1 /* Loop manipulation code for GNU compiler.
2 Copyright (C) 2002, 2003, 2004, 2005, 2007 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
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 COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
25 #include "hard-reg-set.h"
27 #include "basic-block.h"
29 #include "cfglayout.h"
33 static void duplicate_subloops (struct loops
*, struct loop
*, struct loop
*);
34 static void copy_loops_to (struct loops
*, struct loop
**, int,
36 static void loop_redirect_edge (edge
, basic_block
);
37 static bool loop_delete_branch_edge (edge
, int);
38 static void remove_bbs (basic_block
*, int);
39 static bool rpe_enum_p (basic_block
, void *);
40 static int find_path (edge
, basic_block
**);
41 static bool alp_enum_p (basic_block
, void *);
42 static void add_loop (struct loops
*, struct loop
*);
43 static void fix_loop_placements (struct loops
*, struct loop
*, bool *);
44 static bool fix_bb_placement (struct loops
*, basic_block
);
45 static void fix_bb_placements (struct loops
*, basic_block
, bool *);
46 static void place_new_loop (struct loops
*, struct loop
*);
47 static void scale_loop_frequencies (struct loop
*, int, int);
48 static basic_block
create_preheader (struct loop
*, int);
49 static void unloop (struct loops
*, struct loop
*, bool *);
51 #define RDIV(X,Y) (((X) + (Y) / 2) / (Y))
53 /* Checks whether basic block BB is dominated by DATA. */
55 rpe_enum_p (basic_block bb
, void *data
)
57 return dominated_by_p (CDI_DOMINATORS
, bb
, data
);
60 /* Remove basic blocks BBS from loop structure and dominance info,
61 and delete them afterwards. */
63 remove_bbs (basic_block
*bbs
, int nbbs
)
67 for (i
= 0; i
< nbbs
; i
++)
69 remove_bb_from_loops (bbs
[i
]);
70 delete_basic_block (bbs
[i
]);
74 /* Find path -- i.e. the basic blocks dominated by edge E and put them
75 into array BBS, that will be allocated large enough to contain them.
76 E->dest must have exactly one predecessor for this to work (it is
77 easy to achieve and we do not put it here because we do not want to
78 alter anything by this function). The number of basic blocks in the
81 find_path (edge e
, basic_block
**bbs
)
83 gcc_assert (EDGE_COUNT (e
->dest
->preds
) <= 1);
85 /* Find bbs in the path. */
86 *bbs
= XCNEWVEC (basic_block
, n_basic_blocks
);
87 return dfs_enumerate_from (e
->dest
, 0, rpe_enum_p
, *bbs
,
88 n_basic_blocks
, e
->dest
);
91 /* Fix placement of basic block BB inside loop hierarchy stored in LOOPS --
92 Let L be a loop to that BB belongs. Then every successor of BB must either
93 1) belong to some superloop of loop L, or
94 2) be a header of loop K such that K->outer is superloop of L
95 Returns true if we had to move BB into other loop to enforce this condition,
96 false if the placement of BB was already correct (provided that placements
97 of its successors are correct). */
99 fix_bb_placement (struct loops
*loops
, basic_block bb
)
103 struct loop
*loop
= loops
->tree_root
, *act
;
105 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
107 if (e
->dest
== EXIT_BLOCK_PTR
)
110 act
= e
->dest
->loop_father
;
111 if (act
->header
== e
->dest
)
114 if (flow_loop_nested_p (loop
, act
))
118 if (loop
== bb
->loop_father
)
121 remove_bb_from_loops (bb
);
122 add_bb_to_loop (bb
, loop
);
127 /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e.
128 enforce condition condition stated in description of fix_bb_placement. We
129 start from basic block FROM that had some of its successors removed, so that
130 his placement no longer has to be correct, and iteratively fix placement of
131 its predecessors that may change if placement of FROM changed. Also fix
132 placement of subloops of FROM->loop_father, that might also be altered due
133 to this change; the condition for them is similar, except that instead of
134 successors we consider edges coming out of the loops.
136 If the changes may invalidate the information about irreducible regions,
137 IRRED_INVALIDATED is set to true. */
140 fix_bb_placements (struct loops
*loops
, basic_block from
,
141 bool *irred_invalidated
)
144 basic_block
*queue
, *qtop
, *qbeg
, *qend
;
145 struct loop
*base_loop
;
148 /* We pass through blocks back-reachable from FROM, testing whether some
149 of their successors moved to outer loop. It may be necessary to
150 iterate several times, but it is finite, as we stop unless we move
151 the basic block up the loop structure. The whole story is a bit
152 more complicated due to presence of subloops, those are moved using
153 fix_loop_placement. */
155 base_loop
= from
->loop_father
;
156 if (base_loop
== loops
->tree_root
)
159 in_queue
= sbitmap_alloc (last_basic_block
);
160 sbitmap_zero (in_queue
);
161 SET_BIT (in_queue
, from
->index
);
162 /* Prevent us from going out of the base_loop. */
163 SET_BIT (in_queue
, base_loop
->header
->index
);
165 queue
= XNEWVEC (basic_block
, base_loop
->num_nodes
+ 1);
166 qtop
= queue
+ base_loop
->num_nodes
+ 1;
178 RESET_BIT (in_queue
, from
->index
);
180 if (from
->loop_father
->header
== from
)
182 /* Subloop header, maybe move the loop upward. */
183 if (!fix_loop_placement (from
->loop_father
))
188 /* Ordinary basic block. */
189 if (!fix_bb_placement (loops
, from
))
193 FOR_EACH_EDGE (e
, ei
, from
->succs
)
195 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
196 *irred_invalidated
= true;
199 /* Something has changed, insert predecessors into queue. */
200 FOR_EACH_EDGE (e
, ei
, from
->preds
)
202 basic_block pred
= e
->src
;
205 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
206 *irred_invalidated
= true;
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 /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E
241 and update loop structure stored in LOOPS and dominators. Return true if
242 we were able to remove the path, false otherwise (and nothing is affected
245 remove_path (struct loops
*loops
, edge e
)
248 basic_block
*rem_bbs
, *bord_bbs
, *dom_bbs
, from
, bb
;
249 int i
, nrem
, n_bord_bbs
, n_dom_bbs
;
251 bool deleted
, irred_invalidated
= false;
253 if (!loop_delete_branch_edge (e
, 0))
256 /* Keep track of whether we need to update information about irreducible
257 regions. This is the case if the removed area is a part of the
258 irreducible region, or if the set of basic blocks that belong to a loop
259 that is inside an irreducible region is changed, or if such a loop is
261 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
262 irred_invalidated
= true;
264 /* We need to check whether basic blocks are dominated by the edge
265 e, but we only have basic block dominators. This is easy to
266 fix -- when e->dest has exactly one predecessor, this corresponds
267 to blocks dominated by e->dest, if not, split the edge. */
268 if (!single_pred_p (e
->dest
))
269 e
= single_pred_edge (loop_split_edge_with (e
, NULL_RTX
));
271 /* It may happen that by removing path we remove one or more loops
272 we belong to. In this case first unloop the loops, then proceed
273 normally. We may assume that e->dest is not a header of any loop,
274 as it now has exactly one predecessor. */
275 while (e
->src
->loop_father
->outer
276 && dominated_by_p (CDI_DOMINATORS
,
277 e
->src
->loop_father
->latch
, e
->dest
))
278 unloop (loops
, e
->src
->loop_father
, &irred_invalidated
);
280 /* Identify the path. */
281 nrem
= find_path (e
, &rem_bbs
);
284 bord_bbs
= XCNEWVEC (basic_block
, n_basic_blocks
);
285 seen
= sbitmap_alloc (last_basic_block
);
288 /* Find "border" hexes -- i.e. those with predecessor in removed path. */
289 for (i
= 0; i
< nrem
; i
++)
290 SET_BIT (seen
, rem_bbs
[i
]->index
);
291 for (i
= 0; i
< nrem
; i
++)
295 FOR_EACH_EDGE (ae
, ei
, rem_bbs
[i
]->succs
)
296 if (ae
->dest
!= EXIT_BLOCK_PTR
&& !TEST_BIT (seen
, ae
->dest
->index
))
298 SET_BIT (seen
, ae
->dest
->index
);
299 bord_bbs
[n_bord_bbs
++] = ae
->dest
;
301 if (ae
->flags
& EDGE_IRREDUCIBLE_LOOP
)
302 irred_invalidated
= true;
306 /* Remove the path. */
308 deleted
= loop_delete_branch_edge (e
, 1);
309 gcc_assert (deleted
);
310 dom_bbs
= XCNEWVEC (basic_block
, n_basic_blocks
);
312 /* Cancel loops contained in the path. */
313 for (i
= 0; i
< nrem
; i
++)
314 if (rem_bbs
[i
]->loop_father
->header
== rem_bbs
[i
])
315 cancel_loop_tree (loops
, rem_bbs
[i
]->loop_father
);
317 remove_bbs (rem_bbs
, nrem
);
320 /* Find blocks whose dominators may be affected. */
323 for (i
= 0; i
< n_bord_bbs
; i
++)
327 bb
= get_immediate_dominator (CDI_DOMINATORS
, bord_bbs
[i
]);
328 if (TEST_BIT (seen
, bb
->index
))
330 SET_BIT (seen
, bb
->index
);
332 for (ldom
= first_dom_son (CDI_DOMINATORS
, bb
);
334 ldom
= next_dom_son (CDI_DOMINATORS
, ldom
))
335 if (!dominated_by_p (CDI_DOMINATORS
, from
, ldom
))
336 dom_bbs
[n_dom_bbs
++] = ldom
;
341 /* Recount dominators. */
342 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, n_dom_bbs
);
346 /* Fix placements of basic blocks inside loops and the placement of
347 loops in the loop tree. */
348 fix_bb_placements (loops
, from
, &irred_invalidated
);
349 fix_loop_placements (loops
, from
->loop_father
, &irred_invalidated
);
351 if (irred_invalidated
352 && (loops
->state
& LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
) != 0)
353 mark_irreducible_loops (loops
);
358 /* Predicate for enumeration in add_loop. */
360 alp_enum_p (basic_block bb
, void *alp_header
)
362 return bb
!= (basic_block
) alp_header
;
365 /* Given LOOP structure with filled header and latch, find the body of the
366 corresponding loop and add it to LOOPS tree. */
368 add_loop (struct loops
*loops
, struct loop
*loop
)
373 /* Add it to loop structure. */
374 place_new_loop (loops
, loop
);
377 /* Find its nodes. */
378 bbs
= XCNEWVEC (basic_block
, n_basic_blocks
);
379 n
= dfs_enumerate_from (loop
->latch
, 1, alp_enum_p
,
380 bbs
, n_basic_blocks
, loop
->header
);
382 for (i
= 0; i
< n
; i
++)
383 add_bb_to_loop (bbs
[i
], loop
);
384 add_bb_to_loop (loop
->header
, loop
);
389 /* Multiply all frequencies in LOOP by NUM/DEN. */
391 scale_loop_frequencies (struct loop
*loop
, int num
, int den
)
395 bbs
= get_loop_body (loop
);
396 scale_bbs_frequencies_int (bbs
, loop
->num_nodes
, num
, den
);
400 /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting
401 latch to header and update loop tree stored in LOOPS and dominators
402 accordingly. Everything between them plus LATCH_EDGE destination must
403 be dominated by HEADER_EDGE destination, and back-reachable from
404 LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB,
405 FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and
406 TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE.
407 Returns newly created loop. */
410 loopify (struct loops
*loops
, edge latch_edge
, edge header_edge
,
411 basic_block switch_bb
, edge true_edge
, edge false_edge
,
412 bool redirect_all_edges
)
414 basic_block succ_bb
= latch_edge
->dest
;
415 basic_block pred_bb
= header_edge
->src
;
416 basic_block
*dom_bbs
, *body
;
417 unsigned n_dom_bbs
, i
;
419 struct loop
*loop
= XCNEW (struct loop
);
420 struct loop
*outer
= succ_bb
->loop_father
->outer
;
421 int freq
, prob
, tot_prob
;
426 loop
->header
= header_edge
->dest
;
427 loop
->latch
= latch_edge
->src
;
429 freq
= EDGE_FREQUENCY (header_edge
);
430 cnt
= header_edge
->count
;
431 prob
= EDGE_SUCC (switch_bb
, 0)->probability
;
432 tot_prob
= prob
+ EDGE_SUCC (switch_bb
, 1)->probability
;
436 /* Redirect edges. */
437 loop_redirect_edge (latch_edge
, loop
->header
);
438 loop_redirect_edge (true_edge
, succ_bb
);
440 /* During loop versioning, one of the switch_bb edge is already properly
441 set. Do not redirect it again unless redirect_all_edges is true. */
442 if (redirect_all_edges
)
444 loop_redirect_edge (header_edge
, switch_bb
);
445 loop_redirect_edge (false_edge
, loop
->header
);
447 /* Update dominators. */
448 set_immediate_dominator (CDI_DOMINATORS
, switch_bb
, pred_bb
);
449 set_immediate_dominator (CDI_DOMINATORS
, loop
->header
, switch_bb
);
452 set_immediate_dominator (CDI_DOMINATORS
, succ_bb
, switch_bb
);
454 /* Compute new loop. */
455 add_loop (loops
, loop
);
456 flow_loop_tree_node_add (outer
, loop
);
458 /* Add switch_bb to appropriate loop. */
459 add_bb_to_loop (switch_bb
, outer
);
461 /* Fix frequencies. */
462 switch_bb
->frequency
= freq
;
463 switch_bb
->count
= cnt
;
464 FOR_EACH_EDGE (e
, ei
, switch_bb
->succs
)
465 e
->count
= (switch_bb
->count
* e
->probability
) / REG_BR_PROB_BASE
;
466 scale_loop_frequencies (loop
, prob
, tot_prob
);
467 scale_loop_frequencies (succ_bb
->loop_father
, tot_prob
- prob
, tot_prob
);
469 /* Update dominators of blocks outside of LOOP. */
470 dom_bbs
= XCNEWVEC (basic_block
, n_basic_blocks
);
472 seen
= sbitmap_alloc (last_basic_block
);
474 body
= get_loop_body (loop
);
476 for (i
= 0; i
< loop
->num_nodes
; i
++)
477 SET_BIT (seen
, body
[i
]->index
);
479 for (i
= 0; i
< loop
->num_nodes
; i
++)
483 for (ldom
= first_dom_son (CDI_DOMINATORS
, body
[i
]);
485 ldom
= next_dom_son (CDI_DOMINATORS
, ldom
))
486 if (!TEST_BIT (seen
, ldom
->index
))
488 SET_BIT (seen
, ldom
->index
);
489 dom_bbs
[n_dom_bbs
++] = ldom
;
493 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, n_dom_bbs
);
502 /* Remove the latch edge of a LOOP and update LOOPS tree to indicate that
503 the LOOP was removed. After this function, original loop latch will
504 have no successor, which caller is expected to fix somehow.
506 If this may cause the information about irreducible regions to become
507 invalid, IRRED_INVALIDATED is set to true. */
510 unloop (struct loops
*loops
, struct loop
*loop
, bool *irred_invalidated
)
515 basic_block latch
= loop
->latch
;
518 if (loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
)
519 *irred_invalidated
= true;
521 /* This is relatively straightforward. The dominators are unchanged, as
522 loop header dominates loop latch, so the only thing we have to care of
523 is the placement of loops and basic blocks inside the loop tree. We
524 move them all to the loop->outer, and then let fix_bb_placements do
527 body
= get_loop_body (loop
);
529 for (i
= 0; i
< n
; i
++)
530 if (body
[i
]->loop_father
== loop
)
532 remove_bb_from_loops (body
[i
]);
533 add_bb_to_loop (body
[i
], loop
->outer
);
540 flow_loop_tree_node_remove (ploop
);
541 flow_loop_tree_node_add (loop
->outer
, ploop
);
544 /* Remove the loop and free its data. */
545 flow_loop_tree_node_remove (loop
);
546 loops
->parray
[loop
->num
] = NULL
;
547 flow_loop_free (loop
);
549 remove_edge (single_succ_edge (latch
));
551 /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if
552 there is an irreducible region inside the cancelled loop, the flags will
554 fix_bb_placements (loops
, latch
, &dummy
);
557 /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop
558 FATHER of LOOP such that all of the edges coming out of LOOP belong to
559 FATHER, and set it as outer loop of LOOP. Return true if placement of
563 fix_loop_placement (struct loop
*loop
)
569 struct loop
*father
= loop
->pred
[0], *act
;
571 body
= get_loop_body (loop
);
572 for (i
= 0; i
< loop
->num_nodes
; i
++)
573 FOR_EACH_EDGE (e
, ei
, body
[i
]->succs
)
574 if (!flow_bb_inside_loop_p (loop
, e
->dest
))
576 act
= find_common_loop (loop
, e
->dest
->loop_father
);
577 if (flow_loop_nested_p (father
, act
))
582 if (father
!= loop
->outer
)
584 for (act
= loop
->outer
; act
!= father
; act
= act
->outer
)
585 act
->num_nodes
-= loop
->num_nodes
;
586 flow_loop_tree_node_remove (loop
);
587 flow_loop_tree_node_add (father
, loop
);
593 /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that
594 condition stated in description of fix_loop_placement holds for them.
595 It is used in case when we removed some edges coming out of LOOP, which
596 may cause the right placement of LOOP inside loop tree to change.
598 IRRED_INVALIDATED is set to true if a change in the loop structures might
599 invalidate the information about irreducible regions. */
602 fix_loop_placements (struct loops
*loops
, struct loop
*loop
,
603 bool *irred_invalidated
)
610 if (!fix_loop_placement (loop
))
613 /* Changing the placement of a loop in the loop tree may alter the
614 validity of condition 2) of the description of fix_bb_placement
615 for its preheader, because the successor is the header and belongs
616 to the loop. So call fix_bb_placements to fix up the placement
617 of the preheader and (possibly) of its predecessors. */
618 fix_bb_placements (loops
, loop_preheader_edge (loop
)->src
,
624 /* Creates place for a new LOOP in LOOPS structure. */
626 place_new_loop (struct loops
*loops
, struct loop
*loop
)
629 xrealloc (loops
->parray
, (loops
->num
+ 1) * sizeof (struct loop
*));
630 loops
->parray
[loops
->num
] = loop
;
632 loop
->num
= loops
->num
++;
635 /* Copies copy of LOOP as subloop of TARGET loop, placing newly
636 created loop into LOOPS structure. */
638 duplicate_loop (struct loops
*loops
, struct loop
*loop
, struct loop
*target
)
641 cloop
= XCNEW (struct loop
);
642 place_new_loop (loops
, cloop
);
644 /* Initialize copied loop. */
645 cloop
->level
= loop
->level
;
647 /* Set it as copy of loop. */
650 /* Add it to target. */
651 flow_loop_tree_node_add (target
, cloop
);
656 /* Copies structure of subloops of LOOP into TARGET loop, placing
657 newly created loops into loop tree stored in LOOPS. */
659 duplicate_subloops (struct loops
*loops
, struct loop
*loop
, struct loop
*target
)
661 struct loop
*aloop
, *cloop
;
663 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
665 cloop
= duplicate_loop (loops
, aloop
, target
);
666 duplicate_subloops (loops
, aloop
, cloop
);
670 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS,
671 into TARGET loop, placing newly created loops into loop tree LOOPS. */
673 copy_loops_to (struct loops
*loops
, struct loop
**copied_loops
, int n
, struct loop
*target
)
678 for (i
= 0; i
< n
; i
++)
680 aloop
= duplicate_loop (loops
, copied_loops
[i
], target
);
681 duplicate_subloops (loops
, copied_loops
[i
], aloop
);
685 /* Redirects edge E to basic block DEST. */
687 loop_redirect_edge (edge e
, basic_block dest
)
692 redirect_edge_and_branch_force (e
, dest
);
695 /* Deletes edge E from a branch if possible. Unless REALLY_DELETE is set,
696 just test whether it is possible to remove the edge. */
698 loop_delete_branch_edge (edge e
, int really_delete
)
700 basic_block src
= e
->src
;
705 gcc_assert (EDGE_COUNT (src
->succs
) > 1);
707 /* Cannot handle more than two exit edges. */
708 if (EDGE_COUNT (src
->succs
) > 2)
710 /* And it must be just a simple branch. */
711 if (!any_condjump_p (BB_END (src
)))
714 snd
= e
== EDGE_SUCC (src
, 0) ? EDGE_SUCC (src
, 1) : EDGE_SUCC (src
, 0);
716 if (newdest
== EXIT_BLOCK_PTR
)
719 /* Hopefully the above conditions should suffice. */
723 /* Redirecting behaves wrongly wrto this flag. */
724 irr
= snd
->flags
& EDGE_IRREDUCIBLE_LOOP
;
726 if (!redirect_edge_and_branch (e
, newdest
))
728 single_succ_edge (src
)->flags
&= ~EDGE_IRREDUCIBLE_LOOP
;
729 single_succ_edge (src
)->flags
|= irr
;
734 /* Check whether LOOP's body can be duplicated. */
736 can_duplicate_loop_p (struct loop
*loop
)
739 basic_block
*bbs
= get_loop_body (loop
);
741 ret
= can_copy_bbs_p (bbs
, loop
->num_nodes
);
747 /* The NBBS blocks in BBS will get duplicated and the copies will be placed
748 to LOOP. Update the single_exit information in superloops of LOOP. */
751 update_single_exits_after_duplication (basic_block
*bbs
, unsigned nbbs
,
756 for (i
= 0; i
< nbbs
; i
++)
757 bbs
[i
]->flags
|= BB_DUPLICATED
;
759 for (; loop
->outer
; loop
= loop
->outer
)
761 if (!loop
->single_exit
)
764 if (loop
->single_exit
->src
->flags
& BB_DUPLICATED
)
765 loop
->single_exit
= NULL
;
768 for (i
= 0; i
< nbbs
; i
++)
769 bbs
[i
]->flags
&= ~BB_DUPLICATED
;
772 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating
773 LOOPS structure and dominators. E's destination must be LOOP header for
774 this to work, i.e. it must be entry or latch edge of this loop; these are
775 unique, as the loops must have preheaders for this function to work
776 correctly (in case E is latch, the function unrolls the loop, if E is entry
777 edge, it peels the loop). Store edges created by copying ORIG edge from
778 copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to
779 original LOOP body, the other copies are numbered in order given by control
780 flow through them) into TO_REMOVE array. Returns false if duplication is
783 duplicate_loop_to_header_edge (struct loop
*loop
, edge e
, struct loops
*loops
,
784 unsigned int ndupl
, sbitmap wont_exit
,
785 edge orig
, edge
*to_remove
,
786 unsigned int *n_to_remove
, int flags
)
788 struct loop
*target
, *aloop
;
789 struct loop
**orig_loops
;
790 unsigned n_orig_loops
;
791 basic_block header
= loop
->header
, latch
= loop
->latch
;
792 basic_block
*new_bbs
, *bbs
, *first_active
;
793 basic_block new_bb
, bb
, first_active_latch
= NULL
;
795 edge spec_edges
[2], new_spec_edges
[2];
799 int is_latch
= (latch
== e
->src
);
800 int scale_act
= 0, *scale_step
= NULL
, scale_main
= 0;
801 int p
, freq_in
, freq_le
, freq_out_orig
;
802 int prob_pass_thru
, prob_pass_wont_exit
, prob_pass_main
;
803 int add_irreducible_flag
;
804 basic_block place_after
;
806 gcc_assert (e
->dest
== loop
->header
);
807 gcc_assert (ndupl
> 0);
811 /* Orig must be edge out of the loop. */
812 gcc_assert (flow_bb_inside_loop_p (loop
, orig
->src
));
813 gcc_assert (!flow_bb_inside_loop_p (loop
, orig
->dest
));
817 bbs
= get_loop_body_in_dom_order (loop
);
818 gcc_assert (bbs
[0] == loop
->header
);
819 gcc_assert (bbs
[n
- 1] == loop
->latch
);
821 /* Check whether duplication is possible. */
822 if (!can_copy_bbs_p (bbs
, loop
->num_nodes
))
827 new_bbs
= XNEWVEC (basic_block
, loop
->num_nodes
);
829 /* In case we are doing loop peeling and the loop is in the middle of
830 irreducible region, the peeled copies will be inside it too. */
831 add_irreducible_flag
= e
->flags
& EDGE_IRREDUCIBLE_LOOP
;
832 gcc_assert (!is_latch
|| !add_irreducible_flag
);
834 /* Find edge from latch. */
835 latch_edge
= loop_latch_edge (loop
);
837 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
839 /* Calculate coefficients by that we have to scale frequencies
840 of duplicated loop bodies. */
841 freq_in
= header
->frequency
;
842 freq_le
= EDGE_FREQUENCY (latch_edge
);
845 if (freq_in
< freq_le
)
847 freq_out_orig
= orig
? EDGE_FREQUENCY (orig
) : freq_in
- freq_le
;
848 if (freq_out_orig
> freq_in
- freq_le
)
849 freq_out_orig
= freq_in
- freq_le
;
850 prob_pass_thru
= RDIV (REG_BR_PROB_BASE
* freq_le
, freq_in
);
851 prob_pass_wont_exit
=
852 RDIV (REG_BR_PROB_BASE
* (freq_le
+ freq_out_orig
), freq_in
);
854 scale_step
= XNEWVEC (int, ndupl
);
856 for (i
= 1; i
<= ndupl
; i
++)
857 scale_step
[i
- 1] = TEST_BIT (wont_exit
, i
)
858 ? prob_pass_wont_exit
861 /* Complete peeling is special as the probability of exit in last
863 if (flags
& DLTHE_FLAG_COMPLETTE_PEEL
)
865 int wanted_freq
= EDGE_FREQUENCY (e
);
867 if (wanted_freq
> freq_in
)
868 wanted_freq
= freq_in
;
870 gcc_assert (!is_latch
);
871 /* First copy has frequency of incoming edge. Each subsequent
872 frequency should be reduced by prob_pass_wont_exit. Caller
873 should've managed the flags so all except for original loop
874 has won't exist set. */
875 scale_act
= RDIV (wanted_freq
* REG_BR_PROB_BASE
, freq_in
);
876 /* Now simulate the duplication adjustments and compute header
877 frequency of the last copy. */
878 for (i
= 0; i
< ndupl
; i
++)
879 wanted_freq
= RDIV (wanted_freq
* scale_step
[i
], REG_BR_PROB_BASE
);
880 scale_main
= RDIV (wanted_freq
* REG_BR_PROB_BASE
, freq_in
);
884 prob_pass_main
= TEST_BIT (wont_exit
, 0)
885 ? prob_pass_wont_exit
888 scale_main
= REG_BR_PROB_BASE
;
889 for (i
= 0; i
< ndupl
; i
++)
892 p
= RDIV (p
* scale_step
[i
], REG_BR_PROB_BASE
);
894 scale_main
= RDIV (REG_BR_PROB_BASE
* REG_BR_PROB_BASE
, scale_main
);
895 scale_act
= RDIV (scale_main
* prob_pass_main
, REG_BR_PROB_BASE
);
899 scale_main
= REG_BR_PROB_BASE
;
900 for (i
= 0; i
< ndupl
; i
++)
901 scale_main
= RDIV (scale_main
* scale_step
[i
], REG_BR_PROB_BASE
);
902 scale_act
= REG_BR_PROB_BASE
- prob_pass_thru
;
904 for (i
= 0; i
< ndupl
; i
++)
905 gcc_assert (scale_step
[i
] >= 0 && scale_step
[i
] <= REG_BR_PROB_BASE
);
906 gcc_assert (scale_main
>= 0 && scale_main
<= REG_BR_PROB_BASE
907 && scale_act
>= 0 && scale_act
<= REG_BR_PROB_BASE
);
910 /* Loop the new bbs will belong to. */
911 target
= e
->src
->loop_father
;
913 /* Original loops. */
915 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
917 orig_loops
= XCNEWVEC (struct loop
*, n_orig_loops
);
918 for (aloop
= loop
->inner
, i
= 0; aloop
; aloop
= aloop
->next
, i
++)
919 orig_loops
[i
] = aloop
;
923 first_active
= XNEWVEC (basic_block
, n
);
926 memcpy (first_active
, bbs
, n
* sizeof (basic_block
));
927 first_active_latch
= latch
;
930 /* Update the information about single exits. */
931 if (loops
->state
& LOOPS_HAVE_MARKED_SINGLE_EXITS
)
932 update_single_exits_after_duplication (bbs
, n
, target
);
934 /* Record exit edge in original loop body. */
935 if (orig
&& TEST_BIT (wont_exit
, 0))
936 to_remove
[(*n_to_remove
)++] = orig
;
938 spec_edges
[SE_ORIG
] = orig
;
939 spec_edges
[SE_LATCH
] = latch_edge
;
941 place_after
= e
->src
;
942 for (j
= 0; j
< ndupl
; j
++)
945 copy_loops_to (loops
, orig_loops
, n_orig_loops
, target
);
948 copy_bbs (bbs
, n
, new_bbs
, spec_edges
, 2, new_spec_edges
, loop
,
950 place_after
= new_spec_edges
[SE_LATCH
]->src
;
952 if (flags
& DLTHE_RECORD_COPY_NUMBER
)
953 for (i
= 0; i
< n
; i
++)
955 gcc_assert (!new_bbs
[i
]->aux
);
956 new_bbs
[i
]->aux
= (void *)(size_t)(j
+ 1);
959 /* Note whether the blocks and edges belong to an irreducible loop. */
960 if (add_irreducible_flag
)
962 for (i
= 0; i
< n
; i
++)
963 new_bbs
[i
]->flags
|= BB_DUPLICATED
;
964 for (i
= 0; i
< n
; i
++)
968 if (new_bb
->loop_father
== target
)
969 new_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
971 FOR_EACH_EDGE (ae
, ei
, new_bb
->succs
)
972 if ((ae
->dest
->flags
& BB_DUPLICATED
)
973 && (ae
->src
->loop_father
== target
974 || ae
->dest
->loop_father
== target
))
975 ae
->flags
|= EDGE_IRREDUCIBLE_LOOP
;
977 for (i
= 0; i
< n
; i
++)
978 new_bbs
[i
]->flags
&= ~BB_DUPLICATED
;
981 /* Redirect the special edges. */
984 redirect_edge_and_branch_force (latch_edge
, new_bbs
[0]);
985 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
987 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], latch
);
988 latch
= loop
->latch
= new_bbs
[n
- 1];
989 e
= latch_edge
= new_spec_edges
[SE_LATCH
];
993 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
995 redirect_edge_and_branch_force (e
, new_bbs
[0]);
996 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], e
->src
);
997 e
= new_spec_edges
[SE_LATCH
];
1000 /* Record exit edge in this copy. */
1001 if (orig
&& TEST_BIT (wont_exit
, j
+ 1))
1002 to_remove
[(*n_to_remove
)++] = new_spec_edges
[SE_ORIG
];
1004 /* Record the first copy in the control flow order if it is not
1005 the original loop (i.e. in case of peeling). */
1006 if (!first_active_latch
)
1008 memcpy (first_active
, new_bbs
, n
* sizeof (basic_block
));
1009 first_active_latch
= new_bbs
[n
- 1];
1012 /* Set counts and frequencies. */
1013 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1015 scale_bbs_frequencies_int (new_bbs
, n
, scale_act
, REG_BR_PROB_BASE
);
1016 scale_act
= RDIV (scale_act
* scale_step
[j
], REG_BR_PROB_BASE
);
1022 /* Update the original loop. */
1024 set_immediate_dominator (CDI_DOMINATORS
, e
->dest
, e
->src
);
1025 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1027 scale_bbs_frequencies_int (bbs
, n
, scale_main
, REG_BR_PROB_BASE
);
1031 /* Update dominators of outer blocks if affected. */
1032 for (i
= 0; i
< n
; i
++)
1034 basic_block dominated
, dom_bb
, *dom_bbs
;
1040 n_dom_bbs
= get_dominated_by (CDI_DOMINATORS
, bb
, &dom_bbs
);
1041 for (j
= 0; j
< n_dom_bbs
; j
++)
1043 dominated
= dom_bbs
[j
];
1044 if (flow_bb_inside_loop_p (loop
, dominated
))
1046 dom_bb
= nearest_common_dominator (
1047 CDI_DOMINATORS
, first_active
[i
], first_active_latch
);
1048 set_immediate_dominator (CDI_DOMINATORS
, dominated
, dom_bb
);
1052 free (first_active
);
1059 /* A callback for make_forwarder block, to redirect all edges except for
1060 MFB_KJ_EDGE to the entry part. E is the edge for that we should decide
1061 whether to redirect it. */
1063 static edge mfb_kj_edge
;
1065 mfb_keep_just (edge e
)
1067 return e
!= mfb_kj_edge
;
1070 /* A callback for make_forwarder block, to update data structures for a basic
1071 block JUMP created by redirecting an edge (only the latch edge is being
1075 mfb_update_loops (basic_block jump
)
1077 struct loop
*loop
= single_succ (jump
)->loop_father
;
1079 if (dom_computed
[CDI_DOMINATORS
])
1080 set_immediate_dominator (CDI_DOMINATORS
, jump
, single_pred (jump
));
1081 add_bb_to_loop (jump
, loop
);
1085 /* Creates a pre-header for a LOOP. Returns newly created block. Unless
1086 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single
1087 entry; otherwise we also force preheader block to have only one successor.
1088 The function also updates dominators. */
1091 create_preheader (struct loop
*loop
, int flags
)
1095 struct loop
*cloop
, *ploop
;
1098 bool latch_edge_was_fallthru
;
1099 edge one_succ_pred
= 0;
1102 cloop
= loop
->outer
;
1104 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
1106 if (e
->src
== loop
->latch
)
1108 irred
|= (e
->flags
& EDGE_IRREDUCIBLE_LOOP
) != 0;
1110 if (single_succ_p (e
->src
))
1113 gcc_assert (nentry
);
1116 /* Get an edge that is different from the one from loop->latch
1118 e
= EDGE_PRED (loop
->header
,
1119 EDGE_PRED (loop
->header
, 0)->src
== loop
->latch
);
1121 if (!(flags
& CP_SIMPLE_PREHEADERS
) || single_succ_p (e
->src
))
1125 mfb_kj_edge
= loop_latch_edge (loop
);
1126 latch_edge_was_fallthru
= (mfb_kj_edge
->flags
& EDGE_FALLTHRU
) != 0;
1127 fallthru
= make_forwarder_block (loop
->header
, mfb_keep_just
,
1129 dummy
= fallthru
->src
;
1130 loop
->header
= fallthru
->dest
;
1132 /* The header could be a latch of some superloop(s); due to design of
1133 split_block, it would now move to fallthru->dest. */
1134 for (ploop
= loop
; ploop
; ploop
= ploop
->outer
)
1135 if (ploop
->latch
== dummy
)
1136 ploop
->latch
= fallthru
->dest
;
1138 /* Try to be clever in placing the newly created preheader. The idea is to
1139 avoid breaking any "fallthruness" relationship between blocks.
1141 The preheader was created just before the header and all incoming edges
1142 to the header were redirected to the preheader, except the latch edge.
1143 So the only problematic case is when this latch edge was a fallthru
1144 edge: it is not anymore after the preheader creation so we have broken
1145 the fallthruness. We're therefore going to look for a better place. */
1146 if (latch_edge_was_fallthru
)
1151 e
= EDGE_PRED (dummy
, 0);
1153 move_block_after (dummy
, e
->src
);
1156 loop
->header
->loop_father
= loop
;
1157 add_bb_to_loop (dummy
, cloop
);
1161 dummy
->flags
|= BB_IRREDUCIBLE_LOOP
;
1162 single_succ_edge (dummy
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1166 fprintf (dump_file
, "Created preheader block for loop %i\n",
1172 /* Create preheaders for each loop from loop tree stored in LOOPS; for meaning
1173 of FLAGS see create_preheader. */
1175 create_preheaders (struct loops
*loops
, int flags
)
1178 for (i
= 1; i
< loops
->num
; i
++)
1179 create_preheader (loops
->parray
[i
], flags
);
1180 loops
->state
|= LOOPS_HAVE_PREHEADERS
;
1183 /* Forces all loop latches of loops from loop tree LOOPS to have only single
1186 force_single_succ_latches (struct loops
*loops
)
1192 for (i
= 1; i
< loops
->num
; i
++)
1194 loop
= loops
->parray
[i
];
1195 if (loop
->latch
!= loop
->header
&& single_succ_p (loop
->latch
))
1198 e
= find_edge (loop
->latch
, loop
->header
);
1200 loop_split_edge_with (e
, NULL_RTX
);
1202 loops
->state
|= LOOPS_HAVE_SIMPLE_LATCHES
;
1205 /* A quite stupid function to put INSNS on edge E. They are supposed to form
1206 just one basic block. Jumps in INSNS are not handled, so cfg do not have to
1207 be ok after this function. The created block is placed on correct place
1208 in LOOPS structure and its dominator is set. */
1210 loop_split_edge_with (edge e
, rtx insns
)
1212 basic_block src
, dest
, new_bb
;
1213 struct loop
*loop_c
;
1218 loop_c
= find_common_loop (src
->loop_father
, dest
->loop_father
);
1220 /* Create basic block for it. */
1222 new_bb
= split_edge (e
);
1223 add_bb_to_loop (new_bb
, loop_c
);
1224 new_bb
->flags
|= (insns
? BB_SUPERBLOCK
: 0);
1227 emit_insn_after (insns
, BB_END (new_bb
));
1229 if (dest
->loop_father
->latch
== src
)
1230 dest
->loop_father
->latch
= new_bb
;
1235 /* This function is called from loop_version. It splits the entry edge
1236 of the loop we want to version, adds the versioning condition, and
1237 adjust the edges to the two versions of the loop appropriately.
1238 e is an incoming edge. Returns the basic block containing the
1241 --- edge e ---- > [second_head]
1243 Split it and insert new conditional expression and adjust edges.
1245 --- edge e ---> [cond expr] ---> [first_head]
1247 +---------> [second_head]
1251 lv_adjust_loop_entry_edge (basic_block first_head
,
1252 basic_block second_head
,
1256 basic_block new_head
= NULL
;
1259 gcc_assert (e
->dest
== second_head
);
1261 /* Split edge 'e'. This will create a new basic block, where we can
1262 insert conditional expr. */
1263 new_head
= split_edge (e
);
1266 lv_add_condition_to_bb (first_head
, second_head
, new_head
,
1269 /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */
1270 e1
= make_edge (new_head
, first_head
, ir_type () ? EDGE_TRUE_VALUE
: 0);
1271 set_immediate_dominator (CDI_DOMINATORS
, first_head
, new_head
);
1272 set_immediate_dominator (CDI_DOMINATORS
, second_head
, new_head
);
1274 /* Adjust loop header phi nodes. */
1275 lv_adjust_loop_header_phi (first_head
, second_head
, new_head
, e1
);
1280 /* Main entry point for Loop Versioning transformation.
1282 This transformation given a condition and a loop, creates
1283 -if (condition) { loop_copy1 } else { loop_copy2 },
1284 where loop_copy1 is the loop transformed in one way, and loop_copy2
1285 is the loop transformed in another way (or unchanged). 'condition'
1286 may be a run time test for things that were not resolved by static
1287 analysis (overlapping ranges (anti-aliasing), alignment, etc.).
1289 If PLACE_AFTER is true, we place the new loop after LOOP in the
1290 instruction stream, otherwise it is placed before LOOP. */
1293 loop_version (struct loops
*loops
, struct loop
* loop
,
1294 void *cond_expr
, basic_block
*condition_bb
,
1297 basic_block first_head
, second_head
;
1298 edge entry
, latch_edge
, exit
, true_edge
, false_edge
;
1301 basic_block cond_bb
;
1303 /* CHECKME: Loop versioning does not handle nested loop at this point. */
1307 /* Record entry and latch edges for the loop */
1308 entry
= loop_preheader_edge (loop
);
1309 irred_flag
= entry
->flags
& EDGE_IRREDUCIBLE_LOOP
;
1310 entry
->flags
&= ~EDGE_IRREDUCIBLE_LOOP
;
1312 /* Note down head of loop as first_head. */
1313 first_head
= entry
->dest
;
1315 /* Duplicate loop. */
1316 if (!cfg_hook_duplicate_loop_to_header_edge (loop
, entry
, loops
, 1,
1317 NULL
, NULL
, NULL
, NULL
, 0))
1320 /* After duplication entry edge now points to new loop head block.
1321 Note down new head as second_head. */
1322 second_head
= entry
->dest
;
1324 /* Split loop entry edge and insert new block with cond expr. */
1325 cond_bb
= lv_adjust_loop_entry_edge (first_head
, second_head
,
1328 *condition_bb
= cond_bb
;
1332 entry
->flags
|= irred_flag
;
1336 latch_edge
= single_succ_edge (get_bb_copy (loop
->latch
));
1338 extract_cond_bb_edges (cond_bb
, &true_edge
, &false_edge
);
1339 nloop
= loopify (loops
,
1341 single_pred_edge (get_bb_copy (loop
->header
)),
1342 cond_bb
, true_edge
, false_edge
,
1343 false /* Do not redirect all edges. */);
1345 exit
= loop
->single_exit
;
1347 nloop
->single_exit
= find_edge (get_bb_copy (exit
->src
), exit
->dest
);
1349 /* loopify redirected latch_edge. Update its PENDING_STMTS. */
1350 lv_flush_pending_stmts (latch_edge
);
1352 /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */
1353 extract_cond_bb_edges (cond_bb
, &true_edge
, &false_edge
);
1354 lv_flush_pending_stmts (false_edge
);
1355 /* Adjust irreducible flag. */
1358 cond_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
1359 loop_preheader_edge (loop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1360 loop_preheader_edge (nloop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1361 single_pred_edge (cond_bb
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1366 basic_block
*bbs
= get_loop_body_in_dom_order (nloop
), after
;
1369 after
= loop
->latch
;
1371 for (i
= 0; i
< nloop
->num_nodes
; i
++)
1373 move_block_after (bbs
[i
], after
);
1379 /* At this point condition_bb is loop predheader with two successors,
1380 first_head and second_head. Make sure that loop predheader has only
1382 loop_split_edge_with (loop_preheader_edge (loop
), NULL
);
1383 loop_split_edge_with (loop_preheader_edge (nloop
), NULL
);
1388 /* The structure of LOOPS might have changed. Some loops might get removed
1389 (and their headers and latches were set to NULL), loop exists might get
1390 removed (thus the loop nesting may be wrong), and some blocks and edges
1391 were changed (so the information about bb --> loop mapping does not have
1392 to be correct). But still for the remaining loops the header dominates
1393 the latch, and loops did not get new subloobs (new loops might possibly
1394 get created, but we are not interested in them). Fix up the mess.
1396 If CHANGED_BBS is not NULL, basic blocks whose loop has changed are
1400 fix_loop_structure (struct loops
*loops
, bitmap changed_bbs
)
1403 struct loop
*loop
, *ploop
;
1406 /* Remove the old bb -> loop mapping. */
1409 bb
->aux
= (void *) (size_t) bb
->loop_father
->depth
;
1410 bb
->loop_father
= loops
->tree_root
;
1413 /* Remove the dead loops from structures. */
1414 loops
->tree_root
->num_nodes
= n_basic_blocks
;
1415 for (i
= 1; i
< loops
->num
; i
++)
1417 loop
= loops
->parray
[i
];
1421 loop
->num_nodes
= 0;
1427 ploop
= loop
->inner
;
1428 flow_loop_tree_node_remove (ploop
);
1429 flow_loop_tree_node_add (loop
->outer
, ploop
);
1432 /* Remove the loop and free its data. */
1433 flow_loop_tree_node_remove (loop
);
1434 loops
->parray
[loop
->num
] = NULL
;
1435 flow_loop_free (loop
);
1438 /* Rescan the bodies of loops, starting from the outermost. */
1439 loop
= loops
->tree_root
;
1447 && loop
!= loops
->tree_root
)
1449 if (loop
== loops
->tree_root
)
1455 loop
->num_nodes
= flow_loop_nodes_find (loop
->header
, loop
);
1458 /* Now fix the loop nesting. */
1459 for (i
= 1; i
< loops
->num
; i
++)
1461 loop
= loops
->parray
[i
];
1465 bb
= loop_preheader_edge (loop
)->src
;
1466 if (bb
->loop_father
!= loop
->outer
)
1468 flow_loop_tree_node_remove (loop
);
1469 flow_loop_tree_node_add (bb
->loop_father
, loop
);
1473 /* Mark the blocks whose loop has changed. */
1477 && (void *) (size_t) bb
->loop_father
->depth
!= bb
->aux
)
1478 bitmap_set_bit (changed_bbs
, bb
->index
);
1483 if (loops
->state
& LOOPS_HAVE_MARKED_SINGLE_EXITS
)
1484 mark_single_exit_loops (loops
);
1485 if (loops
->state
& LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
)
1486 mark_irreducible_loops (loops
);