Update count_scale for AutoFDO to prevent over-scale.
[official-gcc.git] / gcc-4_8 / gcc / cfgcleanup.c
blob9952a8877ce76374097e6f9011716f28b68c0bf9
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2013 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 "tm.h"
36 #include "rtl.h"
37 #include "hard-reg-set.h"
38 #include "regs.h"
39 #include "insn-config.h"
40 #include "flags.h"
41 #include "recog.h"
42 #include "diagnostic-core.h"
43 #include "cselib.h"
44 #include "params.h"
45 #include "tm_p.h"
46 #include "target.h"
47 #include "function.h" /* For inline functions in emit-rtl.h they need crtl. */
48 #include "emit-rtl.h"
49 #include "tree-pass.h"
50 #include "cfgloop.h"
51 #include "expr.h"
52 #include "df.h"
53 #include "dce.h"
54 #include "dbgcnt.h"
56 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
58 /* Set to true when we are running first pass of try_optimize_cfg loop. */
59 static bool first_pass;
61 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
62 static bool crossjumps_occured;
64 /* Set to true if we couldn't run an optimization due to stale liveness
65 information; we should run df_analyze to enable more opportunities. */
66 static bool block_was_dirty;
68 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
69 static bool try_crossjump_bb (int, basic_block);
70 static bool outgoing_edges_match (int, basic_block, basic_block);
71 static enum replace_direction old_insns_match_p (int, rtx, rtx);
73 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
74 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
75 static bool try_optimize_cfg (int);
76 static bool try_simplify_condjump (basic_block);
77 static bool try_forward_edges (int, basic_block);
78 static edge thread_jump (edge, basic_block);
79 static bool mark_effect (rtx, bitmap);
80 static void notice_new_block (basic_block);
81 static void update_forwarder_flag (basic_block);
82 static int mentions_nonequal_regs (rtx *, void *);
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 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
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
160 || !can_fallthru (jump_block, cbranch_dest_block))
161 return false;
163 /* Invert the conditional branch. */
164 if (!invert_jump (cbranch_insn, block_label (jump_dest_block), 0))
165 return false;
167 if (dump_file)
168 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
169 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
171 /* Success. Update the CFG to match. Note that after this point
172 the edge variable names appear backwards; the redirection is done
173 this way to preserve edge profile data. */
174 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
175 cbranch_dest_block);
176 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
177 jump_dest_block);
178 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
179 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
180 update_br_prob_note (cbranch_block);
182 /* Delete the block with the unconditional jump, and clean up the mess. */
183 delete_basic_block (jump_block);
184 tidy_fallthru_edge (cbranch_jump_edge);
185 update_forwarder_flag (cbranch_block);
187 return true;
190 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
191 on register. Used by jump threading. */
193 static bool
194 mark_effect (rtx exp, regset nonequal)
196 int regno;
197 rtx dest;
198 switch (GET_CODE (exp))
200 /* In case we do clobber the register, mark it as equal, as we know the
201 value is dead so it don't have to match. */
202 case CLOBBER:
203 if (REG_P (XEXP (exp, 0)))
205 dest = XEXP (exp, 0);
206 regno = REGNO (dest);
207 if (HARD_REGISTER_NUM_P (regno))
208 bitmap_clear_range (nonequal, regno,
209 hard_regno_nregs[regno][GET_MODE (dest)]);
210 else
211 bitmap_clear_bit (nonequal, regno);
213 return false;
215 case SET:
216 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
217 return false;
218 dest = SET_DEST (exp);
219 if (dest == pc_rtx)
220 return false;
221 if (!REG_P (dest))
222 return true;
223 regno = REGNO (dest);
224 if (HARD_REGISTER_NUM_P (regno))
225 bitmap_set_range (nonequal, regno,
226 hard_regno_nregs[regno][GET_MODE (dest)]);
227 else
228 bitmap_set_bit (nonequal, regno);
229 return false;
231 default:
232 return false;
236 /* Return nonzero if X is a register set in regset DATA.
237 Called via for_each_rtx. */
238 static int
239 mentions_nonequal_regs (rtx *x, void *data)
241 regset nonequal = (regset) data;
242 if (REG_P (*x))
244 int regno;
246 regno = REGNO (*x);
247 if (REGNO_REG_SET_P (nonequal, regno))
248 return 1;
249 if (regno < FIRST_PSEUDO_REGISTER)
251 int n = hard_regno_nregs[regno][GET_MODE (*x)];
252 while (--n > 0)
253 if (REGNO_REG_SET_P (nonequal, regno + n))
254 return 1;
257 return 0;
259 /* Attempt to prove that the basic block B will have no side effects and
260 always continues in the same edge if reached via E. Return the edge
261 if exist, NULL otherwise. */
263 static edge
264 thread_jump (edge e, basic_block b)
266 rtx set1, set2, cond1, cond2, insn;
267 enum rtx_code code1, code2, reversed_code2;
268 bool reverse1 = false;
269 unsigned i;
270 regset nonequal;
271 bool failed = false;
272 reg_set_iterator rsi;
274 if (b->flags & BB_NONTHREADABLE_BLOCK)
275 return NULL;
277 /* At the moment, we do handle only conditional jumps, but later we may
278 want to extend this code to tablejumps and others. */
279 if (EDGE_COUNT (e->src->succs) != 2)
280 return NULL;
281 if (EDGE_COUNT (b->succs) != 2)
283 b->flags |= BB_NONTHREADABLE_BLOCK;
284 return NULL;
287 /* Second branch must end with onlyjump, as we will eliminate the jump. */
288 if (!any_condjump_p (BB_END (e->src)))
289 return NULL;
291 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
293 b->flags |= BB_NONTHREADABLE_BLOCK;
294 return NULL;
297 set1 = pc_set (BB_END (e->src));
298 set2 = pc_set (BB_END (b));
299 if (((e->flags & EDGE_FALLTHRU) != 0)
300 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
301 reverse1 = true;
303 cond1 = XEXP (SET_SRC (set1), 0);
304 cond2 = XEXP (SET_SRC (set2), 0);
305 if (reverse1)
306 code1 = reversed_comparison_code (cond1, BB_END (e->src));
307 else
308 code1 = GET_CODE (cond1);
310 code2 = GET_CODE (cond2);
311 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
313 if (!comparison_dominates_p (code1, code2)
314 && !comparison_dominates_p (code1, reversed_code2))
315 return NULL;
317 /* Ensure that the comparison operators are equivalent.
318 ??? This is far too pessimistic. We should allow swapped operands,
319 different CCmodes, or for example comparisons for interval, that
320 dominate even when operands are not equivalent. */
321 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
322 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
323 return NULL;
325 /* Short circuit cases where block B contains some side effects, as we can't
326 safely bypass it. */
327 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
328 insn = NEXT_INSN (insn))
329 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
331 b->flags |= BB_NONTHREADABLE_BLOCK;
332 return NULL;
335 cselib_init (0);
337 /* First process all values computed in the source basic block. */
338 for (insn = NEXT_INSN (BB_HEAD (e->src));
339 insn != NEXT_INSN (BB_END (e->src));
340 insn = NEXT_INSN (insn))
341 if (INSN_P (insn))
342 cselib_process_insn (insn);
344 nonequal = BITMAP_ALLOC (NULL);
345 CLEAR_REG_SET (nonequal);
347 /* Now assume that we've continued by the edge E to B and continue
348 processing as if it were same basic block.
349 Our goal is to prove that whole block is an NOOP. */
351 for (insn = NEXT_INSN (BB_HEAD (b));
352 insn != NEXT_INSN (BB_END (b)) && !failed;
353 insn = NEXT_INSN (insn))
355 if (INSN_P (insn))
357 rtx pat = PATTERN (insn);
359 if (GET_CODE (pat) == PARALLEL)
361 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
362 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
364 else
365 failed |= mark_effect (pat, nonequal);
368 cselib_process_insn (insn);
371 /* Later we should clear nonequal of dead registers. So far we don't
372 have life information in cfg_cleanup. */
373 if (failed)
375 b->flags |= BB_NONTHREADABLE_BLOCK;
376 goto failed_exit;
379 /* cond2 must not mention any register that is not equal to the
380 former block. */
381 if (for_each_rtx (&cond2, mentions_nonequal_regs, nonequal))
382 goto failed_exit;
384 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
385 goto failed_exit;
387 BITMAP_FREE (nonequal);
388 cselib_finish ();
389 if ((comparison_dominates_p (code1, code2) != 0)
390 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
391 return BRANCH_EDGE (b);
392 else
393 return FALLTHRU_EDGE (b);
395 failed_exit:
396 BITMAP_FREE (nonequal);
397 cselib_finish ();
398 return NULL;
401 /* Attempt to forward edges leaving basic block B.
402 Return true if successful. */
404 static bool
405 try_forward_edges (int mode, basic_block b)
407 bool changed = false;
408 edge_iterator ei;
409 edge e, *threaded_edges = NULL;
411 /* If we are partitioning hot/cold basic blocks, we don't want to
412 mess up unconditional or indirect jumps that cross between hot
413 and cold sections.
415 Basic block partitioning may result in some jumps that appear to
416 be optimizable (or blocks that appear to be mergeable), but which really
417 must be left untouched (they are required to make it safely across
418 partition boundaries). See the comments at the top of
419 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
421 if (find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX))
422 return false;
424 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
426 basic_block target, first;
427 int counter, goto_locus;
428 bool threaded = false;
429 int nthreaded_edges = 0;
430 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
432 /* Skip complex edges because we don't know how to update them.
434 Still handle fallthru edges, as we can succeed to forward fallthru
435 edge to the same place as the branch edge of conditional branch
436 and turn conditional branch to an unconditional branch. */
437 if (e->flags & EDGE_COMPLEX)
439 ei_next (&ei);
440 continue;
443 target = first = e->dest;
444 counter = NUM_FIXED_BLOCKS;
445 goto_locus = e->goto_locus;
447 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
448 up jumps that cross between hot/cold sections.
450 Basic block partitioning may result in some jumps that appear
451 to be optimizable (or blocks that appear to be mergeable), but which
452 really must be left untouched (they are required to make it safely
453 across partition boundaries). See the comments at the top of
454 bb-reorder.c:partition_hot_cold_basic_blocks for complete
455 details. */
457 if (first != EXIT_BLOCK_PTR
458 && find_reg_note (BB_END (first), REG_CROSSING_JUMP, NULL_RTX))
459 return changed;
461 while (counter < n_basic_blocks)
463 basic_block new_target = NULL;
464 bool new_target_threaded = false;
465 may_thread |= (target->flags & BB_MODIFIED) != 0;
467 if (FORWARDER_BLOCK_P (target)
468 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
469 && single_succ (target) != EXIT_BLOCK_PTR)
471 /* Bypass trivial infinite loops. */
472 new_target = single_succ (target);
473 if (target == new_target)
474 counter = n_basic_blocks;
475 else if (!optimize)
477 /* When not optimizing, ensure that edges or forwarder
478 blocks with different locus are not optimized out. */
479 int new_locus = single_succ_edge (target)->goto_locus;
480 int locus = goto_locus;
482 if (new_locus != UNKNOWN_LOCATION
483 && locus != UNKNOWN_LOCATION
484 && new_locus != locus)
485 new_target = NULL;
486 else
488 rtx last;
490 if (new_locus != UNKNOWN_LOCATION)
491 locus = new_locus;
493 last = BB_END (target);
494 if (DEBUG_INSN_P (last))
495 last = prev_nondebug_insn (last);
497 new_locus = last && INSN_P (last)
498 ? INSN_LOCATION (last) : 0;
500 if (new_locus != UNKNOWN_LOCATION
501 && locus != UNKNOWN_LOCATION
502 && new_locus != locus)
503 new_target = NULL;
504 else
506 if (new_locus != UNKNOWN_LOCATION)
507 locus = new_locus;
509 goto_locus = locus;
515 /* Allow to thread only over one edge at time to simplify updating
516 of probabilities. */
517 else if ((mode & CLEANUP_THREADING) && may_thread)
519 edge t = thread_jump (e, target);
520 if (t)
522 if (!threaded_edges)
523 threaded_edges = XNEWVEC (edge, n_basic_blocks);
524 else
526 int i;
528 /* Detect an infinite loop across blocks not
529 including the start block. */
530 for (i = 0; i < nthreaded_edges; ++i)
531 if (threaded_edges[i] == t)
532 break;
533 if (i < nthreaded_edges)
535 counter = n_basic_blocks;
536 break;
540 /* Detect an infinite loop across the start block. */
541 if (t->dest == b)
542 break;
544 gcc_assert (nthreaded_edges < n_basic_blocks - NUM_FIXED_BLOCKS);
545 threaded_edges[nthreaded_edges++] = t;
547 new_target = t->dest;
548 new_target_threaded = true;
552 if (!new_target)
553 break;
555 counter++;
556 target = new_target;
557 threaded |= new_target_threaded;
560 if (counter >= n_basic_blocks)
562 if (dump_file)
563 fprintf (dump_file, "Infinite loop in BB %i.\n",
564 target->index);
566 else if (target == first)
567 ; /* We didn't do anything. */
568 else
570 /* Save the values now, as the edge may get removed. */
571 gcov_type edge_count = e->count;
572 int edge_probability = e->probability;
573 int edge_frequency;
574 int n = 0;
576 e->goto_locus = goto_locus;
578 /* Don't force if target is exit block. */
579 if (threaded && target != EXIT_BLOCK_PTR)
581 notice_new_block (redirect_edge_and_branch_force (e, target));
582 if (dump_file)
583 fprintf (dump_file, "Conditionals threaded.\n");
585 else if (!redirect_edge_and_branch (e, target))
587 if (dump_file)
588 fprintf (dump_file,
589 "Forwarding edge %i->%i to %i failed.\n",
590 b->index, e->dest->index, target->index);
591 ei_next (&ei);
592 continue;
595 /* We successfully forwarded the edge. Now update profile
596 data: for each edge we traversed in the chain, remove
597 the original edge's execution count. */
598 edge_frequency = ((edge_probability * b->frequency
599 + REG_BR_PROB_BASE / 2)
600 / REG_BR_PROB_BASE);
604 edge t;
606 if (!single_succ_p (first))
608 gcc_assert (n < nthreaded_edges);
609 t = threaded_edges [n++];
610 gcc_assert (t->src == first);
611 update_bb_profile_for_threading (first, edge_frequency,
612 edge_count, t);
613 update_br_prob_note (first);
615 else
617 first->count -= edge_count;
618 if (first->count < 0)
619 first->count = 0;
620 first->frequency -= edge_frequency;
621 if (first->frequency < 0)
622 first->frequency = 0;
623 /* It is possible that as the result of
624 threading we've removed edge as it is
625 threaded to the fallthru edge. Avoid
626 getting out of sync. */
627 if (n < nthreaded_edges
628 && first == threaded_edges [n]->src)
629 n++;
630 t = single_succ_edge (first);
633 t->count -= edge_count;
634 if (t->count < 0)
635 t->count = 0;
636 first = t->dest;
638 while (first != target);
640 changed = true;
641 continue;
643 ei_next (&ei);
646 free (threaded_edges);
647 return changed;
651 /* Blocks A and B are to be merged into a single block. A has no incoming
652 fallthru edge, so it can be moved before B without adding or modifying
653 any jumps (aside from the jump from A to B). */
655 static void
656 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
658 rtx barrier;
660 /* If we are partitioning hot/cold basic blocks, we don't want to
661 mess up unconditional or indirect jumps that cross between hot
662 and cold sections.
664 Basic block partitioning may result in some jumps that appear to
665 be optimizable (or blocks that appear to be mergeable), but which really
666 must be left untouched (they are required to make it safely across
667 partition boundaries). See the comments at the top of
668 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
670 if (BB_PARTITION (a) != BB_PARTITION (b))
671 return;
673 barrier = next_nonnote_insn (BB_END (a));
674 gcc_assert (BARRIER_P (barrier));
675 delete_insn (barrier);
677 /* Scramble the insn chain. */
678 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
679 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
680 df_set_bb_dirty (a);
682 if (dump_file)
683 fprintf (dump_file, "Moved block %d before %d and merged.\n",
684 a->index, b->index);
686 /* Swap the records for the two blocks around. */
688 unlink_block (a);
689 link_block (a, b->prev_bb);
691 /* Now blocks A and B are contiguous. Merge them. */
692 merge_blocks (a, b);
695 /* Blocks A and B are to be merged into a single block. B has no outgoing
696 fallthru edge, so it can be moved after A without adding or modifying
697 any jumps (aside from the jump from A to B). */
699 static void
700 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
702 rtx barrier, real_b_end;
703 rtx label, table;
705 /* If we are partitioning hot/cold basic blocks, we don't want to
706 mess up unconditional or indirect jumps that cross between hot
707 and cold sections.
709 Basic block partitioning may result in some jumps that appear to
710 be optimizable (or blocks that appear to be mergeable), but which really
711 must be left untouched (they are required to make it safely across
712 partition boundaries). See the comments at the top of
713 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
715 if (BB_PARTITION (a) != BB_PARTITION (b))
716 return;
718 real_b_end = BB_END (b);
720 /* If there is a jump table following block B temporarily add the jump table
721 to block B so that it will also be moved to the correct location. */
722 if (tablejump_p (BB_END (b), &label, &table)
723 && prev_active_insn (label) == BB_END (b))
725 BB_END (b) = table;
728 /* There had better have been a barrier there. Delete it. */
729 barrier = NEXT_INSN (BB_END (b));
730 if (barrier && BARRIER_P (barrier))
731 delete_insn (barrier);
734 /* Scramble the insn chain. */
735 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
737 /* Restore the real end of b. */
738 BB_END (b) = real_b_end;
740 if (dump_file)
741 fprintf (dump_file, "Moved block %d after %d and merged.\n",
742 b->index, a->index);
744 /* Now blocks A and B are contiguous. Merge them. */
745 merge_blocks (a, b);
748 /* Attempt to merge basic blocks that are potentially non-adjacent.
749 Return NULL iff the attempt failed, otherwise return basic block
750 where cleanup_cfg should continue. Because the merging commonly
751 moves basic block away or introduces another optimization
752 possibility, return basic block just before B so cleanup_cfg don't
753 need to iterate.
755 It may be good idea to return basic block before C in the case
756 C has been moved after B and originally appeared earlier in the
757 insn sequence, but we have no information available about the
758 relative ordering of these two. Hopefully it is not too common. */
760 static basic_block
761 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
763 basic_block next;
765 /* If we are partitioning hot/cold basic blocks, we don't want to
766 mess up unconditional or indirect jumps that cross between hot
767 and cold sections.
769 Basic block partitioning may result in some jumps that appear to
770 be optimizable (or blocks that appear to be mergeable), but which really
771 must be left untouched (they are required to make it safely across
772 partition boundaries). See the comments at the top of
773 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
775 if (BB_PARTITION (b) != BB_PARTITION (c))
776 return NULL;
778 /* If B has a fallthru edge to C, no need to move anything. */
779 if (e->flags & EDGE_FALLTHRU)
781 int b_index = b->index, c_index = c->index;
783 /* Protect the loop latches. */
784 if (current_loops && c->loop_father->latch == c)
785 return NULL;
787 merge_blocks (b, c);
788 update_forwarder_flag (b);
790 if (dump_file)
791 fprintf (dump_file, "Merged %d and %d without moving.\n",
792 b_index, c_index);
794 return b->prev_bb == ENTRY_BLOCK_PTR ? b : b->prev_bb;
797 /* Otherwise we will need to move code around. Do that only if expensive
798 transformations are allowed. */
799 else if (mode & CLEANUP_EXPENSIVE)
801 edge tmp_edge, b_fallthru_edge;
802 bool c_has_outgoing_fallthru;
803 bool b_has_incoming_fallthru;
805 /* Avoid overactive code motion, as the forwarder blocks should be
806 eliminated by edge redirection instead. One exception might have
807 been if B is a forwarder block and C has no fallthru edge, but
808 that should be cleaned up by bb-reorder instead. */
809 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
810 return NULL;
812 /* We must make sure to not munge nesting of lexical blocks,
813 and loop notes. This is done by squeezing out all the notes
814 and leaving them there to lie. Not ideal, but functional. */
816 tmp_edge = find_fallthru_edge (c->succs);
817 c_has_outgoing_fallthru = (tmp_edge != NULL);
819 tmp_edge = find_fallthru_edge (b->preds);
820 b_has_incoming_fallthru = (tmp_edge != NULL);
821 b_fallthru_edge = tmp_edge;
822 next = b->prev_bb;
823 if (next == c)
824 next = next->prev_bb;
826 /* Otherwise, we're going to try to move C after B. If C does
827 not have an outgoing fallthru, then it can be moved
828 immediately after B without introducing or modifying jumps. */
829 if (! c_has_outgoing_fallthru)
831 merge_blocks_move_successor_nojumps (b, c);
832 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
835 /* If B does not have an incoming fallthru, then it can be moved
836 immediately before C without introducing or modifying jumps.
837 C cannot be the first block, so we do not have to worry about
838 accessing a non-existent block. */
840 if (b_has_incoming_fallthru)
842 basic_block bb;
844 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR)
845 return NULL;
846 bb = force_nonfallthru (b_fallthru_edge);
847 if (bb)
848 notice_new_block (bb);
851 merge_blocks_move_predecessor_nojumps (b, c);
852 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
855 return NULL;
859 /* Removes the memory attributes of MEM expression
860 if they are not equal. */
862 void
863 merge_memattrs (rtx x, rtx y)
865 int i;
866 int j;
867 enum rtx_code code;
868 const char *fmt;
870 if (x == y)
871 return;
872 if (x == 0 || y == 0)
873 return;
875 code = GET_CODE (x);
877 if (code != GET_CODE (y))
878 return;
880 if (GET_MODE (x) != GET_MODE (y))
881 return;
883 if (code == MEM && MEM_ATTRS (x) != MEM_ATTRS (y))
885 if (! MEM_ATTRS (x))
886 MEM_ATTRS (y) = 0;
887 else if (! MEM_ATTRS (y))
888 MEM_ATTRS (x) = 0;
889 else
891 HOST_WIDE_INT mem_size;
893 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
895 set_mem_alias_set (x, 0);
896 set_mem_alias_set (y, 0);
899 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
901 set_mem_expr (x, 0);
902 set_mem_expr (y, 0);
903 clear_mem_offset (x);
904 clear_mem_offset (y);
906 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
907 || (MEM_OFFSET_KNOWN_P (x)
908 && MEM_OFFSET (x) != MEM_OFFSET (y)))
910 clear_mem_offset (x);
911 clear_mem_offset (y);
914 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
916 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
917 set_mem_size (x, mem_size);
918 set_mem_size (y, mem_size);
920 else
922 clear_mem_size (x);
923 clear_mem_size (y);
926 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
927 set_mem_align (y, MEM_ALIGN (x));
930 if (code == MEM)
932 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
934 MEM_READONLY_P (x) = 0;
935 MEM_READONLY_P (y) = 0;
937 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
939 MEM_NOTRAP_P (x) = 0;
940 MEM_NOTRAP_P (y) = 0;
942 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
944 MEM_VOLATILE_P (x) = 1;
945 MEM_VOLATILE_P (y) = 1;
949 fmt = GET_RTX_FORMAT (code);
950 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
952 switch (fmt[i])
954 case 'E':
955 /* Two vectors must have the same length. */
956 if (XVECLEN (x, i) != XVECLEN (y, i))
957 return;
959 for (j = 0; j < XVECLEN (x, i); j++)
960 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
962 break;
964 case 'e':
965 merge_memattrs (XEXP (x, i), XEXP (y, i));
968 return;
972 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
973 different single sets S1 and S2. */
975 static bool
976 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
978 int i;
979 rtx e1, e2;
981 if (p1 == s1 && p2 == s2)
982 return true;
984 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
985 return false;
987 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
988 return false;
990 for (i = 0; i < XVECLEN (p1, 0); i++)
992 e1 = XVECEXP (p1, 0, i);
993 e2 = XVECEXP (p2, 0, i);
994 if (e1 == s1 && e2 == s2)
995 continue;
996 if (reload_completed
997 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
998 continue;
1000 return false;
1003 return true;
1006 /* Examine register notes on I1 and I2 and return:
1007 - dir_forward if I1 can be replaced by I2, or
1008 - dir_backward if I2 can be replaced by I1, or
1009 - dir_both if both are the case. */
1011 static enum replace_direction
1012 can_replace_by (rtx i1, rtx i2)
1014 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1015 bool c1, c2;
1017 /* Check for 2 sets. */
1018 s1 = single_set (i1);
1019 s2 = single_set (i2);
1020 if (s1 == NULL_RTX || s2 == NULL_RTX)
1021 return dir_none;
1023 /* Check that the 2 sets set the same dest. */
1024 d1 = SET_DEST (s1);
1025 d2 = SET_DEST (s2);
1026 if (!(reload_completed
1027 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1028 return dir_none;
1030 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1031 set dest to the same value. */
1032 note1 = find_reg_equal_equiv_note (i1);
1033 note2 = find_reg_equal_equiv_note (i2);
1034 if (!note1 || !note2 || !rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0))
1035 || !CONST_INT_P (XEXP (note1, 0)))
1036 return dir_none;
1038 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1039 return dir_none;
1041 /* Although the 2 sets set dest to the same value, we cannot replace
1042 (set (dest) (const_int))
1044 (set (dest) (reg))
1045 because we don't know if the reg is live and has the same value at the
1046 location of replacement. */
1047 src1 = SET_SRC (s1);
1048 src2 = SET_SRC (s2);
1049 c1 = CONST_INT_P (src1);
1050 c2 = CONST_INT_P (src2);
1051 if (c1 && c2)
1052 return dir_both;
1053 else if (c2)
1054 return dir_forward;
1055 else if (c1)
1056 return dir_backward;
1058 return dir_none;
1061 /* Merges directions A and B. */
1063 static enum replace_direction
1064 merge_dir (enum replace_direction a, enum replace_direction b)
1066 /* Implements the following table:
1067 |bo fw bw no
1068 ---+-----------
1069 bo |bo fw bw no
1070 fw |-- fw no no
1071 bw |-- -- bw no
1072 no |-- -- -- no. */
1074 if (a == b)
1075 return a;
1077 if (a == dir_both)
1078 return b;
1079 if (b == dir_both)
1080 return a;
1082 return dir_none;
1085 /* Examine I1 and I2 and return:
1086 - dir_forward if I1 can be replaced by I2, or
1087 - dir_backward if I2 can be replaced by I1, or
1088 - dir_both if both are the case. */
1090 static enum replace_direction
1091 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx i1, rtx i2)
1093 rtx p1, p2;
1095 /* Verify that I1 and I2 are equivalent. */
1096 if (GET_CODE (i1) != GET_CODE (i2))
1097 return dir_none;
1099 /* __builtin_unreachable() may lead to empty blocks (ending with
1100 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1101 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1102 return dir_both;
1104 /* ??? Do not allow cross-jumping between different stack levels. */
1105 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1106 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1107 if (p1 && p2)
1109 p1 = XEXP (p1, 0);
1110 p2 = XEXP (p2, 0);
1111 if (!rtx_equal_p (p1, p2))
1112 return dir_none;
1114 /* ??? Worse, this adjustment had better be constant lest we
1115 have differing incoming stack levels. */
1116 if (!frame_pointer_needed
1117 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1118 return dir_none;
1120 else if (p1 || p2)
1121 return dir_none;
1123 p1 = PATTERN (i1);
1124 p2 = PATTERN (i2);
1126 if (GET_CODE (p1) != GET_CODE (p2))
1127 return dir_none;
1129 /* If this is a CALL_INSN, compare register usage information.
1130 If we don't check this on stack register machines, the two
1131 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1132 numbers of stack registers in the same basic block.
1133 If we don't check this on machines with delay slots, a delay slot may
1134 be filled that clobbers a parameter expected by the subroutine.
1136 ??? We take the simple route for now and assume that if they're
1137 equal, they were constructed identically.
1139 Also check for identical exception regions. */
1141 if (CALL_P (i1))
1143 /* Ensure the same EH region. */
1144 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1145 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1147 if (!n1 && n2)
1148 return dir_none;
1150 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1151 return dir_none;
1153 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1154 CALL_INSN_FUNCTION_USAGE (i2))
1155 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1156 return dir_none;
1158 /* For address sanitizer, never crossjump __asan_report_* builtins,
1159 otherwise errors might be reported on incorrect lines. */
1160 if (flag_asan)
1162 rtx call = get_call_rtx_from (i1);
1163 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1165 rtx symbol = XEXP (XEXP (call, 0), 0);
1166 if (SYMBOL_REF_DECL (symbol)
1167 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1169 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1170 == BUILT_IN_NORMAL)
1171 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1172 >= BUILT_IN_ASAN_REPORT_LOAD1
1173 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1174 <= BUILT_IN_ASAN_REPORT_STORE16)
1175 return dir_none;
1181 #ifdef STACK_REGS
1182 /* If cross_jump_death_matters is not 0, the insn's mode
1183 indicates whether or not the insn contains any stack-like
1184 regs. */
1186 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1188 /* If register stack conversion has already been done, then
1189 death notes must also be compared before it is certain that
1190 the two instruction streams match. */
1192 rtx note;
1193 HARD_REG_SET i1_regset, i2_regset;
1195 CLEAR_HARD_REG_SET (i1_regset);
1196 CLEAR_HARD_REG_SET (i2_regset);
1198 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1199 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1200 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1202 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1203 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1204 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1206 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1207 return dir_none;
1209 #endif
1211 if (reload_completed
1212 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1213 return dir_both;
1215 return can_replace_by (i1, i2);
1218 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1219 flow_find_head_matching_sequence, ensure the notes match. */
1221 static void
1222 merge_notes (rtx i1, rtx i2)
1224 /* If the merged insns have different REG_EQUAL notes, then
1225 remove them. */
1226 rtx equiv1 = find_reg_equal_equiv_note (i1);
1227 rtx equiv2 = find_reg_equal_equiv_note (i2);
1229 if (equiv1 && !equiv2)
1230 remove_note (i1, equiv1);
1231 else if (!equiv1 && equiv2)
1232 remove_note (i2, equiv2);
1233 else if (equiv1 && equiv2
1234 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1236 remove_note (i1, equiv1);
1237 remove_note (i2, equiv2);
1241 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1242 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1243 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1244 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1245 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1247 static void
1248 walk_to_nondebug_insn (rtx *i1, basic_block *bb1, bool follow_fallthru,
1249 bool *did_fallthru)
1251 edge fallthru;
1253 *did_fallthru = false;
1255 /* Ignore notes. */
1256 while (!NONDEBUG_INSN_P (*i1))
1258 if (*i1 != BB_HEAD (*bb1))
1260 *i1 = PREV_INSN (*i1);
1261 continue;
1264 if (!follow_fallthru)
1265 return;
1267 fallthru = find_fallthru_edge ((*bb1)->preds);
1268 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun)
1269 || !single_succ_p (fallthru->src))
1270 return;
1272 *bb1 = fallthru->src;
1273 *i1 = BB_END (*bb1);
1274 *did_fallthru = true;
1278 /* Look through the insns at the end of BB1 and BB2 and find the longest
1279 sequence that are either equivalent, or allow forward or backward
1280 replacement. Store the first insns for that sequence in *F1 and *F2 and
1281 return the sequence length.
1283 DIR_P indicates the allowed replacement direction on function entry, and
1284 the actual replacement direction on function exit. If NULL, only equivalent
1285 sequences are allowed.
1287 To simplify callers of this function, if the blocks match exactly,
1288 store the head of the blocks in *F1 and *F2. */
1291 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx *f1, rtx *f2,
1292 enum replace_direction *dir_p)
1294 rtx i1, i2, last1, last2, afterlast1, afterlast2;
1295 int ninsns = 0;
1296 rtx p1;
1297 enum replace_direction dir, last_dir, afterlast_dir;
1298 bool follow_fallthru, did_fallthru;
1300 if (dir_p)
1301 dir = *dir_p;
1302 else
1303 dir = dir_both;
1304 afterlast_dir = dir;
1305 last_dir = afterlast_dir;
1307 /* Skip simple jumps at the end of the blocks. Complex jumps still
1308 need to be compared for equivalence, which we'll do below. */
1310 i1 = BB_END (bb1);
1311 last1 = afterlast1 = last2 = afterlast2 = NULL_RTX;
1312 if (onlyjump_p (i1)
1313 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1315 last1 = i1;
1316 i1 = PREV_INSN (i1);
1319 i2 = BB_END (bb2);
1320 if (onlyjump_p (i2)
1321 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1323 last2 = i2;
1324 /* Count everything except for unconditional jump as insn. */
1325 if (!simplejump_p (i2) && !returnjump_p (i2) && last1)
1326 ninsns++;
1327 i2 = PREV_INSN (i2);
1330 while (true)
1332 /* In the following example, we can replace all jumps to C by jumps to A.
1334 This removes 4 duplicate insns.
1335 [bb A] insn1 [bb C] insn1
1336 insn2 insn2
1337 [bb B] insn3 insn3
1338 insn4 insn4
1339 jump_insn jump_insn
1341 We could also replace all jumps to A by jumps to C, but that leaves B
1342 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1343 step, all jumps to B would be replaced with jumps to the middle of C,
1344 achieving the same result with more effort.
1345 So we allow only the first possibility, which means that we don't allow
1346 fallthru in the block that's being replaced. */
1348 follow_fallthru = dir_p && dir != dir_forward;
1349 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1350 if (did_fallthru)
1351 dir = dir_backward;
1353 follow_fallthru = dir_p && dir != dir_backward;
1354 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1355 if (did_fallthru)
1356 dir = dir_forward;
1358 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1359 break;
1361 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1362 if (dir == dir_none || (!dir_p && dir != dir_both))
1363 break;
1365 merge_memattrs (i1, i2);
1367 /* Don't begin a cross-jump with a NOTE insn. */
1368 if (INSN_P (i1))
1370 merge_notes (i1, i2);
1372 afterlast1 = last1, afterlast2 = last2;
1373 last1 = i1, last2 = i2;
1374 afterlast_dir = last_dir;
1375 last_dir = dir;
1376 p1 = PATTERN (i1);
1377 if (!(GET_CODE (p1) == USE || GET_CODE (p1) == CLOBBER))
1378 ninsns++;
1381 i1 = PREV_INSN (i1);
1382 i2 = PREV_INSN (i2);
1385 #ifdef HAVE_cc0
1386 /* Don't allow the insn after a compare to be shared by
1387 cross-jumping unless the compare is also shared. */
1388 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
1389 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1390 #endif
1392 /* Include preceding notes and labels in the cross-jump. One,
1393 this may bring us to the head of the blocks as requested above.
1394 Two, it keeps line number notes as matched as may be. */
1395 if (ninsns)
1397 bb1 = BLOCK_FOR_INSN (last1);
1398 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1399 last1 = PREV_INSN (last1);
1401 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1402 last1 = PREV_INSN (last1);
1404 bb2 = BLOCK_FOR_INSN (last2);
1405 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1406 last2 = PREV_INSN (last2);
1408 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1409 last2 = PREV_INSN (last2);
1411 *f1 = last1;
1412 *f2 = last2;
1415 if (dir_p)
1416 *dir_p = last_dir;
1417 return ninsns;
1420 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1421 the head of the two blocks. Do not include jumps at the end.
1422 If STOP_AFTER is nonzero, stop after finding that many matching
1423 instructions. */
1426 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx *f1,
1427 rtx *f2, int stop_after)
1429 rtx i1, i2, last1, last2, beforelast1, beforelast2;
1430 int ninsns = 0;
1431 edge e;
1432 edge_iterator ei;
1433 int nehedges1 = 0, nehedges2 = 0;
1435 FOR_EACH_EDGE (e, ei, bb1->succs)
1436 if (e->flags & EDGE_EH)
1437 nehedges1++;
1438 FOR_EACH_EDGE (e, ei, bb2->succs)
1439 if (e->flags & EDGE_EH)
1440 nehedges2++;
1442 i1 = BB_HEAD (bb1);
1443 i2 = BB_HEAD (bb2);
1444 last1 = beforelast1 = last2 = beforelast2 = NULL_RTX;
1446 while (true)
1448 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1449 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1451 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1452 break;
1453 i1 = NEXT_INSN (i1);
1456 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1458 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1459 break;
1460 i2 = NEXT_INSN (i2);
1463 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1464 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1465 break;
1467 if (NOTE_P (i1) || NOTE_P (i2)
1468 || JUMP_P (i1) || JUMP_P (i2))
1469 break;
1471 /* A sanity check to make sure we're not merging insns with different
1472 effects on EH. If only one of them ends a basic block, it shouldn't
1473 have an EH edge; if both end a basic block, there should be the same
1474 number of EH edges. */
1475 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1476 && nehedges1 > 0)
1477 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1478 && nehedges2 > 0)
1479 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1480 && nehedges1 != nehedges2))
1481 break;
1483 if (old_insns_match_p (0, i1, i2) != dir_both)
1484 break;
1486 merge_memattrs (i1, i2);
1488 /* Don't begin a cross-jump with a NOTE insn. */
1489 if (INSN_P (i1))
1491 merge_notes (i1, i2);
1493 beforelast1 = last1, beforelast2 = last2;
1494 last1 = i1, last2 = i2;
1495 ninsns++;
1498 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1499 || (stop_after > 0 && ninsns == stop_after))
1500 break;
1502 i1 = NEXT_INSN (i1);
1503 i2 = NEXT_INSN (i2);
1506 #ifdef HAVE_cc0
1507 /* Don't allow a compare to be shared by cross-jumping unless the insn
1508 after the compare is also shared. */
1509 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && sets_cc0_p (last1))
1510 last1 = beforelast1, last2 = beforelast2, ninsns--;
1511 #endif
1513 if (ninsns)
1515 *f1 = last1;
1516 *f2 = last2;
1519 return ninsns;
1522 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1523 the branch instruction. This means that if we commonize the control
1524 flow before end of the basic block, the semantic remains unchanged.
1526 We may assume that there exists one edge with a common destination. */
1528 static bool
1529 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1531 int nehedges1 = 0, nehedges2 = 0;
1532 edge fallthru1 = 0, fallthru2 = 0;
1533 edge e1, e2;
1534 edge_iterator ei;
1536 /* If we performed shrink-wrapping, edges to the EXIT_BLOCK_PTR can
1537 only be distinguished for JUMP_INSNs. The two paths may differ in
1538 whether they went through the prologue. Sibcalls are fine, we know
1539 that we either didn't need or inserted an epilogue before them. */
1540 if (crtl->shrink_wrapped
1541 && single_succ_p (bb1) && single_succ (bb1) == EXIT_BLOCK_PTR
1542 && !JUMP_P (BB_END (bb1))
1543 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1544 return false;
1546 /* If BB1 has only one successor, we may be looking at either an
1547 unconditional jump, or a fake edge to exit. */
1548 if (single_succ_p (bb1)
1549 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1550 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1551 return (single_succ_p (bb2)
1552 && (single_succ_edge (bb2)->flags
1553 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1554 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1556 /* Match conditional jumps - this may get tricky when fallthru and branch
1557 edges are crossed. */
1558 if (EDGE_COUNT (bb1->succs) == 2
1559 && any_condjump_p (BB_END (bb1))
1560 && onlyjump_p (BB_END (bb1)))
1562 edge b1, f1, b2, f2;
1563 bool reverse, match;
1564 rtx set1, set2, cond1, cond2;
1565 enum rtx_code code1, code2;
1567 if (EDGE_COUNT (bb2->succs) != 2
1568 || !any_condjump_p (BB_END (bb2))
1569 || !onlyjump_p (BB_END (bb2)))
1570 return false;
1572 b1 = BRANCH_EDGE (bb1);
1573 b2 = BRANCH_EDGE (bb2);
1574 f1 = FALLTHRU_EDGE (bb1);
1575 f2 = FALLTHRU_EDGE (bb2);
1577 /* Get around possible forwarders on fallthru edges. Other cases
1578 should be optimized out already. */
1579 if (FORWARDER_BLOCK_P (f1->dest))
1580 f1 = single_succ_edge (f1->dest);
1582 if (FORWARDER_BLOCK_P (f2->dest))
1583 f2 = single_succ_edge (f2->dest);
1585 /* To simplify use of this function, return false if there are
1586 unneeded forwarder blocks. These will get eliminated later
1587 during cleanup_cfg. */
1588 if (FORWARDER_BLOCK_P (f1->dest)
1589 || FORWARDER_BLOCK_P (f2->dest)
1590 || FORWARDER_BLOCK_P (b1->dest)
1591 || FORWARDER_BLOCK_P (b2->dest))
1592 return false;
1594 if (f1->dest == f2->dest && b1->dest == b2->dest)
1595 reverse = false;
1596 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1597 reverse = true;
1598 else
1599 return false;
1601 set1 = pc_set (BB_END (bb1));
1602 set2 = pc_set (BB_END (bb2));
1603 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1604 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1605 reverse = !reverse;
1607 cond1 = XEXP (SET_SRC (set1), 0);
1608 cond2 = XEXP (SET_SRC (set2), 0);
1609 code1 = GET_CODE (cond1);
1610 if (reverse)
1611 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1612 else
1613 code2 = GET_CODE (cond2);
1615 if (code2 == UNKNOWN)
1616 return false;
1618 /* Verify codes and operands match. */
1619 match = ((code1 == code2
1620 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1621 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1622 || (code1 == swap_condition (code2)
1623 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1624 XEXP (cond2, 0))
1625 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1626 XEXP (cond2, 1))));
1628 /* If we return true, we will join the blocks. Which means that
1629 we will only have one branch prediction bit to work with. Thus
1630 we require the existing branches to have probabilities that are
1631 roughly similar. */
1632 if (match
1633 && optimize_bb_for_speed_p (bb1)
1634 && optimize_bb_for_speed_p (bb2))
1636 int prob2;
1638 if (b1->dest == b2->dest)
1639 prob2 = b2->probability;
1640 else
1641 /* Do not use f2 probability as f2 may be forwarded. */
1642 prob2 = REG_BR_PROB_BASE - b2->probability;
1644 /* Fail if the difference in probabilities is greater than 50%.
1645 This rules out two well-predicted branches with opposite
1646 outcomes. */
1647 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1649 if (dump_file)
1650 fprintf (dump_file,
1651 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1652 bb1->index, bb2->index, b1->probability, prob2);
1654 return false;
1658 if (dump_file && match)
1659 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1660 bb1->index, bb2->index);
1662 return match;
1665 /* Generic case - we are seeing a computed jump, table jump or trapping
1666 instruction. */
1668 /* Check whether there are tablejumps in the end of BB1 and BB2.
1669 Return true if they are identical. */
1671 rtx label1, label2;
1672 rtx table1, table2;
1674 if (tablejump_p (BB_END (bb1), &label1, &table1)
1675 && tablejump_p (BB_END (bb2), &label2, &table2)
1676 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1678 /* The labels should never be the same rtx. If they really are same
1679 the jump tables are same too. So disable crossjumping of blocks BB1
1680 and BB2 because when deleting the common insns in the end of BB1
1681 by delete_basic_block () the jump table would be deleted too. */
1682 /* If LABEL2 is referenced in BB1->END do not do anything
1683 because we would loose information when replacing
1684 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1685 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1687 /* Set IDENTICAL to true when the tables are identical. */
1688 bool identical = false;
1689 rtx p1, p2;
1691 p1 = PATTERN (table1);
1692 p2 = PATTERN (table2);
1693 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1695 identical = true;
1697 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1698 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1699 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1700 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1702 int i;
1704 identical = true;
1705 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1706 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1707 identical = false;
1710 if (identical)
1712 replace_label_data rr;
1713 bool match;
1715 /* Temporarily replace references to LABEL1 with LABEL2
1716 in BB1->END so that we could compare the instructions. */
1717 rr.r1 = label1;
1718 rr.r2 = label2;
1719 rr.update_label_nuses = false;
1720 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1722 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1723 == dir_both);
1724 if (dump_file && match)
1725 fprintf (dump_file,
1726 "Tablejumps in bb %i and %i match.\n",
1727 bb1->index, bb2->index);
1729 /* Set the original label in BB1->END because when deleting
1730 a block whose end is a tablejump, the tablejump referenced
1731 from the instruction is deleted too. */
1732 rr.r1 = label2;
1733 rr.r2 = label1;
1734 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1736 return match;
1739 return false;
1743 /* Find the last non-debug non-note instruction in each bb, except
1744 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1745 handles that case specially. old_insns_match_p does not handle
1746 other types of instruction notes. */
1747 rtx last1 = BB_END (bb1);
1748 rtx last2 = BB_END (bb2);
1749 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1750 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1751 last1 = PREV_INSN (last1);
1752 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1753 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1754 last2 = PREV_INSN (last2);
1755 gcc_assert (last1 && last2);
1757 /* First ensure that the instructions match. There may be many outgoing
1758 edges so this test is generally cheaper. */
1759 if (old_insns_match_p (mode, last1, last2) != dir_both)
1760 return false;
1762 /* Search the outgoing edges, ensure that the counts do match, find possible
1763 fallthru and exception handling edges since these needs more
1764 validation. */
1765 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1766 return false;
1768 bool nonfakeedges = false;
1769 FOR_EACH_EDGE (e1, ei, bb1->succs)
1771 e2 = EDGE_SUCC (bb2, ei.index);
1773 if ((e1->flags & EDGE_FAKE) == 0)
1774 nonfakeedges = true;
1776 if (e1->flags & EDGE_EH)
1777 nehedges1++;
1779 if (e2->flags & EDGE_EH)
1780 nehedges2++;
1782 if (e1->flags & EDGE_FALLTHRU)
1783 fallthru1 = e1;
1784 if (e2->flags & EDGE_FALLTHRU)
1785 fallthru2 = e2;
1788 /* If number of edges of various types does not match, fail. */
1789 if (nehedges1 != nehedges2
1790 || (fallthru1 != 0) != (fallthru2 != 0))
1791 return false;
1793 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1794 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1795 attempt to optimize, as the two basic blocks might have different
1796 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1797 traps there should be REG_ARG_SIZE notes, they could be missing
1798 for __builtin_unreachable () uses though. */
1799 if (!nonfakeedges
1800 && !ACCUMULATE_OUTGOING_ARGS
1801 && (!INSN_P (last1)
1802 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1803 return false;
1805 /* fallthru edges must be forwarded to the same destination. */
1806 if (fallthru1)
1808 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1809 ? single_succ (fallthru1->dest): fallthru1->dest);
1810 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1811 ? single_succ (fallthru2->dest): fallthru2->dest);
1813 if (d1 != d2)
1814 return false;
1817 /* Ensure the same EH region. */
1819 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1820 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1822 if (!n1 && n2)
1823 return false;
1825 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1826 return false;
1829 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1830 version of sequence abstraction. */
1831 FOR_EACH_EDGE (e1, ei, bb2->succs)
1833 edge e2;
1834 edge_iterator ei;
1835 basic_block d1 = e1->dest;
1837 if (FORWARDER_BLOCK_P (d1))
1838 d1 = EDGE_SUCC (d1, 0)->dest;
1840 FOR_EACH_EDGE (e2, ei, bb1->succs)
1842 basic_block d2 = e2->dest;
1843 if (FORWARDER_BLOCK_P (d2))
1844 d2 = EDGE_SUCC (d2, 0)->dest;
1845 if (d1 == d2)
1846 break;
1849 if (!e2)
1850 return false;
1853 return true;
1856 /* Returns true if BB basic block has a preserve label. */
1858 static bool
1859 block_has_preserve_label (basic_block bb)
1861 return (bb
1862 && block_label (bb)
1863 && LABEL_PRESERVE_P (block_label (bb)));
1866 /* E1 and E2 are edges with the same destination block. Search their
1867 predecessors for common code. If found, redirect control flow from
1868 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1869 or the other way around (dir_backward). DIR specifies the allowed
1870 replacement direction. */
1872 static bool
1873 try_crossjump_to_edge (int mode, edge e1, edge e2,
1874 enum replace_direction dir)
1876 int nmatch;
1877 basic_block src1 = e1->src, src2 = e2->src;
1878 basic_block redirect_to, redirect_from, to_remove;
1879 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1880 rtx newpos1, newpos2;
1881 edge s;
1882 edge_iterator ei;
1884 newpos1 = newpos2 = NULL_RTX;
1886 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1887 to try this optimization.
1889 Basic block partitioning may result in some jumps that appear to
1890 be optimizable (or blocks that appear to be mergeable), but which really
1891 must be left untouched (they are required to make it safely across
1892 partition boundaries). See the comments at the top of
1893 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1895 if (crtl->has_bb_partition && reload_completed)
1896 return false;
1898 /* Search backward through forwarder blocks. We don't need to worry
1899 about multiple entry or chained forwarders, as they will be optimized
1900 away. We do this to look past the unconditional jump following a
1901 conditional jump that is required due to the current CFG shape. */
1902 if (single_pred_p (src1)
1903 && FORWARDER_BLOCK_P (src1))
1904 e1 = single_pred_edge (src1), src1 = e1->src;
1906 if (single_pred_p (src2)
1907 && FORWARDER_BLOCK_P (src2))
1908 e2 = single_pred_edge (src2), src2 = e2->src;
1910 /* Nothing to do if we reach ENTRY, or a common source block. */
1911 if (src1 == ENTRY_BLOCK_PTR || src2 == ENTRY_BLOCK_PTR)
1912 return false;
1913 if (src1 == src2)
1914 return false;
1916 /* Seeing more than 1 forwarder blocks would confuse us later... */
1917 if (FORWARDER_BLOCK_P (e1->dest)
1918 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1919 return false;
1921 if (FORWARDER_BLOCK_P (e2->dest)
1922 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1923 return false;
1925 /* Likewise with dead code (possibly newly created by the other optimizations
1926 of cfg_cleanup). */
1927 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1928 return false;
1930 /* Look for the common insn sequence, part the first ... */
1931 if (!outgoing_edges_match (mode, src1, src2))
1932 return false;
1934 /* ... and part the second. */
1935 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1937 osrc1 = src1;
1938 osrc2 = src2;
1939 if (newpos1 != NULL_RTX)
1940 src1 = BLOCK_FOR_INSN (newpos1);
1941 if (newpos2 != NULL_RTX)
1942 src2 = BLOCK_FOR_INSN (newpos2);
1944 if (dir == dir_backward)
1946 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1947 SWAP (basic_block, osrc1, osrc2);
1948 SWAP (basic_block, src1, src2);
1949 SWAP (edge, e1, e2);
1950 SWAP (rtx, newpos1, newpos2);
1951 #undef SWAP
1954 /* Don't proceed with the crossjump unless we found a sufficient number
1955 of matching instructions or the 'from' block was totally matched
1956 (such that its predecessors will hopefully be redirected and the
1957 block removed). */
1958 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
1959 && (newpos1 != BB_HEAD (src1)))
1960 return false;
1962 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1963 if (block_has_preserve_label (e1->dest)
1964 && (e1->flags & EDGE_ABNORMAL))
1965 return false;
1967 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1968 will be deleted.
1969 If we have tablejumps in the end of SRC1 and SRC2
1970 they have been already compared for equivalence in outgoing_edges_match ()
1971 so replace the references to TABLE1 by references to TABLE2. */
1973 rtx label1, label2;
1974 rtx table1, table2;
1976 if (tablejump_p (BB_END (osrc1), &label1, &table1)
1977 && tablejump_p (BB_END (osrc2), &label2, &table2)
1978 && label1 != label2)
1980 replace_label_data rr;
1981 rtx insn;
1983 /* Replace references to LABEL1 with LABEL2. */
1984 rr.r1 = label1;
1985 rr.r2 = label2;
1986 rr.update_label_nuses = true;
1987 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
1989 /* Do not replace the label in SRC1->END because when deleting
1990 a block whose end is a tablejump, the tablejump referenced
1991 from the instruction is deleted too. */
1992 if (insn != BB_END (osrc1))
1993 for_each_rtx (&insn, replace_label, &rr);
1998 /* Avoid splitting if possible. We must always split when SRC2 has
1999 EH predecessor edges, or we may end up with basic blocks with both
2000 normal and EH predecessor edges. */
2001 if (newpos2 == BB_HEAD (src2)
2002 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2003 redirect_to = src2;
2004 else
2006 if (newpos2 == BB_HEAD (src2))
2008 /* Skip possible basic block header. */
2009 if (LABEL_P (newpos2))
2010 newpos2 = NEXT_INSN (newpos2);
2011 while (DEBUG_INSN_P (newpos2))
2012 newpos2 = NEXT_INSN (newpos2);
2013 if (NOTE_P (newpos2))
2014 newpos2 = NEXT_INSN (newpos2);
2015 while (DEBUG_INSN_P (newpos2))
2016 newpos2 = NEXT_INSN (newpos2);
2019 if (dump_file)
2020 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2021 src2->index, nmatch);
2022 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2025 if (dump_file)
2026 fprintf (dump_file,
2027 "Cross jumping from bb %i to bb %i; %i common insns\n",
2028 src1->index, src2->index, nmatch);
2030 /* We may have some registers visible through the block. */
2031 df_set_bb_dirty (redirect_to);
2033 if (osrc2 == src2)
2034 redirect_edges_to = redirect_to;
2035 else
2036 redirect_edges_to = osrc2;
2038 /* Recompute the frequencies and counts of outgoing edges. */
2039 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2041 edge s2;
2042 edge_iterator ei;
2043 basic_block d = s->dest;
2045 if (FORWARDER_BLOCK_P (d))
2046 d = single_succ (d);
2048 FOR_EACH_EDGE (s2, ei, src1->succs)
2050 basic_block d2 = s2->dest;
2051 if (FORWARDER_BLOCK_P (d2))
2052 d2 = single_succ (d2);
2053 if (d == d2)
2054 break;
2057 s->count += s2->count;
2059 /* Take care to update possible forwarder blocks. We verified
2060 that there is no more than one in the chain, so we can't run
2061 into infinite loop. */
2062 if (FORWARDER_BLOCK_P (s->dest))
2064 single_succ_edge (s->dest)->count += s2->count;
2065 s->dest->count += s2->count;
2066 s->dest->frequency += EDGE_FREQUENCY (s);
2069 if (FORWARDER_BLOCK_P (s2->dest))
2071 single_succ_edge (s2->dest)->count -= s2->count;
2072 if (single_succ_edge (s2->dest)->count < 0)
2073 single_succ_edge (s2->dest)->count = 0;
2074 s2->dest->count -= s2->count;
2075 s2->dest->frequency -= EDGE_FREQUENCY (s);
2076 if (s2->dest->frequency < 0)
2077 s2->dest->frequency = 0;
2078 if (s2->dest->count < 0)
2079 s2->dest->count = 0;
2082 if (!redirect_edges_to->frequency && !src1->frequency)
2083 s->probability = (s->probability + s2->probability) / 2;
2084 else
2085 s->probability
2086 = ((s->probability * redirect_edges_to->frequency +
2087 s2->probability * src1->frequency)
2088 / (redirect_edges_to->frequency + src1->frequency));
2091 /* Adjust count and frequency for the block. An earlier jump
2092 threading pass may have left the profile in an inconsistent
2093 state (see update_bb_profile_for_threading) so we must be
2094 prepared for overflows. */
2095 tmp = redirect_to;
2098 tmp->count += src1->count;
2099 tmp->frequency += src1->frequency;
2100 if (tmp->frequency > BB_FREQ_MAX)
2101 tmp->frequency = BB_FREQ_MAX;
2102 if (tmp == redirect_edges_to)
2103 break;
2104 tmp = find_fallthru_edge (tmp->succs)->dest;
2106 while (true);
2107 update_br_prob_note (redirect_edges_to);
2109 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2111 /* Skip possible basic block header. */
2112 if (LABEL_P (newpos1))
2113 newpos1 = NEXT_INSN (newpos1);
2115 while (DEBUG_INSN_P (newpos1))
2116 newpos1 = NEXT_INSN (newpos1);
2118 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2119 newpos1 = NEXT_INSN (newpos1);
2121 while (DEBUG_INSN_P (newpos1))
2122 newpos1 = NEXT_INSN (newpos1);
2124 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2125 to_remove = single_succ (redirect_from);
2127 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2128 delete_basic_block (to_remove);
2130 update_forwarder_flag (redirect_from);
2131 if (redirect_to != src2)
2132 update_forwarder_flag (src2);
2134 return true;
2137 /* Search the predecessors of BB for common insn sequences. When found,
2138 share code between them by redirecting control flow. Return true if
2139 any changes made. */
2141 static bool
2142 try_crossjump_bb (int mode, basic_block bb)
2144 edge e, e2, fallthru;
2145 bool changed;
2146 unsigned max, ix, ix2;
2148 /* Nothing to do if there is not at least two incoming edges. */
2149 if (EDGE_COUNT (bb->preds) < 2)
2150 return false;
2152 /* Don't crossjump if this block ends in a computed jump,
2153 unless we are optimizing for size. */
2154 if (optimize_bb_for_size_p (bb)
2155 && bb != EXIT_BLOCK_PTR
2156 && computed_jump_p (BB_END (bb)))
2157 return false;
2159 /* If we are partitioning hot/cold basic blocks, we don't want to
2160 mess up unconditional or indirect jumps that cross between hot
2161 and cold sections.
2163 Basic block partitioning may result in some jumps that appear to
2164 be optimizable (or blocks that appear to be mergeable), but which really
2165 must be left untouched (they are required to make it safely across
2166 partition boundaries). See the comments at the top of
2167 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2169 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2170 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2171 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2172 return false;
2174 /* It is always cheapest to redirect a block that ends in a branch to
2175 a block that falls through into BB, as that adds no branches to the
2176 program. We'll try that combination first. */
2177 fallthru = NULL;
2178 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2180 if (EDGE_COUNT (bb->preds) > max)
2181 return false;
2183 fallthru = find_fallthru_edge (bb->preds);
2185 changed = false;
2186 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2188 e = EDGE_PRED (bb, ix);
2189 ix++;
2191 /* As noted above, first try with the fallthru predecessor (or, a
2192 fallthru predecessor if we are in cfglayout mode). */
2193 if (fallthru)
2195 /* Don't combine the fallthru edge into anything else.
2196 If there is a match, we'll do it the other way around. */
2197 if (e == fallthru)
2198 continue;
2199 /* If nothing changed since the last attempt, there is nothing
2200 we can do. */
2201 if (!first_pass
2202 && !((e->src->flags & BB_MODIFIED)
2203 || (fallthru->src->flags & BB_MODIFIED)))
2204 continue;
2206 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2208 changed = true;
2209 ix = 0;
2210 continue;
2214 /* Non-obvious work limiting check: Recognize that we're going
2215 to call try_crossjump_bb on every basic block. So if we have
2216 two blocks with lots of outgoing edges (a switch) and they
2217 share lots of common destinations, then we would do the
2218 cross-jump check once for each common destination.
2220 Now, if the blocks actually are cross-jump candidates, then
2221 all of their destinations will be shared. Which means that
2222 we only need check them for cross-jump candidacy once. We
2223 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2224 choosing to do the check from the block for which the edge
2225 in question is the first successor of A. */
2226 if (EDGE_SUCC (e->src, 0) != e)
2227 continue;
2229 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2231 e2 = EDGE_PRED (bb, ix2);
2233 if (e2 == e)
2234 continue;
2236 /* We've already checked the fallthru edge above. */
2237 if (e2 == fallthru)
2238 continue;
2240 /* The "first successor" check above only prevents multiple
2241 checks of crossjump(A,B). In order to prevent redundant
2242 checks of crossjump(B,A), require that A be the block
2243 with the lowest index. */
2244 if (e->src->index > e2->src->index)
2245 continue;
2247 /* If nothing changed since the last attempt, there is nothing
2248 we can do. */
2249 if (!first_pass
2250 && !((e->src->flags & BB_MODIFIED)
2251 || (e2->src->flags & BB_MODIFIED)))
2252 continue;
2254 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2255 direction. */
2256 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2258 changed = true;
2259 ix = 0;
2260 break;
2265 if (changed)
2266 crossjumps_occured = true;
2268 return changed;
2271 /* Search the successors of BB for common insn sequences. When found,
2272 share code between them by moving it across the basic block
2273 boundary. Return true if any changes made. */
2275 static bool
2276 try_head_merge_bb (basic_block bb)
2278 basic_block final_dest_bb = NULL;
2279 int max_match = INT_MAX;
2280 edge e0;
2281 rtx *headptr, *currptr, *nextptr;
2282 bool changed, moveall;
2283 unsigned ix;
2284 rtx e0_last_head, cond, move_before;
2285 unsigned nedges = EDGE_COUNT (bb->succs);
2286 rtx jump = BB_END (bb);
2287 regset live, live_union;
2289 /* Nothing to do if there is not at least two outgoing edges. */
2290 if (nedges < 2)
2291 return false;
2293 /* Don't crossjump if this block ends in a computed jump,
2294 unless we are optimizing for size. */
2295 if (optimize_bb_for_size_p (bb)
2296 && bb != EXIT_BLOCK_PTR
2297 && computed_jump_p (BB_END (bb)))
2298 return false;
2300 cond = get_condition (jump, &move_before, true, false);
2301 if (cond == NULL_RTX)
2303 #ifdef HAVE_cc0
2304 if (reg_mentioned_p (cc0_rtx, jump))
2305 move_before = prev_nonnote_nondebug_insn (jump);
2306 else
2307 #endif
2308 move_before = jump;
2311 for (ix = 0; ix < nedges; ix++)
2312 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR)
2313 return false;
2315 for (ix = 0; ix < nedges; ix++)
2317 edge e = EDGE_SUCC (bb, ix);
2318 basic_block other_bb = e->dest;
2320 if (df_get_bb_dirty (other_bb))
2322 block_was_dirty = true;
2323 return false;
2326 if (e->flags & EDGE_ABNORMAL)
2327 return false;
2329 /* Normally, all destination blocks must only be reachable from this
2330 block, i.e. they must have one incoming edge.
2332 There is one special case we can handle, that of multiple consecutive
2333 jumps where the first jumps to one of the targets of the second jump.
2334 This happens frequently in switch statements for default labels.
2335 The structure is as follows:
2336 FINAL_DEST_BB
2337 ....
2338 if (cond) jump A;
2339 fall through
2341 jump with targets A, B, C, D...
2343 has two incoming edges, from FINAL_DEST_BB and BB
2345 In this case, we can try to move the insns through BB and into
2346 FINAL_DEST_BB. */
2347 if (EDGE_COUNT (other_bb->preds) != 1)
2349 edge incoming_edge, incoming_bb_other_edge;
2350 edge_iterator ei;
2352 if (final_dest_bb != NULL
2353 || EDGE_COUNT (other_bb->preds) != 2)
2354 return false;
2356 /* We must be able to move the insns across the whole block. */
2357 move_before = BB_HEAD (bb);
2358 while (!NONDEBUG_INSN_P (move_before))
2359 move_before = NEXT_INSN (move_before);
2361 if (EDGE_COUNT (bb->preds) != 1)
2362 return false;
2363 incoming_edge = EDGE_PRED (bb, 0);
2364 final_dest_bb = incoming_edge->src;
2365 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2366 return false;
2367 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2368 if (incoming_bb_other_edge != incoming_edge)
2369 break;
2370 if (incoming_bb_other_edge->dest != other_bb)
2371 return false;
2375 e0 = EDGE_SUCC (bb, 0);
2376 e0_last_head = NULL_RTX;
2377 changed = false;
2379 for (ix = 1; ix < nedges; ix++)
2381 edge e = EDGE_SUCC (bb, ix);
2382 rtx e0_last, e_last;
2383 int nmatch;
2385 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2386 &e0_last, &e_last, 0);
2387 if (nmatch == 0)
2388 return false;
2390 if (nmatch < max_match)
2392 max_match = nmatch;
2393 e0_last_head = e0_last;
2397 /* If we matched an entire block, we probably have to avoid moving the
2398 last insn. */
2399 if (max_match > 0
2400 && e0_last_head == BB_END (e0->dest)
2401 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2402 || control_flow_insn_p (e0_last_head)))
2404 max_match--;
2405 if (max_match == 0)
2406 return false;
2408 e0_last_head = prev_real_insn (e0_last_head);
2409 while (DEBUG_INSN_P (e0_last_head));
2412 if (max_match == 0)
2413 return false;
2415 /* We must find a union of the live registers at each of the end points. */
2416 live = BITMAP_ALLOC (NULL);
2417 live_union = BITMAP_ALLOC (NULL);
2419 currptr = XNEWVEC (rtx, nedges);
2420 headptr = XNEWVEC (rtx, nedges);
2421 nextptr = XNEWVEC (rtx, nedges);
2423 for (ix = 0; ix < nedges; ix++)
2425 int j;
2426 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2427 rtx head = BB_HEAD (merge_bb);
2429 while (!NONDEBUG_INSN_P (head))
2430 head = NEXT_INSN (head);
2431 headptr[ix] = head;
2432 currptr[ix] = head;
2434 /* Compute the end point and live information */
2435 for (j = 1; j < max_match; j++)
2437 head = NEXT_INSN (head);
2438 while (!NONDEBUG_INSN_P (head));
2439 simulate_backwards_to_point (merge_bb, live, head);
2440 IOR_REG_SET (live_union, live);
2443 /* If we're moving across two blocks, verify the validity of the
2444 first move, then adjust the target and let the loop below deal
2445 with the final move. */
2446 if (final_dest_bb != NULL)
2448 rtx move_upto;
2450 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2451 jump, e0->dest, live_union,
2452 NULL, &move_upto);
2453 if (!moveall)
2455 if (move_upto == NULL_RTX)
2456 goto out;
2458 while (e0_last_head != move_upto)
2460 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2461 live_union);
2462 e0_last_head = PREV_INSN (e0_last_head);
2465 if (e0_last_head == NULL_RTX)
2466 goto out;
2468 jump = BB_END (final_dest_bb);
2469 cond = get_condition (jump, &move_before, true, false);
2470 if (cond == NULL_RTX)
2472 #ifdef HAVE_cc0
2473 if (reg_mentioned_p (cc0_rtx, jump))
2474 move_before = prev_nonnote_nondebug_insn (jump);
2475 else
2476 #endif
2477 move_before = jump;
2483 rtx move_upto;
2484 moveall = can_move_insns_across (currptr[0], e0_last_head,
2485 move_before, jump, e0->dest, live_union,
2486 NULL, &move_upto);
2487 if (!moveall && move_upto == NULL_RTX)
2489 if (jump == move_before)
2490 break;
2492 /* Try again, using a different insertion point. */
2493 move_before = jump;
2495 #ifdef HAVE_cc0
2496 /* Don't try moving before a cc0 user, as that may invalidate
2497 the cc0. */
2498 if (reg_mentioned_p (cc0_rtx, jump))
2499 break;
2500 #endif
2502 continue;
2505 if (final_dest_bb && !moveall)
2506 /* We haven't checked whether a partial move would be OK for the first
2507 move, so we have to fail this case. */
2508 break;
2510 changed = true;
2511 for (;;)
2513 if (currptr[0] == move_upto)
2514 break;
2515 for (ix = 0; ix < nedges; ix++)
2517 rtx curr = currptr[ix];
2519 curr = NEXT_INSN (curr);
2520 while (!NONDEBUG_INSN_P (curr));
2521 currptr[ix] = curr;
2525 /* If we can't currently move all of the identical insns, remember
2526 each insn after the range that we'll merge. */
2527 if (!moveall)
2528 for (ix = 0; ix < nedges; ix++)
2530 rtx curr = currptr[ix];
2532 curr = NEXT_INSN (curr);
2533 while (!NONDEBUG_INSN_P (curr));
2534 nextptr[ix] = curr;
2537 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2538 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2539 if (final_dest_bb != NULL)
2540 df_set_bb_dirty (final_dest_bb);
2541 df_set_bb_dirty (bb);
2542 for (ix = 1; ix < nedges; ix++)
2544 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2545 delete_insn_chain (headptr[ix], currptr[ix], false);
2547 if (!moveall)
2549 if (jump == move_before)
2550 break;
2552 /* For the unmerged insns, try a different insertion point. */
2553 move_before = jump;
2555 #ifdef HAVE_cc0
2556 /* Don't try moving before a cc0 user, as that may invalidate
2557 the cc0. */
2558 if (reg_mentioned_p (cc0_rtx, jump))
2559 break;
2560 #endif
2562 for (ix = 0; ix < nedges; ix++)
2563 currptr[ix] = headptr[ix] = nextptr[ix];
2566 while (!moveall);
2568 out:
2569 free (currptr);
2570 free (headptr);
2571 free (nextptr);
2573 crossjumps_occured |= changed;
2575 return changed;
2578 /* Return true if BB contains just bb note, or bb note followed
2579 by only DEBUG_INSNs. */
2581 static bool
2582 trivially_empty_bb_p (basic_block bb)
2584 rtx insn = BB_END (bb);
2586 while (1)
2588 if (insn == BB_HEAD (bb))
2589 return true;
2590 if (!DEBUG_INSN_P (insn))
2591 return false;
2592 insn = PREV_INSN (insn);
2596 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2597 instructions etc. Return nonzero if changes were made. */
2599 static bool
2600 try_optimize_cfg (int mode)
2602 bool changed_overall = false;
2603 bool changed;
2604 int iterations = 0;
2605 basic_block bb, b, next;
2607 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2608 clear_bb_flags ();
2610 crossjumps_occured = false;
2612 FOR_EACH_BB (bb)
2613 update_forwarder_flag (bb);
2615 if (! targetm.cannot_modify_jumps_p ())
2617 first_pass = true;
2618 /* Attempt to merge blocks as made possible by edge removal. If
2619 a block has only one successor, and the successor has only
2620 one predecessor, they may be combined. */
2623 block_was_dirty = false;
2624 changed = false;
2625 iterations++;
2627 if (dump_file)
2628 fprintf (dump_file,
2629 "\n\ntry_optimize_cfg iteration %i\n\n",
2630 iterations);
2632 for (b = ENTRY_BLOCK_PTR->next_bb; b != EXIT_BLOCK_PTR;)
2634 basic_block c;
2635 edge s;
2636 bool changed_here = false;
2638 /* Delete trivially dead basic blocks. This is either
2639 blocks with no predecessors, or empty blocks with no
2640 successors. However if the empty block with no
2641 successors is the successor of the ENTRY_BLOCK, it is
2642 kept. This ensures that the ENTRY_BLOCK will have a
2643 successor which is a precondition for many RTL
2644 passes. Empty blocks may result from expanding
2645 __builtin_unreachable (). */
2646 if (EDGE_COUNT (b->preds) == 0
2647 || (EDGE_COUNT (b->succs) == 0
2648 && trivially_empty_bb_p (b)
2649 && single_succ_edge (ENTRY_BLOCK_PTR)->dest != b))
2651 c = b->prev_bb;
2652 if (EDGE_COUNT (b->preds) > 0)
2654 edge e;
2655 edge_iterator ei;
2657 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2659 if (BB_FOOTER (b)
2660 && BARRIER_P (BB_FOOTER (b)))
2661 FOR_EACH_EDGE (e, ei, b->preds)
2662 if ((e->flags & EDGE_FALLTHRU)
2663 && BB_FOOTER (e->src) == NULL)
2665 if (BB_FOOTER (b))
2667 BB_FOOTER (e->src) = BB_FOOTER (b);
2668 BB_FOOTER (b) = NULL;
2670 else
2672 start_sequence ();
2673 BB_FOOTER (e->src) = emit_barrier ();
2674 end_sequence ();
2678 else
2680 rtx last = get_last_bb_insn (b);
2681 if (last && BARRIER_P (last))
2682 FOR_EACH_EDGE (e, ei, b->preds)
2683 if ((e->flags & EDGE_FALLTHRU))
2684 emit_barrier_after (BB_END (e->src));
2687 delete_basic_block (b);
2688 changed = true;
2689 /* Avoid trying to remove ENTRY_BLOCK_PTR. */
2690 b = (c == ENTRY_BLOCK_PTR ? c->next_bb : c);
2691 continue;
2694 /* Remove code labels no longer used. */
2695 if (single_pred_p (b)
2696 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2697 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2698 && LABEL_P (BB_HEAD (b))
2699 /* If the previous block ends with a branch to this
2700 block, we can't delete the label. Normally this
2701 is a condjump that is yet to be simplified, but
2702 if CASE_DROPS_THRU, this can be a tablejump with
2703 some element going to the same place as the
2704 default (fallthru). */
2705 && (single_pred (b) == ENTRY_BLOCK_PTR
2706 || !JUMP_P (BB_END (single_pred (b)))
2707 || ! label_is_jump_target_p (BB_HEAD (b),
2708 BB_END (single_pred (b)))))
2710 delete_insn (BB_HEAD (b));
2711 if (dump_file)
2712 fprintf (dump_file, "Deleted label in block %i.\n",
2713 b->index);
2716 /* If we fall through an empty block, we can remove it. */
2717 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2718 && single_pred_p (b)
2719 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2720 && !LABEL_P (BB_HEAD (b))
2721 && FORWARDER_BLOCK_P (b)
2722 /* Note that forwarder_block_p true ensures that
2723 there is a successor for this block. */
2724 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2725 && n_basic_blocks > NUM_FIXED_BLOCKS + 1)
2727 if (dump_file)
2728 fprintf (dump_file,
2729 "Deleting fallthru block %i.\n",
2730 b->index);
2732 c = b->prev_bb == ENTRY_BLOCK_PTR ? b->next_bb : b->prev_bb;
2733 redirect_edge_succ_nodup (single_pred_edge (b),
2734 single_succ (b));
2735 delete_basic_block (b);
2736 changed = true;
2737 b = c;
2738 continue;
2741 /* Merge B with its single successor, if any. */
2742 if (single_succ_p (b)
2743 && (s = single_succ_edge (b))
2744 && !(s->flags & EDGE_COMPLEX)
2745 && (c = s->dest) != EXIT_BLOCK_PTR
2746 && single_pred_p (c)
2747 && b != c)
2749 /* When not in cfg_layout mode use code aware of reordering
2750 INSN. This code possibly creates new basic blocks so it
2751 does not fit merge_blocks interface and is kept here in
2752 hope that it will become useless once more of compiler
2753 is transformed to use cfg_layout mode. */
2755 if ((mode & CLEANUP_CFGLAYOUT)
2756 && can_merge_blocks_p (b, c))
2758 merge_blocks (b, c);
2759 update_forwarder_flag (b);
2760 changed_here = true;
2762 else if (!(mode & CLEANUP_CFGLAYOUT)
2763 /* If the jump insn has side effects,
2764 we can't kill the edge. */
2765 && (!JUMP_P (BB_END (b))
2766 || (reload_completed
2767 ? simplejump_p (BB_END (b))
2768 : (onlyjump_p (BB_END (b))
2769 && !tablejump_p (BB_END (b),
2770 NULL, NULL))))
2771 && (next = merge_blocks_move (s, b, c, mode)))
2773 b = next;
2774 changed_here = true;
2778 /* Simplify branch over branch. */
2779 if ((mode & CLEANUP_EXPENSIVE)
2780 && !(mode & CLEANUP_CFGLAYOUT)
2781 && try_simplify_condjump (b))
2782 changed_here = true;
2784 /* If B has a single outgoing edge, but uses a
2785 non-trivial jump instruction without side-effects, we
2786 can either delete the jump entirely, or replace it
2787 with a simple unconditional jump. */
2788 if (single_succ_p (b)
2789 && single_succ (b) != EXIT_BLOCK_PTR
2790 && onlyjump_p (BB_END (b))
2791 && !find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX)
2792 && try_redirect_by_replacing_jump (single_succ_edge (b),
2793 single_succ (b),
2794 (mode & CLEANUP_CFGLAYOUT) != 0))
2796 update_forwarder_flag (b);
2797 changed_here = true;
2800 /* Simplify branch to branch. */
2801 if (try_forward_edges (mode, b))
2803 update_forwarder_flag (b);
2804 changed_here = true;
2807 /* Look for shared code between blocks. */
2808 if ((mode & CLEANUP_CROSSJUMP)
2809 && try_crossjump_bb (mode, b))
2810 changed_here = true;
2812 if ((mode & CLEANUP_CROSSJUMP)
2813 /* This can lengthen register lifetimes. Do it only after
2814 reload. */
2815 && reload_completed
2816 && try_head_merge_bb (b))
2817 changed_here = true;
2819 /* Don't get confused by the index shift caused by
2820 deleting blocks. */
2821 if (!changed_here)
2822 b = b->next_bb;
2823 else
2824 changed = true;
2827 if ((mode & CLEANUP_CROSSJUMP)
2828 && try_crossjump_bb (mode, EXIT_BLOCK_PTR))
2829 changed = true;
2831 if (block_was_dirty)
2833 /* This should only be set by head-merging. */
2834 gcc_assert (mode & CLEANUP_CROSSJUMP);
2835 df_analyze ();
2838 if (changed)
2840 /* Edge forwarding in particular can cause hot blocks previously
2841 reached by both hot and cold blocks to become dominated only
2842 by cold blocks. This will cause the verification below to fail,
2843 and lead to now cold code in the hot section. This is not easy
2844 to detect and fix during edge forwarding, and in some cases
2845 is only visible after newly unreachable blocks are deleted,
2846 which will be done in fixup_partitions. */
2847 fixup_partitions ();
2849 #ifdef ENABLE_CHECKING
2850 verify_flow_info ();
2851 #endif
2854 changed_overall |= changed;
2855 first_pass = false;
2857 while (changed);
2860 FOR_ALL_BB (b)
2861 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2863 return changed_overall;
2866 /* Delete all unreachable basic blocks. */
2868 bool
2869 delete_unreachable_blocks (void)
2871 bool changed = false;
2872 basic_block b, prev_bb;
2874 find_unreachable_blocks ();
2876 /* When we're in GIMPLE mode and there may be debug insns, we should
2877 delete blocks in reverse dominator order, so as to get a chance
2878 to substitute all released DEFs into debug stmts. If we don't
2879 have dominators information, walking blocks backward gets us a
2880 better chance of retaining most debug information than
2881 otherwise. */
2882 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2883 && dom_info_available_p (CDI_DOMINATORS))
2885 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2887 prev_bb = b->prev_bb;
2889 if (!(b->flags & BB_REACHABLE))
2891 /* Speed up the removal of blocks that don't dominate
2892 others. Walking backwards, this should be the common
2893 case. */
2894 if (!first_dom_son (CDI_DOMINATORS, b))
2895 delete_basic_block (b);
2896 else
2898 vec<basic_block> h
2899 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2901 while (h.length ())
2903 b = h.pop ();
2905 prev_bb = b->prev_bb;
2907 gcc_assert (!(b->flags & BB_REACHABLE));
2909 delete_basic_block (b);
2912 h.release ();
2915 changed = true;
2919 else
2921 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2923 prev_bb = b->prev_bb;
2925 if (!(b->flags & BB_REACHABLE))
2927 delete_basic_block (b);
2928 changed = true;
2933 if (changed)
2934 tidy_fallthru_edges ();
2935 return changed;
2938 /* Delete any jump tables never referenced. We can't delete them at the
2939 time of removing tablejump insn as they are referenced by the preceding
2940 insns computing the destination, so we delay deleting and garbagecollect
2941 them once life information is computed. */
2942 void
2943 delete_dead_jumptables (void)
2945 basic_block bb;
2947 /* A dead jump table does not belong to any basic block. Scan insns
2948 between two adjacent basic blocks. */
2949 FOR_EACH_BB (bb)
2951 rtx insn, next;
2953 for (insn = NEXT_INSN (BB_END (bb));
2954 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
2955 insn = next)
2957 next = NEXT_INSN (insn);
2958 if (LABEL_P (insn)
2959 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
2960 && JUMP_TABLE_DATA_P (next))
2962 rtx label = insn, jump = next;
2964 if (dump_file)
2965 fprintf (dump_file, "Dead jumptable %i removed\n",
2966 INSN_UID (insn));
2968 next = NEXT_INSN (next);
2969 delete_insn (jump);
2970 delete_insn (label);
2977 /* Tidy the CFG by deleting unreachable code and whatnot. */
2979 bool
2980 cleanup_cfg (int mode)
2982 bool changed = false;
2984 /* Set the cfglayout mode flag here. We could update all the callers
2985 but that is just inconvenient, especially given that we eventually
2986 want to have cfglayout mode as the default. */
2987 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2988 mode |= CLEANUP_CFGLAYOUT;
2990 timevar_push (TV_CLEANUP_CFG);
2991 if (delete_unreachable_blocks ())
2993 changed = true;
2994 /* We've possibly created trivially dead code. Cleanup it right
2995 now to introduce more opportunities for try_optimize_cfg. */
2996 if (!(mode & (CLEANUP_NO_INSN_DEL))
2997 && !reload_completed)
2998 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3001 compact_blocks ();
3003 /* To tail-merge blocks ending in the same noreturn function (e.g.
3004 a call to abort) we have to insert fake edges to exit. Do this
3005 here once. The fake edges do not interfere with any other CFG
3006 cleanups. */
3007 if (mode & CLEANUP_CROSSJUMP)
3008 add_noreturn_fake_exit_edges ();
3010 if (!dbg_cnt (cfg_cleanup))
3011 return changed;
3013 while (try_optimize_cfg (mode))
3015 delete_unreachable_blocks (), changed = true;
3016 if (!(mode & CLEANUP_NO_INSN_DEL))
3018 /* Try to remove some trivially dead insns when doing an expensive
3019 cleanup. But delete_trivially_dead_insns doesn't work after
3020 reload (it only handles pseudos) and run_fast_dce is too costly
3021 to run in every iteration.
3023 For effective cross jumping, we really want to run a fast DCE to
3024 clean up any dead conditions, or they get in the way of performing
3025 useful tail merges.
3027 Other transformations in cleanup_cfg are not so sensitive to dead
3028 code, so delete_trivially_dead_insns or even doing nothing at all
3029 is good enough. */
3030 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3031 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3032 break;
3033 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
3034 run_fast_dce ();
3036 else
3037 break;
3040 if (mode & CLEANUP_CROSSJUMP)
3041 remove_fake_exit_edges ();
3043 /* Don't call delete_dead_jumptables in cfglayout mode, because
3044 that function assumes that jump tables are in the insns stream.
3045 But we also don't _have_ to delete dead jumptables in cfglayout
3046 mode because we shouldn't even be looking at things that are
3047 not in a basic block. Dead jumptables are cleaned up when
3048 going out of cfglayout mode. */
3049 if (!(mode & CLEANUP_CFGLAYOUT))
3050 delete_dead_jumptables ();
3052 /* ??? We probably do this way too often. */
3053 if (current_loops
3054 && (changed
3055 || (mode & CLEANUP_CFG_CHANGED)))
3057 timevar_push (TV_REPAIR_LOOPS);
3058 /* The above doesn't preserve dominance info if available. */
3059 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3060 calculate_dominance_info (CDI_DOMINATORS);
3061 fix_loop_structure (NULL);
3062 free_dominance_info (CDI_DOMINATORS);
3063 timevar_pop (TV_REPAIR_LOOPS);
3066 timevar_pop (TV_CLEANUP_CFG);
3068 return changed;
3071 static unsigned int
3072 execute_jump (void)
3074 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3075 if (dump_file)
3076 dump_flow_info (dump_file, dump_flags);
3077 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3078 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3079 return 0;
3082 struct rtl_opt_pass pass_jump =
3085 RTL_PASS,
3086 "jump", /* name */
3087 OPTGROUP_NONE, /* optinfo_flags */
3088 NULL, /* gate */
3089 execute_jump, /* execute */
3090 NULL, /* sub */
3091 NULL, /* next */
3092 0, /* static_pass_number */
3093 TV_JUMP, /* tv_id */
3094 0, /* properties_required */
3095 0, /* properties_provided */
3096 0, /* properties_destroyed */
3097 TODO_ggc_collect, /* todo_flags_start */
3098 TODO_verify_rtl_sharing, /* todo_flags_finish */
3102 static unsigned int
3103 execute_jump2 (void)
3105 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3106 return 0;
3109 struct rtl_opt_pass pass_jump2 =
3112 RTL_PASS,
3113 "jump2", /* name */
3114 OPTGROUP_NONE, /* optinfo_flags */
3115 NULL, /* gate */
3116 execute_jump2, /* execute */
3117 NULL, /* sub */
3118 NULL, /* next */
3119 0, /* static_pass_number */
3120 TV_JUMP, /* tv_id */
3121 0, /* properties_required */
3122 0, /* properties_provided */
3123 0, /* properties_destroyed */
3124 TODO_ggc_collect, /* todo_flags_start */
3125 TODO_verify_rtl_sharing, /* todo_flags_finish */