PR tree-optimization/78496
[official-gcc.git] / gcc / cfgcleanup.c
blobf68a964e31e9389944ef9587d99edeb19965b70f
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 /* Array of flags indexed by reg note kind, true if the given
1115 reg note is CFA related. */
1116 static const bool reg_note_cfa_p[] = {
1117 #undef REG_CFA_NOTE
1118 #define DEF_REG_NOTE(NAME) false,
1119 #define REG_CFA_NOTE(NAME) true,
1120 #include "reg-notes.def"
1121 #undef REG_CFA_NOTE
1122 #undef DEF_REG_NOTE
1123 false
1126 /* Return true if I1 and I2 have identical CFA notes (the same order
1127 and equivalent content). */
1129 static bool
1130 insns_have_identical_cfa_notes (rtx_insn *i1, rtx_insn *i2)
1132 rtx n1, n2;
1133 for (n1 = REG_NOTES (i1), n2 = REG_NOTES (i2); ;
1134 n1 = XEXP (n1, 1), n2 = XEXP (n2, 1))
1136 /* Skip over reg notes not related to CFI information. */
1137 while (n1 && !reg_note_cfa_p[REG_NOTE_KIND (n1)])
1138 n1 = XEXP (n1, 1);
1139 while (n2 && !reg_note_cfa_p[REG_NOTE_KIND (n2)])
1140 n2 = XEXP (n2, 1);
1141 if (n1 == NULL_RTX && n2 == NULL_RTX)
1142 return true;
1143 if (n1 == NULL_RTX || n2 == NULL_RTX)
1144 return false;
1145 if (XEXP (n1, 0) == XEXP (n2, 0))
1147 else if (XEXP (n1, 0) == NULL_RTX || XEXP (n2, 0) == NULL_RTX)
1148 return false;
1149 else if (!(reload_completed
1150 ? rtx_renumbered_equal_p (XEXP (n1, 0), XEXP (n2, 0))
1151 : rtx_equal_p (XEXP (n1, 0), XEXP (n2, 0))))
1152 return false;
1156 /* Examine I1 and I2 and return:
1157 - dir_forward if I1 can be replaced by I2, or
1158 - dir_backward if I2 can be replaced by I1, or
1159 - dir_both if both are the case. */
1161 static enum replace_direction
1162 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1164 rtx p1, p2;
1166 /* Verify that I1 and I2 are equivalent. */
1167 if (GET_CODE (i1) != GET_CODE (i2))
1168 return dir_none;
1170 /* __builtin_unreachable() may lead to empty blocks (ending with
1171 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1172 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1173 return dir_both;
1175 /* ??? Do not allow cross-jumping between different stack levels. */
1176 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1177 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1178 if (p1 && p2)
1180 p1 = XEXP (p1, 0);
1181 p2 = XEXP (p2, 0);
1182 if (!rtx_equal_p (p1, p2))
1183 return dir_none;
1185 /* ??? Worse, this adjustment had better be constant lest we
1186 have differing incoming stack levels. */
1187 if (!frame_pointer_needed
1188 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1189 return dir_none;
1191 else if (p1 || p2)
1192 return dir_none;
1194 /* Do not allow cross-jumping between frame related insns and other
1195 insns. */
1196 if (RTX_FRAME_RELATED_P (i1) != RTX_FRAME_RELATED_P (i2))
1197 return dir_none;
1199 p1 = PATTERN (i1);
1200 p2 = PATTERN (i2);
1202 if (GET_CODE (p1) != GET_CODE (p2))
1203 return dir_none;
1205 /* If this is a CALL_INSN, compare register usage information.
1206 If we don't check this on stack register machines, the two
1207 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1208 numbers of stack registers in the same basic block.
1209 If we don't check this on machines with delay slots, a delay slot may
1210 be filled that clobbers a parameter expected by the subroutine.
1212 ??? We take the simple route for now and assume that if they're
1213 equal, they were constructed identically.
1215 Also check for identical exception regions. */
1217 if (CALL_P (i1))
1219 /* Ensure the same EH region. */
1220 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1221 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1223 if (!n1 && n2)
1224 return dir_none;
1226 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1227 return dir_none;
1229 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1230 CALL_INSN_FUNCTION_USAGE (i2))
1231 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1232 return dir_none;
1234 /* For address sanitizer, never crossjump __asan_report_* builtins,
1235 otherwise errors might be reported on incorrect lines. */
1236 if (flag_sanitize & SANITIZE_ADDRESS)
1238 rtx call = get_call_rtx_from (i1);
1239 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1241 rtx symbol = XEXP (XEXP (call, 0), 0);
1242 if (SYMBOL_REF_DECL (symbol)
1243 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1245 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1246 == BUILT_IN_NORMAL)
1247 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1248 >= BUILT_IN_ASAN_REPORT_LOAD1
1249 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1250 <= BUILT_IN_ASAN_STOREN)
1251 return dir_none;
1257 /* If both i1 and i2 are frame related, verify all the CFA notes
1258 in the same order and with the same content. */
1259 if (RTX_FRAME_RELATED_P (i1) && !insns_have_identical_cfa_notes (i1, i2))
1260 return dir_none;
1262 #ifdef STACK_REGS
1263 /* If cross_jump_death_matters is not 0, the insn's mode
1264 indicates whether or not the insn contains any stack-like
1265 regs. */
1267 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1269 /* If register stack conversion has already been done, then
1270 death notes must also be compared before it is certain that
1271 the two instruction streams match. */
1273 rtx note;
1274 HARD_REG_SET i1_regset, i2_regset;
1276 CLEAR_HARD_REG_SET (i1_regset);
1277 CLEAR_HARD_REG_SET (i2_regset);
1279 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1280 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1281 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1283 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1284 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1285 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1287 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1288 return dir_none;
1290 #endif
1292 if (reload_completed
1293 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1294 return dir_both;
1296 return can_replace_by (i1, i2);
1299 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1300 flow_find_head_matching_sequence, ensure the notes match. */
1302 static void
1303 merge_notes (rtx_insn *i1, rtx_insn *i2)
1305 /* If the merged insns have different REG_EQUAL notes, then
1306 remove them. */
1307 rtx equiv1 = find_reg_equal_equiv_note (i1);
1308 rtx equiv2 = find_reg_equal_equiv_note (i2);
1310 if (equiv1 && !equiv2)
1311 remove_note (i1, equiv1);
1312 else if (!equiv1 && equiv2)
1313 remove_note (i2, equiv2);
1314 else if (equiv1 && equiv2
1315 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1317 remove_note (i1, equiv1);
1318 remove_note (i2, equiv2);
1322 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1323 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1324 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1325 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1326 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1328 static void
1329 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1330 bool *did_fallthru)
1332 edge fallthru;
1334 *did_fallthru = false;
1336 /* Ignore notes. */
1337 while (!NONDEBUG_INSN_P (*i1))
1339 if (*i1 != BB_HEAD (*bb1))
1341 *i1 = PREV_INSN (*i1);
1342 continue;
1345 if (!follow_fallthru)
1346 return;
1348 fallthru = find_fallthru_edge ((*bb1)->preds);
1349 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1350 || !single_succ_p (fallthru->src))
1351 return;
1353 *bb1 = fallthru->src;
1354 *i1 = BB_END (*bb1);
1355 *did_fallthru = true;
1359 /* Look through the insns at the end of BB1 and BB2 and find the longest
1360 sequence that are either equivalent, or allow forward or backward
1361 replacement. Store the first insns for that sequence in *F1 and *F2 and
1362 return the sequence length.
1364 DIR_P indicates the allowed replacement direction on function entry, and
1365 the actual replacement direction on function exit. If NULL, only equivalent
1366 sequences are allowed.
1368 To simplify callers of this function, if the blocks match exactly,
1369 store the head of the blocks in *F1 and *F2. */
1372 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1373 rtx_insn **f2, enum replace_direction *dir_p)
1375 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1376 int ninsns = 0;
1377 enum replace_direction dir, last_dir, afterlast_dir;
1378 bool follow_fallthru, did_fallthru;
1380 if (dir_p)
1381 dir = *dir_p;
1382 else
1383 dir = dir_both;
1384 afterlast_dir = dir;
1385 last_dir = afterlast_dir;
1387 /* Skip simple jumps at the end of the blocks. Complex jumps still
1388 need to be compared for equivalence, which we'll do below. */
1390 i1 = BB_END (bb1);
1391 last1 = afterlast1 = last2 = afterlast2 = NULL;
1392 if (onlyjump_p (i1)
1393 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1395 last1 = i1;
1396 i1 = PREV_INSN (i1);
1399 i2 = BB_END (bb2);
1400 if (onlyjump_p (i2)
1401 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1403 last2 = i2;
1404 /* Count everything except for unconditional jump as insn.
1405 Don't count any jumps if dir_p is NULL. */
1406 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1407 ninsns++;
1408 i2 = PREV_INSN (i2);
1411 while (true)
1413 /* In the following example, we can replace all jumps to C by jumps to A.
1415 This removes 4 duplicate insns.
1416 [bb A] insn1 [bb C] insn1
1417 insn2 insn2
1418 [bb B] insn3 insn3
1419 insn4 insn4
1420 jump_insn jump_insn
1422 We could also replace all jumps to A by jumps to C, but that leaves B
1423 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1424 step, all jumps to B would be replaced with jumps to the middle of C,
1425 achieving the same result with more effort.
1426 So we allow only the first possibility, which means that we don't allow
1427 fallthru in the block that's being replaced. */
1429 follow_fallthru = dir_p && dir != dir_forward;
1430 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1431 if (did_fallthru)
1432 dir = dir_backward;
1434 follow_fallthru = dir_p && dir != dir_backward;
1435 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1436 if (did_fallthru)
1437 dir = dir_forward;
1439 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1440 break;
1442 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1443 if (dir == dir_none || (!dir_p && dir != dir_both))
1444 break;
1446 merge_memattrs (i1, i2);
1448 /* Don't begin a cross-jump with a NOTE insn. */
1449 if (INSN_P (i1))
1451 merge_notes (i1, i2);
1453 afterlast1 = last1, afterlast2 = last2;
1454 last1 = i1, last2 = i2;
1455 afterlast_dir = last_dir;
1456 last_dir = dir;
1457 if (active_insn_p (i1))
1458 ninsns++;
1461 i1 = PREV_INSN (i1);
1462 i2 = PREV_INSN (i2);
1465 /* Don't allow the insn after a compare to be shared by
1466 cross-jumping unless the compare is also shared. */
1467 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1468 && ! sets_cc0_p (last1))
1469 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1471 /* Include preceding notes and labels in the cross-jump. One,
1472 this may bring us to the head of the blocks as requested above.
1473 Two, it keeps line number notes as matched as may be. */
1474 if (ninsns)
1476 bb1 = BLOCK_FOR_INSN (last1);
1477 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1478 last1 = PREV_INSN (last1);
1480 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1481 last1 = PREV_INSN (last1);
1483 bb2 = BLOCK_FOR_INSN (last2);
1484 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1485 last2 = PREV_INSN (last2);
1487 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1488 last2 = PREV_INSN (last2);
1490 *f1 = last1;
1491 *f2 = last2;
1494 if (dir_p)
1495 *dir_p = last_dir;
1496 return ninsns;
1499 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1500 the head of the two blocks. Do not include jumps at the end.
1501 If STOP_AFTER is nonzero, stop after finding that many matching
1502 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1503 non-zero, only count active insns. */
1506 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1507 rtx_insn **f2, int stop_after)
1509 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1510 int ninsns = 0;
1511 edge e;
1512 edge_iterator ei;
1513 int nehedges1 = 0, nehedges2 = 0;
1515 FOR_EACH_EDGE (e, ei, bb1->succs)
1516 if (e->flags & EDGE_EH)
1517 nehedges1++;
1518 FOR_EACH_EDGE (e, ei, bb2->succs)
1519 if (e->flags & EDGE_EH)
1520 nehedges2++;
1522 i1 = BB_HEAD (bb1);
1523 i2 = BB_HEAD (bb2);
1524 last1 = beforelast1 = last2 = beforelast2 = NULL;
1526 while (true)
1528 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1529 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1531 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1532 break;
1533 i1 = NEXT_INSN (i1);
1536 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1538 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1539 break;
1540 i2 = NEXT_INSN (i2);
1543 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1544 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1545 break;
1547 if (NOTE_P (i1) || NOTE_P (i2)
1548 || JUMP_P (i1) || JUMP_P (i2))
1549 break;
1551 /* A sanity check to make sure we're not merging insns with different
1552 effects on EH. If only one of them ends a basic block, it shouldn't
1553 have an EH edge; if both end a basic block, there should be the same
1554 number of EH edges. */
1555 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1556 && nehedges1 > 0)
1557 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1558 && nehedges2 > 0)
1559 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1560 && nehedges1 != nehedges2))
1561 break;
1563 if (old_insns_match_p (0, i1, i2) != dir_both)
1564 break;
1566 merge_memattrs (i1, i2);
1568 /* Don't begin a cross-jump with a NOTE insn. */
1569 if (INSN_P (i1))
1571 merge_notes (i1, i2);
1573 beforelast1 = last1, beforelast2 = last2;
1574 last1 = i1, last2 = i2;
1575 if (!stop_after || active_insn_p (i1))
1576 ninsns++;
1579 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1580 || (stop_after > 0 && ninsns == stop_after))
1581 break;
1583 i1 = NEXT_INSN (i1);
1584 i2 = NEXT_INSN (i2);
1587 /* Don't allow a compare to be shared by cross-jumping unless the insn
1588 after the compare is also shared. */
1589 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1590 && sets_cc0_p (last1))
1591 last1 = beforelast1, last2 = beforelast2, ninsns--;
1593 if (ninsns)
1595 *f1 = last1;
1596 *f2 = last2;
1599 return ninsns;
1602 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1603 the branch instruction. This means that if we commonize the control
1604 flow before end of the basic block, the semantic remains unchanged.
1606 We may assume that there exists one edge with a common destination. */
1608 static bool
1609 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1611 int nehedges1 = 0, nehedges2 = 0;
1612 edge fallthru1 = 0, fallthru2 = 0;
1613 edge e1, e2;
1614 edge_iterator ei;
1616 /* If we performed shrink-wrapping, edges to the exit block can
1617 only be distinguished for JUMP_INSNs. The two paths may differ in
1618 whether they went through the prologue. Sibcalls are fine, we know
1619 that we either didn't need or inserted an epilogue before them. */
1620 if (crtl->shrink_wrapped
1621 && single_succ_p (bb1)
1622 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1623 && !JUMP_P (BB_END (bb1))
1624 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1625 return false;
1627 /* If BB1 has only one successor, we may be looking at either an
1628 unconditional jump, or a fake edge to exit. */
1629 if (single_succ_p (bb1)
1630 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1631 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1632 return (single_succ_p (bb2)
1633 && (single_succ_edge (bb2)->flags
1634 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1635 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1637 /* Match conditional jumps - this may get tricky when fallthru and branch
1638 edges are crossed. */
1639 if (EDGE_COUNT (bb1->succs) == 2
1640 && any_condjump_p (BB_END (bb1))
1641 && onlyjump_p (BB_END (bb1)))
1643 edge b1, f1, b2, f2;
1644 bool reverse, match;
1645 rtx set1, set2, cond1, cond2;
1646 enum rtx_code code1, code2;
1648 if (EDGE_COUNT (bb2->succs) != 2
1649 || !any_condjump_p (BB_END (bb2))
1650 || !onlyjump_p (BB_END (bb2)))
1651 return false;
1653 b1 = BRANCH_EDGE (bb1);
1654 b2 = BRANCH_EDGE (bb2);
1655 f1 = FALLTHRU_EDGE (bb1);
1656 f2 = FALLTHRU_EDGE (bb2);
1658 /* Get around possible forwarders on fallthru edges. Other cases
1659 should be optimized out already. */
1660 if (FORWARDER_BLOCK_P (f1->dest))
1661 f1 = single_succ_edge (f1->dest);
1663 if (FORWARDER_BLOCK_P (f2->dest))
1664 f2 = single_succ_edge (f2->dest);
1666 /* To simplify use of this function, return false if there are
1667 unneeded forwarder blocks. These will get eliminated later
1668 during cleanup_cfg. */
1669 if (FORWARDER_BLOCK_P (f1->dest)
1670 || FORWARDER_BLOCK_P (f2->dest)
1671 || FORWARDER_BLOCK_P (b1->dest)
1672 || FORWARDER_BLOCK_P (b2->dest))
1673 return false;
1675 if (f1->dest == f2->dest && b1->dest == b2->dest)
1676 reverse = false;
1677 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1678 reverse = true;
1679 else
1680 return false;
1682 set1 = pc_set (BB_END (bb1));
1683 set2 = pc_set (BB_END (bb2));
1684 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1685 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1686 reverse = !reverse;
1688 cond1 = XEXP (SET_SRC (set1), 0);
1689 cond2 = XEXP (SET_SRC (set2), 0);
1690 code1 = GET_CODE (cond1);
1691 if (reverse)
1692 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1693 else
1694 code2 = GET_CODE (cond2);
1696 if (code2 == UNKNOWN)
1697 return false;
1699 /* Verify codes and operands match. */
1700 match = ((code1 == code2
1701 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1702 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1703 || (code1 == swap_condition (code2)
1704 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1705 XEXP (cond2, 0))
1706 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1707 XEXP (cond2, 1))));
1709 /* If we return true, we will join the blocks. Which means that
1710 we will only have one branch prediction bit to work with. Thus
1711 we require the existing branches to have probabilities that are
1712 roughly similar. */
1713 if (match
1714 && optimize_bb_for_speed_p (bb1)
1715 && optimize_bb_for_speed_p (bb2))
1717 int prob2;
1719 if (b1->dest == b2->dest)
1720 prob2 = b2->probability;
1721 else
1722 /* Do not use f2 probability as f2 may be forwarded. */
1723 prob2 = REG_BR_PROB_BASE - b2->probability;
1725 /* Fail if the difference in probabilities is greater than 50%.
1726 This rules out two well-predicted branches with opposite
1727 outcomes. */
1728 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1730 if (dump_file)
1731 fprintf (dump_file,
1732 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1733 bb1->index, bb2->index, b1->probability, prob2);
1735 return false;
1739 if (dump_file && match)
1740 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1741 bb1->index, bb2->index);
1743 return match;
1746 /* Generic case - we are seeing a computed jump, table jump or trapping
1747 instruction. */
1749 /* Check whether there are tablejumps in the end of BB1 and BB2.
1750 Return true if they are identical. */
1752 rtx_insn *label1, *label2;
1753 rtx_jump_table_data *table1, *table2;
1755 if (tablejump_p (BB_END (bb1), &label1, &table1)
1756 && tablejump_p (BB_END (bb2), &label2, &table2)
1757 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1759 /* The labels should never be the same rtx. If they really are same
1760 the jump tables are same too. So disable crossjumping of blocks BB1
1761 and BB2 because when deleting the common insns in the end of BB1
1762 by delete_basic_block () the jump table would be deleted too. */
1763 /* If LABEL2 is referenced in BB1->END do not do anything
1764 because we would loose information when replacing
1765 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1766 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1768 /* Set IDENTICAL to true when the tables are identical. */
1769 bool identical = false;
1770 rtx p1, p2;
1772 p1 = PATTERN (table1);
1773 p2 = PATTERN (table2);
1774 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1776 identical = true;
1778 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1779 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1780 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1781 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1783 int i;
1785 identical = true;
1786 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1787 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1788 identical = false;
1791 if (identical)
1793 bool match;
1795 /* Temporarily replace references to LABEL1 with LABEL2
1796 in BB1->END so that we could compare the instructions. */
1797 replace_label_in_insn (BB_END (bb1), label1, label2, false);
1799 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1800 == dir_both);
1801 if (dump_file && match)
1802 fprintf (dump_file,
1803 "Tablejumps in bb %i and %i match.\n",
1804 bb1->index, bb2->index);
1806 /* Set the original label in BB1->END because when deleting
1807 a block whose end is a tablejump, the tablejump referenced
1808 from the instruction is deleted too. */
1809 replace_label_in_insn (BB_END (bb1), label2, label1, false);
1811 return match;
1814 return false;
1818 /* Find the last non-debug non-note instruction in each bb, except
1819 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1820 handles that case specially. old_insns_match_p does not handle
1821 other types of instruction notes. */
1822 rtx_insn *last1 = BB_END (bb1);
1823 rtx_insn *last2 = BB_END (bb2);
1824 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1825 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1826 last1 = PREV_INSN (last1);
1827 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1828 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1829 last2 = PREV_INSN (last2);
1830 gcc_assert (last1 && last2);
1832 /* First ensure that the instructions match. There may be many outgoing
1833 edges so this test is generally cheaper. */
1834 if (old_insns_match_p (mode, last1, last2) != dir_both)
1835 return false;
1837 /* Search the outgoing edges, ensure that the counts do match, find possible
1838 fallthru and exception handling edges since these needs more
1839 validation. */
1840 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1841 return false;
1843 bool nonfakeedges = false;
1844 FOR_EACH_EDGE (e1, ei, bb1->succs)
1846 e2 = EDGE_SUCC (bb2, ei.index);
1848 if ((e1->flags & EDGE_FAKE) == 0)
1849 nonfakeedges = true;
1851 if (e1->flags & EDGE_EH)
1852 nehedges1++;
1854 if (e2->flags & EDGE_EH)
1855 nehedges2++;
1857 if (e1->flags & EDGE_FALLTHRU)
1858 fallthru1 = e1;
1859 if (e2->flags & EDGE_FALLTHRU)
1860 fallthru2 = e2;
1863 /* If number of edges of various types does not match, fail. */
1864 if (nehedges1 != nehedges2
1865 || (fallthru1 != 0) != (fallthru2 != 0))
1866 return false;
1868 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1869 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1870 attempt to optimize, as the two basic blocks might have different
1871 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1872 traps there should be REG_ARG_SIZE notes, they could be missing
1873 for __builtin_unreachable () uses though. */
1874 if (!nonfakeedges
1875 && !ACCUMULATE_OUTGOING_ARGS
1876 && (!INSN_P (last1)
1877 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1878 return false;
1880 /* fallthru edges must be forwarded to the same destination. */
1881 if (fallthru1)
1883 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1884 ? single_succ (fallthru1->dest): fallthru1->dest);
1885 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1886 ? single_succ (fallthru2->dest): fallthru2->dest);
1888 if (d1 != d2)
1889 return false;
1892 /* Ensure the same EH region. */
1894 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1895 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1897 if (!n1 && n2)
1898 return false;
1900 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1901 return false;
1904 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1905 version of sequence abstraction. */
1906 FOR_EACH_EDGE (e1, ei, bb2->succs)
1908 edge e2;
1909 edge_iterator ei;
1910 basic_block d1 = e1->dest;
1912 if (FORWARDER_BLOCK_P (d1))
1913 d1 = EDGE_SUCC (d1, 0)->dest;
1915 FOR_EACH_EDGE (e2, ei, bb1->succs)
1917 basic_block d2 = e2->dest;
1918 if (FORWARDER_BLOCK_P (d2))
1919 d2 = EDGE_SUCC (d2, 0)->dest;
1920 if (d1 == d2)
1921 break;
1924 if (!e2)
1925 return false;
1928 return true;
1931 /* Returns true if BB basic block has a preserve label. */
1933 static bool
1934 block_has_preserve_label (basic_block bb)
1936 return (bb
1937 && block_label (bb)
1938 && LABEL_PRESERVE_P (block_label (bb)));
1941 /* E1 and E2 are edges with the same destination block. Search their
1942 predecessors for common code. If found, redirect control flow from
1943 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1944 or the other way around (dir_backward). DIR specifies the allowed
1945 replacement direction. */
1947 static bool
1948 try_crossjump_to_edge (int mode, edge e1, edge e2,
1949 enum replace_direction dir)
1951 int nmatch;
1952 basic_block src1 = e1->src, src2 = e2->src;
1953 basic_block redirect_to, redirect_from, to_remove;
1954 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1955 rtx_insn *newpos1, *newpos2;
1956 edge s;
1957 edge_iterator ei;
1959 newpos1 = newpos2 = NULL;
1961 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1962 to try this optimization.
1964 Basic block partitioning may result in some jumps that appear to
1965 be optimizable (or blocks that appear to be mergeable), but which really
1966 must be left untouched (they are required to make it safely across
1967 partition boundaries). See the comments at the top of
1968 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1970 if (crtl->has_bb_partition && reload_completed)
1971 return false;
1973 /* Search backward through forwarder blocks. We don't need to worry
1974 about multiple entry or chained forwarders, as they will be optimized
1975 away. We do this to look past the unconditional jump following a
1976 conditional jump that is required due to the current CFG shape. */
1977 if (single_pred_p (src1)
1978 && FORWARDER_BLOCK_P (src1))
1979 e1 = single_pred_edge (src1), src1 = e1->src;
1981 if (single_pred_p (src2)
1982 && FORWARDER_BLOCK_P (src2))
1983 e2 = single_pred_edge (src2), src2 = e2->src;
1985 /* Nothing to do if we reach ENTRY, or a common source block. */
1986 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1987 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1988 return false;
1989 if (src1 == src2)
1990 return false;
1992 /* Seeing more than 1 forwarder blocks would confuse us later... */
1993 if (FORWARDER_BLOCK_P (e1->dest)
1994 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1995 return false;
1997 if (FORWARDER_BLOCK_P (e2->dest)
1998 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1999 return false;
2001 /* Likewise with dead code (possibly newly created by the other optimizations
2002 of cfg_cleanup). */
2003 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
2004 return false;
2006 /* Look for the common insn sequence, part the first ... */
2007 if (!outgoing_edges_match (mode, src1, src2))
2008 return false;
2010 /* ... and part the second. */
2011 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
2013 osrc1 = src1;
2014 osrc2 = src2;
2015 if (newpos1 != NULL_RTX)
2016 src1 = BLOCK_FOR_INSN (newpos1);
2017 if (newpos2 != NULL_RTX)
2018 src2 = BLOCK_FOR_INSN (newpos2);
2020 /* Check that SRC1 and SRC2 have preds again. They may have changed
2021 above due to the call to flow_find_cross_jump. */
2022 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
2023 return false;
2025 if (dir == dir_backward)
2027 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
2028 SWAP (basic_block, osrc1, osrc2);
2029 SWAP (basic_block, src1, src2);
2030 SWAP (edge, e1, e2);
2031 SWAP (rtx_insn *, newpos1, newpos2);
2032 #undef SWAP
2035 /* Don't proceed with the crossjump unless we found a sufficient number
2036 of matching instructions or the 'from' block was totally matched
2037 (such that its predecessors will hopefully be redirected and the
2038 block removed). */
2039 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
2040 && (newpos1 != BB_HEAD (src1)))
2041 return false;
2043 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
2044 if (block_has_preserve_label (e1->dest)
2045 && (e1->flags & EDGE_ABNORMAL))
2046 return false;
2048 /* Here we know that the insns in the end of SRC1 which are common with SRC2
2049 will be deleted.
2050 If we have tablejumps in the end of SRC1 and SRC2
2051 they have been already compared for equivalence in outgoing_edges_match ()
2052 so replace the references to TABLE1 by references to TABLE2. */
2054 rtx_insn *label1, *label2;
2055 rtx_jump_table_data *table1, *table2;
2057 if (tablejump_p (BB_END (osrc1), &label1, &table1)
2058 && tablejump_p (BB_END (osrc2), &label2, &table2)
2059 && label1 != label2)
2061 rtx_insn *insn;
2063 /* Replace references to LABEL1 with LABEL2. */
2064 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2066 /* Do not replace the label in SRC1->END because when deleting
2067 a block whose end is a tablejump, the tablejump referenced
2068 from the instruction is deleted too. */
2069 if (insn != BB_END (osrc1))
2070 replace_label_in_insn (insn, label1, label2, true);
2075 /* Avoid splitting if possible. We must always split when SRC2 has
2076 EH predecessor edges, or we may end up with basic blocks with both
2077 normal and EH predecessor edges. */
2078 if (newpos2 == BB_HEAD (src2)
2079 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2080 redirect_to = src2;
2081 else
2083 if (newpos2 == BB_HEAD (src2))
2085 /* Skip possible basic block header. */
2086 if (LABEL_P (newpos2))
2087 newpos2 = NEXT_INSN (newpos2);
2088 while (DEBUG_INSN_P (newpos2))
2089 newpos2 = NEXT_INSN (newpos2);
2090 if (NOTE_P (newpos2))
2091 newpos2 = NEXT_INSN (newpos2);
2092 while (DEBUG_INSN_P (newpos2))
2093 newpos2 = NEXT_INSN (newpos2);
2096 if (dump_file)
2097 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2098 src2->index, nmatch);
2099 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2102 if (dump_file)
2103 fprintf (dump_file,
2104 "Cross jumping from bb %i to bb %i; %i common insns\n",
2105 src1->index, src2->index, nmatch);
2107 /* We may have some registers visible through the block. */
2108 df_set_bb_dirty (redirect_to);
2110 if (osrc2 == src2)
2111 redirect_edges_to = redirect_to;
2112 else
2113 redirect_edges_to = osrc2;
2115 /* Recompute the frequencies and counts of outgoing edges. */
2116 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2118 edge s2;
2119 edge_iterator ei;
2120 basic_block d = s->dest;
2122 if (FORWARDER_BLOCK_P (d))
2123 d = single_succ (d);
2125 FOR_EACH_EDGE (s2, ei, src1->succs)
2127 basic_block d2 = s2->dest;
2128 if (FORWARDER_BLOCK_P (d2))
2129 d2 = single_succ (d2);
2130 if (d == d2)
2131 break;
2134 s->count += s2->count;
2136 /* Take care to update possible forwarder blocks. We verified
2137 that there is no more than one in the chain, so we can't run
2138 into infinite loop. */
2139 if (FORWARDER_BLOCK_P (s->dest))
2141 single_succ_edge (s->dest)->count += s2->count;
2142 s->dest->count += s2->count;
2143 s->dest->frequency += EDGE_FREQUENCY (s);
2146 if (FORWARDER_BLOCK_P (s2->dest))
2148 single_succ_edge (s2->dest)->count -= s2->count;
2149 if (single_succ_edge (s2->dest)->count < 0)
2150 single_succ_edge (s2->dest)->count = 0;
2151 s2->dest->count -= s2->count;
2152 s2->dest->frequency -= EDGE_FREQUENCY (s);
2153 if (s2->dest->frequency < 0)
2154 s2->dest->frequency = 0;
2155 if (s2->dest->count < 0)
2156 s2->dest->count = 0;
2159 if (!redirect_edges_to->frequency && !src1->frequency)
2160 s->probability = (s->probability + s2->probability) / 2;
2161 else
2162 s->probability
2163 = ((s->probability * redirect_edges_to->frequency +
2164 s2->probability * src1->frequency)
2165 / (redirect_edges_to->frequency + src1->frequency));
2168 /* Adjust count and frequency for the block. An earlier jump
2169 threading pass may have left the profile in an inconsistent
2170 state (see update_bb_profile_for_threading) so we must be
2171 prepared for overflows. */
2172 tmp = redirect_to;
2175 tmp->count += src1->count;
2176 tmp->frequency += src1->frequency;
2177 if (tmp->frequency > BB_FREQ_MAX)
2178 tmp->frequency = BB_FREQ_MAX;
2179 if (tmp == redirect_edges_to)
2180 break;
2181 tmp = find_fallthru_edge (tmp->succs)->dest;
2183 while (true);
2184 update_br_prob_note (redirect_edges_to);
2186 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2188 /* Skip possible basic block header. */
2189 if (LABEL_P (newpos1))
2190 newpos1 = NEXT_INSN (newpos1);
2192 while (DEBUG_INSN_P (newpos1))
2193 newpos1 = NEXT_INSN (newpos1);
2195 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2196 newpos1 = NEXT_INSN (newpos1);
2198 while (DEBUG_INSN_P (newpos1))
2199 newpos1 = NEXT_INSN (newpos1);
2201 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2202 to_remove = single_succ (redirect_from);
2204 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2205 delete_basic_block (to_remove);
2207 update_forwarder_flag (redirect_from);
2208 if (redirect_to != src2)
2209 update_forwarder_flag (src2);
2211 return true;
2214 /* Search the predecessors of BB for common insn sequences. When found,
2215 share code between them by redirecting control flow. Return true if
2216 any changes made. */
2218 static bool
2219 try_crossjump_bb (int mode, basic_block bb)
2221 edge e, e2, fallthru;
2222 bool changed;
2223 unsigned max, ix, ix2;
2225 /* Nothing to do if there is not at least two incoming edges. */
2226 if (EDGE_COUNT (bb->preds) < 2)
2227 return false;
2229 /* Don't crossjump if this block ends in a computed jump,
2230 unless we are optimizing for size. */
2231 if (optimize_bb_for_size_p (bb)
2232 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2233 && computed_jump_p (BB_END (bb)))
2234 return false;
2236 /* If we are partitioning hot/cold basic blocks, we don't want to
2237 mess up unconditional or indirect jumps that cross between hot
2238 and cold sections.
2240 Basic block partitioning may result in some jumps that appear to
2241 be optimizable (or blocks that appear to be mergeable), but which really
2242 must be left untouched (they are required to make it safely across
2243 partition boundaries). See the comments at the top of
2244 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2246 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2247 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2248 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2249 return false;
2251 /* It is always cheapest to redirect a block that ends in a branch to
2252 a block that falls through into BB, as that adds no branches to the
2253 program. We'll try that combination first. */
2254 fallthru = NULL;
2255 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2257 if (EDGE_COUNT (bb->preds) > max)
2258 return false;
2260 fallthru = find_fallthru_edge (bb->preds);
2262 changed = false;
2263 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2265 e = EDGE_PRED (bb, ix);
2266 ix++;
2268 /* As noted above, first try with the fallthru predecessor (or, a
2269 fallthru predecessor if we are in cfglayout mode). */
2270 if (fallthru)
2272 /* Don't combine the fallthru edge into anything else.
2273 If there is a match, we'll do it the other way around. */
2274 if (e == fallthru)
2275 continue;
2276 /* If nothing changed since the last attempt, there is nothing
2277 we can do. */
2278 if (!first_pass
2279 && !((e->src->flags & BB_MODIFIED)
2280 || (fallthru->src->flags & BB_MODIFIED)))
2281 continue;
2283 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2285 changed = true;
2286 ix = 0;
2287 continue;
2291 /* Non-obvious work limiting check: Recognize that we're going
2292 to call try_crossjump_bb on every basic block. So if we have
2293 two blocks with lots of outgoing edges (a switch) and they
2294 share lots of common destinations, then we would do the
2295 cross-jump check once for each common destination.
2297 Now, if the blocks actually are cross-jump candidates, then
2298 all of their destinations will be shared. Which means that
2299 we only need check them for cross-jump candidacy once. We
2300 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2301 choosing to do the check from the block for which the edge
2302 in question is the first successor of A. */
2303 if (EDGE_SUCC (e->src, 0) != e)
2304 continue;
2306 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2308 e2 = EDGE_PRED (bb, ix2);
2310 if (e2 == e)
2311 continue;
2313 /* We've already checked the fallthru edge above. */
2314 if (e2 == fallthru)
2315 continue;
2317 /* The "first successor" check above only prevents multiple
2318 checks of crossjump(A,B). In order to prevent redundant
2319 checks of crossjump(B,A), require that A be the block
2320 with the lowest index. */
2321 if (e->src->index > e2->src->index)
2322 continue;
2324 /* If nothing changed since the last attempt, there is nothing
2325 we can do. */
2326 if (!first_pass
2327 && !((e->src->flags & BB_MODIFIED)
2328 || (e2->src->flags & BB_MODIFIED)))
2329 continue;
2331 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2332 direction. */
2333 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2335 changed = true;
2336 ix = 0;
2337 break;
2342 if (changed)
2343 crossjumps_occurred = true;
2345 return changed;
2348 /* Search the successors of BB for common insn sequences. When found,
2349 share code between them by moving it across the basic block
2350 boundary. Return true if any changes made. */
2352 static bool
2353 try_head_merge_bb (basic_block bb)
2355 basic_block final_dest_bb = NULL;
2356 int max_match = INT_MAX;
2357 edge e0;
2358 rtx_insn **headptr, **currptr, **nextptr;
2359 bool changed, moveall;
2360 unsigned ix;
2361 rtx_insn *e0_last_head;
2362 rtx cond;
2363 rtx_insn *move_before;
2364 unsigned nedges = EDGE_COUNT (bb->succs);
2365 rtx_insn *jump = BB_END (bb);
2366 regset live, live_union;
2368 /* Nothing to do if there is not at least two outgoing edges. */
2369 if (nedges < 2)
2370 return false;
2372 /* Don't crossjump if this block ends in a computed jump,
2373 unless we are optimizing for size. */
2374 if (optimize_bb_for_size_p (bb)
2375 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2376 && computed_jump_p (BB_END (bb)))
2377 return false;
2379 cond = get_condition (jump, &move_before, true, false);
2380 if (cond == NULL_RTX)
2382 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2383 move_before = prev_nonnote_nondebug_insn (jump);
2384 else
2385 move_before = jump;
2388 for (ix = 0; ix < nedges; ix++)
2389 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2390 return false;
2392 for (ix = 0; ix < nedges; ix++)
2394 edge e = EDGE_SUCC (bb, ix);
2395 basic_block other_bb = e->dest;
2397 if (df_get_bb_dirty (other_bb))
2399 block_was_dirty = true;
2400 return false;
2403 if (e->flags & EDGE_ABNORMAL)
2404 return false;
2406 /* Normally, all destination blocks must only be reachable from this
2407 block, i.e. they must have one incoming edge.
2409 There is one special case we can handle, that of multiple consecutive
2410 jumps where the first jumps to one of the targets of the second jump.
2411 This happens frequently in switch statements for default labels.
2412 The structure is as follows:
2413 FINAL_DEST_BB
2414 ....
2415 if (cond) jump A;
2416 fall through
2418 jump with targets A, B, C, D...
2420 has two incoming edges, from FINAL_DEST_BB and BB
2422 In this case, we can try to move the insns through BB and into
2423 FINAL_DEST_BB. */
2424 if (EDGE_COUNT (other_bb->preds) != 1)
2426 edge incoming_edge, incoming_bb_other_edge;
2427 edge_iterator ei;
2429 if (final_dest_bb != NULL
2430 || EDGE_COUNT (other_bb->preds) != 2)
2431 return false;
2433 /* We must be able to move the insns across the whole block. */
2434 move_before = BB_HEAD (bb);
2435 while (!NONDEBUG_INSN_P (move_before))
2436 move_before = NEXT_INSN (move_before);
2438 if (EDGE_COUNT (bb->preds) != 1)
2439 return false;
2440 incoming_edge = EDGE_PRED (bb, 0);
2441 final_dest_bb = incoming_edge->src;
2442 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2443 return false;
2444 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2445 if (incoming_bb_other_edge != incoming_edge)
2446 break;
2447 if (incoming_bb_other_edge->dest != other_bb)
2448 return false;
2452 e0 = EDGE_SUCC (bb, 0);
2453 e0_last_head = NULL;
2454 changed = false;
2456 for (ix = 1; ix < nedges; ix++)
2458 edge e = EDGE_SUCC (bb, ix);
2459 rtx_insn *e0_last, *e_last;
2460 int nmatch;
2462 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2463 &e0_last, &e_last, 0);
2464 if (nmatch == 0)
2465 return false;
2467 if (nmatch < max_match)
2469 max_match = nmatch;
2470 e0_last_head = e0_last;
2474 /* If we matched an entire block, we probably have to avoid moving the
2475 last insn. */
2476 if (max_match > 0
2477 && e0_last_head == BB_END (e0->dest)
2478 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2479 || control_flow_insn_p (e0_last_head)))
2481 max_match--;
2482 if (max_match == 0)
2483 return false;
2485 e0_last_head = prev_real_insn (e0_last_head);
2486 while (DEBUG_INSN_P (e0_last_head));
2489 if (max_match == 0)
2490 return false;
2492 /* We must find a union of the live registers at each of the end points. */
2493 live = BITMAP_ALLOC (NULL);
2494 live_union = BITMAP_ALLOC (NULL);
2496 currptr = XNEWVEC (rtx_insn *, nedges);
2497 headptr = XNEWVEC (rtx_insn *, nedges);
2498 nextptr = XNEWVEC (rtx_insn *, nedges);
2500 for (ix = 0; ix < nedges; ix++)
2502 int j;
2503 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2504 rtx_insn *head = BB_HEAD (merge_bb);
2506 while (!NONDEBUG_INSN_P (head))
2507 head = NEXT_INSN (head);
2508 headptr[ix] = head;
2509 currptr[ix] = head;
2511 /* Compute the end point and live information */
2512 for (j = 1; j < max_match; j++)
2514 head = NEXT_INSN (head);
2515 while (!NONDEBUG_INSN_P (head));
2516 simulate_backwards_to_point (merge_bb, live, head);
2517 IOR_REG_SET (live_union, live);
2520 /* If we're moving across two blocks, verify the validity of the
2521 first move, then adjust the target and let the loop below deal
2522 with the final move. */
2523 if (final_dest_bb != NULL)
2525 rtx_insn *move_upto;
2527 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2528 jump, e0->dest, live_union,
2529 NULL, &move_upto);
2530 if (!moveall)
2532 if (move_upto == NULL_RTX)
2533 goto out;
2535 while (e0_last_head != move_upto)
2537 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2538 live_union);
2539 e0_last_head = PREV_INSN (e0_last_head);
2542 if (e0_last_head == NULL_RTX)
2543 goto out;
2545 jump = BB_END (final_dest_bb);
2546 cond = get_condition (jump, &move_before, true, false);
2547 if (cond == NULL_RTX)
2549 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2550 move_before = prev_nonnote_nondebug_insn (jump);
2551 else
2552 move_before = jump;
2558 rtx_insn *move_upto;
2559 moveall = can_move_insns_across (currptr[0], e0_last_head,
2560 move_before, jump, e0->dest, live_union,
2561 NULL, &move_upto);
2562 if (!moveall && move_upto == NULL_RTX)
2564 if (jump == move_before)
2565 break;
2567 /* Try again, using a different insertion point. */
2568 move_before = jump;
2570 /* Don't try moving before a cc0 user, as that may invalidate
2571 the cc0. */
2572 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2573 break;
2575 continue;
2578 if (final_dest_bb && !moveall)
2579 /* We haven't checked whether a partial move would be OK for the first
2580 move, so we have to fail this case. */
2581 break;
2583 changed = true;
2584 for (;;)
2586 if (currptr[0] == move_upto)
2587 break;
2588 for (ix = 0; ix < nedges; ix++)
2590 rtx_insn *curr = currptr[ix];
2592 curr = NEXT_INSN (curr);
2593 while (!NONDEBUG_INSN_P (curr));
2594 currptr[ix] = curr;
2598 /* If we can't currently move all of the identical insns, remember
2599 each insn after the range that we'll merge. */
2600 if (!moveall)
2601 for (ix = 0; ix < nedges; ix++)
2603 rtx_insn *curr = currptr[ix];
2605 curr = NEXT_INSN (curr);
2606 while (!NONDEBUG_INSN_P (curr));
2607 nextptr[ix] = curr;
2610 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2611 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2612 if (final_dest_bb != NULL)
2613 df_set_bb_dirty (final_dest_bb);
2614 df_set_bb_dirty (bb);
2615 for (ix = 1; ix < nedges; ix++)
2617 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2618 delete_insn_chain (headptr[ix], currptr[ix], false);
2620 if (!moveall)
2622 if (jump == move_before)
2623 break;
2625 /* For the unmerged insns, try a different insertion point. */
2626 move_before = jump;
2628 /* Don't try moving before a cc0 user, as that may invalidate
2629 the cc0. */
2630 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2631 break;
2633 for (ix = 0; ix < nedges; ix++)
2634 currptr[ix] = headptr[ix] = nextptr[ix];
2637 while (!moveall);
2639 out:
2640 free (currptr);
2641 free (headptr);
2642 free (nextptr);
2644 crossjumps_occurred |= changed;
2646 return changed;
2649 /* Return true if BB contains just bb note, or bb note followed
2650 by only DEBUG_INSNs. */
2652 static bool
2653 trivially_empty_bb_p (basic_block bb)
2655 rtx_insn *insn = BB_END (bb);
2657 while (1)
2659 if (insn == BB_HEAD (bb))
2660 return true;
2661 if (!DEBUG_INSN_P (insn))
2662 return false;
2663 insn = PREV_INSN (insn);
2667 /* Return true if BB contains just a return and possibly a USE of the
2668 return value. Fill in *RET and *USE with the return and use insns
2669 if any found, otherwise NULL. */
2671 static bool
2672 bb_is_just_return (basic_block bb, rtx_insn **ret, rtx_insn **use)
2674 *ret = *use = NULL;
2675 rtx_insn *insn;
2677 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
2678 return false;
2680 FOR_BB_INSNS (bb, insn)
2681 if (NONDEBUG_INSN_P (insn))
2683 if (!*ret && ANY_RETURN_P (PATTERN (insn)))
2684 *ret = insn;
2685 else if (!*ret && !*use && GET_CODE (PATTERN (insn)) == USE
2686 && REG_P (XEXP (PATTERN (insn), 0))
2687 && REG_FUNCTION_VALUE_P (XEXP (PATTERN (insn), 0)))
2688 *use = insn;
2689 else
2690 return false;
2693 return !!*ret;
2696 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2697 instructions etc. Return nonzero if changes were made. */
2699 static bool
2700 try_optimize_cfg (int mode)
2702 bool changed_overall = false;
2703 bool changed;
2704 int iterations = 0;
2705 basic_block bb, b, next;
2707 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2708 clear_bb_flags ();
2710 crossjumps_occurred = false;
2712 FOR_EACH_BB_FN (bb, cfun)
2713 update_forwarder_flag (bb);
2715 if (! targetm.cannot_modify_jumps_p ())
2717 first_pass = true;
2718 /* Attempt to merge blocks as made possible by edge removal. If
2719 a block has only one successor, and the successor has only
2720 one predecessor, they may be combined. */
2723 block_was_dirty = false;
2724 changed = false;
2725 iterations++;
2727 if (dump_file)
2728 fprintf (dump_file,
2729 "\n\ntry_optimize_cfg iteration %i\n\n",
2730 iterations);
2732 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2733 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2735 basic_block c;
2736 edge s;
2737 bool changed_here = false;
2739 /* Delete trivially dead basic blocks. This is either
2740 blocks with no predecessors, or empty blocks with no
2741 successors. However if the empty block with no
2742 successors is the successor of the ENTRY_BLOCK, it is
2743 kept. This ensures that the ENTRY_BLOCK will have a
2744 successor which is a precondition for many RTL
2745 passes. Empty blocks may result from expanding
2746 __builtin_unreachable (). */
2747 if (EDGE_COUNT (b->preds) == 0
2748 || (EDGE_COUNT (b->succs) == 0
2749 && trivially_empty_bb_p (b)
2750 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2751 != b))
2753 c = b->prev_bb;
2754 if (EDGE_COUNT (b->preds) > 0)
2756 edge e;
2757 edge_iterator ei;
2759 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2761 if (BB_FOOTER (b)
2762 && BARRIER_P (BB_FOOTER (b)))
2763 FOR_EACH_EDGE (e, ei, b->preds)
2764 if ((e->flags & EDGE_FALLTHRU)
2765 && BB_FOOTER (e->src) == NULL)
2767 if (BB_FOOTER (b))
2769 BB_FOOTER (e->src) = BB_FOOTER (b);
2770 BB_FOOTER (b) = NULL;
2772 else
2774 start_sequence ();
2775 BB_FOOTER (e->src) = emit_barrier ();
2776 end_sequence ();
2780 else
2782 rtx_insn *last = get_last_bb_insn (b);
2783 if (last && BARRIER_P (last))
2784 FOR_EACH_EDGE (e, ei, b->preds)
2785 if ((e->flags & EDGE_FALLTHRU))
2786 emit_barrier_after (BB_END (e->src));
2789 delete_basic_block (b);
2790 changed = true;
2791 /* Avoid trying to remove the exit block. */
2792 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2793 continue;
2796 /* Remove code labels no longer used. */
2797 if (single_pred_p (b)
2798 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2799 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2800 && LABEL_P (BB_HEAD (b))
2801 && !LABEL_PRESERVE_P (BB_HEAD (b))
2802 /* If the previous block ends with a branch to this
2803 block, we can't delete the label. Normally this
2804 is a condjump that is yet to be simplified, but
2805 if CASE_DROPS_THRU, this can be a tablejump with
2806 some element going to the same place as the
2807 default (fallthru). */
2808 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2809 || !JUMP_P (BB_END (single_pred (b)))
2810 || ! label_is_jump_target_p (BB_HEAD (b),
2811 BB_END (single_pred (b)))))
2813 delete_insn (BB_HEAD (b));
2814 if (dump_file)
2815 fprintf (dump_file, "Deleted label in block %i.\n",
2816 b->index);
2819 /* If we fall through an empty block, we can remove it. */
2820 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2821 && single_pred_p (b)
2822 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2823 && !LABEL_P (BB_HEAD (b))
2824 && FORWARDER_BLOCK_P (b)
2825 /* Note that forwarder_block_p true ensures that
2826 there is a successor for this block. */
2827 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2828 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2830 if (dump_file)
2831 fprintf (dump_file,
2832 "Deleting fallthru block %i.\n",
2833 b->index);
2835 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2836 ? b->next_bb : b->prev_bb);
2837 redirect_edge_succ_nodup (single_pred_edge (b),
2838 single_succ (b));
2839 delete_basic_block (b);
2840 changed = true;
2841 b = c;
2842 continue;
2845 /* Merge B with its single successor, if any. */
2846 if (single_succ_p (b)
2847 && (s = single_succ_edge (b))
2848 && !(s->flags & EDGE_COMPLEX)
2849 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2850 && single_pred_p (c)
2851 && b != c)
2853 /* When not in cfg_layout mode use code aware of reordering
2854 INSN. This code possibly creates new basic blocks so it
2855 does not fit merge_blocks interface and is kept here in
2856 hope that it will become useless once more of compiler
2857 is transformed to use cfg_layout mode. */
2859 if ((mode & CLEANUP_CFGLAYOUT)
2860 && can_merge_blocks_p (b, c))
2862 merge_blocks (b, c);
2863 update_forwarder_flag (b);
2864 changed_here = true;
2866 else if (!(mode & CLEANUP_CFGLAYOUT)
2867 /* If the jump insn has side effects,
2868 we can't kill the edge. */
2869 && (!JUMP_P (BB_END (b))
2870 || (reload_completed
2871 ? simplejump_p (BB_END (b))
2872 : (onlyjump_p (BB_END (b))
2873 && !tablejump_p (BB_END (b),
2874 NULL, NULL))))
2875 && (next = merge_blocks_move (s, b, c, mode)))
2877 b = next;
2878 changed_here = true;
2882 /* Try to change a branch to a return to just that return. */
2883 rtx_insn *ret, *use;
2884 if (single_succ_p (b)
2885 && onlyjump_p (BB_END (b))
2886 && bb_is_just_return (single_succ (b), &ret, &use))
2888 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2889 PATTERN (ret), 0))
2891 if (use)
2892 emit_insn_before (copy_insn (PATTERN (use)),
2893 BB_END (b));
2894 if (dump_file)
2895 fprintf (dump_file, "Changed jump %d->%d to return.\n",
2896 b->index, single_succ (b)->index);
2897 redirect_edge_succ (single_succ_edge (b),
2898 EXIT_BLOCK_PTR_FOR_FN (cfun));
2899 single_succ_edge (b)->flags &= ~EDGE_CROSSING;
2900 changed_here = true;
2904 /* Try to change a conditional branch to a return to the
2905 respective conditional return. */
2906 if (EDGE_COUNT (b->succs) == 2
2907 && any_condjump_p (BB_END (b))
2908 && bb_is_just_return (BRANCH_EDGE (b)->dest, &ret, &use))
2910 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2911 PATTERN (ret), 0))
2913 if (use)
2914 emit_insn_before (copy_insn (PATTERN (use)),
2915 BB_END (b));
2916 if (dump_file)
2917 fprintf (dump_file, "Changed conditional jump %d->%d "
2918 "to conditional return.\n",
2919 b->index, BRANCH_EDGE (b)->dest->index);
2920 redirect_edge_succ (BRANCH_EDGE (b),
2921 EXIT_BLOCK_PTR_FOR_FN (cfun));
2922 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2923 changed_here = true;
2927 /* Try to flip a conditional branch that falls through to
2928 a return so that it becomes a conditional return and a
2929 new jump to the original branch target. */
2930 if (EDGE_COUNT (b->succs) == 2
2931 && BRANCH_EDGE (b)->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
2932 && any_condjump_p (BB_END (b))
2933 && bb_is_just_return (FALLTHRU_EDGE (b)->dest, &ret, &use))
2935 if (invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2936 JUMP_LABEL (BB_END (b)), 0))
2938 basic_block new_ft = BRANCH_EDGE (b)->dest;
2939 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2940 PATTERN (ret), 0))
2942 if (use)
2943 emit_insn_before (copy_insn (PATTERN (use)),
2944 BB_END (b));
2945 if (dump_file)
2946 fprintf (dump_file, "Changed conditional jump "
2947 "%d->%d to conditional return, adding "
2948 "fall-through jump.\n",
2949 b->index, BRANCH_EDGE (b)->dest->index);
2950 redirect_edge_succ (BRANCH_EDGE (b),
2951 EXIT_BLOCK_PTR_FOR_FN (cfun));
2952 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2953 std::swap (BRANCH_EDGE (b)->probability,
2954 FALLTHRU_EDGE (b)->probability);
2955 update_br_prob_note (b);
2956 basic_block jb = force_nonfallthru (FALLTHRU_EDGE (b));
2957 notice_new_block (jb);
2958 if (!redirect_jump (as_a <rtx_jump_insn *> (BB_END (jb)),
2959 block_label (new_ft), 0))
2960 gcc_unreachable ();
2961 redirect_edge_succ (single_succ_edge (jb), new_ft);
2962 changed_here = true;
2964 else
2966 /* Invert the jump back to what it was. This should
2967 never fail. */
2968 if (!invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2969 JUMP_LABEL (BB_END (b)), 0))
2970 gcc_unreachable ();
2975 /* Simplify branch over branch. */
2976 if ((mode & CLEANUP_EXPENSIVE)
2977 && !(mode & CLEANUP_CFGLAYOUT)
2978 && try_simplify_condjump (b))
2979 changed_here = true;
2981 /* If B has a single outgoing edge, but uses a
2982 non-trivial jump instruction without side-effects, we
2983 can either delete the jump entirely, or replace it
2984 with a simple unconditional jump. */
2985 if (single_succ_p (b)
2986 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2987 && onlyjump_p (BB_END (b))
2988 && !CROSSING_JUMP_P (BB_END (b))
2989 && try_redirect_by_replacing_jump (single_succ_edge (b),
2990 single_succ (b),
2991 (mode & CLEANUP_CFGLAYOUT) != 0))
2993 update_forwarder_flag (b);
2994 changed_here = true;
2997 /* Simplify branch to branch. */
2998 if (try_forward_edges (mode, b))
3000 update_forwarder_flag (b);
3001 changed_here = true;
3004 /* Look for shared code between blocks. */
3005 if ((mode & CLEANUP_CROSSJUMP)
3006 && try_crossjump_bb (mode, b))
3007 changed_here = true;
3009 if ((mode & CLEANUP_CROSSJUMP)
3010 /* This can lengthen register lifetimes. Do it only after
3011 reload. */
3012 && reload_completed
3013 && try_head_merge_bb (b))
3014 changed_here = true;
3016 /* Don't get confused by the index shift caused by
3017 deleting blocks. */
3018 if (!changed_here)
3019 b = b->next_bb;
3020 else
3021 changed = true;
3024 if ((mode & CLEANUP_CROSSJUMP)
3025 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
3026 changed = true;
3028 if (block_was_dirty)
3030 /* This should only be set by head-merging. */
3031 gcc_assert (mode & CLEANUP_CROSSJUMP);
3032 df_analyze ();
3035 if (changed)
3037 /* Edge forwarding in particular can cause hot blocks previously
3038 reached by both hot and cold blocks to become dominated only
3039 by cold blocks. This will cause the verification below to fail,
3040 and lead to now cold code in the hot section. This is not easy
3041 to detect and fix during edge forwarding, and in some cases
3042 is only visible after newly unreachable blocks are deleted,
3043 which will be done in fixup_partitions. */
3044 fixup_partitions ();
3045 checking_verify_flow_info ();
3048 changed_overall |= changed;
3049 first_pass = false;
3051 while (changed);
3054 FOR_ALL_BB_FN (b, cfun)
3055 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
3057 return changed_overall;
3060 /* Delete all unreachable basic blocks. */
3062 bool
3063 delete_unreachable_blocks (void)
3065 bool changed = false;
3066 basic_block b, prev_bb;
3068 find_unreachable_blocks ();
3070 /* When we're in GIMPLE mode and there may be debug insns, we should
3071 delete blocks in reverse dominator order, so as to get a chance
3072 to substitute all released DEFs into debug stmts. If we don't
3073 have dominators information, walking blocks backward gets us a
3074 better chance of retaining most debug information than
3075 otherwise. */
3076 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
3077 && dom_info_available_p (CDI_DOMINATORS))
3079 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3080 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3082 prev_bb = b->prev_bb;
3084 if (!(b->flags & BB_REACHABLE))
3086 /* Speed up the removal of blocks that don't dominate
3087 others. Walking backwards, this should be the common
3088 case. */
3089 if (!first_dom_son (CDI_DOMINATORS, b))
3090 delete_basic_block (b);
3091 else
3093 vec<basic_block> h
3094 = get_all_dominated_blocks (CDI_DOMINATORS, b);
3096 while (h.length ())
3098 b = h.pop ();
3100 prev_bb = b->prev_bb;
3102 gcc_assert (!(b->flags & BB_REACHABLE));
3104 delete_basic_block (b);
3107 h.release ();
3110 changed = true;
3114 else
3116 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3117 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3119 prev_bb = b->prev_bb;
3121 if (!(b->flags & BB_REACHABLE))
3123 delete_basic_block (b);
3124 changed = true;
3129 if (changed)
3130 tidy_fallthru_edges ();
3131 return changed;
3134 /* Delete any jump tables never referenced. We can't delete them at the
3135 time of removing tablejump insn as they are referenced by the preceding
3136 insns computing the destination, so we delay deleting and garbagecollect
3137 them once life information is computed. */
3138 void
3139 delete_dead_jumptables (void)
3141 basic_block bb;
3143 /* A dead jump table does not belong to any basic block. Scan insns
3144 between two adjacent basic blocks. */
3145 FOR_EACH_BB_FN (bb, cfun)
3147 rtx_insn *insn, *next;
3149 for (insn = NEXT_INSN (BB_END (bb));
3150 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
3151 insn = next)
3153 next = NEXT_INSN (insn);
3154 if (LABEL_P (insn)
3155 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
3156 && JUMP_TABLE_DATA_P (next))
3158 rtx_insn *label = insn, *jump = next;
3160 if (dump_file)
3161 fprintf (dump_file, "Dead jumptable %i removed\n",
3162 INSN_UID (insn));
3164 next = NEXT_INSN (next);
3165 delete_insn (jump);
3166 delete_insn (label);
3173 /* Tidy the CFG by deleting unreachable code and whatnot. */
3175 bool
3176 cleanup_cfg (int mode)
3178 bool changed = false;
3180 /* Set the cfglayout mode flag here. We could update all the callers
3181 but that is just inconvenient, especially given that we eventually
3182 want to have cfglayout mode as the default. */
3183 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3184 mode |= CLEANUP_CFGLAYOUT;
3186 timevar_push (TV_CLEANUP_CFG);
3187 if (delete_unreachable_blocks ())
3189 changed = true;
3190 /* We've possibly created trivially dead code. Cleanup it right
3191 now to introduce more opportunities for try_optimize_cfg. */
3192 if (!(mode & (CLEANUP_NO_INSN_DEL))
3193 && !reload_completed)
3194 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3197 compact_blocks ();
3199 /* To tail-merge blocks ending in the same noreturn function (e.g.
3200 a call to abort) we have to insert fake edges to exit. Do this
3201 here once. The fake edges do not interfere with any other CFG
3202 cleanups. */
3203 if (mode & CLEANUP_CROSSJUMP)
3204 add_noreturn_fake_exit_edges ();
3206 if (!dbg_cnt (cfg_cleanup))
3207 return changed;
3209 while (try_optimize_cfg (mode))
3211 delete_unreachable_blocks (), changed = true;
3212 if (!(mode & CLEANUP_NO_INSN_DEL))
3214 /* Try to remove some trivially dead insns when doing an expensive
3215 cleanup. But delete_trivially_dead_insns doesn't work after
3216 reload (it only handles pseudos) and run_fast_dce is too costly
3217 to run in every iteration.
3219 For effective cross jumping, we really want to run a fast DCE to
3220 clean up any dead conditions, or they get in the way of performing
3221 useful tail merges.
3223 Other transformations in cleanup_cfg are not so sensitive to dead
3224 code, so delete_trivially_dead_insns or even doing nothing at all
3225 is good enough. */
3226 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3227 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3228 break;
3229 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occurred)
3230 run_fast_dce ();
3232 else
3233 break;
3236 if (mode & CLEANUP_CROSSJUMP)
3237 remove_fake_exit_edges ();
3239 /* Don't call delete_dead_jumptables in cfglayout mode, because
3240 that function assumes that jump tables are in the insns stream.
3241 But we also don't _have_ to delete dead jumptables in cfglayout
3242 mode because we shouldn't even be looking at things that are
3243 not in a basic block. Dead jumptables are cleaned up when
3244 going out of cfglayout mode. */
3245 if (!(mode & CLEANUP_CFGLAYOUT))
3246 delete_dead_jumptables ();
3248 /* ??? We probably do this way too often. */
3249 if (current_loops
3250 && (changed
3251 || (mode & CLEANUP_CFG_CHANGED)))
3253 timevar_push (TV_REPAIR_LOOPS);
3254 /* The above doesn't preserve dominance info if available. */
3255 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3256 calculate_dominance_info (CDI_DOMINATORS);
3257 fix_loop_structure (NULL);
3258 free_dominance_info (CDI_DOMINATORS);
3259 timevar_pop (TV_REPAIR_LOOPS);
3262 timevar_pop (TV_CLEANUP_CFG);
3264 return changed;
3267 namespace {
3269 const pass_data pass_data_jump =
3271 RTL_PASS, /* type */
3272 "jump", /* name */
3273 OPTGROUP_NONE, /* optinfo_flags */
3274 TV_JUMP, /* tv_id */
3275 0, /* properties_required */
3276 0, /* properties_provided */
3277 0, /* properties_destroyed */
3278 0, /* todo_flags_start */
3279 0, /* todo_flags_finish */
3282 class pass_jump : public rtl_opt_pass
3284 public:
3285 pass_jump (gcc::context *ctxt)
3286 : rtl_opt_pass (pass_data_jump, ctxt)
3289 /* opt_pass methods: */
3290 virtual unsigned int execute (function *);
3292 }; // class pass_jump
3294 unsigned int
3295 pass_jump::execute (function *)
3297 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3298 if (dump_file)
3299 dump_flow_info (dump_file, dump_flags);
3300 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3301 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3302 return 0;
3305 } // anon namespace
3307 rtl_opt_pass *
3308 make_pass_jump (gcc::context *ctxt)
3310 return new pass_jump (ctxt);
3313 namespace {
3315 const pass_data pass_data_jump2 =
3317 RTL_PASS, /* type */
3318 "jump2", /* name */
3319 OPTGROUP_NONE, /* optinfo_flags */
3320 TV_JUMP, /* tv_id */
3321 0, /* properties_required */
3322 0, /* properties_provided */
3323 0, /* properties_destroyed */
3324 0, /* todo_flags_start */
3325 0, /* todo_flags_finish */
3328 class pass_jump2 : public rtl_opt_pass
3330 public:
3331 pass_jump2 (gcc::context *ctxt)
3332 : rtl_opt_pass (pass_data_jump2, ctxt)
3335 /* opt_pass methods: */
3336 virtual unsigned int execute (function *)
3338 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3339 return 0;
3342 }; // class pass_jump2
3344 } // anon namespace
3346 rtl_opt_pass *
3347 make_pass_jump2 (gcc::context *ctxt)
3349 return new pass_jump2 (ctxt);