Add support for ARMv8-R architecture
[official-gcc.git] / gcc / cfgcleanup.c
blob5d00fafffa307b7296f5dc1007cc25b0c76672ce
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2017 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 /* This file contains optimizer of the control flow. The main entry point is
21 cleanup_cfg. Following optimizations are performed:
23 - Unreachable blocks removal
24 - Edge forwarding (edge to the forwarder block is forwarded to its
25 successor. Simplification of the branch instruction is performed by
26 underlying infrastructure so branch can be converted to simplejump or
27 eliminated).
28 - Cross jumping (tail merging)
29 - Conditional jump-around-simplejump simplification
30 - Basic block merging. */
32 #include "config.h"
33 #include "system.h"
34 #include "coretypes.h"
35 #include "backend.h"
36 #include "target.h"
37 #include "rtl.h"
38 #include "tree.h"
39 #include "cfghooks.h"
40 #include "df.h"
41 #include "memmodel.h"
42 #include "tm_p.h"
43 #include "insn-config.h"
44 #include "emit-rtl.h"
45 #include "cselib.h"
46 #include "params.h"
47 #include "tree-pass.h"
48 #include "cfgloop.h"
49 #include "cfgrtl.h"
50 #include "cfganal.h"
51 #include "cfgbuild.h"
52 #include "cfgcleanup.h"
53 #include "dce.h"
54 #include "dbgcnt.h"
55 #include "rtl-iter.h"
57 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
59 /* Set to true when we are running first pass of try_optimize_cfg loop. */
60 static bool first_pass;
62 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
63 static bool crossjumps_occurred;
65 /* Set to true if we couldn't run an optimization due to stale liveness
66 information; we should run df_analyze to enable more opportunities. */
67 static bool block_was_dirty;
69 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
70 static bool try_crossjump_bb (int, basic_block);
71 static bool outgoing_edges_match (int, basic_block, basic_block);
72 static enum replace_direction old_insns_match_p (int, rtx_insn *, rtx_insn *);
74 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
75 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
76 static bool try_optimize_cfg (int);
77 static bool try_simplify_condjump (basic_block);
78 static bool try_forward_edges (int, basic_block);
79 static edge thread_jump (edge, basic_block);
80 static bool mark_effect (rtx, bitmap);
81 static void notice_new_block (basic_block);
82 static void update_forwarder_flag (basic_block);
83 static void merge_memattrs (rtx, rtx);
85 /* Set flags for newly created block. */
87 static void
88 notice_new_block (basic_block bb)
90 if (!bb)
91 return;
93 if (forwarder_block_p (bb))
94 bb->flags |= BB_FORWARDER_BLOCK;
97 /* Recompute forwarder flag after block has been modified. */
99 static void
100 update_forwarder_flag (basic_block bb)
102 if (forwarder_block_p (bb))
103 bb->flags |= BB_FORWARDER_BLOCK;
104 else
105 bb->flags &= ~BB_FORWARDER_BLOCK;
108 /* Simplify a conditional jump around an unconditional jump.
109 Return true if something changed. */
111 static bool
112 try_simplify_condjump (basic_block cbranch_block)
114 basic_block jump_block, jump_dest_block, cbranch_dest_block;
115 edge cbranch_jump_edge, cbranch_fallthru_edge;
116 rtx_insn *cbranch_insn;
118 /* Verify that there are exactly two successors. */
119 if (EDGE_COUNT (cbranch_block->succs) != 2)
120 return false;
122 /* Verify that we've got a normal conditional branch at the end
123 of the block. */
124 cbranch_insn = BB_END (cbranch_block);
125 if (!any_condjump_p (cbranch_insn))
126 return false;
128 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
129 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
131 /* The next block must not have multiple predecessors, must not
132 be the last block in the function, and must contain just the
133 unconditional jump. */
134 jump_block = cbranch_fallthru_edge->dest;
135 if (!single_pred_p (jump_block)
136 || jump_block->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
137 || !FORWARDER_BLOCK_P (jump_block))
138 return false;
139 jump_dest_block = single_succ (jump_block);
141 /* If we are partitioning hot/cold basic blocks, we don't want to
142 mess up unconditional or indirect jumps that cross between hot
143 and cold sections.
145 Basic block partitioning may result in some jumps that appear to
146 be optimizable (or blocks that appear to be mergeable), but which really
147 must be left untouched (they are required to make it safely across
148 partition boundaries). See the comments at the top of
149 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
151 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
152 || (cbranch_jump_edge->flags & EDGE_CROSSING))
153 return false;
155 /* The conditional branch must target the block after the
156 unconditional branch. */
157 cbranch_dest_block = cbranch_jump_edge->dest;
159 if (cbranch_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
160 || jump_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
161 || !can_fallthru (jump_block, cbranch_dest_block))
162 return false;
164 /* Invert the conditional branch. */
165 if (!invert_jump (as_a <rtx_jump_insn *> (cbranch_insn),
166 block_label (jump_dest_block), 0))
167 return false;
169 if (dump_file)
170 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
171 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
173 /* Success. Update the CFG to match. Note that after this point
174 the edge variable names appear backwards; the redirection is done
175 this way to preserve edge profile data. */
176 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
177 cbranch_dest_block);
178 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
179 jump_dest_block);
180 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
181 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
182 update_br_prob_note (cbranch_block);
184 /* Delete the block with the unconditional jump, and clean up the mess. */
185 delete_basic_block (jump_block);
186 tidy_fallthru_edge (cbranch_jump_edge);
187 update_forwarder_flag (cbranch_block);
189 return true;
192 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
193 on register. Used by jump threading. */
195 static bool
196 mark_effect (rtx exp, regset nonequal)
198 rtx dest;
199 switch (GET_CODE (exp))
201 /* In case we do clobber the register, mark it as equal, as we know the
202 value is dead so it don't have to match. */
203 case CLOBBER:
204 dest = XEXP (exp, 0);
205 if (REG_P (dest))
206 bitmap_clear_range (nonequal, REGNO (dest), REG_NREGS (dest));
207 return false;
209 case SET:
210 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
211 return false;
212 dest = SET_DEST (exp);
213 if (dest == pc_rtx)
214 return false;
215 if (!REG_P (dest))
216 return true;
217 bitmap_set_range (nonequal, REGNO (dest), REG_NREGS (dest));
218 return false;
220 default:
221 return false;
225 /* Return true if X contains a register in NONEQUAL. */
226 static bool
227 mentions_nonequal_regs (const_rtx x, regset nonequal)
229 subrtx_iterator::array_type array;
230 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
232 const_rtx x = *iter;
233 if (REG_P (x))
235 unsigned int end_regno = END_REGNO (x);
236 for (unsigned int regno = REGNO (x); regno < end_regno; ++regno)
237 if (REGNO_REG_SET_P (nonequal, regno))
238 return true;
241 return false;
244 /* Attempt to prove that the basic block B will have no side effects and
245 always continues in the same edge if reached via E. Return the edge
246 if exist, NULL otherwise. */
248 static edge
249 thread_jump (edge e, basic_block b)
251 rtx set1, set2, cond1, cond2;
252 rtx_insn *insn;
253 enum rtx_code code1, code2, reversed_code2;
254 bool reverse1 = false;
255 unsigned i;
256 regset nonequal;
257 bool failed = false;
258 reg_set_iterator rsi;
260 if (b->flags & BB_NONTHREADABLE_BLOCK)
261 return NULL;
263 /* At the moment, we do handle only conditional jumps, but later we may
264 want to extend this code to tablejumps and others. */
265 if (EDGE_COUNT (e->src->succs) != 2)
266 return NULL;
267 if (EDGE_COUNT (b->succs) != 2)
269 b->flags |= BB_NONTHREADABLE_BLOCK;
270 return NULL;
273 /* Second branch must end with onlyjump, as we will eliminate the jump. */
274 if (!any_condjump_p (BB_END (e->src)))
275 return NULL;
277 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
279 b->flags |= BB_NONTHREADABLE_BLOCK;
280 return NULL;
283 set1 = pc_set (BB_END (e->src));
284 set2 = pc_set (BB_END (b));
285 if (((e->flags & EDGE_FALLTHRU) != 0)
286 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
287 reverse1 = true;
289 cond1 = XEXP (SET_SRC (set1), 0);
290 cond2 = XEXP (SET_SRC (set2), 0);
291 if (reverse1)
292 code1 = reversed_comparison_code (cond1, BB_END (e->src));
293 else
294 code1 = GET_CODE (cond1);
296 code2 = GET_CODE (cond2);
297 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
299 if (!comparison_dominates_p (code1, code2)
300 && !comparison_dominates_p (code1, reversed_code2))
301 return NULL;
303 /* Ensure that the comparison operators are equivalent.
304 ??? This is far too pessimistic. We should allow swapped operands,
305 different CCmodes, or for example comparisons for interval, that
306 dominate even when operands are not equivalent. */
307 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
308 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
309 return NULL;
311 /* Short circuit cases where block B contains some side effects, as we can't
312 safely bypass it. */
313 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
314 insn = NEXT_INSN (insn))
315 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
317 b->flags |= BB_NONTHREADABLE_BLOCK;
318 return NULL;
321 cselib_init (0);
323 /* First process all values computed in the source basic block. */
324 for (insn = NEXT_INSN (BB_HEAD (e->src));
325 insn != NEXT_INSN (BB_END (e->src));
326 insn = NEXT_INSN (insn))
327 if (INSN_P (insn))
328 cselib_process_insn (insn);
330 nonequal = BITMAP_ALLOC (NULL);
331 CLEAR_REG_SET (nonequal);
333 /* Now assume that we've continued by the edge E to B and continue
334 processing as if it were same basic block.
335 Our goal is to prove that whole block is an NOOP. */
337 for (insn = NEXT_INSN (BB_HEAD (b));
338 insn != NEXT_INSN (BB_END (b)) && !failed;
339 insn = NEXT_INSN (insn))
341 if (INSN_P (insn))
343 rtx pat = PATTERN (insn);
345 if (GET_CODE (pat) == PARALLEL)
347 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
348 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
350 else
351 failed |= mark_effect (pat, nonequal);
354 cselib_process_insn (insn);
357 /* Later we should clear nonequal of dead registers. So far we don't
358 have life information in cfg_cleanup. */
359 if (failed)
361 b->flags |= BB_NONTHREADABLE_BLOCK;
362 goto failed_exit;
365 /* cond2 must not mention any register that is not equal to the
366 former block. */
367 if (mentions_nonequal_regs (cond2, nonequal))
368 goto failed_exit;
370 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
371 goto failed_exit;
373 BITMAP_FREE (nonequal);
374 cselib_finish ();
375 if ((comparison_dominates_p (code1, code2) != 0)
376 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
377 return BRANCH_EDGE (b);
378 else
379 return FALLTHRU_EDGE (b);
381 failed_exit:
382 BITMAP_FREE (nonequal);
383 cselib_finish ();
384 return NULL;
387 /* Attempt to forward edges leaving basic block B.
388 Return true if successful. */
390 static bool
391 try_forward_edges (int mode, basic_block b)
393 bool changed = false;
394 edge_iterator ei;
395 edge e, *threaded_edges = NULL;
397 /* If we are partitioning hot/cold basic blocks, we don't want to
398 mess up unconditional or indirect jumps that cross between hot
399 and cold sections.
401 Basic block partitioning may result in some jumps that appear to
402 be optimizable (or blocks that appear to be mergeable), but which really
403 must be left untouched (they are required to make it safely across
404 partition boundaries). See the comments at the top of
405 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
407 if (JUMP_P (BB_END (b)) && CROSSING_JUMP_P (BB_END (b)))
408 return false;
410 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
412 basic_block target, first;
413 location_t goto_locus;
414 int counter;
415 bool threaded = false;
416 int nthreaded_edges = 0;
417 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
419 /* Skip complex edges because we don't know how to update them.
421 Still handle fallthru edges, as we can succeed to forward fallthru
422 edge to the same place as the branch edge of conditional branch
423 and turn conditional branch to an unconditional branch. */
424 if (e->flags & EDGE_COMPLEX)
426 ei_next (&ei);
427 continue;
430 target = first = e->dest;
431 counter = NUM_FIXED_BLOCKS;
432 goto_locus = e->goto_locus;
434 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
435 up jumps that cross between hot/cold sections.
437 Basic block partitioning may result in some jumps that appear
438 to be optimizable (or blocks that appear to be mergeable), but which
439 really must be left untouched (they are required to make it safely
440 across partition boundaries). See the comments at the top of
441 bb-reorder.c:partition_hot_cold_basic_blocks for complete
442 details. */
444 if (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
445 && JUMP_P (BB_END (first))
446 && CROSSING_JUMP_P (BB_END (first)))
447 return changed;
449 while (counter < n_basic_blocks_for_fn (cfun))
451 basic_block new_target = NULL;
452 bool new_target_threaded = false;
453 may_thread |= (target->flags & BB_MODIFIED) != 0;
455 if (FORWARDER_BLOCK_P (target)
456 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
457 && single_succ (target) != EXIT_BLOCK_PTR_FOR_FN (cfun))
459 /* Bypass trivial infinite loops. */
460 new_target = single_succ (target);
461 if (target == new_target)
462 counter = n_basic_blocks_for_fn (cfun);
463 else if (!optimize)
465 /* When not optimizing, ensure that edges or forwarder
466 blocks with different locus are not optimized out. */
467 location_t new_locus = single_succ_edge (target)->goto_locus;
468 location_t locus = goto_locus;
470 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
471 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
472 && new_locus != locus)
473 new_target = NULL;
474 else
476 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
477 locus = new_locus;
479 rtx_insn *last = BB_END (target);
480 if (DEBUG_INSN_P (last))
481 last = prev_nondebug_insn (last);
482 if (last && INSN_P (last))
483 new_locus = INSN_LOCATION (last);
484 else
485 new_locus = UNKNOWN_LOCATION;
487 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
488 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
489 && new_locus != locus)
490 new_target = NULL;
491 else
493 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
494 locus = new_locus;
496 goto_locus = locus;
502 /* Allow to thread only over one edge at time to simplify updating
503 of probabilities. */
504 else if ((mode & CLEANUP_THREADING) && may_thread)
506 edge t = thread_jump (e, target);
507 if (t)
509 if (!threaded_edges)
510 threaded_edges = XNEWVEC (edge,
511 n_basic_blocks_for_fn (cfun));
512 else
514 int i;
516 /* Detect an infinite loop across blocks not
517 including the start block. */
518 for (i = 0; i < nthreaded_edges; ++i)
519 if (threaded_edges[i] == t)
520 break;
521 if (i < nthreaded_edges)
523 counter = n_basic_blocks_for_fn (cfun);
524 break;
528 /* Detect an infinite loop across the start block. */
529 if (t->dest == b)
530 break;
532 gcc_assert (nthreaded_edges
533 < (n_basic_blocks_for_fn (cfun)
534 - NUM_FIXED_BLOCKS));
535 threaded_edges[nthreaded_edges++] = t;
537 new_target = t->dest;
538 new_target_threaded = true;
542 if (!new_target)
543 break;
545 counter++;
546 target = new_target;
547 threaded |= new_target_threaded;
550 if (counter >= n_basic_blocks_for_fn (cfun))
552 if (dump_file)
553 fprintf (dump_file, "Infinite loop in BB %i.\n",
554 target->index);
556 else if (target == first)
557 ; /* We didn't do anything. */
558 else
560 /* Save the values now, as the edge may get removed. */
561 profile_count edge_count = e->count;
562 profile_probability edge_probability = e->probability;
563 int edge_frequency;
564 int n = 0;
566 e->goto_locus = goto_locus;
568 /* Don't force if target is exit block. */
569 if (threaded && target != EXIT_BLOCK_PTR_FOR_FN (cfun))
571 notice_new_block (redirect_edge_and_branch_force (e, target));
572 if (dump_file)
573 fprintf (dump_file, "Conditionals threaded.\n");
575 else if (!redirect_edge_and_branch (e, target))
577 if (dump_file)
578 fprintf (dump_file,
579 "Forwarding edge %i->%i to %i failed.\n",
580 b->index, e->dest->index, target->index);
581 ei_next (&ei);
582 continue;
585 /* We successfully forwarded the edge. Now update profile
586 data: for each edge we traversed in the chain, remove
587 the original edge's execution count. */
588 edge_frequency = edge_probability.apply (b->frequency);
592 edge t;
594 if (!single_succ_p (first))
596 gcc_assert (n < nthreaded_edges);
597 t = threaded_edges [n++];
598 gcc_assert (t->src == first);
599 update_bb_profile_for_threading (first, edge_frequency,
600 edge_count, t);
601 update_br_prob_note (first);
603 else
605 first->count -= edge_count;
606 first->frequency -= edge_frequency;
607 if (first->frequency < 0)
608 first->frequency = 0;
609 /* It is possible that as the result of
610 threading we've removed edge as it is
611 threaded to the fallthru edge. Avoid
612 getting out of sync. */
613 if (n < nthreaded_edges
614 && first == threaded_edges [n]->src)
615 n++;
616 t = single_succ_edge (first);
619 t->count -= edge_count;
620 first = t->dest;
622 while (first != target);
624 changed = true;
625 continue;
627 ei_next (&ei);
630 free (threaded_edges);
631 return changed;
635 /* Blocks A and B are to be merged into a single block. A has no incoming
636 fallthru edge, so it can be moved before B without adding or modifying
637 any jumps (aside from the jump from A to B). */
639 static void
640 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
642 rtx_insn *barrier;
644 /* If we are partitioning hot/cold basic blocks, we don't want to
645 mess up unconditional or indirect jumps that cross between hot
646 and cold sections.
648 Basic block partitioning may result in some jumps that appear to
649 be optimizable (or blocks that appear to be mergeable), but which really
650 must be left untouched (they are required to make it safely across
651 partition boundaries). See the comments at the top of
652 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
654 if (BB_PARTITION (a) != BB_PARTITION (b))
655 return;
657 barrier = next_nonnote_insn (BB_END (a));
658 gcc_assert (BARRIER_P (barrier));
659 delete_insn (barrier);
661 /* Scramble the insn chain. */
662 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
663 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
664 df_set_bb_dirty (a);
666 if (dump_file)
667 fprintf (dump_file, "Moved block %d before %d and merged.\n",
668 a->index, b->index);
670 /* Swap the records for the two blocks around. */
672 unlink_block (a);
673 link_block (a, b->prev_bb);
675 /* Now blocks A and B are contiguous. Merge them. */
676 merge_blocks (a, b);
679 /* Blocks A and B are to be merged into a single block. B has no outgoing
680 fallthru edge, so it can be moved after A without adding or modifying
681 any jumps (aside from the jump from A to B). */
683 static void
684 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
686 rtx_insn *barrier, *real_b_end;
687 rtx_insn *label;
688 rtx_jump_table_data *table;
690 /* If we are partitioning hot/cold basic blocks, we don't want to
691 mess up unconditional or indirect jumps that cross between hot
692 and cold sections.
694 Basic block partitioning may result in some jumps that appear to
695 be optimizable (or blocks that appear to be mergeable), but which really
696 must be left untouched (they are required to make it safely across
697 partition boundaries). See the comments at the top of
698 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
700 if (BB_PARTITION (a) != BB_PARTITION (b))
701 return;
703 real_b_end = BB_END (b);
705 /* If there is a jump table following block B temporarily add the jump table
706 to block B so that it will also be moved to the correct location. */
707 if (tablejump_p (BB_END (b), &label, &table)
708 && prev_active_insn (label) == BB_END (b))
710 BB_END (b) = table;
713 /* There had better have been a barrier there. Delete it. */
714 barrier = NEXT_INSN (BB_END (b));
715 if (barrier && BARRIER_P (barrier))
716 delete_insn (barrier);
719 /* Scramble the insn chain. */
720 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
722 /* Restore the real end of b. */
723 BB_END (b) = real_b_end;
725 if (dump_file)
726 fprintf (dump_file, "Moved block %d after %d and merged.\n",
727 b->index, a->index);
729 /* Now blocks A and B are contiguous. Merge them. */
730 merge_blocks (a, b);
733 /* Attempt to merge basic blocks that are potentially non-adjacent.
734 Return NULL iff the attempt failed, otherwise return basic block
735 where cleanup_cfg should continue. Because the merging commonly
736 moves basic block away or introduces another optimization
737 possibility, return basic block just before B so cleanup_cfg don't
738 need to iterate.
740 It may be good idea to return basic block before C in the case
741 C has been moved after B and originally appeared earlier in the
742 insn sequence, but we have no information available about the
743 relative ordering of these two. Hopefully it is not too common. */
745 static basic_block
746 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
748 basic_block next;
750 /* If we are partitioning hot/cold basic blocks, we don't want to
751 mess up unconditional or indirect jumps that cross between hot
752 and cold sections.
754 Basic block partitioning may result in some jumps that appear to
755 be optimizable (or blocks that appear to be mergeable), but which really
756 must be left untouched (they are required to make it safely across
757 partition boundaries). See the comments at the top of
758 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
760 if (BB_PARTITION (b) != BB_PARTITION (c))
761 return NULL;
763 /* If B has a fallthru edge to C, no need to move anything. */
764 if (e->flags & EDGE_FALLTHRU)
766 int b_index = b->index, c_index = c->index;
768 /* Protect the loop latches. */
769 if (current_loops && c->loop_father->latch == c)
770 return NULL;
772 merge_blocks (b, c);
773 update_forwarder_flag (b);
775 if (dump_file)
776 fprintf (dump_file, "Merged %d and %d without moving.\n",
777 b_index, c_index);
779 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
782 /* Otherwise we will need to move code around. Do that only if expensive
783 transformations are allowed. */
784 else if (mode & CLEANUP_EXPENSIVE)
786 edge tmp_edge, b_fallthru_edge;
787 bool c_has_outgoing_fallthru;
788 bool b_has_incoming_fallthru;
790 /* Avoid overactive code motion, as the forwarder blocks should be
791 eliminated by edge redirection instead. One exception might have
792 been if B is a forwarder block and C has no fallthru edge, but
793 that should be cleaned up by bb-reorder instead. */
794 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
795 return NULL;
797 /* We must make sure to not munge nesting of lexical blocks,
798 and loop notes. This is done by squeezing out all the notes
799 and leaving them there to lie. Not ideal, but functional. */
801 tmp_edge = find_fallthru_edge (c->succs);
802 c_has_outgoing_fallthru = (tmp_edge != NULL);
804 tmp_edge = find_fallthru_edge (b->preds);
805 b_has_incoming_fallthru = (tmp_edge != NULL);
806 b_fallthru_edge = tmp_edge;
807 next = b->prev_bb;
808 if (next == c)
809 next = next->prev_bb;
811 /* Otherwise, we're going to try to move C after B. If C does
812 not have an outgoing fallthru, then it can be moved
813 immediately after B without introducing or modifying jumps. */
814 if (! c_has_outgoing_fallthru)
816 merge_blocks_move_successor_nojumps (b, c);
817 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
820 /* If B does not have an incoming fallthru, then it can be moved
821 immediately before C without introducing or modifying jumps.
822 C cannot be the first block, so we do not have to worry about
823 accessing a non-existent block. */
825 if (b_has_incoming_fallthru)
827 basic_block bb;
829 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
830 return NULL;
831 bb = force_nonfallthru (b_fallthru_edge);
832 if (bb)
833 notice_new_block (bb);
836 merge_blocks_move_predecessor_nojumps (b, c);
837 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
840 return NULL;
844 /* Removes the memory attributes of MEM expression
845 if they are not equal. */
847 static void
848 merge_memattrs (rtx x, rtx y)
850 int i;
851 int j;
852 enum rtx_code code;
853 const char *fmt;
855 if (x == y)
856 return;
857 if (x == 0 || y == 0)
858 return;
860 code = GET_CODE (x);
862 if (code != GET_CODE (y))
863 return;
865 if (GET_MODE (x) != GET_MODE (y))
866 return;
868 if (code == MEM && !mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
870 if (! MEM_ATTRS (x))
871 MEM_ATTRS (y) = 0;
872 else if (! MEM_ATTRS (y))
873 MEM_ATTRS (x) = 0;
874 else
876 HOST_WIDE_INT mem_size;
878 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
880 set_mem_alias_set (x, 0);
881 set_mem_alias_set (y, 0);
884 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
886 set_mem_expr (x, 0);
887 set_mem_expr (y, 0);
888 clear_mem_offset (x);
889 clear_mem_offset (y);
891 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
892 || (MEM_OFFSET_KNOWN_P (x)
893 && MEM_OFFSET (x) != MEM_OFFSET (y)))
895 clear_mem_offset (x);
896 clear_mem_offset (y);
899 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
901 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
902 set_mem_size (x, mem_size);
903 set_mem_size (y, mem_size);
905 else
907 clear_mem_size (x);
908 clear_mem_size (y);
911 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
912 set_mem_align (y, MEM_ALIGN (x));
915 if (code == MEM)
917 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
919 MEM_READONLY_P (x) = 0;
920 MEM_READONLY_P (y) = 0;
922 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
924 MEM_NOTRAP_P (x) = 0;
925 MEM_NOTRAP_P (y) = 0;
927 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
929 MEM_VOLATILE_P (x) = 1;
930 MEM_VOLATILE_P (y) = 1;
934 fmt = GET_RTX_FORMAT (code);
935 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
937 switch (fmt[i])
939 case 'E':
940 /* Two vectors must have the same length. */
941 if (XVECLEN (x, i) != XVECLEN (y, i))
942 return;
944 for (j = 0; j < XVECLEN (x, i); j++)
945 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
947 break;
949 case 'e':
950 merge_memattrs (XEXP (x, i), XEXP (y, i));
953 return;
957 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
958 different single sets S1 and S2. */
960 static bool
961 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
963 int i;
964 rtx e1, e2;
966 if (p1 == s1 && p2 == s2)
967 return true;
969 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
970 return false;
972 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
973 return false;
975 for (i = 0; i < XVECLEN (p1, 0); i++)
977 e1 = XVECEXP (p1, 0, i);
978 e2 = XVECEXP (p2, 0, i);
979 if (e1 == s1 && e2 == s2)
980 continue;
981 if (reload_completed
982 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
983 continue;
985 return false;
988 return true;
992 /* NOTE1 is the REG_EQUAL note, if any, attached to an insn
993 that is a single_set with a SET_SRC of SRC1. Similarly
994 for NOTE2/SRC2.
996 So effectively NOTE1/NOTE2 are an alternate form of
997 SRC1/SRC2 respectively.
999 Return nonzero if SRC1 or NOTE1 has the same constant
1000 integer value as SRC2 or NOTE2. Else return zero. */
1001 static int
1002 values_equal_p (rtx note1, rtx note2, rtx src1, rtx src2)
1004 if (note1
1005 && note2
1006 && CONST_INT_P (XEXP (note1, 0))
1007 && rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0)))
1008 return 1;
1010 if (!note1
1011 && !note2
1012 && CONST_INT_P (src1)
1013 && CONST_INT_P (src2)
1014 && rtx_equal_p (src1, src2))
1015 return 1;
1017 if (note1
1018 && CONST_INT_P (src2)
1019 && rtx_equal_p (XEXP (note1, 0), src2))
1020 return 1;
1022 if (note2
1023 && CONST_INT_P (src1)
1024 && rtx_equal_p (XEXP (note2, 0), src1))
1025 return 1;
1027 return 0;
1030 /* Examine register notes on I1 and I2 and return:
1031 - dir_forward if I1 can be replaced by I2, or
1032 - dir_backward if I2 can be replaced by I1, or
1033 - dir_both if both are the case. */
1035 static enum replace_direction
1036 can_replace_by (rtx_insn *i1, rtx_insn *i2)
1038 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1039 bool c1, c2;
1041 /* Check for 2 sets. */
1042 s1 = single_set (i1);
1043 s2 = single_set (i2);
1044 if (s1 == NULL_RTX || s2 == NULL_RTX)
1045 return dir_none;
1047 /* Check that the 2 sets set the same dest. */
1048 d1 = SET_DEST (s1);
1049 d2 = SET_DEST (s2);
1050 if (!(reload_completed
1051 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1052 return dir_none;
1054 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1055 set dest to the same value. */
1056 note1 = find_reg_equal_equiv_note (i1);
1057 note2 = find_reg_equal_equiv_note (i2);
1059 src1 = SET_SRC (s1);
1060 src2 = SET_SRC (s2);
1062 if (!values_equal_p (note1, note2, src1, src2))
1063 return dir_none;
1065 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1066 return dir_none;
1068 /* Although the 2 sets set dest to the same value, we cannot replace
1069 (set (dest) (const_int))
1071 (set (dest) (reg))
1072 because we don't know if the reg is live and has the same value at the
1073 location of replacement. */
1074 c1 = CONST_INT_P (src1);
1075 c2 = CONST_INT_P (src2);
1076 if (c1 && c2)
1077 return dir_both;
1078 else if (c2)
1079 return dir_forward;
1080 else if (c1)
1081 return dir_backward;
1083 return dir_none;
1086 /* Merges directions A and B. */
1088 static enum replace_direction
1089 merge_dir (enum replace_direction a, enum replace_direction b)
1091 /* Implements the following table:
1092 |bo fw bw no
1093 ---+-----------
1094 bo |bo fw bw no
1095 fw |-- fw no no
1096 bw |-- -- bw no
1097 no |-- -- -- no. */
1099 if (a == b)
1100 return a;
1102 if (a == dir_both)
1103 return b;
1104 if (b == dir_both)
1105 return a;
1107 return dir_none;
1110 /* Array of flags indexed by reg note kind, true if the given
1111 reg note is CFA related. */
1112 static const bool reg_note_cfa_p[] = {
1113 #undef REG_CFA_NOTE
1114 #define DEF_REG_NOTE(NAME) false,
1115 #define REG_CFA_NOTE(NAME) true,
1116 #include "reg-notes.def"
1117 #undef REG_CFA_NOTE
1118 #undef DEF_REG_NOTE
1119 false
1122 /* Return true if I1 and I2 have identical CFA notes (the same order
1123 and equivalent content). */
1125 static bool
1126 insns_have_identical_cfa_notes (rtx_insn *i1, rtx_insn *i2)
1128 rtx n1, n2;
1129 for (n1 = REG_NOTES (i1), n2 = REG_NOTES (i2); ;
1130 n1 = XEXP (n1, 1), n2 = XEXP (n2, 1))
1132 /* Skip over reg notes not related to CFI information. */
1133 while (n1 && !reg_note_cfa_p[REG_NOTE_KIND (n1)])
1134 n1 = XEXP (n1, 1);
1135 while (n2 && !reg_note_cfa_p[REG_NOTE_KIND (n2)])
1136 n2 = XEXP (n2, 1);
1137 if (n1 == NULL_RTX && n2 == NULL_RTX)
1138 return true;
1139 if (n1 == NULL_RTX || n2 == NULL_RTX)
1140 return false;
1141 if (XEXP (n1, 0) == XEXP (n2, 0))
1143 else if (XEXP (n1, 0) == NULL_RTX || XEXP (n2, 0) == NULL_RTX)
1144 return false;
1145 else if (!(reload_completed
1146 ? rtx_renumbered_equal_p (XEXP (n1, 0), XEXP (n2, 0))
1147 : rtx_equal_p (XEXP (n1, 0), XEXP (n2, 0))))
1148 return false;
1152 /* Examine I1 and I2 and return:
1153 - dir_forward if I1 can be replaced by I2, or
1154 - dir_backward if I2 can be replaced by I1, or
1155 - dir_both if both are the case. */
1157 static enum replace_direction
1158 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1160 rtx p1, p2;
1162 /* Verify that I1 and I2 are equivalent. */
1163 if (GET_CODE (i1) != GET_CODE (i2))
1164 return dir_none;
1166 /* __builtin_unreachable() may lead to empty blocks (ending with
1167 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1168 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1169 return dir_both;
1171 /* ??? Do not allow cross-jumping between different stack levels. */
1172 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1173 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1174 if (p1 && p2)
1176 p1 = XEXP (p1, 0);
1177 p2 = XEXP (p2, 0);
1178 if (!rtx_equal_p (p1, p2))
1179 return dir_none;
1181 /* ??? Worse, this adjustment had better be constant lest we
1182 have differing incoming stack levels. */
1183 if (!frame_pointer_needed
1184 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1185 return dir_none;
1187 else if (p1 || p2)
1188 return dir_none;
1190 /* Do not allow cross-jumping between frame related insns and other
1191 insns. */
1192 if (RTX_FRAME_RELATED_P (i1) != RTX_FRAME_RELATED_P (i2))
1193 return dir_none;
1195 p1 = PATTERN (i1);
1196 p2 = PATTERN (i2);
1198 if (GET_CODE (p1) != GET_CODE (p2))
1199 return dir_none;
1201 /* If this is a CALL_INSN, compare register usage information.
1202 If we don't check this on stack register machines, the two
1203 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1204 numbers of stack registers in the same basic block.
1205 If we don't check this on machines with delay slots, a delay slot may
1206 be filled that clobbers a parameter expected by the subroutine.
1208 ??? We take the simple route for now and assume that if they're
1209 equal, they were constructed identically.
1211 Also check for identical exception regions. */
1213 if (CALL_P (i1))
1215 /* Ensure the same EH region. */
1216 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1217 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1219 if (!n1 && n2)
1220 return dir_none;
1222 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1223 return dir_none;
1225 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1226 CALL_INSN_FUNCTION_USAGE (i2))
1227 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1228 return dir_none;
1230 /* For address sanitizer, never crossjump __asan_report_* builtins,
1231 otherwise errors might be reported on incorrect lines. */
1232 if (flag_sanitize & SANITIZE_ADDRESS)
1234 rtx call = get_call_rtx_from (i1);
1235 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1237 rtx symbol = XEXP (XEXP (call, 0), 0);
1238 if (SYMBOL_REF_DECL (symbol)
1239 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1241 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1242 == BUILT_IN_NORMAL)
1243 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1244 >= BUILT_IN_ASAN_REPORT_LOAD1
1245 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1246 <= BUILT_IN_ASAN_STOREN)
1247 return dir_none;
1253 /* If both i1 and i2 are frame related, verify all the CFA notes
1254 in the same order and with the same content. */
1255 if (RTX_FRAME_RELATED_P (i1) && !insns_have_identical_cfa_notes (i1, i2))
1256 return dir_none;
1258 #ifdef STACK_REGS
1259 /* If cross_jump_death_matters is not 0, the insn's mode
1260 indicates whether or not the insn contains any stack-like
1261 regs. */
1263 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1265 /* If register stack conversion has already been done, then
1266 death notes must also be compared before it is certain that
1267 the two instruction streams match. */
1269 rtx note;
1270 HARD_REG_SET i1_regset, i2_regset;
1272 CLEAR_HARD_REG_SET (i1_regset);
1273 CLEAR_HARD_REG_SET (i2_regset);
1275 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1276 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1277 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1279 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1280 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1281 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1283 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1284 return dir_none;
1286 #endif
1288 if (reload_completed
1289 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1290 return dir_both;
1292 return can_replace_by (i1, i2);
1295 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1296 flow_find_head_matching_sequence, ensure the notes match. */
1298 static void
1299 merge_notes (rtx_insn *i1, rtx_insn *i2)
1301 /* If the merged insns have different REG_EQUAL notes, then
1302 remove them. */
1303 rtx equiv1 = find_reg_equal_equiv_note (i1);
1304 rtx equiv2 = find_reg_equal_equiv_note (i2);
1306 if (equiv1 && !equiv2)
1307 remove_note (i1, equiv1);
1308 else if (!equiv1 && equiv2)
1309 remove_note (i2, equiv2);
1310 else if (equiv1 && equiv2
1311 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1313 remove_note (i1, equiv1);
1314 remove_note (i2, equiv2);
1318 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1319 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1320 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1321 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1322 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1324 static void
1325 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1326 bool *did_fallthru)
1328 edge fallthru;
1330 *did_fallthru = false;
1332 /* Ignore notes. */
1333 while (!NONDEBUG_INSN_P (*i1))
1335 if (*i1 != BB_HEAD (*bb1))
1337 *i1 = PREV_INSN (*i1);
1338 continue;
1341 if (!follow_fallthru)
1342 return;
1344 fallthru = find_fallthru_edge ((*bb1)->preds);
1345 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1346 || !single_succ_p (fallthru->src))
1347 return;
1349 *bb1 = fallthru->src;
1350 *i1 = BB_END (*bb1);
1351 *did_fallthru = true;
1355 /* Look through the insns at the end of BB1 and BB2 and find the longest
1356 sequence that are either equivalent, or allow forward or backward
1357 replacement. Store the first insns for that sequence in *F1 and *F2 and
1358 return the sequence length.
1360 DIR_P indicates the allowed replacement direction on function entry, and
1361 the actual replacement direction on function exit. If NULL, only equivalent
1362 sequences are allowed.
1364 To simplify callers of this function, if the blocks match exactly,
1365 store the head of the blocks in *F1 and *F2. */
1368 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1369 rtx_insn **f2, enum replace_direction *dir_p)
1371 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1372 int ninsns = 0;
1373 enum replace_direction dir, last_dir, afterlast_dir;
1374 bool follow_fallthru, did_fallthru;
1376 if (dir_p)
1377 dir = *dir_p;
1378 else
1379 dir = dir_both;
1380 afterlast_dir = dir;
1381 last_dir = afterlast_dir;
1383 /* Skip simple jumps at the end of the blocks. Complex jumps still
1384 need to be compared for equivalence, which we'll do below. */
1386 i1 = BB_END (bb1);
1387 last1 = afterlast1 = last2 = afterlast2 = NULL;
1388 if (onlyjump_p (i1)
1389 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1391 last1 = i1;
1392 i1 = PREV_INSN (i1);
1395 i2 = BB_END (bb2);
1396 if (onlyjump_p (i2)
1397 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1399 last2 = i2;
1400 /* Count everything except for unconditional jump as insn.
1401 Don't count any jumps if dir_p is NULL. */
1402 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1403 ninsns++;
1404 i2 = PREV_INSN (i2);
1407 while (true)
1409 /* In the following example, we can replace all jumps to C by jumps to A.
1411 This removes 4 duplicate insns.
1412 [bb A] insn1 [bb C] insn1
1413 insn2 insn2
1414 [bb B] insn3 insn3
1415 insn4 insn4
1416 jump_insn jump_insn
1418 We could also replace all jumps to A by jumps to C, but that leaves B
1419 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1420 step, all jumps to B would be replaced with jumps to the middle of C,
1421 achieving the same result with more effort.
1422 So we allow only the first possibility, which means that we don't allow
1423 fallthru in the block that's being replaced. */
1425 follow_fallthru = dir_p && dir != dir_forward;
1426 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1427 if (did_fallthru)
1428 dir = dir_backward;
1430 follow_fallthru = dir_p && dir != dir_backward;
1431 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1432 if (did_fallthru)
1433 dir = dir_forward;
1435 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1436 break;
1438 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1439 if (dir == dir_none || (!dir_p && dir != dir_both))
1440 break;
1442 merge_memattrs (i1, i2);
1444 /* Don't begin a cross-jump with a NOTE insn. */
1445 if (INSN_P (i1))
1447 merge_notes (i1, i2);
1449 afterlast1 = last1, afterlast2 = last2;
1450 last1 = i1, last2 = i2;
1451 afterlast_dir = last_dir;
1452 last_dir = dir;
1453 if (active_insn_p (i1))
1454 ninsns++;
1457 i1 = PREV_INSN (i1);
1458 i2 = PREV_INSN (i2);
1461 /* Don't allow the insn after a compare to be shared by
1462 cross-jumping unless the compare is also shared. */
1463 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1464 && ! sets_cc0_p (last1))
1465 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1467 /* Include preceding notes and labels in the cross-jump. One,
1468 this may bring us to the head of the blocks as requested above.
1469 Two, it keeps line number notes as matched as may be. */
1470 if (ninsns)
1472 bb1 = BLOCK_FOR_INSN (last1);
1473 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1474 last1 = PREV_INSN (last1);
1476 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1477 last1 = PREV_INSN (last1);
1479 bb2 = BLOCK_FOR_INSN (last2);
1480 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1481 last2 = PREV_INSN (last2);
1483 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1484 last2 = PREV_INSN (last2);
1486 *f1 = last1;
1487 *f2 = last2;
1490 if (dir_p)
1491 *dir_p = last_dir;
1492 return ninsns;
1495 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1496 the head of the two blocks. Do not include jumps at the end.
1497 If STOP_AFTER is nonzero, stop after finding that many matching
1498 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1499 non-zero, only count active insns. */
1502 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1503 rtx_insn **f2, int stop_after)
1505 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1506 int ninsns = 0;
1507 edge e;
1508 edge_iterator ei;
1509 int nehedges1 = 0, nehedges2 = 0;
1511 FOR_EACH_EDGE (e, ei, bb1->succs)
1512 if (e->flags & EDGE_EH)
1513 nehedges1++;
1514 FOR_EACH_EDGE (e, ei, bb2->succs)
1515 if (e->flags & EDGE_EH)
1516 nehedges2++;
1518 i1 = BB_HEAD (bb1);
1519 i2 = BB_HEAD (bb2);
1520 last1 = beforelast1 = last2 = beforelast2 = NULL;
1522 while (true)
1524 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1525 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1527 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1528 break;
1529 i1 = NEXT_INSN (i1);
1532 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1534 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1535 break;
1536 i2 = NEXT_INSN (i2);
1539 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1540 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1541 break;
1543 if (NOTE_P (i1) || NOTE_P (i2)
1544 || JUMP_P (i1) || JUMP_P (i2))
1545 break;
1547 /* A sanity check to make sure we're not merging insns with different
1548 effects on EH. If only one of them ends a basic block, it shouldn't
1549 have an EH edge; if both end a basic block, there should be the same
1550 number of EH edges. */
1551 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1552 && nehedges1 > 0)
1553 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1554 && nehedges2 > 0)
1555 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1556 && nehedges1 != nehedges2))
1557 break;
1559 if (old_insns_match_p (0, i1, i2) != dir_both)
1560 break;
1562 merge_memattrs (i1, i2);
1564 /* Don't begin a cross-jump with a NOTE insn. */
1565 if (INSN_P (i1))
1567 merge_notes (i1, i2);
1569 beforelast1 = last1, beforelast2 = last2;
1570 last1 = i1, last2 = i2;
1571 if (!stop_after || active_insn_p (i1))
1572 ninsns++;
1575 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1576 || (stop_after > 0 && ninsns == stop_after))
1577 break;
1579 i1 = NEXT_INSN (i1);
1580 i2 = NEXT_INSN (i2);
1583 /* Don't allow a compare to be shared by cross-jumping unless the insn
1584 after the compare is also shared. */
1585 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1586 && sets_cc0_p (last1))
1587 last1 = beforelast1, last2 = beforelast2, ninsns--;
1589 if (ninsns)
1591 *f1 = last1;
1592 *f2 = last2;
1595 return ninsns;
1598 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1599 the branch instruction. This means that if we commonize the control
1600 flow before end of the basic block, the semantic remains unchanged.
1602 We may assume that there exists one edge with a common destination. */
1604 static bool
1605 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1607 int nehedges1 = 0, nehedges2 = 0;
1608 edge fallthru1 = 0, fallthru2 = 0;
1609 edge e1, e2;
1610 edge_iterator ei;
1612 /* If we performed shrink-wrapping, edges to the exit block can
1613 only be distinguished for JUMP_INSNs. The two paths may differ in
1614 whether they went through the prologue. Sibcalls are fine, we know
1615 that we either didn't need or inserted an epilogue before them. */
1616 if (crtl->shrink_wrapped
1617 && single_succ_p (bb1)
1618 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1619 && !JUMP_P (BB_END (bb1))
1620 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1621 return false;
1623 /* If BB1 has only one successor, we may be looking at either an
1624 unconditional jump, or a fake edge to exit. */
1625 if (single_succ_p (bb1)
1626 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1627 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1628 return (single_succ_p (bb2)
1629 && (single_succ_edge (bb2)->flags
1630 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1631 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1633 /* Match conditional jumps - this may get tricky when fallthru and branch
1634 edges are crossed. */
1635 if (EDGE_COUNT (bb1->succs) == 2
1636 && any_condjump_p (BB_END (bb1))
1637 && onlyjump_p (BB_END (bb1)))
1639 edge b1, f1, b2, f2;
1640 bool reverse, match;
1641 rtx set1, set2, cond1, cond2;
1642 enum rtx_code code1, code2;
1644 if (EDGE_COUNT (bb2->succs) != 2
1645 || !any_condjump_p (BB_END (bb2))
1646 || !onlyjump_p (BB_END (bb2)))
1647 return false;
1649 b1 = BRANCH_EDGE (bb1);
1650 b2 = BRANCH_EDGE (bb2);
1651 f1 = FALLTHRU_EDGE (bb1);
1652 f2 = FALLTHRU_EDGE (bb2);
1654 /* Get around possible forwarders on fallthru edges. Other cases
1655 should be optimized out already. */
1656 if (FORWARDER_BLOCK_P (f1->dest))
1657 f1 = single_succ_edge (f1->dest);
1659 if (FORWARDER_BLOCK_P (f2->dest))
1660 f2 = single_succ_edge (f2->dest);
1662 /* To simplify use of this function, return false if there are
1663 unneeded forwarder blocks. These will get eliminated later
1664 during cleanup_cfg. */
1665 if (FORWARDER_BLOCK_P (f1->dest)
1666 || FORWARDER_BLOCK_P (f2->dest)
1667 || FORWARDER_BLOCK_P (b1->dest)
1668 || FORWARDER_BLOCK_P (b2->dest))
1669 return false;
1671 if (f1->dest == f2->dest && b1->dest == b2->dest)
1672 reverse = false;
1673 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1674 reverse = true;
1675 else
1676 return false;
1678 set1 = pc_set (BB_END (bb1));
1679 set2 = pc_set (BB_END (bb2));
1680 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1681 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1682 reverse = !reverse;
1684 cond1 = XEXP (SET_SRC (set1), 0);
1685 cond2 = XEXP (SET_SRC (set2), 0);
1686 code1 = GET_CODE (cond1);
1687 if (reverse)
1688 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1689 else
1690 code2 = GET_CODE (cond2);
1692 if (code2 == UNKNOWN)
1693 return false;
1695 /* Verify codes and operands match. */
1696 match = ((code1 == code2
1697 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1698 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1699 || (code1 == swap_condition (code2)
1700 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1701 XEXP (cond2, 0))
1702 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1703 XEXP (cond2, 1))));
1705 /* If we return true, we will join the blocks. Which means that
1706 we will only have one branch prediction bit to work with. Thus
1707 we require the existing branches to have probabilities that are
1708 roughly similar. */
1709 if (match
1710 && optimize_bb_for_speed_p (bb1)
1711 && optimize_bb_for_speed_p (bb2))
1713 profile_probability prob2;
1715 if (b1->dest == b2->dest)
1716 prob2 = b2->probability;
1717 else
1718 /* Do not use f2 probability as f2 may be forwarded. */
1719 prob2 = b2->probability.invert ();
1721 /* Fail if the difference in probabilities is greater than 50%.
1722 This rules out two well-predicted branches with opposite
1723 outcomes. */
1724 if (b1->probability.differs_lot_from_p (prob2))
1726 if (dump_file)
1728 fprintf (dump_file,
1729 "Outcomes of branch in bb %i and %i differ too"
1730 " much (", bb1->index, bb2->index);
1731 b1->probability.dump (dump_file);
1732 prob2.dump (dump_file);
1733 fprintf (dump_file, ")\n");
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 s2->dest->count -= s2->count;
2150 s2->dest->frequency -= EDGE_FREQUENCY (s);
2151 if (s2->dest->frequency < 0)
2152 s2->dest->frequency = 0;
2155 if (!redirect_edges_to->frequency && !src1->frequency)
2156 s->probability = s->probability.combine_with_freq
2157 (redirect_edges_to->frequency,
2158 s2->probability, src1->frequency);
2161 /* Adjust count and frequency for the block. An earlier jump
2162 threading pass may have left the profile in an inconsistent
2163 state (see update_bb_profile_for_threading) so we must be
2164 prepared for overflows. */
2165 tmp = redirect_to;
2168 tmp->count += src1->count;
2169 tmp->frequency += src1->frequency;
2170 if (tmp->frequency > BB_FREQ_MAX)
2171 tmp->frequency = BB_FREQ_MAX;
2172 if (tmp == redirect_edges_to)
2173 break;
2174 tmp = find_fallthru_edge (tmp->succs)->dest;
2176 while (true);
2177 update_br_prob_note (redirect_edges_to);
2179 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2181 /* Skip possible basic block header. */
2182 if (LABEL_P (newpos1))
2183 newpos1 = NEXT_INSN (newpos1);
2185 while (DEBUG_INSN_P (newpos1))
2186 newpos1 = NEXT_INSN (newpos1);
2188 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2189 newpos1 = NEXT_INSN (newpos1);
2191 while (DEBUG_INSN_P (newpos1))
2192 newpos1 = NEXT_INSN (newpos1);
2194 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2195 to_remove = single_succ (redirect_from);
2197 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2198 delete_basic_block (to_remove);
2200 update_forwarder_flag (redirect_from);
2201 if (redirect_to != src2)
2202 update_forwarder_flag (src2);
2204 return true;
2207 /* Search the predecessors of BB for common insn sequences. When found,
2208 share code between them by redirecting control flow. Return true if
2209 any changes made. */
2211 static bool
2212 try_crossjump_bb (int mode, basic_block bb)
2214 edge e, e2, fallthru;
2215 bool changed;
2216 unsigned max, ix, ix2;
2218 /* Nothing to do if there is not at least two incoming edges. */
2219 if (EDGE_COUNT (bb->preds) < 2)
2220 return false;
2222 /* Don't crossjump if this block ends in a computed jump,
2223 unless we are optimizing for size. */
2224 if (optimize_bb_for_size_p (bb)
2225 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2226 && computed_jump_p (BB_END (bb)))
2227 return false;
2229 /* If we are partitioning hot/cold basic blocks, we don't want to
2230 mess up unconditional or indirect jumps that cross between hot
2231 and cold sections.
2233 Basic block partitioning may result in some jumps that appear to
2234 be optimizable (or blocks that appear to be mergeable), but which really
2235 must be left untouched (they are required to make it safely across
2236 partition boundaries). See the comments at the top of
2237 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2239 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2240 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2241 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2242 return false;
2244 /* It is always cheapest to redirect a block that ends in a branch to
2245 a block that falls through into BB, as that adds no branches to the
2246 program. We'll try that combination first. */
2247 fallthru = NULL;
2248 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2250 if (EDGE_COUNT (bb->preds) > max)
2251 return false;
2253 fallthru = find_fallthru_edge (bb->preds);
2255 changed = false;
2256 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2258 e = EDGE_PRED (bb, ix);
2259 ix++;
2261 /* As noted above, first try with the fallthru predecessor (or, a
2262 fallthru predecessor if we are in cfglayout mode). */
2263 if (fallthru)
2265 /* Don't combine the fallthru edge into anything else.
2266 If there is a match, we'll do it the other way around. */
2267 if (e == fallthru)
2268 continue;
2269 /* If nothing changed since the last attempt, there is nothing
2270 we can do. */
2271 if (!first_pass
2272 && !((e->src->flags & BB_MODIFIED)
2273 || (fallthru->src->flags & BB_MODIFIED)))
2274 continue;
2276 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2278 changed = true;
2279 ix = 0;
2280 continue;
2284 /* Non-obvious work limiting check: Recognize that we're going
2285 to call try_crossjump_bb on every basic block. So if we have
2286 two blocks with lots of outgoing edges (a switch) and they
2287 share lots of common destinations, then we would do the
2288 cross-jump check once for each common destination.
2290 Now, if the blocks actually are cross-jump candidates, then
2291 all of their destinations will be shared. Which means that
2292 we only need check them for cross-jump candidacy once. We
2293 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2294 choosing to do the check from the block for which the edge
2295 in question is the first successor of A. */
2296 if (EDGE_SUCC (e->src, 0) != e)
2297 continue;
2299 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2301 e2 = EDGE_PRED (bb, ix2);
2303 if (e2 == e)
2304 continue;
2306 /* We've already checked the fallthru edge above. */
2307 if (e2 == fallthru)
2308 continue;
2310 /* The "first successor" check above only prevents multiple
2311 checks of crossjump(A,B). In order to prevent redundant
2312 checks of crossjump(B,A), require that A be the block
2313 with the lowest index. */
2314 if (e->src->index > e2->src->index)
2315 continue;
2317 /* If nothing changed since the last attempt, there is nothing
2318 we can do. */
2319 if (!first_pass
2320 && !((e->src->flags & BB_MODIFIED)
2321 || (e2->src->flags & BB_MODIFIED)))
2322 continue;
2324 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2325 direction. */
2326 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2328 changed = true;
2329 ix = 0;
2330 break;
2335 if (changed)
2336 crossjumps_occurred = true;
2338 return changed;
2341 /* Search the successors of BB for common insn sequences. When found,
2342 share code between them by moving it across the basic block
2343 boundary. Return true if any changes made. */
2345 static bool
2346 try_head_merge_bb (basic_block bb)
2348 basic_block final_dest_bb = NULL;
2349 int max_match = INT_MAX;
2350 edge e0;
2351 rtx_insn **headptr, **currptr, **nextptr;
2352 bool changed, moveall;
2353 unsigned ix;
2354 rtx_insn *e0_last_head;
2355 rtx cond;
2356 rtx_insn *move_before;
2357 unsigned nedges = EDGE_COUNT (bb->succs);
2358 rtx_insn *jump = BB_END (bb);
2359 regset live, live_union;
2361 /* Nothing to do if there is not at least two outgoing edges. */
2362 if (nedges < 2)
2363 return false;
2365 /* Don't crossjump if this block ends in a computed jump,
2366 unless we are optimizing for size. */
2367 if (optimize_bb_for_size_p (bb)
2368 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2369 && computed_jump_p (BB_END (bb)))
2370 return false;
2372 cond = get_condition (jump, &move_before, true, false);
2373 if (cond == NULL_RTX)
2375 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2376 move_before = prev_nonnote_nondebug_insn (jump);
2377 else
2378 move_before = jump;
2381 for (ix = 0; ix < nedges; ix++)
2382 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2383 return false;
2385 for (ix = 0; ix < nedges; ix++)
2387 edge e = EDGE_SUCC (bb, ix);
2388 basic_block other_bb = e->dest;
2390 if (df_get_bb_dirty (other_bb))
2392 block_was_dirty = true;
2393 return false;
2396 if (e->flags & EDGE_ABNORMAL)
2397 return false;
2399 /* Normally, all destination blocks must only be reachable from this
2400 block, i.e. they must have one incoming edge.
2402 There is one special case we can handle, that of multiple consecutive
2403 jumps where the first jumps to one of the targets of the second jump.
2404 This happens frequently in switch statements for default labels.
2405 The structure is as follows:
2406 FINAL_DEST_BB
2407 ....
2408 if (cond) jump A;
2409 fall through
2411 jump with targets A, B, C, D...
2413 has two incoming edges, from FINAL_DEST_BB and BB
2415 In this case, we can try to move the insns through BB and into
2416 FINAL_DEST_BB. */
2417 if (EDGE_COUNT (other_bb->preds) != 1)
2419 edge incoming_edge, incoming_bb_other_edge;
2420 edge_iterator ei;
2422 if (final_dest_bb != NULL
2423 || EDGE_COUNT (other_bb->preds) != 2)
2424 return false;
2426 /* We must be able to move the insns across the whole block. */
2427 move_before = BB_HEAD (bb);
2428 while (!NONDEBUG_INSN_P (move_before))
2429 move_before = NEXT_INSN (move_before);
2431 if (EDGE_COUNT (bb->preds) != 1)
2432 return false;
2433 incoming_edge = EDGE_PRED (bb, 0);
2434 final_dest_bb = incoming_edge->src;
2435 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2436 return false;
2437 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2438 if (incoming_bb_other_edge != incoming_edge)
2439 break;
2440 if (incoming_bb_other_edge->dest != other_bb)
2441 return false;
2445 e0 = EDGE_SUCC (bb, 0);
2446 e0_last_head = NULL;
2447 changed = false;
2449 for (ix = 1; ix < nedges; ix++)
2451 edge e = EDGE_SUCC (bb, ix);
2452 rtx_insn *e0_last, *e_last;
2453 int nmatch;
2455 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2456 &e0_last, &e_last, 0);
2457 if (nmatch == 0)
2458 return false;
2460 if (nmatch < max_match)
2462 max_match = nmatch;
2463 e0_last_head = e0_last;
2467 /* If we matched an entire block, we probably have to avoid moving the
2468 last insn. */
2469 if (max_match > 0
2470 && e0_last_head == BB_END (e0->dest)
2471 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2472 || control_flow_insn_p (e0_last_head)))
2474 max_match--;
2475 if (max_match == 0)
2476 return false;
2478 e0_last_head = prev_real_insn (e0_last_head);
2479 while (DEBUG_INSN_P (e0_last_head));
2482 if (max_match == 0)
2483 return false;
2485 /* We must find a union of the live registers at each of the end points. */
2486 live = BITMAP_ALLOC (NULL);
2487 live_union = BITMAP_ALLOC (NULL);
2489 currptr = XNEWVEC (rtx_insn *, nedges);
2490 headptr = XNEWVEC (rtx_insn *, nedges);
2491 nextptr = XNEWVEC (rtx_insn *, nedges);
2493 for (ix = 0; ix < nedges; ix++)
2495 int j;
2496 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2497 rtx_insn *head = BB_HEAD (merge_bb);
2499 while (!NONDEBUG_INSN_P (head))
2500 head = NEXT_INSN (head);
2501 headptr[ix] = head;
2502 currptr[ix] = head;
2504 /* Compute the end point and live information */
2505 for (j = 1; j < max_match; j++)
2507 head = NEXT_INSN (head);
2508 while (!NONDEBUG_INSN_P (head));
2509 simulate_backwards_to_point (merge_bb, live, head);
2510 IOR_REG_SET (live_union, live);
2513 /* If we're moving across two blocks, verify the validity of the
2514 first move, then adjust the target and let the loop below deal
2515 with the final move. */
2516 if (final_dest_bb != NULL)
2518 rtx_insn *move_upto;
2520 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2521 jump, e0->dest, live_union,
2522 NULL, &move_upto);
2523 if (!moveall)
2525 if (move_upto == NULL_RTX)
2526 goto out;
2528 while (e0_last_head != move_upto)
2530 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2531 live_union);
2532 e0_last_head = PREV_INSN (e0_last_head);
2535 if (e0_last_head == NULL_RTX)
2536 goto out;
2538 jump = BB_END (final_dest_bb);
2539 cond = get_condition (jump, &move_before, true, false);
2540 if (cond == NULL_RTX)
2542 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2543 move_before = prev_nonnote_nondebug_insn (jump);
2544 else
2545 move_before = jump;
2551 rtx_insn *move_upto;
2552 moveall = can_move_insns_across (currptr[0], e0_last_head,
2553 move_before, jump, e0->dest, live_union,
2554 NULL, &move_upto);
2555 if (!moveall && move_upto == NULL_RTX)
2557 if (jump == move_before)
2558 break;
2560 /* Try again, using a different insertion point. */
2561 move_before = jump;
2563 /* Don't try moving before a cc0 user, as that may invalidate
2564 the cc0. */
2565 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2566 break;
2568 continue;
2571 if (final_dest_bb && !moveall)
2572 /* We haven't checked whether a partial move would be OK for the first
2573 move, so we have to fail this case. */
2574 break;
2576 changed = true;
2577 for (;;)
2579 if (currptr[0] == move_upto)
2580 break;
2581 for (ix = 0; ix < nedges; ix++)
2583 rtx_insn *curr = currptr[ix];
2585 curr = NEXT_INSN (curr);
2586 while (!NONDEBUG_INSN_P (curr));
2587 currptr[ix] = curr;
2591 /* If we can't currently move all of the identical insns, remember
2592 each insn after the range that we'll merge. */
2593 if (!moveall)
2594 for (ix = 0; ix < nedges; ix++)
2596 rtx_insn *curr = currptr[ix];
2598 curr = NEXT_INSN (curr);
2599 while (!NONDEBUG_INSN_P (curr));
2600 nextptr[ix] = curr;
2603 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2604 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2605 if (final_dest_bb != NULL)
2606 df_set_bb_dirty (final_dest_bb);
2607 df_set_bb_dirty (bb);
2608 for (ix = 1; ix < nedges; ix++)
2610 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2611 delete_insn_chain (headptr[ix], currptr[ix], false);
2613 if (!moveall)
2615 if (jump == move_before)
2616 break;
2618 /* For the unmerged insns, try a different insertion point. */
2619 move_before = jump;
2621 /* Don't try moving before a cc0 user, as that may invalidate
2622 the cc0. */
2623 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2624 break;
2626 for (ix = 0; ix < nedges; ix++)
2627 currptr[ix] = headptr[ix] = nextptr[ix];
2630 while (!moveall);
2632 out:
2633 free (currptr);
2634 free (headptr);
2635 free (nextptr);
2637 crossjumps_occurred |= changed;
2639 return changed;
2642 /* Return true if BB contains just bb note, or bb note followed
2643 by only DEBUG_INSNs. */
2645 static bool
2646 trivially_empty_bb_p (basic_block bb)
2648 rtx_insn *insn = BB_END (bb);
2650 while (1)
2652 if (insn == BB_HEAD (bb))
2653 return true;
2654 if (!DEBUG_INSN_P (insn))
2655 return false;
2656 insn = PREV_INSN (insn);
2660 /* Return true if BB contains just a return and possibly a USE of the
2661 return value. Fill in *RET and *USE with the return and use insns
2662 if any found, otherwise NULL. All CLOBBERs are ignored. */
2664 static bool
2665 bb_is_just_return (basic_block bb, rtx_insn **ret, rtx_insn **use)
2667 *ret = *use = NULL;
2668 rtx_insn *insn;
2670 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
2671 return false;
2673 FOR_BB_INSNS (bb, insn)
2674 if (NONDEBUG_INSN_P (insn))
2676 rtx pat = PATTERN (insn);
2678 if (!*ret && ANY_RETURN_P (pat))
2679 *ret = insn;
2680 else if (!*ret && !*use && GET_CODE (pat) == USE
2681 && REG_P (XEXP (pat, 0))
2682 && REG_FUNCTION_VALUE_P (XEXP (pat, 0)))
2683 *use = insn;
2684 else if (GET_CODE (pat) != CLOBBER)
2685 return false;
2688 return !!*ret;
2691 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2692 instructions etc. Return nonzero if changes were made. */
2694 static bool
2695 try_optimize_cfg (int mode)
2697 bool changed_overall = false;
2698 bool changed;
2699 int iterations = 0;
2700 basic_block bb, b, next;
2702 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2703 clear_bb_flags ();
2705 crossjumps_occurred = false;
2707 FOR_EACH_BB_FN (bb, cfun)
2708 update_forwarder_flag (bb);
2710 if (! targetm.cannot_modify_jumps_p ())
2712 first_pass = true;
2713 /* Attempt to merge blocks as made possible by edge removal. If
2714 a block has only one successor, and the successor has only
2715 one predecessor, they may be combined. */
2718 block_was_dirty = false;
2719 changed = false;
2720 iterations++;
2722 if (dump_file)
2723 fprintf (dump_file,
2724 "\n\ntry_optimize_cfg iteration %i\n\n",
2725 iterations);
2727 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2728 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2730 basic_block c;
2731 edge s;
2732 bool changed_here = false;
2734 /* Delete trivially dead basic blocks. This is either
2735 blocks with no predecessors, or empty blocks with no
2736 successors. However if the empty block with no
2737 successors is the successor of the ENTRY_BLOCK, it is
2738 kept. This ensures that the ENTRY_BLOCK will have a
2739 successor which is a precondition for many RTL
2740 passes. Empty blocks may result from expanding
2741 __builtin_unreachable (). */
2742 if (EDGE_COUNT (b->preds) == 0
2743 || (EDGE_COUNT (b->succs) == 0
2744 && trivially_empty_bb_p (b)
2745 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2746 != b))
2748 c = b->prev_bb;
2749 if (EDGE_COUNT (b->preds) > 0)
2751 edge e;
2752 edge_iterator ei;
2754 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2756 if (BB_FOOTER (b)
2757 && BARRIER_P (BB_FOOTER (b)))
2758 FOR_EACH_EDGE (e, ei, b->preds)
2759 if ((e->flags & EDGE_FALLTHRU)
2760 && BB_FOOTER (e->src) == NULL)
2762 if (BB_FOOTER (b))
2764 BB_FOOTER (e->src) = BB_FOOTER (b);
2765 BB_FOOTER (b) = NULL;
2767 else
2769 start_sequence ();
2770 BB_FOOTER (e->src) = emit_barrier ();
2771 end_sequence ();
2775 else
2777 rtx_insn *last = get_last_bb_insn (b);
2778 if (last && BARRIER_P (last))
2779 FOR_EACH_EDGE (e, ei, b->preds)
2780 if ((e->flags & EDGE_FALLTHRU))
2781 emit_barrier_after (BB_END (e->src));
2784 delete_basic_block (b);
2785 changed = true;
2786 /* Avoid trying to remove the exit block. */
2787 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2788 continue;
2791 /* Remove code labels no longer used. */
2792 if (single_pred_p (b)
2793 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2794 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2795 && LABEL_P (BB_HEAD (b))
2796 && !LABEL_PRESERVE_P (BB_HEAD (b))
2797 /* If the previous block ends with a branch to this
2798 block, we can't delete the label. Normally this
2799 is a condjump that is yet to be simplified, but
2800 if CASE_DROPS_THRU, this can be a tablejump with
2801 some element going to the same place as the
2802 default (fallthru). */
2803 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2804 || !JUMP_P (BB_END (single_pred (b)))
2805 || ! label_is_jump_target_p (BB_HEAD (b),
2806 BB_END (single_pred (b)))))
2808 delete_insn (BB_HEAD (b));
2809 if (dump_file)
2810 fprintf (dump_file, "Deleted label in block %i.\n",
2811 b->index);
2814 /* If we fall through an empty block, we can remove it. */
2815 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2816 && single_pred_p (b)
2817 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2818 && !LABEL_P (BB_HEAD (b))
2819 && FORWARDER_BLOCK_P (b)
2820 /* Note that forwarder_block_p true ensures that
2821 there is a successor for this block. */
2822 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2823 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2825 if (dump_file)
2826 fprintf (dump_file,
2827 "Deleting fallthru block %i.\n",
2828 b->index);
2830 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2831 ? b->next_bb : b->prev_bb);
2832 redirect_edge_succ_nodup (single_pred_edge (b),
2833 single_succ (b));
2834 delete_basic_block (b);
2835 changed = true;
2836 b = c;
2837 continue;
2840 /* Merge B with its single successor, if any. */
2841 if (single_succ_p (b)
2842 && (s = single_succ_edge (b))
2843 && !(s->flags & EDGE_COMPLEX)
2844 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2845 && single_pred_p (c)
2846 && b != c)
2848 /* When not in cfg_layout mode use code aware of reordering
2849 INSN. This code possibly creates new basic blocks so it
2850 does not fit merge_blocks interface and is kept here in
2851 hope that it will become useless once more of compiler
2852 is transformed to use cfg_layout mode. */
2854 if ((mode & CLEANUP_CFGLAYOUT)
2855 && can_merge_blocks_p (b, c))
2857 merge_blocks (b, c);
2858 update_forwarder_flag (b);
2859 changed_here = true;
2861 else if (!(mode & CLEANUP_CFGLAYOUT)
2862 /* If the jump insn has side effects,
2863 we can't kill the edge. */
2864 && (!JUMP_P (BB_END (b))
2865 || (reload_completed
2866 ? simplejump_p (BB_END (b))
2867 : (onlyjump_p (BB_END (b))
2868 && !tablejump_p (BB_END (b),
2869 NULL, NULL))))
2870 && (next = merge_blocks_move (s, b, c, mode)))
2872 b = next;
2873 changed_here = true;
2877 /* Try to change a branch to a return to just that return. */
2878 rtx_insn *ret, *use;
2879 if (single_succ_p (b)
2880 && onlyjump_p (BB_END (b))
2881 && bb_is_just_return (single_succ (b), &ret, &use))
2883 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2884 PATTERN (ret), 0))
2886 if (use)
2887 emit_insn_before (copy_insn (PATTERN (use)),
2888 BB_END (b));
2889 if (dump_file)
2890 fprintf (dump_file, "Changed jump %d->%d to return.\n",
2891 b->index, single_succ (b)->index);
2892 redirect_edge_succ (single_succ_edge (b),
2893 EXIT_BLOCK_PTR_FOR_FN (cfun));
2894 single_succ_edge (b)->flags &= ~EDGE_CROSSING;
2895 changed_here = true;
2899 /* Try to change a conditional branch to a return to the
2900 respective conditional return. */
2901 if (EDGE_COUNT (b->succs) == 2
2902 && any_condjump_p (BB_END (b))
2903 && bb_is_just_return (BRANCH_EDGE (b)->dest, &ret, &use))
2905 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2906 PATTERN (ret), 0))
2908 if (use)
2909 emit_insn_before (copy_insn (PATTERN (use)),
2910 BB_END (b));
2911 if (dump_file)
2912 fprintf (dump_file, "Changed conditional jump %d->%d "
2913 "to conditional return.\n",
2914 b->index, BRANCH_EDGE (b)->dest->index);
2915 redirect_edge_succ (BRANCH_EDGE (b),
2916 EXIT_BLOCK_PTR_FOR_FN (cfun));
2917 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2918 changed_here = true;
2922 /* Try to flip a conditional branch that falls through to
2923 a return so that it becomes a conditional return and a
2924 new jump to the original branch target. */
2925 if (EDGE_COUNT (b->succs) == 2
2926 && BRANCH_EDGE (b)->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
2927 && any_condjump_p (BB_END (b))
2928 && bb_is_just_return (FALLTHRU_EDGE (b)->dest, &ret, &use))
2930 if (invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2931 JUMP_LABEL (BB_END (b)), 0))
2933 basic_block new_ft = BRANCH_EDGE (b)->dest;
2934 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2935 PATTERN (ret), 0))
2937 if (use)
2938 emit_insn_before (copy_insn (PATTERN (use)),
2939 BB_END (b));
2940 if (dump_file)
2941 fprintf (dump_file, "Changed conditional jump "
2942 "%d->%d to conditional return, adding "
2943 "fall-through jump.\n",
2944 b->index, BRANCH_EDGE (b)->dest->index);
2945 redirect_edge_succ (BRANCH_EDGE (b),
2946 EXIT_BLOCK_PTR_FOR_FN (cfun));
2947 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2948 std::swap (BRANCH_EDGE (b)->probability,
2949 FALLTHRU_EDGE (b)->probability);
2950 update_br_prob_note (b);
2951 basic_block jb = force_nonfallthru (FALLTHRU_EDGE (b));
2952 notice_new_block (jb);
2953 if (!redirect_jump (as_a <rtx_jump_insn *> (BB_END (jb)),
2954 block_label (new_ft), 0))
2955 gcc_unreachable ();
2956 redirect_edge_succ (single_succ_edge (jb), new_ft);
2957 changed_here = true;
2959 else
2961 /* Invert the jump back to what it was. This should
2962 never fail. */
2963 if (!invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2964 JUMP_LABEL (BB_END (b)), 0))
2965 gcc_unreachable ();
2970 /* Simplify branch over branch. */
2971 if ((mode & CLEANUP_EXPENSIVE)
2972 && !(mode & CLEANUP_CFGLAYOUT)
2973 && try_simplify_condjump (b))
2974 changed_here = true;
2976 /* If B has a single outgoing edge, but uses a
2977 non-trivial jump instruction without side-effects, we
2978 can either delete the jump entirely, or replace it
2979 with a simple unconditional jump. */
2980 if (single_succ_p (b)
2981 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2982 && onlyjump_p (BB_END (b))
2983 && !CROSSING_JUMP_P (BB_END (b))
2984 && try_redirect_by_replacing_jump (single_succ_edge (b),
2985 single_succ (b),
2986 (mode & CLEANUP_CFGLAYOUT) != 0))
2988 update_forwarder_flag (b);
2989 changed_here = true;
2992 /* Simplify branch to branch. */
2993 if (try_forward_edges (mode, b))
2995 update_forwarder_flag (b);
2996 changed_here = true;
2999 /* Look for shared code between blocks. */
3000 if ((mode & CLEANUP_CROSSJUMP)
3001 && try_crossjump_bb (mode, b))
3002 changed_here = true;
3004 if ((mode & CLEANUP_CROSSJUMP)
3005 /* This can lengthen register lifetimes. Do it only after
3006 reload. */
3007 && reload_completed
3008 && try_head_merge_bb (b))
3009 changed_here = true;
3011 /* Don't get confused by the index shift caused by
3012 deleting blocks. */
3013 if (!changed_here)
3014 b = b->next_bb;
3015 else
3016 changed = true;
3019 if ((mode & CLEANUP_CROSSJUMP)
3020 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
3021 changed = true;
3023 if (block_was_dirty)
3025 /* This should only be set by head-merging. */
3026 gcc_assert (mode & CLEANUP_CROSSJUMP);
3027 df_analyze ();
3030 if (changed)
3032 /* Edge forwarding in particular can cause hot blocks previously
3033 reached by both hot and cold blocks to become dominated only
3034 by cold blocks. This will cause the verification below to fail,
3035 and lead to now cold code in the hot section. This is not easy
3036 to detect and fix during edge forwarding, and in some cases
3037 is only visible after newly unreachable blocks are deleted,
3038 which will be done in fixup_partitions. */
3039 fixup_partitions ();
3040 checking_verify_flow_info ();
3043 changed_overall |= changed;
3044 first_pass = false;
3046 while (changed);
3049 FOR_ALL_BB_FN (b, cfun)
3050 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
3052 return changed_overall;
3055 /* Delete all unreachable basic blocks. */
3057 bool
3058 delete_unreachable_blocks (void)
3060 bool changed = false;
3061 basic_block b, prev_bb;
3063 find_unreachable_blocks ();
3065 /* When we're in GIMPLE mode and there may be debug insns, we should
3066 delete blocks in reverse dominator order, so as to get a chance
3067 to substitute all released DEFs into debug stmts. If we don't
3068 have dominators information, walking blocks backward gets us a
3069 better chance of retaining most debug information than
3070 otherwise. */
3071 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
3072 && dom_info_available_p (CDI_DOMINATORS))
3074 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3075 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3077 prev_bb = b->prev_bb;
3079 if (!(b->flags & BB_REACHABLE))
3081 /* Speed up the removal of blocks that don't dominate
3082 others. Walking backwards, this should be the common
3083 case. */
3084 if (!first_dom_son (CDI_DOMINATORS, b))
3085 delete_basic_block (b);
3086 else
3088 vec<basic_block> h
3089 = get_all_dominated_blocks (CDI_DOMINATORS, b);
3091 while (h.length ())
3093 b = h.pop ();
3095 prev_bb = b->prev_bb;
3097 gcc_assert (!(b->flags & BB_REACHABLE));
3099 delete_basic_block (b);
3102 h.release ();
3105 changed = true;
3109 else
3111 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3112 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3114 prev_bb = b->prev_bb;
3116 if (!(b->flags & BB_REACHABLE))
3118 delete_basic_block (b);
3119 changed = true;
3124 if (changed)
3125 tidy_fallthru_edges ();
3126 return changed;
3129 /* Delete any jump tables never referenced. We can't delete them at the
3130 time of removing tablejump insn as they are referenced by the preceding
3131 insns computing the destination, so we delay deleting and garbagecollect
3132 them once life information is computed. */
3133 void
3134 delete_dead_jumptables (void)
3136 basic_block bb;
3138 /* A dead jump table does not belong to any basic block. Scan insns
3139 between two adjacent basic blocks. */
3140 FOR_EACH_BB_FN (bb, cfun)
3142 rtx_insn *insn, *next;
3144 for (insn = NEXT_INSN (BB_END (bb));
3145 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
3146 insn = next)
3148 next = NEXT_INSN (insn);
3149 if (LABEL_P (insn)
3150 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
3151 && JUMP_TABLE_DATA_P (next))
3153 rtx_insn *label = insn, *jump = next;
3155 if (dump_file)
3156 fprintf (dump_file, "Dead jumptable %i removed\n",
3157 INSN_UID (insn));
3159 next = NEXT_INSN (next);
3160 delete_insn (jump);
3161 delete_insn (label);
3168 /* Tidy the CFG by deleting unreachable code and whatnot. */
3170 bool
3171 cleanup_cfg (int mode)
3173 bool changed = false;
3175 /* Set the cfglayout mode flag here. We could update all the callers
3176 but that is just inconvenient, especially given that we eventually
3177 want to have cfglayout mode as the default. */
3178 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3179 mode |= CLEANUP_CFGLAYOUT;
3181 timevar_push (TV_CLEANUP_CFG);
3182 if (delete_unreachable_blocks ())
3184 changed = true;
3185 /* We've possibly created trivially dead code. Cleanup it right
3186 now to introduce more opportunities for try_optimize_cfg. */
3187 if (!(mode & (CLEANUP_NO_INSN_DEL))
3188 && !reload_completed)
3189 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3192 compact_blocks ();
3194 /* To tail-merge blocks ending in the same noreturn function (e.g.
3195 a call to abort) we have to insert fake edges to exit. Do this
3196 here once. The fake edges do not interfere with any other CFG
3197 cleanups. */
3198 if (mode & CLEANUP_CROSSJUMP)
3199 add_noreturn_fake_exit_edges ();
3201 if (!dbg_cnt (cfg_cleanup))
3202 return changed;
3204 while (try_optimize_cfg (mode))
3206 delete_unreachable_blocks (), changed = true;
3207 if (!(mode & CLEANUP_NO_INSN_DEL))
3209 /* Try to remove some trivially dead insns when doing an expensive
3210 cleanup. But delete_trivially_dead_insns doesn't work after
3211 reload (it only handles pseudos) and run_fast_dce is too costly
3212 to run in every iteration.
3214 For effective cross jumping, we really want to run a fast DCE to
3215 clean up any dead conditions, or they get in the way of performing
3216 useful tail merges.
3218 Other transformations in cleanup_cfg are not so sensitive to dead
3219 code, so delete_trivially_dead_insns or even doing nothing at all
3220 is good enough. */
3221 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3222 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3223 break;
3224 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occurred)
3225 run_fast_dce ();
3227 else
3228 break;
3231 if (mode & CLEANUP_CROSSJUMP)
3232 remove_fake_exit_edges ();
3234 /* Don't call delete_dead_jumptables in cfglayout mode, because
3235 that function assumes that jump tables are in the insns stream.
3236 But we also don't _have_ to delete dead jumptables in cfglayout
3237 mode because we shouldn't even be looking at things that are
3238 not in a basic block. Dead jumptables are cleaned up when
3239 going out of cfglayout mode. */
3240 if (!(mode & CLEANUP_CFGLAYOUT))
3241 delete_dead_jumptables ();
3243 /* ??? We probably do this way too often. */
3244 if (current_loops
3245 && (changed
3246 || (mode & CLEANUP_CFG_CHANGED)))
3248 timevar_push (TV_REPAIR_LOOPS);
3249 /* The above doesn't preserve dominance info if available. */
3250 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3251 calculate_dominance_info (CDI_DOMINATORS);
3252 fix_loop_structure (NULL);
3253 free_dominance_info (CDI_DOMINATORS);
3254 timevar_pop (TV_REPAIR_LOOPS);
3257 timevar_pop (TV_CLEANUP_CFG);
3259 return changed;
3262 namespace {
3264 const pass_data pass_data_jump =
3266 RTL_PASS, /* type */
3267 "jump", /* name */
3268 OPTGROUP_NONE, /* optinfo_flags */
3269 TV_JUMP, /* tv_id */
3270 0, /* properties_required */
3271 0, /* properties_provided */
3272 0, /* properties_destroyed */
3273 0, /* todo_flags_start */
3274 0, /* todo_flags_finish */
3277 class pass_jump : public rtl_opt_pass
3279 public:
3280 pass_jump (gcc::context *ctxt)
3281 : rtl_opt_pass (pass_data_jump, ctxt)
3284 /* opt_pass methods: */
3285 virtual unsigned int execute (function *);
3287 }; // class pass_jump
3289 unsigned int
3290 pass_jump::execute (function *)
3292 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3293 if (dump_file)
3294 dump_flow_info (dump_file, dump_flags);
3295 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3296 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3297 return 0;
3300 } // anon namespace
3302 rtl_opt_pass *
3303 make_pass_jump (gcc::context *ctxt)
3305 return new pass_jump (ctxt);
3308 namespace {
3310 const pass_data pass_data_jump2 =
3312 RTL_PASS, /* type */
3313 "jump2", /* name */
3314 OPTGROUP_NONE, /* optinfo_flags */
3315 TV_JUMP, /* tv_id */
3316 0, /* properties_required */
3317 0, /* properties_provided */
3318 0, /* properties_destroyed */
3319 0, /* todo_flags_start */
3320 0, /* todo_flags_finish */
3323 class pass_jump2 : public rtl_opt_pass
3325 public:
3326 pass_jump2 (gcc::context *ctxt)
3327 : rtl_opt_pass (pass_data_jump2, ctxt)
3330 /* opt_pass methods: */
3331 virtual unsigned int execute (function *)
3333 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3334 return 0;
3337 }; // class pass_jump2
3339 } // anon namespace
3341 rtl_opt_pass *
3342 make_pass_jump2 (gcc::context *ctxt)
3344 return new pass_jump2 (ctxt);