This patch rewrites the old VEC macro-based interface into a new one based on the...
[gcc.git] / gcc / cfgcleanup.c
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987, 1988, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
3 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2010, 2011
4 Free Software Foundation, Inc.
5
6 This file is part of GCC.
7
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
12
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
17
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
21
22 /* This file contains optimizer of the control flow. The main entry point is
23 cleanup_cfg. Following optimizations are performed:
24
25 - Unreachable blocks removal
26 - Edge forwarding (edge to the forwarder block is forwarded to its
27 successor. Simplification of the branch instruction is performed by
28 underlying infrastructure so branch can be converted to simplejump or
29 eliminated).
30 - Cross jumping (tail merging)
31 - Conditional jump-around-simplejump simplification
32 - Basic block merging. */
33
34 #include "config.h"
35 #include "system.h"
36 #include "coretypes.h"
37 #include "tm.h"
38 #include "rtl.h"
39 #include "hard-reg-set.h"
40 #include "regs.h"
41 #include "insn-config.h"
42 #include "flags.h"
43 #include "recog.h"
44 #include "diagnostic-core.h"
45 #include "cselib.h"
46 #include "params.h"
47 #include "tm_p.h"
48 #include "target.h"
49 #include "function.h" /* For inline functions in emit-rtl.h they need crtl. */
50 #include "emit-rtl.h"
51 #include "tree-pass.h"
52 #include "cfgloop.h"
53 #include "expr.h"
54 #include "df.h"
55 #include "dce.h"
56 #include "dbgcnt.h"
57
58 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
59
60 /* Set to true when we are running first pass of try_optimize_cfg loop. */
61 static bool first_pass;
62
63 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
64 static bool crossjumps_occured;
65
66 /* Set to true if we couldn't run an optimization due to stale liveness
67 information; we should run df_analyze to enable more opportunities. */
68 static bool block_was_dirty;
69
70 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
71 static bool try_crossjump_bb (int, basic_block);
72 static bool outgoing_edges_match (int, basic_block, basic_block);
73 static enum replace_direction old_insns_match_p (int, rtx, rtx);
74
75 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
76 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
77 static bool try_optimize_cfg (int);
78 static bool try_simplify_condjump (basic_block);
79 static bool try_forward_edges (int, basic_block);
80 static edge thread_jump (edge, basic_block);
81 static bool mark_effect (rtx, bitmap);
82 static void notice_new_block (basic_block);
83 static void update_forwarder_flag (basic_block);
84 static int mentions_nonequal_regs (rtx *, void *);
85 static void merge_memattrs (rtx, rtx);
86 \f
87 /* Set flags for newly created block. */
88
89 static void
90 notice_new_block (basic_block bb)
91 {
92 if (!bb)
93 return;
94
95 if (forwarder_block_p (bb))
96 bb->flags |= BB_FORWARDER_BLOCK;
97 }
98
99 /* Recompute forwarder flag after block has been modified. */
100
101 static void
102 update_forwarder_flag (basic_block bb)
103 {
104 if (forwarder_block_p (bb))
105 bb->flags |= BB_FORWARDER_BLOCK;
106 else
107 bb->flags &= ~BB_FORWARDER_BLOCK;
108 }
109 \f
110 /* Simplify a conditional jump around an unconditional jump.
111 Return true if something changed. */
112
113 static bool
114 try_simplify_condjump (basic_block cbranch_block)
115 {
116 basic_block jump_block, jump_dest_block, cbranch_dest_block;
117 edge cbranch_jump_edge, cbranch_fallthru_edge;
118 rtx cbranch_insn;
119
120 /* Verify that there are exactly two successors. */
121 if (EDGE_COUNT (cbranch_block->succs) != 2)
122 return false;
123
124 /* Verify that we've got a normal conditional branch at the end
125 of the block. */
126 cbranch_insn = BB_END (cbranch_block);
127 if (!any_condjump_p (cbranch_insn))
128 return false;
129
130 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
131 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
132
133 /* The next block must not have multiple predecessors, must not
134 be the last block in the function, and must contain just the
135 unconditional jump. */
136 jump_block = cbranch_fallthru_edge->dest;
137 if (!single_pred_p (jump_block)
138 || jump_block->next_bb == EXIT_BLOCK_PTR
139 || !FORWARDER_BLOCK_P (jump_block))
140 return false;
141 jump_dest_block = single_succ (jump_block);
142
143 /* If we are partitioning hot/cold basic blocks, we don't want to
144 mess up unconditional or indirect jumps that cross between hot
145 and cold sections.
146
147 Basic block partitioning may result in some jumps that appear to
148 be optimizable (or blocks that appear to be mergeable), but which really
149 must be left untouched (they are required to make it safely across
150 partition boundaries). See the comments at the top of
151 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
152
153 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
154 || (cbranch_jump_edge->flags & EDGE_CROSSING))
155 return false;
156
157 /* The conditional branch must target the block after the
158 unconditional branch. */
159 cbranch_dest_block = cbranch_jump_edge->dest;
160
161 if (cbranch_dest_block == EXIT_BLOCK_PTR
162 || !can_fallthru (jump_block, cbranch_dest_block))
163 return false;
164
165 /* Invert the conditional branch. */
166 if (!invert_jump (cbranch_insn, block_label (jump_dest_block), 0))
167 return false;
168
169 if (dump_file)
170 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
171 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
172
173 /* Success. Update the CFG to match. Note that after this point
174 the edge variable names appear backwards; the redirection is done
175 this way to preserve edge profile data. */
176 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
177 cbranch_dest_block);
178 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
179 jump_dest_block);
180 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
181 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
182 update_br_prob_note (cbranch_block);
183
184 /* Delete the block with the unconditional jump, and clean up the mess. */
185 delete_basic_block (jump_block);
186 tidy_fallthru_edge (cbranch_jump_edge);
187 update_forwarder_flag (cbranch_block);
188
189 return true;
190 }
191 \f
192 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
193 on register. Used by jump threading. */
194
195 static bool
196 mark_effect (rtx exp, regset nonequal)
197 {
198 int regno;
199 rtx dest;
200 switch (GET_CODE (exp))
201 {
202 /* In case we do clobber the register, mark it as equal, as we know the
203 value is dead so it don't have to match. */
204 case CLOBBER:
205 if (REG_P (XEXP (exp, 0)))
206 {
207 dest = XEXP (exp, 0);
208 regno = REGNO (dest);
209 if (HARD_REGISTER_NUM_P (regno))
210 bitmap_clear_range (nonequal, regno,
211 hard_regno_nregs[regno][GET_MODE (dest)]);
212 else
213 bitmap_clear_bit (nonequal, regno);
214 }
215 return false;
216
217 case SET:
218 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
219 return false;
220 dest = SET_DEST (exp);
221 if (dest == pc_rtx)
222 return false;
223 if (!REG_P (dest))
224 return true;
225 regno = REGNO (dest);
226 if (HARD_REGISTER_NUM_P (regno))
227 bitmap_set_range (nonequal, regno,
228 hard_regno_nregs[regno][GET_MODE (dest)]);
229 else
230 bitmap_set_bit (nonequal, regno);
231 return false;
232
233 default:
234 return false;
235 }
236 }
237
238 /* Return nonzero if X is a register set in regset DATA.
239 Called via for_each_rtx. */
240 static int
241 mentions_nonequal_regs (rtx *x, void *data)
242 {
243 regset nonequal = (regset) data;
244 if (REG_P (*x))
245 {
246 int regno;
247
248 regno = REGNO (*x);
249 if (REGNO_REG_SET_P (nonequal, regno))
250 return 1;
251 if (regno < FIRST_PSEUDO_REGISTER)
252 {
253 int n = hard_regno_nregs[regno][GET_MODE (*x)];
254 while (--n > 0)
255 if (REGNO_REG_SET_P (nonequal, regno + n))
256 return 1;
257 }
258 }
259 return 0;
260 }
261 /* Attempt to prove that the basic block B will have no side effects and
262 always continues in the same edge if reached via E. Return the edge
263 if exist, NULL otherwise. */
264
265 static edge
266 thread_jump (edge e, basic_block b)
267 {
268 rtx set1, set2, cond1, cond2, insn;
269 enum rtx_code code1, code2, reversed_code2;
270 bool reverse1 = false;
271 unsigned i;
272 regset nonequal;
273 bool failed = false;
274 reg_set_iterator rsi;
275
276 if (b->flags & BB_NONTHREADABLE_BLOCK)
277 return NULL;
278
279 /* At the moment, we do handle only conditional jumps, but later we may
280 want to extend this code to tablejumps and others. */
281 if (EDGE_COUNT (e->src->succs) != 2)
282 return NULL;
283 if (EDGE_COUNT (b->succs) != 2)
284 {
285 b->flags |= BB_NONTHREADABLE_BLOCK;
286 return NULL;
287 }
288
289 /* Second branch must end with onlyjump, as we will eliminate the jump. */
290 if (!any_condjump_p (BB_END (e->src)))
291 return NULL;
292
293 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
294 {
295 b->flags |= BB_NONTHREADABLE_BLOCK;
296 return NULL;
297 }
298
299 set1 = pc_set (BB_END (e->src));
300 set2 = pc_set (BB_END (b));
301 if (((e->flags & EDGE_FALLTHRU) != 0)
302 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
303 reverse1 = true;
304
305 cond1 = XEXP (SET_SRC (set1), 0);
306 cond2 = XEXP (SET_SRC (set2), 0);
307 if (reverse1)
308 code1 = reversed_comparison_code (cond1, BB_END (e->src));
309 else
310 code1 = GET_CODE (cond1);
311
312 code2 = GET_CODE (cond2);
313 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
314
315 if (!comparison_dominates_p (code1, code2)
316 && !comparison_dominates_p (code1, reversed_code2))
317 return NULL;
318
319 /* Ensure that the comparison operators are equivalent.
320 ??? This is far too pessimistic. We should allow swapped operands,
321 different CCmodes, or for example comparisons for interval, that
322 dominate even when operands are not equivalent. */
323 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
324 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
325 return NULL;
326
327 /* Short circuit cases where block B contains some side effects, as we can't
328 safely bypass it. */
329 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
330 insn = NEXT_INSN (insn))
331 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
332 {
333 b->flags |= BB_NONTHREADABLE_BLOCK;
334 return NULL;
335 }
336
337 cselib_init (0);
338
339 /* First process all values computed in the source basic block. */
340 for (insn = NEXT_INSN (BB_HEAD (e->src));
341 insn != NEXT_INSN (BB_END (e->src));
342 insn = NEXT_INSN (insn))
343 if (INSN_P (insn))
344 cselib_process_insn (insn);
345
346 nonequal = BITMAP_ALLOC (NULL);
347 CLEAR_REG_SET (nonequal);
348
349 /* Now assume that we've continued by the edge E to B and continue
350 processing as if it were same basic block.
351 Our goal is to prove that whole block is an NOOP. */
352
353 for (insn = NEXT_INSN (BB_HEAD (b));
354 insn != NEXT_INSN (BB_END (b)) && !failed;
355 insn = NEXT_INSN (insn))
356 {
357 if (INSN_P (insn))
358 {
359 rtx pat = PATTERN (insn);
360
361 if (GET_CODE (pat) == PARALLEL)
362 {
363 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
364 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
365 }
366 else
367 failed |= mark_effect (pat, nonequal);
368 }
369
370 cselib_process_insn (insn);
371 }
372
373 /* Later we should clear nonequal of dead registers. So far we don't
374 have life information in cfg_cleanup. */
375 if (failed)
376 {
377 b->flags |= BB_NONTHREADABLE_BLOCK;
378 goto failed_exit;
379 }
380
381 /* cond2 must not mention any register that is not equal to the
382 former block. */
383 if (for_each_rtx (&cond2, mentions_nonequal_regs, nonequal))
384 goto failed_exit;
385
386 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
387 goto failed_exit;
388
389 BITMAP_FREE (nonequal);
390 cselib_finish ();
391 if ((comparison_dominates_p (code1, code2) != 0)
392 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
393 return BRANCH_EDGE (b);
394 else
395 return FALLTHRU_EDGE (b);
396
397 failed_exit:
398 BITMAP_FREE (nonequal);
399 cselib_finish ();
400 return NULL;
401 }
402 \f
403 /* Attempt to forward edges leaving basic block B.
404 Return true if successful. */
405
406 static bool
407 try_forward_edges (int mode, basic_block b)
408 {
409 bool changed = false;
410 edge_iterator ei;
411 edge e, *threaded_edges = NULL;
412
413 /* If we are partitioning hot/cold basic blocks, we don't want to
414 mess up unconditional or indirect jumps that cross between hot
415 and cold sections.
416
417 Basic block partitioning may result in some jumps that appear to
418 be optimizable (or blocks that appear to be mergeable), but which really
419 must be left untouched (they are required to make it safely across
420 partition boundaries). See the comments at the top of
421 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
422
423 if (find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX))
424 return false;
425
426 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
427 {
428 basic_block target, first;
429 int counter, goto_locus;
430 bool threaded = false;
431 int nthreaded_edges = 0;
432 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
433
434 /* Skip complex edges because we don't know how to update them.
435
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)
440 {
441 ei_next (&ei);
442 continue;
443 }
444
445 target = first = e->dest;
446 counter = NUM_FIXED_BLOCKS;
447 goto_locus = e->goto_locus;
448
449 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
450 up jumps that cross between hot/cold sections.
451
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. */
458
459 if (first != EXIT_BLOCK_PTR
460 && find_reg_note (BB_END (first), REG_CROSSING_JUMP, NULL_RTX))
461 return false;
462
463 while (counter < n_basic_blocks)
464 {
465 basic_block new_target = NULL;
466 bool new_target_threaded = false;
467 may_thread |= (target->flags & BB_MODIFIED) != 0;
468
469 if (FORWARDER_BLOCK_P (target)
470 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
471 && single_succ (target) != EXIT_BLOCK_PTR)
472 {
473 /* Bypass trivial infinite loops. */
474 new_target = single_succ (target);
475 if (target == new_target)
476 counter = n_basic_blocks;
477 else if (!optimize)
478 {
479 /* When not optimizing, ensure that edges or forwarder
480 blocks with different locus are not optimized out. */
481 int new_locus = single_succ_edge (target)->goto_locus;
482 int locus = goto_locus;
483
484 if (new_locus != UNKNOWN_LOCATION
485 && locus != UNKNOWN_LOCATION
486 && new_locus != locus)
487 new_target = NULL;
488 else
489 {
490 rtx last;
491
492 if (new_locus != UNKNOWN_LOCATION)
493 locus = new_locus;
494
495 last = BB_END (target);
496 if (DEBUG_INSN_P (last))
497 last = prev_nondebug_insn (last);
498
499 new_locus = last && INSN_P (last)
500 ? INSN_LOCATION (last) : 0;
501
502 if (new_locus != UNKNOWN_LOCATION
503 && locus != UNKNOWN_LOCATION
504 && new_locus != locus)
505 new_target = NULL;
506 else
507 {
508 if (new_locus != UNKNOWN_LOCATION)
509 locus = new_locus;
510
511 goto_locus = locus;
512 }
513 }
514 }
515 }
516
517 /* Allow to thread only over one edge at time to simplify updating
518 of probabilities. */
519 else if ((mode & CLEANUP_THREADING) && may_thread)
520 {
521 edge t = thread_jump (e, target);
522 if (t)
523 {
524 if (!threaded_edges)
525 threaded_edges = XNEWVEC (edge, n_basic_blocks);
526 else
527 {
528 int i;
529
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)
536 {
537 counter = n_basic_blocks;
538 break;
539 }
540 }
541
542 /* Detect an infinite loop across the start block. */
543 if (t->dest == b)
544 break;
545
546 gcc_assert (nthreaded_edges < n_basic_blocks - NUM_FIXED_BLOCKS);
547 threaded_edges[nthreaded_edges++] = t;
548
549 new_target = t->dest;
550 new_target_threaded = true;
551 }
552 }
553
554 if (!new_target)
555 break;
556
557 counter++;
558 target = new_target;
559 threaded |= new_target_threaded;
560 }
561
562 if (counter >= n_basic_blocks)
563 {
564 if (dump_file)
565 fprintf (dump_file, "Infinite loop in BB %i.\n",
566 target->index);
567 }
568 else if (target == first)
569 ; /* We didn't do anything. */
570 else
571 {
572 /* Save the values now, as the edge may get removed. */
573 gcov_type edge_count = e->count;
574 int edge_probability = e->probability;
575 int edge_frequency;
576 int n = 0;
577
578 e->goto_locus = goto_locus;
579
580 /* Don't force if target is exit block. */
581 if (threaded && target != EXIT_BLOCK_PTR)
582 {
583 notice_new_block (redirect_edge_and_branch_force (e, target));
584 if (dump_file)
585 fprintf (dump_file, "Conditionals threaded.\n");
586 }
587 else if (!redirect_edge_and_branch (e, target))
588 {
589 if (dump_file)
590 fprintf (dump_file,
591 "Forwarding edge %i->%i to %i failed.\n",
592 b->index, e->dest->index, target->index);
593 ei_next (&ei);
594 continue;
595 }
596
597 /* We successfully forwarded the edge. Now update profile
598 data: for each edge we traversed in the chain, remove
599 the original edge's execution count. */
600 edge_frequency = ((edge_probability * b->frequency
601 + REG_BR_PROB_BASE / 2)
602 / REG_BR_PROB_BASE);
603
604 do
605 {
606 edge t;
607
608 if (!single_succ_p (first))
609 {
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);
616 }
617 else
618 {
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);
633 }
634
635 t->count -= edge_count;
636 if (t->count < 0)
637 t->count = 0;
638 first = t->dest;
639 }
640 while (first != target);
641
642 changed = true;
643 continue;
644 }
645 ei_next (&ei);
646 }
647
648 free (threaded_edges);
649 return changed;
650 }
651 \f
652
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). */
656
657 static void
658 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
659 {
660 rtx barrier;
661
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.
665
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. */
671
672 if (BB_PARTITION (a) != BB_PARTITION (b))
673 return;
674
675 barrier = next_nonnote_insn (BB_END (a));
676 gcc_assert (BARRIER_P (barrier));
677 delete_insn (barrier);
678
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);
683
684 if (dump_file)
685 fprintf (dump_file, "Moved block %d before %d and merged.\n",
686 a->index, b->index);
687
688 /* Swap the records for the two blocks around. */
689
690 unlink_block (a);
691 link_block (a, b->prev_bb);
692
693 /* Now blocks A and B are contiguous. Merge them. */
694 merge_blocks (a, b);
695 }
696
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). */
700
701 static void
702 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
703 {
704 rtx barrier, real_b_end;
705 rtx label, table;
706
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.
710
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. */
716
717 if (BB_PARTITION (a) != BB_PARTITION (b))
718 return;
719
720 real_b_end = BB_END (b);
721
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))
726 {
727 BB_END (b) = table;
728 }
729
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);
734
735
736 /* Scramble the insn chain. */
737 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
738
739 /* Restore the real end of b. */
740 BB_END (b) = real_b_end;
741
742 if (dump_file)
743 fprintf (dump_file, "Moved block %d after %d and merged.\n",
744 b->index, a->index);
745
746 /* Now blocks A and B are contiguous. Merge them. */
747 merge_blocks (a, b);
748 }
749
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.
756
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. */
761
762 static basic_block
763 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
764 {
765 basic_block next;
766
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.
770
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. */
776
777 if (BB_PARTITION (b) != BB_PARTITION (c))
778 return NULL;
779
780 /* If B has a fallthru edge to C, no need to move anything. */
781 if (e->flags & EDGE_FALLTHRU)
782 {
783 int b_index = b->index, c_index = c->index;
784
785 /* Protect the loop latches. */
786 if (current_loops && c->loop_father->latch == c)
787 return NULL;
788
789 merge_blocks (b, c);
790 update_forwarder_flag (b);
791
792 if (dump_file)
793 fprintf (dump_file, "Merged %d and %d without moving.\n",
794 b_index, c_index);
795
796 return b->prev_bb == ENTRY_BLOCK_PTR ? b : b->prev_bb;
797 }
798
799 /* Otherwise we will need to move code around. Do that only if expensive
800 transformations are allowed. */
801 else if (mode & CLEANUP_EXPENSIVE)
802 {
803 edge tmp_edge, b_fallthru_edge;
804 bool c_has_outgoing_fallthru;
805 bool b_has_incoming_fallthru;
806
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;
813
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. */
817
818 tmp_edge = find_fallthru_edge (c->succs);
819 c_has_outgoing_fallthru = (tmp_edge != NULL);
820
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;
827
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)
832 {
833 merge_blocks_move_successor_nojumps (b, c);
834 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
835 }
836
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. */
841
842 if (b_has_incoming_fallthru)
843 {
844 basic_block bb;
845
846 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR)
847 return NULL;
848 bb = force_nonfallthru (b_fallthru_edge);
849 if (bb)
850 notice_new_block (bb);
851 }
852
853 merge_blocks_move_predecessor_nojumps (b, c);
854 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
855 }
856
857 return NULL;
858 }
859 \f
860
861 /* Removes the memory attributes of MEM expression
862 if they are not equal. */
863
864 void
865 merge_memattrs (rtx x, rtx y)
866 {
867 int i;
868 int j;
869 enum rtx_code code;
870 const char *fmt;
871
872 if (x == y)
873 return;
874 if (x == 0 || y == 0)
875 return;
876
877 code = GET_CODE (x);
878
879 if (code != GET_CODE (y))
880 return;
881
882 if (GET_MODE (x) != GET_MODE (y))
883 return;
884
885 if (code == MEM && MEM_ATTRS (x) != MEM_ATTRS (y))
886 {
887 if (! MEM_ATTRS (x))
888 MEM_ATTRS (y) = 0;
889 else if (! MEM_ATTRS (y))
890 MEM_ATTRS (x) = 0;
891 else
892 {
893 HOST_WIDE_INT mem_size;
894
895 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
896 {
897 set_mem_alias_set (x, 0);
898 set_mem_alias_set (y, 0);
899 }
900
901 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
902 {
903 set_mem_expr (x, 0);
904 set_mem_expr (y, 0);
905 clear_mem_offset (x);
906 clear_mem_offset (y);
907 }
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)))
911 {
912 clear_mem_offset (x);
913 clear_mem_offset (y);
914 }
915
916 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
917 {
918 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
919 set_mem_size (x, mem_size);
920 set_mem_size (y, mem_size);
921 }
922 else
923 {
924 clear_mem_size (x);
925 clear_mem_size (y);
926 }
927
928 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
929 set_mem_align (y, MEM_ALIGN (x));
930 }
931 }
932
933 fmt = GET_RTX_FORMAT (code);
934 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
935 {
936 switch (fmt[i])
937 {
938 case 'E':
939 /* Two vectors must have the same length. */
940 if (XVECLEN (x, i) != XVECLEN (y, i))
941 return;
942
943 for (j = 0; j < XVECLEN (x, i); j++)
944 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
945
946 break;
947
948 case 'e':
949 merge_memattrs (XEXP (x, i), XEXP (y, i));
950 }
951 }
952 return;
953 }
954
955
956 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
957 different single sets S1 and S2. */
958
959 static bool
960 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
961 {
962 int i;
963 rtx e1, e2;
964
965 if (p1 == s1 && p2 == s2)
966 return true;
967
968 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
969 return false;
970
971 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
972 return false;
973
974 for (i = 0; i < XVECLEN (p1, 0); i++)
975 {
976 e1 = XVECEXP (p1, 0, i);
977 e2 = XVECEXP (p2, 0, i);
978 if (e1 == s1 && e2 == s2)
979 continue;
980 if (reload_completed
981 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
982 continue;
983
984 return false;
985 }
986
987 return true;
988 }
989
990 /* Examine register notes on I1 and I2 and return:
991 - dir_forward if I1 can be replaced by I2, or
992 - dir_backward if I2 can be replaced by I1, or
993 - dir_both if both are the case. */
994
995 static enum replace_direction
996 can_replace_by (rtx i1, rtx i2)
997 {
998 rtx s1, s2, d1, d2, src1, src2, note1, note2;
999 bool c1, c2;
1000
1001 /* Check for 2 sets. */
1002 s1 = single_set (i1);
1003 s2 = single_set (i2);
1004 if (s1 == NULL_RTX || s2 == NULL_RTX)
1005 return dir_none;
1006
1007 /* Check that the 2 sets set the same dest. */
1008 d1 = SET_DEST (s1);
1009 d2 = SET_DEST (s2);
1010 if (!(reload_completed
1011 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1012 return dir_none;
1013
1014 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1015 set dest to the same value. */
1016 note1 = find_reg_equal_equiv_note (i1);
1017 note2 = find_reg_equal_equiv_note (i2);
1018 if (!note1 || !note2 || !rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0))
1019 || !CONST_INT_P (XEXP (note1, 0)))
1020 return dir_none;
1021
1022 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1023 return dir_none;
1024
1025 /* Although the 2 sets set dest to the same value, we cannot replace
1026 (set (dest) (const_int))
1027 by
1028 (set (dest) (reg))
1029 because we don't know if the reg is live and has the same value at the
1030 location of replacement. */
1031 src1 = SET_SRC (s1);
1032 src2 = SET_SRC (s2);
1033 c1 = CONST_INT_P (src1);
1034 c2 = CONST_INT_P (src2);
1035 if (c1 && c2)
1036 return dir_both;
1037 else if (c2)
1038 return dir_forward;
1039 else if (c1)
1040 return dir_backward;
1041
1042 return dir_none;
1043 }
1044
1045 /* Merges directions A and B. */
1046
1047 static enum replace_direction
1048 merge_dir (enum replace_direction a, enum replace_direction b)
1049 {
1050 /* Implements the following table:
1051 |bo fw bw no
1052 ---+-----------
1053 bo |bo fw bw no
1054 fw |-- fw no no
1055 bw |-- -- bw no
1056 no |-- -- -- no. */
1057
1058 if (a == b)
1059 return a;
1060
1061 if (a == dir_both)
1062 return b;
1063 if (b == dir_both)
1064 return a;
1065
1066 return dir_none;
1067 }
1068
1069 /* Examine I1 and I2 and return:
1070 - dir_forward if I1 can be replaced by I2, or
1071 - dir_backward if I2 can be replaced by I1, or
1072 - dir_both if both are the case. */
1073
1074 static enum replace_direction
1075 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx i1, rtx i2)
1076 {
1077 rtx p1, p2;
1078
1079 /* Verify that I1 and I2 are equivalent. */
1080 if (GET_CODE (i1) != GET_CODE (i2))
1081 return dir_none;
1082
1083 /* __builtin_unreachable() may lead to empty blocks (ending with
1084 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1085 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1086 return dir_both;
1087
1088 /* ??? Do not allow cross-jumping between different stack levels. */
1089 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1090 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1091 if (p1 && p2)
1092 {
1093 p1 = XEXP (p1, 0);
1094 p2 = XEXP (p2, 0);
1095 if (!rtx_equal_p (p1, p2))
1096 return dir_none;
1097
1098 /* ??? Worse, this adjustment had better be constant lest we
1099 have differing incoming stack levels. */
1100 if (!frame_pointer_needed
1101 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1102 return dir_none;
1103 }
1104 else if (p1 || p2)
1105 return dir_none;
1106
1107 p1 = PATTERN (i1);
1108 p2 = PATTERN (i2);
1109
1110 if (GET_CODE (p1) != GET_CODE (p2))
1111 return dir_none;
1112
1113 /* If this is a CALL_INSN, compare register usage information.
1114 If we don't check this on stack register machines, the two
1115 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1116 numbers of stack registers in the same basic block.
1117 If we don't check this on machines with delay slots, a delay slot may
1118 be filled that clobbers a parameter expected by the subroutine.
1119
1120 ??? We take the simple route for now and assume that if they're
1121 equal, they were constructed identically.
1122
1123 Also check for identical exception regions. */
1124
1125 if (CALL_P (i1))
1126 {
1127 /* Ensure the same EH region. */
1128 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1129 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1130
1131 if (!n1 && n2)
1132 return dir_none;
1133
1134 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1135 return dir_none;
1136
1137 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1138 CALL_INSN_FUNCTION_USAGE (i2))
1139 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1140 return dir_none;
1141 }
1142
1143 #ifdef STACK_REGS
1144 /* If cross_jump_death_matters is not 0, the insn's mode
1145 indicates whether or not the insn contains any stack-like
1146 regs. */
1147
1148 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1149 {
1150 /* If register stack conversion has already been done, then
1151 death notes must also be compared before it is certain that
1152 the two instruction streams match. */
1153
1154 rtx note;
1155 HARD_REG_SET i1_regset, i2_regset;
1156
1157 CLEAR_HARD_REG_SET (i1_regset);
1158 CLEAR_HARD_REG_SET (i2_regset);
1159
1160 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1161 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1162 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1163
1164 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1165 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1166 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1167
1168 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1169 return dir_none;
1170 }
1171 #endif
1172
1173 if (reload_completed
1174 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1175 return dir_both;
1176
1177 return can_replace_by (i1, i2);
1178 }
1179 \f
1180 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1181 flow_find_head_matching_sequence, ensure the notes match. */
1182
1183 static void
1184 merge_notes (rtx i1, rtx i2)
1185 {
1186 /* If the merged insns have different REG_EQUAL notes, then
1187 remove them. */
1188 rtx equiv1 = find_reg_equal_equiv_note (i1);
1189 rtx equiv2 = find_reg_equal_equiv_note (i2);
1190
1191 if (equiv1 && !equiv2)
1192 remove_note (i1, equiv1);
1193 else if (!equiv1 && equiv2)
1194 remove_note (i2, equiv2);
1195 else if (equiv1 && equiv2
1196 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1197 {
1198 remove_note (i1, equiv1);
1199 remove_note (i2, equiv2);
1200 }
1201 }
1202
1203 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1204 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1205 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1206 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1207 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1208
1209 static void
1210 walk_to_nondebug_insn (rtx *i1, basic_block *bb1, bool follow_fallthru,
1211 bool *did_fallthru)
1212 {
1213 edge fallthru;
1214
1215 *did_fallthru = false;
1216
1217 /* Ignore notes. */
1218 while (!NONDEBUG_INSN_P (*i1))
1219 {
1220 if (*i1 != BB_HEAD (*bb1))
1221 {
1222 *i1 = PREV_INSN (*i1);
1223 continue;
1224 }
1225
1226 if (!follow_fallthru)
1227 return;
1228
1229 fallthru = find_fallthru_edge ((*bb1)->preds);
1230 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun)
1231 || !single_succ_p (fallthru->src))
1232 return;
1233
1234 *bb1 = fallthru->src;
1235 *i1 = BB_END (*bb1);
1236 *did_fallthru = true;
1237 }
1238 }
1239
1240 /* Look through the insns at the end of BB1 and BB2 and find the longest
1241 sequence that are either equivalent, or allow forward or backward
1242 replacement. Store the first insns for that sequence in *F1 and *F2 and
1243 return the sequence length.
1244
1245 DIR_P indicates the allowed replacement direction on function entry, and
1246 the actual replacement direction on function exit. If NULL, only equivalent
1247 sequences are allowed.
1248
1249 To simplify callers of this function, if the blocks match exactly,
1250 store the head of the blocks in *F1 and *F2. */
1251
1252 int
1253 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx *f1, rtx *f2,
1254 enum replace_direction *dir_p)
1255 {
1256 rtx i1, i2, last1, last2, afterlast1, afterlast2;
1257 int ninsns = 0;
1258 rtx p1;
1259 enum replace_direction dir, last_dir, afterlast_dir;
1260 bool follow_fallthru, did_fallthru;
1261
1262 if (dir_p)
1263 dir = *dir_p;
1264 else
1265 dir = dir_both;
1266 afterlast_dir = dir;
1267 last_dir = afterlast_dir;
1268
1269 /* Skip simple jumps at the end of the blocks. Complex jumps still
1270 need to be compared for equivalence, which we'll do below. */
1271
1272 i1 = BB_END (bb1);
1273 last1 = afterlast1 = last2 = afterlast2 = NULL_RTX;
1274 if (onlyjump_p (i1)
1275 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1276 {
1277 last1 = i1;
1278 i1 = PREV_INSN (i1);
1279 }
1280
1281 i2 = BB_END (bb2);
1282 if (onlyjump_p (i2)
1283 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1284 {
1285 last2 = i2;
1286 /* Count everything except for unconditional jump as insn. */
1287 if (!simplejump_p (i2) && !returnjump_p (i2) && last1)
1288 ninsns++;
1289 i2 = PREV_INSN (i2);
1290 }
1291
1292 while (true)
1293 {
1294 /* In the following example, we can replace all jumps to C by jumps to A.
1295
1296 This removes 4 duplicate insns.
1297 [bb A] insn1 [bb C] insn1
1298 insn2 insn2
1299 [bb B] insn3 insn3
1300 insn4 insn4
1301 jump_insn jump_insn
1302
1303 We could also replace all jumps to A by jumps to C, but that leaves B
1304 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1305 step, all jumps to B would be replaced with jumps to the middle of C,
1306 achieving the same result with more effort.
1307 So we allow only the first possibility, which means that we don't allow
1308 fallthru in the block that's being replaced. */
1309
1310 follow_fallthru = dir_p && dir != dir_forward;
1311 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1312 if (did_fallthru)
1313 dir = dir_backward;
1314
1315 follow_fallthru = dir_p && dir != dir_backward;
1316 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1317 if (did_fallthru)
1318 dir = dir_forward;
1319
1320 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1321 break;
1322
1323 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1324 if (dir == dir_none || (!dir_p && dir != dir_both))
1325 break;
1326
1327 merge_memattrs (i1, i2);
1328
1329 /* Don't begin a cross-jump with a NOTE insn. */
1330 if (INSN_P (i1))
1331 {
1332 merge_notes (i1, i2);
1333
1334 afterlast1 = last1, afterlast2 = last2;
1335 last1 = i1, last2 = i2;
1336 afterlast_dir = last_dir;
1337 last_dir = dir;
1338 p1 = PATTERN (i1);
1339 if (!(GET_CODE (p1) == USE || GET_CODE (p1) == CLOBBER))
1340 ninsns++;
1341 }
1342
1343 i1 = PREV_INSN (i1);
1344 i2 = PREV_INSN (i2);
1345 }
1346
1347 #ifdef HAVE_cc0
1348 /* Don't allow the insn after a compare to be shared by
1349 cross-jumping unless the compare is also shared. */
1350 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
1351 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1352 #endif
1353
1354 /* Include preceding notes and labels in the cross-jump. One,
1355 this may bring us to the head of the blocks as requested above.
1356 Two, it keeps line number notes as matched as may be. */
1357 if (ninsns)
1358 {
1359 bb1 = BLOCK_FOR_INSN (last1);
1360 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1361 last1 = PREV_INSN (last1);
1362
1363 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1364 last1 = PREV_INSN (last1);
1365
1366 bb2 = BLOCK_FOR_INSN (last2);
1367 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1368 last2 = PREV_INSN (last2);
1369
1370 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1371 last2 = PREV_INSN (last2);
1372
1373 *f1 = last1;
1374 *f2 = last2;
1375 }
1376
1377 if (dir_p)
1378 *dir_p = last_dir;
1379 return ninsns;
1380 }
1381
1382 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1383 the head of the two blocks. Do not include jumps at the end.
1384 If STOP_AFTER is nonzero, stop after finding that many matching
1385 instructions. */
1386
1387 int
1388 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx *f1,
1389 rtx *f2, int stop_after)
1390 {
1391 rtx i1, i2, last1, last2, beforelast1, beforelast2;
1392 int ninsns = 0;
1393 edge e;
1394 edge_iterator ei;
1395 int nehedges1 = 0, nehedges2 = 0;
1396
1397 FOR_EACH_EDGE (e, ei, bb1->succs)
1398 if (e->flags & EDGE_EH)
1399 nehedges1++;
1400 FOR_EACH_EDGE (e, ei, bb2->succs)
1401 if (e->flags & EDGE_EH)
1402 nehedges2++;
1403
1404 i1 = BB_HEAD (bb1);
1405 i2 = BB_HEAD (bb2);
1406 last1 = beforelast1 = last2 = beforelast2 = NULL_RTX;
1407
1408 while (true)
1409 {
1410 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1411 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1412 {
1413 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1414 break;
1415 i1 = NEXT_INSN (i1);
1416 }
1417
1418 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1419 {
1420 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1421 break;
1422 i2 = NEXT_INSN (i2);
1423 }
1424
1425 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1426 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1427 break;
1428
1429 if (NOTE_P (i1) || NOTE_P (i2)
1430 || JUMP_P (i1) || JUMP_P (i2))
1431 break;
1432
1433 /* A sanity check to make sure we're not merging insns with different
1434 effects on EH. If only one of them ends a basic block, it shouldn't
1435 have an EH edge; if both end a basic block, there should be the same
1436 number of EH edges. */
1437 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1438 && nehedges1 > 0)
1439 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1440 && nehedges2 > 0)
1441 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1442 && nehedges1 != nehedges2))
1443 break;
1444
1445 if (old_insns_match_p (0, i1, i2) != dir_both)
1446 break;
1447
1448 merge_memattrs (i1, i2);
1449
1450 /* Don't begin a cross-jump with a NOTE insn. */
1451 if (INSN_P (i1))
1452 {
1453 merge_notes (i1, i2);
1454
1455 beforelast1 = last1, beforelast2 = last2;
1456 last1 = i1, last2 = i2;
1457 ninsns++;
1458 }
1459
1460 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1461 || (stop_after > 0 && ninsns == stop_after))
1462 break;
1463
1464 i1 = NEXT_INSN (i1);
1465 i2 = NEXT_INSN (i2);
1466 }
1467
1468 #ifdef HAVE_cc0
1469 /* Don't allow a compare to be shared by cross-jumping unless the insn
1470 after the compare is also shared. */
1471 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && sets_cc0_p (last1))
1472 last1 = beforelast1, last2 = beforelast2, ninsns--;
1473 #endif
1474
1475 if (ninsns)
1476 {
1477 *f1 = last1;
1478 *f2 = last2;
1479 }
1480
1481 return ninsns;
1482 }
1483
1484 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1485 the branch instruction. This means that if we commonize the control
1486 flow before end of the basic block, the semantic remains unchanged.
1487
1488 We may assume that there exists one edge with a common destination. */
1489
1490 static bool
1491 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1492 {
1493 int nehedges1 = 0, nehedges2 = 0;
1494 edge fallthru1 = 0, fallthru2 = 0;
1495 edge e1, e2;
1496 edge_iterator ei;
1497
1498 /* If we performed shrink-wrapping, edges to the EXIT_BLOCK_PTR can
1499 only be distinguished for JUMP_INSNs. The two paths may differ in
1500 whether they went through the prologue. Sibcalls are fine, we know
1501 that we either didn't need or inserted an epilogue before them. */
1502 if (crtl->shrink_wrapped
1503 && single_succ_p (bb1) && single_succ (bb1) == EXIT_BLOCK_PTR
1504 && !JUMP_P (BB_END (bb1))
1505 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1506 return false;
1507
1508 /* If BB1 has only one successor, we may be looking at either an
1509 unconditional jump, or a fake edge to exit. */
1510 if (single_succ_p (bb1)
1511 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1512 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1513 return (single_succ_p (bb2)
1514 && (single_succ_edge (bb2)->flags
1515 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1516 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1517
1518 /* Match conditional jumps - this may get tricky when fallthru and branch
1519 edges are crossed. */
1520 if (EDGE_COUNT (bb1->succs) == 2
1521 && any_condjump_p (BB_END (bb1))
1522 && onlyjump_p (BB_END (bb1)))
1523 {
1524 edge b1, f1, b2, f2;
1525 bool reverse, match;
1526 rtx set1, set2, cond1, cond2;
1527 enum rtx_code code1, code2;
1528
1529 if (EDGE_COUNT (bb2->succs) != 2
1530 || !any_condjump_p (BB_END (bb2))
1531 || !onlyjump_p (BB_END (bb2)))
1532 return false;
1533
1534 b1 = BRANCH_EDGE (bb1);
1535 b2 = BRANCH_EDGE (bb2);
1536 f1 = FALLTHRU_EDGE (bb1);
1537 f2 = FALLTHRU_EDGE (bb2);
1538
1539 /* Get around possible forwarders on fallthru edges. Other cases
1540 should be optimized out already. */
1541 if (FORWARDER_BLOCK_P (f1->dest))
1542 f1 = single_succ_edge (f1->dest);
1543
1544 if (FORWARDER_BLOCK_P (f2->dest))
1545 f2 = single_succ_edge (f2->dest);
1546
1547 /* To simplify use of this function, return false if there are
1548 unneeded forwarder blocks. These will get eliminated later
1549 during cleanup_cfg. */
1550 if (FORWARDER_BLOCK_P (f1->dest)
1551 || FORWARDER_BLOCK_P (f2->dest)
1552 || FORWARDER_BLOCK_P (b1->dest)
1553 || FORWARDER_BLOCK_P (b2->dest))
1554 return false;
1555
1556 if (f1->dest == f2->dest && b1->dest == b2->dest)
1557 reverse = false;
1558 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1559 reverse = true;
1560 else
1561 return false;
1562
1563 set1 = pc_set (BB_END (bb1));
1564 set2 = pc_set (BB_END (bb2));
1565 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1566 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1567 reverse = !reverse;
1568
1569 cond1 = XEXP (SET_SRC (set1), 0);
1570 cond2 = XEXP (SET_SRC (set2), 0);
1571 code1 = GET_CODE (cond1);
1572 if (reverse)
1573 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1574 else
1575 code2 = GET_CODE (cond2);
1576
1577 if (code2 == UNKNOWN)
1578 return false;
1579
1580 /* Verify codes and operands match. */
1581 match = ((code1 == code2
1582 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1583 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1584 || (code1 == swap_condition (code2)
1585 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1586 XEXP (cond2, 0))
1587 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1588 XEXP (cond2, 1))));
1589
1590 /* If we return true, we will join the blocks. Which means that
1591 we will only have one branch prediction bit to work with. Thus
1592 we require the existing branches to have probabilities that are
1593 roughly similar. */
1594 if (match
1595 && optimize_bb_for_speed_p (bb1)
1596 && optimize_bb_for_speed_p (bb2))
1597 {
1598 int prob2;
1599
1600 if (b1->dest == b2->dest)
1601 prob2 = b2->probability;
1602 else
1603 /* Do not use f2 probability as f2 may be forwarded. */
1604 prob2 = REG_BR_PROB_BASE - b2->probability;
1605
1606 /* Fail if the difference in probabilities is greater than 50%.
1607 This rules out two well-predicted branches with opposite
1608 outcomes. */
1609 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1610 {
1611 if (dump_file)
1612 fprintf (dump_file,
1613 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1614 bb1->index, bb2->index, b1->probability, prob2);
1615
1616 return false;
1617 }
1618 }
1619
1620 if (dump_file && match)
1621 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1622 bb1->index, bb2->index);
1623
1624 return match;
1625 }
1626
1627 /* Generic case - we are seeing a computed jump, table jump or trapping
1628 instruction. */
1629
1630 /* Check whether there are tablejumps in the end of BB1 and BB2.
1631 Return true if they are identical. */
1632 {
1633 rtx label1, label2;
1634 rtx table1, table2;
1635
1636 if (tablejump_p (BB_END (bb1), &label1, &table1)
1637 && tablejump_p (BB_END (bb2), &label2, &table2)
1638 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1639 {
1640 /* The labels should never be the same rtx. If they really are same
1641 the jump tables are same too. So disable crossjumping of blocks BB1
1642 and BB2 because when deleting the common insns in the end of BB1
1643 by delete_basic_block () the jump table would be deleted too. */
1644 /* If LABEL2 is referenced in BB1->END do not do anything
1645 because we would loose information when replacing
1646 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1647 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1648 {
1649 /* Set IDENTICAL to true when the tables are identical. */
1650 bool identical = false;
1651 rtx p1, p2;
1652
1653 p1 = PATTERN (table1);
1654 p2 = PATTERN (table2);
1655 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1656 {
1657 identical = true;
1658 }
1659 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1660 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1661 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1662 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1663 {
1664 int i;
1665
1666 identical = true;
1667 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1668 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1669 identical = false;
1670 }
1671
1672 if (identical)
1673 {
1674 replace_label_data rr;
1675 bool match;
1676
1677 /* Temporarily replace references to LABEL1 with LABEL2
1678 in BB1->END so that we could compare the instructions. */
1679 rr.r1 = label1;
1680 rr.r2 = label2;
1681 rr.update_label_nuses = false;
1682 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1683
1684 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1685 == dir_both);
1686 if (dump_file && match)
1687 fprintf (dump_file,
1688 "Tablejumps in bb %i and %i match.\n",
1689 bb1->index, bb2->index);
1690
1691 /* Set the original label in BB1->END because when deleting
1692 a block whose end is a tablejump, the tablejump referenced
1693 from the instruction is deleted too. */
1694 rr.r1 = label2;
1695 rr.r2 = label1;
1696 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1697
1698 return match;
1699 }
1700 }
1701 return false;
1702 }
1703 }
1704
1705 /* First ensure that the instructions match. There may be many outgoing
1706 edges so this test is generally cheaper. */
1707 if (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2)) != dir_both)
1708 return false;
1709
1710 /* Search the outgoing edges, ensure that the counts do match, find possible
1711 fallthru and exception handling edges since these needs more
1712 validation. */
1713 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1714 return false;
1715
1716 FOR_EACH_EDGE (e1, ei, bb1->succs)
1717 {
1718 e2 = EDGE_SUCC (bb2, ei.index);
1719
1720 if (e1->flags & EDGE_EH)
1721 nehedges1++;
1722
1723 if (e2->flags & EDGE_EH)
1724 nehedges2++;
1725
1726 if (e1->flags & EDGE_FALLTHRU)
1727 fallthru1 = e1;
1728 if (e2->flags & EDGE_FALLTHRU)
1729 fallthru2 = e2;
1730 }
1731
1732 /* If number of edges of various types does not match, fail. */
1733 if (nehedges1 != nehedges2
1734 || (fallthru1 != 0) != (fallthru2 != 0))
1735 return false;
1736
1737 /* fallthru edges must be forwarded to the same destination. */
1738 if (fallthru1)
1739 {
1740 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1741 ? single_succ (fallthru1->dest): fallthru1->dest);
1742 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1743 ? single_succ (fallthru2->dest): fallthru2->dest);
1744
1745 if (d1 != d2)
1746 return false;
1747 }
1748
1749 /* Ensure the same EH region. */
1750 {
1751 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1752 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1753
1754 if (!n1 && n2)
1755 return false;
1756
1757 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1758 return false;
1759 }
1760
1761 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1762 version of sequence abstraction. */
1763 FOR_EACH_EDGE (e1, ei, bb2->succs)
1764 {
1765 edge e2;
1766 edge_iterator ei;
1767 basic_block d1 = e1->dest;
1768
1769 if (FORWARDER_BLOCK_P (d1))
1770 d1 = EDGE_SUCC (d1, 0)->dest;
1771
1772 FOR_EACH_EDGE (e2, ei, bb1->succs)
1773 {
1774 basic_block d2 = e2->dest;
1775 if (FORWARDER_BLOCK_P (d2))
1776 d2 = EDGE_SUCC (d2, 0)->dest;
1777 if (d1 == d2)
1778 break;
1779 }
1780
1781 if (!e2)
1782 return false;
1783 }
1784
1785 return true;
1786 }
1787
1788 /* Returns true if BB basic block has a preserve label. */
1789
1790 static bool
1791 block_has_preserve_label (basic_block bb)
1792 {
1793 return (bb
1794 && block_label (bb)
1795 && LABEL_PRESERVE_P (block_label (bb)));
1796 }
1797
1798 /* E1 and E2 are edges with the same destination block. Search their
1799 predecessors for common code. If found, redirect control flow from
1800 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1801 or the other way around (dir_backward). DIR specifies the allowed
1802 replacement direction. */
1803
1804 static bool
1805 try_crossjump_to_edge (int mode, edge e1, edge e2,
1806 enum replace_direction dir)
1807 {
1808 int nmatch;
1809 basic_block src1 = e1->src, src2 = e2->src;
1810 basic_block redirect_to, redirect_from, to_remove;
1811 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1812 rtx newpos1, newpos2;
1813 edge s;
1814 edge_iterator ei;
1815
1816 newpos1 = newpos2 = NULL_RTX;
1817
1818 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1819 to try this optimization.
1820
1821 Basic block partitioning may result in some jumps that appear to
1822 be optimizable (or blocks that appear to be mergeable), but which really
1823 must be left untouched (they are required to make it safely across
1824 partition boundaries). See the comments at the top of
1825 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1826
1827 if (flag_reorder_blocks_and_partition && reload_completed)
1828 return false;
1829
1830 /* Search backward through forwarder blocks. We don't need to worry
1831 about multiple entry or chained forwarders, as they will be optimized
1832 away. We do this to look past the unconditional jump following a
1833 conditional jump that is required due to the current CFG shape. */
1834 if (single_pred_p (src1)
1835 && FORWARDER_BLOCK_P (src1))
1836 e1 = single_pred_edge (src1), src1 = e1->src;
1837
1838 if (single_pred_p (src2)
1839 && FORWARDER_BLOCK_P (src2))
1840 e2 = single_pred_edge (src2), src2 = e2->src;
1841
1842 /* Nothing to do if we reach ENTRY, or a common source block. */
1843 if (src1 == ENTRY_BLOCK_PTR || src2 == ENTRY_BLOCK_PTR)
1844 return false;
1845 if (src1 == src2)
1846 return false;
1847
1848 /* Seeing more than 1 forwarder blocks would confuse us later... */
1849 if (FORWARDER_BLOCK_P (e1->dest)
1850 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1851 return false;
1852
1853 if (FORWARDER_BLOCK_P (e2->dest)
1854 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1855 return false;
1856
1857 /* Likewise with dead code (possibly newly created by the other optimizations
1858 of cfg_cleanup). */
1859 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1860 return false;
1861
1862 /* Look for the common insn sequence, part the first ... */
1863 if (!outgoing_edges_match (mode, src1, src2))
1864 return false;
1865
1866 /* ... and part the second. */
1867 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1868
1869 osrc1 = src1;
1870 osrc2 = src2;
1871 if (newpos1 != NULL_RTX)
1872 src1 = BLOCK_FOR_INSN (newpos1);
1873 if (newpos2 != NULL_RTX)
1874 src2 = BLOCK_FOR_INSN (newpos2);
1875
1876 if (dir == dir_backward)
1877 {
1878 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1879 SWAP (basic_block, osrc1, osrc2);
1880 SWAP (basic_block, src1, src2);
1881 SWAP (edge, e1, e2);
1882 SWAP (rtx, newpos1, newpos2);
1883 #undef SWAP
1884 }
1885
1886 /* Don't proceed with the crossjump unless we found a sufficient number
1887 of matching instructions or the 'from' block was totally matched
1888 (such that its predecessors will hopefully be redirected and the
1889 block removed). */
1890 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
1891 && (newpos1 != BB_HEAD (src1)))
1892 return false;
1893
1894 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1895 if (block_has_preserve_label (e1->dest)
1896 && (e1->flags & EDGE_ABNORMAL))
1897 return false;
1898
1899 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1900 will be deleted.
1901 If we have tablejumps in the end of SRC1 and SRC2
1902 they have been already compared for equivalence in outgoing_edges_match ()
1903 so replace the references to TABLE1 by references to TABLE2. */
1904 {
1905 rtx label1, label2;
1906 rtx table1, table2;
1907
1908 if (tablejump_p (BB_END (osrc1), &label1, &table1)
1909 && tablejump_p (BB_END (osrc2), &label2, &table2)
1910 && label1 != label2)
1911 {
1912 replace_label_data rr;
1913 rtx insn;
1914
1915 /* Replace references to LABEL1 with LABEL2. */
1916 rr.r1 = label1;
1917 rr.r2 = label2;
1918 rr.update_label_nuses = true;
1919 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
1920 {
1921 /* Do not replace the label in SRC1->END because when deleting
1922 a block whose end is a tablejump, the tablejump referenced
1923 from the instruction is deleted too. */
1924 if (insn != BB_END (osrc1))
1925 for_each_rtx (&insn, replace_label, &rr);
1926 }
1927 }
1928 }
1929
1930 /* Avoid splitting if possible. We must always split when SRC2 has
1931 EH predecessor edges, or we may end up with basic blocks with both
1932 normal and EH predecessor edges. */
1933 if (newpos2 == BB_HEAD (src2)
1934 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
1935 redirect_to = src2;
1936 else
1937 {
1938 if (newpos2 == BB_HEAD (src2))
1939 {
1940 /* Skip possible basic block header. */
1941 if (LABEL_P (newpos2))
1942 newpos2 = NEXT_INSN (newpos2);
1943 while (DEBUG_INSN_P (newpos2))
1944 newpos2 = NEXT_INSN (newpos2);
1945 if (NOTE_P (newpos2))
1946 newpos2 = NEXT_INSN (newpos2);
1947 while (DEBUG_INSN_P (newpos2))
1948 newpos2 = NEXT_INSN (newpos2);
1949 }
1950
1951 if (dump_file)
1952 fprintf (dump_file, "Splitting bb %i before %i insns\n",
1953 src2->index, nmatch);
1954 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
1955 }
1956
1957 if (dump_file)
1958 fprintf (dump_file,
1959 "Cross jumping from bb %i to bb %i; %i common insns\n",
1960 src1->index, src2->index, nmatch);
1961
1962 /* We may have some registers visible through the block. */
1963 df_set_bb_dirty (redirect_to);
1964
1965 if (osrc2 == src2)
1966 redirect_edges_to = redirect_to;
1967 else
1968 redirect_edges_to = osrc2;
1969
1970 /* Recompute the frequencies and counts of outgoing edges. */
1971 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
1972 {
1973 edge s2;
1974 edge_iterator ei;
1975 basic_block d = s->dest;
1976
1977 if (FORWARDER_BLOCK_P (d))
1978 d = single_succ (d);
1979
1980 FOR_EACH_EDGE (s2, ei, src1->succs)
1981 {
1982 basic_block d2 = s2->dest;
1983 if (FORWARDER_BLOCK_P (d2))
1984 d2 = single_succ (d2);
1985 if (d == d2)
1986 break;
1987 }
1988
1989 s->count += s2->count;
1990
1991 /* Take care to update possible forwarder blocks. We verified
1992 that there is no more than one in the chain, so we can't run
1993 into infinite loop. */
1994 if (FORWARDER_BLOCK_P (s->dest))
1995 {
1996 single_succ_edge (s->dest)->count += s2->count;
1997 s->dest->count += s2->count;
1998 s->dest->frequency += EDGE_FREQUENCY (s);
1999 }
2000
2001 if (FORWARDER_BLOCK_P (s2->dest))
2002 {
2003 single_succ_edge (s2->dest)->count -= s2->count;
2004 if (single_succ_edge (s2->dest)->count < 0)
2005 single_succ_edge (s2->dest)->count = 0;
2006 s2->dest->count -= s2->count;
2007 s2->dest->frequency -= EDGE_FREQUENCY (s);
2008 if (s2->dest->frequency < 0)
2009 s2->dest->frequency = 0;
2010 if (s2->dest->count < 0)
2011 s2->dest->count = 0;
2012 }
2013
2014 if (!redirect_edges_to->frequency && !src1->frequency)
2015 s->probability = (s->probability + s2->probability) / 2;
2016 else
2017 s->probability
2018 = ((s->probability * redirect_edges_to->frequency +
2019 s2->probability * src1->frequency)
2020 / (redirect_edges_to->frequency + src1->frequency));
2021 }
2022
2023 /* Adjust count and frequency for the block. An earlier jump
2024 threading pass may have left the profile in an inconsistent
2025 state (see update_bb_profile_for_threading) so we must be
2026 prepared for overflows. */
2027 tmp = redirect_to;
2028 do
2029 {
2030 tmp->count += src1->count;
2031 tmp->frequency += src1->frequency;
2032 if (tmp->frequency > BB_FREQ_MAX)
2033 tmp->frequency = BB_FREQ_MAX;
2034 if (tmp == redirect_edges_to)
2035 break;
2036 tmp = find_fallthru_edge (tmp->succs)->dest;
2037 }
2038 while (true);
2039 update_br_prob_note (redirect_edges_to);
2040
2041 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2042
2043 /* Skip possible basic block header. */
2044 if (LABEL_P (newpos1))
2045 newpos1 = NEXT_INSN (newpos1);
2046
2047 while (DEBUG_INSN_P (newpos1))
2048 newpos1 = NEXT_INSN (newpos1);
2049
2050 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2051 newpos1 = NEXT_INSN (newpos1);
2052
2053 while (DEBUG_INSN_P (newpos1))
2054 newpos1 = NEXT_INSN (newpos1);
2055
2056 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2057 to_remove = single_succ (redirect_from);
2058
2059 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2060 delete_basic_block (to_remove);
2061
2062 update_forwarder_flag (redirect_from);
2063 if (redirect_to != src2)
2064 update_forwarder_flag (src2);
2065
2066 return true;
2067 }
2068
2069 /* Search the predecessors of BB for common insn sequences. When found,
2070 share code between them by redirecting control flow. Return true if
2071 any changes made. */
2072
2073 static bool
2074 try_crossjump_bb (int mode, basic_block bb)
2075 {
2076 edge e, e2, fallthru;
2077 bool changed;
2078 unsigned max, ix, ix2;
2079
2080 /* Nothing to do if there is not at least two incoming edges. */
2081 if (EDGE_COUNT (bb->preds) < 2)
2082 return false;
2083
2084 /* Don't crossjump if this block ends in a computed jump,
2085 unless we are optimizing for size. */
2086 if (optimize_bb_for_size_p (bb)
2087 && bb != EXIT_BLOCK_PTR
2088 && computed_jump_p (BB_END (bb)))
2089 return false;
2090
2091 /* If we are partitioning hot/cold basic blocks, we don't want to
2092 mess up unconditional or indirect jumps that cross between hot
2093 and cold sections.
2094
2095 Basic block partitioning may result in some jumps that appear to
2096 be optimizable (or blocks that appear to be mergeable), but which really
2097 must be left untouched (they are required to make it safely across
2098 partition boundaries). See the comments at the top of
2099 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2100
2101 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2102 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2103 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2104 return false;
2105
2106 /* It is always cheapest to redirect a block that ends in a branch to
2107 a block that falls through into BB, as that adds no branches to the
2108 program. We'll try that combination first. */
2109 fallthru = NULL;
2110 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2111
2112 if (EDGE_COUNT (bb->preds) > max)
2113 return false;
2114
2115 fallthru = find_fallthru_edge (bb->preds);
2116
2117 changed = false;
2118 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2119 {
2120 e = EDGE_PRED (bb, ix);
2121 ix++;
2122
2123 /* As noted above, first try with the fallthru predecessor (or, a
2124 fallthru predecessor if we are in cfglayout mode). */
2125 if (fallthru)
2126 {
2127 /* Don't combine the fallthru edge into anything else.
2128 If there is a match, we'll do it the other way around. */
2129 if (e == fallthru)
2130 continue;
2131 /* If nothing changed since the last attempt, there is nothing
2132 we can do. */
2133 if (!first_pass
2134 && !((e->src->flags & BB_MODIFIED)
2135 || (fallthru->src->flags & BB_MODIFIED)))
2136 continue;
2137
2138 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2139 {
2140 changed = true;
2141 ix = 0;
2142 continue;
2143 }
2144 }
2145
2146 /* Non-obvious work limiting check: Recognize that we're going
2147 to call try_crossjump_bb on every basic block. So if we have
2148 two blocks with lots of outgoing edges (a switch) and they
2149 share lots of common destinations, then we would do the
2150 cross-jump check once for each common destination.
2151
2152 Now, if the blocks actually are cross-jump candidates, then
2153 all of their destinations will be shared. Which means that
2154 we only need check them for cross-jump candidacy once. We
2155 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2156 choosing to do the check from the block for which the edge
2157 in question is the first successor of A. */
2158 if (EDGE_SUCC (e->src, 0) != e)
2159 continue;
2160
2161 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2162 {
2163 e2 = EDGE_PRED (bb, ix2);
2164
2165 if (e2 == e)
2166 continue;
2167
2168 /* We've already checked the fallthru edge above. */
2169 if (e2 == fallthru)
2170 continue;
2171
2172 /* The "first successor" check above only prevents multiple
2173 checks of crossjump(A,B). In order to prevent redundant
2174 checks of crossjump(B,A), require that A be the block
2175 with the lowest index. */
2176 if (e->src->index > e2->src->index)
2177 continue;
2178
2179 /* If nothing changed since the last attempt, there is nothing
2180 we can do. */
2181 if (!first_pass
2182 && !((e->src->flags & BB_MODIFIED)
2183 || (e2->src->flags & BB_MODIFIED)))
2184 continue;
2185
2186 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2187 direction. */
2188 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2189 {
2190 changed = true;
2191 ix = 0;
2192 break;
2193 }
2194 }
2195 }
2196
2197 if (changed)
2198 crossjumps_occured = true;
2199
2200 return changed;
2201 }
2202
2203 /* Search the successors of BB for common insn sequences. When found,
2204 share code between them by moving it across the basic block
2205 boundary. Return true if any changes made. */
2206
2207 static bool
2208 try_head_merge_bb (basic_block bb)
2209 {
2210 basic_block final_dest_bb = NULL;
2211 int max_match = INT_MAX;
2212 edge e0;
2213 rtx *headptr, *currptr, *nextptr;
2214 bool changed, moveall;
2215 unsigned ix;
2216 rtx e0_last_head, cond, move_before;
2217 unsigned nedges = EDGE_COUNT (bb->succs);
2218 rtx jump = BB_END (bb);
2219 regset live, live_union;
2220
2221 /* Nothing to do if there is not at least two outgoing edges. */
2222 if (nedges < 2)
2223 return false;
2224
2225 /* Don't crossjump if this block ends in a computed jump,
2226 unless we are optimizing for size. */
2227 if (optimize_bb_for_size_p (bb)
2228 && bb != EXIT_BLOCK_PTR
2229 && computed_jump_p (BB_END (bb)))
2230 return false;
2231
2232 cond = get_condition (jump, &move_before, true, false);
2233 if (cond == NULL_RTX)
2234 {
2235 #ifdef HAVE_cc0
2236 if (reg_mentioned_p (cc0_rtx, jump))
2237 move_before = prev_nonnote_nondebug_insn (jump);
2238 else
2239 #endif
2240 move_before = jump;
2241 }
2242
2243 for (ix = 0; ix < nedges; ix++)
2244 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR)
2245 return false;
2246
2247 for (ix = 0; ix < nedges; ix++)
2248 {
2249 edge e = EDGE_SUCC (bb, ix);
2250 basic_block other_bb = e->dest;
2251
2252 if (df_get_bb_dirty (other_bb))
2253 {
2254 block_was_dirty = true;
2255 return false;
2256 }
2257
2258 if (e->flags & EDGE_ABNORMAL)
2259 return false;
2260
2261 /* Normally, all destination blocks must only be reachable from this
2262 block, i.e. they must have one incoming edge.
2263
2264 There is one special case we can handle, that of multiple consecutive
2265 jumps where the first jumps to one of the targets of the second jump.
2266 This happens frequently in switch statements for default labels.
2267 The structure is as follows:
2268 FINAL_DEST_BB
2269 ....
2270 if (cond) jump A;
2271 fall through
2272 BB
2273 jump with targets A, B, C, D...
2274 A
2275 has two incoming edges, from FINAL_DEST_BB and BB
2276
2277 In this case, we can try to move the insns through BB and into
2278 FINAL_DEST_BB. */
2279 if (EDGE_COUNT (other_bb->preds) != 1)
2280 {
2281 edge incoming_edge, incoming_bb_other_edge;
2282 edge_iterator ei;
2283
2284 if (final_dest_bb != NULL
2285 || EDGE_COUNT (other_bb->preds) != 2)
2286 return false;
2287
2288 /* We must be able to move the insns across the whole block. */
2289 move_before = BB_HEAD (bb);
2290 while (!NONDEBUG_INSN_P (move_before))
2291 move_before = NEXT_INSN (move_before);
2292
2293 if (EDGE_COUNT (bb->preds) != 1)
2294 return false;
2295 incoming_edge = EDGE_PRED (bb, 0);
2296 final_dest_bb = incoming_edge->src;
2297 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2298 return false;
2299 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2300 if (incoming_bb_other_edge != incoming_edge)
2301 break;
2302 if (incoming_bb_other_edge->dest != other_bb)
2303 return false;
2304 }
2305 }
2306
2307 e0 = EDGE_SUCC (bb, 0);
2308 e0_last_head = NULL_RTX;
2309 changed = false;
2310
2311 for (ix = 1; ix < nedges; ix++)
2312 {
2313 edge e = EDGE_SUCC (bb, ix);
2314 rtx e0_last, e_last;
2315 int nmatch;
2316
2317 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2318 &e0_last, &e_last, 0);
2319 if (nmatch == 0)
2320 return false;
2321
2322 if (nmatch < max_match)
2323 {
2324 max_match = nmatch;
2325 e0_last_head = e0_last;
2326 }
2327 }
2328
2329 /* If we matched an entire block, we probably have to avoid moving the
2330 last insn. */
2331 if (max_match > 0
2332 && e0_last_head == BB_END (e0->dest)
2333 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2334 || control_flow_insn_p (e0_last_head)))
2335 {
2336 max_match--;
2337 if (max_match == 0)
2338 return false;
2339 do
2340 e0_last_head = prev_real_insn (e0_last_head);
2341 while (DEBUG_INSN_P (e0_last_head));
2342 }
2343
2344 if (max_match == 0)
2345 return false;
2346
2347 /* We must find a union of the live registers at each of the end points. */
2348 live = BITMAP_ALLOC (NULL);
2349 live_union = BITMAP_ALLOC (NULL);
2350
2351 currptr = XNEWVEC (rtx, nedges);
2352 headptr = XNEWVEC (rtx, nedges);
2353 nextptr = XNEWVEC (rtx, nedges);
2354
2355 for (ix = 0; ix < nedges; ix++)
2356 {
2357 int j;
2358 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2359 rtx head = BB_HEAD (merge_bb);
2360
2361 while (!NONDEBUG_INSN_P (head))
2362 head = NEXT_INSN (head);
2363 headptr[ix] = head;
2364 currptr[ix] = head;
2365
2366 /* Compute the end point and live information */
2367 for (j = 1; j < max_match; j++)
2368 do
2369 head = NEXT_INSN (head);
2370 while (!NONDEBUG_INSN_P (head));
2371 simulate_backwards_to_point (merge_bb, live, head);
2372 IOR_REG_SET (live_union, live);
2373 }
2374
2375 /* If we're moving across two blocks, verify the validity of the
2376 first move, then adjust the target and let the loop below deal
2377 with the final move. */
2378 if (final_dest_bb != NULL)
2379 {
2380 rtx move_upto;
2381
2382 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2383 jump, e0->dest, live_union,
2384 NULL, &move_upto);
2385 if (!moveall)
2386 {
2387 if (move_upto == NULL_RTX)
2388 goto out;
2389
2390 while (e0_last_head != move_upto)
2391 {
2392 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2393 live_union);
2394 e0_last_head = PREV_INSN (e0_last_head);
2395 }
2396 }
2397 if (e0_last_head == NULL_RTX)
2398 goto out;
2399
2400 jump = BB_END (final_dest_bb);
2401 cond = get_condition (jump, &move_before, true, false);
2402 if (cond == NULL_RTX)
2403 {
2404 #ifdef HAVE_cc0
2405 if (reg_mentioned_p (cc0_rtx, jump))
2406 move_before = prev_nonnote_nondebug_insn (jump);
2407 else
2408 #endif
2409 move_before = jump;
2410 }
2411 }
2412
2413 do
2414 {
2415 rtx move_upto;
2416 moveall = can_move_insns_across (currptr[0], e0_last_head,
2417 move_before, jump, e0->dest, live_union,
2418 NULL, &move_upto);
2419 if (!moveall && move_upto == NULL_RTX)
2420 {
2421 if (jump == move_before)
2422 break;
2423
2424 /* Try again, using a different insertion point. */
2425 move_before = jump;
2426
2427 #ifdef HAVE_cc0
2428 /* Don't try moving before a cc0 user, as that may invalidate
2429 the cc0. */
2430 if (reg_mentioned_p (cc0_rtx, jump))
2431 break;
2432 #endif
2433
2434 continue;
2435 }
2436
2437 if (final_dest_bb && !moveall)
2438 /* We haven't checked whether a partial move would be OK for the first
2439 move, so we have to fail this case. */
2440 break;
2441
2442 changed = true;
2443 for (;;)
2444 {
2445 if (currptr[0] == move_upto)
2446 break;
2447 for (ix = 0; ix < nedges; ix++)
2448 {
2449 rtx curr = currptr[ix];
2450 do
2451 curr = NEXT_INSN (curr);
2452 while (!NONDEBUG_INSN_P (curr));
2453 currptr[ix] = curr;
2454 }
2455 }
2456
2457 /* If we can't currently move all of the identical insns, remember
2458 each insn after the range that we'll merge. */
2459 if (!moveall)
2460 for (ix = 0; ix < nedges; ix++)
2461 {
2462 rtx curr = currptr[ix];
2463 do
2464 curr = NEXT_INSN (curr);
2465 while (!NONDEBUG_INSN_P (curr));
2466 nextptr[ix] = curr;
2467 }
2468
2469 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2470 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2471 if (final_dest_bb != NULL)
2472 df_set_bb_dirty (final_dest_bb);
2473 df_set_bb_dirty (bb);
2474 for (ix = 1; ix < nedges; ix++)
2475 {
2476 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2477 delete_insn_chain (headptr[ix], currptr[ix], false);
2478 }
2479 if (!moveall)
2480 {
2481 if (jump == move_before)
2482 break;
2483
2484 /* For the unmerged insns, try a different insertion point. */
2485 move_before = jump;
2486
2487 #ifdef HAVE_cc0
2488 /* Don't try moving before a cc0 user, as that may invalidate
2489 the cc0. */
2490 if (reg_mentioned_p (cc0_rtx, jump))
2491 break;
2492 #endif
2493
2494 for (ix = 0; ix < nedges; ix++)
2495 currptr[ix] = headptr[ix] = nextptr[ix];
2496 }
2497 }
2498 while (!moveall);
2499
2500 out:
2501 free (currptr);
2502 free (headptr);
2503 free (nextptr);
2504
2505 crossjumps_occured |= changed;
2506
2507 return changed;
2508 }
2509
2510 /* Return true if BB contains just bb note, or bb note followed
2511 by only DEBUG_INSNs. */
2512
2513 static bool
2514 trivially_empty_bb_p (basic_block bb)
2515 {
2516 rtx insn = BB_END (bb);
2517
2518 while (1)
2519 {
2520 if (insn == BB_HEAD (bb))
2521 return true;
2522 if (!DEBUG_INSN_P (insn))
2523 return false;
2524 insn = PREV_INSN (insn);
2525 }
2526 }
2527
2528 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2529 instructions etc. Return nonzero if changes were made. */
2530
2531 static bool
2532 try_optimize_cfg (int mode)
2533 {
2534 bool changed_overall = false;
2535 bool changed;
2536 int iterations = 0;
2537 basic_block bb, b, next;
2538
2539 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2540 clear_bb_flags ();
2541
2542 crossjumps_occured = false;
2543
2544 FOR_EACH_BB (bb)
2545 update_forwarder_flag (bb);
2546
2547 if (! targetm.cannot_modify_jumps_p ())
2548 {
2549 first_pass = true;
2550 /* Attempt to merge blocks as made possible by edge removal. If
2551 a block has only one successor, and the successor has only
2552 one predecessor, they may be combined. */
2553 do
2554 {
2555 block_was_dirty = false;
2556 changed = false;
2557 iterations++;
2558
2559 if (dump_file)
2560 fprintf (dump_file,
2561 "\n\ntry_optimize_cfg iteration %i\n\n",
2562 iterations);
2563
2564 for (b = ENTRY_BLOCK_PTR->next_bb; b != EXIT_BLOCK_PTR;)
2565 {
2566 basic_block c;
2567 edge s;
2568 bool changed_here = false;
2569
2570 /* Delete trivially dead basic blocks. This is either
2571 blocks with no predecessors, or empty blocks with no
2572 successors. However if the empty block with no
2573 successors is the successor of the ENTRY_BLOCK, it is
2574 kept. This ensures that the ENTRY_BLOCK will have a
2575 successor which is a precondition for many RTL
2576 passes. Empty blocks may result from expanding
2577 __builtin_unreachable (). */
2578 if (EDGE_COUNT (b->preds) == 0
2579 || (EDGE_COUNT (b->succs) == 0
2580 && trivially_empty_bb_p (b)
2581 && single_succ_edge (ENTRY_BLOCK_PTR)->dest != b))
2582 {
2583 c = b->prev_bb;
2584 if (EDGE_COUNT (b->preds) > 0)
2585 {
2586 edge e;
2587 edge_iterator ei;
2588
2589 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2590 {
2591 if (BB_FOOTER (b)
2592 && BARRIER_P (BB_FOOTER (b)))
2593 FOR_EACH_EDGE (e, ei, b->preds)
2594 if ((e->flags & EDGE_FALLTHRU)
2595 && BB_FOOTER (e->src) == NULL)
2596 {
2597 if (BB_FOOTER (b))
2598 {
2599 BB_FOOTER (e->src) = BB_FOOTER (b);
2600 BB_FOOTER (b) = NULL;
2601 }
2602 else
2603 {
2604 start_sequence ();
2605 BB_FOOTER (e->src) = emit_barrier ();
2606 end_sequence ();
2607 }
2608 }
2609 }
2610 else
2611 {
2612 rtx last = get_last_bb_insn (b);
2613 if (last && BARRIER_P (last))
2614 FOR_EACH_EDGE (e, ei, b->preds)
2615 if ((e->flags & EDGE_FALLTHRU))
2616 emit_barrier_after (BB_END (e->src));
2617 }
2618 }
2619 delete_basic_block (b);
2620 changed = true;
2621 /* Avoid trying to remove ENTRY_BLOCK_PTR. */
2622 b = (c == ENTRY_BLOCK_PTR ? c->next_bb : c);
2623 continue;
2624 }
2625
2626 /* Remove code labels no longer used. */
2627 if (single_pred_p (b)
2628 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2629 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2630 && LABEL_P (BB_HEAD (b))
2631 /* If the previous block ends with a branch to this
2632 block, we can't delete the label. Normally this
2633 is a condjump that is yet to be simplified, but
2634 if CASE_DROPS_THRU, this can be a tablejump with
2635 some element going to the same place as the
2636 default (fallthru). */
2637 && (single_pred (b) == ENTRY_BLOCK_PTR
2638 || !JUMP_P (BB_END (single_pred (b)))
2639 || ! label_is_jump_target_p (BB_HEAD (b),
2640 BB_END (single_pred (b)))))
2641 {
2642 delete_insn (BB_HEAD (b));
2643 if (dump_file)
2644 fprintf (dump_file, "Deleted label in block %i.\n",
2645 b->index);
2646 }
2647
2648 /* If we fall through an empty block, we can remove it. */
2649 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2650 && single_pred_p (b)
2651 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2652 && !LABEL_P (BB_HEAD (b))
2653 && FORWARDER_BLOCK_P (b)
2654 /* Note that forwarder_block_p true ensures that
2655 there is a successor for this block. */
2656 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2657 && n_basic_blocks > NUM_FIXED_BLOCKS + 1)
2658 {
2659 if (dump_file)
2660 fprintf (dump_file,
2661 "Deleting fallthru block %i.\n",
2662 b->index);
2663
2664 c = b->prev_bb == ENTRY_BLOCK_PTR ? b->next_bb : b->prev_bb;
2665 redirect_edge_succ_nodup (single_pred_edge (b),
2666 single_succ (b));
2667 delete_basic_block (b);
2668 changed = true;
2669 b = c;
2670 continue;
2671 }
2672
2673 /* Merge B with its single successor, if any. */
2674 if (single_succ_p (b)
2675 && (s = single_succ_edge (b))
2676 && !(s->flags & EDGE_COMPLEX)
2677 && (c = s->dest) != EXIT_BLOCK_PTR
2678 && single_pred_p (c)
2679 && b != c)
2680 {
2681 /* When not in cfg_layout mode use code aware of reordering
2682 INSN. This code possibly creates new basic blocks so it
2683 does not fit merge_blocks interface and is kept here in
2684 hope that it will become useless once more of compiler
2685 is transformed to use cfg_layout mode. */
2686
2687 if ((mode & CLEANUP_CFGLAYOUT)
2688 && can_merge_blocks_p (b, c))
2689 {
2690 merge_blocks (b, c);
2691 update_forwarder_flag (b);
2692 changed_here = true;
2693 }
2694 else if (!(mode & CLEANUP_CFGLAYOUT)
2695 /* If the jump insn has side effects,
2696 we can't kill the edge. */
2697 && (!JUMP_P (BB_END (b))
2698 || (reload_completed
2699 ? simplejump_p (BB_END (b))
2700 : (onlyjump_p (BB_END (b))
2701 && !tablejump_p (BB_END (b),
2702 NULL, NULL))))
2703 && (next = merge_blocks_move (s, b, c, mode)))
2704 {
2705 b = next;
2706 changed_here = true;
2707 }
2708 }
2709
2710 /* Simplify branch over branch. */
2711 if ((mode & CLEANUP_EXPENSIVE)
2712 && !(mode & CLEANUP_CFGLAYOUT)
2713 && try_simplify_condjump (b))
2714 changed_here = true;
2715
2716 /* If B has a single outgoing edge, but uses a
2717 non-trivial jump instruction without side-effects, we
2718 can either delete the jump entirely, or replace it
2719 with a simple unconditional jump. */
2720 if (single_succ_p (b)
2721 && single_succ (b) != EXIT_BLOCK_PTR
2722 && onlyjump_p (BB_END (b))
2723 && !find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX)
2724 && try_redirect_by_replacing_jump (single_succ_edge (b),
2725 single_succ (b),
2726 (mode & CLEANUP_CFGLAYOUT) != 0))
2727 {
2728 update_forwarder_flag (b);
2729 changed_here = true;
2730 }
2731
2732 /* Simplify branch to branch. */
2733 if (try_forward_edges (mode, b))
2734 {
2735 update_forwarder_flag (b);
2736 changed_here = true;
2737 }
2738
2739 /* Look for shared code between blocks. */
2740 if ((mode & CLEANUP_CROSSJUMP)
2741 && try_crossjump_bb (mode, b))
2742 changed_here = true;
2743
2744 if ((mode & CLEANUP_CROSSJUMP)
2745 /* This can lengthen register lifetimes. Do it only after
2746 reload. */
2747 && reload_completed
2748 && try_head_merge_bb (b))
2749 changed_here = true;
2750
2751 /* Don't get confused by the index shift caused by
2752 deleting blocks. */
2753 if (!changed_here)
2754 b = b->next_bb;
2755 else
2756 changed = true;
2757 }
2758
2759 if ((mode & CLEANUP_CROSSJUMP)
2760 && try_crossjump_bb (mode, EXIT_BLOCK_PTR))
2761 changed = true;
2762
2763 if (block_was_dirty)
2764 {
2765 /* This should only be set by head-merging. */
2766 gcc_assert (mode & CLEANUP_CROSSJUMP);
2767 df_analyze ();
2768 }
2769
2770 #ifdef ENABLE_CHECKING
2771 if (changed)
2772 verify_flow_info ();
2773 #endif
2774
2775 changed_overall |= changed;
2776 first_pass = false;
2777 }
2778 while (changed);
2779 }
2780
2781 FOR_ALL_BB (b)
2782 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2783
2784 return changed_overall;
2785 }
2786 \f
2787 /* Delete all unreachable basic blocks. */
2788
2789 bool
2790 delete_unreachable_blocks (void)
2791 {
2792 bool changed = false;
2793 basic_block b, prev_bb;
2794
2795 find_unreachable_blocks ();
2796
2797 /* When we're in GIMPLE mode and there may be debug insns, we should
2798 delete blocks in reverse dominator order, so as to get a chance
2799 to substitute all released DEFs into debug stmts. If we don't
2800 have dominators information, walking blocks backward gets us a
2801 better chance of retaining most debug information than
2802 otherwise. */
2803 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2804 && dom_info_available_p (CDI_DOMINATORS))
2805 {
2806 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2807 {
2808 prev_bb = b->prev_bb;
2809
2810 if (!(b->flags & BB_REACHABLE))
2811 {
2812 /* Speed up the removal of blocks that don't dominate
2813 others. Walking backwards, this should be the common
2814 case. */
2815 if (!first_dom_son (CDI_DOMINATORS, b))
2816 delete_basic_block (b);
2817 else
2818 {
2819 vec<basic_block> h
2820 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2821
2822 while (h.length ())
2823 {
2824 b = h.pop ();
2825
2826 prev_bb = b->prev_bb;
2827
2828 gcc_assert (!(b->flags & BB_REACHABLE));
2829
2830 delete_basic_block (b);
2831 }
2832
2833 h.release ();
2834 }
2835
2836 changed = true;
2837 }
2838 }
2839 }
2840 else
2841 {
2842 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2843 {
2844 prev_bb = b->prev_bb;
2845
2846 if (!(b->flags & BB_REACHABLE))
2847 {
2848 delete_basic_block (b);
2849 changed = true;
2850 }
2851 }
2852 }
2853
2854 if (changed)
2855 tidy_fallthru_edges ();
2856 return changed;
2857 }
2858
2859 /* Delete any jump tables never referenced. We can't delete them at the
2860 time of removing tablejump insn as they are referenced by the preceding
2861 insns computing the destination, so we delay deleting and garbagecollect
2862 them once life information is computed. */
2863 void
2864 delete_dead_jumptables (void)
2865 {
2866 basic_block bb;
2867
2868 /* A dead jump table does not belong to any basic block. Scan insns
2869 between two adjacent basic blocks. */
2870 FOR_EACH_BB (bb)
2871 {
2872 rtx insn, next;
2873
2874 for (insn = NEXT_INSN (BB_END (bb));
2875 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
2876 insn = next)
2877 {
2878 next = NEXT_INSN (insn);
2879 if (LABEL_P (insn)
2880 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
2881 && JUMP_TABLE_DATA_P (next))
2882 {
2883 rtx label = insn, jump = next;
2884
2885 if (dump_file)
2886 fprintf (dump_file, "Dead jumptable %i removed\n",
2887 INSN_UID (insn));
2888
2889 next = NEXT_INSN (next);
2890 delete_insn (jump);
2891 delete_insn (label);
2892 }
2893 }
2894 }
2895 }
2896
2897 \f
2898 /* Tidy the CFG by deleting unreachable code and whatnot. */
2899
2900 bool
2901 cleanup_cfg (int mode)
2902 {
2903 bool changed = false;
2904
2905 /* Set the cfglayout mode flag here. We could update all the callers
2906 but that is just inconvenient, especially given that we eventually
2907 want to have cfglayout mode as the default. */
2908 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2909 mode |= CLEANUP_CFGLAYOUT;
2910
2911 timevar_push (TV_CLEANUP_CFG);
2912 if (delete_unreachable_blocks ())
2913 {
2914 changed = true;
2915 /* We've possibly created trivially dead code. Cleanup it right
2916 now to introduce more opportunities for try_optimize_cfg. */
2917 if (!(mode & (CLEANUP_NO_INSN_DEL))
2918 && !reload_completed)
2919 delete_trivially_dead_insns (get_insns (), max_reg_num ());
2920 }
2921
2922 compact_blocks ();
2923
2924 /* To tail-merge blocks ending in the same noreturn function (e.g.
2925 a call to abort) we have to insert fake edges to exit. Do this
2926 here once. The fake edges do not interfere with any other CFG
2927 cleanups. */
2928 if (mode & CLEANUP_CROSSJUMP)
2929 add_noreturn_fake_exit_edges ();
2930
2931 if (!dbg_cnt (cfg_cleanup))
2932 return changed;
2933
2934 while (try_optimize_cfg (mode))
2935 {
2936 delete_unreachable_blocks (), changed = true;
2937 if (!(mode & CLEANUP_NO_INSN_DEL))
2938 {
2939 /* Try to remove some trivially dead insns when doing an expensive
2940 cleanup. But delete_trivially_dead_insns doesn't work after
2941 reload (it only handles pseudos) and run_fast_dce is too costly
2942 to run in every iteration.
2943
2944 For effective cross jumping, we really want to run a fast DCE to
2945 clean up any dead conditions, or they get in the way of performing
2946 useful tail merges.
2947
2948 Other transformations in cleanup_cfg are not so sensitive to dead
2949 code, so delete_trivially_dead_insns or even doing nothing at all
2950 is good enough. */
2951 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
2952 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
2953 break;
2954 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
2955 run_fast_dce ();
2956 }
2957 else
2958 break;
2959 }
2960
2961 if (mode & CLEANUP_CROSSJUMP)
2962 remove_fake_exit_edges ();
2963
2964 /* Don't call delete_dead_jumptables in cfglayout mode, because
2965 that function assumes that jump tables are in the insns stream.
2966 But we also don't _have_ to delete dead jumptables in cfglayout
2967 mode because we shouldn't even be looking at things that are
2968 not in a basic block. Dead jumptables are cleaned up when
2969 going out of cfglayout mode. */
2970 if (!(mode & CLEANUP_CFGLAYOUT))
2971 delete_dead_jumptables ();
2972
2973 /* ??? We probably do this way too often. */
2974 if (current_loops
2975 && (changed
2976 || (mode & CLEANUP_CFG_CHANGED)))
2977 {
2978 bitmap changed_bbs;
2979 timevar_push (TV_REPAIR_LOOPS);
2980 /* The above doesn't preserve dominance info if available. */
2981 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
2982 calculate_dominance_info (CDI_DOMINATORS);
2983 changed_bbs = BITMAP_ALLOC (NULL);
2984 fix_loop_structure (changed_bbs);
2985 BITMAP_FREE (changed_bbs);
2986 free_dominance_info (CDI_DOMINATORS);
2987 timevar_pop (TV_REPAIR_LOOPS);
2988 }
2989
2990 timevar_pop (TV_CLEANUP_CFG);
2991
2992 return changed;
2993 }
2994 \f
2995 static unsigned int
2996 execute_jump (void)
2997 {
2998 delete_trivially_dead_insns (get_insns (), max_reg_num ());
2999 if (dump_file)
3000 dump_flow_info (dump_file, dump_flags);
3001 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3002 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3003 return 0;
3004 }
3005
3006 struct rtl_opt_pass pass_jump =
3007 {
3008 {
3009 RTL_PASS,
3010 "jump", /* name */
3011 OPTGROUP_NONE, /* optinfo_flags */
3012 NULL, /* gate */
3013 execute_jump, /* execute */
3014 NULL, /* sub */
3015 NULL, /* next */
3016 0, /* static_pass_number */
3017 TV_JUMP, /* tv_id */
3018 0, /* properties_required */
3019 0, /* properties_provided */
3020 0, /* properties_destroyed */
3021 TODO_ggc_collect, /* todo_flags_start */
3022 TODO_verify_rtl_sharing, /* todo_flags_finish */
3023 }
3024 };
3025 \f
3026 static unsigned int
3027 execute_jump2 (void)
3028 {
3029 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3030 return 0;
3031 }
3032
3033 struct rtl_opt_pass pass_jump2 =
3034 {
3035 {
3036 RTL_PASS,
3037 "jump2", /* name */
3038 OPTGROUP_NONE, /* optinfo_flags */
3039 NULL, /* gate */
3040 execute_jump2, /* execute */
3041 NULL, /* sub */
3042 NULL, /* next */
3043 0, /* static_pass_number */
3044 TV_JUMP, /* tv_id */
3045 0, /* properties_required */
3046 0, /* properties_provided */
3047 0, /* properties_destroyed */
3048 TODO_ggc_collect, /* todo_flags_start */
3049 TODO_verify_rtl_sharing, /* todo_flags_finish */
3050 }
3051 };