Don't treat zero-sized ranges as overlapping
[official-gcc.git] / gcc / cfgcleanup.c
blobc2b0434f179bde595dea280f9dece9315caeda86
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 first = t->dest;
621 while (first != target);
623 changed = true;
624 continue;
626 ei_next (&ei);
629 free (threaded_edges);
630 return changed;
634 /* Blocks A and B are to be merged into a single block. A has no incoming
635 fallthru edge, so it can be moved before B without adding or modifying
636 any jumps (aside from the jump from A to B). */
638 static void
639 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
641 rtx_insn *barrier;
643 /* If we are partitioning hot/cold basic blocks, we don't want to
644 mess up unconditional or indirect jumps that cross between hot
645 and cold sections.
647 Basic block partitioning may result in some jumps that appear to
648 be optimizable (or blocks that appear to be mergeable), but which really
649 must be left untouched (they are required to make it safely across
650 partition boundaries). See the comments at the top of
651 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
653 if (BB_PARTITION (a) != BB_PARTITION (b))
654 return;
656 barrier = next_nonnote_insn (BB_END (a));
657 gcc_assert (BARRIER_P (barrier));
658 delete_insn (barrier);
660 /* Scramble the insn chain. */
661 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
662 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
663 df_set_bb_dirty (a);
665 if (dump_file)
666 fprintf (dump_file, "Moved block %d before %d and merged.\n",
667 a->index, b->index);
669 /* Swap the records for the two blocks around. */
671 unlink_block (a);
672 link_block (a, b->prev_bb);
674 /* Now blocks A and B are contiguous. Merge them. */
675 merge_blocks (a, b);
678 /* Blocks A and B are to be merged into a single block. B has no outgoing
679 fallthru edge, so it can be moved after A without adding or modifying
680 any jumps (aside from the jump from A to B). */
682 static void
683 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
685 rtx_insn *barrier, *real_b_end;
686 rtx_insn *label;
687 rtx_jump_table_data *table;
689 /* If we are partitioning hot/cold basic blocks, we don't want to
690 mess up unconditional or indirect jumps that cross between hot
691 and cold sections.
693 Basic block partitioning may result in some jumps that appear to
694 be optimizable (or blocks that appear to be mergeable), but which really
695 must be left untouched (they are required to make it safely across
696 partition boundaries). See the comments at the top of
697 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
699 if (BB_PARTITION (a) != BB_PARTITION (b))
700 return;
702 real_b_end = BB_END (b);
704 /* If there is a jump table following block B temporarily add the jump table
705 to block B so that it will also be moved to the correct location. */
706 if (tablejump_p (BB_END (b), &label, &table)
707 && prev_active_insn (label) == BB_END (b))
709 BB_END (b) = table;
712 /* There had better have been a barrier there. Delete it. */
713 barrier = NEXT_INSN (BB_END (b));
714 if (barrier && BARRIER_P (barrier))
715 delete_insn (barrier);
718 /* Scramble the insn chain. */
719 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
721 /* Restore the real end of b. */
722 BB_END (b) = real_b_end;
724 if (dump_file)
725 fprintf (dump_file, "Moved block %d after %d and merged.\n",
726 b->index, a->index);
728 /* Now blocks A and B are contiguous. Merge them. */
729 merge_blocks (a, b);
732 /* Attempt to merge basic blocks that are potentially non-adjacent.
733 Return NULL iff the attempt failed, otherwise return basic block
734 where cleanup_cfg should continue. Because the merging commonly
735 moves basic block away or introduces another optimization
736 possibility, return basic block just before B so cleanup_cfg don't
737 need to iterate.
739 It may be good idea to return basic block before C in the case
740 C has been moved after B and originally appeared earlier in the
741 insn sequence, but we have no information available about the
742 relative ordering of these two. Hopefully it is not too common. */
744 static basic_block
745 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
747 basic_block next;
749 /* If we are partitioning hot/cold basic blocks, we don't want to
750 mess up unconditional or indirect jumps that cross between hot
751 and cold sections.
753 Basic block partitioning may result in some jumps that appear to
754 be optimizable (or blocks that appear to be mergeable), but which really
755 must be left untouched (they are required to make it safely across
756 partition boundaries). See the comments at the top of
757 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
759 if (BB_PARTITION (b) != BB_PARTITION (c))
760 return NULL;
762 /* If B has a fallthru edge to C, no need to move anything. */
763 if (e->flags & EDGE_FALLTHRU)
765 int b_index = b->index, c_index = c->index;
767 /* Protect the loop latches. */
768 if (current_loops && c->loop_father->latch == c)
769 return NULL;
771 merge_blocks (b, c);
772 update_forwarder_flag (b);
774 if (dump_file)
775 fprintf (dump_file, "Merged %d and %d without moving.\n",
776 b_index, c_index);
778 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
781 /* Otherwise we will need to move code around. Do that only if expensive
782 transformations are allowed. */
783 else if (mode & CLEANUP_EXPENSIVE)
785 edge tmp_edge, b_fallthru_edge;
786 bool c_has_outgoing_fallthru;
787 bool b_has_incoming_fallthru;
789 /* Avoid overactive code motion, as the forwarder blocks should be
790 eliminated by edge redirection instead. One exception might have
791 been if B is a forwarder block and C has no fallthru edge, but
792 that should be cleaned up by bb-reorder instead. */
793 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
794 return NULL;
796 /* We must make sure to not munge nesting of lexical blocks,
797 and loop notes. This is done by squeezing out all the notes
798 and leaving them there to lie. Not ideal, but functional. */
800 tmp_edge = find_fallthru_edge (c->succs);
801 c_has_outgoing_fallthru = (tmp_edge != NULL);
803 tmp_edge = find_fallthru_edge (b->preds);
804 b_has_incoming_fallthru = (tmp_edge != NULL);
805 b_fallthru_edge = tmp_edge;
806 next = b->prev_bb;
807 if (next == c)
808 next = next->prev_bb;
810 /* Otherwise, we're going to try to move C after B. If C does
811 not have an outgoing fallthru, then it can be moved
812 immediately after B without introducing or modifying jumps. */
813 if (! c_has_outgoing_fallthru)
815 merge_blocks_move_successor_nojumps (b, c);
816 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
819 /* If B does not have an incoming fallthru, then it can be moved
820 immediately before C without introducing or modifying jumps.
821 C cannot be the first block, so we do not have to worry about
822 accessing a non-existent block. */
824 if (b_has_incoming_fallthru)
826 basic_block bb;
828 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
829 return NULL;
830 bb = force_nonfallthru (b_fallthru_edge);
831 if (bb)
832 notice_new_block (bb);
835 merge_blocks_move_predecessor_nojumps (b, c);
836 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
839 return NULL;
843 /* Removes the memory attributes of MEM expression
844 if they are not equal. */
846 static void
847 merge_memattrs (rtx x, rtx y)
849 int i;
850 int j;
851 enum rtx_code code;
852 const char *fmt;
854 if (x == y)
855 return;
856 if (x == 0 || y == 0)
857 return;
859 code = GET_CODE (x);
861 if (code != GET_CODE (y))
862 return;
864 if (GET_MODE (x) != GET_MODE (y))
865 return;
867 if (code == MEM && !mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
869 if (! MEM_ATTRS (x))
870 MEM_ATTRS (y) = 0;
871 else if (! MEM_ATTRS (y))
872 MEM_ATTRS (x) = 0;
873 else
875 HOST_WIDE_INT mem_size;
877 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
879 set_mem_alias_set (x, 0);
880 set_mem_alias_set (y, 0);
883 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
885 set_mem_expr (x, 0);
886 set_mem_expr (y, 0);
887 clear_mem_offset (x);
888 clear_mem_offset (y);
890 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
891 || (MEM_OFFSET_KNOWN_P (x)
892 && MEM_OFFSET (x) != MEM_OFFSET (y)))
894 clear_mem_offset (x);
895 clear_mem_offset (y);
898 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
900 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
901 set_mem_size (x, mem_size);
902 set_mem_size (y, mem_size);
904 else
906 clear_mem_size (x);
907 clear_mem_size (y);
910 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
911 set_mem_align (y, MEM_ALIGN (x));
914 if (code == MEM)
916 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
918 MEM_READONLY_P (x) = 0;
919 MEM_READONLY_P (y) = 0;
921 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
923 MEM_NOTRAP_P (x) = 0;
924 MEM_NOTRAP_P (y) = 0;
926 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
928 MEM_VOLATILE_P (x) = 1;
929 MEM_VOLATILE_P (y) = 1;
933 fmt = GET_RTX_FORMAT (code);
934 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
936 switch (fmt[i])
938 case 'E':
939 /* Two vectors must have the same length. */
940 if (XVECLEN (x, i) != XVECLEN (y, i))
941 return;
943 for (j = 0; j < XVECLEN (x, i); j++)
944 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
946 break;
948 case 'e':
949 merge_memattrs (XEXP (x, i), XEXP (y, i));
952 return;
956 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
957 different single sets S1 and S2. */
959 static bool
960 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
962 int i;
963 rtx e1, e2;
965 if (p1 == s1 && p2 == s2)
966 return true;
968 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
969 return false;
971 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
972 return false;
974 for (i = 0; i < XVECLEN (p1, 0); i++)
976 e1 = XVECEXP (p1, 0, i);
977 e2 = XVECEXP (p2, 0, i);
978 if (e1 == s1 && e2 == s2)
979 continue;
980 if (reload_completed
981 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
982 continue;
984 return false;
987 return true;
991 /* NOTE1 is the REG_EQUAL note, if any, attached to an insn
992 that is a single_set with a SET_SRC of SRC1. Similarly
993 for NOTE2/SRC2.
995 So effectively NOTE1/NOTE2 are an alternate form of
996 SRC1/SRC2 respectively.
998 Return nonzero if SRC1 or NOTE1 has the same constant
999 integer value as SRC2 or NOTE2. Else return zero. */
1000 static int
1001 values_equal_p (rtx note1, rtx note2, rtx src1, rtx src2)
1003 if (note1
1004 && note2
1005 && CONST_INT_P (XEXP (note1, 0))
1006 && rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0)))
1007 return 1;
1009 if (!note1
1010 && !note2
1011 && CONST_INT_P (src1)
1012 && CONST_INT_P (src2)
1013 && rtx_equal_p (src1, src2))
1014 return 1;
1016 if (note1
1017 && CONST_INT_P (src2)
1018 && rtx_equal_p (XEXP (note1, 0), src2))
1019 return 1;
1021 if (note2
1022 && CONST_INT_P (src1)
1023 && rtx_equal_p (XEXP (note2, 0), src1))
1024 return 1;
1026 return 0;
1029 /* Examine register notes on I1 and I2 and return:
1030 - dir_forward if I1 can be replaced by I2, or
1031 - dir_backward if I2 can be replaced by I1, or
1032 - dir_both if both are the case. */
1034 static enum replace_direction
1035 can_replace_by (rtx_insn *i1, rtx_insn *i2)
1037 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1038 bool c1, c2;
1040 /* Check for 2 sets. */
1041 s1 = single_set (i1);
1042 s2 = single_set (i2);
1043 if (s1 == NULL_RTX || s2 == NULL_RTX)
1044 return dir_none;
1046 /* Check that the 2 sets set the same dest. */
1047 d1 = SET_DEST (s1);
1048 d2 = SET_DEST (s2);
1049 if (!(reload_completed
1050 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1051 return dir_none;
1053 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1054 set dest to the same value. */
1055 note1 = find_reg_equal_equiv_note (i1);
1056 note2 = find_reg_equal_equiv_note (i2);
1058 src1 = SET_SRC (s1);
1059 src2 = SET_SRC (s2);
1061 if (!values_equal_p (note1, note2, src1, src2))
1062 return dir_none;
1064 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1065 return dir_none;
1067 /* Although the 2 sets set dest to the same value, we cannot replace
1068 (set (dest) (const_int))
1070 (set (dest) (reg))
1071 because we don't know if the reg is live and has the same value at the
1072 location of replacement. */
1073 c1 = CONST_INT_P (src1);
1074 c2 = CONST_INT_P (src2);
1075 if (c1 && c2)
1076 return dir_both;
1077 else if (c2)
1078 return dir_forward;
1079 else if (c1)
1080 return dir_backward;
1082 return dir_none;
1085 /* Merges directions A and B. */
1087 static enum replace_direction
1088 merge_dir (enum replace_direction a, enum replace_direction b)
1090 /* Implements the following table:
1091 |bo fw bw no
1092 ---+-----------
1093 bo |bo fw bw no
1094 fw |-- fw no no
1095 bw |-- -- bw no
1096 no |-- -- -- no. */
1098 if (a == b)
1099 return a;
1101 if (a == dir_both)
1102 return b;
1103 if (b == dir_both)
1104 return a;
1106 return dir_none;
1109 /* Array of flags indexed by reg note kind, true if the given
1110 reg note is CFA related. */
1111 static const bool reg_note_cfa_p[] = {
1112 #undef REG_CFA_NOTE
1113 #define DEF_REG_NOTE(NAME) false,
1114 #define REG_CFA_NOTE(NAME) true,
1115 #include "reg-notes.def"
1116 #undef REG_CFA_NOTE
1117 #undef DEF_REG_NOTE
1118 false
1121 /* Return true if I1 and I2 have identical CFA notes (the same order
1122 and equivalent content). */
1124 static bool
1125 insns_have_identical_cfa_notes (rtx_insn *i1, rtx_insn *i2)
1127 rtx n1, n2;
1128 for (n1 = REG_NOTES (i1), n2 = REG_NOTES (i2); ;
1129 n1 = XEXP (n1, 1), n2 = XEXP (n2, 1))
1131 /* Skip over reg notes not related to CFI information. */
1132 while (n1 && !reg_note_cfa_p[REG_NOTE_KIND (n1)])
1133 n1 = XEXP (n1, 1);
1134 while (n2 && !reg_note_cfa_p[REG_NOTE_KIND (n2)])
1135 n2 = XEXP (n2, 1);
1136 if (n1 == NULL_RTX && n2 == NULL_RTX)
1137 return true;
1138 if (n1 == NULL_RTX || n2 == NULL_RTX)
1139 return false;
1140 if (XEXP (n1, 0) == XEXP (n2, 0))
1142 else if (XEXP (n1, 0) == NULL_RTX || XEXP (n2, 0) == NULL_RTX)
1143 return false;
1144 else if (!(reload_completed
1145 ? rtx_renumbered_equal_p (XEXP (n1, 0), XEXP (n2, 0))
1146 : rtx_equal_p (XEXP (n1, 0), XEXP (n2, 0))))
1147 return false;
1151 /* Examine I1 and I2 and return:
1152 - dir_forward if I1 can be replaced by I2, or
1153 - dir_backward if I2 can be replaced by I1, or
1154 - dir_both if both are the case. */
1156 static enum replace_direction
1157 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1159 rtx p1, p2;
1161 /* Verify that I1 and I2 are equivalent. */
1162 if (GET_CODE (i1) != GET_CODE (i2))
1163 return dir_none;
1165 /* __builtin_unreachable() may lead to empty blocks (ending with
1166 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1167 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1168 return dir_both;
1170 /* ??? Do not allow cross-jumping between different stack levels. */
1171 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1172 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1173 if (p1 && p2)
1175 p1 = XEXP (p1, 0);
1176 p2 = XEXP (p2, 0);
1177 if (!rtx_equal_p (p1, p2))
1178 return dir_none;
1180 /* ??? Worse, this adjustment had better be constant lest we
1181 have differing incoming stack levels. */
1182 if (!frame_pointer_needed
1183 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1184 return dir_none;
1186 else if (p1 || p2)
1187 return dir_none;
1189 /* Do not allow cross-jumping between frame related insns and other
1190 insns. */
1191 if (RTX_FRAME_RELATED_P (i1) != RTX_FRAME_RELATED_P (i2))
1192 return dir_none;
1194 p1 = PATTERN (i1);
1195 p2 = PATTERN (i2);
1197 if (GET_CODE (p1) != GET_CODE (p2))
1198 return dir_none;
1200 /* If this is a CALL_INSN, compare register usage information.
1201 If we don't check this on stack register machines, the two
1202 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1203 numbers of stack registers in the same basic block.
1204 If we don't check this on machines with delay slots, a delay slot may
1205 be filled that clobbers a parameter expected by the subroutine.
1207 ??? We take the simple route for now and assume that if they're
1208 equal, they were constructed identically.
1210 Also check for identical exception regions. */
1212 if (CALL_P (i1))
1214 /* Ensure the same EH region. */
1215 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1216 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1218 if (!n1 && n2)
1219 return dir_none;
1221 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1222 return dir_none;
1224 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1225 CALL_INSN_FUNCTION_USAGE (i2))
1226 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1227 return dir_none;
1229 /* For address sanitizer, never crossjump __asan_report_* builtins,
1230 otherwise errors might be reported on incorrect lines. */
1231 if (flag_sanitize & SANITIZE_ADDRESS)
1233 rtx call = get_call_rtx_from (i1);
1234 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1236 rtx symbol = XEXP (XEXP (call, 0), 0);
1237 if (SYMBOL_REF_DECL (symbol)
1238 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1240 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1241 == BUILT_IN_NORMAL)
1242 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1243 >= BUILT_IN_ASAN_REPORT_LOAD1
1244 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1245 <= BUILT_IN_ASAN_STOREN)
1246 return dir_none;
1252 /* If both i1 and i2 are frame related, verify all the CFA notes
1253 in the same order and with the same content. */
1254 if (RTX_FRAME_RELATED_P (i1) && !insns_have_identical_cfa_notes (i1, i2))
1255 return dir_none;
1257 #ifdef STACK_REGS
1258 /* If cross_jump_death_matters is not 0, the insn's mode
1259 indicates whether or not the insn contains any stack-like
1260 regs. */
1262 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1264 /* If register stack conversion has already been done, then
1265 death notes must also be compared before it is certain that
1266 the two instruction streams match. */
1268 rtx note;
1269 HARD_REG_SET i1_regset, i2_regset;
1271 CLEAR_HARD_REG_SET (i1_regset);
1272 CLEAR_HARD_REG_SET (i2_regset);
1274 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1275 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1276 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1278 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1279 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1280 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1282 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1283 return dir_none;
1285 #endif
1287 if (reload_completed
1288 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1289 return dir_both;
1291 return can_replace_by (i1, i2);
1294 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1295 flow_find_head_matching_sequence, ensure the notes match. */
1297 static void
1298 merge_notes (rtx_insn *i1, rtx_insn *i2)
1300 /* If the merged insns have different REG_EQUAL notes, then
1301 remove them. */
1302 rtx equiv1 = find_reg_equal_equiv_note (i1);
1303 rtx equiv2 = find_reg_equal_equiv_note (i2);
1305 if (equiv1 && !equiv2)
1306 remove_note (i1, equiv1);
1307 else if (!equiv1 && equiv2)
1308 remove_note (i2, equiv2);
1309 else if (equiv1 && equiv2
1310 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1312 remove_note (i1, equiv1);
1313 remove_note (i2, equiv2);
1317 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1318 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1319 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1320 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1321 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1323 static void
1324 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1325 bool *did_fallthru)
1327 edge fallthru;
1329 *did_fallthru = false;
1331 /* Ignore notes. */
1332 while (!NONDEBUG_INSN_P (*i1))
1334 if (*i1 != BB_HEAD (*bb1))
1336 *i1 = PREV_INSN (*i1);
1337 continue;
1340 if (!follow_fallthru)
1341 return;
1343 fallthru = find_fallthru_edge ((*bb1)->preds);
1344 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1345 || !single_succ_p (fallthru->src))
1346 return;
1348 *bb1 = fallthru->src;
1349 *i1 = BB_END (*bb1);
1350 *did_fallthru = true;
1354 /* Look through the insns at the end of BB1 and BB2 and find the longest
1355 sequence that are either equivalent, or allow forward or backward
1356 replacement. Store the first insns for that sequence in *F1 and *F2 and
1357 return the sequence length.
1359 DIR_P indicates the allowed replacement direction on function entry, and
1360 the actual replacement direction on function exit. If NULL, only equivalent
1361 sequences are allowed.
1363 To simplify callers of this function, if the blocks match exactly,
1364 store the head of the blocks in *F1 and *F2. */
1367 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1368 rtx_insn **f2, enum replace_direction *dir_p)
1370 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1371 int ninsns = 0;
1372 enum replace_direction dir, last_dir, afterlast_dir;
1373 bool follow_fallthru, did_fallthru;
1375 if (dir_p)
1376 dir = *dir_p;
1377 else
1378 dir = dir_both;
1379 afterlast_dir = dir;
1380 last_dir = afterlast_dir;
1382 /* Skip simple jumps at the end of the blocks. Complex jumps still
1383 need to be compared for equivalence, which we'll do below. */
1385 i1 = BB_END (bb1);
1386 last1 = afterlast1 = last2 = afterlast2 = NULL;
1387 if (onlyjump_p (i1)
1388 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1390 last1 = i1;
1391 i1 = PREV_INSN (i1);
1394 i2 = BB_END (bb2);
1395 if (onlyjump_p (i2)
1396 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1398 last2 = i2;
1399 /* Count everything except for unconditional jump as insn.
1400 Don't count any jumps if dir_p is NULL. */
1401 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1402 ninsns++;
1403 i2 = PREV_INSN (i2);
1406 while (true)
1408 /* In the following example, we can replace all jumps to C by jumps to A.
1410 This removes 4 duplicate insns.
1411 [bb A] insn1 [bb C] insn1
1412 insn2 insn2
1413 [bb B] insn3 insn3
1414 insn4 insn4
1415 jump_insn jump_insn
1417 We could also replace all jumps to A by jumps to C, but that leaves B
1418 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1419 step, all jumps to B would be replaced with jumps to the middle of C,
1420 achieving the same result with more effort.
1421 So we allow only the first possibility, which means that we don't allow
1422 fallthru in the block that's being replaced. */
1424 follow_fallthru = dir_p && dir != dir_forward;
1425 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1426 if (did_fallthru)
1427 dir = dir_backward;
1429 follow_fallthru = dir_p && dir != dir_backward;
1430 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1431 if (did_fallthru)
1432 dir = dir_forward;
1434 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1435 break;
1437 /* Do not turn corssing edge to non-crossing or vice versa after
1438 reload. */
1439 if (BB_PARTITION (BLOCK_FOR_INSN (i1))
1440 != BB_PARTITION (BLOCK_FOR_INSN (i2))
1441 && reload_completed)
1442 break;
1444 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1445 if (dir == dir_none || (!dir_p && dir != dir_both))
1446 break;
1448 merge_memattrs (i1, i2);
1450 /* Don't begin a cross-jump with a NOTE insn. */
1451 if (INSN_P (i1))
1453 merge_notes (i1, i2);
1455 afterlast1 = last1, afterlast2 = last2;
1456 last1 = i1, last2 = i2;
1457 afterlast_dir = last_dir;
1458 last_dir = dir;
1459 if (active_insn_p (i1))
1460 ninsns++;
1463 i1 = PREV_INSN (i1);
1464 i2 = PREV_INSN (i2);
1467 /* Don't allow the insn after a compare to be shared by
1468 cross-jumping unless the compare is also shared. */
1469 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1470 && ! sets_cc0_p (last1))
1471 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1473 /* Include preceding notes and labels in the cross-jump. One,
1474 this may bring us to the head of the blocks as requested above.
1475 Two, it keeps line number notes as matched as may be. */
1476 if (ninsns)
1478 bb1 = BLOCK_FOR_INSN (last1);
1479 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1480 last1 = PREV_INSN (last1);
1482 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1483 last1 = PREV_INSN (last1);
1485 bb2 = BLOCK_FOR_INSN (last2);
1486 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1487 last2 = PREV_INSN (last2);
1489 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1490 last2 = PREV_INSN (last2);
1492 *f1 = last1;
1493 *f2 = last2;
1496 if (dir_p)
1497 *dir_p = last_dir;
1498 return ninsns;
1501 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1502 the head of the two blocks. Do not include jumps at the end.
1503 If STOP_AFTER is nonzero, stop after finding that many matching
1504 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1505 non-zero, only count active insns. */
1508 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1509 rtx_insn **f2, int stop_after)
1511 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1512 int ninsns = 0;
1513 edge e;
1514 edge_iterator ei;
1515 int nehedges1 = 0, nehedges2 = 0;
1517 FOR_EACH_EDGE (e, ei, bb1->succs)
1518 if (e->flags & EDGE_EH)
1519 nehedges1++;
1520 FOR_EACH_EDGE (e, ei, bb2->succs)
1521 if (e->flags & EDGE_EH)
1522 nehedges2++;
1524 i1 = BB_HEAD (bb1);
1525 i2 = BB_HEAD (bb2);
1526 last1 = beforelast1 = last2 = beforelast2 = NULL;
1528 while (true)
1530 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1531 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1533 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1534 break;
1535 i1 = NEXT_INSN (i1);
1538 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1540 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1541 break;
1542 i2 = NEXT_INSN (i2);
1545 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1546 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1547 break;
1549 if (NOTE_P (i1) || NOTE_P (i2)
1550 || JUMP_P (i1) || JUMP_P (i2))
1551 break;
1553 /* A sanity check to make sure we're not merging insns with different
1554 effects on EH. If only one of them ends a basic block, it shouldn't
1555 have an EH edge; if both end a basic block, there should be the same
1556 number of EH edges. */
1557 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1558 && nehedges1 > 0)
1559 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1560 && nehedges2 > 0)
1561 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1562 && nehedges1 != nehedges2))
1563 break;
1565 if (old_insns_match_p (0, i1, i2) != dir_both)
1566 break;
1568 merge_memattrs (i1, i2);
1570 /* Don't begin a cross-jump with a NOTE insn. */
1571 if (INSN_P (i1))
1573 merge_notes (i1, i2);
1575 beforelast1 = last1, beforelast2 = last2;
1576 last1 = i1, last2 = i2;
1577 if (!stop_after || active_insn_p (i1))
1578 ninsns++;
1581 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1582 || (stop_after > 0 && ninsns == stop_after))
1583 break;
1585 i1 = NEXT_INSN (i1);
1586 i2 = NEXT_INSN (i2);
1589 /* Don't allow a compare to be shared by cross-jumping unless the insn
1590 after the compare is also shared. */
1591 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1592 && sets_cc0_p (last1))
1593 last1 = beforelast1, last2 = beforelast2, ninsns--;
1595 if (ninsns)
1597 *f1 = last1;
1598 *f2 = last2;
1601 return ninsns;
1604 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1605 the branch instruction. This means that if we commonize the control
1606 flow before end of the basic block, the semantic remains unchanged.
1608 We may assume that there exists one edge with a common destination. */
1610 static bool
1611 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1613 int nehedges1 = 0, nehedges2 = 0;
1614 edge fallthru1 = 0, fallthru2 = 0;
1615 edge e1, e2;
1616 edge_iterator ei;
1618 /* If we performed shrink-wrapping, edges to the exit block can
1619 only be distinguished for JUMP_INSNs. The two paths may differ in
1620 whether they went through the prologue. Sibcalls are fine, we know
1621 that we either didn't need or inserted an epilogue before them. */
1622 if (crtl->shrink_wrapped
1623 && single_succ_p (bb1)
1624 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1625 && !JUMP_P (BB_END (bb1))
1626 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1627 return false;
1629 /* If BB1 has only one successor, we may be looking at either an
1630 unconditional jump, or a fake edge to exit. */
1631 if (single_succ_p (bb1)
1632 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1633 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1634 return (single_succ_p (bb2)
1635 && (single_succ_edge (bb2)->flags
1636 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1637 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1639 /* Match conditional jumps - this may get tricky when fallthru and branch
1640 edges are crossed. */
1641 if (EDGE_COUNT (bb1->succs) == 2
1642 && any_condjump_p (BB_END (bb1))
1643 && onlyjump_p (BB_END (bb1)))
1645 edge b1, f1, b2, f2;
1646 bool reverse, match;
1647 rtx set1, set2, cond1, cond2;
1648 enum rtx_code code1, code2;
1650 if (EDGE_COUNT (bb2->succs) != 2
1651 || !any_condjump_p (BB_END (bb2))
1652 || !onlyjump_p (BB_END (bb2)))
1653 return false;
1655 b1 = BRANCH_EDGE (bb1);
1656 b2 = BRANCH_EDGE (bb2);
1657 f1 = FALLTHRU_EDGE (bb1);
1658 f2 = FALLTHRU_EDGE (bb2);
1660 /* Get around possible forwarders on fallthru edges. Other cases
1661 should be optimized out already. */
1662 if (FORWARDER_BLOCK_P (f1->dest))
1663 f1 = single_succ_edge (f1->dest);
1665 if (FORWARDER_BLOCK_P (f2->dest))
1666 f2 = single_succ_edge (f2->dest);
1668 /* To simplify use of this function, return false if there are
1669 unneeded forwarder blocks. These will get eliminated later
1670 during cleanup_cfg. */
1671 if (FORWARDER_BLOCK_P (f1->dest)
1672 || FORWARDER_BLOCK_P (f2->dest)
1673 || FORWARDER_BLOCK_P (b1->dest)
1674 || FORWARDER_BLOCK_P (b2->dest))
1675 return false;
1677 if (f1->dest == f2->dest && b1->dest == b2->dest)
1678 reverse = false;
1679 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1680 reverse = true;
1681 else
1682 return false;
1684 set1 = pc_set (BB_END (bb1));
1685 set2 = pc_set (BB_END (bb2));
1686 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1687 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1688 reverse = !reverse;
1690 cond1 = XEXP (SET_SRC (set1), 0);
1691 cond2 = XEXP (SET_SRC (set2), 0);
1692 code1 = GET_CODE (cond1);
1693 if (reverse)
1694 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1695 else
1696 code2 = GET_CODE (cond2);
1698 if (code2 == UNKNOWN)
1699 return false;
1701 /* Verify codes and operands match. */
1702 match = ((code1 == code2
1703 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1704 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1705 || (code1 == swap_condition (code2)
1706 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1707 XEXP (cond2, 0))
1708 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1709 XEXP (cond2, 1))));
1711 /* If we return true, we will join the blocks. Which means that
1712 we will only have one branch prediction bit to work with. Thus
1713 we require the existing branches to have probabilities that are
1714 roughly similar. */
1715 if (match
1716 && optimize_bb_for_speed_p (bb1)
1717 && optimize_bb_for_speed_p (bb2))
1719 profile_probability prob2;
1721 if (b1->dest == b2->dest)
1722 prob2 = b2->probability;
1723 else
1724 /* Do not use f2 probability as f2 may be forwarded. */
1725 prob2 = b2->probability.invert ();
1727 /* Fail if the difference in probabilities is greater than 50%.
1728 This rules out two well-predicted branches with opposite
1729 outcomes. */
1730 if (b1->probability.differs_lot_from_p (prob2))
1732 if (dump_file)
1734 fprintf (dump_file,
1735 "Outcomes of branch in bb %i and %i differ too"
1736 " much (", bb1->index, bb2->index);
1737 b1->probability.dump (dump_file);
1738 prob2.dump (dump_file);
1739 fprintf (dump_file, ")\n");
1741 return false;
1745 if (dump_file && match)
1746 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1747 bb1->index, bb2->index);
1749 return match;
1752 /* Generic case - we are seeing a computed jump, table jump or trapping
1753 instruction. */
1755 /* Check whether there are tablejumps in the end of BB1 and BB2.
1756 Return true if they are identical. */
1758 rtx_insn *label1, *label2;
1759 rtx_jump_table_data *table1, *table2;
1761 if (tablejump_p (BB_END (bb1), &label1, &table1)
1762 && tablejump_p (BB_END (bb2), &label2, &table2)
1763 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1765 /* The labels should never be the same rtx. If they really are same
1766 the jump tables are same too. So disable crossjumping of blocks BB1
1767 and BB2 because when deleting the common insns in the end of BB1
1768 by delete_basic_block () the jump table would be deleted too. */
1769 /* If LABEL2 is referenced in BB1->END do not do anything
1770 because we would loose information when replacing
1771 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1772 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1774 /* Set IDENTICAL to true when the tables are identical. */
1775 bool identical = false;
1776 rtx p1, p2;
1778 p1 = PATTERN (table1);
1779 p2 = PATTERN (table2);
1780 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1782 identical = true;
1784 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1785 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1786 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1787 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1789 int i;
1791 identical = true;
1792 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1793 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1794 identical = false;
1797 if (identical)
1799 bool match;
1801 /* Temporarily replace references to LABEL1 with LABEL2
1802 in BB1->END so that we could compare the instructions. */
1803 replace_label_in_insn (BB_END (bb1), label1, label2, false);
1805 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1806 == dir_both);
1807 if (dump_file && match)
1808 fprintf (dump_file,
1809 "Tablejumps in bb %i and %i match.\n",
1810 bb1->index, bb2->index);
1812 /* Set the original label in BB1->END because when deleting
1813 a block whose end is a tablejump, the tablejump referenced
1814 from the instruction is deleted too. */
1815 replace_label_in_insn (BB_END (bb1), label2, label1, false);
1817 return match;
1820 return false;
1824 /* Find the last non-debug non-note instruction in each bb, except
1825 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1826 handles that case specially. old_insns_match_p does not handle
1827 other types of instruction notes. */
1828 rtx_insn *last1 = BB_END (bb1);
1829 rtx_insn *last2 = BB_END (bb2);
1830 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1831 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1832 last1 = PREV_INSN (last1);
1833 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1834 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1835 last2 = PREV_INSN (last2);
1836 gcc_assert (last1 && last2);
1838 /* First ensure that the instructions match. There may be many outgoing
1839 edges so this test is generally cheaper. */
1840 if (old_insns_match_p (mode, last1, last2) != dir_both)
1841 return false;
1843 /* Search the outgoing edges, ensure that the counts do match, find possible
1844 fallthru and exception handling edges since these needs more
1845 validation. */
1846 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1847 return false;
1849 bool nonfakeedges = false;
1850 FOR_EACH_EDGE (e1, ei, bb1->succs)
1852 e2 = EDGE_SUCC (bb2, ei.index);
1854 if ((e1->flags & EDGE_FAKE) == 0)
1855 nonfakeedges = true;
1857 if (e1->flags & EDGE_EH)
1858 nehedges1++;
1860 if (e2->flags & EDGE_EH)
1861 nehedges2++;
1863 if (e1->flags & EDGE_FALLTHRU)
1864 fallthru1 = e1;
1865 if (e2->flags & EDGE_FALLTHRU)
1866 fallthru2 = e2;
1869 /* If number of edges of various types does not match, fail. */
1870 if (nehedges1 != nehedges2
1871 || (fallthru1 != 0) != (fallthru2 != 0))
1872 return false;
1874 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1875 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1876 attempt to optimize, as the two basic blocks might have different
1877 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1878 traps there should be REG_ARG_SIZE notes, they could be missing
1879 for __builtin_unreachable () uses though. */
1880 if (!nonfakeedges
1881 && !ACCUMULATE_OUTGOING_ARGS
1882 && (!INSN_P (last1)
1883 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1884 return false;
1886 /* fallthru edges must be forwarded to the same destination. */
1887 if (fallthru1)
1889 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1890 ? single_succ (fallthru1->dest): fallthru1->dest);
1891 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1892 ? single_succ (fallthru2->dest): fallthru2->dest);
1894 if (d1 != d2)
1895 return false;
1898 /* Ensure the same EH region. */
1900 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1901 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1903 if (!n1 && n2)
1904 return false;
1906 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1907 return false;
1910 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1911 version of sequence abstraction. */
1912 FOR_EACH_EDGE (e1, ei, bb2->succs)
1914 edge e2;
1915 edge_iterator ei;
1916 basic_block d1 = e1->dest;
1918 if (FORWARDER_BLOCK_P (d1))
1919 d1 = EDGE_SUCC (d1, 0)->dest;
1921 FOR_EACH_EDGE (e2, ei, bb1->succs)
1923 basic_block d2 = e2->dest;
1924 if (FORWARDER_BLOCK_P (d2))
1925 d2 = EDGE_SUCC (d2, 0)->dest;
1926 if (d1 == d2)
1927 break;
1930 if (!e2)
1931 return false;
1934 return true;
1937 /* Returns true if BB basic block has a preserve label. */
1939 static bool
1940 block_has_preserve_label (basic_block bb)
1942 return (bb
1943 && block_label (bb)
1944 && LABEL_PRESERVE_P (block_label (bb)));
1947 /* E1 and E2 are edges with the same destination block. Search their
1948 predecessors for common code. If found, redirect control flow from
1949 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1950 or the other way around (dir_backward). DIR specifies the allowed
1951 replacement direction. */
1953 static bool
1954 try_crossjump_to_edge (int mode, edge e1, edge e2,
1955 enum replace_direction dir)
1957 int nmatch;
1958 basic_block src1 = e1->src, src2 = e2->src;
1959 basic_block redirect_to, redirect_from, to_remove;
1960 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1961 rtx_insn *newpos1, *newpos2;
1962 edge s;
1963 edge_iterator ei;
1965 newpos1 = newpos2 = NULL;
1967 /* Search backward through forwarder blocks. We don't need to worry
1968 about multiple entry or chained forwarders, as they will be optimized
1969 away. We do this to look past the unconditional jump following a
1970 conditional jump that is required due to the current CFG shape. */
1971 if (single_pred_p (src1)
1972 && FORWARDER_BLOCK_P (src1))
1973 e1 = single_pred_edge (src1), src1 = e1->src;
1975 if (single_pred_p (src2)
1976 && FORWARDER_BLOCK_P (src2))
1977 e2 = single_pred_edge (src2), src2 = e2->src;
1979 /* Nothing to do if we reach ENTRY, or a common source block. */
1980 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1981 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1982 return false;
1983 if (src1 == src2)
1984 return false;
1986 /* Seeing more than 1 forwarder blocks would confuse us later... */
1987 if (FORWARDER_BLOCK_P (e1->dest)
1988 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1989 return false;
1991 if (FORWARDER_BLOCK_P (e2->dest)
1992 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1993 return false;
1995 /* Likewise with dead code (possibly newly created by the other optimizations
1996 of cfg_cleanup). */
1997 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1998 return false;
2000 /* Do not turn corssing edge to non-crossing or vice versa after reload. */
2001 if (BB_PARTITION (src1) != BB_PARTITION (src2)
2002 && reload_completed)
2003 return false;
2005 /* Look for the common insn sequence, part the first ... */
2006 if (!outgoing_edges_match (mode, src1, src2))
2007 return false;
2009 /* ... and part the second. */
2010 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
2012 osrc1 = src1;
2013 osrc2 = src2;
2014 if (newpos1 != NULL_RTX)
2015 src1 = BLOCK_FOR_INSN (newpos1);
2016 if (newpos2 != NULL_RTX)
2017 src2 = BLOCK_FOR_INSN (newpos2);
2019 /* Check that SRC1 and SRC2 have preds again. They may have changed
2020 above due to the call to flow_find_cross_jump. */
2021 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
2022 return false;
2024 if (dir == dir_backward)
2026 std::swap (osrc1, osrc2);
2027 std::swap (src1, src2);
2028 std::swap (e1, e2);
2029 std::swap (newpos1, newpos2);
2032 /* Don't proceed with the crossjump unless we found a sufficient number
2033 of matching instructions or the 'from' block was totally matched
2034 (such that its predecessors will hopefully be redirected and the
2035 block removed). */
2036 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
2037 && (newpos1 != BB_HEAD (src1)))
2038 return false;
2040 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
2041 if (block_has_preserve_label (e1->dest)
2042 && (e1->flags & EDGE_ABNORMAL))
2043 return false;
2045 /* Here we know that the insns in the end of SRC1 which are common with SRC2
2046 will be deleted.
2047 If we have tablejumps in the end of SRC1 and SRC2
2048 they have been already compared for equivalence in outgoing_edges_match ()
2049 so replace the references to TABLE1 by references to TABLE2. */
2051 rtx_insn *label1, *label2;
2052 rtx_jump_table_data *table1, *table2;
2054 if (tablejump_p (BB_END (osrc1), &label1, &table1)
2055 && tablejump_p (BB_END (osrc2), &label2, &table2)
2056 && label1 != label2)
2058 rtx_insn *insn;
2060 /* Replace references to LABEL1 with LABEL2. */
2061 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2063 /* Do not replace the label in SRC1->END because when deleting
2064 a block whose end is a tablejump, the tablejump referenced
2065 from the instruction is deleted too. */
2066 if (insn != BB_END (osrc1))
2067 replace_label_in_insn (insn, label1, label2, true);
2072 /* Avoid splitting if possible. We must always split when SRC2 has
2073 EH predecessor edges, or we may end up with basic blocks with both
2074 normal and EH predecessor edges. */
2075 if (newpos2 == BB_HEAD (src2)
2076 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2077 redirect_to = src2;
2078 else
2080 if (newpos2 == BB_HEAD (src2))
2082 /* Skip possible basic block header. */
2083 if (LABEL_P (newpos2))
2084 newpos2 = NEXT_INSN (newpos2);
2085 while (DEBUG_INSN_P (newpos2))
2086 newpos2 = NEXT_INSN (newpos2);
2087 if (NOTE_P (newpos2))
2088 newpos2 = NEXT_INSN (newpos2);
2089 while (DEBUG_INSN_P (newpos2))
2090 newpos2 = NEXT_INSN (newpos2);
2093 if (dump_file)
2094 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2095 src2->index, nmatch);
2096 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2099 if (dump_file)
2100 fprintf (dump_file,
2101 "Cross jumping from bb %i to bb %i; %i common insns\n",
2102 src1->index, src2->index, nmatch);
2104 /* We may have some registers visible through the block. */
2105 df_set_bb_dirty (redirect_to);
2107 if (osrc2 == src2)
2108 redirect_edges_to = redirect_to;
2109 else
2110 redirect_edges_to = osrc2;
2112 /* Recompute the frequencies and counts of outgoing edges. */
2113 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2115 edge s2;
2116 edge_iterator ei;
2117 basic_block d = s->dest;
2119 if (FORWARDER_BLOCK_P (d))
2120 d = single_succ (d);
2122 FOR_EACH_EDGE (s2, ei, src1->succs)
2124 basic_block d2 = s2->dest;
2125 if (FORWARDER_BLOCK_P (d2))
2126 d2 = single_succ (d2);
2127 if (d == d2)
2128 break;
2131 /* Take care to update possible forwarder blocks. We verified
2132 that there is no more than one in the chain, so we can't run
2133 into infinite loop. */
2134 if (FORWARDER_BLOCK_P (s->dest))
2136 s->dest->frequency += EDGE_FREQUENCY (s);
2139 if (FORWARDER_BLOCK_P (s2->dest))
2141 s2->dest->frequency -= EDGE_FREQUENCY (s);
2142 if (s2->dest->frequency < 0)
2143 s2->dest->frequency = 0;
2146 if (!redirect_edges_to->frequency && !src1->frequency)
2147 s->probability = s->probability.combine_with_freq
2148 (redirect_edges_to->frequency,
2149 s2->probability, src1->frequency);
2152 /* Adjust count and frequency for the block. An earlier jump
2153 threading pass may have left the profile in an inconsistent
2154 state (see update_bb_profile_for_threading) so we must be
2155 prepared for overflows. */
2156 tmp = redirect_to;
2159 tmp->count += src1->count;
2160 tmp->frequency += src1->frequency;
2161 if (tmp->frequency > BB_FREQ_MAX)
2162 tmp->frequency = BB_FREQ_MAX;
2163 if (tmp == redirect_edges_to)
2164 break;
2165 tmp = find_fallthru_edge (tmp->succs)->dest;
2167 while (true);
2168 update_br_prob_note (redirect_edges_to);
2170 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2172 /* Skip possible basic block header. */
2173 if (LABEL_P (newpos1))
2174 newpos1 = NEXT_INSN (newpos1);
2176 while (DEBUG_INSN_P (newpos1))
2177 newpos1 = NEXT_INSN (newpos1);
2179 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2180 newpos1 = NEXT_INSN (newpos1);
2182 while (DEBUG_INSN_P (newpos1))
2183 newpos1 = NEXT_INSN (newpos1);
2185 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2186 to_remove = single_succ (redirect_from);
2188 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2189 delete_basic_block (to_remove);
2191 update_forwarder_flag (redirect_from);
2192 if (redirect_to != src2)
2193 update_forwarder_flag (src2);
2195 return true;
2198 /* Search the predecessors of BB for common insn sequences. When found,
2199 share code between them by redirecting control flow. Return true if
2200 any changes made. */
2202 static bool
2203 try_crossjump_bb (int mode, basic_block bb)
2205 edge e, e2, fallthru;
2206 bool changed;
2207 unsigned max, ix, ix2;
2209 /* Nothing to do if there is not at least two incoming edges. */
2210 if (EDGE_COUNT (bb->preds) < 2)
2211 return false;
2213 /* Don't crossjump if this block ends in a computed jump,
2214 unless we are optimizing for size. */
2215 if (optimize_bb_for_size_p (bb)
2216 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2217 && computed_jump_p (BB_END (bb)))
2218 return false;
2220 /* If we are partitioning hot/cold basic blocks, we don't want to
2221 mess up unconditional or indirect jumps that cross between hot
2222 and cold sections.
2224 Basic block partitioning may result in some jumps that appear to
2225 be optimizable (or blocks that appear to be mergeable), but which really
2226 must be left untouched (they are required to make it safely across
2227 partition boundaries). See the comments at the top of
2228 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2230 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2231 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2232 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2233 return false;
2235 /* It is always cheapest to redirect a block that ends in a branch to
2236 a block that falls through into BB, as that adds no branches to the
2237 program. We'll try that combination first. */
2238 fallthru = NULL;
2239 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2241 if (EDGE_COUNT (bb->preds) > max)
2242 return false;
2244 fallthru = find_fallthru_edge (bb->preds);
2246 changed = false;
2247 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2249 e = EDGE_PRED (bb, ix);
2250 ix++;
2252 /* As noted above, first try with the fallthru predecessor (or, a
2253 fallthru predecessor if we are in cfglayout mode). */
2254 if (fallthru)
2256 /* Don't combine the fallthru edge into anything else.
2257 If there is a match, we'll do it the other way around. */
2258 if (e == fallthru)
2259 continue;
2260 /* If nothing changed since the last attempt, there is nothing
2261 we can do. */
2262 if (!first_pass
2263 && !((e->src->flags & BB_MODIFIED)
2264 || (fallthru->src->flags & BB_MODIFIED)))
2265 continue;
2267 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2269 changed = true;
2270 ix = 0;
2271 continue;
2275 /* Non-obvious work limiting check: Recognize that we're going
2276 to call try_crossjump_bb on every basic block. So if we have
2277 two blocks with lots of outgoing edges (a switch) and they
2278 share lots of common destinations, then we would do the
2279 cross-jump check once for each common destination.
2281 Now, if the blocks actually are cross-jump candidates, then
2282 all of their destinations will be shared. Which means that
2283 we only need check them for cross-jump candidacy once. We
2284 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2285 choosing to do the check from the block for which the edge
2286 in question is the first successor of A. */
2287 if (EDGE_SUCC (e->src, 0) != e)
2288 continue;
2290 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2292 e2 = EDGE_PRED (bb, ix2);
2294 if (e2 == e)
2295 continue;
2297 /* We've already checked the fallthru edge above. */
2298 if (e2 == fallthru)
2299 continue;
2301 /* The "first successor" check above only prevents multiple
2302 checks of crossjump(A,B). In order to prevent redundant
2303 checks of crossjump(B,A), require that A be the block
2304 with the lowest index. */
2305 if (e->src->index > e2->src->index)
2306 continue;
2308 /* If nothing changed since the last attempt, there is nothing
2309 we can do. */
2310 if (!first_pass
2311 && !((e->src->flags & BB_MODIFIED)
2312 || (e2->src->flags & BB_MODIFIED)))
2313 continue;
2315 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2316 direction. */
2317 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2319 changed = true;
2320 ix = 0;
2321 break;
2326 if (changed)
2327 crossjumps_occurred = true;
2329 return changed;
2332 /* Search the successors of BB for common insn sequences. When found,
2333 share code between them by moving it across the basic block
2334 boundary. Return true if any changes made. */
2336 static bool
2337 try_head_merge_bb (basic_block bb)
2339 basic_block final_dest_bb = NULL;
2340 int max_match = INT_MAX;
2341 edge e0;
2342 rtx_insn **headptr, **currptr, **nextptr;
2343 bool changed, moveall;
2344 unsigned ix;
2345 rtx_insn *e0_last_head;
2346 rtx cond;
2347 rtx_insn *move_before;
2348 unsigned nedges = EDGE_COUNT (bb->succs);
2349 rtx_insn *jump = BB_END (bb);
2350 regset live, live_union;
2352 /* Nothing to do if there is not at least two outgoing edges. */
2353 if (nedges < 2)
2354 return false;
2356 /* Don't crossjump if this block ends in a computed jump,
2357 unless we are optimizing for size. */
2358 if (optimize_bb_for_size_p (bb)
2359 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2360 && computed_jump_p (BB_END (bb)))
2361 return false;
2363 cond = get_condition (jump, &move_before, true, false);
2364 if (cond == NULL_RTX)
2366 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2367 move_before = prev_nonnote_nondebug_insn (jump);
2368 else
2369 move_before = jump;
2372 for (ix = 0; ix < nedges; ix++)
2373 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2374 return false;
2376 for (ix = 0; ix < nedges; ix++)
2378 edge e = EDGE_SUCC (bb, ix);
2379 basic_block other_bb = e->dest;
2381 if (df_get_bb_dirty (other_bb))
2383 block_was_dirty = true;
2384 return false;
2387 if (e->flags & EDGE_ABNORMAL)
2388 return false;
2390 /* Normally, all destination blocks must only be reachable from this
2391 block, i.e. they must have one incoming edge.
2393 There is one special case we can handle, that of multiple consecutive
2394 jumps where the first jumps to one of the targets of the second jump.
2395 This happens frequently in switch statements for default labels.
2396 The structure is as follows:
2397 FINAL_DEST_BB
2398 ....
2399 if (cond) jump A;
2400 fall through
2402 jump with targets A, B, C, D...
2404 has two incoming edges, from FINAL_DEST_BB and BB
2406 In this case, we can try to move the insns through BB and into
2407 FINAL_DEST_BB. */
2408 if (EDGE_COUNT (other_bb->preds) != 1)
2410 edge incoming_edge, incoming_bb_other_edge;
2411 edge_iterator ei;
2413 if (final_dest_bb != NULL
2414 || EDGE_COUNT (other_bb->preds) != 2)
2415 return false;
2417 /* We must be able to move the insns across the whole block. */
2418 move_before = BB_HEAD (bb);
2419 while (!NONDEBUG_INSN_P (move_before))
2420 move_before = NEXT_INSN (move_before);
2422 if (EDGE_COUNT (bb->preds) != 1)
2423 return false;
2424 incoming_edge = EDGE_PRED (bb, 0);
2425 final_dest_bb = incoming_edge->src;
2426 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2427 return false;
2428 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2429 if (incoming_bb_other_edge != incoming_edge)
2430 break;
2431 if (incoming_bb_other_edge->dest != other_bb)
2432 return false;
2436 e0 = EDGE_SUCC (bb, 0);
2437 e0_last_head = NULL;
2438 changed = false;
2440 for (ix = 1; ix < nedges; ix++)
2442 edge e = EDGE_SUCC (bb, ix);
2443 rtx_insn *e0_last, *e_last;
2444 int nmatch;
2446 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2447 &e0_last, &e_last, 0);
2448 if (nmatch == 0)
2449 return false;
2451 if (nmatch < max_match)
2453 max_match = nmatch;
2454 e0_last_head = e0_last;
2458 /* If we matched an entire block, we probably have to avoid moving the
2459 last insn. */
2460 if (max_match > 0
2461 && e0_last_head == BB_END (e0->dest)
2462 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2463 || control_flow_insn_p (e0_last_head)))
2465 max_match--;
2466 if (max_match == 0)
2467 return false;
2469 e0_last_head = prev_real_insn (e0_last_head);
2470 while (DEBUG_INSN_P (e0_last_head));
2473 if (max_match == 0)
2474 return false;
2476 /* We must find a union of the live registers at each of the end points. */
2477 live = BITMAP_ALLOC (NULL);
2478 live_union = BITMAP_ALLOC (NULL);
2480 currptr = XNEWVEC (rtx_insn *, nedges);
2481 headptr = XNEWVEC (rtx_insn *, nedges);
2482 nextptr = XNEWVEC (rtx_insn *, nedges);
2484 for (ix = 0; ix < nedges; ix++)
2486 int j;
2487 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2488 rtx_insn *head = BB_HEAD (merge_bb);
2490 while (!NONDEBUG_INSN_P (head))
2491 head = NEXT_INSN (head);
2492 headptr[ix] = head;
2493 currptr[ix] = head;
2495 /* Compute the end point and live information */
2496 for (j = 1; j < max_match; j++)
2498 head = NEXT_INSN (head);
2499 while (!NONDEBUG_INSN_P (head));
2500 simulate_backwards_to_point (merge_bb, live, head);
2501 IOR_REG_SET (live_union, live);
2504 /* If we're moving across two blocks, verify the validity of the
2505 first move, then adjust the target and let the loop below deal
2506 with the final move. */
2507 if (final_dest_bb != NULL)
2509 rtx_insn *move_upto;
2511 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2512 jump, e0->dest, live_union,
2513 NULL, &move_upto);
2514 if (!moveall)
2516 if (move_upto == NULL_RTX)
2517 goto out;
2519 while (e0_last_head != move_upto)
2521 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2522 live_union);
2523 e0_last_head = PREV_INSN (e0_last_head);
2526 if (e0_last_head == NULL_RTX)
2527 goto out;
2529 jump = BB_END (final_dest_bb);
2530 cond = get_condition (jump, &move_before, true, false);
2531 if (cond == NULL_RTX)
2533 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2534 move_before = prev_nonnote_nondebug_insn (jump);
2535 else
2536 move_before = jump;
2542 rtx_insn *move_upto;
2543 moveall = can_move_insns_across (currptr[0], e0_last_head,
2544 move_before, jump, e0->dest, live_union,
2545 NULL, &move_upto);
2546 if (!moveall && move_upto == NULL_RTX)
2548 if (jump == move_before)
2549 break;
2551 /* Try again, using a different insertion point. */
2552 move_before = jump;
2554 /* Don't try moving before a cc0 user, as that may invalidate
2555 the cc0. */
2556 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2557 break;
2559 continue;
2562 if (final_dest_bb && !moveall)
2563 /* We haven't checked whether a partial move would be OK for the first
2564 move, so we have to fail this case. */
2565 break;
2567 changed = true;
2568 for (;;)
2570 if (currptr[0] == move_upto)
2571 break;
2572 for (ix = 0; ix < nedges; ix++)
2574 rtx_insn *curr = currptr[ix];
2576 curr = NEXT_INSN (curr);
2577 while (!NONDEBUG_INSN_P (curr));
2578 currptr[ix] = curr;
2582 /* If we can't currently move all of the identical insns, remember
2583 each insn after the range that we'll merge. */
2584 if (!moveall)
2585 for (ix = 0; ix < nedges; ix++)
2587 rtx_insn *curr = currptr[ix];
2589 curr = NEXT_INSN (curr);
2590 while (!NONDEBUG_INSN_P (curr));
2591 nextptr[ix] = curr;
2594 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2595 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2596 if (final_dest_bb != NULL)
2597 df_set_bb_dirty (final_dest_bb);
2598 df_set_bb_dirty (bb);
2599 for (ix = 1; ix < nedges; ix++)
2601 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2602 delete_insn_chain (headptr[ix], currptr[ix], false);
2604 if (!moveall)
2606 if (jump == move_before)
2607 break;
2609 /* For the unmerged insns, try a different insertion point. */
2610 move_before = jump;
2612 /* Don't try moving before a cc0 user, as that may invalidate
2613 the cc0. */
2614 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2615 break;
2617 for (ix = 0; ix < nedges; ix++)
2618 currptr[ix] = headptr[ix] = nextptr[ix];
2621 while (!moveall);
2623 out:
2624 free (currptr);
2625 free (headptr);
2626 free (nextptr);
2628 crossjumps_occurred |= changed;
2630 return changed;
2633 /* Return true if BB contains just bb note, or bb note followed
2634 by only DEBUG_INSNs. */
2636 static bool
2637 trivially_empty_bb_p (basic_block bb)
2639 rtx_insn *insn = BB_END (bb);
2641 while (1)
2643 if (insn == BB_HEAD (bb))
2644 return true;
2645 if (!DEBUG_INSN_P (insn))
2646 return false;
2647 insn = PREV_INSN (insn);
2651 /* Return true if BB contains just a return and possibly a USE of the
2652 return value. Fill in *RET and *USE with the return and use insns
2653 if any found, otherwise NULL. All CLOBBERs are ignored. */
2655 static bool
2656 bb_is_just_return (basic_block bb, rtx_insn **ret, rtx_insn **use)
2658 *ret = *use = NULL;
2659 rtx_insn *insn;
2661 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
2662 return false;
2664 FOR_BB_INSNS (bb, insn)
2665 if (NONDEBUG_INSN_P (insn))
2667 rtx pat = PATTERN (insn);
2669 if (!*ret && ANY_RETURN_P (pat))
2670 *ret = insn;
2671 else if (!*ret && !*use && GET_CODE (pat) == USE
2672 && REG_P (XEXP (pat, 0))
2673 && REG_FUNCTION_VALUE_P (XEXP (pat, 0)))
2674 *use = insn;
2675 else if (GET_CODE (pat) != CLOBBER)
2676 return false;
2679 return !!*ret;
2682 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2683 instructions etc. Return nonzero if changes were made. */
2685 static bool
2686 try_optimize_cfg (int mode)
2688 bool changed_overall = false;
2689 bool changed;
2690 int iterations = 0;
2691 basic_block bb, b, next;
2693 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2694 clear_bb_flags ();
2696 crossjumps_occurred = false;
2698 FOR_EACH_BB_FN (bb, cfun)
2699 update_forwarder_flag (bb);
2701 if (! targetm.cannot_modify_jumps_p ())
2703 first_pass = true;
2704 /* Attempt to merge blocks as made possible by edge removal. If
2705 a block has only one successor, and the successor has only
2706 one predecessor, they may be combined. */
2709 block_was_dirty = false;
2710 changed = false;
2711 iterations++;
2713 if (dump_file)
2714 fprintf (dump_file,
2715 "\n\ntry_optimize_cfg iteration %i\n\n",
2716 iterations);
2718 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2719 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2721 basic_block c;
2722 edge s;
2723 bool changed_here = false;
2725 /* Delete trivially dead basic blocks. This is either
2726 blocks with no predecessors, or empty blocks with no
2727 successors. However if the empty block with no
2728 successors is the successor of the ENTRY_BLOCK, it is
2729 kept. This ensures that the ENTRY_BLOCK will have a
2730 successor which is a precondition for many RTL
2731 passes. Empty blocks may result from expanding
2732 __builtin_unreachable (). */
2733 if (EDGE_COUNT (b->preds) == 0
2734 || (EDGE_COUNT (b->succs) == 0
2735 && trivially_empty_bb_p (b)
2736 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2737 != b))
2739 c = b->prev_bb;
2740 if (EDGE_COUNT (b->preds) > 0)
2742 edge e;
2743 edge_iterator ei;
2745 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2747 if (BB_FOOTER (b)
2748 && BARRIER_P (BB_FOOTER (b)))
2749 FOR_EACH_EDGE (e, ei, b->preds)
2750 if ((e->flags & EDGE_FALLTHRU)
2751 && BB_FOOTER (e->src) == NULL)
2753 if (BB_FOOTER (b))
2755 BB_FOOTER (e->src) = BB_FOOTER (b);
2756 BB_FOOTER (b) = NULL;
2758 else
2760 start_sequence ();
2761 BB_FOOTER (e->src) = emit_barrier ();
2762 end_sequence ();
2766 else
2768 rtx_insn *last = get_last_bb_insn (b);
2769 if (last && BARRIER_P (last))
2770 FOR_EACH_EDGE (e, ei, b->preds)
2771 if ((e->flags & EDGE_FALLTHRU))
2772 emit_barrier_after (BB_END (e->src));
2775 delete_basic_block (b);
2776 changed = true;
2777 /* Avoid trying to remove the exit block. */
2778 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2779 continue;
2782 /* Remove code labels no longer used. */
2783 if (single_pred_p (b)
2784 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2785 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2786 && LABEL_P (BB_HEAD (b))
2787 && !LABEL_PRESERVE_P (BB_HEAD (b))
2788 /* If the previous block ends with a branch to this
2789 block, we can't delete the label. Normally this
2790 is a condjump that is yet to be simplified, but
2791 if CASE_DROPS_THRU, this can be a tablejump with
2792 some element going to the same place as the
2793 default (fallthru). */
2794 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2795 || !JUMP_P (BB_END (single_pred (b)))
2796 || ! label_is_jump_target_p (BB_HEAD (b),
2797 BB_END (single_pred (b)))))
2799 delete_insn (BB_HEAD (b));
2800 if (dump_file)
2801 fprintf (dump_file, "Deleted label in block %i.\n",
2802 b->index);
2805 /* If we fall through an empty block, we can remove it. */
2806 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2807 && single_pred_p (b)
2808 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2809 && !LABEL_P (BB_HEAD (b))
2810 && FORWARDER_BLOCK_P (b)
2811 /* Note that forwarder_block_p true ensures that
2812 there is a successor for this block. */
2813 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2814 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2816 if (dump_file)
2817 fprintf (dump_file,
2818 "Deleting fallthru block %i.\n",
2819 b->index);
2821 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2822 ? b->next_bb : b->prev_bb);
2823 redirect_edge_succ_nodup (single_pred_edge (b),
2824 single_succ (b));
2825 delete_basic_block (b);
2826 changed = true;
2827 b = c;
2828 continue;
2831 /* Merge B with its single successor, if any. */
2832 if (single_succ_p (b)
2833 && (s = single_succ_edge (b))
2834 && !(s->flags & EDGE_COMPLEX)
2835 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2836 && single_pred_p (c)
2837 && b != c)
2839 /* When not in cfg_layout mode use code aware of reordering
2840 INSN. This code possibly creates new basic blocks so it
2841 does not fit merge_blocks interface and is kept here in
2842 hope that it will become useless once more of compiler
2843 is transformed to use cfg_layout mode. */
2845 if ((mode & CLEANUP_CFGLAYOUT)
2846 && can_merge_blocks_p (b, c))
2848 merge_blocks (b, c);
2849 update_forwarder_flag (b);
2850 changed_here = true;
2852 else if (!(mode & CLEANUP_CFGLAYOUT)
2853 /* If the jump insn has side effects,
2854 we can't kill the edge. */
2855 && (!JUMP_P (BB_END (b))
2856 || (reload_completed
2857 ? simplejump_p (BB_END (b))
2858 : (onlyjump_p (BB_END (b))
2859 && !tablejump_p (BB_END (b),
2860 NULL, NULL))))
2861 && (next = merge_blocks_move (s, b, c, mode)))
2863 b = next;
2864 changed_here = true;
2868 /* Try to change a branch to a return to just that return. */
2869 rtx_insn *ret, *use;
2870 if (single_succ_p (b)
2871 && onlyjump_p (BB_END (b))
2872 && bb_is_just_return (single_succ (b), &ret, &use))
2874 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2875 PATTERN (ret), 0))
2877 if (use)
2878 emit_insn_before (copy_insn (PATTERN (use)),
2879 BB_END (b));
2880 if (dump_file)
2881 fprintf (dump_file, "Changed jump %d->%d to return.\n",
2882 b->index, single_succ (b)->index);
2883 redirect_edge_succ (single_succ_edge (b),
2884 EXIT_BLOCK_PTR_FOR_FN (cfun));
2885 single_succ_edge (b)->flags &= ~EDGE_CROSSING;
2886 changed_here = true;
2890 /* Try to change a conditional branch to a return to the
2891 respective conditional return. */
2892 if (EDGE_COUNT (b->succs) == 2
2893 && any_condjump_p (BB_END (b))
2894 && bb_is_just_return (BRANCH_EDGE (b)->dest, &ret, &use))
2896 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2897 PATTERN (ret), 0))
2899 if (use)
2900 emit_insn_before (copy_insn (PATTERN (use)),
2901 BB_END (b));
2902 if (dump_file)
2903 fprintf (dump_file, "Changed conditional jump %d->%d "
2904 "to conditional return.\n",
2905 b->index, BRANCH_EDGE (b)->dest->index);
2906 redirect_edge_succ (BRANCH_EDGE (b),
2907 EXIT_BLOCK_PTR_FOR_FN (cfun));
2908 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2909 changed_here = true;
2913 /* Try to flip a conditional branch that falls through to
2914 a return so that it becomes a conditional return and a
2915 new jump to the original branch target. */
2916 if (EDGE_COUNT (b->succs) == 2
2917 && BRANCH_EDGE (b)->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
2918 && any_condjump_p (BB_END (b))
2919 && bb_is_just_return (FALLTHRU_EDGE (b)->dest, &ret, &use))
2921 if (invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2922 JUMP_LABEL (BB_END (b)), 0))
2924 basic_block new_ft = BRANCH_EDGE (b)->dest;
2925 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2926 PATTERN (ret), 0))
2928 if (use)
2929 emit_insn_before (copy_insn (PATTERN (use)),
2930 BB_END (b));
2931 if (dump_file)
2932 fprintf (dump_file, "Changed conditional jump "
2933 "%d->%d to conditional return, adding "
2934 "fall-through jump.\n",
2935 b->index, BRANCH_EDGE (b)->dest->index);
2936 redirect_edge_succ (BRANCH_EDGE (b),
2937 EXIT_BLOCK_PTR_FOR_FN (cfun));
2938 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2939 std::swap (BRANCH_EDGE (b)->probability,
2940 FALLTHRU_EDGE (b)->probability);
2941 update_br_prob_note (b);
2942 basic_block jb = force_nonfallthru (FALLTHRU_EDGE (b));
2943 notice_new_block (jb);
2944 if (!redirect_jump (as_a <rtx_jump_insn *> (BB_END (jb)),
2945 block_label (new_ft), 0))
2946 gcc_unreachable ();
2947 redirect_edge_succ (single_succ_edge (jb), new_ft);
2948 changed_here = true;
2950 else
2952 /* Invert the jump back to what it was. This should
2953 never fail. */
2954 if (!invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2955 JUMP_LABEL (BB_END (b)), 0))
2956 gcc_unreachable ();
2961 /* Simplify branch over branch. */
2962 if ((mode & CLEANUP_EXPENSIVE)
2963 && !(mode & CLEANUP_CFGLAYOUT)
2964 && try_simplify_condjump (b))
2965 changed_here = true;
2967 /* If B has a single outgoing edge, but uses a
2968 non-trivial jump instruction without side-effects, we
2969 can either delete the jump entirely, or replace it
2970 with a simple unconditional jump. */
2971 if (single_succ_p (b)
2972 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2973 && onlyjump_p (BB_END (b))
2974 && !CROSSING_JUMP_P (BB_END (b))
2975 && try_redirect_by_replacing_jump (single_succ_edge (b),
2976 single_succ (b),
2977 (mode & CLEANUP_CFGLAYOUT) != 0))
2979 update_forwarder_flag (b);
2980 changed_here = true;
2983 /* Simplify branch to branch. */
2984 if (try_forward_edges (mode, b))
2986 update_forwarder_flag (b);
2987 changed_here = true;
2990 /* Look for shared code between blocks. */
2991 if ((mode & CLEANUP_CROSSJUMP)
2992 && try_crossjump_bb (mode, b))
2993 changed_here = true;
2995 if ((mode & CLEANUP_CROSSJUMP)
2996 /* This can lengthen register lifetimes. Do it only after
2997 reload. */
2998 && reload_completed
2999 && try_head_merge_bb (b))
3000 changed_here = true;
3002 /* Don't get confused by the index shift caused by
3003 deleting blocks. */
3004 if (!changed_here)
3005 b = b->next_bb;
3006 else
3007 changed = true;
3010 if ((mode & CLEANUP_CROSSJUMP)
3011 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
3012 changed = true;
3014 if (block_was_dirty)
3016 /* This should only be set by head-merging. */
3017 gcc_assert (mode & CLEANUP_CROSSJUMP);
3018 df_analyze ();
3021 if (changed)
3023 /* Edge forwarding in particular can cause hot blocks previously
3024 reached by both hot and cold blocks to become dominated only
3025 by cold blocks. This will cause the verification below to fail,
3026 and lead to now cold code in the hot section. This is not easy
3027 to detect and fix during edge forwarding, and in some cases
3028 is only visible after newly unreachable blocks are deleted,
3029 which will be done in fixup_partitions. */
3030 fixup_partitions ();
3031 checking_verify_flow_info ();
3034 changed_overall |= changed;
3035 first_pass = false;
3037 while (changed);
3040 FOR_ALL_BB_FN (b, cfun)
3041 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
3043 return changed_overall;
3046 /* Delete all unreachable basic blocks. */
3048 bool
3049 delete_unreachable_blocks (void)
3051 bool changed = false;
3052 basic_block b, prev_bb;
3054 find_unreachable_blocks ();
3056 /* When we're in GIMPLE mode and there may be debug insns, we should
3057 delete blocks in reverse dominator order, so as to get a chance
3058 to substitute all released DEFs into debug stmts. If we don't
3059 have dominators information, walking blocks backward gets us a
3060 better chance of retaining most debug information than
3061 otherwise. */
3062 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
3063 && dom_info_available_p (CDI_DOMINATORS))
3065 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3066 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3068 prev_bb = b->prev_bb;
3070 if (!(b->flags & BB_REACHABLE))
3072 /* Speed up the removal of blocks that don't dominate
3073 others. Walking backwards, this should be the common
3074 case. */
3075 if (!first_dom_son (CDI_DOMINATORS, b))
3076 delete_basic_block (b);
3077 else
3079 vec<basic_block> h
3080 = get_all_dominated_blocks (CDI_DOMINATORS, b);
3082 while (h.length ())
3084 b = h.pop ();
3086 prev_bb = b->prev_bb;
3088 gcc_assert (!(b->flags & BB_REACHABLE));
3090 delete_basic_block (b);
3093 h.release ();
3096 changed = true;
3100 else
3102 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3103 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3105 prev_bb = b->prev_bb;
3107 if (!(b->flags & BB_REACHABLE))
3109 delete_basic_block (b);
3110 changed = true;
3115 if (changed)
3116 tidy_fallthru_edges ();
3117 return changed;
3120 /* Delete any jump tables never referenced. We can't delete them at the
3121 time of removing tablejump insn as they are referenced by the preceding
3122 insns computing the destination, so we delay deleting and garbagecollect
3123 them once life information is computed. */
3124 void
3125 delete_dead_jumptables (void)
3127 basic_block bb;
3129 /* A dead jump table does not belong to any basic block. Scan insns
3130 between two adjacent basic blocks. */
3131 FOR_EACH_BB_FN (bb, cfun)
3133 rtx_insn *insn, *next;
3135 for (insn = NEXT_INSN (BB_END (bb));
3136 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
3137 insn = next)
3139 next = NEXT_INSN (insn);
3140 if (LABEL_P (insn)
3141 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
3142 && JUMP_TABLE_DATA_P (next))
3144 rtx_insn *label = insn, *jump = next;
3146 if (dump_file)
3147 fprintf (dump_file, "Dead jumptable %i removed\n",
3148 INSN_UID (insn));
3150 next = NEXT_INSN (next);
3151 delete_insn (jump);
3152 delete_insn (label);
3159 /* Tidy the CFG by deleting unreachable code and whatnot. */
3161 bool
3162 cleanup_cfg (int mode)
3164 bool changed = false;
3166 /* Set the cfglayout mode flag here. We could update all the callers
3167 but that is just inconvenient, especially given that we eventually
3168 want to have cfglayout mode as the default. */
3169 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3170 mode |= CLEANUP_CFGLAYOUT;
3172 timevar_push (TV_CLEANUP_CFG);
3173 if (delete_unreachable_blocks ())
3175 changed = true;
3176 /* We've possibly created trivially dead code. Cleanup it right
3177 now to introduce more opportunities for try_optimize_cfg. */
3178 if (!(mode & (CLEANUP_NO_INSN_DEL))
3179 && !reload_completed)
3180 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3183 compact_blocks ();
3185 /* To tail-merge blocks ending in the same noreturn function (e.g.
3186 a call to abort) we have to insert fake edges to exit. Do this
3187 here once. The fake edges do not interfere with any other CFG
3188 cleanups. */
3189 if (mode & CLEANUP_CROSSJUMP)
3190 add_noreturn_fake_exit_edges ();
3192 if (!dbg_cnt (cfg_cleanup))
3193 return changed;
3195 while (try_optimize_cfg (mode))
3197 delete_unreachable_blocks (), changed = true;
3198 if (!(mode & CLEANUP_NO_INSN_DEL))
3200 /* Try to remove some trivially dead insns when doing an expensive
3201 cleanup. But delete_trivially_dead_insns doesn't work after
3202 reload (it only handles pseudos) and run_fast_dce is too costly
3203 to run in every iteration.
3205 For effective cross jumping, we really want to run a fast DCE to
3206 clean up any dead conditions, or they get in the way of performing
3207 useful tail merges.
3209 Other transformations in cleanup_cfg are not so sensitive to dead
3210 code, so delete_trivially_dead_insns or even doing nothing at all
3211 is good enough. */
3212 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3213 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3214 break;
3215 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occurred)
3216 run_fast_dce ();
3218 else
3219 break;
3222 if (mode & CLEANUP_CROSSJUMP)
3223 remove_fake_exit_edges ();
3225 /* Don't call delete_dead_jumptables in cfglayout mode, because
3226 that function assumes that jump tables are in the insns stream.
3227 But we also don't _have_ to delete dead jumptables in cfglayout
3228 mode because we shouldn't even be looking at things that are
3229 not in a basic block. Dead jumptables are cleaned up when
3230 going out of cfglayout mode. */
3231 if (!(mode & CLEANUP_CFGLAYOUT))
3232 delete_dead_jumptables ();
3234 /* ??? We probably do this way too often. */
3235 if (current_loops
3236 && (changed
3237 || (mode & CLEANUP_CFG_CHANGED)))
3239 timevar_push (TV_REPAIR_LOOPS);
3240 /* The above doesn't preserve dominance info if available. */
3241 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3242 calculate_dominance_info (CDI_DOMINATORS);
3243 fix_loop_structure (NULL);
3244 free_dominance_info (CDI_DOMINATORS);
3245 timevar_pop (TV_REPAIR_LOOPS);
3248 timevar_pop (TV_CLEANUP_CFG);
3250 return changed;
3253 namespace {
3255 const pass_data pass_data_jump =
3257 RTL_PASS, /* type */
3258 "jump", /* name */
3259 OPTGROUP_NONE, /* optinfo_flags */
3260 TV_JUMP, /* tv_id */
3261 0, /* properties_required */
3262 0, /* properties_provided */
3263 0, /* properties_destroyed */
3264 0, /* todo_flags_start */
3265 0, /* todo_flags_finish */
3268 class pass_jump : public rtl_opt_pass
3270 public:
3271 pass_jump (gcc::context *ctxt)
3272 : rtl_opt_pass (pass_data_jump, ctxt)
3275 /* opt_pass methods: */
3276 virtual unsigned int execute (function *);
3278 }; // class pass_jump
3280 unsigned int
3281 pass_jump::execute (function *)
3283 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3284 if (dump_file)
3285 dump_flow_info (dump_file, dump_flags);
3286 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3287 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3288 return 0;
3291 } // anon namespace
3293 rtl_opt_pass *
3294 make_pass_jump (gcc::context *ctxt)
3296 return new pass_jump (ctxt);
3299 namespace {
3301 const pass_data pass_data_jump2 =
3303 RTL_PASS, /* type */
3304 "jump2", /* name */
3305 OPTGROUP_NONE, /* optinfo_flags */
3306 TV_JUMP, /* tv_id */
3307 0, /* properties_required */
3308 0, /* properties_provided */
3309 0, /* properties_destroyed */
3310 0, /* todo_flags_start */
3311 0, /* todo_flags_finish */
3314 class pass_jump2 : public rtl_opt_pass
3316 public:
3317 pass_jump2 (gcc::context *ctxt)
3318 : rtl_opt_pass (pass_data_jump2, ctxt)
3321 /* opt_pass methods: */
3322 virtual unsigned int execute (function *)
3324 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3325 return 0;
3328 }; // class pass_jump2
3330 } // anon namespace
3332 rtl_opt_pass *
3333 make_pass_jump2 (gcc::context *ctxt)
3335 return new pass_jump2 (ctxt);