Improve max_insns_skipped logic
[official-gcc.git] / gcc / cfgcleanup.c
blob365c971effb3da1e05525a85e8b12954531933cc
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2017 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 profile_probability edge_probability = e->probability;
563 int edge_frequency;
564 int n = 0;
566 e->goto_locus = goto_locus;
568 /* Don't force if target is exit block. */
569 if (threaded && target != EXIT_BLOCK_PTR_FOR_FN (cfun))
571 notice_new_block (redirect_edge_and_branch_force (e, target));
572 if (dump_file)
573 fprintf (dump_file, "Conditionals threaded.\n");
575 else if (!redirect_edge_and_branch (e, target))
577 if (dump_file)
578 fprintf (dump_file,
579 "Forwarding edge %i->%i to %i failed.\n",
580 b->index, e->dest->index, target->index);
581 ei_next (&ei);
582 continue;
585 /* We successfully forwarded the edge. Now update profile
586 data: for each edge we traversed in the chain, remove
587 the original edge's execution count. */
588 edge_frequency = edge_probability.apply (b->frequency);
592 edge t;
594 if (!single_succ_p (first))
596 gcc_assert (n < nthreaded_edges);
597 t = threaded_edges [n++];
598 gcc_assert (t->src == first);
599 update_bb_profile_for_threading (first, edge_frequency,
600 edge_count, t);
601 update_br_prob_note (first);
603 else
605 first->count -= edge_count;
606 first->frequency -= edge_frequency;
607 if (first->frequency < 0)
608 first->frequency = 0;
609 /* It is possible that as the result of
610 threading we've removed edge as it is
611 threaded to the fallthru edge. Avoid
612 getting out of sync. */
613 if (n < nthreaded_edges
614 && first == threaded_edges [n]->src)
615 n++;
616 t = single_succ_edge (first);
619 t->count -= edge_count;
620 first = t->dest;
622 while (first != target);
624 changed = true;
625 continue;
627 ei_next (&ei);
630 free (threaded_edges);
631 return changed;
635 /* Blocks A and B are to be merged into a single block. A has no incoming
636 fallthru edge, so it can be moved before B without adding or modifying
637 any jumps (aside from the jump from A to B). */
639 static void
640 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
642 rtx_insn *barrier;
644 /* If we are partitioning hot/cold basic blocks, we don't want to
645 mess up unconditional or indirect jumps that cross between hot
646 and cold sections.
648 Basic block partitioning may result in some jumps that appear to
649 be optimizable (or blocks that appear to be mergeable), but which really
650 must be left untouched (they are required to make it safely across
651 partition boundaries). See the comments at the top of
652 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
654 if (BB_PARTITION (a) != BB_PARTITION (b))
655 return;
657 barrier = next_nonnote_insn (BB_END (a));
658 gcc_assert (BARRIER_P (barrier));
659 delete_insn (barrier);
661 /* Scramble the insn chain. */
662 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
663 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
664 df_set_bb_dirty (a);
666 if (dump_file)
667 fprintf (dump_file, "Moved block %d before %d and merged.\n",
668 a->index, b->index);
670 /* Swap the records for the two blocks around. */
672 unlink_block (a);
673 link_block (a, b->prev_bb);
675 /* Now blocks A and B are contiguous. Merge them. */
676 merge_blocks (a, b);
679 /* Blocks A and B are to be merged into a single block. B has no outgoing
680 fallthru edge, so it can be moved after A without adding or modifying
681 any jumps (aside from the jump from A to B). */
683 static void
684 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
686 rtx_insn *barrier, *real_b_end;
687 rtx_insn *label;
688 rtx_jump_table_data *table;
690 /* If we are partitioning hot/cold basic blocks, we don't want to
691 mess up unconditional or indirect jumps that cross between hot
692 and cold sections.
694 Basic block partitioning may result in some jumps that appear to
695 be optimizable (or blocks that appear to be mergeable), but which really
696 must be left untouched (they are required to make it safely across
697 partition boundaries). See the comments at the top of
698 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
700 if (BB_PARTITION (a) != BB_PARTITION (b))
701 return;
703 real_b_end = BB_END (b);
705 /* If there is a jump table following block B temporarily add the jump table
706 to block B so that it will also be moved to the correct location. */
707 if (tablejump_p (BB_END (b), &label, &table)
708 && prev_active_insn (label) == BB_END (b))
710 BB_END (b) = table;
713 /* There had better have been a barrier there. Delete it. */
714 barrier = NEXT_INSN (BB_END (b));
715 if (barrier && BARRIER_P (barrier))
716 delete_insn (barrier);
719 /* Scramble the insn chain. */
720 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
722 /* Restore the real end of b. */
723 BB_END (b) = real_b_end;
725 if (dump_file)
726 fprintf (dump_file, "Moved block %d after %d and merged.\n",
727 b->index, a->index);
729 /* Now blocks A and B are contiguous. Merge them. */
730 merge_blocks (a, b);
733 /* Attempt to merge basic blocks that are potentially non-adjacent.
734 Return NULL iff the attempt failed, otherwise return basic block
735 where cleanup_cfg should continue. Because the merging commonly
736 moves basic block away or introduces another optimization
737 possibility, return basic block just before B so cleanup_cfg don't
738 need to iterate.
740 It may be good idea to return basic block before C in the case
741 C has been moved after B and originally appeared earlier in the
742 insn sequence, but we have no information available about the
743 relative ordering of these two. Hopefully it is not too common. */
745 static basic_block
746 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
748 basic_block next;
750 /* If we are partitioning hot/cold basic blocks, we don't want to
751 mess up unconditional or indirect jumps that cross between hot
752 and cold sections.
754 Basic block partitioning may result in some jumps that appear to
755 be optimizable (or blocks that appear to be mergeable), but which really
756 must be left untouched (they are required to make it safely across
757 partition boundaries). See the comments at the top of
758 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
760 if (BB_PARTITION (b) != BB_PARTITION (c))
761 return NULL;
763 /* If B has a fallthru edge to C, no need to move anything. */
764 if (e->flags & EDGE_FALLTHRU)
766 int b_index = b->index, c_index = c->index;
768 /* Protect the loop latches. */
769 if (current_loops && c->loop_father->latch == c)
770 return NULL;
772 merge_blocks (b, c);
773 update_forwarder_flag (b);
775 if (dump_file)
776 fprintf (dump_file, "Merged %d and %d without moving.\n",
777 b_index, c_index);
779 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
782 /* Otherwise we will need to move code around. Do that only if expensive
783 transformations are allowed. */
784 else if (mode & CLEANUP_EXPENSIVE)
786 edge tmp_edge, b_fallthru_edge;
787 bool c_has_outgoing_fallthru;
788 bool b_has_incoming_fallthru;
790 /* Avoid overactive code motion, as the forwarder blocks should be
791 eliminated by edge redirection instead. One exception might have
792 been if B is a forwarder block and C has no fallthru edge, but
793 that should be cleaned up by bb-reorder instead. */
794 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
795 return NULL;
797 /* We must make sure to not munge nesting of lexical blocks,
798 and loop notes. This is done by squeezing out all the notes
799 and leaving them there to lie. Not ideal, but functional. */
801 tmp_edge = find_fallthru_edge (c->succs);
802 c_has_outgoing_fallthru = (tmp_edge != NULL);
804 tmp_edge = find_fallthru_edge (b->preds);
805 b_has_incoming_fallthru = (tmp_edge != NULL);
806 b_fallthru_edge = tmp_edge;
807 next = b->prev_bb;
808 if (next == c)
809 next = next->prev_bb;
811 /* Otherwise, we're going to try to move C after B. If C does
812 not have an outgoing fallthru, then it can be moved
813 immediately after B without introducing or modifying jumps. */
814 if (! c_has_outgoing_fallthru)
816 merge_blocks_move_successor_nojumps (b, c);
817 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
820 /* If B does not have an incoming fallthru, then it can be moved
821 immediately before C without introducing or modifying jumps.
822 C cannot be the first block, so we do not have to worry about
823 accessing a non-existent block. */
825 if (b_has_incoming_fallthru)
827 basic_block bb;
829 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
830 return NULL;
831 bb = force_nonfallthru (b_fallthru_edge);
832 if (bb)
833 notice_new_block (bb);
836 merge_blocks_move_predecessor_nojumps (b, c);
837 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
840 return NULL;
844 /* Removes the memory attributes of MEM expression
845 if they are not equal. */
847 static void
848 merge_memattrs (rtx x, rtx y)
850 int i;
851 int j;
852 enum rtx_code code;
853 const char *fmt;
855 if (x == y)
856 return;
857 if (x == 0 || y == 0)
858 return;
860 code = GET_CODE (x);
862 if (code != GET_CODE (y))
863 return;
865 if (GET_MODE (x) != GET_MODE (y))
866 return;
868 if (code == MEM && !mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
870 if (! MEM_ATTRS (x))
871 MEM_ATTRS (y) = 0;
872 else if (! MEM_ATTRS (y))
873 MEM_ATTRS (x) = 0;
874 else
876 HOST_WIDE_INT mem_size;
878 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
880 set_mem_alias_set (x, 0);
881 set_mem_alias_set (y, 0);
884 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
886 set_mem_expr (x, 0);
887 set_mem_expr (y, 0);
888 clear_mem_offset (x);
889 clear_mem_offset (y);
891 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
892 || (MEM_OFFSET_KNOWN_P (x)
893 && MEM_OFFSET (x) != MEM_OFFSET (y)))
895 clear_mem_offset (x);
896 clear_mem_offset (y);
899 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
901 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
902 set_mem_size (x, mem_size);
903 set_mem_size (y, mem_size);
905 else
907 clear_mem_size (x);
908 clear_mem_size (y);
911 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
912 set_mem_align (y, MEM_ALIGN (x));
915 if (code == MEM)
917 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
919 MEM_READONLY_P (x) = 0;
920 MEM_READONLY_P (y) = 0;
922 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
924 MEM_NOTRAP_P (x) = 0;
925 MEM_NOTRAP_P (y) = 0;
927 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
929 MEM_VOLATILE_P (x) = 1;
930 MEM_VOLATILE_P (y) = 1;
934 fmt = GET_RTX_FORMAT (code);
935 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
937 switch (fmt[i])
939 case 'E':
940 /* Two vectors must have the same length. */
941 if (XVECLEN (x, i) != XVECLEN (y, i))
942 return;
944 for (j = 0; j < XVECLEN (x, i); j++)
945 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
947 break;
949 case 'e':
950 merge_memattrs (XEXP (x, i), XEXP (y, i));
953 return;
957 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
958 different single sets S1 and S2. */
960 static bool
961 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
963 int i;
964 rtx e1, e2;
966 if (p1 == s1 && p2 == s2)
967 return true;
969 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
970 return false;
972 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
973 return false;
975 for (i = 0; i < XVECLEN (p1, 0); i++)
977 e1 = XVECEXP (p1, 0, i);
978 e2 = XVECEXP (p2, 0, i);
979 if (e1 == s1 && e2 == s2)
980 continue;
981 if (reload_completed
982 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
983 continue;
985 return false;
988 return true;
992 /* NOTE1 is the REG_EQUAL note, if any, attached to an insn
993 that is a single_set with a SET_SRC of SRC1. Similarly
994 for NOTE2/SRC2.
996 So effectively NOTE1/NOTE2 are an alternate form of
997 SRC1/SRC2 respectively.
999 Return nonzero if SRC1 or NOTE1 has the same constant
1000 integer value as SRC2 or NOTE2. Else return zero. */
1001 static int
1002 values_equal_p (rtx note1, rtx note2, rtx src1, rtx src2)
1004 if (note1
1005 && note2
1006 && CONST_INT_P (XEXP (note1, 0))
1007 && rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0)))
1008 return 1;
1010 if (!note1
1011 && !note2
1012 && CONST_INT_P (src1)
1013 && CONST_INT_P (src2)
1014 && rtx_equal_p (src1, src2))
1015 return 1;
1017 if (note1
1018 && CONST_INT_P (src2)
1019 && rtx_equal_p (XEXP (note1, 0), src2))
1020 return 1;
1022 if (note2
1023 && CONST_INT_P (src1)
1024 && rtx_equal_p (XEXP (note2, 0), src1))
1025 return 1;
1027 return 0;
1030 /* Examine register notes on I1 and I2 and return:
1031 - dir_forward if I1 can be replaced by I2, or
1032 - dir_backward if I2 can be replaced by I1, or
1033 - dir_both if both are the case. */
1035 static enum replace_direction
1036 can_replace_by (rtx_insn *i1, rtx_insn *i2)
1038 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1039 bool c1, c2;
1041 /* Check for 2 sets. */
1042 s1 = single_set (i1);
1043 s2 = single_set (i2);
1044 if (s1 == NULL_RTX || s2 == NULL_RTX)
1045 return dir_none;
1047 /* Check that the 2 sets set the same dest. */
1048 d1 = SET_DEST (s1);
1049 d2 = SET_DEST (s2);
1050 if (!(reload_completed
1051 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1052 return dir_none;
1054 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1055 set dest to the same value. */
1056 note1 = find_reg_equal_equiv_note (i1);
1057 note2 = find_reg_equal_equiv_note (i2);
1059 src1 = SET_SRC (s1);
1060 src2 = SET_SRC (s2);
1062 if (!values_equal_p (note1, note2, src1, src2))
1063 return dir_none;
1065 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1066 return dir_none;
1068 /* Although the 2 sets set dest to the same value, we cannot replace
1069 (set (dest) (const_int))
1071 (set (dest) (reg))
1072 because we don't know if the reg is live and has the same value at the
1073 location of replacement. */
1074 c1 = CONST_INT_P (src1);
1075 c2 = CONST_INT_P (src2);
1076 if (c1 && c2)
1077 return dir_both;
1078 else if (c2)
1079 return dir_forward;
1080 else if (c1)
1081 return dir_backward;
1083 return dir_none;
1086 /* Merges directions A and B. */
1088 static enum replace_direction
1089 merge_dir (enum replace_direction a, enum replace_direction b)
1091 /* Implements the following table:
1092 |bo fw bw no
1093 ---+-----------
1094 bo |bo fw bw no
1095 fw |-- fw no no
1096 bw |-- -- bw no
1097 no |-- -- -- no. */
1099 if (a == b)
1100 return a;
1102 if (a == dir_both)
1103 return b;
1104 if (b == dir_both)
1105 return a;
1107 return dir_none;
1110 /* Array of flags indexed by reg note kind, true if the given
1111 reg note is CFA related. */
1112 static const bool reg_note_cfa_p[] = {
1113 #undef REG_CFA_NOTE
1114 #define DEF_REG_NOTE(NAME) false,
1115 #define REG_CFA_NOTE(NAME) true,
1116 #include "reg-notes.def"
1117 #undef REG_CFA_NOTE
1118 #undef DEF_REG_NOTE
1119 false
1122 /* Return true if I1 and I2 have identical CFA notes (the same order
1123 and equivalent content). */
1125 static bool
1126 insns_have_identical_cfa_notes (rtx_insn *i1, rtx_insn *i2)
1128 rtx n1, n2;
1129 for (n1 = REG_NOTES (i1), n2 = REG_NOTES (i2); ;
1130 n1 = XEXP (n1, 1), n2 = XEXP (n2, 1))
1132 /* Skip over reg notes not related to CFI information. */
1133 while (n1 && !reg_note_cfa_p[REG_NOTE_KIND (n1)])
1134 n1 = XEXP (n1, 1);
1135 while (n2 && !reg_note_cfa_p[REG_NOTE_KIND (n2)])
1136 n2 = XEXP (n2, 1);
1137 if (n1 == NULL_RTX && n2 == NULL_RTX)
1138 return true;
1139 if (n1 == NULL_RTX || n2 == NULL_RTX)
1140 return false;
1141 if (XEXP (n1, 0) == XEXP (n2, 0))
1143 else if (XEXP (n1, 0) == NULL_RTX || XEXP (n2, 0) == NULL_RTX)
1144 return false;
1145 else if (!(reload_completed
1146 ? rtx_renumbered_equal_p (XEXP (n1, 0), XEXP (n2, 0))
1147 : rtx_equal_p (XEXP (n1, 0), XEXP (n2, 0))))
1148 return false;
1152 /* Examine I1 and I2 and return:
1153 - dir_forward if I1 can be replaced by I2, or
1154 - dir_backward if I2 can be replaced by I1, or
1155 - dir_both if both are the case. */
1157 static enum replace_direction
1158 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1160 rtx p1, p2;
1162 /* Verify that I1 and I2 are equivalent. */
1163 if (GET_CODE (i1) != GET_CODE (i2))
1164 return dir_none;
1166 /* __builtin_unreachable() may lead to empty blocks (ending with
1167 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1168 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1169 return dir_both;
1171 /* ??? Do not allow cross-jumping between different stack levels. */
1172 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1173 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1174 if (p1 && p2)
1176 p1 = XEXP (p1, 0);
1177 p2 = XEXP (p2, 0);
1178 if (!rtx_equal_p (p1, p2))
1179 return dir_none;
1181 /* ??? Worse, this adjustment had better be constant lest we
1182 have differing incoming stack levels. */
1183 if (!frame_pointer_needed
1184 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1185 return dir_none;
1187 else if (p1 || p2)
1188 return dir_none;
1190 /* Do not allow cross-jumping between frame related insns and other
1191 insns. */
1192 if (RTX_FRAME_RELATED_P (i1) != RTX_FRAME_RELATED_P (i2))
1193 return dir_none;
1195 p1 = PATTERN (i1);
1196 p2 = PATTERN (i2);
1198 if (GET_CODE (p1) != GET_CODE (p2))
1199 return dir_none;
1201 /* If this is a CALL_INSN, compare register usage information.
1202 If we don't check this on stack register machines, the two
1203 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1204 numbers of stack registers in the same basic block.
1205 If we don't check this on machines with delay slots, a delay slot may
1206 be filled that clobbers a parameter expected by the subroutine.
1208 ??? We take the simple route for now and assume that if they're
1209 equal, they were constructed identically.
1211 Also check for identical exception regions. */
1213 if (CALL_P (i1))
1215 /* Ensure the same EH region. */
1216 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1217 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1219 if (!n1 && n2)
1220 return dir_none;
1222 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1223 return dir_none;
1225 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1226 CALL_INSN_FUNCTION_USAGE (i2))
1227 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1228 return dir_none;
1230 /* For address sanitizer, never crossjump __asan_report_* builtins,
1231 otherwise errors might be reported on incorrect lines. */
1232 if (flag_sanitize & SANITIZE_ADDRESS)
1234 rtx call = get_call_rtx_from (i1);
1235 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1237 rtx symbol = XEXP (XEXP (call, 0), 0);
1238 if (SYMBOL_REF_DECL (symbol)
1239 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1241 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1242 == BUILT_IN_NORMAL)
1243 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1244 >= BUILT_IN_ASAN_REPORT_LOAD1
1245 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1246 <= BUILT_IN_ASAN_STOREN)
1247 return dir_none;
1253 /* If both i1 and i2 are frame related, verify all the CFA notes
1254 in the same order and with the same content. */
1255 if (RTX_FRAME_RELATED_P (i1) && !insns_have_identical_cfa_notes (i1, i2))
1256 return dir_none;
1258 #ifdef STACK_REGS
1259 /* If cross_jump_death_matters is not 0, the insn's mode
1260 indicates whether or not the insn contains any stack-like
1261 regs. */
1263 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1265 /* If register stack conversion has already been done, then
1266 death notes must also be compared before it is certain that
1267 the two instruction streams match. */
1269 rtx note;
1270 HARD_REG_SET i1_regset, i2_regset;
1272 CLEAR_HARD_REG_SET (i1_regset);
1273 CLEAR_HARD_REG_SET (i2_regset);
1275 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1276 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1277 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1279 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1280 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1281 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1283 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1284 return dir_none;
1286 #endif
1288 if (reload_completed
1289 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1290 return dir_both;
1292 return can_replace_by (i1, i2);
1295 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1296 flow_find_head_matching_sequence, ensure the notes match. */
1298 static void
1299 merge_notes (rtx_insn *i1, rtx_insn *i2)
1301 /* If the merged insns have different REG_EQUAL notes, then
1302 remove them. */
1303 rtx equiv1 = find_reg_equal_equiv_note (i1);
1304 rtx equiv2 = find_reg_equal_equiv_note (i2);
1306 if (equiv1 && !equiv2)
1307 remove_note (i1, equiv1);
1308 else if (!equiv1 && equiv2)
1309 remove_note (i2, equiv2);
1310 else if (equiv1 && equiv2
1311 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1313 remove_note (i1, equiv1);
1314 remove_note (i2, equiv2);
1318 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1319 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1320 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1321 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1322 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1324 static void
1325 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1326 bool *did_fallthru)
1328 edge fallthru;
1330 *did_fallthru = false;
1332 /* Ignore notes. */
1333 while (!NONDEBUG_INSN_P (*i1))
1335 if (*i1 != BB_HEAD (*bb1))
1337 *i1 = PREV_INSN (*i1);
1338 continue;
1341 if (!follow_fallthru)
1342 return;
1344 fallthru = find_fallthru_edge ((*bb1)->preds);
1345 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1346 || !single_succ_p (fallthru->src))
1347 return;
1349 *bb1 = fallthru->src;
1350 *i1 = BB_END (*bb1);
1351 *did_fallthru = true;
1355 /* Look through the insns at the end of BB1 and BB2 and find the longest
1356 sequence that are either equivalent, or allow forward or backward
1357 replacement. Store the first insns for that sequence in *F1 and *F2 and
1358 return the sequence length.
1360 DIR_P indicates the allowed replacement direction on function entry, and
1361 the actual replacement direction on function exit. If NULL, only equivalent
1362 sequences are allowed.
1364 To simplify callers of this function, if the blocks match exactly,
1365 store the head of the blocks in *F1 and *F2. */
1368 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1369 rtx_insn **f2, enum replace_direction *dir_p)
1371 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1372 int ninsns = 0;
1373 enum replace_direction dir, last_dir, afterlast_dir;
1374 bool follow_fallthru, did_fallthru;
1376 if (dir_p)
1377 dir = *dir_p;
1378 else
1379 dir = dir_both;
1380 afterlast_dir = dir;
1381 last_dir = afterlast_dir;
1383 /* Skip simple jumps at the end of the blocks. Complex jumps still
1384 need to be compared for equivalence, which we'll do below. */
1386 i1 = BB_END (bb1);
1387 last1 = afterlast1 = last2 = afterlast2 = NULL;
1388 if (onlyjump_p (i1)
1389 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1391 last1 = i1;
1392 i1 = PREV_INSN (i1);
1395 i2 = BB_END (bb2);
1396 if (onlyjump_p (i2)
1397 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1399 last2 = i2;
1400 /* Count everything except for unconditional jump as insn.
1401 Don't count any jumps if dir_p is NULL. */
1402 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1403 ninsns++;
1404 i2 = PREV_INSN (i2);
1407 while (true)
1409 /* In the following example, we can replace all jumps to C by jumps to A.
1411 This removes 4 duplicate insns.
1412 [bb A] insn1 [bb C] insn1
1413 insn2 insn2
1414 [bb B] insn3 insn3
1415 insn4 insn4
1416 jump_insn jump_insn
1418 We could also replace all jumps to A by jumps to C, but that leaves B
1419 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1420 step, all jumps to B would be replaced with jumps to the middle of C,
1421 achieving the same result with more effort.
1422 So we allow only the first possibility, which means that we don't allow
1423 fallthru in the block that's being replaced. */
1425 follow_fallthru = dir_p && dir != dir_forward;
1426 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1427 if (did_fallthru)
1428 dir = dir_backward;
1430 follow_fallthru = dir_p && dir != dir_backward;
1431 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1432 if (did_fallthru)
1433 dir = dir_forward;
1435 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1436 break;
1438 /* Do not turn corssing edge to non-crossing or vice versa after
1439 reload. */
1440 if (BB_PARTITION (BLOCK_FOR_INSN (i1))
1441 != BB_PARTITION (BLOCK_FOR_INSN (i2))
1442 && reload_completed)
1443 break;
1445 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1446 if (dir == dir_none || (!dir_p && dir != dir_both))
1447 break;
1449 merge_memattrs (i1, i2);
1451 /* Don't begin a cross-jump with a NOTE insn. */
1452 if (INSN_P (i1))
1454 merge_notes (i1, i2);
1456 afterlast1 = last1, afterlast2 = last2;
1457 last1 = i1, last2 = i2;
1458 afterlast_dir = last_dir;
1459 last_dir = dir;
1460 if (active_insn_p (i1))
1461 ninsns++;
1464 i1 = PREV_INSN (i1);
1465 i2 = PREV_INSN (i2);
1468 /* Don't allow the insn after a compare to be shared by
1469 cross-jumping unless the compare is also shared. */
1470 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1471 && ! sets_cc0_p (last1))
1472 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1474 /* Include preceding notes and labels in the cross-jump. One,
1475 this may bring us to the head of the blocks as requested above.
1476 Two, it keeps line number notes as matched as may be. */
1477 if (ninsns)
1479 bb1 = BLOCK_FOR_INSN (last1);
1480 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1481 last1 = PREV_INSN (last1);
1483 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1484 last1 = PREV_INSN (last1);
1486 bb2 = BLOCK_FOR_INSN (last2);
1487 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1488 last2 = PREV_INSN (last2);
1490 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1491 last2 = PREV_INSN (last2);
1493 *f1 = last1;
1494 *f2 = last2;
1497 if (dir_p)
1498 *dir_p = last_dir;
1499 return ninsns;
1502 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1503 the head of the two blocks. Do not include jumps at the end.
1504 If STOP_AFTER is nonzero, stop after finding that many matching
1505 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1506 non-zero, only count active insns. */
1509 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1510 rtx_insn **f2, int stop_after)
1512 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1513 int ninsns = 0;
1514 edge e;
1515 edge_iterator ei;
1516 int nehedges1 = 0, nehedges2 = 0;
1518 FOR_EACH_EDGE (e, ei, bb1->succs)
1519 if (e->flags & EDGE_EH)
1520 nehedges1++;
1521 FOR_EACH_EDGE (e, ei, bb2->succs)
1522 if (e->flags & EDGE_EH)
1523 nehedges2++;
1525 i1 = BB_HEAD (bb1);
1526 i2 = BB_HEAD (bb2);
1527 last1 = beforelast1 = last2 = beforelast2 = NULL;
1529 while (true)
1531 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1532 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1534 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1535 break;
1536 i1 = NEXT_INSN (i1);
1539 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1541 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1542 break;
1543 i2 = NEXT_INSN (i2);
1546 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1547 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1548 break;
1550 if (NOTE_P (i1) || NOTE_P (i2)
1551 || JUMP_P (i1) || JUMP_P (i2))
1552 break;
1554 /* A sanity check to make sure we're not merging insns with different
1555 effects on EH. If only one of them ends a basic block, it shouldn't
1556 have an EH edge; if both end a basic block, there should be the same
1557 number of EH edges. */
1558 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1559 && nehedges1 > 0)
1560 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1561 && nehedges2 > 0)
1562 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1563 && nehedges1 != nehedges2))
1564 break;
1566 if (old_insns_match_p (0, i1, i2) != dir_both)
1567 break;
1569 merge_memattrs (i1, i2);
1571 /* Don't begin a cross-jump with a NOTE insn. */
1572 if (INSN_P (i1))
1574 merge_notes (i1, i2);
1576 beforelast1 = last1, beforelast2 = last2;
1577 last1 = i1, last2 = i2;
1578 if (!stop_after || active_insn_p (i1))
1579 ninsns++;
1582 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1583 || (stop_after > 0 && ninsns == stop_after))
1584 break;
1586 i1 = NEXT_INSN (i1);
1587 i2 = NEXT_INSN (i2);
1590 /* Don't allow a compare to be shared by cross-jumping unless the insn
1591 after the compare is also shared. */
1592 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1593 && sets_cc0_p (last1))
1594 last1 = beforelast1, last2 = beforelast2, ninsns--;
1596 if (ninsns)
1598 *f1 = last1;
1599 *f2 = last2;
1602 return ninsns;
1605 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1606 the branch instruction. This means that if we commonize the control
1607 flow before end of the basic block, the semantic remains unchanged.
1609 We may assume that there exists one edge with a common destination. */
1611 static bool
1612 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1614 int nehedges1 = 0, nehedges2 = 0;
1615 edge fallthru1 = 0, fallthru2 = 0;
1616 edge e1, e2;
1617 edge_iterator ei;
1619 /* If we performed shrink-wrapping, edges to the exit block can
1620 only be distinguished for JUMP_INSNs. The two paths may differ in
1621 whether they went through the prologue. Sibcalls are fine, we know
1622 that we either didn't need or inserted an epilogue before them. */
1623 if (crtl->shrink_wrapped
1624 && single_succ_p (bb1)
1625 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1626 && !JUMP_P (BB_END (bb1))
1627 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1628 return false;
1630 /* If BB1 has only one successor, we may be looking at either an
1631 unconditional jump, or a fake edge to exit. */
1632 if (single_succ_p (bb1)
1633 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1634 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1635 return (single_succ_p (bb2)
1636 && (single_succ_edge (bb2)->flags
1637 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1638 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1640 /* Match conditional jumps - this may get tricky when fallthru and branch
1641 edges are crossed. */
1642 if (EDGE_COUNT (bb1->succs) == 2
1643 && any_condjump_p (BB_END (bb1))
1644 && onlyjump_p (BB_END (bb1)))
1646 edge b1, f1, b2, f2;
1647 bool reverse, match;
1648 rtx set1, set2, cond1, cond2;
1649 enum rtx_code code1, code2;
1651 if (EDGE_COUNT (bb2->succs) != 2
1652 || !any_condjump_p (BB_END (bb2))
1653 || !onlyjump_p (BB_END (bb2)))
1654 return false;
1656 b1 = BRANCH_EDGE (bb1);
1657 b2 = BRANCH_EDGE (bb2);
1658 f1 = FALLTHRU_EDGE (bb1);
1659 f2 = FALLTHRU_EDGE (bb2);
1661 /* Get around possible forwarders on fallthru edges. Other cases
1662 should be optimized out already. */
1663 if (FORWARDER_BLOCK_P (f1->dest))
1664 f1 = single_succ_edge (f1->dest);
1666 if (FORWARDER_BLOCK_P (f2->dest))
1667 f2 = single_succ_edge (f2->dest);
1669 /* To simplify use of this function, return false if there are
1670 unneeded forwarder blocks. These will get eliminated later
1671 during cleanup_cfg. */
1672 if (FORWARDER_BLOCK_P (f1->dest)
1673 || FORWARDER_BLOCK_P (f2->dest)
1674 || FORWARDER_BLOCK_P (b1->dest)
1675 || FORWARDER_BLOCK_P (b2->dest))
1676 return false;
1678 if (f1->dest == f2->dest && b1->dest == b2->dest)
1679 reverse = false;
1680 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1681 reverse = true;
1682 else
1683 return false;
1685 set1 = pc_set (BB_END (bb1));
1686 set2 = pc_set (BB_END (bb2));
1687 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1688 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1689 reverse = !reverse;
1691 cond1 = XEXP (SET_SRC (set1), 0);
1692 cond2 = XEXP (SET_SRC (set2), 0);
1693 code1 = GET_CODE (cond1);
1694 if (reverse)
1695 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1696 else
1697 code2 = GET_CODE (cond2);
1699 if (code2 == UNKNOWN)
1700 return false;
1702 /* Verify codes and operands match. */
1703 match = ((code1 == code2
1704 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1705 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1706 || (code1 == swap_condition (code2)
1707 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1708 XEXP (cond2, 0))
1709 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1710 XEXP (cond2, 1))));
1712 /* If we return true, we will join the blocks. Which means that
1713 we will only have one branch prediction bit to work with. Thus
1714 we require the existing branches to have probabilities that are
1715 roughly similar. */
1716 if (match
1717 && optimize_bb_for_speed_p (bb1)
1718 && optimize_bb_for_speed_p (bb2))
1720 profile_probability prob2;
1722 if (b1->dest == b2->dest)
1723 prob2 = b2->probability;
1724 else
1725 /* Do not use f2 probability as f2 may be forwarded. */
1726 prob2 = b2->probability.invert ();
1728 /* Fail if the difference in probabilities is greater than 50%.
1729 This rules out two well-predicted branches with opposite
1730 outcomes. */
1731 if (b1->probability.differs_lot_from_p (prob2))
1733 if (dump_file)
1735 fprintf (dump_file,
1736 "Outcomes of branch in bb %i and %i differ too"
1737 " much (", bb1->index, bb2->index);
1738 b1->probability.dump (dump_file);
1739 prob2.dump (dump_file);
1740 fprintf (dump_file, ")\n");
1742 return false;
1746 if (dump_file && match)
1747 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1748 bb1->index, bb2->index);
1750 return match;
1753 /* Generic case - we are seeing a computed jump, table jump or trapping
1754 instruction. */
1756 /* Check whether there are tablejumps in the end of BB1 and BB2.
1757 Return true if they are identical. */
1759 rtx_insn *label1, *label2;
1760 rtx_jump_table_data *table1, *table2;
1762 if (tablejump_p (BB_END (bb1), &label1, &table1)
1763 && tablejump_p (BB_END (bb2), &label2, &table2)
1764 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1766 /* The labels should never be the same rtx. If they really are same
1767 the jump tables are same too. So disable crossjumping of blocks BB1
1768 and BB2 because when deleting the common insns in the end of BB1
1769 by delete_basic_block () the jump table would be deleted too. */
1770 /* If LABEL2 is referenced in BB1->END do not do anything
1771 because we would loose information when replacing
1772 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1773 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1775 /* Set IDENTICAL to true when the tables are identical. */
1776 bool identical = false;
1777 rtx p1, p2;
1779 p1 = PATTERN (table1);
1780 p2 = PATTERN (table2);
1781 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1783 identical = true;
1785 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1786 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1787 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1788 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1790 int i;
1792 identical = true;
1793 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1794 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1795 identical = false;
1798 if (identical)
1800 bool match;
1802 /* Temporarily replace references to LABEL1 with LABEL2
1803 in BB1->END so that we could compare the instructions. */
1804 replace_label_in_insn (BB_END (bb1), label1, label2, false);
1806 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1807 == dir_both);
1808 if (dump_file && match)
1809 fprintf (dump_file,
1810 "Tablejumps in bb %i and %i match.\n",
1811 bb1->index, bb2->index);
1813 /* Set the original label in BB1->END because when deleting
1814 a block whose end is a tablejump, the tablejump referenced
1815 from the instruction is deleted too. */
1816 replace_label_in_insn (BB_END (bb1), label2, label1, false);
1818 return match;
1821 return false;
1825 /* Find the last non-debug non-note instruction in each bb, except
1826 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1827 handles that case specially. old_insns_match_p does not handle
1828 other types of instruction notes. */
1829 rtx_insn *last1 = BB_END (bb1);
1830 rtx_insn *last2 = BB_END (bb2);
1831 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1832 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1833 last1 = PREV_INSN (last1);
1834 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1835 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1836 last2 = PREV_INSN (last2);
1837 gcc_assert (last1 && last2);
1839 /* First ensure that the instructions match. There may be many outgoing
1840 edges so this test is generally cheaper. */
1841 if (old_insns_match_p (mode, last1, last2) != dir_both)
1842 return false;
1844 /* Search the outgoing edges, ensure that the counts do match, find possible
1845 fallthru and exception handling edges since these needs more
1846 validation. */
1847 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1848 return false;
1850 bool nonfakeedges = false;
1851 FOR_EACH_EDGE (e1, ei, bb1->succs)
1853 e2 = EDGE_SUCC (bb2, ei.index);
1855 if ((e1->flags & EDGE_FAKE) == 0)
1856 nonfakeedges = true;
1858 if (e1->flags & EDGE_EH)
1859 nehedges1++;
1861 if (e2->flags & EDGE_EH)
1862 nehedges2++;
1864 if (e1->flags & EDGE_FALLTHRU)
1865 fallthru1 = e1;
1866 if (e2->flags & EDGE_FALLTHRU)
1867 fallthru2 = e2;
1870 /* If number of edges of various types does not match, fail. */
1871 if (nehedges1 != nehedges2
1872 || (fallthru1 != 0) != (fallthru2 != 0))
1873 return false;
1875 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1876 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1877 attempt to optimize, as the two basic blocks might have different
1878 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1879 traps there should be REG_ARG_SIZE notes, they could be missing
1880 for __builtin_unreachable () uses though. */
1881 if (!nonfakeedges
1882 && !ACCUMULATE_OUTGOING_ARGS
1883 && (!INSN_P (last1)
1884 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1885 return false;
1887 /* fallthru edges must be forwarded to the same destination. */
1888 if (fallthru1)
1890 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1891 ? single_succ (fallthru1->dest): fallthru1->dest);
1892 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1893 ? single_succ (fallthru2->dest): fallthru2->dest);
1895 if (d1 != d2)
1896 return false;
1899 /* Ensure the same EH region. */
1901 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1902 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1904 if (!n1 && n2)
1905 return false;
1907 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1908 return false;
1911 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1912 version of sequence abstraction. */
1913 FOR_EACH_EDGE (e1, ei, bb2->succs)
1915 edge e2;
1916 edge_iterator ei;
1917 basic_block d1 = e1->dest;
1919 if (FORWARDER_BLOCK_P (d1))
1920 d1 = EDGE_SUCC (d1, 0)->dest;
1922 FOR_EACH_EDGE (e2, ei, bb1->succs)
1924 basic_block d2 = e2->dest;
1925 if (FORWARDER_BLOCK_P (d2))
1926 d2 = EDGE_SUCC (d2, 0)->dest;
1927 if (d1 == d2)
1928 break;
1931 if (!e2)
1932 return false;
1935 return true;
1938 /* Returns true if BB basic block has a preserve label. */
1940 static bool
1941 block_has_preserve_label (basic_block bb)
1943 return (bb
1944 && block_label (bb)
1945 && LABEL_PRESERVE_P (block_label (bb)));
1948 /* E1 and E2 are edges with the same destination block. Search their
1949 predecessors for common code. If found, redirect control flow from
1950 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1951 or the other way around (dir_backward). DIR specifies the allowed
1952 replacement direction. */
1954 static bool
1955 try_crossjump_to_edge (int mode, edge e1, edge e2,
1956 enum replace_direction dir)
1958 int nmatch;
1959 basic_block src1 = e1->src, src2 = e2->src;
1960 basic_block redirect_to, redirect_from, to_remove;
1961 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1962 rtx_insn *newpos1, *newpos2;
1963 edge s;
1964 edge_iterator ei;
1966 newpos1 = newpos2 = NULL;
1968 /* Search backward through forwarder blocks. We don't need to worry
1969 about multiple entry or chained forwarders, as they will be optimized
1970 away. We do this to look past the unconditional jump following a
1971 conditional jump that is required due to the current CFG shape. */
1972 if (single_pred_p (src1)
1973 && FORWARDER_BLOCK_P (src1))
1974 e1 = single_pred_edge (src1), src1 = e1->src;
1976 if (single_pred_p (src2)
1977 && FORWARDER_BLOCK_P (src2))
1978 e2 = single_pred_edge (src2), src2 = e2->src;
1980 /* Nothing to do if we reach ENTRY, or a common source block. */
1981 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1982 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1983 return false;
1984 if (src1 == src2)
1985 return false;
1987 /* Seeing more than 1 forwarder blocks would confuse us later... */
1988 if (FORWARDER_BLOCK_P (e1->dest)
1989 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1990 return false;
1992 if (FORWARDER_BLOCK_P (e2->dest)
1993 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1994 return false;
1996 /* Likewise with dead code (possibly newly created by the other optimizations
1997 of cfg_cleanup). */
1998 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1999 return false;
2001 /* Do not turn corssing edge to non-crossing or vice versa after reload. */
2002 if (BB_PARTITION (src1) != BB_PARTITION (src2)
2003 && reload_completed)
2004 return false;
2006 /* Look for the common insn sequence, part the first ... */
2007 if (!outgoing_edges_match (mode, src1, src2))
2008 return false;
2010 /* ... and part the second. */
2011 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
2013 osrc1 = src1;
2014 osrc2 = src2;
2015 if (newpos1 != NULL_RTX)
2016 src1 = BLOCK_FOR_INSN (newpos1);
2017 if (newpos2 != NULL_RTX)
2018 src2 = BLOCK_FOR_INSN (newpos2);
2020 /* Check that SRC1 and SRC2 have preds again. They may have changed
2021 above due to the call to flow_find_cross_jump. */
2022 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
2023 return false;
2025 if (dir == dir_backward)
2027 std::swap (osrc1, osrc2);
2028 std::swap (src1, src2);
2029 std::swap (e1, e2);
2030 std::swap (newpos1, newpos2);
2033 /* Don't proceed with the crossjump unless we found a sufficient number
2034 of matching instructions or the 'from' block was totally matched
2035 (such that its predecessors will hopefully be redirected and the
2036 block removed). */
2037 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
2038 && (newpos1 != BB_HEAD (src1)))
2039 return false;
2041 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
2042 if (block_has_preserve_label (e1->dest)
2043 && (e1->flags & EDGE_ABNORMAL))
2044 return false;
2046 /* Here we know that the insns in the end of SRC1 which are common with SRC2
2047 will be deleted.
2048 If we have tablejumps in the end of SRC1 and SRC2
2049 they have been already compared for equivalence in outgoing_edges_match ()
2050 so replace the references to TABLE1 by references to TABLE2. */
2052 rtx_insn *label1, *label2;
2053 rtx_jump_table_data *table1, *table2;
2055 if (tablejump_p (BB_END (osrc1), &label1, &table1)
2056 && tablejump_p (BB_END (osrc2), &label2, &table2)
2057 && label1 != label2)
2059 rtx_insn *insn;
2061 /* Replace references to LABEL1 with LABEL2. */
2062 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2064 /* Do not replace the label in SRC1->END because when deleting
2065 a block whose end is a tablejump, the tablejump referenced
2066 from the instruction is deleted too. */
2067 if (insn != BB_END (osrc1))
2068 replace_label_in_insn (insn, label1, label2, true);
2073 /* Avoid splitting if possible. We must always split when SRC2 has
2074 EH predecessor edges, or we may end up with basic blocks with both
2075 normal and EH predecessor edges. */
2076 if (newpos2 == BB_HEAD (src2)
2077 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2078 redirect_to = src2;
2079 else
2081 if (newpos2 == BB_HEAD (src2))
2083 /* Skip possible basic block header. */
2084 if (LABEL_P (newpos2))
2085 newpos2 = NEXT_INSN (newpos2);
2086 while (DEBUG_INSN_P (newpos2))
2087 newpos2 = NEXT_INSN (newpos2);
2088 if (NOTE_P (newpos2))
2089 newpos2 = NEXT_INSN (newpos2);
2090 while (DEBUG_INSN_P (newpos2))
2091 newpos2 = NEXT_INSN (newpos2);
2094 if (dump_file)
2095 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2096 src2->index, nmatch);
2097 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2100 if (dump_file)
2101 fprintf (dump_file,
2102 "Cross jumping from bb %i to bb %i; %i common insns\n",
2103 src1->index, src2->index, nmatch);
2105 /* We may have some registers visible through the block. */
2106 df_set_bb_dirty (redirect_to);
2108 if (osrc2 == src2)
2109 redirect_edges_to = redirect_to;
2110 else
2111 redirect_edges_to = osrc2;
2113 /* Recompute the frequencies and counts of outgoing edges. */
2114 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2116 edge s2;
2117 edge_iterator ei;
2118 basic_block d = s->dest;
2120 if (FORWARDER_BLOCK_P (d))
2121 d = single_succ (d);
2123 FOR_EACH_EDGE (s2, ei, src1->succs)
2125 basic_block d2 = s2->dest;
2126 if (FORWARDER_BLOCK_P (d2))
2127 d2 = single_succ (d2);
2128 if (d == d2)
2129 break;
2132 s->count += s2->count;
2134 /* Take care to update possible forwarder blocks. We verified
2135 that there is no more than one in the chain, so we can't run
2136 into infinite loop. */
2137 if (FORWARDER_BLOCK_P (s->dest))
2139 single_succ_edge (s->dest)->count += s2->count;
2140 s->dest->count += s2->count;
2141 s->dest->frequency += EDGE_FREQUENCY (s);
2144 if (FORWARDER_BLOCK_P (s2->dest))
2146 single_succ_edge (s2->dest)->count -= s2->count;
2147 s2->dest->count -= s2->count;
2148 s2->dest->frequency -= EDGE_FREQUENCY (s);
2149 if (s2->dest->frequency < 0)
2150 s2->dest->frequency = 0;
2153 if (!redirect_edges_to->frequency && !src1->frequency)
2154 s->probability = s->probability.combine_with_freq
2155 (redirect_edges_to->frequency,
2156 s2->probability, src1->frequency);
2159 /* Adjust count and frequency for the block. An earlier jump
2160 threading pass may have left the profile in an inconsistent
2161 state (see update_bb_profile_for_threading) so we must be
2162 prepared for overflows. */
2163 tmp = redirect_to;
2166 tmp->count += src1->count;
2167 tmp->frequency += src1->frequency;
2168 if (tmp->frequency > BB_FREQ_MAX)
2169 tmp->frequency = BB_FREQ_MAX;
2170 if (tmp == redirect_edges_to)
2171 break;
2172 tmp = find_fallthru_edge (tmp->succs)->dest;
2174 while (true);
2175 update_br_prob_note (redirect_edges_to);
2177 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2179 /* Skip possible basic block header. */
2180 if (LABEL_P (newpos1))
2181 newpos1 = NEXT_INSN (newpos1);
2183 while (DEBUG_INSN_P (newpos1))
2184 newpos1 = NEXT_INSN (newpos1);
2186 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2187 newpos1 = NEXT_INSN (newpos1);
2189 while (DEBUG_INSN_P (newpos1))
2190 newpos1 = NEXT_INSN (newpos1);
2192 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2193 to_remove = single_succ (redirect_from);
2195 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2196 delete_basic_block (to_remove);
2198 update_forwarder_flag (redirect_from);
2199 if (redirect_to != src2)
2200 update_forwarder_flag (src2);
2202 return true;
2205 /* Search the predecessors of BB for common insn sequences. When found,
2206 share code between them by redirecting control flow. Return true if
2207 any changes made. */
2209 static bool
2210 try_crossjump_bb (int mode, basic_block bb)
2212 edge e, e2, fallthru;
2213 bool changed;
2214 unsigned max, ix, ix2;
2216 /* Nothing to do if there is not at least two incoming edges. */
2217 if (EDGE_COUNT (bb->preds) < 2)
2218 return false;
2220 /* Don't crossjump if this block ends in a computed jump,
2221 unless we are optimizing for size. */
2222 if (optimize_bb_for_size_p (bb)
2223 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2224 && computed_jump_p (BB_END (bb)))
2225 return false;
2227 /* If we are partitioning hot/cold basic blocks, we don't want to
2228 mess up unconditional or indirect jumps that cross between hot
2229 and cold sections.
2231 Basic block partitioning may result in some jumps that appear to
2232 be optimizable (or blocks that appear to be mergeable), but which really
2233 must be left untouched (they are required to make it safely across
2234 partition boundaries). See the comments at the top of
2235 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2237 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2238 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2239 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2240 return false;
2242 /* It is always cheapest to redirect a block that ends in a branch to
2243 a block that falls through into BB, as that adds no branches to the
2244 program. We'll try that combination first. */
2245 fallthru = NULL;
2246 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2248 if (EDGE_COUNT (bb->preds) > max)
2249 return false;
2251 fallthru = find_fallthru_edge (bb->preds);
2253 changed = false;
2254 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2256 e = EDGE_PRED (bb, ix);
2257 ix++;
2259 /* As noted above, first try with the fallthru predecessor (or, a
2260 fallthru predecessor if we are in cfglayout mode). */
2261 if (fallthru)
2263 /* Don't combine the fallthru edge into anything else.
2264 If there is a match, we'll do it the other way around. */
2265 if (e == fallthru)
2266 continue;
2267 /* If nothing changed since the last attempt, there is nothing
2268 we can do. */
2269 if (!first_pass
2270 && !((e->src->flags & BB_MODIFIED)
2271 || (fallthru->src->flags & BB_MODIFIED)))
2272 continue;
2274 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2276 changed = true;
2277 ix = 0;
2278 continue;
2282 /* Non-obvious work limiting check: Recognize that we're going
2283 to call try_crossjump_bb on every basic block. So if we have
2284 two blocks with lots of outgoing edges (a switch) and they
2285 share lots of common destinations, then we would do the
2286 cross-jump check once for each common destination.
2288 Now, if the blocks actually are cross-jump candidates, then
2289 all of their destinations will be shared. Which means that
2290 we only need check them for cross-jump candidacy once. We
2291 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2292 choosing to do the check from the block for which the edge
2293 in question is the first successor of A. */
2294 if (EDGE_SUCC (e->src, 0) != e)
2295 continue;
2297 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2299 e2 = EDGE_PRED (bb, ix2);
2301 if (e2 == e)
2302 continue;
2304 /* We've already checked the fallthru edge above. */
2305 if (e2 == fallthru)
2306 continue;
2308 /* The "first successor" check above only prevents multiple
2309 checks of crossjump(A,B). In order to prevent redundant
2310 checks of crossjump(B,A), require that A be the block
2311 with the lowest index. */
2312 if (e->src->index > e2->src->index)
2313 continue;
2315 /* If nothing changed since the last attempt, there is nothing
2316 we can do. */
2317 if (!first_pass
2318 && !((e->src->flags & BB_MODIFIED)
2319 || (e2->src->flags & BB_MODIFIED)))
2320 continue;
2322 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2323 direction. */
2324 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2326 changed = true;
2327 ix = 0;
2328 break;
2333 if (changed)
2334 crossjumps_occurred = true;
2336 return changed;
2339 /* Search the successors of BB for common insn sequences. When found,
2340 share code between them by moving it across the basic block
2341 boundary. Return true if any changes made. */
2343 static bool
2344 try_head_merge_bb (basic_block bb)
2346 basic_block final_dest_bb = NULL;
2347 int max_match = INT_MAX;
2348 edge e0;
2349 rtx_insn **headptr, **currptr, **nextptr;
2350 bool changed, moveall;
2351 unsigned ix;
2352 rtx_insn *e0_last_head;
2353 rtx cond;
2354 rtx_insn *move_before;
2355 unsigned nedges = EDGE_COUNT (bb->succs);
2356 rtx_insn *jump = BB_END (bb);
2357 regset live, live_union;
2359 /* Nothing to do if there is not at least two outgoing edges. */
2360 if (nedges < 2)
2361 return false;
2363 /* Don't crossjump if this block ends in a computed jump,
2364 unless we are optimizing for size. */
2365 if (optimize_bb_for_size_p (bb)
2366 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2367 && computed_jump_p (BB_END (bb)))
2368 return false;
2370 cond = get_condition (jump, &move_before, true, false);
2371 if (cond == NULL_RTX)
2373 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2374 move_before = prev_nonnote_nondebug_insn (jump);
2375 else
2376 move_before = jump;
2379 for (ix = 0; ix < nedges; ix++)
2380 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2381 return false;
2383 for (ix = 0; ix < nedges; ix++)
2385 edge e = EDGE_SUCC (bb, ix);
2386 basic_block other_bb = e->dest;
2388 if (df_get_bb_dirty (other_bb))
2390 block_was_dirty = true;
2391 return false;
2394 if (e->flags & EDGE_ABNORMAL)
2395 return false;
2397 /* Normally, all destination blocks must only be reachable from this
2398 block, i.e. they must have one incoming edge.
2400 There is one special case we can handle, that of multiple consecutive
2401 jumps where the first jumps to one of the targets of the second jump.
2402 This happens frequently in switch statements for default labels.
2403 The structure is as follows:
2404 FINAL_DEST_BB
2405 ....
2406 if (cond) jump A;
2407 fall through
2409 jump with targets A, B, C, D...
2411 has two incoming edges, from FINAL_DEST_BB and BB
2413 In this case, we can try to move the insns through BB and into
2414 FINAL_DEST_BB. */
2415 if (EDGE_COUNT (other_bb->preds) != 1)
2417 edge incoming_edge, incoming_bb_other_edge;
2418 edge_iterator ei;
2420 if (final_dest_bb != NULL
2421 || EDGE_COUNT (other_bb->preds) != 2)
2422 return false;
2424 /* We must be able to move the insns across the whole block. */
2425 move_before = BB_HEAD (bb);
2426 while (!NONDEBUG_INSN_P (move_before))
2427 move_before = NEXT_INSN (move_before);
2429 if (EDGE_COUNT (bb->preds) != 1)
2430 return false;
2431 incoming_edge = EDGE_PRED (bb, 0);
2432 final_dest_bb = incoming_edge->src;
2433 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2434 return false;
2435 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2436 if (incoming_bb_other_edge != incoming_edge)
2437 break;
2438 if (incoming_bb_other_edge->dest != other_bb)
2439 return false;
2443 e0 = EDGE_SUCC (bb, 0);
2444 e0_last_head = NULL;
2445 changed = false;
2447 for (ix = 1; ix < nedges; ix++)
2449 edge e = EDGE_SUCC (bb, ix);
2450 rtx_insn *e0_last, *e_last;
2451 int nmatch;
2453 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2454 &e0_last, &e_last, 0);
2455 if (nmatch == 0)
2456 return false;
2458 if (nmatch < max_match)
2460 max_match = nmatch;
2461 e0_last_head = e0_last;
2465 /* If we matched an entire block, we probably have to avoid moving the
2466 last insn. */
2467 if (max_match > 0
2468 && e0_last_head == BB_END (e0->dest)
2469 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2470 || control_flow_insn_p (e0_last_head)))
2472 max_match--;
2473 if (max_match == 0)
2474 return false;
2476 e0_last_head = prev_real_insn (e0_last_head);
2477 while (DEBUG_INSN_P (e0_last_head));
2480 if (max_match == 0)
2481 return false;
2483 /* We must find a union of the live registers at each of the end points. */
2484 live = BITMAP_ALLOC (NULL);
2485 live_union = BITMAP_ALLOC (NULL);
2487 currptr = XNEWVEC (rtx_insn *, nedges);
2488 headptr = XNEWVEC (rtx_insn *, nedges);
2489 nextptr = XNEWVEC (rtx_insn *, nedges);
2491 for (ix = 0; ix < nedges; ix++)
2493 int j;
2494 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2495 rtx_insn *head = BB_HEAD (merge_bb);
2497 while (!NONDEBUG_INSN_P (head))
2498 head = NEXT_INSN (head);
2499 headptr[ix] = head;
2500 currptr[ix] = head;
2502 /* Compute the end point and live information */
2503 for (j = 1; j < max_match; j++)
2505 head = NEXT_INSN (head);
2506 while (!NONDEBUG_INSN_P (head));
2507 simulate_backwards_to_point (merge_bb, live, head);
2508 IOR_REG_SET (live_union, live);
2511 /* If we're moving across two blocks, verify the validity of the
2512 first move, then adjust the target and let the loop below deal
2513 with the final move. */
2514 if (final_dest_bb != NULL)
2516 rtx_insn *move_upto;
2518 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2519 jump, e0->dest, live_union,
2520 NULL, &move_upto);
2521 if (!moveall)
2523 if (move_upto == NULL_RTX)
2524 goto out;
2526 while (e0_last_head != move_upto)
2528 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2529 live_union);
2530 e0_last_head = PREV_INSN (e0_last_head);
2533 if (e0_last_head == NULL_RTX)
2534 goto out;
2536 jump = BB_END (final_dest_bb);
2537 cond = get_condition (jump, &move_before, true, false);
2538 if (cond == NULL_RTX)
2540 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2541 move_before = prev_nonnote_nondebug_insn (jump);
2542 else
2543 move_before = jump;
2549 rtx_insn *move_upto;
2550 moveall = can_move_insns_across (currptr[0], e0_last_head,
2551 move_before, jump, e0->dest, live_union,
2552 NULL, &move_upto);
2553 if (!moveall && move_upto == NULL_RTX)
2555 if (jump == move_before)
2556 break;
2558 /* Try again, using a different insertion point. */
2559 move_before = jump;
2561 /* Don't try moving before a cc0 user, as that may invalidate
2562 the cc0. */
2563 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2564 break;
2566 continue;
2569 if (final_dest_bb && !moveall)
2570 /* We haven't checked whether a partial move would be OK for the first
2571 move, so we have to fail this case. */
2572 break;
2574 changed = true;
2575 for (;;)
2577 if (currptr[0] == move_upto)
2578 break;
2579 for (ix = 0; ix < nedges; ix++)
2581 rtx_insn *curr = currptr[ix];
2583 curr = NEXT_INSN (curr);
2584 while (!NONDEBUG_INSN_P (curr));
2585 currptr[ix] = curr;
2589 /* If we can't currently move all of the identical insns, remember
2590 each insn after the range that we'll merge. */
2591 if (!moveall)
2592 for (ix = 0; ix < nedges; ix++)
2594 rtx_insn *curr = currptr[ix];
2596 curr = NEXT_INSN (curr);
2597 while (!NONDEBUG_INSN_P (curr));
2598 nextptr[ix] = curr;
2601 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2602 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2603 if (final_dest_bb != NULL)
2604 df_set_bb_dirty (final_dest_bb);
2605 df_set_bb_dirty (bb);
2606 for (ix = 1; ix < nedges; ix++)
2608 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2609 delete_insn_chain (headptr[ix], currptr[ix], false);
2611 if (!moveall)
2613 if (jump == move_before)
2614 break;
2616 /* For the unmerged insns, try a different insertion point. */
2617 move_before = jump;
2619 /* Don't try moving before a cc0 user, as that may invalidate
2620 the cc0. */
2621 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2622 break;
2624 for (ix = 0; ix < nedges; ix++)
2625 currptr[ix] = headptr[ix] = nextptr[ix];
2628 while (!moveall);
2630 out:
2631 free (currptr);
2632 free (headptr);
2633 free (nextptr);
2635 crossjumps_occurred |= changed;
2637 return changed;
2640 /* Return true if BB contains just bb note, or bb note followed
2641 by only DEBUG_INSNs. */
2643 static bool
2644 trivially_empty_bb_p (basic_block bb)
2646 rtx_insn *insn = BB_END (bb);
2648 while (1)
2650 if (insn == BB_HEAD (bb))
2651 return true;
2652 if (!DEBUG_INSN_P (insn))
2653 return false;
2654 insn = PREV_INSN (insn);
2658 /* Return true if BB contains just a return and possibly a USE of the
2659 return value. Fill in *RET and *USE with the return and use insns
2660 if any found, otherwise NULL. All CLOBBERs are ignored. */
2662 static bool
2663 bb_is_just_return (basic_block bb, rtx_insn **ret, rtx_insn **use)
2665 *ret = *use = NULL;
2666 rtx_insn *insn;
2668 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
2669 return false;
2671 FOR_BB_INSNS (bb, insn)
2672 if (NONDEBUG_INSN_P (insn))
2674 rtx pat = PATTERN (insn);
2676 if (!*ret && ANY_RETURN_P (pat))
2677 *ret = insn;
2678 else if (!*ret && !*use && GET_CODE (pat) == USE
2679 && REG_P (XEXP (pat, 0))
2680 && REG_FUNCTION_VALUE_P (XEXP (pat, 0)))
2681 *use = insn;
2682 else if (GET_CODE (pat) != CLOBBER)
2683 return false;
2686 return !!*ret;
2689 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2690 instructions etc. Return nonzero if changes were made. */
2692 static bool
2693 try_optimize_cfg (int mode)
2695 bool changed_overall = false;
2696 bool changed;
2697 int iterations = 0;
2698 basic_block bb, b, next;
2700 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2701 clear_bb_flags ();
2703 crossjumps_occurred = false;
2705 FOR_EACH_BB_FN (bb, cfun)
2706 update_forwarder_flag (bb);
2708 if (! targetm.cannot_modify_jumps_p ())
2710 first_pass = true;
2711 /* Attempt to merge blocks as made possible by edge removal. If
2712 a block has only one successor, and the successor has only
2713 one predecessor, they may be combined. */
2716 block_was_dirty = false;
2717 changed = false;
2718 iterations++;
2720 if (dump_file)
2721 fprintf (dump_file,
2722 "\n\ntry_optimize_cfg iteration %i\n\n",
2723 iterations);
2725 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2726 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2728 basic_block c;
2729 edge s;
2730 bool changed_here = false;
2732 /* Delete trivially dead basic blocks. This is either
2733 blocks with no predecessors, or empty blocks with no
2734 successors. However if the empty block with no
2735 successors is the successor of the ENTRY_BLOCK, it is
2736 kept. This ensures that the ENTRY_BLOCK will have a
2737 successor which is a precondition for many RTL
2738 passes. Empty blocks may result from expanding
2739 __builtin_unreachable (). */
2740 if (EDGE_COUNT (b->preds) == 0
2741 || (EDGE_COUNT (b->succs) == 0
2742 && trivially_empty_bb_p (b)
2743 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2744 != b))
2746 c = b->prev_bb;
2747 if (EDGE_COUNT (b->preds) > 0)
2749 edge e;
2750 edge_iterator ei;
2752 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2754 if (BB_FOOTER (b)
2755 && BARRIER_P (BB_FOOTER (b)))
2756 FOR_EACH_EDGE (e, ei, b->preds)
2757 if ((e->flags & EDGE_FALLTHRU)
2758 && BB_FOOTER (e->src) == NULL)
2760 if (BB_FOOTER (b))
2762 BB_FOOTER (e->src) = BB_FOOTER (b);
2763 BB_FOOTER (b) = NULL;
2765 else
2767 start_sequence ();
2768 BB_FOOTER (e->src) = emit_barrier ();
2769 end_sequence ();
2773 else
2775 rtx_insn *last = get_last_bb_insn (b);
2776 if (last && BARRIER_P (last))
2777 FOR_EACH_EDGE (e, ei, b->preds)
2778 if ((e->flags & EDGE_FALLTHRU))
2779 emit_barrier_after (BB_END (e->src));
2782 delete_basic_block (b);
2783 changed = true;
2784 /* Avoid trying to remove the exit block. */
2785 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2786 continue;
2789 /* Remove code labels no longer used. */
2790 if (single_pred_p (b)
2791 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2792 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2793 && LABEL_P (BB_HEAD (b))
2794 && !LABEL_PRESERVE_P (BB_HEAD (b))
2795 /* If the previous block ends with a branch to this
2796 block, we can't delete the label. Normally this
2797 is a condjump that is yet to be simplified, but
2798 if CASE_DROPS_THRU, this can be a tablejump with
2799 some element going to the same place as the
2800 default (fallthru). */
2801 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2802 || !JUMP_P (BB_END (single_pred (b)))
2803 || ! label_is_jump_target_p (BB_HEAD (b),
2804 BB_END (single_pred (b)))))
2806 delete_insn (BB_HEAD (b));
2807 if (dump_file)
2808 fprintf (dump_file, "Deleted label in block %i.\n",
2809 b->index);
2812 /* If we fall through an empty block, we can remove it. */
2813 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2814 && single_pred_p (b)
2815 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2816 && !LABEL_P (BB_HEAD (b))
2817 && FORWARDER_BLOCK_P (b)
2818 /* Note that forwarder_block_p true ensures that
2819 there is a successor for this block. */
2820 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2821 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2823 if (dump_file)
2824 fprintf (dump_file,
2825 "Deleting fallthru block %i.\n",
2826 b->index);
2828 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2829 ? b->next_bb : b->prev_bb);
2830 redirect_edge_succ_nodup (single_pred_edge (b),
2831 single_succ (b));
2832 delete_basic_block (b);
2833 changed = true;
2834 b = c;
2835 continue;
2838 /* Merge B with its single successor, if any. */
2839 if (single_succ_p (b)
2840 && (s = single_succ_edge (b))
2841 && !(s->flags & EDGE_COMPLEX)
2842 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2843 && single_pred_p (c)
2844 && b != c)
2846 /* When not in cfg_layout mode use code aware of reordering
2847 INSN. This code possibly creates new basic blocks so it
2848 does not fit merge_blocks interface and is kept here in
2849 hope that it will become useless once more of compiler
2850 is transformed to use cfg_layout mode. */
2852 if ((mode & CLEANUP_CFGLAYOUT)
2853 && can_merge_blocks_p (b, c))
2855 merge_blocks (b, c);
2856 update_forwarder_flag (b);
2857 changed_here = true;
2859 else if (!(mode & CLEANUP_CFGLAYOUT)
2860 /* If the jump insn has side effects,
2861 we can't kill the edge. */
2862 && (!JUMP_P (BB_END (b))
2863 || (reload_completed
2864 ? simplejump_p (BB_END (b))
2865 : (onlyjump_p (BB_END (b))
2866 && !tablejump_p (BB_END (b),
2867 NULL, NULL))))
2868 && (next = merge_blocks_move (s, b, c, mode)))
2870 b = next;
2871 changed_here = true;
2875 /* Try to change a branch to a return to just that return. */
2876 rtx_insn *ret, *use;
2877 if (single_succ_p (b)
2878 && onlyjump_p (BB_END (b))
2879 && bb_is_just_return (single_succ (b), &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 jump %d->%d to return.\n",
2889 b->index, single_succ (b)->index);
2890 redirect_edge_succ (single_succ_edge (b),
2891 EXIT_BLOCK_PTR_FOR_FN (cfun));
2892 single_succ_edge (b)->flags &= ~EDGE_CROSSING;
2893 changed_here = true;
2897 /* Try to change a conditional branch to a return to the
2898 respective conditional return. */
2899 if (EDGE_COUNT (b->succs) == 2
2900 && any_condjump_p (BB_END (b))
2901 && bb_is_just_return (BRANCH_EDGE (b)->dest, &ret, &use))
2903 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2904 PATTERN (ret), 0))
2906 if (use)
2907 emit_insn_before (copy_insn (PATTERN (use)),
2908 BB_END (b));
2909 if (dump_file)
2910 fprintf (dump_file, "Changed conditional jump %d->%d "
2911 "to conditional return.\n",
2912 b->index, BRANCH_EDGE (b)->dest->index);
2913 redirect_edge_succ (BRANCH_EDGE (b),
2914 EXIT_BLOCK_PTR_FOR_FN (cfun));
2915 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2916 changed_here = true;
2920 /* Try to flip a conditional branch that falls through to
2921 a return so that it becomes a conditional return and a
2922 new jump to the original branch target. */
2923 if (EDGE_COUNT (b->succs) == 2
2924 && BRANCH_EDGE (b)->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
2925 && any_condjump_p (BB_END (b))
2926 && bb_is_just_return (FALLTHRU_EDGE (b)->dest, &ret, &use))
2928 if (invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2929 JUMP_LABEL (BB_END (b)), 0))
2931 basic_block new_ft = BRANCH_EDGE (b)->dest;
2932 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2933 PATTERN (ret), 0))
2935 if (use)
2936 emit_insn_before (copy_insn (PATTERN (use)),
2937 BB_END (b));
2938 if (dump_file)
2939 fprintf (dump_file, "Changed conditional jump "
2940 "%d->%d to conditional return, adding "
2941 "fall-through jump.\n",
2942 b->index, BRANCH_EDGE (b)->dest->index);
2943 redirect_edge_succ (BRANCH_EDGE (b),
2944 EXIT_BLOCK_PTR_FOR_FN (cfun));
2945 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2946 std::swap (BRANCH_EDGE (b)->probability,
2947 FALLTHRU_EDGE (b)->probability);
2948 update_br_prob_note (b);
2949 basic_block jb = force_nonfallthru (FALLTHRU_EDGE (b));
2950 notice_new_block (jb);
2951 if (!redirect_jump (as_a <rtx_jump_insn *> (BB_END (jb)),
2952 block_label (new_ft), 0))
2953 gcc_unreachable ();
2954 redirect_edge_succ (single_succ_edge (jb), new_ft);
2955 changed_here = true;
2957 else
2959 /* Invert the jump back to what it was. This should
2960 never fail. */
2961 if (!invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2962 JUMP_LABEL (BB_END (b)), 0))
2963 gcc_unreachable ();
2968 /* Simplify branch over branch. */
2969 if ((mode & CLEANUP_EXPENSIVE)
2970 && !(mode & CLEANUP_CFGLAYOUT)
2971 && try_simplify_condjump (b))
2972 changed_here = true;
2974 /* If B has a single outgoing edge, but uses a
2975 non-trivial jump instruction without side-effects, we
2976 can either delete the jump entirely, or replace it
2977 with a simple unconditional jump. */
2978 if (single_succ_p (b)
2979 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2980 && onlyjump_p (BB_END (b))
2981 && !CROSSING_JUMP_P (BB_END (b))
2982 && try_redirect_by_replacing_jump (single_succ_edge (b),
2983 single_succ (b),
2984 (mode & CLEANUP_CFGLAYOUT) != 0))
2986 update_forwarder_flag (b);
2987 changed_here = true;
2990 /* Simplify branch to branch. */
2991 if (try_forward_edges (mode, b))
2993 update_forwarder_flag (b);
2994 changed_here = true;
2997 /* Look for shared code between blocks. */
2998 if ((mode & CLEANUP_CROSSJUMP)
2999 && try_crossjump_bb (mode, b))
3000 changed_here = true;
3002 if ((mode & CLEANUP_CROSSJUMP)
3003 /* This can lengthen register lifetimes. Do it only after
3004 reload. */
3005 && reload_completed
3006 && try_head_merge_bb (b))
3007 changed_here = true;
3009 /* Don't get confused by the index shift caused by
3010 deleting blocks. */
3011 if (!changed_here)
3012 b = b->next_bb;
3013 else
3014 changed = true;
3017 if ((mode & CLEANUP_CROSSJUMP)
3018 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
3019 changed = true;
3021 if (block_was_dirty)
3023 /* This should only be set by head-merging. */
3024 gcc_assert (mode & CLEANUP_CROSSJUMP);
3025 df_analyze ();
3028 if (changed)
3030 /* Edge forwarding in particular can cause hot blocks previously
3031 reached by both hot and cold blocks to become dominated only
3032 by cold blocks. This will cause the verification below to fail,
3033 and lead to now cold code in the hot section. This is not easy
3034 to detect and fix during edge forwarding, and in some cases
3035 is only visible after newly unreachable blocks are deleted,
3036 which will be done in fixup_partitions. */
3037 fixup_partitions ();
3038 checking_verify_flow_info ();
3041 changed_overall |= changed;
3042 first_pass = false;
3044 while (changed);
3047 FOR_ALL_BB_FN (b, cfun)
3048 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
3050 return changed_overall;
3053 /* Delete all unreachable basic blocks. */
3055 bool
3056 delete_unreachable_blocks (void)
3058 bool changed = false;
3059 basic_block b, prev_bb;
3061 find_unreachable_blocks ();
3063 /* When we're in GIMPLE mode and there may be debug insns, we should
3064 delete blocks in reverse dominator order, so as to get a chance
3065 to substitute all released DEFs into debug stmts. If we don't
3066 have dominators information, walking blocks backward gets us a
3067 better chance of retaining most debug information than
3068 otherwise. */
3069 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
3070 && dom_info_available_p (CDI_DOMINATORS))
3072 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3073 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3075 prev_bb = b->prev_bb;
3077 if (!(b->flags & BB_REACHABLE))
3079 /* Speed up the removal of blocks that don't dominate
3080 others. Walking backwards, this should be the common
3081 case. */
3082 if (!first_dom_son (CDI_DOMINATORS, b))
3083 delete_basic_block (b);
3084 else
3086 vec<basic_block> h
3087 = get_all_dominated_blocks (CDI_DOMINATORS, b);
3089 while (h.length ())
3091 b = h.pop ();
3093 prev_bb = b->prev_bb;
3095 gcc_assert (!(b->flags & BB_REACHABLE));
3097 delete_basic_block (b);
3100 h.release ();
3103 changed = true;
3107 else
3109 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3110 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3112 prev_bb = b->prev_bb;
3114 if (!(b->flags & BB_REACHABLE))
3116 delete_basic_block (b);
3117 changed = true;
3122 if (changed)
3123 tidy_fallthru_edges ();
3124 return changed;
3127 /* Delete any jump tables never referenced. We can't delete them at the
3128 time of removing tablejump insn as they are referenced by the preceding
3129 insns computing the destination, so we delay deleting and garbagecollect
3130 them once life information is computed. */
3131 void
3132 delete_dead_jumptables (void)
3134 basic_block bb;
3136 /* A dead jump table does not belong to any basic block. Scan insns
3137 between two adjacent basic blocks. */
3138 FOR_EACH_BB_FN (bb, cfun)
3140 rtx_insn *insn, *next;
3142 for (insn = NEXT_INSN (BB_END (bb));
3143 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
3144 insn = next)
3146 next = NEXT_INSN (insn);
3147 if (LABEL_P (insn)
3148 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
3149 && JUMP_TABLE_DATA_P (next))
3151 rtx_insn *label = insn, *jump = next;
3153 if (dump_file)
3154 fprintf (dump_file, "Dead jumptable %i removed\n",
3155 INSN_UID (insn));
3157 next = NEXT_INSN (next);
3158 delete_insn (jump);
3159 delete_insn (label);
3166 /* Tidy the CFG by deleting unreachable code and whatnot. */
3168 bool
3169 cleanup_cfg (int mode)
3171 bool changed = false;
3173 /* Set the cfglayout mode flag here. We could update all the callers
3174 but that is just inconvenient, especially given that we eventually
3175 want to have cfglayout mode as the default. */
3176 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3177 mode |= CLEANUP_CFGLAYOUT;
3179 timevar_push (TV_CLEANUP_CFG);
3180 if (delete_unreachable_blocks ())
3182 changed = true;
3183 /* We've possibly created trivially dead code. Cleanup it right
3184 now to introduce more opportunities for try_optimize_cfg. */
3185 if (!(mode & (CLEANUP_NO_INSN_DEL))
3186 && !reload_completed)
3187 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3190 compact_blocks ();
3192 /* To tail-merge blocks ending in the same noreturn function (e.g.
3193 a call to abort) we have to insert fake edges to exit. Do this
3194 here once. The fake edges do not interfere with any other CFG
3195 cleanups. */
3196 if (mode & CLEANUP_CROSSJUMP)
3197 add_noreturn_fake_exit_edges ();
3199 if (!dbg_cnt (cfg_cleanup))
3200 return changed;
3202 while (try_optimize_cfg (mode))
3204 delete_unreachable_blocks (), changed = true;
3205 if (!(mode & CLEANUP_NO_INSN_DEL))
3207 /* Try to remove some trivially dead insns when doing an expensive
3208 cleanup. But delete_trivially_dead_insns doesn't work after
3209 reload (it only handles pseudos) and run_fast_dce is too costly
3210 to run in every iteration.
3212 For effective cross jumping, we really want to run a fast DCE to
3213 clean up any dead conditions, or they get in the way of performing
3214 useful tail merges.
3216 Other transformations in cleanup_cfg are not so sensitive to dead
3217 code, so delete_trivially_dead_insns or even doing nothing at all
3218 is good enough. */
3219 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3220 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3221 break;
3222 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occurred)
3223 run_fast_dce ();
3225 else
3226 break;
3229 if (mode & CLEANUP_CROSSJUMP)
3230 remove_fake_exit_edges ();
3232 /* Don't call delete_dead_jumptables in cfglayout mode, because
3233 that function assumes that jump tables are in the insns stream.
3234 But we also don't _have_ to delete dead jumptables in cfglayout
3235 mode because we shouldn't even be looking at things that are
3236 not in a basic block. Dead jumptables are cleaned up when
3237 going out of cfglayout mode. */
3238 if (!(mode & CLEANUP_CFGLAYOUT))
3239 delete_dead_jumptables ();
3241 /* ??? We probably do this way too often. */
3242 if (current_loops
3243 && (changed
3244 || (mode & CLEANUP_CFG_CHANGED)))
3246 timevar_push (TV_REPAIR_LOOPS);
3247 /* The above doesn't preserve dominance info if available. */
3248 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3249 calculate_dominance_info (CDI_DOMINATORS);
3250 fix_loop_structure (NULL);
3251 free_dominance_info (CDI_DOMINATORS);
3252 timevar_pop (TV_REPAIR_LOOPS);
3255 timevar_pop (TV_CLEANUP_CFG);
3257 return changed;
3260 namespace {
3262 const pass_data pass_data_jump =
3264 RTL_PASS, /* type */
3265 "jump", /* name */
3266 OPTGROUP_NONE, /* optinfo_flags */
3267 TV_JUMP, /* tv_id */
3268 0, /* properties_required */
3269 0, /* properties_provided */
3270 0, /* properties_destroyed */
3271 0, /* todo_flags_start */
3272 0, /* todo_flags_finish */
3275 class pass_jump : public rtl_opt_pass
3277 public:
3278 pass_jump (gcc::context *ctxt)
3279 : rtl_opt_pass (pass_data_jump, ctxt)
3282 /* opt_pass methods: */
3283 virtual unsigned int execute (function *);
3285 }; // class pass_jump
3287 unsigned int
3288 pass_jump::execute (function *)
3290 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3291 if (dump_file)
3292 dump_flow_info (dump_file, dump_flags);
3293 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3294 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3295 return 0;
3298 } // anon namespace
3300 rtl_opt_pass *
3301 make_pass_jump (gcc::context *ctxt)
3303 return new pass_jump (ctxt);
3306 namespace {
3308 const pass_data pass_data_jump2 =
3310 RTL_PASS, /* type */
3311 "jump2", /* name */
3312 OPTGROUP_NONE, /* optinfo_flags */
3313 TV_JUMP, /* tv_id */
3314 0, /* properties_required */
3315 0, /* properties_provided */
3316 0, /* properties_destroyed */
3317 0, /* todo_flags_start */
3318 0, /* todo_flags_finish */
3321 class pass_jump2 : public rtl_opt_pass
3323 public:
3324 pass_jump2 (gcc::context *ctxt)
3325 : rtl_opt_pass (pass_data_jump2, ctxt)
3328 /* opt_pass methods: */
3329 virtual unsigned int execute (function *)
3331 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3332 return 0;
3335 }; // class pass_jump2
3337 } // anon namespace
3339 rtl_opt_pass *
3340 make_pass_jump2 (gcc::context *ctxt)
3342 return new pass_jump2 (ctxt);