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 loop
*, struct loop
*);
34 static void copy_loops_to (struct loop
**, int,
36 static void loop_redirect_edge (edge
, basic_block
);
37 static void remove_bbs (basic_block
*, int);
38 static bool rpe_enum_p (const_basic_block
, const void *);
39 static int find_path (edge
, basic_block
**);
40 static void fix_loop_placements (struct loop
*, bool *);
41 static bool fix_bb_placement (basic_block
);
42 static void fix_bb_placements (basic_block
, bool *);
43 static void unloop (struct loop
*, bool *);
45 #define RDIV(X,Y) (((X) + (Y) / 2) / (Y))
47 /* Checks whether basic block BB is dominated by DATA. */
49 rpe_enum_p (const_basic_block bb
, const void *data
)
51 return dominated_by_p (CDI_DOMINATORS
, bb
, (const_basic_block
) data
);
54 /* Remove basic blocks BBS. NBBS is the number of the basic blocks. */
57 remove_bbs (basic_block
*bbs
, int nbbs
)
61 for (i
= 0; i
< nbbs
; i
++)
62 delete_basic_block (bbs
[i
]);
65 /* Find path -- i.e. the basic blocks dominated by edge E and put them
66 into array BBS, that will be allocated large enough to contain them.
67 E->dest must have exactly one predecessor for this to work (it is
68 easy to achieve and we do not put it here because we do not want to
69 alter anything by this function). The number of basic blocks in the
72 find_path (edge e
, basic_block
**bbs
)
74 gcc_assert (EDGE_COUNT (e
->dest
->preds
) <= 1);
76 /* Find bbs in the path. */
77 *bbs
= XCNEWVEC (basic_block
, n_basic_blocks
);
78 return dfs_enumerate_from (e
->dest
, 0, rpe_enum_p
, *bbs
,
79 n_basic_blocks
, e
->dest
);
82 /* Fix placement of basic block BB inside loop hierarchy --
83 Let L be a loop to that BB belongs. Then every successor of BB must either
84 1) belong to some superloop of loop L, or
85 2) be a header of loop K such that K->outer is superloop of L
86 Returns true if we had to move BB into other loop to enforce this condition,
87 false if the placement of BB was already correct (provided that placements
88 of its successors are correct). */
90 fix_bb_placement (basic_block bb
)
94 struct loop
*loop
= current_loops
->tree_root
, *act
;
96 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
98 if (e
->dest
== EXIT_BLOCK_PTR
)
101 act
= e
->dest
->loop_father
;
102 if (act
->header
== e
->dest
)
103 act
= loop_outer (act
);
105 if (flow_loop_nested_p (loop
, act
))
109 if (loop
== bb
->loop_father
)
112 remove_bb_from_loops (bb
);
113 add_bb_to_loop (bb
, loop
);
118 /* Fix placement of LOOP inside loop tree, i.e. find the innermost superloop
119 of LOOP to that leads at least one exit edge of LOOP, and set it
120 as the immediate superloop of LOOP. Return true if the immediate superloop
124 fix_loop_placement (struct loop
*loop
)
128 VEC (edge
, heap
) *exits
= get_loop_exit_edges (loop
);
129 struct loop
*father
= current_loops
->tree_root
, *act
;
132 for (i
= 0; VEC_iterate (edge
, exits
, i
, e
); i
++)
134 act
= find_common_loop (loop
, e
->dest
->loop_father
);
135 if (flow_loop_nested_p (father
, act
))
139 if (father
!= loop_outer (loop
))
141 for (act
= loop_outer (loop
); act
!= father
; act
= loop_outer (act
))
142 act
->num_nodes
-= loop
->num_nodes
;
143 flow_loop_tree_node_remove (loop
);
144 flow_loop_tree_node_add (father
, loop
);
146 /* The exit edges of LOOP no longer exits its original immediate
147 superloops; remove them from the appropriate exit lists. */
148 for (i
= 0; VEC_iterate (edge
, exits
, i
, e
); i
++)
149 rescan_loop_exit (e
, false, false);
154 VEC_free (edge
, heap
, exits
);
158 /* Fix placements of basic blocks inside loop hierarchy stored in loops; i.e.
159 enforce condition condition stated in description of fix_bb_placement. We
160 start from basic block FROM that had some of its successors removed, so that
161 his placement no longer has to be correct, and iteratively fix placement of
162 its predecessors that may change if placement of FROM changed. Also fix
163 placement of subloops of FROM->loop_father, that might also be altered due
164 to this change; the condition for them is similar, except that instead of
165 successors we consider edges coming out of the loops.
167 If the changes may invalidate the information about irreducible regions,
168 IRRED_INVALIDATED is set to true. */
171 fix_bb_placements (basic_block from
,
172 bool *irred_invalidated
)
175 basic_block
*queue
, *qtop
, *qbeg
, *qend
;
176 struct loop
*base_loop
;
179 /* We pass through blocks back-reachable from FROM, testing whether some
180 of their successors moved to outer loop. It may be necessary to
181 iterate several times, but it is finite, as we stop unless we move
182 the basic block up the loop structure. The whole story is a bit
183 more complicated due to presence of subloops, those are moved using
184 fix_loop_placement. */
186 base_loop
= from
->loop_father
;
187 if (base_loop
== current_loops
->tree_root
)
190 in_queue
= sbitmap_alloc (last_basic_block
);
191 sbitmap_zero (in_queue
);
192 SET_BIT (in_queue
, from
->index
);
193 /* Prevent us from going out of the base_loop. */
194 SET_BIT (in_queue
, base_loop
->header
->index
);
196 queue
= XNEWVEC (basic_block
, base_loop
->num_nodes
+ 1);
197 qtop
= queue
+ base_loop
->num_nodes
+ 1;
209 RESET_BIT (in_queue
, from
->index
);
211 if (from
->loop_father
->header
== from
)
213 /* Subloop header, maybe move the loop upward. */
214 if (!fix_loop_placement (from
->loop_father
))
219 /* Ordinary basic block. */
220 if (!fix_bb_placement (from
))
224 FOR_EACH_EDGE (e
, ei
, from
->succs
)
226 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
227 *irred_invalidated
= true;
230 /* Something has changed, insert predecessors into queue. */
231 FOR_EACH_EDGE (e
, ei
, from
->preds
)
233 basic_block pred
= e
->src
;
236 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
237 *irred_invalidated
= true;
239 if (TEST_BIT (in_queue
, pred
->index
))
242 /* If it is subloop, then it either was not moved, or
243 the path up the loop tree from base_loop do not contain
245 nca
= find_common_loop (pred
->loop_father
, base_loop
);
246 if (pred
->loop_father
!= base_loop
248 || nca
!= pred
->loop_father
))
249 pred
= pred
->loop_father
->header
;
250 else if (!flow_loop_nested_p (from
->loop_father
, pred
->loop_father
))
252 /* No point in processing it. */
256 if (TEST_BIT (in_queue
, pred
->index
))
259 /* Schedule the basic block. */
264 SET_BIT (in_queue
, pred
->index
);
271 /* Removes path beginning at edge E, i.e. remove basic blocks dominated by E
272 and update loop structures and dominators. Return true if we were able
273 to remove the path, false otherwise (and nothing is affected then). */
278 basic_block
*rem_bbs
, *bord_bbs
, from
, bb
;
279 VEC (basic_block
, heap
) *dom_bbs
;
280 int i
, nrem
, n_bord_bbs
, nreml
;
282 bool irred_invalidated
= false;
283 struct loop
**deleted_loop
;
285 if (!can_remove_branch_p (e
))
288 /* Keep track of whether we need to update information about irreducible
289 regions. This is the case if the removed area is a part of the
290 irreducible region, or if the set of basic blocks that belong to a loop
291 that is inside an irreducible region is changed, or if such a loop is
293 if (e
->flags
& EDGE_IRREDUCIBLE_LOOP
)
294 irred_invalidated
= true;
296 /* We need to check whether basic blocks are dominated by the edge
297 e, but we only have basic block dominators. This is easy to
298 fix -- when e->dest has exactly one predecessor, this corresponds
299 to blocks dominated by e->dest, if not, split the edge. */
300 if (!single_pred_p (e
->dest
))
301 e
= single_pred_edge (split_edge (e
));
303 /* It may happen that by removing path we remove one or more loops
304 we belong to. In this case first unloop the loops, then proceed
305 normally. We may assume that e->dest is not a header of any loop,
306 as it now has exactly one predecessor. */
307 while (loop_outer (e
->src
->loop_father
)
308 && dominated_by_p (CDI_DOMINATORS
,
309 e
->src
->loop_father
->latch
, e
->dest
))
310 unloop (e
->src
->loop_father
, &irred_invalidated
);
312 /* Identify the path. */
313 nrem
= find_path (e
, &rem_bbs
);
316 bord_bbs
= XCNEWVEC (basic_block
, n_basic_blocks
);
317 seen
= sbitmap_alloc (last_basic_block
);
320 /* Find "border" hexes -- i.e. those with predecessor in removed path. */
321 for (i
= 0; i
< nrem
; i
++)
322 SET_BIT (seen
, rem_bbs
[i
]->index
);
323 for (i
= 0; i
< nrem
; i
++)
327 FOR_EACH_EDGE (ae
, ei
, rem_bbs
[i
]->succs
)
328 if (ae
->dest
!= EXIT_BLOCK_PTR
&& !TEST_BIT (seen
, ae
->dest
->index
))
330 SET_BIT (seen
, ae
->dest
->index
);
331 bord_bbs
[n_bord_bbs
++] = ae
->dest
;
333 if (ae
->flags
& EDGE_IRREDUCIBLE_LOOP
)
334 irred_invalidated
= true;
338 /* Remove the path. */
343 /* Cancel loops contained in the path. */
344 deleted_loop
= XNEWVEC (struct loop
*, nrem
);
346 for (i
= 0; i
< nrem
; i
++)
347 if (rem_bbs
[i
]->loop_father
->header
== rem_bbs
[i
])
348 deleted_loop
[nreml
++] = rem_bbs
[i
]->loop_father
;
350 remove_bbs (rem_bbs
, nrem
);
353 for (i
= 0; i
< nreml
; i
++)
354 cancel_loop_tree (deleted_loop
[i
]);
357 /* Find blocks whose dominators may be affected. */
359 for (i
= 0; i
< n_bord_bbs
; i
++)
363 bb
= get_immediate_dominator (CDI_DOMINATORS
, bord_bbs
[i
]);
364 if (TEST_BIT (seen
, bb
->index
))
366 SET_BIT (seen
, bb
->index
);
368 for (ldom
= first_dom_son (CDI_DOMINATORS
, bb
);
370 ldom
= next_dom_son (CDI_DOMINATORS
, ldom
))
371 if (!dominated_by_p (CDI_DOMINATORS
, from
, ldom
))
372 VEC_safe_push (basic_block
, heap
, dom_bbs
, ldom
);
377 /* Recount dominators. */
378 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, true);
379 VEC_free (basic_block
, heap
, dom_bbs
);
382 /* Fix placements of basic blocks inside loops and the placement of
383 loops in the loop tree. */
384 fix_bb_placements (from
, &irred_invalidated
);
385 fix_loop_placements (from
->loop_father
, &irred_invalidated
);
387 if (irred_invalidated
388 && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
))
389 mark_irreducible_loops ();
394 /* Creates place for a new LOOP in loops structure. */
397 place_new_loop (struct loop
*loop
)
399 loop
->num
= number_of_loops ();
400 VEC_safe_push (loop_p
, gc
, current_loops
->larray
, loop
);
403 /* Given LOOP structure with filled header and latch, find the body of the
404 corresponding loop and add it to loops tree. Insert the LOOP as a son of
408 add_loop (struct loop
*loop
, struct loop
*outer
)
412 struct loop
*subloop
;
416 /* Add it to loop structure. */
417 place_new_loop (loop
);
418 flow_loop_tree_node_add (outer
, loop
);
420 /* Find its nodes. */
421 bbs
= XNEWVEC (basic_block
, n_basic_blocks
);
422 n
= get_loop_body_with_size (loop
, bbs
, n_basic_blocks
);
424 for (i
= 0; i
< n
; i
++)
426 if (bbs
[i
]->loop_father
== outer
)
428 remove_bb_from_loops (bbs
[i
]);
429 add_bb_to_loop (bbs
[i
], loop
);
435 /* If we find a direct subloop of OUTER, move it to LOOP. */
436 subloop
= bbs
[i
]->loop_father
;
437 if (loop_outer (subloop
) == outer
438 && subloop
->header
== bbs
[i
])
440 flow_loop_tree_node_remove (subloop
);
441 flow_loop_tree_node_add (loop
, subloop
);
445 /* Update the information about loop exit edges. */
446 for (i
= 0; i
< n
; i
++)
448 FOR_EACH_EDGE (e
, ei
, bbs
[i
]->succs
)
450 rescan_loop_exit (e
, false, false);
457 /* Multiply all frequencies in LOOP by NUM/DEN. */
459 scale_loop_frequencies (struct loop
*loop
, int num
, int den
)
463 bbs
= get_loop_body (loop
);
464 scale_bbs_frequencies_int (bbs
, loop
->num_nodes
, num
, den
);
468 /* Make area between HEADER_EDGE and LATCH_EDGE a loop by connecting
469 latch to header and update loop tree and dominators
470 accordingly. Everything between them plus LATCH_EDGE destination must
471 be dominated by HEADER_EDGE destination, and back-reachable from
472 LATCH_EDGE source. HEADER_EDGE is redirected to basic block SWITCH_BB,
473 FALSE_EDGE of SWITCH_BB to original destination of HEADER_EDGE and
474 TRUE_EDGE of SWITCH_BB to original destination of LATCH_EDGE.
475 Returns the newly created loop. Frequencies and counts in the new loop
476 are scaled by FALSE_SCALE and in the old one by TRUE_SCALE. */
479 loopify (edge latch_edge
, edge header_edge
,
480 basic_block switch_bb
, edge true_edge
, edge false_edge
,
481 bool redirect_all_edges
, unsigned true_scale
, unsigned false_scale
)
483 basic_block succ_bb
= latch_edge
->dest
;
484 basic_block pred_bb
= header_edge
->src
;
486 VEC (basic_block
, heap
) *dom_bbs
;
489 struct loop
*loop
= alloc_loop ();
490 struct loop
*outer
= loop_outer (succ_bb
->loop_father
);
496 loop
->header
= header_edge
->dest
;
497 loop
->latch
= latch_edge
->src
;
499 freq
= EDGE_FREQUENCY (header_edge
);
500 cnt
= header_edge
->count
;
502 /* Redirect edges. */
503 loop_redirect_edge (latch_edge
, loop
->header
);
504 loop_redirect_edge (true_edge
, succ_bb
);
506 /* During loop versioning, one of the switch_bb edge is already properly
507 set. Do not redirect it again unless redirect_all_edges is true. */
508 if (redirect_all_edges
)
510 loop_redirect_edge (header_edge
, switch_bb
);
511 loop_redirect_edge (false_edge
, loop
->header
);
513 /* Update dominators. */
514 set_immediate_dominator (CDI_DOMINATORS
, switch_bb
, pred_bb
);
515 set_immediate_dominator (CDI_DOMINATORS
, loop
->header
, switch_bb
);
518 set_immediate_dominator (CDI_DOMINATORS
, succ_bb
, switch_bb
);
520 /* Compute new loop. */
521 add_loop (loop
, outer
);
523 /* Add switch_bb to appropriate loop. */
524 if (switch_bb
->loop_father
)
525 remove_bb_from_loops (switch_bb
);
526 add_bb_to_loop (switch_bb
, outer
);
528 /* Fix frequencies. */
529 if (redirect_all_edges
)
531 switch_bb
->frequency
= freq
;
532 switch_bb
->count
= cnt
;
533 FOR_EACH_EDGE (e
, ei
, switch_bb
->succs
)
535 e
->count
= (switch_bb
->count
* e
->probability
) / REG_BR_PROB_BASE
;
538 scale_loop_frequencies (loop
, false_scale
, REG_BR_PROB_BASE
);
539 scale_loop_frequencies (succ_bb
->loop_father
, true_scale
, REG_BR_PROB_BASE
);
541 /* Update dominators of blocks outside of LOOP. */
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 VEC_safe_push (basic_block
, heap
, dom_bbs
, ldom
);
564 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, false);
568 VEC_free (basic_block
, heap
, dom_bbs
);
573 /* Remove the latch edge of a LOOP and update loops 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 If this may cause the information about irreducible regions to become
578 invalid, IRRED_INVALIDATED is set to true. */
581 unloop (struct loop
*loop
, bool *irred_invalidated
)
586 basic_block latch
= loop
->latch
;
589 if (loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
)
590 *irred_invalidated
= true;
592 /* This is relatively straightforward. The dominators are unchanged, as
593 loop header dominates loop latch, so the only thing we have to care of
594 is the placement of loops and basic blocks inside the loop tree. We
595 move them all to the loop->outer, and then let fix_bb_placements do
598 body
= get_loop_body (loop
);
600 for (i
= 0; i
< n
; i
++)
601 if (body
[i
]->loop_father
== loop
)
603 remove_bb_from_loops (body
[i
]);
604 add_bb_to_loop (body
[i
], loop_outer (loop
));
611 flow_loop_tree_node_remove (ploop
);
612 flow_loop_tree_node_add (loop_outer (loop
), ploop
);
615 /* Remove the loop and free its data. */
618 remove_edge (single_succ_edge (latch
));
620 /* We do not pass IRRED_INVALIDATED to fix_bb_placements here, as even if
621 there is an irreducible region inside the cancelled loop, the flags will
623 fix_bb_placements (latch
, &dummy
);
626 /* Fix placement of superloops of LOOP inside loop tree, i.e. ensure that
627 condition stated in description of fix_loop_placement holds for them.
628 It is used in case when we removed some edges coming out of LOOP, which
629 may cause the right placement of LOOP inside loop tree to change.
631 IRRED_INVALIDATED is set to true if a change in the loop structures might
632 invalidate the information about irreducible regions. */
635 fix_loop_placements (struct loop
*loop
, bool *irred_invalidated
)
639 while (loop_outer (loop
))
641 outer
= loop_outer (loop
);
642 if (!fix_loop_placement (loop
))
645 /* Changing the placement of a loop in the loop tree may alter the
646 validity of condition 2) of the description of fix_bb_placement
647 for its preheader, because the successor is the header and belongs
648 to the loop. So call fix_bb_placements to fix up the placement
649 of the preheader and (possibly) of its predecessors. */
650 fix_bb_placements (loop_preheader_edge (loop
)->src
,
656 /* Copies copy of LOOP as subloop of TARGET loop, placing newly
657 created loop into loops structure. */
659 duplicate_loop (struct loop
*loop
, struct loop
*target
)
662 cloop
= alloc_loop ();
663 place_new_loop (cloop
);
665 /* Mark the new loop as copy of LOOP. */
666 set_loop_copy (loop
, cloop
);
668 /* Add it to target. */
669 flow_loop_tree_node_add (target
, cloop
);
674 /* Copies structure of subloops of LOOP into TARGET loop, placing
675 newly created loops into loop tree. */
677 duplicate_subloops (struct loop
*loop
, struct loop
*target
)
679 struct loop
*aloop
, *cloop
;
681 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
683 cloop
= duplicate_loop (aloop
, target
);
684 duplicate_subloops (aloop
, cloop
);
688 /* Copies structure of subloops of N loops, stored in array COPIED_LOOPS,
689 into TARGET loop, placing newly created loops into loop tree. */
691 copy_loops_to (struct loop
**copied_loops
, int n
, struct loop
*target
)
696 for (i
= 0; i
< n
; i
++)
698 aloop
= duplicate_loop (copied_loops
[i
], target
);
699 duplicate_subloops (copied_loops
[i
], aloop
);
703 /* Redirects edge E to basic block DEST. */
705 loop_redirect_edge (edge e
, basic_block dest
)
710 redirect_edge_and_branch_force (e
, dest
);
713 /* Check whether LOOP's body can be duplicated. */
715 can_duplicate_loop_p (const struct loop
*loop
)
718 basic_block
*bbs
= get_loop_body (loop
);
720 ret
= can_copy_bbs_p (bbs
, loop
->num_nodes
);
726 /* Sets probability and count of edge E to zero. The probability and count
727 is redistributed evenly to the remaining edges coming from E->src. */
730 set_zero_probability (edge e
)
732 basic_block bb
= e
->src
;
734 edge ae
, last
= NULL
;
735 unsigned n
= EDGE_COUNT (bb
->succs
);
736 gcov_type cnt
= e
->count
, cnt1
;
737 unsigned prob
= e
->probability
, prob1
;
740 cnt1
= cnt
/ (n
- 1);
741 prob1
= prob
/ (n
- 1);
743 FOR_EACH_EDGE (ae
, ei
, bb
->succs
)
748 ae
->probability
+= prob1
;
753 /* Move the rest to one of the edges. */
754 last
->probability
+= prob
% (n
- 1);
755 last
->count
+= cnt
% (n
- 1);
761 /* Duplicates body of LOOP to given edge E NDUPL times. Takes care of updating
762 loop structure and dominators. E's destination must be LOOP header for
763 this to work, i.e. it must be entry or latch edge of this loop; these are
764 unique, as the loops must have preheaders for this function to work
765 correctly (in case E is latch, the function unrolls the loop, if E is entry
766 edge, it peels the loop). Store edges created by copying ORIG edge from
767 copies corresponding to set bits in WONT_EXIT bitmap (bit 0 corresponds to
768 original LOOP body, the other copies are numbered in order given by control
769 flow through them) into TO_REMOVE array. Returns false if duplication is
773 duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
774 unsigned int ndupl
, sbitmap wont_exit
,
775 edge orig
, VEC (edge
, heap
) **to_remove
,
778 struct loop
*target
, *aloop
;
779 struct loop
**orig_loops
;
780 unsigned n_orig_loops
;
781 basic_block header
= loop
->header
, latch
= loop
->latch
;
782 basic_block
*new_bbs
, *bbs
, *first_active
;
783 basic_block new_bb
, bb
, first_active_latch
= NULL
;
785 edge spec_edges
[2], new_spec_edges
[2];
789 int is_latch
= (latch
== e
->src
);
790 int scale_act
= 0, *scale_step
= NULL
, scale_main
= 0;
791 int scale_after_exit
= 0;
792 int p
, freq_in
, freq_le
, freq_out_orig
;
793 int prob_pass_thru
, prob_pass_wont_exit
, prob_pass_main
;
794 int add_irreducible_flag
;
795 basic_block place_after
;
796 bitmap bbs_to_scale
= NULL
;
799 gcc_assert (e
->dest
== loop
->header
);
800 gcc_assert (ndupl
> 0);
804 /* Orig must be edge out of the loop. */
805 gcc_assert (flow_bb_inside_loop_p (loop
, orig
->src
));
806 gcc_assert (!flow_bb_inside_loop_p (loop
, orig
->dest
));
810 bbs
= get_loop_body_in_dom_order (loop
);
811 gcc_assert (bbs
[0] == loop
->header
);
812 gcc_assert (bbs
[n
- 1] == loop
->latch
);
814 /* Check whether duplication is possible. */
815 if (!can_copy_bbs_p (bbs
, loop
->num_nodes
))
820 new_bbs
= XNEWVEC (basic_block
, loop
->num_nodes
);
822 /* In case we are doing loop peeling and the loop is in the middle of
823 irreducible region, the peeled copies will be inside it too. */
824 add_irreducible_flag
= e
->flags
& EDGE_IRREDUCIBLE_LOOP
;
825 gcc_assert (!is_latch
|| !add_irreducible_flag
);
827 /* Find edge from latch. */
828 latch_edge
= loop_latch_edge (loop
);
830 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
832 /* Calculate coefficients by that we have to scale frequencies
833 of duplicated loop bodies. */
834 freq_in
= header
->frequency
;
835 freq_le
= EDGE_FREQUENCY (latch_edge
);
838 if (freq_in
< freq_le
)
840 freq_out_orig
= orig
? EDGE_FREQUENCY (orig
) : freq_in
- freq_le
;
841 if (freq_out_orig
> freq_in
- freq_le
)
842 freq_out_orig
= freq_in
- freq_le
;
843 prob_pass_thru
= RDIV (REG_BR_PROB_BASE
* freq_le
, freq_in
);
844 prob_pass_wont_exit
=
845 RDIV (REG_BR_PROB_BASE
* (freq_le
+ freq_out_orig
), freq_in
);
848 && REG_BR_PROB_BASE
- orig
->probability
!= 0)
850 /* The blocks that are dominated by a removed exit edge ORIG have
851 frequencies scaled by this. */
852 scale_after_exit
= RDIV (REG_BR_PROB_BASE
* REG_BR_PROB_BASE
,
853 REG_BR_PROB_BASE
- orig
->probability
);
854 bbs_to_scale
= BITMAP_ALLOC (NULL
);
855 for (i
= 0; i
< n
; i
++)
857 if (bbs
[i
] != orig
->src
858 && dominated_by_p (CDI_DOMINATORS
, bbs
[i
], orig
->src
))
859 bitmap_set_bit (bbs_to_scale
, i
);
863 scale_step
= XNEWVEC (int, ndupl
);
865 for (i
= 1; i
<= ndupl
; i
++)
866 scale_step
[i
- 1] = TEST_BIT (wont_exit
, i
)
867 ? prob_pass_wont_exit
870 /* Complete peeling is special as the probability of exit in last
872 if (flags
& DLTHE_FLAG_COMPLETTE_PEEL
)
874 int wanted_freq
= EDGE_FREQUENCY (e
);
876 if (wanted_freq
> freq_in
)
877 wanted_freq
= freq_in
;
879 gcc_assert (!is_latch
);
880 /* First copy has frequency of incoming edge. Each subsequent
881 frequency should be reduced by prob_pass_wont_exit. Caller
882 should've managed the flags so all except for original loop
883 has won't exist set. */
884 scale_act
= RDIV (wanted_freq
* REG_BR_PROB_BASE
, freq_in
);
885 /* Now simulate the duplication adjustments and compute header
886 frequency of the last copy. */
887 for (i
= 0; i
< ndupl
; i
++)
888 wanted_freq
= RDIV (wanted_freq
* scale_step
[i
], REG_BR_PROB_BASE
);
889 scale_main
= RDIV (wanted_freq
* REG_BR_PROB_BASE
, freq_in
);
893 prob_pass_main
= TEST_BIT (wont_exit
, 0)
894 ? prob_pass_wont_exit
897 scale_main
= REG_BR_PROB_BASE
;
898 for (i
= 0; i
< ndupl
; i
++)
901 p
= RDIV (p
* scale_step
[i
], REG_BR_PROB_BASE
);
903 scale_main
= RDIV (REG_BR_PROB_BASE
* REG_BR_PROB_BASE
, scale_main
);
904 scale_act
= RDIV (scale_main
* prob_pass_main
, REG_BR_PROB_BASE
);
908 scale_main
= REG_BR_PROB_BASE
;
909 for (i
= 0; i
< ndupl
; i
++)
910 scale_main
= RDIV (scale_main
* scale_step
[i
], REG_BR_PROB_BASE
);
911 scale_act
= REG_BR_PROB_BASE
- prob_pass_thru
;
913 for (i
= 0; i
< ndupl
; i
++)
914 gcc_assert (scale_step
[i
] >= 0 && scale_step
[i
] <= REG_BR_PROB_BASE
);
915 gcc_assert (scale_main
>= 0 && scale_main
<= REG_BR_PROB_BASE
916 && scale_act
>= 0 && scale_act
<= REG_BR_PROB_BASE
);
919 /* Loop the new bbs will belong to. */
920 target
= e
->src
->loop_father
;
922 /* Original loops. */
924 for (aloop
= loop
->inner
; aloop
; aloop
= aloop
->next
)
926 orig_loops
= XCNEWVEC (struct loop
*, n_orig_loops
);
927 for (aloop
= loop
->inner
, i
= 0; aloop
; aloop
= aloop
->next
, i
++)
928 orig_loops
[i
] = aloop
;
930 set_loop_copy (loop
, target
);
932 first_active
= XNEWVEC (basic_block
, n
);
935 memcpy (first_active
, bbs
, n
* sizeof (basic_block
));
936 first_active_latch
= latch
;
939 spec_edges
[SE_ORIG
] = orig
;
940 spec_edges
[SE_LATCH
] = latch_edge
;
942 place_after
= e
->src
;
943 for (j
= 0; j
< ndupl
; j
++)
946 copy_loops_to (orig_loops
, n_orig_loops
, target
);
949 copy_bbs (bbs
, n
, new_bbs
, spec_edges
, 2, new_spec_edges
, loop
,
951 place_after
= new_spec_edges
[SE_LATCH
]->src
;
953 if (flags
& DLTHE_RECORD_COPY_NUMBER
)
954 for (i
= 0; i
< n
; i
++)
956 gcc_assert (!new_bbs
[i
]->aux
);
957 new_bbs
[i
]->aux
= (void *)(size_t)(j
+ 1);
960 /* Note whether the blocks and edges belong to an irreducible loop. */
961 if (add_irreducible_flag
)
963 for (i
= 0; i
< n
; i
++)
964 new_bbs
[i
]->flags
|= BB_DUPLICATED
;
965 for (i
= 0; i
< n
; i
++)
969 if (new_bb
->loop_father
== target
)
970 new_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
972 FOR_EACH_EDGE (ae
, ei
, new_bb
->succs
)
973 if ((ae
->dest
->flags
& BB_DUPLICATED
)
974 && (ae
->src
->loop_father
== target
975 || ae
->dest
->loop_father
== target
))
976 ae
->flags
|= EDGE_IRREDUCIBLE_LOOP
;
978 for (i
= 0; i
< n
; i
++)
979 new_bbs
[i
]->flags
&= ~BB_DUPLICATED
;
982 /* Redirect the special edges. */
985 redirect_edge_and_branch_force (latch_edge
, new_bbs
[0]);
986 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
988 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], latch
);
989 latch
= loop
->latch
= new_bbs
[n
- 1];
990 e
= latch_edge
= new_spec_edges
[SE_LATCH
];
994 redirect_edge_and_branch_force (new_spec_edges
[SE_LATCH
],
996 redirect_edge_and_branch_force (e
, new_bbs
[0]);
997 set_immediate_dominator (CDI_DOMINATORS
, new_bbs
[0], e
->src
);
998 e
= new_spec_edges
[SE_LATCH
];
1001 /* Record exit edge in this copy. */
1002 if (orig
&& TEST_BIT (wont_exit
, j
+ 1))
1005 VEC_safe_push (edge
, heap
, *to_remove
, new_spec_edges
[SE_ORIG
]);
1006 set_zero_probability (new_spec_edges
[SE_ORIG
]);
1008 /* Scale the frequencies of the blocks dominated by the exit. */
1011 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale
, 0, i
, bi
)
1013 scale_bbs_frequencies_int (new_bbs
+ i
, 1, scale_after_exit
,
1019 /* Record the first copy in the control flow order if it is not
1020 the original loop (i.e. in case of peeling). */
1021 if (!first_active_latch
)
1023 memcpy (first_active
, new_bbs
, n
* sizeof (basic_block
));
1024 first_active_latch
= new_bbs
[n
- 1];
1027 /* Set counts and frequencies. */
1028 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1030 scale_bbs_frequencies_int (new_bbs
, n
, scale_act
, REG_BR_PROB_BASE
);
1031 scale_act
= RDIV (scale_act
* scale_step
[j
], REG_BR_PROB_BASE
);
1037 /* Record the exit edge in the original loop body, and update the frequencies. */
1038 if (orig
&& TEST_BIT (wont_exit
, 0))
1041 VEC_safe_push (edge
, heap
, *to_remove
, orig
);
1042 set_zero_probability (orig
);
1044 /* Scale the frequencies of the blocks dominated by the exit. */
1047 EXECUTE_IF_SET_IN_BITMAP (bbs_to_scale
, 0, i
, bi
)
1049 scale_bbs_frequencies_int (bbs
+ i
, 1, scale_after_exit
,
1055 /* Update the original loop. */
1057 set_immediate_dominator (CDI_DOMINATORS
, e
->dest
, e
->src
);
1058 if (flags
& DLTHE_FLAG_UPDATE_FREQ
)
1060 scale_bbs_frequencies_int (bbs
, n
, scale_main
, REG_BR_PROB_BASE
);
1064 /* Update dominators of outer blocks if affected. */
1065 for (i
= 0; i
< n
; i
++)
1067 basic_block dominated
, dom_bb
;
1068 VEC (basic_block
, heap
) *dom_bbs
;
1074 dom_bbs
= get_dominated_by (CDI_DOMINATORS
, bb
);
1075 for (j
= 0; VEC_iterate (basic_block
, dom_bbs
, j
, dominated
); j
++)
1077 if (flow_bb_inside_loop_p (loop
, dominated
))
1079 dom_bb
= nearest_common_dominator (
1080 CDI_DOMINATORS
, first_active
[i
], first_active_latch
);
1081 set_immediate_dominator (CDI_DOMINATORS
, dominated
, dom_bb
);
1083 VEC_free (basic_block
, heap
, dom_bbs
);
1085 free (first_active
);
1088 BITMAP_FREE (bbs_to_scale
);
1093 /* A callback for make_forwarder block, to redirect all edges except for
1094 MFB_KJ_EDGE to the entry part. E is the edge for that we should decide
1095 whether to redirect it. */
1099 mfb_keep_just (edge e
)
1101 return e
!= mfb_kj_edge
;
1104 /* Creates a pre-header for a LOOP. Returns newly created block. Unless
1105 CP_SIMPLE_PREHEADERS is set in FLAGS, we only force LOOP to have single
1106 entry; otherwise we also force preheader block to have only one successor.
1107 The function also updates dominators. */
1110 create_preheader (struct loop
*loop
, int flags
)
1116 bool latch_edge_was_fallthru
;
1117 edge one_succ_pred
= NULL
, single_entry
= NULL
;
1120 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
1122 if (e
->src
== loop
->latch
)
1124 irred
|= (e
->flags
& EDGE_IRREDUCIBLE_LOOP
) != 0;
1127 if (single_succ_p (e
->src
))
1130 gcc_assert (nentry
);
1133 if (/* We do not allow entry block to be the loop preheader, since we
1134 cannot emit code there. */
1135 single_entry
->src
!= ENTRY_BLOCK_PTR
1136 /* If we want simple preheaders, also force the preheader to have
1137 just a single successor. */
1138 && !((flags
& CP_SIMPLE_PREHEADERS
)
1139 && !single_succ_p (single_entry
->src
)))
1143 mfb_kj_edge
= loop_latch_edge (loop
);
1144 latch_edge_was_fallthru
= (mfb_kj_edge
->flags
& EDGE_FALLTHRU
) != 0;
1145 fallthru
= make_forwarder_block (loop
->header
, mfb_keep_just
, NULL
);
1146 dummy
= fallthru
->src
;
1147 loop
->header
= fallthru
->dest
;
1149 /* Try to be clever in placing the newly created preheader. The idea is to
1150 avoid breaking any "fallthruness" relationship between blocks.
1152 The preheader was created just before the header and all incoming edges
1153 to the header were redirected to the preheader, except the latch edge.
1154 So the only problematic case is when this latch edge was a fallthru
1155 edge: it is not anymore after the preheader creation so we have broken
1156 the fallthruness. We're therefore going to look for a better place. */
1157 if (latch_edge_was_fallthru
)
1162 e
= EDGE_PRED (dummy
, 0);
1164 move_block_after (dummy
, e
->src
);
1169 dummy
->flags
|= BB_IRREDUCIBLE_LOOP
;
1170 single_succ_edge (dummy
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1174 fprintf (dump_file
, "Created preheader block for loop %i\n",
1180 /* Create preheaders for each loop; for meaning of FLAGS see create_preheader. */
1183 create_preheaders (int flags
)
1191 FOR_EACH_LOOP (li
, loop
, 0)
1192 create_preheader (loop
, flags
);
1193 loops_state_set (LOOPS_HAVE_PREHEADERS
);
1196 /* Forces all loop latches to have only single successor. */
1199 force_single_succ_latches (void)
1205 FOR_EACH_LOOP (li
, loop
, 0)
1207 if (loop
->latch
!= loop
->header
&& single_succ_p (loop
->latch
))
1210 e
= find_edge (loop
->latch
, loop
->header
);
1214 loops_state_set (LOOPS_HAVE_SIMPLE_LATCHES
);
1217 /* This function is called from loop_version. It splits the entry edge
1218 of the loop we want to version, adds the versioning condition, and
1219 adjust the edges to the two versions of the loop appropriately.
1220 e is an incoming edge. Returns the basic block containing the
1223 --- edge e ---- > [second_head]
1225 Split it and insert new conditional expression and adjust edges.
1227 --- edge e ---> [cond expr] ---> [first_head]
1229 +---------> [second_head]
1231 THEN_PROB is the probability of then branch of the condition. */
1234 lv_adjust_loop_entry_edge (basic_block first_head
, basic_block second_head
,
1235 edge e
, void *cond_expr
, unsigned then_prob
)
1237 basic_block new_head
= NULL
;
1240 gcc_assert (e
->dest
== second_head
);
1242 /* Split edge 'e'. This will create a new basic block, where we can
1243 insert conditional expr. */
1244 new_head
= split_edge (e
);
1246 lv_add_condition_to_bb (first_head
, second_head
, new_head
,
1249 /* Don't set EDGE_TRUE_VALUE in RTL mode, as it's invalid there. */
1250 e
= single_succ_edge (new_head
);
1251 e1
= make_edge (new_head
, first_head
,
1252 current_ir_type () == IR_GIMPLE
? EDGE_TRUE_VALUE
: 0);
1253 e1
->probability
= then_prob
;
1254 e
->probability
= REG_BR_PROB_BASE
- then_prob
;
1255 e1
->count
= RDIV (e
->count
* e1
->probability
, REG_BR_PROB_BASE
);
1256 e
->count
= RDIV (e
->count
* e
->probability
, REG_BR_PROB_BASE
);
1258 set_immediate_dominator (CDI_DOMINATORS
, first_head
, new_head
);
1259 set_immediate_dominator (CDI_DOMINATORS
, second_head
, new_head
);
1261 /* Adjust loop header phi nodes. */
1262 lv_adjust_loop_header_phi (first_head
, second_head
, new_head
, e1
);
1267 /* Main entry point for Loop Versioning transformation.
1269 This transformation given a condition and a loop, creates
1270 -if (condition) { loop_copy1 } else { loop_copy2 },
1271 where loop_copy1 is the loop transformed in one way, and loop_copy2
1272 is the loop transformed in another way (or unchanged). 'condition'
1273 may be a run time test for things that were not resolved by static
1274 analysis (overlapping ranges (anti-aliasing), alignment, etc.).
1276 THEN_PROB is the probability of the then edge of the if. THEN_SCALE
1277 is the ratio by that the frequencies in the original loop should
1278 be scaled. ELSE_SCALE is the ratio by that the frequencies in the
1279 new loop should be scaled.
1281 If PLACE_AFTER is true, we place the new loop after LOOP in the
1282 instruction stream, otherwise it is placed before LOOP. */
1285 loop_version (struct loop
*loop
,
1286 void *cond_expr
, basic_block
*condition_bb
,
1287 unsigned then_prob
, unsigned then_scale
, unsigned else_scale
,
1290 basic_block first_head
, second_head
;
1291 edge entry
, latch_edge
, true_edge
, false_edge
;
1294 basic_block cond_bb
;
1296 /* Record entry and latch edges for the loop */
1297 entry
= loop_preheader_edge (loop
);
1298 irred_flag
= entry
->flags
& EDGE_IRREDUCIBLE_LOOP
;
1299 entry
->flags
&= ~EDGE_IRREDUCIBLE_LOOP
;
1301 /* Note down head of loop as first_head. */
1302 first_head
= entry
->dest
;
1304 /* Duplicate loop. */
1305 if (!cfg_hook_duplicate_loop_to_header_edge (loop
, entry
, 1,
1306 NULL
, NULL
, NULL
, 0))
1309 /* After duplication entry edge now points to new loop head block.
1310 Note down new head as second_head. */
1311 second_head
= entry
->dest
;
1313 /* Split loop entry edge and insert new block with cond expr. */
1314 cond_bb
= lv_adjust_loop_entry_edge (first_head
, second_head
,
1315 entry
, cond_expr
, then_prob
);
1317 *condition_bb
= cond_bb
;
1321 entry
->flags
|= irred_flag
;
1325 latch_edge
= single_succ_edge (get_bb_copy (loop
->latch
));
1327 extract_cond_bb_edges (cond_bb
, &true_edge
, &false_edge
);
1328 nloop
= loopify (latch_edge
,
1329 single_pred_edge (get_bb_copy (loop
->header
)),
1330 cond_bb
, true_edge
, false_edge
,
1331 false /* Do not redirect all edges. */,
1332 then_scale
, else_scale
);
1334 /* loopify redirected latch_edge. Update its PENDING_STMTS. */
1335 lv_flush_pending_stmts (latch_edge
);
1337 /* loopify redirected condition_bb's succ edge. Update its PENDING_STMTS. */
1338 extract_cond_bb_edges (cond_bb
, &true_edge
, &false_edge
);
1339 lv_flush_pending_stmts (false_edge
);
1340 /* Adjust irreducible flag. */
1343 cond_bb
->flags
|= BB_IRREDUCIBLE_LOOP
;
1344 loop_preheader_edge (loop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1345 loop_preheader_edge (nloop
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1346 single_pred_edge (cond_bb
)->flags
|= EDGE_IRREDUCIBLE_LOOP
;
1351 basic_block
*bbs
= get_loop_body_in_dom_order (nloop
), after
;
1354 after
= loop
->latch
;
1356 for (i
= 0; i
< nloop
->num_nodes
; i
++)
1358 move_block_after (bbs
[i
], after
);
1364 /* At this point condition_bb is loop predheader with two successors,
1365 first_head and second_head. Make sure that loop predheader has only
1367 split_edge (loop_preheader_edge (loop
));
1368 split_edge (loop_preheader_edge (nloop
));
1373 /* The structure of loops might have changed. Some loops might get removed
1374 (and their headers and latches were set to NULL), loop exists might get
1375 removed (thus the loop nesting may be wrong), and some blocks and edges
1376 were changed (so the information about bb --> loop mapping does not have
1377 to be correct). But still for the remaining loops the header dominates
1378 the latch, and loops did not get new subloobs (new loops might possibly
1379 get created, but we are not interested in them). Fix up the mess.
1381 If CHANGED_BBS is not NULL, basic blocks whose loop has changed are
1385 fix_loop_structure (bitmap changed_bbs
)
1388 struct loop
*loop
, *ploop
;
1390 bool record_exits
= false;
1391 struct loop
**superloop
= XNEWVEC (struct loop
*, number_of_loops ());
1393 /* Remove the old bb -> loop mapping. Remember the depth of the blocks in
1394 the loop hierarchy, so that we can recognize blocks whose loop nesting
1395 relationship has changed. */
1399 bb
->aux
= (void *) (size_t) loop_depth (bb
->loop_father
);
1400 bb
->loop_father
= current_loops
->tree_root
;
1403 if (loops_state_satisfies_p (LOOPS_HAVE_RECORDED_EXITS
))
1405 release_recorded_exits ();
1406 record_exits
= true;
1409 /* Remove the dead loops from structures. We start from the innermost
1410 loops, so that when we remove the loops, we know that the loops inside
1411 are preserved, and do not waste time relinking loops that will be
1413 FOR_EACH_LOOP (li
, loop
, LI_FROM_INNERMOST
)
1420 ploop
= loop
->inner
;
1421 flow_loop_tree_node_remove (ploop
);
1422 flow_loop_tree_node_add (loop_outer (loop
), ploop
);
1425 /* Remove the loop and free its data. */
1429 /* Rescan the bodies of loops, starting from the outermost ones. We assume
1430 that no optimization interchanges the order of the loops, i.e., it cannot
1431 happen that L1 was superloop of L2 before and it is subloop of L2 now
1432 (without explicitly updating loop information). At the same time, we also
1433 determine the new loop structure. */
1434 current_loops
->tree_root
->num_nodes
= n_basic_blocks
;
1435 FOR_EACH_LOOP (li
, loop
, 0)
1437 superloop
[loop
->num
] = loop
->header
->loop_father
;
1438 loop
->num_nodes
= flow_loop_nodes_find (loop
->header
, loop
);
1441 /* Now fix the loop nesting. */
1442 FOR_EACH_LOOP (li
, loop
, 0)
1444 ploop
= superloop
[loop
->num
];
1445 if (ploop
!= loop_outer (loop
))
1447 flow_loop_tree_node_remove (loop
);
1448 flow_loop_tree_node_add (ploop
, loop
);
1453 /* Mark the blocks whose loop has changed. */
1458 if ((void *) (size_t) loop_depth (bb
->loop_father
) != bb
->aux
)
1459 bitmap_set_bit (changed_bbs
, bb
->index
);
1465 if (loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS
))
1466 create_preheaders (CP_SIMPLE_PREHEADERS
);
1468 if (loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES
))
1469 force_single_succ_latches ();
1471 if (loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
))
1472 mark_irreducible_loops ();
1475 record_loop_exits ();
1477 #ifdef ENABLE_CHECKING
1478 verify_loop_structure ();