PR middle-end/79665
[official-gcc.git] / gcc / cfgcleanup.c
blob9663b68fb3281b57658c258f72d2f5d325af494e
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 gcov_type edge_count = e->count;
562 int 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 = apply_probability (b->frequency, edge_probability);
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 if (first->count < 0)
607 first->count = 0;
608 first->frequency -= edge_frequency;
609 if (first->frequency < 0)
610 first->frequency = 0;
611 /* It is possible that as the result of
612 threading we've removed edge as it is
613 threaded to the fallthru edge. Avoid
614 getting out of sync. */
615 if (n < nthreaded_edges
616 && first == threaded_edges [n]->src)
617 n++;
618 t = single_succ_edge (first);
621 t->count -= edge_count;
622 if (t->count < 0)
623 t->count = 0;
624 first = t->dest;
626 while (first != target);
628 changed = true;
629 continue;
631 ei_next (&ei);
634 free (threaded_edges);
635 return changed;
639 /* Blocks A and B are to be merged into a single block. A has no incoming
640 fallthru edge, so it can be moved before B without adding or modifying
641 any jumps (aside from the jump from A to B). */
643 static void
644 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
646 rtx_insn *barrier;
648 /* If we are partitioning hot/cold basic blocks, we don't want to
649 mess up unconditional or indirect jumps that cross between hot
650 and cold sections.
652 Basic block partitioning may result in some jumps that appear to
653 be optimizable (or blocks that appear to be mergeable), but which really
654 must be left untouched (they are required to make it safely across
655 partition boundaries). See the comments at the top of
656 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
658 if (BB_PARTITION (a) != BB_PARTITION (b))
659 return;
661 barrier = next_nonnote_insn (BB_END (a));
662 gcc_assert (BARRIER_P (barrier));
663 delete_insn (barrier);
665 /* Scramble the insn chain. */
666 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
667 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
668 df_set_bb_dirty (a);
670 if (dump_file)
671 fprintf (dump_file, "Moved block %d before %d and merged.\n",
672 a->index, b->index);
674 /* Swap the records for the two blocks around. */
676 unlink_block (a);
677 link_block (a, b->prev_bb);
679 /* Now blocks A and B are contiguous. Merge them. */
680 merge_blocks (a, b);
683 /* Blocks A and B are to be merged into a single block. B has no outgoing
684 fallthru edge, so it can be moved after A without adding or modifying
685 any jumps (aside from the jump from A to B). */
687 static void
688 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
690 rtx_insn *barrier, *real_b_end;
691 rtx_insn *label;
692 rtx_jump_table_data *table;
694 /* If we are partitioning hot/cold basic blocks, we don't want to
695 mess up unconditional or indirect jumps that cross between hot
696 and cold sections.
698 Basic block partitioning may result in some jumps that appear to
699 be optimizable (or blocks that appear to be mergeable), but which really
700 must be left untouched (they are required to make it safely across
701 partition boundaries). See the comments at the top of
702 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
704 if (BB_PARTITION (a) != BB_PARTITION (b))
705 return;
707 real_b_end = BB_END (b);
709 /* If there is a jump table following block B temporarily add the jump table
710 to block B so that it will also be moved to the correct location. */
711 if (tablejump_p (BB_END (b), &label, &table)
712 && prev_active_insn (label) == BB_END (b))
714 BB_END (b) = table;
717 /* There had better have been a barrier there. Delete it. */
718 barrier = NEXT_INSN (BB_END (b));
719 if (barrier && BARRIER_P (barrier))
720 delete_insn (barrier);
723 /* Scramble the insn chain. */
724 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
726 /* Restore the real end of b. */
727 BB_END (b) = real_b_end;
729 if (dump_file)
730 fprintf (dump_file, "Moved block %d after %d and merged.\n",
731 b->index, a->index);
733 /* Now blocks A and B are contiguous. Merge them. */
734 merge_blocks (a, b);
737 /* Attempt to merge basic blocks that are potentially non-adjacent.
738 Return NULL iff the attempt failed, otherwise return basic block
739 where cleanup_cfg should continue. Because the merging commonly
740 moves basic block away or introduces another optimization
741 possibility, return basic block just before B so cleanup_cfg don't
742 need to iterate.
744 It may be good idea to return basic block before C in the case
745 C has been moved after B and originally appeared earlier in the
746 insn sequence, but we have no information available about the
747 relative ordering of these two. Hopefully it is not too common. */
749 static basic_block
750 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
752 basic_block next;
754 /* If we are partitioning hot/cold basic blocks, we don't want to
755 mess up unconditional or indirect jumps that cross between hot
756 and cold sections.
758 Basic block partitioning may result in some jumps that appear to
759 be optimizable (or blocks that appear to be mergeable), but which really
760 must be left untouched (they are required to make it safely across
761 partition boundaries). See the comments at the top of
762 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
764 if (BB_PARTITION (b) != BB_PARTITION (c))
765 return NULL;
767 /* If B has a fallthru edge to C, no need to move anything. */
768 if (e->flags & EDGE_FALLTHRU)
770 int b_index = b->index, c_index = c->index;
772 /* Protect the loop latches. */
773 if (current_loops && c->loop_father->latch == c)
774 return NULL;
776 merge_blocks (b, c);
777 update_forwarder_flag (b);
779 if (dump_file)
780 fprintf (dump_file, "Merged %d and %d without moving.\n",
781 b_index, c_index);
783 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
786 /* Otherwise we will need to move code around. Do that only if expensive
787 transformations are allowed. */
788 else if (mode & CLEANUP_EXPENSIVE)
790 edge tmp_edge, b_fallthru_edge;
791 bool c_has_outgoing_fallthru;
792 bool b_has_incoming_fallthru;
794 /* Avoid overactive code motion, as the forwarder blocks should be
795 eliminated by edge redirection instead. One exception might have
796 been if B is a forwarder block and C has no fallthru edge, but
797 that should be cleaned up by bb-reorder instead. */
798 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
799 return NULL;
801 /* We must make sure to not munge nesting of lexical blocks,
802 and loop notes. This is done by squeezing out all the notes
803 and leaving them there to lie. Not ideal, but functional. */
805 tmp_edge = find_fallthru_edge (c->succs);
806 c_has_outgoing_fallthru = (tmp_edge != NULL);
808 tmp_edge = find_fallthru_edge (b->preds);
809 b_has_incoming_fallthru = (tmp_edge != NULL);
810 b_fallthru_edge = tmp_edge;
811 next = b->prev_bb;
812 if (next == c)
813 next = next->prev_bb;
815 /* Otherwise, we're going to try to move C after B. If C does
816 not have an outgoing fallthru, then it can be moved
817 immediately after B without introducing or modifying jumps. */
818 if (! c_has_outgoing_fallthru)
820 merge_blocks_move_successor_nojumps (b, c);
821 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
824 /* If B does not have an incoming fallthru, then it can be moved
825 immediately before C without introducing or modifying jumps.
826 C cannot be the first block, so we do not have to worry about
827 accessing a non-existent block. */
829 if (b_has_incoming_fallthru)
831 basic_block bb;
833 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
834 return NULL;
835 bb = force_nonfallthru (b_fallthru_edge);
836 if (bb)
837 notice_new_block (bb);
840 merge_blocks_move_predecessor_nojumps (b, c);
841 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
844 return NULL;
848 /* Removes the memory attributes of MEM expression
849 if they are not equal. */
851 static void
852 merge_memattrs (rtx x, rtx y)
854 int i;
855 int j;
856 enum rtx_code code;
857 const char *fmt;
859 if (x == y)
860 return;
861 if (x == 0 || y == 0)
862 return;
864 code = GET_CODE (x);
866 if (code != GET_CODE (y))
867 return;
869 if (GET_MODE (x) != GET_MODE (y))
870 return;
872 if (code == MEM && !mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
874 if (! MEM_ATTRS (x))
875 MEM_ATTRS (y) = 0;
876 else if (! MEM_ATTRS (y))
877 MEM_ATTRS (x) = 0;
878 else
880 HOST_WIDE_INT mem_size;
882 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
884 set_mem_alias_set (x, 0);
885 set_mem_alias_set (y, 0);
888 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
890 set_mem_expr (x, 0);
891 set_mem_expr (y, 0);
892 clear_mem_offset (x);
893 clear_mem_offset (y);
895 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
896 || (MEM_OFFSET_KNOWN_P (x)
897 && MEM_OFFSET (x) != MEM_OFFSET (y)))
899 clear_mem_offset (x);
900 clear_mem_offset (y);
903 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
905 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
906 set_mem_size (x, mem_size);
907 set_mem_size (y, mem_size);
909 else
911 clear_mem_size (x);
912 clear_mem_size (y);
915 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
916 set_mem_align (y, MEM_ALIGN (x));
919 if (code == MEM)
921 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
923 MEM_READONLY_P (x) = 0;
924 MEM_READONLY_P (y) = 0;
926 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
928 MEM_NOTRAP_P (x) = 0;
929 MEM_NOTRAP_P (y) = 0;
931 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
933 MEM_VOLATILE_P (x) = 1;
934 MEM_VOLATILE_P (y) = 1;
938 fmt = GET_RTX_FORMAT (code);
939 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
941 switch (fmt[i])
943 case 'E':
944 /* Two vectors must have the same length. */
945 if (XVECLEN (x, i) != XVECLEN (y, i))
946 return;
948 for (j = 0; j < XVECLEN (x, i); j++)
949 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
951 break;
953 case 'e':
954 merge_memattrs (XEXP (x, i), XEXP (y, i));
957 return;
961 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
962 different single sets S1 and S2. */
964 static bool
965 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
967 int i;
968 rtx e1, e2;
970 if (p1 == s1 && p2 == s2)
971 return true;
973 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
974 return false;
976 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
977 return false;
979 for (i = 0; i < XVECLEN (p1, 0); i++)
981 e1 = XVECEXP (p1, 0, i);
982 e2 = XVECEXP (p2, 0, i);
983 if (e1 == s1 && e2 == s2)
984 continue;
985 if (reload_completed
986 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
987 continue;
989 return false;
992 return true;
996 /* NOTE1 is the REG_EQUAL note, if any, attached to an insn
997 that is a single_set with a SET_SRC of SRC1. Similarly
998 for NOTE2/SRC2.
1000 So effectively NOTE1/NOTE2 are an alternate form of
1001 SRC1/SRC2 respectively.
1003 Return nonzero if SRC1 or NOTE1 has the same constant
1004 integer value as SRC2 or NOTE2. Else return zero. */
1005 static int
1006 values_equal_p (rtx note1, rtx note2, rtx src1, rtx src2)
1008 if (note1
1009 && note2
1010 && CONST_INT_P (XEXP (note1, 0))
1011 && rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0)))
1012 return 1;
1014 if (!note1
1015 && !note2
1016 && CONST_INT_P (src1)
1017 && CONST_INT_P (src2)
1018 && rtx_equal_p (src1, src2))
1019 return 1;
1021 if (note1
1022 && CONST_INT_P (src2)
1023 && rtx_equal_p (XEXP (note1, 0), src2))
1024 return 1;
1026 if (note2
1027 && CONST_INT_P (src1)
1028 && rtx_equal_p (XEXP (note2, 0), src1))
1029 return 1;
1031 return 0;
1034 /* Examine register notes on I1 and I2 and return:
1035 - dir_forward if I1 can be replaced by I2, or
1036 - dir_backward if I2 can be replaced by I1, or
1037 - dir_both if both are the case. */
1039 static enum replace_direction
1040 can_replace_by (rtx_insn *i1, rtx_insn *i2)
1042 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1043 bool c1, c2;
1045 /* Check for 2 sets. */
1046 s1 = single_set (i1);
1047 s2 = single_set (i2);
1048 if (s1 == NULL_RTX || s2 == NULL_RTX)
1049 return dir_none;
1051 /* Check that the 2 sets set the same dest. */
1052 d1 = SET_DEST (s1);
1053 d2 = SET_DEST (s2);
1054 if (!(reload_completed
1055 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1056 return dir_none;
1058 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1059 set dest to the same value. */
1060 note1 = find_reg_equal_equiv_note (i1);
1061 note2 = find_reg_equal_equiv_note (i2);
1063 src1 = SET_SRC (s1);
1064 src2 = SET_SRC (s2);
1066 if (!values_equal_p (note1, note2, src1, src2))
1067 return dir_none;
1069 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1070 return dir_none;
1072 /* Although the 2 sets set dest to the same value, we cannot replace
1073 (set (dest) (const_int))
1075 (set (dest) (reg))
1076 because we don't know if the reg is live and has the same value at the
1077 location of replacement. */
1078 c1 = CONST_INT_P (src1);
1079 c2 = CONST_INT_P (src2);
1080 if (c1 && c2)
1081 return dir_both;
1082 else if (c2)
1083 return dir_forward;
1084 else if (c1)
1085 return dir_backward;
1087 return dir_none;
1090 /* Merges directions A and B. */
1092 static enum replace_direction
1093 merge_dir (enum replace_direction a, enum replace_direction b)
1095 /* Implements the following table:
1096 |bo fw bw no
1097 ---+-----------
1098 bo |bo fw bw no
1099 fw |-- fw no no
1100 bw |-- -- bw no
1101 no |-- -- -- no. */
1103 if (a == b)
1104 return a;
1106 if (a == dir_both)
1107 return b;
1108 if (b == dir_both)
1109 return a;
1111 return dir_none;
1114 /* Examine I1 and I2 and return:
1115 - dir_forward if I1 can be replaced by I2, or
1116 - dir_backward if I2 can be replaced by I1, or
1117 - dir_both if both are the case. */
1119 static enum replace_direction
1120 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1122 rtx p1, p2;
1124 /* Verify that I1 and I2 are equivalent. */
1125 if (GET_CODE (i1) != GET_CODE (i2))
1126 return dir_none;
1128 /* __builtin_unreachable() may lead to empty blocks (ending with
1129 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1130 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1131 return dir_both;
1133 /* ??? Do not allow cross-jumping between different stack levels. */
1134 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1135 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1136 if (p1 && p2)
1138 p1 = XEXP (p1, 0);
1139 p2 = XEXP (p2, 0);
1140 if (!rtx_equal_p (p1, p2))
1141 return dir_none;
1143 /* ??? Worse, this adjustment had better be constant lest we
1144 have differing incoming stack levels. */
1145 if (!frame_pointer_needed
1146 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1147 return dir_none;
1149 else if (p1 || p2)
1150 return dir_none;
1152 p1 = PATTERN (i1);
1153 p2 = PATTERN (i2);
1155 if (GET_CODE (p1) != GET_CODE (p2))
1156 return dir_none;
1158 /* If this is a CALL_INSN, compare register usage information.
1159 If we don't check this on stack register machines, the two
1160 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1161 numbers of stack registers in the same basic block.
1162 If we don't check this on machines with delay slots, a delay slot may
1163 be filled that clobbers a parameter expected by the subroutine.
1165 ??? We take the simple route for now and assume that if they're
1166 equal, they were constructed identically.
1168 Also check for identical exception regions. */
1170 if (CALL_P (i1))
1172 /* Ensure the same EH region. */
1173 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1174 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1176 if (!n1 && n2)
1177 return dir_none;
1179 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1180 return dir_none;
1182 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1183 CALL_INSN_FUNCTION_USAGE (i2))
1184 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1185 return dir_none;
1187 /* For address sanitizer, never crossjump __asan_report_* builtins,
1188 otherwise errors might be reported on incorrect lines. */
1189 if (flag_sanitize & SANITIZE_ADDRESS)
1191 rtx call = get_call_rtx_from (i1);
1192 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1194 rtx symbol = XEXP (XEXP (call, 0), 0);
1195 if (SYMBOL_REF_DECL (symbol)
1196 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1198 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1199 == BUILT_IN_NORMAL)
1200 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1201 >= BUILT_IN_ASAN_REPORT_LOAD1
1202 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1203 <= BUILT_IN_ASAN_STOREN)
1204 return dir_none;
1210 #ifdef STACK_REGS
1211 /* If cross_jump_death_matters is not 0, the insn's mode
1212 indicates whether or not the insn contains any stack-like
1213 regs. */
1215 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1217 /* If register stack conversion has already been done, then
1218 death notes must also be compared before it is certain that
1219 the two instruction streams match. */
1221 rtx note;
1222 HARD_REG_SET i1_regset, i2_regset;
1224 CLEAR_HARD_REG_SET (i1_regset);
1225 CLEAR_HARD_REG_SET (i2_regset);
1227 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1228 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1229 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1231 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1232 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1233 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1235 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1236 return dir_none;
1238 #endif
1240 if (reload_completed
1241 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1242 return dir_both;
1244 return can_replace_by (i1, i2);
1247 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1248 flow_find_head_matching_sequence, ensure the notes match. */
1250 static void
1251 merge_notes (rtx_insn *i1, rtx_insn *i2)
1253 /* If the merged insns have different REG_EQUAL notes, then
1254 remove them. */
1255 rtx equiv1 = find_reg_equal_equiv_note (i1);
1256 rtx equiv2 = find_reg_equal_equiv_note (i2);
1258 if (equiv1 && !equiv2)
1259 remove_note (i1, equiv1);
1260 else if (!equiv1 && equiv2)
1261 remove_note (i2, equiv2);
1262 else if (equiv1 && equiv2
1263 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1265 remove_note (i1, equiv1);
1266 remove_note (i2, equiv2);
1270 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1271 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1272 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1273 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1274 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1276 static void
1277 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1278 bool *did_fallthru)
1280 edge fallthru;
1282 *did_fallthru = false;
1284 /* Ignore notes. */
1285 while (!NONDEBUG_INSN_P (*i1))
1287 if (*i1 != BB_HEAD (*bb1))
1289 *i1 = PREV_INSN (*i1);
1290 continue;
1293 if (!follow_fallthru)
1294 return;
1296 fallthru = find_fallthru_edge ((*bb1)->preds);
1297 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1298 || !single_succ_p (fallthru->src))
1299 return;
1301 *bb1 = fallthru->src;
1302 *i1 = BB_END (*bb1);
1303 *did_fallthru = true;
1307 /* Look through the insns at the end of BB1 and BB2 and find the longest
1308 sequence that are either equivalent, or allow forward or backward
1309 replacement. Store the first insns for that sequence in *F1 and *F2 and
1310 return the sequence length.
1312 DIR_P indicates the allowed replacement direction on function entry, and
1313 the actual replacement direction on function exit. If NULL, only equivalent
1314 sequences are allowed.
1316 To simplify callers of this function, if the blocks match exactly,
1317 store the head of the blocks in *F1 and *F2. */
1320 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1321 rtx_insn **f2, enum replace_direction *dir_p)
1323 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1324 int ninsns = 0;
1325 enum replace_direction dir, last_dir, afterlast_dir;
1326 bool follow_fallthru, did_fallthru;
1328 if (dir_p)
1329 dir = *dir_p;
1330 else
1331 dir = dir_both;
1332 afterlast_dir = dir;
1333 last_dir = afterlast_dir;
1335 /* Skip simple jumps at the end of the blocks. Complex jumps still
1336 need to be compared for equivalence, which we'll do below. */
1338 i1 = BB_END (bb1);
1339 last1 = afterlast1 = last2 = afterlast2 = NULL;
1340 if (onlyjump_p (i1)
1341 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1343 last1 = i1;
1344 i1 = PREV_INSN (i1);
1347 i2 = BB_END (bb2);
1348 if (onlyjump_p (i2)
1349 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1351 last2 = i2;
1352 /* Count everything except for unconditional jump as insn.
1353 Don't count any jumps if dir_p is NULL. */
1354 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1355 ninsns++;
1356 i2 = PREV_INSN (i2);
1359 while (true)
1361 /* In the following example, we can replace all jumps to C by jumps to A.
1363 This removes 4 duplicate insns.
1364 [bb A] insn1 [bb C] insn1
1365 insn2 insn2
1366 [bb B] insn3 insn3
1367 insn4 insn4
1368 jump_insn jump_insn
1370 We could also replace all jumps to A by jumps to C, but that leaves B
1371 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1372 step, all jumps to B would be replaced with jumps to the middle of C,
1373 achieving the same result with more effort.
1374 So we allow only the first possibility, which means that we don't allow
1375 fallthru in the block that's being replaced. */
1377 follow_fallthru = dir_p && dir != dir_forward;
1378 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1379 if (did_fallthru)
1380 dir = dir_backward;
1382 follow_fallthru = dir_p && dir != dir_backward;
1383 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1384 if (did_fallthru)
1385 dir = dir_forward;
1387 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1388 break;
1390 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1391 if (dir == dir_none || (!dir_p && dir != dir_both))
1392 break;
1394 merge_memattrs (i1, i2);
1396 /* Don't begin a cross-jump with a NOTE insn. */
1397 if (INSN_P (i1))
1399 merge_notes (i1, i2);
1401 afterlast1 = last1, afterlast2 = last2;
1402 last1 = i1, last2 = i2;
1403 afterlast_dir = last_dir;
1404 last_dir = dir;
1405 if (active_insn_p (i1))
1406 ninsns++;
1409 i1 = PREV_INSN (i1);
1410 i2 = PREV_INSN (i2);
1413 /* Don't allow the insn after a compare to be shared by
1414 cross-jumping unless the compare is also shared. */
1415 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1416 && ! sets_cc0_p (last1))
1417 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1419 /* Include preceding notes and labels in the cross-jump. One,
1420 this may bring us to the head of the blocks as requested above.
1421 Two, it keeps line number notes as matched as may be. */
1422 if (ninsns)
1424 bb1 = BLOCK_FOR_INSN (last1);
1425 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1426 last1 = PREV_INSN (last1);
1428 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1429 last1 = PREV_INSN (last1);
1431 bb2 = BLOCK_FOR_INSN (last2);
1432 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1433 last2 = PREV_INSN (last2);
1435 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1436 last2 = PREV_INSN (last2);
1438 *f1 = last1;
1439 *f2 = last2;
1442 if (dir_p)
1443 *dir_p = last_dir;
1444 return ninsns;
1447 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1448 the head of the two blocks. Do not include jumps at the end.
1449 If STOP_AFTER is nonzero, stop after finding that many matching
1450 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1451 non-zero, only count active insns. */
1454 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1455 rtx_insn **f2, int stop_after)
1457 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1458 int ninsns = 0;
1459 edge e;
1460 edge_iterator ei;
1461 int nehedges1 = 0, nehedges2 = 0;
1463 FOR_EACH_EDGE (e, ei, bb1->succs)
1464 if (e->flags & EDGE_EH)
1465 nehedges1++;
1466 FOR_EACH_EDGE (e, ei, bb2->succs)
1467 if (e->flags & EDGE_EH)
1468 nehedges2++;
1470 i1 = BB_HEAD (bb1);
1471 i2 = BB_HEAD (bb2);
1472 last1 = beforelast1 = last2 = beforelast2 = NULL;
1474 while (true)
1476 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1477 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1479 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1480 break;
1481 i1 = NEXT_INSN (i1);
1484 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1486 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1487 break;
1488 i2 = NEXT_INSN (i2);
1491 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1492 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1493 break;
1495 if (NOTE_P (i1) || NOTE_P (i2)
1496 || JUMP_P (i1) || JUMP_P (i2))
1497 break;
1499 /* A sanity check to make sure we're not merging insns with different
1500 effects on EH. If only one of them ends a basic block, it shouldn't
1501 have an EH edge; if both end a basic block, there should be the same
1502 number of EH edges. */
1503 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1504 && nehedges1 > 0)
1505 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1506 && nehedges2 > 0)
1507 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1508 && nehedges1 != nehedges2))
1509 break;
1511 if (old_insns_match_p (0, i1, i2) != dir_both)
1512 break;
1514 merge_memattrs (i1, i2);
1516 /* Don't begin a cross-jump with a NOTE insn. */
1517 if (INSN_P (i1))
1519 merge_notes (i1, i2);
1521 beforelast1 = last1, beforelast2 = last2;
1522 last1 = i1, last2 = i2;
1523 if (!stop_after || active_insn_p (i1))
1524 ninsns++;
1527 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1528 || (stop_after > 0 && ninsns == stop_after))
1529 break;
1531 i1 = NEXT_INSN (i1);
1532 i2 = NEXT_INSN (i2);
1535 /* Don't allow a compare to be shared by cross-jumping unless the insn
1536 after the compare is also shared. */
1537 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1538 && sets_cc0_p (last1))
1539 last1 = beforelast1, last2 = beforelast2, ninsns--;
1541 if (ninsns)
1543 *f1 = last1;
1544 *f2 = last2;
1547 return ninsns;
1550 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1551 the branch instruction. This means that if we commonize the control
1552 flow before end of the basic block, the semantic remains unchanged.
1554 We may assume that there exists one edge with a common destination. */
1556 static bool
1557 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1559 int nehedges1 = 0, nehedges2 = 0;
1560 edge fallthru1 = 0, fallthru2 = 0;
1561 edge e1, e2;
1562 edge_iterator ei;
1564 /* If we performed shrink-wrapping, edges to the exit block can
1565 only be distinguished for JUMP_INSNs. The two paths may differ in
1566 whether they went through the prologue. Sibcalls are fine, we know
1567 that we either didn't need or inserted an epilogue before them. */
1568 if (crtl->shrink_wrapped
1569 && single_succ_p (bb1)
1570 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1571 && !JUMP_P (BB_END (bb1))
1572 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1573 return false;
1575 /* If BB1 has only one successor, we may be looking at either an
1576 unconditional jump, or a fake edge to exit. */
1577 if (single_succ_p (bb1)
1578 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1579 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1580 return (single_succ_p (bb2)
1581 && (single_succ_edge (bb2)->flags
1582 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1583 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1585 /* Match conditional jumps - this may get tricky when fallthru and branch
1586 edges are crossed. */
1587 if (EDGE_COUNT (bb1->succs) == 2
1588 && any_condjump_p (BB_END (bb1))
1589 && onlyjump_p (BB_END (bb1)))
1591 edge b1, f1, b2, f2;
1592 bool reverse, match;
1593 rtx set1, set2, cond1, cond2;
1594 enum rtx_code code1, code2;
1596 if (EDGE_COUNT (bb2->succs) != 2
1597 || !any_condjump_p (BB_END (bb2))
1598 || !onlyjump_p (BB_END (bb2)))
1599 return false;
1601 b1 = BRANCH_EDGE (bb1);
1602 b2 = BRANCH_EDGE (bb2);
1603 f1 = FALLTHRU_EDGE (bb1);
1604 f2 = FALLTHRU_EDGE (bb2);
1606 /* Get around possible forwarders on fallthru edges. Other cases
1607 should be optimized out already. */
1608 if (FORWARDER_BLOCK_P (f1->dest))
1609 f1 = single_succ_edge (f1->dest);
1611 if (FORWARDER_BLOCK_P (f2->dest))
1612 f2 = single_succ_edge (f2->dest);
1614 /* To simplify use of this function, return false if there are
1615 unneeded forwarder blocks. These will get eliminated later
1616 during cleanup_cfg. */
1617 if (FORWARDER_BLOCK_P (f1->dest)
1618 || FORWARDER_BLOCK_P (f2->dest)
1619 || FORWARDER_BLOCK_P (b1->dest)
1620 || FORWARDER_BLOCK_P (b2->dest))
1621 return false;
1623 if (f1->dest == f2->dest && b1->dest == b2->dest)
1624 reverse = false;
1625 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1626 reverse = true;
1627 else
1628 return false;
1630 set1 = pc_set (BB_END (bb1));
1631 set2 = pc_set (BB_END (bb2));
1632 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1633 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1634 reverse = !reverse;
1636 cond1 = XEXP (SET_SRC (set1), 0);
1637 cond2 = XEXP (SET_SRC (set2), 0);
1638 code1 = GET_CODE (cond1);
1639 if (reverse)
1640 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1641 else
1642 code2 = GET_CODE (cond2);
1644 if (code2 == UNKNOWN)
1645 return false;
1647 /* Verify codes and operands match. */
1648 match = ((code1 == code2
1649 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1650 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1651 || (code1 == swap_condition (code2)
1652 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1653 XEXP (cond2, 0))
1654 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1655 XEXP (cond2, 1))));
1657 /* If we return true, we will join the blocks. Which means that
1658 we will only have one branch prediction bit to work with. Thus
1659 we require the existing branches to have probabilities that are
1660 roughly similar. */
1661 if (match
1662 && optimize_bb_for_speed_p (bb1)
1663 && optimize_bb_for_speed_p (bb2))
1665 int prob2;
1667 if (b1->dest == b2->dest)
1668 prob2 = b2->probability;
1669 else
1670 /* Do not use f2 probability as f2 may be forwarded. */
1671 prob2 = REG_BR_PROB_BASE - b2->probability;
1673 /* Fail if the difference in probabilities is greater than 50%.
1674 This rules out two well-predicted branches with opposite
1675 outcomes. */
1676 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1678 if (dump_file)
1679 fprintf (dump_file,
1680 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1681 bb1->index, bb2->index, b1->probability, prob2);
1683 return false;
1687 if (dump_file && match)
1688 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1689 bb1->index, bb2->index);
1691 return match;
1694 /* Generic case - we are seeing a computed jump, table jump or trapping
1695 instruction. */
1697 /* Check whether there are tablejumps in the end of BB1 and BB2.
1698 Return true if they are identical. */
1700 rtx_insn *label1, *label2;
1701 rtx_jump_table_data *table1, *table2;
1703 if (tablejump_p (BB_END (bb1), &label1, &table1)
1704 && tablejump_p (BB_END (bb2), &label2, &table2)
1705 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1707 /* The labels should never be the same rtx. If they really are same
1708 the jump tables are same too. So disable crossjumping of blocks BB1
1709 and BB2 because when deleting the common insns in the end of BB1
1710 by delete_basic_block () the jump table would be deleted too. */
1711 /* If LABEL2 is referenced in BB1->END do not do anything
1712 because we would loose information when replacing
1713 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1714 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1716 /* Set IDENTICAL to true when the tables are identical. */
1717 bool identical = false;
1718 rtx p1, p2;
1720 p1 = PATTERN (table1);
1721 p2 = PATTERN (table2);
1722 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1724 identical = true;
1726 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1727 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1728 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1729 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1731 int i;
1733 identical = true;
1734 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1735 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1736 identical = false;
1739 if (identical)
1741 bool match;
1743 /* Temporarily replace references to LABEL1 with LABEL2
1744 in BB1->END so that we could compare the instructions. */
1745 replace_label_in_insn (BB_END (bb1), label1, label2, false);
1747 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1748 == dir_both);
1749 if (dump_file && match)
1750 fprintf (dump_file,
1751 "Tablejumps in bb %i and %i match.\n",
1752 bb1->index, bb2->index);
1754 /* Set the original label in BB1->END because when deleting
1755 a block whose end is a tablejump, the tablejump referenced
1756 from the instruction is deleted too. */
1757 replace_label_in_insn (BB_END (bb1), label2, label1, false);
1759 return match;
1762 return false;
1766 /* Find the last non-debug non-note instruction in each bb, except
1767 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1768 handles that case specially. old_insns_match_p does not handle
1769 other types of instruction notes. */
1770 rtx_insn *last1 = BB_END (bb1);
1771 rtx_insn *last2 = BB_END (bb2);
1772 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1773 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1774 last1 = PREV_INSN (last1);
1775 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1776 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1777 last2 = PREV_INSN (last2);
1778 gcc_assert (last1 && last2);
1780 /* First ensure that the instructions match. There may be many outgoing
1781 edges so this test is generally cheaper. */
1782 if (old_insns_match_p (mode, last1, last2) != dir_both)
1783 return false;
1785 /* Search the outgoing edges, ensure that the counts do match, find possible
1786 fallthru and exception handling edges since these needs more
1787 validation. */
1788 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1789 return false;
1791 bool nonfakeedges = false;
1792 FOR_EACH_EDGE (e1, ei, bb1->succs)
1794 e2 = EDGE_SUCC (bb2, ei.index);
1796 if ((e1->flags & EDGE_FAKE) == 0)
1797 nonfakeedges = true;
1799 if (e1->flags & EDGE_EH)
1800 nehedges1++;
1802 if (e2->flags & EDGE_EH)
1803 nehedges2++;
1805 if (e1->flags & EDGE_FALLTHRU)
1806 fallthru1 = e1;
1807 if (e2->flags & EDGE_FALLTHRU)
1808 fallthru2 = e2;
1811 /* If number of edges of various types does not match, fail. */
1812 if (nehedges1 != nehedges2
1813 || (fallthru1 != 0) != (fallthru2 != 0))
1814 return false;
1816 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1817 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1818 attempt to optimize, as the two basic blocks might have different
1819 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1820 traps there should be REG_ARG_SIZE notes, they could be missing
1821 for __builtin_unreachable () uses though. */
1822 if (!nonfakeedges
1823 && !ACCUMULATE_OUTGOING_ARGS
1824 && (!INSN_P (last1)
1825 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1826 return false;
1828 /* fallthru edges must be forwarded to the same destination. */
1829 if (fallthru1)
1831 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1832 ? single_succ (fallthru1->dest): fallthru1->dest);
1833 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1834 ? single_succ (fallthru2->dest): fallthru2->dest);
1836 if (d1 != d2)
1837 return false;
1840 /* Ensure the same EH region. */
1842 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1843 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1845 if (!n1 && n2)
1846 return false;
1848 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1849 return false;
1852 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1853 version of sequence abstraction. */
1854 FOR_EACH_EDGE (e1, ei, bb2->succs)
1856 edge e2;
1857 edge_iterator ei;
1858 basic_block d1 = e1->dest;
1860 if (FORWARDER_BLOCK_P (d1))
1861 d1 = EDGE_SUCC (d1, 0)->dest;
1863 FOR_EACH_EDGE (e2, ei, bb1->succs)
1865 basic_block d2 = e2->dest;
1866 if (FORWARDER_BLOCK_P (d2))
1867 d2 = EDGE_SUCC (d2, 0)->dest;
1868 if (d1 == d2)
1869 break;
1872 if (!e2)
1873 return false;
1876 return true;
1879 /* Returns true if BB basic block has a preserve label. */
1881 static bool
1882 block_has_preserve_label (basic_block bb)
1884 return (bb
1885 && block_label (bb)
1886 && LABEL_PRESERVE_P (block_label (bb)));
1889 /* E1 and E2 are edges with the same destination block. Search their
1890 predecessors for common code. If found, redirect control flow from
1891 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1892 or the other way around (dir_backward). DIR specifies the allowed
1893 replacement direction. */
1895 static bool
1896 try_crossjump_to_edge (int mode, edge e1, edge e2,
1897 enum replace_direction dir)
1899 int nmatch;
1900 basic_block src1 = e1->src, src2 = e2->src;
1901 basic_block redirect_to, redirect_from, to_remove;
1902 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1903 rtx_insn *newpos1, *newpos2;
1904 edge s;
1905 edge_iterator ei;
1907 newpos1 = newpos2 = NULL;
1909 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1910 to try this optimization.
1912 Basic block partitioning may result in some jumps that appear to
1913 be optimizable (or blocks that appear to be mergeable), but which really
1914 must be left untouched (they are required to make it safely across
1915 partition boundaries). See the comments at the top of
1916 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1918 if (crtl->has_bb_partition && reload_completed)
1919 return false;
1921 /* Search backward through forwarder blocks. We don't need to worry
1922 about multiple entry or chained forwarders, as they will be optimized
1923 away. We do this to look past the unconditional jump following a
1924 conditional jump that is required due to the current CFG shape. */
1925 if (single_pred_p (src1)
1926 && FORWARDER_BLOCK_P (src1))
1927 e1 = single_pred_edge (src1), src1 = e1->src;
1929 if (single_pred_p (src2)
1930 && FORWARDER_BLOCK_P (src2))
1931 e2 = single_pred_edge (src2), src2 = e2->src;
1933 /* Nothing to do if we reach ENTRY, or a common source block. */
1934 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1935 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1936 return false;
1937 if (src1 == src2)
1938 return false;
1940 /* Seeing more than 1 forwarder blocks would confuse us later... */
1941 if (FORWARDER_BLOCK_P (e1->dest)
1942 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1943 return false;
1945 if (FORWARDER_BLOCK_P (e2->dest)
1946 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1947 return false;
1949 /* Likewise with dead code (possibly newly created by the other optimizations
1950 of cfg_cleanup). */
1951 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1952 return false;
1954 /* Look for the common insn sequence, part the first ... */
1955 if (!outgoing_edges_match (mode, src1, src2))
1956 return false;
1958 /* ... and part the second. */
1959 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1961 osrc1 = src1;
1962 osrc2 = src2;
1963 if (newpos1 != NULL_RTX)
1964 src1 = BLOCK_FOR_INSN (newpos1);
1965 if (newpos2 != NULL_RTX)
1966 src2 = BLOCK_FOR_INSN (newpos2);
1968 if (dir == dir_backward)
1970 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1971 SWAP (basic_block, osrc1, osrc2);
1972 SWAP (basic_block, src1, src2);
1973 SWAP (edge, e1, e2);
1974 SWAP (rtx_insn *, newpos1, newpos2);
1975 #undef SWAP
1978 /* Don't proceed with the crossjump unless we found a sufficient number
1979 of matching instructions or the 'from' block was totally matched
1980 (such that its predecessors will hopefully be redirected and the
1981 block removed). */
1982 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
1983 && (newpos1 != BB_HEAD (src1)))
1984 return false;
1986 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1987 if (block_has_preserve_label (e1->dest)
1988 && (e1->flags & EDGE_ABNORMAL))
1989 return false;
1991 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1992 will be deleted.
1993 If we have tablejumps in the end of SRC1 and SRC2
1994 they have been already compared for equivalence in outgoing_edges_match ()
1995 so replace the references to TABLE1 by references to TABLE2. */
1997 rtx_insn *label1, *label2;
1998 rtx_jump_table_data *table1, *table2;
2000 if (tablejump_p (BB_END (osrc1), &label1, &table1)
2001 && tablejump_p (BB_END (osrc2), &label2, &table2)
2002 && label1 != label2)
2004 rtx_insn *insn;
2006 /* Replace references to LABEL1 with LABEL2. */
2007 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2009 /* Do not replace the label in SRC1->END because when deleting
2010 a block whose end is a tablejump, the tablejump referenced
2011 from the instruction is deleted too. */
2012 if (insn != BB_END (osrc1))
2013 replace_label_in_insn (insn, label1, label2, true);
2018 /* Avoid splitting if possible. We must always split when SRC2 has
2019 EH predecessor edges, or we may end up with basic blocks with both
2020 normal and EH predecessor edges. */
2021 if (newpos2 == BB_HEAD (src2)
2022 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2023 redirect_to = src2;
2024 else
2026 if (newpos2 == BB_HEAD (src2))
2028 /* Skip possible basic block header. */
2029 if (LABEL_P (newpos2))
2030 newpos2 = NEXT_INSN (newpos2);
2031 while (DEBUG_INSN_P (newpos2))
2032 newpos2 = NEXT_INSN (newpos2);
2033 if (NOTE_P (newpos2))
2034 newpos2 = NEXT_INSN (newpos2);
2035 while (DEBUG_INSN_P (newpos2))
2036 newpos2 = NEXT_INSN (newpos2);
2039 if (dump_file)
2040 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2041 src2->index, nmatch);
2042 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2045 if (dump_file)
2046 fprintf (dump_file,
2047 "Cross jumping from bb %i to bb %i; %i common insns\n",
2048 src1->index, src2->index, nmatch);
2050 /* We may have some registers visible through the block. */
2051 df_set_bb_dirty (redirect_to);
2053 if (osrc2 == src2)
2054 redirect_edges_to = redirect_to;
2055 else
2056 redirect_edges_to = osrc2;
2058 /* Recompute the frequencies and counts of outgoing edges. */
2059 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2061 edge s2;
2062 edge_iterator ei;
2063 basic_block d = s->dest;
2065 if (FORWARDER_BLOCK_P (d))
2066 d = single_succ (d);
2068 FOR_EACH_EDGE (s2, ei, src1->succs)
2070 basic_block d2 = s2->dest;
2071 if (FORWARDER_BLOCK_P (d2))
2072 d2 = single_succ (d2);
2073 if (d == d2)
2074 break;
2077 s->count += s2->count;
2079 /* Take care to update possible forwarder blocks. We verified
2080 that there is no more than one in the chain, so we can't run
2081 into infinite loop. */
2082 if (FORWARDER_BLOCK_P (s->dest))
2084 single_succ_edge (s->dest)->count += s2->count;
2085 s->dest->count += s2->count;
2086 s->dest->frequency += EDGE_FREQUENCY (s);
2089 if (FORWARDER_BLOCK_P (s2->dest))
2091 single_succ_edge (s2->dest)->count -= s2->count;
2092 if (single_succ_edge (s2->dest)->count < 0)
2093 single_succ_edge (s2->dest)->count = 0;
2094 s2->dest->count -= s2->count;
2095 s2->dest->frequency -= EDGE_FREQUENCY (s);
2096 if (s2->dest->frequency < 0)
2097 s2->dest->frequency = 0;
2098 if (s2->dest->count < 0)
2099 s2->dest->count = 0;
2102 if (!redirect_edges_to->frequency && !src1->frequency)
2103 s->probability = (s->probability + s2->probability) / 2;
2104 else
2105 s->probability
2106 = ((s->probability * redirect_edges_to->frequency +
2107 s2->probability * src1->frequency)
2108 / (redirect_edges_to->frequency + src1->frequency));
2111 /* Adjust count and frequency for the block. An earlier jump
2112 threading pass may have left the profile in an inconsistent
2113 state (see update_bb_profile_for_threading) so we must be
2114 prepared for overflows. */
2115 tmp = redirect_to;
2118 tmp->count += src1->count;
2119 tmp->frequency += src1->frequency;
2120 if (tmp->frequency > BB_FREQ_MAX)
2121 tmp->frequency = BB_FREQ_MAX;
2122 if (tmp == redirect_edges_to)
2123 break;
2124 tmp = find_fallthru_edge (tmp->succs)->dest;
2126 while (true);
2127 update_br_prob_note (redirect_edges_to);
2129 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2131 /* Skip possible basic block header. */
2132 if (LABEL_P (newpos1))
2133 newpos1 = NEXT_INSN (newpos1);
2135 while (DEBUG_INSN_P (newpos1))
2136 newpos1 = NEXT_INSN (newpos1);
2138 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2139 newpos1 = NEXT_INSN (newpos1);
2141 while (DEBUG_INSN_P (newpos1))
2142 newpos1 = NEXT_INSN (newpos1);
2144 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2145 to_remove = single_succ (redirect_from);
2147 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2148 delete_basic_block (to_remove);
2150 update_forwarder_flag (redirect_from);
2151 if (redirect_to != src2)
2152 update_forwarder_flag (src2);
2154 return true;
2157 /* Search the predecessors of BB for common insn sequences. When found,
2158 share code between them by redirecting control flow. Return true if
2159 any changes made. */
2161 static bool
2162 try_crossjump_bb (int mode, basic_block bb)
2164 edge e, e2, fallthru;
2165 bool changed;
2166 unsigned max, ix, ix2;
2168 /* Nothing to do if there is not at least two incoming edges. */
2169 if (EDGE_COUNT (bb->preds) < 2)
2170 return false;
2172 /* Don't crossjump if this block ends in a computed jump,
2173 unless we are optimizing for size. */
2174 if (optimize_bb_for_size_p (bb)
2175 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2176 && computed_jump_p (BB_END (bb)))
2177 return false;
2179 /* If we are partitioning hot/cold basic blocks, we don't want to
2180 mess up unconditional or indirect jumps that cross between hot
2181 and cold sections.
2183 Basic block partitioning may result in some jumps that appear to
2184 be optimizable (or blocks that appear to be mergeable), but which really
2185 must be left untouched (they are required to make it safely across
2186 partition boundaries). See the comments at the top of
2187 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2189 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2190 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2191 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2192 return false;
2194 /* It is always cheapest to redirect a block that ends in a branch to
2195 a block that falls through into BB, as that adds no branches to the
2196 program. We'll try that combination first. */
2197 fallthru = NULL;
2198 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2200 if (EDGE_COUNT (bb->preds) > max)
2201 return false;
2203 fallthru = find_fallthru_edge (bb->preds);
2205 changed = false;
2206 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2208 e = EDGE_PRED (bb, ix);
2209 ix++;
2211 /* As noted above, first try with the fallthru predecessor (or, a
2212 fallthru predecessor if we are in cfglayout mode). */
2213 if (fallthru)
2215 /* Don't combine the fallthru edge into anything else.
2216 If there is a match, we'll do it the other way around. */
2217 if (e == fallthru)
2218 continue;
2219 /* If nothing changed since the last attempt, there is nothing
2220 we can do. */
2221 if (!first_pass
2222 && !((e->src->flags & BB_MODIFIED)
2223 || (fallthru->src->flags & BB_MODIFIED)))
2224 continue;
2226 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2228 changed = true;
2229 ix = 0;
2230 continue;
2234 /* Non-obvious work limiting check: Recognize that we're going
2235 to call try_crossjump_bb on every basic block. So if we have
2236 two blocks with lots of outgoing edges (a switch) and they
2237 share lots of common destinations, then we would do the
2238 cross-jump check once for each common destination.
2240 Now, if the blocks actually are cross-jump candidates, then
2241 all of their destinations will be shared. Which means that
2242 we only need check them for cross-jump candidacy once. We
2243 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2244 choosing to do the check from the block for which the edge
2245 in question is the first successor of A. */
2246 if (EDGE_SUCC (e->src, 0) != e)
2247 continue;
2249 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2251 e2 = EDGE_PRED (bb, ix2);
2253 if (e2 == e)
2254 continue;
2256 /* We've already checked the fallthru edge above. */
2257 if (e2 == fallthru)
2258 continue;
2260 /* The "first successor" check above only prevents multiple
2261 checks of crossjump(A,B). In order to prevent redundant
2262 checks of crossjump(B,A), require that A be the block
2263 with the lowest index. */
2264 if (e->src->index > e2->src->index)
2265 continue;
2267 /* If nothing changed since the last attempt, there is nothing
2268 we can do. */
2269 if (!first_pass
2270 && !((e->src->flags & BB_MODIFIED)
2271 || (e2->src->flags & BB_MODIFIED)))
2272 continue;
2274 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2275 direction. */
2276 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2278 changed = true;
2279 ix = 0;
2280 break;
2285 if (changed)
2286 crossjumps_occurred = true;
2288 return changed;
2291 /* Search the successors of BB for common insn sequences. When found,
2292 share code between them by moving it across the basic block
2293 boundary. Return true if any changes made. */
2295 static bool
2296 try_head_merge_bb (basic_block bb)
2298 basic_block final_dest_bb = NULL;
2299 int max_match = INT_MAX;
2300 edge e0;
2301 rtx_insn **headptr, **currptr, **nextptr;
2302 bool changed, moveall;
2303 unsigned ix;
2304 rtx_insn *e0_last_head;
2305 rtx cond;
2306 rtx_insn *move_before;
2307 unsigned nedges = EDGE_COUNT (bb->succs);
2308 rtx_insn *jump = BB_END (bb);
2309 regset live, live_union;
2311 /* Nothing to do if there is not at least two outgoing edges. */
2312 if (nedges < 2)
2313 return false;
2315 /* Don't crossjump if this block ends in a computed jump,
2316 unless we are optimizing for size. */
2317 if (optimize_bb_for_size_p (bb)
2318 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2319 && computed_jump_p (BB_END (bb)))
2320 return false;
2322 cond = get_condition (jump, &move_before, true, false);
2323 if (cond == NULL_RTX)
2325 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2326 move_before = prev_nonnote_nondebug_insn (jump);
2327 else
2328 move_before = jump;
2331 for (ix = 0; ix < nedges; ix++)
2332 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2333 return false;
2335 for (ix = 0; ix < nedges; ix++)
2337 edge e = EDGE_SUCC (bb, ix);
2338 basic_block other_bb = e->dest;
2340 if (df_get_bb_dirty (other_bb))
2342 block_was_dirty = true;
2343 return false;
2346 if (e->flags & EDGE_ABNORMAL)
2347 return false;
2349 /* Normally, all destination blocks must only be reachable from this
2350 block, i.e. they must have one incoming edge.
2352 There is one special case we can handle, that of multiple consecutive
2353 jumps where the first jumps to one of the targets of the second jump.
2354 This happens frequently in switch statements for default labels.
2355 The structure is as follows:
2356 FINAL_DEST_BB
2357 ....
2358 if (cond) jump A;
2359 fall through
2361 jump with targets A, B, C, D...
2363 has two incoming edges, from FINAL_DEST_BB and BB
2365 In this case, we can try to move the insns through BB and into
2366 FINAL_DEST_BB. */
2367 if (EDGE_COUNT (other_bb->preds) != 1)
2369 edge incoming_edge, incoming_bb_other_edge;
2370 edge_iterator ei;
2372 if (final_dest_bb != NULL
2373 || EDGE_COUNT (other_bb->preds) != 2)
2374 return false;
2376 /* We must be able to move the insns across the whole block. */
2377 move_before = BB_HEAD (bb);
2378 while (!NONDEBUG_INSN_P (move_before))
2379 move_before = NEXT_INSN (move_before);
2381 if (EDGE_COUNT (bb->preds) != 1)
2382 return false;
2383 incoming_edge = EDGE_PRED (bb, 0);
2384 final_dest_bb = incoming_edge->src;
2385 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2386 return false;
2387 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2388 if (incoming_bb_other_edge != incoming_edge)
2389 break;
2390 if (incoming_bb_other_edge->dest != other_bb)
2391 return false;
2395 e0 = EDGE_SUCC (bb, 0);
2396 e0_last_head = NULL;
2397 changed = false;
2399 for (ix = 1; ix < nedges; ix++)
2401 edge e = EDGE_SUCC (bb, ix);
2402 rtx_insn *e0_last, *e_last;
2403 int nmatch;
2405 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2406 &e0_last, &e_last, 0);
2407 if (nmatch == 0)
2408 return false;
2410 if (nmatch < max_match)
2412 max_match = nmatch;
2413 e0_last_head = e0_last;
2417 /* If we matched an entire block, we probably have to avoid moving the
2418 last insn. */
2419 if (max_match > 0
2420 && e0_last_head == BB_END (e0->dest)
2421 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2422 || control_flow_insn_p (e0_last_head)))
2424 max_match--;
2425 if (max_match == 0)
2426 return false;
2428 e0_last_head = prev_real_insn (e0_last_head);
2429 while (DEBUG_INSN_P (e0_last_head));
2432 if (max_match == 0)
2433 return false;
2435 /* We must find a union of the live registers at each of the end points. */
2436 live = BITMAP_ALLOC (NULL);
2437 live_union = BITMAP_ALLOC (NULL);
2439 currptr = XNEWVEC (rtx_insn *, nedges);
2440 headptr = XNEWVEC (rtx_insn *, nedges);
2441 nextptr = XNEWVEC (rtx_insn *, nedges);
2443 for (ix = 0; ix < nedges; ix++)
2445 int j;
2446 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2447 rtx_insn *head = BB_HEAD (merge_bb);
2449 while (!NONDEBUG_INSN_P (head))
2450 head = NEXT_INSN (head);
2451 headptr[ix] = head;
2452 currptr[ix] = head;
2454 /* Compute the end point and live information */
2455 for (j = 1; j < max_match; j++)
2457 head = NEXT_INSN (head);
2458 while (!NONDEBUG_INSN_P (head));
2459 simulate_backwards_to_point (merge_bb, live, head);
2460 IOR_REG_SET (live_union, live);
2463 /* If we're moving across two blocks, verify the validity of the
2464 first move, then adjust the target and let the loop below deal
2465 with the final move. */
2466 if (final_dest_bb != NULL)
2468 rtx_insn *move_upto;
2470 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2471 jump, e0->dest, live_union,
2472 NULL, &move_upto);
2473 if (!moveall)
2475 if (move_upto == NULL_RTX)
2476 goto out;
2478 while (e0_last_head != move_upto)
2480 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2481 live_union);
2482 e0_last_head = PREV_INSN (e0_last_head);
2485 if (e0_last_head == NULL_RTX)
2486 goto out;
2488 jump = BB_END (final_dest_bb);
2489 cond = get_condition (jump, &move_before, true, false);
2490 if (cond == NULL_RTX)
2492 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2493 move_before = prev_nonnote_nondebug_insn (jump);
2494 else
2495 move_before = jump;
2501 rtx_insn *move_upto;
2502 moveall = can_move_insns_across (currptr[0], e0_last_head,
2503 move_before, jump, e0->dest, live_union,
2504 NULL, &move_upto);
2505 if (!moveall && move_upto == NULL_RTX)
2507 if (jump == move_before)
2508 break;
2510 /* Try again, using a different insertion point. */
2511 move_before = jump;
2513 /* Don't try moving before a cc0 user, as that may invalidate
2514 the cc0. */
2515 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2516 break;
2518 continue;
2521 if (final_dest_bb && !moveall)
2522 /* We haven't checked whether a partial move would be OK for the first
2523 move, so we have to fail this case. */
2524 break;
2526 changed = true;
2527 for (;;)
2529 if (currptr[0] == move_upto)
2530 break;
2531 for (ix = 0; ix < nedges; ix++)
2533 rtx_insn *curr = currptr[ix];
2535 curr = NEXT_INSN (curr);
2536 while (!NONDEBUG_INSN_P (curr));
2537 currptr[ix] = curr;
2541 /* If we can't currently move all of the identical insns, remember
2542 each insn after the range that we'll merge. */
2543 if (!moveall)
2544 for (ix = 0; ix < nedges; ix++)
2546 rtx_insn *curr = currptr[ix];
2548 curr = NEXT_INSN (curr);
2549 while (!NONDEBUG_INSN_P (curr));
2550 nextptr[ix] = curr;
2553 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2554 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2555 if (final_dest_bb != NULL)
2556 df_set_bb_dirty (final_dest_bb);
2557 df_set_bb_dirty (bb);
2558 for (ix = 1; ix < nedges; ix++)
2560 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2561 delete_insn_chain (headptr[ix], currptr[ix], false);
2563 if (!moveall)
2565 if (jump == move_before)
2566 break;
2568 /* For the unmerged insns, try a different insertion point. */
2569 move_before = jump;
2571 /* Don't try moving before a cc0 user, as that may invalidate
2572 the cc0. */
2573 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2574 break;
2576 for (ix = 0; ix < nedges; ix++)
2577 currptr[ix] = headptr[ix] = nextptr[ix];
2580 while (!moveall);
2582 out:
2583 free (currptr);
2584 free (headptr);
2585 free (nextptr);
2587 crossjumps_occurred |= changed;
2589 return changed;
2592 /* Return true if BB contains just bb note, or bb note followed
2593 by only DEBUG_INSNs. */
2595 static bool
2596 trivially_empty_bb_p (basic_block bb)
2598 rtx_insn *insn = BB_END (bb);
2600 while (1)
2602 if (insn == BB_HEAD (bb))
2603 return true;
2604 if (!DEBUG_INSN_P (insn))
2605 return false;
2606 insn = PREV_INSN (insn);
2610 /* Return true if BB contains just a return and possibly a USE of the
2611 return value. Fill in *RET and *USE with the return and use insns
2612 if any found, otherwise NULL. */
2614 static bool
2615 bb_is_just_return (basic_block bb, rtx_insn **ret, rtx_insn **use)
2617 *ret = *use = NULL;
2618 rtx_insn *insn;
2620 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
2621 return false;
2623 FOR_BB_INSNS (bb, insn)
2624 if (NONDEBUG_INSN_P (insn))
2626 if (!*ret && ANY_RETURN_P (PATTERN (insn)))
2627 *ret = insn;
2628 else if (!*ret && !*use && GET_CODE (PATTERN (insn)) == USE
2629 && REG_P (XEXP (PATTERN (insn), 0))
2630 && REG_FUNCTION_VALUE_P (XEXP (PATTERN (insn), 0)))
2631 *use = insn;
2632 else
2633 return false;
2636 return !!*ret;
2639 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2640 instructions etc. Return nonzero if changes were made. */
2642 static bool
2643 try_optimize_cfg (int mode)
2645 bool changed_overall = false;
2646 bool changed;
2647 int iterations = 0;
2648 basic_block bb, b, next;
2650 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2651 clear_bb_flags ();
2653 crossjumps_occurred = false;
2655 FOR_EACH_BB_FN (bb, cfun)
2656 update_forwarder_flag (bb);
2658 if (! targetm.cannot_modify_jumps_p ())
2660 first_pass = true;
2661 /* Attempt to merge blocks as made possible by edge removal. If
2662 a block has only one successor, and the successor has only
2663 one predecessor, they may be combined. */
2666 block_was_dirty = false;
2667 changed = false;
2668 iterations++;
2670 if (dump_file)
2671 fprintf (dump_file,
2672 "\n\ntry_optimize_cfg iteration %i\n\n",
2673 iterations);
2675 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2676 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2678 basic_block c;
2679 edge s;
2680 bool changed_here = false;
2682 /* Delete trivially dead basic blocks. This is either
2683 blocks with no predecessors, or empty blocks with no
2684 successors. However if the empty block with no
2685 successors is the successor of the ENTRY_BLOCK, it is
2686 kept. This ensures that the ENTRY_BLOCK will have a
2687 successor which is a precondition for many RTL
2688 passes. Empty blocks may result from expanding
2689 __builtin_unreachable (). */
2690 if (EDGE_COUNT (b->preds) == 0
2691 || (EDGE_COUNT (b->succs) == 0
2692 && trivially_empty_bb_p (b)
2693 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2694 != b))
2696 c = b->prev_bb;
2697 if (EDGE_COUNT (b->preds) > 0)
2699 edge e;
2700 edge_iterator ei;
2702 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2704 if (BB_FOOTER (b)
2705 && BARRIER_P (BB_FOOTER (b)))
2706 FOR_EACH_EDGE (e, ei, b->preds)
2707 if ((e->flags & EDGE_FALLTHRU)
2708 && BB_FOOTER (e->src) == NULL)
2710 if (BB_FOOTER (b))
2712 BB_FOOTER (e->src) = BB_FOOTER (b);
2713 BB_FOOTER (b) = NULL;
2715 else
2717 start_sequence ();
2718 BB_FOOTER (e->src) = emit_barrier ();
2719 end_sequence ();
2723 else
2725 rtx_insn *last = get_last_bb_insn (b);
2726 if (last && BARRIER_P (last))
2727 FOR_EACH_EDGE (e, ei, b->preds)
2728 if ((e->flags & EDGE_FALLTHRU))
2729 emit_barrier_after (BB_END (e->src));
2732 delete_basic_block (b);
2733 changed = true;
2734 /* Avoid trying to remove the exit block. */
2735 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2736 continue;
2739 /* Remove code labels no longer used. */
2740 if (single_pred_p (b)
2741 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2742 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2743 && LABEL_P (BB_HEAD (b))
2744 && !LABEL_PRESERVE_P (BB_HEAD (b))
2745 /* If the previous block ends with a branch to this
2746 block, we can't delete the label. Normally this
2747 is a condjump that is yet to be simplified, but
2748 if CASE_DROPS_THRU, this can be a tablejump with
2749 some element going to the same place as the
2750 default (fallthru). */
2751 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2752 || !JUMP_P (BB_END (single_pred (b)))
2753 || ! label_is_jump_target_p (BB_HEAD (b),
2754 BB_END (single_pred (b)))))
2756 delete_insn (BB_HEAD (b));
2757 if (dump_file)
2758 fprintf (dump_file, "Deleted label in block %i.\n",
2759 b->index);
2762 /* If we fall through an empty block, we can remove it. */
2763 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2764 && single_pred_p (b)
2765 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2766 && !LABEL_P (BB_HEAD (b))
2767 && FORWARDER_BLOCK_P (b)
2768 /* Note that forwarder_block_p true ensures that
2769 there is a successor for this block. */
2770 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2771 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2773 if (dump_file)
2774 fprintf (dump_file,
2775 "Deleting fallthru block %i.\n",
2776 b->index);
2778 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2779 ? b->next_bb : b->prev_bb);
2780 redirect_edge_succ_nodup (single_pred_edge (b),
2781 single_succ (b));
2782 delete_basic_block (b);
2783 changed = true;
2784 b = c;
2785 continue;
2788 /* Merge B with its single successor, if any. */
2789 if (single_succ_p (b)
2790 && (s = single_succ_edge (b))
2791 && !(s->flags & EDGE_COMPLEX)
2792 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2793 && single_pred_p (c)
2794 && b != c)
2796 /* When not in cfg_layout mode use code aware of reordering
2797 INSN. This code possibly creates new basic blocks so it
2798 does not fit merge_blocks interface and is kept here in
2799 hope that it will become useless once more of compiler
2800 is transformed to use cfg_layout mode. */
2802 if ((mode & CLEANUP_CFGLAYOUT)
2803 && can_merge_blocks_p (b, c))
2805 merge_blocks (b, c);
2806 update_forwarder_flag (b);
2807 changed_here = true;
2809 else if (!(mode & CLEANUP_CFGLAYOUT)
2810 /* If the jump insn has side effects,
2811 we can't kill the edge. */
2812 && (!JUMP_P (BB_END (b))
2813 || (reload_completed
2814 ? simplejump_p (BB_END (b))
2815 : (onlyjump_p (BB_END (b))
2816 && !tablejump_p (BB_END (b),
2817 NULL, NULL))))
2818 && (next = merge_blocks_move (s, b, c, mode)))
2820 b = next;
2821 changed_here = true;
2825 /* Try to change a branch to a return to just that return. */
2826 rtx_insn *ret, *use;
2827 if (single_succ_p (b)
2828 && onlyjump_p (BB_END (b))
2829 && bb_is_just_return (single_succ (b), &ret, &use))
2831 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2832 PATTERN (ret), 0))
2834 if (use)
2835 emit_insn_before (copy_insn (PATTERN (use)),
2836 BB_END (b));
2837 if (dump_file)
2838 fprintf (dump_file, "Changed jump %d->%d to return.\n",
2839 b->index, single_succ (b)->index);
2840 redirect_edge_succ (single_succ_edge (b),
2841 EXIT_BLOCK_PTR_FOR_FN (cfun));
2842 single_succ_edge (b)->flags &= ~EDGE_CROSSING;
2843 changed_here = true;
2847 /* Try to change a conditional branch to a return to the
2848 respective conditional return. */
2849 if (EDGE_COUNT (b->succs) == 2
2850 && any_condjump_p (BB_END (b))
2851 && bb_is_just_return (BRANCH_EDGE (b)->dest, &ret, &use))
2853 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2854 PATTERN (ret), 0))
2856 if (use)
2857 emit_insn_before (copy_insn (PATTERN (use)),
2858 BB_END (b));
2859 if (dump_file)
2860 fprintf (dump_file, "Changed conditional jump %d->%d "
2861 "to conditional return.\n",
2862 b->index, BRANCH_EDGE (b)->dest->index);
2863 redirect_edge_succ (BRANCH_EDGE (b),
2864 EXIT_BLOCK_PTR_FOR_FN (cfun));
2865 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2866 changed_here = true;
2870 /* Try to flip a conditional branch that falls through to
2871 a return so that it becomes a conditional return and a
2872 new jump to the original branch target. */
2873 if (EDGE_COUNT (b->succs) == 2
2874 && BRANCH_EDGE (b)->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
2875 && any_condjump_p (BB_END (b))
2876 && bb_is_just_return (FALLTHRU_EDGE (b)->dest, &ret, &use))
2878 if (invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2879 JUMP_LABEL (BB_END (b)), 0))
2881 basic_block new_ft = BRANCH_EDGE (b)->dest;
2882 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2883 PATTERN (ret), 0))
2885 if (use)
2886 emit_insn_before (copy_insn (PATTERN (use)),
2887 BB_END (b));
2888 if (dump_file)
2889 fprintf (dump_file, "Changed conditional jump "
2890 "%d->%d to conditional return, adding "
2891 "fall-through jump.\n",
2892 b->index, BRANCH_EDGE (b)->dest->index);
2893 redirect_edge_succ (BRANCH_EDGE (b),
2894 EXIT_BLOCK_PTR_FOR_FN (cfun));
2895 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2896 std::swap (BRANCH_EDGE (b)->probability,
2897 FALLTHRU_EDGE (b)->probability);
2898 update_br_prob_note (b);
2899 basic_block jb = force_nonfallthru (FALLTHRU_EDGE (b));
2900 notice_new_block (jb);
2901 if (!redirect_jump (as_a <rtx_jump_insn *> (BB_END (jb)),
2902 block_label (new_ft), 0))
2903 gcc_unreachable ();
2904 redirect_edge_succ (single_succ_edge (jb), new_ft);
2905 changed_here = true;
2907 else
2909 /* Invert the jump back to what it was. This should
2910 never fail. */
2911 if (!invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2912 JUMP_LABEL (BB_END (b)), 0))
2913 gcc_unreachable ();
2918 /* Simplify branch over branch. */
2919 if ((mode & CLEANUP_EXPENSIVE)
2920 && !(mode & CLEANUP_CFGLAYOUT)
2921 && try_simplify_condjump (b))
2922 changed_here = true;
2924 /* If B has a single outgoing edge, but uses a
2925 non-trivial jump instruction without side-effects, we
2926 can either delete the jump entirely, or replace it
2927 with a simple unconditional jump. */
2928 if (single_succ_p (b)
2929 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2930 && onlyjump_p (BB_END (b))
2931 && !CROSSING_JUMP_P (BB_END (b))
2932 && try_redirect_by_replacing_jump (single_succ_edge (b),
2933 single_succ (b),
2934 (mode & CLEANUP_CFGLAYOUT) != 0))
2936 update_forwarder_flag (b);
2937 changed_here = true;
2940 /* Simplify branch to branch. */
2941 if (try_forward_edges (mode, b))
2943 update_forwarder_flag (b);
2944 changed_here = true;
2947 /* Look for shared code between blocks. */
2948 if ((mode & CLEANUP_CROSSJUMP)
2949 && try_crossjump_bb (mode, b))
2950 changed_here = true;
2952 if ((mode & CLEANUP_CROSSJUMP)
2953 /* This can lengthen register lifetimes. Do it only after
2954 reload. */
2955 && reload_completed
2956 && try_head_merge_bb (b))
2957 changed_here = true;
2959 /* Don't get confused by the index shift caused by
2960 deleting blocks. */
2961 if (!changed_here)
2962 b = b->next_bb;
2963 else
2964 changed = true;
2967 if ((mode & CLEANUP_CROSSJUMP)
2968 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
2969 changed = true;
2971 if (block_was_dirty)
2973 /* This should only be set by head-merging. */
2974 gcc_assert (mode & CLEANUP_CROSSJUMP);
2975 df_analyze ();
2978 if (changed)
2980 /* Edge forwarding in particular can cause hot blocks previously
2981 reached by both hot and cold blocks to become dominated only
2982 by cold blocks. This will cause the verification below to fail,
2983 and lead to now cold code in the hot section. This is not easy
2984 to detect and fix during edge forwarding, and in some cases
2985 is only visible after newly unreachable blocks are deleted,
2986 which will be done in fixup_partitions. */
2987 fixup_partitions ();
2988 checking_verify_flow_info ();
2991 changed_overall |= changed;
2992 first_pass = false;
2994 while (changed);
2997 FOR_ALL_BB_FN (b, cfun)
2998 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
3000 return changed_overall;
3003 /* Delete all unreachable basic blocks. */
3005 bool
3006 delete_unreachable_blocks (void)
3008 bool changed = false;
3009 basic_block b, prev_bb;
3011 find_unreachable_blocks ();
3013 /* When we're in GIMPLE mode and there may be debug insns, we should
3014 delete blocks in reverse dominator order, so as to get a chance
3015 to substitute all released DEFs into debug stmts. If we don't
3016 have dominators information, walking blocks backward gets us a
3017 better chance of retaining most debug information than
3018 otherwise. */
3019 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
3020 && dom_info_available_p (CDI_DOMINATORS))
3022 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3023 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3025 prev_bb = b->prev_bb;
3027 if (!(b->flags & BB_REACHABLE))
3029 /* Speed up the removal of blocks that don't dominate
3030 others. Walking backwards, this should be the common
3031 case. */
3032 if (!first_dom_son (CDI_DOMINATORS, b))
3033 delete_basic_block (b);
3034 else
3036 vec<basic_block> h
3037 = get_all_dominated_blocks (CDI_DOMINATORS, b);
3039 while (h.length ())
3041 b = h.pop ();
3043 prev_bb = b->prev_bb;
3045 gcc_assert (!(b->flags & BB_REACHABLE));
3047 delete_basic_block (b);
3050 h.release ();
3053 changed = true;
3057 else
3059 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3060 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3062 prev_bb = b->prev_bb;
3064 if (!(b->flags & BB_REACHABLE))
3066 delete_basic_block (b);
3067 changed = true;
3072 if (changed)
3073 tidy_fallthru_edges ();
3074 return changed;
3077 /* Delete any jump tables never referenced. We can't delete them at the
3078 time of removing tablejump insn as they are referenced by the preceding
3079 insns computing the destination, so we delay deleting and garbagecollect
3080 them once life information is computed. */
3081 void
3082 delete_dead_jumptables (void)
3084 basic_block bb;
3086 /* A dead jump table does not belong to any basic block. Scan insns
3087 between two adjacent basic blocks. */
3088 FOR_EACH_BB_FN (bb, cfun)
3090 rtx_insn *insn, *next;
3092 for (insn = NEXT_INSN (BB_END (bb));
3093 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
3094 insn = next)
3096 next = NEXT_INSN (insn);
3097 if (LABEL_P (insn)
3098 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
3099 && JUMP_TABLE_DATA_P (next))
3101 rtx_insn *label = insn, *jump = next;
3103 if (dump_file)
3104 fprintf (dump_file, "Dead jumptable %i removed\n",
3105 INSN_UID (insn));
3107 next = NEXT_INSN (next);
3108 delete_insn (jump);
3109 delete_insn (label);
3116 /* Tidy the CFG by deleting unreachable code and whatnot. */
3118 bool
3119 cleanup_cfg (int mode)
3121 bool changed = false;
3123 /* Set the cfglayout mode flag here. We could update all the callers
3124 but that is just inconvenient, especially given that we eventually
3125 want to have cfglayout mode as the default. */
3126 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3127 mode |= CLEANUP_CFGLAYOUT;
3129 timevar_push (TV_CLEANUP_CFG);
3130 if (delete_unreachable_blocks ())
3132 changed = true;
3133 /* We've possibly created trivially dead code. Cleanup it right
3134 now to introduce more opportunities for try_optimize_cfg. */
3135 if (!(mode & (CLEANUP_NO_INSN_DEL))
3136 && !reload_completed)
3137 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3140 compact_blocks ();
3142 /* To tail-merge blocks ending in the same noreturn function (e.g.
3143 a call to abort) we have to insert fake edges to exit. Do this
3144 here once. The fake edges do not interfere with any other CFG
3145 cleanups. */
3146 if (mode & CLEANUP_CROSSJUMP)
3147 add_noreturn_fake_exit_edges ();
3149 if (!dbg_cnt (cfg_cleanup))
3150 return changed;
3152 while (try_optimize_cfg (mode))
3154 delete_unreachable_blocks (), changed = true;
3155 if (!(mode & CLEANUP_NO_INSN_DEL))
3157 /* Try to remove some trivially dead insns when doing an expensive
3158 cleanup. But delete_trivially_dead_insns doesn't work after
3159 reload (it only handles pseudos) and run_fast_dce is too costly
3160 to run in every iteration.
3162 For effective cross jumping, we really want to run a fast DCE to
3163 clean up any dead conditions, or they get in the way of performing
3164 useful tail merges.
3166 Other transformations in cleanup_cfg are not so sensitive to dead
3167 code, so delete_trivially_dead_insns or even doing nothing at all
3168 is good enough. */
3169 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3170 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3171 break;
3172 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occurred)
3173 run_fast_dce ();
3175 else
3176 break;
3179 if (mode & CLEANUP_CROSSJUMP)
3180 remove_fake_exit_edges ();
3182 /* Don't call delete_dead_jumptables in cfglayout mode, because
3183 that function assumes that jump tables are in the insns stream.
3184 But we also don't _have_ to delete dead jumptables in cfglayout
3185 mode because we shouldn't even be looking at things that are
3186 not in a basic block. Dead jumptables are cleaned up when
3187 going out of cfglayout mode. */
3188 if (!(mode & CLEANUP_CFGLAYOUT))
3189 delete_dead_jumptables ();
3191 /* ??? We probably do this way too often. */
3192 if (current_loops
3193 && (changed
3194 || (mode & CLEANUP_CFG_CHANGED)))
3196 timevar_push (TV_REPAIR_LOOPS);
3197 /* The above doesn't preserve dominance info if available. */
3198 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3199 calculate_dominance_info (CDI_DOMINATORS);
3200 fix_loop_structure (NULL);
3201 free_dominance_info (CDI_DOMINATORS);
3202 timevar_pop (TV_REPAIR_LOOPS);
3205 timevar_pop (TV_CLEANUP_CFG);
3207 return changed;
3210 namespace {
3212 const pass_data pass_data_jump =
3214 RTL_PASS, /* type */
3215 "jump", /* name */
3216 OPTGROUP_NONE, /* optinfo_flags */
3217 TV_JUMP, /* tv_id */
3218 0, /* properties_required */
3219 0, /* properties_provided */
3220 0, /* properties_destroyed */
3221 0, /* todo_flags_start */
3222 0, /* todo_flags_finish */
3225 class pass_jump : public rtl_opt_pass
3227 public:
3228 pass_jump (gcc::context *ctxt)
3229 : rtl_opt_pass (pass_data_jump, ctxt)
3232 /* opt_pass methods: */
3233 virtual unsigned int execute (function *);
3235 }; // class pass_jump
3237 unsigned int
3238 pass_jump::execute (function *)
3240 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3241 if (dump_file)
3242 dump_flow_info (dump_file, dump_flags);
3243 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3244 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3245 return 0;
3248 } // anon namespace
3250 rtl_opt_pass *
3251 make_pass_jump (gcc::context *ctxt)
3253 return new pass_jump (ctxt);
3256 namespace {
3258 const pass_data pass_data_jump2 =
3260 RTL_PASS, /* type */
3261 "jump2", /* name */
3262 OPTGROUP_NONE, /* optinfo_flags */
3263 TV_JUMP, /* tv_id */
3264 0, /* properties_required */
3265 0, /* properties_provided */
3266 0, /* properties_destroyed */
3267 0, /* todo_flags_start */
3268 0, /* todo_flags_finish */
3271 class pass_jump2 : public rtl_opt_pass
3273 public:
3274 pass_jump2 (gcc::context *ctxt)
3275 : rtl_opt_pass (pass_data_jump2, ctxt)
3278 /* opt_pass methods: */
3279 virtual unsigned int execute (function *)
3281 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3282 return 0;
3285 }; // class pass_jump2
3287 } // anon namespace
3289 rtl_opt_pass *
3290 make_pass_jump2 (gcc::context *ctxt)
3292 return new pass_jump2 (ctxt);