1 /* Loop manipulation code for GNU compiler.
2 Copyright (C) 2002, 2003 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 struct loop * duplicate_loop (struct loops *, struct loop *,
34 static void duplicate_subloops (struct loops *, struct loop *, struct loop *);
35 static void copy_loops_to (struct loops *, struct loop **, int,
37 static void loop_redirect_edge (edge, basic_block);
38 static bool loop_delete_branch_edge (edge, int);
39 static void copy_bbs (basic_block *, int, edge, edge, basic_block **,
40 struct loops *, edge *, edge *, int);
41 static void remove_bbs (dominance_info, basic_block *, int);
42 static bool rpe_enum_p (basic_block, void *);
43 static int find_path (edge, dominance_info, basic_block **);
44 static bool alp_enum_p (basic_block, void *);
45 static void add_loop (struct loops *, struct loop *);
46 static void fix_loop_placements (struct loop *);
47 static bool fix_bb_placement (struct loops *, basic_block);
48 static void fix_bb_placements (struct loops *, basic_block);
49 static void place_new_loop (struct loops *, struct loop *);
50 static void scale_loop_frequencies (struct loop *, int, int);
51 static void scale_bbs_frequencies (basic_block *, int, int, int);
52 static void record_exit_edges (edge, basic_block *, int, edge *, unsigned *,
54 static basic_block create_preheader (struct loop *, dominance_info, int);
55 static void fix_irreducible_loops (basic_block);
57 /* Splits basic block BB after INSN, returns created edge. Updates loops
60 split_loop_bb (struct loops *loops, basic_block bb, rtx insn)
66 /* Split the block. */
67 e = split_block (bb, insn);
69 /* Add dest to loop. */
70 add_bb_to_loop (e->dest, e->src->loop_father);
73 add_to_dominance_info (loops->cfg.dom, e->dest);
74 n_dom_bbs = get_dominated_by (loops->cfg.dom, e->src, &dom_bbs);
75 for (i = 0; i < n_dom_bbs; i++)
76 set_immediate_dominator (loops->cfg.dom, dom_bbs[i], e->dest);
78 set_immediate_dominator (loops->cfg.dom, e->dest, e->src);
83 /* Checks whether basic block BB is dominated by RPE->DOM, where
84 RPE is passed through DATA. */
92 rpe_enum_p (basic_block bb, void *data)
94 struct rpe_data *rpe = data;
95 return dominated_by_p (rpe->doms, bb, rpe->dom);
98 /* Remove basic blocks BBS from loop structure and dominance info,
99 and delete them afterwards. */
101 remove_bbs (dominance_info dom, basic_block *bbs, int nbbs)
105 for (i = 0; i < nbbs; i++)
107 remove_bb_from_loops (bbs[i]);
108 delete_from_dominance_info (dom, bbs[i]);
109 delete_block (bbs[i]);
113 /* Find path -- i.e. the basic blocks dominated by edge E and put them
114 into array BBS, that will be allocated large enough to contain them.
115 E->dest must have exactly one predecessor for this to work (it is
116 easy to achieve and we do not put it here because we do not want to
117 alter anything by this function). The number of basic blocks in the
120 find_path (edge e, dominance_info doms, basic_block **bbs)
124 if (e->dest->pred->pred_next)
127 /* Find bbs in the path. */
130 *bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
131 return dfs_enumerate_from (e->dest, 0, rpe_enum_p, *bbs,
132 n_basic_blocks, &rpe);
135 /* Fix placement of basic block BB inside loop hierarchy stored in LOOPS --
136 Let L be a loop to that BB belongs. Then every successor of BB must either
137 1) belong to some superloop of loop L, or
138 2) be a header of loop K such that K->outer is superloop of L
139 Returns true if we had to move BB into other loop to enforce this condition,
140 false if the placement of BB was already correct (provided that placements
141 of its successors are correct). */
143 fix_bb_placement (struct loops *loops, basic_block bb)
146 struct loop *loop = loops->tree_root, *act;
148 for (e = bb->succ; e; e = e->succ_next)
150 if (e->dest == EXIT_BLOCK_PTR)
153 act = e->dest->loop_father;
154 if (act->header == e->dest)
157 if (flow_loop_nested_p (loop, act))
161 if (loop == bb->loop_father)
164 remove_bb_from_loops (bb);
165 add_bb_to_loop (bb, loop);
170 /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e.
171 enforce condition condition stated in description of fix_bb_placement. We
172 start from basic block FROM that had some of its successors removed, so that
173 his placement no longer has to be correct, and iteratively fix placement of
174 its predecessors that may change if placement of FROM changed. Also fix
175 placement of subloops of FROM->loop_father, that might also be altered due
176 to this change; the condition for them is simmilar, except that instead of
177 successors we consider edges coming out of the loops. */
179 fix_bb_placements (struct loops *loops, basic_block from)
182 basic_block *queue, *qtop, *qbeg, *qend;
183 struct loop *base_loop;
186 /* We pass through blocks back-reachable from FROM, testing whether some
187 of their successors moved to outer loop. It may be necessary to
188 iterate several times, but it is finite, as we stop unless we move
189 the basic block up the loop structure. The whole story is a bit
190 more complicated due to presence of subloops, those are moved using
191 fix_loop_placement. */
193 base_loop = from->loop_father;
194 if (base_loop == loops->tree_root)
197 in_queue = sbitmap_alloc (last_basic_block);
198 sbitmap_zero (in_queue);
199 SET_BIT (in_queue, from->index);
200 /* Prevent us from going out of the base_loop. */
201 SET_BIT (in_queue, base_loop->header->index);
203 queue = xmalloc ((base_loop->num_nodes + 1) * sizeof (basic_block));
204 qtop = queue + base_loop->num_nodes + 1;
215 RESET_BIT (in_queue, from->index);
217 if (from->loop_father->header == from)
219 /* Subloop header, maybe move the loop upward. */
220 if (!fix_loop_placement (from->loop_father))
225 /* Ordinary basic block. */
226 if (!fix_bb_placement (loops, from))
230 /* Something has changed, insert predecessors into queue. */
231 for (e = from->pred; e; e = e->pred_next)
233 basic_block pred = e->src;
236 if (TEST_BIT (in_queue, pred->index))
239 /* If it is subloop, then it either was not moved, or
240 the path up the loop tree from base_loop do not contain
242 nca = find_common_loop (pred->loop_father, base_loop);
243 if (pred->loop_father != base_loop
245 || nca != pred->loop_father))
246 pred = pred->loop_father->header;
247 else if (!flow_loop_nested_p (from->loop_father, pred->loop_father))
249 /* No point in processing it. */
253 if (TEST_BIT (in_queue, pred->index))
256 /* Schedule the basic block. */
261 SET_BIT (in_queue, pred->index);
268 /* Basic block from has lost one or more of its predecessors, so it might
269 mo longer be part irreducible loop. Fix it and proceed recursively
270 for its successors if needed. */
272 fix_irreducible_loops (basic_block from)
281 if (!(from->flags & BB_IRREDUCIBLE_LOOP))
284 on_stack = sbitmap_alloc (last_basic_block);
285 sbitmap_zero (on_stack);
286 SET_BIT (on_stack, from->index);
287 stack = xmalloc (from->loop_father->num_nodes * sizeof (basic_block));
293 bb = stack[--stack_top];
294 RESET_BIT (on_stack, bb->index);
296 for (e = bb->pred; e; e = e->pred_next)
297 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
302 bb->flags &= ~BB_IRREDUCIBLE_LOOP;
303 if (bb->loop_father->header == bb)
304 edges = get_loop_exit_edges (bb->loop_father, &n_edges);
308 for (e = bb->succ; e; e = e->succ_next)
310 edges = xmalloc (n_edges * sizeof (edge));
312 for (e = bb->succ; e; e = e->succ_next)
313 edges[n_edges++] = e;
316 for (i = 0; i < n_edges; i++)
317 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
319 if (!flow_bb_inside_loop_p (from->loop_father, e->dest))
322 e->flags &= ~EDGE_IRREDUCIBLE_LOOP;
323 if (TEST_BIT (on_stack, e->dest->index))
326 SET_BIT (on_stack, e->dest->index);
327 stack[stack_top++] = e->dest;
336 /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E
337 and update loop structure stored in LOOPS and dominators. Return true if
338 we were able to remove the path, false otherwise (and nothing is affected
341 remove_path (struct loops *loops, edge e)
344 basic_block *rem_bbs, *bord_bbs, *dom_bbs, from, bb;
345 int i, nrem, n_bord_bbs, n_dom_bbs;
348 if (!loop_delete_branch_edge (e, 0))
351 /* We need to check whether basic blocks are dominated by the edge
352 e, but we only have basic block dominators. This is easy to
353 fix -- when e->dest has exactly one predecessor, this corresponds
354 to blocks dominated by e->dest, if not, split the edge. */
355 if (e->dest->pred->pred_next)
356 e = loop_split_edge_with (e, NULL_RTX, loops)->pred;
358 /* It may happen that by removing path we remove one or more loops
359 we belong to. In this case first unloop the loops, then proceed
360 normally. We may assume that e->dest is not a header of any loop,
361 as it now has exactly one predecessor. */
362 while (e->src->loop_father->outer
363 && dominated_by_p (loops->cfg.dom,
364 e->src->loop_father->latch, e->dest))
365 unloop (loops, e->src->loop_father);
367 /* Identify the path. */
368 nrem = find_path (e, loops->cfg.dom, &rem_bbs);
371 bord_bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
372 seen = sbitmap_alloc (last_basic_block);
375 /* Find "border" hexes -- i.e. those with predecessor in removed path. */
376 for (i = 0; i < nrem; i++)
377 SET_BIT (seen, rem_bbs[i]->index);
378 for (i = 0; i < nrem; i++)
381 for (ae = rem_bbs[i]->succ; ae; ae = ae->succ_next)
382 if (ae->dest != EXIT_BLOCK_PTR && !TEST_BIT (seen, ae->dest->index))
384 SET_BIT (seen, ae->dest->index);
385 bord_bbs[n_bord_bbs++] = ae->dest;
389 /* Remove the path. */
391 if (!loop_delete_branch_edge (e, 1))
393 dom_bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
395 /* Cancel loops contained in the path. */
396 for (i = 0; i < nrem; i++)
397 if (rem_bbs[i]->loop_father->header == rem_bbs[i])
398 cancel_loop_tree (loops, rem_bbs[i]->loop_father);
400 remove_bbs (loops->cfg.dom, rem_bbs, nrem);
403 /* Find blocks whose dominators may be affected. */
406 for (i = 0; i < n_bord_bbs; i++)
411 bb = get_immediate_dominator (loops->cfg.dom, bord_bbs[i]);
412 if (TEST_BIT (seen, bb->index))
414 SET_BIT (seen, bb->index);
416 nldom = get_dominated_by (loops->cfg.dom, bb, &ldom);
417 for (j = 0; j < nldom; j++)
418 if (!dominated_by_p (loops->cfg.dom, from, ldom[j]))
419 dom_bbs[n_dom_bbs++] = ldom[j];
425 /* Recount dominators. */
426 iterate_fix_dominators (loops->cfg.dom, dom_bbs, n_dom_bbs);
429 /* These blocks have lost some predecessor(s), thus their irreducible
430 status could be changed. */
431 for (i = 0; i < n_bord_bbs; i++)
432 fix_irreducible_loops (bord_bbs[i]);
435 /* Fix placements of basic blocks inside loops and the placement of
436 loops in the loop tree. */
437 fix_bb_placements (loops, from);
438 fix_loop_placements (from->loop_father);
443 /* Predicate for enumeration in add_loop. */
445 alp_enum_p (basic_block bb, void *alp_header)
447 return bb != (basic_block) alp_header;
450 /* Given LOOP structure with filled header and latch, find the body of the
451 corresponding loop and add it to LOOPS tree. */
453 add_loop (struct loops *loops, struct loop *loop)
458 /* Add it to loop structure. */
459 place_new_loop (loops, loop);
462 /* Find its nodes. */
463 bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
464 n = dfs_enumerate_from (loop->latch, 1, alp_enum_p,
465 bbs, n_basic_blocks, loop->header);
467 for (i = 0; i < n; i++)
468 add_bb_to_loop (bbs[i], loop);
469 add_bb_to_loop (loop->header, loop);
474 /* Multiply all frequencies of basic blocks in array BBS of lenght NBBS
477 scale_bbs_frequencies (basic_block *bbs, int nbbs, int num, int den)
482 for (i = 0; i < nbbs; i++)
484 bbs[i]->frequency = (bbs[i]->frequency * num) / den;
485 bbs[i]->count = (bbs[i]->count * num) / den;
486 for (e = bbs[i]->succ; e; e = e->succ_next)
487 e->count = (e->count * num) /den;
491 /* Multiply all frequencies in LOOP by NUM/DEN. */
493 scale_loop_frequencies (struct loop *loop, int num, int den)
497 bbs = get_loop_body (loop);
498 scale_bbs_frequencies (bbs, loop->num_nodes, num, den);
502 /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting
503 latch to header and update loop tree stored in LOOPS and dominators
504 accordingly. Everything between them plus LATCH_EDGE destination must
505 be dominated by HEADER_EDGE destination, and back-reachable from
506 LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB,
507 SWITCH_BB->succ to original destination of LATCH_EDGE and
508 SWITCH_BB->succ->succ_next to original destination of HEADER_EDGE.
509 Returns newly created loop. */
511 loopify (struct loops *loops, edge latch_edge, edge header_edge, basic_block switch_bb)
513 basic_block succ_bb = latch_edge->dest;
514 basic_block pred_bb = header_edge->src;
515 basic_block *dom_bbs, *body;
516 unsigned n_dom_bbs, i, j;
518 struct loop *loop = xcalloc (1, sizeof (struct loop));
519 struct loop *outer = succ_bb->loop_father->outer;
520 int freq, prob, tot_prob;
524 loop->header = header_edge->dest;
525 loop->latch = latch_edge->src;
527 freq = EDGE_FREQUENCY (header_edge);
528 cnt = header_edge->count;
529 prob = switch_bb->succ->probability;
530 tot_prob = prob + switch_bb->succ->succ_next->probability;
534 /* Redirect edges. */
535 loop_redirect_edge (latch_edge, loop->header);
536 loop_redirect_edge (header_edge, switch_bb);
537 loop_redirect_edge (switch_bb->succ->succ_next, loop->header);
538 loop_redirect_edge (switch_bb->succ, succ_bb);
540 /* Update dominators. */
541 set_immediate_dominator (loops->cfg.dom, switch_bb, pred_bb);
542 set_immediate_dominator (loops->cfg.dom, loop->header, switch_bb);
543 set_immediate_dominator (loops->cfg.dom, succ_bb, switch_bb);
545 /* Compute new loop. */
546 add_loop (loops, loop);
547 flow_loop_tree_node_add (outer, loop);
549 /* Add switch_bb to appropriate loop. */
550 add_bb_to_loop (switch_bb, outer);
552 /* Fix frequencies. */
553 switch_bb->frequency = freq;
554 switch_bb->count = cnt;
555 for (e = switch_bb->succ; e; e = e->succ_next)
556 e->count = (switch_bb->count * e->probability) / REG_BR_PROB_BASE;
557 scale_loop_frequencies (loop, prob, tot_prob);
558 scale_loop_frequencies (succ_bb->loop_father, tot_prob - prob, tot_prob);
560 /* Update dominators of blocks outside of LOOP. */
561 dom_bbs = xcalloc (n_basic_blocks, sizeof (basic_block));
563 seen = sbitmap_alloc (last_basic_block);
565 body = get_loop_body (loop);
567 for (i = 0; i < loop->num_nodes; i++)
568 SET_BIT (seen, body[i]->index);
570 for (i = 0; i < loop->num_nodes; i++)
575 nldom = get_dominated_by (loops->cfg.dom, body[i], &ldom);
576 for (j = 0; j < nldom; j++)
577 if (!TEST_BIT (seen, ldom[j]->index))
579 SET_BIT (seen, ldom[j]->index);
580 dom_bbs[n_dom_bbs++] = ldom[j];
585 iterate_fix_dominators (loops->cfg.dom, dom_bbs, n_dom_bbs);
594 /* Remove the latch edge of a LOOP and update LOOPS tree to indicate that
595 the LOOP was removed. After this function, original loop latch will
596 have no successor, which caller is expected to fix somehow. */
598 unloop (struct loops *loops, struct loop *loop)
603 basic_block latch = loop->latch;
607 /* This is relatively straigtforward. The dominators are unchanged, as
608 loop header dominates loop latch, so the only thing we have to care of
609 is the placement of loops and basic blocks inside the loop tree. We
610 move them all to the loop->outer, and then let fix_bb_placements do
613 body = get_loop_body (loop);
614 edges = get_loop_exit_edges (loop, &n_edges);
616 for (i = 0; i < n; i++)
617 if (body[i]->loop_father == loop)
619 remove_bb_from_loops (body[i]);
620 add_bb_to_loop (body[i], loop->outer);
627 flow_loop_tree_node_remove (ploop);
628 flow_loop_tree_node_add (loop->outer, ploop);
631 /* Remove the loop and free its data. */
632 flow_loop_tree_node_remove (loop);
633 loops->parray[loop->num] = NULL;
634 flow_loop_free (loop);
636 remove_edge (latch->succ);
637 fix_bb_placements (loops, latch);
639 /* If the loop was inside an irreducible region, we would have to somehow
640 update the irreducible marks inside its body. While it is certainly
641 possible to do, it is a bit complicated and this situation should be
642 very rare, so we just remark all loops in this case. */
643 for (i = 0; i < n_edges; i++)
644 if (edges[i]->flags & EDGE_IRREDUCIBLE_LOOP)
647 mark_irreducible_loops (loops);
651 /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop
652 FATHER of LOOP such that all of the edges comming out of LOOP belong to
653 FATHER, and set it as outer loop of LOOP. Return 1 if placement of
656 fix_loop_placement (struct loop *loop)
661 struct loop *father = loop->pred[0], *act;
663 body = get_loop_body (loop);
664 for (i = 0; i < loop->num_nodes; i++)
665 for (e = body[i]->succ; e; e = e->succ_next)
666 if (!flow_bb_inside_loop_p (loop, e->dest))
668 act = find_common_loop (loop, e->dest->loop_father);
669 if (flow_loop_nested_p (father, act))
674 if (father != loop->outer)
676 for (act = loop->outer; act != father; act = act->outer)
677 act->num_nodes -= loop->num_nodes;
678 flow_loop_tree_node_remove (loop);
679 flow_loop_tree_node_add (father, loop);
685 /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that
686 condition stated in description of fix_loop_placement holds for them.
687 It is used in case when we removed some edges coming out of LOOP, which
688 may cause the right placement of LOOP inside loop tree to change. */
690 fix_loop_placements (struct loop *loop)
697 if (!fix_loop_placement (loop))
703 /* Creates place for a new LOOP in LOOPS structure. */
705 place_new_loop (struct loops *loops, struct loop *loop)
708 xrealloc (loops->parray, (loops->num + 1) * sizeof (struct loop *));
709 loops->parray[loops->num] = loop;
711 loop->num = loops->num++;
714 /* Copies copy of LOOP as subloop of TARGET loop, placing newly
715 created loop into LOOPS structure. */
717 duplicate_loop (struct loops *loops, struct loop *loop, struct loop *target)
720 cloop = xcalloc (1, sizeof (struct loop));
721 place_new_loop (loops, cloop);
723 /* Initialize copied loop. */
724 cloop->level = loop->level;
726 /* Set it as copy of loop. */
729 /* Add it to target. */
730 flow_loop_tree_node_add (target, cloop);
735 /* Copies structure of subloops of LOOP into TARGET loop, placing
736 newly created loops into loop tree stored in LOOPS. */
738 duplicate_subloops (struct loops *loops, struct loop *loop, struct loop *target)
740 struct loop *aloop, *cloop;
742 for (aloop = loop->inner; aloop; aloop = aloop->next)
744 cloop = duplicate_loop (loops, aloop, target);
745 duplicate_subloops (loops, aloop, cloop);
749 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS,
750 into TARGET loop, placing newly created loops into loop tree LOOPS. */
752 copy_loops_to (struct loops *loops, struct loop **copied_loops, int n, struct loop *target)
757 for (i = 0; i < n; i++)
759 aloop = duplicate_loop (loops, copied_loops[i], target);
760 duplicate_subloops (loops, copied_loops[i], aloop);
764 /* Redirects edge E to basic block DEST. */
766 loop_redirect_edge (edge e, basic_block dest)
771 redirect_edge_and_branch_force (e, dest);
774 /* Deletes edge E from a branch if possible. Unless REALLY_DELETE is set,
775 just test whether it is possible to remove the edge. */
777 loop_delete_branch_edge (edge e, int really_delete)
779 basic_block src = e->src;
783 if (src->succ->succ_next)
787 /* Cannot handle more than two exit edges. */
788 if (src->succ->succ_next->succ_next)
790 /* And it must be just a simple branch. */
791 if (!any_condjump_p (src->end))
794 snd = e == src->succ ? src->succ->succ_next : src->succ;
796 if (newdest == EXIT_BLOCK_PTR)
799 /* Hopefully the above conditions should suffice. */
803 /* Redirecting behaves wrongly wrto this flag. */
804 irr = snd->flags & EDGE_IRREDUCIBLE_LOOP;
806 if (!redirect_edge_and_branch (e, newdest))
808 src->succ->flags &= ~EDGE_IRREDUCIBLE_LOOP;
809 src->succ->flags |= irr;
815 /* Cannot happen -- we are using this only to remove an edge
820 return false; /* To avoid warning, cannot get here. */
823 /* Duplicates N basic blocks stored in array BBS (they form a body of
824 duplicated loop). Newly created basic blocks are placed into array NEW_BBS
825 that we allocate. Edges from basic blocks in BBS are also duplicated and
826 copies of those of them that lead into BBS are redirected to appropriate
827 newly created block. The function also assigns bbs into loops and updates
828 dominators. If ADD_IRREDUCIBLE_FLAG is set, newly created basic blocks that
829 are not members of any inner loop are marked irreducible.
831 Additionally, we perform following manipulation with edges:
832 We have two special edges given. LATCH_EDGE is the latch edge of the
833 duplicated loop and leads into its header (one of blocks in BBS);
834 it does not have neccessarily lead from one of the blocks, because
835 we may be copying the loop body several times in unrolling.
836 Edge ENTRY leads also leads to header, and it is either latch or entry
837 edge. Copy of LATCH_EDGE is redirected to header and is stored in
838 HEADER_EDGE, the ENTRY edge is redirected into copy of header and
839 returned as COPY_HEADER_EDGE. The effect is following:
840 if LATCH_EDGE == ENTRY, then the loop is unrolled by one copy,
841 HEADER_EDGE is latch of a new loop, COPY_HEADER_EDGE leads from original
842 latch source to first block in copy.
843 if LATCH_EDGE != ENTRY, then the loop is peeled by one copy,
844 HEADER_EDGE is entry edge of the loop, COPY_HEADER_EDGE leads from
845 original entry block to first block in peeled copy.
848 copy_bbs (basic_block *bbs, int n, edge entry, edge latch_edge,
849 basic_block **new_bbs, struct loops *loops, edge *header_edge,
850 edge *copy_header_edge, int add_irreducible_flag)
853 basic_block bb, new_bb, header = entry->dest, dom_bb;
856 /* Duplicate bbs, update dominators, assign bbs to loops. */
857 (*new_bbs) = xcalloc (n, sizeof (basic_block));
858 for (i = 0; i < n; i++)
862 new_bb = (*new_bbs)[i] = cfg_layout_duplicate_bb (bb, NULL);
863 RBI (new_bb)->duplicated = 1;
865 add_bb_to_loop (new_bb, bb->loop_father->copy);
866 add_to_dominance_info (loops->cfg.dom, new_bb);
867 /* Possibly set header. */
868 if (bb->loop_father->header == bb && bb != header)
869 new_bb->loop_father->header = new_bb;
871 if (bb->loop_father->latch == bb &&
872 bb->loop_father != header->loop_father)
873 new_bb->loop_father->latch = new_bb;
874 /* Take care of irreducible loops. */
875 if (add_irreducible_flag
876 && bb->loop_father == header->loop_father)
877 new_bb->flags |= BB_IRREDUCIBLE_LOOP;
880 /* Set dominators. */
881 for (i = 0; i < n; i++)
884 new_bb = (*new_bbs)[i];
887 /* For anything else than loop header, just copy it. */
888 dom_bb = get_immediate_dominator (loops->cfg.dom, bb);
889 dom_bb = RBI (dom_bb)->copy;
893 /* Copy of header is dominated by entry source. */
898 set_immediate_dominator (loops->cfg.dom, new_bb, dom_bb);
901 /* Redirect edges. */
902 for (i = 0; i < n; i++)
905 new_bb = (*new_bbs)[i];
907 for (e = bb->pred; e; e = e_pred)
909 basic_block src = e->src;
911 e_pred = e->pred_next;
913 if (!RBI (src)->duplicated)
916 /* Leads to copied loop and it is not latch edge, redirect it. */
918 loop_redirect_edge (e, new_bb);
920 if (add_irreducible_flag
921 && (bb->loop_father == header->loop_father
922 || RBI (src)->original->loop_father == header->loop_father))
923 e->flags |= EDGE_IRREDUCIBLE_LOOP;
927 /* Redirect header edge. */
928 bb = RBI (latch_edge->src)->copy;
929 for (e = bb->succ; e->dest != latch_edge->dest; e = e->succ_next);
931 loop_redirect_edge (*header_edge, header);
933 /* Redirect entry to copy of header. */
934 loop_redirect_edge (entry, RBI (header)->copy);
935 *copy_header_edge = entry;
937 /* Clear information about duplicates. */
938 for (i = 0; i < n; i++)
939 RBI ((*new_bbs)[i])->duplicated = 0;
942 /* Check whether LOOP's body can be duplicated. */
944 can_duplicate_loop_p (struct loop *loop)
949 bbs = get_loop_body (loop);
951 for (i = 0; i < loop->num_nodes; i++)
955 /* In case loop contains abnormal edge we can not redirect,
956 we can't perform duplication. */
958 for (e = bbs[i]->succ; e; e = e->succ_next)
959 if ((e->flags & EDGE_ABNORMAL)
960 && flow_bb_inside_loop_p (loop, e->dest))
966 if (!cfg_layout_can_duplicate_bb_p (bbs[i]))
977 /* Record edges, leading from NBBS basic blocks stored in BBS, that were created
978 by copying ORIG edge (or just ORIG edge if IS_ORIG is set).
979 If ORIG is NULL, then record all edges coming outside of BBS. Store them
980 into TO_REMOVE array that must be large enough to hold them all; their
981 number is returned in N_TO_REMOVE. */
983 record_exit_edges (edge orig, basic_block *bbs, int nbbs, edge *to_remove,
984 unsigned int *n_to_remove, int is_orig)
994 to_remove[(*n_to_remove)++] = orig;
998 for (e = RBI (orig->src)->copy->succ; e; e = e->succ_next)
999 if (e->dest == orig->dest)
1004 to_remove[(*n_to_remove)++] = e;
1008 my_blocks = sbitmap_alloc (last_basic_block);
1009 sbitmap_zero (my_blocks);
1010 for (i = 0; i < nbbs; i++)
1011 SET_BIT (my_blocks, bbs[i]->index);
1013 for (i = 0; i < nbbs; i++)
1014 for (e = bbs[i]->succ; e; e = e->succ_next)
1015 if (e->dest == EXIT_BLOCK_PTR ||
1016 !TEST_BIT (my_blocks, e->dest->index))
1017 to_remove[(*n_to_remove)++] = e;
1024 #define RDIV(X,Y) (((X) + (Y) / 2) / (Y))
1026 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of
1027 updating LOOPS structure and dominators. E's destination must be LOOP
1028 header for this to work, i.e. it must be entry or latch edge of this loop;
1029 these are unique, as the loops must have preheaders for this function to
1030 work correctly (in case E is latch, the function unrolls the loop, if E is
1031 entry edge, it peels the loop). Store edges created by copying ORIG edge
1032 (if NULL, then all edges leaving loop) from copies corresponding to set
1033 bits in WONT_EXIT bitmap (bit 0 corresponds to original LOOP body, the
1034 other copies are numbered in order given by control flow through them)
1035 into TO_REMOVE array. Returns false if duplication is impossible. */
1037 duplicate_loop_to_header_edge (struct loop *loop, edge e, struct loops *loops,
1038 unsigned int ndupl, sbitmap wont_exit,
1039 edge orig, edge *to_remove,
1040 unsigned int *n_to_remove, int flags)
1042 struct loop *target, *aloop;
1043 struct loop **orig_loops;
1044 unsigned n_orig_loops;
1045 basic_block header = loop->header, latch = loop->latch;
1046 basic_block *new_bbs, *bbs, *first_active;
1047 basic_block new_bb, bb, first_active_latch = NULL;
1048 edge ae, latch_edge, he;
1050 int is_latch = (latch == e->src);
1051 int scale_act = 0, *scale_step = NULL, scale_main = 0;
1052 int p, freq_in, freq_le, freq_out_orig;
1053 int prob_pass_thru, prob_pass_wont_exit, prob_pass_main;
1054 int add_irreducible_flag;
1056 if (e->dest != loop->header)
1063 /* Orig must be edge out of the loop. */
1064 if (!flow_bb_inside_loop_p (loop, orig->src))
1066 if (flow_bb_inside_loop_p (loop, orig->dest))
1070 bbs = get_loop_body (loop);
1072 /* Check whether duplication is possible. */
1074 for (i = 0; i < loop->num_nodes; i++)
1076 if (!cfg_layout_can_duplicate_bb_p (bbs[i]))
1083 add_irreducible_flag = !is_latch && (e->flags & EDGE_IRREDUCIBLE_LOOP);
1085 /* Find edge from latch. */
1086 latch_edge = loop_latch_edge (loop);
1088 if (flags & DLTHE_FLAG_UPDATE_FREQ)
1090 /* Calculate coefficients by that we have to scale frequencies
1091 of duplicated loop bodies. */
1092 freq_in = header->frequency;
1093 freq_le = EDGE_FREQUENCY (latch_edge);
1096 if (freq_in < freq_le)
1098 freq_out_orig = orig ? EDGE_FREQUENCY (orig) : freq_in - freq_le;
1099 if (freq_out_orig > freq_in - freq_le)
1100 freq_out_orig = freq_in - freq_le;
1101 prob_pass_thru = RDIV (REG_BR_PROB_BASE * freq_le, freq_in);
1102 prob_pass_wont_exit =
1103 RDIV (REG_BR_PROB_BASE * (freq_le + freq_out_orig), freq_in);
1105 scale_step = xmalloc (ndupl * sizeof (int));
1107 for (i = 1; i <= ndupl; i++)
1108 scale_step[i - 1] = TEST_BIT (wont_exit, i)
1109 ? prob_pass_wont_exit
1114 prob_pass_main = TEST_BIT (wont_exit, 0)
1115 ? prob_pass_wont_exit
1118 scale_main = REG_BR_PROB_BASE;
1119 for (i = 0; i < ndupl; i++)
1122 p = RDIV (p * scale_step[i], REG_BR_PROB_BASE);
1124 scale_main = RDIV (REG_BR_PROB_BASE * REG_BR_PROB_BASE, scale_main);
1125 scale_act = RDIV (scale_main * prob_pass_main, REG_BR_PROB_BASE);
1129 scale_main = REG_BR_PROB_BASE;
1130 for (i = 0; i < ndupl; i++)
1131 scale_main = RDIV (scale_main * scale_step[i], REG_BR_PROB_BASE);
1132 scale_act = REG_BR_PROB_BASE - prob_pass_thru;
1134 for (i = 0; i < ndupl; i++)
1135 if (scale_step[i] < 0 || scale_step[i] > REG_BR_PROB_BASE)
1137 if (scale_main < 0 || scale_main > REG_BR_PROB_BASE
1138 || scale_act < 0 || scale_act > REG_BR_PROB_BASE)
1142 /* Loop the new bbs will belong to. */
1143 target = find_common_loop (e->src->loop_father, e->dest->loop_father);
1145 /* Original loops. */
1147 for (aloop = loop->inner; aloop; aloop = aloop->next)
1149 orig_loops = xcalloc (n_orig_loops, sizeof (struct loop *));
1150 for (aloop = loop->inner, i = 0; aloop; aloop = aloop->next, i++)
1151 orig_loops[i] = aloop;
1153 loop->copy = target;
1155 /* Original basic blocks. */
1156 n = loop->num_nodes;
1158 first_active = xcalloc(n, sizeof (basic_block));
1161 memcpy (first_active, bbs, n * sizeof (basic_block));
1162 first_active_latch = latch;
1165 /* Record exit edges in original loop body. */
1166 if (TEST_BIT (wont_exit, 0))
1167 record_exit_edges (orig, bbs, n, to_remove, n_to_remove, true);
1169 for (j = 0; j < ndupl; j++)
1172 copy_loops_to (loops, orig_loops, n_orig_loops, target);
1175 copy_bbs (bbs, n, e, latch_edge, &new_bbs, loops,
1176 &e, &he, add_irreducible_flag);
1178 loop->latch = RBI (latch)->copy;
1180 /* Record exit edges in this copy. */
1181 if (TEST_BIT (wont_exit, j + 1))
1182 record_exit_edges (orig, new_bbs, n, to_remove, n_to_remove, false);
1184 /* Set counts and frequencies. */
1185 for (i = 0; i < n; i++)
1187 new_bb = new_bbs[i];
1190 if (flags & DLTHE_FLAG_UPDATE_FREQ)
1192 new_bb->count = RDIV (scale_act * bb->count, REG_BR_PROB_BASE);
1193 new_bb->frequency = RDIV (scale_act * bb->frequency,
1198 new_bb->count = bb->count;
1199 new_bb->frequency = bb->frequency;
1202 for (ae = new_bb->succ; ae; ae = ae->succ_next)
1203 ae->count = RDIV (new_bb->count * ae->probability,
1206 if (flags & DLTHE_FLAG_UPDATE_FREQ)
1207 scale_act = RDIV (scale_act * scale_step[j], REG_BR_PROB_BASE);
1209 if (!first_active_latch)
1211 memcpy (first_active, new_bbs, n * sizeof (basic_block));
1212 first_active_latch = RBI (latch)->copy;
1217 /* Original loop header is dominated by latch copy
1218 if we duplicated on its only entry edge. */
1219 if (!is_latch && !header->pred->pred_next->pred_next)
1220 set_immediate_dominator (loops->cfg.dom, header, RBI (latch)->copy);
1221 if (is_latch && j == 0)
1223 /* Update edge from latch. */
1224 for (latch_edge = RBI (header)->copy->pred;
1225 latch_edge->src != latch;
1226 latch_edge = latch_edge->pred_next);
1229 /* Now handle original loop. */
1231 /* Update edge counts. */
1232 if (flags & DLTHE_FLAG_UPDATE_FREQ)
1234 for (i = 0; i < n; i++)
1237 bb->count = RDIV (scale_main * bb->count, REG_BR_PROB_BASE);
1238 bb->frequency = RDIV (scale_main * bb->frequency, REG_BR_PROB_BASE);
1239 for (ae = bb->succ; ae; ae = ae->succ_next)
1240 ae->count = RDIV (bb->count * ae->probability, REG_BR_PROB_BASE);
1246 /* Update dominators of other blocks if affected. */
1247 for (i = 0; i < n; i++)
1249 basic_block dominated, dom_bb, *dom_bbs;
1253 n_dom_bbs = get_dominated_by (loops->cfg.dom, bb, &dom_bbs);
1254 for (j = 0; j < n_dom_bbs; j++)
1256 dominated = dom_bbs[j];
1257 if (flow_bb_inside_loop_p (loop, dominated))
1259 dom_bb = nearest_common_dominator (
1260 loops->cfg.dom, first_active[i], first_active_latch);
1261 set_immediate_dominator (loops->cfg.dom, dominated, dom_bb);
1265 free (first_active);
1272 /* Creates a pre-header for a LOOP. Returns newly created block. Unless
1273 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single
1274 entry; otherwise we also force preheader block to have only one successor.
1275 The function also updates dominators stored in DOM. */
1277 create_preheader (struct loop *loop, dominance_info dom, int flags)
1281 basic_block jump, src = 0;
1282 struct loop *cloop, *ploop;
1286 cloop = loop->outer;
1288 for (e = loop->header->pred; e; e = e->pred_next)
1290 if (e->src == loop->latch)
1298 for (e = loop->header->pred; e->src == loop->latch; e = e->pred_next);
1299 if (!(flags & CP_SIMPLE_PREHEADERS)
1300 || !e->src->succ->succ_next)
1304 insn = first_insn_after_basic_block_note (loop->header);
1306 insn = PREV_INSN (insn);
1308 insn = get_last_insn ();
1309 if (insn == loop->header->end)
1311 /* Split_block would not split block after its end. */
1312 emit_note_after (NOTE_INSN_DELETED, insn);
1314 fallthru = split_block (loop->header, insn);
1315 dummy = fallthru->src;
1316 loop->header = fallthru->dest;
1318 /* The header could be a latch of some superloop(s); due to design of
1319 split_block, it would now move to fallthru->dest. */
1320 for (ploop = loop; ploop; ploop = ploop->outer)
1321 if (ploop->latch == dummy)
1322 ploop->latch = fallthru->dest;
1324 add_to_dominance_info (dom, fallthru->dest);
1326 /* Redirect edges. */
1327 for (e = dummy->pred; e; e = e->pred_next)
1330 if (src == loop->latch)
1336 dummy->frequency -= EDGE_FREQUENCY (e);
1337 dummy->count -= e->count;
1338 fallthru->count -= e->count;
1339 jump = redirect_edge_and_branch_force (e, loop->header);
1342 add_to_dominance_info (dom, jump);
1343 set_immediate_dominator (dom, jump, src);
1344 add_bb_to_loop (jump, loop);
1348 /* Update structures. */
1349 redirect_immediate_dominators (dom, dummy, loop->header);
1350 set_immediate_dominator (dom, loop->header, dummy);
1351 loop->header->loop_father = loop;
1352 add_bb_to_loop (dummy, cloop);
1354 fprintf (rtl_dump_file, "Created preheader block for loop %i\n",
1360 /* Create preheaders for each loop from loop tree stored in LOOPS; for meaning
1361 of FLAGS see create_preheader. */
1363 create_preheaders (struct loops *loops, int flags)
1366 for (i = 1; i < loops->num; i++)
1367 create_preheader (loops->parray[i], loops->cfg.dom, flags);
1368 loops->state |= LOOPS_HAVE_PREHEADERS;
1371 /* Forces all loop latches of loops from loop tree LOOPS to have only single
1374 force_single_succ_latches (struct loops *loops)
1380 for (i = 1; i < loops->num; i++)
1382 loop = loops->parray[i];
1383 if (!loop->latch->succ->succ_next)
1386 for (e = loop->header->pred; e->src != loop->latch; e = e->pred_next)
1389 loop_split_edge_with (e, NULL_RTX, loops);
1391 loops->state |= LOOPS_HAVE_SIMPLE_LATCHES;
1394 /* A quite stupid function to put INSNS on edge E. They are supposed to form
1395 just one basic block. Jumps in INSNS are not handled, so cfg do not have to
1396 be ok after this function. The created block is placed on correct place
1397 in LOOPS structure and its dominator is set. */
1399 loop_split_edge_with (edge e, rtx insns, struct loops *loops)
1401 basic_block src, dest, new_bb;
1402 struct loop *loop_c;
1408 loop_c = find_common_loop (src->loop_father, dest->loop_father);
1410 /* Create basic block for it. */
1412 new_bb = create_basic_block (NULL_RTX, NULL_RTX, EXIT_BLOCK_PTR->prev_bb);
1413 add_to_dominance_info (loops->cfg.dom, new_bb);
1414 add_bb_to_loop (new_bb, loop_c);
1415 new_bb->flags = insns ? BB_SUPERBLOCK : 0;
1417 new_e = make_edge (new_bb, dest, EDGE_FALLTHRU);
1418 new_e->probability = REG_BR_PROB_BASE;
1419 new_e->count = e->count;
1420 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
1422 new_bb->flags |= BB_IRREDUCIBLE_LOOP;
1423 new_e->flags |= EDGE_IRREDUCIBLE_LOOP;
1426 new_bb->count = e->count;
1427 new_bb->frequency = EDGE_FREQUENCY (e);
1428 redirect_edge_and_branch_force (e, new_bb);
1430 alloc_aux_for_block (new_bb, sizeof (struct reorder_block_def));
1435 insns = get_insns ();
1437 emit_insn_after (insns, new_bb->end);
1440 set_immediate_dominator (loops->cfg.dom, new_bb, src);
1441 set_immediate_dominator (loops->cfg.dom, dest,
1442 recount_dominator (loops->cfg.dom, dest));
1444 if (dest->loop_father->latch == src)
1445 dest->loop_father->latch = new_bb;