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 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 remove_bbs (basic_block
*, int);
40 static bool rpe_enum_p (basic_block
, void *);
41 static int find_path (edge
, basic_block
**);
42 static bool alp_enum_p (basic_block
, void *);
43 static void add_loop (struct loops
*, struct loop
*);
44 static void fix_loop_placements (struct loop
*);
45 static bool fix_bb_placement (struct loops
*, basic_block
);
46 static void fix_bb_placements (struct loops
*, basic_block
);
47 static void place_new_loop (struct loops
*, struct loop
*);
48 static void scale_loop_frequencies (struct loop
*, int, int);
49 static void scale_bbs_frequencies (basic_block
*, int, int, int);
50 static basic_block
create_preheader (struct loop
*, int);
51 static void fix_irreducible_loops (basic_block
);
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 if (e
->dest
->pred
->pred_next
)
102 /* Find bbs in the path. */
103 *bbs
= xcalloc (n_basic_blocks
, sizeof (basic_block
));
104 return dfs_enumerate_from (e
->dest
, 0, rpe_enum_p
, *bbs
,
105 n_basic_blocks
, e
->dest
);
108 /* Fix placement of basic block BB inside loop hierarchy stored in LOOPS --
109 Let L be a loop to that BB belongs. Then every successor of BB must either
110 1) belong to some superloop of loop L, or
111 2) be a header of loop K such that K->outer is superloop of L
112 Returns true if we had to move BB into other loop to enforce this condition,
113 false if the placement of BB was already correct (provided that placements
114 of its successors are correct). */
116 fix_bb_placement (struct loops
*loops
, basic_block bb
)
119 struct loop
*loop
= loops
->tree_root
, *act
;
121 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
123 if (e
->dest
== EXIT_BLOCK_PTR
)
126 act
= e
->dest
->loop_father
;
127 if (act
->header
== e
->dest
)
130 if (flow_loop_nested_p (loop
, act
))
134 if (loop
== bb
->loop_father
)
137 remove_bb_from_loops (bb
);
138 add_bb_to_loop (bb
, loop
);
143 /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e.
144 enforce condition condition stated in description of fix_bb_placement. We
145 start from basic block FROM that had some of its successors removed, so that
146 his placement no longer has to be correct, and iteratively fix placement of
147 its predecessors that may change if placement of FROM changed. Also fix
148 placement of subloops of FROM->loop_father, that might also be altered due
149 to this change; the condition for them is similar, except that instead of
150 successors we consider edges coming out of the loops. */
152 fix_bb_placements (struct loops
*loops
, basic_block from
)
155 basic_block
*queue
, *qtop
, *qbeg
, *qend
;
156 struct loop
*base_loop
;
159 /* We pass through blocks back-reachable from FROM, testing whether some
160 of their successors moved to outer loop. It may be necessary to
161 iterate several times, but it is finite, as we stop unless we move
162 the basic block up the loop structure. The whole story is a bit
163 more complicated due to presence of subloops, those are moved using
164 fix_loop_placement. */
166 base_loop
= from
->loop_father
;
167 if (base_loop
== loops
->tree_root
)
170 in_queue
= sbitmap_alloc (last_basic_block
);
171 sbitmap_zero (in_queue
);
172 SET_BIT (in_queue
, from
->index
);
173 /* Prevent us from going out of the base_loop. */
174 SET_BIT (in_queue
, base_loop
->header
->index
);
176 queue
= xmalloc ((base_loop
->num_nodes
+ 1) * sizeof (basic_block
));
177 qtop
= queue
+ base_loop
->num_nodes
+ 1;
188 RESET_BIT (in_queue
, from
->index
);
190 if (from
->loop_father
->header
== from
)
192 /* Subloop header, maybe move the loop upward. */
193 if (!fix_loop_placement (from
->loop_father
))
198 /* Ordinary basic block. */
199 if (!fix_bb_placement (loops
, from
))
203 /* Something has changed, insert predecessors into queue. */
204 for (e
= from
->pred
; e
; e
= e
->pred_next
)
206 basic_block pred
= e
->src
;
209 if (TEST_BIT (in_queue
, pred
->index
))
212 /* If it is subloop, then it either was not moved, or
213 the path up the loop tree from base_loop do not contain
215 nca
= find_common_loop (pred
->loop_father
, base_loop
);
216 if (pred
->loop_father
!= base_loop
218 || nca
!= pred
->loop_father
))
219 pred
= pred
->loop_father
->header
;
220 else if (!flow_loop_nested_p (from
->loop_father
, pred
->loop_father
))
222 /* No point in processing it. */
226 if (TEST_BIT (in_queue
, pred
->index
))
229 /* Schedule the basic block. */
234 SET_BIT (in_queue
, pred
->index
);
241 /* Basic block from has lost one or more of its predecessors, so it might
242 mo longer be part irreducible loop. Fix it and proceed recursively
243 for its successors if needed. */
245 fix_irreducible_loops (basic_block from
)
254 if (!(from
->flags
& BB_IRREDUCIBLE_LOOP
))
257 on_stack
= sbitmap_alloc (last_basic_block
);
258 sbitmap_zero (on_stack
);
259 SET_BIT (on_stack
, from
->index
);
260 stack
= xmalloc (from
->loop_father
->num_nodes
* sizeof (basic_block
));
266 bb
= stack
[--stack_top
];
267 RESET_BIT (on_stack
, bb
->index
);
269 for (e
= bb
->pred
; e
; e
= e
->pred_next
)
270 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
275 bb
->flags
&= ~BB_IRREDUCIBLE_LOOP
;
276 if (bb
->loop_father
->header
== bb
)
277 edges
= get_loop_exit_edges (bb
->loop_father
, &n_edges
);
281 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
283 edges
= xmalloc (n_edges
* sizeof (edge
));
285 for (e
= bb
->succ
; e
; e
= e
->succ_next
)
286 edges
[n_edges
++] = e
;
289 for (i
= 0; i
< n_edges
; i
++)
293 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
295 if (!flow_bb_inside_loop_p (from
->loop_father
, e
->dest
))
298 e
->flags
&= ~EDGE_IRREDUCIBLE_LOOP
;
299 if (TEST_BIT (on_stack
, e
->dest
->index
))
302 SET_BIT (on_stack
, e
->dest
->index
);
303 stack
[stack_top
++] = e
->dest
;
313 /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E
314 and update loop structure stored in LOOPS and dominators. Return true if
315 we were able to remove the path, false otherwise (and nothing is affected
318 remove_path (struct loops
*loops
, edge e
)
321 basic_block
*rem_bbs
, *bord_bbs
, *dom_bbs
, from
, bb
;
322 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 if (!loop_delete_branch_edge (e
, 1))
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 (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
= (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 loop
*loop
)
676 if (!fix_loop_placement (loop
))
682 /* Creates place for a new LOOP in LOOPS structure. */
684 place_new_loop (struct loops
*loops
, struct loop
*loop
)
687 xrealloc (loops
->parray
, (loops
->num
+ 1) * sizeof (struct loop
*));
688 loops
->parray
[loops
->num
] = loop
;
690 loop
->num
= loops
->num
++;
693 /* Copies copy of LOOP as subloop of TARGET loop, placing newly
694 created loop into LOOPS structure. */
696 duplicate_loop (struct loops
*loops
, struct loop
*loop
, struct loop
*target
)
699 cloop
= xcalloc (1, sizeof (struct loop
));
700 place_new_loop (loops
, cloop
);
702 /* Initialize copied loop. */
703 cloop
->level
= loop
->level
;
705 /* Set it as copy of loop. */
708 /* Add it to target. */
709 flow_loop_tree_node_add (target
, cloop
);
714 /* Copies structure of subloops of LOOP into TARGET loop, placing
715 newly created loops into loop tree stored in LOOPS. */
717 duplicate_subloops (struct loops
*loops
, struct loop
*loop
, struct loop
*target
)
719 struct loop
*aloop
, *cloop
;
721 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
723 cloop
= duplicate_loop (loops
, aloop
, target
);
724 duplicate_subloops (loops
, aloop
, cloop
);
728 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS,
729 into TARGET loop, placing newly created loops into loop tree LOOPS. */
731 copy_loops_to (struct loops
*loops
, struct loop
**copied_loops
, int n
, struct loop
*target
)
736 for (i
= 0; i
< n
; i
++)
738 aloop
= duplicate_loop (loops
, copied_loops
[i
], target
);
739 duplicate_subloops (loops
, copied_loops
[i
], aloop
);
743 /* Redirects edge E to basic block DEST. */
745 loop_redirect_edge (edge e
, basic_block dest
)
750 redirect_edge_and_branch_force (e
, dest
);
753 /* Deletes edge E from a branch if possible. Unless REALLY_DELETE is set,
754 just test whether it is possible to remove the edge. */
756 loop_delete_branch_edge (edge e
, int really_delete
)
758 basic_block src
= e
->src
;
762 if (src
->succ
->succ_next
)
766 /* Cannot handle more than two exit edges. */
767 if (src
->succ
->succ_next
->succ_next
)
769 /* And it must be just a simple branch. */
770 if (!any_condjump_p (BB_END (src
)))
773 snd
= e
== src
->succ
? src
->succ
->succ_next
: src
->succ
;
775 if (newdest
== EXIT_BLOCK_PTR
)
778 /* Hopefully the above conditions should suffice. */
782 /* Redirecting behaves wrongly wrto this flag. */
783 irr
= snd
->flags
& EDGE_IRREDUCIBLE_LOOP
;
785 if (!redirect_edge_and_branch (e
, newdest
))
787 src
->succ
->flags
&= ~EDGE_IRREDUCIBLE_LOOP
;
788 src
->succ
->flags
|= irr
;
794 /* Cannot happen -- we are using this only to remove an edge
799 return false; /* To avoid warning, cannot get here. */
802 /* Check whether LOOP's body can be duplicated. */
804 can_duplicate_loop_p (struct loop
*loop
)
807 basic_block
*bbs
= get_loop_body (loop
);
809 ret
= can_copy_bbs_p (bbs
, loop
->num_nodes
);
815 #define RDIV(X,Y) (((X) + (Y) / 2) / (Y))
817 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating
818 LOOPS structure and dominators. E's destination must be LOOP header for
819 this to work, i.e. it must be entry or latch edge of this loop; these are
820 unique, as the loops must have preheaders for this function to work
821 correctly (in case E is latch, the function unrolls the loop, if E is entry
822 edge, it peels the loop). Store edges created by copying ORIG edge from
823 copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to
824 original LOOP body, the other copies are numbered in order given by control
825 flow through them) into TO_REMOVE array. Returns false if duplication is
828 duplicate_loop_to_header_edge (struct loop
*loop
, edge e
, struct loops
*loops
,
829 unsigned int ndupl
, sbitmap wont_exit
,
830 edge orig
, edge
*to_remove
,
831 unsigned int *n_to_remove
, int flags
)
833 struct loop
*target
, *aloop
;
834 struct loop
**orig_loops
;
835 unsigned n_orig_loops
;
836 basic_block header
= loop
->header
, latch
= loop
->latch
;
837 basic_block
*new_bbs
, *bbs
, *first_active
;
838 basic_block new_bb
, bb
, first_active_latch
= NULL
;
840 edge spec_edges
[2], new_spec_edges
[2];
844 int is_latch
= (latch
== e
->src
);
845 int scale_act
= 0, *scale_step
= NULL
, scale_main
= 0;
846 int p
, freq_in
, freq_le
, freq_out_orig
;
847 int prob_pass_thru
, prob_pass_wont_exit
, prob_pass_main
;
848 int add_irreducible_flag
;
850 if (e
->dest
!= loop
->header
)
857 /* Orig must be edge out of the loop. */
858 if (!flow_bb_inside_loop_p (loop
, orig
->src
))
860 if (flow_bb_inside_loop_p (loop
, orig
->dest
))
864 bbs
= get_loop_body (loop
);
866 /* Check whether duplication is possible. */
867 if (!can_copy_bbs_p (bbs
, loop
->num_nodes
))
872 new_bbs
= xmalloc (sizeof (basic_block
) * loop
->num_nodes
);
874 /* In case we are doing loop peeling and the loop is in the middle of
875 irreducible region, the peeled copies will be inside it too. */
876 add_irreducible_flag
= e
->flags
& EDGE_IRREDUCIBLE_LOOP
;
877 if (is_latch
&& add_irreducible_flag
)
880 /* Find edge from latch. */
881 latch_edge
= loop_latch_edge (loop
);
883 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
885 /* Calculate coefficients by that we have to scale frequencies
886 of duplicated loop bodies. */
887 freq_in
= header
->frequency
;
888 freq_le
= EDGE_FREQUENCY (latch_edge
);
891 if (freq_in
< freq_le
)
893 freq_out_orig
= orig
? EDGE_FREQUENCY (orig
) : freq_in
- freq_le
;
894 if (freq_out_orig
> freq_in
- freq_le
)
895 freq_out_orig
= freq_in
- freq_le
;
896 prob_pass_thru
= RDIV (REG_BR_PROB_BASE
* freq_le
, freq_in
);
897 prob_pass_wont_exit
=
898 RDIV (REG_BR_PROB_BASE
* (freq_le
+ freq_out_orig
), freq_in
);
900 scale_step
= xmalloc (ndupl
* sizeof (int));
902 for (i
= 1; i
<= ndupl
; i
++)
903 scale_step
[i
- 1] = TEST_BIT (wont_exit
, i
)
904 ? prob_pass_wont_exit
909 prob_pass_main
= TEST_BIT (wont_exit
, 0)
910 ? prob_pass_wont_exit
913 scale_main
= REG_BR_PROB_BASE
;
914 for (i
= 0; i
< ndupl
; i
++)
917 p
= RDIV (p
* scale_step
[i
], REG_BR_PROB_BASE
);
919 scale_main
= RDIV (REG_BR_PROB_BASE
* REG_BR_PROB_BASE
, scale_main
);
920 scale_act
= RDIV (scale_main
* prob_pass_main
, REG_BR_PROB_BASE
);
924 scale_main
= REG_BR_PROB_BASE
;
925 for (i
= 0; i
< ndupl
; i
++)
926 scale_main
= RDIV (scale_main
* scale_step
[i
], REG_BR_PROB_BASE
);
927 scale_act
= REG_BR_PROB_BASE
- prob_pass_thru
;
929 for (i
= 0; i
< ndupl
; i
++)
930 if (scale_step
[i
] < 0 || scale_step
[i
] > REG_BR_PROB_BASE
)
932 if (scale_main
< 0 || scale_main
> REG_BR_PROB_BASE
933 || scale_act
< 0 || scale_act
> REG_BR_PROB_BASE
)
937 /* Loop the new bbs will belong to. */
938 target
= e
->src
->loop_father
;
940 /* Original loops. */
942 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
944 orig_loops
= xcalloc (n_orig_loops
, sizeof (struct loop
*));
945 for (aloop
= loop
->inner
, i
= 0; aloop
; aloop
= aloop
->next
, i
++)
946 orig_loops
[i
] = aloop
;
952 first_active
= xmalloc (n
* sizeof (basic_block
));
955 memcpy (first_active
, bbs
, n
* sizeof (basic_block
));
956 first_active_latch
= latch
;
959 /* Record exit edge in original loop body. */
960 if (orig
&& TEST_BIT (wont_exit
, 0))
961 to_remove
[(*n_to_remove
)++] = orig
;
963 spec_edges
[SE_ORIG
] = orig
;
964 spec_edges
[SE_LATCH
] = latch_edge
;
966 for (j
= 0; j
< ndupl
; j
++)
969 copy_loops_to (loops
, orig_loops
, n_orig_loops
, target
);
972 copy_bbs (bbs
, n
, new_bbs
, spec_edges
, 2, new_spec_edges
, loop
);
974 /* Note whether the blocks and edges belong to an irreducible loop. */
975 if (add_irreducible_flag
)
977 for (i
= 0; i
< n
; i
++)
978 new_bbs
[i
]->rbi
->duplicated
= 1;
979 for (i
= 0; i
< n
; i
++)
982 if (new_bb
->loop_father
== target
)
983 new_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
985 for (ae
= new_bb
->succ
; ae
; ae
= ae
->succ_next
)
986 if (ae
->dest
->rbi
->duplicated
987 && (ae
->src
->loop_father
== target
988 || ae
->dest
->loop_father
== target
))
989 ae
->flags
|= EDGE_IRREDUCIBLE_LOOP
;
991 for (i
= 0; i
< n
; i
++)
992 new_bbs
[i
]->rbi
->duplicated
= 0;
995 /* Redirect the special edges. */
998 redirect_edge_and_branch_force (latch_edge
, new_bbs
[0]);
999 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
1001 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], latch
);
1002 latch
= loop
->latch
= new_bbs
[1];
1003 e
= latch_edge
= new_spec_edges
[SE_LATCH
];
1007 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
1009 redirect_edge_and_branch_force (e
, new_bbs
[0]);
1010 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], e
->src
);
1011 e
= new_spec_edges
[SE_LATCH
];
1014 /* Record exit edge in this copy. */
1015 if (orig
&& TEST_BIT (wont_exit
, j
+ 1))
1016 to_remove
[(*n_to_remove
)++] = new_spec_edges
[SE_ORIG
];
1018 /* Record the first copy in the control flow order if it is not
1019 the original loop (i.e. in case of peeling). */
1020 if (!first_active_latch
)
1022 memcpy (first_active
, new_bbs
, n
* sizeof (basic_block
));
1023 first_active_latch
= new_bbs
[1];
1026 /* Set counts and frequencies. */
1027 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1029 scale_bbs_frequencies (new_bbs
, n
, scale_act
, REG_BR_PROB_BASE
);
1030 scale_act
= RDIV (scale_act
* scale_step
[j
], REG_BR_PROB_BASE
);
1036 /* Update the original loop. */
1038 set_immediate_dominator (CDI_DOMINATORS
, e
->dest
, e
->src
);
1039 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1041 scale_bbs_frequencies (bbs
, n
, scale_main
, REG_BR_PROB_BASE
);
1045 /* Update dominators of outer blocks if affected. */
1046 for (i
= 0; i
< n
; i
++)
1048 basic_block dominated
, dom_bb
, *dom_bbs
;
1052 n_dom_bbs
= get_dominated_by (CDI_DOMINATORS
, bb
, &dom_bbs
);
1053 for (j
= 0; j
< n_dom_bbs
; j
++)
1055 dominated
= dom_bbs
[j
];
1056 if (flow_bb_inside_loop_p (loop
, dominated
))
1058 dom_bb
= nearest_common_dominator (
1059 CDI_DOMINATORS
, first_active
[i
], first_active_latch
);
1060 set_immediate_dominator (CDI_DOMINATORS
, dominated
, dom_bb
);
1064 free (first_active
);
1071 /* A callback for make_forwarder block, to redirect all edges except for
1072 MFB_KJ_EDGE to the entry part. E is the edge for that we should decide
1073 whether to redirect it. */
1075 static edge mfb_kj_edge
;
1077 mfb_keep_just (edge e
)
1079 return e
!= mfb_kj_edge
;
1082 /* A callback for make_forwarder block, to update data structures for a basic
1083 block JUMP created by redirecting an edge (only the latch edge is being
1087 mfb_update_loops (basic_block jump
)
1089 struct loop
*loop
= jump
->succ
->dest
->loop_father
;
1091 if (dom_computed
[CDI_DOMINATORS
])
1092 set_immediate_dominator (CDI_DOMINATORS
, jump
, jump
->pred
->src
);
1093 add_bb_to_loop (jump
, loop
);
1097 /* Creates a pre-header for a LOOP. Returns newly created block. Unless
1098 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single
1099 entry; otherwise we also force preheader block to have only one successor.
1100 The function also updates dominators. */
1103 create_preheader (struct loop
*loop
, int flags
)
1107 struct loop
*cloop
, *ploop
;
1111 cloop
= loop
->outer
;
1113 for (e
= loop
->header
->pred
; e
; e
= e
->pred_next
)
1115 if (e
->src
== loop
->latch
)
1117 irred
|= (e
->flags
& EDGE_IRREDUCIBLE_LOOP
) != 0;
1124 for (e
= loop
->header
->pred
; e
->src
== loop
->latch
; e
= e
->pred_next
);
1125 if (!(flags
& CP_SIMPLE_PREHEADERS
)
1126 || !e
->src
->succ
->succ_next
)
1130 mfb_kj_edge
= loop_latch_edge (loop
);
1131 fallthru
= make_forwarder_block (loop
->header
, mfb_keep_just
,
1133 dummy
= fallthru
->src
;
1134 loop
->header
= fallthru
->dest
;
1136 /* The header could be a latch of some superloop(s); due to design of
1137 split_block, it would now move to fallthru->dest. */
1138 for (ploop
= loop
; ploop
; ploop
= ploop
->outer
)
1139 if (ploop
->latch
== dummy
)
1140 ploop
->latch
= fallthru
->dest
;
1142 /* Reorganize blocks so that the preheader is not stuck in the middle of the
1144 for (e
= dummy
->pred
; e
; e
= e
->pred_next
)
1145 if (e
->src
!= loop
->latch
)
1147 move_block_after (dummy
, e
->src
);
1149 loop
->header
->loop_father
= loop
;
1150 add_bb_to_loop (dummy
, cloop
);
1154 dummy
->flags
|= BB_IRREDUCIBLE_LOOP
;
1155 dummy
->succ
->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1159 fprintf (dump_file
, "Created preheader block for loop %i\n",
1165 /* Create preheaders for each loop from loop tree stored in LOOPS; for meaning
1166 of FLAGS see create_preheader. */
1168 create_preheaders (struct loops
*loops
, int flags
)
1171 for (i
= 1; i
< loops
->num
; i
++)
1172 create_preheader (loops
->parray
[i
], flags
);
1173 loops
->state
|= LOOPS_HAVE_PREHEADERS
;
1176 /* Forces all loop latches of loops from loop tree LOOPS to have only single
1179 force_single_succ_latches (struct loops
*loops
)
1185 for (i
= 1; i
< loops
->num
; i
++)
1187 loop
= loops
->parray
[i
];
1188 if (loop
->latch
!= loop
->header
1189 && !loop
->latch
->succ
->succ_next
)
1192 for (e
= loop
->header
->pred
; e
->src
!= loop
->latch
; e
= e
->pred_next
)
1195 loop_split_edge_with (e
, NULL_RTX
);
1197 loops
->state
|= LOOPS_HAVE_SIMPLE_LATCHES
;
1200 /* A quite stupid function to put INSNS on edge E. They are supposed to form
1201 just one basic block. Jumps in INSNS are not handled, so cfg do not have to
1202 be ok after this function. The created block is placed on correct place
1203 in LOOPS structure and its dominator is set. */
1205 loop_split_edge_with (edge e
, rtx insns
)
1207 basic_block src
, dest
, new_bb
;
1208 struct loop
*loop_c
;
1214 loop_c
= find_common_loop (src
->loop_father
, dest
->loop_father
);
1216 /* Create basic block for it. */
1218 new_bb
= split_edge (e
);
1219 add_bb_to_loop (new_bb
, loop_c
);
1220 new_bb
->flags
= insns
? BB_SUPERBLOCK
: 0;
1222 new_e
= new_bb
->succ
;
1223 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
1225 new_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
1226 new_e
->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1230 emit_insn_after (insns
, BB_END (new_bb
));
1232 if (dest
->loop_father
->latch
== src
)
1233 dest
->loop_father
->latch
= new_bb
;