2014-07-29 Ed Smith-Rowland <3dw4rd@verizon.net>
[official-gcc.git] / gcc / cfgcleanup.c
blob01fd9ff72aedc84574a90642e0e3e58e92fe87af
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 (JUMP_P (BB_END (b)) && CROSSING_JUMP_P (BB_END (b)))
423 return false;
425 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
427 basic_block target, first;
428 location_t goto_locus;
429 int counter;
430 bool threaded = false;
431 int nthreaded_edges = 0;
432 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
434 /* Skip complex edges because we don't know how to update them.
436 Still handle fallthru edges, as we can succeed to forward fallthru
437 edge to the same place as the branch edge of conditional branch
438 and turn conditional branch to an unconditional branch. */
439 if (e->flags & EDGE_COMPLEX)
441 ei_next (&ei);
442 continue;
445 target = first = e->dest;
446 counter = NUM_FIXED_BLOCKS;
447 goto_locus = e->goto_locus;
449 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
450 up jumps that cross between hot/cold sections.
452 Basic block partitioning may result in some jumps that appear
453 to be optimizable (or blocks that appear to be mergeable), but which
454 really must be left untouched (they are required to make it safely
455 across partition boundaries). See the comments at the top of
456 bb-reorder.c:partition_hot_cold_basic_blocks for complete
457 details. */
459 if (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
460 && JUMP_P (BB_END (first))
461 && CROSSING_JUMP_P (BB_END (first)))
462 return changed;
464 while (counter < n_basic_blocks_for_fn (cfun))
466 basic_block new_target = NULL;
467 bool new_target_threaded = false;
468 may_thread |= (target->flags & BB_MODIFIED) != 0;
470 if (FORWARDER_BLOCK_P (target)
471 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
472 && single_succ (target) != EXIT_BLOCK_PTR_FOR_FN (cfun))
474 /* Bypass trivial infinite loops. */
475 new_target = single_succ (target);
476 if (target == new_target)
477 counter = n_basic_blocks_for_fn (cfun);
478 else if (!optimize)
480 /* When not optimizing, ensure that edges or forwarder
481 blocks with different locus are not optimized out. */
482 location_t new_locus = single_succ_edge (target)->goto_locus;
483 location_t locus = goto_locus;
485 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
486 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
487 && new_locus != locus)
488 new_target = NULL;
489 else
491 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
492 locus = new_locus;
494 rtx last = BB_END (target);
495 if (DEBUG_INSN_P (last))
496 last = prev_nondebug_insn (last);
497 if (last && INSN_P (last))
498 new_locus = INSN_LOCATION (last);
499 else
500 new_locus = UNKNOWN_LOCATION;
502 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
503 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
504 && new_locus != locus)
505 new_target = NULL;
506 else
508 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
509 locus = new_locus;
511 goto_locus = locus;
517 /* Allow to thread only over one edge at time to simplify updating
518 of probabilities. */
519 else if ((mode & CLEANUP_THREADING) && may_thread)
521 edge t = thread_jump (e, target);
522 if (t)
524 if (!threaded_edges)
525 threaded_edges = XNEWVEC (edge,
526 n_basic_blocks_for_fn (cfun));
527 else
529 int i;
531 /* Detect an infinite loop across blocks not
532 including the start block. */
533 for (i = 0; i < nthreaded_edges; ++i)
534 if (threaded_edges[i] == t)
535 break;
536 if (i < nthreaded_edges)
538 counter = n_basic_blocks_for_fn (cfun);
539 break;
543 /* Detect an infinite loop across the start block. */
544 if (t->dest == b)
545 break;
547 gcc_assert (nthreaded_edges
548 < (n_basic_blocks_for_fn (cfun)
549 - NUM_FIXED_BLOCKS));
550 threaded_edges[nthreaded_edges++] = t;
552 new_target = t->dest;
553 new_target_threaded = true;
557 if (!new_target)
558 break;
560 counter++;
561 target = new_target;
562 threaded |= new_target_threaded;
565 if (counter >= n_basic_blocks_for_fn (cfun))
567 if (dump_file)
568 fprintf (dump_file, "Infinite loop in BB %i.\n",
569 target->index);
571 else if (target == first)
572 ; /* We didn't do anything. */
573 else
575 /* Save the values now, as the edge may get removed. */
576 gcov_type edge_count = e->count;
577 int edge_probability = e->probability;
578 int edge_frequency;
579 int n = 0;
581 e->goto_locus = goto_locus;
583 /* Don't force if target is exit block. */
584 if (threaded && target != EXIT_BLOCK_PTR_FOR_FN (cfun))
586 notice_new_block (redirect_edge_and_branch_force (e, target));
587 if (dump_file)
588 fprintf (dump_file, "Conditionals threaded.\n");
590 else if (!redirect_edge_and_branch (e, target))
592 if (dump_file)
593 fprintf (dump_file,
594 "Forwarding edge %i->%i to %i failed.\n",
595 b->index, e->dest->index, target->index);
596 ei_next (&ei);
597 continue;
600 /* We successfully forwarded the edge. Now update profile
601 data: for each edge we traversed in the chain, remove
602 the original edge's execution count. */
603 edge_frequency = apply_probability (b->frequency, edge_probability);
607 edge t;
609 if (!single_succ_p (first))
611 gcc_assert (n < nthreaded_edges);
612 t = threaded_edges [n++];
613 gcc_assert (t->src == first);
614 update_bb_profile_for_threading (first, edge_frequency,
615 edge_count, t);
616 update_br_prob_note (first);
618 else
620 first->count -= edge_count;
621 if (first->count < 0)
622 first->count = 0;
623 first->frequency -= edge_frequency;
624 if (first->frequency < 0)
625 first->frequency = 0;
626 /* It is possible that as the result of
627 threading we've removed edge as it is
628 threaded to the fallthru edge. Avoid
629 getting out of sync. */
630 if (n < nthreaded_edges
631 && first == threaded_edges [n]->src)
632 n++;
633 t = single_succ_edge (first);
636 t->count -= edge_count;
637 if (t->count < 0)
638 t->count = 0;
639 first = t->dest;
641 while (first != target);
643 changed = true;
644 continue;
646 ei_next (&ei);
649 free (threaded_edges);
650 return changed;
654 /* Blocks A and B are to be merged into a single block. A has no incoming
655 fallthru edge, so it can be moved before B without adding or modifying
656 any jumps (aside from the jump from A to B). */
658 static void
659 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
661 rtx barrier;
663 /* If we are partitioning hot/cold basic blocks, we don't want to
664 mess up unconditional or indirect jumps that cross between hot
665 and cold sections.
667 Basic block partitioning may result in some jumps that appear to
668 be optimizable (or blocks that appear to be mergeable), but which really
669 must be left untouched (they are required to make it safely across
670 partition boundaries). See the comments at the top of
671 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
673 if (BB_PARTITION (a) != BB_PARTITION (b))
674 return;
676 barrier = next_nonnote_insn (BB_END (a));
677 gcc_assert (BARRIER_P (barrier));
678 delete_insn (barrier);
680 /* Scramble the insn chain. */
681 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
682 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
683 df_set_bb_dirty (a);
685 if (dump_file)
686 fprintf (dump_file, "Moved block %d before %d and merged.\n",
687 a->index, b->index);
689 /* Swap the records for the two blocks around. */
691 unlink_block (a);
692 link_block (a, b->prev_bb);
694 /* Now blocks A and B are contiguous. Merge them. */
695 merge_blocks (a, b);
698 /* Blocks A and B are to be merged into a single block. B has no outgoing
699 fallthru edge, so it can be moved after A without adding or modifying
700 any jumps (aside from the jump from A to B). */
702 static void
703 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
705 rtx barrier, real_b_end;
706 rtx label, table;
708 /* If we are partitioning hot/cold basic blocks, we don't want to
709 mess up unconditional or indirect jumps that cross between hot
710 and cold sections.
712 Basic block partitioning may result in some jumps that appear to
713 be optimizable (or blocks that appear to be mergeable), but which really
714 must be left untouched (they are required to make it safely across
715 partition boundaries). See the comments at the top of
716 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
718 if (BB_PARTITION (a) != BB_PARTITION (b))
719 return;
721 real_b_end = BB_END (b);
723 /* If there is a jump table following block B temporarily add the jump table
724 to block B so that it will also be moved to the correct location. */
725 if (tablejump_p (BB_END (b), &label, &table)
726 && prev_active_insn (label) == BB_END (b))
728 BB_END (b) = table;
731 /* There had better have been a barrier there. Delete it. */
732 barrier = NEXT_INSN (BB_END (b));
733 if (barrier && BARRIER_P (barrier))
734 delete_insn (barrier);
737 /* Scramble the insn chain. */
738 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
740 /* Restore the real end of b. */
741 BB_END (b) = real_b_end;
743 if (dump_file)
744 fprintf (dump_file, "Moved block %d after %d and merged.\n",
745 b->index, a->index);
747 /* Now blocks A and B are contiguous. Merge them. */
748 merge_blocks (a, b);
751 /* Attempt to merge basic blocks that are potentially non-adjacent.
752 Return NULL iff the attempt failed, otherwise return basic block
753 where cleanup_cfg should continue. Because the merging commonly
754 moves basic block away or introduces another optimization
755 possibility, return basic block just before B so cleanup_cfg don't
756 need to iterate.
758 It may be good idea to return basic block before C in the case
759 C has been moved after B and originally appeared earlier in the
760 insn sequence, but we have no information available about the
761 relative ordering of these two. Hopefully it is not too common. */
763 static basic_block
764 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
766 basic_block next;
768 /* If we are partitioning hot/cold basic blocks, we don't want to
769 mess up unconditional or indirect jumps that cross between hot
770 and cold sections.
772 Basic block partitioning may result in some jumps that appear to
773 be optimizable (or blocks that appear to be mergeable), but which really
774 must be left untouched (they are required to make it safely across
775 partition boundaries). See the comments at the top of
776 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
778 if (BB_PARTITION (b) != BB_PARTITION (c))
779 return NULL;
781 /* If B has a fallthru edge to C, no need to move anything. */
782 if (e->flags & EDGE_FALLTHRU)
784 int b_index = b->index, c_index = c->index;
786 /* Protect the loop latches. */
787 if (current_loops && c->loop_father->latch == c)
788 return NULL;
790 merge_blocks (b, c);
791 update_forwarder_flag (b);
793 if (dump_file)
794 fprintf (dump_file, "Merged %d and %d without moving.\n",
795 b_index, c_index);
797 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
800 /* Otherwise we will need to move code around. Do that only if expensive
801 transformations are allowed. */
802 else if (mode & CLEANUP_EXPENSIVE)
804 edge tmp_edge, b_fallthru_edge;
805 bool c_has_outgoing_fallthru;
806 bool b_has_incoming_fallthru;
808 /* Avoid overactive code motion, as the forwarder blocks should be
809 eliminated by edge redirection instead. One exception might have
810 been if B is a forwarder block and C has no fallthru edge, but
811 that should be cleaned up by bb-reorder instead. */
812 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
813 return NULL;
815 /* We must make sure to not munge nesting of lexical blocks,
816 and loop notes. This is done by squeezing out all the notes
817 and leaving them there to lie. Not ideal, but functional. */
819 tmp_edge = find_fallthru_edge (c->succs);
820 c_has_outgoing_fallthru = (tmp_edge != NULL);
822 tmp_edge = find_fallthru_edge (b->preds);
823 b_has_incoming_fallthru = (tmp_edge != NULL);
824 b_fallthru_edge = tmp_edge;
825 next = b->prev_bb;
826 if (next == c)
827 next = next->prev_bb;
829 /* Otherwise, we're going to try to move C after B. If C does
830 not have an outgoing fallthru, then it can be moved
831 immediately after B without introducing or modifying jumps. */
832 if (! c_has_outgoing_fallthru)
834 merge_blocks_move_successor_nojumps (b, c);
835 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
838 /* If B does not have an incoming fallthru, then it can be moved
839 immediately before C without introducing or modifying jumps.
840 C cannot be the first block, so we do not have to worry about
841 accessing a non-existent block. */
843 if (b_has_incoming_fallthru)
845 basic_block bb;
847 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
848 return NULL;
849 bb = force_nonfallthru (b_fallthru_edge);
850 if (bb)
851 notice_new_block (bb);
854 merge_blocks_move_predecessor_nojumps (b, c);
855 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
858 return NULL;
862 /* Removes the memory attributes of MEM expression
863 if they are not equal. */
865 void
866 merge_memattrs (rtx x, rtx y)
868 int i;
869 int j;
870 enum rtx_code code;
871 const char *fmt;
873 if (x == y)
874 return;
875 if (x == 0 || y == 0)
876 return;
878 code = GET_CODE (x);
880 if (code != GET_CODE (y))
881 return;
883 if (GET_MODE (x) != GET_MODE (y))
884 return;
886 if (code == MEM && MEM_ATTRS (x) != MEM_ATTRS (y))
888 if (! MEM_ATTRS (x))
889 MEM_ATTRS (y) = 0;
890 else if (! MEM_ATTRS (y))
891 MEM_ATTRS (x) = 0;
892 else
894 HOST_WIDE_INT mem_size;
896 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
898 set_mem_alias_set (x, 0);
899 set_mem_alias_set (y, 0);
902 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
904 set_mem_expr (x, 0);
905 set_mem_expr (y, 0);
906 clear_mem_offset (x);
907 clear_mem_offset (y);
909 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
910 || (MEM_OFFSET_KNOWN_P (x)
911 && MEM_OFFSET (x) != MEM_OFFSET (y)))
913 clear_mem_offset (x);
914 clear_mem_offset (y);
917 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
919 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
920 set_mem_size (x, mem_size);
921 set_mem_size (y, mem_size);
923 else
925 clear_mem_size (x);
926 clear_mem_size (y);
929 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
930 set_mem_align (y, MEM_ALIGN (x));
933 if (code == MEM)
935 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
937 MEM_READONLY_P (x) = 0;
938 MEM_READONLY_P (y) = 0;
940 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
942 MEM_NOTRAP_P (x) = 0;
943 MEM_NOTRAP_P (y) = 0;
945 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
947 MEM_VOLATILE_P (x) = 1;
948 MEM_VOLATILE_P (y) = 1;
952 fmt = GET_RTX_FORMAT (code);
953 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
955 switch (fmt[i])
957 case 'E':
958 /* Two vectors must have the same length. */
959 if (XVECLEN (x, i) != XVECLEN (y, i))
960 return;
962 for (j = 0; j < XVECLEN (x, i); j++)
963 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
965 break;
967 case 'e':
968 merge_memattrs (XEXP (x, i), XEXP (y, i));
971 return;
975 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
976 different single sets S1 and S2. */
978 static bool
979 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
981 int i;
982 rtx e1, e2;
984 if (p1 == s1 && p2 == s2)
985 return true;
987 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
988 return false;
990 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
991 return false;
993 for (i = 0; i < XVECLEN (p1, 0); i++)
995 e1 = XVECEXP (p1, 0, i);
996 e2 = XVECEXP (p2, 0, i);
997 if (e1 == s1 && e2 == s2)
998 continue;
999 if (reload_completed
1000 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
1001 continue;
1003 return false;
1006 return true;
1009 /* Examine register notes on I1 and I2 and return:
1010 - dir_forward if I1 can be replaced by I2, or
1011 - dir_backward if I2 can be replaced by I1, or
1012 - dir_both if both are the case. */
1014 static enum replace_direction
1015 can_replace_by (rtx i1, rtx i2)
1017 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1018 bool c1, c2;
1020 /* Check for 2 sets. */
1021 s1 = single_set (i1);
1022 s2 = single_set (i2);
1023 if (s1 == NULL_RTX || s2 == NULL_RTX)
1024 return dir_none;
1026 /* Check that the 2 sets set the same dest. */
1027 d1 = SET_DEST (s1);
1028 d2 = SET_DEST (s2);
1029 if (!(reload_completed
1030 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1031 return dir_none;
1033 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1034 set dest to the same value. */
1035 note1 = find_reg_equal_equiv_note (i1);
1036 note2 = find_reg_equal_equiv_note (i2);
1037 if (!note1 || !note2 || !rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0))
1038 || !CONST_INT_P (XEXP (note1, 0)))
1039 return dir_none;
1041 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1042 return dir_none;
1044 /* Although the 2 sets set dest to the same value, we cannot replace
1045 (set (dest) (const_int))
1047 (set (dest) (reg))
1048 because we don't know if the reg is live and has the same value at the
1049 location of replacement. */
1050 src1 = SET_SRC (s1);
1051 src2 = SET_SRC (s2);
1052 c1 = CONST_INT_P (src1);
1053 c2 = CONST_INT_P (src2);
1054 if (c1 && c2)
1055 return dir_both;
1056 else if (c2)
1057 return dir_forward;
1058 else if (c1)
1059 return dir_backward;
1061 return dir_none;
1064 /* Merges directions A and B. */
1066 static enum replace_direction
1067 merge_dir (enum replace_direction a, enum replace_direction b)
1069 /* Implements the following table:
1070 |bo fw bw no
1071 ---+-----------
1072 bo |bo fw bw no
1073 fw |-- fw no no
1074 bw |-- -- bw no
1075 no |-- -- -- no. */
1077 if (a == b)
1078 return a;
1080 if (a == dir_both)
1081 return b;
1082 if (b == dir_both)
1083 return a;
1085 return dir_none;
1088 /* Examine I1 and I2 and return:
1089 - dir_forward if I1 can be replaced by I2, or
1090 - dir_backward if I2 can be replaced by I1, or
1091 - dir_both if both are the case. */
1093 static enum replace_direction
1094 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx i1, rtx i2)
1096 rtx p1, p2;
1098 /* Verify that I1 and I2 are equivalent. */
1099 if (GET_CODE (i1) != GET_CODE (i2))
1100 return dir_none;
1102 /* __builtin_unreachable() may lead to empty blocks (ending with
1103 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1104 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1105 return dir_both;
1107 /* ??? Do not allow cross-jumping between different stack levels. */
1108 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1109 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1110 if (p1 && p2)
1112 p1 = XEXP (p1, 0);
1113 p2 = XEXP (p2, 0);
1114 if (!rtx_equal_p (p1, p2))
1115 return dir_none;
1117 /* ??? Worse, this adjustment had better be constant lest we
1118 have differing incoming stack levels. */
1119 if (!frame_pointer_needed
1120 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1121 return dir_none;
1123 else if (p1 || p2)
1124 return dir_none;
1126 p1 = PATTERN (i1);
1127 p2 = PATTERN (i2);
1129 if (GET_CODE (p1) != GET_CODE (p2))
1130 return dir_none;
1132 /* If this is a CALL_INSN, compare register usage information.
1133 If we don't check this on stack register machines, the two
1134 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1135 numbers of stack registers in the same basic block.
1136 If we don't check this on machines with delay slots, a delay slot may
1137 be filled that clobbers a parameter expected by the subroutine.
1139 ??? We take the simple route for now and assume that if they're
1140 equal, they were constructed identically.
1142 Also check for identical exception regions. */
1144 if (CALL_P (i1))
1146 /* Ensure the same EH region. */
1147 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1148 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1150 if (!n1 && n2)
1151 return dir_none;
1153 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1154 return dir_none;
1156 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1157 CALL_INSN_FUNCTION_USAGE (i2))
1158 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1159 return dir_none;
1161 /* For address sanitizer, never crossjump __asan_report_* builtins,
1162 otherwise errors might be reported on incorrect lines. */
1163 if (flag_sanitize & SANITIZE_ADDRESS)
1165 rtx call = get_call_rtx_from (i1);
1166 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1168 rtx symbol = XEXP (XEXP (call, 0), 0);
1169 if (SYMBOL_REF_DECL (symbol)
1170 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1172 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1173 == BUILT_IN_NORMAL)
1174 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1175 >= BUILT_IN_ASAN_REPORT_LOAD1
1176 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1177 <= BUILT_IN_ASAN_STOREN)
1178 return dir_none;
1184 #ifdef STACK_REGS
1185 /* If cross_jump_death_matters is not 0, the insn's mode
1186 indicates whether or not the insn contains any stack-like
1187 regs. */
1189 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1191 /* If register stack conversion has already been done, then
1192 death notes must also be compared before it is certain that
1193 the two instruction streams match. */
1195 rtx note;
1196 HARD_REG_SET i1_regset, i2_regset;
1198 CLEAR_HARD_REG_SET (i1_regset);
1199 CLEAR_HARD_REG_SET (i2_regset);
1201 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1202 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1203 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1205 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1206 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1207 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1209 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1210 return dir_none;
1212 #endif
1214 if (reload_completed
1215 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1216 return dir_both;
1218 return can_replace_by (i1, i2);
1221 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1222 flow_find_head_matching_sequence, ensure the notes match. */
1224 static void
1225 merge_notes (rtx i1, rtx i2)
1227 /* If the merged insns have different REG_EQUAL notes, then
1228 remove them. */
1229 rtx equiv1 = find_reg_equal_equiv_note (i1);
1230 rtx equiv2 = find_reg_equal_equiv_note (i2);
1232 if (equiv1 && !equiv2)
1233 remove_note (i1, equiv1);
1234 else if (!equiv1 && equiv2)
1235 remove_note (i2, equiv2);
1236 else if (equiv1 && equiv2
1237 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1239 remove_note (i1, equiv1);
1240 remove_note (i2, equiv2);
1244 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1245 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1246 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1247 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1248 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1250 static void
1251 walk_to_nondebug_insn (rtx *i1, basic_block *bb1, bool follow_fallthru,
1252 bool *did_fallthru)
1254 edge fallthru;
1256 *did_fallthru = false;
1258 /* Ignore notes. */
1259 while (!NONDEBUG_INSN_P (*i1))
1261 if (*i1 != BB_HEAD (*bb1))
1263 *i1 = PREV_INSN (*i1);
1264 continue;
1267 if (!follow_fallthru)
1268 return;
1270 fallthru = find_fallthru_edge ((*bb1)->preds);
1271 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1272 || !single_succ_p (fallthru->src))
1273 return;
1275 *bb1 = fallthru->src;
1276 *i1 = BB_END (*bb1);
1277 *did_fallthru = true;
1281 /* Look through the insns at the end of BB1 and BB2 and find the longest
1282 sequence that are either equivalent, or allow forward or backward
1283 replacement. Store the first insns for that sequence in *F1 and *F2 and
1284 return the sequence length.
1286 DIR_P indicates the allowed replacement direction on function entry, and
1287 the actual replacement direction on function exit. If NULL, only equivalent
1288 sequences are allowed.
1290 To simplify callers of this function, if the blocks match exactly,
1291 store the head of the blocks in *F1 and *F2. */
1294 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx *f1, rtx *f2,
1295 enum replace_direction *dir_p)
1297 rtx i1, i2, last1, last2, afterlast1, afterlast2;
1298 int ninsns = 0;
1299 enum replace_direction dir, last_dir, afterlast_dir;
1300 bool follow_fallthru, did_fallthru;
1302 if (dir_p)
1303 dir = *dir_p;
1304 else
1305 dir = dir_both;
1306 afterlast_dir = dir;
1307 last_dir = afterlast_dir;
1309 /* Skip simple jumps at the end of the blocks. Complex jumps still
1310 need to be compared for equivalence, which we'll do below. */
1312 i1 = BB_END (bb1);
1313 last1 = afterlast1 = last2 = afterlast2 = NULL_RTX;
1314 if (onlyjump_p (i1)
1315 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1317 last1 = i1;
1318 i1 = PREV_INSN (i1);
1321 i2 = BB_END (bb2);
1322 if (onlyjump_p (i2)
1323 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1325 last2 = i2;
1326 /* Count everything except for unconditional jump as insn.
1327 Don't count any jumps if dir_p is NULL. */
1328 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1329 ninsns++;
1330 i2 = PREV_INSN (i2);
1333 while (true)
1335 /* In the following example, we can replace all jumps to C by jumps to A.
1337 This removes 4 duplicate insns.
1338 [bb A] insn1 [bb C] insn1
1339 insn2 insn2
1340 [bb B] insn3 insn3
1341 insn4 insn4
1342 jump_insn jump_insn
1344 We could also replace all jumps to A by jumps to C, but that leaves B
1345 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1346 step, all jumps to B would be replaced with jumps to the middle of C,
1347 achieving the same result with more effort.
1348 So we allow only the first possibility, which means that we don't allow
1349 fallthru in the block that's being replaced. */
1351 follow_fallthru = dir_p && dir != dir_forward;
1352 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1353 if (did_fallthru)
1354 dir = dir_backward;
1356 follow_fallthru = dir_p && dir != dir_backward;
1357 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1358 if (did_fallthru)
1359 dir = dir_forward;
1361 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1362 break;
1364 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1365 if (dir == dir_none || (!dir_p && dir != dir_both))
1366 break;
1368 merge_memattrs (i1, i2);
1370 /* Don't begin a cross-jump with a NOTE insn. */
1371 if (INSN_P (i1))
1373 merge_notes (i1, i2);
1375 afterlast1 = last1, afterlast2 = last2;
1376 last1 = i1, last2 = i2;
1377 afterlast_dir = last_dir;
1378 last_dir = dir;
1379 if (active_insn_p (i1))
1380 ninsns++;
1383 i1 = PREV_INSN (i1);
1384 i2 = PREV_INSN (i2);
1387 #ifdef HAVE_cc0
1388 /* Don't allow the insn after a compare to be shared by
1389 cross-jumping unless the compare is also shared. */
1390 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
1391 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1392 #endif
1394 /* Include preceding notes and labels in the cross-jump. One,
1395 this may bring us to the head of the blocks as requested above.
1396 Two, it keeps line number notes as matched as may be. */
1397 if (ninsns)
1399 bb1 = BLOCK_FOR_INSN (last1);
1400 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1401 last1 = PREV_INSN (last1);
1403 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1404 last1 = PREV_INSN (last1);
1406 bb2 = BLOCK_FOR_INSN (last2);
1407 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1408 last2 = PREV_INSN (last2);
1410 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1411 last2 = PREV_INSN (last2);
1413 *f1 = last1;
1414 *f2 = last2;
1417 if (dir_p)
1418 *dir_p = last_dir;
1419 return ninsns;
1422 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1423 the head of the two blocks. Do not include jumps at the end.
1424 If STOP_AFTER is nonzero, stop after finding that many matching
1425 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1426 non-zero, only count active insns. */
1429 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx *f1,
1430 rtx *f2, int stop_after)
1432 rtx i1, i2, last1, last2, beforelast1, beforelast2;
1433 int ninsns = 0;
1434 edge e;
1435 edge_iterator ei;
1436 int nehedges1 = 0, nehedges2 = 0;
1438 FOR_EACH_EDGE (e, ei, bb1->succs)
1439 if (e->flags & EDGE_EH)
1440 nehedges1++;
1441 FOR_EACH_EDGE (e, ei, bb2->succs)
1442 if (e->flags & EDGE_EH)
1443 nehedges2++;
1445 i1 = BB_HEAD (bb1);
1446 i2 = BB_HEAD (bb2);
1447 last1 = beforelast1 = last2 = beforelast2 = NULL_RTX;
1449 while (true)
1451 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1452 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1454 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1455 break;
1456 i1 = NEXT_INSN (i1);
1459 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1461 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1462 break;
1463 i2 = NEXT_INSN (i2);
1466 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1467 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1468 break;
1470 if (NOTE_P (i1) || NOTE_P (i2)
1471 || JUMP_P (i1) || JUMP_P (i2))
1472 break;
1474 /* A sanity check to make sure we're not merging insns with different
1475 effects on EH. If only one of them ends a basic block, it shouldn't
1476 have an EH edge; if both end a basic block, there should be the same
1477 number of EH edges. */
1478 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1479 && nehedges1 > 0)
1480 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1481 && nehedges2 > 0)
1482 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1483 && nehedges1 != nehedges2))
1484 break;
1486 if (old_insns_match_p (0, i1, i2) != dir_both)
1487 break;
1489 merge_memattrs (i1, i2);
1491 /* Don't begin a cross-jump with a NOTE insn. */
1492 if (INSN_P (i1))
1494 merge_notes (i1, i2);
1496 beforelast1 = last1, beforelast2 = last2;
1497 last1 = i1, last2 = i2;
1498 if (!stop_after || active_insn_p (i1))
1499 ninsns++;
1502 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1503 || (stop_after > 0 && ninsns == stop_after))
1504 break;
1506 i1 = NEXT_INSN (i1);
1507 i2 = NEXT_INSN (i2);
1510 #ifdef HAVE_cc0
1511 /* Don't allow a compare to be shared by cross-jumping unless the insn
1512 after the compare is also shared. */
1513 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && sets_cc0_p (last1))
1514 last1 = beforelast1, last2 = beforelast2, ninsns--;
1515 #endif
1517 if (ninsns)
1519 *f1 = last1;
1520 *f2 = last2;
1523 return ninsns;
1526 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1527 the branch instruction. This means that if we commonize the control
1528 flow before end of the basic block, the semantic remains unchanged.
1530 We may assume that there exists one edge with a common destination. */
1532 static bool
1533 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1535 int nehedges1 = 0, nehedges2 = 0;
1536 edge fallthru1 = 0, fallthru2 = 0;
1537 edge e1, e2;
1538 edge_iterator ei;
1540 /* If we performed shrink-wrapping, edges to the exit block can
1541 only be distinguished for JUMP_INSNs. The two paths may differ in
1542 whether they went through the prologue. Sibcalls are fine, we know
1543 that we either didn't need or inserted an epilogue before them. */
1544 if (crtl->shrink_wrapped
1545 && single_succ_p (bb1)
1546 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1547 && !JUMP_P (BB_END (bb1))
1548 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1549 return false;
1551 /* If BB1 has only one successor, we may be looking at either an
1552 unconditional jump, or a fake edge to exit. */
1553 if (single_succ_p (bb1)
1554 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1555 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1556 return (single_succ_p (bb2)
1557 && (single_succ_edge (bb2)->flags
1558 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1559 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1561 /* Match conditional jumps - this may get tricky when fallthru and branch
1562 edges are crossed. */
1563 if (EDGE_COUNT (bb1->succs) == 2
1564 && any_condjump_p (BB_END (bb1))
1565 && onlyjump_p (BB_END (bb1)))
1567 edge b1, f1, b2, f2;
1568 bool reverse, match;
1569 rtx set1, set2, cond1, cond2;
1570 enum rtx_code code1, code2;
1572 if (EDGE_COUNT (bb2->succs) != 2
1573 || !any_condjump_p (BB_END (bb2))
1574 || !onlyjump_p (BB_END (bb2)))
1575 return false;
1577 b1 = BRANCH_EDGE (bb1);
1578 b2 = BRANCH_EDGE (bb2);
1579 f1 = FALLTHRU_EDGE (bb1);
1580 f2 = FALLTHRU_EDGE (bb2);
1582 /* Get around possible forwarders on fallthru edges. Other cases
1583 should be optimized out already. */
1584 if (FORWARDER_BLOCK_P (f1->dest))
1585 f1 = single_succ_edge (f1->dest);
1587 if (FORWARDER_BLOCK_P (f2->dest))
1588 f2 = single_succ_edge (f2->dest);
1590 /* To simplify use of this function, return false if there are
1591 unneeded forwarder blocks. These will get eliminated later
1592 during cleanup_cfg. */
1593 if (FORWARDER_BLOCK_P (f1->dest)
1594 || FORWARDER_BLOCK_P (f2->dest)
1595 || FORWARDER_BLOCK_P (b1->dest)
1596 || FORWARDER_BLOCK_P (b2->dest))
1597 return false;
1599 if (f1->dest == f2->dest && b1->dest == b2->dest)
1600 reverse = false;
1601 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1602 reverse = true;
1603 else
1604 return false;
1606 set1 = pc_set (BB_END (bb1));
1607 set2 = pc_set (BB_END (bb2));
1608 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1609 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1610 reverse = !reverse;
1612 cond1 = XEXP (SET_SRC (set1), 0);
1613 cond2 = XEXP (SET_SRC (set2), 0);
1614 code1 = GET_CODE (cond1);
1615 if (reverse)
1616 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1617 else
1618 code2 = GET_CODE (cond2);
1620 if (code2 == UNKNOWN)
1621 return false;
1623 /* Verify codes and operands match. */
1624 match = ((code1 == code2
1625 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1626 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1627 || (code1 == swap_condition (code2)
1628 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1629 XEXP (cond2, 0))
1630 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1631 XEXP (cond2, 1))));
1633 /* If we return true, we will join the blocks. Which means that
1634 we will only have one branch prediction bit to work with. Thus
1635 we require the existing branches to have probabilities that are
1636 roughly similar. */
1637 if (match
1638 && optimize_bb_for_speed_p (bb1)
1639 && optimize_bb_for_speed_p (bb2))
1641 int prob2;
1643 if (b1->dest == b2->dest)
1644 prob2 = b2->probability;
1645 else
1646 /* Do not use f2 probability as f2 may be forwarded. */
1647 prob2 = REG_BR_PROB_BASE - b2->probability;
1649 /* Fail if the difference in probabilities is greater than 50%.
1650 This rules out two well-predicted branches with opposite
1651 outcomes. */
1652 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1654 if (dump_file)
1655 fprintf (dump_file,
1656 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1657 bb1->index, bb2->index, b1->probability, prob2);
1659 return false;
1663 if (dump_file && match)
1664 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1665 bb1->index, bb2->index);
1667 return match;
1670 /* Generic case - we are seeing a computed jump, table jump or trapping
1671 instruction. */
1673 /* Check whether there are tablejumps in the end of BB1 and BB2.
1674 Return true if they are identical. */
1676 rtx label1, label2;
1677 rtx table1, table2;
1679 if (tablejump_p (BB_END (bb1), &label1, &table1)
1680 && tablejump_p (BB_END (bb2), &label2, &table2)
1681 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1683 /* The labels should never be the same rtx. If they really are same
1684 the jump tables are same too. So disable crossjumping of blocks BB1
1685 and BB2 because when deleting the common insns in the end of BB1
1686 by delete_basic_block () the jump table would be deleted too. */
1687 /* If LABEL2 is referenced in BB1->END do not do anything
1688 because we would loose information when replacing
1689 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1690 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1692 /* Set IDENTICAL to true when the tables are identical. */
1693 bool identical = false;
1694 rtx p1, p2;
1696 p1 = PATTERN (table1);
1697 p2 = PATTERN (table2);
1698 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1700 identical = true;
1702 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1703 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1704 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1705 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1707 int i;
1709 identical = true;
1710 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1711 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1712 identical = false;
1715 if (identical)
1717 replace_label_data rr;
1718 bool match;
1720 /* Temporarily replace references to LABEL1 with LABEL2
1721 in BB1->END so that we could compare the instructions. */
1722 rr.r1 = label1;
1723 rr.r2 = label2;
1724 rr.update_label_nuses = false;
1725 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1727 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1728 == dir_both);
1729 if (dump_file && match)
1730 fprintf (dump_file,
1731 "Tablejumps in bb %i and %i match.\n",
1732 bb1->index, bb2->index);
1734 /* Set the original label in BB1->END because when deleting
1735 a block whose end is a tablejump, the tablejump referenced
1736 from the instruction is deleted too. */
1737 rr.r1 = label2;
1738 rr.r2 = label1;
1739 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1741 return match;
1744 return false;
1748 /* Find the last non-debug non-note instruction in each bb, except
1749 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1750 handles that case specially. old_insns_match_p does not handle
1751 other types of instruction notes. */
1752 rtx last1 = BB_END (bb1);
1753 rtx last2 = BB_END (bb2);
1754 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1755 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1756 last1 = PREV_INSN (last1);
1757 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1758 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1759 last2 = PREV_INSN (last2);
1760 gcc_assert (last1 && last2);
1762 /* First ensure that the instructions match. There may be many outgoing
1763 edges so this test is generally cheaper. */
1764 if (old_insns_match_p (mode, last1, last2) != dir_both)
1765 return false;
1767 /* Search the outgoing edges, ensure that the counts do match, find possible
1768 fallthru and exception handling edges since these needs more
1769 validation. */
1770 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1771 return false;
1773 bool nonfakeedges = false;
1774 FOR_EACH_EDGE (e1, ei, bb1->succs)
1776 e2 = EDGE_SUCC (bb2, ei.index);
1778 if ((e1->flags & EDGE_FAKE) == 0)
1779 nonfakeedges = true;
1781 if (e1->flags & EDGE_EH)
1782 nehedges1++;
1784 if (e2->flags & EDGE_EH)
1785 nehedges2++;
1787 if (e1->flags & EDGE_FALLTHRU)
1788 fallthru1 = e1;
1789 if (e2->flags & EDGE_FALLTHRU)
1790 fallthru2 = e2;
1793 /* If number of edges of various types does not match, fail. */
1794 if (nehedges1 != nehedges2
1795 || (fallthru1 != 0) != (fallthru2 != 0))
1796 return false;
1798 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1799 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1800 attempt to optimize, as the two basic blocks might have different
1801 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1802 traps there should be REG_ARG_SIZE notes, they could be missing
1803 for __builtin_unreachable () uses though. */
1804 if (!nonfakeedges
1805 && !ACCUMULATE_OUTGOING_ARGS
1806 && (!INSN_P (last1)
1807 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1808 return false;
1810 /* fallthru edges must be forwarded to the same destination. */
1811 if (fallthru1)
1813 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1814 ? single_succ (fallthru1->dest): fallthru1->dest);
1815 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1816 ? single_succ (fallthru2->dest): fallthru2->dest);
1818 if (d1 != d2)
1819 return false;
1822 /* Ensure the same EH region. */
1824 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1825 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1827 if (!n1 && n2)
1828 return false;
1830 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1831 return false;
1834 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1835 version of sequence abstraction. */
1836 FOR_EACH_EDGE (e1, ei, bb2->succs)
1838 edge e2;
1839 edge_iterator ei;
1840 basic_block d1 = e1->dest;
1842 if (FORWARDER_BLOCK_P (d1))
1843 d1 = EDGE_SUCC (d1, 0)->dest;
1845 FOR_EACH_EDGE (e2, ei, bb1->succs)
1847 basic_block d2 = e2->dest;
1848 if (FORWARDER_BLOCK_P (d2))
1849 d2 = EDGE_SUCC (d2, 0)->dest;
1850 if (d1 == d2)
1851 break;
1854 if (!e2)
1855 return false;
1858 return true;
1861 /* Returns true if BB basic block has a preserve label. */
1863 static bool
1864 block_has_preserve_label (basic_block bb)
1866 return (bb
1867 && block_label (bb)
1868 && LABEL_PRESERVE_P (block_label (bb)));
1871 /* E1 and E2 are edges with the same destination block. Search their
1872 predecessors for common code. If found, redirect control flow from
1873 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1874 or the other way around (dir_backward). DIR specifies the allowed
1875 replacement direction. */
1877 static bool
1878 try_crossjump_to_edge (int mode, edge e1, edge e2,
1879 enum replace_direction dir)
1881 int nmatch;
1882 basic_block src1 = e1->src, src2 = e2->src;
1883 basic_block redirect_to, redirect_from, to_remove;
1884 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1885 rtx newpos1, newpos2;
1886 edge s;
1887 edge_iterator ei;
1889 newpos1 = newpos2 = NULL_RTX;
1891 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1892 to try this optimization.
1894 Basic block partitioning may result in some jumps that appear to
1895 be optimizable (or blocks that appear to be mergeable), but which really
1896 must be left untouched (they are required to make it safely across
1897 partition boundaries). See the comments at the top of
1898 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1900 if (crtl->has_bb_partition && reload_completed)
1901 return false;
1903 /* Search backward through forwarder blocks. We don't need to worry
1904 about multiple entry or chained forwarders, as they will be optimized
1905 away. We do this to look past the unconditional jump following a
1906 conditional jump that is required due to the current CFG shape. */
1907 if (single_pred_p (src1)
1908 && FORWARDER_BLOCK_P (src1))
1909 e1 = single_pred_edge (src1), src1 = e1->src;
1911 if (single_pred_p (src2)
1912 && FORWARDER_BLOCK_P (src2))
1913 e2 = single_pred_edge (src2), src2 = e2->src;
1915 /* Nothing to do if we reach ENTRY, or a common source block. */
1916 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1917 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1918 return false;
1919 if (src1 == src2)
1920 return false;
1922 /* Seeing more than 1 forwarder blocks would confuse us later... */
1923 if (FORWARDER_BLOCK_P (e1->dest)
1924 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1925 return false;
1927 if (FORWARDER_BLOCK_P (e2->dest)
1928 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1929 return false;
1931 /* Likewise with dead code (possibly newly created by the other optimizations
1932 of cfg_cleanup). */
1933 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1934 return false;
1936 /* Look for the common insn sequence, part the first ... */
1937 if (!outgoing_edges_match (mode, src1, src2))
1938 return false;
1940 /* ... and part the second. */
1941 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1943 osrc1 = src1;
1944 osrc2 = src2;
1945 if (newpos1 != NULL_RTX)
1946 src1 = BLOCK_FOR_INSN (newpos1);
1947 if (newpos2 != NULL_RTX)
1948 src2 = BLOCK_FOR_INSN (newpos2);
1950 if (dir == dir_backward)
1952 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1953 SWAP (basic_block, osrc1, osrc2);
1954 SWAP (basic_block, src1, src2);
1955 SWAP (edge, e1, e2);
1956 SWAP (rtx, newpos1, newpos2);
1957 #undef SWAP
1960 /* Don't proceed with the crossjump unless we found a sufficient number
1961 of matching instructions or the 'from' block was totally matched
1962 (such that its predecessors will hopefully be redirected and the
1963 block removed). */
1964 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
1965 && (newpos1 != BB_HEAD (src1)))
1966 return false;
1968 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1969 if (block_has_preserve_label (e1->dest)
1970 && (e1->flags & EDGE_ABNORMAL))
1971 return false;
1973 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1974 will be deleted.
1975 If we have tablejumps in the end of SRC1 and SRC2
1976 they have been already compared for equivalence in outgoing_edges_match ()
1977 so replace the references to TABLE1 by references to TABLE2. */
1979 rtx label1, label2;
1980 rtx table1, table2;
1982 if (tablejump_p (BB_END (osrc1), &label1, &table1)
1983 && tablejump_p (BB_END (osrc2), &label2, &table2)
1984 && label1 != label2)
1986 replace_label_data rr;
1987 rtx insn;
1989 /* Replace references to LABEL1 with LABEL2. */
1990 rr.r1 = label1;
1991 rr.r2 = label2;
1992 rr.update_label_nuses = true;
1993 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
1995 /* Do not replace the label in SRC1->END because when deleting
1996 a block whose end is a tablejump, the tablejump referenced
1997 from the instruction is deleted too. */
1998 if (insn != BB_END (osrc1))
1999 for_each_rtx (&insn, replace_label, &rr);
2004 /* Avoid splitting if possible. We must always split when SRC2 has
2005 EH predecessor edges, or we may end up with basic blocks with both
2006 normal and EH predecessor edges. */
2007 if (newpos2 == BB_HEAD (src2)
2008 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2009 redirect_to = src2;
2010 else
2012 if (newpos2 == BB_HEAD (src2))
2014 /* Skip possible basic block header. */
2015 if (LABEL_P (newpos2))
2016 newpos2 = NEXT_INSN (newpos2);
2017 while (DEBUG_INSN_P (newpos2))
2018 newpos2 = NEXT_INSN (newpos2);
2019 if (NOTE_P (newpos2))
2020 newpos2 = NEXT_INSN (newpos2);
2021 while (DEBUG_INSN_P (newpos2))
2022 newpos2 = NEXT_INSN (newpos2);
2025 if (dump_file)
2026 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2027 src2->index, nmatch);
2028 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2031 if (dump_file)
2032 fprintf (dump_file,
2033 "Cross jumping from bb %i to bb %i; %i common insns\n",
2034 src1->index, src2->index, nmatch);
2036 /* We may have some registers visible through the block. */
2037 df_set_bb_dirty (redirect_to);
2039 if (osrc2 == src2)
2040 redirect_edges_to = redirect_to;
2041 else
2042 redirect_edges_to = osrc2;
2044 /* Recompute the frequencies and counts of outgoing edges. */
2045 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2047 edge s2;
2048 edge_iterator ei;
2049 basic_block d = s->dest;
2051 if (FORWARDER_BLOCK_P (d))
2052 d = single_succ (d);
2054 FOR_EACH_EDGE (s2, ei, src1->succs)
2056 basic_block d2 = s2->dest;
2057 if (FORWARDER_BLOCK_P (d2))
2058 d2 = single_succ (d2);
2059 if (d == d2)
2060 break;
2063 s->count += s2->count;
2065 /* Take care to update possible forwarder blocks. We verified
2066 that there is no more than one in the chain, so we can't run
2067 into infinite loop. */
2068 if (FORWARDER_BLOCK_P (s->dest))
2070 single_succ_edge (s->dest)->count += s2->count;
2071 s->dest->count += s2->count;
2072 s->dest->frequency += EDGE_FREQUENCY (s);
2075 if (FORWARDER_BLOCK_P (s2->dest))
2077 single_succ_edge (s2->dest)->count -= s2->count;
2078 if (single_succ_edge (s2->dest)->count < 0)
2079 single_succ_edge (s2->dest)->count = 0;
2080 s2->dest->count -= s2->count;
2081 s2->dest->frequency -= EDGE_FREQUENCY (s);
2082 if (s2->dest->frequency < 0)
2083 s2->dest->frequency = 0;
2084 if (s2->dest->count < 0)
2085 s2->dest->count = 0;
2088 if (!redirect_edges_to->frequency && !src1->frequency)
2089 s->probability = (s->probability + s2->probability) / 2;
2090 else
2091 s->probability
2092 = ((s->probability * redirect_edges_to->frequency +
2093 s2->probability * src1->frequency)
2094 / (redirect_edges_to->frequency + src1->frequency));
2097 /* Adjust count and frequency for the block. An earlier jump
2098 threading pass may have left the profile in an inconsistent
2099 state (see update_bb_profile_for_threading) so we must be
2100 prepared for overflows. */
2101 tmp = redirect_to;
2104 tmp->count += src1->count;
2105 tmp->frequency += src1->frequency;
2106 if (tmp->frequency > BB_FREQ_MAX)
2107 tmp->frequency = BB_FREQ_MAX;
2108 if (tmp == redirect_edges_to)
2109 break;
2110 tmp = find_fallthru_edge (tmp->succs)->dest;
2112 while (true);
2113 update_br_prob_note (redirect_edges_to);
2115 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2117 /* Skip possible basic block header. */
2118 if (LABEL_P (newpos1))
2119 newpos1 = NEXT_INSN (newpos1);
2121 while (DEBUG_INSN_P (newpos1))
2122 newpos1 = NEXT_INSN (newpos1);
2124 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2125 newpos1 = NEXT_INSN (newpos1);
2127 while (DEBUG_INSN_P (newpos1))
2128 newpos1 = NEXT_INSN (newpos1);
2130 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2131 to_remove = single_succ (redirect_from);
2133 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2134 delete_basic_block (to_remove);
2136 update_forwarder_flag (redirect_from);
2137 if (redirect_to != src2)
2138 update_forwarder_flag (src2);
2140 return true;
2143 /* Search the predecessors of BB for common insn sequences. When found,
2144 share code between them by redirecting control flow. Return true if
2145 any changes made. */
2147 static bool
2148 try_crossjump_bb (int mode, basic_block bb)
2150 edge e, e2, fallthru;
2151 bool changed;
2152 unsigned max, ix, ix2;
2154 /* Nothing to do if there is not at least two incoming edges. */
2155 if (EDGE_COUNT (bb->preds) < 2)
2156 return false;
2158 /* Don't crossjump if this block ends in a computed jump,
2159 unless we are optimizing for size. */
2160 if (optimize_bb_for_size_p (bb)
2161 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2162 && computed_jump_p (BB_END (bb)))
2163 return false;
2165 /* If we are partitioning hot/cold basic blocks, we don't want to
2166 mess up unconditional or indirect jumps that cross between hot
2167 and cold sections.
2169 Basic block partitioning may result in some jumps that appear to
2170 be optimizable (or blocks that appear to be mergeable), but which really
2171 must be left untouched (they are required to make it safely across
2172 partition boundaries). See the comments at the top of
2173 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2175 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2176 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2177 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2178 return false;
2180 /* It is always cheapest to redirect a block that ends in a branch to
2181 a block that falls through into BB, as that adds no branches to the
2182 program. We'll try that combination first. */
2183 fallthru = NULL;
2184 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2186 if (EDGE_COUNT (bb->preds) > max)
2187 return false;
2189 fallthru = find_fallthru_edge (bb->preds);
2191 changed = false;
2192 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2194 e = EDGE_PRED (bb, ix);
2195 ix++;
2197 /* As noted above, first try with the fallthru predecessor (or, a
2198 fallthru predecessor if we are in cfglayout mode). */
2199 if (fallthru)
2201 /* Don't combine the fallthru edge into anything else.
2202 If there is a match, we'll do it the other way around. */
2203 if (e == fallthru)
2204 continue;
2205 /* If nothing changed since the last attempt, there is nothing
2206 we can do. */
2207 if (!first_pass
2208 && !((e->src->flags & BB_MODIFIED)
2209 || (fallthru->src->flags & BB_MODIFIED)))
2210 continue;
2212 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2214 changed = true;
2215 ix = 0;
2216 continue;
2220 /* Non-obvious work limiting check: Recognize that we're going
2221 to call try_crossjump_bb on every basic block. So if we have
2222 two blocks with lots of outgoing edges (a switch) and they
2223 share lots of common destinations, then we would do the
2224 cross-jump check once for each common destination.
2226 Now, if the blocks actually are cross-jump candidates, then
2227 all of their destinations will be shared. Which means that
2228 we only need check them for cross-jump candidacy once. We
2229 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2230 choosing to do the check from the block for which the edge
2231 in question is the first successor of A. */
2232 if (EDGE_SUCC (e->src, 0) != e)
2233 continue;
2235 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2237 e2 = EDGE_PRED (bb, ix2);
2239 if (e2 == e)
2240 continue;
2242 /* We've already checked the fallthru edge above. */
2243 if (e2 == fallthru)
2244 continue;
2246 /* The "first successor" check above only prevents multiple
2247 checks of crossjump(A,B). In order to prevent redundant
2248 checks of crossjump(B,A), require that A be the block
2249 with the lowest index. */
2250 if (e->src->index > e2->src->index)
2251 continue;
2253 /* If nothing changed since the last attempt, there is nothing
2254 we can do. */
2255 if (!first_pass
2256 && !((e->src->flags & BB_MODIFIED)
2257 || (e2->src->flags & BB_MODIFIED)))
2258 continue;
2260 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2261 direction. */
2262 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2264 changed = true;
2265 ix = 0;
2266 break;
2271 if (changed)
2272 crossjumps_occured = true;
2274 return changed;
2277 /* Search the successors of BB for common insn sequences. When found,
2278 share code between them by moving it across the basic block
2279 boundary. Return true if any changes made. */
2281 static bool
2282 try_head_merge_bb (basic_block bb)
2284 basic_block final_dest_bb = NULL;
2285 int max_match = INT_MAX;
2286 edge e0;
2287 rtx *headptr, *currptr, *nextptr;
2288 bool changed, moveall;
2289 unsigned ix;
2290 rtx e0_last_head, cond, move_before;
2291 unsigned nedges = EDGE_COUNT (bb->succs);
2292 rtx jump = BB_END (bb);
2293 regset live, live_union;
2295 /* Nothing to do if there is not at least two outgoing edges. */
2296 if (nedges < 2)
2297 return false;
2299 /* Don't crossjump if this block ends in a computed jump,
2300 unless we are optimizing for size. */
2301 if (optimize_bb_for_size_p (bb)
2302 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2303 && computed_jump_p (BB_END (bb)))
2304 return false;
2306 cond = get_condition (jump, &move_before, true, false);
2307 if (cond == NULL_RTX)
2309 #ifdef HAVE_cc0
2310 if (reg_mentioned_p (cc0_rtx, jump))
2311 move_before = prev_nonnote_nondebug_insn (jump);
2312 else
2313 #endif
2314 move_before = jump;
2317 for (ix = 0; ix < nedges; ix++)
2318 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2319 return false;
2321 for (ix = 0; ix < nedges; ix++)
2323 edge e = EDGE_SUCC (bb, ix);
2324 basic_block other_bb = e->dest;
2326 if (df_get_bb_dirty (other_bb))
2328 block_was_dirty = true;
2329 return false;
2332 if (e->flags & EDGE_ABNORMAL)
2333 return false;
2335 /* Normally, all destination blocks must only be reachable from this
2336 block, i.e. they must have one incoming edge.
2338 There is one special case we can handle, that of multiple consecutive
2339 jumps where the first jumps to one of the targets of the second jump.
2340 This happens frequently in switch statements for default labels.
2341 The structure is as follows:
2342 FINAL_DEST_BB
2343 ....
2344 if (cond) jump A;
2345 fall through
2347 jump with targets A, B, C, D...
2349 has two incoming edges, from FINAL_DEST_BB and BB
2351 In this case, we can try to move the insns through BB and into
2352 FINAL_DEST_BB. */
2353 if (EDGE_COUNT (other_bb->preds) != 1)
2355 edge incoming_edge, incoming_bb_other_edge;
2356 edge_iterator ei;
2358 if (final_dest_bb != NULL
2359 || EDGE_COUNT (other_bb->preds) != 2)
2360 return false;
2362 /* We must be able to move the insns across the whole block. */
2363 move_before = BB_HEAD (bb);
2364 while (!NONDEBUG_INSN_P (move_before))
2365 move_before = NEXT_INSN (move_before);
2367 if (EDGE_COUNT (bb->preds) != 1)
2368 return false;
2369 incoming_edge = EDGE_PRED (bb, 0);
2370 final_dest_bb = incoming_edge->src;
2371 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2372 return false;
2373 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2374 if (incoming_bb_other_edge != incoming_edge)
2375 break;
2376 if (incoming_bb_other_edge->dest != other_bb)
2377 return false;
2381 e0 = EDGE_SUCC (bb, 0);
2382 e0_last_head = NULL_RTX;
2383 changed = false;
2385 for (ix = 1; ix < nedges; ix++)
2387 edge e = EDGE_SUCC (bb, ix);
2388 rtx e0_last, e_last;
2389 int nmatch;
2391 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2392 &e0_last, &e_last, 0);
2393 if (nmatch == 0)
2394 return false;
2396 if (nmatch < max_match)
2398 max_match = nmatch;
2399 e0_last_head = e0_last;
2403 /* If we matched an entire block, we probably have to avoid moving the
2404 last insn. */
2405 if (max_match > 0
2406 && e0_last_head == BB_END (e0->dest)
2407 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2408 || control_flow_insn_p (e0_last_head)))
2410 max_match--;
2411 if (max_match == 0)
2412 return false;
2414 e0_last_head = prev_real_insn (e0_last_head);
2415 while (DEBUG_INSN_P (e0_last_head));
2418 if (max_match == 0)
2419 return false;
2421 /* We must find a union of the live registers at each of the end points. */
2422 live = BITMAP_ALLOC (NULL);
2423 live_union = BITMAP_ALLOC (NULL);
2425 currptr = XNEWVEC (rtx, nedges);
2426 headptr = XNEWVEC (rtx, nedges);
2427 nextptr = XNEWVEC (rtx, nedges);
2429 for (ix = 0; ix < nedges; ix++)
2431 int j;
2432 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2433 rtx head = BB_HEAD (merge_bb);
2435 while (!NONDEBUG_INSN_P (head))
2436 head = NEXT_INSN (head);
2437 headptr[ix] = head;
2438 currptr[ix] = head;
2440 /* Compute the end point and live information */
2441 for (j = 1; j < max_match; j++)
2443 head = NEXT_INSN (head);
2444 while (!NONDEBUG_INSN_P (head));
2445 simulate_backwards_to_point (merge_bb, live, head);
2446 IOR_REG_SET (live_union, live);
2449 /* If we're moving across two blocks, verify the validity of the
2450 first move, then adjust the target and let the loop below deal
2451 with the final move. */
2452 if (final_dest_bb != NULL)
2454 rtx move_upto;
2456 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2457 jump, e0->dest, live_union,
2458 NULL, &move_upto);
2459 if (!moveall)
2461 if (move_upto == NULL_RTX)
2462 goto out;
2464 while (e0_last_head != move_upto)
2466 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2467 live_union);
2468 e0_last_head = PREV_INSN (e0_last_head);
2471 if (e0_last_head == NULL_RTX)
2472 goto out;
2474 jump = BB_END (final_dest_bb);
2475 cond = get_condition (jump, &move_before, true, false);
2476 if (cond == NULL_RTX)
2478 #ifdef HAVE_cc0
2479 if (reg_mentioned_p (cc0_rtx, jump))
2480 move_before = prev_nonnote_nondebug_insn (jump);
2481 else
2482 #endif
2483 move_before = jump;
2489 rtx move_upto;
2490 moveall = can_move_insns_across (currptr[0], e0_last_head,
2491 move_before, jump, e0->dest, live_union,
2492 NULL, &move_upto);
2493 if (!moveall && move_upto == NULL_RTX)
2495 if (jump == move_before)
2496 break;
2498 /* Try again, using a different insertion point. */
2499 move_before = jump;
2501 #ifdef HAVE_cc0
2502 /* Don't try moving before a cc0 user, as that may invalidate
2503 the cc0. */
2504 if (reg_mentioned_p (cc0_rtx, jump))
2505 break;
2506 #endif
2508 continue;
2511 if (final_dest_bb && !moveall)
2512 /* We haven't checked whether a partial move would be OK for the first
2513 move, so we have to fail this case. */
2514 break;
2516 changed = true;
2517 for (;;)
2519 if (currptr[0] == move_upto)
2520 break;
2521 for (ix = 0; ix < nedges; ix++)
2523 rtx curr = currptr[ix];
2525 curr = NEXT_INSN (curr);
2526 while (!NONDEBUG_INSN_P (curr));
2527 currptr[ix] = curr;
2531 /* If we can't currently move all of the identical insns, remember
2532 each insn after the range that we'll merge. */
2533 if (!moveall)
2534 for (ix = 0; ix < nedges; ix++)
2536 rtx curr = currptr[ix];
2538 curr = NEXT_INSN (curr);
2539 while (!NONDEBUG_INSN_P (curr));
2540 nextptr[ix] = curr;
2543 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2544 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2545 if (final_dest_bb != NULL)
2546 df_set_bb_dirty (final_dest_bb);
2547 df_set_bb_dirty (bb);
2548 for (ix = 1; ix < nedges; ix++)
2550 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2551 delete_insn_chain (headptr[ix], currptr[ix], false);
2553 if (!moveall)
2555 if (jump == move_before)
2556 break;
2558 /* For the unmerged insns, try a different insertion point. */
2559 move_before = jump;
2561 #ifdef HAVE_cc0
2562 /* Don't try moving before a cc0 user, as that may invalidate
2563 the cc0. */
2564 if (reg_mentioned_p (cc0_rtx, jump))
2565 break;
2566 #endif
2568 for (ix = 0; ix < nedges; ix++)
2569 currptr[ix] = headptr[ix] = nextptr[ix];
2572 while (!moveall);
2574 out:
2575 free (currptr);
2576 free (headptr);
2577 free (nextptr);
2579 crossjumps_occured |= changed;
2581 return changed;
2584 /* Return true if BB contains just bb note, or bb note followed
2585 by only DEBUG_INSNs. */
2587 static bool
2588 trivially_empty_bb_p (basic_block bb)
2590 rtx insn = BB_END (bb);
2592 while (1)
2594 if (insn == BB_HEAD (bb))
2595 return true;
2596 if (!DEBUG_INSN_P (insn))
2597 return false;
2598 insn = PREV_INSN (insn);
2602 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2603 instructions etc. Return nonzero if changes were made. */
2605 static bool
2606 try_optimize_cfg (int mode)
2608 bool changed_overall = false;
2609 bool changed;
2610 int iterations = 0;
2611 basic_block bb, b, next;
2613 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2614 clear_bb_flags ();
2616 crossjumps_occured = false;
2618 FOR_EACH_BB_FN (bb, cfun)
2619 update_forwarder_flag (bb);
2621 if (! targetm.cannot_modify_jumps_p ())
2623 first_pass = true;
2624 /* Attempt to merge blocks as made possible by edge removal. If
2625 a block has only one successor, and the successor has only
2626 one predecessor, they may be combined. */
2629 block_was_dirty = false;
2630 changed = false;
2631 iterations++;
2633 if (dump_file)
2634 fprintf (dump_file,
2635 "\n\ntry_optimize_cfg iteration %i\n\n",
2636 iterations);
2638 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2639 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2641 basic_block c;
2642 edge s;
2643 bool changed_here = false;
2645 /* Delete trivially dead basic blocks. This is either
2646 blocks with no predecessors, or empty blocks with no
2647 successors. However if the empty block with no
2648 successors is the successor of the ENTRY_BLOCK, it is
2649 kept. This ensures that the ENTRY_BLOCK will have a
2650 successor which is a precondition for many RTL
2651 passes. Empty blocks may result from expanding
2652 __builtin_unreachable (). */
2653 if (EDGE_COUNT (b->preds) == 0
2654 || (EDGE_COUNT (b->succs) == 0
2655 && trivially_empty_bb_p (b)
2656 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2657 != b))
2659 c = b->prev_bb;
2660 if (EDGE_COUNT (b->preds) > 0)
2662 edge e;
2663 edge_iterator ei;
2665 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2667 if (BB_FOOTER (b)
2668 && BARRIER_P (BB_FOOTER (b)))
2669 FOR_EACH_EDGE (e, ei, b->preds)
2670 if ((e->flags & EDGE_FALLTHRU)
2671 && BB_FOOTER (e->src) == NULL)
2673 if (BB_FOOTER (b))
2675 BB_FOOTER (e->src) = BB_FOOTER (b);
2676 BB_FOOTER (b) = NULL;
2678 else
2680 start_sequence ();
2681 BB_FOOTER (e->src) = emit_barrier ();
2682 end_sequence ();
2686 else
2688 rtx last = get_last_bb_insn (b);
2689 if (last && BARRIER_P (last))
2690 FOR_EACH_EDGE (e, ei, b->preds)
2691 if ((e->flags & EDGE_FALLTHRU))
2692 emit_barrier_after (BB_END (e->src));
2695 delete_basic_block (b);
2696 changed = true;
2697 /* Avoid trying to remove the exit block. */
2698 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2699 continue;
2702 /* Remove code labels no longer used. */
2703 if (single_pred_p (b)
2704 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2705 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2706 && LABEL_P (BB_HEAD (b))
2707 /* If the previous block ends with a branch to this
2708 block, we can't delete the label. Normally this
2709 is a condjump that is yet to be simplified, but
2710 if CASE_DROPS_THRU, this can be a tablejump with
2711 some element going to the same place as the
2712 default (fallthru). */
2713 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2714 || !JUMP_P (BB_END (single_pred (b)))
2715 || ! label_is_jump_target_p (BB_HEAD (b),
2716 BB_END (single_pred (b)))))
2718 delete_insn (BB_HEAD (b));
2719 if (dump_file)
2720 fprintf (dump_file, "Deleted label in block %i.\n",
2721 b->index);
2724 /* If we fall through an empty block, we can remove it. */
2725 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2726 && single_pred_p (b)
2727 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2728 && !LABEL_P (BB_HEAD (b))
2729 && FORWARDER_BLOCK_P (b)
2730 /* Note that forwarder_block_p true ensures that
2731 there is a successor for this block. */
2732 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2733 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2735 if (dump_file)
2736 fprintf (dump_file,
2737 "Deleting fallthru block %i.\n",
2738 b->index);
2740 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2741 ? b->next_bb : b->prev_bb);
2742 redirect_edge_succ_nodup (single_pred_edge (b),
2743 single_succ (b));
2744 delete_basic_block (b);
2745 changed = true;
2746 b = c;
2747 continue;
2750 /* Merge B with its single successor, if any. */
2751 if (single_succ_p (b)
2752 && (s = single_succ_edge (b))
2753 && !(s->flags & EDGE_COMPLEX)
2754 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2755 && single_pred_p (c)
2756 && b != c)
2758 /* When not in cfg_layout mode use code aware of reordering
2759 INSN. This code possibly creates new basic blocks so it
2760 does not fit merge_blocks interface and is kept here in
2761 hope that it will become useless once more of compiler
2762 is transformed to use cfg_layout mode. */
2764 if ((mode & CLEANUP_CFGLAYOUT)
2765 && can_merge_blocks_p (b, c))
2767 merge_blocks (b, c);
2768 update_forwarder_flag (b);
2769 changed_here = true;
2771 else if (!(mode & CLEANUP_CFGLAYOUT)
2772 /* If the jump insn has side effects,
2773 we can't kill the edge. */
2774 && (!JUMP_P (BB_END (b))
2775 || (reload_completed
2776 ? simplejump_p (BB_END (b))
2777 : (onlyjump_p (BB_END (b))
2778 && !tablejump_p (BB_END (b),
2779 NULL, NULL))))
2780 && (next = merge_blocks_move (s, b, c, mode)))
2782 b = next;
2783 changed_here = true;
2787 /* Simplify branch over branch. */
2788 if ((mode & CLEANUP_EXPENSIVE)
2789 && !(mode & CLEANUP_CFGLAYOUT)
2790 && try_simplify_condjump (b))
2791 changed_here = true;
2793 /* If B has a single outgoing edge, but uses a
2794 non-trivial jump instruction without side-effects, we
2795 can either delete the jump entirely, or replace it
2796 with a simple unconditional jump. */
2797 if (single_succ_p (b)
2798 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2799 && onlyjump_p (BB_END (b))
2800 && !CROSSING_JUMP_P (BB_END (b))
2801 && try_redirect_by_replacing_jump (single_succ_edge (b),
2802 single_succ (b),
2803 (mode & CLEANUP_CFGLAYOUT) != 0))
2805 update_forwarder_flag (b);
2806 changed_here = true;
2809 /* Simplify branch to branch. */
2810 if (try_forward_edges (mode, b))
2812 update_forwarder_flag (b);
2813 changed_here = true;
2816 /* Look for shared code between blocks. */
2817 if ((mode & CLEANUP_CROSSJUMP)
2818 && try_crossjump_bb (mode, b))
2819 changed_here = true;
2821 if ((mode & CLEANUP_CROSSJUMP)
2822 /* This can lengthen register lifetimes. Do it only after
2823 reload. */
2824 && reload_completed
2825 && try_head_merge_bb (b))
2826 changed_here = true;
2828 /* Don't get confused by the index shift caused by
2829 deleting blocks. */
2830 if (!changed_here)
2831 b = b->next_bb;
2832 else
2833 changed = true;
2836 if ((mode & CLEANUP_CROSSJUMP)
2837 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
2838 changed = true;
2840 if (block_was_dirty)
2842 /* This should only be set by head-merging. */
2843 gcc_assert (mode & CLEANUP_CROSSJUMP);
2844 df_analyze ();
2847 if (changed)
2849 /* Edge forwarding in particular can cause hot blocks previously
2850 reached by both hot and cold blocks to become dominated only
2851 by cold blocks. This will cause the verification below to fail,
2852 and lead to now cold code in the hot section. This is not easy
2853 to detect and fix during edge forwarding, and in some cases
2854 is only visible after newly unreachable blocks are deleted,
2855 which will be done in fixup_partitions. */
2856 fixup_partitions ();
2858 #ifdef ENABLE_CHECKING
2859 verify_flow_info ();
2860 #endif
2863 changed_overall |= changed;
2864 first_pass = false;
2866 while (changed);
2869 FOR_ALL_BB_FN (b, cfun)
2870 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2872 return changed_overall;
2875 /* Delete all unreachable basic blocks. */
2877 bool
2878 delete_unreachable_blocks (void)
2880 bool changed = false;
2881 basic_block b, prev_bb;
2883 find_unreachable_blocks ();
2885 /* When we're in GIMPLE mode and there may be debug insns, we should
2886 delete blocks in reverse dominator order, so as to get a chance
2887 to substitute all released DEFs into debug stmts. If we don't
2888 have dominators information, walking blocks backward gets us a
2889 better chance of retaining most debug information than
2890 otherwise. */
2891 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2892 && dom_info_available_p (CDI_DOMINATORS))
2894 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2895 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2897 prev_bb = b->prev_bb;
2899 if (!(b->flags & BB_REACHABLE))
2901 /* Speed up the removal of blocks that don't dominate
2902 others. Walking backwards, this should be the common
2903 case. */
2904 if (!first_dom_son (CDI_DOMINATORS, b))
2905 delete_basic_block (b);
2906 else
2908 vec<basic_block> h
2909 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2911 while (h.length ())
2913 b = h.pop ();
2915 prev_bb = b->prev_bb;
2917 gcc_assert (!(b->flags & BB_REACHABLE));
2919 delete_basic_block (b);
2922 h.release ();
2925 changed = true;
2929 else
2931 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2932 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2934 prev_bb = b->prev_bb;
2936 if (!(b->flags & BB_REACHABLE))
2938 delete_basic_block (b);
2939 changed = true;
2944 if (changed)
2945 tidy_fallthru_edges ();
2946 return changed;
2949 /* Delete any jump tables never referenced. We can't delete them at the
2950 time of removing tablejump insn as they are referenced by the preceding
2951 insns computing the destination, so we delay deleting and garbagecollect
2952 them once life information is computed. */
2953 void
2954 delete_dead_jumptables (void)
2956 basic_block bb;
2958 /* A dead jump table does not belong to any basic block. Scan insns
2959 between two adjacent basic blocks. */
2960 FOR_EACH_BB_FN (bb, cfun)
2962 rtx insn, next;
2964 for (insn = NEXT_INSN (BB_END (bb));
2965 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
2966 insn = next)
2968 next = NEXT_INSN (insn);
2969 if (LABEL_P (insn)
2970 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
2971 && JUMP_TABLE_DATA_P (next))
2973 rtx label = insn, jump = next;
2975 if (dump_file)
2976 fprintf (dump_file, "Dead jumptable %i removed\n",
2977 INSN_UID (insn));
2979 next = NEXT_INSN (next);
2980 delete_insn (jump);
2981 delete_insn (label);
2988 /* Tidy the CFG by deleting unreachable code and whatnot. */
2990 bool
2991 cleanup_cfg (int mode)
2993 bool changed = false;
2995 /* Set the cfglayout mode flag here. We could update all the callers
2996 but that is just inconvenient, especially given that we eventually
2997 want to have cfglayout mode as the default. */
2998 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2999 mode |= CLEANUP_CFGLAYOUT;
3001 timevar_push (TV_CLEANUP_CFG);
3002 if (delete_unreachable_blocks ())
3004 changed = true;
3005 /* We've possibly created trivially dead code. Cleanup it right
3006 now to introduce more opportunities for try_optimize_cfg. */
3007 if (!(mode & (CLEANUP_NO_INSN_DEL))
3008 && !reload_completed)
3009 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3012 compact_blocks ();
3014 /* To tail-merge blocks ending in the same noreturn function (e.g.
3015 a call to abort) we have to insert fake edges to exit. Do this
3016 here once. The fake edges do not interfere with any other CFG
3017 cleanups. */
3018 if (mode & CLEANUP_CROSSJUMP)
3019 add_noreturn_fake_exit_edges ();
3021 if (!dbg_cnt (cfg_cleanup))
3022 return changed;
3024 while (try_optimize_cfg (mode))
3026 delete_unreachable_blocks (), changed = true;
3027 if (!(mode & CLEANUP_NO_INSN_DEL))
3029 /* Try to remove some trivially dead insns when doing an expensive
3030 cleanup. But delete_trivially_dead_insns doesn't work after
3031 reload (it only handles pseudos) and run_fast_dce is too costly
3032 to run in every iteration.
3034 For effective cross jumping, we really want to run a fast DCE to
3035 clean up any dead conditions, or they get in the way of performing
3036 useful tail merges.
3038 Other transformations in cleanup_cfg are not so sensitive to dead
3039 code, so delete_trivially_dead_insns or even doing nothing at all
3040 is good enough. */
3041 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3042 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3043 break;
3044 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
3045 run_fast_dce ();
3047 else
3048 break;
3051 if (mode & CLEANUP_CROSSJUMP)
3052 remove_fake_exit_edges ();
3054 /* Don't call delete_dead_jumptables in cfglayout mode, because
3055 that function assumes that jump tables are in the insns stream.
3056 But we also don't _have_ to delete dead jumptables in cfglayout
3057 mode because we shouldn't even be looking at things that are
3058 not in a basic block. Dead jumptables are cleaned up when
3059 going out of cfglayout mode. */
3060 if (!(mode & CLEANUP_CFGLAYOUT))
3061 delete_dead_jumptables ();
3063 /* ??? We probably do this way too often. */
3064 if (current_loops
3065 && (changed
3066 || (mode & CLEANUP_CFG_CHANGED)))
3068 timevar_push (TV_REPAIR_LOOPS);
3069 /* The above doesn't preserve dominance info if available. */
3070 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3071 calculate_dominance_info (CDI_DOMINATORS);
3072 fix_loop_structure (NULL);
3073 free_dominance_info (CDI_DOMINATORS);
3074 timevar_pop (TV_REPAIR_LOOPS);
3077 timevar_pop (TV_CLEANUP_CFG);
3079 return changed;
3082 namespace {
3084 const pass_data pass_data_jump =
3086 RTL_PASS, /* type */
3087 "jump", /* name */
3088 OPTGROUP_NONE, /* optinfo_flags */
3089 TV_JUMP, /* tv_id */
3090 0, /* properties_required */
3091 0, /* properties_provided */
3092 0, /* properties_destroyed */
3093 0, /* todo_flags_start */
3094 0, /* todo_flags_finish */
3097 class pass_jump : public rtl_opt_pass
3099 public:
3100 pass_jump (gcc::context *ctxt)
3101 : rtl_opt_pass (pass_data_jump, ctxt)
3104 /* opt_pass methods: */
3105 virtual unsigned int execute (function *);
3107 }; // class pass_jump
3109 unsigned int
3110 pass_jump::execute (function *)
3112 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3113 if (dump_file)
3114 dump_flow_info (dump_file, dump_flags);
3115 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3116 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3117 return 0;
3120 } // anon namespace
3122 rtl_opt_pass *
3123 make_pass_jump (gcc::context *ctxt)
3125 return new pass_jump (ctxt);
3128 namespace {
3130 const pass_data pass_data_jump2 =
3132 RTL_PASS, /* type */
3133 "jump2", /* name */
3134 OPTGROUP_NONE, /* optinfo_flags */
3135 TV_JUMP, /* tv_id */
3136 0, /* properties_required */
3137 0, /* properties_provided */
3138 0, /* properties_destroyed */
3139 0, /* todo_flags_start */
3140 0, /* todo_flags_finish */
3143 class pass_jump2 : public rtl_opt_pass
3145 public:
3146 pass_jump2 (gcc::context *ctxt)
3147 : rtl_opt_pass (pass_data_jump2, ctxt)
3150 /* opt_pass methods: */
3151 virtual unsigned int execute (function *)
3153 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3154 return 0;
3157 }; // class pass_jump2
3159 } // anon namespace
3161 rtl_opt_pass *
3162 make_pass_jump2 (gcc::context *ctxt)
3164 return new pass_jump2 (ctxt);