1 /* Loop manipulation code for GNU compiler.
2 Copyright (C) 2002, 2003, 2004 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 2, or (at your option) any later
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
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING. If not, write to the Free
18 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
23 #include "coretypes.h"
26 #include "hard-reg-set.h"
27 #include "basic-block.h"
29 #include "cfglayout.h"
32 static void duplicate_subloops (struct loops *, struct loop *, struct loop *);
33 static void copy_loops_to (struct loops *, struct loop **, int,
35 static void loop_redirect_edge (edge, basic_block);
36 static bool loop_delete_branch_edge (edge, int);
37 static void remove_bbs (basic_block *, int);
38 static bool rpe_enum_p (basic_block, void *);
39 static int find_path (edge, basic_block **);
40 static bool alp_enum_p (basic_block, void *);
41 static void add_loop (struct loops *, struct loop *);
42 static void fix_loop_placements (struct loops *, struct loop *);
43 static bool fix_bb_placement (struct loops *, basic_block);
44 static void fix_bb_placements (struct loops *, basic_block);
45 static void place_new_loop (struct loops *, struct loop *);
46 static void scale_loop_frequencies (struct loop *, int, int);
47 static void scale_bbs_frequencies (basic_block *, int, int, int);
48 static basic_block create_preheader (struct loop *, int);
49 static void fix_irreducible_loops (basic_block);
51 #define RDIV(X,Y) (((X) + (Y) / 2) / (Y))
53 /* Splits basic block BB after INSN, returns created edge. Updates loops
56 split_loop_bb (basic_block bb, rtx insn)
60 /* Split the block. */
61 e = split_block (bb, insn);
63 /* Add dest to loop. */
64 add_bb_to_loop (e->dest, e->src->loop_father);
69 /* Checks whether basic block BB is dominated by DATA. */
71 rpe_enum_p (basic_block bb, void *data)
73 return dominated_by_p (CDI_DOMINATORS, bb, data);
76 /* Remove basic blocks BBS from loop structure and dominance info,
77 and delete them afterwards. */
79 remove_bbs (basic_block *bbs, int nbbs)
83 for (i = 0; i < nbbs; i++)
85 remove_bb_from_loops (bbs[i]);
86 delete_basic_block (bbs[i]);
90 /* Find path -- i.e. the basic blocks dominated by edge E and put them
91 into array BBS, that will be allocated large enough to contain them.
92 E->dest must have exactly one predecessor for this to work (it is
93 easy to achieve and we do not put it here because we do not want to
94 alter anything by this function). The number of basic blocks in the
97 find_path (edge e, basic_block **bbs)
99 gcc_assert (!e->dest->pred->pred_next);
101 /* Find bbs in the path. */
102 *bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
103 return dfs_enumerate_from (e->dest, 0, rpe_enum_p, *bbs,
104 n_basic_blocks, e->dest);
107 /* Fix placement of basic block BB inside loop hierarchy stored in LOOPS --
108 Let L be a loop to that BB belongs. Then every successor of BB must either
109 1) belong to some superloop of loop L, or
110 2) be a header of loop K such that K->outer is superloop of L
111 Returns true if we had to move BB into other loop to enforce this condition,
112 false if the placement of BB was already correct (provided that placements
113 of its successors are correct). */
115 fix_bb_placement (struct loops *loops, basic_block bb)
118 struct loop *loop = loops->tree_root, *act;
120 for (e = bb->succ; e; e = e->succ_next)
122 if (e->dest == EXIT_BLOCK_PTR)
125 act = e->dest->loop_father;
126 if (act->header == e->dest)
129 if (flow_loop_nested_p (loop, act))
133 if (loop == bb->loop_father)
136 remove_bb_from_loops (bb);
137 add_bb_to_loop (bb, loop);
142 /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e.
143 enforce condition condition stated in description of fix_bb_placement. We
144 start from basic block FROM that had some of its successors removed, so that
145 his placement no longer has to be correct, and iteratively fix placement of
146 its predecessors that may change if placement of FROM changed. Also fix
147 placement of subloops of FROM->loop_father, that might also be altered due
148 to this change; the condition for them is similar, except that instead of
149 successors we consider edges coming out of the loops. */
151 fix_bb_placements (struct loops *loops, basic_block from)
154 basic_block *queue, *qtop, *qbeg, *qend;
155 struct loop *base_loop;
158 /* We pass through blocks back-reachable from FROM, testing whether some
159 of their successors moved to outer loop. It may be necessary to
160 iterate several times, but it is finite, as we stop unless we move
161 the basic block up the loop structure. The whole story is a bit
162 more complicated due to presence of subloops, those are moved using
163 fix_loop_placement. */
165 base_loop = from->loop_father;
166 if (base_loop == loops->tree_root)
169 in_queue = sbitmap_alloc (last_basic_block);
170 sbitmap_zero (in_queue);
171 SET_BIT (in_queue, from->index);
172 /* Prevent us from going out of the base_loop. */
173 SET_BIT (in_queue, base_loop->header->index);
175 queue = xmalloc ((base_loop->num_nodes + 1) * sizeof (basic_block));
176 qtop = queue + base_loop->num_nodes + 1;
187 RESET_BIT (in_queue, from->index);
189 if (from->loop_father->header == from)
191 /* Subloop header, maybe move the loop upward. */
192 if (!fix_loop_placement (from->loop_father))
197 /* Ordinary basic block. */
198 if (!fix_bb_placement (loops, from))
202 /* Something has changed, insert predecessors into queue. */
203 for (e = from->pred; e; e = e->pred_next)
205 basic_block pred = e->src;
208 if (TEST_BIT (in_queue, pred->index))
211 /* If it is subloop, then it either was not moved, or
212 the path up the loop tree from base_loop do not contain
214 nca = find_common_loop (pred->loop_father, base_loop);
215 if (pred->loop_father != base_loop
217 || nca != pred->loop_father))
218 pred = pred->loop_father->header;
219 else if (!flow_loop_nested_p (from->loop_father, pred->loop_father))
221 /* No point in processing it. */
225 if (TEST_BIT (in_queue, pred->index))
228 /* Schedule the basic block. */
233 SET_BIT (in_queue, pred->index);
240 /* Basic block from has lost one or more of its predecessors, so it might
241 mo longer be part irreducible loop. Fix it and proceed recursively
242 for its successors if needed. */
244 fix_irreducible_loops (basic_block from)
253 if (!(from->flags & BB_IRREDUCIBLE_LOOP))
256 on_stack = sbitmap_alloc (last_basic_block);
257 sbitmap_zero (on_stack);
258 SET_BIT (on_stack, from->index);
259 stack = xmalloc (from->loop_father->num_nodes * sizeof (basic_block));
265 bb = stack[--stack_top];
266 RESET_BIT (on_stack, bb->index);
268 for (e = bb->pred; e; e = e->pred_next)
269 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
274 bb->flags &= ~BB_IRREDUCIBLE_LOOP;
275 if (bb->loop_father->header == bb)
276 edges = get_loop_exit_edges (bb->loop_father, &n_edges);
280 for (e = bb->succ; e; e = e->succ_next)
282 edges = xmalloc (n_edges * sizeof (edge));
284 for (e = bb->succ; e; e = e->succ_next)
285 edges[n_edges++] = e;
288 for (i = 0; i < n_edges; i++)
292 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
294 if (!flow_bb_inside_loop_p (from->loop_father, e->dest))
297 e->flags &= ~EDGE_IRREDUCIBLE_LOOP;
298 if (TEST_BIT (on_stack, e->dest->index))
301 SET_BIT (on_stack, e->dest->index);
302 stack[stack_top++] = e->dest;
312 /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E
313 and update loop structure stored in LOOPS and dominators. Return true if
314 we were able to remove the path, false otherwise (and nothing is affected
317 remove_path (struct loops *loops, edge e)
320 basic_block *rem_bbs, *bord_bbs, *dom_bbs, from, bb;
321 int i, nrem, n_bord_bbs, n_dom_bbs;
325 if (!loop_delete_branch_edge (e, 0))
328 /* We need to check whether basic blocks are dominated by the edge
329 e, but we only have basic block dominators. This is easy to
330 fix -- when e->dest has exactly one predecessor, this corresponds
331 to blocks dominated by e->dest, if not, split the edge. */
332 if (e->dest->pred->pred_next)
333 e = loop_split_edge_with (e, NULL_RTX)->pred;
335 /* It may happen that by removing path we remove one or more loops
336 we belong to. In this case first unloop the loops, then proceed
337 normally. We may assume that e->dest is not a header of any loop,
338 as it now has exactly one predecessor. */
339 while (e->src->loop_father->outer
340 && dominated_by_p (CDI_DOMINATORS,
341 e->src->loop_father->latch, e->dest))
342 unloop (loops, e->src->loop_father);
344 /* Identify the path. */
345 nrem = find_path (e, &rem_bbs);
348 bord_bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
349 seen = sbitmap_alloc (last_basic_block);
352 /* Find "border" hexes -- i.e. those with predecessor in removed path. */
353 for (i = 0; i < nrem; i++)
354 SET_BIT (seen, rem_bbs[i]->index);
355 for (i = 0; i < nrem; i++)
358 for (ae = rem_bbs[i]->succ; ae; ae = ae->succ_next)
359 if (ae->dest != EXIT_BLOCK_PTR && !TEST_BIT (seen, ae->dest->index))
361 SET_BIT (seen, ae->dest->index);
362 bord_bbs[n_bord_bbs++] = ae->dest;
366 /* Remove the path. */
368 deleted = loop_delete_branch_edge (e, 1);
369 gcc_assert (deleted);
370 dom_bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
372 /* Cancel loops contained in the path. */
373 for (i = 0; i < nrem; i++)
374 if (rem_bbs[i]->loop_father->header == rem_bbs[i])
375 cancel_loop_tree (loops, rem_bbs[i]->loop_father);
377 remove_bbs (rem_bbs, nrem);
380 /* Find blocks whose dominators may be affected. */
383 for (i = 0; i < n_bord_bbs; i++)
387 bb = get_immediate_dominator (CDI_DOMINATORS, bord_bbs[i]);
388 if (TEST_BIT (seen, bb->index))
390 SET_BIT (seen, bb->index);
392 for (ldom = first_dom_son (CDI_DOMINATORS, bb);
394 ldom = next_dom_son (CDI_DOMINATORS, ldom))
395 if (!dominated_by_p (CDI_DOMINATORS, from, ldom))
396 dom_bbs[n_dom_bbs++] = ldom;
401 /* Recount dominators. */
402 iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, n_dom_bbs);
405 /* These blocks have lost some predecessor(s), thus their irreducible
406 status could be changed. */
407 for (i = 0; i < n_bord_bbs; i++)
408 fix_irreducible_loops (bord_bbs[i]);
411 /* Fix placements of basic blocks inside loops and the placement of
412 loops in the loop tree. */
413 fix_bb_placements (loops, from);
414 fix_loop_placements (loops, from->loop_father);
419 /* Predicate for enumeration in add_loop. */
421 alp_enum_p (basic_block bb, void *alp_header)
423 return bb != (basic_block) alp_header;
426 /* Given LOOP structure with filled header and latch, find the body of the
427 corresponding loop and add it to LOOPS tree. */
429 add_loop (struct loops *loops, struct loop *loop)
434 /* Add it to loop structure. */
435 place_new_loop (loops, loop);
438 /* Find its nodes. */
439 bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
440 n = dfs_enumerate_from (loop->latch, 1, alp_enum_p,
441 bbs, n_basic_blocks, loop->header);
443 for (i = 0; i < n; i++)
444 add_bb_to_loop (bbs[i], loop);
445 add_bb_to_loop (loop->header, loop);
450 /* Multiply all frequencies of basic blocks in array BBS of length NBBS
453 scale_bbs_frequencies (basic_block *bbs, int nbbs, int num, int den)
458 for (i = 0; i < nbbs; i++)
460 bbs[i]->frequency = (bbs[i]->frequency * num) / den;
461 bbs[i]->count = RDIV (bbs[i]->count * num, den);
462 for (e = bbs[i]->succ; e; e = e->succ_next)
463 e->count = (e->count * num) /den;
467 /* Multiply all frequencies in LOOP by NUM/DEN. */
469 scale_loop_frequencies (struct loop *loop, int num, int den)
473 bbs = get_loop_body (loop);
474 scale_bbs_frequencies (bbs, loop->num_nodes, num, den);
478 /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting
479 latch to header and update loop tree stored in LOOPS and dominators
480 accordingly. Everything between them plus LATCH_EDGE destination must
481 be dominated by HEADER_EDGE destination, and back-reachable from
482 LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB,
483 FALLTHRU_EDGE (SWITCH_BB) to original destination of HEADER_EDGE and
484 BRANCH_EDGE (SWITCH_BB) to original destination of LATCH_EDGE.
485 Returns newly created loop. */
488 loopify (struct loops *loops, edge latch_edge, edge header_edge,
489 basic_block switch_bb)
491 basic_block succ_bb = latch_edge->dest;
492 basic_block pred_bb = header_edge->src;
493 basic_block *dom_bbs, *body;
494 unsigned n_dom_bbs, i;
496 struct loop *loop = xcalloc (1, sizeof (struct loop));
497 struct loop *outer = succ_bb->loop_father->outer;
498 int freq, prob, tot_prob;
502 loop->header = header_edge->dest;
503 loop->latch = latch_edge->src;
505 freq = EDGE_FREQUENCY (header_edge);
506 cnt = header_edge->count;
507 prob = switch_bb->succ->probability;
508 tot_prob = prob + switch_bb->succ->succ_next->probability;
512 /* Redirect edges. */
513 loop_redirect_edge (latch_edge, loop->header);
514 loop_redirect_edge (BRANCH_EDGE (switch_bb), succ_bb);
516 loop_redirect_edge (header_edge, switch_bb);
517 loop_redirect_edge (FALLTHRU_EDGE (switch_bb), loop->header);
519 /* Update dominators. */
520 set_immediate_dominator (CDI_DOMINATORS, switch_bb, pred_bb);
521 set_immediate_dominator (CDI_DOMINATORS, loop->header, switch_bb);
523 set_immediate_dominator (CDI_DOMINATORS, succ_bb, switch_bb);
525 /* Compute new loop. */
526 add_loop (loops, loop);
527 flow_loop_tree_node_add (outer, loop);
529 /* Add switch_bb to appropriate loop. */
530 add_bb_to_loop (switch_bb, outer);
532 /* Fix frequencies. */
533 switch_bb->frequency = freq;
534 switch_bb->count = cnt;
535 for (e = switch_bb->succ; e; e = e->succ_next)
536 e->count = (switch_bb->count * e->probability) / REG_BR_PROB_BASE;
537 scale_loop_frequencies (loop, prob, tot_prob);
538 scale_loop_frequencies (succ_bb->loop_father, tot_prob - prob, tot_prob);
540 /* Update dominators of blocks outside of LOOP. */
541 dom_bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
543 seen = sbitmap_alloc (last_basic_block);
545 body = get_loop_body (loop);
547 for (i = 0; i < loop->num_nodes; i++)
548 SET_BIT (seen, body[i]->index);
550 for (i = 0; i < loop->num_nodes; i++)
554 for (ldom = first_dom_son (CDI_DOMINATORS, body[i]);
556 ldom = next_dom_son (CDI_DOMINATORS, ldom))
557 if (!TEST_BIT (seen, ldom->index))
559 SET_BIT (seen, ldom->index);
560 dom_bbs[n_dom_bbs++] = ldom;
564 iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, n_dom_bbs);
573 /* Remove the latch edge of a LOOP and update LOOPS tree to indicate that
574 the LOOP was removed. After this function, original loop latch will
575 have no successor, which caller is expected to fix somehow. */
577 unloop (struct loops *loops, struct loop *loop)
582 basic_block latch = loop->latch;
586 /* This is relatively straightforward. The dominators are unchanged, as
587 loop header dominates loop latch, so the only thing we have to care of
588 is the placement of loops and basic blocks inside the loop tree. We
589 move them all to the loop->outer, and then let fix_bb_placements do
592 body = get_loop_body (loop);
593 edges = get_loop_exit_edges (loop, &n_edges);
595 for (i = 0; i < n; i++)
596 if (body[i]->loop_father == loop)
598 remove_bb_from_loops (body[i]);
599 add_bb_to_loop (body[i], loop->outer);
606 flow_loop_tree_node_remove (ploop);
607 flow_loop_tree_node_add (loop->outer, ploop);
610 /* Remove the loop and free its data. */
611 flow_loop_tree_node_remove (loop);
612 loops->parray[loop->num] = NULL;
613 flow_loop_free (loop);
615 remove_edge (latch->succ);
616 fix_bb_placements (loops, latch);
618 /* If the loop was inside an irreducible region, we would have to somehow
619 update the irreducible marks inside its body. While it is certainly
620 possible to do, it is a bit complicated and this situation should be
621 very rare, so we just remark all loops in this case. */
622 for (i = 0; i < n_edges; i++)
623 if (edges[i]->flags & EDGE_IRREDUCIBLE_LOOP)
626 mark_irreducible_loops (loops);
630 /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop
631 FATHER of LOOP such that all of the edges coming out of LOOP belong to
632 FATHER, and set it as outer loop of LOOP. Return 1 if placement of
635 fix_loop_placement (struct loop *loop)
640 struct loop *father = loop->pred[0], *act;
642 body = get_loop_body (loop);
643 for (i = 0; i < loop->num_nodes; i++)
644 for (e = body[i]->succ; e; e = e->succ_next)
645 if (!flow_bb_inside_loop_p (loop, e->dest))
647 act = find_common_loop (loop, e->dest->loop_father);
648 if (flow_loop_nested_p (father, act))
653 if (father != loop->outer)
655 for (act = loop->outer; act != father; act = act->outer)
656 act->num_nodes -= loop->num_nodes;
657 flow_loop_tree_node_remove (loop);
658 flow_loop_tree_node_add (father, loop);
664 /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that
665 condition stated in description of fix_loop_placement holds for them.
666 It is used in case when we removed some edges coming out of LOOP, which
667 may cause the right placement of LOOP inside loop tree to change. */
669 fix_loop_placements (struct loops *loops, struct loop *loop)
676 if (!fix_loop_placement (loop))
679 /* Changing the placement of a loop in the loop tree may alter the
680 validity of condition 2) of the description of fix_bb_placement
681 for its preheader, because the successor is the header and belongs
682 to the loop. So call fix_bb_placements to fix up the placement
683 of the preheader and (possibly) of its predecessors. */
684 fix_bb_placements (loops, loop_preheader_edge (loop)->src);
689 /* Creates place for a new LOOP in LOOPS structure. */
691 place_new_loop (struct loops *loops, struct loop *loop)
694 xrealloc (loops->parray, (loops->num + 1) * sizeof (struct loop *));
695 loops->parray[loops->num] = loop;
697 loop->num = loops->num++;
700 /* Copies copy of LOOP as subloop of TARGET loop, placing newly
701 created loop into LOOPS structure. */
703 duplicate_loop (struct loops *loops, struct loop *loop, struct loop *target)
706 cloop = xcalloc (1, sizeof (struct loop));
707 place_new_loop (loops, cloop);
709 /* Initialize copied loop. */
710 cloop->level = loop->level;
712 /* Set it as copy of loop. */
715 /* Add it to target. */
716 flow_loop_tree_node_add (target, cloop);
721 /* Copies structure of subloops of LOOP into TARGET loop, placing
722 newly created loops into loop tree stored in LOOPS. */
724 duplicate_subloops (struct loops *loops, struct loop *loop, struct loop *target)
726 struct loop *aloop, *cloop;
728 for (aloop = loop->inner; aloop; aloop = aloop->next)
730 cloop = duplicate_loop (loops, aloop, target);
731 duplicate_subloops (loops, aloop, cloop);
735 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS,
736 into TARGET loop, placing newly created loops into loop tree LOOPS. */
738 copy_loops_to (struct loops *loops, struct loop **copied_loops, int n, struct loop *target)
743 for (i = 0; i < n; i++)
745 aloop = duplicate_loop (loops, copied_loops[i], target);
746 duplicate_subloops (loops, copied_loops[i], aloop);
750 /* Redirects edge E to basic block DEST. */
752 loop_redirect_edge (edge e, basic_block dest)
757 redirect_edge_and_branch_force (e, dest);
760 /* Deletes edge E from a branch if possible. Unless REALLY_DELETE is set,
761 just test whether it is possible to remove the edge. */
763 loop_delete_branch_edge (edge e, int really_delete)
765 basic_block src = e->src;
770 gcc_assert (src->succ->succ_next);
772 /* Cannot handle more than two exit edges. */
773 if (src->succ->succ_next->succ_next)
775 /* And it must be just a simple branch. */
776 if (!any_condjump_p (BB_END (src)))
779 snd = e == src->succ ? src->succ->succ_next : src->succ;
781 if (newdest == EXIT_BLOCK_PTR)
784 /* Hopefully the above conditions should suffice. */
788 /* Redirecting behaves wrongly wrto this flag. */
789 irr = snd->flags & EDGE_IRREDUCIBLE_LOOP;
791 if (!redirect_edge_and_branch (e, newdest))
793 src->succ->flags &= ~EDGE_IRREDUCIBLE_LOOP;
794 src->succ->flags |= irr;
799 /* Check whether LOOP's body can be duplicated. */
801 can_duplicate_loop_p (struct loop *loop)
804 basic_block *bbs = get_loop_body (loop);
806 ret = can_copy_bbs_p (bbs, loop->num_nodes);
812 /* The NBBS blocks in BBS will get duplicated and the copies will be placed
813 to LOOP. Update the single_exit information in superloops of LOOP. */
816 update_single_exits_after_duplication (basic_block *bbs, unsigned nbbs,
821 for (i = 0; i < nbbs; i++)
822 bbs[i]->rbi->duplicated = 1;
824 for (; loop->outer; loop = loop->outer)
826 if (!loop->single_exit)
829 if (loop->single_exit->src->rbi->duplicated)
830 loop->single_exit = NULL;
833 for (i = 0; i < nbbs; i++)
834 bbs[i]->rbi->duplicated = 0;
838 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating
839 LOOPS structure and dominators. E's destination must be LOOP header for
840 this to work, i.e. it must be entry or latch edge of this loop; these are
841 unique, as the loops must have preheaders for this function to work
842 correctly (in case E is latch, the function unrolls the loop, if E is entry
843 edge, it peels the loop). Store edges created by copying ORIG edge from
844 copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to
845 original LOOP body, the other copies are numbered in order given by control
846 flow through them) into TO_REMOVE array. Returns false if duplication is
849 duplicate_loop_to_header_edge (struct loop *loop, edge e, struct loops *loops,
850 unsigned int ndupl, sbitmap wont_exit,
851 edge orig, edge *to_remove,
852 unsigned int *n_to_remove, int flags)
854 struct loop *target, *aloop;
855 struct loop **orig_loops;
856 unsigned n_orig_loops;
857 basic_block header = loop->header, latch = loop->latch;
858 basic_block *new_bbs, *bbs, *first_active;
859 basic_block new_bb, bb, first_active_latch = NULL;
861 edge spec_edges[2], new_spec_edges[2];
865 int is_latch = (latch == e->src);
866 int scale_act = 0, *scale_step = NULL, scale_main = 0;
867 int p, freq_in, freq_le, freq_out_orig;
868 int prob_pass_thru, prob_pass_wont_exit, prob_pass_main;
869 int add_irreducible_flag;
871 gcc_assert (e->dest == loop->header);
872 gcc_assert (ndupl > 0);
876 /* Orig must be edge out of the loop. */
877 gcc_assert (flow_bb_inside_loop_p (loop, orig->src));
878 gcc_assert (!flow_bb_inside_loop_p (loop, orig->dest));
881 bbs = get_loop_body (loop);
883 /* Check whether duplication is possible. */
884 if (!can_copy_bbs_p (bbs, loop->num_nodes))
889 new_bbs = xmalloc (sizeof (basic_block) * loop->num_nodes);
891 /* In case we are doing loop peeling and the loop is in the middle of
892 irreducible region, the peeled copies will be inside it too. */
893 add_irreducible_flag = e->flags & EDGE_IRREDUCIBLE_LOOP;
894 gcc_assert (!is_latch || !add_irreducible_flag);
896 /* Find edge from latch. */
897 latch_edge = loop_latch_edge (loop);
899 if (flags & DLTHE_FLAG_UPDATE_FREQ)
901 /* Calculate coefficients by that we have to scale frequencies
902 of duplicated loop bodies. */
903 freq_in = header->frequency;
904 freq_le = EDGE_FREQUENCY (latch_edge);
907 if (freq_in < freq_le)
909 freq_out_orig = orig ? EDGE_FREQUENCY (orig) : freq_in - freq_le;
910 if (freq_out_orig > freq_in - freq_le)
911 freq_out_orig = freq_in - freq_le;
912 prob_pass_thru = RDIV (REG_BR_PROB_BASE * freq_le, freq_in);
913 prob_pass_wont_exit =
914 RDIV (REG_BR_PROB_BASE * (freq_le + freq_out_orig), freq_in);
916 scale_step = xmalloc (ndupl * sizeof (int));
918 for (i = 1; i <= ndupl; i++)
919 scale_step[i - 1] = TEST_BIT (wont_exit, i)
920 ? prob_pass_wont_exit
925 prob_pass_main = TEST_BIT (wont_exit, 0)
926 ? prob_pass_wont_exit
929 scale_main = REG_BR_PROB_BASE;
930 for (i = 0; i < ndupl; i++)
933 p = RDIV (p * scale_step[i], REG_BR_PROB_BASE);
935 scale_main = RDIV (REG_BR_PROB_BASE * REG_BR_PROB_BASE, scale_main);
936 scale_act = RDIV (scale_main * prob_pass_main, REG_BR_PROB_BASE);
940 scale_main = REG_BR_PROB_BASE;
941 for (i = 0; i < ndupl; i++)
942 scale_main = RDIV (scale_main * scale_step[i], REG_BR_PROB_BASE);
943 scale_act = REG_BR_PROB_BASE - prob_pass_thru;
945 for (i = 0; i < ndupl; i++)
946 gcc_assert (scale_step[i] >= 0 && scale_step[i] <= REG_BR_PROB_BASE);
947 gcc_assert (scale_main >= 0 && scale_main <= REG_BR_PROB_BASE
948 && scale_act >= 0 && scale_act <= REG_BR_PROB_BASE);
951 /* Loop the new bbs will belong to. */
952 target = e->src->loop_father;
954 /* Original loops. */
956 for (aloop = loop->inner; aloop; aloop = aloop->next)
958 orig_loops = xcalloc (n_orig_loops, sizeof (struct loop *));
959 for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++)
960 orig_loops[i] = aloop;
966 first_active = xmalloc (n * sizeof (basic_block));
969 memcpy (first_active, bbs, n * sizeof (basic_block));
970 first_active_latch = latch;
973 /* Update the information about single exits. */
974 if (loops->state & LOOPS_HAVE_MARKED_SINGLE_EXITS)
975 update_single_exits_after_duplication (bbs, n, target);
977 /* Record exit edge in original loop body. */
978 if (orig && TEST_BIT (wont_exit, 0))
979 to_remove[(*n_to_remove)++] = orig;
981 spec_edges[SE_ORIG] = orig;
982 spec_edges[SE_LATCH] = latch_edge;
984 for (j = 0; j < ndupl; j++)
987 copy_loops_to (loops, orig_loops, n_orig_loops, target);
990 copy_bbs (bbs, n, new_bbs, spec_edges, 2, new_spec_edges, loop);
992 for (i = 0; i < n; i++)
993 new_bbs[i]->rbi->copy_number = j + 1;
995 /* Note whether the blocks and edges belong to an irreducible loop. */
996 if (add_irreducible_flag)
998 for (i = 0; i < n; i++)
999 new_bbs[i]->rbi->duplicated = 1;
1000 for (i = 0; i < n; i++)
1002 new_bb = new_bbs[i];
1003 if (new_bb->loop_father == target)
1004 new_bb->flags |= BB_IRREDUCIBLE_LOOP;
1006 for (ae = new_bb->succ; ae; ae = ae->succ_next)
1007 if (ae->dest->rbi->duplicated
1008 && (ae->src->loop_father == target
1009 || ae->dest->loop_father == target))
1010 ae->flags |= EDGE_IRREDUCIBLE_LOOP;
1012 for (i = 0; i < n; i++)
1013 new_bbs[i]->rbi->duplicated = 0;
1016 /* Redirect the special edges. */
1019 redirect_edge_and_branch_force (latch_edge, new_bbs[0]);
1020 redirect_edge_and_branch_force (new_spec_edges[SE_LATCH],
1022 set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], latch);
1023 latch = loop->latch = new_bbs[1];
1024 e = latch_edge = new_spec_edges[SE_LATCH];
1028 redirect_edge_and_branch_force (new_spec_edges[SE_LATCH],
1030 redirect_edge_and_branch_force (e, new_bbs[0]);
1031 set_immediate_dominator (CDI_DOMINATORS, new_bbs[0], e->src);
1032 e = new_spec_edges[SE_LATCH];
1035 /* Record exit edge in this copy. */
1036 if (orig && TEST_BIT (wont_exit, j + 1))
1037 to_remove[(*n_to_remove)++] = new_spec_edges[SE_ORIG];
1039 /* Record the first copy in the control flow order if it is not
1040 the original loop (i.e. in case of peeling). */
1041 if (!first_active_latch)
1043 memcpy (first_active, new_bbs, n * sizeof (basic_block));
1044 first_active_latch = new_bbs[1];
1047 /* Set counts and frequencies. */
1048 if (flags & DLTHE_FLAG_UPDATE_FREQ)
1050 scale_bbs_frequencies (new_bbs, n, scale_act, REG_BR_PROB_BASE);
1051 scale_act = RDIV (scale_act * scale_step[j], REG_BR_PROB_BASE);
1057 /* Update the original loop. */
1059 set_immediate_dominator (CDI_DOMINATORS, e->dest, e->src);
1060 if (flags & DLTHE_FLAG_UPDATE_FREQ)
1062 scale_bbs_frequencies (bbs, n, scale_main, REG_BR_PROB_BASE);
1066 /* Update dominators of outer blocks if affected. */
1067 for (i = 0; i < n; i++)
1069 basic_block dominated, dom_bb, *dom_bbs;
1073 bb->rbi->copy_number = 0;
1075 n_dom_bbs = get_dominated_by (CDI_DOMINATORS, bb, &dom_bbs);
1076 for (j = 0; j < n_dom_bbs; j++)
1078 dominated = dom_bbs[j];
1079 if (flow_bb_inside_loop_p (loop, dominated))
1081 dom_bb = nearest_common_dominator (
1082 CDI_DOMINATORS, first_active[i], first_active_latch);
1083 set_immediate_dominator (CDI_DOMINATORS, dominated, dom_bb);
1087 free (first_active);
1094 /* A callback for make_forwarder block, to redirect all edges except for
1095 MFB_KJ_EDGE to the entry part. E is the edge for that we should decide
1096 whether to redirect it. */
1098 static edge mfb_kj_edge;
1100 mfb_keep_just (edge e)
1102 return e != mfb_kj_edge;
1105 /* A callback for make_forwarder block, to update data structures for a basic
1106 block JUMP created by redirecting an edge (only the latch edge is being
1110 mfb_update_loops (basic_block jump)
1112 struct loop *loop = jump->succ->dest->loop_father;
1114 if (dom_computed[CDI_DOMINATORS])
1115 set_immediate_dominator (CDI_DOMINATORS, jump, jump->pred->src);
1116 add_bb_to_loop (jump, loop);
1120 /* Creates a pre-header for a LOOP. Returns newly created block. Unless
1121 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single
1122 entry; otherwise we also force preheader block to have only one successor.
1123 The function also updates dominators. */
1126 create_preheader (struct loop *loop, int flags)
1130 struct loop *cloop, *ploop;
1134 cloop = loop->outer;
1136 for (e = loop->header->pred; e; e = e->pred_next)
1138 if (e->src == loop->latch)
1140 irred |= (e->flags & EDGE_IRREDUCIBLE_LOOP) != 0;
1143 gcc_assert (nentry);
1146 for (e = loop->header->pred; e->src == loop->latch; e = e->pred_next);
1147 if (!(flags & CP_SIMPLE_PREHEADERS)
1148 || !e->src->succ->succ_next)
1152 mfb_kj_edge = loop_latch_edge (loop);
1153 fallthru = make_forwarder_block (loop->header, mfb_keep_just,
1155 dummy = fallthru->src;
1156 loop->header = fallthru->dest;
1158 /* The header could be a latch of some superloop(s); due to design of
1159 split_block, it would now move to fallthru->dest. */
1160 for (ploop = loop; ploop; ploop = ploop->outer)
1161 if (ploop->latch == dummy)
1162 ploop->latch = fallthru->dest;
1164 /* Reorganize blocks so that the preheader is not stuck in the middle of the
1166 for (e = dummy->pred; e; e = e->pred_next)
1167 if (e->src != loop->latch)
1169 move_block_after (dummy, e->src);
1171 loop->header->loop_father = loop;
1172 add_bb_to_loop (dummy, cloop);
1176 dummy->flags |= BB_IRREDUCIBLE_LOOP;
1177 dummy->succ->flags |= EDGE_IRREDUCIBLE_LOOP;
1181 fprintf (dump_file, "Created preheader block for loop %i\n",
1187 /* Create preheaders for each loop from loop tree stored in LOOPS; for meaning
1188 of FLAGS see create_preheader. */
1190 create_preheaders (struct loops *loops, int flags)
1193 for (i = 1; i < loops->num; i++)
1194 create_preheader (loops->parray[i], flags);
1195 loops->state |= LOOPS_HAVE_PREHEADERS;
1198 /* Forces all loop latches of loops from loop tree LOOPS to have only single
1201 force_single_succ_latches (struct loops *loops)
1207 for (i = 1; i < loops->num; i++)
1209 loop = loops->parray[i];
1210 if (loop->latch != loop->header
1211 && !loop->latch->succ->succ_next)
1214 for (e = loop->header->pred; e->src != loop->latch; e = e->pred_next)
1217 loop_split_edge_with (e, NULL_RTX);
1219 loops->state |= LOOPS_HAVE_SIMPLE_LATCHES;
1222 /* A quite stupid function to put INSNS on edge E. They are supposed to form
1223 just one basic block. Jumps in INSNS are not handled, so cfg do not have to
1224 be ok after this function. The created block is placed on correct place
1225 in LOOPS structure and its dominator is set. */
1227 loop_split_edge_with (edge e, rtx insns)
1229 basic_block src, dest, new_bb;
1230 struct loop *loop_c;
1236 loop_c = find_common_loop (src->loop_father, dest->loop_father);
1238 /* Create basic block for it. */
1240 new_bb = split_edge (e);
1241 add_bb_to_loop (new_bb, loop_c);
1242 new_bb->flags = insns ? BB_SUPERBLOCK : 0;
1244 new_e = new_bb->succ;
1245 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
1247 new_bb->flags |= BB_IRREDUCIBLE_LOOP;
1248 new_e->flags |= EDGE_IRREDUCIBLE_LOOP;
1252 emit_insn_after (insns, BB_END (new_bb));
1254 if (dest->loop_father->latch == src)
1255 dest->loop_father->latch = new_bb;
1260 /* Uses the natural loop discovery to recreate loop notes. */
1262 create_loop_notes (void)
1264 rtx insn, head, end;
1267 basic_block *first, *last, bb, pbb;
1268 struct loop **stack, **top;
1270 #ifdef ENABLE_CHECKING
1271 /* Verify that there really are no loop notes. */
1272 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
1273 gcc_assert (!NOTE_P (insn) ||
1274 NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
1277 flow_loops_find (&loops, LOOP_TREE);
1278 free_dominance_info (CDI_DOMINATORS);
1281 last = xcalloc (loops.num, sizeof (basic_block));
1285 for (loop = bb->loop_father; loop->outer; loop = loop->outer)
1286 last[loop->num] = bb;
1289 first = xcalloc (loops.num, sizeof (basic_block));
1290 stack = xcalloc (loops.num, sizeof (struct loop *));
1295 for (loop = bb->loop_father; loop->outer; loop = loop->outer)
1297 if (!first[loop->num])
1300 first[loop->num] = bb;
1303 if (bb == last[loop->num])
1305 /* Prevent loops from overlapping. */
1306 while (*--top != loop)
1307 last[(*top)->num] = EXIT_BLOCK_PTR;
1309 /* If loop starts with jump into it, place the note in
1310 front of the jump. */
1311 insn = PREV_INSN (BB_HEAD (first[loop->num]));
1313 && BARRIER_P (insn))
1314 insn = PREV_INSN (insn);
1318 && any_uncondjump_p (insn)
1319 && onlyjump_p (insn))
1321 pbb = BLOCK_FOR_INSN (insn);
1322 gcc_assert (pbb && pbb->succ && !pbb->succ->succ_next);
1324 if (!flow_bb_inside_loop_p (loop, pbb->succ->dest))
1325 insn = BB_HEAD (first[loop->num]);
1328 insn = BB_HEAD (first[loop->num]);
1330 head = BB_HEAD (first[loop->num]);
1331 emit_note_before (NOTE_INSN_LOOP_BEG, insn);
1332 BB_HEAD (first[loop->num]) = head;
1334 /* Position the note correctly wrto barrier. */
1335 insn = BB_END (last[loop->num]);
1336 if (NEXT_INSN (insn)
1337 && BARRIER_P (NEXT_INSN (insn)))
1338 insn = NEXT_INSN (insn);
1340 end = BB_END (last[loop->num]);
1341 emit_note_after (NOTE_INSN_LOOP_END, insn);
1342 BB_END (last[loop->num]) = end;
1351 flow_loops_free (&loops);