PR middle-end/83575
[official-gcc.git] / gcc / cfgcleanup.c
blobc33022c1bb5c6d75736d80af75f0b1469e654e1e
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2018 Free Software Foundation, Inc.
4 This file is part of GCC.
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.
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.
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/>. */
20 /* This file contains optimizer of the control flow. The main entry point is
21 cleanup_cfg. Following optimizations are performed:
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. */
32 #include "config.h"
33 #include "system.h"
34 #include "coretypes.h"
35 #include "backend.h"
36 #include "target.h"
37 #include "rtl.h"
38 #include "tree.h"
39 #include "cfghooks.h"
40 #include "df.h"
41 #include "memmodel.h"
42 #include "tm_p.h"
43 #include "insn-config.h"
44 #include "emit-rtl.h"
45 #include "cselib.h"
46 #include "params.h"
47 #include "tree-pass.h"
48 #include "cfgloop.h"
49 #include "cfgrtl.h"
50 #include "cfganal.h"
51 #include "cfgbuild.h"
52 #include "cfgcleanup.h"
53 #include "dce.h"
54 #include "dbgcnt.h"
55 #include "rtl-iter.h"
57 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
59 /* Set to true when we are running first pass of try_optimize_cfg loop. */
60 static bool first_pass;
62 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
63 static bool crossjumps_occurred;
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;
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_insn *, rtx_insn *);
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 void merge_memattrs (rtx, rtx);
85 /* Set flags for newly created block. */
87 static void
88 notice_new_block (basic_block bb)
90 if (!bb)
91 return;
93 if (forwarder_block_p (bb))
94 bb->flags |= BB_FORWARDER_BLOCK;
97 /* Recompute forwarder flag after block has been modified. */
99 static void
100 update_forwarder_flag (basic_block bb)
102 if (forwarder_block_p (bb))
103 bb->flags |= BB_FORWARDER_BLOCK;
104 else
105 bb->flags &= ~BB_FORWARDER_BLOCK;
108 /* Simplify a conditional jump around an unconditional jump.
109 Return true if something changed. */
111 static bool
112 try_simplify_condjump (basic_block cbranch_block)
114 basic_block jump_block, jump_dest_block, cbranch_dest_block;
115 edge cbranch_jump_edge, cbranch_fallthru_edge;
116 rtx_insn *cbranch_insn;
118 /* Verify that there are exactly two successors. */
119 if (EDGE_COUNT (cbranch_block->succs) != 2)
120 return false;
122 /* Verify that we've got a normal conditional branch at the end
123 of the block. */
124 cbranch_insn = BB_END (cbranch_block);
125 if (!any_condjump_p (cbranch_insn))
126 return false;
128 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
129 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
131 /* The next block must not have multiple predecessors, must not
132 be the last block in the function, and must contain just the
133 unconditional jump. */
134 jump_block = cbranch_fallthru_edge->dest;
135 if (!single_pred_p (jump_block)
136 || jump_block->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
137 || !FORWARDER_BLOCK_P (jump_block))
138 return false;
139 jump_dest_block = single_succ (jump_block);
141 /* If we are partitioning hot/cold basic blocks, we don't want to
142 mess up unconditional or indirect jumps that cross between hot
143 and cold sections.
145 Basic block partitioning may result in some jumps that appear to
146 be optimizable (or blocks that appear to be mergeable), but which really
147 must be left untouched (they are required to make it safely across
148 partition boundaries). See the comments at the top of
149 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
151 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
152 || (cbranch_jump_edge->flags & EDGE_CROSSING))
153 return false;
155 /* The conditional branch must target the block after the
156 unconditional branch. */
157 cbranch_dest_block = cbranch_jump_edge->dest;
159 if (cbranch_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
160 || jump_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
161 || !can_fallthru (jump_block, cbranch_dest_block))
162 return false;
164 /* Invert the conditional branch. */
165 if (!invert_jump (as_a <rtx_jump_insn *> (cbranch_insn),
166 block_label (jump_dest_block), 0))
167 return false;
169 if (dump_file)
170 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
171 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
173 /* Success. Update the CFG to match. Note that after this point
174 the edge variable names appear backwards; the redirection is done
175 this way to preserve edge profile data. */
176 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
177 cbranch_dest_block);
178 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
179 jump_dest_block);
180 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
181 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
182 update_br_prob_note (cbranch_block);
184 /* Delete the block with the unconditional jump, and clean up the mess. */
185 delete_basic_block (jump_block);
186 tidy_fallthru_edge (cbranch_jump_edge);
187 update_forwarder_flag (cbranch_block);
189 return true;
192 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
193 on register. Used by jump threading. */
195 static bool
196 mark_effect (rtx exp, regset nonequal)
198 rtx dest;
199 switch (GET_CODE (exp))
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 dest = XEXP (exp, 0);
205 if (REG_P (dest))
206 bitmap_clear_range (nonequal, REGNO (dest), REG_NREGS (dest));
207 return false;
209 case SET:
210 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
211 return false;
212 dest = SET_DEST (exp);
213 if (dest == pc_rtx)
214 return false;
215 if (!REG_P (dest))
216 return true;
217 bitmap_set_range (nonequal, REGNO (dest), REG_NREGS (dest));
218 return false;
220 default:
221 return false;
225 /* Return true if X contains a register in NONEQUAL. */
226 static bool
227 mentions_nonequal_regs (const_rtx x, regset nonequal)
229 subrtx_iterator::array_type array;
230 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
232 const_rtx x = *iter;
233 if (REG_P (x))
235 unsigned int end_regno = END_REGNO (x);
236 for (unsigned int regno = REGNO (x); regno < end_regno; ++regno)
237 if (REGNO_REG_SET_P (nonequal, regno))
238 return true;
241 return false;
244 /* Attempt to prove that the basic block B will have no side effects and
245 always continues in the same edge if reached via E. Return the edge
246 if exist, NULL otherwise. */
248 static edge
249 thread_jump (edge e, basic_block b)
251 rtx set1, set2, cond1, cond2;
252 rtx_insn *insn;
253 enum rtx_code code1, code2, reversed_code2;
254 bool reverse1 = false;
255 unsigned i;
256 regset nonequal;
257 bool failed = false;
258 reg_set_iterator rsi;
260 if (b->flags & BB_NONTHREADABLE_BLOCK)
261 return NULL;
263 /* At the moment, we do handle only conditional jumps, but later we may
264 want to extend this code to tablejumps and others. */
265 if (EDGE_COUNT (e->src->succs) != 2)
266 return NULL;
267 if (EDGE_COUNT (b->succs) != 2)
269 b->flags |= BB_NONTHREADABLE_BLOCK;
270 return NULL;
273 /* Second branch must end with onlyjump, as we will eliminate the jump. */
274 if (!any_condjump_p (BB_END (e->src)))
275 return NULL;
277 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
279 b->flags |= BB_NONTHREADABLE_BLOCK;
280 return NULL;
283 set1 = pc_set (BB_END (e->src));
284 set2 = pc_set (BB_END (b));
285 if (((e->flags & EDGE_FALLTHRU) != 0)
286 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
287 reverse1 = true;
289 cond1 = XEXP (SET_SRC (set1), 0);
290 cond2 = XEXP (SET_SRC (set2), 0);
291 if (reverse1)
292 code1 = reversed_comparison_code (cond1, BB_END (e->src));
293 else
294 code1 = GET_CODE (cond1);
296 code2 = GET_CODE (cond2);
297 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
299 if (!comparison_dominates_p (code1, code2)
300 && !comparison_dominates_p (code1, reversed_code2))
301 return NULL;
303 /* Ensure that the comparison operators are equivalent.
304 ??? This is far too pessimistic. We should allow swapped operands,
305 different CCmodes, or for example comparisons for interval, that
306 dominate even when operands are not equivalent. */
307 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
308 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
309 return NULL;
311 /* Short circuit cases where block B contains some side effects, as we can't
312 safely bypass it. */
313 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
314 insn = NEXT_INSN (insn))
315 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
317 b->flags |= BB_NONTHREADABLE_BLOCK;
318 return NULL;
321 cselib_init (0);
323 /* First process all values computed in the source basic block. */
324 for (insn = NEXT_INSN (BB_HEAD (e->src));
325 insn != NEXT_INSN (BB_END (e->src));
326 insn = NEXT_INSN (insn))
327 if (INSN_P (insn))
328 cselib_process_insn (insn);
330 nonequal = BITMAP_ALLOC (NULL);
331 CLEAR_REG_SET (nonequal);
333 /* Now assume that we've continued by the edge E to B and continue
334 processing as if it were same basic block.
335 Our goal is to prove that whole block is an NOOP. */
337 for (insn = NEXT_INSN (BB_HEAD (b));
338 insn != NEXT_INSN (BB_END (b)) && !failed;
339 insn = NEXT_INSN (insn))
341 if (INSN_P (insn))
343 rtx pat = PATTERN (insn);
345 if (GET_CODE (pat) == PARALLEL)
347 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
348 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
350 else
351 failed |= mark_effect (pat, nonequal);
354 cselib_process_insn (insn);
357 /* Later we should clear nonequal of dead registers. So far we don't
358 have life information in cfg_cleanup. */
359 if (failed)
361 b->flags |= BB_NONTHREADABLE_BLOCK;
362 goto failed_exit;
365 /* cond2 must not mention any register that is not equal to the
366 former block. */
367 if (mentions_nonequal_regs (cond2, nonequal))
368 goto failed_exit;
370 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
371 goto failed_exit;
373 BITMAP_FREE (nonequal);
374 cselib_finish ();
375 if ((comparison_dominates_p (code1, code2) != 0)
376 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
377 return BRANCH_EDGE (b);
378 else
379 return FALLTHRU_EDGE (b);
381 failed_exit:
382 BITMAP_FREE (nonequal);
383 cselib_finish ();
384 return NULL;
387 /* Attempt to forward edges leaving basic block B.
388 Return true if successful. */
390 static bool
391 try_forward_edges (int mode, basic_block b)
393 bool changed = false;
394 edge_iterator ei;
395 edge e, *threaded_edges = NULL;
397 /* If we are partitioning hot/cold basic blocks, we don't want to
398 mess up unconditional or indirect jumps that cross between hot
399 and cold sections.
401 Basic block partitioning may result in some jumps that appear to
402 be optimizable (or blocks that appear to be mergeable), but which really
403 must be left untouched (they are required to make it safely across
404 partition boundaries). See the comments at the top of
405 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
407 if (JUMP_P (BB_END (b)) && CROSSING_JUMP_P (BB_END (b)))
408 return false;
410 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
412 basic_block target, first;
413 location_t goto_locus;
414 int counter;
415 bool threaded = false;
416 int nthreaded_edges = 0;
417 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
419 /* Skip complex edges because we don't know how to update them.
421 Still handle fallthru edges, as we can succeed to forward fallthru
422 edge to the same place as the branch edge of conditional branch
423 and turn conditional branch to an unconditional branch. */
424 if (e->flags & EDGE_COMPLEX)
426 ei_next (&ei);
427 continue;
430 target = first = e->dest;
431 counter = NUM_FIXED_BLOCKS;
432 goto_locus = e->goto_locus;
434 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
435 up jumps that cross between hot/cold sections.
437 Basic block partitioning may result in some jumps that appear
438 to be optimizable (or blocks that appear to be mergeable), but which
439 really must be left untouched (they are required to make it safely
440 across partition boundaries). See the comments at the top of
441 bb-reorder.c:partition_hot_cold_basic_blocks for complete
442 details. */
444 if (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
445 && JUMP_P (BB_END (first))
446 && CROSSING_JUMP_P (BB_END (first)))
447 return changed;
449 while (counter < n_basic_blocks_for_fn (cfun))
451 basic_block new_target = NULL;
452 bool new_target_threaded = false;
453 may_thread |= (target->flags & BB_MODIFIED) != 0;
455 if (FORWARDER_BLOCK_P (target)
456 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
457 && single_succ (target) != EXIT_BLOCK_PTR_FOR_FN (cfun))
459 /* Bypass trivial infinite loops. */
460 new_target = single_succ (target);
461 if (target == new_target)
462 counter = n_basic_blocks_for_fn (cfun);
463 else if (!optimize)
465 /* When not optimizing, ensure that edges or forwarder
466 blocks with different locus are not optimized out. */
467 location_t new_locus = single_succ_edge (target)->goto_locus;
468 location_t locus = goto_locus;
470 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
471 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
472 && new_locus != locus)
473 new_target = NULL;
474 else
476 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
477 locus = new_locus;
479 rtx_insn *last = BB_END (target);
480 if (DEBUG_INSN_P (last))
481 last = prev_nondebug_insn (last);
482 if (last && INSN_P (last))
483 new_locus = INSN_LOCATION (last);
484 else
485 new_locus = UNKNOWN_LOCATION;
487 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
488 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
489 && new_locus != locus)
490 new_target = NULL;
491 else
493 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
494 locus = new_locus;
496 goto_locus = locus;
502 /* Allow to thread only over one edge at time to simplify updating
503 of probabilities. */
504 else if ((mode & CLEANUP_THREADING) && may_thread)
506 edge t = thread_jump (e, target);
507 if (t)
509 if (!threaded_edges)
510 threaded_edges = XNEWVEC (edge,
511 n_basic_blocks_for_fn (cfun));
512 else
514 int i;
516 /* Detect an infinite loop across blocks not
517 including the start block. */
518 for (i = 0; i < nthreaded_edges; ++i)
519 if (threaded_edges[i] == t)
520 break;
521 if (i < nthreaded_edges)
523 counter = n_basic_blocks_for_fn (cfun);
524 break;
528 /* Detect an infinite loop across the start block. */
529 if (t->dest == b)
530 break;
532 gcc_assert (nthreaded_edges
533 < (n_basic_blocks_for_fn (cfun)
534 - NUM_FIXED_BLOCKS));
535 threaded_edges[nthreaded_edges++] = t;
537 new_target = t->dest;
538 new_target_threaded = true;
542 if (!new_target)
543 break;
545 counter++;
546 target = new_target;
547 threaded |= new_target_threaded;
550 if (counter >= n_basic_blocks_for_fn (cfun))
552 if (dump_file)
553 fprintf (dump_file, "Infinite loop in BB %i.\n",
554 target->index);
556 else if (target == first)
557 ; /* We didn't do anything. */
558 else
560 /* Save the values now, as the edge may get removed. */
561 profile_count edge_count = e->count ();
562 int n = 0;
564 e->goto_locus = goto_locus;
566 /* Don't force if target is exit block. */
567 if (threaded && target != EXIT_BLOCK_PTR_FOR_FN (cfun))
569 notice_new_block (redirect_edge_and_branch_force (e, target));
570 if (dump_file)
571 fprintf (dump_file, "Conditionals threaded.\n");
573 else if (!redirect_edge_and_branch (e, target))
575 if (dump_file)
576 fprintf (dump_file,
577 "Forwarding edge %i->%i to %i failed.\n",
578 b->index, e->dest->index, target->index);
579 ei_next (&ei);
580 continue;
583 /* We successfully forwarded the edge. Now update profile
584 data: for each edge we traversed in the chain, remove
585 the original edge's execution count. */
588 edge t;
590 if (!single_succ_p (first))
592 gcc_assert (n < nthreaded_edges);
593 t = threaded_edges [n++];
594 gcc_assert (t->src == first);
595 update_bb_profile_for_threading (first, edge_count, t);
596 update_br_prob_note (first);
598 else
600 first->count -= edge_count;
601 /* It is possible that as the result of
602 threading we've removed edge as it is
603 threaded to the fallthru edge. Avoid
604 getting out of sync. */
605 if (n < nthreaded_edges
606 && first == threaded_edges [n]->src)
607 n++;
608 t = single_succ_edge (first);
611 first = t->dest;
613 while (first != target);
615 changed = true;
616 continue;
618 ei_next (&ei);
621 free (threaded_edges);
622 return changed;
626 /* Blocks A and B are to be merged into a single block. A has no incoming
627 fallthru edge, so it can be moved before B without adding or modifying
628 any jumps (aside from the jump from A to B). */
630 static void
631 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
633 rtx_insn *barrier;
635 /* If we are partitioning hot/cold basic blocks, we don't want to
636 mess up unconditional or indirect jumps that cross between hot
637 and cold sections.
639 Basic block partitioning may result in some jumps that appear to
640 be optimizable (or blocks that appear to be mergeable), but which really
641 must be left untouched (they are required to make it safely across
642 partition boundaries). See the comments at the top of
643 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
645 if (BB_PARTITION (a) != BB_PARTITION (b))
646 return;
648 barrier = next_nonnote_insn (BB_END (a));
649 gcc_assert (BARRIER_P (barrier));
650 delete_insn (barrier);
652 /* Scramble the insn chain. */
653 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
654 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
655 df_set_bb_dirty (a);
657 if (dump_file)
658 fprintf (dump_file, "Moved block %d before %d and merged.\n",
659 a->index, b->index);
661 /* Swap the records for the two blocks around. */
663 unlink_block (a);
664 link_block (a, b->prev_bb);
666 /* Now blocks A and B are contiguous. Merge them. */
667 merge_blocks (a, b);
670 /* Blocks A and B are to be merged into a single block. B has no outgoing
671 fallthru edge, so it can be moved after A without adding or modifying
672 any jumps (aside from the jump from A to B). */
674 static void
675 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
677 rtx_insn *barrier, *real_b_end;
678 rtx_insn *label;
679 rtx_jump_table_data *table;
681 /* If we are partitioning hot/cold basic blocks, we don't want to
682 mess up unconditional or indirect jumps that cross between hot
683 and cold sections.
685 Basic block partitioning may result in some jumps that appear to
686 be optimizable (or blocks that appear to be mergeable), but which really
687 must be left untouched (they are required to make it safely across
688 partition boundaries). See the comments at the top of
689 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
691 if (BB_PARTITION (a) != BB_PARTITION (b))
692 return;
694 real_b_end = BB_END (b);
696 /* If there is a jump table following block B temporarily add the jump table
697 to block B so that it will also be moved to the correct location. */
698 if (tablejump_p (BB_END (b), &label, &table)
699 && prev_active_insn (label) == BB_END (b))
701 BB_END (b) = table;
704 /* There had better have been a barrier there. Delete it. */
705 barrier = NEXT_INSN (BB_END (b));
706 if (barrier && BARRIER_P (barrier))
707 delete_insn (barrier);
710 /* Scramble the insn chain. */
711 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
713 /* Restore the real end of b. */
714 BB_END (b) = real_b_end;
716 if (dump_file)
717 fprintf (dump_file, "Moved block %d after %d and merged.\n",
718 b->index, a->index);
720 /* Now blocks A and B are contiguous. Merge them. */
721 merge_blocks (a, b);
724 /* Attempt to merge basic blocks that are potentially non-adjacent.
725 Return NULL iff the attempt failed, otherwise return basic block
726 where cleanup_cfg should continue. Because the merging commonly
727 moves basic block away or introduces another optimization
728 possibility, return basic block just before B so cleanup_cfg don't
729 need to iterate.
731 It may be good idea to return basic block before C in the case
732 C has been moved after B and originally appeared earlier in the
733 insn sequence, but we have no information available about the
734 relative ordering of these two. Hopefully it is not too common. */
736 static basic_block
737 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
739 basic_block next;
741 /* If we are partitioning hot/cold basic blocks, we don't want to
742 mess up unconditional or indirect jumps that cross between hot
743 and cold sections.
745 Basic block partitioning may result in some jumps that appear to
746 be optimizable (or blocks that appear to be mergeable), but which really
747 must be left untouched (they are required to make it safely across
748 partition boundaries). See the comments at the top of
749 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
751 if (BB_PARTITION (b) != BB_PARTITION (c))
752 return NULL;
754 /* If B has a fallthru edge to C, no need to move anything. */
755 if (e->flags & EDGE_FALLTHRU)
757 int b_index = b->index, c_index = c->index;
759 /* Protect the loop latches. */
760 if (current_loops && c->loop_father->latch == c)
761 return NULL;
763 merge_blocks (b, c);
764 update_forwarder_flag (b);
766 if (dump_file)
767 fprintf (dump_file, "Merged %d and %d without moving.\n",
768 b_index, c_index);
770 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
773 /* Otherwise we will need to move code around. Do that only if expensive
774 transformations are allowed. */
775 else if (mode & CLEANUP_EXPENSIVE)
777 edge tmp_edge, b_fallthru_edge;
778 bool c_has_outgoing_fallthru;
779 bool b_has_incoming_fallthru;
781 /* Avoid overactive code motion, as the forwarder blocks should be
782 eliminated by edge redirection instead. One exception might have
783 been if B is a forwarder block and C has no fallthru edge, but
784 that should be cleaned up by bb-reorder instead. */
785 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
786 return NULL;
788 /* We must make sure to not munge nesting of lexical blocks,
789 and loop notes. This is done by squeezing out all the notes
790 and leaving them there to lie. Not ideal, but functional. */
792 tmp_edge = find_fallthru_edge (c->succs);
793 c_has_outgoing_fallthru = (tmp_edge != NULL);
795 tmp_edge = find_fallthru_edge (b->preds);
796 b_has_incoming_fallthru = (tmp_edge != NULL);
797 b_fallthru_edge = tmp_edge;
798 next = b->prev_bb;
799 if (next == c)
800 next = next->prev_bb;
802 /* Otherwise, we're going to try to move C after B. If C does
803 not have an outgoing fallthru, then it can be moved
804 immediately after B without introducing or modifying jumps. */
805 if (! c_has_outgoing_fallthru)
807 merge_blocks_move_successor_nojumps (b, c);
808 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
811 /* If B does not have an incoming fallthru, then it can be moved
812 immediately before C without introducing or modifying jumps.
813 C cannot be the first block, so we do not have to worry about
814 accessing a non-existent block. */
816 if (b_has_incoming_fallthru)
818 basic_block bb;
820 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
821 return NULL;
822 bb = force_nonfallthru (b_fallthru_edge);
823 if (bb)
824 notice_new_block (bb);
827 merge_blocks_move_predecessor_nojumps (b, c);
828 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
831 return NULL;
835 /* Removes the memory attributes of MEM expression
836 if they are not equal. */
838 static void
839 merge_memattrs (rtx x, rtx y)
841 int i;
842 int j;
843 enum rtx_code code;
844 const char *fmt;
846 if (x == y)
847 return;
848 if (x == 0 || y == 0)
849 return;
851 code = GET_CODE (x);
853 if (code != GET_CODE (y))
854 return;
856 if (GET_MODE (x) != GET_MODE (y))
857 return;
859 if (code == MEM && !mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
861 if (! MEM_ATTRS (x))
862 MEM_ATTRS (y) = 0;
863 else if (! MEM_ATTRS (y))
864 MEM_ATTRS (x) = 0;
865 else
867 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
869 set_mem_alias_set (x, 0);
870 set_mem_alias_set (y, 0);
873 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
875 set_mem_expr (x, 0);
876 set_mem_expr (y, 0);
877 clear_mem_offset (x);
878 clear_mem_offset (y);
880 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
881 || (MEM_OFFSET_KNOWN_P (x)
882 && maybe_ne (MEM_OFFSET (x), MEM_OFFSET (y))))
884 clear_mem_offset (x);
885 clear_mem_offset (y);
888 if (!MEM_SIZE_KNOWN_P (x))
889 clear_mem_size (y);
890 else if (!MEM_SIZE_KNOWN_P (y))
891 clear_mem_size (x);
892 else if (known_le (MEM_SIZE (x), MEM_SIZE (y)))
893 set_mem_size (x, MEM_SIZE (y));
894 else if (known_le (MEM_SIZE (y), MEM_SIZE (x)))
895 set_mem_size (y, MEM_SIZE (x));
896 else
898 /* The sizes aren't ordered, so we can't merge them. */
899 clear_mem_size (x);
900 clear_mem_size (y);
903 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
904 set_mem_align (y, MEM_ALIGN (x));
907 if (code == MEM)
909 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
911 MEM_READONLY_P (x) = 0;
912 MEM_READONLY_P (y) = 0;
914 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
916 MEM_NOTRAP_P (x) = 0;
917 MEM_NOTRAP_P (y) = 0;
919 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
921 MEM_VOLATILE_P (x) = 1;
922 MEM_VOLATILE_P (y) = 1;
926 fmt = GET_RTX_FORMAT (code);
927 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
929 switch (fmt[i])
931 case 'E':
932 /* Two vectors must have the same length. */
933 if (XVECLEN (x, i) != XVECLEN (y, i))
934 return;
936 for (j = 0; j < XVECLEN (x, i); j++)
937 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
939 break;
941 case 'e':
942 merge_memattrs (XEXP (x, i), XEXP (y, i));
945 return;
949 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
950 different single sets S1 and S2. */
952 static bool
953 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
955 int i;
956 rtx e1, e2;
958 if (p1 == s1 && p2 == s2)
959 return true;
961 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
962 return false;
964 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
965 return false;
967 for (i = 0; i < XVECLEN (p1, 0); i++)
969 e1 = XVECEXP (p1, 0, i);
970 e2 = XVECEXP (p2, 0, i);
971 if (e1 == s1 && e2 == s2)
972 continue;
973 if (reload_completed
974 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
975 continue;
977 return false;
980 return true;
984 /* NOTE1 is the REG_EQUAL note, if any, attached to an insn
985 that is a single_set with a SET_SRC of SRC1. Similarly
986 for NOTE2/SRC2.
988 So effectively NOTE1/NOTE2 are an alternate form of
989 SRC1/SRC2 respectively.
991 Return nonzero if SRC1 or NOTE1 has the same constant
992 integer value as SRC2 or NOTE2. Else return zero. */
993 static int
994 values_equal_p (rtx note1, rtx note2, rtx src1, rtx src2)
996 if (note1
997 && note2
998 && CONST_INT_P (XEXP (note1, 0))
999 && rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0)))
1000 return 1;
1002 if (!note1
1003 && !note2
1004 && CONST_INT_P (src1)
1005 && CONST_INT_P (src2)
1006 && rtx_equal_p (src1, src2))
1007 return 1;
1009 if (note1
1010 && CONST_INT_P (src2)
1011 && rtx_equal_p (XEXP (note1, 0), src2))
1012 return 1;
1014 if (note2
1015 && CONST_INT_P (src1)
1016 && rtx_equal_p (XEXP (note2, 0), src1))
1017 return 1;
1019 return 0;
1022 /* Examine register notes on I1 and I2 and return:
1023 - dir_forward if I1 can be replaced by I2, or
1024 - dir_backward if I2 can be replaced by I1, or
1025 - dir_both if both are the case. */
1027 static enum replace_direction
1028 can_replace_by (rtx_insn *i1, rtx_insn *i2)
1030 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1031 bool c1, c2;
1033 /* Check for 2 sets. */
1034 s1 = single_set (i1);
1035 s2 = single_set (i2);
1036 if (s1 == NULL_RTX || s2 == NULL_RTX)
1037 return dir_none;
1039 /* Check that the 2 sets set the same dest. */
1040 d1 = SET_DEST (s1);
1041 d2 = SET_DEST (s2);
1042 if (!(reload_completed
1043 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1044 return dir_none;
1046 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1047 set dest to the same value. */
1048 note1 = find_reg_equal_equiv_note (i1);
1049 note2 = find_reg_equal_equiv_note (i2);
1051 src1 = SET_SRC (s1);
1052 src2 = SET_SRC (s2);
1054 if (!values_equal_p (note1, note2, src1, src2))
1055 return dir_none;
1057 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1058 return dir_none;
1060 /* Although the 2 sets set dest to the same value, we cannot replace
1061 (set (dest) (const_int))
1063 (set (dest) (reg))
1064 because we don't know if the reg is live and has the same value at the
1065 location of replacement. */
1066 c1 = CONST_INT_P (src1);
1067 c2 = CONST_INT_P (src2);
1068 if (c1 && c2)
1069 return dir_both;
1070 else if (c2)
1071 return dir_forward;
1072 else if (c1)
1073 return dir_backward;
1075 return dir_none;
1078 /* Merges directions A and B. */
1080 static enum replace_direction
1081 merge_dir (enum replace_direction a, enum replace_direction b)
1083 /* Implements the following table:
1084 |bo fw bw no
1085 ---+-----------
1086 bo |bo fw bw no
1087 fw |-- fw no no
1088 bw |-- -- bw no
1089 no |-- -- -- no. */
1091 if (a == b)
1092 return a;
1094 if (a == dir_both)
1095 return b;
1096 if (b == dir_both)
1097 return a;
1099 return dir_none;
1102 /* Array of flags indexed by reg note kind, true if the given
1103 reg note is CFA related. */
1104 static const bool reg_note_cfa_p[] = {
1105 #undef REG_CFA_NOTE
1106 #define DEF_REG_NOTE(NAME) false,
1107 #define REG_CFA_NOTE(NAME) true,
1108 #include "reg-notes.def"
1109 #undef REG_CFA_NOTE
1110 #undef DEF_REG_NOTE
1111 false
1114 /* Return true if I1 and I2 have identical CFA notes (the same order
1115 and equivalent content). */
1117 static bool
1118 insns_have_identical_cfa_notes (rtx_insn *i1, rtx_insn *i2)
1120 rtx n1, n2;
1121 for (n1 = REG_NOTES (i1), n2 = REG_NOTES (i2); ;
1122 n1 = XEXP (n1, 1), n2 = XEXP (n2, 1))
1124 /* Skip over reg notes not related to CFI information. */
1125 while (n1 && !reg_note_cfa_p[REG_NOTE_KIND (n1)])
1126 n1 = XEXP (n1, 1);
1127 while (n2 && !reg_note_cfa_p[REG_NOTE_KIND (n2)])
1128 n2 = XEXP (n2, 1);
1129 if (n1 == NULL_RTX && n2 == NULL_RTX)
1130 return true;
1131 if (n1 == NULL_RTX || n2 == NULL_RTX)
1132 return false;
1133 if (XEXP (n1, 0) == XEXP (n2, 0))
1135 else if (XEXP (n1, 0) == NULL_RTX || XEXP (n2, 0) == NULL_RTX)
1136 return false;
1137 else if (!(reload_completed
1138 ? rtx_renumbered_equal_p (XEXP (n1, 0), XEXP (n2, 0))
1139 : rtx_equal_p (XEXP (n1, 0), XEXP (n2, 0))))
1140 return false;
1144 /* Examine I1 and I2 and return:
1145 - dir_forward if I1 can be replaced by I2, or
1146 - dir_backward if I2 can be replaced by I1, or
1147 - dir_both if both are the case. */
1149 static enum replace_direction
1150 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1152 rtx p1, p2;
1154 /* Verify that I1 and I2 are equivalent. */
1155 if (GET_CODE (i1) != GET_CODE (i2))
1156 return dir_none;
1158 /* __builtin_unreachable() may lead to empty blocks (ending with
1159 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1160 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1161 return dir_both;
1163 /* ??? Do not allow cross-jumping between different stack levels. */
1164 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1165 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1166 if (p1 && p2)
1168 p1 = XEXP (p1, 0);
1169 p2 = XEXP (p2, 0);
1170 if (!rtx_equal_p (p1, p2))
1171 return dir_none;
1173 /* ??? Worse, this adjustment had better be constant lest we
1174 have differing incoming stack levels. */
1175 if (!frame_pointer_needed
1176 && known_eq (find_args_size_adjust (i1), HOST_WIDE_INT_MIN))
1177 return dir_none;
1179 else if (p1 || p2)
1180 return dir_none;
1182 /* Do not allow cross-jumping between frame related insns and other
1183 insns. */
1184 if (RTX_FRAME_RELATED_P (i1) != RTX_FRAME_RELATED_P (i2))
1185 return dir_none;
1187 p1 = PATTERN (i1);
1188 p2 = PATTERN (i2);
1190 if (GET_CODE (p1) != GET_CODE (p2))
1191 return dir_none;
1193 /* If this is a CALL_INSN, compare register usage information.
1194 If we don't check this on stack register machines, the two
1195 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1196 numbers of stack registers in the same basic block.
1197 If we don't check this on machines with delay slots, a delay slot may
1198 be filled that clobbers a parameter expected by the subroutine.
1200 ??? We take the simple route for now and assume that if they're
1201 equal, they were constructed identically.
1203 Also check for identical exception regions. */
1205 if (CALL_P (i1))
1207 /* Ensure the same EH region. */
1208 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1209 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1211 if (!n1 && n2)
1212 return dir_none;
1214 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1215 return dir_none;
1217 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1218 CALL_INSN_FUNCTION_USAGE (i2))
1219 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1220 return dir_none;
1222 /* For address sanitizer, never crossjump __asan_report_* builtins,
1223 otherwise errors might be reported on incorrect lines. */
1224 if (flag_sanitize & SANITIZE_ADDRESS)
1226 rtx call = get_call_rtx_from (i1);
1227 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1229 rtx symbol = XEXP (XEXP (call, 0), 0);
1230 if (SYMBOL_REF_DECL (symbol)
1231 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1233 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1234 == BUILT_IN_NORMAL)
1235 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1236 >= BUILT_IN_ASAN_REPORT_LOAD1
1237 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1238 <= BUILT_IN_ASAN_STOREN)
1239 return dir_none;
1245 /* If both i1 and i2 are frame related, verify all the CFA notes
1246 in the same order and with the same content. */
1247 if (RTX_FRAME_RELATED_P (i1) && !insns_have_identical_cfa_notes (i1, i2))
1248 return dir_none;
1250 #ifdef STACK_REGS
1251 /* If cross_jump_death_matters is not 0, the insn's mode
1252 indicates whether or not the insn contains any stack-like
1253 regs. */
1255 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1257 /* If register stack conversion has already been done, then
1258 death notes must also be compared before it is certain that
1259 the two instruction streams match. */
1261 rtx note;
1262 HARD_REG_SET i1_regset, i2_regset;
1264 CLEAR_HARD_REG_SET (i1_regset);
1265 CLEAR_HARD_REG_SET (i2_regset);
1267 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1268 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1269 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1271 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1272 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1273 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1275 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1276 return dir_none;
1278 #endif
1280 if (reload_completed
1281 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1282 return dir_both;
1284 return can_replace_by (i1, i2);
1287 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1288 flow_find_head_matching_sequence, ensure the notes match. */
1290 static void
1291 merge_notes (rtx_insn *i1, rtx_insn *i2)
1293 /* If the merged insns have different REG_EQUAL notes, then
1294 remove them. */
1295 rtx equiv1 = find_reg_equal_equiv_note (i1);
1296 rtx equiv2 = find_reg_equal_equiv_note (i2);
1298 if (equiv1 && !equiv2)
1299 remove_note (i1, equiv1);
1300 else if (!equiv1 && equiv2)
1301 remove_note (i2, equiv2);
1302 else if (equiv1 && equiv2
1303 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1305 remove_note (i1, equiv1);
1306 remove_note (i2, equiv2);
1310 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1311 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1312 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1313 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1314 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1316 static void
1317 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1318 bool *did_fallthru)
1320 edge fallthru;
1322 *did_fallthru = false;
1324 /* Ignore notes. */
1325 while (!NONDEBUG_INSN_P (*i1))
1327 if (*i1 != BB_HEAD (*bb1))
1329 *i1 = PREV_INSN (*i1);
1330 continue;
1333 if (!follow_fallthru)
1334 return;
1336 fallthru = find_fallthru_edge ((*bb1)->preds);
1337 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1338 || !single_succ_p (fallthru->src))
1339 return;
1341 *bb1 = fallthru->src;
1342 *i1 = BB_END (*bb1);
1343 *did_fallthru = true;
1347 /* Look through the insns at the end of BB1 and BB2 and find the longest
1348 sequence that are either equivalent, or allow forward or backward
1349 replacement. Store the first insns for that sequence in *F1 and *F2 and
1350 return the sequence length.
1352 DIR_P indicates the allowed replacement direction on function entry, and
1353 the actual replacement direction on function exit. If NULL, only equivalent
1354 sequences are allowed.
1356 To simplify callers of this function, if the blocks match exactly,
1357 store the head of the blocks in *F1 and *F2. */
1360 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1361 rtx_insn **f2, enum replace_direction *dir_p)
1363 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1364 int ninsns = 0;
1365 enum replace_direction dir, last_dir, afterlast_dir;
1366 bool follow_fallthru, did_fallthru;
1368 if (dir_p)
1369 dir = *dir_p;
1370 else
1371 dir = dir_both;
1372 afterlast_dir = dir;
1373 last_dir = afterlast_dir;
1375 /* Skip simple jumps at the end of the blocks. Complex jumps still
1376 need to be compared for equivalence, which we'll do below. */
1378 i1 = BB_END (bb1);
1379 last1 = afterlast1 = last2 = afterlast2 = NULL;
1380 if (onlyjump_p (i1)
1381 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1383 last1 = i1;
1384 i1 = PREV_INSN (i1);
1387 i2 = BB_END (bb2);
1388 if (onlyjump_p (i2)
1389 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1391 last2 = i2;
1392 /* Count everything except for unconditional jump as insn.
1393 Don't count any jumps if dir_p is NULL. */
1394 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1395 ninsns++;
1396 i2 = PREV_INSN (i2);
1399 while (true)
1401 /* In the following example, we can replace all jumps to C by jumps to A.
1403 This removes 4 duplicate insns.
1404 [bb A] insn1 [bb C] insn1
1405 insn2 insn2
1406 [bb B] insn3 insn3
1407 insn4 insn4
1408 jump_insn jump_insn
1410 We could also replace all jumps to A by jumps to C, but that leaves B
1411 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1412 step, all jumps to B would be replaced with jumps to the middle of C,
1413 achieving the same result with more effort.
1414 So we allow only the first possibility, which means that we don't allow
1415 fallthru in the block that's being replaced. */
1417 follow_fallthru = dir_p && dir != dir_forward;
1418 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1419 if (did_fallthru)
1420 dir = dir_backward;
1422 follow_fallthru = dir_p && dir != dir_backward;
1423 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1424 if (did_fallthru)
1425 dir = dir_forward;
1427 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1428 break;
1430 /* Do not turn corssing edge to non-crossing or vice versa after
1431 reload. */
1432 if (BB_PARTITION (BLOCK_FOR_INSN (i1))
1433 != BB_PARTITION (BLOCK_FOR_INSN (i2))
1434 && reload_completed)
1435 break;
1437 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1438 if (dir == dir_none || (!dir_p && dir != dir_both))
1439 break;
1441 merge_memattrs (i1, i2);
1443 /* Don't begin a cross-jump with a NOTE insn. */
1444 if (INSN_P (i1))
1446 merge_notes (i1, i2);
1448 afterlast1 = last1, afterlast2 = last2;
1449 last1 = i1, last2 = i2;
1450 afterlast_dir = last_dir;
1451 last_dir = dir;
1452 if (active_insn_p (i1))
1453 ninsns++;
1456 i1 = PREV_INSN (i1);
1457 i2 = PREV_INSN (i2);
1460 /* Don't allow the insn after a compare to be shared by
1461 cross-jumping unless the compare is also shared. */
1462 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1463 && ! sets_cc0_p (last1))
1464 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1466 /* Include preceding notes and labels in the cross-jump. One,
1467 this may bring us to the head of the blocks as requested above.
1468 Two, it keeps line number notes as matched as may be. */
1469 if (ninsns)
1471 bb1 = BLOCK_FOR_INSN (last1);
1472 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1473 last1 = PREV_INSN (last1);
1475 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1476 last1 = PREV_INSN (last1);
1478 bb2 = BLOCK_FOR_INSN (last2);
1479 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1480 last2 = PREV_INSN (last2);
1482 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1483 last2 = PREV_INSN (last2);
1485 *f1 = last1;
1486 *f2 = last2;
1489 if (dir_p)
1490 *dir_p = last_dir;
1491 return ninsns;
1494 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1495 the head of the two blocks. Do not include jumps at the end.
1496 If STOP_AFTER is nonzero, stop after finding that many matching
1497 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1498 non-zero, only count active insns. */
1501 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1502 rtx_insn **f2, int stop_after)
1504 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1505 int ninsns = 0;
1506 edge e;
1507 edge_iterator ei;
1508 int nehedges1 = 0, nehedges2 = 0;
1510 FOR_EACH_EDGE (e, ei, bb1->succs)
1511 if (e->flags & EDGE_EH)
1512 nehedges1++;
1513 FOR_EACH_EDGE (e, ei, bb2->succs)
1514 if (e->flags & EDGE_EH)
1515 nehedges2++;
1517 i1 = BB_HEAD (bb1);
1518 i2 = BB_HEAD (bb2);
1519 last1 = beforelast1 = last2 = beforelast2 = NULL;
1521 while (true)
1523 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1524 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1526 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1527 break;
1528 i1 = NEXT_INSN (i1);
1531 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1533 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1534 break;
1535 i2 = NEXT_INSN (i2);
1538 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1539 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1540 break;
1542 if (NOTE_P (i1) || NOTE_P (i2)
1543 || JUMP_P (i1) || JUMP_P (i2))
1544 break;
1546 /* A sanity check to make sure we're not merging insns with different
1547 effects on EH. If only one of them ends a basic block, it shouldn't
1548 have an EH edge; if both end a basic block, there should be the same
1549 number of EH edges. */
1550 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1551 && nehedges1 > 0)
1552 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1553 && nehedges2 > 0)
1554 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1555 && nehedges1 != nehedges2))
1556 break;
1558 if (old_insns_match_p (0, i1, i2) != dir_both)
1559 break;
1561 merge_memattrs (i1, i2);
1563 /* Don't begin a cross-jump with a NOTE insn. */
1564 if (INSN_P (i1))
1566 merge_notes (i1, i2);
1568 beforelast1 = last1, beforelast2 = last2;
1569 last1 = i1, last2 = i2;
1570 if (!stop_after || active_insn_p (i1))
1571 ninsns++;
1574 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1575 || (stop_after > 0 && ninsns == stop_after))
1576 break;
1578 i1 = NEXT_INSN (i1);
1579 i2 = NEXT_INSN (i2);
1582 /* Don't allow a compare to be shared by cross-jumping unless the insn
1583 after the compare is also shared. */
1584 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1585 && sets_cc0_p (last1))
1586 last1 = beforelast1, last2 = beforelast2, ninsns--;
1588 if (ninsns)
1590 *f1 = last1;
1591 *f2 = last2;
1594 return ninsns;
1597 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1598 the branch instruction. This means that if we commonize the control
1599 flow before end of the basic block, the semantic remains unchanged.
1601 We may assume that there exists one edge with a common destination. */
1603 static bool
1604 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1606 int nehedges1 = 0, nehedges2 = 0;
1607 edge fallthru1 = 0, fallthru2 = 0;
1608 edge e1, e2;
1609 edge_iterator ei;
1611 /* If we performed shrink-wrapping, edges to the exit block can
1612 only be distinguished for JUMP_INSNs. The two paths may differ in
1613 whether they went through the prologue. Sibcalls are fine, we know
1614 that we either didn't need or inserted an epilogue before them. */
1615 if (crtl->shrink_wrapped
1616 && single_succ_p (bb1)
1617 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1618 && !JUMP_P (BB_END (bb1))
1619 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1620 return false;
1622 /* If BB1 has only one successor, we may be looking at either an
1623 unconditional jump, or a fake edge to exit. */
1624 if (single_succ_p (bb1)
1625 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1626 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1627 return (single_succ_p (bb2)
1628 && (single_succ_edge (bb2)->flags
1629 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1630 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1632 /* Match conditional jumps - this may get tricky when fallthru and branch
1633 edges are crossed. */
1634 if (EDGE_COUNT (bb1->succs) == 2
1635 && any_condjump_p (BB_END (bb1))
1636 && onlyjump_p (BB_END (bb1)))
1638 edge b1, f1, b2, f2;
1639 bool reverse, match;
1640 rtx set1, set2, cond1, cond2;
1641 enum rtx_code code1, code2;
1643 if (EDGE_COUNT (bb2->succs) != 2
1644 || !any_condjump_p (BB_END (bb2))
1645 || !onlyjump_p (BB_END (bb2)))
1646 return false;
1648 b1 = BRANCH_EDGE (bb1);
1649 b2 = BRANCH_EDGE (bb2);
1650 f1 = FALLTHRU_EDGE (bb1);
1651 f2 = FALLTHRU_EDGE (bb2);
1653 /* Get around possible forwarders on fallthru edges. Other cases
1654 should be optimized out already. */
1655 if (FORWARDER_BLOCK_P (f1->dest))
1656 f1 = single_succ_edge (f1->dest);
1658 if (FORWARDER_BLOCK_P (f2->dest))
1659 f2 = single_succ_edge (f2->dest);
1661 /* To simplify use of this function, return false if there are
1662 unneeded forwarder blocks. These will get eliminated later
1663 during cleanup_cfg. */
1664 if (FORWARDER_BLOCK_P (f1->dest)
1665 || FORWARDER_BLOCK_P (f2->dest)
1666 || FORWARDER_BLOCK_P (b1->dest)
1667 || FORWARDER_BLOCK_P (b2->dest))
1668 return false;
1670 if (f1->dest == f2->dest && b1->dest == b2->dest)
1671 reverse = false;
1672 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1673 reverse = true;
1674 else
1675 return false;
1677 set1 = pc_set (BB_END (bb1));
1678 set2 = pc_set (BB_END (bb2));
1679 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1680 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1681 reverse = !reverse;
1683 cond1 = XEXP (SET_SRC (set1), 0);
1684 cond2 = XEXP (SET_SRC (set2), 0);
1685 code1 = GET_CODE (cond1);
1686 if (reverse)
1687 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1688 else
1689 code2 = GET_CODE (cond2);
1691 if (code2 == UNKNOWN)
1692 return false;
1694 /* Verify codes and operands match. */
1695 match = ((code1 == code2
1696 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1697 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1698 || (code1 == swap_condition (code2)
1699 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1700 XEXP (cond2, 0))
1701 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1702 XEXP (cond2, 1))));
1704 /* If we return true, we will join the blocks. Which means that
1705 we will only have one branch prediction bit to work with. Thus
1706 we require the existing branches to have probabilities that are
1707 roughly similar. */
1708 if (match
1709 && optimize_bb_for_speed_p (bb1)
1710 && optimize_bb_for_speed_p (bb2))
1712 profile_probability prob2;
1714 if (b1->dest == b2->dest)
1715 prob2 = b2->probability;
1716 else
1717 /* Do not use f2 probability as f2 may be forwarded. */
1718 prob2 = b2->probability.invert ();
1720 /* Fail if the difference in probabilities is greater than 50%.
1721 This rules out two well-predicted branches with opposite
1722 outcomes. */
1723 if (b1->probability.differs_lot_from_p (prob2))
1725 if (dump_file)
1727 fprintf (dump_file,
1728 "Outcomes of branch in bb %i and %i differ too"
1729 " much (", bb1->index, bb2->index);
1730 b1->probability.dump (dump_file);
1731 prob2.dump (dump_file);
1732 fprintf (dump_file, ")\n");
1734 return false;
1738 if (dump_file && match)
1739 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1740 bb1->index, bb2->index);
1742 return match;
1745 /* Generic case - we are seeing a computed jump, table jump or trapping
1746 instruction. */
1748 /* Check whether there are tablejumps in the end of BB1 and BB2.
1749 Return true if they are identical. */
1751 rtx_insn *label1, *label2;
1752 rtx_jump_table_data *table1, *table2;
1754 if (tablejump_p (BB_END (bb1), &label1, &table1)
1755 && tablejump_p (BB_END (bb2), &label2, &table2)
1756 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1758 /* The labels should never be the same rtx. If they really are same
1759 the jump tables are same too. So disable crossjumping of blocks BB1
1760 and BB2 because when deleting the common insns in the end of BB1
1761 by delete_basic_block () the jump table would be deleted too. */
1762 /* If LABEL2 is referenced in BB1->END do not do anything
1763 because we would loose information when replacing
1764 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1765 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1767 /* Set IDENTICAL to true when the tables are identical. */
1768 bool identical = false;
1769 rtx p1, p2;
1771 p1 = PATTERN (table1);
1772 p2 = PATTERN (table2);
1773 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1775 identical = true;
1777 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1778 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1779 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1780 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1782 int i;
1784 identical = true;
1785 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1786 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1787 identical = false;
1790 if (identical)
1792 bool match;
1794 /* Temporarily replace references to LABEL1 with LABEL2
1795 in BB1->END so that we could compare the instructions. */
1796 replace_label_in_insn (BB_END (bb1), label1, label2, false);
1798 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1799 == dir_both);
1800 if (dump_file && match)
1801 fprintf (dump_file,
1802 "Tablejumps in bb %i and %i match.\n",
1803 bb1->index, bb2->index);
1805 /* Set the original label in BB1->END because when deleting
1806 a block whose end is a tablejump, the tablejump referenced
1807 from the instruction is deleted too. */
1808 replace_label_in_insn (BB_END (bb1), label2, label1, false);
1810 return match;
1813 return false;
1817 /* Find the last non-debug non-note instruction in each bb, except
1818 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1819 handles that case specially. old_insns_match_p does not handle
1820 other types of instruction notes. */
1821 rtx_insn *last1 = BB_END (bb1);
1822 rtx_insn *last2 = BB_END (bb2);
1823 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1824 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1825 last1 = PREV_INSN (last1);
1826 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1827 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1828 last2 = PREV_INSN (last2);
1829 gcc_assert (last1 && last2);
1831 /* First ensure that the instructions match. There may be many outgoing
1832 edges so this test is generally cheaper. */
1833 if (old_insns_match_p (mode, last1, last2) != dir_both)
1834 return false;
1836 /* Search the outgoing edges, ensure that the counts do match, find possible
1837 fallthru and exception handling edges since these needs more
1838 validation. */
1839 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1840 return false;
1842 bool nonfakeedges = false;
1843 FOR_EACH_EDGE (e1, ei, bb1->succs)
1845 e2 = EDGE_SUCC (bb2, ei.index);
1847 if ((e1->flags & EDGE_FAKE) == 0)
1848 nonfakeedges = true;
1850 if (e1->flags & EDGE_EH)
1851 nehedges1++;
1853 if (e2->flags & EDGE_EH)
1854 nehedges2++;
1856 if (e1->flags & EDGE_FALLTHRU)
1857 fallthru1 = e1;
1858 if (e2->flags & EDGE_FALLTHRU)
1859 fallthru2 = e2;
1862 /* If number of edges of various types does not match, fail. */
1863 if (nehedges1 != nehedges2
1864 || (fallthru1 != 0) != (fallthru2 != 0))
1865 return false;
1867 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1868 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1869 attempt to optimize, as the two basic blocks might have different
1870 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1871 traps there should be REG_ARG_SIZE notes, they could be missing
1872 for __builtin_unreachable () uses though. */
1873 if (!nonfakeedges
1874 && !ACCUMULATE_OUTGOING_ARGS
1875 && (!INSN_P (last1)
1876 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1877 return false;
1879 /* fallthru edges must be forwarded to the same destination. */
1880 if (fallthru1)
1882 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1883 ? single_succ (fallthru1->dest): fallthru1->dest);
1884 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1885 ? single_succ (fallthru2->dest): fallthru2->dest);
1887 if (d1 != d2)
1888 return false;
1891 /* Ensure the same EH region. */
1893 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1894 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1896 if (!n1 && n2)
1897 return false;
1899 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1900 return false;
1903 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1904 version of sequence abstraction. */
1905 FOR_EACH_EDGE (e1, ei, bb2->succs)
1907 edge e2;
1908 edge_iterator ei;
1909 basic_block d1 = e1->dest;
1911 if (FORWARDER_BLOCK_P (d1))
1912 d1 = EDGE_SUCC (d1, 0)->dest;
1914 FOR_EACH_EDGE (e2, ei, bb1->succs)
1916 basic_block d2 = e2->dest;
1917 if (FORWARDER_BLOCK_P (d2))
1918 d2 = EDGE_SUCC (d2, 0)->dest;
1919 if (d1 == d2)
1920 break;
1923 if (!e2)
1924 return false;
1927 return true;
1930 /* Returns true if BB basic block has a preserve label. */
1932 static bool
1933 block_has_preserve_label (basic_block bb)
1935 return (bb
1936 && block_label (bb)
1937 && LABEL_PRESERVE_P (block_label (bb)));
1940 /* E1 and E2 are edges with the same destination block. Search their
1941 predecessors for common code. If found, redirect control flow from
1942 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1943 or the other way around (dir_backward). DIR specifies the allowed
1944 replacement direction. */
1946 static bool
1947 try_crossjump_to_edge (int mode, edge e1, edge e2,
1948 enum replace_direction dir)
1950 int nmatch;
1951 basic_block src1 = e1->src, src2 = e2->src;
1952 basic_block redirect_to, redirect_from, to_remove;
1953 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1954 rtx_insn *newpos1, *newpos2;
1955 edge s;
1956 edge_iterator ei;
1958 newpos1 = newpos2 = NULL;
1960 /* Search backward through forwarder blocks. We don't need to worry
1961 about multiple entry or chained forwarders, as they will be optimized
1962 away. We do this to look past the unconditional jump following a
1963 conditional jump that is required due to the current CFG shape. */
1964 if (single_pred_p (src1)
1965 && FORWARDER_BLOCK_P (src1))
1966 e1 = single_pred_edge (src1), src1 = e1->src;
1968 if (single_pred_p (src2)
1969 && FORWARDER_BLOCK_P (src2))
1970 e2 = single_pred_edge (src2), src2 = e2->src;
1972 /* Nothing to do if we reach ENTRY, or a common source block. */
1973 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1974 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1975 return false;
1976 if (src1 == src2)
1977 return false;
1979 /* Seeing more than 1 forwarder blocks would confuse us later... */
1980 if (FORWARDER_BLOCK_P (e1->dest)
1981 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1982 return false;
1984 if (FORWARDER_BLOCK_P (e2->dest)
1985 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1986 return false;
1988 /* Likewise with dead code (possibly newly created by the other optimizations
1989 of cfg_cleanup). */
1990 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1991 return false;
1993 /* Do not turn corssing edge to non-crossing or vice versa after reload. */
1994 if (BB_PARTITION (src1) != BB_PARTITION (src2)
1995 && reload_completed)
1996 return false;
1998 /* Look for the common insn sequence, part the first ... */
1999 if (!outgoing_edges_match (mode, src1, src2))
2000 return false;
2002 /* ... and part the second. */
2003 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
2005 osrc1 = src1;
2006 osrc2 = src2;
2007 if (newpos1 != NULL_RTX)
2008 src1 = BLOCK_FOR_INSN (newpos1);
2009 if (newpos2 != NULL_RTX)
2010 src2 = BLOCK_FOR_INSN (newpos2);
2012 /* Check that SRC1 and SRC2 have preds again. They may have changed
2013 above due to the call to flow_find_cross_jump. */
2014 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
2015 return false;
2017 if (dir == dir_backward)
2019 std::swap (osrc1, osrc2);
2020 std::swap (src1, src2);
2021 std::swap (e1, e2);
2022 std::swap (newpos1, newpos2);
2025 /* Don't proceed with the crossjump unless we found a sufficient number
2026 of matching instructions or the 'from' block was totally matched
2027 (such that its predecessors will hopefully be redirected and the
2028 block removed). */
2029 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
2030 && (newpos1 != BB_HEAD (src1)))
2031 return false;
2033 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
2034 if (block_has_preserve_label (e1->dest)
2035 && (e1->flags & EDGE_ABNORMAL))
2036 return false;
2038 /* Here we know that the insns in the end of SRC1 which are common with SRC2
2039 will be deleted.
2040 If we have tablejumps in the end of SRC1 and SRC2
2041 they have been already compared for equivalence in outgoing_edges_match ()
2042 so replace the references to TABLE1 by references to TABLE2. */
2044 rtx_insn *label1, *label2;
2045 rtx_jump_table_data *table1, *table2;
2047 if (tablejump_p (BB_END (osrc1), &label1, &table1)
2048 && tablejump_p (BB_END (osrc2), &label2, &table2)
2049 && label1 != label2)
2051 rtx_insn *insn;
2053 /* Replace references to LABEL1 with LABEL2. */
2054 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2056 /* Do not replace the label in SRC1->END because when deleting
2057 a block whose end is a tablejump, the tablejump referenced
2058 from the instruction is deleted too. */
2059 if (insn != BB_END (osrc1))
2060 replace_label_in_insn (insn, label1, label2, true);
2065 /* Avoid splitting if possible. We must always split when SRC2 has
2066 EH predecessor edges, or we may end up with basic blocks with both
2067 normal and EH predecessor edges. */
2068 if (newpos2 == BB_HEAD (src2)
2069 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2070 redirect_to = src2;
2071 else
2073 if (newpos2 == BB_HEAD (src2))
2075 /* Skip possible basic block header. */
2076 if (LABEL_P (newpos2))
2077 newpos2 = NEXT_INSN (newpos2);
2078 while (DEBUG_INSN_P (newpos2))
2079 newpos2 = NEXT_INSN (newpos2);
2080 if (NOTE_P (newpos2))
2081 newpos2 = NEXT_INSN (newpos2);
2082 while (DEBUG_INSN_P (newpos2))
2083 newpos2 = NEXT_INSN (newpos2);
2086 if (dump_file)
2087 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2088 src2->index, nmatch);
2089 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2092 if (dump_file)
2093 fprintf (dump_file,
2094 "Cross jumping from bb %i to bb %i; %i common insns\n",
2095 src1->index, src2->index, nmatch);
2097 /* We may have some registers visible through the block. */
2098 df_set_bb_dirty (redirect_to);
2100 if (osrc2 == src2)
2101 redirect_edges_to = redirect_to;
2102 else
2103 redirect_edges_to = osrc2;
2105 /* Recompute the counts of destinations of outgoing edges. */
2106 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2108 edge s2;
2109 edge_iterator ei;
2110 basic_block d = s->dest;
2112 if (FORWARDER_BLOCK_P (d))
2113 d = single_succ (d);
2115 FOR_EACH_EDGE (s2, ei, src1->succs)
2117 basic_block d2 = s2->dest;
2118 if (FORWARDER_BLOCK_P (d2))
2119 d2 = single_succ (d2);
2120 if (d == d2)
2121 break;
2124 /* Take care to update possible forwarder blocks. We verified
2125 that there is no more than one in the chain, so we can't run
2126 into infinite loop. */
2127 if (FORWARDER_BLOCK_P (s->dest))
2128 s->dest->count += s->count ();
2130 if (FORWARDER_BLOCK_P (s2->dest))
2131 s2->dest->count -= s->count ();
2133 /* FIXME: Is this correct? Should be rewritten to count API. */
2134 if (redirect_edges_to->count.nonzero_p () && src1->count.nonzero_p ())
2135 s->probability = s->probability.combine_with_freq
2136 (redirect_edges_to->count.to_frequency (cfun),
2137 s2->probability, src1->count.to_frequency (cfun));
2140 /* Adjust count for the block. An earlier jump
2141 threading pass may have left the profile in an inconsistent
2142 state (see update_bb_profile_for_threading) so we must be
2143 prepared for overflows. */
2144 tmp = redirect_to;
2147 tmp->count += src1->count;
2148 if (tmp == redirect_edges_to)
2149 break;
2150 tmp = find_fallthru_edge (tmp->succs)->dest;
2152 while (true);
2153 update_br_prob_note (redirect_edges_to);
2155 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2157 /* Skip possible basic block header. */
2158 if (LABEL_P (newpos1))
2159 newpos1 = NEXT_INSN (newpos1);
2161 while (DEBUG_INSN_P (newpos1))
2162 newpos1 = NEXT_INSN (newpos1);
2164 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2165 newpos1 = NEXT_INSN (newpos1);
2167 while (DEBUG_INSN_P (newpos1))
2168 newpos1 = NEXT_INSN (newpos1);
2170 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2171 to_remove = single_succ (redirect_from);
2173 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2174 delete_basic_block (to_remove);
2176 update_forwarder_flag (redirect_from);
2177 if (redirect_to != src2)
2178 update_forwarder_flag (src2);
2180 return true;
2183 /* Search the predecessors of BB for common insn sequences. When found,
2184 share code between them by redirecting control flow. Return true if
2185 any changes made. */
2187 static bool
2188 try_crossjump_bb (int mode, basic_block bb)
2190 edge e, e2, fallthru;
2191 bool changed;
2192 unsigned max, ix, ix2;
2194 /* Nothing to do if there is not at least two incoming edges. */
2195 if (EDGE_COUNT (bb->preds) < 2)
2196 return false;
2198 /* Don't crossjump if this block ends in a computed jump,
2199 unless we are optimizing for size. */
2200 if (optimize_bb_for_size_p (bb)
2201 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2202 && computed_jump_p (BB_END (bb)))
2203 return false;
2205 /* If we are partitioning hot/cold basic blocks, we don't want to
2206 mess up unconditional or indirect jumps that cross between hot
2207 and cold sections.
2209 Basic block partitioning may result in some jumps that appear to
2210 be optimizable (or blocks that appear to be mergeable), but which really
2211 must be left untouched (they are required to make it safely across
2212 partition boundaries). See the comments at the top of
2213 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2215 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2216 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2217 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2218 return false;
2220 /* It is always cheapest to redirect a block that ends in a branch to
2221 a block that falls through into BB, as that adds no branches to the
2222 program. We'll try that combination first. */
2223 fallthru = NULL;
2224 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2226 if (EDGE_COUNT (bb->preds) > max)
2227 return false;
2229 fallthru = find_fallthru_edge (bb->preds);
2231 changed = false;
2232 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2234 e = EDGE_PRED (bb, ix);
2235 ix++;
2237 /* As noted above, first try with the fallthru predecessor (or, a
2238 fallthru predecessor if we are in cfglayout mode). */
2239 if (fallthru)
2241 /* Don't combine the fallthru edge into anything else.
2242 If there is a match, we'll do it the other way around. */
2243 if (e == fallthru)
2244 continue;
2245 /* If nothing changed since the last attempt, there is nothing
2246 we can do. */
2247 if (!first_pass
2248 && !((e->src->flags & BB_MODIFIED)
2249 || (fallthru->src->flags & BB_MODIFIED)))
2250 continue;
2252 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2254 changed = true;
2255 ix = 0;
2256 continue;
2260 /* Non-obvious work limiting check: Recognize that we're going
2261 to call try_crossjump_bb on every basic block. So if we have
2262 two blocks with lots of outgoing edges (a switch) and they
2263 share lots of common destinations, then we would do the
2264 cross-jump check once for each common destination.
2266 Now, if the blocks actually are cross-jump candidates, then
2267 all of their destinations will be shared. Which means that
2268 we only need check them for cross-jump candidacy once. We
2269 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2270 choosing to do the check from the block for which the edge
2271 in question is the first successor of A. */
2272 if (EDGE_SUCC (e->src, 0) != e)
2273 continue;
2275 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2277 e2 = EDGE_PRED (bb, ix2);
2279 if (e2 == e)
2280 continue;
2282 /* We've already checked the fallthru edge above. */
2283 if (e2 == fallthru)
2284 continue;
2286 /* The "first successor" check above only prevents multiple
2287 checks of crossjump(A,B). In order to prevent redundant
2288 checks of crossjump(B,A), require that A be the block
2289 with the lowest index. */
2290 if (e->src->index > e2->src->index)
2291 continue;
2293 /* If nothing changed since the last attempt, there is nothing
2294 we can do. */
2295 if (!first_pass
2296 && !((e->src->flags & BB_MODIFIED)
2297 || (e2->src->flags & BB_MODIFIED)))
2298 continue;
2300 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2301 direction. */
2302 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2304 changed = true;
2305 ix = 0;
2306 break;
2311 if (changed)
2312 crossjumps_occurred = true;
2314 return changed;
2317 /* Search the successors of BB for common insn sequences. When found,
2318 share code between them by moving it across the basic block
2319 boundary. Return true if any changes made. */
2321 static bool
2322 try_head_merge_bb (basic_block bb)
2324 basic_block final_dest_bb = NULL;
2325 int max_match = INT_MAX;
2326 edge e0;
2327 rtx_insn **headptr, **currptr, **nextptr;
2328 bool changed, moveall;
2329 unsigned ix;
2330 rtx_insn *e0_last_head;
2331 rtx cond;
2332 rtx_insn *move_before;
2333 unsigned nedges = EDGE_COUNT (bb->succs);
2334 rtx_insn *jump = BB_END (bb);
2335 regset live, live_union;
2337 /* Nothing to do if there is not at least two outgoing edges. */
2338 if (nedges < 2)
2339 return false;
2341 /* Don't crossjump if this block ends in a computed jump,
2342 unless we are optimizing for size. */
2343 if (optimize_bb_for_size_p (bb)
2344 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2345 && computed_jump_p (BB_END (bb)))
2346 return false;
2348 cond = get_condition (jump, &move_before, true, false);
2349 if (cond == NULL_RTX)
2351 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2352 move_before = prev_nonnote_nondebug_insn (jump);
2353 else
2354 move_before = jump;
2357 for (ix = 0; ix < nedges; ix++)
2358 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2359 return false;
2361 for (ix = 0; ix < nedges; ix++)
2363 edge e = EDGE_SUCC (bb, ix);
2364 basic_block other_bb = e->dest;
2366 if (df_get_bb_dirty (other_bb))
2368 block_was_dirty = true;
2369 return false;
2372 if (e->flags & EDGE_ABNORMAL)
2373 return false;
2375 /* Normally, all destination blocks must only be reachable from this
2376 block, i.e. they must have one incoming edge.
2378 There is one special case we can handle, that of multiple consecutive
2379 jumps where the first jumps to one of the targets of the second jump.
2380 This happens frequently in switch statements for default labels.
2381 The structure is as follows:
2382 FINAL_DEST_BB
2383 ....
2384 if (cond) jump A;
2385 fall through
2387 jump with targets A, B, C, D...
2389 has two incoming edges, from FINAL_DEST_BB and BB
2391 In this case, we can try to move the insns through BB and into
2392 FINAL_DEST_BB. */
2393 if (EDGE_COUNT (other_bb->preds) != 1)
2395 edge incoming_edge, incoming_bb_other_edge;
2396 edge_iterator ei;
2398 if (final_dest_bb != NULL
2399 || EDGE_COUNT (other_bb->preds) != 2)
2400 return false;
2402 /* We must be able to move the insns across the whole block. */
2403 move_before = BB_HEAD (bb);
2404 while (!NONDEBUG_INSN_P (move_before))
2405 move_before = NEXT_INSN (move_before);
2407 if (EDGE_COUNT (bb->preds) != 1)
2408 return false;
2409 incoming_edge = EDGE_PRED (bb, 0);
2410 final_dest_bb = incoming_edge->src;
2411 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2412 return false;
2413 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2414 if (incoming_bb_other_edge != incoming_edge)
2415 break;
2416 if (incoming_bb_other_edge->dest != other_bb)
2417 return false;
2421 e0 = EDGE_SUCC (bb, 0);
2422 e0_last_head = NULL;
2423 changed = false;
2425 for (ix = 1; ix < nedges; ix++)
2427 edge e = EDGE_SUCC (bb, ix);
2428 rtx_insn *e0_last, *e_last;
2429 int nmatch;
2431 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2432 &e0_last, &e_last, 0);
2433 if (nmatch == 0)
2434 return false;
2436 if (nmatch < max_match)
2438 max_match = nmatch;
2439 e0_last_head = e0_last;
2443 /* If we matched an entire block, we probably have to avoid moving the
2444 last insn. */
2445 if (max_match > 0
2446 && e0_last_head == BB_END (e0->dest)
2447 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2448 || control_flow_insn_p (e0_last_head)))
2450 max_match--;
2451 if (max_match == 0)
2452 return false;
2454 e0_last_head = prev_real_insn (e0_last_head);
2455 while (DEBUG_INSN_P (e0_last_head));
2458 if (max_match == 0)
2459 return false;
2461 /* We must find a union of the live registers at each of the end points. */
2462 live = BITMAP_ALLOC (NULL);
2463 live_union = BITMAP_ALLOC (NULL);
2465 currptr = XNEWVEC (rtx_insn *, nedges);
2466 headptr = XNEWVEC (rtx_insn *, nedges);
2467 nextptr = XNEWVEC (rtx_insn *, nedges);
2469 for (ix = 0; ix < nedges; ix++)
2471 int j;
2472 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2473 rtx_insn *head = BB_HEAD (merge_bb);
2475 while (!NONDEBUG_INSN_P (head))
2476 head = NEXT_INSN (head);
2477 headptr[ix] = head;
2478 currptr[ix] = head;
2480 /* Compute the end point and live information */
2481 for (j = 1; j < max_match; j++)
2483 head = NEXT_INSN (head);
2484 while (!NONDEBUG_INSN_P (head));
2485 simulate_backwards_to_point (merge_bb, live, head);
2486 IOR_REG_SET (live_union, live);
2489 /* If we're moving across two blocks, verify the validity of the
2490 first move, then adjust the target and let the loop below deal
2491 with the final move. */
2492 if (final_dest_bb != NULL)
2494 rtx_insn *move_upto;
2496 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2497 jump, e0->dest, live_union,
2498 NULL, &move_upto);
2499 if (!moveall)
2501 if (move_upto == NULL_RTX)
2502 goto out;
2504 while (e0_last_head != move_upto)
2506 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2507 live_union);
2508 e0_last_head = PREV_INSN (e0_last_head);
2511 if (e0_last_head == NULL_RTX)
2512 goto out;
2514 jump = BB_END (final_dest_bb);
2515 cond = get_condition (jump, &move_before, true, false);
2516 if (cond == NULL_RTX)
2518 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2519 move_before = prev_nonnote_nondebug_insn (jump);
2520 else
2521 move_before = jump;
2527 rtx_insn *move_upto;
2528 moveall = can_move_insns_across (currptr[0], e0_last_head,
2529 move_before, jump, e0->dest, live_union,
2530 NULL, &move_upto);
2531 if (!moveall && move_upto == NULL_RTX)
2533 if (jump == move_before)
2534 break;
2536 /* Try again, using a different insertion point. */
2537 move_before = jump;
2539 /* Don't try moving before a cc0 user, as that may invalidate
2540 the cc0. */
2541 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2542 break;
2544 continue;
2547 if (final_dest_bb && !moveall)
2548 /* We haven't checked whether a partial move would be OK for the first
2549 move, so we have to fail this case. */
2550 break;
2552 changed = true;
2553 for (;;)
2555 if (currptr[0] == move_upto)
2556 break;
2557 for (ix = 0; ix < nedges; ix++)
2559 rtx_insn *curr = currptr[ix];
2561 curr = NEXT_INSN (curr);
2562 while (!NONDEBUG_INSN_P (curr));
2563 currptr[ix] = curr;
2567 /* If we can't currently move all of the identical insns, remember
2568 each insn after the range that we'll merge. */
2569 if (!moveall)
2570 for (ix = 0; ix < nedges; ix++)
2572 rtx_insn *curr = currptr[ix];
2574 curr = NEXT_INSN (curr);
2575 while (!NONDEBUG_INSN_P (curr));
2576 nextptr[ix] = curr;
2579 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2580 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2581 if (final_dest_bb != NULL)
2582 df_set_bb_dirty (final_dest_bb);
2583 df_set_bb_dirty (bb);
2584 for (ix = 1; ix < nedges; ix++)
2586 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2587 delete_insn_chain (headptr[ix], currptr[ix], false);
2589 if (!moveall)
2591 if (jump == move_before)
2592 break;
2594 /* For the unmerged insns, try a different insertion point. */
2595 move_before = jump;
2597 /* Don't try moving before a cc0 user, as that may invalidate
2598 the cc0. */
2599 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2600 break;
2602 for (ix = 0; ix < nedges; ix++)
2603 currptr[ix] = headptr[ix] = nextptr[ix];
2606 while (!moveall);
2608 out:
2609 free (currptr);
2610 free (headptr);
2611 free (nextptr);
2613 crossjumps_occurred |= changed;
2615 return changed;
2618 /* Return true if BB contains just bb note, or bb note followed
2619 by only DEBUG_INSNs. */
2621 static bool
2622 trivially_empty_bb_p (basic_block bb)
2624 rtx_insn *insn = BB_END (bb);
2626 while (1)
2628 if (insn == BB_HEAD (bb))
2629 return true;
2630 if (!DEBUG_INSN_P (insn))
2631 return false;
2632 insn = PREV_INSN (insn);
2636 /* Return true if BB contains just a return and possibly a USE of the
2637 return value. Fill in *RET and *USE with the return and use insns
2638 if any found, otherwise NULL. All CLOBBERs are ignored. */
2640 static bool
2641 bb_is_just_return (basic_block bb, rtx_insn **ret, rtx_insn **use)
2643 *ret = *use = NULL;
2644 rtx_insn *insn;
2646 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
2647 return false;
2649 FOR_BB_INSNS (bb, insn)
2650 if (NONDEBUG_INSN_P (insn))
2652 rtx pat = PATTERN (insn);
2654 if (!*ret && ANY_RETURN_P (pat))
2655 *ret = insn;
2656 else if (!*ret && !*use && GET_CODE (pat) == USE
2657 && REG_P (XEXP (pat, 0))
2658 && REG_FUNCTION_VALUE_P (XEXP (pat, 0)))
2659 *use = insn;
2660 else if (GET_CODE (pat) != CLOBBER)
2661 return false;
2664 return !!*ret;
2667 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2668 instructions etc. Return nonzero if changes were made. */
2670 static bool
2671 try_optimize_cfg (int mode)
2673 bool changed_overall = false;
2674 bool changed;
2675 int iterations = 0;
2676 basic_block bb, b, next;
2678 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2679 clear_bb_flags ();
2681 crossjumps_occurred = false;
2683 FOR_EACH_BB_FN (bb, cfun)
2684 update_forwarder_flag (bb);
2686 if (! targetm.cannot_modify_jumps_p ())
2688 first_pass = true;
2689 /* Attempt to merge blocks as made possible by edge removal. If
2690 a block has only one successor, and the successor has only
2691 one predecessor, they may be combined. */
2694 block_was_dirty = false;
2695 changed = false;
2696 iterations++;
2698 if (dump_file)
2699 fprintf (dump_file,
2700 "\n\ntry_optimize_cfg iteration %i\n\n",
2701 iterations);
2703 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2704 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2706 basic_block c;
2707 edge s;
2708 bool changed_here = false;
2710 /* Delete trivially dead basic blocks. This is either
2711 blocks with no predecessors, or empty blocks with no
2712 successors. However if the empty block with no
2713 successors is the successor of the ENTRY_BLOCK, it is
2714 kept. This ensures that the ENTRY_BLOCK will have a
2715 successor which is a precondition for many RTL
2716 passes. Empty blocks may result from expanding
2717 __builtin_unreachable (). */
2718 if (EDGE_COUNT (b->preds) == 0
2719 || (EDGE_COUNT (b->succs) == 0
2720 && trivially_empty_bb_p (b)
2721 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2722 != b))
2724 c = b->prev_bb;
2725 if (EDGE_COUNT (b->preds) > 0)
2727 edge e;
2728 edge_iterator ei;
2730 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2732 if (BB_FOOTER (b)
2733 && BARRIER_P (BB_FOOTER (b)))
2734 FOR_EACH_EDGE (e, ei, b->preds)
2735 if ((e->flags & EDGE_FALLTHRU)
2736 && BB_FOOTER (e->src) == NULL)
2738 if (BB_FOOTER (b))
2740 BB_FOOTER (e->src) = BB_FOOTER (b);
2741 BB_FOOTER (b) = NULL;
2743 else
2745 start_sequence ();
2746 BB_FOOTER (e->src) = emit_barrier ();
2747 end_sequence ();
2751 else
2753 rtx_insn *last = get_last_bb_insn (b);
2754 if (last && BARRIER_P (last))
2755 FOR_EACH_EDGE (e, ei, b->preds)
2756 if ((e->flags & EDGE_FALLTHRU))
2757 emit_barrier_after (BB_END (e->src));
2760 delete_basic_block (b);
2761 changed = true;
2762 /* Avoid trying to remove the exit block. */
2763 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2764 continue;
2767 /* Remove code labels no longer used. */
2768 if (single_pred_p (b)
2769 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2770 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2771 && LABEL_P (BB_HEAD (b))
2772 && !LABEL_PRESERVE_P (BB_HEAD (b))
2773 /* If the previous block ends with a branch to this
2774 block, we can't delete the label. Normally this
2775 is a condjump that is yet to be simplified, but
2776 if CASE_DROPS_THRU, this can be a tablejump with
2777 some element going to the same place as the
2778 default (fallthru). */
2779 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2780 || !JUMP_P (BB_END (single_pred (b)))
2781 || ! label_is_jump_target_p (BB_HEAD (b),
2782 BB_END (single_pred (b)))))
2784 delete_insn (BB_HEAD (b));
2785 if (dump_file)
2786 fprintf (dump_file, "Deleted label in block %i.\n",
2787 b->index);
2790 /* If we fall through an empty block, we can remove it. */
2791 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2792 && single_pred_p (b)
2793 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2794 && !LABEL_P (BB_HEAD (b))
2795 && FORWARDER_BLOCK_P (b)
2796 /* Note that forwarder_block_p true ensures that
2797 there is a successor for this block. */
2798 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2799 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2801 if (dump_file)
2802 fprintf (dump_file,
2803 "Deleting fallthru block %i.\n",
2804 b->index);
2806 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2807 ? b->next_bb : b->prev_bb);
2808 redirect_edge_succ_nodup (single_pred_edge (b),
2809 single_succ (b));
2810 delete_basic_block (b);
2811 changed = true;
2812 b = c;
2813 continue;
2816 /* Merge B with its single successor, if any. */
2817 if (single_succ_p (b)
2818 && (s = single_succ_edge (b))
2819 && !(s->flags & EDGE_COMPLEX)
2820 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2821 && single_pred_p (c)
2822 && b != c)
2824 /* When not in cfg_layout mode use code aware of reordering
2825 INSN. This code possibly creates new basic blocks so it
2826 does not fit merge_blocks interface and is kept here in
2827 hope that it will become useless once more of compiler
2828 is transformed to use cfg_layout mode. */
2830 if ((mode & CLEANUP_CFGLAYOUT)
2831 && can_merge_blocks_p (b, c))
2833 merge_blocks (b, c);
2834 update_forwarder_flag (b);
2835 changed_here = true;
2837 else if (!(mode & CLEANUP_CFGLAYOUT)
2838 /* If the jump insn has side effects,
2839 we can't kill the edge. */
2840 && (!JUMP_P (BB_END (b))
2841 || (reload_completed
2842 ? simplejump_p (BB_END (b))
2843 : (onlyjump_p (BB_END (b))
2844 && !tablejump_p (BB_END (b),
2845 NULL, NULL))))
2846 && (next = merge_blocks_move (s, b, c, mode)))
2848 b = next;
2849 changed_here = true;
2853 /* Try to change a branch to a return to just that return. */
2854 rtx_insn *ret, *use;
2855 if (single_succ_p (b)
2856 && onlyjump_p (BB_END (b))
2857 && bb_is_just_return (single_succ (b), &ret, &use))
2859 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2860 PATTERN (ret), 0))
2862 if (use)
2863 emit_insn_before (copy_insn (PATTERN (use)),
2864 BB_END (b));
2865 if (dump_file)
2866 fprintf (dump_file, "Changed jump %d->%d to return.\n",
2867 b->index, single_succ (b)->index);
2868 redirect_edge_succ (single_succ_edge (b),
2869 EXIT_BLOCK_PTR_FOR_FN (cfun));
2870 single_succ_edge (b)->flags &= ~EDGE_CROSSING;
2871 changed_here = true;
2875 /* Try to change a conditional branch to a return to the
2876 respective conditional return. */
2877 if (EDGE_COUNT (b->succs) == 2
2878 && any_condjump_p (BB_END (b))
2879 && bb_is_just_return (BRANCH_EDGE (b)->dest, &ret, &use))
2881 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2882 PATTERN (ret), 0))
2884 if (use)
2885 emit_insn_before (copy_insn (PATTERN (use)),
2886 BB_END (b));
2887 if (dump_file)
2888 fprintf (dump_file, "Changed conditional jump %d->%d "
2889 "to conditional return.\n",
2890 b->index, BRANCH_EDGE (b)->dest->index);
2891 redirect_edge_succ (BRANCH_EDGE (b),
2892 EXIT_BLOCK_PTR_FOR_FN (cfun));
2893 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2894 changed_here = true;
2898 /* Try to flip a conditional branch that falls through to
2899 a return so that it becomes a conditional return and a
2900 new jump to the original branch target. */
2901 if (EDGE_COUNT (b->succs) == 2
2902 && BRANCH_EDGE (b)->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
2903 && any_condjump_p (BB_END (b))
2904 && bb_is_just_return (FALLTHRU_EDGE (b)->dest, &ret, &use))
2906 if (invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2907 JUMP_LABEL (BB_END (b)), 0))
2909 basic_block new_ft = BRANCH_EDGE (b)->dest;
2910 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2911 PATTERN (ret), 0))
2913 if (use)
2914 emit_insn_before (copy_insn (PATTERN (use)),
2915 BB_END (b));
2916 if (dump_file)
2917 fprintf (dump_file, "Changed conditional jump "
2918 "%d->%d to conditional return, adding "
2919 "fall-through jump.\n",
2920 b->index, BRANCH_EDGE (b)->dest->index);
2921 redirect_edge_succ (BRANCH_EDGE (b),
2922 EXIT_BLOCK_PTR_FOR_FN (cfun));
2923 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2924 std::swap (BRANCH_EDGE (b)->probability,
2925 FALLTHRU_EDGE (b)->probability);
2926 update_br_prob_note (b);
2927 basic_block jb = force_nonfallthru (FALLTHRU_EDGE (b));
2928 notice_new_block (jb);
2929 if (!redirect_jump (as_a <rtx_jump_insn *> (BB_END (jb)),
2930 block_label (new_ft), 0))
2931 gcc_unreachable ();
2932 redirect_edge_succ (single_succ_edge (jb), new_ft);
2933 changed_here = true;
2935 else
2937 /* Invert the jump back to what it was. This should
2938 never fail. */
2939 if (!invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2940 JUMP_LABEL (BB_END (b)), 0))
2941 gcc_unreachable ();
2946 /* Simplify branch over branch. */
2947 if ((mode & CLEANUP_EXPENSIVE)
2948 && !(mode & CLEANUP_CFGLAYOUT)
2949 && try_simplify_condjump (b))
2950 changed_here = true;
2952 /* If B has a single outgoing edge, but uses a
2953 non-trivial jump instruction without side-effects, we
2954 can either delete the jump entirely, or replace it
2955 with a simple unconditional jump. */
2956 if (single_succ_p (b)
2957 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2958 && onlyjump_p (BB_END (b))
2959 && !CROSSING_JUMP_P (BB_END (b))
2960 && try_redirect_by_replacing_jump (single_succ_edge (b),
2961 single_succ (b),
2962 (mode & CLEANUP_CFGLAYOUT) != 0))
2964 update_forwarder_flag (b);
2965 changed_here = true;
2968 /* Simplify branch to branch. */
2969 if (try_forward_edges (mode, b))
2971 update_forwarder_flag (b);
2972 changed_here = true;
2975 /* Look for shared code between blocks. */
2976 if ((mode & CLEANUP_CROSSJUMP)
2977 && try_crossjump_bb (mode, b))
2978 changed_here = true;
2980 if ((mode & CLEANUP_CROSSJUMP)
2981 /* This can lengthen register lifetimes. Do it only after
2982 reload. */
2983 && reload_completed
2984 && try_head_merge_bb (b))
2985 changed_here = true;
2987 /* Don't get confused by the index shift caused by
2988 deleting blocks. */
2989 if (!changed_here)
2990 b = b->next_bb;
2991 else
2992 changed = true;
2995 if ((mode & CLEANUP_CROSSJUMP)
2996 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
2997 changed = true;
2999 if (block_was_dirty)
3001 /* This should only be set by head-merging. */
3002 gcc_assert (mode & CLEANUP_CROSSJUMP);
3003 df_analyze ();
3006 if (changed)
3008 /* Edge forwarding in particular can cause hot blocks previously
3009 reached by both hot and cold blocks to become dominated only
3010 by cold blocks. This will cause the verification below to fail,
3011 and lead to now cold code in the hot section. This is not easy
3012 to detect and fix during edge forwarding, and in some cases
3013 is only visible after newly unreachable blocks are deleted,
3014 which will be done in fixup_partitions. */
3015 if ((mode & CLEANUP_NO_PARTITIONING) == 0)
3017 fixup_partitions ();
3018 checking_verify_flow_info ();
3022 changed_overall |= changed;
3023 first_pass = false;
3025 while (changed);
3028 FOR_ALL_BB_FN (b, cfun)
3029 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
3031 return changed_overall;
3034 /* Delete all unreachable basic blocks. */
3036 bool
3037 delete_unreachable_blocks (void)
3039 bool changed = false;
3040 basic_block b, prev_bb;
3042 find_unreachable_blocks ();
3044 /* When we're in GIMPLE mode and there may be debug bind insns, we
3045 should delete blocks in reverse dominator order, so as to get a
3046 chance to substitute all released DEFs into debug bind stmts. If
3047 we don't have dominators information, walking blocks backward
3048 gets us a better chance of retaining most debug information than
3049 otherwise. */
3050 if (MAY_HAVE_DEBUG_BIND_INSNS && current_ir_type () == IR_GIMPLE
3051 && dom_info_available_p (CDI_DOMINATORS))
3053 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3054 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3056 prev_bb = b->prev_bb;
3058 if (!(b->flags & BB_REACHABLE))
3060 /* Speed up the removal of blocks that don't dominate
3061 others. Walking backwards, this should be the common
3062 case. */
3063 if (!first_dom_son (CDI_DOMINATORS, b))
3064 delete_basic_block (b);
3065 else
3067 vec<basic_block> h
3068 = get_all_dominated_blocks (CDI_DOMINATORS, b);
3070 while (h.length ())
3072 b = h.pop ();
3074 prev_bb = b->prev_bb;
3076 gcc_assert (!(b->flags & BB_REACHABLE));
3078 delete_basic_block (b);
3081 h.release ();
3084 changed = true;
3088 else
3090 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3091 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3093 prev_bb = b->prev_bb;
3095 if (!(b->flags & BB_REACHABLE))
3097 delete_basic_block (b);
3098 changed = true;
3103 if (changed)
3104 tidy_fallthru_edges ();
3105 return changed;
3108 /* Delete any jump tables never referenced. We can't delete them at the
3109 time of removing tablejump insn as they are referenced by the preceding
3110 insns computing the destination, so we delay deleting and garbagecollect
3111 them once life information is computed. */
3112 void
3113 delete_dead_jumptables (void)
3115 basic_block bb;
3117 /* A dead jump table does not belong to any basic block. Scan insns
3118 between two adjacent basic blocks. */
3119 FOR_EACH_BB_FN (bb, cfun)
3121 rtx_insn *insn, *next;
3123 for (insn = NEXT_INSN (BB_END (bb));
3124 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
3125 insn = next)
3127 next = NEXT_INSN (insn);
3128 if (LABEL_P (insn)
3129 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
3130 && JUMP_TABLE_DATA_P (next))
3132 rtx_insn *label = insn, *jump = next;
3134 if (dump_file)
3135 fprintf (dump_file, "Dead jumptable %i removed\n",
3136 INSN_UID (insn));
3138 next = NEXT_INSN (next);
3139 delete_insn (jump);
3140 delete_insn (label);
3147 /* Tidy the CFG by deleting unreachable code and whatnot. */
3149 bool
3150 cleanup_cfg (int mode)
3152 bool changed = false;
3154 /* Set the cfglayout mode flag here. We could update all the callers
3155 but that is just inconvenient, especially given that we eventually
3156 want to have cfglayout mode as the default. */
3157 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3158 mode |= CLEANUP_CFGLAYOUT;
3160 timevar_push (TV_CLEANUP_CFG);
3161 if (delete_unreachable_blocks ())
3163 changed = true;
3164 /* We've possibly created trivially dead code. Cleanup it right
3165 now to introduce more opportunities for try_optimize_cfg. */
3166 if (!(mode & (CLEANUP_NO_INSN_DEL))
3167 && !reload_completed)
3168 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3171 compact_blocks ();
3173 /* To tail-merge blocks ending in the same noreturn function (e.g.
3174 a call to abort) we have to insert fake edges to exit. Do this
3175 here once. The fake edges do not interfere with any other CFG
3176 cleanups. */
3177 if (mode & CLEANUP_CROSSJUMP)
3178 add_noreturn_fake_exit_edges ();
3180 if (!dbg_cnt (cfg_cleanup))
3181 return changed;
3183 while (try_optimize_cfg (mode))
3185 delete_unreachable_blocks (), changed = true;
3186 if (!(mode & CLEANUP_NO_INSN_DEL))
3188 /* Try to remove some trivially dead insns when doing an expensive
3189 cleanup. But delete_trivially_dead_insns doesn't work after
3190 reload (it only handles pseudos) and run_fast_dce is too costly
3191 to run in every iteration.
3193 For effective cross jumping, we really want to run a fast DCE to
3194 clean up any dead conditions, or they get in the way of performing
3195 useful tail merges.
3197 Other transformations in cleanup_cfg are not so sensitive to dead
3198 code, so delete_trivially_dead_insns or even doing nothing at all
3199 is good enough. */
3200 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3201 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3202 break;
3203 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occurred)
3204 run_fast_dce ();
3206 else
3207 break;
3210 if (mode & CLEANUP_CROSSJUMP)
3211 remove_fake_exit_edges ();
3213 /* Don't call delete_dead_jumptables in cfglayout mode, because
3214 that function assumes that jump tables are in the insns stream.
3215 But we also don't _have_ to delete dead jumptables in cfglayout
3216 mode because we shouldn't even be looking at things that are
3217 not in a basic block. Dead jumptables are cleaned up when
3218 going out of cfglayout mode. */
3219 if (!(mode & CLEANUP_CFGLAYOUT))
3220 delete_dead_jumptables ();
3222 /* ??? We probably do this way too often. */
3223 if (current_loops
3224 && (changed
3225 || (mode & CLEANUP_CFG_CHANGED)))
3227 timevar_push (TV_REPAIR_LOOPS);
3228 /* The above doesn't preserve dominance info if available. */
3229 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3230 calculate_dominance_info (CDI_DOMINATORS);
3231 fix_loop_structure (NULL);
3232 free_dominance_info (CDI_DOMINATORS);
3233 timevar_pop (TV_REPAIR_LOOPS);
3236 timevar_pop (TV_CLEANUP_CFG);
3238 return changed;
3241 namespace {
3243 const pass_data pass_data_jump =
3245 RTL_PASS, /* type */
3246 "jump", /* name */
3247 OPTGROUP_NONE, /* optinfo_flags */
3248 TV_JUMP, /* tv_id */
3249 0, /* properties_required */
3250 0, /* properties_provided */
3251 0, /* properties_destroyed */
3252 0, /* todo_flags_start */
3253 0, /* todo_flags_finish */
3256 class pass_jump : public rtl_opt_pass
3258 public:
3259 pass_jump (gcc::context *ctxt)
3260 : rtl_opt_pass (pass_data_jump, ctxt)
3263 /* opt_pass methods: */
3264 virtual unsigned int execute (function *);
3266 }; // class pass_jump
3268 unsigned int
3269 pass_jump::execute (function *)
3271 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3272 if (dump_file)
3273 dump_flow_info (dump_file, dump_flags);
3274 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3275 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3276 return 0;
3279 } // anon namespace
3281 rtl_opt_pass *
3282 make_pass_jump (gcc::context *ctxt)
3284 return new pass_jump (ctxt);
3287 namespace {
3289 const pass_data pass_data_jump2 =
3291 RTL_PASS, /* type */
3292 "jump2", /* name */
3293 OPTGROUP_NONE, /* optinfo_flags */
3294 TV_JUMP, /* tv_id */
3295 0, /* properties_required */
3296 0, /* properties_provided */
3297 0, /* properties_destroyed */
3298 0, /* todo_flags_start */
3299 0, /* todo_flags_finish */
3302 class pass_jump2 : public rtl_opt_pass
3304 public:
3305 pass_jump2 (gcc::context *ctxt)
3306 : rtl_opt_pass (pass_data_jump2, ctxt)
3309 /* opt_pass methods: */
3310 virtual unsigned int execute (function *)
3312 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3313 return 0;
3316 }; // class pass_jump2
3318 } // anon namespace
3320 rtl_opt_pass *
3321 make_pass_jump2 (gcc::context *ctxt)
3323 return new pass_jump2 (ctxt);