* lto-partition.c (add_symbol_to_partition_1,
[official-gcc.git] / gcc / cfgcleanup.c
blob77196ee6bf71e7ddd35e2589379bfd6da9932d84
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 /* This file contains optimizer of the control flow. The main entry point is
21 cleanup_cfg. Following optimizations are performed:
23 - Unreachable blocks removal
24 - Edge forwarding (edge to the forwarder block is forwarded to its
25 successor. Simplification of the branch instruction is performed by
26 underlying infrastructure so branch can be converted to simplejump or
27 eliminated).
28 - Cross jumping (tail merging)
29 - Conditional jump-around-simplejump simplification
30 - Basic block merging. */
32 #include "config.h"
33 #include "system.h"
34 #include "coretypes.h"
35 #include "tm.h"
36 #include "rtl.h"
37 #include "tree.h"
38 #include "hard-reg-set.h"
39 #include "regs.h"
40 #include "insn-config.h"
41 #include "flags.h"
42 #include "recog.h"
43 #include "diagnostic-core.h"
44 #include "cselib.h"
45 #include "params.h"
46 #include "tm_p.h"
47 #include "target.h"
48 #include "function.h" /* For inline functions in emit-rtl.h they need crtl. */
49 #include "emit-rtl.h"
50 #include "tree-pass.h"
51 #include "cfgloop.h"
52 #include "expr.h"
53 #include "df.h"
54 #include "dce.h"
55 #include "dbgcnt.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_occured;
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, rtx);
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 int mentions_nonequal_regs (rtx *, void *);
84 static void merge_memattrs (rtx, rtx);
86 /* Set flags for newly created block. */
88 static void
89 notice_new_block (basic_block bb)
91 if (!bb)
92 return;
94 if (forwarder_block_p (bb))
95 bb->flags |= BB_FORWARDER_BLOCK;
98 /* Recompute forwarder flag after block has been modified. */
100 static void
101 update_forwarder_flag (basic_block bb)
103 if (forwarder_block_p (bb))
104 bb->flags |= BB_FORWARDER_BLOCK;
105 else
106 bb->flags &= ~BB_FORWARDER_BLOCK;
109 /* Simplify a conditional jump around an unconditional jump.
110 Return true if something changed. */
112 static bool
113 try_simplify_condjump (basic_block cbranch_block)
115 basic_block jump_block, jump_dest_block, cbranch_dest_block;
116 edge cbranch_jump_edge, cbranch_fallthru_edge;
117 rtx cbranch_insn;
119 /* Verify that there are exactly two successors. */
120 if (EDGE_COUNT (cbranch_block->succs) != 2)
121 return false;
123 /* Verify that we've got a normal conditional branch at the end
124 of the block. */
125 cbranch_insn = BB_END (cbranch_block);
126 if (!any_condjump_p (cbranch_insn))
127 return false;
129 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
130 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
132 /* The next block must not have multiple predecessors, must not
133 be the last block in the function, and must contain just the
134 unconditional jump. */
135 jump_block = cbranch_fallthru_edge->dest;
136 if (!single_pred_p (jump_block)
137 || jump_block->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
138 || !FORWARDER_BLOCK_P (jump_block))
139 return false;
140 jump_dest_block = single_succ (jump_block);
142 /* If we are partitioning hot/cold basic blocks, we don't want to
143 mess up unconditional or indirect jumps that cross between hot
144 and cold sections.
146 Basic block partitioning may result in some jumps that appear to
147 be optimizable (or blocks that appear to be mergeable), but which really
148 must be left untouched (they are required to make it safely across
149 partition boundaries). See the comments at the top of
150 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
152 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
153 || (cbranch_jump_edge->flags & EDGE_CROSSING))
154 return false;
156 /* The conditional branch must target the block after the
157 unconditional branch. */
158 cbranch_dest_block = cbranch_jump_edge->dest;
160 if (cbranch_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 (cbranch_insn, block_label (jump_dest_block), 0))
166 return false;
168 if (dump_file)
169 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
170 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
172 /* Success. Update the CFG to match. Note that after this point
173 the edge variable names appear backwards; the redirection is done
174 this way to preserve edge profile data. */
175 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
176 cbranch_dest_block);
177 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
178 jump_dest_block);
179 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
180 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
181 update_br_prob_note (cbranch_block);
183 /* Delete the block with the unconditional jump, and clean up the mess. */
184 delete_basic_block (jump_block);
185 tidy_fallthru_edge (cbranch_jump_edge);
186 update_forwarder_flag (cbranch_block);
188 return true;
191 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
192 on register. Used by jump threading. */
194 static bool
195 mark_effect (rtx exp, regset nonequal)
197 int regno;
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 if (REG_P (XEXP (exp, 0)))
206 dest = XEXP (exp, 0);
207 regno = REGNO (dest);
208 if (HARD_REGISTER_NUM_P (regno))
209 bitmap_clear_range (nonequal, regno,
210 hard_regno_nregs[regno][GET_MODE (dest)]);
211 else
212 bitmap_clear_bit (nonequal, regno);
214 return false;
216 case SET:
217 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
218 return false;
219 dest = SET_DEST (exp);
220 if (dest == pc_rtx)
221 return false;
222 if (!REG_P (dest))
223 return true;
224 regno = REGNO (dest);
225 if (HARD_REGISTER_NUM_P (regno))
226 bitmap_set_range (nonequal, regno,
227 hard_regno_nregs[regno][GET_MODE (dest)]);
228 else
229 bitmap_set_bit (nonequal, regno);
230 return false;
232 default:
233 return false;
237 /* Return nonzero if X is a register set in regset DATA.
238 Called via for_each_rtx. */
239 static int
240 mentions_nonequal_regs (rtx *x, void *data)
242 regset nonequal = (regset) data;
243 if (REG_P (*x))
245 int regno;
247 regno = REGNO (*x);
248 if (REGNO_REG_SET_P (nonequal, regno))
249 return 1;
250 if (regno < FIRST_PSEUDO_REGISTER)
252 int n = hard_regno_nregs[regno][GET_MODE (*x)];
253 while (--n > 0)
254 if (REGNO_REG_SET_P (nonequal, regno + n))
255 return 1;
258 return 0;
260 /* Attempt to prove that the basic block B will have no side effects and
261 always continues in the same edge if reached via E. Return the edge
262 if exist, NULL otherwise. */
264 static edge
265 thread_jump (edge e, basic_block b)
267 rtx set1, set2, cond1, cond2, insn;
268 enum rtx_code code1, code2, reversed_code2;
269 bool reverse1 = false;
270 unsigned i;
271 regset nonequal;
272 bool failed = false;
273 reg_set_iterator rsi;
275 if (b->flags & BB_NONTHREADABLE_BLOCK)
276 return NULL;
278 /* At the moment, we do handle only conditional jumps, but later we may
279 want to extend this code to tablejumps and others. */
280 if (EDGE_COUNT (e->src->succs) != 2)
281 return NULL;
282 if (EDGE_COUNT (b->succs) != 2)
284 b->flags |= BB_NONTHREADABLE_BLOCK;
285 return NULL;
288 /* Second branch must end with onlyjump, as we will eliminate the jump. */
289 if (!any_condjump_p (BB_END (e->src)))
290 return NULL;
292 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
294 b->flags |= BB_NONTHREADABLE_BLOCK;
295 return NULL;
298 set1 = pc_set (BB_END (e->src));
299 set2 = pc_set (BB_END (b));
300 if (((e->flags & EDGE_FALLTHRU) != 0)
301 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
302 reverse1 = true;
304 cond1 = XEXP (SET_SRC (set1), 0);
305 cond2 = XEXP (SET_SRC (set2), 0);
306 if (reverse1)
307 code1 = reversed_comparison_code (cond1, BB_END (e->src));
308 else
309 code1 = GET_CODE (cond1);
311 code2 = GET_CODE (cond2);
312 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
314 if (!comparison_dominates_p (code1, code2)
315 && !comparison_dominates_p (code1, reversed_code2))
316 return NULL;
318 /* Ensure that the comparison operators are equivalent.
319 ??? This is far too pessimistic. We should allow swapped operands,
320 different CCmodes, or for example comparisons for interval, that
321 dominate even when operands are not equivalent. */
322 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
323 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
324 return NULL;
326 /* Short circuit cases where block B contains some side effects, as we can't
327 safely bypass it. */
328 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
329 insn = NEXT_INSN (insn))
330 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
332 b->flags |= BB_NONTHREADABLE_BLOCK;
333 return NULL;
336 cselib_init (0);
338 /* First process all values computed in the source basic block. */
339 for (insn = NEXT_INSN (BB_HEAD (e->src));
340 insn != NEXT_INSN (BB_END (e->src));
341 insn = NEXT_INSN (insn))
342 if (INSN_P (insn))
343 cselib_process_insn (insn);
345 nonequal = BITMAP_ALLOC (NULL);
346 CLEAR_REG_SET (nonequal);
348 /* Now assume that we've continued by the edge E to B and continue
349 processing as if it were same basic block.
350 Our goal is to prove that whole block is an NOOP. */
352 for (insn = NEXT_INSN (BB_HEAD (b));
353 insn != NEXT_INSN (BB_END (b)) && !failed;
354 insn = NEXT_INSN (insn))
356 if (INSN_P (insn))
358 rtx pat = PATTERN (insn);
360 if (GET_CODE (pat) == PARALLEL)
362 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
363 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
365 else
366 failed |= mark_effect (pat, nonequal);
369 cselib_process_insn (insn);
372 /* Later we should clear nonequal of dead registers. So far we don't
373 have life information in cfg_cleanup. */
374 if (failed)
376 b->flags |= BB_NONTHREADABLE_BLOCK;
377 goto failed_exit;
380 /* cond2 must not mention any register that is not equal to the
381 former block. */
382 if (for_each_rtx (&cond2, mentions_nonequal_regs, nonequal))
383 goto failed_exit;
385 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
386 goto failed_exit;
388 BITMAP_FREE (nonequal);
389 cselib_finish ();
390 if ((comparison_dominates_p (code1, code2) != 0)
391 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
392 return BRANCH_EDGE (b);
393 else
394 return FALLTHRU_EDGE (b);
396 failed_exit:
397 BITMAP_FREE (nonequal);
398 cselib_finish ();
399 return NULL;
402 /* Attempt to forward edges leaving basic block B.
403 Return true if successful. */
405 static bool
406 try_forward_edges (int mode, basic_block b)
408 bool changed = false;
409 edge_iterator ei;
410 edge e, *threaded_edges = NULL;
412 /* If we are partitioning hot/cold basic blocks, we don't want to
413 mess up unconditional or indirect jumps that cross between hot
414 and cold sections.
416 Basic block partitioning may result in some jumps that appear to
417 be optimizable (or blocks that appear to be mergeable), but which really
418 must be left untouched (they are required to make it safely across
419 partition boundaries). See the comments at the top of
420 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
422 if (find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX))
423 return false;
425 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
427 basic_block target, first;
428 int counter, goto_locus;
429 bool threaded = false;
430 int nthreaded_edges = 0;
431 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
433 /* Skip complex edges because we don't know how to update them.
435 Still handle fallthru edges, as we can succeed to forward fallthru
436 edge to the same place as the branch edge of conditional branch
437 and turn conditional branch to an unconditional branch. */
438 if (e->flags & EDGE_COMPLEX)
440 ei_next (&ei);
441 continue;
444 target = first = e->dest;
445 counter = NUM_FIXED_BLOCKS;
446 goto_locus = e->goto_locus;
448 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
449 up jumps that cross between hot/cold sections.
451 Basic block partitioning may result in some jumps that appear
452 to be optimizable (or blocks that appear to be mergeable), but which
453 really must be left untouched (they are required to make it safely
454 across partition boundaries). See the comments at the top of
455 bb-reorder.c:partition_hot_cold_basic_blocks for complete
456 details. */
458 if (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
459 && find_reg_note (BB_END (first), REG_CROSSING_JUMP, NULL_RTX))
460 return changed;
462 while (counter < n_basic_blocks_for_fn (cfun))
464 basic_block new_target = NULL;
465 bool new_target_threaded = false;
466 may_thread |= (target->flags & BB_MODIFIED) != 0;
468 if (FORWARDER_BLOCK_P (target)
469 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
470 && single_succ (target) != EXIT_BLOCK_PTR_FOR_FN (cfun))
472 /* Bypass trivial infinite loops. */
473 new_target = single_succ (target);
474 if (target == new_target)
475 counter = n_basic_blocks_for_fn (cfun);
476 else if (!optimize)
478 /* When not optimizing, ensure that edges or forwarder
479 blocks with different locus are not optimized out. */
480 int new_locus = single_succ_edge (target)->goto_locus;
481 int locus = goto_locus;
483 if (new_locus != UNKNOWN_LOCATION
484 && locus != UNKNOWN_LOCATION
485 && new_locus != locus)
486 new_target = NULL;
487 else
489 rtx last;
491 if (new_locus != UNKNOWN_LOCATION)
492 locus = new_locus;
494 last = BB_END (target);
495 if (DEBUG_INSN_P (last))
496 last = prev_nondebug_insn (last);
498 new_locus = last && INSN_P (last)
499 ? INSN_LOCATION (last) : 0;
501 if (new_locus != UNKNOWN_LOCATION
502 && locus != UNKNOWN_LOCATION
503 && new_locus != locus)
504 new_target = NULL;
505 else
507 if (new_locus != UNKNOWN_LOCATION)
508 locus = new_locus;
510 goto_locus = locus;
516 /* Allow to thread only over one edge at time to simplify updating
517 of probabilities. */
518 else if ((mode & CLEANUP_THREADING) && may_thread)
520 edge t = thread_jump (e, target);
521 if (t)
523 if (!threaded_edges)
524 threaded_edges = XNEWVEC (edge,
525 n_basic_blocks_for_fn (cfun));
526 else
528 int i;
530 /* Detect an infinite loop across blocks not
531 including the start block. */
532 for (i = 0; i < nthreaded_edges; ++i)
533 if (threaded_edges[i] == t)
534 break;
535 if (i < nthreaded_edges)
537 counter = n_basic_blocks_for_fn (cfun);
538 break;
542 /* Detect an infinite loop across the start block. */
543 if (t->dest == b)
544 break;
546 gcc_assert (nthreaded_edges
547 < (n_basic_blocks_for_fn (cfun)
548 - NUM_FIXED_BLOCKS));
549 threaded_edges[nthreaded_edges++] = t;
551 new_target = t->dest;
552 new_target_threaded = true;
556 if (!new_target)
557 break;
559 counter++;
560 target = new_target;
561 threaded |= new_target_threaded;
564 if (counter >= n_basic_blocks_for_fn (cfun))
566 if (dump_file)
567 fprintf (dump_file, "Infinite loop in BB %i.\n",
568 target->index);
570 else if (target == first)
571 ; /* We didn't do anything. */
572 else
574 /* Save the values now, as the edge may get removed. */
575 gcov_type edge_count = e->count;
576 int edge_probability = e->probability;
577 int edge_frequency;
578 int n = 0;
580 e->goto_locus = goto_locus;
582 /* Don't force if target is exit block. */
583 if (threaded && target != EXIT_BLOCK_PTR_FOR_FN (cfun))
585 notice_new_block (redirect_edge_and_branch_force (e, target));
586 if (dump_file)
587 fprintf (dump_file, "Conditionals threaded.\n");
589 else if (!redirect_edge_and_branch (e, target))
591 if (dump_file)
592 fprintf (dump_file,
593 "Forwarding edge %i->%i to %i failed.\n",
594 b->index, e->dest->index, target->index);
595 ei_next (&ei);
596 continue;
599 /* We successfully forwarded the edge. Now update profile
600 data: for each edge we traversed in the chain, remove
601 the original edge's execution count. */
602 edge_frequency = apply_probability (b->frequency, edge_probability);
606 edge t;
608 if (!single_succ_p (first))
610 gcc_assert (n < nthreaded_edges);
611 t = threaded_edges [n++];
612 gcc_assert (t->src == first);
613 update_bb_profile_for_threading (first, edge_frequency,
614 edge_count, t);
615 update_br_prob_note (first);
617 else
619 first->count -= edge_count;
620 if (first->count < 0)
621 first->count = 0;
622 first->frequency -= edge_frequency;
623 if (first->frequency < 0)
624 first->frequency = 0;
625 /* It is possible that as the result of
626 threading we've removed edge as it is
627 threaded to the fallthru edge. Avoid
628 getting out of sync. */
629 if (n < nthreaded_edges
630 && first == threaded_edges [n]->src)
631 n++;
632 t = single_succ_edge (first);
635 t->count -= edge_count;
636 if (t->count < 0)
637 t->count = 0;
638 first = t->dest;
640 while (first != target);
642 changed = true;
643 continue;
645 ei_next (&ei);
648 free (threaded_edges);
649 return changed;
653 /* Blocks A and B are to be merged into a single block. A has no incoming
654 fallthru edge, so it can be moved before B without adding or modifying
655 any jumps (aside from the jump from A to B). */
657 static void
658 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
660 rtx barrier;
662 /* If we are partitioning hot/cold basic blocks, we don't want to
663 mess up unconditional or indirect jumps that cross between hot
664 and cold sections.
666 Basic block partitioning may result in some jumps that appear to
667 be optimizable (or blocks that appear to be mergeable), but which really
668 must be left untouched (they are required to make it safely across
669 partition boundaries). See the comments at the top of
670 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
672 if (BB_PARTITION (a) != BB_PARTITION (b))
673 return;
675 barrier = next_nonnote_insn (BB_END (a));
676 gcc_assert (BARRIER_P (barrier));
677 delete_insn (barrier);
679 /* Scramble the insn chain. */
680 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
681 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
682 df_set_bb_dirty (a);
684 if (dump_file)
685 fprintf (dump_file, "Moved block %d before %d and merged.\n",
686 a->index, b->index);
688 /* Swap the records for the two blocks around. */
690 unlink_block (a);
691 link_block (a, b->prev_bb);
693 /* Now blocks A and B are contiguous. Merge them. */
694 merge_blocks (a, b);
697 /* Blocks A and B are to be merged into a single block. B has no outgoing
698 fallthru edge, so it can be moved after A without adding or modifying
699 any jumps (aside from the jump from A to B). */
701 static void
702 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
704 rtx barrier, real_b_end;
705 rtx label, table;
707 /* If we are partitioning hot/cold basic blocks, we don't want to
708 mess up unconditional or indirect jumps that cross between hot
709 and cold sections.
711 Basic block partitioning may result in some jumps that appear to
712 be optimizable (or blocks that appear to be mergeable), but which really
713 must be left untouched (they are required to make it safely across
714 partition boundaries). See the comments at the top of
715 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
717 if (BB_PARTITION (a) != BB_PARTITION (b))
718 return;
720 real_b_end = BB_END (b);
722 /* If there is a jump table following block B temporarily add the jump table
723 to block B so that it will also be moved to the correct location. */
724 if (tablejump_p (BB_END (b), &label, &table)
725 && prev_active_insn (label) == BB_END (b))
727 BB_END (b) = table;
730 /* There had better have been a barrier there. Delete it. */
731 barrier = NEXT_INSN (BB_END (b));
732 if (barrier && BARRIER_P (barrier))
733 delete_insn (barrier);
736 /* Scramble the insn chain. */
737 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
739 /* Restore the real end of b. */
740 BB_END (b) = real_b_end;
742 if (dump_file)
743 fprintf (dump_file, "Moved block %d after %d and merged.\n",
744 b->index, a->index);
746 /* Now blocks A and B are contiguous. Merge them. */
747 merge_blocks (a, b);
750 /* Attempt to merge basic blocks that are potentially non-adjacent.
751 Return NULL iff the attempt failed, otherwise return basic block
752 where cleanup_cfg should continue. Because the merging commonly
753 moves basic block away or introduces another optimization
754 possibility, return basic block just before B so cleanup_cfg don't
755 need to iterate.
757 It may be good idea to return basic block before C in the case
758 C has been moved after B and originally appeared earlier in the
759 insn sequence, but we have no information available about the
760 relative ordering of these two. Hopefully it is not too common. */
762 static basic_block
763 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
765 basic_block next;
767 /* If we are partitioning hot/cold basic blocks, we don't want to
768 mess up unconditional or indirect jumps that cross between hot
769 and cold sections.
771 Basic block partitioning may result in some jumps that appear to
772 be optimizable (or blocks that appear to be mergeable), but which really
773 must be left untouched (they are required to make it safely across
774 partition boundaries). See the comments at the top of
775 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
777 if (BB_PARTITION (b) != BB_PARTITION (c))
778 return NULL;
780 /* If B has a fallthru edge to C, no need to move anything. */
781 if (e->flags & EDGE_FALLTHRU)
783 int b_index = b->index, c_index = c->index;
785 /* Protect the loop latches. */
786 if (current_loops && c->loop_father->latch == c)
787 return NULL;
789 merge_blocks (b, c);
790 update_forwarder_flag (b);
792 if (dump_file)
793 fprintf (dump_file, "Merged %d and %d without moving.\n",
794 b_index, c_index);
796 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
799 /* Otherwise we will need to move code around. Do that only if expensive
800 transformations are allowed. */
801 else if (mode & CLEANUP_EXPENSIVE)
803 edge tmp_edge, b_fallthru_edge;
804 bool c_has_outgoing_fallthru;
805 bool b_has_incoming_fallthru;
807 /* Avoid overactive code motion, as the forwarder blocks should be
808 eliminated by edge redirection instead. One exception might have
809 been if B is a forwarder block and C has no fallthru edge, but
810 that should be cleaned up by bb-reorder instead. */
811 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
812 return NULL;
814 /* We must make sure to not munge nesting of lexical blocks,
815 and loop notes. This is done by squeezing out all the notes
816 and leaving them there to lie. Not ideal, but functional. */
818 tmp_edge = find_fallthru_edge (c->succs);
819 c_has_outgoing_fallthru = (tmp_edge != NULL);
821 tmp_edge = find_fallthru_edge (b->preds);
822 b_has_incoming_fallthru = (tmp_edge != NULL);
823 b_fallthru_edge = tmp_edge;
824 next = b->prev_bb;
825 if (next == c)
826 next = next->prev_bb;
828 /* Otherwise, we're going to try to move C after B. If C does
829 not have an outgoing fallthru, then it can be moved
830 immediately after B without introducing or modifying jumps. */
831 if (! c_has_outgoing_fallthru)
833 merge_blocks_move_successor_nojumps (b, c);
834 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
837 /* If B does not have an incoming fallthru, then it can be moved
838 immediately before C without introducing or modifying jumps.
839 C cannot be the first block, so we do not have to worry about
840 accessing a non-existent block. */
842 if (b_has_incoming_fallthru)
844 basic_block bb;
846 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
847 return NULL;
848 bb = force_nonfallthru (b_fallthru_edge);
849 if (bb)
850 notice_new_block (bb);
853 merge_blocks_move_predecessor_nojumps (b, c);
854 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
857 return NULL;
861 /* Removes the memory attributes of MEM expression
862 if they are not equal. */
864 void
865 merge_memattrs (rtx x, rtx y)
867 int i;
868 int j;
869 enum rtx_code code;
870 const char *fmt;
872 if (x == y)
873 return;
874 if (x == 0 || y == 0)
875 return;
877 code = GET_CODE (x);
879 if (code != GET_CODE (y))
880 return;
882 if (GET_MODE (x) != GET_MODE (y))
883 return;
885 if (code == MEM && MEM_ATTRS (x) != MEM_ATTRS (y))
887 if (! MEM_ATTRS (x))
888 MEM_ATTRS (y) = 0;
889 else if (! MEM_ATTRS (y))
890 MEM_ATTRS (x) = 0;
891 else
893 HOST_WIDE_INT mem_size;
895 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
897 set_mem_alias_set (x, 0);
898 set_mem_alias_set (y, 0);
901 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
903 set_mem_expr (x, 0);
904 set_mem_expr (y, 0);
905 clear_mem_offset (x);
906 clear_mem_offset (y);
908 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
909 || (MEM_OFFSET_KNOWN_P (x)
910 && MEM_OFFSET (x) != MEM_OFFSET (y)))
912 clear_mem_offset (x);
913 clear_mem_offset (y);
916 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
918 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
919 set_mem_size (x, mem_size);
920 set_mem_size (y, mem_size);
922 else
924 clear_mem_size (x);
925 clear_mem_size (y);
928 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
929 set_mem_align (y, MEM_ALIGN (x));
932 if (code == MEM)
934 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
936 MEM_READONLY_P (x) = 0;
937 MEM_READONLY_P (y) = 0;
939 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
941 MEM_NOTRAP_P (x) = 0;
942 MEM_NOTRAP_P (y) = 0;
944 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
946 MEM_VOLATILE_P (x) = 1;
947 MEM_VOLATILE_P (y) = 1;
951 fmt = GET_RTX_FORMAT (code);
952 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
954 switch (fmt[i])
956 case 'E':
957 /* Two vectors must have the same length. */
958 if (XVECLEN (x, i) != XVECLEN (y, i))
959 return;
961 for (j = 0; j < XVECLEN (x, i); j++)
962 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
964 break;
966 case 'e':
967 merge_memattrs (XEXP (x, i), XEXP (y, i));
970 return;
974 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
975 different single sets S1 and S2. */
977 static bool
978 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
980 int i;
981 rtx e1, e2;
983 if (p1 == s1 && p2 == s2)
984 return true;
986 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
987 return false;
989 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
990 return false;
992 for (i = 0; i < XVECLEN (p1, 0); i++)
994 e1 = XVECEXP (p1, 0, i);
995 e2 = XVECEXP (p2, 0, i);
996 if (e1 == s1 && e2 == s2)
997 continue;
998 if (reload_completed
999 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
1000 continue;
1002 return false;
1005 return true;
1008 /* Examine register notes on I1 and I2 and return:
1009 - dir_forward if I1 can be replaced by I2, or
1010 - dir_backward if I2 can be replaced by I1, or
1011 - dir_both if both are the case. */
1013 static enum replace_direction
1014 can_replace_by (rtx i1, rtx i2)
1016 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1017 bool c1, c2;
1019 /* Check for 2 sets. */
1020 s1 = single_set (i1);
1021 s2 = single_set (i2);
1022 if (s1 == NULL_RTX || s2 == NULL_RTX)
1023 return dir_none;
1025 /* Check that the 2 sets set the same dest. */
1026 d1 = SET_DEST (s1);
1027 d2 = SET_DEST (s2);
1028 if (!(reload_completed
1029 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1030 return dir_none;
1032 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1033 set dest to the same value. */
1034 note1 = find_reg_equal_equiv_note (i1);
1035 note2 = find_reg_equal_equiv_note (i2);
1036 if (!note1 || !note2 || !rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0))
1037 || !CONST_INT_P (XEXP (note1, 0)))
1038 return dir_none;
1040 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1041 return dir_none;
1043 /* Although the 2 sets set dest to the same value, we cannot replace
1044 (set (dest) (const_int))
1046 (set (dest) (reg))
1047 because we don't know if the reg is live and has the same value at the
1048 location of replacement. */
1049 src1 = SET_SRC (s1);
1050 src2 = SET_SRC (s2);
1051 c1 = CONST_INT_P (src1);
1052 c2 = CONST_INT_P (src2);
1053 if (c1 && c2)
1054 return dir_both;
1055 else if (c2)
1056 return dir_forward;
1057 else if (c1)
1058 return dir_backward;
1060 return dir_none;
1063 /* Merges directions A and B. */
1065 static enum replace_direction
1066 merge_dir (enum replace_direction a, enum replace_direction b)
1068 /* Implements the following table:
1069 |bo fw bw no
1070 ---+-----------
1071 bo |bo fw bw no
1072 fw |-- fw no no
1073 bw |-- -- bw no
1074 no |-- -- -- no. */
1076 if (a == b)
1077 return a;
1079 if (a == dir_both)
1080 return b;
1081 if (b == dir_both)
1082 return a;
1084 return dir_none;
1087 /* Examine I1 and I2 and return:
1088 - dir_forward if I1 can be replaced by I2, or
1089 - dir_backward if I2 can be replaced by I1, or
1090 - dir_both if both are the case. */
1092 static enum replace_direction
1093 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx i1, rtx i2)
1095 rtx p1, p2;
1097 /* Verify that I1 and I2 are equivalent. */
1098 if (GET_CODE (i1) != GET_CODE (i2))
1099 return dir_none;
1101 /* __builtin_unreachable() may lead to empty blocks (ending with
1102 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1103 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1104 return dir_both;
1106 /* ??? Do not allow cross-jumping between different stack levels. */
1107 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1108 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1109 if (p1 && p2)
1111 p1 = XEXP (p1, 0);
1112 p2 = XEXP (p2, 0);
1113 if (!rtx_equal_p (p1, p2))
1114 return dir_none;
1116 /* ??? Worse, this adjustment had better be constant lest we
1117 have differing incoming stack levels. */
1118 if (!frame_pointer_needed
1119 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1120 return dir_none;
1122 else if (p1 || p2)
1123 return dir_none;
1125 p1 = PATTERN (i1);
1126 p2 = PATTERN (i2);
1128 if (GET_CODE (p1) != GET_CODE (p2))
1129 return dir_none;
1131 /* If this is a CALL_INSN, compare register usage information.
1132 If we don't check this on stack register machines, the two
1133 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1134 numbers of stack registers in the same basic block.
1135 If we don't check this on machines with delay slots, a delay slot may
1136 be filled that clobbers a parameter expected by the subroutine.
1138 ??? We take the simple route for now and assume that if they're
1139 equal, they were constructed identically.
1141 Also check for identical exception regions. */
1143 if (CALL_P (i1))
1145 /* Ensure the same EH region. */
1146 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1147 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1149 if (!n1 && n2)
1150 return dir_none;
1152 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1153 return dir_none;
1155 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1156 CALL_INSN_FUNCTION_USAGE (i2))
1157 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1158 return dir_none;
1160 /* For address sanitizer, never crossjump __asan_report_* builtins,
1161 otherwise errors might be reported on incorrect lines. */
1162 if (flag_sanitize & SANITIZE_ADDRESS)
1164 rtx call = get_call_rtx_from (i1);
1165 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1167 rtx symbol = XEXP (XEXP (call, 0), 0);
1168 if (SYMBOL_REF_DECL (symbol)
1169 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1171 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1172 == BUILT_IN_NORMAL)
1173 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1174 >= BUILT_IN_ASAN_REPORT_LOAD1
1175 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1176 <= BUILT_IN_ASAN_REPORT_STORE16)
1177 return dir_none;
1183 #ifdef STACK_REGS
1184 /* If cross_jump_death_matters is not 0, the insn's mode
1185 indicates whether or not the insn contains any stack-like
1186 regs. */
1188 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1190 /* If register stack conversion has already been done, then
1191 death notes must also be compared before it is certain that
1192 the two instruction streams match. */
1194 rtx note;
1195 HARD_REG_SET i1_regset, i2_regset;
1197 CLEAR_HARD_REG_SET (i1_regset);
1198 CLEAR_HARD_REG_SET (i2_regset);
1200 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1201 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1202 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1204 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1205 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1206 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1208 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1209 return dir_none;
1211 #endif
1213 if (reload_completed
1214 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1215 return dir_both;
1217 return can_replace_by (i1, i2);
1220 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1221 flow_find_head_matching_sequence, ensure the notes match. */
1223 static void
1224 merge_notes (rtx i1, rtx i2)
1226 /* If the merged insns have different REG_EQUAL notes, then
1227 remove them. */
1228 rtx equiv1 = find_reg_equal_equiv_note (i1);
1229 rtx equiv2 = find_reg_equal_equiv_note (i2);
1231 if (equiv1 && !equiv2)
1232 remove_note (i1, equiv1);
1233 else if (!equiv1 && equiv2)
1234 remove_note (i2, equiv2);
1235 else if (equiv1 && equiv2
1236 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1238 remove_note (i1, equiv1);
1239 remove_note (i2, equiv2);
1243 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1244 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1245 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1246 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1247 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1249 static void
1250 walk_to_nondebug_insn (rtx *i1, basic_block *bb1, bool follow_fallthru,
1251 bool *did_fallthru)
1253 edge fallthru;
1255 *did_fallthru = false;
1257 /* Ignore notes. */
1258 while (!NONDEBUG_INSN_P (*i1))
1260 if (*i1 != BB_HEAD (*bb1))
1262 *i1 = PREV_INSN (*i1);
1263 continue;
1266 if (!follow_fallthru)
1267 return;
1269 fallthru = find_fallthru_edge ((*bb1)->preds);
1270 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1271 || !single_succ_p (fallthru->src))
1272 return;
1274 *bb1 = fallthru->src;
1275 *i1 = BB_END (*bb1);
1276 *did_fallthru = true;
1280 /* Look through the insns at the end of BB1 and BB2 and find the longest
1281 sequence that are either equivalent, or allow forward or backward
1282 replacement. Store the first insns for that sequence in *F1 and *F2 and
1283 return the sequence length.
1285 DIR_P indicates the allowed replacement direction on function entry, and
1286 the actual replacement direction on function exit. If NULL, only equivalent
1287 sequences are allowed.
1289 To simplify callers of this function, if the blocks match exactly,
1290 store the head of the blocks in *F1 and *F2. */
1293 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx *f1, rtx *f2,
1294 enum replace_direction *dir_p)
1296 rtx i1, i2, last1, last2, afterlast1, afterlast2;
1297 int ninsns = 0;
1298 enum replace_direction dir, last_dir, afterlast_dir;
1299 bool follow_fallthru, did_fallthru;
1301 if (dir_p)
1302 dir = *dir_p;
1303 else
1304 dir = dir_both;
1305 afterlast_dir = dir;
1306 last_dir = afterlast_dir;
1308 /* Skip simple jumps at the end of the blocks. Complex jumps still
1309 need to be compared for equivalence, which we'll do below. */
1311 i1 = BB_END (bb1);
1312 last1 = afterlast1 = last2 = afterlast2 = NULL_RTX;
1313 if (onlyjump_p (i1)
1314 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1316 last1 = i1;
1317 i1 = PREV_INSN (i1);
1320 i2 = BB_END (bb2);
1321 if (onlyjump_p (i2)
1322 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1324 last2 = i2;
1325 /* Count everything except for unconditional jump as insn.
1326 Don't count any jumps if dir_p is NULL. */
1327 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1328 ninsns++;
1329 i2 = PREV_INSN (i2);
1332 while (true)
1334 /* In the following example, we can replace all jumps to C by jumps to A.
1336 This removes 4 duplicate insns.
1337 [bb A] insn1 [bb C] insn1
1338 insn2 insn2
1339 [bb B] insn3 insn3
1340 insn4 insn4
1341 jump_insn jump_insn
1343 We could also replace all jumps to A by jumps to C, but that leaves B
1344 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1345 step, all jumps to B would be replaced with jumps to the middle of C,
1346 achieving the same result with more effort.
1347 So we allow only the first possibility, which means that we don't allow
1348 fallthru in the block that's being replaced. */
1350 follow_fallthru = dir_p && dir != dir_forward;
1351 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1352 if (did_fallthru)
1353 dir = dir_backward;
1355 follow_fallthru = dir_p && dir != dir_backward;
1356 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1357 if (did_fallthru)
1358 dir = dir_forward;
1360 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1361 break;
1363 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1364 if (dir == dir_none || (!dir_p && dir != dir_both))
1365 break;
1367 merge_memattrs (i1, i2);
1369 /* Don't begin a cross-jump with a NOTE insn. */
1370 if (INSN_P (i1))
1372 merge_notes (i1, i2);
1374 afterlast1 = last1, afterlast2 = last2;
1375 last1 = i1, last2 = i2;
1376 afterlast_dir = last_dir;
1377 last_dir = dir;
1378 if (active_insn_p (i1))
1379 ninsns++;
1382 i1 = PREV_INSN (i1);
1383 i2 = PREV_INSN (i2);
1386 #ifdef HAVE_cc0
1387 /* Don't allow the insn after a compare to be shared by
1388 cross-jumping unless the compare is also shared. */
1389 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
1390 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1391 #endif
1393 /* Include preceding notes and labels in the cross-jump. One,
1394 this may bring us to the head of the blocks as requested above.
1395 Two, it keeps line number notes as matched as may be. */
1396 if (ninsns)
1398 bb1 = BLOCK_FOR_INSN (last1);
1399 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1400 last1 = PREV_INSN (last1);
1402 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1403 last1 = PREV_INSN (last1);
1405 bb2 = BLOCK_FOR_INSN (last2);
1406 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1407 last2 = PREV_INSN (last2);
1409 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1410 last2 = PREV_INSN (last2);
1412 *f1 = last1;
1413 *f2 = last2;
1416 if (dir_p)
1417 *dir_p = last_dir;
1418 return ninsns;
1421 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1422 the head of the two blocks. Do not include jumps at the end.
1423 If STOP_AFTER is nonzero, stop after finding that many matching
1424 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1425 non-zero, only count active insns. */
1428 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx *f1,
1429 rtx *f2, int stop_after)
1431 rtx i1, i2, last1, last2, beforelast1, beforelast2;
1432 int ninsns = 0;
1433 edge e;
1434 edge_iterator ei;
1435 int nehedges1 = 0, nehedges2 = 0;
1437 FOR_EACH_EDGE (e, ei, bb1->succs)
1438 if (e->flags & EDGE_EH)
1439 nehedges1++;
1440 FOR_EACH_EDGE (e, ei, bb2->succs)
1441 if (e->flags & EDGE_EH)
1442 nehedges2++;
1444 i1 = BB_HEAD (bb1);
1445 i2 = BB_HEAD (bb2);
1446 last1 = beforelast1 = last2 = beforelast2 = NULL_RTX;
1448 while (true)
1450 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1451 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1453 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1454 break;
1455 i1 = NEXT_INSN (i1);
1458 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1460 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1461 break;
1462 i2 = NEXT_INSN (i2);
1465 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1466 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1467 break;
1469 if (NOTE_P (i1) || NOTE_P (i2)
1470 || JUMP_P (i1) || JUMP_P (i2))
1471 break;
1473 /* A sanity check to make sure we're not merging insns with different
1474 effects on EH. If only one of them ends a basic block, it shouldn't
1475 have an EH edge; if both end a basic block, there should be the same
1476 number of EH edges. */
1477 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1478 && nehedges1 > 0)
1479 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1480 && nehedges2 > 0)
1481 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1482 && nehedges1 != nehedges2))
1483 break;
1485 if (old_insns_match_p (0, i1, i2) != dir_both)
1486 break;
1488 merge_memattrs (i1, i2);
1490 /* Don't begin a cross-jump with a NOTE insn. */
1491 if (INSN_P (i1))
1493 merge_notes (i1, i2);
1495 beforelast1 = last1, beforelast2 = last2;
1496 last1 = i1, last2 = i2;
1497 if (!stop_after || active_insn_p (i1))
1498 ninsns++;
1501 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1502 || (stop_after > 0 && ninsns == stop_after))
1503 break;
1505 i1 = NEXT_INSN (i1);
1506 i2 = NEXT_INSN (i2);
1509 #ifdef HAVE_cc0
1510 /* Don't allow a compare to be shared by cross-jumping unless the insn
1511 after the compare is also shared. */
1512 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && sets_cc0_p (last1))
1513 last1 = beforelast1, last2 = beforelast2, ninsns--;
1514 #endif
1516 if (ninsns)
1518 *f1 = last1;
1519 *f2 = last2;
1522 return ninsns;
1525 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1526 the branch instruction. This means that if we commonize the control
1527 flow before end of the basic block, the semantic remains unchanged.
1529 We may assume that there exists one edge with a common destination. */
1531 static bool
1532 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1534 int nehedges1 = 0, nehedges2 = 0;
1535 edge fallthru1 = 0, fallthru2 = 0;
1536 edge e1, e2;
1537 edge_iterator ei;
1539 /* If we performed shrink-wrapping, edges to the exit block can
1540 only be distinguished for JUMP_INSNs. The two paths may differ in
1541 whether they went through the prologue. Sibcalls are fine, we know
1542 that we either didn't need or inserted an epilogue before them. */
1543 if (crtl->shrink_wrapped
1544 && single_succ_p (bb1)
1545 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1546 && !JUMP_P (BB_END (bb1))
1547 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1548 return false;
1550 /* If BB1 has only one successor, we may be looking at either an
1551 unconditional jump, or a fake edge to exit. */
1552 if (single_succ_p (bb1)
1553 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1554 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1555 return (single_succ_p (bb2)
1556 && (single_succ_edge (bb2)->flags
1557 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1558 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1560 /* Match conditional jumps - this may get tricky when fallthru and branch
1561 edges are crossed. */
1562 if (EDGE_COUNT (bb1->succs) == 2
1563 && any_condjump_p (BB_END (bb1))
1564 && onlyjump_p (BB_END (bb1)))
1566 edge b1, f1, b2, f2;
1567 bool reverse, match;
1568 rtx set1, set2, cond1, cond2;
1569 enum rtx_code code1, code2;
1571 if (EDGE_COUNT (bb2->succs) != 2
1572 || !any_condjump_p (BB_END (bb2))
1573 || !onlyjump_p (BB_END (bb2)))
1574 return false;
1576 b1 = BRANCH_EDGE (bb1);
1577 b2 = BRANCH_EDGE (bb2);
1578 f1 = FALLTHRU_EDGE (bb1);
1579 f2 = FALLTHRU_EDGE (bb2);
1581 /* Get around possible forwarders on fallthru edges. Other cases
1582 should be optimized out already. */
1583 if (FORWARDER_BLOCK_P (f1->dest))
1584 f1 = single_succ_edge (f1->dest);
1586 if (FORWARDER_BLOCK_P (f2->dest))
1587 f2 = single_succ_edge (f2->dest);
1589 /* To simplify use of this function, return false if there are
1590 unneeded forwarder blocks. These will get eliminated later
1591 during cleanup_cfg. */
1592 if (FORWARDER_BLOCK_P (f1->dest)
1593 || FORWARDER_BLOCK_P (f2->dest)
1594 || FORWARDER_BLOCK_P (b1->dest)
1595 || FORWARDER_BLOCK_P (b2->dest))
1596 return false;
1598 if (f1->dest == f2->dest && b1->dest == b2->dest)
1599 reverse = false;
1600 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1601 reverse = true;
1602 else
1603 return false;
1605 set1 = pc_set (BB_END (bb1));
1606 set2 = pc_set (BB_END (bb2));
1607 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1608 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1609 reverse = !reverse;
1611 cond1 = XEXP (SET_SRC (set1), 0);
1612 cond2 = XEXP (SET_SRC (set2), 0);
1613 code1 = GET_CODE (cond1);
1614 if (reverse)
1615 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1616 else
1617 code2 = GET_CODE (cond2);
1619 if (code2 == UNKNOWN)
1620 return false;
1622 /* Verify codes and operands match. */
1623 match = ((code1 == code2
1624 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1625 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1626 || (code1 == swap_condition (code2)
1627 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1628 XEXP (cond2, 0))
1629 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1630 XEXP (cond2, 1))));
1632 /* If we return true, we will join the blocks. Which means that
1633 we will only have one branch prediction bit to work with. Thus
1634 we require the existing branches to have probabilities that are
1635 roughly similar. */
1636 if (match
1637 && optimize_bb_for_speed_p (bb1)
1638 && optimize_bb_for_speed_p (bb2))
1640 int prob2;
1642 if (b1->dest == b2->dest)
1643 prob2 = b2->probability;
1644 else
1645 /* Do not use f2 probability as f2 may be forwarded. */
1646 prob2 = REG_BR_PROB_BASE - b2->probability;
1648 /* Fail if the difference in probabilities is greater than 50%.
1649 This rules out two well-predicted branches with opposite
1650 outcomes. */
1651 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1653 if (dump_file)
1654 fprintf (dump_file,
1655 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1656 bb1->index, bb2->index, b1->probability, prob2);
1658 return false;
1662 if (dump_file && match)
1663 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1664 bb1->index, bb2->index);
1666 return match;
1669 /* Generic case - we are seeing a computed jump, table jump or trapping
1670 instruction. */
1672 /* Check whether there are tablejumps in the end of BB1 and BB2.
1673 Return true if they are identical. */
1675 rtx label1, label2;
1676 rtx table1, table2;
1678 if (tablejump_p (BB_END (bb1), &label1, &table1)
1679 && tablejump_p (BB_END (bb2), &label2, &table2)
1680 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1682 /* The labels should never be the same rtx. If they really are same
1683 the jump tables are same too. So disable crossjumping of blocks BB1
1684 and BB2 because when deleting the common insns in the end of BB1
1685 by delete_basic_block () the jump table would be deleted too. */
1686 /* If LABEL2 is referenced in BB1->END do not do anything
1687 because we would loose information when replacing
1688 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1689 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1691 /* Set IDENTICAL to true when the tables are identical. */
1692 bool identical = false;
1693 rtx p1, p2;
1695 p1 = PATTERN (table1);
1696 p2 = PATTERN (table2);
1697 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1699 identical = true;
1701 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1702 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1703 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1704 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1706 int i;
1708 identical = true;
1709 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1710 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1711 identical = false;
1714 if (identical)
1716 replace_label_data rr;
1717 bool match;
1719 /* Temporarily replace references to LABEL1 with LABEL2
1720 in BB1->END so that we could compare the instructions. */
1721 rr.r1 = label1;
1722 rr.r2 = label2;
1723 rr.update_label_nuses = false;
1724 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1726 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1727 == dir_both);
1728 if (dump_file && match)
1729 fprintf (dump_file,
1730 "Tablejumps in bb %i and %i match.\n",
1731 bb1->index, bb2->index);
1733 /* Set the original label in BB1->END because when deleting
1734 a block whose end is a tablejump, the tablejump referenced
1735 from the instruction is deleted too. */
1736 rr.r1 = label2;
1737 rr.r2 = label1;
1738 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1740 return match;
1743 return false;
1747 /* Find the last non-debug non-note instruction in each bb, except
1748 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1749 handles that case specially. old_insns_match_p does not handle
1750 other types of instruction notes. */
1751 rtx last1 = BB_END (bb1);
1752 rtx last2 = BB_END (bb2);
1753 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1754 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1755 last1 = PREV_INSN (last1);
1756 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1757 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1758 last2 = PREV_INSN (last2);
1759 gcc_assert (last1 && last2);
1761 /* First ensure that the instructions match. There may be many outgoing
1762 edges so this test is generally cheaper. */
1763 if (old_insns_match_p (mode, last1, last2) != dir_both)
1764 return false;
1766 /* Search the outgoing edges, ensure that the counts do match, find possible
1767 fallthru and exception handling edges since these needs more
1768 validation. */
1769 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1770 return false;
1772 bool nonfakeedges = false;
1773 FOR_EACH_EDGE (e1, ei, bb1->succs)
1775 e2 = EDGE_SUCC (bb2, ei.index);
1777 if ((e1->flags & EDGE_FAKE) == 0)
1778 nonfakeedges = true;
1780 if (e1->flags & EDGE_EH)
1781 nehedges1++;
1783 if (e2->flags & EDGE_EH)
1784 nehedges2++;
1786 if (e1->flags & EDGE_FALLTHRU)
1787 fallthru1 = e1;
1788 if (e2->flags & EDGE_FALLTHRU)
1789 fallthru2 = e2;
1792 /* If number of edges of various types does not match, fail. */
1793 if (nehedges1 != nehedges2
1794 || (fallthru1 != 0) != (fallthru2 != 0))
1795 return false;
1797 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1798 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1799 attempt to optimize, as the two basic blocks might have different
1800 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1801 traps there should be REG_ARG_SIZE notes, they could be missing
1802 for __builtin_unreachable () uses though. */
1803 if (!nonfakeedges
1804 && !ACCUMULATE_OUTGOING_ARGS
1805 && (!INSN_P (last1)
1806 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1807 return false;
1809 /* fallthru edges must be forwarded to the same destination. */
1810 if (fallthru1)
1812 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1813 ? single_succ (fallthru1->dest): fallthru1->dest);
1814 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1815 ? single_succ (fallthru2->dest): fallthru2->dest);
1817 if (d1 != d2)
1818 return false;
1821 /* Ensure the same EH region. */
1823 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1824 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1826 if (!n1 && n2)
1827 return false;
1829 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1830 return false;
1833 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1834 version of sequence abstraction. */
1835 FOR_EACH_EDGE (e1, ei, bb2->succs)
1837 edge e2;
1838 edge_iterator ei;
1839 basic_block d1 = e1->dest;
1841 if (FORWARDER_BLOCK_P (d1))
1842 d1 = EDGE_SUCC (d1, 0)->dest;
1844 FOR_EACH_EDGE (e2, ei, bb1->succs)
1846 basic_block d2 = e2->dest;
1847 if (FORWARDER_BLOCK_P (d2))
1848 d2 = EDGE_SUCC (d2, 0)->dest;
1849 if (d1 == d2)
1850 break;
1853 if (!e2)
1854 return false;
1857 return true;
1860 /* Returns true if BB basic block has a preserve label. */
1862 static bool
1863 block_has_preserve_label (basic_block bb)
1865 return (bb
1866 && block_label (bb)
1867 && LABEL_PRESERVE_P (block_label (bb)));
1870 /* E1 and E2 are edges with the same destination block. Search their
1871 predecessors for common code. If found, redirect control flow from
1872 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1873 or the other way around (dir_backward). DIR specifies the allowed
1874 replacement direction. */
1876 static bool
1877 try_crossjump_to_edge (int mode, edge e1, edge e2,
1878 enum replace_direction dir)
1880 int nmatch;
1881 basic_block src1 = e1->src, src2 = e2->src;
1882 basic_block redirect_to, redirect_from, to_remove;
1883 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1884 rtx newpos1, newpos2;
1885 edge s;
1886 edge_iterator ei;
1888 newpos1 = newpos2 = NULL_RTX;
1890 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1891 to try this optimization.
1893 Basic block partitioning may result in some jumps that appear to
1894 be optimizable (or blocks that appear to be mergeable), but which really
1895 must be left untouched (they are required to make it safely across
1896 partition boundaries). See the comments at the top of
1897 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1899 if (crtl->has_bb_partition && reload_completed)
1900 return false;
1902 /* Search backward through forwarder blocks. We don't need to worry
1903 about multiple entry or chained forwarders, as they will be optimized
1904 away. We do this to look past the unconditional jump following a
1905 conditional jump that is required due to the current CFG shape. */
1906 if (single_pred_p (src1)
1907 && FORWARDER_BLOCK_P (src1))
1908 e1 = single_pred_edge (src1), src1 = e1->src;
1910 if (single_pred_p (src2)
1911 && FORWARDER_BLOCK_P (src2))
1912 e2 = single_pred_edge (src2), src2 = e2->src;
1914 /* Nothing to do if we reach ENTRY, or a common source block. */
1915 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1916 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1917 return false;
1918 if (src1 == src2)
1919 return false;
1921 /* Seeing more than 1 forwarder blocks would confuse us later... */
1922 if (FORWARDER_BLOCK_P (e1->dest)
1923 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1924 return false;
1926 if (FORWARDER_BLOCK_P (e2->dest)
1927 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1928 return false;
1930 /* Likewise with dead code (possibly newly created by the other optimizations
1931 of cfg_cleanup). */
1932 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1933 return false;
1935 /* Look for the common insn sequence, part the first ... */
1936 if (!outgoing_edges_match (mode, src1, src2))
1937 return false;
1939 /* ... and part the second. */
1940 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1942 osrc1 = src1;
1943 osrc2 = src2;
1944 if (newpos1 != NULL_RTX)
1945 src1 = BLOCK_FOR_INSN (newpos1);
1946 if (newpos2 != NULL_RTX)
1947 src2 = BLOCK_FOR_INSN (newpos2);
1949 if (dir == dir_backward)
1951 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1952 SWAP (basic_block, osrc1, osrc2);
1953 SWAP (basic_block, src1, src2);
1954 SWAP (edge, e1, e2);
1955 SWAP (rtx, newpos1, newpos2);
1956 #undef SWAP
1959 /* Don't proceed with the crossjump unless we found a sufficient number
1960 of matching instructions or the 'from' block was totally matched
1961 (such that its predecessors will hopefully be redirected and the
1962 block removed). */
1963 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
1964 && (newpos1 != BB_HEAD (src1)))
1965 return false;
1967 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1968 if (block_has_preserve_label (e1->dest)
1969 && (e1->flags & EDGE_ABNORMAL))
1970 return false;
1972 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1973 will be deleted.
1974 If we have tablejumps in the end of SRC1 and SRC2
1975 they have been already compared for equivalence in outgoing_edges_match ()
1976 so replace the references to TABLE1 by references to TABLE2. */
1978 rtx label1, label2;
1979 rtx table1, table2;
1981 if (tablejump_p (BB_END (osrc1), &label1, &table1)
1982 && tablejump_p (BB_END (osrc2), &label2, &table2)
1983 && label1 != label2)
1985 replace_label_data rr;
1986 rtx insn;
1988 /* Replace references to LABEL1 with LABEL2. */
1989 rr.r1 = label1;
1990 rr.r2 = label2;
1991 rr.update_label_nuses = true;
1992 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
1994 /* Do not replace the label in SRC1->END because when deleting
1995 a block whose end is a tablejump, the tablejump referenced
1996 from the instruction is deleted too. */
1997 if (insn != BB_END (osrc1))
1998 for_each_rtx (&insn, replace_label, &rr);
2003 /* Avoid splitting if possible. We must always split when SRC2 has
2004 EH predecessor edges, or we may end up with basic blocks with both
2005 normal and EH predecessor edges. */
2006 if (newpos2 == BB_HEAD (src2)
2007 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2008 redirect_to = src2;
2009 else
2011 if (newpos2 == BB_HEAD (src2))
2013 /* Skip possible basic block header. */
2014 if (LABEL_P (newpos2))
2015 newpos2 = NEXT_INSN (newpos2);
2016 while (DEBUG_INSN_P (newpos2))
2017 newpos2 = NEXT_INSN (newpos2);
2018 if (NOTE_P (newpos2))
2019 newpos2 = NEXT_INSN (newpos2);
2020 while (DEBUG_INSN_P (newpos2))
2021 newpos2 = NEXT_INSN (newpos2);
2024 if (dump_file)
2025 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2026 src2->index, nmatch);
2027 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2030 if (dump_file)
2031 fprintf (dump_file,
2032 "Cross jumping from bb %i to bb %i; %i common insns\n",
2033 src1->index, src2->index, nmatch);
2035 /* We may have some registers visible through the block. */
2036 df_set_bb_dirty (redirect_to);
2038 if (osrc2 == src2)
2039 redirect_edges_to = redirect_to;
2040 else
2041 redirect_edges_to = osrc2;
2043 /* Recompute the frequencies and counts of outgoing edges. */
2044 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2046 edge s2;
2047 edge_iterator ei;
2048 basic_block d = s->dest;
2050 if (FORWARDER_BLOCK_P (d))
2051 d = single_succ (d);
2053 FOR_EACH_EDGE (s2, ei, src1->succs)
2055 basic_block d2 = s2->dest;
2056 if (FORWARDER_BLOCK_P (d2))
2057 d2 = single_succ (d2);
2058 if (d == d2)
2059 break;
2062 s->count += s2->count;
2064 /* Take care to update possible forwarder blocks. We verified
2065 that there is no more than one in the chain, so we can't run
2066 into infinite loop. */
2067 if (FORWARDER_BLOCK_P (s->dest))
2069 single_succ_edge (s->dest)->count += s2->count;
2070 s->dest->count += s2->count;
2071 s->dest->frequency += EDGE_FREQUENCY (s);
2074 if (FORWARDER_BLOCK_P (s2->dest))
2076 single_succ_edge (s2->dest)->count -= s2->count;
2077 if (single_succ_edge (s2->dest)->count < 0)
2078 single_succ_edge (s2->dest)->count = 0;
2079 s2->dest->count -= s2->count;
2080 s2->dest->frequency -= EDGE_FREQUENCY (s);
2081 if (s2->dest->frequency < 0)
2082 s2->dest->frequency = 0;
2083 if (s2->dest->count < 0)
2084 s2->dest->count = 0;
2087 if (!redirect_edges_to->frequency && !src1->frequency)
2088 s->probability = (s->probability + s2->probability) / 2;
2089 else
2090 s->probability
2091 = ((s->probability * redirect_edges_to->frequency +
2092 s2->probability * src1->frequency)
2093 / (redirect_edges_to->frequency + src1->frequency));
2096 /* Adjust count and frequency for the block. An earlier jump
2097 threading pass may have left the profile in an inconsistent
2098 state (see update_bb_profile_for_threading) so we must be
2099 prepared for overflows. */
2100 tmp = redirect_to;
2103 tmp->count += src1->count;
2104 tmp->frequency += src1->frequency;
2105 if (tmp->frequency > BB_FREQ_MAX)
2106 tmp->frequency = BB_FREQ_MAX;
2107 if (tmp == redirect_edges_to)
2108 break;
2109 tmp = find_fallthru_edge (tmp->succs)->dest;
2111 while (true);
2112 update_br_prob_note (redirect_edges_to);
2114 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2116 /* Skip possible basic block header. */
2117 if (LABEL_P (newpos1))
2118 newpos1 = NEXT_INSN (newpos1);
2120 while (DEBUG_INSN_P (newpos1))
2121 newpos1 = NEXT_INSN (newpos1);
2123 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2124 newpos1 = NEXT_INSN (newpos1);
2126 while (DEBUG_INSN_P (newpos1))
2127 newpos1 = NEXT_INSN (newpos1);
2129 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2130 to_remove = single_succ (redirect_from);
2132 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2133 delete_basic_block (to_remove);
2135 update_forwarder_flag (redirect_from);
2136 if (redirect_to != src2)
2137 update_forwarder_flag (src2);
2139 return true;
2142 /* Search the predecessors of BB for common insn sequences. When found,
2143 share code between them by redirecting control flow. Return true if
2144 any changes made. */
2146 static bool
2147 try_crossjump_bb (int mode, basic_block bb)
2149 edge e, e2, fallthru;
2150 bool changed;
2151 unsigned max, ix, ix2;
2153 /* Nothing to do if there is not at least two incoming edges. */
2154 if (EDGE_COUNT (bb->preds) < 2)
2155 return false;
2157 /* Don't crossjump if this block ends in a computed jump,
2158 unless we are optimizing for size. */
2159 if (optimize_bb_for_size_p (bb)
2160 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2161 && computed_jump_p (BB_END (bb)))
2162 return false;
2164 /* If we are partitioning hot/cold basic blocks, we don't want to
2165 mess up unconditional or indirect jumps that cross between hot
2166 and cold sections.
2168 Basic block partitioning may result in some jumps that appear to
2169 be optimizable (or blocks that appear to be mergeable), but which really
2170 must be left untouched (they are required to make it safely across
2171 partition boundaries). See the comments at the top of
2172 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2174 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2175 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2176 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2177 return false;
2179 /* It is always cheapest to redirect a block that ends in a branch to
2180 a block that falls through into BB, as that adds no branches to the
2181 program. We'll try that combination first. */
2182 fallthru = NULL;
2183 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2185 if (EDGE_COUNT (bb->preds) > max)
2186 return false;
2188 fallthru = find_fallthru_edge (bb->preds);
2190 changed = false;
2191 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2193 e = EDGE_PRED (bb, ix);
2194 ix++;
2196 /* As noted above, first try with the fallthru predecessor (or, a
2197 fallthru predecessor if we are in cfglayout mode). */
2198 if (fallthru)
2200 /* Don't combine the fallthru edge into anything else.
2201 If there is a match, we'll do it the other way around. */
2202 if (e == fallthru)
2203 continue;
2204 /* If nothing changed since the last attempt, there is nothing
2205 we can do. */
2206 if (!first_pass
2207 && !((e->src->flags & BB_MODIFIED)
2208 || (fallthru->src->flags & BB_MODIFIED)))
2209 continue;
2211 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2213 changed = true;
2214 ix = 0;
2215 continue;
2219 /* Non-obvious work limiting check: Recognize that we're going
2220 to call try_crossjump_bb on every basic block. So if we have
2221 two blocks with lots of outgoing edges (a switch) and they
2222 share lots of common destinations, then we would do the
2223 cross-jump check once for each common destination.
2225 Now, if the blocks actually are cross-jump candidates, then
2226 all of their destinations will be shared. Which means that
2227 we only need check them for cross-jump candidacy once. We
2228 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2229 choosing to do the check from the block for which the edge
2230 in question is the first successor of A. */
2231 if (EDGE_SUCC (e->src, 0) != e)
2232 continue;
2234 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2236 e2 = EDGE_PRED (bb, ix2);
2238 if (e2 == e)
2239 continue;
2241 /* We've already checked the fallthru edge above. */
2242 if (e2 == fallthru)
2243 continue;
2245 /* The "first successor" check above only prevents multiple
2246 checks of crossjump(A,B). In order to prevent redundant
2247 checks of crossjump(B,A), require that A be the block
2248 with the lowest index. */
2249 if (e->src->index > e2->src->index)
2250 continue;
2252 /* If nothing changed since the last attempt, there is nothing
2253 we can do. */
2254 if (!first_pass
2255 && !((e->src->flags & BB_MODIFIED)
2256 || (e2->src->flags & BB_MODIFIED)))
2257 continue;
2259 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2260 direction. */
2261 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2263 changed = true;
2264 ix = 0;
2265 break;
2270 if (changed)
2271 crossjumps_occured = true;
2273 return changed;
2276 /* Search the successors of BB for common insn sequences. When found,
2277 share code between them by moving it across the basic block
2278 boundary. Return true if any changes made. */
2280 static bool
2281 try_head_merge_bb (basic_block bb)
2283 basic_block final_dest_bb = NULL;
2284 int max_match = INT_MAX;
2285 edge e0;
2286 rtx *headptr, *currptr, *nextptr;
2287 bool changed, moveall;
2288 unsigned ix;
2289 rtx e0_last_head, cond, move_before;
2290 unsigned nedges = EDGE_COUNT (bb->succs);
2291 rtx jump = BB_END (bb);
2292 regset live, live_union;
2294 /* Nothing to do if there is not at least two outgoing edges. */
2295 if (nedges < 2)
2296 return false;
2298 /* Don't crossjump if this block ends in a computed jump,
2299 unless we are optimizing for size. */
2300 if (optimize_bb_for_size_p (bb)
2301 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2302 && computed_jump_p (BB_END (bb)))
2303 return false;
2305 cond = get_condition (jump, &move_before, true, false);
2306 if (cond == NULL_RTX)
2308 #ifdef HAVE_cc0
2309 if (reg_mentioned_p (cc0_rtx, jump))
2310 move_before = prev_nonnote_nondebug_insn (jump);
2311 else
2312 #endif
2313 move_before = jump;
2316 for (ix = 0; ix < nedges; ix++)
2317 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2318 return false;
2320 for (ix = 0; ix < nedges; ix++)
2322 edge e = EDGE_SUCC (bb, ix);
2323 basic_block other_bb = e->dest;
2325 if (df_get_bb_dirty (other_bb))
2327 block_was_dirty = true;
2328 return false;
2331 if (e->flags & EDGE_ABNORMAL)
2332 return false;
2334 /* Normally, all destination blocks must only be reachable from this
2335 block, i.e. they must have one incoming edge.
2337 There is one special case we can handle, that of multiple consecutive
2338 jumps where the first jumps to one of the targets of the second jump.
2339 This happens frequently in switch statements for default labels.
2340 The structure is as follows:
2341 FINAL_DEST_BB
2342 ....
2343 if (cond) jump A;
2344 fall through
2346 jump with targets A, B, C, D...
2348 has two incoming edges, from FINAL_DEST_BB and BB
2350 In this case, we can try to move the insns through BB and into
2351 FINAL_DEST_BB. */
2352 if (EDGE_COUNT (other_bb->preds) != 1)
2354 edge incoming_edge, incoming_bb_other_edge;
2355 edge_iterator ei;
2357 if (final_dest_bb != NULL
2358 || EDGE_COUNT (other_bb->preds) != 2)
2359 return false;
2361 /* We must be able to move the insns across the whole block. */
2362 move_before = BB_HEAD (bb);
2363 while (!NONDEBUG_INSN_P (move_before))
2364 move_before = NEXT_INSN (move_before);
2366 if (EDGE_COUNT (bb->preds) != 1)
2367 return false;
2368 incoming_edge = EDGE_PRED (bb, 0);
2369 final_dest_bb = incoming_edge->src;
2370 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2371 return false;
2372 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2373 if (incoming_bb_other_edge != incoming_edge)
2374 break;
2375 if (incoming_bb_other_edge->dest != other_bb)
2376 return false;
2380 e0 = EDGE_SUCC (bb, 0);
2381 e0_last_head = NULL_RTX;
2382 changed = false;
2384 for (ix = 1; ix < nedges; ix++)
2386 edge e = EDGE_SUCC (bb, ix);
2387 rtx e0_last, e_last;
2388 int nmatch;
2390 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2391 &e0_last, &e_last, 0);
2392 if (nmatch == 0)
2393 return false;
2395 if (nmatch < max_match)
2397 max_match = nmatch;
2398 e0_last_head = e0_last;
2402 /* If we matched an entire block, we probably have to avoid moving the
2403 last insn. */
2404 if (max_match > 0
2405 && e0_last_head == BB_END (e0->dest)
2406 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2407 || control_flow_insn_p (e0_last_head)))
2409 max_match--;
2410 if (max_match == 0)
2411 return false;
2413 e0_last_head = prev_real_insn (e0_last_head);
2414 while (DEBUG_INSN_P (e0_last_head));
2417 if (max_match == 0)
2418 return false;
2420 /* We must find a union of the live registers at each of the end points. */
2421 live = BITMAP_ALLOC (NULL);
2422 live_union = BITMAP_ALLOC (NULL);
2424 currptr = XNEWVEC (rtx, nedges);
2425 headptr = XNEWVEC (rtx, nedges);
2426 nextptr = XNEWVEC (rtx, nedges);
2428 for (ix = 0; ix < nedges; ix++)
2430 int j;
2431 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2432 rtx head = BB_HEAD (merge_bb);
2434 while (!NONDEBUG_INSN_P (head))
2435 head = NEXT_INSN (head);
2436 headptr[ix] = head;
2437 currptr[ix] = head;
2439 /* Compute the end point and live information */
2440 for (j = 1; j < max_match; j++)
2442 head = NEXT_INSN (head);
2443 while (!NONDEBUG_INSN_P (head));
2444 simulate_backwards_to_point (merge_bb, live, head);
2445 IOR_REG_SET (live_union, live);
2448 /* If we're moving across two blocks, verify the validity of the
2449 first move, then adjust the target and let the loop below deal
2450 with the final move. */
2451 if (final_dest_bb != NULL)
2453 rtx move_upto;
2455 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2456 jump, e0->dest, live_union,
2457 NULL, &move_upto);
2458 if (!moveall)
2460 if (move_upto == NULL_RTX)
2461 goto out;
2463 while (e0_last_head != move_upto)
2465 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2466 live_union);
2467 e0_last_head = PREV_INSN (e0_last_head);
2470 if (e0_last_head == NULL_RTX)
2471 goto out;
2473 jump = BB_END (final_dest_bb);
2474 cond = get_condition (jump, &move_before, true, false);
2475 if (cond == NULL_RTX)
2477 #ifdef HAVE_cc0
2478 if (reg_mentioned_p (cc0_rtx, jump))
2479 move_before = prev_nonnote_nondebug_insn (jump);
2480 else
2481 #endif
2482 move_before = jump;
2488 rtx move_upto;
2489 moveall = can_move_insns_across (currptr[0], e0_last_head,
2490 move_before, jump, e0->dest, live_union,
2491 NULL, &move_upto);
2492 if (!moveall && move_upto == NULL_RTX)
2494 if (jump == move_before)
2495 break;
2497 /* Try again, using a different insertion point. */
2498 move_before = jump;
2500 #ifdef HAVE_cc0
2501 /* Don't try moving before a cc0 user, as that may invalidate
2502 the cc0. */
2503 if (reg_mentioned_p (cc0_rtx, jump))
2504 break;
2505 #endif
2507 continue;
2510 if (final_dest_bb && !moveall)
2511 /* We haven't checked whether a partial move would be OK for the first
2512 move, so we have to fail this case. */
2513 break;
2515 changed = true;
2516 for (;;)
2518 if (currptr[0] == move_upto)
2519 break;
2520 for (ix = 0; ix < nedges; ix++)
2522 rtx curr = currptr[ix];
2524 curr = NEXT_INSN (curr);
2525 while (!NONDEBUG_INSN_P (curr));
2526 currptr[ix] = curr;
2530 /* If we can't currently move all of the identical insns, remember
2531 each insn after the range that we'll merge. */
2532 if (!moveall)
2533 for (ix = 0; ix < nedges; ix++)
2535 rtx curr = currptr[ix];
2537 curr = NEXT_INSN (curr);
2538 while (!NONDEBUG_INSN_P (curr));
2539 nextptr[ix] = curr;
2542 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2543 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2544 if (final_dest_bb != NULL)
2545 df_set_bb_dirty (final_dest_bb);
2546 df_set_bb_dirty (bb);
2547 for (ix = 1; ix < nedges; ix++)
2549 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2550 delete_insn_chain (headptr[ix], currptr[ix], false);
2552 if (!moveall)
2554 if (jump == move_before)
2555 break;
2557 /* For the unmerged insns, try a different insertion point. */
2558 move_before = jump;
2560 #ifdef HAVE_cc0
2561 /* Don't try moving before a cc0 user, as that may invalidate
2562 the cc0. */
2563 if (reg_mentioned_p (cc0_rtx, jump))
2564 break;
2565 #endif
2567 for (ix = 0; ix < nedges; ix++)
2568 currptr[ix] = headptr[ix] = nextptr[ix];
2571 while (!moveall);
2573 out:
2574 free (currptr);
2575 free (headptr);
2576 free (nextptr);
2578 crossjumps_occured |= changed;
2580 return changed;
2583 /* Return true if BB contains just bb note, or bb note followed
2584 by only DEBUG_INSNs. */
2586 static bool
2587 trivially_empty_bb_p (basic_block bb)
2589 rtx insn = BB_END (bb);
2591 while (1)
2593 if (insn == BB_HEAD (bb))
2594 return true;
2595 if (!DEBUG_INSN_P (insn))
2596 return false;
2597 insn = PREV_INSN (insn);
2601 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2602 instructions etc. Return nonzero if changes were made. */
2604 static bool
2605 try_optimize_cfg (int mode)
2607 bool changed_overall = false;
2608 bool changed;
2609 int iterations = 0;
2610 basic_block bb, b, next;
2612 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2613 clear_bb_flags ();
2615 crossjumps_occured = false;
2617 FOR_EACH_BB_FN (bb, cfun)
2618 update_forwarder_flag (bb);
2620 if (! targetm.cannot_modify_jumps_p ())
2622 first_pass = true;
2623 /* Attempt to merge blocks as made possible by edge removal. If
2624 a block has only one successor, and the successor has only
2625 one predecessor, they may be combined. */
2628 block_was_dirty = false;
2629 changed = false;
2630 iterations++;
2632 if (dump_file)
2633 fprintf (dump_file,
2634 "\n\ntry_optimize_cfg iteration %i\n\n",
2635 iterations);
2637 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2638 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2640 basic_block c;
2641 edge s;
2642 bool changed_here = false;
2644 /* Delete trivially dead basic blocks. This is either
2645 blocks with no predecessors, or empty blocks with no
2646 successors. However if the empty block with no
2647 successors is the successor of the ENTRY_BLOCK, it is
2648 kept. This ensures that the ENTRY_BLOCK will have a
2649 successor which is a precondition for many RTL
2650 passes. Empty blocks may result from expanding
2651 __builtin_unreachable (). */
2652 if (EDGE_COUNT (b->preds) == 0
2653 || (EDGE_COUNT (b->succs) == 0
2654 && trivially_empty_bb_p (b)
2655 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2656 != b))
2658 c = b->prev_bb;
2659 if (EDGE_COUNT (b->preds) > 0)
2661 edge e;
2662 edge_iterator ei;
2664 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2666 if (BB_FOOTER (b)
2667 && BARRIER_P (BB_FOOTER (b)))
2668 FOR_EACH_EDGE (e, ei, b->preds)
2669 if ((e->flags & EDGE_FALLTHRU)
2670 && BB_FOOTER (e->src) == NULL)
2672 if (BB_FOOTER (b))
2674 BB_FOOTER (e->src) = BB_FOOTER (b);
2675 BB_FOOTER (b) = NULL;
2677 else
2679 start_sequence ();
2680 BB_FOOTER (e->src) = emit_barrier ();
2681 end_sequence ();
2685 else
2687 rtx last = get_last_bb_insn (b);
2688 if (last && BARRIER_P (last))
2689 FOR_EACH_EDGE (e, ei, b->preds)
2690 if ((e->flags & EDGE_FALLTHRU))
2691 emit_barrier_after (BB_END (e->src));
2694 delete_basic_block (b);
2695 changed = true;
2696 /* Avoid trying to remove the exit block. */
2697 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2698 continue;
2701 /* Remove code labels no longer used. */
2702 if (single_pred_p (b)
2703 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2704 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2705 && LABEL_P (BB_HEAD (b))
2706 /* If the previous block ends with a branch to this
2707 block, we can't delete the label. Normally this
2708 is a condjump that is yet to be simplified, but
2709 if CASE_DROPS_THRU, this can be a tablejump with
2710 some element going to the same place as the
2711 default (fallthru). */
2712 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2713 || !JUMP_P (BB_END (single_pred (b)))
2714 || ! label_is_jump_target_p (BB_HEAD (b),
2715 BB_END (single_pred (b)))))
2717 delete_insn (BB_HEAD (b));
2718 if (dump_file)
2719 fprintf (dump_file, "Deleted label in block %i.\n",
2720 b->index);
2723 /* If we fall through an empty block, we can remove it. */
2724 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2725 && single_pred_p (b)
2726 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2727 && !LABEL_P (BB_HEAD (b))
2728 && FORWARDER_BLOCK_P (b)
2729 /* Note that forwarder_block_p true ensures that
2730 there is a successor for this block. */
2731 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2732 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2734 if (dump_file)
2735 fprintf (dump_file,
2736 "Deleting fallthru block %i.\n",
2737 b->index);
2739 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2740 ? b->next_bb : b->prev_bb);
2741 redirect_edge_succ_nodup (single_pred_edge (b),
2742 single_succ (b));
2743 delete_basic_block (b);
2744 changed = true;
2745 b = c;
2746 continue;
2749 /* Merge B with its single successor, if any. */
2750 if (single_succ_p (b)
2751 && (s = single_succ_edge (b))
2752 && !(s->flags & EDGE_COMPLEX)
2753 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2754 && single_pred_p (c)
2755 && b != c)
2757 /* When not in cfg_layout mode use code aware of reordering
2758 INSN. This code possibly creates new basic blocks so it
2759 does not fit merge_blocks interface and is kept here in
2760 hope that it will become useless once more of compiler
2761 is transformed to use cfg_layout mode. */
2763 if ((mode & CLEANUP_CFGLAYOUT)
2764 && can_merge_blocks_p (b, c))
2766 merge_blocks (b, c);
2767 update_forwarder_flag (b);
2768 changed_here = true;
2770 else if (!(mode & CLEANUP_CFGLAYOUT)
2771 /* If the jump insn has side effects,
2772 we can't kill the edge. */
2773 && (!JUMP_P (BB_END (b))
2774 || (reload_completed
2775 ? simplejump_p (BB_END (b))
2776 : (onlyjump_p (BB_END (b))
2777 && !tablejump_p (BB_END (b),
2778 NULL, NULL))))
2779 && (next = merge_blocks_move (s, b, c, mode)))
2781 b = next;
2782 changed_here = true;
2786 /* Simplify branch over branch. */
2787 if ((mode & CLEANUP_EXPENSIVE)
2788 && !(mode & CLEANUP_CFGLAYOUT)
2789 && try_simplify_condjump (b))
2790 changed_here = true;
2792 /* If B has a single outgoing edge, but uses a
2793 non-trivial jump instruction without side-effects, we
2794 can either delete the jump entirely, or replace it
2795 with a simple unconditional jump. */
2796 if (single_succ_p (b)
2797 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2798 && onlyjump_p (BB_END (b))
2799 && !find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX)
2800 && try_redirect_by_replacing_jump (single_succ_edge (b),
2801 single_succ (b),
2802 (mode & CLEANUP_CFGLAYOUT) != 0))
2804 update_forwarder_flag (b);
2805 changed_here = true;
2808 /* Simplify branch to branch. */
2809 if (try_forward_edges (mode, b))
2811 update_forwarder_flag (b);
2812 changed_here = true;
2815 /* Look for shared code between blocks. */
2816 if ((mode & CLEANUP_CROSSJUMP)
2817 && try_crossjump_bb (mode, b))
2818 changed_here = true;
2820 if ((mode & CLEANUP_CROSSJUMP)
2821 /* This can lengthen register lifetimes. Do it only after
2822 reload. */
2823 && reload_completed
2824 && try_head_merge_bb (b))
2825 changed_here = true;
2827 /* Don't get confused by the index shift caused by
2828 deleting blocks. */
2829 if (!changed_here)
2830 b = b->next_bb;
2831 else
2832 changed = true;
2835 if ((mode & CLEANUP_CROSSJUMP)
2836 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
2837 changed = true;
2839 if (block_was_dirty)
2841 /* This should only be set by head-merging. */
2842 gcc_assert (mode & CLEANUP_CROSSJUMP);
2843 df_analyze ();
2846 if (changed)
2848 /* Edge forwarding in particular can cause hot blocks previously
2849 reached by both hot and cold blocks to become dominated only
2850 by cold blocks. This will cause the verification below to fail,
2851 and lead to now cold code in the hot section. This is not easy
2852 to detect and fix during edge forwarding, and in some cases
2853 is only visible after newly unreachable blocks are deleted,
2854 which will be done in fixup_partitions. */
2855 fixup_partitions ();
2857 #ifdef ENABLE_CHECKING
2858 verify_flow_info ();
2859 #endif
2862 changed_overall |= changed;
2863 first_pass = false;
2865 while (changed);
2868 FOR_ALL_BB_FN (b, cfun)
2869 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2871 return changed_overall;
2874 /* Delete all unreachable basic blocks. */
2876 bool
2877 delete_unreachable_blocks (void)
2879 bool changed = false;
2880 basic_block b, prev_bb;
2882 find_unreachable_blocks ();
2884 /* When we're in GIMPLE mode and there may be debug insns, we should
2885 delete blocks in reverse dominator order, so as to get a chance
2886 to substitute all released DEFs into debug stmts. If we don't
2887 have dominators information, walking blocks backward gets us a
2888 better chance of retaining most debug information than
2889 otherwise. */
2890 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2891 && dom_info_available_p (CDI_DOMINATORS))
2893 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2894 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2896 prev_bb = b->prev_bb;
2898 if (!(b->flags & BB_REACHABLE))
2900 /* Speed up the removal of blocks that don't dominate
2901 others. Walking backwards, this should be the common
2902 case. */
2903 if (!first_dom_son (CDI_DOMINATORS, b))
2904 delete_basic_block (b);
2905 else
2907 vec<basic_block> h
2908 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2910 while (h.length ())
2912 b = h.pop ();
2914 prev_bb = b->prev_bb;
2916 gcc_assert (!(b->flags & BB_REACHABLE));
2918 delete_basic_block (b);
2921 h.release ();
2924 changed = true;
2928 else
2930 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2931 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2933 prev_bb = b->prev_bb;
2935 if (!(b->flags & BB_REACHABLE))
2937 delete_basic_block (b);
2938 changed = true;
2943 if (changed)
2944 tidy_fallthru_edges ();
2945 return changed;
2948 /* Delete any jump tables never referenced. We can't delete them at the
2949 time of removing tablejump insn as they are referenced by the preceding
2950 insns computing the destination, so we delay deleting and garbagecollect
2951 them once life information is computed. */
2952 void
2953 delete_dead_jumptables (void)
2955 basic_block bb;
2957 /* A dead jump table does not belong to any basic block. Scan insns
2958 between two adjacent basic blocks. */
2959 FOR_EACH_BB_FN (bb, cfun)
2961 rtx insn, next;
2963 for (insn = NEXT_INSN (BB_END (bb));
2964 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
2965 insn = next)
2967 next = NEXT_INSN (insn);
2968 if (LABEL_P (insn)
2969 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
2970 && JUMP_TABLE_DATA_P (next))
2972 rtx label = insn, jump = next;
2974 if (dump_file)
2975 fprintf (dump_file, "Dead jumptable %i removed\n",
2976 INSN_UID (insn));
2978 next = NEXT_INSN (next);
2979 delete_insn (jump);
2980 delete_insn (label);
2987 /* Tidy the CFG by deleting unreachable code and whatnot. */
2989 bool
2990 cleanup_cfg (int mode)
2992 bool changed = false;
2994 /* Set the cfglayout mode flag here. We could update all the callers
2995 but that is just inconvenient, especially given that we eventually
2996 want to have cfglayout mode as the default. */
2997 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2998 mode |= CLEANUP_CFGLAYOUT;
3000 timevar_push (TV_CLEANUP_CFG);
3001 if (delete_unreachable_blocks ())
3003 changed = true;
3004 /* We've possibly created trivially dead code. Cleanup it right
3005 now to introduce more opportunities for try_optimize_cfg. */
3006 if (!(mode & (CLEANUP_NO_INSN_DEL))
3007 && !reload_completed)
3008 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3011 compact_blocks ();
3013 /* To tail-merge blocks ending in the same noreturn function (e.g.
3014 a call to abort) we have to insert fake edges to exit. Do this
3015 here once. The fake edges do not interfere with any other CFG
3016 cleanups. */
3017 if (mode & CLEANUP_CROSSJUMP)
3018 add_noreturn_fake_exit_edges ();
3020 if (!dbg_cnt (cfg_cleanup))
3021 return changed;
3023 while (try_optimize_cfg (mode))
3025 delete_unreachable_blocks (), changed = true;
3026 if (!(mode & CLEANUP_NO_INSN_DEL))
3028 /* Try to remove some trivially dead insns when doing an expensive
3029 cleanup. But delete_trivially_dead_insns doesn't work after
3030 reload (it only handles pseudos) and run_fast_dce is too costly
3031 to run in every iteration.
3033 For effective cross jumping, we really want to run a fast DCE to
3034 clean up any dead conditions, or they get in the way of performing
3035 useful tail merges.
3037 Other transformations in cleanup_cfg are not so sensitive to dead
3038 code, so delete_trivially_dead_insns or even doing nothing at all
3039 is good enough. */
3040 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3041 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3042 break;
3043 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
3044 run_fast_dce ();
3046 else
3047 break;
3050 if (mode & CLEANUP_CROSSJUMP)
3051 remove_fake_exit_edges ();
3053 /* Don't call delete_dead_jumptables in cfglayout mode, because
3054 that function assumes that jump tables are in the insns stream.
3055 But we also don't _have_ to delete dead jumptables in cfglayout
3056 mode because we shouldn't even be looking at things that are
3057 not in a basic block. Dead jumptables are cleaned up when
3058 going out of cfglayout mode. */
3059 if (!(mode & CLEANUP_CFGLAYOUT))
3060 delete_dead_jumptables ();
3062 /* ??? We probably do this way too often. */
3063 if (current_loops
3064 && (changed
3065 || (mode & CLEANUP_CFG_CHANGED)))
3067 timevar_push (TV_REPAIR_LOOPS);
3068 /* The above doesn't preserve dominance info if available. */
3069 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3070 calculate_dominance_info (CDI_DOMINATORS);
3071 fix_loop_structure (NULL);
3072 free_dominance_info (CDI_DOMINATORS);
3073 timevar_pop (TV_REPAIR_LOOPS);
3076 timevar_pop (TV_CLEANUP_CFG);
3078 return changed;
3081 static unsigned int
3082 execute_jump (void)
3084 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3085 if (dump_file)
3086 dump_flow_info (dump_file, dump_flags);
3087 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3088 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3089 return 0;
3092 namespace {
3094 const pass_data pass_data_jump =
3096 RTL_PASS, /* type */
3097 "jump", /* name */
3098 OPTGROUP_NONE, /* optinfo_flags */
3099 false, /* has_gate */
3100 true, /* has_execute */
3101 TV_JUMP, /* tv_id */
3102 0, /* properties_required */
3103 0, /* properties_provided */
3104 0, /* properties_destroyed */
3105 0, /* todo_flags_start */
3106 TODO_verify_rtl_sharing, /* todo_flags_finish */
3109 class pass_jump : public rtl_opt_pass
3111 public:
3112 pass_jump (gcc::context *ctxt)
3113 : rtl_opt_pass (pass_data_jump, ctxt)
3116 /* opt_pass methods: */
3117 unsigned int execute () { return execute_jump (); }
3119 }; // class pass_jump
3121 } // anon namespace
3123 rtl_opt_pass *
3124 make_pass_jump (gcc::context *ctxt)
3126 return new pass_jump (ctxt);
3129 static unsigned int
3130 execute_jump2 (void)
3132 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3133 return 0;
3136 namespace {
3138 const pass_data pass_data_jump2 =
3140 RTL_PASS, /* type */
3141 "jump2", /* name */
3142 OPTGROUP_NONE, /* optinfo_flags */
3143 false, /* has_gate */
3144 true, /* has_execute */
3145 TV_JUMP, /* tv_id */
3146 0, /* properties_required */
3147 0, /* properties_provided */
3148 0, /* properties_destroyed */
3149 0, /* todo_flags_start */
3150 TODO_verify_rtl_sharing, /* todo_flags_finish */
3153 class pass_jump2 : public rtl_opt_pass
3155 public:
3156 pass_jump2 (gcc::context *ctxt)
3157 : rtl_opt_pass (pass_data_jump2, ctxt)
3160 /* opt_pass methods: */
3161 unsigned int execute () { return execute_jump2 (); }
3163 }; // class pass_jump2
3165 } // anon namespace
3167 rtl_opt_pass *
3168 make_pass_jump2 (gcc::context *ctxt)
3170 return new pass_jump2 (ctxt);