2))
1998 {
1999 /* Skip possible basic block header. */
2000 if (LABEL_P (newpos2))
2001 newpos2 = NEXT_INSN (newpos2);
2002 while (DEBUG_INSN_P (newpos2))
2003 newpos2 = NEXT_INSN (newpos2);
2004 if (NOTE_P (newpos2))
2005 newpos2 = NEXT_INSN (newpos2);
2006 while (DEBUG_INSN_P (newpos2))
2007 newpos2 = NEXT_INSN (newpos2);
2008 }
2009
2010 if (dump_file)
2011 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2012 src2->index, nmatch);
2013 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2014 }
2015
2016 if (dump_file)
2017 fprintf (dump_file,
2018 "Cross jumping from bb %i to bb %i; %i common insns\n",
2019 src1->index, src2->index, nmatch);
2020
2021 /* We may have some registers visible through the block. */
2022 df_set_bb_dirty (redirect_to);
2023
2024 if (osrc2 == src2)
2025 redirect_edges_to = redirect_to;
2026 else
2027 redirect_edges_to = osrc2;
2028
2029 /* Recompute the frequencies and counts of outgoing edges. */
2030 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2031 {
2032 edge s2;
2033 edge_iterator ei;
2034 basic_block d = s->dest;
2035
2036 if (FORWARDER_BLOCK_P (d))
2037 d = single_succ (d);
2038
2039 FOR_EACH_EDGE (s2, ei, src1->succs)
2040 {
2041 basic_block d2 = s2->dest;
2042 if (FORWARDER_BLOCK_P (d2))
2043 d2 = single_succ (d2);
2044 if (d == d2)
2045 break;
2046 }
2047
2048 s->count += s2->count;
2049
2050 /* Take care to update possible forwarder blocks. We verified
2051 that there is no more than one in the chain, so we can't run
2052 into infinite loop. */
2053 if (FORWARDER_BLOCK_P (s->dest))
2054 {
2055 single_succ_edge (s->dest)->count += s2->count;
2056 s->dest->count += s2->count;
2057 s->dest->frequency += EDGE_FREQUENCY (s);
2058 }
2059
2060 if (FORWARDER_BLOCK_P (s2->dest))
2061 {
2062 single_succ_edge (s2->dest)->count -= s2->count;
2063 if (single_succ_edge (s2->dest)->count < 0)
2064 single_succ_edge (s2->dest)->count = 0;
2065 s2->dest->count -= s2->count;
2066 s2->dest->frequency -= EDGE_FREQUENCY (s);
2067 if (s2->dest->frequency < 0)
2068 s2->dest->frequency = 0;
2069 if (s2->dest->count < 0)
2070 s2->dest->count = 0;
2071 }
2072
2073 if (!redirect_edges_to->frequency && !src1->frequency)
2074 s->probability = (s->probability + s2->probability) / 2;
2075 else
2076 s->probability
2077 = ((s->probability * redirect_edges_to->frequency +
2078 s2->probability * src1->frequency)
2079 / (redirect_edges_to->frequency + src1->frequency));
2080 }
2081
2082 /* Adjust count and frequency for the block. An earlier jump
2083 threading pass may have left the profile in an inconsistent
2084 state (see update_bb_profile_for_threading) so we must be
2085 prepared for overflows. */
2086 tmp = redirect_to;
2087 do
2088 {
2089 tmp->count += src1->count;
2090 tmp->frequency += src1->frequency;
2091 if (tmp->frequency > BB_FREQ_MAX)
2092 tmp->frequency = BB_FREQ_MAX;
2093 if (tmp == redirect_edges_to)
2094 break;
2095 tmp = find_fallthru_edge (tmp->succs)->dest;
2096 }
2097 while (true);
2098 update_br_prob_note (redirect_edges_to);
2099
2100 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2101
2102 /* Skip possible basic block header. */
2103 if (LABEL_P (newpos1))
2104 newpos1 = NEXT_INSN (newpos1);
2105
2106 while (DEBUG_INSN_P (newpos1))
2107 newpos1 = NEXT_INSN (newpos1);
2108
2109 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2110 newpos1 = NEXT_INSN (newpos1);
2111
2112 while (DEBUG_INSN_P (newpos1))
2113 newpos1 = NEXT_INSN (newpos1);
2114
2115 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2116 to_remove = single_succ (redirect_from);
2117
2118 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2119 delete_basic_block (to_remove);
2120
2121 update_forwarder_flag (redirect_from);
2122 if (redirect_to != src2)
2123 update_forwarder_flag (src2);
2124
2125 return true;
2126 }
2127
2128 /* Search the predecessors of BB for common insn sequences. When found,
2129 share code between them by redirecting control flow. Return true if
2130 any changes made. */
2131
2132 static bool
2133 try_crossjump_bb (int mode, basic_block bb)
2134 {
2135 edge e, e2, fallthru;
2136 bool changed;
2137 unsigned max, ix, ix2;
2138
2139 /* Nothing to do if there is not at least two incoming edges. */
2140 if (EDGE_COUNT (bb->preds) < 2)
2141 return false;
2142
2143 /* Don't crossjump if this block ends in a computed jump,
2144 unless we are optimizing for size. */
2145 if (optimize_bb_for_size_p (bb)
2146 && bb != EXIT_BLOCK_PTR
2147 && computed_jump_p (BB_END (bb)))
2148 return false;
2149
2150 /* If we are partitioning hot/cold basic blocks, we don't want to
2151 mess up unconditional or indirect jumps that cross between hot
2152 and cold sections.
2153
2154 Basic block partitioning may result in some jumps that appear to
2155 be optimizable (or blocks that appear to be mergeable), but which really
2156 must be left untouched (they are required to make it safely across
2157 partition boundaries). See the comments at the top of
2158 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2159
2160 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2161 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2162 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2163 return false;
2164
2165 /* It is always cheapest to redirect a block that ends in a branch to
2166 a block that falls through into BB, as that adds no branches to the
2167 program. We'll try that combination first. */
2168 fallthru = NULL;
2169 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2170
2171 if (EDGE_COUNT (bb->preds) > max)
2172 return false;
2173
2174 fallthru = find_fallthru_edge (bb->preds);
2175
2176 changed = false;
2177 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2178 {
2179 e = EDGE_PRED (bb, ix);
2180 ix++;
2181
2182 /* As noted above, first try with the fallthru predecessor (or, a
2183 fallthru predecessor if we are in cfglayout mode). */
2184 if (fallthru)
2185 {
2186 /* Don't combine the fallthru edge into anything else.
2187 If there is a match, we'll do it the other way around. */
2188 if (e == fallthru)
2189 continue;
2190 /* If nothing changed since the last attempt, there is nothing
2191 we can do. */
2192 if (!first_pass
2193 && !((e->src->flags & BB_MODIFIED)
2194 || (fallthru->src->flags & BB_MODIFIED)))
2195 continue;
2196
2197 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2198 {
2199 changed = true;
2200 ix = 0;
2201 continue;
2202 }
2203 }
2204
2205 /* Non-obvious work limiting check: Recognize that we're going
2206 to call try_crossjump_bb on every basic block. So if we have
2207 two blocks with lots of outgoing edges (a switch) and they
2208 share lots of common destinations, then we would do the
2209 cross-jump check once for each common destination.
2210
2211 Now, if the blocks actually are cross-jump candidates, then
2212 all of their destinations will be shared. Which means that
2213 we only need check them for cross-jump candidacy once. We
2214 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2215 choosing to do the check from the block for which the edge
2216 in question is the first successor of A. */
2217 if (EDGE_SUCC (e->src, 0) != e)
2218 continue;
2219
2220 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2221 {
2222 e2 = EDGE_PRED (bb, ix2);
2223
2224 if (e2 == e)
2225 continue;
2226
2227 /* We've already checked the fallthru edge above. */
2228 if (e2 == fallthru)
2229 continue;
2230
2231 /* The "first successor" check above only prevents multiple
2232 checks of crossjump(A,B). In order to prevent redundant
2233 checks of crossjump(B,A), require that A be the block
2234 with the lowest index. */
2235 if (e->src->index > e2->src->index)
2236 continue;
2237
2238 /* If nothing changed since the last attempt, there is nothing
2239 we can do. */
2240 if (!first_pass
2241 && !((e->src->flags & BB_MODIFIED)
2242 || (e2->src->flags & BB_MODIFIED)))
2243 continue;
2244
2245 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2246 direction. */
2247 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2248 {
2249 changed = true;
2250 ix = 0;
2251 break;
2252 }
2253 }
2254 }
2255
2256 if (changed)
2257 crossjumps_occured = true;
2258
2259 return changed;
2260 }
2261
2262 /* Search the successors of BB for common insn sequences. When found,
2263 share code between them by moving it across the basic block
2264 boundary. Return true if any changes made. */
2265
2266 static bool
2267 try_head_merge_bb (basic_block bb)
2268 {
2269 basic_block final_dest_bb = NULL;
2270 int max_match = INT_MAX;
2271 edge e0;
2272 rtx *headptr, *currptr, *nextptr;
2273 bool changed, moveall;
2274 unsigned ix;
2275 rtx e0_last_head, cond, move_before;
2276 unsigned nedges = EDGE_COUNT (bb->succs);
2277 rtx jump = BB_END (bb);
2278 regset live, live_union;
2279
2280 /* Nothing to do if there is not at least two outgoing edges. */
2281 if (nedges < 2)
2282 return false;
2283
2284 /* Don't crossjump if this block ends in a computed jump,
2285 unless we are optimizing for size. */
2286 if (optimize_bb_for_size_p (bb)
2287 && bb != EXIT_BLOCK_PTR
2288 && computed_jump_p (BB_END (bb)))
2289 return false;
2290
2291 cond = get_condition (jump, &move_before, true, false);
2292 if (cond == NULL_RTX)
2293 {
2294 #ifdef HAVE_cc0
2295 if (reg_mentioned_p (cc0_rtx, jump))
2296 move_before = prev_nonnote_nondebug_insn (jump);
2297 else
2298 #endif
2299 move_before = jump;
2300 }
2301
2302 for (ix = 0; ix < nedges; ix++)
2303 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR)
2304 return false;
2305
2306 for (ix = 0; ix < nedges; ix++)
2307 {
2308 edge e = EDGE_SUCC (bb, ix);
2309 basic_block other_bb = e->dest;
2310
2311 if (df_get_bb_dirty (other_bb))
2312 {
2313 block_was_dirty = true;
2314 return false;
2315 }
2316
2317 if (e->flags & EDGE_ABNORMAL)
2318 return false;
2319
2320 /* Normally, all destination blocks must only be reachable from this
2321 block, i.e. they must have one incoming edge.
2322
2323 There is one special case we can handle, that of multiple consecutive
2324 jumps where the first jumps to one of the targets of the second jump.
2325 This happens frequently in switch statements for default labels.
2326 The structure is as follows:
2327 FINAL_DEST_BB
2328 ....
2329 if (cond) jump A;
2330 fall through
2331 BB
2332 jump with targets A, B, C, D...
2333 A
2334 has two incoming edges, from FINAL_DEST_BB and BB
2335
2336 In this case, we can try to move the insns through BB and into
2337 FINAL_DEST_BB. */
2338 if (EDGE_COUNT (other_bb->preds) != 1)
2339 {
2340 edge incoming_edge, incoming_bb_other_edge;
2341 edge_iterator ei;
2342
2343 if (final_dest_bb != NULL
2344 || EDGE_COUNT (other_bb->preds) != 2)
2345 return false;
2346
2347 /* We must be able to move the insns across the whole block. */
2348 move_before = BB_HEAD (bb);
2349 while (!NONDEBUG_INSN_P (move_before))
2350 move_before = NEXT_INSN (move_before);
2351
2352 if (EDGE_COUNT (bb->preds) != 1)
2353 return false;
2354 incoming_edge = EDGE_PRED (bb, 0);
2355 final_dest_bb = incoming_edge->src;
2356 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2357 return false;
2358 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2359 if (incoming_bb_other_edge != incoming_edge)
2360 break;
2361 if (incoming_bb_other_edge->dest != other_bb)
2362 return false;
2363 }
2364 }
2365
2366 e0 = EDGE_SUCC (bb, 0);
2367 e0_last_head = NULL_RTX;
2368 changed = false;
2369
2370 for (ix = 1; ix < nedges; ix++)
2371 {
2372 edge e = EDGE_SUCC (bb, ix);
2373 rtx e0_last, e_last;
2374 int nmatch;
2375
2376 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2377 &e0_last, &e_last, 0);
2378 if (nmatch == 0)
2379 return false;
2380
2381 if (nmatch < max_match)
2382 {
2383 max_match = nmatch;
2384 e0_last_head = e0_last;
2385 }
2386 }
2387
2388 /* If we matched an entire block, we probably have to avoid moving the
2389 last insn. */
2390 if (max_match > 0
2391 && e0_last_head == BB_END (e0->dest)
2392 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2393 || control_flow_insn_p (e0_last_head)))
2394 {
2395 max_match--;
2396 if (max_match == 0)
2397 return false;
2398 do
2399 e0_last_head = prev_real_insn (e0_last_head);
2400 while (DEBUG_INSN_P (e0_last_head));
2401 }
2402
2403 if (max_match == 0)
2404 return false;
2405
2406 /* We must find a union of the live registers at each of the end points. */
2407 live = BITMAP_ALLOC (NULL);
2408 live_union = BITMAP_ALLOC (NULL);
2409
2410 currptr = XNEWVEC (rtx, nedges);
2411 headptr = XNEWVEC (rtx, nedges);
2412 nextptr = XNEWVEC (rtx, nedges);
2413
2414 for (ix = 0; ix < nedges; ix++)
2415 {
2416 int j;
2417 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2418 rtx head = BB_HEAD (merge_bb);
2419
2420 while (!NONDEBUG_INSN_P (head))
2421 head = NEXT_INSN (head);
2422 headptr[ix] = head;
2423 currptr[ix] = head;
2424
2425 /* Compute the end point and live information */
2426 for (j = 1; j < max_match; j++)
2427 do
2428 head = NEXT_INSN (head);
2429 while (!NONDEBUG_INSN_P (head));
2430 simulate_backwards_to_point (merge_bb, live, head);
2431 IOR_REG_SET (live_union, live);
2432 }
2433
2434 /* If we're moving across two blocks, verify the validity of the
2435 first move, then adjust the target and let the loop below deal
2436 with the final move. */
2437 if (final_dest_bb != NULL)
2438 {
2439 rtx move_upto;
2440
2441 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2442 jump, e0->dest, live_union,
2443 NULL, &move_upto);
2444 if (!moveall)
2445 {
2446 if (move_upto == NULL_RTX)
2447 goto out;
2448
2449 while (e0_last_head != move_upto)
2450 {
2451 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2452 live_union);
2453 e0_last_head = PREV_INSN (e0_last_head);
2454 }
2455 }
2456 if (e0_last_head == NULL_RTX)
2457 goto out;
2458
2459 jump = BB_END (final_dest_bb);
2460 cond = get_condition (jump, &move_before, true, false);
2461 if (cond == NULL_RTX)
2462 {
2463 #ifdef HAVE_cc0
2464 if (reg_mentioned_p (cc0_rtx, jump))
2465 move_before = prev_nonnote_nondebug_insn (jump);
2466 else
2467 #endif
2468 move_before = jump;
2469 }
2470 }
2471
2472 do
2473 {
2474 rtx move_upto;
2475 moveall = can_move_insns_across (currptr[0], e0_last_head,
2476 move_before, jump, e0->dest, live_union,
2477 NULL, &move_upto);
2478 if (!moveall && move_upto == NULL_RTX)
2479 {
2480 if (jump == move_before)
2481 break;
2482
2483 /* Try again, using a different insertion point. */
2484 move_before = jump;
2485
2486 #ifdef HAVE_cc0
2487 /* Don't try moving before a cc0 user, as that may invalidate
2488 the cc0. */
2489 if (reg_mentioned_p (cc0_rtx, jump))
2490 break;
2491 #endif
2492
2493 continue;
2494 }
2495
2496 if (final_dest_bb && !moveall)
2497 /* We haven't checked whether a partial move would be OK for the first
2498 move, so we have to fail this case. */
2499 break;
2500
2501 changed = true;
2502 for (;;)
2503 {
2504 if (currptr[0] == move_upto)
2505 break;
2506 for (ix = 0; ix < nedges; ix++)
2507 {
2508 rtx curr = currptr[ix];
2509 do
2510 curr = NEXT_INSN (curr);
2511 while (!NONDEBUG_INSN_P (curr));
2512 currptr[ix] = curr;
2513 }
2514 }
2515
2516 /* If we can't currently move all of the identical insns, remember
2517 each insn after the range that we'll merge. */
2518 if (!moveall)
2519 for (ix = 0; ix < nedges; ix++)
2520 {
2521 rtx curr = currptr[ix];
2522 do
2523 curr = NEXT_INSN (curr);
2524 while (!NONDEBUG_INSN_P (curr));
2525 nextptr[ix] = curr;
2526 }
2527
2528 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2529 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2530 if (final_dest_bb != NULL)
2531 df_set_bb_dirty (final_dest_bb);
2532 df_set_bb_dirty (bb);
2533 for (ix = 1; ix < nedges; ix++)
2534 {
2535 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2536 delete_insn_chain (headptr[ix], currptr[ix], false);
2537 }
2538 if (!moveall)
2539 {
2540 if (jump == move_before)
2541 break;
2542
2543 /* For the unmerged insns, try a different insertion point. */
2544 move_before = jump;
2545
2546 #ifdef HAVE_cc0
2547 /* Don't try moving before a cc0 user, as that may invalidate
2548 the cc0. */
2549 if (reg_mentioned_p (cc0_rtx, jump))
2550 break;
2551 #endif
2552
2553 for (ix = 0; ix < nedges; ix++)
2554 currptr[ix] = headptr[ix] = nextptr[ix];
2555 }
2556 }
2557 while (!moveall);
2558
2559 out:
2560 free (currptr);
2561 free (headptr);
2562 free (nextptr);
2563
2564 crossjumps_occured |= changed;
2565
2566 return changed;
2567 }
2568
2569 /* Return true if BB contains just bb note, or bb note followed
2570 by only DEBUG_INSNs. */
2571
2572 static bool
2573 trivially_empty_bb_p (basic_block bb)
2574 {
2575 rtx insn = BB_END (bb);
2576
2577 while (1)
2578 {
2579 if (insn == BB_HEAD (bb))
2580 return true;
2581 if (!DEBUG_INSN_P (insn))
2582 return false;
2583 insn = PREV_INSN (insn);
2584 }
2585 }
2586
2587 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2588 instructions etc. Return nonzero if changes were made. */
2589
2590 static bool
2591 try_optimize_cfg (int mode)
2592 {
2593 bool changed_overall = false;
2594 bool changed;
2595 int iterations = 0;
2596 basic_block bb, b, next;
2597
2598 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2599 clear_bb_flags ();
2600
2601 crossjumps_occured = false;
2602
2603 FOR_EACH_BB (bb)
2604 update_forwarder_flag (bb);
2605
2606 if (! targetm.cannot_modify_jumps_p ())
2607 {
2608 first_pass = true;
2609 /* Attempt to merge blocks as made possible by edge removal. If
2610 a block has only one successor, and the successor has only
2611 one predecessor, they may be combined. */
2612 do
2613 {
2614 block_was_dirty = false;
2615 changed = false;
2616 iterations++;
2617
2618 if (dump_file)
2619 fprintf (dump_file,
2620 "\n\ntry_optimize_cfg iteration %i\n\n",
2621 iterations);
2622
2623 for (b = ENTRY_BLOCK_PTR->next_bb; b != EXIT_BLOCK_PTR;)
2624 {
2625 basic_block c;
2626 edge s;
2627 bool changed_here = false;
2628
2629 /* Delete trivially dead basic blocks. This is either
2630 blocks with no predecessors, or empty blocks with no
2631 successors. However if the empty block with no
2632 successors is the successor of the ENTRY_BLOCK, it is
2633 kept. This ensures that the ENTRY_BLOCK will have a
2634 successor which is a precondition for many RTL
2635 passes. Empty blocks may result from expanding
2636 __builtin_unreachable (). */
2637 if (EDGE_COUNT (b->preds) == 0
2638 || (EDGE_COUNT (b->succs) == 0
2639 && trivially_empty_bb_p (b)
2640 && single_succ_edge (ENTRY_BLOCK_PTR)->dest != b))
2641 {
2642 c = b->prev_bb;
2643 if (EDGE_COUNT (b->preds) > 0)
2644 {
2645 edge e;
2646 edge_iterator ei;
2647
2648 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2649 {
2650 if (BB_FOOTER (b)
2651 && BARRIER_P (BB_FOOTER (b)))
2652 FOR_EACH_EDGE (e, ei, b->preds)
2653 if ((e->flags & EDGE_FALLTHRU)
2654 && BB_FOOTER (e->src) == NULL)
2655 {
2656 if (BB_FOOTER (b))
2657 {
2658 BB_FOOTER (e->src) = BB_FOOTER (b);
2659 BB_FOOTER (b) = NULL;
2660 }
2661 else
2662 {
2663 start_sequence ();
2664 BB_FOOTER (e->src) = emit_barrier ();
2665 end_sequence ();
2666 }
2667 }
2668 }
2669 else
2670 {
2671 rtx last = get_last_bb_insn (b);
2672 if (last && BARRIER_P (last))
2673 FOR_EACH_EDGE (e, ei, b->preds)
2674 if ((e->flags & EDGE_FALLTHRU))
2675 emit_barrier_after (BB_END (e->src));
2676 }
2677 }
2678 delete_basic_block (b);
2679 changed = true;
2680 /* Avoid trying to remove ENTRY_BLOCK_PTR. */
2681 b = (c == ENTRY_BLOCK_PTR ? c->next_bb : c);
2682 continue;
2683 }
2684
2685 /* Remove code labels no longer used. */
2686 if (single_pred_p (b)
2687 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2688 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2689 && LABEL_P (BB_HEAD (b))
2690 /* If the previous block ends with a branch to this
2691 block, we can't delete the label. Normally this
2692 is a condjump that is yet to be simplified, but
2693 if CASE_DROPS_THRU, this can be a tablejump with
2694 some element going to the same place as the
2695 default (fallthru). */
2696 && (single_pred (b) == ENTRY_BLOCK_PTR
2697 || !JUMP_P (BB_END (single_pred (b)))
2698 || ! label_is_jump_target_p (BB_HEAD (b),
2699 BB_END (single_pred (b)))))
2700 {
2701 delete_insn (BB_HEAD (b));
2702 if (dump_file)
2703 fprintf (dump_file, "Deleted label in block %i.\n",
2704 b->index);
2705 }
2706
2707 /* If we fall through an empty block, we can remove it. */
2708 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2709 && single_pred_p (b)
2710 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2711 && !LABEL_P (BB_HEAD (b))
2712 && FORWARDER_BLOCK_P (b)
2713 /* Note that forwarder_block_p true ensures that
2714 there is a successor for this block. */
2715 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2716 && n_basic_blocks > NUM_FIXED_BLOCKS + 1)
2717 {
2718 if (dump_file)
2719 fprintf (dump_file,
2720 "Deleting fallthru block %i.\n",
2721 b->index);
2722
2723 c = b->prev_bb == ENTRY_BLOCK_PTR ? b->next_bb : b->prev_bb;
2724 redirect_edge_succ_nodup (single_pred_edge (b),
2725 single_succ (b));
2726 delete_basic_block (b);
2727 changed = true;
2728 b = c;
2729 continue;
2730 }
2731
2732 /* Merge B with its single successor, if any. */
2733 if (single_succ_p (b)
2734 && (s = single_succ_edge (b))
2735 && !(s->flags & EDGE_COMPLEX)
2736 && (c = s->dest) != EXIT_BLOCK_PTR
2737 && single_pred_p (c)
2738 && b != c)
2739 {
2740 /* When not in cfg_layout mode use code aware of reordering
2741 INSN. This code possibly creates new basic blocks so it
2742 does not fit merge_blocks interface and is kept here in
2743 hope that it will become useless once more of compiler
2744 is transformed to use cfg_layout mode. */
2745
2746 if ((mode & CLEANUP_CFGLAYOUT)
2747 && can_merge_blocks_p (b, c))
2748 {
2749 merge_blocks (b, c);
2750 update_forwarder_flag (b);
2751 changed_here = true;
2752 }
2753 else if (!(mode & CLEANUP_CFGLAYOUT)
2754 /* If the jump insn has side effects,
2755 we can't kill the edge. */
2756 && (!JUMP_P (BB_END (b))
2757 || (reload_completed
2758 ? simplejump_p (BB_END (b))
2759 : (onlyjump_p (BB_END (b))
2760 && !tablejump_p (BB_END (b),
2761 NULL, NULL))))
2762 && (next = merge_blocks_move (s, b, c, mode)))
2763 {
2764 b = next;
2765 changed_here = true;
2766 }
2767 }
2768
2769 /* Simplify branch over branch. */
2770 if ((mode & CLEANUP_EXPENSIVE)
2771 && !(mode & CLEANUP_CFGLAYOUT)
2772 && try_simplify_condjump (b))
2773 changed_here = true;
2774
2775 /* If B has a single outgoing edge, but uses a
2776 non-trivial jump instruction without side-effects, we
2777 can either delete the jump entirely, or replace it
2778 with a simple unconditional jump. */
2779 if (single_succ_p (b)
2780 && single_succ (b) != EXIT_BLOCK_PTR
2781 && onlyjump_p (BB_END (b))
2782 && !find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX)
2783 && try_redirect_by_replacing_jump (single_succ_edge (b),
2784 single_succ (b),
2785 (mode & CLEANUP_CFGLAYOUT) != 0))
2786 {
2787 update_forwarder_flag (b);
2788 changed_here = true;
2789 }
2790
2791 /* Simplify branch to branch. */
2792 if (try_forward_edges (mode, b))
2793 {
2794 update_forwarder_flag (b);
2795 changed_here = true;
2796 }
2797
2798 /* Look for shared code between blocks. */
2799 if ((mode & CLEANUP_CROSSJUMP)
2800 && try_crossjump_bb (mode, b))
2801 changed_here = true;
2802
2803 if ((mode & CLEANUP_CROSSJUMP)
2804 /* This can lengthen register lifetimes. Do it only after
2805 reload. */
2806 && reload_completed
2807 && try_head_merge_bb (b))
2808 changed_here = true;
2809
2810 /* Don't get confused by the index shift caused by
2811 deleting blocks. */
2812 if (!changed_here)
2813 b = b->next_bb;
2814 else
2815 changed = true;
2816 }
2817
2818 if ((mode & CLEANUP_CROSSJUMP)
2819 && try_crossjump_bb (mode, EXIT_BLOCK_PTR))
2820 changed = true;
2821
2822 if (block_was_dirty)
2823 {
2824 /* This should only be set by head-merging. */
2825 gcc_assert (mode & CLEANUP_CROSSJUMP);
2826 df_analyze ();
2827 }
2828
2829 if (changed)
2830 {
2831 /* Edge forwarding in particular can cause hot blocks previously
2832 reached by both hot and cold blocks to become dominated only
2833 by cold blocks. This will cause the verification below to fail,
2834 and lead to now cold code in the hot section. This is not easy
2835 to detect and fix during edge forwarding, and in some cases
2836 is only visible after newly unreachable blocks are deleted,
2837 which will be done in fixup_partitions. */
2838 fixup_partitions ();
2839
2840 #ifdef ENABLE_CHECKING
2841 verify_flow_info ();
2842 #endif
2843 }
2844
2845 changed_overall |= changed;
2846 first_pass = false;
2847 }
2848 while (changed);
2849 }
2850
2851 FOR_ALL_BB (b)
2852 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2853
2854 return changed_overall;
2855 }
2856 \f
2857 /* Delete all unreachable basic blocks. */
2858
2859 bool
2860 delete_unreachable_blocks (void)
2861 {
2862 bool changed = false;
2863 basic_block b, prev_bb;
2864
2865 find_unreachable_blocks ();
2866
2867 /* When we're in GIMPLE mode and there may be debug insns, we should
2868 delete blocks in reverse dominator order, so as to get a chance
2869 to substitute all released DEFs into debug stmts. If we don't
2870 have dominators information, walking blocks backward gets us a
2871 better chance of retaining most debug information than
2872 otherwise. */
2873 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2874 && dom_info_available_p (CDI_DOMINATORS))
2875 {
2876 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2877 {
2878 prev_bb = b->prev_bb;
2879
2880 if (!(b->flags & BB_REACHABLE))
2881 {
2882 /* Speed up the removal of blocks that don't dominate
2883 others. Walking backwards, this should be the common
2884 case. */
2885 if (!first_dom_son (CDI_DOMINATORS, b))
2886 delete_basic_block (b);
2887 else
2888 {
2889 vec<basic_block> h
2890 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2891
2892 while (h.length ())
2893 {
2894 b = h.pop ();
2895
2896 prev_bb = b->prev_bb;
2897
2898 gcc_assert (!(b->flags & BB_REACHABLE));
2899
2900 delete_basic_block (b);
2901 }
2902
2903 h.release ();
2904 }
2905
2906 changed = true;
2907 }
2908 }
2909 }
2910 else
2911 {
2912 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2913 {
2914 prev_bb = b->prev_bb;
2915
2916 if (!(b->flags & BB_REACHABLE))
2917 {
2918 delete_basic_block (b);
2919 changed = true;
2920 }
2921 }
2922 }
2923
2924 if (changed)
2925 tidy_fallthru_edges ();
2926 return changed;
2927 }
2928
2929 /* Delete any jump tables never referenced. We can't delete them at the
2930 time of removing tablejump insn as they are referenced by the preceding
2931 insns computing the destination, so we delay deleting and garbagecollect
2932 them once life information is computed. */
2933 void
2934 delete_dead_jumptables (void)
2935 {
2936 basic_block bb;
2937
2938 /* A dead jump table does not belong to any basic block. Scan insns
2939 between two adjacent basic blocks. */
2940 FOR_EACH_BB (bb)
2941 {
2942 rtx insn, next;
2943
2944 for (insn = NEXT_INSN (BB_END (bb));
2945 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
2946 insn = next)
2947 {
2948 next = NEXT_INSN (insn);
2949 if (LABEL_P (insn)
2950 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
2951 && JUMP_TABLE_DATA_P (next))
2952 {
2953 rtx label = insn, jump = next;
2954
2955 if (dump_file)
2956 fprintf (dump_file, "Dead jumptable %i removed\n",
2957 INSN_UID (insn));
2958
2959 next = NEXT_INSN (next);
2960 delete_insn (jump);
2961 delete_insn (label);
2962 }
2963 }
2964 }
2965 }
2966
2967 \f
2968 /* Tidy the CFG by deleting unreachable code and whatnot. */
2969
2970 bool
2971 cleanup_cfg (int mode)
2972 {
2973 bool changed = false;
2974
2975 /* Set the cfglayout mode flag here. We could update all the callers
2976 but that is just inconvenient, especially given that we eventually
2977 want to have cfglayout mode as the default. */
2978 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2979 mode |= CLEANUP_CFGLAYOUT;
2980
2981 timevar_push (TV_CLEANUP_CFG);
2982 if (delete_unreachable_blocks ())
2983 {
2984 changed = true;
2985 /* We've possibly created trivially dead code. Cleanup it right
2986 now to introduce more opportunities for try_optimize_cfg. */
2987 if (!(mode & (CLEANUP_NO_INSN_DEL))
2988 && !reload_completed)
2989 delete_trivially_dead_insns (get_insns (), max_reg_num ());
2990 }
2991
2992 compact_blocks ();
2993
2994 /* To tail-merge blocks ending in the same noreturn function (e.g.
2995 a call to abort) we have to insert fake edges to exit. Do this
2996 here once. The fake edges do not interfere with any other CFG
2997 cleanups. */
2998 if (mode & CLEANUP_CROSSJUMP)
2999 add_noreturn_fake_exit_edges ();
3000
3001 if (!dbg_cnt (cfg_cleanup))
3002 return changed;
3003
3004 while (try_optimize_cfg (mode))
3005 {
3006 delete_unreachable_blocks (), changed = true;
3007 if (!(mode & CLEANUP_NO_INSN_DEL))
3008 {
3009 /* Try to remove some trivially dead insns when doing an expensive
3010 cleanup. But delete_trivially_dead_insns doesn't work after
3011 reload (it only handles pseudos) and run_fast_dce is too costly
3012 to run in every iteration.
3013
3014 For effective cross jumping, we really want to run a fast DCE to
3015 clean up any dead conditions, or they get in the way of performing
3016 useful tail merges.
3017
3018 Other transformations in cleanup_cfg are not so sensitive to dead
3019 code, so delete_trivially_dead_insns or even doing nothing at all
3020 is good enough. */
3021 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3022 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3023 break;
3024 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
3025 run_fast_dce ();
3026 }
3027 else
3028 break;
3029 }
3030
3031 if (mode & CLEANUP_CROSSJUMP)
3032 remove_fake_exit_edges ();
3033
3034 /* Don't call delete_dead_jumptables in cfglayout mode, because
3035 that function assumes that jump tables are in the insns stream.
3036 But we also don't _have_ to delete dead jumptables in cfglayout
3037 mode because we shouldn't even be looking at things that are
3038 not in a basic block. Dead jumptables are cleaned up when
3039 going out of cfglayout mode. */
3040 if (!(mode & CLEANUP_CFGLAYOUT))
3041 delete_dead_jumptables ();
3042
3043 /* ??? We probably do this way too often. */
3044 if (current_loops
3045 && (changed
3046 || (mode & CLEANUP_CFG_CHANGED)))
3047 {
3048 timevar_push (TV_REPAIR_LOOPS);
3049 /* The above doesn't preserve dominance info if available. */
3050 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3051 calculate_dominance_info (CDI_DOMINATORS);
3052 fix_loop_structure (NULL);
3053 free_dominance_info (CDI_DOMINATORS);
3054 timevar_pop (TV_REPAIR_LOOPS);
3055 }
3056
3057 timevar_pop (TV_CLEANUP_CFG);
3058
3059 return changed;
3060 }
3061 \f
3062 static unsigned int
3063 execute_jump (void)
3064 {
3065 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3066 if (dump_file)
3067 dump_flow_info (dump_file, dump_flags);
3068 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3069 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3070 return 0;
3071 }
3072
3073 namespace {
3074
3075 const pass_data pass_data_jump =
3076 {
3077 RTL_PASS, /* type */
3078 "jump", /* name */
3079 OPTGROUP_NONE, /* optinfo_flags */
3080 false, /* has_gate */
3081 true, /* has_execute */
3082 TV_JUMP, /* tv_id */
3083 0, /* properties_required */
3084 0, /* properties_provided */
3085 0, /* properties_destroyed */
3086 0, /* todo_flags_start */
3087 TODO_verify_rtl_sharing, /* todo_flags_finish */
3088 };
3089
3090 class pass_jump : public rtl_opt_pass
3091 {
3092 public:
3093 pass_jump (gcc::context *ctxt)
3094 : rtl_opt_pass (pass_data_jump, ctxt)
3095 {}
3096
3097 /* opt_pass methods: */
3098 unsigned int execute () { return execute_jump (); }
3099
3100 }; // class pass_jump
3101
3102 } // anon namespace
3103
3104 rtl_opt_pass *
3105 make_pass_jump (gcc::context *ctxt)
3106 {
3107 return new pass_jump (ctxt);
3108 }
3109 \f
3110 static unsigned int
3111 execute_jump2 (void)
3112 {
3113 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3114 return 0;
3115 }
3116
3117 namespace {
3118
3119 const pass_data pass_data_jump2 =
3120 {
3121 RTL_PASS, /* type */
3122 "jump2", /* name */
3123 OPTGROUP_NONE, /* optinfo_flags */
3124 false, /* has_gate */
3125 true, /* has_execute */
3126 TV_JUMP, /* tv_id */
3127 0, /* properties_required */
3128 0, /* properties_provided */
3129 0, /* properties_destroyed */
3130 0, /* todo_flags_start */
3131 TODO_verify_rtl_sharing, /* todo_flags_finish */
3132 };
3133
3134 class pass_jump2 : public rtl_opt_pass
3135 {
3136 public:
3137 pass_jump2 (gcc::context *ctxt)
3138 : rtl_opt_pass (pass_data_jump2, ctxt)
3139 {}
3140
3141 /* opt_pass methods: */
3142 unsigned int execute () { return execute_jump2 (); }
3143
3144 }; // class pass_jump2
3145
3146 } // anon namespace
3147
3148 rtl_opt_pass *
3149 make_pass_jump2 (gcc::context *ctxt)
3150 {
3151 return new pass_jump2 (ctxt);
3152 }